The Dynamic Morse Theory of Control Systems - Souza CAMBRIGE 2020

The Dynamic Morse
Theory of Control
Systems
The Dynamic Morse
Theory of Control
Systems
By
Josiney A. Souza
The Dynamic Morse Theory of Control Systems
By Josiney A. Souza
This book first published 2020
Cambridge Scholars Publishing
Lady Stephenson Library, Newcastle upon Tyne, NE6 2PA, UK
British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library
Copyright © 2020 by Josiney A. Souza
All rights for this book reserved. No part of this book may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording or otherwise, without
the prior permission of the copyright owner.
ISBN (10): 1-5275-4507-5

ISBN (13): 978-1-5275-4507-6
For Priscila, Larissa, and Henrique.
“The LORD by wisdom hath founded the earth; by understanding hath he
established the heavens. By his knowledge the depths are broken up, and the
clouds drop down the dew. My son, let not them depart from thine eyes; keep
sound wisdom and discretion: So shall they be life unto thy soul, and grace to
thy neck”. Proverbs 3,19-22
Contents
Preface xiii
Introduction 1
1 Fundamental Theory of Control systems 11

1.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Shift Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Control affine systems . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Control flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5 Group structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.7 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . 40
2 Elementary Concepts 41
2.1 Invariant sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2 Orbits and semi-orbits . . . . . . . . . . . . . . . . . . . . . . . 44
2.3 Critical points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.4 Periodic points . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3 Minimal sets and transitivity 69

3.1 Minimal sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2 Weak control sets . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.3 Dynamic order . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
ix
x Contents
4 Limit sets and prolongations 81

4.1 Limit sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 Prolongational limit sets . . . . . . . . . . . . . . . . . . . . . . 86
4.3 Minimal equistable sets . . . . . . . . . . . . . . . . . . . . . . 94
4.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5 Asymptotic transitivity 103

5.1 Holding sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Asymptotic transitive sets . . . . . . . . . . . . . . . . . . . . . 109
5.3 Poincaré recurrence theorem . . . . . . . . . . . . . . . . . . . . 118
5.4 Control sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.5 Invariant control sets . . . . . . . . . . . . . . . . . . . . . . . . 132
5.6 Control sets for linear control systems . . . . . . . . . . . . . . 136
5.7 Prolongational transitive sets . . . . . . . . . . . . . . . . . . . 146
5.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6 Attractors and repellers 169

6.1 General concepts of attraction . . . . . . . . . . . . . . . . . . . 170
6.2 Uniform attractor . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.3 Conley attractor . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.4 Global attractors . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
7 Chain transitivity 215

7.1 Chain transitive sets . . . . . . . . . . . . . . . . . . . . . . . . 215
7.2 Attractor-repeller pair . . . . . . . . . . . . . . . . . . . . . . . 230
7.3 The Conley main theorem . . . . . . . . . . . . . . . . . . . . . 240
7.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
8 Morse decompositions 251

8.1 Dynamic Morse decompositions . . . . . . . . . . . . . . . . . . 252
8.2 Attractor-repeller Morse decomposition . . . . . . . . . . . . . 257
8.3 Morse decomposition and chain recurrence . . . . . . . . . . . . 261
8.4 Complete Lyapunov functions . . . . . . . . . . . . . . . . . . . 265
Contents xi
8.5 Final comments and open questions . . . . . . . . . . . . . . . 271

8.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
A Dynamical systems 279

A.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
A.2 Morse decompositions . . . . . . . . . . . . . . . . . . . . . . . 285
A.3 Attractors and repellers . . . . . . . . . . . . . . . . . . . . . . 291
A.4 Chain transitive sets . . . . . . . . . . . . . . . . . . . . . . . . 300
A.5 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . 311
B Differentiable manifolds 313
Bibliography 319
Index 329
xii Contents
Preface
This book is designed for students or researchers who are interested in dynamic
concepts of control systems. It presents classical concepts of control theory
integrated with a report about ongoing research on Conley theory for control
systems. We assume the reader knows the rudiments of differential equations,
real analysis, and topology. For theoretical purposes, we consider control
systems on differentiable manifolds. In this setting, we may include many
interesting cases of control systems in non-Euclidean spaces, for instance, tori,
spheres, and projective spaces. Readers who are not familiar with manifolds
may consider Euclidean state space instead. Otherwise, a brief appendix on
differentiable manifolds is provided for elementary definitions, notations, and
references. An appendix on dynamical systems is provided for references to
results from the Conley theory of flows.
This volume expands on the material presented in A Course on Geometric
Control Theory: Transitivity and Minimal Sets [93]. Its contents includes a
presentation of the fundamental theory of control systems, an exposition of
elementary concepts of orbits, invariance, periodicity, and a broad discussion
on various aspects of transitivity and controllability. The main part deals
with attractors and repellers, Morse decompositions, and chain transitivity.
Most concepts presented are illustrated by means of comprehensive examples
and figures. A list of problems is given in each chapter with the intention
of reinforcing the reader’s grasp of the material, amplifying and completing
proofs, applying theorems, and enabling the reader to discover important
facts, examples, and counterexamples. Notes and references are included at
the end of each chapter to indicate results not discussed in the text, remarks,
references for further reading, and historical notes.
The main contribution of the present work is the improvement of results
in dynamic Morse theory for control systems, which are now integrated in
xiii
xiv Preface
this unique volume. The text adds to our knowledge of various dynamical
concepts which compose the full ingredients of the central notion of Morse
decomposition. Having read this book, the reader will have the opportunity
to expand Conley’s ideas by means of open questions on Morse decompositions
for control system on noncompact manifolds.
Parts of the book may be useful in courses or seminars in mathematics
as well as control-theoretic engineering. The material may also be used as
a reference source for various topics in control systems, and serve as a basic
reference for academic or research projects.
Thanks are given above all to God for life and for the sciences. Special
thanks are due to José V. de Souza and Rita C. A. de Souza for moral support;
to Prof. Carlos J. Braga Barros, his collaboration and the donation of the
Colonius–Kliemann book; and to Prof. Luiz A. B. San Martin, his instructions
and our discussions on control theory and transformation semigroups. We
acknowledge all our coworkers: Prof. Ronan A. Reis, Prof. Hélio V. M.
Tozatti, Prof. Victor H. L. Rocha, and Stephanie A. Raminelli. We also
give thanks to João A. N. Cossich for part of the material on differentiable
manifolds.
Maringá, Brazil
March 2019
Introduction
The global dynamical behavior of a compact system is described by Morse

decomposition. This was stated by Charles C. Conley (1933–1984) in his
famous work, “Isolated Invariant Sets and the Morse Index” ([36]), and is
currently one of the most important statements in the study of asymptotic
behavior of dynamical systems. A Morse decomposition contains essential
information about the long-term behavior of a system, since each state con-
verges in forward as well as in backward time to some Morse set. This is
due to the attractor-repeller configuration of a Morse decomposition, which
provides attractive and repulsive properties for its components. Besides, the
attractor-repeller configuration implies an order among the Morse sets such
that the flow can be interpreted via its movement from the Morse sets with
lower indices toward those with higher ones. This means that, outside a Morse
decomposition, every trajectory of the system comes near some Morse set in
forward time, and other distinct Morse set in backward time. Consequently,
the system does not admit asymptotic transitivity outside a Morse decompo-
sition, or, in other words, any asymptotic transitive set resides in some Morse
set. Thus periodicity and recurrence occur only in the components of a Morse
decomposition; outside them, however, the system is transient.
In general, the transient part of a dynamical system consists of all states
which are not chain recurrent. The chain recurrence is a general type of recur-
rence based on returning trajectories with jumps. This defines an
equivalence relation - the chain transitivity - whose equivalence classes lie
in the components of a Morse decomposition. In his report entitled “The
gradient structure of a flow” ([37]), Conley proved the Fundamental Theorem
of Dynamical Systems that any flow on a compact metric space decomposes
into the chain recurrent part and the transient or ‘gradient-like’ part. This
means that, if each equivalence class of chain transitivity is identified to a
1
2 Introduction
point, then the resulting flow has a Lyapunov function that decreases along
all trajectories except the fixed points. This was the reason for Conley coining
the term, the ‘Morse decomposition’ of a system. The central result in this
direction characterizes a system with finitely many chain equivalence classes
as a system admitting the finest Morse decomposition. Conley’s ideas about
Morse decomposition concentrated in the topological considerations of con-
necting orbits between the Morse sets. His later studies result in the notion of
a ‘connection matrix’ ([78]), with the main applications being for the theory
of shock waves ([38, 39, 40, 41]).
In the Conley global view, the fundamental elements of a dynamical system
are the isolated invariant sets. An invariant set is called ‘isolated’ if it is the
maximal invariant set in some neighborhood of itself. This extends the well-
known concept of isolated singularity. Attractors, repellers, and Morse sets
are the main isolated invariant sets of a system. The ‘Morse index’ of an
isolated invariant set is the homotopy type of an n-dimensional sphere, where
n is the dimension of the unstable manifold, the set of points along the flow
which are sent away from the isolated invariant set as time moves forward.
Conley expanded the ideas about the Morse index by introducing the notions
of isolating blocks. An ‘isolating block’ is a set whose boundary has no internal
tangencies to the flow, that is, if the flow is tangential to the boundary of the
block, then the trajectories leave the block both forward and backward in
time. The ‘exit set’ is the set of points on the boundary where the flow leaves
the block. On the one hand, every isolated invariant set can be surrounded
by an isolating block and, on the other hand, every isolating block contains
an isolated invariant set in its interior. The ‘Conley index’ is the homotopy
type of the isolating block with the exit set identified to a point. If an isolated
invariant set is a non-degenerate rest point for a smooth flow on a compact
manifold, then the Conley index and the Morse index coincide.
An isolating block is structurally stable in the sense that it persists under
perturbations of the system. Conley believed that this property of isolating
blocks meant that they were the only dynamical objects which could be de-
tected in nature and that their properties reflected the important properties
of natural systems. This inspires his interest to explore new areas of science
with the intention of applying his theories. The Conley index was applied
to prove conjectures on the number of fixed points of sympletic maps ([42]).
It generalized the Hopf index theorem that predicts the existence of fixed
points of a flow inside a planar region in terms of information about its be-
Introduction 3
havior on the boundary. Other applications in the study of dynamics include

the existence of periodic orbits in Hamiltonian systems and traveling wave
solutions for partial differential equations, structure of global attractors for
reaction-diffusion equations and delay differential equations, proof of chaotic
behavior in dynamical systems, and bifurcation theory. We use the works
[8, 52, 62, 76, 77] to ensure complete information on and references for the
Conley index theory.
The proposal of the present book is to reproduce the Morse decomposition
part of the Conley theory for control systems. Put more simply, we start
with a dynamical system associated with a differential equation ẋ = X0 (x)
on a manifold (or Euclidean state space), and we then consider a family of
differential equations of the form
n
X
ẋ = X0 (x) + ui (t) Xi (x)
i=1
where X1 , ..., Xn are vector fields and u1 , ..., un are real-valued functions such
that u (t) = (u1 (t) , ..., un (t)) is valued in a given subset U ⊂ Rn . The addi-
tional terms on the right hand side can be interpreted as control or perturba-
tion. In the control point of view, the set U ⊂ Rn is the given control range
and the vector fields X1 , ..., Xn determine the input structure. This family
of differential equations defines a control affine system, or in other words, a
nominal dynamical system with additional control inputs. A fundamental fact
in this definition is that the solutions for the control system do not define a
flow on the adjacent state space. We then sought a sense or interpretation for
Conley theory in the control system set-up. The motivation for responding
to this question was the controllability, the main subject in control theory.
Indeed, a set is ‘controllable’ if its points are pairwise connected by trajec-
tories of the control system. This means that any point in a controllable set
has the property of returning trajectories. Thus dynamical concepts, such as
periodicity and recurrence could be related to both notions of controllability
for control systems and Morse decomposition for dynamical systems. This
perception indicated that controllability must relate to some notion of Morse
decomposition for control systems.
The book written by Fritz Colonius and Wolfgang Kliemann [34] con-
tained the first attempt to reproduce the Conley theory for control systems.
These authors used the strategy of associating a dynamical system with a
given control affine system, studying control-theoretic aspects together with
the analysis of the associated dynamical system. The so called ‘control flow’
4 Introduction
is an infinite dimensional dynamical system that lives on the product space of

admissible inputs and state space. It was extensively used in later works with
the purpose of transferring dynamical objects and results from the theory of
dynamical systems to the theory of control systems. For instance, the Con-
ley definition of attractor was transferred to control systems by means of the
control flow ([35]). Nevertheless, the Colonius–Kliemann definition for chain
transitivity of control systems does not use the control flow. Actually, chain
transitivity can be defined in general situations of control systems, where the
required conditions to constitute the control flow need not be satisfied. In fact,
Carlos C. Braga Barros and Luiz A. B. San Martin [20] realized that the chain
controllability depends on a family of ideals in the system semigroup associ-
ated with the control system. In a straightforward way, they extended the con-
cept of a chain control set to more general semigroup actions, where the chain
recurrence came to depend on a family of subsets of the acting semigroup.
This idea has been extensively used to generalize dynamical concepts for semi-
group actions, which, in particular, composed a basis for the Morse decom-
position part of the Conley theory for control systems (as references sources
we mention the papers [22, 23, 25, 27, 83, 94, 95, 96, 97, 99, 100, 101, 102]).
The present book unites the ideas from these papers and the Colonius–
Kliemann book to define the elements of the Conley theory for control sys-
tems. Although dynamical concepts are easily generalized by the semigroup
methodology, control systems need not satisfy important conditions as, for in-
stance, invariance of limit sets and attractors. Thus, extending Conley results
from dynamical systems to control systems is a nontrivial work. Faced with
the possibility of noninvariant attractors and repellers, two distinct notions
of Morse decomposition have been considered in the control system set up
([25, 96]). A ‘dynamic Morse decomposition’ is a finite collection of compact
isolated invariant sets (Morse sets) which consist of the residence of limit sets,
while the trajectories of an external point can not come near the same Morse
set for both forward and backward times. An ‘attractor-repeller Morse de-
composition’ is a finite collection of sets given by intersections of sets in a
sequence of attractors and complementary repellers. In a classical dynami-
cal system, Conley shows that these two notions of Morse decomposition are
equivalent. In the paper [25], the Conley results were proved in a special sit-
uation where control systems satisfied certain translation hypothesis. One of
the main tasks in this book is to relate dynamic and attractor-repeller Morse
decompositions for control systems without assuming the translation hypoth-
Introduction 5
esis. The strategy is based on invariance issues of attractor-repeller pairs.

In compact state space, one verifies that any dynamic Morse decomposition
has an attractor-repeller configuration. On the other hand, Morse decom-
positions determined by invariant attractor-repeller pairs are dynamic Morse
decompositions. Thus, dynamic and attractor-repeller Morse decompositions
are equivalent concepts for control systems with invariant attractor-repeller
pairs.
The other principal intention of this book is to reproduce the Conley
theorem that connects Morse decomposition and chain recurrence. Every
attractor-repeller Morse decomposition contains the chain recurrence set, which
is the set of all chain recurrent points of the control system. In a special case,
if the finest attractor-repeller Morse decomposition exists, then its Morse
sets coincide with the connected components of the chain recurrence set (or,
simply, chain components). On the other hand, if the chain recurrence set ad-
mits finitely many connected components then it determines a dynamic Morse
decomposition. Consequently, for a control system with invariant attractor-
repeller pairs, the existence of finitely many chain components is a necessary
and sufficient condition for the existence of the finest Morse decomposition.
The Conley theorems are concluded by constructing a Lyapunov function for
Morse decomposition and then proving the existence of a complete Lyapunov
function for control systems on compact manifolds.
Many things should be explained before proving the Conley theorems.
This effort provides particular results which are interesting in themselves. In
fact, outside a Morse decomposition, every trajectory of the system comes
close to some Morse set, in forward as well as in backward time. Thus the
control system does not admit asymptotic transitivity outside a Morse de-
composition, or in other words, any asymptotic transitive set resides in some
Morse set. This means that periodicity, recurrence, and controllability occur
in the components of a Morse decomposition. In view of this fact, full chap-
ters are dedicated to study all transitivity concepts, with the main intention
being to describe technically what happens inside Morse sets. The amount of
information about Morse decomposition will be completed with the study of
chain transitivity, contributing to an estimate of the global dynamic behavior
of a control system.
In view of the attractor-repeller configuration of Morse decomposition, a
full chapter is devoted to explaining attractors. There are various notions of
attraction in the theory of dynamical systems. The notion of an attractor
6 Introduction
for a singular point was used by E. Coddington and N. Levinson [32]. The
conception of an attractor for a closed set was first studied by J. Auslander,
N. Bhatia, and P. Siebert [4]. The concepts of attractors related to stability
theory were extensively studied by N. Bhatia et al. [11, 12, 13, 14], including
the notions of global weak attractors and global uniform attractors. Alterna-
tive concepts of attractors and global attractors were used by J. Hale [56] and
O. Ladyzhenskaya [70]. Finally, C. Conley [36] defined a special notion of an
attractor that generates Morse decompositions. All these concepts of attrac-
tors are studied in the control system framework. The main task consists of
proving the connection between the Conley attractor and the uniform attrac-
tor. In general, every compact Conley attractor is an asymptotically stable
set, which means a stable uniform attractor. This result yields an important
statement that the existence of the finest Morse decomposition implies the
existence of a chain component that is asymptotically stable.
This book has been written with a wide audience in mind: control-theoretic
engineers or mathematicians, post-graduate students, and graduate students
researching dynamical systems and geometry. Control theorists may go di-
rectly to Chapter 4. Readers who are not familiar with dynamical systems
are invited to consult Appendix A. The two first chapters of the book con-
tain elementary concepts of control affine systems, but they are not mere
preliminaries. Chapter 1 provides an introduction to the basic definitions and
properties of control affine systems. It presents detailed mathematical for-
mulations of integral curves, shift space, control flow, and system semigroup.
Chapter 2 studies the elementary concepts of invariance, orbits, equilibria, and
periodicity in the control system setting. Topological properties of invariant
sets and orbits are investigated. Characterizations of critical and periodic
points are presented with some new features (Theorems 2.3.1, 2.3.2, 1.2.1,
and Proposition 2.4.1).
The middle part of this book consists of studies into various aspects of
transitivity for control systems. The weak transitivity relation is defined to-
gether with the concept of minimal sets in Chapter 3. The main result shows
that the minimal sets are upper bounds for a dynamic order among the equiv-
alence classes of weak transitivity. Chapter 4 treats the concepts of limit sets
and prolongational limit sets for control systems. These are crucial for the
concepts of asymptotic transitivity and attraction. The notion of prolonga-
tion is used to describe compact equistable sets. Special attention is given to
minimal equistable sets, an extension of minimal sets. The notion of asymp-
Introduction 7
totic transitivity is presented in Chapter 5 as a relation among the Poincaré

recurrent points of the control system. The studies include the classical con-
trol sets and invariant control sets. A state point is positively recurrent if it
lies in its positive limit set. Two positively recurrent points are equivalent if
each one is a limit point of the other. This relation constitutes the asymptotic
transitivity and is proved to be an equivalence relation among the positively
recurrent points. The Poincaré recurrence theorem for control systems is re-
produced, stating that µ-almost every point is positively recurrent, with µ
an invariant probability measure. The notion of a nonwandering point is also
studied in Chapter 5. By definition, a nonwandering point lies in its prolonga-
tional limit set. The main result shows that every point closed to a recurrent
point is nonwandering. In view of this relation, a more general notion of con-
trollability is defined by prolongations and its relative prolongational control
sets are studied.
A new feature in this middle section concerns the relation between Poincaré
recurrence and periodicity. It is clear that a periodic point is recurrent. The
converse does not hold, except in very special situations. The Poincaré–
Bendixson theorem states that a nonempty compact limit set of a C 1 planar
dynamical system, which contains no equilibrium point, is a periodic tra-
jectory ([60, Chapter 11]). In a higher dimension, however, it has no gen-
eralization or counterpart. A famous example of dynamical system on the
2-torus shows recurrent points which are not periodic (see a version for con-
trol system in Example 5.2.1). In the general set-up of control systems on
manifolds, recurrence and periodicity are equivalent concepts for points with
closed semi-orbits (Theorems 5.2.1, 5.2.2, 5.2.3, and 5.2.4).
Attractors, Morse decompositions, and chain transitive sets make up the
main part of the book. Chapter 6 deals with various notions of attrac-
tors and repellers for control systems. Since the Conley definition of at-
tractor approaches the uniform attraction, one gives special attention to the
properties of uniform attractor. Both the notions of global attractor and
global uniform attractor are studied in the chapter. A measure of noncom-
pactness is used to describe the asymptotic behavior of a control system ad-
mitting a global attractor. Conditions for the existence of a global attractor
are discussed. Conley theorems on chain recurrence are presented in Chapter
7. By contemplating trajectories with jumps, the chain transitivity gener-
alizes the asymptotic transitivity. This is also an equivalence relation and
its equivalence classes are described as maximal regions of chain transitiv-
8 Introduction
ity. One shows that asymptotic transitive and prolongational control sets are
contained in some chain transitive set. In order to present the main Conley
theorem of chain recurrence, the attractor-repeller pair paradigm is defined.
The Conley theorem describes the chain recurrence set as the intersection
of all attractor-repeller pairs of the control system. The last chapter of the
book studies dynamic and attractor-repeller Morse decompositions of control
systems separately. In compact state space, every dynamic Morse decompo-
sition admits an attractor-repeller configuration, while, on the other hand,
every invariant attractor-repeller Morse decomposition is a dynamic Morse
decomposition. The main Conley theorem of chain recurrence implies that
any Morse decomposition contains all chain recurrent points of the system.
The existence of finitely many chain transitivity classes is equivalent to the
existence of the finest Morse decomposition. The last important result shows
that the components of a Morse decomposition are connected by trajectories
which go through decreasing levels of some Lyapunov function. This implies
the existence of a complete Lyapunov function that decreases strictly on tra-
jectories outside the chain recurrence set and maps each chain equivalence
class onto a critical value.
Although these theorems on Morse decompositions are not new, they are
proved with the absence of translation hypothesis. This extends and improves
the results on dynamic Morse theory for control systems. The notion of a limit
compact control system is a new, and the relation between asymptotically
compact and limit compact control systems is thus a new result (Proposition
6.4.6). Another new feature is the use of a prolongational limit set to describe
the complementary repeller of a Conley attractor as well as the properties of
attractor-repeller pair (Theorem 7.2.3 and Proposition 7.2.4). Finally, it was
not known that nonwandering points are chain recurrent points (Proposition
7.1.7).
The dynamical concepts of a control system do not require compact state
space, but the main Conley theorems are assured under compactness. The
dynamic Morse theory on noncompact space the has many open questions
to be investigated. At the end of Chapter 8, a formulation of generalized
Morse decomposition is suggested for further discussion on Conley’s ideas in
noncompact state space.
Parts of the book may be used for one-semester courses or seminars in
mathematics or control-theoretic engineering, with the following suggestions:
• A first course in control system theory: Chapters 1 and 2, for students

Introduction 9
who are familiar with ordinary differential equations and basic notions
of real analysis and general topology.
• Controllability and asymptotic transitivity: Chapters 3, 4, and 5.
• Morse decompositions and chain transitivity: Chapters 6, 7, and 8.
• Seminar on periodicity: Sections 2.2, 2.3, 2.4, and 5.1.
• Seminar on Poincaré recurrence: Sections 4.1, 5.2, and 5.3.
• Seminar on attractors: Chapter 6.
• Introduction to Morse theory of dynamical systems: Appendix A.
The book may be also used as a reference source for various topics in
dynamical and control systems, contributing to the research of students in-
terested in the dynamics of control systems. Its contents may integrate basic
references for academic dissertations.
Implicitly, the book contains a survey about ongoing research into the
dynamics of control systems. It reproduces the Morse decomposition part of
the Conley theory, leaving the Conley index theorems to future work. The
dynamical concepts presented here might be viewed from different mathe-
matical problems. For instance, the family of differential equations given
Pn
by the formula ẋ = X0 (x) + i=1 ui (t) Xi (x) can be interpreted as time-
dependent perturbations of an ordinary differential equation by the vector
fields X1 , ..., Xn . In this case, one wants to study the Conley concepts of the
perturbed system relative to the Conley concepts of the nominal dynamical
system ẋ = X0 (x). An advanced mathematical problem considers X0 , ..., Xn
as invariant vector fields on a Lie group. A task in this case is to investigate a
homogeneous structure for Morse decompositions and chain transitivity. An-
other mathematical problem discusses the relationship between periodicity
and Poincaré recurrence of control systems with the intention of establishing
a general Poincaré-Bendixson theorem.
10 Introduction
Chapter 1
Fundamental Theory of
Control systems
In the language of mathematics, the concept of a control system was formu-

lated with the purpose of regulating dynamical systems. Intuitively, a control
system is an undetermined dynamical system, in which appropriate functions
can be chosen to a given criterion, resulting in a system with desired proper-
ties. The possible model classes of control systems are, for instance, algebraic
differential models (e.g., Flies et al. [49]), input-output operators (e.g., Fran-
cis [51]), the behavioral approach (e.g., Willems [113]), or the classical system
model (e.g., Colonius and Kliemann [34]). The classical mathematical formu-
lation of a control system consists of a family of differential equations, which
are interpreted as models for distinct forms of operation of the same device.
A set of control functions (inputs) determines the differential equations and
then the behaviors (outputs) of the system.
This book considers the classical mathematical paradigm of control sys-
tems. This first chapter introduces the basic definitions and presents the main
properties of control systems used afterwards. Section 1.1 contains the general
definition of control systems. Section 1.2 defines the shift space, which often
determines the control index set. Section 1.3 introduces the class of control
affine systems, which is the model of control system studied throughout the
book. The flow point of view of a control affine system is described in Section
1.4 via the control flow. Section 1.5 describes the group structure of a control
affine system.
11
12 Chapter 1 · Fundamental Theory of Control systems
1.1 Basic definitions

This first section contains basic notations and definitions of control sys-
tems. Throughout the book, the symbol M denotes a differentiable manifold
equipped with a compatible distance d; Rn stands for the n-dimensional Eu-
clidean space endowed with an inner product h·, ·i and associated norm k·k.
The state space of a control system consists of a manifold M , where we
often deal with topological objects. In order to establish the basic notations
for topological concepts, for x ∈ M and > 0, we define the open -ball
B (x, ) and the closed -ball B [x, ] centered at x respectively as
B (x, ) = {y ∈ M : d (x, y) < } ,

B [x, ] = {y ∈ M : d (x, y) ≤ } .
For a given subset X ⊂ M , the notations int (X), cl (X), and fr (X) means
respectively the interior of X, the closure of X, and the boundary of X in M .
The same notations are used in any other case of metric space that appears
throughout the book.
In order to define admissible control inputs, we need the following ingre-
dients.
Definition 1.1.1 For a given real number s ∈ R and two functions u, v :

R → Rn , the s-concatenation of u and v is the function w : R → Rn defined
by
u(t), if t 6 s
w(t) = .
v(t − s), if t > s
The s-shift of u is the function u · s : R → Rn given by u · s (t) = u (s + t)
for all t ∈ R.
For sequences of functions u1 , ..., uk and numbers s1 , s2 , ..., sk−1 with s1 <
s2 < ... < sk−1 , we may define the (s1 , ..., sk−1 )-concatenation of u1 , ..., uk by

 u1 (t), if t 6 s1

 u2 (t − s1 ), if s1 < t ≤ s2

u (t) = .. .


 .

uk (t − sk−1 ), if t > sk−1
Definition 1.1.2 Let U ⊂ Rn . A family of functions U = {u : R → U} is

said to be admissible if it satisfies the following properties:
Section 1.1 · Basic definitions 13
1. Each function u ∈ U is locally integrable, that is, u is Lebesgue integrable

on every bounded interval.
2. For each u ∈ U and s ∈ R, the s-shift u · s is contained in U.
3. For u, v ∈ U and s ∈ R, the s-concatenation of u and v is contained in

U.
A trivial admissible family is given by one constant function u (t) ≡ u0 ∈

Rn . A non-trivial admissible family is basically formed by piecewise constant
functions. A function u : R → U to be piecewise constant means that the real
line R is decomposed into subintervals of length bounded below by a positive
number such that u is constant on each subinterval. The admissible family of
all piecewise constant functions is denoted by Upc .
The general definition of control systems is given in the following.
Definition 1.1.3 A control system Σ = (M, U, U, X) is formed by
1. A state space M that is a d-dimensional C ∞ -differentiable manifold.
2. A control range U ⊂ Rn and an admissible family of control functions

U = {u : R → U}.
3. A family of ordinary differential equations
ẋ = X (x, u (t))
depending on the control functions, where X : M × Rn → T M is a C ∞

map from the product manifold M × Rn into the tangent bundle T M of
M.
In the case M = Rd , one may consider a C ∞ map X : Rd × Rn → Rd in

the definition of control systems.
Then each element u in the control range U corresponds to a constant
control function. In the trivial case of a unique constant control function, the
control system corresponds to a classical dynamical system. In the nontrivial
case, each control function u ∈ U determines a nonautonomous differential
equation ẋ = X (x, u (t)) on the manifold M . To guarantee solutions for this
equation, we can not apply the usual procedure to reduce the nonautonomous
differential equation to an autonomous one by introducing t as an additional
state variable, since the dependence of u on t need not be differentiable.

However, we may locally analyze the associated integrated equation
Z t
x (t) = x + X (x (s) , u (s)) ds
0
and proceed with Carathéodory’s theory allowing measurable dependence on

t. In this case, a solution x (t) of the differential equation ẋ = X (x, u (t)) with
initial condition x (0) = x0 is a broken integral curve, that is, x (t) is absolutely
continuous, satisfies the differential equation almost everywhere, and satisfies
the initial condition. Here, x (t) absolutely continuous means that x (t) is
differentiable almost everywhere, its derivative is Lebesgue integrable, and
Rt
x (t) = x (0) + 0 x0 (s) ds. We refer to Sontag [92] for the technical details in
Rd and to Coddington [32] for Carathéodory’s theory of differential equations.
The theory developed in this book requires solutions defined forward on
all positive times t ∈ R+ . In fact, we study concepts which depend on the
action of a semigroup defined by positive time transitions. This is explained
in Section 1.5 and commented in the notes throughout the book. For technical
simplifications, however, we assume throughout that for each control function
u ∈ U and each point x ∈ M the preceding equation has a unique solution
ϕ (t, x, u), defined for all time t ∈ R, with ϕ (0, x, u) = x. In particular, we
assume that for every u in the control range U the C ∞ vector field Xu : x ∈
M 7→ X (x, u) ∈ T M is complete, that is, the corresponding flow (t, x) →
ϕ (t, x, u) is defined globally on R × M .
Three special classes of control systems are extensively studied in mathe-
matics. They are:
Example 1.1.1 A linear control system takes the form
ẋ = A (x) + B (u(t))
where A is a d×d real matrix, B is a d×n real matrix, M = Rd , and U = Rn .
Example 1.1.2 A bilinear control system takes the form

n
X
ẋ = A0 (x) + ui (t)Ai (x)
i=1
n
where u(t) = (u1 (t), ..., un (t)) ∈ R , A0 , A1 , ..., An are d × d real matrices,
M = Rd , and U = Rn .
Section 1.2 · Shift Space 15
Example 1.1.3 A control affine system on a C ∞ -manifold M takes the

form
n
X
ẋ = X0 (x) + ui (t)Xi (x)
i=1
where X0 , X1 , ..., Xn are C ∞ -vector fields on M . Both linear and bilinear
control systems are special cases of control affine system.
1.2 Shift Space

We shall now describe the shift space, the theoretic set of control inputs
considered throughout the book. It is a metrizable compactification of the
piecewise control functions and the shift map defines a dynamical system on
it. The results presented in this section are extracted from [34, Chapter 4].
Readers who are not familiar with concepts of functional analysis may consult
[84] or just consider the statements as preliminary assumptions.
Let Upc be a family of piecewise constant functions with a compact and
convex control range U ⊂ Rn . Let L∞ (R, Rn ) be the vector space of all mea-
surable functions from R into Rn which are essentially bounded, i.e. bounded
up to a set of measure zero. Since the control range U ⊂ Rn is compact,
∗
we have Upc ⊂ L∞ (R, Rn ). It is well-known that L∞ (R, Rn ) = L1 (R, Rn ) ,
n n
where L1 (R, R ) is the space of all functions from R into R for which the
∗
absolute value is Lebesgue integrable, and L1 (R, Rn ) is the dual space of
L1 (R, Rn ) (see e.g. [33, Theorem 4.5.1]). For each α ∈ L1 (R, Rn ), the
L1 -norm of α is given by
Z
kαk1 = kα (t)k dt.
R
Define the linear functional fα : L∞ (R, Rn ) → R by

Z
fα (u) = hu (t) , α (t)i dt.
R
The weak* topology on L∞ (R, Rn ) is the weak topology induced on

L∞ (R, Rn ) by the collection {fα : α ∈ L1 (R, Rn )}, that is, the weakest topol-
ogy such that fα is continuous for all α ∈ L1 (R, Rn ). This is the topology for
which the sets fα−1 ((a, b)), for α ∈ L1 (R, Rn ) and (a, b) open interval in R,
form a subbase.
From now on, we assume that L∞ (R, Rn ) is endowed with the weak*
topology and define the subspace U = cl (Upc ) ⊂ L∞ (R, Rn ).
Lemma 1.2.1 The shift is an internal operation in U.

Tk
Proof. Let u ∈ U and t ∈ R. Suppose that u · t ∈ i=1 fα−1 i
((ai , bi )) with
α1 , ..., αk ∈ L1 (R, Rn ). Then fαi (u · t) ∈ (ai , bi ) for each i. Since
Z Z
fαi (u · t) = hu (t + s) , αi (s)i ds = hu (τ ) , αi (τ − t)i dτ = fαi ·(−t) (u)
R R
Tk
we have u ∈ i=1 fα−1 i ·(−t)
((ai , bi )). Hence there is some v ∈ Upc such that
Tk −1 Tk
v ∈ i=1 fαi ·(−t) ((ai , bi )). This implies that v · t ∈ i=1 fα−1 i
((ai , bi )). Since
v · t ∈ Upc , it follows that u · t ∈ U.
This enables us to define the shift map Θ : R × U → U by Θ (t, u) = u · t.

The shift space is given by the space U equipped with the shift map. We
can show that the shift space is a compact metrizable space and the shift map
is continuous. We need the following:
Lemma 1.2.2 The subspace U ⊂ L∞ (R, Rn ) given by
U = {u : R → Rn : u (t) ∈ U for a.a. t ∈ R, measurable}
is a compact metrizable space.
Proof. It should be noticed that L1 (R, Rn ) is a separable Banach space,

which means there is a countable and dense subset {αk : k ∈ N} of L1 (R, Rn )
(see [33, Proposition 3.4.5]). Define
∞
X 1 |fαk (u − v)|
d (u, v) = (1.1)
2k 1 + |fαk (u − v)|
k=1
for every pair u, v ∈ L∞ (R, Rn ). By Alaoglu’s theorem, the unit ball of

L∞ (R, Rn ) is compact and metrizable, and a metric is given by 1.1 (see [44,
Theorem 3, p. 434]). Since the control range U is compact, the set U is
bounded in L∞ (R, Rn ). It remains to show that U is closed (weak* closed)
in L∞ (R, Rn ). Indeed, for any compact interval I ⊂ R, consider the set
U|I = { u|I : u ∈ U} ⊂ L2 (I, Rn )

2
where L2 (I, Rn ) is endowed with the L2 -norm given by kuk2 = R ku (t)k dt.
R
n
Since U is convex, U|I is a convex set in L2 (I, R ). Moreover, for a given se-
quence (un ) in U|I that converges to u ∈ L2 (I, Rn ) in the L2 -norm, there
Section 1.2 · Shift Space 17
is a subsequence (unk ) such that unk (t) → u (t) for a.a. t ∈ I. As U

is compact, u (t) ∈ U for a.a. t ∈ I, and hence u ∈ U|I . Thus U|I
is closed in the L2 -norm of L2 (I, Rn ). Since every closed and convex set
in a locally convex space is weakly closed, it follows that U|I is closed in
the weak topology of L2 (I, Rn ) (see [44, Theorem 13, p. 422]). Hence
U|I is closed in the weak* topology, since L2 (I, Rn ) is a reflexive Banach
space. We now consider a sequence (un ) in U that converges (weak* con-
verges) to u in L∞ (R, Rn ). Then fα (un ) → fα (u) for all α ∈ L1 (R, Rn ).

Hence I un (t) , α (t) dt → I u (t) , α (t) dt for all α ∈ L2 (I, Rn ), since
L2 (I, Rn ) ⊂ L1 (I, Rn ). This means that un |I → u|I in the weak* topology
of L2 (I, Rn ), and therefore u|I ∈ U|I . Since the compact interval I ⊂ R is
arbitrary, it follows that u (t) ∈ U for a.a. t ∈ R, that is, u ∈ U. Thus U is
closed in L∞ (R, Rn ).
We are now able to prove the main result of this section.
Theorem 1.2.1 The shift space U is a compact metrizable space and the shift
map Θ : R × U → U is continuous.
Proof. Since U ⊂ U is weak* closed, it follows that U is a compact

metrizable space, by Lemma 1.2.2. To show that Θ is continuous, we need
to prove that fα ◦ Θ is continuous for every α ∈ L1 (R, Rn ). In fact, suppose
that tk → t in R and uk → u in U . We have
|fα ◦ Θ (tk , uk ) − fα ◦ Θ (t, u)|

= [uk (tk + s) , α (s) − u (t + s) , α (s)] dt

R

≤ [uk (tk + s) , α (s) − uk (t + s) , α (s)] dt

R

+ [uk (t + s) , α (s) − u (t + s) , α (s)] dt
R

= [uk (τ ) , α (τ − tk ) − uk (τ ) , α (τ − t)] dτ

R

+ [uk (τ ) , α (τ − t) − u (τ ) , α (τ − t)] dτ .
R
By continuity, we have fα·(−t) (uk ) → fα·(−t) (u), hence

[uk (τ ) , α (τ − t) − u (τ ) , α (τ − t)] dτ → 0.

R
By defining K = supv∈U v, we have

uk (τ ) , α (τ − tk ) − α (τ − t) dτ

R
≤ |uk (τ ) , α (τ − tk ) − α (τ − t)| dτ
R
≤ uk (τ ) α (τ − tk ) − α (τ − t) dτ
R

≤ K α (τ − tk ) − α (τ − t) dτ
R
= K α · (−tk ) − α · (−t)1 .
Since −tk → −t, we have α · (−tk ) → α · (−t) in the L1 -norm, hence

α · (−tk ) − α · (−t)1 converges to zero. It follows that
|fα ◦ Θ (tk , uk ) − fα ◦ Θ (t, u)| → 0
and therefore fα ◦ Θ is continuous.
Thus the shift map Θ : R × U → U defines a dynamical system on the

shift space U . The following proposition is crucial for the properties of control
affine systems studied in the next section.
Proposition 1.2.1 For u, v ∈ U and s ∈ R, let w be the s-concatenation of

u and v. Then w ∈ U.
Proof. Take sequences (un ) , (vn ) in U such that un → u and vn → v and

denote by wn the s-concatenation of un and vn . We claim that (wn ) → w
in the weak* topology, and then w ∈ U. In fact, for a given α ∈ L1 (R, Rn ),
define β, γ : R → Rn by

0, if t ≤ s α (t) , if t ≤ s
β (t) = , γ (t) = .
α (t) , if t > s 0, if t > s
Section 1.3 · Control affine systems 19
Then β, γ ∈ L1 (R, Rn ) and we have
|fα (wn − w)|

= wn (t) − w (t) , α (t) dt

R
s
+∞
= u (t) − u (t) , α (t) dt + vn (t − s) − v (t − s) , α (t) dt
n
−∞ s
≤ |fα (un − u)| + |fα (vn · (−s) − v · (−s))|
+∞

+ un (t) − u (t) , α (t) dt
s s

+ vn · (−s) (t) − v · (−s) (t) , α (t) dt
−∞
= |fα (un − u)| + |fα (vn · (−s) − v · (−s))|
+ |fβ (un − u)| + |fγ (vn · (−s) − v · (−s))| .
Now, by continuity, we have fα (un ) → fα (u), fα (vn ) → fα (v), fβ (un ) →

fβ (u), and fγ (vn · (−s)) → fγ (v · (−s)), hence |fα (wn − w)| converges to 0.
This means that wn → w, as claimed.
1.3 Control affine systems

For the purposes of this book, we consider only control affine systems with
piecewise control functions. This section is dedicated to studying the main
properties of this class of control systems.
Consider a control affine system

n
ẋ (t) = X (x (t) , u (t)) = X0 (x (t)) + ui (t) Xi (x (t)) (Σ)
i=1
on a d-dimensional C ∞ -manifold M , with compact and convex control range

U ⊂ Rn , and control inputs u in the shift space U . Assume that, for each
u ∈ U and x ∈ M , the preceding equation has a unique solution ϕ (t, x, u),
t ∈ R, with ϕ (0, x, u) = x, and the vector fields X (·, u), u ∈ U, are complete.
The solutions of the control system Σ define a map ϕ : R × M × U → M ,
(t, x, u)
→ ϕ (t, x, u), which is called the phase map of the control system.
The partial maps of ϕ determine other important maps. For t ∈ R fixed, the
map ϕt : M × U → M given by ϕt (x, u) = ϕ (t, x, u) is called t-projection;
for u ∈ U and t ∈ R fixed, the map ϕut : M → M defined by ϕut (x) = ϕ (t, x, u)
is called a transition; for u ∈ U and x ∈ M fixed, the map ϕux : R → M
given by ϕux (t) = ϕ (t, x, u) is said to be a motion through x.
The image set ϕux (R) is called a trajectory through x. Both the motion
ϕux and the trajectory ϕux (R) through x are said to be periodic if there is
a T = 0 such that ϕux (t) = ϕux (t + T ) for all t ∈ R. The point x is called a
critical point (or equilibrium point) with respect to the control function
u ∈ U if the trajectory ϕux (R) reduces to the point x, that is, ϕ (t, x, u) = x
for all t ∈ R.
We often consider the time-reversed control system defined by
ẋ (t) = −X (x (t) , u (t)) .
The solutions for the time-reversed system can be described by means of the
solutions for the system Σ. For each control function u ∈ U we define the
time reversed control function u∗ (t) = u (−t), t ∈ R. Note that u∗ ∈ U and
∗
(u∗ ) = u. For a given x0 ∈ M , the solution for the initial value problem
ẋ = −X (x, u (t)), x (0) = x0 , is the function ϕ∗ (t, x0 , u) = ϕ (−t, x0 , u∗ ). In
fact, we have ϕ∗ (0, x0 , u) = ϕ (0, x0 , u∗ ) = x0 and
d ∗ d d (−t) d
ϕ (t, x0 , u) = ϕ (−t, x0 , u∗ ) = ϕ (−t, x0 , u∗ )
dt dt dt d (−t)
= −X (ϕ (−t, x0 , u∗ ) , u∗ (−t))
= −X (ϕ∗ (t, x0 , u) , u (t))
almost everywhere. We may then use the time-reversed phase map ϕ∗ :

R × M × U → M defined as (t, x, u)
→ ϕ∗ (t, x, u). All the properties of
the control system proved for positive time can be immediately proved for
negative time by using the time-reversed control system.
For u ∈ U and t ∈ R, the transition ϕut : M → M is a diffeomorphism of
M , since the flow for the vector field Xu is differentiable. For convenience, we
might use the usual notation exp (tXu ) instead of ϕut to distinguish between
transition by u ∈ U and transition by u ∈ U.
Let us see some examples.
Example 1.3.1 Consider the control affine system
ẋ = u1 (t) A1 (x) + u2 (t) A2 (x)

on the unit 2-sphere S2 ⊂ R3 where

⎛ ⎞ ⎛ ⎞
0 1 0 0 0 1
⎝
A1 = −1 0 0⎠ and A2 = ⎝ 0 0 0⎠
0 0 0 −1 0 0
with the control range U = [0, 1] × [0, 1]. The trajectories through x =
(x1 , x2 , x3 ) ∈ S2 with respect to the constant control functions u ≡ (1, 0)
and v ≡ (0, 1) are given by
exp (tXu ) (x) = exp (tA1 ) (x) = (x1 cos t + x2 sin t, −x1 sin t + x2 cos t, x3 ) ,
exp (tXv ) (x) = exp (tA2 ) (x) = (x1 cos t + x3 sin t, x2 , −x1 sin t + x3 cos t) ,
that is, exp (tXu ) (x) is a rotation around the z axis and exp (tXv ) (x) is a
rotation around the y axis (see Fig. 1.1).
Figure 1.1: Trajectories of a control affine system on the 2-sphere.
Example 1.3.2 Let M = S1 × S1 be the 2-torus identified with a subset of

C2 and consider the control affine system on S1 × S1 given by
ẋ = u1 X1 (x) + u2 X2 (x)
with control range U = [−1, 1] × [−1, 1] ⊂ R2 , where X1 (z1 , z2 ) = (iz1 , 0) and

X2 (z1 , z2 ) = (0, iz2 ). The trajectories of X1 and X2 are circles of the form

exp (tX1 ) (z1 , z2 ) = eti z1 , z2 and exp (tX2 ) (z1 , z2 ) = z1 , eti z2 .
These trajectories are illustrated in Fig. 1.2.

Figure 1.2: Trajectories of a control affine system on the 2-torus.
We now present the main property of solutions for control systems.
Proposition 1.3.1 The solutions for the control system Σ satisfy the cocycle
property
ϕ(t + s, x, u) = ϕ(t, ϕ(s, x, u), u · s).
Proof. For x0 ∈ M , u ∈ U, and s ∈ R, denote y = ϕ(s, x0 , u). Define the

curves φ (t) and ψ (t) in M by
φ (t) = ϕ (t + s, x0 , u) and ψ (t) = ϕ (t, y, u · s)
for t ∈ R. Consider the initial value problem
ẋ = X (x, u · s(t)) , x (0) = y. (IV P )
Then both functions φ and ψ are solutions for IV P . Indeed, it is immediately

clear that both φ and ψ are absolutely continuous and φ(0) = ψ(0) = y.
Moreover
d d
ψ (t) = ϕ (t, y, u · s) = X (ϕ (t, y, u · s) , u · s (t))
dt dt
= X (ψ (t) , u · s (t))
almost everywhere and

d d (t + s) d
φ (t) = φ (t)
dt dt d (t + s)
d
= ϕ (t + s, x0 , u)
d (t + s)
= X (ϕ (t + s, x0 , u) , u (t + s))
= X (φ (t) , u · s (t)) .
almost everywhere. By uniqueness of solution, φ (t) = ψ (t), that is
ϕ (t + s, x0 , u) = ϕ (t, ϕ (s, x0 , u) , u · s)
as desired.
The cocycle property says that the trajectory through x at time t + s and
control function u coincides with the trajectory through y = ϕ(s, x, u) at time
t and control function u · s. In terms of transitions, this means that
ϕut+s = ϕu·s
t ◦ ϕus , t, s ∈ R, u ∈ U.
The cocycle property yields a general property of transitions: for t, s ∈ R

and u ∈ U , there is u ∈ U such that ϕut+s = ϕut ◦ ϕus . We have a converse
property by considering either t, s ≥ 0 or t, s ≤ 0. We need the following
technical result:
Proposition 1.3.2 For s ≥ 0 and u, v ∈ U, let w ∈ U be the s-concatenation

of u and v
u(t), if t ≤ s
w (t) = .
v(t − s), if t > s
Then
ϕ (t, x, u) , if t ≤ s
ϕ (t, x, w) =
ϕ (t − s, ϕ (s, x, u) , v) , if t > s
for every x ∈ M .
Proof. For a given x0 ∈ M , consider the initial value problem
ẋ = X (x, w (t)) , x (0) = x0 .
Define the curve φ (t) in M by

ϕ (t, x0 , u) , if t ≤ s
φ (t) = .
ϕ (t − s, ϕ (s, x0 , u) , v) , if t > s
Then φ (t) is absolutely continuous and φ (0) = ϕ (0, x0 , u) = x0 since 0 ≤ s.

Moreover, for t < s we have
d d
φ (t) = ϕ (t, x0 , u) = X (ϕ (t, x0 , u) , u (t))
dt dt
= X (φ (t) , w (t))
almost everywhere, and for t > s we have

d d
φ (t) = ϕ (t − s, ϕ (s, x0 , u) , v)
dt dt
d
= ϕ (t, ϕ (−s, ϕ (s, x0 , u) , v) , v · (−s))
dt
= X (ϕ (t, ϕ (−s, ϕ (s, x0 , u) , v) , v · (−s)) , v · (−s) (t))
= X (ϕ (t − s, ϕ (s, x0 , u) , v) , v (t − s))
= X (φ (t) , w (t))
where we used the cocycle property. Thus φ (t) satisfies the equation ẋ =
X (x, w (t)) almost everywhere and φ (0) = x0 . By uniqueness of solution, it
follows that φ (t) = ϕ (t, x, w) and the proposition is proved.
The following result is a consequence of Proposition 1.3.2.
Corollary 1.3.1 For s, τ ≥ 0 and u, v ∈ U, there is w ∈ U such that
ϕ (τ , ϕ (s, x, u) , v) = ϕ (τ + s, x, w)
for every x ∈ M .
Proof. By Proposition 1.3.2, there is w ∈ U such that

ϕ (t, x, u) , if t ≤ s
ϕ (t, x, w) = .
ϕ (t − s, ϕ (s, x, u) , v) , if t > s
Since ϕ (s, x, u) = ϕ (s, x, w), the equality holds for τ = 0. For τ > 0, we have
τ + s > s, and then ϕ (τ + s, x, w) = ϕ (τ , ϕ (s, x, u) , v).
This property extends naturally to finitely many control functions.
Corollary 1.3.2 For t1 , ..., tk ≥ 0 and u1 , ..., uk ∈ U, there is w ∈ U such

that
ϕ (tk , ϕ (tk−1 , ..., ϕ (t1 , x, u1 ) , ..., uk−1 ) , uk ) = ϕ (t1 + · · · + tk , x, w)
for every x ∈ M .
In terms of transitions, for t1 , ..., tk ≥ 0 and u1 , ..., uk ∈ U , there is w ∈ U

such that
uk−1
ϕutkk ◦ ϕtk−1 ◦ · · · ◦ ϕut11 = ϕw
t1 +···+tk .
By using the time reversed control system, we can obtain a similar result
for negative times t1 , ..., tk ≤ 0.
Piecing together the solutions for constant control functions we obtain all
the solutions by piecewise constant functions, as in the following:
Proposition 1.3.3 For a given u ∈ Upc and 0 = t0 < t1 , ..., tN such that
⎧
⎪ u1 , for 0 ≤ t ≤ t1
⎪
⎪
⎪
⎪ u 2 for t1 < t ≤ t1 + t2
,
⎨
u (t) = ..
⎪ .
⎪
⎪ N
⎪
⎪
−1
N
⎩ uN , for ti < t ≤ ti
i=0 i=0
one has

k−1

ϕ (t, x, u) = exp t− ti Xuk ◦ exp tk−1 Xuk−1 ◦ · · · ◦ exp (t1 Xu1 ) (x)
i=0

k−1
k
for ti ≤ t ≤ ti , k = 1, ..., N .
i=0 i=0
Proof. Define the curve ψ (t) in M by

k−1

ψ (t) = exp t− ti Xuk ◦ exp tk−1 Xuk−1 ◦ · · · ◦ exp (t1 Xu1 ) (x)
i=0

k−1
k
for ti ≤ t ≤ ti , k = 1, ..., N . Then ψ is absolutely continuous in
i=0 i=0

N
0, ti and ψ (0) = exp (0Xu1 ) (x) = x. Moreover
i=0
d
ψ (t)
dt

d
k−1

= exp t− ti Xuk ◦ exp tk−1 Xuk−1 ◦ · · · ◦ exp (t1 Xu1 ) (x)
dt i=0
k−1
d
= exp (tXuk ) exp − ti Xuk ◦ · · · ◦ exp (t1 Xu1 ) (x)
dt i=0
k−1

= Xuk exp (tXuk ) exp − ti Xuk ◦ · · · ◦ exp (t1 Xu1 ) (x)
i=0
= Xuk (ψ (t))

k−1
k d
for ti < t < ti , k = 1, ..., N . Hence ψ (t) = X (ψ (t) , u (t)) al-
i=0 dt
i=0

N
most everywhere in 0, ti . By uniqueness of solution, it follows that
Ni=0

ψ (t) = ϕ (t, x, u) in 0, ti .
i=0
By using the time-reversed control system, we can obtain a similar result

for negative times tN , ..., t1 < t0 = 0, that is, if we have
⎧
⎪ N N−1
⎪
⎪ u , for t ≤ t ≤ ti
⎪
⎪
N i
⎨ i=0 i=0
..
u (t) = .
⎪
⎪
⎪
⎪ u , for t + t2 < t ≤ t1
⎪
⎩ 2 1
u1 , for t1 < t ≤ 0
then

k−1

ϕ (t, x, u) = exp t− ti Xuk ◦ exp tk−1 Xuk−1 ◦ · · · ◦ exp (t1 Xu1 ) (x)
i=0

k
k−1
for ti ≤ t ≤ ti , k = 1, ..., N .
i=0 i=0
An interesting consequence from Proposition 1.3.3 is that all transitions
by piecewise constant functions are diffeomorphisms of M , as the following:
Corollary 1.3.3 For any s ∈ R and u ∈ Upc , the transition ϕus : M → M is
−1
a diffeomorphism of M and (ϕus ) = ϕu·s
−s .
Proof. We may assume that s > 0. Then there is a sequence of times

N
0 = t0 < t1 , ..., tN , with ti = s, such that
i=0
⎧
⎪ u1 , for 0 ≤ t ≤ t1
⎪
⎪
⎪
⎪ u , for t1 < t ≤ t1 + t2
⎨ 2
u (t) = .. .
⎪ .
⎪
⎪
⎪
⎪
N−1
⎩ uN , for ti < t ≤ s
i=0
By Proposition 1.3.3 we have

ϕus (x) = exp (tN XuN ) ◦ exp tN −1 XuN −1 ◦ · · · ◦ exp (t1 Xu1 ) (x)
for every x ∈ M . Hence ϕus is a composition of diffeomorphisms of M , and

therefore ϕus is a diffeomorphism of M . Now, by Proposition 1.3.1, we have
−s ◦ ϕs (x) = ϕ (−s, ϕ(s, x, u), u · s) = ϕ(s − s, x, u) = ϕ (0, x, u) = x,

ϕu·s u
and
ϕus ◦ ϕu·s
−s (x) = ϕ (s, ϕ(−s, x, u · s), u) = ϕ (s, ϕ(−s, x, u · s), (u · s) · (−s))
= ϕ(s − s, x, u · s)
= ϕ (0, x, u · s) = x.
−1
for every x ∈ M . Hence (ϕus ) = ϕu·s
−s .
Below we present other example of control affine system.
Example 1.3.3 Consider the bilinear control system on M = R2 given by
ẋ = A0 (x) + u (t) A1 (x)
where
1 0 0 1
A0 = , A1 = ,
0 −1 −1 0
and U = [−2, 2]. For a given control u ∈ [−2, 2], we have the autonomous
differential equation ẋ = A0 (x) + uA1 (x). For u ≡ 0, the trajectory through
x = (x1 , x2 ) is

x1 et
exp (tA0 ) (x) = .
x2 e−t
In this case, the origin (0, 0) is a saddle point (see √
Fig. 1.3).
For u = 0, the eigenvalues of A0 + uA1 are ± 1 − u2 . If |u| ≤ 1 then
these eigenvalues are real numbers. For u = ±1, the trajectories through
x = (x1 , x2 ) are

(x1 + x2 ) t + x1
exp (t [A0 + A1 ]) (x) = ,
− (x1 + x2 ) t + x2

(x1 − x2 ) t + x1
exp (t [A0 − A1 ]) (x) = .
(x1 − x2 ) t + x2
Figure 1.3: The origin (0, 0) is a saddle point.
Figure 1.4: On the left are stationary points in the line x2 = −x1 ; on the
right are stationary points in the line x2 = x1 .
In the case u = 1 all the points in the line x2 = −x1 are stationary, while
in the case u = −1 all the points in the line x2 = x1 are stationary (see Fig.
1.4).
For 0 < |u| < 1, the solutions curves are of the form
√ √
√ u t 1−u2 √−u −t 1−u2
φ (t) = C1 e + C 2 e .
1 − u2 − 1 1 + 1 − u2
In this case, the equilibrium point at the origin is a saddle point (see Fig.
1.5). √
Finally, if |u| > 1 then the eigenvalues are the complex numbers ±i u2 − 1.
The solution curves are of the form

u cos t u 2−1
φ (t) = C1 +
− cos t u2 − 1 − u2 − 1 sin t u2 − 1

u sin
t u2 − 1
+C2
u2 − 1 cos t u2 − 1 − sin t u2 − 1
√
2
Figure 1.5: Portrait for u = . The equilibrium point at the origin is a
2
saddle point.
which describe periodic trajectories around the origin (see Fig. 1.6).
√
Figure 1.6: Portrait for u = 2. Periodic trajectories around the origin.
Piecing together these trajectories for constant controls we obtain all the
trajectories with piecewise constant functions. See an example of trajectory
with piecewise constant function in Fig. 1.7).
The last important property of transitions we show in this section is the

equicontinuity. It should be remembered that a family of continuous maps
F ⊂ C (M, M ) is equicontinuous at x ∈ M if, for a given ε > 0, there is
δ > 0 such that d (y, x) < δ implies d (f (y) , f (x)) < ε for every f ∈ F. The
family F is equicontinuous if it is equicontinuous at every x ∈ M , and F is
uniformly equicontinuous if the number δ does not depend on the point x.
An equicontinuous family on compact space is uniformly equicontinuous (see
e.g. [84]).
Proposition 1.3.4 For compact sets [−T, T ] ⊂ R and K ⊂ M , the family of

transitions F = {ϕut : t ∈ [−T, T ] , u ∈ U} is uniformly equicontinuous on K.
Figure 1.7: A broken integral curve of a bilinear control system.
Proof. It is enough to prove the theorem for the compact interval [0, T ].
By covering K with finitely many compact local coordinate systems, we may
assume that the C ∞ map X is lipschitzian on K × U, with Lipschitz constant
L. Then, for t ∈ [0, T ], u ∈ U , and x, y ∈ K in a local coordinate system, we
have
ϕ (t, x, u) − ϕ (t, y, u)
t
= x−y+ (X (ϕ (τ , x, u) , u (τ )) − X (ϕ (τ , y, u) , u (τ ))) dτ
0
which implies
ϕ (t, x, u) − ϕ (t, y, u)

t
≤ x − y + X (ϕ (τ , x, u) , u (τ )) − X (ϕ (τ , y, u) , u (τ )) dτ
0
t
≤ x − y + L (ϕ (τ , x, u) , u (τ )) − (ϕ (τ , y, u) , u (τ )) dτ
0
t
= x − y + L ϕ (τ , x, u) − ϕ (τ , y, u) dτ .
0
By Gronwall’s inequality (see [60, Chapter 8, Section 4]), it follows that
ϕ (t, x, u) − ϕ (t, y, u) ≤ etL x − y
for all t ∈ [0, T ] and u ∈ U. Now, for a given ε > 0, we can take δ > 0
such that δ < e−tL ε for all t ∈ [0, T ]. Then ϕ (t, x, u) − ϕ (t, y, u) < ε, for
Section 1.4 · Control flow 31
all t ∈ [0, T ], u ∈ U, and x, y ∈ K with x − y < δ. Finally, for a given

ε > 0, we can use the continuity of the charts to obtain δ > 0 such that
d (ϕ (t, x, u) , ϕ (t, y, u)) < ε, for all t ∈ [0, T ], u ∈ U , and x, y ∈ K with
d (x, y) < δ. Thus F is uniformly equicontinuous on K.
1.4 Control flow
Every control affine system associates with a dynamical system, the control
flow. In this section, we show the continuity of the phase map. We then
defines the control flow as the skew-product flow of the phase map and the
shift map. Throughout the book we explore the control flow to study various
dynamical concepts of a control system. The results presented here are based
on Colonius and Kliemann [34, Chapter 4]. The main results of flows we use
in this book are shown in Appendix A.
Consider the control affine system Σ. In order to show that the phase map
ϕ : R × M × U → M is continuous, we need the following lemmas:
Lemma 1.4.1 If (tk ) , (xk ) , (uk ) are sequences respectively in R, M , U , with

tk → 0 and xk → x, then ϕ (tk , xk , uk ) → x.
Proof. Suppose by contradiction that the sequence ϕ (tk , xk , uk ) does not

converge to x. There is then a number ε > 0 and a subsequence ϕ (tkm , xkm , ukm )
such that ϕ (tkm , xkm , ukm ) ∈
/ B [x, ε] for all m ∈ N. We may assume that ei-
ther tkm ≥ 0, for all m, or tkm ≤ 0, for all m. We assume the case tkm ≥ 0;
the other case is similar. For each m ∈ N, define the number
sm = sup {s > 0 : ϕ (s, xkm , ukm ) ∈ B [x, ε]} .
Since xkm → x, we may assume that xkm ∈ B [x, ε] for all m ∈ N. As

ϕ (tkm , xkm , ukm ) ∈
/ B [x, ε], it follows that sm is finite. Moreover, sm < tkm
and ϕ (sm , xkm , ukm ) ∈ fr (B [x, ε]). We may consider this situation in a local
coordinate system at x, which means that ϕ (sm , xkm , ukm ) − x = ε, with
sm → 0, as tkm → 0. This implies that
sm

ε = ϕ (sm , xkm , ukm ) − x = X (ϕ (τ , x km , u km ) , u km (τ )) dτ

0
sm
≤ X (ϕ (τ , xkm , ukm ) , ukm (τ )) − X (x, ukm (τ )) dτ
0
sm
+ X (x, ukm (τ )) dτ .
0
Since the control range U is compact, there is a number K > 0 such that
X (x, ukm (τ )) ≤ K for all m. If L > 0 is the Lipschitz constant of X on
B [x, ε] × U, it follows that
sm sm
ε ≤ L ϕ (τ , xkm , ukm ) − x dτ + K dτ
0 sm sm 0
≤ Lε dτ + K dτ
0 0

s
where 0 m (Lε + K) dτ converges to 0 as sm → 0. This contradiction implies
ϕ (tk , xk , uk ) → x.
Lemma 1.4.2 If xk → x in M and uk → u in U then ϕ (t, xk , uk ) →

ϕ (t, x, u) for every t ∈ R.
Proof. We firstly consider the system on Rd . For T > 0 and t ∈ [0, T ],

Section 1.4 · Control flow 33
we have
ϕ (t, xk , uk ) − ϕ (t, x, u)

t

= xk − x + [X (ϕ (τ , x , u ) , u (τ )) − X (ϕ (τ , x, u) , u (τ ))] dτ
k k k
0
t
≤ xk − x + X0 (ϕ (τ , xk , uk )) − X0 (ϕ (τ , x, u)) dτ
0
n
t
+ (uk )i (τ ) Xi (ϕ (τ , xk , uk )) − (uk )i (τ ) Xi (ϕ (τ , x, u))
i=1 0
n t
+ (uk )i (τ ) Xi (ϕ (τ , x, u)) − ui (τ ) Xi (ϕ (τ , x, u))
i=1 0
t
≤ xk − x + L0 ϕ (τ , xk , uk ) − ϕ (τ , x, u) dτ
0

n t
+ Li |(uk )i (τ )| ϕ (τ , xk , uk ) − ϕ (τ , x, u)
i=1 0
n t
+ |(uk )i (τ ) − ui (τ )| Xi (ϕ (τ , x, u)) dτ
i=1 0
where L0 , L1 , ..., Ln are respectively the Lipschitz constants for the vector
fields X0 , X1 , ..., Xn . If K = sup max {|v1 | , ..., |vn |}, the second and
(v1 ,...,vn )∈U
the third summands are bounded from above by
t
L ϕ (τ , xk , uk ) − ϕ (τ , x, u) dτ
0
where L = max {L0 , L1 K, ..., Ln K}. As uk → u, we have |(uk )i (τ ) − ui (τ )| →

0, and then the fourth summand converges to zero. Since xk → x, we also
have xk − x → 0. By writing
n
t
Ck = |(uk )i (τ ) − ui (τ )| Xi (ϕ (τ , x, u)) dτ
i=1 0
it follows that
ϕ (t, xk , uk ) − ϕ (t, x, u) ≤ xk − x + Ck

t
+L ϕ (τ , xk , uk ) − ϕ (τ , x, u) dτ .
0
By Gronwall’s inequality, this means that
ϕ (t, xk , uk ) − ϕ (t, x, u) ≤ eLt (xk − x + Ck ) .
As xk − x + Ck → 0, we have ϕ (t, xk , uk ) − ϕ (t, x, u) → 0. Hence, for k

large enough, we can consider ϕ (t, xk , uk ) and ϕ (t, x, u) in a local coordinate
system, and therefore ϕ (t, xk , uk ) → ϕ (t, x, u) for all t ∈ [0, T ]. We may re-
peat the proof for t ∈ [−T, 0], and then ϕ (t, xk , uk ) → ϕ (t, x, u) for all t ∈ R.
We now prove the continuity of the phase map.
Proposition 1.4.1 The phase map ϕ : R × M × U → M is continuous.
Proof. Suppose that tk → t in R, xk → x in M , and uk → u in U . By

Lemma 1.4.2, we have ϕ (t, xk , uk ) → ϕ (t, x, u). By Theorem 1.2.1, uk · t →
u · t. As tk − t → 0, we have
ϕ (tk , xk , uk ) = ϕ (tk − t, ϕ (t, xk , uk ) , uk · t) → ϕ (t, x, u)
by Lemma 1.4.1. Thus ϕ is continuous.
Finally, we have the main result of this section.
Theorem 1.4.1 Define the map Φ : R × M × U → M × U by
Φ (t, x, u) = (ϕ (t, x, u) , Θ (t, u)) .
Then Φ defines a dynamical system on M × U .
Proof. By Theorem 1.2.1 and Proposition 1.4.1, Φ is a continuous map.

Moreover, we have Φ (0, x, u) = (ϕ (0, x, u) , Θ (0, u)) = (x, u) for every (x, u) ∈
M × U . For t, s ∈ R and (x, u) ∈ M × U , we have
Φ (t + s, x, u) = (ϕ (t + s, x, u) , Θ (t + s, u))
= (ϕ (t, ϕ (s, x, u) , Θ (s, u)) , Θ (t, Θ (s, u)))
= Φ (t, Φ (s, x, u)) .
Hence the group property is satisfied, and therefore Φ is a dynamical system.

Section 1.5 · Group structure 35
The dynamical system Φ : R × M × U → M × U is said to be the control

flow of the control affine system Σ.
The relationships between the control system and the control flow are
given by projections and lifts. Consider the first-projection π : M × U → M .
A trajectory Φ(x,u) (R) of the control flow projects onto the trajectory ϕux (R)
of the control system, since π (Φ (t, x, u)) = ϕ (t, x, u) for all t ∈ R. On the
other hand, a trajectory of the control system lifts to M × U. For a given
subset A ⊂ M , the lift of A is defined as
L (A) = {(x, u) ∈ M × U : ϕ (t, x, u) ∈ A for all t ∈ R} .
By considering the t-projections ϕt : M × U → M , it can be seen that

L (A) = ϕ−1
t (A)
t∈R
and L (A) ⊂ A × U . Then (y, v) ∈ L (ϕux (R)) if and only if there is a function
γ : R → R such that ϕ (t, y, v) = ϕ (γ (t) , x, u). We have π (L (ϕux (R))) =
ϕux (R) and, in particular, Φ(x,u) (R) ⊂ L (ϕux (R)).
The control flow is often used to interpret dynamical concepts of the con-
trol system.
1.5 Group structure

In addition to the flow structure, a control affine system can be interpreted
as a transformation group. Proposition 1.3.3 states that the trajectories of
the control system Σ by piecewise constant controls are determined by the
corresponding family of vector fields F = {Xu : u ∈ U}. This reveals a struc-
ture of transformation group for the control system. It remains to extend this
group structure to all transitions of the system.
Proposition 1.5.1 For t, s ≥ 0 and u, v ∈ U , there is w ∈ U such that

ϕvs ◦ ϕut = ϕw
s+t .
Proof. Take sequences (uk ) , (vk ) in Upc such that uk → u and vk → v.

For each k, there is wk ∈ Upc such that ϕvsk ◦ ϕut k = ϕw k
s+t , by Proposition
1.3.2. By compactness, we may assume that (wk ) converges to some w ∈ U .
Now, by Proposition 1.4.1, it follows that
ϕvs ◦ ϕut (x) = ϕ (s, ϕ (t, x, u) , v) = lim ϕ (s, ϕ (t, x, uk ) , vk )

k→∞
= lim ϕ (s + t, x, wk )
k→∞
= ϕ (s + t, x, w)
= ϕw
s+t (x)
for every x ∈ M , and the result is proved.
Definition 1.5.1 The system group G and the system semigroup S for
the control affine system Σ are defined respectively by

G = ϕut11 ◦ · · · ◦ ϕutkk : ui ∈ U, ti ∈ R, k ∈ N ,
u1
S = ϕt1 ◦ · · · ◦ ϕutkk : ui ∈ U, ti ≥ 0, k ∈ N .
The system group G is the group of transformations of M generated by

the transitions of the system. The system semigroup S is a subsemigroup of

G. Then the set S −1 = φ−1 : φ ∈ S is also a subsemigroup of G. By Propo-
sition 1.5.1, the system semigroup coincides with the set of all transitions by
positive time, that is
S = {ϕut : u ∈ U, t ≥ 0} .
Hence the inverse system semigroup is given by S −1 = {ϕut : u ∈ U, t ≤ 0}.

We therefore have
S (x) = {ϕ (t, x, u) : t ≥ 0, u ∈ U} ,
−1
S (x) = {ϕ (t, x, u) : t ≤ 0, u ∈ U} ,
for any x ∈ M . In general, S (x) ∪ S −1 (x) is a proper subset of G (x) (see

Problem 1.6(4)). The sets S (x) and S −1 (x) are called positive semi-orbit
and negative semi-orbit of x, and they are usually denoted by O+ (x) and
O− (x), respectively.
By the continuity of the phase map, the closure of O+ (x) in M coincides
with the closure of the set of all positive semi-trajectories through x with
piecewise constant functions

cl O+ (x) = cl {ϕ (t, x, u) : t ≥ 0, u ∈ Upc }
Section 1.6 · Problems 37
and analogously

cl O− (x) = cl {ϕ (t, x, u) : t ≤ 0, u ∈ Upc } .
Hence properties which only use the closures of the orbits with piecewise con-
stant functions do not change under the extension to the shift space. Thus
we may view a control affine system as a family of vector fields on a mani-
fold. Sometimes, this convenience is useful to construct examples of control
systems. Problem 1.6(8) brings a general formulation of a control system
determined by family of local vector fields.
1.6 Problems
1. Show that a function u ∈ U is periodic if and only if u is a periodic
point of the shift Θ. (Hint: see Lemma 4.2.2 in [34])
2. Check that the system group G is a group with respect to the function
composition and the system semigroup S is a subsemigroup of G.
3. Define x ∼ y if and only if y ∈ G(x). Show that “ ∼” is an equivalence

relation on M and the equivalence classes are the orbits of G.
4. Consider the control affine system on Rn given by ẋ = u (t), with control

n
range U = [0, 1] . For x = (x1 , ..., xn ) ∈ Rn , show that
S (x) = {y ∈ Rn : yi ≥ xi } ,
S −1 (x) = {y ∈ Rn : yi ≤ xi } ,
G (x) = Rn .
Conclude that S (x) ∪ S −1 (x) = G (x).
5. Let T2 be the 2-torus identified with the quotient space R2 Z2 and
endowed with the sympletic canonical form ω = dx∧dy. This means that
T2 is a 2-sympletic manifold. If H : T2 → R is a C ∞ function, define the
∂H ∂ ∂H ∂
vector field XH on T2 by XH = − . This vector field
∂x2 ∂x1 ∂x1 ∂x2
XH is called Hamiltonian vector field with the Hamiltonian H in the
sense that dH (·) = ω (XH , ·). Consider the C ∞ functions hi : R2 → R
given by
h1 (x1 , x2 ) = cos 2πx1 sin 2πx2 ,

h2 (x1 , x2 ) = sin 2πx1 sin 2πx2 ,
h3 (x1 , x2 ) = cos 2πx2 ,

5 7
h4 (x1 , x2 ) = f (x1 ) with supp (f ) ⊂ , ,
8 8

1 3 dg 1
h5 (x1 , x2 ) = g (x2 ) with supp (g) ⊂ , and = 0,
4 4 dx2 2

where supp (f ) = cl x ∈ R2 : f (x) = 0 is the support of the function
f . Let Hi : T2 → R be the function induced by hi on T2 . This means
that Hi (ρ (x)) = hi (x), where ρ : R2 → T2 is the standard projection.
5
Consider the (Hamiltonian) control affine system ẋ = i=1 ui (t) Xi (x)
4
on T2 , with control range U = [0, 1] × [−1, 1] , where Xi = XHi for
i = 1, ..., 5. Show that G (x) = T2 for every x ∈ T2 .
6. Let G + be the subset of the system group G defined as

⎧ ⎫
⎨
n ⎬
G + = ϕutnn ◦ · · · ◦ ϕut11 : ui ∈ U, tj ≥ 0, n ∈ N .
⎩ ⎭
j=0
−1
Prove that G + is a subsemigroup of G and G = G + ∪ (G + ) (this means
that G + is total in G, see e.g., [71]).
7. Assume that M = Gl (d, R) is the group of all d × d invertible real

2
matrices. It is an open submanifold of Rd called linear Lie group. For
x, y ∈ Gl (d, R), denote by xy the matrix product of x and y. Consider
a control affine system

m
ẋ(t) = A0 (x(t)) + ui (t)Ai (x(t))
i=1
on Gl (d, R), where A0 , A1 , ..., Am are d × d real matrices.
(a) Prove that the solutions satisfy ϕ (t, x, u) = ϕ (t, 1, u) x, for all
x ∈ Gl (d, R), u ∈ U , and t ∈ R, where 1 stands for the identity
matrix in Gl (d, R). (Hint: note that the solutions for constant
controls are matrix exponential maps)
(b) Conclude that
G = {ϕ (t, 1, u) : t ∈ R, u ∈ U} ⊂ Gl (d, R)
and S = {ϕ (t, 1, u) : t ≥ 0, u ∈ U}.
8. A local vector field on a manifold M is a smooth map F : U → T M

with U an open subset of M . In this case, U is an open submanifold of
M . Let F be a family of local vector fields of M and define the sets:
GF = {exp (tk Fk ) ◦ · · · ◦ exp (t1 F1 ) : Fi ∈ F, ti ∈ R, k ∈ N} ,

SF = {exp (tk Fk ) ◦ · · · ◦ exp (t1 F1 ) : Fi ∈ F, ti ≥ 0, k ∈ N} ,
where only reasonable compositions are considered. The set GF is a

local group in the sense that it is a group of local diffeomorphisms of
M and the set SF is a local semigroup. Sometimes F is called a control
system, with system group GF and system semigroup SF . A family of
local vector fields may be interpreted as a control system with obstacles.
!
∂ ∂
Consider the family of vector fields F = , |U on R2 where
∂x ∂y
U = {(x, y) : x < 0}. Show that
(a) SF (x0 , y0 ) = {(x, y) : x ≥ x0 , y ≥ y0 } whenever (x0 , y0 ) ∈ R2 \ U ;

(b) SF (x0 , y0 ) = {(x, y0 ) : x ≥ x0 } whenever (x0 , y0 ) ∈ U ;
(c) GF (x0 , y0 ) = R2 for every (x0 , y0 ) ∈ R2 .
9. Consider the control affine system ẋ = u (t) on Rn , as defined in Problem

4 above, and the control system on Rn determined ! by the family of
∂ ∂
partial derivatives operators F = , ..., . Check that S (x) =
∂x1 ∂xn
SF (x) for all x ∈ Rn .
10. Let F = {A1 , ..., Am } be a family of pairwise commutative n × n real

m
matrices and consider the control affine system ẋ = i=1 ui (t) Ai (x)
m
on Rn with control range U = [0, a] , a > 0. Prove that cl (S (x)) =
cl (SF (x)) for all x ∈ Rn , or in other words, the semi-orbit SF (x)
coincides with all positive semi-trajectories through x with piecewise
constant functions. (We often consider control affine systems in this
form.)
1.7 Notes and references

The geometric control theory is a set of concepts and ideas from control the-
ory and dynamical systems, where the main control-theoretic notions are con-
trol sets, chain control sets, linearization, stabilization, and spectrum, while
the main dynamical concepts are recurrence, chain recurrence, and Lyapunov
stability. The theory of transformation groups unites all these concepts.
The concept of a system semigroup as considered by Colonius–Kliemann
[34] is determined by the family of vector fields of the control affine system,
hence the system semigroup corresponds to the transitions with piecewise
constant functions. This present book extends the system semigroup to the
transitions with control functions in the shift space.
The theory of control systems determined by families of local vector fields
was developed by San Martin in the monograph [88].
Chapter 2
Elementary Concepts
This chapter focuses on invariance, orbits, equilibria, and periodicity in

control theory. The concept of invariance concerns the analysis of trajectories
which start in certain subsets of states. This is connected with the study of
orbits. Critical points have stationary trajectories. Periodic trajectories are
closed integral curves of the system. All these concepts are essential to the
study of stability and controllability. In Section 2.1, we introduce various
notions of invariance and study topological properties of invariant sets. In
Section 2.2 we study the properties of orbit, positive semi-orbit, and negative
semi-orbit. Section 2.3 contains characterizations of critical points. Various
properties and characterizations of periodic points are presented in Section
2.4. Throughout the book, there is a fixed control affine system Σ.
2.1 Invariant sets

In this section we introduce the notion of invariance and study the funda-
mental properties of invariant sets.
Definition 2.1.1 A set X ⊂ M is called
1. positively invariant if ϕ (t, x, u) ∈ X for all x ∈ X, u ∈ U, and t > 0;
2. negatively invariant if ϕ (t, x, u) ∈ X for all x ∈ X, u ∈ U , and

t < 0;
3. invariant if it is both positively and negatively invariant;
41
42 Chapter 2 · Elementary Concepts
4. isolated invariant if it is invariant and there is a neighborhood V of

X with the following property: if x ∈ V and ϕ (t, x, u) ∈ V , for all t ∈ R
and u ∈ U, then x ∈ X.
In terms of transitions, a set X ⊂ M is positively invariant if and only

if ϕut (X) ⊂ X for all u ∈ U and t > 0; it is negatively invariant if and only
if ϕut (X) ⊂ X for all u ∈ U and t < 0; and it is invariant if and only if
ϕut (X) ⊂ X for all u ∈ U and t ∈ R. A set is isolated invariant if and only
if it is the maximal invariant set in some neighborhood of itself. In this case,
the neighborhood is called an isolating neighborhood. Isolated invariant
sets will be discussed in Chapter 6.
For closed sets, we may replace U by Upc in Definition 2.1.1 (see Problem
2.5(3)).
Note that the whole state space M is invariant itself. The empty set ∅ is
also invariant by vacuousness.
Proposition 2.1.1 Let {Xγ }γ∈Γ be a collection of positively invariant, nega-

tively invariant, or invariant subsets of M . Their intersection and their union
then have the same property.
Proof. Suppose that {Xγ }γ∈Γ is a collection of positively invariant sets.

"
Let x ∈ Xγ , t > 0, and u ∈ U. Then x ∈ Xγ for some γ. As Xγ is posi-
γ∈Γ "
tively invariant, we have ϕ (t, x, u) ∈ Xγ . Hence the union Xγ is positively
γ∈Γ
invariant. Now suppose that the intersection Xγ is nonempty and take
γ∈Γ
x∈ Xγ , t > 0, and u ∈ U. Since any Xγ is positively invariant, we have
γ∈Γ
ϕ (t, x, u) ∈ Xγ for every γ ∈ Γ. Hence ϕ (t, x, u) ∈ Xγ , and therefore
γ∈Γ
Xγ is positively invariant. The proof for any other case is analogous.
γ∈Γ
Proposition 2.1.2 Let X ⊂ M be positively invariant, negatively invariant,

or invariant. Its closure cl (X) and its interior int (X) then have the same
property.
Proof. Suppose that X ⊂ M is positively invariant. Let t > 0 and

u ∈ U. As the transition ϕut is a homeomorphism of M , we have ϕut (cl (X)) =
cl (ϕut (X)) ⊂ cl (X). Hence, cl (X) is positively invariant. Analogously,
Section 2.1 · Invariant sets 43
ϕut (int (X)) = int (ϕut (X)) ⊂ int (X), and therefore int (X) is positively in-
variant. Similarly, X negatively invariant implies that cl (X) and int (X) are
negatively invariant, and therefore X invariant implies that cl (X) and int (X)
are invariant.
Proposition 2.1.3 A set X ⊂ M is positively invariant if and only if its

complement set M \ X is negatively invariant. Consequently, M is invariant
if and only if M \ X is invariant.
Proof. Suppose that X is positively invariant and take y ∈ M \ X, t > 0,

and u ∈ U . If x ∈ X then ϕ (t, x, u · (−t)) ∈ X, by positive invariance.
This means that ϕ (−t, y, u) ∈
/ X, since ϕ (t, ϕ (−t, y, u) , u · (−t)) = y ∈
/ X.
Thus M \X is negatively invariant. The proof for the converse is analogous.
Consequently, the boundary of any invariant set is invariant.
Corollary 2.1.1 If X ⊂ M is invariant then its boundary fr (X) is invariant.
Proof. Since fr (X) = cl (X)∩cl (M \ X) and X is invariant, Propositions

2.1.1, 2.1.2, and 2.1.3 together imply that fr (X) is invariant.
Thus both int (X) and fr (X) are invariant whenever X is invariant. We
have a partial converse for this result.
Proposition 2.1.4 Let X ⊂ M be an open or closed set. If both fr (X) and

int (X) are invariant then X is invariant.
Proof. If X is open then X = int (X) is invariant by hypothesis. If X is

closed then X = fr (X) ∪ int (X), and therefore X is invariant by Proposition
2.1.1.
We can describe the invariant sets of the control system by means of the
invariant sets of the associated control flow. Recall that the lift of a set B ⊂ M

to the product M × U is given by L (B) = t∈R ϕ−1 t (B) .
Proposition 2.1.5 Let B ⊂ M be a nonempty set. Then
1. L (B) is invariant for the control flow.
2. B is invariant if and only if B × U = L (B).

Proof. (1) Let (x, u) ∈ L (B) and t ∈ R, and v ∈ U. We have ϕ (s, x, u) ∈

B, for all s ∈ R, and Φ (t, x, u) = (ϕ (t, x, u) , u · t). It follows that
ϕ (s, ϕ (t, x, u) , u · t) = ϕ (s + t, x, u) ∈ B for all s ∈ R. Hence Φ (t, x, u) ∈
L (B), and therefore L (B) is invariant.
(2) Suppose that B is invariant. Then ϕ (s, x, u) ∈ B for all s ∈ R and
(x, u) ∈ B × U . Hence B × U ⊂ L (B). The other inclusion L (B) ⊂ B × U
holds in general. Thus B × U = L (B). As to the converse, suppose that
B × U = L (B). Then B × U is invariant by item (1). For any x ∈ B, it
follows that (ϕ (t, x, u) , u · t) ∈ B × U for all t ∈ R and u ∈ U. Hence B is
invariant.
2.2 Orbits and semi-orbits

In this section we investigate properties of orbits. Recall that the positive
and negative semi-orbits of a point x ∈ M are respectively defined as
O+ (x) = {ϕ (t, x, u) : t ≥ 0, u ∈ U} ,
−
O (x) = {ϕ (t, x, u) : t ≤ 0, u ∈ U} .
We also define the orbit of x by the set union
O (x) = O+ (x) ∪ O− (x) .
Note that the orbit of x need not correspond to the orbit of the system group
G (x).
In line with these notations, for any x ∈ M and T ≥ 0 we define the sets
# #
−
O>T
+
(x) = ϕux ((T, +∞)) , O>T (x) = ϕux ((−∞, −T )) ,
u∈U u∈U
# #
−
O≥T
+
(x) = ϕux ([T, +∞)) , O≥T (x) = ϕux ((−∞, −T ]) ,
u∈U u∈U
# #
−
O≤T
+
(x) = ϕux ([0, T ]) , O≤T (x) = ϕux ([−T, 0]) .
u∈U u∈U
All these sets define maps of the form F : M → 2M , where 2M denotes

the set of all subsets of M . Then, for a given map F : M → 2M and a subset
"
X ⊂ M , we may consider its image F (X) = F (x).
x∈X
Section 2.2 · Orbits and semi-orbits 45
For further considerations, we define the subsets of the system semigroup
S>T = {ϕut : t > T, u ∈ U} and S≥T = {ϕut : t ≥ T, u ∈ U}
for each T ≥ 0. Then S≥0 = S and S≥T is a two-sided ideal of S. This means
that both translates φS≥T and S≥T φ are contained in S≥T , for all φ ∈ S. The
set S>T has the same property. In particular, S>T and S≥T are semigroups.
For any x ∈ M and T ≥ 0 we have
− −1
O>T
+
(x) = S>T (x) , O>T (x) = S>T (x) ,
− −1
O≥T
+
(x) = S≥T (x) , O≥T (x) = S≥T (x) .
Figure 2.1: A broken integral curve.
Example 2.2.1 Consider the family of linear vector fields F = {X1 , X2 } on

M = R2 where

−1 0 1 0
X1 = , X2 = .
0 −1 0 −1
Since X1 , X2 are commutative matrices, the control system determined by F

corresponds to the control affine system ẋ = u1 (t) X1 (x) + u2 (t) X2 (x) on
2
R2 , with control range U = [0, a] , a > 0 (cf. Problems 1.6(8, 10)). The origin
0 is a stationary point. The trajectories of X1 are straight lines moving to
the origin. The trajectories of X2 move asymptotically to the axes x and y,
respectively, for t → +∞ and t → −∞. Fig. 2.1 illustrates a trajectory of
the control system by piecing together trajectories of X1 and X2 . The positive
semi-orbit of any point in one of the axes is thin while the positive semi-orbit
of any point out of the axes is thick. The same holds for negative semi-orbit
and orbit (see Fig. 2.2). Note that any one of the quadrants is an invariant
set.
Figure 2.2: A positive semi-orbit in the interior of the first quadrant; a neg-
ative semi-orbit in the interior of the second quadrant; and an orbit in the
interior of the third quadrant.
Figure 2.3: A broken integral curve.
Example 2.2.2 Consider the control affine system on M = R2 determined

by the linear commutative vector fields X1 , X2 given by

0 1 1 1
X1 = , X2 = .
−1 0 −1 1
The origin 0 is a stationary point. The trajectories of X1 move on circles

centered at 0. The trajectories of X2 move on spirals centered at 0. Fig. 2.3
illustrates a trajectory of the control system by piecing together trajectories of
X1 and X2 . The positive semi-orbit of any point x = 0 is the exterior region
of the circle containing x while the negative semi-orbit of x is the interior
region of the circle containing x, excluding the origin (see Fig. 2.4). The
orbit of any point distinct from the origin coincides with R2 \ {0}.
Figure 2.4: On the left, the positive semi-orbit of a point x; on the right, the
negative semi-orbit of x.
It can be easily seen that a set X ⊂ M is positively invariant, negatively

invariant, or invariant if and only if respectively O+ (X) ⊂ X, O− (X) ⊂ X,
or O (X) ⊂ X, which are respectively equivalent to O+ (X) = X, O− (X) =
X, or O (X) = X. The set X is isolated invariant if and only if it has a
neighborhood V such that x ∈ V and O (x) ⊂ V implies x ∈ X.
The semi-orbits O+ (x) and O− (x) are, respectively, positively invariant
and negatively invariant themselves. Before proving these facts, we show a
useful technical lemma.
Lemma 2.2.1 For any x ∈ M , t, T ≥ 0, and u ∈ U, one has

+ + +
1. ϕut O>T (x) ⊂ O>t+T +
(x), ϕut cl O>T (x) ⊂ cl O>t+T (x) , and
+ +
ϕut int O>T (x) ⊂ int O>t+T (x) .
$ % $ $ %% $ %
2. ϕut O≥T
+
(x) ⊂ O≥t+T +
(x), ϕut cl O≥T+
(x) ⊂ cl O≥t+T
+
(x) , and
$ $ %% $ %
ϕut int O≥T+
(x) ⊂ int O≥t+T
+
(x) .
− −
− −
3. ϕu−t O>T (x) ⊂ O>t+T (x), ϕu−t cl O>T (x) ⊂ cl O>t+T (x) , and
− −
ϕu−t int O>T (x) ⊂ int O>t+T (x)
$ % $ $ %% $ %
− − − −
4. ϕu−t O≥T (x) ⊂ O≥t+T (x), ϕu−t cl O≥T (x) ⊂ cl O≥t+T (x) ,
$ $ %% $ %
− −
and ϕu−t int O≥T (x) ⊂ int O≥t+T (x) .
Proof. We will only prove item (1). The other items can be proved
in the same way. If y ∈ O>T +
(x) then there are s > T and v ∈ U such
that ϕ (s, x, v) = y. Proposition 1.3.2 shows that there is w ∈ U such that
ϕ (t, ϕ (s, x, v) , u) = ϕ (t + s, x, w), with t + s > t + T . Hence ϕ (t, y, u) =
+
ϕ (t + s, x, w) ∈ O>t+T+
(x). This proves that ϕut O>T (x) ⊂ O>t+T+
(x).
u
Now, since the transition ϕt is a homeomorphism of M , we have
+ + +
ϕut cl O>T (x) = cl ϕut O>T (x) ⊂ cl O>t+T (x) ,
+ + +
ϕut int O>T (x) = int ϕut O>T (x) ⊂ int O>t+T (x) .
Proposition 2.2.1 For any x ∈ M and T ≥ 0,

+ +
1. O>T
+
(x), cl O>T (x) , and int O>T (x) are positively invariant.
$ % $ %
2. O≥T
+
(x), cl O≥T
+
(x) , and int O≥T
+
(x) are positively invariant.
−
− −
3. O>T (x), cl O>T (x) , and int O>T (x) are negatively invariant.
$ % $ %
− − −
4. O≥T (x), cl O≥T (x) , and int O≥T (x) are negatively invariant.
Proof. We only prove item (1). Let t > 0 and u ∈ U. By Lemma 2.2.1,
+
ϕut O>T (x) ⊂ O>t+T
+
(x) ⊂ O>T
+
(x), and therefore O>T
+
(x) is positively
+ +
invariant. Now both the sets cl O>T (x) and int O>T (x) are positively
invariant by Proposition 2.1.2.
This result implies an important property of the semi-orbits.

