Studies in Systems, Decision and Control 185

Andrea Bacciotti

Stability
and Control
of Linear
Systems
Studies in Systems, Decision and Control

Volume 185

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: kacprzyk@ibspan.waw.pl
The series “Studies in Systems, Decision and Control” (SSDC) covers both new
developments and advances, as well as the state of the art, in the various areas of
broadly perceived systems, decision making and control–quickly, up to date and
with a high quality. The intent is to cover the theory, applications, and perspectives
on the state of the art and future developments relevant to systems, decision
making, control, complex processes and related areas, as embedded in the fields of
engineering, computer science, physics, economics, social and life sciences, as well
as the paradigms and methodologies behind them. The series contains monographs,
textbooks, lecture notes and edited volumes in systems, decision making and
control spanning the areas of Cyber-Physical Systems, Autonomous Systems,
Sensor Networks, Control Systems, Energy Systems, Automotive Systems,
Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace
Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power
Systems, Robotics, Social Systems, Economic Systems and other. Of particular
value to both the contributors and the readership are the short publication timeframe
and the world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.

More information about this series at http://www.springer.com/series/13304

Andrea Bacciotti

Stability and Control

of Linear Systems

123
Andrea Bacciotti
Dipartimento di Scienze Matematiche
“G.L. Lagrange” (DISMA: Dipartimento
di eccellenza 2018–22)
Politecnico di Torino
Turin, Italy

ISSN 2198-4182 ISSN 2198-4190 (electronic)

Studies in Systems, Decision and Control
ISBN 978-3-030-02404-8 ISBN 978-3-030-02405-5 (eBook)
https://doi.org/10.1007/978-3-030-02405-5

Library of Congress Control Number: 2018957665

© Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To Giannina
Preface

This book is the natural outcome of a course I taught for many years at the
Technical University of Torino, first for students enrolled in the aerospace engi-
neering curriculum, and later for students enrolled in the applied mathematics
curriculum. The aim of the course was to provide an introduction to the main
notions of system theory and automatic control, with a rigorous theoretical
framework and a solid mathematical background.
Throughout the book, the reference model is a finite-dimensional, time-invariant,
multivariable linear system. The exposition is basically concerned with the
time-domain approach, but also the frequency-domain approach is taken into
consideration. In fact, the relationship between the two approaches is discussed,
especially for the case of single-input–single-output systems. Of course, there are
many other excellent handbooks on the same subject (just to quote a few of them,
[3, 6, 8, 11, 14, 23, 25, 27, 28, 32]). The distinguishing feature of the present book
is the treatment of some specific topics which are rare to find elsewhere at a
graduate level. For instance, bounded-input–bounded-output stability (including a
characterization in terms of canonical decompositions), static output feedback
stabilization (for which a simple criterion in terms of generalized inverse matrices is
proposed), controllability under constrained controls.
The mathematical theories of stability and controllability of linear systems are
essentially based on linear algebra, and it has reached today a high level of
advancement. During the last three decades of the past century, a great effort was
done, in order to develop an analogous theory for nonlinear systems, based on
differential geometry (see [7] for a historical overview). For this development,
usually referred to as geometric control theory, we have today a rich literature ([2,
5, 13, 18–20, 26, 30]). However, I believe that the starting point for a successful
approach to nonlinear systems is a wide and deep knowledge of the linear case. For
this reason, while this book is limited to the linear context, in the presentation and
organization of the material, as well as in the selection of topics, the final goal I had
in mind is to prepare the reader for such a nonlinear extension.

vii
viii Preface

Concerning the prerequisites, I assume that the reader is familiar with basic
differential and integral calculus (for real functions of several real variables) and
linear algebra. Some notions of complex analysis are required in the
frequency-domain approach. The book can be used as a reference book for basic
courses at a doctoral (or also upper undergraduate) level in mathematical control
theory and in automatic control. More generally, parts of this book can be used in
applied mathematics courses, where an introduction to the point of view of system
theory and control philosophy is advisable. The perspective of control systems and
the stability problem are indeed ubiquitous in applied sciences and witness a rapidly
increasing importance in modern engineering. At a postdoctoral level, this book can
be recommended for reading courses both for mathematician oriented to engi-
neering applications and engineers with theoretical interests. To better focus on the
main concepts and results, some more technical proofs are avoided or limited to
special situations. However, in these cases, appropriate bibliographic references are
supplied for the curious reader.
It follows a short description of the contents. The first chapter aims to introduce
the reader to the “point of view” of system theory: In particular, the notions of input–
output operator and external stability are given. The second chapter deals with
systems without external forces which reduce, according to a more classical ter-
minology, to homogeneous systems of linear differential equations. In view of the
application, we are interested in, the representation of the general integral in terms of
exponential matrix and Jordan form is crucial, and it is treated in detail. Chapter 3 is
devoted to Lyapunov stability theory of the equilibrium position of a linear unforced
system. The results reported in this chapter are classical but very important for the
following chapters. In Chap. 4, we present some alternative approaches to the rep-
resentation of solutions of a nonhomogeneous (i.e., with forcing term) system of
linear differential equations: variation of constants, undetermined coefficients,
Laplace transform. In Chap. 5 we finally begin the study of linear systems in a
control perspective. We discuss the notions of controllability and observability, their
analogies and characterizations, and the corresponding canonical forms. The final
section treats shortly the controllability problem under constrained control, in view
of possible applications to optimization theory. In Chap. 6, we address the
bounded-input–bounded-output stability problem, and we propose a characterization
using the canonical decompositions introduced in Chap. 5. Chapter 7 is devoted to
various aspects of the stabilization problem: asymptotic controllability, static state
feedback stabilization, static output feedback stabilization, dynamic output feedback
stabilization. In particular, we re-propose in a new setting some old results about
static output feedback stabilization. In author’s opinion, these results are very
interesting, but neglected in the current literature. Finally, in Chap. 8, we introduce
the frequency-domain approach and study the relationship with the time-domain
approach. Two appendices follow. In the first one, the notions of internal stability are
introduced. These notions are formulated with respect to a system of nonlinear
ordinary differential equations. In fact, only in this part of the book nonlinear
systems came into play. The reason of this choice is that all the aspects of the
Preface ix

stability notions became more evident in the nonlinear context. The second appendix
is a short list of useful facts about Laplace transform.
Finally, I wish to thank students, colleagues, and coworkers who contributed in
many ways to improve the content of this book. A special thanks to Luisa Mazzi
and Francesca Ceragioli.

Turin, Italy Andrea Bacciotti

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The Abstract Notion of System . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 The Input-Output Operator . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Discrete Time and Continuous Time . . . . . . . . . . . . . . . . 3
1.1.3 Input Space and Output Space . . . . . . . . . . . . . . . . . . . . 3
1.1.4 State Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.5 Finite Dimensional Systems . . . . . . . . . . . . . . . . . . . . . . 4
1.1.6 Connection of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.7 System Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.8 Control System Design . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.9 Properties of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Impulse Response Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Initial Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.1 Deterministic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.2 Time Invariant Systems . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.4 External Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.5 Zero-Initialized Systems and Unforced Systems . . . . . . . . 15
1.4 Differential Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.1 Admissible Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 State Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.3 Linear Differential Systems . . . . . . . . . . . . . . . . . . . . . . . 18
2 Unforced Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 The Exponential Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 The Diagonal Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 The Nilpotent Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 The Block Diagonal Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.6 Linear Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

xi
xii Contents

2.7 The Diagonalizable Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.8 Jordan Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.9 Asymptotic Estimation of the Solutions . . . . . . . . . . . . . . . . . . . 33
2.10 The Scalar Equation of Order n . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.11 The Companion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 Stability of Unforced Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . 43
3.1 Equilibrium Positions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Conditions for Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Lyapunov Matrix Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Routh-Hurwitz Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4 Linear Systems with Forcing Term . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1 Nonhomogeneous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.1 The Variation of Constants Method . . . . . . . . . . . . . . . . 54
4.1.2 The Method of Undetermined Coefficients . . . . . . . . . . . 55
4.2 Transient and Steady State . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 The Nonhomogeneous Scalar Equation of Order n . . . . . . . . . . . 59
4.4 The Laplace Transform Method . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4.1 Transfer Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4.2 Frequency Response Analysis . . . . . . . . . . . . . . . . . . . . . 66
5 Controllability and Observability of Linear Systems . . . . . . . . . . . . 69
5.1 The Reachable Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1.1 Structure of the Reachable Sets . . . . . . . . . . . . . . . . . . . 72
5.1.2 The Input-Output Map . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1.3 Solution of the Reachability Problem . . . . . . . . . . . . . . . 73
5.1.4 The Controllability Matrix . . . . . . . . . . . . . . . . . . . . . . . 75
5.1.5 Hautus’ Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2 Observability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.1 The Unobservability Space . . . . . . . . . . . . . . . . . . . . . . . 81
5.2.2 The Observability Matrix . . . . . . . . . . . . . . . . . . . . . . . . 82
5.2.3 Reconstruction of the Initial State . . . . . . . . . . . . . . . . . . 84
5.2.4 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3 Canonical Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3.1 Linear Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.2 Controlled Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.3 Controllability Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.3.4 Observability Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.3.5 Kalman Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3.6 Some Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4 Constrained Controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Contents xiii

6 External Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2 Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.3 The Case C ¼ I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.4 The General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7 Stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.1 Static State Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.1.1 Controllability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.1.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.1.3 Systems with Scalar Input . . . . . . . . . . . . . . . . . . . . . . . 114
7.1.4 Stabilizability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.1.5 Asymptotic Controllability . . . . . . . . . . . . . . . . . . . . . . . 122
7.2 Static Output Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.1 Reduction of Dimension . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.2.2 Systems with Stable Zero Dynamics . . . . . . . . . . . . . . . . 128
7.2.3 A Generalized Matrix Equation . . . . . . . . . . . . . . . . . . . 128
7.2.4 A Necessary and Sufficient Condition . . . . . . . . . . . . . . . 130
7.3 Dynamic Output Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.3.1 Construction of an Asymptotic Observer . . . . . . . . . . . . . 133
7.3.2 Construction of the Dynamic Stabilizer . . . . . . . . . . . . . . 134
7.4 PID Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8 Frequency Domain Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.1 The Transfer Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.2 Properties of the Transfer Matrix . . . . . . . . . . . . . . . . . . . . . . . . 142
8.3 The Realization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.4 SISO Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.4.1 The Realization Problem for SISO Systems . . . . . . . . . . . 148
8.4.2 External Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.4.3 Nyquist Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.4.4 Stabilization by Static Output Feedback . . . . . . . . . . . . . 157
8.5 Disturbance Decoupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Appendix A: Internal Stability Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Appendix B: Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Notations and Terminology

• We denote by N, Z, R, and C, respectively, the set of natural, integer, real, and

complex numbers. The symbol jaj denotes the modulus of a if a 2 C, and the
absolute value of a if a 2 R. The symbol sgn a denotes the sign function (i.e., a
if a  0, a if a\0). If z 2 C, z̄ Re z, im z denote respectively the conjugate, the
real part and the imaginary part of z.
• If V is a (real or complex) vector space of finite dimension with dim V = n, the
components of v 2 V with respect to a given basis are usually denoted by
v1 ; . . .; vn and we also write v ¼ ðv1 ; . . .; vn Þ. This notation is simple, but it may
give rise to some ambiguity when we deal with several vectors distinguished by
indices; in these cases, we write ðvi Þj to denote the j-th component of the i-th
vector. The subspace of V generated by a subset U  V is denoted by span U.
• If v is an element of a finite-dimensional normed vector space V, the norm of v is
generically denoted by jjvjjV or, when the space is clear from the context, simply
by jjvjj. If v is a vector of Rn, and if not differently stated, jjvjj denotes the
P
Euclidean norm, i.e., jjvjj ¼ ð ni¼1 v2i Þ1=2 .
• Let m and n be fixed. We denote MðRÞ the vector space of all the matrices with
m rows and n columns, with real entries. A similar notation with R replaced by
C is adopted for matrices with complex entries. If M 2 MðRÞ, we may also say
that M is a m  n matrix. Of course, MðRÞ can be identified with Rmn :
However, note that to this end, Rmn should be considered different from Rmn .
A matrix M is said to be square if n = m. In this case, we may also say that
M has dimension n.
To specify the entries of a matrix, we write

M ¼ ðmij Þi ¼ 1; ...; n; j ¼ 1; ...; m

(the first index specifies the row, the second one the column). As for vectors, we
may assign a norm to a matrix. In this book, we use the so-called Frobenius
norm

xv
xvi Notations and Terminology

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X
jjMjj ¼ jmij j2 :
i;j

• The identity matrix of dimension n is denoted by In , or simply by I when

the dimension is clear from the context. If M is a square matrix, we denote by M t
the transpose of M. We also denote respectively by tr M, det M, and rank M the
trace, the determinant, and the rank of M. The symbol ker M represents the
kernel of M, the symbol im M the image (range) of M. The characteristic
polynomial of a n  n matrix M is written pM ð‚Þ. It is defined by

pM ð‚Þ ¼ det ðM  ‚IÞ ¼ ð1Þn det ð‚I  MÞ:

Note in particular that deg pM ð‚Þ ¼ n (where deg Pð‚Þ denotes the degree of a
polynomial Pð‚Þ) and that for each n, ð1Þn pM ð‚Þ is a monic polynomial
(which means that the coefficient of ‚n is 1).
• Recall that the eigenvalues of a square matrix M are the roots of the charac-
teristic polynomial of M, that is, the solutions of the algebraic equation
pM ð‚Þ ¼ 0. The set of distinct eigenvalues of M constitutes the spectrum of M. It
is denoted by rðMÞ, and it is, in general, a subset of the complex plane C. Recall
also that the matrices A and B are similar if there exists a nonsingular matrix
P such that B ¼ P1 AP.
An eigenvector of M corresponding to an eigenvalue ‚ is a nontrivial solution
of the linear algebraic system ðM  ‚IÞv0 ¼ 0. The dimension of the subspace
generated by all the eigenvalues of an eigenvalue ‚ of A is called the geometric
multiplicity of ‚. The geometric multiplicity is less than or equal to the algebraic
multiplicity of ‚.
Let v0 be an eigenvector of M; the finite sequence of vectors v1 ; . . .; vk forms a
chain of generalized eigenvectors generated by v0 if ðM  ‚IÞv1 ¼ v0 ;
ðM  ‚IÞv2 ¼ v1 ; . . .; ðM  ‚IÞvk ¼ vk1 .

• If A is a subset of Rn , we denote respectively by A, A, @A the set of the interior
points of A, the closure of A, the boundary of A (in the topology of Rn ).
• If A and B are two arbitrary sets, F ðA; BÞ denotes the set of all the functions
from A to B. In particular:
– CðI; UÞ denotes the set of all the continuous functions defined in I with
values in U, where I is an interval (open or closed, bounded or unbounded)
of real numbers and U  Rn ;
– PCð½a; b; U Þ denotes the set of all the piecewise continuous,1
right-continuous functions defined on ½a; b with values in U, where a and
b are real numbers (a\b) and U  Rn ;

1
Recall that a function is piecewise continuous on a compact interval ½a; b if in this interval it has
at most finitely many discontinuity points, and each possible discontinuity point is a jump.
Notations and Terminology xvii

– PCð½a; þ 1Þ; U Þ, where a 2 R and U  Rn , denotes the set of all the

piecewise continuous2 right-continuous functions defined on the interval
½a; þ 1Þ with values in U (the sets PCðð1; þ 1Þ; UÞ and PCðð1; b;
UÞ are defined in analogous way);
– BðI; Rn Þ denotes the set of all the bounded functions defined in the interval I
with values in Rn ;
– we may use also the notations

CBð½a; þ 1Þ; Rn Þ and PCBð½a; þ 1Þ; Rn Þ

to denote the sets of all the bounded functions which belong respectively to
the sets

Cð½a; þ 1Þ; Rn Þ and PCð½a; þ 1Þ; Rn Þ:

• If the function f ð Þ is an element of a functional normed vector space V, its norm

is denoted by jj f ð ÞjjV . In particular, if f ð Þ 2 BðI; Rn Þ, we will write jj f ð Þjj1 ¼
supt2I jj f ðtÞjj (norm of the uniform convergence).
• Depending on the circumstances, for the derivative of a function f ðtÞ : R ! Rn ,
the following symbols can be used: dfdt , f 0 ðtÞ, f_ ðtÞ, ðDf ÞðtÞ. For higher-order
derivatives, we write f ðkÞ ðtÞ.
• A rational function has the form RðsÞ ¼ NðsÞ=DðsÞ where NðsÞ and DðsÞ are
polynomials. It is usually thought of as a function from C to C. A rational
function is said to be proper if degNðsÞ \ degDðsÞ. Other agreements about
rational functions will be specified later in Chap. 8 (see in particular Remark 8.2)

2
Recall that a function is piecewise continuous on a unbounded interval I if it is piecewise
continuous on every compact interval ½c; b  I.
Chapter 1
Introduction

Many phenomena observed in the real world, regardless to their different nature,
involve several physical quantities and result from the interaction of various compo-
nents: for these reasons, in these situations the term “system” is generally used.
The experimental information obtained by studying a physical system gives often
rise to the construction of a mathematical model. In this way, it can be easily commu-
nicated and elaborated qualitatively or numerically, and possibly employed to control
the evolution of the system. In this book, the term system will be often referred to
the mathematical model, rather than the represented real phenomenon.
Without any pretence of giving an axiomatic definition, the present introductory
chapter aims to describe informally the main features of the notion of system, and
the way we can take advantages of them.

1.1 The Abstract Notion of System

Direct experience shows that a system is often subject to time evolution. This means
that the numerical values of the physical quantities characterizing the state of the
system change while time passes. For this reason, they will be treated as variables.
The changes are due, in general, to the action of internal forces and constraints, as
well as of possible external forces or signals.

1.1.1 The Input-Output Operator

In order to provide an abstract description of the evolution of a system, we need to

assign the following objects:

© Springer Nature Switzerland AG 2019 1

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_1
2 1 Introduction

(1) a set T representing the time;

(2) a set U, whose members represent the numerical values of the possible external
variables (forces or signals);
(3) a set X , whose members represent the numerical values of the internal state
variables;
(4) a set Y, whose members represent the numerical values of the variables carrying
the information provided by the system about its internal state.

The action exerted on the system by the external world during the evolution
is therefore represented by a function u(·) ∈ F(T , U): it is called the input map.
The response of the system, that is the available information about the state of the
system during the evolution, is represented by a function y(·) ∈ F(T , Y): it is called
the output map. Finally, the internal state of the system, during the evolution, is
represented by a function x(·) ∈ F(T , X ), called the state evolution map. The sets
U, Y and X are respectively called the input set, the output set and the state set.
The system acts as on operator R transforming elements u(·) ∈ F(T , U) to ele-
ments y(·) ∈ F(T , Y). We will write

y(·) = R(u(·)) , with R : F(T , U) → F(T , Y) .

The operator R is called the input-output operator. In system theory, the action
of an input-output operator R is often represented graphically by means of a flow
chart, as the following figure shows.

An input-output operator is not necessarily defined for each element u(·) ∈

F(T , U). For instance, in some applications the input maps need to satisfy some
constraints, imposed by the nature of the problem either on their values or on their
functional character. The subset of F(T , U) formed by all the maps satisfying the
required conditions constitutes the domain of R. It is called the set of admissible
inputs.
The following subsections aim to specify the nature of the sets T , U, X , Y.

Remark 1.1 When we assume that the output map can be uniquely and exactly
determined by applying the operator R, we are implicitly assuming that we have a
full knowledge about the structure of the system and the physical laws governing its
evolution. But in real situations this is not always true. In practice, it might happen that
repeated experiments (with the same input map) give rise to different outputs, or that
the output is affected by imprecisions, due to one or more of the following reasons:
neglecting or simplifying some details during the modeling process; measurement
errors; uncertainty in parameters identification; random phenomena. To face these or
similar situations, suitable extensions of the theory need to be developed. But these
will not be considered in this book.
1.1 The Abstract Notion of System 3

Remark 1.2 A comment about terminology is in order. The word “system” is often
used with lightly different meanings in the common language, and sometimes also
in the technical literature. For instance, in Mathematics, “system” classically means
“set of coupled equations”. A system corresponding to an input-output operator as
above, should be more properly called an input-output system. However, throughout
this book, we prefer to use for simplicity the term “system” also in this case. The
ambiguity is not serious. The right sense can be easily understood every time from
the context.

1.1.2 Discrete Time and Continuous Time

The time can be represented by any totally ordered set T , endowed with a group
structure. In practice, we have two possible choices: either T = Z or T = R. In
the former case we speak about discrete time systems: the functions representing the
input, the state and the output are actually sequences. In the latter case we speak about
continuous time systems. It may happens that a physical system can be modeled both
as a discrete time system and as a continuous time system. This may depends on the
purposes of the search, on the measure scales and on the measure devices. Some-
times, different representations of the same physical system provide complementary
information.

1.1.3 Input Space and Output Space

We can distinguish several types of inputs variables. A disturbance is a undesired

signal, which cannot be governed, and sometimes not even measured in real time. A
reference signal is an input representing the ideal evolution, to be tracked by the real
evolution of the system. A control is an input completely determined by the decisions
of a supervisor, which can be used to modify the behavior of the system during the
evolution.
In general, we assume that the value of each single physical quantity involved in
the model can be expressed by a real number. Moreover, it is convenient to order
the input variables (and, separately, the output variables), and to rewrite them as the
components of a vector. It is therefore natural to assume that the input set U and the
output set Y are endowed with the structure of a real vector spaces.

1.1.4 State Space

In common applications, the number of the state variables is usually greater than the
number of the input and output variables. Moreover, the state variables are difficult to
identify, since in general they are not directly available to the observation. Sometimes,
4 1 Introduction

one should think of the state variables as mathematical idealizations, inherent to the
model. We will assume that also the state set X has the structure of a real vector space.

1.1.5 Finite Dimensional Systems

We say that a system is finite dimensional when the input variables, the output
variables and the state variables can be represented as vectors with finitely many real
components. Thus, for a finite dimensional system, it is natural to assume X = Rn ,
U = Rm , Y = R p , where n, m, p are given integers, greater than or equal to 1. The
sets of functions representing the input, the output and the state maps will be therefore
respectively denoted by F(R, Rm ), F(R, R p ), F(R, Rn ). In particular, the system is
said to be SISO (single-input-single-output) when m = p = 1; otherwise, the system
is said to be MIMO (multi-input-multi-output).

Remark 1.3 From now on, by the term system we mean a finite dimensional, time
continuous system.

1.1.6 Connection of Systems

In some applications, it is necessary to manage simultaneously several systems, and

to enable connections among them. The result of these manipulations may be often
reviewed as a new system. On the contrary, it may be sometimes convenient to
decompose a given system as the connection of certain subsystems. Let two systems,
whose representing operators are denoted respectively by R1 and R2 , be given. We
describe below three basic types of connections.

(1) Cascade connection. The input of the second system coincides with the output
of the first system.

If we denote by R the operator representing respectively the new resulting sys-

tem, we have R = R2 ◦ R1 , where ◦ denotes the composition of maps.
(2) Parallel connection. The first and the second system have the same input and
both contribute to determine the output.
1.1 The Abstract Notion of System 5

If we denote as before by R the operators representing the resulting system, a

typical parallel connection is obtained by taking R = R1 + R2 .
(3) Feedback connection. The output of the first system is conveyed into the second
system; it is elaborated, possibly combined with other external inputs and finally
re-injected into the input channel of the first system.

In this case, the operator R representing the result of the connection of R1 and
R2 , is implicitly defined by the relation y(·) = R1 (R2 (y(·)) + u(·)).

These types of connection can be combined, to obtain very general patterns. As

a rule, the complexity of a system becomes greater and greater, as the number of
connections increases.

1.1.7 System Analysis

The purpose of the analysis of a system is the study of the properties of the input-
output operator. For instance, it is interesting to estimate how the energy carried by
the output signal depends on the energy carried by the input signal. To this end, it
is necessary to assume that the spaces of the input maps and of the output maps
are endowed with a structure of normed vector space. For the moment we do not
need to chose a specific norm, which may depend on the particular application. For
simplicity, we continue to use the notation F(R, Rm ) and F(R, R p ) for the space
of the input maps and the space of the output maps, but remember that from now
on they are normed space. The norms on these spaces are respectively denoted by
|| · ||F (R,Rm ) and || · ||F (R,R p ) .
Informally, it is used to say that a system is externally stable when each bounded
input map generates a bounded output map. More precisely, we give the following
definition.

Definition 1.1 A system, or its representing operator R, is said to be BIBO-stable

(i.e., bounded-input-bounded-output-stable) with respect to the norms || · ||F (R,Rm )
and || · ||F (R,R p ) if for each real number R > 0 there exists a real number S > 0 such
that for each input map u(·) ∈ F(R, Rm ) one has

||u(·)||F (R,Rm ) ≤ R =⇒ ||y(·)||F (R,Rm ) ≤ S

where y(·) is the output map of the system corresponding to the input map u(·).
6 1 Introduction

Notice that according to Definition 1.1, the output is allowed to be different from
zero, even if the input vanishes.

Proposition 1.1 A system, or its representing operator R, is BIBO-stable if and

only if there exists a continuous and non-decreasing function

α(r ) : [0, +∞) → [0, +∞)

such that for each input map u(·) ∈ F(R, Rm ) one has:

||y(·)||F (R,R p ) ≤ α(||u(·)||F (R,Rm ) ). (1.1)

The meaning of (1.1) can be explained as follows: if the energy carried by the
input signal is bounded, then the energy of the output signal can be estimated in
terms of the energy of the input signal. The value of α(0) is sometimes called the
bias term, while the function α(r ) − α(0) is called the gain function.
In system analysis, it is very important to know the conditions under which a
system is BIBO-stable and, in the positive case, to give information about the shape
of the function α.

1.1.8 Control System Design

The control system design consists in the development of a control strategy, to be

exerted throughout the input channel, in such a way that the output matches a refer-
ence signal as well as possible. Roughly speaking, we can distinguish two kinds of
control strategies.

(1) Open loop control. The control is realized as a function of time u(·) ∈ F(R, Rm ),
and directly injected into the system.
(2) Closed loop control. The control is implemented by constructing a second system
and establishing a feedback connection.
The closed loop control strategy is also called automatic control. It provides
some advantages. Indeed, it enables the system to self-regulate, also in presence of
unpredictable perturbations, without the need of intervention of a human super-
visor. Let us use the term plant to denote the system to be controlled, and let
us denote by R P : F(R, Rm ) → F(R, R p ) the corresponding operator. Let us
call compensator or controller the system to be designed, and let us denote by
RC : F(R, R p ) → F(R, Rm ) the representing operator. The closed loop control
strategy consists basically in the following procedure. The output of the plant is
monitored and compared with the reference signal; when a unacceptable difference
between the two signals is detected, the compensator is activated and the necessary
corrections are sent to the plant.
1.1 The Abstract Notion of System 7

When it is possible to observe directly the state of the system, the compensator
can be realized as an operator RC : F(R, Rn ) → F(R, Rm ). We will use the terms
output feedback or state feedback when we need to distinguish the two situations.

1.1.9 Properties of Systems

In this section we aim to discuss the properties that a generic operator R : F

(R, Rm ) → F(R, R p ) is expected to satisfy in order to represent a real physical
system.

1.1.9.1 Causal Systems

Usually, systems encountered in applications are causal (or non anticipative). This
means that for each t ∈ R and for each pair of input maps u 1 (·), u 2 (·) ∈ F(R, Rm ), if

u 1 (τ ) = u 2 (τ ) for each τ ≤ t ,

then
y1 (t) = y2 (t)

where y1 (·) = R(u 1 (·)) and y2 (·) = R(u 2 (·)). In other words, the value of the output
at any instant t is determined only by the values that the input map takes at the interval
(−∞, t].

1.1.9.2 Time Invariant Systems

We say that a system, or its representing operator, is time invariant if for each t, T ∈ R
and for each input map u(·) ∈ F(R, Rm ), one has

z(t) = y(t − T )

where
v(t) = u(t − T ), y(·) = R(u(·)), z(·) = R(v(·)).

In other words, if the input signal is delayed (or anticipated) of a fixed duration,
also the output signal is delayed (or anticipated) of the same duration, but its shape
is unchanged. Time invariant systems are also called stationary, or autonomous.
8 1 Introduction

1.1.9.3 Linear Systems

A system is said to be linear when its input-output operator is linear, that is

a1 R(u 1 (·)) + a2 R(u 2 (·)) = R(a1 u 1 (·) + a2 u 2 (·))

for each pair of input maps u 1 (·), u 2 (·) ∈ F(R, Rm ) and each pair of scalars a1 , a2 .
Notice that this definition makes sense, since the input and the output sets are
vector spaces.

1.2 Impulse Response Systems

In this section we try to make more concrete the description of a continuous time,
finite dimensional system. More precisely, here we assume the existence of a matrix
h(t) with p rows and m columns, whose elements are continuous functions defined
for each t ∈ R, such that the response y(·) = R(u(·)) corresponding to an input map
u(·) ∈ F(R, Rm ) admits the representation
 +∞
y(t) = h(t − τ )u(τ ) dτ . (1.2)
−∞

Of course, here we are implicitly assuming that the integral is absolutely conver-
gent.1 A system for which such a matrix exists is called an impulse response system,
and the matrix h(t) is called an impulse response matrix. This terminology can be
explained in the following way.
Let e1 , . . . , em be the canonical basis of Rm , and let u(t) = δ(t)ei (for some
i ∈ {1, . . . , m}), where δ(t) represents the Dirac delta function (see Appendix B).
We have:
 +∞  +∞
y(t) = h(t − τ )u(τ ) dτ = h(t − τ )δ(τ )ei dτ = h(t)ei .
−∞ −∞

This shows that the response of the system to the unit impulse in the direction of
the vector ei coincides with the i-th column of the matrix h(t). Notice that for SISO
systems (i.e., with p = m = 1), h(t) is simply a real function of one real variable.
The proof of the following proposition is straightforward.
Proposition 1.2 For any impulse response system, the associated input-output oper-
ator R is linear.
In particular, it follows from Proposition 1.2 that for an impulse response system
with a vanishing input map, the output is zero for each t.

1 This may require some restrictions on the nature of the system and the set of admissible inputs.
1.2 Impulse Response Systems 9

Proposition 1.3 Each impulse response system is time invariant.

Proof Let u(t) be an input map, and let y(t) be the corresponding output map. Let
moreover T ∈ R, v(t) = u(t − T ), and let z(t) be the output corresponding to the
input v(t). We have:
 +∞  +∞
z(t) = h(t − τ )v(τ ) dτ = h(t − τ )u(τ − T ) dτ .
−∞ −∞

Setting τ − T = θ, we finally obtain:

 +∞
z(t) = h(t − T − θ)u(θ) dθ = y(t − T ).
−∞


Remark 1.4 The two previous propositions provide a complete characterization of

the class of impulse response systems. Indeed, it is possible to prove that each
time invariant, linear system is an impulse response system (see for instance [32],
pp. 152–154).


However, in general, an impulse response system is not causal.

Proposition 1.4 Let an impulse response system be given, and let h(t) be its impulse
response matrix. The following properties are equivalent.

(1) The system is causal.

(2) h(t) = 0 for each t < 0.
(3) For each input map u(·) ∈ F(R, Rm ) and each t ∈ R
 t
y(t) = h(t − τ )u(τ ) dτ .
−∞

Proof We start to prove that (1) =⇒ (2). For convenience of exposition, we first
discuss the case m = p = 1. Let t > 0 be fixed. Let

0 if τ < t
u 1 (τ ) = 0 and u 2 (τ ) =
sgn h(t − τ ) if τ ≥ t

for τ ∈ R, be two input maps (note that u 2 depends on t).

Since u 1 (τ ) = u 2 (τ ) for τ < t and the system is causal, we have y1 (t) = y2 (t).
On the other hand, it is evident that y1 (t) = 0, while
 +∞  +∞
y2 (t) = h(t − τ )u 2 (τ ) dτ = |h(t − τ )| dτ .
−∞ t
10 1 Introduction

We are so led to conclude that h(t − τ ) = 0 for each τ > t, that is h(r ) = 0 for
r < 0.
If m or p (or both) are not equal to 1, the proof is technically more complicated, but
the basic idea is the same. One starts by fixing a pair of indices i, j, (i = 1, . . . , p, j =
1, . . . , m). As before, we chose u 1 (τ ) = 0 for each τ ∈ R, which implies that the
corresponding output vanishes identically. Next we define u 2 (τ ) component-wise,
according to the following rule: if l = j, then (u 2 )l (τ ) = 0 for each τ ∈ R, while

0 if τ < t
(u 2 ) j (τ ) =
sgn h i j (t − τ ) if τ ≥ t.

The causality assumption implies that the output corresponding to u 2 (τ ) must

be identically zero, as well. On the other hand, the i-th component of the output
corresponding to u 2 (τ ) is
 +∞
(y2 )i (t) = h i j (t − τ ) · (u 2 ) j (τ ) dτ
−∞
 +∞  +∞
= h i j (t − τ ) · sgn h i j (t − τ ) dτ = |h i j (t − τ )| dτ
t t

which is zero only if h i j (r ) vanishes for each r ∈ (−∞, t). The conclusion is
achieved, by repeating the argument for each choice of i, j.
The proof that (2) =⇒ (3) is straightforward. Thus, it remains to prove that (3)
=⇒ (1). Let t ∈ R be fixed. If u 1 , u 2 are input maps such that u 1 (τ ) = u 2 (τ ) for
each τ ≤ t, then the corresponding output maps y1 , y2 satisfy
 t  t
y1 (t) = h(t − τ )u 1 (τ ) dτ = h(t − τ )u 2 (τ ) dτ = y2 (t).
−∞ −∞

Hence, the system is causal. 

A further simplification in the representation of impulse response systems is pos-

sible, if we limit ourselves to input maps u(·) ∈ F(R, Rm ) satisfying the following
condition: there exists t0 ∈ R such that u(τ ) = 0 for each τ < t0 . Indeed, in such a
case, we will have y(t0 ) = 0 and, for each t > t0 ,
 t
y(t) = h(t − τ )u(τ ) dτ .
t0

For impulse response systems, there is also a simple characterization of the

external stability property. Recall that the Frobenius norm of a real matrix M =
(m i j )i=1,..., p, j=1,...,m satisfies the following inequality:

||M|| ≤ |m i j |. (1.3)
i, j
1.2 Impulse Response Systems 11

Proposition 1.5 Let an impulse response system be given, and let h(t) be its impulse
response matrix. Let us assume in addition that the system is causal. Let B(R, Rm )
and B(R, R p ) be respectively, the input maps and the output maps space, both
endowed with the uniform convergence norm. The system is BIBO-stable if and only
if the integral  +∞
||h(r )|| dr
0

t
is convergent or, equivalently, if and only if the function 0 ||h(r )|| dr is bounded for
t ∈ [0, +∞).

Proof Since the system is causal, for each t ∈ R we have:

 t  t
||y(t)|| = || h(t − τ )u(τ ) dτ || ≤ ||h(t − τ )u(τ )|| dτ
−∞ −∞
 t
≤ ||h(t − τ )|| · ||u(τ )|| dτ
−∞
 t
≤ ||h(t − τ )|| dτ · ||u(·)||∞ .
−∞

By the substitution t − τ = r , we get

 t  +∞
||h(t − τ )|| dτ = ||h(r )|| dr.
−∞ 0

 +∞
Hence, if 0 ||h(r )|| dr =  < ∞, from the previous computation we obtain

||y(t)|| ≤ ||u(·)||∞

for each t ∈ R and, finally, ||y(·)||∞ ≤ ||u(·)||∞ . The BIBO-stability condition will
be therefore satisfied taking, for each R > 0, S = R.
As far as the reverse implication is concerned, let us consider first the case m =
p = 1. Assuming that the system is BIBO-stable, let us fix t > 0 and define the input
map 
0 if τ < 0
ũ(τ ) =
sgn h(t − τ ) if τ ∈ [0, t]

(notice that ũ(τ ) depends on t). Let ỹ(t) be the corresponding output. Invoking again
the causality assumption, we have:
 t  t  t
ỹ(t) = h(t − τ )ũ(τ ) dτ = |h(t − τ )| dτ = |h(r )| dr. (1.4)
−∞ 0 0
12 1 Introduction

Since |ũ(τ )| ≤ 1 for each t and τ , applying the BIBO-stability condition with
R = 1 we find a constant S > 0 such that

| ỹ(t)| ≤ || ỹ(·)||∞ ≤ S (1.5)

 +∞
for each t > 0. From (1.4) and (1.5) we infer that the integral 0 |h(r )| dr is
convergent.
With some notational complication, the proof extends to the general case where m
or p (or both) are greater than 1. Since there are similarities with the proof of Propo-
sition 1.4, we limit ourselves to sketch the reasoning. Let h i j (t) (i = 1, . . . , p, j =
1, . . . , m) be the elements of the matrix h(t), and let t > 0. Let a pair of indices i, j
be fixed, and define the input map ũ(τ ) = (ũ 1 (τ ), . . . , ũ m (τ )) by taking ũ l (τ ) ≡ 0
if l = j and 
0 if τ < 0
ũ j (τ ) =
sgn h i j (t − τ ) if τ ∈ [0, t].

Let finally ỹ(t) = ( ỹ1 (t), . . . , ỹ p (t)) the corresponding output map. Using the
causality hypothesis we have
 t  t
ỹi (t) = h i j (t − τ )ũ j (τ ) dτ = |h i j (r )| dr. (1.6)
0 0

Since ũ(·)∞ ≤ 1, the BIBO-stability condition allows us to determine a real

number S such that
|| ỹ(t)|| ≤ S (1.7)

for each t > 0. Clearly, ỹi (t) = | ỹi (t)| ≤  ỹ(t). As a consequence of (1.6) and (1.7)
we conclude that  t
|h i j (r )| dr ≤ S
0

for each t > 0. Finally, by virtue of (1.3), we have

 t  t
h(r ) dr ≤ |h i j (r )| dr ≤ pm S
0 i, j 0

for each t > 0. The conclusion easily follows. 

1.3 Initial Conditions

Representing a system as an operator R : F(R, Rm ) → F(R, R p ) is a very simple

and attractive idea, but it is not realistic. In common applications indeed, the inputs
are not known on the whole time axis, but only starting from some instant t0 ∈ R,
1.3 Initial Conditions 13

assumed as the initial instant. Moreover, we are interested to study the behavior of
the system in the future, that is for t ≥ t0 . In these cases, in order to compensate the
loss of information about the inputs for t < t0 , we need to assume the assignment of
the initial state, that is the value x0 ∈ Rn assumed by the state variable at the initial
instant t0 .

1.3.1 Deterministic Systems

We may image that the initial condition i.e., the pair (t0 , x0 ) ∈ R × Rn , summarizes
the past history of the system. It is also reasonable to presume that the assignment of
the initial data, together with the assignment of the input map for t ≥ t0 , is sufficient to
determine uniquely the future evolution of the system. This is actually an assumption,
similar to the causality assumption, but more appropriate to the new point of view.
Definition 1.2 We say that a system, or its representing operator, is deterministic if
for each t0 ∈ R,

u 1 (t) = u 2 (t) ∀t ≥ t0 and x1 (t0 ) = x2 (t0 ) =⇒ y1 (t) = y2 (t) ∀t ≥ t0

where xi (t), yi (t) are respectively the state evolution map and the output map cor-
responding to the input map u i (t), i = 1, 2.
Note that the deterministic hypothesis basically differs from the causality assump-
tion, since it explicitly involves the state of the system. When a system is determinis-
tic, it is convenient to interpret the input-output operator as an “initialized” operator
R(t0 , x0 )(u(·)), mapping functions2

u(·) ∈ F([t0 , +∞), Rm )

to functions
y(·) ∈ F([t0 , +∞), R p ).

We write also y(·) = R(t0 , x0 )(u(·)).

Remark 1.5 According to Definition 1.2, the so-called delayed systems (systems
whose behavior for t ≥ t0 depends not only on the state of the system at the initial
instant t0 , but also on the values assumed by the state variable on some interval
[t0 − θ, t0 ], (θ > 0) and, more generally, systems with memory, cannot be considered
deterministic. 
When we want to make use of the notion of initialized operator, the definitions of
time invariant system and of linear system need to be appropriately modified.

2 Alternatively, we may agree that the admissible inputs are restricted to functions u(·) ∈ F (R, Rm )
vanishing for t < t0 .
14 1 Introduction

1.3.2 Time Invariant Systems

A system represented by a deterministic initialized operator is time invariant if for

each t0 , T ∈ R, each x0 ∈ Rn , and each input map u(·), denoting

v(t) = u(t − T ), y(·) = R(t0 , x0 )(u(·)), z(·) = R(t0 + T, x0 )(v(·))

one has
z(t) = y(t − T ).

Proposition 1.6 Let R be a time invariant, deterministic initialized operator, and

let y(·) = R(t0 , x0 )(u(·)). Then,

y(t) = z(t − t0 )

where z(·) = R(0, x0 )(v(·)), and v(t) = u(t + t0 ).

In other words, dealing with a time invariant operator, we may assume, without
loss of generality, that the initial instant coincides with the origin of the time axis.

1.3.3 Linear Systems

A system represented by means of a deterministic initialized operator R, is lin-

ear if for each t0 ∈ R, R is linear as a map from Rn × F([t0 , +∞), Rm ) to
F([t0 , +∞), R p ), that is if for each t0 ∈ R, and for each choice of the pairs
x1 , x2 ∈ Rn , u 1 (·), u 2 (·) ∈ F([t0 , +∞), Rm ), and a1 , a2 ∈ R one has

R (t0 , a1 x1 + a2 x2 )(a1 u 1 (·) + a2 u 2 (·))

= a1 R(t0 , x1 )(u 1 (·)) + a2 R(t0 , x2 )(u 2 (·)).

Proposition 1.7 Let R be a linear, time invariant, deterministic initialized operator.

For each t0 ∈ R, each x0 ∈ Rn and each u(·) ∈ F([t0 , +∞), Rm ) one has:

y(·) = R(t0 , x0 )(0) + R(t0 , 0)(u(·)).

Proof By applying the definition of linear initialized system with x1 = x0 , x2 = 0,

u 1 (·) = 0, u 2 (·) = u(·), a1 = a2 = 1, we immediately have

y(·) = R(t0 , x0 )(u(·)) = R(t0 , x0 )(0) + R(t0 , 0)(u(·))

as required. 
1.3 Initial Conditions 15

1.3.4 External Stability

We now update Definition 1.1, for the case of systems represented by initialized
operators.
Definition 1.3 A system represented by a deterministic initialized operator R, is
BIBO-stable (uniformly with respect to the initial instant) if for each real number
R > 0 there exists a real number S > 0 such that for each t0 ∈ R and each input map
u(·) ∈ B([t0 , +∞), Rm ) we have

||x0 || ≤ R, ||u(·)||∞ ≤ R =⇒ ||y(·)||∞ ≤ S

where y(·) = R(t0 , x0 )(u(·)).

1.3.5 Zero-Initialized Systems and Unforced Systems

We may interpret Proposition 1.7 by saying that the response of a linear system
corresponding to some initial state x0 and some input map u(·) can be decomposed
as the sum of
• the response corresponding to the initial state x0 when the the input is set to be
zero;
• the response corresponding to the input u(·) when the initial state is set to be zero.
In other words, when analyzing the behavior of a linear system, and more precisely
when we are interested in the study of the external stability, the way the response is
affected by the initial data and the way it is affected by the inputs can be analyzed
separately. We will frequently refer to this principle in this book.
Therefore, in the study of linear systems we may conveniently distinguish two
different steps. In the first step we may assume that the input vanishes, while in the
second step we may assume that the initial state vanishes. In this way, we will be
also able to recover some analogies with the theory of the impulse response systems.
We say that a deterministic system represented by a time invariant initialized
operator is zero-initialized (or initialized at zero) if the initial state x0 at the instant
t0 = 0 is set to be zero. We say that a deterministic system represented by a time
invariant initialized operator is unforced if the input map is set to be equal to zero
for each t ≥ 0.
Unforced systems may present a non-zero evolution in time: indeed, because of the
energy stored in the system at the initial instant, the initial state does not coincide,
in general, with a rest point. In these circumstances, we expect that the unforced
system evolves in such a way that the initial energy is dissipated, by approaching
a rest point asymptotically. If this really happens, we will say informally that the
system is internally stable. A more precise and formal definition of internal stability
will be given later.
16 1 Introduction

In the analysis of the qualitative properties of a system, the study of the behavior
when the forcing terms are provisionally suppressed, is an essential preliminary step.
As we shall see, the properties of internal stability and external stability are intimately
related.

1.4 Differential Systems

In this section we focus on systems which are modeled by means of ordinary differ-
ential equations; they will be called differential systems. This class of systems is very
important, because of the variety and the large amount of applications. Moreover,
a well developed and complete theory is available, for these systems. However, its
introduction requires some restrictions.

