Applied Mathematical Sciences

Volume 173
Editors
S.S Antman
Department of Mathematics
and
Institute for Physical Science and Technology
University of Maryland
College Park, MD 20742-4015
USA
[email protected]
J.E. Marsden
Control and Dynamical Systems, 107-81
California Institute of Technology
Pasadena, CA 91125
USA
[email protected]
L. Sirovich
Laboratory of Applied Mathematics
Department of Biomathematical Sciences
Mount Sinai School of Medicine
New York, NY 10029-6574
[email protected]
Advisors
L. Greengard P. Holmes J. Keener
J. Keller R. Laubenbacher B.J. Matkowsky
A. Mielke C.S. Peskin K.R. Sreenivasan A. Stevens A. Stuart

For other titles published in this series, go to

http://www.springer.com/series/34
Ying-Cheng Lai · Tamás Tél

Transient Chaos
Complex Dynamics on Finite-Time Scales

123
Ying-Cheng Lai Tamás Tél
Department of Electrical Engineering Department of Theoretical Physics
Arizona State University Institute of Physics
Tempe Arizona Eötvös University
USA 1117 Budapest
[email protected] Hungary
[email protected]

ISSN 0066-5452
ISBN 978-1-4419-6986-6 e-ISBN 978-1-4419-6987-3
DOI 10.1007/978-1-4419-6987-3
Springer New York Dordrecht Heidelberg London

Library of Congress Control Number: 2011920691

Mathematics Subject Classification (2010): 37-XX

c Springer Science+Business Media, LLC 2011

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.

Printed on acid-free paper

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

Preface

In a dynamical system, transients are temporal evolutions preceding the asymptotic

dynamics. Transient dynamics can be more relevant than the asymptotic states of
the system in terms of the observation, modeling, prediction, and control of the sys-
tem. As a result, transients are important to dynamical systems arising from a wide
range of disciplines such as physics, chemistry, biology, engineering, economics,
and even social sciences. Research on nonlinear dynamical systems has revealed
that sustained chaos, as characterized by a random-like yet structured dynamics
with sensitive dependence on initial conditions, is ubiquitous in nature. A question
is, then, can chaos be transient?
A common perception, as conveyed in many existing books on nonlinear dynam-
ics, is that chaos is an asymptotic property that manifests itself only after a long
observation. Indeed, standard characteristics of chaos, such as the Lyapunov expo-
nents that measure the exponential separation rates of nearby trajectories and hence
quantify the degree of the sensitivity to initial conditions, are defined in the infinite
time limit. These features seem to be incompatible with the possibility of chaotic
transients.
Research on nonlinear dynamics has shown, however, that the essential feature of
chaos is the existence of so-called chaotic sets in the phase space, and quantitative
characterization of chaos is meaningful with respect to the dynamics on such sets
only. Since this does not imply that trajectories from random initial conditions would
necessarily approach these sets asymptotically, transient chaos can arise. Transient
chaos is associated with the existence of nonattracting chaotic sets. Research has
also revealed that transient chaos is in fact more common and possibly richer than
sustained or permanent chaos, since the latter can be regarded merely as a limit of
transient chaos when the average lifetime of the underlying chaotic set becomes
infinite. Transient chaos thus plays a similar role in the realm of complex dynamics
to that of a weakly unstable equilibrium state in regular dynamics. In fact, transient
chaos can be regarded as a kind of metastable state. The concept of transient chaos
is ideally suited to the description of nonequilibrium processes.
The aim of this book is to give an overview, based on the results of nearly three
decades of intensive research, of transient chaos. One belief that motivates us to
write this book is that transient chaos may not have been appreciated even within
the nonlinear-science community, let alone other scientific disciplines. During the

v
vi Preface

course of research and interactions with various scientific communities, we have

become increasingly convinced that knowledge of transient chaos can be particu-
larly important and useful as we witness a proliferation of applications in various
branches of science and engineering based on or motivated by nonlinear dynamics.
We shall show in this book that the basic concepts required to understand tran-
sient chaos are actually fairly easily generalized from concepts of standard nonlinear
dynamics. One special emphasis will be on the fact that certain interesting dynami-
cal phenomena can be understood only in the framework of transient chaos.
That transient chaos can arise in a broad array of fields can be illustrated by the
following examples:
• Chemical reactions in closed containers can lead to thermal equilibrium only.
However, the transients can be chaotic if one begins sufficiently far from equilib-
rium states.
• Certain epidemiological data, e.g., on the spread of chickenpox, can be consis-
tently and meaningfully interpreted only in terms of transient chaos.
• The so-called shimmy (an irregular dancing motion) of the front wheels of mo-
torcycles and airplanes, which can lead to disastrous incidents, turns out to be a
manifestation of transient chaos.
• Satellite encounters and the escapes from major planets are chaotic transients.
• The trapping of advected material or pollutant around obstacles, often seen in the
wake of pillars or piers, is a consequence of transient chaos.
• In nanostructures, today a cutting-edge field of science and engineering, the clas-
sical dynamics of electrons bear the signature of transient chaos.
This book should be regarded as a research monograph and is intended for
graduate students and researchers in science and engineering who are interested in
understanding and applying this extended concept of chaotic dynamics to their re-
spective areas of research. Preliminary knowledge of sustained chaos, e.g., chaotic
attractors, Lyapunov exponents, fractals, periodic orbits, stable and unstable mani-
folds, is assumed. These concepts can be found in almost any existing textbook on
chaotic dynamics.1
Our Book not only gives an introduction to the novel concepts needed for under-
standing and for properly treating transient chaos, but also provides an overview of
various transient-chaos-related phenomena. The book is organized as follows.
Part I: Basics of Transient Chaos. The first part covers the basic concepts, notions,
ideas, theories, and algorithms required for understanding transient chaos.
• Chapter 1: Introduction to Transient Chaos. This chapter is devoted to a prelim-
inary acquaintance with transient chaos, where basic properties of nonattracting
chaotic sets are presented. To underline the relevance of transient chaos, a brief
presentation of a number of experiments is given, which also illustrate different
aspects of the applicability.

1 The textbooks [564, 773] also provide an elementary treatment of transient chaos.
Preface vii

• Chapter 2: Transient Chaos in Low-Dimensional Systems. Dynamics from a

one-dimensional map mimic those along the unstable manifold of, for exam-
ple, a two-dimensional invertible map associated with a three-dimensional flow.
Many fundamental insights into transient chaos can be gained by investigating
one- and two-dimensional dynamical systems.
• Chapter 3: Crises. Transient chaos often precedes the birth of permanent chaos.
Attractor destructions, explosions, and merging are often accompanied by tran-
sient chaos. Dynamical properties of transient chaos are partially inherited by the
enlarged attractor. Transient chaos thus provides the backbone of the motion on
composed attractors. Periodic windows, in spite of their name, are in fact param-
eter regions in which transient chaos is typically present.
• Chapter 4: Noise and Transient Chaos. In systems subject to external random
forces, the attractor and the associated dynamics depend on the noise intensity.
The phenomenon that a dynamical system with simple periodic attractors be-
comes chaotic in the presence of noise is noise-induced chaos. It is due to the
transient chaotic dynamics coexisting with the periodic attractors in the noise-
free system, which become stabilized by noise. This chapter presents an extensive
treatment of the effects of noise on dynamical systems exhibiting transient chaos,
which is physically important because noise is inevitable in any realistic dynam-
ical systems.
Part II: Physical Manifestations of Transient Chaos. This part presents physical
manifestations of transient chaos in various natural systems. A striking aspect
of transient chaos is that it can lead to fundamental difficulties in predictabil-
ity. Chaotic scattering, the manifestation of transient chaos in open Hamiltonian
systems, will also be described both in classical and in quantum mechanics.
• Chapter 5: Fractal Basin Boundaries. If two or more periodic or chaotic attrac-
tors coexist, a trajectory may wander for a long time before approaching one
of the attractors asymptotically. When there is transient chaos on the boundaries
separating the basins of attraction, prediction of the final (asymptotic) state of the
system may not be possible. There can also be situations in which the boundaries
are severely interwoven (riddled basins), so that the motions on the boundaries
dominate the dynamics. Fractal basin boundaries or riddled basins cause a fun-
damental difficulty in predicting the asymptotic state of the system.
• Chapter 6: Chaotic Scattering. For scattering processes in open conservative sys-
tems the only way chaos can appear is in the form of transients, as a consequence
of the asymptotic freedom of the incoming and outgoing motions. Physical tra-
jectories are usually trapped in a scattering region of the configuration space for a
finite amount of time before leaving the system. Applications range from chemi-
cal reactions to celestial mechanics.
• Chapter 7: Quantum Chaotic Scattering and Conductance Fluctuations in
Nanostructures. This chapter deals with signatures of chaotic scattering when
the same system is treated quantum-mechanically in the semiclassical regime.
Scattering-matrix elements exhibit random fluctuations as some physical pa-
rameters of the system change. Depending on whether the classical scattering
viii Preface

is hyperbolic or nonhyperbolic, statistical properties of the fluctuations can be

quite distinct. One area in which quantum chaotic scattering finds significant
applications is electronic transport in semiconductor nanostructures.
Part III: High-Dimensional Transient Chaos. Although low-dimensional transient
chaos for which the underlying nonattracting chaotic sets have only one posi-
tive Lyapunov exponent is relatively well understood, high-dimensional transient
chaos generated by chaotic sets with multiple positive Lyapunov exponents re-
mains a forefront area of research in nonlinear dynamics. This part summarizes
what is known so far about high-dimensional transient chaos.
• Chapter 8: Transient Chaos in Higher Dimensions. The increase in the unstable
dimension from dimension one represents a highly nontrivial extension in terms
of what has been known about transient chaos. Topics treated include the di-
mension formulas, algorithms for computing high-dimensional chaotic saddles,
and chaotic scattering in physical systems with three degrees of freedom. In
high-dimensional dynamical systems, transients can differ from those in low-
dimensional systems in that the average lifetime is often extremely long before
the system settles into a final attractor, which is usually nonchaotic. The presence
of such transients implies that observation of the actual attractors of the system
is practically impossible. The basic scaling law characterizing the so-called su-
perpersistent chaotic transients and the effect of noise are treated.
• Chapter 9: Transient Chaos in Spatially Extended Systems. In a spatially ex-
tended system, transient lifetime often grows with the system size, and this
growth can be as fast as exponential, or even faster. The presence of such super-
long transients implies that the observed spatiotemporal behavior is not related
to chaotic attractors. Certain phenomena such as pipe turbulence may thus turn
out to exist on finite time scales only. An overview of transient chaos in spatially
extended dynamical systems and open issues is presented in this chapter.
Part IV: Applications of Transient Chaos. This part focuses on different aspects of
applications of transient chaos in physical, chemical, biological, and engineering
systems. A physical context in which transient chaos is ubiquitous is fluid sys-
tems. Another broad area of application is control and maintenance of transient
chaos for desirable system performance. The collection and analysis of transient
chaotic time series for probing the underlying system are also applicable in many
areas of science and engineering.
• Chapter 10: Chaotic Advection in Fluid Flows. The passive advection of tracer
particles (e.g., small dye droplets) in open hydrodynamical flows with uniform
inflow and outflow velocities turns out to be an appealing application of chaotic
scattering. The unstable manifold of the nonattracting chaotic set becomes a di-
rect physical observable in such cases as this manifold is traced out by particles
or pollutants while being advected downstream. These manifolds form the back-
bone of possible chemical and biological reactions taking place in the flow. The
transient-chaos-based approach to advection in fluid flows can have significant
applications in engineering and environmental sciences.
Preface ix