+
Corollary 2.2.1 If x ∈ int O>T (x) then O>T+
(x) is an open set in M . The
− −
sets O≥T (x), O>T (x), and O≥T (x) have the same property.
+
+ +
Proof. If x ∈ int O>T (x) then O>T +
(x) ⊂ int O>T (x) , by Proposition

2.2.1. Hence O>T+
(x) = int O>T +
(x) , that is, O>T
+
(x) is open. The cases
− −
O≥T (x), O>T (x), and O≥T (x) are analogously proved.
+

Similarly, we can show that the orbit O (x) is an open set in M if x ∈

int (O(x)). In this case, O (x) is an open submanifold of M with dim O (x) =
dim M .
As a consequence of Propositions 2.1.1 and 2.2.1, for any subset X ⊂ M ,
−
both O>T+
(X) and O≥T+
(X) are positively invariant while both O>T (X) and
−
O≥T (X) are negatively invariant. However, the orbit O (x) may be neither
positively invariant nor negatively invariant (see Problem 1.6(4)).
The following property of semi-orbit is used afterwards:
Proposition 2.2.2 For all times t, s > 0 and any set X ⊂ M , the following
equalities hold
$ % $ $ %% $ $ $ %%%
cl O≥s+t
+
(X) = cl O≥s
+
O≥t
+
(X) = cl O≥s +
cl O≥t
+
(X) .
$ %
Proof. By the cocycle property we have O≥s
+
O≥t
+
(X) = O≥s+t
+
(X),
$ % $ $ %%
hence cl O≥s+t (X) = cl O≥s O≥t (X) . It is clear that
+ + +
$ $ %% $ $ $ %%%
cl O≥s
+
O≥t
+
(X) ⊂ cl O≥s
+
cl O≥t
+
(X) .
On the other hand

$ $ %% # $ $ %%
O≥s
+
cl O≥t
+
(X) = ϕuT cl O≥t
+
(X)
T ≥s,u∈U
# $ $ %%
= cl ϕuT O≥t
+
(X)
T ≥s,u∈U
# $ %
⊂ cl O≥T
+
+t (X)
T ≥s
$ %
= cl O≥s+t
+
(X) .
$ $ $ %%% $ %
Hence cl O≥s +
cl O≥t
+
(X) ⊂ cl O≥s+t
+
(X) , which implies the
equalities.
The following presents a topological property of orbits.

−
Proposition 2.2.3 For any x ∈ M and T ≥ 0, the sets O≤T +
(x), O≤T (x),
−
O≥T (x), and O≥T (x) are pathwise connected. In particular, the sets O+ (x),
+
O− (x), and O (x) are pathwise connected.

Proof. We first prove that O≤T +

(x) is pathwise connected for all x ∈ M
and T > 0. Let y, z ∈ O≤T (x). Then there are t1 , t2 ∈ [0, T ] and u1 , u2 ∈ Upc
+
such that y = ϕ (t1 , x, u1 ) and z = ϕ (t2 , x, u2 ). We may assume that t1 , t2 >

0. Define a path γ : [0, 2] → M by

ϕux1 (t1 − tt1 ) , if 0 ≤ t < 1
γ (t) = .
ϕux2 (tt2 − t2 ) , if 1 ≤ t ≤ 2
Then γ (0) = ϕux1 (t1 ) = y and γ (2) = ϕux2 (t2 ) = z. Since the motions ϕux1
and ϕux2 are continuous, it follows that γ is continuous. As 0 < t1 − tt1 ≤ t1 ,
for 0 ≤ t < 1, and 0 ≤ tt2 − t2 ≤ t2 , for 1 ≤ t ≤ 2, we have γ (t) ∈ O≤T
+
(x) for
all t ∈ [0, 2]. Hence γ is a path connecting y with z in O≤T (x) (see Fig. 2.5).
+
Thus O≤T +
(x) is pathwise connected for all x ∈ M and T > 0. Now, since
" +
O (x) =
+
O≤T (x), we have O+ (x) is pathwise connected. Analogously,
T ≥0
−
we prove that O≤T (x) is pathwise connected, for all x ∈ M and T > 0, and
then O (x) is also pathwise connected. Since O (x) = O+ (x) ∪ O− (x) with
−
x ∈ O+ (x)∩O− (x), it follows that the orbit O (x) is pathwise connected. We

now prove that O≥T +
(x) is pathwise connected for all x ∈ M and T > 0. For
y, z ∈ O≥T (x) there are t1 , t2 ≥ T and u1 , u2 ∈ U such that y = ϕ (t1 , x, u1 )
+
and z = ϕ (t2 , x, u2 ). Take s1 , s2 ≥ 0 such that t1 = T + s1 and t2 = T + s2 .

By the first part of the proof, there is a path γ : [0, 2] → O+ (x) connecting
ϕ (s1 , x, u1 ) and ϕ (s2 , x, u2 ). Let u3 ∈ U be the T -concatenation of u1 · s1
and u2 · t2 . Then

u1 · s1 (t) , if t ≤ T
u3 (t) =
u2 · t2 (t − T ) , if t > T

u1 · s1 (t) , if t ≤ T
= .
u2 · s2 (t) , if t > T
Define the path α : [0, 2] → M by α (t) = ϕ (T, γ (t) , u3 ). We have
α (0) = ϕ (T, γ (0) , u3 )

= ϕ (T, ϕ (s1 , x, u1 ) , u3 )
= ϕ (T, ϕ (s1 , x, u1 ) , u1 · s1 )
= ϕ (T + s1 , x, u1 )
= ϕ (t1 , x, u1 )
= y
and
α (2) = ϕ (T, γ (2) , u3 )

= ϕ (T, ϕ (s2 , x, u2 ) , u3 )
= lim ϕ (T + δ n , ϕ (s2 , x, u2 ) , u3 )
δ n →0+
= lim ϕ (T + δ n , ϕ (s2 , x, u2 ) , u2 · s2 )
δ n →0+
= lim ϕ (T + δ n + s2 , x, u2 )
δ n →0+
= ϕ (T + s2 , x, u2 )
= ϕ (t2 , x, u2 )
= z.
Moreover, as γ (t) ∈ O+ (x), we have α (t) ∈ ϕuT3 (O+ (x)) ⊂ O≥T

+
(x). Hence
α is a path in O≥T (x) connecting y with z. Therefore O≥T (x) is pathwise
+ +
connected.
Figure 2.5: A path γ connecting the points y and z in the positive semi-orbit
O+ (x).
− −
Consequently, the closures of the sets O≤T
+
(x), O≤T (x), O≥T
+
(x), O≥T (x),
−
O (x), O (x), and O (x) are connected. The following extends this property
+
to the greater orbit G (x).
Proposition 2.2.4 The orbit G (x) is pathwise connected and its closure
cl (G (x)) is connected for all x ∈ M .
Proof. Let y, z ∈ G (x). Then y = ϕutkk ◦ · · · ◦ ϕut11 (x) and z = ϕvsll ◦ · · · ◦

ϕvs11 (x), with ti , sj ∈ R and ui , vj ∈ U. Hence
x = ϕu−t1 ·t1 1 ◦ · · · ◦ ϕu−tk ·t

k
k
(y)
and then
z = ϕvsll ◦ · · · ◦ ϕvs11 ◦ ϕu−t1 ·t1 1 ◦ · · · ◦ ϕu−tk ·t
k
k
(y) .
Define the path γ : [0, 1] → G (x) by
1 ·t1 k ·tk
γ (t) = ϕvtsl l ◦ · · · ◦ ϕvts11 ◦ ϕu−tt 1
◦ · · · ◦ ϕu−tt k
(y) .
We have γ (0) = y and γ (1) = z. Thus G (x) is pathwise connected.
The following result expresses the transitivity property of the semi-orbits.
Proposition 2.2.5 Let x, y, z ∈ M and T, S 0.
1. If x ∈ O>T
+
(y) and y ∈ O>S
+
(z) then x ∈ O>T
+
+S (z).
2. If x ∈ O≥T
+
(y) and y ∈ O≥S
+
(z) then x ∈ O≥T
+
+S (z).
− − −
3. If x ∈ O>T (y) and y ∈ O>S (z) then x ∈ O>T +S (z) .
− − −
4. If x ∈ O≥T (y) and y ∈ O≥S (z) then x ∈ O≥T +S (z).
Proof. For item (1), let x ∈ O>T

+
(y) and y ∈ O>S
+
(z). By Lemma 2.2.1,
O>T
+
(y) ⊂ O>T +S (z). Hence, x ∈ O>T (y) ⊂ O>T +S (z). The other cases
+ + +
are analogously proved.
In the same way, we can show that closure and interior of semi-orbits also
have the transitivity property.
Proposition 2.2.6 Let x, y, z ∈ M and T, S ≥ 0.

+ + +
1. If x ∈ cl O>T (y) and y ∈ cl O>S (z) then x ∈ cl O>T +S (z) .
$ % $ % $ %
2. If x ∈ cl O≥T
+
(y) and y ∈ cl O≥S
+
(z) then x ∈ cl O≥T
+
+S (z) .
− − −
3. If x ∈ cl O>T (y) and y ∈ cl O>S (z) then x ∈ cl O>T +S (z) .
$ % $ % $ %
− − −
4. If x ∈ cl O≥T (y) and y ∈ cl O≥S (z) then x ∈ cl O≥T +S (z) .
The same property holds replacing closure by interior.
The following result reveals a property of periodicity.
Proposition 2.2.7 Let x ∈ M and T ≥ 0.
1. If x ∈ O>T
+
(x) then x ∈ O>t
+
(x) for all t ≥ 0.
2. If x ∈ O≥T
+
(x) with T > 0 then x ∈ O>t
+
(x) for all t ≥ 0.
− −
3. If x ∈ O>T (x) then x ∈ O>t (x) for all t ≥ 0.
− −
4. If x ∈ O≥T (x) with T > 0 then x ∈ O>t (x) for all t ≥ 0.
Proof. We will only prove item (1). If x ∈ O>T +

(x) then there is t1 > T
and u1 ∈ U such that ϕ (t1 , x, u1 ) = x. For any t ≥ 0, there is k ∈ N
such that kt1 > t. By taking the t1 -concatenation u2 of u1 and u1 we have
ϕ (2t1 , x, u2 ) = x. Following by induction we obtain a control function uk
such that ϕ (kt1 , x, uk ) = x. Hence, x ∈ O>t
+
(x).
The properties of the orbits of the system group G are similar to the
properties of the semi-orbits, as in the following:
Proposition 2.2.8 1. G (x), cl (G (x)), and int (G (x)) are invariant sets
for any x ∈ M .
2. If x ∈ G (y) and y ∈ G (z) then x ∈ G (x).
3. φG (x) = Gφ (x) = G (x) for all x ∈ M and φ ∈ G.
4. If int (G(x)) = ∅ then G (x) is an open set.
Proof. Items (1) and (2) are similar to the case of semi-orbit. Item (3) is
obvious since φG = Gφ = G for every φ ∈ G. We then prove item (4). Indeed,
take y ∈ int (G(x)). Then G(y) ⊂ int (G(x)) since int (G(x)) is invariant by
item (1). As G(y) = G(x) by item (3), it follows that G(x) ⊂ int (G(x)), and
therefore G(x) = int (G(x)).
By the Orbit Theorem, the orbits of the system group are submanifolds of
M ([104, 105, 106]). If int (G(x)) = ∅ then G (x) is an open submanifold of M
with dim G(x) = dim M . The semi-orbits need not be submanifolds of M .
We now investigate the orbits of the control system from the flow point
of view. Basically, the trajectories through the points in the fiber π −1 (x) ⊂
M × U project onto the orbit of x. For T ≥ 0 and X ⊂ M × U , we define the
sets
Φ (X) = {Φ (t, x, u) : t ∈ R, (x, u) ∈ X} ,

Φ+
≥T (X) = {Φ (t, x, u) : t ≥ T, (x, u) ∈ X} ,
Φ−
≥T (X) = {Φ (−t, x, u) : t ≥ T, (x, u) ∈ X} ,
Φ+
>T (X) = {Φ (t, x, u) : t > T, (x, u) ∈ X} ,
Φ−
>T (X) = {Φ (−t, x, u) : t > T, (x, u) ∈ X} .
Proposition 2.2.9 Let x ∈ M and t ≥ 0. One has

$ −1 %
1. π Φ+≥T π (x) = O≥T
+
(x).
$ −1 %
2. π Φ−
≥T π (x) −
= O≥T (x).
−1
3. π Φ+
>T π (x) = O>T
+
(x).
− −1 −
4. π Φ>T π (x) = O>T (x).
−1
5. π Φ π (x) = O (x).
$ −1 %
Proof. To prove (1), take y ∈ π Φ+ ≥T π (x) . There is then some
−1
u ∈ U such that (y, u) ∈ Φ+ ≥T π (x) . Hence there are v ∈ U and t ≥ T
such that (y, u) = Φ (t, x, v) = (ϕ (t, x, v) , v · t), and therefore y = ϕ (t, x, v) ∈
O≥T
+
(x). On the other hand, if z ∈ O≥T +
(x) there are then u ∈ U and t ≥ T
−1
such that z = ϕ (t, x, u), hence (z, u · t) = Φ (t, x, u) ∈ Φ+ π (x) . Thus
$ −1 %
≥T
z ∈ π Φ+ ≥T π (x) . This proves (1). Items (2) − (4) can be proved in a
similar way. Item (5) is a consequence of items (1) and (2).
2.3 Critical points

In this section we present various characterizations of critical points for
control affine systems. The results unite and clarify important properties
about equilibria. The idea is to show common properties of control affine
systems and classical dynamical systems.
Section 2.3 · Critical points 55
Definition 2.3.1 A point x ∈ M is called a critical point (or equilibrium

point) with respect to the control function u ∈ U if the trajectory ϕux (R)
reduces to the point x, that is, ϕ (t, x, u) = x for all t ∈ R. The point x ∈ M
is said to be a critical point (or equilibrium point) if it is a critical point with
respect to some control function u ∈ U; x is called a global critical point if
it is a critical point with respect to all u ∈ U.
In other words, a point x ∈ M is critical if and only if there is some

constant u ∈ U such that X (x, u) = 0. Indeed, if x is critical there is then a
function u ∈ U such that ϕ (t, x, u) = x for all t ∈ R. Hence
d
0= ϕ (t, x, u)|t=0 = X (x, u (0)) .
dt
On the other hand, suppose that X (x, u) = 0 for some constant u ∈ U. We

d
have ϕ (t, x, u)|t=0 = X (x, u) = 0, and hence
dt
d d
ϕ (t, x, u) = ϕ (t + s, x, u)|s=0
dt ds
d
= ϕ (t, ϕ (s, x, u) , u)|s=0
ds
d u
= ϕ (ϕ (s, x, u))|s=0
dt t
d
= d (ϕut )x ϕ (s, x, u)|s=0
dt
= 0.
We shall prove other interesting characterizations of critical points. In

order to prove them, we need the following technical lemma.
Lemma 2.3.1 Let x ∈ M such that ϕ (τ , x, u) = x for some τ ∈ R and

u ∈ U. There is then v ∈ U such that ϕ (kτ , x, v) = x for all integers k ∈ Z.
Proof. For each k ∈ N, define the function vk : R → U by
⎧
⎪
⎪ 0, if t ≤ −kτ
⎪
⎪
⎪
⎪ u (t + kτ ) , if − kτ < t ≤ − (k − 1) τ
⎪
⎪
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎨ u (t + τ ) , if −τ <t≤0
vk (t) = u (t) , if 0 < t ≤ τ .
⎪
⎪
⎪
⎪ u (t − τ ) , if τ < t ≤ 2τ
⎪
⎪
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎪ u (t − kτ ) , if kτ < t ≤ (k + 1) τ
⎪
⎩
0, if t > (k + 1) τ
Then vk is given by finitely many concatenations of control functions in U ,

hence vk ∈ U. Now define the function v : R → U by
⎧
⎪
⎪
..
⎪
⎪ .
⎪
⎪
⎪
⎪ u (t + (k + 1) τ ) , if − (k + 1) τ < t ≤ −kτ
⎪
⎪
⎪
⎪ u (t + kτ ) , if − kτ < t ≤ − (k − 1) τ
⎪
⎪
⎪
⎪ .
⎪
⎪ ..
⎪
⎪
⎪
⎪ u (t + τ ) , if − τ < t ≤ 0
⎨
v (t) = u (t) , if 0 < t ≤ τ .
⎪
⎪
⎪
⎪ u (t − τ ) , if τ < t ≤ 2τ
⎪
⎪
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎪ u (t − kτ ) , if kτ < t ≤ (k + 1) τ
⎪
⎪
⎪
⎪ − 1) τ < t ≤ (k + 2) τ
⎪
⎪ u (t (k + 1) τ ) , if (k +
⎪
⎪ .
⎩ ..
This function is periodic with period τ . Indeed we have
⎧
⎪
⎪
..
⎪
⎪ .
⎪
⎪
⎪
⎪ u (t + (k + 2) τ ) , if − (k + 2) τ < t ≤ − (k + 1) τ
⎪
⎪
⎪
⎪ u (t + (k + 1) τ ) , if − (k + 1) τ < t ≤ −kτ
⎪
⎪
⎪
⎪ .
⎪
⎪ ..
⎪
⎪
⎪
⎪ u (t + 2τ ) , if − 2τ < t ≤ −τ
⎨
v · τ (t) = v (τ + t) = u (τ + t) , if − τ < t ≤ 0
⎪
⎪
⎪
⎪ u (t) , if 0 < t ≤ τ
⎪
⎪
⎪
⎪ ..
⎪
⎪ .
⎪
⎪
⎪
⎪ u (t − (k − 1) τ ) , if (k − 1) τ < t ≤ kτ
⎪
⎪
⎪
⎪ − t ≤ (k + 1) τ
⎪
⎪ u (t kτ ) , if kτ <
⎪
⎪ .
⎩ ..
= v (t)
for all t ∈ R, hence v · τ = v. Moreover, the sequence (vk ) converges to v. In

fact, for a given α ∈ L1 (R, Rn ), we have

|fα (v) − fα (vk )| = v (t) − vk (t) , α (t) dt

R

≤ v (t) , α (t) dt
R\[−(k+1)τ ,kτ ]

≤ K α (t) dt
R\[−(k+1)τ ,kτ ]
where K = supw∈U w. Hence |fα (v) − fα (vk )| → 0 as k → +∞, and thus
vk → v in U . Finally, for any natural number k ∈ N, we have
ϕ (kτ , x, v) = ϕ ((k − 1) τ + τ , x, v)
= ϕ ((k − 1) τ , ϕ (τ , x, v) , v · τ )
= ϕ ((k − 1) τ , ϕ (τ , x, u) , v)
= ϕ ((k − 1) τ , x, v)
and
ϕ (−kτ , x, v) = ϕ (− (k − 1) τ − τ , x, v)
= ϕ (− (k − 1) τ , ϕ (−τ , x, v) , v · (−τ ))
= ϕ (− (k − 1) τ , ϕ (τ , x, u) , v)
= ϕ (− (k − 1) τ , x, v) .
Following by induction we obtain ϕ (±kτ , x, v) = x, as desired.
The following theorem contains various equivalences of critical point:
Theorem 2.3.1 For any x ∈ M , the following sentences are equivalent:
1. x is a critical point.
2. There is u ∈ U such that ϕux (R) = {x}.
3. There is u ∈ U such that ϕux (R+ ) = {x}.
4. There is u ∈ U such that ϕux (R− ) = {x}.
5. There is u ∈ U such that ϕux ([a, b]) = {x} for some a < b.
6. There is u ∈ U and a sequence (tn ) in R such that tn > 0, tn → 0, and

ϕ (tn , x, u) = x for all n ∈ N.
Proof. The equivalence (1) ⇔ (2) and the implications (2) ⇒ (3) ⇒ (5),
(2) ⇒ (4), and (2) ⇒ (6) are clear. We then first prove (5) ⇒ (2). Suppose
that ϕux ([a, b]) = {x} for some a < b and u ∈ U . We may assume that
a = 0 since ϕu·ax ([0, b − a]) = {x}. Then suppose ϕx ([0, τ ]) = {x}. Take the
u
function v : R → U as in Lemma 2.3.1. If kτ < t ≤ (k + 1) τ , with k ∈ Z, we

have 0 < t − kτ ≤ τ . Since ϕ (kτ , x, v) = x by the lemma, it follows that
ϕ (t, x, v) = ϕ (t − kτ + kτ , x, v) = ϕ (t − kτ , ϕ (kτ , x, v) , v · kτ )
= ϕ (t − kτ , x, v)
= ϕ (t − kτ , x, u)
= x
and therefore ϕvx (R) = {x}. The implication (4) ⇒ (2) is analogously proved,
and we then have the equivalence of (1) with (2) − (5). Finally, we prove
(6) ⇒ (2). Suppose that (6) holds and choose τ > 0 such that tn ≤ τ for all
n. Take the function v : R → U as in Lemma 2.3.1. We first assume t = ktn
for some integers k and n. For k = 1, we have ϕ (tn , x, v) = ϕ (tn , x, u) = x.
Following by induction, suppose that ϕ (ktn , x, v) = x for some k > 1. We
then have
ϕ ((k + 1) tn , x, v) = ϕ (tn , ϕ (ktn , x, v) , v · ktn ) = ϕ (tn , x, v) = x.
Hence ϕ (ktn , x, v) = x for all positive integer k. For a negative integer k, we

have
ϕ (ktn , x, v) = ϕ (ktn , ϕ (−ktn , x, v) , v · (−ktn )) = ϕ (0, x, v) = x.
Thus ϕ (ktn , x, v) = x for all integers k and n. Now, let t ∈ R with t = ktn
for all integers k and n. For each integer n we can take an integer kn such
that kn tn < t < (kn + 1) tn . Furthermore there is an integer m > n such
that kn tn < km tm < t < (km + 1) tm < (kn + 1) tn . It follows that kn tn → t,
and then ϕ (kn tn , x, v) → ϕ (t, x, v) by continuity. Since ϕ (kn tn , x, v) = x, we
have ϕ (t, x, v) = x. Thus ϕvx (R) = {x}.
Next we present more theorems on critical points.
Theorem 2.3.2 For any x ∈ M , the following statements are equivalent:
1. x is critical.
2. There is u ∈ U such that every neighborhood of x contains a positive

semi-trajectory with respect to u.
3. There is u ∈ U such that every neighborhood of x contains a negative

semi-trajectory with respect to u.
Proof. The implications (1) ⇒ (2) and (1) ⇒ (3) are clear, since {x}
is a trajectory whenever x is critical. To prove the implication (2) ⇒ (1),
assume that every neighborhood of x contains a positive semi-trajectory with
respect to u ∈ U . Suppose by contradiction that x is not critical. Accord-
ing to Theorem 2.3.1, this means that ϕux (R+ ) = {x}, hence there is t > 0
such that ϕ (t, x, u) = x. Let U, V ⊂ M be disjoint open sets such that
x ∈ U and ϕ (t, x, u) ∈ V . Since the transition map ϕut is a homeomor-
phism of M , it follows that ϕut (U ) is an open set containing ϕ (t, x, u). Define
W1 = V ∩ ϕut (U ) and W2 = ϕu·t −t (V ∩ ϕt (U )). Then W1 ⊂ V and W2 ⊂ U ,
u
and hence W1 ∩ W2 = ∅, ϕ (t, x, u) ∈ W1 and x ∈ W2 . Now, if y ∈ W2 then

ϕ (t, y, u) ∈ ϕut (W2 ) ⊂ V , hence the neighborhood W2 of x can not contain
a positive semi-trajectory with respect to u. This contradiction implies that
ϕ (t, x, u) = x for all t ∈ R+ , and therefore x is critical according to Theorem
2.3.1. Similarly we can prove the implication (3) ⇒ (1).
This theorem yields the following criterium for critical points:
Corollary 2.3.1 Let x ∈ M . If lim d (ϕ (t, y, u) , x) = 0 for some y ∈ M

t→+∞
and u ∈ U then x is critical. The same holds if lim d (ϕ (t, y, u) , x) = 0.
t→−∞
Proof. Suppose lim d (ϕ (t, y, u) , x) = 0 and let U be a neighborhood

t→+∞
of x. There is T ≥ 0 such that ϕ (t, y, u) ∈ U whenever t ≥ T . Hence,
for any t > 0, we have ϕ (t, ϕ (T, y, u) , u · T ) = ϕ (t + T, y, u) ∈ U , and
therefore U contains the positive semi-trajectory through ϕ (T, y, u) with re-
spect to u · T ∈ U. According to Theorem 2.3.2, x is critical. The case
lim d (ϕ (t, y, u) , x) = 0 follows analogously.
t→+∞
2.4 Periodic points

We shall now unite and explain the main properties of periodic points for
control affine systems. The intention is to show how periodicity of control
systems approaches periodicity of dynamical systems.
Definition 2.4.1 A point x ∈ M is called periodic with respect to the control

u ∈ U if the motion ϕux (or the trajectory ϕux (R)) is periodic, that is, there is
a τ = 0 such that ϕux (t) = ϕux (t + τ ) for all t ∈ R. In this case, the number τ
is called a period of x. The point x ∈ M is said to be periodic if it is periodic
with respect to some control function; x is called a global periodic point if
it is periodic with respect to all control functions.
Remark 2.4.1 If τ is a period of x with respect to u ∈ U and k is an integer

number, we have
ϕ (t + kτ , x, u) = ϕ (t + (k − 1) τ + τ , x, u) = ϕ (t + (k − 1) τ , x, u) .
Following by induction we have ϕ (t + kτ , x, u) = ϕ (t, x, u) for all t ∈ R.

Hence kτ is a period of x for all integer k = 0.
Section 2.4 · Periodic points 61
The following characterizations of periodic point are very useful.
Theorem 2.4.1 Let x ∈ M . The following sentences are equivalent:
1. x is a periodic point.
2. x ∈ O>0
+
(x).
−
3. x ∈ O>0 (x).
4. x ∈ O>T
+
(x) for all T ≥ 0.
5. x ∈ O>T (x) for all T ≥ 0.
Proof. Note that (2) ⇔ (3), (4) ⇔ (5), (1) ⇒ (2), and (4) ⇒ (2) are clear,
and (2) ⇒ (4) follows by Remark 2.4.1. We then show (2) ⇒ (1). Indeed,
suppose that ϕuτ (x) = x with u ∈ U and τ > 0. Define the function v : R → U
as in Lemma 2.3.1. We have ϕ (τ , x, v) = x, and v · τ = v. For any t ∈ R, it
follows that
ϕ (t + τ , x, v) = ϕ (t, ϕ (τ , x, v) , v · τ ) = ϕ (t, x, v) .
Hence x is periodic with respect to v.
The condition (2) in Theorem 2.4.1 is the usual definition of periodic point
for control systems (e.g., [85]). Below we study other properties of periodic
points.
Theorem 2.4.2 Let x ∈ M be not critical. If x is periodic with respect to

u ∈ U then there is τ > 0 such that τ is the smallest positive period of x with
respect to u. Moreover if τ is any other period of x then τ = kτ for some
integer k.
Proof. Set P = {t > 0 : t is a period of x with respect to u}. According

to Remark 2.4.1, P is nonempty. Define τ = inf (P ). We claim that τ > 0 and
τ is a period of x. Indeed, we have τ ≥ 0 and there is a sequence (tn ) in P such
that tn → τ . Since ϕ (tn , x, u) = x for all n, it follows that ϕ (τ , x, u) = x
by continuity. As x is not critical, Theorem 2.3.1 implies that τ > 0, and
therefore τ is a period of x. Now let τ be any period of x with respect to
u. Suppose by contradiction that τ = kτ for all integer k. There is then an
integer k with kτ < τ < (k + 1) τ . This means that 0 < τ − kτ < τ and, for
any t ∈ R, we have
ϕ (t + τ − kτ , x, u) = ϕ (t + τ , x, u) = ϕ (t, x, u)
since both τ and −kτ are periods of x. Hence τ − kτ is a positive period of

x with respect to u and τ − kτ < τ , which contradicts the fact that τ was
the smallest positive period of x. Thus τ = kτ for some integer k.
Because of this theorem, the smallest positive period of a periodic point

x with respect to u ∈ U is called the fundamental period of x relative to
the control u. Evidently, if x is also periodic with respect to another control
function v the fundamental period of x relative to v may be different from the
fundamental period of x relative to u.
Example 2.4.1 For the control affine systems on the 2-sphere and the 2-
torus defined respectively in Examples 1.3.1 and 1.3.2, all points are periodic.
Example 2.4.2 Consider the control system on the plane as defined in

Example 2.2.2. The origin is a critical point and all other points are peri-
odic.
Figure 2.6: Trajectories with controls u = (0, 1).
ẋ = u1 (t) X1 (x) + u2 (t) X2 (x)

Section 2.4 · Periodic points 63
Figure 2.7: Trajectories with controls u = (u1 , u2 ) with 0 < u1 < 1.
on R2 , with control range

U = (u1 , u2 ) ∈ R2 : u1 + u2 = 1, u1 ∈ [0, 1]
where

X1 (x1 , x2 ) = x2 , x1 − x21 , X2 (x1 , x2 ) = x2 , x1 − x31 .
The origin 0 is a global hyperbolic critical point and (1, 0) is a global stable
1
critical point. The other critical points are , 0 , with 0 ≤ p < 1.
p−1
The set of trajectories with respect to the constant control u = (u1 , u2 ), with
0 ≤ u1 < 1, contains two tear drops, which are invariant sets bounded by
orbits homoclinic
to the origin 0, and filled out with periodic solutions around,
1
respectively, , 0 and (0, 1). The exterior of the set formed by the
u1 − 1
union of these two tear drops is also filled out with periodic solutions (see Figs.
2.6 and 2.7). The set of trajectories with respect to the constant control u =
(1, 0) contains a tear drop in the right half plane {(x1 , x2 ) : x1 ≥ 0} bounded
by orbit homoclinic to the origin 0 and filled out with periodic solutions around
(0, 1), but it does not contain periodic solutions outside the tear drop (see Fig.
2.8). Thus the dynamics of the system has a bifurcation at the parameter
u1 = 1. Now, for each constant control u ∈ U, denote by Tu the tear drop in

the right half plane with respect to u, and then set T = u∈U Tu . Any two
Figure 2.8: Trajectories with control u = (1, 0).
periodic trajectories around T can be connected by using the control u = (1, 0)

in the left half plane. Furthermore, any point outside T can be attainable to
a periodic trajectory around T . Thus M \ T ⊂ O>0 +
(x) for every x ∈ M \ T .
Fig. 2.9 illustrates a trajectory connecting two points outside T .
Figure 2.9: Trajectory connecting two points outside the set T .
The following relates critical and periodic points.

Proposition 2.4.1 Let (xn ) be a sequence in M such that each xn is a pe-
riodic point with respect to a control function un ∈ U. If xn → x and their
positive period τ n → 0 then x is a critical point.
Proof. We may assume that un → u in U . For a given t ∈ R, there are
integers kn such that
k n τ n ≤ t < kn τ n + τ n .
As τ n → 0, we have kn τ n → t. According to 2.4.1, ϕ (kn τ n , xn , un ) = xn ,

and then ϕ (t, x, u) = limn→∞ ϕ (kn τ n , xn , un ) = x. Thus x is critical.
We now use the control flow to provide other characterizations of critical

and periodic points.
Proposition 2.4.2 Let x ∈ M . Then
1. x is a critical point of the control system if and only if there is a control

function u ∈ U such that (x, u) is a critical point for the control flow.
2. x is a periodic point of the control system if and only if there is a control

function u ∈ U such that (x, u) is a periodic point for the control flow.
Proof. (1) Suppose that (x, u) is a critical point of the control flow Φ.
Then Φ (t, x, u) = (x, u) for all t ∈ R. This means that ϕ (t, x, u) = x for all
t ∈ R, and hence x is a critical point. On the other hand, suppose that x is a
critical point with respect to the control function u ∈ U . Take a sequence of
positive numbers τ n → 0. We have ϕ (τ n , x, u) = x, for every n. According to
Theorem 2.4.1, x is periodic with respect to a control function un ∈ U with
period τ n . Assume that un → u in U . For any integer k, we have un ·kτ n = un
for every n. Now, for a given t ∈ R, there are integers kn such that kn τ n → t.
We then have
Φ (t, x, u) = (ϕ (t, x, u) , u · t) = lim (ϕ (kn τ n , x, un ) , un · kn τ n )

n→∞
= lim (x, un )
n→∞
= (x, u) .
Hence (x, u) is a critical point of the control flow.

(2) It follows by Proposition 2.2.9 and Theorem 2.4.1.
2.5 Problems
1. Prove that a set X is positively invariant, negatively invariant, or invari-
ant, if and only if, for each x ∈ M , respectively, SX = X, S −1 X = X,
or GX = X.
2. Let X ⊂ M . Show that the following sentences are equivalent:

(a) X is invariant.
(b) ϕut (X) = X for all u ∈ U and t ∈ R.
(c) φ (X) = X for all φ ∈ G.
3. Let X ⊂ M be a closed set. Show that X is positively invariant, neg-

atively invariant, or invariant, if and only if, respectively, ϕut (X) ⊂ X
for all u ∈ Upc and t > 0, ϕut (X) ⊂ X for all u ∈ Upc and t < 0, or
ϕut (X) ⊂ X for all u ∈ Upc and t ∈ R.
4. Give an example of a positively invariant set that is not negatively
invariant.
5. Show that the set of all critical points of a control affine system is closed.
6. Let x be a periodic point with respect to a control function u ∈ U. Show
that ϕux (R) = ϕux ([0, T ]) for some T > 0. Does the converse hold?
7. Given any τ > 0, prove that the set of all periodic points with positive
period T ≤ τ is closed.
8. Consider the control affine system ẋ = u (t) on Rn . Show that the
system has periodic point if and only if 0 ∈ U. (It should be remembered
that we assume convex control range.)
9. Prove that X ⊂ M is invariant, positively invariant, or negatively in-
variant if and only if each of its connected components has the same
property.
10. Let H (M ) be the hyperspace of all compact subsets of a manifold M
endowed with the Hausdorff distance
!
dH (A, B) = max sup d (a, B) , sup d (b, A)
a∈A b∈B
where d (y, X) = inf d (y, x). For each t ∈ R+ and x ∈ M , define the
x∈X
sets
Ot+ (x) = {y ∈ M : y = ϕ (t, x, u) for some u ∈ U} ,
Ot− (x) = {y ∈ M : y = ϕ (−t, x, u) for some u ∈ U} .
(a) Prove that Ot+ (x) and Ot− (x) are compact sets in M .
(b) Show that the functions O+ , O− : R+ × M → H (M ), given respec-
tively by O+ (t, x) = Ot+ (x) and O− (t, x) = Ot+ (x), are continu-
ous.
Section 2.6 · Notes and references 67

More concepts and results on control systems may be found in Colonius–
Kliemann [34], Jurdjevic [63], and Sontag [92]. In these books, one can find
many other problems in different lines of investigation, for instance, bifurca-
tion, linearization, optimal control, spectrum, and stabilization.
Chapter 3
Minimal sets and

transitivity
In this chapter we study an important class of invariant set and show how
the orbits of a control system yield a partition of the state space. We deal
with the notions of minimal set and transitive set, which are regions of the
state space where the control system is transitive. The transitivity is based
on the relations among the positive semi-orbits. A transitive set is a region
of the state space where each state x can be attained by any other state y
in positive time. The control system is called transitive if O+ (x) coincides
with the whole state space, which means that each state x can be attainable
forward to any other state y. In classical control theory, this property is
usually called complete controllability. In general, however, we can not expect
transitivity on the whole state space. We then investigate a weaker notion of
transitivity and ask the basic question: In which regions of the state space
is the system weak transitive? A weak transitive set is a region where each
state x can be closely attained by any other state y in positive time. The
maximal weak control sets decompose the state space in equivalence classes
of the weak transitivity relation. Besides, they are ordered by a dynamic
order relation that has the positive minimal sets as upper bounds. In Section
3.1, we introduce the notions of minimal sets, which are very special subsets
of the state space satisfying invariance properties. In Section 3.2, we define
the weak transitivity relation and show that their equivalence classes form
maximal regions of transitivity. In the last section, we define the notion
69
70 Chapter 3 · Minimal sets and transitivity
of a domain of attraction of the maximal regions of transitivity and study a

dynamic order among them. We show that the minimal sets are upper bounds
for this order. We only consider transitivity with respect to positive semi-
trajectories. Otherwise, transitivity with respect to negative semi-trajectory
may be studied in the same way by using the time-reversed system.
3.1 Minimal sets

We start by defining the notions of minimal sets, which are minimal invari-
ant sets for the control system.
Definition 3.1.1 Let X ⊂ M be a nonempty subset.
1. X is said to be a positive minimal set if it is closed, positively invari-

ant, and does not contain any nonempty proper subset that is closed
and positively invariant.
2. X is said to be a negative minimal set if it is closed, negatively

invariant, and does not contain any nonempty proper subset that is
closed and negatively invariant.
3. X is called a minimal set if it is closed, invariant, and does not contain

any nonempty proper subset that is closed and invariant.
The following theorem presents the asymptotic property of the minimal

sets.
Theorem 3.1.1 Let X ⊂ M be a nonempty subset. Then

+
1. X is a positive minimal set if and only if X = cl (O+ (x)) = cl O>T (x)
for every x ∈ X and T ≥ 0.
−
2. X is a negative minimal set if and only if X = cl (O− (x)) = cl O>T (x)
for every x ∈ X and T ≥ 0.
3. X is a minimal set if and only if X = cl (G (x)) for every x ∈ X.
Proof. We first prove item (3). Suppose that X is a minimal set. As

X is invariant, we have G (x) ⊂ X for every x ∈ X. Since X is closed, it
follows that cl (G (x)) ⊂ X for every x ∈ X. According to Proposition 2.2.8,
cl (G (x)) is invariant. By the minimality of X, it follows that cl (G (x)) = X
Section 3.1 · Minimal sets 71
for every x ∈ X. As to the converse, suppose that X = cl (G (x)) for every

x ∈ X. According to Proposition 2.2.8, X is closed and invariant. Let Y ⊂ X
be a nonempty subset that is closed and invariant. For a given x ∈ Y , we
have X = cl (G (x)) ⊂ Y , hence Y = X. Thus X is a minimal set. We now
prove item (1). Suppose that X is a positive minimal set. As X is positively
invariant, we have O>T+
(x) ⊂ O+ (x) ⊂ X for every x ∈ X and T ≥ 0.
+
Since X is closed, it follows that cl O>T (x) ⊂ cl (O+ (x)) ⊂ X for every
+
x ∈ X and T ≥ 0. According to Proposition 2.2.1, cl O>T (x) is positively
+
invariant. By the minimality of X, it follows that cl O>T (x) = X, and
+
therefore X = cl (O+ (x)) = cl O>T (x) for every x ∈ X and T ≥ 0. As
to the converse, suppose that X = cl (O+ (x)) for every x ∈ X. According
to Proposition 2.2.1, X is closed and positively invariant. Let Y ⊂ X be a
nonempty subset that is closed and positively invariant. For a given x ∈ Y ,
we have X = cl (O+ (x)) ⊂ Y , hence Y = X. Thus X is a positive minimal
set. The proof for item (2) is similar.
According to Propositions 2.2.3 and 2.2.4, positive minimal sets, negative

minimal sets, and minimal sets are connected.
The following presents a sufficient condition for the existence and unique-
ness of a minimal set.

Theorem 3.1.2 1. If X = cl (O+ (x)) is nonempty then X is the
x∈M
unique positive minimal set.

2. If X = cl (O− (x)) is nonempty then X is the unique negative min-
x∈M
imal set.

3. If X = cl (G (x)) is nonempty then X is the unique minimal set.
x∈M
Proof. The proofs for the three items are similar. We will only prove

item (1). Suppose that X = cl (O+ (x)) = ∅. It is immediate that X is
x∈M
closed. According to Propositions 2.2.3 and 2.2.4, X is positively invariant.
Hence O+ (x) ⊂ X for every x ∈ X. Since X ⊂ O+ (x) for every x ∈ X, it
follows that cl (O+ (x)) = X for every x ∈ X. According to Theorem 3.1.1,
X is a positive minimal set. Now, suppose that Y ⊂ M is a positive minimal
set. Then Y = cl (O+ (y)) for every y ∈ Y , which implies that X ⊂ Y . Hence
Y = X, and therefore X is the unique positive minimal set of the system.
Compactness is other sufficient condition for the existence of minimal sets.

This is a particular case of a famous theorem proved by R. Ellis [48] in the
general setting of semigroup actions.
Theorem 3.1.3 Assume that M is a compact manifold. There is then a
minimal set, a positive minimal set, and a negative minimal set for the control
system.
Proof. We only prove the existence of a positive minimal set. For indeed,
let C be the collection of nonempty, closed, positively invariant subsets in M .
As M ∈ C, we have C = ∅. Consider C partially ordered downward by the set
inclusion. Let {Cα }γ∈Γ ⊂ C be a totally ordered subcollection. Since M is

compact, we have X = Cγ = ∅. According to Proposition 2.1.1, X is posi-
γ∈Γ
tively invariant. Hence X ∈ C, and therefore any totally ordered subcollection
of C has a lower bound in C. Then the Zorn lemma guarantees the existence
of a minimal element of C. Let X ∈ C be such a minimal element. If Y ⊂ X
is nonempty, closed, and positively invariant, then Y ∈ C. By the minimal-
ity of X in C, it follows that Y = X. Therefore X is a positive minimal set.
This result can be straightforwardly extended in the following way.

Corollary 3.1.1 Let X ⊂ M be a nonempty compact subset.
1. If X is positively invariant then there is a positive minimal set contained
in X.
2. If X is negatively invariant then there is a negative minimal set con-
tained in X.
3. If X is invariant then there is a minimal set contained in X.
Proof. Repeat the proof for Theorem 3.1.3 restricted to X.
In particular, if cl (O+ (x)) is compact then there is a positive minimal

set in cl (O+ (x)); if cl (O− (x)) is compact then there is a negative minimal
set in cl (O− (x)); and if cl (G (x)) is compact then there is a minimal set in
cl (G (x)). The property of compact closure of semi-orbit is known as Lagrange
stability and plays an important role in studies of minimal sets (see the notes
and Problem 5.8(60)).
Theorems 3.1.2 and 3.1.3 together yield a necessary and sufficient condition
for the existence and uniqueness of minimal sets.
Section 3.1 · Minimal sets 73
Theorem 3.1.4 Assume that M is a compact manifold.
1. There is a unique positive minimal set if and only if the intersection

cl (O+ (x)) is nonempty. In this case, cl (O+ (x)) is the unique
x∈M x∈M
positive minimal set.
2. There is a unique negative minimal set if and only if the intersection

cl (O− (x)) is nonempty. In this case, cl (O− (x)) is the unique
x∈M x∈M
negative minimal set.
3. There is a unique minimal set if and only if the intersection

cl (G (x)) is nonempty. In this case, cl (G (x)) is the unique
x∈M x∈M
minimal set.
Proof. We only prove item (1). Suppose that X ⊂ M is the unique

positive minimal set of the system. For any x ∈ M , the set cl (O+ (x)) is
compact and positively invariant. According to Theorem 3.1.3, there is a pos-
itive minimal set Px ⊂ cl (O+ (x)). By uniqueness, Px = X for all x ∈ M .

Hence X ⊂ cl (O+ (x)), and therefore X = cl (O+ (x)) according to
x∈M x∈M
Theorem 3.1.2. The converse follows according to Theorem 3.1.2.
Example 3.1.1 Let the unit circle S1 be identified with the interval [0, 2π).
Consider the control affine system ẋ = u1 (t) sin x + u2 (t) cos x sin x on S1

with control range U = (u1 , u2 ) ∈ R2 : u1 + u2 = 1, u2 ∈ [1/2, 1] . Fig. 3.1
illustrates the trajectories of the system with respect to constant controls u2 =
1/2, 1/2 < u2 < 1, and u2 = 1. The points 0 and π in S1 are global critical
points. Thus {0} and {π} are minimal sets. All points in the set [π/2, 3π/2] ⊂
S1 are critical. Both upper and lower semicircles are compact invariant sets,
but they are not minimal. Every set [θ, π] with θ ≤ π/2 is compact and
positively invariant but not negatively invariant. The sets [π, θ], with θ ≥
3π/2, have the same property.
It is clear that compactness is not a necessary condition for the existence

of minimal sets. In fact, any periodic trajectory for a dynamical system is a
minimal set even if the state space is not compact.
The control system is called minimal if the whole state space M is itself a
minimal set. In this case, M is the unique minimal set and coincides with its
interior set. A partial converse for this fact holds, as in the following:
Figure 3.1: From left to right, trajectories with controls u2 = 1/2, 1/2 < u2 <
1, and u2 = 1.
Proposition 3.1.1 If X ⊂ M is a minimal set with nonvoid interior then X

is open in M .
Proof. Suppose that X ⊂ M is a minimal set with int (X) = ∅ and

take any x ∈ X. According to Theorem 3.1.1, we have X = cl (G (x)), and
then int (X) ∩ G (x) = ∅. Hence there is a homeomorphism φ ∈ G such that
φ (x) ∈ int (X). Since X is invariant, we have

x ∈ φ−1 (int (X)) = int φ−1 (X) = int (X)
and therefore X = int (X).
Thus a minimal set X ⊂ M with nonvoid interior is open and closed in M .

If M is a connected manifold, it follows that X = M . The reader is invited
to prove that the control systems in Examples 1.3.2 and 1.3.1 are minimal.
3.2 Weak control sets

We now define the notion of weak transitivity and study the maximal re-
gions of M where the control system is weak transitive.
In the following we define the weak transitivity relation.
Definition 3.2.1 We say that two points x, y ∈ M are weak equivalent if

+ +
x ∈ cl O>0 (y) and y ∈ cl O>0 (x) .
+
According to Proposition 2.2.1, the set cl O>0 (x) is positively invari-
+ +
ant for every x ∈ M . Hence x ∈ cl O>0 (y) and y ∈ cl O>0 (x) implies
+ + + +
cl O>0 (x) ⊂ cl O>0 (y) and cl O>0 (y) ⊂ cl O>0 (x) .
Section 3.2 · Weak control sets 75
Proposition 3.2.1 The weak transitivity relation is an equivalence relation.
Proof. The weak transitivity relation is clearly reflexive and symmet-

ric. Besides, if x is weak equivalent to y and y is weak equivalent to z then
+ + +
cl O>0 (x) = cl O>0 (y) = cl O>0 (z) , and hence x is equivalent to z.
Definition 3.2.2 A nonempty subset X ⊂ M is said to be a weak transitive

+
set if X ⊂ cl O>0 (x) for all x ∈ X.
For each x ∈ M , we denote by Dx the equivalence class of x with respect

to the weak transitivity relation. The following theorem characterizes the sets
Dx .
Proposition 3.2.2 Let D ⊂ M be a nonempty set. D is then an equivalence

+
class of the weak transitivity relation if and only if D ⊂ cl O>0 (x) for all
x ∈ D and D is maximal satisfying this property.
Proof. Suppose that D = Dz for some z ∈ M . It is immediate that

+
D ⊂ cl O>0 (x) for all x ∈ D. Let D ⊂ M such that D ⊂ D and D ⊂
+ +
cl O>0 (x) for all x ∈ D . For any x ∈ D , we have x ∈ cl O>0 (z) and
+
z ∈ cl O>0 (x ) , that is, x and z are weak equivalent. Hence x ∈ Dz = D,
and therefore D ⊂ D, proving that D is maximal. As to the converse, sup-
+
pose that D ⊂ cl O>0 (x) for all x ∈ D and D is maximal satisfying this
+ +
property. For any x, y ∈ D, we have x ∈ cl O>0 (y) and y ∈ cl O>0 (x) ,
hence x is weak equivalent to y. By taking x ∈ D, it follows that D ⊂ Dx .
Now, for any z ∈ Dx , we have Dx ⊂ O>0 +
(z). By the maximality of D, it
follows that D = Dx .
Each equivalence class Dx is called a weak control set. Note that each
weak transitive set is contained in a weak control set and the state space M
decomposes into weak control sets.
Example 3.2.1 Consider the control affine system on S1 as defined in Exam-

ple 3.1.1. Both minimal sets {0} and {π} are compact invariant weak control
sets. The other weak control sets are [π/2, π) and (π, 3π/2].
The following result states the non-regressing property of trajectories start-

ing at weak control sets.
Proposition 3.2.3 Let D ⊂ M be a weak control set. Assume that for a

given point x ∈ D there are T > 0 and u ∈ U such that ϕ (T, x, u) ∈ D. Then
ϕ (t, x, u) ∈ D for all t ∈ [0, T ].
Proof. Denote D = D ∪ {ϕ (t, x, u)} with 0 ≤ t ≤ T . We show

+
that D satisfies D ⊂ cl O>0 (y) for every y ∈ D . In fact, we have
+
ϕ (t, x, u) ∈ O>0
+
(x) ⊂ cl O>0
+
(x) . For any y ∈ D, we have x ∈ cl O>0 (y) .
+
According to Proposition 2.2.6, it follows that ϕ(t, x, u) ∈ cl O>0 (y) . Since
+ +
D ⊂ cl O>0 (y) , we have D ⊂ cl O>0 (y) for every y ∈ D. Now, as
+
ϕ (T, x, u) ∈ D, we have y ∈ cl O>0 (ϕ (T, x, u)) for any y ∈ D. Since t ≤ T ,
there is τ ≥ 0 such that T = t + τ . We then have
O>0
+
(ϕ(T, x, u)) = O>0
+
(ϕ(t + τ , x, u))
= O>0
+
(ϕ(τ , ϕ(t, x, u), u · t))
⊂ O>0
+
(ϕ(t, x, u))
+ +
which implies cl O>0 (ϕ(T, x, u)) ⊂ cl O>0 (ϕ(t, x, u)) . It follows that
+ +
y ∈ cl O>0 (ϕ(t, x, u)) , for any y ∈ D. Moreover, as ϕ (t, x, u) ∈ cl O>0 (y)
+ +
and y ∈ cl O>0 (ϕ(t, x, u)) , we have ϕ (t, x, u) ∈ cl O>0 (ϕ(t, x, u)) , ac-

cording to Proposition 2.2.6. Hence D ⊂ cl O>0 +
(ϕ(t, x, u)) , and therefore

+
D ⊂ cl O>0 (y) for every y ∈ D . By the maximality of D, it follows that
D = D, and therefore ϕ (t, x, u) ∈ D for all t ∈ [0, T ].
Thus a positive semi-trajectory that starts at a weak control set, and goes
out of this set, can not return to it.
3.3 Dynamic order

In this section we define the notion of a domain of attraction and study a
dynamic order among the weak control sets. We think about the possibility
+
of the closure cl O>0 (x) to intersect a weak control set D with x ∈ / D.
Such a point x is said to be weakly attracted to D. The following definition
establishes this notion of weak attraction.
Definition 3.3.1 Let D be a weak control set. The domain of weak at-
traction of D is defined by
+
Aw (D) = x ∈ M : cl O>0 (x) ∩ D = ∅ .
Section 3.3 · Dynamic order 77
By means of weak attraction, we define the following relation among the

weak control sets: for two weak control sets D1 and D2
D1 w D2 if and only if D1 ∩ Aw (D2 ) = ∅.

+
It is immediate that D ⊂ Aw (D) since D ⊂ cl O>0 (x) for every x ∈ D.
This property is satisfied by any point in Aw (D), as the following.
Proposition 3.3.1 Let D ⊂ M be a weak control set. Then Aw (D) =

+
x ∈ M : D ⊂ cl O>0 (x) .
+
Proof. The inclusion x ∈ M : D ⊂ cl O>0 (x) ⊂ Aw (D) is clear.
+
Then let x ∈ Aw (D) and take y ∈ cl O>0 (x) ∩ D. For any z ∈ D, we
+ +
have z ∈ cl O>0 (y) , according to Proposition 3.2.2. Hence z ∈ cl O>0 (x) ,
+
according to Proposition 2.2.6, and therefore D ⊂ cl O>0 (x) .
Thus, for two weak control sets D1 and D2 , we have D1 w D2 if and only
+ +
if there is x ∈ D1 such that D2 ⊂ cl O>0 (x) , which implies D2 ⊂ cl O>0 (x)
for all x ∈ D1 .
Proposition 3.3.2 The relation “w ” among the weak control sets is a pre-
order.
+
Proof. Let D ⊂ M be a weak control set. Then D ⊂ cl O>0 (x) for all
x ∈ D, hence D w D. Now, suppose that D1 , D2 , D3 are weak control sets
with D1 w D2 and D2 w D3 . We show that D1 w D3 . Indeed, we have
+ +
D2 ⊂ cl O>0 (x) , for all x ∈ D1 , and D3 ⊂ cl O>0 (y) for all y ∈ D2 . Ac-
+
cording to Proposition 2.2.6, it follows that D3 ⊂ cl O>0 (x) , for all x ∈ D1 .
Hence D1 w D3 .
The preorder “w ” is called the weak preorder. We can present upper
bounds for this preorder. We need the following characterization of positive
a minimal set.
Theorem 3.3.1 A set X ⊂ M is a positive minimal set if and only if it is a

closed and positively invariant weak control set.
Proof. Let X ⊂ M be a positive minimal set. Then X is closed and pos-

+
itively invariant. According to Proposition 3.1.1, we have X = cl O>0 (x)
for every x ∈ X. Suppose that Y is a set containing X and satisfying Y ⊂
+ +
cl O>0 (y) for every y ∈ Y . For x ∈ X ⊂ Y , we have Y ⊂ cl O>0 (x) = X.
Hence X = Y , and therefore X is a weak control set according to Theorem
5.7.2. On the other hand, suppose that X is a closed and positively invariant
weak control set. If Y ⊂ X is nonempty, closed, and positively invariant,
+
then cl O>0 (x) ⊂ Y for any y ∈ Y . As X is weak transitive, it follows that
+
X ⊂ cl O>0 (y) ⊂ Y for any y ∈ Y . Hence X = Y and therefore X is a
positive minimal set.
Moreover, for a positive minimal set X, Proposition 3.1.1 says that X =

cl O>T
+
(x) for every x ∈ X and T ≥ 0. This reveals the asymptotic property
of minimal sets, which will be studied in the Chapter 5.
We now show that positive minimal sets are upper bounds for the weak
preorder.
Proposition 3.3.3 If D is a positive minimal set then D is maximal with

respect to the weak preorder among the weak control sets.
Proof. Suppose that D w D for some weak control set D . There is

+
then x ∈ D such that D ⊂ cl O>0 (x) = D. Hence D = D , and therefore
D is maximal with respect to the weak preorder.
In a compact state space, the converse to Proposition 3.3.3 holds, that is,
the positive minimal sets cover the upper bounds for the weak preorder.
Theorem 3.3.2 Assume that M is a compact manifold and let X ⊂ M be a

nonempty subset. Then X is a positive minimal set if and only if X is a weak
control set that is maximal with respect to the weak preorder.
Proof. Suppose that X is a weak control set that is maximal with

+
respect to the weak preorder and take x ∈ X. Since cl O>0 (x) is compact
+
and positively invariant, there is a positive minimal set N ⊂ cl O>0 (x) .
According to Theorem 3.3.1, N is a weak control set and X w N . Since X
is maximal with respect to the weak preorder, we have X = N . Thus X is
a positive minimal set. The converse follows according to Proposition 3.3.3.
Example 3.3.1 Let us explore Example 3.1.1 again. We have four weak
control sets: D1 = {0}, D2 = {π}, D3 = [π/2, π), D4 = (π, 3π/2]. The
minimal sets D1 and D2 are the upper bounds for the weak preorder. Note
that Aw (D2 ) = (0, 2π), hence D3 w D2 and D4 w D2 .
3.4 Problems
+
1. Give an example of a control system in what Dx = cl O>0 (x) ∩
−
cl O>0 (x) .
2. Consider the control affine system ẋ = u (t) on the plane R2 with control
range U = [1, 2] × [−1, 1]. Show that the weak control sets satisfy
−
D = O>0
+
(x) ∩ O>0 (x) for every x ∈ D. Describe the dynamic preorder
among the weak control sets.
3. Let M = R (mod 3) × R and consider an affine control system
ẋ (t) = X0 (x (t)) + u (t) X1 (x (t))
on M , with control range U = [0, 1], where

x2 , x22 , for x2 ≥ 0
X0 (x) =
(0, 0) , for x2 < 0
and X1 (x) = (0, f (x)) with f a non-negative C ∞ function such that

1 1
f (x) > 0 if and only if x − (1, 0) < or x − (2, 0) < . Show that
4 ! 4
1 1
the set D = (x1 , 0) : |x1 − 1| < or |x1 − 2| < is a disconnected
4 4
weak control set of the system.
4. Show that the control systems in Examples 1.3.2 and 1.3.1 are minimal.
5. Give an example of minimal set for a control affine system on a non-

compact manifold.
6. Consider the control affine system ẋ = u (t) on Rn with control range

n
U = [1, 2] .
+ −
(a) Find the sets cl O>T (x) and cl O>T (x) for T ≥ 0.
(b) Check that no orbit is invariant.
(c) Show that there is no minimal set.

u (t) 1
7. Consider the control affine system ẋ = x on R2 with control
−1 0
range U = [−1, 0]. Show that the circles centered at the origin are the
weak control sets of the system. Describe the dynamic order among
them.

Sometimes, in the literature, a control system is called transitive if G (x) =
M for some (and then for every) x ∈ M . In this case, one says that the system
group G is transitive on M and M is called a homogeneous space of G. The
adjective “controllable” has been introduced to indicate the transitivity by
the system semigroup, which means a transitivity with respect to the positive
semi-orbits. A classification of matrix Lie group which are transitive in Rn \
{0} was determined by Boothby [17] and Boothby–Wilson [18]. Transitivity
of semigroups on homogeneous spaces was investigated by San Martin [89].
The minimal set is one of the most fruitful concepts in the theory of
topological dynamics. This concept was introduced in the classical theory of
dynamical systems by Birkhoff [15], who proved the well-known relationship
between compact minimal sets and recurrent points. The introduction of
minimal sets of topological transformation groups is due to Gottschalk [55].
Ellis and Nerurkar [48] extended the notion of a minimal set to the setting of
semigroup actions on topological spaces. The concept of a universal minimal
set was stated by Ellis [46, 47]. Souza–Tozatti [102] studied the relationship
between Lagrange stability and existence of minimal sets for semigroup actions
and control systems.
Chapter 4
Limit sets and

prolongations
This chapter is dedicated to the concepts of limit sets and prolongational

limit sets for control systems. They are fundamental elements in the study of
limit behavior problems, asymptotic transitivity, recursiveness, and
attraction. We explain how the limit sets of a control system relate to the
limit sets of its control flow. In Section 4.1 we define the concepts of posi-
tive and negative limit sets in the sense of Conley [36]. Our intention is to
present the main properties of these dynamical concepts. Section 4.2 contains
an explanation of prolongations and prolongational limit sets. We show how
the Conley definition of limit set relates to prolongational limit set. Section
4.3 deals with equistability for control systems. This is an aspect of stability
described by prolongations. The minimal equistable sets are the main objects
of investigation.
4.1 Limit sets

We start extending the Conley formulation of limit set for sets.
Definition 4.1.1 The positive limit set of a set X ⊂ M is defined by

+
ω + (X) = cl O>T (X)
T >0
81
82 Chapter 4 · Limit sets and prolongations
and the negative limit set of X by

−
ω − (X) = cl O>T (X) .
T >0
For a point x ∈ M , we define ω + (x) = ω + ({x}) and ω − (x) = ω − ({x}).

In this case, the Conley definition of limit set coincides with the usual one.
$ %
Remark 4.1.1 It can be easily seen that ω + (X) = n∈N cl O≥n +
(X) and
$ %
−
ω − (X) = n∈N cl O≥n (X) for every subset X ⊂ M .
In the following we list the main properties of limit sets.

Proposition 4.1.1 Let x ∈ M and X ⊂ M . Then:
1. ω + (X) and ω − (X) are closed subsets of M .
2. ω + (X) is positively invariant and ω − (X) is negatively invariant.
3. ω + (X) ⊂ X if X is closed and positively invariant and ω − (X) ⊂ X if
X is closed and negatively invariant.
+
4. ω + (X) = cl O>T (X) for all T > 0 if X ⊂ ω + (X) and ω − (X) =
−
cl O>T (X) for all T > 0 if X ⊂ ω − (X).
Proof. Item (1) follows by definition and item (2) follows according to
Propositions 2.1.1 and 2.2.1. If X is closed and positively invariant then
+
cl O>T (X) ⊂ X for all T > 0, hence ω + (X) ⊂ X; if X is closed and neg-
−
atively invariant then cl O>T (X) ⊂ X for all T > 0, hence ω − (X) ⊂ X.
This proves item (3). To prove item (4), note that the inclusion ω + (X) ⊂
+
cl O>T (X) is clear. On the other hand, if X ⊂ ω + (X), we have O>T
+
(X) ⊂
+
ω (X), by item (2), hence cl O>T (X) ⊂ ω (X) for all T > 0. Thus
+ +
+
ω + (X) = cl O>T (X) for all T > 0 if X ⊂ ω + (X). Analogously, we prove
−

that ω − (X) = cl O>T (X) for all T > 0 if X ⊂ ω − (X).
The following proposition characterizes the limits sets via limit of se-
quences.
Proposition 4.1.2 Let X ⊂ M . Then
!
+ y ∈ M : there are sequences (xn ) in X, (un ) in U ,
ω (X) = ,
and tn → +∞ such that ϕ (tn , xn , un ) → y
!
y ∈ M : there are sequences (xn ) in X, (un ) in U ,
ω − (X) = .
and tn → −∞ such that ϕ (tn , xn , un ) → y
Section 4.1 · Limit sets 83
Proof. Suppose that ϕ (tn , xn , un ) → y with (xn ) in X and tn → +∞.

For T, ε > 0, there is n0 such that tn > T and d (ϕ (tn , xn , un ) , y) < ε
+
whenever n > n0 . Hence y ∈ cl O>T (X) , for all T > 0, and therefore
y ∈ ω + (X). On the other hand, if z ∈ ω + (X) then B (z, 1/n) ∩ O>n +
(X) = ∅
for all n ∈ N. Hence there is tn > n, xn ∈ X, and un ∈ U such that
ϕ (tn , xn , un ) ∈ B (z, 1/n), that is, d (ϕ (tn , xn , un ) , z) < 1/n for all n ∈ N.
This means that ϕ (tn , xn , un ) → z with (xn ) in X and tn → +∞. This proves
the equality with respect to the positive limit set. The case of negative limit
set is proved similarly.
In certain senses, the limit set “closes” the semi-orbit, as in the following
Proposition 4.1.3 For any x ∈ M , one has

cl O+ (x) = O+ (x) ∪ ω + (x) and cl O− (x) = O− (x) ∪ ω − (x) .
Proof. We prove only the case cl (O+ (x)) = O+ (x) ∪ ω + (x). Indeed, it
can be easily seen that O+ (x)∪ω + (x) ⊂ cl (O+ (x)). Take any y ∈ cl (O+ (x))
and suppose that y ∈ / O+ (x). There are sequences (tn ) in R+ and (un ) in U
such that ϕ (tn , x, un ) → y. If (tn ) is bounded, we may assume that tn → t
with t ≥ 0. By compactness, we may assume that un → u in U . Hence
ϕ (tn , x, un ) → ϕ (t, x, u), and than y = ϕ (t, x, u) ∈ O+ (x), a contradiction.
This means that (tn ) is unbounded, and therefore y ∈ ω + (x).
The following example shows that limit sets need not be invariant.
Example 4.1.1 Consider the control affine system defined in Example 2.2.2.
As the origin 0 is a stationary point for both vector fields X1 and X2 , we have
ω + (0) = ω − (0) = {0}. For x = 0 we have
−
cl O>T (x) = {y ∈ M : y ≤ x} ,
+
cl O>T (x) = {y ∈ M : y ≥ x} ,
for all T > 0, hence
ω − (x) = {y ∈ M : y ≤ x} ,

+
ω (x) = {y ∈ M : y ≥ x} .
Note that the negative limit set ω − (x) is not positively invariant and the
positive limit set ω + (x) is not negatively invariant.
Example 4.1.2 Go back to Example 3.1.1. We have ω + (0) = ω − (0) = {0}

and ω + (π) = ω − (π) = {π}. For any x ∈ (0, π/2), we have ω + (x) = [π/2, π]
and ω − (x) = {0}. For x ∈ [π/2, π), we have ω + (x) = [π/2, π] and ω − (x) =
[0, π]. For any x ∈ (π, 3π/2), we have ω + (x) = [π, 3π/2] and ω − (x) is the
lower semicircle. For x ∈ [3π/2, 2π), we have ω + (x) = [π, 3π/2] and ω − (x) =
{0}. The intervals [π/2, π] and [π, 3π/2] are not negatively invariant.
There is a general situation where limit sets are invariant (see Problem
4.4(8)).
Note that the limit sets of a set X ⊂ M may be empty, even if X is
nonempty. This possibility does not occur in compact manifolds, as in the
following:
Proposition 4.1.4 Assume that the manifold M is compact. If X ⊂ M is a

nonempty subset then the limit sets of X are nonempty and compact.
Proof. We only prove the case of a positive limit set. Let x ∈ X, u ∈ U ,

and tn → +∞. By the compactness of M , a subsequence of (ϕ (tn , x, u))n∈N
converges to some y ∈ M . According to Proposition 4.1.2, y ∈ ω + (x) ⊂
ω + (X), and therefore ω + (X) is nonempty. Moreover, ω + (X) is compact
since it is closed.
In general, if cl (O+ (x)) or cl (O− (x)) is compact then respectively ω + (x)

or ω − (x) is nonempty and compact, since ω + (x) ⊂ cl (O+ (x)) and ω − (x) ⊂
cl (O− (x)). The converse holds, as follows:
Proposition 4.1.5 The positive limit set ω + (x) is nonempty and compact if
and only if cl (O+ (x)) is compact. The negative limit set ω − (x) is nonempty
and compact if and only if cl (O− (x)) is compact.
Proof. It remains to show that cl (O+ (x)) is compact whenever ω + (x)

is nonempty and compact, and cl (O− (x)) is compact whenever ω − (x) is
nonempty and compact. Suppose that ω + (x) is nonempty and compact.
Let K be a compact neighborhood of ω + (x) such that fr (K) ∩ ω + (x) = ∅.
There is some t > 0 with O≥t +
(x) ⊂ K. In fact, suppose by contradic-
tion that for every n ∈ N, O≥n +
(x) K. Since O≥n +
(x) is connected,
there are tn ≥ n and un ∈ U such that ϕ (tn , x, un ) ∈ fr (K). Since fr (K)
is compact, there is a subsequence (ϕ (tnk , x, unk )) that converges to some
point y ∈ fr (K). As tnk → +∞, we have y ∈ fr (K) ∩ ω + (x), which
Section 4.1 · Limit sets 85
$ %
is a contradiction. Take t > 0 with O≥t
+
(x) ⊂ K. Then cl O≥t
+
(x)
is compact. As O≤t+
(x) = ϕ ([0, t] × {x} × U ) is compact, it follows that
$ %
cl (O (x)) = O≤t (x) ∪ cl O≥t
+ + +
(x) is compact. The proof for cl (O− (x)) is
analogous.
As a consequence, compact limit sets are continuum.
Corollary 4.1.1 Every compact limit set is connected.
Proof. Suppose that ω + (x) is nonempty and compact. According to

+
Proposition 4.1.5, cl (O+ (x)) is compact, and then cl O>t (x) is compact
+
for all t > 0. According to Proposition 2.2.3, each cl O>t (x) is connected,
and hence this is a continuum set. It is well-known that every intersection of
continuum sets is a continuum set. Thus ω + (x) is a continuum set, that is, a
compact and connected set. The proof for ω − (x) is analogous.
In the case of compact state space, every limit set is a continuum set. This
result can be easily extended in the following way:
Proposition 4.1.6 If x lies in some compact positively invariant subset X ⊂

M then the positive limit set ω + (x) is a continuum. Analogously, if x lies in
some compact negatively invariant set X then the negative limit set ω − (x) is
a continuum.
The following shows how limit set map relates to transition map:
Proposition 4.1.7 For every set X ⊂ M , t > 0, and u ∈ U, the following

inclusions hold:

ω + (ϕut (X)) ⊂ ω + (X) ⊂ ω + ϕu−t (X) ,

ω − ϕu−t (X) ⊂ ω − (X) ⊂ ω − (ϕut (X)) .
Proof. We only prove the case of positive limit set. If x ∈ ω + (ϕut (X))
then there are sequences (xn ) in X, (un ) in U , and tn → +∞ such that
ϕ (tn , ϕut (xn ) , un ) → x. According to Proposition 1.3.2, there is vn ∈ U such
that ϕ (tn , ϕ (t, xn , u) , un ) = ϕ (tn + t, xn , vn ). As ϕ (tn + t, xn , vn ) → x and
tn + t → +∞, we have x ∈ ω + (X), and thus ω + (ϕut (X)) ⊂ ω + (X). Now,
if y ∈ ω + (X) then there are sequences (xn ) in X, (un ) in U , and tn →
+∞ such that ϕ (tn , xn , un ) → y. Since xn = ϕ (t, ϕ (−t, xn , u) , u · (−t)), we
can write ϕ (tn , xn , un ) = ϕ (tn , ϕ (t, ϕ (−t, xn , u) , u · (−t)) , un ). According

to Proposition 1.3.2 again, there is wn ∈ U such that
ϕ (tn , ϕ (t, ϕ (−t, xn , u) , u · (−t)) , un ) = ϕ (tn + t, ϕ (−t, xn , u) , wn ) .
It follows that ϕ (tn + t, ϕ (−t, xn , u) , wn ) → y with tn + t → +∞ and

ϕ (−t, xn , u) ∈ ϕu−t (X). Hence y ∈ ω + ϕu−t (X) , and therefore ω + (X) ⊂

ω + ϕu−t (X) .
See Problem 4.4(8) for a situation where the inclusions stated in Proposi-
tion 4.1.7 become equalities.
We now describes the limit sets of the control system by means of the limit
sets of the control flow via the first-projection π : M × U → M .

Proposition 4.1.8 For any set X ⊂ M , one has π ω + π −1 (X) = ω + (X)
− −1
and π ω π (X) = ω − (X).

Proof. If z ∈ π ω + π −1 (X) then there is some u ∈ U such that

(z, u) ∈ ω + π −1 (X) . Hence there are sequences tn → +∞ and ((xn , un )) in
π −1 (X) such that Φ (tn , xn , un ) → (z, u). This means that (xn ) is a sequence
in X such that ϕ (tn , xn , un ) → z, and therefore z ∈ ω + (X). This proves

the inclusion π ω + π −1 (X) ⊂ ω + (X). On the other hand, suppose that
y ∈ ω + (X). Then there are sequences tn → +∞, (xn ) in X, and (un ) in U
such that ϕ (tn , xn , un ) → y. By compactness, we may assume that un ·tn → u
in U . This implies that
Φ (tn , xn , un ) = (ϕ (tn , xn , un ) , un · tn ) → (y, u)

where ((xn , un )) is a sequence in π −1 (X). Hence (y, u) ∈ ω + π −1 (x) ,

and therefore y ∈ π ω + π −1 (X) . This finishes the proof for the equal-
+ −1
ity π ω π (X) = ω + (X). The equality π ω − π −1 (X) = ω − (X) is
similarly proved.
4.2 Prolongational limit sets

This section defines the concepts of prolongations and prolongational limit
sets for control systems. They are concerned with general types of limit
behavior problems and recursiveness.
Section 4.2 · Prolongational limit sets 87
Definition 4.2.1 Let x ∈ M and t ≥ 0. The first positive and the first
negative t-prolongations of x are respectively the sets
$ % $ %
−
D+ (x, t) = cl O≥t
+
(B(x, ε)) and D− (x, t) = cl O≥t (B(x, ε)) .
ε>0 ε>0
Note that

D+ (x, 0) = cl O+ (B(x, ε)) ,
ε>0

−
D (x, 0) = cl O− (B(x, ε)) .
ε>0
We denote simply D+ (x) = D(x, 0) and D− (x) = D− (x, 0).

The first positive t-prolongation of a set X ⊂ M is defined by the union
" + " −
D+ (X, t) = D (x, t) and D− (X, t) = D (x, t). In special, D+ (X) =
" + x∈X " − x∈X
D (x) and D− (X) = D (x).
x∈X x∈X
The following proposition characterizes prolongations by limit of sequences.
Proposition 4.2.1 Let x ∈ M and t ≥ 0. Then

⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in [t, +∞) , ⎬
D+ (x, t) = (un ) in U , and (xn ) in M such that xn → x ,
⎩ ⎭
and ϕ (tn , xn , un ) → y
⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in (−∞, t] , ⎬
D− (x, t) = (un ) in U , and (xn ) in M such that xn → x .
⎩ ⎭
and ϕ (tn , xn , un ) → y
Proof. If y ∈ D+ (x, t) then B (y, 1/n) ∩ O≥t +

(B(x, 1/n)) = ∅ for all
n ∈ N. Hence there are sequences (tn ) in [t, +∞), (un ) in U , and (xn ) in M
such that xn ∈ B (x, 1/n) and ϕ (tn , xn , un ) ∈ B (y, 1/n) ∩ O≥t
+
(B(x, 1/n)) for
every n ∈ N. This means that xn → x and ϕ (tn , xn , un ) → y. On the other
hand, suppose that there are sequences (tn ) in [t, +∞), (un ) in U , and (xn )
in M such that xn → x and ϕ (tn , xn , un ) → y. For any neighborhood V of
y and ε > 0, there is n such that xn ∈ B (x, ε) $and ϕ (tn , xn , u%n ) ∈ V , hence
ϕ (tn , xn , un ) ∈ V ∩ O≥t
+
(B(x, ε)). Thus y ∈ cl O≥t +
(B(x, ε)) for all ε > 0.
This finishes the proof for the first equality. The second equality is proved in
the same way by considering the interval (−∞, t] instead of [t, +∞).
Next we define prolongational limit sets.

Definition 4.2.2 The first positive prolongational limit set of x ∈ M

is defined as
$ %
J + (x) = D+ (x, t) = cl O≥t
+
(B(x, ε))
t>0 t>0 ε>0
and the first negative prolongational limit set of x as

$ %
−
J − (x) = D− (x, t) = cl O≥t (B(x, ε)) .
t>0 t>0 ε>0
"
For a given set X ⊂ M , we define J + (X) = J + (x) and J − (X) =
" x∈X
J − (x).
x∈X
As we did with the prolongations, we can characterize the prolongational
limit sets by limit of sequences, but now with asymptotic sequences. The
proof is just an adaptation from the proof for Proposition 4.2.1, and thus it
is omitted (see Problem 4.4(4)).
Proposition 4.2.2 For any x ∈ M , one has

⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in R+ , ⎬
+
J (x) = (u ) in U , and (xn ) in M such that xn → x, ,
⎩ n ⎭
tn → +∞, and ϕ (tn , xn , un ) → y
⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in R− , ⎬
−
J (x) = (u ) in U , and (xn ) in M such that xn → x, .
⎩ n ⎭
tn → −∞, and ϕ (tn , xn , un ) → y
Note that ω + (x) ⊂ J + (x) and ω − (x) ⊂ J − (x) for every x ∈ M . These
inclusions may be proper. For instance, consider the control system on the
plane as defined in Example 2.2.1. The origin 0 is a global critical point, hence
ω + (0) = ω − (0) = {0}. It is easily seen that ω + (0) = {0} J + (0) = R × {0}
and ω − (0) = {0} J − (0) = R2 . In Example 3.1.1, we have ω + (0) =
ω − (0) = {0} and ω + (π) = ω − (π) = {π}, however, J + (0) = S1 = J − (π).
In the following we list the basic properties of prolongations and prolon-
gational limit sets.
Proposition 4.2.3 For any x ∈ M and t ≥ 0
1. D+ (x, t) and J + (x) are closed and positively invariant;
2. D− (x, t) and J − (x) are closed and negatively invariant.

Proof. This is just an application of Proposition 2.1.1, 2.1.2, and 2.2.1.
There exists a general situation in which the prolongational limit sets are
invariant (see Problem 4.4(8)).
In a certain sense, the prolongational limit set covers the excess of the
prolongation over the semi-orbit, as in the following:
Proposition 4.2.4 For any x ∈ M , the following equalities hold

D+ (x) = O+ (x) ∪ J + (x) = cl O+ (x) ∪ J + (x) ,

D− (x) = O− (x) ∪ J − (x) = cl O− (x) ∪ J − (x) .
Proof. It is easily seen that cl (O+ (x)) ∪ J + (x) ⊂ D+ (x). On the other
hand, if y ∈ D+ (x), then there are sequences (tn ) in R+ , (un ) in U , and (xn )
in M such that xn → x and ϕ (tn , xn , un ) → y. Suppose that y ∈ / J + (x), that
is, for any sequences (sn ) in R , (vn ) in U , and (yn ) in M , with yn → x and
+
sn → +∞, the sequence (ϕ (sn , yn , vn )) does not converge to y. If (tn ) ⊂ [0, a]

for some a > 0, then there is a subsequence (tnk ) that converges to some t in
[0, a] and a subsequence (unk ) that converges to some u ∈ U . It follows that
ϕ (tnk , xnk , unk ) → ϕ (t, x, u). Hence, y = ϕ (t, x, u) ∈ O+ (x). In other case,
if (tn ) [0, a] for every a > 0, there is a subsequence tnk → +∞. It follows
that ϕ (tnk , xnk , unk ) → y, with tnk → +∞ and xnk → x, which contradicts
y ∈/ J + (x). Thus, (tn ) ⊂ [0, a] for some a > 0, which implies y ∈ O+ (x).
Therefore D+ (x) = O+ (x) ∪ J + (x). Consequently, since D+ (x) is closed,
we have D+ (x) = cl (O+ (x)) ∪ J + (x). Analogous proof works for D− (x).
There may be a set of states that does not have prolongation. The follow-
ing describes this situation:
Proposition 4.2.5 Let X ⊂ M be a nonempty set. Then D+ (X) = X

if and only if for every y ∈ / X and x ∈ X there is some ε > 0 such that
B (y, ε) ∩ O (B (x, ε)) = ∅. Analogously, D− (X) = X if and only if for every
+
y∈ / X and x ∈ X there is some ε > 0 such that B (y, ε) ∩ O− (B (x, ε)) = ∅.
Proof. Suppose that D+ (X) = X and take y ∈ / X and x ∈ X. Then

y ∈
/ D+ (x), and hence there is δ > 0 such that y ∈ / cl (O+ (B (x, δ))).
Thus M \ cl (O (B (x, δ))) is an open neighborhood of y that does not meet
+
O+ (B (x, δ)). Choose ε > 0 such that ε < δ and B (y, ε) ⊂ M \cl (O+ (B (x, δ))).
This means that B (y, ε) ∩ O+ (B (x, ε)) = ∅. As to the converse, suppose that
for every y ∈
/ X and x ∈ X there is ε > 0 such that B (y, ε)∩O+ (B (x, ε)) = ∅.
Since X ⊂ D+ (X) in general, we need to show the inclusion D+ (X) ⊂ X.
Indeed, take y ∈
/ X and x ∈ X. By this hypothesis, there is ε > 0 such that
B (y, ε)∩O+ (B (x, ε)) = ∅. Hence y ∈ / cl (O+ (B (x, ε))), and thus y ∈
/ D+ (x).
Since x ∈ X is arbitrary, it follows that y ∈
/ D (X). Therefore D (X) ⊂ X.
+ +
In the next section we explain that the property of a set not admitting
prolongation approaches an aspect of stability.
Note that all prolongations and prolongational limit sets define maps of
"
the form F : M → 2M , where we consider the image F (X) = F (x) for
x∈X
any subset X ⊂ M . However, observe that the limit sets of a set X are defined
in a distinct way. Problem 4.4(2) asks about this difference. Otherwise, we
have an elementary relation between limit set and prolongational limit set for
sets, as in the following.
Proposition 4.2.6 Let K ⊂ M be a compact set and V ⊂ M an open set.