1.4.1 Admissible Inputs

Dealing with differential systems, by admissible input map we mean a function

u(·) ∈ PC([t0 , +∞), Rm ), for some t0 ∈ R. For some applications, it is necessary
to limit further the admissible inputs: a typical choice is u(·) ∈ PC([t0 , +∞), U ),
where U is a given nonempty, bounded subset of Rm . The role of U is to represent
possible limitations on the energy available in order to exert the control. Notice that
if U is bounded, PC([t0 , +∞), U ) is not a vector space.

1.4.2 State Equations

Let f (t, x, u) : R × Rn × Rm → Rn and h(t, x) : R × Rn → R p be given func-

tions. A differential system is defined by the equations

dx
ẋ = = f (t, x, u) (1.8)
dt

y = h(t, x). (1.9)

Equation (1.8) is also called the state equation, while h(t, x) is called the obser-
vation map. For each admissible input map u(t), (1.8) becomes a system of ordinary
differential equations of the first order in normal form

ẋ = f (t, x, u(t)) = g(t, x). (1.10)

1.4 Differential Systems 17

Concerning the functions f and h, it is customary to make the following assump-

tions:
(A1) f is continuous with respect to the pair of variables (x, u); the first partial
derivatives of f with respect to all the components xi of the state vector x
exist and are continuous with respect to the pair of variables (x, u); moreover,
f is piecewise continuous with respect to t;
(A2) h is continuous with respect to x, and piecewise continuous with respect to t;
(A3) for each admissible input u(t), there exist continuous, positive real functions
a(t), b(t) such that

|| f (t, x, u(t))|| ≤ a(t)||x|| + b(t)

for each (t, x) ∈ R × Rn .

Under these assumptions, for each pair of initial values (t0 , x0 ) ∈ R × Rn and for
each admissible input map u(t) there exists a unique solution of the Cauchy problem

ẋ = g(t, x)
(1.11)
x(t0 ) = x0

defined on the whole interval3 [t0 , +∞). When we want to emphasize the dependence
of the solution of the problem (1.11) on the initial conditions and on the input map,
we will use the notation
x = x(t; t0 , x0 , u(·)). (1.12)

When the dependence on the initial conditions and on the input map is clear
from the context, we may also use the simplified notation x = x(t). The initialized
input-output operator associated to the differential system (1.8), (1.9)

y(·) = R(t0 , x0 )(u(·)) (1.13)

is given by y(t) = h(t, x(t; t0 , x0 , u(·))) for t ≥ t0 . By analogy with (1.12), some-
times we may use the notation

y = y(t; t0 , x0 , u(·)). (1.14)

The following proposition summarizes the previous remarks.

Proposition 1.8 Under the hypotheses (A1), (A2), (A3), the differential system
(1.8), (1.9) defines a deterministic input-output operator on the set of admissible
input maps. Moreover, the output map is continuous.

3 Provided that the input is defined for each t ∈ R, existence and uniqueness of solutions is actually
guaranteed on (−∞, +∞).
18 1 Introduction

Next proposition characterizes the differential systems which possess the time
invariance property.

Proposition 1.9 Assume that (A1), (A2), (A3) hold. The input-output operator
(1.13) defined by the differential system (1.8), (1.9) is time invariant if the functions
f e h do not depend explicitly on t, that is f (t, x, u) = f (x, u) and h(t, x) = h(x).

Proof Let t0 ∈ R and let u(t) ∈ PC([t0 , +∞), Rm ). Assume that an initial state x0
is given; let x(t) be the corresponding solution of (1.8) and let y(t) = h(x(t)). Let
finally T be a fixed real number. Setting v(t) = u(t − T ) and ξ(t) = x(t − T ), we
have
d d
ξ(t) = x(t − T ) = f (x(t − T ), u(t − T )) = f (ξ(t), v(t)).
dt dt

In other words, ξ(t) coincides with the solution corresponding to the translated
input map v(t) and to the initial condition (t0 + T, x0 ). Setting finally z(t) = h(ξ(t)),
it is clear that z(t) = y(t − T ). 

By virtue of Propositions 1.6 and 1.9, if the functions f and h do not depend
explicitly on t we may assume t0 = 0 without loss of generality. In this case, the
notation (1.12) and (1.14) can be simplified, by avoiding the explicit indication of
the initial instant.

1.4.3 Linear Differential Systems

In the mathematical theory of differential systems, a prominent role is played by

systems whose state equations are linear. The importance of linear systems is also
supported by their interest in applications.

Definition 1.4 A time invariant differential system is said to be linear if there

exist real matrices A, B, C of respective dimensions n × n, n × m, p × n, such
that f (x, u) = Ax + Bu and h(x) = C x.

In other words, a system is linear in the sense of Definition 1.4 when it can be
written in the form 
ẋ = Ax + Bu
(1.15)
y = C x.

Proposition 1.10 If a system is linear in the sense of Definition 1.4, then the asso-
ciated input-output initialized operator (1.13) is linear.

The proof of Proposition 1.10 will be given later. Beginning with Chap. 2, we
focus our attention on the study of linear, time invariant differential systems.
1.4 Differential Systems 19

Chapter Summary
The first part of this chapter constitutes a short introduction to systems theory. The
basic notions of input, output and state variables are presented in abstract terms, as
well as the notion of input-output operator. We discuss the main properties involved
in the investigation of a system, and illustrate how distinct systems can be combined
to give rise to a new system. In this framework, we also introduce the main concern
of this book: how to exploit the input channel in order to control the evolution of a
system.
The exposition becomes more concrete in the remaining part of the chapter, where
we explain how a system can be represented by certain mathematical models: impulse
response, state space equations. The role of initial conditions is emphasized, in a
deterministic philosophy, in connection with the notion of state variable.
Chapter 2
Unforced Linear Systems

In this chapter we undertake a systematic study of finite dimensional, unforced,

linear, time invariant differential systems. They are defined by a system of ordinary
differential equations of the form

ẋ = Ax , x ∈ Rn . (2.1)

According to the mathematical tradition, (2.1) is called a linear homogeneous

system of differential equations (with constant coefficients). In extended notation,
(2.1) reads ⎧
⎪
⎨ẋ1 = a11 x1 + · · · + a1n xn
... ... ...
⎪
⎩
ẋn = an1 x1 + · · · + ann xn .

For a general system of ordinary differential equations, the notion of solution is

recalled in Appendix A. In force of the special form of (2.1) the solutions enjoy some
special properties.

2.1 Prerequisites

In this section we recall some important facts, concerning a system of equations of

type (2.1) and its solutions.
Fact 1. For each initial state x0 there exists a unique solution x = ϕ(t) of system
(2.1) such that ϕ(0) = x0 ; moreover, ϕ(t) is defined for each t ∈ R.
Fact 2. If v ∈ Rn (v = 0) is an eigenvector of A corresponding to the eigenvalue
λ ∈ R, then ϕ(t) = eλt v represents the solution of (2.1) corresponding to the
initial state v.
© Springer Nature Switzerland AG 2019 21
A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_2
22 2 Unforced Linear Systems

Fact 3. If ϕ1 (·), ϕ2 (·) are solutions of (2.1) and α1 , α2 ∈ R, then also α1 ϕ1 (·) +
α2 ϕ2 (·) is a solution of (2.1).
Fact 4. Let ϕ1 (·), . . . , ϕk (·) be k solutions of (2.1). The following statements are
equivalent:
• there exists t̄ ∈ R such that the vectors ϕ1 (t̄), . . . , ϕk (t̄) are linearly independent
in Rn ;
• the functions ϕ1 (·), . . . , ϕk (·) are linearly independent, as elements of the space
C(−∞, +∞, Rn );
• for each t ∈ R, the vectors ϕ1 (t), . . . , ϕk (t) are linearly independent, as elements
of the space Rn .
When one of the above equivalent conditions holds, we simply say that ϕ1 (·), . . . ,
ϕk (·) are linearly independent.
Fact 5. The set of all the solutions of the system (2.1) forms a subspace S of
C(−∞, +∞, Rn ). The dimension of S is finite and, more precisely, it is equal to
n. The subspace S is also called the general integral of system (2.1).
Notice that system (2.1) makes sense even if we allow that x takes value into the
n-dimensional complex space Cn , and that the entries of A are complex numbers:
apart from some obvious modifications, all the previous facts remain valid.1 Actually,
to this respect we may list some further properties.
Fact 6. If the elements of A are real, and if ϕ(·) is a solution of (2.1) with nonzero
imaginary part, then the conjugate function ϕ(·) is a solution of (2.1), as well.
Fact 7. If the elements of A are real, and if ϕ(·) is a solution of (2.1) with nonzero
imaginary part, then ϕ(·) and ϕ(·) are linearly independent; in addition,

ϕ(·) + ϕ(·) ϕ(·) − ϕ(·)

ϕ1 (·) = and ϕ2 (·) = (2.2)
2 2i
are two real and linearly independent solutions of (2.1).
If A is a matrix with real elements and with a complex eigenvalue λ = α + i β
(β = 0) associated to an eigenvector v = u + i w, we dispose of the complex solution
ϕ(t) = eλt v. Then, using (2.2), we obtain the representation

ϕ1 (t) = eαt [(cos βt)u − (sin βt)w] , ϕ2 (t) = eαt [(cos βt)w + (sin βt)u] .

Remark 2.1 The existence of non-real eigenvalues implies therefore the existence
of real oscillatory solutions. In particular, if α = 0 and β = 0, the eigenvalues
are purely imaginary, and we have periodic solutions with minimal period equal
to 2π/β. 

1 The convenience of extending the search for the solutions to the complex field even if the elements

of A are real numbers, is suggested by Fact 2: possible eigenvalues of A represented by conjugate

complex numbers (with nonzero imaginary part) generates solutions which, otherwise, would be
difficult to identify.
2.1 Prerequisites 23

We are now able to conclude that the general integral of system (2.1) can be
written as a linear combination

ϕ(t) = c1 ϕ1 (t) + · · · + cn ϕn (t) (2.3)

where c1 , . . . , cn represent arbitrary constants, and ϕ1 , . . . , ϕn represent n arbitrary

solutions, provided that they are linearly independent. If A is real, it is not restrictive
to assume that ϕ1 , . . . , ϕn are real valued: hence, Eq. (2.3) describes either the space
of all the real solutions when the constants c1 , . . . , cn are taken in the real field, or
the space of all the complex solutions when the constants c1 , . . . , cn are taken in the
complex field.
A set formed by n linearly independent solutions of the system (2.1) is called a
fundamental set of solutions. To each fundamental set of solutions ϕ1 , ..., ϕn , we
associate a fundamental matrix

(t) = (ϕ1 (t), . . . , ϕn (t))

whose columns are formed by the components of the vectors ϕ1 (t), ..., ϕn (t). Notice
that if (t) is a fundamental matrix and Q is a constant, nonsingular matrix, then
also (t)Q is a fundamental matrix. From this remark, it follows easily that, for each
t0 ∈ R, there exists a unique fundamental matrix such that (t0 ) = I . This is also
called the principal fundamental matrix relative to t0 . The principal fundamental
matrix relative to t0 = 0 will be simply called principal fundamental matrix.
Let us introduce the constant vector c = (c1 , . . . , cn )t . If (t) is any fundamental
matrix, we can rewrite (2.3) as
ϕ(t) = (t)c (2.4)

The particular solution satisfying the initial conditions ϕ(t0 ) = x0 can be recov-
ered by solving the algebraic system

(t0 )c = x0

with respect to the unknown vector c. If (t) is the principal fundamental matrix
relative to t0 , we simply have c = x0 .

2.2 The Exponential Matrix

Let M(C) be the finite dimensional vector space formed by the square matrices M =
(mij )i,j=1,...,n of dimensions n × n with complex entries, endowed with the Frobenius
norm. It is possible to prove that the series
24 2 Unforced Linear Systems

∞
 Mk
,
k!
k=0

converges for each M ∈ M(C) (see for instance [17], p. 83). Its sum is denoted eM
and it is called the exponential matrix of M . We list below the main properties of the
exponential matrix.
• If the entries of M are real, then the entries of eM are real.
• e0 = I , where 0 denotes here a matrix whose entries are all equal to zero and I is
the identity matrix.
• eN +M = eM eN , provided that MN = NM .
• The eigenvalues of eM are the complex numbers of the form eλ , where λ is an
eigenvalue of M .
• eM M = MeM .
• det eM = etr M . As a consequence, det eM = 0 for each M .
−1
• If P is a nonsingular matrix, eP MP = P −1 eM P.

Let us come back to system (2.1). For each t ∈ R, all the entries of the matrix etA
are of class C 1 . Moreover, the following proposition holds.

Proposition 2.1 For each A ∈ M(C) and each t ∈ R, we have

d tA
e = AetA .
dt
Thus, the exponential matrix provides a useful formalism, which allows us to
represent the solutions of the system (2.1). Indeed, if x = ϕ(t) is the solution of (2.1)
such that ϕ(t0 ) = x0 , then by using the uniqueness of solutions and the properties of
the exponential matrix, we get

ϕ(t) = e(t−t0 )A x0 .

If t0 = 0, we simply have
ϕ(t) = etA x0 (2.5)

for each t ∈ R. In other words, computing the exponential matrix is equivalent to

compute a fundamental matrix of the system (actually, the principal fundamental
matrix).
In the following sections, we will see how to realize an explicit construction of
the exponential matrix. The final result will be achieved through several steps. We
start by examining some special situations.
2.3 The Diagonal Case 25

2.3 The Diagonal Case

Let ⎛ ⎞
λ1 0 ... 0
⎜0 λ2 ... 0⎟
⎜
A=⎝ ⎟ = diag (λ1 , . . . , λn )
0 ... ... 0⎠
0 0 ... λn

where λ1 , . . . , λn are not necessarily distinct numbers (real or complex).

Remark 2.2 For such a matrix, λ is an eigenvalue if and only if λ = λi for some
i = 1, . . . , n, and the algebraic multiplicity of λ indicates how many λi ’s are equal to
λ. The eigenvectors corresponding to λ1 , . . . , λn can be taken respectively coincident
with the vectors of the canonical basis
⎛ ⎞ ⎛ ⎞
1 0
⎜0⎟ ⎜0 ⎟
⎜ ⎟ ⎜ ⎟
v1 = ⎜ . ⎟ , . . . , vn = ⎜ . ⎟ . (2.6)
⎝ .. ⎠ ⎝ .. ⎠
0 1


A fundamental set of solutions of (2.1) can be therefore written in the form

ϕ1 (t) = eλ1 t v1 , . . . , ϕn (t) = eλn t vn .

A system (2.1) defined by a diagonal matrix A is called decoupled, since the

evolution of each component xi of x depends on xi , but not on xj with j = i. A system
of this type can be trivially solved by integrating separately the single equations.
The fundamental set of solutions obtained by this method obviously coincides by
the previous one. The same fundamental set of solutions can be obtained also by
computing the exponential matrix. Indeed, it is easy to check that for each positive
integer k,
Ak = diag (λk1 , . . . , λkn ) ,

hence
etA = diag (eλ1 t , . . . , eλn t ) .

2.4 The Nilpotent Case

If A is nilpotent, there exists a positive integer q such that Ak = 0 for each k ≥ q.

Thus, the power series which defines the exponential matrix reduces to a polynomial
and can be computed in elementary way. A typical nilpotent matrix (for which q = n)
is
26 2 Unforced Linear Systems
⎛ ⎞
0 1 0 ... 0
⎜0 0 1 ... 0⎟
⎜ ⎟
⎜. . . ... ... ... . . .⎟
A=⎜
⎜0
⎟ (2.7)
⎜ 0 0 ... 1⎟ ⎟
⎝0 0 0 ... 0⎠
.

The direct computation of the exponential matrix shows that if A has the form
(2.7), then ⎛ ⎞
t2 t n−1
1 t 2! . . . (n−1)!
⎜ t n−2 ⎟
⎜0 1 t . . . (n−2)! ⎟
⎜ ⎟
⎜ .
e = ⎜. . . . . .. ⎟
.
⎟ .
tA
⎜ . . . ⎟
⎝0 0 0 . . . t ⎠
0 0 0 ... 1

Alternatively, we can write the corresponding system

⎧
⎪
⎪ ẋ1 = x2
⎪
⎨ẋ = x
2 3
⎪
⎪ ...
⎪
⎩
ẋn = 0

and solve it by cascaded integration (from down to top). The two approaches obvi-
ously lead to the same result. A fundamental set of solutions can be written in the
form

t n−1
ϕ1 (t) = v1 , ϕ2 (t) = tv1 + v2 , . . . , ϕn (t) = v1 + · · · + tvn−1 + vn
(n − 1)!
(2.8)
where the vectors v1 , . . . , vn are as in (2.6) the vector of the canonical basis.

Remark 2.3 Notice that zero is the unique eigenvalue of the matrix (2.7); the cor-
responding proper subspace is one dimensional. Moreover, Av1 = 0 (which means
that v1 is an eigenvector of A), Av2 = v1 , Av3 = v2 and so on. 

The general integral of the system defined by the matrix (2.7) can be written as

t n−1
ϕ(t) = c1 ϕ1 (t) + · · · + cn ϕn (t) = d1 + td2 + · · · + dn
(n − 1)!
2.4 The Nilpotent Case 27

where ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
c1 c2 cn
⎜ ⎟
c2 ⎜c3 ⎟ ⎜0⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟.. ⎜.⎟ ⎜.⎟
d1 = ⎜ ⎟ , d2 = ⎜ .. ⎟ , . . . , dn = ⎜ .. ⎟ .
.
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝cn−1 ⎠ ⎝cn ⎠ ⎝0⎠
cn 0 0

Notice that Ad1 = d2 , Ad2 = d3 , . . . , Adn = 0. Notice also that d1 can be arbitrar-
ily chosen, and that dn is an eigenvector of A, regardless to the choice of d1 .

Remark 2.4 Combining the methods used for the cases of diagonal and nilpotent
matrices, we are able to compute the exponential matrix for each matrix A of the
form λI + T where λ is any real number, I is the identity matrix of dimensions n × n,
and T is nilpotent. In particular, if T has the form (2.7), then
⎛ ⎞
t2 t n−1
1 t ... (n−1)!
⎜ 2!
t n−2 ⎟
⎜0 1 t ... ⎟
⎜
λt ⎜ .
(n−2)! ⎟
et(λI +T ) = e ⎜. .. .. .. ⎟ . (2.9)
⎟
⎜. . . . ⎟
⎝0 0 0 ... t ⎠
0 0 0 ... 1

2.5 The Block Diagonal Case

If M is block diagonal, that is

⎛ ⎞
M1 0 ... 0
⎜0 M2 . . . 0 ⎟
⎜ ⎟
M =⎜ . .. .. ⎟ = diag (M1 , . . . , Mk ) ,
⎝ .. . . ⎠
0 0 . . . Mk

then also its exponential matrix is block diagonal

eM = diag (eM1 , . . . , eMk ) .

The exponential matrix of M is easily obtained, provided that we are able to

construct the exponential matrix of every block Mi .
28 2 Unforced Linear Systems

2.6 Linear Equivalence

To address the problem of computing the exponential matrix in the general case, we
need to introduce the concept of linear equivalence.
Let us image system (2.1) as the mathematical model of a process evolving in a
real vector space V of dimension n, where a basis has been fixed. The state of the
system is represented, in this basis, by the n-tuple x = (x1 , . . . , xn )t .
Assume that a new basis of V is given, and let y = (y1 , . . . , yn )t be the components
of the state in this new basis. As well known, there exists a nonsingular matrix P
such that for each element of V ,
x = Py.

We want to see how (2.1) changes, when the state is represented in the new basis.
We have
ẏ = P −1 ẋ = P −1 APy = By . (2.10)

We therefore obtain again a linear system, defined by a matrix B which is similar

to the given matrix A. Vice versa, two systems of the type (2.1) defined by similar
matrices can be always thought of as two representations of the same system in two
different systems of coordinates.

Definition 2.1 Two systems

ẋ = Ax and ẏ = By , x ∈ Rn , y ∈ Rn

are said to be linearly equivalent if A and B are similar, that is if B = P −1 AP for

some nonsingular matrix P.

The previous definition is actually an equivalence relation. It is clear that each

solution x = ϕ(t) of the first system is of the form ϕ(t) = Pψ(t) where y = ψ(t) is
a solution of the second one and vice-versa. On the other hand, it is easy to see that

etB = P −1 etA P (or, equivalently, etA = PetB P −1 ) . (2.11)

Hence, as far as we are interested in solution representation, we can work with

any system linearly equivalent to the given one, and finally we can use (2.11) in order
to come back to the original coordinates.
The notion of linear equivalence, as well as the notion of similar matrices, can be
immediately generalized to the case where x ∈ Cn . Of course, if A and B are similar
matrices, A is real and B contains complex elements, then the matrix P determining
the change of basis must contain complex elements, as well.
2.7 The Diagonalizable Case 29

2.7 The Diagonalizable Case

It is well known that a matrix A of dimensions n × n is diagonalizable (that is,

similar to a diagonal matrix) if and only if there exist n linearly independent vectors
v1 , . . . , vn such that each vi , i = 1, . . . , n, is an eigenvector of A. In such a case,
we say that the vectors v1 , . . . , vn constitute a proper basis of A. In particular, A is
diagonalizable if it admits n distinct eigenvalues.
Denoting by P the matrix whose columns are v1 , . . . , vn (in this order), we have

P −1 AP = diag (λ1 , . . . , λn ) = D

where λ1 is the eigenvalue of A corresponding to v1 , λ2 is the eigenvalue of A

corresponding to v2 and so on (it is not required that the numbers λ1 , . . . , λn are
distinct).
To compute etA we can proceed in the following way: first we diagonalize A by
means of the change of coordinates determined by P, then we compute etD , and
finally we come back to the original coordinates, making use of (2.11).

Remark 2.5 If A is real but it admits complex eigenvalues, then P and D will have
complex elements, as well. However, by construction, the elements of etA must be
real.
Notice that (t) = PetD is a fundamental matrix; its computation do not require
to know the inverse of P. However, in general the elements of PetD are not real, not
even if A is real.
In conclusion, to determine explicitly the elements of the matrix etA and hence
the general integral of (2.1) in the diagonalizable case, it is sufficient to know the
eigenvalues of A and the corresponding eigenvectors.

Example 2.1 Let us consider the system

ẋ1 = −x2
ẋ2 = x1

defined by the matrix  

0 −1
A= .
1 0
 
i
The eigenvalues of A are +i , with eigenvector , and −i , with eigenvector
1
 
−i
. It is easy to identify two (complex conjugate) linearly independent solutions
1
     
i −sint cos t
ϕ1 (t) = e it
= +i
1 cos t sin t
30 2 Unforced Linear Systems

and      
i −sint cos t
ϕ2 (t) = e−i t = −i .
1 cos t sin t

Taking their real and imaginary parts we obtain two linearly independent real
solutions    
−sint cos t
ψ1 (t) = and ψ2 (t) = .
cos t sin t

Alternatively, we can apply the diagonalization procedure. To this end, we need

to compute the inverse matrix of
 
−i i
P=
1 1

given by  
1 1 −i
P −1 = − .
2i −1 −i

We easily get  
i 0
D = P −1 AP = ,
0 −i

and  it 
e 0
etD = .
0 e−i t

Finally,  
−1 cos t −sint
e = Pe P
tA tD
= .
sin t cos t

In this case, the exponential matrix could be also obtained directly, by applying
the definition; indeed, it is not difficult to see that
 
10
A4 = .
01

2.8 Jordan Form

In this section, for any n × n matrix A, we denote by λ1 , . . . , λk (1 ≤ k ≤ n) its

distinct eigenvalues. For each eigenvalue λi of A, by μi and νi we denote respectively
the algebraic and geometric multiplicity of λi (1 ≤ νi ≤ μi ). Moreover, we will write
λi = αi + i βi .
2.8 Jordan Form 31

If A possesses eigenvalues with algebraic multiplicity greater than one and with
geometric multiplicity less than the algebraic multiplicity, then A is not diagonaliz-
able. In other words, the number of linearly independent eigenvectors is not sufficient
to form a basis of the space. To overcome the difficulty, we resort to generalized
eigenvectors. The following theorem holds (see for instance [4]).

Theorem 2.1 Each matrix A of dimension n × n is similar to a block-diagonal

matrix of the form ⎛ ⎞
C1,1 0 . . . 0
⎜ 0 C1,2 . . . 0 ⎟
⎜ ⎟
J =⎜ . .. .. ⎟
⎝ .. . . ⎠
0 0 . . . Ck,νk

where the blocks Ci,j are square matrices of the form

⎛ ⎞
λi 1 0 ... 0
⎜0 λi 1 ... 0⎟
⎜ ⎟
⎜ .. .. .. .. ⎟
Ci,j = ⎜
⎜. . . .⎟⎟ .
⎜ .. .. .. ⎟
⎝. . . ... 1⎠
0 0 0 . . . λi

Only one eigenvalue appears in each block, but a single eigenvalue can appear
in more than one block. More precisely, for each eigenvalue λi there are exactly νi
blocks, and each block is associated to one proper eigenvector. The dimension of a
block Ci,j equals the length of the chain of generalized eigenvectors originating from
the j-th eigenvector associated to λi . The eigenvalue λi appears exactly μi times on
the principal diagonal of J .

The matrix J is called a Jordan form of A. From our point of view, it is important
to remark that each block J has the form λi I + T , where I is the identity matrix
(of appropriate dimension), and T is the nilpotent matrix of type (2.7). Taking into
account the conclusions of Sect. 2.5, the strategy illustrated for the case of a diago-
nalizable matrix can be therefore extended to the present situation: we transform the
given system (2.1) to the system
ẏ = Jy (2.12)

by means of a suitable change of coordinates, then we find the solutions of (2.12)

directly (or, alternatively, we compute etJ , and we come back to the original coor-
dinates). It remains only the problem of identifying the matrix P which determines
the similarity between A and J . To this purpose, as already sketched, we need to
determine for each eigenvalue λi , a number of linearly independent eigenvectors and
generalized eigenvectors equal to μi . These vectors must be indexed in accordance
to the order of the indices of the eigenvalues and, for each eigenvector, in accordance
with the order of generation of the generalized eigenvectors of a same chain.
32 2 Unforced Linear Systems

λ1
  
v1,1,0 v1,1,1 . . . v1,2,0 v1,2,1 . . . v1,3,0 v1,3,1 . . . . . .
        
eigenvector eigenvector eigenvector
      
first chain second chain third chain
λ2
  
v2,1,0 v2,1,1 . . . . . . . . .
  
eigenvector
  
first chain

The set of all these vectors constitutes a basis of the space, called again a proper
basis. The columns of the matrix P are formed by the vectors of a proper basis in the
aforementioned order, that is

P = (v1,1,0 | v1,1,1 | . . . | v1,2,0 | v1,2,1 | . . . | v1,3,0 | v1,3,1 | . . .

. . . | v2,1,0 | v2,1,1 | . . .) .

Another proper basis and another corresponding Jordan form can be obtained by
permutations of the order of the eigenvalues or, for each eigenvalue, permutations
of the order of the corresponding eigenvectors (but leaving unchanged the order of
generation of the generalized eigenvectors). In this sense, the Jordan form is not
unique.
After that a proper basis has been constructed and provided that the order of the
various indices is correctly settled out, we have all the information we need in order
to explicitly write the Jordan form. In fact, we do not need to perform the change of
coordinates. However, the computation of P and P −1 is inevitable in order to recover
etA in the original coordinates. The computation of P −1 can be avoided, if we may
limit ourselves to write the fundamental (in general, complex) matrix PetJ .
Keeping in mind (2.9), and the procedure illustrated in Sect. 2.1 (Fact 7), we can
resume the conclusions achieved so far in the following proposition.

Proposition 2.2 The generic element ϕr,s (t) of the matrix etA (r, s = 1, . . . , n)
reads as

k
ϕr,s (t) = (Zr,s )i (t)eλi t
i=1

where each term (Zr,s )i (t) is a polynomial (in general, with complex coefficients)
whose degree is (strictly) less than the algebraic multiplicity of λi , and λi is an
eigenvalue of A (i = 1, . . . , k).
If A is real, the generic element ϕr,s (t) of the matrix etA can be put in the form


k
 
ϕr,s (t) = eαi t (pr,s )i (t) cos βi t + (qr,s )i (t) sin βi t (2.13)
i=1
2.8 Jordan Form 33

where(pr,s )i and (qr,s )i are polynomials with real coefficients whose degree is
(strictly) less than the algebraic multiplicity of λi (of course, the previous formula
includes also the contributions of the real eigenvalues, for which βi = 0).

2.9 Asymptotic Estimation of the Solutions

To our purposes, one of the main applications of the conclusions of the previous
section is the estimation of the asymptotic behavior of the solutions of (2.1) for
t → +∞.

Lemma 2.1 For each ε > 0 and each integer m ∈ N there exists a constant k > 0
such that t m < keεt , for each t ≥ 0.

Proof The proof can be carried on by mathematical induction. If m = 1 we can take

k = 1ε . Indeed, setting
eεt
f (t) = −t
ε

we have f (0) = 1ε and f  (t) = eεt − 1 > 0 for t > 0. Let us assume that the result
holds for m − 1, with k = k̄. The function

f (t) = keεt − t m

is such that
 
 εt kε εt
f (0) = k and f (t) = kεe − mt m−1
=m e − t m−1 >0
m

for t > 0, provided that we choose k = mk̄

ε
. 

Let α0 be the maximum of the real parts αi of the eigenvalues λi of the matrix A
(i = 1, . . . , k) and let α be any real number greater than α0 :

α > α0 ≥ αi for each (i = 1, . . . , k) .

Since | sin βi t| ≤ 1 and | cos βi t| ≤ 1 for each i = 1, . . . , k, starting from (2.13)

and using repeatedly the triangular inequality we get, for t ≥ 0,


k
  k
|ϕr,s (t)| ≤ eαi t |(pr,s )i (t)| + |(qr,s )i (t)| ≤ (Qr,s )i (t)eαi t
i=1 i=1

where (Qr,s )i is a polynomial whose coefficients are nonnegative real numbers, which
majorize the absolute values of the corresponding coefficients of the polynomials
34 2 Unforced Linear Systems

(pr,s )i (t) and (qr,s )i (t). Even if not essential for the subsequent developments, we
note that the degree of (Qr,s )i is less than the algebraic multiplicity of λi .
Let 0 < ε < α − α0 . By Lemma 2.1, there are constants  kr,s such that |ϕr,s (t)| ≤

(α0 +ε)t αt 2 αt
kr,s e ≤ kr,s e for each t ≥ 0. Hence e = tA
r,s ϕr,s (t) ≤
2
r,s kr,s e
and, finally,
etA ≤ k0 eαt ∀t ≥ 0

where k0 is a new constant.

Note that if all the eigenvalues λi whose real parts are exactly equal to α0 (i.e.,
αi = Re λi = α0 ) have algebraic multiplicity coincident with the geometric multi-
plicity, then the previous inequality holds even when α is replaced by α0 . Indeed,
in this case the corresponding polynomials (pr,s )i (t) and (qr,s )i (t) reduce to con-
stants. Hence, the term (Qr,s )i (t)eαi t can be directly majorized by eα0 t , apart from
a multiplicative constant, without need of using Lemma 2.1. Concerning the eigen-
values λi for which αi = Re λi < α0 , we may apply Lemma 2.1 with ε = α0 − αi .
The corresponding terms (Qr,s )i (t)eαi t can therefore be majorized, apart form some
multiplicative constants, by eαi t eεt = eα0 t . Summing up, we can state the following
proposition.

Proposition 2.3 Let A be a real matrix. For each α > α0 , there exists k0 > 0 such
that
etA ≤ k0 eαt ∀t ≥ 0 . (2.14)

If all the eigenvalues of A with real part equal to α0 have the algebraic multiplicity
coincident with the geometric multiplicity, then in (2.14) we can take α = α0 .

From (2.14) it follows

etA c ≤ k0 c eαt , t ≥ 0 (2.15)

for each real constant vector c.

2.10 The Scalar Equation of Order n

The scalar differential equation (with constant coefficients, n > 1)

y(n) + a1 y(n−1) + · · · + an−1 y + an y = 0 (2.16)

can be thought of as a particular case of (2.1). Indeed, setting

y = x1 , y = x2 , . . . , y(n−1) = xn
2.10 The Scalar Equation of Order n 35

and using (2.16) we have

x1 = y = x2
x2 = y = x3
............
xn = y(n) = −a1 xn − . . . − an x1

that is, with vector notation,

ẋ = Cx (2.17)

where we set x = (x1 , . . . , xn )t , and

⎛ ⎞
0 1 0 ... 0
⎜ 0 0 1 ... 0 ⎟
⎜ ⎟
⎜ .. .. .. .. ⎟ .
C=⎜ . . . . ⎟ (2.18)
⎜ ⎟
⎝ 0 0 0 ... 1 ⎠
−an −an−1 . . . . . . −a1

By solution of (2.16) we obviously mean a n-times continuously differentiable

function y(t) : R → R for which (2.16) is identically satisfied for each t ∈ R. The
problem of determining the general integral (that is, the set of all the solutions) of
(2.16) is clearly equivalent to the problem of finding n linearly independent solutions
of (2.17).
A matrix exhibiting the structure (2.18) is called a companion matrix. More pre-
cisely, we say that (2.18) is the companion matrix associated to the Eq. (2.16).
The special structure of C displayed by (2.18) allows us to identify immediately
an important algebraic object, invariant under similarity. Indeed, using mathematical
induction,
 it is easy to check that the characteristic polynomial of C is pC (λ) =
(−1)n λn + a1 λn−1 + · · · + an . We are especially interested in the eigenvalues of
C, which are the roots of pC (λ); hence, the coefficient (−1)n can be neglected. In
fact, it is customary (even is slightly confusing) to call (−1)n pC (λ) the characteristic
polynomial of the differential equation (2.16). From now on, we adopt the notation

pch (λ) = (−1)n pC (λ) = λn + a1 λn−1 + · · · + an . (2.19)

Note that pch (λ) is monic for each n, and that it can be immediately written,
without need of transforming (2.16) in the equivalent system (2.17), by replacing
formally y by λ and reinterpreting the orders of the derivatives as powers. It is also
customary to say that

pch (λ) = λn + a1 λn−1 + · · · + an = 0 (2.20)

 36 2 Unforced Linear Systems

is the characteristic equation of the differential equation (2.16), and that its solutions
are the characteristic roots of (2.16).
Now we change the point of view. Let A be an arbitrary n × n real matrix, and let
 
pA (λ) = (−1)n λn + a1 λn−1 + · · · + an . (2.21)

its characteristic polynomial. Write a matrix CA of the form (2.18), reporting in the
last row the coefficients a1 , . . . , an taken from (2.21). In this way, A and CA will have
the same characteristic polynomial and hence the same eigenvalues (with the same
algebraic multiplicity). The matrix CA is called the the companion matrix associated
to A. Unfortunately, in general, A and CA need not to be similar. For instance,   the
n
i n
characteristic polynomial of the identity matrix is pI (λ) = (−1) λi . We
i=0 i
may write the associated companion matrix CI ; however the identity matrix I is not
similar to CI , the class of equivalence of I under the similarity relation being just the
singleton {I }. It follows that not all the systems of linear differential equations in Rn
can be reduced by linear transformations to a scalar equation of order n.
The following theorem provides conditions ensuring that a given matrix A is
similar to its associated matrix CA in companion form. This theorem has its own
interest from an algebraic point of view, but it is also very important for our future
developments.

Theorem 2.2 Let A be a square matrix of dimensions n × n. The following properties

are equivalent.

(i) A is similar to its associated matrix in companion form.

(ii) rank (A − λI ) = n − 1 for each eigenvalue λ of A.
(iii) The geometric multiplicity of each eigenvalue of A is equal to 1.
(iv) The characteristic polynomial of A coincides with its minimal polynomial.
(v) There exists a vector v = 0 such that the n vectors

v, Av, A2 v, . . . , An−1 v

are linearly independent.

The complete proof of Theorem 2.2 can be found for instance in [22]. To our future
purposes, the equivalence between (i) and (v) is especially important.2 A vector v
enjoying the property stated in (v) is said to be cyclic for A.
Thus, if the n × n matrix A satisfies one of the assumptions of Theorem 2.2, solving
the linear system defined by A is actually equivalent to solve a differential equation
of order n of the form (2.16). Property (iii) of Theorem 2.2 implies in particular that
for each eigenvalue λ of A there is a unique eigenvector v and hence a unique chain
of possible generalized eigenvectors engendered by v.

2 For reader’s convenience, a proof of this equivalence will be given in the next section.
2.10 The Scalar Equation of Order n 37

Of course, statements (i),…,(v) are fulfilled by any matrix C assigned in compan-

ion form. It follows that the general integral of (2.16) can be obtained as a linear
combination of the n functions

eλ1 t , teλ1 t , . . . , t μi −1 eλ1 t

......................
eλk t , teλk t , . . . , t μk −1 eλk t

where λ1 , . . . , λk are the distinct roots of the characteristic Eq. (2.20) and μ1 , . . . , μk
the respective algebraic multiplicities.
Example 2.2 Let us consider in detail the case of a linear equation of order 2

y + ay + by = 0 . (2.22)

To write the general integral y(t), first we need to discuss the characteristic equa-
tion
λ2 + aλ + b = 0 . (2.23)

If (2.23) has two distinct real solutions λ1 , λ2 , then we have

y(t) = c1 eλ1 t + c2 eλ2 t . (2.24)

If (2.23) has a unique real solution λ1 = λ2 = λ of multiplicity 2, then we have

y(t) = (c1 + tc2 )eλt . (2.25)

Finally, if (2.23) has complex (not real) conjugate solutions3 α ± i β, then we

have

y(t) = (c1 cos βt + c2 sin βt)eαt . (2.26)

The behavior of the solutions for t ≥ 0 depends on the signs of λ1 and λ2 in the
case (2.24) and, respectively, an the signs of λ and α in the cases (2.25) (2.26).
For instance, if λ1 , λ2 < 0 [respectively, λ < 0, α < 0] the energy initially stored
in the system (measured by the initial conditions) is dissipated:

3 The expression (2.26) results from the application of formulæ (2.2). Alternatively, (2.26) can be
obtained starting from the complex version of (2.24)

k1 eλt + k2 eλt

using the fact that eα±i β = eαt (cos βt ± i sin βt) and setting

c 1 = k1 + k2 c2 = −i (k1 − k2 ) .

In particular, c1 and c2 turn out to be real if we take k2 = k 1 .

38 2 Unforced Linear Systems

Fig. 2.1 Dissipation in the 2

case of real characteristic
roots 1.5

0.5

−0.5

−1

−1.5

−2
0 2 4 6 8 10 12 14 16 18 20

Fig. 2.2 Dissipation in the 2

case of complex
characteristic roots 1.5

0.5

−0.5

−1

−1.5
0 2 4 6 8 10 12 14 16 18 20

• monotonically in the cases (2.24) and (2.25), after possible initial picks, whose
occurrence depends on the choice of c1 and c2 (Fig. 2.1);
• with oscillatory decay in the case (2.26) (Fig. 2.2).

√ The case a = 0 and b > 0 is a particular instance of (2.26) with α = 0 and β =

b: the solutions are periodic with minimal period 2π/β. The general integral takes
the form

y(t) = c1 cos βt + c2 sin βt = ρ cos(βt + θ) (2.27)

∈ [0, 2π) are identified by the relations c1 = ρ cos θ, c2 = ρ sin θ. The

where ρ and θ 
numbers ρ = c12 + c22 and θ are called amplitude and phase of the periodic function
(2.27). The inverse of the minimal period is called the frequency. Note that in (2.27),
2.10 The Scalar Equation of Order n 39

the frequency depends on b while the amplitude depends on the initial conditions and
remains constant. In other words, in this case we have conservation of the energy.
The reader can easily check that these conclusions agree with those of Example 2.1.

Remark 2.6 Let us denote by D the derivative operator. Formally, (2.16) can be
rewritten as
L(D)y = (Dn + a1 Dn−1 + · · · + an )y = 0

where (−1)n L(D) = pch (D). Notice that L(D) acts as a linear operator.

2.11 The Companion Matrix

In this section we show that a matrix A is similar to the associated matrix in companion
form if and only if there exists a cyclic vector for A, that is a vector v = 0 such that
v, Av, . . . , An−1 v form a basis of Rn (this proves the equivalence of statements (i)
and (v) of Theorem 2.2).

Lemma 2.2 Let C be a matrix in companion form, and let λ1 , . . . , λn its eigenvalues
(not necessarily distinct). Then, C is similar to a matrix of the form
⎛ ⎞
λ1 1 0 ... 0 0
⎜0 λ2 1 ... 0 0⎟
⎜ ⎟
⎜ .. ⎟ .
M = ⎜ ... ..
.
..
.
..
. .⎟ (2.28)
⎜ ⎟
⎝0 0 0 . . . λn−1 1⎠
0 0 0 ... ... λn

Proof Let us start with Eq. (2.16). Let us show that by means of suitable linear
substitutions, (2.16) can be transformed in a system of first order linear equations
defined by the matrix (2.28). Let us set

ξ1 = y, ξ2 = p2,1 y + y , . . . , ξn = pn,1 y + · · · + pn,n−1 y(n−2) + y(n−1)

where the coefficients pi,j are recovered by the relations

p2,1 y + y = (−λ1 + D)y

p3,1 y + p3,2 y + y = (−λ1 + D)(−λ2 + D)y
.............
pn,1 y + · · · + pn,n−1 y(n−2) + y(n−1) = (−λ1 + D) · · · · · (−λn−1 + D)y
40 2 Unforced Linear Systems

and D is the derivation operator. We have

ξ1 = y = y + p2,1 y − p2,1 y = (−λ1 + D)y + λ1 y = λ1 ξ1 + ξ2

ξ2 = (p2,1 y + y ) =
= D(−λ1 + D)y + λ2 (−λ1 + D)y − λ2 (−λ1 + D)y =
= (−λ2 + D)(−λ1 + D)y + λ2 ξ2 = λ2 ξ2 + ξ3
...............
ξn = D(pn,1 y + · · · + pn,n−1 y(n−2) + y(n−1) ) =
= D(−λn−1 + D) · · · · · (−λ1 + D)y =
= (−λn + D) · · · · · (−λ1 + D)y + λn (−λn−1 + D) · · · · · (−λ1 + D)y .

The term (−λn + D) · · · · · (−λ1 + D)y vanishes, since it coincides with (2.16).
Hence we get
ξn = λn ξn .

The statement easily follows.

We emphasize that (2.28) is not a Jordan form of C (it coincides with the Jordan
form of C, only in the case where λ1 = λ2 = · · · = λn ). Let P be the matrix such
that C = P −1 MP. Then P has the form
⎛ ⎞
1 0 0 ... 0 0
⎜p2,1 1 0 . . . 0 0⎟
⎜ ⎟
⎜ .. .. .. .. .. ⎟
⎝ . . . . .⎠
pn,1 pn,2 pn,3 . . . pn,n−1 1

where the numbers pij are the same as in the proof of Lemma 2.2. Let us remark that
the companion form is not the unique way to rewrite (2.16) as a system of first order
equations. We can take for instance
⎧
⎪
⎪ z1 = an−1 y + an−2 y + · · · + a1 y(n−2) + y(n−1)
⎪
⎪
⎪
⎪  (n−3)
+ y(n−2)
⎨z2 = an−2 y + an−3 y + · · · + a1 y
.............
⎪
⎪
⎪zn−1 = a1 y + y
⎪
⎪
⎪
⎩z = y .
n
2.11 The Companion Matrix 41

Then we have
⎧
⎪
⎪ z1 = an−1 y + an−2 y + · · · + a1 y(n−1) + y(n) =
⎪
⎪
⎪
⎪ = −an y = −an zn
⎪
⎪
⎪
⎪
⎪
⎪ z2 = an−2 y + · · · + a1 y(n−2) + y(n−1) =
⎪
⎨ = an−2 y + · · · + a1 y(n−2) + z1 −
⎪
⎪ −(an−1 y + an−2 y + · · · + a1 y(n−2) ) =
⎪
⎪
⎪
⎪ = z1 − an−1 y = z1 − an−1 zn
⎪
⎪
⎪
⎪
⎪
⎪ ............
⎪
⎩
zn = y = zn−1 − a1 y = zn−1 − a1 zn .