• Chapter 11: Controlling Transient Chaos and Applications. We demonstrate in

this chapter that transient chaos can be controlled by small perturbations. As
in the control of permanent chaos, an unstable orbit on the chaotic set can be
stabilized. A different form of control is to convert transient chaos into perma-
nent chaos. Applications presented include voltage collapse in electrical-power
systems and prevention, population control in ecology, and digital-information
encoding.
• Chapter 12: Transient Chaotic Time Series Analysis. For transient chaos, only
short time series are available, which makes the application of the methods de-
veloped in data analysis nontrivial. This chapter is devoted to basic issues in
transient chaotic time series analysis, which include delay-coordinate embed-
ding, and estimation of fractal dimension and Lyapunov exponents.
The main text is closed by a few final remarks. In the appendices, we treat a
number of technical issues such as multifractal spectra, open random baker maps,
semiclassical theory of chaotic scattering, and scattering cross sections.
To preview the applicability of the subject, we give in Table 1 a list of applications
of transient chaos in various disciplines, all of which will be treated (although not
in the same depth) in different chapters, including those outside of Part IV of this
book.
We try to give as broad as possible an overview. The field is, however, actively
developing, and full coverage of the literature is hardly possible by now. The selec-
tion of the material is therefore unavoidably biased, influenced by the authors’ own
experience.
We wish to thank all of our of colleagues with whom we had an opportunity for an
exchange of ideas on transient chaos. We are particularly grateful to our coworkers
for collaborative research. A particularly long record of joint publications binds both
of us to C. Grebogi. We thank E.G. Altmann, G. Csernák, A. Csordás, B. Eckhardt,
U. Feudel, M. Gruiz, G. Haller, D. Hensley, I.M. Jánosi, C. Jung, G. Károlyi,
Z. Kaufmann, A.P.S. de Moura, G. Stépán, and K.G. Szabó for insightful com-
ments on different chapters of the book during its preparation. E.G. Altmann, Y. Do,
M. Gruiz, I. Mezić, Sz. Hadobás, and M. Pattantyús-Ábrahám helped us by prepar-
ing some of the figures. In addition, YCL would like to thank Dr. Arje Nachman, at
the Air Force office of Scientific Research, for his wonderful support for research on
nonlinear dynamics and chaos. TT is grateful to the Hungarian Science Foundation
for its support by grant NK72037. We would like to express our thanks to the staff
of Springer Science and Media.

Phoenix and Budapest, 2009 Ying-Cheng Lai

Tamás Tél
x Preface

Table 1 Applications of transient chaos in different disciplines

Discipline Subject Chapters
Mathematics Continued fraction 2
Transfer operators 2
Almost invariant sets 2,10
Snapshot attractors and saddles, random maps 4,10
Leaked dynamics 2,10
Astronomy Escape of celestial bodies 6
Statistical physics Poincaré recurrences 2,6,7
Random systems and noise 2,4,11
Lobe dynamics 6,10
Transport processes 6,7,10
Optics Dielectric cavities 6,7
Lasers 6,12
Quantum mechanics Open quantum systems 7
Quantum echoes 7
Fractal Weyl law 7
Nanoscience Quantum dots 7
Graphene 7
Microfluidics 10
Fluid dynamics Stirring and mixing 10
Vortex dynamics 10
von Kármán vortex street 8,10
Turbulence 9
Engineering Shimmying wheels 2
Voltage collapse 11
Encoding digital information 11
Chemistry Classical molecular reactions 6,8
Reactions in open flows 10
Reaction–diffusion systems 9
Biology Population and plankton dynamics 6,10,11
Epidemiology and ecology 4
Food chains 11
Species extinction 11
Environmental sciences Spreading of pollutants 10
Lagrangian coherent structures 10
Convection in the Earth’s mantle 10
Advection of finite-size particles 10
Contents

Part I Basics of Transient Chaos

1 Introduction to Transient Chaos .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 3

1.1 Basic Notions of Transient Chaos . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 6
1.2 Characterizing Transient Chaos . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 9
1.3 Experimental Evidence of Transient Chaos . . . . . . . . . .. . . . . . . . . . . . . . . . . 25
1.4 A Brief History of Transient Chaos . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 34

2 Transient Chaos in Low-Dimensional Systems . . . . . . . . . . .. . . . . . . . . . . . . . . . . 37

2.1 One-Dimensional Maps, Natural Measures, and c-Measures . . . . . . . . 38
2.2 General Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 44
2.3 Examples of Transient Chaos in One Dimension .. . .. . . . . . . . . . . . . . . . . 48
2.4 Nonhyperbolic Transient Chaos in One Dimension
and Intermittency .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 55
2.5 Analytic Example of Transient Chaos in Two Dimensions . . . . . . . . . . 58
2.6 General Properties of Chaotic Saddles
in Two-Dimensional Maps .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 62
2.7 Leaked Dynamical Systems and Poincaré Recurrences . . . . . . . . . . . . . . 70

3 Crises. . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 79
3.1 Boundary Crises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 80
3.2 Interior Crises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 90
3.3 Crisis-Induced Intermittency . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 98
3.4 Gap-Filling and Growth of Topological Entropy . . . .. . . . . . . . . . . . . . . . .103

4 Noise and Transient Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .107

4.1 Effects of Noise on Lifetime of Transient Chaos . . . .. . . . . . . . . . . . . . . . .108
4.2 Quasipotentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .111
4.3 Noise-Induced Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .119
4.4 General Properties of Noise-Induced Chaos . . . . . . . . .. . . . . . . . . . . . . . . . .128
4.5 Noise-Induced Crisis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .132
4.6 Random Maps and Transient Phenomena .. . . . . . . . . . .. . . . . . . . . . . . . . . . .134

xi
xii Contents

Part II Physical Manifestations of Transient Chaos

5 Fractal Basin Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .147

5.1 Basin Boundaries: Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .148
5.2 Types of Fractal Basin Boundaries . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .149
5.3 Fractal Basin Boundaries and Predictability . . . . . . . . .. . . . . . . . . . . . . . . . .153
5.4 Emergence of Fractal Basin Boundaries . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .158
5.5 Wada Basin Boundaries .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .165
5.6 Sporadically Fractal Basin Boundaries .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .170
5.7 Riddled Basins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .175
5.8 Catastrophic Bifurcation of a Riddled Basin . . . . . . . . .. . . . . . . . . . . . . . . . .179

6 Chaotic Scattering .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .187

6.1 Occurrence of Scattering .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .188
6.2 A Paradigmatic Example of Chaotic Scattering . . . . .. . . . . . . . . . . . . . . . .190
6.3 Transitions to Chaotic Scattering .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .195
6.4 Nonhyperbolic Chaotic Scattering.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .211
6.5 Fluctuations of the Algebraic-Decay Exponent
in Nonhyperbolic Chaotic Scattering .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .222
6.6 Effect of Dissipation and Noise on Chaotic Scattering .. . . . . . . . . . . . . .230
6.7 Application of Nonhyperbolic Chaotic Scattering:
Dynamics in Deformed Optical Microlasing Cavities . . . . . . . . . . . . . . . .232

7 Quantum Chaotic Scattering and Conductance

Fluctuations in Nanostructures . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .239
7.1 Quantum Manifestation of Chaotic Scattering .. . . . . .. . . . . . . . . . . . . . . . .240
7.2 Hyperbolic Chaotic Scattering .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .242
7.3 Nonhyperbolic Chaotic Scattering.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .245
7.4 Conductance Fluctuations in Quantum Dots . . . . . . . . .. . . . . . . . . . . . . . . . .247
7.5 Dynamical Tunneling in Nonhyperbolic Quantum Dots . . . . . . . . . . . . .254
7.6 Dynamical Tunneling and Quantum Echoes in Scattering . . . . . . . . . . .259
7.7 Leaked Quantum Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .261

Part III High-Dimensional Transient Chaos

8 Transient Chaos in Higher Dimensions . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .265

8.1 Three-Dimensional Open Baker Map . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .266
8.2 Escape Rate, Entropies, and Fractal Dimensions
for Nonattracting Chaotic Sets in Higher Dimensions . . . . . . . . . . . . . . .268
8.3 Models Testing Dimension Formulas . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .274
8.4 Numerical Method for Computing High-Dimensional
Chaotic Saddles: Stagger-and-Step .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .282
8.5 High-Dimensional Chaotic Scattering.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .287
8.6 Superpersistent Transient Chaos: Basics . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .298
8.7 Superpersistent Transient Chaos: Effect of Noise
and Applications.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .305
Contents xiii

9 Transient Chaos in Spatially Extended Systems . . . . . . . . .. . . . . . . . . . . . . . . . .311

9.1 Basic Characteristics of Spatiotemporal Chaos . . . . . .. . . . . . . . . . . . . . . . .312
9.2 Supertransients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .316
9.3 Effect of Noise and Nonlocal Coupling on Supertransients .. . . . . . . . .321
9.4 Crises in Spatiotemporal Dynamical Systems . . . . . . .. . . . . . . . . . . . . . . . .323
9.5 Fractal Properties of Supertransients . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .329
9.6 Turbulence in Pipe Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .333
9.7 Closing Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .338