Then
1. ω + (K) ⊂ J + (K) and ω − (K) ⊂ J − (K).
2. J + (V ) ⊂ ω + (V ) and J − (V ) ⊂ ω − (V ).
Proof. (1) Let x ∈ ω + (K) with K compact. There are then sequences
tn → +∞, (xn ) in K, and (un ) in U such that ϕ (tn , xn , un ) → x. By taking
a subsequence, we can assume that xn → y with y ∈ K. Hence, x ∈ J + (y),
and therefore ω + (K) ⊂ J + (K). The inclusion ω − (K) ⊂ J − (K) is similarly
proved.
(2) Let x ∈ J + (V ) with V open. Then x ∈ J + (z) for some z ∈ V . There
are then sequences tn → +∞, (un ) in U , and (xn ) in M such that xn → z
and ϕ (tn , xn , un ) → x. Since V is open, there is n0 such that xn ∈ V for all
n ≥ n0 . Hence x ∈ ω + (V ) and therefore J + (V ) ⊂ ω + (V ). Analogously we
prove the inclusion J − (V ) ⊂ ω − (V ).
Thus, for K compact and V open, we have the relation

# #
ω + (x) ⊂ ω + (K) ⊂ J + (K) , ω − (x) ⊂ ω − (K) ⊂ J − (K) ,
x∈K x∈K
# #
+
ω (x) ⊂ J (V ) ⊂ ω (V ) ,
+ +
ω − (x) ⊂ J − (V ) ⊂ ω − (V ) .
x∈V x∈V
The following property of prolongational limit sets need not be satisfied

by limit sets.
Proposition 4.2.7 For any x, y ∈ M , y ∈ J + (x) is and only if x ∈ J − (y).
Proof. Suppose that y ∈ J + (x) and take sequences tn → +∞, (un ) in

U , and (xn ) in M such that xn → x and ϕ (tn , xn , un ) → y. Then
ϕ (−tn , ϕ (tn , xn , un ) , un · tn ) = xn → x
with −tn → −∞ and ϕ (tn , xn , un ) → y, hence x ∈ J − (y). The converse is

similarly proved.
We now relate the prolongational limit set maps to transition maps.
Proposition 4.2.8 For any set X ⊂ M , t > 0, and u ∈ U , one has the
inclusions

J + (ϕut (X)) ⊂ J + (X) ⊂ J + ϕu−t (X) ,

J − ϕu−t (X) ⊂ J − (X) ⊂ J − (ϕut (X)) .
Proof. It is enough to show the inclusions above for a single set {x}. We
only prove the inclusions
J + (ϕ (t, x, u)) ⊂ J + (x) ⊂ J + (ϕ (−t, x, u))
for a given point x ∈ M . Indeed, if y ∈ J + (ϕ (t, x, u)) then there are se-
quences tn → +∞, (un ) in U , and (xn ) in M such that xn → ϕ (t, x, u) and
ϕ (tn , xn , un ) → y. This implies that ϕ (−t, xn , u · t) → x by continuity. We
can write
ϕ (tn , xn , un ) = ϕ (tn , ϕ (t, ϕ (−t, xn , u · t) , u) , un )

= ϕ (tn + t, ϕ (−t, xn , u · t) , vn )
for some vn ∈ U. We then have tn + t → +∞, ϕ (−t, xn , u · t) → x,

and ϕ (tn + t, ϕ (−t, xn , u · t) , vn ) → y. Hence y ∈ J + (x), and therefore
J + (ϕ (t, x, u)) ⊂ J + (x). Now take z ∈ J + (x). There are then sequences
tn → +∞, (un ) in U , and (xn ) in M such that xn → x and ϕ (tn , xn , un ) → z.
It follows that ϕ (−t, xn , u) → ϕ (−t, x, u), and we can write
ϕ (tn , xn , un ) = ϕ (tn , ϕ (t, ϕ (−t, xn , u) , u · (−t)) , un )

= ϕ (tn + t, ϕ (−t, xn , u) , vn )
which implies that ϕ (tn + t, ϕ (−t, xn , u) , vn ) → z. Hence we have z ∈

J + (ϕ (−t, x, u)), and therefore J + (x) ⊂ J + (ϕ (−t, x, u)).
See Problem 4.4(8) for a situation where the inclusions stated in Proposi-
tion 4.2.8 can be replaced by equalities.
Some properties of the prolongational limit set are transferred to the limit
set, as follows.
Proposition 4.2.9 The limit set ω + (x) is nonempty and compact whenever
the prolongational limit set J + (x) is nonempty and compact. Analogously,
ω − (x) is nonempty and compact whenever J − (x) is nonempty and compact.
Proof. Suppose that J + (x) is nonempty and compact, and ω + (x) =

∅. Since cl (O+ (x)) = O+ (x) ∪ ω + (x), it follows that O+ (x) is closed.
Moreover, O+ (x) is disjoint with J + (x). Indeed, if O+ (x) ∩ J + (x) = ∅
then there is s > 0 and u ∈ U such that ϕ (s, x, u) ∈ J + (x), and since
J + (x) is compact and positively invariant, we have ω + (ϕ (s, x, u)) = ∅.
According to Proposition 4.1.7, it follows that ω + (x) = ∅, which contra-
dicts the assumption ω + (x) = ∅. Thus O+ (x) is closed and disjoint with
J + (x). As J + (x) is compact, there is a compact neighborhood K of J + (x)
with fr (K) ∩ J + (x) = ∅, such that K ∩ O+ (x) = ∅. For y ∈ J + (x),
there are sequences tn → +∞, (un ) in U , and (xn ) in M such that xn →
x and ϕ (tn , xn , un ) → y. We may assume that ϕ (tn , xn , un ) ∈ K for
every n. The trajectory segments ϕ ([0, tn ] , xn , un ) then intersect fr (K),
hence we obtain a sequence (ϕ (τ n , xn , un )) ∈ fr (K) with τ n ∈ [0, tn ]. As
fr (K) is compact, there is a subsequence ϕ (τ nk , xnk , unk ) that converges to
some point z ∈ fr (K). If the sequence (τ nk ) is bounded, we may assume
that τ nk → t in R+ and unk → u in U , and therefore ϕ (τ nk , xnk , unk ) →
ϕ (t, x, u) = z. Since O+ (x) is closed, it follows that z ∈ O+ (x) ∩ fr (K)
which contradicts K ∩ O+ (x) = ∅. If τ nk → +∞, then z ∈ J + (x) ∩ fr (K)
which contradicts fr (K) ∩ J + (x) = ∅. Therefore, J + (x) being nonempty
and compact with ω + (x) = ∅ is impossible. The proof for ω − (x) and J − (x)
follows in a similar way.
Consequently, the compactness of the prolongational limit set implies the

compactness of the prolongation.
Corollary 4.2.1 The prolongational limit set J + (x) is nonempty and com-
pact if and only if the prolongation D+ (x) is compact. Analogously, J − (x)
is nonempty and compact if and only if D− (x) is compact.
Proof. If D+ (x) is compact then ω + (x) is nonempty, because x ∈ D+ (x)

and D+ (x) is a compact positively invariant set. Hence J + (x) is nonempty.
Since J + (x) ⊂ D+ (x), it follows that J + (x) is also compact. As to the con-
verse, suppose that J + (x) is nonempty and compact. According to Proposi-
tion 4.2.9, ω + (x) is nonempty and compact. Proposition 4.1.5 now says that
cl (O+ (x)) is compact. Finally, D+ (x) is compact according to Proposition
4.2.4.
We shall now discuss the connectedness of the prolongations and the pro-
longational limit sets. Compact prolongations and compact prolongational
limit sets are continuum sets.
Theorem 4.2.1 All the compact prolongations and compact prolongational

limit sets of the control system are connected.
Proof. Suppose by contradiction that D+ (x, t) is compact and discon-

nected for some t ∈ R and x ∈ M . There are then two compact nonempty
sets A and B such that A ∩ B = ∅ and A ∪ B = D+ (x, t). Let KA and KB be
compact neighborhoods of A and B, respectively, with KA ∩ KB = ∅. Since
O≥t+
(x) is connected and O≥t +
(x) ⊂ D+ (x, t), it follows that O≥t
+
(x) ⊂ A
or O≥t (x) ⊂ B. Let O≥t (x) ⊂ A. For y ∈ B, there are sequences (tn ) in
+ +
[t, +∞), (un ) in U , and (xn ) in M such that xn → x and ϕ (tn , xn , un ) → y.

For some n0 , we have xn ∈ KA and ϕ (tn , xn , un ) ∈ KB for all n ≥ n0 .
Then the trajectory segments ϕ ([0, tn ] , xn , un ) intersect fr (KA ) for every
n ≥ n0 . Hence, there is a sequence (τ n ), with τ n ∈ [0, tn ], such that
ϕ (τ n , xn , un ) ∈ fr (KA ). As fr (KA ) is compact, there is a subsequence
(ϕ (τ nk , xnk , unk )) that converges to some point z in fr (KA ). Then z ∈
D+ (x, t) but z ∈ / A ∪ B, which is a contradiction. Therefore, D+ (x, t)
is connected whenever it is compact. A similar proof works for the first
negative t-prolongations. Now suppose that J + (x) is nonempty and compact.
Then the prolongation D+ (x) is compact, and hence D+ (x, t) is compact for
all t > 0. By the first part of the theorem, D+ (x, t) connected, hence it is
a continuum. It follows that J + (x) is an intersection of the continuum, and
therefore it is a continuum. Similar proof works for J − (x).
The following result describes the prolongational limit sets of the control
system by means of the prolongational limit sets of the control flow.

Proposition 4.2.10 For any point x ∈ M , one has π J + π −1 (x) =

J + (x) and π J − π −1 (x) = J − (x).

Proof. If z ∈ π J + π −1 (x) then there are u, v ∈ U such that (z, u) ∈
J + ((x, v)). Hence there are sequences tn → +∞ and (xn , vn ) → (x, v) such
that Φ (tn , xn , vn ) → (z, u). This means that xn → x and ϕ (tn , xn , vn ) →

z, and therefore z ∈ J + (x). This proves the inclusion π J + π −1 (x) ⊂
J + (x). On the other hand, suppose that y ∈ J + (x). There are then sequences
tn → +∞, xn → x in X, and (un ) in U such that ϕ (tn , xn , un ) → y. By
compactness, we may assume that un → u and un · tn → v in U . This implies
that (xn , un ) → (x, u) and
Φ (tn , xn , un ) = (ϕ (tn , xn , un ) , un · tn ) → (y, v)

hence (y, v) ∈ J + ((x, u)).Therefore y ∈ π J + π −1 (x) . This finishes the
+ −1
proof for the equality π J π (x) = J + (x). The other equality can be
similarly proved.
Consequently, we have

π J + π −1 (X) = J + (X) ,

π J − π −1 (X) = J − (X) ,
for any subset X ⊂ M .
4.3 Minimal equistable sets

In this section we deal with the concept of equistability for control systems.
This is an aspect of stability directly linked to prolongations. In particular,
we study the minimal equistable sets.
Definition 4.3.1 A set E ⊂ M is said to be equistable if for each x ∈

/ E
there exists some ε > 0 such that x ∈ +
/ cl(O (B (E, ε))).
In the following we show the main properties of equistable sets.
Proposition 4.3.1 Let E ⊂ M be an equistable set. Then
1. E is positively invariant.
Section 4.3 · Minimal equistable sets 95
2. D+ (x) ⊂ E for every x ∈ E.
Proof. To show (1), let x ∈ E, t > 0, and u ∈ U . Then ϕ (t, x, u) ∈

O+ (B (E, ε)) for any ε > 0. Since E is equistable, it follows that ϕ (t, x, u) ∈
E, and therefore E is positively invariant. To show (2), suppose that there
is x ∈ E such that D+ (x) E and take y ∈ D+ (x) \ E. Then there
are sequences (tn ) in R+ , (un ) in U , and (xn ) in M such that xn → x and
ϕ (tn , xn , un ) → y. Moreover, there is an ε > 0 satisfying y ∈
/ cl(O+ (B (E, ε))).
We may consider xn ∈ B (x, ε) for all n, which implies that ϕ (tn , xn , un ) ∈
O+ (B (x, ε)) for all n. It follows that y ∈ cl (O+ (B (x, ε))) with x ∈ E, which
is a contradiction. Thus D+ (x) ⊂ E for every x ∈ E.
In the case of compact sets, the equistable sets are characterized as the
sets which do not admit prolongation, as in the following:
Theorem 4.3.1 A compact set K ⊂ M is equistable if and only if D+ (K) =

K.
Proof. By Proposition 4.2.5, it is enough to show that K is equistable

if and only if for every x ∈ / K and y ∈ K there is ε > 0 such that B (x, ε) ∩
O (B (y, ε)) = ∅. In fact, suppose that K is equistable. For x ∈
+
/ K there
is δ > 0 with x ∈ / cl (O+ (B (K, δ))). Then we can find δ > 0 such that

B x, δ ∩ O+ (B (K, δ)) = ∅. For ε = min δ, δ , it follows that B (x, ε) ∩
O+ (B (y, ε)) = ∅ for any y ∈ K. As to the converse, suppose that (2) holds
and take x ∈ / K. For each y ∈ K there is εy > 0 such that B (x, εy ) ∩
O+ (B (y, εy )) = ∅. Since K is compact we can find finitely many points
"k
y1 , ..., yk ∈ K such that K ⊂ i=1 B (yi , εyi ). Take ε > 0 such that B (K, ε) ⊂
"k k
i=1 B (yi , εyi ) and define U = i=1 B (x, εyi ). Then U is an open neigh-
borhood of x with U ∩ O+ (B (K, ε)) = ∅. Hence x ∈ / cl (O+ (B (K, ε))), and
therefore K is equistable.
In particular, a point x ∈ M being equistable means that D+ (x) = {x}.

As D+ (x) is positively invariant, it follows that x is a global critical point.
We now define the concept of a minimal equistable set.
Definition 4.3.2 A nonempty set E ⊂ M is said to be minimal equistable

if it is closed, equistable, and does not contain a nonempty proper subset
satisfying these properties.
Contrast the definition of minimal equistable set with the definition of min-
imal set (Definition 3.1.1). In the following, we present a sufficient condition
for an equistable set to be minimal equistable.
Proposition 4.3.2 Assume that E ⊂ M is an equistable set. If D+ (x) = E

for all x ∈ E then E is minimal equistable.
Proof. Suppose that E ⊂ E is equistable. According to Proposition

4.3.1, we have D+ (y) ⊂ E for every y ∈ E . Hence E = D+ (y) ⊂ E for any
y ∈ E , and therefore E = E. This proves that E is minimal equistable.
Evidently, the converse to Proposition 4.3.2 holds if the prolongations are

equistable sets themselves. This fact occurs under compactness and equicon-
tinuity. We say that the control system is equicontinuous if the family of all
transitions of the system is equicontinuous. Proposition 1.3.4 proves that the
family of transitions for bounded time is uniformly equicontinuous on any
compact subset of M . In general, however, the family of all transitions with
unbounded time need not be equicontinuous, unless they are functions with a
same Lipschitz constant. In order to show the converse for Proposition 4.3.2,
we need the following lemma:
Lemma 4.3.1 Assume that the control system is equicontinuous at y ∈ M .

If x ∈ D+ (y, t) and y ∈ D+ (z, s) then x ∈ D (z, t + s).
Proof. Since x ∈ D+ (y, t), there are sequences (tn ) in [t, +∞), (un ) in
U , and (yn ) in M such that yn → y and ϕ (tn , yn , un ) → x. As y ∈ D+ (z, s),
there are sequences (sn ) in [s, +∞), (vn ) in U , and (zn ) in M such that
zn → z and ϕ (sn , zn , v$n ) →
$ y. Let
%% ε, δ > 0. By the equicontinuity at y, there
is δ > 0 such that ϕuτ B y, δ ⊂ B (ϕuτ (y) , δ/4) for all u ∈ U and τ ∈ R.

We may consider that yn , ϕ (sn , zn , vn ) ∈ B y, δ , ϕ (tn , yn , un ) ∈ B (x, δ/2),
and zn ∈ B (z, ε). Hence
ϕ (tn , yn , un ) , ϕ (tn , ϕ (sn , zn , vn ) , un ) ∈ B (ϕ (tn , y, un ) , δ/4)
for all n, and therefore
d (ϕ (tn , ϕ (sn , zn , vn ) , un ) , x) ≤ d (ϕ (tn , ϕ (sn , zn , vn ) , un ) , ϕ (tn , yn , un ))

+d (ϕ (tn , yn , un ) , x)
δ δ
< + = δ.
2 2
Section 4.3 · Minimal equistable sets 97
As ϕ (tn , ϕ (sn , zn , vn ) , un ) = ϕ (tn + sn , zn , wn ) for some wn ∈ U, it follows

that ϕ (tn + sn , zn , wn ) ∈ B (x, δ) with tn +sn ∈ [t + s, +∞) and zn ∈ B (z, ε).
Hence, B (x, δ) ∩ O≥t+s +
B (z, ε) = ∅, and therefore x ∈ D+ (z, t + s).
Proposition 4.3.3 Assume that the control system is equicontinuous. If the

prolongational limit set J + (x) is a nonempty compact set then it is equistable.
In particular, if M is a compact manifold then J + (x) is a compact equistable
set for all x ∈ M .
Proof. Suppose by contradiction that J + (x) is not equistable. There is

then y ∈ M \ J + (x) such that y ∈ cl (O+ (B (J + (x) , ε))) for every ε > 0.
Hence for each n ∈ N there are tn ∈ R+ , un ∈ U , xn ∈ B (J + (x) , 1/n), and
zn ∈ J + (x) such that xn ∈ B (zn , 1/n) and ϕ (tn , xn , un ) ∈ B (y, 1/n). By the
compactness of J + (x), we may assume that zn → z with z ∈ J + (x). Since
d (xn , zn ) < 1/n, we have xn → z. If (tn ) is bounded, we may assume that
tn → t, with t ≥ 0, and un → u in U , hence ϕ (tn , xn , un ) → ϕ (t, z, u) = y.
Since z ∈ J + (x) and J + (x) is positively invariant, it follows that y ∈ J + (x),
which is a contradiction. Thus (tn ) is unbounded, and we may then as-
sume that tn → +∞. As ϕ (tn , xn , un ) → y, we have y ∈ J + (z). Hence
y ∈ D+ (z, t) and z ∈ D+ (x, t) for every t ≥ 0. According to Lemma 4.3.1, it
follows that y ∈ J + (x), which is a contradiction again.
Thus, under equicontinuity, a compact prolongation J + (x) is equistable.

Propositions 4.3.1 and 4.3.3 imply the following result.
Corollary 4.3.1 Assume that the control system is equicontinuous. Let K

be a compact equistable subset of M . Then J + (x) is a compact equistable set
for all x ∈ K.
The next result characterizes the compact minimal equistable sets by

means of prolongational limit sets.
Proposition 4.3.4 Assume that the control system is equicontinuous. A

compact set K ⊂ M is then minimal equistable if and only if J + (x) = K
for every x ∈ K.
Proof. If K is minimal equistable and x ∈ K, Proposition 4.3.3 assures

that J + (x) is a compact equistable subset of K. The minimality of K implies
that J + (x) = K. As to the converse, if J + (x) = K, for all x ∈ K, then K is

equistable according to Proposition 4.3.3. Now, let K ⊂ K be a nonempty,
closed, and equistable subset. For x ∈ K , we have J + (x) ⊂ K . It follows
that K = J + (x) ⊂ K , and therefore K = K .
Every equistable set is positively invariant but need not be a positive

minimal set. For instance, consider the dynamical system on the plane whose
phase portrait is as in Fig. 4.1. The unit circle S1 consists of an equilibrium
point p = (−1, 0) and a homoclinic trajectory Γ such that ω + (x) = ω − (x) =
{p} for all x ∈ Γ. Hence the unit circle is not a minimal set. However, every
trajectory in the unit disk has the same property as Γ, and every trajectory
outside the unit disk spiral to S1 as t → +∞. Hence we have J + (x) = S1
for all x ∈ S1 . According to Proposition 4.3.4, the unit circle is a minimal
equistable set.
Figure 4.1: A minimal equistable set that is not a positive minimal set.
We now prove that every compact equistable set contains an equistable

minimal set. We need the following lemma.
Lemma 4.3.2 Let (Ei )i∈I be a collection of equistable subsets of M . Then

i∈I Ei is equistable if it is nonempty.

Proof. Take z ∈ / i∈I Ei . Then there is j ∈ I such that z ∈ / Ej . By
the equistability of Ej there is ε > 0 such that z ∈
/ cl (O+ (B (Ej , ε))). Hence

z∈/ cl O+ B i∈I Ei , ε , and therefore i∈I Ei is equistable.
Proposition 4.3.5 Every compact equistable set in M contains a minimal

equistable set.
Proof. Let K ⊂ M be a compact equistable set and consider the collection
C = {C ⊂ M : C is a closed equistable subset of K}
ordered by reverse inclusion, that is, C2 C1 if and only if C2 ⊂ C1 . Note

that C = ∅ because K ∈ C. Take an arbitrary chain (Ci )i∈I in C and let

C = i∈I Ci . The compactness of K and Lemma 4.3.2 imply that C is a
nonempty upper bound of (Ci )i∈I contained in C. By the Zorn lemma we
obtain a maximal element of C, which is a minimal equistable subset of K.
4.4 Problems
1. Check that
$ %
ω + (X) = cl O≥n
+
(X) ,
n∈N
$ %
−
ω − (X) = cl O≥n (X) ,
n∈N
for every subset X ⊂ M .

" + " −
2. Prove that ω (x) ⊂ ω + (X) and ω (x) ⊂ ω − (X) for any
x∈X x∈X
subset X ⊂ M . When do the equalities hold?
3. Show that if cl (O+ (x)) is compact then lim d (ϕ (t, x, u) , ω + (x)) = 0

t→+∞
for any u ∈ U. Prove a similar result for ω − (x).
4. Check the equalities

⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in R+ , ⎬
J + (x) = (un ) in U , and (xn ) in M such that xn → x, ,
⎩ ⎭
tn → +∞, and ϕ (tn , xn , un ) → y
⎧ ⎫
⎨ y ∈ M : there are sequences (tn ) in R− , ⎬
J − (x) = (un ) in U , and (xn ) in M such that xn → x, .
⎩ ⎭
tn → −∞, and ϕ (tn , xn , un ) → y

5. Check the inclusions J − ϕu−t (X) ⊂ J − (X) ⊂ J − (ϕut (X)) for any set
X ⊂ M , t > 0, and u ∈ U.

6. Show that J + (x) = ε>0 ω + (B (x, ε)) and J − (x) = ε>0 ω − (B (x, ε))
for every x ∈ M .
7. Show that D+ (K, t) and D− (K, t) are closed sets, for any compact
K ⊂ M and t ≥ 0.
8. Recall that S≥T = {ϕut : t ≥ T, u ∈ U} are ideals of the system semi-

group S. A control system satisfies the translation hypothesis if for
any φ ∈ S and t > 0, there is s > 0 such that S≥s ⊂ φS≥t ∩ S≥t φ.
Assuming the translation hypothesis, show the following statements:
(a) All limit sets and prolongational limit sets are invariant.
(b) ω + (ϕut (X)) = ω + (X) and ω − (ϕut (X)) = ω − (X) for all X ⊂ M ,
t > 0, and u ∈ U.
(c) J + (ϕut (X)) = J + (X) and J − (ϕut (X)) = J − (X) for all X ⊂ M ,
t > 0, and u ∈ U.
9. Consider the control affine system ẋ = u (t) x on M = Rn with control

range U = [−2, −1] ⊂ R. Show that this control system satisfies the
translation hypothesis. (Hint: Describe the system semigroup as S =
{eun tn . . . eu1 t1 : uj ∈ [−2, −1] , tj ≥ 0, n ∈ N})
n
10. Consider the bilinear control system ẋ = i=1 ui (t) Xi (x) on Rd , where
U = {u ∈ Rn : a ≤ u ≤ b}, with a > 0, and X1 , ..., Xn ∈ Rd×d are
pairwise commutative matrices. Verify that this control system sat-
isfies the translation hypothesis. (Hint: Observe that exp (tXu ) =
exp (tu1 X1 + · · · + tun Xn ) for all u = (u1 , ..., un ) ∈ U and t ∈ R.
For any t, s > 0 and v = (v1 , ..., vn ) ∈ U, take T > 0 such that
b T − s > t. For a given exp $ (rXu ) ∈%S≥T , find w ∈ U such that
a

ru−sv
exp (rXu ) exp (−sXu ) = exp b Xw .
11. Show that the control affine system satisfies the translation hypothesis
if the system semigroup S coincides with the total subsemigroup G + of
G as defined in Problem 1.6(6). (In the general theory of semigroup
actions, the dynamical behavior of G + may be independently studied.)
12. Let M = Gl (n, R) be the linear Lie group and a the vector space of n×n
diagonal matrices of the form diag {λ, ..., λ}, λ ∈ R. Fix a nonzero n × n
matrix X and consider the control affine system on Gl (n, R) determined
by the family of vector fields F = {X + Y : Y ∈ a}. Prove that:
(a) the system semigroup is S = {exp t (X + Y ) : t ≥ 0, Y ∈ a};

(b) S is not a total subsemigroup in G (Hint: Check that exp (a)
S ∪ S −1 );
(c) the control system satisfies the translation hypothesis.

Many questions on dynamical systems and control systems depend only on
the action of a semigroup: the additive semigroup R+ (or Z+ ) for dynami-
cal systems and the system semigroup S for control affine systems. So these
questions can be abstracted to arbitrary semigroup actions and solved in a
more general setting. In such a generalization, the direction for limit behavior
is established by a filter basis on the subsets of the semigroup. This method-
ology was used to study limit sets and prolongational limit sets for semigroup
actions respectively by Braga Barros–Souza [22] and Souza–Tozatti [102].
The term ‘first’ prolongation in Definition 4.2.1 comes from the theory of
dynamical systems, where notions of higher prolongations are defined. The
first positive prolongation is an extension of the positive semi-trajectory, the
second and higher prolongations are extensions of the first prolongation (see
e.g. Bhatia [14]). Higher prolongations for control systems are in progress.
The first positive prolongation is used to characterize the stability of a com-
pact set (Theorem 4.3.1) and the first positive prolongational limit set is
applied to characterize dispersive control systems ([101]).
The concept of equistability is an aspect of Lyapunov stability extended
from dynamical systems to the general set-up of semigroup actions by Braga
Barros–Rocha–Souza [26]. In general, every closed uniformly stable set is
equistable. For compact sets in control systems, all Lyapunov notions of
stability, uniform stability, orbital stability, and equistability have the same
meaning.
Most results in this chapter were first published in [101, 102, 103].
+
In hyperspace, it is possible to show that the sequences cl O>n (x) and
+

cl(O>n (B (x, 1/n))) converge respectively to ω + (x) and J + (x)
whenever the manifold M is complete and the control system is limit com-
+
pact. In general, the set-valued function x ∈ M
→ cl O>t (x) is lower
semicontinuous and x ∈ M
→ J + (x) is upper semicontinuous ([1]).
Chapter 5
Asymptotic transitivity
An asymptotic transitive set is a region of the state space where each state
x is a limit point of any other state y. This means that points in asymp-
totic transitive sets are recurrent. Control sets are special cases of asymptotic
transitive sets. This chapter deals with the general notion of asymptotic tran-
sitivity. In Section 5.1 we study the transitivity relation among the periodic
points of a control system. The set of periodic points is partitioned in equiv-
alence classes where two periodic points are equivalent if they are in the same
periodic trajectory. These equivalence classes are called holding sets and con-
stitute special cases of asymptotic transitive sets. The minimal sets play an
important role in the studies of asymptotic transitivity and controllability. In
Section 5.2 we define the general notion of asymptotic transitivity by means
of Poincaré recurrence. A state point is positively recurrent if it lies in its
positive limit set. Two positively recurrent points are equivalent if each one is
a limit point of the other. We show that this is an equivalence relation whose
equivalence classes are maximal regions of asymptotic transitivity. Section 5.3
contains a brief discussion on ergodic theory. We reproduce the Poincaré re-
currence theorem that μ-almost every point is positively recurrent, with μ an
invariant probability measure. Section 5.4 is dedicated to the classical notion
of a control set. A control set is an asymptotic transitive set that contains
entire positive semi-trajectories through each one of its points. We introduce
the concept of accessibility and show how this concept influences the control-
lability in view of asymptotic transitivity. In Section 5.5 we specialize the
study on invariant control sets by indicating their relationship with positive
minimal sets. The controllability of the linear control system takes place sep-
103
104 Chapter 5 · Asymptotic transitivity
arately in Section 5.6. We use the Kalman condition to describe the effective
control sets for linear systems. In the last section we introduce another re-
cursive concept called a nonwandering point. By definition, a nonwandering
point lies in its prolongational limit set; recurrent points are then nonwander-
ing. More generally, we can show that every point closed to a recurrent point
is nonwandering. We also define a notion of controllability by prolongation,
whose relative prolongational control set means a region where each state lies
in the prolongational limit set of any other state.
5.1 Holding sets

This section sets out to study transitivity relations among the periodic
points of the control system. Throughout the book, P ⊂ M denotes the set
of all periodic points.
Definition 5.1.1 We say that two periodic points x, y ∈ P are equivalent if

x ∈ O>0
+
(y) and y ∈ O>0
+
(x).
This relation among the periodic points is called the transitivity re-
lation. It can be easily seen that the transitivity relation is reflexive and
symmetric. Furthermore, if x, y, z ∈ P with x equivalent to y and y equiv-
alent to z then x ∈ O>0+
(y), y ∈ O>0 +
(x), y ∈ O>0+
(z), and z ∈ O>0+
(y).
According to Proposition 2.2.5, x ∈ O>0 (z) and z ∈ O>0 (x), and hence x
+ +
is equivalent to z. Thus the transitivity relation is an equivalence relation in

the set P.
The following theorem characterizes the equivalence classes of the transi-
tivity relation.
Theorem 5.1.1 Let H ⊂ M be a nonempty set. The following sentences are

equivalent:
1. H is an equivalence class of the transitivity relation.
2. H ⊂ O>0 +
(x) for every x ∈ H and H is maximal satisfying this
property.
−
3. H = O>0
+
(x) ∩ O>0 (x) for every x ∈ H.
−
4. H = O>T
+
(x) ∩ O>T (x) for all x ∈ H and T ≥ 0.
Section 5.1 · Holding sets 105
Proof. (1) ⇒ (3) Take x, y ∈ H. Then x ∈ O>0 +

(y) and y ∈ O>0 +
(x),
− −
which means y ∈ O>0 (x) ∩ O>0 (x). Hence H ⊂ O>0 (x) ∩ O>0 (x) for every
+ +
−
x ∈ H. On the other hand, if z ∈ O>0 +
(x) ∩ O>0 (x) then z ∈ O>0+
(x) and
x ∈ O>0 (z). It follows that z ∈ O>0 (x) ⊂ O>0 (z), hence z is periodic and
+ + +
−
equivalent to x. Thus O>0 +
(x) ∩ O>0 (x) ⊂ H.
(3) ⇒ (2) Clearly, H ⊂ O>0 (x) for every x ∈ H. Let H ⊂ M such that
+
H ⊂ H and H ⊂ O>0 +
(x ) for all x ∈ H . For any x ∈ H and x ∈ H , we
−
have x ∈ O>0 (x) and x ∈ O>0
+ +
(x ), hence x ∈ O>0
+
(x) ∩ O>0 (x) = H, and

therefore H ⊂ H, proving that H is maximal.
(2) ⇒ (1) For any x, y ∈ H, we have x ∈ O>0 +
(y) and y ∈ O>0+
(x), hence x
and y are equivalent periodic points. Thus H is contained in some equivalence
class E of the transitivity relation in P. For any y ∈ E, we have E ⊂ O>0 +
(y).
By the maximality of H, it follows that H = E.
(3) ⇔ (4) It follows according to Theorem 2.4.1.
An equivalence class of the transitivity relation in P is usually called a

holding set, as named by E. Roxin [85]. Motivated by the properties stated
in Theorem 5.1.1, any holding set is also called a maximal transitive set.
Item (4) of Theorem 5.1.1 expresses the asymptotic property of the holding
sets.
−
For each x ∈ M , we define Hx = O>0 +
(x)∩O>0 (x). According to Theorem
2.4.1, we have the following alternative characterization of a periodic point.
Corollary 5.1.1 A point x ∈ M is periodic if and only if Hx = ∅.
Thus the non-periodic points of the control system correspond to the points
x ∈ M with Hx = ∅. This extends the transitivity relation to the whole state
space M .
Sometimes the holding set Hx reduces to the single set {x}, as in the
following example.
Example 5.1.1 Consider the control affine system ẋ = u (t) x on Rn with
t
control range U = [1, 2]. The solutions are of the form ϕ (t, x, u) = e 0 u(s)ds x.
In particular, ϕ (t, 0, u) = 0 for all t ∈ R and u ∈ U. Hence the origin 0 is a
global critical point, that is, O (0) = {0}. This means that H0 = {0}. Now,
take x = 0 and suppose that ϕ (t, x, u) = x for some t > 0 and u ∈ U . Then
t
t
t
t
e 0 u(s)ds x = x, which implies 0 u (s) ds = 0. As 0 u (s) ds ≥ 0 1ds = t, it
follows that t ≤ 0, a contradiction. Thus x is not a periodic point.
In other situations, the holding set Hx is a neighborhood of x. In this

case, Hx is an open set. The following characterizes an open holding set.
Proposition 5.1.1 Let x ∈ M be a periodic point. The following statements
are equivalent:
1. Hx is an open set.
2. int (Hx ) = ∅.
3. x ∈ int (Hx ).
Proof. The implications (1) ⇒ (3) and (3) ⇒ (2) are obvious. We then
+ −
show (2) ⇒ (1). Indeed, if y ∈ int (Hx ) then y ∈ int O>0 (y) ∩ int O>0 (y) .
−
According to Corollary 2.2.1, both sets O>0+
(y) and O>0 (y) are open. Hence
−
Hx = Hy = O>0+
(y) ∩ O>0 (y) is open.
Example 5.1.2 Consider the control affine system ẋ = x + u (t) on Rn with

control range U = B [0, 1], the closed unit ball centered at the origin 0 ∈

t
Rn . The solutions are of the form ϕ (t, x, u) = et x + et 0 e−s u (s) ds. Take
x, y ∈ Rn with x < c < 1. We claim that there is a constant function
u (t) = z and a time T > 0 such that ϕ (T, x, u) = y. It is enough to analyze
y − et x
the equation et x + (et − 1) z = y for t > 0. This means that z = t .
e −1
1 et
As t → 0 and t → 1 as t → +∞, we can find a T > 0 such that
e −1 e −1
y e x
T
y eT x
< 1−c and < c. Hence z ≤ + < 1. By taking
eT − 1 eT − 1 eT − 1 eT − 1
y−e xT
the constant function u (t) = T ∈ B [0, 1], we have ϕ (T, x, u) = eT x +
e −1
T y − eT x
e −1 +
= y, as claimed. Since y is arbitrary, we have O>0 (x) = Rn
eT − 1
for every x ∈ B (0, 1). Therefore the origin 0 is periodic and B (0, 1) ⊂ H0 .
According to Proposition 5.1.1, H0 is an open set. We show that H0 = B (0, 1).
Indeed, assume that x ≥ 1 and define ϕ (t, x, u) = y with t > 0. This means

t
that x = e−t y − 0 e−s u (s) ds, and then

−t t −s

1 ≤ x ≤ e y + e u (s) ds
0
t
≤ e−t y + e−s u (s) ds
0
≤ e−t y − e−t + 1.
Section 5.1 · Holding sets 107
It follows that y ≥ 1. Hence O>0

+
(x) ∩ B (0, 1) = ∅ whenever x ≥ 1. This
−
proves that O>0 (0) = B (0, 1) and therefore H0 = B (0, 1).
Example 5.1.3 Consider the control system on M = R2 as defined in Ex-

ample 2.4.3. The points (0, 0) and (1, 0) are global stationary points. Hence
H(0,0) = {(0, 0)} and H(1,0) = {(1, 0)}. The points of R2 \ T are in the same
holding set H(−1,0) . According to Proposition 5.1.1, H(−1,0) is open. The
reader is persuaded to discuss what happens inside T .
We shall now discuss the connectedness of the holding sets. The following
result states the non-regressing property of trajectories starting at holding
set.
Proposition 5.1.2 Let x ∈ P and assume that there are T > 0 and u ∈ U
such that ϕ (T, x, u) ∈ Hx . Then ϕ (t, x, u) ∈ Hx for all t ∈ [0, T ].
Proof. Denote z = ϕ (T, x, u). For a given t ∈ (0, T ), denote y =

ϕ (t, x, u). Then y ∈ O>0
+
(x). We also have
z = ϕ(T − t + t, x, u)
= ϕ (T − t, ϕ(t, x, u), u · t))
= ϕ (T − t, y, u · t)
where T −t > 0. Hence z ∈ O>0 +

(y). On the other hand, we have x ∈ O>0 +
(z),
since x and z are equivalent. It follows that y ∈ O>0 (x) and x ∈ O>0 (y),
+ +
that is, y is equivalent to x. Thus ϕ (t, x, u) = y ∈ Hx for arbitrary t ∈ (0, T ).

Thus a positive semi-trajectory that starts at a holding set, and goes out-
of it, can not return to it. This enables us to show that holding sets are
connected sets, as follows.
Proposition 5.1.3 Every holding set is pathwise connected.
Proof. Let x ∈ P and y, z ∈ Hx . Then there are T > 0 and u ∈ U

such that ϕ (T, y, u) = z. According to Proposition 5.1.2, it follows that
ϕ (t, y, u) ∈ Hx for all t ∈ [0, T ]. Hence Hx is pathwise connected.
There exists a dynamic order among the holding sets which describes the
behavior of the trajectories outside them. In order to define this order, we
introduce the domain of attraction of a holding set. Note that the positive
semi-orbit O>0
+
(x) may intersect a holding set H even if x ∈
/ H (see the prob-
lems). Such a point x is said to be attracted to H. The following definition
establishes this notion of attraction.
Definition 5.1.2 Let H be a holding set. The domain of attraction of H

is defined by

A (H) = x ∈ M : O>0 +
(x) ∩ H = ∅ .
By means of the notion of attraction, we define the following relation

among the holding sets: for two holding sets H1 and H2
H1 H2 if and only if H1 ∩ A (H2 ) = ∅.
Note that H ⊂ A (H) since H ⊂ O>0 +

(x) for all x ∈ H. Actually, any
point in A (H) has this property, as in the following:
Proposition 5.1.4 Let H ⊂ M be a holding set. Then

A (H) = x ∈ M : H ⊂ O>0
+
(x)

= x ∈ M : H ⊂ O>T
+
(x) for all T > 0 .
Proof. The inclusions

x ∈ M : H ⊂ O>T
+
(x) for all T > 0 ⊂ x ∈ M : H ⊂ O>0
+
(x) ⊂ A (H)
are clear. On the other hand, let x ∈ A (H). Then there is y ∈ O>0
+
(x) ∩ H.
For any z ∈ H and T > 0, we have z ∈ O>T (y) according to Theorem
+
4.3.1. Hence z ∈ O>T +

(x), according to Proposition 2.2.5, and therefore
H ⊂ O>T (x) for all T > 0.
+

Thus, for two holding sets H1 and H2 , we have H1 H2 if and only if

there is x ∈ H1 such that H2 ⊂ O>0
+
(x), which means that H2 ⊂ O>T
+
(x) for
all x ∈ H1 and T > 0.
Proposition 5.1.5 The relation “” among the holding sets is a preorder.
+
Proof. Let H ⊂ M be a holding set. Then H ⊂ cl O>0 (x) for all x ∈ H,
hence H H. Now, suppose that H1 , H2 , H3 are holding sets with H1 H2
and H2 H3 . We show that H1 H3 . Indeed, we have H2 ⊂ O>0 +
(x), for
all x ∈ H1 , and H3 ⊂ O>0 (y) for all y ∈ H2 . According to Proposition 2.2.5,
+
it follows that H3 ⊂ O>0

+
(x), for all x ∈ H1 . Hence H1 H3 .
Section 5.2 · Asymptotic transitive sets 109
5.2 Asymptotic transitive sets

This section generalizes the asymptotic transitivity of holding sets by means
of Poincaré recurrence. We study the asymptotic transitive sets, which are sets
of equivalent recurrent points. We show the relation between periodic point
and recurrent point, and then connect the holding sets to the asymptotic
transitive sets.
Definition 5.2.1 A point x ∈ M is said to be positively recurrent or neg-
atively recurrent if respectively x ∈ ω + (x) or x ∈ ω − (x); x is called re-
current if it is positively and negatively recurrent.
−
If x ∈ M is a periodic point then x ∈ O>T +
(x) ∩ O>T (x), for all T ≥ 0,
according to Proposition 5.1.1. This implies that x ∈ ω + (x)∩ω − (x), and thus
x is recurrent. In order to study other relations between recurrent points and
periodic points, we characterize the recurrent points by means of semi-orbits.
Theorem 5.2.1 Let x ∈ M . The following statements are equivalent:
1. x is positively recurrent.
2. cl (O+ (x)) = ω + (x).
+
3. cl (O+ (x)) = cl O>T (x) for all T ≥ 0.
4. O+ (x) ⊂ ω + (x).
5. For every > 0, there exist t ≥ 1 and u ∈ U such that ϕ (t, x, u) ∈
B (x, ).
Proof. The equivalence (1) ⇔ (4) is clear since ω + (x) is positively in-
variant. The equivalence (2) ⇔ (4) follows by the equality cl (O+ (x)) =
O+ (x) ∪ ω + (x) given in Proposition 4.1.3. The equivalence (2) ⇔ (3) and
the implication (4) ⇒ (5) are trivial. We then show (5) ⇒ (1). Sup-
pose that (5) holds. For each n ∈ N, take tn ≥ 1 and un ∈ U such that
ϕ (tn , x, un ) ∈ B (x, 1/n). Then ϕ (tn , x, un ) → x. If there is a subsequence
tnk → +∞, we have x ∈ ω + (x). If (tn ) is bounded, we may assume that
tn → t, and then t ≥ 1. We may also assume that un → u in U . It follows
that ϕ (tn , x, un ) → ϕ (t, x, u), and therefore ϕ (t, x, u) = x. As t ≥ 1, this
means that x is periodic, and therefore x ∈ ω + (x).
Analogously, we have the corresponding result for a negatively recurrent

point, as follows:
Theorem 5.2.2 Let x ∈ M . Then the following statements are equivalent:
1. x is negatively recurrent.
2. cl (O− (x)) = ω − (x).

−
3. cl (O− (x)) = cl O>T (x) for all T ≥ 0.
4. O− (x) ⊂ ω − (x).
5. For every > 0, there exist t ≥ 1 and u ∈ U such that ϕ (−t, x, u) ∈

B (x, ).
Note that x ∈ ω + (x) does not mean x ∈ ω − (x). In fact, a positively

recurrent point need not be negatively recurrent (see Problem 5.8(18)).
We now show a precise relationship between Poincaré recurrence and pe-
riodicity.
1. O+ (x) = ω + (x).
2. x is periodic and the positive semi-orbit O+ (x) is closed.
3. x is positively recurrent and the positive semi-orbit O+ (x) is closed.
Proof. (1) ⇔ (2) Suppose that O+ (x) = ω + (x). Then O+ (x) is closed,
since ω + (x) is closed. As x ∈ O+ (x) = ω + (x), there are sequences tn → +∞
and (un ) in U such that ϕ (tn , x, un ) → x. We may assume that un · tn → u
in U . For a given t > 0, it follows that
ϕ (−t, ϕ (tn , x, un ) , un · tn ) → ϕ (−t, x, u) .
As ϕ (−t, ϕ (tn , x, un ) , un · tn ) = ϕ (−t + tn , x, un ) and −t + tn → +∞, we

have ϕ (−t, x, u) ∈ ω + (x). Hence there is some T ≥ 0 and v ∈ U such that
ϕ (−t, x, u) = ϕ (T, x, v), and therefore x = ϕ (t + T, x, w) for some w ∈ U .
Since t + T > 0, it follows that x is periodic. As to the converse, suppose
that x is periodic and O+ (x) is closed. There are τ > 0 and u ∈ U such that
ϕ (kτ , x, u) = x for all integer k. For any t ≥ 0, v ∈ U , and n ∈ N, we can
find vn ∈ U such that
ϕ (t, x, v) = ϕ (t, ϕ (nτ , x, u) , v) = ϕ (t + nτ , x, vn ) .

As t + nτ → +∞, we have ϕ (t, x, v) ∈ ω + (x), and therefore O+ (x) ⊂ ω + (x).

Since O+ (x) is closed, it follows that O+ (x) = cl (O+ (x)) = ω + (x).
(1) ⇔ (3) If O+ (x) = ω + (x) then it follows immediately that x is posi-
tively recurrent and O+ (x) is closed. Conversely, if x is positively recurrent
and O+ (x) is closed then O+ (x) = ω + (x), according to Theorem 5.2.1.
We have an analogous result for negative limit sets and negatively recur-
rent points, as shown in the following:
1. O− (x) = ω − (x).
2. x is periodic and the negative semi-orbit O− (x) is closed.
3. x is negatively recurrent and the negative semi-orbit O− (x) is closed.
The following example shows recurrent points which are not periodic.
Example 5.2.1 Consider the control affine system on the plane R2 given by
the equations ẋ (t) = u (t) x0 , with control range U = [1, 2], where x0 = (a, b)
is such that b/a is an irrational number.$ The solution % of this system with

t
respect to the control u is ϕ (t, x, u) = x + 0 u (s) ds x0 . Take the quotient
&
map π : R2 → T2 of R2 onto the torus T2 = R2 Z2 and induce the solutions
of the system on T2 by putting ϕ (t, π (x) , u) = π (ϕ (t, x, u)). By taking the
constant function u = 1, define the vector field X on T2 by
d
X (π (x)) = ϕ (t, π (x) , 1)|t=0 = dπ x (x0 ) .
dt
Then the induced functions ϕ (t, π (x) , u) are solutions for the control affine
system ẏ (t) = u (t) X (y (t)) on T2 . Indeed we have
d
ϕ (t, π (x) , u) = dπ ϕ(t,x,u) (u (t) x0 ) = u (t) X (ϕ (t, π (x) , u)) .
dt
Note that the trajectory ϕuπ(x) (R) does not depend on the control function u.
The controls only influence the velocities of the motions. We claim that no
point of T2 is periodic. In fact, if ϕ (t, π (x) , v) = π (x)
$
for some% t > 0 and
t
v ∈ U then π (ϕ (t, x, v)) = π (x), which means that x+ 0 v (s) ds x0 = x+z

t b d
for some z = (c, d) ∈ Z2 . Since 0 v (s) ds ≥ t > 0, it follows that = is
a c
rational, which is a contradiction. We now claim that every point of the torus
is recurrent. Indeed, since b/a is irrational, the set A = {pb/a + q : p, q ∈ Z}
is dense in the real line R. For given (x1 , x2 ) , (y1 , y2 ) ∈ R2 and > 0,
b b b
there are p, q ∈ Z such that y2 − x2 + x1 − y1 − p − q < . By taking
a a a
y1 − x 1 + p
t= and the constant function u (t) = 1, we have
a
(y1 , y2 ) − ϕ (t, x, u) + (p, −q)
= (y1 , y2 ) − (x1 , x2 ) − t (a, b) + (p, q)

= (y1 − x1 − y1 + x1 − p + p, y2 − x2 − b y1 + b x1 − p b − q)
a a a

= (0, y2 − x2 − b y1 + b x1 − p b − q)
a a a

= y2 − x2 − b y1 + b x1 − p b − q) < .
a a a
This means that the set ϕux (R) + Z2 is dense in R2 , hence ϕuπ(x) (R) is dense
in T2 for every π (x) ∈ T2 . Thus the flow ϕu is minimal on T2 , which means
that ω + (π (x)) = ω − (π (x)) = T2 for every π (x) ∈ T2 . So any point of T2 is
recurrent but no one is periodic. Note that any positive semi-orbit O+ (π (x))
is not closed (compare with Theorem 5.2.3).
The flow point of view brings up an interesting problem on Poincaré re-
currence. In fact, the main characteristic of a control system is the possibility
of choosing an appropriate control that results a solution with the desired
qualitative properties. In the special case of recurrence, the intention is to
find a control function u ∈ U such that the point x is recurrent with respect
to the solution determined by u. In order to formulate this idea, we study the
limit behavior of motions by means of relative limit sets.
Definition 5.2.2 Let x ∈ M and u ∈ U. The positive limit set and the
negative limit set of x ∈ M relative to u are respectively defined by
!
y ∈ M : there is a sequence tn → +∞ such that
ω+ (x) = ,
u
ϕ (tn , x, u) → y
!
y ∈ M : there is a sequence tn → −∞ such that
ω− (x) = .
u
ϕ (tn , x, u) → y
The point x is said to be positively recurrent, negatively recurrent, or recurrent
− −
relative to u if x ∈ ω +
u (x), x ∈ ω u (x), or x ∈ ω u (x) ∩ ω u (x).
+
By taking the first-projection π : M × U → M , a simple computation

shows that

π ω + (x, u) = ω +
u (x) and π ω − (x, u) = ω −
u (x)
where ω + (x, u) and ω − (x, u) are respectively the positive limit set and neg-
ative limit set of (x, u) for the control flow Φ on M × U . Thus we have the
following result of recurrence in view of the control flow:
Proposition 5.2.1 Let (x, u) ∈ M × U . Then
1. (x, u) is a positively recurrent point of the control flow Φ if and only if

x is a positively recurrent point of the control system relative to u.
2. (x, u) is a negatively recurrent point of the control flow Φ if and only if

x is a negatively recurrent point of the control system relative to u.
3. (x, u) is a recurrent point of the control flow Φ if and only if x is a

recurrent point of the control system relative to u.
From now on, we concentrate the studies on the positive recurrence. Sim-
ilar results can be reproduced for the negative recurrence.
Let R+ denote the set of all positively recurrent points of the control
system. The closure of R+ in M is called the Birkhoff center of the control
system.
Definition 5.2.3 We say that two positively recurrent points x, y ∈ R+ are

equivalent if x ∈ ω + (y) and y ∈ ω + (x).
This relation among the positively recurrent points is called the asymp-
totic transitivity relation.
Proposition 5.2.2 The asymptotic transitivity relation is an equivalence re-

lation in R+ .
Proof. The reflexive and symmetric properties are clear. It remains for
us to prove the transitive property. Take x, y, z ∈ R+ such that x is equiv-
alent to y and y is equivalent to z. Then x ∈ ω + (y) and y ∈ ω + (z). Let
V be any open neighborhood of x and T > 0. Then V ∩ O>0 +
(y) = ∅, and
u −1 −1
hence there is t > 0 and u ∈ U such that y ∈ (ϕt ) (V ). Since (ϕut ) (V )
−1
is an open neighborhood of y, we have (ϕut ) (V ) ∩ O>T +
(z) = ∅. Hence
+
∅ = V ∩ ϕut O>T (z) ⊂ V ∩ O>T +
(z). Thus x ∈ ω + (z). Similarly, we show
that x ∈ ω + (z), and therefore x and z are equivalent.
We shall now describe the equivalence classes of the asymptotic transitivity

relation.
Proposition 5.2.3 A set R ⊂ R+ is an equivalence class of the asymptotic
transitivity relation if and only if R satisfies the following conditions:
1. R ⊂ ω + (x) for every x ∈ R;
2. R is maximal satisfying property (1).
Proof. Suppose that R is an equivalence class of the asymptotic transi-
tivity relation. For any x, y ∈ R, we have y ∈ ω + (x). Hence R ⊂ ω + (x) for
every x ∈ R. Let R be a set containing R and satisfying condition (1). For
x ∈ R and y ∈ R , we have y ∈ ω + (x) and x ∈ ω + (y), hence y and x are
equivalent. This means that y ∈ R, and therefore R is maximal satisfying
condition (1). As to the converse, if R satisfies condition (1) and (2) then x
and y are equivalent for all x, y ∈ R. Hence R is contained in some equiva-
lence class R of the asymptotic transitivity relation. As R satisfies condition
(1), the maximality of R implies that R = R.
A set R ⊂ M that satisfies the property (1) of Proposition 5.2.3 is called

an asymptotic transitive set. Each equivalence class of the asymptotic
transitivity relation is called a maximal asymptotic transitive set. For
each x ∈ R+ , we denote by Rx the maximal asymptotic transitive set of x.
We may extend the asymptotic transitivity relation to the whole state space
M by including the class Rx = M \ R+ for every x ∈ M \ R+ .
Maximal asymptotic transitive sets need not be closed or positively in-
variant. In fact, the following theorem shows that the maximal asymptotic
transitive sets which are closed and positively invariant are the positive min-
imal sets. It should be observed that a set X ⊂ M is positive minimal if and
only if X = ω + (x) for every x ∈ X, according to Theorem 3.1.1.
Theorem 5.2.5 A set X ⊂ M is a closed and positively invariant maximal
asymptotic transitive set if and only if X is a positive minimal set.
Proof. If X ⊂ M is a positive minimal set then ω + (x) = X for every
x ∈ X. Hence X is closed, positively invariant, and is contained in some max-
imal asymptotic transitive set R. Then R ⊂ ω + (x) = X, for any x ∈ X, and
thus X = R is a maximal asymptotic transitive set. As to the converse, sup-

pose that X is a closed and positively invariant maximal asymptotic transitive
set. If Y ⊂ X is nonempty, closed, and positively invariant, then ω + (y) ⊂ Y
for every y ∈ Y . It follows that X ⊂ ω + (y) ⊂ Y for every y ∈ Y . Hence
X = Y and therefore X is a positive minimal set.
Since positive limit sets are positively invariant, we have ω + (x) = ω + (y)
whenever x and y are equivalent positively recurrent points. Hence, if R is an
asymptotic transitive set then ω + (x) = ω + (y) for all x, y ∈ R, which means
there is a set R ⊂ M such that R ⊂ R and R = ω + (x) for all x ∈ R. It is
easily seen that R is a positive minimal set if and only if R = R.
The following example shows maximal asymptotic transitive sets which
are not positively invariant.
Example 5.2.2 Go back to Example 4.1.1. Any point x ∈ R2 is recurrent.

Since the origin 0 is a global stationary point, the set R0 = {0} is a minimal
set. For any pair of points x, y distinct from the origin, they are equivalent if
and only if x = y. Thus each circle centered at the origin is an asymptotic
transitive set. However, since the circles are not positively invariant, they are
not positive minimal sets.
Note that a limit set ω + (x) may intersect a maximal asymptotic transitive
set R even if x ∈/ R. Such a point x is said to be asymptotically attracted to
R. The following definition establishes this notion of asymptotic attraction.
Definition 5.2.4 Let R be a maximal asymptotic transitive set. The domain

of asymptotic attraction of R is defined by

Aa (R) = x ∈ M : ω + (x) ∩ R = ∅ .
This notion of asymptotic attraction yields the definition of a dynamic

preorder among the asymptotic transitive sets: for two asymptotic transitive
sets R1 and R2 ,
R1 a R2 if and only if R1 ∩ Aa (R2 ) = ∅.
Since R ⊂ ω + (x) for all x ∈ R, we have R ⊂ Aa (R). The proof for the
following proposition is similar to the proof for Proposition 5.1.4, by using
Proposition 2.2.6.
Proposition 5.2.4 Let R ⊂ M be an asymptotic transitive set. Then Aa (R) =

{x ∈ M : R ⊂ ω + (x)}.
For two asymptotic transitive sets R1 and R2 , we have R1 a R2 if and

only if there is x ∈ R1 such that R2 ⊂ ω + (x), which implies R2 ⊂ ω + (x) for
all x ∈ R1 . The proof for the following proposition is similar to the proof for
Proposition 5.1.5, by using Proposition 2.2.6.
Proposition 5.2.5 The relation “a ” among the asymptotic transitive sets
is a preorder.
The preorder “a ” is called the asymptotic transitivity preorder.

Positive minimal sets are upper bounds for the asymptotic transitivity pre-
order, as in the following:
Proposition 5.2.6 If R is a positive minimal set then R is maximal with

respect to the asymptotic transitivity preorder among the maximal asymptotic
transitive sets.
Proof. Suppose that R a R for some maximal asymptotic transitive

set R . Then there is x ∈ R such that R ⊂ ω + (x) = R. Hence R = R , and
therefore R is maximal with respect to the asymptotic preorder.
In compact state space, the converse for Proposition 5.2.6 holds, that is,
the positive minimal sets cover the upper bounds for the asymptotic transi-
tivity preorder.
Theorem 5.2.6 Assume that M is a compact manifold and let X ⊂ M be

a nonempty subset. Then X is a positive minimal set if and only if X is
an asymptotic transitive set that is maximal with respect to the asymptotic
preorder.
Proof. Suppose that X is an asymptotic transitive set that is maximal

with respect to the asymptotic preorder and take x ∈ X. Since ω + (x) is
compact and positively invariant, there is a positive minimal set N ⊂ ω + (x).
According to Theorem 5.2.5, N is an asymptotic transitive set and X a N .
Since X is maximal with respect to the asymptotic preorder, we have X = N .
Thus X is a positive minimal set. The converse is clear from Proposition
5.2.6.
As a consequence of Theorems 5.2.5 and 5.2.6, we have the following char-

acterizations of positive minimal sets in compact state space.
Theorem 5.2.7 Assume that M is a compact manifold and let X ⊂ M . The

following sentences are equivalent:
1. X is a positive minimal set.
2. X is a closed and positively invariant maximal asymptotic transitive set.
3. X is an asymptotic transitive set that is maximal with respect to the

asymptotic preorder.
Example 5.2.3 In Example 4.1.1, the circles centered at the origin are the
asymptotic transitive sets. The origin is a minimal set. For any circle Sr [0] =

x ∈ R2 : x = r , with r > 0, we have

Aa (Sr [0]) = x ∈ R2 : 0 < x ≤ r .
Hence Sδ [0] ⊂ Aa (Sr [0]) for all 0 < δ ≤ r. Thus Sδ [0] a Sr [0], whenever
0 < δ ≤ r, and the origin does not relate to any other asymptotic transitive
set.
Example 5.2.4 In Example 2.2.1, the origin 0 is a global stationary point

and {0} is then a minimal set. The four quadrants Q1 , Q2 , Q3 , Q4 are in-
variant sets. The semi-axes A1 = {(x1 , 0) : x1 ≥ 0}, A2 = {(0, x2 ) : x2 ≥ 0},
A3 = {(x1 , 0) : x1 ≤ 0}, and A4 = {(0, x2 ) : x2 ≤ 0} are also invariant sets.
For any point x ∈ A2 ∪ A4 (axe y), we have ω + (x) = {0}, ω − (x) =
∅ if x = 0, and ω − (0) = {0}. For x ∈ A1 \ {0}, we have ω + (x) =
ω − (x) = A1 . For x ∈ A3 \ {0}, we have ω + (x) = ω − (x) = A3 . If x ∈
int (Q1 ) = {(x1 , x2 ) : x1 , x2 > 0} or x ∈ int (Q4 ) = {(x1 , x2 ) : x1 > 0, x2 < 0}
then ω + (x) = A1 and ω − (x) = ∅. If x ∈ int (Q2 ) = {(x1 , x2 ) : x1 < 0, x2 > 0}
or x ∈ int (Q3 ) = {(x1 , x2 ) : x1 < 0, x2 < 0} then ω + (x) = A3 and ω − (x) =
∅. Thus the sets R0 = {0}, R1 = A1 \ {0}, and R3 = A3 \ {0} are the
asymptotic transitive sets of the control system and they are invariant sets.
Moreover, Aa (R0 ) = R2 , Aa (R1 ) = {(x1 , x2 ) : x1 > 0}, and Aa (R3 ) =
{(x1 , x2 ) : x1 < 0}. Thus R1 a R0 and R2 a R0 , but there is no relation
between R1 and R2 .
5.3 Poincaré recurrence theorem

We shall now present the Poincaré recurrence theorem for control systems.
This is based on the recurrence theorem for discrete dynamical systems (see
[74]). We refer to [50] for unexplained concepts of measure theory.
The Borel σ-algebra on the manifold M is a collection BM of subsets of M
which includes all open sets, closed sets, countable intersections of open sets,
countable unions of closed sets, and so forth, satisfying the basic properties
of being closed under countable unions and complements. The sets in BM are
called (Borel) measurable sets. A measure on BM is a function μ : BM →
∞
[0, ∞] such that (i) μ (∅) = 0, and (ii) if {En }n=1 is a sequence of disjoint sets
"∞ ∞
in BM then μ ( n=1 En ) = n=1 μ (En ). The basic properties of measures
are listed below:
1. If E, F ∈ BM and E ⊂ F then μ (E) ≤ μ (F ).
2. If E, F ∈ BM then μ (E ∪ F ) = μ (E) + μ (F ) − μ (E ∩ F ).
∞ "∞ ∞
3. If {En }n=1 is a sequence in BM then μ ( n=1 En ) ≤ n=1 μ (En ).
∞
4. If {E } is an increasing sequence in BM then
"∞ n n=1
μ ( n=1 En ) = limn→∞ μ (En ).
∞
5. If {En }n=1 is a decreasing sequence in BM and μ (Em ) < ∞ for some
∞
m then μ ( n=1 En ) = limn→∞ μ (En )
Let μ be a measure on BM . Then μ is called outer regular on the mea-

surable set E if μ (E) = inf {μ (U ) : U ⊃ E, U open} and inner regular on E
if μ (E) = sup {μ (K) : K ⊂ E, K compact}; μ is a Radon measure if it is
finite on compact sets, outer regular on all measurable sets, and inner regular
on all open sets. Let μ be a Radon measure and N the union of all open sets
U ⊂ M with μ (U ) = 0. Then N is open, μ (N ) = 0, and if V is open and
V \ N = ∅ then μ (V ) > 0. The complement of N is called the support of μ,
which is denoted by supp (μ).
Example 5.3.1 The Lebesgue measure on R is the unique measure μ on

BR such that μ ((a, b]) = b − a for all a, b ∈ R. The Lebesgue measure on Rn
is the n-fold product μn = μ × · · · × μ on BRn . The Lebesgue measure can be
straightforwardly extended to manifolds (see Appendix B).
Section 5.3 · Poincaré recurrence theorem 119
Example 5.3.2 A matrix Lie group is a closed subgroup G of the linear

Lie group Gl (n, R). Every matrix Lie group G is a submanifold of Gl (n, R).
A (left) Haar measure on G is a nonzero Radon measure μ on BG that
satisfies μ (gE) = μ (E), for all g ∈ G and E ∈ BG , and μ (U ) > 0 for all
nonempty open set U ⊂ G. For instance, a Haar measure on Gl (n, R) is
−n 2
defined as μ (g) = |det g| ν (g), where ν is the Lebesgue measure on Rn .
Every matrix Lie group G admits a positive Haar measure (see [50, Theorem
10.5]).
If μ (M ) = 1 then μ is called a probability measure. For the following,

we assume a probability measure μ on BM that is invariant by the control
system, which means that μ (ϕut (E)) = μ (E) for all E ∈ BM , t ∈ R, and
u ∈ U.
Lemma 5.3.1 Let μ be a probability measure on BM that is invariant by the

control system. If U is an open set in M then the set

' = x ∈ U : O+ (x) ∩ U = ∅ for all T ≥ 0
U >T
$ %
' .
is measurable and μ (U ) = μ U

Proof. Note that U' = x ∈ U : O>n +
(x) ∩ U = ∅ for all n ∈ N . For each
− − "
n ∈ N, define the set Un∗ = U \ O>n (U ). Since O>n (U ) = ϕu−t (U ) is
t>n,u∈U
open, Un∗ is measurable. It follows that U' is measurable, as U ' = U \ " U ∗.
n
n∈N
Now, denote by IM the identity map of M and set Sn = {ϕut : t > n, u ∈ U} ∪

{IM }. For t > n and u ∈ U , we have Un∗ ⊂ Sn−1 (U ) \ ϕu−t Sn−1 (U ) and

ϕu−t Sn−1 (U ) ⊂ Sn−1 (U ). Hence

μ (Un∗ ) ≤ μ Sn−1 U − μ ϕu−t Sn−1 (U ) = 0.
It follows that μ (Un∗ ) = 0 for all n ∈ N, and therefore

$ % #
μ U ' = μ (U ) − μ U ∗
= μ (U ) .
n
n∈N
The following result is the Poincaré recurrence theorem for control systems.
Theorem 5.3.1 Let μ be a probability measure on the Borel sets of M that

is invariant by the control system. Then μ (R+ ) = 1 or, in other words,
μ-almost every point is positively recurrent.
Proof. Let X = {xn }n∈N be a countable dense subset of M . Define
(
∞ "

M= B (x, 1/n), where
n=1x∈M

B (x, 1/n) = y ∈ B (x, 1/n) : O>k
+
(y) ∩ B (x, 1/n) = ∅ for all k ∈ N
as in Lemma 5.3.1. For any z ∈ M ( and a neighborhood V of z, there is n

"
such that B (z, 1/n) ⊂ V . Since z ∈ B (x, 1/2n), it follows that there is
x∈M
x ∈ M such that B (x, 1/2n) ⊂ V and O>k +
(z) ∩ B (x, 1/2n) = ∅ for all k ∈ N.
Hence O>k (z) ∩ V = ∅ for all k ∈ N. It follows that
+
$ z ∈% ω + (z), and therefore
( ⊂ R+ . According to Lemma 5.3.1, we have μ U \ U
M ' = 0 for all open set
"
U . Since M = B (x, 1/n), we have
x∈M
# #
M\
B (x, 1/n) ⊂
B (x, 1/n) \ B (x, 1/n)
x∈M x∈M
and then

# $ %
μ (M ) − μ
B (x, 1/n) ≤
μ B (x, 1/n) \ B (x, 1/n) = 0.
x∈M x∈M

"
Hence, μ B (x, 1/n) = μ (M ) = 1. It follows that
x∈M

$ % #
( = lim μ
μ M
B (x, 1/n) =1
n→∞
x∈M
and therefore μ (R+ ) = 1.
We ask about a situation where the Birkhoff center of a control system

covers the whole state space, that is, the positively recurrent points are dense
in M . The Poincaré recurrence theorem provides an answer for this question.
Corollary 5.3.1 Let the assumptions be as in the above theorem and assume
further that μ is a Radon measure. Then supp (μ) ⊂ cl (R+ ). In particular,
if μ (U ) > 0 for every nonempty open set U ⊂ M then R+ is dense in M ,
that is, the Birkhoff center coincides with M .
Section 5.4 · Control sets 121
Proof. The set U = M \ cl (R+ ) is open and μ (U ) = 0, since μ (R+ ) = 1,

according to Theorem 5.3.1. Hence supp (μ) ⊂ M \ U = cl (R+ ). Now, if
μ (U ) > 0 for any nonempty open set U then supp (μ) = M and therefore
cl (R+ ) = M .
Example 5.3.3 Let M = O (n, R) be the matrix Lie group of the orthog-
onal compact matrices in Gl (n, R). A left Haar measure on G is finite,
and it may then be normalized, becoming a probability measure μ. Consider
m
a control affine system ẋ = X0 (x) + i=1 ui (t) Xi (x) on O (n, R) where
X0 , X1 , ..., Xm are n × n antisymmetric matrices. The solutions of the system
satisfy ϕ (t, g, u) = ϕ (t, 1, u) g for all t ∈ R, g ∈ O (n, R), and u ∈ U . We
then have μ (ϕut (E)) = μ (ϕut (1) E) = μ (E) for all E ∈ BM , t ∈ R, and
u ∈ U. According to the Poincaré recurrence theorem, the Birkhoff center of
the system coincides with O (n, R).
5.4 Control sets

We shall now study the special class of control sets. A control set is an
asymptotic transitive set that contains entire positive semi-trajectories. We
are specially interested in control sets given by asymptotic transitive sets with
nonvoid interior. We introduce the concept of accessibility and show how this
concept influences the controllability.
We consider again relative limit sets and denote by R the set of all relative
positively recurrent points. Then R is the subset of R+ given by

R = x ∈ M : there is u ∈ U such that x ∈ ω +
u (x) .
For each x ∈ R , Rx = Rx ∩ R is the equivalence class of x with respect

to the asymptotic transitivity relation restricted to R . In general, R is a
proper subset of R+ (see Problem 5.8(21)). Remembering that P denotes the
set of all periodic points of the system, we have P ⊂ R ⊂ R+ , and then

Hx ⊂ R x ⊂ R x .
The following result shows a special property of the sets Rx .
Proposition 5.4.1 Let x ∈ R . For each y ∈ Rx , there is a control function

u ∈ U with ϕ (t, y, u) ∈ Rx for all t ≥ 0, that is, there is a positive semi-
trajectory through y entirely contained in Rx .
Proof. For y ∈ Rx there is u ∈ U such that y ∈ ω + u (y). Since ω u (y) ⊂

+
ω + (y) and ω + (y) is positively invariant, it follows that ϕ (t, y, u) ∈ ω + (y)

for all t ≥ 0. Take a sequence tn → +∞ such that ϕ (tn , y, u) → y. As
ϕ (tn , y, u) = ϕ (tn − t, ϕ (t, y, u) , u · t), with tn − t → +∞, we have y ∈
ω + (ϕ (t, y, u)) for all t ≥ 0. Since ω + (ϕ (t, y, u)) is positively invariant, we
have ϕ (t, y, u) ∈ ω + (ϕ (t, y, u)) for all t ≥ 0, that is, ϕ (t, y, u) ∈ R for all
t ≥ 0. Thus Rϕ(t,y,u) = Ry = Rx , for all t ≥ 0. This proves the proposition.
This property of relative recurrent points joins with the weak transitivity
to define the concept of a control set, as in the following:
Definition 5.4.1 A nonempty subset D ⊂ M is called a control set if it

satisfies the following conditions:
1. For each x ∈ D, there is a control function u ∈ U with ϕ (t, x, u) ∈ D for

all t ≥ 0, that is, there is a positive semi-trajectory through x entirely
contained in D.
+
2. D ⊂ cl O>0 (x) for every x ∈ D (weak transitivity).
3. D is maximal satisfying both properties (1) and (2), that is, if D con-
tains D and satisfies both properties (1) and (2) then D = D.
For positively invariant sets, the concepts of control set and weak control
set have the same meaning. If the whole state space M is a control set itself
(or equivalently, a weak control set) then it is the unique control set. In this
case we say that the control system is controllable.
Example 5.4.1 A control system may have no control set. Consider the
control system ẋ = u (t) on R with control range U ⊂ (0, +∞). This system
does not have a control set. Indeed, the solutions of the system are of the form
t
ϕ (t, x, u) = x + u (s) ds
0

t
where 0 u (s) ds is positive and increasing with respect to t. Suppose that
D is a control set of the system. Then, for a given x ∈ D, there is u ∈ U
such that ϕ (t, x, u) ∈ D for all t ≥ 0. Take y > x such that y = ϕ (t1 , x, u)
with t1 > 0. Then y ∈ D. Now, for any z ∈ O>0 +
(y), we have z > y,
+
hence O>0 (y) ⊂ (y, +∞). Thus x ∈
+
/ cl O>0 (y) , which contradicts the weak
transitivity of D.
ẋ = u1 (t) X1 (x) + u2 (t) X2 (x)
on M = R2 , with control range

U = (u1 , u2 ) ∈ R2 : u1 + u2 = 1, u1 ∈ [0, 1]

∂ 0 −1
where X1 = and X2 = . The trajectories of X1 are vertical
∂x2 1 0
lines and the trajectories of X2 are circles centered at 0. By piecing together
the trajectories of X1 and X2 we can connect any pair of points in the plane
(see Fig. 5.1). Thus the control system is controllable.
Figure 5.1: A controllable system.
Example 5.4.3 Consider the control affine system on R2 determined by the

commutative linear vector fields

−1 0 1 1
X= Y = .
0 −1 −1 1
The origin 0 is a global stationary point. The trajectories of X are straight

lines moving to 0 and the trajectories of Y move on spirals from 0. By piecing
together the trajectories of X and Y we can connect any pair of points distinct
from the origin (see Fig. 5.2). By deleting the origin, the resulting control
system on the manifold M = R2 \ {0} is controllable.
Figure 5.2: An attaining broken integral curve.
In order to avoid trivial cases of control sets, for instance a global critical
point, one may require control sets with a nonempty interior. We afterward
show that a weak control set with interior points is a control set. We shall
first relate the control sets to the asymptotic transitive sets and the set R .
Proposition 5.4.2 Each control set is contained in some maximal asymp-

totic transitive set.
Proof. If D is a control set and x ∈ D then there is a control func-

tion u ∈ U with ϕ (t, x, u) ∈ D for all t ≥ 0. This implies that D ⊂
+ +
cl O>0 (ϕ (t, x, u)) ⊂ cl O>t (x) for all x ∈ D and t ≥ 0. Thus x ∈ R+
and D ⊂ Rx .
Similarly to the asymptotic transitive sets, the control sets also define a
partition of the state space M , as follows.
Proposition 5.4.3 If D and D are two control sets then either D ∩ D = ∅

or D = D .
Proof. Suppose that D ∩ D = ∅ and take x ∈ D ∩ D . We show that

the set D ∪ D satisfies items (1) and (2) of Definition 5.4.1. In fact, for a
given y ∈ D ∪ D , we may assume that y ∈ D. Since D is a control set, there
is a control function u ∈ U such that ϕ (t, y, u) ∈ D for all t ≥ 0. Hence
ϕ (t, y, u) ∈ D ∪ D for all t ≥ 0, and therefore D ∪ D satisfies item (1).
Now let y, z ∈ D ∪ D be arbitrary. We may assume that y ∈ D. If z ∈ D,
+
we have z ∈ cl O>0 (y) , because D is a control set. If z ∈ D , we have
+ +
z ∈ cl O>0 (x) and x ∈ cl O>0 (y) , since x ∈ D ∩ D and both D and D are
+
control sets. According to Proposition 2.2.6, it follows that z ∈ cl O>0 (y) .
+
Hence D∪D ⊂ cl O>0 (y) for every y ∈ D∪D , and therefore D∪D satisfies
item (2). By the maximality of both D and D , we have D = D ∪ D = D .
In general, however, a maximal asymptotic transitive set need not be a

control set (see Problem 5.8(21)).
In order to relate the control sets to the set R , we need the following:
Proposition 5.4.4 Let D ⊂ M be a set satisfying both conditions (1) and

(2) of Definition 5.4.1. D is then contained in a control set.
Proof. Let D be the collection of all subsets of M containing D and

satisfying both condition (1) and (2) of Definition 5.4.1. Consider D ordered
by the set inclusion. Note that D is nonempty since D ∈ D. Let {Ci }i∈I be
"
a totally ordered subcollection of D. The reunion C = Ci is a set of D.
i∈I
In fact, for a given x ∈ C, we have x ∈ Cj for some j ∈ I. As Cj satisfies
condition (1), there is a control function u ∈ U such that ϕ (t, x, u) ∈ Cj for
all t ≥ 0. Hence ϕ (t, x, u) ∈ C for all t ≥ 0, and therefore C satisfies the
condition (1). Now for any pair of points x, y ∈ C, we have x ∈ Ci and y ∈ Cj
for some i, j ∈ I. Choose a point z ∈ D. As D ⊂ Ci and D ⊂ Cj , we have
+ + +
x ∈ cl O>0 (z) and z ∈ cl O>0 (y) , and therefore x ∈ cl O>0 (y) . Hence
C also satisfies the condition (2), and therefore C ∈ D. According to the
Zorn lemma, D has a maximal element Cm . To show that Cm is a control
set, it remains to prove that Cm is maximal satisfying both conditions (1)
and (2). Indeed, if C ⊂ C and C satisfies both the conditions (1) and (2)
then C ∈ D. By the maximality of Cm in D, it follows that C = C and the
proposition is proved.
We now prove that the reunion of the control sets contains the set R .
Proposition 5.4.5 Each equivalence class Rx = Rx ∩ R is contained in a

control set.
Proof. For each y ∈ Rx , we use Proposition 5.4.1 to find uy ∈ U with

" uy +
ϕ (t, y, uy ) ∈ Rx for all t ≥ 0. Then Rx ⊂ ϕy (R ). By defining D =

y∈Rx
" u
ϕy y (R+ ), we have D ⊂ Rx . If z ∈ D then z = ϕ (t, y, uy ) for some

y∈Rx
y ∈ Rx and t ≥ 0. It follows that
ϕ (s, z, uy · t) = ϕ (s, ϕ (t, y, uy ) , uy · t) = ϕ (t + s, y, uy ) ∈ D

+
for all s ≥ 0. Moreover, since z ∈ Rx , we have D ⊂ Rx ⊂ cl O>0 (z) . Hence
D satisfies properties (1) and (2) of Definition 5.4.1. The result follows ac-
cording to Proposition 5.4.4.
Thus, for any positively recurrent point x ∈ R+ , we have the inclusions
Hx ⊂ Rx ⊂ D ⊂ Rx
where D is the control set that contains x.

From now on, we study control sets with interior points. They are cru-
cial for controllability problems. We might assume the following accessibility
condition on the semi-orbits.
Definition 5.4.2 The control system is said to be accessible from x ∈ M

if both O+ (x) and O− (x) have nonvoid interior$ in %M ; the system
$ is called
%
−
locally accessible from x ∈ M if both int O≤T (x) and int O≤T (x) are
+
nonempty for all T > 0. If these properties hold for all x ∈ M the system is
called accessible or locally accessible, respectively.
In the following we have an important property of local accessibility:
Proposition 5.4.6 If the control system is locally accessible, then for every
x ∈ M the interior int (O+ (x)) is dense in O+ (x) and int (O− (x)) is dense
in O− (x).
Proof. Suppose by contradiction that int (O+ (x)) is not dense in O+ (x).
There are then y ∈ O+ (x) and ε > 0 such that B (y, ε)∩int (O+ (x)) is empty.
By continuity, we can find T > 0 such that O≤T+
(y) ⊂ B (y, ε). According to
$ %
the hypothesis, we have int O≤T (y)+
= ∅. However
$ %
int O≤T
+
(y) ⊂ int (O+ (x)), and we then have a contradiction. Similarly
we can prove that int (O− (x)) is dense in O− (x).
This result is used afterwards. Let us see some examples of accessibility

and local accessibility.
$ %
Example 5.4.4 In Example 2.2.2, both the sets int O≤T +
(x) and
$ %
−
int O≤T (x) are nonempty for all T > 0 and x distinct from the origin.
$ %
Example 5.4.5 In Example 2.2.1, both the sets int O≤T +
(x) and
$ %
−
int O≤T (x) are nonempty for all T > 0 and x out of the axes. Never-
theless, the system is not locally accessible from any point on the axes.
ẋ = u1 (t) X1 (x) + u2 (t) X2 (x)
on R2 , with control range

U = (u1 , u2 ) ∈ R2 : u1 + u2 = 1, u1 ∈ [0, 1]
∂ ∂
where X1 = and X2 = f (x1 ) , with f a real-valued C ∞ function such
∂x1 ∂x2
that f (x1 ) = 0 for x1 ≤ 0 and f (x1 ) > 0 for x1 > 0. An example of vector
field like X2 is illustrated in Fig. 5.3. A point (x1 , x2 ) with x1 ≤ 0 is an equi-
librium point for X2 . The trajectory of X2 through a point (x1 , x2 ) with x1 > 0
is a vertical line. The trajectories of X1 are horizontal lines. The system is
then locally accessible from every point x = $ (x1 , x2 )% with x1 ≥
$ 0. However,
%
−
for x = (x1 , x2 ) with x1 < 0, the sets int O≤T +
(x) and int O≤T (x) are
empty for T > 0 sufficiently small (see Fig. 5.4). Thus the control system is
not locally accessible from x = (x1 , x2 ) with x1 < 0. Otherwise, this system is
accessible.
The following result shows that the condition of a control set to have entire
positive semi-trajectories may be omitted if it has a nonempty interior.
Proposition 5.4.7 If D ⊂ M is a weak control set with int (D) = ∅ then D

is a control set.
Proof. It is enough to prove that D satisfies the item (1) of Definition

+
5.4.1. In fact, for a given x ∈ D we have int (D) ⊂ D ⊂ cl O>0 (x) . By
choosing two distinct points y, z ∈ int (D), we can take a neighborhood Vy of
y and a neighborhood Vz of z such that Vy , Vz ⊂ int (D) and Vy ∩ Vz = ∅.
There is then x1 ∈ O>0 +
(x) ∩ Vy . Hence there are t0 > 0 and u0 ∈ U such that
+
x1 = ϕ (t0 , x, u0 ) ∈ Vy . Since D ⊂ cl O>0 (x1 ) , there is x2 ∈ O>0
+
(x1 ) ∩ Vz ,
∂
Figure 5.3: Plot of the vector field Y = f (x1 ) with f (x1 ) = 0 for x1 ≤ 0
∂x2
and f (x1 ) = x21 for x1 > 0.
Figure 5.4: Positive semi-orbit through a point (x1 , x2 ) with x1 < 0.
and hence there are t1 > 0 and v1 ∈ U such that x2 = ϕ (t1 , x1 , v1 ) ∈ Vz .

By taking the t0 -concatenation u1 of u0 and v1 , we have ϕ (t0 + t1 , x, u1 ) =
+
ϕ (t1 , ϕ (t0 , x, u0 ) , v1 ) = x2 . Now, as D ⊂ cl O>0 (x2 ) , there are t2 > 0
and v2 ∈ U such that x3 = ϕ (t2 , x2 , v2 ) ∈ Vy . By taking the (t0 + t1 )-
concatenation u2 of u1 and v2 , we have ϕ (t0 + t1 + t2 , x, u2 ) = x3 . Following

∞
by induction we obtain a control function u and a series ti such that
k k i=0

ϕ ti , x, u ∈ Vy , for k even, and ϕ ti , x, u ∈ Vz , for k odd. Ac-
i=0 i=0

∞
cording to Proposition 3.2.3, ϕ (t, x, u) ∈ D for all 0 ≤ t < ti . Moreover,
i=0

∞
since Vy ∩ Vz = ∅, the series ti diverges, and therefore ϕ (t, x, u) ∈ D for
i=0
all t ≥ 0.
A control set with nonempty interior is called a main control set. This
motivates the notion of a control set for general semigroup actions (see Prob-
lem 5.8(9)).
Definition 5.4.3 Let D ⊂ M be a main control set. The transitivity set

of D is the set D0 defined by
+ −
D0 = x ∈ D : x ∈ int O>0 (x) ∩ int O>0 (x) .
The main control set D is called effective if D0 = ∅.

+ −
By Corollary 2.2.1, x ∈ int O>0 (x) ∩ int O>0 (x) implies that both
−
semi-orbits O>0
+
(x) and O>0 (x) are open sets in M . Thus the transitivity
set of a control set D is
−

D0 = x ∈ D : x ∈ O>0
+
(x) ∩ O>0 (x)
−
with O>0
+
(x) ∩ O>0 (x) open in M . The transitivity set of a main control set
is an open holding set that is dense and positively invariant in D, as stated
by the following theorem.
Theorem 5.4.1 Let D ⊂ M be an effective control set. Then

−
1. D ⊂ O>0 (x) for every x ∈ D0 .
2. D0 is an open holding set.
3. For any pair x, y ∈ D0 , there are t > 0 and u ∈ U such that ϕ (t, x, u) =
y.
4. D0 is dense in D, that is, cl (D0 ) = cl (D).
5. D0 is positively invariant in D, that is, if ϕ (t, x, u) ∈ D for some t > 0,

u ∈ U, and x ∈ D0 , then ϕ (t, x, u) ∈ D0 .
−
Proof. (1) Let y ∈ D and x ∈ D0 . Since O>0 (x) is an open set and
+ − −
x ∈ cl O>0 (y) , we have O>0 (x) ∩ O>0 (y) = ∅. Take z ∈ O>0
+
(x) ∩ O>0
+
(y).
− −
Then y ∈ O>0 (z) and z ∈ O>0 (x). According to Proposition 2.2.5, we have
− −
y ∈ O>0 (x). Therefore D ⊂ O>0 (x).
−
(2) If we show that D0 = O>0 +
(x) ∩ O>0 (x), for every x ∈ D0 , then
D0 is an open holding set according to Proposition 5.1.1. Let x, y ∈ D0 .
− −
By item (1), we have y ∈ O>0 (x) and x ∈ O>0 (y). Hence y ∈ O>0 +
(x) ∩
− −
O>0 (x), and therefore D0 ⊂ O>0 (x) ∩ O>0 (x). On the other hand, let
+
−
x ∈ D0 and y ∈ O>0 +
(x) ∩ O>0 (x). We have y ∈ D. For indeed, there are
t1 , t2 > 0 and u1 , u2 ∈ U such that ϕ (t1 , x, u1 ) = y and ϕ (t2 , y, u2 ) = x. By
taking the t1 -concatenation u3 of u1 and u2 , we have ϕ (t1 + t2 , x, u3 ) = x.
According to Proposition 3.2.3, ϕ (t, x, u3 ) ∈ D for all t ∈ [0, t1 + t2 ]. Hence
−
y = ϕ (t1 , x, u3 ) ∈ D. Now, for any other z ∈ O>0 +
(x) ∩ O>0 (x), we have
z ∈ O>0 (x) and x ∈ O>0 (y). According to Proposition 2.2.5, it follows
+ +
− −
that z ∈ O>0 +
(y). Moreover, we have z ∈ O>0 (x) and x ∈ O>0 (y), hence z ∈
− − − −
O>0 (y). Thus, O>0 (x)∩O>0 (x) ⊂ O>0 (y)∩O>0 (y). Since O>0
+ + +
(x)∩O>0 (x)
+ −
is an open set, it follows that y ∈ int O>0 (y) ∩ int O>0 (y) , and therefore
−
y ∈ D0 . It follows that O>0 +
(x) ∩ O>0 (x) = D0 .
(3) For any pair x, y ∈ D0 , we have y ∈ O>0 +
(x), by item (2). Hence there
are t > 0 and u ∈ U such that ϕ (t, x, u) = y.
−
(4) For any x ∈ D0 , we have D ⊂ O>0 (x), by item (1). As x ∈ D, we also
+ + −
have D ⊂ cl O>0 (x) . Hence D ⊂ cl O>0 (x) ∩ O>0 (x). Now, by item (2),
−
we have D0 = O>0 (x) ∩ O>0 (x). Hence
+
+ −
+ −

D ⊂ cl O>0 (x) ∩ O>0 (x) ⊂ cl O>0 (x) ∩ O>0 (x) = cl (D0 )
and therefore D0 is dense in D.

(5) Suppose that ϕ (t, x, u) ∈ D for some t > 0, u ∈ U , and x ∈ D0 . By
−
item (1), ϕ (t, x, u) ∈ O>0 (x). As ϕ (t, x, u) ∈ O>0
+
(x), we have ϕ (t, x, u) ∈
−
O>0 (x) ∩ O>0 (x). By item (2), it follows that ϕ (t, x, u) ∈ D0 .
+

This theorem guarantees that a control set D contains at most one holding
set, that is, its transitivity set D0 . On the other hand, any open holding set
is a transitivity set of some effective control set, as follows:
Proposition 5.4.8 If x ∈ P then the holding set Hx is contained in some

control set D. Furthermore, if Hx is open then Hx = D0 .
Proof. For the first part of the theorem, it is enough to prove that Hx
satisfies the item (1) of Definition 5.4.1, since Hx is an asymptotic transi-
tive set. Indeed, for a given y ∈ Hx , there is τ > 0 and u1 ∈ U such that
ϕ (τ , y, u1 ) = y. By taking the τ -concatenation u2 of u1 and u1 , we have

ϕ (2τ , y, u2 ) = y. By taking the 2τ -concatenation u3 of u2 and u1 , we have
ϕ (3τ , y, u3 ) = y. By induction, we can define a control function u such
that ϕ (kτ , y, u) = y for all k ∈ N. According to Proposition 5.1.2, we have
ϕ (t, y, u) ∈ Hx for all t ≥ 0. Thus for any y ∈ Hx there is a positive semi-
trajectory through y entirely contained in Hx . For the second part of the
theorem, we suppose that Hx is open (that is equivalent to int (Hx ) = ∅).
According to Theorem 2.3.1, Hx ⊂ D0 . According to Theorem 5.4.1, D0 is
also a holding set, and therefore Hx = D0 .
Thus the open holding sets coincide with the transitivity sets of the effec-
tive control sets. Another consequence of Theorem 5.4.1 is that any effective
control set D is connected, since its transitivity set D0 is dense in D and D0
is pathwise connected by Proposition 5.1.3. However, control sets with an
empty interior need not be connected, for the same reason that weak control
sets need not (Problem 3.4(3)).
Finishing the explanation on the fundamental properties of the effective
control sets, note that D0 ⊂ int (D) for any effective control set D. We then
ask about the possibility of D0 = int (D). We explore the local accessibility
in order to state a condition for this equality.
Theorem 5.4.2 Let D ⊂ M be a main control set. If the control system is
locally accessible from some y ∈ int (D) then D is an effective control set,
int (D) = D0 , and D = cl (O+ (x)) ∩ O− (x) for all x ∈ D0 .
Proof. By repeating the proof for Proposition 5.4.6, int (O− (y)) is dense
in O− (y) for every y ∈ int (D). Hence int (D) ∩ int (O− (y)) = ∅ for ev-
ery y ∈ int (D). Since int (D) ∩ int (O− (y)) ⊂ D ⊂ cl (O+ (x)), there is
z ∈ int (D) ∩ int (O− (y)) ∩ O+ (x). Then y ∈ O+ (z) and z ∈ O+ (x), hence
y ∈ O+ (x). Thus int (D) ⊂ O+ (x) for every x ∈ D, or in other words,
D ⊂ O− (y) for all y ∈ int (D). In particular, int (D) ⊂ O+ (y) ∩ O− (y) for
every y ∈ int (D). Hence int (D) ⊂ D0 and therefore int (D) = D0 . Now,
as D ⊂ cl (O+ (x)) for all x ∈ D, we have D ⊂ cl (O+ (y)) ∩ O− (y) for all
y ∈ int (D). On the other hand, if z ∈ cl (O+ (y))∩O− (y) for y ∈ int (D) then
z and y are weak equivalent, hence z ∈ D. Therefore D = cl (O+ (y))∩O− (y)
for all y ∈ int (D).
Thus, for locally accessible control systems, the effective control sets coin-
cide with the main control sets.

