Hyperbolic Chaos

Sergey P.
Kuznetsov
Hyperbolic Chaos
A Physicist’s View
Author
Sergey P. Kuznetsov
Kotel’nikov’s Institute of Radio-Engineering and
Electronics of RAS, Saratov Branch
410019, Zelenaya 38, Saratov, Russian Federation
Email: spkuz@rambler.ru
ISBN 978-7-04-031964-4
Higher Education Press, Beijing
ISBN 978-3-642-23665-5 e-ISBN 978-3-642-23666-2

Springer Heidelberg Dordrecht London New York
Library of Congress Control Number: 2011935039
c Higher Education Press, Beijing and Springer-Verlag Berlin Heidelberg 2012

This work is subject to copyright. All rights are reserved, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data
banks. Duplication of this publication or parts thereof is permitted only under the provisions
of the German Copyright Law of September 9, 1965, in its current version, and permission
for use must always be obtained from Springer. Violations are liable to prosecution under the
German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Sergey P. Kuznetsov
Hyperbolic Chaos
A Physicist’s View
With 214 figures

Preface
One of intensively developing research directions is exploration of complex dynam-

ics and chaos in nonlinear systems. We regard an object as a dynamical system, if
its state at any given time may be obtained from the initial state according to a cer-
tain rule defined for the given system. It is remarkable that such definition, although
representing ideal of deterministic vision, does not exclude a possibility of chaotic
behavior of the object, with dependence of the state on time looking like a random
process. Chaos occurs in systems of different nature, e.g. relating to mechanics, hy-
drodynamics, electronics, laser physics and nonlinear optics, chemical kinetics, and
biomedical disciplines. The main attribute of the dynamical chaos is sensitivity to
small perturbations of initial conditions, which makes it impossible to predict the
future states actually for times longer than some characteristic scale, which usu-
ally depends logarithmically on the inaccuracy of the initial conditions (“horizon of
predictability”).
For dissipative dynamical systems, chaos is associated with presence in the state
space of a curious object called the strange attractor. At present, the collection of
models with strange attractors is very rich, including artificial mathematical exam-
ples, models of physical, chemical, and biological systems.
A classic example of chaotic attractor is Lorenz attractor. It occurs in a set of
three first-order differential equations modeling fluid convection or dynamics of a
single-mode laser. The Lorenz model was a subject of extensive and careful stud-
ies in recent decades. Accurate mathematical foundation of chaotic nature of dy-
namics of the Lorenz model was found to be a subtle and delicate problem. It was
announced as the 14-th Smale problem (from the list of difficult mathematical prob-
lems suggested by Steven Smale in 1998, as challenging mathematicians of the 21st
century, similar to the list of Hilbert problems advanced for the 20th century). The
solution was given by W. Tucker in 2002 by combination of technique of computer-
assisted proof and careful analytic considerations.
Maybe, there was a lost opportunity for the nonlinear science to find out an al-
ternative, a less painful way to discover physically meaningful and mathematically
validated examples of chaotic behavior.
vi Preface
About 40 years ago, in mathematical studies a special type of chaotic attractors

was introduced, namely, the uniformly hyperbolic attractors. They relate to systems
belonging to the so-called axiom A class and are considered in the hyperbolic the-
ory associated with the names of Anosov, Alekseev, Smale, Williams, Sinai, Plykin,
Ruelle, Pesin, Newhouse and others. Chaotic nature of dynamics on the uniformly
hyperbolic attractors is proved rigorously. They possess a property of structural sta-
bility, i.e., the phase space arrangement, character of dynamics, and its statistical
characteristics are insensitive to variation of parameters and functions in the gov-
erning equations.
Originally, it was expected that the uniformly hyperbolic attractors will be rel-
evant to many physical situations of occurrence of dynamical chaos. However, as
time passed and many examples of chaotic systems of different nature had been
suggested and studied, it became clear that these examples do not fit the narrow
frames of the early hyperbolic theory. Therefore, people started to think of the uni-
formly hyperbolic attractors rather as of refined abstract image of chaos that has
no relation to real systems. Efforts of mathematicians were redirected at developing
generalizations applicable to broader classes of systems. For example, notions of
singular-hyperbolic (or quasi-hyperbolic) attractors, non-uniformly hyperbolic at-
tractors, partially hyperbolic attractors, and quasi-attractors were introduced, and
certain progress in their exploration was achieved.
Abandoned for a long time and not clarified until recently is a question on a
possibility of occurrence of the dynamical behavior associated with uniformly hy-
perbolic attractors in real world systems or, at least, in specially designed systems
in physics and technology.
In textbooks on and reviews of nonlinear dynamics, the uniformly hyperbolic
attractors are represented usually by artificial discrete time models based on geo-
metric constructions, often explained qualitatively, in words, or by graphic images.
Of course, for a physicist this is only a starting point for a research.
First of all, in addition to the geometric constructions it is desirable to have exam-
ples in a form of explicitly written equations, which allow application of computer
methods for analyzing the dynamics and calculation of quantitative characteristics
interesting to possible applications.
For some physical systems description in discrete time is very natural, and it
would be worth examining possibilities of occurrence of the hyperbolic attractors in
maps relating to such systems.
Next, it is important to turn to continuous time systems, which are of the first
place importance, e.g., in physics and technology.
It is desirable to have a clear vision of how to implement the dynamics on uni-
formly hyperbolic attractors using combinations of structural elements known in
the context of the theory of oscillations and in applications (oscillators, coupled
systems, and feedback loops).
Finally, the proposed models should be implemented as real operating devices,
for example, in electronics, mechanics, nonlinear optics, and technical applications
of such devices should be indicated with explanation of advantages over alternative
possible solutions.
Preface vii
In the theory of oscillations, since classic works of Andronov and his school,
rough or structurally stable systems are regarded as those subjected to priority
research, and as the most important for practice. It seems natural that the same
should relate to systems with structurally stable uniformly hyperbolic attractors.
The lack of physical examples in this regard looks as an evident dissonance. From
a methodological point of view, the situation is similar to that in the early 20th
century concerning the limit cycles, before their role as mathematical images for
self-oscillations was established. In a similar way, the uniformly hyperbolic chaotic
attractors should find their place as images for phenomena in real systems. It will
help to link the abstract hyperbolic theory developed by mathematicians with de-
scription of real systems and attribute this theory to physical content.
This book is devoted to review the modern state of the outlined research program.
The material is presented in four parts.
The introductory Part I consists of two chapters. Chapter 1 is devoted to the nec-
essary basic concepts, including the notion of dynamical system, attractor, Poincaré
map, and hyperbolicity. We introduce and discuss classic examples of uniformly
hyperbolic attractors: Smale-Williams solenoid, DA-attractor of Smale, and attrac-
tors of Plykin type. An overview is given to the content of the hyperbolic theory
(the cone criterion, structural stability, Markov partitions and symbolic dynamics,
measures of Sinai-Ruelle-Bowen, etc.). Chapter 2 is a review of the literature indi-
cating various possible situations of occurrence of uniformly hyperbolic attractors
in dynamical systems.
Part II is the basic one. Here, we introduce a number of examples of systems,
which allow physical implementation, and possess uniformly hyperbolic attrac-
tors with one-dimensional unstable manifolds (one positive Lyapunov exponent).
In Chap. 3 we consider models, which admit a natural stroboscopic discrete time
description. These are systems operating under periodic pulse kicks and systems
whose dynamics is composed of a periodic sequence of stages, each of which is gov-
erned by a particular form of the differential equations. Chapter 4 discusses models
designed on a base of two self-oscillators which are excited alternately because of
forced external parameter modulation and transmit the excitation to each other in
such way that its phase transforms in accordance with expanding circle map. Chap-
ter 5 is devoted to autonomous systems operating according to the same principle. In
Chap. 6 we consider schemes of parametric generators of chaos with hyperbolic at-
tractors. Chapter 7 is devoted to methods of computer verification of the hyperbolic
nature of attractors illustrated with the examples suggested in the previous chapters.
Particularly, we consider technique of visualization of mutual location of stable and
unstable manifolds, statistics of angles of intersections between them, visualization
of the natural invariant measure distributions on the attractors, and verification of
the cone criterion.
Part III is devoted to model systems, for which mathematical justification of the
hyperbolic nature of the attractors is more problematic. Hypothetically, however, the
hyperbolicity apparently takes place. This assertion is based on the fact that from a
physical point of view, those models are built following the same principles as the
systems considered in Chaps. 4 and 5, for which the hyperbolicity is justified at
viii Preface
the level of computations. In Chap. 8 we consider non-autonomous systems of four

alternately excited oscillators. Among them, there is a model, in which the trans-
formation of phases during a period of parameter modulation follows the Anosov
map on the torus, and a model associated with a map on the torus with hyperchaotic
dynamics (more than one positive Lyapunov exponent). Also a system is considered
constructed of two coupled elements, with each possessing a hyperbolic attractor.
For this case, some details are studied for the transition corresponding to complete
synchronization of chaos. In Chap. 9, an autonomous system is considered func-
tioning due to dynamics nearby a heteroclinic cycle in the amplitude equations. On
this basis, we implement an attractor of Smale-Williams type in the Poincaré map,
an attractor with dynamics of cyclic (phase) variables governed by the Anosov map,
and a hyperchaotic attractor. Chapter 10 is devoted to systems with delayed feed-
back, in which a chaotic map for the phases of successively generated oscillation
trains is arranged due to the transfer of excitation from the previous stages of activ-
ity of a single oscillatory element (self-oscillator) to the next stages. In Chap. 11 we
consider a high-dimensional system built on two ensembles of self-oscillators with
global coupling within each of them. Due to the turning the coupling on and off
alternately in two ensembles, they undergo periodic Kuramoto transitions from syn-
chronization to desynchronization and back; the interaction between the ensembles
is organized in such way that the phase of the oscillatory excitation for the mean
field transferred to each other, follows the chaotic map.
Part IV is devoted to examples of hyperbolic (or hypothetically hyperbolic) at-
tractors in electronic experiments available to date. In Chap. 12 we consider ex-
periments with a non-autonomous system of two alternately excited self-oscillators
manifesting dynamics on the attractor of Smale-Williams type. In Chap. 13 two vari-
ants of non-autonomous time-delay systems with periodic parameter modulation are
discussed. One possesses presumably an attractor of Smale-Williams type, embed-
ded in an infinite-dimensional state space of the respective stroboscopic Poincaré
map, and the other corresponds to a partially hyperbolic attractor.
The Appendix address some matters, which are essential to the whole presenta-
tion and illustrations, but dropping out of the basic structure of the book. We con-
sider algorithm for computing Lyapunov exponents based on the Gram-Schmidt
orthogonalization, and derivation of the Hénon and Ikeda maps from physical con-
siderations for systems with impulse kicks. Further, the construction of the Smale
horseshoe is presented, which gives a nontrivial example of a non-attractive compli-
cated invariant set. The Kaplan-Yorke formula is derived and explained connecting
approximately the Lyapunov exponents and dimensions of chaotic attractors. For-
mal definition of the model of Hunt is reproduced, which implements suspension
of attractor of Plykin type. A simple model of hyperbolic dynamics on a compact
surface of negative curvature is considered. The final Appendix is devoted to the
shadowing property of hyperbolic chaotic attractors, which is illustrated in compu-
tations for a model from Chap. 4 with added noise.
The book may be useful to physicists and engineers interested in the practical ap-
plication of the theory of deterministic chaos, particularly in obtaining robust chaos
insensitive to parameters and characteristics of components, fluctuations, interfer-
Preface ix
ences, etc. This may relate to various disciplines——mechanics, hydrodynamics,

electronics, laser physics, and nonlinear optics.
Contents of the book can probably help to think about possibilities of occurrence
of structurally stable chaos in systems of different physical nature. Intriguing area
for contemplations may be the significance of such chaos in neurodynamics.
On the other hand, the book can be useful for mathematicians interested in con-
crete applications of the hyperbolic theory. For them, it may be of interest to see
how the mathematical theories are refracted from the perspective of the applied dis-
ciplines.
The author tried to present the material in a style available to graduate and post
graduate students of non-mathematical specialties and to make it, as far as possible,
self-consistent to allow a study without recourse to other sources. I tried to avoid for-
mal definitions and formulations with abusing mathematical symbolism, replacing
them with rather intuitive qualitative arguments. Perhaps, a part of mathematically
oriented readers will find it insufficient; they are referred to rich literature on math-
ematical theory of dynamical systems. Also, I must warn that the consideration of
general content of nonlinear dynamics is restricted here by the minimum needed for
understanding the substantive material of the book. So, it can not be regarded as a
substitute for a systematic study of integral general courses of nonlinear dynamics,
to which the interested reader is referred for this purpose.
I am grateful, for useful discussions, assistance, and constructive criticism, to
V.S. Anishchenko, V.S. Afraimovich, B.P. Bezruchko, V.N. Belykh, A.Yu. Jalnine,
A.Yu. Zhirov, O.B. Isaeva, A.P. Kuznetsov, P.V. Kuptsov, J. Kurths, A.Yu. Losku-
tov, R. MacKay, L.A. Mel’nikov, A. Pikovsky, V.I. Ponomarenko, A.G. Rozhnev,
N.M. Ryskin, I.R. Sataev, E.P. Seleznev, W. Tucker, D.V. Treshchev, D.I. Trubet-
skov, L.V. Turukina.
Also, I thank all colleagues and administrative staff of Saratov Branch of Ko-
tel’nikov’s Institute of Russian Academy of Sciences for many years of friendly
collaboration, support, and creative atmosphere stimulating the scientific work.
The researches laying the foundation of this book were supported, in part, by
RFBR grants 06-02-16619, 09-02-00426, 09-02-00707, and DFG-RFBR grants
04-02-04011, 08-02-91963. Additionally, the author acknowledges a support from
program of Development of Scientific Potential of Higher Education of Ministry
of Education of Ministry of Education and Science of Russian Federation (grant
2.1.1/1738).
Sergey P. Kuznetsov
Saratov, February 2011
Contents
Part I Basic Notions and Review
1 Dynamical Systems and Hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.1 Dynamical systems: basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Systems with continuous and discrete time, and their
mutual relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Dynamics in terms of phase fluid: Conservative and
dissipative systems and attractors . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.3 Rough systems and structural stability . . . . . . . . . . . . . . . . . . . 8
1.1.4 Lyapunov exponents and their computation . . . . . . . . . . . . . . 10
1.2 Model examples of chaotic attractors . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.1 Chaos in terms of phase fluid and baker’s map . . . . . . . . . . . . 12
1.2.2 Smale-Williams solenoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.3 DA-attractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.4 Plykin type attractors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3 Notion of hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Content and conclusions of the hyperbolic theory . . . . . . . . . . . . . . . . 22
1.4.1 Cone criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4.2 Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4.3 Transversal Cantor structure and Kaplan-Yorke
dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4.4 Markov partition and symbolic dynamics . . . . . . . . . . . . . . . . 26
1.4.5 Enumerating of orbits and topological entropy . . . . . . . . . . . . 27
1.4.6 Structural stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.4.7 Invariant measure of Sinai-Ruelle-Bowen . . . . . . . . . . . . . . . . 29
1.4.8 Shadowing and effect of noise . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.4.9 Ergodicity and mixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.4.10 Kolmogorov-Sinai entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
xii Contents
2 Possible Occurrence of Hyperbolic Attractors . . . . . . . . . . . . . . . . . . . . . 35

2.1 The Newhouse-Ruelle-Takens theorem and its relation to the
uniformly hyperbolic attractors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2 Lorenz model and its modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Some maps with uniformly hyperbolic attractors . . . . . . . . . . . . . . . . 40
2.4 From DA to the Plykin type attractor . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5 Hunt’s example: Suspending the Plykin type attractor . . . . . . . . . . . . 46
2.6 The triple linkage: A mechanical system with hyperbolic
dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.7 A possible occurrence of a Plykin type attractor in
Hindmarsh-Rose neuron model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.8 Blue sky catastrophe and birth of the Smale-Williams attractor . . . . . 52
2.9 Taffy-pulling machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Part II Low-Dimensional Models
3 Kicked Mechanical Models and Differential Equations with

Periodic Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.1 Smale-Williams solenoid in mechanical model: Motion of a
particle on a plane under periodic kicks . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 A set of switching differential equations with attractor of
Smale-Williams type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.3 Explicit dynamical system with attractor of Plykin type . . . . . . . . . . . 68
3.3.1 Plykin type attractor on a sphere . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3.2 Plykin type attractor on the plane . . . . . . . . . . . . . . . . . . . . . . . 73
3.4 Plykin-like attractor in smooth non-autonomous system . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 Non-Autonomous Systems of Coupled Self-Oscillators . . . . . . . . . . . . . . 81

4.1 Van der Pol oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 Smale-Williams attractor in a non-autonomous system of
alternately excited van der Pol oscillators . . . . . . . . . . . . . . . . . . . . . . . 84
4.3 System of alternately excited van der Pol oscillators in terms of
slow complex amplitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.4 Non-resonance excitation transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5 Plykin-like attractor in non-autonomous coupled oscillators . . . . . . . 95
4.5.1 Representation of states on a sphere and equations of the
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.5.2 Numerical results for the coupled oscillators . . . . . . . . . . . . . 98
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5 Autonomous Low-dimensional Systems with Uniformly Hyperbolic

Attractors in the Poincar é Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.1 Autonomous system of two coupled oscillators with
self-regulating alternating excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Contents xiii
5.2 System constructed on a base of the predator-prey model . . . . . . . . . 107

5.3 Example of blue sky catastrophe accompanied by a birth of
Smale-Williams attractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6 Parametric Generators of Hyperbolic Chaos . . . . . . . . . . . . . . . . . . . . . . 119

6.1 Parametric excitation of coupled oscillators. Three-frequency
parametric generator and its operation . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.2 Hyperbolic chaos in parametric oscillator with Q-switch and
pump modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2.1 Dynamical equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2.2 Qualitative explanation of the operation . . . . . . . . . . . . . . . . . 126
6.2.3 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2.4 Numerical results in the frame of method of slow complex
amplitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.3 Parametric generator of hyperbolic chaos based on four coupled
oscillators with pump modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.3.1 Model, operation principle and basic equations . . . . . . . . . . . 132
6.3.2 Chaotic dynamics: results of computer simulation . . . . . . . . . 134
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7 Recognizing the Hyperbolicity: Cone Criterion and Other

Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.1 Verification of transversality for manifolds . . . . . . . . . . . . . . . . . . . . . 141
7.1.1 Visualization of the manifolds . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.1.2 Distributions of angles of the manifold intersections . . . . . . . 144
7.2 Visualization of invariant measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.3 Cone criterion and examples of its application . . . . . . . . . . . . . . . . . . 155
7.3.1 Procedure of verification of the cone criterion . . . . . . . . . . . . 155
7.3.2 Examples of application of the cone criterion . . . . . . . . . . . . . 161
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Part III Higher-Dimensional Systems and Phenomena
8 Systems of Four Alternately Excited Non-autonomous Oscillators . . . 173

8.1 Arnold’s cat map dynamics in a system of coupled
non-autonomous van der Pol oscillators . . . . . . . . . . . . . . . . . . . . . . . . 173
8.2 Dynamics corresponding to hyperchaotic maps . . . . . . . . . . . . . . . . . . 180
8.2.1 System implementing toral hyperchaotic map . . . . . . . . . . . . 180
8.2.2 Model with cascade transfer of excitation upward the
frequency spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
8.3 Hyperchaos and synchronous chaos in a system of coupled
non-autonomous oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.3.1 Equations and basic modes of operation . . . . . . . . . . . . . . . . . 188
8.3.2 Equations for slow complex amplitudes . . . . . . . . . . . . . . . . . 193
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
xiv Contents
9 Autonomous Systems Based on Dynamics Close to Heteroclinic

Cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.1 Heteroclinic connection: an example of Guckenheimer and
Holmes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
9.2 Attractor of Smale-Williams type in a system of three coupled
self-oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.3 Attractor with dynamics governed by the Arnold cat map . . . . . . . . . 207
9.4 Model with hyperchaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
9.5 An autonomous system with attractor of Smale-Williams type
with resonance transfer of excitation in a ring array of van der Pol
oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
10 Systems with Time-delay Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

10.1 Some notions concerning differential equations with deviating
argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
10.2 Van der Pol oscillator with delayed feedback, parameter
modulation and auxiliary signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
10.2.1 Attractor of Smale-Williams type in the time-delayed
system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
10.2.2 Hyperchaotic attractors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
10.3 Van der Pol oscillator with two delayed feedback loops and
parameter modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
10.4 Autonomous time-delay system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
11 Chaos in Co-operative Dynamics of Alternately Synchronized

Ensembles of Globally Coupled Self-oscillators . . . . . . . . . . . . . . . . . . . . 243
11.1 Kuramoto transition in ensemble of globally coupled oscillators . . . 243
11.2 Model of two alternately synchronized ensembles of oscillators . . . . 247
11.2.1 Collective chaos in ensemble of van der Pol oscillators . . . . . 248
11.2.2 Slow-amplitude approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
11.2.3 Description of the dynamics in terms of ensembles of
phase oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Part IV Experimental Studies
12 Electronic Device with Attractor of Smale-Williams Type . . . . . . . . . . . 259

12.1 Scheme of the device and the principle of operation . . . . . . . . . . . . . . 259
12.2 Experimental observation of the Smale-Williams attractor . . . . . . . . . 260
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Contents xv
13 Delay-time Electronic Devices Generating Trains of Oscillations

with Phases Governed by Chaotic Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 265
modulation and auxiliary signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
parameter modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
14 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Appendix A Computation of Lyapunov Exponents:

The Benettin Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Appendix B Hénon and Ikeda Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Appendix C Smale’s Horseshoe and Homoclinic Tangle . . . . . . . . . . . . . . . . 289

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
Appendix D Fractal Dimensions and Kaplan-Yorke Formula . . . . . . . . . . . 293

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Appendix E Hunt’s Model: Formal Definition . . . . . . . . . . . . . . . . . . . . . . . . 299

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Appendix F Geodesics on a Compact Surface of Negative Curvature . . . . 305

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Appendix G Effect of Noise in a System with a Hyperbolic Attractor . . . . 311

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Part I
Basic Notions and Review
Chapter 1
Dynamical Systems and Hyperbolicity
Abstract In this chapter we review basic notions of the theory of dynamical sys-
tems essential for understanding subsequent chapters of the book. Particularly, a
definition of the dynamical system is discussed and some important classes like
continuous-time and discrete-time systems, conservative and dissipative systems,
autonomous and non-autonomous systems are introduced. Interpretation of dynam-
ics as evolution of a cloud of representative points in the state space is considered
and implemented for explanation of chaos and, particularly, for the uniformly hy-
perbolic attractors, like Smale-Williams solenoid, DA attractor of Smale, and Plykin
type attractors. Lyapunov exponents as a tool for quantitative approach to analyzing
chaotic or regular dynamics are examined, and methodic of their computation is dis-
cussed. Main notions of the hyperbolic theory are considered, and its substational
content is briefly reviewed.
1.1 Dynamical systems: basic notions
1.1.1 Systems with continuous and discrete time, and their mutual
relation
A finite-dimensional dynamical system is an object, for which one can specify a

state by a collection of a finite number N of real variables named a state vector,
supposing that there exists a definite rule called the evolution operator that makes
it possible to indicate precisely the state vector resulting from the initial one at any
latter time instant (Birkhoff, 1927; Schuster and Just, 2005; Thompson and Stewart,
1986; Strogatz, 2001; Hasselblatt and Katok, 2003). A set of all possible states is
a phase space, or state space; its dimension equals just the number of variables N
needed to specify the state vector. In other words, this is a space with coordinate
axes, each of which is associated with one dynamical variable from the set of N of
them.
4 1 Dynamical Systems and Hyperbolicity
Both continuous-time and discrete-time systems are introduced and considered;

in mathematical literature they are called flows and cascades, respectively.
Evolution of a state in time corresponds to motion of the representative point
in the phase space along the phase trajectory, or orbit. A set in the phase space is
called invariant set if all representative points from this set are transformed by the
evolution operator again to the points belonging to the same set.
In order to describe continuous-time autonomous systems, one can use differen-
tial equations of the form
dx/dt = F(x). (1.1)
Here x = (x1 , x2 , . . . , xN ) is a state vector of dimension N, and F = (F1 (x), F2 (x), . . . ,
FN (x)) is a vector function. In coordinate notation,
dx1 /dt = F1 (x1 , x2 , . . . , xN ),

dx2 /dt = F2 (x1 , x2 , . . . , xN ),
(1.2)
............
dxN /dt = FN (x1 , x2 , . . . , xN ).
By virtue of the theorem of existence and uniqueness of solutions for sets of dif-
ferential equations (Arnold, 1978), with a given state at some instant t 0 , one can
determine states in the future t > t 0 , as well as in the past t < t0 . In other words, the
evolution of a state may be monitored both forward and backward in time.
If the right-hand side function F in the differential equation depends on time ex-
plicitly, the system is called non-autonomous. Physically, it corresponds to systems
operating in the presence of external time-dependent forcing. To specify the state
in this case, besides the vector x one needs to indicate the time instant it relates to.
Therefore, we introduce the space of dimension N+1 with an additional coordinate
axis t, which in this context is called the extended phase space. In this book, refer-
ring to non-autonomous systems, we will consider only the class of systems with
the functions periodic in time: F(x,t + T ) = F(x,t).
Discrete time systems are described by evolution rules defined by iterated maps,
which correspond to transformation of the states step by step,
xn+1 = g(xn ). (1.3)
Here x is a state vector, and g is a vector function specifying the evolution operator.
In this case, a phase trajectory is a discrete sequence of points in the phase space.
Evolution that takes place for k steps corresponds to k-fold iteration of the map; it is
designated as
xn+k = gk (xn ) ≡ g(g(· · · g(xn ) · · · )) ≡ g ◦ g ◦ · · ·◦ g(xn ). (1.4)

k times k times
Continuous time systems and discrete time systems are closely related. Proce-
dure of passage from one to another is known as the Poincar é section construction.
In phase space of a continuous time system one selects some fixed surface in such
a way that the phase trajectories cross it again and again in the course of the time
1.1 Dynamical systems: basic notions 5
evolution (Fig. 1.1(a)). Then we introduce a function g, which maps a point on the
cross-section to another point at the next cross of the same surface by the trajec-
tory emitting from the initial point. This function determines the Poincar é map, or
return map. If the coefficients in differential equations have smooth dependence on
the dynamical variables, then the Poincaré map is a diffeomorphism, a continuously
differentiable function, having a uniquely defined inverse function, which is contin-
uously differentiable too.
For non-autonomous systems with periodic functions of time in the right-hand
part of the differential equations, a standard method of constructing the Poincaré
map corresponds to stroboscopic description: the dynamics is monitored at discrete
time instants with the period T . It means selection of the cross-section in the ex-
tended phase space with a family of planes t=nT, and the mapping is performed
from one plane to the next one (Fig. 1.1(b)). All these planes may be regarded as
identified because of the assumed time periodicity of the function F.
Fig. 1.1 Poincaré map construction for 3-dimensional autonomous system (a) and for a non-
autonomous system with 3-dimensional extended phase space (b).
It is rather a rare case when the Poincaré map may be derived in explicit
form; alternatively, one can construct it as a numerical algorithm.
For periodically non-autonomous systems the procedure is easy: we sim-
ply select a constant step of the finite-difference integration in such a way that
the period of variation of coefficients in the differential equations contains an
integer number of these steps. Starting with some initial state, after this num-
ber of steps we get the state corresponding to the result of the transformation
of the original state under the stroboscopic Poincaré map.
For the case of autonomous dynamics, one has to define the preliminary
Poincaré section by some equation, say S(x 1 , x2 , . . . , xN ) = 0. Starting from
initial state, which satisfies this condition, we should perform the integration
till this condition will be valid again. Practically, of course, it will happen
somewhere between the consecutive integration steps; here we have to stop
and apply some interpolation procedure. An elegant recipe to do this in perfect

consistency in accuracy with the finite-difference method was designed by
Hénon. Shortly, we supplement the procedure with one non-standard final step
of integration, which is produced with the same finite-difference method, but
applied to the differential equations reformulated by taking S(x 1 , x2 , . . . , , xN )
as the independent variable (Hénon, 1982). The length of this additional step
just corresponds to the value of S obtained in the previous step.
A continuous time system restored from a given invertible map is referred to as

a suspension.1
1.1.2 Dynamics in terms of phase fluid: Conservative and

dissipative systems and attractors
Traditionally, two classes of dynamical systems are distinguished, the conservative

and dissipative ones. In physics, the conservation means conserving the energy. For
example, a mechanical oscillatory system without friction relates to the conservative
class. In the presence of friction, it becomes a dissipative system. In the last case,
the mechanical energy is not conserved, but gradually dissipates being transformed
to heat, i.e. to the energy of microscopic motion of molecules of the physical system
itself, and of its environment. Strictly speaking, now the temporal evolution should
be determined not only by the macroscopic state of the system, but also by variables
relating to the molecules. Nevertheless, in many cases, description of the frame-
work of the theory of dynamical system in terms of macroscopic variables remains
sufficient, reasonable and accurate; then, we get a dissipative dynamical system.
It would be good to have a definition appealing not to a particular class of objects
of specific physical nature, but to the basic concepts of the theory of dynamical
systems.
Let us consider an ensemble composed of a very large number of identical non-
interacting dynamical systems, differing only in initial conditions. In the phase space
the ensemble corresponds to a cloud of representative points. This cloud evolves in
time, changing size and shape due to motion of all individual points in accordance
with the dynamic equations of a single underlying system.
It may occur that each finite-size element of the cloud evolves in such a way
that its phase volume remains constant (at least, with description of some properly
chosen variables). This is the case of the conservative systems.
1 The term suspension is preferable for flow systems in which the mapping of the required form
takes place for a fixed time period, which corresponds to dynamics in a non-autonomous system
governed by differential equations with coefficients periodic in time. In more general case, with
respect to the system, for which the given map corresponds to the Poincaré map, the term special
flow is used (Hasselblatt and Pesin, 2008).
In classic mechanics, Lagrangian and Hamiltonian formalisms are devel-

oped for the conservative systems of even dimension of the phase space N
(Landau and Lifshitz, 1976; Arnold, 1989). Traditionally, the twice less num-
ber N/2 is called the number of degrees of freedom.
Lagrangian description is based on a scalar-valued function L(q 1 , . . . , qN/2 ,
q̇1 , . . . , q̇N/2 ) depending on generalized coordinates q i and generalized veloc-
ities q̇i . For mechanical systems it has a sense of difference of potential and
kinetic energy. Differential equations are obtained from the Lagrange function
in a standard way; they are called the Euler-Lagrange equations:
d ∂L ∂L
− = 0 (i = 1, . . . , N/2).
dt ∂ ẋi ∂ xi
Considered in terms of variables (q i , q̇i ), the dynamics of a conservative sys-

tem, in general, does not obey the property of preserving phase volume. To
have this property, another set of variables has to be used. Namely, instead of
the generalized velocities the generalized momenta are introduced by means
∂L
of the relations p i = q̇i . To reformulate the description in terms of the
∂ q̇i
canonical variables (q i , pi ), the Hamiltonian function is used
N/2
∂L
H(q1 , . . . , qN/2 , p1 , . . . , pN/2 ) = ∑ q̇i ∂ q̇i − L.
i=1
Then, the dynamics is governed by the Hamilton equations

∂ H(p1 , . . . , pN/2 , q1 , . . . , qN/2 )
ṗi = − ,
∂ qi
∂ H(p1 , . . . , pN/2 , q1 , . . . , qN/2 )
q̇i = , i = 1, . . . , N/2.
∂ pi
Conservation of phase volume of an element evolving in the state space
(qi , pi ) is stated by the Liouville theorem.
For non-autonomous systems the formalism remains the same, although
the Lagrangian and Hamiltonian functions depend additionally on the time
variable t in explicit way.
For the case of discrete time, correct description of Hamiltonian systems is
defined by an implicit set of equations expressed via one scalar-valued func-
tion of N variables F(q 1 , . . . , qN/2 , q1 , . . . , qN/2 ) called the generating func-
tion:
∂ F(q1 , . . . , qN/2 , q1 , . . . , qN/2 )
pi = − ,
∂ qi
∂ F(q1 , . . . , qN/2 , q1 , . . . , qN/2 )
pi = , i = 1, . . . , N/2.
∂ qi
(Here the values marked with prime relate to the next step of the discrete
time.)
In dissipative systems the phase volume decreases in the course of time evolu-
tion, at least, in average. So, the cloud of representative points eventually settles on
some subset in the phase space, called attractor (or, possibly, on several attractors).
In conservative case the cloud of the representative points may be thought of as
composed of incompressible fluid, while in the dissipative case it must be regarded
as compressible substance, like vapor able to condensate with essential decrease of
its volume.
Attractors represent sustained dynamical regimes in dissipative systems arising
due to long-time evolution. The simplest example of attractor is a stable fixed point,
or stable equilibrium. Another simple example is a stable limit cycle, a closed or-
bit, representing the periodic self-oscillations. An attractor associated with sustained
quasiperiodic motion, which is a composition of two or more components of incom-
mensurable frequencies is torus of respective dimension. Chaotic sustained regimes
correspond to specific nontrivial sets in the phase space called the strange attractors
(Ruelle and Takens, 1971).
In the case of coexistence of two or more attractors in the phase space, we talk
about bistability or multistability. A set of states in the phase space, from which the
orbits arrive at the given attractor, is called the basin of this attractor.
It appears to be a hard problem to give a satisfactory formal definition of attractor
embracing all cases that may occur (Milnor, 1985). For aims of this book, sufficient
is a definition of the so-called maximal attractor, which uses a notion of absorbing
domain.
For a discrete time case, with evolution operator defined by a function g(x), con-
sider a domain in the phase space mapped inside itself: g(D) ⊂ Int(D). (Here g(D)
means a set of images of representative points, which relate to D.) Then, the attractor
A certainly belongs to D, and may be defined as intersection of images of D under
successive n iterations of the map, namely,
∞

A= gn (D). (1.5)
n=1
For a continuous time case the notion of the maximal attractor implies that there
exists a connected domain D bounded by a surface B of such kind that all orbits
cross B transversally (with non-zero angle) being directed inside the domain D.
1.1.3 Rough systems and structural stability
In 1937 Andronov and Pontryagin introduced the concept of rough systems (An-
dronov and Pontrjagin, 1937; Andronov et al., 1966), later reformulated and devel-
oped under the name of structural stability (Smale, 1967; Shilnikov, 1997; Guck-
enheimer and Holmes, 1983; Katok and Hasselblatt, 1995; Shilnikov et al., 1998;
2002). The roughness is a property of a dynamical system to preserve the qualitative
behavior of trajectories under small variation of parameters and functions in the def-
inition of the evolution operator. A more accurate formulation is as follows: While
those variations are small in the class of continuous functions with first derivatives
(in the common notation, the functions of class C 1 ), the system remains topologi-
cally equivalent to the original one. It means that there exists a continuous invertible
change of variables (a homeomorphism, class C 0 ), which transforms directed phase
trajectories of the perturbed system to those of the original one.
Andronov and Pontryagin indicated a necessary and sufficient condition for
roughness of two-dimensional flow systems in a bounded domain. (It is the presence
of a finite number of fixed points and periodic orbits with eigenvalues of respective
linearized problems distant from zero, and the absence of orbits connecting saddle
fixed points.)
The notion of roughness is essential for physical and technical applications be-
cause one can be sure that the mathematical models for the case of rough systems
describe correctly all relevant qualitative features of dynamical behavior in spite of
inaccuracy of the models or of the presence of inevitable external interferences and
noises.
The idea of roughness, or structural stability, became a foundation for a great
research program in oscillation theory advanced and developed by Andronov and
his school (Andronov et al., 1966; Shilnikov et al., 1998; Shilnikov et al, 2002).
According to this program, rough systems are regarded as the most typical (that’s
true, at least, for the two-dimensional flow systems!) and, hence, subjected to pri-
ority research as the most important for practice. The next step is examination of
systems of the first level of non-roughness, which may be thought of as separating
rough classes in a space of systems, that corresponds to theory of bifurcations of
codimension one. Next, one can turn to situations separating the systems of the first
degree of non-roughness, and it will be theory of codimension-two bifurcations, and
so on (Kuznetsov Yu., 1998).
The outlined research program was completely successful in application to the
flow systems of dimension two, but attempts of generalization for dimension three
and more met great difficulties.
For higher dimensions a straightforward generalization of the Andronov-Pontr-
yagin approach was found appropriate only for a restricted class called the Morse-
Smale systems (Smale, 1967; Palis, 1969; Guckenheimer and Holmes, 1983; De-
vaney, 2003), which are characterized by a finite number of fixed points and periodic
orbits in a bounded domain of the phase space. Beside this class, the general picture
for the multidimensional case is extremely complex, and the notion of roughness
seems not helpful and elucidative. Particularly, it is known that whole regions may
exist in parameter spaces, where structurally unstable systems occur on a dense set
(the Newhouse regions) (Newhouse, 1979; Guckenheimer and Holmes, 1983). All
these problems were actively studied in last several decades.
However, in the framework of the present book we must specially outline one
more class of higher-dimensional rough systems (beside the Morse-Smale ones):
these are just the systems with uniformly hyperbolic chaos. For them, the consistent
theory was developed exclusively in mathematics (Smale, 1967; Williams, 1974;
Katok and Hasselblatt, 1995; Afraimovich and Hsu, 2003; Bonatti et al., 2005; Has-
selblatt et al., 2008). Lack of physical examples of this kind having existed until
recently is a great dissatisfaction. It seems natural that general argumentation con-
cerning theoretical and practical significance of rough systems must relate to the
systems with hyperbolic chaos too; so, the problem deserves attention in concern of
physical systems and technical devices.
1.1.4 Lyapunov exponents and their computation
To distinguish regular and chaotic dynamics on a quantitative level, a useful and

constructive method is concept of the Lyapunov exponents (Pesin, 1977; Rabi-
novich and Trubetskov, 1989; Guckenheimer and Holmes, 1983; Barreira and Pesin,
2001; Hilborn, 2000; Anishchenko, 2002; Ott, 2002; Schuster, and Just, 2005; An-
ishchenko et al., 2007).
Let x(t)be some phase trajectory of the dynamical system (1.1). Consider an en-
semble associated with a cloud of representative points forming initially a small ball
of radius ε centered at the starting point of the reference orbit x(t 0 ). How will this
cloud evolve in time? While its size is small, it will take a form of an N-dimensional
ellipsoid, like that shown in Fig. 1.2. Let us expressthe semi-principal axes of this
ellipsoid at some time instant t as {l 1 , l2 , . . . , lN } = ε eλ1t , ε eλ2t , . . . , ε eλN t . Alge-
braically, they are related with the singular values of the matrix governing the evo-
lution of the perturbation vectors along the reference orbit (Greene and Kim, 1987).
In the limit of small ε and large t, a set of the quantities {λ 1 , λ2 , . . . , λN } determines
Fig. 1.2 Geometric interpretation of spectrum of Lyapunov exponents: each exponent character-
izes scale change along one of principle axes of ellipsoid arising from an infinitesimal ball centered
at the starting point of the considered orbit.
the spectrum of Lyapunov characteristic exponents. A common convention is to or-

der them in such a way that λ 1 λ2 · · · λN . Each exponent may be thought
of as indicating rate of expansion (if positive) or compression (if negative) for the
cloud of representative points along the principal axes of the ellipsoid.
It should be understood that the above definition speaks about Lyapunov expo-
nents of an individual reference orbit. Can this notion be applied to attractor, as a
whole? Certainly, it is so for attractors consisting of a single orbit, like a limit cycle,
or a stable fixed point. However, attractors being of more complex nature are com-
posed of a set of trajectories, like a torus, or a strange attractor. Then, the question
is not so trivial because distinct orbits may have different Lyapunov exponents.
In fact, appropriate generalization of the notion of Lyapunov spectra is possi-
ble: we refer to the Lyapunov spectrum of an attractor as defined for a typical tra-
jectory on the attractor. This approach is justified mathematically by the so-called
multiplicative ergodic theorem (Oseledets, 1968; Pesin, 1977; Raghunathan, 1979;
Guckenheimer and Holmes, 1983).
Analytic evaluation of the Lyapunov exponents for dynamical systems is avail-
able rarely, so, usually we need to turn to the computational estimates.
To be concrete, consider a continuous-time autonomous system (1.1). General-
izations to a non-autonomous case and to diffeomorphisms are simple and straight-
forward. The computations are based on simultaneous solution of the dynamical
equation (1.1) on the reference trajectory belonging to the attractor and the equation
for small perturbation vector, linearized near that trajectory. To derive the last one,
we substitute x(t) + x̃(t) into (1.1) and expand the right-hand part in powers of the
perturbation x̃ = (x̃ 1 , x̃2 , . . . , x̃N ) neglecting terms of order higher than one. It yields
the so-called variation equation
d x̃
= A (x(t)) x̃, (1.6)
dt
where A(x(t)) is a matrix composed of all partial derivatives of the components of
the vector function F(x) over the components of the vector x:
⎛ ⎞
∂ F1 ∂ F1 ∂ F1
⎜ ∂ x1 ∂ x2 · · ·
⎜ ∂ xN ⎟
⎟
⎜ ∂F ∂F ∂ F2 ⎟
⎜ 2 2
··· ⎟
⎜ ⎟
A = ⎜ ∂ x1 ∂ x2 ∂ xN ⎟ . (1.7)
⎜ . .. .. ⎟
⎜ .. ⎟
⎜ . . ⎟
⎝ ∂ F1 ∂ F2 ∂ FN ⎠
···
∂ xN ∂ xN ∂ xN
In general, the matrix A depends on time via the variable x relating to the unper-
turbed reference trajectory. A linear vector space associated with each point on the
reference orbit x, where the perturbation vectors x̃ live, is called the tangent space.
To compute more than one Lyapunov exponent, one needs to monitor a collection
of the respective number of perturbation vectors evolving in time along the reference
phase trajectory. However, without special measures, each vector in the collection
typically contains a component characterized by the maximum Lyapunov exponent
that will dominate at large time scale, and then evaluation of other Lyapunov expo-
nents becomes impossible. Moreover, in the case of positive maximal exponent the
norms of the vectors will manifest unbound growth with overfilling the memory reg-
isters in computer. To overcome these difficulties, it is recommended in the course
of the algorithm operation to redefine from time to time the perturbation vectors ap-
plying the Gram-Schmidt orthogonalization process and renormalization (Benettin
et al., 1980). The Lyapunov exponents are determined as mean rates for increase or
decrease of accumulating sums for logarithms of ratios of norms of the vectors after
orthogonalization but before the normalization (See explanation of the procedure in
more detail in Appendix A). 2
1.2 Model examples of chaotic attractors
1.2.1 Chaos in terms of phase fluid and baker’s map
Let us turn again to representation of dynamics in terms of ensemble of a large

number of identical systems.
Chaos in dynamics occurs if the cloud of representative points in the course

of evolution in the phase space undergoes repetitive deformations of stretch-
ing, folding, and transversal compression.
To illustrate, let us consider a dissipative version of the so-called baker map

(Manneville, 1995; Schuster and Just, 1995; Tél and Gruiz, 2006), see Fig. 1.3. Take
a unit square on the plane (x, y) and deform it, like baker rolls a piece of dough. Next,
cut the piece in half, and place the halves, one above the other, along the top and
the bottom sides of the unit square, leaving a gap in the middle. Let us assume that
Fig. 1.3 Geometric illustration of the dissipative baker’s map.

2 Now, some alternative approaches to computation of Lyapunov exponents have been suggested,
e.g. (Geist et al., 1990; Christiansen and Rugh, 1997; Rangarajan et al., 1998).
1.2 Model examples of chaotic attractors 13
in this process the horizontal size is doubled, and the vertical width becomes three
times narrower. The total area then is decreasing; it reflects the dissipative nature
of the mapping (the “dough” in this case should be imagined as a compressible
substance). When the transformation is repeated many times, a layered structure
appears containing horizontal bands, a number of which grows proportionally to
2n (n is the number of iterations), whereas the total width of the bands decreases
proportionally to 3 −n .
Attractor is the object arising in the limit of infinitely large number of iterations.
In the cross-section it has structure of the Cantor set. Note that in this example
distribution of the density of the “dough” along the layers tends to the uniform one.
It is worth stressing that the individual dynamics of particles of the “dough” is
chaotic. Indeed, formally a one-step transformation of the instant state of a single
particle in coordinates (x, y) is governed by the equations
⎧ ⎧
⎨ xn+1 = 2xn 1 ⎨ xn+1 = 2xn − 1 1
at x , at xn > . (1.8)
⎩ yn+1 = 1 yn ⎩ yn+1 = 1 yn + 2
n
2 2
3 3 3
To explain chaotic nature of the dynamics, represent an initial value of x by a real
number in binary notation, for example, 0.0110010110111 . . .. One step evolution
of the system corresponds to a shift of the binary sequence by one position to the
left discarding the leftmost digit. Note that a character at the first position after the
binary point indicates residence of the particle at the present moment in the left or
in the right half of the unit square for 0 or 1, respectively. When taking the initial
condition as a random real number from the unit interval, we get a random set of
zeros and ones in the binary notation; so the particle will appear in the left and right
halves of the square in accordance with the prescribed random sequence of digits.
Despite its simplicity, the baker map shows many features common for chaotic
attractors, for example, the transverse Cantor structure and specific combination of
stability, in the sense of attraction of the orbits to the limit set in transverse direction,
and instability, in the sense of mutual divergence of close orbits along the fibers of
the attractor.
In this construction a natural dissatisfaction is the involved “cutting” the piece
of the “dough”. One can try to repair this unlikable feature, assuming the piece is
not cut, but folded and placed again in the unit square (Fig. 1.4). Then, on the first
steps one observes again formation of the layered structure. However, in the limit
of a large number of iterations, no uniform distribution along the fibers (or at least
close to that) arises! What is observed, the substance has a tendency to concentrate
at folds, forming singularity of the density distribution. The reason is that at the
place of folding the direction of the compression coincides with the vertical tangent
at the edge of the “dough” piece, and the flattening gives rise to appearance of a
local region of high density. In nonlinear dissipative systems, such situations may
Fig. 1.4 An attempt of modification of the dissipative baker map with folding instead of cut of
the “dough”; two steps of successive iteration are shown. Formation of the layered structure is
accompanied in this case with appearance of local singularities of density at places of folding.
be associated with non-hyperbolic chaos, like that in Hénon map (Appendix B), or
with subsidence of the cloud on regular attractors (fixed points and cycles). 3
Hyperbolic chaos corresponds to situations when the transformation of a

cloud of representative points, including the longitudinal stretch and transver-
sal compression, is carried out perfectly, without cuts, discontinuities, and
formation of local singular concentrations.
A principal possibility of occurrence of attractors with such properties may be

proven by consideration of artificial examples. In each of them we, first, define the
phase space, then indicate an absorbing domain D, and formulate a rule of evolution
in discrete time g(x) in such a way that on one step this domain is mapped inside
itself: g(D) ⊂ Int(D). Then, the attractor is obviously located inside this domain
∞

being formally defined as A = gn (D).
n=1
3This type of dynamics is associated with the construction called Smale’s horseshoe (see Appendix
C).
1.2.2 Smale-Williams solenoid
The first example is attractor of Smale-Williams (Smale, 1967; Williams, 1974;

Shilnikov, 1997; Guckenheimer and Holmes, 1983; Katok and Hasselblatt, 1995;
Afraimovich and Hsu, 2003; Hasselblatt and Katok, 2003; Hasselblatt et al., 2008).
The simplest case corresponds to a three-dimensional map. Consider a domain in
the form of torus in three-dimensional space. Let us think of this torus as a plas-
tic doughnut. One step of transformation is the following: we stretch the doughnut
twice, squeeze it in the transversal direction, then fold it to get the double loop, and
insert this loop within the original torus (Fig. 1.5). To fit it there, we must assume
that the transverse size is reduced more than twice. At each next step of the trans-
formation, the total volume decreases (that means the map is dissipative) and the
number of coils is doubled. In the limit of infinite number of steps, it tends to infin-
ity, and the object is formed, which is known as solenoid. In the transversal cross
sections it has a Cantor-like structure. 4
The essential point of the construction is that the angular coordinate undergoes
doubling at each discrete time step: ϕ n+1 = 2ϕn (Fig. 1.5). An obvious generaliza-
tion is to carry out the procedure making a loop with a number of coils M. It corre-
sponds to the mapping for the angular coordinate of the form ϕ n+1 = M ϕn ( mod 2π ).
With M 2 it is termed an expanding circle map, or a Bernoulli map.
Fig. 1.5 Absorbing toroidal domain in three-dimensional phase space, results of its transformation
in the first two kinds of iteration, and the Smale-Williams solenoid, obtained after a large number
of repetitive applications of the map (right).
4 As a topological object, the solenoid was introduced by Vietoris and van Dantzig (Shilnikov,
1997), but as a chaotic attractor it appears due to Smale (Smale, 1967) and Williams (Williams,
1974). Interestingly, the solenoid occurs as well in a very different context in flow dynamical
systems, as a non-chaotic attracting set at the Feigenbaum critical point of accumulation of period-
doubling bifurcations (Vul et al., 1984). In the latter case, the formation of each new level of the
intrinsic fractal structure is associated with doubling time scale, not with one step of discrete time
as it is for the Smale-Williams attractor.
1.2.3 DA-attractor
The next example is the so-called DA-attractor proposed and named by Smale (DA
stands for “Derived from Anosov”). It takes place for a two-dimensional map de-
fined on a surface of a torus (Smale, 1967; Shilnikov, 1997; Katok and Hasselblatt,
1995).
Let us start with a map of Anosov, also known as the Arnold cat map (Arnold
and Avez, 1968; Anosov et al., 1995; Devaney, 2003)
pn+1 = pn + qn, qn+1 = pn + 2qn ( mod 1), (1.9)
where ϕ = 2π p, θ = 2π q are angular coordinates on the torus. Graphically, it is

convenient to represent the phase space as a unit square, having in mind periodic
continuation of all structures in the horizontal and vertical directions (Fig. 1.6).
The Anosov map is conservative and has a fixed point of saddle type in the
√ directions given by the vectors a 1 = (1,W ) and
origin, with stable and unstable
a2 = (−W, 1), where W = ( 5 + 1)/2.
Fig. 1.6 Illustration of the hyperbolic Anosov mapping, also known as the Arnold cat map, on the
unit square representing the unfolding of the torus (a). The Smale surgery performed in the vicinity
of the fixed point at the origin (b). Portrait of the DA-attractor depicted in the unit square (c) and
on the surface of torus (d).
To get a dissipative map with attractor, Smale suggested modifying the map by
means of a “surgical operation” in a small neighborhood of the fixed point. The pro-
cedure is carried out in such a way that the motion along the unstable direction is not
perturbed, but in the orthogonal direction it modifies; the fixed point becomes repul-
sive, and, hence, two saddle points S 1 and S2 appear nearby (Fig. 1.6(b)). Outside
the region subjected the “surgery”, the mapping does not change.
In this setup, the absorbing domain is the entire surface of the torus, except for
a truncated circular neighborhood of the origin (the saddles S 1 and S2 are inside
this excluded neighborhood). Under iteration, the cut in the cloud of representative
points elongates along the vector a 1 with simultaneous contraction in the orthogonal
direction a2 , so that the cloud becomes a narrow strip stretching along the unstable
direction. Since the slope coefficient is irrational, it covers the torus densely (Fig. 1.6
(c), (d)). This is the way the Cantor-like transverse structure of the hyperbolic at-
tractor arises in this case. One particular variant of such attractor is described and
visualized in the article (Coudene, 2006), devoted to computer illustrations of the
hyperbolic theory.
1.2.4 Plykin type attractors
The Plykin attractor takes place in a special two-dimensional map on a plane

(Plykin, 1974; Sinai, 1979; Anosov et al., 1995; Katok and Hasselblatt, 1995;
Shilnikov, 1997; Devaney, 2003). Figure 1.7 shows a region composed of three
semi-discs with small semi-circular cutouts. The sub-domains on the figure are cov-
ered with hatching indicating two fields of directions defined in this area. The map
is constructed in such a way that action on the states represented by points of the
region produces a figure shown to the right. Note that the fields of directions af-
ter the transformation coincide with the initial fields. One corresponds to direction
of stretch, and the other to direction of compression. This provides the hyperbolic
nature of the attractor.
Understanding dynamics on the Plykin attractor becomes clearer due to the fol-
lowing observation. Consider representative points belonging to a common straight
line segment of the radial hatching, or foliation, in the first figure. It is easy to see
that all of them exhibit the same behavior in the sense that under iteration of the map
they remain on a common segment of the foliation and visit simultaneously certain
sub-areas on the diagram. We may agree not to distinguish points belonging to one
and the same radial segment; then, instead of the dynamics in the two-dimensional
phase space we can restrict consideration with the one-dimensional dynamics. The
phase space for this dynamics is the so-called one-dimensional branched manifold
(Williams, 1974). It can be thought of as a rubber thread; in the present case of the
Plykin attractor it is a loop with two attached smaller loops. Initially, this thread
is stretched to three nails, as shown in Fig. 1.7(b) (left). One iteration step implies
stretching the thread and putting it again on the same nails, as shown in Fig. 1.7
(right).
Fig. 1.7 Absorbing domain and its transformation under application of the map (a) and description
of the dynamics with the one-dimensional branched manifold (b) for the original Plykin attractor.
Description of the dynamics with a help of branched manifolds was used by

Williams to develop a classification scheme for hyperbolic attractors in bounded
two-dimensional domains with three or more holes. Now it is known that many
distinct attractors of such kind exist (Williams, 1974; Zhirov, 1995; Grines and
Zhuzhoma, 2006). They are called the Plykin type attractors. One of them, shown
in Fig. 1.8, is of special interest for further consideration in the present book.
Fig. 1.8 Absorbing domain and its transformation under application of the map (a) and description
of the dynamics with the one-dimensional branched manifold (b) for an attractor of Plykin type
essential for further consideration. Symbol R indicates an unstable fixed point belonging to the
attractor, which is of importance for the further consideration.
1.3 Notion of hyperbolicity 19
Attractors of Plykin type may be regarded as being placed on a sphere rather

than on a plane. Indeed, mapping from the plane to the sphere and back may be
performed with a change of variables called stereographic projection, well known
from elementary geometry (Fig. 1.9).
Fig. 1.9 Correspondence of points A, B, C on the sphere and points A , B , C on the plane estab-
lished by the stereographic projection.
On the sphere four holes are required minimally for a uniformly hyperbolic at-
tractor of Plykin type could exist. This is consistent with the previous statements.
Actually, if the center of the stereographic projection (mapped to infinity) is placed
in one of the holes, then the transformed attractor is localized in a bounded domain
with three holes on the plane.
1.3 Notion of hyperbolicity
A good way to begin introduction to the hyperbolic theory is discussion about a

hyperbolic fixed point, the saddle. Figure 1.10 shows a pendulum and its phase
Fig. 1.10 A picture of pendulum with unstable top equilibrium (a) and phase portrait of this system
on the plane (b). S is a stable separatrix of the saddle point, and U is the unstable separatrix. Locally
near the saddle point trajectories look like hyperbolae, so it is called a hyperbolic fixed point.
portrait on the phase plane (Andronov et al., 1966; Sagdeev et al., 1988; Rabinovich
and Trubetskov, 1989). The horizontal axis corresponds to the angular coordinate,
and the vertical axis to instantaneous angular velocity. The pendulum can oscillate
near the lower equilibrium ϕ = 0, but now an unstable state of equilibrium ϕ = π
is of interest for us. In the phase plane it corresponds to a saddle point A, located at
the intersection of curves, called separatrices. From the state of equilibrium in the
presence of a small perturbation the representative point departs along the unstable
separatrix U. On the other hand, let the pendulum in the down position be kicked
with a certain properly chosen momentum, so that it reaches the top point and stops
there. In this case, the representative point moves along the stable separatrix S. The
saddle point itself is a phase trajectory, and this trajectory is hyperbolic. The name
stands for the fact that depicted on the phase plane phase trajectories near the saddle
locally are of hyperbolic form.
A slightly more complex example is a closed hyperbolic trajectory, or a saddle
periodic orbit (Fig. 1.11). It is an unstable limit cycle in the phase space, to which
trajectories approach on the surface S, or depart being situated on the surface U. If
we make a cross-section of the flow with a plane, as shown in the figure, we see two
curves formed by the intersections of the plane with these surfaces. The point where
these curves intersect each other belongs to the limit cycle, and it is a saddle point
for the Poincaré map.
Fig. 1.11 Phase space arrangement near a saddle cycle. S is a surface representing the stable man-
ifold, and U is the unstable manifold. An inset shows a picture in the Poincaré section.
The next step is generalization of the notion of hyperbolicity for non-periodic

phase trajectories (Fig. 1.12).
A phase trajectory is called hyperbolic if at each point of this orbit in the vector
space of all possible infinitesimal perturbations V one can define a subspace of vec-
tors decreasing in norm in the course of evolution forward in time and a subspace
of vectors decreasing in norm in reverse time. Moreover, the vectors belonging to
these subspaces VS , or VU must be bounded in norm by exponential functions of
time, which decrease in forward, or inverted time, respectively. In the discrete-time
case of a system governed by a diffeomorphism, all vectors in the space V must al-
low representation as linear combinations of vectors belonging to V S and VU . Math-
ematically, it means that the vector space V is a direct sum of the subspaces V S and
1.3 Notion of hyperbolicity 21
Fig. 1.12 An illustration explaining the phase space structure near a general saddle phase trajec-
tory; S is the stable manifold and U is the unstable manifold.
VU : V = VU ⊕VS . In flow systems, a one-dimensional neutral subspace V N has to be

added, corresponding to perturbations directed along the trajectory; such perturba-
tions, on average, do not grow or decrease in the course of time evolution. In this
case it is required that V = VU ⊕ VN ⊕ VS .
The epithet uniformly hyperbolic means that the rates of exponential growth of
decay of magnitudes of vectors relating to the stable and unstable manifolds are
bounded and detached from zero by some (globally defined) constants.
In the phase space a set of trajectories, which approaches the reference orbit in
the course of forward evolution in time, is called the stable manifold. Similarly, the
unstable manifold is a set of trajectories, which approaches the reference orbit in
reverse time. For hyperbolic orbits these sets are indeed manifolds, that means they
are smooth objects like curves, surfaces or hyper-surfaces in the phase space; this is
a conclusion of special theorem (known as the Hadamard-Perron theorem) (Anosov,
1967; Katok and Hasselblatt, 1995; Barreira and Pesin, 2001).
Uniformly hyperbolic saddle trajectories, and invariant sets composed of such
trajectories may occur in phase spaces of both conservative and dissipative systems,
but in this book we concentrate on the dissipative case. Hence, we will deal with
such a kind of the hyperbolic invariant sets as the uniformly hyperbolic attractors.
The uniformly hyperbolic attractor is a bounded attracting invariant set in the
phase space of a dissipative system, composed exclusively of uniformly hyperbolic
saddle trajectories, and near all these trajectories the phase space is arranged locally
in one and the same manner. Manifolds for all trajectories belonging to the attrac-
tor must have the same dimension. The intersections between stable and unstable
manifolds are allowed only at nonzero angles (touches are excluded).
For the above example of Smale-Williams attractor, the stable manifolds are two-
dimensional; they are represented by a family of meridianal plane sections of the
torus domain (Fig. 1.5). By contrast, the unstable manifolds are one-dimensional.
For each point on the attractor the unstable manifold is a fiber of the solenoid con-
taining that point. As a whole, the unstable manifold actually coincides with the
attractor.
For the DA-attractor (Fig. 1.6) the stable manifolds are one-dimensional, so are
the unstable ones. The unstable manifolds follow approximately the direction of
the expansion vector a 1 (with some deflections nears the surgery area), and stable
manifolds go along the orthogonal direction of the vector a 2 .
For attractors of Plykin type both stable and unstable manifolds are one-dimen-
sional. The stable manifolds correspond to the family of lines, along which the com-
pression of the phase volume takes place. The unstable manifolds run along the
second family of lines associated with direction of stretching for the phase volume
elements (Figs. 1.7 and 1.8).
Besides the uniform hyperbolicity we have to mention a notion of partial

hyperbolicity developed as one of the generalizations in the mathematical the-
ory (Pesin,2004; Bonatti et al., 2005; Pesin, 2007). It relates to a case when the
vector space associated with perturbations at points of the reference orbit ad-
mits representation as a direct sum of subspaces V = VU ⊕ VS ⊕ VC , where VU
is an unstable subspace (corresponding to expansion), V S is a stable subspace
(corresponding to contraction), and VC is called the central subspace. The last
one may correspond to expansion and/or contraction, but with rates in inter-
val between the minimal expansion in VU and minimal contraction in VS . In
analysis, the dynamical properties appearing due to presence of the strong
expansion and contraction may be distinguished and considered separately,
and they provide features of dynamics analogous to those of uniformly hyper-
bolic systems. Meanwhile, dynamical properties associated with the central
subspace may be of special nature for concrete systems. Although the current
book is devoted to uniform hyperbolicity, we will need to refer to the partial
hyperbolicity concerning some examples discussed in subsequent chapters.
1.4 Content and conclusions of the hyperbolic theory
Let us turn now to a brief overview of the substational part of the hyperbolic the-
ory (Anosov, 1967; Smale, 1967; Williams, 1974; Sinai, 1979; Anosov et al., 1995;
Eckmann and Ruelle, 1985; Shilnikov, 1997; Katok and Hasselblatt, 1995; Guck-
enheimer and Holmes, 1983; Afraimovich and Hsu, 2003; Hasselblatt and Katok,
2003; Devaney, 2003; Hasselblatt et al., 2008). For simplicity, we talk in this sec-
tion about discrete time systems represented by maps (diffeomorphisms). For sys-
tems with continuous time the notions and results may be extended by means of
appealing to the description in terms of the Poincaré map.
We begin with discussion of the axiom A, a statement formulated by Smale
(Smale, 1967) which allows (together with the additional requirement of strict
transversality) one to single out a class of structurally stable systems among the
dynamical systems of arbitrary finite dimension.
A point in phase space of a dynamical system is called wandering if it has a
neighborhood, into which a trajectory starting from the reference point never returns
1.4 Content and conclusions of the hyperbolic theory 23
at time exceeding some fixed constant. From a physical point of view, such dynamics
corresponds to transient motions. If this property does not hold, the point is called
non-wandering; this may be regarded as some preliminary formalization for notion
of sustained or stationary dynamical behavior (Birkhoff, 1927).
A system with discrete time, given by a diffeomorphism g, is said to obey the
Axiom A if, first, a set of non-wandering points NW is hyperbolic, and, second,
periodic points of g form a dense subset in NW. The latter means that an arbitrarily
small neighborhood of any point of the set NW necessarily contains a point related
to a periodic orbit, belonging to the same set. 5 Hyperbolicity is understood just as
explained above: the linear space of vectors of perturbations (tangent space) is com-
posed of spaces, VU and VS corresponding to exponential expansion or compression
of vectors during their time evolution. 6 Strong transversality condition requires that
the stable and unstable manifolds for an arbitrary pair of points belonging to NW
are in a generic mutual disposition, i.e. crossings are allowed only with non-zero
angles, and touches are excluded.
For the class of systems with Axiom A, Smale proved a spectral decomposition
theorem (Smale, 1967; Katok and Hasselblatt, 1995; Afraimovich and Hsu, 2003).
It asserts that the set of non-wandering points can be represented as a union of a
finite number of disjoint invariant sets B i , obeying the property of transitivity (i.e.,
each such set includes an orbit visiting an arbitrarily small neighborhood of any
point of this set). They are called basic or locally maximal sets. For various con-
crete systems, it may be stable or unstable fixed points, periodic orbits, nontrivial
attractive, repulsive, or saddle invariant sets. Definition of the local maximal set de-
clares that it is an intersection of all images of some neighborhood U containing
∞

this set under iteration of the map forward and backward in time: B = gn (U).
n=−∞
This is similar to the definition of a maximal attractor as intersection of images of
the absorbing region in iteration in forward time; such attractors surely can occur
as a particular case of the basic sets. Further, each basic set can be decomposed to
a union of a finite number k i 1 of disjoint subsets Xi, j , visited in a certain fixed
order under iteration of the map. Each set X i, j is an invariant set for the map iterated
ki times. Uniformly hyperbolic attractors of Plykin type, Smale-Williams solenoid,
DA-attractor are basic sets for certain model dynamical systems with Axiom A, for
which the number k=1 (i.e., they do not allow further decomposition). Unless oth-
erwise is not indicated specially (Part III of the present book), we will have in mind
only attractors, whose unstable manifolds are one-dimensional.
5 One particular class of systems with axiom A is represented by Anosov systems, like the Arnold
cat map; their specificity is that for them the set of non-wandering points is the whole phase space.
6 It does not exclude situations where one of the subspaces V or V is empty set; this makes it
U S
possible to incorporate simple structurally stable systems (with non-wandering sets of the type of
fixed points or periodic orbits) in the framework of the axiom A reasoning.
1.4.1 Cone criterion
Now, it should be noted that there exists a rigorous criterion of hyperbolicity formu-
lated in such a way that it may be verified in computations (Sinai, 1979; Hasselblatt
et al., 2008).
Let us assume that dynamics in discrete time is determined by a smooth invert-
ible map g(x). The criterion requires that for some choice of a constant γ >1, for
each point x relating to the analyzed invariant set one can define expanding and
contracting cones in the space of vectors of infinitesimal perturbations (the tangent
space) (Fig. 1.13). An expanding cone is a set of vectors, norms of which increase
by application of the mapping by factor γ or more. A contracting cone is a set of
vectors, norms of which grow by factor γ or more by application of the inverse map-
ping. Bearing in mind the smooth dependence of all relevant objects on the position
of the reference point in the phase space, we may talk on the fields of the expanding
and contracting cones.
Fig. 1.13 Explaining the criterion of hyperbolicity. The image of the expanding cone from a point
x is located inside the expanding cone, defined for the image point x̄, and the pre-image of the
compressing cone is situated inside the contracting cone defined at the pre-image point x̃. The
symbols Dg and Dg−1 denote the derivative matrices of the direct and inverse mappings, which
define the transformation of the perturbation vectors on one step forward and backward in time.
The criterion holds if the expanding and contracting cones are invariant in
the sense that for all points on the reference trajectory the image of the ex-
panding cone belongs to interior of the expanding cone defined for the image
point, and the pre-image of the contracting cone belongs to interior of the
contracting cone defined for the pre-image point.
In subsequent chapters we consider in some detail technique of application of

this criterion for verification of hyperbolic nature of attractors in concrete systems.
1.4.2 Instability
The fact that unstable manifold of any point on the attractor belongs to the attrac-
tor provides sensitivity of motion on the attractor in respect to small perturbations
of initial conditions, which is the main attribute of dynamical chaos. Indeed, for
two representative points, slightly shifted relative to each other along the unstable
manifold, the images will evolve in time in such a way that they are moving apart
from each other. While the distance between the images is small enough, it grows
exponentially (in average), in accordance with the relation Δx ∼ e Λn . Here n is
the discrete time, and Λ is the largest Lyapunov exponent.
If the unstable manifolds are one-dimensional, the spectrum of Lyapunov ex-
ponents contains one positive Lyapunov exponent, while all other exponents are
negative. For maps all the Lyapunov exponents of a uniformly hyperbolic attractor
are distanced from zero. Indeed, for individual trajectories on the attractor the posi-
tive exponents correspond to the perturbation vectors belonging to expanding cones
defined at the points in these trajectories. Since the minimal degree of expansion
is limited by a constant γ >1 (used in the cone criterion), the positive Lyapunov
exponent is not less than a positive number Λ +min = ln γ . This is true for all trajecto-
ries belonging to the attractor, and hence the same estimate holds for the exponent
corresponding to the attractor as a whole. Similarly, the negative Lyapunov expo-
nents, which are associated with the vectors belonging to the contracting cones, are
bounded from above by a constant Λ − max = − ln γ .
1.4.3 Transversal Cantor structure and Kaplan-Yorke dimension
As seen from considered examples, essential attribute of the chaotic attractors is

intrinsic Cantor-like transversal structure provided by the repetitive transformations
of stretching, folding, and transversal compression for the phase space volume. For
quantitative characterization of this structure, one has to use the fractal dimension.
The main idea of the definition is that a number of elements of size ε needed to
cover the set grows with decrease of ε according to the power law N ∼ ε −D , where
D is the dimension. It generalizes the familiar notion of dimension (1 for curves, 2
for surfaces, 3 for volumes); for such non-trivial sets as the strange attractors, it is
usually a non-integer number. (See Appendix D for more details.) In due time, a for-
mula was proposed to expresses the fractal dimension via the Lyapunov exponents
(Kaplan and Yorke, 1979):
m
DKY = m + Sm |λm+1 | , Sm = ∑ λi . (1.10)
i=1
Here the Lyapunov exponents λ 1 , . . . , λm , . . . , λN are enumerated in decreasing order,

and m is such an integer that the sum S m is positive while Sm+1 is already negative.
Empirically, this formula yields good results, remarkably close usually to correct
dimensions of attractors. However, mathematically it is proved only for the case

of attractors of two-dimensional maps. Then, as Solomon’s solution, the scientific
community used to regard the estimate (1.10) as a special kind of dimension called
the Kaplan-Yorke dimension, or Lyapunov dimension. Great practical advantage of
this dimension is simplicity of its evaluation, since it requires only the spectrum of
Lyapunov exponents of the attractor.
Derivation of formula (1.10) explained in Appendix D is based on assumption
of uniform contraction and expansion of phase space volume in the course of the
dynamical evolution. This is precisely what occurs in a neighborhood of a uniformly
hyperbolic attractor along the stable and unstable manifolds, respectively. Hence,
we can assume that this formula has the most reliable foundation especially for the
class of uniformly hyperbolic attractors, and it will be used widely in this book for
operative estimates of their dimensions.
1.4.4 Markov partition and symbolic dynamics
Due to intrinsic structure formed by stable and unstable manifolds in a neighbor-

hood of the uniformly hyperbolic attractor, it appears possible to construct a parti-
tion of a domain containing the attractor by a finite number of non-overlapping con-
nected sub-domains, whose boundaries follow the curves or surfaces corresponding
to these manifolds. Under some conditions (particularly, the images of boundaries
coinciding with the stable manifolds under action of the mapping must fall back
precisely on the boundaries of the same nature), it is called the Markov partition
and can be used for complete description of the trajectories belonging to the at-
tractor in terms of symbolic dynamics (Hasselblatt and Young, 2005; Sinai, 1979;
Anosov et al., 1995; Katok and Hasselblatt, 1995; Hasselblatt and Katok, 2003;
Devaney, 2003). Denoting each element of the partition with a certain symbol, a
letter of the finite alphabet, a trajectory is encoded by these characters written in a
string in order of visiting respective elements of the partition by this trajectory. For
concrete attractors, these sequences may be subjected to certain restrictions, “the
rules of grammar”, which may be visualized by means of respective graph with a
finite number of vertices and directed edges. The graph is represented formally by
an adjacency matrix whose elements are zeros and ones. A unit at the position (i, j)
indicates the presence of a directed edge connecting the i-th and the j-th vertices of
the graph, and zero means the absence of such an allowed transition.
Figure 1.14 illustrates the Markov partitions for the Smale-Williams attractor and
for the attractor of Plykin type. Also, the respective graphs are drawn, which specify
the grammar rules. In the first case the configuration is rather evident: one segment
of the border between the partition elements is a plane cross-section mapped to
itself; the other segment is a plane cross-section mapped to the first one. In the
case of Plykin type attractor the border between the partition elements consists of
a vertical line, which coincides with the stable manifold of the saddle-fixed point R
belonging to the attractor, and segments mapped to that vertical line by iteration of
the map.
Fig. 1.14 Markov partitions and graphs determining the allowed transitions for the attractor of
Smale-Williams (a) and for the Plykin type attractor (b). In the last case the border separating the
partition elements is represented by the stable manifold of the saddle-fixed point R.
As seen, the structure of these graphs implies presence of transitions from each
vertex to any other along the directed edges in a finite number of steps. Moreover
there are vertices emitting more than one edge. 7 The last means that constructing a
symbolic sequence we can select a transition from such vertex randomly from the
allowed alternatives. Any code built in such a way will correspond to an orbit on
the attractor with some definite initial conditions. It will manifest chaotic behavior
visiting the elements of the partition in accordance with the prescribed code. It may
be interpreted as the Markov chain (Feller, 1968), a random process with discrete
time and finite number of states that correspond to the symbols of the alphabet. It
justifies the term “Markov partition”.
1.4.5 Enumerating of orbits and topological entropy
The set of trajectories belonging to a hyperbolic attractor is in one-to-one correspon-

dence with the set of all possible infinite sequences composed of symbols of the
finite alphabet with certain grammar rules. It has the cardinality of continuum. The
7 Adjacency matrix in this case has a property of transitivity. It means that for some power of
this matrix all elements are nonzero. This is the case of the attractor having a single transitive
component.
set of periodic orbits on the attractor corresponds to periodic symbolic sequences

and has the cardinality of an infinite countable set. With the increase of period T
the number of periodic orbits of N grows exponentially as N ∼ e hT . The character-
istic index h is called the topological entropy (Adler et al., 1965; Beck and Schlogl,
1993). It is interpreted as a quantity characterizing complexity of the set of orbits on
the attractor.
It is clear that the possibility of complete symbolic description of trajectories on a
uniformly hyperbolic attractor sustains a fundamental relationship of this issue with
the information theory and coding theory. This may be of importance from the point
of view of possible exploitation of systems with the uniformly hyperbolic attractors
in information and communication applications.
1.4.6 Structural stability
Structural stability of a uniformly hyperbolic attractor occurs due to the fact that the
expansions and contractions of elements of phase volume always take place in di-
rections, which have a non-zero angle between them (Smale, 1967; Williams, 1974;
Sinai, 1979; Eckmann and Ruelle, 1985; Anosov et al., 1995; Katok and Hassel-
blatt, 1995; Shilnikov, 1997; Guckenheimer and Holmes, 1983; Afraimovich and
Hsu, 2003; Devaney, 2003; Hasselblatt and Young, 2005). In terms of stable and
unstable manifolds, it corresponds to their transversal (without touch) relative po-
sition at the intersections. If you slightly vary parameters, or functions used in the
definition of the evolution operator, then, due to the transversal intersection of the
manifolds, this perturbation does not destroy the associated topology of the phase
space structure, at least while the perturbation is not too large. The same relates to
a set of curves and surfaces used in the construction of Markov partitions. It means
that the symbolic dynamics remains unchanged.
A more accurate formulation is as follows: While the considered perturbations
are small in the class of continuous functions with first derivatives (of class C 1 ),
the system allows reduction to the original form by a homeomorphism that is a
continuous invertible change of variables (of class C 0 ), with the reversal variable
change being continuous too. In other words, the perturbed system is topologically
equivalent to the original one.
The Lyapunov exponents are obtained from the variation equations derived by
linearization of the original equations near the reference trajectory, which implies
the use of operation of differentiation. Therefore, under the change of variables,
which may be not smooth, the Lyapunov exponents, generally speaking, are not
invariant. Yet, due to the structural stability, with respect to their behavior under
small variations of the system one can draw certain conclusions. Until the attractor
remains uniformly hyperbolic, retaining the intrinsic topological structure of stable
and unstable manifolds, the number of positive and negative exponents can not be
changed. Moreover, those and others definitely remain distant from zero (we speak
about maps here). Apparently, even stronger statement has to be valid that for a
small variation of the system, at least the positive Lyapunov exponent also varies
weakly. This assertion is confirmed by numerical data for models with uniformly
hyperbolic attractors, discussed in subsequent chapters, but a mathematical result,
to which in this context one could refer, is unknown to the author.
1.4.7 Invariant measure of Sinai-Ruelle-Bowen
For the uniformly hyperbolic attractors there exists an absolutely continuous invari-
ant measure of Sinai-Ruelle-Bowen (Sinai, 1972; Bowen, 1975; Ruelle, 1976). An
abbreviation, the SRB-measure, is commonly used.
Let us turn again to interpretation of the dynamics as evolution of the cloud of
representative points in the phase space. Assuming that initially the ensemble is
characterized by a “good” distribution function (continuous and decreasing rapidly
on the sides), let us ask, what will be the nature of the distribution at large times? It
is easy to see that even in the simplest cases the arising distributions are singular; for
example, for a stable attractive fixed point we have in the limit t → ∞ a distribution
represented by the Dirac delta function. A generalization embracing both smooth
and singular distributions is delivered by a mathematical construct called a mea-
sure. This may be regarded as a generalization of familiar notions of length, area,
and volume, which relate to simple one-, two-, and three-dimensional objects. In the
context of dynamical systems, to define a measure means to prescribe non-negative
numbers to subsets of phase space (not to any subset, but to those belonging to a suf-
ficiently rich class of measurable subsets), which satisfy some natural requirements,
say, a union of disjoint measurable sets must have a measure equal to the sum of the
measures of its elements.
Let us consider a phase trajectory on the attractor and agree to prescribe to a
domain of the phase space a measure equal to the relative duration of residence of
the representative point in this domain as the observation time tends to infinity. It is
called the natural invariant measure. For uniformly hyperbolic attractors a measure
defined in such a way for any typical orbit relating to the attractor appears to be one
and the same, and is called a measure of Sinai-Ruelle-Bowen. This measure may be
thought of as a distribution of the phase fluid on the fibers of the attractor, i.e. along
the unstable manifolds. Although across the structure of the fibers the distribution is
of singular nature, it is described by a continuous function of coordinate measured
along each fiber. In other words, the phase fluid substance is distributed on the fibers
in a smooth manner, without pronounced local concentrations.
One more remark concerns relation of the Sinai-Ruelle-Bowen measure to sys-
tems subjected by noise.
In the presence of noise, instead of singular distributions we get the smooth ones,
say, near a stable fixed point, instead of the Dirac delta function we will observe
a finite-width Gauss-like distribution. Now, the assertion is that the distributions
arising in systems with uniformly hyperbolic attractors under effect of small noise,
in the limit of zero intensity of the noise tend precisely to those associated with the
Sinai-Ruelle-Bowen measures. So, these measures may be regarded as physically

relevant.
1.4.8 Shadowing and effect of noise
The trajectories on uniformly hyperbolic attractors possess a remarkable property of

shadowing (Guckenheimer and Holmes, 1983; Katok and Hasselblatt, 1995; Palmer,
2009). It refers to relationship between a true trajectory on the attractor and an ap-
proximate obtained, say, in the presence of small noise of bounded magnitude, or
in computations with round-off at finite accuracy. The true orbit shadows a noisy
trajectory in the sense that it remains close to it for arbitrarily long time. For the dy-
namics in the presence of noise this result was proven in due time by Kiefer (Kiefer,
1974). It shows that the stochastic behavior of dynamical origin is dominating in a
sense over the stochastic motion caused by small random external force.
1.4.9 Ergodicity and mixing
Motion on a uniformly hyperbolic attractor (with one transitive component) is char-

acterized by the properties of ergodicity and mixing (Sinai, 1979; Eckmann and
Ruelle, 1985; Anosov et al., 1995; Katok and Hasselblatt, 1995; Shilnikov, 1997;
Guckenheimer and Holmes, 1983; Afraimovich and Hsu, 2003; Devaney, 2003;
Hasselblatt and Young, 2005). Ergodicity means that a typical trajectory on the at-
tractor visits during the time evolution any neighborhood of any point on the at-
tractor. It corresponds to equivalence of averaging over time and averaging over
invariant measure and provides a statistical approach to the analysis of sustained
regimes of dynamics. The mixing property is stronger. It means that for any element
of phase volume for a sufficiently long time the corresponding cloud of representa-
tive points will be distributed throughout the attractor. The mixing allows describing
statistically the process of approach of the ensemble to a state corresponding to the
invariant distribution. The property of mixing is also associated with decay of cor-
relations: the correlation function, calculated for the signal generated by dynamics
of the discrete-time system on a uniformly hyperbolic attractor, is characterized by
exponential decay. For continuous-time systems representing suspension of diffeo-
morphisms with uniformly hyperbolic attractors, the question of mixing and decay
of correlations requires a special analysis in each case, as it depends on the nature
of the distribution for times of returns to the Poincaré section.
References 31
1.4.10 Kolmogorov-Sinai entropy
Consider a set of all “words” occurring in symbolic representations of trajectories

on the attractor and containing n characters. For each i-th word one can determine a
probability of its occurrence p i . This value can be interpreted as a measure of Sinai-
Ruelle-Bowen for a set of points in phase space, which correspond to the trajectory,
having visited the elements of the Markov partition marked with the respective sym-
bols in the previous n steps. Now, let us define a sum S n = ∑ pi log pi over all words
admissible by the grammar, divide it by n, and consider the limit n → ∞. The quan-
tity hKS = − lim Sn /n is called the metric entropy or entropy of Kolmogorov-Sinai
n→∞
(Kolmogorov, 1959; Sinai, 1959). As seen from the definition, it can be interpreted
as amount of information generated by the dynamics on the attractor per unit of time.
What is essential, the definition is appealing to an invariant measure associated with
the attractor.
Positive metric entropy is a criterion of chaotic nature of the dynamics on the
attractor. The metric entropy h KS and the topological entropy h are related with
an inequality h h KS . The result proved by Pesin relates spectrum of Lyapunov
exponents and the Kolmogorov-Sinai entropy (Pesin, 1977). In general case, a sum
of all positive Lyapunov exponents is the upper estimate for the metric entropy. For
the uniformly hyperbolic attractors, the case of equality holds in the Pesin formula.
References
Adler, R.L., Konheim, A.G., McAndrew, M.H.: Topological entropy. Trans. Amer. Math. Soc. 114,
309–319 (1965).
Afraimovich, V., Hsu, S.-B.: Lectures on Chaotic Dynamical Systems. International Press,
Somerville, MA (2003).
Andronov, A., Pontrjagin, L.: Systèmes Grossiers. Dokl. Akad. Nauk. SSSR 14, 247–251 (1937).
Andronov, A.A., Vitt, A.A., Khaı̌kin, S.Ė: Theory of oscillators. Pergamon Press (1966).
Anishchenko, V.S.: Nonlinear Dynamics of Chaotic and Stochastic Systems: Tutorial and Modern
Developments. Springer, Heidelberg (2002).
Anishchenko, V.S., Astakhov, V., Vadivasova, T., Neiman, A., Schimansky-Geier, L.: Nonlinear
Dynamics of Chaotic and Stochastic Systems. Tutorial and Modern Development. Springer,
Berlin, Heidelberg (2007).
Anosov, D.V.: Geodesic flows on closed Riemannian manifolds of negative curvature. Trudy Mat.
Inst. Steklov 90, 3–210 (1967).
Anosov, D.V., Gould, G.G., Aranson, S.K., Grines, V.Z., Plykin, R.V., Safonov, A.V., Sataev,
E.A., Shlyachkov, S.V., Solodov, V.V., Starkov, A.N., Stepin, A.M.: Dynamical Systems IX:
Dynamical Systems with Hyperbolic Behaviour (Encyclopaedia of Mathematical Sciences) (v.
9). Springer, New York (1995).
Arnold, V.I.: Ordinary Differential Equations. The MIT Press, Somerville, MA (1978).
Arnold, V.I.: Mathematical Methods of Classical Mechanics. Springer-Verlag, Heidelberg (1989).
Arnold, V.I., Avez, A.: Ergodic Problems in Classical Mechanics. Benjamin, New York (1968).
Barreira, L, Pesin, Y.: Lectures on Lyapunov exponents and smooth ergodic theory. In: Smooth
Ergodic Theory and Its Applications. Proceedings of Symposia in Pure Mathematics, pp. 3–90.
AMS (2001).
Beck, C., Schlogl, F.: Thermodynamics of Chaotic Systems: An Introduction. Cambridge Univer-
sity Press, Cambridge (1993).
Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.-M.: Lyapunov characteristic exponents for
smooth dynamical systems and for Hamiltonian systems: A method for computing all of them.
Meccanica 15, 9–30 (1980).
Birkhoff, G.D.: Dynamical Systems. American Mathematical Society, Provindence RI (1927).
Bonatti, C., Diaz, L.J., Viana, M.: Dynamics Beyond Uniform Hyperbolicity. A Global Geomet-
ric and Probobalistic Perspective. Encyclopedia of Mathematical Sciences. Vol.102. Springer,
Berlin, Heidelberg, New York (2005).
Bowen R.: Equilibrium States and Ergodic Theory of Anosov Diffeomorphisms. Lecture Notes in
Math. 470, Springer, Heidelberg (1975).
Coudene, Y.: Pictures of hyperbolic dynamical systems. Notices of the American Mathematical
Society 53, 8–13 (2006).
Devaney, R.L.: An Introduction to Chaotic Dynamical Systems. Westview Press, New York (2003).
Eckmann, J.-P., Ruelle, D.: Ergodic theory of chaos and strange attractors. Rev. Mod. Phys. 57,
617–656 (1985).
Feller, W.: An Introduction to Probability Theory and Its Applications 1. Wiley, New York (1968).
Greene, J.M., Kim, J.-S.: The calculation of Lyapunov spectra. Physica D 24, 213–225 (1987).
Grines, V.Z., Zhuzhoma, E.V.: Expanding attractors. Regular and Chaotic Dynamics 11 (2), 225–
246 (2006).
Guckenheimer, J., Holmes, P.: Nonlinear Oscillations, Dynamical Systems, and Bifurcations of
Vector Fields. Springer-Verlag, New York (1983).
Hasselblatt, B., Katok, A.A.: First Course in Dynamics: with a Panorama of Recent Developments.
Cambridge University Press, Cambridge (2003).
Hasselblatt, B., Pesin, Y., Schmeling, J.: Pointwise Hyperbolicity Implies Uniform Hyperbolicity.
Preprint ESI 2067, Vienna (2008).
Hasselblatt, B., Young, L.S.: Anosov and axiom a systems. In: Scott, A.(ed) Encyclopedia of
Nonlinear Science, pp. 11–13. Routledge, New York (2005).
Hénon, M.: A two-dimensional mapping with a strange attractor. Commun. Math. Phys. 50, 69–77
(1976).
Hilborn, R.C.: Chaos and Nonlinear Dynamics. Oxford University Press, Cambridge (2000).
Kaplan, J.L., Yorke, J.A.: A chaotic behavior of multi-dimensional differential equations. In: Peit-
gen, H.-O. and Walther, H.-O. (eds.) Functional Differential Equations and Approximations of
Fixed Points. Lecture Notes in Mathematics, 730, pp. 204–227. Springer, Berlin, New York
(1979).
Katok, A., Hasselblatt, B.: Introduction to the Modern Theory of Dynamical Systems. Cambridge
University Press, Cambridge (1995).
Kifer, Yu.I.: On small random perturbations of some smooth dynamical systems. Math. USSR
Ivestija 8, 1083–1107 (1974).
Kolmogorov, A.N.: Entropy per unit time as a metric invariant of automorphism. Doklady Acad.
Sci. SSSR 124, 754–755 (1959).
Kuznetsov, Yu.A.: Elements of Applied Bifurcation Theory. Springer, New York, Berlin, Heidel-
berg (1998).
Landau, L.D., Lifshitz, E.M.: Mechanics (Course of Theoretical Physics, vol.1). Butterworth-
Heinemann , Oxford (1976).
Manneville, P.: Dissipative structures and weak turbulence. In: Chaos — The Interplay Between
Stochastic and Deterministic Behaviour. Lecture Notes in Physics 457, 257–272 (1995).
Milnor, J.: On the concept of attractor. Communications in Mathematical Physics, 99, 177–195
(1985).
Newhouse, S.E.: The abundance of wild hyperbolic sets and non-smooth stable sets for diffeomor-
phisms. Publications Mathématiques de L’IHÉS 50, 101–151 (1979).
Oseledets, V.I.: A multiplicative ergodic theorem: Lyapunov characteristic numbers for dynamical
systems. Trans. Moscow Math. Soc. 19, 197–231 (1968).
Ott, E.: Chaos in dynamical systems. Cambridge University Press, Cambridge (2002).
References 33
Palis, J.: On Morse-Smale dynamical systems. Topology 8, 385–405 (1969).