This system corresponds to the matrix

⎛ ⎞
0 0 . . . 0 −an
⎜1 0 . . . 0 −an−1 ⎟
⎜ ⎟
Ct = ⎜ . .. .. .. ⎟ .
⎝ .. . . . ⎠
0 0 . . . 1 −a1

Since all these substitutions are linear and invertible, we have actually proved that
C and C t are similar (as a matter of fact, this is true for every square matrix).
We are now able to conclude the proof. Let us assume that A is similar to its
companion form CA . We know by Lemma 2.2 that A is similar to the matrix M given
by (2.28), as well.
Let w = (0, . . . , 0, 1)t . The result of the multiplication M w is a vector coin-
ciding with the last column of M . Let us perform the iterated multiplications
M 2 w = M (M w), M 3 w = M (M 2 w), . . . and let us form a new matrix whose
columns are given by the vectors w, M w, . . . , M n−1 w, in this order:
⎛ ⎞
0 0 0 ... 1
⎜0 0 0 . . . ∗⎟
⎜ ⎟
⎜ .. .. .. .. ⎟
⎜. . . .⎟
⎜ ⎟
⎜0 0 1 . . . ∗⎟
⎜ ⎟
⎝0 1 λn + λn−1 . . . ∗⎠
1 λn λ2n ... ∗

where we denoted by ∗ some unessential functions of λ1 , . . . , λn . This matrix is not

singular, which means that w, M w, . . . , M n−1 w are linearly independent. Now let
P be the matrix transforming A in M , and let v = Pw. We have

(w|M w| . . . |M n−1 w) = (P −1 Pw|P −1 APw| . . . |P −1 An−1 Pw) =

= P −1 (v|Av| . . . |An−1 v)
42 2 Unforced Linear Systems

which yields the desired conclusion. Vice versa, we finally prove that A is similar to CAt
(the transpose of the companion form of A) provided that the condition (v) of Theorem
2.2 holds. Setting R = (v, Av, . . . , An−1 v), we have to prove that R−1 AR = CAt or,
equivalently,

AR = (Av|A2 v| . . . |An v) = RCAt =

⎛ ⎞
0 0 ... 0 −an
⎜1 0 . . . 0 −an−1 ⎟
⎜ ⎟
⎜ 0 −an−2 ⎟
= (v|Av| . . . |An−1 v) · ⎜0 1 . . . ⎟ .
⎜ .. .. .. .. ⎟
⎝. . . . ⎠
0 0 . . . 1 −a1

The computations are not difficult (for the last column we need to apply the
Cayley-Hamilton Theorem). We already know that a matrix in companion form and
its transpose are similar, but we can also proceed in a direct way. Indeed, it is sufficient
to remark that a matrix in companion form satisfies (v) with v = (0, . . . , 0, 1)t . By
repeating the same computations as before, we recover the required similarity.
Chapter Summary
This chapter is devoted to the mathematical problem of representing the solutions of
a homogeneous system of linear differential equations by means of suitable explicit
formulæ. This corresponds to the study of the qualitative behavior of a system when
the evolution depends only on the internal forces and the external inputs are switched
off. It is actually the first step in the investigation of the properties of a system.
Chapter 3
Stability of Unforced Linear Systems

In this chapter we continue our investigation of the properties of linear unforced

differential systems
ẋ = Ax (3.1)

where A is a square n × n matrix with real entries, and x ∈ Rn . We focus in particular

on the stability problem.

3.1 Equilibrium Positions

The equilibrium positions of system (3.1) coincide with the solutions of the algebraic
equation Ax = 0. Hence, a system of the form (3.1) always have an equilibrium
position for x = 0. Such an equilibrium position is unique (and hence isolated) if
and only if det A = 0. Otherwise, there are infinitely many equilibrium positions
(none of which isolated): more precisely, the set of all the equilibrium positions of
(3.1) constitutes a subspace of Rn .

Remark 3.1 Assume that there is a point x̄ = 0 such that A x̄ = 0. Then, x̄ is a stable
equilibrium position for system (3.1) if and only if the origin is a stable equilibrium
position for system (3.1). Indeed, setting y = x − x̄, we have

ẏ = ẋ = Ax = Ax − A x̄ = Ay.


The displacements of x with respect to x̄, that is the displacement of y with respect
to y = 0, are determined by the same system which determines the displacements of
x with respect to x = 0.

© Springer Nature Switzerland AG 2019 43

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_3
44 3 Stability of Unforced Linear Systems

Remark 3.2 If an equilibrium point is attractive for system (3.1), then it must be
isolated. Hence, if A is singular, there exist no attractive equilibrium positions. In
other words, if system (3.1) possesses an attractive equilibrium position x̄, then
x̄ = 0, and there are no other equilibria x̄ = 0. 

According to the previous remarks, when studying stability and asymptotic sta-
bility of linear systems, it is not restrictive to limit ourselves to the origin.

Proposition 3.1 If the origin is stable [respectively, asymptotically stable] for sys-
tem (3.1), then the origin is stable [respectively, asymptotically stable] for all the
systems linearly equivalent to (3.1).

Proof Let B = P −1 A P and let ψ(t) be any solution of the system ẏ = By. Let us
fix ε > 0, and let ε = ε/P −1 . Since (3.1) is stable at the origin, there exists δ  > 0
such that ϕ(0) < δ  =⇒ ϕ(t) < ε for each t ≥ 0 and each solution ϕ(t) of
(3.1). Let δ = δ  /P and ϕ(t) = Pψ(t). Then,

ψ(0) < δ =⇒ ϕ(0) = Pψ(0) ≤ P · ψ(0) < δ 

so that
ψ(t) = ||P −1 ϕ(t)|| ≤ P −1  · ϕ(t) < ε .

Finally, assume that limt→+∞ ϕ(t) = 0 for a given solution of (3.1). Then, for
each σ > 0 there exists T > 0 such that

t > T =⇒ ϕ(t) < σ 

where σ  = σ/P −1 , and this implies that ψ(t) ≤ P −1  · ϕ(t) < σ. The rea-
soning is easily completed. 

3.2 Conditions for Stability

For linear time invariant systems, the analysis of the stability properties can be carried
on by means of purely algebraic tools.
Definition 3.1 We say that a real square matrix A possesses the Hurwitz property if
all the eigenvalues of A have (strictly) negative real part.
In short, when A possesses the Hurwitz property we shall also say that A is a
Hurwitz matrix. Note that every Hurwitz matrix is nonsingular.

Theorem 3.1 If A is a Hurwitz matrix then the origin is a globally and exponen-
tially stable equilibrium point for system (3.1). If the origin is a locally attractive
equilibrium point for system (3.1), then A is a Hurwitz matrix.
3.2 Conditions for Stability 45

Proof Assume that all the eigenvalues of A have negative real part. Then we can
choose α and k0 in (2.15) in such a way that α0 < α < 0. The global and exponential
attraction of the origin trivially follows. As far as the stability property is concerned,
we may use again (2.15). As already noticed, it is not restrictive to take α < 0;
for t ≥ 0, we have therefore eαt ≤ 1. Hence, for each ε > 0, it is sufficient to take
δ = ε/k0 .
We now pass to the second statement. Being the origin locally attractive, there
exists a neighborhood  of the origin such that all the solutions issuing from a point
of  asymptotically approach zero when t → +∞. We proceed by distinguishing
several cases.
Assume first that there is an eigenvalue λ with strictly positive real part. If λ is
real and if v is a corresponding (real) eigenvector, then we can construct a solution
of the form eλt v. Note that the norm of v can be taken arbitrarily small. Instead, if
λ = α + i β is not real, then we can construct a solution of the form

eαt [(cos βt)u + (sin βt)w] ,

where u, w are certain real vectors, whose norm can be taken arbitrarily small, and
α > 0. In both cases, these solutions are unbounded for t ≥ 0. This contradicts the
assumptions.
In similar way we exclude the existence of eigenvalues λ with zero real part.
Indeed, in this case either λ = 0, so that there is a nonzero constant solution, or λ is
purely imaginary, so that we can construct a periodic solution (cos βt)u + (sin βt)w,
which is bounded but does not approach zero. 
From Theorem 3.1 and its proof we can infer other information, which can be
resumed in the following way.
• For the linear system (3.1), the condition that A possesses the Hurwitz property is
necessary and sufficient for the asymptotic stability of the origin.
• If the origin is locally attractive for the linear system (3.1), then it is globally and
exponentially attractive, as well.
• For a linear system, if the origin is locally attractive then it is also stable.
Instead, even in the case of a linear system it may happen that the origin is stable
but not attractive; very simple examples are given by the system ẋ = 0 with x ∈ R,
whose solutions are constant, and by the system in Example 2.1, whose solutions are
periodic.
From the proof of Theorem 3.1, we also immediately see that if there exists an
eigenvalue of A with strictly positive real part then the origin is unstable. Thus, it
remains to discuss the case where all the eigenvalues of A have non-positive real
part, and at least one among them has a real part exactly equal to zero.
Theorem 3.2 The following statements are equivalent.
(i) All the eigenvalues of A have non-positive real part, and for each possible
eigenvalue with zero real part, the algebraic multiplicity and the geometric
multiplicity coincide.
46 3 Stability of Unforced Linear Systems

(ii) The norm of the exponential matrix et A is bounded for t ≥ 0.

(iii) The origin is stable for system (3.1).
Proof (i) =⇒ (ii). If all the eigenvalues of A have nonpositive real part and the
possible eigenvalues with zero real part have the same algebraic and geometric mul-
tiplicity, then we can apply (2.14) with α = α0 = 0. The conclusion is straightfor-
ward.
(ii) =⇒ (iii). If there is a constant k0 > 0 such that ||et A || ≤ k0 for t ≥ 0, then
for every x0 ∈ Rn we have
et A x0  ≤ k0 x0  . (3.2)

The definition of stability is recovered taking δ = ε/k0 .

Finally we prove by contradiction that (iii) =⇒ (i). We already know that if the
origin is stable, there exist no eigenvalues with strictly positive real part. Assume
that there is an eigenvalue λ with zero real part and whose geometric multiplicity is
less than its algebraic multiplicity.
If λ = i β con β = 0, we could construct a complex solution of the form (cos βt +
i sin βt)(tu + v), where v is an eigenvector corresponding to λ, and u is a generalized
eigenvector; both v and u can be chosen of arbitrarily small norm. But then, we could
also find a real solution

(cos βt)[tu 1 + v1 ] − (sin βt)[tu 2 + v2 ]

where v1 , v2 , u 1 , u 2 are some real vectors. This solution corresponds to the initial
state x0 = v1 . Since u 1 and u 2 cannot be both zero, the solution exhibits an oscillatory
behavior and the amplitude of the oscillations increases as t increases. This solution
is not bounded for t ≥ 0, so that the stability assumption is contradicted.
The case λ = 0 can be ruled out in similar way. 
We may also prove the following proposition by analogous arguments.
Proposition 3.2 The following statements are equivalent.
(i) All the eigenvalues of A have nonpositive real part, and for each possible eigen-
value with zero real part, the algebraic multiplicity and the geometric multiplicity
coincide.
(ii) All the solutions of the system are bounded on t ≥ 0.
Remark 3.3 If the system at hand is defined by a scalar differential equation of order
n like (2.16), the stability conditions of the equilibrium position y = y  = . . . =
y (n−1) = 0 can be stated in terms of its characteristic roots. 

3.3 Lyapunov Matrix Equation

In this section we present a different characterization of stable linear systems. Recall

that a real symmetric matrix P is said to be:
3.3 Lyapunov Matrix Equation 47

• positive definite if for each 0 = x ∈ Rn we have x t P x > 0.

• positive semidefinite if x t P x ≥ 0 for each x ∈ Rn .
If the conditions above are fulfilled by the matrix −P, then we say that, respec-
tively, P is negative definite, negative semidefinite. Finally, we say that P is indefinite
if x t P x takes both strictly positive and strictly negative values.

Theorem 3.3 The following statements are equivalent.

(i) A possesses the Hurwitz property.
(ii) There exists a positive definite, real symmetric matrix P such that for each
nontrivial solution ϕ(t) of (3.1), we have

d 
g(t) = −||ϕ(0)||2 (3.3)
dt t=0

where V (x) = x t P x and g(t) = V (ϕ(t)).

(iii) There exists a positive definite, real symmetric matrix P such that

At P + P A = −I . (3.4)

Proof (i) =⇒ (ii). Assume that (i) holds and denote by pi j the unknown elements
of the matrix P. Let ψ1 (t), . . . , ψn (t) be the columns of the exponential matrix et A ,
and let us define  +∞
pi j = ψit (s)ψ j (s)ds .
0

The numbers pi j are well defined: indeed, if the eigenvalues of A have negative
real part, all the entries of et A go to zero exponentially and so, the integral converges.
Let us check that the matrix P fulfils the required properties. Clearly, P is symmetric.
The solution ϕ(t) corresponding to the initial state x = (x1 , . . . , xn ) ∈ Rn , can be
written as
n
ϕ(t) = et A x = ψi (t)xi .
i=1

Thus,


n n 
 +∞ 
V (x) = x t P x = pi j xi x j = ψit (s)ψ j (s)ds xi x j
i, j=1 i, j=1 0
 +∞ 
n
 t 
= ψi (s)ψ j (s) xi x j ds
0 i, j=1

t ⎛ n ⎞
+∞ 
n 
= ψi (s)xi ⎝ ψ j (s)x j ⎠ ds
0 i=1 j=1
48 3 Stability of Unforced Linear Systems
 +∞
= ϕ(s)2 ds .
0

The properties V (0) = 0 and V (x) > 0 for each x = 0 are easily checked. It
remains to prove (3.3). Since
 +∞  +∞
V (ϕ(t)) = ϕ(t + s)2 ds = ϕ(σ)2 dσ
0 t

we have, for each x ∈ Rn ,

 +∞
d  d 

V (ϕ(t)) t=0 = ϕ(σ)2 dσ t=0
dt dt t

= −ϕ(t)2 t=0 = −x2 .

Now we prove that (ii) =⇒ (iii). We reconsider the already defined function
V (ϕ(t)) and we compute its derivative in a different way. We have

d d 
V (ϕ(t)) = (ϕ(t))t Pϕ(t) = (ϕ̇(t))t Pϕ(t) + (ϕ(t))t P ϕ̇(t)
dt dt
= (ϕ(t))t At Pϕ(t) + (ϕ(t))t P Aϕ(t) .

Setting t = 0 and ϕ(0) = x, and taking the assumption into account, the identity
above yields
x t [At P + P A]x = −x t x .

Since the solution ϕ(t) is arbitrary, we obtain At P + P A = −I , as required.

Finally, we prove that (iii) =⇒ (i). Let λ be a (real or complex) eigenvalue of
A, and let Av = λv (v = 0). we have

−v̄ t v = v̄ t (At P + P A)v = (Av̄)t Pv + v̄ t P Av

= λ̄v̄ t Pv + λv̄ t Pv = (λ̄ + λ)v̄ t Pv = 2α v̄ t Pv

where α denotes the real part of λ. Now, it is not difficult to check that if P is any
positive definite, real symmetric matrix and if v is any (real or complex) nonzero
vector, then v̄ t Pv > 0. Hence we must have α < 0. 

In what follows, we refer to (3.4) as the Lyapunov matrix equation. Indeed, it

can be conveniently interpreted as an equation in the unknown P. It is equivalent
to a system with n(n + 1)/2 algebraic linear equations, whose unknowns are the
elements of P; Theorem 3.3 states in particular that if A is a Hurwitz matrix, then
such a system admits a solution. More precisely, under this condition there exists a
unique positive definite solution of (3.4).
A function of the form V (x) = x t P x where P is a positive definite, real symmetric
matrix enjoying one of the properties (ii) or (iii) listed in Theorem 3.3, is called a
3.3 Lyapunov Matrix Equation 49

quadratic Lyapunov function. It is indeed a homogeneous real polynomial of degree

2 with n variables.
The following corollary provides a generalized version of (3.4). It enlightens in
particular the flexibility of the matrix Lyapunov equation.

Corollary 3.1 If there exists a positive definite, real symmetric matrix Q such that
the matrix equation
At P + P A = −Q (3.5)

admits a (positive definite, real symmetric) solution P, then A is a Hurwitz matrix.

If A is a Hurwitz matrix, then for each positive definite, real symmetric matrix
Q, there exists a unique (positive definite, real symmetric) solution P of the matrix
equation (3.5).

Proof The proof of the first statement is a slight modification of the proof that
(iii) =⇒ (i) of Theorem 3.3. As far as the second statement is concerned, we start
by writing Q = R t R, where R is some nonsingular symmetric matrix (see [21] Chap.
11, or [6]). Since the eigenvalues of A have strictly negative real part, the same is true
for the matrix à = R A R −1 . According to Theorem 3.3 (iii), there exists a matrix P̃
such that
Ãt P̃ + P̃ Ã = −I .

This implies
(R −1 )t At R t P̃ + P̃ R A R −1 = −I .

The conclusion follows, multiplying both sides of this equation by R t and R

(respectively, on the left and on the right), and setting P = R t P̃ R. 
∞ t
An elegant representation of the solution of (3.5) is (see [31]) P = 0 et A
Qet A dt. In order to characterize the stability property, we may use a weakened
version of Theorem 3.3.

Theorem 3.4 The following properties are equivalent.

(i) All the eigenvalues of A have nonpositive real part, and for each possible eigen-
value with zero real part the algebraic multiplicity and the geometric multiplicity
coincide.
(ii) There exists a positive definite, real symmetric matrix P such that the matrix
At P + P A is negative semidefinite.

A function of the form V (x) = x t P x with P positive definite, real symmetric,

enjoying one of the properties listed in Theorem 3.4 is called a quadratic weak
Lyapunov function.
50 3 Stability of Unforced Linear Systems

3.4 Routh-Hurwitz Criterion

The results presented in this chapter emphasize the interest of criteria which enable
us to predict the sign of the roots of a polynomial, without need of computing them
explicitly. Recall that the eigenvalues of a matrix A coincide with the roots of the
characteristic polynomial of A. Let

P(λ) = λn + a1 λn−1 + · · · + an−1 λ + an

be a monic polynomial of degree n with real coefficients.

Proposition 3.3 If all the roots of P(λ) have strictly negative real part, then all the
coefficients ai must be strictly grater than zero.
Proof Let λ1 , . . . , λk be the (negative) real roots of P(λ) and let α1 ± i β1 , . . . , αh ±
i βh be the complex conjugate roots, with α1 < 0, . . . , αh < 0. Then we have

P(λ) = (λ − λ1 )μ1 · . . . · (λ − λk )μk

·(λ − (α1 + i β1 ))ν1 · (λ − (α1 − i β1 ))ν1 · . . .
·(λ − (αh + i βh ))νh · (λ − (αh − i βh ))νh .

Every pair of linear factors where the complex roots appear can be replaced by a
unique factor of degree 2

λ2 + p 1 λ + q 1 , . . . , λ2 + p h λ + q h

where, being α1 < 0, . . . , αh < 0, all the coefficients p1 , . . . , ph , q1 , . . . , qh are

positive.
Recovering the expanded form of the polynomial, we find of course that all the
coefficients of P(λ) are positive. 
The necessary condition provided by Proposition 3.3 is also sufficient if the degree
of the polynomial is n = 1 or n = 2, but not in general. For instance,
 √   √ 
1 + i 11 1 − i 11
λ− · λ− · [λ + 2] = λ3 + λ2 + λ + 6 .
2 2

There are several necessary and sufficient conditions which allow us to establish
when the roots of a polynomial belong to the half plane {z ∈ C : Rez < 0}. They
are generally referred to as Routh and Hurwitz criteria. We state one of these criteria
without proof. It is based on the examination of the sign of the determinants of n
matrices 1 , 2 , . . . , n of order 1, 2, . . . , n, respectively. These matrices are com-
puted starting from the coefficients of P(λ), according to the following procedure.
First of all, for sake of convenience, we agree to write a j = 0 for each value of
j > n, and a0 = 1. We define
3.4 Routh-Hurwitz Criterion 51

1 = a 1 ,
 
a1 a3
2 = ,
a0 a2
⎛ ⎞
a1 a3 a5
3 = ⎝ a 0 a 2 a 4 ⎠ ,
0 a1 a3
⎛ ⎞
a1 a3 a5 a7
⎜ a0 a2 a4 a6 ⎟
4 = ⎜
⎝0
⎟,
a1 a3 a5 ⎠
0 1 a2 a4
⎛ ⎞
a1 a3 a5 a7 a9
⎜ a0 a2 a4 a6 a8 ⎟
⎜ ⎟
5 = ⎜
⎜0 a1 a3 a5 a7 ⎟
⎟ ,
⎝0 1 a2 a4 a6 ⎠
0 0 a1 a3 a5

and so on, finishing with n . Let us remark that on the “odd” rows of these matrices
(the first row, the third row, etc.) we find the coefficients with odd index, displayed
in increasing order, while on the “even” rows we find the coefficients of even index.
The elements which appear in the first two rows are repeated in the following rows,
shifted of one position each time. The free positions at the beginning of any new row
are filled with zeros, while the last element on the right of any row is eliminated at
a new repetition.
Theorem 3.5 All the roots of the polynomial P(λ) belong to the half plane {z ∈ C :
Rez < 0} if and only if all the determinants of the matrices 1 , . . . , n are positive.
For instance, in the case n = 3 the condition of Theorem 3.5 reduces to

a1 > 0, a3 > 0, a1 a2 − a3 > 0 .

This form of the Routh-Hurwitz criterion can be found in [24] or in [10], where
the reader can also find a proof of Theorem 3.5.
Chapter Summary
In this chapter the study of unforced linear systems is continued. We focus in particu-
lar on the stability properties of the equilibrium position (the origin). This corresponds
to the study of the internal stability properties of a system with input and output. We
state and prove the classical Lyapunov Theorem which allows us to reduce the stabil-
ity analysis to an algebraic problem (computation of the eigenvalues of a matrix). We
also introduce the quadratic Lyapunov functions and the Lyapunov matrix equation.
The Routh-Hurwitz criterion is given without proof.
Chapter 4
Linear Systems with Forcing Term

The simplest way to model an external input is to introduce an additive term in

the system equations. In this chapter we shall see how the solutions of a system
of differential equations, whose right-hand side is the sum of a linear part and a
time-varying term, can be explicitly found.

4.1 Nonhomogeneous Systems

A linear nonhomogeneous system1 of differential equations has the general form

ẋ = Ax + b(t) (4.1)

where b(t), frequently referred to as the forcing term, belongs to the space PC(I, R).
Here, I denotes in general any interval of R with nonempty interior, although for
our purposes, the relevant cases are I = R and I = [0, +∞). We report below some
basic facts.
Fact 1. For each initial instant t0 ∈ I and each initial state x0 ∈ Rn there exists a
unique solution x = ψ(t) of (4.1) such that ψ(t0 ) = x0 . Moreover, ψ(t) is
defined for each t ∈ I .
Fact 2. If ψ1 (t), ψ2 (t) are solutions of (4.1) defined on I , then ψ1 (t) − ψ2 (t) is a
solution of the so-called associated homogeneous system

ẋ = Ax. (4.2)

1 Accordingto a more correct terminology, a system of the form (4.1) should be called an “affine”
system; however, the term “linear nonhomogeneous” is very frequent in the literature.
© Springer Nature Switzerland AG 2019 53
A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_4
54 4 Linear Systems with Forcing Term

Fact 3. If ϕ(t) is any solution of the associated homogeneous system (4.2) and
ψ ∗ (t) is any solution of the nonhomogeneous system (4.1), then ϕ(t) +
ψ ∗ (t) is a solution of the nonhomogeneous system (4.1).
Fact 4. (Superposition principle) If ψ1 (t) is a solution of system (4.1) with b(t) =
b1 (t) and ψ2 (t) is a solution of system (4.1) with b(t) = b2 (t), then ψ1 (t) +
ψ2 (t) is a solution of system (4.1) with b(t) = b1 (t) + b2 (t).

From Facts 2 and 3 it follows that in order to determine the set of all the solutions
of system (4.1), we need to find:

(a) a fundamental matrix (t) of (4.2);

(b) a particular solution ψ ∗ (t) of (4.1).

The set of all the solutions of system (4.1) can be therefore represented by the
formula
x = ψ(t) = (t)c + ψ ∗ (t) (4.3)

where c is a vector of arbitrary constants. It is called the general integral of system

(4.1). The particular solution corresponding to a given initial condition (t0 , x0 ) can
be obtained solving the algebraic system

x0 − ψ ∗ (t0 ) = (t0 )c.

If (t) = e(t−t0 )A , then c = x0 − ψ ∗ (t0 ).

4.1.1 The Variation of Constants Method

The problem of determining a fundamental matrix of system (4.2) has been solved
in Chap. 2. As far as point (b) is concerned, we have the following general result.

Proposition 4.1 The function

 t
ψ0∗ (t) = e(t−τ )A b(τ ) dτ (4.4)
t0

provides the solution of (4.1) such that ψ0∗ (t0 ) = 0.

Taking into account this result, we can write the solution corresponding to the
initial state (t0 , x0 ) as
 t  t
(t−t0 )A (t−τ )A (t−t0 )A
ψ(t) = e x0 + e b(τ )dτ = e (x0 + e(t0 −τ )A b(τ )dτ ). (4.5)
t0 t0

This is called Lagrange formula or variation of constants formula. This formula is

very well suited for theoretical purposes but sometimes not so convenient in practice,
4.1 Nonhomogeneous Systems 55

because of the presence of the integral that, for certain functions b(t), might be hard
or even impossible to compute explicitly.
Remark 4.1 Formula (4.5) is often used with t0 = 0 (provided of course that 0 ∈ I ),
that is in the form  t
ψ(t) = et A x0 + e(t−τ )A b(τ ) dτ . (4.6)
0

There is an interesting interpretation of (4.6): it shows that each solution is the

sum of two contributions. The first one depends on the initial state but not on the
forcing term. On the contrary, the second one depends on the forcing term but not
on the initial state. This suggests that the analysis of the dynamical behavior of a
linear system can be carried out by investigating separately the effect of the initial
conditions (with zeroed forcing term) and the effect of the forcing term (with zeroed
initial state).2 
Nonhomogeneous equations arise frequently in applications, both in classical
physics and in system theory. In any case, it is natural to presume that b(t) represents
a signal generated by an exosystem, that is an external system connected to the main
plant by a cascade connection. Exosystems are often simple linear devices without
forcing terms. Hence it is reasonable to focus on forcing terms of the form


H
b(t) = Ph (t)eγh t
h=1

where each Ph (t) is a polynomial with vector coefficients and γh ∈ C. As illustrated

in the next section, in such cases the computation of the integral in (4.5) can be
avoided by virtue of the superposition principle and the use of some practical rules
which allows us to find a particular solution in a more direct way. These rules are
presented in the next section.

4.1.2 The Method of Undetermined Coefficients

We can limit ourselves to assume b(t) = P(t)eγt , where P(t) is a polynomial with
vector coefficients. We distinguish two cases.
Case 1: γ is not an eigenvalue of A. Then, there exists a particular solution of (4.1)
with the following structure: ψ ∗ (t) = Q(t)eγt where Q is a polynomial with vector
coefficients and the same degree of P.
Case 2: γ is an eigenvalue of A, with algebraic multiplicity μ ≥ 1. Then, there exists
a particular solution of (4.1) with the following structure: ψ ∗ (t) = Q(t)eγt where Q

2 This agrees with the conclusions of Chap. 1 (Sect. 1.3.5) provided that the forcing term is interpreted

as an input and taking into account Proposition 1.10.

56 4 Linear Systems with Forcing Term

is a polynomial with vector coefficients, and deg Q = deg P + μ: in this case, we

say that the system exhibits resonance.
In both cases, the coefficients of Q depend on A, γ and the coefficients of P(t);
they can be determined by exploiting the identity ψ̇ ∗ (t) = Aψ ∗ (t) + P(t)eγt , which
leads to a system of merely algebraic equations. This is the reason why this procedure
is called the method of undetermined coefficients. We emphasize that these rules hold
even if γ is a complex number. By virtue of the formulæ

ei ωt + e−i ωt ei ωt − e−i ωt
cos ωt = , sin ωt =
2 2i
the method of undetermined coefficients can be therefore extended to forcing terms
of the form b(t) = P1 (t) cos ωt + P2 (t) sin ωt (P1 (t) and P2 (t) being polynomials
with real vector coefficients).

Example 4.1 We are especially interested in the case where the forcing term is a
periodic function of the form

b(t) = (cos ωt)u + (sin ωt)v (4.7)

where u, v ∈ Rn are constant vectors. This case occurs frequently in applications,

and it will be further developed later in this chapter, in different situations. Let us
apply the method of undetermined coefficients separately to the systems

ei ωt e−i ωt
ẋ = Ax + (u − i v) and ẋ = Ax + (u + i v).
2 2
Assuming for simplicity that resonance does not occur and taking into account
that the elements of A and u are real, we find respectively particular solutions of the
type
ψ1 (t) = ei ωt c and ψ2 (t) = e−i ωt c̄

for some constant vector c ∈ Cn . According to the superposition principle, a partic-

ular solution of the system will be found of the form

ψ ∗ (t) = ei ωt c + e−i ωt c̄.

This solution is actually real, since it is the sum of two conjugate terms. It can be
rewritten as
ψ ∗ (t) = (cos ωt)a + (sin ωt)b (4.8)

for some vector constants a, b ∈ Rn . 

Remark 4.2 It is important to notice that (4.8) is a periodic solution, with the same
frequency as the forcing term (4.7). It should be also noticed that in (4.8) a and b
may be both nonzero, even if in (4.7) one between u and v is zero. 
4.1 Nonhomogeneous Systems 57

Remark 4.3 Notice that in general it is not possible to preassign the initial state of the
particular solution obtained by the method of undetermined coefficients. In general,
we will have ψ ∗ (t0 ) = 0, so that it does not coincide with the solution introduced
in Proposition 4.1. More precisely, let ψ ∗ (t) be a particular solution obtained by the
method of undetermined coefficients, and let for simplicity t0 = 0. We may rewrite
(4.3), as
x = et A (x0 − ψ ∗ (0)) + ψ ∗ (t) (4.9)

where x0 stands for the desired initial state. The particular solution provided by the
method of variation of constants can be recovered as

ψ0∗ (t) = ψ ∗ (t) − et A ψ ∗ (0). (4.10)

4.2 Transient and Steady State

Throughout this section, we assume that I = [0, +∞) and t0 = 0. In addition, we

assume that the matrix A in (4.1) possesses the Hurwitz property (Definition 3.1).
Using (2.14), we may give an asymptotic estimation of the solutions also for
nonhomogeneous systems of type (4.1).
Proposition 4.2 If A is a Hurwitz matrix and b(t) is bounded on the interval
[0, +∞), then for each solution ψ(t) of (4.1) we have

ψ(t) ≤ k0 x0 eαt + k1 · b0 , t ≥ 0,

where k0 and k1 are positive constants, α < 0, b0 = supτ ≥0 b(τ ), and x0 = ψ(0).
Proof The assumptions imply the existence of constants α < 0 and k0 > 0 such that
for each t and each τ ∈ [0, t]

e(t−τ )A b(τ ) ≤ k0 b(τ )e(t−τ )α .

Since the initial state x0 is assigned for t0 = 0, we may use the version (4.6) of
the variation of constants formula. We have:
 t
αt b0 k0  (t−τ )α t
ψ(t) ≤ k0 x0 e + b0 k0 e(t−τ )α dτ = k0 x0 eαt − e
0 α 0


b0 k0 αt 
= k0 x0 eαt + e −1 .
α

Being α < 0, we have eαt ≤ 1 for t > 0. Setting k1 = k0

|α|
, the previous inequality
reduces to the desired one. 
58 4 Linear Systems with Forcing Term

Proposition 4.2 implies in particular that if the matrix A is Hurwitz and the forcing
term is bounded, then every solution is bounded on [0, +∞). We want to focus on
the following two particular cases:
(1) the forcing term b(t) is constant;
(2) the forcing term b(t) is a periodic function of the form (4.7).
Note that in force of the Hurwitz property, resonance does not occur neither in
case (1) nor in case (2). As a consequence, the system admits a unique constant
solution in case (1) and, respectively, a unique periodic solution3 in case (2). One
such solution ψ ∗ (t) can be used in (4.9), in order to represent a generic solution.
Recalling again that A is Hurwitz and using (2.14) with α < 0, we have that

lim et A c = 0
t→+∞

for each c ∈ Rn . This means that in (4.9), for sufficiently large t, the term et A (x0 −
ψ ∗ (0)) can be neglected and the evolution of the system “becomes independent”
of the initial state x0 . It is approximately constant or periodic, and it is essentially
determined by the forcing term. It is customary to distinguish two stages in the
time evolution of the system. The first stage, where the evolution is appreciably
affected by the initial state x0 , is called the transient. The subsequent stage, where
the effect of the initial state is no more perceptible, is called the steady state. Of course,
the distinction between the transient and the steady state is not rigorous, since the
term et A (x0 − ψ ∗ (0)) in (4.9) will never be exactly equal to zero. Distinguishing
the two stages depends on the admitted error margins and on the precision of the
measurements, but it is very impressive and convenient, at least from the heuristic
point of view.
Remark 4.4 The steady state solution is be more correctly thought of as a “limit”
solution, asymptotically approached by all the solutions of system (4.1). As already
noticed, such a limit solution does not necessarily vanish for t = 0, and so it does
not coincide, in general, with the particular solution appearing in the variation of
constants formula (4.6). Indeed, as we can understand from (4.10), further terms
vanishing when t → +∞, could be hidden in the particular solution appearing in
(4.6). These terms compensate for the gap between the assigned initial state and the
initial state of the steady state solution. The steady state solution is found in a natural
way when the method of undetermined coefficients is adopted. 
Example 4.2 Consider the system represented by the scalar differential equation

ẋ = −x + 2 (4.11)

with the initial condition x(0) = 1. The general integral of the associated homoge-
neous system is x = e−t c, with c an arbitrary constant. A solution of the

3 On the other hand, it is easy to check that there exists a constant or periodic solution only if the
forcing term is, respectively, constant or periodic.
4.2 Transient and Steady State 59

2.5

1.5

0.5

−0.5

−1
−1 0 1 2 3 4 5

Fig. 4.1 The curve in bold represents the graph of the solution of (4.11) such that x(0) = 1; the
curve marked by A represents the graph of the solution of the associated homogeneous equation with
the same initial condition x(0) = 1; the curve marked by R represents the graph of the steady state
solution; the curve marked by N represents the graph of the solution of (4.11) such that x(0) = 0

nonhomogeneous
 t equation (4.11) can be found by applying Proposition 4.1: we
obtain x = 2 0 e−(t−τ ) dτ = 2 − 2e−t . Since it vanishes for t = 0, we set c =
x(0) = 1. According to (4.6), the required solution writes

x = e−t − 2e−t + 2. (4.12)

Alternatively, we can use the method of undetermined coefficients. In this way

we find directly the steady state solution x = 2. The general integral of (4.11) takes
therefore the form
x = e−t c + 2

and imposing the condition x(0) = 1, now we find c = −1. Of course, the two
approaches lead to the same result. The graphs of the various components of the sum
(4.12) are shown in Fig. 4.1. 

4.3 The Nonhomogeneous Scalar Equation of Order n

By the same procedure illustrated in Sect. 2.10, the nonhomogeneous scalar equation
of order n (with constant coefficients)

y (n) + a1 y (n−1) + · · · + an−1 y

+ an y = g(t) (4.13)
60 4 Linear Systems with Forcing Term

can be rewritten as a system of the form (4.1) with a matrix A in companion form,
and ⎛ ⎞
0
⎜ .. ⎟
⎜ ⎟
b(t) = ⎜ . ⎟ .
⎝ 0 ⎠
g(t)

Thus, (4.13) can be considered a particular case of (4.1), the function g(t) playing
the role of the forcing term. It follows that for each function g(·) ∈ PC(I, R) and for
each set of initial conditions

y(t0 ) = y0 , y
(t0 ) = y1 , . . . , y (n−1) (t0 ) = yn−1 (4.14)

(t0 ∈ I ) there is a unique solution defined for t ∈ I . We emphasize that by this

procedure, we are led to identify the state of the system with the vector whose
components are (y, y
, . . . , y (n−1) ).
In order to determine the solutions of (4.13), the methods described in the previ-
ous sections can be applied. However, if g(t) = p(t)eγt where γ ∈ C and p(t) is a
polynomial with real or complex coefficients, it is more convenient to work directly
with (4.13). Indeed, we can write the general integral as

y(t) = c1 y1 (t) + · · · + cn yn (t) + χ∗ (t) (4.15)

where y1 (t), . . . , yn (t) are linearly independent solutions of the associated homoge-
neous (or unforced) equation

y (n) + a1 y (n−1) + · · · + an−1 y

+ an y = 0, (4.16)

c1 , . . . , cn are arbitrary constants, and χ∗ (t) is a particular solution of (4.13). Now,

the method of undetermined coefficients gives rise to the following simplified rules:
a particular solution χ∗ (t) can be sought of the form

1. χ∗ (t) = q(t)eγt provided that γ is not a characteristic root of (4.16);

2. χ∗ (t) = t μ q(t)eγt provided that γ is a characteristic root of (4.16) with algebraic
multiplicity μ (case of resonance).

In both cases, q(t) represents a polynomial of the same degree as p(t). Recall
that the solution χ∗ (t) obtained by the method of undetermined coefficients does not
coincide, in general, with the solution of (4.13) vanishing at t = t0 .

Remark 4.5 If all the characteristic roots have strictly negative real part, then all the
solutions of the associated homogeneous system (4.16) (and all their derivatives)
go to zero when t → +∞. Hence, if the forcing term g(t) is constant or periodic,
the particular solution χ∗ (t) can be interpreted, also in this case, as the steady state
solution. 
4.3 The Nonhomogeneous Scalar Equation of Order n 61

1 0.5
0.9 0.4
0.8 0.3
0.7 0.2
0.6 0.1
0.5 0

0.4 −0.1

0.3 −0.2

0.2 −0.3

0.1 −0.4

0 −0.5
0 5 10 15 0 5 10 15

Fig. 4.2 Examples 4.3 and 4.4: steady state solution and transient

Example 4.3 The linear equation of order 2 in general form

+ ay
+ by = g(t) (4.17)

constitutes a model for a large variety of physical problems, and it is appropriate to

illustrate the transient and the steady state phenomena. To this end, we assume that
the characteristic polynomial of the associated homogeneous equation has a pair of
complex conjugate roots α ± i β with α = −a/2 < 0 and β = 0 (which necessarily
yields b = 0).
If the forcing term is constant, say g(t) = g0 , the unique constant solution is
χ∗ (t) = g0 /b. Its graph is drawn in Fig. 4.2 (left), for the case a = 1, b = 6,
g0 = 3. The figure shows also the graph of a solution corresponding to different
initial conditions. The transient stage can be recognized in the interval where the two
graphs can be clearly distinguished. 

Example 4.4 Considered again the general second order equation (4.17) under the
same assumptions about the coefficients a, b, but now with a periodic forcing term

g(t) = p1 cos ωt + p2 sin ωt, p1 , p2 ∈ R.

By the same procedure of Example 4.1, we can find a particular solution of the
form
χ∗ (t) = q1 cos ωt + q2 sin ωt, q1 , q2 ∈ R (4.18)

which can be recognized as the steady state solution. The general integral can be
written as

y(t) = (c1 cos βt + c2 sin βt)eαt + q1 cos ωt + q2 sin ωt. (4.19)

62 4 Linear Systems with Forcing Term

The transient will be shorter and shorter, as the absolute value of α becomes larger
and larger.
The coefficients q1 and q2 in (4.18) depend on p1 , p2 and ω (as well as on a and
b) and can be computed by direct substitution.4 Sometimes, it may be convenient to
rewrite (4.18) as
χ∗ (t) = ρ cos(ωt + θ) (4.20)

where q1 = ρ cos θ, q2 = ρ sin θ. The quantities ρ and θ represent the amplitude and,
respectively, the phase of the periodic function at hand (compare with Example 2.2).
Also the forcing term can be rewritten in a similar way. Note that the initial conditions
contribute to determine the values of c1 and c2 in (4.19), but not the values of q1 , q2
(equivalently, ρ, θ) characterizing the shape of (4.18).
Consistently with Remark 4.2, we see that the frequency of the steady state solution
is unchanged, when compared with the frequency of the forcing term. On the contrary,
while the signal goes through the system, the phase and the amplitude may undergo
a variation.
A simulation is presented in Fig. 4.2 (right), for the case a = 1, b = 6, g0 = sin t.
The periodic steady state solution is χ∗ (t) = (− cos t + 5 sin t)/26. 
Example 4.5 Let us consider again the Eq. (4.17), with the same forcing term but
now with a = 0. If ω 2 = b then i ω is a solution of the characteristic equation. The
system resonates. The form of the general integral is

χ∗ (t) = (c1 + tq1 ) cos ωt + (c2 + tq2 ) sin ωt. (4.21)

The constants q1 , q2 characterizing the particular solution can be easily determined

by direct substitution. The solutions exhibit an oscillatory behavior, and the amplitude
of the oscillations goes to +∞ when t → ∞. 

4.4 The Laplace Transform Method

In this section we discuss a different approach to the problem of determining the

solutions of (4.13), based on the Laplace transform (see Appendix B for notation
and properties of the Laplace transform).

4.4.1 Transfer Function

Let us assume that the forcing term g(·) is defined for t ≥ 0, and that it belongs to
the set of subexponential functions of class PC([0, +∞), Rm ). According to what

4 Recall that pi = 0 (i = 1, 2) does not imply in general qi = 0.

4.4 The Laplace Transform Method 63

exposed in the previous sections, we know that the solution y(t) is defined for t ≥ 0,
and that it is a subexponential function, as well. This justifies the use of the Laplace
transform.
Let us apply the Laplace transform to both side of (4.13). Recalling (4.14) and
(B.10), from

L[y (n) + a1 y (n−1) + · · · + an−1 y

+ an y] = L[g(t)]

we have

(s n + a1 s n−1 + · · · + an )Y (s)
n−1
− s y0 + s n−2 y1 + s n−3 y2 + · · · + yn−1
 
+ s n−2 y0 + s n−3 y1 + · · · + yn−2 a1
 
+ s n−3 y0 + · · · + yn−3 a2
+.................................

+y0 an−1 = G(s).

We recognize that s n + a1 s n−1 + · · · + an is nothing else but the characteristic

polynomial pch (s) of the homogeneous equation (4.16) associated to (4.13). Thus
we can write
pch (s)Y (s) − P0 (s) = G(s) (4.22)

where
P0 (s) = A0 s n−1 + A1 s n−2 + · · · + An−1 (4.23)

with

A0 = y0 , A1 = y1 + a1 y0 , . . . , An−1 = yn−1 + a1 yn−2 + · · · + an−1 y0 .

We remark that:

(i) pch (s) is independent of both the forcing term and the initial conditions;
(ii) deg P0 (s) < deg pch (s);
(iii) P0 (s) vanishes if and only if y0 = · · · = yn−1 = 0.

From (4.22) we obtain formally

P0 (s) G(s)
Y (s) = + . (4.24)
pch (s) pch (s)

Formula (4.24) is well defined provided that s is not a solution of the characteristic
equation pch (s) = 0. Since the characteristic equation has finitely many solutions,
there exists a real number σ0 such that (4.24) holds in the half plane {s ∈ C : Re s >
σ0 }.
64 4 Linear Systems with Forcing Term

Formula (4.24) provides in a purely algebraic way the Laplace transform of the
solution y(t) corresponding to the given initial conditions. Therefore, the solution
y(t) can be now determined for t ≥ 0 by applying the inverse of the Laplace transform
L−1 . It is convenient to set
   
P0 (s) G(s)
ϕ(t) = L−1 and χ(t) = L−1
pch (s) pch (s)

so that y(t) = ϕ(t) + χ(t). The following remarks point out the analogy between
the structures of (4.24) and of (4.15).

Remark 4.6 The first summand of (4.24) contains the information about the initial
conditions: it coincides with the solution of the homogeneous equation (4.16) associ-
ated to (4.13), with the same initial conditions. This term is a proper rational function:
once it has been decomposed as a sum of partial fractions, we may easily go back to
ϕ(t) by means of the table of inverse Laplace transforms.
Of course, in this way we recover the well known conclusions about the structure
of the form of the general integral of a linear homogeneous differential equation.
Indeed, the inverse transform of the rational function P0 (s)/ pch (s) is given by the
sum of functions of the form Q 1 (t)eαt cos βt and Q 2 (t)eαt sin βt where Q 1 (t), Q 2 (t)
are polynomials of degree less than n, whose coefficients depend on the initial con-
ditions. 

Remark 4.7 The second summand of (4.24) depends on the forcing term. It coin-
cides with the solution obtained solving (4.13) with zero initial state (instead of
the conditions (4.14)). It is written as a product H (s)G(s), where the function
H (s) = 1/ pch (s) (defined on the half plane {s ∈ C : Re s > σ0 }) is called the trans-
fer function. Let h(t) be the function which coincides with the inverse Laplace
transform of H (s) for t ≥ 0, and vanishes for t < 0. Then the solution of (4.13)
corresponding to the initial conditions y0 = · · · = yn−1 = 0 can be represented by
the formula  t
χ(t) = h(t − τ )g(τ ) dτ per t ≥ 0 (4.25)
0

(recall (B.12)). In particular, if we interpret g(t) as an input and we agree that it

vanishes for t < 0, then h(t) can be reviewed as the impulse response function of
the system defined by (4.13).
Formula (4.25) can be considered as an extension of the variation of constants
formula to the differential equation (4.13). 