Theorem 5.4.2 reveals a situation where an equivalence class Rx is an
asymptotic transitive set, as in the following:
Theorem 5.4.3 Let x ∈ R+ and assume that the asymptotic transitive set
Rx has interior points. If the control system is locally accessible from some
y ∈ int (Rx ) then Rx = Rx and int (Rx ) = Hy .
Proof. Note that Rx is contained in some weak control set D. Hence
y ∈ int (Rx ) ⊂ int (D). According to Proposition 5.4.7, D is a main control
set, and thus D ⊂ Rx . This means that D = Rx . According to Theorem 5.4.2,
int (Rx ) = D0 = Hy , and therefore cl (int (Rx )) = cl (D0 ) = cl (Rx ) according
to Theorem 5.4.1. Now let z ∈ Rx and take a sequence of positive numbers
εn → 0. Since int (Rx ) is dense in Rx , we have B (z, ε1 ) ∩ int (Rx ) = ∅, and
there are then t1 > 0 and u1 ∈ U such that z1 = ϕ (t1 , z, u1 ) ∈ B (z, ε1 ) ∩ Rx .
Since z1 ∈ Rx , there are t2 > 0 and u2 ∈ U such that z2 = ϕ (t2 , z1 , u2 ) ∈
B (z, ε2 ) ∩ Rx . Following by induction, for each n ∈ N, we can find tn > 0
and un ∈ U such that zn = ϕ (tn , zn−1 , un ) ∈ B (z, εn ) ∩ Rx . We then define
the function u (t) ∈ U by
n−1 )

n−1 n
u (t) = un t − ti if t ∈ ti , ti ,
i=0 i=0 i=0
where t0 = 0. We may choose the sequence (ti ) in such a way that the series

∞
ti diverges. For t ≤ 0, u (t) = 0. By construction we have
i=0

n
ϕ ti , z, u = ϕ (tn , zn−1 , un ) → z
i=0
and hence z ∈ ω + (z). Thus Rx ⊂ R , which means that Rx = Rx .
Under the assumptions and notations of Theorem 5.4.3, we have the rela-
tions
Hy = D0 ⊂ Rx = D = Rx ,
cl (Hy ) = cl (Rx ) = cl (Rx ) .
5.5 Invariant control sets

We shall now study a special class of control sets called invariant control
sets. This approaches the notion of positive minimal sets.
Section 5.5 · Invariant control sets 133
Definition 5.5.1 A nonempty subset C ⊂ M is called an invariant control

set if it satisfies the following properties:
1. For each x ∈ C, there is a control function u ∈ U such that ϕ (t, x, u) ∈

C for all t ≥ 0.
+
2. For every x ∈ C, cl (C) = cl O>0 (x) .
3. C is maximal satisfying the properties (1) and (2).
An invariant control set need not be invariant, although it is named ‘in-

variant’. See Example 5.5.1 below for an invariant control set that is not
invariant. Otherwise, the closure of an invariant control set is positively in-
variant, by definition. The closed invariant control sets relates to the positive
minimal sets, as follows:
Theorem 5.5.1 A subset X ⊂ M is a positive minimal set if and only if it

is a closed invariant control set.
+
Proof. Suppose that X is a positive minimal set. Then X = cl O>0 (x)
for all x ∈ X. For a given x ∈ X, it follows that ϕ (t, x, u) ∈ O>0
+
(x) ⊂ X for
any t > 0 and u ∈ U . Since X is an asymptotic transitive set, it follows that
X is a closed invariant control set. As to the converse, suppose that X is a
+
closed invariant control set. Then X = cl (X) = cl O>0 (x) for all x ∈ X.
Thus X is a positive minimal set.
An invariant control set may not be closed, even in compact state space
(see Problem 5.8(58)). The existence of (closed) invariant control sets is dis-
cussed in Problems 5.8(26, 27).
There exists the possibility of an invariant control set to be contained in
the closure of another invariant control set (see Example 5.1.3). This fact
does not occur under locally accessibility, as in the following:
Theorem 5.5.2 Assume that the control system is locally accessible from all
x in the closure of an invariant control set C. Then
1. C is closed, connected, and positively invariant.
2. int (C) is nonempty, positively invariant, and cl (int (C)) = C.
3. int (C) ⊂ O+ (x) for all x ∈ C and int (C) = O+ (x) for x ∈ int (C).
Proof. (1) According to Propositions 2.2.1, 2.1.2, and 2.2.3, the closure
of a positive semi-orbit is connected and positively invariant, hence cl (C) is
+
connected and positively invariant. If y ∈ cl (C) then y ∈ cl O>0 (x) for all
+ + +
x ∈ C and int O>0 (y) = ∅. Hence cl O>0 (y) ⊂ cl O>0 (x) = cl (C) for
+
all x ∈ C. In particular, ∅ = int O>0 (y) ⊂ cl (C) and then there is a point
+
z ∈ int O>0 (y) ∩ C. Thus we have
+ +
cl (C) = cl O>0 (z) ⊂ cl O>0 (y)
+
which implies cl (C) = cl O>0 (y) , since cl (C) is positively invariant. Hence
cl (C) satisfies the properties (1) and (2) of Definition 5.5.1. By the maximal-
ity of C, it follows that cl (C) = C.
(2) Let x ∈ C. By item (1) and local accessibility, we have
+
∅ = int O>0 (x) ⊂ C, hence int (C) = ∅. By item (1) and Proposi-
tion 2.1.2, int (C) is positively invariant. Now, since C is closed, we have
cl (int (C)) ⊂ C. For the converse, take x ∈ C. Since C is positively invari-
ant, we have O+ (x) ⊂ C and hence int (O+ (x)) ⊂ int (C). By Proposition
5.4.6, int (O+ (x)) is dense in O+ (x). Hence
+ +
C = cl O>0 (x) = cl int O>0 (x) ⊂ cl (int (C))
and therefore cl (int (C)) = C.

+ +
(3) For x ∈ C and y ∈ int (C) we have C = cl O>0 (x) = cl O>0 (y)
and int (C ∩ O− (y)) = ∅. Then there is z ∈ O>0 +
(x) ∩ int (C) ∩ int (O− (y)),
which means y ∈ O (x). Thus int (C) ⊂ O (x) for all x ∈ C. In particular,
+ +
for y ∈ int (C), we have O>0

+
(y) ⊂ int (C), since int (C) is positively invariant
by item (2). Therefore int (C) = O+ (y) for y ∈ int (C).
We finally show a relation between the closed main control sets and the
invariant control sets.
Theorem 5.5.3 Let D ⊂ M be a main control set and assume that the con-
trol system is locally accessible from all x in the closure of D. Then D is an
invariant control set if and only if it is closed.
Proof. Assume that D is closed. Suppose by contradiction that D is not

an invariant control set. There are then x ∈ D, t0 > 0, and u ∈ U such that
ϕ (t0 , x, u) ∈
/ D. Define
t∗ = sup {t < t0 : ϕ (t, x, u) ∈ D} .

Section 5.5 · Invariant control sets 135
Then ϕ (t∗ , x, u) ∈ fr (D). According to Theorem 5.4.2, we have

ϕ (t∗ , x, u) ∈ cl (D) = D = cl O+ (y) ∩ O− (y)
for y ∈ D0 = int (D). Since O− (y) is an open set whenever y ∈ D0 , we can

take t > t∗ such that ϕ (t, x, u) ∈ O− (y). As cl (O+ (y)) is positively invari-
ant, it follows that ϕ (t, x, u) ∈ cl (O+ (y)) ∩ O− (y) = D, which contradicts
the definition of t∗ . Therefore D is an invariant control set. The converse
follows by Theorem 5.5.2.
The following result is an immediate consequence of Theorems 5.5.1, 5.5.2,

and 5.5.3.
Corollary 5.5.1 Let D ⊂ M be a main control set and assume that the
control system is locally accessible from all x in the closure of D. The following
statements are equivalent:
1. D is closed.
2. D is an invariant control set.
3. D is a positive minimal set.
Figure 5.5: A positive minimal set not negatively invariant.
ẋ = u1 (t) X1 (x) + u2 (t) X2 (x)
on M = R2 \ {0}, with control range

U = (u1 , u2 ) ∈ R2 : u1 + u2 = 1, u1 ∈ [0, 1/2]
where
X1 (x1 , x2 ) = (x1 + x2 , x2 − x1 )

x1 − x31 − x1 x22 x2 − x21 x2 − x32
X2 (x1 , x2 ) = , .
x21 + x22 x21 + x22
By using polar coordinates (r, θ), the corresponding system is
θ −u1 (t)
=
1 − r2 2r2 − 1
r = + u1 (t) .
r r

For u1 = 0, we have θ = α and 1 − r2 = e−2t eβ , where α, β are constant.
In this case, the points in the unit circle S1 are equilibrium points and the
trajectories outside S1 are straight lines converging to S1 as t → +∞.
For u1 = 1/2, we have θ = −t/2 + α and r2 = t + β, with α, β constant.
In this case, the trajectories move on spirals to infinity as t → +∞.
2 − 2u1
In the case 0 < u1 < 1/2, we have θ = −u1 t + α and r2 = +
2 − 4u1
2 − 2u1
βe(4u1 −2)t , with α, β constant, −2 < 4u1 − 2 < 0, and 1 < . The
* 2 − 4u1
2 − 2u1
circle C with radius is a periodic trajectory and the trajectories
2 − 4u1
outside C move on spirals coming near C as t → +∞.
All these cases are illustrated in Fig. 5.5. The system is locally accessible

and the set D = x ∈ R2 : x ≥ 1 satisfies cls (O+ (x)) = D for every x ∈
D. Hence D is a positive minimal set (or closed invariant control set), with
D0 = int (D). Note that D is not negatively invariant.
5.6 Control sets for linear control systems

In this section we study control sets for the special case of linear control
systems. We introduce the Kalman condition to describe the effective control
sets.
Consider a linear control system
ẋ = Ax + Bu (t)
on Rd with piecewise constant functions, where A is a d × d real matrix and
⎛ ⎞
b1i
⎜ .. ⎟
B is a d × m real matrix with columns bi = ⎝ . ⎠, i = 1, ..., m. This system
bdi
Section 5.6 · Control sets for linear control systems 137
corresponds to the control affine system

m
ẋ = Ax + ui (t) bi (LΣ)
i=1
where each bi defines a constant vector field.

For a given control function u, the solution is of the form
t
tA
ϕ (t, x, u) = e x + e tA
e−sA Bu (s) ds
0
for all x ∈ R , where e

d tA
denotes the exponential matrix. In particular, for
x = 0 we have t
ϕ (t, 0, u) = etA e−sA Bu (s) ds
0
and thus
ϕ (t, x, u) = etA x + ϕ (t, 0, u) .
Assume that this linear system has unrestricted controls, that is, the con-
trol range is U = Rm . Let Δ ⊂ Rd be the subspace given by

Δ = span Ak bi : k = 0, ..., d − 1, i = 1, ..., m .

d−1
Then Δ = Im Ak B .
k=0
Definition 5.6.1 The linear control system LΣ satisfies the Kalman con-
dition if dim Δ = d, that is, Δ = Rd .
The Kalman condition is sufficient for controllability, as in the following:
Theorem 5.6.1 The linear control system LΣ is controllable if it satisfies

the Kalman condition.
Proof. Suppose that the system satisfies the Kalman condition. For t > 0
and x ∈ Rd , consider the affine map ϕ(t,x) : Rm → Rd given by the partial
map ϕ(t,x) (u) = ϕ (t, x, u). We claim that ϕ(t,x) is surjective, and the system
is then controllable. Indeed, since ϕ(t,x) (u) = etA x + ϕ(t,0) (u), it is enough to

t
show that the linear part ϕ(t,0) (u) = etA 0 e−sA Bu ds is surjective. Suppose,
$ %
by contradiction, that ϕ(t,0) is not surjective. Then Im ϕ(t,0) is a proper
subspace of Rd . Since
$ % - ∗ $ % .
Im ϕ(t,0) = Ker (λ) : λ ∈ Rd , Im ϕ(t,0) ⊂ Ker (λ)
$ %
there is a nontrivial linear functional λ : Rd → R with Im ϕ(t,0) ⊂ Ker (λ).
$
t %
Then λ etA 0 e−sA Bu ds = 0 for all u ∈ Rm . By derivating this expression
we have
t
d tA −sA
0 = λ e e Bu ds
dt 0
t
= λ AetA e−sA Bu ds + etA e−tA Bu
0
t
= λ AetA e−sA Bu ds + Bu .
0
By derivating successively we obtain

t
0 = λ Ak etA e−sA Bu ds + Ak−1 Bu
0

for all k ∈ N. The derivatives for t = 0 imply 0 = λ Ak Bu for all u ∈ Rm
and k ∈ N. Hence λ (Δ) = 0 with λ = 0, which contradicts Δ = Rd .
Note that the Kalman condition implies Rd = ϕ(t,x) (Rm ) for all t > 0
and x ∈ Rd . This is stronger than controllability and is in fact equivalent to
the Kalman condition. In order to formulate this stronger controllability, we
consider the natural extension ϕ(t,x) : Upc → Rd for all t ∈ R and x ∈ Rd .
Definition 5.6.2 The linear control system LΣ is called completely con-

trollable at time t > 0 if ϕ(t,x) (Upc ) = Rd for every x ∈ Rd ; the system is
$ strongly %accessible if for each x ∈ M there is some t > 0 such
said to be
that int ϕ(t,x) (Upc ) = ∅.
Note that a control system is strongly accessible whenever it is completely

controllable at some time t > 0.
Theorem 5.6.2 The linear control system LΣ satisfies the Kalman condition
if and only if it is completely controllable at some time t > 0.
Proof. Assume that ϕ(t,0) (Upc ) = Rd for some t > 0 and suppose, by
contradiction, that Δ = Rd . Then there is a nontrivial linear functional
λ : Rd → R with Δ ⊂ Ker (λ). Let p (r) = det (rId − A) be the characteristic
polynomial of A and write
p (r) = rd + αd−1 rd−1 + · · · + α1 r + α0

with αi ∈ R. By the Cayley-Hamilton theorem ([3]), we have
Ad + αd−1 Ad−1 + · · · + α1 A + α0 Id = 0

d−1
and we can then write Ad = β k Ak . It follows that
k=0

d−1
d−1
Ad+1 = AAd = A β k Ak = β k Ak+1
k=0 k=0

d−2
= β d−1 Ad + β k Ak+1
k=0

d−1
d−1
= β d−1 β k Ak + β k−1 Ak
k=0 k=1

d−1
d−1
= β d−1 β 0 Id + β d−1 β k Ak + β k−1 Ak
k=1 k=1

d−1
= γ 1k Ak .
k=0
n k
d−1
Following by induction we obtain Ad+n = γ k A for every n ∈ N. Thus
k k=0
Δ = span A bi : k ∈ N, i = 1, ..., m . Since Δ ⊂ Ker (λ), it follows that
∞ n
∞ n
t n k t n+k
tA k
λ e A Bu = λ A A Bu = λ A Bu = 0
n=0
n! n=0
n!
for all u ∈ Rm . Then, for a given u ∈ Upc , we have

t
tA −sA
λ (ϕ (t, 0, u)) = λ e e Bu (s) ds
0
t ∞

n
(−s)
= λ etA An Bu (s) ds
0 n=0 n!
∞ t
(−1) n
n
tA n
= λ e A B s u (s) ds
n=0
n! 0
⎛ ⎞
∞ n l tj
(−1) n
= λ ⎝etA A B sn uj ds⎠
n=0
n! j=1 tj−1
⎛ ⎞
∞ l
(−1) n
n
t n+1
− t n+1
= λ ⎝etA uj ⎠
j j−1
A B
n=0
n! j=1
n + 1
⎛ ⎞
∞
(−1)
n l

= λ ⎝etA An B tn+1
j − tn+1
j−1 uj
⎠
n=0
(n + 1)! j=1
⎛ ⎞
∞ n
(−1)
l

= λ ⎝etA tn+1
j − tn+1 n
j−1 A Buj
⎠
n=0
(n + 1)! j=1
∞
(−1) n+1 tA n
l n
= tj − tn+1
j−1 λ e A Buj
n=0 j=1
(n + 1)!
= 0.
Hence ϕ(t,0) (Upc ) ⊂ Ker (λ) with λ = 0, which contradicts ϕ(t,0) (Upc ) = Rd .
Thus the system satisfies the Kalman condition. The converse follows by The-
orem 5.6.1.
The Kalman condition provides a simple criterion for the controllability

of linear control systems.
Example 5.6.1 Consider the linear control system

−2 1 1
ẋ = x + u (t)
2 −1 0
on R2 with control range U = R. Since Δ = span {(1, 0) , (−2, 2)} = R2 , the

control system is controllable.
From now one, we consider a linear control system LΣ with compact and
convex control range U ⊂ Rm such that 0 ∈ int (U). For t ∈ R and x ∈ Rd ,
we take the continuous extension ϕ(t,x) : U → Rd and define the sets
Ot+ (x) = ϕ(t,x) (U ) , Ot− (x) = ϕ(−t,x) (U )
for t ≥ 0 and x ∈ Rd . These sets were defined in Problem 2.5(9) for general
control affine systems. Note that both Ot+ (x) and Ot− (x) are compact in Rd ,
Ot+ (x) = O≤t
+
(x) ∩ O≥t
+
(x) and Ot− (x) = O≤t
− −
(x) ∩ O≥t (x).
Proposition 5.6.1 Assume that the linear control system LΣ satisfies the
Kalman condition. Then

1. The origin 0 of Rd is a periodic point satisfying 0 ∈ int Ot+ (0) ∩
−
int Ot (0) for all t > 0.
2. The system is locally accessible.
Proof. Let t > 0. Since the linear map ϕ(t,0) : Rm → Rd is surjec-

tive, ϕ(t,0) is an open map, by the open mapping theorem. Hence 0 =
$ %
ϕ(t,0) (0) ∈ int ϕ(t,0) (U) , and then 0 ∈ int Ot+ (0) . In particular, the

system is strongly accessible from 0. Similarly, we show that 0 ∈ int Ot− (0) ,
proving the item (1). Now, for any x ∈ Rd , we have Ot+ (x) = etA x + Ot+ (0),

and then int Ot+ (x) = ∅. This means that the system is strongly accessible

from any x ∈ Rd at any time t > 0.$ Analogously,
% have int %Ot− (x) = ∅ for
we $
−
all x ∈ Rd and t > 0. Hence int O≤t
+
(x) and int O≤t (x) are nonempty
for all x ∈ Rd and t > 0, and thus the system is locally accessible according
to Definition 5.4.2.
Therefore, under the Kalman condition, the effective control sets and the
main control sets have the same meaning (see Theorem 5.4.2). Moreover, we
+ −
have 0 ∈ int O>0 (0) ∩int O>0 (0) , which means that the origin is contained
in some transitivity set of an effective control set. By Corollary 2.2.1, both
semi-orbits O+ (0) and O− (0) are open sets in Rd , hence the transitivity set
containing 0 is the open holding set H0 = O+ (0) ∩ O− (0). The following
theorem shows that H0 is the unique open holding set of the system.
Theorem 5.6.3 Assume that the linear control system LΣ, with compact
and convex control range U ⊂ Rm such that 0 ∈ int (U), satisfies the Kalman
condition. There then exists a unique effective control set, D = cl (O+ (0)) ∩
O− (0). Furthermore, the following statements hold:
1. D is an invariant control set if the origin 0 is a sink for A, that is,

Re λ < 0 for all eigenvalues λ of A. In this case D = cl (O+ (0)).
2. D is open if the origin 0 is a source for A, that is, Re λ > 0 for all
eigenvalues of A. In this case D = O− (0).
Proof. Let D ⊂ Rd be the effective control set with transitivity set

D0 = H0 . Since the system is locally accessible, it follows from Theorem
5.4.2 that int (D) = D0 and D = cl (O+ (0)) ∩ O− (0). Suppose that D ⊂ Rd
is an effective control set of the system. By Theorem 5.4.2 again, D is an
effective control set with int (D ) = D0 . By taking x ∈ D0 , we can find
0 < a < 1 such that ax ∈ D0 . As D0 = O+ (x) ∩ O− (x), there are t, s > 0
and u1 , v1 ∈ U such that ϕ (t, x, u1 ) = ax and ϕ (−s, x, v1 ) = ax. Since the
control range U is convex, 0 ∈ U, and 0 < a < 1, we have au1 ∈ U. From the
equality ϕ (t, x, u1 ) = ax we obtain by linearity
ϕ (t, ϕ (t, x, u1 ) , au1 ) = ϕ (t, ax, au1 ) = aϕ (t, x, u1 ) = a2 x.
By Proposition 1.3.2, there is u2 ∈ U such that ϕ (t, ϕ (t, x, u1 ) , au1 ) =

ϕ (2t, x, u2 ), hence ϕ (2t, x, u2 ) = a2 x. Following by induction we obtain a se-
quence of control functions (un ) in U satisfying ϕ (nt, x, un ) = an x for all n ∈
N. Similarly, we can obtain a sequence (vn ) in U satisfying ϕ (−ns, x, vn ) =
an x for all n ∈ N. Now, as an → 0, there is some n0 such that an0 x ∈ D0 , since
D0 is a neighborhood of the origin 0 ∈ Rd . As an0 x = ϕ (n0 t, x, un0 ) ∈ O>0
+
(x)
−
and a x = ϕ (−n0 s, x, vn0 ) ∈ O>0 (x), it follows that
n0
−
an0 x ∈ D0 ∩ O>0
+
(x) ∩ O>0 (x) = D0 ∩ D0 .
This means that D = D, and therefore D is the unique effective control set of
the system. We now show statements (1) and (2). For statement (1) suppose
that the origin 0 ∈ Rd is a sink for the linear system ẋ = Ax. For a given
−
x ∈ Rd there is t > 0 with etA x ∈ D0 . Hence ϕ (t, x, 0) = etA x ∈ O>0 (0)
− −
which implies x ∈ O>0 (0). Thus O>0 (0) = R and D = cl (O (0)). It
d +
follows that D is positively invariant and closed, which means that D is an

invariant control set. For statement (2) suppose that 0 ∈ Rd is a source for
ẋ = Ax. For a given x ∈ Rd there is t > 0 with e−tA x ∈ O>0+
(0), and then
−
x ∈ O>0 (0). Hence R = O>0 (0) and D = O (0). Therefore D is open.
+ d +
Now, consider the linear part ẋ = Ax of the control system LΣ. By a

suitable change of coordinates x = P y we transform A into a real canonical
form C = P AP −1 . The solutions of ẋ = Ax are of the form x (t) = P y (t)
where y (t) is a solution to y = Cy. This equation y = Cy breaks up into
a set of uncoupled equations of the form z = Cλ z where Cλ is one of the
Jordan λ-blocks in the real canonical form C of A. If λ ∈ R and Cλ is a r × r
matrix, the solution to z = Cλ z with initial value z (0) = v is
⎛ ⎞⎛ ⎞
1 v
⎜ t 1 ⎟ ⎜ 1⎟
⎜ ⎟⎜ ⎟
⎜ .. .. ⎟⎜ ⎟
tλ ⎜ ⎟ ⎜ .. ⎟
2
t . .
z (t) = e ⎜ ⎟⎜ . ⎟.
⎜ 2!
.. ⎟⎜ ⎟
⎜ .. .. .. ⎟⎝ ⎠
⎝ . . . . ⎠
tr−1 t2 vr
(r−1)! ··· 2! t 1
In coordinates one has

i−1 k
t
zi (t) = eλt vi−k .
k!
k=0
In the case λ = a + bi with b = 0, each coordinate zi (t) of z (t) is two-

dimensional of the form zi (t) = zi1 (t) , zi2 (t) with
t 1
i−1 k

zi1 (t) = eat vi−k cos bt − vi−k
2
sin bt ,
k!
k=0
t 2
i−1 k

zi2 (t) = eat v 1
cos bt + vi−k sin bt ,
k! i−k
k=0
where v = v 1 + v 2 i. We refer to [60, Chapter 6] for details.
Theorem 5.6.4 Assume that the linear control system LΣ, with compact
and convex control range U ⊂ Rm such that 0 ∈ int (U), satisfies the Kalman
condition. The system is then controllable if and only if Re λ = 0 for all
eigenvalues λ of A.
Proof. We write the system in the coordinate system corresponding to the

Jordan canonical form C of A. Suppose that one of the eigenvalues λ = a + bi
of A has real part a > 0. We firstly consider b = 0 and take the Jordan
λ-block ⎛ ⎞
a
⎜1 a ⎟
⎜ ⎟
Cλ = ⎜ .. .. ⎟.
⎝ . . ⎠
1 a
In the generalized eigenspace for λ, the first component solution xk (t) satisfies
the equation
d m
xk (t) = axk (t) + ui (t) bik
dt i=1
where bik is the corresponding component of B. Hence xk (t) = ceat +

t
m
eat 0 e−as ui (t) bik ds. Since U is compact, we can choose a number K > 0
i=1
m
such that ui bik < K for all u = (u1 , .., um ) ∈ U. For xk (0) = c > K/a,

i=1
it follows that
t
d
xr (t) > Keat + Keat −ae−as ds − K = 0
dt 0

and then the subset Xk = x ∈ Rd : xk > K/a is positively invariant. Hence,
for x ∈ Xk , we have O+ (x) ⊂ Xk = Rd , and therefore the system is not
controllable. In the case a < 0, we conclude the same result by considering
the initial value xk (0) > K/ |a|. We now consider the case λ = a + bi with
b = 0 and take the Jordan λ-block
⎛ ⎞
a −b
⎜b a ⎟
⎜ ⎟
⎜1 a −b ⎟
⎜ ⎟
⎜ ⎟
Cλ = ⎜ ⎜ 1 b a ⎟.
⎟
⎜ .. .. ⎟
⎜ . . ⎟
⎜ ⎟
⎝ 1 a −b⎠
1 b a
In the generalized eigenspace for λ, the first two-dimensional component so-

lution xk (t) satisfies the equation

m
d a −b
xk (t) = xk (t) + ui (t) bik
dt b a
i=1
where bik is the corresponding two-dimensional component of B. We then

have
/ 0
d 2 a −b m
xk (t) = 2 xk (t) + ui (t) bik , xk (t)
dt b a
i=1
1 /m 02
2

= 2 a xk (t) + ui (t) bik , xk (t) .
i=1
3 m 4 m
m

As
ui bik , xk (t) ≤ ui bik xk (t) with ui bik

< K for all
i=1 i=1 i=1
u = (u1 , .., um ) ∈ U and some constant K > 0, it follows that
d - .
2 2
xk (t) > 2 a xk (t) − K xk (t) .
dt
2
By considering a > 0 and xk (0) = x0 > K/a we have dt d
xk (t) > 0,

and then the subset Xk = x ∈ R : xk > K/a is positively invariant.
d
Thus the system is not controllable. The case a < 0 is analogous. Therefore
Re λ = 0 for all eigenvalues λ of A if the system is controllable. As to the
converse, suppose that all eigenvalues of A have zero real part. Consider the
eigenspace E 0 corresponding to the real eigenvalue 0 ∈ R. For x ∈ E 0 we

have ϕ (t, x, u) = x + ϕ (t, 0, u) for all t ∈ R and u ∈ U. As 0 ∈ int Ot+ (0)

for all t > 0, we have x ∈ int x + Ot+ (0) ⊂ int Ot+ (x) for all t > 0.
−
Analogously, x ∈ int Ot (x) for all t > 0, and hence x is contained in the
transitivity set D0 of some effective control set D. By Theorem 5.6.3, D is
the unique effective control set, and therefore E 0 ⊂ D0 . Now, let bi be a
complex eigenvalue of A and consider the Jordan bi-block Cbi . The solutions
for z = Cbi z are of the form
⎛ ⎞
v11 cos bt − v12 sin bt
⎜ v12 cos bt + v11 sin bt ⎟
⎜ ⎟
⎜v 1 cos bt − v 2 sin bt + t v 1 cos bt − v 2 sin bt⎟
⎜ 2 2 1 1 ⎟
⎜ 2 ⎟
⎜v2 cos bt + v21 sin bt + t v12 cos bt + v11 sin bt ⎟
⎜ ⎟
z (t) = ⎜
⎜
..
.
⎟
⎟
⎜ ⎟
⎜ ti 1
k−1 ⎟
⎜ ⎟
i! vk−i cos bt − vk−i sin bt
2
⎜ ⎟
⎜ i=0 ⎟
⎝ ti 2
k−1 ⎠
1
i! vk−i cos bt + vk−i sin bt
i=0

with initial value z (0) = v11 , v12 , v21 , v22 , ..., vk1 , vk2 . If v = (0, ..., 0, 1, 0, ..., 0)
is an eigenvector of bi then the solution with initial value z (0) = v satisfies

2kπ 2kπ
z = v for all k ∈ Z. It follows that ϕ , v, 0 = v for all k ∈ Z.
b b 2π
For a given y ∈ O+ 2π (0) there is u ∈ U such that ϕ 2 , 0, u = y, which
b
2 , v, u = v + y. Hence v + O 2π (0) ⊂ O+
+
implies ϕ 2π 2π (v) and therefore
$ % $ % b b $ %
v ∈ int v + O+ 2π (0) ⊂ int O+ 2π (v) . Similarly, v ∈ int v + O−
2π (0) ⊂
$ % b +b b
− −
int O 2π (v) . Thus v ∈ int O>0 (v) ∩ O>0 (v) and then v ∈ D0 . It follows
b
that all eigenvectors lie in D0 , which means that D0 = Rd .
5.7 Prolongational transitive sets

We now introduce the notion of a nonwandering point. It is another concept
connected with recursiveness. We give some characterizations of nonwander-
ing points in terms of prolongational limit sets.
Definition 5.7.1 A point x ∈ M is said to be nonwandering if x ∈ J + (x)

(or equivalently x ∈ J − (x)).
Note that x ∈ J + (x) if and only if x ∈ J − (x) (Proposition 4.2.7). Since

ω (x) ⊂ J + (x) and ω − (x) ⊂ J − (x), any positively or negatively recurrent
+
point is nonwandering. However, a nonwandering point need not be positively

or negatively recurrent. In Example 2.4.3, the state x = 0 is a nonwandering
point of the system, but x ∈ / ω + (x) ∪ ω − (x). In the same example, 0 ∈
ω + (x) = {0} but x ∈ / ω − (0) = {0}. This fact shows that the property of
the prolongational limit sets as presented in Proposition 4.2.7 is generally not
possessed by the limit sets.
The following result shows that the points which reside in the limit sets
are nonwandering.
Proposition 5.7.1 Let x ∈ M and assume that ω + (x) is nonempty. Every

y ∈ ω + (x) is nonwandering. The limit set ω − (x) has the same property.
Proof. If y ∈ ω + (x) then there are sequences tn → +∞ and (un ) in U

such that ϕ (tn , x, un ) → y. Take a subsequence (tnk ) such that tnk+1 −tnk ≥ k
for each k. Define τ k = tnk+1 − tnk and xk = ϕ (tnk , x, unk ). Then τ k → +∞
and xk → y . For each k, define a function vk ∈ U by

unk · tnk (t) , if t ≤ −tnk
vk (t) = .
unk+1 (t + tnk ) , if t > −tnk
Section 5.7 · Prolongational transitive sets 147
Note that
unk (t) , if t ≤ 0
vk · (−tnk ) (t) = .
unk+1 (t) , if t > 0
Then we have
ϕ (−tnk , ϕ (tnk , x, unk ) , vk ) = ϕ (−tnk , ϕ (tnk , x, unk ) , unk · tnk ) = x,

ϕ tnk+1 , x, vk · (−tnk ) = ϕ tnk+1 , x, unk+1 ,
and hence
ϕ (τ k , xk , vk )

= ϕ tnk+1 − tnk , ϕ (tnk , x, unk ) , vk

= ϕ tnk+1 , ϕ (−tnk , ϕ (tnk , x, unk ) , vk ) , vk · (−tnk )

= ϕ tnk+1 , x, unk+1 .
This means that ϕ (τ k , xk , vk ) → y, and therefore, y ∈ J + (y). The proof for

ω − (x) is analogous.
The following example illustrates a situation in which the equality ω + (x) =

+
J (x) holds and all nonwandering points are then recurrent points.
Figure
$√ % $5.6: % Examples
$√ % of trajectories through the points
√
2 π
2 , 4 ,
2 3π
2 , 4 , 2 , 2 , u ∈ 1, 12 , 20
2 3π 1
.
ẋ (t) = X0 (x (t)) + u (t) X1 (x (t))

5 6
on the unit disk M = {x = (x1 , x2 ) : x ≤ 1}, where U = 1
20 , 1 and X0 , X1
are vector fields given by

X0 (x, y) = (y, −x) and X1 (x, y) = x − xy 2 − x3 , y − yx2 − y 3 .
The phase portrait is as shown in Fig. 5.6. The origin 0 is a singular point
and the trajectories through the points in the unit sphere S1 are periodic and
coincide with S1 . Hence, ω + (0) = ω − (0) = {0} and ω + (x) = ω − (x) = S1 for
every x ∈ S1 . Thus the points in S1 ∪ {0} are recurrent. For x ∈ M \ S1 ∪ {0},
we have ω + (x) = S1 and ω − (x) = {0}. Thus the points $in M \S1 ∪{0} % are not
recurrent. Now, given ε > 0 small enough, we have cl O≥t (B (0, ε)) = M
+
$ %
−
for every t > 0, hence J + (0) = M . For x ∈ S1 , we have cl O≥t (B (x, ε)) =
−
M for every
$ % J (x) = M . Nevertheless,
t > 0, hence, $ for%x ∈ M \S1 ∪{0}, we
−
have cl O≥t (B (x, ε)) ⊂ M \{0} and cl O≥t (B (x, ε)) ⊂ M \S1 . Only the
+
sets S1 and {0} are positively (and negatively) invariant, hence J + (x) = S1
and J − (x) = {0}. Therefore the points in S1 ∪ {0} are the nonwandering
points of the system, and for x ∈ M \ S1 ∪ {0}, we have ω + (x) = J + (x) = S1
and ω − (x) = J − (x) = {0}.
We now obtain a connection between nonwandering points and recurrent

points.
Proposition 5.7.2 If R ⊂ M is a set of positively recurrent points then any

point in cl (R) is nonwandering.
Proof. Take a point x ∈ cl (R) and consider a sequence (xn ) in R such

that xn → x. For a given ε > 0, we can take some xn ∈ B (x, ε). Since
xn ∈ ω + (xn ), we have B (x, ε) ∩ O≥t+
(xn ) = ∅ for all t > 0. It follows that
∅= B (x, ε) ∩ O≥t (xn ) ⊂ B (x, ε) ∩ O≥t
+ +
(B (x, ε)). Thus x ∈ J + (x).
Analogously, if Q ⊂ M is a set of negatively recurrent points then any

point in cl (Q) is nonwandering.
If M admits an invariant probability measure μ on the Borel sets of M ,
the Poincaré recurrence theorem assures that μ-almost every point in M is
positively recurrent (Theorem 5.3.1). In particular, if μ is a Radon measure
such that μ (U ) > 0 for any nonempty open set U ⊂ M then R+ is dense
in M , that is, the Birkhoff center coincides with M . The following theorem
presents a necessary and sufficient condition for R+ to be dense in M .
Theorem 5.7.1 Assume that M is a complete metric space. The following

statements are equivalent:
1. Every point in M is nonwandering.
2. The set R+ of all positively recurrent points is dense in M .
3. The set R− of all negatively recurrent points is dense in M .
4. The set R of all recurrent points is dense in M.
Proof. (1) ⇔ (2) Suppose that R+ is dense in M . According to Propo-

sition 5.7.2, every point in M = cl (R) is nonwandering. As to the converse,
suppose that every point in M is nonwandering. Let x ∈ M and ε > 0. Define
− −
U0 = B (x, ε). Since U0 ∩ O≥1 (U0 ) = ∅, there are x1 ∈ U0 ∩ O≥1 (U0 ) and
− −
0 < ε1 < 2 such that U1 = B (x1 , ε1 ) ⊂ U0 ∩ O≥1 (U0 ). As U1 ∩ O≥2 (U1 ) = ∅,
1
−
there are x2 ∈ U1 ∩ O≥2 (U1 ) and 0 < ε2 < 212 such that U2 = B (x2 , ε2 ) ⊂
−
U1 ∩ O≥2 (U1 ). By following this process, we construct a sequence of open
−
sets {Un } such that Un+1 ⊂ Un ∩ O≥n+1 (U ) and diam (Un ) → 0 as n → ∞.
∞ n
Since M is complete, it follows that n=0 Un = {y}, by the Cantor intersec-
tion theorem. As y ∈ Un for all n, there are sn ≥ n, un ∈ U , and yn ∈ Un−1
such that ϕ (sn , y, un ) = yn . Hence ϕ (sn , y, un ) → y with sn → +∞, and
therefore y is a positively recurrent point in B (x, ε). This proves that R+ is
dense in M .
(1) ⇔ (3) It follows similarly to the proof for (1) ⇔ (2).
(1) ⇔ (4) We have (4) ⇒ (2) ⇒ (1). It then remains to show that
(1) ⇒ (4). In fact, suppose that every point in M is nonwandering, take
x ∈ M and ε > 0. Define U0 = B (x, ε). Since U0 ∩ O≥1 +
(U0 ) = ∅ and
O≥1 (U0 ) is open, there are x1 ∈ U0 ∩ O≥1 (U0 ) and 0 < ε1 < 12 such
+ +
−
that U1 = B (x1 , ε1 ) ⊂ U0 ∩ O≥1 +
(U0 ). As U1 ∩ O≥2 (U1 ) = ∅, there are
− −
x2 ∈ U1 ∩O≥2 (U1 ) and 0 < ε2 < 22 such that U2 = B (x2 , ε2 ) ⊂ U1 ∩O≥2
1
(U1 ).
Since U2 ∩ O≥3 (U2 ) = ∅, there are x3 ∈ U2 ∩ O≥3 (U2 ) and 0 < ε3 < 213 such
+ +
−
that U3 = B (x3 , ε3 ) ⊂ U2 ∩ O≥3 +
(U2 ). As U3 ∩ O≥4 (U3 ) = ∅, there are
− −
x4 ∈ U3 ∩ O≥4 U3 and 0 < ε4 < 24 such that U4 = B (x4 , ε4 ) ⊂ U3 ∩ O≥1
1
(U3 ).
By following this process, we obtain a sequence of open sets {Un } such that
−
U2n+1 ⊂ U2n ∩ O≥2n+1+
(U2n ) and U2n ⊂ U2n−1 ∩ O≥2n (U ), for all n > 0,
∞2n−1
and diam (Un ) → 0 as n → ∞. It follows that n=0 Un = {y}. For
2n + 1, there are s2n+1 ≥ 2n + 1, u2n+1 ∈ U, and y2n+1 ∈ U2n such that
ϕ (s2n+1 , y2n+1 , u2n+1 ) = y. Then ϕ (−s2n+1 , y, u2n+1 · s2n+1 ) = y2n+1 → y
with −s2n+1 → −∞, which means y ∈ ω − (y). For 2n, there are s2n ≥ 2n,
u2n ∈ U, and y2n ∈ U2n−1 such that ϕ (s2n , y, u2n ) = y2n → y, with s2n →
+∞, hence y ∈ ω + (x). Thus y is a recurrent point B (x, ε), which proves that
R is dense in M .
In Example 5.3.3, the Birkhoff center of the control system on O (n, R)

covers all the state space, and therefore every state is nonwandering.
We often denote by J the set of all nonwandering points of the control
system. We shall define an equivalence relation in J . We start with the
following:
Definition 5.7.2 We say that two nonwandering points x, y ∈ J are equiv-

alent if x ∈ J + (y) and y ∈ J + (x).
This relation among the nonwandering points is called prolongational

transitivity relation. A set of equivalent nonwandering points is called a
prolongational transitive set. Then E ⊂ M is a prolongational transitive
set if and only if E ⊂ J + (x) for every x ∈ E. It is easily seen that an
asymptotic transitive set is contained in some prolongational transitive set.
It should be observed that the prolongational transitivity relation may not
be an equivalence relation, as in the following example.
Example 5.7.2 Consider the control affine system on the unit circle S1 as
defined in Example 3.1.1. We have J + (0) = S1 , J + (x) = [π/2, π] for all
x ∈ (0, π), J + (π) = [π/2, 3π/2], and J + (x) = [π, 3π/2] for all x ∈ (π, 2π).
Then 0 and all points in [π/2, 3π/2] are nowandering, however, π/2 ∈ J + (π),
π ∈ J + (3π/2), but π/2 ∈ / J + (3π/2). Hence the prolongational transitivity
relation is not an equivalence relation in J .
The prolongational transitivity relation is an equivalence relation if the

control system is equicontinuous on J , as follows.
Proposition 5.7.3 Assume that the control system is equicontinuous at ev-

ery nonwandering point. The prolongational transitivity relation in J is then
an equivalence relation.
Proof. Let x, y, z ∈ J . By Lemma 4.3.1, if x ∈ D+ (y, t) and y ∈ D+ (z, s)

then x ∈ D+ (z, t + s). Hence x ∈ J + (y) and y ∈ J + (z) implies x ∈ J + (z).
Thus the prolongational transitivity relation is transitive. Since this relation
is clearly reflexive and symmetric, it is an equivalence relation.
From now on, in this section, we assume that the prolongational transi-
tivity relation is an equivalence relation among the nonwandering points. We
shall describe its equivalence classes.
Definition 5.7.3 For t ≥ 0, we say that a nonempty set E ⊂ M is a weak

t-prolongation control set if
1. E ⊂ D+ (x, t) for all x ∈ E,
2. E is maximal satisfying the property (1).
Contrast the condition (1) of a weak 0-prolongation control set with the
property of an equistable set stated in Proposition 4.3.1.
Definition 5.7.4 A nonempty set E ⊂ M is said to be a weak prolonga-

tional control set if
1. E ⊂ J + (x) for all x ∈ E,
2. E is maximal satisfying the property (1).
For each x ∈ M , we set Ex = J + (x)∩J − (x). Note that x is nonwandering

if and only if x ∈ Ex . We then expect that the sets Ex , with x ∈ J , are the
weak prolongational control sets of the system and coincide with the equiva-
lence classes of the prolongational transitivity relation. We also expect that
every weak prolongational control set is an intersection of weak t-prolongation
control sets.
Example 5.7.3 Return to Example 5.5.1 and consider the set

D = x ∈ R2 : x ≥ 1 .
It is not difficult to see that D = J + (x) for every x ∈ D. Thus D is a weak

prolongational control set. If x, y ∈ D then x ∈ J + (y), hence y ∈ J − (x).
This means that D ⊂ J − (x) for every x ∈ D. Therefore D = J + (x) ∩ J − (x)
for every x ∈ D.
The next proposition shows that a set satisfying the item (1) of Definition
5.7.4 can be extended to a weak prolongational control set. A proof for it is
similar to the proof for Proposition 5.4.4, and then it is omitted (see Problem
5.8(62)).
Proposition 5.7.4 Assume that E ⊂ M is a nonempty subset satisfying

E ⊂ J + (x) for all x ∈ E. There is then a weak prolongational control set
containing E.
This means that every nonwandering point is contained in some weak

prolongational control set. In the following, we characterize the equivalence
classes of the prolongational transitivity relation.
Proposition 5.7.5 Let E ⊂ M be a nonempty set. The following statements

are equivalent:
1. E is an equivalence class of the prolongational transitivity relation in

J.
2. E is a weak prolongational control set.
3. E = Ex = J + (x) ∩ J − (x) for every x ∈ E.
Proof. (1) ⇒ (3) If x, y ∈ E then x ∈ J + (y) and y ∈ J + (x), which means

y ∈ J − (x) and y ∈ J + (x), by Proposition 4.2.7. Hence E ⊂ J + (x) ∩ J − (x)
for every x ∈ E. On the other hand, if z ∈ J + (x) ∩ J − (x), with x ∈ E, then
z is nonwandering, by the transitive property, and z ∈ J + (x) and x ∈ J + (z).
Hence x and z are equivalent, and therefore z ∈ E. This proves that E =
J + (x) ∩ J − (x) for every x ∈ E.
(3) ⇒ (2) It is clear that E ⊂ J + (x) for every x ∈ E. By Proposition 5.7.4,
there is a weak prolongational control set E containing E. For x ∈ E and
x ∈ E, it follows that x ∈ J + (x) and x ∈ J + (x ), which means x ∈ J + (x)
and x ∈ J − (x). Hence E ⊂ J + (x) ∩ J − (x) = E, and therefore E = E is a
weak prolongational control set.
(2) ⇒ (1) If x, y ∈ E then x ∈ E ⊂ J + (y) and y ∈ E ⊂ J + (x), hence
x and y are equivalent. This implies that E is contained in an equivalence
class E of the prolongational transitivity relation. By the first part of the
proof, E ⊂ J + (x ) for every x ∈ E . By the maximality of E, it follows that
E = E.
The next result describes a weak prolongational control set as an intersec-

tion of weak t-prolongation control sets.
Proposition 5.7.6 If E ⊂ M is a weak prolongational control set then for

each t ≥ 0 there is a weak t-prolongation control set Et such that

E= Et .
t≥0
The converse holds if distinct weak t-prolongation control sets are disjoint.
Proof. Let E be a weak prolongational control set. For each t ≥ 0, we

have E ⊂ J + (x) ⊂ D+ (x, t) for all t ≥ 0. Then there is a weak t-prolongation

control set Et such that E ⊂ Et . Hence, E ⊂ t≥0 Et . On the other hand,

for any x, y ∈ t≥0 Et , we have y ∈ t≥0 Et ⊂ t≥0 D+ (x, t) = J + (x).

Hence the points of t≥0 Et lie in the same equivalence class of the prolon-

gational transitivity relation, and therefore E = t≥0 Et . As to the converse,
suppose that the weak t-prolongation control sets are pairwise disjoint. Let

E = t≥0 Et , where each Et is a weak t-prolongation control set. For x ∈ E,
we have E ⊂ J + (x). Hence there is a weak prolongational control set E such

that E ⊂ E . By the first part of this proof, it follows that E = t≥0 Et ,
where each Et is a weak t-prolongation control set. Since distinct weak t-
prolongation control sets are disjoint, we have Et = Et , for all t ≥ 0. There-
fore E = E .
Note that the weak t-prolongation control sets are pairwise disjoint if the
control system is equicontinuous, by Lemma 4.3.1.
We now define a dynamic order among the weak prolongational control
sets. We need the following definition of a domain of weak uniform attraction:
Definition 5.7.5 Let E ⊂ M be a weak prolongational control set. The

domain of weak uniform attraction of E is defined by

Awu (E) = x ∈ M : J + (x) ∩ E = ∅ .
Any point x in Awu (E) is said to be weakly uniformly attracted to E.
A dynamic order among the weak prolongational control sets is defined as

follows: for two weak prolongational control sets E and E
let E wu E if and only if E ∩ Awu (E ) = ∅.
Proposition 5.7.7 The relation “wu ” among the weak prolongational con-
trol sets is a preoder.
Proof. If E is a weak prolongational control set then E ⊂ J + (x) for all

x ∈ E. Hence E ⊂ Awu (E) and therefore E wu E. If E and E are two
weak prolongational control sets such that E wu E and E wu E, then
there are x ∈ E ∩ Awu (E ) and y ∈ E ∩ Awu (E), that is, J + (x) ∩ E = ∅
and J + (y) ∩ E = ∅. Take u ∈ J + (x) ∩ E and v ∈ J + (y) ∩ E. As x ∈ E and
y ∈ E , we have x ∈ J + (y) and y ∈ J + (x). Hence x and y are equivalent,
and therefore E = E . Finally, suppose that E wu E and E wu E .
Then E ∩ Awu (E ) = ∅ and E ∩ Awu (E ) = ∅. Take x ∈ E ∩ Awu (E ) and
y ∈ E ∩ Awu (E ). There are then u ∈ J + (x) ∩ E and v ∈ J + (y) ∩ E .
As y ∈ E , it follows that y ∈ J + (x) and therefore v ∈ J + (x). Thus
v ∈ J + (x) ∩ E , which means E wu E .
The preorder “wu ” is called the prolongational transitivity pre-

order. In order to describe the upper bounds for the prolongational tran-
sitivity preorder, it should be observed that we may omit the equicontinuity
in Propositions 4.3.3 and 4.3.4, since we are assuming that the prolongational
transitivity relation is an equivalence relation. This is very simple to check
and is left to the reader. So we can prove the following:
Theorem 5.7.2 If K ⊂ M is a compact minimal equistable set then K is a
weak prolongational control set that is maximal with respect to the prolonga-
tional transitivity preorder. The converse holds if the manifold M is compact.
Proof. Suppose that K is a compact minimal equistable set. By Propo-
sition 4.3.4, it follows that J + (x) = K for all x ∈ K. Hence there is a weak
prolongational control set E such that K ⊂ E. For x ∈ K and y ∈ E, we
have y ∈ E ⊂ J + (x) = K. Thus K = E is a weak prolongational control set.
Suppose that E is a weak prolongational control set with K wu E . There
is then x ∈ K ∩ Awu (E), that is, J + (x) ∩ E = ∅. As J + (x) = K, it follows
that K = E . Thus K is maximal with respect to the prolongational transi-
tivity preorder. Now, suppose that M is a compact manifold and K is a weak
prolongational control set that is maximal with respect to the prolongational
transitivity preorder. For x ∈ K, J + (x) is compact and equistable according
to Proposition 4.3.3. According to Proposition 4.3.5, there is a minimal eq-
uistable set K contained in J + (x). By the first part of the proof, K is a
weak prolongational control set. Since x ∈ K ∩ Awu (K ), we have K wu K .
From the maximality of K, it follows that K = K . Therefore K is a minimal
equistable set.
5.8 Problems
1. Let x, y ∈ M be two periodic points. Show that x and y are equivalent
− −
if and only if O>0
+
(x) = O>0
+
(y) and O>0 (x) = O>0 (y).
2. Give examples of control systems with P = ∅.
3. Consider the control system on the plane

R2 determined
! by the family
∂ ∂
of partial derivatives operators F = ,± . Describe the holding
∂x ∂y
sets and the dynamic preorder among them.
4. Find other holding sets for the control system in Example 5.1.2.
5. Consider the
7 control system on R given by ẋ = sin x + u (t), with control
√ √ 8
range U = − 22 , 22 . Prove that
7 π π8
Hx = ∅, if x ∈/ kπ − , kπ + ,
4 4
π π
Hx = {x} , if x = kπ − or x = kπ + ,
$ 4 4
π π% $ π π%
Hx = kπ − , kπ + , if x ∈ kπ − , kπ + ,
4 4 4 4
where k ∈ Z.
$x%
6. Consider the control system on the unit circle S1 given by ẋ = sin2 +
$ % 2
2 x
u (t) cos , with control range U = [−1, 1]. Show that Hx = S for 1
2
every x ∈ S .
1
7. Consider the control affine system on the torus T2 defined as in

Example 5.2.1 but with b/a a rational number. Show that any point of
T2 is periodic.
8. Show that the control systems in Examples 1.3.1 and 1.3.2 have global
holding sets, that is, Hx = M for all x ∈ M . Conclude that they are
minimal control systems.
9. In general topological dynamics, a semigroup S acts on the left on a

topological space X if s (tx) = (st) x for all x ∈ X and s, t ∈ S. A
nonempty subset D ⊂ X is a weak control set of S if D ⊂ cl (Sx) for all
x ∈ D and D is maximal satisfying this property. A control set of S is
a weak control set with nonempty interior. For a given control set D,
we define the transitivity set

D0 = x ∈ D : x ∈ int (Sx) ∩ int S −1 x .
If D0 = ∅ then D is called an effective control set of S. The domain of

attraction of a weak control set D is defined by
A (D) = {x ∈ M : cl (Sx) ∩ D = ∅} .
Recall the two-sided ideals S≥T ⊂ S defined in Section 2.2. Show that
a nonempty set R ⊂ M is a maximal asymptotic transitive set if and
only if for each T > 0 there is a weak control set DT of S≥T such that

R= DT . In this case, prove that A (R) = A (DT ).
T >0 T >0
10. Let M = R2 and consider

the control system determined
! by the family
∂ ∂ ∂
of vector fields F = , f (x1 ) ,± where f (x1 ) < 0 for
∂x1 ∂x1 ∂x2
x1 > 0 and f (x1 ) = 0 for x1 ≤ 0. Prove the following statements:
∂
(a) The trajectories of the vector field f (x1 ) in the set
∂x1
{(x1 , x2 ) : x1 > 0} are horizontal lines converging to the vertical
axis for t → +∞.
(b) The points in the left half plane {(x1 , x2 ) : x1 ≤ 0} are
∂
equilibrium points for f (x1 ) .
∂x1
(c) The vertical lines in {(x1 , x2 ) : x1 ≤ 0} are weak control sets with
empty interior.
(d) The right half plane D = {(x1 , x2 ) : x1 ≥ 0} is a closed effective
control set with transitivity set D0 = {(x1 , x2 ) : x1 > 0}.
11. Consider the control affine system on M = R2 determined by the vector

fields X, Y given by

X (x1 , x2 ) = x2 + x1 1 − x21 − x22 , −x1 + x2 1 − x21 − x22 ,
Y (x1 , x2 ) = (x1 , x2 ) .
Discuss the control sets of the system.


12. Consider the control system on the unit disk M = x ∈ R2 : x ≤ 1
determined by the vector fields X = (X1 , X2 ) and Y = (Y1 , Y2 ) given
by
2 π 2 π
X1 (x) = −x2 + x1 x sin , X2 (x) = x1 + x2 x sin
x x
Y1 (x) = α (x) x1 + (α (x) − β (x)) x2

2 π
+x1 x β (x) sin − 4α (x) ,
x
Y2 (x) = α (x) x2 + (β (x) − α (x)) x1

2 π
+x2 x β (x) sin − 4α (x)
x
and X (0) = Y (0) = 0, where

1 if x ≤ 1/2 0 if x ≤ 1/2
α (x) = , β (x) = .
0 if 1/2 < x ≤ 1 1 if 1/2 < x ≤ 1
Discuss the control sets of this system.
13. A nonempty subset X ⊂ M is said to be positively recursive with

respect to the nonempty subset Y ⊂ M if for every t > 0, O≥t +
(Y )∩X =
∅; the set X is said to be negatively recursive with respect to the set
−
Y if for every t > 0, O≥t (Y ) ∩ X = ∅. A point x ∈ M is said to be
positively Poisson stable if every neighborhood V of x is positively
recursive with respect to the single set {x}; x is said to be negatively
Poisson stable if every neighborhood V of x is negatively recursive
with respect to the set {x}; x is said to be Poisson stable if it is
positively and negatively Poisson stable. Prove that:
(a) x is positively Poisson stable if and only if x is positively recurrent.

(b) x is negatively Poisson stable if and only if x is negatively recurrent.
(c) x is Poisson stable if and only if x is recurrent.
14. Let x ∈ R+ . Show that Rx = {y ∈ ω + (x) : ω + (y) = ω + (x)}.

+
15. Define the set R+ 1 = x ∈ M : x ∈ cl O>1 (x) endowed with the fol-
lowing relation: two point x, y ∈ R+ are equivalent if there are numbers
+ 1 +
t1 , t2 > 0 such that x ∈ cl O>t1 (y) and y ∈ cl O>t2 (x) .
(a) Show that this is an equivalence relation in R+

1.
1 = R and the equivalence classes in R1 coincide

(b) Prove that R+ + +
with the maximal asymptotic transitive sets.
16. Let R ⊂ M be a maximal asymptotic transitive set. Assume that for a

given point x ∈ R there are T > 0 and u ∈ U such that ϕ (T, x, u) ∈ R.
Prove that ϕ (t, x, u) ∈ R for all t ∈ [0, T ].
17. Let X ⊂ M be any type of asymptotic transitive set (holding set, control
set, invariant control set, and so forth). Discuss the lift L (X).
18. Consider the control affine system on the plane R2 given by

2
ẋ = u (t) x x0 , with control range U = [1, 2], where x0 = (a, b) is
such that b/a is an irrational number. Induce this control system on
&
the torus T2 = R2 Z2 by the quotient map π : R2 → T2 (see Example
5.2.1). Prove the following:
(a) The point π (0) is the unique equilibrium point of the control sys-
tem on T2 .
(b) There is exactly one point π (x1 ) distinct from π (0) such that
ω + (π (x1 )) = {π (0)}, and exactly one point π (x2 ) distinct from
both π (0) and π (x1 ) such that ω − (π (x2 )) = {π (0)}. Further
ω − (π (x1 )) = ω + (π (x2 )) = T2 .
(c) ω − (π (x)) = ω + (π (x)) = T2 if π (x) ∈
/ {π (0) , π (x1 ) , π (x2 )}.
(d) x1 is negatively recurrent, but not positively recurrent, and x2 is
positively recurrent, but not negatively recurrent.
! control system on M = R determined by the vector fields

2
19.
Consider the
∂ ∂
, . Prove that the weak control sets are Dx = {x}, x ∈ R2 ,
∂x1 ∂x2
and there is no control set.
20. Let the asymptotic transitivity relation be restricted to the set R of

all relative positively recurrent points of the control system. Show that
Rx = Rx ∩ R is the equivalence class for every x in R .
21. Consider the unit circle S1 [0, 2π) and the control system on the
cylinder S1 × R given by ẋ = X0 (x) + u (t) X1 (x), with control range
U = [0, 1] and vector fields

X0 (x1 .x2 ) = sin2 x1 + x22 , x2 sin2 x1 + x22 ,
X1 (x1 , x2 ) = (0, f (x1 x2 )) ,
where f*is a non-negative C ∞ function such that f (x1 , x2 ) > 0 if and

$ π %2 π
only if x1 − + x22 < .
2 4

3π
(a) Let A = 0, × {0} ⊂ S1 × R. Show that A ⊂ R+ but
4
A ∩ R = ∅.
$π %
(b) Prove that , 0 is a positively recurrent point and A ⊂ R( π ,0) .
2 2
(c) Show that the maximal asymptotic transitive set R( π ,0) is not a
2
control set.
22. Consider the control system on R given by the equations
ẋ = (3 − u (t) − x) (u (t) − x) (−3 − u (t) − x)
with control range U = [−1, 1]. Show that the set [−5, 5] is positively
invariant and their subsets D1 = [−4, −2], D2 = (1−, 1), and D3 = [2, 4]
are the control sets of the system.
23. Prove that the control systems in Examples 1.3.1 and 1.3.2 are control-
lable.
! system on M = R determined by the vector

2
24. Show
that the control
∂ ∂
fields ± ,± is controllable.
∂x1 ∂x2
25. Prove that the number of main control sets is at most countable (finite
if M is compact).
+
26. Suppose that X = cl O>0 (x) is a nonempty set. Check that X is
x∈M
the unique closed invariant control set of the control system.
27. Assume that M is a compact manifold. Prove that the intersection

+
cl O>0 (x) is nonempty if and only if there is a unique compact
x∈M
+
invariant control set. In this case, cl O>0 (x) is the compact in-
x∈M
variant control set.
28. Consider the control system on M = R2 determined by the vector

fields X (x1 , x2 ) = (−x1 , −x2 ) and Y (x1 , x2 ) = x2 , x1 − x21 (Exam-
ple 5.1.3). Show that C0 = {0} and C1 = T \ {0} are the invariant
control sets of the system, where T is the tear drop.
3
29. Consider the control affine system ẋ = X0 (x) + i=1 ui (t) Xi (x) on
5 6
the plane M = R2 , with control range U = B 0, 103 , where
∂ ∂ ∂
X0 (x) = α (x2 ) + [1 − α (x2 )] , X1 (x) = β (x1 , x2 ) ,
∂x1 ∂x2 ∂x1
∂ ∂
X2 (x) = β (x1 , x2 ) , X3 (x) = γ (x1 , x2 ) ,
∂x2 ∂x2
and
⎧
⎨ =0 if x2 ≥ 1
α (x2 ) ∈ (0, 1) if 0 < x2 < 1 ,
⎩
=1 if x2 ≤ 0

>0 if x1 > 0 and x2 > 0
β (x1 , x2 ) ,
=0 otherwise
1
2
>0 if x21 + (x2 − 1) < 1
4
γ (x1 , x2 ) .
=0 otherwise
Discuss the control sets of this system. (Hint: See an analysis of this
system with unrestricted control range in [34, Example 3.2.9].)
30. Show that a locally accessible control system has countably many in-
variant control sets.
31. Let K ⊂ M be a positively invariant compact set. Prove that, for each
+
x ∈ K, there is an invariant control set Cx ⊂ cl O>0 (x) . Conclude
that there are finitely many invariant control sets in K if the system is
locally accessible from every point in K. (Hint: Use Theorems 3.1.3,
5.5.1, and 5.5.2.)
32. Consider a linear control system LΣ with compact and convex control
range U ⊂ Rm such that 0 ∈ int (U). Assume that the system sat-
isfies the Kalman condition. Show that, for every x ∈ Rd and every
open neighborhood V of x, the sets O+ (x) ∩ V and O− (x) ∩ V have
nonvoid interior. (Hint: For 0 ∈ U, check that ϕ (t, x, 0) = etA x ∈
$ %
int etA x + ϕ(t,0) (U) and then ϕ (t, x, 0) ∈ int Ot+ (x) for all x ∈ Rd
and t > 0.)
33. Show that the control system ẋ = u(t)x on Rn is not accessible from
any point.
34. Consider the control system on R2 determined by the vector fields X =

∂ ∂
and Y = f (x1 ) , where
∂x1 ∂x2
⎧
⎪
⎪ 0, if x1 ≤ 0
⎪
⎨ 1 − cos x , if 0 < x ≤ 2π
1 1
f (x1 ) = .
⎪
⎪ 0, if 2π < x ≤ 3π
⎪ 1
⎩ (x − 3π)2 , if x > 3π
1
Discuss the accessibility and the controllability of the system.
35. Let a < b < c and consider

a control system
! on R2 determined by
∂ ∂
the vector fields F = ± , f (x1 ) where f (x1 ) = 0 for x1 ∈
∂x1 ∂x2
(−∞, a] ∪ (b, c], f (x1 ) > 0 for x1 ∈ (a, b], and f (x1 ) < 0 for x1 ∈
(c, +∞). Show that this system is controllable.
36. Sometimes a control system is called controllable from x ∈ M if

O+ (x) = M . Let A = (x1 , x2 ) ∈ R2 : x1 < 0 and consider a control
system on R2 determined by the family of vector fields
!
∂ ∂ ∂ ∂
F= , f (x) , g (x) , h (x)
∂x1 ∂x1 ∂x2 ∂x2
where f (x) = g (x) = h (x) = 0 for x ∈ R2 \ A, f (x) , g (x) < 0 for

x ∈ A, and h (x) > 0 for x ∈ A. Show that this system is controllable
from every point in the set A but is not accessible from any point outside
A.
37. Study the control sets of the control system

0 u(t)
ẋ(t) = x(t)
−u(t) 0
on M = R2 with control range [0, 2] ⊂ R.

38. Consider the control system on M = R2 determined by the vector fields

∂
X= and Y (x1 , x2 ) = x2 , x1 − x21 . Study the control sets of this
∂x2
system.

∂
X= and Y (x1 , x2 ) = x2 , x1 − x31 . Study the control sets of this
∂x1
system.
40. Consider the control system on R2 determined by the vector fields

X (x1 , x2 ) = (x1 , −x2 ) and Y (x1 , x2 ) = x2 , x1 − x31 . Show that this
system is controllable on M = R2 \ {0}.
41. Study the controllability of the control system on R2 determined by the

vector fields
X (x1 , x2 ) = (−x1 − 2, −x2 ) ,

Y (x1 , x2 ) = (x1 − 2 + x2 , −x1 + 2 + x2 ) .
42. Let K ⊂ M be a compact positively invariant set. Assume that the

"
l
system is accessible from all points of K. Show that K = Aw (Ci )∩K
i=1
where C1 , ..., Cl are invariant control sets in K.
43. Show that a control affine system ẋ = X (x, u (t)) on a one-dimensional

manifold M is accessible if and only if for each x ∈ M there is some
u ∈ U such that X (x, u) = 0.

44. Let X : R2 ×R → R2 be the function X (x1 , x2 , u) = u, x21 and consider
the control system ẋ = X (x, u (t)) on R2 . Prove that this system is
locally accessible but not controllable.
45. Suppose that the manifold M is connected and the control affine system
on M is accessible. Show that the system group is transitive on M , that
is, G (x) = M for all x ∈ M . (Hint: use Proposition 2.2.8.)
46. Assume that the control affine system Σ is controllable and the time-
reversed control system Σ∗ is accessible. Show that both systems Σ and
Σ∗ are controllable (Lobry [73]).
47. Assume that the system group G is transitive on M and the set of
recurrent points is dense in M .
(a) If V ⊂ M is nonempty, open, and positively invariant, show that

V is dense in M .
(b) If both systems Σ and Σ∗ are accessible, prove that Σ is control-
lable. (Hint: Use item (a) for V = int (O+ (x)) and then apply
Problem 46.)
48. Let F be a family of local vector fields on the manifold M (see Problem
1.6(8)). For a given open set V ⊂ M , FV denotes the family of the vector
fields of F restricted to V . The control system determined by F is called
accessible from x ∈ M if int (SF (x)) = ∅; it is called locally accessible
from x if FV is accessible from x for every open neighborhood V of x.
If these properties hold for all x ∈ M the system is called accessible or
locally accessible, respectively.
(a) Assume that F is locally accessible from x ∈ M . Show that

cls (int (SF (x))) = cls (SF (x)).
!
∂ ∂
(b) Let F = , |V , where V = (x, y) ∈ R2 : x > 0 . Prove
∂x ∂y
that the control system on R2 determined by F is accessible but
not locally accessible.

49. Let α = X 1 , ..., X k be a finite sequence of local vector fields on the
manifold M . Define the map

ρα (t1 , ..., tk , x) = exp t1 X 1 ◦ · · · ◦ exp tk X k (x)
for (t1 , ..., tk ) ∈ Rk and x ∈ M such that the compositions make sense.
The domain of definition of ρα is then an open subset of Rk × M
and ρα is differentiable. By fixing a point x, one has the partial map
ρα,x (t1 , ..., tk ) = ρα (t1 , ..., tk , x). Let F be a family of local vector fields
on M such that the reunion of their domains of definition covers M .
The rank of F at x is the number rkF (x) defined by
- $ % .
rkF (x) = max rk d ρα,x (t ,...,t ) : t1 , ..., tk ≥ 0
α⊂F 1 k
$ %
where rk d ρα,x (t is the rank of the derivative d ρα,x (t
1 ,...,tk ) 1 ,...,tk )
(a) Show that the map x ∈ M

−→ rankF (x) is lower semicontinuous.
(b) Prove that the control system determined by F is accessible from

x if rkF (x) = dim M . (In this case we say that the control system
is normally accessible from x. It is possible to show that the
control system is accessible if and only if it is normally accessible,
that is, rkF (x) = dim M for every x ∈ M . See the proof in [107].)

∂ ∂
X= and Y = f (x) , with f (x) a C ∞ function such that f (x) = 0
∂x ∂y
dk f
if x = 0 and (0) = 0 for all k ≥ 0. Show that FV is normally
dxk
accessible for every open set V ⊂ M (or in other words, the system is
locally normally accessible).
∂ ∂
X= and Y = f (x, y) , with f a C ∞ function such that f (x, y) =
∂x ∂y
0 if and only if x ∈ [0, 1]. Prove that this system is normally accessible
but is not locally accessible from the origin (0, 0).
52. Consider the control system on R3 determined by the family of local
vector fields F = {X, Y, Z} where
∂
X = ,
∂x
#∞ !
∂ 1 1
Y = |V , with V = (x, y, z) : <x< ,
∂y 2k 2k − 1
k=1
#∞ !
∂ 1 1
Z = |W , with W = (x, y, z) : <x< .
∂z 2k + 1 2k
k=1
Show that this system is locally normally accessible from the origin
(0, 0, 0) but is not locally accessible from any point (x, y, z) = (0, 0, 0).
53. Let F be a countable family of C ∞ local vector fields on the manifold
M . Show that F is normally accessible from x ∈ M if F is accessible
from x ∈ M . (Hint: Note that the collection of all finite sequences
α ⊂ F is countable. Use the accessibility together with the Baire and
Sard theorems.)
54. Find an example of a control system with countably many control sets
with nonvoid interior. (Hint: See [34, Example 3.2.11] for a compact
case.)
55. Consider a linear control system ẋ = Ax + Bu (t) on Rd with compact

and convex control range U ⊂ Rm such that 0 ∈ int (U). Assume that
the system satisfies the Kalman condition. Show that O− (0) = Rd if
and only if Re λ ≤ 0 for every eigenvalue λ of A.
56. Discuss the controllability of the control system

a1 a2 b
ẋ = x + u (t) 1
a3 a4 b2
on R2 with restricted control range U = [−1, 1].
57. Consider a linear control system ẋ = Ax + Bu (t) on R2 where A =

1 0 1
,B= , and U = [−1, 1]. Prove that D = (−1, 1) × [−1, 1]
0 −1 1
is the unique effective control set of the system. Is there any control set
with empty interior?
58. Sometimes a control system is called approximately controllable

from x ∈ M if cls (O+ (x)) = M . Go back to Problem 1.6(5) and
define the sets

P = [x, 0] ∈ T2 : 0 ≤ x ≤ 1/2 and Q = T2 \ P.
Show that:
(a) P is positively invariant but not negatively invariant, hence Q is

negatively invariant but not positively invariant.
(b) Σ∗ is normally accessible.
(c) Σ is controllable from every point in Q but is neither approximately
controllable nor accessible from any point in P .
(d) Σ∗ is controllable from every point in P and approximately con-
trollable from every point in Q.
(e) P and Q are invariant control sets.
59. Show that the following sentences are equivalent:
(a) x ∈ M is a nonwandering point.

(b) every neighborhood U of x is positively recursive with respect to
itself, that is, O≥t
+
(U ) ∩ U = ∅ for all t > 0.
(c) every neighborhood U of x is negatively recursive with respect to

−
itself, that is, O≥t (U ) ∩ U = ∅ for all t > 0.
60. For any x ∈ M and u ∈ U , the motion ϕux : R → M is said to be

positively Lagrange stable if cl (ϕux (R+ )) is compact; it is called
negatively Lagrange stable if cl (ϕux (R− )) is compact; and it is called
Lagrange stable if cl (ϕux (R)) is compact. Show that:
(a) ϕux is positively Lagrange stable if and only if ω + (x) is a nonempty

compact set;
(b) ϕux is negatively Lagrange stable if and only if ω − (x) is a nonempty
compact set;
(c) ϕux is Lagrange stable if and only if both ω + (x) and ω − (x) are
nonempty compact sets.
61. Show that a point x is nonwandering if and only if x ∈ ω + (U ) for every

neighborhood U of x.
62. Assume that the prolongational transitivity relation is an equivalence

relation among the nonwandering points. If E ⊂ M is a nonempty
subset satisfying E ⊂ J + (x), for all x ∈ E, prove that there is a weak
prolongational control set containing E. Conclude that any weak control
set or asymptotic transitive set is contained in some weak prolongational
control set.

Periodic points are crucial for studies of stability and controllability. An
equivalence class of periodic points was termed a ‘holding set’ by Roxin [85].
Gayek and Vincent [54] named this equivalence class a ‘maneuverable set’ in a
specific examination of controllability. In [86], Roxin displayed the importance
of the holding sets to the study of optimal control. Kang [67] implicitly
dealt with holding sets by studying controllability near bifurcation equilibrium
points. Holding sets also relate to the concept of local controllability (e.g.,
Basto [9], Byrnes and Crouch [43], Hermes [57, 58], Sussman [108, 109]).
Problem 5.8(13) discusses the equivalence between Poincaré recurrence
and Poisson stability of control systems ([101],[102]). This equivalence is one
of the main connections between stability and controllability for semigroup

actions and control systems ([103]).
Most results on control sets treated in this chapter, mainly in Section 5.6,
are based on the work by Colonius and Kliemann [34], and the most proofs
are due to these authors. In their book, one can find many other problems on
control sets, for instance, dependence on control ranges, bifurcations, hyper-
bolicity, multiplicity, robustness, and semicontinuity.
The set R of relative positively recurrent points was studied by Bellicanta
[10].
The notion of an invariant control set emerged in studies of dynamical
problems of stochastic differential equations developed by Arnold and Klie-
mann [2] and Kliemann [68].
One of the open problems in control theory is to establish necessary and
sufficient conditions for the controllability of bilinear control systems. In
this line of investigation, sufficient conditions were presented by Jurdjevic
and Kupka [64, 65]. The controllability of two-dimensional bilinear control
systems was characterized by Braga Barros, Gonçalves F., do Rocio, and San
Martin [19].
More information about controllability of linear control systems can be
found in Lee and Markus [72].
Controllability of control systems on Lie groups has received special
attention in the literature. Jurdjevic and Sussmann [66] started the stud-
ies in this line of investigation. By extending the results of [64], Jurdjevic and
Kupka [65] established sufficient conditions for the controllability of invariant
systems on semisimple Lie groups. Subsequently, these conditions were inves-
tigated by El Assoudi and Gauthier [45], Gauthier, Kupka, and Sallet [53],
and Silva Leite and Crouch [91]. Ayala B. [5] and Hilgert and Neeb [59] pre-
sented necessary and sufficient conditions for the controllability of invariant
control systems on nilpotent Lie groups. In the setting of semigroup actions,
control sets for systems induced by semisimple Lie groups were studied by
San Martin [87] and San Martin and Tonelli [90].
Sometimes a control system is called locally accessible from a state x if
for every open neighborhood V of x the sets O+ (x) ∩ V and O− (x) ∩ V have
nonvoid interior. This notion comes from the accessibility rank condition ,
which means the full rank condition of the Lie algebra L (F) generated by the
vector fields of the control system, corresponding to the local accessibility as
defined in Definition 5.4.2 ([34, Theorems A.4.4 and A.4.5],[69]). In the real
analytic and nonsingular C ∞ cases, the accessibility is equivalent to the local

accessibility and to the accessibility rank condition ([34, Theorem A.4.6]). In
the case of linear control systems, the Kalman condition corresponds to the
accessibility rank condition.
The term ‘prolongational control set’ was reserved to indicate a weak pro-
longational control set with interior points, which might appear in future
studies on stability and controllability of control systems. A previous discus-
sion can be found in the paper [103].
Chapter 6
Attractors and repellers
The purpose of this chapter is: to study various notions of attractors and
repellers for control systems; to clarify the relations between them; and to
investigate control systems with global attractors.
The global attractor means an invariant compact set that attracts every
bounded set of states. A central result shows that the global attractor is a
Conley attractor of the system. Section 6.1 contains the basic definitions and
properties of attractors, weak attractors, uniform attractors, and weak uni-
form attractors. Section 6.2 is devoted to the characterization of the compact
invariant uniform attractors by means of Lyapunov functions. The Conley
definition for attractors is studied in Section 6.3. The relationship between
the notion of a uniform attractor and the notion of a Conley attractor is care-
fully explained. Finally, Section 6.4 deals with the notions of both a global
attractor and a global uniform attractor. While the global attractor is the
smallest compact set that attracts every bounded set, the global uniform at-
tractor is a uniform attractor whose region of uniform attraction coincides
with the whole state space. We prove that the global attractor is the global
invariant uniform attractor, and the converse holds under completeness. Nec-
essary and sufficient conditions for the existence of the global attractor are
discussed. A measure of noncompactness is used to describe the asymptotic
behavior of a control system with a global attractor.
Since we often explore the control flow of the system, some of Conley’s
results of flows are needed. We then suggest that the reader consult Appendix
A for unexplained properties of the Conley concepts of limit sets, attractors,
and repellers for dynamical systems.
169
170 Chapter 6 · Attractors and repellers
6.1 General concepts of attraction

In this section, we define regions of attraction of the sets in the state space
and present various notions of attractors. Analogously, we define regions of
repellency and repellers.
Definition 6.1.1 For a given subset X ⊂ M , the region of weak attrac-

tion of X is defined as
- .
Aw (X) = x ∈ M : O≥T+
(x) ∩ B (X, ε) = ∅ for every ε, T > 0 ;
the region of attraction of X is the set

1 2
x ∈ M : for each ε > 0 there is T > 0 such that
A (X) = ;
O≥T
+
(x) ⊂ B (X, ε)
the region of uniform attraction of X is the set

1 2
x ∈ M : for each ε > 0 there are δ, T > 0 such that
Au (X) = ;
O≥T
+
(B (x, δ)) ⊂ B (X, ε)
the region of weak uniform attraction of X is defined as

- .
Awu (X) = x ∈ M : O≥T+
(B (x, δ)) ∩ B (X, ε) = ∅ for every all ε, δ, T > 0 .
Any set Y in Aw (X), A (X), Au (X), or Awu (X) may be respectively said
to be weakly attracted, attracted, uniformly attracted, or weakly uni-
formly attracted to X.
Attractors are sets in the state space which lie in the interior of their
corresponding regions of attraction. In other words, an attractor is a set that
attracts a neighborhood of it, as in the following.
Definition 6.1.2 The set X ⊂ M is called weak attractor, attractor,

uniform attractor, or weak uniform attractor if there is ε > 0 such
that respectively B (X, ε) ⊂ Aw (X), B (X, ε) ⊂ A (X), B (X, ε) ⊂ Au (X), or
B (X, ε) ⊂ Awu (X).
We have the inclusions Au (X) ⊂ A (X) ⊂ Aw (X, F) ⊂ Awu (X) directly

from the definitions. Hence every uniform attractor is an attractor, every
Section 6.1 · General concepts of attraction 171
attractor is a weak attractor, and any weak attractor is a weak uniform at-
tractor. All these notions of attractors have distinct meaning (see [14, Chapter
V, Examples 1.7] for an explanation of illustrations in dynamical systems).
In the following we present the general invariance properties of regions of
attraction.
Proposition 6.1.1 Let X ⊂ M be a nonempty set. Then
1. Aw (X) is negatively invariant.
2. A (X) is positively invariant.
3. Au (X) is positively invariant.
4. Awu (X, F) is negatively invariant.
Proof. (1) Let x ∈ Aw (X), t > 0, and u ∈ U . For ε, T > 0, we

have O≥T +
(x) ∩ B (X, ε) = ∅. Then there are s ≥ T and v ∈ U such that
ϕ (s, x, v) ∈ B (X, ε). We can write
ϕ (s, x, v) = ϕ (s, ϕ (t, ϕ (−t, x, u) , u · (−t)) , v)

= ϕ (s + t, ϕ (−t, x, u) , w)
for some w ∈ U. As s + t > T , we have O≥T +

(ϕ (−t, x, u)) ∩ B (X, ε) = ∅, and
hence ϕ (−t, x, u) ∈ Aw (X). Therefore Aw (X) is negatively invariant.
(2) Let x ∈ A (X), t > 0, and u ∈ U. For ε > 0 there exists T > 0 such that
O≥T
+
(x) ⊂ B (X, ε). This implies that O≥T+
(ϕ (t, x, u)) ⊂ O≥T
+
(x) ⊂ B (X, ε),
and then ϕ (t, x, u) ∈ A (X). Thus A (X) is positively invariant.
(3) Let x ∈ Au (X), t > 0, and u ∈ U . For ε > 0 there are δ, T > 0 such
that O≥T +
(B (x, δ)) ⊂ B (X, ε). Since ϕut (B (x, δ)) is an open set containing

ϕ (t, x, u), we can find δ > 0 such that B ϕ (t, x, u) , δ ⊂ ϕut (B (x, δ)). It
follows that

O≥T
+
B ϕ (t, x, u) , δ ⊂ O≥T
+
(ϕut (B (x, δ)))
⊂ O≥T
+
(B (x, δ)) ⊂ B (X, ε)
and hence ϕ (t, x, u) ∈ Au (X). Thus Au (X) is positively invariant.