Palmer, K. J.: Shadowing lemma for flows. Scholarpedia 4(4), 7918 (2009).
Pesin, Y.B.: Characteristic Lyapunov exponents and smooth ergodic theory. Russian Math. Surveys
32, 55–114 (1977).
Pesin, Ya.: Existence and genericity problems for dynamical systems with nonzero Lyapunov ex-
ponents. Regular and Chaotic Dynamics 12, 476–489 (2007).
Pesin, Ya.B.: Lectures on Partial Hyperbolicity and Stable Ergodicity. Zurich lectures in advanced
mathematics. European Mathematical Society (2004).
Plykin, R.V.: Sources and sinks of A-diffeomorphisms of surfaces. Math. USSR Sb. 23, 233–253
(1974).
Rabinovich, M.I., Trubetskov, D.I.: Oscillations and Waves: In Linear and Nonlinear Systems,
Kluwer Academic Publication, The Netherlands (1989).
Raghunathan, M.S. A proof of Oseledec’s multiplicative ergodic theorem. Israel Journal of Math-
ematics 32 (4), 356–362 (1979)
Ruelle, D.: A measure associated with Axiom A attractors. Amer. J. Math. 98, 619–654 (1976).
Ruelle, D., Takens, F.: On the nature of turbulence Commun. Math. Phys. 20, 167–192 (1971).
Sagdeev, R.Z., Usikov, D.A., Zaslavsky, G.M.: Nonlinear physics: from the pendulum to turbulence
and chaos. CRC Press, Boca Raton, FLA (1988).
Schuster, H.G., Just, W. Deterministic Chaos: An Introduction. Wiley-VCH Verlag GmbH & Co.
KGaA, Weinheim, FRG (2005) .
Shilnikov, L.: Mathematical problems of nonlinear dynamics: a tutorial. Int. J. of Bifurcation and
Chaos 7, 1353–2001 (1997).
Shilnikov, L.P., Shilnikov, A.L., Turaev, D.V., Chua, L.O.: Methods of Qualitative Theory in Non-
linear Dynamics (Part 1). World Scientific, Singapore (1998).
Shilnikov, A.L., Turaev, D.V., Chua, L.O., Shilnikov, L.P.: Methods of Qualitative Theory in Non-
linear Dynamics (Part 2). World Scientific, Singapore (2002).
Sinai, Y.G.: On the Notion of Entropy of a Dynamical System. Doklady Acad. Sci. SSSR 124,
768–771 (1959).
Sinai, Y.G.: Gibbs measures in ergodic theory. Russ. Math. Surv. 27, 21–70 (1972).
Sinai, Y.G.: Stochasticity of dynamical systems. In: Gaponov-Grekhov, A.V. (ed.) Nolinear waves,
pp.192–212. Moscow, Nauka (1979).
Schuster, H.G. and Just, W. Deterministic chaos: an introduction. Wiley-VCH (2005).
Smale, S.: Differentiable dynamical systems. Bull. Amer. Math. Soc. (NS) 73, 747–817 (1967).
Strogatz, S.H.: Nonlinear Dynamics and Chaos: with Applications to Physics, Biology, Chemistry,
and Engineering. Westview Press, MA (2001).
Tél, T. and Gruiz, M.: Chaotic dynamics: an Introduction Based on Classical Mechanics. Cam-
bridge University Press, Cambridge (2006).
Thompson, J.M.T., Stewart, H.B.: Nonlinear Dynamics and Chaos. Wiley, New York (1986).
Vul, E.B., Sinai, Y.G., Khanin, K.M.: Feigenbaum universality and the thermodynamic formalism.
Russ. Math. Surv. 39, 1–40 (1984).
Williams, R.F.: Expanding attractors. Publications mathématiques de l’I.H.É.S. 43, 169–203
(1974).
Zhirov, A.Yu.: Hyperbolic attractors of diffeomorphisms of orientable surfaces. Russ. Ac. Sci. Sb.
Math. 83, 23–66 (1995).
Chapter 2
Possible Occurrence of Hyperbolic Attractors
Abstract In this chapter we review some situations considered or mentioned in lit-

erature that may have relation to occurrence of uniformly hyperbolic attractors in
dynamical systems represented by maps or differential equations. To date, the ap-
proaches under discussion are developed to different extents. Only few of them are
elaborated up to such degree that may allow illustrating on a level of numerical
computations. Nevertheless, the collected material is of undoubted significance for
prospects of search for real-world systems with uniformly hyperbolic attractors or
design of physical and technical devices operating due to the presence of such at-
tractors.
2.1 The Newhouse-Ruelle-Takens theorem and its relation to the

uniformly hyperbolic attractors
In the early 1970’s, Ruelle and Takens advanced their scenario of origin of turbu-
lence. They argued that under variation of a control parameter (the dimensionless
Reynolds number in hydrodynamics) the transition to chaos happens after birth of a
relatively small number of oscillatory components and is associated with emergence
of a strange attractor (Ruelle and Takens, 1971). Detailed description and critical
evaluation of this concept go far beyond the scope of the present review. However,
it is worth emphasizing that Ruelle and Takens in their formulations meant specif-
ically and definitely a uniformly hyperbolic attractor (not a strange attractor in a
broad sense, as this term is commonly adopted now). So, it is interesting to ask to
what extent their argumentation can be useful for design of concrete examples of
dynamical systems with such attractors.
The basic theorem of Newhouse, Ruelle and Takens (Newhouse et al., 1978)
establishes a possibility of occurrence of chaotic uniformly hyperbolic attractors for
autonomous differential equations obtained by arbitrarily small perturbations from
equations for quasiperiodic motions on tori of dimension n 3 or n 4.
36 2 Possible Occurrence of Hyperbolic Attractors
Here we concentrate on their intermediate result. The formulated theorem relates

to three cases: for diffeomorphisms (i) on a two-dimensional compact manifold, (ii)
on a two-dimensional torus, and (iii) on a compact manifold of dimension m 3. In
case (i) the theorem says that a map having a uniformly hyperbolic attractor can be
obtained by a perturbation of the identity map with a variation arbitrarily small in
the class of functions with first derivatives C 1 . In case (ii) a similar assertion holds,
but the perturbation is small in the class of functions with two derivatives C 2 , and in
case (iii) it relates to the class of infinitely differentiable functions C ∞ .
As a starting point for reasoning, the authors refer to existence of a two-
dimensional map F(X), which has a uniformly hyperbolic chaotic attractor in some
bounded domain. This map is supposed to allow representation as a composition of
a large number of maps close to the identity map:
(N−1) (N−2) (0)
F(X) = F1/N ◦ F1/N ◦ · · · ◦ F1/N (X). (2.1)
Here N is a large integer, and

(k) (k)
F1/N (X) = X + N −1F̃1/N (X). (2.2)
In case (i) the authors suggest to consider a ring area of width 2π /N partitioned
into N sectors (Fig. 2.1(a)). By a shift and rescaling, one can place the domain of
localization of the attractor of F(X) in the sector k=0. Next, a mapping of the ring
into itself is constructed to map each sector of number k into the sector k+1. This
(k)
map is obtained from F 1/N by means of smooth variable changes for images and
pre-images, with rescaling by factor N. Besides this, certain conditions are provided
to arrange smooth matching at the boundaries of the sectors. The resulting map pos-
sesses a uniformly hyperbolic attractor. Indeed, after N iterations, an orbit starting
from the sector k=0 after visiting all other sectors returns to the original one, and the
respective composite mapping is equivalent to the initial map F (up to the variable
change).
The map from the k-th to the k+1-th sector contains, besides the trivial part, an ad-
(k)
ditive perturbation term, which after the rescaling looks like N −2 F̃1/N (N ξ ), where
ξ designates the state vector transformed by the variable change. The perturbation
is of order N −2 , and its derivative is of order N −1 . Hence, by increase of N one
can make the perturbation arbitrarily small together with its derivative. So, the map
defined on the ring becomes arbitrarily close to the identity in the class of functions
C1 as N is large enough, while the N-fold composition is equivalent to the map F
that has the uniformly hyperbolic attractor.
For case (ii) consider a surface of two-dimensional torus partitioned into N = M 2
areas, small “squares” (Fig. 2.1(b)). Again, let us design a map, which ensures an
orbit to visit successively all the elements of the partition, in such a way that after
bypassing of all the elements (that requires M 2 steps) the resulting transformation
is equivalent to the map (2.1). Thus, we represent the map equivalent to F as a com-
position of M 2 mappings, which contain the additive perturbation terms expressed
2.2 Lorenz model and its modifications 37
Fig. 2.1 Geometric illustrations for constructions considered in cases (i), (ii), and (iii) of the theo-
rem of Ruelle, Takens, and Newhouse; see the panels (a), (b), and (c), respectively.
(k)
in the rescaled variables as M −3 F̃1/M2 (M ξ ), k = 0, 1, . . . , M 2 − 1. The perturbation
is of order M −3 , its first and second derivatives are of orders M −2 and M −1 , respec-
tively. Increasing M we get the perturbation arbitrarily small in the class of functions
C2 .
Finally, we turn to the situation (iii). Restricting discussion for simplicity with the
three-dimensional case, m=3, let us consider a torus subdivided by cross-sections
with meridianal planes into N equal sectors (Fig. 2.1(c)). Assume that the dynam-
ics along the azimuthal coordinate is governed by a one-dimensional map, which
transforms each sector to the neighboring one. The map for the rest coordinates is
supposed to be of form (2.1), where X specifies a position of the representative point
in the transversal cross-section. Under N-fold application of the map an orbit arrives
again in the original sector it started from, and the resulting mapping is equivalent
to F in respect to variables associated with the coordinates in the transversal cross-
section.
In the last case the rescaling of the dynamical variables is not needed, so, in the
(k)
limit N → ∞ the additive perturbation term N −1 F̃1/N tends to zero together with all
derivatives, i.e. becomes arbitrarily small in the class of functions C ∞ .
What can be said about potential application of the above considerations for prac-
tical design of physically realizable models with uniformly hyperbolic attractors?
Certainly, such prospects seem very doubtful. As the reasoning starts from a given
map with the uniformly hyperbolic attractor, it can not help, obviously, in finding
out concrete examples of such maps. Moreover, the mathematical constructions, at
least, in cases (i) and (ii), look absolutely non-physical, because of involvement of
functions of extremely small characteristic scale in the phase space in the definitions
of the dynamical systems.
2.2 Lorenz model and its modifications
One of the most famous systems in the context of nonlinear dynamics and chaos is
the Lorenz model (Lorenz, 1963; Afraimovich et al., 1977; Sparrow, 1982; Guck-
enheimer and Holmes, 1983; Alligood et al., 1996; Ott, 2002; Schuster and Just,
2005)
ẋ = σ (y − x), ẏ = rx − y − xz, ż = −bz + xy. (2.3)
Here x, y, z are dynamical variables, and σ , r, b are parameters. Such equations orig-
inate in the context of convection in the layer of fluid heated from bottom, in the
theory of single-mode laser, in a model of water wheel, in the dissipative oscillator
with inertial excitation and in some other problems.
In the work published in 1963 Lorenz reported observation of non-periodic sus-
tained oscillations in the numerical solution of these equations, and interpreted these
regimes as relating to the own complex dynamics of the system. Particularly, the au-
thor emphasized importance of sensitivity of the trajectories to small perturbations
of initial conditions. Attractor of the Lorenz system is shown in Fig. 2.2. So far,
due to deep and pervasive analysis made by many authors, the dynamical nature
of chaos in the Lorenz model has been studied exhaustively. An important final
advance in the long-term explorations is a computer-assisted proof published by
Tucker (Tucker, 2002).
Historically, an important step in explaining many qualitative aspects of chaotic
dynamics in the Lorenz model was a construction of a simplified model, the so-
called geometric Lorenz attractor (Sparrow, 1982; Guckenheimer and Holmes,
1983) (Fig. 2.2(b)).
Fig. 2.2 Portrait of the Lorenz attractor obtained from numerical solution of Eq. (2.3) at r = 28,
σ = 10, b = 8/3 (a), the geometric Lorenz model (b), and illustration of the transverse structure
of the attractor, the “Cantor book” (c).
For organizing structure of the flow in the phase space of the Lorenz model a
saddle fixed point O at the origin is important. This point is characterized by one
positive (unstable) and two negative (stable) eigenvalues of linearized equations that
2.2 Lorenz model and its modifications 39
means it has one-dimensional unstable manifold and two-dimensional stable mani-

fold. This point belongs to the attractor together with its one-dimensional unstable
manifold. In domain D, shown in gray in Fig. 2.2(b), the flow is defined as that
corresponding to linearization of the Lorenz system near the origin. By this flow
F0 two halves of the “ceiling” C of the domain D are mapped to two curvilinear
triangles on the sides of the domain. Further, by a flow defined in outside of D by
some field F these sides are transformed to two narrow elongated triangles inside
the rectangle C. It corresponds to one step of the Poincaré map. In the Poincaré sec-
tion the stretching is directed along these elongated triangles, and the compression
takes place across them. The construction is built in such a way that the direction
of the compression remains invariant under the action of the map. Attractor is not
uniformly hyperbolic because of specific behavior of the trajectories near the singu-
lar orbit that is the unstable manifold of the point O. In the Poincaré cross-section
the singular orbit corresponds to tips of the triangular regions. With decrease of
distance from a certain trajectory to the singular one, the return time tends to infin-
ity. The Cantor-like transversal structure of the attractor near the singular trajectory
is arranged in a way illustrated in Fig. 2.2(c) that is the so-called “Cantor book”.
Attractor in the geometric model is classified as quasi-hyperbolic or singular hy-
perbolic. It does not possess structural stability in the usual sense. Markov partition
can be constructed, but it requires an infinite alphabet. Following the Tucker result
(Tucker, 2002), attractor in the original Lorenz equations (2.3) is of the same nature.
Now, we wonder whether it is possible to modify the system in such a way that it
would have a uniformly hyperbolic attractor. A positive answer is given by Morales,
though only on the level of the geometric constructions (Morales, 1996).
Let us turn to a model, in which the fixed point O undergoes a kind of saddle-
node bifurcation, and in other respects the system is similar to the geometric Lorenz
model. In parameter range before the bifurcation, there is a pair of unstable fixed
points. One of them is analogous to the fixed point in the Lorentz model and has
one unstable and two stable directions, while the second fixed point has one stable
and two unstable directions. As the parameter is varied, both fixed points move to
meet each other at the bifurcation, and then annihilate. In the parameter range of
coexistence of the pair of the fixed points the system has attractor of the Lorenz
type. In the parameter range where the pair has disappeared, one of two options
occurs, depending on the arrangement of the vector field in the remote parts of the
phase space, through which the trajectory travels. The first variant is emergence
of non-hyperbolic dynamics (somewhat like that in the Hénon map), and the second
corresponds to appearance of attractor of Plykin type in the Poincaré map. In Fig. 2.3
the pictures in the Poincaré map are shown for the usual Lorenz attractor (a), and
for the two cases indicated by Morales, (b) and (c). The last two are depicted for the
dynamics arising after the bifurcation of merging and disappearance of the pair of
the fixed points.
Possibility of occurrence of Plykin-type attractors in the modified Lorenz model
seems interesting, and it may be significant for elaboration of physical examples, be-
cause the Lorenz model itself certainly has relation to a number of concrete systems
Fig. 2.3 Qualitative illustrations of the action of the Poincaré map for the original geometric
Lorenz attractor (a) and for the modified model in the work of Morales (b and c). The variant
(b) corresponds to the emergent dynamics of Hénon type, and the variant (c) to the formation of
the Plykin type attractor.
of different nature. However, so far no explicit examples of systems with uniformly

hyperbolic chaos have been designed on this base.
2.3 Some maps with uniformly hyperbolic attractors
It is easy to write down a map possessing attractor of Smale-Williams type; it may

be found in many textbooks (e.g. (Katok and Hasselblatt, 1995)). One can derive
the map using variables (ϕ , X,Y ) of the following geometric sense. Let X and Y be
Cartesian coordinates of a point in a meridianal plane section of a torus assuming
X 2 + Y 2 1, and ϕ is an angular coordinate specifying the given section. Coordi-
nates in the three-dimensional phase space are expressed as
x = (R + X) cos ϕ , y = (R + X) sin ϕ , z = Y, (2.4)
where R is an additional parameter concretizing the geometry. Now, we define the

map on one step of discrete time as
3 1 3 1
ϕn+1 = 2ϕn , Xn+1 = Xn + cos ϕn , Yn+1 = Yn + sin ϕn .
10 2 10 2
It may be reformulated in terms of Cartesian coordinates (x,y,z) as follows:

7 3 1 xn x2n − y2n
xn+1 = R+ xn + yn +
2 2 ,
10 10 2 x2n + y2n x2n + y2n

7 3 1 xn 2xn yn
yn+1 = R+ xn + yn +
2 2 , (2.5)
10 10 2 xn + yn x2n + y2n
2 2
3 1 yn
zn+1 = zn + .
10 2 x2n + y2n
2.3 Some maps with uniformly hyperbolic attractors 41
Fig. 2.4 Portrait of attractor of Smale-Williams type obtained from iterations of the map (2.5) at
R = 2 drawn on a background of the toroidal absorbing domain shown in gray.
Figure 2.4 shows a toroidal absorbing domain with a border expressed by the
relation X 2 + Y 2 = 1 in the original variables, and portrait of the attractor obtained
from iterations of the map (2.5) at R = 2. Observe visible transversal Cantor-like
structure intrinsic to the Smale-Williams solenoid. The Lyapunov exponents for
this attractor are Λ1 = ln 2 = 0.6931 · · ·, Λ 2,3 = ln(3/10) = −1.2039 · · ·, because
one step of the iterations is accompanied by twofold stretch along the filaments of
the attractor, i.e. along the angular coordinate ϕ , and by compression along two
transversal directions by factor 3/10. The Kaplan-Yorke dimension of this attractor
equals DKY = 1 + Λ1/|Λ2 | ≈ 1.576.
Unfortunately, topologically this is not a kind of attractor that could occur

in Poincaré maps of flow systems (Katok and Hasselblatt, 1995).
The geometric construction truly associated with the map (2.5) may be
thought of as follows. Take a closed tube representing the surface of the torus,
cut it at some place by a meridianal plane, then stretch the tube in length,
compress it in width, fold in double loop, and glue the edges accurately at the
place of the former cut (Ilyashenko, 2005).
If trying to perform the torus deformation continuously, without the cut
(say, with a real rubber tube), one can observe that the tube inevitably un-
dergoes additional azimuthal torsion. It means that we should glue the edges
not directly, but with their relative rotation by the angle of 2π . Hence, the
map (2.5) is not appropriate as starting point for constructing physical flow
systems suspending the Smale-Williams attractor, and alternative approaches
must be developed.
Let us consider now an explicit example of a discrete-time system with DA-

attractor on a torus (Kuznetsov, 2009). To start, we notice that the Arnold cat map
(1.9) can be decomposed to two successively applied maps, which may be inter-
preted naturally as corresponding to two halves of a time step. Namely, if
pn+ 1 = pn + qn, qn+ 1 = qn , (2.6)

2 2
and
pn+1 = pn+ 1 , qn+1 = pn+ 1 + qn+ 1 , (2.7)

2 2 2
then, on the whole step we get p n+1 = pn + qn, qn+1 = pn + 2qn. The symbols mod 1
are omitted for brevity.
As alternative to the Smale “surgery”, let us perform the DA modification adding
some analytical functions into the right-hand parts of (2.6) and (2.7). On the first half
of the step the additional term is introduced in the first equation (2.6), as a function
depending only on x. To preserve the fixed point in the origin, this function must
be zero at p=0, and we assume the function to be odd. Then, as the dynamics takes
place on the torus, it is natural to require this function to be of period 1 and represent
it by the Fourier series. Accounting only the first and the second harmonics, we
select their amplitudes in such a way that the function is as close to zero as possible
1
at the middle of the unit interval; so, we select it as sin 2π p + sin 4π p. On the
2
second half-step analogous modification is performed in symmetric way by adding
the function of q in the second equation. Combining both half-steps, we arrive at the
map

1
pn+1 = pn + qn + ε sin 2π pn + sin 4π pn 2π ( mod 1),
2

1
qn+1 = pn + 2qn + ε sin 2π pn + sin 4π pn + sin2π qn (2.8)
2

1
+ sin 4π qn 2π ( mod 1).
2
Here ε characterizes relative value of the added terms. The Jacobin matrix of the
map (2.8)

1 + ε (cos2π p + cos4π p) 1
M(p, q) = (2.9)
1 + ε (cos2π p + cos4π p) 2 + ε cos 2π q + ε cos 4π q
8
has determinant strongly positive and detached from zero globally while ε < .
9
Hence, under this condition the mapping remains invertible (the diffeomorphism).
The map (2.8) has a fixed point in the origin, and for the matrix of linearization
at this point M(0,0) the eigenvalues called the multipliers are

1 5
μ1,2 = (3 + 4ε ) ± + 2ε . (2.10)
2 4
1 √ 1 √
At ε =0 we have μ1 = (3 + 5) > 1 and μ2 = (3 − 5) < 1, that is a saddle. With
2 2
1
increase of ε both the multipliers grow, and the second one exceeds 1 at ε > ε 1 = .
2
2.4 From DA to the Plykin type attractor 43
It corresponds to the situation that the fixed point accepts the properties required
1 8
from the Smale surgery. So, we have to select < ε < .
2 9
Computations indicate that attractor of the map (2.8) is uniformly hyperbolic,
e.g. at ε = 0.7. In Fig. 2.5 portraits of the attractor are shown in the unit square (a)
and on the surface of torus (b). Observe transversal Cantor-like structure to the DA
attractor. Lyapunov exponents computed by means of the Benettin √ algorithm are
3+ 5
Λ1 = 0.9618, Λ 2 = −1.3445. The largest exponent is close to ln ≈ 0.9624,
2
the value associated with the Arnold cat map. Kaplan-Yorke dimension, respec-
tively, is DKY = 1 + Λ1/|Λ2 | ≈ 1.715.
Fig. 2.5 Attractor of the map (2.8) at ε = 0.7 depicted in the unit square (unfolding of torus) (a),
and in the torus surface (b).
One more way to construct explicitly written maps with uniformly hyperbolic at-
tractors was indicated by Lopez (Lopez, 1983). The author wonders what the degree
of algebraic polynomials may be, representing a two-dimensional map with attrac-
tor of Plykin type. In fact, an example he has considered contains a function defined
as a composition of polynomials of degree up to 9, and the result is represented by
polynomials of very high order (over 10 5 ). The author believes, however, that the
lower limit of the order of polynomials defining the maps with the Plykin attractor
is given by the number 7.
Regrettably, no one of the above examples indicates a straightforward way to
construct continuous-time systems suspending the uniformly hyperbolic attractors.
2.4 From DA to the Plykin type attractor
It occurs that it is possible to come to a Plykin-type attractor on a plane or on a

sphere departing from the DA-attractor of Smale (Katok and Hasselblatt, 1995;
Hunt, 2000; Coudene, 2006). The initial step is introducing additional symmetry
in such a way that the portrait of the attractor appears to be composed of four copies
of the original version, as shown in Fig. 2.6 (left). To do so, in the map
pn+1 = pn + qn + f (pn , qn ), qn+1 = pn + 2qn + g(pn, qn ), (2.11)
obtained as result of the Smale surgery, we enlarge domain of definition in coordi-

nates (p, q) up to the square 2×2, and let this square represent the unfolding of the
torus now. In other words, we regard p and q as variables defined modulo 2 while the
functions f and g are evaluated in the same way as earlier, with the arguments de-
fined modulo 1 (see Subsection 1.2.3). Next, let us perform consecutively the steps
illustrated in Fig. 2.6. First, we take a top (or bottom) half of the square (a). Imaging
the picture drawn on a transparent film, we fold it along the vertical middle line
and glue along the left, top and bottom edges (b). Due to the symmetry, with such
gluing the filaments of the attractor match accurately and correctly. Next, inflate the
resulting square “envelope” like a balloon, and turn it into a sphere (c). Note that the
corner singularities fall into the “holes”, so, they are not relevant to the dynamics on
the attractor. At this step we arrive at the Plykin type attractor on the sphere. Finally,
by the variable change corresponding to the stereographic projection we obtain the
attractor represented on the plane (e).
Fig. 2.6 Transformations of the DA-attractor on a torus (left) to the Plykin attractor on a plane
(left): (a) selecting of a rectangular that is the top half of the square 2 × 2; (b) double folding; (c)
inflating the resulting “envelope” to the sphere; (d) transformation onto the plane with variable
change involving the Weierstrass function; (e) mapping from the plane to the sphere or back by
means of the stereographic projection.
Let us turn to more formal description. As the square 2×2 represents the unfold-
ing of torus, the change of variables must be expressed by a function of complex
variable Z = p + iq, with the fundamental periods of 2 and 2i. As known, the doubly
periodic analytic functions of complex variable relate to the class of elliptic func-
tions (Abramowitz and Stegun, 1972; Akhiezer, 1990). Particularly, the Weierstrass
function℘ with the periods 2ω 1 and 2ω2 is defined as
∞
℘(z) = z−2 + ∑
m,n=−∞
(z − 2ω1m − 2ω2n)−2 − (2ω1m + 2ω2n)−2 . (2.12)
(m,n)=(0,0)
2.4 From DA to the Plykin type attractor 45
The evaluation of this function is inverse operation to the evaluation of the elliptic
integral ∞
dx
z= . (2.13)
℘ 4x − g2 x − g3
3
Here g2 and g3 are parameters called the invariants, and their values are selected to
provide the required periods. Particularly, the half-periods ω 1 = 1 and ω2 = i are
obtained at g 2 = 11.817045 and g 3 = 0.
The variable change
2
X + iY = √ ℘(p + iq) (2.14)
g2
transforms a bottom part of the square 2 × 2 to the plane (X,Y ). The fixed point
in the origin is mapped to infinity, and three special points of the map (2.11) rep-
resenting the period-three orbit · · · → (0, 1) → (1, 0) → (1, 1) → · · · are mapped,
respectively, to (−1, 0), (1, 0) and (0, 0). In the new variables, the map formally
may be written in analytical form as 1
2
Xn+1 = √ Re [℘((1 + i)p + (1 + 2i)q + ε (1 + i) f (p, q) + ε i f (p, q))] ,
g2
2
Yn+1 = √ Im [℘((1 + i)p + (1 + 2i)q + ε (1 + i) f (p, q) + ε i f (p, q))] , (2.15)
g2
√ √
−1 g2 −1 g2
p = Re ℘ (Xn + iYn ) , q = Im ℘ (Xn + iYn ) .
2 2
Here ℘−1 designates the inverse for the Weierstrass function ℘. Surely, the analyt-
ical expression for the map is cumbersome. Practically, of course, it is much easier
to perform the iteration in the variables p and q, and convert them for graphical rep-
resentation by the variable change (2.14). In Fig. 2.6 it corresponds to the step (d).
Finally, by the variable change,
x + iy = 2℘(Z)/(1 + |℘(Z)|2), z = (1 − |℘(Z)|2)/(1 + |℘(Z)|2), (2.16)
we may represent the picture in coordinates x, y, z satisfying x 2 + y2 + z2 = 1 and

consider the attractor as being placed on the sphere; this is the step (e) in Fig. 2.6. 2
Perhaps, it is difficult to conclude the topological nature of the attractor from
the portrait on the plane shown in Fig. 2.6 (right) because essential fine details are
1 It should be noted that ℘−1 is the double-valued function (as we regard the square 2 × 2 as the
definition domain for ℘). For a given argument Z = X + iY the function ℘−1 defines two images,
one in the top and other in the bottom part of the square 2 × 2, which are linked through the
symmetry relations (p, q) ↔ (2 − p, 2 − q). With an additional agreement not to distinguish such
points, this peculiarity is not essential from the point of view of the dynamics of the system.
2 Interestingly, the relation between the sphere and the torus exploited in the above reasoning has
had an application in cartography. It lies in the basis of the so-called Guyou projection, through
which the Earth surface is depicted in a rectangle with aspect ratio 2:1, or on a plane, as periodically
reproduced pattern.
indistinguishable. Nevertheless, careful analysis indicates that this attractor is topo-

logically equivalent to the Plykin type attractor drawn in Fig. 1.8.
Explicit design of a suspension for the obtained map with the Plykin type attrac-
tor is not trivial, as noticed in the PhD thesis of Hunt (Hunt, 2000). On the basis of
this construction the problem was not resolved, but Hunt suggested an alternative
approach for suspending the Plykin type attractor discussed in the next section.
The above considerations and the variable change (2.14) may be naturally applied
to the map (2.8) as well; then we get an implicitly defined analytical map possessing
the same kind of Plykin type attractor on the plane. Its portrait is shown in Fig. 2.7.
Because of the use of the smooth variable change, the Lyapunov exponents and
Kaplan-Yorke dimension of this attractor are the same as those for the DA-attractor
of the map (2.8), namely, Λ 1 = 0.9618, Λ 2 = −1.3445, and D KY = 1.715.
Fig. 2.7 Attractor of Plykin type obtained by application of the approach discussed in the cur-
rent section to the analytical map with DA attractor (2.8) by the variable change involving the
Weierstrass function (2.14) at ε = 0.7.
.
2.5 Hunt’s example: Suspending the Plykin type attractor
The Hunt model is a non-autonomous system governed by differential equations for

two variables x and y, with the right-hand parts depending on x, y, and the time
variable t:
dx/dt = f ∗ (x, y,t), dy/dt = g∗ (x, y,t). (2.17)
Here the functions f ∗ and g∗ are continuously differentiable. They are assumed to
be 2π -periodic in respect to the argument t. Formal description and mathematical
relations for computation of these functions are given in Appendix E.
Qualitatively the evolution rule constructed by Hunt is explained in Fig. 2.8.
Panel (a) shows the initial configuration of the absorbing domain U, which con-
sists of three sub-domains Ui . In each of them special curvilinear coordinates
2.5 Hunt’s example: Suspending the Plykin type attractor 47
(r, θ )i (i = 1, 2, 3) are defined; the respective families of the coordinate lines are
shown. Filled small black circles on the panel (a) indicate positions of the origins
for the curvilinear coordinates.
Fig. 2.8 Evolution of the domain U = U1 ∪U2 ∪U3 in time for one period Δt = 2π for the Hunt
model.
The cloud of representative points from the domain U evolves in time as shown
in the figure. On one time period 2π the dynamics consists of three consecutive
stages, with each duration 2π /3. At the first stage, the domain U undergoes ver-
tical compression and horizontal extension (see panel (b)). More accurately, com-
pressing takes place along the coordinate lines of the curvilinear coordinate system
θ = const, and stretching along the lines r = const. In the second stage, the right-
hand sub-domain U 2 smoothly bends up, left, and down (c), so that finally it is
placed along the border of the domain U 3 (d). At the third stage, the left-hand sub-
domain U1 undergoes similar deformations, bending up, right, and down (e). The
result is that its lower border is placed on the edge of the domain U 2 transformed
in the previous stage (see panel (f)). The whole area takes the form of the original
domain. Moreover, the coordinate lines in their new positions run along the lines of
the original coordinates.
The described process of continuous deformation represents, say, the main con-
tent of the dynamics. Besides, the model includes some additional adjustments, by
which the origins of the curvilinear coordinates become repelling, and the transverse
compression of the figure is stronger, with the result that the transformed domain
appears to be inside the original one. These adjustments are characterized by some
parameter ε (see details in the Appendix E). Additionally, relations are formulated
for definition of the vector field ( f ∗ (x, y,t), g∗ (x, y,t)) outside the evolving domain
U(t).
Formal definition of the Hunt model is complicated and cumbersome, so it seems
very doubtful that such a construction may be designed as a real physical device. In
Ref. (Aidarova and Kuznetsov, 2009) this model was used for tests of computa-
tional methods applicable to analysis of dynamics of hyperbolic attractors. Results
of computer simulations were presented and discussed, and a technical error in the
computer program in Hunt’s work was noticed and corrected.
Figure 2.9 shows plots for observables x and y versus time, as obtained from
numerical solution of the equations for the Hunt model. Irregular behavior of these
curves indicates the chaotic nature of dynamics. The dependencies are characterized
by the presence of matched parts of distinct behavior: the horizontal plateaus and
sharp bursts. This is so due to specific nature of the Hunt model dynamics evolving
on each period 2π in the consecutive three stages.
Fig. 2.9 Dynamical variables x and y versus time in the Hunt model, ε = 0.17.
Portraits of the attractor in the Hunt model are shown in Fig. 2.10. Panel (a)
depicts the attractor in the extended phase space. Panel (b) shows the same attractor
in the projection on the plane (x, y). Like the Lorenz attractor, it is composed of two
“wings of a butterfly”, which, however, are asymmetrical. For their formation, the
second and third stages of the dynamics of the Hunt model are responsible, which
include the deformations with rotations of the lateral sub-domains. Finally, the panel
(c) shows the portrait of the attractor in the Poincaré section.
Clearly visible are the first few levels of the transverse fractal structure: the object
is built of strips, each of which contains strips of the next level, and so on. The same
object can be seen in the diagram (a) in the top plane cross-section of the figure. In
accordance with Hunt’s construction of the evolution rule, this attractor corresponds
to the Plykin-type attractor shown in Fig. 2.10.
Because of complexity of the formal definition of the model, for computation of
the Lyapunov exponents a version of the Benettin algorithm was applied basing on
numerical integration of the differential equations along neighboring phase trajecto-
ries instead of the use of equations in variations. Namely, three sets of Equations
(2.17) are integrated. One corresponds to the reference trajectory, and the other
two correspond to close orbits, with fixed initial norm of the perturbation vector

(xi − x0 )2 + (yi − y0 )2 = ε0 1, i=1, 2. After each time interval of duration 2π the
program performs orthogonalization of the vectors with the Gram-Schmidt process
2.6 The triple linkage: A mechanical system with hyperbolic dynamics 49
Fig. 2.10 Attractor of the Hunt model in the extended phase space (a), in projection on the plane
of dynamical variables x and y (b), and in the Poincaré cross-section (c); the parameter value is
ε = 0.17.
and normalization for the perturbation vectors to get again the norms ε 0 . The Lya-
punov exponents are evaluated as average rates for growth or decrease of accumu-
lated sums of logarithms for norm ratios at the end and at the beginning of the inter-
vals. It yields λ1 = 0.1532, λ 2 = −0.1930. The Lyapunov exponents for the strobo-
scopic Poincaré map are Λ 1,2 = 2πλ1,2 , and we get Λ1 = 0.9625, Λ√2 = −1.213. Note
that the largest exponent agrees with the estimate Λ ∼ = ln((3 + 5)/2) = 0.9624,
which may be derived directly from the procedure of constructing the Hunt model.
Sum of the Lyapunov exponents is negative that corresponds to an exponential de-
crease of the volume of a cloud of representative points in the course of the approach
to the attractor. Estimate of the attractor dimension with the Kaplan-Yorke formula
yields dL = 1 + λ1/|λ2 | ≈ 1.79. For the attractor as an object in the extended three-
dimensional phase space, the dimension is larger by one: D L ≈ 2.79.
2.6 The triple linkage: A mechanical system with hyperbolic

dynamics
Hunt and MacKay (Hunt and MacKay, 2003; Hunt, 2000; MacKay, 2005) estab-
lished the hyperbolic nature of the dynamics for a mechanical system called the
triple linkage that is a crank mechanism, earlier discussed in the popular mathemat-
ical article (Thurston and Weeks, 1984).
The triple linkage consists of three rods placed on a plane, free to rotate about
pivots fixed at vertices of an equilateral triangle (Fig. 2.11(a)). The opposite ends are
linked by floating pivots with three other rods, which have the opposite ends con-
nected together with one more floating pivot. An instant configuration may be spec-
ified by the angular coordinates θ 1 , θ2 , θ3 , but only two of them are independent due
to the mechanical constraints. Thus, the configuration space is a two-dimensional
manifold. Topologically, as noted by Thurston and Weeks, this is a surface of genus
3 (a “pretzel” with three holes).
Fig. 2.11 A mechanical system of Thurston and Weeks, the triple linkage (a), and a trajectory
in the three-dimensional space (θ1 , θ2 , θ3 ) (b) obtained in computations for values of lengths
41 7 9 13 1
and masses l = , l0 = ,m = ,M = , M0 = , m0 = 0. In accordance with the data
40 40 100 100 5
of MacKay and Hunt, at these parameters the motion is hyperbolic because of globally negative
curvature of the metric on the two-dimensional manifold in the configuration space.
In the absence of friction and external forces the motion by inertia takes place,
with conservation of the kinetic energy. The kinetic energy is expressed as a
quadratic form in terms of the generalized velocities, which are the time derivatives
of the local coordinates on the two-dimensional manifold. This quadratic form with
coefficients depending on the ratios of lengths and masses of the structural elements
defines a metric on the two-dimensional manifold, and the motion takes place fol-
lowing the geodesics of this metric. By enumerating variants, the authors have found
a set of parameter values, for which the metric is characterized by negative curvature
everywhere in the configuration space (the two-dimensional manifold). As known
due to Hadamard and others (Anosov, 1967; Balazs and Voros, 1986), in this situa-
tion a hyperbolic chaos occurs in its conservative version (see a simpler example in
Appendix F). A sample trajectory is shown in Fig. 2.11(b) in the three-dimensional
space (θ1 , θ2 , θ3 ) as obtained from numerical solution of the mechanical equations
of motion. The computations were performed for lengths and masses, at which the
hyperbolicity must occur in accordance with the conclusions of MacKay and Hunt
because of globally negative curvature of the metric on the manifold.
2.7 A possible occurrence of a Plykin type attractor in Hindmarsh-Rose neuron model 51
As MacKay and Hunt argued, providing additionally friction, potential field, and
appropriate feedback control, it is possible to get a system with a hyperbolic at-
tractor. So far, no concerned studies have been made, in which such attractors were
demonstrated explicitly in computer simulations or in experiments.
2.7 A possible occurrence of a Plykin type attractor in

Hindmarsh-Rose neuron model
In a rapidly developing field of neuroscience, a very popular and well studied is

the Hindmarsh-Rose model (Hindmarsh and Rose,1984; Gonzalez-Miranda, 2007;
Shilnikov and Kolomiets, 2008), which describes many features of dynamics of a
single neuron. It is represented by an autonomous set of three differential equations
ẋ = y − x3 + 3x2 − z,
ẏ = 1 − 5x2 − y, (2.18)
ż = μ (4(x − c) − z − δ ).
Here x, y, z are dynamical variables, and δ , μ , c are parameters. At fixed values of μ

and x0 (e.g., μ = 0.005, c = −1.56) the model can demonstrate various dynamical
regimes in dependence on the parameter δ . In some range δ < δ 1 , the relatively
long-period self-oscillations occur, which are referred to as the bursting activity. In
another range δ > δ 2 , the relatively fast, short-period oscillations take place, and
this is generation of spikes. In the intermediate interval δ 1 < δ < δ2 the spikes occur
on a background of the bursts (Fig. 2.12), and this is just the parameter range where
nontrivial dynamical phenomena may arise.
In the work of V. Belykh, I. Belykh, and E. Mosekilde (Belykh et al., 2005) the
authors argue in favor of occurrence of the Plykin-type attractor in Poincaré map of
a system of such kind. In fact, they examine equations in more formal generalized
form
ẋ = X(x, z), ż = μ (Z(x) − z − δ ).
Here x is a two-dimensional vector, and functions X(x, z) and Z(x) are concretized
only on a qualitative level to justify the considerations.
The authors indicate a way to introduce the Poincaré map ensuring adequate de-
scription of the dynamics in the parameter range of interest, where the bursting and
spiking activities coexist. In accordance with the undertaken qualitative analysis,
in this parameter interval the two-dimensional Poincaré map acts in a domain in a
form of a disc with several cut-offs (the holes). Particularly, a situation may occur in
mapping on a disk with three holes that is a main non-trivial condition for appear-
ance of attractor of Plykin type. The conclusion is that in the respective parameter
domain an attractor of the three-dimensional flow system is hyperbolic, and appears
as result of a secondary two-loop homoclinic bifurcation of a saddle.
These results are interesting because they relate to a system of natural origin
(neuron model). It gives a motive to speculate about possible significance of the
Fig. 2.12 Bursts and spikes on a plot of the dynamical variable versus time in the model of
Hindmarsh-Rose at μ = 0.005, c = −1.56, δ = 2.8.
uniform hyperbolicity for information processes in neurodynamics. Unfortunately,

so far no concrete equations, expressions for the functions, and concrete recommen-
dations for parameter selection have been presented associated with occurrence of
the uniformly hyperbolic attractor. Obviously, setup of more detailed researches is
needed, which would include detecting and demonstrating the uniformly hyperbolic
attractors in computations, estimating their quantitative characteristics (like Lya-
punov exponents, dimensions, etc.), and verifying the cone criterion. So far, no such
studies have been made.
2.8 Blue sky catastrophe and birth of the Smale-Williams

attractor
Shilnikov and Turaev (Turaev and Shilnikov, 1995; Shilnikov and Turaev, 1997) in-
dicated a possibility of appearance of attractors of Smale-Williams type in Poincaré
section of continuous-time systems undergoing a kind of the so-called blue sky
catastrophe.
In the simplest version, a blue sky catastrophe occurs in three-dimensional phase
space. At the bifurcation, a saddle-node limit cycle exists, from which the trajecto-
ries depart along the unstable manifold and return to the same cycle, but from the
opposite side (Fig. 2.13(a)). With a shift of the control parameter value in one direc-
tion, the saddle-node cycle transforms to a pair of close limit cycles, one stable and
2.9 Taffy-pulling machine 53
the other unstable. With a parameter shift in the opposite direction, the saddle-node
cycle disappears; instead, an attracting large-scale limit cycle emerges containing
helical coils near the former saddle-node cycle.
Having initially an angular coordinate ϕ near the saddle-node cycle, after a travel
along the unstable manifold and subsequent return, the angular coordinate is ex-
pressed by a relation containing an additive term mϕ , where an integer m may
be either 0, or 1 for the three-dimensional case. However, at higher dimensions,
any integer can occur. Particularly, m = 2 corresponds to a birth of a hyperbolic
strange attractor represented by a Smale-Williams solenoid in the Poincaré section
(Fig. 2.13(b)). The authors stress that due to such bifurcation, a system of Morse-
Smale class with simple dynamical behavior immediately converts to a system with
complex dynamics associated with the structurally stable chaotic attractor.
Fig. 2.13 The phase space arrangement at the blue sky catastrophe in its simplest form in three-
dimensional phase space (a), and schematic representation of the situation of birth of the attractor
of Smale-Williams, which is possible in the phase space of dimension 4 and more (b).
To my knowledge, no concrete systems with the last variant of the blue sky bi-
furcation were described previously in the literature. Known examples have been
presented so far only for the case of three-dimensional phase space (Gavrilov and
Shilnikov, 1999; Shilnikov and Cymbalyuk, 2005; Glyzin et al., 2008), in which the
transition to the attractor of Smale-Williams is not possible. A system of differen-
tial equations of the fourth order, in which such attractor indeed arises, has been
suggested recently by the author (Kuznetsov, 2010) and is discussed in Chap. 5.
2.9 Taffy-pulling machine
A funny demonstration of feasibility of permanent stretching deformations for a

continuous medium needed for a uniformly hyperbolic attractor, with no cuts and
without appearance of local density singularities, is delivered by the so-called taffy-
pulling machine used for preparation of cotton candy (Fig. 2.14) (MacKay, 2001;
Halbert and Yorke, 2003; Hasselblatt and Katok, 2003). Due to the rotational motion
of a set of transverse rods, the substance undergoes continuous stretching and fold-
ing with formation of a layered mixed pattern. In the paper by Halbert and Yorke a
mathematical model was constructed in the form of a map describing the dynamics
of local elements under deformation of the substance (Halbert and Yorke, 2003). It
was said that such a machine can be regarded as a variant of the physical realization
of a hyperbolic attractor. However, this is not appropriate physical situation we mean
in the context of the present book. In fact, the whole device is not a low-dimensional
dynamical system, as it contains distributed medium that is the continuum undergo-
ing deformations, while its local elements are governed by the map possessing the
hyperbolic dynamics. In other words, here we deal with dynamics of a real sub-
stance in the real space, rather than dynamics of a phase fluid in the phase space of
a low-dimensional dynamical system.
Fig. 2.14 The taffy-pulling machine for preparing cotton candy. The reader can find photos and
videos illustrating operation of such machines in the Internet using the key word “taffy-pulling
machine” for the search.
References
Abramowitz, M., Stegun, I.A.(eds.): Handbook of Mathematical Functions With Formulas, Graphs,
and Mathematical Tables. National Bureau of Standards USA (1972).
Afraimovich, V.S., Bykov, V.V., Shilnikov, L. P.: Origin and structure of the Lorenz attractor. Akad.
Nauk SSSR Doklady 234, 336–339 (1977).
Aidarova, Yu.S., Kuznetsov, S.P.: Chaotic dynamics of the Hunt model, an artificially constructed
flow system with a hyperbolic attractor. Preprint. arXiv: 0901.2727 (2009).
Akhiezer, N.I.: Elements of the Theory of Elliptic Functions. American Mathematical Society,
Providence, RI (1990).
Alligood, K., Sauer, T., Yorke, J.A.: Chaos. An Introduction to Dynamical Systems. Springer-
Verlag, New York (1996).
Inst. Steklov 90, 3–210 (1967).
Balazs, N.L., Voros, A.: Chaos on the pseudosphere. Physics Reports 143, 109–240 (1986).
Belykh, V., Belykh, I., Mosekilde, E.: The hyperbolic Plykin attractor can exist in neuron models.
Int. J. of Bifurcation and Chaos 15, 3567–3578 (2005).
Coudene, Y.: Pictures of hyperbolic dynamical systems. Notices of the American Mathematical
Society 53, 8–13 (2006).
References 55
Gavrilov, N.K., Shilnikov, A.L.: An example of blue sky catastrophe. In: Methods of qualitative
theory of differential equations and related topics. Amer. Math. Soc. Transl., II Ser. 200, pp.
165–188. AMS, Providence, RI (1999).
Glyzin, S.D., Kolesov, A.Yu., Rozov, N.Kh.: Blue Sky Catastrophe in Relaxation Systems with
One Fast and Two Slow Variables. Differential Equations 44, 161–175 (2008).
Gonzalez-Miranda, J.M.: Complex bifurcation structures in the Hindmarsh-Rose neuron model.
Int. J. Bifurcation and Chaos 17, 3071–3083 (2007).
Halbert, J.T., Yorke, J.A.: Modeling a chaotic machine’s dynamics as a linear map on a “square
sphere”. http://www.math.umd.edu/∼halbert/taffy-paper-1.pdf (2003).
Hindmarsh, J.L., Rose, R.M.: A model of neuronal bursting using three coupled first order dif-
ferential equations. Proc. Roy. Soc. London. Ser. B, Biological Sciences. 221 (1222), 87–102
(1984).
Hunt, T.J.: Low Dimensional Dynamics: Bifurcations of Cantori and Realisations of Uniform
Hyperbolicity. PhD Thesis. Univercity of Cambridge (2000).
Hunt, T.J., MacKay, R.S.: Anosov parameter values for the triple linkage and a physical system
with a uniformly chaotic attractor. Nonlinearity 16, 1499–1510 (2003).
Ilyashenko, Y.S.: Attractors and their fractal dimension. MCCME, Moscow (2005).
Kuznetsov, S.P.: Plykin-type attractor in nonautonomous coupled oscillators. Chaos 19, 013114
(2009).
Kuznetsov, S.P.: Example of blue sky catastrophe accompanied by a birth of Smale-Williams
attractor. Regular and Chaotic Dynamics 15, 348–353 (2010).
Lopez, A.O.: An Example of Polynomial Interpolation of an Hyperbolic Attractor. In: Lecture
Notes in Mathematics 1007, pp. 498–511. Springer-Verlag, Berlin (1983).
Lorenz, E.N.: Deterministic nonperiodic flow. J. Atmos. Sci. 20, 130–141 (1963).
MacKay, R.S.: Chaos in three physical systems. In: Dumortier, F., Broer, H., Mawhin, J., Van-
derbauvohede, A., Verduyn Lunel S. (eds) Equadiff 2003, pp. 59–72, World Science Press,
Singapore (2005).
MacKay, R.S.: Complicated dynamics from simple topological hypotheses. Phil. Trans. R. Soc.
Lond. A 359, 1479–1496 (2001).
Vanderbauwhede, A., Verduyn Lunel S. (eds.) Equadiff 2003, pp. 59–72, World Science Press,
Singapore (2005)
Morales, C.A.: Lorenz attractor through saddle-node bifurcations. Ann. de l’Inst. Henri Poincaré
13, 589–617 (1996).
Newhouse, S., Ruelle, D., Takens, F.: Occurrence of strange Axiom-A attractors near quasi periodic
flows on Tm , m> = 3. Comm. Math. Phys. 64, 35–40 (1978).
Ott, E.: Chaos in Dynamical Systems. Cambridge University Press, Cambridge (2002).
Ruelle, D., Takens, F.: On the nature of turbulence Commun. Math. Phys. 20, 167–192 (1971).
Schuster, H.G., Just, W.: Deterministic Chaos: An Introduction. Wiley-VCH Verlag GmbH & Co.
KGaA, Weinheim, FRG (2005).
Shilnikov, A., Cymbalyuk, G.: Transition between tonic spiking and bursting in a neuron model
via the blue-sky catastrophe. Phys. Rev. Lett. 94, 048101 (2005).
Shilnikov, A., Kolomiets, M.: Methods of the qualitative theory for the Hindmarsh-Rose model: a
case study. Int. J. Bifurcation and Chaos 18, 2141–2168 (2008).
Shilnikov, L.P., Turaev, D.V.: Simple bifurcations leading to hyperbolic attractors. Computers
Math. Appl. 34, 173–193 (1997).
Sparrow, C.: The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors. Springer-Verlag,
New York, Heidelberg, Berlin (1982).
Thurston, W.P., Weeks, J.R.: The mathematics of three-dimensional manifolds. Sci. Am. 251, 94–
106 (1984).
Tucker, W.: A rigorous ODE solver and Smale’s 14th problem. Comp. Math. 2, 53–117 (2002).
Turaev, D.V., Shilnikov, L.P.: On blue sky catastrophes. Doklady Akademii Nauk 342, 596–599
(1995).
Part II
Low-Dimensional Models
Chapter 3
Kicked Mechanical Models and Differential
Equations with Periodic Switch
Abstract Some physical systems admit very natural description in discrete time.
For example, it relates to mechanical or electronic systems under action of periodic
pulses (kicks). Instead considering the motion in continuous time, one can derive a
map, which expresses the state variables just before (or after) the next kick via the
state just before (or after) the previous kick. In papers and textbooks on nonlinear
dynamics, maps of Hénon, Ikeda, and Chirikov-Taylor-Zaslavsky are derived and
studied in this context (Heagy, 1992; Kuznetsov et al., 2008; Sagdeev et al., 1988).
Is it possible for the uniformly hyperbolic attractors to occur for diffeomorphisms
arising in description of periodically kicked systems? In this chapter we introduce
some simple physically motivated toy models with dynamics described by maps
possessing attractors of Smale-Williams type.
One more approach to design of non-autonomous systems with uniformly hy-
perbolic attractors is to organize evolution in time as certain periodically repeated
distinct stages, in such a way that the right-hand parts in the differential equations
are defined in a special form for each of the successive stages. In other words,
the right-hand parts may be thought of as being represented by functions piece-
wise continuous and periodic in time. It may be referred to as differential equations
with periodic switch. One example of such kind will be suggested with attractor of
Smale-Williams type.
Earlier, the approach based on the periodically switching equations was devel-
oped by Hunt (Hunt, 2000) in the context of the problem of suspending attractor
of Plykin type, as mentioned in Chap. 2. We introduce a simpler model with analo-
gous attractor governed by equations with right-hand parts represented by periodic
piece-wise continuous functions of time. Then, exploiting the structural stability, we
modify the model to obtain differential equations with smooth coefficients possess-
ing attractor of the same type.
60 3 Kicked Mechanical Models and Differential Equations with Periodic Switch
3.1 Smale-Williams solenoid in mechanical model: Motion of a

particle on a plane under periodic kicks
Let us turn to mechanical setup giving rise to maps with uniformly hyperbolic at-
tractors of Smale-Williams type (Kuznetsov and Turukina, 2010).
Consider motion of a particle of unit mass on a plane (x, y) with friction force
proportional to the instant velocity. Let the motion take place in the presence of a sta-
1 1
tionary potential field; the potential function is U(x, y) = − μ (x2 + y2 ) + μ (x2 +
2 4
y2 )2 . This function possesses circular symmetry being minimal on the unit circle.
Let us assume that short pulses kick the particle with period T , and the direction and
magnitude of the kicks depend on the instant position of the particle. In other words,
some force field is switched on for a short time, and then switched off, again and
again. Let us assume that momentum accepted by the particle from a single kick is
expressed as P = Px (x, y) i + Py (x, y) j. Here x and y are the instant coordinates of the
particle, the functions Px,y (x, y) describe the spatial distribution of the force field, i
and j are unit vectors in the x and y directions, respectively. For simplicity we set
the unite friction coefficient and write down the equations of the motion in the form
∞
ẍ + ẋ = μ x(1 − x2 − y2 ) + Px (x, y) ∑ δ (t − nT ),
n=−∞
∞ (3.1)
ÿ + ẏ = μ y(1 − x2 − y2 ) + Py (x, y) ∑ δ (t − nT ).
n=−∞
To select appropriate spatial distribution for the field P(x, y) we proceed as follows.
Let us assume that initially we set a ring of particles resting on a unit circle, with
coordinates x = cos ϕ and y = sin ϕ (0 ϕ < 2π ). After a kick caused by the force
field, each particle characterized by initial angular coordinate ϕ accepts the momen-
tum components Px (x, y) and Py (x, y) that entails certain displacement of the particle
after some time. If the potential field is not taken into account (μ = 0), the particle
would stop because of the friction at the coordinates
x = x + Px (x, y), y = y + Py(x, y). (3.2)
Let us specify the functions Px (x, y) and Py (x, y) requiring the particles to arrive
again on the unit circle, but in such a manner that a single bypass of the original
ring would correspond to an M-fold bypass for the new disposition, with an integer
M 2. In other words, the angular coordinate has to undergo the multiplication by
M
ϕ = M ϕ ( mod 2π ). (3.3)
For this, the final coordinates of the particle must satisfy
x = cos ϕ = cos M ϕ , y = sin ϕ = sin M ϕ . (3.4)

3.1 Smale-Williams solenoid in mechanical model 61
It will be the case if the functions characterizing the distribution of the pulse force
field are selected in such a way that
Px (x, y) = x − x = cos M ϕ − cos ϕ = Re(eiMϕ − eiϕ ),

(3.5)
Py (x, y) = y − y = sin M ϕ − sin ϕ = Im(eiMϕ − eiϕ ).
Taking into account the relation of the Cartesian coordinates and the angular variable
on the unit circle (x = cos ϕ and y = sin ϕ ), one easily obtains
Px (x, y) + i Py (x, y) = (x + iy)M − (x + iy). (3.6)
In the relations for Px,y (x, y) it is allowable to express even powers of one variable
through the other, as on the unit circle the relation y 2 = 1 − x2 is valid. Then, we
obtain for M=2:
Px (x, y) = 2x2 − x − 1, Py (x, y) = 2xy − y, (3.7)
and for M = 3 :
Px (x, y) = 4x3 − 4x, Py (x, y) = −4y3 + 2y. (3.8)
Note that in the last case the field of the pulse force is potential; moreover, the
components x and y depend only on one single respective coordinate, that may be
regarded as an advantage of this model from the point of view of simplicity of its
implementation.
Parameter μ may be selected as relatively small value; then, during the character-
istic time of the motion caused by the kick, a displacement of the particle caused by
the potential field U(x, y) is rather negligible. On the other hand, the period of kicks
T is assumed large enough to have time for the particle to approach the potential
minimum. These conditions are not strong and have to be valid at least in a rough
approximation. It is so because of structural stability of the hyperbolic attractor we
intend to construct. With initial state just before the n-th kick x n = {x, ẋ, y, ẏ}t=nT −0 ,
one can determine the state before the next, n + 1-th kick from solution of Eqs. (3.1)
in the period T :
ẍ + ẋ = μ x(1 − x2 − y2), ÿ + ẏ = μ y(1 − x2 − y2 ) (3.9)
with the initial conditions specified by the state just after the kick:
x|t=nT +0 = xn , ẋ|t=nT +0 = ẋn + Px (xn , yn ),

(3.10)
y|t=nT +0 = yn , ẏ|t=nT +0 = ẏn + Py (xn , yn ).
It is easy to write down additionally the variation equations linearized near the
reference trajectory needed for computation of Lyapunov exponents. Substituting
(x + x̃, y + ỹ) into Eqs. (3.9), for small perturbations x̃, ỹ we obtain equations
x̃¨ + x̃˙ = μ x̃ − μ (y2 x̃ + 3x2x̃ + 2xyỹ), ỹ¨ + ỹ˙ = μ ỹ − μ (x2 ỹ + 3y2ỹ + 2xyx̃) (3.11)
supplemented by the initial conditions just after the kick following from (3.10):

x̃|t=nT +0 = x̃n , x̃˙t=nT +0 = x̃˙n + x̃n∂x Px (xn , yn ) + ỹn∂y Px (xn , yn ),
(3.12)
ỹ|t=nT +0 = ỹn , ỹ˙t=nT +0 = ỹn + x̃n ∂x Py (xn , yn ) + ỹn ∂y Py (xn , yn ).
Relations (3.9) and (3.10) give rise to a four-dimensional Poincar é map xn+1 =
f(xn ). This map is invertible because all the state transformations produced by the
kicks and by the continuous time evolution governed by the differential equations
are invertible. Although the exact form for the map is not derived analytically, the
action of this map can be easily reproduced by means of numerical solution of the
differential equations (3.9) by a computer program.
What is of relevance to occurrence of the Smale-Williams attractor, that’s the
topological property of the ensemble after the transformation, namely, emergence of
a configuration encircling the origin two or three times. Compression in the transver-
sal direction in the phase space is provided by the friction and by the effect of the
potential field, due to which the particle drifts towards the potential minimum at
the unit circle. In contrast to the classic construction discussed in Chap. 1, in these
models the attractors of Smale-Williams type are embedded in the four-dimensional
phase space of the Poincaré map rather than in the three-dimensional one.
In Figs. 3.1 — 3.4 illustrations are shown obtained from numerical solution of
the equations of motion of the particle under pulsing driving force for the case
M=2, with the force field specified by (3.7). Figure 3.1 shows a trajectory of the
particle in the course of its motion on the plane (x, y) that corresponds to the
dynamics of the system on the attractor. Portrait of the attractor in the strobo-
scopic cross-section is presented in Fig. 3.2 in projection on the plane (x, y). Its
visual similarity to the Smale-Williams solenoid is obvious. The inset shows a mag-
nified fragment of the picture, and the intrinsic Cantor-like transversal structure
may be seen. In Fig. 3.3 an iteration diagram is drawn for the angular coordinate
ϕn = arg(x(nT − 0) + iy(nT − 0)) determined just before each kick. Observe that
in very good approximation the angular coordinate behaves in accordance with the
expanding circle map, or Bernoulli map. One bypass of the circle for pre-image
corresponds to a two-fold bypass for the image.
Fig. 3.1 Trajectory of the particle on the plane (x, y) in the models (3.1), (3.7) at μ = 0.44, T = 5
for 30 periods of the driving. Arrows indicate directions of the motion.
3.1 Smale-Williams solenoid in mechanical model 63
Fig. 3.2 Portrait of attractor for the systems (3.1), (3.7) in projection on the plane (x, y) in the
stroboscopic cross-section at μ = 0.44, T = 5. A fragment is shown with magnification to resolve
the transversal Cantor-like structure of the object.
Fig. 3.3 Iteration diagram for the angular coordinate obtained from numerical solution of Eqs.
(3.1) with the distribution of the force field (3.7) at μ = 0.44, T = 5.
Fig. 3.4 Plot for Lyapunov exponents of the models (3.1), (3.7) versus parameter μ at T = 5.
In Fig. 3.4 plots for the Lyapunov exponents are shown versus the parameter μ at
fixed T . The Lyapunov exponents were computed by means of joint numerical solu-
tion of Eq. (3.9) for reference orbit with relations (3.10) at the kicks and a collection
of four variation equations (3.11) with relations (3.12) at the kicks; Gram-Schmidt
orthogonalization and normalization of the perturbation vectors were performed af-
ter each kick. Observe that the largest Lyapunov exponent remains nearly constant,
and for the stroboscopic map it is very close to the estimate Λ 1 ≈ ln2. It corresponds
to the Lyapunov exponent for the Bernoulli map describing approximately the dy-
namics of the angular coordinate. Other Lyapunov exponents are negative. Estimate
of the fractal dimension of the attractor in the Poincaré section at μ =0.44 yields
D=1.328.
In Figs. 3.5 — 3.8 analogous illustrations are presented for the model with M=3,
and the force field is specified by (3.8). Figure 3.5 shows a trajectory of the par-
ticle on the plane (x, y). Portrait of the attractor in the stroboscopic cross section
is shown in Fig. 3.6 that corresponds to a variant of the Smale-Williams solenoid
(with tripling of number of coils at each next step of the construction) and manifests
Cantor-like transversal structure distinguished in the inset. In Fig. 3.7 an iteration
diagram is shown for the angular coordinate determined just before the kicks. It
behaves in very good approximation in accordance with the three-fold expanding
circle map. In Fig. 3.8 the Lyapunov exponents are plotted versus the parameter
μ at fixed T . The largest Lyapunov exponent is nearly constant, and for the stro-
boscopic map it is close to the estimate Λ 1 ≈ ln 3. Other Lyapunov exponents are
negative. Estimate of the fractal dimension of the attractor in the Poincaré section at
μ =0.22 yields D=1.47.
Fig. 3.5 Trajectory of the particle on the planes (x, y) in the models (3.1), (3.8) at μ = 0.22, T = 8
for 30 periods of the driving. Arrows indicate directions of the motion.
Fig. 3.6 Portrait of attractor for the system (3.1) with distribution of the force field (3.8) in projec-
tion on the plane (x, y) in the stroboscopic cross-section at μ = 0.22, T = 8. A fragment is shown
with magnification to resolve the transversal Cantor-like structure of the object.
3.2 A set of switching differential equations with attractor of Smale-Williams type 65
Fig. 3.7 Iteration diagram for the angular coordinate obtained from numerical solution of Eqs.
(3.1) with the distribution of the force field (3.8) at μ = 0.22, T = 8.
Fig. 3.8 Plot for Lyapunov exponents of the models (3.1), (3.8) versus parameter μ at T = 8.
3.2 A set of switching differential equations with attractor of

Smale-Williams type
The advantage of the above mechanical models is their simplicity. However, the
dimension of the phase space is larger by one than the minimal dimension required
for maps with attractors of Smale-Williams type.
Let us construct now a non-autonomous system in which the state at fixed time
is given by a three-dimensional vector (x, y, z), and the extended phase space is
four-dimensional. Consider four periodically repeated stages of continuous trans-
formations, with each unit time duration governed by the differential equations, as
follows:
I. Differential rotation around xx-axis:
π π
ẋ = 0, ẏ = − xz, ż = xy. (3.13)
2 2
II. Non-uniform shift with compression in y -direction:
ẋ = 0, ẏ = −d1 (y − 2x2 + 1), ż = 0. (3.14)
III. Rotation by angle π /2 around y -axis:

π π
ẋ = − z, ẏ = 0, ż = x, (3.15)
2 2
IV. Compression towards unit circle in the plane z =0 :
ẋ = μ x(1 − x2 − y2 ), ẏ = μ y(1 − x2 − y2 ), ż = −d2 z. (3.16)
The dynamical equations may be written in the unified form as

π
ẋ = − z f3 (t) + μ x(1 − x2 − y2 ) f4 (t),
2
π
ẏ = − xz f1 (t) − d1(y − 2x2 + 1) f2 (t) + μ y(1 − x2 − y2 ) f4 (t), (3.17)
2
π π
ż = xy f1 (t) + x f3 (t) − d2z f4 (t),
2 2
where f k (t) = 1, if k − 1 t − 4[t/4] < k, and 0 otherwise. Considering all four
stages successively, one can derive the following map from analytic solution of Eqs.
(3.13) —- (3.16):
xn+1 = b(xn ), (3.18)
where
⎛ eμ z
⎞
−
⎜ 1+(e −1) z +(e 1 (y cos 2 x−z sin π2 x)+(2x2 −1)(1−e−d1 ))
2 μ 2 −d π 2
⎟
⎜ ⎟
⎜ eμ −d1 (y cos π2 x−z sin π2 x)+eμ (2x2 −1)(1−e−d1 ) ⎟
b(x) = ⎜ ⎟.
⎜ ⎟
1+(e −1) z2 +(e−d1 (y cos π2 x−z sin π2 x)+(2x2 −1)(1−e−d1 ))
2
⎝ 2 μ
⎠
e−d2 x
As well, the inverse map may be written down explicitly

⎛ ⎞
zed2
⎜ d−μ ⎟
⎜ ye cos( π2 ed2 z)−xe−μ sin( π2 ed2 z) 2 e2d2 − 1)(1 − ed ) cos( π ed2 z) ⎟
−1 ⎜ √ + (2z ⎟
b (x) = ⎜ 1+(x2 +y2 )(e−2μ −1) 2 ⎟.
⎜ ⎟
⎝ −yed−μ sin( π ed2 z)−xe−μ cos( π ed2 z) ⎠
√ π
2 2
− (2z e − 1)(1 − e ) sin( 2 e z)
2 2d 2 d d 2
1+(x2 +y2 )(e−2μ −1)
(3.19)
Figure 3.9(a) shows an absorbing domain appropriatelyselected at the parame-
ter values d1 = d2 =2, μ =2 defined by the inequality (p x2 + y2 − 1)2 + qz2 1,
p = 12.5, q = 25, and its image under application of the Poincaré map (3.18). The
action of the map agrees qualitatively with the Smale-Williams construction. Dia-
gram (b) shows portrait of the attractor of the map. In the inset one can observe
intrinsic transversal Cantor-like structure of filaments. Although local directions of
expansion do not coincide precisely with directions of the filaments, and local com-
pressions are directed with different angles dependent on position in the phase space,
the present degree of correspondence with the geometric construction of Smale-
3.2 A set of switching differential equations with attractor of Smale-Williams type 67
Williams appears to be sufficient for existence of the uniformly hyperbolic attractor.