Part IV Applications of Transient Chaos

10 Chaotic Advection in Fluid Flows . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .343

10.1 General Setting of Passive Advective Dynamics . . . .. . . . . . . . . . . . . . . . .344
10.2 Passive Advection in von Kármán Vortex Streets. . . .. . . . . . . . . . . . . . . . .346
10.3 Point Vortex Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .351
10.4 Dye Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .358
10.5 Advection in Aperiodic Flows . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .362
10.6 Advection in Closed Flows with Leaks . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .370
10.7 Advection of Finite-Size Particles .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .373
10.8 Reactions in Open Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .377

11 Controlling Transient Chaos and Applications. . . . . . . . . . .. . . . . . . . . . . . . . . . .385

11.1 Controlling Transient Chaos: General Introduction .. . . . . . . . . . . . . . . . .386
11.2 Maintaining Chaos: General Introduction.. . . . . . . . . . .. . . . . . . . . . . . . . . . .392
11.3 Voltage Collapse and Prevention . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .395
11.4 Maintaining Chaos to Prevent Species Extinction .. .. . . . . . . . . . . . . . . . .399
11.5 Maintaining Chaos in the Presence of Noise, Safe Sets . . . . . . . . . . . . . .405
11.6 Encoding Digital Information Using Transient Chaos . . . . . . . . . . . . . . .407

12 Transient Chaotic Time-Series Analysis . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .413

12.1 Reconstruction of Phase Space . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .414
12.2 Detection of Unstable Periodic Orbits. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .421
12.3 Computation of Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .426
12.4 Computing Lyapunov Exponents from Transient
Chaotic Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .430

Final Remarks . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .435

A Multifractal Spectra .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .437

A.1 Definition of Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .437
A.2 Multifractal Spectra for Repellers of One-Dimensional Maps .. . . . . .437
A.3 Multifractal Spectra of Saddles of Two-Dimensional Maps . . . . . . . . .441
A.4 Zeta Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .442
xiv Contents

B Open Random Baker Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .445

B.1 Single Scale Baker Map.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .445
B.2 General Baker Map.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .447

C Semiclassical Approximation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .449

C.1 Semiclassical S-Matrix in Action-Angle Representation.. . . . . . . . . . . .449
C.2 Stationary Phase Approximation and the Maslov Index.. . . . . . . . . . . . .450

D Scattering Cross Sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .455

D.1 Scattering Cross Sections in Classical Chaotic Scattering . . . . . . . . . . .455
D.2 Semiclassical Scattering Cross Sections . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .457

References .. . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .459

Index . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .491
 Part I
Basics of Transient Chaos
Chapter 1
Introduction to Transient Chaos

In numerical or experimental investigations one never has infinitely long time

intervals at one’s disposal. In fact, what is needed for the observation of chaos is
a well-defined separation of time scales. Let t0 denote the internal characteristic
time of the system. In continuous-time problems, t0 can be the average turnover
time of trajectories on a Poincaré map in the phase space. In a driven system, it
is the driving period. In discrete-time dynamics, t0 can be the time step itself.
Suppose one observes signals that appear random for an average lifetime τ . Since
chaos is characterized by a sensitive dependence on initial conditions, which is
meaningful only on sufficiently long time scales, the appearance of chaotic signals
requires that τ be much greater than the internal characteristic time:

τ  t0 . (1.1)

The difference between sustained and transient chaos lies in the actual value of τ :
for the former, τ is infinite, but it is finite for the latter. As a matter of practicality,
one cannot exclude the possibility that a system apparently exhibiting a chaotic at-
tractor may turn out to be transiently chaotic if a much longer period of observation
is allowed. It is therefore useful to consider an additional time scale: the observa-
tion time TO . The sustained or transient nature of chaos then depends on how τ is
compared with TO . We can speak of transient chaos if

τ < TO . (1.2)

In the numerical investigation of attractors, a general habit is to discard a long

sequence of the trajectory in order to concentrate on the asymptotic properties.
A much richer dynamics may be observed, however, if one follows the trajecto-
ries from the beginning, i.e., if transients are not thrown out. One often finds then
complex dynamics over some time, different from the dynamics governed by the
attractor. The lifetime of a chaotic transient depends on the initial condition. An ex-
ample can be seen in Fig. 1.1, where transiently chaotic trajectories are shown from
the Hénon map [325, 564] at a parameter set where the attractor is a limit cycle.
Such signals can also be observed in experiments. An example is shown in
Fig. 1.2, where the measured quantity is the temperature difference between two

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 3
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 1,

c Springer Science+Business Media, LLC 2011
4 1 Introduction to Transient Chaos

a
xn
1

0,5

–0,5 50 100 150 200 250 300 350 400 450 n

–1

b xn
1

0,5

0
50 100 150 200 250 300 350 400 450 n
–0,5

–1

Fig. 1.1 Transient chaotic signals from the Hénon map xn+1 = 1 − ax2n + byn , yn+1 = xn for
parameters a = 1.25 and b = 0.3, with a period-7 attractor. For clear visualization, only every
seventh iterate is shown. (a) Trajectory initiated at x0 = 0.738816, y0 = 0.893088 exhibits chaotic
behavior over 441 iterates. (b) The initial condition is shifted by 2 · 10−19 in the x direction and the
length of the chaotic transient is only 126

Fig. 1.2 Transient chaotic signal of the temperature difference observed between two points of
an experimental loop of fluid heated from below with a constant heat flux (see Sect. 1.3 for more
details). In this run, chaotic oscillations last up to nearly 40 min [823] (with kind permission from
Elsevier Science)
1 Introduction to Transient Chaos 5

points in a fluid loop. Over some time chaotic temperature oscillations are observed,
which are accompanied by chaotic velocity oscillations of the laminar flow in the
loop, and then, rather suddenly, a crossover takes place towards a nearly constant
temperature difference corresponding to a uniform rotation of the fluid motion.
(For a list of other representative experiments, see Sect. 1.3.)
Based on these and many other examples, one concludes that transiently chaotic
signals (whose precise characterization will be discussed in Sect. 1.2) have the fol-
lowing characteristic properties:
1. For a fixed initial condition the signal appears chaotic up to certain time and then
switches over, often quite abruptly, into a different, often nonchaotic, behavior
that governs all the rest of the signal. The average lifetime, τ , can be obtained
from an ensemble of such observations, although for individual observations,
the actual lengths of transients depend sensitively on initial conditions: nearby
trajectories typically have drastically different lifetimes.
2. The probability distribution, P(t), of finding lifetimes longer than t is a smooth
function, which satisfies P(t) → 0 for t → ∞.
3. There exist infinitely long transients. Mathematically, however, the set of ini-
tial conditions leading to infinite transients has zero volume in the phase space
(has Lebesgue measure zero). Physically, this means that such infinite tran-
sients cannot be realized by initial conditions chosen randomly. In fact, for a
typical (i.e., randomly chosen) initial condition, the transient lifetime is finite.
Nonetheless, it is the presence of the measure-zero set of the initial conditions
with infinite transients which causes the random distribution of the transient life-
times for typical initial conditions.
4. It is known [564] that in a parameter region where chaotic attractors arise,
periodic windows are dense. That is, for a specific parameter value that leads to
a chaotic attractor, an arbitrarily small perturbation in the parameter can lead to
a periodic attractor. In this sense, chaotic attractors are not structurally stable.
Transient chaos is, however, robust against small parameter perturbations.
Similar to the fact that sustained chaotic signals are due to chaotic attractors
in the phase space, there exist chaotic invariant sets that are responsible for tran-
siently chaotic signals. Globally, such a chaotic set does not attract trajectories from
its neighborhood, and hence it is nonattracting. Nonattracting chaotic sets (chaotic
saddles or repellers; see Sect. 1.1.2) are therefore the phase-space objects that un-
derly transient chaos. We thus accept the following definition: transient chaos is the
form of chaos due to nonattracting chaotic sets in the phase space.
This chapter serves as a “first acquaintance” with transient chaos. The basic
properties of nonattracting chaotic sets will be described. The average lifetime and
the escape rate from these sets will then be introduced. Different methods for nu-
merically constructing nonattracting chaotic sets will be given. The construction of
the natural probability distribution on these sets will also be discussed, and an im-
portant related distribution, the conditionally invariant measure (c-Measure), will be
introduced, from which characterizing quantities such as the Lyapunov exponents
of the transients and dimensions of the nonattracting chaotic sets can be defined
6 1 Introduction to Transient Chaos

and calculated. To underline the scientific relevance of transient chaos, a list of

experiments taken from different disciplines will be presented, which also illustrate
different aspects of transient chaos. Finally, a brief history of transient chaos will
be given.

1.1 Basic Notions of Transient Chaos

1.1.1 Dynamical Systems

Dynamical systems are usually described by a set of ordinary differential equations:

dx
= F(x, p), (1.3)
dt
where x(t) is the vector characterizing the state of the system at time t and p repre-
sents a set of parameters. Alternatively, discrete-time dynamical systems, or maps,
of the form
xn+1 = f(xn , p) (1.4)
can be investigated, where xn is the state vector at discrete time n. Unless otherwise
stated, the map is assumed to be autonomous, i.e., f does not depend explicitly on n.
Maps can always be deduced from flows (1.3) by taking an appropriately defined
Poincaré surface of section or stroboscopic map [564], the latter corresponding to
repeatedly taking snapshots of the system at the multiples of some characteristic
time t0 . Using such maps, the phase-space dimension is reduced effectively by one,
facilitating visualization and analysis. In fact, Poincaré or stroboscopic maps have
been used commonly in numerical and laboratory experiments on transient chaos
(see Sect. 1.3). In order to have a consistent terminology, maps will be used for the
rest of the chapter to illustrate the basic dynamical properties of transient chaos, but
the main results apply also to flows (see also [398]).