(4) Let x ∈ Awu (X), t > 0, and u ∈ U. For ε, δ, T > 0, there is δ > 0

such that B x, δ ⊂ ϕt
u·(−t)
(B (ϕ (−t, x, u) , δ)). Since x ∈ Awu (X), we have

O≥T
+
B x, δ ∩ B (X, ε) = ∅, hence

∅ = O≥T
+
B x, δ ∩ B (X, ε)
$ %
u·(−t)
⊂ O≥T
+
ϕt (B (ϕ (−t, x, u) , δ)) ∩ B (X, ε)
⊂ O≥T
+
(B (ϕ (−t, x, u) , δ)) ∩ B (X, ε) .
Thus ϕ (−t, x, u) ∈ Awu (X).
The regions of attraction of the compact sets have special properties, which
we will now discuss.
Proposition 6.1.2 If K ⊂ M is a compact weak attractor then Aw (K) is

an open set.
Proof. Take r > 0 such that B (K, r) ⊂ Aw (K) and fix x ∈ Aw (K). It

follows from the definition of Aw (K) that there are t > 0 and u ∈ U such that
ϕ (t, x, u) ∈ B (K, r). Since Aw (K) is negatively invariant, by Proposition
6.1.1, it follows that
x ∈ ϕu·t
−t (B (K, r)) ⊂ ϕ−t (Aw (K)) ⊂ Aw (K) .
u·t
−t (B (K, r)) is a neighborhood of x contained in Aw (K), and there-

Hence ϕu·t
fore Aw (K) is an open set in M .
The following theorem characterizes the regions of attraction of the com-

pact sets.
Proposition 6.1.3 Let K be a compact subset of M . Then

Aw (K) = x ∈ M : K ∩ ω + (x) = ∅ ,

A (K) = x ∈ M : ω + (x) = ∅ and ω + (x) ⊂ K ,

Au (K) = x ∈ M : J + (x) = ∅ and J + (x) ⊂ K ,

Awu (K) = x ∈ M : J + (x) ∩ K = ∅ .
$ %
Proof. Suppose that K ∩ ω + (x) = ∅. Then K ∩ cl O≥T
+
(x) = ∅ for
all T > 0. Hence, for every ε, T > 0, we have O≥T
+
(x) ∩ B (K, ε) = ∅, and
therefore x ∈ Aw (K). As to the converse, if x ∈ Aw (K) then for each
n ∈ N, we can find tn > n, un ∈ U , and xn ∈ K such that ϕ (tn , x, un ) ∈
B (xn , 1/n). Since K is compact, we may assume that xn → y in K, and then

ϕ (tn , x, un ) → y. Since tn → +∞, we have y ∈ K ∩ ω + (x). This completes
the proof for the first equality of the theorem.
Let x ∈ A (K). As A (K) ⊂ Aw (K), we have ω + (x) = ∅. Take y ∈ ω + (x)
and ε > 0. Since K is compact, there is δ > 0 such that cl (B (K, δ)) ⊂
B (K, ε). As x ∈ A (K), there is T > 0 such that O≥T
+
(x) ⊂ B (K, δ). Hence
$ %
y ∈ cl O≥T (x) ⊂ cl (B (K, δ)), and therefore y ∈ B (K, ε). Since ε > 0
+
is arbitrary and K is compact, it follows that y ∈ K. Thus ω + (x) = ∅ and

ω + (x) ⊂ K. Conversely, suppose that the point x ∈ M
satisfies ω + (x) = ∅ and ω + (x) ⊂ K. Let N ⊂ M be a compact neighborhood
of K. For a given ε > 0, take δ > 0 such that δ < ε and B (K, δ) ⊂ N . Then
O≥T
+
(x) ∩ B (K, δ) = ∅ for all T > 0. We claim that O≥T +
(x) ⊂ B (K, δ) for
some T > 0. Indeed, suppose by contradiction that O≥T (x) B (K, δ) for
+
every T > 0. Since O≥T +

(x) is connected, we can obtain tn > n and un ∈ U
such that ϕ (tn , x, un ) ∈ N \ B (K, δ). As N is compact and tn → +∞, it
follows that there is a point y ∈ ω + (x) ∩ (N \ B (K, δ)), which is a contra-
diction. Hence there is T > 0 such that O≥T +
(x) ⊂ B (K, δ) ⊂ B (K, U ), and
therefore x ∈ A (K). This finishes the proof for the second equality.
For the third equality, take x ∈ Au (K). Since Au (K) ⊂ A (K, F), we have
ω + (x) = ∅, hence J + (x) = ∅. Take y ∈ J + (x) and ε > 0. Choose δ > 0 such
that cl (B (K, δ)) ⊂ B (K, ε). As x ∈ Au (K), $ there are δ ,%T > 0 such that

O≥T
+
B x, δ ⊂ B (K, δ). Hence y ∈ cl O≥T +
B x, δ ⊂ cl (B (K, δ)),
and therefore y ∈ B (K, ε). Since ε > 0 is arbitrary, we have y ∈ K. Thus
J + (x) = ∅ and J + (x) ⊂ K. As to the converse, suppose that the point
x ∈ M satisfies J + (x) = ∅ and J + (x) ⊂ K. Let N ⊂ M be a compact
neighborhood of K. Since J + (x) is nonempty and compact, it follows that
ω + (x) is nonempty. For a given ε > 0, take δ > 0 with δ < ε and B (K, δ) ⊂
N . Then O≥T +
(B (x, r)) ∩ B (K, δ) = ∅ for all r, T > 0. Now, for each n ∈ N,
define the set Un ⊂ M by
- .
Un = y ∈ B (x, 1/n) : O≥n +
(y) ∩ B (K, δ) = ∅ .
Note that n ≥ m implies Um ⊂ Un . If y ∈ Un then there are t > 0 and

u ∈ U such that ϕ (t, y, u) ∈ B (K, δ). Hence ϕu·t
−t (B (K, δ)) ∩ B (x, 1/n) is
an open neighborhood of y. Moreover, if z ∈ ϕu·t
−t (B (K, δ)) ∩ B (x, 1/n) then
ϕ (t, z, u) ∈ B (K, δ), hence z ∈ Un . Thus ϕu·t
−t (B (K, δ)) ∩ B (x, 1/n) is an
open neighborhood of y contained in Un . This means that Un is an open
set contained in B (x, 1/n). There is n such that O≥n+
(Un ) ⊂ B (K, δ). In-
deed, suppose by contradiction that O≥n +

(Un ) B (K, δ) for every n. There
is then xn ∈ Un such that O≥n (xn ) B (K, δ). Since O≥n
+ +
(xn ) is con-
nected and O≥n (xn ) ∩ B (K, δ) = ∅, there are tn > n and un ∈ U such
+
that ϕ (tn , xn , un ) ∈ N \ B (K, δ). Since N is compact, we may assume

that the sequence (ϕ (tn , xn , un )) converges to a point z ∈ N \ B (K, δ). As
Un ⊂ B (x, 1/n) and tn → +∞, it follows that z ∈ J + (x) ∩ (N \ B (K, δ)),
which is a contradiction. Thus we can take n0 with O≥n +
0
(Un0 ) ⊂ B (K, δ).
If x ∈ Un for some n ≥ n0 then O≥n (Un ) ⊂ O≥n0 (Un0 ) ⊂ B (K, δ), with Un
+ +
an open neighborhood of x. If x ∈ / Un , for all n ≥ n0 , then there is T > n0

such that O≥T (x) ⊂ M \ N . In fact, x ∈
+
/ Un implies O≥n +
(x) ∩ B (K, δ) = ∅,
$ %
and then cl O≥n (x) ∩ B (K, δ) = ∅. This means that ω (x) ∩ B (K, δ) = ∅,
+ +
which contradicts ω + (x) ⊂ J + (x) ⊂ K. Thus there is n such that x ∈ Un

and O≥n+
(Un ) ⊂ B (K, δ). This proves that x ∈ Au (K).
We now show the last equality. Suppose that K ∩ J + (x) = ∅ and take
y ∈ K ∩ J + (x). Then B (y, δ) ∩ O≥T +
(B (x, ε)) = ∅ for all δ, ε, T > 0. Since
B (y, δ) ⊂ B (K, δ), we have x ∈ Awu (K). Conversely, take x ∈ Awu (K). For
each n ∈ N, we can take tn > n, un ∈ U , yn ∈ K, and xn ∈ B (x, 1/n) such
that ϕ (tn , xn , un ) ∈ B (yn , 1/n). By the compactness of K we may assume
that yn → y in K. Hence ϕ (tn , xn , un ) → y, with xn → x and tn → +∞. It
follows that y ∈ J + (x) ∩ K, and therefore K ∩ J + (x) = ∅.
Note the similarity between a region of weak attraction and a domain of

attraction of an asymptotic transitive set (Definition 5.2.4), and the similarity
between a region of weak uniform attraction and a domain of weak attraction
of a prolongational transitive set (Definition 5.7.5). These similarities came
up from the unification of dynamical systems and control systems by the
transformation semigroup theory. A possible connection between the notions
of attractors and prolongational transitive sets may be the concept of an
asymptotically stable set, which is defined in the next section.
Example 6.1.1 Consider the control affine system ẋ = u (t) X (x) on M =

R3 \ {0}, with control range U = [1, 10] and vector field

2 2 2
x1 − x1 x x2 − x2 x x3 − x3 x
X (x) = 2 , 2 , 2 .
x x x
Note that the points in the unit sphere S2 are global critical points. By using
spherical coordinates x = ρ sin φ cos θ, y = ρ sin φ sin θ, z = ρ cos φ, we have
the corresponding system
1 − ρ2
ρ = u (t) θ = 0 φ = 0.
ρ

Then
the variables θ and φ are constant and the variable ρ satisfies 1 − ρ2 =
e−2 u(t) dt . For any point x outside S2 , the trajectory through x is a straight
x
line converging to ∈ S2 as t → +∞, and it has no limit as t → −∞.
x
The controls only determine the velocities of the trajectories.
! Thus ω + (x) =
x
ω − (x) = x for every x ∈ S2 , while ω + (x) = and ω − (x) = ∅ for all
x
x ∈ M \ S2 . This means that the unit sphere is a compact attractor of the
system. Actually, it is easily seen that J + (x) = ω + (x) for all x ∈ M , and
therefore S2 is a compact uniform attractor.
All definitions and results on the concepts of attraction may be transferred

to the time-reversed control system, obtaining the notions of repellency.
Definition 6.1.3 For a given subset X ⊂ M , the region of weak repel-

lency of X is defined as
- .
−
Rw (X) = x ∈ M : O≥T (x) ∩ B (X, ε) = ∅ for every ε, T > 0 ;
the region of repellency of X is the set

1 2
x ∈ M : for each ε > 0 there is T > 0 such that
R (X) = − ;
O≥T (x) ⊂ B (X, ε)
the region of uniform repellency of X is the set

1 2
x ∈ M : for each ε > 0 there are δ, T > 0 such that
Ru (X) = − ;
O≥T (B (x, δ)) ⊂ B (X, ε)
the region of weak uniform repellency of X is defined as

- .
−
Rwu (X) = x ∈ M : O≥T (B (x, δ)) ∩ B (X, ε) = ∅ for every all ε, δ, T > 0 .
Any set Y in Rw (X), R (X), Ru (X), or Rwu (X) may be respectively said to
be weakly repelled, repelled, uniformly repelled, or weakly uniformly
repelled from X.
Definition 6.1.4 The set X ⊂ M is called weak repeller, repeller, uni-

form repeller, or weak uniform repeller if there is ε > 0 such that
respectively B (X, ε) ⊂ Rw (X), B (X, ε) ⊂ R (X), B (X, ε) ⊂ Ru (X), or
B (X, ε) ⊂ Rwu (X).
We have the inclusions Ru (X) ⊂ R (X) ⊂ Rw (X, F) ⊂ Rwu (X), hence

every uniform repeller is a repeller, every repeller is a weak repeller, and
any weak repeller is a weak uniform repeller. About invariance of regions
of repellency, Rw (X) is positively invariant, R (X) is negatively invariant,
Ru (X) is negatively invariant, and Rwu (X, F) is positively invariant.
For any compact set K ⊂ M , we have

Rw (K) = x ∈ M : K ∩ ω − (x) = ∅ ,

R (K) = x ∈ M : ω − (x) = ∅ and ω − (x) ⊂ K ,

Ru (K) = x ∈ M : J − (x) = ∅ and J − (x) ⊂ K ,

Rwu (K) = x ∈ M : J − (x) ∩ K = ∅ .
Moreover, if K is a compact weak repeller then Rw (K) is an open set.
6.2 Uniform attractor

In this section we explain the main properties of the uniform attractors.
We provide a characterization of the compact uniform attractors by means of
Lyapunov functions. All the results may be clearly reproduced for the uniform
repellers.
We need the following technical lemmas.
Lemma 6.2.1 Let P ⊂ M be a positively invariant compact set. If U and V

% in P with V ⊂ U and ω (U ) ⊂ V then there is t > 0
+
are open sets$ contained
such that cl O≥t+
(U ) ⊂ V .
$ %
Proof. Suppose to the contrary that cl O≥t +
(U ) V for all t > 0.
$ %
As cl O≥t+
(U ) ⊂ P , for each n ∈ N, there is xn ∈ P such that xn ∈
$ %
cl O≥n
+
(U ) \ V . By compactness, we may assume that xn → x in P . This
$ %
implies that x ∈ P \ V , as V is open, and x ∈ cl O≥n+
(U ) for all n ∈ N.
Hence x ∈ ω + (U ) \ V , which is impossible.
Section 6.2 · Uniform attractor 177
Lemma 6.2.2 Let K ⊂ M be a positively invariant compact set and U an

% such that ω $(U ) ⊂ K. Then
+
open neighborhood
$ of K % there are ε, t > 0 such
that cl O≥t
+
(B (K, ε)) ⊂ U , with cl O≥t
+
(B (K, ε)) compact in M .
Proof. Let U be a compact neighborhood of K such that U ⊂ U . There

are ε, t > 0 such that O≥t +
(B (K, ε)) ⊂ U . Indeed, suppose by contradiction

that O≥t (B (K, ε)) U for all ε, t > 0. Then O≥n
+ +
(B (K, 1/n)) U for all

n ∈ N with B (K, 1/n) ⊂ int (U ). Then there are sequences (tn ) in R+ , (xn )
in M , and (un ) in U such that tn ≥ n, xn ∈ B (K, 1/n), and ϕ (tn , xn , un ) ∈
M \ U . We may assume that xn ∈ int (U ) for all n ∈ N. It follows that
ϕ (τ n , xn , un ) ∈ fr (U ) for some τ n ≤ tn . Since fr (U ) is compact, we may as-
sume that ϕ (τ n , xn , un ) → z with z ∈ fr (U ). If (τ n ) is unbounded, it follows
that z ∈ ω + (U ) ⊂ ω + (U ) ⊂ K, which contradicts z ∈ fr (U ). Hence (τ n ) is
bounded, and therefore we can assume that τ n → τ . By the compactness of K,
we may assume that xn → x in K. Moreover, by the compactness of U , we may
assume that un → u in U . Thus z = limn→∞ ϕ (τ n , xn , un ) = ϕ (τ , x, u). Nev-
ertheless, z = ϕ (τ , x, u) ∈ K, since K is positively invariant, and then we have

a contradiction again.
$ Thus there% are ε, t > 0 such that O≥t (B (K, ε)) ⊂ U ,
which implies cl O≥t
+
(B (K, ε)) ⊂ U ⊂ U .
We now have the following consequence from Lemmas 6.2.1 and 6.2.2.
Proposition 6.2.1 Let K ⊂ M be a positively invariant compact set and U

an open neighborhood of K such that ω + (U ) ⊂ K. There
$ is then a% t > 0 and

an open neighborhood U of K such that U ⊂ U and cl O≥t +
(U ) ⊂ U .
$ %
Proof. By Lemma 6.2.2, there are ε, t > 0 such that cl O≥t
+
(B (K, ε)) is
$ %
compact and cl O≥t
+
(B (K, ε)) ⊂ U . As K is positively invariant, we have
O≤t
+
(K) ⊂ K. By continuity, since [0, t] and U are compact sets, there is
$ % $ %
δ > 0 such that cl O≤t+
(B (K, δ)) is compact and cl O≤t +
(B (K, δ)) ⊂ U .
For ε0 = min {ε, δ}, we have
$$ %%
cl O+ (B (K, ε0 )) = cl O≤t +
(B (K, ε0 )) ∪ O≥t
+
(B (K, ε0 )) ⊂ U.
Now, set P = cl (O+ (B (K, ε0 ))). Then P is a positively invariant compact

neighborhood of K such that ω + (int (P ))$ ⊂ ω + (U ) ⊂%K ⊂ int (P ). By
Lemma 6.2.1, there is t > 0 such that cl O≥t +
(int (P )) ⊂ int (P ), which
finishes the proof.
As a consequence, we have the following properties of the compact uniform

attractors.
Corollary 6.2.1 Let K ⊂ M be a compact and positively invariant uniform

attractor of the control
$ system.
% There is then a t > 0 and a neighborhood U
of K such that cl O≥t +
(U ) ⊂ U .
Proof. We have K ⊂ int (Au (K)), where

Au (K) = x ∈ M : J + (x) = ∅ and J + (x) ⊂ K
according to Proposition 6.1.3. Let V ⊂ int (Au (K)) be a compact neighbor-

"
hood of K. We have J + (V ) = x∈V J + (x) ⊂ K. According to Proposition
4.2.6, we have ω + (V ) ⊂ J + (V ), and then ω + (int (V )) ⊂ K. The result fol-
lows according to Proposition 6.2.1.
For the following, recall that a set X ⊂ M is isolated invariant if and only
if it has a neighborhood V such that x ∈ V and O (x) ⊂ V implies x ∈ X.
Proposition 6.2.2 Every compact and invariant uniform attractor of the

control system is isolated invariant.
Proof. Let K ⊂ M be a compact and invariant uniform attractor and U

an open neighborhood of K such that ω + (U ) ⊂ K. If x ∈ U and O (x) ⊂ U
−
then O≥t (x) ⊂ U for all t > 0. Hence x ∈ O≥t
+
(U ) for all t > 0, and therefore
x ∈ ω (U ) ⊂ K. Thus K is isolated invariant.
+

Compact uniform attractors are also asymptotically stable, as in the fol-

lowing definition.
Definition 6.2.1 A set X ⊂ M is said to be uniformly stable if for each

ε > 0 there is δ > 0 such that O+ (B(X, δ)) ⊂ B (K, ε). The set X is called
asymptotically stable if it is a uniformly stable attractor.
For compact sets, uniform stability and equistability have the same mean-
ing (see the notes for Chapter 4 and Problem 6.5(13)).
Proposition 6.2.3 Every compact and positively invariant uniform attractor

is asymptotically stable.
Proof. Let K ⊂ M be a compact and positively invariant uniform at-

tractor. By definition, K ⊂ int (Au (K)), hence for every x ∈ K and ε > 0
there are δ x , Tx > 0 such that O≥T
+
x
(B (x, δ x )) ⊂ B (K, ε). As K is positively
invariant, we have O≤Tx (x) ⊂ K. By continuity, there is δ x > 0 such that
+

O≤T
+
x
B x, δ x ⊂ B (K, ε). By taking δ x = min δ x , δ x , it follows that

O+ B(x, δ x ) ⊂ O≤T+
B x, δ x ∪ O≥T +
(B (x, δ x )) ⊂ B (K, ε) .
"
Now take δ > 0 such that B (K, δ) ⊂ x∈K B x, δ x . We then have
#
O+ (B (K, δ)) ⊂ O+ B x, δ x ⊂ B (K, ε) .
x∈K
In the following we exhibit a property of the uniform attractor that is

important to describe the global dynamical behavior of the control system.
We use the description of regions of attraction for a compact set given by
Proposition 6.1.3.
Proposition 6.2.4 Let A ⊂ M be a compact uniform attractor. For any

x ∈ M , one has ω − (x) , ω + (x) ⊂ A ∪ (M \ int (Au (A))), with the following
conditions:
/ A ∪ (M \ int (Au (A))) then ω − (x) ⊂ J − (x) ⊂ M \ int (Au (A))

1. If x ∈
and ω + (x) ⊂ J + (x) ⊂ A.
2. If x ∈ A then ω + (x) ⊂ J + (x) ⊂ A, and either ω − (x) ⊂ M \int (Au (A))

or ω − (x) ⊂ A if ω − (x) is nonempty and connected.
3. If x ∈ M \ int (Au (A)) then ω − (x) ⊂ J − (x) ⊂ M \ int (Au (A)), and
either ω + (x) ⊂ M \ int (Au (A)) or ω + (x) ⊂ A if ω + (x) is nonempty
and connected.
Proof. If x ∈ / A ∪ M \ int (Au (A)) then x ∈ Au (A), hence ω + (x) ⊂

J (x) ⊂ A. Suppose to the contrary that there is y ∈ J − (x) ∩ int (Au (A)).
+
Then x ∈ J + (y) ⊂ A, which is a contradiction. Thus ω − (x) ⊂ J − (x) ⊂

M \ int (Au (A)). This proves item (1) and the inclusion ω − (x) , ω + (x) ⊂
A ∪ (M \ int (Au (A))) in the case x ∈
/ A ∪ M \ int (Au (A)). Now, if x ∈ A
then clearly J + (x) ⊂ A. We claim that ω − (x) ⊂ A ∪ (M \ int (Au (A))).
In fact, suppose that ω − (x) M \ int (Au (A)) and take y ∈ ω − (x) \
(M \ int (Au (A))). By Proposition 5.7.1, y is nonwandering. As y ∈ Au (A), it

follows that y ∈ J + (y) ⊂ A. Hence ω − (x) ⊂ A ∪ (M \ int (Au (A))), proving
the claim. If x ∈ M \int (Au (A)) then J − (x) ⊂ M \int (Au (A)), as in the first
part of the proof. It remains to show that ω + (x) ⊂ A ∪ (M \ int (Au (A))).
In fact, suppose that ω + (x) A and take y ∈ ω + (x) \ A. Then y is nonwan-
dering, which implies y ∈ M \ int (Au (A)). This proves the first part of the
proposition. To show item (2), let x ∈ A and suppose that ω + (x) is nonempty
and connected. Then J + (x) ⊂ A and ω − (x) ⊂ A ∪ (M \ int (Au (A))). Since
A and M \ int (Au (A)) are disjoint and ω − (x) is nonempty and connected,
it follows that either ω − (x) ⊂ M \ int (Au (A)) or ω − (x) ⊂ A. Item (3) is
proved in the same way.
Thus, by means of a uniform attractor, we may estimate a region of M

that contains all the limit sets of points. Note that the condition of connected
limit sets is satisfied under compactness (Corollary 4.1.1). This has some
interesting implications.
Corollary 6.2.2 Let A ⊂ M be a compact uniform attractor and x ∈ M .

Assume that both limit sets ω − (x) and ω + (x) are nonempty and connected:
1. If ω − (x) , ω + (x) ⊂ A then x ∈ A.
2. If ω − (x) , ω + (x) ⊂ M \ int (Au (A)) then x ∈ M \ int (Au (A)).
Proof. It follows immediately from Proposition 6.2.4.
Corollary 6.2.3 Assume that all positive limit sets of points are connected.
If A ⊂ M is a compact uniform attractor then A (A) is an open invariant set
and coincides with Aw (A).
Proof. According to Propositions 6.1.1 and 6.1.2, it is enough to show that

A (A) = Aw (A). In fact, if x ∈ Aw (A) then ω + (x) ∩ A = ∅. This means that
ω + (x) is nonempty and connected. According to Proposition 6.2.4, it follows
that ω + (x) ⊂ A, hence x ∈ A (A). This proves the inclusion Aw (A) ⊂ A (A).
Since A (A) ⊂ Aw (A) holds in general, we have A (A) = Aw (A).

In Example 5.5.1, the set D = x ∈ R2 : x ≥ 1 is a positively invariant
uniform attractor not negatively invariant. In the following we present an
example of an invariant uniform attractor.
ẋ (t) = X0 (x (t)) + u (t) X1 (x (t))
on M = R2 , with control range U = [1, 2], where X0 , X1 are vector fields given
by

X0 (x, y) = (y, −x) and X1 (x, y) = x − xy 2 − x3 , y − yx2 − y 3 .
In order to study the trajectories of this system, we use polar coordinates r

and θ with x = r cos θ and y = r sin θ. Note that the origin 0 is the unique
equilibrium point of the system. For u ∈ [1, 2] and (r, θ), r > 0, the trajectory
of the vector field Xu through (r, θ) is computed by
* −1
e−2ut
exp (tXu ) (r, θ) = 1− e−2ut + ,θ − t
r2
for all t ∈ R if 0 < r ≤ 1, and

* −1
e−2ut
exp (tXu ) (r, θ) = 1 − e−2ut + ,θ − t
r2
for t ∈ (−ε, +∞) if r > 1,where ε > 0 depends on u and r. A general

transition determined by the vector fields is of the form
exp (tn Xun ) ◦ · · · ◦ exp (t1 Xu1 ) (r, θ)

⎛9 n −1 ⎞
n −2u j t j ⎟
n
⎜ −2uj tj e j=1
= ⎝ 1 − e j=1 + 2
,θ − tj ⎠ .
r j=1
: −1
−2ut
Note that 1 − e−2ut + e r2 → 1 as t → +∞, the unit disk
D = {(r, θ) : r ≤ 1} is a compact invariant set, and the unit circle S1 is
2
a compact invariant set that is a periodic trajectory. Some trajectories of

the control system are plotted in Fig. 6.1. Consider the neighborhood V =
{(r, θ) : r < 1} of the origin 0. For sequences ((rn , θn ))n∈N in V , (un ) in U ,
and tn → −∞, we have ϕ (tn , (rn , θn ) , un ) → 0. Hence ω − (V ) = {0},
which means that J − (x) = {0} for every x ∈ V , and therefore {0} is an
invariant uniform repeller with a region of uniform repellency Ru ({0}) = V .
Now consider the open set W = R2 \ {0}. Then W is a neighborhood of the

unit circle S1 . By invariance, we have S1 = ω + S1 ⊂ ω + (W ). On the other
hand, if y ∈ ω + (W ) then there are sequences ((rn , θn ))n∈N in W , (un ) in U ,
and tn → +∞ such that ϕ (tn , (rn , θn ) , un ) → y. As ϕ (tn , (rn , θn ) , un ) →
1, it follows that y ∈ S1 . Hence ω + (W ) = S1 and therefore S1 is an invariant

uniform attractor with a region of uniform attraction Au S1 = W . Thus

M \ Au S1 = {0}. By similar calculations, we can see that the unit disk D2

is also a uniform attractor with global region of attraction Au D2 = R2 .
Figure 6.1: The unit circle is an invariant uniform attractor and the origin is
an invariant uniform repeller for this control system on the plane.
In general, we have Au (A) ⊂ A (A) if A is a compact uniform attrac-

tor. The following result presents a necessary and sufficient condition for the
equalities Au (A) = A (A) = Aw (A).
Proposition 6.2.5 Let A ⊂ M be a compact uniform attractor of the control
system. Then Au (A) = A (A) = Aw (A) if and only if Au (A) is invariant.
Proof. Since A is a weak attractor, the set Aw (A) is open according
to Proposition 6.1.2. Suppose that Au (A) = A (A) = Aw (A). According
to Proposition 6.1.1, Au (A) is invariant. Conversely, suppose that Au (A) is
open and invariant. According to Proposition 2.1.2, int (Au (A)) is invariant,
hence M \ int (Au (A)) is invariant and closed. If x ∈ M \ int (Au (A)), it
follows that ω + (x) ⊂ M \ int (Au (A)). Since A ⊂ int (Au (A)), it follows
that ω + (x) ∩ A = ∅. Thus x ∈ M \ Aw (A), which means that Aw (A) ⊂
int (Au (A)). As the inclusions Au (A) ⊂ A (A) ⊂ Aw (A) hold in general, we
have Au (A) = Au (A) = Aw (A).
Thus Au (A) is open if it is invariant. The following result characterizes

the compact invariant uniform attractors with invariant regions of uniform
attraction.
Proposition 6.2.6 Assume that all limit sets of points are nonempty. Let
A ⊂ M be a compact uniform attractor. Then
1. A is invariant if and only if ω − (x) , ω + (x) ⊂ A for every x ∈ A.
2. Au (A) is invariant if and only if ω − (x) , ω + (x) ⊂ M \ Au (A) for every

x ∈ M \ Au (A).
Proof. (1) Suppose that ω − (x) , ω + (x) ⊂ A for every x ∈ A. Let x ∈ A,

t > 0, and u ∈ U . Suppose by contradiction that ϕ (−t, x, u) ∈ / A. Since
A and M \ int (Au (A)) are disjoint closed sets, we can find s > 0 such
that ϕ (−s, x, u) ∈ / A ∪ (M \ int (Au (A))). According to Proposition 6.2.4,
it follows that ω − (ϕ (−s, x, u)) ⊂ M \ int (Au (A)) and ω + (ϕ (−s, x, u)) ⊂
A. According to Proposition 4.1.7 and the condition ω − (x) ⊂ A, we have
ω − (ϕ (−s, x, u)) ⊂ ω − (x) ⊂ A. Since ω − (ϕ (−s, x, u)) is nonempty, this
contradicts A ∩ (M \ int (Au (A))) = ∅. Hence ϕ (−t, x, u) ∈ A, and therefore
A is negatively invariant. Analogously, we can prove that A is positively in-
variant, and thus A is invariant. Conversely, if A is invariant then it is clear
that ω − (x) , ω + (x) ⊂ A for every x ∈ A.
(2) Suppose that ω − (x) , ω + (x) ⊂ M \ Au (A) for every x ∈ M \ Au (A).
We show that Au (A) is invariant. Since Au (A) is positively invariant ac-
cording to Proposition 6.1.1, it is enough to prove that Au (A) is negatively
invariant, or equivalently M \ Au (A) is positively invariant. In fact, take
x ∈ M \ Au (A), t > 0, and u ∈ U. Then ω + (x) ⊂ M \ Au (A). According to
Proposition 4.1.7, ω + (ϕ (t, x, u)) ⊂ ω + (x) ⊂ M \ Au (A). As ω + (ϕ (t, x, u))
is nonempty by hypothesis and ω + (ϕ (t, x, u)) ⊂ J + (ϕ (t, x, u)), it follows
that J + (ϕ (t, x, u)) A, that is, ϕ (t, x, u) ∈ M \ Au (A). Thus M \ Au (A)
is positively invariant, as desired. As to the converse, if Au (A) is invariant
then Au (A) is open, according to Proposition 6.2.5. Hence M \ Au (A) is
invariant and closed, and therefore ω − (x) , ω + (x) ⊂ M \ Au (A) for every
x ∈ M \ Au (A).
Proposition 6.2.6 applies to the special case of a compact state space, since
all the limit sets of points are nonempty under compactness.
We shall now prove one of the main results of this section that charac-
terizes the compact and positively invariant uniform attractors via Lyapunov
functions.
Definition 6.2.2 A Lyapunov function for a compact set K ⊂ M is a

real-valued function φ defined on a neighborhood N of K such that
1. φ (x) = 0 if x ∈ K, and φ (x) > 0 if x ∈ N \ K.
2. φ (ϕ (t, x, u)) < φ (x) for every x ∈ N \ K, t > 0, and u ∈ U.
3. φ (ϕ (t, x, u)) → 0 as t → +∞ for each x ∈ N and u ∈ U.
Lyapunov functions may be continuous or smooth. Let us see some exam-

ples.

u (t) 1
ẋ (t) = x (t)
−1 0
on M = R2 with control range U = [−2, −1]. The origin 0 in R2 is a global

equilibrium point. Hence the set K = {0} is invariant. Define the smooth
2
function φ : R2 → R by φ (x) = x . Then φ (0) = 0 and φ (x) > 0 if x = 0.
For x = 0 and u ∈ U, write ϕ (t, x, u) = (ϕ1 (t) , ϕ2 (t)). We have
d 2
φ (ϕ (t, x, u)) = 2 (ϕ1 (t)) u (t) < 0
dt
for all t ∈ R. Hence φ is decreasing along the trajectories of the system. The
solution for the constant function u = −2 is

−t a + bt + b
ϕ (t, x, u) = e
a + bt + 2b
and the solution for u ∈ (−2, −1] is

√
2
√
2

ut/2 $ √ %a cos 4−u2 +$b sin 4−u %
ϕ (t, x, u) = e √ √ 2 √ ,
b 4−u2 2 2 4−u2
2 − au
2 cos 4−u
2 + − a 4−u2 − bu
2 sin 2
where a, b ∈ R. In any case, we have φ (ϕ (t, x, u)) → 0 as t → +∞. Hence,

φ (ϕ (t, x, u)) → 0 as t → +∞, for each x ∈ R2 and u ∈ U. Thus φ is a
smooth Lyapunov function for K = {0}.
Example 6.2.3 Lyapunov functions need not be smooth. Consider the con-
trol affine system ẋ = u (t) x on M = Rn with control range U = [−2, −1].

t
For u ∈ U and x ∈ Rn , the solution is ϕ (t, x, u) = e 0 u(s)ds x. The origin 0 in
Rn is a global equilibrium point. Hence the set K = {0} is invariant. Define
φ : Rn → R by φ (x) = x. It is easily seen that (1) φ (x) = 0 if x = 0 and
φ (x) > 0 if x = 0; (2) φ (ϕ (t, x, u)) < φ (x) for x = 0, t > 0, and u ∈ U ;
and (3) φ (ϕ (t, x, u)) → 0 as t → +∞ for each x ∈ Rn and u ∈ U. Hence φ
is a continuous Lyapunov function for K = {0}, but φ is not differentiable at
x = 0.
We now prove the main theorem.
Theorem 6.2.1 A compact set K ⊂ M is a positively invariant uniform

attractor if and only if there exists a continuous Lyapunov function φ defined
on a positively invariant neighborhood N of K.
Proof. Suppose that K is a positively invariant uniform attractor. By

Proposition 6.2.3, K is asymptotically stable. Since K ⊂ int (Au (K)), there
is η > 0 such that cl (O+ (B (K, η))) is compact and cl (O+ (B (K, η))) ⊂
Au (K). As K and M \ O+ (B (K, η)) are disjoint closed sets and M is a
perfectly normal space, there is a continuous function f : M → [0, 1] such
that f −1 (0) = K and f −1 (1) = M \ O+ (B (K, η)). Define h : M → [0, 1] by
h (x) = sup {f (ϕ (s, x, v)) : s ≥ 0, v ∈ U} .
Since f −1 (0) = K, it is easily seen that h−1 (0) ⊂ K. As K is positively

invariant, K ⊂ h−1 (0), and therefore h−1 (0) = K. Then h (x) > 0 if x ∈
/ K.
For x ∈ K, t > 0, and u ∈ U, we have
h (ϕ (t, x, u)) = sup {f (ϕ (s, ϕ (t, x, u) , v)) : s ≥ 0, v ∈ U}

≤ sup {f (ϕ (s + t, x, w)) : s ≥ 0, w ∈ U}
≤ h (x) .
Hence, h is nonincreasing on trajectories.

We claim that h is continuous. It is enough to show that h−1 ([0, ε)) and
h ((ε, 1]) are open sets in M for every 0 < ε < 1. Let x ∈ h−1 ((ε, 1]).
−1
Then there are t ≥ 0 and u ∈ U such that f (ϕ (t, x, u)) > ε. Hence
−1
ϕu·t
−t f ((ε, 1]) ⊂ h−1 ((ε, 1]) is an open neighborhood of x. Thus h−1 ((ε, 1])
is an open set in M . Now, let x ∈ h−1 ([0, ε)). Then, h (x) < ε and
O+ (x) ⊂ f −1 ([0, h (x))). By taking δ > 0 with h (x) < δ < ε, f −1 ([0, δ)) is
an open neighborhood of K ∪ {x} and f −1 ([0, δ)) ⊂ O+ (B (K, η)) because
f −1 (1) = M \ O+ (B (K, η)). By Proposition 4.2.6, we have

ω + f −1 ([0, δ)) ⊂ J + cl O+ (B (K, η)) ⊂ K.
−1
By Lemma 6.2.1, there is t > 0 such that O≥t +
f ([0, δ)) ⊂ f −1 ([0, δ)).
Moreover, by continuity, there is a neighborhood V of x such that V ⊂
f −1 ([0, δ)) and O≤t +
(V ) ⊂ f −1 ([0, ε )) for some ε , with δ ≤ ε < ε. We
then have O (V ) = O≤t
+ +
(V ) ∪ O≥t
+
(V ) ⊂ f −1 ([0, ε )), which implies V ⊂
h−1 ([0, ε)). Hence h−1 ([0, ε)) is an open set in M , and therefore h is contin-
uous.
We claim that h−1 (1) = M \O+ (B (K, η)). In fact, if x ∈ M \O+ (B (X, η)) =
f −1 (1) then h (x) = 1, hence M \O+ (B (K, η)) ⊂ h−1 (1). On the other hand,
if h (y) = 1 and y ∈ O+ (B (K, η)) then for each n ∈ N there are tn ≥ n and
un ∈ U such that f (ϕ (tn , y, un )) > 12 , because f < 1 at O+ (B (K, η)). As
ϕ (tn , y, un ) ∈ O+ (B (K, η)) and cl (O+ (B (K, η))) is compact, we may as-
sume that ϕ (tn , y, un ) → z in cl (O+ (B (K, η))). Because tn → +∞, we have
z ∈ ω + (y) ⊂ J + (y) ⊂ K, and therefore f (ϕ (tn , y, un )) → f (z) = 0, which
is a contradiction. Hence y ∈ M \ O+ (B (K, η)), and we have h−1 (1) =
M \ O+ (B (K, η)).
We now define the function l : M × U → [0, 1] by
+∞
l (x, u) = e−t h (ϕ (t, x, u)) dt.
0
Since h and ϕ are continuous, it follows that l is continuous. As h−1 (0) = K

and h−1 (1) = M \ O+ (B (K, η)), we have l−1 (0) = K × U and l−1 (1) =
(M \ O+ (B (K, η))) × U . Because h is nonincreasing on trajectories, we have
+∞
l (ϕ (s, x, u) , u · s) = e−t h (ϕ (t, ϕ (s, x, u) , u · s)) dt
0
+∞
= e−t h (ϕ (s, ϕ (t, x, u) , u · t)) dt
0
+∞
≤ e−t h (ϕ (t, x, u)) dt
0
= l (x, u)
for all s > 0 and x ∈ M . If x ∈ O+ (B (K, η)) \ K, there is ε > 0 such that
h (x) ≥ ε. Since ω + (x) ⊂ K, there is T > 0 such that O≥T
+
(x) ⊂ h−1 ([0, ε)).
For u ∈ U, it follows that
t = sup {t > 0 : h (ϕ (t, x, u)) ≥ ε}
is finite and h (ϕ (t , x, u)) ≥ ε. If s > 0, the function t ∈ (0, +∞)

→
h (ϕ (t + s, x, u)) − h (ϕ (t, x, u)) is not identically zero, since
h (ϕ (t + s, x, u)) − h (ϕ (t , x, u)) < 0.
Hence
+∞
l (ϕ (s, x, u) , u · s) − l (x, u) = e−t (h (ϕ (t + s, x, u)) − h (ϕ (t, x, u))) dt
0
is strictly negative, that is, l (ϕ (s, x, u) , u · s) < l (x, u).

Finally, the Lyapunov function φ : M → [0, 1] is defined as
φ (x) = sup l (x, u) .

u∈U
Indeed, since l−1 (0) = K × U and l−1 (1) = (M \ O+ (B (K, η))) × U , we have
φ−1 (0) = K and φ−1 (1) = M \ O+ (B (K, η)). To show that φ is continuous,
we choose ε such that 0 < ε < 1. If x ∈ φ−1 ([0, ε)), we can take ε such that
φ (x) < ε < ε. Then l (x, u) < ε for all u ∈ U . Since l is continuous, there is
a neighborhood V of x in M such that V × U ⊂ l−1 ([0, ε )). If y ∈ V , then
l (y, u) < ε for all u ∈ U, hence φ (y) ≤ ε < ε. It follows that V ⊂ φ−1 ([0, ε)),
and therefore φ−1 ([0, ε)) is open. If x ∈ φ−1 ((ε, 1]), there is some u such that
l (x, u) > ε. Then there is a neighborhood V of x with V × {u} ⊂ l−1 ((ε, 1]).
If y ∈ V , we have l (y, u) > ε, hence φ (y) > ε. It follows that V ⊂ φ−1 ((ε, 1]),
and therefore φ−1 ((ε, 1]) is open. Thus φ is continuous. We now show that φ is
strictly decreasing on trajectories in O+ (B (K, η)) \ K. Let s > 0, v ∈ U, and
x ∈ O+ (B (K, η))\K. Since φ (ϕ (s, x, v)) = l (ϕ (s, x, v) , u0 ) for some u0 ∈ U,
and ϕ (t, ϕ (s, x, v) , u0 ) = ϕ (t + s, x, w), where w ∈ U is the s-concatenation
of v and u0 , we have
φ (ϕ (s, x, v)) = l (ϕ (s, x, v) , u0 )

+∞
= e−t h (ϕ (t, ϕ (s, x, v) , u0 )) dt
0
+∞
= e−t h (ϕ (t + s, x, w)) dt
0
+∞
= e−t h (ϕ (t, ϕ (s, x, w) , w · s)) dt
0
= l (ϕ (s, x, w) , w · s)
< l (x, w)
≤ φ (x) .
Therefore, φ is a Lyapunov function.

By denoting N = O+ (B (K, η)), it remains to show that φ (ϕ (t, x, u)) → 0
as t → +∞ for each x ∈ N and u ∈ U . Indeed, if x ∈ N then J + (x) ⊂ K
and ϕ (t, x, u) ∈ N for all t > 0. As cl (N ) is compact, we can take a sequence
(ϕ (tn , x, u)) that converges to some y ∈ cl (N ) as tn → +∞. Then y ∈ J + (x),
which implies φ (y) = 0. By continuity, φ (ϕ (tn , x, u)) → 0. For a given ε > 0,
choose n such that φ (ϕ (tn , x, u)) < ε. For t > tn , we have
φ (ϕ (t, x, u)) = φ (ϕ (t − tn , ϕ (tn , x, u) , u · tn ))

< φ (ϕ (tn , x, u))
< ε.
Hence φ (ϕ (t, x, u)) → 0 as t → +∞.

We now show the converse. Suppose there exists a continuous Lyapunov
function φ defined on a positively invariant neighborhood N of K. We
claim that K is positively invariant. In fact, suppose to the contrary that
ϕ (τ , x, u) ∈
/ K, with x ∈ K, τ > 0, and u ∈ U. Because of the property (1) of
Definition 6.2.2, φ (ϕ (τ , x, u)) = δ > 0. Let t∗ = inf {t > 0 : φ (ϕ (t, x, u)) ≥ δ}.
Then, t∗ > 0, φ (ϕ (t∗ , x, u)) ≥ δ, and 0 < φ (ϕ (t∗ − ε, x, u)) < δ for some
small ε > 0. But this shows by using (2) that
δ ≤ φ (ϕ (t∗ , x, u)) = φ (ϕ (ε, ϕ (t∗ − ε, x, u) , u · (t∗ − ε)))

< φ (ϕ (t∗ − ε, x, u)) < δ
which is a contradiction. Thus K is positively invariant. Now choose η > 0

such that cl (B (K, η)) ⊂ N is compact. Define
m = min (φ (x) : x ∈ fr (B (K, η))) .
Because of the property (1), m > 0. By the property (2), it follows that P =
φ−1 ([0, m]) ∩ cl (B (X, η)) is a compact and positively invariant neighborhood
of K. If x ∈ int (P ) then ∅ = J + (x) ⊂ P . Take any y ∈ J + (x). Then there
are sequences tn → +∞, xn → x, and (un ) in U such that ϕ (tn , xn , un ) →
y. For each k ∈ N, we can take nk such that tn > k for all n ≥ nk . If
ϕ (k, xn , un ) ∈ K, then φ (ϕ (tn , xn , un )) = 0 = φ (ϕ (k, xn , un )), since K is
positively invariant. If ϕ (k, xn , un ) ∈
/ K then
φ (ϕ (tn , xn , un )) = φ (ϕ (tn − k, ϕ (k, xn , un ) , un · k)) < φ (ϕ (k, xn , un )) .
Thus we have φ (ϕ (tn , xn , un )) ≤ φ (ϕ (k, xn , un )) for all n ≥ nk . We may

assume that un → u. It follows that φ (y) ≤ φ (ϕ (k, x, u)) for all k ∈ N. Be-
cause of the property (3), φ (y) = 0, and therefore y ∈ K. Thus J + (x) ⊂ K
for all x ∈ int (P ), which implies that K is a uniform attractor.
In Examples 6.2.2 and 6.2.3, because of the existence of a Lyapunov func-

tion, the origin of the Euclidean state space is an invariant uniform attractor
of the control system. In Example 6.2.1, the unit disk D2 = {(r, θ) : r ≤ 1} is
an invariant uniform attractor with the Lyapunov function

0 if r ≤ 1
φ (r, θ) =
r − 1 if r > 1
(see Problem 6.5(3)).

We finish this section by presenting a general smooth case of Lyapunov
functions for uniform attractors. For the following, dφx denotes the derivative
of a function φ at the point x.
Theorem 6.2.2 Let K ⊂ M be a compact set. Assume that there exists a

continuously differentiable real-valued function φ defined on a compact neigh-
borhood N of K that satisfies the following properties:
1. φ (x) = 0 if x ∈ K, and φ (x) > 0 if x ∈ N \ K.
2. dφx (X (x, u)) ≤ −f (d (x, K)) for all x ∈ N \ K and u ∈ U, where

f : [0, +∞) → [0, +∞) is a continuous strictly increasing function with
f (0) = 0.
Then φ is a Lyapunov function and K is a positively invariant uniform

attractor.
Proof. By condition (2), we have
d
φ (ϕ (t, x, u)) = dφϕ(t,x,u) (X (ϕ (t, x, u) , u (t)))
dt
≤ −f (d (ϕ (t, x, u) , K)) ≤ 0
for every x ∈ N , u ∈ U , and t > 0 such that ϕ (t, x, u) ∈ N , hence the real
function φ (ϕ (t, x, u)) is decreasing. In particular, if x ∈ K then φ (ϕ (t, x, u)) ≤
φ (x) = 0, hence φ (ϕ (t, x, u)) = 0, which means ϕ (t, x, u) ∈ K by condition
(1). This means that K is positively invariant in N , that is, if x ∈ K and
ϕ (t, x, u) ∈ N then ϕ (t, x, u) ∈ K. Since N is compact and K = φ−1 (0) ⊂
int (N ), we can find δ > 0 such that U = φ−1 ([0, δ)) ⊂ int (N ). Then U is
an open neighborhood of K in M . We claim that U is positively invariant.
Indeed, for x ∈ U and u ∈ U, define
t∗ = sup {s > 0 : ϕux ([0, s]) ⊂ U } .
Suppose by contradiction that t∗ < +∞. By the compactness of N , we have

ϕ (t∗ , x, u) ∈ N . Then φ (ϕ (t∗ , x, u)) ≤ φ (x) < δ, and hence ϕ (t∗ , x, u) ∈ U .
This contradicts the definition of t∗ , since U is open. Thus t∗ must be infinite,
which means ϕux ([0, +∞)) ⊂ U . This proves that U is positively invariant.
As a consequence, K is positively invariant. Now, take x ∈ U \ K, t > 0, and
u ∈ U . If ϕ (t, x, u) ∈ K then φ (ϕ (t, x, u)) = 0 < φ (x). If ϕ (t, x, u) ∈/ K we
have
d
φ (ϕ (t, x, u)) ≤ −f (d (ϕ (t, x, u) , K)) < 0
dt
hence φ (ϕ (t, x, u)) is strictly decreasing when ϕ (t, x, u) ∈ U \ K. In any
case we have φ (ϕ (t, x, u)) < φ (x) for every x ∈ N \ K, t > 0, and u ∈ U.
Hence φ satisfies the items (1) and (2) of Definition 6.2.2. Take x ∈ N and
u ∈ U. As φ is decreasing on trajectories and φ (y) ≥ 0 for all y ∈ N , the
limit L = lim φ (ϕ (t, x, u)) exists. This limit is L = 0. Indeed, suppose by
t→+∞
contradiction that L > 0. For ε > 0, with ε < L, there is T > 0 such that L −
< φ (ϕ (t, x, u)) < L + ε for all t ≥ T Hence ϕ (t, x, u) ∈ φ−1 ((L − , L + ε))
for every t ∈ [T, +∞). Take η > 0 such that cl (B (K, η)) is a compact and
cl (B (K, η)) ∩ φ−1 ((L − ε, L + ε)) = ∅. Then d (ϕ (t, x, u) , K) > η for all
t ∈ [T, +∞). As f is strictly increasing, it follows that f (d (ϕ (t, x, u) , K)) >
f (η) for all t ∈ [T, +∞). We then have

dφϕ(t,x,u) (X (ϕ (t, x, u) , u (t))) ≤ −f (d (ϕ (t, x, u) , K)) < −f (η)
for every t ≥ T , and therefore
t
φ (ϕ (t, x, u)) − φ (ϕ (T, x, u)) ≤ − f (η) ds = − (t − T ) f (η)
T
for every t ≥ T , which is a contradiction, since f (η) > 0 and
lim φ (ϕ (t, x, u)) = L exists. Hence φ (ϕ (t, x, u)) → 0 as t → +∞, and
t→+∞
therefore φ is a smooth Lyapunov function defined on a positively invariant
neighborhood U of K. The result follows according to Theorem 6.2.1.
Example 6.2.4 Consider the control system defined in Example 6.2.3, but
2
now take the function φ (x) = x . For x ∈ Rn and u ∈ U, we have
2 2
dφx (X (x, u)) = ∇φ (x) , ux = 2u x ≤ −2 x = −f (x)
where f (r) = 2r2 is strictly increasing and f (0) = 0. Thus φ is a smooth
Lyapunov function for the uniform attractor point {0}.
Example 6.2.5 Consider the control affine system ẋ = X0 (x) + u (t) X1 (x)

on M = x = (x1 , x2 ) ∈ R2 : x ≥ 1 , with control range U = [0, 4] ⊂ R,
where X0 and X1 are given by
X0 (x1 , x2 ) = (−x2 , x1 ) ,
2
1 − x
X1 (x1 , x2 ) = 2 (x1 , x2 ) .
x
2
Let φ : M → R be the function φ (x) = x − 1. We have φ (x) = 0 if and
only if x ∈ S1 . Furthermore, for x ∈ M \ S1 and u ∈ U, we have
/ 0
2
1 − x
dφx (X (x, u)) = 2 (x1 , x2 ) , u 2 (x1 , x2 )
x
$ %
2
= 2u 1 − x
$ %
2
≤ −8 x − 1 .
Define f (r) = 8r for r ≥ 0. Then f is strictly increasing and f (0) = 0. For

x $∈ M \ S1%, we have d x, S1 = x − 1 and f (x − 1) = 8 (x − 1) ≤
2
8 x − 1 . Hence dφx (X (x, u)) ≤ −f d x, S1 , and therefore the unit
circle S1 is a uniform attractor by Theorem 6.2.2.
6.3 Conley attractor

This section goes into the Conley definitions of attractors and repellers.
We show the relationship between the Conley attractors and the uniform
attractors.
Definition 6.3.1 A Conley attractor for the control system is a set A ⊂ M

that admits a neighborhood U such that ω + (U ) = A. A Conley repeller is
a set R ⊂ M that has a neighborhood V with ω − (V ) = R. The neighborhoods
U and V are called respectively attractor neighborhood of A and repeller
neighborhood of R. We consider both the empty set and the whole state
space M as Conley attractors and Conley repellers.
Since the positive limit sets and the negative limit sets are respectively
positively invariant and negatively invariant, the Conley attractors and the
Conley repellers are respectively positively invariant sets and negatively in-
variant sets. However, they need not be invariant (see Example 6.3.1). Oth-
erwise, invariant Conley attractors and invariant Conley repellers are isolated
invariant sets. The proof for this fact is similar to the proof for Proposition
6.2.2.
Proposition 6.3.1 Every invariant Conley attractor or invariant Conley re-

peller is isolated invariant.
The following proposition shows that a finite union of Conley attractors

is a Conley attractor. The repellers have the same property.
Proposition 6.3.2 If A and B are two Conley attractors or two Conley re-
pellers then their union A ∪ B is respectively a Conley attractor or a Conley
repeller.
Proof. Let V and W be attractor neighborhoods of A and B respectively.

Then V ∪ W is a neighborhood of A ∪ B and
$ % $ % $ %
ω + (V ∪ W ) = cl O≥t
+
(V ∪ W ) = cl O≥t
+
(V ) ∪ cl O≥t
+
(W ) .
t>0 t>0
Thus A ∪ B = ω + (V ) ∪ ω + (W ) ⊂ ω + (V ∪ W ). On the other hand, take

x ∈ ω + (V ∪$ W ) and %suppose that x ∈/$ A. There
% is then T > 0 such
that x ∈
/ cl O≥T+
(V ) . Hence x ∈ cl O≥T +
(W ) . For t > 0, we have
Section 6.3 · Conley attractor 193
$ % $ % $ %
x ∈
/ cl O≥t+T
+
(V ) , hence x ∈ cl O≥t+T
+
W ⊂ cl O≥t
+
(W ) , and there-
fore x ∈ B. This proves that ω + (V ∪ W ) ⊂ A ∪ B, and therefore A ∪ B =
ω + (V ∪ W ) is a Conley attractor with attractor neighborhood V ∪ W . The
proof for the Conley repellers is analogous.
The Conley attractor is different from the attractor essentially because the
"
set inclusion x∈X ω + (x) ⊂ ω + (X) need not be an equality. Actually, due
to the properties stated in Proposition 4.2.6, a compact Conley attractor is a
uniform attractor, as in the following.
Proposition 6.3.3 Let A ⊂ M be a compact Conley attractor with attractor

neighborhood U . Then

1. There is a t > % open neighborhood U of A such that U ⊂
$ 0 and an

int (U ) and cl O≥t (U ) ⊂ U .
+
2. There is a t >$ 0 and a%compact neighborhood K of A such that K ⊂

int (U ) and cl O≥t
+
(K) ⊂ K.
3. A is a positively invariant uniform attractor.
4. A is asymptotically stable.
Proof. (1) Let V ⊂ int (U ) be a compact neighborhood of A. Then

ω (V ) ⊂ ω + (U ) = A. According to Proposition 6.2.1, $there is a%t > 0 and
+
an open neighborhood U of A such that U ⊂ V and cl O≥t

+
(U ) ⊂ U .
(2) Take U of item (1) and set K = cl (U ). Then K is a compact
neighborhood of A and
#
O≥t+
(K) = O≥t +
(cl (U )) = ϕus (cl (U ))
s≥t,u∈U
#
= cl (ϕus (U ))
s≥t,u∈U
⎛ ⎞
#
⊂ cl ⎝ ϕus (U )⎠
s≥t,u∈U
$ %
= cl O≥t
+
(U )
$ %
hence cl O≥t
+
(K) ⊂ cl (U ) = K.
(3) It remains to show that A is a uniform attractor. By the

$ item (2),
% there
is a t > 0 and a compact neighborhood K of A such that cl O≥t (K) ⊂ U .
+
$ % $ %
For x ∈ K and u ∈ U, we have ϕ (t, x, u) ∈ cl O≥t +
(K) . Since cl O≥t
+
(K)
is compact and positively invariant, the set ω + (ϕ (t, x, u)) is nonempty. By
Proposition 4.1.7, we have ω + (ϕ (t, x, u)) ⊂ ω + (x), hence ω + (x) = ∅. This
implies that J + (x) = ∅ for all x ∈ K. Moreover, Proposition 4.2.6 assures us
that J + (int (K)) ⊂ ω + (K) ⊂ A. Thus A ⊂ int (K) ⊂ Au (A), and therefore
A is a uniform attractor.
(4) It follows according to Proposition 6.2.3.
On the other hand, the negatively invariant compact uniform attractors

are Conley attractors.
Proposition 6.3.4 If K ⊂ M is a compact uniform attractor that is nega-

tively invariant then K is an invariant Conley attractor.
Proof. Since K is compact and K ⊂ int (Au (K)), there is a compact

neighborhood V of K such that V ⊂ int (Au (K)). According to Proposition
4.2.6, it follows that ω + (V ) ⊂ J + (V ) ⊂ K. On the other hand, take x ∈ K,
tn → +∞, and u ∈ U. As K is negatively invariant, ϕ (−tn , x, u) ∈ K for
all n. Hence x = ϕ (tn , ϕ (−tn , x, u) , u · (−tn )), and therefore x ∈ ω + (K). It
follows that K ⊂ ω + (K) ⊂ ω + (V ), and thus ω + (V ) = K. This means that
K is a Conley attractor. Since every Conley attractor is positively invariant,
it follows that K is invariant.
Propositions 6.3.3 and 6.3.4 state the following result.
Corollary 6.3.1 Let K ⊂ M be an invariant compact set. K is then a

Conley attractor if and only if it is a uniform attractor.
Since the compact Conley attractors are positively invariant uniform at-
tractors, Theorem 6.2.1 yields the existence of Lyapunov functions for com-
pact Conley attractors. Concerning invariant compact sets, Corollary 6.3.1
and Theorem 6.2.1 state the characterization of the Conley attractors in terms
of Lyapunov functions.
Corollary 6.3.2 Let A ⊂ M be an invariant compact set. A is then a Conley

attractor if and only if there exists a continuous Lyapunov function defined
on a positively invariant neighborhood N of A.
Section 6.3 · Conley attractor 195
Another interesting characterization of the compact Conley attractors is

obtained when all limit sets of the system are invariant. According to Proposi-
tions 6.3.1 and 6.3.3, an invariant compact Conley attractor is isolated invari-
ant and uniformly stable. These properties characterize the compact Conley
attractors provided all positive limit sets are invariant.
Proposition 6.3.5 Assume that the positive limit sets of the control system
are invariant. Then a compact set A ⊂ M is a Conley attractor if and only
if it is isolated invariant and uniformly stable.
Proof. Suppose that A is nonempty, isolated invariant, and uniformly

stable. Let U ⊂ M be an isolating neighborhood of A and take a compact
neighborhood U of A with U ⊂ U . Since A is compact and invariant, we have
ω + (A) = A. As A is uniformly stable, there is ε > 0 such that O+ (B (A, ε)) ⊂
U . This implies that ω + (B (A, ε)) ⊂ cl (O+ (B (A, ε))) ⊂ U . Note that
A = ω + (A) ⊂ ω + (B (A, ε)). On the other hand, if x ∈ ω + (B (A, ε)) then
O+ (x) ⊂ ω + (B (A, ε)), because ω + (B (A, ε)) is invariant. This means that
x ∈ U and O+ (x) ⊂ U , hence x ∈ A since U is isolating. It follows that
ω + (B (A, ε)) = A and that A is, therefore, a Conley attractor.
In general, a compact Conley attractor is an attractor, but the converse

does not hold. This is because an attractor need not be a uniform attractor.
Fig. 6.2 shows the trajectories of a dynamical system on the unit disk with two
stationary points (0, 0) and (1, 0). All points in the unit circle have {(1, 0)} as
positive and negative limit sets. Every point p = (0, 0) has ω + (p) = {(1, 0)}.
This means that {(1, 0)} is an invariant attractor. However, for any point p =
(δ, 0) with 0 < δ ≤ 1, J + (p) is the unit circle. Hence in every neighborhood
of (1, 0) there is a point p such that J + (p) {(1, 0)}. Thus {(1, 0)} is not a
uniform attractor.
Thus, for compact sets, we have the general diagram in Fig. 6.3.
The following examples illustrate nontrivial Conley attractors and Conley
repellers.
Example 6.3.1 Consider the control affine system on the unit circle S1 as
defined in Example 3.1.1. The set A = [π/2, 3π/2] is a compact Conley attrac-
tor not negatively invariant. The single set R = {0} is an invariant Conley
repeller.
Figure 6.2: Example of an attractor that is not a uniform attractor.
Figure 6.3: Relations between attraction and stability
Example 6.3.2 Consider the control affine system ẋ = u (t) X (x) on M =

2 0
R \ {0}, where X =
2
and U = [a, b], with a > 0. Project this
0 −1
x (t)
system down to the unit circle S1 by defining y (t) = . One has the
x (t)
induced control system
5 6
y = u (t) X − y T Xy y
on S1 . Fig. 6.4 illustrates the trajectories of this system. Note that the points
(1, 0) and (−1, 0) are Conley attractor points and (0, 1) and (0, −1) are Conley
repeller points. Both upper and lower semicircles are Conley attractors, while
both left and right half circles are Conley repellers.
Example 6.3.3 Go back to Example 6.2.1. The unit circle S1 is a compact

invariant Conley attractor with attractor neighborhood U = B (0, 2) \ {0}.
Section 6.4 · Global attractors 197
Figure 6.4: Trajectories of a control system on the unit circle with two at-
tractor points and two repeller points.
The unit disk D2 is also a compact invariant Conley attractor with attractor
neighborhood B (0, 2). The single set {0} is an invariant compact Conley
repeller with repeller neighborhood B (0, 1).
6.4 Global attractors

In this section we study two notions of global attractors and show how
the measure of noncompactness is used to describe the dynamics of a control
system with the presence of the global attractor.
The notion of a global attractor means an invariant compact set that at-
tracts every bounded set or, in other words, the smallest compact set having
this property. The measure of noncompactness associates numbers to sets in
such a way that every compact sets has measure zero while any noncompact
closed set has measure greater than zero. A control system with global at-
tractor satisfies the asymptotic compactness, a property characterized by the
measure of noncompactness.
For two subsets Y and Z of M , the notation dist (Y, Z) means the Haus-
dorff semidistance dist (Y, Z) = supy∈Y (inf z∈Z d (y, z)). We also need the
following formalization of the measure of noncompactness.
Definition 6.4.1 Let Y ⊂ M be a nonempty set. The measure of non-

compactness of Y is defined as
1 2
#
n
β (Y ) = inf δ > 0 : Y admits a finite cover Y ⊂ B (xi , δ) .
i=1
If Y is bounded then β (Y ) < +∞. If cl (Y ) is compact then β (Y ) = 0, and

the converse holds if M is a complete metric space. The Cantor–Kuratowski
intersection theorem assures us that M is a complete metric space if and
only if every decreasing sequence (Fn ) of nonempty bounded closed sets, with
β (Fn ) → 0, has nonempty compact intersection.
In the following we define an alternative notion of attraction for control
systems.
Definition 6.4.2 Let Y, Z ⊂ M be nonempty sets. We say that the set Y

attracts the set Z if for each > 0 there is t > 0 such that O≥t
+
(Z) ⊂ B (Y, ε);
we say that Y absorbs Z if there is t > 0 such that O≥t +
(Z) ⊂ Y .
It can be easily seen that Y attracts Z if Y absorbs Z. The following

result characterizes this notion of attraction.
Proposition 6.4.1 Let Y, Z ⊂ M be nonempty sets. The set Y attracts the

set Z if and only if dist (ϕt (Z × U ) , Y ) → 0 as t → +∞.
Proof. Suppose that Y attracts Z. For a given ε > 0 there is T > 0

such that O≥T +
(Z) ⊂ B (Y, ε). Hence ϕt (Z × U ) ⊂ O≥T +
(Z) ⊂ B (Y, ε)
for all t > T , that is, dist (ϕt (Z × U ) , Y ) < ε for all t > T . Therefore
dist (ϕt (Z × U ) , Y ) → 0 as t → +∞. As to the converse, assume that
dist (ϕt (Z × U ) , Y ) → 0 as t → +∞. For a given ε > 0, there is T > 0
such that
dist (ϕt (Z × U ) , Y ) < ε
for all t ≥ T . If x ∈ O≥T

+
(Z) then x ∈ ϕt (Z × U ) for some t ≥ T , hence
dist (x, Y ) < ε. Thus O≥T (Z) ⊂ B (Y, ε). This proves that Y attracts Z.
+
The global attractor is a region of M that attracts every bounded set in

M , as follows.
Definition 6.4.3 A set A ⊂ M is called global attractor if A is nonempty,

compact, invariant, and it attracts every bounded set in M .
There exists at most one global attractor for the control system.
Theorem 6.4.1 If the control system admits a global attractor A then it is

the unique one and coincides with the reunion of all bounded invariant sets.
Proof. Let A and A be global attractors. Since A is compact, it is

bounded. Hence A attracts A , that is, for > 0 there is t > 0 such that
O≥t+
(A ) ⊂ B (A, ). As A is invariant, we have A = O≥t +
(A ) ⊂ B (A, ).
Hence dist (A , A) < . Since is arbitrary, it follows that dist (A , A) = 0,

which means A ⊂ cl (A) = A. In the same way we show that A ⊂ A . There-

fore, A = A . Now, let {Yλ }λ∈Λ be the set of all bounded invariant sets in M .
"
Since A is bounded, we have A ⊂ λ∈Λ Yλ . On the other hand, A attracts
every Yλ . Hence, for > 0 there is t > 0 such that O≥t +
(Yλ ) ⊂ B (A, ). As Yλ
is invariant, it follows that Yλ ⊂ B (A, ) for all ε > 0. Hence dist (Yλ , A) = 0,
"
which implies Yλ ⊂ A for all λ ∈ Λ. Therefore, λ∈Λ Yλ ⊂ A.
The following examples illustrate global attractors for control systems.
Example 6.4.1 Consider the control affine system ẋ = u (t) x on M = Rn

with control range U = [−2, −1] ⊂ R. This system has the global attractor
A = {0}. In fact, let Y ⊂ Rn be a bounded set and ε > 0. There is L > 0
such that x ≤ L for all x ∈ Y . Take T > 0 such that e−T < ε/L. For
x ∈ Y , t ≥ T , and u ∈ U, we have
t

ϕ (t, x, u) = e 0 u(s) ds x ≤ e−t L < ε.
Hence O≥T+
(Y ) ⊂ B (0, ε), and therefore A = {0} is the global attractor of the
control system.
Example 6.4.2 Return to Example 6.2.1. Recall that

⎛9 k −1 ⎞
k −2u j t j ⎟
k
⎜ −2uj tj e j=1
ϕutkk ◦ · · · ◦ ϕut11 (r, θ) = ⎝ 1 − e j=1 + 2
,θ − tj ⎠
r j=1
for u1 , ..., uk ∈ U and note that the unit disk D2 is an invariant compact set.
This is the global attractor for the control system. Indeed, let Y ⊂ R2 be a
bounded set and δ > 0. Take ε > 0 such : that ε < δ. There is L > 1 such that
2 (1+ε)2
r ≤ L for all (r, θ) ∈ Y . Take t > ln LL−1 2
(1+ε)2 −1
, ϕutkk ◦ · · · ◦ ϕut11 ∈ S with
k
u1 , ..., uk ∈ U, j=1 tj ≥ t, and (r, θ) ∈ Y such that r > 1. Since uj ≥ 1, we
have
9 9
k k L2 − 1 (1 + ε)
2
r2 − 1 (1 + ε)
2
u j tj ≥ tj ≥ t > ln ≥ ln .
j=1 j=1 L2 (1 + ε)2 − 1 r2 (1 + ε)2 − 1
Then 9

k r2 − 1 (1 + ε)
2
−2uj tj < −2 ln
j=1 r2 (1 + ε)2 − 1
and therefore
k 2
−2uj tj r2 (1 + ε) − 1
−e j=1 >− .
r − 1 (1 + ε)2
2
It follows that
k
k −2uj tj k
−2uj tj e j=1
−2uj tj 1
1−e j=1 + = 1−e j=1 1− 2
r2 r
2
r2 (1 + ε) − 1 1
> 1− 1−
r −1
2
(1 + ε)
2 r2
2
(1 + ε) − 1
= 1− 2
(1 + ε)
1
= 2.
(1 + ε)
Finally,
9 k −1
k −2uj tj
−2uj tj e j=1
1−e j=1 + <1+ε
r2

and therefore ϕutkk ◦ · · · ◦ ϕut11 (r, θ) ∈ B D2 , ε . This means that O≥t
+
(Y ) ⊂
2
B D , δ , and thus D is the global attractor of the system.
2
We now introduce the notions of eventually compact, eventually bounded,

bounded dissipative, asymptotically compact, and limit compact control sys-
tems. All these concepts relate to the control systems which admit global
attractors. It should be remembered that a subset X ⊂ M is precompact if
cl (X) is compact in M .
Definition 6.4.4 The control system is called:
1. eventually compact if there is t0 > 0 such that the t-projection ϕt :

M × U → M is a compact map, that is, the image ϕt (W ) ⊂ M is
precompact for every bounded set W ⊂ M × U ;
2. eventually bounded if for each bounded set Y ⊂ M there is t > 0 such

that O≥t
+
(Y ) is a bounded set in M ;
3. bounded dissipative if there is a bounded set D ⊂ M that absorbs

every bounded set in M ;
4. point dissipative if there is a bounded set D ⊂ M that absorbs every

point x ∈ M ;
5. asymptotically compact if for all bounded sequence (xn ) in M , (un )

in U , and tn → +∞ in R, the sequence (ϕ (tn , xn , un )) in M has con-
vergent subsequence;
6. limit compact if$for every%bounded set Y ⊂ M and any ε > 0 there is

t > 0 such that β O≥t
+
(Y ) < ε.
A control system admitting the global attractor is eventually bounded as

well as bounded dissipative and asymptotically compact. This is part of our
main theorem in this section. Before proving it, we study the properties and
the relations of all type of control systems in Definition 6.4.4.
We can characterize bounded dissipative control systems as in the follow-
ing.
Proposition 6.4.2 The control system is bounded dissipative if and only if

there is a bounded set D ⊂ M which attracts every bounded set in M .
Proof. Suppose that the system is bounded dissipative. There is then

a bounded set D ⊂ M that absorbs every bounded set in M . Let Y be a
bounded set in M . There is then t > 0 such that O≥t +
(Y ) ⊂ D. Hence
O≥t (Y ) ⊂ B (D, ) for ε > 0, and therefore D attracts Y . As to the converse,
+
suppose the existence of a bounded set Y ⊂ M which attracts every bounded

set in M . Choose ε > 0 and define D = B (Y, ). For a given bounded set
Z ⊂ M , there is t > 0 such that O≥t+
(Z) ⊂ B (Y, ) = D, that is, D absorbs
Z. Therefore the control system is bounded dissipative.
We have the following relation among eventually compact, eventually

bounded, and asymptotically compact control systems.
Proposition 6.4.3 If the control system is both eventually compact and even-
tually bounded then it is asymptotically compact.
Proof. Let (xn ) be a bounded sequence in M , (un ) sequence in U ,

and tn → +∞. Since the system is eventually bounded and the set Y0 =
{xn ; n ∈ N} is bounded, there is t > 0 such that O≥t +

(Y0 ) is bounded. As
the system is eventually compact, there is t0 > 0 such that the map ϕt0
is compact. Since tn → +∞, we may assume that tn > t + t0 for ev-
ery n. Hence we can write tn = t + t0 + sn , with sn > 0. Now, define
Y = {ϕ (t + sn , xn , un ) : n ∈ N}. As Y ⊂ O≥t +
(Y0 ), it follows that Y is
bounded. Then the set ϕt0 (Y × U ) is precompact, which implies that the
subset {ϕ (tn , xn , un ) : n ∈ N} ⊂ ϕt0 (Y × U ) is precompact. Hence the se-
quence (ϕ (tn , xn , un )) has convergent subsequence, and therefore the control
system is asymptotically compact.
Asymptotically compact control systems have relevant dynamical and topo-

logical properties, as stated in the following propositions.
Proposition 6.4.4 Assume that the control system is asymptotically compact

and let Y ⊂ M be a bounded set. Then the limit set ω + (Y ) is nonempty,
compact, and attracts Y . Furthermore, ω + (Y ) is the smallest closed set in
M that attracts Y .
Proof. Let (xn ) be a sequence in Y , (un ) a sequence in U , and tn → +∞.

Since the system is asymptotically compact, we may assume that the sequence
(ϕ (tn , xn , un )) converges to some point x ∈ M . This means that x ∈ ω + (Y ),
and therefore ω + (Y ) is nonempty. Now, let (xn ) be a sequence in ω + (Y ). For
each n ∈ N, take zn ∈ O≥n +
(Y ) ∩ B (xn , 1/n). We can write zn = ϕ (tn , yn , un )
with tn ≥ n, yn ∈ Y , and un ∈ U. Since Y is a bounded set and the system
is asymptotically compact, the sequence (ϕ (tn , yn , un )) has a subsequence
(ϕ (tnk , ynk , unk )) that converges to some point x ∈ M . This means that
x ∈ ω + (Y ). Moreover, for each k we have
d (xnk , x) ≤ d (xnk , ϕ (tnk , ynk , unk )) + d (ϕ (tnk , ynk , unk ) , x)

1
< + d (ϕ (tnk , ynk , unk ) , x)
nk
hence xnk → x, and therefore any sequence in ω + (Y ) has convergent subse-

quence. This proves that ω + (Y ) is compact. Finally, we show that ω + (Y )
attracts Y and is the smallest closed set in M with this property. Suppose by
contradiction that, for a given ε > 0, O≥t+
(Y ) B (ω + (Y ) , ε) for all t > 0.
Then, for each n ∈ N, there are tn ≥ n, yn ∈ Y , and un ∈ U such that

d ϕ (tn , yn , un ) , ω + (Y ) ≥ ε.
As tn → +∞, we may assume that the sequence (ϕ (tn , yn , un )) converges

to some point x ∈ M , by asymptotic compactness. Hence x ∈ ω + (Y ).
On the other hand, we have d (ϕ (tn , yn , un ) , ω + (Y )) ≥ for all n, hence
d (x, ω + (Y )) ≥ , which contradicts x ∈ ω + (Y ). Therefore ω + (Y ) attracts
Y . Now, let K ⊂ M be a closed set that attracts Y . Suppose by contradic-
tion that ω + (Y ) K. There is then x ∈ ω + (Y ) such that x ∈ / K. Since
K is closed, we have d (x, K) = δ > 0. As K attracts Y , there is t > 0 such
that O≥t +
(Y ) ⊂ B (K, δ/2). Then for all s ≥ t, y ∈ Y , and u ∈ U , we have
d (ϕ (t, y, u) , K) < δ/2. On the other hand, since x ∈ ω + (Y ), there are se-
quences (xn ) in Y , (un ) in U , and tn → +∞ with ϕ (tn , yn , un ) → x. We may
then assume that tn > t and d (ϕ (tn , yn , un ) , K) < δ/2 for all n. It follows
that d (x, K) ≤ δ/2, which contradicts d (x, K) = δ. Therefore ω + (Y ) ⊂ K
and the proposition is proved.
Proposition 6.4.5 Assume that the control system is asymptotically com-

pact. If Y ⊂ M is a bounded set and ω + (Y ) attracts a connected set C
containing Y then ω + (Y ) is connected.
Proof. Suppose by contradiction that ω + (Y ) is disconnected. Then

ω (Y ) = K1 ∪ K2 , where K1 and K2 are two nonempty closed sets with
+
K1 ∩ K2 = ∅. By Proposition 6.4.4, ω + (Y ) is compact, and then both K1

and K2 are compact. Hence d (K1 , K2 ) = δ > 0. Since ω + (Y ) attracts C,
there is t > 0 such that

O≥t
+
(C) ⊂ B ω + (Y ) , δ/2 = B (K1 , δ/2) ∪ B (K2 , δ/2) .
As d (K1 , K2 ) = δ, the sets B (K1 , δ/2) and B (K2 , δ/2) are disjoint. Since
O≥t
+
(C) is connected, it follows that either O≥t +
(C) ⊂ B (K1 , δ/2) or
O≥t (C) ⊂ B (K2 , δ/2). Suppose that O≥t (C) ⊂ B (K1 , δ/2). We then have
+ +
$ % $ %
K2 ⊂ ω + (Y ) ⊂ cl O≥t
+
(Y ) ⊂ cl O≥t
+
(C) ⊂ cl (B (K1 , δ/2))
hence d (K1 , K2 ) ≤ 2δ , which contradicts d (K1 , K2 ) = δ. Therefore ω + (Y ) is

connected.
Limit compact control systems seem quite a bit like asymptotically com-
pact control systems, as in the following.
Proposition 6.4.6 If the control system is asymptotically compact then it is

limit compact. The converse holds if M is a complete metric space.
Proof. Suppose that the system is asymptotically compact and let Y ⊂
M be a nonempty bounded set. According to Proposition 6.4.4, ω + (Y ) is
nonempty, compact, and attracts Y . For a given δ > 0, the open covering
" "
n
ω + (Y ) ⊂ B (x, δ) admits a finite subcovering ω + (Y ) ⊂ B (xi , δ).
x∈ω + (Y ) i=1
"
n
Take ε > 0 such that B (ω + (Y ) , ε) ⊂ B (xi , δ). Since ω + (Y ) attracts
i=1
"
n
Y , there is t > 0 such that O≥t +
(Y ) ⊂ B (ω + (Y ) , ε) ⊂ B (xi , δ), hence
$ % i=1
β O≥t+
(Y ) < δ. Thus the system is limit compact. Conversely, suppose
that M is complete and the system is limit compact. Let (xn ) be a bounded
sequence in M , (un ) sequence in U , and tn → +∞. Set Y = {xn : n ∈ N}.
For
$ each
$ ε > 0,%% we can use the limit compactness to find tε > 0 such that
β cl O≥tε (Y ) < ε. For each n ∈ N, set Fn = cl {ϕ (tn , xn , un ) : n ≥ n}.
+
$ a given ε% > 0, there is nε such that n ≥ nε implies $

For tn >$ tε . Hence %%Fn ⊂
cl O≥t+
ε
(Y ) for all n ≥ n ε , and therefore β (F n ) < β cl O +
≥tε (Y ) <ε
for all n ≥ nε . This means that β (Fn ) → 0. Moreover, if n ≥ n then
Fn ⊂ Fn , and then (Fn ) is a decreasing sequence of closed sets. By the

Cantor-Kuratowski theorem, the intersection n∈N Fn is nonempty, and we

can then get x ∈ n∈N Fn . Now, for each k ∈ N, there is nk ≥ k such that
ϕ (tnk , xnk , unk ) ∈ B (x, 1/k), because x ∈ Fk . Hence (ϕ (tnk , xnk , unk )) is a
subsequence of (ϕ (tn , xn , un )) such that ϕ (tnk , xnk , unk ) → x. Therefore the
system is asymptotically compact.
We now present the main result of this section.

Theorem 6.4.2 If the control system has the global attractor A, then the
system is eventually bounded, bounded dissipative, asymptotically compact,
and
#
A = ω + (Y )
Y ∈B
!
x ∈ M : there is a bounded sequence (xn ) in M , (un ) in U ,
=
and tn → +∞ such that ϕ (tn , xn , un ) → x
where B is the collection of all bounded sets in M . The converse holds if the
positive limit set ω + (Y ) is invariant for every bounded set Y ⊂ M .
Proof. By Proposition 6.4.2, the system is bounded dissipative. Let Y

be a bounded set in M . For > 0, there is t > 0 such that O≥t +
(Y ) ⊂
B (A, ). Since A is bounded, B (A, ) is also bounded, and hence O≥t (Y ) is
+
bounded. This proves that the system is eventually bounded. Now let (xn )
be a bounded sequence in M , (un ) a sequence in U , and tn → +∞. Take
ε > 0 such that cl (B (A, )) is compact. Since the set Y = {xn : n ∈ N} is
bounded, there is t > 0 such that O≥t +
(Y ) ⊂ B (A, ). For tn > t we have
ϕ (tn , xn , un ) ∈ cl (B (A, )), and hence the sequence (ϕ (tn , xn , un )) admits
a convergent subsequence (in fact a subsequence that converges to a point
in A, see Problem 6.5(5)). Thus the system is asymptotically compact. We
" +
now show that A = ω (Y ). Indeed, if Y ∈ B then ω + (Y ) is the small-
Y ∈B
est closed set in M that attracts Y , according to Proposition 6.4.4. Hence
ω + (Y ) ⊂ A for every Y ∈ B. Since A is invariant and closed, we have
" +
A = ω + (A). As A ∈ B, it follows that A = ω (Y ). For the converse,
Y ∈B
we assume that ω + (Y ) is invariant for every bounded set Y ⊂ M and the
system is asymptotically compact and bounded dissipative (and eventually
bounded). Let B be the collection of all nonempty bounded sets in M and
" +
define A = ω (Y ). By Proposition 6.4.4, ω + (Y ) is nonempty, compact,
Y ∈B
and attracts Y whenever Y ∈ B. Hence A is nonempty and attracts every
bounded set. Moreover A is invariant, because each ω + (Y ) is invariant by
hypothesis. Since the system is bounded dissipative, there is a bounded set
D ⊂ X which absorbs Y , whenever Y ∈ B. It follows that ω + (Y ) ⊂ cl (D)
for all Y ∈ B. Hence A ⊂ cl (D), and ω + (A) ⊂ ω + (D). Since A is invari-
ant, we have A = ω + (A) ⊂ ω + (D). On the other hand, as D is bounded,
ω + (D) ⊂ A, and therefore A = ω + (D). By Proposition 6.4.4 again, A is
compact, and thus A is the global attractor for the control system.
In some cases, we may consider point dissipative control systems instead

of bounded dissipative control systems, as follows:
Proposition 6.4.7 Assume that the limit set ω + (Y ) is invariant for

every bounded set Y ⊂ M . If the control system is eventually compact, even-
tually bounded, and point dissipative, then it has global attractor.
Proof. Since the system is point dissipative, there is a bounded set D0 ⊂

M that absorbs every point x ∈ M . For a given > 0 we define the bounded
set D1 = B (D0 , ). As the system is eventually bounded, there is t > 0 such
that the set D = O≥t +
(D1 ) is bounded. This set D absorbs every bounded
subset Y of M . In fact, since the action is eventually compact, there is t0 > 0

such that cl ϕt0 (Y × U ) is compact for every bounded set Y ⊂ M . For each

x ∈ cl ϕt0 (Y × U ) , there is tx > 0 such that
O≥t
+
x
(x) ⊂ D0 ⊂ D1 .
Hence ϕtx ({x} × U ) ⊂ D1 , and thus there is δ x > 0 such that
ϕtx (B (x, δ x ) × U ) ⊂ D1 .
It follows that

O≥t+t
+
x
(B (x, δ x )) = O≥t
+
ϕtx (B (x, δ x ) × U ) ⊂ O≥t
+
(D1 ) = D.
"
Now the open covering cl ϕt0 (Y × U ) ⊂ B (x, δ x ) admits a finite
x∈cl(ϕt0 (Y ))
"n
subcovering cl ϕt0 (Y × U ) ⊂ j=1 B xj , δ xj . If τ = max {tx1 , ..., txn }, we
have
⎛ ⎞
#
n

O+ cl ϕt (Y × U ) ⊂ O+ ⎝ B xj , δ xj ⎠
≥t+τ 0 ≥t+τ
j=1
#
n

= O≥t+τ
+
B xj , δ xj ⊂ D.
j=1
It follows that

O≥t+τ
+
+t0 (Y ) = O≥t+τ ϕt0 (Y × U ) ⊂ D
+
hence D absorbs Y . Thus the control system is bounded dissipative. Now,

according to Proposition 6.4.3, the system is also asymptotically compact.
Finally, the system has global attractor according to Theorem 6.4.2.
The control systems studied in Examples 6.4.1, 6.4.2, and 6.2.1 are asymp-
totically compact, because they have global attractors.
We shall now discuss what type of attractor the global attractor is. In the
following, we show that the global attractor is a compact invariant uniform
attractor, and therefore it is a Conley attractor according to Proposition 6.3.4.
Proposition 6.4.8 If A ⊂ M is the global attractor of the control system

then A is an invariant compact uniform attractor with Au (A) = M .
Proof. According to Theorem 6.4.2, the system is asymptotically com-

pact. According to Proposition 6.4.4, it follows that the positive limit sets
are nonempty. This implies that J + (x) = ∅ for every x ∈ M . Suppose by
contradiction that J + (x) A for some x ∈ M and take y ∈ J + (x) \ A.
Since A is compact, we have d (y, A) = δ > 0. As y ∈ J + (x), there are
sequences (tn ) in R, (xn ) in M , and (un ) in U with tn → +∞, xn → x,
and ϕ (tn , xn , un ) → y. For ε > 0, we may assume that xn ∈ B (x, ε). Set
Y = {xn : n ∈ N}. Since Y is bounded, A attracts Y , hence there is t > 0
such that O≥t+
(Y ) ⊂ B (A, δ/2). As tn → +∞, we may assume that tn > t for
all n. Then d (ϕ (tn , xn , un ) , A) < δ/2, and therefore d (y, A) ≤ δ/2, contra-
dicting d (y, A) = δ. Thus we have J + (x) ⊂ A for every x ∈ M , and therefore
A is a uniform attractor with M = Au (A).
We say that a set A ⊂ M is a global uniform attractor if it is a compact

uniform attractor and its region of uniform attraction Au (A) coincides with
the whole state space M . Unlike the definition of the global attractor, the
invariance is not required for a global uniform attractor. This is only to yield
generality, since noninvariant uniform attractors with Au (A) = M are very
common. Otherwise, the control system admits at most one invariant global
uniform attractor (see Problem 6.5(6)).
According to Theorem 6.4.2 and Proposition 6.4.8, the control system
admits an invariant global uniform attractor and is asymptotically compact
whenever it has a global attractor. The converse holds if the manifold M is
complete or the system is eventually compact, as in the following:
Proposition 6.4.9 Assume that A ⊂ M is the invariant global uniform at-

tractor of the control system. A is then the global attractor if any one of the
following conditions holds:
1. the control system is asymptotically compact and eventually compact, or
2. the control system is asymptotically compact and the manifold M is a

complete metric space.
Proof. Suppose that the system is asymptotically compact and eventually

compact. Let Y ⊂ M be a bounded set. There is then t0 > 0 such that

cl ϕt0 (Y × U ) is compact. We then have

ω + ϕt0 (Y × U ) ⊂ ω + cl ϕt0 (Y × U ) ⊂ J + cl ϕt0 (Y × U ) ⊂ A
and
$ %
ω + ϕt0 (Y × U ) = cl O≥t
+
ϕt0 (Y × U )
t>0
$ %
= cl O≥t+t
+
0
(Y ) = ω + (Y ) .
t>0
Hence ω + (Y ) ⊂ A. By Proposition 6.4.4, ω + (Y ) attracts Y , hence A attracts

Y . Thus A attracts every bounded set of M . Since A is invariant and compact,
it follows that A is the global attractor. Now suppose that the system is
asymptotically compact and M is complete. This means that every bounded
and closed set in M is compact. If Y ⊂ M is a bounded set then cl (Y ) is
compact, and hence
ω + (Y ) ⊂ ω + (cl (Y )) ⊂ J + (cl (Y )) ⊂ A.
The proof follows similarly to the first part.
From Propositions 6.4.3 and 6.4.9, we have the following result.
Corollary 6.4.1 Assume that the control system is eventually bounded and
eventually compact. A set A ⊂ M is then the global attractor if and only if it
is the invariant global uniform attractor.
Example 6.4.3 Consider the vector field X on R3 \ {x3 = 1} given by

⎛ ⎞T
2x1 ((x3 −1)2 −x21 −x22 )
−2x −
(x3 − 1) − x21 − x22 ⎜ ⎟
2 2 (x3 −1) 2
⎜ 2x2 ((x3 −1)2 −x21 −x22 ) ⎟
X (x1 , x2 , x3 ) = ⎜ 2x − ⎟ .
(x3 − 1) + x21 + x22 ⎝ 1 ⎠
2 (x3 −1)2
2 2
4x1 +4x2
x3 −1
For (x1 , x2 , x3 ) in the unit sphere S2 , we have

4x1 x23 4x2 x23
X (x1 , x2 , x3 ) = −2x2 − , 2x1 − 2
, 4x3 + 4x3 .
x3 − 1 x3 − 1
Consider the control affine system on M = S2 \ {(0, 0, 1)} given by ẋ =

u (t) X (x), with control range U = [1, 10]. We have X (0, 0, −1) = (0, 0, 0),
hence the point p = (0, 0, −1) is a critical point of the system. For x =
(x1 , x2 , 0) in the unit circle S1 ⊂ S2 , we have uX (x) = (−2ux2 , 2ux1 , 0),
hence the trajectories through x are periodic, coinciding with S1 . To see the

other trajectories of the system, we take x ∈ M \ S1 ∪ {p} and consider
spherical coordinates x1 = sin φ cos θ, x2 = sin φ sin θ, x3 = cos φ. For u ∈ U,
1
the variables φ and θ satisfy cos φ = and θ = 2ut + β, with α, β
αe4ut − 1
constant, α = 0. If α < 0, the trajectory through x moves on spirals, coming
near S1 as t → +∞ and closing to p as t → −∞. If α > 0, the trajectory
through x moves on spirals coming near S1 as t → +∞, but ω − (x) is empty
in this case. All the trajectories of the system are illustrated in Fig. 6.5. Thus
the circle S1 is an invariant compact uniform attractor not global, because it
does not attract the point p. The global attractor is the lower hemisphere.
Figure 6.5: Trajectories of a control system on the bored sphere.
To finish this section, we discuss the existence of the global attractor from
the flow point of view. Consider the maximum metric on the product space
M × U given by
d ((x, u) , (y, v)) = max {d (x, y) , d (u, v)} .
We have a strong connection between the global attractor of the control system
and the global attractor of the control flow, as in the following:
Proposition 6.4.10 A set A ⊂ M is the global attractor of the control system
if and only if A × U is the global attractor of the control flow.
Proof. Suppose that A is the global attractor for the control system.
Then A × U is compact and Φ-invariant, because A is compact and invariant.
Let W ⊂ M ×U be a bounded set in M ×U and ε > 0. The projection π (W ) is

a bounded set in M . Hence there is T > 0 such that O≥T
+
(π (W )) ⊂ B (A, ε).
For (x, u) ∈ W and t ≥ T , we have
Φ (t, x, u) = (ϕ (t, x, u) , u · t)
∈ O≥T
+
(x) × U ⊂ O≥T
+
(π (W )) × U
⊂ B (A, ε) × U
hence Φ ([T, +∞) × W ) ⊂ B (A, ε) × U. This means that A × U attracts any
bounded set in M × U , and therefore A × U is the global attractor of the
control flow. As to the converse, suppose that A × U is the global attractor
of the control flow. Since A × U is compact and invariant, it follows that A
is compact and invariant. Let Y ⊂ M be a bounded set and ε > 0. Then
Y × U is a bounded set in M × U , and hence there is T > 0 such that
Φ ([T, +∞) × Y × U ) ⊂ B (A, ε) × U . It follows that
O≥T
+
(Y ) ⊂ π (Φ ([T, +∞) × Y × U )) ⊂ B (A, ε)
and thus A attracts Y . This finishes the proof.
6.5 Problems
1. Let X, Y ⊂ M and prove the following:
(a) X ⊂ Aw (Y ) (X is weakly attracted to Y ) if and only if there

are sequences (tn ) in R+ and (un ) in U such that tn → +∞ and

d ϕutnn (X) , Y → 0.

(b) X ⊂ A (Y ) (X is attracted to Y ) if and only if d ϕutnn (X) , Y → 0
for all sequences (tn ) in R+ and (un ) in U with tn → +∞.
2. Consider the control affine system defined in Example 6.3.2. Check that
the single sets R(1,0) = {(1, 0)}, R(−1,0) = {(−1, 0)}, R(0,1) = {(0, 1)},
R(0,−1) = {(0, −1)} are the maximal asymptotic transitive sets of the
system, and

Aa R(1,0) = {(x1 , x2 ) : x1 > 0} ,

Aa R(−1,0) = {(x1 , x2 ) : x1 < 0} ,

Aa R(0,1) = {(0, 1)} ,

Aa R(0,−1) = {(0, −1)} .
3. Consider the control system in Example 6.2.1.
(a) Check that both sets {0} and S1 are maximal asymptotic transitive

sets with Aa ({0}) = {0} and Aa S1 = R2 \ {0}. Conclude that
they do not relate to each other.
(b) Describe all negative prolongational limit sets.
(c) Check that the function

0 if r ≤ 1
φ (r, θ) =
r − 1 if r > 1
is a continuous Lyapunov function for the unit disk D2 .

(d) Find a Lyapunov function for the unit circle S1 .
4. Assume that the control system is asymptotically compact and let A ⊂

M be a compact Conley attractor. Prove that there is a neighborhood
+
V of A such that for each δ > 0 there is a T > 0 with cl O>T (V ) ⊂
B (A, δ). Prove a similar result for a Conley repeller.
5. Assume that the control system has the global attractor A and let (xn )
be a bounded sequence in M , (un ) a sequence in U , and tn → +∞.
Show that a subsequence of (ϕ (tn , xn , un )) converges to a point x ∈ A.
(Hint: for each k ∈ N, take sk > k such that O≥s +
k
(Y ) ⊂ B (A, 1/k),
where Y = {xn }, and then use the compactness of A.)
6. Prove that a control system has at most one global invariant uniform
attractor.
7. Assume that the manifold M is compact and let A ⊂ M ×U be a Conley
attractor for the control flow. What is the relation between A and the
projection-lift L (π (A))?
8. Verify that the Conley attractors as well as the Conley repellers are in-
variant if the control system satisfies the translation hypothesis (Prob-
lem 4.4(8)).
9. Check that Awu (K) = J − (K) and Rwu (K) = J + (X) for any compact
set K ⊂ M .
10. Assume that the control system satisfies the translation hypothesis. For
any set X ⊂ M , prove that all regions Aw (X), A (X), Au (X), Awu (X)
are invariant.
11. Assume that the control system satisfies the translation hypothesis and
let K ⊂ M be a compact set. Prove that A (K) or Au (K) is an open
neighborhood of K if K is respectively an attractor or a uniform attrac-
tor of the system.
12. Assume that the control system satisfies the translation hypothesis.
Prove that a compact and invariant set A ⊂ M is a Conley attractor of
the control system if and only if there exists a compact neighborhood
V of A such that O− (x) V for all x ∈ V \ A. (Hint: adapt the proof
for Theorem A.3.2 in Appendix A)
13. Let K ⊂ M be a compact set. Prove that K is equistable if and only if

it is uniformly stable.
14. Let U ⊂ M be a compact set such that Ot+ (U ) ⊂ int (U ) for some t > 0.
Show that ω + (U ) is a Conley attractor contained in the interior of U .
15. (Conjecture) Assume that the control system satisfies the translation hy-
pothesis and every trajectory intersects some given bounded set B ⊂ M .
Then the set of all bounded trajectories is a compact Conley attractor.