The panel (c) illustrates the iteration diagram for the angular variable determined
at the n-th iteration of the map by the relation ϕ n = arg(xn + iyn ). Its form corre-
sponds to expanding circle map, slightly deformed Bernoulli map: variation of the
argument by 2π implies variation of the function by 4π .
Fig. 3.9 Absorbing domain in the phase space and its image after application of the map (3.18) (a);
portrait of the attractor (b) and iteration diagram for the angular coordinate (c). Parameter values
are d1 = d2 =2, μ = 2.
The Lyapunov exponents are found to be Λ 1 = 0.661, Λ2 = −2.601, Λ3 =

−2.650, and the Kaplan-Yorke dimension of the attractor is D L ≈ 1 + Λ1 /|Λ2 | ≈
1.25. Numerical computations confirm the cone criterion in the entire absorbing
domain.
Figures 3.10 and 3.11 illustrate some features of the dynamics in continuous
time in accordance with the differential equation (3.17). In Fig. 3.10 the depen-
dence of the dynamical variables versus time is plotted manifesting visually chaotic
dynamics. Figure 3.11 shows three-dimensional projection of the attractor from the
extended phase space of the continuous-time system. The picture is drawn in the
gray-scale technique: the deepness of the gray tone indicates relative duration of
Fig. 3.10 Dependence of variable x, y, z on time obtained from numerical solution of Eqs. (3.17)
at d1 = d2 =2, μ = 2.
Fig. 3.11 Portrait of the attractor of the continuous-time system (3.17) in three-dimensional pro-
jection from the four-dimensional extended phase space. Gray tones indicate relative duration of
presence of the phase trajectories in the respective pixels.
presence of the phase trajectories in the respective pixels. In cross-section of the

object by the hyper-plane t=const the solenoid shown in Fig. 3.9b takes place.
3.3 Explicit dynamical system with attractor of Plykin type
The present section is aimed at explicit construction of a non-autonomous flow sys-

tem with Plykin type attractor (Kuznetsov, 2009a,b,c, 2011). The starting point is
some map composed of successive periodically repeating stages of continuous geo-
metrically evident transformations defined on a two-dimensional sphere.
3.3.1 Plykin type attractor on a sphere
Let us consider a sphere of unit radius (Fig. 3.12). A point on the sphere can be
specified in angular coordinates (θ , ϕ ), or in Cartesian coordinates
x = cos ϕ sin θ , y = sin ϕ sin θ , z = cos θ , (3.20)
which satisfy the relation x 2 + y2 + z2 = 1. As proven by Plykin, a map on a sphere

can possess hyperbolic attractor in the presence of at least four holes, the areas not
visited by trajectories belonging to the attractor. In our construction the holes will
be represented
√ √neighborhoods of points A, B, C, D with coordinates (x, y, z) =
by
(±1/ 2, 0, ±1/ 2). North and south poles of the sphere are denoted by N and S,
respectively.
3.3 Explicit dynamical system with attractor of Plykin type 69
Fig. 3.12 A sphere with marked points A, B, C, D, neighborhoods of which will correspond to the
holes not visited by trajectories on the attractor. The north and south poles are indicated with N
and S, respectively. The angular coordinates (θ , ϕ ) are shown for some point M, and axes of the
Cartesian coordinates x, y, z are depicted.
Let us consider a periodic sequence of four successive continuous transforma-

tions, each unit time duration.
I. Flow down along circles of latitude: This is the motion of representative
points on the sphere away from the meridians NABS and NDCB towards the merid-
ians equally distant from the arcs AB and CD. In Cartesian coordinates it is governed
by equations
ẋ = −ε xy2 , ẏ = ε x2 y, ż = 0, (3.21)
where ε is a parameter.
II. Differential rotation around z-axis with angular velocity depending on z
linearly, in such a way that points B and C do not move, while A and D exchange
their positions; it corresponds to equations
√ √
ẋ = π (z/ 2 + 1/2)y, ẏ = −π (z/ 2 + 1/2)x, ż = 0. (3.22)
III. Flow down to the equator: That is motion of representative points along
circles centered on the x-axis; the motion takes place on the sphere from the great
circle ABCD, towards the equator:
ẋ = 0, ẏ = ε yz2 , ż = −ε y2 z. (3.23)
IV. Differential rotation around x -axis with angular velocity depending on x

linearly, in such a way that in the plane orthogonal to x-axis and containing the
point C the representative points do not move, while those in the plane containing
the point B undergo a turn by 180 ◦:
√ √
ẋ = 0, ẏ = −π (x/ 2 + 1/2)z, ż = π (x/ 2 + 1/2)y. (3.24)
Note symmetry of the procedure: The first and the second pairs of the transfor-
mations are identical to exchange of the variables x and z.
Intuitively, it looks reasonable that this sequence of transformations will generate
a flow on the sphere accompanying with formation of filaments of fine transversal
structure that is a characteristic feature of the Plykin type attractors.
One can accept an interpretation that an instant speed of a representative point on
the sphere is determined by combination of two vector fields switched on and off,
turn by turn. One corresponds to dynamics during the stages of flow down, and the
other to the stages of differential rotation. We can introduce angles ᾱ and β̄ deter-
mining directions of the fields, and coefficients p and q responsible for switching
them on and off:
π π 1 1 1 1
ᾱ = β̄ = (−1)[t/2] − , p= + (−1)[t] , q = (−1)[t/2] − (−1)[t/2+1/2] .
4 4 2 2 2 2
√
(Here [τ ] designates the integer part of τ .) Setting K = π / 2, we write down
Eqs. (3.21) —- (3.24) in a compact unified form, which will be used for modifi-
cation considering in the next section:

1
ẋ = −pε y2 (x cos ᾱ + z sin ᾱ ) cos ᾱ + qKy −x sin β̄ + z cos β̄ + √ cos β̄ ,
2
1
ẏ = pε y(x cos ᾱ + z sin ᾱ ) − qK(x cos β̄ + z sin β̄ ) −x sin β̄ + z cos β̄ + √ ,
2
2
1
ż = −pε y (x cos ᾱ + z sin ᾱ ) sin ᾱ + qKy −x sin β̄ + z cos β̄ + √ sin β̄ .
2
2
(3.25)
The Poincaré map, which determines transformation of a state vector x n =
(xn , yn , zn ) per period T = 4, may be derived explicitly. From successive solution
of the differential eqs. (3.21) — (3.24) with account of the mentioned symmetry,
one can express the resulting state vector x n+1 as
xn+1 = f+ (f− (xn )) ≡ f(xn ), (3.26)
where
⎛ ⎞
±z
⎜ ε2 (x2 +y2 ) π √ ε 2 π √ ⎟
cos (z 2 + 1) ± xe− 2 (x +y ) sin (z 2 + 1) ⎟
2
⎜ ye
⎜ 2 2 ⎟
⎜ ⎟
⎜ sinh ε (x 2 + y 2 ) ⎟
⎜ cosh ε (x2 + y2 ) + ε (y2 − x2 ) ⎟
⎜ ε (x + ) ⎟
⎜
f± (x) = ⎜
2 y 2
⎟.
⎟
⎜ ε (x2 +y2 ) π √ √
− 2ε (x2 +y2 ) cos π (z 2 + 1) ⎟
⎟
⎜ ye 2 sin (z 2 + 1) ∓ xe
⎜ 2 2 ⎟
⎜ ⎟
⎜ ⎟
⎝ sinh ε (x 2
+ y 2
) ⎠
cosh ε (x + y ) + ε (y − x )
2 2 2 2
ε (x2 + y2 )
The relation (3.26) determines a map of the sphere to itself f, for which C is a fixed
point, while A, B and D compose a period-3 unstable orbit: A → D → B → A. The
map f is invertible. The inverse map appears as a result of the same transformations
in backward order, with reversed directions of the rotations.
Figure 3.13 shows attractor of the map f at ε = 0.77. Observe specific fractal-like
transversal structure of the attractor: the object looks like composed of strips, each
of which contains narrower strips of the next level, etc.
Fig. 3.13 Attractor of the map (3.26) at ε = 0.77 on the unit square in an axonometric projection
(a) and in polar azimuthal projection (b).
Two Lyapunov exponents of the attractor are plotted in Fig. 3.14 versus the pa-
rameter ε . The procedure of computation of the Lyapunov exponents is based on the
Benettin algorithm (see Appendix A), with the joint iteration of the mapping (3.26)
together with two sets of the variational equations
x̃n+1 = f + (f− (xn ))f − (xn )x̃n (3.27)
with orthogonalization of the perturbation vectors to the radius vector x = (x, y, z)

and to each other, and their normalization to a fixed constant.
Fig. 3.14 Dependence of Lyapunov exponents on the parameter ε (solid line) and the Kaplan-
Yorke dimension of the attractor (dotted line) for the map (3.26). The range of hyperbolic chaos
according to the numerical verification of the cone criterion is approximately 0.35 < ε < 2. At
ε > εc ≈ 2.023 there is no chaos; the dynamics become periodic.
In the range ε < ε c ≈ 2.03 one of two Lyapunov exponents is positive, which in-
dicates the presence of chaos, and the other is negative. If ε > ε c (strong dissipation
on the stages I and III), chaos disappears. Note smooth dependence of the positive
Lyapunov exponent on the parameter, without drops to negative values intrinsic to
many systems with non-hyperbolic attractors (see Appendix B). Additionally, a plot
is shown for the Kaplan-Yorke dimension, which in this case is D KY = 1 + Λ1 /|Λ2 |.
With growth of ε , the dimension gradually decreases from 2 (that corresponds to
extremely small dissipation) approximately to 1. Particularly, at ε =0.77 the Lya-
punov exponents are Λ 1 = 0.959 and Λ 2 = −1.141, so the Kaplan-Yorke dimension
is DKY ≈ 1.84.
Numerical verification of the cone criterion confirms hyperbolic nature of the
attractor in the parameter interval 0.35 < ε < 2 (Kuznetsov, 2009a,b). Figure 3.15
illustrates how the attractor structure evolves as the parameter is varied in this range:
the strips forming the transversal Cantor-like structure look fat at smaller ε and
becomes narrower with increase of it.
Fig. 3.15 Portraits of attractors of the map (3.26) depicted in the polar azimuthal projection in
the northern hemisphere for different values of ε in the range of hyperbolicity. (The maps for the
southern hemisphere look similarly northern, being obtained as mirror images about the vertical
axis on the plots.)
3.3.2 Plykin type attractor on the plane
Description of the dynamics can be reformulated to represent states of the system

on a plane. To do this, the following variable change is appropriate
√
x − z + iy 2
W = X + iY = √ . (3.28)
x+z+ 2
It corresponds to stereographic projection
√ from the √ sphere to the plane, with selec-
tion of the projection point at C(−1/ 2, 0, −1/ 2). This point does not belong to
the attractor (it is in the “hole”), so the attractor on the plane appears to be located
in a bounded domain. The backward variable change is expressed as
1 − |W |2 + 2Re W 2Im W 1 − |W|2 − 2Re W

x= √ , y= , z= √ . (3.29)
2(1 + |W|2 ) 1 + |W|2 2(1 + |W|2 )
Now, the dynamical eqs. (3.21) — (3.24) may be rewritten in new variables. The
advantage is that the third redundant variable is excluded. Computing time deriva-
tive of (3.29), substituting derivatives of the old variables from (3.21) — (3.24),
and expressing the result via X and Y we obtain the following equations in the four
stages.
Stage I:
(1 + 2X − X 2 − Y 2 )(X − 1)
Ẋ = ε Y 2 ,
(1 + X 2 + Y 2 )2
(3.30)
(1 + 2X − X 2 − Y 2 )(1 + 2X − X 2 + Y 2 )
Ẏ = ε Y .
2(1 + X 2 + Y 2 )2
Stage II:
π Y (1 − X)2 π (1 − X)(1 + 2X − X 2 + Y 2 )
Ẋ = √ , Ẏ = − √ . (3.31)
2 1 + X2 +Y2 2 2 1 + X2 +Y2
Stage III:
(1 − 2X − X 2 − Y 2 )(X + 1)
Ẋ = ε Y 2 ,
(1 + X 2 + Y 2 )2
(3.32)
(1 − 2X − X 2 − Y 2 )(1 − 2X − X 2 + Y 2 )
Ẏ = ε Y .
2(1 + X 2 + Y 2 )2
Stage IV:
π Y (1 + X)2 π (1 + X)(1 − 2X − X 2 + Y 2 )
Ẋ = − √ , Ẏ = − √ . (3.33)
2 1 + X2 +Y2 2 2 1 + X2 +Y2
The Poincaré map may be derived either from direct analytic solution of the dy-
namical equations (step by step on the four stages), or from (3.26), by means of the
variable changes (3.28) and (3.29). The resulting map is expressed as
Wn+1 = T+ (T− (Wn )) ≡ T(Wn ), (3.34)

⎛ √ ρ /2 ⎞
sinΦ∓(1−|W |2 +2X)e−ρ /2 cosΦ
±(1 − |W|2 − 2X) − 2 2Ye
√
⎜ cosh ρ +γ sinh ρ /ρ ⎟
⎜ √ ⎟
⎜ ±(1 − |W|2 − 2X) + 2 2Yeρ /2 sinΦ∓(1−|W
√ |2 +2X)e−ρ /2 cosΦ
+ 2 + 2|W|2 ⎟
⎜ ⎟
T± (W) = ⎜ √ 2√2Yeρ /2 cos Φ±(1−|W |2 +2X)e−ρ /2 sinΦ ⎟,
⎜ √ ⎟
⎜ 2 ⎟
⎝ √
ρ /2 sinΦ∓(1−|W |2 +2X)e−ρ /2 cosΦ ⎠
±(1 − |W|2 − 2X) + 2 2Ye
√ + 2 + 2|W|2
cosh ρ +γ sinh ρ /ρ
where
8Y 2 + (1 − |W|2 − 2X)2
|W |2 = X 2 + Y 2 , ρ =ε ,
2(1 + |W|2 )2
8Y 2 − (1 − |W|2 − 2X)2 1−X
γ =ε , Φ=π .
2(1 + |W|2 )2 1 + |W|2
Portrait of the attractor on the plane obtained from iteration of the map (3.34) at
ε =0.77 is shown in Fig. 3.16(a).
To reveal that this attractor is of Plykin type, and to confirm the uniform hyper-
bolicity, it is appropriate to turn to graphical representation of the stable and unstable
foliations. To draw the stable and unstable manifolds with computer, we do the fol-
lowing. First, for a given point on the attractor x we obtain the image by iteration
of the Poincaré map x̄ = f N (x), and the pre-image by iteration of the inverse map
x̃ = f−N (x), where N is some empirically chosen integer. Then, with random initial
conditions ỹ in a small neighborhood of x̃, we get by iteration of the Poincaré map a
set of points y = f N (ỹ), which mark the unstable manifold. In a similar way, starting
with random initial conditions ȳ in a small neighborhood of x̄, we draw the stable
manifold by a set of points y = f −N (ȳ). The accuracy the manifolds are depicted
grows fast with increase of N. Actually, already selected N = 6 is enough to get so
small errors that they are visually indistinguishable in the plot.
In Fig. 3.16(b) the stable and unstable foliations are drawn in the domain D con-
taining the attractor.The stable and unstable manifolds run along the black and gray
curves, respectively. As seen, the unstable manifolds follow filaments of the attrac-
tor, while the stable ones are transversal to them. Mutual disposition of the stable
and unstable manifolds in the domain D certainly excludes their tangencies.
Markov partition of the domain D can be obtained. A border separating the ele-
ments of the partition is just the stable manifold of the saddle-fixed point R, which
belongs to the attractor (at ε =0.77 it is located at X=0, Y = −3.45524253). The
partition is depicted in Fig. 3.16(c) as obtained from computations using the above
technique of drawing the stable manifolds. Comparison of topology of the partition
(mutual location of the sub-areas) with that shown in Fig. 1.14(b) for the Plykin type
attractor clearly indicates that the attractor relates to the same class. Respectively,
Fig. 3.16 (a) Portrait of the attractor of the map (3.34) at ε =0.77. (b) Stable and unstable foliation
obtained in computations in the absorbing domain D shown in white. The stable and unstable man-
ifolds run along the black and gray curves, respectively. (c) Markov partition: the border between
the sub-areas is the stable manifold of the saddle fixed point R belonging to the attractor. The par-
tition may be compared to that shown in Fig. 1.14(b) that indicates relation of the attractor to the
same Plykin type class. For clearness of the comparison, direction of the horizontal coordinate axis
is inverted here, and the same colors and letters are used for the sub areas as those in Fig. 1.14(b).
the attractor discussed in the previous section has to be interpreted as the Plykin
type attractor on the sphere.
Both representations, on the sphere and on the plane, differ only by a smooth
variable change. Hence, the Lyapunov exponents for the attractor on the plane cor-
respond exactly to those plotted in Fig. 3.14. The attractor is uniformly hyperbolic
as follows from the numerical verification of the cone criterion at least in the pa-
rameter range 0.35 < ε < 2. Figure 3.17 illustrates transformation of structure of
the attractor portraits drawn on the plane as the parameter is varied in this interval.
Fig. 3.17 Portraits of attractors on the plane for different values of ε in the range of hyperbolicity.
3.4 Plykin-like attractor in smooth non-autonomous system
One can object that the above examples are not completely satisfactory from the
point of view of the classic hyperbolic theory as they are based on discontinuous
evolution in time. Nevertheless, their evolution in a finite time (the period of the
sequence of the stages) is governed by a diffeomorphism, which is a composition
of the diffeomorphisms relating to the stages. So, they deliver examples of discrete-
time evolution rules possessing the hyperbolic attractors. Moreover, we may try to
modify slightly the evolution rules to exclude the discontinuities. If the modification
is in some sense weak, one may expect that the variations in the stroboscopic map
3.4 Plykin-like attractor in smooth non-autonomous system 77
are also small, and, by virtue of the structural stability, the uniformly hyperbolic
attractor will survive. Here we consider such approach in application to the system
discussed in Sect. 3.3.
So, we intend to construct a version of the non-autonomous model containing
only smooth functions. Let us return to Eqs. (3.25), which define a flow on the
sphere, and set there
π π π π π π π
ᾱ = cos t − , β̄ = sin t − , p = 1, q = sin t. (3.35)
4 2 4 4 2 4 2
Now, two vector fields, responsible for the flow down and for the differential rota-
tion, vary in time continuously and smoothly, undergoing rotations in space in such
a way that the original configuration is repeated with the period T = 4. The second
field oscillates, reversing the direction twice per period. To retain the alternating
stages of flow down and of differential rotation, the variations of one and other
fields are shifted in time relatively by a quarter of period. So, the extremal values of
the fields are achieved by turns. The action of the fields on the motion of the repre-
sentative points is mostly significant at the extremal values, hence, it is reasonable
to specify them as the same like those in the original version of the model. Because
of structural stability of the hyperbolic attractor, one can hope that it survives the
modification, at least in a properly selected range of the parameters. As seen from
the computations, it is so, e.g. at ε = 0.72 and K = 1.9.
In contrast to the previous version of the model, the Poincaré map can not be ex-
pressed analytically. However, it may be easily performed with a computer program
integrating the equations with a finite-difference method in a time period T = 4.
We can exclude a redundant variable in the equations by means of the variable
change (3.28), and rewrite them in terms of W = X + iY. After separation of the real
and imaginary parts, we obtain
dX π π π π
= −2ε Y 2 Ω1 (X,Y,t) cos cos t − X sin cos t
dt 4 2 4 2
π π π π πt
+KY Ω2 (X,Y,t) cos sin t − X sin sin t sin ,
4 2 4 2 2

dY π π 1 π π
= 2ε Y Ω 1 (X,Y,t) X cos cos t + (1 − X + Y ) sin
2 2 cos t
dt 4 2 2 4 2
π
π 1 π π πt
−K Ω2 (X,Y,t) X cos sin t + (1 − X + Y ) sin
2 2 sin t sin ,
4 2 2 4 2 2
(3.36)
where
π π π π
2X cos cos t + (1 − X 2 − Y 2 ) sin cos t
Ω1 (X,Y,t) = 4 2 4 2 ,
(1 + X 2 + Y 2 )2
π π π π
(3.37)
−2X sin sin t + (1 − X 2 − Y 2 ) cos sin t
Ω2 (X,Y,t) = 4 2 4 2 + √1 .
1 + X +Y
2 2
2
The expressions (3.36), (3.37) look awkward, but this is the first explicit example of
a set of differential equations with smooth coefficients, which has attractor of Plykin
type on a plane in the Poincar é cross-section.
Figure 3.18 shows variables X and Y in dependence on time as obtained from
numerical integration of the differential equations (3.36) at ε = 0.72 and K = 1.9,
after exclusion of transients. Visually they look like samples of random processes,
as they should be for the dynamics on the chaotic attractor.
Fig. 3.18 Variables X and Y versus time obtained numerically from integration of the differential
equations (3.36) at ε =0.72 and K=1.9. The plot relates to sustained chaotic regime associated with
motion on the attractor; the transients are excluded.
Figure 3.19 shows portrait of the attractor in the three-dimensional extended

phase space (X, Y , t). To make the inherent structure visible, the picture is pre-
sented with the gray-scale technique. Brighter tones correspond to relatively larger
probability of visiting pixels by orbits on the attractor. Time interval on the verti-
cal axis corresponds just to a period of variation of coefficients in the equations.
In the cross-section of the attractor with the horizontal plane, one can observe an
object with fractal structure, remarkably similar to that discussed in the previous
section (Fig. 3.16(a)). At a qualitative level, it may be regarded as an argument in
favor of persistence of the hyperbolic attractor under modification of the model we
undertake.
Computation of the Lyapunov exponents for the model (3.36) by means of the
Benettin algorithm (Appendix A) at ε = 0.72 and K = 1.9 yields λ 1 ≈ 0.221 and
Λ2 ≈ −0.315. It corresponds to the Lyapunov exponents of the Poincaré map Λ 1 =
λ1 T ≈ 0.884 and Λ 2 = λ2 T ≈ −1.260. Estimate of the attractor dimension in the
Poincaré section by the Kaplan-Yorke formula is D KY ≈ 1.70.
The hyperbolic nature of the attractor was verified by means of graphical repre-
sentation of manifolds in the Poincaré section (like in Fig. 3.16(b)) and the compu-
tations based on the cone criterion (Kuznetsov, 2009a,b). Following those results,
the attractor is hyperbolic in some parameter range around ε = 0.72 and K = 1.9.
References 79
Fig. 3.19 Portrait of the attractor of the model (3.36) at ε = 0.72 and K = 1.9 in the extended three-
dimensional phase space. The gray-scale technique is used: brighter tones correspond to areas of
relatively large probability of visiting by orbits on the attractor.
References
Heagy J.F.: A physical interpretation of the Hénon map. Physica D 57, 436–446 (1992).
Hunt, T.J.: Low dimensional dynamics: bifurcations of Cantori and realisations of uniform hyper-
bolicity. PhD Thesis. Univercity of Cambridge (2000).
Kuznetsov, A.P., Savin, A.V., Savin, D.V.: On some properties of nearly conservative dynamics of
Ikeda map and its relation with the conservative case. Physica A 387, 1464-1474 (2008).
Kuznetsov, S.P.: Hyperbolic strange attractors of physically realizable systems. Izvestija VUZov-
Applied Nonlinear Dynamics (Saratov) 17 (4), 5-34 (2009a).
Kuznetsov, S.P.: A non-autonomous flow system with Plykin type attractor. CNSNS 14, 3487–3491
(2009b).
Kuznetsov, S.P.: An example of a non-autonomous continuous-time system with attractor of Plykin
type in the Poincaré map. Nonlinear Dynamics 5, 403-424 (2009c).
Kuznetsov, S.P., Turukina, L.V.: Attractors of Smale-Williams type in periodically kicked model
systems. Izvestija VUZov-Applied Nonlinear Dynamics (Saratov) 18, 80-92 (2010).
Kuznetsov, S.P.: Dynamical chaos and uniformly hyperbolic attractors: from mathematics to
physics. Physics-Uspekhi 54 (2), 119 (2011).
Sagdeev, R.Z., Usikov, D.A., Zaslavsky, G.M.: Nonlinear physics: from the pendulum to turbulence
and chaos. CRC Press, Boca Raton, FLA (1988).
Chapter 4
Non-Autonomous Systems of Coupled
Self-Oscillators
Abstract In this chapter we examine models composed of coupled limit-cycle self-

oscillators. We start with discussion of a single van der Pol oscillator and pay at-
tention, particularly, to its description in terms of slow complex amplitudes. This
oscillator will be used as a building block for design of models with hyperbolic
attractors. Some years ago the author suggested a system admitting physical real-
ization, in which the Poincaré map possesses the uniformly hyperbolic attractor of
Smale-Williams type (Kuznetsov, 2005). It is composed of two van der Pol oscil-
lators, which become active by turns because of external modulation of parameters
responsible for the birth of the limit cycle. The partial oscillators pass the excitation
each other in such a way that on a whole cycle of this transfer the phase of the os-
cillations undergoes transformation corresponding to the expanding circle map. The
last section of the chapter is devoted to another approach: it will be shown how the
dynamics associated with the Plykin type attractor considered in the previous chap-
ter may be realized in a system of two coupled self-oscillators described in terms of
complex amplitudes.
4.1 Van der Pol oscillator
Let us start with discussion of the van der Pol oscillator, which will be used as a
building block for the systems discussed in the main part of this chapter. This model
is well known, popular and significant in nonlinear science because it represents
the simplest example of self-oscillatory system able to manifest sustained periodic
oscillations of magnitude, period, and wave-form determined by the system itself,
not by some external force. Originally, this model was designed by Dutch physicist
and radio-engineer B. van der Pol in application to an electronic generator of non-
dumping radio-frequency oscillations. Later, it acquired status of a universal model
of theory of oscillations used for description of systems of different physical nature
(Andronov et al., 1966; Davis, 1962; Hayashi, 1964; Rabinovich and Trubetskov,
1989; Landa, 1996; Strogatz, 2001; Pikovsky et al., 2002).
82 4 Non-Autonomous Systems of Coupled Self-Oscillators
The van der Pol oscillator is a system governed by differential equation of the
second order
ẍ − (A − x2)ẋ + ω02 x = 0. (4.1)
Here ω0 is characteristic frequency; A is control parameter that determines character
of dynamical regime of the system. An instant state of the system is specified by two
dynamical variables, the generalized coordinate x (its concrete physical sense de-
pends on the problem in consideration, say, in electronics it is some time-dependent
voltage) and the generalized velocity ẋ. So, the states are represented by points on
the phase plane (x, ẋ).
At A <0 the system has a unique attractor on the phase plane, the stable fixed
point at the origin (Fig. 4.1(a)). At A >0 the equilibrium state in the origin becomes
unstable; an attractive closed orbit appears nearby, which is the stable limit cycle
corresponding to the self-oscillations in the system. At small positive A the oscil-
√
lations are close to sinusoid with frequency ω 0 and amplitude proportional to A
(Fig. 4.1(b)). At large A the oscillations accept some peculiar form (called the re-
laxation oscillations), and their main frequency undergoes graduate decrease with
growth of A (Fig. 4.1(c)). The phenomenon of birth of the limit cycle at A=0 is a par-
ticular case of the normal (or supercritical) Andronov-Hopf bifurcation (Andronov
et al., 1966; Kuznetsov, 1998).
Fig. 4.1 Phase portraits for the van der Pol oscillator (4.1) at ω0 = 2π , A = −1 (a), A = 1 (b),
and A = 5 (c) on the phase plane (x, ẋ). Transient trajectories are shown in gray, and attractor (the
stable fixed point in the first, and the limit circles on the second and the third pictures) in black.
A useful and luminous approach to analytical description of the model is based

on the method of slow complex amplitudes. The early version of this approach was
developed by van der Pol himself; later important contributions were made by An-
dronov, Bogolyubov and Mitropolsky, and others (Andronov et al., 1966; Nayfeh,
2000; Grebenikov, 2004). It is convenient to explain this method in the version we
use hereafter.
Let us set
x = a(t)eiω0t + a∗ (t)e−iω0t , (4.2)
where a(t) is a complex-valued function of time, and assume
4.1 Van der Pol oscillator 83
ẋ = iω0 aeiω0t − ω0 a∗ e−iω0t , (4.3)
which implies that

ȧeiω0t + ȧ∗ e−iω0t = 0. (4.4)
(The relation (4.2) introduces two unknown functions, the real and the imaginary
parts of the complex amplitude, instead of one real function x(t); so, we are in our
right to impose one additional functional condition that is just the relation (4.4).)
Then, accounting (4.4), we get
ẍ + ω02 x = iω0 ȧeiω0t − iω0 ȧ∗ e−iω0t = 2iω0 ȧeiω0t . (4.5)
So, Equation (4.1) may be rewritten through the complex amplitudes as

1
ȧ = [A − (aeiω0t + a∗e−iω0t )2 ](aeiω0t − a∗e−iω0t )e−iω0t . (4.6)
2
So far, all the algebra transformations have been rigorous, but now we will turn to
approximation and assume that the complex amplitudes are slow-varying functions
of time. To account this, we perform averaging in Eq. (4.6) over a period of the
relatively fast oscillations of frequency ω 0 . Practically, this may be done simply by
dropping all terms on the right-hand side, which contain fast time dependencies like
e±iω0t , e±2iω0t , etc. Then, we obtain what is called the amplitude equation, or the
shortened van der Pol equation
ȧ = 12 (A − |a|2)a. (4.7)
Advantage over the original model is that the last equation is analytically solvable.
Namely, if the initial complex amplitude at t=0 equals a 0 , then the explicit solution
is expressed as
a0 eAt/2
a(t) = . (4.8)
1 + |a0|2 (eAt − 1)/A
Observe that in the limit t → ∞ in accordance with this relation √the solution at A <0
decays to zero, and at A >0 it saturates at finite amplitude |a| = A that corresponds
to the stable limit cycle.
One can regard Eq. (4.7) not only as a certain approximation for the original
system (4.1), but as a self-contained object for studies and analysis. Sometimes,
these or similar equations can appear in much more wide context than the original
van der Pol model. A good example is the so-called Landau-Stewart equation for
description of initial stages of development of turbulence (Landau and Lifshitz, 1959
Kuramoto, 1984; Strogatz, 2001).
4.2 Smale-Williams attractor in a non-autonomous system of

alternately excited van der Pol oscillators
For an attractor of Smale-Williams type to occur, of principal significance is pres-

ence of some angular variable, which is multiplied by factor of two (or by a larger
integer) under the action of the Poincaré map, while in other directions the phase
volume undergoes compression.
Let us assume that the angular variable is a phase of some oscillatory process.
Can we design a non-autonomous system with expanding transformation for the
phase after each period of variation of coefficients in the differential equations? It is
not a so trivial task because transformations of oscillatory signals produce usually
only a shift of the output phase, equivalent to a shift of the input phase; so, realiza-
tion of the expanding circle map seems problematic. Nevertheless, a solution does
exist: we can use two or more oscillatory elements, which transfer excitation each
other in such a way that the required transformation for phases takes place on the
whole cycle of the transfer. 1 It is worth noticing that in this situation each oscil-
lator cannot be characterized by globally defined phase during the whole process;
otherwise expanding map for the phase would be impossible.
Consider two coupled van der Pol oscillators of frequencies ω 0 and 2ω0 , respec-
tively, governed by equations
ẍ − (A cos(2π t/T ) − x2 )ẋ + ω02 x = ε y cos ω0t,

(4.9)
ÿ − (−A cos(2π t/T ) − y2 )ẏ + 4ω02y = ε x2 .
The generalized coordinates x and y relate to two subsystems activated by turns due
to the externally forced modulation of the parameter responsible for the bifurcation
of birth of the limit cycle in a single oscillator. Its magnitude is determined by the
parameter A. Interaction between the subsystems is characterized by coupling con-
stant ε . The first oscillator affects the second one due to the quadratic term ε x 2 in
the right-hand part of the second equation. The backward coupling is introduced
by a product of the dynamical variable and the auxiliary signal of frequency ω 0
(the right-hand term in the first equation). It is assumed that a period of the pa-
rameter modulation contains an integer number of periods of the auxiliary signal
T = 2π N/ω0. Thus, the external driving is periodic, and description of the dynam-
ics in terms of the stroboscopic Poincaré map is appropriate.
Equations (4.9) may be rewritten as a set of equations of the first order
1 On a base of analogous approach, feasible systems were suggested also for implementation of
some phenomena of complex dynamics discussed earlier only on a level of abstract constructions.
It relates to a robust strange non-chaotic attractor of Hunt and Ott type (Hunt and Ott, 2001; Kim
et al., 2003) considered in (Jalnine and Kuznetsov, 2007), and phenomena of complex analytic
dynamics, like Mandelbrot and Julia sets (Devaney, 2003) considered in (Isaeva et al., 2007; 2008).
4.2 Smale-Williams attractor in a non-autonomous system of . . . 85
ẋ = ω0 u,
u̇ = (A cos(2π t/T ) − x2 )u − ω0x + εω0−1 y cos ω0t,
ẏ = 2ω0 v, (4.10)
1
v̇ = (−A cos(2π t/T ) − y2)v − 2ω0y + εω0−1 x2 ,
2
where the variables u and v represent normalized generalized velocities of two os-
cillators. To introduce the Poincaré map, assume the state at t = t n = nT is specified
by a vector x n = {x(nT ), u(nT ), y(nT ), v(nT )}. Solution of Eqs. (4.10) in the time
interval T with the initial conditions x n generates a new vector x n+1 , and so we de-
fine the four-dimensional map x n+1 = T(xn ). Geometrically, in the five-dimensional
extended phase space of the non-autonomous system {x, u, y, v,t} it corresponds
to cross-section of the flow of trajectories by the four-dimensional hyper-planes
t = tn = nT .
Qualitatively, the system operates as follows. Let the first oscillator have some
phase ϕ at a stage of activity: x ∼ cos(ω 0t + ϕ ). The squared variable x 2 contains
the second harmonic component because of the trigonometric equality
1 1
x2 ∼ cos2 (ω0t + ϕ ) = + cos(2ω0t + 2ϕ ), (4.11)
2 2
and the phase of this component is 2ϕ . As the half-period comes to the end, the term
x2 stimulates excitation of the second oscillator, and oscillations of the variable y
get the phase 2ϕ . Half a period later, the mixture of these oscillations with the
auxiliary signal stimulates excitation of the first oscillator, which accepts now this
phase 2ϕ . Thus, in subsequent periods of the parameter modulation, the phase of
the first oscillator follows approximately the relation
ϕn+1 = 2ϕn + const ( mod 2π ). (4.12)
Here the constant term accounts phase shifts in the course of the transfer of the
excitation from one oscillator to another; this term may be removed by displacement
of the origin for the phase variable. The relation (4.12) is the expanding circle map,
or Bernoulli map, which is commonly known as the simplest model example in the
chaos theory.
We emphasize that the phase ϕ cannot be determined as continuous variable in
the whole time interval T . Indeed, at the stage when the oscillator is suppressed, its
amplitude is small, and the phase is not well defined.
The process we have described indeed occurs in numerical simulations in a wide
parameter range. Figure 4.2 shows a typical sample of time dependence for variables
x and y obtained from numerical solution of Eqs. (4.9) at particular parameter values
ω0 = 2π , T = 6, A = 5, ε = 0.5. Being represented by periodic alternation of exci-
tation and suppression for the first and the second oscillators, the process, in fact, is
chaotic. Chaos reveals itself in variations of phases of the oscillators at successive
stages of activity. More accurate analysis shows that these variations of the phases
follow the Bernoulli map (in reasonable approximation). In Fig. 4.3 an empirical
Fig. 4.2 Typical sample of time dependence for the generalized coordinates x and y obtained from
numerical solution of Eqs. (4.9) at ω0 = 2π , T = 6, A = 5, ε = 0.5.
diagram is shown for the phases determined in the process of numerical simulations
at successive excitation stages for the first oscillator as ϕ n = arg(x(nT ) − iu(nT ))
and plotted for sufficiently large number of the basic periods. The mapping for the
phase surely looks topologically equivalent to the Bernoulli map (4.12). Some dis-
tortions arise due to imperfection of the above qualitative considerations and of the
definition of the phase; the correspondence becomes better at larger period ratios
N. Of principal significance is topological nature of the phase transformation: one
complete bypass for the pre-image ϕ n (that is variation by 2π ) corresponds to the
two-fold bypass for the image ϕ n+1 .
Fig. 4.3 Diagram illustrating transformation of the phase for the first oscillator from one stage of
activity to the next one obtained from numerical solution of Eqs. (4.9) at ω0 = 2π , T = 6, A = 5,
ε = 0.5.
For computation of Lyapunov exponents one integrates Eqs. (4.10) together with
collection of four replicas of variation equations
x̃˙ = ω0 ũ,
ũ˙ = (A cos(2π t/T ) − x2 )ũ − 2xx̃u − ω 0x̃ + εω0−1 ỹ cos ω0t,
(4.13)
ỹ˙ = 2ω0 ṽ,
ṽ˙ = (−A cos(2π t/T ) − y2 )ṽ − 2yỹv − 2ω0ỹ + εω0−1 xx̃,
and the procedure is supplemented with normalization and orthogonalization of four

perturbation vectors (as explained in Appendix A) at each period T . The Lyapunov
exponents are evaluated as average rates of growth or decrease of accumulating
sums of logarithms of norms of the perturbation vectors before the normaliza-

tions. According to the calculations, the Lyapunov exponents for the attractor of
the Poincaré map at the specified parameters are the following
Λ1 = 0.6832, Λ2 = −2.602, Λ3 = −4.605, Λ4 = −6.538. (4.14)
The positive exponent Λ 1 indicates chaotic nature of the dynamics. It is close to

ln 2 = 0.693 . . ., which is in good correspondence with the approximate description
of the evolution for the phase variable in terms of the one-dimensional Bernoulli
map (4.12). Estimate of the Kaplan-Yorke dimension of the attractor through the
Lyapunov exponents yields D ≈ 1.263.
Action of the Poincaré map T in the four-dimensional space is accompanied by
expansion in the directions locally associated with the angular variable, or phase, in
Eq. (4.12), and by compression in the rest three phase space directions. Empirically,
one can indicate a toroidal absorbing domain D, which is topologically a product of
a one-dimensional circle and a three-dimensional ball. An iteration of the Poincaré
map in application to the representative points from this domain generates an object
T(D) in a form of a closed “tube” that is stretched in length, compressed in width and
folded in double loop inside the original toroidal domain. It precisely corresponds
to the Smale-Williams construction, but in the four-dimensional phase space.
To write down an analytic expression for the domain D it is convenient to redefine
the coordinate system and introduce new variables {x 0 , x1 , x2 , x3 } as follows:
x0 = x/r0 , x1 = (u − cuxx)/r1 ,
(4.15)
x2 = y − cyxx − cyu u, x3 = v − cvx x − cvuu − cvyy,
where cux , cyx , cyu , cvx , cvu , cvy , r0 , r1 are some constants, which can be determined
for concrete mode of operation of the system as follows. First, by numerical solution
of Eqs. (4.9) we accumulate a large number of points {x, u, y, v} on the attractor in
the Poincaré section. Then, by the least square method we determine the coefficients
to minimize the mean-square values < (u − c ux x)2 >, < (y − cyx x − cyu u)2 >, < (v −
cvx x − cvu u − cvy y)2 >. Geometrically, it means directing the coordinate axes along
the principal axes of ellipsoid that roughly approximates the attractor. Additionally,
we normalize the first two variables x 0 and x1 by factors r0 and r1 to have < x20 >=<
x21 >≈ 1/2. Finally, at the parameters of the system ω 0 = 2π , T = 6, A = 5, ε = 0.5
we get
cux = 0.438, cyx = −0.042, cyu = 0.226, cvx = −0.218,

cvu = 0.029, cvy = −0.118, r0 = 0.812, r1 = 0.721.
In the new coordinates the absorbing domain D is defined by the inequality:

2
x20 + x21 − r dr + (x2 /d)2 + (x3 /d)2 1. (4.16)
Here the empirically selected constants are r = 0.94, d r = 0.4, d = 0.15.

Figure 4.4 shows a three-dimensional projection of the objects in the four-
dimensional space to give impression of how the image T(D) is embedded inside
the original toroid D. The mutual location of the domains D and T(D) is analogous
to that at the first step of the construction of the Smale-Williams attractor. The do-
main T(D) looks like a narrow band because of strong transversal compression in
the course of evolution in one period of the parameter modulation.
Fig. 4.4 Absorbing domain D and its image T(D) under the Poincaré map in three-dimensional
projection.
It is interesting to examine process of deformation of the toroidal domain in

continuous time that ensures the embedding of the image inside the original domain
being folded in a double loop. It is illustrated by a series of pictures in Fig. 4.5.
Each of them represents projection of the four-dimensional object into the three-
dimensional space of variables {x 0 , x1 , x2 } along the axis of the fourth variable x 3 .
The first picture (a) corresponds to initial time instant t = 0 and depicts the original
toroidal domain. Then, this toroid is compressed strongly in transversal directions
(b). The compression differs in different directions, and the image takes a form
of a closed narrow band. Then, the image undergoes expansion in the longitudinal
direction and deformation of such kind that the double loop is formed ((c), (d)).
At final stages of the process, this double loop widens in size and appears to be
placed inside the original domain ((e), (f)). Actually, the process reproduces the
continuous deformations of plastic doughnut used for popular explanation of the
Smale-Williams construction as considered in Chap. 1.
Figure 4.6 shows portraits of attractor on a plane of variables of the first oscillator.
The panel (a) represents a projection of the attractor from the five-dimensional ex-
tended phase space on the plane of original variables (x, u). The portrait is depicted
in gray scales (the darkness reflects a relative duration of presence of the orbit inside
a given pixel). Black dots relate to the stroboscopic Poincaré cross-section, i.e., to
the instants tn = nT . The panel (b) shows the attractor in the Poincaré cross-section
Fig. 4.5 Illustration of transformation of the original toroidal domain (a) in the course of the
continuous time evolution on one period T ((b)-(f)). Each picture is a three-dimensional projection
for the four-dimensional object into the space of variables {x0 , x1 , x2 } along the axis of the fourth
variable x3 . Notice essentially different scales along the coordinate axes on the diagrams.
on the plane of the redefined coordinates (x 0 , x1 ) (see (4.15)). The transverse Cantor-
like structure is illustrated separately by magnified fragments of the initial picture.
Note evident visual similarity with the Smale-Williams attractor (Fig. 1.5).
Fig. 4.6 Attractor of the system of coupled non-autonomous van der Pol oscillator in projection
onto phase plane of the first oscillator in the original variables and portrait in the stroboscopic
cross-section on the plane (x0 , x1 ).
For quantitative characterization of the fractal structure of the attractor in the

Poincaré cross-section, the correlation dimension was estimated in computations by
means of the algorithm of Grassberger and Procaccia (Grassberger and Procaccia,
1983; Tong, 1994; Schuster, H.G. and Just, 2005; Hilborn,2000; Ott, 2002). A 4-
component time series x n = {x0 (tn ), x1 (tn ), x2 (tn ), x3 (tn )} (n=1÷M, M=40000) was
obtained from numerical iteration of the Poincaré map. The resulting dimension
D2 ≈ 1.252 may be compared with the estimate from the Kaplan-Yorke formula
D ≈ 1.263.
In Fig. 4.7 all four Lyapunov exponents are plotted versus the amplitude of
slow modulation A at fixed other parameters. As seen, the largest exponent re-
mains almost constant in a wide range of the parameter, and is close to the value
ln 2 = 0.693 . . . characteristic to the Bernoulli map. Such behavior of the largest
Lyapunov exponent responsible for the chaotic nature of the dynamics agrees with
structural stability of the hyperbolic attractor, which evidently persists in this pa-
rameter interval (for comparison, see plots of the Lyapunov exponents for models,
which do not possess the structural stability in Appendix B). The rest exponents are
large negative. They correspond to strong compression of the phase volume along
three of four dimensions of the phase space of the Poincaré map.
For the system (4.9) with the used parameter values the computations were car-
ried out to confirm hyperbolicity (see Chap. 7). It was demonstrated in computations
that the cone criterion is valid in the toroidal domain, containing the attractor of the
Poincaré map (Kuznetsov and Sataev, 2007). Recently, the hyperbolic nature of the
attractor in this system has been established rigorously, on the basis of the computer-
assisted proof (Wilczak, 2010).
Fig. 4.7 Lyapunov exponents of the Poincaré map for the system (4.9) plotted versus the amplitude
of slow modulation A at fixed other parameters ω0 = 2π , T = 6, ε = 0.5.
Fig. 4.8 Power spectrum generated by the system (4.9) at ω0 = 2π , T = 6, ε = 0.5 obtained by
processing data of the numerical simulation: the square root of the spectral power density (a) and
the spectral power density in logarithmic scale (b).
Figure 4.8 shows power spectra of two oscillators constituting the system (4.9)
as obtained from processing time series for x and y sampled from data of numerical
simulation of the dynamics on the attractor. The spectrum is computed according to
the method of statistical evaluation of the spectral power density recommended in
the theory of random processes (Jenkins and Watts, 1968). The procedure includes
breaking the time series onto pieces of certain length, performing the Fourier trans-
form for each segment, and then averaging the squares of the amplitudes of spectral
components over all pieces. Panel (a) shows a plot for square root of the power
density versus frequency, and panel (b) plots this quantity in logarithmic scale, in
decibels traditionally used in electronics (10 dB correspond to 10-fold ratio of the
compared power levels). 2
As can be seen from the figure, the spectrum of one oscillator is concentrated
in a certain range near the natural frequency ω 0 , and the second near the doubled
frequency 2ω 0 . The continuous component in the power spectrum is an attribute of
chaotic dynamics. Discrete peaks in the left part of the picture appear due to the
presence of periodic component in the dynamics of the system because of the slow
modulation of the parameter.
In Appendix G features of dynamics on hyperbolic attractor under effect of ex-
ternal noise are considered using this model as an example.
It was found out that in the considering system the hard excitation and hys-
teresis can be observed in a quite wide range of parameters. To demonstrate this
phenomenon, let us supplement Eqs. (4.9) with an additional parameter h, which
regulates relative duration of excitation and damping stages of the oscillators:
ẍ − (h + A cos(2π t/T ) − x2 )ẋ + ω02 x = ε y cos ω0t,

(4.17)
ÿ − (h − A cos(2π t/T ) − y2 )ẏ + 4ω02y = ε x2 .
The above-described mode of operation corresponds to excitation of the first os-

cillator entering the activity stage by stimulation from the partner. The respective
driving force is proportional to the squared amplitude of oscillations of the partner
at the previous stage of activity. If one sets the initial conditions with characteristic
amplitude small enough, the excitation of the whole system does not occur because
of negligible level of the squared amplitude, and the oscillations do not develop.
Generation regime can be restored by increasing parameter h, i.e. by increase of
the part of the modulation period, during which the oscillators are linearly unstable.
With gradual growth of h, at some point one observes an abrupt transition to the
large amplitude chaotic regime. Now, if one starts to decrease h, it does not lead
immediately to breakdown of the chaotic oscillations since the excitation in the sys-
tem is transferred from the previous stages of activity to the next ones. Only at large
enough negative values of h the oscillations decay. The hysteresis phenomenon is
illustrated in Fig. 4.9. In the course of computation, small random perturbation was
added at each step of the numeric integration on a level of order Δx ∼ 0.0001 to
ensure departure of the system from the trivial stationary state at zero as it becomes
unstable. The insets in Fig. 4.9 show portraits of attractors in the Poincaré map at
representative points of the hysteresis loop, from which one can see that its type
corresponds to the Smale-Williams solenoid.
2 The first form of presentation is selected to make possible a qualitative comparison with ex-
perimental data in Chap. 12 available in this scale; the second is used to make possible visual
comparison with data relating to time-delay systems in experiment in Chap. 13 and in numerics in
Chap. 10.
4.3 System of alternately excited van der Pol oscillators in terms of slow complex amplitudes 93
Fig. 4.9 A diagram illustrating the hard excitation and hysteresis in the system: Dependence of
the maximum amplitude of the first oscillator on the parameter h in (4.17) at ω0 = 2π , T = 6,
A = 5, ε =0.5. The parameter h is varied, first, from left to right that corresponds to motion along
the lower branch, and then from right to left that corresponds to motion along the upper branch of
the hysteresis loop.
4.3 System of alternately excited van der Pol oscillators in terms

of slow complex amplitudes
In assumption that N = ω 0 T /2π 1, and for parameters A and ε being not too
large, the problem may be reformulated in terms of slow complex amplitudes, with
decrease of a number of relevant control parameters. For this, we set
x = aeiω0t + a∗e−iω0t , ẋ = iω0 aeiω0t − iω0 a∗ e−iω0t ,

(4.18)
y = be2iω0t + b∗ e−2iω0t , ẏ = 2iω0 be2iω0t − 2iω0b∗ e−2iω0t ,
and up on substitution in Eqs. (4.9) and averaging we obtain

1
ȧ = (aA cos(2π t/T ) − |a|2a) − iε̄ b,
2
(4.19)
1
ḃ = (−bA cos(2π t/T ) − |b|2b) − iε̄ a2 ,
2
where ε̄ = ε /4ω0 . A study of this system was made in (Kuptsov et al., 2008). It was
shown that in a wide parameter range attractor of Smale-Williams type takes place
in the stroboscopic Poincaré map, particularly, it is the case at A=3, T =10, ε̄ = 0.05.
It may be speculated that such equations will arise in a broader context than the
approximating dynamics of the concrete model. Arguing in the spirit of theory of
onset of turbulence of Landau and Hopf (Landau and Lifshitz, 1959; Hopf, 1948),
one can imagine the following situation. Suppose we have some spatially extended
system, in which two oscillatory modes occur with frequency ratio close to 1:2 on
the background of an additional slow periodic motion; they alternately come beyond
the threshold of excitation and back, because the nonlinear interactions transfer exci-
tation to each other. Then, we have the conditions for the realization of a hyperbolic
attractor of Smale-Williams type.
4.4 Non-resonance excitation transfer
As mentioned above, the system (4.9) is designed in such a way that the coupling
terms contain frequency components in resonance with the oscillator to which the
excitation is passed. It occurs that attractor of Smale-Williams type can arise in the
system of two alternately active self-oscillators as well with non-resonance transfer
of the excitation (Kuznetsov et al., 2007). Indeed, accounting finite characteristic
time of excitation of the partner oscillator, the spectrum of signal ensuring the exci-
tation transfer occupies a band of some width around the central frequency having
decaying tails on both sides. If, say, the central frequency is 2ω 0 , and natural fre-
quency of the excited oscillator is ω 0 , then, under certain conditions presence of the
components on the tail of the spectral distribution near ω 0 may arrange the excitation
transfer, accounting that on the activity stage at finite super-criticality the growth of
the perturbation is exponential. Of course, in this case conditions of parameters of
the system are essentially more restricted. 3
As an example, consider a system
ẍ − (A cos(2π t/T ) − x2)ẋ + ω02 x = ε y,

(4.20)
ÿ − (−A cos(2π t/T ) − y2)ẏ + 4ω02y = ε x2 ,
which operates without the auxiliary signal in contrast to (4.9). According to re-
sults of numerical simulation of the dynamics at T = 10, ω 0 = 2π , ε = 0.5, A = 6,
attractor of Smale-Williams type occurs in the stroboscopic Poincaré map, although
the transfer of the excitation from the second oscillator to the first one is of non-
resonance character. Figure 4.10 shows the iteration diagram for phases at succes-
sive stages of activity of the first oscillator, the attractor in projection from extended
phase space onto the plane (x, ẋ), and the portrait of the attractor in the stroboscopic
Poincaré map. Lyapunov exponents estimated numerically for the Poincaré map in
this regime are Λ1 = 0.6808, Λ 2 = −3.625, Λ3 = −8.326, Λ4 = −17.633, and esti-
mate of the Kaplan-Yorke dimension yields D KY = 1.188.
3 Apparently, in many applications, say, in microwave electronics, laser physics, nonlinear optics,
the resonance mechanism should be regarded as preferable. Indeed, it seems problematic that non-
resonance transfer may be arranged on a level significantly exceeding noises under frequency larger
in order or more than the modulation frequency.
4.5 Plykin-like attractor in non-autonomous coupled oscillators 95
Fig. 4.10 Diagram for phases determined from formula ϕ = arg(x − iẋ/ω0 ) at successive instants
t = nT (a), attractor in projection onto the phase plane of the first oscillator (b), and portrait in the
Poincaré section (c) for the system with non-resonance excitation transfer governed by Eqs. (4.20)
at T = 10, ω0 = 2π , ε = 0.5, A = 6.
4.5 Plykin-like attractor in non-autonomous coupled oscillators
In the present section it will be shown how to implement the dynamics associated
with attractor of Plykin type in a system of coupled non-autonomous oscillators.
As believed, it opens prospects for design of physical and technical systems, e.g.
electronic devices with Plykin type attractors (Kuznetsov, 2009).
4.5.1 Representation of states on a sphere and equations of the

model
Let us start with a system of two self-oscillators with compensation of losses from
the common energy source; the equations for slow complex amplitudes a and b read
1
ȧ = μ (1 − |a|2 − |b|2)a,
2
(4.21)
1
ḃ = μ (1 − |a|2 − |b|2)b,
2
where μ is a positive parameter. Let us set
√ iϕ /2+iψ √ −iϕ /2+iψ
b= ρe sin(θ /2), a= ρe cos(θ /2). (4.22)
Clearly, in the sustained regime of the self-oscillations, the equality ρ = |a| 2 + |b|2 =
1 has to be valid. Identifying the states of the system, differing only in the overall
phase, under condition ρ = 1 one can associate them with points on a unit sphere
(Fig. 3.12). Also, the Cartesian coordinates may be introduced
x = ρ cos ϕ sin θ , y = ρ sin ϕ sin θ , z = ρ cos θ . (4.23)

Via the complex amplitudes they are expressed as
x + iy = 2a∗ b, z = |a|2 − |b|2 . (4.24)
We are going to modify the model (4.21) in order to obtain a set of equations
with coefficients periodically varying in time, in such a way that in the stroboscopic
description the Plykin type attractor will occur.
Let us consider a sequence of continuous transformations on the sphere, which
is equivalent in a whole period to those discussed in Sect. 3.3. However, in order
to simplify the equations written for the complex amplitudes, it is more convenient
here to perform the transformations in six rather than in four stages. Duration of
each of the stages is accepted equal to a unit time interval.
I. Flow down along circles of latitude. At this stage we require the angular
velocity of representative points on the sphere to be proportional to sin 2ϕ . An ap-
propriate form of the differential equations is
ȧ = −iε aIm(a ∗ b)2 , ḃ = iε bIm(a∗ b)2 . (4.25)
Indeed, substituting b = e iϕ /2+iψ sin(θ /2), a = e−iϕ /2+iψ cos(θ /2), after some sim-
1
ple manipulations we get ϕ̇ = ε sin2 θ sin 2ϕ , θ̇ = 0. Physically, the right-hand
2
terms in (4.25) ensure a frequency shift of opposite sign for two oscillators. Magni-
tude of the shift is proportional to the amplitude of a low-frequency signal produced
from mixing of the second harmonic components of both oscillators on a quadratic
nonlinear element.
II. Differential rotation around z-axis with angular velocity depending on z. At
this stage we set
1 √ √ 1 √ √
ȧ = − iπ ( 2 − 1 − 2 2|a|2 )a, ḃ = − iπ ( 2 + 1 − 2 2|b|2 )b. (4.26)
4 4
1 √
In angular variables (ϕ , θ ), these equations reduce to ϕ̇ = − π ( 2 cos θ + 1),
2
θ̇ = 0. Note√that the angular velocity ϕ̇ depends linearly on z = cos θ and vanishes
at z = −1/ 2. At this stage two subsystems behave like uncoupled classic non-
isochronous oscillators. At small amplitudes their frequencies are detuned from the
1 √
reference point by Δω a,b = − iπ ( 2 ∓ 1). With growth of the amplitudes, the fre-
4
quencies undergo a shift proportional to the squared amplitude for both subsystems.
III. Rotation of the sphere by 90 ◦ around the y-axis in accordance with the
equations
1 1
ȧ = π b, ḃ = − π a. (4.27)
4 4
It corresponds to a conservative system of coupled oscillators with equal frequen-
cies, with the coupling coefficient of such value that the energy exchange between
the partial oscillators corresponds precisely to the duration of the stage.
IV. Flow down along circles of latitude, like at the first stage:
ȧ = −iε aIm(a ∗ b)2 , ḃ = iε bIm(a∗ b)2 . (4.28)
V. Inverse differential rotation around z-axis:

1 √ √ 1 √ √
ȧ = iπ ( 2 − 1 − 2 2|a|2 )a, ḃ = iπ ( 2 + 1 − 2 2|b|2 )b. (4.29)
4 4
VI. Inverse rotation by 90 ◦ around the y-axis:
1 1
ȧ = − π b, ḃ = π a. (4.30)
4 4
The procedure is symmetric in the sense that the operations for the stages (I) and
(IV) are identical, while the stages (II) and (III) differ from (V) and (VI) only by
directions of the rotations.
Now, we can write down equations for the complex amplitudes embracing the
complete time period T = 6. For this, we compose the right-hand sides as combina-
tions of terms (4.25), (4.26), (4.27), which are supposed to be switched on, or off,
during the respective stages of the time evolution. As to the terms from (4.21), we
account them only at the stages of rotation. Their exclusion for other stages is not so
significant, but we do so, because it simplifies the derivation of the Poincaré map.
Thus, we arrive at the equations
ȧ = −iε (1 − σ 2 − s2 )Im(a2 b∗2 )a

1 √ √ π 1
+ iσ π ( 2 − 1 − 2 2|a|2 )a − sb + s2 μ (1 − |a|2 − |b|2)a,
4 4 2
(4.31)
ḃ = iε (1 − σ 2 − s2 )Im(a2 b∗2 )b
1 √ √ π 1
+ iσ π ( 2 + 1 − 2 2|b|2 )b + sa + s2 μ (1 − |a|2 − |b|2)b.
4 4 2
Here the factors σ and s depend on time with period T = 6, and on a single period
are defined by the relations
⎧ ⎧
⎨ −1, 1 t < 2 ⎨ −1, 2t <3
σ= 1, 4 t < 5, s = 1, 5 t < 6, (4.32)
⎩ ⎩
0, otherwise, 0, otherwise.
Let us derive the Poincaré map, which determines transformation of the state
over one period T =6 and delivers stroboscopic description of the time evolution.
Let the initial conditions for Eqs. (4.31) at t n = nT be defined as the state vector
Xn = (an , bn ). Solving the equations in each successive unit interval analytically,
we can represent the map explicitly:
Xn+1 = G+ (G− (X)) ≡ G(X), (4.33)
where
⎧ √ √ √ √ ⎫
aD(a, b)e± 4 iπ ( 2−1−2 2|a| ) ∓ bD∗ (a, b)e± 4 iπ ( 2+1−2 2|b| )
1 2 1 2
⎪
⎪ 1μ ⎪
⎪
⎪
⎪ e 2 ⎪
⎪
⎨ 1 + (|a|2 + |b|2)(eμ − 1) ⎬
G± (X) = √ √ √ √ ,
⎪
⎪ ⎪
1μ ⎪
± 41 iπ ( 2−1−2 2|a|2 ) ∗ (a, b)e± 41 iπ ( 2+1−2 2|b|2 )
⎪
⎪ ±aD(a, b)e + bD ⎪
⎩ e2 ⎪⎭
1 + (|a|2 + |b|2)(eμ − 1)
(4.34)
and
14
1 |a|2 |b|2 − (a∗ b)2 tanh(2ε |a|2 |b|2 )
D(a, b) = √ . (4.35)
2 |a|2 |b|2 − (ab∗)2 tanh(2ε |a|2 |b|2 )
4.5.2 Numerical results for the coupled oscillators
Figure 4.11 shows plots for the amplitudes of the coupled oscillators |a| and |b|
versus time obtained from numerical solution of the differential equations (4.31).
As initial conditions, some small in absolute value and random in phase complex
amplitudes a and b are taken, so the plot depicts the transient evolution prior to
the regime of chaotic self-oscillations. In the right-hand part of the diagram the
dependences look like samples of a random process, that associates with motion on
the chaotic attractor. Locally, some features can be seen because of the piecewise
continuous nature of the process composed of successive stages. Particularly, the
horizontal plateaus relate to the stages of evolution (I), (II), (IV), (V), on which the
amplitudes |a| and |b| remain constant. Note that the dependence on |a| and |b| is
interconnected: in the sustained regime they obey the relation |a| 2 + |b|2 = 1.
Fig. 4.11 Plots of the real amplitudes |a| and |b| versus time in the transient process obtained from
numerical solution of the differential equations (4.31) at ε = 0.77 and μ = 1.
Fig. 4.12(a) shows portrait of the attractor of the stroboscopic Poincaré map cor-
responding to the instants t n = nT, T = 6. To make visible the intrinsic transversal
Cantor-like structure, special coordinates Re(a ∗ b), Im(a∗ b) are used. The points are
shown in black, if the amplitude of the first oscillator is larger than that of the second
oscillator, and in gray otherwise. As seen from relations (4.22), such visualization
corresponds to a definite plane projection of the attractor from the sphere. It may be
observed that this attractor exactly corresponds to the attractor of the formal model
from the previous chapter (compare Fig. 4.12(a) with 3.13(b)).
Figure 4.12(b) presents a portrait of the attractor for the continuous-time system
(4.31). As the dimension of the state space is sufficiently high (vector X = (a, b)
is four-dimensional, and the extended state space of the non-autonomous system is
five-dimensional), depicting the image to resolve subtle fractal transverse structure
intrinsic to the attractor is not a trivial task. The portrait is shown on the plane of
two values of real amplitude |a(t)| and |a(t − 3)|, which relate to instants separated
by a half of time period of variation of the coefficients in Eqs. (4.31). Presentation
of the object in gray scales is used: brighter tones correspond to pixels visited by
the representative point with higher probability. Here, again one can distinguish the
fractal-like transverse structure linked with the dynamics of the Plykin type. This
method of visualization may be appropriate in experiments with systems of the class
under consideration.
Fig. 4.12 Portrait of attractor for the Poincaré map (4.33) in the coordinates Re(a∗ b), Im(a∗ b) (a)
and attractor of the continuous-time system (4.31) on the plane of real amplitudes of one oscillator
separated by time interval T /2 = 3 in gray scales: brighter tones correspond to higher probability
of visiting the pixels by the orbit (b). The parameter values are ε = 0.77, μ = 1.
Two Lyapunov exponents for the Poincaré map are equal to those for the map
(3.24). Besides, there is a zero exponent associated with a simultaneous equal phase
shift for both oscillators, and a negative exponent corresponding to approach of
the trajectories to the invariant set |a| 2 + |b|2 = 1. Figure 4.13 presents the results
graphically. The first plot (a) shows four Lyapunov exponents Λ i in dependence on
parameter ε at fixed μ = 1. In the range ε < ε c ≈ 2.03 one of the exponents is posi-
tive that means chaos. Among other exponents one is zero (up to numerical errors),
and two are negative. Note a smooth dependence of the largest exponent on the pa-
rameter in that range. For larger ε (strong dissipation is brought in at the stages I and
IV) chaos disappears. The second plot (b) shows the Lyapunov exponents versus μ
at fixed ε = 1. Observe that variation of μ notably effects only one exponent, which
corresponds, obviously, to approach of orbits to the invariant sphere. Presence of
Fig. 4.13 Four Lyapunov exponents versus parameter ε at μ = 1 (a) and versus parameter μ at
ε = 1 (b) for the Poincaré map represented in terms of complex amplitudes. A zero exponent
occurs due to invariance of the equations in respect to the overall phase shift.
a zero exponent reflects invariance of the equations in respect to the overall phase
shift. Of course, results of computations agree with the data from the previous chap-
ter: at identical ε two of the nonzero exponents coincide (up to numerical errors)
with those obtained for the two-dimensional map.
Formally, in our complex amplitude equations the attractor should not be inter-
preted as uniformly hyperbolic, because of presence of a neutral direction in the
state space associated with the overall phase (presence of the zero-value Lyapunov
exponent in the Poincaré map). Instead, it has to be related to the class of partially
hyperbolic attractors (Pesin,2004; Bonatti et al., 2005; Pesin, 2007). Nevertheless,
in the form used here, the invariance of the equations in respect to the overall phase
is exact; it means that one can accept a legal agreement not to distinguish states
distinct only in the phase. 4 In a frame of this convention, we can treat the dynam-
ics as being true hyperbolic. However, in systems, for which description in terms
of slow complex amplitudes will be an approximation, one can expect appearance
of peculiarities associated with features of the partially hyperbolic attractor. If the
deflections from the slow-amplitude approximation are small, the overall phase will
manifest slow random walk, while the dynamics of the rest variables will retain its
character because of intrinsic robustness.
In conclusion of this section, it is worth mentioning that there are some other sit-
uations in physics where the states are naturally associated with points on a sphere.
For example, it relates to representation of two-level systems and, particularly, of
particles of spin 1/2, on the Bloch sphere (Mark Fox, 2006), or to the description
of light polarization in terms of Stokes parameters (Goldstein and Collet, 2003). We
can assume that under certain manipulations in such systems the dynamics corre-
sponding to attractors of Plykin type could be implemented.
4 Anologous reasoning was used to decrease the phase-space dimension handled in many models
considered in literature, including the derivation of the celebrating Lorenz model.
References 101
References
Andronov, A.A., Vitt, A.A., Khaı̌kin, S.Ė.: Theory of Oscillators, Pergamon Press, Oxford (1966).
Davis, H.T.: Introduction to Nonlinear Differential and Integral Equations. Dover Publications,
New York (1962).
Goldstein, D.H., Collet, E.: Polarized Light, Revised and Expanded. CRC Press, Boca Raton
(2003).
Grassberger, P., Procaccia, I.: Measuring the strangeness of strange attractors. Physica D 9, 189–
208 (1983).
Grebenikov, E., Mitropolsky, Yu.A., Ryabov, Y.A.: Asymptotic Methods in Resonance Analytical
Dynamics. CRC Press, Boca Raton (2004).
Hayashi, C.: Nonlinear Oscillations in Physical Systems. McGraw-Hill, New York (1964).
Hopf, E.: A mathematical example displaying features of turbulence. Communications on Pure
and Applied Mathematics 1, 303–322 (1948)
Hunt, B.R., Ott, E.: Fractal properties of robust strange nonchaotic attractors. Phys. Rev. Lett. 87,
254101 (2001).
Isaeva, O.B., Kuznetsov, S.P., Osbaldestin, A.H.: Complex Analytic Dynamics Phenomena in a
System of Coupled Nonautonomous Oscillators with Alternative Excitation. Tech. Phys. Lett.
33, 748–751 (2007)
Isaeva, O.B., Kuznetsov, S.P., Osbaldestin, A.H.: A system of alternately excited coupled non-
autonomous oscillators manifesting phenomena intrinsic to complex analytical maps. Physica
D 237, 873–884 (2008).
Jalnine, A.Yu., Kuznetsov, S.P.: On the realization of the hunt-ott strange nonchaotic attractor in a
physical system. Technical Physics 52, 401–408 (2007).
Jenkins, G.M., Watts, D.G.: Spectral Analysis and its Application. Holden-Day, Inc., San Francisco
(1968).
Kim, J-W., Kim, S.-Y. Hunt, B., Ott, E.: Fractal properties of robust strange nonchaotic attractors
in maps of two or more dimensions. Phys. Rev. E 67, 036211 (2003).
Kuptsov, P.V., Kuznetsov, S.P., Sataev, I.R.: Hyperbolic attractor of Smale-Williams type in a
system of two coupled non-autonomous amplitude equations. Preprint nlin. arXiv: 0804.3677
(2008).
Kuramoto, Y.: Chemical Oscillations, Waves, and Turbulence. Springer, Berlin (1984).
Kuznetsov, A.P., Kuznetsov, S.P., Pikovsky, A.S., Turukina, L.V.: Chaotic dynamics in the system
of coupling nonautonomous oscillators with resonance and nonresonance transfer of excitation.
Izvestija VUZov-Applied Nonlinear Dynamics (Saratov) 15 (6), 75–85 (2007) (In Russian).
Kuznetsov, S.P.: Example of a physical system with a hyperbolic attractor of the Smale-Williams
type. Phys. Rev. Lett. 95, 144101 (2005).
Kuznetsov, S.P., Sataev, I.R.: Hyperbolic attractor in a system of coupled non-autonomous van
der Pol oscillators: Numerical test for expanding and contracting cones. Physics Letters A 365,
97–104 (2007).
Kuznetsov, Yu.A.: Elements of Applied Bifurcation Theory. Springer, New York, Berlin, Heidel-
berg (1998).
Kuznetsov, S.P. and Sataev, I.R.: Hyperbolic attractor in a system of coupled non-autonomous van
97–104 (2007).
Landa, P.S.: Nonlinear Oscillations and Waves in Dynamical Systems. Kluwer Academic Pub-
lisher, The Netherlands (1996).
Landau, L.D., Lifshitz, E.M.: Fluid Mechanics (Course of Theoretical Physics, vol.6). Pergamon
Press, London (1959).
Mark Fox, A.: Quantum Optics: An Introduction. Oxford University Press, Cambridge (2006).
Nayfeh, A.H.: Perturbation Methods. Wiley, NewYork (2000).
Pesin, Ya.B.: Lectures on Partial Hyperbolicity and Stable Ergodicity. Zurich Lectures in Advanced
Mathematics. European Mathematical Society (2004).
Pesin, Ya.: Existence and genericity problems for dynamical systems with nonzero Lyapunov ex-
ponents. Regular and Chaotic Dynamics 12, 476–489 (2007).
Pikovsky, A., Rosenblum, M., Kurtz, J.: Synchronization: A Universal Concept in Nonlinear Sci-
ences. Cambridge University Press, Cambridge (2002).
Schuster, H.G., Just, W.: Deterministic Chaos: An Introduction. Wiley-VCH Verlag GmbH & Co.
Strogatz, S.H.: Nonlinear Dynamics and Chaos: with Applications to Physics, Biology, Chemistry,
and Engineering. Westview Press, MA (2001).
Tong, H.(Ed.): Dimension Estimation and Models (Nonlinear Time Series and Chaos). World Sci-
entific Publication, Singapore (1994).
Wilczak, D.: Uniformly Hyperbolic attractor of the Smale-Williams type for a Poincaré map in the
Kuznetsov system. SIAM J. Applied Dynamical Systems 9, 1263–1283 (2010).
Chapter 5
Autonomous Low-dimensional Systems with
Uniformly Hyperbolic Attractors in the Poincaré
Maps
Abstract The idea of manipulating by phases in the course of the excitation trans-
fer between the alternately active oscillators may be applied to design of au-
tonomous systems with hyperbolic chaos. To develop such models one needs to
ensure self-modulation accompanying with alternating activating and damping of
self-oscillations in two interacting subsystems. Several examples will be considered
in the present chapter. One of them may serve as illustration of intriguing prob-
lem of origin of hyperbolic attractors under variation of parameters of the system; it
relates to the scenario of birth of the Smale-Williams attractor via the so-called blue-
sky catastrophe in accordance with mechanism suggested by Shilnikov and Turaev.
All systems we examine here are based on non-resonance transfer of excitation; in
some respect, it may be regarded as a disadvantage because it may be problematic
to implement these models, say, in electronics and nonlinear optics, at large ratios
of characteristic frequencies of the self-oscillation and self-modulation.
5.1 Autonomous system of two coupled oscillators with

self-regulating alternating excitation
Let us try to construct an autonomous continuous-time system with Smale-Williams

attractor in the Poincaré map departing from the coupled alternately excited oscilla-
tors discussed in the previous chapter.
There are two kinds of time dependence of coefficients, which make Eqs. (4.9)
non-autonomous. The first is slow periodic variation of parameters responsible for
the Andronov-Hopf bifurcation, and the second is the auxiliary signal of frequency
ω0 . In the averaged equations for complex amplitudes (4.17) the last kind of the
time dependence is excluded, so, to get a set of autonomous differential equations
we need only one step: to replace the external slow periodic driving by appropriate
evolution of an additional real internal variable z, which enters equations for two
oscillators with opposite signs:
104 5 Autonomous Low-dimensional Systems with . . .
1
ȧ = (z − |a|2 )a − iε b,
2
(5.1)
1
ḃ = (−z − |b|2)b − iε a2.
2
This additional variable has to decrease while the first oscillator is active, and to
grow as the second one is active, in symmetric way. To obtain the alternating excita-
tion of the oscillators as a self-supporting process, we supplement Eqs. (5.1) with a
differential equation of the first order for the variable z; the right-hand part contains
the difference of the squared amplitudes of two oscillators and a term linear in z,
with some positive coefficient. So, we assume
ż = α z + |b|2 − |a|2. (5.2)
Now, setting a = X + iY, b = U + iV in (5.1) and (5.2) we arrive at an autonomous

five-dimensional system in real variables
1
Ẋ = (z − X 2 − Y 2 )X + ε V,
2
1
Ẏ = (z − X 2 − Y 2 )Y − ε U,
2
1 (5.3)
U̇ = (−z − U 2 − V 2 )U + 2ε XY,
2
1
V̇ = (−z − U 2 − V 2 )V − ε (X 2 − Y 2 ),
2
ż = α z + U 2 + V 2 − X 2 − Y 2 .
It delivers an explicit example of an autonomous set of differential equations

possessing attractor of Smale-Williams type in the Poincaré map. 1
Figure 5.1 shows time dependence on amplitudes and phases obtained from nu-
merical simulation of the dynamics at appropriately selected parameter values. Ob-
serve alternating self-oscillating excitations of the first and the second partial oscil-
lators. As will be shown below, they are accompanied with the transformation of
phases, like in the prototypical system corresponding to the Bernoulli type map on
each cycle of oscillations of the variable z.
To proceed, we have to define Poincaré section in the five-dimensional phase
space; it appears appropriate to use a hyper-plane z=0 accounting cross-sections of
it by orbits, say, in direction of increase of the variable z, i.e., with the additional
condition ż > 0.
1 Although we have derived equations departing from a model associated with physically non-
autonomous system, what is worth outlining here is formal relation of the differential equations
(5.1) by their mathematical entity to the autonomous class. At the moment, we do not concern the
question of concrete physical realization of the dynamics.
5.1 Autonomous system of two coupled oscillators with self-regulating alternating excitation 105
Fig. 5.1 Plots of the real amplitudes |a| and |b| versus time in the sustained regime obtained from
numerical solution of the differential equations (5.3), at ε = 0.1 and α = 0.6.
Fig. 5.2 Portrait of attractor for the system (5.3) in projection on the plane (U, V ) in the Poincaré
section with the hyper-plane z = 0 passed by the orbits with ż > 0. Parameter values are ε = 0.1 and
α = 0.6. A fragment is shown with magnification to resolve the transversal Cantor-like structure
of the object. A mean time interval between successive cross-sections of the hyper-plane in right
direction is < T >≈ 13.77.
Figure 5.2 shows portrait of attractor of the Poincaré map in projection on a

plane of variables relating to one of the alternately exciting subsystems. In fact, the
attractor is a kind of the Smale-Williams solenoid embedded in the four-dimensional
phase space of the Poincaré map. In the inset a fragment of the picture is magnified
that allows resolving the intrinsic transversal Cantor-like structure of the filaments
of the attractor. Figure 5.3 shows the diagram illustrating transformation of the an-
gular coordinate, or phase for the second subsystem defined at successive returns
at the cross-sections as ϕn = arg(U + iV )|z=0,ż>0 . It may be seen that for the cyclic
coordinate ϕ the expanding circle map takes place with high accuracy producing
the doubling of the phase. In other directions in the phase space the phase volume
undergoes compression.
Fig. 5.3 Iteration diagram for the phase of the second partial oscillator ϕn = arg(U + iV )|z=0,ż>0
obtained from numerical solution of Eqs. (5.3) at ε = 0.1 and α = 0.6.
Computation of the Lyapunov exponents for the attractor of the flow system (5.3)
at ε = 0.1 and α = 0.6 yields
λ1 = 0.05036, λ 2 = 0.00000, λ 3 = −0.3351,

(5.4)
λ4 = −0.3933, λ 5 = −0.4787.
Accounting that the average period of transfer of the excitation according to the
computations is found to be Tav ≈ 13.77, the largest Lyapunov exponent for the
Poincaré map is Λ 1 = λ1 Tav ≈ 0.6934, which is in excellent agreement with the es-
timate based on the Bernoulli map, Λ 1 = ln 2. The second exponent λ 2 is zero up
to numerical errors and corresponds to perturbation vector directed along the refer-
ence trajectory as it always takes place in autonomous systems. All other Lyapunov
exponents are negative being responsible for compression of the phase volume and
approach of phase trajectories to the attractor. The estimate of the Kaplane-Yorke
dimension for the attractor embedded in the five-dimensional phase space of Eqs.
(5.3) yields D KY = 2 + (λ1 + λ2)/|λ3 | ≈ 2.15.
The numerical simulations agree with the structural stability expected for the
attractor. Qualitatively the same behavior with Bernoulli type dynamics of the phase
for successive epochs of activity persists in a rather wide area on the parameter plane
(Fig. 5.4).
5.2 System constructed on a base of the predator-prey model 107
Fig. 5.4 Domain of persistence of dynamics associated with Bernoulli type transformation of
phases for successive epoch of activity of subsystems of the model (5.3) shown as gray area on
the parameter plane.
5.2 System constructed on a base of the predator-prey model
A number of examples of autonomous systems with hyperbolic attractors were sug-

gested in our joint work with Pikovsky (Kuznetsov and Pikovsky, 2007). Let us
turn to one, which is characterized by the minimal phase space dimension allow-
ing existence of attractor of Smale-Williams (other examples will be discussed in
Chap. 8).
We begin with a two-dimensional predator-prey system with instant state speci-
fied by two non-negative variables r 1 , r2 :

1 1 2
ṙ1 = 2 1 − r2 + r1 − r1 r1 , ṙ2 = 2(r1 − 1)r2 . (5.5)
2 50
The system has a limit cycle which passes very close to the saddle fixed point in
the origin, goes around the unstable focus A (at r 1 = 1, r2 = 37/25), passes near
the saddle point B (at r 1 = 26.86, r2 = 0) and then the loop is closed (see the phase
√ √
portrait of trajectories on the plane of variables r1 and r2 in Fig. 5.5(a)). During
1 1
this process, the growth rates represented by the factors m 1 = 1 − r2 + r1 − r12
2 50
and m2 = r1 − 1 in Eqs. (5.5) alternately change sign.
Let us regard r1 and r2 as squared amplitudes of two oscillators of some fre-
quency ω0 , and represent them as r 1 = x2 + u2 and r2 = y2 + v2 . The new variables
x, u and y, v correspond to real and imaginary parts of two complex amplitudes:
a1 = x + iu and a2 = y + iv. Then, we write down
ȧ1 = −iω0 a1 + m1 (|a1 |, |a2 |)a1 , ȧ2 = −iω0 a2 + m2(|a1 |, |a2 |)a2 , (5.6)
or
ẋ = ω0 u + m1(x2 + u2 , y2 + v2 )x,
u̇ = −ω0 x + m1 (x2 + u2 , y2 + v2 )u,
(5.7)
ẏ = ω0 v + m2(x2 + u2, y2 + v2)y,
v̇ = −ω0 y + m2(x2 + u2 , y2 + v2 )v.
At this point, what we have is a system of two oscillators alternately undergoing
growth and decay, and their phases are uncoupled. Finally, we complement (5.7)
with additional coupling terms in the first and the third equations, which are propor-
tional to vy and x, respectively, and arrive at the following model
1 1
ẋ = ω0 u + [1 − (y2 + v2 ) + (x2 + u2 ) − (x2 + u2 )2 ]x + ε yv,
2 50
1 1
u̇ = −ω0 x + [1 − (y2 + v2 ) + (x2 + u2) − (x2 + u2 )2 ]u, (5.8)
2 50
ẏ = ω0 v + (x2 + u2 − 1)y + ε x,
v̇ = −ω0 y + (x2 + u2 − 1)v.
Here ε is the coupling constant. In the numerical computations discussed below, we

fix the parameters ω 0 = 2π , ε = 0.3.
At ε = 0 the equations for variables r 1 = |a1 |2 = x2 + u2 and r2 = |a2 |2 = y2 + v2
coincide precisely with relations (5.5). At ε small enough, motion of a representa-
tive point on the plane r 1 , r2 follows approximately a closed limit cycle, like that
in Fig. 5.5(a), visiting again and again a neighborhood of the origin. Qualitatively,
for each such passage, the following stages may be specified: excitation of the first
oscillator (i), excitation of the second oscillator (ii), damping of the first oscilla-
tor (iii), and slower damping of the second oscillator (iv). Activation of the sec-
√ √
Fig. 5.5 Portrait of trajectories on the phase plane of Eqs. (5.5) in variables r1 and r2 (a)
and portrait of a trajectory on attractor of the model (5.8) at ω0 = 2π and ε = 0.3 depicted on a
plane of the same variables expressed as absolute values of the complex amplitudes (b). The arrow
indicates direction of the motion. Segment S corresponds to the hyper-plane used in the definition
of the Poincaré map.
ond oscillator occurs in the presence of driving from the partner, due to the cou-
pling term proportional to ε in the second equation, so, it inherits the phase from
the first oscillator. During the damping stage of the second oscillator, its residual
oscillations stimulate the first one. The corresponding term in the equations ε yv
is selected in such a way that the stimulation is determined by the second har-
monic of the oscillations. Indeed, if y ∼ sin(ω 0t + ϕ ) and v ∼ cos(ω0t + ϕ ), then
1
yk vk ∼ sin(ω0t + ϕ ) cos(ω0t + ϕ ) = sin(2ω0t + 2ϕ ). As a result, the first oscillator
2
inherits the doubled phase 2ϕ on its active stage. Observe that the frequency of driv-
ing is essentially different from the natural frequency of the stimulated oscillator:
2ω0 versus ω0 ; so we talk of the non-resonant transfer of the excitation. As time
passes, it is the turn for the first oscillator to decay, but it transfers the excitation
back to the second oscillator and so on. This successive exchange by the excitation
repeats again and again, and each complete cycle of the process corresponds to the
expanding circle map of Bernoulli for the phase variable ϕ n+1 = 2ϕn + const, at
least in a rough approximation. In spite of the non-resonant nature of the excitation
transfer on a certain stage, everything works; the dynamics in accordance with the
explained mechanism is clearly observed in computations, at least while one sets
the frequency parameter ω 0 not too large. 2
Figure 5.5(b) shows portrait of the attractor as it looks in projection onto the
plane of the same variables as those used for depicting the limit cycle of the initial
version of Eqs. (5.5).
In Fig. 5.6 the panel (a) presents time dependence for oscillating variables x and y
relating to the first and the second oscillators. Observe alternating stages of activity
of the partial oscillators. The mean time interval of repetitions of the stages esti-
mated in this operation mode is found to be < T >= 7.248. The panel (b)√shows a
three-dimensional portrait of the attractor in the space (y, v, ρ ), where ρ = x2 + u2.
Rotation of the orbit approximately on the horizontal plane corresponds to epochs
of excitation and slow decay for the second oscillator and of silence for the first one.
Fig. 5.6 Plots for time dependence for oscillating variables x and y relating to the first and the
second oscillators (a) and portrait of attractor in 3D projection (b) for the model (5.8) at ω0 = 2π
and ε = 0.3.
2 In computations at different initial conditions one and the same chaotic attractor is always ob-
served; no phenomena of multistability have been noticed. The same kind of dynamics is observed
in a rather wide parameter range.
The excitation of the first oscillator gives rise to the departure of the orbit from the
origin with the subsequent excitation of the second oscillator, its decay, and the next
return into a neighborhood of the origin.
To proceed, we introduce the Poincaré map in the four-dimensional phase space
of the system (x, u, y, v). From Fig. 5.5(b) it is clear that it is appropriate to define the
Poincaré section with a hyper-surface specified by the equation r 1 = c that is y2 +
v2 = c taking into account crossings in direction of increase of r 1 . At the parameters
assigned here, an appropriate constant is c = 2.5. In Fig. 5.5(b) it corresponds to a
horizontal segment S. In the three-dimensional cross-section we use the coordinates
{x, u, z}, here x and u are just variables appearing in the original equations, and the
third variable is determined as
z = arg(y + iv) − arg(x + iu) (5.9)
in such a way that z ∈ (−π , π ].