1.1.2 Saddles and Repellers

The actual form of a nonattracting chaotic set depends on whether the dynamics is
invertible. A dynamical system is invertible if its motion can be uniquely followed
when time is reversed. This does not imply, however, that the time-reversed dy-
namics can actually occur in reality (although this is true for Hamiltonian systems,
which are invariant under time reversal if no external magnetic field or Corio-
lis effect is present). Dynamical systems described by differential equations are
typically invertible due to the uniqueness of solutions. Invertible dynamical sys-
tems are thus physically relevant. Noninvertible systems such as those described
by one-dimensional maps can, however, be quite useful models for understanding
specific features of transient chaos, and we shall consider them as well.
1.1 Basic Notions of Transient Chaos 7

In an invertible dynamical system, a typical nonattracting chaotic set repels

trajectories only along some special hypersurface in the phase space, which is called
the unstable manifold. Along a different invariant hypersurface, or the stable man-
ifold, the set can actually attract nearby trajectories. Usually, the local phase space
at a point in the chaotic set can be decomposed into the stable and the unstable
subspaces. For this reason, nonattracting chaotic sets in invertible dynamical sys-
tems are called chaotic saddles. Because differential equations are, in general,
invertible, and many real-life phenomena are described by differential equations,
transient chaos in experiments is typically related to chaotic saddles.
In contrast, for noninvertible dynamical systems in which the inverse is not
unique, nonattracting chaotic sets are often chaotic repellers, objects that are re-
pellent in all possible directions of the phase space. Chaotic repellers possess only
unstable manifolds. These considerations are summarized in Table 1.1. The geomet-
rical appearances of chaotic saddles and chaotic repellers can be quite different, as
Fig. 1.3 illustrates.
The dynamical difference between chaotic repellers and saddles is that long-lived
trajectories can start only from a neighborhood of the repeller, but for saddles

Table 1.1 Types of typical Dynamics Nonattracting chaotic set

nonattracting chaotic sets in
Invertible Chaotic saddle
nonlinear dynamical systems
Noninvertible Chaotic repeller

a b
y y
1 1

0.5 0.5

0 0

− 0.5 − 0.5

−1 −1

−1 − 0.5 0 0.5 1 x −1 −0.5 0 0.5 1 x

Fig. 1.3 Comparison of a chaotic saddle and a chaotic repeller. (a) A chaotic saddle from a
periodically kicked harmonic oscillator. On a stroboscopic plane the position xn and the veloc-
ity yn of the oscillator evolve according to the map [773] xn+1 = yn , yn+1 = 1 − 3.2y2n − 0.49xn .
(b) A chaotic repeller of the quadratic map zn+1 = zn 2 + 0.2 in the complex plane z = x + iy, which
is in fact a Julia set [824]. The saddle in (a) appears as a fractal set of points, which is in fact the
direct product of two Cantor-like sets, while the repeller in (b) is a complicated but nonetheless
continuous curve in the plane
8 1 Introduction to Transient Chaos

they can also start from a neighborhood of the stable manifold, a typically much
larger set. If a chaotic repeller and saddle coexist,1 transient chaos is primarily gov-
erned by the chaotic saddle.
Because a nonattracting chaotic set is invariant, trajectories starting from points
on the set never leave the set and in fact exhibit chaotic motion for infinitely long
time. However, because the Lebesgue measure of the set is zero, the probability
that a randomly chosen point of the phase space is in the set is zero. What is
observable is not the nonattracting set but a a small neighborhood of it. In par-
ticular, trajectories can originate from points in the vicinity of the set and can then
stay in the neighborhood of the set for a long but finite amount of time, and they
eventually leave the nonattracting chaotic set. These are the trajectories that gener-
ate transiently chaotic signals. The phenomenon of transient chaos thus illustrates
that the existence of a set of Lebesgue measure zero can be observed via finite-time
properties. As a consequence, we shall also see that the fractal features of a nonat-
tracting chaotic set are different from those of a chaotic attractor.
A related point is that the natural measure, a special invariant distribution char-
acterizing the dynamics on a nonattracting chaotic set, not only exists but can
be obtained approximately in numerical or actual experiments. In particular, the
distribution can be approximately specified on a small neighborhood of the set.
The approximate natural measure can then be used to perform ensemble averages of
physical quantities of interest, similar to the situation with chaotic attractors. Since
the distribution is only approximate, any ensemble average will contain errors, but
they can be controlled.
Transient chaotic dynamics can also be classified according to whether the
process is dissipative or conservative. In a strictly dissipative system where the
local phase-space volume contracts everywhere, the asymptotic states of the sys-
tem are attractors that may be regular, but transient chaos provides a “platform” for
approaching the attractors. In such a case the transient dynamics before the final
state of the system is reached is chaotic. In dissipative systems, transient chaos ap-
pears in the form of chaotic transients. In conservative or Hamiltonian systems, the
phase-space volume is constant under time evolution. As a result, there are no attrac-
tors, but some simple asymptotic states of the system can still be defined. Consider,
for example, a particle-scattering experiment in which the underlying dynamics is
Hamiltonian. Particles coming from far away approach the scattering region, and af-
ter a finite amount of time, they leave the region and escape to “infinity.” There can,
however, be qualitatively different exit routes to infinity. In this case, the different
exit routes can be regarded as asymptotic states (but not attractors) of the system.
The dynamics in the scattering region can, however, be regular or chaotic, where
the latter, i.e., transient chaos in Hamiltonian systems, defines the phenomenon of
chaotic scattering. Hamiltonian systems are invertible, so the nonattracting set un-
derlying chaotic scattering is typically a saddle.

1 For instance, in the time-reversed dynamics of an invertible system possessing a chaotic attractor

and a coexisting chaotic saddle.

1.2 Characterizing Transient Chaos 9

It should be noted that nonchaotic transients may also exist in dynamical

systems. An example is provided by trajectories that approach an attractor but are
far away from any nonattracting chaotic set. These transients are typically short and
do not exhibit chaotic features, although the actual asymptotic state may be chaotic.
Thus, transients to chaos can be quite different from chaotic transients, since the
latter, but not the former, are due to an underlying nonattracting chaotic set.

1.1.3 Types of Transient Chaos

According to the type of attractor(s) with which a nonattracting chaotic set coex-
ists, we can distinguish two main types of transient chaos. The first type is for the
case in which the coexisting attractor is simple, e.g., a periodic attractor. While the
asymptotic behavior of the system is relatively simple, the transients are chaotic.
Transient chaos arising in situations in which there is an attractor at infinity, and in
open Hamiltonian systems in which attractors are replaced by different exit routes
also exhibit this type of transient chaos.
The second type occurs when a nonattracting chaotic set coexists with a chaotic
attractor. In this case, there are two distinct forms of chaotic behavior. A signal
from the system typically exhibits one form of chaotic behavior, the one due to the
nonattracting set, on time scale τ , and then switches over to another form of chaos
asymptotically. A common situation is that the motion determined by the nonat-
tracting set is more chaotic than that due to the chaotic attractor (for more detail see
Fig. 1.16 and Chap. 3). Thus, focusing on the asymptotic properties will “miss” the
dominant chaotic part of the full complex dynamics that contains important infor-
mation about the underlying dynamical system.

1.2 Characterizing Transient Chaos

Having introduced the basic concepts of transient chaos in a qualitative manner,

we now discuss its quantitative characterization. A natural question is whether
there is actually chaos in the seemingly chaotic signals observed over finite time
scales. There are different levels of characterization of increasing complexity, as
follows:
1. Measurement of the lifetime distribution, the escape rate, and the average life-
time.
2. Construction of nonattracting chaotic sets in the phase space.
3. Construction of invariant measures on the chaotic set.
4. Determination of dynamical invariants such as the Lyapunov exponents and the
fractal dimensions of the nonattracting chaotic set and its natural measure.
10 1 Introduction to Transient Chaos

Following this hierarchy, one can find criteria to address the question of whether
the system is indeed chaotic and if so, to calculate some measure of the strength of
chaoticity. In the following we discuss these levels of characterization.

1.2.1 Escape Rate

In transient chaos, typical trajectories, i.e., trajectories initiated from random initial
conditions, escape any neighborhood of the nonattracting chaotic set. A quantity
measuring how quickly this occurs is the escape rate [824]. To define the escape
rate, imagine distributing a large number N0 of initial points according to some
initial density ρ0 in a phase-space region R that does not contain any attractor
or asymptotic state of the system. The density ρ0 is often chosen to be uniform,
and the geometry of R can be chosen to be simple, e.g., a rectangle in a two-
dimensional phase space. Many trajectories from the initial points may come close
to the nonattracting chaotic set at some later time. We define a restraining region Γ
as a bounded, compact region containing the nonattracting set. Once a point leaves
the restraining region, it cannot return to it. After visiting a neighborhood of the
set, almost all trajectories eventually leave Γ . Let N(n) denote the number of trajec-
tories remaining inside Γ after n steps, and choose N0 to be sufficiently large that
N(n)  1. As n is increased, one observes in general an exponential decay in the
number of trajectory points that are still in Γ (surviving points) [373, 596, 843]:
N(n) ∼ e−κ n for n  1, (1.5)
where κ is called the escape rate.2 A small value of κ implies weak “repulsion” of
typical trajectories by the nonattracting chaotic set. The escape rate turns out to be
independent of the distribution ρ0 of the initial conditions, of its support R, and of
the choice of the restraining region Γ . The escape rate κ is thus a property solely
of the nonattracting chaotic set. However, the prefactor of the exponential form in
(1.5), and the behavior of the system preceding the exponential decay do depend on
details such as the choices of ρ0 , R, and Γ .
A practical issue concerns about the choice of the support R of the initial density.
In a noninvertible system, R should overlap with the chaotic repeller, while in an in-
vertible system it is sufficient to choose R so that it overlaps with the stable manifold
of the chaotic saddle. In any case, if an exponential decay is found, its rate should
be given by the escape rate κ . In practice, the initial density is often distributed on
the restraining region, implying R = Γ .
In a realistic physical system, the exponential decay can be observed with high
accuracy after a finite, often short, time n∗ , i.e.,
N(n) = Ne−κ n for n ≥ n∗ , (1.6)

2 There are situations in which the decay follows a power law for certain types of nonhyperbolic

chaotic sets, which will be treated in Sect. 2.4 and Chap. 6. Such decays cannot be characterized
by escape rates.
1.2 Characterizing Transient Chaos 11

Fig. 1.4 Survival in the Hénon map xn+1 = 1 − ax2n + byn , yn+1 = xn for parameters a = 2.0 and
b = 0.3. Number N(n) of surviving trajectory points in the square defined by Γ : |xn |, |yn | ≤ 1.0,
obtained from N0 = 106 initial points distributed uniformly in the same square (R = Γ ). The fitted
dashed line has slope approximately −0.36, giving κ ≈ 0.36. The value of n∗ is approximately 4.
The survival probability P(n) is approximately N(n)/N0

where the value of n∗ and the prefactor N may also depend on ρ0 , R, and Γ .3
An example is shown in Fig. 1.4, where we see that the value of n∗ is relatively
small.
The definition of the escape rate indicates that the number of surviving points
is decreased by a factor of 1/e after about 1/κ time steps. This implies that most
trajectories do not live longer than 1/κ in the restraining region. It is thus reasonable
to estimate the average lifetime τ of the chaotic transient as

1
τ≈ . (1.7)
κ
Since the escape rate can be obtained by following the decay law over a finite
time interval, cf. (1.5), transient chaos of short average lifetime may be difficult to
identify. A condition for the practical observability of transient chaos is thus that κ
be small.
In a more general context, for any initial distribution on R and choice of Γ , we can
define the probability P(n) of finding survival times larger than n ≥ 1. The survival
probability P(n) is thus the probability of finding initial points that have not escaped
Γ up to time n, which can be approximated by N(n)/N0 for large N0 . In view of
(1.6), the decay of P(n) is exponential:

P(n) = ge−κ n for n ≥ n∗ . (1.8)

3 The prefactor N yields what the number of initial points would be if the decay were exponential

from the very beginning. Therefore N is different from N0 .