The Conley attractor is also known as a strong attractor, the name originally
given to it by Conley himself in [36]. But it is now known as the “Conley
attractor” in his memory.
Control affine systems with solutions depending continuously on inputs
may be interpreted as nonautonomous dynamical systems. In this setting,
Cheban, Kloeden, and Schmalfu [29] related the concept of a global attractor
to pullback and forward attractors. Cheban [28] studied the existence of the
global attractor for special classes of control systems.
In classical dynamical systems, the compact and positively invariant uni-
form attractors correspond to the asymptotically (Lyapunov) stable sets ([14,
Chapter V, Section 1]). Proposition 6.2.3 shows that every compact and pos-
itively invariant uniform attractor of a control affine system is asymptotically
stable. The converse was proved in the general set-up of semigroup actions by
assuming the translation hypothesis ([26, Theorem 3.6]). Bacciotti et al. [6, 7]
discussed this correspondence between uniform attractors and asymptotically
stable sets in the setting of control systems viewed as dynamical polysystems.
By assuming that the reachable map of the system is upper semicontinuous,

an asymptotically stable compact set is a uniform attractor ([7, Proposition
17]).
The paper [7] also introduced the notion of a uniformly asymptotically
stable set and studied the description of it in terms of the Lyapunov func-
tion. The authors considered higher uniformity described by the higher-order
prolongational extensions of the reachable map, and the former definition
of a Lyapunov function was modified with a further condition of preserving
limit sets. Basically, in this paradigm, the property (3) of a Lyapunov func-
tion in Definition 6.2.2 of this book is changed by the following one: for any
y, z ∈ ω + (x) one has φ (y) = φ (z), for every x ∈ N (see [7, Definition 2.2]).
Asymptotic stability for differential inclusions was related to Lyapunov
functions in the works [30, 75, 110].
Theorem 6.2.2 shows that the existence of a smooth Lyapunov function
implies the uniform attraction. The converse is an open problem.
We refer to Raminelli and Souza [83] for the general theory of global
attractors for semigroup actions.
Chapter 7
Chain transitivity
In this chapter we study a more general aspect of transitivity defined by

means of trajectories with jumps. Chain transitivity is a concept of the Conley
theory that generalizes the asymptotic transitivity and relates to the Conley
attractors. This is an equivalence relation whose equivalence classes are max-
imal regions of chain transitivity. The asymptotic transitive sets and the
prolongational control sets are then chain transitive. The concept of a chain
control set is a special case of a chain transitive set. Section 7.1 contains the
basic definitions and properties of chain transitivity. We show a crucial result
that the compact limit sets of points are chain transitive. In Section 7.2 we
reproduce the attractor-repeller pair paradigm, which plays an important role
in Conley’s theorems. Section 7.3 is dedicated to the Conley main theorem of
chain recurrence that describes the chain recurrence set as the intersection of
all attractor-repeller pairs of the system.
7.1 Chain transitive sets

This section contains the basic definitions and properties of chain transitiv-
ity and chain controllability. We show that the concept of a chain transitive
set extends the concepts of asymptotic and prolongational transitive sets.
Definition 7.1.1 Let x, y ∈ M and , T > 0. An (, T )-chain from x to

y consists of sequences x0 = x, x1 , ..., xk = y in M , u0 , ..., uk−1 in U , and
215
216 Chapter 7 · Chain transitivity
t0 , ..., tk−1 ≥ T such that
d (ϕ (ti , xi , ui ) , xi+1 ) <
for every i = 0, ..., k − 1.
A chain is basically a trajectory with gauged jumps (Fig. 7.1).
Figure 7.1: Illustration of chain from a point x to a point y.
Definition 7.1.2 For a given nonempty subset X ⊂ M and , T > 0 we

define the sets
Ω+ (X, , T ) = {y ∈ M : there is x ∈ X and an (, T ) -chain from x to y} ,

−
Ω (X, , T ) = {y ∈ M : there is x ∈ X and an (, T ) -chain from y to x} .
The positive chain limit set of X is defined by

Ω+ (X) = Ω+ (X, , T )
,T >0
and the negative chain limit set of X is defined as

Ω− (X) = Ω− (X, , T ) .
,T >0
For any set X ⊂ M , it is easily seen that

!
+ y ∈ M : for a given pair , T > 0 there is some x ∈ X
Ω (X) = ,
and an (, T ) -chain from x to y
!
y ∈ M : for a given pair , T > 0 there is some x ∈ X
Ω− (X) = .
and an (, T ) -chain from y to x
Section 7.1 · Chain transitive sets 217
For any point x ∈ M , we have
Ω+ (x) = {y ∈ M : there is an (ε, T ) -chain from x to y for all ε, T > 0} ,

−
Ω (x) = {y ∈ M : there is an (ε, T ) -chain from y to x for all ε, T > 0} .
Note that x ∈ Ω+ (y) if and only if y ∈ Ω− (x). Recall that this property is
also satisfied by the prolongational limit sets, but not by the limit sets.
We often use concatenation of chains. If x ∈ Ω+ (y, , T ) and y ∈ Ω+ (z, , T ),
there is an (, T )-chain from x to y consisting of sequences x0 = x, x1 , ..., xk =
y ∈ M , u0 , ..., uk−1 ∈ U, and t0 , ..., tk−1 ≥ T ; and there is an (, T )-chain
from y to z consisting of sequences y0 = y, y1 , ..., yl = z ∈ M , v0 , ..., vl−1 ∈ U,
and s0 , ..., sl−1 ≥ T . By denoting xk+i = yi for i = 0, ..., l, uk+i = vi , and
tk+i = si for i = 0, ..., l − 1, the sequences x0 = x, x1 , ..., xk+l = z ∈ M ,
u0 , ..., uk+l−1 ∈ U, and t0 , ..., tk+l−1 ≥ T form an (, T )-chain from x to z.
Hence, x ∈ Ω+ (z, , T ).
Definition 7.1.3 A set X ⊂ M is said to be chain transitive if X ⊂

Ω+ (x) for all x ∈ X. A point x ∈ M is called chain recurrent (or chain
transitive) if x ∈ Ω+ (x). The set of all chain recurrent points is called the
chain recurrence set and is denoted by C.
We say that a set X ⊂ M is chain recurrent if every point in X is chain

recurrent. If the whole state space M is chain recurrent itself, that is C = M ,
we say that the control system is chain recurrent. In the following we define
the notion of chain controllability.
Definition 7.1.4 The control system is said to be chain controllable from

x ∈ M if Ω+ (x) = M ; the system is called chain controllable if Ω+ (x) = M
for every x ∈ M .
If the system is chain controllable from x, then x can be attainable to any

other point y in M by trajectories with arbitrary jumps and times. In a chain
transitive set X, any point x can be attainable to another y by trajectories
with arbitrary jumps and times. It can be easily seen that the control system
is chain controllable if and only if M is a chain transitive set itself.
In the compact case, we can simplify the definition of a chain transitive
set, as follows:
Proposition 7.1.1 Let C ⊂ M be a precompact set. C is then chain transi-

tive if and only if C ⊂ Ω+ (x, ε, 1) for all x ∈ E and ε > 0.
Proof. Suppose that C ⊂ Ω+ (x, ε, 1) for every x ∈ C and ε > 0. Let

x, y ∈ C, ε > 0, and n ∈ N. Since cl (C) is compact, we can use Proposi-
ε
tion 1.3.4 to obtain δ > 0 with δ < such that d (ϕ (t, z, u) , ϕ (t, w, u)) <
2
ε
, for all t ∈ [0, 4n − 2], u ∈ U , and z, w ∈ C with d (z, w) < δ.
4 (n − 1)
Consider a (δ, 1)-chain from x to y given by x1 = x, x2 , ..., xk+1 = y in M ,
t1 , ..., tk+1 ≥ 1, and u1 , ..., uk+1 ∈ U . We then have d (ϕ (ti , xi , ui ) , xi+1 ) < δ
for all i = 1, ..., k. By introducing artificial jumps, we may assume that
ti ∈ [1, 2] for all i = 1, ..., k + 1. Now, by the Euclidean algorithm, there are
q, r ∈ N such that k = qn + r with r < n. We then define the following
sequence of points and times:

n
z1 = x1 = x τ1 = ti
i=1
z2 = xn+1 2n
z3 = x2n+1 τ2 = ti
i=n+1
.. ..
. .
zq = x(q−1)n+1
k
zq+1 = y τq = ti
i=(q−1)n+1
and consider the sequence of control functions vj ∈ U such that they coincide
on the respective intervals with the controls u1 , ..., uk . These sequences form
an (ε, n)-chain from x to y. Indeed, since ti ∈ [1, 2], we have τ j ≥ n for all
j = 1, ..., q. For j = 1, ..., q − 1, we have τ j ≤ 2 (2n − 1) = 4n − 2, hence
d (ϕ (τ j , zj , vj ) , zj+1 )
⎛ ⎛ ⎞ ⎞
jn
= ⎝
d ϕ ⎝ ti , x(j−1)n+1 , vj , xjn+1 ⎠
⎠
i=(j−1)n+1
⎛ ⎛ ⎞ ⎛ ⎞⎞
jn
jn

≤ d ⎝ϕ ⎝t(j−1)n+1 + ti , x(j−1)n+1 , vj ⎠ , ϕ ⎝ ti , x(j−1)n+2 , vj ⎠⎠
i=(j−1)n+2 i=(j−1)n+2
⎛ ⎛ ⎞ ⎛ ⎞⎞
jn
jn

+d ⎝ϕ ⎝t(j−1)n+2 + ti , x(j−1)n+2 , vj ⎠ , ϕ ⎝ ti , x(j−1)n+3 , vj ⎠⎠
i=(j−1)n+3 i=(j−1)n+3
+ · · · + d (ϕ (tjn−1 + tjn , xjn−1 , vj ) , ϕ (tjn , xjn−1 , vj )) +

+d (ϕ (tjn , xjn , vj ) , xjn+1 )
ε
< (n − 1) +δ
4 (n − 1)
ε ε
< + < ε.
4 2
Similarly, for τ q we have

⎛ ⎛ ⎞ ⎞

k
d (ϕ (τ q , zq , vq ) , zq+1 ) = d ⎝ϕ ⎝ ti , x(q−1)n+1 , vq ⎠ , y ⎠
i=(q−1)n+1
ε
≤ (n − 1 + r) +δ
4 (n − 1)
ε ε
< (2n − 2) + = ε.
4 (n − 1) 2
Since x, y ∈ C, ε > 0, and n ∈ N are arbitrary, we have C ⊂ Ω+ (x) for every
x ∈ C. Hence C is a chain transitive set. The converse is clear.
It can be easily seen that a chain transitive set is a chain recurrent set. The
converse does not hold in general. However, a compact and connected chain
recurrent set is chain transitive. This condition is included in the following
criteria for chain transitivity:
Proposition 7.1.2 A compact and connected set C ⊂ M is chain transitive

if one of the following conditions holds:
1. C is chain recurrent.
2. The set of chain recurrent points of C is dense in C and O>T

+
(x) ∩
B (C, δ) = ∅ for all x ∈ C and δ, T > 0.
3. The set of chain recurrent points of C is dense in C and C is positively

invariant.
Proof. Suppose that C is chain recurrent. Let x, y ∈ C and ε, T >

ε
0. Since C is compact and connected, we can take a finite sequence of -
4
balls B (x1 , ε/4) , ..., B (xk , ε/4) centered at points x1 , ..., xk ∈ C such that
B (xi , ε/4)∩B (xi+1 , ε/4)∩N = ∅, for every i = 1, ..., k−1, with x ∈ B (x1 , ε/4)
and y ∈ B (xk , ε/4). For each i = 1, ..., k, take an (ε/2, T )-chain from xi
to xi formed by sequences of points xi0 = xi , xi1 , ..., xipi = xi ∈ M , times
ti0 , ..., tipi −1 ≥ T , and control functions ui0 , ..., uipi −1 ∈ U . For i = 1, ..., k − 1,
we have

d ϕ tipi −1 , xipi −1 , uipi −1 , xi+1

≤ d ϕ tipi −1 , xipi −1 , uipi −1 , xi + d (xi , xi+1 )
ε ε
< + =ε
2 2
and for i = k we have

d ϕ tkpk −1 , xkpk −1 , ukpk −1 , y

≤ d ϕ tkpk −1 , xkpk −1 , ukpk −1 , xk + d (xk , y)
ε ε
< + < ε.
2 4
Hence we obtain an (ε, T )-chain from x1 to y, by concatenations of chains.
Finally, by taking an (ε/2, T )-chain from x to x, we similarly construct an
(ε, T )-chain from x to x1 , and then we obtain an (ε, T )-chain from x to y.
Therefore, C is chain transitive.
We now suppose that the set of chain recurrent points of C is dense in
C and O>T +
(x) ∩ B (C, δ) = ∅ for all x ∈ C and δ, T > 0. Let x, y ∈ C
and ε, T > 0. Since O>T +
(x) ∩ B (C, /4) = ∅, there is t0 > T , x0 ∈ C, and
u0 ∈ U such that ϕ (t0 , x, u0 ) ∈ B (x0 , /4). There is then a finite sequence of

-balls B (x1 , /4) , ..., B (xk , /4) centered at points x1 , ..., xk ∈ C such that
4
B (xi , /4)∩B (xi+1 , /4)∩C = ∅, for every i = 1, ..., k−1, with x0 ∈ B (x1 , /4)
and y ∈ B (xk , /4). As the chain recurrent points of C are dense in C,
there is a chain recurrent point yi ∈ B (xi , /4) ∩ B (xi+1 , /4) ∩ C for every
i = 1, ..., k − 1. For each i = 1, ..., k − 1, take an (ε/4, T )-chain from yi
to yi formed by sequences of points y0i = yi , y1i , ..., ypi i = yi ∈ M , times
ti0 , ..., tipi −1 ≥ T , and control functions ui0 , ..., uipi −1 ∈ U. We then have
$ $ % %
d ϕ tk−1 , y k−1
pk−1 −1 pk−1 −1 , u k−1
pk−1 −1 , y
$ $ % %
≤ d ϕ tk−1 k−1 k−1
pk−1 −1 , ypk−1 −1 , upk−1 −1 , yk−1 + d (yk−1 , xk ) + d (xk , y)
ε ε ε
< + + <ε
4 4 4
and for i < k − 1 we have

d ϕ tipi −1 , ypi i −1 , uipi −1 , yi+1

≤ d ϕ tipi −1 , ypi i −1 , uipi −1 , yi + d (yi , xi+1 ) + d (xi+1 , yi+1 )
ε ε ε
< + + < ε.
4 4 4
Moreover
d (ϕ (t0 , x, u0 ) , y1 ) ≤ d (ϕ (t0 , x, u0 ) , x0 ) + d (x0 , x1 ) + d (x1 , y1 )

ε ε ε
< + + <ε
4 4 4
and then we obtain an (ε, T )-chain from x to y, by concatenations of chains.

Therefore C is a chain transitive set.
Finally, note that condition (3) implies condition (2).
In particular, a compact connected set filled with periodic or homoclinic

trajectories is chain transitive.
Example 7.1.1 Consider a control affine system ẋ = u (t) X (x) on Rn where

X is a rotation matrix. For every > 0, the closed ball B [0, ] is compact,
connected, and filled with periodic trajectories. According to Proposition 7.1.2,
B [0, ] is chain transitive. Now, for a given pair of points x, y ∈ Rn , we can
find > 0 such that x, y ∈ B [0, ]. Thus the system is chain controllable.
Example 7.1.2 Consider the control affine system on the plane as defined
in Example 2.4.3. The intersection T of all tear drops is filled with periodic
and homoclinic trajectories. Hence T is chain transitive. The complement
R2 \ T is contained in a holding set, and it is the chain transitive. By using
concatenations of chains, we can construct chains from T to R2 \ T , and from
R2 \ T to T , with arbitrary jumps and times. Thus the control system is chain
controllable.
We now define the following chain transitivity relation in the chain

recurrence set C:
x ∼ y whenever x ∈ Ω+ (y) and y ∈ Ω+ (x) .
For a given x ∈ C, it is immediate that x ∼ x since x ∈ Ω+ (x) by the

definition. We also have by the definition that x ∼ y implies y ∼ x. Finally,
if x ∼ y and y ∼ z, then x ∼ z by concatenation of chains. Thus the chain
transitivity relation is an equivalence relation in the chain recurrence set C.
If E is an equivalence class of the chain transitivity relation and x, y ∈ E
then x ∈ Ω+ (y) and y ∈ Ω+ (x). For any z ∈ Ω+ (y), it follows that z ∈
Ω+ (x), and therefore Ω+ (y) ⊂ Ω+ (x). Similarly, we have Ω+ (x) ⊂ Ω+ (y).
Thus Ω+ (x) = Ω+ (y). The following result describes the equivalence classes
of the chain transitivity relation.
Proposition 7.1.3 Let C ⊂ M be a nonempty set. The following statements

are equivalent:
1. C is an equivalence class of the chain transitivity relation.

2. C = Ω+ (x) ∩ Ω− (x) for all x ∈ E.
3. C is a maximal chain transitive set.
Proof. (1) ⇒ (2) Suppose that C is an equivalence class of the chain

transitivity relation and let x, y ∈ C. Then x ∈ Ω+ (y) and y ∈ Ω+ (x), that
is, y ∈ Ω− (x) and y ∈ Ω+ (x). Hence y ∈ Ω+ (x) ∩ Ω− (x). Since x, y ∈ C are
arbitraries, it follows that C ⊂ Ω+ (x) ∩ Ω− (x) for all x ∈ C. On the other
hand, if z ∈ Ω+ (x) ∩ Ω− (x) with x ∈ C then z ∈ Ω+ (x) and x ∈ Ω+ (z), that
is, x ∼ z. Hence z ∈ C, and therefore C = Ω+ (x) ∩ Ω− (x).
(2) ⇒ (3) Suppose that C = Ω+ (x) ∩ Ω− (x) for all x ∈ E. Then it is
immediate that C ⊂ Ω+ (x) for every x ∈ C, that is, C is chain transitive.
Let C be a chain transitive set containing C. For x ∈ C and y ∈ C , we have
y ∈ Ω+ (x) and x ∈ Ω+ (y), hence y ∈ Ω+ (x) ∩ Ω− (x) = C. It follows that
C = C and therefore C is a maximal chain transitive set.
(3) ⇒ (1) If C is a maximal chain transitive set then x ∼ y for all x, y ∈ C.
Hence C is contained in some equivalence class C of the chain transitivity
relation. As C is chain transitive, the maximality of C implies that C = C.
Any equivalence class of the chain transitivity relation is called a maximal

chain transitive set. We denote Cx = Ω+ (x) ∩ Ω− (x), for each x ∈ C.
Example 7.1.3 The maximal chain transitive sets need not be invariant. In
fact, consider the control affine system ẋ = X0 (x) + u (t) X1 (x) on R2 , with

∂ 1 0
control range U = [0, 1], where X0 = and X1 = . For u = 0,
∂x2 0 −1
the trajectories are vertical lines oriented by the vector (0, 1). For u ∈ (0, 1],
the solution is of the form

1 −ut e−ut
ut
ϕ (t, x, u) = e x1 , + x2 e −
u u

1
and then the trajectories have a saddle point in , 0 . The line L =
u
{(0, x2 ) : x2 ≥ 1} is then a control set that is the unique chain transitive set
of the system (see Fig. 7.2). It can be easily seen that L is positively invariant
but not negatively invariant.
By Proposition 7.1.1, we have the following characterization of the com-

pact maximal chain transitive sets.
Figure 7.2: A noninvariant maximal chain transitive set.
Proposition 7.1.4 A compact set C ⊂ M is a maximal chain transitive set

if and only if C is maximal satisfying C ⊂ Ω+ (x, ε, 1) for all x ∈ C and
ε > 0.
Proof. Suppose that C is a maximal chain transitive set. Then C ⊂
Ω+ (x, ε, 1) for all x ∈ C and ε > 0. Let C ⊂ M be a set such that C ⊂ C
and C ⊂ Ω+ (x , ε, 1) for every x ∈ C and ε > 0. For a given x ∈ C , define
N = C ∪ {x }. N is then a compact set satisfying N ⊂ C ⊂ Ω+ (y, ε, 1) for
every y ∈ N and ε > 0. According to Proposition 7.1.1, N is chain transi-
tive, and therefore C = N by the maximality of C. It follows that x ∈ C
for every x ∈ C , which means that C = C . Thus C is maximal satisfying
C ⊂ Ω+ (x, ε, 1) for every x ∈ C and ε > 0. On the other hand, if C is
maximal satisfying C ⊂ Ω+ (x, ε, 1) for every x ∈ C and ε > 0 then C is chain
transitive, according to Proposition 7.1.1. If C ⊃ E is a chain transitive set
then C ⊂ Ω+ (x , ε, 1) for every x ∈ C and ε > 0. By the maximality of C,
it follows that C = C . Thus C is a maximal chain transitive set.
We shall now discuss the topological and dynamical properties of the chain
recurrence set C. We first show that C is closed.
Proposition 7.1.5 The chain recurrence set C of the control system is a
closed set in M .
Proof. Let x ∈ cl (C) and ε > 0. According to Proposition 1.3.4, for
ε
a given compact neighborhood K of x, there is δ > 0 with δ < such
2
ε
that d (ϕ (t, y, u) , ϕ (t, z, u)) < for all t ∈ [0, 2], u ∈ U, and y, z ∈ K
2
ε
with d (y, z) < δ. Choose 0 < η < such that B (x, η) ⊂ K and take
2
y ∈ B (x, η) ∩ C. Consider an (ε/2, 1)-chain from y to y composed by points
x0 = y, x1 , ..., xk = y in M , times t0 , ..., tk−1 ∈ [1, 2], and control functions
ε
u0 , u1 , ..., uk−1 . We have d (ϕ (t0 , x, u0 ) , ϕ (t0 , y, u0 )) < by the equicontinu-
2
ity, and then
d (ϕ (t0 , x, u0 ) , x1 ) ≤ d (ϕ (t0 , x, u0 ) , ϕ (t0 , y, u0 )) + d (ϕ (t0 , y, u0 ) , x1 )

ε ε
< + = ε.
2 2
Moreover we have
d (ϕ (tk−1 , xk−1 , uk−1 ) , x) ≤ d (ϕ (tk−1 , xk−1 , uk−1 ) , y) + d (y, x)

ε
< + η < ε.
2
Hence we obtain an (ε, 1)-chain from x to x composed of the points x0 =
x, x1 , ..., xk = x in M , times t0 , ..., tk−1 ∈ [1, 2], and control functions
u0 , u1 , ..., uk−1 . According to Proposition 7.1.1, it follows that x is a chain
recurrent point, hence cl (C) = C.
We may show that the closure of a chain transitive set is chain transitive
We now define the domain of chain attraction for a chain transitive set
and present a dynamic order among the chain transitivity classes.
Definition 7.1.5 Let C be a maximal chain transitive set. The domain of

chain attraction of C is defined by

Ac (C) = x ∈ M : Ω+ (x) ∩ C = ∅ .
Since C ⊂ Ω+ (x) for all x ∈ C, we have C ⊂ Ac (C). It can be easily seen

that Ac (C) = {x ∈ M : C ⊂ Ω+ (x)}. For two maximal chain transitive sets
C1 and C2 , we define the relation
C1 c C2 if and only if C1 ∩ Ac (C2 ) = ∅.
Note that C1 c C2 if and only if there is x ∈ C1 such that C2 ⊂ Ω+ (x),

which implies C2 ⊂ Ω+ (x) for all x ∈ C1 . It is not difficult to see that the
chain transitivity relation is a preorder among the maximal chain transitive
sets. This preorder is called chain transitivity preorder.
Example 7.1.4 Consider the control system as defined in Example 6.3.2.

The stationary points (1, 0) and (−1, 0) are attractor points and (0, 1) and
(0, −1) are repeller points. We have
Ω+ ((1, 0)) = {(1, 0)} , Ω− ((1, 0)) = {(x1 , x2 ) : x1 ≥ 0} ,

Ω+ ((−1, 0)) = {(−1, 0)} , Ω− ((−1, 0)) = {(x1 , x2 ) : x1 ≤ 0} ,
Ω+ ((0, 1)) = {(x1 , x2 ) : x2 ≥ 0} , Ω− ((0, 1)) = {(0, 1)} ,
Ω+ ((0, −1)) = {(x1 , x2 ) : x2 ≤ 0} , Ω− ((0, −1)) = {(0, −1)} .
We only prove that Ω+ ((0, 1)) = {(x, y) : y ≥ 0}. Indeed, if

x ∈ {(x1 , x2 ) : x2 > 0} and u ∈ U, then ϕ (−t, x, u) → (0, 1) as t → +∞. For
, T > 0, take t1 ≥ T such that d (ϕ (−t1 , x, u) , (0, 1)) < . Then the sequences
x0 = (0, 1) , x1 = ϕ (−t1 , x, u) , x2 = x, t0 = t1 ≥ T , and u0 = u1 = u ∈ U
form an (, T )-chain from (0, 1) to x. Thus x ∈ Ω+ ((0, 1)). Now, choose
x = (x1 , x2 ) with x1 , x2 > 0. Since ϕ (−t, x, u) → (0, 1) and ϕ (t, x, u) → (1, 0)
as t → +∞, we can take t ≥ T such that d (ϕ (−t, x, u) , (0, 1)) < and
d (ϕ (t, x, u) , (1, 0)) < . Then the sequences x0 = (0, 1) , x1 = ϕ (−t, x, u) , x2 =
x, x3 = (1, 0), t0 = t1 = t2 = t ≥ T , and u0 = u1 = u2 = u ∈ U form an
(, T )-chain from (0, 1) to (1, 0). Analogously, we obtain an (, T )-chain from
(0, 1) to (−1, 0). Therefore Ω+ ((0, 1)) = {(x, y) : y ≥ 0}. Thus C(1,0) =
{(1, 0)}, C(−1,0) = {(−1, 0)}, C(0,1) = {(0, 1)}, and C(0,−1) = {(0, −1)} are
the maximal chain transitive sets of the system with C(0,1) , C(0,−1) c C(1,0)
and C(0,1) , C(0,−1) c C(−1,0) . Note that the attractors C(1,0) and C(−1,0) are
maximal while the repellers C(0,1) and C(0,−1) are minimal with respect to the
chain transitivity preorder.
The following proposition relates the limit sets and the chain limit sets:
Proposition 7.1.6 For any subset X ⊂ M , one has ω + (X) ⊂ Ω+ (X) and
ω − (X) ⊂ Ω− (X). Consequently, any asymptotic transitive set is chain tran-
sitive and Rx ⊂ Cx .
Proof. For x ∈ ω + (X) and , T > 0, we have B (x, ε) ∩ O>T+

(X) = ∅.
Hence there are y ∈ X, t > T , and u ∈ U , such that d (ϕ (t, y, u) , x) < ,
which shows that x ∈ Ω+ (X). The inclusion ω − (X) ⊂ Ω− (X) is proved
in the same way. Now, let X ⊂ M be an asymptotic transitive set. Then
X ⊂ ω + (x) for all x ∈ X. Hence X ⊂ Ω+ (x), for every x ∈ X, thus X is
chain transitive.
Thus the holding sets, the positive minimal sets, and the weak control
sets are chain transitive. By Proposition 4.1.3, ω + (x) = M for all x ∈
M if the control system is controllable, hence a controllable system is chain
controllable. The converse does not hold (see Example 7.1.5).
The following result relates the prolongational limit sets and the chain
limit sets:
Proposition 7.1.7 For any point x ∈ M , one has J + (x) ⊂ Ω+ (x) and
J − (x) ⊂ Ω− (x). Consequently, every nonwandering point is chain recurrent.
Proof. Let y ∈ J + (x) and ε, T > 0. There are then sequences xn → x,

tn → +∞, and un → u, such that ϕ (tn , xn , un ) → y. Write
ϕ (tn , xn , un ) = ϕ (tn − T, ϕ (T, xn , un ) , un · T ) .
We have tn − T → +∞ and ϕ (T, xn , un ) → ϕ (T, x, u). There are then n such

that
tn − T > T,
d (ϕ (T, x, u) , ϕ (T, xn , un )) < ε,
d (ϕ (tn − T, ϕ (T, xn , un ) , un · T ) , y) < ε.
Hence the points x0 = x, x1 = ϕ (T, xn , un ) , x2 = y in M , the times s0 =

T, s1 = tn − T ≥ T , and the controls u0 = u, u1 = un · T form an (ε, T )-chain
from x to y. Thus J + (x) ⊂ Ω+ (x). The inclusion J − (x) ⊂ Ω− (x) is proved
in a similar way.
Hence the weak prolongational control sets are chain transitive. This shows
that the chain transitivity is the more general notion of transitivity for control
systems, with the following inclusions of classes
P ⊂ R+ ⊂ J ⊂ C.
According to Propositions 7.1.2, 7.1.6, and 7.1.7, we have the following

simpler conditions for the chain transitivity of a compact and connected set:
Corollary 7.1.1 Let C ⊂ M be a compact and connected set. C is then a

chain transitive set if any one of the following conditions holds:
1. C is a set of recurrent points (nonwandering points).

2. The set of recurrent points (nonwandering points) of C is dense in C

and O>T
+
(x) ∩ B (C, δ) = ∅ for all x ∈ C and δ, T > 0.
3. The set of recurrent points (nonwandering points) of C is dense in C

and C is positively invariant.
Since the nonempty compact limit sets of points are connected, Proposi-
tions 5.7.1 and 7.1.7 imply the following:
Corollary 7.1.2 Every nonempty and compact limit set of a point is chain
transitive.
This property of the limit sets is fundamental to the relationship between

the chain transitive sets and the Morse decompositions, described in the next
chapter.
Example 7.1.5 Consider the control system

−u (t) 1
ẋ = x
−1 0
on M = R2 with control range U = [0, 1] ⊂ R. For u = 0, the trajectories

of the system move on circles centered at the origin 0. For given x, y ∈ R2 ,
there is a ball B centered at 0 such that x, y ∈ B. Since B is compact,
connected, and filled with periodic trajectories, B is chain transitive according
to Proposition 7.1.2. Hence x ∈ Ω+ (y) and y ∈ Ω+ (x). Thus the system is
chain controllable. Now, for u ∈ (0, 1], the trajectories of the system move
on spirals around the origin. Hence the circles centered at 0 are the control
sets of the system (continuum of control sets), and then the system is not
controllable.
We shall now discuss the projection and the lift of a chain transitive set.
It should be remembered that the product space M × U is equipped with the
maximum metric.

Proposition 7.1.8 For any x ∈ M , one has π Ω+ π −1 (x) = Ω+ (x) and
− −1
π Ω π (x) = Ω− (x).

Proof. We only prove the equality π Ω+ π −1 (x) = Ω+ (x), the other

one may be proved similarly. Let (y, u) ∈ Ω+ π −1 (x) and , T > 0. There
are then points (x0 , u0 ) = (x, u0 ) , ..., (xk , uk ) = (y, u) in M × U and times
t0 , ..., tk−1 ≥ T such that d (Φ (ti , xi , ui ) , (xi+1 , ui+1 )) < , for i = 0, ..., k − 1,
where
d (Φ (ti , xi , ui ) , (xi+1 , ui+1 )) = max {d (ϕ (ti , xi , ui ) , xi+1 ) , d (ui · ti , ui+1 )} .
Hence d (ϕ (ti , xi , ui ) , xi+1 ) < for i = 0, ..., k − 1, and therefore the points
x0 = x, ..., xk = y in M , the times t0 , ..., tk−1 ≥ T , and the controls u0 , ..., uk−1 ∈
U form an (, T )-chain from x to y. This proves that y ∈ Ω+ (x), and thus

Ω+ π −1 (x) projects into Ω+ (x). On the other hand, take y ∈ Ω+ (x). For
, T > 0, there are points x0 = x, ..., xk = y in M , times t0 , ..., tk−1 ≥ T ,
and the controls u0 , ..., uk−1 ∈ U such that d (ϕ (ti , xi , ui ) , xi+1 ) < for
i = 0, ..., k − 1. Take u ∈ U defined as
⎧
⎪
⎪ u0 (t) , if t ≤ t0
⎪
⎪
⎪
⎪ u 1 (t − t0 ) , if t0 < t ≤ t0 + t1
⎪
⎪
⎪
⎪ ..
⎨ $ % .
u (t) = k−j−1 k−j−1 k−j
⎪ uk−j t − i=0 ti , if ti < t ≤ i=0 ti .
⎪
⎪
i=0
⎪
⎪ ..
⎪
⎪
⎪ $ .
⎪
⎪
⎩ k−2 % k−2
u k−1 t− i=0 t , if
i t <t i=0 i
Then
⎧ $ k−j % k−j
⎪
⎪ u0 t + i=0 ti , if t ≤ − i=1 ti
⎪
⎪
⎪
⎪ ..
⎪
⎪
⎪
⎪
.

k−j ⎨
uk−j (t + tk−j ) , if − tk−j < t ≤ 0
u· ti (t) =
⎪
⎪ uk−j+1 (t) , if 0 < t ≤ tk−j+1
i=0 ⎪
⎪
⎪
⎪ ..
⎪
⎪ .%
⎪
⎪ $ k−2 k−2
⎩ uk−1 t + i=k−j+1 ti , if i=k−j+1 ti < t
for j = 2, ..., k. It follows that Φ (t0 , x0 , u) = (ϕ (t0 , x0 , u0 ) , u · t0 ) and

k−j
Φ tk−j+1 , xk−j+1 , u · ti
i=0

k−j
k−j+1
= ϕ tk−j+1 , xk−j+1 , u · ti ,u · ti
i=0 i=0

k−j+1
= ϕ (tk−j+1 , xk−j+1 , uk−j+1 ) , u · ti
i=0
for j = 2, ..., k. Hence we have
d (Φ (t0 , x0 , u) , (x1 , u · t0 )) = d (ϕ (t0 , x1 , u0 ) , x1 ) <
and

k−j
k−j+1
d Φ tk−j+1 , xk−j+1 , u · ti , xk−j+2 , u · ti
i=0 i=0
= d (ϕ (tk−j+1 , xk−j+1 , uk−j+1 ) , xk−j+2 ) <
$ k−1 %
for all j = 2, ..., k. This means that y, u · i=0 ti ∈ Ω+ ((x, u) , , T ). Thus,
for each n ∈ N, we obtain un , vn ∈ U such that
(y, vn ) ∈ Ω+ ((x, un ) , 1/n, n) .
We may assume that the sequence (vn ) converges to v in U . For , T > 0

we can find n such that 1/n < /2, n > T , and d (vn , v) < /2. Then
(y, vn ) ∈ Ω+ ((x, un ) , 1/n, n) implies (y, v) ∈ Ω+ ((x, un ) , , T ), and hence

(y, v) ∈ Ω+ π −1 (x) , , T for all , T > 0. Thus (y, v) ∈ Ω+ π −1 (x) and
the proof is finished.
Corollary 7.1.3 Each maximal chain transitive set of the control flow projects
into a maximal chain transitive set of the control system.
Proof. If (x, u) ∈ M × U is a chain recurrent point of the control flow

then C(x,u) = Ω+ (x, u) ∩ Ω− (x, u), hence

π C(x,u) ⊂ π Ω+ (x, u) ∩ π Ω− (x, u) ⊂ Ω+ (x) ∩ Ω− (x) = Cx .
We end this section by introducing the classical concept of a chain control

set.
Definition 7.1.6 A set C ⊂ M is called a chain control set if it satisfies

the following properties:
1. For each x ∈ C there is u ∈ U such that ϕ (t, x, u) ∈ C for all t ∈ R.
2. C ⊂ Ω+ (x) for every x ∈ C (chain transitivity).

3. C is maximal satisfying both properties (1) and (2).
If C satisfies both conditions (1) and (2) of Definition 7.1.6 then C is

contained in some chain control set. A proof for this fact is similar to the
proof for Proposition 5.4.4 in the case of a control set (Problem 7.4(1)). Note
that condition (3) guarantees that distinct chain control sets do not intersect
each other (Problem 7.4(2)).
It can be readily seen that a chain control set is contained in some max-
imal chain transitive set. On the other hand, if a maximal chain transitive
set satisfies condition (1) of Definition 7.1.6 then it is a chain control set. For
instance, an invariant maximal chain transitive set is a chain control set. How-
ever, a chain control set may be a maximal chain transitive set not invariant
The next theorem simplifies the definition of a chain control set. The proof
for it is similar to the proof for Proposition 7.1.4.
Proposition 7.1.9 Let C ⊂ M be a precompact set. Then C is a chain

control set if and only if C is maximal satisfying the following properties:
1. For each x ∈ C there is u ∈ U such that ϕ (t, x, u) ∈ C for all t ∈ R.
2. C ⊂ Ω+ (x, ε, 1) for every x ∈ C and ε > 0.
7.2 Attractor-repeller pair

In this section we present the notion of an attractor-repeller pair. This pair
plays a fundamental role in the Conley theory, especially for the description
of the chain recurrence set of the control system.
We first show that the positive chain limit sets are described by intersec-
tion of Conley attractors while the negative chain limit sets are described by
intersection of Conley repellers. These facts reveal the asymptotic property
of the chain limit sets. We need the following.
Proposition 7.2.1 For any , T > 0 and X ⊂ M , the sets Ω+ (X, , T ) and
Ω− (X, , T ) are open.
Proof. For y ∈ Ω (X, , T ), there is x0 = x, ..., xk = y ∈ M , with x ∈ X,

t0 , ..., tk−1 ≥ T , and u0 , ..., uk−1 ∈ U such that d (ϕ (ti , xi , ui ) , xi+1 ) < for
i = 0, ..., k − 1. As d (ϕ (tk−1 , xk−1 , uk−1 ) , y) < , we have
Section 7.2 · Attractor-repeller pair 231
y ∈ B (ϕ (tk−1 , xk−1 , uk−1 ) , ), where B (ϕ (tk−1 , xk−1 , uk−1 ) , ) denotes the
-ball centered at ϕ (tk−1 , xk−1 , uk−1 ). For any z ∈ B (ϕ (tk−1 , xk−1 , uk−1 ) , ),
we can obtain an (, T )- chain from x to z. Hence B (ϕ (tk−1 , xk−1 , uk−1 ) , ) ⊂
Ω (X, , T ), and therefore Ω+ (X, , T ) is open. The case Ω− (X, , T ) is proved
by the same way.
Theorem 7.2.1 Let X ⊂ M be a nonempty subset. Then
1. ω + (Ω+ (X, , T )) is a Conley attractor containing ω + (X), for all , T >

0, and
Ω+ (X) = ω + Ω+ (X, , T ) .
,T >0
2. ω − (Ω− (X, , T )) is a Conley repeller containing ω − (X), for all , T >

0, and
Ω− (X) = ω − Ω− (X, , T ) .
,T >0
Proof. We only prove item (1). If y ∈ ω + (Ω+ (X, , T )), then there are
t > T , u ∈ U , and x ∈ Ω+ (X, , T ) such that ϕ (t, x, u) ∈ B (y, ). By taking
x0 ∈ X and an (, T )-chain from x0 to x, we obtain an (, T )-chain from x0
to y. Hence y ∈ Ω+ (X, , T ), and therefore ω + (Ω+ (X, , T )) ⊂ Ω+ (X, , T ).
Since Ω+ (X, , T ) is open, it follows that ω + (Ω+ (X, , T )) is a Conley at-
tractor with attractor neighborhood Ω+ (X, , T ). Now, let z ∈ Ω+ (X) and
, T > 0 fixed. For any δ, t > 0, take δ < min {, δ} and t > max {T, t}.
There are points x0 , ..., xk = z ∈ M , x0 ∈ X, times t0 , ..., tk−1 ≥ t , and

controls u0 , ..., uk−1 ∈ U which define a δ , t -chain from x0 to z. Then

ϕ (tk−1 , xk−1 , uk−1 ) ∈ B z, δ ⊂ B (z, δ), with tk−1 > t. Moreover, the
sequences x0 , ..., xk−1 ∈ M , t0 , ..., tk−2 ≥ t , and u0 , ..., uk−2 ∈ U, form a

δ , t -chain from x0 to xk−1 . Since x0 ∈ X, δ < , and t > T , it follows that
xk−1 ∈ Ω+ (X, , T ). Hence ϕ (tk−1 , xk−1 , uk−1 ) ∈ O>t+
(Ω (X, , T )) ∩ B (z, δ).
Since δ, t are arbitrary, we have z ∈ ω (Ω (X, , T )). Thus Ω+ (X) ⊂
+ +
ω + (Ω+ (X, , T )) for all , T > 0. By Proposition 7.1.6, we now have ω + (Y ) ⊂

Ω+ (Y ) ⊂ ω + (Ω+ (X, , T )) for all , T > 0. Finally,

ω + Ω+ (X, , T ) ⊂ Ω+ (X, , T ) = Ω+ (X)
,T >0 ,T >0

and therefore Ω+ (X) = ω + (Ω+ (X, , T )).
,T >0
Thus a maximal chain transitive set Cx = Ω+ (x)∩Ω− (x) is an intersection

between Conley attractors and Conley repellers. Since the Conley attractors
and the Conley repellers are closed sets, it follows that each Cx is a closed
set.
Although the Conley attractors are positively invariant and the Conley
repellers are negatively invariant, Cx may not have these properties (Example
7.1.3). In compact state space, however, maximal chain transitive sets are
invariant. We will prove this fact later.
There are cases in which a maximal chain transitive set is a Conley at-
tractor or a Conley repeller. In Example 7.1.4, C(1,0) and C(−1,0) are Con-
ley attractors while C(0,1) and C(0,−1) are Conley repellers; both the Conley
attractors are maximal and both the Conley repellers are minimal with re-
spect to the chain transitivity preorder. These properties hold, in general, for
asymptotically compact control systems.
Proposition 7.2.2 Assume that the control system is asymptotically com-

pact. Then
1. A maximal chain transitive set that is a compact Conley attractor is

maximal with respect to the chain transitivity preorder.
2. A maximal chain transitive set that is a compact Conley repeller is min-

imal with respect to the chain transitivity preorder.
Proof. (1) Let C be a maximal chain transitive set that is a compact Con-
ley attractor. Then there is > 0 $
such that cl (B (C,
% )) is compact and for
any δ > 0 there is a T > 0 with cl O≥T (B (C, )) ⊂ B (C, δ/2). Let δ <
+
and suppose that C ⊂ M is a maximal chain transitive set with C C .

Then C ⊂ Ω+ (x) for every x ∈ C. Hence, for y ∈ C and x ∈ C, there is a
(δ/2, T )-chain from x to y formed by sequences x0 = x, x1 , ..., xn = y in M ,
u0 , ..., un−1 in U , and t0 , ..., tn−1 ≥ T . Since x0 = x ∈ C and t0 ≥ T , we have
δ
ϕ (t0 , x0 , u0 ) ∈ O≥T
+
(B (C, )) ⊂ B (C, δ/2). As d (ϕ (t0 , x0 , u0 ) , x1 ) < , it
2
follows that x1 ∈ B (C, δ), by triangular inequality. Now, since x1 ∈ B (C, ),
δ
ϕ (t1 , x1 , u1 ) ∈ O≥T+
(B (C, )) ⊂ B (C, δ/2) and d (ϕ (t1 , x1 , u1 ) , x2 ) < .
2
Hence x2 ∈ B (C, δ). Following by induction, we have y = xn ∈ B (C, δ).
Since δ is arbitrary, we have y ∈ C. It follows that C = C , and therefore C

is maximal with respect to the chain transitivity preorder.
(2) Let C be a maximal chain transitive set that is a compact Conley
repeller. Then there is > 0 such
$ that cl (B (C,
% )) is compact and for any
−
δ > 0 there is a T > 0 with cl O≥T (B (C, )) ⊂ B (C, δ/2). Let δ < and
suppose that C ⊂ M is a maximal chain transitive set with C C. Then
C ⊂ Ω+ (x) for every x ∈ C . Hence, for x ∈ C and y ∈ C , there is a (δ/2, T )-
chain from y to x formed by sequences y0 = y, y1 , ..., yn = x in M , u0 , ..., un−1
in U , and t0 , ..., tn−1 ≥ T . Since yn = x ∈ E and d (ϕ (tn−1 , yn−1 , un−1 ) , x) <
δ
, we have ϕ (tn−1 , yn−1 , un−1 ) ∈ B (C, δ). As tn−1 ≥ T , it follows that
2
yn−1 = ϕ (−tn−1 , ϕ (tn−1 , yn−1 , un−1 ) , un−1 · tn−1 )
−
∈ O≥T (B (C, )) ⊂ B (C, δ/2) .
δ
Since d (ϕ (tn−2 , yn−2 , un−2 ) , yn−1 ) < and tn−2 ≥ T , we have
2
ϕ (tn−1 , yn−1 , un−1 ) ∈ B (C, δ)
hence yn−2 ∈ B (C, δ/2). Following by induction, we have y = x0 ∈ B (C, δ/2).

Since δ is arbitrary, we have y ∈ C, which implies y ∈ C ∩ C. This means
that C = C , and therefore C is minimal with respect to the chain transitivity
preorder.
In the next chapter, we will see that the Conley attractors and the Conley
repellers are respectively upper and lower bounds in Morse decompositions.
From now on, in order to define the notion of an attractor-repeller pair, we
assume that M is a compact manifold. The Conley attractors and the Conley
repellers are then compact, all limit sets of nonempty sets are nonempty and
compact, and the limit sets of points are connected.
Outside an attractor neighborhood of a Conley attractor, there exists a
peculiar Conley repeller. To describe it, we need the following.
Theorem 7.2.2 Let Y ⊂ M be a nonempty set. Then Ω+ (Y ) is the intersec-

tion of all Conley attractors containing ω + (Y ) and Ω− (Y ) is the intersection
of all Conley repellers containing ω − (Y ).
Proof. Let A be a Conley attractor that contains ω + (Y$ ) and V %an

attractor neighborhood of A. There is then T > 0 such that cl O≥T
+
(V ) ⊂
$ $ % %
int (V ). Take ε > 0 such that B cl O≥T
+
(V ) , ε ⊂ int (V ). Since ω + (Y ) ⊂
A ⊂ int (V ), there is T0 > 0 such that O≥T
+
0
(Y ) ⊂ V . We then have
$ %
O≥T+
+T0 (Y ) = O≥T O≥T0 (Y ) ⊂ O≥T (V ) ⊂ int (V ) .
+ + +
We claim that Ω+ (Y, ε, T + T0 ) ⊂ int (V ). In fact, take z ∈ Ω+ (Y, ε, T + T0 )

and x0 = y, ..., xn = z, with y ∈ Y , u0 , ..., un−1 in U , and t0 , ..., tn−1 ≥ T + T0
forming an (ε, T + T0 )-chain from y to z. Since ϕ (t0 , y, u0 ) ∈ O≥T +
+T0 (Y )
and d (ϕ (t0 , y, u0 ) , x1 ) < ε, it follows that
$ $ % %
x1 ∈ B cl O≥T +
(V ) , ε ⊂ int (V ) .
Hence ϕ (t1 , x1 , u1 ) ∈ O≥T

+
(V ) and d (ϕ (t1 , x1 , u1 ) , x2 ) < ε, and thus
$ $ % % +T0
x2 ∈ B cl O≥T (V ) , ε ⊂ int (V ). Following by induction we obtain
+
xn = z ∈ int (V ), and thus Ω+ (Y, ε, T + T0 ) ⊂ int (V ), proving the claim.

Now, ω + (Ω+ (Y, ε, T + T0 )) ⊂ ω + (V ) = A. Hence Ω+ (Y ) ⊂ A, according
to Theorem 7.2.1, and therefore Ω+ (Y ) is contained in the intersection of all
Conley attractors containing ω + (Y ). The other inclusion follows from Theo-
rem 7.2.1. The proof for Ω− (Y ) is analogous.
Proposition 7.2.3 If A ⊂ M is a Conley attractor of the control system

then Au (A) is invariant.
Proof. The result is immediate if A is a trivial Conley attractor. Suppose

that A is nontrivial. We only need to show that Au (A) is negatively invari-
ant. Indeed, take x ∈ Au (A), t > 0, and u ∈ U . Then J + (x) ⊂ A, hence
ω + (x) ⊂ A. According to Proposition 4.1.7, it follows that ω + (ϕ (−t, x, u)) ∩
A = ∅. According to Proposition 6.2.4, this means that ω + (ϕ (−t, x, u)) ⊂ A.
Now, according to Theorem 7.2.2, we have Ω+ (ϕ (−t, x, u)) ⊂ A. Hence
J + (ϕ (−t, x, u)) ⊂ A according to Proposition 7.1.7. This means that
ϕ (−t, x, u) ⊂ Au (A), and therefore Au (A) is negatively invariant.
According to Propositions 6.2.5 and 7.2.3, we have the equalities Aw (A) =

A (A) = Au (A) for every Conley attractor A.
For the following, we consider an attractor neighborhood U of a Con-
ley$ attractor
% A and use Lemma 6.2.1 to obtain a time t > 0 such that
cl O≥t+
(U ) ⊂ int (U ).
Theorem 7.2.3 Let A ⊂ M be a Conley $ attractor

% with attractor neighbor-
hood U and consider t > 0 such that cl O≥t (U ) ⊂ int (U ). Define the open
+
$ %
set U ∗ = M \ cl O≥t+
(U ) . Then ω − (U ∗ ) is an invariant Conley repeller,
disjoint from A, with repeller neighborhood U ∗ . Furthermore
ω − (U ∗ ) = M \ Au (A) = M \ A (A) = M \ Aw (A) .
Proof. Since Aw (A) = A (A) = Au (A), it is enough to show that

ω − (U ∗ ) = M \ Au (A). We assume the nontrivial cases A = ∅ and A = M .
−
Since O≥t +
(U ) ∩ U ∗ = ∅, we have O≥t (U ∗ ) ⊂ M \ U . Hence
$ % $ %
− −
ω − (U ∗ ) = cl O≥s (U ∗ ) ⊂ cl O≥t (U ∗ ) ⊂ M \ U ⊂ U ∗
s>0
which means$that ω − (U ∗ ∗
% ) is a Conley repeller with repeller neighborhood U .
Since A ⊂ cl O≥t
+
(U ) , we have A∩U ∗ = ∅, and therefore A and ω − (U ∗ ) are
disjoint sets. We also have M \ Au (A) ⊂ M \ int (U ), because U ⊂ Au (A).
Hence M \ Au (A) ⊂ U ∗ . We claim that U ∗ ⊂ Ru (M \ Au (A)). Indeed,
take x ∈ U ∗ and suppose by contradiction that x ∈/ Ru (M \ Au (A)). Then
J − (x) M \ Au (A), which means there is some y ∈ J − (x) ∩ Au (A). This
implies that x ∈ J + (y) ⊂ A, which contradicts x ∈ U ∗ , proving the claim.
The set Ru (M \ Au (A)) is then a neighborhood of M \ Au (A), and therefore
M \ Au (A) is an invariant and compact uniform repeller. Now, according to
Proposition 2.2.2, we have
$ $ %% $ $ $ %%%
− − −
ω − cl O≥t (U ∗ ) = cl O≥s cl O≥t (U ∗ )
s>0
$ %
−
= cl O≥s+t (U ∗ )
s>0
ω (U ∗ ) .
= −
$ % $ %
− −
Since cl O≥t (U ∗ ) is compact and cl O≥t (U ∗ ) ⊂ U ∗ , we have
$ $ %% $ $ %%
− −
ω − cl O≥t (U ∗ ) ⊂ J − cl O≥t (U ∗ ) ⊂ J − (U ∗ ) ⊂ M \ Au (A)
hence ω − (U ∗ ) ⊂ M \ Au (A). On the other hand, if x ∈ M \ Aw (A) then

ω + (x) ∩ A = ∅. This implies that ω + (x) ∩ U = ∅, because ω + (U ) = A. Then
ω + (x) ⊂ M \ U ⊂ U ∗ . As U ∗ is open, we can find sequences tn → +∞ and
(un ) in U such that ϕ (tn , x, un ) ∈ U ∗ . By denoting xn = ϕ (tn , x, un ), we have

x = ϕ (−tn , xn , un · tn ) with −tn → −∞ and (xn ) in U ∗ , hence x ∈ ω − (U ∗ ).
Thus M \ Aw (A) ⊂ ω − (U ∗ ) and the proof is finished.
The set M \ Au (A) is often called the complementary repeller of the

Conley attractor A and is denoted by A∗ . The pair (A, A∗ ) is called an
attractor-repeller pair of the control system. Then
A∗ = ω − (U ∗ ) = M \ Au (A) = M \ A (A) = M \ Aw (A) .
The main property of the attractor-repeller pair (A, A∗ ) is obtained from

Proposition 6.2.4, as in the following:
Proposition 7.2.4 Let A ⊂ M be a Conley attractor. For any x ∈ M , one

has ω − (x) , ω + (x) ⊂ A ∪ A∗ , with the following conditions:
/ A ∪ A∗ then ω − (x) ⊂ J − (x) ⊂ A∗ and ω + (x) ⊂ J + (x) ⊂ A.

1. If x ∈
2. If x ∈ A then ω + (x) ⊂ J + (x) ⊂ A, and either ω − (x) ⊂ A∗ or ω − (x) ⊂

A.
3. If x ∈ A∗ then ω − (x) ⊂ J − (x) ⊂ A∗ and ω + (x) ⊂ A∗ .
Thus an attractor-repeller pair contains all limit sets of points. This

means that every trajectory of the system has “beginning” and “end” in the
attractor-repeller pair. This global dynamical behavior does not admit cycles,
as follows:
Corollary 7.2.1 Let A ⊂ M be a Conley attractor and x ∈ M .
1. If ω − (x) , ω + (x) ⊂ A then x ∈ A.
2. If ω − (x) , ω + (x) ⊂ A∗ then x ∈ A∗ .
According to Proposition 6.2.6, a nonempty Conley attractor A is invariant

if and only if x ∈ A implies ω − (x) , ω + (x) ⊂ A.
Example 7.2.1 Consider the control system on the unit circle S1 defined in
Example 6.3.2. The upper and lower semicircles are invariant Conley attrac-
tors and their complementary repellers are respectively the Conley repeller
points {(0, −1)} and {(0, 1)}. The single sets {(1, 0)} and {(−1, 0)} are Con-
ley attractors and their complementary repellers are respectively the left-side
and right-side semicircles.
Example 7.2.2 Consider the control system defined in Example 6.2.1

restricted to the unit disk D2 . The unit circle S1 is an invariant Conley
attractor and its complementary repeller is the Conley repeller point {0}.
We shall now discuss the lift and the projection of a Conley attractor. The
following shows that every Conley attractor for the control system lifts to a
Conley attractor for the control flow:
Proposition 7.2.5 If A ⊂ M is a Conley attractor for the control system

with attractor neighborhood U ⊂ M then L (A) is a Conley attractor for the
control flow with attractor neighborhood U × U . The converse holds if A is
invariant.
Proof. According to Proposition 4.1.8, we have

π ω + (U × U ) = π ω + π −1 (U ) = ω + (U ) = A.
As A ⊂ int (U ), we have ω + (U × U ) ⊂ int (U ) × U , hence ω + (U × U ) is a

Conley attractor of the control flow with attractor neighborhood U × U . Take
(x, u) ∈ ω + (U × U ) and t ∈ R. Since ω + (U × U ) is Φ-invariant, we have
Φ (t, x, u) ∈ ω + (U × U ) for every t ∈ R. Hence ϕ (t, x, u) ∈ A for every t ∈ R,
and thus (x, u) ∈ L (A). This proves that ω + (U × U ) ⊂ L (A). On the other
hand, we have L (A) ⊂ L (U ) ⊂ U × U, and since L (A) is invariant according
to Proposition 2.1.5, we have
L (A) = ω + (L (A)) ⊂ ω + (U × U ) .
It follows that ω + (U × U ) = L (A), and therefore L (A) is a Conley attractor

for the control flow with attractor neighborhood U × U . For the converse,
assume that A is invariant and L (A) is a Conley attractor for the control
flow with attractor neighborhood U × U. According to Proposition 2.1.5
again, L (A) = A × U . Hence we have

A = π (A × U ) = π ω + (U × U ) = ω + (U ) .
As A × U ⊂ int (U ) × U , it follows that A = ω + (U ) ⊂ int (U ), and the result

is proved.
This result implies the following properties of attractors.

Corollary 7.2.2 Let B ⊂ M be an invariant Conley attractor. Assume that

A ⊂ B is an invariant Conley attractor for the control system restricted to
B. Then A is a Conley attractor in M .
Proof. According to Proposition 7.2.5, L (B) is a Conley attractor for

the control flow in M × U and L (A) is a Conley attractor for the control flow
restricted to L (B). According to Corollary A.3.1 in Appendix A, this means
that L (A) is a Conley attractor for the control flow in M × U . According to
Proposition 7.2.5 again, it follows that A is a Conley attractor for the control
system on M .
Proposition 7.2.5 shows that a Conley attractor in M with attractor neigh-

borhood U lifts to a Conley attractor in M × U with attractor neighborhood
U ×U . An important property of the control flow is that any Conley attractor
in M ×U has an attractor neighborhood of the form U ×U, as in the following:
Theorem 7.2.4 Every Conley attractor for the control flow in M × U has
an attractor neighborhood of the form U × U. Consequently, every Conley
attractor for the control flow projects onto a Conley attractor for the control
system.
Proof. The trivial cases of Conley attractors are clear. Suppose that
A ⊂ M × U is a nontrivial Conley attractor for the control flow. We claim
that π (A) ∩ π (A∗ ) = ∅. Indeed, suppose by contradiction there is x ∈
π (A) ∩ π (A∗ ). There are then u, v ∈ U such that (x, u) ∈ A and (x, v) ∈ A∗ .
Take w ∈ U given by

u (t) , if t ≤ 0
w (t) =
v (t) , if t > 0
and let (tn ) be a sequence in R. For each n we have

0, if t ≤ −tn
w · tn (t) − u · tn (t) = .
v · tn (t) − u · tn (t) , if t > −tn
For any α ∈ L1 (R, Rn ), it follows that

fα (w · tn − u · tn ) = w · tn (t) − u · tn (t) , α (t) dt
R
+∞
= w · tn (t) − u · tn (t) , α (t) dt
−tn
and then fα (w · tn − u · tn ) → 0 as tn → −∞. This means that

lim w · tn = lim u · tn , and hence ω − (x, w) = ω − (x, u). On the other
tn →−∞ tn →−∞
hand, we have

u · tn (t) − v · tn (t) , if t ≤ −tn
w · tn (t) − v · tn (t) =
0, , if t > −tn
and then
−tn
fα (w · tn − v · tn ) = w · tn (t) − u · tn (t) , α (t) dt
−∞
which implies fα (w · tn − v · tn ) → 0 as tn → +∞. Hence lim w · tn =

tn →+∞
∗
lim v·tn , and therefore ω + (x, w) = ω + (x, v). Since A and A are invariant
tn →+∞
and disjoint, the equalities ω − (x, w) = ω − (x, u) and ω + (x, w) = ω + (x, v)
imply (x, w) ∈ / A ∪ A∗ . This means that ω − (x, w) ⊂ A∗ and ω + (x, w) ⊂ A,
and hence ω (x, u) ⊂ A∗ and ω + (x, v) ⊂ A. This contradicts (x, u) ∈ A
−
and (x, v) ∈ A∗ . Thus π (A) and π (A∗ ) are disjoint sets, proving the claim.
Now, take a compact neighborhood K of π (A) such that K ∩ π (A∗ ) = ∅. If
(x, u) ∈ π −1 (K) then J + (x, u) ⊂ A, hence π −1 (K) ⊂ Au (A). Since π −1 (K)

is compact in M × U , this implies that ω + π −1 (K) ⊂ J + π −1 (K) ⊂ A.

As A = ω + (A) ⊂ ω + π −1 (K) , it follows that π −1 (K) = K × U is an at-

tractor neighborhood of A. Finally, we have ω + (K) = π ω + π −1 (K) =
π (A) ⊂ K, hence π (A) is a Conley attractor for the control system.
Analogously, we have the following relations between the Conley repellers

of the control system and the Conley repellers for the control flow:
1. If R ⊂ M is a Conley repeller for the control system with repeller

neighborhood V ⊂ M then L (R) is a Conley repeller for the control
flow with repeller neighborhood V × U. The converse holds if R is
invariant.
2. If S is an invariant Conley repeller in M and R is an invariant Conley

repeller in S then R is a Conley repeller in M .
3. Every Conley repeller for the control flow in M × U has a repeller neigh-
borhood of the form V × U .
We are now able to describe the lift of a complementary repeller, as follows.

Proposition 7.2.6 If A ⊂ M is a Conley attractor for the control system

∗
then L (A) = L (A∗ ).
∗
Proof. Let U be an attractor neighborhood of A. If (x, u) ∈ L (A) then
ω (x, u) ∩ L (A) = ∅. This implies that ω + (x, u) ∩ U × U = ∅, and therefore
+
π (ω + (x, u)) ∩ A = ∅. As π (ω + (x, u)) ⊂ ω + (x), we have ω + (x) A, hence

∗
x ∈ A∗ . Since A∗ is invariant, we have (x, u) ∈ L (A∗ ). Thus L (A) ⊂ L (A∗ ).
On the other hand, if (x, u) ∈ L (A∗ ) then x ∈ A∗ , and hence ω + (x) ⊂ A∗ .

This means that ω + (x) ∩ U = ∅, and then π ω + π −1 (x) ∩ U = ∅. Thus

ω + π −1 (x) ∩ (U × U ) = ∅, which implies ω + (x, u) ∩ L (A) = ∅. It follows
∗ ∗
that L (A∗ ) ⊂ L (A) and therefore L (A) = L (A∗ ).
Thus if (A, A∗ ) is an attractor-repeller pair of the control system then

(L (A) , L (A∗ )) is an attractor-repeller pair of the control flow.
Note that we start with a Conley attractor to define the attractor-repeller
pair. Evidently, by considering the time-reversed control system, we may
start with a Conley repeller to define the repeller-attractor pair. Indeed,
if R ⊂ M is a compact Conley repeller with repeller neighborhood V then
R $is a negatively
% invariant uniform repeller and there
$ is a %t > 0 such that
− −
cl O≥t (V ) ⊂ int (V ). By setting V∗ = M \ cl O≥t (V ) , ω + (V∗ ) is an
invariant Conley attractor, disjoint from R, and
ω + (V∗ ) = M \ Rw (R) = M \ R (R) = M \ Ru (R) .
Then ω + (V∗ ) is called the complementary attractor of the Conley repeller

R and is denoted by R∗ . The pair (R, R∗ ) is called a repeller-attractor pair
of the control system. A natural question asks about the equalities A = (A∗ )∗
∗
and R = (R∗ ) . They hold if all Conley attractors and Conley repellers of
the system are invariant sets (Problem 7.4(11)).
7.3 The Conley main theorem

A central Conley’s theorem describes the chain recurrence set by means of
the attractor-repeller pairs. This reveals the invariance and the rigid topolog-
ical property of the maximal chain transitive sets.
In order to show the Conley main theorem, we provide the following al-
ternative description of a negative chain limit set.
Section 7.3 · The Conley main theorem 241
Proposition 7.3.1 For each x ∈ M , the set Ω− (x) coincides with the inter-
section of all complementary repellers of the Conley attractors which do not
contain x.
Proof. Let A ⊂ M be a Conley attractor not containing x and take

y ∈ Ω− (x). Then x ∈ Ω+ (y). According to Theorem 7.2.2, Ω+ (y) is the
intersection of all Conley attractors containing ω + (y). As x ∈
/ A and x ∈
Ω (y), it follows that ω (y) A, hence y ∈
+ +
/ A (A). Thus we have the
inclusion
#
Ω− (x) ⊂ M \ {A (A) : A is a Conley attractor with x ∈ / A}

= {A∗ : A is a Conley attractor with x ∈
/ A} .

On the other hand, let z ∈ {A∗ : A is a Conley attractor with x ∈ / A}. We
have to show that z ∈ Ω− (x), which is equivalent to x ∈ Ω+ (z). Indeed, let
A be a Conley attractor containing ω + (z). If x ∈
/ A then z ∈/ A (A), hence
ω (z) A, a contradiction. Thus x ∈ A, where A is an arbitrary Conley
+
attractor containing ω + (z). According to Theorem 7.2.2, x ∈ Ω+ (z), as de-

sired.
According to Proposition 2.1.1 and Corollary 6.2.3, the reunion

#
{A (A) : A is a Conley attractor with x ∈/ A}
is an invariant (and open) set. According to Proposition 7.3.1, it follows that

the negative chain limit set Ω− (x) is invariant. By repeating all the procedure
with Conley repeller instead of Conley attractor, we can show that

Ω+ (x) = {R∗ : R is a Conley repeller with x ∈/ R}
and therefore Ω+ (x) is an invariant set. Thus every maximal chain transitive
set Cx = Ω+ (x) ∩ Ω− (x) is invariant.
Proposition 7.3.2 Every maximal chain transitive set of the control system
is invariant. Consequently, the chain recurrence set C is invariant.
We now have an interesting description of the maximal chain transitive

sets by means of the connected components of the chain recurrence set C.
Since C is invariant, its connected components are invariant (Problem 2.5(9)).
Theorem 7.3.1 The connected components of the chain recurrence set C co-
incide with the maximal chain transitive sets of the control system.
Proof. Let C ⊂ M be a connected component of C. Then C is a compact

invariant set of M , since C is compact, according to Proposition 7.1.5, and
invariant, according to Proposition 7.3.2. As C is connected, compact, and
chain recurrent, C is chain transitive, according to Proposition 7.1.2. Thus C
is contained in a maximal chain transitive set. Now, if E ⊂ M is a maximal
"
chain transitive set then E ⊂ Cα where each Cα is a connected component
α∈I
of C with Cα ∩ E = ∅. Since Cα is chain transitive and E is maximal, it
"
follows that Cα ⊂ E, and therefore E = Cα . It remains to show that E is
α∈I
connected. Suppose to the contrary that E is not connected. There are then
nonempty compact sets A, B ⊂ M such that E = A ∪ B and A ∩ B = ∅. This
means that dH (A, B) > 0, that is,
ε0 = inf {d (a, b) : a ∈ A, b ∈ B} > 0.
Let x ∈ A and y ∈ B. According to Proposition 7.3.3, there is an (ε0 /2, 1)-

chain from x to y formed by sequences of points x0 = x, x1 , ..., xk = y in
E, times t0 , t1 , ..., tk−1 ≥ 1, and control functions u0 , u1 , ..., uk−1 . We then
ε0
have d (ϕ (tk−1 , xk−1 , uk−1 ) , y) < . Since E is invariant and xk−1 ∈ E,
2
we have ϕ (tk−1 , xk−1 , uk−1 ) ∈ E. It follows that ϕ (tk−1 , xk−1 , uk−1 ) ∈ B,
because of the definition of ε0 . Note that B is a reunion of connected com-
ponent of C, which implies that B is an invariant set. Hence xk−1 ∈ B. As
ε0
d (ϕ (tk−2 , xk−2 , uk−2 ) , xk−1 ) < , the same argument shows that xk−2 ∈ B.
2
By using this argument successively, we obtain x ∈ B, which contradicts
x ∈ A. Thus E is connected and the theorem is proved.
The following is an immediate consequence of Theorem 7.3.1.
Corollary 7.3.1 Let C ⊂ M . The following statements are equivalent:
1. C is a maximal chain transitive set.
2. C is a connected component of C.
3. C is a chain control set.
Theorem 7.3.1 suggests that a maximal chain transitive set is usually called
a chain component of the control system.
We finally present the main result of this chapter which characterizes the
chain recurrence set by means of the attractor-repeller pairs.
Theorem 7.3.2 (Conley main theorem) The chain recurrence set C co-
incides with the set

{A ∪ A∗ : A is a Conley attractor} .
Proof. If x ∈ C then x belongs to the intersection of all Conley attractors

containing ω + (x), according to Theorem 7.2.2. Let A be a Conley attractor
such that ω + (x) A. This means that x ∈ A∗ . Hence x ∈ A ∪ A∗ for ev-

ery Conley attractor A. Therefore C ⊂ {A ∪ A∗ : A is a Conley attractor}.
On the other hand, suppose that y ∈ A ∪ A∗ for all Conley attractors A.
If ω + (y) ⊂ A for a Conley attractor A, then y ∈ / A∗ . Hence y ∈ A, and
thus y belongs to all Conley attractors containing ω + (y). This means that
y ∈ Ω+ (y), that is, y ∈ C, finishing the proof.
Consequently, a chain controllable control system only admits the trivial

attractor-repeller pairs, as follows:
Corollary 7.3.2 The control system is chain controllable if and only if the
whole state space M is the unique nonempty Conley attractor.
Proof. Suppose that the system is chain controllable, that is, M = Ω+ (x)
for every x ∈ M . Let A be a Conley attractor containing x. It follows
that ω + (x) ⊂ A. According to Theorem 7.2.2, we have Ω+ (x) ⊂ A, hence
M = Ω+ (x) = A. Conversely, assume that M is the unique nonempty Conley
attractor. According to Proposition 7.2.2, Ω+ (x) = M for every x ∈ M , thus
the system is chain controllable.
We finish the chapter by proving that the chain components of the con-
trol system are internally chain transitive, which means that the control
system restricted to each chain component is chain controllable. In order to
prove this fact, it should be remembered that the hyperspace H (M ) off all
nonempty compact subsets of M is a compact metric space endowed with the
Hausdorff distance given by
dH (A, B) = max {dist (A, B) , dist (B, A)}
!
= max sup inf d (a, b) , sup inf d (a, b) .
a∈A b∈B b∈B a∈A
Theorem 7.3.3 Every chain component of the control system is internally

chain transitive.
Proof. Let C ⊂ M be a chain component of the system. It is enough to

show that for any pair of points x, y ∈ C and ε, T > 0, there is an (ε, T )-chain
from x to yformed by a sequence of points in C. In fact, for x, y ∈ C and p ∈ N
1
there is a , 1 -chain from x to y formed by points x0 = x, x1 , ..., xnp = y
p
in M , control functions u0 , u1 , ..., unp −1 , and times tp0 , ..., tpnp −1 ∈ [1, 2], and

1
there is a , 1 -chain from y to x formed by points y0 = x, y1 , ..., ymp = x
p
in M , control functions v0 , v1 , ..., vmp −1 , and times sp0 , ..., spmp −1 ∈ [1, 2]. We
then define the compact sets
#
np
Kp,x = {xi , ϕ (tpi , xi , ui ) , y} ,
i=0
#
mp

Kp,y = yj , ϕ spj , yj , vj , x ,
j=0
and obtain a sequence of compact sets (Kp )p∈N with Kp = Kp,x ∪ Kp,y . We
may assume that (Kp ) converges on the Hausdorff topology to a nonempty
compact set K ⊂ M with x, y ∈ K. Now let ε > 0. According to Proposition
ε ε
1.3.4, there is δ > 0 with δ < such that d (ϕ (t, z, u) , ϕ (t, w, u)) < for
4 4
all t ∈ [0, 2], u ∈ U , and z, w ∈ M with d (z, w) < δ. For any z ∈ K, there
is an (ε, 1)-chain in K from z to y and there! is an (ε, 1)-chain in K from
4 1
y to z. Indeed, we choose p > max , with dH (Kp , K) < δ. Then,
ε δ
for each i = 1, ..., np − 1, we take xi ∈ K such that d (xi , xi ) < δ, and for

each j = 1, ..., mp − 1, we take yj ∈ K such that d yj , yj < δ. Hence we
obtain sequences of points x0 = x, x1 , ..., xnp = y ∈ K, control functions
u0 , u1 , ..., unp −1 , and times tp0 , ..., tpnp −1 ∈ [1, 2] such that

d ϕ (tpi , xi , ui ) , xi+1 ≤ d (ϕ (tpi , xi , ui ) , ϕ (tpi , xi , ui )) +

+d (ϕ (tpi , xi , ui ) , xi+1 ) + d xi+1 , xi+1
ε 1
< + +δ <ε
4 p
which form an (ε, 1)-chain in K from x to y. We also obtain sequences of
points y0 = y, y1 , ..., ym

p
= x ∈ K, control functions v0 , v1 , ..., vnp −1 , and
times sp0 , ..., spmp −1 ∈ [1, 2] such that

d ϕ spj , yj , vj , yj+1 ≤ d ϕ spj , yj , vj , ϕ spj , yj , vj +

+d ϕ spj , yj , vj , yj+1 + d yj+1 , yj+1
ε 1
< + +δ <ε
4 p
which determine an (ε, 1)-chain in K from y to x. Thus we have a cyclic
(ε, 1)-chain passing through x and y. By assuming z = x and z = y, there is
w ∈ Kp such that d (w, z) < δ. If w = xi for some i, we have

d ϕ (tpi , z, ui ) , xi+1 ≤ d (ϕ (tpi , z, ui ) , ϕ (tpi , xi , ui )) +

+d (ϕ (tpi , xi , ui ) , xi+1 ) + d xi+1 , xi+1
ε 1
< + + δ < ε.
4 p
Hence the sequence of points z0 = z, z1 = xi+1 , ..., znp −i = y ∈ K determines
an (ε, 1)-chain from z to y in K. Moreover, we have

d ϕ tpi−1 , xi−1 , ui−1 , z ≤ d ϕ tpi−1 , xi−1 , ui−1 , ϕ tpi−1 , xi−1 , ui−1 +

+d ϕ tpi−1 , xi−1 , ui−1 , xi + d (xi , z)
ε 1
< + +δ <ε
4 p
and then we obtain an (ε, 1)-chain from y to z in K, by concatenation of
chains. Now, if w = ϕ (tpi , xi , ui ) for some i we then have
d (ϕ (tpi , xi , ui ) , z) ≤ d (ϕ (tpi , xi , ui ) , ϕ (tpi , xi , ui )) + d (ϕ (tpi , xi , ui ) , z)

ε
< +δ <ε
4
hence there is an (ε, 1)-chain from y to z in K, by concatenation of chains.
1
On the other hand, since d (w, xi+1 ) = d (ϕ (tpi , xi , ui ) , xi+1 ) < , we have
p

d ϕ tpi+1 , z, ui+1 , xi+2

≤ d ϕ tpi+1 , z, ui+1 , ϕ tpi+1 , w, ui+1 +

+d ϕ tpi+1 , w, ui+1 , ϕ tpi+1 , xi+1 , ui+1 +

+d ϕ tpi+1 , xi+1 , ui+1 , xi+2 + d xi+2 , xi+2
ε ε 1
< + + +δ <ε
4 4 p
where xi+2 = y1 , xi+2 = y1 , tpi+1 = sp0 , and ui+1 = v0 in the case i = np − 1.
By taking a concatenation of chains, we obtain an (ε, 1)-chain from z to y in

K. For the cases w = yj and w = ϕ spj , xj , vj , we analogously construct
an (ε, 1)-chain from z to y in K and an (ε, 1)-chain from y to z in K. By
Proposition 7.1.1, C = C ∪ {z} is a chain transitive set, and then C = C by
the maximality of C. Hence K ⊂ C, and therefore we have an (ε, 1)-chain in
C from x to y. This proves the theorem.
More examples of chain controllable systems are found in the problems.
7.4 Problems
1. Assume that a set C ⊂ M satisfies both conditions (1) and (2) of Defi-
nition 7.1.6. Prove that C is contained in a chain control set.
2. Not assuming M compact, show that the chain control sets are pairwise
disjoint.
3. Let X ⊂ M be a chain transitive set. Check that cl (X) is chain transi-
tive.
on R determined
2
4.
Consider thecontrol system by the vector fields X =
−1 0 1 1
and Y = . Prove that this system is chain
0 −1 −1 1
controllable but not controllable.
5. Consider the control system on R2 determined by the vector fields

X = (0, 1) and Y (x1 , x2 ) = x2 , x1 − x21 . Show that the tear drop
T ⊂ {(x1 , x2 ) : x1 ≥ 0} is a chain control set that is a maximal chain
transitive set not invariant.
6. Let X : R3 → R be the function X (x, y, z) = −y sin2 x + z sin x cos x +
y cos2 x. Consider the unit circle S1 parameterized by the angle x ∈
the control system ẋ = X (x, u (t)) on S with control
1
[0, 2π) and define
1
range U = 0, × [0, 1].
2
7π π8
3π
(a) Show that the sets D1 = , , D2 = , π , D3 = D1 + π,
4 2 4
and D4 = D2 + π are all the main control sets (effective control
sets).
"
4
(b) Show that the sets Dx = {x} with x ∈ S1 \ Di are one-point
i=1
control sets.
(c) Prove that the system is chain controllable.
7. Let X : R2 → R be the function X (x, y) = − sin2 x + b cos2 x − y cos2 x,

where b > 0 is a constant number. Consider the unit circle S1 pa-
rameterized by the angle x ∈ [0, 2π) and define the control system
ẋ = X (x, u (t)) on S1 with control range U = [a, b].
(a) Show that the sets

7 8 $ %
D1 = 0, arctan (b − a) , D2 = π − arctan (b − a), π ,
D3 = D1 + π, D4 = D 2 + π
are all the control sets of the system.

(b) Prove that the control system is chain controllable and then con-
"
4
clude that the points in the set S1 \ cl (Di ) are chain recurrent
i=1
but are contained in no control set.
8. Consider the linear oscillator on R2
ẋ = A0 (x) + u (t) A1 (x)

0 1
with control range U = [−3, 3], where A0 = and A1 =
−1 0

0 0
. The projected system on the projective space P1 is given by
1 0
5 6 5 6
y = A0 − y T A0 yId (y) + u (t) A1 − y T A1 yId (y) .
Show that this system is chain controllable.
9. Consider the bilinear control system on R2

1 0 0 1 0 0
ẋ = (x) + u1 (t) (x) + u1 (t) (x)
0 0 1 0 0 1

1
with control range U = 0, × [1, 2]. Let P1 be parameterized by the
$ π π8 2
angle θ ∈ − , . Prove that the projected system on P1 is chain
2 2
$ π % 7π π8
controllable and has two control sets D1 = − , 0 and D2 = , .
4 4 2
What is the dynamic relation between D1 and D2 ?
10. Consider the control system
ẋ (t) = u1 (t) cos (2x (t)) + u2 (t) f (x (t)) + u3 (t) g (x (t))

3
on the projective space P1 identified with R (mod π), where U = [0, 1] ,
and f (t) , g (t) are C ∞ -functions such that
⎧
⎪ >0 for x ∈ π8 , π2
⎪
⎨
=0 for x = π2
f (x) ,
⎪
⎪ >0 for x ∈ π2 , 7π
⎩ 8
= 0 elsewhere in [0, π)
⎧
⎪
⎪
=0 for x=0

⎨
<0 for x ∈ 0, 3π
g (x) 8 .
⎪
⎪ <0 for x ∈ 5π 8 ,π
⎩
= 0 elsewhere in [0, π)
Show that this system is chain controllable and has two control sets
D1 = [0, π/2] and D2 = (π/2, π). (Hint: Check that all states are
periodic. See [34, Example 4.4.5].)
11. Assume that all Conley attractors and Conley repellers of the control
∗
system are invariant sets. Show that A = (A∗ )∗ and R = (R∗ ) for
every Conley attractor A and every Conley repeller R.
12. Let A1 , A2 ⊂ M be two Conley attractors. Discuss the following sen-

tences and decide if they are true or false:
∗
(a) A1 ∩ A2 is a Conley attractor and (A1 ∩ A2 ) = A∗1 ∪ A∗2 .
∗
(b) (A1 ∪ A2 ) = A∗1 ∩ A∗2 .
13. Let A ⊂ M be a compact Conley attractor and K ⊂ M any compact

neighborhood of A disjoint from A∗ . Show that ω + (K) = A.
14. Let A ⊂ M be an invariant Conley attractor and X ⊂ M a closed

invariant set. Is (A ∩ X, A∗ ∩ X) an attractor-repeller pair in X?
15. Let X ⊂ M be a closed invariant set and (B, B ∗ ) an attractor-repeller

pair in X. Is there an attractor-repeller pair (A, A∗ ) in M such that
(B, B ∗ ) = (A ∩ X, A∗ ∩ X)?
16. A dynamical system φ : R × M → M is called gradient-like if there is

some continuous real valued function which is strictly decreasing on non-
constant trajectories. This generalizes the systems determined by gra-
dient equations of the form ẋ = grad F (x), whose Lyapunov function is
−F . Consider a control affine system on a compact
manifold M and let (A, A∗ ) be a nontrivial attractor-repeller pair of
the system. Take the quotient space Q obtained from M × U by collaps-
∗
ing the attractor-repeller pair L (A) , L (A) to distinct points q1 , q2
and let ρ : M × U → Q be the quotient map. Define φ : R × Q → Q by
φ (t, ρ (x, u)) = ρ (Φ (t, x, u)) t ∈ R, (x, u) ∈ M × U .
Show that φ is a gradient-like dynamical system (with two rest points).
17. Consider a control affine system on a compact manifold admitting finitely

many chain components {C1 , ..., Cn } ordered under the general condi-
tion
xi ∈
/ Ω+ (xk ) if i > k, k = 1, ..., n − 1.
Compare this order with the chain transitivity preorder.

For more information on chain controllability we refer to Colonius–Kliemann
[34], where we can find many other problems on chain control sets which are
not considered in the present book, for instance, bifurcations, dependence on
control ranges, hyperbolicity, multiplicity, and semicontinuity.
Braga Barros and San Martin [20] extended the notion of a chain control
set to the setting of semigroup actions. They considered the specific case of
a system that is induced by a subsemigroup of a transitive Lie group. Under
broad conditions, they proved that a chain control set is the intersection of
control sets for the shadowing semigroups of the system. Thus the chain tran-
sitivity of the control system can be interpreted as the asymptotic transitivity
of some transformation semigroup.
A notion of a gradient-like control system may be obviously defined: the
control system is called gradient-like if there is some continuous real-valued
function that is strictly decreasing on nonconstant solutions. However, quo-
tient space obtained by collapsing sets need not admit a differentiable struc-
ture such that the control system can be induced on it. Therefore, we can not
simply adapt the strategy used in Problem 16 in order to obtain an associated

gradient-like control system; another methodology should be employed.
Chapter 8
Morse decompositions
This chapter is devoted to the central concept in the Conley theory. Morse
decompositions for control systems are studied in two distinct formulations.
We draw attention to their particular properties and then provide a connec-
tion between them. Section 8.1 studies the properties of a dynamic Morse
decomposition, which means an ordered collection of sets containing all limit
sets of points and satisfying the no-cycle condition. The main characterization
of a dynamic Morse decomposition shows that it describes the limit behavior
of the system via its movement from the Morse sets with lower indices to-
ward those with higher ones. Section 8.2 treats the attractor-repeller Morse
decomposition, that is a collection of sets given by intersections of Conley
attractors with complementary repellers. This configuration implies that the
components of an attractor-repeller Morse decomposition produce attractive
and repulsive effects in such a way that every trajectory of the system con-
verges in forward as well as in backward time to some of these components.
Every dynamic Morse decomposition has an attractor-repeller configuration.
On the other hand, any invariant attractor-repeller Morse decomposition is
a dynamic Morse decomposition. In general, however, an attractor-repeller
Morse decomposition need not be invariant. Section 8.3 treats the relation-
ship between Morse decomposition and chain transitivity. By the Conley main
theorem on chain recurrence, every Morse decomposition contains the chain
recurrence set. This implies that the existence of the finest Morse decomposi-
tion is equivalent to the existence of finitely many chain components. The last
section presents a Lyapunov function that describes the global structure of a
control system in the sense that it is strictly decreasing on trajectories outside
251
252 Chapter 8 · Morse decompositions
the chain recurrence set and maps each chain component onto a critical value.
8.1 Dynamic Morse decompositions

In this section we define the concept of a dynamic Morse decomposition of a
control system. Along this chapter, we assume a control affine system on a
compact manifold M .
It should be remembered that a set X ⊂ M is isolated invariant if every
orbit that remains close to X actually belongs to X.
Definition 8.1.1 A dynamic Morse decomposition for the control affine

system on M is a finite collection M = {C1 , ..., Cn } of nonempty, pairwise
disjoint, isolated invariant, and compact sets such that:
1. For all x ∈ M there are integers i, j such that ω + (x) ⊂ Ci and ω − (x) ⊂
Cj ;
2. (No-cycle condition) If there are sequences Cj0 , Cj1 , ..., Cjl and x1 , ..., xl
"n
in M \ Ci with ω − (xk ) ⊂ Cjk−1 and ω + (xk ) ⊂ Cjk , for k = 1, ..., l;
i=1
then Cj0 = Cjl .
Each set Ci is called a dynamic Morse set (or simply a Morse set).
A dynamic Morse decomposition M contains all limit sets of points and

"
n
“cycles” are not allowed. In the case x ∈ M \ Ci , we have ω + (x) , ω − (x) ⊂
i=1
"
n
Ci , although the orbit of x does not intersect the Morse sets. Otherwise,
i=1
the orbit O (x) and the limit sets ω + (x) , ω − (x) are contained in Ci whenever
x ∈ Ci .
Proposition 8.1.1 Let M = {Ci : i = 1, ..., n} be a dynamic Morse decom-

position. If ω + (x) ∪ ω − (x) ⊂ Cj then x ∈ Cj .
"
n
Proof. Suppose by contradiction that x ∈ M \ Ci . By the no-cycle
i=1
condition, ω − (x) ⊂ Cj and ω + (x) ⊂ Cj imply Cj = Cj , which is a contra-
diction. Therefore x ∈ Ck , for some k ∈ {1, ..., n}. We claim that Ck = Cj .
Indeed, as Ck is compact and invariant, it follows that ω + (x) , ω − (x) ⊂ Ck .
Since the limit sets are nonempty and the dynamic Morse sets are pairwise
Section 8.1 · Dynamic Morse decompositions 253
disjoint, we have Ck = Cj , and therefore x ∈ Cj .
This result implies a dynamic order between the Morse sets, as in the
following:
Definition 8.1.2 Let {C1 , ..., Cn } be a collection of nonempty, compact, iso-

lated invariant, and pairwise disjoint sets of M . Then Cα Cβ if there are se-
quences Cj0 = Cα , Cj1 , ..., Cjm = Cβ and x1 , ..., xm ∈ M with ω − (xk ) ⊂ Cjk−1
and ω + (xk ) ⊂ Cjk , for k = 1, ..., m.
Proposition 8.1.2 The relation “” defined above is an order among the
components of a dynamic Morse decomposition M = {C1 , ..., Cn }.
Proof. Take x ∈ Ci . Since Ci is compact and invariant, it follows that

−
ω (x) ⊂ Ci and ω + (x) ⊂ Ci . Thus Ci Ci , for all i = 1, ..., n. Now, suppose
that Cα Cβ and Cβ Cα . We obtain Cj0 = Cα , Cj1 , ..., Cjm = Cα and
x1 , ..., xm ∈ M with ω − (xk ) ⊂ Cjk−1 and ω + (xk ) ⊂ Cjk , for k = 1, ..., m,
where Cβ = Cjk , for some k. Since Cj0 = Cα = Cjm , the no-cycle condition
"
n
implies that there is some xkl ∈ Ci . Thus xkl ∈ Cl , for some l ∈ {1, ..., n},
i=1
and therefore, ω − (xkl ) ∪ ω + (xkl ) ⊂ Cl . However, ω − (xkl ) ⊂ Cjkl −1 and
ω + (xkl ) ⊂ Cjkl , and since the dynamic Morse sets are pairwise disjoint, we
have Cjkl −1 = Cl = Cjkl . Continuing with this argument, we can show that
Cα = Cjk , for all k = 1, ..., m, and then Cα = Cβ . The transitivity property
of “” follows immediately from the definition.
The following theorem characterizes the dynamic Morse decompositions.
Theorem 8.1.1 A collection {Ci : i = 1, ..., n} of nonempty, compact, iso-

lated invariant, and pairwise disjoint sets of M is a dynamic Morse decom-
position if and only if it satisfies the following properties:
"
n
1. ω + (x) , ω − (x) ⊂ Ci for every x ∈ M .
i=1
2. ω + (x) ∪ ω − (x) ⊂ Ci implies x ∈ Ci .
3. the relation “” defined above is an order among the sets Ci .
Proof. Suppose that the collection {Ci : i = 1, ..., n} satisfies the proper-
ties (1) − (3). To show that it is a dynamic Morse decomposition, it is enough
to prove that it satisfies the no-cycle condition. Indeed, suppose that there are
"
n
sequences Cj0 , Cj1 , ..., Cjl and x1 , ..., xl ∈ M \ Ci with ω − (xk ) ⊂ Cjk−1 and
i=1
ω + (xk ) ⊂ Cjk , for k = 1, ..., l, and Cj0 = Cjl . For all k = 1, ..., l − 1, we have
Cj0 Cjk and Cjk Cjl = Cj0 . Since “” is an order according to property
(3), this implies that Cjk = Cj0 , for all k = 1, ..., l. Thus ω − (xk ) ⊂ Cj0 and
ω + (xk ) ⊂ Cj0 , for all k = 1, ..., l. According to property (2), it follows that
"
n
xk ∈ Cj0 , which contradicts xk ∈ M \ Ci . Hence Cj0 = Cjl , and therefore
i=1
the no-cycle condition is satisfied. The converse follows by Propositions 8.1.1
and 8.1.2.
Enumerating the dynamic Morse sets Ci in such a way that Ci Cj

implies i ≤ j, the dynamic Morse decomposition describes the limit behavior
of the control system via its movement from the sets Ci with lower indices
toward those with higher ones.
In the following we show that the lift of a dynamic Morse decomposition
is a dynamic Morse decomposition of the control flow.
Proposition 8.1.3 If M = {C1 , ..., Cn } is a dynamic Morse decomposition

for the control system in M , then L (M) = {L (C1 ) , ..., L (Cn )} is a dynamic
Morse decomposition for the control flow in M × U .
Proof. Since each Morse set Ci is invariant, we have L (Ci ) = Ci × U .