The procedure for computation of the Poincaré map for a given initial vector
x = {x, u, z} is arranged as a computer program based on numerical solution of
differential equations (5.8) complemented with Hénon’s interpolation procedure,
consistent in accuracy with the applied finite-difference scheme (Hénon, 1982). As
the trajectory segment starting and finishing at the Poincaré section is computed, the
updated variables x̄ = (x̄, ū, z̄) = T(x) are determined at the final point.
Performing the computations for some representative trajectory on the attractor
we can determine sequence of phases at successive intersections of the surface S
at tn as ϕn = arg(x(tn ) + iu(tn )) to check whether it satisfies the Bernoulli map as
it should be according to the above qualitative intuitive argumentation. Figure 5.7
shows a plot for the phases obtained from iteration of the Poincaré map T in the com-
putations for the model (5.8) at ω 0 = 2π , ε = 0.3. Observe that the plot is indeed
close to that of the Bernoulli map. The most important is the intrinsic visible topo-
logical nature of the map: a single bypass of the complete circle by the pre-image
implies the two-fold bypass for the image. It gives a foundation for suggestion that
the Smale-Williams attractor occurs because the Poincaré map ensures the two-fold
Fig. 5.7 Iteration diagram for the phases of the first partial oscillator at successive crossing the
Poincaré section ϕn = arg(x+ iu)|z=0,ż>0 obtained from numerical solution of Eqs. (5.8) at ω0 = 2π
and ε = 0.3.
stretching along the angular phase coordinate, and compressions in other directions
for the evolving elements of phase volume.
Let us consider a toroidal domain D in the three-dimensional phase space of the
Poincaré map T(x) specified by the following parametric equations:
x = (r + ρ b sin ϑ ) cos ψ ,
u = (r + ρ b sin ϑ ) sin ψ , (5.10)
z = ρ d cos ϑ ,
where 0 ρ 1, 0 ψ < 2π , 0 θ π , and the constants are r = 4.05, b = 0.3,

d = 0.1. As it was validated in computations, this is an absorbing domain for the
attractor of the Poincaré map T(x). Figure 5.8 shows this domain and its image
under a single application of the Poincaré map. Observe that the image looks exactly
as it is assumed in the construction of the Smale-Williams attractor: this is a narrow
double folded loop embedded inside the toroidal absorbing domain. In Fig. 5.9 a
portrait of the attractor of the Poincaré map is plotted in projection onto the phase
Fig. 5.8 Absorbing domain parameterized by Eqs. (5.10) is image in the 3D phase space of the
Poincaré map of the system (5.8) at ω0 = 2π , ε = 0.3.
Fig. 5.9 Attractor of the Poincaré map of the system (5.8) at ω0 = 2π , ε = 0.3 in projection onto
the phase plane of the first oscillator (x, u).
plane of the first oscillator (x, u). Observe visible Cantor-like transversal structure
of the attractor, which is a characteristic feature of the Smale-Williams solenoid.
Calculation of Lyapunov exponents using the algorithm of Benettin in applica-
tion to the model (5.7) yields the following data:
λ1 = 0.0918, λ 2 = 0.0000, λ 3 = −0.982, λ4 = −1.330. (5.11)
Accounting the autonomous nature of the system, it is natural to regard the expo-
nent λ2 as precisely zero. Next, as mentioned, the average period between successive
passages of the Poincaré section is Tav = 7.248. Then, in accordance with the ap-
proach based on the Bernoulli map, the largest Lyapunov exponent must be equal to
Tav−1 ln2 ≈ 0.096. This is in reasonable agreement with the value of λ 1 obtained in
numerical calculations. Estimate of the dimension of the attractor according to the
Kaplan-Yorke formula yields D L = 2 + (λ1 + λ2)/λ3 ≈ 2.094.
For the system governed by Eqs. (5.8) with the above-mentioned parameters val-
ues the computations were carried out, confirming that the cone criterion is valid in
the toroidal domain, containing the attractor of the Poincaré map (see (Kuznetsov et
al., 2007) and Sect. 7.3 of this book); so, this is really the hyperbolic attractor.
5.3 Example of blue sky catastrophe accompanied by a birth of

Smale-Williams attractor
As mentioned in Chap. 2, attractors of Smale-Williams type in Poincaré sections of

continuous-time systems may arise under variation of a parameter via the so-called
blue sky catastrophe (Turaev and Shilnikov, 1995; Shilnikov and Turaev, 1997).
To construct an example with birth of the Smale-Williams attractor, one must
have the phase space dimension at least equal to four. The phase-space flow has to
be organized in such a way that a toroidal domain close initially to the saddle-node
cycle, in the course of travel and returning to that cycle would form a double-folded
loop, with thinner coils. To do this, the model considered in the previous section
may be applied with certain modifications.
Let us revise the prototype model (5.5) as follows (Kuznetsov, 2010):

1 1 2 1 1 2
ṙ1 = 2 1 − r2 + r1 − r1 r1 , ṙ2 = 2 r1 − μ + r2 − r2 r2 . (5.12)
2 50 2 50
These equations differ from (5.5) with additional nonlinear terms in the second
(“predator”) equation, and contain now a control parameter μ . If the value of μ
1
is slightly less than μ0 = 3 , a picture of orbits on the phase plane r 1 , r2 looks like
8
that shown in Fig. 5.10(a). There are four fixed points: an unstable focus A, saddles
B and C1 , and a node C 2 . With increase of μ , the fixed points C 1 and C2 move to
meet each other at μ = μ 0 , and then disappear (see the panels (b) and (c), respec-
tively). Instead of the former pair of fixed points, a domain of relatively slow motion
5.3 Example of blue sky catastrophe accompanied by a birth of Smale-Williams attractor 113
appears, and the attractor is now a limit cycle, which passes close to the origin and
to the saddle B.
Fig. 5.10 Phase portraits of the system (5.12) with increase of μ from (a) to (c) show that two
fixed points, a stable node C2 and a saddle C1 (a) meet together at the bifurcation (b), and a limit
cycle appears shown with a bold curve (c).
Like in the previous section, let us regard the quantities r 1 , r2 as squared absolute
values of complex amplitudes for two oscillators and introduce additional coupling
between them. The equations read

1 2 1 2
ẋ = ω0 u + 1 − (y + v ) + (x + u ) − (x + u ) x + ε yv,
2 2 2 2 2
2 50

1 1
u̇ = −ω0 x + 1 − (y2 + v2 ) + (x2 + u2) − (x2 + u2)2 u,
2 50
(5.13)
1 2 1 2
ẏ = ω0 v + x + u − μ + (y + v ) − (y + v ) y + ε x,
2 2 2 2 2
2 50

1 2 1 2
v̇ = −ω0 y + x + u − μ + (y + v ) − (y + v ) v.
2 2 2 2 2
2 50
At ε = 0, equations for r 1 = |a1 |2 = x2 + u2 and r2 = |a2 |2 = y2 + v2 can be eas-

ily derived from (5.13), which coincide with Eqs. (5.12). At ε small enough, and at
values of μ notably less than μ 0 , the sustained dynamics of the model (5.12), be-
ing represented graphically on the plane r 1 , r2 , takes place nearby the node C 2 . For
nonzero ε , this is a limit cycle of such kind that the second oscillator has some no-
table amplitude, while the amplitude for the first one is very small. Besides, there is
a close unstable limit cycle close to C 1 . With gradual increase of the parameter, both
cycles come closer, meet together and coincide at some μ = μ c (ε ) ≈ μ0 , forming
a semi-stable limit cycle. At μ > μc (ε ) it disappears. Now, motion of a representa-
tive point on the plane r 1 , r2 follows approximately a closed large-scale path, like
in Fig. 5.10(c), visiting again and again a neighborhood of the origin. As explained
in the previous section, transformation of the phase at each next cycle of the exci-
tation exchange in this regime will correspond approximately to the Bernoulli map,
ϕn+1 = 2ϕn + const.
These considerations suggest that at μ = μ c a kind of blue sky bifurcation oc-
curs, which corresponds to the case m = 2 (see Sect. 2.7). So, in accordance with
argumentation of (Shilnikov and Turaev, 1997) it has to be accompanied by a birth
of hyperbolic attractor, the Smale-Williams solenoid in the Poincaré section.
Figure 5.11 shows portraits of phase trajectories in projection on the plane of real
amplitudes of two oscillators as obtained from numerical solution of Eqs. (5.13).
Parameter values in the computations are specified as ω 0 = 2π and ε = 0.5.
Taking into account configuration of the attractor, in the supercritical domain
μ > μc it is appropriate to define the Poincaré section with a hypersurface S in
the 4D phase space by equation |a 2 | = c with c = 2.5, taking into to account only
crossings in direction of increase of |a 2 | (see a horizontal segment and an arrow in
panel (b) of Fig. 5.11).
Fig. 5.11 A transient trajectory (gray) and limit cycle (black) before the bifurcation at μ = 3.14
(a) and portrait of attractor after the bifurcation, at μ = 3.1442 (b), projected onto a plane of real
amplitudes |a1 | and |a2 |. Other parameters are ω0 = 2π and ε = 0.5.
As mentioned in (Shilnikov and Turaev, 1997), near the bifurcation, the char-
acteristic time of return of orbits at the Poincaré section depends on the control
√
parameter like T (μ ) ∼ 1/ μ − μc . In computations, the numerical data are found
to agree well with this relation. Evidence is delivered by a plot of squared inverse
averaged time of returns against parameter μ shown in Fig. 5.12. The bifurcation
threshold corresponds to intersection of the curve with the horizontal coordinate
axis, and locally the dependence is linear (see the dotted line). From extrapolation,
the bifurcation is located with high accuracy at μ c ≈ 3.144196.
Diagrams in Fig. 5.13 illustrate evolution of phases at successive crossings of
the Poincaré section S in the course of the motion on the attractor. The phase Φ n
relates to the first oscillator at t = tn , which is the n-th crossing of the orbit with
the hypersurface S in the correct direction. It is determined as Φ n = arg(x(tn ) +
iu(tn )). Observe that topologically the discrete-step evolution of phases corresponds
to Bernoulli map. Such topological nature persists in a wide range of parameter μ
above the bifurcation threshold. It supports the qualitative argument that the case
5.3 Example of blue sky catastrophe accompanied by a birth of Smale-Williams attractor 115
Fig. 5.12 Squared inverse averaged time of returns at the Poincaré section for orbits on the attractor
versus control parameter μ ; the label μc marks the blue sky bifurcation point. Other parameters are
ω0 = 2π and ε = 0.5.
Fig. 5.13 Diagrams for discrete-step evolution of phases for the first oscillator at passages of the
Poincaré section S. Parameters are ω0 = 2π , ε = 0.5, μ = 3.1442 (a) and μ = 3.15 (b).
m=2 does occur associated with presence of structurally stable hyperbolic attractor
represented by the Smale-Williams solenoid in the Poincaré map. Figure 5.14 shows
the attractor in the Poincaré section in projection on the phase plane of the first
oscillator. Although it looks like a simple closed curve, a high enough zoom resolves
intrinsic transversal Cantor-like structure.
Fig. 5.14 Attractor of the system (5.13) at μ = 3.15 in the Poincaré section projected onto the
phase plane of the first oscillator. Transversal Cantor-like structure becomes visible at great reso-
lution, as shown in the inset.
In Fig. 5.15 a plot for the Lyapunov exponents is shown, as they behave un-
der parameter variation through the blue sky catastrophe, which occurs at μ = μ c
(see panel (a)). In the left part of the plot, the largest Lyapunov exponent is zero,
and three others are negative. Here attractor is a stable limit cycle, like that in
Fig. 5.11(a). At the bifurcation, both larger exponents meet at zero, and two others
remain negative. After the bifurcation, there is one positive exponent, which grows
with increase of μ . The second exponent remains zero, and two others are negative,
with gradually decreasing absolute value. In this range the Smale-Williams attractor
persist.
Fig. 5.15 Lyapunov exponents of the system (5.13) plotted versus parameter μ (a), a magnified
part for the largest Lyapunov exponent near the bifurcation threshold in supercritical domain (b),
and the plot for the Lyapunov exponent of the Poincaré map (c). Observe that the Lyapunov expo-
nent for the Poincaré map is approximately constant close to ln 2 that corresponds to the Bernoulli
map.
Particularly, at μ = 3.15 the Lyapunov exponents are
λ1 = 0.0434, λ 2 = 0.0000, λ 3 = −8.75, λ4 = −8.85. (5.14)
The second exponent is zero (up to numerical error) and must be regarded as
being associated with neutral infinitesimal perturbation along the trajectory on the
attractor. Estimate of the attractor dimension from the Kaplan-Yorke formula (Ap-
pendix D) yields D L = 2 + λ1/|λ3 | ≈ 2.005.
The growth of the positive Lyapunov exponent in the supercritical domain,
μ > μc (see a separate plot (b)), occurs because of a decrease of the characteristic
time scale observed with departure from the bifurcation point. In the supercritical
domain, one can determine Lyapunov exponents Λ 1 , Λ2 , Λ3 for the Poincaré map
as well. The zero-value exponent is excluded, and others are in obvious relation
with those for the differential equations, namely, Λ 1 = λ1 < T >, Λ2 = λ3 < T >,
References 117
Λ3 = λ4 < T >. Here < T > is characteristic time of return at the Poincaré sec-
tion, averaged over orbits on the attractor. Observe that the positive exponent for the
Poincaré map remains almost constant in a wide parameter range (up to fluctuations
because of the numerical inaccuracy). It is close to ln 2, the value corresponding to
the Bernoulli map (see panel (c) in Fig. 5.15).
The present model essentially supplements a collection of known explicit exam-
ples of blue sky catastrophes (Gavrilov and Shilnikov, 1999; Shilnikov and Cymba-
lyuk, 2005; Glyzin et al., 2008), which were restricted, until now, with transitions
involving only regular attractors, the limit cycles. This and similar examples may
be of interest for understanding variety of dynamical behavior, e.g. in neurodynam-
ics. Indeed, some models with the blue sky catastrophes relate to this discipline
(Shilnikov and Cymbalyuk, 2005; Shilnikov, A. and Kolomiets, 2008), as well as a
hypothetical example of a uniformly hyperbolic strange attractor suggested in (Be-
lykh et al., 2005).
References
Gavrilov, N.K., Shilnikov, A.L.: An example of blue sky catastrophe. In: Methods of qualitative
theory of differential equations and related topics. Amer. Math. Soc. Transl., II Ser. 200, pp.165–
188. AMS, Providence, RI (1999).
Glyzin, S.D., Kolesov, A.Yu., Rozov, N.Kh.: Blue sky catastrophe in relaxation systems with one
fast and two slow variables. Differential Equations 44, 161–175 (2008).
Hénon, M.: On the numerical computation of Poincaré maps. Physica D 5, 412–414 (1982).
Kuznetsov, S.P.: Example of blue sky catastrophe accompanied by a birth of Smale-Williams at-
tractor. Regular and Chaotic Dynamics 15, 348–353 (2010).
Kuznetsov, S.P., Pikovsky, A.: Autonomous coupled oscillators with hyperbolic strange attractors.
Physica D 232, 87–102 (2007).
Kuznetsov, S.P., Pikovsky, A.S., Sataev, I.R.: Hyperbolic Smale–Williams attractor in Poincaré
map of a four-dimensional autonomous system. In: Proc. of the III Int. Conf. “Frontiers of
Nonlinear Physics”. Nizhny Novgorod – Saratov – Nizhny Novgorod, pp. 66-67 (2007).
Shilnikov, A., Cymbalyuk, G.: Transition between tonic spiking and bursting in a neuron model
via the blue-sky catastrophe. Phys. Rev. Lett. 94, 048101 (2005).
Shilnikov, L.P., Turaev, D.V.: Simple bifurcations leading to hyperbolic attractors. Computers
Math. Appl. 34, 173–193 (1997).
case study. Int. J. Bifurcation and Chaos 18, 2141–2168 (2008).
Turaev, D.V., Shilnikov, L.P.: On blue sky catastrophes. Doklady Akademii Nauk 342, 596–599
(1995).
Chapter 6
Parametric Generators of Hyperbolic Chaos
Abstract Excitation transfer between alternately active oscillators with phases

transformed in accordance with chaotic maps, like that in the previous two chap-
ters, may be regarded as some general principle of design of systems with attrac-
tors of the Smale-Williams type. Appropriate and convenient for implementation
of this principle, are parametric oscillatory systems (Mandelshtam, 1972; Louisell,
1960; Akhmanov and Khokhlov, 1966; Rabinovich and Trubetskov, 1989; Damgov,
2004). The term relates to a class of systems with excitation of oscillations caused
by periodic variation of some parameter. Parametric oscillations occur in systems of
mechanics, electronics, acoustics, nonlinear optics, etc. In mechanics, a commonly
known example relates to a person on a swing, who gradually increases the am-
plitude of the oscillations due to variation of the body position that corresponds to
periodic change of the effective length of the equivalent pendulum (Fig. 6.1 (a) and
(b)). In electric LC-circuit one can increase amplitude of the oscillations step by step
if vary periodically the capacity: increasing the distance between the plates of the
capacitor at the instants of its maximal charge, and decreasing it back at the instants
of maximal current in the inductance, when the charge of the capacitor is close to
zero (Fig. 6.1(c)). Clearly, the mechanical work against the force of attraction of the
charged plates in the course of the process converts into the energy of electromag-
netic oscillations in the circuit; so, the amplitude will grow gradually. The main part
of this chapter starts with consideration of the three-frequency parametric oscillator
(Louisell, 1960), in which the excitation takes place in two coupled oscillators of fre-
quencies ω1 and ω2 due to periodic variation of the coupling strength with the pump
frequency ω 3 = ω1 + ω2 . Saturation of the excitation may be ensured by nonlinear
damping in the scheme, which manifests then regular sustained regime of paramet-
ric generation. After that, we turn to designs, in which phases of the parametrically
excited oscillators undergo transformations corresponding to the expanding circle
map in some period of time. Accounting compression of the phase volume in the
state space along other directions, these will be systems with attractors of Smale-
Williams type in the stroboscopic Poincaré maps. In the systems examined here,
the phase manipulation is performed as energy is transferred between oscillators,
whereas the key role in systems discussed in the previous two chapters is played by
120 6 Parametric Generators of Hyperbolic Chaos
a non-oscillatory external source of energy that compensates for oscillator losses,

rather than by the energy transfer.
Fig. 6.1 Simple examples of parametric excitation: (a) a child increases amplitude of the swing
by variation of the body position that is equivalent to periodic variation of the effective length of
a pendulum (b); in LC-circuit the electromagnetic oscillations grow in amplitude due to periodic
variation of the capacity (c).
6.1 Parametric excitation of coupled oscillators. Three-frequency

parametric generator and its operation
Let us start with a system composed of two oscillators of frequencies ω 1 and ω2

(Fig. 6.2) coupled via a linear element characterized by strength of coupling period-
ically varied with frequency ω 3 , and the frequencies are related as
ω3 = ω1 + ω2 . (6.1)
The values ω1 and ω2 are called the signal and the idler frequency, and ω 3 is the
pump frequency. The relation (6.1) is the parametric resonance condition.
Fig. 6.2 Block diagram of a three-frequency parametric generator. Blocks marked with ω1 and ω2
represent the signal and idler oscillators, and the block with the wavy crossing is the non-dissipative
coupling element characterized by parameter oscillating with a frequency of pump ω3 .
To derive differential equations for systems of some physical nature one should
use relations specific for each respective sub-discipline; say, Newton equations in
6.1 Parametric excitation of coupled oscillators. . . . 121
mechanics, Kirchhoff equations in electric circuits, etc. Alternatively, we may turn

to formalism elaborated for general description of conservative dynamics, like the
Lagrangian or Hamiltonian method. Then, the equations are derived in elegant and
straightforward way in correct and general form. To apply them to concrete para-
metric systems, one needs only to specify physical sense of the used dynamical
variables and parameters.
1 1
A Lagrange function in our case is composed of terms ẋ2 − ω12 x2 and
2 2
1 2 1 2 2
ẏ − ω2 y representing differences of the kinetic and potential energy for two
2 2
partial oscillators, expressed via generalized coordinates and velocities, and of the
coupling term containing product of the generalized coordinates and the pump sig-
nal of frequency ω 3 :
1
L(x, ẋ, y, ẏ,t) = (ẋ2 + ẏ2 − ω12 x2 − ω22 y2 ) + κ xy sin ω3t. (6.2)
2
Here parameter κ characterizes strength of the coupling and intensity of the pump.
Explicit dependence of the Lagrangian on time indicates that we deal with a non-
autonomous system.
The differential equations governing dynamics of the system are obtained from
the Lagrange function in standard way, as a set of Lagrange-Euler equations
d ∂L ∂L
− = 0,
dt ∂ ẋ ∂ x
(6.3)
d ∂L ∂L
− = 0.
dt ∂ ẏ ∂ y
With the Lagrangian (6.2) we get the linear equations
ẍ + ω12 x = κ y sin ω3t,

(6.4)
ÿ + ω22 y = κ x sin ω3t.
As the parametric excitation occurs, in these linear equations the amplitude will
grow unboundedly. To avoid this, we have to introduce into the model some fac-
tor responsible for saturation. The simplest and physically justified way is to add
nonlinear damping; moreover, it is reasonable because we wish to have a model
possessing an attractor in the state space.
To account dissipation, the Lagrangian formalism is complemented with intro-
duction of the Rayleigh dissipation function (Marsden and Raţiu, 1999), which we
specify in the form
1 1
R(ẋ, ẏ) = (α1 ẋ2 + α2 ẏ2 ) + (β1 ẋ4 + β2ẏ4 ). (6.5)
2 4
Here the parameters α 1,2 and β1,2 will be responsible for the linear and nonlinear
dissipation, respectively. Now, instead of Eqs. (6.3) we have to use
d ∂L ∂L ∂R
− =− ,
dt ∂ ẋ ∂ x ∂ ẋ
(6.6)
d ∂L ∂L ∂R
− =− ,
dt ∂ ẏ ∂ y ∂ ẏ
that yields
ẍ + ω12 x = κ y sin ω3t − α1 ẋ − β1ẋ3 ,
(6.7)
ÿ + ω22 y = κ x sin ω3t − α2 ẏ − β2ẏ3 .
To illustrate the dynamics, we present results of numerical solution of Eqs. (6.7)
in Fig. 6.3. In the computations we set ω 1 = 2π , ω2 = 4π , ω3 = 6π (the natural pe-
riod of the first oscillator is the time unit), and the resonance condition is obviously
valid. Other parameters are κ = 35, α 1 = α2 = 0.6, β1 = β2 = 0.1. The variables x 1 ,
x2 , and the pump signal κ sin ω 3t are shown in dependence on time. Observe satura-
tion of the transient process resulting in stationary regime of parametric generation
with fixed amplitudes of the signal and idler oscillations. In stroboscopic Poincaré
map attractor is represented by a stable fixed point. Also, there is an unstable fixed
point in the origin: departing orbits from that point correspond to transients.
Fig. 6.3 Plots for signal and idler oscillations versus time, and the pump signal ω3 for the model
(6.7) at κ = 35, α1 = α2 = 0.6, β1 = β2 = 0.1, ω1 = 2π , ω2 = 4π , ω3 = 6π .
A useful approach to description of the dynamics of the system under consid-

eration is the method of slow complex amplitudes, analogous to that mentioned in
Chap. 4. For parametric systems this method is commonly exploited giving rise to
very special, easily recognizable form of equations common for parametric systems
of different physical nature. Here we explain this method in application to the sys-
tem (6.7).
Let us set
x = aeiω1t + a∗e−iω1t , y = beiω2t + b∗ e−iω2t , (6.8)
where a(t) = a + ia and b(t) = b + ib are complex functions of time variable
obeying the additional condition
ȧeiω1t + ȧ∗e−iω1t = 0, ḃeiω2t + ḃ∗ e−iω2t = 0. (6.9)

6.2 Hyperbolic chaos in parametric oscillator with Q-switch and pump modulation 123
With substitution of these expression in (6.7), multiplication of the first equation by

factor e−iω1t and the second by factor e −iω2t , accounting the frequency resonance
relation, and performing averaging over a period of fast oscillations, we obtain
κ ∗ 1 3
ȧ = − b − α1 a − ω12 β1 a|a|2 ,
4ω1 2 2
(6.10)
κ ∗ 1 3 2
ḃ = − a − α2 b − ω2 β2 b|b|2 .
4ω2 2 2
To give evidence for the parametric instability, consider these equations for a
case of equal coefficients of linear dissipation (α 1 = α2 = α ) and without nonlinear
dissipation (β1,2 = 0). Then, they can be solved analytically; the solution is
√ √
a = (C+ eκ t/4 ω1 ω2 + C− e−κ t/4 ω1 ω2 )e−α t ,
√ √
∗ eκ t/4 ω1 ω2 + C∗ e−κ t/4 ω1 ω2 )e−α t ,
b = − ω2 /ω1 (C+ −
where C+ and C− are complex constants determined by initial conditions. There is a

finite threshold: at κ > α unbounded growth of both amplitudes |a| and |b| occurs.
Saturation takes place only with account of the nonlinear dissipative terms in the
equation.
6.2 Hyperbolic chaos in parametric oscillator with Q-switch and

pump modulation
Let us consider a system similar to that discussed in the previous section but with pe-
riodic modulation of parameters, including the pump amplitudes and the dissipation
parameters in both oscillators (Kuznetsov et al., 2010). The system composed of two
oscillators with natural frequencies ω 1 and ω2 = 2ω1 undergoes parametric excita-
tion due to a sequence of pulses of pump signal of frequency ω 3 = ω1 + ω2 = 3ω1
following one after another with a period T 1/ω 3 . With the same period, the
damping strengths (Q-factors) for both oscillators are modulated, as well as a co-
efficient of the additional quadratic coupling of the oscillators. For clearness of the
presentation, we consider modulation of all mentioned parameters with a rectan-
gular waveform, assuming instantaneous switching from one parameter value to an-
other at certain points in time (at zeros of the pump signal). In other words, evolution
in time in one period T is supposed to consist of four stages of equal duration.
6.2.1 Dynamical equations
Let x and y be generalized coordinates of two oscillators with natural frequencies

ω1 and ω2 = 2ω1 . The pump frequency will be ω 3 = ω1 + ω2 = 3ω1 .
In terms of the Lagrange and Rayleigh functions we describe the dynami-

cal evolution as follows. At the first stage, parametric excitation takes place be-
cause of oscillations of the coupling parameter with the pump frequency. Non-
linear dissipation is present ensuring saturation of the parametric instability. The
1
Lagrangian is L = (ẋ2 + ẏ2 − ω12 x2 − ω22 y2 ) + κ xy sin ω3t, and the Rayleigh func-
2
1
tion is R = (β1 ẋ4 + β2 ẏ4 ). At the second stage, interaction of the oscillators is
4
switched off, and linear damping of the second oscillator takes place. It corre-
1 1
sponds to L = (ẋ2 + ẏ2 − ω12 x2 − ω22 y2 ), and R = α2 ẏ2 . At the third stage, there
2 2
is no dissipation, and the oscillators are coupled via quadratic nonlinear element.
Here we account a term proportional to x 2 y in the Lagrangian, and assume L =
1 2 2
(ẋ + ẏ − ω12 x2 − ω22 y2 )+ ε x2 y, while R = 0. At the fourth stage, linear damping of
2
1
the first oscillator takes place that corresponds to L = (ẋ2 + ẏ2 − ω12 x2 − ω22 y2 ), and
2
1
R = α1 ẋ2 . The durations of all stages are assumed equal and containing an integer
2
number of the high-frequency oscillations of the pump signal, i.e. T /4 = 2π N/ω 3 .
The dynamical equations for each of the four stages obtained from the relations
(6.6) are the following.
I. Stage of parametric excitation:
ẍ + ω12 x = κ y sin ω3t − β1 ẋ3 , ÿ + ω22 y = κ x sin ω3t − β2 ẏ3 . (6.11)
II. Stage of damping of the second oscillator:
ẍ + ω12x = 0, ÿ + ω22 y = −α2 ẏ. (6.12)
III. Stage of interaction through the quadratic nonlinearity:
ẍ + ω12 x = 2ε xy, ÿ + ω22 y = ε x2 . (6.13)
IV. Stage of damping of the first oscillator:
ẍ + ω12x = −α1 ẋ, ÿ + ω22 y = 0. (6.14)
Alternatively, the dynamics may be represented by a single set of differential

equations with piecewise-continuous coefficients:
ẍ + ω12 x = f1 (t)(κ y sin ω3t − β1ẋ3 ) − α1 f4 (t)ẋ + 2ε f 3 (t)xy,

(6.15)
ÿ + ω22y = f1 (t)(κ x sin ω3t − β2 ẏ3 ) − α2 f2 (t)ẏ + ε f3 (t)x2 ,
where f k (t) = 1, if k − 1 4{t/T } < k, and 0 otherwise. Here the braces desig-
nate the fractional part of a number. Parameter κ determines strength of parametric
coupling because of presence of the pump; ε is parameter of quadratic coupling;
parameters α1,2 are responsible for linear, and β 1,2 for nonlinear dissipation in the
first and the second oscillators, respectively.
A state of the system at fixed time instant t is defined by a four-dimensional
vector x = (x, u, y, v), where u = ẋ/ω 1 and v = ẏ/ω2 are normalized generalized
velocities of the oscillators. Transformation of the vector in a full period of variation
of the coefficients in the equations from t = nT up to t = (n + 1)T corresponds to
the stroboscopic Poincaré map,
xn+1 = f(xn ). (6.16)
The mapping may be performed by means of the computer program solving the set
of differential eqs. (6.11) — (6.14) in the period T at four successive stages by a
finite-difference integration method. It is natural to specify the integration step to
have an integer number of steps at each stage.
If the period of parameter modulation is much larger than the period of high-
frequency oscillations, N 1, and parameters of coupling and dissipation are small
enough, we may reformulate the problem in terms of slow complex amplitudes of
the same kind as considered in Sect. 4.1. For this, we set
x = aeiω1t + a∗e−iω1t , ẋ = iω1 aeiω1t − iω1 a∗ e−iω1t , (6.17)
and
y = beiω2t + b∗e−iω2t , ẏ = iω2 beiω2t − iω2 b∗ e−iω2t . (6.18)
Substitution into (6.11) — (6.14) with multiplication by e −iω1t
(in the equations
for a), and by e −iω2t (in the equations for b), and averaging over period of the fast
oscillations accounting the frequency relation (6.1) yield the shortened equations
for slow amplitudes at four stages of the dynamics as follows:
3 3
Stage I. ȧ = −(κ /4ω 1)b∗ − β1 ω12 |a|2 a, ḃ = −(κ /4ω2)a∗ − β2 ω22 |b|2 b.
2 2
1
Stage II. ȧ = 0, ḃ = − α2 b.
2
1
Stage III. ȧ = −iεω1 a b, ḃ = − iεω2−1 a2 .
−1 ∗
2
1
Stage IV. ȧ = − α1 a, ḃ = 0.
2
Equivalently, it may be represented by a set of differential equations
3 1
ȧ = −(κ /4ω1)b∗ f1 (t) − f1 (t)β1 ω12 |a|2 a − i f3 (t)εω1−1 a∗ b − f4 (t)α1 a,
2 2
3 1 1
ḃ = −(κ /4ω2)a∗ f1 (t) − f1 (t)β2 ω22 |b|2 b − i f3 (t)εω2−1 a2 − f2 (t)α2 b,
2 2 2
(6.19)
where, again, f k (t) = 1, if k − 1 4{t/T } < k, and 0 otherwise.
6.2.2 Qualitative explanation of the operation
At the first stage of parametric excitation amplitudes of both oscillators grow, and
their final phases appear to be fixed in certain relation. To explain this point, let us
assume for simplicity that the dissipation is absent. Then, the complex amplitudes
∗ ∗
evolve
in accordance √with equations ȧ = √ −κ b /4ω1 , ḃ = −κ √a /4ω2 , and we get
√ a=
− ω1 /ω2 (C+ e ∗ κ t/4 ω1 ω 2 + C− e ∗ − κ t/4 ω1 ω2 ), b = C+ e κ t/4 ω1 ω2 + C e−κ t/4 ω1 ω2 .
−
The constants C+ = Reiϕ and C− = Re−iϕ are determined by initial conditions at
the onset of the stage. The second terms in both expressions decay,√and at the end
ofthe stage the first terms √ dominate; so, we have b ∼ R · √e iϕ eκ t/4 ω1 ω2 and a ∼
− ω1 / −i ϕ
ω2 R · e e √ κ t/4 ω 1 ω 2 . It corresponds to y ∼ R · e κ t/4 ω1 ω2 cos(ω t + ϕ ) and
2
x ∼ R · ω1 / ω2 e κ t/4 ω 1 ω2 cos(ω1t − ϕ − π ). Thus, the oscillations arising due to
the parametric instability on the first harmonic (ω 1 ) and the second harmonic (ω 2 =
2ω1 ) have related phases determined by a single constant ϕ , which depends on the
initial conditions. With account of the nonlinear dissipation, the amplitudes undergo
saturation, but the phase relation is retained, at least approximately.
At the second stage the oscillators are decoupled. The first one demonstrates free
oscillations of constant amplitude, and in the second one the oscillations decay up to
some very low level. The phase shift of the first oscillator remains nearly constant,
and we have a ∼ −e −iϕ and x ∼ cos(ω1t − ϕ − π ).
At the third stage, interaction of the oscillators through the quadratic nonlinearity
takes place. The second oscillator is driven by the resonance action from the second
harmonic component of the first oscillator because of the assumed relation ω 2 =
1
2ω1 . From the equation ḃ = − iεω2−1 a2 , with amplitude of the first oscillator a ∼
2
−e−iϕ , we obtain b ∼ −ia 2 = −ie−2iϕ , and y ∼ cos(ω2t − 2ϕ − π /2). Hence, the
value of the phase for the second oscillator appears to be doubled (up to a sign and
a constant shift).
At the fourth stage, the first oscillator decays up to negligible level while the
second one retains its amplitude and phase shift, namely, b ∼ −ie −2iϕ and y ∼
cos(ω2t − 2ϕ − π /2). At the end of the stage, it determines the initial conditions
for the next stage of parametric excitation.
Updated coefficients in the expression for the solution determining growth of
oscillations at the next stage of parametric excitation C + and C are expressed as
−
b = C+ + C ∼ −ie−2iϕ , a ∼ C − C ≈ 0. Then, C = R eiϕ ∼ −ie−2iϕ and it
− + − +
means that in the approximation we use the new and the old phase shifts ϕ related
as
ϕ = −2ϕ − π /2. (6.20)
So, transformation of the system state in a full period of four stages corresponds
to expanding map for the phase. It is a kind of chaotic expanding circle map, which
differs from the usual Bernoulli map only in inversion of the direction of the bypass
for the image around the circle. A perturbation of the variable ϕ is multiplied by
factor 2, so, the positive Lyapunov exponent is Λ = ln 2.
Each iteration of the Poincaré map is accompanied with expansion of elemen-

tary phase volume by factor 2 along a cyclic coordinate associated with the phase
variable ϕ , and with compression in all the rest directions in the phase space. It
corresponds to situation, where presence of the Smale-Williams attractor has to be
expected.
6.2.3 Numerical results
Let us turn to results of numerical simulation of the dynamics of the system. It is

convenient to set ω 1 = 2π , ω2 = 4π , and ω3 = 6π that correspond to the choice of
the unit time period of natural oscillations for the first oscillator. Figure 6.4 shows
typical examples of time dependence in the sustained regime of parametric genera-
tion obtained by solving Eqs. (6.15) with computer finite-difference method at the
parameters
Fig. 6.4 Plots of the time dependence for the pump (a) and for variables associated with signal
and idler oscillations in the sustained regime of parametric generation (b). Roman numbers denote
the four stages of the system operation. Presence of chaos reveals itself in irregular evolution of
the phase of the carrier signal relatively to the envelope; it becomes visually evident with plotting
several samples of the signals on one diagram (c).
T = 16, α1 = α2 = 1.6, β1 = β2 = 0.02, ε = 2.2, κ = 25. (6.21)
The time interval depicted contains three periods of the pump modulation; the plot
of the pump signal is shown in the upper diagram (a). As seen from the diagram (b),
each of the two oscillators that constitute the system generates oscillatory trains,
which follow with time interval corresponding to the period of pump modulation.
However, both signal and idler oscillations are in fact non-periodic: the phases of
the high-frequency filling vary from pulse to pulse chaotically. The non-periodicity
is visually evident from the diagram (c), where several successive sections of the
signals are superimposed on one and the same plot.
Figure 6.5(a) shows the iteration diagram obtained from the numerical solution
of Eqs. (6.15); the phases are determined at the end of the stages of the paramet-
ric excitation from the formula ϕ = arg(x − iu) and are reduced to fit the interval
[0, 2π ). Observe that the phase diagram is consistent with the qualitative description
in the previous subsection. Indeed, it is obvious that one complete bypass of a full
cycle for pre-image gives rise to two-fold bypass for the image (in the reverse direc-
tion). Figure 6.5(b) shows a portrait of the attractor in the Poincaré section projected
on the phase plane of the first oscillator (x, u), which corresponds to the expected
solenoid of Smale-Williams type. The transversal Cantor-like structure character-
istic for such attractor is noticeable. The hyperbolic nature of this attractor was
verified numerically with the technique considered in the next chapter (Kuznetsov
et al., 2010).
Fig. 6.5 Iteration diagram for the phases of the first oscillator in the final parts of the stages of
the parametric excitation (a) and a portrait of the attractor in the Poincaré section (stroboscopic
section), corresponding to the same points in time (b).
To have evidence for chaos on the quantitative level, we turn to the calculation
of Lyapunov exponents with a standard method discussed in Appendix A and carry
out a joint solution of Eqs. (6.15) and four replicas of variation equations
x̃¨ + ω12 x̃ = f1 (t)(κ ỹ sin ω3t − 3β1ẋ2 x̃˙) − α1 f4 (t)x̃˙ + 2ε f 3 (t)(x̃y + xỹ),
(6.22)
ỹ¨ + ω22ỹ = f1 (t)(κ x̃ sin ω3t − 3β2ẏ2 ỹ˙) − α2 f2 (t)ỹ˙ + 2ε f 3 (t)xx̃,
accompanied by the Gram-Schmidt orthogonalization and normalization for the per-

turbation vectors x̃ i = {x̃i , x̃˙i , ỹi , ỹ˙i }, i = 1, 2, 3, 4. The Lyapunov exponents are deter-
mined as mean rates for increase or decrease of accumulating sums for logarithms
of ratios of norms of the vectors (after orthogonalization but before the normaliza-
tion). Full spectrum of Lyapunov exponents for the attractor at the parameter values
(6.21), according to the calculations, is the following:
λ1 ≈ 0.0402, λ 2 ≈ −0.0631, λ 3 ≈ −0.559, λ4 ≈ −0.675. (6.23)
For the stroboscopic Poincaré map, the Lyapunov exponents are expressed as
Λk = λk T . Then, Λ1 ≈ 0.643 that agrees well with the estimate Λ = ln 2 in the
approximation corresponding to the one-dimensional expanding circle map for the
phase variable.
The presence of a positive exponent Λ 1 is a quantitative confirmation of chaos.
The remaining exponents Λ 2 , Λ3 , Λ4 are all negative. It means that an element of
phase volume undergoes stretching in one direction associated with the cyclic phase
variable, and contraction in three other directions in the four-dimensional phase
space of the Poincaré map. The Kaplan-Yorke dimension of the attractor of the
Poincaré map (see Appendix D) is D = 1 + Λ 1/|Λ2 | ≈ 1.64.
6.2.4 Numerical results in the frame of method of slow complex

amplitudes
Equations obtained in Subsection 6.2.1 for the complex amplitudes can be regarded
not only as an approximation for the original system (6.15), but also as a separate
object of study. It is essential that this formulation reveals the universal nature of
the model: the combinations of complex amplitudes and conjugate values are char-
acteristic for equations describing parametric systems of different physical nature.
Note that the number of essential parameters in this formulation is less than in the
original equations, because the absolute value of the fundamental frequency ω 1 does
not enter the equations, except in combination with other parameters. The evolution
equation can be rewritten as follows:
3 1
ȧ = −κ̄ b∗ f1 (t) − f1 (t)β̄1 |a|2 a − i f3 (t)ε̄ a∗ b − f4 (t)α1 a,
2 2
(6.24)
1 ∗ 3 1 1
ḃ = − κ̄ a f1 (t) − f1 (t)β̄2 |b|2 b − i f3 (t)ε̄ a2 − f2 (t)α2 b,
2 2 4 2
where κ̄ = κ /4ω1, β̄1,2 = β1,2 ω1,2
2
, ε̄ = ε /ω1 . Hereafter we select the parameters to
have correspondence to the previous subsection, namely,
T = 16, α1 = α2 = 1.6, β1 = 0.7896,

(6.25)
β2 = 3.1583, ε = 0.3501, κ = 0.9947.
Figure 6.6 provides time dependence of the amplitudes and phases of the com-
plex variables a and b, obtained by computer numerical solution of Eqs. (6.24). As
seen, the amplitude dependence corresponds well to the envelopes for the plots of
oscillations of two subsystems in Fig. 6.4. The dependence of the phases on time is
irregular (see panel (b)). One can check that the phase transformation in each period
of repetition of stages appears to be in good correspondence with the expanding
circle map discussed above. It is seen from the diagram shown in Fig. 6.7(a), which
should be compared with Fig. 6.5(a). Portrait of the attractor in the Poincaré section
in Fig. 6.7(b) demonstrates an obvious correspondence with attractor of the original
model in Fig. 6.5(b).
Fig. 6.6 Plots for amplitudes (a) and phases (b) of complex variables for the two partial oscillators
obtained from computer numerical solution of Eqs. (6.24). Roman numerals designate the four
stages of evolution of the system.
Fig. 6.7 Iteration diagram for the phases of the first oscillator in the final parts of the stages of
the parametric excitation (a) and a portrait of the attractor in the stroboscopic Poincaré section (b)
obtained from numerical solution of the amplitude equations (6.24).
6.3 Parametric generator of hyperbolic chaos based on four coupled oscillators with . . . 131
Computation of Lyapunov exponents for the attractor by the method of Benettin

(see Appendix A) is based on solving the equations for slow amplitudes, supple-
mented by four replicas of linearized equations for perturbations of the reference
phase trajectory, and yields the spectrum of Lyapunov exponents
λ1 ≈ 0.0413, λ 2 ≈ −0.0656, λ 3 ≈ −0.558, λ4 ≈ −0.685, (6.26)
which is in good agreement with the results (6.23). Observe that the largest exponent
for the Poincaré map Λ 1 = λ1 T ≈ 0.661 is close to the value ln 2 corresponding to
approximate description of the evolution of the phase in terms of the expanding
circle map.
Figure 6.8 shows a plot for Lyapunov exponents of the Poincaré map versus the
parameter of the pump intensity κ̄ . As seen, the largest Lyapunov exponent in a wide
parameter range remains almost constant and close to the value ln 2. This character
of the dependence is consistent with the suggested presence of a structurally stable
hyperbolic attractor.
Fig. 6.8 Lyapunov exponents for the Poincaré map plotted against the parameter of intensity of
pump for the model (6.24) at T = 16, α1 = α2 = 1.6, β1 = 0.7896, β2 = 3.1583, ε = 0.3501.
6.3 Parametric generator of hyperbolic chaos based on four

coupled oscillators with pump modulation
Consider a system composed of two subsystems A and B, each of which represents

a parametric generator consisting of two parametrically coupled oscillators with fre-
quencies ω1 and ω2 = 2ω1 , like what discussed in Sect. 6.1 (Kuznetsov, 2008). The
block diagram is shown in Fig. 6.9. Pumping frequency is assumed to be equal to
ω3 = ω1 + ω2 = 3ω1 ; so, the condition of parametric resonance is satisfied. The
oscillator of frequency ω 1 is coupled to the oscillator of frequency ω 2 in the other

subsystem via an element with quadratic nonlinearity. The pump is switched on and
off alternately in both subsystems, so the parametrically excited oscillations grow
and decay in them by turns. Each excitation of a subsystem is stimulated by a second
harmonic component generated by the other subsystem. Due to this, the oscillation
phase is doubled in the course of each signal transfer; hence, it is multiplied by a
factor of 4 over a whole period of pump modulation. As a result, the system gen-
erates a sequence of oscillatory trains following with the modulation period, whose
carrier phase chaotically varies from pulse to pulse.
Fig. 6.9 Block diagram of the parametric chaos generator. Blocks labeled as ω1 and ω2 represent
the oscillators with these natural frequencies; those labeled with a wave-like crossing are reactive
elements with coupling parameter oscillating at the pump carrier frequency; the gray squares are
coupling elements characterized by the quadratic nonlinearity.
6.3.1 Model, operation principle and basic equations
Consider the Lagrangian function

1 2
L= (ẋ + ẋ22 + ẏ21 + ẏ22 − ω12 x21 − ω22x22 − ω12 y21 − ω22 y22 )
2 1 (6.27)
+κ [x1 x2 f (t) + y1 y2 g(t)] sin ω3t + ε (x21 y2 + y21 x2 ),
where x1,2 and y1,2 are generalized coordinates of two coupled pairs of oscillators,
ẋ1,2 and ẏ1,2 are the corresponding generalized velocities, ω 1 and ω2 = 2ω1 are
the respective natural frequencies of the oscillators in both pairs. The parameter κ
quantifies the pump strength and the functions f (t) and g(t) determine the respective
slowly varying amplitudes of the pump signals that drive the two pairs of oscillators.
The pump carrier frequency is ω 3 = ω1 + ω2 = 3ω1 . The parameter ε characterizes
the coupling between oscillators belonging to different pairs.
The nonlinear damping needed for saturation of parametric instability and for
existence of an attractor is introduced by means of the Rayleigh dissipation function
defined as follows:
1 1
R = (α1 ẋ21 + α2 ẋ22 + α1 ẏ21 + α2 ẏ22 ) + (β1 ẋ41 + β2 ẋ42 + β1 ẏ41 + β2ẏ42 ), (6.28)
2 4
where α1,2 , β1,2 are positive constants. The equations of motion for the system

d ∂L ∂L ∂R d ∂L ∂L ∂R
= − , = − , i = 1, 2 (6.29)
dt ∂ ẋi ∂ xi ∂ ẋi dt ∂ ẏi ∂ yi ∂ ẏi
are then written as

ẍ1 + ω12 x1 = κ x2 f (t) sin ω3t + 2ε x1 y2 − α1 ẋ1 − β1ẋ31 ,
ẍ2 + ω22 x2 = κ x1 f (t) sin ω3t + ε y21 − α2 ẋ2 − β2 ẋ32 ,
(6.30)
ÿ1 + ω12 y1 = κ y2 g(t) sin ω3t + 2ε y1 x2 − α1 ẏ1 − β1 ẏ31 ,
ÿ2 + ω22 y2 = κ y1 g(t) sin ω3t + ε x21 − α2 ẏ2 − β2 ẏ32 .
Let us specify the functions f (t) and g(t) so that the pump signals for two subsys-
tems should be modulated in counter-phase:
f (t) = sin2 (π t/T ), g(t) = cos2 (π t/T ). (6.31)
The modulation period is assumed to be an integer multiple of the pump carrier

period, i.e. T = 2π N/ω 3, where N is an integer. This condition implies that (6.30)
is just a system of equations with T -periodic coefficients, and its dynamics can
be rigorously represented by the stroboscopic Poincaré map using sections of the
extended phase space with time step T . 1
In the case N 1, following a standard technique of the method of slow ampli-
tudes, we set
x1 = A1 eiω1t + A∗1 e−iω1t , x2 = A2 eiω2t + A∗2 e−iω2t ,

(6.32)
y1 = B1 eiω1t + B∗1 e−iω1t , y2 = B2 eiω2t + B∗2 e−iω2t ,
where A1,2 (t) = A1,2 + iA1,2 and B1,2 (t) = B1,2 + i B1,2 are slowly varying complex
functions of time subject to the additional conditions
Ȧ1 eiω1t + Ȧ∗1e−iω1t = 0, Ȧ2 eiω2t + Ȧ∗2 e−iω2t = 0,

(6.33)
Ḃ1 eiω1t + Ḃ∗1e−iω1t = 0, Ḃ2 eiω2t + Ḃ∗2 e−iω2t = 0.
Substituting these relations into (6.30), multiplying the equations for A 1 and B1 by
e−iω1t , and for A 2 and B2 by e−iω2t , averaging the resulting equations over a period
of fast oscillations, and using the assumed relation between ω 1,2,3 we obtain
1 This condition can be ignored in the approach based on slowly varying amplitudes considered
below.
κ ε 1 3
Ȧ1 = − f (t)A∗2 − i A∗1 B2 − α1 A1 − ω12 β1 A1 |A1 |2 ,
4ω1 ω1 2 2
κ ε 2 1 3
Ȧ2 = − f (t)A∗1 − i B − α2 A2 − ω22 β2 A2 |A2 |2 ,
4ω2 2ω2 1 2 2
(6.34)
κ ε 1 3
Ḃ1 = − g(t)B∗2 − i B∗1 A2 − α1 B1 − ω12 β1 B1 |B1 |2 ,
4ω1 ω1 2 2
κ ε 2 1 3
Ḃ2 = − g(t)B∗1 − i A − α2 B2 − ω22 β2 B2 |B2 |2 .
4ω2 2ω2 1 2 2
If ε = 0, the system splits into two symmetric isolated subsystems A and B, each
comprising two parametrically coupled oscillators. Consider one of the subsystems.
For the constant pump amplitude, as explained in the previous section, the paramet-
rically excited oscillations at the frequencies ω 1 and ω2 = 2ω1 are characterized by
a common phase shift constant ϕ , depending on the initial conditions.
Now, consider the regime when the coupled subsystems A and B with nonzero ε
are driven by turns due to the modulated pump.
In each subsystem, the oscillator with natural frequency ω 2 = 2ω1 is excited,
stimulated by the second harmonic of the oscillator of frequency ω 1 from the oppo-
site subsystem, produced by the quadratic nonlinear coupling element. The dou-
ble phase shift in the second harmonic is inherited then by the excited oscilla-
tor. During the next modulation half-period, the subsystems A and B exchange
roles, and the phase shift of the passed signal doubles once again. Thus, the
phase shift after a modulation period can be approximately represented by the map
ϕnew = 4ϕold + const( mod2π ). It is a Bernoulli type map with chaotic dynamics
characterized by the Lyapunov exponent Λ = ln 4 ≈ 1.386.
Description of the dynamics may be done in discrete time by sampling the eight-
dimensional state vectors X n at time instants tn = nT. From solution of Eqs. (6.30) or
(6.34) at a time interval T starting from a particular t n , we obtain a new state vector
Xn+1 . The function
Xn+1 = T(Xn ), (6.35)
which maps the eight-dimensional phase space into itself, is the stroboscopic
Poincaré map associated with the system of differential evolution equations whose
right-hand sides are smooth and bounded in a finite domain in the phase space. By
virtue of the existence, uniqueness, continuity, and differentiability of the solution
to the system, the map T is a diffeomorphism (Arnold, 1978).
6.3.2 Chaotic dynamics: results of computer simulation
Hereafter we assign ω1 = 2π , ω2 = 4π , ω3 = 6π that correspond to measuring of

time in units of the natural period of one of the involved oscillators.
Figure 6.10 shows typical samples of waveforms obtained from numerical so-
lution of Eqs. (6.30) for the generalized coordinates x and y, with modulated am-
plitudes (6.31) of the pump signals plotted separately on the panel (a). Parameter
values are assigned as follows:
T = 40, ε = 0.5, κ = 35, a = 0.6, b = 0.01. (6.36)
Fig. 6.10 Time-domain waveforms obtained from numerical solution of Eqs. (6.30) with param-
eter values (6.36): (a) pump signals; (b), (c) single samples; (d) superimposed samples of the
same signal realization. The envelopes in panels (b) and (c) are obtained by numerical solution of
equations for complex amplitudes (6.34).
As observed, in the course of the evolution in time both subsystems are active
alternately, in correspondence with the action of the pump (panels (a), (b), and (c)).
Each subsystem produces a train of oscillations following with the pump modula-
tion period; however, the waveform is not periodic: the carrier phase varies irreg-
ularly from pulse to pulse. An illuminating visual illustration is provided by panel
(d), which shows several superimposed signal samples corresponding to successive
modulation periods. Panels (b) and (c) illustrate additionally the time dependence
obtained from numerical solution of the amplitude equations (6.34). 2 The figure
demonstrates good agreement between the results obtained by the slowly varying
amplitude method and the exact solution.
Figure 6.11 shows iteration diagram for phases of the carrier in successive oscil-
latory trains generated by one of the subsystems. The abscissa and ordinate are the
phases of oscillations in the subsystem B relating to the time instants t n = nT and
tn+1 , respectively. The plot in panel (a) is obtained for the original equations (6.30),
and that in panel (b) for the amplitude equations (6.34). In the first case the phase
is determined from the formula ϕ n = arg(y1 (tn ) + ẏ1 (tn )/iω1 ), and in the second
case from ϕn = argB1 (tn ). Note that the phase is defined only within a time interval
of activity of the subsystem, when the output amplitude does not approach zero. 3
According to Fig. 6.11, as ϕ n varies from 0 to 2π , the point associated with the
phase ϕn+1 goes around the unit circle four times; i.e., the mapping is topologically
equivalent to the Bernoulli type map.
Fig. 6.11 Iteration diagram for phases at the stage of activity of one of the subsystems, composed
on the basis of the numerical solution of the original equations (6.30) (a) and the equations for
complex amplitudes (6.34) (b); the parameters assigned in accordance with (6.36).
Because of good agreement between the results obtained with the original and
the amplitude equations, the analysis below is restricted to the amplitude equations.
Quantitative evidence of chaotic behavior is obtained by calculating the Lya-
punov exponents. To find the Lyapunov spectrum by Benettin algorithm (Appendix
A), Equations (6.34) are computed simultaneously with eight replicas of variation
equations
2 To correlate the waveforms, the initial conditions set at the starting point in time were calculated
for the corresponding state of the system (6.30) by using relations (6.32) and (6.33).
3 The phase cannot be defined globally on the entire time axis, because the amplitude is nearly zero
when the subsystem is not active. If such a definition were possible, then the map for the phase
could not belong to the appropriate topological class.
κ ε 1 3
Ã˙ 1 = − f (t)Ã∗2 − i (Ã∗1 B2 + A∗1 B̃2 ) − α1 Ã1 − ω12 β1 A21 Ã∗1 − 3ω12β1 A∗1 A1 Ã1 ,
4ω1 ω1 2 2
κ ε 1 3
Ã˙ 2 = − f (t)Ã∗1 − i B1 B̃1 − α2 Ã2 − ω22 β2 A22 Ã∗2 − 3ω22β2 A∗2 A2 Ã2 ,
4ω2 ω2 2 2
˙ κ ε 1 3
B̃1 = − f (t)B̃2 − i (A2 B̃1 + B1 Ã2 ) − α1 B̃1 − ω12 β1 B21 B̃∗1 − 3ω12β1 B∗1 B1 B̃1 ,
∗ ∗ ∗
4ω1 ω1 2 2
κ ε 1 3
B̃˙ 2 = − f (t)B̃∗1 − i A1 Ã1 − α2 B̃2 − ω22 β2 B22 B̃∗2 − 3ω22β2 B∗2 B2 B̃2 .
4ω2 ω2 2 2
(6.37)
Each time after a certain number of integration steps have been performed, the
Gram-Schmidt orthonormalization process is applied to the perturbation vectors
{Ã 1 , A˜ 1 , Ã 2 , A˜ 2 , B̃ 1 , B˜ 1 , B̃ 2 , B˜ 2 }. (Prime and double prime designate real and
imaginary parts, respectively.) The Lyapunov exponents are calculated by averag-
ing the growth rates of the sum of logarithms of norms of the perturbation vectors
after orthogonalization, but before normalization. The Lyapunov spectrum obtained
for the attractor corresponding to parameter set (6.36) is as follows:
λ1 = 0.03456, λ 2 = −0.1320, λ 3 = −0.2247, λ 4 = −0.5220,

(6.38)
λ5 = −0.6826, λ 6 = −0.9012, λ 7 = −1.4189, λ 8 = −2.3248.
The corresponding Lyapunov exponents of the stroboscopic Poincaré map are

determined from (6.38) as Λ k = λk T . Particularly, Λ1 = 1.3823, which is in good
agreement with the value ln 4 = 1.3862· · · obtained from the Bernoulli type map
ϕn+1 = 4ϕn + const( mod 2π ). A positive Λ 1 is an indicator of chaos. Since the re-
maining exponents Λ 2 · · ·Λ8 are negative, an element of phase space volume under-
goes expanding in one direction in the phase space of the Poincaré map, associated
with the angular coordinate ϕ , and the others are contracting. Figure 6.12 plots two
larger Lyapunov exponents of the Poincaré map calculated as functions of the cou-
Fig. 6.12 The largest two Lyapunov exponents of stroboscopic Poincaré map vs. parameter ε
computed for ω1 = 2π , ω2 = 4π , ω3 = 6π , T = 40, κ = 35, a = 0.6, b = 0.01. The largest exponent
is consistent with the estimate Λ1 = ln 4.
pling strength ε while the remaining parameters are held constant. It is clear from
Fig. 6.12 that the value of Λ 1 is close to ln 4 over a wide interval of ε . The positive
exponent smoothly depends on the parameter and has no sharp dips characteristic
for non-hyperbolic attractors (See for comparison of Figs. B. 4 and B. 7 in Appendix
B.)
The Kaplan-Yorke dimension of the attractor of the stroboscopic Poincaré map
with the Lyapunov spectrum (6.38) is
D = 1 + Λ1/|Λ2 | ≈ 1.26 (6.39)
(since Λ1 > 0 and Λ1 + Λ2 < 0). Accordingly, the total dimension of the attractor
embedded in the nine-dimensional extended phase space is D = D + 1 ≈ 2.26.
Figure 6.13(a) shows portrait of the attractor in the Poincaré section obtained
by projection onto the phase plane of oscillator ω 1 in subsystem B. An enlarged
fragment is shown to resolve details of the intrinsic transverse Cantor-like fractal
structure. Figure 6.13(b) presents an analogous phase portrait corresponding to a
coupling strength value near the threshold for existence of the chaotic attractor,
where the transverse fractal structure is visible much better.
Fig. 6.13 Poincaré section of the attractor portrait projected onto the complex amplitude plane of
the first oscillator in subsystem B for ε = 0.5 (a) and ε = 0.125 (b). Remaining parameters are
assigned in accordance with (6.36).
The underlying mechanism of operating of the parametric generator of chaos,

as well as numerical results confirming implementation of this mechanism allow
suggesting that the observed chaotic attractor is uniformly hyperbolic. Interpreting
the action of the Poincaré map geometrically, we can imagine a toroid embedded
in an eight-dimensional space (that is a direct product of one-dimensional circle
and the seven-dimensional ball) and associate a single iteration of the map with the
longitudinal expanding and transversal compression of the object, then folded to
a quadruple loop deposited inside the original domain. At each step of repetition
of this procedure, the number of turns is increased by four times. In the limit, we
obtain a solenoid containing an infinite number of coils and transverse structure of
the Cantor set, that is one of the variants of attractors of the Smale-Williams type.
Confirmation of relation of this attractor to the uniformly hyperbolic class could be
based on the use of computational procedures described in the next chapter, although
References 139
in this case the phase space dimension is relatively high, and the problem requires
excessive computational resources.
Despite the hypothetical nature of the conclusion of the hyperbolic nature of
the chaotic attractor in the last model, the schemes of the parametric generators of
chaos considered in this chapter certainly open up possibility for generation of ro-
bust chaos insensitive to the variation of parameters and characteristics of elements,
e.g. in the systems of radio-engineering and electronics, mechanics, acoustics, non-
linear optics.
References
Akhmanov, S.A., Khokhlov, R.V.: Parametric amplifiers and generators of light. Soviet Physics
Uspekhi 9 (2), 210 (1966).
Damgov, V.: Nonlinear and parametric phenomena: theory and applications in radiophysical and
mechanical systems . World Scientific Publ. (2004).
Kuznetsov, S.P.: On feasibility of a parametric generator of hyperbolic chaos. JETP 106, 380–387
(2008).
Kuznetsov, A.S., Kuznetsov, S.P., Sataev, I.R.: Parametric generator of hyperbolic chaos based on
two coupled oscillators with nonlinear dissipation. Technical Physics 55, 1707–1715 (2010).
Louisell, W.H.: Coupled Mode and Paramagnetic Electronics. Wiley, New York (1960).
Mandelshtam, L.I.: Lectures on oscillation theory. Nauka, Moscow (1972) (In Russian).
Marsden, J.E., Raţiu, T.S.: Introduction To Mechanics And Symmetry: A Basic Exposition Of
Classical Mechanical Systems. Springer, New York (1999).
Chapter 7
Recognizing the Hyperbolicity: Cone Criterion
and Other Approaches
Abstract Physical and technical devices usually are not specially suited to admitting
simple mathematical proofs. So, it is vital to employ numerical tools for verification
of the hyperbolicity in systems, which potentially may possess uniformly hyper-
bolic chaotic attractors. Substantiation of hyperbolicity is essential to accounting
relevant conclusions of the mathematical theory, like availability of description in
terms of Markov partitions with a finite alphabet, or structural stability of the chaotic
attractors, which may be of natural practical significance. Section 7.1 is devoted to
approaches based on verification of a fundamental property of the hyperbolic invari-
ant sets that is transversality of mutual location of stable and unstable manifolds for
all orbits belonging to the invariant set. In Sect. 7.2 we turn to a technique of vi-
sualization of natural invariant measures along filaments of attractors, which shows
presence or absence of singularities in the distributions. For comparison, besides
models with uniformly hyperbolic attractors, some systems not relating to this class
are considered in terms of the same approaches. Final part of the chapter is devoted
to the cone criterion based on a rigorous mathematical result and appropriate for
validation of hyperbolicity in computations. Here it is reexamined in some detail,
with explanation of technical hints, and with concrete examples of its application.
7.1 Verification of transversality for manifolds
One of the main properties, which must be valid for a hyperbolic invariant set, is
transversal mutual location of stable and unstable manifolds for all orbits relating
to the invariant set (Smale, 1967; Williams, 1974; Sinai, 1979; Shilnikov, 1997;
Guckenheimer and Holmes, 1983; Katok and Hasselblatt, 1995; Afraimovich and
Hsu, 2003; Hasselblatt and Katok, 2003; Hasselblatt and Pesin, 2008). It means that
intersections between the manifolds can occur exclusively without tangencies (or
there are no intersections at all, but that is not the case for the attractors we deal
with).
In this section we consider two approaches to verification of the transversality.
142 7 Recognizing the Hyperbolicity: Cone Criterion and Other Approaches
The first one is based on visualization of the manifolds by their graphical repre-
sentation. For this purpose it is sufficient to have an accuracy about one pixel of the
picture; it is achieved easily with a simple method explained in Subsection 7.2.1. It
is worth noting, however, that more accurate methods are also available (Krauskopf,
2005). For application of this technique the most appropriate is the case of two-
dimensional Poincaré maps, when a nontrivial attractor has one positive and one
negative Lyapunov exponents.
The second method is based on computation of angles between vectors tangent to
the stable and unstable manifolds at their intersections for a representative collection
of points on the examined invariant set, with subsequent inspection of the statistical
distribution of these angles. If the uniform hyperbolicity holds, all the observed an-
gles have to be distant from zero. The procedure may be easily expanded to analyze
attractors with dimension of the stable manifolds larger than one. This method was
suggested in (Lai et al., 1993) and then used with modifications in (Hirata et al.,
1999; Anishchenko et al., 2000; Kuznetsov, 2005; Kuznetsov and Seleznev, 2006;
Ginelli et al., 2007; Kuptsov and Kuznetsov, 2009).
7.1.1 Visualization of the manifolds
Let us consider a dissipative discrete-time system governed by a map x n+1 = f(xn );

alternatively, it may be a Poincaré map for a continuous time system. The map is
assumed to be a diffeomorphism, with nontrivial attractor A. In our considerations it
is supposed that the map is available for evaluation by means of a computer program
together with the respective inverse map.
To have a possibility of simple graphical representation, we restrict ourselves
here with two-dimensional maps assuming that for the attractor A both stable and
unstable manifolds are one-dimensional: N S = 1 and NU = 1. Hence, one Lyapunov
exponent is negative, and the other one is positive.
To draw the stable and unstable manifolds with computer, we proceed as fol-
lows. First, for an arbitrarily given point on the attractor x ∈ A we obtain its image
under N-fold application of the map, i.e. x̄ = f N (x), and pre-image by means of the
inverse map x̃ = f −N (x), where N is some empirically chosen integer. Then, with
random initial conditions ỹ in a small neighborhood of x̃, by forward iteration of
the map we get a set of points y = f N (ỹ) close to the unstable manifold containing
the initial point x, which marks this manifold on the plot. In a similar way, starting
with random initial conditions ȳ in a small neighborhood of x̄, we can draw the sta-
ble manifold with a set of points y = f −N (ȳ). Formally speaking, the accuracy the
manifolds depict grows fast with increase of N; an actual selection of this number
implicates a compromise between the graphical accuracy and the numerical errors
arising as N becomes too large.
In the cases we consider here, the unstable manifolds are one-dimensional and
coincide with filaments of the attractor; so, to judge of the transversality it is suffi-
cient to plot only stable manifolds for some representative set of points on the attrac-
7.1 Verification of transversality for manifolds 143
tor. The pictures obtained in this way are illuminating and convincing to conclude
qualitatively presence or absence of touches of the manifolds, and respectively, ab-
sence or presence of the uniform hyperbolicity. Nevertheless, this should not be
regarded as a substitute for legitimate mathematical proofs!
The pictures shown in Fig. 7.1 relate to two uniformly hyperbolic attractors men-
tioned in the previous chapters. Diagrams (a) and (b) correspond, respectively, to
attractor of Plykin type in the map on the plane (3.34) and to DA-attractor of the
map on the torus (2.6). Portraits of the attractors are shown in gray, and families of
stable manifolds are depicted in black on its background. One can see clearly that
they are transversal (form nonzero angles) with the filaments of the attractors. Note
that the families of the stable manifolds visually look representative: one can easily
imagine intermediately missed curves.
Fig. 7.1 Portraits of chaotic uniformly hyperbolic attractors shown in gray with families of stable
manifolds computed as explained in the text and plotted in black. Observe clearly visible transver-
sality of the stable manifolds globally over the whole attractor to the filaments forming its structure.
The diagram (a) relates to the Plykin type attractor of the map (3.34) at ε =0.77, and (b) corresponds
to the DA-attractor of the map (2.6) at ε =0.7.
For comparison, in Fig. 7.2 the pictures are shown for another kind of attrac-
tors, which do not relate to the uniformly hyperbolic class. One is attractor of the
Hénon map (a), and the other for the Ikeda map (b), at parameters corresponding
to chaotic dynamics. There is one positive and one negative Lyapunov exponents.
The numerical technique applied to plotting the diagrams is exactly the same as
explained. Again, the families of the stable manifolds are representative; one can
easily supplement them mentally with intermediate curves. However, the pictures
look essentially distinct in comparison to those relating to the uniformly hyperbolic
case. Namely, the stable manifolds are disposed in such a way that the touches with
filaments of the attractors occur inevitably.
It is interesting to consider the Lorenz attractor, which is in some sense special,
being classified as quasi-hyperbolic, or singular hyperbolic. In Fig. 7.3 the respec-
tive diagram is shown for the attractor in the Lorenz model (2.1) at the traditional
Fig. 7.2 Portraits of chaotic attractors for Hénon map (F.4) at a=1.4, b=0.3, and Ikeda map (F.35)
at A=4.5, B=0.2, ϕ =0 shown in gray with families of stable manifolds computed as explained
in the text and plotted in black. Observe tangencies, which clearly occur either for shown or for
intermediate (mentally supplemented) curves of the stable manifold families.
Fig. 7.3 Lorenz attractor of Eqs. (2.1) at r=28, b=8/3, σ =10 (gray) and the family of stable mani-
folds (black) in the Poincaré section by the plane z = r − 1 (only crossings in direction of decrease
of z are accounted to be plotted).
values of parameters r=28, b=8/3, σ =10. The Poincaré section is used with the plane
z = r − 1. In this case, no visible tangencies occur, but there is a special situation of
mutual location of the attractor and the stable manifolds near the sharp ends of the
curved segments representing two branches of the attractor in the Poincaré section.
These are just the features, which do not allow relating the Lorenz attractor to the
uniformly hyperbolic class.
7.1.2 Distributions of angles of the manifold intersections
To my knowledge, the idea of verification of hyperbolicity by analysis of statistical

distribution of angles between local stable and unstable manifolds for dynamical
systems having one stable direction and one unstable direction at each point of the
invariant set was advanced in (Lai et al., 1993). Such approach was applied to some
saddle hyperbolic sets (Hirata et al., 1999) and to attractors not relating to the uni-
formly hyperbolic class (Anishchenko et al., 2000; Ginelli et al., 2007). Later, it was
developed in application to the case of attractor of Smale-Williams type for strobo-
scopic maps of coupled non-autonomous oscillators to embrace the case of higher-
dimensional stable manifold (Kuznetsov, 2005; Kuznetsov and Seleznev, 2006). 1
In this procedure, the angles between the directions of small perturbations are
evaluated at points of one and the same trajectory calculated forward and backward
in time. If the set of the angles does not contain values close to zero, it indicates that
the invariant set is hyperbolic. If the distribution demonstrates nonzero probability
of zero angles, then the tangencies between stable and unstable manifolds occur
that implies non-hyperbolicity. These tangencies may be interpreted as indicator of
presence of a quasi-attractor in a dissipative system under consideration that is a
complex invariant set containing beside chaotic orbits longperiod stable periodic
trajectories with extremely narrow domains of attraction (Shilnikov, 1997).
For the case of a two-dimensional map x n+1 = f(xn ) the procedure is very simple.
First, by iteration we generate a long orbit x n belonging to the attractor we analyze.
Then, with random initial conditions from the variation equations we produce two
sequences of perturbation vectors, one by iteration forward in time, ũ n+1 = f (xn )ũn ,
and the other by backward iteration, ṽ n = (f (xn ))−1 ṽn+1 , where f (xn ) designates
the matrix derivative of the map, and (f (xn ))−1 is the inverse matrix. To avoid
divergences, the vectors at each step are normalized: ũ n+1 = ũn+1 /ũn+1 and
ṽn = ṽn /ṽn. The angle between each pair of the vectors ũ n and ṽn is determined
from their inner product (dot product): θ n = arccos(ũn · ṽn ). For the collection of the
angles {θn } we can plot a histogram and judge whether there is or not statistically
significant occurrences of angles close to zero.
In Fig. 7.4 the histograms are shown obtained in computations for uniformly
hyperbolic attractors, one for the Plykin type attractor in the map (3.34) and the
other for the DA-attractor of the map on torus (2.6). Observe clearly visible gap
separating the distribution from zero angle on both plots.
It is illuminating to compare these diagrams with those relating to attractors,
which are not uniformly hyperbolic (Fig. 7.5). One is plotted for the Hénon attractor
in the map (F.4) and the other for the attractor of the Ikeda map (F.35) at their
parameters corresponding to chaotic dynamics (with one positive and one negative
Lyapunov exponents).
One more object for comparison is the Lorenz attractor classified as the quasi-
hyperbolic, or singular hyperbolic one. The computations were performed for the
Poincaré map obtained by means of numerical solution of the Lorenz equations
(2.1) at classic parameter set r=28, σ =10, b=8/3 supplemented with the Hénon in-
terpolation procedure at the intersections (Hénon, 1982). First, a sufficiently long
reference orbit belonging to the Lorenz attractor was generated, and then respec-
1 An advanced version of the approach is based on computation of covariant Lyapunov vectors and
evaluation of principal angles between the stable and unstable manifolds (Kuptsov and Kuznetsov,
2009).
Fig. 7.4 Chaotic uniformly hyperbolic attractors and histograms for distributions of angles for
intersections of local stable and unstable manifolds: (a) Plykin type attractor of the map (3.34) at
ε = 0.77 and (b) DA-attractor on the torus for the map (2.6) at ε = 0.7.
tive linearized equations were solved along this orbit in forward and inverse time to
obtain normalized perturbation vectors at crossings with the Poincaré section (the
plane z = r − 1 in the phase space; crossings by the orbits are accounted only in
the direction of decrease of z). The histogram for the distribution of the angles of
the respective vectors is shown in Fig. 7.6. In this case the distribution looks distant
from zero, like in the above-discussed cases of uniform hyperbolicity, in contrast
to the non-hyperbolic situations of Fig. 7.5. So, this method fails to distinguish the
Lorenz attractor from uniformly hyperbolic ones.
Now, let us turn to generalization of the method, which would be appropriate for
the case of larger dimensions of the stable manifolds. Then, the technique of compu-
tations for the local stable manifolds and determination of angles with the unstable
manifolds has to be reconsidered. Examples are systems with attractors of Smale-
Williams type described in the previous chapters. For instance, in the Poincaré map
of the models (4.9) and (5.5) the stable manifolds are three-dimensional, i.e. N s = 3;
for the models (3.18) and (4.17) they are two-dimensional, N s = 2.
Again, we start with computation of a representative orbit x(t) on the attractor
for sufficiently long time interval. Then, we consider linearized equations for per-
Fig. 7.5 Chaotic non-hyperbolic attractors and histograms for distributions of angles for intersec-
tions of local stable and unstable manifolds for: (a) Hénon map (F.4) at a=1.4, b=0.3, and (b) Ikeda
map (F.35) at A=4.5, B=0.2, ϕ =0. Observe non-vanishing probability of occurrence of angles close
to zero.
Fig. 7.6 A histograms for distributions of angles for intersections of local stable and unstable
manifolds for the singular hyperbolic attractor in the Poincaré map of the Lorenz model governed
by Eqs. (2.1) at r=28, b=8/3, σ =10. The Poincaré section is defined with the plane z = r − 1
accounting crossings by the orbits in the direction of decrease of the variable z.
turbations and integrate them in computations along the reference orbit, initially in
forward time, with renormalization of the vector at each next step of the Poincaré
map to have a fixed norm to preclude divergence. The result is a set of vectors {u n }.
Next, we perform solution of the linearized equations backward in time, along the
same reference orbit, for a collection of N s perturbation vectors. For them, at each
step of iteration for the Poincaré map we perform the Gram-Schmidt orthogonal-
ization and normalization procedure (like that used in the computations of the Lya-
punov exponents) to avoid divergence and predominance of one of these vectors.
(1) (2) (N )
The result is a set of vectors {v n , vn , . . . , vn s }. Now, at each n the vector u n de-
(1) (2) (N )
fines the local unstable direction, and the span of the vectors {v n , vn , . . . , vn s }
defines the local Ns -dimensional stable manifold. To evaluate the angle θ n between
the manifolds, we, first, determine a vector z n orthogonal to the stable manifold by
solving the linear system of Ns equations
(1)
zn · vn = 0,
(2)
zn · vn = 0,
............
(Ns )
zn · v n = 0.
Then, we calculate the angle β n ∈[0,π /2] between u n and zn from cos βn = |un · zn |/
|un ||zn | and, finally, set θn = π /2 − βn . Plotting the histograms for the computed
angles θn one can judge whether the distribution is distant from zero that testifies in
favor of hyperbolicity, or there is non-vanishing probability of occurrence of near-
zero angles that means non-hyperbolicity.
Examples are presented in Fig. 7.7, where the panels (a) and (b) correspond to
models with three-dimensional Poincaré maps, and the panels (c), (d) to the models
with four-dimensional Poincaré maps.
Panel (a) relates to the model with attractor of Smale-Williams type constructed
in Sect. 3.2 as a flow system governed by a set of switched differential equations
of the third order. The stroboscopic Poincaré map is expressed explicitly by (3.18).
Parameter values are d 1 = d2 = 2, μ = 2.
Panel (b) presents data for attractor of Smale-Williams type in the three-dimensi-
onal Poincaré map of the autonomous model developed in (Kuznetsov and Pikovsky,
Fig. 7.7 Attractors of Smale-Williams type and histograms for distributions of angles between
local stable and unstable manifolds: (a) model map (3.16) at d1 = d2 = 2, μ = 2; (b) three-
dimensional Poincaré map of the autonomous model (5.13); (c) the stroboscopic map of alternately
excited non-autonomous oscillators (4.9) at ω0 = 2π , T = 6, A = 5, ε = 0.5; (d) Poincaré map of
the model (5.5) at ε = 0.1, a = 0.6; developed on the base of predator-prey system, ω0 = 2π ,
ε = 0.3.
2007) on the base of predator-prey system and discussed in Sect. 5.3 (see Eqs.
(5.13)). Parameter values are ω 0 =2π , ε =0.3.
Panel (c) corresponds to the basic model governed by Eqs. (4.9) of two non-
autonomous alternately excited van der Pol oscillators transferring the excitation
each other with transformation of the phase approximating by the expanding cir-
cle map. The stroboscopic Poincaré map is four-dimensional. Parameter values are
ω0 =2π , T =6, A=5, ε =0.5.
Panel (d) relates to the autonomous model with attractor of Smale-Williams type
introduced in Sect. 5.1 (see Eqs. (5.3)). The Poincaré map is four-dimensional. Pa-
rameter values are ω 0 =2π , ε =0.3.
Observe that all histograms of Fig. 7.7 manifest a well expressed gap separating
the distributions from the domain of angles close to zero. It indicates absence of
touches of the stable manifolds with unstable ones associated with hyperbolic nature
of the attractors in agreement with other methods of their analysis.
7.2 Visualization of invariant measures
If we consider dynamics in terms of evolution of a cloud of representative points in