Example 4.6 We want to find the solution of the system defined by the linear differ-
ential equation of second order

+ 3y
+ 2y = 1 (4.26)
4.4 The Laplace Transform Method 65

with initial conditions y(0) = 1, y

(0) = 0. We apply to both sides of (4.26) the
operator L. We have
       
L y

+ 3L y
+ 2L [y] = −y
(0) + sL y
+ 3L y
+ 2L [y]
= −y
(0) + (s + 3)(sY (s) − y(0)) + 2Y (s)
= (s 2 + 3s + 2)Y (s) − y(0)(s + 3) − y
(0)
1
= .
s
The Laplace transform of the forcing term requires the restriction Re s > 0. In
this region of the complex plane there is no solutions of the characteristic equation

s 2 + 3s + 2 = 0

which are both real and negative. So we obtain

s+3 1 s 2 + 3s + 1
Y (s) = + =
s 2 + 3s + 2 s(s 2 + 3s + 2) s(s + 2)(s + 1)
 
1 1 1 2
= − + .
2 s s+2 s+1

By applying the inverse transform L−1 , we easily get

1 
y(t) = 1 − e−2t + 2e−t
2
for t ≥ 0. We recognize in this last expression the sum of a particular solution of (4.26)
and a particular solution of the associated homogeneous equation. The computations
above deserve some comments. In particular, we remark that Y (s) was obtained as
the sum of two terms: then we passed to a single rational expression and finally we
performed the partial fraction decomposition. This approach is the most natural and
convenient for practical purposes. However, we may also rearrange the computation
in a different way. Consistently with the previous analysis (Remarks 4.6 and 4.7), we
now maintain separate the term carrying the information about the initial conditions
and the term carrying the information about the forcing term. We have
   
1 2 1 1 1 2
Y (s) = − + + + −
s+2 s+1 2 s s+2 s+1

which yields
 
 −2t −t
 1  −2t −t
 1
y(t) = −e + 2e + e − 2e + .
2 2
66 4 Linear Systems with Forcing Term

Now it is easier to interpret the structure of y(t). The first summand represents the
solution corresponding to the zero input (that is, the solution of the associated homo-
geneous equation) and the same initial conditions. Since the roots of the characteristic
polynomial are negative, this part affects only the transient.
The second summand represents the solution corresponding to zero initial condi-
tions. In turn, it is formed by a constant term (the steady state solution) plus other
terms whose effect can be appreciated only in the transient. As already mentioned,
the presence of these terms is due to the need of compensating the difference between
the initial data of the actual solution and the steady state solution.
When the forcing term is not constant, the problem of the factorization of a
polynomial of higher degree arises, For instance, if we take an input signal g(t) =
sin t, we have:

s+3 1
Y (s) = +
(s + 2)(s + 1) (s + 2)(s + 1)(s 2 + 1)
 
s 2 + 3s 2 + s + 4 1 −12 25 1 − 3s
= = + +
(s + 2)(s + 1)(s 2 + 1) 10 s + 2 s + 1 s2 + 1

and so
1  
y(t) = −12e−2t + 25e−t + sin t − 3 cos t .
10


4.4.2 Frequency Response Analysis

In this section we present some further developments about the study of a linear
differential equation (4.13), with a periodic forcing term of the form

g(t) = p1 cos ωt + p2 sin ωt (4.27)

under the assumption that all the solutions of the characteristic equation pch (s) = 0
have strictly negative real part. As well known, under these conditions the steady
state solution is periodic, with the same frequency as the forcing term (4.27). One of
the classical problems at the origin of system theory is the analysis of the solution
(response) corresponding to a periodic forcing term (input) of this form.
The problem has been already studied in the case where the order of the equation is
n = 2 (Example 4.4), as an application of the method of undetermined coefficients.
For the general case, the method illustrated in this section, based on the Laplace
transform, provides a very efficient tool which allows us to obtain further information,
and in particular to determine the parameters of the system and of the forcing term in
such a way that the solutions have preassigned amplitude and phase. This approach
is the so-called frequency response analysis. Taking into account (4.27), we may
rewrite (4.24) as
4.4 The Laplace Transform Method 67

P0 (s) 1 p1 s + p2 ω
Y (s) = + .
pch (s) pch (s) s 2 + ω 2

Since the solutions of the characteristic equation pch (s) = 0 lies in the negative
complex half-plane, we have pch (i ω) = 0; hence s 2 + ω 2 is not a divisor of pch (s).
We can rewrite the right-hand side as

P0 (s) P(s) q1 s + q2
Y (s) = + + 2
pch (s) pch (s) s + ω2

where P(s) is a polynomial, and q1 , q2 are constants such that

P(s)s 2 + q1 pch (s)s + P(s)ω 2 + q2 pch (s) = p1 s + p2 ω . (4.28)

We already know (see Remark 4.6) that the inverse transform of the rational
function P0 (s)/ pch (s) is the sum of functions of the form Q 1 (t)eαt cos βt and
Q 2 (t)eαt sin βt where Q 1 (t), Q 2 (t) are polynomials of degree less than n. Our
hypothesis that all the characteristic roots have negative real part implies that these
terms go to zero when t → +∞. Formula (4.28) shows in particular that deg P <
deg pch . Thus, the same reasoning can be repeated about the term P(s)/ pch (s),
as well. We finally conclude that the contributions of the terms P0 (s)/ pch (s) and
P(s)/ pch (s) can be ultimately neglected, and the steady state response depends
essentially on the third summand (q1 s + q2 )/(s 2 + ω 2 ), whose inverse transform is
 
q1 s + q2 q2
L−1 = q1 cos ωt + sin ωt = k sin(ωt + θ)
s +ω
2 2 ω

being q1 = k sin θ and q2 = kω cos θ. Recall that the term P0 (s)/ pch (s) represents
the solution of the unforced system with the same initial conditions of the given
system. The term P(s)/ pch (s) compensates the difference between the assigned
initial conditions and the (in general, different) initial conditions of the steady state
solution (compare these comments with those in Remark 4.4).
Finally, we show how to compute q1 and q2 , and hence k and θ. Replacing s = i ω,
from (4.28) we find
ω
q2 + i q1 ω = ( p2 + i p1 ),
pch (i ω)

which yields q2 = Re [ pchω(i ω) ( p2 + i p1 )] and q1 = Im [ pch 1(i ω) ( p2 + i p1 )]. Alterna-

tively, we can compute p1 and p2 as functions of some desired values of q1 and q2 .
Chapter Summary
This chapter constitutes a different development of Chap. 2. We consider the prob-
lem of representing the solutions of nonhomogeneous (i.e., with forcing term) sys-
tems of linear differential equations. We present the variation of constants formula
and the method of undetermined coefficients. Moreover, we illustrate the qualitative
68 4 Linear Systems with Forcing Term

notions of transient and steady state solution. Finally we present the Laplace trans-
form method and, as an application, we discuss the frequency response analysis of a
system under periodic input.
Chapter 5
Controllability and Observability
of Linear Systems

In this chapter we begin to study differential systems with inputs and outputs. We
focus in particular on the so-called structural properties of finite-dimensional, time
invariant linear systems, that is systems of the form

ẋ = Ax + Bu
(5.1)
y = Cx

where x ∈ Rn represents the state of the system, u ∈ Rm represents the input

and y ∈ R p represents the output (n, m and p arbitrary integers greater than or
equal to 1). Throughout this chapter, the admissible inputs are functions u(·) ∈
PC([0, +∞), Rm ). Indeed, the qualifier “structural” identifies properties which de-
pend only on the matrices A, B, C, and so are not affected by possible restrictions
on the inputs variables.

5.1 The Reachable Sets

For each admissible input u(·) ∈ PC([0, +∞), Rm ) and for each initial state x(0) =
x0 , there is a unique solution of the system

ẋ = Ax + Bu(t) (5.2)

denoted by x(t, x0 , u(·)), and defined for t ≥ 0. System (5.2) can be thought of as a
linear nonhomogeneous system with forcing term b(t) = Bu(t). Hence, the variation
of constants formula applies and we can represent the solution as

© Springer Nature Switzerland AG 2019 69

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_5
70 5 Controllability and Observability of Linear Systems
  t 
x(t, x0 , u(·)) = et A x0 + e−τ A Bu(τ ) dτ . (5.3)
0

As already mentioned (Chap. 4), it is natural to think of (5.3) as the sum of

x(t, x0 , 0) = et A x0 (5.4)

also called the free (or unforced) solution, and

 t
x(t, 0, u(·)) = e(t−τ )A Bu(τ )dτ . (5.5)
0

We emphasize that (5.4) represents the solution corresponding to the input u = 0,

while (5.5) represents the solution corresponding to the actual input but with zeroed
initial state. To this respect, there is an analogue of Proposition A.1.

Proposition 5.1 For each pair of real numbers t, τ ∈ [0, +∞), for each admissible
input u(·) : [0, +∞) → Rm , and for each initial state x0 , we have

x(0, x0 , u(·)) = x0

and
x(t + τ , x0 , u(·)) = x(t, x(τ , x0 , u(·)), w(·))

where we set w(t) = u(t + τ ) for t ∈ [0, +∞). 

We now introduce the first important notion of this chapter.

Definition 5.1 Let x0 , η0 ∈ Rn . We say that η0 is reachable from x0 at time T > 0
(or also that x0 is controllable to η0 at time T ) if there exists an admissible input
u(·) : [0, T ] → Rm such that

η0 = x(T, x0 , u(·)) . (5.6)

For fixed x0 and T , the set of points reachable from x0 at time T is denoted by
R(T, x0 ) and it is called the reachable set.
Intuitively, the “size” of the set R(T, x0 ) provides a measure of our ability to con-
trol the performances of the system. We are in particular interested in the following
definitions.
Definition 5.2 A system of the form (5.1) is said to be:
• globally reachable from x0 at time T if R(T, x0 ) = Rn ;
• globally reachable at time T if R(T, x0 ) = Rn for each x0 .
In fact, the two notions introduced in the previous definition are equivalent.
5.1 The Reachable Sets 71

Proposition 5.2 If there exists a state x0 such that the system is globally reachable
from x0 at time T , then the system is globally reachable at time T from the origin, as
well. If the system is globally reachable from the origin at time T , then it is globally
reachable at time T .

Proof Assume that there exists a point x0 such that the system is globally reachable
at time T from x0 , and let η0 be an arbitrary point of Rn . Let η̄ = η0 + e T A x0 . By
assumption, there exists an input function u(·) such that
 T
η̄ = η0 + e T A x0 = e T A x0 + e(T −τ )A Bu(τ ) dτ .
0

This yields
 T
η0 = e(T −τ )A Bu(τ ) dτ ,
0

meaning that η0 is reachable from the origin at time T .

Vice versa, assume that the system is globally reachable at time T from the origin.
Let x0 and η0 be two arbitrary points of Rn . Setting η̄ = η0 − e T A x0 , we can find an
input function u(·) such that
 T
η̄ = η0 − e TA
x0 = e(T −τ )A Bu(τ ) dτ
0

that is  T
η0 = e TA
x0 + e(T −τ )A Bu(τ ) dτ
0

and this means that η0 is reachable from x0 at time T . 

Remark 5.1 Analogously, we may fix η0 and T and then we may consider the set
of points x0 for which there exists an admissible input u(·) : [0, T ] → Rm such that
(5.6) holds. This is called the controllable set and it is denoted by C(T, η0 ). Clearly,
C(T, η0 ) is nothing else that R(T, η0 ) for the reversed time system, obtained replacing
A, B by −A, −B in (5.1). Indeed, multiplying by e−T A both sides of the equality
 T
η0 = e TA
x0 + e(T −τ )A Bu(τ ) dτ
0

and transforming the integral by the substitution τ = T − θ, we get

 T
T (−A)
x0 = e η0 + e(T −θ)(−A) (−B)u(T − θ) dθ .
0


72 5 Controllability and Observability of Linear Systems

5.1.1 Structure of the Reachable Sets

The reachability property introduced in the previous section involves uniquely the
input variables and the state variables. It does not depend on the matrix C, and
therefore it is natural to argue that it can be characterized only in terms of the matrices
A and B. Moreover, Proposition 5.2 suggests that our attention can be focused on
the set of points reachable from the origin.
Theorem 5.1 Let the linear system (5.1) be given. The map which associates to each
u(·) ∈ PC([0, +∞), Rm ) the function
 t
t → x(t, 0, u(·)) = e(t−τ )A Bu(τ ) dτ ∈ C([0, +∞), Rn ) (5.7)
0

is linear.

Proof If u 1 (·), u 2 (·) ∈ PC([0, +∞), Rm ) and α, β ∈ R, then also αu 1 (·) + βu 2 (·) ∈
PC([0, +∞), Rm ) and, according to the basic properties of the integral,
 t
e(t−τ )A B(αu 1 (τ ) + βu 2 (τ )) dτ
0
 t  t
=α e(t−τ )A Bu 1 (τ ) dτ + β e(t−τ )A Bu 2 (τ ) dτ .
0 0

Fix now T > 0. We can reinterpret (5.7) as a map  which associates to each
input function u(·) ∈ PC([0, T ], Rm ) the element of Rn
 T
x = (u(·)) = x(T, 0, u(·)) = e(T −τ )A Bu(τ ) dτ ∈ Rn . (5.8)
0

Corollary 5.1 The map  : PC([0, T ], Rm ) → Rn is linear.

Corollary 5.2 For each fixed T > 0, the set R(T, 0) is a linear subspace of Rn . For
each T > 0 and each x0 = 0, the set R(T, x0 ) is a linear manifold of Rn .

Proof For each fixed T > 0, the set R(T, 0) coincides with the image of the operator
 and hence it is a linear subspace of Rn . As far as the second statement is concerned,
it is sufficient to remark that R(T, x0 ) is the translation of R(T, 0) by means of the
vector v = e T A x0 . 

According to these conclusions, it is natural to assume as a measure of the “size”

of the set R(T, x0 ) the dimension of R(T, x0 ) as a linear manifold of Rn . Moreover,
R(T, 0) will be often referred to as the reachable space.
5.1 The Reachable Sets 73

Corollary 5.3 System (5.1) is globally reachable at time T if and only if R(T, 0) =
Rn , that is if and only if the dimension of R(T, 0) is maximal.

5.1.2 The Input-Output Map

The previous results enable us to prove Proposition 1.10. From (5.3), given any
admissible input u(·) : [0, +∞) → Rm and any initial state x0 , the following repre-
sentation for the output function of system (5.1)

 t 
y(t, x0 , u(·)) = C x(t, x0 , u(·)) = Ce tA
x0 + e−τ A Bu(τ ) dτ (5.9)
0

can be readily deduced. Of course, y(0, x0 , u(·)) = C x0 and

y(t, x0 , u(·)) = y(t, x0 , 0) + y(t, 0, u(·)) . (5.10)

Proof of Proposition 1.10 The map which associates to each u(·) ∈ PC([0, +∞),
Rm ) the function x(t, 0, u(·)) is linear, by virtue of Theorem 5.1. Hence, the map
which associates to u(·) the function y(t, 0, u(·)) = C x(t, 0, u(·)) is linear, as well.
On the other hand, also the map which associates to x0 the function y(t, x0 , 0) =
Cet A x0 is linear. To finish, it is sufficient to take into account (5.10) and the fact that
if f 1 : V1 → W , f 2 : V2 → W are linear maps, then f 1 + f 2 : V1 × V2 → W is a
linear map. 

5.1.3 Solution of the Reachability Problem

Next theorem provides a first necessary and sufficient condition for the global reach-
ability of a linear system.

Theorem 5.2 System (5.1) is globally reachable at time T > 0 if and only if the
matrix  T
e−τ A B B t e−τ A dτ
t
(T ) =
0

is nonsingular.

Proof First we show that if (T ) is nonsingular, then for each pair of states x0 , η0 ∈
Rn there exists an input function u(·) for which (5.6) holds. Let, for τ ∈ [0, T ],

u(τ ) = −B t e−τ A  −1 (T )[x0 − e−T A η0 ]

t
(5.11)
74 5 Controllability and Observability of Linear Systems

and compute
 T
e TA
x0 + e(T −τ )A Bu(τ ) dτ (5.12)
0
 T
e−τ A B B t e−τ A dτ  −1 (T )[x0 − e−T A η0 ]
t
= eT A x0 − eT A
0
=e TA
x0 − e TA
(T ) −1 (T )[x0 − e−T A η0 ] = e T A x0 − e T A x0 + η0 .

In conclusion,  T
e TA
x0 + e(T −τ )A Bu(τ ) dτ = η0 .
0

In order to prove the converse, we need some preliminary remarks. Clearly, (T )
is symmetric, and the quadratic form
 T
||B t e−τ A ξ||2 dτ
t
ξ (T )ξ =
t
(5.13)
0

is, in general, positive semidefinite. If (T ) is singular, then there exists a point
x0 ∈ Rn (x0 = 0) such that x0 t (T )x0 = 0. Therefore, taking into account (5.13), we
have B t e−τ A x0 = 0, identically for τ ∈ [0, T ]. The global reachability assumption
t

implies that starting from x0 it is possible to reach the origin at time T . This yields
 T
eT A x0 = − e(T −τ )A Bu(τ ) dτ (5.14)
0

for some admissible input u(·). From (5.14) it follows

 T
x0 = − e−τ A Bu(τ ) dτ
0

and so
 T 
||x0 ||2 = x0 t x0 = − e−τ A Bu(τ ) dτ t
x0
0
 T
u t (τ )B t e−τ A x0 dτ = 0 .
t
=−
0

This contradicts the assumption x0 = 0. 

Remark 5.2 Formula (5.11) provides an answer to the problem of determining a

control function which allows us to steer the system from the state x0 to the state η0 . 
5.1 The Reachable Sets 75

5.1.4 The Controllability Matrix

The condition stated in Theorem 5.2 is useful for theoretical developments, but not
easy to apply in practice. From this point of view, the criterion we are going to present
in this section is more convenient, since it amounts to purely algebraic computations
involving only the matrices A and B which define the system.
Theorem 5.3 For a system of the form (5.1), the set R(T, 0) is independent of T .
Moreover, for each T > 0 we have

R(T, 0) = V (5.15)

where


V = span b1 , . . . , bm , Ab1 , . . . , Abm , . . . , An−1 b1 , . . . , An−1 bm (5.16)

and b1 , . . . , bm denote the columns of B.

Proof Since both sides of (5.15) are subspaces of Rn , it is sufficient to prove that the
respective orthogonal spaces coincide. First we prove that V ⊂ R(T, 0). Let μ = 0
be a vector orthogonal to R(T, 0). Then we have, for each admissible input,
 T
0 = μt e(T −τ )A Bu(τ ) dτ (5.17)
0
 T  T
= μt e(T −τ )A Bu(τ ) dτ = μt eθ A Bu(T − θ) dθ .
0 0

Taking into account (5.17), now we show that

μt eθ A Bu = 0 (5.18)

for each θ ∈ (0, T ) and each u ∈ Rm . Assume that this is false. Then we can find
θ̄ ∈ (0, T ) and ū ∈ Rm such that μt eθ̄ A B ū = 0 (say for instance, μt eθ̄ A B ū > 0).
Then, by continuity, there exists δ > 0 such that (θ̄ − δ, θ̄ + δ) ⊂ (0, T ) and the
function
θ → μt eθ A B ū

takes positive values for θ̄ − δ < θ < θ̄ + δ. Setting τ̄ = T − θ̄, we can therefore
define 
ū for τ̄ − δ < τ < τ̄ + δ
u(τ ) = (5.19)
0 otherwise .

Then, 
ū for θ̄ − δ < θ < θ̄ + δ
u(T − θ) = (5.20)
0 otherwise .
76 5 Controllability and Observability of Linear Systems

This yields
 T  θ̄+δ
μt eθ A Bu(T − θ) dθ = μt eθ A B ū dθ > 0
0 θ̄−δ

and we have a contradiction to (5.17). Hence, (5.18) is true. Taking the limit for
θ → 0+ , we get
μt Bu = 0 ∀u ∈ Rm .
⎛ ⎞ ⎛ ⎞
1 0
⎜0⎟ ⎜ .. ⎟
⎜ ⎟ ⎜ ⎟
Choosing respectively u = ⎜ . ⎟ , . . . , u = ⎜ . ⎟, this last expression indicates
⎝ .. ⎠ ⎝0⎠
0 1
that μ is orthogonal to b1 , . . . , bm . Moreover, taking the derivative of (5.18) with
respect to θ, we have

μt eθ A ABu = 0 ∀θ ∈ (0, T ) and ∀u ∈ Rm

which in turn implies, for θ → 0+ ,

μt ABu = 0 .

Repeating the previous reasoning, we see that μ is orthogonal to the vectors

Ab1 , . . . , Abm , as well. The procedure can be iterated, until the conclusion is
achieved.
Now we prove the opposite inclusion. Let μ be orthogonal to

b1 , . . . , bm , Ab1 , . . . , Abm , . . . , An−1 b1 , . . . , An−1 bm .

For each u ∈ Rm , we have

μt Bu = · · · = μt An−1 Bu = 0 .

Moreover


n−1 i i ∞

θ A θ i Ai
μt eθ A Bu = μt Bu + μt Bu
i=0
i! i=n
i!

n−1 i ∞

θ θi
= μt Ai Bu + μt Ai Bu .
i=0
i! i=n
i!

Clearly, the terms of the first sum vanish. But also the terms of the second sum
vanish since, by Cayley-Hamilton Theorem, for each i ≥ n, the vector Ai Bu is a
5.1 The Reachable Sets 77

linear combination of the vectors Ai Bu with i < n. In conclusion, μt eθ A Bu = 0, for

all θ ∈ [0, T ] and all u ∈ Rm . But then also
 T
μ t
e(T −τ )A Bu(τ )dτ = 0
0

for every u(·) ∈ PC([0, T ], Rm ). The theorem is finally proved. 

Definition 5.3 System (5.1) is said to be completely controllable when

rank (B|AB| . . . |An−1 B) = n (5.21)

where (B|AB| . . . |An−1 B) is the matrix with n rows and nm columns formed by the
columns of the matrices B, AB, . . . , An−1 B.

The matrix (B|AB| . . . |An−1 B) is called the controllability matrix of system

(5.1). The following corollary is a straightforward consequence of Theorem 5.3.

Corollary 5.4 System (5.1) is completely controllable if and only if it is globally

controllable for some (and hence for each) T > 0.

Remark 5.3 The vectors v1 , . . . , vn of Rn form a linearly independent set if and only
if
det (v1 | . . . |vn ) = 0 .

Since the determinant depends continuously on the entries of the matrix, replacing
the vectors v1 , . . . , vn by some other vectors ṽ1 , . . . , ṽn such that ṽk is sufficiently
close to vk (for every k = 1, . . . , n), then also the vectors ṽ1 , . . . , ṽn form a linearly
independent set.
From this remark it follows that if system (5.1) is completely controllable and if
the matrices Ã, B̃ are sufficiently close to, respectively, A and B, then the system
defined by the matrices Ã, B̃ is completely controllable, as well. It is also clear
that if system (5.1) is not completely controllable, then there exist pairs of matrices
Ã, B̃ arbitrarily close to A, B, such that the system defined by Ã, B̃ is completely
controllable. In other words, we can say that “generically”, any linear system is
completely controllable, in the sense that:
• the complete controllability property is preserved under arbitrary small perturba-
tions of the coefficients;
• the complete controllability property can be achieved by means of suitable small
perturbations of the coefficients.
These considerations can be also resumed by saying that complete controllability
is an open-dense property. 

Remark 5.4 If the input of system (5.1) is scalar i.e., m = 1, matrix B reduces to
a single column b and the controllability matrix is square. Checking the complete
78 5 Controllability and Observability of Linear Systems

controllability condition reduces to compute the determinant of the controllability

matrix. It is easily seen that such a system is completely controllable if and only if
b is cyclic for A (see Sect. 2.10).
Consider in particular a SISO system defined by a linear nonhomogeneous equa-
tion of order n
y (n) + a1 y (n−1) + · · · + an−1 y
+ an y = u(t) (5.22)

where the forcing term plays the role of a (scalar) input, and y is reviewed as a
(scalar) output. According to the procedure illustrated in Sects. 2.10 and 4.3, (5.22)
can be rewritten in the form (5.1) with A a companion matrix,
⎛ ⎞
0
⎜ .. ⎟
⎜ ⎟
b = ⎜.⎟
⎝0 ⎠
1

and C reduced to the row (1 0 . . . 0). The state variable coincides with the vector
x = (y, y
, . . . , y (n−1) ).
We can easily check that every system of the form (5.22) is completely controllable
regardless the choice of the coefficients a1 , . . . , an . 

In force of the conclusions of Theorem 5.3, we can slightly simplify our notation:
from now on, we write R instead of R(T, 0).

5.1.5 Hautus’ Criterion

Conditions equivalent to complete controllability of the system (5.1) can be given in

several different forms. In this section we present a criterion which will be sometimes
recalled in our future developments.
Theorem 5.4 (Hautus’ criterion) System (5.1) is completely controllable if and only
if
∀λ ∈ C, rank (A − λI |B) = n . (5.23)

We remark that (5.23) is trivially fulfilled if λ is not an eigenvalue of A. Note also

that in general, (A − λI |B) is a matrix with complex entries. In order to prove the
theorem, it is therefore advisable to interpret also A and B as operators acting on
complex spaces.
Definition 5.4 A subspace V of Cn is called a (complex) algebraic invariant for A
if AV ⊆ V .

Lemma 5.1 If a subspace V is an algebraic invariant for A, then there exists an

eigenvector v = 0 of A such that v ∈ V .
5.1 The Reachable Sets 79

Proof Let q = dim V and let v1 , . . . , vn be a basis of Cn , such that its first q elements
v1 , . . . , vq constitute a basis of V . With respect to this basis, A takes the form
 
A11 A12
0 A22

by virtue of the invariance assumption. The operator acting from V to V and defined
by the matrix A11 will necessarily have at least one eigenvector v ∈ V . It is not
difficult to check that the same vector v, reinterpreted as an element of Cn , is an
eigenvector of A corresponding to the same eigenvalue λ. 

Proof of Theorem 5.4 We show that (5.21) implies (5.23). Assume by contradiction
that for some λ ∈ C, the n rows of matrix (A − λI |B) are linearly independent.
Then, there exists a vector η ∈ Cn (η = 0) such that

η t A = λη t and η t B = 0 .

In particular, the function

λ2 t 2 t
ϕ(t) = eλt η t B = (eλt η)t B = η t B + λtη t B + η B + ···
2
must vanish. Thus we will have

ϕ(0) = ϕ
(0) = ϕ

(0) = · · · = 0 .

By applying the theorem about the derivative of a power series, and taking into
account
η t A = λη t =⇒ η t A2 = λη t A = λ2 η t etc.

we finally obtain
η t B = η t AB = η t A2 B = · · · = 0 .

This implies that the n rows of matrix (B|AB| . . . |An−1 B) are linearly dependent,
so that its rank is not equal to n.
Finally, we show that (5.23) implies (5.21). According to Cayley-Hamilton Theo-
rem, if (5.21) is false then the rows of all the matrices of the form A j B ( j = 0, 1, . . .)
will belong to a same proper subspace of Cn . In other words, we could find a vector
v = 0 such that
v t B = v t AB = v t A2 B = · · · = 0 . (5.24)

Setting w = At v, we have

w t B = (At v)t B = v t AB = 0, w t AB = (At v)t AB = v t A2 B = 0, etc.

80 5 Controllability and Observability of Linear Systems

Denoted by V the subspace of Cn constituted by all vectors v for which (5.24)

holds, we have so proved that if v ∈ V , also w = At v ∈ V , and this in turn implies
that V is an algebraic invariant with respect to the linear operator associated to matrix
At . Then by Lemma 5.1, it must exists a nonzero vector η ∈ V and a complex number
λ ∈ C such that
At η = λη that is η t A = λη t .

As a consequence of the definition of V , we have in particular that η t B = 0. In

conclusion, the rows of (A − λI |B) are linearly independent and (5.23) does not
hold. 

5.2 Observability

In common applications, the output variable does not coincide with the state variable.
In these cases, the observability function plays an essential role.

Definition 5.5 We say that two points x0 , η0 ∈ Rn are indistinguishable at time T

if for each admissible input u(·) : [0, T ] → Rn one has

y(t, x0 , u(·)) = y(t, η0 , u(·)) ∀t ∈ [0, T ] .

The previous definition is inspired by the following idea: for each fixed input
function u(·), if the initial state x0 is replaced by η0 , then the system response remains
unchanged. In other words, it is not possible in general to reconstruct exactly the
initial state on the base of information obtained uniquely by monitoring the output
corresponding to a known input.
Example 5.1 Consider the system

ẋ1 = x1 + u
ẋ2 = x2

with y = x1 − x2 . The solution corresponding to an initial state of the form (a, a) is

easily found:   
t
x 1 = et a + e−τ u(τ )dτ , x2 = aet .
0

t
Hence, we see that y(t) = et 0 e−τ u(τ )dτ is independent of a. In other words,
two distinct arbitrary points on the line x1 = x2 are indistinguishable. 
Let us emphasize that in practical applications, the knowledge of the initial state is
an important issue. Assume that we have a physical system, and that a mathematical
model has been constructed. In principle, the mathematical model should be used to
5.2 Observability 81

simulate the evolution of the physical system and to predict the future behavior. To
this end, we need to integrate analytically or numerically the system equations. But
this is impossible, if we do not known how to set the initial state of the model, which
should be the same as the initial state of the physical system.

5.2.1 The Unobservability Space

Our aim now is to characterize those systems for which there exist no pairs of
indistinguishable points. First of all, we remark that in Definition 5.5, the role of the
input function is unessential, in the sense explained by the following proposition.

Proposition 5.3 The points x0 and η0 are indistinguishable at time T for the system
(5.1) if and only if they are indistinguishable at time T for the unforced system

ẋ = Ax
(5.25)
y = Cx .

Proof If x0 and η0 are indistinguishable at time T for the system (5.1), then for each
u(·) ∈ PC([0, T ], Rm ) and each t ∈ [0, T ] we have

y(t, x0 , u(·)) = y(t, η0 , u(·))

that is
 t  t
−τ A
C[e (x0 +
tA
e Bu(τ ) dτ )] = C[e (η0 +
tA
e−τ A Bu(τ ) dτ )] .
0 0

Getting rid of the common term, we obtain the identity

Cet A x0 = Cet A η0

for each t ∈ [0, T ]. This actually means that x0 and η0 are indistinguishable with
respect to the system (5.25). The reverse argument proves the vice versa. 

Next proposition points out that in order to characterize the set of points which
are indistinguishable from a fixed x ∈ Rn , it is sufficient to characterize the set of
points which are indistinguishable from the origin.

Proposition 5.4 If x0 , η0 are indistinguishable at the time T > 0, then ξ = x0 − η0

is indistinguishable from the origin at time T . Vice versa, if ξ is indistinguishable
from the origin at the time T and if x0 is any vector of Rn , then x0 and η0 = x0 + ξ
are indistinguishable each other at time T .
82 5 Controllability and Observability of Linear Systems

Proof By virtue of Proposition 5.3, we can refer to system (5.25). From the assump-
tion that x0 and η0 are indistinguishable at time T , we deduce that

Cet A x0 = Cet A η0

for each t ∈ [0, T ]. This last equality rewrites

Cet A (η0 − x0 ) = 0 = Cet A 0

for each t ∈ [0, T ]. The statement follows, setting ξ = x0 − η0 . Vice versa, if ξ is

indistinguishable from the origin (which means that Cet A ξ = 0 for each t ∈ [0, T ]),
for each x0 ∈ Rn the equality

Cet A (ξ + x0 ) = Cet A ξ + Cet A x0 = Cet A x0

holds for each t ∈ [0, T ]. This means that ξ + x0 and x0 are indistinguishable at time
T for system (5.25), and so also for system (5.1). 

We denote by N(T, 0) the set of the states indistinguishable from the origin at
time T . A first characterization of N(T, 0) is provided by the following Theorem.

Theorem 5.5 The following statements are equivalent.

(i) ξ is indistinguishable from the origin at time T for system (5.1);
(ii) ξ ∈ ker Cet A , ∀t ∈ [0, T ];
(iii) the output function of system (5.1) corresponding to the input u(t) = 0 for each
t ∈ [0, T ) and to the initial state ξ, vanishes on [0, T ].

Proof The equivalence between (i) and (ii) follows from Proposition 5.3. The equiv-
alence between (ii) and (iii) is straightforward. 

Theorem 5.5 (ii) implies in particular that the set N(T, 0) coincides with

ker Cet A . (5.26)
t∈[0,T ]

But (5.26) is a subspace of Rn . Hence N(T, 0) is a subspace of Rn . It is called the

unobservability space.

5.2.2 The Observability Matrix

Now, consider the matrices

C t , At C t , . . . , (At )n−1 C t .
5.2 Observability 83

Their columns can be interpreted as vectors of Rn . Let V ⊂ Rn be the space

engendered by these vectors.
Theorem 5.6 Given the system (5.1), the unobservability space is independent of
T . In fact, for each T > 0 one has N(T, 0) = V ⊥ .

Proof We limit ourselves to sketch the main steps, since the proof is similar to that
of Theorem 5.3. Let v ∈ N(T, 0). Then for all θ ∈ [0, T ] and all μ ∈ R p

(Ceθ A v)t μ = 0

or
v t eθ A C t μ = 0 .
t

For θ = 0, we find v t C t μ = 0 and, being μ arbitrary, v is orthogonal to all the

columns of C t . Next step is to compute iteratively the derivatives of any order with
respect to θ. Each derivative is evaluated at θ = 0. Vice versa, if v ∈ V ⊥ , then for
each μ ∈ R p
v t C t μ = · · · = v t (At )n−1 C t μ = 0 .

Making use, as in Theorem 5.3, of the series expansion of the exponential and
of Cayley-Hamilton Theorem, this implies in turn that v t eθ A C t μ = 0 for each θ.
t

Finally,
(Ceθ A v)t μ = 0 ∀θ ∈ R

which implies Ceθ A v = 0, for each θ ∈ R. 

Matrix (C t |At C t | . . . |(At )n−1 C t ) is called the observability matrix of system

(5.1). From now on, since N(T, 0) is independent of T , we write simply N.

Definition 5.6 The system is said to be completely observable when

rank (C t |At C t | . . . |(At )n−1 C t ) = n . (5.27)

Corollary 5.5 System (5.1) is completely observable if and only if for each pair of
indistinguishable states x0 , η0 we have x0 = η0 , or, equivalently, when N = {0}.

Remark 5.5 Neither (ii) of Theorem 5.5 nor (5.27) depend on matrix B. This is not
surprising, if we have in mind Proposition 5.3. 

Remark 5.6 Since the unobservability space N does not depend on T , we may say
that a point x0 ∈ N if and only if Ceθ A x0 = 0 for each θ ≥ 0. On the other hand, from
Theorem 5.6 it follows that if we replace A by −A, the space N does not change.
Hence, the previous statement can be strengthened, writing that x0 ∈ N if and only
if Ceθ A x0 = 0 for each θ ∈ R. This implies in turn that N is dynamically invariant
(compare with Definition A.4) with respect to the unforced system. Indeed, if x0 ∈ N
and η = et A x0 for t ∈ R, we have
84 5 Controllability and Observability of Linear Systems

Ceθ A η = Ceθ A et A x0 = Ce(θ+t)A x0 = 0

for each θ ∈ R, and so η ∈ N. 

Remark 5.7 It is easily checked that a system (5.1) defined by means of a scalar
linear differential equation of order n,

y (n) + a1 y (n−1) + · · · + an−1 y

+ an y = u(t)

where u is interpreted as the input and y as the output, is completely observable,

regardless the choice of the coefficients a1 , . . . , an . 

5.2.3 Reconstruction of the Initial State

For a fixed input function u(·) : [0, +∞) → Rm , let us look at the map which asso-
ciates the output function y(t) to a given initial state x(0) = x0 . Complete observ-
ability of system (5.1) implies that such a map is injective. Hence, we expect that
monitoring y(t) on the interval [0, T ] (for some T > 0) provides sufficient informa-
tion in order to recover the exact value of x0 . Next we show how this can be actually
done. Assume that u(t) and y(t) are known for t ∈ [0, T ]. Recall that
  t
(t−τ )A
y(t) = C e x0 +
tA
e Bu(τ ) dτ .
0

t
Multiplying both sides by et A C t we get
 t
e(t−τ )A Bu(τ ) dτ
t t t
et A C t y(t) = et A C t Cet A x0 + et A C t C
0

and integrating from 0 to T :

 T  T  t 
e(t−τ )A Bu(τ ) dτ
t t
E(T )x0 = et A C t y(t) dt − et A C t C dt , (5.28)
0 0 0

T t
where we set E(T ) = 0 et A C t Cet A dt.
Theorem 5.7 The following properties are equivalent.
(i) System (5.1) is completely observable.
(ii) Matrix E(T ) is positive definite for each T > 0, and so invertible.
Proof A simple computation shows that
 T
ξ E(T )ξ =
t
||Cet A ξ||2 dt ≥ 0 .
0
5.2 Observability 85

The integral on the right-hand side is zero if and only if the integrand vanishes,
that is if and only if ξ ∈ ker Cet A for each t ∈ [0, T ]. If the system is completely
observable, this may happen only if ξ = 0. The opposite statement can be easily
proven by contradiction. 

Therefore, if (5.1) is completely observable, from (5.28) it is possible to exactly

compute x0 taking the inverse of the matrix E(T ). We emphasize that in this procedure
the input is absolutely arbitrary: the more natural choice is, of course, u(t) = 0 for
each t ∈ [0, T ].
The approach to the observability problem described in this section has a draw-
back: indeed, (5.28) may be sometimes hard to apply, because of the need of com-
puting the exponential matrix, some integrals and an inverse matrix.

5.2.4 Duality

The analogies between the notions of controllability and observability are evident.
We point out that the observability matrix of system (5.1) coincides with the con-
trollability matrix of system

ẋ = At x + C t u
(5.29)
y = Bt x

where u ∈ R p and y ∈ Rm . Note that with respect to (5.1), the roles of B and C are
exchanged. Thus, (5.1) is completely controllable if and only if (5.29) is completely
observable and vice versa. System (5.29) is called the dual of (5.1). The proper-
ties of complete controllability and complete observability are also said to be dual
properties.
We emphasize also the analogies (and the differences) between Theorems 5.2
and 5.7.

5.3 Canonical Decompositions

In the analysis of a system, it is important to find out certain canonical forms; they
are particular representations which make possible to understand at a first glance the
main structural properties of the system. This requires the search for suitable changes
of coordinates.
86 5 Controllability and Observability of Linear Systems

5.3.1 Linear Equivalence

The systems  
ẋ = Ax + Bu ż = Ãz + B̃u
and
y = Cx y = C̃ z

are called linearly equivalent if there exists a linear change of coordinates x = Pz

(det P = 0) such that à = P −1 A P, B̃ = P −1 B, and C̃ = C P. We recognize in this
definition a generalization of a notion already introduced for linear unforced system,
and fruitfully applied in Chap. 2.
We remark that such a transformation does not affect the controllability properties
of the system. Indeed, one has

( B̃| Ã B̃| . . . | Ãn−1 B̃) = (P −1 B|P −1 AB| . . . |P −1 An−1 B)

= P −1 (B|AB| . . . |An−1 B)

so that the controllability matrices of any pair of linearly equivalent systems have the
same rank. Moreover, the subspaces engendered by the columns of these matrices
(that is, the controllability spaces of these systems) are consistently transformed each
other by the change of coordinates.
Similar conclusions can be achieved, of course, about the observability property
and the unobservability space.

5.3.2 Controlled Invariance

Before to introduce the first important canonical form, we still need a definition. A
subspace W ⊂ Rn is said to be a controlled invariant for system (5.1) if for each
x0 ∈ W and for each admissible input we have:

x(t, x0 , u(·)) ∈ W ∀t > 0 .

We recognize in this definition an extension of Definition A.4. The space R is

an example of controlled invariant. Indeed, assume by contradiction that there exist
a point x0 ∈ R and an input function u(·) : [0, T ] → Rm such that x(T, x0 , u(·)) =
η∈/ R. If u 0 (·) : [0, t0 ] → Rm is an input function for which x(t0 , 0, u 0 (·)) = x0 (at
least one such input function exists by hypothesis) we can take the new input

u 0 (τ ) for 0 ≤ τ < t0
ũ(τ ) =
u(τ − t0 ) for t0 ≤ τ ≤ t0 + T .

Clearly x(t0 + T, 0, ũ(·)) = η, and this is a contradiction.

5.3 Canonical Decompositions 87

Lemma 5.2 Let the subspace W ⊂ Rn be a controlled invariant for system (5.1)
and let
x(t, x0 , u(·))

a trajectory with x0 ∈ W . Then the tangent vector of the curve t → x(t, x0 , u(·)) for
t = 0 belongs to W , as well.

Proof Let us consider the difference quotient

x(t, x0 , u(·)) − x0
q(t) =
t

for t = 0. The tangent vector is defined as limt→0 q(t). Since W is a subspace,

q(t) ∈ W for each t = 0. Hence, the limit must belong to W , as well, being a subspace
a closed set. 

5.3.3 Controllability Form

Theorem 5.8 There exist a change of coordinates x = Pz and an integer q (0 ≤

q ≤ n) such that in the new coordinates z the system (5.1) takes the form

ż 1 = A11 z 1 + A12 z 2 + B1 u
(5.30)
ż 2 = A22 z 2

where z = (z 1 , z 2 ) with z 1 ∈ Rq , z 2 ∈ Rn−q , and where A11 , A12 , A22 , and B1 are
matrices of suitable dimensions. Moreover, the system

ż 1 = A11 z 1 + B1 u (5.31)

with state variable z 1 ∈ Rq , is completely controllable.

Proof Let q = dim R. The limit cases q = 0 and q = n correspond respectively to

the cases where (5.1) is completely uncontrollable (that is B = 0) and the case where
(5.1) is completely controllable. So, we can limit ourselves to assume (0 < q < n).
Consider a basis of the state space, such that the first q vectors form a basis of R. Let z
be the coordinates in this new basis, partitioned in such a way that R = {z : z 2 = 0}.
In general, the representation of the system in these new coordinates can be written

ż 1 = A11 z 1 + A12 z 2 + B1 u
ż 2 = A21 z 1 + A22 z 2 + B2 u .

We show that, according to the particular choice of the basis, A21 = B2 = 0. Recall
that R is a controlled invariant, and notice that this property does not depend on the
88 5 Controllability and Observability of Linear Systems

choice of the coordinates. Assume that A21 = 0. Let us take an initial state (z 1 , 0) ∈ R
with z 1 = 0, A21 z 1 = 0. Let moreover u(t) = 0 for each t ≥ 0. The tangent vector
to the corresponding solution, evaluated at t = 0, is
 
A11 z 1
∈
/ R.
A21 z 1

This is a contradiction to Lemma 5.2. Hence A21 = 0. In a similar way, it is

possible to show also that B2 = 0. In practice, the matrix P which determines the
change of coordinates can be written as

P = (v1 | . . . |vq |vq+1 | . . . |vn )

where v1 , . . . , vq are chosen in such a way that they form a basis of R, and
vq+1 , . . . , vn in such a way that they form, together with v1 , . . . , vq , a basis of
Rn . Notice that the basis v1 , . . . , vn is not uniquely determined by this construc-
tion. For instance, it is not restrictive (in fact, for future developments, it is strongly
recommended) to take the vectors v1 , . . . , vn pairwise orthogonal.
It remains to prove that (5.31) is completely controllable. Let z 1 ∈ Rq be given.
By construction, there exists an input function u(·) such that the corresponding
solution of system (5.30) steers the origin of Rn to the state (z 1 , 0) ∈ R. Obviously,
the same input applied to system (5.31) steers the origin of Rq in z 1 . Hence, (5.31)
is completely controllable. 

The Eq. (5.31) can be obtained from (5.30) setting z 2 = 0. It can be therefore in-
terpreted as a subsystem: it is called the controllable part of the overall system (in the
figure above, it is denoted by 1 ). Notice that the evolution of the component z 2 of the
state in (5.30) does not depend at all on the action of the input function. It represents
the uncontrollable part of the system (in the figure above, it is denoted by 2 ).
The form (5.30) reveals the structure of the system, and in particular it allows us
to separate and recognize the controllable and uncontrollable parts.

