Professional Documents
Culture Documents
The Dynamic Morse Theory of Control Systems - Souza CAMBRIGE 2020
The Dynamic Morse Theory of Control Systems - Souza CAMBRIGE 2020
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
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.
Preface xiii
Introduction 1
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
2.6 Notes and references . . . . . . . . . . . . . . . . . . . . . . . . 67
ix
x Contents
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
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
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
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
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
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:
who are familiar with ordinary differential equations and basic notions
of real analysis and general topology.
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
11
12 Chapter 1 · Fundamental Theory of Control systems
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.
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
ẋ = X (x, u (t))
ẋ = A (x) + B (u(t))
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
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 ).
Theorem 1.2.1 The shift space U is a compact metrizable space and the shift
map Θ : R × U → U is continuous.
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 .
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
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
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))
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).
ẋ = u1 X1 (x) + u2 X2 (x)
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).
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.
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.
ϕ (τ , ϕ (s, x, u) , v) = ϕ (τ + s, x, w)
for every x ∈ M .
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).
for every x ∈ M .
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
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))
26 Chapter 1 · Fundamental Theory of Control systems
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
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:
−1
a diffeomorphism of M and (ϕus ) = ϕu·s
−s .
⎧
⎪ u1 , for 0 ≤ t ≤ t1
⎪
⎪
⎪
⎪ u , for t1 < t ≤ t1 + t2
⎨ 2
u (t) = .. .
⎪ .
⎪
⎪
⎪
⎪
N−1
⎩ uN , for ti < t ≤ s
i=0
Section 1.3 · Control affine systems 27
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 .
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
28 Chapter 1 · Fundamental Theory of Control systems
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
Section 1.3 · Control affine systems 29
√
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).
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
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
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:
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.
we have
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
it follows that
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)) .
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.
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 .
S (x) = {ϕ (t, x, u) : t ≥ 0, u ∈ U} ,
−1
S (x) = {ϕ (t, x, u) : t ≤ 0, u ∈ U} ,
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.
S (x) = {y ∈ Rn : yi ≥ xi } ,
S −1 (x) = {y ∈ Rn : yi ≤ xi } ,
G (x) = Rn .
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
38 Chapter 1 · Fundamental Theory of Control systems
given by
−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]).
(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)
Section 1.6 · Problems 39
G = {ϕ (t, 1, u) : t ∈ R, u ∈ U} ⊂ Gl (d, R)
Elementary Concepts
41
42 Chapter 2 · Elementary Concepts
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.
Thus both int (X) and fr (X) are invariant whenever X is invariant. We
have a partial converse for this result.
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) .
O+ (x) = {ϕ (t, x, u) : t ≥ 0, u ∈ U} ,
−
O (x) = {ϕ (t, x, u) : t ≤ 0, u ∈ U} .
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
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) .
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 2.4: On the left, the positive semi-orbit of a point x; on the right, the
negative semi-orbit of 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) .
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.
+ +
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.
+
Section 2.2 · Orbits and semi-orbits 49
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) .
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
−
and
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
+
Proposition 2.2.4 The orbit G (x) is pathwise connected and its closure
cl (G (x)) is connected for all x ∈ M .
52 Chapter 2 · Elementary Concepts
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) .
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).
In the same way, we can show that closure and interior of semi-orbits also
have the transitivity property.
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.
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 .
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 .
54 Chapter 2 · Elementary Concepts
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
d
0= ϕ (t, x, u)|t=0 = X (x, u (0)) .
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.
⎧
⎪
⎪ 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
⎧
⎪
⎪
..
⎪
⎪ .
⎪
⎪
⎪
⎪ 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 +
⎪
⎪ .
⎩ ..
Section 2.3 · Critical points 57
⎧
⎪
⎪
..
⎪
⎪ .
⎪
⎪
⎪
⎪ 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)
|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)
58 Chapter 2 · Elementary Concepts
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) .
1. x is a critical point.
5. There is u ∈ U such that ϕux ([a, b]) = {x} for some a < b.
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
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
Section 2.3 · Critical points 59
(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
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}.
1. x is critical.