It can be easily seen that L (M) = {L (C1 ) , ..., L (Cn )} is a collection of
nonempty, compact, invariant, and pairwise disjoint sets in M × U . For each
i, let Vi be a compact isolating neighborhood of Ci such that Vi ∩ Vj =
∅, for i = j. Suppose that (x, u) ∈ Vi × Ui and Φ (t, x, u) ∈ Vi × U for
all t ∈ R. It follows that ω − (x, u) , ω + (x, u) ⊂ Vi × U , as Vi is compact.
Hence π (ω − (x, u)) ∪ π (ω + (x, u)) ⊂ Vi , with π (ω − (x, u)) ⊂ ω − (x) and
π (ω + (x, u)) ⊂ ω + (x). This means that ω − (x) ∩ Vi and ω + (x) ∩ Vi are
nonempty, which implies ω − (x) , ω + (x) ⊂ Ci . Thus x ∈ Ci and (x, u) ∈
Ci × U . This proves that Ci × U is an isolated invariant set. Now, for any
(x, u) ∈ M × U , we have
#
n
π ω − (x, u) ∪ π ω + (x, u) ⊂ ω − (x) ∪ ω + (x) ⊂ Ci
i=1
−
"n
hence ω (x, u) , ω (x, u) ⊂ i=1 Ci × U . Suppose there are sequences Cj0 ×
+
"n
U , Cj1 × U , ..., Cjl × U and (x1 , u1 ) , ..., (xl , ul ) ∈ M × U \ i=1 Ci × U with
Section 8.1 · Dynamic Morse decompositions 255
ω − (xk , uk ) ⊂ Cjk−1 × U and ω + (xk , uk ) ⊂ Cjk × U , for k = 1, ..., l. It follows

that ω − (xk ) ⊂ Cjk−1 and ω + (xk ) ⊂ Cjk for k = 1, ..., l. Hence Cj0 = Cjl ,
and therefore Cj0 × U = Cjl × U . This proves that L (M) is a dynamic Morse
decomposition for the control flow in M × U .
One dynamic Morse decomposition may be refined by an other. In this

sense, we may discuss the existence of the finest dynamic Morse decomposi-
tion. We introduce this notion in the following:
Definition 8.1.3 A dynamic Morse decomposition M = {C1 , ..., Cn } is said

to be finer than a dynamic Morse decomposition M = {C1 , ..., Cm
} if for

each Morse set Cj there is a Morse set Ci with Ci ⊂ Cj . A dynamic Morse
decomposition is called the finest dynamic Morse decomposition if it is
finer than any other dynamic Morse decomposition.

Let M = {Ci : i = 1, ..., n} and M = Cj : j = 1, ..., m be two dynamic
Morse decompositions of the control system. For the following, we denote by
M ∩ M the collection

M ∩ M = Ci ∩ Cj : i, j
where only those indices i and j with Ci ∩ Cj = ∅ are allowed.

Proposition 8.1.4 If M = {Ci : i = 1, ..., n} and M = Cj : j = 1, ..., m
are dynamic Morse decompositions then M ∩ M is a dynamic Morse decom-
position finer than both M and M .
Proof. It follows immediately from the definition that the sets Ci ∩ Cj
are nonempty, compact, invariant, and pairwise disjoint. Let Vi ⊂ M be
an isolating neighborhood of Ci and Vj ⊂ M an isolating neighborhood of
Cj . Them Vi ∩ Vj is a neighborhood of Ci ∩ Cj . Moreover, if x ∈ Vi ∩ Vj and
O (x) ⊂ Vi ∩Vj , then x ∈ Ci ∩Cj , hence Vi ∩Vj is an isolating neighborhood of
Ci ∩Cj . Therefore Ci ∩Cj is isolated invariant for all i, j. Now, for every x ∈ M
there are i, j such that ω + (x) ⊂ Ci and ω + (x) ⊂ Cj . Thus ω + (x) ⊂ Ci ∩ Cj .
Analogously, there are k, l such that ω − (x) ⊂ Ck ∩ Cl . Therefore M ∩ M
satisfies condition (1) of Definition 8.1.1. Finally, suppose there are sequences
"
m
Ck0 ∩ Ck 0 , Ck1 ∩ Ck 1 , ..., Ckm ∩ Ck m and x1 , ..., xm ∈ M \ Ckl ∩ Ck l with
l=0
ω − (xl ) ⊂ Ckl−1 ∩ Ck l−1 and ω + (xl ) ⊂ Ckl ∩ Ck l , for l = 1, ..., m. Since
Ckl ∩ Ck l ⊂ Ckl , for all l = 1, ..., m, we have Ck0 = Ckm , and hence Ck0 and
Ckm are disjoint. Thus Ck0 ∩Ck 0 = Ckm ∩Ck m , showing that M∩M satisfies
condition (2) of a dynamic Morse decomposition.
We now show that M∩M is finer than both M and M . In fact, take any
Ci in M and pick x ∈ Ci . Then ω + (x) ⊂ Ci . There exists some j such that
ω + (x) ⊂ Cj . Hence ω + (x) ⊂ Ci ∩ Cj , which means that Ci ∩ Cj ∈ M ∩ M
and Ci ∩ Cj ⊂ Ci . Thus M ∩ M is finer than M. Analogously, we may show
that M ∩ M is finer than M .
By taking successive refinements we want to know if this process stabilizes

in a minimal dynamic Morse decomposition. The following result responds to
this question:
Proposition 8.1.5 If the control system admits the finest dynamic Morse
decomposition then it is the intersection of all dynamic Morse decomposition.
Proof. Let {Mλ }λ∈Λ be the family of all dynamic Morse decomposition
and suppose that F = {Fi : i = 1, ..., n} is the finest Morse decomposition of
the control system. For each λ, choose Cλ ∈ Mλ . There then exists iλ such

that Fiλ ⊂ Cλ . As F ∈ {Mλ }λ∈Λ , we have Fi = iλ =i Cλ , for all i = 1, ..., n.

Thus F = λ∈Λ Mλ .
However, the intersection of all dynamic Morse decompositions need not

be a dynamic Morse decomposition. See Example A.2.2 in Appendix A for a
typical case of nonexistence of the finest Morse decomposition.
Example 8.1.1 Consider the control affine system on the unit disk M =

x ∈ R2 : x ≤ 1 as defined in Example 5.7.1. The collection M = S1 , {0}
is the unique nontrivial dynamic Morse decomposition of the control system
on M . The corresponding (dynamic) Morse decomposition of the control flow

is L (M) = S1 × U , {0} × U .
Example 8.1.2 Go back to Example 6.3.2 and identify the unit circle S1 with
[0, 2π). The finest dynamic Morse decomposition of the control system is M =

{0} , π2 , {π} , 3π
2 , and the corresponding finest Morse decomposition of
the control flow is
-π. ! !
3π
L (M) = {0} × U , × U , {π} × U , ×U .
2 2
Section 8.2 · Attractor-repeller Morse decomposition 257
8.2 Attractor-repeller Morse decomposition

We now present the attractor-repeller configuration of a Morse decompo-
sition. We investigate its main properties and discuss its connection with a
dynamic Morse decomposition.
If ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M is an increasing sequence of Conley
attractors then ∅ = A∗n ⊂ A∗n−1 ⊂ · · · ⊂ A∗0 = M is an increasing sequence
of complementary repellers. In fact, Ai−1 ⊂ Ai implies Au (Ai−1 ) ⊂ Au (Ai ),
hence
A∗i = M \ Au (Ai ) ⊂ M \ Au (Ai−1 ) = A∗i−1
for every i = 1, ..., n.
Definition 8.2.1 Let ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M be an increasing se-

quence of Conley attractors in M and define Cn−i = Ai+1 ∩A∗i , i = 0, · · · , n−
1. Then the ordered collection M = {C1 , · · · , Cn } is called an attractor-
repeller Morse decomposition of the control system. Each component Ci
is called an attractor-repeller Morse set (or simply a Morse set).
As in the case of dynamic Morse decompositions, we define the notion of

a finer attractor-repeller Morse decomposition.
Note that the first attractor-repeller Morse set C1 is the complementary
repeller A∗n−1 and the last one Cn is the Conley attractor A1 . In general,
since Conley attractors are positively invariant and complementary repellers
are invariant, the attractor-repeller Morse sets are positively invariant. The
following properties of the attractor-repeller Morse decompositions approach
the properties of the dynamic Morse decompositions.
Proposition 8.2.1 Let M = {C1 , ..., Cn } be an attractor-repeller Morse de-

composition with Cn−i = Ai+1 ∩ A∗i , for 0 ≤ i ≤ n − 1, where ∅ = A0 ⊂
A1 ⊂ · · · ⊂ An = M is an increasing sequence of Conley attractors. Then M
satisfies the following properties:
1. The sets Ci are nonempty, compact, and pairwise disjoint.
2. For every x ∈ M there are integers i, j such that ω + (x) ⊂ Ci and

ω − (x) ⊂ Cj .

Proof. (1) Each attractor-repeller Morse set Ci is compact, since it is

an intersection of compact sets. The lift L (Ai+1 ) is a Conley attractor and
∗
L (A∗i ) = L (Ai ) is a Conley repeller for the control flow (Propositions 7.2.5
and 7.2.6). Since L (Ai ) ⊂ L (Ai+1 ), we have L (Ai+1 ) ∩ L (A∗i ) = ∅, and then
Cn−i = Ai+1 ∩ A∗i = ∅ according to Theorem 7.2.4. Now, take i, j ∈ {1, ..., n}
such that i = j and suppose that i < j. Then i + 1 ≤ j and hence
Cn−i ∩ Cn−j = Ai+1 ∩ A∗i ∩ Aj+1 ∩ A∗j = Ai+1 ∩ A∗j ⊂ Ai+1 ∩ A∗i+1 = ∅.
Therefore Cn−i ∩ Cn−j = ∅.
(2) For x ∈ M , let i be the smallest integer such that ω + (x) ⊂ Ai . Then
i > 0 and ω + (x) Ai−1 . This means that ω + (x) ⊂ A∗i−1 , according to
Proposition 7.2.4. Thus ω + (x) ⊂ Ai ∩ A∗i−1 = Cn−i+1 . Similarly, let j
be the largest integer such that ω − (x) ⊂ A∗j . Then ω − (x) A∗j+1 , hence
ω − (x) ⊂ Aj+1 . Thus ω − (x) ⊂ Aj+1 ∩ A∗j = Cn−j .
(3) Suppose that ω + (x) ∪ ω − (x) ⊂ Cn−i = Ai+1 ∩ A∗i . According to
Proposition 7.2.4, it follows that x ∈ Cn−i .
Each dynamic Morse decomposition has an attractor-repeller configura-

tion, as follows:
Theorem 8.2.1 Every dynamic Morse decomposition M = {C1 , ..., Cn } is
an attractor-repeller Morse decomposition.
Proof. According to Proposition 8.1.3, the collection
L (M) = {C1 × U , ..., Cn × U} is a dynamic Morse decomposition for the
control flow in M ×U . For flows, dynamic Morse decomposition and attractor-
repeller Morse decomposition are equivalent concepts. Thus there is an in-
creasing sequence of Conley attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M × U
such that Cn−i × U = Ai+1 ∩ A∗i for 0 ≤ i ≤ n − 1. According to Theorem
7.2.4, ∅ = π (A0 ) ⊂ π (A1 ) ⊂ · · · ⊂ π (An ) = M is an increasing sequence
of Conley attractors for the control system, with π (Ai ) ∩ π (A∗i ) = ∅. If
∗
x∈ / π (Ai ) then ω + (x) ⊂ π (Ai ). As π ω + π −1 (x) = ω + (x), it follows

that ω + π −1 (x) ⊂ π −1 (π (Ai )), hence π −1 (x) ∩ A∗i = ∅. This means that
∗ ∗
x∈/ π (A∗i ), and therefore π (A∗i ) ⊂ π (Ai ) . On the other hand, if y ∈ π (Ai )
∗
then ω (y) ⊂ π (Ai ) by invariance. Hence ω (y) ∩ π (Ai ) = ∅, which means
+ +

π ω + π −1 (y) ∩ π (Ai ) = ∅. It follows that ω + π −1 (y) ∩ Ai = ∅, and
∗
thus π −1 (y) ⊂ A∗i , that is, y ∈ π (A∗i ). Then π (Ai ) ⊂ π (A∗i ), and therefore
∗ ∗
π (Ai ) = π (Ai ). We now have
∗
Cn−i ⊂ π (Ai+1 ) ∩ π (A∗i ) = π (Ai+1 ) ∩ π (Ai )
Section 8.2 · Attractor-repeller Morse decomposition 259
∗
for 0 ≤ i ≤ n − 1. We claim that Cn−i = π (Ai+1 ) ∩ π (Ai ) . Indeed, if
∗ ∗
x ∈ π (Ai+1 ) ∩ π (Ai ) then ω + (x) ⊂ π (Ai+1 ) ∩ π (Ai ) by the positive in-
variance of the attractor-repeller Morse set. This implies that ω + (x) ⊂ Cn−i .
Take u ∈ U such that (x, u) ∈ Ai+1 . By the invariance of Ai+1 , we have
ω − (x, u) ⊂ Ai+1 , and then ω − (x) ∩ π (Ai+1 ) = ∅. According to Proposition
∗
7.2.4, it follows that ω − (x) ⊂ π (Ai+1 ). As ω − (x) ⊂ π (Ai ) by the invari-
∗
ance of the complementary repeller, we have ω − (x) ⊂ π (Ai+1 ) ∩ π (Ai ) .
−
This means that ω (x) , ω (x) ⊂ Cn−i , and hence x ∈ Cn−i by the property
+
∗
of the Morse set. Therefore Cn−i = π (Ai+1 ) ∩ π (Ai ) , for 0 ≤ i ≤ n − 1.
It is clear that the dynamic Morse decompositions differ from the attractor-
repeller Morse decompositions by the invariance condition on the Morse sets.
In fact, an attractor-repeller Morse decomposition need not be invariant (see
Example 8.2.1). Otherwise, any invariant attractor-repeller Morse decompo-
sition is a dynamic Morse decomposition, as in the following:
Theorem 8.2.2 Assume that M = {C1 , ..., Cn } is an attractor-repeller Morse

decomposition with Cn−i = Ai+1 ∩ A∗i , for 0 ≤ i ≤ n − 1, where ∅ = A0 ⊂
A1 ⊂ · · · ⊂ An = M is an increasing sequence of invariant Conley attractors.
Then M is a dynamic Morse decomposition.
Proof. According to Proposition 6.3.1, both the sets Ai and A∗i are iso-
lated invariant. Then the attractor-repeller Morse sets C1 , ..., Cn are iso-
lated invariant, since they are intersections of isolated invariant sets. We
claim that each Ci is nonempty. In fact, by taking x ∈ Ai+1 \ Ai , we have
ω + (x) ∪ ω − (x) ⊂ Ai+1 , because Ai+1 is invariant, and ω + (x) ∪ ω − (x)
Ai , according to Proposition 7.2.4. Then at least ω − (x) ⊂ A∗i , and hence
ω − (x) ⊂ Ai+1 ∩ A∗i = Cn−i . This proves that every attractor-repeller Morse
set is nonempty. According to Proposition 8.2.1, item (1), M is a collection
of nonempty, compact, invariant isolated, and pairwise disjoint sets in M .
Now, according to Proposition 8.2.1, items (2) and (3), M satisfies items (1)
and (2) of Theorem 8.1.1. Thus it remains to show that the relation “” is
an order in M. It is enough to show that the conditions ω − (x) ⊂ Cn−i and
ω + (x) ⊂ Cn−j , with i = j, imply n−i < n−j, that is, i > j. Indeed, suppose
by contradiction that i < j. Then
ω − (x) ⊂ Cn−i = Ai+1 ∩ A∗i ⊂ Aj

ω + (x) ⊂ Cn−j = Aj+1 ∩ A∗j ⊂ A∗j
and hence x must lie in Aj ∪A∗j , according to Proposition 7.2.4. If x ∈ Aj then

ω + (x) ⊂ Aj , which contradicts ω + (x) ⊂ A∗j . If x ∈ A∗j then ω − (x) ⊂ A∗j ,
which contradicts ω − (x) ⊂ Aj . Therefore i > j.
Thus, if all Conley attractors of the control system are invariant, the con-
cepts of dynamic and attractor-repeller Morse decompositions are equivalent.
Theorem 8.2.3 Assume that all Conley attractors of the control system are
invariant. Then a finite collection M = {C1 , ..., Cn } of sets in M is a dy-
namic Morse decomposition if and only if it is an attractor-repeller Morse
decomposition.
Example 8.2.1 Consider the control affine system on the unit circle as de-
fined in Example 3.1.1. The left half semicircle A = [π/2, 3π/2] is the unique
nontrivial Conley attractor of the system, and its complementary repeller is
A∗ = {0}. Hence the attractor-repeller pair (A, A∗ ) defines the unique non-
trivial attractor-repeller Morse decomposition of the system. Since A is not
negatively invariant, (A, A∗ ) is not a dynamic Morse decomposition.
Example 8.2.2 In Example 5.7.1, the unit circle S1 is the unique

∗
nontrivial Conley attractor of the control system on the unit disk, and S1 =
{0}. Then the sequence of attractors ∅ ⊂ S1 ⊂ M describes the unique non-

trivial attractor-repeller Morse decomposition M = S1 , {0} . In this case,
M is a dynamic Morse decomposition.
Example 8.2.3 In Example 6.3.2, we have the sequence of Conley attractors
A0 = ∅ ⊂ A1 = {0} ⊂ A2 = {0, π} ⊂ A3 = [0, π] ⊂ A4 = M
with corresponding sequence of complementary repellers
A∗a = ∅ ⊂ A∗3 = {3π/2} ⊂ A∗2 = {π/2, 3π/2} ⊂ A∗1 = [π/2, 3π/2] ⊂ A∗0 = M.
This is an attractor-repeller configuration for the finest Morse decomposition

of the system. In fact, we have
C1 = C4−3 = A4 ∩ A∗3 = {3π/2} ,

C2 = C4−2 = A3 ∩ A∗2 = {π/2} ,
C3 = C4−1 = A2 ∩ A∗1 = {π} ,
C4 = C4−0 = A1 ∩ A∗0 = {0} .
Section 8.3 · Morse decomposition and chain recurrence 261
8.3 Morse decomposition and chain recurrence

We shall now show the relationship between the chain transitive sets and
the Morse decompositions. We discuss the existence of the finest Morse de-
composition in view of the cardinality of the chain components of the control
system.
Following Theorem 8.2.2, we use the name ‘Morse decomposition’ instead
of ‘attractor-repeller Morse decomposition’ and ‘invariant Morse decomposi-
tion’ instead of ‘dynamic Morse decomposition’.
Proposition 8.3.1 Every Morse decomposition contains the chain recurrence

set of the control system.
Proof. Let M = {C1 , · · · , Cn } be a Morse decomposition defined by a

sequence of Conley attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M . According to
Theorem 7.3.2, it is enough to show that
#
n
n
Ci = Ai ∪ A∗i .
i=1 i=1
"n
Indeed, for x ∈ i=1 Ci and j ∈ {1, ..., n}, suppose that x ∈ / Aj . Take
k ∈ {1, ..., n} such that x ∈ Ak ∩ A∗k−1 . It follows that k > j and k − 1 ≥ j.
Hence x ∈ A∗k−1 ⊂ A∗j . Therefore x ∈ Aj ∪ A∗j . On the other hand, take
n
y ∈ i=1 Ai ∪ A∗i and let j be the smaller integer such that y ∈ Aj . Since
y∈/ Aj−1 , we have y ∈ A∗j−1 . Hence y ∈ Aj ∩ A∗j−1 .
Combining Propositions 8.1.5 and 8.3.1 with the Conley main theorem,
the existence of the finest Morse decomposition must be equivalent to the
existence of only finitely many chain components, as the following.
Theorem 8.3.1 The finest Morse decomposition exists if and only if the
chain recurrence set has only finitely many chain components. In this case,
the chain components form the finest Morse decomposition.
Proof. Suppose that M = {C1 , · · · , Cn } is the finest Morse decomposi-

tion determined by the sequence of Conley attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂
An = M . For a given x ∈ Ci , take a Conley attractor A containing ω + (x).
We claim that Ci ⊂ A. In fact, let V and Vi be attractor neighborhoods of A
and Ai respectively. Since
ω + (V ∩ Vi ) ⊂ A ∩ Ai ⊂ V ∩ Vi
the set B = ω + (V ∩ Vi ) is a Conley attractor. According to Proposition

6.3.2, Ai−1 ∪ B is a Conley attractor. Now, consider the increasing sequence
of Conley attractors
A0 ⊂ A1 ⊂ · · · ⊂ Ai−1 ⊂ Ai−1 ∪ B ⊂ Ai ⊂ ... ⊂ An .
Since M is the finest Morse decomposition, there is Cj such that Cj ⊂

(Ai−1 ∪ B) ∩ A∗i−1 = B ∩ A∗i−1 . Then B ∩ A∗i−1 ⊂ Ai ∩ A∗i−1 . Since Morse
sets are pairwise disjoint, we have Cj = Ci . Hence Ci ⊂ B ∩ A∗i−1 , and there-
fore Ci ⊂ A, proving the claim. According to Theorem 7.2.2, it follows that
Ci ⊂ Ω+ (x), thus Ci is chain transitive. Now, since ω + (x) ⊂ Ai , we have
Ω+ (x) ⊂ Ai , and since x ∈ A∗i−1 , it follows that Ω− (x) ⊂ A∗i−1 , according
to Proposition 7.3.1. Hence Ω+ (x) ∩ Ω− (x) ⊂ Ai ∩ A∗i−1 = Ci , and since
Ci is chain transitive, we conclude that Ci = Ω+ (x) ∩ Ω− (x). Therefore Ci
is a maximal chain transitive set. The conclusion follows according to The-
orem 7.3.2. As to the converse, suppose that {C1 , ..., Cn } is the set of all
chain components of the system. This is a collection of compact, invariant,
and pairwise disjoint sets. Since the limit sets are chain transitive, we have
"n
ω + (x) , ω − (x) ⊂ i=1 Ci for all x ∈ M . Suppose that ω + (x) ∪ ω − (x) ⊂ Ci .
Then x ∈ / Cj , if j = i, because each Cj is invariant and disjoint from Ci .
By taking y ∈ ω + (x) and z ∈ ω − (x), we have Ω+ (y) = Ω+ (z), y ∈ Ω+ (x),
and z ∈ Ω− (x), hence x ∈ Ω+ (x) by the transitive property of the chain
transitivity relation. The unique possibility for this fact is x ∈ Ci . Now, take
a compact neighborhood Vi of Ci such that Vi ∩ Cj = ∅ if j = i. If x ∈ Vi
and O (x) ⊂ Vi then ω + (x) ∪ ω − (x) ⊂ Vi , hence ω + (x) ∪ ω − (x) ⊂ Ci . This
implies that x ∈ Ci , and therefore Ci is isolated invariant. It remains to show
that the relation of Definition 8.1.2 is an order among the chain components.
Indeed, consider the chain components enumerated as follows. First, write
Ci = Ω+ (xi ) ∩ Ω− (xi ). Then for some xi we have xj ∈ / Ω+ (xi ) if j = i, since
otherwise we would have a cycle
xi1 ∈ Ω+ (xi2 ) , ..., xik ∈ Ω+ (xi1 )
which means Ci1 = · · · = Cik with i1 , ..., ik pairwise distinct. We may change
the order and consider xi ∈ / Ω+ (x1 ) if i > 1. For the same reason, for some
i > 1 we have xj ∈ / Ω (xi ) if j > 1 and j = i, and hence we may consider
+
xi ∈
/ Ω+ (x2 ) if i > 2. Following this process by induction we obtain the
ordered collection of chain components {C1 , ..., Cn } with the general condition
xi ∈
/ Ω+ (xk ) if i > k, k = 1, ..., n − 1.
Section 8.3 · Morse decomposition and chain recurrence 263
Completing the proof, it is enough to show that ω − (x) ⊂ Ci and ω + (x) ⊂ Cj ,

"n
with x ∈ M \ i=1 Ci , implies j ≤ i. In fact, by taking y ∈ ω − (x) and
z ∈ ω + (x), we have y ∈ Ω− (x) ∩ Ω+ (xi ) and z ∈ Ω+ (x) ∩ Ω− (xj ), hence
xj ∈ Ω+ (z) , z ∈ Ω+ (x) , x ∈ Ω+ (y) , y ∈ Ω+ (xi ) ,
and therefore xj ∈ Ω+ (xi ). This is possible only if j ≤ i, as desired.
We shall finish this section by presenting the relationship between the

chain components of the control system and the chain components of the
control flow.
Proposition 8.3.2 A point x ∈ M is a chain recurrent point of the control

system if and only if there exists a chain recurrent point of the control flow in
the fiber π −1 (x).
Proof. Suppose that x is a chain recurrent point of the control sys-

tem. Let (x, u) ∈ π −1 (x) and take a Conley attractor A ⊂ M × U such
that ω + (x, u) ⊂ A. According to Proposition 4.1.8 and Theorem 7.2.4,
it follows that π (A) ⊂ M is a Conley attractor of the control system and
ω + (x) ∩ π (A) = ∅. This means that ω + (x) ⊂ π (A), and then x ∈ Ω+ (x) ⊂
π (A) according to Theorem 7.2.2. In particular, we have x ∈ Ω+ (x) ⊂
π (ω + (Ω+ ((x, u) , , T ))), for all , T > 0, where ω + (Ω+ ((x, u) , , T )) is a
Conley attractor with attractor neighborhood Ω+ ((x, u) , , T ). Hence, for
each n ∈ N, we can take un ∈ U such that (x, un ) ∈ Ω+ ((x, u) , 1/n, 1),
and we may assume that un → v in U . For a given > 0, there is n such
that 1/n < /2 and d ((x, v) , (x, un )) < /2. As (x, un ) ∈ Ω+ ((x, u) , 1/n, 1),
we have (x, v) ∈ Ω+ ((x, u) , , 1). Thus (x, v) ∈ Ω+ ((x, u)). Now we show
that there exists a chain recurrent point in the fiber π −1 (x). Start with a
point (x, u1 ) and take (x, u2 ) ∈ Ω+ ((x, u1 )) as in the first part. Following
by induction, we obtain a sequence (un ) in U with (x, un+1 ) ∈ Ω+ ((x, un ))
for all n. We may assume that un → u in U . For a given > 0, take un
such that d ((x, u) , (x, un )) < /2. Since (x, un ) ∈ Ω+ ((x, un−1 ) , /2, 1), we
have (x, u) ∈ Ω+ ((x, un−1 ) , , 1). Hence, for each k ∈ N, we can choose
unk with nk > k and (x, u) ∈ Ω+ ((x, unk ) , 1/k, 1). Note that the sub-
sequence (unk ) also converges to u. For a given > 0 there is δ > 0
such that d ((y, v) , (z, w)) < δ implies d (Φ (t, y, v) , Φ (t, z, w)) < /2 for all
t ∈ [1, 2]. Choose k such that 1/k < /2 and d ((x, u) , (x, unk )) < δ. As
(x, u) ∈ Ω+ ((x, unk ) , 1/k, 1), we can take an (1/k, 1)-chain from (x, unk ) to
(x, u) formed by sequences (x0 , v0 ) = (x, unk ) , (x1 , v1 ) , ..., (xl , vl ) = (x, u)
in M × U and t0 , ..., tl−1 in [1, 2]. Since d ((x, u) , (x, unk )) < δ, we have
d (Φ (t0 , x, u) , Φ (t0 , x, unk )) < ε/2. As d (Φ (t0 , x, unk ) , (x1 , v1 )) < 1/k <
/2, it follows that d (Φ (t0 , x, u) , (x1 , v1 )) < ε. Hence the sequences
(x, u) , (x1 , v1 ) , ..., (xl , vl ) = (x, u) in M × U and t0 , ..., tl−1 in [1, 2] form an
(, 1)- chain from (x, u) to (x, u). Therefore (x, u) ∈ Ω+ ((x, u)), as desired.
The converse follows according to Proposition 7.1.8.
The following shows that there is an one-to-one correspondence between

the chain components of the control system and the chain components of the
control flow.
Proposition 8.3.3 For each chain component C ⊂ M of the control system

there exists a unique chain component C ⊂ M × U of the control flow such
that π (C) = C.
Proof. Let x, y ∈ C. According to Proposition 8.3.2, there are u, v ∈ U

such that (x, u) and (y, v) are chain recurrent points of the control flow.
Let A ⊂ M × U be a Conley attractor with ω + (x, u) ⊂ A. According
to Theorem 7.2.4, A has an attractor neighborhood of the form V × U .
Since ω + (x) ⊂ π (A), it follows that y ∈ Ω+ (x) ⊂ π (A) ⊂ V , hence
π −1 (y) ⊂ V × U . This means that ω + (y, v) ⊂ ω + (V × U ) = A, and thus
(y, v) ∈ Ω+ (y, v) ⊂ A. Since A is arbitrary, we have (y, v) ∈ Ω+ (x, u). In
the same way, we can show that (x, u) ∈ Ω+ (y, v), hence (x, u) and (y, v) are
in the same chain component C ⊂ M × U. Since x, y ∈ C are arbitrary, C is
the unique chain component of the control flow with π (C) = C.
As a consequence of Theorem 8.3.1 and Proposition 8.3.3, we have the

following:
Corollary 8.3.1 The control system has the finest Morse decomposition if
and only if the control flow has the finest Morse decomposition.

Example 8.3.1 In Example 5.7.1, M = S1 , {0} is the finest Morse de-
composition of the control system on the unit disk, hence S1 , {0} are the chain

components of the system. The lift L (M) = S1 × U , {0} × U is the finest
Morse decomposition of the control flow and S1 × U, {0} × U are its chain
components.
Section 8.4 · Complete Lyapunov functions 265
Example 8.3.2 Go back to Example 6.3.2 and identify the unit circle S1
with [0, 2π). The finest dynamic Morse decomposition of the control system is

M = {0} , π2 , {π} , 3π 2 , hence the single sets {0} , π2 , {π} , 3π
2 are
the chain components of the control system. The corresponding finest Morse
decomposition of the control flow is
-π. ! !
3π
L (M) = {0} × U , × U , {π} × U , ×U
2 2
π 3π
and therefore the fibers {0} × U, 2 × U, {π} × U, 2 × U are the chain
components of the control flow.
8.4 Complete Lyapunov functions

We finish the book by studying the complete Lyapunov functions of control
systems. The purpose is to determine a continuous real-valued function that
describes the global structure of a control system. This means that the chain
components are connected by orbits which go through decreasing levels of
some Lyapunov function. The manifold M is assumed to be compact.
Definition 8.4.1 A Morse-Lyapunov function associated to a Morse de-

composition M = {C1 , ..., Cn } is a continuous real-valued function LM : M →
"n
R that is strictly decreasing on trajectories outside i=1 Ci and for each Ci
"n
there is ci ∈ R such that L−1M (ci ) = Ci . The set LM ( i=1 Ci ) is called the
critical value set of LM .
"n
The function LM to be strictly decreasing on trajectories outside Ci
"n i=1
means that LM (ϕ (t, x, u)) < LM (x) for all x ∈ M \ i=1 Ci .
Remark 8.4.1 For g1 , g2 ∈ S, define g1 ≥ g2 if g1 ∈ Sg2 and g1 > g2 if there

is t > 0 such that g1 ∈ S≥t g2 . The relation “≥” in S is the reverse of Green’s
L-preorder of the semigroup theory (cf. [31]). A function f : X ⊂ M → R is
strictly decreasing on trajectories if and only if f (g1 (x)) < f (g2 (x)) whenever
g1 > g2 . Indeed, suppose that f is strictly decreasing on trajectories and let
g1 > g2 . Write g1 = ϕut11 and g2 = ϕut22 . Then there are t > 0 and u ∈ U such
that ϕut11 = ϕut ◦ ϕut22 . For any x ∈ X, we have

f (g1 (x)) = f ϕut ϕut22 (x) < f ϕut22 (x) = f (g2 (x)) .
On the other hand, suppose that f (g1 (x)) < f (g2 (x)) whenever g1 > g2 . For
t > 0 and u ∈ U, we have ϕut = ϕut ◦ Id, hence ϕut > Id. It follows that
f (ϕ (t, x, u)) < f (x) for all x ∈ X.
A Morse-Lyapunov function associated to an attractor-repeller pair (A, A∗ )

can be described as a real-valued function LA : M → [0, 1] such that L−1
A (0) =
−1
A, LA (1) = A , and LA is strictly decreasing on trajectories in C (A, A∗ ) =
∗
M \ (A ∪ A∗ ). The set C (A, A∗ ) is usually called the set of connecting

orbits of the attractor-repeller pair (A, A∗ ).
The following theorem assures the existence of a Morse-Lyapunov function
of an attractor-repeller pair.
Theorem 8.4.1 Let A ⊂ M be a Conley attractor of the control system.

There then exists a Morse-Lyapunov function LA for (A, A∗ ).
Proof. Suppose that A is a nontrivial Conley attractor. Since A and A∗

are disjoint closed sets and M is a perfectly normal space, there is a continuous
function f : M → [0, 1] such that f −1 (0) = A and f −1 (1) = A∗ . We define
h : M → [0, 1] by h (x) = sup {f (g (x)) : g ∈ S}. If g1 ≥ g2 , then g1 = kg2 for
some k ∈ S. We then have
h (g2 x) = sup {f (gg2 (x)) : g ∈ S}

≥ sup {f (gkg2 (x)) : g ∈ S}
= sup {f (gg1 (x)) : g ∈ S}
= h (g1 (x)) .
Hence h is nonincreasing on trajectories. We claim that h is continuous. It is

enough to show that h−1 ([0, ε)) and h−1 ((ε, 1]) are open sets in M for every
0 < ε < 1, since h−1 ((a, b)) = h−1 ([0, b)) ∩ h−1 ((a, 1]), for any basic open
set of the form (a, b) ⊂ [0, 1]. For 0 < ε < 1, let x ∈ h−1 ((ε, 1]). There is

then g ∈ S such that f (g (x)) > ε. Hence g −1 f −1 ((ε, 1]) ⊂ h−1 ((ε, 1])
is an open neighborhood of x. Thus h−1 ((ε, 1]) is an open set in M . Now,
let x ∈ h−1 ([0, ε)). Then h (x) < ε and O+ (x) ⊂ f −1 ([0, h (x))). By taking
δ > 0 with h (x) < δ < ε, it follows that f −1 ([0, δ)) is an open neighborhood of
−1
A and f −1 ([0, δ))∩A∗ = ∅. Then there is t > 0 such that O≥t +
f ([0, δ)) ⊂
−1
f −1 ([0, δ)). If y ∈ O≥t
+
f ([0, δ)) and g ∈ S then
−1 −1
g (y) ∈ gO≥t
+
f ([0, δ)) ⊂ O≥t
+
f ([0, δ)) ⊂ f −1 ([0, δ)) .
Hence
h (y) = sup {f (g (y)) : g ∈ S} ≤ δ < ε,
−1
that is, y ∈ h−1 ([0, ε)). Thus O≥t +
f ([0, δ)) ⊂ h−1 ([0, ε)). Now, since
h (x) < δ, we have O≤t +
(x) ⊂ f −1 ([0, δ)), and hence [0, t] × {x} × U ⊂
−1
−1
ϕ f ([0, δ)) . This means that there is an open neighborhood V of x

such that V ⊂ f −1 ([0, δ)) and [0, t] × V × U ⊂ ϕ−1 f −1 ([0, δ)) . Hence
O≤t
+
(V ) ⊂ f −1 ([0, δ)), and therefore
O+ (V ) = O≤t
+
(V ) ∪ O≥t
+
(V ) ⊂ f −1 ([0, δ)) .
This means that V ⊂ h−1 ([0, ε)), and then h−1 ([0, ε)) is an open set in M .
This proves the claim. Now, since f −1 (0) = A and A is positively invariant,
we have f (g (x)) = 0 for all g ∈ S, hence h (A) = 0. If h (y) = 0 then
f (y) = 0, hence y ∈ A. Thus h−1 (0) = A. If x ∈ A∗ = f −1 (1) then
h (x) ≥ f (x) = 1, hence h (x) = 1. On the other hand, if h (y) = 1 then there
are sequences (tn ) in R+ and (un ) in U such that f (ϕ (tn , x, un )) → 1. If
(tn ) is bounded, we may assume that tn → t with t ≥ 0. Assuming also that
un → u, the continuity of f implies f (ϕ (t, x, u)) = 1. Hence ϕ (t, x, u) ∈ A∗ ,
and since A∗ is invariant, it follows that x ∈ A∗ . Thus if x ∈ / A∗ then (tn ) must
be unbounded, and we may then assume that tn → +∞. By compactness,
we may also assume that ϕ (tn , x, un ) → y. This means that y ∈ ω + (x). As
x∈ / A∗ , ω + (x) ⊂ A, hence f (y) = 0, and thus f (ϕ (tn , x, un )) → f (y) = 0,
a contradiction. Therefore h (x) = 1 implies x ∈ A∗ , which proves that
h−1 (1) = A∗ . Now we define the function l : M × U → [0, 1] by
+∞
l (x, u) = e−t h (ϕ (t, x, u)) dt.
0
Since h and ϕ are continuous, it follows that l is continuous. We have l (x, u) =

0 if and only if h (x) = 0, hence l−1 (0) = A × U . We also have l (x, u) = 1 if
and only if h (x) = 1, and then l−1 (1) = A∗ × U. Because h is nonincreasing
on orbits, we have
+∞
l (ϕ (s, x, u) , u · s) = e−t h (ϕ (t, ϕ (s, x, u) , u · s)) dt
0
+∞
= e−t h (ϕ (s, ϕ (t, x, u) , u · t)) dt
0
+∞
≤ e−t h (ϕ (t, x, u)) dt
0
= l (x, u)
for all s > 0 and x ∈ M . If x ∈ C (A, A∗ ), there is ε > 0 such that h (x) ≥ ε.
Since ω + (x) ⊂ A, there is T > 0 such that O≥T+
(x) ⊂ h−1 ([0, ε)). For u ∈ U,
it follows that
t = sup {t > 0 : h (ϕ (t, x, u)) ≥ ε}
is finite and h (ϕ (t , x, u)) ≥ ε. Since h (ϕ (t + s, x, u)) − h (ϕ (t , x, u)) < 0,
the function t ∈ (0, +∞)
→ h (ϕ (t + s, x, u)) − h (ϕ (t, x, u)) is not identically
zero if s > 0. Hence
+∞
l (ϕ (s, x, u) , u · s) − l (x, u) = e−t (h (ϕ (t + s, x, u)) − h (ϕ (t, x, u))) dt
0
is strictly negative, that is, l (ϕ (s, x, u) , u · s) < l (x, u). Finally, a Morse-
Lyapunov function LA : M → [0, 1] for (A, A∗ ) is defined as
LA (x) = sup l (x, u) .

u∈U
Indeed, since l−1 (0) = A × U and l−1 (1) = A∗ × U , we have L−1 A (0) = A
−1 ∗
and LA (1) = A . We claim that LA is continuous. We choose ε such that
0 < ε < 1. If x ∈ L−1
A ([0, ε)), we can take ε such that LA (x) < ε < ε. Then

l (x, u) < ε for all u ∈ U . Since l is continuous, there is a neighborhood V
of x in M such that V × U ⊂ l−1 ([0, ε )). If y ∈ V then l (y, u) < ε for all
u ∈ U , hence LA (y) ≤ ε < ε. It follows that V ⊂ L−1 A ([0, ε)), and therefore
L−1A ([0, ε)) is open. If x ∈ L −1
A ((ε, 1]), there is some u such that l (x, u) > ε.
There is then a neighborhood V of x with V × {u} ⊂ l−1 ((ε, 1]). If y ∈ V ,
we have l (y, u) > ε, hence LA (y) > ε. It follows that V ⊂ L−1 A ((ε, 1]), and
therefore L−1 A ((ε, 1]) is open. Thus L A is continuous. It remains to show
that LA is strictly decreasing on trajectories in C (A, A∗ ). Let x ∈ C (A, A∗ ),
s > 0, and v ∈ U . Since LA (ϕ (s, x, v)) = l (ϕ (s, x, v) , u0 ) for some u0 ∈ U ,
and ϕ (t, ϕ (s, x, v) , u0 ) = ϕ (t + s, x, w), where w ∈ U is the s-concatenation
of v and u0 , we have
LA (ϕ (s, x, v)) = l (ϕ (s, x, v) , u0 )

+∞
= e−t h (ϕ (t, ϕ (s, x, v) , u0 )) dt
0
+∞
= e−t h (ϕ (t + s, x, w)) dt
0
+∞
= e−t h (ϕ (t, ϕ (s, x, w) , w · s)) dt
0
= l (ϕ (s, x, w) , w · s)
< l (x, w)
≤ LA (x) .
Therefore LA is a Morse-Lyapunov function for (A, A∗ ).
We now prove the existence of a Morse-Lyapunov function for a general

Morse decomposition, as follows:
Theorem 8.4.2 Every Morse decomposition of the control system admits a

Morse-Lyapunov function.
Proof. Let M = {C1 , ..., Cn } be a Morse decomposition and let ∅ =

A0 ⊂ A1 ⊂ · · · ⊂ An = M be the increasing sequence of Conley attractors
such that Ci = Ai ∩ A∗i−1 , i = 1, ..., n. Define LM = Σni=1 3−i LAi , where
LAi is a Morse-Lyapunov function for (Ai , A∗i ) given by Theorem 8.4.1. The
function LM is a Morse-Lyapunov function for M. In fact, it is easily seen
"n
that LM is continuous. If x ∈ M \ i=1 Ci , we take j = min {i : x ∈ Ai }.
Since x ∈/ Aj ∩ A∗j−1 , we have x ∈ / A∗j−1 , hence x ∈/ Aj−1 ∪ A∗j−1 . It fol-
lows that LAj−1 (ϕ (t, x, u)) < LAj−1 (x), for all t > 0 and u ∈ U, and hence
LM (ϕ (t, x, u)) < LM (x). Thus LM is strictly decreasing on trajectories out-
"n "n
side i=1 Ci . For a given critical value c ∈ LM ( i=1 Ci ), we take y ∈ L−1 M (c).
−i
Then LM (y) = LM (Cj ) for some Cj , that is, LM (y) = Σj−1 i=1 3 . This means
that LAj (y) = 0 and LAj−1 (y) = 1, and hence y ∈ Aj ∩A∗j−1 = Cj . Therefore
L−1
M (c) = Cj , and the proof is completed.
Finally, we define complete Lyapunov functions for control systems.
Definition 8.4.2 A complete Lyapunov function for the control system

is a continuous real-valued function L : M → R that is strictly decreasing on
trajectories outside the chain recurrence set C and such that the set L (C) of
critical values of L is nowhere dense in R and, for each critical value c, the
set L−1 (c) is a chain component.
According to Theorems 8.3.1 and 8.4.2, if the finest Morse decomposition

exists then its Morse-Lyapunov function is a complete Lyapunov function for
the control system. Otherwise, the existence of a complete Lyapunov function
of the control system is assured if every Conley attractor is invariant.
Theorem 8.4.3 Assume that the Conley attractors are invariant. There then
exists a complete Lyapunov function for the control system on the compact
manifold M .
Proof. It should be remembered that the hyperspace H (M ) is a metric

space endowed with the Hausdorff metric dH . For X ∈ H (M ) and > 0, the
-ball centered at X is given by
BH (X, ) = {Y ∈ H (M ) : Y ⊂ B (X, ) and X ⊂ B (Y, )} .
Suppose that A is a nontrivial Conley attractor. We can find > 0 such

that B (A, ) and B (A∗ , ) are disjoint sets. We claim that (A, A∗ ) is the
unique attractor-repeller pair such that A ⊂ B (A, ) and A∗ ⊂ B (A∗ , ). In-
deed, suppose that (B, B ∗ ) is an attractor-repeller pair with B ⊂ B (A, )
and B ∗ ⊂ B (A∗ , ). If x ∈ B then ω + (x) ∪ ω − (x) ⊂ B by invariance. As
B (A, ) and B (A∗ , ) are disjoint, it follows that ω + (x) ∪ ω − (x) ⊂ A, and
hence x ∈ A. Thus B ⊂ A. Analogously, we can show that A ⊂ B, and
therefore A = B. Since the complementary repellers are invariant, we can
show that A∗ = B ∗ in a similar way, proving the claim. Thus (A, A∗ ) is the
unique attractor-repeller pair in the basic open set BH (A, ) × BH (A∗ , ) of
the product space H (M ) × H (M ). Since H (M ) × H (M ) has a countable
base, it follows that there are at most countably many attractor-repeller pairs
for the control system. Let A1 , A2 , ..., An , ... be the sequence of all Conley
attractors of the system. Define L = Σ∞ n=1 3
−n
LAn where LA1 , LA2 , ... are
Morse-Lyapunov functions for the attractor-repeller pairs of the control sys-
tem. Then L is clearly a continuous function on M . If x ∈ / An ∪A∗n
/ C, then x ∈
for some Conley attractor An , by Theorem 7.3.2. It follows that L is strictly
decreasing on trajectories outside the chain recurrence set C. Since each LAn
is either 0 or 1 at a point of C, each critical value in L (C) lies in the “middle
third” Cantor set, hence L (C) is nowhere dense in R. Now, if C is a chain
Section 8.5 · Final comments and open questions 271
component and x, y ∈ C then x ∈ Ω+ (y) and y ∈ Ω+ (x). This means that

every Conley attractor containing x also contains y. Hence L (C) = c for some
critical value c. On the other hand, if z ∈ L−1 (c) then L (z) = L (x). If An
is a Conley attractor containing x then LAn (x) = 0, and hence LAn (z) = 0.
If Am is such that x ∈ A∗m then LAm (z) = LAm (x) = 1. Hence z lies in any
Conley attractor containing x and in any complementary repeller containing
x. It follows that z ∈ Ω+ (x) ∩ Ω− (x) = C. Therefore L−1 (c) = C and the
theorem is proved.
8.5 Final comments and open questions

This book extends the Conley theory to the framework of control affine sys-
tems, reproducing the Morse decompositions and showing their relationship
with chain controllability. The dynamical concepts of control systems do not
require compact state space, although the main theorems are assured under
compactness. A compact state space is required in order to guarantee the
main properties of the limit sets and the attractor-repeller pairs, which are
crucial for the Conley theorems. However, the definition of a dynamic Morse
decomposition can be adapted for control systems on noncompact manifolds,
extending Conley’s ideas to a more general framework.
In order to discuss the Morse decompositions on noncompact spaces, we
first notice that the Conley theorems can be clearly applied to control systems
with global attractors. Indeed, if a control system on a noncompact manifold
M has the global attractor A then the behavior of the system outside A is
well-known, since all trajectories come near A in forward time and go away
to infinity in backward time. Thus we may only study the dynamics of the
system inside the compact invariant set A, where all the Conley theorems
can be used. For instance, in Example 6.1.1, the sphere S2 ⊂ R3 \ {0} is the
global attractor of the control system, and since its points are critical, S2 is
chain controllable. In Example 6.2.1, the control system on the plane has the
global attractor D2 , the unit disk, and the dynamics in D2 have the finest

Morse decomposition M = {0} , S1 . Similarly, in Example 6.4.3, the lower
hemisphere of the bored sphere S2 \ {(0, 0, 1)} is the global attractor of the

system, and M = {(0, 0, −1)} , S1 × {0} is the finest Morse decomposition.
For another example, consider the control affine system

ẋ = u (t) x x2 − 1 x2 − 4
on the real line R, with control range U = [−10, −1]. This system has the
global attractor A = [−2, 2] and M = {{1} , {2} , {−1} , {0} , {−2}} is the
(ordered) finest Morse decomposition (see Fig. 8.1).
Figure 8.1: Control system on the real line with global attractor and finest
Morse decomposition.
Figs. 8.2 and 8.3 illustrate other typical systems whose global dynamical
behavior is determined by the attractive and the repulsive properties of the
critical points, but now with absence of the global attractor.
Figure 8.2: Image from Wikimedia Commons. The magnetic field with one
attractor point (negative pole) and one repeller point (positive pole).
Sometimes, infinitely many sets influence the dynamics of the system. For
instance, consider the control affine system ẋ = u (t) sin (πx) on the real line
R, with control range U = [1, 2]. The set Z of integer numbers corresponds
to the set of all critical points, which attract or repel all trajectories of the
system. In another case, consider the control affine system

x u1 (t) cos2 x
=
y u2 (t) sin x
2
on R2 , with control range U = [1, 2] . The dynamics of this system have
Section 8.5 · Final comments and open questions 273
Figure 8.3: Dynamics with two attractor points and one repeller point.
(2k + 1) π
vertical lines x = , with k ∈ Z. Between any two consecutive lines
2
(2k + 1) π (2k + 3) π
x= and x = , the trajectories with respect to u ∈ U are
2 2 !
u2
given by (x, y) : y + c = sec x where c is some constant. The dynamics
u1
kπ
of the system are illustrated in Fig. 8.4. The vertical lines x = attract or
2
repel all trajectories of the system, but the limit sets are all empty.
Figure 8.4: Dynamics with infinitely many noncompact attractors and re-
pellers.
In general, when the global dynamical behavior depends on the attractive

or repulsive effects from some invariant closed sets, the system has a Morse
decomposition. We then define a generalized dynamic Morse decompo-
sition as a sequence M = {Ci }i∈N of nonempty, isolated invariant, and closed

sets such that:
1. There is a δ > 0 such that B (Ci , δ) ∩ B (Cj , δ) = ∅ whenever i = j.

"
2. M = i∈N A (Ci ) ∪ R (Ci ).
"
3. If there are finite sequences Cj0 , Cj1 , ..., Cjl and x1 , ..., xl ∈ M \ Ci
i∈N
with xk ∈ R Cjk−1 and xk ∈ A (Cjk ), for k = 1, ..., l; then Cj0 = Cjl .
Condition (1) means that the (generalized) Morse sets are uniformly pair-
wise disjoint; condition (2) requires that all orbits of the system are attracted
or repelled by the Morse sets; condition (3) means that cycles are not al-
lowed. It is clear that this definition in fact generalizes the notion of a Morse
decomposition from Conley’s paradigm.
It should be mentioned that the study of Morse decompositions in non-
compact spaces is not a novelty (see the notes). Nevertheless, this present
proposal of a generalized dynamic Morse decomposition has not been devel-
oped anywhere else. The obvious questions are the following:
• Is the chain recurrence set contained in any generalized dynamic Morse

decomposition?
• Does the finest generalized dynamic Morse decomposition coincide with

the collection of the chain components?
• Is the finest generalized dynamic Morse decomposition the intersection

of all generalized dynamic Morse decompositions?
• Does a generalized dynamic Morse decomposition admit an attractor-

repeller configuration?
• Is there any Morse-Lyapunov function for a generalized dynamic Morse

decomposition?
• Is there any complete Lyapunov function for a control system on a non-

compact manifold?
Other interesting questions may come up during the studies of the dynamic
Morse theory for control systems on noncompact manifolds. The definition of
a generalized Morse decomposition might be changed or enhanced for better
discussion.
8.6 Problems
1. Recall the control system on the unit circle as defined in Example 6.3.2.
(a) Find all dynamic Morse decompositions of the control system and
describe their attractor-repeller configurations.
(b) Discuss the possible attractor-repeller configurations for the finest
Morse decomposition.
2. Let the unit circle S1 be identified with the interval [0, 2π). Study the
Morse decompositions of the following control affine systems on the 2-
torus T2 = S1 × S1 .

ẋ u (t) sin x
(a) = , with control range U = [1, 2].
ẏ 1

ẋ u (t) cos x
(b) = , with control range U = [1, 2].
ẏ u (t)

ẋ u1 (t) sin x 2
(c) = , with control range U = [1, 2] .
ẏ u2 (t)

ẋ u1 (t) cos x 2
(d) = , with control range U = [1, 2] .
ẏ u2 (t) sin y

ẋ u1 (t) sin x + u2 (t) cos x sin x
(e) = , with control range U =
ẏ 1

(u1 , u2 ) ∈ R2 : u1 + u2 = 1, u2 ∈ [1/2, 1] .
3. Consider the control affine system ẋ = u (t) X (x) on the unit disk M =

x ∈ R2 : x ≤ 1 , with control range U = [1, 10] and vector field
⎛ ⎞
2
−x + x x + x 2
sin √ π
x1 ⎜ 2 1 1 2 2 2 ⎟
X = ⎜ x1 +x2 ⎟
x2 ⎝ 2 ⎠
x1 + x2 x1 + x22 sin √ 2π 2
x1 +x2
X (0) = 0.
Discuss the Morse decompositions of this system. Does the finest one
exist?
4. Let M = {C1 , ..., Ck } be a dynamic Morse decomposition and X ⊂ M a

closed invariant set. Define MX = {Ci ∩ X : i ∈ {1, ..., k}} where only
those i with Ci ∩ X = ∅ are allowed. Prove that MX is a dynamic
Morse decomposition of the control system restricted to X.
5. In Problem 4, if we consider attractor-repeller Morse decomposition

instead of dynamic Morse decomposition, does the result hold?
6. Let M = {C1 , ..., Ck } be a dynamic Morse decomposition of the control

system in M and Mi = C1i , ..., Cki i a dynamic Morse decomposition
of the control system restricted to Ci , for each i = 1, ..., k. Show that
"k
M = i=1 Mi is a dynamic Morse decomposition of the control system
in M .
7. Consider a control system on a compact manifold M and let M =

{C1 , ..., Ck } be an attractor-repeller Morse decomposition of the system.
Take the quotient space Q obtained from M × U by collapsing the lifts
L (Ci ) to distinct points q1 , ..., qk and let ρ : M × U → Q be the quotient
map. Define φ : R × Q → Q as in Problem 7.4(16). Show that φ is a
gradient-like dynamical system with k rest points, and such that the
ordering of the Morse sets corresponds to that of the points induced by
the Morse-Lyapunov function.
8. A dynamical system φ : R × M → M is called strongly gradient-like

if the chain recurrence set is totally disconnected. Consider a control
affine system on a compact manifold M . Show that its control flow is
uniquely represented as the extension of a chain recurrent dynamical
system by a strongly gradient-like dynamical system. (Hint: Note that
a totally disconnected chain recurrence set must be the rest point set of
the system.)
9. Let a1 < a2 < · · · < an be an increasing sequence of real numbers.

Discuss the generalized dynamic Morse decompositions of the control
affine system ẋ = u (t) (x − a1 ) (x − a2 ) ... (x − an ) on the real line R,
with control range U = [b, c].
10. Find a generalized dynamic Morse decomposition for the control system
in Example 5.5.1.
11. Study the generalized dynamic Morse decompositions of the control sys-
tem ẋ = u (t) X (x) on R2 , with control range U = [1, 2] and vector field
⎛ $ %⎞

x1 −x2 + x1 x21 + x22 cos π2 x21 + x22
X =⎝ $ % ⎠.
x2 x1 + x2 x21 + x22 cos π2 x21 + x22
What can you say about the finest Morse decomposition?
12. Recall the control systems in Examples 2.2.1 and 7.1.3. Discuss the
generalized dynamic Morse decompositions of these systems.
13. May a controllable system have a generalized dynamic Morse decompo-

sition?
14. May a chain controllable system have a generalized dynamic Morse de-
composition?
15. Consider a control system on a noncompact manifold and let M =

{C1 , ..., Cn } be a finite generalized dynamic Morse decomposition of the
system. Is it necessary that C1 is a repeller and Cn is an attractor?

The Morse decomposition part of the Conley theory can be extended to
the control systems on noncompact manifolds by means of the compactifi-
cation methodologies. The paper [95] uses the Stone-Čech compactification
by reproducing the Conley main theorem for semigroup actions on noncom-
pact spaces. The paper [100] applies the Alexandroff compactification by
discussing the Conley theorems for semigroup actions and control systems on
noncompact state spaces. These papers do not treat the generalized dynamic
Morse decomposition as defined in this chapter, but they may help to answer
some of the open questions.
The Selgrade theorem on linear systems on vector bundles is one of the
most important applications of the Conley theorems. A control system Σ on
a compact manifold M lifts to a control system T Σ on the tangent bundle
T M . This linearized control system induces a control system on the projective
bundle PT M by the fibration P : T M \Z → PT M , where Z is the zero section
of T M . If Σ is chain controllable on M then the Selgrade theorem assures the
existence of the finest Morse decomposition {C1 , ..., Cn } for the induced system
on PTM , and furthermore each Morse set Ci defines an invariant subbundle

P −1 (Ci ) of T M such that T M = P −1 (C1 ) ⊕ · · · ⊕ P −1 (Cn ) as a Whitney sum
([98]).
Appendix A
Dynamical systems
The theory of dynamical systems derived from the qualitative studies of

differential equations. A dynamical system was presented as a flow associated
to a differential equation of the form ẋ = X (x), where X is a vector field on
the Euclidean space Rn . This concept was extended to general topological
spaces, becoming a special type of homeomorphism group. This appendix is
dedicated to the dynamic Morse theory of dynamical systems developed by
Conley ([36]). The exposition is based on Colonius and Kliemann [34]. In
the preliminaries, we study the asymptotic behavior of a dynamical system
by means of the limit sets. In sequence, we present the concepts of Morse
decompositions and attractor-repeller pairs for dynamical systems on compact
metric spaces. The main result shows that all Morse decompositions are
obtained by increasing sequences of attractors. We also define the notion of
chain transitivity of dynamical systems and show its relation to the Morse
decompositions. The main theorem shows that the existence of the finest
Morse decomposition is equivalent to the existence of finitely many connected
components of the chain recurrence set.
A.1 Preliminaries
In this section we introduce the fundamental concepts of dynamical systems.
They are all extracted from the works [34, 36].
A dynamical system is determined by a continuous map that defines a
homeomorphism group parameterized by real numbers.
279
280 Chapter A · Dynamical systems
Definition A.1.1 Let M be a metric space with metric d. A dynamical

system or flow on M is a continuous map φ : R × M → M that satisfies
the conditions φ (0, x) = x and φ (t + s, x) = φ (t, φ (s, x)) for all x ∈ M and
t, s ∈ R.
For x ∈ M and t ∈ R fixed, we have the partial maps
φx : R → M φt : M → M
s
→ φx (s) = φ (s, x) y
→ φt (y) = φ (t, y)
which are continuous. It is easily seen that the map φ−t is the inverse map
of φt , hence each map φt is a homeomorphism of M . The set φx (R) is called
trajectory or orbit of x.
Exercise 1 Prove that two trajectories of a dynamical system are either co-
incident or disjoint.
Exercise 2 Let M be a differential manifold (or Rn for simplicity) and X

a differentiable vector field on M . According to the Fundamental Theorem
of Differential Equations, there is a unique solution φ (t, x0 ) for the initial
value problem ẋ = X (x), x (0) = x0 . Show that φ satisfies the group prop-
erty φ (t + s, x0 ) = φ (t, φ (s, x0 )), whenever φ (t + s, x0 ) and φ (t, φ (s, x0 ))
are defined.
Example A.1.1 Let M = Rn and consider the map
φ : R × Rn → Rn
(t, x)
→ φ (t, x) = (t + x1 , ..., t + xn )
where x = (x1 , ..., xn ). This map is clearly continuous and also satisfies
φ (0, x) = x,
φ (t + s, x) = (t + s + x1 , ..., t + s + xn ) = φ (t, φ (s, x)) ,
for all x ∈ Rn and t, s ∈ R. Hence φ is a dynamical system on Rn . The

trajectory of x ∈ Rn is a line containing x. For t ∈ R, the map φt is a
translation in Rn . Note that φ satisfies the system of differential equations
ẋi = 1, i = 1, ..., n, or, in other words, it satisfies the equation ẋ = X (x)
where X is the constant vector field X ≡ (1, ..., 1).
Section A.1 · Preliminaries 281
Example A.1.2 Let M = Rn and consider the map
φ : R × Rn → Rn .
(t, x)
→ t
φ (t, x) = e x
This map is clearly continuous and also satisfies
φ (0, x) = x,
φ (t + s, x) = et+s x = et es x = φ (t, φ (s, x)) ,
for all x ∈ Rn and t, s ∈ R. Hence φ is a dynamical system on Rn . The origin

0 of Rn is an equilibrium point. The trajectory of x = 0 is a line converging
to 0 in backward time. For t ∈ R, the map φt is a translation in Rn . Note
that φ satisfies the equation ẋ = x.
Example A.1.3 Let A be an n × n real matrix and consider the map
φ : R × Rn → Rn
(t, x)
→ φ (t, x) = etA (x)
where etA is the exponential matrix. It can be easily seen that φ is a dynamical
system on Rn associated to the linear differential equation ẋ = A (x).
Example A.1.4 Let M = GL (n, R) be the group of all invertible n × n real

2
matrices endowed with the relative topology of Rn . Consider the map
φ : R × GL (n, R) → GL (n, R)
(t, x)
→ φ (t, x) = etA x
where A is a n × n real matrix and etA x is the product of the matrices etA
and x. It is not difficult to show that this map defines a dynamical system on
GL (n, R). Note that φ satisfies the equation ẋ = Ax on GL (n, R).
Next, we define limit sets.
Definition A.1.2 Let X ⊂ M be a nonempty set. The sets

!
x ∈ M : there are sequences (xn ) in X and tn → +∞
ω + (X) =
such that φ (tn , xn ) → x
!
− x ∈ M : there are sequences (xn ) in X and tn → −∞
ω (X) =
such that φ (tn , xn ) → x
are respectively called positive limit set and negative limit set of X.
For t ∈ R and x ∈ M , we define φ∗ (t, x) = φ (−t, x). Then φ∗ also defines

a dynamical system on M . Note that the negative limit sets of φ coincide
with the positive limit sets of φ∗ . Thus, in order to obtain the properties of
the limit sets of a dynamical system, it is enough to verify the properties of
its positive limit sets. The map φ∗ is called the time-reversed dynamical
system of φ.
Definition A.1.3 A set X ⊂ M is said to be invariant if φ (R × X) ⊂ X.
Note that an invariant set contains all the trajectories of its points.
Exercise 3 Let X ⊂ M be invariant and closed. Show that ω + (X) =

ω − (X) = X.
Proposition A.1.1 The limit sets of a dynamical system are closed and in-
variant. In particular, if M is compact then the limit sets are nonempty,
compact, and invariant.
Proof. Let X ⊂ M be a nonempty set. The result is clear if ω + (X) is

empty. Suppose that ω + (X) = ∅ and take y ∈ cl (ω + (X)). Then there is a
sequence (xn ) in ω + (X) that converges to y. For each n, there are sequences
(xnk ) in X and tnk → +∞ such that φ (tnk , xnk ) → xn . Hence we can find kn
1
such that tnkn > n and d φ tnkn , xnkn , xn < . It follows that tnkn → +∞,
n
xnkn ∈ X, and

d φ tnkn , xnkn , y ≤ d φ tnkn , xnkn , xn + d (xn , y)
1
< + d (xn , y) .
n

This means that φ tnkn , xnkn → y as n → +∞, and therefore y ∈ ω + (X).
This proves that ω + (X) is closed in M . By showing that ω + (X) is invariant,
take any z ∈ ω + (X) and t ∈ R. Then there are sequences φ (tk , xk ), tk ∈ R,
and xk ∈ X, with tk → +∞ and φ (tk , xk ) → z. By continuity we have
φt (φ (tk , xk )) → φt (z), that is, φ (t + tk , xk ) → φt (z), with t + tk → +∞.
Hence φt (z) ∈ ω + (X), and therefore ω + (X) is invariant. Now, suppose that
M is compact. Take x ∈ X and consider the sequence (φ (n, x))n∈N . By
compactness, there is a subsequence (φ (nk , x)) that converges to some point
y ∈ M . This means that y ∈ ω + (X), and thus ω + (X) is nonempty. By the
first part of the proof, ω + (X) is compact and invariant.
Section A.1 · Preliminaries 283
Exercise 4 Prove that ω + (x) = ω + (φ (t, x)) and ω − (x) = ω − (φ (t, x)) for
all x ∈ M and t ∈ R.
Exercise 5 Suppose that φx (R) is a periodic trajectory with period p, that

is, φx (t + p) = φx (t) for all t ∈ R. Show that ω + (x) = ω − (x) = φx (R).
Figure A.1: Two limit sets of a dynamical system on the plane.
In the following we present some examples of limit sets.
Example A.1.5 Consider the dynamical system on M = R2 generated by

the vector field X = (X1 , X2 ) with

X1 (x) = x2 + x1 1 − x21 − x22 ,

X2 (x) = −x1 + x2 1 − x21 − x22 .
For every x ∈ M , the inner product of x and X (x) is

2 4
x, X (x) = x − x .
If x = 1, the trajectory of x is the unit circle C1 . Hence ω + (x) = ω − (x) =

C1 . In the case 0 < x < 1, the trajectory of x is a spiral around the origin
0 of R2 , coming near to the circle C1 . Hence we have ω + (x) = C1 and
ω − (x) = {0}. As the origin is a fixed point of the system, we have ω + (0) =
ω − (0) = {0}. Now, if x > 1, the trajectory of x spirals coming near to
C1 . Hence ω + (x) = C1 and ω − (x) = ∅. The trajectories of the system are
illustrated in Fig. A.1.
Figure A.2: Infinitely many limit sets of a dynamical system on the plane.
Example A.1.6 Consider the dynamical system on M = R2 determined by

the vector field X = (X1 , X2 ) with

2 π
X1 (x) = −x2 + x1 x1 + x22 sin 2 ,
x1 + x22

2 π
X2 (x) = x1 + x2 x1 + x2 sin 2
2
.
x1 + x22
For every x ∈ R2 , the inner product of x and X (x) is

4 π
x, X (x) = x sin .
x
The origin 0 of R2 is the unique fixed point of the system. Then ω + (0) =
1
ω − (0) = {0}. If x = , with n ∈ N, then the trajectory of x is the
n
1
circle centered at 0 with radius . By denoting this circle by Cn we have
n
ω + (x) = ω − (x) = Cn for every x ∈ Cn . If x > 1 then the trajectory of x
spirals to the circle C1 in positive time, and goes away to infinity in negative
1 1
time. Hence ω − (x) = ∅ and ω + (x) = C1 . In the case < x < ,
n+1 n
the trajectory of x is a spiral inside the region limited by the circles Cn and
Cn+1 . If n is even, the spiral comes near the circle Cn+1 in positive time, and
comes near the circle Cn in negative time. This means that ω + (x) = Cn and
ω − (x) = Cn+1 . If n is odd, the spiral comes near Cn in positive time, and
comes near Cn+1 in negative time. Thus ω − (x) = Cn+1 and ω + (x) = Cn .
Thus, for points in the unit disk, the positive limit sets are the circles Cj ,
Section A.2 · Morse decompositions 285
with j odd, and the negative limit sets are the circles Ci , with i even. The
trajectories of the system are illustrated in Fig. A.2.
Exercise 6 Consider the system of differential equations

x1 = x2
x2 = −ax2
on the plane with a constant a > 0. Find the limit sets of this system.
A.2 Morse decompositions

This section is dedicated to the Morse decompositions of dynamical systems
on compact metric spaces. A Morse decomposition contains all the limit sets
and describes the global limit behavior of a system.
From now on, there is a fixed dynamical system φ on a compact metric
space M .
Definition A.2.1 A set X ⊂ M is said to be isolated invariant if it is

invariant and there is an isolating neighborhood V of X such that φx (R) ⊂ V
implies x ∈ X.
Definition A.2.2 A Morse decomposition of the dynamical system on M

is a finite collection M = {Ci : i = 1, ..., n} of nonempty, compact, isolated
invariant, and pairwise disjoint sets such that:
"
n
1. For every x ∈ M , ω + (x) , ω − (x) ⊂ Ci ;
i=1
"
n
2. If there are sets Cj0 , Cj1 , ..., Cjl and points x1 , ..., xl ∈ M \ Ci with
i=1
ω − (xk ) ⊂ Cjk−1 and ω + (xk ) ⊂ Cjk , for k = 1, ..., l, then Cj0 = Cjl .
Each element of M is called a Morse set.
The condition 2 of Definition A.2.2 is usually called no-cycle condition. In

the following we present other properties of a Morse decomposition.
Proposition A.2.1 Let M = {Ci : i = 1, ..., n} be a Morse decomposition.

If ω + (x) ∪ ω − (x) ⊂ Cj then x ∈ Cj .
Proof. Since M is compact, the limit sets ω + (x) and ω − (x) are nonempty.
Let x ∈ M such that ω + (x) ∪ ω − (x) ⊂ Cj . According to item 2 of Definition
A.2.2, it follows that x ∈ Ck for some k ∈ {1, ..., n}. As Ck is compact and
invariant, we have ω + (x) , ω − (x) ⊂ Ck . Since the Morse sets are pairwise
disjoint, it follows that Ck = Cj , and therefore x ∈ Cj .
The no-cycle condition of a Morse decomposition enables to define an order

among the Morse sets, as follows.
Definition A.2.3 Let M = {Ci : i = 1, ..., n} be a Morse decomposition of

the dynamical system on M . For Cα , Cβ ∈ M, we define Cα Cβ if there
are sequences Cj0 = Cα , Cj1 , ..., Cjm = Cβ and x1 , ..., xm ∈ M with ω − (xk ) ⊂
Cjk−1 and ω + (xk ) ⊂ Cjk , for k = 1, ..., m.
Proposition A.2.2 The relation “” is an order among the Morse sets of
a Morse decomposition M = {Ci : i = 1, ..., n}.
Proof. For any x ∈ Ci , we have ω − (x) , ω + (x) ⊂ Ci , hence Ci

Ci . Suppose that Cα Cβ and Cβ Cα . There are then sequences
Cj0 = Cα , Cj1 , ..., Cjm = Cα , and x1 , ..., xm ∈ M with ω − (xk ) ⊂ Cjk−1 ,
and ω + (xk ) ⊂ Cjk , for k = 1, ..., m, where Cβ = Cjk for some k. As Cj0 =
"
n
Cα = Cjm , there must be some xkl ∈ Ci , because of the no-cycle condition.
i=1
Hence xkl ∈ Cl for some l ∈ {1, ..., n}, and thus ω + (xkl ) , ω − (xkl ) ⊂ Cl . Nev-
ertheless, ω − (xkl ) ⊂ Cjkl −1 and ω + (xkl ) ⊂ Cjkl , hence Cjkl −1 = Cl = Cjkl .
Following by induction we obtain Cα = Cjk for all k = 1, ..., m, and thus
Cα = Cβ . This proves that the relation “” is antisymmetric. The transitive
property is clear.
The properties of the last two propositions above characterize the Morse
decompositions.
Theorem A.2.1 A finite collection {Ci : i = 1, ..., n} of nonempty, compact,

isolated invariant, and pairwise disjoint sets in M is a Morse decomposition
if and only if the following conditions are satisfied:
"
n
1. ω + (x) , ω − (x) ⊂ Ci , for every x ∈ M .
i=1

3. the relation “” is an order among the sets Ci .
Proof. Suppose that the collection {Ci : i = 1, ..., n} satisfies the condi-
tions (1) − (3). By checking that it is a Morse decomposition, it is enough
to show that it satisfies the no-cycle condition. Suppose to the contrary
"
n
that there are sequences Cj0 , Cj1 , ..., Cjl and x1 , ..., xl ∈ M \ Ci with
i=1
ω − (xk ) ⊂ Cjk−1 and ω + (xk ) ⊂ Cjk , for k = 1, ..., l, and Cj0 = Cjl . For
any k = 1, ..., l − 1, we have Cj0 Cjk and Cjk Cjl = Cj0 . Since “” is
an order, it follows that Cjk = Cj0 for all k = 1, ..., l. Hence ω − (xk ) ⊂ Cj0
and ω + (xk ) ⊂ Cj0 for all k = 1, ..., l, and thus xk ∈ Cj0 by the condition (2).
"n
This contradicts xk ∈ M \ Ci , and therefore the collection {Ci : i = 1, ..., n}
i=1
satisfies the no-cycle condition.
By enumerating the Morse sets in such a way that Ci Cj means i ≤ j, a

Morse decomposition describes the global dynamical behavior of the system
by means of its movements from the Morse sets with lower indices toward
those with higher ones.
A dynamical system may admit a lot of Morse decompositions. We then
need a manner of comparing or relating the Morse decompositions.
Definition A.2.4 A Morse decomposition M = {Ci : i = 1, ..., n} is finer

than the Morse decomposition M = Cj : j = 1, ..., m if for each j there is
i with Ci ⊂ Cj . A Morse decomposition M is the finest Morse decomposi-
tion of the dynamical system if it is finer than any other Morse decomposition
of the system.
The finest Morse decomposition may not exist (see Example A.2.2). If
it exists then it must be the intersection of all the Morse decompositions
of the system. For two Morse decompositions M = {Ci : i = 1, ..., n} and

M = Cj : j = 1, ..., m , we define the intersection

M ∩ M = Ci ∩ Cj : i, j
where only the indexes i, j with Ci ∩ Cj = ∅ are allowed.

Proposition A.2.3 If M = {Ci : i = 1, ..., n} and M = Cj : j = 1, ..., m
are two Morse decompositions of the dynamical system then the intersection
M ∩ M is also a Morse decomposition.
Proof. It is clear that the sets Ci ∩ Cj are nonempty, compact, and
pairwise disjoint. If x ∈ Ci ∩ Cj then φx (R) ⊂ Ci ∩ Cj , because both
Ci e Cj are invariant. Hence Ci ∩ Cj is invariant. Moreover, the intersec-
tion of isolating neighborhoods of Ci and Cj is an isolating neighborhood of
Ci ∩ Cj . Thus Ci ∩ Cj is isolated invariant for all i, j. Now, for a given
x ∈ M and y ∈ ω + (x), we have y ∈ Ci and y ∈ Cj , for some i, j, hence
" "
ω + (x) ⊂ Ci ∩ Cj . Analogously, ω − (x) ⊂ Ci ∩ Cj . Thus M ∩ M sat-
i,j i,j
isfies the condition 1 of Definition A.2.2. Finally, suppose that there are
"
sequences Ck0 ∩ Ck 0 , Ck1 ∩ Ck 1 , ..., Ckm ∩ Ck m and x1 , ..., xm ∈ M \ Ci ∩ Cj
i,j
with ω − (xl ) ⊂ Ckl−1 ∩ Ck l−1 and ω + (xl ) ⊂ Ckl ∩ Ck l , for l = 1, ..., m. As
Ckl ∩ Ck l ⊂ Ckl , for all l = 1, ..., m, we have Ck0 = Ckm , hence Ck0 and Ckm
are disjoint. This means that Ck0 ∩ Ck 0 = Ckm ∩ Ck m , and thus M ∩ M also
satisfies the no-cycle condition.
In the following we present some examples of Morse decompositions.
Example A.2.1 Consider the dynamical system defined in Example A.1.5.

Let S ⊂ R3 be the sphere of radius 1/2 centered at the point (0, 0, 1/2). Denote
p = (0, 0, 1) and p∗ = (0, 0, 0). Take the stereographic projection ρ : S \ {p} →
R2 and define the map F : R × S \ {p} → S \ {p} by
F (t, x) = ρ−1 ◦ φ ◦ (Id × ρ) (t, x) .
For x ∈ S \ {p} and t, s ∈ R, we have
F (0, x) = ρ−1 ◦ φ (0, ρ (x)) = ρ−1 (ρ (x)) = x
and

F (t, F (s, x)) = ρ−1 ◦ φ ◦ (Id × ρ) t, ρ−1 ◦ φ (s, ρ (x))
= ρ−1 ◦ φ (t, φ (s, ρ (x)))
= ρ−1 ◦ φ (t + s, ρ (x))
= ρ−1 ◦ φ ◦ (Id × ρ) (t + s, x)
= F (t + s, x) .
Now we define the map Φ : R × S → S by

F (t, x) , if x = p
Φ (t, x) = .
p, if x = p
Figure A.3: Dynamical system on the sphere with three limit sets.
Suppose that (tn , xn ) → (t, p) in S. This means that ρ (xn ) → +∞, hence
φ (tn , ρ (xn )) → +∞, and therefore ρ−1 (φ (tn , ρ (xn ))) → p. This means
that F (tn , xn ) → p, and therefore Φ is continuous. Since p is a fixed point, it
follows that Φ define a dynamical system on S. We have
ω + (p) = ω − (p) = {p} and ω + (p∗ ) = ω − (p∗ ) = {p∗ } .

Let S1 ⊂ R2 be the unit cycle and define E = ρ−1 S1 ⊂ S. If x ∈ E then
ω + (x) = ω − (x) = E.

Let D ⊂ R2 be the unit disk. If x ∈ ρ−1 R2 \ D then
ω + (x) = E and ω − (x) = {p} .
If x ∈ ρ−1 (int (D)) we have
ω + (x) = E and ω − (x) = {p∗ } .
Fig. A.3 illustrates the trajectories of the dynamical system. We claim that
the collection M = {E, {p} , {p∗ }} is the finest Morse decomposition of the
system. Indeed, the sets E, {p} , {p∗ } are compact, isolated invariant, and
pairwise disjoint. Moreover, ω + (x) , ω − (x) ⊂ E ∪ {p} ∪ {p∗ } for every x ∈ S.
The possible relations among the sets of M are the following
E E, {p} {p} , {p∗ } {p∗ } , {p} E, {p∗ } E.

Thus the relation “” is an order among the elements of M. Finally,

ω + (x) , ω − (x) ⊂ E implies x ∈ E, because ω − (x) ∈
/ E if x ∈
/ E. More-
over ω (x) , ω (x) ⊂ {p} means x = p, and ω (x) , ω (x) ⊂ {p∗ } im-
+ − + −
plies x = p∗ . Therefore M is a Morse decomposition. Now, denoting Hp =

ρ−1 R2 \ int (D) ∪ {p} and Hp∗ = ρ−1 (D), the other possible Morse decom-
positions of the system are the following
M1 = {E, {p, p∗ }} , M2 = {Hp , {p∗ }} ,

M3 = {Hp∗ , {p}} , M4 = {S} .
See Fig. A.4. Then M is the intersection of all the Morse decompositions of
the system, and therefore it is the finest one.
Figure A.4: Possible nontrivial Morse decompositions of a dynamical system

on the sphere.
Example A.2.2 Let M be the unit disk in R2 and consider the dynamical
system defined in Example A.1.6 restricted to M . For each natural number
n, define the sets
!
1
Cn = x ∈ M : x = ,
n
!
1 1
Kn = x∈M : ≤ x ≤ ,
n+1 n
!
1
Dn = x ∈ M : x ≤ .
n
Fix an even number n and define the sets
C1 = C1 ,
Ci = K2i−2 , for i = 2, ..., n/2,
C n2 +1 = Dn .
Section A.3 · Attractors and repellers 291
By denoting kn = n2 + 1, the collection Mkn = {C1 , C2 , ..., Ckn } is a Morse

decomposition of the system on M . In fact, the elements of Mkn are compact,
isolated invariant, and pairwise disjoint sets. Moreover, ω + (x) , ω − (x) ⊂
"n
k
Ci for every x ∈ M . Suppose that there are Ci0 , Ci1 , ..., Cil and x1 , ..., xl ∈
i=1
"n
k
M\ Ci with ω − (xk ) ⊂ Cik−1 and ω + (xk ) ⊂ Cik for k = 1, ..., l. Note that
i=1
ω − (xk ) ⊂ Cik−1 implies ω + (xk ) ⊂ Cik−1 +1 , hence Cik = Cik−1 +1 . This means
that Ci0 = Cil , and therefore Mkn is a Morse decomposition (see Fig. A.5).
This dynamical system does not have finest Morse decomposition. By fixing
any natural number n, the collection Mn = {C1 , ..., Cn }, where Ci = Ci for
i = 1, ..., n − 1 and Cn = Dn , is a Morse decomposition. Hence, for any
natural number n, we can find a Morse decomposition with n elements, and
therefore the system does not admit the finest Morse decomposition.
Figure A.5: Examples of Morse decompositions of a dynamical system on the

unit disk.
A.3 Attractors and repellers

This section contains a study on the concepts of attractors and repellers as
defined by Conley [36]. We define the attractor-repeller pairs and relate them
to the Morse decompositions.
Definition A.3.1 A set A ⊂ M is called an attractor if it has an attractor

neighborhood U such that ω + (U ) = A. A set R ⊂ M is called a repeller if
it admits a repeller neighborhood V such that ω − (V ) = R.
It is immediate that attractors and repellers are compact invariant sets.

A repeller of the dynamical system φ is an attractor of the time-reversed

dynamical system φ∗ .
The whole set M is an attractor as well as a repeller. We assume that
the empty set ∅ is also an attractor and a repeller. The empty set ∅ and the
whole set M are the trivial attractors and repellers of the system.
Proposition A.3.1 If A ⊂ M is an attractor and X ⊂ M is a compact

invariant set then the intersection A ∩ X is an attractor for the dynamical
system restricted to X.
Proof. Let U ⊂ M be an attractor neighborhood of A. If A ∩ X = ∅ then

the result is obvious. Suppose A ∩ X = ∅. Note that U ∩ X is a neighborhood
of A∩X in X. As ω + (U ) = A, we have ω + (U )∩X = A∩X. If x ∈ ω + (U )∩X
then x ∈ intX (U ∩ X) and there is a sequence (φ (tn , vn )) that converges to
x, with tn → +∞ and vn ∈ U . We may then assume that φ (tn , vn ) ∈ U ∩ X
for all n. By invariance, we have vn ∈ X, and hence φ (tn , vn ) → x, with
tn → +∞ and vn ∈ U ∩ X. This means that x ∈ ω + (U ∩ X), and thus
ω + (U ) ∩ X ⊂ ω + (U ∩ X). It follows that A ∩ X ⊂ ω + (U ∩ X). On
the other hand, since ω + (X) ⊂ X, we have ω + (U ∩ X) ⊂ X. More-
over, ω + (U ∩ X) ⊂ ω + (U ) = A, hence ω + (U ∩ X) ⊂ A ∩ X. Therefore
ω + (U ∩ X) = A ∩ X is an attractor with attractor neighborhood U ∩ X.
Exercise 7 Let A ⊂ M be an attractor with attractor neighborhood U . Show

that any neighborhood V of A contained in U is also an attractor neighborhood
of A.
In order to define the notion of an attractor-repeller pair, we need the

following lemma.
Lemma A.3.1 For every attractor neighborhood V of an attractor A there

is t∗ > 0 such that cl (φ ([t∗ , ∞) × V )) ⊂ int (V ).
Proof. Suppose by contradiction that for every t > 0 there is a point

x ∈ cl (φ ([t, ∞) × V )) such that x ∈ / int (V ). For each natural number n, we
take xn ∈ cl (φ ([n, ∞) × V )) with xn ∈ / int (V ). Since M \ int (V ) is compact,
we may assume that xn → x for some x ∈ M \ int (V ). For each n fixed,
there is a sequence φ (tnk , vkn ) → xn as k → +∞, with tnk ∈ [n, ∞) and vkn ∈ V .
1
Hence there is k0 such that k > k0 implies d (φ (tnk , vkn ) , xn ) < n. By taking
tnkn such that kn > k0 , we have
1
d φ tnkn , vknn , x ≤ d φ tnkn , vknn , xn + d (xn , x) < + d (xn , x) .
n
n n
Hence φ tkn , vkn → x as n → +∞, with tkn → +∞. This means that
n
x ∈ ω + (V ) = A, which contradicts A ⊂ int (V ) and x ∈ / int (V ).
Theorem A.3.1 If A ⊂ M is an attractor then the set

A∗ = x ∈ M : ω + (x) ∩ A = ∅
is a repeller.
Proof. Let V ⊂ M be an attractor neighborhood of A. According to

Lemma A.3.1, there is t∗ > 0 such that cl (φ ([t∗ , ∞) × V )) ⊂ int (V ). Define
the open set
V ∗ = M \ cl (φ ([t∗ , ∞) × V )) .
We claim that φ ((−∞, −t∗ ] × V ∗ ) ⊂ M \ V . Suppose to the contrary that
for some t ∈ (−∞, −t∗ ] there is v ∈ V ∗ with φ (t , v ) = x ∈ V . Then
v = φ (−t , x), with −t ∈ [t∗ , ∞) and x ∈ V , that is, v ∈ φ ([t∗ , ∞) × V ),
which contradicts v ∈ V ∗ . Hence we have ω − (V ∗ ) ⊂ M \ V , and since
M \V ⊂ M \cl (φ ([t∗ , ∞) × V )) = V ∗ , it follows that ω − (V ∗ ) ⊂ V ∗ . Thus V ∗
is an open neighborhood of ω − (V ∗ ). It remains to show that ω − (V ∗ ) = A∗ .
Indeed, if y ∈ ω − (V ∗ ) then ω + (y) ⊂ ω − (V ∗ ), hence ω + (y) ⊂ M \ V . Since
A ⊂ int (V ), it follows that ω + (y) ∩ A = ∅, thus y ∈ A∗ . On the other hand,
for any z ∈ A∗ , we have ω + (z) ∩ V = ∅, because if there was w ∈ ω + (z) ∩ V
we would have ω + (w) ⊂ ω + (z) and ω + (w) ⊂ ω + (V ) = A, which means
ω + (z) ∩ A = ∅. Thus ω + (z) ⊂ M \ V ⊂ V ∗ . Now, take w ∈ ω + (z). Then
w ∈ V ∗ and there is a sequence φ (tk , z) → w with tk → +∞. Since V ∗
is open, we may assume that φ (tk , z) ∈ V ∗ . Evidently, φ (−tk , (tk , z)) → z,
with −tk → −∞ and φ (tk , z) ∈ V ∗ , hence z ∈ ω − (V ∗ ). Thus ω − (V ∗ ) = A∗ .
The repeller A∗ is called the complementary repeller of the attractor

A. Then pair (A, A∗ ) is called an attractor-repeller pair. Since A ⊂ V
and A∗ ⊂ M \ V , A and A∗ are disjoint. For the trivial attractor M , we have
M ∗ = ∅.
The following result presents an important property of the attractor-

repeller pairs.
Proposition A.3.2 Let A ⊂ M be an attractor of the dynamical system. If

/ A ∪ A∗ then ω − (x) ⊂ A∗ and ω + (x) ⊂ A.
x∈
Proof. Let x ∈ / A ∪ A∗ . As x ∈ / A∗ , we have ω + (x) ∩ A = ∅. Take

a ∈ ω (x) ∩ A. Then there is a sequence (φ (tn , x)) that converges to a, with
+
tn → +∞. For a given attractor neighborhood V of A, we have a ∈ int (V ).