5.3.4 Observability Form

An analogous construction, based on linear equivalence, allows us to obtain a form of

system (5.1) which reveals the observability properties. More precisely, it is possible
5.3 Canonical Decompositions 89

to prove the existence of an integer r (0 ≤ r ≤ n) and a nonsingular matrix P such

that the change of coordinates x = Pz gives rise to the form
⎧
⎪
⎨ż 1 = A11 z 1 + A12 z 2 + B1 u
ż 2 = A22 z 2 + B2 u (5.32)
⎪
⎩
y = C2 z 2

where z = (z 1 , z 2 ), z 1 ∈ Rr , z 2 ∈ Rn−r , and the reduced order system


ż 2 = A22 z 2 + B2 u
(5.33)
y = C2 z 2

with state variable z 2 ∈ Rn−r , is completely observable.

Note that if we put z 2 = 0 in the differential part of (5.32), the resulting reduced
order system
ż 1 = A11 z 1 + B1 u (5.34)

with state variable z 1 ∈ Rr , does not produce any output. The reduced order systems
(5.33) and (5.34) are called, respectively, the observable part (denoted by 2 in the
figure below) and the unobservable part (denoted by 1 in the figure below) of the
system.

However, this time the construction of the matrix P which determines the change
of coordinates is more delicate. We start by computing the observability matrix. Let
n − r be its rank. Choose n − r linearly independent columns of the observability
matrix, and let us denote them by vr +1 , . . . , vn . Choose finally r linearly independent
vectors v1 , . . . , vr such that the subspace generated by v1 , . . . , vr is orthogonal1 to the
subspace generated by vr +1 , . . . , vn . According to Theorem 5.6, the vector v1 , . . . , vr
constitute a basis of the non-observability subspace N. A possible choice of P is the
matrix whose columns are

P = (v1 | . . . |vr |vr +1 | . . . |vn ) .

1 We stress that in general, the construction does not work if the orthogonality requirement is
neglected: this is an important difference with respect to the construction of the controllability
form. The reason of this fact is implicit in the statement of Theorem 5.6.
90 5 Controllability and Observability of Linear Systems

The construction can be easily modified, in order to replace P by a new matrix

whose columns are all pairwise orthogonal. To prove that, after the change of coor-
dinates, the block A21 becomes zero, we may argue as in the proof of Theorem 5.8,
setting u = 0 and taking into account Remark 5.6.

5.3.5 Kalman Decomposition

The controllability form and the observability form discussed in the previous sections
can be combined, giving rise to the form
⎧
⎪
⎪ ż 1 = A11 z 1 + A12 z 2 + A13 z 3 + A14 z 4 + B1 u
⎪
⎪
⎪
⎪
⎨ż 2 = A22 z 2 + A24 z 4 + B2 u
ż 3 = A33 z 3 + A34 z 4 (5.35)
⎪
⎪
⎪
⎪ ż 4 = A44 z 4
⎪
⎪
⎩y = C z + C z .
2 2 4 4

The special structure exhibited by (5.35) corresponds to the connections displayed

in the figure below, where by 1 , 2 , 3 , 4 we have denoted the reduced order
systems which determine the evolutions of the blocks of coordinates z 1 , z 2 , z 3 , z 4 ,
respectively.

We remark in particular that:

• the block of coordinates (z 1 , z 2 ) identifies the completely controllable part: the
form (5.30) is recovered as a particular case, by collecting the blocks of coordinates
(z 1 , z 2 ), (z 3 , z 4 );
• the block of coordinates (z 2 , z 4 ) identifies the completely observable part: the form
(5.32) is recovered by collecting the blocks of coordinates (z 1 , z 3 ), (z 2 , z 4 ) and
rewriting the equations, after reordering the indices in the following way: 1, 3, 2, 4;
• the block of coordinates z 2 identified the completely controllable and completely
observable part;
• the block of coordinates z 3 identifies the uncontrollable and unobservable part.
5.3 Canonical Decompositions 91

These remarks are trivial, with the exception of the third one, for which we provide
a sketch of proof. Because of the block-triangular form of the matrix, we realize that
the controllability matrix for the block of coordinates z 2 can be obtained taking suit-
able submatrices of the controllability matrix for the block (z 1 , z 2 ). Such a matrix has
a maximal rank, since the block (z 1 , z 2 ) corresponds to the completely controllable
part of the overall system. But this is possible only if the rank of the controllability
matrix for the block z 2 is maximal. The complete observability is proved in similar
way.

5.3.6 Some Examples

In order to illustrate the construction of the canonical forms presented in the previous
sections we present some examples.
Example 5.2 Consider the system with scalar input defined by the matrices
   
1 4 4
A= b= .
2 −6 1

We do not need to specify C, since in this example we will beinterested

 only
4 8
in the controllability form. The controllability matrix is (b|Ab) = and its
1 2
rank is equal to 1. The system is not completely controllable and we can proceed to
the determination of a controllability form. We perform a change of coordinates by
means of the matrices
   
4 −1 1 4 1
P= P −1 =
1 4 17 −1 4

(notice that the columns of P are orthogonal). We obtain, as desired,

   
−1 2 −2 −1 1
P AP = P b= .
0 −7 0

If we chose a different matrix, with the first column parallel to b and the second
column linearly independent (but not necessarily orthogonal) to the first one, we
obtain again a controllability form, which in general may differ from the previous
one for some unessential details. For instance, with
 
4 1
Q=
1 0

we have    
2 2 1
Q −1 AQ = Q −1 b = .
0 −7 0


92 5 Controllability and Observability of Linear Systems

Example 5.3 Now consider the system with scalar input and scalar output defined
by the matrices    
3 2 1
A= b= c = (1 1) .
3 4 0

This system
 is completely controllable but not completely observable. Indeed,
6
since A c =
t t
, the rank of the observability matrix is 1. We can proceed to
6
determine an observability canonical form. Define a change of coordinates by the
matrices
   
−1 1 −1 1 −1 1
P= P = .
1 1 2 1 1

We get
   
1 1 1 −1
P −1 A P = P −1 b = c P = (0 2) .
0 6 2 1

Note that the columns of P are orthogonal. Note also that with this procedure, the
unobservability space becomes coincident with the first component. 

Example 5.4 Consider finally the system with scalar input and scalar output defined
by the matrices
   
3 −2 1
A= b= c = (−1 2) .
1 0 1

In this case, both the controllability and the observability matrices have rank 1. The
system is neither completely controllable nor completely observable. It is convenient
to start by computing an observability form. However, now it is preferable to proceed
in a lightly different way. Making use of the change of coordinates defined by the
matrices    
−1 2 1 −1 2
P= P −1 = , (5.36)
2 1 5 2 1

we may align, in the new coordinates, the unobservability space with the second
component instead of the first one, as we did in Example 5.3. We have:
   
−1 1 0 −1 1 1
P AP = = Â P b= = b̂ c P = (5 0) = ĉ .
−3 2 5 3

Notice that the columns of P are orthogonal. Notice also that this form does
not allow to identify immediately the controllability space. Thus, we need to apply
a further change of coordinates, to the purpose of achieving a complete Kalman
5.3 Canonical Decompositions 93

decomposition. More precisely, we need to find a change of coordinates which,

while keeping unchanged the position of the unobservability space, superposes the
controllability space to the subspace orthogonal to the unobservability one (that is,
with the first component). We can take:
   
1 0 −1 1 0
Q= Q =
3 1 −3 1

(the first column of Q is parallel to b̂, the second generates the unobservability space
in the new coordinates). We finally obtain the required form:
   
−1 1 0 −1 1 1
Q ÂQ = Q b̂ = ĉ P = (5 0) .
0 2 5 0

It is now evident, in particular, that the system possesses a completely controllable

and completely observable part, and a uncontrollable and unobservable part.
Of course, the transformation can be accomplished by a unique change of coor-
dinates defined by the matrix
 
5 2
PQ = .
5 1


Example 5.5 As a last example we take the same matrices A and c as in Example
5.4, but  
2
b= .
1

By the first change of coordinates given by (5.36), we get

 
−1 0
P b= .
1

The controllability space is already coincident with the unobservability space. Of

course, now it is not possible, by a further change of coordinates, to move the control-
lability space in such a way that it becomes orthogonal to the unobservability space.
The complete Kalman form has been obtained with the first change of coordinates.
The system possesses a controllable part which is not observable, and an observable
part which is not controllable. 

5.4 Constrained Controllability

One of the most important developments of control theory is optimization. Typical

examples are the minimal time problem [16] and the quadratic regulator problem
[6]. In particular, minimal time problems are strictly related to the controllability
94 5 Controllability and Observability of Linear Systems

properties of the system. However, a minimal time problem makes sense only if the
admissible control functions are constrained to take values in a bounded set. Although
optimization is beyond the purposes of this book, in this section we discuss shortly
how the geometric properties of the reachable sets change when the admissible
controls are subject to constraints. Thus, in this section we consider linear systems

ẋ = Ax + Bu (5.37)

with x ∈ Rn and u ∈ U , where U represents a nonempty, proper (in general, bounded)

subset of Rm . To begin with, we need to update the notion of reachability introduced
in the previous Sect. 5.1 and the related notation.
Let x0 , η0 ∈ Rn . We say that η0 is reachable from x0 at time T > 0 with constrained
controls if there exists u(·) ∈ PC([0, +∞), U ) such that x(T, x0 , u(·)) = η0 , and we
denote R(T, x0 , U ) the set of such points. This is called the reachable set of (5.37)
with constrained controls.
The geometric properties of R(T, x0 , U ) depend not only on the pair of matrices A
and B, but also on the set U . For instance, it is no more true in general that R(T, 0, U )
is a subspace of Rn . As a consequence, we should enrich the notion of reachability
introducing some new definitions.

Definition 5.7 A system (5.37), with admissible inputs constrained to a set U , is

said to be:
◦
• accessible from x0 at time T when R(T, x0 , U ) = ∅;
◦
• locally reachable from x0 at time T when x0 ∈ R(T, x0 , U );
• locally reachable along the free solution from x0 at time T if x(T, x0 , 0) ∈
◦
R(T, x0 , U ).

Notice that when x0 = 0, the definitions of local reachability and local reachability
along the free solution coincide.

Example 5.6 Consider the simple scalar system

ẋ = x + u, x, u ∈ R.

Let u 0 > 0 be fixed and assume that the control functions are subject to the
constraint |u(t)| ≤ u 0 . Solving the equation, we have for t ≥ 0
 t
x(t) = et x0 + e(t−τ ) u(τ ) dτ
0

that is  t  t
et x 0 − e(t−τ ) u 0 dτ ≤ x(t) ≤ et x0 + e(t−τ ) u 0 dτ
0 0
5.4 Constrained Controllability 95

namely
et (x0 − u 0 ) + u 0 ≤ x(t) ≤ et (x0 + u 0 ) − u 0

and finally
et x0 − (et − 1)u 0 ≤ x(t) ≤ et x0 + (et − 1)u 0 .

This shows that the system is accessible and locally reachable along the free
solution, for each choice of x0 and u 0 (recall that if t ≥ 0 then et − 1 ≥ 0).
If −u 0 < x0 < u 0 , then x0 belongs to the interval (et (x0 − u 0 ) + u 0 , et (x0 +
◦
u 0 ) − u 0 ) = R(t, x0 , U ) for each t. Hence, the system is also locally reachable at x0 .
Moreover, x0 − u 0 < 0 < x0 + u 0 so that R(x0 , U ) = R. On the contrary, if x0 ≥ u 0
then R(x0 , U ) coincides with the half line [x0 , +∞). In this case the system is not
locally reachable. The conclusion is the same for x0 ≤ −u 0 . 

The study of the reachability properties of a linear system with constrained input
can be actually reduced to the case x0 = 0.

Proposition 5.5 Let a system of the form (5.37) be given, with admissible control
functions constrained to a subset U ⊂ Rm .
(i) The system is accessible from x0 at time T if and only if it is accessible from the
origin at the time T .
(ii) The system is locally reachable from x0 at time T along the free solution if
and only if it is locally reachable from the origin at time T or, equivalently,
◦
0 ∈ R(T, 0, U ).

The proof is a straightforward application of the variation of constants formula.

Notice that if the constraints are relaxed, system (5.37) is accessible from any initial
state if and only if R(T, 0, Rm ) = Rn , that is if and only if the system is globally
reachable at time T .
We already noticed that if U is a proper subset of Rm , then R(T, 0, U ) is no more,
in general, a subspace of Rn . However, it preserves an important property.

Proposition 5.6 Consider the system (5.37). For each T ≥ 0 and for each nonempty
constraint set U , the set R(T, 0, U ) is convex.

The proof is trivial if U is convex. Otherwise, some advanced results of measure

theory are needed (see [23] p. 163, [12] p. 11).

Theorem 5.9 Assume that system (5.37) is completely controllable. Assume in ad-
◦ ◦
dition that 0 ∈ U . Then, 0 ∈ R(T, 0, U ).

Proof Let B be a ball centered at the origin. Let r be radius of B, chosen in such a
way B ⊂ U . Since the system is completely controllable, for each unit vector ei of
the canonical basis of Rn there exists a control function u i (t) : [0, T ] → Rm which
steers the system from the origin to ei at time T . The control functions u i (t) need
96 5 Controllability and Observability of Linear Systems

not to met the prescribed constraints but, since they are piecewise continuous and
hence bounded on [0, T ], there exists M > 0 such that

|u i (t)| < M ∀t ∈ [0, T ], ∀i = 1, . . . , n .

Since the system is linear, the controls

r
ũ i (t) = u i (t)
M

steer the system from the origin to some vectors ẽi which still constitute a basis
of Rn . These new control functions satisfy the required constraints. Invoking again
the linearity of the system, we finally see that the points −ẽi can be reached by means
of the control functions −ũ i (t). The conclusion follows, since R(T, 0, U ) is
convex. 
Corollary 5.6 Under the assumptions of Theorem 5.9, the system possesses the
property of local reachability along the free solution.
Corollary 5.7 Assume that system (5.37) is completely controllable. Moreover, as-
◦
sume that U = ∅. Then, the system possesses the property of accessibility from the
origin for each T > 0.
◦
Proof Let u 0 ∈ U . By assumption, there exists a ball of positive radius centered at
u 0 , which is contained in U . Replacing U by

U − {u 0 } = {v : v = u − u 0 with u ∈ U }

we obtain a system which satisfies the assumptions of Theorem 5.9. Thus, it is

sufficient to remark that
 T
R(T, 0, U ) = R(T, 0, U − {u 0 }) + e(T −s)A Bu 0 ds .
0


Chapter Summary

In this chapter we deal with the so-called structural properties of a linear system
with input and output. These properties depend only on the coefficients of the math-
ematical model. We study in particular controllability (which provides a measure
of our ability to control the system) and observability (which provides a measure
of our ability of extracting information about the state of the system). We obtain
algebraic characterizations of these properties. We also study canonical forms i.e.,
linear transformations of the state space which allow us to rewrite the model. This
makes more evident, in this way, recognizing the controllability and observability
properties.
Chapter 6
External Stability

The classical notions of stability and asymptotic stability are no more sufficient to
describe the behavior of a system, in the presence of external input. Consider for
instance a system for which the origin is stable when the external inputs are switched
off. Likely, restoring the external forces, a deviation from the equilibrium will be
observed. It seems natural to expect that the amplitude of the deviation is related to
the amplitude of the input signal, and that it will be “small” in some sense, if the
input signal is “small”. As already mentioned in Chap. 1, this kind of behavior is
informally qualified as external stability. However, simple stability is not sufficient
to guarantee such a natural behavior, as shown for instance by the simple example
ẋ = u 0 , where x ∈ R and u 0 is a nonzero constant.
In Chap. 4 we proved that a linear differential system with a constant (respectively,
periodic) forcing term has a constant (respectively, periodic) solution, provided that
resonance does not occur: such solutions are, of course, bounded. For constant (or
periodic) forcing terms, resonance is surely avoided if all the eigenvalues of the
unforced system have negative real part. Moreover, in this case all the solutions
approach the constant (or periodic) one and are so bounded. This remark points out
that the right property to be considered in order to characterize external stability is
internal stability.
In this chapter we are actually interested in studying the relationship between the
external stability and the internal stability of a linear system of the form

ẋ = Ax + Bu
(6.1)
y = Cx

where, with the usual notation, x ∈ Rn , u ∈ Rm , y ∈ R p . Recall that informally, it

is used to say that the system (6.1) is internally stable when the associate unforced
system
ẋ = Ax (6.2)

© Springer Nature Switzerland AG 2019 97

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_6
98 6 External Stability

has an asymptotically stable equilibrium point at the origin. We will see in particular
that internal stability actually implies external stability, but the converse does not
hold in general.
To the purposes of this chapter, the space of the input maps is restricted to
PCB([0, +∞), Rm ) endowed with the norm of the uniform convergence. The norm
of the uniform convergence will be adopted also for the output maps, provided that
they belong to CB([0, +∞), R p ). Moreover, we denote respectively by x(t, x0 , u(·))
and y(t, x0 , u(·)) the solution of the differential system and the output map, corre-
sponding to the initial state x0 and the input u(t).

6.1 Definitions

For reader’s convenience, we recall the definition of BIBO-stability already given in

Chap. 1 (see Definition 1.3).
Definition 6.1 We say that system (6.1) is BIBO-stable if for each R > 0 there exists
S > 0 such that

x0  ≤ R, u(·)∞ ≤ R =⇒ y(t, x0 , u(·)) ≤ S (6.3)

for each t ≥ 0, where x0 ∈ Rn and u(·) ∈ PCB([0, +∞), Rm ).

From (6.3) we infer in particular that for any BIBO-stable system the output map
y(·, x0 , u(·)) necessarily belongs to CB([0, +∞), R p ), and ||y(·, x0 , u(·))||∞ ≤ S.
If C = I (the unit matrix), inequality (6.3) holds with y(t, x0 , u(·)) replaced by
x(t, x0 , u(·)); in such a case we say, more appropriately, that the system is BIBS-
stable (bounded-input-bounded-state-stable). There is another possible definition of
external stability for system (6.1).

Definition 6.2 We say that system (6.1) has the finite gain stability property if there
exist two positive constant γ1 and γ2 such that

y(t, x0 , u(·)) ≤ γ1 x0  + γ2 u(·)∞ (6.4)

for each t ≥ 0, each x0 ∈ Rn and each input u(·) ∈ PCB([0, +∞), Rm ) (without
loss of generality, we may assume that γ1 = γ2 ).

At a first glance, inequality (6.4) appears more restrictive than (6.3): indeed, it
requires that the norm of the output is proportional to the norms of the input and of
the initial state. Our first task is to prove that, for linear systems, Definitions 6.1–6.2
are actually equivalent.

Theorem 6.1 Given a system of the form (6.1), the finite gain stability property is
necessary and sufficient for the BIBO-stability property.
6.1 Definitions 99

The proof of Theorem 6.1 requires several steps. The sufficient part is almost
trivial.
Proof of the sufficient part Assume that (6.4) holds. Then, (6.3) is fulfilled, if we
take for each R > 0, S = R(γ1 + γ2 ). 
To prove the necessary part, we need to establish some preliminary lemmas.
Lemma 6.1 Let a system of the form (6.1) be given. If it is BIBO-stable, then there
exists M > 0 such that ||Cet A || < M for t ≥ 0.
Proof By assumption, there exists a constant S1 such that if ||x0 || ≤ 1 then ||y(t, x0 , 0)
|| ≤ S1 for each t ≥ 0. Let us denote by ηi j (t) a generic entry of the matrix Cet A
and let us assume by contradiction that there exists a pair of indices i, j for which
ηi j (t) is not bounded for t ≥ 0. Take as x0 the j-th vector of the canonic basis, that
is the vector e j whose components are zero, except the j-th which is 1. The i-th
component of the function ψ(t) = Cet A e j is equal to ηi j (t) and we have

|ηi j (t)| ≤ ||ψ(t)|| .

The function ψ(t) = Cet A e j is therefore unbounded for t ≥ 0. But ψ(t) coincides
with y(t, e j , 0), and ||e j || = 1. We get in this way a contradiction to the BIBO-
stability assumption. Thus, we are led to conclude that all the entries of the matrix
Cet A are bounded, and the conclusion easily follows. 
Lemma 6.2 If the system (6.1) is BIBO-stable, then there exists L > 0 such that
 t
W (τ ) dτ < L (6.5)
0

for each t ≥ 0, where W (τ ) = Ceτ A B.

Remark 6.1 Note that W (τ ) is a matrix with p rows and m columns. Note also that
Eq. (6.5) holds if and only if the integral
 ∞
W (τ ) dτ
0

is convergent. 
Proof of Lemma 6.2 A system of the form (6.1) can be interpreted as an impulse
response system provided that:
(a) the initial state is equal to zero;
(b) the admissible input maps are assumed to vanish for t < 0.
Under these conditions, the impulse response matrix can be defined as

W (t) if t ≥ 0
h(t) = (6.6)
0 if t < 0.
100 6 External Stability

Indeed, by virtue of the variation of constants formula, when the system is initial-
ized to zero, the input-output operator writes
 t  t
(t−τ )A
y(t) = Ce Bu(τ ) dτ = W (t − τ )u(τ ) dτ
0 0

and hence, taking into account (b) and (6.6),

 +∞
y(t) = h(t − τ )u(τ ) dτ .
−∞

The proof can be therefore carried out by repeating the same arguments as in
Chap. 1 (necessary part of the proof of Proposition 1.5). 
In what follows, we sometimes refer the term “impulse response matrix” to the
matrix W (t). This is a little abuse, justified by the reasons explained in the previous
proof.

Remark 6.2 Lemma 6.2 can be restated by saying that in the case of linear systems,
BIBO-stability implies the boundedness of the impulse response. However in general,
boundedness of the impulse response and BIBO-stability are not equivalent. As an
example, we can consider the system

ẋ1 = x1
ẋ2 = −x2 + u

with C = I . This system has a bounded impulse response matrix, but it is not BIBO-
stable.
Thus we see that the sufficient part of Proposition 1.5 of Chap. 1 does not hold
for systems of the form (6.1): the reason of course is that the behavior of the system
depends here also on the initial conditions, and not only on the input (as for the
systems considered in Proposition 1.5). 

We are now in a position to complete the proof of Theorem 6.1.

Proof of the necessary part Assuming that the system (6.1) is BIBO-stable, we
show that it has the finite gain stability property, as well.
We first consider the output of the system (6.1) corresponding to the initial state
x0 = 0 and any admissible (bounded) input u(·). According to Lemma 6.2, with
L = γ2 , we get
 t  t
y(t, 0, u(·)) ≤ W (t − τ )u(τ ) dτ ≤ W (t − τ ) u(·)∞ dτ
0 0
 t
≤ u(·)∞ W (σ) dσ ≤ γ2 u(·)∞
0
6.1 Definitions 101

for each t ≥ 0. Then we consider the output of system (6.1) corresponding to any
initial state x0 and the constant input u(t) = 0. According to Lemma 6.1, with M =
γ1 , we have
||Cet A || ≤ γ1

for each t ≥ 0. This yields

||y(t, x0 , 0)|| ≤ γ1 ||x0 ||

for each x0 ∈ Rn and each t ≥ 0. Since for a linear system

y(t, x0 , u(·)) = y(t, x0 , 0) + y(t, 0, u(·))

the conclusion is readily achieved. 

From now on, when we say that a linear system is externally stable we mean that
one of the two Definitions 6.1 and 6.2 is (and hence both are) fulfilled. To conclude
this section, we show that external stability is invariant under linear changes of
coordinates in the state space Rn .

Proposition 6.1 Let the system


ż = Ãz + B̃u
(6.7)
y = C̃ z

be linearly equivalent to (6.1). System (6.7) is externally stable if and only if system
(6.1) is externally stable.

Proof Since the systems are linearly equivalent, there exists a nonsingular matrix P
such that à = P −1 A P, B̃ = P −1 B, C̃ = C P. Assume that (6.1) has the finite gain
stability property. By the variation of constants formula, the output of the system
(6.7) writes
 t  t
y(t) = C̃et à z 0 + C̃e(t−τ ) à B̃u(τ ) dτ = Cet A Pz 0 + Ce(t−τ )A Bu(τ ) dτ .
0 0

It coincides with the response of system (6.1) corresponding to the initial state
P z 0 and the same input map. We have

||y(t)|| ≤ γ1 ||Pz 0 || + γ2 ||u(·)||∞ ≤ γ̃1 ||z 0 || + γ2 ||u(·)||∞

where γ̃1 = ||P||γ1 . The proof of the converse statement can be achieved by exchang-
ing the roles of the systems. 
102 6 External Stability

6.2 Internal Stability

Recall that the origin is asymptotically stable for the unforced system (6.2) if and
only if all the eigenvalues of A have negative real part. In this case, we use to say that
A has the Hurwitz property or, in short, that A is a Hurwitz matrix (Definition 3.1).
We show that in general the internal stability property implies the external stability
one.
Theorem 6.2 Assume that the matrix A is Hurwitz. Then, the system (6.1) has the
finite gain stability property.
Proof By virtue of Proposition 4.2, for some constants k0 , α, b̄ we have

||x(t, x0 , u(·))|| ≤ k0 ||x0 ||eαt + b̄||u(·)||∞

for t ≥ 0. If A is Hurwitz, α can be taken negative, and so eαt < 1. The conclusion
follows, since ||y|| ≤ ||C|| · ||x||. 
Unfortunately, the converse of Theorem 6.2 is false in general.
Example 6.1 Let us consider the system

ẋ1 = 0
ẋ2 = −x2 + u

with the identity as observation function. It is easy to check that this system has the
finite gain stability property, but it is not internally stable. 
Next we address the problem of identifying additional conditions which allow us
to prove a partial converse of Theorem 6.2 or, alternatively, a property weaker than
asymptotic stability of the unforced system, which could be proved to be equivalent
to the external stability of (6.1). In the next section, we solve this problem for the
particular case where the matrix C is the identity. The general case will be studied
later in Sect. 6.4.

6.3 The Case C = I

If C = I , the output coincides with the state, and the system takes the form

ẋ = Ax + Bu
(6.8)
y=x.

Recall that under this condition the BIBO-stability property reduces to the BIBS-
stability one and, moreover, W (t) = et A B.
6.3 The Case C = I 103

Proposition 6.2 If system (6.8) is BIBO-stable, then the origin is a (in general, not
asymptotically) stable equilibrium point for the unforced system (6.2).

Proof Of course, if the system is BIBO-stable then the solutions of the unforced
system (6.2) must be bounded for t ≥ 0. Thus, the origin is stable (in Lyapunov
sense) for, the unforced system (6.2). 

On the other hand, simple examples (take for instance the scalar system ẋ = 1)
point out that the simple (not asymptotic) stability of the unforced system is not
sufficient for the BIBO-stability of system (6.8). The property of the unforced system
we are looking for, should be therefore intermediate between simple stability and
asymptotic stability.

Lemma 6.3 Assume that the system (6.8) is BIBO-stable, and let u(·) : [0, +∞] →
Rm be a piecewise continuous function. If there exists a number t¯ > 0 such that
u(t) = 0 for every t ≥ t¯, then

lim x(t, 0, u(·)) = 0 . (6.9)

t→+∞

Proof Let ū be a constant such that ||u(t)|| ≤ ū for t ∈ [0, t¯]. If t > t¯, we have
 t¯  t
x(t, 0, u(·)) = W (t − τ )u(τ ) dτ = W (σ)u(t − σ) dσ
0 t−t¯

which implies  t
||x(t, 0, u(·))|| ≤ ū ||W (σ)|| dσ . (6.10)
t−t¯

∞
Since the system is BIBO-stable, the integral 0 ||W (σ)|| dσ converges (Lemma
6.2). Hence, the integral in (6.10) can be rendered arbitrarily small for sufficiently
large t. 

We are now ready to prove a partial converse of Theorem 6.2 for systems of the
form (6.8).

Proposition 6.3 Assume that system (6.8) is completely controllable. If in addition

the system is BIBO-stable, then the matrix A is Hurwitz.

Proof We will show that the origin is globally attractive for the unforced system
(6.2). In other words, we will show that for each x0 ∈ Rn ,

lim x(t, x0 , 0) = 0 .
t→+∞

Since (6.8) is completely controllable, for any fixed instant T0 > 0 there exists
an admissible input map u 0 : [0, T0 ) → Rm such that x(T0 , 0, u 0 (·)) = x0 . Let us
define
104 6 External Stability

u 0 (t) if t ∈ [0, T0 ]
u(t) =
0 if t > T0 .

According to Lemma 6.3, we must have

lim x(t, 0, u(·)) = 0 .

t→+∞

Since x(t, 0, u(·)) ≡ x(t, 0, u 0 (·)) for t ∈ [0, T0 ], we have x(T0 , 0, u(·)) = x0 .
Moreover, using time invariance,

x(t, 0, u(·)) = x(t − T0 , x0 , 0) for t > T0 .

The conclusion is straightforward. 

So far, we have established that under the complete controllability assumption,
and continuing to assume C = I , internal stability is necessary and sufficient for
external stability. However, a system may be internally stable even if not completely
controllable (this happens for instance with the scalar system ẋ = −x + bu, if b = 0).
A more precise necessary and sufficient condition of external stability for systems of
the form (6.8) can be achieved by weakening the complete controllability assumption.
Recall that, apart from a linear change of coordinates, we can rewrite system (6.8)
under the controllability canonical form

ż 1 = A11 z 1 + A12 z 2 + B1 u
(6.11)
ż 2 = A22 z 2

where z 1 ∈ Rq e z 2 ∈ Rn−q (the notation is that of Chap. 5) and the pair of matrices
A11 , B1 defines a completely controllable system in Rq .
Theorem 6.3 System (6.8) is BIBO-stable if and only if both the following conditions
hold:
(i) all the eigenvalues of the matrix A11 of its controllable part have negative real
part;
(ii) all the eigenvalues of the matrix A22 of its uncontrollable part have nonpositive
real part and, for each possible eigenvalue of A22 with zero real part, the
algebraic and geometric multiplicity coincide.
Proof If system (6.11) is BIBO-stable, the same is true for the subsystem

ż 1 = A11 z 1 + B1 u . (6.12)

To prove statement (i), it is sufficient to remark that since (6.12) is completely

controllable, Proposition 6.3 applies. On the other hand, the eigenvalues of A22 form
a subset of the set of the eigenvalues of A. Therefore, statement (ii) follows from
Proposition 6.2.
6.3 The Case C = I 105

Next we prove the converse. Let us fix a positive number R. Consider a point
z̄ ∈ Rn and an input map u(t) such that

z̄ < R, u(·)∞ < R .

Denote by z(t) = (z 1 (t), z 2 (t)) the solution of the system, written in the form
(6.11), corresponding to the initial condition z̄ = (z̄ 1 , z̄ 2 ) and the input u(t). Notice
that
z(t) ≤ z 1 (t) + z 2 (t)

and
z̄ 1  ≤ z̄, z̄ 2  ≤ z̄

where, depending on the vectors, the norms are taken in Rn , in Rq or in Rn−q . The
assumption about A22 implies the existence of a constant γ > 0 such that

z 2 (t) ≤ γz̄ 2  < γ R .

Let v(t) = A12 z 2 (t) + B1 u(t). We have

v(t) ≤ A12 γ R + B1 R .

Let now R̃ = max{R, A12 γ R + B1 R}. Since A11 is Hurwitz, the subsystem

ż 1 = A11 z 1 + v

is BIBO-stable. Hence, there exists a constant S̃ such that if

v(t) < R̃, z̄ 1  < R̃ (6.13)

then z 1 (t) < S̃. Note that in our case, conditions (6.13) are valid by construction.
Hence,
z(t) ≤ z 1 (t) + z 2 (t) ≤ S̃ + γ R

and the inequality (6.3) is recovered taking S = S̃ + γ R. 

6.4 The General Case

Now we come back to the general case (6.1). First, we establish a generalized version
of Proposition 6.2.
106 6 External Stability

Proposition 6.4 Assume that system (6.1) is completely observable and, in addition,
BIBO-stable. Then, the origin is a stable (in general, not asymptotically) equilibrium
point for the unforced system (6.2).

Proof By the external stability assumption, for each R > 0 there exists S > 0 such
that if ||x0 || < R and u(t) = 0 for each t ≥ 0, then we have

||Cet A x0 || < S

for t ≥ 0. We distinguish several cases. First, we assume that A possesses a real

eigenvalue λ > 0, and let v
= 0 be an eigenvector of λ such that ||v|| < R. We have

||Cet A v|| = ||C(eλt v)|| = eλt ||Cv|| < S

for t ≥ 0, which is possible only if Cv = 0. The properties of the exponential matrix

can be invoked to infer that
d
Cet A v = C Aet A v .
dt
On the other hand, we also have

d d
Cet A v = eλt Cv = λeλt Cv .
dt dt

From Cv = 0 we therefore obtain C Aet A v = 0 for t ≥ 0, and in particular, for

t = 0, C Av = 0. This procedure can be iterated, until we arrive to conclude that:
⎛ ⎞
C
⎜ CA ⎟
⎜ ⎟
⎝ ... ⎠v = 0.
C An−1

This implies in turn that the rank of the matrix (C t |At C t | . . . |(At )n−1 C t ) is strictly
less than n, which is impossible since the system is completely observable.
Second, assume that A has an eigenvalue λ = 0, whose geometric multiplicity is
strictly less than the algebraic multiplicity. Then there exist an eigenvector v0
= 0
and a generalized eigenvector v1
= 0 such that

et A v1 = tv0 + v1 .

It is not restrictive to assume ||v1 || < R, so that ||Cet A v1 || < S per t ≥ 0. But

||Cet A v1 || = ||tCv0 + Cv1 || ≥ t||Cv0 || − ||Cv1 ||

which is bounded for t ≥ 0 only if Cv0 = 0. Of course,

6.4 The General Case 107

d tA d
e v1 = (tv0 + v1 ) = v0
dt dt
that is
d
Cet A v1 = Cv0 = 0 .
dt
On the other hand, we also have

d
Cet A v1 = C Aet A v1 = C A(tv0 + v1 ) .
dt
This last expression vanishes only if C Av0 = 0. The remaining part of the proof
can be carried out as in the previous case.
In analogous way, we exclude the cases where A possesses complex eigenvalues
with positive real part, or imaginary eigenvalues with algebraic multiplicity strictly
greater than their geometric multiplicity. 

Lemma 6.4 Assume that system (6.1) is completely observable and, in addition,
BIBO-stable. Then system (6.8), where the matrices A and B are the same as those
of system (6.1), is BIBO-stable.

Proof Let us start by a preliminary remark, which will be used later. Let E(t) be
the matrix defined in Sect. 5.2.3. By assumption, E(t) is positive definite for each
t ≥ 0. Let M = min||x||=1 x t E(1)x. This minimum exists and it is strictly positive.
Moreover, for each x ∈ Rn ,
 
xt x
x E(1)x =
t
E(1) ||x||2 ≥ M||x||2 .
||x||2 ||x||2

The proof of the Lemma is by contradiction. Assume that there exists R > 0
such that for each S > 0 we can find an initial state x0 , an admissible input u(·) :
[0, +∞] → Rm , and an instant τ > 0 such that

S
||x0 || < R, ||u(·)||∞ < R, but ||x(τ )|| > .
M

Consider the output y(t) of system (6.1), corresponding to the same initial state
x0 and the admissible input

u(t) for t ≤ τ
ũ(t) =
0 for t > τ .

Of course, ||ũ(·)||∞ < R. For t ≥ τ we have

y(t) = C x(t) = Ce(t−τ )A x(τ ) .

108 6 External Stability

Taking the norms, and integrating over the interval [τ , τ + 1], we have
 τ +1  τ +1
||y(t)|| dt = ||Ce(t−τ )A x(τ )|| dt
τ τ
 τ +1
x t (τ )e(t−τ )A C t Ce(t−τ )A x(τ ) dt
t
=
τ
 1
eσ A C t Ceσ A dσx(τ ) = x t (τ )E(1)x(τ ) > 0 .
t
= x t (τ )
0

 τ +1
On the other hand, τ ||y(t)|| dt ≤ max[τ ,τ +1] ||y(t)|| = ||y(τ̃ )|| for some τ̃ ≥
τ . Comparing these two conclusions, we get
 τ +1
||y(τ̃ )|| ≥ ||y(t)|| dt ≥ x t E(1)x ≥ M||x||2 > S .
τ

This means that (6.1) is not externally stable, which contradicts one of the hypothe-
ses. 

Thanks to the previous Lemma, we can state a partial converse of Theorem 6.2
for the systems of the form (6.1); this is actually a generalization of Proposition 6.3.

Proposition 6.5 Assume that system (6.1) is completely controllable and completely
observable. Assume also that the system is BIBO-stable. Then, the matrix A is Hur-
witz.

Proof According to Lemma 6.4, system (6.8) is externally stable. If system (6.1) is
also completely controllable, then system (6.8) is completely controllable, as well.
The conclusion is easily established as a consequence of Proposition 6.3. 

We have therefore realized that under the hypotheses of complete controllability

and complete observability, a system of the form (6.1) is externally stable if and only
if it is internally stable.
By a last effort, we can finally relax any restrictive assumption and formulate a
general necessary and sufficient condition. Referring again to the Kalman decompo-
sition and the related notation studied in Sect. 5.3.5, we consider the system

ż 2 = A22 z 2 + A24 z 4 + B2 u
(6.14)
ż 4 = A44 z 4

with the observation function y = C2 z 2 + C4 z 4 . By construction, (6.14) is com-

pletely observable.

Lemma 6.5 System (6.1) is externally stable if and only if system (6.14) is externally
stable.
6.4 The General Case 109

Proof First, we assume that system (6.14) possesses the BIBO-stability property. By
Proposition 6.1, we may assume without loss of generality that (6.1) is in the Kalman
canonical form
⎧
⎪
⎪ ż 1 = A11 z 1 + A12 z 2 + A13 z 3 + A14 z 4 + B1 u
⎪
⎪
⎪
⎪
⎨ż 2 = A22 z 2 + A24 z 4 + B2 u
ż 3 = A33 z 3 + A34 z 4
⎪
⎪
⎪
⎪ ż 4 = A44 z 4
⎪
⎪
⎩y = C z + C z .
2 2 4 4

Let R > 0. Focusing on system (6.1), we chose an initial state z̄ = (z̄ 1 , z̄ 2 , z̄ 3 , z̄ 4 )

and an admissible input u(·) such that

||z̄|| < R and ||u(·)||∞ < R .

The components of the vector (z̄ 2 , z̄ 4 ) are a subset of the components of z̄; hence
we have
||(z̄ 2 , z̄ 4 )|| < R .

Then by hypothesis, there exists S such that ||y(t)|| < S for each t ≥ 0. To con-
clude, it is sufficient to remark that (6.14) and (6.1) have the same output.
Vice versa, assume that (6.1) possesses the BIBO-stability property. Let us intro-
duce R and consider the system (6.14). Chose an initial state (z̄ 2 , z̄ 4 ) and an admis-
sible input u(·) such that

||(z̄ 2 , z̄ 4 )|| < R and ||u(·)||∞ < R .

Apply the same input to the system (6.1), choosing the initial condition z̄ =
(0, z̄ 2 , 0, z̄ 4 ). Clearly ||z̄|| = ||(z̄ 2 , z̄ 4 )||. Hence, there exists S such that ||y(t)|| < S
for each t ≥ 0. Again, we get the conclusion noticing that the systems have the same
output. 

Lemma 6.6 System (6.14) is BIBO-stable if and only if

(i) all the eigenvalues of A22 have strictly negative real part;
(ii) all the eigenvalues of A44 have nonpositive real part, and for each possible
eigenvalue with zero real part, the algebraic and geometric multiplicities coin-
cide.

Proof Assume that the conditions (i) and (ii) hold. Matrix A44 defines a unforced
system for which the origin is stable (in general, not asymptotically). In particular, the
z 4 component of (6.14) remains bounded for t ≥ 0. Let us consider v = A24 z 4 + B2 u
as a new input. If u(·) is bounded, v(·) is bounded, as well. Since the subsystem

ż 2 = A22 z 2 + v (6.15)
110 6 External Stability

is completely controllable, also z 2 remains bounded for t ≥ 0. The conclusion is

readily obtained.
Vice versa, assume that the system is BIBO-stable. Since it is completely observ-
able, by Proposition 6.4, the matrix
 
A22 A24
0 A44

defines an unforced system for which the origin is stable (in general, not asymptot-
ically). This implies in particular the validity of (ii). Condition (i) follows applying
Proposition 6.5 to the completely controllable and completely observable subsystem

ż 2 = A22 z 2 + B1 u
(6.16)
y = C2 z 2

initialized to (z̄ 2 , 0). 

The desired necessary and sufficient condition is finally obtained by combining

Lemmas 6.5 and 6.6.

Theorem 6.4 A system of the form (6.1) is BIBO-stable if and only if the following
two conditions hold:
(i) all the eigenvalues of the matrix A22 of the controllable and observable part of
the system have strictly negative real part;
(ii) all the eigenvalues of the matrix A44 of the observable but not controllable part
of the system have nonpositive real part, and for the possible eigenvalues of
A44 with zero real part the algebraic and geometric multiplicities coincide.

Chapter Summary
In this chapter we address the study of the external stability. External stability is a
natural consequence of the internal stability properties studied in Chap. 3. However,
there are systems which are externally stable but not internally stable. Necessary
and sufficient conditions for external stability are proven, in terms of the canonical
decomposition introduced in Chap. 5.
Chapter 7
Stabilization

As already pointed out in Chap. 1, the behavior of a system can be regulated, without
need of radical changes in its internal plant, by the construction of a suitable device
which interacts with the system by means of a feedback connection. The action of
such a device may have a static nature (and, in this case, it can be mathematically
represented as a function) or a dynamic one (and so being interpreted as an auxiliary
system). The feedback connection allows us to exert the control action in an auto-
matic way (i.e., without need of the presence of a human operator), and requires the
installation of sensors and actuators.
When all the state variables can be monitored and measured at each time, and all
the information about their evolution can be used by the control device, we speak
about state feedback. On the contrary, when the information about the state is only
partially available (since they are, for instance, obtained by means of an observation
function) we speak about output feedback.

7.1 Static State Feedback

In the static state feedback stabilization problem, the observation function is not
involved. Hence, in this section we can limit ourselves to systems of the form

ẋ = Ax + Bu , x ∈ Rn , u ∈ Rm . (7.1)

First of all, we try to understand what happens when the feedback connection is
implemented. Let v(t) be an external signal, injected into the system through the input
channel u, and let x(t) be the solution representing the resulting state evolution. The
feedback map k(x) : Rn → Rm generates another signal w(t) = k(x(t)). The signal
actually received by the system is the sum of v(t) and w(t), that is u = w(t) + v(t).

© Springer Nature Switzerland AG 2019 111

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_7
112 7 Stabilization

A static state feedback is said to be linear if k(x) = F x, F being a m × n matrix.

The implementation of a linear feedback can be mathematically interpreted as a
substitution
u = Fx + v (7.2)

which gives rise to a transformation of the system. Indeed, replacing (7.2) in (7.1),
we obtain
ẋ = (A + B F)x + Bv , (7.3)

the so-called closed loop system.

We stress that by virtue of the particular structure of the control (7.2), the trans-
formed system is still of the form (7.1), with the matrix B unchanged, and the matrix
A replaced by the new matrix à = A + B F. We also notice that the transformation
induced by (7.2) is invertible; indeed, if we apply the feedback law v = −F x + u to
(7.3) we recover the form (7.1) of the system. Thus, the transformation (7.2) defines
an equivalence relation on the set of all the systems of the form (7.1); this fact can
be formalized by the following definition.
Definition 7.1 We say that systems (7.1) and

ẋ = Ãx + Bu

are feedback equivalent if there exists a matrix F such that à = A + B F.

In this perspective, we can formulate the following problem pattern: assume that
we are interested in a certain property, and that this property is not satisfied by the
given system (7.1). We wonder whether the property is satisfied by system (7.3), for
a suitable choice of the matrix F. More precisely, we want to find conditions under
which the qualitative behavior of the given system can be modified in the desired
way by means of a convenient feedback connection.

7.1.1 Controllability

As a first example, we ask whether a system can achieve the complete controllability
property by means of a feedback transformation (diversely stated, whether in the
same feedback equivalence class there may exist systems whose reachable spaces
have different dimensions). The answer is negative; indeed, the following theorem
holds.
7.1 Static State Feedback 113

Theorem 7.1 For each matrix F, the reachable spaces of the systems (7.1) and
(7.3), denoted here respectively by R(7.1) and R(7.3) , coincide. As a consequence, we
have

rank (B|AB| . . . |An−1 B) = rank (B|(A + B F)B| . . . |(A + B F)n−1 B) .

Proof According to Theorem 5.3

R(7.1) = span {b1 , . . . , bm ; Ab1 , . . . , Abm ; . . . ; An−1 b1 , . . . , An−1 bm } .