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
60 Chapter 2 · Elementary Concepts
ϕ (t + kτ , x, u) = ϕ (t + (k − 1) τ + τ , x, u) = ϕ (t + (k − 1) τ , x, u) .
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.
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) .
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.
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)
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.
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
64 Chapter 2 · Elementary Concepts
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
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.
(a) X is invariant.
(b) ϕut (X) = X for all u ∈ U and t ∈ R.
(c) φ (X) = X for all φ ∈ G.
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
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
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.
72 Chapter 3 · Minimal sets and transitivity
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.
Figure 3.1: From left to right, trajectories with controls u2 = 1/2, 1/2 < u2 <
1, and u2 = 1.
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.
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.
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
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.
+ +
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.
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.
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 .
Section 3.4 · Problems 79
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.
4. Show that the control systems in Examples 1.3.2 and 1.3.1 are minimal.
81
82 Chapter 4 · Limit sets and prolongations
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
In certain senses, the limit set “closes” the semi-orbit, as in the following
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} ,
Note that the negative limit set ω − (x) is not positively invariant and the
positive limit set ω + (x) is not negatively invariant.
84 Chapter 4 · Limit sets and prolongations
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 4.1.5 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.
$ %
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.
In the case of compact state space, every limit set is a continuum set. This
result can be easily extended in the following way:
The following shows how limit set map relates to transition map:
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
86 Chapter 4 · Limit sets and prolongations
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
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
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.
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:
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
+
There may be a set of states that does not have prolongation. The follow-
ing describes this situation:
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
90 Chapter 4 · Limit sets and prolongations
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.
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 ).
ϕ (−tn , ϕ (tn , xn , un ) , un · tn ) = xn → x
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
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
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.
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)
Section 4.2 · Prolongational limit sets 93
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.
The following result describes the prolongational limit sets of the control
system by means of the prolongational limit sets of the control flow.
94 Chapter 4 · Limit sets and prolongations
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
Consequently, we have
π J + π −1 (X) = J + (X) ,
π J − π −1 (X) = J − (X) ,
1. E is positively invariant.
Section 4.3 · Minimal equistable sets 95
In the case of compact sets, the equistable sets are characterized as the
sets which do not admit prolongation, as in the following:
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.
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
Figure 4.1: A minimal equistable set that is not a positive minimal set.
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.
Section 4.4 · Problems 99
4.4 Problems
1. Check that
$ %
ω + (X) = cl O≥n
+
(X) ,
n∈N
$ %
−
ω − (X) = cl O≥n (X) ,
n∈N
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.
(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.
ru−sv
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.)
Section 4.5 · Notes and references 101
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:
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.
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
+ +
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
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.
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.
106 Chapter 5 · Asymptotic transitivity
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 ].
z = ϕ(T − t + t, x, u)
= ϕ (T − t, ϕ(t, x, u), u · t))
= ϕ (T − t, y, u · 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.
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
108 Chapter 5 · Asymptotic transitivity
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.
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
+
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,
+
1. x is negatively recurrent.
4. O− (x) ⊂ ω − (x).
1. O+ (x) = ω + (x).
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
We have an analogous result for negative limit sets and negatively recur-
rent points, as shown in the following:
1. O− (x) = ω − (x).
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
112 Chapter 5 · Asymptotic transitivity
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).
+
Section 5.2 · Asymptotic transitive sets 113
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:
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.
This relation among the positively recurrent points is called the asymp-
totic transitivity relation.
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
114 Chapter 5 · Asymptotic transitivity
+
∅ = 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.
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.
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.
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.
116 Chapter 5 · Asymptotic transitivity
Proposition 5.2.5 The relation “a ” among the asymptotic transitive sets
is a 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.
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.
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 )
The following result is the Poincaré recurrence theorem for control systems.
120 Chapter 5 · Asymptotic transitivity
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
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).
This property of relative recurrent points joins with the weak transitivity
to define the concept of a control set, as in the following:
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.
Section 5.4 · Control sets 123
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 .
Similarly to the asymptotic transitive sets, the control sets also define a
partition of the state space M , as follows.
+ +
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 .
We now prove that the reunion of the control sets contains the set R .