12 1 Introduction to Transient Chaos

A related probability is the escape-time distribution, p(n), the probability that a

particle escapes region Γ exactly in the nth iterate. This quantity can be estimated
as [N(n − 1) − N(n)]/N0 and is therefore the “density” of the cumulative distribu-
tion P(n). We have
∞
P(n) = ∑ p(n ). (1.9)
n =n+1

Being the “derivative” of an exponential function, the long-time behavior of p(n) is

also exponential and can be written in the form of (1.8) (with a different n∗ , but the
same escape rate).4
The average lifetime τ is defined as the average escape time, i.e.,
∞
τ ≡ n̄ = ∑ np(n). (1.10)
n=1

Since the distribution is not exponential for n < n∗ , the exact average lifetime τ
does depend on the choices of ρ0 , R, and Γ . Note that the estimate (1.7) does not
reflect this property.5 Since the average lifetime depends on many details, the escape
rate κ is a more appropriate characteristic of the decay process than τ . The escape
rate is a unique property of the underlying nonattracting chaotic set, in contrast to
the average lifetime, which also contains information about, e.g., the initial distri-
bution of particles. While the values of τ and 1/κ are typically different even for
slow decays, their scaling properties in terms of, for example, parameter changes
are usually the same.
There can be situations in which two (or more) nonattracting chaotic sets coex-
ist with different escape rates κ1 and κ2 . In such a case, the number of surviving
trajectory points in a given restraining region Γ is the sum of two exponentials for
large n:
N(n) ∼ N1 e−κ1 n + N2 e−κ2 n , (1.11)
and the prefactors Ni depend on the choices of ρ0 , R, and Γ .
It should be emphasized that the existence of a positive escape rate κ for
transients does not at all imply their chaoticity. One should also measure, for ex-
ample, the Lyapunov exponents on time scale 1/κ [714] and check whether at least
one of the exponents is positive. A complication is that even simple nonattracting
sets, for instance a single, regular saddle point (also called a hyperbolic point) are
at least partially repelling. Trajectories deviate from them exponentially. Regular

4 For continuous-time systems, (1.5)–(1.8) remain valid under the transform n → t. The escape-
time distribution becomes then a probability density, and the sum in (1.9) is replaced by an integral.
The escape rate in the corresponding continuous-time system is κ /t0 , where t0 denotes the internal
characteristic time mentioned in the introduction to this chapter. Analogously, the average lifetime
can be estimated as t0 /κ .
5 Equation (1.7) is a rough estimate, since even in the ideal case of n∗ = 1, when p(n) =

(exp (κ ) − 1) exp (−κ n), we obtain τ = (1 − exp (−κ ))−1 from (1.10) [147], which is consistent
with (1.7) for κ 1 only.
1.2 Characterizing Transient Chaos 13

Fig. 1.5 Lifetime function: n

dependence of the lifetime n
35
on the initial position x along
the interval defined by 30
y = −1.5 and |x| ≤ 1 in the
Hénon map at the parameters 25
of Fig. 1.4. (For the
corresponding phase-space 20
patterns, see Figs. 1.7 and 15
1.9.) The fractal irregularity
of this lifetime function is a 10
sign of transient chaos

−1 −0.5 0 0.5 1 x

nonattracting sets are therefore characterized by a positive Lyapunov exponent,

although the dynamics about them are not chaotic. The positivity of at least one
Lyapunov exponent is thus not sufficient for the chaotic behavior of transients. This
is why we accept the definition, used throughout the book, that transient chaos is the
dynamics associated with nonattracting chaotic sets.
To determine whether the transients are truly chaotic, one therefore needs more
information than the mere positivity of the Lyapunov exponent. Qualitatively, the
visual appearance of the signal can be helpful: about chaotic nonattracting sets
trajectories should be complicated. This is, nonetheless, only a hint. A property
uniquely indicating the chaotic nature of the transients is the irregular dependence
of lifetimes on initial conditions, as illustrated by Fig. 1.5. Suppose one starts tra-
jectories along a smooth curve in the phase space that intersects a chaotic repeller
or the stable manifold of a chaotic saddle. One then finds that for some points the
lifetimes are large. In principle, points of infinitely large lifetimes belong to a fractal
subset of initial conditions, since these must be points of the chaotic repeller or of
the saddle’s stable manifold. A fingerprint in a finite-accuracy numerical simulation
is large lifetimes separated by small values in between.

1.2.2 Constructing Nonattracting Chaotic Sets

Repellers are straightforward to construct, since they are the attractors of the in-
verted dynamical systems. Noninvertibility is generally due to the existence of more
than one inverted branch. When following the time-reversed dynamics, all possible
inverses should be taken into account.
For an invertible dynamical system, the calculation of chaotic saddles is more
delicate. While such a system can be inverted, the inverted dynamics still results in a
chaotic saddle. This feature can in fact be viewed as an illustration of the robustness
of the hyperbolic structure that is often seen for chaotic saddles. Roughly, a chaotic
saddle is the set of intersections between the stable and the unstable manifolds, and
14 1 Introduction to Transient Chaos

in hyperbolic cases, the angles at the intersecting points are bounded away from
zero. In what follows, we will describe an intuitive numerical procedure for calcu-
lating chaotic saddles, which serves to further illustrate their dynamical structures.
More practical numerical methods will then be introduced.

1.2.2.1 Horseshoe Construction

The intuitive method is based on the observation that a chaotic saddle has typically
embedded within itself a dense set of unstable periodic orbits, a property of any
chaotic set. Imagine that we choose an unstable periodic orbit in an invertible two-
dimensional map and plot its stable and unstable manifolds, which are the curves
along which the orbit is attracting in the direct and in the inverted dynamics, re-
spectively. If these curves cross each other once at a point (a homoclinic point),
they must do so infinitely many times, since the images and the preimages of such
an intersection are of the same type. All the homoclinic points form a homoclinic
orbit. Since it belongs simultaneously to the stable and the unstable manifolds of
the original periodic orbit, a homoclinic orbit approaches asymptotically, but can
never reach, the periodic orbit. As a result, the stable and unstable manifolds exhibit
a complex, intertwined structure, as shown schematically in Fig. 1.6. The horse-
shoe structure of the manifolds and the existence of homoclinic orbits have been
known since the works of Smale [300, 721]. Thus, mathematically, chaotic saddles
are closed, bounded, and invariant sets with dense orbits. They are the “soul” of
chaotic dynamics [721]. Similar to the formation of homoclinic orbits, the stable
(unstable) manifold of a periodic orbit can intersect with the unstable (stable) man-
ifold of a different orbit, forming a heteroclinic orbit. The stable and the unstable
manifolds of different periodic orbits of a chaotic saddle are usually close to each
other in the phase space, and all the resulting homoclinic and heteroclinic orbits
belong to the chaotic saddle.

Fig. 1.6 Horseshoe

structure: schematic
illustration of horseshoes
formed by the stable and the
unstable manifolds of a fixed
point (period-1 orbit) denoted
by the dot. The set of
intersection points
(homoclinic points) between
the manifolds belongs to the
chaotic saddle
1.2 Characterizing Transient Chaos 15

Fig. 1.7 A horseshoe 1.5

construction: a few branches x
of the stable and the unstable
manifolds of the fixed point 1
H+ of the Hénon map H+
xn+1 = 1 − 2.0x2n + 0.3yn ,
yn+1 = xn 0.5

− 0.5

−1

−1.5
−1.5 −1 − 0.5 0 0.5 1 1.5
y

The above discussion suggests the following procedure for numerically calculat-
ing a chaotic saddle. One first finds a simple hyperbolic orbit, such as a fixed point
or a periodic orbit of low period, and then calculates its stable and unstable mani-
folds. In particular, the unstable (stable) manifold can be obtained by distributing a
large number of initial points in a small neighborhood of the hyperbolic orbit and
iterating them under the forward (inverted) dynamics. The set of intersecting points
between the manifolds is part of the chaotic saddle. Since in practice, only a finite
number of branches of the manifolds can be constructed, the intersections provide
an approximate representation of the saddle. If the number of initial points used in
the calculation is reasonably large, the fractal nature of the saddle and its stable and
unstable manifolds can be revealed. An example is shown in Fig. 1.7. In general,
the appearance of a fractal geometry along both the stable and the unstable mani-
folds and the existence of a horseshoe type of structure are indications that a chaotic
saddle exists in the phase space of interest. Note that if the manifolds of the hyper-
bolic orbit chosen do not intersect each other, the orbit does not belong to a chaotic
saddle. In this case, it is necessary to choose a different periodic orbit to start with.