Thus, v ∈ R(7.1) if and only if v is a linear combination of the vectors

b1 , . . . , bm ; Ab1 , . . . , Abm ; . . . ; An−1 b1 , . . . , An−1 bm .

Analogously,

R(7.3) = span {b1 , . . . , bm ; (A + B F)b1 , . . . , (A + B F)bm ;

. . . ; (A + B F)n−1 b1 , . . . , (A + B F)n−1 bm } .

Notice that (A + B F)b j = Ab j + B Fb j . The vector Ab j belongs to R(7.1) and

the vector
B Fb j = (b1 | . . . |bm )Fb j

belongs to R(7.1) as well, since it is a linear combination of b1 , . . . , bm . Continuing

in this way, we notice that

(A + B F)2 b j = (A + B F)(A + B F)b j = A2 b j + AB Fb j + B F Ab j + B F B Fb j .

The first term is in R(7.1) by construction; the second because it is a linear com-
bination of Ab1 , . . . , Abm ; the third and the fourth term because they are linear
combination of b1 , . . . , bm . The same reasoning applies to each term of the form
(A + B F)k b j .
In conclusion, R(7.3) ⊆ R(7.1) , since all the vectors of R(7.3) are linear combinations
of vectors of R(7.1) .
The opposite inclusion can be achieved by exchanging the roles of the systems
(recall that (7.1) can be recovered from (7.3) by the inverse feedback transformation
v = −F x + u). 

In other words, Theorem 7.1 states that the complete controllability property is
invariant under feedback equivalence.
114 7 Stabilization

7.1.2 Stability

In the previous chapter we tried to characterize those systems of the form (7.1) which
enjoy the external stability property. We noticed that this property is intimately linked
to the internal stability properties of the system (Hurwitz property). This motivates
the effort to elaborate models for which the eigenvalues of the system matrix A lie
in the open left half of the complex plane and, in case this condition is not fulfilled,
the interest in devising appropriate corrections.
The main purpose of this chapter is to show that feedback connections represent
a convenient tool in order to improve the internal stability properties of a system.

7.1.3 Systems with Scalar Input

Consider first the case of a system with scalar input (i.e., with m = 1 and B reduced
to a column vector b). Our approach is based on the following theorem.
Theorem 7.2 Assume that m = 1, and that system (7.1) is completely controllable.
Then, there exists a change of coordinates x = Pζ for which the system takes the
form
ζ̇ = A0 ζ + ub0 (7.4)

where A0 is the companion matrix

⎛ ⎞ ⎛ ⎞
0 1 0 ... 0 0
⎜ 0 ⎜0⎟
⎜ 0 1 ... 0 ⎟⎟ ⎜ ⎟
⎜ ⎟
⎜ ... ...
⎜ ... ... ... ⎟
⎟ and b0 = ⎜ ... ⎟ .
⎝ 0 ⎜ ⎟
0 0 ... 1 ⎠ ⎝0⎠
−an −an−1 ... ... −a1 1

Proof In Sects. 2.10 and 2.11 we saw that if there exists a cyclic vector for A (that
is a vector v such that
v, Av, . . . , An−1 v

are linearly independent), then A is similar to the matrix

⎛ ⎞
0 0 ... 0 −an
⎜ 1 0 ... 0 −an−1 ⎟
⎜ ⎟
⎜ 0 1 ... 0 −an−2 ⎟
⎜ ⎟
⎝... ... ... ... ... ⎠
0 ... ... 1 −a1
7.1 Static State Feedback 115

where the numbers a1 , . . . , an are the coefficients of the characteristic polynomial of

A, apart for a possible change of sign. Such a matrix is the transpose of the companion
form.
The matrix associated to the change of coordinates is formed by the columns
v, Av, . . . , An−1 v. The complete controllability hypothesis states that the rank of
the matrix
R = (b|Ab| . . . |An−1 b)

is n. Hence, b is cyclic for A and R −1 A R = A0 t . Moreover,

⎛ ⎞ ⎛ ⎞
1 1
⎜0⎟ ⎜0⎟
⎜ ⎟ ⎜ ⎟
R⎜.⎟=b that is R −1 b = ⎜ . ⎟ .
⎝ .. ⎠ ⎝ .. ⎠
0 0

We have so proved that our system is linearly equivalent to

⎛ ⎞
1
⎜0⎟
⎜ ⎟
ẇ = At0 w + u ⎜ . ⎟ . (7.5)
⎝ .. ⎠
0

On the other hand, the matrix Q = b0 |A0 b0 | . . . |An−1
0 b0 has the form

⎛ ⎞
0 ... ... 0 1
⎜ 0 ... 0 1 ∗ ⎟
⎜ ⎟
⎜... ... ... ... ...⎟
⎜ ⎟
⎝ 0 1 ∗ ... ∗ ⎠
1 ∗ ... ... ∗

and so it is nonsingular (the stars stand for some numbers whose explicit expression
is unessential). By the same arguments as before, we must have Q −1 A0 Q = At0 as
well. Moreover, ⎛ ⎞ ⎛ ⎞
1 1
⎜0⎟ ⎜0⎟
⎜ ⎟ ⎜ ⎟
Q ⎜ . ⎟ = b0 that is Q −1 b0 = ⎜ . ⎟ .
⎝ .. ⎠ ⎝ .. ⎠
0 0

Thus, the system (7.4) is linearly equivalent to (7.5), as well. Finally, (7.4) and
the given system, being both linearly equivalent to (7.5), are equivalent each other.

Recall that the companion form characterizes the system representation of scalar
linear differential equations. Theorem 7.2 states therefore that any completely
116 7 Stabilization

controllable linear system with single input and state space dimension n is linearly
equivalent to a system represented by a single linear differential equation of order n.
We emphasize that the proof of Theorem 7.2 supplies an explicit expression for the
matrix P which determines the similarity between A and its companion form A0 .
Indeed, it is immediately seen that P = R Q −1 .
We rewrite for convenience system (7.4) as
⎧
⎪ζ̇1 = ζ2
⎪
⎪
⎪
⎪ ..
⎪
⎨ .
ζ̇n−1 = ζn (7.6)
⎪
⎪
⎪
⎪
⎪
⎪
⎩
ζ̇n = −an ζ1 − · · · − a1 ζn + u .

Definition 7.2 A 2k-tuple {λ1 , . . . , λk , μ1 , . . . , μk } is said to be consistent if:

(1) 1 ≤ k ≤ n;
(2) λ1 , . . . , λk are distinct complex numbers;
(3) μ1 , . . . , μk are (not necessarily distinct) positive integers such that μ1 + · · · +
μk = n;
(4) for each i ∈ {1, . . . , k} there exists j ∈ {1, . . . , k} such that λ j = λi (the
conjugate of λi ) and μi = μ j .

Given any consistent 2k-tuple, it is easy to construct a monic polynomial with

real coefficients
λn + b1 λn−1 + · · · + bn

whose roots are exactly λ1 , . . . , λk , with respective multiplicities μ1 , . . . , μk . Now,

let us apply to system (7.6) the feedback

u = (−bn + an )ζ1 + · · · + (−b1 + a1 )ζn + v . (7.7)

The resulting closed-loop system is

⎧
⎪
⎨ζ̇1 = ζ2
⎪
..
⎪ .
⎪
⎩ζ̇ = −b ζ − · · · − b ζ + v .
n n 1 1 n

Setting finally v = 0, we obtain an unforced linear system whose characteristic

equation has exactly the roots λ1 , . . . , λk . Of course, the roots of the characteristic
equation coincide with the eigenvalues of the matrix. If the numbers λi have been
chosen in such a way that Reλi < 0 for each i, we have obtained, by means of the
feedback (7.7), a system for which the origin is asymptotically stable.
7.1 Static State Feedback 117

If fact, we have proven something more. For any preassigned real n × n matrix
M, a completely controllable system with scalar input can be always transformed in
a new system, such that the eigenvalues of the matrix of the new system coincide
exactly with those of M.

7.1.3.1 System with Multiple Inputs

The discussion of the previous section motivates the following general definitions.
Definition 7.3 We say that (7.1) is stabilizable if there exists a static state feedback
u = F x such that all the eigenvalues of the matrix (A + B F) have negative real part.
We say that (7.1) is superstabilizable if for each α > 0 there exists a static state
feedback u = F x (with F dependent on α) such that the real part of each eigenvalue
of the matrix (A + B F) is less than −α.
We say that (7.1) has the pole assignment property if for each given consistent 2k-
tuple there exists a static state feedback u = F x such that the eigenvalues of A + B F
are exactly the numbers λ1 , . . . , λk , with respective multiplicities μ1 , . . . , μk .
Systems which are superstabilizable are particularly interesting for applications.
Indeed for these systems, it is not only possible to construct stabilizing feedback
laws, but also to assign an arbitrary decay rate.
We already know that any completely controllable system with a scalar input
possesses the pole assignment property, and hence it is stabilizable and superstabi-
lizable. This result can be extended, with some technical complications in the proof,
to systems with multiple input.
Theorem 7.3 For any system of the form (7.1), the following properties are equiv-
alent:
(i) complete controllability
(ii) pole assignment
(iii) superstabilizability.
The reader interested in the full proof of Theorem 7.3 is referred, for instance,
to [11], p. 145 or [28], p. 58. It follows in particular from Theorem 7.3 that for any
system in the general form (7.1), complete controllability implies stabilizability by
static state feedback. We give below an independent and direct proof of this fact.
Proposition 7.1 If (7.1) is completely controllable, then it is stabilizable.

Proof The completely controllability assumption amounts to say that for each T > 0
the matrix
 T
e−τ A B B t e−τ A dτ
t
(T ) =
0
118 7 Stabilization

is positive definite (Theorem 5.2). Write for simplicity  = (1). Let us show that
the feedback law u = F x = −B t  −1 x actually stabilizes the system. Compute the
derivative
d  −t A 
e B B t e−t A = −Ae−t A B B t e−t A − e−t A B B t e−t A At
t t t

dt
which yields
 
1
d  −t A  1 
e B B t e−t A dt = − Ae−t A B B t e−t A
t t
(7.8)
dt
0 0

− e−t A B B t e−t A At dt .
t

Clearly, the expression on the left hand side can be rewritten as

1
−t A

t −t At  −A t −At
e BB e  = e BB e − B Bt .
0

By the definition of , the right hand side of (7.8) is equal to

 1  1
e−t A B B t e−t A dt − e−t A B B t e−t A dt At = −A −  At .
t t
−A
0 0

Hence,
e−A B B t e−A − B B t = −A −  At .
t
(7.9)

On the other hand,  being a symmetric matrix,

(A − B B t  −1 ) + (A − B B t  −1 )t = A +  At − 2B B t . (7.10)

From (7.9) and (7.10) we infer

(A − B B t  −1 ) + (A − B B t  −1 )t = −e−A B B t e−A − B B t .

t
(7.11)

The matrix at the right hand side is (at least) negative semidefinite. According to
Theorem 3.4, we can conclude that the origin is stable for the system

ẋ = (A − B B t  −1 )t x (7.12)

and so also for the system

ẋ = (A − B B t  −1 )x (7.13)

since any square matrix has the same eigenvalues as its transpose. However, on the
base of (7.11), we are not able to conclude that (7.13) is asymptotically stable: there
7.1 Static State Feedback 119

are indeed simple examples of completely controllable linear systems for which the
matrix at the right hand side of (7.11) is actually not positive definite.1 In other words,
we cannot be sure that V (x) = x t x is a strict Lyapunov function for (7.12).
To finish the proof, we need therefore to try another way. We will resort directly
to Theorem 3.1. More precisely, we will show that all the eigenvalues of (A −
B B t  −1 )t have strictly negative real part. To this end, we take advantage of the
previous computations.
Let λ be an eigenvalue (real or complex) of (A − B B t  −1 )t , and let v = 0 be a
corresponding eigenvector. We have

(A − B B t  −1 )t v = λv . (7.14)

From (7.11) we obtain

v̄ t [(A − B B t  −1 ) + (A − B B t  −1 )t ]v = −v̄ t [e−A B B t e−A + B B t ]v . (7.15)

t

On the other hand,

v̄ t [(A − B B t  −1 ) + (A − B B t  −1 )t ]v (7.16)

= λ̄v̄ t v + λv̄ t v = (λ̄ + λ)v̄ t v = 2 Re λ(v̄ t v) .

Hence,
v̄ t [e−A B B t e−A + B B t ]v = −2 Re λ(v̄ t v) .
t
(7.17)

Since  is real and positive definite, we can easily check that v̄ t v > 0 (notice the
analogies between this argument and the computation in the proof of Theorem 3.3).
We recover in this way the previous conclusion that Re λ ≤ 0. Now, if it happens
that Re λ = 0 for some λ, then we should also have that

v̄ t [e−A B B t e−A + B B t ]v = 0
t

and so in particular
v̄ t B B t v = v̄ t B̄ B t v = 0 .

This implies that B t v = 0. It follows

(A − B B t  −1 )t v = At v − ( −1 )t B B t v = At v . (7.18)

Comparing (7.14) and (7.18), we conclude that v is also an eigenvector of At

corresponding to the same eigenvalue λ. But then e−t A v = e−λt v. Finally we get
t

   
0 −1 0
1 One such example can be obtained taking A = ,b= .
0 0 1
120 7 Stabilization
 1  1
t −t A t −t At
v̄ v =
t
v̄ e BB e v dt = e−λt v̄ t e−t A B B t v dt = 0
0 0

which is impossible since  is positive definite and v = 0. 

7.1.4 Stabilizability

The stabilizability property is actually weaker than complete controllability; this can
be easily realized looking at a system for which A is Hurwitz and B = 0. In this
section we aim to characterize the stabilizability property by means of suitable and
easy-to-check conditions.
To this end, it is convenient to apply a preliminary change of coordinates in order
to put the system in the controllability canonical form. In other words, without loss
of generality, we can assume for our system the form

ż 1 = A11 z 1 + A12 z 2 + B1 u
(7.19)
ż 2 = A22 z 2

where the subsystem corresponding to the block of coordinates z 1 is completely

controllable.
Theorem 7.4 System (7.19) is stabilizable if and only if the matrix A22 is Hurwitz.
 
A11 A12
Proof The set of the eigenvalues of a matrix A = is the union of the sets
0 A22
of the eigenvalues of the matrices A11 and A22 . By means of a feedback, there is no
way to modify the eigenvalues of the matrix A22 . Hence, if the system is stabilizable,
A22 must be Hurwitz.
Vice versa, since the subsystem corresponding to the components z 1 is completely
controllable, we can construct a feedback u = F1 z 1 (for instance, by the method
illustrated in the proof of Proposition 7.1) in such a way that matrix A11 + B1 F1 is
Hurwitz. The matrix A22 is Hurwitz by hypothesis. Hence, the matrix A is Hurwitz,
as well. 

Next we present (without proof) other necessary and sufficient conditions for
stabilization.
Theorem 7.5 Let V be the subspace of Cn generated by all the eigenvectors (includ-
ing the generalized ones) associated to all the eigenvalues λ of A, having nonnegative
real part. Moreover, let U be the subspace of Rn generated by all the vectors of the
form Rev and im v, with v ∈ V . System (7.1) is stabilizable if and only if U is
contained in its reachable space R.
7.1 Static State Feedback 121

Theorem 7.6 System (7.1) is stabilizable if and only if

rank (A − λI | B) = n

for every complex number λ with nonnegative real part.

It is interesting to compare Theorem 7.6 and Hautus’ controllability criterion

(Theorem 5.4).
Theorem 7.7 System (7.1) is stabilizable if and only if there exists a symmetric,
positive definite matrix P such that

At P + P A − P B B t P = −I . (7.20)

In this case, a stabilizing feedback can be found of the form u = F x, with

F = −αB t P and α ≥ 21 .
It is not difficult to see that (7.20) is sufficient for stabilizability. Indeed, replacing
the feedback F = −αB t P into the equation we obtain the closed-loop system in the
form
ẋ = Ax − αB B t P x . (7.21)

Taking into account (7.20), we have

(A − αB B t P)t P + P(A − αB B t P) = At P + P A − 2αP B B t P

= −I + (1 − 2α)P B B t P .

The matrix P B B t P = (B t P)t B t P corresponds to a positive semidefinite

quadratic form. Hence, if α ≥ 21 , P solves a Lyapunov matrix equation for sys-
tem (7.21). The conclusion follows by Corollary 3.1.
On the contrary, proving that the same condition is also necessary for stabilizability
is more difficult (a proof can be found in [11], p. 133).
Equation (7.20) is called the algebraic Riccati matrix equation associated to sys-
tem (7.1). We emphasize that (7.20) is nonlinear with respect to the entries of the
unknown matrix P. We emphasize also that, once a solution of (7.20) has been found,
the feedback law provided by Theorem 7.7 is explicit and simpler than the feedback
provided in the proof of Proposition 7.1.
Corollary 7.1 If there exists a symmetric, positive definite matrix Q such that the
matrix equation
At P + P A − P B B t P = −Q (7.22)

admits a symmetric, positive definite solution P, then system (7.1) is stabilizable.

On the other hand, if system (7.1) is stabilizable, then for each symmetric, positive
definite matrix Q there exists a symmetric, positive definite solution P of the matrix
equation (7.22).
122 7 Stabilization

Proof The proof of the first statement is similar to the proof of the sufficient part
of Theorem 7.7. In order to prove the second statement, we write Q = R t R, with
R nonsingular and symmetric. If the feedback u = F x stabilizes the given system,
then the system
ẋ = Ãx + B̃u

where à = R A R −1 and B̃ = R B, is stabilized by the feedback u = F R −1 . Hence,

according to the necessary part of Theorem 7.7, there must exists a symmetric,
positive definite matrix P̃ such that

Ãt P̃ + P̃ Ã − P̃ B̃ B̃ t P̃ = −I .

The remaining part of the proof is similar to that of Corollary 3.1. 

7.1.5 Asymptotic Controllability

If system (7.1) is stabilizable by means of a feedback u = F x, then for each initial

state x0 ∈ Rn we can consider the solution x(t, x0 ) of the problem

ẋ = (A + B F)x
x(0) = x0 .

This solution x(t, x0 ) obviously coincides with the solution x(t) of the problem

ẋ = Ax + Bu x0 (t)
(7.23)
x(0) = x0

where u x0 (t) = F x(t, x0 ).

Definition 7.4 We say that the system (7.1) is asymptotically controllable if for each
x0 ∈ Rn there exists an input map u x0 (t) such that the corresponding solution x(t)
of the problem (7.23) approaches the origin for t → +∞.

The previous reasoning shows that a stabilizable system is asymptotically con-

trollable. But also the converse is true. Indeed, by an argument similar to that used in
the proof of Theorem 7.4, it is not difficult to see that if system (7.1) is asymptotically
controllable, then the uncontrollable part of the associated canonical controllability
form must be asymptotically stable. Then, the conclusion follows in force of The-
orem 7.4. We can therefore state a further necessary and sufficient condition for
stabilizability.

Theorem 7.8 System (7.1) is stabilizable if and only if it is asymptotically control-

lable.
7.2 Static Output Feedback 123

7.2 Static Output Feedback

In the previous section we studied the problem of stabilizability by means of a static

state feedback. Obviously, this way is not feasible when the whole state of the system
is not available. This usually happens when we deal with a system with an observation
function 
ẋ = Ax + Bu
(7.24)
y = Cx

(x ∈ Rn , u ∈ Rm , y ∈ R p ), and the requested feedback law has of the form u = K y.

The problem of stabilization by means of output feedback is more natural in view of
the applications, but also much more difficult to study.
To become familiar with these new difficulties, we examine some simple exam-
ples.
Example 7.1 The two dimensional system

ẋ1 = x2
(7.25)
ẋ2 = u

is completely controllable, and hence stabilizable by means of a feedback of the form

u = k1 x1 + k2 x2 . In fact, provided that the parameters k1 and k2 can chosen freely
and independently each other, the system is superstabilizable.
However, it is not possible to stabilize the system if the choice is limited to
feedback laws of the form u = kx1 . Indeed, by applying such a feedback, the system
becomes 
ẋ1 = x2
ẋ2 = kx1

whose eigenvalues are both either on the imaginary axis, or on the real axis and, in this
second case, they have opposite sign. In other words, since now only one parameter
can be arbitrarily chosen, we do not have the degrees of freedom necessary to solve
the problem.
As already suggested, the impossibility of implementing a feedback which uses
all the state variables typically arises when we have an observation function. In this
example, the feedback u = kx1 can be interpreted as an output feedback, if we
assume an observation function y = ct x with c = (1 0). We emphasize that the
system, with respect to this observation function, is completely observable, as well;
nevertheless, the system is not stabilizable by an output feedback. 

Example 7.2 Consider again the system (7.25), but this time with the observation
function y = x1 + x2 . By applying the feedback u = −ky = −k(x1 + x2 ), we obtain
the system
124 7 Stabilization

Fig. 7.1 Graph of 3

Re (λ2 (k))

−1

−2

−3
−1 0 1 2 3 4 5 6 7 8


ẋ1 = x2
ẋ2 = −kx1 − kx2

whose characteristic equation is λ2 + kλ + k = 0. Thus, the system is stabilized

if k > 0. However, the system is not superstabilizable. Indeed, if 0 < k < 4 the
characteristic roots are not real. Their real part is Re λ = −k/2; it decreases when
k increases, and
lim− Re λ = −2 .
k→4

If k = 4 we have λ1 = λ2 = −2. If k > 4, the characteristic roots are real and

can be represented as
√ √
−k − k 2 − 4k −k + k 2 − 4k
λ1 = and λ2 = .
2 2
Clearly λ2 > λ1 : moreover, λ2 increases when k increases, and

lim λ2 = −2 , lim λ2 = −1 .
k→4+ k→+∞

In conclusion, Re λ2 (k) ≥ −2 for each k ≥ 0, and it attains the minimum for

k = −2 (the graph of the real part of λ2 as a function of k is shown in Fig. 7.1).
Notice that again in this example, the system is completely controllable and com-
pletely observable. 

Of course, if a system is stabilizable by a static output feedback, it is stabilizable

also by a static state feedback. Hence, all the static state feedback stabilizability
7.2 Static Output Feedback 125

conditions listed in the previous section can be reviewed as necessary (but no more
sufficient) conditions for static output feedback stabilizability.

7.2.1 Reduction of Dimension

In this section we present a theorem which allows us to simplify, under particular

conditions, the study of the static output feedback stabilization problem.
Given a system (7.24), we begin by applying a linear change of coordinates, with
the purpose of rewriting the matrices A, B, C in Kalman’s canonical form (5.35),
that is ⎧
⎪
⎪ ż 1 = A11 z 1 + A12 z 2 + A13 z 3 + A14 z 4 + B1 u
⎪
⎪
⎪
⎪
⎨ż 2 = A22 z 2 + A24 z 4 + B2 u
ż 3 = A33 z 3 + A34 z 4 (7.26)
⎪
⎪
⎪
⎪ ż 4 = A44 z 4
⎪
⎪
⎩y = C z + C z .
2 2 4 4

Recall that according to the usual notation, the controllable part of the system is
identified by the indices 1 and 2, while the completely observable part is identified
by the indices 2 and 4.
Theorem 7.9 The overall system (7.24) is stabilizable by static output feedback if
and only if the following conditions are both satisfied:

(1) the matrices A11 , A33 and A44 have all the eigenvalues with negative real part;
(2) the completely controllable and completely observable part of the system, that
is the part corresponding to the subsystem

ż 2 = A22 z 2 + B2 u
(7.27)
y = C2 z 2

is stabilizable by static output feedback.

Proof We start by a preliminary remark. Let K be a matrix with p columns and m

rows. If we apply the feedback u = K y = K C2 z 2 + K C4 z 4 to system (7.26), the
system matrix becomes ⎛ ⎞
A11 Ã12 A13 Ã14
⎜ 0 Ã22 0 Ã24 ⎟
⎜ ⎟ (7.28)
⎝ 0 0 A33 A34 ⎠
0 0 0 A44

where Ã12 = A12 + B1 K C2 , Ã14 = A14 + B1 K C4 , Ã22 = A22 + B2 K C2 , Ã24 =

A24 + B2 K C4 .
126 7 Stabilization

By virtue of the triangular block form of (7.28), it is clear that the feedback
u = K y stabilizes the system if and only if the matrices A11 , Ã22 , A33 and A44 have
all their eigenvalues with negative real part. Taking into account of condition (1),
this actually happens if and only if the feedback u = K y = K C2 z 2 stabilizes the
reduced order system (7.27). 
In view of Theorem 7.9, as far as we are interested in the static output feedback
stabilization problem, it is not restrictive to assume that the system at hand is com-
pletely controllable as well as completely observable. Then, the following sufficient
condition may be of some help.
Proposition 7.2 Let the system (7.24) be given. Assume that it is completely con-
trollable, and that the matrix C is invertible. Then, the system is stabilizable by a
static output feedback.
Proof By the complete controllability hypothesis, there exists a matrix K such that
the system is stabilizable by a static state feedback u = K x. We can write u =
K C −1 C x = K C −1 y. We obtain in this way a static output feedback u = F y with
F = K C −1 whose effect on the system is the desired one. 
In the previous statement, the assumption that C is invertible implies of course
that p = n and that the system is completely observable, as well.
Example 7.3 Consider the system
⎧
⎪
⎨ẋ1 = x1 + 4x2
ẋ2 = 2x1 − 6x2 + u
⎪
⎩
y = 2x1 + x2 .

It is clear that it is completely controllable, but not completely observable. Using

the change of coordinates x = Pz where P is defined as
 
−1 2
P=
2 1

(according to the method explained in Sect. 5.3) we recover the observability canon-
ical form ⎧
⎪
⎨ż 1 = −7z 1 − 2z 2 + 5 u
2

ż 2 = 2z 2 + 5 u
1
⎪
⎩
y = 5z 2 .

The unobservable part (first equation), when we set u = 0, is asymptotically

stable. By Theorem 7.9, now it is clear that the system, in the new coordinates, is
stabilizable by means of a static output feedback u = ky. Convenient values for the
parameter k can be determined by direct computation: we find k < −2. Coming
back to the original coordinates, the feedback to be applied is u = ky = k(2x1 + x2 )
(again, k < −2). 
7.2 Static Output Feedback 127

As suggested by the previous example, once the reduction of dimension has been
performed, if the dimension of the completely controllable and observable part turns
out to be small, the existence of static output stabilizers can be checked by direct
computation. An other example is given below.
Example 7.4 Consider a two-dimensional, completely controllable and completely
observable system (in controllability canonical form)
⎧
⎪
⎨ẋ1 = x2
ẋ2 = −a2 x1 − a1 x2 + u (7.29)
⎪
⎩
y = c0 x1 + c1 x2 .

The complete observability assumption amount to say that c02 −c0 c1 a1 +c12 a2 = 0,
which in turn implies that c0 and c1 cannot be both zero. By the substitution u = ky,
the system matrix takes the form
 
0 1
−a2 + kc0 −a1 + kc1

whose characteristic equation is

λ2 − λ(−a1 + kc1 ) + a2 − kc0 = 0 .

By the Routh-Hurwitz criterion, it is easy to see that the system is stabilizable

with static output feedback if and only if there exists k ∈ R such that kc0 < a2 , and
kc1 < a1 . 
The following statement is a dual version of Proposition 7.2.
Proposition 7.3 Let the system (7.24) with m = n be given. Assume that it is com-
pletely observable, and that the matrix B is invertible. Then, the system is stabilizable
by a static output feedback.

Proof The observability assumption about (7.24) implies that the dual system

ẋ = At x + C t v
y = Bt x

is completely controllable. Hence, there is a matrix K such that At +C t K is Hurwitz.

The matrix A + K t C is Hurwitz, as well. Thus, if we apply to (7.24) the feedback
law u = B −1 K t y, then the matrix of the closed loop system is A + B B −1 K t C =
A + K t C. The statement is proven. 
128 7 Stabilization

7.2.2 Systems with Stable Zero Dynamics

Concerning systems represented by Eq. (7.26), there is a further interesting remark.

For a moment, let us limit ourselves to look at the differential part of the system,
operated in open loop. The solution corresponding to an initial state of the form z̄ =
(z̄ 1 , 0, z̄ 3 , 0) and a vanishing input u = 0 evolves inside the subspace of equations
z 2 = z 4 = 0, which is therefore invariant (in the sense of Definition A.4) for the
unforced system. We point out that each solution lying in this subspace gives rise to
a vanishing output. For this reason, the subspace of equations z 2 = z 4 = 0 is called
the space of the zero dynamics, and the system

ż 1 = A11 z 1 + A13 z 3
ż 3 = A33 z 3

is called the system of the zero dynamics. The following statement is a straightforward
consequence of Theorem 7.9 and Proposition 7.2.

Corollary 7.2 Let the dimension of the observable but not controllable part of the
system (7.24) be zero. Assume in addition that the matrix C2 is invertible. Then, the
system is stabilizable by static output feedback if and only if the origin is asymptoti-
cally stable for the system of the zero dynamics.

7.2.3 A Generalized Matrix Equation

Other sufficient conditions for static output stabilization can be obtained by suitable
generalizations of the Riccati matrix equation (7.20). Next we present one such
generalization.
Let C be a matrix with p rows and n columns. A generalized inverse (or pseu-
doinverse) of C is any matrix C † with n rows and p columns such that

CC † C = C and C † CC † = C † .

If in addition we require that CC † is symmetric, then C † is uniquely determined

and it is called the Moore-Penrose generalized inverse of C (see [4] for properties of
generalized inverse matrices). The space Rn can be decomposed as im (C t )⊕ker (C).
Moreover, the subspaces im (C t ) and ker (C) are orthogonal each other. The square
n × n matrix E im = C † C represents the orthogonal projection on im (C t ), while the
orthogonal projection on ker (C) is given by E ker = I − E im . The matrices E im and
E ker are uniquely determined and symmetric.

Theorem 7.10 Consider the system (7.24), and assume that there exist symmetric
and positive definite matrices P and Q such that

At P + P A − E im P B B t P E im + E ker P B B t P E ker = −Q . (7.30)

7.2 Static Output Feedback 129

Then, the system is stabilizable by the static output feedback u = K y, with

K = −B t PC † .
Before proving the theorem, we point out the following matrix identity:

E im P B B t P E im − E ker P B B t P E ker + P B B t P (7.31)

= E im P B B P + P B B P E im
t t

which can be easily checked taking into account that E ker = I − E im .

Proof of Theorem 7.10 Applying the static output feedback u = −B t PC † y, the
closed-loop system takes the form

ẋ = (A − B B t PC † C)x .

We show that, by virtue of (7.30) and (7.31), P solves the Lyapunov matrix
equation for this system. Indeed, we have:

(A − B B t PC † C)t P + P(A − B B t PC † C)
= At P + P A − [E im P B B t P + P B B t P E im ]
= At P + P A − E im P B B t P E im + E ker P B B t P E ker − P B B t P
= −Q − P B B t P .

The conclusion follows by Theorem 3.3, since the matrix P B B t P is clearly

positive semidefinite. 
Notice that (7.30) reduces to (7.20) when C = I .
Example 7.5 The condition of Theorem 7.10 works, for the case considered in the
previous Example 7.2. Since C = (1 1), as a generalized inverse we can take for
instance  
1/2
C† =
1/2

so that    
1/2 1/2 1/2 −1/2
E im = and E ker = .
1/2 1/2 −1/2 1/2

Writing  
p11 p12
P= (7.32)
p12 p22

the left-hand side of (7.30) takes the form

⎛ 2 ⎞
2
p12 + p22
⎜ − p12 p22 p11 − ⎟
⎝ 2 ⎠ .
2
p12 + p22
2
p11 − 2 p12 − p12 p22
2
130 7 Stabilization

A solution of (7.30) with the required properties is obtained taking p11 = 6, p22 =
3, p12 = 2. The corresponding output feedback is u = − 25 y. 

Note that assuming Q = I in (7.30) would be restrictive, contrary to what happens

in the case of the Lyapunov matrix equation (see Theorem 3.3 and Corollary 3.1)
and in the case of the Riccati matrix equation. For instance, it is not difficult to check
that in the previous example, there is no solutions for (7.30) if we set Q = I .

Example 7.6 The condition of Theorem 7.10 is not necessary for the existence of
an output feedback stabilizer. Consider the two-dimensional system with
   
1 1 0
A= , B= , C = (1 0) .
0 −2 1

Clearly, this time we have

   
10 00
E im = and E ker = .
00 01

Assume that (7.30) has a positive definite solution P, that we write again in the
form (7.32). Then it would be possible to construct an output stabilizer of the form
u = −B t PC † y, which reduces in this case to u = − p12 y. By direct substitution, it
is easily seen that the system can be actually stabilized by an output feedback of this
form, provided that p12 > 2.
On the other hand, the left-hand side of (7.30) takes now the form
 
2 p11 − p12
2
p11 − p12
.
p11 − p12 2 p12 − 4 p22 + p22
2

This matrix is definite negative only if the term 2 p12 − 4 p22 + p22
2
is negative: as
easily seen, this requires that p12 < 2.
In conclusion, the system is stabilizable by an output feedback, but the coefficient
of the feedback cannot be determined on the base of Theorem 7.10. 

7.2.4 A Necessary and Sufficient Condition

The static output feedback stabilization problem is sometimes referred to in the

literature as an unsolved problem. From a practical point of view, a numerical solution
of a nonlinear matrix equation like (7.30) can be indeed very hard to find. On the
contrary, theoretical characterizations of systems admitting static output stabilizing
feedbacks expressed in the form of nonlinear generalized Riccati equations can be
actually found in the existing literature. For instance, the following theorem appears
in [29].
7.2 Static Output Feedback 131

Theorem 7.11 System (7.24) is stabilizable by a static output linear feedback if and
only if there exist symmetric, positive definite n × n matrices P, Q and a matrix M
(with m rows and n columns) such that

At P + P A − E im M t B t P − P B M E im = −Q . (7.33)

Moreover, when (7.33) holds, a stabilizing feedback can be taken of the form
u = K y, with K = −MC † .

Proof To prove the necessity of the condition, let us assume that the system is
stabilized by a feedback of the form u = K y for some matrix K . Then we have also
the static state stabilizer

u = K C x = K CC † C x = −MC † C x = −M E im x

where we used the definition of generalized inverse and we set M = −K C. Hence,

the closed loop system must satisfy the Lyapunov matrix equation (Theorem 3.3) for
some symmetric, positive definite matrix P; namely

(A − B M E im )t P + P(A − B M E im )
= At P + P A − E im M t B t P − P B M E im = −I

which is (7.33) with Q = I . As far as the sufficiency is concerned, assuming

that (7.33) holds for some matrices P, Q, M, we apply the static output feedback
u = −MC † y = −M E im x. The conclusion can be easily achieved by repeating the
same computation as before and using the Lyapunov matrix equation as a sufficient
condition. 

Remark 7.1 In [29], condition (7.33) is written in a different, but equivalent, way:
indeed, the authors do not use the formalism of generalized inverse. 

Remark 7.2 It is not difficult to see that (7.30) implies (7.33), setting M = B t P,
and using the matrix identity (7.31). However, we notice that with respect to (7.30),
the matrix equation (7.33) contains the additional unknown M. 

Remark 7.3 Another necessary and sufficient condition for the existence of static
output stabilizing feedbacks was given in [9]. Reformulated in terms of generalized
inverse, this condition reads: there exist matrices P, Q, M (of the same dimensions
as before) such that

At P + P A − P B B t P + (B t P − M E im )t (B t P − M E im ) = −Q . (7.34)

Of course, (7.34) is equivalent to (7.33), but not with the same P and Q, in
general. 
132 7 Stabilization

7.3 Dynamic Output Feedback

The practical difficulties encountered in the static output stabilization problem can be
overcome resorting to a different approach, provided that the system is, in principle,
stabilizable by means of a static state feedback law and a suitable (but natural)
technical condition is met. The new approach we are going to describe in this section
is dynamic output feedback.
Definition 7.5 We say that system (7.24) is stabilizable by dynamic output feedback
if there exists a system 
ż = F z + Gv
(7.35)
ξ = Dz

with z ∈ Rν , v ∈ R p , ξ ∈ Rm such that the composed system obtained by means of

the substitutions u = ξ and v = y

ẋ = Ax + B Dz
ż = F z + GC x

has an asymptotically stable equilibrium position at the origin (x, z) = (0, 0) ∈

Rn+ν . The system (7.35) represents the compensator, or controller.

In the figure above, 1 and 2 denote respectively the differential parts of (7.24)
and (7.35).
Example 7.7 The system (7.25) (Example 7.1) with the observation function y = x1 ,
can be dynamically stabilized by means of the compensator
⎧
⎪
⎨ż 1 = −z 1 + z 2 + v
ż 2 = −2z 1 − z 2 + v
⎪
⎩
ξ = −z 1 − z 2 .

Indeed, establishing the connection as explained above, we obtain the closed-loop

system ⎧
⎪
⎪ ẋ1 = x2
⎪
⎨ẋ = −z − z
2 1 2
⎪
⎪ ż = x − z + z2
⎪
⎩
1 1 1
ż 2 = x1 − 2z 1 − z 2
7.3 Dynamic Output Feedback 133

whose characteristic equation is λ4 + 2λ3 + 3λ2 + 2λ + 1 = 0. It is not difficult to

check, by the aid of the Routh-Hurwitz criterion, that all the roots have negative real
part. 

The remaining part of this section is devoted to illustrate how the stabilizing
compensator can be constructed in practice, for a general system of the form (7.24).

7.3.1 Construction of an Asymptotic Observer

Definition 7.6 We say that system (7.24) has the detectability property (or that it is
detectable) if there exists a matrix K of appropriate dimensions such that the matrix
L t = At − C t K t is Hurwitz.

A system possesses the detectability property if and only if its dual system is
stabilizable by static state feedback. In particular, each completely observable system
is detectable.

Proposition 7.4 Assume that the given system (7.24) is detectable. For each admis-
sible open loop input u(t) : [0, +∞) → Rm and for each pair of vectors x0 , z 0 ∈ Rn ,
we denote by x(t) the solution of the system

ẋ = Ax + Bu

corresponding to the input u(t) and the initial state x0 , and by z(t) the solution of
the system

ż = Lz + K y + Bu (7.36)

corresponding to the input u(t) and the initial state z 0 . Then we have

lim (x(t) − z(t)) = 0 .

t→+∞

Proof Denote by e(t) = x(t) − z(t) the difference between z(t) and the state x(t).
We have:

ė = ẋ − ż = Ax + Bu − Lz − K C x − Bu = (L + K C)x − Lz − K C x = Le . (7.37)

Recall that the eigenvalues of a matrix are the same as the eigenvalues of its
transpose. Since by assumption all the eigenvalues of L have negative real part, we
conclude that limt→+∞ e(t) = 0 as desired. 

Remark 7.4 Notice that the input of (7.36) is the sum of the same external input
received by the given system and the output of the given system. Proposition 7.4
states that, regardless the initialization of the two systems and assuming that the
134 7 Stabilization

external input is the same, the solutions of (7.36) asymptotically approximate the
solutions of the given system. For this reason, system (7.36) is called an asymptotic
observer and the quantity e(t) introduced in the previous proof is called the error
between the true state x(t) and the observed state z(t). 

7.3.2 Construction of the Dynamic Stabilizer

Now assume that system (7.24) is stabilizable by static state feedback, as well as
detectable. Under this additional hypothesis, we may find a matrix H such that the
matrix (A + B H ) is Hurwitz.
If the full state vector is measurable and available for control purposes, we could
directly apply the feedback u = H x and solve in this way the stabilization problem.
Otherwise, it is natural to try the control law u = H z, where z is the approximation
of x provided by the asymptotic observer (7.36).
Replacing u = H z in (7.24) and in (7.36), and recalling that y = C x, we obtain
the two systems of differential equations

ẋ = Ax + B H z , (7.38)

ż = Lz + K C x + B H z = K C x + (A − K C + B H )z . (7.39)

Lemma 7.1 The system composed by (7.38) and (7.39) is asymptotically stable at
the origin.
Proof Let us introduce, as above, the variable e = x − z. System (7.38) is equivalent
to
ẋ = (A + B H )x − B H e (7.40)

while system (7.39) is equivalent to

ė = Ax + B H z − Az + K C z − B H z − K C x (7.41)
= Ax − Ax + Ae + K C x − K Ce − K C x = Le .

Systems (7.40) and (7.41) can be reviewed as a unique unforced system, whose
matrix is  
A + B H −B H
. (7.42)
0 A − KC

The set of the eigenvalues of the matrix (7.42) is the union of the sets of the
eigenvalues of the matrices A + B H and A − K C which, by construction, are
Hurwitz. The statement is so proven. 
Theorem 7.12 If system (7.24) is stabilizable by static state feedback, and if it is
detectable, then it is stabilizable by a dynamic output feedback, as well.
7.3 Dynamic Output Feedback 135

Proof The system composed by (7.38) and (7.39) can be interpreted as the result of
the connection of the given system (7.24) and the dynamic compensator

ż = (A − K C + B H )z + K v
ξ = Hz .

We have so really constructed a dynamic output feedback: the consistency with

the notation of Definition 7.5, is easily recovered setting F = A − K C + B H ,
G = K , D = H . The argument is finished, thanks to Lemma 7.1. 

We therefore see that the construction of a stabilizing dynamic output feedback

reduces to the construction of a stabilizing static state feedback u = H x for the given
system, and the construction of a stabilizing static state feedback w = −K t z for the
dual system
ż = At z + C t w .

This conclusion is known as the separation principle. We emphasize once more

that in order to construct H and K we need to know A, B and C, but in order to
practically implement the connection, it is sufficient to dispose of the output y.
At a first glance, Theorem 7.12 seems to suggest that the method of dynamic
feedback is more general than the method of static state feedback. As a matter of
fact, these two methods are (theoretically but, recall, not practically) equivalent.

Theorem 7.13 Let the system (7.24) be given, and assume that it is stabilizable by
means of a dynamic output feedback (7.35). Then, the system is stabilizable by means
of a static state feedback, as well.

Proof Assume that a stabilizing dynamic output feedback exists. The closed-loop
system writes as
        
ẋ A 0 x B 0 0 D x
= + .
ż 0 F z 0 G C 0 z

Hence, the system in Rn+ν defined by the matrices

   
A 0 B 0
à = and B̃ =
0 F 0 G

is stabilizable by a static state feedback. Then, according to Theorem 7.6, we must

have  
A + λI 0 |B 0
rank =n+ν
0 F + λI | 0 G
136 7 Stabilization

for each complex number λ with nonnegative real part. This yields

rank (A + λI | B) = n

for each complex number λ with nonnegative real part. 

7.4 PID Control

As already mentioned, many physical systems of interest in practical applications

can be represented by a single linear differential equation

ξ¨ + a ξ˙ + bξ = u . (7.43)

In (7.43) ξ represents the main variable of interest, while u is a scalar input. In

the early literature, the control of such systems is based on the following ideas.

1. A feedback proportional to the main variable, that is

u = k0 ξ . (7.44)

2. A feedback proportional to the derivative of the main variable, that is

u = k1 ξ˙ . (7.45)

3. An input proportional to the integral of ξ(t), that is

 t
u = k2 ξ(τ ) dτ . (7.46)
0

Here, k0 , k1 , k2 are suitable real constants, often referred to as the gains. The
feedback (7.44) is called a P control. It can be reviewed as a static output feedback,
assuming that (7.43) is associated to the observation function y = ξ.
The feedback (7.45) is called a D control. The sum of (7.44) and (7.45), that is
the feedback
u = k0 ξ + k1 ξ˙ (7.47)

˙ (7.47) can be
is called a PD control. Since the full state of (7.43) is the pair (ξ, ξ),
reviewed as a static state feedback for (7.43).
The function defined in (7.46) is called a I control. Notice that (7.46) can be
thought of as a signal generated by the dynamic compensator

ż = k2 ξ . (7.48)

The combination of (7.44), (7.45), (7.46), that is

7.4 PID Control 137
 t
u = k0 ξ + k1 ξ˙ + k2 ξ(τ ) dτ (7.49)
0

is called a PID control. A PID control can be reviewed as a static state feedback for
the system formed by the composition of (7.43) and (7.48). The following example
illustrates the use of the PID control.

Consider the equation

L
ξ¨ + εξ˙ − ξ = u (0 < ε << 1) . (7.50)
g

It represents the linearized equation of an inverted pendulum with (small) friction.

Here, L denotes the length of the bar of the pendulum (of mass m), and g is the gravity
constant.
The main variable ξ = π − θ represents the angle with respect to the upward
oriented vertical line (see the figure). The control u is exerted by a torque applied to
the pivot.