" 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
Hx ⊂ Rx ⊂ D ⊂ Rx
nonempty for all T > 0. If these properties hold for all x ∈ M the system is
called accessible or locally accessible, respectively.
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).
$ %
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.
∂ ∂
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.
∂
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.
∞
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)).
3. For any pair x, y ∈ D0 , there are t > 0 and u ∈ U such that ϕ (t, x, u) =
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 )
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:
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
Section 5.4 · Control sets 131
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.
132 Chapter 5 · Asymptotic transitivity
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
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 ) .
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
3. int (C) ⊂ O+ (x) for all x ∈ C and int (C) = O+ (x) for x ∈ int (C).
134 Chapter 5 · Asymptotic transitivity
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))
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.
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.
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.
Definition 5.6.1 The linear control system LΣ satisfies the Kalman con-
dition if dim Δ = d, that is, Δ = Rd .
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)
138 Chapter 5 · Asymptotic transitivity
$ %
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
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 .
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
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!
140 Chapter 5 · Asymptotic transitivity
∞
(−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.
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
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.
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
142 Chapter 5 · Asymptotic transitivity
condition. There then exists a unique effective control set, D = cl (O+ (0)) ∩
O− (0). Furthermore, the following statements hold:
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).
−
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 +
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
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.
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
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
146 Chapter 5 · Asymptotic transitivity
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 .
Note that
unk (t) , if t ≤ 0
vk · (−tnk ) (t) = .
unk+1 (t) , if t > 0
Then we have
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 .
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
.
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}.
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 .
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 .
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.
Contrast the condition (1) of a weak 0-prolongation control set with the
property of an equistable set stated in Proposition 4.3.1.
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)).
152 Chapter 5 · Asymptotic transitivity
The converse holds if distinct weak t-prolongation control sets are disjoint.
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:
Proposition 5.7.7 The relation “wu ” among the weak prolongational con-
trol sets is a preoder.
154 Chapter 5 · Asymptotic transitivity
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).
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
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.
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 .
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
∂
(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}.
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
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).
(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.
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
Section 5.8 · Problems 159
(c) Show that the maximal asymptotic transitive set R( π ,0) is not a
2
control set.
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.
∂ ∂ ∂
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 ∈
Section 5.8 · Problems 161
$ %
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.
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 .
Section 5.8 · Problems 163
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.
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 )
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.)
Section 5.8 · Problems 165
Show that:
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
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.
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.
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) .
The regions of attraction of the compact sets have special properties, which
we will now discuss.
x ∈ ϕu·t
−t (B (K, r)) ⊂ ϕ−t (Aw (K)) ⊂ Aw (K) .
u·t
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
Section 6.1 · General concepts of attraction 175
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.
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.
176 Chapter 6 · Attractors and repellers
We now have the following consequence from Lemmas 6.2.1 and 6.2.2.
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.
For compact sets, uniform stability and equistability have the same mean-
ing (see the notes for Chapter 4 and Problem 6.5(13)).
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.
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).
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 .
: −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
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.
Proposition 6.2.6 Assume that all limit sets of points are nonempty. Let
A ⊂ M be a compact uniform attractor. Then
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.
184 Chapter 6 · Attractors and repellers
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.
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.
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
186 Chapter 6 · Attractors and repellers
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
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)).
Section 6.2 · Uniform attractor 187
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
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
188 Chapter 6 · Attractors and repellers
of v and u0 , we have
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
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
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 .
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 6.3.2 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.
$ % $ % $ %
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.
4. A is asymptotically stable.
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.
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.
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.
196 Chapter 6 · Attractors and repellers
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.
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).
There exists at most one global attractor for the control system.
Hence O≥T+
(Y ) ⊂ B (0, ε), and therefore A = {0} is the global attractor of the
control system.
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
200 Chapter 6 · Attractors and repellers
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
Proposition 6.4.3 If the control system is both eventually compact and even-
tually bounded then it is asymptotically compact.
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))
Limit compact control systems seem quite a bit like asymptotically com-
pact control systems, as in the following.
204 Chapter 6 · Attractors and repellers
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.
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 .
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
+
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.