phase space of a dissipative dynamical system, there is some stationary distribution
arising in a limit of large observation time; according to the accepted terminology,
this is the invariant measure associated with the attractor. In accordance with theo-
rem of Krylov and Bogolyubov (Anosov, 1994), natural invariant measure, as rela-
tive probability of residence of a representative orbit on the measured set in phase
space, may be introduced under very general circumstances; but for non-hyperbolic
attractors it is characterized by presence of local singularities. Hyperbolic chaos
takes place in a situation of evolution of the clouds including stretching, folding,
and compression in a perfect way, without breaks and without appearance of lo-
cal seals. In these cases we deal with the invariant measure of Sinai-Ruelle-Bowen
(Sinai, 1972; Bowen, 1975; Ruelle, 1976). It is absolutely continuous that means
the smooth distribution along the filaments of the attractor occurs. Hence, visualiza-
tion of the distributions of the invariant measures is a useful and illuminating tool to
distinguish hyperbolic and non-hyperbolic attractors.
In computations the simplest way to visualize the invariant measure is to define
a grid of cells in the phase space domain containing the attractor, and in the course
of the simulation to accumulate array of the quantities indicating relative time of
residence in all the cells. Afterwards, probabilities of the residence are estimated
and plotted in appropriate coordinates (Tél and Gruiz, 2006). A disadvantage of this
method is presence of relatively large statistical fluctuations of the probability esti-
mates, that makes the distinction between smooth and singular distributions visually
not so impressive. For this reason, it is worth turning to modification of the method,
which we consider here in application to attractors with one-dimensional unstable
manifolds.
The modified approach consists in plotting just the one-dimensional probability
distribution along the one-dimensional unstable manifold, that is the same, along
a filament of the attractor. In application to some map g(x) corresponding to a
discrete-time system, or to a Poincaré map of a continuous-time system, the al-
gorithm operates as follows (Kuznetsov and Sataev, 2011).
7.2 Visualization of invariant measures 151
(0) (1)
1. Select a point on the attractor x 0 , and another one x 0 at a small distance
Δx = ε0 (practically, it may be ε 0 ∼ 0.001).
2. Consider iterates of the map g(x) starting from both initial points, while the
(0)
distance between them becomes greater than 10ε 0 ; the resulting points will be x n
(1)
and xn .
(0) (0) (1) (0)
3. Then, update the initial pair of points and set x 0 = xn and x0 = xn +
(1) (0) (1) (0)
ε (xn − xn )/xn − xn .
4. Repeat the steps 2 and 3 quite a lot of times (practically, about 10 3 ); then both
initial points will sit on the unstable manifold.
5. The segment between the initial points obtained is divided onto a large number
of equal small segments by inserting intermediate points (the number of them may
be about 10 5 )
6. Starting from all these points iterate the map, while the maximum distance for
points of some adjacent pair remains less than a prescribed constant, say, 50ε .
7. Each of the resulting segments is attributed to the value of the inverse of its
length, which is proportional to the density measure in this interval.
8. For visualization, the resulting values are normalized to the maximal one to get
the numbers in the unit interval from 0 to 1, and plotted in appropriate coordinates
in proportional, or logarithmic scale.
The pictures obtained are representative, if the finally obtained segment of the un-
stable manifold covered by the examined points is long enough to cover the attractor
for many times, revealing its thin fractal transverse structure up to a sufficiently deep
level.
Diagrams obtained by means of the above computational scheme in application
to model systems with uniformly hyperbolic attractors discussed in the previous
chapters are shown in Figs. 7.8—7.10.
Figure 7.8 shows the distribution of the invariant measure for DA-attractor in the
map (2.6) at ε =0.7. The unit square in the horizontal plane corresponds to unfolding
of the torus surface, on which the map is defined. Figure 7.9 relates to the Plykin
type attractor on the plane for the model governed by the map (3.34) at ε =0.77.
Figure 7.10 corresponds to the attractor of Smale-Williams type in the stroboscopic
Poincaré map of the system of two alternately excited non-autonomous van der Pol
equations (4.9) at ω 0 = 2π , T =6, A=5, ε =0.5. In this case the Poincaré map is four-
dimensional; so, the distribution of the invariant measure is shown over a projection
of the attractor onto the plane of two dynamical variables (x, ẋ/ω 0 ). Observe that in
all these cases, which relate to the uniformly hyperbolic attractors, smooth depen-
dence of the probability distribution takes place along the filaments of the attractors.
In the used technique of visualization it is seen very clear.
For comparison, let us discuss examples of non-hyperbolic attractors, e.g. those
relating to the Hénon and Ikeda maps. The diagrams obtained for them by means of
the above computational scheme are shown in Figs.7.11 and 7.12. The distributions
of probability along the filaments of these attractors look essentially non-uniform,
and contain singularities represented by narrow and high peaks. To show them in
comparable scale with the rest part of the distribution functions we use the logarith-
Fig. 7.8 Diagram illustrating the distribution of the invariant measure of Sinai-Ruelle-Bowen
along unstable manifolds of a uniformly hyperbolic attractor of DA type for the map (2.6) at ε =0.7.
along unstable manifolds of Plykin type attractor on the plane obtained for the map (3.34) at
ε =0.77.
mic scale along the vertical axis. The plots are in impressive contrast with the case
of the uniformly hyperbolic attractors.
Finally, it is interesting to consider a special case of quasi-hyperbolic (or singular-
hyperbolic) Lorenz attractor. The respective diagram is shown in Fig. 7.13 obtained
for the two-dimensional Poincaré map of the Lorenz model (2.3) at parameter values
7.2 Visualization of invariant measures 153
along unstable manifolds of attractor of Smale-Williams type in the stroboscopic Poincaré map
of the system of two alternately excited non-autonomous van der Pol equations (4.9) at ω0 = 2π ,
T =6, A=5, ε =0.5.
Fig. 7.11 Diagram illustrating the distribution of the invariant measure for the attractor of the
Hénon map (F.4) at a=1.4, b=0.3. Logarithmic scale is used along the vertical axis to visualize
both singular and non-singular components of the distribution.
r=28, b=8/3, σ =6. The Poincaré section is defined with the plane z = r − 1 account-
ing orbits crossings this surface in the direction of decrease of the variable z. In
Fig. 7.12 Diagram illustrating the distribution of the invariant measure for the attractor of the Ikeda
map (F.35) at A=3, B=0.2, ϕ =0. Logarithmic scale is used along the vertical axis to visualize both
singular and non-singular components of the distribution.
Fig. 7.13 Diagram illustrating the distribution of the invariant measure for the Lorenz attractor of
the model (2.3) at r=28, b=8/3, σ =6 in the Poincaré section by the plane z = r − 1.
this case the picture does not show singularities of the distribution of the invariant
measure, like in two previous examples. However, there is a visible difference in
comparison with the case of uniformly hyperbolic attractors. It is vanishing of the
distribution function at some end points of the fragments representing the unstable
7.3 Cone criterion and examples of its application 155
manifolds in the Poincaré section; for uniformly hyperbolic attractors the distribu-
tions are locally smooth and not vanishing along the entire unstable manifolds.
To conclude, the developed technique provides an effective tool for detection of
hyperbolic or non-hyperbolic nature of attractors in numerical calculations, at least
on a qualitative level.
7.3 Cone criterion and examples of its application
In Chap. 1, reviewing content and conclusions of the hyperbolic theory, we told

about the cone criterion (Sinai, 1979; Hasselblatt and Pesin, 2008), which provides
a necessary and sufficient condition for hyperbolicity of an invariant set, appropriate
for computer verification. The first known attempt of application of this criterion to
verification of chaotic nature of attractor in computations was presented in (Sinai
and Vul, 1981). Recently, this criterion was applied to computer validation of hy-
perbolicity of attractors in models suggested in Chap. 3—6 (e.g. (Kuznetsov and
Sataev, 2007; Kuznetsov et al., 2007; Kuznetsov, 2009; Kuznetsov et al., 2010)).
Let us return to this issue and consider it in more constructive manner to explain
concretely the content of the computations.
7.3.1 Procedure of verification of the cone criterion
We will discuss here the cone criterion only for attractors with one positive Lya-
punov exponent in application to maps represented by diffeomorphisms. They may
be thought of as relating either to discrete-time dynamics, or to description of
continues-time systems in terms of Poincaré sections.
Let us have a dissipative map x n+1 = g(xn ), where x is an N-dimensional vector,
and this map possesses attractor A, which has to be analyzed. Additionally, we define
the linear map un+1 = g (xn )un governing infinitesimal perturbations. Here g (x) is
the matrix derivative of the map g at the point x; it is a square matrix of size N × N
called the Jacobian matrix. In our considerations it is assumed that the map g and
the matrices g (x) are available for evaluation for the dynamical system under study
by means of a computer program.
Let us consider details of the procedure of verification for the cone criterion at
some reference point in the phase space x. To proceed, we need to specify the image
point x̄ = g(x), and the pre-image point x̃, so that x = g(x̃). Let γ be some fixed
constant larger than 1.
The matrix derivative of the map g at the reference point v = g (x) acts in the
linear space of vectors u = {u 0 , u2 , . . . , uN−1 } called the tangent space at x. Via the
auxiliary symmetric matrix b̂ = vT v (the superscript T means the transpose), the
Euclidean norm of the vector ū = vu may be expressed as
ū2 = uT b̂u. (7.1)
A set of vectors
Sx = {u|uT b̂u γ 2 uT u} (7.2)
represents the expanding cone at the point x. With the same matrix b̂ one can define
a set
Cx̄ = {u|γ 2 uT b̂u uT u}. (7.3)
This set corresponds to the pre-image for the contracting cone defined at the point
x̄ = g(x).
Now, consider the matrix derivative for the inverse map w = (g −1 (x)) . Note
that this matrix may be evaluated without handling the inverse map itself, via the
inverse matrix of the derivative of the map g at the pre-image point x̃: w = (g (x̃))−1 .
Introducing symmetric matrix â = wT w we represent the norm of the vector ũ = wu
as
ũ2 = uT âu. (7.4)
With the matrix â one can define a set
Cx = {u|uT âu γ 2 uT u}, (7.5)
which represents the contracting cone at x, and a set
Sx̃ = {u|γ 2 uT âu uT u}, (7.6)
which corresponds to an image for the expanding cone relating to the point x̃ =
g−1 (x).
In computations it is necessary to check, first, existence of nonempty cones sat-
isfying the definitions, and, second, invariance of the cones that means validity of
the inclusions S x̃ ⊂ Sx and Cx̄ ⊂ Cx .
The required inclusions may be formulated in terms of quadratic forms associated
with the matrices b = b̂ − Γ2 ê and a = â− Γ−2 ê, where ê is a unit matrix. A constant
factor Γ is assumed to be equal to either γ , or 1/γ , considering the expanding, or
contracting cones, respectively. Obviously, the equations
N−1
∑ bmn um un = 0 (7.7)
m,n=0
and
N−1
∑ amn um un = 0 (7.8)
m,n=0
determine the borders of the cones.

The N × N matrix b̂ is symmetric and positively definite (as it is a product of two
mutually transposed matrices). Eigenvalues of this matrix are all positive, and one
can determine respective orthogonal basis, the set of unit vectors d 0 , d1 , . . . , dN−1 .
We suppose that the eigenvalues are enumerated in the decreasing order. Then, using
the matrix D = (d0 , d1 , . . . , dN−1 ) we transform the matrix b̂ to the diagonal form
⎛ 2 ⎞
Λ0 0 · · · 0
⎜ 0 Λ2 · · · 0 ⎟
⎜ 1 ⎟
D b̂D = ⎜ . .. ⎟ .
T
. (7.9)
⎝ .. .. . ⎠
0 0 · · · Λ2N−1
For the case we deal with, there should be one expanding direction, and others con-
tracting, so, we must check in computations that Λ 20 > Γ2 , Λ21 , . . . , Λ2N−1 < Γ2 . If this
is the case, the matrix
⎛ 2 ⎞
Λ0 − Γ2 0 ··· 0
⎜ 0 Λ21 − Γ2 · · · 0 ⎟
⎜ ⎟
DT (b̂ − Γ2 ê)D = ⎜ .. .. .. ⎟ (7.10)
⎝ . . . ⎠
0 0 · · · Λ2N−1 − Γ2
has one positive and all other negative elements on the diagonal. By an additional
scale change along the coordinate axes, the matrix is transformed to the canonical
form: ⎛ ⎞
1 0 ··· 0
⎜ 0 −1 · · · 0 ⎟
⎜ ⎟
ST DT (b̂ − Γ2ê)DS = ⎜ . . .. ⎟ , (7.11)
⎝ .. .. . ⎠
0 0 · · · −1
where
⎛ ⎞
1/ Λ20 − Γ2 0 ··· 0
⎜ ⎟
⎜ ⎟
⎜ 0 1/ Γ2 − Λ21 · · · 0 ⎟
S=⎜
⎜ .. .. ..
⎟.
⎟ (7.12)
⎜ . . . ⎟
⎝ ⎠
0 0 · · · 1/ Γ2 − Λ2N−1
In the new variables u = ST DT u the quadratic form (7.7) reduces to
u2 2 2
0 − u1 − · · · − uN−1 = 0. (7.13)
Under the same transformations, Equation (7.8) converts to

N−1
∑ amn um un = 0, (7.14)
m,n=0
where amn are elements of the matrix a = SDT (â − Γ−2ê)DS.

Let us examine vectors u = {1, c1 , c2 , . . . , cN−1 }. The relation (7.13) corresponds
to a unit sphere in the space of N − 1 coefficients: c 21 + c22 + · · · + c2N−1 = 1. On the
other hand, the expression (7.14) for such vectors can be rewritten as
N−1 N−1
a00 + ∑ (a0m cm + am0cm ) + ∑ amn cm cn = 0. (7.15)
m=1 m,n=1
The required inclusion of the cones means that the last equation has to determine
a surface in a form of ellipsoid placed inside the unit sphere. We prefer to use a
sufficient condition for such situation, which is easy for derivation and application.
First, we determine a center point c̄ for the second-order hyper-surface from the
set of equations
a11 c̄1 + a12c̄2 + · · · + a1,N−1 c̄N−1 = −a10 ,

a21 c̄1 + a22c̄2 + · · · + a2,N−1 c̄N−1 = −a20 ,
(7.16)
············
aN−1,1 c̄1 + aN−1,2 c̄2 + · · · + aN−1,N−1, c̄N−1 = −aN−1,0 .

and evaluate its distance from the center of the unit sphere:

ρ = c̄21 + c̄22 + · · · + c̄2N−1 . (7.17)
By transfer of the origin to the point c̄ we rewrite Eq. (7.15) in variables c̃ m =

cm − c̄m as
N−1
∑ amn c̃m c̃n = R2 . (7.18)
m,n=1
For the symmetric matrix

⎛ ⎞
a11 a12 ··· a1,N−1
⎜ a21 a21 ··· a2,N−1 ⎟
⎜ ⎟
h=⎜ .. .. .. ⎟ (7.19)
⎝ . . . ⎠
aN−1,1 aN−1,1 · · · aN−1,N−1
of size (N − 1) × (N − 1), all eigenvalues l 1 , l2 , . . ., lN−1 are real. Let q1 , . . . , qN−1

be respective eigenvectors. Diagonal representation of this matrix is obtained by
orthogonal coordinate transformation. To do this, we set h = QT hQ and c = QT c,
where Q = (q1 , . . . , qN−1 ). Then, Equation (7.15) reduces to
N−1
∑ lm c2m = H, (7.20)
m=1
where
N−1 N−1
H = −a00 − ∑ (a0mc̄m + am0c̄m )− ∑ amn c̄m c̄n . (7.21)
m=1 m,n=1
If all the eigenvalues l m are of the same sign as the sign of the quantity H, Equation
(7.20) determines an ellipsoid with the largest semi-axis
expressed via the eigen-
value of minimal magnitude l min ; the length equals H/lmin . A sufficient condition
for the ellipsoid to belong to the interior of the unit sphere is the inequality

ρ + H/lmin < 1. (7.22)
If all the above conditions are true simultaneously both for Γ = γ and Γ = 1/γ
with some γ > 1, one can conclude about correct inclusions for the expanding and
contracting cones at the reference point x.
Indeed, in the variables c the cross-section of the expanding cone S x is the closed
unit ball in the space of dimension N − 1, and cross-section of the cone S x̃ is rep-
resented by closure of the interior of the small ellipsoid obtained at Γ = γ . So, the
required inclusion S x̃ ⊂ Sx is valid, as shown schematically in Fig. 7.14(a). On the
other hand, cross-section of the contracting cone C x is a closure of exterior of the
small ellipsoid obtained at Γ = 1/γ . Cross-section of the cone C x̄ corresponds to
a closure of exterior of the unit sphere. Hence, the inclusion C x̄ ⊂ Cx is valid as
illustrated in Fig. 7.14(b).
Fig. 7.14 Schematic illustration of disposition of cross-sections of the expanding cones Sx̃ ⊂ Sx
(a) and contracting cones Cx̄ ⊂ Cx (b). A circle circumscribed around the ellipse corresponds to
circumscribed ball around ellipsoid that should be located inside the unit ball if the condition (7.22)
is valid.
For the particular case of two-dimensional maps (N=2), the last part of the pro-
cedure is simplified essentially. In this case the cross-sections of the cones corre-
spond to one-dimensional intervals, for which the appropriate inclusions have to be
checked. The matrix h degenerates to a single real number a 11 , and lmin = l1 = a11 .
Then, the relation (7.22) represents correct necessary and sufficient condition for
the required inclusions to hold.
The procedure we have described relates to a single reference point x, but we
need to be sure that all the required conditions are true on the whole attractor A.
One way to organize the calculations may be used if an absorbing domain D is

specified from some preliminary analysis, or in computations, namely, g(D) ⊂ Int D,
∞

and A = gm (D). We may arrange scanning of this domain D on a grid in the N-
m=1
dimensional phase space and perform the above computational procedure at each
node of the grid, with one and the same constant γ > 1. What is essential is selec-
tion of sufficiently small size of the grid cells. All relevant involved characteristics
(elements of the matrices, eigenvalues, and eigenvectors) have to be nearly constant
at the scale of a single cell. If all the imposed conditions are valid at the nodes of
the grid covering the domain D, we may be sure that they are true on the attractor
A as it belongs to this domain. One disadvantage of this method is a necessity of
preliminary knowledge of the absorbing domain. Another is a rapid increase of the
volume of computations with the growth of the phase space dimension N.
Alternative approach consists in verification of the conditions of the cone crite-
rion for a representative set of points on the attractor itself. These points are obtained
step by step from iteration of the map g. In this case we do not need to determinate
the absorbing domain, and the method can be easily applied to systems of high di-
mensions N.
At first glance, the second method might seem much less convincing than the first
one. However, the practice of calculations shows that being essentially easier in use,
it allows reliable recognizing of situations of hyperbolic chaos, not more worthful
than the first approach. It may be explained as follows. For smooth systems, the
objects involved in the procedure of verification of the cone criterion in the region
of interest depend on the state variables in smooth manner as they are determined
by the dynamics in finite time intervals. It means that validity of the conditions at
some point x with constant γ distant from 1 implies that the conditions of the cone
criterion hold as well in a neighborhood of x (the wider, the larger the value |γ − 1|).
Hence, a positive result of the test for a representative set of points implies validity
of the criterion on the entire attractor, if it is covered completely by the union of the
mentioned neighborhoods. Practically, such situation is achieved by increase of the
number of representative points on the attractor subjected to the test, i.e. the number
of iterations of the map under consideration.
It is worth noting that the computational procedure may be recommended in
some cases for application to k-fold iteration of the map, where k is some integer, as
the hyperbolicity recognition often becomes better with increase of k. On the other
hand, at large k the involved operations of linear algebra become ill-conditioned,
because of coexistence of divergence and convergence for nearby orbits at the hy-
perbolic trajectories. So, the most appropriate selection of k may require some com-
promise.
7.3.2 Examples of application of the cone criterion
Let us turn to concrete examples of application of the computational verification of

the cone criterion to model systems discussed in previous chapters. For all of them
we present materials in a unified form, basing on the computations at representa-
tive sets of points on the attractors. For some of the models the computations were
performed as well with the technique of scanning the absorbing domains over ap-
propriate grids; for these results we refer the reader to publications (Kuznetsov and
Sataev, 2007; Kuznetsov et al., 2007; Kuznetsov, 2009; Kuznetsov et al., 2010).
First, let us consider the system governed by switched differential equations in-
troduced in Sect. 3.2. It is rather a simple example for application of the computa-
tional procedure because the stroboscopic Poincaré map may be written down ex-
plicitly so does the inverse map (see formulas (3.18) and (3.19)). The map is three-
dimensional, and it corresponds to the minimal dimension, for which the Smale-
Williams type attractor can occur. The matrix derivatives may be obtained analyti-
cally:
⎛ ⎞ ⎛ ⎞
∂ xn+1 ∂ xn+1 ∂ xn+1 ∂ xn−1 ∂ xn−1 ∂ xn−1
⎜ ∂ xn ∂ yn ∂ zn ⎟ ⎜ ∂ xn ∂ yn ∂ zn ⎟
⎜ ⎟ ⎜ ⎟
⎜ ∂ y ∂ y ∂ yn+1 ⎟ ⎜ ∂ y ∂ y ∂ yn−1 ⎟
g (xn ) = ⎜ ⎟ and (g−1 (x)) = ⎜ n−1 ⎟.
n+1 n+1 n−1
⎜ ∂ xn ∂ yn ∂ zn ⎟ ⎜ ∂ xn ∂ yn ∂ zn ⎟
⎜ ⎟ ⎜ ⎟
⎝ ∂z ⎠ ⎝ ∂z ⎠
n+1 ∂ zn+1 ∂ zn+1 n−1 ∂ zn−1 ∂ zn−1
∂ xn ∂ yn ∂ zn ∂ xn ∂ yn ∂ zn
Explicit expressions are too cumbersome, and we omit them for brevity.
Figure 7.15 presents data illustrating all essential components of the procedure
of verification of the cone criterion. The quantities are plotted obtained in the course
of the computations at successively visited points on the attractor in the sustained
chaotic regime. The horizontal axis corresponds to the angular coordinate counted
around the toroid containing the attractor (Fig. 3.9). In panel (a) one can see a plot
for three eigenvalues of the matrix b̂ = vT v = g T g . Observe that the first eigen-
value is straightly larger than 1 on the entire attractor, while the others are straightly
less than 1. It indicates presence of the one-dimensional expansion and the two-
dimensional contraction in the phase space, and the families of expanding and con-
tracting cones may be defined all over the attractor. Panel (b) presents the data for
eigenvalues of the auxiliary matrix h (see (7.19)), which is of size 2 × 2 in this case.
Observe that the ratios of the eigenvalues to the quantity H (see (7.20)) for both of
them are positive all over the attractor, that means that in the cross-section the cones
produce the ellipses. Although the diagram is drawn for Γ=1, all filaments observed
on the plot are clearly separated from zero, so, the situation will be qualitatively the
same for values of Γ slightly larger or less than 1. Finally, panel (c) gives the evi-
dence for the correct inclusions for the cones. The value ρ + H/lmin is plotted (see
(7.22)), which is the maximal distance between the centre of the unit circle and a
point on the small ellipse that must be inside the unit circle for the criterion to hold.
Observe that this positive quantity is straightly less than 1. This situation occurs
Fig. 7.15 Relevant quantities computed in the course of verification of the cone criterion in depen-
dence on the angular coordinate counted around the coils of the solenoid for the system governed
by the three-dimensional map (3.18): (a) eigenvalues of symmetric matrix used for definition of
the cones; (b) eigenvalues of the auxiliary matrix used for analysis of mutual location of cross-
sections of the cones; (c) the distance, which is less than 1 in the case of correct inclusions for the
expanding cones (the left diagram) and for the contracting cones (the right diagram). Parameters
values assigned are d = d2 =2, μ =2.
for Γ = γ and 1/γ with some γ larger than 1. For example, it is true at Γ 2 = 1.36;
so, the
√ required inclusions for the expanding and the contracting cones are valid at
γ = 1.36 > 1.
In Fig. 7.16 portrait of the attractor in the stroboscopic Poincaré section is shown
supplemented on the periphery with diagrams indicating the mutual location for
the expanding cones and the images of the expanding cones at several points of
the attractor. The expanding cone is represented by a unit disc, and the image of
the expanding cone from the previous point is a small ellipse inside the disc. The
circumscribed circle around the ellipse in all cases belongs to the unit disc as well;
this is the content of the sufficient condition (7.22).
The next example we consider is the model map with DA attractor (2.6). This
is a two-dimensional map for the state vector x n = {pn , qn }. The matrix derivative
may be obtained analytically,
Fig. 7.16 Illustrations of validity of the criterion for the expanding cones for γ2 =1.36 checked at
seven representative points for the attractor of map (3.18) depicted in the centre. Parameters values
assigned are d = d2 = 2, μ = 2. The expanding cone at each point is represented by a unit disc,
and the image of the expanding cone from the previous point is a small ellipse inside the disc.
The circumscribed circle around the ellipse is used in formulation of the sufficient condition of the
correct inclusion for the cones (7.22).
⎛ ⎞
∂ pn+1 ∂ xn+1
⎜ ∂ xn ∂ xn ⎟
g (xn ) = ⎜
⎝ ∂ xn+1
⎟.
∂ xn+1 ⎠
∂ xn ∂ xn
While ε < 89 , its Jacobian determinant is positive, so, the map is invertible, but the
analytical expression for the inverse map is not available.
Figure 7.17 presents data relating to the computational procedure of verification
of the cone criterion for the map (2.6) at ε = 0.7. The abscissa axis corresponds to
the coordinate p counted around the torus, on which the map is defined. In panel
(a) one can see a plot for two eigenvalues of the matrix b̂ = vT v = g T g . The first
eigenvalue is everywhere larger than 1, and the second is straightly less than 1. In the
two-dimensional case, instead of the matrix h we have simply a single number l 1 ; its
ratio to the quantity H (see (7.20)) is plotted in panel (b). Observe that it is positive
all over the attractor. The
panel (c) gives the evidence for the correct inclusions for
the cones. The value ρ + H/lmin in this case is the maximal distance between the
centre of the unit interval and a border of smaller interval that must be inside the
unit interval for the criterion to hold. This situation occurs for Γ = γ and 1/γ , for
example, at γ 2 = 5, as seen from the right diagram (c).
Fig. 7.17 Relevant quantities for the cone criterion in dependence on the coordinate around the
torus for the system governed by the map (2.6) at ε = 0.7: (a) two eigenvalues of the symmetric
matrix used for definition of the cones; (b) the auxiliary value used for analysis of mutual location
of intervals representing cross-sections of the two-dimensional cones; (c) the distance, which is
less than 1 in the case of correct inclusions for the expanding cones (the left diagram) and for the
contracting cones (the right diagram).
Let us examine the Plykin type attractor on the plane in the two-dimensional
map introduced in Sect. 3.2. In this case, again the map has an analytical expression
(3.34). Also, an inverse map can be derived analytically, so do the matrix derivatives
T and (T )−1 . In the course of computations performed at ε =0.77 it appears that
degree of non-uniformity of transformations over the attractor in the representation
on the plane is quite high; so, the cone criterion in the version we formulate here
is successful only for at least four-fold iteration of the original map. Panel (a) of
Fig. 7.18 shows a plot for two eigenvalues of the matrix b̂ = vT v = g T g , where
g = T4 , and the abscissa axis corresponds to the coordinate X used in the definition
of the map. Note that the first eigenvalue is straightly higher than 1, and the second
is straightly less then 1. The ratio of the auxiliary values l and H is plotted in panel
(b) and is positive all over the attractor. Finally,
panel (c) gives evidence for the
correct inclusions for the cones. The value ρ + H/lmin is straightly less than 1 for
both Γ = γ and 1/γ with γ 2 = 1.2.
Now we turn to more complex examples, for which the Poincaré maps and
derivative matrices have to be obtained in computations from numerical integration
of the respective differential equations.
Fig. 7.18 Relevant quantities for the cone criterion in dependence on the coordinate for the Plykin
type attractor of the map (3.34) at ε = 0.77: (a) two eigenvalues of the symmetric matrix used
for definition of the cones; (b) the auxiliary value used for analysis of mutual location of intervals
representing cross-sections of the two-dimensional cones; (c) the distance, which is less than 1 in
the case of correct inclusions for the expanding cones (the left diagram) and for the contracting
cones (the right diagram).
The model constructed on a base of the predator-prey equations was intro-

duced in the joint work (Kuznetsov and Pikovsky, 2007) and discussed in Chap.
5. It is governed by Eqs. (5.8), which may be rewritten as a set of four differen-
tial equations for real variables x = Re a 1 , u = Im a1 , y = Re a2 , v = Im a2 . The
parameter values assigned are ω 0 = 2π , ε = 0.3. As mentioned in Chap. 5, it is
appropriate to define the Poincaré section with a hyper-surface in the 4D phase
space specified by the equation y 2 + v2 = c, with account of crossings in direction
of increase of y 2 + v2 . In the parameter range of interest, an appropriate value of
the constant is c = 2.5. In the three-dimensional hyper-surface of the cross-section
we use coordinates {x, u, z}, where z = (2π ) −1 arg[(y + iv)/(x + iu)]. The procedure
for computation of the Poincaré map is arranged as a computer program basing on
numerical solution of differential equations (5.8) complemented with interpolation
consistent in accuracy with the applied finite-difference scheme (Hénon, 1982). The
initial conditions are formulated in terms of the variables {x, u, z} at the Poincaré
section. As the part of the trajectory starting and finishing at the Poincaré section
has been computed, this results in the updated variables x̄, ū, z̄. Then, the variational
equations are integrated along the same trajectory with initial conditions x̃, ũ, z̃; in
the course of integration the original variables and their variations are used. After
returning to the Poincaré section, the updated perturbation vector is obtained. From
the last one the component associated with a shift along the trajectory is subtracted.
The rest vector is expressed in terms of the updated variations x̃, ũ, z̃.
For evaluation of the matrix derivative one performs the procedure of compu-
tation three times along the same reference trajectory with initial conditions for
the variations (1,0,0), (0,1,0), and (0,0,1). Three resulting final vectors form three
columns of the derivative matrix.
Figure 7.19 illustrates verification of the cone criterion for the attractor of Smale-
Williams type in the Poincaré map of the system (5.3). Panel (a) shows a plot for
three eigenvalues of the matrix b̂ = vT v. One of them is straightly higher than 1,
and others are straightly less than 1. Panel (b) presents the data for eigenvalues of
the auxiliary matrix h of size 2 × 2 in this case. The ratios of the eigenvalues to the
quantity H are positive all over the attractor. The diagram is drawn for Γ = 1, but all
observed filaments of the formation on the picture are clearly separated from zero,
so, the situation will be the same for Γ at least slightly higher or less than 1. Panel (c)
Fig. 7.19 Relevant quantities computed in the course of verification of the cone criterion in depen-
dence on the angular coordinate counted around the coils of the solenoid for the system governed
by differential equations (5.3): (a) eigenvalues of symmetric matrix used for definition of the cones;
(b) eigenvalues of the auxiliary matrix used for analysis of mutual location of cross-sections of the
cones; (c) the distance, which is less than 1 in the case of correct inclusions for the expanding
cones (the gray plot) and for the contracting cones (the black plot). Parameters values assigned are
ω0 = 2π , ε = 0.3.

indicates that the value ρ + H/lmin is straightly less than 1. This situation occurs
at Γ2 = 1.3, as well as at Γ2 = 1/1.3. It means that the required √ inclusions for both
the expanding and the contracting cones are valid at γ = 1.3 > 1. In (Kuznetsov et
al., 2007) the cone criterion for this model was verified by scanning the absorbing
domain on a grid of small cell size and was found valid at γ =1.3.
The next example is the system of two coupled alternately excited non-autonomous
self-oscillators introduced by the author in (Kuznetsov, 2005). It is governed by
Eqs. (4.9). The parameter values selected for detailed analysis are ω 0 = 2π , A =
5, T = 6, ε = 0.5. All the time-dependent coefficients in the equations have the
least common period T , so, the dynamics may be described in terms of strobo-
scopic Poincaré map, which transforms the initial four-dimensional state vector x n =
(x, ẋ/ω0 , y, ẏ/ω0 )|t=nT = (x, u, y, v)|t=nT to the state vector x n+1 = (x, u, y, v)|t=(n+1)T .
It is easily performed with numerical solution of the differential equations by a stan-
dard finite-difference method of Runge-Kutta assuming the time interval T contains
an integer number of time steps of the finite-difference scheme. Together with the in-
tegration, along the reference orbit, a collection of four linearized equations for vari-
ations (x̃, ũ, ỹ, ṽ) is integrated with initial conditions (1,0,0,0), (0,1,0,0), (0,0,1,0),
Fig. 7.20 Relevant quantities for the cone criterion in dependence on the angular coordinate
counted around the solenoid for the system governed by differential equations (4.9): (a) eigen-
values of the symmetric matrix used for definition of the cones; (b) eigenvalues of the auxiliary
matrix used for analysis of mutual location of cross-sections of the cones; the distance, which is
less than 1 in the case of correct inclusions for the expanding cones (c) and for the contracting
cones (d). Parameters values assigned are A = 5, T = 6, ε = 0.5.
and (0,0,0,1) to obtain columns of the derivative matrix for the stroboscopic four-
dimensional map.
As stated in Sect. 4.2, attractor of the Poincaré map at the selected parameters is
a kind of Smale-Williams solenoid embedded in the four-dimensional state space.
Fig. 7.20 illustrates verification of the cone criterion for a set of representative points
on the attractor obtained by successive iteration of the Poincaré map for the system
(4.9). Panel (a) shows a plot for four eigenvalues of the matrix b̂ = vT v, one is
straightly higher than 1, and three others are straightly less than 1. Panel (b) presents
eigenvalues for the matrix h of size 3×3; the ratios of the eigenvalues to H are
positive all over the attractor. Panels (c) and (d) show that the value ρ + H/lmin is
√ less than 1 at Γ = 1.35 and at Γ = 1/1.35. So, the cone criterion is valid
2 2
straightly
at γ = 1.35 > 1.
An absorbing domain for this attractor was specified (see formula (4.16) and
Fig. 4.3). For expanding cones the criterion was verified on a grid covering this
absorbing domain D, and for contracting cones that was verified for the domain
T2 (D) (Kuznetsov and Sataev, 2007). It was found valid, e.g. at γ 2 = 1.1.
Fig. 7.21 Relevant quantities for the cone criterion in dependence on the angular coordinate
counted around the solenoid for the system governed by differential equations (5.5): (a) eigen-
values of the symmetric matrix used for definition of the cones; (b) eigenvalues of the auxiliary
matrix used for analysis of mutual location of cross-sections of the cones; (c) the distance, which
is less than 1 in the case of correct inclusions for the expanding cones (the gray plot) and for the
contracting cones (the black plot). Parameters values assigned are α = 0.6, ε = 0.1.
References 169
A final example of attractor of Smale-Williams type relates to the five-dimensi-

onal autonomous system introduced in Sect. 5.2 and governed by differential equa-
tions (5.5). The Poincaré map is four-dimensional, like in the previous system, but
because of its autonomous nature the construction of the Poincaré map requires a lit-
tle bit more care. The computation of the Poincaré map is performed by a computer
program basing on numerical solution of differential equations (5.5) complemented
with procedure of (Hénon, 1982). The initial conditions are specified by the vari-
ables {X,Y,U,V } at z = 0. As the part of the trajectory starting and finishing at
the Poincaré section has been computed, the updated values are obtained. Then, the
variation equations are integrated along the same trajectory; there is a collection of
four vectors to construct columns of the derivative matrix for the Poincaré map. Af-
ter returning to the Poincaré section, from each perturbation vector the component
associated with a shift along the trajectory is subtracted, to ensure zero value of
variation of the variable z; the resulting vectors determine columns of the derivative
matrix.
Figure. 7.21 illustrates verification of the cone criterion for a set of representative
points on the attractor obtained by successive iteration of the Poincaré map. The
panels have the same sense as for the previous √ system. Panel (c) shows that the
required cone inclusions take place at γ = 3 > 1.
References
Afraimovich, V., Hsu, S.-B.: Lectures on Chaotic Dynamical Systems. International Press,
Somerville, MA (2003).
Anishchenko, V.S., Kopeikin, A.S., Kurths, J., Vadivasova, T.E., Strelkova G.I.: Studying hyper-
bolicity in chaotic systems. Physics Letters A 270, 301–307 (2000).
Anosov, D.V.: On the contribution of N.N. Bogolyubov to the theory of dynamical systems. Russ.
Math. Surv. 49, 1–18 (1994).
Bowen R.: Equilibrium states and ergodic theory of Anosov diffeomorphisms. Lecture Notes in
Math. 470, Springer, Heidelberg (1975).
Ginelli, F., Poggi, P., Turchi, A., Chaté, H., Livi, R., Politi, A.: Characterizing dynamics with
covariant Lyapunov vectors. Phys. Rev. Lett. 99, 130601 (2007).
Hasselblatt, B., Pesin, Y.: Hyperbolic dynamics. Scholarpedia 3, 2208 (2008).
Hénon, M.: On the numerical computation of Poincaré maps. Physica D 5, 412–414 (1982).
Hirata, Y., Nozaki, K., Konishi, T.: The intersection angles between N-dimensional stable and
unstable manifolds in 2N-dimensional symplectic mappings. Prog. Theor. Phys. 102, 701–706
(1999).
Krauskopf, B., Osinga, H.M., Doedel E.J., Henderson, M.E., Guckenheimer, J., Vladimirsky, A.,
Dellnitz, M., Junge, O.: A survey of methods for computing (un) stable manifolds of vector
fields. Int. J. Bifurcation and Chaos 15, 763–791(2005).
Kuptsov, P.V., Kuznetsov, S.P.: Violation of hyperbolicity in a diffusive medium with local hyper-
bolic attractor. Phys. Rev. E 80, 016205 (2009).
Kuznetsov, A.S., Kuznetsov, S.P., Sataev, I.R.: Parametric generator of hyperbolic chaos based on
two coupled oscillators with nonlinear dissipation. Technical Physics 55, 1707–1715 (2010).
Kuznetsov, S.P.: A non-autonomous flow system with Plykin type attractor. CNSNS 14, 3487–3491
(2009).
Kuznetsov, S.P. and Pikovsky, A.: Autonomous coupled oscillators with hyperbolic strange attrac-
tors. Physica D 232, 87–102 (2007).
Kuznetsov, S.P., Sataev, I.R.: Hyperbolic attractor in a system of coupled non-autonomous van
97–104 (2007).
Kuznetsov, S.P., Sataev, I.R.: Visualization of invariant measures and comparison of hyperbolic
and non-hyperbolic chaotic attractors. Preprint. SB IRE RAS, Saratov (2011).
Kuznetsov, S.P., Seleznev, E.P.: A strange attractor of the Smale-Williams type in the chaotic
dynamics of a physical system. JETP 102, 355–364 (2006).
Kuznetsov, S.P., Pikovsky, A.S., Sataev, I.R.: Hyperbolic Smale-Williams attractor in Poincaré map
of a four-dimensional autonomous system. In: Proc. of the III Int. Conf. “Frontiers of Nonlinear
Physics”. Nizhny Novgorod-Saratov-Nizhny Novgorod, pp. 66–67 (2007)
Lai, Y.-C., Grebogi, C., Yorke, J. A., Kan, I.: How often are chaotic saddles nonhyperbolic? Non-
linearity 6, 779–798 (1993).
Ruelle, D.: A measure associated with Axiom A attractors. Amer. J. Math. 98, 619–654 (1976).
Shilnikov, L.: Mathematical problems of nonlinear dynamics: a tutorial. Int. J. Bifurcation and
Chaos 7, 1353–2001 (1997).
Sinai, Y.G.: Gibbs measures in ergodic theory. Russ. Math. Surv. 27, 21–70 (1972).
Sinai, Y.G.: Stochasticity of dynamical systems. In: Gaponov-Grekhov, A.V. (ed.) Nolinear waves,
pp. 192–212. Moscow, Nauka (1979).
Sinai, J.G., Vul, E.B.: Hyperbolicity conditions for the Lorenz model. Physica D 2, 3–7 (1981).
Tél, T., Gruiz, M.: Chaotic dynamics: an Introduction Based on Classical Mechanics. Cambridge
Williams, R.F.: Expanding attractors. Publications mathématiques de l’I.H.É.S. 43, 169–203
(1974).
Part III
Higher-Dimensional
Systems and Phenomena
Chapter 8
Systems of Four Alternately Excited
Non-autonomous Oscillators
Abstract This chapter is devoted to non-autonomous models generating successive

trains of oscillations with phase chaotically varying from one train to the other. In
contrast to Chap. 4, here we examine dynamics in the phase space of larger dimen-
sions; so, the models are composed of four oscillators activating by turns (usually
in pairs). Particularly, we consider a model, in which evolution of the phases in suc-
cessive epochs of activity is described by the Arnold cat map and by a hyperchaotic
torus map that has two positive Lyapunov exponents. Then, a system will be dis-
cussed that can be thought of as being composed of two coupled pairs of oscillators,
each of which corresponds to a system with hyperbolic attractor of Smale-Williams
type in the Poincaré map. This may be regarded as initial step of a research program
for ensembles of coupled elements with hyperbolic chaotic dynamics, like coupled
map lattices considered, e.g. in (Bunimovich and Sinai, 1988; 1993; Bricmont and
Kupiainen, 1997; Järvenpää, 2005), but on a base of concrete physically realizable
systems.
8.1 Arnold’s cat map dynamics in a system of coupled

non-autonomous van der Pol oscillators
As mentioned in Chap. 1, one special class of objects with chaotic dynamics cor-
responds to Anosov hyperbolic automorphisms of torus (Arnold and Avez, 1968;
Anosov et al., 1995; Devaney, 2003). The most popular example is known also as
Arnold’s cat map:
pn+1 = pn + qn ( mod 1),
(8.1)
qn+1 = pn + 2qn ( mod 1).
For graphical representation one can use a unit square with the opposite pairs of
sides identified (Fig. 1.6(a)). The map (8.1) is a conservative system: any domain in
the (p, q) plane (say, that cat face) conserves its area under the iteration.
174 8 Systems of Four Alternately Excited Non-autonomous Oscillators
It is known that the map (8.1) demonstrates chaotic dynamics in a sense of hy-
perbolic theory, with such attributes as existence of continuous invariant measure,
description in terms of Markov partition and symbolic dynamics, positive topolog-
ical and metric entropies (Devaney, 2003). Two Lyapunov exponents are expressed
via eigenvalues of the matrix associated with the map, namely,
√ √
Λ1 = ln(3 + 5)/2 ≈ 0.9624, Λ 2 = − ln(3 + 5)/2 ≈ −0.9624. (8.2)
The largest exponent is positive that reflects occurrence of chaos characterized by

exponential sensitivity to initial conditions.
The Arnold cat map may be represented as a two-fold composition of a simpler
map:
pn+1 = qn ( mod 1),
(8.3)
qn+1 = pn + qn ( mod 1).
Expressed in terms of one variable, as p n+1 = pn + pn−1 , it is called the Fibonacci
map.
To implement the corresponding dynamics, we consider the following set of cou-
pled non-autonomous van der Pol oscillators (Isaeva et al., 2006):
ẍ − (A cos(2π t/T) − x2 )ẋ + ω02 x = ε z cos ω0t,

ÿ − (A cos(2π t/T ) − y2 )ẏ + ω02 y = ε w,
(8.4)
z̈ − (−A cos(2π t/T ) − z2 )ż + 4ω02 z = ε xy,
ẅ − (−A cos(2π t/T ) − w2 )ẇ + ω02 w = ε x.
Three of them associated with variables x, y, and w have equal natural frequencies
ω0 , and the rest one (z) has a frequency 2ω 0 . The control parameter responsible
for the Andronov-Hopf bifurcation in the partial systems is forced to vary slowly
with some period T . In each half-period, at the beginning of activity stage for one
pair of the oscillators (x, y), or (z, w), the other pair stimulates its excitation because
of presence of respective coupling terms in the equations proportional to ε , one
of which contains also an auxiliary signal of frequency ω 0 . A period of the slow
parameter modulation is assumed to contain an integer number of periods of the
auxiliary signal N0 = ω0 T /2π , so the external driving is periodic.
Let us derive approximate relations for transformation of phases of the oscillators
in the course of the process. Suppose that the first and the second oscillators at a
stage of activity are characterized by some initial phases:
x ∼ cos(ω0t + ϕx ), y ∼ cos(ω0t + ϕy ). (8.5)
The coupling term in the third equation contains a product xy and, hence, may be
represented as
1 1
cos(ω0 t + ϕx ) cos(ω0t + ϕy ) = cos(ϕx − ϕy ) + cos(2ω0t + ϕx + ϕy ). (8.6)
2 2
8.1 Arnold’s cat map dynamics in a system of . . . 175
The last additive term in this expression is a resonance drive for the oscillator z and
stimulates its excitation at the beginning of the activity stage. So, the third oscillator
inherits the phase ϕz ≈ ϕx + ϕy + const, while the fourth one simply accepts the
phase from the first oscillator: ϕ w ≈ ϕx + const. At the next stage of activity for the
first pair of the oscillators, the coupling term in the first equation
1
z cos 2π t ∼ (cos(6π t + ϕx + ϕy ) + cos(2π t + ϕx + ϕy )) (8.7)
2
stimulates the oscillator x, and it acquires the phase ϕ x ≈ ϕz + const ≈ ϕx + ϕy +
const. The oscillator y inherits the phase from w: ϕ y ≈ ϕw ≈ ϕx + const. So, the
mapping for the phases is expressed as follows:
ϕx = ϕx + ϕy + const,
(8.8)
ϕy = ϕx + const.
Accounting that the phases are defined in the interval from 0 to 2π and using the
normalization q = ϕ x /2π and p = ϕy /2π , we obtain exactly the map (8.3), up to
additive constants; they obviously may be removed by appropriate shift of the origin.
Hence, for the phases of oscillators (x, y) the stroboscopic dynamics with sampling
with time interval 2T follows the Arnold cat map, at least in rough approximation.
To support the drawn conclusion, let us turn to direct numerical computations
for the set of non-autonomous differential equations (8.4). Figure 8.1 shows the
dynamical variables versus time in the system (8.4), at parameter values
ω0 = 2π , N0 = T = 20, A = 2, ε = 0.4. (8.9)
Observe that the system operates in correspondence with the discussed mechanism.
Chaos reveals itself in random phases of the carrier oscillations in respect to the
envelope in subsequent periods of the slow modulation.
Fig. 8.1 Time dependence for the dynamical variables of the system (8.4) at ω0 = 2π , T = 20, A =
2, ε = 0.4.
To demonstrate correspondence of the transformation of phases in the system of

oscillators with the Arnold cat map, we implement the following procedure. In the
course of numerical integration of Eqs. (8.4) we determine the phases for the first
and the second oscillators in the middle of their activity stage 1
ϕx = arg(x + iω0−1 ẋ), ϕy = arg(y + iω0−1ẏ). (8.10)
If the point (q, p) = (ϕ x /2π , ϕy /2π ) belongs to the area of the cat face drawn in the
unit square, we mark the dot on the diagram, and also mark the dot for the images
after time intervals 2T and 4T on two succeeding plots. If it is not the case, the dot
is not marked. By accumulation of a great number of dots, we can see the cat face on
the first picture, and images corresponding to once and twice iterating of the Arnold
map on the second and the third diagrams, respectively (Fig. 8.2(a)).
The pictures may be compared with those obtained directly from iteration of the
Arnold cat map. For better correspondence, we must account the additive constant
terms in the equations for phases, which appear in the course of the transfer of exci-
tations from oscillators to their partners. With these constants empirically selected
(for the particular case under consideration, see (8.9)) the map reads
p = p + q + 0.07 ( mod 1),

(8.11)
q = p + 2q − 0.38 ( mod 1).
The plots illustrating evolution of the cat face in accordance with this map are shown
in Fig. 8.2(b). Observe a nice correspondence of the phase dynamics in the system of
non-autonomous van der Pol oscillators and in the modified Arnold cat map (8.11).
For more accurate discrete-time description, we turn to the stroboscopic Poincaré
map. Let us have a certain state of the system at t n = 2nT associated with a vector
xn = {x, ẋ/ω0 , y, ẏ/ω0 , z, ż/2ω0 , w, ẇ/ω0 }. From solution of the differential equa-
tions (8.4) with the initial state x n , after time interval 2T , we get a new vector x n+1
and introduce a function that maps the eight-dimensional space into itself:
xn+1 = T(xn ). (8.12)
Geometrically, in the nine-dimensional extended phase space of our non-autono-

mous system {x,t} we have a cross-section of the flow by the eight-dimensional
hyper-planes t = t n = 2nT. The Poincaré map T is a diffeomorphism in R 8 , a one-
to-one differentiable map of class C ∞ (Arnold, 1978).
Figure 8.3 shows portrait of the attractor in projection from the extended phase
space onto the plane of variables of the first oscillator (x, ẋ). The attractor is shown in
gray scales reflecting relative duration of residence of the orbit inside pixels. Black
dots relate to the Poincaré cross-section, i.e., to the instants t n = 2nT .
1 The phases ϕx and ϕy cannot be defined globally, in the whole time interval T : they are attributed
to a stage of excitation of the pair of oscillators and may be used in the context of the discrete time
description. Indeed, when this pair of oscillators is not active, the respective amplitudes are small,
and the phases are not well defined.
Fig. 8.2 Transformation of an area in the form of a cat face in a plane of phase variables in the
system of van der Pol oscillators (8.4) after time intervals 2T and 4T (a) and analogous pictures
obtained from the map (8.10) (b).
Fig. 8.3 Portrait of attractor at ω0 = 2π , T = 6, A = 6.5, ε = 0.4: projection from the extended
phase space on the plane of variables (x, ẋ) and the Poincaré cross-section corresponding to the
instants tn = 2nT (black dots), which looks like projection of a narrow two-dimensional torus.
The spectrum of Lyapunov exponents is evaluated with a help of algorithm based

on simultaneous numerical integration of Eqs. (8.1) together with eight replicas of
the linearized equations for perturbations:
x̃¨ + 2xẋx̃ − (A cos(2π t/T ) − x 2)x̃˙ + ω02 x̃ = ε z̃ cos ω0t,

ỹ¨ + 2yẏỹ − (A cos(2π t/T) − y 2 )ỹ˙ + ω02ỹ = ε w̃,
(8.13)
z̃¨ + 2zżz̃ − (−A cos(2π t/T ) − z2 )z̃˙ + 4ω02z̃ = ε (x̃y + xỹ),
w̃¨ + 2wẇw̃ − (−A cos(2π t/T ) − w2 )w̃˙ + ω02 w̃ = ε x̃.
In the course of the solution, at each next period 2T , the Gram-Schmidt orthogonal-
ization and normalization were performed for eight vectors x̃ = {x̃, x̃˙/ω0 , ỹ, ỹ˙/ω0 , z̃,
z̃˙/2ω0 , w̃, w̃˙ /ω0 }, and the mean rates of growth or decrease of the accumulated sums
of logarithms of the norm ratios are estimated (Appendix A). As found, at the pa-
rameters (8.9) the Lyapunov exponents for the attractor of the Poincaré map are 2
Λ1 = 0.962, Λ2 = −0.970, Λ3 = −15.525, Λ4 = −19.074,

(8.14)
Λ5 = −20.053, Λ6 = −21.315, Λ7 = −32.444, Λ8 = −32.898.
Figure 8.4 shows a plot for all eight Lyapunov exponents of the system versus the
amplitude A of slow modulation at fixed other parameters. As seen, the largest two
exponents remain almost constant in a wide range and are close to the values char-
acteristic of the Arnold cat map (8.2). The rest exponents are largely negative. They
correspond to strong compression of the phase volume along six remaining dimen-
sions in the phase space of the Poincaré map.
Fig. 8.4 Lyapunov exponents for the system (8.4) versus the amplitude of parameter modulation
at ω0 = 2π , T = 20, ε = 0.4.
Figure 8.5 illustrates distribution of invariant measure on the attractor as it looks

on the plane of two phase variables used to show the Arnold cat face. Here one
can see subtle fractal-like structure intrinsic to the distribution, which appears due
to some deflections from the linear toral map. Such distributions were discussed in
the context of the problem of the so-called “non-strange chaotic attractor” (Farmer
et al., 1983; Anishchenko et al., 2002). In fact, attractor in the Poincaré map of
our system is not “non-strange”: it certainly has a transverse fractal structure in the
eight-dimensional phase space, but distribution of the invariant measure over the
torus approximately representing the attractor is of the nature that has been sug-
gested for the “non-strange chaos”.
To have quantitative characteristics for the fractal invariant measure on the attrac-
tor, we have estimated the correlation dimensions from the Grassberger-Procaccia
2 We use here 2T as the normalization time interval in definition of the Lyapunov exponents to
have a correspondence with the cat map model.
Fig. 8.5 Distributions of points on the attractor in projection on a plane of phases of the first pair
of oscillators in Poincaré cross-section obtained from computations or the system (8.1).
algorithm (Grassberger and Procaccia, 1983) by processing a two-component time

series composed of data for phases (8.10) form 10 5 iteration of the Poincaré map
obtained in numerical solution of Eqs. (8.4) (Fig. 8.6). At the present parameter val-
ues (8.18) we get the dimension D 2 =1.98 distinct from the value 2, which would
correspond to a uniform distribution of the invariant density over a surface of 2D
torus. The estimate from Kaplan-Yorke formula (Appendix D) using the Lyapunov
exponents (8.19) yields D L = 1 + Λ1/|Λ2 | ≈ 1.99.
Fig. 8.6 Plot of the correlation integral versus the partition size in double logarithmic scale ob-
tained from processing of two-component time series of 105 pairs of phases (8.10) in the Poincaré
cross-section from numerical solution of (8.4). The respective estimate for the correlation dimen-
sion corresponds to a slope in the range of linear dependence.
The considered system of four coupled non-autonomous van der Pol oscillators
delivers a realistic example of dynamics approximately described by hyperbolic
maps on a torus. In fact, we deal with an attractor in the eight-dimensional phase
space of Poincaré map. Due to strong phase volume compression, it reduces ap-
proximately to a two-dimensional map for the phases of a pair of the oscillators that
corresponds to the Arnold cat map. It may be hypothesized that the attractor in the
Poincaré map is uniformly hyperbolic and, hence, structurally stable, although an
accurate check of this assumption seems not easy because of high dimension of the
phase space.
8.2 Dynamics corresponding to hyperchaotic maps
The term hyperchaos was introduced by Rössler and relates to dynamical regimes
characterized by two or more positive Lyapunov exponents (Rössler, 1979; Mat-
sumoto et al., 1986; Stoop et al., 1989; Reiterer et al., 1998; Letellier and Rössler,
2007). To date, no compelling reasons have been seen for relating the known in
the literature examples of hyperchaotic behavior to uniformly hyperbolic attractors.
Here we turn to models, which can pretend to deliver appropriate examples.
8.2.1 System implementing toral hyperchaotic map
First, let us consider a system of four coupled van der Pol oscillators
ẍ − (A cos(2π t/T ) − x2 )ẋ + ω02 x = ε z cos ω0t,

ÿ − (A cos(2π t/T ) − y2)ẏ + ω02y = ε w cos 2ω0t,
(8.15)
z̈ − (−A cos(2π t/T ) − z2 )ż + 4ω02z = ε xy,
ẅ − (−A cos(2π t/T ) − w2 )ẇ + 9ω02w = ε x3 ,
two of which have the natural frequencies ω 0 , and the third and the fourth have,
respectively, the frequencies 2ω 0 and 3ω0 . Control parameters, responsible for the
Andronov-Hopf bifurcation in the autonomous subsystems, are slowly modulated
periodically in time. In the first half-period of modulation the first pair of oscilla-
tors is active, and the second is damped. In the second half-period the situation is
reversed.
Suppose that the first oscillator in the epoch of activity is characterized by an ini-
tial phase ϕx , and the second by the phase ϕ y ; so, the respective dynamical variables
oscillate as x ∼ cos(ω0t + ϕx ) and y ∼ cos(ω0t + ϕy ). Then, when the second pair
is activating, the third oscillator is stimulated by the coupling term in the equation
proportional to
1 1
xy ∼ cos(ω0t + ϕx ) cos(ω0t + ϕy ) = cos(ϕx − ϕy ) + cos(2ω0t + ϕx + ϕy ).
2 2
Then, it accepts the phase ϕ z = ϕx + ϕy + const because the last term in the sum is
resonant for this oscillator. The fourth oscillator, in turn, accepts the tripled phase
from the first one, as can be deduced from the relation
3 1
x3 ∼ cos3 (ω0t + ϕx ) = cos(ω0t + ϕx ) + cos(3ω0t + 3ϕx ).
4 4
8.2 Dynamics corresponding to hyperchaotic maps 181
The resonance stimulation occurs due to the last term of the tripled frequency. So,
we have ϕw = 3ϕx + const. Then, at the next transfer of the excitation, these phases
are transmitted to the first and the second oscillators at their next activity stage. So,
the mapping for the phases in the complete period of modulation is
ϕx = ϕx + ϕy + const, ϕy = 3ϕx + const, (8.16)
where the constants may be removed by a shift of the origin on the plane (ϕ x , ϕy ).
Accounting that the phases are defined modulo 2π , for normalized variables q =
ϕx /2π i and p = ϕy /2π , we get
pn+1 = pn + qn ( mod 1).

(8.17)
qn+1 = 3pn ( mod 1),
or
p 1 1 p
= . (8.18)
q n+1 3 0 q n
√
Eigenvalues of the matrix in the last expressions are (1 ± 13)/2. Respectively, the
Lyapunov exponents for the map are
√ √
Λ1 = ln( 13 + 1)/2 = 0.834, Λ 2 = ln( 13 − 1)/2 = 0.264. (8.19)
Both of them are positive, so, the map generates hyperchaos.

Figure 8.7 depicts chaotic attractor in the projection from the eight-dimensional
phase space in the stroboscopic section on the plane of phase variables (ϕ x , ϕy ). Note
that the points fill the unit square almost uniformly. Figure 8.8 shows the Lyapunov
exponents of the attractor versus magnitude of the parameter modulation. Observe
that two larger Lyapunov exponents remain approximately constant and close to the
approximated values (8.19) in a wide range. This indirectly confirms the robustness
of the observed hyperchaotic attractor.
Fig. 8.7 Portrait of the trajectory on the hyperchaotic attractor in stroboscopic map of the model
(8.15) onto the plane of phase variables (ϕx , ϕy ) at ω0 = 2π , T = 20, A = 2.5, ε = 0.5.
Fig. 8.8 Lyapunov exponents of the stroboscopic Poincaré map of the model (8.15) at ω0 =
2π , T = 20, A = 2.5, ε = 0.5.
8.2.2 Model with cascade transfer of excitation upward the

frequency spectrum
The following example is inspired by the concept that underlies understanding of

developed hydrodynamic turbulence. As known, it is based on the idea of a cascade
energy transfer through the spectrum, from formations in the turbulent flow charac-
terized by large spatial and temporal scales to smaller scales (Landau and Lifshitz,
1959; Monin and Yaglom, 2007).
Obviously this idea in nature is general and, apparently, it makes sense to use it
as a guiding principle for design of various systems with chaotic dynamics even in
situations where equations are not similar to the hydrodynamic ones.
Consider a system of four van der Pol oscillators with natural frequencies
ω0 , 2ω0 , 4ω0 , 8ω0 shown schematically in Fig. 8.9, operation of which is provided
by a cascade transfer of excitation from one oscillator to another with succes-
sive frequency doubling (Kuznetsov and Sokha, 2010). Parameters controlling the
Andronov-Hopf bifurcation in each oscillator are forced to vary slowly in time, with
amplitude A and period T = 2π /Ω, where Ω ω 0 . The modulation takes place in
the same phase for the first and the third oscillators, and in the counter-phase for the
second and the fourth oscillators. An appropriate model is the system of differential
Fig. 8.9 Block-diagram of the system with cascade transfer of excitation with frequency doubling.
equations
ẍ − (A cosΩt − x2 )ẋ + ω02x = ε w cos 7ω0t,
ÿ − (−A cosΩt − y2 )ẏ + 4ω02y = ε x2 ,
(8.20)
z̈ − (A cosΩt − z2 )ż + 16ω02z = ε y2 ,
ẅ − (−A cosΩt − w2 )ẇ + 64ω02w = ε z2 ,
where the variables x, y, z, w are the generalized coordinates of the oscillators from
the first to the fourth, respectively. We assume that the relation ω 0 T = 2π N holds,
where N is an integer, so that the non-autonomous set of equations has periodic time
dependence of the coefficients.
Due to the assumed parameter modulation, two pairs of oscillators become active
in turn. Stimulations for the oscillators 2, 3 and 4 at the beginning of each new stage
of activity are provided by the excitation transfer from the previous oscillators of
twice less frequencies through the quadratic nonlinear element. In the course of
the excitation transfer, the phase shift constant intrinsic to the signal is doubled.
1
Indeed, from the signal cos(ω t + ϕ ) we obtain cos 2 (ω t + ϕ ) = cos(2ω t + 2ϕ ) +
2
1
; here one can account only the component of doubled frequency resonant with the
2
newly excited oscillator. The feedback is closed with final nonlinear transformation
of mixing of the output signal from the chain with a reference signal of frequency
7ω0 and with fixed phase shift to obtain the difference frequency response in the
1
resonance band of the first oscillator. Indeed, cos(8ω 0t + θ ) cos7ω0t = cos(ω0t +
2
θ ) + · · · here the non-resonance term is omitted, which is negligible for stimulation
of the oscillator of frequency ω 0 .
For a clearer understanding of the phase transformations in the course of opera-
tion of the system, refer to Table 8.1. The columns correspond to oscillators 1, 2, 3,
4, and the rows relate to successive half-periods of slow modulation. The first row
corresponds to the stage of activity of the oscillators 1 and 3 characterized by phases
ϕ0 and θ0 . The oscillators 2 and 4 at this stage are damped and do not possess a well
defined phase, so that the corresponding cells are left empty. The oscillators 2 and
4 are excited on the next half-period of modulation. Following the above explana-
tions, they accept the phases, respectively, 2ϕ 0 and 2θ0 (up to additive constants).
Then, the process is repeated. In the table, for clarity, a few lines corresponding to
successive steps of the conversion are given.
Table 8.1 Transformation of phases for oscillators 1–4 at successive stages of activity.
1 2 3 4
0 ϕ0 θ0
1
2ϕ0 + const 2θ0 + const
2
1 2θ0 + const 4ϕ0 + const
3
4θ0 + const 8ϕ0 + const
2
2 8ϕ0 + const 8θ0 + const
As seen from the table, after each next period of slow modulation of parame-
ters the phases of two initially active oscillators are transformed according to the
relations
ϕn+1 = 2θn + const ( mod 2π ),
(8.21)
θn+1 = 4ϕn + const ( mod 2π ).
In other words, a two-dimensional vector v n = (ϕn , θn ) whose components are
defined modulo 2π , is mapped to another vector v n+1 = (ϕn+1 , θn+1 ) of the same
type, i.e. this may be interpreted as a map defined on a torus. It has specific degener-
ation, namely, the complete phase sequence of ϕ n (or θn ) is composed of two inde-
pendent subsequences, so that in the course of the time evolution their members al-
ternate. Subsequences manifest chaotic dynamics. In two steps, small perturbations
to each of them increase by eight times. Therefore, the map (8.21) is characterized
1
by presence of two equal positive Lyapunov exponents Λ 1,2 = ln8 ≈ 1.0397.
2
The results of the numerical solution of differential equations (8.20) show that
the expected type of dynamics indeed occurs in a wide range of parameters. For a
detailed analysis we select the case ω 0 = 2π , T = 8, A = 4.5, ε = 0.6. Figure 8.10
shows the dynamic variables for the four oscillators plotted as functions of time.
Fig. 8.10 Dynamical variables for four oscillators constituting the system (8.20) plotted as func-
tions of time at ω0 = 2π , T = 8, A = 4.5, ε = 0.6.
Each oscillator generates a sequence of pulses following one another with time in-
tervals T . The carrier frequencies correspond to the natural frequencies of the partial
oscillators. Chaos reveals itself in chaotic variation of the phases of the carrier in re-
spect to envelops of the signals. The presence of chaos is caused by the described
mechanism of phase transfer between the oscillators from the previous stages of the
process to the next ones.
Figure 8.11 shows the iteration diagrams for the phases, which indicate that their
dynamics actually correspond in certain approximation to the map (8.21). Phases,
each corresponding to the next stage of the activity of the oscillators 1 and 3, are
determined at a time instant fixed with respect to the modulation signal according
to the formulas
ϕn = arg(x − iẋ/ω0 ), θn = arg(z − iż/ω0 ). (8.22)
Fig. 8.11 Iteration diagrams for phases of the oscillators 1 and 3, respectively, ϕn and θn at suc-
cessive stages of their activity at ω0 = 2π , T = 8, A = 4.5, ε = 0.6.
Fig. 8.12 Attractor projection onto the phase plane of the oscillator 1 shown in gray scale. White
dots correspond to the stroboscopic Poincaré section.
Figure 8.12 shows portrait of the attractor in the projection from the extended
nine-dimensional phase space to the phase plane of the first oscillator. The picture is
presented in gray scale, so that lighter pixels correspond to a relatively great prob-
ability for the residence of the representative orbit there. White dots correspond to
instants of reaching the maximum values of the modulated bifurcation parameter in
the first oscillator; they form a portrait of the attractor in the Poincaré section.
To evaluate the Lyapunov exponents, in a computer program we calculate the
Poincaré map from numerical solution of Eqs. (8.20) and the Jacobian matrix by
simultaneous solution of eight replicas of the linearized variation equations. With
multiple iteration of this procedure, the evolution of the eight vectors of small per-
turbations is monitored by multiplying them by the Jacobian matrices. After each
step of the iterations the Gram-Schmidt orthogonalization process is applied and
normalization is performed for the eight vectors (Appendix A). Lyapunov expo-
nents are defined as the average rates of growth or decrease the accumulated sums
of logarithms of the norms of the vectors.
Figure 8.13 shows the dependence of the Lyapunov exponents for the strobo-
scopic Poincaré map on the parameter modulation magnitude A at fixed other pa-
rameters. In a wide range of parameter variation the largest two exponents remain
almost constant and close to (ln 8)/2. Particularly, at A = 4.5 the Lyapunov expo-
nents are
Λ1 = 1.038, Λ2 = 1.037, Λ3 = −4.622, Λ4 = −5.077,
(8.23)
Λ5 = −7.914, Λ6 = −8.403, Λ7 = −10.335, Λ8 = −10.722,
and the attractor dimension in the Poincaré section from the Kaplan-Yorke formula
is 2.449.
Fig. 8.13 Lyapunov exponents versus the parameter of the modulation depth A at ω0 = 2π , T =
8, ε = 0.6.
Figure 8.14 shows the Fourier power spectrum of the signal represented as the
sum of the variables related to all four oscillators X = x + y + z + w. Spectrum is
continuous that reflects the chaotic nature of the dynamics on the attractor. In the
8.3 Hyperchaos and synchronous chaos in a system of coupled non-autonomous oscillators 187
spectrum one can see four well-defined widened peaks, which correspond to the
natural frequencies of four oscillators constituting the system. Discrete component,
which is distinguishable on the left side of the spectrum corresponds to the fre-
quency of slow periodic modulation of parameters.
Fig. 8.14 The power spectrum of the signal X = x + y + z + w generated by the system at ω0 =
2π , T = 8, A = 4.5, ε = 0.6.
Summarizing, we outline a fascinating moment that the considered scheme ex-

ploits one of the key ideas of the theory of hydrodynamic turbulence——a cascade
transfer of energy over the spectrum. The analogy, of course, is rather qualitative
and heuristic in nature: in this case we talk about the transfer of excitation along a
chain of oscillators that become active alternately with successive doublings of the
frequency. A feature of the system is a quite broad total frequency band of the gener-
ated signal. This is due to the presence of a set of oscillators with essentially differ-
ent natural frequencies. Number of oscillators in the chain can be increased that will
further expand the frequency band of the generated chaotic signal. In a wide range
of parameters the dynamics is characterized by the presence of two positive Lya-
punov exponents, which are close in value and weakly depend on the parameters.
The scheme obviously can be realized as an electronic device.
8.3 Hyperchaos and synchronous chaos in a system of coupled

non-autonomous oscillators
Important research direction in nonlinear science concerns lattice models composed

of blocks, or cells, manifesting certain types of dynamical behavior. Particularly,
these studies are relevant to some fundamental challenging problems like founda-
tion of statistical mechanics or origin of turbulence. In this context, one can mention
seminal work of Fermi, Pasta and Ulam (Ford, 1992), cellular automata considered
by Wolfram and others (Wolfram, 1986), coupled map lattices suggested and studied
by Kaneko and others (Kaneko, 1993). As directly relating to the framework of the
present book, one has to outline the work of Bunimovich and Sinai and others (Buni-
movich and Sinai, 1988; 1993; Bricmont and Kupiainen, 1997; Järvenpää, 2005)
devoted to investigation of lattices composed of elements demonstrating hyperbolic
chaos. Particularly, it may be lattices of coupled one-dimensional expanding maps,

or lattices composed of systems with uniformly hyperbolic attractors. It was shown
that hyperbolic nature of chaos and strong chaotic properties, like existence of the
absolutely continuous SRB measure, positive topological and metric entropy, persist
for the coupled systems while the coupling strength is not too great. These results
become essential now, when we can handle feasible systems with uniformly hyper-
bolic chaotic attractors. To start, it is natural to turn to a situation of minimal number
of coupled subsystems manifesting the hyperbolic chaos (Kuptsov and Kuznetsov,
2006). Here we consider a model composed of two pairs of non-autonomous al-
ternately excited van der Pol oscillators discussed in Chap. 4, the whole system
consists of four oscillators, and just fits the common topic of the present chapter.
8.3.1 Equations and basic modes of operation
Let us examine a system of two coupled elements, each of which is a pair of al-
ternately excited van der Pol oscillators; so, the model (4.9) just corresponds to a
single block. In the following set of equations
ẍ1 − (A cos(2π t/T ) − x21 )ẋ1 + ω02 x1 = ε y1 cos ω0t + d(ẋ2 − ẋ1 ),
ÿ1 − (−A cos(2π t/T ) − y21)ẏ1 + 4ω02 y1 = ε x21 + d(ẏ2 − ẏ1 ),
(8.24)
ẍ2 − (A cos(2π t/T ) − x22 )ẋ2 + ω02 x2 = ε y2 cos ω0t + d(ẋ1 − ẋ2 ),
ÿ2 − (−A cos(2π t/T ) − y22)ẏ2 + 4ω02 y2 = ε x22 + d(ẏ2 − ẏ1 ),
the variables relating to two subsystems are indicated with subscripts 1 and 2, re-
spectively, and d is parameter of coupling of the subsystems. We assume, as usual,
that the period of parameter modulation T is an integer multiple of the period of the
auxiliary signal 2π /ω 0.
Coupling in Eqs. (8.3) is defined in such a way that the terms responsible for the
interaction of the subsystems vanish at the coincidence of their states. The corre-
sponding condition determines an invariant manifold of symmetric motions in the
phase space of the joined system. This manifold contains an invariant set, which is
identical in structure to the attractor of the partial system (4.9), i.e. it is the suspen-
sion of the Smale-Williams solenoid.
The coupling terms in Eqs. (8.3) are proportional to differences of time deriva-
tives (the generalized velocities). Hence, the coupling is dissipative in the sense that
it provides tendency to convergence and equalization of instantaneous states of the
subsystems. As individual subsystems are chaotic, the dynamics is determined by
competition of two factors: the divergence of states of the subsystems because of in-
trinsic sensitivity to initial conditions, and the tendency to convergence of the states
due to the coupling. At zero, as well as at small values of the coupling parame-
ter d, the first factor dominates; chaotic motions in the subsystems are independent
or weakly correlated. At large values of the coupling parameter, the second factor
prevails, and the complete chaotic synchronization occurs: both subsystems evolve
chaotically in time, but in perfect synchrony where the instant states at any time are
identical. It is called the complete chaos synchronization (Pikovsky et al., 2002).
Figure 8.15 illustrates dynamics of the model (8.24) at ω 0 = 2π , T = 10, ε =
0.5, A = 3 for two values of the coupling parameter. The smaller value corresponds
to chaotic, roughly independent, motions in both subsystems (panel (a)), and the
larger one to emergency of chaos synchronization in the course of the time evolution
(panel (b)). The relevant features of the processes reveal themselves in large time
scales; so, we present fragments of the waveforms separated by relatively long time
intervals (marked by three dots). Observe that in the panel (a) the phase differences
between the simultaneously active partner oscillators of two subsystems fluctuate
from one to other fragment of the waveform. In the second case, the phase difference
vanishes in the course of time evolution, and the motions of the subsystems become
synchronous.
Fig. 8.15 Fragments of typical time dependence for the dynamical variables of the model (8.24)
illustrating nature of behavior in asynchronous (a) and synchronous (b) chaotic regimes at ω0 =
2π , T = 10, A = 3, ε = 0.5.
Figure 8.16 shows Lyapunov exponents evaluated for the stroboscopic eight-
dimensional Poincaré map of the system and plotted against the coupling parameter
at ω0 = 2π , T = 10, ε = 0.5, A = 3. The data were obtained in computations based
on the Benettin algorithm from numerical solution of Eqs. (8.24) supplemented
with eight sets of linearized equations for perturbation vectors, which are subjected
orthogonalization and normalization at each period of the parameter modulation
(Appendix A). Observe that the largest Lyapunov exponent in the whole parameter
range remains almost constant close to the value ln2 = 0.693 · · ·. The second Lya-
punov exponent is of the same value at zero coupling, but gradually decreases as
d grows. Its passage through zero is associated with transition from asynchronous
to synchronous behavior approximately at d c =0.069. Other Lyapunov exponents are
negative everywhere.
Chaotic attractors in the composite system (8.24) surely relate to the uniformly
hyperbolic class in two limit cases.
Certainly, it is so at zero coupling. Here the system decomposes to two indepen-
dent subsystems, each possessing the hyperbolic attractors of Smale-Williams type
Fig. 8.16 Lyapunov exponents for the system (8.24) versus parameter of strength of dissipative
coupling ω0 = 2π , T = 10, ε = 0.5, A = 3. The subscripts (, ⊥) designate the exponents relating,
respectively, to symmetric and anti-symmetric (transversal) parts of the Lyapunov spectrum. The
synchronization-desynchronization transition is indicated with an arrow.
in stroboscopic map, and the attractor of the whole system is also uniformly hy-
perbolic. Because of its structural stability, such nature of the attractor persists for
close situations of non-zero coupling while it is small enough. The same conclusion
follows from the studies (Bunimovich and Sinai, 1988; 1993).
At large couplings, in the situation of synchronous chaotic motion, attractor of
the composite system is of the same structure as that of a single block, which is
uniformly hyperbolic. The only difference is presence of additional directions asso-
ciated with strong compression in the phase space of doubled dimension.
On the road from the first to the second situations, under gradual increase of the
coupling parameter, the hyperbolicity is violated somewhere because the transition
from the first to the second types of dynamics implies change of dimensions of the
stable and unstable manifolds; this transition is a separate subject for the analysis.
According to Chap. 4, an illuminating approach to understanding dynamics of a
single subsystem is monitoring transformations of phase shifts on successive stages
of activity of the oscillators. The same hint is appropriate for the coupled systems.
The top row of Fig. 8.17 shows portraits for attractors of the model (8.24) in
the stroboscopic cross-section in projection onto the plane of the phase variables
(ϕ1 , ϕ2 ) evaluated for oscillators of frequency ω 1 relating to the first and the second
subsystems. The diagrams correspond to four values of coupling parameter d indi-
cated on the plots; the remaining parameters are ω 0 = 2π , T = 20, A = 2, ε = 0.5.
Panels (a-c) relate to asynchronous dynamics, and panel (d) relates to synchronous
regime.
Accounting symmetry of the model in respect to exchange of indices 1 and 2, it is
appropriate to introduce the symmetric and anti-symmetric variables defined mod-
ulo 2π instead of two phases ϕ 1 and ϕ2 relating to simultaneously active oscillators
of both subsystems, namely,
Fig. 8.17 Portraits of a trajectory on the attractors and histograms for distribution of the phase
shifts between the subsystems obtained in computations with model (8.24) in chaotic asynchronous
(a-c) and synchronous (d) regimes at ω0 = 2π , T = 20, A = 2, ε = 0.5.
φ = ϕ1 + ϕ2 and ψ = ϕ1 − ϕ2 . (8.25)
The bottom row of Fig. 8.17 shows histograms for distribution of the phase shifts ψ
between the subsystems obtained in computations for model (8.24) at the same pa-
rameters. Observe gradual formation of a distribution with pronounced maximum at
zero under graduate increase of coupling in non-synchronous regimes, and appear-
ance of Dirac delta function distribution after the synchronization transition.
Figure 8.18 shows iteration diagrams for the phase variables φ and ψ . In dia-
grams in the top row the symmetric and anti-symmetric phases are plotted at sub-
critical coupling. In this hyperchaotic regime both variables evolve according to
Bernoulli-like maps; two Lyapunov exponents are positive here. The second row
shows diagrams for anti-symmetric variables in the nearly critical situation (left)
and in the case of coupling strength greater than the critical one. In the last case spe-
cial procedure was used to obtain the plot from processing the transients; attractor
of the synchronous chaotic regime corresponds in these coordinates to a point in the
origin as indicated by arrow.
Slightly below the synchronization threshold, a dynamical behavior occurs called
the on-off intermittency (Platt et al., 1993; Hilborn, 2000; Mosekilde et al., 2002;
Schuster and Just, 2005). The system spends most time getting very close to the
synchronous state, but from time to time appears to be far from that state; these
periods occupy the smaller proportion of time at which we are closer to the point of
the transition. Figure 8.19 illustrates this phenomenon for the model (8.24) with the
plot of the phase difference between the subsystems depending on time.
Besides the on-off intermittency, transition from non-synchronous to synchronous
chaos is accompanied by occurrence of some specific phenomena, like riddling
Fig. 8.18 Iteration diagrams for the model (8.24) at tn = nT for three values of coupling parameter
d and ω0 = 2π , T = 20, A = 2, ε = 0.5. The top row shows diagrams for symmetric and anti-
symmetric phase variables determined according to (8.25) for the case of subcritical coupling.
The second row shows diagrams for anti-symmetric variables in nearly critical situation (left) and
in the case of coupling larger than the critical value (right). In the last diagram attractor of the
synchronous chaotic regime is a point in the origin indicated by the arrow.
Fig. 8.19 On-off intermittency in the model (8.24) at ω0 = 2π , T = 20, A = 2, ε = 0.5 and d =
0.0347. The phase differences between two subsystems are plotted evaluated at time instants nT.
basins and bubbling attractor, examined in the literature in the context of the chaos
synchronization with examples of coupled systems manifesting non-hyperbolic at-
tractors (Anishchenko, 2002; Mosekilde et al., 2002; Ott, 2002; Ashwin, 2005).
Now, they have to be reconsidered to reveal features that may be intrinsic to the
case of uniformly hyperbolic chaos in the subsystems.
8.3.2 Equations for slow complex amplitudes
More convenient model for detailed studies may be obtained with reformulation
of the problem in terms of equations for slow complex amplitudes. Applying this
approach to Eqs. (8.24), like described in Sect. 4.1 and 4.3, one can derive the fol-
lowing set of shortened equations
1 1 1
ȧ1 = (A cos(2π t/T ) − |a1|2 )a1 − iεω0−1 b1 + d(a2 − a1),
2 4 2
1 1 1
ḃ1 = (−A cos(2π t/T ) − |b1|2 )b1 − iεω0−1 a21 + d(b2 − b1 ),
2 4 2
(8.26)
1 1 1
ȧ2 = (A cos(2π t/T ) − |a2|2 )a2 − iεω0−1 b2 + d(a1 − a2),
2 4 2
1 1 −1 2 1
ḃ2 = (−A cos(2π t/T ) − |b2| )b2 − iεω0 a2 + d(b1 − b2 ).
2
2 4 2
Computations confirm that dynamics governed by these equations agrees well with
that for the original model (at least at the parameter values assigned in the previous
subsection). So we can use this version of the model to explain essential features of
the observed behavior.
First, let us discuss Lyapunov exponents of the symmetric invariant set S defined
by the relations a 1 ≡ a2 , b1 ≡ b2 , which represents attractor of the composite system
associated with chaos synchronization at large couplings. In the subcritical range
of the coupling parameter this set is not attractive. This set is the Smale-Williams
solenoid in the stroboscopic Poincaré section.
Setting a1,2 (t) = a(t)+ ã1,2 (t), b1,2 (t) = b(t)+ b̃1,2 (t) we obtain in linear approx-
imation for perturbations of the reference trajectory the following equations:
1 1 1
ã˙1 = (A cos(2π t/T ) − 3|a|2)ã1 − a2ã∗1 − iεω0−1 b̃1 + d(ã2 − ã1),
2 4 2
1 1 −1 1
ḃ1 = (−A cos(2π t/T ) − 3|b|2)b̃1 − b2 b̃∗1 − iεω0 aã1 + d(b̃2 − b̃1),
2 2 2
(8.27)
˙ã2 = 1 (A cos(2π t/T ) − 3|a|2)ã2 − a2ã∗2 − 1 iεω0−1 b̃2 + 1 d(ã1 − ã2),
2 4 2
1 ∗ 1 −1 1
ḃ2 = (−A cos(2π t/T ) − 3|b| )b̃2 − b b̃2 − iεω0 aã2 + d(b̃1 − b̃2).
2 2
2 2 2
It may be rewritten for symmetric variables ã || = ã1 + ã2, b̃|| = b̃1 + b̃2 as
1 1
ã˙|| = (A cos(2π t/T ) − 3|a|2)ã|| − a2ã∗|| − iεω0−1 b̃|| ,
2 4
(8.28)
1 1
b̃˙ || = (−A cos(2π t/T ) − 3|b|2)b̃|| − b2 b̃∗|| − iεω0−1 aã|| ,
2 2
and for anti-symmetric ones ã ⊥ = ã1 − ã2 , b̃⊥ = b̃1 − b̃2 as
1 1
ã˙⊥ = (A cos(2π t/T ) − 3|a|2)ã⊥ − a2ã∗⊥ − iεω0−1 b̃⊥ − d ã⊥,
2 4
(8.29)
1 1
b̃˙ ⊥ = (−A cos(2π t/T ) − 3|b|2)b̃⊥ − b2b̃∗⊥ − iεω0−1 aã⊥ − d b̃⊥.
2 2
From these equations it is obvious that spectrum of Lyapunov exponents decom-
poses onto two parts, the symmetric one obtained from Eqs. (8.28) and anti-
symmetric or transversal exponents from (8.29).
The equations for symmetric perturbations (8.28) coincide exactly with respec-
tive equations for a single subsystem (4.19). Thus, the Lyapunov exponents are sim-
ply equal to those obtained for the single system.
As to the second part of the spectrum, one can notice that substitution ã ⊥ =
ãe−dt , b̃⊥ = b̃e−dt transforms (8.29) exactly to (8.28). It means that the transversal
Lyapunov exponents are connected with those relating to symmetric perturbations
by relations λ⊥i = λ||i − d, or, for the Poincaré map, Λ i⊥ = Λi|| − dT where T is the
period of modulation, and i is index numbering the Lyapunov exponents from 1 to
4. The dependence of Λ i⊥ on coupling parameter is linear (this is clearly visible in
Fig. 8.16).
Transition from synchronous to non-synchronous chaos observed with decrease
of the coupling parameter d occurs at zero value of the largest transversal exponent,
so, the critical coupling is d c = Λ1|| /T . Accounting that Λ 1|| ≈ ln 2, we obtain an
approximate estimate for the critical coupling parameter d c ≈ 0.693/T. It agrees
well with the computations.
Let us derive an approximate map describing dynamics in terms of the phases.
To do this, we note that the effect of coupling appears on time scale of order T
while the phase doubling in individual subsystems is characterized by essentially
less time scale. So, the phase doubling may be accounted neglecting the coupling
and regarded as happening independently in both subsystems.
Due to the coupling, a relatively slow process of evolution of the phase difference
between the subsystems takes place for each pair of simultaneously active oscilla-
tors, which can be described as follows. Accounting identical time dependence of
the bifurcation parameter modulation in coupled pairs of oscillators a 1 and a2 , it is
reasonable to assume that their amplitudes are approximately equal to each other
in the whole period of activity: a 1 ≈ A(t)eiϕ , a2 ≈ A(t)eiθ . Substituting these ex-
pressions into respective equations (8.26), multiplying them by the exponentials
e−iϕ , e−iθ , and subtracting the equations, we take the imaginary part and obtain the
equation for the evolution of the anti-symmetric phase variable ψ = ϕ − θ of the
form
ψ̇ = −d sin ψ . (8.30)
It may be solved analytically; the convenient representation of the solution is