1.2.2.2 Ensemble Method

The idea of this method, introduced by Kantz and Grassberger [380], is to follow
an ensemble of trajectories and select the pieces that remain in the vicinity of the
saddle. In particular, one first chooses a region R close to the suspected chaotic
saddle but not containing any attractor, distributes uniformly a large number N0
of points in R, and iterates these initial conditions under the forward dynamics.
A criterion is needed for deciding when a trajectory is away from the saddle, which
can simply be that the trajectory moves out of a restraining region Γ surrounding the
16 1 Introduction to Transient Chaos

saddle (regions R and Γ can be the same as the respective ones used for computing
the escape rate). Another criterion can be [380] to calculate the effective Lyapunov
exponents over a finite number of time steps and examine whether they are close
to the corresponding exponents characterizing an attractor. In the case of a point
attractor, it is simply the negativity of all local Lyapunov exponents that can be used
as an indicator of the trajectory’s having left the saddle. All trajectories leaving the
saddle earlier than n0 steps are discarded, and trajectories of lifetime longer than
or equal to n0 are kept. The choice of the value of n0 can be somewhat arbitrary,
but some large value should be chosen if the lifetime τ of the chaotic saddle is
large. (Experience indicates that choosing n0 a few multiples of 1/κ is proper.)
One can then select long-lived trajectories in the neighborhood of the saddle to
approximate it. For example, if the desirable number of trajectories whose lengths
are not less than n0 is M0 , the number N0 of initial points should be of the order of
n0 M0 exp (κ n0 ), which can be a few orders of magnitude larger than M0 . To ensure
that trajectories close to the saddle are selected, the long-lived trajectories need to be
truncated at both the beginning and the end. For example, for a trajectory of length
larger than n0 , one can discard the first n1 and the last n2 points so that the resulting
trajectory is close to the saddle but not close to its stable and unstable manifolds,
respectively, where n1 and n2 are each a fraction of n0 . A representative example is
shown in Fig. 1.8.

1.5
x
1

0.5

−0.5

−1

−1.5
−1.5 −1 − 0.5 0 0.5 1
y 1.5

Fig. 1.8 Chaotic saddle in the Hénon map (a = 2.0, b = 0.3) obtained by the ensemble method,
where N0 = 106 initial points are distributed uniformly in the interval R = (| y0 |< 0.5, x0 = 0).
The restraining region is Γ =| xn |≤ 1.2. The first 10 and the last 20 steps of long-lived trajectories
are discarded (n0 = 30). Observe that the pattern is practically the same as the one formed by the
set of homoclinic points in Fig. 1.7. The direct product structure of two Cantor-like sets is a generic
characteristic of chaotic saddles of two-dimensional maps
1.2 Characterizing Transient Chaos 17

1.2.2.3 Sprinkler Method

A variant of the ensemble method, the sprinkler method [341], can be used to
calculate not only a chaotic saddle but also its stable and unstable manifolds [453].
Again, one starts from N0  1 trajectories distributed uniformly over a region R en-
closing at least a part of the saddle. One then chooses an iteration number n0 that is
several times larger than the estimated lifetime (1.7) of the saddle, and follows the
time evolution of each initial point up to exactly time n0 . Only trajectories that do
not escape R in n0 steps are kept, whose number is approximately N0 e−n0 κ . If n0 κ
is sufficiently large (but not so large that only a few points remain inside R), trajec-
tories with this long lifetime come close to the saddle in the course of dynamical
evolution, implying that their initial points will be in the immediate vicinity of the
stable manifold of the saddle, or of the saddle itself, and that their end points will be
close to the unstable manifold of the saddle. The latter is so because most points still
inside after n0 steps might already be in the stage of leaving the region. The points
from the middle of these trajectories (with n ≈ n0 /2) are then in the vicinity of the
saddle. In general, the initial, middle, and end points of trajectories of lifetimes of
at least n0 approximate the stable manifold, the saddle, and the unstable manifold,
respectively, within the region R, as exemplified by Fig. 1.9. In order to obtain the
full saddle, R should be chosen to fully contain the saddle (which corresponds to
the choice R = Γ in the ensemble method). An advantage of the sprinkler method is
that it is computationally fast and is easy to apply to high-dimensional cases.

1.2.2.4 Single-Trajectory (PIM-Triple) Method

The PIM-triple method, proposed by Nusse and Yorke [557], sets out to find a single
and arbitrarily long trajectory very much near the chaotic saddle. The procedure is

1.5
x a b c
1

0.5

−0.5

−1

−1.5
−1.5 −1 − 0.5 0 0.5 1y −1 −0.5 0 0.5 1 y −1 − 0.5 0 0.5 1 y 1.5

Fig. 1.9 Sprinkler method: an example of finding the chaotic saddle and its manifolds for the
Hénon map in Figs. 1.7 and 1.8 by the sprinkler method. Region R = Γ is a square of size 2 × 2
centered at the origin, N0 = 107 , κ = 0.36. Parts (a), (b), and (c) show points of trajectories with a
minimum lifetime n0 = 16 at iteration numbers n = 0, n = 8, and n = 16, well approximating the
stable manifold, the saddle, and the unstable manifold, respectively. Note that the stable manifold
here shows a more detailed structure than in Fig. 1.7
18 1 Introduction to Transient Chaos

based on the observation that trajectories starting close to the stable manifold of the
saddle stay for a long time in the vicinity of the saddle. The closer they start out
to the stable manifold, the longer their lifetime is. One begins by taking an interval
AB somewhere in a restraining region Γ so that it intersects the stable manifold of
the chaotic saddle. One next chooses initial points uniformly distributed on AB and
measures their lifetimes in Γ . A triplet of points, i.e., three points on AB such that
the midpoint has a lifetime longer than those of the two neighboring points, can then
be identified. Such a triplet is called a PIM-triple, where “PIM” stands for proper
interior maximum. For the PIM-triple whose middle point has the maximum life-
time, the two external points are expected to lie on two different sides of a branch of
the stable manifold. These two points then define a new interval A B that intersects
the stable manifold. One can then repeat this refining procedure to find a PIM-triple
whose size is smaller than a prescribed value δ 1. The triple can then be iter-
ated forward under the dynamics. Points on the triple approach the saddle along the
stable manifold but simultaneously move apart along the unstable manifold. When
the size of the evolving triplet becomes larger than δ , the refining procedure is ac-
tivated to reduce the size of the triple to within δ , and the new triple is iterated
forward again, and so on. One thus finds a set of PIM-triples of size less than δ ,
and the set of middle points can be taken as an approximation of a typical trajectory
on the chaotic saddle. While the PIM-triple procedure is computationally expensive
(as compared with the previous methods), it is applicable to chaotic saddles even
with relatively short lifetimes. The desirable feature of the PIM-triple method is that
it enables a long trajectory near the saddle to be found, facilitating characteriza-
tions of the saddle by dynamical invariants such as the Lyapunov exponents and the
fractal dimensions.
On a given interval, several PIM-triples can usually be found. If one selects from
those the PIM-triple for which the lifetime at the middle is the largest, not only can
the geometry of the saddle be revealed, but such a long trajectory also generates a
good approximation to the natural measure on the saddle [353].

1.2.3 The Invariant Measures of Transient Chaos

1.2.3.1 Natural Measure

The natural measure (or natural invariant distribution) μ characterizes how often
different pieces of a nonattracting chaotic set are visited by a dense trajectory on the
set [380]. This distribution can be obtained from a smooth distribution about the set
by selecting long-lived trajectories on the set. In practice, however, infinitely long
trajectories cannot be found, nor can such trajectories be calculated with arbitrarily
high precision. It is thus necessary to approximate the natural measure. For example,
one can choose a fine but finite coverage of the chaotic set by boxes is the phase
space of dimensionless linear size ε 1, and consider trajectory points that stay for
a long time close to the set.
1.2 Characterizing Transient Chaos 19

Fig. 1.10 Distribution of the

natural measure μ of the
Hénon chaotic saddle in
Fig. 1.8 in a two-dimensional
region in the phase space,
specified on a grid of size
ε = 1/400 (Picture by M.
Gruiz and Sz. Hadobás)

Let N be the number of such points. All nonempty boxes define a coarse-grained
version of the chaotic set at the resolution ε . The approximate natural measure on
this coarse-grained set is given by the frequencies of visits to different boxes by long
trajectories in the vicinity of the chaotic set. More precisely, the natural measure
Pi (ε ) of nonempty box i is
Ni
Pi (ε ) = , (1.12)
N
where Ni is the total number of trajectory points falling into box i. It is desirable to
take N large enough that the condition Ni  1 is satisfied in all nonempty boxes. The
quantity Pi (ε ) is also called the box probability and represents the natural measure
μ inside box i with an accuracy of order ε :

Pi (ε ) = dμ (x). (1.13)
x∈ box i

Points of long-lived trajectories can be generated either by the ensemble, the sprin-
kler, or the PIM-triple method. An example of the natural measure of a chaotic
saddle specified with finite resolution is shown in Fig. 1.10.

1.2.3.2 Conditionally Invariant Measure

A concept that is closely related to the natural measure is the c-measure intro-
duced by Pianigiani and Yorke [595, 596]. For a nonattracting chaotic set, this
measure is defined with respect to its unstable manifold. While there can be dif-
ferent c-measures for a given system [400], the “natural” c-measure is particularly
relevant; it can be obtained as the limiting distribution of iterating trajectories start-
ing from a smooth distribution of initial conditions about the chaotic set. In this
sense, the c-measure is the analogue of the SRB measure [564] for attractors. For
20 1 Introduction to Transient Chaos

transient chaos, the c-measure also describes how points deviate from the chaotic
set, and can be regarded as being maintained by supplying new points into the re-
gion of interest according to the rate at which trajectories escape from the region
asymptotically. Formally, this can be achieved by multiplying the number of points
everywhere by the constant eκ at each time step. Because of the contraction along
the stable direction (if it exists), the limiting distribution will be nonzero along the
unstable manifold only. This distribution, the c-measure, thus characterizes how
points leave the neighborhood of the underlying nonattracting chaotic set asymp-
totically, and how often certain regions are visited by trajectories in the process of
escaping. The natural measure μ can also be considered as the c-measure μc re-
stricted to a small neighborhood of the nonattracting chaotic set itself.
Under the map f, the region not escaping the restraining region Γ within one
iterate is the preimage f−1 (Γ ) of Γ . Since the c-measure characterizes how points
are distributed along the unstable manifold before escaping, from a probabilistic
point of view, it is the c-measure of f−1 (Γ ) which is proportional to the number of
trajectory points not escaping in one time step. By normalizing the c-measure of the
restraining region to be unity, μc (Γ ) = 1, one finds, as pointed out by Pianigiai and
Yorke [596], that the compensation factor is the reciprocal of μc [f−1 (Γ )], i.e.,

e−κ = μc [f−1 (Γ )]. (1.14)