Section 6.4 · Global attractors 207
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
ω + (Y ) ⊂ ω + (cl (Y )) ⊂ J + (cl (Y )) ⊂ A.
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.
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.
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.
210 Chapter 6 · Attractors and repellers
6.5 Problems
1. Let X, Y ⊂ M and prove the following:
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)} .
Section 6.5 · Problems 211
(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
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)
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.
Chain transitivity
215
216 Chapter 7 · Chain transitivity
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 ).
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
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:
1. C is chain recurrent.
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.
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.
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
224 Chapter 7 · Chain transitivity
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
We may show that the closure of a chain transitive set is chain transitive
(see Problem 7.4(3)).
We now define the domain of chain attraction for a chain transitive set
and present a dynamic order among the chain transitivity classes.
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 .
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.
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 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.
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.
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
228 Chapter 7 · Chain transitivity
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
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
Corollary 7.1.3 Each maximal chain transitive set of the control flow projects
into a maximal chain transitive set of the control system.
Proposition 7.2.1 For any , T > 0 and X ⊂ M , the sets Ω+ (X, , T ) and
Ω− (X, , T ) are open.
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.
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) ⊂
+ +
and therefore Ω+ (X) = ω + (Ω+ (X, , T )).
,T >0
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 <
+
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)
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.
$ $ % %
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 ) .
+ + +
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)
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.
Section 7.2 · Attractor-repeller pair 237
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:
L (A) = ω + (L (A)) ⊂ ω + (U × U ) .
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
and then
−tn
fα (w · tn − v · tn ) = w · tn (t) − u · tn (t) , α (t) dt
−∞
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.
3. Every Conley repeller for the control flow in M × U has a repeller neigh-
borhood of the form V × U .
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.
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.
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.
2. C is a connected component of C.
Theorem 7.3.1 suggests that a maximal chain transitive set is usually called
a chain component of the control system.
Section 7.3 · The Conley main theorem 243
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} .
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
244 Chapter 7 · Chain transitivity
#
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
Section 7.3 · The Conley main theorem 245
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.
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).
Section 7.4 · Problems 247
"
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.
$ 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 ?
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.
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.
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).
This result implies a dynamic order between the Morse sets, as in the
following:
Proposition 8.1.2 The relation “” defined above is an order among the
components of a dynamic Morse decomposition M = {C1 , ..., Cn }.
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
254 Chapter 8 · Morse decompositions
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.
#
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
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
256 Chapter 8 · Morse decompositions
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 .
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 .
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
∗
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:
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
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.
A∗a = ∅ ⊂ A∗3 = {3π/2} ⊂ A∗2 = {π/2, 3π/2} ⊂ A∗1 = [π/2, 3π/2] ⊂ A∗0 = M.
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.
ω + (V ∩ Vi ) ⊂ A ∩ Ai ⊂ V ∩ Vi
262 Chapter 8 · Morse decompositions
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
(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.
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.
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.
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
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
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
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
Section 8.4 · Complete Lyapunov functions 269
of v and u0 , we have
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.
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 .
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.
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:
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.
Section 8.6 · Problems 275
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?
276 Chapter 8 · Morse decompositions
10. Find a generalized dynamic Morse decomposition for the control system
in Example 5.5.1.
Section 8.7 · Notes and references 277
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
12. Recall the control systems in Examples 2.2.1 and 7.1.3. Discuss the
generalized dynamic Morse decompositions of these systems.
14. May a chain controllable system have a generalized dynamic Morse de-
composition?
Dynamical systems
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
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.
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)) ,
φ : R × Rn → Rn .
(t, x)
→ t
φ (t, x) = e x
φ (0, x) = x,
φ (t + s, x) = et+s x = et es x = φ (t, φ (s, x)) ,
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).
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).
are respectively called positive limit set and negative limit set of X.
282 Chapter A · Dynamical systems
Note that an invariant set contains all the trajectories of its points.
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.
Exercise 4 Prove that ω + (x) = ω + (φ (t, x)) and ω − (x) = ω − (φ (t, x)) for
all x ∈ M and t ∈ R.
Figure A.2: Infinitely many limit sets of a dynamical system on the plane.