(1 − e−dt ) tan ψ0
ψ = ψ0 + arctan , where ψ0 designates the initial value of ψ . The
1 + e−dt tan2 ψ0
same is true in the second half of the modulation period in respect to the com-
plex amplitudes b 1 and b2 . Summarizing, the transformation of the phase in one
modulation period consists of the slow evolution of the anti-symmetric variable in
accordance with Eq. (4.19) of full duration T , and of a fast event of doubling of both
phases ϕ and θ . So, we have to write

1 (1 − e−dT ) tan(θn − ϕn )
ϕn+1 = 2 ϕn + arctan + c( mod 2π ),
2 1 + e−dT tan2 (θn − ϕn )
(8.31)
1 (1 − e−dT ) tan(ϕn − θn )
θn+1 = 2 θn + arctan + c( mod 2π ),
2 1 + e−dT tan2 (θn − ϕn )
where c is some constant. 3

Figure 8.20 shows the plots for the anti-symmetric variable ψ n = ϕn − θn ac-
cording to the relations (8.31). They have to be compared with empirical plots in
Fig. 8.18. Observe that the forms of the curve evolve with growth of the coupling pa-
rameter in agreement with qualitative explanations and numerical results. At small
coupling, the map for the anti-symmetric phase looks equivalent to Bernoulli map;
here the slope of the curve in the origin is greater than 1. At the critical situation
d = dc the curve touches the angle bisector line, and for d > d c one observes inter-
sections of the curve with that line; in the origin it corresponds to stable fixed point
associated with the regime of complete phase synchronization. Figure 8.21 shows
a plot illustrating the on-off intermittency close to the critical value of the coupling
parameter in the simplified model (8.31). Observe good qualitative correspondence
with Fig. 8.19.
Fig. 8.20 Plots of the analytic approximation for the map for the anti-symmetric phase according
to Eqs. (8.31) for three values of coupling parameter d at T = 20. The left diagram corresponds to
subcritical situation (no synchronization), the right one to supercritical coupling (synchronization
takes place associated with the fixed point on the plot indicated with arrow), and the central diagram
relates to the intermediate critical situation.
3 From this presentation one can easily see how to compose coupled map lattices associated with
systems constituted by large number of the elementary blocks, like those studied in (Bunimovich
and Sinai, 1988; 1993).
Fig. 8.21 On-off intermittency in the map (8.31) at T = 20 and d = 0.03462. The anti-symmetric
phase variable is plotted versus the discrete time n.
One has to recognize that the model (8.31) is not perfect because it does not
embrace some subtle details of the synchronization transition, like riddling basins.
In the case of coupled elements with hyperbolic attractor this phenomenon is not so
pronounced, but it can be revealed in a narrow parameter range near the transition.
To illustrate this, we turn to computations based on Eqs. (8.26).
To proceed we note that the symmetric invariant set S of the system, which is
the Smale-Williams solenoid in the stroboscopic Poincaré section, contains invari-
ant subsets, including, particularly, the unstable periodic orbits of different periods
(cycles), and the degrees of instability of these subsets in respect to the transver-
sal perturbations differ, generally speaking, from that averaged over the whole set
S.4 Close to the synchronization threshold this circumstance determines effects of
riddling basins as well as appearance of the so-called bubbling attractor.
To be concrete, several periodic orbits were determined in computations for the
model (8.26) at T = 4, ε = 0.2π , A = 8. The relatively small modulation period T
is selected here to make the examined effects more pronounced. For these orbits
Floquet multipliers μ were evaluated, and the Lyapunov exponents λ are expressed
as λ = P−1 ln μ , where P = pT is the period of the orbit.
Figure 8.22 plots the largest transversal Lyapunov exponents for the periodic or-
bits with numbers p ranging from 1 to 7. Observe that with decrease of the coupling
parameter, the periodic orbits become transversely unstable at different parameter
values in some range around the critical point where it happens for the symmet-
ric invariant set in the sense of a typical trajectory belonging to it. In the region
slightly above the synchronization threshold, due to existence of transversally un-
stable invariant subsets, the basin of attraction for the symmetric attractor is riddled
(Mosekilde et al., 2002; Ashwin, 2005). To visualize this basin we consider a two-
dimensional section in the phase space, on which dots will be marked belonging
and not belonging to the basin in white and black, respectively. For this, we scan a
square domain on the plane of variables Re a 1 and Re a2 , assigning other variables
some values associated with an arbitrarily chosen point on the symmetric invari-
ant set. In the plot, the diagonal Re a 1 = Re a2 represents the symmetric attractor.
Then, the computations are performed at each pixel for sufficiently large number
4 In the approximate model (8.38) this difference is neglected.

of steps of iteration of the stroboscopic Poincaré map, and presence or absence of

synchronization is diagnosed by monitoring the norm of the state difference for the
subsystem, starting from certain step. 5
Fig. 8.22 The largest transversal Lyapunov exponents for cycles with periods ranging from 1 to 7
versus the coupling parameter at A = 8, T = 2, ε = 0.05. The dotted line marks the synchronization
threshold for the symmetric attractor.
It was found possible to detect confidently the effect of the basin riddling for
symmetric attractor of the system (8.26) only in very small neighborhood of the
synchronization transition. Figure 8.23 shows the results of the procedure of visual-
Fig. 8.23 Riddling of the basin of symmetric attractor of the system (8.26) depicted with low (a)
and high (b) resolution of the recognition procedure. Black color marks the points, from which
the system is detected not to arrive at the symmetric attractor. Parameters assigned are A = 8, T =
2, ε = 0.05, d = 0.173. Im a1 (0) = Im a2 (0) = 0.37, b1 (0) = b2 (0) = −0.05 + 0.07i.
5 Rigorously speaking, this procedure guarantees the recognition of convergence or divergence
from the symmetric attractor with some probability; it is not absolutely reliable. However, for
qualitative illustrations we need here, it is sufficient.
ization of the riddling at d = 0.173. One can see that intrinsic feature of the basin is
absence of any areas of notable size which would not contain points from the basin.
With increase of resolution of the recognition procedure each structural element
of the picture breaks down onto smaller disconnected elements, and so on. Such
structure may be explained by a rather low density of the invariant subsets in the
symmetric attractor responsible for origin of the riddling. In turn, this is connected
with hyperbolic nature of the attractor of a single subsystem.
We now consider the effect of “bubbling” of the attractor, which is also associ-
ated with the presence of transverse instability of the invariant subsets of the set S
(Anishchenko, 2002; Mosekilde et al., 2002; Ott, 2002). Suppose the system is in
the supercritical region near the threshold of synchronization. Choose initial condi-
tions so that the system goes into a symmetric attractor, i.e. the distance between
the subsystems |a1 − a2|2 + |b1 − b2 |2 decreases for a sufficiently large observa-
tion time. Then, we retain the previous initial conditions and parameters, but in the
course of the numerical solution add a little random noise to the variables. Under
its influence, the system can fall in a neighborhood of transverse unstable set, and
it will lead to a burst of desynchronized oscillations. After some time the system
returns to a symmetric attractor and will remain in his neighborhood as long as the
other random shift will cause a new burst.
The “bubbling” is illustrated in Fig. 8.24. In panel (b) one can see that the pres-
ence of noise of small amplitude (of order 0.001) causes bursts of oscillations whose
amplitude exceeds the amplitude of noise by orders of magnitude. However, the sys-
tem turns out to be quite rare in the vicinity of transversely unstable invariant sets.
Fig. 8.24 Illustration of “bubbling attractor”: dependence of the distance between instant states
of the subsystems in model (8.26) versus number of steps of the stroboscopic Poincaré map at
A = 8, T = 2, ε = 0.05. Small noise is added in the course of integration at each step of the finite-
difference scheme of order 0.001. The coupling parameter is d = 0.173 (a) and d = 0.174 (b).
References 199
As seen from panel (b), a little increase of the coupling parameter leads to essential
decrease in probability of appearance of bursts, and the peaks in this figure are lower
than those in the previous plot. In other words, in the case of riddling, the effect of
bubbling attractor also disappears rapidly with increasing d.
References
Anishchenko, V.S.: Nonlinear Dynamics of Chaotic and Stochastic Systems: Tutorial and Modern
Developments. Springer, Heidelberg (2002).
Anishchenko, V.S., Kopeikin, A.S., Kurths, J., Vadivasova, T.E., Strelkova G.I.: Studying hyper-
bolicity in chaotic systems. Physics Letters A 270, 301–307 (2000).
9). Springer-Verlag, New York (1995).
Arnold, V. I., Avez, A.: Ergodic Problems in Classical Mechanics. Benjamin, New York (1968).
Ashwin, P.: Riddled basins and coupled dynamical systems. In: Chazottes, J.-R. and Fernandez,
B.(eds.): Dynamics of Coupled Map Lattices and of Related Spatially Extended Systems, pp.
181–208. Springer, Berlin (2005).
Bricmont, J., Kupiainen, A.: Infinite-dimensional SRB measures. Physica D 103, 18–33 (1997).
Bunimovich, L.A, Sinai, Ya.G.: Spacetime chaos in coupled map lattices. Nonlinearity 1, 491–516
(1988).
Bunimovich, L.A., Sinai, Ya.G.: Statistical mechanics of coupled map lattices. In: Kaneko, K. (ed.)
Theory and Application of Coupled Map Lattices, pp. 169–189. John Wiley & Sons Ltd., New
York (1993).
Farmer, J., Ott, E., Yorke, J.: The dimension of chaotic attractors. Physica D 7, 153–180 (1983).
Ford, J.: The Fermi-Pasta-Ulam problem: Paradox turns discovery. Physics Reports 213, 271–310
(1992).
208 (1983).
Isaeva, O.B., Jalnine, A.Yu., Kuznetsov, S.P.: Arnold’s cat map dynamics in a system of coupled
nonautonomous van der Pol oscillators. Phys. Rev. E 74, 046207 (2006).
Järvenpää, E.: SRB-measures for coupled map lattices. In: Chazottes, J.-R. and Fernandez,
B.(eds.): Dynamics of Coupled Map Lattices and of Related Spatially Extended Systems, pp.
95–114. Springer (2005).
Kaneko, K. (ed.): Theory and Application of Coupled Map Lattices. John Wiley & Sons Ltd.,
Chichester (1993).
Kuptsov, P.V., Kuznetsov, S.P.: Transition to a synchronous chaos regime in a system of coupled
non-autonomous oscillators presented in terms of amplitude equations. Nonlinear Dynamics
(Izhevsk) 2(3), 307–331 (2006) (In Russian).
Kuznetsov, S.P., Sokha, Y.I.: Hyperchaos in model nonautonomous system with a cascade ex-
citation transmission through the spectrum. Izvestija VUZov – Applied Nonlinear Dynamics
(Saratov) 18, 24–32 (2010).
Landau, L.D., Lifshitz, E.M.: Fluid Mechanics (Course of Theoretical Physics, vol.6). Pergamon
Press, London (1959).
Letellier C., Rössler, O.E.: Hyperchaos. Scholarpedia, 2, 1936 (2007).
Matsumoto, T., Chua, L.O., Kobayashi, K.: Hyperchaos laboratory experiment and numerical con-
firmation. IEEE Trans, Circuits & Syst. 33, 1143–1147 (1986).
Monin, A.S., Yaglom, A.M.: Statistical Fluid Mechanics, Volume II: Mechanics of Turbulence.
Dover Publication, New York (2007).
Mosekilde, E., Maistrenko, Yu., Postnov, D.: Chaotic Synchronization: Applications To Living
Systems. World Scientific Publication, Singapore (2002).
Platt, N., Spiegel, E.A., Tresser C.: On-off intermittency: A mechanism for bursting. Phys. Rev.
Lett. 70, 279–282 (1993).
Reiterer, P., Lainscsek, C., Schürrer, F., Letellier C., Maquet, J.: A nine-dimensional Lorenz system
to study high-dimensional chaos. Journal of Physics A 31, 7121–7139 (1998).
Rössler, O.E.: An equation for hyperchaos. Physics Letters A 71, 155–157 (1979).
Stoop, R., Peinke, J., Parisi, J., Röhricht B., Hübener, R. P.: A p-Ge semiconductor experiment
showing chaos and hyperchaos. Physica D 35, 425–435 (1989).
Wolfram, S.: Theory and applications of cellular automata. World Scientific Publication, Singapore
(1986).
Chapter 9
Autonomous Systems Based on Dynamics Close
to Heteroclinic Cycle
Abstract This chapter is devoted to autonomous models generating successive

trains of oscillations with chaotic phase, which are designed on a base of sys-
tems with heteroclinic cycles in the phase space as suggested in our joint work
(Kuznetsov and Pikovsky, 2007). The heteroclinic cycle, or the heteroclinic loop
consists of several saddle fixed points and of trajectories joining them. The paradigm
example for a heteroclinic cycle (Guckenheimer and Holmes, 1988) is discussed in
Sect. 9.1. In a frame of the present research, it is interpreted as a set of equations
for real amplitudes of interacting oscillators. The equations are supplemented with
appropriate additional coupling terms in such a way that the alternating cyclic exci-
tation of the oscillators occurs accompanied by transformation of the phases in ac-
cordance with some chaotic map. The number of oscillators may be three or more;
so, in the models constructed in this way, the minimal phase space dimension is six.
In Sect. 9.2—9.4 we discuss three models, one with attractor of Smale-Williams
type, one possessing attractor with dynamics governed approximately by the Arnold
cat map, and an example of hyperchaos that corresponds to attractor with two pos-
itive Lyapunov exponents. All these examples are designed on a base of equations
written for complex amplitudes, and they implement non-resonance mechanism of
the excitation transfer between the alternately exciting oscillators. In Sect. 9.5 we
advance a model composed of a large number of van der Pol oscillators. Because
of slow variation of the natural frequencies of the oscillators around the ring struc-
ture of the device, it appears possible to use resonance mechanism for the excitation
transfer; so, the system may have prospects for implementing high-frequency chaos
generators.
9.1 Heteroclinic connection: an example of Guckenheimer and

Holmes
Normalized equation for complex amplitude of a single van der Pol oscillator above
the threshold of the Andronov-Hopf bifurcation is ȧ = (1 − |a| 2 )a, and respective
202 9 Autonomous Systems Based on Dynamics Close to Heteroclinic Cycle
equation for the squared real amplitude is ṙ = 2(1 − r)r, where r = |a| 2 . Consider
a symmetric system of some number of elements of such kind, which are arranged
in a ring and mutually affect each other via nonlinear multiplying factor, and each
amplitude variable suppresses those for the previous and the next elements in the
ring in different degrees. A simple example composed of three elements is a set of
equations
1
ṙ1 = 2 1 − r1 − r3 − 2r2 r1 ,
2

1
ṙ2 = 2 1 − r2 − r1 − 2r3 r2 , (9.1)
2

1
ṙ3 = 2 1 − r3 − r2 − 2r1 r3 .
2
This just corresponds to the paradigm model of Guckenheimer and Holmes (Guck-
enheimer and Holmes, 1988) rewritten in terms of squared variables with con-
cretized values of the coefficients.
Qualitatively, the behavior of the system may be explained as follows. Let us
suppose that initially all three variables are relatively small, but, say, r 1 is the largest.
Then, it grows in accordance with the first equation, reaches a level of order 1 and
nearly saturates. Now, it suppresses strongly the variable r 3 , because r1 enters the
third equation with factor 2, but it does not restrain the growth of the variable r 2 ,
1
as r1 is present in the second equation with factor . At the next stage, the variable
2
r2 grows, saturates, and suppresses the variable r 1 , but allows growth of r 3 , and so
on. So, the elements become active by turns in cyclic order r 1 → r2 → r3 → r1 →
· · · . Each next repetition of the process is longer in time because the successive
suppressions of the excitations become deeper.
In the phase space of the system (9.1) there exist three saddle fixed points
(r1 , r2 , r3 ) = (1, 0, 0), (0, 1, 0), and (0, 0, 1). They are connected by a contour con-
sisting of three smooth curved segments, each representing simultaneously the out-
going separatrix of one saddle and the ingoing separatrix of another (Fig. 9.1(a)).
Motion close to the contour corresponds to bypasses around all three segments in
the phase space visiting narrower neighborhoods of the saddles; there the motions
are very slow, so, the successive time intervals of the bypasses grow gradually and
tend to infinity (Fig. 9.1(b)).
Of course, concrete coefficients in (9.1) may be of other values. Moreover, anal-
ogous dynamics may be arranged as well with larger number of the elements and,
respectively, of the involved saddles.
Currently, there exists quite extensive literature devoted to dynamical phenomena
associated with the heteroclinic cycles and to its applications, e.g. in the context of
fluid convection, biological populations, neuroscience, etc. (Mercader et al., 2002;
Huisman and Weissing, 2001; Rabinovich et al., 2006).
Now we proceed with interpretation of r 1 , r2 , r3 as squared amplitudes of three
oscillators, rk = |ak |2 , k = 1, 2, 3. The equations for the complex amplitudes are
9.2 Attractor of Smale-Williams type in a system of three coupled self-oscillators 203
Fig. 9.1 A heteroclinic cycle in the phase space of the system (9.1) (a), and state variables r1,2,3
versus time; observe oscillations with a growing period (b).

1
ȧ1 + iω1 a1 = 1 − |a1|2 − |a3 |2 − 2|a2|2 a1 ,
2

1
ȧ2 + iω2 a2 = 1 − |a2| − |a1 | − 2|a3| a2 ,
2 2 2 (9.2)
2

1
ȧ3 + iω3 a3 = 1 − |a3|2 − |a2 |2 − 2|a1|2 a3 ,
2
where ω1,2,3 designate the frequency parameters. Obviously, the dynamics for the
squared amplitudes r k = |ak |2 will be precisely the same as that governed by Eqs.
(9.1).
9.2 Attractor of Smale-Williams type in a system of three

coupled self-oscillators
Let us modify Eqs. (9.2) to obtain an autonomous system with attractor of Smale-
Williams type in the Poincaré map (Kuznetsov and Pikovsky, 2007). We turn to a
symmetric case and set ω 1,2,3 = ω0 . Next, we introduce some additional coupling
terms in the equations expressed via quadratic nonlinear terms to ensure transfer of
the excitation to each next oscillator after its suppression stage from the previously
active partner. Setting a k = xk + iyk , in terms of the real and imaginary parts, we
write down

1
ẋ1 = ω0 y1 + 1 − (x21 + y21 ) − (x23 + y23 ) − 2(x22 + y22 ) x1 + ε x3 y3 ,
2

1 2
ẏ1 = −ω0 x1 + 1 − (x1 + y1 ) − (x3 + y3) − 2(x2 + y2 ) y1 ,
2 2 2 2 2
2

1
ẋ2 = ω0 y2 + 1 − (x22 + y22 ) − (x21 + y21 ) − 2(x23 + y23 ) x2 + ε x1 y1 ,
2
(9.3)
1 2
ẏ2 = −ω0 x2 + 1 − (x2 + y2 ) − (x1 + y1) − 2(x3 + y3 ) y2 ,
2 2 2 2 2
2

1
ẋ3 = ω0 y3 + 1 − (x3 + y3 ) − (x2 + y2 ) − 2(x1 + y1 ) x3 + ε x2 y2 ,
2 2 2 2 2 2
2

1
ẏ3 = −ω0 x3 + 1 − (x23 + y23 ) − (x22 + y22) − 2(x21 + y21 ) y3 .
2
Here ε is the coefficient of the additional coupling. In the process of the dynamics,
the added terms serve as germs for excitation of each next oscillator at the beginning
of the epochs of its activity. It occurs just at passages of the orbit near the saddle
fixed points on the underlying heteroclinic cycle. The oscillators become active fol-
lowing the order 1 → 2 → 3 → 1 → · · · . Each new transfer of the excitation is ac-
companied by the phase doubling. Indeed, let x k ∼ sin(ω0t + ϕ ), yk ∼ cos(ω0t + ϕ );
then the coupling term in the equation for the next oscillator contains the doubled
1
phase: xk yk ∼ sin(ω0t + ϕ ) cos(ω0t + ϕ ) = sin(2ω0t + 2ϕ ). That oscillator under-
2
goes excitation in the presence of this stimulating oscillating force; so it will inherit
this doubled phase at the stage of activity, and so on. The mechanism of transfer
of the excitation is non-resonance: indeed, the frequency of the stimulating force is
twice as large as the natural frequency of the oscillator. Everything works, however,
if one sets the frequency parameters relatively small.
To illustrate functioning of the model in computations, we fix ω 0 =1 and ε =0.03.
At these parameter values, testing different initial conditions, only one attractor
could be observed.
Figure 9.2 shows typical time dependence for the variables x 1 , x2 , and x3 as ob-
tained from the computer solution of Eqs. (9.3). Observe that in the course of time
evolution, three oscillators become excited one after another in accordance with the
above qualitative description. The process is not periodic, although there is a rather
well-defined mean period of the amplitude dynamics (in contrast to the model with-
out the added coupling terms, cf. Fig. 9.1).
To give evidence for the phase doubling, which is crucial for occurrence of the
attractor of Smale-Williams type, let us construct the Poincaré map. Here it will be
done with some trick taking into account the symmetry of the system. Namely, in the
9.2 Attractor of Smale-Williams type in a system of three coupled self-oscillators 205
Fig. 9.2 Time dependence for x1 (black curve), x2 (light gray curve), and x3 (dark gray curve),
obtained from the numerical computer solution of Eqs. (9.3) ω0 = 1 and ε = 0.03.
six-dimensional phase space {x 1 , y1 , x2 , y2 , x3 , y3 } let us define three hyper-surfaces

by the equations
S1 = x21 + y21 − x23 + y23 = 0,
S2 = x22 + y22 − x21 + y21 = 0, (9.4)
S3 = x23 + y23 − x22 + y22 = 0,
and examine successive intersections of them in the direction of growth of the values
of the quantities S j . The intersection of a surface S j = 0 by a trajectory corresponds
to the condition that amplitude of the rising oscillator (x j , y j ) just overtakes that of
the previously excited partner. The intersections follow in the order S 1 → S2 →
S3 → · · · . A given point on a surface S j = 0, j = 1, 2, 3 corresponds to a five-
dimensional vector v n . On the next intersection of a surface of the family (9.4)
by the trajectory started at v n we obtain a new vector; this determines the map
vn+1 = T(vn ). The spaces of images and pre-images may be identified because of
the cyclic symmetry of the system. The usual Poincaré map will correspond to a
threefold iteration of the map T.
At each intersection of a surface S j = 0 at time instant tn we define a phase for
the j-th oscillator as ϕn = arg(x j (tn ) + iy j (tn )). The diagram for ϕ n+1 versus ϕn is
shown in Fig. 9.3. The presence of the phase doubling mechanism is evident from
Fig. 9.3 Iteration diagram for phases obtained in computations for the model (9.3) at ω0 = 1 and
ε = 0.03; the phases are determined at each successive intersection of the family of surfaces Sj .
the picture: the diagram looks topologically similar to the expanding circle map
ϕn+1 = 2ϕn + const ( mod 2π ).
Figure 9.4 shows projections of the attractor in the Poincaré map onto three
planes of the variables (x j , y j ) corresponding to three partial oscillators. Because
of strong transversal compression in the phase space, the fractal structure of the
attractor is not distinguishable on diagrams (a) and (c) for the oscillators, whose
amplitudes are equal in the condition of the Poincaré section, but it can be seen
on the portrait for the rest oscillator (b) possessing the smallest amplitude at this
moment in time.
Fig. 9.4 Portraits of attractor for the model (9.3) in the cross-section of the phase space by the
surface S1 = 0 (see (9.4)) depicted in projection onto planes of the variables (x1 , y1 ) (a), (x2 , y2 )
(b), and (x3 , y3 ) (c) at ω0 = 1 and ε = 0.03.
Computation of the spectrum of Lyapunov exponents for the attractor in accor-

dance with the Benettin algorithm requires simultaneous numerical solution of Eqs.
(9.3) together with collection of six replicas of the linearized variation equations
for perturbation vectors accompanied by Gram-Schmidt orthogonalization and nor-
malization at successive steps of the algorithm with estimate of rates of variation of
respective accumulating sums for logarithms of the norm ratios (Appendix A). In
the dynamical regime under consideration it yields
λ1 = 0.06305, λ 2 = 0.00000, λ 3 = −0.2101,

(9.5)
λ4 = −0.2259, λ 5 = −0.3395, λ 6 = −1.99993.
One Lyapunov exponent is positive, the next one is close to zero (it corresponds,
clearly, to the perturbation directed along the orbit), and the remaining four expo-
nents are negative.
For the map T, the Lyapunov exponents may be expressed via the non-zero expo-
nents of the flow system by means of the relation Λ m = Tav λm , where Tav is the aver-
age time interval between successive intersections of the surfaces (9.4). Following
from the numerical simulations, in the dynamical regime we consider, T av ≈ 10.96.
Hence, Λ1 ≈ 0.691 that agrees well with the value ln 2 ≈ 0.693 expected from the
approximation based on the expanding circle map.
9.3 Attractor with dynamics governed by the Arnold cat map 207
Calculation of the attractor dimension from the Kaplan-Yorke formula (Appendix

D) yields for the flow system D KY = 2 + λ1/|λ3 | ≈ 2.30.
9.3 Attractor with dynamics governed by the Arnold cat map
Let us start again with the system (9.2) at ω 1,2,3 = ω0 , but introduce the additional
coupling terms in other way, namely, each partial oscillator is excited by a product
of variables relating to two other oscillators. Then, we arrive at the equations

1 2
ẋ1 = ω0 y1 + 1 − (x1 + y1 ) − (x3 + y3 ) − 2(x2 + y2 ) x1 + ε (x2 y3 + x3 y2 ),
2 2 2 2 2
2

1
ẏ1 = −ω0 x1 + 1 − (x21 + y21 ) − (x23 + y23) − 2(x22 + y22 ) y1 ,
2

1
2 2 2 2 2 2
2
(9.6)
1
ẏ2 = −ω0 x2 + 1 − (x22 + y22 ) − (x21 + y21) − 2(x23 + y23 ) y2 ,
2

1
2
1
ẏ3 = −ω0 x3 + 1 − (x3 + y3 ) − (x22 + y22) − 2(x21 + y21 ) y3 .
2 2
2
The added coupling terms are selected in symmetric form to exclude the presence of
the non-oscillating components, which could disturb the required phase transfer. To
illustrate the dynamics we fix ω 0 =1 and ε = 0.03. The computations indicate that a
unique attractor is observed regardless of initial conditions.
The oscillators become excited following the order 1 → 2 → 3 → 1. Each new
transfer of the excitation is accompanied by assigning new phase with a sum of two
previous phases. Indeed, let us suppose that two partner elements transferring the
excitation to the given one oscillate as x j ∼ sin(ω0t + ϕ ), y j ∼ cos(ω0t + ϕ ) and
xk ∼ sin(ω0t + φ ), yk ∼ cos(ω0t + φ ). Then the respective coupling term contains a
sum of the phases
x j yk + xk y j ∼ sin(ω0t + ϕ ) cos(ω0t + φ ) + sin(ω0t + φ ) cos(ω0t + ϕ )

= sin(2ω0t + ϕ + φ ).
As the excitation takes place in the presence of this stimulating oscillating force, the
oscillator will inherit the sum of previous phases; so, the successive transformations
will follow the relation ϕ n+1 = ϕn + ϕn−1 ( mod2π ) called the Fibonacci map. It
is connected with the Arnold cat map. Indeed, setting q n = 2πϕn, pn = 2πϕn−1 , we
have pn+1 = qn , qn+1 = pn + qn ( mod 1), and the two-fold iteration just corresponds
to the map
pn+2 = pn + qn, qn+2 = pn + 2qn ( mod 1). (9.7)
Figure 9.5 illustrates evolution of the variables x 1 , x2 , and x3 , as functions of time

obtained from the computer solution of Eqs. (9.6) after the decay of transients. Ob-
serve that the whole process is chaotic and manifests successive excitation of the
partial oscillators in proper order with some characteristic average time period.
obtained from the numerical computer solution of Eqs. (9.6) at ω0 = 1 and ε = 0.03.
Fig. 9.6 Portraits of the attractors of the model (9.6) in the cross-section of the phase space by the
surface S j = 0 (see (9.4)) depicted in projection onto planes of the variables (xj , y j ) (a), (x j+1 , y j+1 )
(b), and (x j+2 , y j+2 ) (c) at ω0 = 1 and ε = 0.03.
We can construct the Poincaré map in the same way as described in the previous
section, namely, in the six-dimensional phase space {x 1 , y1 , x2 , y2 , x3 , y3 } we define
three surfaces determined by Eqs. (9.6) and examine successive intersections of
these surfaces by a trajectory. In this way we define a five-dimensional map v n+1 =
T(vn ): the pre-image is a vector v n associated with a point on the surface S j , and
the image is a vector v n+1 on the next intersected surface S k = 0, which follows the
order S1 → S2 → S3 → · · · . Like in the previous subsection, the spaces of images and
pre-images may be identified due to the symmetry of the system. Figure 9.6 shows
projections of the attractor in the Poincaré map onto three planes of the variables
(x j , y j ) corresponding to three partial oscillators.
For a sequence of successive intersections of the surfaces S j = 0 we can define
phases of the oscillations as ϕ n = arg[x j (tn ) + iy j (tn )]. Figure 9.7 demonstrates that
they indeed obey the Fibonacci map ϕ n+1 = ϕn + ϕn−1 + const with a good accu-
9.3 Attractor with dynamics governed by the Arnold cat map 209
racy. As mentioned above, the second iteration of the Fibonacci map is the Arnold
cat map. Figure 9.8 illustrates the action of the map obtained from the numerical
solution of Eqs. (9.6) with a traditional picture of the cat face. It may be compared
to Fig. 1.6 in Chap. 1.
Fig. 9.7 Three-dimensional representation of the dependence of ϕn+1 on ϕn and ϕn−1 in model
(9.6) with ω0 = 1 and ε = 0.03.
Fig. 9.8 Diagram for phases of partial oscillators in the model (9.6) with ω0 = 1 and ε = 0.03
representing successive steps of evolution of the cloud of representative points in a form of the cat
face.
Computation of the Lyapunov exponents yields
λ1 = 0.04786, λ 2 = 0.00000, λ 3 = −0.04840,

(9.8)
λ4 = −0.3184, λ 5 = −0.3673, λ 6 = −2.00002.
There is one positive exponent, one zero exponent associated with a shift along the
trajectory, and others negative. Taking into account that the average time between
the intersections of the surfaces S j = 0 at the given parameters is approximately
Tav ≈ 9.94, one can evaluate the Lyapunov exponents for the map v n+1 = T(vn ): as
Λm = Tav λm , Λ1 ≈ 0.476, and Λ 3 ≈ −0.481. These numbers should be compared
1 √
with logarithms of square roots of the eigenvalues for the map (9.7) ln (1 + 5) ≈
2
1 √
0.481 and ln ( 5 − 1) ≈ −0.481. Note that for the cat map a sum of the Lyapunov
2
exponents vanishes. In the system (9.6) the attractor in the Poincaré map is concen-
trated close to a two-dimensional torus surface, the dynamics on which are governed
approximately by the cat map. This assertion agrees with the fact that the sum of the
largest two non-zero Lyapunov exponents is very small: λ 1 + λ3 ≈ −0.0005. In fact,
due to numerical statistical errors it is hard to distinguish it from zero. Therefore we
conclude that the Lyapunov dimension of the attractor is nearly 3, and the dimension
of the attractor in the Poincaré map is close to 2. This explains the visible absence
of a fractal structure in Fig. 9.6 (c).
9.4 Model with hyperchaos
Hyperchaos is a dynamical regime with two or more positive Lyapunov exponents

(Rössler, 1979; Matsumoto et al., 1986; Stoop et al., 1989; Reiterer et al., 1998;
Letellier and Rössler, 2007). Let us introduce a system with a hyperchaotic at-
tractor basing on the system of three interacting oscillators evolving close to the
heteroclinic cycle. In contrast to two previous models, the present one will be non-
symmetric. In Eqs. (9.2) we assign ω 1 = 3ω0 , ω2,3 = ω0 , and introduce additional
coupling characterized by constants ε 1,2,3 as follows:

1 2
ẋ1 = 3ω0 y1 + 1 − (x1 + y1 ) − (x3 + y3 ) − 2(x2 + y2 ) x1 + ε1 x22 y2 ,
2 2 2 2 2
2

1
ẏ1 = −3ω0 x1 + 1 − (x21 + y21 ) − (x23 + y23 ) − 2(x22 + y22 ) y1 ,
2

1
ẋ2 = ω0 y2 + 1 − (x2 + y2 ) − (x1 + y1 ) − 2(x3 + y3) x2 + ε2 x3 ,
2 2 2 2 2 2
2
(9.9)
1 2
ẏ2 = −ω0 x2 + 1 − (x2 + y2 ) − (x1 + y1 ) − 2(x3 + y3 ) y2 ,
2 2 2 2 2
2

1
ẋ3 = ω0 y3 + 1 − (x23 + y23 ) − (x22 + y22 ) − 2(x21 + y21) x3 + ε3 (x1 y2 + x2 y1 ),
2

1 2
ẏ3 = −ω0 x3 + 1 − (x3 + y3 ) − (x2 + y2 ) − 2(x1 + y1 ) y3 .
2 2 2 2 2
2
In computations we set ω 0 = 1, ε1 = ε2 = 0.004, ε3 = 0.1. Starting with various

initial conditions the system was observed to arrive always to one and the same
attractor. Figure 9.9 shows the variables x 1 , x2 , and x3 , as functions of time obtained
from the computer solution of Eqs. (9.9) after the decay of transients. Successive
excitations of the partial oscillators follow, like in the previous models, the sequence
9.4 Model with hyperchaos 211
1 → 2 → 3 → 1 → 2 → 3 → ··· .
obtained from the numerical computer solution of Eqs. (9.9) at ω0 = 1, ε1 = ε3 = 0.004, ε2 = 0.1.
Let us consider the phases the oscillators accept at the cross-sections of the sur-
faces determined by the conditions (9.4); then we will have a sequence
ϕ1 → ϕ2 → ϕ3 → ϕ1 → ϕ2 → ϕ3 → · · · .
Excitation of the oscillator 1 is stimulated by the force proportional to the term

x22 y2 . The dominant contribution is due to the third harmonic of the basic frequency
ω0 as it is in resonance with the oscillator 1. Hence, at the beginning of the epoch
of activity the oscillator 1 adopts the phase
ϕ1 = 3ϕ2 + const. (9.10)
Oscillator 2 in the course of the excitation is driven in resonance by the force pro-
portional to x 3 , so it takes up the phase from the oscillator 3:
ϕ2 = ϕ3 + const. (9.11)
Finally, oscillator 3 is driven by the signal x 1 y2 + x2 y1 . At the beginning of the active

epoch it is excited in a non-resonance manner (this is the reason why we assign a
relatively large value for ε 3 ), and accepts the phase
ϕ3 = ϕ1 + ϕ2 + const. (9.12)
Figure 9.10 demonstrates that the phase relations indeed occur in the dynamics gov-
erned by Eqs. (9.9) following from computations.
One may remove the constants in (9.9)—(9.11) by appropriate shifts of origins
for the variables ϕ 1,2,3 and write down the relation between the pairs (ϕ 3 , ϕ2 ) and
(ϕ3 , ϕ2 ) in the matrix form

ϕ3 1 3 ϕ3
= . (9.13)
ϕ2 1 0 ϕ2
Fig. 9.10 Phase transformation plots obtained in computations for the model (9.9) at ω0 = 1, ε1 =
ε3 = 0.004, ε2 = 0.1.
The matrix here has two eigenvalues larger than 1, namely,

√ √
s1 = ( 13 + 1)/2 = 2.30277 · · · , s 2 = ( 13 − 1)/2 = 1.30277 · · · . (9.14)
Hence, the Poincaré map corresponding to a complete bypass around the hetero-
clinic contour is expected to have two positive Lyapunov exponents associated with
the dynamics of the phase variables.
The Lyapunov exponents for the model (9.9) evaluated from the computations
are the following:
λ1 = 0.01878, λ 2 = 0.00621, λ 3 = 0.00000,

(9.15)
λ4 = −0.3680, λ 5 = −0.4313, λ 6 = −2.00003.
The presence of two positive Lyapunov exponents affirms the hyperchaotic nature
of the dynamics. The third exponent is close to zero and has to be interpreted as
being linked with infinitesimal shift along the orbit. Other exponents are negative;
they correspond to approach of trajectories to the attractor.
In the present case, all consecutive transformations ϕ 1 → ϕ2 → ϕ3 → ϕ1 are
different; so, one should consider the overall transformation with the return onto a
cross-section, say, S 1 = 0, as the Poincaré map. As evaluated from the computations,
at the assigned parameters the average time between successive returns is T av ≈
41.475. The largest Lyapunov exponents for the Poincaré map are evaluated via the
exponents of the flow system as Λ 1 = Tav λ1 ≈ 0.779 and Λ 2 = Tav λ2 ≈ 0.258. They
may be compared with the estimates obtained from the qualitative consideration of
the dynamics of phases as ln s 1 = 0.834 and ln s 2 = 0.265. Observe quite a good
correspondence.
The Lyapunov dimension of the attractor estimated from the Kaplan-Yorke for-
mula is 3.068.
It may be suggested that attractor in the Poincaré map of the considered model
is hyperbolic, but verification of this assertion would require reformulation of the
algorithm for cone criterion presented in Chap. 7 because the unstable manifold for
this attractor is two-dimensional.
9.5 An autonomous system with attractor of Smale-Williams type with . . . 213
9.5 An autonomous system with attractor of Smale-Williams

type with resonance transfer of excitation in a ring array of
van der Pol oscillators
Examples examined in Sect. 9.2—9.4 are designed departing from equations of in-
teracting oscillators written for complex amplitudes. If one intends to organize such
dynamics in practice, it would be possible on a base of high-frequency oscillations
in situation of applicability of description in terms of slow complex amplitudes.
On the other hand, non-resonance mechanism of transfer of excitation between the
alternately active oscillators is able to ensure phase doubling or other chaotic trans-
formation for phases only at relatively low natural frequencies. These two condi-
tions contradict each other; so, practical implementation is questionable. One way
to overcome this difficulty is to consider a large number of oscillators arranged in
such a way that the neighbors transferring each other the excitation possess close
natural frequencies, which decrease gradually from the beginning to the end of the
chain. The final stage of the excitation transfer at the closure of the chain is of special
kind being accompanied by doubling of the respective frequency and the oscillation
phase. In the present section this idea is considered in application to the ring model
composed of van der Pol oscillators, following (Kruglov and Kuznetsov, 2011).
As mentioned in Chap. 4, a single autonomous van der Pol oscillator is governed
by a differential equation
ẍ − (A − x2)ẋ + ω02 x = 0,
where ω0 is the natural frequency. At positive values of the parameter A it demon-

strates self-oscillations associated with the attractive limit cycle. Let us consider the
following ring model composed of N + 1 coupled van der Pol oscillators:
3
ẍ0 − (A + x2N + x20 − 2S)ẋ0 + ω02 x0 = ε x2N ,
2
3 2
ẍ1 − (A + x0 + x21 − 2S)ẋ1 + 2−2/N ω02 x1 = ε x0 ,
2
3 2
ẍ2 − (A + x1 + x22 − 2S)ẋ2 + 2−4/N ω02 x2 = ε x1 ,
2 (9.16)
············
3
ẍN − (A + x2N−1 + x2N − 2S)ẋN + 2−2ω02 xN = ε xN−1 ,
2
N
S= ∑ x2k .
k=0
Structure of the factors at the first derivatives ensures a situation that each given
oscillator contributes to strong suppression of all other partners, except one num-
bered with the next index. Hence, the system manifests cyclic propagation of the
localized excitation around the ring in direction of increasing indices, in accordance
with mechanism explained in Sect. 9.1. Due to presence of the additional coupling
characterized by parameter ε the excitation of each j-th oscillator is stimulated by
oscillations from ( j − 1)-th partner. As N is large enough, the natural frequencies of
the neighbor oscillators are close, and the stimulation is practically resonance. Only
at the edge of the chain, the transfer from N-th oscillator to zero one is arranged in
other way: the first one has the natural frequency twice less that the last one, and
the resonance stimulation is produced by the second harmonic component of the
coupling term proportional to x 2N in the first eq. (9.16).
Figure 9.11 shows in panel (a) the time dependence for generalized coordinates
of oscillators in the ring structure for N=14 obtained from numerical computer so-
lution of Eqs. (9.16) after exclusion of the transients.
Fig. 9.11 Time dependence for x j , j = 0, . . ., N and auxiliary variable z used for construction of the
return map obtained from the numerical computer solution of Eqs. (9.16) at N = 14, A = 1, ω0 =
2π , ε = 0.2.
In accordance with the idea of the design, the oscillatory excitation has to un-
dergo doubling of phase with each bypass of the ring. This property is crucial for
the realization of attractor of Smale-Williams type, and it is highly desirable to il-
lustrate it in the numerical computations.
It is not so trivial to construct appropriate Poincaré map: simple definitions sim-
ilar to (9.4) appear to be unsuccessful. The single van der Pol system is not so close
to perfect symmetry intrinsic to the amplitude equations, and it is problematic to
locate the instant of the cross-section relatively to envelope of the oscillations with
inaccuracy essentially less than the period associated with the natural frequency to
have good definition of the phases.
To settle this problem, let us introduce an auxiliary variable z obeying the differ-
ential equation
ż + γ z = +x20 + ω0−2ẋ20 − x21 − 22/N ω0−2 ẋ21 . (9.17)
Here the right-hand part is the difference of squared amplitudes for the oscillatory
elements 0 and 1, and γ is positive constant. The behavior of the variable z is il-
lustrated by panel (b) of Fig. 9.11 at particular γ = 0.1. The value z grows during
the epoch of activity of the oscillator 0, then decreases below zero during epoch of
9.5 An autonomous system with attractor of Smale-Williams type with . . . 215
activity for the oscillator 1, and then decays nearly to zero until the next activation
of the oscillator 0 occurs. For the record of the system states and evaluation of the
phases ϕ = arg(x0 + iẋ0 /ω0 ) we take sequence of the instants of the transitions from
positive to negative values of the variable.
Figure 9.12 shows the diagram for ϕ n+1 versus ϕn . Observe that it looks topolog-
ically corresponding to the Bernoulli map: one complete revolution for pre image
gives rise to two revolutions for image. Accounting strong compression in other
directions in the phase space (excluding the cyclic phase variable and the neutral
direction along the phase trajectory), it gives a reason for the conjecture that hyper-
bolic attractor does exist, which corresponds to suspension (special flow) over the
Smale-Williams solenoid.
Fig. 9.12 Iteration diagram for phases obtained in computations for the model (9.16) at N =
14, A = 1, ω0 = 2π , ε = 0.2. The phases are determined at instants of each successive passage of
them by the auxiliary variable z evolving in accordance with Eq. (9.17) from positive to negative
values.
Figure 9.13 presents portraits of the attractor in projection onto the phase plane
of oscillator j = 0. Panel (a) shows the attractor for the flow system in projection
from the phase space of dimension 2(N + 1). In panel (b) attractor in the cross-
section defined with the help of the auxiliary variable is plotted. The last one is, in
fact, a picture of solenoid; the intrinsic fine transversal Cantor-like structure may be
resolved under magnification as shown by the inset.
Figure 9.14 shows portraits of the attractor for the flow system in some other
projections, with coordinate axes corresponding to neighboring oscillators in the
ring. The first picture on the plane (x N , x0 ) is of specific kind; all the rest ones are
qualitatively similar to each other. It is so because just transfer of the excitation
from the N-th oscillator to zero one is arranged through the quadratic coupling term
and the second harmonic; all others are stimulated directly by linear terms in the
equations.
By means of the Benettin algorithm (Appendix A) applied to the set of Eqs.
(9.16) and respective variation equations, the larger Lyapunov exponents were eval-
uated:
λ1 = 0.003044, λ 2 = 0.000002, λ 3 = −0.3353, λ 4 = −0.6318 · · · . (9.18)

Fig. 9.13 Attractor of the model (9.16) at N = 14, A = 1, ω0 = 2π , ε = 0.2 in projection onto the
phase plane of the oscillator j = 0 from the multidimensional phase space of the flow system (a)
and in the cross-section constructed using the auxiliary variable z as explained in the text (b).
Fig. 9.14 Portraits of the attractor of the model (9.16) at N = 14, A = 1, ω0 = 2π , ε = 0.2 in pro-
jections from the multidimensional phase space of the flow system onto the planes of generalized
coordinates of two neighboring oscillators un the ring structure: N and 0 (a), 0 and 1 (b), N − 1
and N (c). For other neighboring pairs the diagrams look qualitatively similar to (b) and (c).
The largest Lyapunov exponent is positive that means the dynamics is chaotic. Ac-
counting that intersections of the axis of the auxiliary variable z=0 (see (9.17)) used
in the analysis of the phase map occur with average time period T av = 227.85, it
is worth considering normalization of the Lyapunov exponent to this time interval
that yields Λ1 = λ1 Tav ≈ 0.6934. Observe a good correspondence with the estimate
following from the Bernoulli map for the phases, ln 2 ≈ 0.6931. So, we interpret the
largest exponent as being associated with the cyclic coordinate in the phase space of
dimension 2(N + 1) along which the expansion takes place. The second exponent is
close to zero, up to the numeric error, and should be related to perturbations of a shift
along the reference phase trajectory responsible for appearance of a zero term in the
spectrum of Lyapunov exponents of the autonomous system. The third exponent is
negative, and its absolute value is large enough to ensure overall compression of the
phase volume sufficient to evaluate the Kaplan-Yorke dimension of the attractor:
References 217
DKY ≈ 2.009 (in application to the flow system). Although the full spectrum of Lya-
punov exponents contains 2(N + 1) numbers, it seems not necessary to compute all
of them because the rest ones correspond to yet stronger compression in the phase
space and do not influence the estimate of the dimension.
In conclusion, we can assume that based on similar ring structures, using other
methods of introducing of coupling, the dynamics associated with the Arnold cat
map and with hyper-chaotic attractors also can be realized in analogy to the previous
sections of this chapter.
References
Guckenheimer, J., Holmes, P.: Structurally stable heteroclinic cycles. Math. Proc. Cambridge Phi-
los. Soc. 103, 189–192 (1988).
Huisman, J., Weissing, F.J.: Biological conditions for oscillations and chaos generated by multi-
species competition. Ecology 82, 2682–2695 (2001).
Kruglov, V.P., Kuznetsov, S.P.: An autonomous system with attractor of Smale-Williams type with
resonance transfer of excitation in a ring array of van der Pol oscillators. CNSNS 16, 3219–3223
(2011).
Physica D 232, 87–102 (2007).
Letellier C., Rössler, O.E.: Hyperchaos. Scholarpedia 2, 1936 (2007).
Matsumoto, T., Chua, L.O., Kobayashi, K.: Hyperchaos laboratory experiment and numerical con-
firmation. IEEE Trans, Circuits & Syst. 33, 1143–1147 (1986).
Mercader, I., Prat, J., Knobloch, E.: Robust heteroclinic cycles in two-dimensional Rayleigh-
Benard convection without Boussinesq symmetry. Int. J. Bifurcation and Chaos 12, 2501–2522
(2002).
Rabinovich, M.I., Varona, P., Selverston, A.I., Abarbanel, H.D.I.: Dynamical principles in neuro-
science. Rev. Mod. Phys. 78, 1213–1265 (2006).
Reiterer, P., Lainscsek, C., Schürrer, F., Letellier C., Maquet, J.: A nine-dimensional Lorenz system
to study high-dimensional chaos. Journal of Physics A 31, 7121–7139 (1998).
Rössler, O.E.: An equation for hyperchaos. Physics Letters A 71, 155–157 (1979).
Stoop, R., Peinke, J., Parisi, J., Röhricht B., Hübener, R. P.: A p-Ge semiconductor experiment
showing chaos and hyperchaos. Physica D 35, 425–435 (1989).
Chapter 10
Systems with Time-delay Feedback
Abstract Besides the alternately excited oscillators, the principle of manipulation

by phases in the course of the excitation transfer may be applied to systems with
delayed feedback. In this case, it is sufficient to have a single self-oscillator man-
ifesting successive stages of activity and suppression, while the excitation transfer
accompanied with appropriate phase transformation is carried out through the de-
layed feedback loop, from one stage of activity to another. In practical implementa-
tion, these systems can be even simpler than the alternately excited oscillators. From
a mathematical point of view, they are more complex, because presence of the delay
implies formally infinite dimension of the state space. Careful mathematical analy-
sis of attractors in such systems, including rigorous foundation of hyperbolicity, is
a challenging problem that goes far beyond the scope of this book. In the present
chapter, we start with some necessary material concerning the mathematical na-
ture of the differential equations with delay. The next two sections consider models
of systems with delay, generating sequences of pulses with the phases of the carrier
transformed from pulse to pulse in accordance with expanding circle maps. First, the
non-autonomous systems are discussed that operate due to the externally imposed
parameter modulation with or without an auxiliary reference signal. Afterwards, we
consider an autonomous system with time delay. On the physical level of reasoning
it seems confident that chaotic attractors in these systems are of the same nature as
those in the above discussed alternately excited oscillatory systems being associ-
ated with expanding circle maps for some cyclic variable. Particularly, it is believed
that the generated chaos is structurally stable. As one can hypothesize, attractors in
these systems as they are treated in a framework of discrete-time description (like
the Poincaré maps) belong to the class of solenoids of Smale-Williams embedded
in the infinite-dimensional phase space.
220 10 Systems with Time-delay Feedback
10.1 Some notions concerning differential equations with

deviating argument
Time-delay occurs in many systems in nature and technology, particularly, in au-

tomatic control theory, electronics, nonlinear optics, physiology, etc. (Kuznetsov,
1982; Vallee et al., 1984; Glass and Mackey, 1988; Hu and Wang, 2002; Chiasson
and Loiseaum, 2007). Analysis of the time-delay systems is, therefore, of both the-
oretical and practical importance. Unlike systems governed by ordinary differential
equations, the delay systems are characterized by infinite dimension of their state
spaces.
Natural mathematical tool for description of the time-delay systems is the the-
ory of differential equations with deviating argument (Bellman and Cooke, 1963;
Myskis, 1972; El’sgol’ts and Norkin, 1973; Hu and Wang, 2002). In general, these
are equations connecting an unknown vector function and its derivatives taken for
different values of the argument. A particular, but sufficiently common form of such
equation may be written as
x(n) (t) = f(t, x(t), x(1)(t), . . . , x(n−1) (t), x(t − τ0 ), x(1) (t − τ1 ), . . . , x(m) (t − τm )),
(10.1)
where x is a vector of some finite dimension N, f is a vector function, x (k) (t) is the
derivative d k x/dt k evaluated at the instant t. The values τ i , i = 0, . . . , m are assumed
to be positive constants.
Equation (10.1) is said to be of retarded type if the maximal order of derivatives
at shifted arguments is less than the order of the derivative in the left-hand part:
0 m < n. If m = n it is called an equation of neutral type, and for m > n it is
an equation of leading type. Well studied is the retarded type; to less extent it may
be said about the neutral type. For the leading type development of the consistent
theory is difficult and problematic because the initial-value problem appears to be
ill-posed. Below, we deal only with equations of retarded type.
To explain the main feature of the delay systems, consider a simple particular
case of (10.1), namely, an equation with N = 1, n = 1, m = 0:
ẋ = f (x(t), x(t − τ )). (10.2)
To start integration at a given time t =0 and proceed at 0 < t τ we need to eval-

uate step by step the right-hand part, which depends on the variable x at delayed
time instants from the interval [−τ , 0]. Hence, the unknown function x(t) must be
specified somehow on this interval [−τ , 0]. This function must be regarded as a kind
of initial condition. The same reasoning is suitable for any other point in time t,
after which we wish to continue the solution. In a framework of dynamical system
theory, we have to say that the instant state of the system (10.2) at some t is de-
termined by a function defined in a time interval of duration equal to the retarding
constant: x(t )|t ∈[t−τ ,t] = ξ (t ). Thus, the phase space for this system is the space of
functions of one real argument defined in interval of length τ . This space is infinite-
10.1 Some notions concerning differential equations with deviating argument 221
dimensional. Indeed, one can think of these functions as vectors; the components of
such vector are values of the function at the set of points of the interval of length τ .
Integration of the equation forward in time is not difficult; in principle it is re-
duced to step-by-step solution of successive set of ordinary differential equations:
ẋ1 = f (x1 , ξ ), 0 t τ ,
ẋ2 = f (x2 , x1 ), τ t 2τ , (10.3)
ẋ3 = f (x3 , x2 ), 2τ t 3τ ,
and so on. In the case of several retarding constants one should use the maximal one
to specify a length of the interval determining domain of definition for the initial
functions.
Numerical integration of the equations of retarded type by finite-difference
method is organized similar to that for ordinary differential equations.
A feature is that we have to deal with array of values of the unknown function
sampled with the integration step and relating to a time interval of required dura-
tion, which represents the initial function. With data from this array, we evaluate
the right-hand part of the equation and perform a time step of integration. One can
use methods analogous to those of Euler or Runge-Kutta of different orders; if nec-
essary, an interpolation procedure consistent in accuracy with the finite-difference
scheme may be applied over the array of the data. As a step has been made, the array
is updated, and the procedure goes on.
In contrast to ordinary differential equations, really hard is a question on possi-
bility to trace dynamics backward in time. If trying to do so, we have to deal actually
with the equation of leading type that is the ill-posed problem, as mentioned above.
Is it possible to generalize the hyperbolic theory and to introduce uniformly hy-
perbolic attractors in systems governed by the differential equations with deviating
argument? It seems that this question was not examined in mathematical literature,
but it certainly deserves attention of mathematicians. On the other hand, account-
ing close relation of the differential-delay equations with ordinary equations via the
“step method” (expressed by relations (10.3)) and applicability of approaches devel-
oped for design of systems with uniformly hyperbolic attractors to the time-delayed
systems (see below), one can expect that the answer will be positive.
To get impression on possible behavior of neighboring orbits in the time-delayed
system, let us consider a particular question on stability analysis of a fixed point in
the model (10.2). Let the fixed point solution be x = x 0 , and we consider dynamics
close to it setting x(t) = x0 + x̃(t). Substitution in (10.2) yields
x̃˙ = ax̃(t) + bx̃(t − τ ), (10.4)
where a = f 1 (x0 , x0 ) and b = f 2 (x0 , x0 ) are derivatives of the function f in respect

to the first and the second arguments evaluated at these arguments equal to x 0 . The
exponential substitution x̃(t) ∼ e λ t reduces (10.4) to the transcendent characteristic
equation:
λ = a + be−λ τ . (10.5)
Equations of such kind for the case of retarding type problems can have only a
restricted number of roots in the right half-plane of complex variable λ , which are
associated with instability. However, there is an infinite set of complex roots in the
left half-plane; their real parts asymptotically tend to −∞. 1 It means that a saddle-
fixed point in the time-delayed system has a finite-dimensional unstable subspace,
and an infinite-dimensional stable subspace. Of analogous type will be structure of
other typical saddle orbits in the state space of time-delayed systems.
Methods of computation of Lyapunov exponents for the time-delay systems were
discussed, e.g. in (Farmer, 1982; Giacomelli and Politi, 1996; Le Berre et al., 1987;
Lepri et al., 1994; Balyakin and Ryskin, 2007; Pazó, D. and López, 2010).
The Lyapunov exponents may be introduced in a similar way as for ordinary
differential equations. Say, for a system (10.2) the largest Lyapunov exponent is
obtained from simultaneous numerical solution of that equation along the reference
trajectory and the variation equation
x̃˙ = f1 (x(t), x(t − τ ))x̃(t) + f 2 (x(t), x(t − τ ))x̃(t − τ ). (10.6)
The perturbation vector at some instant t in the computations is represented as an

array of values (x̃ 0 , x̃1 , x̃2 , . . . , x̃M ), where x̃k = x̃(t − kh), h is a step of the finite-
difference scheme, and M = τ /h is a number of steps in the delay time interval.
Using the norm definition x = x20 + x21 + · · · + x2M one renormalizes the pertur-
bation vector to have a fixed norm and continues the procedure accumulating the
sums of logarithms of the norm ratios before and after the normalization. The rate
of growth of the accumulating sums determines the Lyapunov exponent.
Supplementing the procedure with additional equations for perturbation vec-
tors, and with Gram-Schmidt orthogonalization (using the dot product defined as
√
x · y = x0 y0 + x1 y1 + · · · + xM yM ) in straight analogy with the method described
in Appendix A, one can evaluate more than one Lyapunov exponents, but not the
complete spectrum (as it is infinite in the original equation although restricted in
the finite-difference numerical scheme). It is reasonable to agree that the number
of terms in the spectrum of Lyapunov exponents really deserving consideration and
catching essential features of the dynamics is just the number sufficient to get a
correct estimate of the dimension for the attractor from Kaplan-Yorke formula (Ap-
pendix D) (Farmer, 1982; Blokhina et al., 2007).
It is appropriate for the time-delayed systems to consider construction analo-
gous to the Poincaré maps in ordinary differential equations (Kuznetsov, 1982; Gi-
acomelli and Politi, 1996).
In this respect, the simplest is the case of non-autonomous systems represented
by equation
ẋ = f(t, x(t), x(t − τ )), (10.7)
1 For equations of neutral type the infinite set of roots has asymptotically equal real parts, and may
be placed either in right, or in a left half-plane depending on concrete equations and parameters. In
leading type case the placement is inversed in comparison with the retarding equation that implies
infinite number of unstable roots.
10.2 Van der Pol oscillator with delayed feedback, parameter modulation and . . . 223
with the right-hand parts periodic in time: f(t, x, y) ≡ f(t + T, x, y), where T is a
period. Then, the description may be formulated in terms of functional stroboscopic
map
Xn+1 (t) = F(Xn (t)). (10.8)
Here the functions X n (t) are defined in an interval of length τ and are related with
the original variable x(t) as X n (t) = x(t + nT ). Concrete examples of such systems
will be discussed in Sect. 10.2.
In autonomous systems one has to introduce some condition on the function x(t),
which is valid at once in each characteristic period of the dynamics, and introduce
a functional map, which generates a sequence of the states at successive instants of
validity of this condition. An example will be discussed in Sect. 10.3.

modulation and auxiliary signal
Consider a model system described by the equation of the following form (Kuznetsov
and Ponomarenko, 2008):
ẍ − (A cos(2π t/T ) − x2 )ẋ + ω02 x = ε x(t − τ )ẋ(t − τ ) cos ω0t. (10.9)
Here, x is a dynamic variable of van der Pol oscillator with the operating frequency
ω0 , in which the parameter controlling the bifurcation of limit cycle varies slowly
in time with period T and amplitude A, so that the oscillator alternately manifests
activation and damping. To the right-hand side of the equation a term is added re-
sponsible for the delayed feedback. It is a product of the dynamical variable delayed
by τ , its derivative, and the auxiliary signal at the frequency ω 0 . Parameter ε deter-
mines strength of the delayed feedback. We assume that N = ω 0 T /2π is an integer,
so that the external driving is periodic. A block-diagram of the system is shown in
Fig. 10.1.
Fig. 10.1 Block diagram of the system with delay.

10.2.1 Attractor of Smale-Williams type in the time-delayed system
Functioning of the system (10.9) as a generator of chaos can be explained as follows.