The escape rate is thus uniquely expressed by the c-measure of the nonescaping
points within one iterate. Since the c-measure describes the asymptotic escape pro-
cess, which is purely exponential, in choosing the density ρc of the c-measure as
an initial density ρ0 on Γ , the exponential form of the survival probability (1.8) is
valid by the first time step. We see that ρc is thus the special initial distribution for
which n∗ = 1.
The c-measures of one-dimensional maps will be studied in detail in the next
chapter. Here we briefly describe the construction of the c-measure for invertible
two-dimensional maps. For such a system, the c-measure can be defined as fol-
lows [341]. Since trajectories escape the chaotic saddle along the unstable manifold,
after n  1 iterations, the surviving trajectory points of number Nn in Γ will be in the
vicinity of the unstable manifold. Let B be a small box within Γ that contains part
of the unstable manifold. The c-measure along the unstable manifold in B is thus

Nu,n (B)
μc (B) = , for n  1, (1.15)
Nn

where Nu,n (B) is the number of the Nn points that fall in B at time n. An example of
the c-measure is shown in Fig. 1.11. A comparison with the natural distribution in
Fig. 1.10 indicates that the singular spikes of the c-measure fall outside the saddle,
and that the natural distribution is nothing but the c-measure restricted to the saddle
(with a proper normalization).
(r)
The natural measure μ of the chaotic saddle can also be defined based on Nn (B),
the number of trajectory points in B at a time rn between zero and n:
1.2 Characterizing Transient Chaos 21

Fig. 1.11 Conditionally

invariant measure μc of the
Hénon saddle (a = 2, b = 0.3)
on the restraining region
Γ :| x |, | y |≤ 1.5, identified
on a grid of size ε = 1/400.
The support of the c-measure
is the unstable manifold of
the saddle (cf. Fig. 1.7)
(Picture by M. Gruiz and
Sz. Hadobás)

(r)
Nn (B)
μ (B) = , where 0 < r < 1, n  1. (1.16)
Nn

(1)
We have Nn (B) = Nu,n (B). For large N0 and n, trajectories remaining in Γ will
stay near the chaotic saddle for most of the time between zero and n, except at the
beginning, when they are attracted toward the saddle along the stable manifold, and
at the end, when they exit along the unstable manifold. Thus, the measure defined
in (1.16) is independent of r, insofar as r deviates considerably from both 0 and 1.
A measure along the stable manifold can be defined as

Ns,n (B)
μs (B) = , for n  1, (1.17)
Nn

where Ns,n (B) is the number of initial conditions in B whose trajectories do not
leave Γ before time n (Fig. 1.12). Formally, this corresponds to evaluating (1.16)
at r = 0. Measure μs can in fact be regarded as the c-measure of the time-reversed
dynamics,6 and a restriction of this measure to the saddle provides the natural mea-
sure of the saddle in the time-reversed dynamics. When plotting the measure of the
two manifolds together, as in Fig. 1.13, one notes the following two features: (1) the
singularities of the c-measure fall outside the saddle, an indication of the hyperbolic
nature of the saddle, and (2) the restriction of the stable manifold measure to the sad-
dle is different from the natural measure (Fig. 1.10, and red columns in Fig. 1.13).
Both restricted measures are defined on the chaotic saddle, but they agree only if the
dynamics is invariant under time reversal (e.g., as in Hamiltonian problems).
Note that the sprinkler method (Sect. 1.2.2.3) for the construction of the saddle
(r)
and its manifolds is based on the concept of numbers Nn , and corresponds to the
particular choice r ≈ 1/2 for plotting points of the saddle.

6 Because of the analogy between μc and μs , we can also call the conditionally invariant measure
the measure of the unstable manifold.
22 1 Introduction to Transient Chaos

Fig. 1.12 Measure along the

stable manifold (cf. Fig. 1.7)
of the Hénon saddle for a = 2,
b = 0.3, identified on a grid
of size ε = 1/400. The region
Γ is | x |, | y |≤ 1.5 (Picture by
M. Gruiz and Sz. Hadobás)

Fig. 1.13 Measure of the

stable and the unstable
manifolds. The natural
measure of the Hénon saddle
is shown in red. The
distribution in red is the same
as that of Fig. 1.10 but the
spatial view is different. The
restriction of the stable
manifold’s measure to the
saddle differs from the natural
measure. (Picture by
M. Gruiz and Sz. Hadobás)

1.2.3.3 Characterization of the Natural Measure

Both the nonattracting set and its natural measure can possess complicated struc-
tures. To characterize the natural measure by certain numbers or simple functions is
therefore of interest. In fact, such convenient characteristics are used widely in the
study of chaotic attractors. While the characteristics can be worked out for any type
of invariant distributions on the nonattracting set, we discuss here characterization
of the natural measure, since it is physically most relevant. The typical character-
istics are the Lyapunov exponents, the box-counting and information dimensions,
the metric and the topological entropies, which are the dynamical invariants that we
shall focus on in this book. For a more detailed analysis, full spectra of Lyapunov
exponents, dimensions, and entropies can be introduced (see Appendix A). For
simplicity we assume that there is only one expanding direction in the system. More-
complicated cases will be treated in Chap. 8.
Consider first the Lyapunov exponent. Take a small interval of length Δ0 along
the unstable direction in a nonempty box i. It will be mapped after n steps onto a
larger interval of some length Δn . The stretching factor Δn /Δ0 can be written as
exp [Λ1i (n)], where the positive quantity Λ1i (n) is the stretching exponent belonging
to box i. Here Λ1i (n)/n plays the role of a local Lyapunov exponent. The average
1.2 Characterizing Transient Chaos 23

Lyapunov exponent λ1 is simply the average of the stretching exponent with respect
to the natural measure divided by n, i.e.,
1
n∑
λ1 = Λ1i (n)Pi (ε ) for n  1. (1.18)
i

It should be noted again that the positivity of the Lyapunov exponent is not a defining
characteristic for transient chaos, since any kind of nonattracting sets, e.g., unstable
fixed points, can have a positive Lyapunov exponent. For an isolated saddle (hyper-
bolic) point one has, e.g., λ1 = κ > 0.
The fractal properties of the nonattracting chaotic set and its natural measure
describe how quantities scale with the box size ε . The box-counting dimension D0
reflects how rapidly the number N(ε ) of nonempty boxes of dimensionless size ε
covering the set increases with refining resolution:

N(ε ) ∼ ε −D0 for ε 1. (1.19)

If the set is covered by boxes of different sizes εi , i = 1, . . . , N(ε ), which are all
bounded from above by an ε 1, then (1.19) can be generalized to yield [45] the
following implicit equation for D0 :
N(ε )
∑ εiD0 = 1 for ε 1. (1.20)
i=1

This dimension characterizes only the geometry of the nonattracting set.

The information dimension is a measure of the inhomogeneity of the natural
distribution. It measures how the information content of the box probabilities Pi (ε )
changes with the resolution:
N(ε )
− ∑ Pi (ε ) ln Pi(ε ) = D1 ln (1/ε ) for ε 1. (1.21)
i=1

When the coverage consists of unequal small boxes of different sizes εi , but all
bounded from above by some ε 1, the information dimension D1 can be expressed
[45] by Pi , the probability that box i of size εi is visited, as

∑i Pi ln Pi
D1 = for ε 1. (1.22)
∑i Pi ln εi

The information dimension in fact belongs to a subset of the nonattracting set, the
one that contributes dominantly to the information − ∑i Pi ln Pi . It therefore cannot
be greater than the box-counting dimension:

D1 ≤ D0 . (1.23)

The equality holds only when the distribution is uniform: Pi (ε ) = constant on the
nonattracting set.
24 1 Introduction to Transient Chaos

Typically, one can associate a few symbols to different regions containing the
nonattracting set, and assign a symbol when the trajectory visits a given region. This
defines a symbolic representation of trajectories on the set [45,220,564]. By follow-
ing trajectories of length m about the nonattracting set, one can specify how often a
given symbolic sequence {S j }m 1 occurs. These path probabilities P({S j }) provide a
complementary characterization of the chaotic set: entropies.
In particular, the metric entropy K1 is defined as the growth rate of the in-
formation content of the path probabilities with length m of symbolic sequences
[45, 220, 564]:

− ∑ P({S j }) ln P({S j }) = K1 m for m  1, (1.24)

{S j }

where the summation is taken over all symbolic sequences. Since the path proba-
bilities depend on the natural measure, the metric entropy is also a characteristic
of this measure. In terms of an information-theoretic interpretation [283], the metric
entropy is the rate at which information stored in the insignificant digits of the initial
condition flows toward the significant ones with time. The Lyapunov exponent λ1 is
in fact the mean velocity of this flow.
The topological entropy [2] K0 reflects how complicated the organization of the
symbolic encoding is: it is the growth rate of the number Ωm of all allowed symbolic
sequences of length m:
Ωm ∼ eK0 m for m  1. (1.25)
Here the different symbolic sequences are not weighted in terms of the path proba-
bilities, whence the term “topological.” The topological entropy can be also defined
as the growth rate of the number Nm of all points of unstable periodic orbits of
length m:
Nm ∼ eK0 m for m  1. (1.26)
A straightforward method for numerically calculating the topological entropy
is due to Newhouse and Pignataro [548], which is based on the stretching of line
segments in two-dimensional maps. Let Ln denote the length of the nth image of a
line segment of initial length L0 falling within some restraining region Γ enclosing
the nonattracting set. One has

Ln
∼ eK0 n for n  1. (1.27)
L0

Similar to the relation between the fractal and the information dimensions (1.23), we
have the following inequality between the topological and the metric entropies [45]:

K1 ≤ K0 , (1.28)

where the equality holds only for the special case in which all the symbolic se-
quences are equally probable.
1.3 Experimental Evidence of Transient Chaos 25

An important feature of transient chaos is that there are exact and simple relations
among the escape rate of the underlying nonattracting chaotic set, the information
dimension, the Lyapunov exponents, and the metric entropy. In particular, one has

κ = g(λ1 , K1 ) (1.29)

and
D1 = gD (λ1 , K1 ). (1.30)
The particular forms of the right-hand sides of (1.29) and (1.30) depend on the
dimensionality of the system. For maps of arbitrary phase-space dimensions, ex-
plicit expressions for g and gD can be obtained, which we shall derive in later
chapters.
Summarizing briefly, so far we have given, in terms of the quantities introduced,
the criteria based on which the existence of a nonattracting chaotic set and conse-
quently transient chaos can be established: (1) positivity of the topological entropy
(K0 > 0), and (2) fractality of the nonattracting set (noninteger values for the box-
counting dimension D0 or the information dimension D1 ).7
That period three implies chaos [483], or equivalently, the existence of unstable
cycles of infinite length, in fact implies the positivity of the topological entropy. It is
then true that period-3 implies transient chaos.