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.
on the plane with a constant a > 0. Find the limit sets of this system.
"
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 .
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 .
Proposition A.2.2 The relation “” is an order among the Morse sets of
a Morse decomposition M = {Ci : i = 1, ..., n}.
The properties of the last two propositions above characterize the Morse
decompositions.
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.
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
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.
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) .
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 + (x) = ω − (x) = E.
Let D ⊂ R2 be the unit disk. If x ∈ ρ−1 R2 \ D then
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
See Fig. A.4. Then M is the intersection of all the Morse decompositions of
the system, and therefore it is the finest one.
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
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
is a repeller.
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.
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.
∅ = A0 ⊂ A1 = E ⊂ A2 = Hp ⊂ A3 = S
∅ = A 0 ⊂ A1 = E ⊂ A 2 = S
∅ = 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}}.
- .
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 .
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).
300 Chapter A · Dynamical systems
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.
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.
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
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
302 Chapter A · Dynamical systems
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
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.
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
Corollary A.4.1 Every maximal chain transitive set is compact and invari-
ant.
Section A.4 · Chain transitive sets 303
"
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.
This means that, for = 0 /2, there is no (, T )-chain from A to B, which
contradicts the hypothesis on X.
#
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
Section A.4 · Chain transitive sets 305
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.
The set
!
+ y ∈ M : for any , T > 0, there is an (, T ) -chain
Ω (X) =
from a point x of X to y
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 )).
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).
We then have
+
+
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
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.
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.
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
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.
Differentiable manifolds
313
314 Chapter B · Differentiable manifolds
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
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
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)
318 Chapter B · Differentiable manifolds
[2] Arnold, L. and Kliemann, W., On unique ergodicity for degenerate dif-
fusions, Stochastics and Stochastics Reperts, 21 (1984), 41–61.
[4] Auslander J.; Bhatia, N.P.; Seibert, P., Attractors in Dynamical Sys-
tems. Bol. Soc. Mat. Mex., 9 (1964), 55–66.
[7] Bacciotti, A.; Kalouptsidis, N.; Tsinias, J., Lyapunov functions and
stability of dynamical polysystems, Math. Systems Theory, 19 (1987),
333–354.
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.
[15] Birkhoff, G. D., Dynamical Systems, Amer. Math. Soc. Colloquium Pub-
lications, vol. 9, New York, 1927.
[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.
[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.
[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
[40] Conley, C. and Smoller. J., The MHD version of a theorem of H. Weyl,
Proc. Amer. Math. Soc., 42 (1974), 248–250.
[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.
[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.
[53] Gauthier, P.; Kupka, I.; Sallet, G., Controllability of right invariant
systems on real simple Lie groups, Systems & Control Letters, 5 (1984),
187–190.
[55] Gottschalk, W. H., A survey of minimal sets, Ann. Inst. Fourier (Greno-
ble), 14 (1964), 53–60.
[59] Hilgert, J. and Neeb, K.-H., Lie semigroups and their applications, Lec-
ture Notes in Mathematics, vol. 1552, Springer-Verlag, 1993.
[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.
[65] Jurdjevic, V. and Kupka, I., Control systems on semisimple Lie groups
and their homogeneous spaces, Ann. Inst. Fourier (Grenoble), 31 (1981),
186–211.
[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.
[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.
[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.
[78] Moeckel, R., Orbits near triple collision in the three-body problem, PH.D.
thesis, 1980.
BIBLIOGRAPHY 325
[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.
[87] San Martin, L. A. B., Invariant control sets on flag manifolds, Math. of
Control, Signals, and Systems, 6 (1993), 41–61.
[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
[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.
[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.
[104] Stefan, P., Accessible sets, orbits and foliations with singularities, Proc.
London Math. Soc., 29 (1974), 699–713.
[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.
[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.
[113] Willems, J., Models for dynamics, Dynamics Reported, 2 (1989), 171–
269.
[115] Zenhxin, L., The random case of Conley’s theorem III. Random semiflow
case and Morse decomposition, Nonlinearity, 20 (2007), 2773–2791.
Index
329
330 INDEX
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