Due to the periodic variation of the parameter responsible for the emergence of
the generation, the oscillator alternately manifests stages of activity and damping.
3
Having made proper selection of the retarding time, for example τ = T , one can
4
ensure such a situation that the emergence of self-sustaining oscillations at each next
stage of activity is stimulated by a signal emitted at the previous activity epoch.
Suppose that it was characterized by some phase ϕ , i.e. x(t) ∼ sin(ω 0t + ϕ ) and
ẋ(t) ∼ cos(ω0t + ϕ ). Then the term in the right-hand part of the equation will contain
a component at the fundamental frequency ω 0 with the doubled phase shift constant.
Indeed,
x(t − τ )ẋ(t − τ ) cos ω0t ∼ sin 2(ω0 (t − τ ) + ϕ ) cos ω0t

1 (10.10)
= sin(ω0t − 2ω0τ + 2ϕ ) + · · · ,
2
where the dots designate a non-resonant term. Therefore, at a new stage of activity
the phase of the oscillator will be determined by the phase of the resonant com-
ponent of the stimulating signal, and at successive stages of activity we obtain the
Bernoulli map ϕn+1 = 2ϕn + const(mod 2π ), in the state space that corresponds to
expansion along the angular coordinate ϕ . If in other directions compression of the
phase volume takes place, such dynamics must correspond to the presence of at-
tractor of Smale-Williams type for the mapping of the phase space in one period of
external driving.
For the numerical solution the finite difference method of Runge-Kutta was used,
modified with respect to the system with time delay. The retarded values for vari-
able x required to perform steps of the difference scheme are extracted from the
stored array of data obtained in the previous steps of the integration. An interpola-
tion scheme was used providing the required order of accuracy. At the beginning
of the computations, the array of data is usually filled with random values; then the
long enough initial stage of calculations is performed for the system to reach the
attractor, and only after that processing procedures including plotting the data are
performed.
In Fig. 10.2 panel (a) shows the dynamical variable plotted versus time as ob-
tained from numerical solution of Eqs. (10.9) for oscillations corresponding to the
motion on the attractor. Chaos manifests itself in randomness of phases of the oscil-
lations relative to the envelope at successive stages of activity. Panel (b) is a diagram
for the phases identified at successive stages of activity; it demonstrates nice agree-
ment with the Bernoulli map, as expected from the qualitative consideration of the
operation of the scheme. Panels (b) and (c) show portraits of the attractor. One is
projection of the attractor from the extended (including the time axis) phase space
of the non-autonomous infinite-dimensional flow system, the other is a portrait of
the attractor in the stroboscopic Poincaré section. In the last diagram the transversal
Fig. 10.2 Waveform of the dynamical variable x according to the results of the numerical solution
3
of Eq. (10.9) at ω0 = 2π , T = 6, τ = T, A = 5.5, ε = 0.2 (a). Diagram illustrating the transforma-
4
tion of phases in the successive stages of the activity (b). Attractor of the system in the projection
from the infinite dimensional space of states on the phase plane of the oscillator (c) and a portrait
of the attractor in the Poincaré section (d).
fractal structure is visible, which is a characteristic feature of attractor of Smale-

Williams type.
To calculate the Lyapunov exponents the method is used based on the Benettin
algorithm with modifications corresponding to the class of delay systems (Farmer,
1982; Giacomelli and Politi, 1996; Le Berre et al., 1987; Lepri et al., 1994; Balyakin
and Ryskin, 2007; Pazó, D. and López, 2010). It exploits data of joint numerical
integration of Eq. (10.9) and a collection of m variation equations
x̃¨ − (A cos(2π t/T ) − x2 )x̃˙ + 2xẋx̃ + ω02 x̃

(10.11)
= ε [x̃(t − τ )ẋ(t − τ ) + x(t − τ )x̃˙(t − τ )] cos ω0t
to get m larger Lyapunov exponents. In numerical calculations, a perturbation vector

is given by the instantaneous values of x̃, x̃˙ and an array of data corresponding to the
values of the variable x̃ in the interval of delay.
Since the system is non-autonomous, it is convenient to organize the compu-
tations, considering the stroboscopic dynamics, with time intervals of duration T .
Upon completion of each stage, the perturbation vectors are orthogonalized by
means of the Gram-Schmidt process and normalized to a fixed constant. Lyapunov
exponents are obtained from the rate of growth of the cumulative sums of logarithms
of the ratios characterizing increase or decrease of the norms of the vectors.
At the chosen parameters the computations yield the following values normalized
as for the stroboscopic map (the time scale is the period of modulation T ):
Λ1 = 0.688, Λ2 = −0.837, Λ3 = −4.287 . . .. (10.12)
The only positive exponent of the attractor is close to ln2 = 0.693 . . ., which is con-
sistent with the approximate description of the dynamics in terms of Bernoulli map
for the phase variable. Estimate of the dimension with Kaplan-Yorke formula yields
for the attractor of the stroboscopic map D KY = 1 + Λ1 /|Λ2 | ≈ 1.82. Dimension of
the attractor in the extended phase space of the continuous time system is greater by
one (D ≈ 2.82).
It was discovered that in this system hard excitation occurs with hysteresis in a
quite wide range of parameters. To study the question, it is convenient to supple-
ment Eq. (10.9) with an additional parameter h, which regulates relative duration of
excitation and damping stages of the oscillator:
ẍ − (h + A cos(2π t/T ) − x2)ẋ + ω02 x = ε x(t − τ )ẋ(t − τ ) cos ω0t. (10.13)
The above-described mode of operation corresponds to excitation of each new

train of oscillations by the stimulating signals transmitted in the delayed feedback
loop. Its value is proportional to the squared amplitude of oscillations at the previous
stage of activity. If one sets the initial conditions with characteristic amplitude small
enough, the excitation does not occur because of negligible level of the squared am-
plitude, and the oscillations do not develop. Generation regime can be restored by
increasing parameter h, i.e. by increase of the part of the modulation period, during
which the oscillations are linearly unstable. With gradual growth of h, at some point
one observes an abrupt transition to the large-amplitude chaotic regime. Now, if one
starts to decrease h, it does not lead immediately to breakdown of the chaotic oscil-
lations since the excitation in the system is transferred from the previous stages of
activity to the following ones. Only at large enough negative values of h the oscilla-
tions disappear. The hysteresis phenomenon we have described is illustrated in Fig.
10.3. The insets show iteration diagrams for phases associated with certain places
on the hysteresis loop. In the course of computation, small random perturbation was
added at each step of the numeric integration on a level of order Δx ∼ 0.001 to
ensure departure of the system from the trivial stationary state at zero.
The mechanism of functioning of the model and numerical results, confirming its
implementation, suggest that the observed low-dimensional chaotic attractor is uni-
formly hyperbolic, namely, it is a suspension of attractor of Smale-Williams embed-
ded in an infinite state space of the system with delay. Despite hypothetical nature
of this conclusion, the considered scheme of the chaotic generator is interesting in
itself, because it is prominent for generation of chaotic regimes insensitive to varia-
tions of parameters, technical fluctuations and design details, e.g. in the systems of
radio-engineering and electronics, nonlinear optics, etc.
Fig. 10.3 A diagram illustrating the hard excitation and hysteresis in the system: dependence of the
maximum amplitude of the parameter h in (10.13) at ω0 = 2π , T = 24, τ = 12, A = 3, ε = 0.3. The
parameter h is varied, first, from left to right that corresponds to motion along the lower branch,
and then from right to left that corresponds to motion along the upper branch of the hysteresis loop.
10.2.2 Hyperchaotic attractors
An interesting question is the following. Accounting infinite dimensional nature

of the dynamics, can we obtain hyperchaos characterized by two or more positive
Lyapunov exponents? In the hyperbolic theory it corresponds to attractors with di-
mension of the unstable manifolds two or more. Remarkably, it may be done in the
above model without structural modification, simply by means of appropriate selec-
tion of parameters. Namely, one has to increase the ratio of the delay time to the
period of parameter modulation in such a way that the signal arriving through the
feedback loop would stimulate excitation not at the nearest next stage of activity, but
with elapse of one, two, or more activity stages (Fig. 10.4). The mode of operation
discussed in the previous section corresponds to delay time marked with the arrow
Fig. 10.4 A diagram illustrating selection of delay time for implementation of regimes with dif-
ferent numbers of positive Lyapunov exponents.
τ1 while larger delay times, which give rise to hyperchaotic regimes, are indicated
with the arrows τ2 , τ3 , τ4 .
Let us hereafter concretize the delay times as follows:

1
τ = k− T, (10.14)
2
where k is a positive integer. Then, with argumentation similar to that in the previous
section, we arrive at the mapping for phases of successive trains of oscillations
ϕn = 2ϕn−k + const. (10.15)
It reduces to the Bernoulli map in the case k = 1. At k 2 this relation must be

interpreted as a map on a k-dimensional torus: a vector v n−1 = (ϕn−1 , . . . , ϕn−k ) is
mapped to vector v n = (ϕn , . . . , ϕn−k+1 ). The map possesses a special degeneracy,
namely, the complete sequence of phase variables ϕ n is composed of k indepen-
dent subsequences in such a way that in the course of the time evolution their terms
alternate. For each subsequence, at k steps of discrete time small perturbations un-
dergo doubling, being clear from (10.15). Thus, the map (10.13) has k Lyapunov
exponents of equal value Λ i = k−1 ln 2, i = 1, 2, . . . , k.
Relating the original system (10.9) the map (10.15) is an approximation. Never-
theless, in application to the regimes of active operation (pulsations), in a wide range
of parameters, the conclusion concerning a number of positive Lyapunov exponents
must remain valid, although the consecutive terms of the sequence of the phases will
no longer be completely independent. It means that the degeneracy is lifted, and the
Lyapunov exponents will not be precisely equal. In the infinite-dimensional phase
space of the delay system, such dynamics will give rise to a stretch of phase volume
along k directions, associated with cyclic coordinates ϕ 1 , . . . , ϕk , and the remaining
directions will correspond to compression. It will be hyperchaotic attractor with k
positive and others negative Lyapunov exponents.
Figure 10.5 shows waveforms for the oscillating variable versus time plotted in
sustained chaotic regimes at different ratios of the delay time and the modulation
period selected in accordance with (10.14) for k = 1, 2, 3, 4. Thus, panel (a) corre-
sponds to situation of stimulation of excitation by the feedback signal just at the
next activity stage, and panels (b)—(d) correspond to stimulation through one, two,
and three elapsed intermediate activity stages.
As can be seen, in all cases the process looks like a sequence of oscillating trains
following each other with time interval T , and the carrier phase evolves from pulse
to pulse in a chaotic manner. To demonstrate that chaos indeed appears due to the
above-described mechanism of the phase transfer from the previous stages of ac-
tivity, in Fig. 10.6 the iteration diagrams are shown for the phases. Phases corre-
sponding to each next stage of activity are determined at certain instants fixed with
respect to the waveform of slow parameter modulation, according to the formula
ϕn = arg(x − iω0−1 ẋ). The same parameter values were assigned as those for compu-
tation of the waveforms in Fig. 10.5. The plots are shown in coordinates (ϕ n , ϕn−k ),
where k = 1, 2, 3, 4 for panels (a), (b), (c) and (d), respectively. Observe that the di-
Fig. 10.5 Waveforms for the oscillating variable versus time obtained from numerical simulation
for model (10.9) at ω0 = 2π , ε = 0.3, T = 24, A = 3, and k = 1, τ = 12 (a); k = 2, τ = 36 (b);
k = 3, τ = 60 (c); k = 4, τ = 84 (d).
agrams confirm efficiency of the description in terms of the maps (10.13). Indeed,
variation of the phase ϕ n−k implies roughly doubled variation of ϕ n . In Fig. 10.7 we
show diagrams for phases in coordinates (ϕ n , ϕn−1 ). The fact that the dots nearly
uniformly fill a square indicates closeness of the situation to statistical independence
of the subsequences. Slightly distinguishable non-uniform structures indicate that a
weak statistical dependence is still present, which is not accounted in the simplified
description by maps (10.15).
Computations support the conclusions concerning a number of positive Lya-
punov exponents in regimes at delay times selected in accordance with (10.15)
at different k. It is convenient to use Lyapunov exponents normalized to the pe-
riod kT : Λ̄i = kΛi = kT λi . In the framework of description by the map (10.15),
the first k Lyapunov exponents with this normalization should be equal precisely to
ln 2 ≈ 0.693.
Fig. 10.6 Iteration diagrams for phases illustrating applicability of the approximated description
by the maps (10.15) at different k; parameters are the same as in Fig. 10.5.
In computations, in each regime of operation of the system we evaluated a num-

ber of Lyapunov exponents for the stroboscopic mapping sufficient for estimate of
the attractor dimension from the Kaplan-Yorke formula D = m + S m /|Λ̄m+1 |, here
m
the integer m is selected in such a way that the sum S m = ∑ Λ̄i is positive, while
i=1
Sm+1 is already negative.
An insightful way of representing the spectrum of Lyapunov exponents, which
allows immediate distinguishing the type of hyperchaos regime, is to plot the sums
Sm against the number m. These graphs are shown in Fig. 10.8. On such a diagram all
essential quantifiers of chaos are presented: the value of the plotted function at unit
argument is the largest Lyapunov exponent Λ̄1 ; the argument value at the maximum
is the number of positive exponents. The maximum of the function corresponds to
the Kolmogorov-Sinai entropy according to the Pesin formula h̄ = ∑ Λ̄k . Intersec-
Λ̄k >0
tion of the broken line with the abscissa determines the Kaplan-York dimension of
the attractor. Observe that in all the cases positive Lyapunov exponents are close to
ln 2 ≈ 0.693 that agrees with the conclusion based on the approximate maps (10.15).
10.3 Van der Pol oscillator with two delayed feedback loops and parameter modulation 231
Fig. 10.7 Diagrams illustrating weak statistical dependence of phases of nearest neighboring ac-
tivity stages at k = 2, 3, and 4; parameters are the same as in Fig. 10.5.
Fig. 10.8 Plots of sums of Lyapunov exponents on the number of them with indicated values of the
largest exponent Λ̄1 , Kaplan-Yorke dimension D, and Kolmogorov-Sinai entropy h̄. The data are
obtained in computations for the model (10.9) at ω0 = 2π , ε = 0.3, T = 24, A = 3 for k = 1, τ = 12
(a); k = 2, τ = 36 (b); k = 3, τ = 60 (c); k = 4, τ = 84 (d).
parameter modulation
The system discussed in the previous section is non-autonomous because of two

kinds of time dependence of the coefficients. The first is slow variation of a param-
eter responsible for the birth of self-oscillation in the van der Pol oscillator, and the
second is the auxiliary signal of frequency ω 0 equal to the natural frequency of the
oscillator. Now, we intend to introduce a model, in which the auxiliary signal is not
needed; it makes the scheme simpler in practical implementation (Baranov et al.,
2010).
Consider the block diagram in Fig. 10.9. Again, the main element is the van der
Pol oscillator, in which the parameter controlling the Andronov-Hopf bifurcation, is
slowly modulated in time with period T , so that it alternately manifests the excita-
tion and damping. Now, however, the signal from the oscillator passes through two
loops of delayed feedback, in which the retarding time τ 1 and τ2 differ by T . In one
of the feedback circuits the signal undergoes quadratic nonlinear transformation and
differentiation, so, the low-frequency component is filtered out (see blocks marked
on the diagram, respectively x 2 and d/dt), and then it is mixed on a quadratic non-
linear element (marked on the diagram as skew cross) with the signal received from
the second feedback circuit. The resulting signal contains a component at the dif-
ference frequency resonant with the oscillator, and it stimulates the excitation at the
beginning of the next stage of activity. The delays are chosen in such a way that to
stimulate excitation at the n + 1-th stage of activity, the signals arrive at the right
time from the (n − 1)-th and n-th stages of activity, respectively, through the first
and the second feedback loops.
Fig. 10.9 Block diagram of the system with two delayed feedback loops.
A formal model for the device is a differential equation with two delays
ẍ − (A cos(2π t/T ) + h − x2)ẋ + ω02x = ε x(t − τ1 )x(t − τ2 )ẋ(t − τ2 ). (10.16)
Here, x is an oscillating variable of the van der Pol oscillator; ε is a parameter

characterizing depth of the delayed feedback, the value of A determines magnitude
of the modulation of parameter responsible for the excitation of the oscillator in
respect to the average level determined by a value of h. To be concrete, we select
1 3
the delay time constants as τ 1 = T and τ2 = T .
2 2
Let us assume for a while that h =0. Then, the oscillator will be alternately in
active and passive stages of equal duration, respectively, while A cos(2π t/T ) > 0 or
A cos(2π t/T ) < 0. Suppose the oscillations at the activity stages numbered with n
and n − 1 were ϕ n and ϕn−1 , respectively, i.e. x ∼ sin(ω 0 t + ϕn ) and x ∼ sin(ω0t +
ϕn−1 ). Then, substitution in the right-hand part of (10.16) yields
1 3 3
x(t − T )x(t − T )ẋ(t − T )
2 2 2

1 3 3
∼ sin ω0t − ω0 T + ϕn sin ω0t − ω0t + ϕn−1 cos ω0t − ω0 T + ϕn−1
2 2 2

1 1
= sin ω0t − ω0 T + ϕn sin(2ω0t − 3ω0T + 2ϕn−1 )
2 2

1 5
= − cos ω0t − ω0 T + 2ϕn−1 − ϕn + · · · ,
4 2
where three dots designate a non-resonance term (the third harmonic). Hence, the
phase of oscillations at the next activity stage will be expressed as
ϕn+1 = −ϕn + 2ϕn−1 + const ( mod 2π ). (10.17)
In the following, for the phase difference Δϕ n = ϕn − ϕn−1 the expanding circle
map, or Bernoulli map, holds:
Δϕn+1 = −2Δϕn ( mod 2π ). (10.18)
The last map is chaotic, with the Lyapunov exponent Λ 1 = ln 2 ≈ 0.693. The map
(10.17) has the same value of the largest Lyapunov exponent, and additionally, the
zero value exponent Λ 2 = 0.
Fig. 10.10 Waveform for the oscillating variable obtained in computations for the model (10.16)
1 3
at ω0 = 2π , T = 8, A = 4, h = 0, ε = 0.05, τ1 = T and τ2 = T .
2 2
Calculations show that the expected type of chaotic dynamics occurs in a quite
wide range of parameters. For detailed analysis, the following parameter values are
selected: ω0 = 2π , T = 8, A = 4, h = 0, ε = 0.05. Figure 10.10 shows the general-
ized coordinate of the oscillator versus time for these parameters. The process is a
sequence of trains of oscillations following each other after a time interval T . How-
ever, the filling phase of the carrier varies from pulse to pulse in a chaotic manner.
The presence of chaos is caused by the above-described mechanism of the transfer
of phase from the previous stages of the process. Figure 10.11 (a) shows the iteration
diagram for phases in the form of three-dimensional graphics, which demonstrates
approximate correspondence of the respective function of two variables and expres-

sion (10.17). Figure 10.11 (b) shows the iteration diagram for the phase differences
between neighboring activity stages, which is observed to be in good agreement
with relation (10.18).
Fig. 10.11 Diagrams illustrating the transformation of phases in the successive stages of the activ-
ity of oscillator (a) and for phase differences (b) in model (10.16) at ω0 = 2π , T = 8, A = 4, h =
1 3
0, ε = 0.05, τ1 = T and τ2 = T .
2 2
Figure 10.12 shows in panel (a) a portrait of a trajectory on the attractor plotted
on the phase plane of the van der Pol oscillator. In fact, the attractor of the system
with delay is an object in an infinite-dimensional state space; so the picture should be
viewed as a two-dimensional projection of the trajectory. Panel (b) shows a portrait
of the attractor in the stroboscopic cross-section.
Fig. 10.12 Two-dimensional projection of a trajectory on the attractor (a) and portrait of the
attractor in stroboscopic cross-section (b) obtained in the numerical solution of Eq. (10.16) at
ω0 = 2π , T = 8, A = 4, h = 0, ε = 0.05.
To calculate the Lyapunov exponents the method was used similar to that dis-
cussed in the previous section. It exploits data of joint numerical integration of Eq.
(10.16) and a collection of the variation equations for the perturbation vectors
x̃¨ + 2xx̃ẋ − (A cos(2π t/T ) + h − x 2)x̃˙ + ω02 x

1 3 3 1 3 3
= ε x̃ t − T x t − T ẋ t − T + ε x t − T x̃ t − T ẋ t − T
2 2 2 2 2 2

1 3 3
+ε x t − T x t − T x̃˙ t − T .
2 2 2
(10.19)
Figure 10.13 shows plots for the largest three Lyapunov exponents versus the
parameter of modulation depth A at fixed other parameters. It should be noted that
the largest two exponents in a wide range remains almost constant and close to
the values ln2 and 0, respectively, in agreement with analyses based on the map
(10.17) for transformation of phases. Particularly, at A = 4 the first five exponents are
found Λ1 = 0.693, Λ2 = 0.000, Λ3 = −1.140, Λ4 = −1.333, Λ5 = −3.807. Hence,
the Kaplan-Yorke dimension for attractor in stroboscopic cross-section is D KY = 2+
Λ1 /|Λ3 | ≈ 2.61, and for attractor embedded in the extended state space is D ≈ 3.61.
Fig. 10.13 Dependence of the first four of the Lyapunov exponents on parameter modulation A
for stroboscopic map of the model (10.16) at ω0 = 2π , T = 8, ε = 0.05, h = 0. The computations
were performed with gradual decrease of A, with the inheritance of the instantaneous states of the
system at each step of the parameter variation.
Figure 10.14 shows power spectrum of the oscillator in the present mode of gen-
eration of chaos. For its construction a long enough time series was processed sam-
pled from data of numerical simulation of the model (10.16) for motion on the at-
tractor. The spectrum is computed according to the method of statistical evaluation
of the spectral power density known in the theory of random processes (Jenkins
and Watts, 1968). The procedure is to break the time series into pieces of certain
length, performing the Fourier transform for each segment, and then averaging the
squares of the amplitudes of spectral components. As can be seen from the fig-
ure, the spectrum is concentrated in a certain band near the frequency, f ≈ 1, i.e.
ω = 2π f ≈ ω0 = 2π . Note the presence of continuous part of the spectrum that
is characteristic for chaotic dynamics. Discrete peaks in the left part of the diagram
appear due to the periodic component present in the dynamics of the system because
of the slow modulation of the excitation parameter.
Fig. 10.14 Power spectrum of the signal generated by the system at ω0 = 2π , T = 8, A = 4, ε =

0.05, h = 0 obtained by processing the numerical results of the model (10.16).
Fig. 10.15 A diagram illustrating the soft-hard excitation and hysteresis in the system: dependence
of the maximum amplitude of the parameter h in (10.16) at ω0 = 2π , T = 8, A = 3.9, ε = 0.05.
It was observed that the system manifests soft-hard excitation and hysteresis.
This phenomenon is illustrated in Fig. 10.15. The amplitude of oscillations is plotted
10.4 Autonomous time-delay system 237
against parameter h, which is varied, first, from the left to the right that corresponds
to motion along the lower branch, and then from the right to the left that corresponds
to motion along the upper branch of the hysteresis loop. Insets in the top part of
the picture show the iteration diagrams for phases, corresponding to representative
points on the hysteresis loop. When performing calculations in the system a small
noise of level 0.001 was added to ensure departure of the system from the trivial
zero state. The mechanism for the phenomenon is similar to that discussed in the
previous section.
As one can guess, this system should be related to a class of systems with par-
tially hyperbolic attractors (Pesin, 2004; Bonatti et al., 2005; Pesin, 2007) due to
the presence of a neutral direction at points of the attractor associated with the near-
zero Lyapunov exponent. A rigorous justification of this conjecture seems to be
rather complex mathematical problem because of infinite-dimensional nature of the
system.
It is worth mentioning that another version of a scheme of generator of chaos
based on van der Pol oscillator with two delayed feedback loops was advanced and
analyzed in (Kuznetsov and Pikovsky, 2008). It differs from (10.16) with the de-
gree of the nonlinear transformations of signals in the feedback circuits; in model
(Kuznetsov and Pikovsky, 2008) these transformations are cubic and quadratic,
while those in the model (10.16) are linear and quadratic, respectively.
10.4 Autonomous time-delay system
To construct an autonomous model with the mechanism of generation of chaos sim-

ilar to that in the previous sections, but without external driving, an appropriate
starting point is the logistic delay equation. It was offered in due time in the context
of population biology (Fowler, 1982) and reads
ṙ = μ (1 − r(t − τ ))r(t). (10.20)
Here r is positive population variable normalized in such a way that the satura-
tion occurs at r = 1; μ > 0 is parameter of the birth rate at small populations; τ
is parameter of the delay time characterizing lag of the effect of saturation. Under
condition τ < π /2μ the system has a stable stationary state r = 1. For τ > π /2μ
self-oscillations of the population arise. At large values of delay τ they have a form
of periodic sequence of pulses (Fig. 10.16). The period grows with parameter τ as
P∼ = (1 + e μτ )/ μ , and minimal level of the population reached in between the pulses
is estimated as rmin ∼ = μτ exp(−e μτ + 2μτ − 1), i.e. manifests the double exponen-
tially decrease with increase of τ (Fowler, 1982).
Now, we regard the positive variable r as squared amplitude of some oscillatory
process with frequency ω 0 . For this, we set r = x2 + y2 and require the new variables
to satisfy the equations
Fig. 10.16 Self-pulsations in the logistic-delay equation (10.20) at μ = 1.6, τ = 2; the observed
period of pulsations is Tav ≈ 16.0.
1
ẋ = −ω0 y + μ (1 − x2(t − τ ) − y2(t − τ ))x,
2
(10.21)
1
ẏ = ω0 x + μ (1 − x2(t − τ ) − y2(t − τ ))y.
2
Next, we add in the first equation an additional term ε x(t − τ )y(t − τ ), where ε is
a small parameter, and arrive at the set of equations (Kuznetsov and Pikovsky, 2010)
1
ẋ = −ω0 y + μ (1 − x2 (t − τ ) − y2(t − τ ))x + ε x(t − τ )y(t − τ ),
2
(10.22)
1
ẏ = ω0 x + μ (1 − x2(t − τ ) − y2(t − τ ))y.
2
In the case of generation of pulses with extremely low level of minimal amplitude
between them, just the additional term will initiate formation of the next pulse of
oscillations in the system. Due to this, the period of pulses becomes less than that
in system (10.20). If we assume that during a current pulse the variables behave as
x ≈ f (t) cos(ω0t + ϕ ), y ≈ f (t) sin(ω0t + ϕ ), the additional term is expressed as
1
ε x(t − τ )y(t − τ ) ≈ ε f 2 (t − τ ) sin(2ω0 (t − τ ) + 2ϕ ). (10.23)
2
It represents a stimulating signal, which initiates oscillations corresponding to
the next stage of activity and the phase shift of this signal is transferred to these
oscillations. Hence, the phase for the train of oscillations is transformed from the
previous train with doubling:
ϕn+1 = 2ϕn + const. (10.24)
It corresponds to expanding circle map, or the Bernoulli map, which is chaotic

and characterizing by the Lyapunov exponent Λ = ln 2 ≈ 0.693.
As it is common in time-delayed systems, the phase space for the model (10.22)
is infinite-dimensional. Indeed, to identify an instant state of the system at some t =
t0 and determine uniquely further evolution in time one has to specify functions x(t)
and y(t) in time interval t ∈ [t 0 − τ ,t0 ]; it corresponds to a point in the phase space.
Attractor will be an object embedded in this infinite-dimensional space. One can
introduce a mapping of the state space into itself that corresponds to transformation
from one oscillation train to the next one. It will be the infinite-dimensional Poincaré
10.4 Autonomous time-delay system 239
map for the system. Dynamics determined by this map is of such nature that on the
attractor there is expansion along a direction associated with the phase variable ϕ ,
and compression in other directions. Accounting Bernoulli type of the map for the
phase, it must be attractor of Smale-Williams type.
Figure 10.17 shows the time dependence for dynamical variables which illus-
trate operation of the system (10.22) in accordance with the above qualitative con-
siderations. Indeed, the process looks like a sequence of pulses, trains of oscilla-
tions. An average period of their appearance is less than that in model (10.20) and
equals T ∼ = 11.657 in the regime under discussion. Portrait of the attractor pro-
jected from the infinite-dimensional
phase space is presented in Fig. 10.18 in coor-
dinates (x, y, ρτ ), where ρτ = x2 (t − τ ) + y2(t − τ ). It looks similar to analogous
representation of attractor in ordinary differential equations constructed on a base
of predator-prey model in Sect. 5.2 (Fig. 5.6(b)).
Fig. 10.17 Waveforms for the dynamical variables x (black) and y (gray) according to the results of
the numerical solution of Eq. (10.22) at ω0 = 2π , μ = 1.6, ω0 = 2π , τ = 2, ε = 0.05; the observed
period of pulsations is Tav ≈ 11.657.
Fig. 10.18 Portrait of the attractor projected from the infinite-dimensional phase space of the
model (10.22) at ω0 = 2π , μ = 1.6, ω0 = 2π , τ = 2, ε = 0.05.
Three larger Lyapunov exponents evaluated numerically for the system (10.22)
are
λ1 = 0.0544, λ 2 = 0.0005, λ 3 = −1.982. (10.25)
The largest exponent λ 1 is positive and corresponds with good accuracy to that ob-
tained from the approximation based on the Bernoulli map. Indeed, accounting that
the characteristic period of pulses is about T ∼ = 11.657, for the stroboscopic map
the respective exponent is Λ 1 = 0.646, close to the expected value ln 2. The sec-
ond exponent is close to zero up to numerical precision; accounting the autonomous
nature of the system it is interpreted as being associated with perturbations of in-
finitesimal shift along the phase trajectory. Other exponents are negative. Estimate
of the attractor dimension for the Poincaré map from the Kaplan-Yorke formula
yields DKY ≈ 2.027 for the flow system.
Figure 10.19 demonstrates that transformation of phases of the oscillations from
pulse to pulse follows the Bernoulli map. A phase associated with certain pulse is
determined at each instant of maximal squared amplitude x 2 + y2 from the relation
φ = arg(x − iy), and the data are plotted in coordinates (φ n , φn+1 ).
Fig. 10.19 Diagram illustrating the transformation of phases in the successive stages of the activity
plotted according to the results of the numerical solution of Eq. (10.22) at ω0 = 2π , μ = 1.6, ω0 =
2π , τ = 2, ε = 0.05.
References
Balyakin, A.A., Ryskin, N.M.: Features of computation of spectrum of Lyapunov exponents in

distributed self-oscillatory systems with delayed feedback. Izvestija VUZov-Applied Nonlinear
Dynamics (Saratov) 15, 3–21 (2007).
Baranov, S.V., Kuznetsov, S.P., Ponomarenko, V.I.: Chaos in the phase dynamics of Q-switched van
der Pol oscillator with additional delayed-feedback loop. Izvestija VUZov-Applied Nonlinear
Dynamics (Saratov) 18 (1), 11–23 (2010) .
Bellman, R., Cooke, C.L.: Differential-difference Equations. Academic Press, New York (1963).
Blokhina, E.V., Kuznetsov, S.P., Rozhnev, A.G.: High-dimensional chaos in a gyrotron. IEEE
Transactions on Electron Devices 54, 188–193 (2007).
References 241
Chiasson, J.N., Loiseaum, J.J. (eds.): Applications of Time Delay Systems (Lecture Notes in Con-
trol and Information Sciences, Vol. 352). Springer, Heidelberg (2007).
El’sgol’ts, L.E., Norkin, S.B.: Introduction to the Theory and Application of Differential Equations
with Deviating Arguments. Academic Press, New York (1973).
Farmer, J.D.: Chaotic attractors of an infinite-dimensional dynamical system. Physica D 4, 366–
393 (1982).
Fowler, A.C.: An asymptotic analysis of the delayed logistic equation when the delay is large. IMA
Journal of Applied Mathematics 28, 41–49 (1982).
Giacomelli, G., Politi, A.: Relationship between delayed and spatially extended dynamical sys-
tems. Phys. Rev. Lett. 76, 2686–2689 (1996).
Glass, L., Mackey, M.C.: From Clocks to Chaos: The Rhythms of Life. Princeton University Press,
Princeton (1988).
Hu, H.Y., Wang, Z.H.: Dynamics of Controlled Mechanical Systems with Delayed Feedback.
Springer, Berlin (2002).
Jenkins, G.M., Watts, D.G.: Spectral Analysis and its Application. Holden-Day, Inc., San Francisco
(1968).
Kuznetsov, S.P.: Complex dynamics of oscillators with delayed feedback. Radiophysics & Quan-
tum Electronics 25, 996–1009 (1982).
Kuznetsov, S.P., Pikovsky, A.: Hyperbolic chaos in the phase dynamics of a Q-switched oscillator
with delayed nonlinear feedbacks. Europhysics Letters 84, 10013 (2008).
Kuznetsov, S.P., Pikovsky, A.: Attractor of Smale-Williams type in an autonomous time-delay
system. Preprint nlin. arXiv: 1011.5972 (2010).
Kuznetsov, S.P., Ponomarenko, V.I.: Realization of a strange attractor of the Smale–Williams type
in a radiotechnical delay-fedback oscillator. Tech. Phys. Lett. 34, 771–773 (2008).
Le Berre, M., Ressayre, E., Tallet, A., Gibbs, H.M., Kaplan, D.L., Rose M.H.: Conjecture on
the dimensions of chaotic attractors of delayed-feedback dynamical systems.Phys. Rev. A 35,
4020–4022 (1987) .
Lepri, S., Giacomelli, G., Politi, A., Arecchi, F.T.: High-dimensional chaos in delayed dynamical
systems. Physica D 70, 235–249 (1994).
Myskis, A.D.: Linear differential equations with retarded argument, Nauka, Moscow (1972).
Pazó, D., López, J.M.: Characteristic Lyapunov vectors in chaotic time-delayed systems. Phys.
Rev. E 82, 056201 (2010).
Pesin, Ya.B.: Lectures on Partial Hyperbolicity and Stable Ergodicity. Zurich lectures in advanced
mathematics. European Mathematical Society (2004).
Pesin, Ya.: Existence and genericity problems for dynamical systems with nonzero Lyapunov
exponents. Regular and Chaotic Dynamics 12, 476–489 (2007).
Vallee R., Delisle, C., Chrostowski J.: Noise versus chavs in acousto-optic bistability. Physical
Review A 30, 336–342 (1984).
Chapter 11
Chaos in Co-operative Dynamics of Alternately
Synchronized Ensembles of Globally Coupled
Self-oscillators
Abstract In many cases dynamics of spatially extended systems can be thought of as

cooperative action of a large number of relatively simple elements, like interacting
limit-cycle oscillators, and treated as low-dimensional dynamics in terms of appro-
priate collective modes (Cross and Hohenberg, 1993). In this context, a paradigm
concept is the Kuramoto model of ensemble of globally coupled phase oscilla-
tors distributed in some range of the natural frequencies, and the synchronization-
desynchronization transition in such ensemble (Kuramoto, 1984; Strogatz, 2000;
Pikovsky et al., 2002; Acebrón, 2005). Now, an interesting question arises about
a possibility of occurrence of hyperbolic attractors on the level of the collective
modes. This problem may be of interest in various fields, e.g. in hydrodynamics, in
laser physics and nonlinear optics, electronics, neurodynamics. In this chapter we
start from a short introduction to dynamics of a globally coupled ensemble of limit-
cycle oscillators and discuss different levels of reduction, including slow complex
amplitude equations and phase equations; the last represents exactly the Kuramoto
model. Then, we turn to some artificial, but feasible and analytically convenient
model of two alternately synchronized and desynchronized interacting ensembles
of globally coupled oscillators, which demonstrate transformation of the phase of
collective excitation in the course of dynamic evolution in accordance with the ex-
panding circle map at successive stages of synchronization of the subsystems. The
consideration is based on a joint work with Pikovsky and Rosenblum (Kuznetsov et
al., 2010).
11.1 Kuramoto transition in ensemble of globally coupled

oscillators
Consider an ensemble of van der Pol oscillators characterized by some distribution

over their natural frequencies and interacting through a common field arising due to
the global dissipative coupling of all the oscillators. A set of governing equations
for the model reads:
244 11 Chaos in Co-operative Dynamics of Alternately Synchronized Ensembles of . . .
ẍω − (A − x2ω )ẋω + ω 2 xω = κ < ẋω >, (11.1)
where κ is the coupling parameter, A is the parameter responsible $ for the birth of
the limit cycle in a single oscillator, and the notion < ẋ ω >= ẋω f (ω )d ω means
averaging of the variable ẋ ω over the ensemble. The distribution of natural frequen-
cies of the members of the ensemble over some frequency
$ range is characterized by
a function f (ω ), normalized in such a way that f (ω )d ω = 1. The distribution is
unimodal; it means that the function has a unique maximum and decays monotoni-
cally on both sides of this point with deflection of the frequency.
Supposing that parameters A and κ are relatively small, the central frequency ω 0
is large enough, and the range of natural frequencies is relatively narrow, one can
reformulate the problem using the method of slow amplitudes. To do so, introduce
the complex amplitude a ω for each oscillator in such a way that the coordinate and
velocity variables are expressed as
xω = aω eiω0t + a∗ω e−iω0t and ẋω = iω0 aω eiω0t − iω0 a∗ω e−iω0t . (11.2)
It implies that the additional condition ȧ ω eiω0t + ȧ∗ω e−iω0t = 0 is assumed. Then,
substitution in Eq. (11.1) yields
2iω0 ȧω eiω0t + (ω 2 − ω02)(aω eiω0t + a∗ω e−iω0t )

= (A − a2ω e2iω0t − 2|aω |2 − a∗2
ωe
−2iω0t )(iω a eiω0 t − iω a∗ e−iω0 t ) + κ < ẋ > .
0 ω 0 ω ω
(11.3)
Now, we multiply the relation by e −iω0t and carry out the averaging over a period of
fast oscillations. It is easily to find out that < ẋ ω > e−iω0t = iω0 < a >, where the
overline designates the time averaging. Instead of the frequency of the oscillators it
is convenient hereafter to introduce and use the quantity characterizing the detuning
Ω = (ω 2 − ω02)/2ω0 ≈ ω − ω0 . Then, we get
1 1
ȧΩ = iΩaΩ + (A − |aΩ|2 )aΩ + κ < a >, (11.4)
2 2
$
where < a >= aΩ f˜(Ω)dΩ. The distribution function for the detuning values is
expressed $in obvious way as f˜(Ω) ≈ f (ω0 + Ω), and satisfies the normalization
condition f (Ω)dΩ = 1.
One more step in simplifying the model is reduction to consideration of ensemble
of phase oscillators. The individual oscillators are assumed to have constant ampli-
tude corresponding to their limit cycles throughout the process, while the dynamics
manifests itself only in evolution of their phases. In the classical theory of synchro-
nization, this corresponds to description in terms of the so-called Adler equation
(Pikovsky et al., 2002; Osipov et al., 2007). It is justified if the limit cycle is well
pronounced at the selected value of the parameter A, distant from the threshold of
birth of the limit cycle although it is not too large: the self-oscillations are quasi-
harmonic. To reduce the description to the phase equations, we substitute in (11.4)
aΩ ≈ ReiΦΩ , (11.5)
11.1 Kuramoto transition in ensemble of globally coupled oscillators 245
√
where R ≈ A being constant for all oscillators. After multiplying the equation by
e−iΦΩ and separating the imaginary part, we finally obtain
1
Φ̇Ω = Ω + κ Im(< eiΦ > e−iΦΩ ). (11.6)
2
$
Here < eiΦ >= eiΦΩ f˜(Ω)dΩ, where f˜(Ω) is the same distribution function of
detuning, which had appeared in the model (11.4). Equation (11.6) corresponds to
the Kuramoto model, which is extensively examined in (Kuramoto, 1984; Strogatz,
2000; Pikovsky et al., 2002; Acebrón, 2005).
The most interesting feature of the Kuramoto model is the synchronization-
desynchronization transition, which occurs at some critical value of parameter κ c as
a collective effect in the ensemble of globally coupled oscillators. The coupling may
be regarded as a mean field affecting all oscillators. If the coupling constant is above
some threshold, this mean field synchronizes some macroscopically significant part
of the ensemble, which generates this mean field in self-consistent manner. If the
coupling is below the threshold, the ensemble desynchronizes due to the intrinsic
frequency difference of the oscillators constituting the ensemble, and the mean field
vanishes. As found analytically by Kuramoto, the transition occurs at κ c = 4/π f˜(0).
As the model (11.6) is obtained as approximation for the systems (11.1) and
(11.4), one can conclude that the same kind of transition occurs in those models too,
although the exact position of the threshold on the parameter axis may be slightly
distinct from that evaluated for the Kuramoto model (11.6).
To be concrete, hereafter we assume that the natural frequencies of the oscillators
are spread in a finite range (ω min , ωmax ) in accordance with the distribution function
⎧
⎪ π π (ω − ωmin)
⎨ cos , ω ∈ [ωmin , ωmax ],
f (ω ) = 2 ωmax − ωmin (11.7)
⎪
⎩
0, otherwise.
The plot of this function is shown in Fig. 11.1. This distribution can be obtained
with a help of an auxiliary variable ξ spreading on the unit interval uniformly; one
has to set simply ω = ωmin + (ωmax − ωmin) arccos ξ . For the versions of the model
based on slow-amplitude and phase approximations, we set, respectively, f˜(Ω) =
f (ω0 + Ω).
Fig. 11.1 A plot of the distribution function for the oscillator frequencies in the ensemble.
Equations (11.1), (11.4), or (11.6) correspond to asymptotic of infinitely large

number of members of the ensemble called the thermodynamic limit. To organize
numerical simulations, we assume a finite (though large enough) number of ensem-
ble members K 1. Their natural frequencies ω k are selected to correspond to the
prescribed distribution function f (ω ) in accordance with the relation
(ωmax − ωmin )
ωk = ωmin + arccos((2k − 1)/K − 1), k = 1, 2, . . ., K. (11.8)
π
The averaging operation over the ensemble for any function g(x, ẋ) is performed
now in accordance with the relation
1 K
< g(x, ẋ) >= ∑ g(xk , ẋk ).
K k=1
(11.9)
In computations the central frequency is specified as ω 0 = 2π , so, the unit time

corresponds to period of the fast oscillations. Interval of the frequency distribution
1
is of width Δω = ωmax − ωmin = π .
4
Figure 11.2 illustrates the synchronization-desynchronization transition. Time
dependence is plotted for the variables characterizing the mean field versus time
for three versions of the model for ensembles of K = 1000 units.
Panel (a) relates to the original ensemble of van der Pol oscillators with global
dissipative coupling (11.1) at A = 3 with frequencies distributed in the range
(2π − π /8, 2π + π /8) in accordance with the function (11.7) obtained from the
formula (11.8). Initial conditions are assigned with randomized phases of all os-
cillators, while their amplitudes correspond to the limit cycle of a single oscillator.
The coupling parameter initially equals a constant κ = 1, which is above the crit-
ical value. Thus, one can observe development of the collective synchronization
accompanied by arising oscillations of the mean field < ẋ > of significant magni-
tude. At the middle of the time interval shown on the diagram, the coupling constant
is switched to a sub-critical value κ = 0.5; this event starts the process of gradual
destruction of the synchronized state due to the frequency detuning intrinsic to the
members of the ensemble.
Panels (b) and (c) correspond, respectively, to description of the phenomenon
in terms of slow complex amplitudes with Eqs. (11.4) at A=3, and in terms of the
ensemble of phase oscillators (11.6). The frequency detuning parameters for the
oscillators are obtained from the formula
Ωk = −(π /8) + (1/4) arccos((2k − 1)/K − 1), k = 1, 2, . . ., K. (11.10)
It corresponds to the distribution of frequencies in accordance to (11.7) with Δω =

π /4. The initial conditions correspond to randomized phases of oscillators, and the
coupling parameter is switched from the supercritical to the subcritical values just
in the same way as in the panel (a). Again one can observe development of collec-
tive synchronization accompanied by arising significant amplitude of the mean field
11.2 Model of two alternately synchronized ensembles of oscillators 247
Fig. 11.2 Time dependence for the variables characterizing the mean field versus time for three
versions of the model of K = 1000 globally coupled oscillators with frequencies distributed in
the range (2π − π /8, 2π + π /8). Panel (a) corresponds to the original system (11.1), (b) to the
slow-amplitude version (11.4), and (c) to the system of phase oscillators (11.6). Initial conditions
relate to states of randomized phases of all oscillators and their amplitudes corresponding to the
limit cycle of a single oscillator. The coupling parameter on the first half of the time interval is
supercritical, κ = 1, and in the middle it is switched to a subcritical value κ = 0.5. Parameter value
used in the computations for panels (a) and (b): A = 3.
represented by the averages | < a > | and | < e iΦ > |, respectively, and desynchro-
nization after the switch. Geometrically, the scales for the mean-field variables on
all three panels are selected to make them visually comparable; so, one can judge
about degree of correspondence between the three levels of the description.
11.2 Model of two alternately synchronized ensembles of

oscillators
Let us turn to the main part of the present chapter aiming to give evidence of possi-
bility of dynamics associated with hyperbolic attractor of Smale-Williams in coop-
erative behavior of ensemble of oscillators (Kuznetsov et al., 2010).
11.2.1 Collective chaos in ensemble of van der Pol oscillators
In Chap. 4 it was shown how to implement the dynamics described by the Bernoulli
map in low-dimensional system of alternately excited self-oscillators. Chaotic na-
ture of the dynamics reveals itself in chaotic evolution of the phases of oscillations
generated at successive stages of excitation of the oscillators.
Now, consider a system composed of two ensembles of self-oscillators, similar
to each other with natural frequencies distributed in some range, like assumed in
the previous section. Oscillator coordinate variables for members of one and the
second sub-ensemble will be denoted x and y, respectively. To get immediately the
model appropriate for numerical simulation, we formulate equations for subsystems
composed of a finite (although large) number of oscillators K:
d 2 xk 2) dxk + ω 2 x = κ f (t) d < x > + ε d < y > sin ω t,

2
− (A − x k k ω 1 0
dt 2 dt dt dt
d 2 yk 2) dyk + ω 2 y = κ f (t) d < y > + ε d < x > sin ω t,
2 (11.11)
− (A − y k k k 2 0
dt 2 dt dt dt
k = 1, 2, . . . , K.
The averaged quantities in the right-hand parts of the equations are defined as
1 K 1 K 2 1 K 1 K 2
< x >= ∑ xk ,
K k=1
< x2 >= ∑ xk ,
K k=1
< y >= ∑ yk ,
K k=1
< y2 >= ∑ yk .
K k=1
(11.12)
Parameter A responsible for the birth of the limit cycle is of one and the same value
for all members of both sub-ensembles. Parameter κ is the coupling constant as-
sociated with global dissipative coupling within each sub-ensemble. The strength
of this coupling is modulated in time with period T 2π /ω 0 in counter-phase
in two sub-ensembles; this process is accounted in the equations by the functions
f1 (t) = sin2 (π t/T − π /4) and f 2 (t) = sin2 (π t/T + π /4). Interaction between the
sub-ensembles is introduced by means of a field, obtained by averaged quadratic
terms. The field responsible for the interaction oscillates on a doubled mean fre-
quency, but, due to multiplication of the respective terms by the auxiliary signal of
frequency ω 0 , it provides resonance effect onto each sub-ensemble from the partner.
The ratio N = ω0 T /2π is assumed to be an integer. The distributions of the oscilla-
tors in both ensembles over frequencies are identical, characterized by the function
(11.7); it is so if the frequencies ω k are assigned in accordance with relations (11.8).
Due to presence of the global coupling, each sub-ensemble can undergo the Ku-
ramoto transition: as oscillators synchronize, the collective field emerges with no-
table amplitude and definite phase of oscillations. As the coupling is turned off, the
oscillators desynchronize because of the frequency detuning of individual elements,
and the collective field disappears. The operation of the whole system consists of
alternating activities of two ensembles with the corresponding alternating growth
and decay of their collective mean fields.
Usually, as a single ensemble of oscillators passes through the Kuramoto transi-

tion from a non-synchronous state to a synchronous one while the global coupling
increases, the phase of the arising collective mode, potentially, of arbitrary value, is
determined by fluctuations. In our setup, however, the excitation of collective field
in one sub-ensemble is stimulated by a small driving force (proportional to ε ) be-
cause of action of the partner sub-ensemble; the last is synchronous at that moment
and generates notable mean field. This stimulation determines the phase on the ap-
pearing collective mode, which accepts this externally designated phase. This is the
mechanism of the phase transfer. Because the mutual couplings in our model are
proportional to the second-order mean fields of doubled characteristic frequency,
the phase transfer is accompanied with doubling of the phase. To explain this, let us
assume that at the transfer of excitation from the first to the second sub-ensembles
we have d < x > /dt ∼ cos(ω 0t + Φ) and d < x2 > /dt ∼ cos(2(ω0t + Φ) + const).
Then, the driving force affecting the second sub-ensemble contains the resonance
component of (d < x 2 > /dt) sin ω0t ∼ cos(ω0t + 2Φ + const). Respectively, the
arising means field will be of the form < y >∼ cos(ω 0t + 2Φ + const). As dou-
bling of the phase occurs at each transfer from one subsystem to another and back
through the full cycle, we expect that the phase is multiplied by the factor of 4 (up
to an additive constant).
In computations the central frequency is ω 0 = 2π , so, the unit time corresponds
to the period of the fast oscillations. The range of frequency distribution is of width
1
Δω = ωmax − ωmin = π . Number of elements in each of two ensembles is K =
4
1000. Other parameters are T = 100, κ = 1, A = 3, ε = 0.025. Initial distribution of
phases of the oscillators is uniform in the interval [0, 2π ) obtained with a standard
procedure of pseudo-random numbers generation. Initially, sufficiently long-time
interval of transient behavior is excluded from consideration to ensure that we deal
with sustained process of the collective dynamics of the two interacting ensembles.
Figure 11.3 shows the graphs of the averaged generalized velocity over each
of the ensembles versus time. The presence of intense oscillations of this variable
indicates synchronization of the ensemble of oscillators. As seen from the figure,
the epochs of intense oscillations for one and the other subsystem alternate with
Fig. 11.3 Graphs for the generalized velocities of the oscillators averaged over the first and the
second interacting ensembles, obtained from numerical solution of Eqs. (11.11). The number of
oscillators in each ensemble K = 1000, the distribution function of frequencies corresponds to the
formula (11.7), at ω0 = 2π , ωmin /ω0 = 15/16, ωmax /ω0 = 17/16, the values of other parameters
T = 100, κ = 1, A = 3, ε = 0.025.
the period T . Nevertheless, the dynamics in time is chaotic since the carrier phase
relating to the envelope varies from one epoch of the ensemble synchronization to
the other being governed by a chaotic map.
According to the approximate qualitative analysis, the phase should be trans-
formed from one epoch to the next synchronization in accordance with the four-fold
expanding circle map. Figure 11.4 shows the phase diagram obtained in compu-
tations for the collective excitation of one of the ensembles relating to successive
epochs of synchronization. The phases are determined by a formula Φ = arg(< x >
−iω0−1 < ẋ >) at the time instants tn = nT, and Φn+1 versus Φn are plotted. As
seen, the diagram clearly indicates that the transformation is indeed close to the
Bernoulli-type map. The observed scatter of points decreases with increasing num-
ber of oscillators K, as confirmed in computations.
Fig. 11.4 Diagram for the phases of the collective excitation of one of the ensembles relating to
the moments of time tn = nT . The parameters are the same as in Fig. 11.3.
Fig. 11.5 Portrait of attractor in the stroboscopic section, corresponding to the instants tn = nT , in
projection on the plane of averaged generalized coordinates and velocities. The parameters are the
same as in Fig. 11.3.
Figure 11.5 shows a portrait of the attractor of the system in the stroboscopic
section, in projection on the plane of averaged generalized coordinates and velocities
of the oscillators. Transversal Cantor structure, which was expected to be observable
on this portrait, is masked by the spread of points, which, apparently, is mainly due
to the finite size of the ensemble.
The question about features of Lyapunov spectra of systems of globally cou-
pled oscillators is too multifaceted to discuss here (Nakagawa and Kuramoto, 1995;
Strogatz, 2000; Maistrenko et al., 2005); so, we consider only the largest Lyapunov
exponent. Its quantitative estimate via the Benettin algorithm at the parameters spec-
ified above yields λ = 0.01069±0.00002. For the stroboscopic map, accounting that
the period of parameter modulation is T = 100, we get Λ = 1.069. It is notably less
than the value ln 4 = 1.386 · · · expected from qualitative analysis of the phase dy-
namics according to the Bernoulli type map. Nevertheless, it does not contradict the
hypothesis of occurrence of suspension of attractor of Smale-Williams type in the
phase space of the system responsible for the collective behavior of the oscillatory
ensembles.
11.2.2 Slow-amplitude approach
Under some additional assumptions (relatively small A, κ , and ε ; large enough av-
erage frequency ω 0 ; relatively narrow frequency range Δω ) the dynamics of two
alternately synchronized ensembles may be treated in terms of slow complex ampli-
tudes, like that for a single ensemble in Sect. 11.1. To do so, introduce the complex
amplitudes ak and bk to oscillators relating to two sub-ensembles in such a way that
the coordinate and velocity variables are expressed as
xk = ak eiω0t + a∗k e−iω0t , yk = bk eiω0t + b∗k e−iω0t ,

(11.13)
ẋk = iω0 ak eiω0t − iω0 a∗k e−iω0t , yk = iω0 bk eiω0t − iω0 b∗k e−iω0t .
It implies that the complex amplitudes satisfy the additional conditions
ȧk eiω0t + ȧ∗k e−iω0t = 0, bk eiω0t + b∗k e−iω0t = 0. (11.14)
Substitution of relations (11.13) in Eqs. (11.11) with taking into account (11.14)
yields
2iω0 ȧk eiω0t + (ω 2 − ω02 )(ak eiω0t + a∗k e−iω0t )

= (A − a2k e2iω0t − 2|ak |2 − a∗2
k e
−2iω0 t
)(iω0 ak eiω0t − iω0 a∗k e−iω0t )
+κ f1 (t) < ẋk > −iε < yk ẏk > (eiω0t − e−iω0t ),
(11.15)
2iω0 ḃk eiω0t + (ωk2 − ω02 )(bk eiω0t + b∗k e−iω0t )
= (A − b2k e2iω0t − 2|bk |2 − b∗2
k e
−2iω0 t )(iω b eiω0 t − iω b∗ e−iω0 t )
0 ω 0 ω
+κ f2 (t) < ẏk > −iε < xk ẋk > (eiω0t − e−iω0t ).
Now, multiply these expressions by e −iω0t and carry out the averaging operation
over a period of fast oscillations, which is indicated with the overline. It is easily to
see that
< ẋω > e−iω0t = iω0 < a >, < ẏω > e−iω0t = iω0 < b >,
< xω ẋω > (eiω0 t − e−iω0t )e−iω0t
(11.16)
= < iω0 (a2ω e2iω0t − a∗2
ωe
−2iω0t ) > (1 − e−2iω0t ) = −iω < a2 >,
0
< yω ẏω > (eiω0 t − e−iω0t )e−iω0t = −iω0 < b2 > .
Instead of the frequencies of the oscillators ω k it is convenient now to use the de-
tuning Ωk = (ωk2 − ω02 )/2ω0 ≈ ωk − ω0 . Finally, we get
1 1 1
ȧk = iΩk ak + (A − |ak |2 )ak + κ f1 (t) < a > − iε < b2 >,
2 2 2
(11.17)
1 1 1
ḃk = iΩk bk + (A − |bk |2 )bk + κ f2 (t) < b > − iε < a2 >,
2 2 2
where
1 K 1 K 2 1 K 1 K 2
< a >= ∑ ak ,
K k=1
< a2 >= ∑ ak ,
K k=1
< b >= ∑ bk ,
K k=1
< b2 >= ∑ bk .
K k=1
(11.18)
The detuning values Ω k are determined in accordance with (11.10).
Fig. 11.6 Graphs of time dependence of modulus (a) and phase (b) of the complex amplitude
averaged over the ensemble for two subsystems, obtained from numerical solution of Eqs. (11.17).
The number of oscillators in each ensemble is K = 1000, the distribution function of frequencies
corresponds to the formula (11.7), the values of other parameters T = 100, κ = 1, A = 3, ε = 0.025.
Figure 11.6 shows the graphs for the modulus and phase averaged over the
ensemble of the complex amplitude versus time for two subsystems: < a(t) >=
| < a(t) > |eiΦ(t) , < b(t) >= | < b(t) > |eiΘ(t) , see, respectively, diagrams (a) and
(b). A number of oscillators in each of two ensembles is K = 1000. Other param-
eters are T = 100, κ = 1, A = 3, ε = 0.1. Epochs of synchronization correspond to
high absolute values of the respective amplitude. As seen, the phase remains roughly
constant over time during those epochs. Although the time dependence of the am-
plitude appears closely periodic in time, in fact, the dynamics is chaotic. Indeed, the
phases change from one epoch of synchronization of the ensemble to another, in ac-
cordance with the chaotic map. In Fig. 11.7 empirical diagram is shown obtained in
the computations phase for the phase Φ n = arg(< a >) of the collective excitation of
one of the ensembles at the times t n = nT relating to the epochs of synchronization
of the ensemble.
Fig. 11.7 Diagram for the phases of the collective excitation for one of the ensembles relating to
Figure 11.8 shows a portrait of the attractor in the stroboscopic section in a pro-
jection on the plane averaged over the ensemble values of generalized coordinates
and velocities of the oscillators. By zoom of the image noticeable transverse Cantor
structure of the attractor becomes visible (see box in the center of the diagram). It
may be conjectured that it is attractor of Smale-Williams type in the reduced low-
dimensional phase space of collective variables of the system.
Fig. 11.8 Portrait of attractor in the stroboscopic section at tn = nT in projection on the plane of
the complex amplitude averaged over the ensemble. The parameters are the same as in Fig. 11.6.
11.2.3 Description of the dynamics in terms of ensembles of phase

oscillators
Let us undertake a final step in simplifying the model and consider the alternately
synchronized ensembles of phase oscillators. Now, the individual oscillators will be
characterized by one and the same constant amplitude R, and the dynamics mani-
fests itself only in evolution of their phases. To reduce the description to the phase
equations, we substitute in (11.17)
ak ≈ ReiΦk , bk ≈ ReiΘk . (11.19)
In selection of R there is certain arbitrariness; a sufficiently good

√ approximation at
moderate values of parameter A and relatively small κ is R = A.
Substituting (11.19) into Eqs. (11.17), after multiplying the first and the second
equations, respectively, by e −iΦk , e−iΘk , and taking the imaginary part responsible
for the dynamics of phase, we obtain
1
Φ̇k = Ωk + Im((κ < eiΦ > −iε R < e2iΘ >)e−iΦk ),
2
(11.20)
1
Θ̇k = Ωk + Im((κ < eiΘ > −iε R < e2iΦ >)e−iΘk ),
2
where
K K
< eimΦ >= K −1 ∑ eimΦk , < eimΘ >= K −1 ∑ eimΘk , m = 1, 2. (11.21)
k=1 k=1
Again, to be concrete, the quantities Ω k are supposed to be determined from (11.10).

Figure 11.9 shows graphs for two subsystems for time dependence of the mod-
ulus and phase of the complex order parameter. The number of oscillators in each
ensemble K = 1000, and other parameters are the same as in the previous subsec-
tion, T = 100, κ = 1, A = 3, ε = 0.025. Fig. 11.10 is a diagram for the phases of
the collective excitation Θ n = arg(< b >) at tn = nT . In Fig. 11.11, a portrait of
the attractor of the system is shown in the stroboscopic section, in projection on the
complex plane of the variable < b(t) >.
To conclude, we have proposed and studied a model system, which realizes
chaotic dynamics of the phases described by expanding circle map transformation in
the course of the transfer of collective excitation alternately between two synchro-
nizing and desynchronizing groups of oscillators. As hoped, it may stimulate search
for and discovery of other examples of this kind, perhaps closer to problems of spe-
cial disciplines such as fluid dynamics, and neurodynamics. It is interesting to note
in this context again that the concept of turbulence presented by Ruelle and Takens,
the radical for its time, postulated the emergence of structurally stable hyperbolic
strange attractors (Ruelle and Takens, 1971), not chaotic attractors in broader sense,
as this term is commonly used now.
Fig. 11.9 Graphs of dependence for modulus (a) and phase (b) of complex order parameters for
the two subsystems, obtained by numerical solution of Eqs. (11.20). The number of oscillators in
each ensemble is K = 1000, the distribution function of frequencies corresponds to the formula
(4), the values of other parameters T = 100, κ = 1, A = 3, ε = 0.025.
Fig. 11.10 Diagram for the phases of collective excitation of an ensemble of (11.20) relating to
Fig. 11.11 Portrait of attractor in the stroboscopic section, corresponding to the moment of time
tn = nT , projected on the plane of the complex order parameter of one of the ensembles of the
phase oscillators (11.20). The parameters are the same as in Fig. 11.9.
Furthermore, it seems realistic to construct artificial systems with collective

chaotic phase dynamics based on ensembles of individual elements that can show
only regular dynamics. This can be done, say, on the basis of electronic devices,
such as arrays of Josephson contacts (Wiesenfeld and Swift, 1995), or with nonlin-
ear optical systems, like arrays of semiconductor lasers (Glova, 2003). Such systems
will represent generators of strong chaos providing the power level much higher
than that characteristic to the individual elements. Systems of this kind may be of
interest for applications requiring the generation of chaotic signals, such as commu-
nication schemes (Dmitriev and Panas, 2002; Koronovskii et al., 2009), and noise
radar (Lukin, 2001).
References
Acebrón, J.A., Bonilla, L.L., Pérez Vicente, C.J., Ritort, F.: The Kuramoto model: A simple
paradigm for synchronization phenomena. Rev. Mod. Phys. 77, 137–185 (2005).
Cross, M.C., Hohenberg, P.C.: Pattern formation outside of equilibrium. Rev. Mod. Phys. 65, 851–
1112 (1993).
Dmitriev, A.S., Panas, A.I. Dynamical Chaos: New Information Carriers for Communication Sys-
tems. Fizmatlit, Moscow (2002).
Glova, A.F.: Phase locking of optically coupled lasers. Quantum Electronics 33, 283–306 (2003).
Koronovskii, A.A., Moskalenko, O.I., Hramov, A.E.: On the use of chaotic synchronization for
secure communication. Physics-Uspekhi 52, 1213–1238 (2009).
Kuramoto, Y.: Chemical Oscillations, Waves, and Turbulence, Springer, Berlin (1984).
Kuznetsov, S.P., Pikovsky, A., Rosenblum, M.: Collective phase chaos in the dynamics of interact-
ing oscillator ensembles. Chaos 20, 043134 (2010).
Lukin, K.A.: Noise radar technology. Telecommunications and Radio-Engineering 16, 8–16
(2001).
Maistrenko, Y.L., Popovych, O.V., Tass, P.A.: Desynchronization and chaos in the Kuramoto
model. In: Chazottes, J.-R and Fernandez, B. (eds.): Dynamics of Coupled Map Lattices and
of Related Spatially Extended Systems, pp.285–306. Springer (2005).
Nakagawa, N. and Kuramoto, Y.: Anomalous lyapunov spectrum in globally coupled oscillators.
Physica D 80, 307–316 (1995).
Osipov, G.V., Kurths, J., Zhou, C.: Synchronization in Oscillatory Networks. Springer, Berlin
(2007).
Ruelle, D., Takens, F.: On the nature of turbulence. Commun. Math. Phys. 20, 167–192 (1971).
Strogatz, S.H.: From Kuramoto to Crawford: exploring the onset of synchronization in populations
of coupled oscillators. Physica D 143, 1–20 (2000).
Wiesenfeld, K., Swift, J.W.: Averaged equations for Josephson junction series arrays. Phys. Rev.
E 51, 1020–1025 (1995).
Part IV
Experimental Studies
Chapter 12
Electronic Device with Attractor of
Smale-Williams Type
Abstract A very interesting question concerns implementation of models discussed

in previous chapters (as well as other systems with hyperbolic attractors) as real
devices, e.g. in electronics, laser physics, mechanics, etc. In the present chapter
an electronic device is considered inspired by the model examined in Sect. 4.2.
An advantage of such chaos generator is robustness of the dynamics because of
structural stability of the hyperbolic attractor, which may be important for many
applications.
12.1 Scheme of the device and the principle of operation
Consider an electronic device schematized by the circuit diagram in Fig. 12.1

(Kuznetsov and Seleznev, 2006). It is a non-autonomous system combining two
van der Pol oscillators of natural frequencies ω 0 and 2ω0 subjected to appropriate
periodic external driving. Each oscillator contains a coil of inductance L 1,2 and a
capacitor of capacitance C1,2 selected in such a way that

ω0 = 1/ L1C1 , 2ω0 = 1/ L2C2 .
Fig. 12.1 Circuit diagram of a device composed of two coupled van der Pol oscillators with peri-
odically varying parameters: a physical system having a Smale-Williams type strange attractor.
260 12 Electronic Device with Attractor of Smale-Williams Type
Negative-resistance components based on operational amplifiers are introduced

whose respective resistances (−R 1,2 ) can be treated as constant parameters in the
entire ranges of operating voltage of the corresponding oscillating circuits. A non-
linear conductance that ensures increase in energy loss with oscillation amplitude
is implemented by circuit components consisting of two oppositely poled parallel
arrays of series-connected semiconductor diodes. A field-effect transistor is used as
an almost linearly conducting component, with drain current controlled by the gate
voltage slowly varying as a periodic function of time with period T = 2π N/ω 0 ,
where N is an integer. During successive half-periods of its variation, one oscil-
lator is active while the other is suppressed and vice versa. Quadratic nonlinear
element A1 transforms the signal of the first oscillator into the second harmonic sig-
nal, which stimulates activation of the second oscillator in a frequency range around
2ω0 . Backward coupling is arranged via the nonlinear element A 2 mixing signal
from the second oscillator of frequency about 2ω 0 with an auxiliary external sig-
nal of frequency ω 0 to produce a difference-frequency signal resonant with the first
oscillator. It stimulates excitation of the first oscillator as it enters its activity stage.
Thus, the excitation is alternately transferred between the oscillators.
The circuit operates as a chaos generator as follows. Suppose that the signal
produced by the active first oscillator currently has a phase characterized by the
shift constant ϕ : U1 ∼ cos(ω0t + ϕ ). Then, the output of the quadratic nonlinear
transformer A 1 contains the second harmonic cos(2ω 0t + 2ϕ ) with phase 2ϕ . When
the second oscillator becomes active, the generated signal U 2 has the phase 2ϕ .
After this signal is heterodyned with the auxiliary signal by means of the element
A2 , the resulting signal component of frequency ω 0 has the same phase 2ϕ . Thus,
the signal generated by the first oscillator during the next half-period of modulation
has the phase 2ϕ . It is obvious that the phases of the oscillation trains generated
during subsequent half-periods can be represented, at least, approximately, by the
Bernoulli map
ϕn+1 = 2ϕn + const ( mod 2π ). (12.1)
12.2 Experimental observation of the Smale-Williams attractor
The circuit was implemented as a laboratory device (Kuznetsov and Seleznev,

2006). The capacities used were C 1 = 20 nF, C2 = 5 nF. Ferrite-core coils had equal
inductances L1 and L2 of approximately 1 H. The natural frequencies of the os-
cillators were, respectively, f 1 = ω0 /2π = 1090Hz and f 2 = 2 f1 = 2180 Hz. The
negative-resistance amplifier and nonlinear conductance were implemented by using
a 140UD26 operational amplifier and KD102 diodes, respectively. A time-varying
conductance was introduced by using KP303G field-effect transistors. The nonlin-
ear components responsible for the coupling between the subsystems were based on
525PS2 analog frequency multipliers. The output voltages U 1 and U2 were fed into
a measuring device (oscilloscope or spectrum analyzer), or into a computer via an
ADM12-3 analog-to-digital (ADC) converter with 12-bit resolution and a maximum
12.2 Experimental observation of the Smale-Williams attractor 261
sampling frequency of 3 MHz. The functions U̇1 and U̇2 were obtained by using a
standard analog differential amplifier consisting of a 500 pF capacitor and a 62 kΩ
resistor combined with a 140UD26 operational amplifier.
In a wide parameter range the experimental system exhibited chaotic oscillations
as excitation was alternately transferred between the oscillators by the mechanism
discussed above. The operation of the system is illustrated in Fig. 12.2. Panels (a)
and (c) show typical waveforms recorded for N = 8 and N = 4. The time series were
obtained with the ADC converter operating at a sampling rate of 200 kHz (approxi-
mately 200 data points per period corresponding to the main natural frequency ω 0 ).
Panels (b) and (d) show iteration diagrams obtained from processing the recorded
data sampled with time interval equal to a period of the parameter modulation
T = 2π N/ω0 . The phases were calculated from the formula ϕ = arg(U 1 − iω0−1U̇1 )
by substituting the values of the voltage U 1 and the derivative U̇1 at the same instants
produced by the analog differentiator. The topological equivalence of the maps to
the Bernoulli map (12.1) is obvious; it is essential to infer the hyperbolicity of the
attractor from experimental data.
Fig. 12.2 Samples of waveforms for signals generated by the first (solid curves) and the second
(dotted curves) oscillators operating in the chaotic regime, and the empirical iteration diagrams for
stroboscopic records of phases of the first oscillator at N = 8 ((a),(b)) and 4 ((c),(d)).
262 12 Electronic Device with Attractor of Smale-Williams Type
Figure 12.3(a) shows an oscilloscopic image of the chaotic attractor correspond-

ing to N = 4 (with horizontal and vertical deflections proportional to U 1 and U̇1 ,
respectively) photographed with an exposure time of a few seconds to capture a
sufficiently large number of the recurrent loops of a trajectory on the attractor.
The image obviously resembles that shown in Fig. 4.6(a). Figure 12.3(b) shows
the stroboscopic section of the attractor projected onto the plane (U 1 , U̇1 ) obtained
by processing the bivariate time series used to calculate the diagrams for phases in
Fig. 12.2(d). It is obviously similar to the analogous portrait of the Smale-Williams
solenoid (Fig. 4.6(b)). An enlarged fragment is shown in inset demonstrating the fine
Cantor-like structure of the attractor. The largest Lyapunov exponent of the strobo-
scopic map evaluated by the method of (Wolf et al., 1985) and by processing a time
series sampled with the period T = 2π N/ω 0 is Λ ≈ 0.73, in fair agreement with
the estimated value ln 2 ≈ 0.693. The Grassberger-Procaccia correlation dimension
calculated by processing a time series generated at a sampling rate of 200 kHz is
d ≈ 2.3.
Fig. 12.3 Portraits of the attractor projected onto the plane of variables U1 and U̇1 for N = 4: (a)
photograph of an oscilloscopic trace image; (b) stroboscopic section with period T = 2π N/ω0 at
instants corresponding to near-maximum values of U1 with enlarged fragment demonstrating the
fine fractal structure of the attractor.
Fig. 12.4 Spectra of signals produced by the two oscillators operating in the chaotic regime corre-
sponding to N = 4 redrawn from photographs of the spectrum-analyzer display.
Figure 12.4 shows power spectra of signals generated by both oscillators oper-
ating in the chaotic regime at N = 4. Note that the spectra are continuous and are
References 263
localized around the natural frequencies ω 0 and 2ω0 , respectively. They are obvi-
ously similar to those obtained for the model discussed in Chap. 4. Unfortunately,
some numerical results relating to that model cannot be verified by the experiment.
For example, it is hard to determine the complete Lyapunov spectrum or to verify
the hyperbolicity criterion. Nevertheless, all in all, the results obtained strongly sup-
port the assertion that the dynamics of the experimental device correspond to those
of the Smale-Williams type attractor in the non-autonomous system similar to that
analyzed in the theoretical study and in numerical simulations.
Additionally, it is worth noting that a device implementing similar operation
mechanism was suggested on a base of microwave klystron generator (Emel’yanov
et al., 2009; 2010). For that device at the moment, only numerical simulation results
are available.
References
Emel’yanov, V.V., Kuznetsov, S.P., Ryskin, N.M.: Hyperbolic chaos generator based on coupled
drift klystrons. Technical Physics Letters 35, 773–776 (2009).
Emel’yanov, V.V., Kuznetsov, S.P., Ryskin, N.M.: Hyperbolic chaos in the klystron-type mi-
crowave vacuum tube oscillator. Chaos 20, 043104 (2010).
Kuznetsov, S.P., Seleznev, E.P.: A strange attractor of the Smale-Williams type in the chaotic
Wolf, A., Swift, J.B., Swinney, H.L., Vastano, J.A.: Determining Lyapunov exponents from a time
series. Physica D 16, 285–317 (1985).
Chapter 13
Delay-time Electronic Devices Generating Trains
of Oscillations with Phases Governed by Chaotic
Maps
Abstract Here we turn to consideration of systems with time delayed feedback

corresponding to models of Cha. 10, which have been implemented as electronic
laboratory devices and studied in experiments described in (Kuznetsov and Pono-
marenko, 2008; Baranov et al., 2010). In a frame of the hyperbolic theory the status
of dynamics observed in these systems is not so well defined because the classic
formulation of the theory relates to finite-dimensional systems, while for the time-
delay systems the state space is formally infinite-dimensional. Nevertheless, phys-
ically operation principle and observed dynamical phenomena look very similar to
those in systems of alternately excited oscillators, for which the hyperbolic nature
of the chaotic attractor is well established (see Chaps. 4, 7 and 12). On one hand, it
implies a challenging problem for mathematicians to give rigorous definitions and
foundation for existence of uniformly hyperbolic chaotic attractors embedded in the
infinite dimensional state space of the time-delay systems. On the other hand, ac-
counting transparent principle of operation and evident analogy with the alternately
excited oscillators, on physical level of reasoning one can confidently suggest that
the chaotic dynamics in time-delay systems of the considered class will have the
same features relevant to technical applications like structural stability (robustness).

modulation and auxiliary signal
A time-delay system inspired by the model considered in Subsection 10.2.1 was im-
plemented as a laboratory electronic experimental device and studied in (Kuznetsov
and Ponomarenko, 2008). It is based on a single van der Pol oscillator with peri-
odically modulated control parameter to arrange alternating stages of activity and
damping, and contains delayed feedback loop with quadratic nonlinear transforma-
tion of the signal and mixing with auxiliary external reference signal. At each next
stage of activity excitation of the van der Pol oscillator is stimulated by signal from
266 13 Delay-time Electronic Devices Generating Trains of Oscillations with . . .
the delay feedback emitted at a previous stage of activity and transfers the doubled
phase of the oscillations.
The electronic circuit is schematically depicted in Fig. 13.1. The active ele-
ment is a single self-oscillator of van der Pol type implemented as an LC circuit.
A negative-resistance element is arranged on the operational amplifier DA1, and a
nonlinear dissipative element is composed of semiconductor diodes D1–D6. The
natural frequency of this oscillator was f 0 = ω0 /2π = 3 kHz. The modulation of
a parameter responsible for the excitation of self-oscillations was organized by
introducing additional dissipation via the field-effect transistor VT1, whose resis-
tance slowly varies with time due to the action of the external periodic signal as
ΔR ∼ A cos Ωt = A cos(2π t/T ). During successive periods of the parameter mod-
ulation, the oscillator alternately manifests active oscillatory stages and stages of
damping as it is below the generation threshold. The generation at each new activity
stage is stimulated by a signal arriving via the delay feedback line from the output
of the multiplier DA3. In the feedback line, the signal undergoes quadratic nonlinear
transformation (by a multiplier implemented on the amplifier DA2) and differenti-
ation (by a standard scheme of differentiator implemented on resistor R2, capacitor
C2, and operational amplifier DA4). Then, the signal is passed via a delay line com-
prising an analog-to-digital converter (ADC), a computer, and a digital-to-analog
converter (DAC), where the delay is realized by a computer program algorithm. The
delayed signal is multiplied by an auxiliary signal of frequency f 0 supplied from an
external generator. Owing to the selection of the delay time τ = 3T /4, the signal
emitted at the moment of maximum excitation arrives at the oscillator through the
delay line approximately at the beginning of the next stage of activity, thus provid-
ing stimulation for the onset of generation. Owing to the presence of the quadratic
transformation, the system produces phase doubling on each passage from one to
another train of oscillations. The ratio of frequencies of the driving signals in the
discussed experiments was N = ω 0 /Ω = f0 T = 6.
Fig. 13.1 Circuit diagram of a chaos generator based on a single van der Pol oscillator with pe-
riodically modulated parameter and additional time-delayed feedback loop, in which the signal
undergoes quadratic nonlinear transformation and mixing with auxiliary signal.
Figures 13.2—13.4 present some experimental data.

Figure 13.2 shows the waveform of the dynamic variable representing sequential
trains of oscillations. The chaos is manifested in irregular variations of the position
of the carrier oscillations within the envelope in different trains. Figure 13.3 shows
a diagram for the phases of oscillations for sequential trains. To depict the diagram,
the signal generated by the device was recorded as time series in computer via the
ADC for the voltage Un and the derivative U̇n obtained from output of the differen-
tiator. The sampling was performed with discrete time step equal to the period of
slow parameter modulation at time instants corresponding to the middle of the ac-
tive stages of the oscillator. The phases were calculated as ϕ n = arg(Un + iω0−1U̇n ).
As seen, the phase transformation over one period T corresponds topologically to
the expanding circle map, or the Bernoulli map: one bypass for pre-image implies
two-fold bypass of the full circle.
Fig. 13.2 Waveform of signal generated by the laboratory device as photographed from the oscil-
loscope screen.
Fig. 13.3 Diagram for the phases of oscillations for sequential trains obtained by processing the
time series of generated signal from the experimental device.
Figure 13.4 shows portraits of the attractor photographed from the oscilloscope
screen (a,b). Panel (a) corresponds to projection of the attractor from the multidi-
mensional phase space of the time-delay system onto plane of two variables (U, U̇).
Fig. 13.4 Portrait of attractor photographed from the oscilloscope screen that corresponds to pro-
jection onto plane of two variables (U,U̇) from the multidimensional phase space of the time-delay
system; stroboscopic portrait of the same attractor.
Panel (b) is stroboscopic cross-section of this attractor; the sampling is performed

with time interval equal to period of slow parameter modulation. The stroboscopic
image resembles the Smale-Williams solenoid with a noticeable transverse fractal
structure characteristic of the attractors of this type.
Figure 13.5 shows the spectral power density obtained by processing the time
series of the generated signal recorded in the computer. The spectrum is surely con-
tinuous and has a full width at half maximum of about 20% at an average frequency
of 3 kHz. Discrete peaks observed in the left-hand part of the plot originate due to
presence of periodic component due to the external parameter modulation.
Fig. 13.5 Plot of spectral power density in the chaotic regime of operation obtained by processing
the recorded time series from the experimental device.
As believed, this experiment demonstrates chaotic dynamics associated with at-

tractor of Smale-Williams type embedded in the multidimensional phase space of
the time-delay system. Rigorous foundation of this hypothesis, as mentioned, is a
challenging problem.
parameter modulation
Now we turn to a version of the time-delay system discussed in Subsection 10.2.2.

It is based on a single van der Pol oscillator with two delayed feedback loops, one
with linear transmission characteristic, and the other with quadratic nonlinearity
producing the second harmonic; both transmitted signals are mixed to produce the
difference-frequency signal stimulating excitation at the next stage of activity of the
van der Pol oscillator. The arrangement of the device ensures doubling of phase
differences of two neighboring oscillation trains in the course of the process.
Figure 13.6 shows the circuit diagram for the system (Baranov et al., 2010).
The basic element is the self-oscillator designed using the LC-circuit (L = 10 mH,
C1 = 5 uF) and the operational amplifier DA1, which introduces a negative resis-
tance. Nonlinearity is provided by two parallel counter-directed chains of semicon-
ductor diodes. The natural frequency of oscillations in the concrete experimental
setup was about 700 Hz. Parameter responsible for excitation of the self-oscillator
was modulated by an external signal with a frequency of 71 Hz supplied to the gate
of the field-effect transistor VT1. The signal from the oscillator was squared with an
analog multiplier circuit DA2 and differentiated at the operational amplifier DA4.
The signals from the output circuit of differentiation and from the oscillator itself
served as two inputs of an analog-digital converter (ADC) and a computer imple-
menting two-channel delay line (delay time was 21 ms and 7 ms, respectively, which
is approximately 3/2 and 1/2 of a period of parameter modulation). The output sig-
nals transformed back to analog form in the two-channel digital-analog converter
(DAC) are input to the analog multiplier whose output is connected to the coil cir-
cuit inductance of the self-oscillator. This device corresponds qualitatively to Eq.
(10.15). For analysis of the phase relations, the signal from the self-oscillator (the
voltage on the capacitor C1) was recorded in a computer together with the second
component (derivative of the voltage) obtained by analog differentiator by means of
an additional (not shown in the diagram) analog-digital converter.
Fig. 13.6 Schematic diagram of the experimental device.