Assume for simplicity that Lg = 1. The free system (i.e., with u = 0) is clearly
unstable. The system can be stabilized by means of a P control, with gain k0 < −1.
However, by means of such a control the decay rate cannot be improved, since it
˙ which it is not affected by a P control.
depends on the coefficient of the derivative ξ,
Now we try a PI control (that is, a linear combination of P and I controls). To this
end, we add the Eq. (7.48) to the system, and write a new system with variables x1 = z,
˙ The matrices involved in this three-dimensional representation are
x2 = ξ, x3 = ξ.
⎛ ⎞ ⎛ ⎞
01 0 0
A = ⎝0 0 1 ⎠ , b = ⎝0 ⎠ , c = (1 1 0) .
0 1 −ε 1

The choice of c is related to the form u = k2 x1 + k0 x2 of the desired feedback.

The system is completely controllable and observable, and can be actually stabilized
by static output feedback. However, as before, the decay rate cannot be improved.
138 7 Stabilization

Finally, we can easily see that the system is superstabilizable if a PID control
is used. Unfortunately, feedbacks involving a D control are not easy to implement,
because measuring the derivative of a variable is usually in practice a hard task.
Nevertheless, even today PID control is very popular in industrial applications.

Chapter Summary

In this chapter, the two main topics studied in this book (stability and control)
encounter each other. We address the stabilization problem, that is the problem of
improving the stability performances of a system by applying a suitable feedback
law. We consider several approaches: static state feedback, static output feedback
and dynamic feedback. Finally we revisit in this framework the classical PID control
method.
Chapter 8
Frequency Domain Approach

As illustrated in the last section of Chap. 4, one possible approach to the analysis
of linear differential equations of any order with forcing term makes use of the
Laplace transform. This approach reveals a great significance when the forcing term
is interpreted as a control. In this chapter, we first generalize the Laplace transform
approach to MIMO systems with external inputs. For the particular case of SISO
systems which are completely controllable and completely observable, we show
how, by this approach, we can obtain a simple, alternative solution of the synthesis
problem, which means the explicit construction of static output stabilizing feedback
laws.

8.1 The Transfer Matrix

Consider a finite dimensional, time invariant linear system represented by the equa-
tions 
ẋ = Ax + Bu
(8.1)
y = Cx ,

where, as usual, x ∈ Rn , u ∈ Rm and y ∈ R p , with n, m and p positive integers.

Making use of the variation of constants formula, the output corresponding to an
initial state x0 and an input u(·) can be written
 t
y(t) = C x(t) where x(t) = e(t−τ )A Bu(τ ) dτ + et A x0 . (8.2)
0

An alternative representation can be obtained by applying the vector Laplace

transform (Sect. B.4) to both side of the first equation in (8.1). To this end, we

© Springer Nature Switzerland AG 2019 139

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5_8
140 8 Frequency Domain Approach

assume, as we did in Sect. 4.4, that u(·), and hence also x(·) and y(·), are defined for
t ≥ 0. Moreover, we restrict the set of admissible inputs to subexponential functions
of the class PC([0, +∞), Rm ). By virtue of (8.2), this implies in turn that also x(t)
and y(t) are subexponential.
Let X (s) = L[x(t)], U (s) = L[u(t)] and Y (s) = L[y(t)]. We have

L[ẋ(t)] = s X (s) − x0 = L[Ax(t) + Bu(t)] = AX (s) + BU (s)

which yields
−(A − s I )X (s) = x0 + BU (s) .

The matrix (A − s I ) is invertible, except for those values of s ∈ C which coin-

cides with some eigenvalue of A. Let us denote by σ0 the maximal real part of the
eigenvalues of A. Therefore, if Re s > σ0 , we may write

X (s) = −(A − s I )−1 x0 − (A − s I )−1 BU (s)

and
Y (s) = C X (s) = T (s)U (s) + C(s I − A)−1 x0 (8.3)

where we set
T (s) = C(s I − A)−1 B . (8.4)

The analogy between (8.2) and (8.3) is not surprising: both contain the sum of
two terms; one of them depends on the initial state, the other depends on the input
function. As a matter of fact, (8.3) can be alternatively obtained by applying the
Laplace transform to (8.2) for s > σ0 , and making use of Proposition B.4.
Comparing (8.3) with (4.24) of Chap. 4, we see that the role of the polynomial
pch (s) is now played by the matrix (A − s I ).

Remark 8.1 Summing up, we have at our disposal two ways in order to represent a
physical system with input and output: the matrix (8.4) and the Eq. (8.1) identified
in short, in what follows, by the triplet of matrices (A, B, C). When (8.4) is used,
we say that the system is represented according to the frequency domain approach.
This terminology comes from the classical problem of frequency response analysis
illustrated in Sect. 4.4.2. When (8.1) is used, we say that the system is represented
according to the time domain approach. In principle, we should expect that both
representations supply the same information about the behavior of the system but,
as we shall see later, this is only partially true.
Notice also that in the frequency domain approach, the notion of “state” of the
system is not explicitly involved. On the other hand, (8.3) requires purely algebraic
computations, while in order to solve (8.1) integral calculus is needed. 
8.1 The Transfer Matrix 141

The matrix T (s) given by (8.4) is called the transfer matrix. Notice that T (s) is
independent of the initial conditions, so that in order to compute it, we may assume
x0 = 0. In the case of a SISO system i.e., when p = m = 1, the transfer matrix
reduces to a unique element. When in addition the system is defined by a single
linear differential equation of order n, it coincides with the transfer function already
introduced in Remark 4.7.
In principle, it is possible to compute explicitly the transfer matrix making use
of the formula (8.4): the main difficulty rests on the computation of the inverse
of (s I − A). Such inverse matrix is sometimes called the resolvent of A. For our
purposes, the following proposition is sufficient.

Proposition 8.1 If s ∈ C is not an eigenvalue of A, then

1
(s I − A)−1 = M(s) (8.5)
s n + a1 s n−1 + · · · + an

where s n + a1 s n−1 + · · · + an = (−1)n p A (s), and M(s) is a matrix, whose entries

are polynomials of degree less than or equal to n − 1 (recall that p A (s) denotes the
characteristic polynomial of A).

The proof of (8.5) is easily obtained, having in mind the construction of the inverse
of a matrix based on cofactors. A more precise formula for (s I − A)−1 can be found
for instance in [25], p. 12.

Example 8.1 Consider the system


ẋ1 = x1 − x2 + u 1
ẋ2 = x1 + u 2

with observation map y = (x1 + x2 )/2. The input, state and output spaces have
respectively dimension equal to 2, 2, 1. The matrices which define the system are
   
1 −1 10 1 
A= B= C= 2
, 1
2
.
1 0 01

The system is completely controllable and completely observable. The matrix

 
s−1 1
sI − A =
−1 s
√
is invertible for s = (1 ± i 3)/2, and
 
−1 1 s −1
(s I − A) = 2 .
s −s+1 1 s−1
142 8 Frequency Domain Approach

The matrix T (s) coincides with the row-vector

1  
T (s) = s+1, s−2
2(s 2 − s + 1)

8.2 Properties of the Transfer Matrix

Given a system of the form (8.1), we now list some important properties of its transfer
matrix T (s) defined in (8.4).

Property 1 Each element of T (s) is a proper rational function of s ∈ C. Moreover,

the denominator of each element of T (s) is a polynomial of degree less than or equal
to n.

Property 1 is a straightforward consequence of Proposition 8.1.

Remark 8.2 Property 1 points out that rational functions play a relevant role in the
frequency domain approach. Thus, it is convenient to fix some terminology. Let
R1 (s), R2 (s) be two rational functions of the complex variable s. We agree that the
expression “R1 (s) and R2 (s) are equal (or coincide, or are the same function)” means
that there exists a real number r0 such that

R1 (s) = R2 (s) ∀s ∈ C with Re s > r0 . (8.6)

To understand the meaning of this definition, we may look at the following exam-
ple. Let
s 1
R1 (s) = , R2 (s) = .
s(s − 1) s−1

These functions cannot be considered “the same function” in the usual sense, since
they do not have the same domain. However, they are “equal” in the aforementioned
sense. Notice that if the numerator and the denominator of a rational function are
polynomial of high degree, it may be very hard to recognize the existence of possible
common factors and to get rid of them.
Let N (s), D(s) be two polynomials. We say that N (s) and D(s) are coprime
polynomials if they do not have common factors. In this case, we also say that the
rational function N (s)/D(s) is written in lowest terms.
This terminology extends in the obvious way to matrices whose elements are
rational functions. For instance, we say that a matrix T (s) is written in lowest terms
if the common factors have been canceled in all its entries. 
8.2 Properties of the Transfer Matrix 143

Property 2 Let σ0 be the maximum of the real parts of the eigenvalues of A, and let
W (τ ) = Ceτ A B. Then
T (s) = L[W (t)] (8.7)

for each s ∈ C such that s > σ0 .

Formula (8.7) is an easy consequence of (B.26). The matrix W (t) has been already
encountered in Chap. 6 (Lemma 6.2 and subsequent comments). It can be interpreted
as the matrix of the impulse response (Sect. 1.2): indeed, the columns of W (t)
coincide with the output functions of the system corresponding to the initial state
x0 = 0 and the impulsive inputs δ(t)e1 , . . . , δ(t)em .
Property 1 could be also recovered from Property 2 taking into account the rules
of the Laplace transform. Indeed, the elements of et A are obtained as sum of terms of
the type q1 (t)eμt cos ωt and q2 (t)eμt sin ωt, where q1 (t), q2 (t) are polynomials and
λ = μ + i ω is an eigenvalue of A.
Notice that T (s) coincides with the Laplace transform of W (t) only if s > σ0
but its natural domain, as a rational function, contains all the points of the complex
plane, with finitely many exceptions.

Property 3 Let the system 

x̃˙ = Ã x̃ + B̃u
(8.8)
y = C̃ x̃

be linearly equivalent to (8.1), according to the definition of Sect. 5.3. Then, (8.1)
and (8.8) have the same transfer matrix.

Indeed, if (8.1) and (8.8) are linearly equivalent, then there exists a nonsingular
matrix P such that à = P −1 A P, B̃ = P −1 B, and C̃ = C P. To compute the
transfer matrix of system (8.8) we may apply the usual procedure, starting with the
differential part of the system. Without loss of generality, we assume a vanishing
initial state. We have

s X̃ (s) = Ã X̃ (s) + B̃U (s) = P −1 A P X̃ (s) + P −1 BU (s)

which yields
(s I − P −1 A P) X̃ (s) = P −1 BU (s)

or
P −1 (s I − A)P X̃ (s) = P −1 BU (s) .

Then we proceed in the following way: we multiply both sides on the left first by
P, then by (s I − A)−1 , and finally by P −1 . We obtain

X̃ (s) = P −1 (s I − A)−1 BU (s) .

144 8 Frequency Domain Approach

In conclusion,

Y (s) = C̃ X̃ (s) = C P P −1 (s I − A)−1 BU (s) = C(s I − A)−1 BU (s) .

Property 4 The transfer matrix T (s) depends only on the completely controllable
and completely observable part of system (8.1).

The proof of the statement above makes use of the decomposition of Sect. 5.3,
which can be recovered by linear equivalence. Recall that if the matrix A has a
triangular block structure, the exponential matrix et A has an analogous triangular
block structure, too. Then, it is not difficult to see that W (t) = Cet A B = C2 et A22 B2
(see Sect. 5.3 for the notation).

Definition 8.1 Assume that every element of T (s) has been reduced to lowest terms.
We say that the complex number s0 is a pole of the system with multiplicity μ ≥ 1
if:
1. the denominator of at least one element of T (s) can be exactly divided by (s −
s 0 )μ ;
2. there exists no element of T (s) whose denominator can be exactly divided by
(s − s0 )μ+1 .

In other words, the poles of a system are the points of the complex plane where
at least one of the elements of the matrix T (s) is not defined.

Property 5 If s0 is a pole of T (s) with multiplicity μ, then s0 is an eigenvalue of A

with algebraic multiplicity greater than or equal to μ.

Property 5 follows directly from Proposition 8.1. On the contrary, it may happens
that A possesses some eigenvalue s0 which is not a pole T (s). We shall come on this
point very soon.

Remark 8.3 Properties 4 and 5 imply that if s0 is a pole of T (s) with multiplicity
μ, then s0 is an eigenvalue of the matrix A22 (i.e., the matrix of the completely
controllable and completely observable part of system) with algebraic multiplicity
greater than or equal to μ. 

8.3 The Realization Problem

In the previous section we saw how to determine the transfer matrix of a system
given under the form (8.1). Now we address the inverse problem. Namely, we want
to know if (and how) it is possible to recover the representation (8.1), when the
system is assigned by means of its transfer matrix.
8.3 The Realization Problem 145

Definition 8.2 Let T (s) be a matrix with p rows and m columns, whose elements
are proper rational functions of the variable s ∈ C. The triplet of matrices (A, B, C)
whose dimensions are respectively n × n, n × m, p × n, is said to be a realization
of T (s) if T (s) coincides with the transfer function of the system (8.1) defined by
means of the matrices A, B, C. The number n is said to be the dimension of the
realization.

In order to illustrate some difficulties of the problem, we propose two examples.

Example 8.2 Let us consider the SISO system defined by the equations
⎧
⎪
⎨ẋ1 = x2
ẋ2 = −2x1 + 3x2 + u
⎪
⎩
y = x1 − x2 .

The differential part of the system is equivalent to the scalar equation ξ  − 3ξ  +

2ξ = u, where we set ξ = x1 . The transfer function is

1−s 1−s 1
T (s) = = = .
s2 − 3s + 2 (s − 1)(s − 2) 2−s

The eigenvalues of the system matrix



0 1
A=
−2 3

are 1 and 2, while the unique pole of the transfer function is 2. This example shows
that the number of the poles can be actually less than the number of the eigenvalues:
this is related to the cancelation of common factors appearing at the numerator and at
the denominator of the transfer function. Notice that the system at hand is completely
controllable but not completely observable. We also notice that the given system and
the system represented by the single equation

ξ˙ = 2ξ − u
y=ξ

have the same transfer function. 

Example 8.3 Consider the SISO system

⎧
⎪
⎨ẋ1 = u
ẋ2 = x2
⎪
⎩
y = x1 + x2 .
146 8 Frequency Domain Approach

Denoting by A the system matrix, we readily obtain

 
s 0
sI − A = .
0 s−1

This matrix is invertible for s ∈

/ {0, 1}, and
 
1 s−1 0
(s I − A)−1 = .
s(s − 1) 0 s

It follows that T (s) = 1/s. This time, we notice that the system is completely
observable, but not completely controllable. The transfer matrix is the same as the
transfer matrix of the system 
ξ˙ = u
y=ξ.

Examples 8.2 and 8.3 point out that systems with different time domain represen-
tation may have the same transfer function. In other words, the realization problem
does not have, in general, a unique solution. Of course, taking into account that
the transfer function depends only on the completely controllable and completely
observable part of the system (Property 4 above), this is not surprising. Examples
8.2 and 8.3 suggest also that the presence of uncontrollable or unobservable parts
may lead to the cancelation of common factors at the numerator and the denomina-
tor of some element of the transfer matrix: this implies in turn a loss of information
about the evolution of the internal state of the system. To distinguish among different
realizations of the matrix T (s), the following definitions are useful.

Definition 8.3 Let T (s) be a matrix with p rows and m columns, whose elements
are proper rational functions of the variable s ∈ C. A realization (A, B, C) of T (s)
is said to be minimal if, denoting by n its state space dimension, the state space
dimension of any other realization of T (s) is greater than (or equal to) n.
A realization (A, B, C) is said to be canonical if the system (8.1) defined by the
triplet A, B, C is completely controllable and completely observable.

Theorem 8.1 Let T (s) be a matrix with p rows and m columns, whose elements are
proper rational functions of the variable s ∈ C. A realization of T (s) is minimal if
and only if it is canonical.

Proof If the realization (A, B, C) is minimal then it must be completely controllable

and completely observable. Indeed, assume the contrary. Then according to Property
4, the completely controllable and completely observable part of the system defined
by the triplet (A, B, C) provides another realization of T (s), whose state space
dimension is strictly less than the previous one.
8.3 The Realization Problem 147

To prove the converse, let us first recall (see Sects. 5.1.3 and 5.2.3) that a system
of the form (8.1) is:
(1) completely controllable if and only if the matrix
 T
e−σ A B B t e−σ A dσ
t
Γ (T ) =
0

is nonsingular for some (and hence for each) T > 0;

(2) completely observable if and only if the matrix
 T
eσ A C t Ceσ A dσ
t
E(T ) =
0

is nonsingular for some (and hence for each) T > 0.

Even in this case we may argue by contradiction. Let (A, B, C) be a canonical
realization of dimension n and let us assume that for some system

x̃˙ = Ã x̃ + B̃u
y = C̃ x̃

where à is a square matrix of dimensions ν × ν with ν < n, we have

T (s) = L[Cet A B] = L[C̃et à B̃] .

By applying L−1 to both sides, we get Cet A B = C̃et à B̃ for each t ≥ 0, and so
also for each t ∈ R. Now let τ , σ be two arbitrary real numbers. We have:

Ceτ A e−σ A B = Ce(τ −σ)A B = C̃e(τ −σ) Ã B̃ = C̃eτ Ã e−σ Ã B̃ .

Multiplying by eτ A C t to the left and by B t e−σ A to the right we obtain

t t

eτ A C t Ceτ A e−σ A B B t e−σ A = eτ A C t C̃eτ Ã e−σ Ã B̃ B t e−σ A .

t t t t

Next we integrate both sides on the square 0 ≤ τ ≤ T , 0 ≤ σ ≤ T . The result

can be written as
E(T ) · Γ (T ) = Ẽ(T ) · Γ˜ (T ) (8.9)
T T
where Ẽ(T ) = 0 eτ A C t C̃eτ Ã dτ and Γ˜ (T ) = 0 e−σ Ã B̃ B t e−σ A dσ.
t t

The matrix E(T ) · Γ (T ) is, by hypothesis, the product of two nonsingular matri-
ces. Hence, it has maximal rank equal to n. On the other hand, Ẽ(T ) has only ν
columns, and Γ˜ (T ) only ν rows. Hence, the rank of their product cannot be greater
than ν < n. Therefore, we conclude that the identity (8.9) is false and the statement
is proved. 
148 8 Frequency Domain Approach

The realization problem has not a unique solution, not even if we limit ourselves
to minimal realizations. The reason is clear, having in mind Property 3. Indeed, the
first essential step in the construction of a realization is the choice of a real vector
space to serve as state space. But, in order to write the Eq. (8.1), one needs also to fix
a system of coordinates in this space. This choice is, of course, absolutely arbitrary:
for different choices of the system of coordinates the system equations will look
formally different, although of the same dimension. The following results clear up
the situation.

Theorem 8.2 Let T (s) be a matrix with p rows and m columns, whose elements
are proper rational functions of the variable s ∈ C. Then there exists at least one
realization of T (s).

Of course, if there exists one realization of T (s), then there exists also a minimal
(and canonical) realization.

Proposition 8.2 Under the same assumptions of Theorem 8.2, if (A, B, C) (with
state space dimension equal to n) and ( Ã, B̃, C̃) (with state space dimension equal
to ñ) are two canonical realizations of T (s), then n = ñ and, moreover, the systems
respectively defined by the triplets (A, B, C) and ( Ã, B̃, C̃) are linearly equivalent.

For the proofs of these results we refer to [3, 6, 23].

8.4 SISO Systems

In this section we focus our attention on SISO systems. Thus in what follows, we
always have m = p = 1. Moreover, the transfer matrix reduces to a unique element
represented by a proper rational function, referred to as the transfer function.

8.4.1 The Realization Problem for SISO Systems

Consider the rational function of the variable s ∈ C

N (s) c0 + c1 s + · · · + ck s k
T (s) = = n (8.10)
D(s) s + a1 s n−1 + · · · + an

where N (s) and D(s) may possibly have common factors. If k < n, T (s) is proper,
and so according to Theorem 8.2 there must exist a realization. Next proposition
provides a simple, explicit construction for such a realization, and so it provides also
a proof of Theorem 8.2 for the particular case of SISO systems.
8.4 SISO Systems 149

Proposition 8.3 If k < n, the system

⎧
⎪
⎪ ẋ1 = x2
⎪
⎨. . .
(8.11)
⎪
⎪ ẋn−1 = xn
⎪
⎩
ẋn = −an x1 − · · · − a1 xn + u(t)

with the observation function

y = c0 x1 + c1 x2 + · · · + ck xk+1 (8.12)

represents a realization of (8.10).

Remark 8.4 We emphasize that the dimension of realization (8.11) is equal to the
degree of the polynomial D(s). 

Proof of Proposition 8.3 Let C be the matrix of system (8.11). Since C is in companion
form, system (8.11) is equivalent to the differential equation of order n

ξ (n) + a1 ξ (n−1) + · · · + an−1 ξ  + an ξ = u(t) (8.13)

where we set ξ = x1 . Taking the Laplace transform of both sides, with the usual
notation, we get
1
(s) = U (s)
pch (s)

where pch (s) = s n + a1 s n−1 + · · · + an = D(s) coincides with the characteris-

tic polynomial of (8.13) and, by a possible change of sign, with the characteristic
polynomial pC (s) of the matrix C. Taking into account (8.12), we easily get

c0 + c1 s + · · · + ck s k N (s)
Y (s) = U (s) = U (s) = T (s)U (s) .
pch (s) D(s)

Thus we see that T (s) coincides with the transfer function of system (8.11), (8.12).

Clearly, the realization (8.11), (8.12) provided by Proposition 8.3 is completely
controllable, but not necessarily completely observable. The following proposition
concludes the reasoning.
Proposition 8.4 Let the rational function (8.10) be given, and let k < n. The poly-
nomials N (s) and D(s) are coprime if and only if system (8.11) with the observation
function (8.12) represents a minimal realization of (8.10).
Proof Let us assume that numerator and denominator of T (s) do not have common
factors. We already know that the system (8.11) with observation function (8.12) is
150 8 Frequency Domain Approach

a completely controllable realization of (8.10). If it is not completely observable,

taking the observable part we can obtain another realization for which the dimen-
sion of the state space is ñ < n. Because of Property 1, we could rewrite T (s) as
T (s) = Ñ (s)/ D̃(s), where D̃(s) is a polynomial of degree not greater than ñ. This
is impossible, since by assumption there is no common factors to cancel in N (s) and
D(s).
Vice versa, let us assume that the system (8.11), (8.12) satisfies the complete
observability condition, in addition to the complete controllability one. By Propo-
sition 8.3, the transfer function of system (8.11), (8.12) coincides with T (s) =
N (s)/D(s). Recall that the dimension of the state space of (8.11) is, by construction,
equal to the degree of D(s).
If there are common factors to cancel in N (s) and D(s), we could rewrite T (s)
as Ñ (s)/ D̃(s), where D̃(s) is a polynomial of degree ñ < n. But then, using again
Proposition 8.3, we can construct another realization of dimension ñ < n. This would
imply that the realization (8.11), (8.12) is not minimal, and hence not canonical. This
contradicts the assumption. 

Remark 8.5 In particular, if in (8.10) we have k = 0 and c0 = 1, (8.12) reduces to

y = x1 and the realization provided by Proposition 8.3 is canonical. 

Finally, we show that in the case of SISO systems, Property 5 of the previous
section admits a partial converse.
Proposition 8.5 Let a system (8.1) with m = p = 1 be given, and let T (s) be its
transfer function. Assume that the system is completely controllable and completely
observable. If λ is an eigenvalue of A of algebraic multiplicity μ, then λ is a pole of
T (s) of multiplicity μ.
Proof According to Property 1, the transfer function can be written in the form

N (s)
T (s) =
D(s)

where the degree of the polynomial D(s) is not greater than n. If the degree of D(s)
is strictly less than n, then T (s) would admit realizations of dimension strictly less
than n, a contradiction to the complete controllability and complete observability
assumptions. The unique possibility is therefore that the degree of D(s) is exactly
equal to n.
It follows that T (s) has exactly n poles (counting possible multiplicities). Let us
denote by s1 , . . . , sk the distinct poles of T (s), and by μ1 , . . . , μk their multiplicities,
so that μ1 + · · · + μk = n. By Property 5, every si is an eigenvalue of A and its
algebraic multiplicity is greater than or equal to μi . But the eigenvalues of A (counting
multiplicities) cannot be more than n. Then, if there is some eigenvalue of A different
from s1 , . . . , sk , or if for some index i, si regarded as an eigenvalue of A would have
multiplicity strictly greater than μi , we get a contradiction. 
8.4 SISO Systems 151

Example 8.4 Consider a system for which the set of the admissible inputs is restricted
to the set of functions u(t) of class C k on the interval [0, +∞). Assume that the
evolution of the system is determined by a linear differential equation of order n

ξ (n) + a1 ξ (n−1) + · · · + an−1 ξ  + an ξ = ck u (k) + · · · + c0 u (8.14)

where a1 , . . . , an , c0 , . . . , ck are real numbers, with n > k. Moreover, we assume

that the output y coincides with ξ. Such a model, because of the presence of the
derivative of the input, seems not to be covered by the form (8.1).
Let us apply the Laplace transform to both sides of (8.14). Assuming ξ(0) =
ξ  (0) = · · · = ξ (n−1) (0) = 0 and u(0) = u  (0) = · · · = u (k) (0) = 0, we obtain

(s n + a1 s n−1 + · · · + an )(s) = (ck s k + · · · + c0 )U (s)

that is
ck s k + · · · + c0
Y (s) = (s) = U (s)
s n + a1 s n−1 + · · · + an

where, according to the assumption m = p = 1, the transfer matrix reduces to the

scalar function
ck s k + · · · + c0
T (s) = n . (8.15)
s + a1 s n−1 + · · · + an

This is a proper rational function, possibly with some common factors in the
numerator and the denominator. Note that the denominator coincides with the charac-
teristic polynomial pch (s) of the homogeneous equation associated to (8.14). There-
fore, system (8.14) can be realized by means of (8.11) with observation function
(8.12). Recall that such a realization is completely controllable, but could be not
completely observable.
In the modeling of physical systems, it is not rare the case where the derivative
of the input appears explicitly in the equations; this happens for instance when a
nonholonomic constraint is modeled as an input. 

8.4.2 External Stability

As a consequence of Proposition 8.5, and recalling the conclusions of Chaps. 3 and 6

(in particular, Theorem 3.1, Theorem 6.2 and Proposition 6.5), we can immediately
state the following theorem.

Theorem 8.3 Let a system (8.1), with m = p = 1, be given. Let T (s) be its transfer
function. Assume that (8.1) is completely controllable and completely observable.
Then the following statements are equivalent:
(1) all the eigenvalues of A have negative real part;
(2) all the poles of the transfer function have negative real part;
152 8 Frequency Domain Approach

(3) the system is internally stable;

(4) the system is BIBO-stable.
Remark 8.6 Recall that for systems which are not completely controllable or not
completely observable, the previous statements are no more equivalent, in general.
For instance, the system ⎧
⎪
⎨ẋ1 = −x1 + u
ẋ2 = x2
⎪
⎩
y = x1 + x2

is not BIBO-stable, but its transfer function T (s) = 1/(s + 1) has the unique pole
s = −1. 

Remark 8.7 Theorem 8.3 applies in particular to BIBO systems of the special form

y (n) + a1 y (n−1) + · · · + an−1 y  + an y = u(t) (8.16)

where y is taken as the output variable. Indeed, systems of this form are recognized to
be completely controllable and completely observable (Remarks 5.4 and 5.7). Hence,
for such systems, external stability and internal stability are equivalent properties. 

The equivalence (2) ⇐⇒ (4) of Theorem 8.3 can be proved by using only fre-
quency domain methods. Consider, for simplicity, the case of a system defined by
the second order equation
y  + a1 y  + a2 y = u (8.17)

with output variable y, under the assumption that the characteristic roots s1 , s2 of the
associated unforced equation are real and distinct. Recall that the solutions of (8.17)
can be put in the form
y(t) = ϕ(t) + χ(t) (8.18)

where ϕ(t) and χ(t) represent respectively the general solution of the associated
homogeneous equation i.e., the unforced solution, and χ(t) represents the inverse
Laplace transform of the function U (s)/ pch (s) i.e., the solution corresponding to
the zeroed initial state (compare with (4.24)). Here, with the usual notation, U (s)
denotes the Laplace transform of the input u(t). The following lemma exploits the
linearity of the system; the argument is similar to that used in the proof of Theorem
6.1.
Lemma 8.1 The solution y(t) is bounded on the interval [0, +∞) for each initial
condition and each bounded input u(t) if and only if both the following conditions
are met:
(i) the function ϕ(t) is bounded on [0, +∞) for each initial condition;
(ii) the function χ(t) is bounded on [0, +∞) for each bounded input u(t).
8.4 SISO Systems 153

Proof The sufficient part is straightforward, by virtue of the triangular inequality

|y(t)| ≤ |ϕ(t)| + |χ(t)| .

The necessary part can be proved by contradiction. Indeed, assume that there
is a choice of the initial conditions y0 , y1 , for which ϕ(t) is not bounded. Let us
apply the zero input and let the system evolve with these initial conditions. We have
y(t) = ϕ(t) for each t ≥ 0, so that y(t) is unbounded, as well. To prove that also
χ(t) must be bounded, we can argue in a similar way. 

The poles of the transfer function coincide with the characteristic roots of the
unforced equation associated to (8.17), and the condition that all the characteristic
roots have negative real part is necessary and sufficient for internal stability. There-
fore, under this condition, the function ϕ(t) is bounded in [0, +∞).
Let us show that the same condition implies the boundedness of χ(t) as well,
provided that the inputs are bounded. Recalling that χ(t) can be written in the form
(4.25) we recall Proposition 1.5, which states that proving the boundedness of χ(t) is
equivalent to proving that the integral of the function h(ρ) is absolutely convergent.
To this end, we need an explicit expression of h(ρ). Since in this case pch (s) =
s 2 + a1 s + a2 = (s − s2 )(s − s1 ), the Laplace transform of χ(t) is
1 1
· · U (s) . (8.19)
s − s1 s − s2

This last expression tells us that the system acts as a cascade connection of two
systems of the first order

ẏ = s2 y + v and v̇ = s1 v + u .

Solving independently these two systems, we obtain

 t  t
y(t) = e(t−τ2 )s2 v(τ2 ) dτ2 and v(t) = e(t−τ1 )s1 u(τ1 ) dτ1 .
0 0

Combining these two expressions leads to

 t  τ2 
(t−τ2 )s2 (τ2 −τ1 )s1
y(t) = e e u(τ1 ) dτ1 dτ2 .
0 0

We can eliminate the variable τ2 by changing the order of integration and applying
the substitution τ2 − τ1 = r . We find
 t  t 
y(t) = e(t−τ2 )s2 e(τ2 −τ1 )s1 dτ2 u(τ1 ) dτ1
0 τ1
 t  t−τ1 
(t−τ1 −r )s2 r s1
= e e dr u(τ1 ) dτ1 .
0 0
154 8 Frequency Domain Approach

Comparing this last expression with (4.25) and applying further the substitution
ρ = t − τ1 , we finally get1
 ρ
h(ρ) = e(ρ−r )s2 er s1 dr . (8.20)
0
∞
Now it is straightforward to see that the integral 0 h(ρ) dρ is absolutely con-
vergent, by virtue of the assumption s1 < 0, s2 < 0. To finish the proof, we show
that the negativity of the real part of the characteristic roots of the unforced system
associated to (8.17) is also necessary for the boundedness of χ(t). Computing the
integral in (8.20), we immediately find

es1 ρ − es2 ρ
h(ρ) = . (8.21)
s1 − s2

If both s1 and s2 (that are distinct by assumption) are not zero, we have with a
further integration
 ∞  
1 es1 T es2 T 1 1
h(ρ) dρ = lim − − + .
0 T →∞ s1 − s2 s1 s2 s1 s2

Hence, if at least one is positive, the integral is not convergent. If one is equal to
zero (say, s1 ), we have:
 ∞  
1 es2 T 1
h(ρ) dρ = lim T− + .
0 T →∞ s1 − s2 s2 s2
∞
and also in this case the integral 0 h(ρ) dρ does not converge, regardless the sign
of s2 .

1 Alternatively, performing a decomposition to partial fractions, we can rewrite (8.19) as

 
1 U (s) U (s)
− .
s2 − s1 s − s2 s − s1

Taking the inverse transform we obtain

 t  t 
1
es2 (t−τ ) u(τ ) dτ − es1 (t−τ ) u(τ ) dτ
s2 − s1 0 0
 t 
1 s2 (t−τ ) s1 (t−τ )
= [e −e ]u(τ ) dτ
s2 − s1 0

and we can recover (8.20), comparing again with (4.25).

8.4 SISO Systems 155

8.4.3 Nyquist Diagram

The so-called Nyquist diagram is a graphic criterion which, applied to the transfer
function, allows us to recognize whether a given system possesses the BIBO stability
property.
Assume that the system is given by means of its state equations (8.1), where
m = p = 1, and that it is completely controllable and completely observable, so
that the conclusions of Theorem 8.3 hold. Moreover, by virtue of Proposition 8.4,
under the same conditions as before, there is no loss of generality assuming that the
numerator and the denominator of the transfer function are coprime.
Now let T (s) be a proper rational function of the variable s ∈ C, with no common
factors. A number s0 ∈ C is said to be a zero2 of T (s) if T (s0 ) = 0.
Let us denote by w = T (s) ∈ C the dependent variable. Any complex number s
can be thought of as a point of a plane, where a system of coordinates has been fixed
(Re s, Ims). Analogously, any complex number w will be thought of as a point in a
plane referred to the coordinates (Re w, Imw). A continuous map s = γ(t) from R
to C can be interpreted as a planar curve. Analogously, the image of γ(t) throughout
T can be interpreted as a planar curve w = δ(t) = T (γ(t)).
If s = γ(t) is simple and closed, it surrounds an open and bounded region Γ ⊂ C.
Of course, if s = γ(t) is simple and closed, δ(t) is closed, but it is not necessarily
simple.
Example 8.5 Let T (s) = 1/(s − 1)(s − 2). Figure 8.1 shows the curve δ obtained
by applying T to the circumference

Re s = 2 + 2 cos t
Ims = 2 sin t .

Example 8.6 Let T (s) = 1/(s − 1)2 , and let δ be now obtained by applying T to
the circumference 
Re s = 1 + cos t
Ims = sin t .

Figure 8.2 may give the wrong impression of a simple and closed curve. Actually
the curve is run twice. 

Let s = γ(t) be a simple and closed curve. Let us denote by Z the number of zeros
of T (s) lying in Γ and by P the number of poles of T (s) lying in Γ . For simplicity,
we assume that there is neither zeros nor poles on the contour of Γ . We need the

2 IfT (s) is the transfer function of a system, its zeros give useful information about the behavior of
the system: the interested reader is referred to [8].
156 8 Frequency Domain Approach

Fig. 8.1 The curve δ of 0.8

Example 8.5
0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8

Fig. 8.2 The curve δ of

Example 8.6 1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1

−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

following, classical result from the theory of functions of a complex variable (see
for instance [1]).

Argument principle Let Q be the integer number denoting how many times the
curve δ(t) encircles the origin in counterclockwise sense, while the contour of Γ is
run once in the counterclockwise sense. Then, Q = Z − P.

Definition 8.4 The Nyquist diagram of a proper rational function T (s) is the image
of the curve w = T (γ(t)) = δ(t), when γ(t) = −i t (t ∈ R).

The curve γ(t) = −i t generating the Nyquist diagram is not closed. Nevertheless,
the image δ of γ obtained by composition with T , surrounds a bounded region of the
complex plane. Indeed, since T is proper, we have
8.4 SISO Systems 157

lim δ(t) = 0 .
t→±∞

In fact, we may also think of γ(t) as a closed curve, by adding to its domain the
infinity point: completed in this way, we may imagine that γ surrounds the right half
plane of C (the contour being run in the counterclockwise sense). Notice that by
construction, δ(t) = (Re T (−i t), ImT (−i t)).
Let T (s) be a proper rational function without zeros or poles on the imaginary
axis. Drawing the Nyquist diagram and assuming that Z is known, we can now easily
check whether the right half plane of C contains some poles of T (s).
By some suitable modifications, these conclusions can be extended to the case
where T (s) possesses purely imaginary poles or zeros.

8.4.4 Stabilization by Static Output Feedback

Continuing to deal with a SISO system of the form (8.1) satisfying the complete
controllability and the complete observability assumption, in this section we show
how to take advantages of the Nyquist criterion in order to determine a static output
feedback which stabilizes the given system in the BIBO (and hence also in the
internal) sense.
As usual, we denote by u ∈ R the input variable and by y ∈ R the output variable.
First, we examine how the transfer function changes, when a feedback of the form
−ky is added to the external input u: here, k is a positive constant, sometimes called
the gain; the choice of the minus sign is conventional.

Let T (s) be the transfer function of the given system. Let v = u − ky. By the aid
of the figure above, we easily see that

Y (s) = T (s)V (s) = T (s)(U (s) − kY (s))

so that
Y (s) + kT (s)Y (s) = T (s)U (S) .

As a consequence, for each s ∈ C such that 1 + kT (s) = 0,

T (s) 1 T (s)
Y (s) = G(s)U (s) = U (s) = · U (s)
1 + kT (s) k 1
k
+ T (s)
158 8 Frequency Domain Approach

where G(s) denotes the transfer function of the closed loop system. While the value
of parameter k varies, the positions of the poles of the resulting transfer function
G(s) vary in a continuous way. Thus, to accomplish the desired goal, we need to find
a value of k, if any, in such a way that all the poles of the G(s) are moved to the left
half of the complex plane.
For simplicity, we assume that G(s) does not have poles on the imaginary axis,
and we write
1
H (s) = + T (s) .
k

Lemma 8.2 The poles of G(s) coincide with the zeros of H (s).

Proof Write T (s) = N (s)/D(s), where N (s) and D(s) are polynomial. We have

1 N (s) D(s) + k N (s)

H (s) = + = .
k D(s) k D(s)

Hence, s0 is a zero of H (s) if and only if D(s0 ) + k N (s0 ) = 0. On the other hand

N (s) 1 N (s)
G(s) = · = .
D(s) N (s) D(s) + k N (s)
1+k
D(s)

Hence, s0 is a pole of G(s) if and only if D(s0 ) + k N (s0 ) = 0. 

In addition, it is straightforward to realize that the poles of H (s) coincide with

the poles of T (s). Applying the Argument Principle to the rational function H (s)
leads to the following conclusions:

number o f the poles o f G with positive r eal par t

= number o f poles o f T with positive r eal par t + Q

where Q denotes the number of times the curve H (−i t) encircles the origin in the
counterclockwise sense, while the parameter t moves from −∞ to +∞.
On the other hand, it is evident that −Q represents the number of times the curve
T (−i t) encircles the point of coordinates (− k1 , 0) of the complex plane in clockwise
sense, while the parameter t moves from −∞ to +∞. The following statement
resumes the conclusions.

Proposition 8.6 The static output feedback −ky stabilizes in BIBO (and so also in
internal) sense the SISO system (8.1) if the number of times the Nyquist diagram of
its transfer function T (s) encircles the point (− k1 , 0) in clockwise sense while the
parameter t moves from −∞ to +∞, is equal to the number of poles of the given
system lying in the open right half of the complex plane.
8.4 SISO Systems 159

In practical applications, one draws the Nyquist diagram of the given system, and
then checks whether there exists a region D encircled by the diagram the required
number of times. If this region exists and intersects the negative real axis, the sys-
tem is stabilizable. A stabilizing feedback is provided by any value of k such that
(− k1 , 0) ∈ D.

8.5 Disturbance Decoupling

In this last section we discuss an important application which involves both frequency
domain and time domain techniques. Consider the system

ẋ = Ax + Gd
(8.22)
y = Cx

where x ∈ Rn , y ∈ R p , d ∈ Rq . The input d(t) : [0, +∞) → Rq is now interpreted

as a disturbance. In other words, d(t) is a unknown and undesired input; we just
assume that it is piecewise continuous and right continuous, in order to guarantee
existence of solutions. For each initial state x0 , the variation of constants formula
yields  t
y(t, x0 , d(·)) = Cet A x0 + Ce(t−τ )A Gd(τ ) dτ
0

which reduces to
y0 (t) = Cet A x0

when d(t) = 0 for each t ≥ 0. The function y0 (t) is called the uncorrupted output
signal. It may happen that y(t) = y0 (t) even for not vanishing disturbances d(t).

Example 8.7 Clearly, the output of the (not completely observable) system
⎧
⎪
⎪ ẋ1 = x1 − x2 + d
⎪
⎨ẋ = x
2 2
⎪
⎪
⎪
⎩
y = x2

is not affected by the disturbance. 

Definition 8.5 Let us denote, as before, by y0 (t) the uncorrupted output, that is the
output corresponding to some initial state x0 and the vanishing input d(t) = 0. We
say that the system is disturbance decoupled if we have y(t, x0 , d(·)) = y0 (t) for
each t ≥ 0, each initial state x0 and each input d(t).
160 8 Frequency Domain Approach

Proposition 8.7 The following statements are equivalent:

(i) the system (8.22) is disturbance decoupled;
(ii) the impulse response matrix W (t) = Cet A G vanishes for t ≥ 0 (and hence,
being a real analytic function, for each t ∈ R);
(iii) the transfer matrix T (s) = C(s I − A)−1 G vanishes for s ∈ C;
(iv) for each integer k ≥ 0, one has C Ak G = 0 .

Proof The equivalences (i) ⇐⇒ (ii) ⇐⇒ (iii) are straightforward. Thus, we focus
on the statement (iv), and we will prove that it is equivalent to (ii). Assume first that
the identity
W (t) = Cet A G = 0 (8.23)

holds for each t ∈ R. To begin with, the substitution t = 0 yields C G = 0. Coming

back to (8.23) and taking the derivative, we obtain

C Aet A G = 0 (8.24)

for t ∈ R, which implies C AG = 0 by the substitution t = 0. We repeat the

procedure, taking now the derivative of (8.24) and letting again t = 0. This time we
obtain C A2 G = 0. Continuing in this way, we conclude finally that C Ak G = 0 for
each integer k ≥ 0. The converse implication is immediate, since
∞ k
t
W (t) = Cet A G = C Ak G
k=0
k!

for each t ∈ R. 

Remark 8.8 According to the Cayley-Hamilton Theorem, it is sufficient to check

condition (iv) of Proposition 8.7 for k = 0, . . . , n − 1. 

Next we establish a necessary and sufficient condition.

Definition 8.6 Let A be a real matrix of dimensions n × n. A subspace V of Rn is

said to be an algebraic (o geometric) invariant for A if AV ⊆ V .
The subspace V is said to be a dynamic invariant for A if from x0 ∈ V it follows
et A x0 ∈ V for each t ≥ 0 (and hence for each t ∈ R).

Proposition 8.8 The subspace V is an algebraic invariant for A if and only if it is a

dynamic invariant for A.

Proof Let V be an algebraic invariant. For each x0 ∈ V , we clearly have Ax0 ∈ V ,


A2 x0 ∈ V , and so on. Hence, et A x0 = ∞ tk k
k=0 k! A x 0 belongs to V . On the other
hand, let V be a dynamic invariant, and let x0 ∈ V . Then, for each t = 0, we also
have
8.5 Disturbance Decoupling 161

et A x 0 − x 0
∈V .
t
Taking the limit for t → 0, we get Ax0 ∈ V . 

Theorem 8.4 The given system is disturbance decoupled if and only if there exists
a subspace V of Rn which is an algebraic invariant for A, and such that im G ⊆
V ⊆ ker C.

Proof Assume that the system is disturbance decoupled. Let us introduce a matrix H ,
whose columns coincide with the columns of the matrices G, AG, A2 G, . . . , An−1 G,
in this order. The matrix H can be interpreted as a linear map from Rn×q in Rn . Let
V = im H . By the Cayley-Hamilton Theorem, V is an algebraic invariant. The
inclusion im G ⊆ V is obvious, while the other one V ⊆ ker C follows from
Proposition 8.7, (iv).
To prove the converse, we first remark that if a subspace V is an algebraic invari-
ant and im G ⊆ V , then clearly im (Ak G) ⊆ V for each positive integer k. As a
consequence, since V ⊆ ker C, we also have C Ak Gx = 0 for each integer k ≥ 0
and each x ∈ Rn . The conclusion follows, using again Proposition 8.7, (iv). 

There is an other characterization of disturbance decoupled systems. By means of

a linear change of coordinates, we can put the system in the observability canonical
form ⎧
⎪
⎪ ż 1 = A11 z 1 + A12 z 2 + G 1 d
⎪
⎨ż = A z + G d
2 22 2 2
(8.25)
⎪
⎪
⎪
⎩
y = C2 z 2

where z 1 ∈ Rn−r , z 2 ∈ Rr for some nonnegative integer r ≤ n, and the subsystem

⎧
⎪
⎨ż 2 = A22 z 2 + G 2 d
(8.26)
⎪
⎩
y = C2 z 2

is completely observable. The case r = 0 is trivial, so we can assume r > 0.