Hence there is tn0 > 0 with φ (tn0 , x) ∈ int (V ). For any y ∈ ω + (x), there is
a sequence φ (tk , x) → y, with tk → +∞. Since
φ (tk , x) = φ (tk − tn0 + tn0 , x) = φ (tk − tn0 , φ (tn0 , x))
it follows that φ (tk − tn0 , φ (tn0 , x)) → y, with tk − tn0 → +∞. As ω + (V ) =

A, we have y ∈ A, and therefore ω + (x) ⊂ A. Now, suppose by contradiction
that there is y ∈ ω − (x) such that y ∈ / A∗ . Then there is a point a ∈ A and
a sequence φ (tk , y) converging to a, with tk → +∞, and there is a sequence
φ (tn , x) converging to y, with tn → −∞. As a ∈ int (V ), there is N ∈ N such
that φ (tN , y) ∈ int (V ). By continuity, we have φ (tN + tn , x) → φ (tN , y),
with tN + tn → −∞. Then we may assume that φ (tN + tn , x) ∈ int (V ) for
all n. We have φ (− (tN + tn ) , φ (tN + tn , x)) → x, with − (tN + tn ) → +∞
and φ (tN + tn , x) ∈ V . Hence x ∈ ω + (V ) = A, which is a contradiction.
Therefore ω − (x) ⊂ A∗ .
If (A, A∗ ) is an attractor-repeller pair of the dynamical system φ then

∗
(A , A) is an attractor-repeller pair of the time-reversed dynamical system
φ∗ .
The following proposition presents a necessary and sufficient condition for
a compact invariant set to be an attractor.
Theorem A.3.2 A compact and invariant set A ⊂ M is an attractor if and

only if there exists a compact neighborhood V of A such that φx ((−∞, 0]) V
for all x ∈ V \ A.
Proof. Suppose that A is an attractor. Since A is compact and in-

variant, we can take a compact attractor neighborhood V of A. If x ∈ V
and φx ((−∞, 0]) ⊂ V then x ∈ φ ([t, +∞) × V ) for all t > 0, and hence
x ∈ ω + (V ) = A. Thus φx ((−∞, 0]) V for all x ∈ V \ A. As to the
converse, let V be a compact neighborhood of A such that φx ((−∞, 0]) V

for all x ∈ V \ A. Thus there is t∗ > 0 such that φx ((−∞, −t∗ ]) V
for all x ∈ fr (V ) = V ∩ cl (M \ V ). In fact, suppose by contradiction that
such a t∗ does not exist. Then, for each n ∈ N, there is xn ∈ fr (V ) such
that φxn ((−∞, −n]) ⊂ V . Since fr (V ) is compact, we may assume that the
sequence (xn ) converges to a point x in fr (V ). Hence, for any t ≥ 0, we have
φ (−t, x) = lim φ (−t, xn ) ∈ V

n→∞
and therefore φx ((−∞, 0]) ⊂ V with x ∈ V \ A, which contradicts the hy-

pothesis. Thus we can choose t∗ > 0 such that φx ((−∞, −t∗ ]) V for
all x ∈ fr (V ). Since A is invariant, we have φy ([0, t∗ ]) ⊂ A for every
y ∈ A. By continuity, there is an open neighborhood Vy of y such that
φ ([0, t∗ ] × Vy ) ⊂ int (V ). Hence, for x ∈ Vy , we have φx ([0, t∗ ]) ⊂ int (V ).
We claim that φx ([0, +∞)) ⊂ V for every x ∈ Vy . Indeed, we define t =

sup t : φy ([0, t]) ⊂ V and suppose by contradiction that t < +∞. Since V
is compact, we have z = φ (t , x) ∈ fr (V ). As φz ([−t∗ , 0]) ⊂ φx ([0, t ]) ⊂ V ,
we have z ∈ / fr (V ), which is a contradiction. Thus we have φx ([0, +∞)) ⊂ V .
Since x is arbitrary in Vy , we have φ ([0, +∞) × Vy ) ⊂ V . Hence the set
#
N= Vy
y∈A
is an open neighborhood of A such that φ ([0, +∞) × N ) ⊂ V , and therefore

ω + (N ) ⊂ V . It remains to show that ω + (N ) = A. Take z ∈ ω + (N ). Since
ω + (N ) is invariant, we have φz ((−∞, 0]) ⊂ ω + (N ) ⊂ V . This implies that
z ∈ A, and then ω + (N ) ⊂ A. On the other hand, since A is compact and
invariant, we have A = ω + (A) ⊂ ω + (N ). Thus A is an attractor.
This result reveals other properties of the attractors.
Corollary A.3.1 If S is an attractor and A ⊂ S is an attractor for the

dynamical system restricted to S then A is an attractor in M .
Proof. Suppose that A is an attractor in S. By Theorem A.3.2, there

is a compact neighborhood VS of A in S such that φx ((−∞, 0]) VS for
every x ∈ VS \ A. Let V ⊂ M be a compact neighborhood of A in M
such that V ∩ S = VS and N ⊂ M a compact neighborhood of S such that
φx ((−∞, 0]) N for every x ∈ N \ S. Then V ∩ N is a compact neigh-
borhood of A in M . Let x ∈ (V ∩ N ) \ A. If x ∈ S then x ∈ (V ∩ S) \ A,
hence φx ((−∞, 0]) V ∩ S. Since S is invariant, we have φx ((−∞, 0]) ⊂ S,

hence φx ((−∞, 0]) V , and therefore φx ((−∞, 0]) V ∩ N . If x ∈
/ S then
φx ((−∞, 0]) N , because x ∈ N \ S. Hence φx ((−∞, 0]) V ∩ N for every
x ∈ (V ∩ N ) \ A, and thus A is an attractor in M , according to Theorem
A.3.2.
Corollary A.3.2 Every attractor is an isolated invariant set.
Proof. Let A be an attractor. According to Theorem A.3.2, there is a

compact neighborhood V of A such that φx ((−∞, 0]) V for every x ∈ V \A.
Suppose that x ∈ V and φx (R) ⊂ V . In particular, φx ((−∞, 0]) ⊂ V , hence
x ∈ A. Thus V is an isolating neighborhood of A.
Next, we show the relationship between the attractor-repeller pairs and the
Morse decompositions. In the following, a Morse decomposition is enumerated
according to the order relation between the Morse sets.
Theorem A.3.3 A collection {C1 , ..., Cn } of subsets of M is a Morse de-

composition if and only if there exists an increasing sequence of attractors
∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M such that Cn−i = Ai+1 ∩ A∗i for i = 0, ..., n − 1.
Proof. Let M = {C1 , ..., Cn } be a Morse decomposition. For n = 1,

we have M = {M } and the sequence of attractors is ∅ = A0 ⊂ A1 = M .
Following by induction on the numbers of Morse sets, suppose that the result
holds for m with m < n. Let V be a compact neighborhood of Cn such that
V ∩ Ci = ∅ for i ∈ {1, ..., n − 1}. If x ∈ V \ Cn then ω − (x) ⊂ C1 ∪ · · · ∪ Cn−1 ,
because otherwise we would have ω − (x) , ω + (x) ⊂ Cn , which means x ∈ Cn .
Thus φx ((−∞, 0]) V for every x ∈ V \ Cn . According to Theorem A.3.2, it
follows that Cn is an attractor. For each i ∈ {1, ..., n − 1}, the Morse set Ci is
contained in the complementary repeller Cn∗ of Cn . It can be easily seen that
the collection {C1 , ..., Cn−1 } defines a Morse decomposition of the dynamical
system restricted to Cn∗ . By hypothesis, there exists an increasing sequence
of attractors
∅ = Â1 ⊂ Â2 ⊂ · · · ⊂ Ân = Cn∗
such that Cn−j = Âj+1 ∩ Â∗j for j = 1, ..., n − 1. Note that Â∗1 = Cn∗ and
Â∗n = ∅. According to Corollary A.3.1, as Â∗i is a repeller in Cn∗ and Cn∗
is a repeller in M , Â∗i is also a repeller in M for all i = 1, ..., n. Let Ai
be the complementary attractor of Â∗i in M . We obtain the sequence of

attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An , with A1 = Cn and An = M . Since
Cn ⊂ Ai , for i = 1, ..., n, we have A∗i ⊂ Cn∗ . According to Proposition A.3.1,
(Ai ∩ Cn∗ , A∗i ∩ Cn∗ ) is an attractor-repeller pair in Cn∗ . As A∗i ∩ Cn∗ = Â∗i is
the complementary repeller of Âi in Cn∗ , we have Ai ∩ Cn∗ = Âi . Hence
Ai+1 ∩ A∗i = Ai+1 ∩ Cn∗ ∩ A∗i = Âi+1 ∩ Â∗i = Cn−i
for i = 1, ..., n − 1. For i = 0, we have A1 ∩ A∗0 = Cn , and therefore Cn−i =

Ai+1 ∩ A∗i for i = 0, ..., n − 1.
As to the converse, let ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M be a sequence of
attractors such that Cn−i = Ai+1 ∩A∗i , for i = 0, ..., n−1. The elements of the
collection {C1 , ..., Cn } are the intersections of compact and isolated invariant
sets, hence they are compact and isolated invariant sets. If i < j then
Cn−i ∩ Cn−j = Ai+1 ∩ A∗i ∩ Aj+1 ∩ A∗j

= Ai+1 ∩ A∗j
⊂ Aj ∩ A∗j = ∅
hence the sets C1 , ..., Cn are pairwise disjoint. Now, for a given x ∈ M , let i
be the smallest index such that ω + (x) ⊂ Ai . Since ∅ = A∗n ⊂ A∗n−1 ⊂ · · · ⊂
A∗0 = M , we can also take the largest index j such that ω − (x) ⊂ A∗j . We have
i > 0 and j < n, and since ω + (x) Ai−1 , we have ω + (x) ∩ Ai−1 = ∅, that is,
x ∈ A∗i−1 . This implies ω + (x) ⊂ A∗i−1 , hence ω + (x) ⊂ Ai ∩ A∗i−1 = Cn−i+1 .
Analogously, as ω − (x) A∗j+1 , we have x ∈ Aj+1 , hence ω − (x) ⊂ Aj+1 , and
thus ω − (x) ⊂ Aj+1 ∩ A∗j = Cn−j . If j < i − 1 then j + 1 ≤ i − 1, hence
A∗i−1 ⊂ A∗j+1 , and thus x ∈ Aj+1 ∩ A∗i−1 ⊂ Aj+1 ∩ A∗j+1 = ∅, a contradiction.
Therefore j ≥ i − 1, which means n − j ≤ n − i + 1. In the case j = i − 1,
we have ω − (x) ∪ ω + (x) ⊂ Cn−j , which implies x ∈ Aj+1 ∩ A∗j = Cn−j .
According to Theorem A.2.1, the collection {C1 , ..., Cn } is a dynamic Morse
decomposition.
Corollary A.3.3 If M = {C1 , ..., Cn } is a Morse decomposition then there

exists an increasing sequence of attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M
such that
#n n
Ci = Ai ∪ A∗i .
i=1 i=0
Proof. By Theorem A.3.3, there exists a sequence of attractors ∅ = A0 ⊂

A1 ⊂ · · · ⊂ An = M such that Cn−j = Aj+1 ∩ A∗j for j = 0, ..., n − 1. Let
"
n
x∈ Ci . Then x ∈ Cn−j for some 0 ≤ j ≤ n − 1, hence x ∈ Aj+1 ∩ A∗j .
i=1
This implies that x ∈ A∗k , for every k ≤ j, and x ∈ Al , for every l ≥ j + 1.
"
n
n
Hence x ∈ Ai ∪ A∗i for every i = 0, 1, ..., n. Thus Ci ⊂ Ai ∪ A∗i . On the
i=1 i=0

n
other hand, take y ∈ Ai ∪ A∗i . There exists a j = 0 that is the smallest
i=0
index with ω + (y) ⊂ Aj . Then ω + (y) ∩ Aj−1 = ∅, and hence y ∈ A∗j−1 .
As ω + (y) ∩ Aj = ∅, we have y ∈ / A∗j , which means that y ∈ Aj . Hence

n "
n
y ∈ Aj ∩ A∗j−1 = Cn−(j−1) , and therefore Ai ∪ A∗i ⊂ Ci .
i=0 i=1
In the following examples, we determine some Morse decompositions by

means of attractor-repeller pairs:

The attractors in the sphere S are the following: ∅, E, Hp , Hp∗ , and S. The
complementary repellers of E, Hp , and Hp∗ are respectively E ∗ = {p, p∗ },
Hp∗ = {p∗ }, and Hp∗∗ = {p}. Taking the sequence of attractors
∅ = A0 ⊂ A1 = E ⊂ A2 = Hp ⊂ A3 = S
we have A1 ∩ A∗0 = E, A2 ∩ A∗1 = {p}, and A3 ∩ A∗2 = {S}. We then obtain

the Morse decomposition M = {E, {p} , {p∗ }}. Taking the sequence
∅ = A 0 ⊂ A1 = E ⊂ A 2 = S
we have A1 ∩ A∗0 = E and A2 ∩ A∗1 = {p, p∗ }, we then obtain the Morse

decomposition M1 = {E, {p, p∗ }}. The sequence
∅ = A 0 ⊂ A1 = H p ⊂ A2 = S
defines the Morse decomposition M2 = {Hp , {p∗ }}. Finally, the sequence of
attractors
∅ = A 0 ⊂ A1 = H p ∗ ⊂ A2 = S
determines the Morse decomposition M3 = {Hp∗ , {p}}.

The disk Dn = x ∈ M : x ≤ n1 is an attractor with attractor neighborhood
- .
Vn = x ∈ M : x < n−1 1
, for any even number n. The complementary
repeller of Dn is the set
!
∗ 1
D n = M \ Vn = x ∈ M : ≤ x ≤ 1 .
n−1
Fix an even number n and denote Ai = Dn−2(i−1) , for i = 1, ..., n/2, and
A n2 +1 = M . We obtain the sequence of attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂
A n2 +1 = M such that
C n2 +1−i = Ai+1 ∩ A∗i = Dn−2i ∩ M \Vn−2(i−1)
!
1 1
= x∈M : ≤ x ≤
n − 2i + 1 n − 2i
= Kn−2i
for i = 1, ..., n2 − 1, and
C1 = A n2 +1 ∩ A∗n2 = M ∩ C1 = C1 ,
C n2 +1A1 ∩ A∗0 = Dn ∩ M = Dn .
=

We obtain the Morse decomposition Mkn = C1 , C2 , ..., C n2 +1 .
Figure A.6: Morse decompositions of a dynamical system on the unit disk.
Now, for any natural number n > 1 (the case n = 1 is trivial), define

the set Ai = x ∈ M : 2i 1
≤ x ≤ 1 , for i = 1, ..., n − 1, and An = M .
Each set Ai is an attractor
- . and its complementary repeller is the set A∗i =
x ∈ M : x ≤ 2i+1
1
, for i = 1, ..., n − 1, and A∗n = ∅. We then obtain the
sequence of attractors ∅ = A0 ⊂ A1 ⊂ · · · ⊂ An = M such that
!
1 1
Cn−i = Ai+1 ∩ A∗i = x ∈ M : ≤ x ≤ = K2i+1
2i + 2 2i + 1
for i = 0, ..., n − 2, and C1 = An ∩ A∗n−1 = D2n−1 . Thus M = {C1 , ..., Cn } is
a Morse decomposition (see Fig. A.6).
A.4 Chain transitive sets

In this last section we study the concept of chain transitivity of dynamical
systems. The main theorem relates the chain transitivity and the Morse
decompositions.
Definition A.4.1 For x, y ∈ M and , T > 0, an (, T )-chain from x to

y is formed by sequences of points x0 = x, x1 , ..., xn = y in M and times
t0 , ..., tn−1 ≥ T such that d (φ (ti , xi ) , xi+1 ) < for every i = 0, 1, ..., n − 1.
Then an (, T )-chain from x to y is a trajectory with bounded jumps. If

there is an (, T )-chain from x to y and another from y to z then we can
concatenate them by obtaining an (, T )-chain from x to z. In particular, if
z = x then we obtain an (, T )-chain from x to x.
Definition A.4.2 A subset X ⊂ M is said to be chain transitive if for any

x, y ∈ X and , T > 0, there is an (, T )-chain from x to y. A point x ∈ M
is called chain recurrent if the single set {x} is chain transitive. The set of
all chain recurrent points is called chain recurrence set and is denoted by C.
Every chain transitive set is contained in the chain recurrence set C. In

Example A.2.1, the recurrence set consists
of the reunion E ∪ {p, p∗ }. In
"
Example A.1.6, we have C = Cn ∪ {0}.
n∈N
The dynamical system is called chain transitive if the whole set M is chain
transitive; the system is called chain recurrent if C = M .
Remark A.4.1 In Definition A.4.2, the intermediate points of an (, T )-

chain need not be contained in the chain transitive set X. In the case of
X being invariant and containing all the points of its chains, the dynamical
system restricted to X is chain transitive and X is called internally chain
transitive.
The limit sets are important cases of chain transitive sets, as in the fol-
lowing:
Proposition A.4.1 For every x ∈ M , the limit sets ω + (x) and ω − (x) are
chain transitive.
Proof. Let a, b ∈ ω + (x) and , T > 0. There are sequences φ (tk , x) → a

and φ (tl , x) → b with tk , tl → +∞. Choose T0 > T . We have φ (T0 + tk , x) →
Section A.4 · Chain transitive sets 301
φ (T0 , a). Hence there is k0 such that d (φ (T0 + tk0 , x) , φ (T0 , a)) < . We
can find l0 ∈ N such that tl0 > T + T0 + tk0 and d (φ (tl0 , x) , b) < . Denote
T1 = tl0 −T0 −tk0 > T and x1 = φ (T0 + tk0 , x). We have φ (tl0 , x) = φ (T1 , x1 ),
hence the points x0 = a, x1 , x2 = b and the times T0 , T1 > T form an (, T )-
chain form a to b. Thus ω + (x) is chain transitive. Now let a, b ∈ ω − (x)
and , T > 0. Then there are sequences φ (tk , x) → a and φ (tl , x) → b with
tk , tl → −∞. We can take l0 ∈ N such that d (φ (tl0 , x) , b) < and −tl0 > T .
As φ (−tl0 + tk , x) → φ (−tl0 , a), we can take k0 ∈ N such that tk0 < 2tl0 − T
and d (φ (−tl0 + tk0 , x) , φ (−tl0 , a)) < . Define x1 = φ (−tl0 + tk0 , x), T0 =
−tl0 , and T1 = 2tl0 −tk0 > T . We have φ (tl0 , x) = φ (T1 , x1 ), hence the points
x0 = a, x1 , x2 = b and the times T0 , T1 > T define a (, T )-chain form a to b.
Thus ω − (x) is chain transitive.
The dynamical system is called topologically transitive if ω + (x) = M

for some x ∈ M . Then every topologically transitive dynamical system is
chain transitive.
The following presents an important property of the chain recurrence set
C.
Proposition A.4.2 The chain recurrence set C is compact and invariant.
Proof. We first show that C is closed. Let x ∈ cl (C) and , T > 0. By con-
tinuity, there is δ () > 0 such that d (x, y) < δ () implies d (φ (T, x) , φ (T, y)) <
. Define δ = min {δ () , /2}. The open ball B (x, δ) contains a recur-
rent point y. Take a (δ, 2T )-chain from y to y formed by sequences y1 =
y, y2 , ..., yn = y in M and times T1 , ..., Tn−1 ≥ 2T . We have
d (φ (Tn−1 , yn−1 ) , x) ≤ d (φ (Tn−1 , yn−1 ) , y) + d (y, x) < δ + δ <
and
d (φ (T1 − T, φ (T, y)) , y2 ) = d (φ (T1 , y) , y2 ) < δ <
with T1 − T ≥ T . Then the points x0 = xn = x, x1 = φ (T, y) , xj = yj and
the times t0 = T, t1 = T1 − T, tj = Tj , for j = 2, ..., n − 1, form an (, T )-chain
from x to x. Hence x ∈ C, and therefore cl (C) = C. We now show that C is
invariant. Let x ∈ C and t ∈ R. By continuity and compactness, the map φt
is uniformly continuous. For , T > 0, there is δ > 0 such that d (y, z) < δ
implies d (φ (t, y) , φ (t, z)) < . Moreover, for δ, T > 0 there are sequences
of points x0 = x, x1 , ..., xk = x in M and times T0 , ..., Tk−1 ≥ T such that
d (φ (Ti , xi ) , xi+1 ) < δ for i = 0, ..., k − 1. We then obtain the sequence of
points φ (t, x0 ) = φ (t, x) , φ (t, x1 ) , ..., φ (t, xk ) = φ (t, x) in M and the times
T0 , ..., Tk−1 ≥ T such that
d (φ (Ti , φ (t, xi )) , φ (t, xi+1 )) = d (φ (t, φ (Ti , xi )) , φ (t, xi+1 )) < .
Hence there is an (, T )-chain from φ (t, x) to φ (t, x), and therefore φ (t, x) ∈
C. This proves that C is invariant.
A set X ⊂ M is called a maximal chain transitive set if it is a chain

transitive set that is not contained properly in another chain transitive set.
The following result shows that every maximal chain transitive set is compact.
Proposition A.4.3 If a set X ⊂ M is chain transitive then its closure has

the same property.
Proof. If X ⊂ C then cl (X) ⊂ C, since C is compact by Proposition A.4.2.

For x, y ∈ cl (X) and , T > 0, take x ∈ X such that d (x , x) < /2. Consider
an (/2, T )-chain from x to x formed by the sequences x0 = x, x1 , ..., xn = x
in M and T0 , ..., Tn−1 ≥ T . Replace xn = x by xn = x . We then have

d (φ (Tn−1 , xn−1 ) , x ) ≤ d (φ (Tn−1 , xn−1 ) , x) + d (x, x ) < + =
2 2
hence we obtain an (, T )-chain from x to x . On the other hand, we can choose
y ∈ X such that d (y , y) < /2. By taking an (/2, T )-chain from x to y
formed by points x0 = x , x1 , ..., xm = y in M and times T0 , ..., Tm−1

≥ T,

we replace xm = y by xm = y, and then we have

d φ Tm−1 , xm−1 , y ≤ d φ Tm−1 , xm−1 , y + d (y , y) < + = .
2 2
This provides an (, T )-chain from x to y. Finally, we can construct an (, T )-

chain from x to y by taking the concatenation of the (, T )-chains from x to
x and from x to y. Therefore cl (X) is chain transitive.
According to Propositions A.4.2 and A.4.3, we have the following proper-

ties of the maximal chain transitive sets.
Corollary A.4.1 Every maximal chain transitive set is compact and invari-
ant.
Proof. Let X ⊂ M be a maximal chain transitive set. According to

Proposition A.4.3, cl (X) is chain transitive, the maximality of X then im-
plies cl (X) = X. To show that X is invariant, let x ∈ X and t ∈ R. According
to Proposition A.4.2, we have φ (t, x) ∈ C. Let , T > 0. Since the map φt
is uniformly continuous, there is δ () > 0 such that d (y, z) < δ () implies
d (φt (y) , φt (z)) < . Define δ = min {δ () , } and t = max {T, T − t}. Con-
sider a (δ, t )-chain from x to x formed by sequences x0 = x, x1 , ..., xn = x
in M and t0 , ..., tn−1 ≥ t . As d (φ (tn−1 , xn−1 ) , x) < δ < δ (), we have
d (φ (t + tn−1 , xn−1 ) , φ (t, x)) < with t + tn−1 ≥ T . Replacing xn = x by
xn = φ (t, x), and tn−1 by t+tn−1 , we obtain an (, T )-chain from x to φ (t, x).
Similarly, we can obtain an (, T )-chain from φ (t, x) to x. By the maximality
of X, this means that φ (t, x) ∈ X. Thus X is invariant.
The following sequence of results shows that the connected components of

the chain recurrence set C coincide with the maximal chain transitive sets.
Proposition A.4.4 Let X ⊂ M be a nonempty, compact, and connected set.

If X is chain recurrent then it is chain transitive.
"
n
Proof. Let x, y ∈ X and , T > 0. Take a finite cover X ⊂ B (xi , /4),
i=1
with x1 , ..., xn ∈ X. For each i, consider an (/4, T )-chain from xi to xi , that
is, points xi0 = xi , xi1 , ..., xiki = xi in M and times T0i , ..., Tkii −1 ≥ T such that

d φ Tjii , xiji , xiji +1 <
4
for ji = 0, ..., ki − 1. Since

d φ Tkii −1 , xiki −1 , xi <
4
the connectedness of X implies that there is a point xj with j ∈ {1, ..., n}
such that
3
d φ Tkii −1 , xiki −1 , xj < .
4
Denoting xiki = xj , we obtain an (, T )-chain from xi to xj . By taking con-
catenations, we can construct an (, T )-chain from x to y. Therefore, X is
chain transitive.
As a converse for this result we have the following.

Proposition A.4.5 Let X ⊂ M be a compact invariant set. If the dynamical

system restricted to X is chain transitive, then X is connected.
Proof. Suppose that the system is chain transitive in X and X is discon-

nected. There are then nonempty sets A, B ⊂ M which are disjoint, compact,
and X = A ∪ B. Hence we can define
0 = inf {d (a, b) : a ∈ A, b ∈ B} > 0.
This means that, for = 0 /2, there is no (, T )-chain from A to B, which
contradicts the hypothesis on X.
In particular, if the dynamical system on M is chain transitive then M

must be connected.
For the following result, it should be remembered that the hyperspace
H (M ) of all nonempty compact sets of M is a compact metric space endowed
with the Hausdorff metric
!
dH (A, B) = max max mind (a, b) , max mind (a, b) .
a∈A b∈B b∈B a∈A
Proposition A.4.6 Let X ⊂ M be a maximal chain transitive set. The

dynamical system restricted to X is chain transitive. In particular, the dy-
namical system restricted to C is chain recurrent.
Proof. Let x, x ∈ X. For each p ∈ N and T > 0, there are points x0 =

x, x1 , ..., xnp = x ∈ M and times T0p , ..., Tnpp −1 ≥ T such that
d (φ (Tip , xi ) , xi+1 ) < 1/p for i = 0, ..., np − 1. Define
#
np
Kp = φxi ([0, Tip ])
i=0
with Tnpp ≥ 0. We obtain a sequence of compact sets {Kp }p∈N such that
x, x ∈ Kp for all p. We may assume that the sequence (Kp )p∈N in H (M )
converges to K ∈ H (M ). This means that x, x ∈ K. For y, z ∈ K and q ∈ N,
we claim there is an (1/q, T )-chain in K from y to z. In fact, by the uniform
continuity of the maps φTip , for i = 0, ..., np − 1, there is a δ > 0 such that
δ < 1/3q and d (a, b) < δ implies
1
d (φ (Tip , a) , φ (Tip , b)) < .
3q
By the Hausdorff convergence of Kp , we can take a p such that

p > max {3q, 1/δ} and dH (Kp , K) < δ. Hence we can find xi ∈ K such
that d (xi , xi ) < δ, for any i = 0, ..., np − 1, and we can choose xnp ∈ K with
$ %
d xnp , xnp < δ. The points x0 , x1 , ..., xnp ∈ K and the times T0p , ..., Tnpp −1 ≥
T satisfy

d φ (Tip , xi ) , xi+1 ≤ d (φ (Tip , xi ) , φ (Tip , xi )) + d (φ (Tip , xi ) , xi+1 ) +

+d xi+1 , xi+1
1 1 1 1 1 1
< + +δ < + + =
3q p 3q 3q 3q q
thus they form an (1/q, T )-chain in K form x0 to xnp . In particular, we

can consider x0 = x and xnp = x . Hence, for any , T > 0, we can find
q ∈ N with 1/q < by obtaining an (, T )-chain from x to x . Note that
both the points x and x must be contained in the same connected compo-
nent of K. We may then assume that K is connected. Now, take x0 , xnp ∈ K
$ %
such that d (x0 , x0 ) < δ and d xnp , xnp < δ , with δ = min {δ, 1/5q} and

p > max 5q, 1/δ . We obtain an (1/q, T )-chain from x0 to xnp . We can
cover K by 1/q-balls and apply this process successively by constructing an
(1/q, T )-chain in K from y to z, for all q ∈ N. Thus K is chain transitive.
Finally, the condition X ∩ K = ∅ implies that X ∪ K is chain transitive. The
maximality of X then implies that K ⊂ X. This means that there is an
(, T )-chain in X from x to x , for any , T > 0. Therefore, the dynamical
system restricted to X is chain transitive.
We now show a topological characterization of the maximal chain transi-

tive sets.
Theorem A.4.1 The maximal chain transitive sets of the dynamical system
coincide with the connected components of the chain recurrence set C. Con-
sequently, the dynamical system restricted to a connected component of C is
chain transitive.
Proof. Let C be a connected component of C. Then C is compact,

connected, and chain recurrent. According to Proposition A.4.4, C is chain
transitive. Let D be a maximal chain transitive set containing C. As D is
compact and invariant, the dynamical system restricted to D is chain tran-
sitive, according to Proposition A.4.6. According to Proposition A.4.5, D is
connected, and thus D = C. Conversely, let Y be a maximal chain transitive

set. Y is then connected and is contained in C. This means that Y is con-
tained in some connected component C of C. As C is chain transitive, the
maximality of Y implies C = Y .
Because of this result, a connected component of C is usually called a

chain component of the dynamical system.
Finishing this appendix, we show the relationship between the notions
of chain transitive sets and Morse decompositions. We need the following
definition.
Definition A.4.3 For a given set X ⊂ M we define
Ω+ (X, , T ) = {y ∈ M : there is an (, T ) -chain from a point x of X to y} .
The set
!
+ y ∈ M : for any , T > 0, there is an (, T ) -chain
Ω (X) =
from a point x of X to y
is called positive chain limit set of X.

It can be easily seen that Ω+ (X) = Ω+ (X, , T ), and if X is chain
,T >0
recurrent then X ⊂ Ω+ (X). In particular, a point x is chain recurrent if and
only if x ∈ Ω+ (x).
The negative chain limit set Ω− (x) may be defined in such a way that
y ∈ Ω− (x) if and only if x ∈ Ω+ (y). Thus we concentrate the study on the
positive chain limit sets.
In the following results we present the main properties of the sets Ω+ (X, , T )
and Ω+ (X).
Proposition A.4.7 For any X ⊂ M and , T > 0, the set Ω+ (X, , T ) is

open.
Proof. Let y ∈ Ω+ (X, , T ). There exists a point x ∈ X together with

sequences x0 = x, ..., xn = y in M and times T0 , ..., Tn−1 ≥ T such that
d (φ (Ti , xi ) , xi+1 ) < , for i = 0, ..., n−1. By defining r = d (φ (Tn−1 , xn−1 ) , y),
let r < r < . For any z ∈ B (y, r − r), we have
d (φ (Tn−1 , xn−1 ) , z) ≤ d (φ (Tn−1 , xn−1 ) , y) + d (y, z) < r + r − r = r < .

Hence we obtain an (, T )-chain from x to z, with x ∈ X, that is, z ∈

Ω+ (X, , T ). This means that B (y, r − r) ⊂ Ω+ (X, , T ), and therefore
Ω+ (X, , T ) is open.
Proposition A.4.8 For any X ⊂ M , one has ω + (X) ⊂ Ω+ (X), and for
any , T > 0, one has Ω+ (X) ⊂ ω + (Ω+ (X, , T )).
Proof. If y ∈ ω + (X) there is then a sequence φ (tk , xk ) → y with xk ∈ X

and tk → +∞. For , T > 0, there is a K ∈ N such that k ≥ K implies
d (φ (tk , xk ) , y) < . As tk → +∞, we can choose tk0 > max {tK , T }. Hence
xk0 ∈ X, tk0 > T , and d (φ (tk0 , xk0 ) , y) < , which form an (, T )-chain from
xk0 to y. This proves that y ∈ Ω+ (X), and thus ω + (X) ⊂ Ω+ (X). Now, let
z ∈ Ω+ (X). For each n ∈ N, there is a point xn ∈ X and an (1/n, n)-chain
from xn to z formed by sequences xn0 = xn , xn1 , ..., xnkn = z in M and times

T0n , ..., Tknn −1 ≥ n. The points φ Tknn −1 , xnkn −1 then define a sequence such
that
1
d φ Tknn −1 , xnkn −1 , z < .
n

Hence we have φ Tknn −1 , xnkn −1 → z, as n → +∞, with Tknn −1 → +∞ and
xnkn −1 ∈ Ω+ (X, 1/n, n). Finally, for any , T > 0, we can take n such that

1/n < and n > T , and we then obtain a sequence φ Tknn −1 , xnkn −1 → z,
with Tknn −1 → +∞ and xnkn −1 ∈ Ω+ (X, 1/n, n) ⊂ Ω+ (X, , T ). This proves
that z ∈ ω + (Ω+ (X, , T )), and thus Ω+ (X) ⊂ ω + (Ω+ (X, , T )), for any
, T > 0.
Proposition A.4.9 For any set X ⊂ M , the chain limit set Ω+ (X) is the
intersection of all attractors containing the limit set ω + (X).
Proof. For any , T > 0, define the set

V,T = cl Ω+ (X, , T ) .
If y ∈ ω + (V,T ) then there is a sequence φ (tn , xn ) → y with xn ∈ V,T and

tn → +∞. Choose n0 ∈ N such that tn0 > T and d (φ (tn0 , xn0 ) , y) < /2.
By the uniform continuity of the map φtn0 , there is δ > 0 such that d (a, b) <
δ implies d (φ (tn0 , a) , φ (tn0 , b)) < /2. As xn0 ∈ V,T , the ball B (xn0 , δ)
contains a point z of Ω+ (X, , T ) and

d (φ (tn0 , xn0 ) , φ (tn0 , z)) < .
2
We then have
d (φ (tn0 , z) , y) ≤ d (φ (tn0 , z) , φ (tn0 , xn0 )) + d (φ (tn0 , xn0 ) , y)

< + = .
2 2
Now, taking x ∈ X and an (, T )-chain from x to z, we obtain an (, T )-
chain from x to y. Hence y ∈ Ω+ (X, , T ), and therefore ω + (V,T ) ⊂
Ω+ (X, , T ). According to Proposition A.4.7, the set Ω+ (X, , T ) is open.
Hence Ω+ (X, , T ) ⊂ int (V,T ), and thus ω + (V,T ) ⊂ int (V,T ). This means
that ω + (V,T ) is an attractor with attractor neighborhood V,T . Moreover we
have

ω + (X) ⊂ ω + Ω+ (X, , T ) ⊂ ω + cl Ω+ (X, , T ) = ω + (V,T )
for all , T > 0, and by Proposition A.4.8 we have

Ω+ (X) ⊂ ω + Ω+ (X, , T ) ⊂ ω + (V,T ) .
,T >0 ,T >0
On the other hand, as ω + (V,T ) ⊂ Ω+ (X, , T ), we have

ω + (V,T ) ⊂ Ω+ (X, , T ) = Ω+ (X) .
,T >0 ,T >0
+
+
Hence Ω (X) = ω (V,T ). It remains to show that every attractor A
,T >0
containing ω + (X) also contains Ω+ (X). Indeed, take an attractor neighbor-
hood V of A. By Lemma A.3.1, there is t∗ > 0 such that cl (φ ([t∗ , ∞) × V )) ⊂
int (V ). We have A ⊂ cl (φ ([t∗ , ∞) × V )), since ω + (V ) = A. Consider
0 < < inf {d (a, b) : a ∈ cl (φ ([t∗ , ∞) × V )) , b ∈ M \ int (V )}
and an /2-neighborhood N of cl (φ ([t∗ , ∞) × V )). As ω + (X) ⊂ A, we can

choose a time T such that φ ([T, ∞) × V ) ⊂ N and T > t∗ . Let y ∈ Ω+ (X).
For each n ∈ N with n > T and 1/n < /2, there are points xn0 , xn1 , ..., xnkn = y
in M with xn0 ∈ X and times T0n , ..., Tknn −1 ≥ n such that
1
d φ Tinn , xnin , xnin +1 < < .
n 2
Since φ (T0n , xn0 ) ∈ N , we can take a point z ∈ cl (φ ([t∗ , ∞) × V )) such that
d (φ (T0n , xn0 ) , z) < /2. We then have

d (z, xn1 ) ≤ d (z, φ (T0n , xn0 )) + d (φ (T0n , xn0 ) , xn1 ) < + = .
2 2
This means that xn1 ∈ int (V ). Now, φ (T1n , xn1 ) ∈ cl (φ ([t∗ , ∞) × V )), because
T1n > t∗ and xn1 ∈ V . Hence d (φ (T1n , xn1 ) , xn2 ) < 1/n < /2, and thus we have

xn2 ∈ V . Following by induction we obtain d φ Tknn −1 , xnkn −1 , y < 1/n,

with xnkn −1 ∈ V . We have φ Tknn −1 , xnkn −1 → y, as n → +∞, with Tknn −1 →
+∞ and xnkn −1 ∈ V . Hence y ∈ ω + (V ) = A, and therefore Ω+ (X) ⊂ A.
Corollary A.4.2 The dynamical system on M is chain transitive if and only

if the whole space M is the unique nonempty attractor.
Proof. Suppose that the system is chain transitive and let A ⊂ M be

a nonempty attractor. For a given x ∈ A, we have ω + (x) ⊂ A. By chain
transitivity, we have Ω+ (x) = M , and M is then the intersection of all the
attractors containing ω + (x), according to Proposition A.4.9. Hence A = M .
As to the converse, suppose that M is the unique nonempty attractor of the
system. According to Proposition A.4.9, it follows that Ω+ (x) = M for every
x ∈ M . Thus the system is chain transitive.
The last theorem of this appendix shows the relationship between chain
transitivity and Morse decomposition. This is known as the Conley main
theorem of dynamical systems.
Theorem A.4.2 The chain recurrence set of the dynamical system is de-
scribed by
C= {A ∪ A∗ : A is an attractor} .
Consequently, there are finitely many chain components if and only if there
exists the finest Morse decomposition of the dynamical system. In this case,
the Morse sets coincide with the chain components and the dynamical system
restricted to each Morse set is chain transitive.
Proof. Let x ∈ C. According to Proposition A.4.9, x belongs to the

intersection of all the attractors which contain ω + (x). Let B be an attractor
that does not contain ω + (x). Then x ∈ / B ∗ , we have ω + (x) ⊂
/ B. If also x ∈
B, according to Proposition A.3.2, which is a contradiction. Hence x ∈ B ∗ .
This means that x ∈ A ∪ A∗ for every attractor A of the system, and thus

C ⊂ {A ∪ A∗ : A is an attractor}. On the other hand, take a point y in the

intersection {A ∪ A∗ : A is an attractor}. If ω + (y) is contained in some
attractor B then y ∈ / B ∗ , hence y ∈ B. Thus y belongs to all attractors which
contain ω + (y). According to Proposition A.4.9, it follows that y ∈ Ω+ (y).

Thus {A ∪ A∗ : A is an attractor} ⊂ C, and we have the desired description
of C.
Now, suppose there is the finest Morse decomposition M = {C1 , ..., Cn }.
Then M is the intersection of all Morse decompositions of the system. Ac-
cording to Corollary A.3.3, it follows that
#
n
Ci = {A ∪ A∗ : A is an attractor} = C.
i=1
We claim that the Morse sets coincide with the chain components. Let x, y ∈
Ci for some i. Then ω + (x) ⊂ Ci , and hence Ci is contained in the intersection
of all attractors which contain ω + (x). By Proposition A.4.9, this means that
Ci ⊂ Ω+ (x), and hence y ∈ Ω+ (x). Thus Ci is chain transitive, which implies
"
n
that Ci is contained in some chain component. Since Ci = C, this means
i=1
that the Morse sets coincide with the chain components of C. As to the
converse, suppose that C has a finite number of chain components C1 , ..., Cn .
Then {C1 , ..., Cn } is a collection of compact, invariant, and pairwise disjoint
sets. Since every limit set is chain transitive, we also have ω + (x) , ω − (x) ⊂
"
n
Ci for every x ∈ M . Now, suppose there are sequences Ci0 , ..., Cik and
i=1
"
n
x1 , ..., xk ∈ M \ Ci such that ω − (xj ) ⊂ Cij −1 and ω + (xj ) ⊂ Cij , for
i=1
j = 1, ..., k, and Ci0 = Cik . Then ω − (x1 ) , ω + (xk ) ⊂ Ci0 . Choose a ∈ ω − (x1 )
and take a sequence φ (tl , x1 ) → a with tl → −∞. For , Ṫ > 0, take T0 > T .
We have φ (T0 + tl , x1 ) → φ (T0 , a), with T0 + tl → −∞. We can choose l0
such that −T0 − tl0 > T and d (φ (T0 + tl0 , x1 ) , φ (T0 , a)) < . By denoting
y1 = φ (T0 + tl0 , x1 ) and T1 = −T0 − tl0 > T , the points y0 = a, y1 , y2 = x1
and the times T0 , T1 > T satisfy
d (φ (T0 , a) , y1 ) < and d (φ (T1 , y1 ) , y2 ) = d (x1 , x1 ) = 0 <
which form an (, T )-chain from a to x1 . We now take b ∈ ω + (xk ). Then

b ∈ Ω+ (xk ) ∩ Ci0 . As Ci0 is chain transitive and contains a, there is an (, T )-
chain from b to a. Hence we can construct an (, T )-chain from xk to x1 for
any , T > 0. In a similar way, since ω − (xj+1 ) , ω + (xj ) ⊂ Cij , there is an
(, T )-chain from xj to xj+1 , for all , T > 0 and j = 1, ..., k − 1. Thus we ob-
tain an (, T )-chain from x1 to xk for all , T > 0. This means that both points
"
n
x1 and xk are chain recurrent, contradicting the condition x1 , xk ∈ / Ci = C.
i=1
Section A.5 · Notes and references 311
Thus the chain components satisfy the no-cycle condition. It remains to show
that the chain components are isolated invariant. Indeed, for each i = 1, ..., n,
let Vi be an open neighborhood of Ci such that V1 , ..., Vn are pairwise disjoint.
Suppose that φx (R) ⊂ Vi . As ω + (x) ⊂ Cj for some j, there is t > 0 such that
φ (t, x) ∈ Vj . This implies that Cj = Ci , and thus ω + (x) ⊂ Ci . Analogously,
we can show that ω − (x) ⊂ Ci . Hence ω + (x) , ω − (x) ⊂ Ci , and therefore
x ∈ Ci . This means that Vi is an isolating neighborhood of Ci . Therefore
{C1 , ..., Cn } is a Morse decomposition. Finally, we show that this is the finest
Morse decomposition. Suppose that M = {C1 , ..., Cm } is a Morse decompo-
sition finer than {C1 , ..., Cn }. For a given Morse set Cj , there is some chain
"
n "
m
component Ci such that Cj ⊂ Ci . Since Ci = C ⊂ Ci , we have Ci ⊂ Cj
i=1 i=1
for some j . Hence Cj ⊂ Ci ⊂ Cj , and thus Cj = Cj = Ci . This means that
M = {C1 , ..., Cm }, and therefore the chain components form the finest Morse
decomposition of the system.
In Example A.2.1, the finest Morse decomposition of the dynamical system

on the sphere S is given by M = {E, {p} , {p∗ }}. In this case, the sets
E, {p} , {p∗ } are the chain components of the system. In Example A.2.2, the
dynamical system on the unit disk does not have finest Morse decomposition.
The chain components are the cycles with radius 1/n together with the origin
0 of R2 , and thus there are infinite chain components.
A.5 Notes and references

The Conley main theorem of dynamical systems was extended to the setting
of semiflows on topological spaces (see e.g. Hirch [61], Patrão and San Martin
[79, 80]). For linear flows on projective spaces, Rasmussen [82] defined the
all-time Morse decompositions. More recently, Zhenxin Liu [114] obtained
Conley’s decomposition theorem of infinite dimensional random dynamical
systems and introduced the Morse decompositions for random semiflows (see
also [115] for results on random flows). Many questions on flows or semiflows
were solved throughout the theory of semigroup actions (as reference sources
we mention [21],[23],[24],[48],[79],[80],[81]).
Appendix B
Differentiable manifolds
Differential equations can be extended straightforwardly to the framework

of the differentiable manifolds. In fact, the differentiable manifolds generalize
the Euclidean spaces and were specially idealized in order to provide basic
structures for vector fields, metrics, and measures, in a larger variety of spaces.
The intention of this appendix is to provide a brief description of differentiable
manifolds. We refer to Boothby [16] or Warner [112] for the unexplained
concepts and results.
Let M be a topological space. A chart for M is a pair (φ, U ) such that
U is an open subset of M and the map φ : U → V is a homeomorphism onto
an open subset V of Rd for some d ∈ N which is called the dimension of
(φ, U ). A chart (φ, U ) is said to be a chart around x ∈ M if x ∈ U . A family
of d-dimensional charts A = {(φα , Uα )}α∈Λ on M is called a C ∞ -atlas of
dimension d if
1. {Uα }α∈Λ covers M ;
2. For all α, β ∈ Λ, the map φβ ◦ φ−1 ∞
α : φα (Uα ∩ Uβ ) → φβ (Uα ∩ Uβ ) is C .
A d-dimensional chart (φ, U ) for M is admissible relative to an atlas A =

{(φα , Uα )}α∈Λ of dimension d on M if for all α ∈ Λ with U ∩ Uα = ∅, the
maps φ ◦ φ−1 α and φα ◦ φ
−1
are C ∞ . We say that a d-dimensional atlas
A = {(φα , Uα )}α∈Λ on M is maximal if all admissible d-dimensional charts
relative to A are contained in A.
Definition B.0.1 A pair (M, A) is called a d-dimensional differentiable
manifold (of class C ∞ ) or simply manifold if the set M is a second-
313
314 Chapter B · Differentiable manifolds
countable Hausdorff space provided with a maximal C ∞ -atlas A of dimension

d. The natural number d is called the dimension of M and A is a differen-
tiable structure on M . Each chart (φα , Uα ) is called a local coordinate
system for M .
The differentiable structure enables the definition of differentiable maps
between manifolds, as in the following:
Definition B.0.2 A map f : M → N between two manifolds (M, A) and
(N, B) is said to be differentiable at x ∈ M if there are charts (φ, U ) ∈ A
around x and (ψ, V ) ∈ B around f (x) such that f (U ) ⊂ V and the local
representation ψ ◦ f ◦ φ−1 : φ (U ) → ψ (V ) of f is differentiable (of class
C ∞ ) on φ (x). The map f is differentiable if it is differentiable at all x ∈ M .
A differentiable bijection f : M → N whose inverse f −1 : N → M is also
differentiable is called a diffeomorphism.
A fundamental fact is that the definition of a differentiable map does not
depend on the choice of the charts.
We select below some classical examples of manifolds.
1. It is easily seen that the d-dimensional Euclidean space Rd is a manifold.
2. Every open subset N of a d-dimensional manifold (M, A) is itself a d-
dimensional manifold (open submanifold) with atlas
{( φ|U ∩N , U ∩ N ) : (φ, U ) ∈ A} .
3. Given two manifolds (M, A) and (N, B) of dimensions k and l, respec-

tively, their Cartesian product M ×N (endowed with the product topol-
ogy) becomes a (k + l)-dimensional manifold with the maximal atlas
which contains the product atlas
{(φ × ψ, U × V ) : (φ, U ) ∈ A, (ψ, V ) ∈ B} .
A manifold of this type is called a product manifold.
4. Let Sn ⊂ Rn+1 be the unit sphere (endowed with the relative topol-
ogy of Rn+1 ). Take the stereographic projections φ1 : U1 = Sn \
{(0, ..., 0, 1)} → Rn and φ−1 : U−1 = Sn \ {(0, ..., 0, −1)} → Rn given by
1
φ1 (x1 , ..., xn+1 ) = − (x1 , ..., xn ) ,
xn+1 − 1
1
φ−1 (x1 , ..., xn+1 ) = (x1 , ..., xn ) .
xn+1 + 1
315
Then Sn is a manifold of dimension n with the maximal atlas which

contains the atlas (φ1 , U1 ) , φ−1 , U−1 . In special, the unit circle
S1 can be viewed as a subset of the complex plane C and it can be
then identified with the interval [0, 2π) by the map f : [0, 2π) → S1 ,
f (x) = eix .
5. The n-torus is the product manifold Tn = S1 × · · · × S1 . It can be

identified with the quotient space Rn / Zn with respect to the following
equivalence relation: x ∼ y ⇔ y = x + z for some z ∈ Zn . For n = 2,
T2 can be represented by a square [0, 1] × [0, 1] ⊂ R2 whose right and
left edges are identified, as well as its top and bottom edges.
6. Consider the following equivalence relation in Rn+1 \ {0}: x ∼ y if

and only if there is λ = 0 such that y = λx. The set of all equiv-

alence classes Pn = [x] : x ∈ Rn+1 \ {0} endowed with the quotient
topology is called real projective space of dimension n. For each
i ∈ {1, ..., n + 1}, the set Ui = {[x1 , ..., xn+1 ] : xi = 0} is an open set in
Pn and the function φi : Ui → Rn given by

x1 xi−1 xi+1 xn+1
φi ([x1 , ..., xn+1 ]) = , ..., , , ...,
xi xi xi xi
define a chart for Pn . Thus Pn is a manifold with the maximal

atlas which contains the atlas {(φi , Ui ) : i = 1, ..., n + 1}. In the one-
dimensional case, P1 is diffeomorphic with S1 .
7. A Lie group is a topological group provided with a differentiable struc-

ture such that the product p : G × G → G is a differentiable map from
the product manifold G × G onto G. A classical Lie group is the lin-
ear group Gl (n, R) of all invertible n × n real matrices with the usual
matrix product. This is an open submanifold of the space of n × n ma-
2
trices Mn×n (R) Rn . According to the Cartan theorem, every closed
subgroup of a Lie group is itself a Lie group (see [111, 112]).
A curve on a manifold M is a differentiable map c : I → M defined on

an open interval I of R. Let x be an element of a d-dimensional manifold
(M, A) and denote by Cx the set of all curves c : I → M such that 0 ∈ I and
c (0) = x. For a given c ∈ Cx and a chart (φ, U ) ∈ A around x, every time we
write φ ◦ c we are admitting that the domain of c was consciously reduced to
a smaller open interval I , containing 0, such that c (I ) ⊂ U . On Cx we can

define the following equivalence relation:
c1 ∼ c2 ⇐⇒ there is (φ, U ) ∈ A such that

d d
(φ ◦ c1 ) (t)|t=0 = (φ ◦ c2 ) (t)|t=0 .
dt dt
d d
It is worth noting that the equality (φ ◦ c1 ) (t)|t=0 = (φ ◦ c2 ) (t)|t=0 is
dt dt
independent of the choice of the chart. We then say that the equivalence class
[c] of a curve c ∈ Cx is a tangent vector at x. The quotient set Cx / ∼ is
indicated by Tx M and is called tangent space at x.
The set Tx M has a natural structure of real vector space by requiring that
for a given (φ, U ) ∈ A the well-defined bijection φ̄ : Tx M → Rd given by
d
φ̄ ([c]) = (φ ◦ c) (t)|t=0
dt
is an isomorphism. The operation so defined does not depend on the choice
of (φ, U ). The preimages of the standard basis vectors e1 , . . . , ed ∈ Rd under
∂ ∂
φ̄ form a basis of Tx M . They are denoted by ,..., . For a given dif-
∂x1 ∂xd
ferentiable map f : M → N between the manifolds M and N , the derivative
of f at x ∈ M is the well-defined linear map dfx : Tx M → Tf (x) N given by
dfx ([c]) = [f ◦ c].
The set #
T M := ({x} × Tx M )
x∈M
is called the tangent bundle of M . Each chart (φ, U ) ∈ A induces a map

T φ (x, [c]) = (φ (x) , dφx ([c])) ∈ φ (U ) × Rd . The preimages

U = T φ−1 φ (U ) × Rd for all (φ, U ) ∈ A generate a topology on T M such
that T M becomes a 2d-dimensional manifold with the maximal atlas which

contains the atlas T φ, U : (φ, U ) ∈ A .
A vector field on a manifold M is a map X : M → T M such that
for each x ∈ M , X(x) ∈ Tx M . In a local coordinate system, we can write
d ∂
any vector fields X as X = i=1 ai , where each coefficient ai is a real-
∂xi
valued function. The vector field X on M is differentiable (of class C ∞ ) if
X : M → T M is differentiable as a map between the manifolds M and T M .
This concept allows us to study ordinary differential equations on a manifold.
Indeed, for given a vector field X on M , there is an ordinary differential
317
equation associated to X in the natural way
d
x (t) = X (x(t))
dt
which is indicated in the usual form ẋ = X (x).
A Riemannian metric on a connected manifold M is a correspondence
g which associates to each point x ∈ M an inner product g (·, ·)x , that is, a
symmetric, bilinear, positive-definite form on the tangent space Tx M which
varies differentially in the following sense: if (φ, U ) is a chart around x with
∂
y = φ−1 (x1 , . . . , xd ) and dφ−1 y (0, . . . , 1, . . . , 0) = (y), then
∂xi

∂ ∂
gi,j (x1 , . . . , xd ) := g (y), (y)
∂xi ∂xj y
is a differentiable function on φ (U ). This definition does not depend on the

choice of the chart. It is possible to show that every differentiable manifold
admits a Riemannian metric. A Riemannian manifold is a manifold M
endowed with a Riemannian metric g, and we represent it by a pair (M, g).
Let (M, g) be a Riemannian manifold and define the norm ·x on Tx M
by :
vx = g (v, v)x .
For a given curve α : I → M and an interval [a, b] ⊂ I, the length of the

segment α ([a, b]) is defined as

b d
lab (α) = α (t) dt.
dt
a α(t)
The Riemannian distance d (x, y) between the points x, y ∈ M is given by

d (x, y) := inf l01 (α)
α
where the infimum is taken over all curves α : [0, 1] → M with α (0) = x and
α (1) = y. This metric is compatible with the topology of M .
The Riemannian volume of a Borel set A of a d-dimensional Riemannian
manifold (M, g) which is contained in the domain of a chart (φ, U ) is defined
as :
vol (A) := det [gi,j (x))] dx,
φ(A)
where the integral is the usual Lebesgue integral on Rd . This definition is

independent of the chosen chart. The Riemannian volume is then extended
naturally to all Borel sets of M and

ϕ dvol = ϕ ◦ f |det df | dvol,
f (A) A
for any diffeomorphism f : M → M and any integrable function ϕ : M → R.

Bibliography
[1] Alves, R. W. M. and Souza, J. A., Hyperconvergence in topological

dynamics, arXiv:1806.09434.
[2] Arnold, L. and Kliemann, W., On unique ergodicity for degenerate dif-
fusions, Stochastics and Stochastics Reperts, 21 (1984), 41–61.
[3] Atiyah, M. F. and MacDonald, I. G., Introduction to Commutative Al-

gebra, Westview Press, 1969.
[4] Auslander J.; Bhatia, N.P.; Seibert, P., Attractors in Dynamical Sys-
tems. Bol. Soc. Mat. Mex., 9 (1964), 55–66.
[5] Ayala, V., Controllability of Nilpotent Systems, Banach Center Publi-

cations, 32 (1995), 35–46.
[6] Bacciotti, A. and Kalouptsidis, N., Topological dynamics of control

systems: stability and attraction, J. Nonlinear Anal. Theory, Methods
Appl., 10 (1986), 547–565.
[7] Bacciotti, A.; Kalouptsidis, N.; Tsinias, J., Lyapunov functions and
stability of dynamical polysystems, Math. Systems Theory, 19 (1987),
333–354.
[8] Bartsch, T., ”Conley index”, in Hazewinkel, M., Encyclopedia of Math-

ematics, Springer Science+Business Media B.V. / Kluwer Academic
Publishers, 2001.
[9] Basto Gonçalves, J., Geometric conditions for local controllability, J.

Diff. Equations, 89 (1991), 388–395.
[10] Bellicanta, L. S., Trajetórias periódicas em sistemas de controle afins.

Universidade de São Paulo, thesis, 2002.
319
320 BIBLIOGRAPHY
[11] Bhatia N.P. and Hajek, O., Local Semi-Dynamical Systems, Lecture
Notes in Mathematics, vol. 90, Berlin-Heidelberg, Springer, New York,
1969.
[12] Bhatia N.P.; Lazer, A. C.; Szegö G.P., On global weak attractors in
dynamical systems, J. Math. Anal. Appl., 16 (1966), 544–552.
[13] Bhatia N.P. and Szegö G.P., Dynamical systems: stability theory and
applications, Lecture Notes in Mathematics 35. Springer-Verlag, 1967.
[14] Bhatia, N. P. and Szegö, G. P. Stability Theory of Dynamical Systems,

Springer-Verlag, Berlin, 1970.
[15] Birkhoff, G. D., Dynamical Systems, Amer. Math. Soc. Colloquium Pub-
lications, vol. 9, New York, 1927.
[16] Boothby, W. M., An Introduction to Differential Manifolds and Rie-

mannian Geometry, Academic Press, New York, 1975.
[17] Boothby, W. M., A transitivity problem from control theory, J. Diff.

Equations, 17 (1975), 296–307.
[18] Boothby, W. M. and Wilson, E. N., Determination of the transitivity of

bilinear systems, SIAM J. Control Optim., 17 (1979), 212–221.
[19] Braga Barros, C. J.; Gonçalves F., J. R.; do Rocio, O. G.; San Martin, L.
A. B., Controllability of two-dimensional bilinear systems, Proyecciones,
15 (1996), 111–139.
[20] Braga Barros, C. J. and San Martin, L. A. B., Chain control sets for
semigroup actions, Comp. Appl. Math., 15 (1996), 257–276.
[21] Braga Barros, C. J. and San Martin, L. A. B., Chain transitive sets for
flows on flag bundles, Forum Math., 19 (2007), 19–60.
[22] Braga Barros, C. J. and Souza, J. A., Attractors and chain recurrence
for semigroup actions, J. Dyn. Diff. Eq., 22 (2010), 723–740.
[23] Braga Barros, C. J. and Souza, J. A., Finest Morse decompositions

for semigroup actions on fiber bundles, J. Dyn. Diff. Eq., 22 (2010),
741–750.
[24] Braga Barros, C. J. and Souza, J. A., On the number of maximal chain
transitive sets in fiber bundles, Forum Math., 25 (2013), 363–381.
BIBLIOGRAPHY 321
[25] Braga Barros, C. J.; Souza, J. A.; Reis, R. A., Dynamic Morse Decom-
positions for Semigroups of Homeomorphisms and Control Systems, J.
Dyn. Cont. Systems, 18 (2012), 1–19.
[26] Braga Barros, C. J.; Rocha, V. H. L.; Souza, J. A., Lyapunov stability
for semigroup actions, Semigroup Forum, 88 (2014), 227–249.
[27] Braga Barros, C. J.; Rocha, V. H. L.; Souza, J. A., On attractors and
stability for semigroup actions and control systems, Math. Nachr., 289
(2016), 272–287.
[28] Cheban, D., Compact global attractors of control systems, J. Dyn. Cont.
Systems, 16 (2010), 23–44.
[29] Cheban, D.; Kloeden, P. E.; Schmalfu, B., The relationship between
pullback, forwards and global attractors of nonautonomous dynamical
systems, Nonlinear Dyn. Syst. Theory, 2 (2002), 125–144.
[30] Clarke, F. H.; Ledyaev, Y. S.; Stern, R. J., Asymptotic stability and
smooth Lyapunov functions, J. Diff. Equations, 149 (1998), 69–114.
[31] Clifford, A. H. and Preston, G. B., The Algebraic Theory of Semigroups

I, American Mathematical Society, Providence, 1967.
[32] Coddington, E. A. and Levinson, N., Theory of Ordinary Differential

Equations, McGraw-Hill, New York, 1955.
[33] Cohen, D., Measure Theory, Birkhäuser Verlag, 1980.
[34] Colonious, F. and Klieman, W., The dynamics of control, Birkhäuser,

Boston, 2000.
[35] Colonious, F. and Klieman, W., Limits of input-to-state stability, Sys-

tems & Control Letters, 49 (2003), 111–120.
[36] Conley, C., Isolated Invariant Sets and the Morse Index, CBMS Re-
gional Conf. Ser. in Math. 38, American Mathematical Society, 1978.
[37] Conley, C., The gradient structure of a flow: I, Ergodic Theory Dynam.
Systems, 8 (1988), 11–26.
[38] Conley, C. and Smoller. J., Topological methods in the theory of shock
waves, Partial differential Equations, Proceedings of Symposia in Pure
Mathematics XXIII. AMS, Providence (1973), 293–302.
322 BIBLIOGRAPHY
[39] Conley, C. and Smoller. J., On the structure of magnetohydrodynamic

shock waves, Comm. Pure Appl. Math., 27 (1974), 367–375.
[40] Conley, C. and Smoller. J., The MHD version of a theorem of H. Weyl,
Proc. Amer. Math. Soc., 42 (1974), 248–250.
[41] Conley, C. and Smoller. J., On the structure of magnetohydrodynamic

shock waves II, Math. Pures et Appl., 54 (1975), 429–444.
[42] Conley, C. and Zehnder, E., A global fixed point theorem for symplec-
tic maps and subharmonic solutions of Hamiltonian equations on tori,
Nonlinear Functional analysis and its Applications, ed. F. Browder. Pro-
ceedings of Symposia in Pure Mathematics, 45, Part I. AMS, Providence
(1986), 283–299.
[43] Crouch, P. E. and Byrnes, C., Local accessibility, local reachability, and
representations of compact groups, Math. Systems Theory, 19 (1986),
43–65.
[44] Dunford, N. and Schwartz, J. T., Linear Operators, Part I: General

Theory, John Wiley & Sons, 1988.
[45] El Assoudi, R. and Gauthier, J. P., Controllability of right invariant

systems on real simple Lie groups of type F4 , G2 , Cn and Bn , Math. of
Control, Signals, and Systems, 1 (1988), 292–391.
[46] Ellis, R., Universal minimal sets, Proc. Amer. Math. Soc., 11 (1960),
540–543.
[47] Ellis, R., Lectures on Topological Dynamics, W.A. Benjamin Inc., New
York, 1969.
[48] Ellis, D. B.; Ellis, R.; Nerurkar, M., The topological dynamics of semi-
group actions, Trans. Amer. Math. Soc., 353 (2000), 1279–1320.
[49] Flies, M.; Levine, J.; Martin, P.; Rouchon, P., Flatness and defect of
non-linear-systems: introductory theory and examples, Int. J. Control,
61 (1995), 1327–1361.
[50] Folland, G. B., Real Analysis. John Wiley & Sons, New York, 1984.
[51] Francis, B., A Course in H∞ Control Theory, vol. 88 of LN Control

and Information Sciences, Springer-Verlag, 1987.
BIBLIOGRAPHY 323
[52] Franks, J. and Misiurewicz, M., Topological methods in dynamics.

Chapter 7 in Handbook of Dynamical Systems, vol 1, part 1, 547–598,
Elsevier, 2002.
[53] Gauthier, P.; Kupka, I.; Sallet, G., Controllability of right invariant
systems on real simple Lie groups, Systems & Control Letters, 5 (1984),
187–190.
[54] Gayek, J. E. and Vincent, T. L., On the intersection of controllable

and reachable sets, Journal of Optimization Theory and Applications,
50 (1986), 267–278.
[55] Gottschalk, W. H., A survey of minimal sets, Ann. Inst. Fourier (Greno-
ble), 14 (1964), 53–60.
[56] Hale, J. K., Asymptotic behavior of dissipative systems, Mathematical

Surveys and Monographs Number 25. American Mathematical Society,
Providence, RI., 1988.
[57] Hermes, H., Local controllability and sufficient conditions in singular

problems, J. Diff. Equations, 20 (1976), 213–232.
[58] Hermes, H., On local controllability, SIAM J. Control and Optim., 20

(1982), 211–220.
[59] Hilgert, J. and Neeb, K.-H., Lie semigroups and their applications, Lec-
ture Notes in Mathematics, vol. 1552, Springer-Verlag, 1993.
[60] Hirsch, M. W. and Smale, S., Differential Equations, Dynamical Sys-

tems, and Linear Algebra, Academic Press, New York, 1974.
[61] Hirsch, W. M.; Smith, H. L.; Zhao, X., Chain transitivity, attractivity
and strong repellers for semidynamical systems, J. Dyn. Diff. Eq., 13
(2001), 107–131.
[62] Jost, J., Dynamical systems. Examples of complex behaviour, Universi-

text. Springer-Verlag, Berlin, 2005.
[63] Jurdjevic, V., Geometric Control Theory, Cambridge University Press,

1997.
[64] Jurdjevic, V. and Kupka, I., Control systems subordinated to a group

action: accessibility, J. Diff. Equations, 39 (1981), 151–179.
324 BIBLIOGRAPHY
[65] Jurdjevic, V. and Kupka, I., Control systems on semisimple Lie groups
and their homogeneous spaces, Ann. Inst. Fourier (Grenoble), 31 (1981),
186–211.
[66] Jurdjevic, V. and Sussmann, H. J., Control systems on Lie groups, J.

of Diff. Eq., 12 (1972), 313–329.
[67] Kang, W., Bifurcation and normal form of nonlinear control systems I,
SIAM J. Control Optim., 36 (1998), 193–212.
[68] Kliemann, W., Recurrence and invariant measures for degenerate diffu-
sions, Annals of Probability, 15 (1987), 690–707.
[69] Krener, A. J., A generalization of Chow’s theorem and the bang-bang

theorem to nonlinear control systems, SIAM J. Control and Optim., 12
(1974), 43–52.
[70] Ladyzhenskaya, O. A., Attractors for semigroups and evolution equa-

tions, Cambridge University Press, Cambridge, 1991.
[71] Lawson, J. D., Maximal subsemigroups of Lie groups that are total,
Edinb. Math. Soc., 30 (1987), 479–501.
[72] Lee, L. and Markus, L., Foundations of optimal control theory, Wiley,
1967.
[73] Lobry, C., Controllability of nonlinear systems onn compact manifolds,

SIAM J. Control and Optim., 12 (1974), 1–4.
[74] Mañé, R., Ergodic Theory and Differentiable Dynamics, Springer-

Verlag, 1987.
[75] Mayhew, C.G. and Teel, A.R., On the topological structure of attraction
basins for differential inclusions, Systems & Control Letters, 60 (2011),
1045–1050.
[76] McGehee, R., Charles C. Conley, 1933–1984, Ergod. Th. & Dynam. Sys.,
8* (1988), 1–7.
[77] Mischaikow, K. and Mrozek, M., Conley index, Chapter 9 in Handbook

of Dynamical Systems, vol 2, 393–460, Elsevier, 2002.
[78] Moeckel, R., Orbits near triple collision in the three-body problem, PH.D.
thesis, 1980.
BIBLIOGRAPHY 325
[79] Patrão, M., Morse decomposition of semiflows on topological spaces, J.

of Dyn. Diff. Eq., 19 (2007), 181–198.
[80] Patrão, M. and San Martin, L. A. B., Semiflows on topological spaces:

chain transitivity and semigroups, J. of Dyn. Diff. Eq., 19 (2007), 155–
180.
[81] Patrão, M. and San Martin, L. A. B., Morse decomposition of semiflows

on fiber bundles, Discrete and Continuous Dynamical Systems (Series
A), 17 (2007), 113–139.
[82] Rasmussen, M., All-time Morse decompositions of linear nonau-

tonomous dynamical systems, Proc. Amer. Math. Soc., 136 (2008),
1045–1055.
[83] Raminelli, S.A. and Souza, J. A., Global attractors for semigroup ac-
tions, J. Math. Anal. Appl., 407 (2013), 316–327.
[84] Reed, M. and Simon, B., Functional Analysis, Academic Press, Boston,
1980.
[85] Roxin, E. O., Reachable sets, limit sets and holding sets in control sys-
tems, Nonlinear Analysis and Applications, V, V. Lakshmikantham ed.,
Marcel Dekker Inc. Publ., New York, 1987.
[86] Roxin, E. O., Application of Holding Sets to Optimal Control, Proceed-

ings of The International Conference on Theory and Applications of
Differential Equations, Ohio University, 1988.
[87] San Martin, L. A. B., Invariant control sets on flag manifolds, Math. of
Control, Signals, and Systems, 6 (1993), 41–61.
[88] San Martin, L. A. B., Distribuições singulares e órbitas de famı́lias de

campo de vetores, Universidade Estadual de Campinas, Brazil, 1996.
[89] San Martin, L. A. B., Homogeneous spaces admitting transitive semi-

groups, J. Lie Theory, 8 (1998), 111–128.
[90] San Martin, L. A. B. and Tonelli, P. A., Semigroup actions on homoge-

neous spaces, Semigroup Forum, 50 (1995), 59–88.
[91] Silva Leite, F. and Crouch, P. E., Controllability on classical Lie groups,
Math. of Control, Signals, and Systems, 1 (1988), 31–42.
326 BIBLIOGRAPHY
[92] Sontag, E. D., Mathematical Control Theory, Springer-Verlag, 1990.
[93] Souza, J. A., A Course on Geometric Control Theory: Transitivity and

Minimal Sets, Deutschland: LAP Lambert Academic Publishing, 2016.
[94] Souza, J. A., Complete Lyapunov functions of control systems, Systems

& Control Letters, 61 (2012), 322–326.
[95] Souza, J. A., On limit behavior of semigroup actions on noncompact

spaces, Proc. Amer. Math. Soc., 140 (2012), 3959–3972.
[96] Souza, J. A., On Morse decomposition of control systems, Int. J. Con-

trol, 85 (2012), 815–821.
[97] Souza, J. A., Characterization of uniform attractors for control sytems,

Int. J. Control, 88 (2015), 2403–2411.
[98] Souza, J. A., Chain transitivity for semigroup actions on flag bundles,
Annali di Matematica Pura ed Applicata, 193 (2014), 817–836.
[99] Souza, J. A., Recurrence theorem for semigroup actions, Semigroup Fo-
rum, 83 (2011), 351–370.
[100] Souza, J. A., Semigroup actions under compactifications: a view from

control systems, Semigroup Forum, 92 (2015), 494–510.
[101] Souza, J. A. and Tozatti, H. V. M., Prolongational limit sets of control

systems, J. Diff. Equations, 254 (2013), 2183–2195.
[102] Souza, J. A. and Tozatti, H. V. M., Some aspects of stability for semi-
group actions and control systems, J. Dyn. Diff. Eq., 26 (2014), 631–654.
[103] Souza, J. A.; Rocha, V. H. L.; Tozatti, H. V. M., On stability and

controllability for semigroup actions, Topological Methods in Nonlinear
Analysis, 48 (2016), 1–29.
[104] Stefan, P., Accessible sets, orbits and foliations with singularities, Proc.
London Math. Soc., 29 (1974), 699–713.
[105] Stefan, P., Integrability of systems of vector fields, J. London Math.

Soc., 21 (1980), 544–556.
[106] Sussmann, H. J., Orbits of families of vector fields and integrability of

distribuitions, Trans. Amer. Math. Soc., 180 (1973), 171–188.
BIBLIOGRAPHY 327
[107] Sussmann, H. J., Some properties of vector field systems that are not
altered by small perturbations, J. Diff. Equations, 20 (1976), 292–315.
[108] Sussmann, H. J., Lie brackets and local controllability, SIAM J. Control
and Optim., 21 (1983), 687–713.
[109] Sussmann, H. J., On a general theorem on local controllability, SIAM

J. Control and Optim., 25 (1987), 158–194.
[110] Teel, A. R. and Praly, L., A smooth Lyapunov function from a class-KL
estimate involving two positive semidefinite functions, ESAIM: Control,
Optim. Calculus Variations, 5 (2000), 313–367.
[111] Varadarajan, V. S., Lie groups, Lie algebras and their representations,
Prentice-Hall, 1974.
[112] Warner, F., Foundations of Differentiable Manifolds and Lie Groups,

Scott Foresman and Company, Glenview, Illinois, 1971.
[113] Willems, J., Models for dynamics, Dynamics Reported, 2 (1989), 171–
269.
[114] Zenhxin, L., The random case of Conley’s theorem, Nonlinearity, 19

(2006), 277–291.
[115] Zenhxin, L., The random case of Conley’s theorem III. Random semiflow
case and Morse decomposition, Nonlinearity, 20 (2007), 2773–2791.
Index
accessibility, 126 for flows, 293

accessibility rank condition
Kalman condition, 168 Birkhoff center, 113
accessibility rank condition, 167 bounded dissipative control system, 201
accessible control system, 126 broken integral curve, 14
admissible family of functions, 12
chain, 215
all-time Morse decomposition, 311
for flows, 300
asymptotic transitive preorder, 116
chain component, 242
asymptotic transitive set, 114
control system vs. control flow,
maximal, 114
264
positive minimal set, 114
Chain control set
vs. chain transitive set, 225 shadowing semigroups, 249
vs. weak prolongational control chain control set, 229
set, 166 for general semigroup actions, 249
asymptotic transitivity, 103 vs. maximal chain transitive set,
asymptotic transitivity relation, 113 230
asymptotically compact control sys- chain controllable control system, 217
tem, 201 chain controllable system
vs. limit compact control sys- not controllable, 246
tem, 204 chain recurrence set, 217
asymptotically stable set, 178 vs. attractor-repeller pair, 243
attraction chain recurrent control system, 217
region of, 170 chain recurrent point, 217
attractor, 170 chain recurrent set, 217
attractor-repeller Morse decomposition, chain transitive set, 217
257 vs. chain recurrent set, 219
not dynamic Morse decomposition, vs. recurrent set, 226
260 chain transitivity preorder, 224
attractor-repeller pair, 236 chain transitivity relation, 221
329
330 INDEX
cocycle property, 22 bilinear control system, 14

compactification, 277 control affine system, 15
complete Lyapunov function, 269 linear control system, 14
completely controllable control system, with obstacles, 39
138 controllable system, 122
concatenation of functions, 12 critical point, 20, 55
Conley attractor, 192
complementary repeller of, 236 differentiable manifold, 313
lift of, 240 curve, 315
for flows, 291 differentiable map, 314
lift of, 237 derivative, 316
Lyapunov function for, 194 local coordinate system, 314
projection of, 238 open submanifold, 314
product of, 314
strong attractor, 212
Riemannian metric, 317
union of, 192
tangent bundle of, 316
vs. asymptotically stable set, 193
vector field, 316
vs. isolated invariant set, 192
differentiable menifold
vs. uniform attractor, 193, 194
Riemannian distance, 317
vs. uniformly stable set, 195
disconnected weak control set, 79
Conley main theorem, 243
domain of asymptotic attraction, 115
for flows, 309
domain of attraction, 108
Conley repeller, 192
domain of chain attraction, 224
complementary attractor of, 240
domain of weak attraction, 76, 153
for flows, 291
dynamic Morse decomposition, 252
lift and projection of, 239 lift of, 254
vs. isolated invariant set, 192 vs. attractor-repeller Morse de-
control affine system composition, 258–260
vs. nonautonomous dynamical sys- dynamical system, 280
tem, 212
control flow, 35 effective control set, 129
control range, 13 vs. holding set, 129
control set, 122 equicontinuous family, 29
for general semigroup action, 156 equilibrium point, 20, 55
vs. invariant control set, 134 equistable set, 94
vs. weak control set, 127 vs. prolongation, 95
with nonvoid interior, 129 eventually bounded control system, 200
control system, 13 eventually compact control system, 200
INDEX 331
finest Morse decomposition, 255, 256 non-regressing property of trajec-

in projective bundle, 278 tories, 107
vs. chain component, 261 vs. control set, 130
first negative prolongational limit set, with interior points, 106
88 homogeneous space, 80
projection of, 94 hyperspace, 66, 243
vs. closure of negative semi-orbit,
ideal, 45
89
internally chain transitive set, 243
vs. negative limit set, 90, 92
invariant control set, 133
first positive prolongational limit set,
not invariant, 136
88
under local accessibility, 133
vs. closure of positive semi-orbit,
vs. stochastic differential equa-
89
tion, 167
vs. positive limit set, 90, 92
invariant control system, 167
fist positive prolongational limit set
invariant Morse decomposition, 261
projection of, 94
invariant set, 41
flow, 280 isolated invariant set, 42
fundamental period, 62 isolating neighborhood, 42
Fundamental Theorem of
Differential Equations, 280 Kalman condition, 137
complete controllability, 138
generalized dynamic Morse decompo-
Lie group, 315
sition, 274
lift of a set, 35
global attractor, 198
limit compact control system, 201
existence of, 204, 205
linear control system
vs. Conley attractor, 206
under Kalman condition, 137
vs. global uniform attractor, 207 linear Lie group, 38
vs. uniform attractor, 207 linear oscillator, 247
global uniform attractor, 207 linearized control system, 277
gradient-like dynamical system, 1, 249 local controllability, 166
strongly, 276 local group, 39
group, 167 local semigroup, 39
local vector field, 39
Hamiltonian vector field, 37 locally accessible control system, 126
Hausdorff distance, 66, 243 locally integrable function, 13
higher prolongations, 101 locally normally accessible control sys-
holding set, 105 tem, 164
332 INDEX
Lyapunov function, 184 positively Lagrange stable, 166

for attractor-repeller pair, 266
negative chain limit set, 216
main control set, 129 projection of, 227
of linear control system, 142 vs. complementary repeller, 241
maneuverable set, 166 vs. Conley repeller, 231
matrix Lie group in compact state space, 233
control affine system on, 121 negative limit set, 82
Haar measure, 119 for flows, 281
maximal chain transitive set, 222 not positively invariant, 83
being Conley attractor, 232 projection of, 86
being Conley repeller, 232 under negative Lagrange stabil-
invariance of, 241 ity, 166
is closed, 232 vs. chain transitivity, 227
not invariant, 246 vs. closure of negative semi-orbit,
vs. chain recurrent component, 83
242 vs. continuum set, 85
vs. internally chain transitive set, vs. negative chain limit set, 225
244 negative minimal set, 70
maximal transitive set, 105 existence, 72
measure of noncompactness, 197 uniqueness, 71
minimal control system, 73 negative semi-orbit, 36, 44
minimal equistable set, 95 negatively invariant set, 41
vs. prolongation, 96 negatively Poisson stable point, 157
vs. prolongational transitivity pre- negatively recurrent point, 109
order, 154 not positively recurrent, 158
minimal set, 70 periodic point, 111
existence, 72 vs. negative limit set, 110
Lagrange stability, 80 vs. negative semi-orbit, 110
uniqueness, 71 negatively recursive set, 157
with nonvoid interior, 74 nonwandering point, 146
Morse decomposition, 261 chain recurrent point, 226
for flows, 285 in limit sets, 146
vs. chain recurrence set, 261 Poincaré recurrence, 149
Morse-Lyapunov function, 265 vs. negatively recurrent point,
motion map, 20 148
Lagrange stable, 166 vs. positively recurrent point, 148
negatively Lagrange stable, 166 normally accessible control system, 164
INDEX 333
optimal control, 166 vs. asymptotic transitivity pre-

orbit, 44 order, 116
Orbit Theorem, 53 vs. control set, 135
vs. invariant control set, 133
periodic point, 60 vs. weak control set, 77, 78
periodic trajectory of a control sys- vs. weak preorder, 78
tem, 20 positive semi-orbit, 36, 44
phase map, 19 positively invariant set, 41
piecewise constant function, 13 positively Poisson stable point, 157
Poicaré recurrence, 109 positively recurrent point, 109
Poincaré recurrence not negatively recurrent, 158
vs. Poisson stability, 157 vs. periodic point, 110
vs. positive limit set, 109
Poincaré recurrence theorem, 120
vs. positive semi-orbit, 109
point dissipative control system, 201
positively recursive set, 157
Poisson stable point, 157
preorder among holding sets, 108
positive chain limit set, 216
preorder among weak control sets, 77
for flows, 306
projected system
projection of, 227
on the projective space, 247
vs. Conley attractor, 231
on the unit sphere, 196
in compact state space, 233
projective space, 315
positive limit set, 81
prolongational control set, 168
for flows, 281
prolongational transitive set, 150
not negatively invariant, 83
prolongational transitivity preorder, 154
projection of, 86
prolongational transitivity relation, 150
under asymptotic compactness, 202 prolongations, 87
under positive Lagrange stability, vs. continuum sets, 93
166
vs. chain transitivity, 227 random dynamical system, 311
vs. closure of positive semi-orbit, rank of a control system, 163
83 recurrent point, 109
vs. continuum set, 85 region of weak attraction
vs. positive chain limit set, 225 vs. domain of attraction, 174
positive minimal set, 70 region of weak uniform attraction
existence, 72 domain of weak attraction, 174
uniqueness, 71 relative negative limit set, 112
vs. asymptotic transitive preorder, relative positive limit set, 112
116 repellency
334 INDEX
region of, 175 uniform repeller, 176

repeller, 176 uniformly equicontinuous family, 29
repeller-attractor pair, 240 uniformly stable set, 178
universal minimal set, 80
Selgrade theorem, 277
set of central motions, 113 weak attraction
set of relative positively recurrent points, region of, 170
121 weak attractor, 170
shift of a function, 12 weak control set, 75
shift space, 16 non-regressing property of trajec-
smooth Lyapunov function tories, 75
uniform attractor of, 189 vs. weak prolongational control
strongly accessible control system, 138 set, 166
system group, 36 weak prolongation control set, 151
system semigroup, 36 weak prolongational control set, 151
Green’s preorder, 265 weak prolongation control set, 153
weak repellency
t-projection, 19 region of, 175
time reversed control system, 20 weak repeller, 176
time reversed phase map, 20 weak transitive set, 75
torus, 315 weak transitivity relation, 74
total semigroup, 38 weak uniform attraction
trajectory of a control system, 20 region, 170
transition map, 20 weak uniform attractor, 170
transitivity relation, 104 weak uniform repeller, 176
transitivity set , 129 weak* topology, 15
translation hypothesis, 100
uniform attraction
region of, 170
uniform attractor, 170
compact and positively invariant,
178
Lyapunov function for, 185
vs. asymptotically stable set, 178
vs. isolated invariant set, 178
uniform repellency
region of, 175

The Dynamic Morse Theory of Control Systems - Souza CAMBRIGE 2020

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Dynamic Morse Theory of Control Systems - Souza CAMBRIGE 2020

Uploaded by

Copyright:

Available Formats

The Dynamic Morse

This book first published 2020

Cambridge Scholars Publishing

Lady Stephenson Library, Newcastle upon Tyne, NE6 2PA, UK

British Library Cataloguing in Publication Data

Copyright © 2020 by Josiney A. Souza

ISBN (10): 1-5275-4507-5

1 Fundamental Theory of Control systems 11

3 Minimal sets and transitivity 69

4 Limit sets and prolongations 81

5 Asymptotic transitivity 103

6 Attractors and repellers 169

7 Chain transitivity 215

8 Morse decompositions 251

8.5 Final comments and open questions . . . . . . . . . . . . . . . 271

A Dynamical systems 279

B Differentiable manifolds 313

The global dynamical behavior of a compact system is described by Morse

havior on the boundary. Other applications in the study of dynamics include

is an infinite dimensional dynamical system that lives on the product space of

esis. The strategy is based on invariance issues of attractor-repeller pairs.

totic transitivity is presented in Chapter 5 as a relation among the Poincaré

• A first course in control system theory: Chapters 1 and 2, for students

• Controllability and asymptotic transitivity: Chapters 3, 4, and 5.

• Morse decompositions and chain transitivity: Chapters 6, 7, and 8.

• Seminar on periodicity: Sections 2.2, 2.3, 2.4, and 5.1.

• Seminar on Poincaré recurrence: Sections 4.1, 5.2, and 5.3.

• Seminar on attractors: Chapter 6.

• Introduction to Morse theory of dynamical systems: Appendix A.

In the language of mathematics, the concept of a control system was formu-

1.1 Basic definitions

B (x, ) = {y ∈ M : d (x, y) < } ,

Definition 1.1.1 For a given real number s ∈ R and two functions u, v :

Definition 1.1.2 Let U ⊂ Rn . A family of functions U = {u : R → U} is

1. Each function u ∈ U is locally integrable, that is, u is Lebesgue integrable

2. For each u ∈ U and s ∈ R, the s-shift u · s is contained in U.

3. For u, v ∈ U and s ∈ R, the s-concatenation of u and v is contained in

A trivial admissible family is given by one constant function u (t) ≡ u0 ∈

Definition 1.1.3 A control system Σ = (M, U, U, X) is formed by

1. A state space M that is a d-dimensional C ∞ -differentiable manifold.

2. A control range U ⊂ Rn and an admissible family of control functions

3. A family of ordinary differential equations

depending on the control functions, where X : M × Rn → T M is a C ∞

In the case M = Rd , one may consider a C ∞ map X : Rd × Rn → Rd in

state variable, since the dependence of u on t need not be differentiable.

and proceed with Carathéodory’s theory allowing measurable dependence on

Example 1.1.1 A linear control system takes the form

where A is a d×d real matrix, B is a d×n real matrix, M = Rd , and U = Rn .

Example 1.1.2 A bilinear control system takes the form

Example 1.1.3 A control affine system on a C ∞ -manifold M takes the

1.2 Shift Space

Define the linear functional fα : L∞ (R, Rn ) → R by

The weak* topology on L∞ (R, Rn ) is the weak topology induced on

Lemma 1.2.1 The shift is an internal operation in U.

This enables us to define the shift map Θ : R × U → U by Θ (t, u) = u · t.

Lemma 1.2.2 The subspace U ⊂ L∞ (R, Rn ) given by

U = {u : R → Rn : u (t) ∈ U for a.a. t ∈ R, measurable}

is a compact metrizable space.

Proof. It should be noticed that L1 (R, Rn ) is a separable Banach space,

for every pair u, v ∈ L∞ (R, Rn ). By Alaoglu’s theorem, the unit ball of

U|I = { u|I : u ∈ U} ⊂ L2 (I, Rn )

is a subsequence (unk ) such that unk (t) → u (t) for a.a. t ∈ I. As U

We are now able to prove the main result of this section.

Proof. Since U ⊂ U is weak* closed, it follows that U is a compact

B (x, ) = {y ∈ M : d (x, y) < } ,

By deﬁning K = supv∈U v, we have

ϕ (t, x, u) − ϕ (t, y, u)

ϕ (t, x, u) − ϕ (t, y, u) ≤ etL x − y

all t ∈ [0, T ], u ∈ U, and x, y ∈ K with x − y < δ. Finally, for a given