Figure 13.7 shows a waveform of chaotic signal obtained in the experiment, and
the power spectrum of the oscillations in the regime of chaos generation. They are in
good qualitative agreement, respectively, with Fig. 10.10 and Fig. 10.14. In Fig. 13.8
iteration diagrams for phases are shown obtained by processing the experimental
data recorded from ADC in computer as bivariate time series for the voltage on the
capacitor C1 and its time derivative. Panel (a) is three-dimensional plot for phases
of three successive oscillation trains, which is in qualitative agreement with relation
(10.17). The phases are evaluated at time instants corresponding to the middles of
the active stages of the oscillator ϕ n = arg(Un + iω0−1U̇n ). Panel (b) shows a diagram
for phase differences Δϕ n = ϕn − ϕn−1 and corresponds well to the relation (10.18).
Fig. 13.7 Waveform of chaotic signal generated by the experimental device photographed from
the oscilloscope screen (a) and power spectrum of the signal in this regime (b).
Fig. 13.8 Iteration diagrams for phases as three-dimensional plot obtained by processing of
recorded experimental data (a) and the diagram for phase differences (b).
Figure 13.9 shows portraits of the attractor obtained in the experiment. Panel (a)
is photography from the oscilloscope screen; the horizontal and vertical coordinates
correspond to voltage on the capacitor C1 and to its time derivative, respectively. Re-
call that the attractor of the system with delay is an object in an infinite dimensional
state space, so the image has to be regarded as a two-dimensional projection of the
object. Panel (b) provides a picture of the attractor in the stroboscopic section. This
Fig. 13.9 Portrait of the attractor in the projection onto the phase plane of the oscillator, pho-
tographed from the oscilloscope screen (a), and a portrait of the attractor in the stroboscopic sec-
tion, built by processing the experimental data (b).
attractor embedded in the multidimensional state space of the time-delay system ap-
parently relates to the class of partially hyperbolic attractors, like that discussed in
Sect. 10.3.
The experiment confirmed occurrence of the effect of hard excitation. When ad-
justing the negative resistance introduced into the oscillator circuit due to the pres-
ence of operational amplifier, it was possible to observe hysteresis, as illustrated in
Fig. 13.10. With gradual motion in the course of parameter variation from the left to
the right along the lower branch of the plot, there are no intense oscillations in the
system. They arise via abrupt transition at certain threshold, but they persist in the
course of parameter variation in reverse direction in a relatively wide range that cor-
responds to motion from the right to the left along the top branch of the hysteresis
loop on the plot.
Fig. 13.10 Diagram illustrating hysteresis accompanying excitation of oscillations in the experi-
mental device. The voltage Vpp (“peak-to-peak”) is plotted for the signal generated by the system
versus the negative resistance introduced into the oscillator circuit due to the presence of an oper-
ational amplifier.
References
Baranov, S.V., Kuznetsov, S.P., Ponomarenko, V.I.: Chaos in the phase dynamics of Q-switched van
der Pol oscillator with additional delayed-feedback loop. Izvestija VUZov-Applied Nonlinear
Dynamics (Saratov) 18 (1), 11–23 (2010) .
Kuznetsov, S.P., Ponomarenko, V.I.: Realization of a strange attractor of the Smale-Williams type
Chapter 14
Conclusion
The material presented in this book shows significant progress in the main directions
of the research program outlined in the Preface. We can say now that we have a col-
lection of realistic concrete examples of physically realizable systems with chaotic
dynamics, to which the principles of the hyperbolic theory are applicable (“systems
with axiom A”).
Several approaches to constructing systems with uniformly hyperbolic chaotic at-
tractors are developed. Among them are: exploiting models driven by short periodic
pulses; organizing dynamics in successive stages, each governed by special form
of differential equations; composing systems of oscillators excited alternately and
transferring the excitation of each other with transformations of phases described
by maps with chaotic dynamics. As well, the delayed feedback systems may be de-
signed to implement the last principle. Also we have shown how one can apply the
parametric excitation principle to generation of the hyperbolic chaos.
We have presented and examined a number of concrete systems in which the
presence of uniformly hyperbolic chaotic attractors is confirmed at the level of com-
puter validation of cone criterion or is assumed on the basis of qualitative arguments
and data of computations. Many examples belong to a class of non-autonomous sys-
tems with periodic variation in time of coefficients in the governing equations, but
there are also examples of autonomous systems with attractors of Smale-Williams
type in the Poincaré map. Due to the intrinsic structural stability, these models can
be transformed to a certain extent, without violating the hyperbolic nature of the dy-
namics, and this is a way to get many other new systems with uniformly hyperbolic
chaos. Based on some of the proposed schemes, electronic devices were imple-
mented, in which the dynamics on a hyperbolic attractor has been demonstrated in
experiments.
At the moment, the examples of dynamics on the uniformly hyperbolic chaotic
attractors we have discussed relate to purposefully designed devices, not to systems
of natural origin (where they occur, as thought, extremely rare). However, a border
here is conditional, of course. You may recall that the van der Pol oscillator com-
monly recognized now as a universal model in the theory of oscillations originally
came into use in the application to an electronic device, the vacuum-tube generator.
274 14 Conclusion
It is hoped that models with hyperbolic attractors also may be useful for descrip-
tion of some systems of natural origin, for example, in neurodynamics. Above we
mentioned the work, which discussed a possibility of occurrence of an attractor of
Plykin type in the neuron model of Hindmarsh-Rose (Belykh et al., 2005). Also
in the context of neurodynamics systems with blue sky catastrophes were consid-
ered (Shilnikov and Cymbalyuk, 2005; Shilnikov and Kolomiets, 2008); in extended
models with the phase space dimension four and more situations may arise, corre-
sponding to the presence of attractors of Smale-Williams type.
A possibility of physical realization of hyperbolic chaos opens up prospects for
applications for the well-developed mathematical theory and provides a basis for
comparative studies of hyperbolic and non-hyperbolic chaos, including those in
computations and in experiments. Also, it becomes interesting and informative to
re-examine a field of research, consisting of the construction of complex systems on
the basis of the elements with hyperbolic chaos, like chains, lattices, and networks
(Bunimovich and Sinai, 1988; 1993; Bricmont and Kupiainen, 1997; Järvenpää,
2005; Kuptsov and Kuznetsov, 2009). One might think that models, specifically de-
signed in order to realize the hyperbolic chaos, will be useful for understanding
some fundamental issues still challenging researchers, for example, the problem of
turbulence.
An interesting question is about the diagnostic of hyperbolic chaos in physical
experiment. Although in this respect the mathematical techniques, such as the cone
criterion, seem to be unsuitable, the problem does not look hopeless. At least in
the case of the Smale-Williams attractor a productive and convincing approach to
recognize the respective type of dynamics is plotting diagrams for the cyclic phase
variables demonstrating agreement with expanding mapping of the circle or Anosov
torus maps. This may be applied in both computer calculations, and processing of
experimental data.
Speaking about possible technical applications of systems with hyperbolic attrac-
tors, special attention should be paid to their most important property of structural
stability, or roughness.
Recently, a problem of generation of the so-called robust chaos is widely dis-
cussed (Banerjee et al., 1998; Elhadj and Sprott, 2008; Drutarovský and Galajda P.,
2007; Deshpande et al., 2010). This refers to chaotic dynamics of such kind, when
parameter variation is not accompanied with “windows” of regularity (periodicity),
and the dependence of the largest Lyapunov exponent in a wide range is a smooth
function. It is emphasized that such chaos is desirable in applications, including in
the communication systems, the random number generators, encryption schemes.
To implement such signals, it was suggested to use components having the charac-
teristics as functions with break points. In practice, the ideal breaks are unfeasible,
and it is problematic to eliminate completely the windows of regularity in this way.
At the same time, in systems with uniformly hyperbolic attractors the needed prop-
erties of chaos appear as a natural attribute because of inherent structural stability.
During the last 20 years the active researches are directed on elaboration of appli-
cations of chaotic signals for communication and information systems (Yang, 2001,
2004; Dmitriev and Panas, 2002; Argyris et al., 2006; Koronovskii et al., 2009).
References 275
While the arguments in favor of this development look convincing (great informa-
tion capacity of signals; possibility of controlling the dynamics by small external
perturbations; a variety of methods to enclose information in a signal; rich possibil-
ities for encryption of information), it should be noted that the expected potential
benefits are far not achieved yet. A possible reason is that with the signals generated
by non-hyperbolic attractors, we can not stand on a strong theoretical foundation.
On the other hand, for hyperbolic attractors a complete mathematical description of
chaos does exist; for example, one can mention a possibility of exhaustive enumera-
tion of orbits belonging to the attractor using finite Markov partitions and respective
symbolic dynamics. Therefore, it is likely that involvement of the hyperbolic chaos
may realize the expected advantages of the information and communication systems
based on chaotic signals.
Of special interest for communication systems may be the principle of manip-
ulation of the phases (Kuznetsov, 2005; Kuznetsov and Seleznev, 2006; Isaeva
et al., 2006; Kuznetsov and Pikovsky, 2007; Kuznetsov and Ponomarenko, 2008;
Kuznetsov and Pikovsky, 2008; Kuznetsov, 2011). Chaos manifests itself in this
case as irregular variation of phase of carrier in a generated sequence of oscillation
trains. So, on this basis it is possible to implement systems in which the signal trans-
mission in communication channel will be much less sensitive to noises, losses and
distortions than in schemes proposed so far (Yang, 2001, 2004; Dmitriev and Panas,
2002; Argyris et al., 2006; Koronovskii et al., 2009). In this connection it is appro-
priate to recall the advantages of frequency or phase modulations over amplitude
modulation in the traditional communication technologies.
References
Argyris, A., Syvridis, D., Larger, L. et al.: Chaos-based communications at high bit rates using.
commercial fibre-optic links. Nature 438, 343–346 (2006).
Banerjee, S., Yorke, J.A., Grebogi, C.: Robust chaos. Phys. Rev. Lett. 80, 3049–3052 (1998).
Bricmont, J., Kupiainen, A.: Infinite-dimensional SRB measures. Physica D 103, 18–33 (1997).
Bunimovich, L.A, Sinai, Ya.G.: Spacetime chaos in coupled map lattices. Nonlinearity 1, 491–516
(1988).
Bunimovich, L.A., Sinai, Ya.G.: Statistical mechanics of coupled map lattices. In: Kaneko, K.
(ed.) Theory and application of coupled map lattices, pp. 169–189. John Wiley & Sons Ltd.,
New York (1993).
Deshpande, A., Chen, Q., Wang,Y., Lai, Y.-C., Do Y.: Effect of smoothing on robust chaos. Phys.
Rev. E 82, 026209 (2010).
Dmitriev, A.S., Panas, A.I. Dynamical Chaos: New Information Carriers for Communication Sys-
tems. Fizmatlit, Moscow (2002).
Drutarovský, M., Galajda, P.: A Robust Chaos-Based True Random Number Generator Embedded
in Reconfigurable Switched-Capacitor Hardware. Radioengineering 16 , 120–127 (2007).
Elhadj, Z., Sprott, J.C.: On the robustness of chaos in dynamical systems: Theories and applica-
tions. Front. Phys. China 3, 195–204 (2008).
276 14 Conclusion
Isaeva, O.B., Jalnine, A.Yu., Kuznetsov, S.P.: Arnold’s cat map dynamics in a system of coupled
nonautonomous van der Pol oscillators. Phys. Rev. E 74, 046207 (2006).
Järvenpää, E.: SRB-measures for coupled map lattices. In: Chazottes, J.-R. and Fernandez,
B.(eds.): Dynamics of Coupled Map Lattices and of Related Spatially Extended Systems,
pp.95–114. Springer, (2005).
Koronovskii, A.A., Moskalenko, O.I., Hramov, A.E.: On the use of chaotic synchronization for
secure communication. Physics-Uspekhi 52, 1213–1238 (2009).
Kuptsov, P.V., Kuznetsov, S.P.: Violation of hyperbolicity in a diffusive medium with local hyper-
bolic attractor. Phys. Rev. E 80, 016205 (2009).
Kuznetsov, S.P.: On Feasibility of a Parametric Generator of Hyperbolic Chaos. JETP 106, 380–
387 (2008)
Kuznetsov, S.P.: Dynamical chaos and uniformly hyperbolic attractors: from mathematics to
physics. Physics-Uspekhi 54 (2), 119 (2011).
Physica D 232, 87–102 (2007).
Kuznetsov, S.P., Pikovsky, A.: Hyperbolic chaos in the phase dynamics of a Q-switched oscillator
with delayed nonlinear feedbacks. Europhysics Letters 84, 10013 (2008).
Kuznetsov, S.P., Ponomarenko, V.I.: Realization of a strange attractor of the Smale–Williams type
Kuznetsov, S.P., Seleznev, E.P.: A strange attractor of the Smale–Williams type in the chaotic
Shilnikov, A., Cymbalyuk, G.: Transition between Tonic Spiking and Bursting in a Neuron Model
via the Blue-Sky Catastrophe. Phys. Rev. Lett. 94, 048101 (2005).
case study. Int. J. of Bifurcation and Chaos 18, 2141–2168 (2008).
Yang, T.: Chaotic Communication Systems. Nova Biomedical (2001).
Yang, T.: A survey of chaotic secure communication systems. International Journal of Computa-
tional Cognition 2, 81–130 (2004).
Appendix A
Computation of Lyapunov Exponents: The
Benettin Algorithm
A convenient and widely used algorithm for evaluation of Lyapunov exponents

(Benettin et al., 1980) is shortly explained here. To be concrete, we discuss a
continuous-time autonomous system governed by equation
dx/dt = F(x). (A.1)
Generalizations to the non-autonomous case and to diffeomorphisms are simple and

straightforward.
First, we undertake the numerical solution of Eq. (A.1) up to arrival of the orbit
at the attractor.
Then, to compute the largest Lyapunov exponent we proceed with simultaneous
integration of Eq. (A.1) together with the variation equation
d x̃
= A (x(t)) x̃. (A.2)
dt
Here A(x(t)) is a matrix derivative composed of the partial derivatives of compo-
nents of the vector function F(x) over the components of the vector x. We perform
starting at t = t 0 from x(t0 )
integration on some appropriately chosen time interval T
N
and with some vector x̃(t 0 ) of certain norm x̃(t 0 ) = ∑ x̃2i (t0 ). As x̃ obeys the
i=1
linear equation, this norm can be chosen arbitrarily, say, x̃(t 0 ) = 1. At t1 = t0 + T
we redefine the perturbation to have the norm equal to the initial value but keep-
ing direction of the vector unchanged, namely, x̃ 0 (t1 ) = x̃(t1 )/ x̃(t1 ) (Fig. A.1).
Now, continue the procedure with numerical solution from the initial point x(t 1 )
with the perturbation x̃ 0 (t1 ). Next, obtain the state x(t 2 ) at t2 = t0 + 2T and pertur-
bation x̃(2T ), and redefine the perturbation as x̃ 0 (t2 ) = x̃(t2 )/x̃(t2 ), and so on. For
M steps of the algorithm the total factor determining the change of the perturbation
M
norm will be P = ∏ x̃k (t0 + kT ). Obviously, evolution of the magnitude of the
k=1
perturbation is determined by the largest Lyapunov exponent, so, it is estimated as
278 Appendix A Computation of Lyapunov Exponents: The Benettin Algorithm
M
1 1
λ1 ∼
= lnP = ∑ ln x̃k (t0 + kT ), (A.3)
MT MT k=1
where M is supposed to be large enough. This successive renormalization of the

perturbation vectors is essential feature of the computational algorithm.
Fig. A.1 An illustration for Benettin’s algorithm of evaluation of the largest Lyapunov exponent
(see text).
To compute more than one Lyapunov exponent we monitor a collection of the re-
spective number of perturbation vectors evolving in time along the reference phase
trajectory and at each step of the algorithm redefine the perturbation vectors apply-
ing the Gram-Schmidt orthogonalization process.
It works as follows. Suppose, at some step of algorithm the perturbation vectors
evolve from x 01 (tk−1 ), x02 (tk−1 ), x03 (tk−1 ), . . . at tk−1 = t0 + (k − 1)T to x1 (tk ), x2 (tk ),
x3 (tk ), . . . at tk = t0 + kT. Then, we set
x̃01 (tk ) = x̃1 (tk )/x̃1 (tk ),

x̃02 (tk ) = a2 /a2 , a2 = x̃2 (tk ) − x̃2 (tk ), x̃1 (tk ) x̃01 (tk ),
x̃03 (tk )
= a3 /a3 , a3 = x̃3 (tk ) − x̃3 (tk ), x̃1 (tk ) x̃01 (tk ) − x̃3 (tk ), x̃2 (tk ) x̃02 (tk ),
············
(A.4)
N
Here the brackets designate the dot product: u, v = u · v = ∑ ui vi . The obtained
i=1
vectors x01 , x02 , . . . are used to start a next step of the algorithm at t k = kT . In the
M
course of the computations, the accumulated sums S m (M) = ∑ ln x̃m (kT ) are
i=1
calculated, and the Lyapunov exponents are evaluated from the relations
λm ≈ Sm (M)/MT. (A.5)
References 279
References
Benettin, G., Galgani, L., Giorgilli, A., Strelcyn, J.-M.: Lyapunov characteristic exponents for
smooth dynamical systems and for Hamiltonian systems: A method for computing all of them.
Meccanica 15, 9–30 (1980).
Appendix B
Hénon and Ikeda Maps
In some physical systems description by iterated maps appears in a very natural

way. Say, considering a particle moving under action of periodic pulses (kicks),
one can express the coordinate and velocity just before the next kick via the state
before the previous kick and consider dynamics in discrete time with this mapping.
Particularly, well-known maps of Hénon (Hénon, 1976) and Ikeda (Ikeda et al.,
1980) may be obtained in appropriate setup; it reveals physical significance of these
models.
Accounting framework of the current book, it is worth stressing that chaotic at-
tractors in the Hénon and Ikeda maps do not relate to the class of uniformly hyper-
bolic ones. Transformation of the phase space volumes in the course of dynamical
evolution is similar to that mentioned in Chap. 1 in the context of Fig. 1.4 accom-
panied by formation of local singularities of the density of the “phase fluid”. It is
associated with an inevitable presence of tangencies between stable and unstable
manifolds of orbits on the attractors as illustrated in Chap. 7.
Consider the following simple physical system (Fig. B.1). Let a particle of mass
m be able to move along the x axis with friction force proportional to the velocity,
f = −kv. Suppose further that pulsed kicks effect the particle with period T of in-
tensity depending on the instantaneous position of the particle, i.e. the momentum
transferred is expressed by some function P(x).
Fig. B.1 Mechanical system whose dynamics are described by the Hénon map.
Assume that just before the n-th kick the particle coordinate is x n , and the ve-
locity is vn . Immediately after the kick the velocity is v n + P(xn )/m and then it
282 Appendix B Hénon and Ikeda Maps
decreases exponentially according to the expression v(t) = (v n + P(xn)/m)e−kt/m .

At the instant before the next kick we get
vn+1 = (vn + P(xn)/m) e−kT /m ,

T (B.1)
xn+1 = xn + v(t)dt = xn + (mvn + P(xn )) 1 − e−kt/m k−1 .
0
In terms of variable y = x − mk −1 (ekT /m − 1)v introduced instead of v, Equations

(B.1) may be rewritten as
xn+1 = f (xn ) + byn, yn+1 = xn , (B.2)
where
b = −e−kT /m and f (x) = x(1 − b) + P(x)(1 + b)k −1. (B.3)
Now, let us concretize the spatial distribution of the force field in such a way that
f (x) = 1 − ax2. Then, (B.2) reads
xn+1 = 1 − ax2n + byn , yn+1 = xn . (B.4)
In 1976 Hénon introduced this map from purely abstract motivation to get a sim-
ple model with chaotic attractor. Now, it is evident that this map may serve for de-
scription of simple physical systems, one of which just has been considered. Other
examples are dissipative oscillator and rotator under pulsed periodic driving (Heagy,
1992). In physical setup negative values are natural for the parameter b, although in
the original work of Hénon occasionally the sign of that parameter was set opposite.
Observe that the map is invertible: one can easily express x n and yn via xn+1 and
yn+1 uniquely.
Evaluation of the Jacobi determinant for the Hénon map yields

∂ xn+1 /∂ xn ∂ xn+1 /∂ xn −2a b
J= = = −b. (B.5)
∂ xn+1 /∂ xn ∂ xn+1 ∂ xn 1 0
If we consider a cloud of representative points occupying an area S on the phase

plane (x,y), then, at each successive step of iterations the area will be multiplied by
factor |b| so, at |b| < 1 the system is dissipative.
Figure B.2 shows attractor of the map at parameters a = 1.4, b = 0.3 correspond-
ing to the original work of Hénon.
Figure B.3 shows a chart of dynamical regimes on the parameter plane (a, b).
Distinct gray scales correspond to observed periodic dynamics of different peri-
ods. One special gray tone is used for the chaos. The bottom part of the diagram
corresponds to presence of attractive stable fixed point. As the parameter a grows,
attractors corresponding to periodic oscillations arise out of period 2, then periods 4,
8, etc.; there is a sequence of period doubling bifurcations. The bifurcation lines in
the parameter plane represent borders of respective areas of these periodic regimes.
They accumulate according to the Feigenbaum law to the limit critical line, which
represents the border of chaos (Feigenbaum, 1979; Vul et al., 1984; Schuster and
Appendix B Hénon and Ikeda Maps 283
Just, 2005). Observe small areas of periodic motions scattered over the region oc-
cupied by chaos; because of their characteristic shape, they are called the “shrimps”
(Gallas, 1993, 1994). Presence of these formations is a visually illustrative feature
of non-hyperbolic nature of chaos.
Fig. B.2 Strange attractor of the Hénon map (B.4) at a = 1.4, b = 0.3.
Fig. B.3 Chart of dynamical regimes on the parameter plane of the Hénon map (B.4). White area
in the top left part of the diagram corresponds to divergence of iteration of the map. The numbers
1, 2, 4,. . . indicate the periods of motion associated with the respective areas.
Figure B.4 shows the larger one of two Lyapunov exponents of the Hénon map
(B.4) obtained numerically and plotted versus parameter a at constant b = 0.3.
Observe a characteristic feature of the dependence intrinsic to the non-hyperbolic
chaos. On a background of a set of positive Lyapunov exponent indicating chaos
there occur densely deep narrow drops to negative values of the Lyapunov exponent
in the windows of periodicity (or, windows of regularity). In the chart of dynamical
regimes of Fig. B.3 the diagram B.4 corresponds to a path on the parameter plane
along a vertical line b = −0.3; there occur the drops to negative Lyapunov exponent
values as the path crosses the “shrimps”.
Now, let us turn to another system, an oscillator with cubic nonlinearity driven
by periodic short pulses. The model equation is
Fig. B.4 The largest Lyapunov exponent of the Hénon map (B.4) versus parameter a at constant
b = 0.3.
ẍ + γ ẋ + ω02 x + β x3 = ΣCδ (t − nT ), (B.6)
where x is the dynamical variable, ω 0 is natural frequency, γ is a coefficient of

dissipation, and β is parameter of nonlinearity. External driving is represented by
a sequence of pulses accounted by Dirac delta function, which follow with period
T and have magnitude characterized by constant C. We suppose for simplicity that
the period of pulses contains an integer number of periods of natural frequency,
ω0 T /2π = N.
In assumption that ω 0 and T are relatively large, while γ , β , and C are relatively
small, one can use the slow amplitude method in intervals between the kicks. For
this, we introduce the complex amplitude a(t) and set
x = aeiω0t + a∗e−iω0t , ẋ = iω0 (aeiω0t − a∗e−iω0t ). (B.7)
These relations suggest that the complex amplitude satisfies a condition
ȧeiω0t + ȧ∗ e−iω0t = 0. (B.8)
Substituting the relations (B.7) into (B.6) with zero right-hand part, and accounting
(B.8), after averaging over a period of fast oscillations we obtain
γ 3 iβ 2
ȧ = − a + |a| a. (B.9)
2 2 ω0
Now, suppose that just after the n-th kick at t=nT the complex amplitude is a n =
1
(xn − iω0−1 ẋn ). To obtain the amplitude just before the next pulse, we integrate
2
Eq. (B.9) in period T with the initial condition a(nT + 0) = a n . It can be done
1 3
analytically. Indeed, substitution a = re iϕ yields ṙ + irϕ̇ = − γ r + iβ ω0−1 r3 , and
2 2
after separation of the real and imaginary parts,
Appendix B Hénon and Ikeda Maps 285
1 3
ṙ = − γ r, ϕ̇ = β ω0−1 r2 . (B.10)
2 2
The following hold
r(Δt) = rn e− 2 γ Δt ,
1
Δt (B.11)
3 1 1
ϕ (Δt) = ϕn + β ω0−1 rn2 e− 2 γτ d τ = ϕn + 3β ω0−1γ −1 (1 − e− 2 γ Δt )rn2 .
2 0
Substituting Δt = T , we obtain
1
γ (1−e− 2 γ Δt )rn2
−1 −1
a((n + 1)T − 0) = rn e− 2 γ T eiϕn +3iβ ω0
1
1
γ (1−e− 2 γ Δt )|an |2
−1 −1
= an e− 2 γ T e3iβ ω0
1
. (B.12)
The delta-function kick is accompanied by instant transfer of the momentum; as

seen from (B.6) the velocity ẋ is changed by the value C, while the coordinate x just
after the kick retains its value. Thus, we have to set
1 1
an+1 = (x(nT − 0) − iω0−1(ẋ(nT − 0) + C)) = a(nT − 0) − iω0−1C. (B.13)
2 2
Combining with (B.12) implies that
1 −1 −1 − 1 γT
an+1 = − iω0−1C + ane− 2 γ T e3iβ ω0 γ (1−e 2 )|an | .
1 2
(B.14)
2
Finally, by the change of variable and parameters
1/2 1/2
3β (1 − e−γ T/2 ) C 3β (1 − e−γ T/2 )
Z = ia , A= , B = e−γ T /2 .
ω0 γ 2ω0 ω0 γ
(B.15)
we arrive at the convenient form of the map
Zn+1 = A + BZnei|Zn | .
2
(B.16)
This map was introduced firstly for rather different physical setup, namely, for
optical circular resonator filled by medium with the refractive index varying pro-
portional to squared amplitude of electric field of the propagating electromagnetic
wave (Ikeda et al., 1980). After that, this map called the Ikeda map became a popu-
lar model; it was studied extensively and used often in literature for illustrations of
phenomena of chaotic dynamics.
It is worth noticing that applicability of the map (B.16) to quantitative de-

scription of the nonlinear optical system is doubtful. Correctly, that system
is described by the time-delayed equations written out and examined by the
same authors (Ikeda et al., 1980). As to the map (B.16), it was introduced
heuristically, and never was derived for the nonlinear optical system on a firm
theoretical basis. By contrast, for the nonlinear oscillator this map really de-
livers a reasonable description valid in appropriate asymptotic (Kuznetsov et
al., 2008).
The map (B.16) is represented as a single relation for the complex variable Z, but
in terms of real variables it is in fact two-dimensional. Evaluation of Jacobi matrix
for this map yields J = B 2 ; so, at B <1 the map is dissipative (area compressing).
Figure B.5 shows portraits of attractors for the Ikeda map at different values of
parameter A at B=0.2.
Fig. B.5 Strange chaotic attractor of the Ikeda map (B.16) at B = 0.2 for three different values of
the intensity of external driving.
In Fig. B.6 a chart of dynamical regimes on the parameter plane (A, B) is pre-
sented. If one increases the parameter of intensity of driving A at fixed dissipation
B, the transition to chaos occurs, as a rule, through the period-doubling bifurca-
Fig. B.6 Chart of dynamic regimes on the parameter plane for the Ikeda map (B.16).
References 287
tion cascade. However, globally the border of chaos is arranged in complex manner.
Particularly, there are narrow strips of periodic dynamics entering deeply inside
area occupied by chaos and containing scattered small domains of periodicity, the
“shrimps”.
Figure B.7 shows the larger one of two Lyapunov exponents of the Ikeda map
(B.4) obtained numerically and plotted versus parameter A at constant B = 0.2.
Again, one can see, in a wide range mostly occupied by chaos, a set of deep narrow
drops to negative values of the Lyapunov exponent associated with windows of peri-
odicity. In the chart of dynamical regimes it corresponds to a path on the parameter
plane along a horizontal line with the drops to negative Lyapunov exponent values
as the pass crosses the “strips” and the “shrimps”.
Fig. B.7 The largest Lyapunov exponent of the Ikeda map (B.16) versus parameter A at constant
value of B = 0.2.
In Hénon map and Ikeda map under appropriate selection of parameters the con-
figuration of mapping corresponding to the Smale horseshoe (Appendix C) can be
verified. It proves complex and chaotic nature of trajectories in the phase space, but
chaotic attractors, if occur, do not relate to the uniformly hyperbolic class. A visible
indicator of non-hyperbolicity is presence of tiny domains of periodicity surrounded
by chaos on the charts of dynamical regimes of Figs. B.3 and B.5. It indicates the
absence of the structural stability. In Chap. 7 the Hénon map and Ikeda map are
used as examples of non-hyperbolic behavior to oppose them against the systems
with uniformly hyperbolic attractors.
References
Gallas, J.A.C.: Structure of the parameter space of the Hénon map. Phys.Rev.Lett.70, 2714 (1993).
Gallas, J.A.C.: Dissecting shrimps: results for some one-dimensional physical systems. Physica A
202, 196 (1994).
Feigenbaum, M.J.: The universal metric properties of nonlinear transformations. J. Stat. Phys. 21,
669–706 (1979).
Heagy J.F.: A physical interpretation of the Hénon map. Physica D 57, 436–446 (1992).
Hénon, M.: A two-dimensional mapping with a strange attractor. Commun. Math. Phys. 50, 69–77
(1976).
Ikeda, K., Daido, H., Akimoto, O.: Optical turbulence: chaotic behavior of transmitted light from
a ring cavity. Phys. Rev. Lett. 45, 709–712 (1980).
Kuznetsov, A.P., Savin, A.V., Savin, D.V.:On some properties of nearly conservative dynamics of
Ikeda map and its relation with the conservative case. Physica A 387, 1464–1474 (2008).
Vul, E.B., Sinai, Y.G., Khanin, K.M.: Feigenbaum universality and the thermodynamic formalism.
Russ. Math. Surv. 39, 1–40 (1984).
Appendix C
Smale’s Horseshoe and Homoclinic Tangle
In Cha. 1 we mentioned complex dynamics associated with stretching, folding, and

compression of phase space volume in two-dimensional setting (Fig. 1.4). Formal-
ization of that construction is the so-called Smale horseshoe (Smale, 1967; Guck-
enheimer and Holmes, 1983; Anosov et al., 1995; Devaney, 2003; Schuster and
Just, 2005). It is a simple model map invented by Smale as a counterexample to
assumption that in high dimensions structurally stable systems are typical. There is
no uniformly hyperbolic attractor in this map. Instead, there is a non-trivial chaotic
invariant set of saddle type. Presence of the horseshoe is an indicator of complexity
of dynamics and chaos on some set of orbits, although physically it may be unob-
servable as the invariant set is not attractive.
Consider a region in the form of a stadium, which consists of three parts, the
square S, and half-disks, D 0 and D1 docked on two sides. Let the action of the map
correspond to stretching this area horizontally, compression in the vertical direction,
so it becomes long and narrow strip. Next, we deform it so that it takes a form of a
horseshoe, and impose it into the original area, as shown in Fig. C.1. This is Smale’s
horseshoe map.
All the points belonging to the domains D 0 and D1 , are mapped inside D 0 and
eventually attracted to a stable fixed point located in this area. Non-trivial dynamics
corresponds to a set of orbits starting from domain S. It is clear that after the first
iteration of the map some of them leave the domain of S and fall into the D 0 or D1 ;
they are not our interest.
Where are the starting points placed, which remain inside the domain S after
the first iteration? Obviously, they occupy two narrow vertical bands V 0 and V1 ,
which, when stretched horizontally, just become equal in width to the area S. And
how do placed images of points belong to S at the previous time step? They occupy
two narrow horizontal bands H 0 and H1 , which correspond to the two halves of the
horseshoe. At the intersection of these bands the points live that remain in S for
iterating twice.
Now consider the horseshoe map in two steps of the iterations (see bottom picture
in Fig. C.2). Now, there are four vertical stripes, two of which are located in a strip
V0 (denote them by V00 and V01 ) and two are situated in the band V1 (denote them by
290 Appendix C Smale’s Horseshoe and Homoclinic Tangle
Fig. C.1 Smale’s horseshoe map: transformation of the initial stadium domain with iterating once
and twice.
V10 and V11 ). Points belonging to these strips survive in the region S with twice iter-
ating. Next, there are four horizontal stripes, two of which are located in the strip H 0
(they are H00 and H10 ) and two in the strip H1 (they are H01 and H11 ). They represent
a set where the orbits arrive started in S after twice iterating. The union of overlaps
of these strips determines a set of starting points of orbits, which do not leave the
region S with four times iterating. With increase of number of considered iterations,
each time within each vertical and horizontal strip two more narrow sub-strips are
selected and left for following steps of the construction. In the limit of infinite num-
ber of iterating, the set of vertical strips and of the horizontal ones form a Cantor
set. The intersection of these two sets is a “Cantor grid”, which is an invariant set Ω,
whose elements survive in the area of S forever, for an arbitrarily large number of
iterating of the horseshoe map. Complete description of the invariant set Ω in terms
of symbolic dynamics is achieved by introducing appropriate binary numbering of
horizontal and vertical lines, like that introduced in Fig. C.2 for the first two steps
of the construction.
Elements of the set Ω are encoded by two-side infinite binary sequences, and
dynamics in this representation is determined as a shift by one step at each once
iterating (Bernoulli shift). It implies the presence of chaos. Indeed, given an ele-
ment of this set by a random, say, with a coin toss, we can obtain two sequences of
random digits and use them as head and tail of the two-side binary sequence deter-
mining position of some point belonging to the set Ω in the domain S. Starting at
that point, and considering dynamics in forward and in inverse time we observe vis-
iting the horizontal and vertical bands of Fig. C.2(a) in accordance with prescribed
random sequences. Periodic symbolic sequences correspond to periodic orbits of
Appendix C Smale’s Horseshoe and Homoclinic Tangle 291
Fig. C.2 Initial two steps of constructing invariant set Ω of points surviving in the region S. The
right-hand pictures on panels (a) and (b) show in dark gray the sets of survivors with twice and
four times iterating, respectively.
the horseshoe map; they form an infinite countable set. The set of all nonperiodic
sequences is of cardinality of continuum.
It should be noted that all conclusions about the complex dynamics of the horse-
shoe map are based essentially only on topological arguments. Configuration of
regions may look not exactly like that in Fig. C.1; only topological consistency
is needed. Deduction for complex dynamics associated with existence of chaotic
orbits and of an infinite number of periodic orbits is in any case a significant re-
sult, although, in the case of chaos established by the horseshoe map argumentation
chaotic motion as sustained physically relevant regime may not occur because of
non-attractive nature of the invariant set Ω.
Existence of the invariant set Ω is closely linked with the remarkable ob-
ject called the homoclinic tangle (Guckenheimer and Holmes, 1983; Neimark and
Landa, 1992; Anosov et al., 1995; Hilborn, 2000; Ott, 2002). Figure C.3 shows dis-
position of stable and unstable manifolds for a saddle-fixed point existing in the
domain S for the horseshoe map. Existence of intersection of these manifolds at the
homoclinic point Γ implies existence of an infinite collection of other homoclinic
points! Indeed, the point Γ is mapped to some other point, which must lay at in-
tersection of the stable and unstable manifolds; that point, in turn, is mapped to a
point of the same nature and so on. These multiple intersections are seen in panel
(b), where prolonged segments of the manifolds are shown.
292 Appendix C Smale’s Horseshoe and Homoclinic Tangle
Occurrence of the homoclinic tangle is an equivalent indicator of complex dy-

namics in a sense of presence of non-attractive chaotic invariant set in the phase
space of the system. Simple examples of physically relevant models, in which these
phenomena take place, are the Hénon map and Ikeda map discussed in Appendix B.
Fig. C.3 Disposition of stable and unstable manifolds for a saddle-fixed point of the horseshoe
map intersecting at the homoclinic point Γ (a) and illustration of formation of the homoclinic
tangle (b).
References
9). Springer-Verlag, New York (1995).
Vector Fields. Springer-Verlag (1990).
Neimark, Yu.I., Landa, P.S.: Stochastic and Chaotic Oscillations. Springer-Verlag, Berlin (1992).
Appendix D
Fractal Dimensions and Kaplan-Yorke Formula
Essential attribute of chaotic attractors is intrinsic Cantor-like transversal structure

provided by the repetitive transformations of stretching, folding, and transversal
compression for the phase space volume (see Chap. 1). A tool for quantitatively
characterizing this structure is delivered by the fractal geometry (Mandelbrot, 1982;
Beck and Schlogl, 1993; Ott, 2002; Schuster and Just, 2005). Fractal dimension
generalizes the familiar notion of dimension (1 for curves, 2 for surfaces, 3 for
volumes), but for non-trivial sets, like strange attractors, it may be non-integer con-
taining a fractional part.
The simplest definition relates the so-called box-counting dimension, or capacity.
It is based on covering the set by boxes of equal size ε without overlap (Fig. D.1),
and considering dependence on the needed number of the elements of this size. The
dimension is defined then as
log N(ε )
D0 = − lim . (D.1)
ε →0 log ε
Fig. D.1 Covering of an attractor by boxes of fixed size ε (a) and constructing classic Cantor set
by successive steps of removing the middle thirds of segments obtained in the previous steps (b).
294 Appendix D Fractal Dimensions and Kaplan-Yorke Formula
For example, a classic Cantor set (which appears just in the above dissipative
baker map in cross-section of the picture in Fig. 1.3 by a vertical line) obviously
can be covered by N = 2 n intervals of length ε = 3 −n (it just corresponds to the n-th
step of its construction). Taking the limit n → ∞, which corresponds to ε → 0, we
obtain from (1.13) D = log2/ log3 = 0.6309 . . . Dimension of the attractor of the
baker map may be easily found to be larger by one, namely, D b = 1.6309 . . . .
As it is, to characterize strange attractors with sufficient completeness, it is neces-
sary to introduce not one but a set of dimensions. Why is it so that a single dimension
D0 is not enough?
Imagine that the attractor is essentially not uniform: some of elements of cov-
ering are visited more frequently than others. This is in no way reflected in the
definition of D 0 , although, evidently, it should be significant for such a quantitative
characterization of the attractor, which pretends to be a dimension.
To account the different probabilities of visiting the cells one needs to account
the invariant measure. As noted in Chap. 1, the natural invariant measure of any
region of phase space S may be defined as the probability of residence in the area
for a typical orbit: p(S) = lim (τ (S, T )/T ), where τ (S,T ) is the residence time in
T →∞
S at observation interval T . Assuming that the invariant measure is defined for the
attractor, we attribute certain measure p i to each i-th cell, which is probability of
N(ε )
residence there of a typical orbit. Now, consider a sum I(ε ) = − ∑ pi log pi . It may
i=1
be interpreted as amount of information needed to specify a cell of residence at given
instant. It is clear that the sum will increase with decreasing size of the cells. We
can assume that this growth should obey a power law, I(ε ) ∼ = ε −D1 , or, equivalently
that one can define a limit

N(ε )
D1 = − lim (I(ε )/ log ε ) = lim
ε →0 ε →0
∑ pi logε pi . (D.2)
i=1
It is called the information dimension.

Moreover, in 1983 an infinite family of generalized dimensions was introduced
in use in the nonlinear dynamics called the R ényi dimensions on behalf of the Hun-
garian mathematician who proposed them earlier in another context (Grassberger,
1983; Hentschel and Procaccia, 1983).
Assume that the size of boxes covering the attractor is ε , and the probability of
residence in the i-th cell is p i . Then for any real number q, the dimension D q is
defined as the limit
N(ε )
1
Dq = lim logε ∑ pi .q
(D.3)
q − 1 ε →0 i=1
Note that at q = 0 it turns to the capacity dimension, and at q = 1 it is the infor-

mation dimension. The last follows from evaluating the limit q → 1 by means of the
l’Hôpital’s rule.
Appendix D Fractal Dimensions and Kaplan-Yorke Formula 295
One more special case is the so-called correlation dimension D 2 , which is distin-
guished because of existence of relatively simple procedure of estimate in compu-
tations suggested by (Grassberger and Procaccia, 1983).
In more accurate mathematical analysis, improved definitions were developed
accounting possibility of covering sets with elements of different forms and sizes
(Hausdorff dimension) (Guckenheimer and Holmes, 1983; Beck and Schlogl, 1993;
Ott, 2002).
Remind now that formation of fractal Cantor-like structure in strange attractors
occurs due to stretching and compressions of phase volume, which are associated
with presence of positive and negative Lyapunov exponents, respectively. Hence,
one can ask about a relation between the dimensions and the Lyapunov spectrum. An
answer for this question is somewhat given by the Kaplan-Yorke formula (Kaplan
and Yorke, 1979).
Let the dimension of the phase space of a dissipative system be N, then we have
N Lyapunov exponents, which are assumed to be numbered in descending order:
λ1 , λ2 , . . . , λN . If the attractor is chaotic, the largest exponent is positive. On the
other hand, the sum of all Lyapunov exponents is negative. Therefore, calculating
m
the sums Sm = ∑ λk for m = 1, 2, . . . we will get a result at first positive and then
k=1
negative values (Fig. D.2). We select such m for which S m > 0, but Sm+1 < 0. Con-
sider the subspace spanning over the perturbation vectors associated with the first m
exponents. In this space, the phase volume will increase in the course of the dynam-
ics. On the other hand, in the subspace spanning over m+1 vectors the phase volume
will be compressed. Along other dimensions the compression is even stronger, so,
we will ignore those dimensions.
Fig. D.2 Illustration of dependence of a sum of Lyapunov exponents on the number of terms
accounted. Particularly, S1 is the largest Lyapunov exponent λ1 , the maximum value of Sm is the
Kolmogorov-Sinai entropy h, SN is negative constant d, characterizing the contraction rate for
approach to the attractor. A point of intersection of the graph with the horizontal axis yields the
Kaplan-Yorke dimension according to formula (D.7).
In the subspace of dimension m + 1 consider covering of the attractor by boxes

of size ε assuming the orientations of the edges correspond to the directions as-
sociated with the Lyapunov exponents λ 1 , λ2 , . . . , λm+1 . Let the number of cells
296 Appendix D Fractal Dimensions and Kaplan-Yorke Formula
needed be N(ε ). Due to dynamics of representative points along respective orbits,

the domain covered by the boxes after some time T transforms to another domain,
strongly elongated in the direction corresponding to the largest exponent λ 1 and
flattened in the direction associated with the last one taken into consideration, λ m+1
(Fig. D.3). Each box transforms to a parallelepiped of size ε 1 × ε2 × ε3 × · · · × εm+1 ,
where εi = ε eλi T . Covering of one parallelepiped by boxes of size ε m+1 requires
m m
n = ∏ (εi /εm+1 ) = ∏ ε eλi T /ε e−|λm+1 |T cells. Doing so with all N(ε ) paral-
i=1 i=1
lelepipeds we obtain a new covering of the attractor with the number of cells
m
N(εm+1 ) = nN(ε ) = N(ε ) ∏ ε eλi T /ε e−|λm+1 |T
i=1
m
= N(ε ) ∏ eλi T /e−|λm+1|T . (D.4)
i=1
Fig. D.3 To the derivation of the Kaplan-Yorke formula.
Estimate of the dimension according to (D.1) yields

m
logN(ε ) + log ∏ eλi T − m loge−|λm+1 |T
logN(εm+1 )
D∼
i=1
=− = . (D.5)
log εm+1 − log ε + |λm+1 |T
Multiplying the left-and the right-hand parts by − log ε + |λ m+1 |T we obtain

m
−D log ε + D|λm+1 |T = logN(ε ) + ∑ λi T + m|λm+1|T. (D.6)
i=1
Accounting that D ∼ = −logN(ε )/log ε , we cancel terms in both parts of the relation
and finally obtain the Kaplan-Yorke formula
m
∑ λi
i=1
D = m+ . (D.7)
|λm+1 |
References 297
m
Here, remind that the number m is determined in such a way that S m = ∑ λi > 0,
i=1
m+1
but already S m+1 = ∑ λi < 0.
i=1
The initial conjecture was that this formula evaluates the information dimension
of strange attractors. Indeed, it was proven for attractors of two-dimensional maps.
But in general case the conjecture was not confirmed and, generally speaking, it is
not valid. Nevertheless, empirically, this formula provides usually sufficiently good
results, remarkably close to correct dimensions of attractors in many cases.
It the framework of the present book, one should specially outline that the above
derivation exploits essentially assumption of uniform expansion and contraction of
the phase volume elements in the course of the dynamical evolution. Hence, we
expect that among various chaotic attractors, this formula is especially appropriate
just for the uniformly hyperbolic attractors.
To conclude, observe how meaningful the diagram of Fig. D.2 is, which depicts
the sums Sm versus the number of accounted Lyapunov exponents m. From this plot
one can see immediately a set of relevant quantifiers for chaotic attractors: S 1 equals
the positive Lyapunov exponent responsible for chaos, S max is the Kolmogorov-
Sinai entropy (in accordance with the Pesin formula), the intersection of the plot
with the abscissa axis determines the Kaplan-Yorke dimension, and the value S N
determines the rate of compression of the phase volume in the course of approach
to the attractor.
References
Beck, C., Schlogl, F.: Thermodynamics of Chaotic Systems: An Introduction. Cambridge Univer-
sity Press, Cambridge (1993).
208 (1983).
Grassberger, P.: Generalized dimensions of strange attractors. Phys. Lett. A 97, 227–230 (1983).
Hentschel, H.G.E., Procaccia, I.: The infinite number of generalized dimensions of fractals and
strange attractors. Physica D 8, 435–444 (1983).
Kaplan, J.L., Yorke, J.A.: A chaotic behavior of multi-dimensional differential equations. In: Peit-
gen, H.-O. and Walther, H.-O. (eds.) Functional Differential Equations and Approximations of
Fixed Points. Lecture Notes in Mathematics, 730, pp. 204–227. Springer-Verlag, Berlin, N.Y.
(1979).
Mandelbrot, B.B.: The Fractal Geometry of Nature. W.H. Freeman and Company, New York
(1982).
KGaA, Weinheim, FRG (2005).
Appendix E
Hunt’s Model: Formal Definition
Hunt’s model is an artificially constructed example of a flow non-autonomous sys-

tem with attractor of Plykin type in stroboscopic Poincaré section developed by
T. Hunt in his PhD thesis under supervision of Prof. Robert MacKay (Hunt, 2000).
The set of differential equations is formulated for two variables x and y, with the
right-hand parts depending on x, y, and the time variable t:
dx/dt = f ∗ (x, y,t), dy/dt = g∗ (x, y,t). (E.1)
Here the functions f ∗ and g∗ are continuously differentiable, and have period 2π in
respect to the argument t.
Dynamics of variables x, y in the time interval Δt = 2π is represented in three
successive stages, the duration of each is 2π /3. In formulation of the relations, fic-
titious time s is used, which varies in each stage from 0 to 1. To ensure smoothness
of the flow, relationship between s and t is chosen in such a way that the motion of
the representative points on the plane of x and y is of zero velocity at the edges of
the stages.
In description of the flow, three special curvilinear coordinate systems (r, θ ) i ,
i = 1, 2, 3 are used in domains D 1,2,3 (Fig. 2.6), being linked with the Cartesian
coordinates x, y by relations
x = X + Si (π /2 − |θ |), y = Y + rsgnθ , |θ | π /2,

(E.2)
x = X + Si r cos fr (θ ), y = Y + r sin fr (θ ), |θ | < π /2,
where f r (θ ) = 2π 2rθ /(π 2 (r + 1) + 4(r − 1)θ 2 ), S1 = −1 and S2,3 = 1. Variables

X and Y determine location of the origin for the i-th coordinate system in the
Cartesian coordinates. For brevity, we designate the transformations to new co-
ordinates and back, respectively, as (r, θ ) i = (R(x, y, X,Y, Si ), Θ(x, y, X,Y, Si )) and
(x, y) = (F i (r, θ , X,Y, Si ), Gi (r, θ , X,Y, Si )). The explicit expressions are derived eas-
ily from (E.2). √ √
Let us set λ = (3 + 5)/2, μ = (3 − 5)/2, and introduce some other constants
determining geometrical disposition of the construction on the plane (x, y) :
300 Appendix E Hunt’s Model: Formal Definition
1 √ √ 1 √ √
X1 = 6 5 − 6 − (3 − 5)π , X2 = 3, X3 = 18 − 6 5 − ( 5 − 1)π ,
4 4
1 √ 1 √
Y1 = (1 + 5), Y2 = 1, Y3 = (3 + 5), (E.3)
2 2
1 √ 1 √
R1 = (1 + 5), R2 = 1, R3 = ( 5 − 1).
2 2

2 3
1. At the first stage t ∈ [0, 2π /3] we set s = sin t . The origins for the
4
special coordinates move in dependence on s in accordance with
Xi1 (s) = λ s (Xi + π /2) − π /2, Yi1 = μ sYi , i = 1, 2, 3. (E.4)
Besides, we set Ri1 (s) = μ s Ri . Given certain s, x, y, we determine the vector

f(x, y, s) = ( f (x, y, s), g(x, y, s)) via the following procedure.
(a) If x 0, perform transformation to the coordinate system number 1:
(r, θ )1 = (R(x, y, −X11 (s),Y11 (s), −1), Θ(x, y, −X11 (s),Y11 (s), −1)) , (E.5)
and, designating by dot the derivative in respect to s, set
θ̇ = (ln λ )θ , ṙ = (γ (θ , X11 (s))h12 (r, R11 (s)) + (1 − γ (θ , X11(s)))) (ln μ )r. (E.6)
Here the functions are introduced:

⎧ π
⎪
⎪ 1, |θ | ,
⎪
⎪ 2
⎨ π π
21 −1
γ (θ , X) = cos 4 π (2|θ | − π )X , θ + X,
⎪
⎪ 2 2
⎪
⎪ π
⎩ 0, |θ | + X,
⎧ 2
⎪ cos π (1 − r/ε ), r < ε, (E.7)
⎪
⎪
⎪
⎪ 1, ε r R − ε,
⎪
⎨
h12 (r, R) = 5 + 1 cos 1 π (r − R), |r − R| < ε ,
⎪
⎪ 4 4 2ε
⎪
⎪
⎪
⎪
⎩ 3, r R + ε.
2
Then, compute components of the vector field in Cartesian coordinates ( f 1 , g1 ) as
%
Ẋ − Ssgn(θ )θ̇ , |θ | π /2,
fi =
Ẋ + Sṙ cos f r (θ ) − Sr sin fr (θ )(ṙ ∂ fr (θ )/∂ r + θ̇ ∂ fr (θ )/∂ θ ), |θ | π /2,
%
Ẏ + sgn(θ )ṙ, |θ | π /2,
gi = & '
Ẏ + ṙ sin fr (θ ) + r cos fr (θ ) ṙ∂ fr (θ )/∂ r + θ̇ ∂ fr (θ )/∂ θ , |θ | π /2,
(E.8)
where i = 1, S = S 1 = −1, Ẋ = −(ln λ )(X11 + π /2), Ẏ = (ln μ )Y11 .
Appendix E Hunt’s Model: Formal Definition 301
(b) If x > 0, perform analogous computations in the coordinate systems 2 and 3.

First, find out
(r, θ )2 = (R(x, y, X21 (s),Y21 (s), 1), Θ(x, y, X21 (s),Y21 (s), 1)) . (E.9)
Then, set
θ̇ = (ln λ )θ ,
ṙ = (γ (θ , X21 (s) − X31 (s))h12 (r, R21 (s)) (E.10)
+(1 − γ (θ , X21 (s) − X31 (s)))) (ln μ )r,
and with the formula (E.8), where i = 2, S = S 2 = 1, Ẋ = (ln λ )(X21 + π /2), Ẏ =
(ln μ )Y21 compute the components of the vector field ( f 2 , g2 ).
Next, find out
(r, θ )3 = (R(x, y, X31 (s),Y31 (s), 1), Θ(x, y, X31 (s),Y31 (s), 1)) (E.11)
and set
θ̇ = (ln λ )θ , ṙ = (γ (θ , X31 (s))h3 (r, R31 (s)) + (1 − γ (θ , X31(s)))) (ln μ )r, (E.12)
where %
cos π (1 − r/ε ), r < ε ,
h3 (r) = (E.13)
1, r ε.
With the formula (E.8), where i = 3, S = S 3 = 1, Ẋ = (ln λ )(X31 + π /2), Ẏ =

(ln μ )Y31 , obtain components of the vector field ( f 3 , g3 ).
(c) Now, determine the vector field
%
( f1 , g1 ), x 0,
f̃(x, y, s) = (E.14)
w(d3 , d2 ) · ( f2 , g2 ) + w(d2 , d3 ) · ( f3 , g3 ), x > 0,
where the function is introduced

1
w = w(u, v) = sin2 π u(u + v)−1 , (E.15)
2
with arguments expressed as d α = max{Rα (x, y, Xα 1 (s),Yα 1 (s)) − μ s Rα , 0}, α =

2, 3.
(d) As a final step of computations at the stage 1 set
3 & '
f(x, y, s) = f̃(x, y, s) + ∑ β (x − Xi1) · Ẋi1 (s) − f̃(x, y, s) + (ln λ )(x − Xi1) .
i=1
(E.16)
Here x = (x, y), Xi1 (s) = (Si Xi1 ,Yi1 ), Ẋi1 (s) = (Si Ẋi1 , Ẏi1 ), and the function is intro-
duced
302 Appendix E Hunt’s Model: Formal Definition
⎧
⎪
⎨ 1, ρ ε /4,
β (ρ ) = cos π (2ρ /ε − 1/2), ε /4 < ρ < ε /2,
2 (E.17)
⎪
⎩
0, ρ ε /2.
Here an inaccuracy occurs in the Hunt’s work (Aidarova and Kuznetsov,

2009). While the definitions in his text correspond to (E.16) and (E.17), in
the code of the program in Mathematical the argument of the function β has
been incorrectly written as squared norm x − X i12 . Computations indicate
its significance: with substitution of the squared norm the hyperbolicity disap-
pears! On the other hand, with the definitions (E.16) and (E.17), the attractor
really is detected as a hyperbolic one. However, at ε = 0.05 selected by Hunt,
it appears difficult to observe the fractal transversal structure of the attractor in
illustrations. Thus, in computations discussed in Chap. 2, an increased param-
eter value ε = 0.17 was taken, although within the allowable range specified
in the work of Hunt.

3 1
2. At the second stage t ∈ [2π /3, 4π /3] set s = sin 2 t − π . Taking as the
4 2
origin the point
X22 (s) = X2 + (1 − s)D,
Y22 (s) = Y2 + (1 − s)D, (E.18)
D = (λ − 1)(X2 + π /2) + μ R2,
we represent an instantaneous state (x, y) in new coordinates as
(r, θ ) = (R(x, y, X22 (s),Y22 (s), 1), Θ(x, y, X22 (s),Y22 (s), 1)) (E.19)
and set
ṙ = −D, θ̇ = D. (E.20)
Backward transformation to Cartesian coordinates is performed with formula
(E.8), where S = 1, Ẋ = −D, Ẏ = −D, and components of the vector field ( f 0 , g0 )
are obtained. Finally, set
f(x, y, s) = w(a2 , b2 ) · ( f0 , g0 ), (E.21)
where the function w is given by (E.14), and its arguments are expressed as
( )
1 1
a2 = max R(x, y, X22 (s), (Y2 + Y22 (s)), 1) − (−Y2 + Y22 (s)) − μ R3 , 0 ,
2 2 (E.22)
b2 = max{Y22 (s) − 2R2 μ − R(x, y, X22 (s),Y22 (s), 1), 0} .

3
3. At the third stage t ∈ [4π /3, 2π ] set s = sin 2 t − π . Take the point
4
Appendix F Geodesics on a Compact Surface of Negative Curvature 303
X13 (s) = X1 + (1 − s)D, Y13 (s) = Y1 + (1 − s)D, D = (λ − 1)(X1 + π /2) + μ R1

(E.23)
as the origin. Transform initial values (x, y) to coordinates (r, θ ):
(r, θ ) = (R(x, y, −X13 (s),Y13 (s), −1), Θ(x, y, −X13 (s),Y13 (s), −1)) . (E.24)
Next, define the flow by formula (E.20) and pass to the Cartesian coordinates by
means of (E.8), setting S = −1, Ẋ = −D, Ẏ = −D. Now,
f(x, y, s) = w(a3 , b3 ) · ( f0 , g0 ), (E.25)
where the function w is defined by (E.15), and

( )
1 1
a3 = max R(x, y, −X13 (s), (Y1 + Y13 (s)), −1) − (−Y1 + Y13 (s)) − μ R2 , 0 ,
2 2
b3 = max{Y13 (s) − 2R1 μ − R(x, y, −X13 (s),Y13 (s), −1), 0} .
4. Finally, setting f(x, y, s) = ( f (x, y, s), g(x, y, s)), we obtain the equations for the
continuous-time model valid at all three stages as follows:
3 3 3 3
dx/dt = | sin t| f (x, y, s(t)), dy/dt = | sin t|g(x, y, s(t)). (E.26)
4 2 4 2
3 3
In the right-hand parts a factor s (t) = | sin t| is taken into account, arising be-
4 2
cause of the passage from fictitious time s to natural time t.
References
Aidarova, Yu. S., Kuznetsov, S.P.: Chaotic dynamics of the Hunt model, an artificially constructed
flow system with a hyperbolic attractor. Preprint. arXiv: 0901.2727 (2009).
Hunt, T.J.: Low Dimensional Dynamics: Bifurcations of Cantori and Realisations of Uniform
Hyperbolicity. PhD Thesis. Univercity of Cambridge (2000).
Appendix F
Geodesics on a Compact Surface of Negative
Curvature
Free mechanical motion of a particle in curved space takes place along the geodesic
lines of metric identified by a positive-definite quadratic form with smooth coor-
dinate-dependent coefficients. In two dimensions, the square of the infinitesimal
displacement is expressed by the relation
ds2 = E(x, y)dx2 + F(x, y)dxdy + G(x, y)dy 2 . (F.1)
Through the coefficients of the quadratic form we can obtain the Gaussian curvature
using the Brioschi formula (Struik, 1988)
1
K=
(EG − F 2 )2
⎛ ⎞
1 1 1 1 1 1
− Eyy + Fxy − Gxx E F − E 0 E G
⎜ 2 2 2 ⎟
x x y y x
⎜ 2 2 2 ⎟
⎜ 1 1 ⎟
·⎜ ⎟
⎜ F y − G x E F −
2 y E E F ⎟ . (F.2)
⎜ 2 ⎟
⎝ 1 ⎠
1 G G
Gy F G x F
2 2
If the curvature is negative, the geodesics are characterized by instability with
respect to transversal perturbations; then, if the motion takes place in bounded do-
main, it is chaotic. Analysis of the dynamics in this situation goes back to Hadamard
and others (Anosov, 1967). This problem was one of the main sources of inspiration
for the creation of the hyperbolic theory. The purpose of this Appendix is to illus-
trate the main features of the motion on a compact manifold of negative curvature
for a specific example, which follows (Balazs and Voros, 1986).
Surface of constant negative curvature (the Lobachevsky plane or pseudosphere)
can be represented by a model of Poincaré, in which points of the surface are
mapped as points of the unit disk, and straight lines (i.e. geodesics) are represented
by circular arcs orthogonal to the edge of the disc (Fig. F.1(a)). Metric given by the
Poincaré model
306 Appendix F Geodesics on a Compact Surface of Negative Curvature
Fig. F.1 Unit disc representing the Lobachevsky plane in the model of Poincaré with shown rep-
resentatives of the geodesics. Domain D for consideration of motion of a particle and a set of
equivalent domains obtained from D via the variable change (F.7) with m = 0, 1, . . ., 7.
4(dx2 + dy2)
ds2 = (F.3)
1 − x 2 − y2
corresponds to a constant negative curvature K = −1 that may be verified by formula
(F.2).
The Lagrange function for the free particle motion contains only the term of the
kinetic energy; accounting (F.3) it is expressed as
2μ (ẋ2 + ẏ2)
L =W = , (F.4)
(1 − x2 − y2)2
where μ is the mass of the particle. The Euler-Lagrange equations governing the
motion of the particle then are obtained in the form
xẏ2 − xẋ2 − 2yẋẏ yẋ2 − yẏ2 − 2xẋẏ

ẍ = 2 , ÿ = 2 . (F.5)
(1 − x2 − y2 )2 (1 − x2 − y2)2
Motion according these equations conserves the kinetic energy W = 4μ (ẋ 2 + ẏ2 )(1−
x2 − y2 )−2 and, consequently, occurs along the geodesics of the metric determined
by relation (F.3). The velocity, as the rate of increase of distance traveling along the
geodesic line, is expressed as
4(ẋ2 + ẏ2)
ṡ2 = = 2W μ −1 . (F.6)
(1 − x2 − y2 )2
Consider motion of the particle in a domain of an octagon D bounded by circular

arcs (Fig. F.1(b)). Because of symmetry inherent to the pseudosphere, the octagon
can be considered equivalent to each of the eight regions attached to the sides, which
in the Poincaré model look like distorted octagons of smaller size. The variable
changes corresponding to the transformation of each of these regions in the original
octagon are given by the relations
Appendix F Geodesics on a Compact Surface of Negative Curvature 307
√ √
1 + 2(x + iy) + e iπ m/4 2 + 2 2

x + iy = √ √ , (F.7)
1 + 2 − 2 + 2e−iπ m/4 (x + iy)
where m is an index of the respective domain. In the course of the motion, each
time when the particle crosses a border of the octagon D, one can think of it as
appearing on the opposite side with coordinates reassigned according to (F.7) and
velocity components obtained as a derivative of this relation:
√
(3 + 3 2)(ẋ + iẏ)
x + iy = √ √ 2 . (F.8)
1 + 2 − 2 + 2e−iπ m/4 (x + iy)
Points of every two opposite segments of the border of the domain D may be
identified; one can think of them as being glued to each other. Topologically, the
resulting object may be thought of as a doughnut with two holes, or a sphere with
two attached handles. According to the classification adopted in the topology, this
is a manifold of genus 2.
Figure F.2 shows a trajectory of free motion of a particle in the domain D ob-
tained from numerical solution of the differential equations (F.5) supplemented with
the rules of the border cross-sections (F.7) and (F.8). For longer integration time the
orbit is observed to cover the domain densely.
To be sure in its chaotic nature it is worth examining the Lyapunov exponents.
Numerically, they may be computed with standard Benettin method solving the
equations of motion (F.5) together with a collection of four sets of linearized equa-
tions
x̃ẏ2 + 2xẏỹ˙ − x̃ẋ2 − 2xẋx̃˙ − 2ỹẋẏ − 2y x̃˙ẏ − 2yẋỹ˙
x̃¨ = 2
(1 − x2 − y2 )2
(xẏ2 − xẋ2 − 2yẋẏ)(xx̃ + yỹ)
+4 ,
(1 − x2 − y2 )3
(F.9)
ỹẋ2 + 2yẋx̃˙ − ỹẏ2 − 2yẏỹ˙ − 2x̃ẋẏ − 2x x̃˙ẏ − 2xẋỹ˙
ỹ¨ = 2
(1 − x2 − y2 )2
(yẋ − yẏ − 2xẋẏ)(xx̃ + yỹ)
2 2
+4 .
(1 − x2 − y2 )3
Fortunately, due to special nature of the problem under consideration, there is an
alternative approach. One can represent the vector x̃ = (x̃, ỹ) in locally determined
special basis, (e , e⊥ ), where e is a unit vector directed along the geodesic line, and
e⊥ is a unit vector orthogonal to the previous one. Setting x̃ = J e + J⊥ e⊥ one can
derive the following equations for the coefficients:
J¨ = 0, J¨⊥ = −K ṡ2 J⊥ , (F.10)
where K is the Gaussian curvature. These equations hold even for a general case of
free motion of a particle in metrics (F.1), with arbitrary dependence of the coeffi-
308 Appendix F Geodesics on a Compact Surface of Negative Curvature
cients on coordinates; then K should be understood as the local curvature depending

on the instant coordinates x, y in accordance with the formula (F.2).
Fig. F.2 A representative trajectory of free motion of a particle in compact domain on a surface
of constant curvature in accordance with Eqs. (F.5) supplemented with the rules of border cross-
sections (F.7) and (F.8).
General solution of the first equation (F.10) is a linear combination of functions

1 and t. Hence, it contributes to the spectrum of Lyapunov exponents with two zero
values. As to the second equation, for the system under consideration the value
ṡ2 is equal to the doubled ratio of a constant kinetic energy of the particle to its
mass, 2W μ −1 √ (see (F.3)). With√ the constant negative curvature K, the fundamental
−1 −1
solutions are e KW μ t and e− KW μ t . So, the Lyapunov spectrum contains one
positive exponent, two zero ones, and one negative one, namely,

λ1 = 2KW μ −1 , λ2 = 0, λ3 = 0, λ4 = − 2KW μ −1 . (F.11)
Presence of a positive exponent indicates chaotic nature of the dynamics. Observe

that the sum of all the exponents is zero that reflects the conservative nature of the
system we consider here.
Results of direct numerical calculation of the Lyapunov exponents by the Benet-
tin method are in good agreement with the theoretical result (F.11). For example, for
x = 0, y = 0, ẋ = 0.6, ẏ = 0.8 we have W = 2μ ,
a trajectory with initial conditions
and accounting K = −1, obtain 2KW μ −1 = 2. The result of computation of the
Lyapunov exponents in time interval Δt = 1000 is
λ1 = 2.0004, λ 2 = 0.0038, λ 3 = −0.0038, λ 4 = −2.0004. (F.12)
Obviously, deflections from (F.11) should be interpreted as inaccuracy of the nu-

merical computations.
Appendix G Effect of Noise in a System with a Hyperbolic Attractor 309
References
Inst. Steklov 90, 3–210 (1967).
Balazs, N.L., Voros, A.: Chaos on the pseudosphere. Physics Reports 143, 109–240 (1986).
Struik, D.J.: Lectures on classical differential geometry. Dover Publications, NewYork (1988).
Appendix G
Effect of Noise in a System with a Hyperbolic
Attractor
In this appendix some results are presented concerning effect of noise in a physically
realizable flow system with a hyperbolic chaotic attractor of the Smale-Williams
type (Jalnine and Kuznetsov, 2008).
Let us examine a system considered in Sect. 4.2 and add random external driving
to get a set of stochastic equations
ẍ − (A cos(2π t/T ) − x2 )ẋ + ω02 x = ε y cos ω0t + Dξ (t),

(G.1)
ÿ − (−A cos(2π t/T ) − y2 )ẏ + 4ω02y = ε x2 + Dξ (t).
Here ξ (t) represents the Gaussian noise; it is assumed that < ξ (t) >= 0 and
< ξ (t)ξ (t − τ ) >= δ (τ ). Parameter D characterizes the noise intensity and can be
varied in a wide range.
For numerical solution of the stochastic equations (G.1) we exploited a second-
order method described in (Mannella and Palleschi, 1989). A plot in Fig. G.1(a)
shows the results of computation for the noisy system at the parameters A = 3, ε =
0.5, ω0 = 2π , T = 10 and at the noise intensity D = 0.01. In gray color we show 100
superimposed samples of the process under effect of noise starting from identical
initial conditions. Due to the noise, in the final part of the interval of observation the
states appear to be essentially different because of instability intrinsic to the orbits
on the chaotic attractor. As a result, the picture becomes fuzzy. For comparison, the
curve shown in black color is related to the system without noise, starting from the
same initial conditions.
It is easy to demonstrate qualitatively that a very similar picture is observed in
the noiseless system if one considers an ensemble of samples with a slight deviation
of the initial conditions. In Fig. G.1(b) we show a set of 100 samples for the system
without noise launched from initial conditions with a small random variation near
the same initial state as in the previous diagram. The range of the random variations
is specially selected to obtain close degree of mutual divergence of the orbits in the
considered time interval in comparison with that produced by the effect of noise.
As shown, for the noiseless system the phases determined at successive stages
of activity for one of the subsystems obey approximately the Bernoulli map. It is
312 Appendix G Effect of Noise in a System with a Hyperbolic Attractor
Fig. G.1 Time dependence for the variable x of the model (G.1): (a) superimposed 100 samples
obtained at the noise level D = 0.02 (gray) and the trajectory without noise (black), all with the
same initial conditions; (b) superimposed 100 samples for the system without noise (D = 0) for
the ensemble of slightly different initial conditions. Here and hereafter the parameters are chosen
to be A = 3.0, ε = 0.5, ω0 = 2π , and T = 10.
interesting to investigate the influence of noise on the iteration diagrams for phases.
Taking into account that during the active stage the oscillations of x(t) are close
to sinusoidal ones with modulated amplitude and floating phase, we determine the
phases at the time moments t n = t0 + nT as ϕn = arg(ẋ(tn ) + iω0 x(tn )). Figure G.2
shows the iteration diagrams for the phases in the presence of noise (gray) and with-
out noise (black). Note that the presence of noise of sufficiently low intensity does
not change the topological nature of the phase map, which remains in the same class
as the Bernoulli map ϕ n+1 = 2ϕn + const.
Fig. G.2 Iteration diagrams for the phase of the first subsystem of the model (G.1) with noise
D = 0.1.
Another approach that allows us to compare the noisy and the noiseless dynamics
is based on the analysis of the Lyapunov exponents. For the model with noise they
can be computed using the same linearized variation equations as in Chap. 4, but
integrated along the reference trajectory in the noisy system (G.1).
In Fig. G.3 the Lyapunov exponents are plotted versus the noise intensity param-
eter D. At low intensity of noise, the largest Lyapunov exponent is close to T −1 ln 2,
which corresponds to the approximate description of the dynamics by the Bernoulli
map. A notable deflection appears only at a sufficiently high level of noise, namely,
at D 0.2. At D ∼ 0.5 the effect of noise is already very relevant, the largest Lya-
punov exponent crosses zero and becomes negative indicating suppression of the
intrinsic dynamical chaos by the external noise. Dependence of other Lyapunov ex-
ponents on the noise intensity is not noticeable at all.
Fig. G.3 Lyapunov exponents versus intensity of noise for the model (G.1).
For a system with hyperbolic chaotic attractor, presence of noise of low intensity
is in some sense inessential: under the slightly varying initial conditions, we can
ensure that the phase trajectory of the system without noise remains close to the
trajectory of the system with noise over arbitrarily large time interval. This assertion,
based on the so-called shadowing lemma (Guckenheimer and Holmes, 1983), is
justified in the context of the mathematical analysis of noise in (Kiefer, 1974). We
have demonstrated this property in computations for the system (G.1).
Suppose two trajectories are launched from identical initial conditions, one in the
“pure” system without noise and the other in the system with noise. Obviously, the
“noisy” orbit will diverge from the “pure” one. Now, let us try to vary the initial con-
ditions for the pure trajectory to get an approximation for the noisy orbit in the best
way at a long time interval. The whole construction is performed for a definite sam-
ple of the noisy orbit obtained with the same sample of noise. To explain the method
of selection of the initial conditions, it is appropriate to consider the Poincaré map
produced by a period-T stroboscopic section of the flow system (G.1) without noise
xn+1 = T(xn ). In Fig. G.4(a) one can see a portrait of the attractor of the map pro-
jected onto the plane (x, ẋ/ω 0 ). This is solenoid manifesting filaments of an infinite
number of wraps possessing the Cantor-like structure in the cross-section. Solid
black dots in the picture denote four successive points of the stroboscopic section
for one specially chosen trajectory on the attractor. The black lines passing through
these dots designate the respective unstable direction D u , which are tangent to the
unstable manifolds W u at the given points. These directions can be approximated
numerically from long-time evolution of an arbitrarily chosen unit perturbation vec-
tor Vn = (x̃, x̃˙/ω0 , ỹ, ỹ˙/2ω0 ) at tn = t0 + nT , since such a vector governed by the
linearized variation equations tends to the unstable direction associated with a sin-
gle positive Lyapunov exponent.
Now one should take a note of how a cloud of representative points evolves
from the same initial conditions in the presence of noise. An illustration is shown in
Fig. G.4(b). The dots pictured in gray are obtained in the Poincaré cross-section, i.e.,
stroboscopically at each next period T from 104 sample orbits of the noisy system
at D = 0.02. The black dots correspond to the trajectory of the system without noise
(D = 0) launched from the same initial conditions. Note that the cloud stretches
along the unstable direction tangent to the filaments forming the attractor, but it
does not grow in the radial direction.
Fig. G.4 The stroboscopic cross-section of the attractor of the system (G.1) at t0 = 0 (a). The
attractor is projected on the phase plane of the first subsystem. Evolution of the cloud of repre-
sentative points launched from identical initial conditions in the presence of noise (D = 0.02) (b).
The labels 1,2,3,4 indicate the number of steps of the Poincaré map. An ensemble of 104 orbits is
shown. Black dots correspond to the system without noise starting from the same initial conditions.
The algorithm for localization of the “pure” trajectory, which approximates a

noisy orbit, consists in the following. Let us denote the original noisy orbit as
Vnoisy (t), while the pure trajectory in zero approximation is denoted as V (0) (t). We
start first from the initial conditions at t = t 0 identical to the noisy and the pure orbit:
Vnoisy (t0 ) = V(0) (t0 ). The starting point is supposed to belong to the attractor of the
system without noise. Then, computing the two trajectories V noisy (t) and V(0)(t),
we consider the norm of the difference vector ΔV (0) (t1 ) = Vnoisy (t1 ) − V(0) (t1 )
at t1 = t0 + T . Next, we slightly modify the initial condition for the pure orbit
from V(0) (t0 ) to V(G.1) (t0 ) with the purpose to minimize the norm of the differ-
ence Vnoisy (t1 ) − V(1) (t1 ). It is provided by variation of the phase of the partial
oscillator active at t = t0 . For our system it corresponds to variation of the initial
state along the unstable direction associated with the point on the attractor, or, that
is the same, along a filament of the attractor containing the initial point. The pro-
cedure of the search for new initial conditions is explained in detail in (Jalnine and
Kuznetsov, 2008).
Suppose we take a new initial condition for the pure orbit V (1) (t0 ). Now with
the same sample of the noise we again trace two trajectories, the noisy one start-
ing from V noisy (t0 ) and the pure one starting from V (1) (t0 ) for a longer time inter-
val up to t2 = t0 + 2T , and obtain the norm of the difference vector ΔV (1) (t2 ) =
Vnoisy (t2 ) − V(1) (t2 ) Then, modifying again the initial condition for the pure orbit
by variation of the initial phase of the active oscillator, at this time in closer neigh-
borhood of V (1) (t0 ), we try to minimize V noisy (t2 ) − V(2) (t2 ) and obtain the new
initial condition V (2) (t0 ). The procedure of variation of the initial conditions is the
same as the above-described, with only difference being that the maximum variation
ΔV(2) is chosen from the relation of ΔV (2) = ΔV(1)(t2 ) exp(−2λ1 T ). Step by step
we successively increase the time interval nT and select the initial conditions for the
pure orbit V (n) (t0 ), which deliver minimal values for the norms of the differences
Vnoisy (tn ) − V(n)(tn ). Note that the maximum variation of the initial condition at
the n-th step of the algorithm decays as ΔV (n) ∼ exp(−nλ1 T ). Due to the recur-
rent nature of the algorithm, it appears that the approximation of the noisy orbit by
the pure one holds uniformly during the whole time interval [t 0 ,t0 + nT ].
Figure G.5 illustrates results of application of several steps of the above algo-
rithm. Parameters of the system are taken the same as those in the previous examples
of computations, and the noise intensity is D = 0.1.
The first plot in Fig. G.5 corresponds to the initial step of the algorithm: the noisy
(gray) and pure (black) trajectories start from identical initial conditions. One can
see their sufficiently fast divergence: the phase synchronism disappears already after
one period of the parameter modulation T . After the first modification of the initial
conditions for the pure orbit the divergence is delayed and the phase synchronism
persists over 1 − 2 periods of T , but then the orbits diverge. At the next steps of the
algorithm the time intervals of existence of the phase synchronism become longer
and longer and occupy finally the whole range in the diagram. Figure G.6 shows in
gray color 20 dots of the projection of the noisy 5 orbit at D = 0.04 on the plane
(x, ẋ/ω0 ). In black color 20 dots are shown in the same stroboscopic cross-section
of the approximating pure trajectory recovered by the above method.
To characterize the degree of closeness of a noisy orbit to a shadowing pure orbit,
we use the following value
t0 +kT * *
1 * *
ρk = *Vnoisy (t) − V (k) (t)* dt, (G.2)
kT t0
where k designates the duration of the considered time interval in units of the modu-
lation period T . Averaging this quantity over the ensemble of initial conditions and
Fig. G.5 Successive steps of constructing the shadowing trajectory (black) to the given noisy orbits
(gray). At the initial step both orbits start from the identical initial conditions. At the last presented
diagram the shadowing takes place over the time range of six modulation periods.
the noise samples we obtain a mean deviation < ρ k >, characterizing the degree
of closeness of the noisy and shadowing trajectories on the attractor. In Fig. G.7(a)
we present plots of the mean deviation < ρ k > versus k for different noise levels. It
can be seen in the figure that the mean deviation depends on k very weakly in the
range k = 8, . . . , 50. At k ∼ 50 ÷ 60 the errors become noticeable due to the finite-
digital arithmetic and for the considered level of accuracy further observation of the
shadowing becomes impossible.
Figure G.7(b) shows a plot of the mean deviation < ρ k > computed for k = 9 in
dependence on the parameter of the noise intensity D. For each D we consider a set
of 100 orbits launched from different initial conditions and calculated for different
noise samples. Each noisy orbit was approximated by the shadowing trajectory with
the method described above, and the mean value < ρ 9 > was evaluated by averaging
over the set of orbits. As seen from the figure, the dependence < ρ 9 > versus D is
nearly linear in the range from zero to 0.1.
References 317
Fig. G.6 Stroboscopic cross-section of the noisy (light gray) and the shadowing noiseless (black)
trajectories on a background of the attractor (dark gray) in projection onto the phase plane of the
first subsystem; the noise intensity D = 0.04.
Fig. G.7 Mean average mutual deviation of noise and shadowing trajectories versus the number of
modulation periods taken for the computations at the three noise levels D = 0.002, 0.01 and 0.02
(a). The dependence of < ρk > on the noise intensity D for k = 9 (b).
References
Jalnine, A.Yu., Kuznetsov, S.P.: Effect of noise in a nonautonomous system of alternately excited
oscillators with a hyperbolic strange attractor. Phys. Rev. E 77, 036220 (2008).
Kifer, Yu.I.: On small random perturbations of some smooth dynamical systems. Math. USSR
Ivestija 8, 1083–1107 (1974).
Mannella, R. and Palleschi, V.: Fast and precise algorithm for computer simulation of stochastic
differential equations. Phys. Rev. A 40, 3381–3386 (1989).
Index
A Delay equation 221

Deviating argument differential equation
Absorbing domain 8, 14, 17, 41, 67, 87, 111 220-221
Andronov-Hopf bifurcation 82, 103, 174, 180, Discrete-time system 4
182 Dissipative system 6
Anosov map 16 Dissipative coupling 190, 243, 248
Anosov system 23 Dynamical system 3
Arnold cat map 16, 23, 41, 173–179, 207–209
Attractor 8 E
Autonomous system 4
Averaging method 83, 93, 123, 125, 133, 244, Electronic device 259, 266
251, 284 Ergodicity 30
Axiom A 22–23 Evolution operator 3
Expanding circle map, see Bernoulli map
B
F
Baker’s map 12
Basic set 23 Fibonacci map 174
Benettin algorithm 43, 277 Flows 4
Bernoulli map 15, 62, 67, 85, 106, 110, 126, Fourier power spectrum 186
195, 215, 240, 261, 267 Fractal dimensions 293–296
Bistability 8
Blue sky catastrophe 52, 112–116 G
Brioschi formula 305
Bubbling attractor 192 Geodesic 50, 305
Global coupling 248–249
C Grassberger-Procaccia dimension 178, 262
Cascades 4 H
Chaotic synchronization 188
Cone criterion 24, 155–160 Hamiltonian 7
Conservative system 6–7 Hénon map 14, 143–144, 147, 153, 281–284
Curvature 50, 305 Hénon method 5–6
Heteroclinic cycle 201–203
D Hindmarsh-Rose model 51
Homoclinic tangle 291–292
DA-attractor 16, 43–46 Hunt model 46–49
320 Index
Hyperbolic chaos 14 Phase trajectory 4

Hyperbolicity 19–22 Plykin attractor 17-19, 43–46
Hyperchaos 180, 187, 210, 227-231 Poincaré map 5
Hysteresis 93, 226, 236, 271 Poincaré section 4–5
Predator-prey model 107
I
R
Ikeda map 143–144, 147, 154, 284–287
Invariant measure 29, 150–155 Rayleigh dissipation function 121–124, 133
Invariant set 4 Return map 5
Riddled basin 196–198
K Roughness 9
Kaplan-Yorke dimension 25, 295–297 Ruelle-Takens scenario 35–37
Kolmogorov-Sinai entropy 31 S
Kuramoto model 245
Kuramoto transition 243 Saddle orbit 222
Saddle point 17, 20
L Shadowing 30, 313–317
Lagrangian 7, 121, 124, 132 Slow amplitudes 125, 133–134, 244, 251–253,
Limit cycle 8, 20, 52, 82, 107, 113, 244 284–285
Liouville theorem 7 Smale-Williams attractor 15, 41, 52, 62,
Locally maximal set 23 88–90, 103, 111, 127, 260
Logistic delay equation 237–238 Smale horseshoe 287, 289–292
Lorenz model 37–39, 143, 147, 152 Solenoid 15
Lyapunov exponent 10–12, 25, 28, 277–278 Spectral decomposition theorem 23
Lyapunov exponent in time-delay system 222, SRB measure 29–30, 188
225, 230, 235, 239 Stable manifold 21, 141–150
State vector 3
M State space 3
Stereographic projection 19
Map 4 Structural stability 8–10, 28–29
Markov partition 26–27, 74–75 Suspension 6
Maximal attractor 8 Symbolic dynamics 26–27
Metric entropy 31
Mixing 30 T
Multistability 8
Taffy-pulling machine 53–54
N Thermodynamic limit 246
Topological entropy 27–28
Newhouse-Ruelle-Takens theorem 35–37 Triple linkage 49–51
Noise 9, 29–30, 198, 237, 311-317
Non-autonomous system 4 U
Non-wandering point 23
Uniformly hyperbolic attractor 21
O Unstable manifold 21, 141–150
On-off intermittency 191–192, 195–196 V
Orbit 4
Van der Pol oscillator 81–83
P Van der Pol oscillator with delay 223, 265
Variation equation 11, 61, 86, 128, 136, 206,
Parametric generator 120–123
222, 277
Parametric oscillations 119–120
Partial hyperbolicity 22 W
Phase oscillator 244, 254–256
Phase space 3 Windows of periodicity 283, 287

Hyperbolic Chaos

Uploaded by

Document Informationclick to expand document informationHyperbolic Chaos

Document Informationclick to expand document information

Copyright:

Available Formats

Hyperbolic Chaos

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Hyperbolic Chaos

Uploaded by

Copyright:

Available Formats

Sergey P.

ISBN 978-3-642-23665-5 e-ISBN 978-3-642-23666-2

Library of Congress Control Number: 2011935039

c Higher Education Press, Beijing and Springer-Verlag Berlin Heidelberg 2012

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

With 214 figures

One of intensively developing research directions is exploration of complex dynam-

About 40 years ago, in mathematical studies a special type of chaotic attractors

the level of computations. In Chap. 8 we consider non-autonomous systems of four

ences, etc. This may relate to various disciplines——mechanics, hydrodynamics,

Part I Basic Notions and Review

1 Dynamical Systems and Hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Possible Occurrence of Hyperbolic Attractors . . . . . . . . . . . . . . . . . . . . . 35

Part II Low-Dimensional Models

3 Kicked Mechanical Models and Differential Equations with

4 Non-Autonomous Systems of Coupled Self-Oscillators . . . . . . . . . . . . . . 81

5 Autonomous Low-dimensional Systems with Uniformly Hyperbolic

5.2 System constructed on a base of the predator-prey model . . . . . . . . . 107

6 Parametric Generators of Hyperbolic Chaos . . . . . . . . . . . . . . . . . . . . . . 119

7 Recognizing the Hyperbolicity: Cone Criterion and Other

Part III Higher-Dimensional Systems and Phenomena

8 Systems of Four Alternately Excited Non-autonomous Oscillators . . . 173

9 Autonomous Systems Based on Dynamics Close to Heteroclinic

10 Systems with Time-delay Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

11 Chaos in Co-operative Dynamics of Alternately Synchronized

Part IV Experimental Studies

12 Electronic Device with Attractor of Smale-Williams Type . . . . . . . . . . . 259

13 Delay-time Electronic Devices Generating Trains of Oscillations

Appendix A Computation of Lyapunov Exponents:

Appendix B Hénon and Ikeda Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281

Appendix C Smale’s Horseshoe and Homoclinic Tangle . . . . . . . . . . . . . . . . 289

Appendix D Fractal Dimensions and Kaplan-Yorke Formula . . . . . . . . . . . 293

Appendix E Hunt’s Model: Formal Definition . . . . . . . . . . . . . . . . . . . . . . . . 299

Appendix F Geodesics on a Compact Surface of Negative Curvature . . . . 305

Appendix G Effect of Noise in a System with a Hyperbolic Attractor . . . . 311

1.1 Dynamical systems: basic notions

A finite-dimensional dynamical system is an object, for which one can specify a

Both continuous-time and discrete-time systems are introduced and considered;

dx1 /dt = F1 (x1 , x2 , . . . , xN ),

xn+1 = g(xn ). (1.3)

xn+k = gk (xn ) ≡ g(g(· · · g(xn ) · · · )) ≡ g ◦ g ◦ · · ·◦ g(xn ). (1.4)

and apply some interpolation procedure. An elegant recipe to do this in perfect

A continuous time system restored from a given invertible map is referred to as

1.1.2 Dynamics in terms of phase fluid: Conservative and

Traditionally, two classes of dynamical systems are distinguished, the conservative

In classic mechanics, Lagrangian and Hamiltonian formalisms are devel-

Considered in terms of variables (q i , q̇i ), the dynamics of a conservative sys-

Then, the dynamics is governed by the Hamilton equations

1.1.3 Rough systems and structural stability

1.1.4 Lyapunov exponents and their computation

To distinguish regular and chaotic dynamics on a quantitative level, a useful and

the spectrum of Lyapunov characteristic exponents. A common convention is to or-

1.2 Model examples of chaotic attractors

1.2.1 Chaos in terms of phase fluid and baker’s map

Let us turn again to representation of dynamics in terms of ensemble of a large

Chaos in dynamics occurs if the cloud of representative points in the course

To illustrate, let us consider a dissipative version of the so-called baker map