Theorem 8.5 The system (8.22) is disturbance decoupled if and only if G 2 = 0,

where G 2 is the matrix appearing in (8.25).

Proof The sufficient part is evident (to be formal, it can be easily obtained as an
application of Theorem 8.4). Let us prove the necessary part.
Assume that the system is disturbance decoupled. Taking into account the form
(8.25), for each integer k ≥ 0, we see that C Ak can be written as a row block matrix
(0 | C2 Ak22 ), where 0 denotes here a block of n − r zero columns. From this, it easily
162 8 Frequency Domain Approach

follows that C Ak G = C2 Ak22 G 2 for each integer k ≥ 0. Since the subsystem (8.26)
is completely observable, the matrix
⎛ ⎞
C2
⎜ C2 A22 ⎟
M =⎜ ⎝ ... ⎠
⎟

C2 Ar22−1

has a maximal rank i.e., rank M = r . Now, assume by contradiction that v = G 2 d =

0 for some d ∈ Rq (note that v ∈ Rr and that, necessarily, d = 0). The vector
Mv ∈ R p×r is a linear combination of the r linearly independent columns of M, so
that being v = 0, we also have Mv = 0. But
⎛ ⎞ ⎛ ⎞
C2 G 2 d C2 G 2
⎜ C2 A22 G 2 d ⎟ ⎜ C2 A22 G 2 ⎟
Mv = ⎜
⎝
⎟=⎜
⎠ ⎝
⎟d
⎠ (8.27)
... ...
C2 Ar22−1 G 2 d C2 Ar22−1 G 2

with d = 0. On the other hand, the disturbance decoupling assumption implies

C2 G 2 = C2 A22 G 2 = · · · = C2 Ar22−1 G 2 = 0 . (8.28)

Clearly, (8.27) and (8.28) are in contradiction. Therefore, we must have G 2 d = 0

for each d ∈ Rq , and this means that G 2 = 0. 

If the given system is not disturbance decoupled, we can try to achieve this property
by the use of a suitable feedback law. In other words, we add a control term in the
system equation 
ẋ = Ax + Bu + Gd
(8.29)
y = Cx

where with the usual notation u ∈ Rm , and we ask whether it is possible to find a
static state feedback of the form u = F x such that the closed-loop system

ẋ = (A + B F)x + Gd
y = Cx

is disturbance decoupled. The conditions for answering this question rest on the
introduction of a new notion of invariance, concerning the state equation

ẋ = Ax + Bu . (8.30)
8.5 Disturbance Decoupling 163

Definition 8.7 A subspace V ⊆ Rn is said to be a strong controlled invariant for the

system (8.30) if for each x0 ∈ V and each admissible input u(t) : [0, +∞) → Rm
we have x(t, x0 , u(·)) ∈ V for each t ≥ 0.

Apart from the modified terminology, the definition above coincides with the
notion already introduced in Sect. 5.3.2.

Definition 8.8 A subspace V ⊆ Rn is said to be a weak controlled invariant for the

system (8.30) if for each x0 ∈ V there exists an admissible input u(t) : [0, +∞) →
Rm such that x(t, x0 , u(·)) ∈ V for each t ≥ 0.

Example 8.8 The subspace V = {(x1 , x2 ) : x2 = 0} ⊆ R2 is a weak controlled

invariant, but not a strong controlled invariant, for the system

ẋ1 = x1 + x2
ẋ2 = u .

Note that this system is completely controllable. 

The weak controlled invariant subspaces can be characterized in the following

way.

Proposition 8.9 The following statements are equivalent.

(i) V is a weak controlled invariant;
(ii) AV ⊆ V + im B;
(iii) there exists a matrix F with n columns and m rows such that (A + B F)V ⊆ V .

Proof First we prove that (i) =⇒ (ii). Let x0 ∈ V and let u(t) : [0, +∞) → Rm be
an input such that x(t, x0 , u(·)) ∈ V for each t ≥ 0. Without loss of generality, we
can extend continuously u(t) on a small interval (−ε, 0), so that x(t, x0 , u(·)) can
be considered of class C 1 at t = 0. Then

x(t, x0 , u(·)) − x0
lim = ẋ(0) ∈ V
t→0+ t

that is Ax0 + Bu(0) ∈ V , or Ax0 ∈ V − Bu(0).

Next we prove that (ii) =⇒ (iii). Let dim V = k ≤ n. Let e1 , . . . , en be a basis
of Rn , such that its first k elements e1 , . . . , ek constitute a basis of V . Then for each
i = 1, . . . , k one has Aei = gi + Bu i for some gi ∈ V and some u i ∈ Rm . Let us
chose other vectors u k+1 , . . . , u n ∈ Rm in arbitrary way, and define the matrix F by
the relations Fe j = −u j , for j = 1, . . . , n. Then we have, for i = 1, . . . , k,

(A + B F)ei = Aei + B Fei = gi + Bu i − Bu i = gi ∈ V .

Finally we prove that (iii) =⇒ (i). Let x0 ∈ V and let x(t) be the solution of the
closed loop system
164 8 Frequency Domain Approach

ẋ = (A + B F)x
(8.31)
x(0) = x0 .

Of course, x(t) is also a solution of the problem


ẋ = Ax + Bu(t)
x(0) = x0

where u(t) = F x(t). The proof is completed, by noticing that V is a dynamic

invariant with respect to system (8.31). 

We are finally able to state the main result of this section.

Theorem 8.6 System (8.29) can be rendered disturbance decoupled by means of

a linear feedback if and only if there exists a subspace V ⊆ Rn which is a weak
controlled invariant for the system (8.30) and such that im G ⊆ V ⊆ ker C.

Proof Let us prove first the necessary part. So let F be a matrix such that the system

ẋ = (A + B F)x + Gd
(8.32)
y = Cx

is disturbance decoupled. According to Theorem 8.4, there exists an algebraic invari-

ant subspace V , such that im G ⊆ V ⊆ ker C. This implies that (A + B F)V ⊆ V ,
and this in turn means that V is a weak controlled invariant, by Proposition 8.9.
Then we prove the sufficient part. If V is weak controlled invariant, then by
Proposition 8.9 there exists F such that (A + B F)V ⊆ V . Together with the inclu-
sions im G ⊆ V ⊆ ker C, this implies finally that the system (8.32) is disturbance
decoupled by Theorem 8.4. 

Chapter Summary

The subject of the last chapter is the relationship between two possible approaches
to the analysis of a system: the time domain approach (developed in the previous
chapters) and the more traditional frequency domain approach based on the Laplace
transform. We study in particular the realization problem. For the case of SISO sys-
tems, we also give a different solution to the stabilization problem by output feedback.
Finally, we illustrate the decoupling problem, whose solution takes advantages of
both approaches.
Appendix A
Internal Stability Notions

The purpose of this appendix is to introduce the basic properties arising in the
characterization of the long-term qualitative behavior of solutions of unforced,
time invariant differential systems. Notation and terminology are those of the Intro-
duction (Chap. 1). However, since the interest is focused on the state variable, in this
appendix the observation map is ignored. From the mathematical point of view, the
systems considered in this chapter reduce therefore to systems of ordinary differential
equations (in general, nonlinear)1

ẋ = f (x) (A.1)

where x ∈ Rn . Recall that a solution of (A.1) is any differentiable function x = ϕ(t)

defined on some interval I ⊆ R such that ϕ̇(t) = f (ϕ(t)) for each t ∈ I . We will
assume that the function f in (A.1) is defined and continuous together with its
first partial derivatives, for each x ∈ Rn ; moreover, we assume that it satisfies the
inequality
|| f (x)|| ≤ a||x|| + b

for some positive constants a, b. Under these assumptions, for each initial pair (t0 , x0 )
existence and uniqueness of solutions are guaranteed, and we may further take I = R
without loss of generality [24]. Moreover, since the function f does not depend
explicitly on t, according to Proposition 1.9, the system (A.1) is time invariant;
therefore it is not restrictive to assume t0 = 0.
The notions introduced in this appendix are often referred to as internal stability
notions, in order to emphasize the difference with the notion of external stability
introduced in Chap. 1 and studied in detail in Chap. 6.

1 Thenotions we are going to introduce are applied in this book essentially for the case of linear
systems; however, they can be better understood when referred to a general system of the type (A.1).
© Springer Nature Switzerland AG 2019 165
A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5
166 Appendix A: Internal Stability Notions

A.1 The Flow Map

A solution of (A.1) can be regarded as a parameterized curve x = ϕ(t) of Rn . For

each t ∈ R, the tangent vector to such a curve at the point x coincides with f (x). For
this reason, the function f : Rn → Rn which defines (A.1) is also called a vector
field.
The image of a solution x = ϕ(t) of (A.1) is called an orbit or a trajectory. It is
important do not confuse the graph of a solution ϕ(t), which is a subset of R×Rn ,
with the orbit of ϕ(t), which coincides with the set ϕ(R) and it is a subset of Rn . We
may also view the orbit of ϕ as the orthogonal projection of the graph of ϕ on Rn ,
along the time axis (see Fig. A.1).
We already mentioned that under the stated assumptions, for each initial condi-
tion (A.1) has a unique solution. In order to emphasize its global validity, we my
reformulate this property writing that if x = ϕ(t) and x = ψ(t) are two arbitrary
solutions of (A.1), then

∃t¯ : ϕ(t¯) = ψ(t¯) =⇒ ϕ(t) = ψ(t) ∀t ∈ R . (A.2)

The geometric interpretation of (A.2) is that if the graphs of the two solutions
have a common point, then they must coincide. We may also interpret the time
invariance property from a geometrical point of view: the time translation of the
graph of a solution is again the graph of a (in general, different) solution. All the
solutions obtained as time translation of a fixed solution obviously are equivalent
parametrization of the same curve, and so they define the same orbit (see again
Fig. A.1). This fact admits a converse.

Fig. A.1 Two solutions and

their projections

10
y
5

0
t

−5

−10
10 x
5
20
0
15
−5 10
5
−10 0
Appendix A: Internal Stability Notions 167

Lemma A.1 Let ϕ(t) and ψ(t) be two arbitrary solutions of (A.1) defined for each
t ∈ R. Then,

∃t1 , t2 : ϕ(t1 ) = ψ(t2 ) =⇒ ψ(t) = ϕ(t + T ) ∀t ∈ R , (A.3)

where we set T = t1 − t2 .

Proof Let T = t1 − t2 and χ(t) = ϕ(t + T ). Clearly, χ(t) is a solution. It satisfies

the initial condition
χ(t2 ) = ϕ(t2 + t1 − t2 ) = ϕ(t1 ) .

But also ψ(t) is a solution which, by hypothesis, satisfies the same condition.
Because of the uniqueness property, we have

χ(t) = ϕ(t + T ) = ψ(t) ∀t ∈ R .

The meaning of Lemma A.1 is that if two orbits have a common point, then
they must coincide (the reader is warned to notice the difference between (A.2) and
(A.3)). In other words, there is a unique orbit passing through every point of Rn .
The orbits of the system (A.1) fill the space and are displayed in such a way to form
a partition of the space. We might define an equivalence relation, saying that two
points are equivalent when they lie on the same orbit. In the particular case n = 2,
we can image that the orbits form a picture in the plane. This picture is also called
state configuration or phase portrait. To denote the solution of the Cauchy problem

ẋ = f (x)
(A.4)
x(0) = x0

we use the notation2

x = x(t, x0 ) (A.5)

which has the advantage of emphasizing, beside the time variable t, also the initial
state x0 . Equation (A.5) define a function from R × Rn to Rn : this is called the flow
map generated by the vector field f . It can be interpreted as a function of t for each
fixed x0 , or as a function from Rn to Rn , parameterized by t.

Remark A.1 In (A.5), the variable t should be thought of not as the indication of a
precise instant of time, but rather as the indication of the duration of a process, that
is the length of the time interval needed to transfer the state of the system from x0 to
x(t, x0 ). 

2 Note that (A.5) is nothing else than (1.12) adapted to the case of (A.1).
168 Appendix A: Internal Stability Notions

Proposition A.1 The flow map of the vector field f satisfies the following properties:

x(0, x0 ) = x0 (A.6)

x(t, x(τ , x0 )) = x(t + τ , x0 ) (A.7)

for each t, τ ∈ R and x0 ∈ Rn .

A.2 Equilibrium Points and Stability in Lyapunov Sense

Roughly speaking, internal stability means that in the absence of external energy
supply, the state of a system evolves remaining in a neighborhood of a rest point, and
eventually approaches a rest point.
Let the unforced, time invariant differential system (A.1) be given. We say that x̄ ∈
Rn is an equilibrium point if the constant function ϕ(t) ≡ x̄ is a solution. Sometimes,
equilibrium points are also called rest or singular, or even critical points. If x̄ is an
equilibrium point, then the orbit issuing from x̄ reduces to the singleton {x̄}.

Proposition A.2 The point x̄ is an equilibrium point if and only if f (x̄) = 0.

We say that the equilibrium point x̄ is isolated if there exists a neighborhood O

of x̄ such that f (x)
= 0 for each x ∈ O, x
= x̄.

Definition A.1 Let x̄ be an equilibrium point. We say that x̄ is stable (in Lyapunov
sense) for the system (A.1) if for each ε > 0 there exists δ > 0 such that

x0 − x̄ < δ =⇒ x(t; x0 ) − x̄ < ε, ∀t ≥ 0.

Definition A.2 Let x̄ be an equilibrium point. We say that x̄ is attractive if there

exists δ0 > 0 such that, for each initial state x0 for which x0 − x̄ < δ0 , one has

lim x(t; x0 ) = x̄ . (A.8)

t→+∞

Definition A.3 Let x̄ be an equilibrium point. If x̄ is both stable and attractive, we

say that it is asymptotically stable. Moreover, if (A.8) holds for each x0 ∈ Rn , x̄ is
called globally asymptotically stable. Finally, if the solutions converge to x̄ with an
exponential decay i.e., there exist M > 0, α > 0 such that

∀x0 with x0 − x̄ < δ0 we have x(t; x0 ) − x̄ ≤ Me−αt , (A.9)

we speak about exponential stability. The supremum of the numbers α such that
(A.9) holds for some suitable M, is called the decay rate.
Appendix A: Internal Stability Notions 169

We end this chapter by the following notion, very useful in the analysis of the
qualitative behavior of the unforced system (A.1).

Definition A.4 Let K be a closed subset of Rn . We say that K is dynamically

invariant for the system (A.1) if for each x0 ∈ K one has x(t; x0 ) ∈ K for every
t ∈ R.

Appendix Summary
Appendix A recalls some mathematical definitions concerning stability. The informal
term “stability” actually involves the notion of stability in the sense of Lyapunov and
the notion of attraction. In general, these notions are mutually independent, but in the
case of linear systems the latter implies the former. Moreover, in the case of linear
systems there is no way to distinguish the local and global aspects. For these reasons,
in this Appendix, and only in this Appendix, we refer to general (nonlinear) systems
of ordinary differential equations.
Appendix B
Laplace Transform

In this appendix we recall some basic facts about Laplace transform, that are needed
for the applications considered in this book. In view of our limited goals and for
sake of simplicity, the subject will not be treated with the maximal generality and
mathematical rigor. In particular, the Dirac-delta function and its Laplace transform
will be introduced only at heuristic level. For a more formal presentation, the reader
can be addressed to one of many existing books on this topic, for instance [15].

B.1 Definition and Main Properties

Let f : [0, +∞) → R be a piecewise continuous function.

Definition B.1 We say that f is a subexponential function if there exist real constants
M > 0 and α such that

| f (t)| ≤ Meαt ∀t ∈ [0, +∞) . (B.1)

To each piecewise continuous, subexponential function we can associate a real

number σ0 , defined as the infimum of the numbers α for which there exists M such
that (B.1) holds. This number σ0 is called the order of f .

Lemma B.1 Let f be a subexponential, piecewise continuous function, and let s be

any complex number such that Re s > σ0 . Then,

lim f (ξ)e−sξ = 0 .
ξ→+∞

Proof If H (ξ) is a complex function of one real variable, lim H (ξ) = 0 is equiv-
ξ→+∞
alent to
lim Re H (ξ) = lim Im H (ξ) = 0 .
ξ→+∞ ξ→+∞

© Springer Nature Switzerland AG 2019 171

A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5
172 Appendix B: Laplace Transform

But for each z ∈ C, we have

|Re z| ≤ |z| and |Im z| ≤ |z| .

Thus, it is sufficient to prove that

lim | f (ξ)e−sξ | = lim | f (ξ)|e−ξRe s = 0 .

ξ→+∞ ξ→+∞

Let α be a real number such that σ0 < α < Re s. We have

| f (ξ)|e−ξRe s ≤ Meξ(α−Re s)

and the statement follows since α − Re s < 0.


The Laplace transform allows us to associate a function F : C → C to each piece-
wise continuous, subexponential function f : [0, +∞) → R. Before giving the for-
mal definition, we still need some preliminary results.

Lemma B.2 Let f be a piecewise continuous, subexponential function, whose order

is σ0 . For each complex number s such that Re s > σ0 , the improper integral
 +∞
f (t)e−st dt
0

is absolutely convergent.
 +∞
Proof The absolute convergence of the improper integral H (t)dt of a function
0
H : R → C is equivalent to the convergence of both the integrals
 +∞  +∞
|Re H (t)| dt and |Im H (t)| dt . (B.2)
0 0

As already noticed in the proof of Lemma B.1, it is therefore sufficient to show

that the integral  +∞
|H (t)|dt
0

is convergent. In our case,

| f (t)e−st | = | f (t)||e−st | = | f (t)|e−tRe s ≤ Met (α−Re s) .

According to the definition of σ0 , we can chose α in such a way that σ0 < α <
Re s, so that α − Re s < 0. The convergence of the two integrals (B.2) is guaranteed
by comparison. Notice that if we set s = σ + i ω, we have
Appendix B: Laplace Transform 173
 +∞  +∞
|Re ( f (t)e−st )|dt = | f (t)e−tσ cos(−ωt)|dt
0 0

and  +∞  +∞
−st
|Im ( f (t)e )|dt = | f (t)e−tσ sin(−ωt)|dt .
0 0


We are finally ready to introduce the main definition of this Appendix.
Definition B.2 Let f be a piecewise continuous, subexponential function, defined
in [0, +∞), whose order is σ0 . The Laplace transform of f is the complex function
 +∞
s → F(s) = f (t)e−st dt (B.3)
0

defined on the domain {s ∈ C : Re s > σ0 }.

It is convenient to remark that F could be coincident with the restriction to the half-
plane {s ∈ C : Re s > σ0 } of a function F̃ : C → C defined in a broader domain
(to this respect, see Remark B.1). We also remark that (B.3) is meaningful even if f
is a complex function of one real variable.
The operator defined by (B.3) will be denoted by the symbol L. We will also
agree to denote by the same letter (respectively, small and capital) the function to be
transformed and its Laplace transform. Hence, we write
 +∞
F(s) = L[ f (t)] = f (t)e−st dt.
0

Next we review some important properties of the Laplace transform.

Property 1 (Linearity) Let f and g be two piecewise continuous, subexponential
functions, defined on the interval [0, +∞), of order respectively σ1 and σ2 . Then, for
each a, b ∈ R the function a f + bg is subexponential of order σ0 = max{σ1 , σ2 }.
Moreover,
a F(s) + bG(s) = L[a f (t) + bg(t)] (B.4)

for Re s > σ0 , where F(s) = L[ f (t)] and G(s) = L[g(t)].

Proof It is immediate to check that a f + bg is subexponential (take a value of α
greater than both σ1 and σ2 ). Formula (B.4) is a trivial consequence of the properties
of the integrals.

Property 2 (Rescaling) Let f be a piecewise continuous, subexponential function,
defined in [0, +∞), of order σ0 , and let F(s) = L[ f (t)] for Re s > σ0 . Then, for
each a > 0, g(t) = f (at) is a subexponential function of order aσ0 , and
174 Appendix B: Laplace Transform

1 s 
L[ f (at)] = F f or Re s > aσ0 . (B.5)
a a

Proof From (B.1) we have easily | f (at)| ≤ Meαat = Meβt for β = αa > aσ0 . Set-
ting τ = at, we therefore have
 +∞  +∞  +∞
−st 1 − as τ 1 1
f (at)e dt = f (τ )e dτ = f (τ )e−r τ dτ = F(r )
0 a 0 a 0 a

provided that Re r > σ0 , where r = s/a. But Re r = Re (s/a) = (Re s)/a, and hence
requiring Re r > σ0 it is equivalent to require Re s > aσ0 .

Property 3 (Right translation) Let f be a piecewise continuous, subexponential
function, defined in [0, +∞), of order σ0 , and let F(s) = L[ f (t)] for Re s > σ0 . In
addition, let 
0 for 0 ≤ t ≤ c
g(t) =
f (t − c) for t > c ,

where c > 0. Then g is a subexponential function of order σ0 , and L[g(t)] =

e−cs F(s) for Re s > σ0 .
Proof The reader can easily check that g is a subexponential function. Moreover, by
the definition of g, we have
 +∞  +∞
g(t)e−st dt = f (t − c)e−st dt .
0 c

Finally, the substitution τ = t − c yields

 +∞  +∞
−st
g(t)e dt = f (τ )e−s(τ +c) dτ = e−cs F(s).
0 0

n
Property 4 (Multiplication by t ) Let f be a piecewise continuous, subexponential
function, defined in [0, +∞), of order σ0 , and let F(s) = L[ f (t)] for Re s > σ0 . For
each n ∈ N, the function t n f (t) is a subexponential function of order σ0 , and

d n F(s)
L[t n f (t)] = (−1)n , Re s > σ0 . (B.6)
ds n

Proof The subexponential property of t n f (t) is a consequence of Lemma 2.1. For-

mula (B.6) can be proved by induction. The case n = 0 follows immediately by
definition. Assuming that (B.6) holds for n = k, the case n = k + 1 can be obtained
by computing the derivative3 of both sides with respect to s.


3 Here and in other following proofs, the crucial point consists in exchanging the order of certain
operations like limits, derivatives, integrals. The correctness of such exchanges requires some uni-
Appendix B: Laplace Transform 175

Property 5 (Multiplication by eat ) Let f be a piecewise continuous, subexponential

function, defined in [0, +∞), of order σ0 , and let F(s) = L[ f (t)] for Re s > σ0 . Let
moreover a ∈ R. Then, eat f (t) is a subexponential function of order σ0 + a, and

L[eat f (t)] = F(s − a) Re s > σ0 + a . (B.7)

Proof It is easy to check that eat f (t) is subexponential. Moreover,

 +∞  +∞  +∞
−st −t (s−a)
e f (t)e
at
dt = f (t)e dt = f (t)e−tr dt = F(r )
0 0 0

with Re r > σ0 , where r = s − a. But Re r = Re s − a and so Re r > σ0 is equiva-

lent to Re s > σ0 + a.

Property 5 extends in the case a ∈ C, with Re s > σ0 + Re a.
Next properties are the most important from our point of view, since they refer to
the behavior of the operator L with respect to the operations of the differential and
integral calculus.
Property 6 Assume that f is a piecewise continuous, subexponential function
defined in [0, +∞), of order σ0 . Assume further that its derivative f  exists and
it is piecewise continuous in [0, +∞). If F(s) = L[ f (t)] for Re s > σ0 , then the
Laplace transform of f  exists, and it is given by

L[ f  (t)] = − f (0) + s F(s) f or Re s > σ0 . (B.8)

Proof We want to compute

 ξ
lim f  (t)e−st dt .
ξ→+∞ 0

Integrating by parts we have

 
ξ
ξ  ξ

f (t)e −st
dt = f (t)e−st 0 +s f (t)e−st dt
0 0
 ξ
−sξ
= f (ξ)e − f (0) + s f (t)e−st dt .
0

The statement is proved, taking the limit for ξ → +∞, and taking into account
Lemma B.1.


formity assumptions, which are not difficult to ensure when we work with continuous functions
defined on compact intervals. In our framework (complex variables, unbounded intervals) there are
some additional technical difficulties. We do not enter in these details..
176 Appendix B: Laplace Transform

Property 7 Let f be a piecewise continuous, subexponential function, defined in

[0, +∞), of order σ0 , and let F(s) = L[ f (t)] for Re s > σ0 . Then, each antideriva-
tive of f is subexponential of order max{σ0 , 0} and we have
 t 
F(s)
L f (ρ) dρ = f or Re s > max{σ0 , 0} . (B.9)
0 s

Proof The first statement is left as an exercise. As far as (B.9) is concerned, we can
apply again the integration by part rule:
 ξ  t
 ξ
−st
f (ρ) dρ e dt = h(t)e−st dt
0 0 0

h(t)e−st ξ 1 ξ 
= + h (t)e−st dt
−s 0 s 0

h(ξ)e−sξ h(0) 1 ξ
= − + f (t)e−st dt .
−s −s s 0

Noticing that h(0) = 0, the conclusion follows by taking the limit for ξ → +∞.


Obviously, (B.8) and (B.9) can be iterated, which gives:

L[ f (k) (t)] = − f (k−1) (0) − s f (k−2) (0) − · · · − s k−1 f (0) + s k F(s) , (B.10)
 t  t1  tk−1 
F(s)
L ... f (tk ) dtk dtk−1 . . . dt1 = . (B.11)
0 0 0 sk

Property 8 Let f and g be two piecewise continuous, subexponential functions,

defined in [0, +∞), of order respectively σ1 and σ2 . Let F(s) = L[ f (t)] for Re s >
σ1 and G(s) = L[g(t)] for Re s > σ2 . Let moreover
 t
h(t) = f (t − ρ)g(ρ) dρ . (B.12)
0

Then, h is a subexponential function, and L[h(t)] = F(s)G(s) for Re s >

max{σ1 , σ2 }.

Property 8 answers the question of finding a function h(t) such that L[h(t)] =
F(s)G(s), assuming that F(s) = L[ f (t)] and G(s) = L[g(t)] are known. We
remark that (B.12) is well defined, since for ρ ∈ [0, t] we have t − ρ ≥ 0. Intro-
ducing the following extensions of the functions f and g:
 
0 if t < 0 0 if t < 0
f˜(t) = and g̃(t) =
f (t) if t ≥ 0 g(t) if t ≥ 0 ,
Appendix B: Laplace Transform 177

we can write
 t  +∞
h(t) = f (t − ρ)g(ρ) dρ = f˜(t − ρ)g̃(ρ) dρ .
0 −∞

Now let p(·) and q(·) be two piecewise continuous arbitrary functions defined on
the whole of R. The convolution between p and q is defined by
 +∞
( p ∗ q)(t) = p(t − ρ)q(ρ) dρ ,
−∞

provided that the integral is convergent. Thus, we may reformulate (B.12) by writing

L[( f˜ ∗ g̃)(t)] = F(s)G(s) .

Proof of Property 8 To prove that h(t) is a subexponential function of order

max{σ1, σ2 } is a simple exercise. With the notation above, we may also write
+∞
h(t) = f˜(t − ρ)g(ρ) dρ. Thus
0

 +∞  +∞

L[h(t)] = ˜
f (t − ρ)g(ρ) dρ e−st dt
0 0
 +∞  +∞

= g(ρ) f˜(t − ρ)e−st dt dρ .

0 0

By virtue of Property 3, we finally get

 +∞  +∞
−ρs
L[h(t)] = g(ρ)e F(s) dρ = F(s) g(ρ)e−ρs dρ = F(s)G(s) .
0 0

B.2 A List of Laplace Transforms

B.2.1 Elementary Functions

We now compute the Laplace transform of some elementary functions.

Proposition B.1 Let f (t) ≡ 1 for t ≥ 0. Then

1
L[ f (t)] = f or Re s > 0 .
s
178 Appendix B: Laplace Transform

Proof We have
 
+∞
−st
ξ
est ξ e−sξ 1
L[ f (t)] = e dt = lim e−st dt = lim = lim + .
0 ξ→+∞ 0 ξ→+∞ −s 0 ξ→+∞ −s s

The conclusion follows from the remark that if Re s > 0, then we have
lim e−sξ = 0.
ξ→+∞


Proposition B.2 The Laplace transform of the restrictions to the interval [0, +∞)
of the power functions, the exponential function, and the trigonometric functions are
given by:
a
L[at] = 2 f or Re s > 0 ; (B.13)
s

n!
L[t n ] = f or Re s > 0 (n ∈ N) ; (B.14)
s n+1
1
L[eat ] = f or Re s > a ; (B.15)
s−a
s
L[cos ωt] = f or Re s > 0 ; (B.16)
s 2 + ω2
ω
L[sin ωt] = f or Re s > 0 . (B.17)
s 2 + ω2

Proof We will prove formulæ (B.13), (B.14) and (B.15) as applications of Proposi-
tion B.1 and Properties 1, 2, 4 and 5. We begin with (B.13). We have

L[at] = aL[t] = aL[t · 1] .

By applying Property 4 and recalling that L[1] = 1/s, we conclude

a
L[at] = .
s2
Now consider formula (B.14). We have

d n F(s) 1
L[t n ] = L[t n · 1] = (−1)n where F(s) = .
ds n s

From this, by mathematical induction, we get F (n) (s) = (−1)n n!s −(n+1) . As far
as (B.15) is concerned, we just need to remark that

1
L[eat ] = L[eat · 1] = .
s−a
Appendix B: Laplace Transform 179

Now we consider the Laplace transform of the trigonometric functions. As already

noticed, the operator L applies to functions f : R → C, as well. Hence we can
compute L[cos ωt] making use of the Euler formula

ei ωt + e−i ωt
cos ωt = .
2
We have

1  i ωt  1 1 1
L[cos ωt] = L[e ] + L[e−i ωt ] = +
2 2 s +iω s −iω

for Re s > 0. This yields

s
L[cos ωt] = .
s2 + ω2

The proof of (B.17) is similar. 

B.2.2 Discontinuous functions

From Propositions B.1 and B.2 it is possible to deduce the Laplace transform of some
functions which are commonly used in signal theory. For instance,

0 if t < 0
U (t) =
1 if t ≥ 0

is called the unit step or also the Heaviside function. It represents a signal which
instantaneously jumps from zero to 1 (switch-on). The function f (t) ≡ 1 considered
in Proposition B.1 coincides with the restriction of U (t) to [0, +∞). Taking into
account the definition of Laplace transform, with a little abuse of notation we will
write
1
L(U (t)) = (Res > 0) .
s
Remark B.1 Let us remark that the complex function of a complex variable which
associates s to its inverse 1/s is defined for each s
= 0. Nevertheless, it is not correct
to say that such a function is the Laplace transform of U (t). Indeed the identity
L[U (t)] = 1/s holds only for Re s > 0. In other words, L[U (t)] coincides with the
restriction to the positive complex half plane of the function 1/s.


The function U (t) allows us to represent other types of discontinuous signals,
whose Laplace transform can be easily computed by applying Properties 1 and 3.
For instance,
180 Appendix B: Laplace Transform

(i) switch-on for t = c 

0 for t < c
f (t) =
1 for t ≥ c

is equivalent to f (t) = U (t − c);

(ii) switch-off for t = c 
1 for t ≤ c
f (t) =
0 for t > 0

is equivalent to f (t) = 1 − U (t − c) (or = U (c − t));

(iii) rectangular impulse ⎧
⎪0 for t < a
⎨
f (t) = 1 for a ≤ t ≤ b
⎪
⎩
0 for t > b

is equivalent to f (t) = U (t − a) − U (t − b).

The function U (t) is also useful to represent piecewise elementary functions. For
instance the function 
t for t < 1
f (t) = 2
t for t ≥ 1

can be written as

f (t) = t[1 − U (t − 1)] + t 2 U (t − 1) .

With the same abuse of notation as above, we can think of (B.13) as the Laplace
transform of a signal of the form

0 for t < 0
f (t) =
at for t ≥ 0

while (B.16), (B.17) provide the Laplace transform of signal of sinusoidal shape (but
vanishing for t < 0).

B.2.3 Dirac Delta Function

One of the most important signals typically employed in system theory is the unit
impulse function, denoted by the symbol δ(t) and also called Dirac δ function.
Appendix B: Laplace Transform 181

+∞ if t = 0
δ(t) = (B.18)
0 if t
= 0 .

This definition is an ideal representation of a signal of very large energy, concen-

trated at a single point. Of course, (B.18) is nonsense from a rigorous point of view,
since it really does not define a function R → R. There is a theory, called distribu-
tion theory, based on a generalization of the notion of function, which allows us to
formally introduce and study objects like (B.18). To our purposes, it is sufficient to
think of the function δ(t) as the limit of suitable sequences; for instance

1
δ(t) = lim+ (U (t + ε) − U (t − ε)) . (B.19)
ε→0 2ε

From (B.19) we infer in particular that

 +∞  k
δ(t) dt = δ(t) dt = 1 ∀k > 0 . (B.20)
−∞ −k

It is possible to define the sum and the multiplication between generalized func-
tions like δ(t). It is also possible to give a sense to certain operators of the differential
calculus for generalized functions, but this is not required in this book. We limit our-
selves to recall some facts and properties related to the Dirac delta function.
An impulse of intensity k concentrated at a point a ∈ R is represented by k · δ
(t − a). We have  +∞
k · δ(t − a)dt = k (B.21)
−∞

and  +∞
f (t)δ(t − a)dt = f (a) (B.22)
−∞

provided that the function f is continuous at the point t = a. Finally,

 t
δ(τ )dτ = U (t) (B.23)
−∞

and
L[δ(t)] = 1 . (B.24)

Formulæ (B.20)–(B.24) can be formally proved in the context of distribution the-

ory. They can be also justified heuristically on the base of (B.19). For instance,
concerning (B.24) we suggest the following argument.
182 Appendix B: Laplace Transform
 
1 
L[δ(t)] = L lim+ U (t) − U (t − ε)
ε→0 ε
 
1 1 1 e−εs
= lim+ {L[U (t)] − L[U (t − ε)]} = lim+ −
ε→0 ε ε→0 ε s s
−εs −εs
1 1−e e −1
= lim+ · = − lim+ =1.
ε→0 s ε ε→0 εs

B.3 Inverse Transform

The following proposition guarantees the existence of the inverse transformation

L−1 .
Proposition B.3 Let f and g be two continuous functions defined on [0, +∞), and
let F(s) = L[ f (t)], G(s) = L[g(t)], both defined for Re s > σ0 . If F(s) = G(s) for
all s such that Re s > σ0 , then f (t) = g(t) for all t ≥ 0.
This proposition states that f (t) can be uniquely reconstructed from F(s), and so
it allows us to define L−1 . However, it should be noted that if f (t) is the restriction to
[0, +∞) of a function ϕ(t) defined on a larger interval [a, +∞) with −∞ ≤ a < 0,
in general it is not possible to reconstruct ϕ(t) on the interval (a, 0) by applying L−1
to F(s). Nevertheless, in many applications this problem can be overcome if it is
known a priori that f (t) is real analytic.
The inverse transformation L−1 can be explicitly represented by a suitable for-
mula, but the use of this formula is not needed in this book. We limit ourselves to
remark that L−1 is, like L, a linear operator.

B.4 The Laplace Transform of a Vector Function

The extension of the Laplace transform to vector functions

f : [0, +∞) → Rn or f : [0, +∞) → Cn

where f = ( f 1 , . . . , f n ), is straightforward: under the assumption that each compo-

nent is a subexponential function, we set L[ f ] = (L[ f 1 ], . . . , L[ f n ]).
The aforementioned properties of the Laplace transform can be easily extended,
as well. In addition, we have

L[M f (t)] = ML[ f (t)] (B.25)

for each matrix M with real or complex constant entries. We are especially interested
in the transform of the exponential matrix and in the convolution product formula
(Property 8).
Appendix B: Laplace Transform 183

Proposition B.4 Let A be a square matrix (real or complex), c be a constant vector

and b(t) be a vector function whose components are piecewise continuous, subex-
ponential functions. Let σ0 be the maximal real part of the eigenvalues of A. Then,
for each s ∈ C such that Re s > σ0 , we have:

L[et A c] = −(A − s I )−1 c , (B.26)

 t 
(t−τ )A
L e b(τ )dτ = −(A − s I )−1 B(s) . (B.27)
0

Proof By definition,
 +∞  +∞  ξ
−st t A t (A−s I )
L[e c] =
tA
e e cdt = e c dt = lim et (A−s I ) c dt .
0 0 ξ→∞ 0

If Re s > σ0 , then det (A − s I )

= 0 and the inverse (A − s I )−1 exists. On the
other hand, it is well known that for each square matrix M, the exponential matrix et M
admits
 t M a derivative and (et M ) = Met M = et M M. This implies that if M is invertible,
e = M e = et M M −1 . We can therefore proceed in the following way:
−1 t M

ξ

L[et A c] = lim (A − s I )−1 et (A−s I ) c
ξ→∞ 0
 −1 ξ(A−s I )

= lim (A − s I ) e c − (A − s I )−1 c .
ξ→∞

The assumption that Re s > σ0 also implies that all the eigenvalues of A − s I
have negative real part. Indeed, it is clear that the eigenvalues μ of A − s I have the
form μ = λ − s where λ is an eigenvalue of A . But then Re μ = Re λ − Re s < 0.
We know that if all the eigenvalues of a matrix M have negative real part, then
for each c we have lim eξ M c = 0. In conclusion,
ξ→+∞

L[et A c] = −(A − s I )−1 c

as required. As far as (B.27) is concerned, we remark that

 t   +∞  t

L e(t−τ )A b(τ )dτ = e−st e(t−τ )A b(τ )dτ dt

0 0 0
 +∞  +∞

−τ A
= e et (A−s I ) dt b(τ ) dτ .
0 τ

Note the change of the integration interval due to the change of integration order.
Making use of the assumption that Re s > σ0 , we finally conclude
184 Appendix B: Laplace Transform
 t   +∞
L e(t−τ )A b(τ )dτ = − e−τ A (A − s I )−1 eτ (A−s I ) b(τ )dτ
0 0
 +∞
=− (A − s I )−1 e−sτ b(τ )dτ
0
 +∞
= −(A − s I )−1 e−sτ b(τ )dτ
0
= −(A − s I )−1 B(s) .

Appendix Summary
Appendix B recalls the definition of Laplace transform and its main properties.
Moreover, we give a list of the Laplace transforms of some elementary functions.

References

1. Ahlfors LV (1979) Complex analysis: an introduction to the theory of analytic functions of one
complex variable. McGraw-Hill, New York
2. Bacciotti A, Rosier L (2005) Liapunov functions and stability in control theory. Springer, Berlin
3. Barnett S (1975) Introduction to mathematical control theory. Clarendon Press, Oxford
4. Bernstein DS (2005) Matrix mathematics: theory facts and formulas with application to linear
systems theory. Princeton University Press, Princeton
5. Bressan A, Piccoli B (2007) An introduction to the mathematical theory of control. American
Institute of Mathematical Sciences, Springfields
6. Brockett RW (1970) Finite dimensional linear systems. Wiley, New York
7. Brockett R (2014) The early days of geometric nonlinear control. Automatica 50(9):2203–2224
8. Callier FM, Desoer CA (1982) Multivariable feedback systems. Springer, New York
9. Cao YY, Lam J, Sun YX (1998) Static output feedback stabilization: an ILMI approach. Auto-
matica 34(12):1641–1645
10. Coppel WA (1965) Stability and asymptotic behaviour of differential equations. Heath Math-
ematical Monographs, Boston
11. Conti R (1976) Linear differential equations and control. Academic Press, London
12. Conti R (1985) Processi di controllo lineari in R n , quaderni UMI 30. Pitagora Editrice, Bologna
(in Italian)
13. Coron JM (2007) Control and nonlinearity. American Mathematical Society, Providence
14. Dragan V, Halanay A (1999) Stabilization of linear systems. Birkhäuser, Boston
15. Guest PB (1991) Laplace transforms and an introduction to distributions. Ellis Horwood, New
York
16. Hermes H, LaSalle JP (1969) Functional analysis and time optimal control. Academic Press,
New York
17. Hirsch MW, Smale S (1974) Differential equations, dynamical systems and linear algebra.
Academic Press, New York
18. Isidori A (1989) Nonlinear control systems. Springer, Berlin
19. Isidori A (2017) Lectures in feedback design for multivariable systems. Springer
Appendix B: Laplace Transform 185

20. Jurdjevic V (1997) Geometric control theory. Cambridge University Press, Cambridge
21. Lang S (1966) Linear algebra. Addison Wesley, Reading Mass
22. LaSalle J (1986) The stability and control of discrete process. Springer, New York
23. Lee EB, Markus L (1967) Foundations of optimal control theory. Wiley, New York
24. Pontrjagin L (1975) Équations différentielles ordinaires. MIR, Moscou
25. Rosenbrock HH (1970) State-space and multivariable theory. Nelson, London
26. Sontag ED (1990) Mathematical control theory. Springer, New York
27. Terrel WJ (2009) Stability and stabilization: an introduction. Princeton press, Princeton
28. Trentelman HJ, Storvogel AA, Hautus M (2001) Control theory for linear systems. Springer,
London
29. Tsinias J, Kalouptsidis N (1990) Output feedback stabilization. IEEE Trans Auto Control
35(8):951–954
30. Vidyasagar M (1993) Nonlinear systems analysis. Prentice Hall, Englewood Cliffs
31. Wonham WM (1979) Linear multivariable control: a geometric approach. Springer, New York
32. Zadeh LA, Desoer CA (1963) Linear system theory. McGraw-Hill, New York
Index

A F
Accessibility, 94 Feedback
dynamic, 111
linear, 112
output, 7
C
state, 7
Cauchy problem, 17, 167
static, 111
Compensator, 6, 132
Flow map, 167
Connection
Formula
cascade, 4 variation of constants, 54
feedback, 5 Frequency, 38
parallel, 4 Function
Control, 3 Dirac delta, 8, 180
closed loop, 6 Heaviside, 179
open loop, 6 subexponential, 171
PID, 137 transfer, 64
Controllability, 70
asymptotic, 122
Controller, 6, 132 G
Convolution, 177 Gain, 6, 136, 157
Cyclic vector, 36 General integral, 22, 54

D H
Disturbance, 3 Hurwitz property, 44
Domain
frequency, 140
time, 140 I
Indistinguishability, 80
Initial condition, 13
Input, 2
E admissible, 2, 16
Equilibrium point, 168 Invariance
Equivalence algebraic, 78, 160
feedback, 112 controlled, 86
linear, 28, 86 dynamic, 160, 169
Exosystem, 55 strong controlled, 163
© Springer Nature Switzerland AG 2019 187
A. Bacciotti, Stability and Control of Linear Systems, Studies in Systems,
Decision and Control 185, https://doi.org/10.1007/978-3-030-02405-5
188 Index

weak controlled, 163 Plant, 6

Polynomial
coprime, 142
L Principle
Lagrange formula, 54 argument, 156
Lyapunov separation, 135
equation, 48 superposition, 54
function, 49 Property
stability, 168 detectability, 133
weak function, 49 pole assignment, 117

M R
Matrix Reachability, 70
companion, 35 along the free solution, 94
controllability, 77 global, 70
diagonalizable, 29 local, 94
exponential, 24 Realization, 145
fundamental, 23 canonical, 146
generalized inverse, 128 minimal, 146
Hurwitz, 44 Reference signal, 3
impulse response, 100
Resolvent, 141
indefinite, 47
Resonance, 56
Jordan form, 31
Riccati
Lyapunov equation, 48
algebraic equation, 121
positive definite, 47
positive semidefinite, 47
transfer, 141
S
Solution
N free or unforced, 70
Nyquist diagram, 156 fundamental set, 23
Space
reachable, 72
O unobsevable, 82
Observation map, 16 Stability
Observer asymptotic, 168
asymptotic, 134 BIBO, 5, 15, 98
Operator BIBS, 98
input-output, 2 exponential, 168
Orbit, 166 external, 5, 101, 165
Order finite gain, 98
of a function, 171 in Lyapunov sense, 168
Output, 2 internal, 15, 165, 168
uncorrupted, 159 State, 2
State equation, 16
Steady state, 58
P System, 1, 4
Part causal, 7
controllable, 88 closed loop, 112
observable, 89 completely controllable, 77
uncontrollable, 88 completely observable, 83
unobservable, 89 continuous time, 3
Phase portrait, 167 decoupled, 25
Index 189

deterministic, 13 zero dynamics, 128

differential, 16
discrete time, 3
disturbance decoupled, 159 T
dual, 85 Trajectory, 166
homogeneous associated, 53 Transient, 58
impulse response, 8
input-output, 3
linear, 8, 14 U
MIMO, 4 Undetermined coefficients, 56
SISO, 4 Uniform convergence, xvii
stabilizable, 117
superstabilizable, 117
time invariant, 7 V
unforced, 15 Vector field, 166