1.3 Experimental Evidence of Transient Chaos

There has been ample experimental evidence of transient chaos. Here we shall
present results from a few such experiments. In terms of quantification, earlier ex-
periments mainly focused on determining the average transient lifetime, while more
recent works have extended to characterizing the nonattracting chaotic set and its
natural measure.

1.3.1 Convection Loop Experiment

The apparatus consists of a loop-shaped reservoir filled with water, which is

heated from below along the lower semicircle and cooled from above, as shown in
Fig. 1.14. The system was investigated from the point of view of chaotic dynamics
by Widmann, Gorman, and Robbins [273, 274, 823], and by Bau and coworkers
[822]. The system is the one-dimensional analogue of the Rayleigh–Bénard con-
vection problem, and its dynamics can be described in certain parameter regimes

7 The positivity of the largest Lyapunov exponent cannot be taken as a criterion because of the

example of an isolated saddle point.

26 1 Introduction to Transient Chaos

Fig. 1.14 Schematic diagram of the convection loop experiment. Representative results from the
experiment are shown in Fig. 1.2 [823] (with kind permission from Elsevier Science)

by the Lorenz model [488]. After convection sets in at sufficiently high heat flux,
the velocity of the fluid along the loop changes its sign and magnitude in a chaotic
manner. There is a broad range of system parameters for which the chaotic oscilla-
tions last for a finite amount of time before settling into a state in which the system
ceases to oscillate, as exemplified by Fig. 1.2. The duration of the transient chaotic
oscillations depends sensitively on the initial state of the system.

1.3.2 Chemical Reactions Preceding Thermal Equilibrium

Stirred chemical reactions in closed containers cannot be chaotic in a sustained

manner, since the system typically approaches thermal equilibrium after a transient
period of time, corresponding to a fixed-point attractor in the space of the concen-
trations. Scott, Showalter and coworkers conjectured, however, that the approach
toward equilibrium, i.e., the reaction dynamics before settling into the final fixed-
point attractor, can be chaotic [700]. The conjecture was experimentally verified by
Wang, Sorensen, and Hynne [820, 821] using the Belousov–Zhabotinsky reaction.
Chaotic oscillations as the light transmission have been observed (Fig. 1.15) over
several hours before a stage close to thermal equilibrium is reached.

1.3.3 Nuclear Magnetic Resonance Laser Experiment

The high quality of the data from an nuclear magnetic resonance (NMR) laser makes
it appropriate for experimental investigations of a variety of chaotic phenomena. The
laser output is a time-dependent voltage signal, and the control parameter is usually
the modulation amplitude. At a bifurcation called a crisis (see Chap. 3), a chaotic
1.3 Experimental Evidence of Transient Chaos 27

Fig. 1.15 Transiently chaotic oscillations (which start after about 1.5 h and continue for about 7 h)
in a closed Belousov–Zhabotinsky reaction ending with small-scale oscillations that ultimately
diminish as the system approaches thermal equilibrium [820] (Copyright 1994 by the American
Chemical Society)

attractor undergoes a sudden explosion in its size. In particular, slightly before the
crisis only a small-size attractor exists. In this regime, before settling into the small
attractor, trajectories started from random initial conditions exhibit chaotic motion
in the region where the postcrisis attractor lives, signifying transient chaos coex-
isting with permanent chaos. Time-series analysis of transient chaos [356] led to a
successful reconstruction of the chaotic saddle responsible for the observed tran-
sient behavior, which was accomplished by considering only the parts of the signals
that do not belong to the small attractor. By combining a number of such truncated
signals, a long time series can be generated that is amenable to chaotic time-series
analysis [382]. In this way, not only can the chaotic saddle be reconstructed, but
also its dynamical characteristics, such as the average Lyapunov exponent, can be
determined (Fig. 1.16).

1.3.4 Driven Pendulum

The parametrically driven damped pendulum is another example in which high-

precision experiments [479, 480] on transient chaos can be carried out. In such
an experiment, the angle and the angular velocity of the pendulum are measured,
with the damping constant as a bifurcation parameter. The chaotic saddle can be
reconstructed using a stroboscopic map. The experiment demonstrates, explicitly,
fractality along both the stable and the unstable manifolds (Fig. 1.17). The box-
counting dimension of the saddle was determined to be about D0 = 1.7. For a more
recent pendulum experiment, see [170]
28 1 Introduction to Transient Chaos

Fig. 1.16 Reconstruction of the chaotic sets from a nuclear magnetic resonance (NMR) laser
experiment via time series analysis. (a) Chaotic saddle, (b) coexisting chaotic attractor. Note that
the saddle not only appears larger, it is also more chaotic: its average Lyapunov exponent is about
twice as large as that of the attractor [356] (Copyright 1994 by the American Physical Society)

Fig. 1.17 A chaotic saddle reconstructed from the driven-pendulum experimental data. The double
Cantor set character can be seen through the blank saps amid the points [480] (with kind permission
from Elsevier Science)

1.3.5 Fractal Basin Boundaries

The boundaries between basins of attractions can often be fractal in nonlinear dy-
namical systems, and transient chaos can arise in a phase-space region containing
such boundaries. Experimental observation of fractal basin boundaries can be quite
1.3 Experimental Evidence of Transient Chaos 29

Fig. 1.18 For an electric

circuit system of four coupled
chaotic oscillators: basin of
the synchronous chaotic
attractor in black, and basin
of one of the periodic
attractors in white. The
horizontal and vertical
coordinates are proportional
to the initial deviations of two
coordinates of the oscillators
from a point on their common
chaotic attractor. Points of the
white basin are dense in the
black basin: the black basin is
riddled [322] (copyright
1994, the American Physical
Society)

challenging because of the necessity to choose initial conditions on a fine scale and
to monitor each trajectory until it approaches one attractor. An experimental inves-
tigation of an extreme type of fractal boundaries is due to Heagy et al. [322], who
studied a system of a set of four weakly coupled, nearly identical oscillators. Each
isolated oscillator exhibits a chaotic attractor, and the attractor associated with the
whole coupled system corresponds to a synchronous motion of the chaotic oscilla-
tors. The coupled system also possesses periodic attractors. Figure 1.18 shows the
basins of the synchronized state (set of black points) and of one of the periodic at-
tractors (white regions). The black basin exhibits a rather special property: it is the
riddled basin where every point in the black basin has points of the white basin
arbitrarily nearby.

1.3.6 Advection in the Wake of a Cylinder

In two-dimensional laminar flows around some obstacles, von Kármán vortex streets
are typically formed. Due to the periodic detachment of vortices in the wake, the
flow is periodic in time. The advection of particles is generally chaotic in time-
periodic flows. In particular, since strong time-dependence is restricted to the wake
only, advective chaos is transient, as illustrated in an experiment with a towed cylin-
der by Sommerer, Ku, and Gilreath [725]. The physical space of the fluid motion
happens to coincide with the phase space, rendering directly observable fractal pat-
terns that usually exist in the phase space. Of particular interest is the unstable
manifold of the chaotic saddle in the wake, since dye particles flow away along this
manifold. The experiment not only illustrated that dye (or pollution) often spreads
out along fractal patterns, but also provided an example in which a fractal unstable
manifold can be seen even by the naked eye, as shown in Fig. 1.19.
30 1 Introduction to Transient Chaos

Fig. 1.19 For the experiment

of advection in the wake of a
cylinder (black disk), fractal
pattern traced out by
spreading a dye droplet the
unstable manifold of a chaotic
saddle existing in the wake.
The flow is from left to right,
and the droplet is injected into
the upstream of the flow. The
lower panel shows the results
of a numerical simulation of
the same problem [725]
(copyright 1996, the
American Physical Society)

1.3.7 Semiclassical Fluctuations in Chaotic Scattering

Interference effects of the scattering process become important in the semiclassical

regime where wave properties are observable. Chaotic wave scattering, the scatter-
ing of waves from systems for which the underlying classical dynamics is chaotic, is
observable not only in nanoscale electronic devices, but also in microwave scattering
from macroscopic objects. Doron, Smilansky, and Frenkel [204] investigated the re-
flection of microwaves from an elbow-shaped cavity, where the underlying classical
ray dynamics is chaotic with an exponential decay of rate κ in the survival probabil-
ity of particles in the cavity. The basic quantity characterizing the wave-scattering
process is the scattering matrix S(ω ) as a function of the frequency. According to
the semiclassical theory, the squared modulus of its Fourier transform S(t) tends to
decay with the classical escape rate κ . This is consistent with the experimental find-
ings, as shown in Fig. 1.20. It can also be seen that the absolute value of the squared
frequency-dependent autocorrelation function is a Lorentzian of half-width κ .
1.3 Experimental Evidence of Transient Chaos 31

Fig. 1.20 For microwave scattering from an an elbow-shaped cavity, squared modulus of the
time-dependent S-matrix, | S(t) |2 (oscillating curve), exhibiting an overall decay that agrees with
that of the classical escape-time distribution (curve in background). Inset: squared modulus of the
frequency- (energy-) dependent autocorrelation function from the measured data (continuous line)
and from the semiclassical theory (dashed line). The dimensionless half-width is κ = 0.1 [204]
(copyright 1990, the American Physical Society)

1.3.8 Emission of Light from Dielectric Cavities

Dielectric cavities of cylindrical or spherical geometry are of technological impor-

tance because they can keep light trapped for a long time. When light circulates
almost tangent to the surface of such a cavity via total internal reflection, it suffers
minimal loss. Slightly deformed cavities emit light of nonnegligible intensity, and
it was observed with surprise that these emissions are peaked in certain directions.
A measurement by Schwefel and coworkers [699] records the light intensity for
billiard-shaped cavities as a function of two angles: the angle Φ along the sidewall
of the billiard, and an angle θ by which the camera is rotated in the far field from
the major axis. Figure 1.21 shows the result for three different shapes with a defor-
mation corresponding to approximately the same major-to-minor-axis ratio. Light
intensity is localized to certain regions of the (Φ , θ ) plane, meaning that light is
emitted at specific points of the sidewall only, and in an approximately predeter-
mined direction. Although the billiard shapes (shown as insets) appear similar in
the last two cases, the emission patterns are rather different: the intensity integrated
over the sidewall angle is peaked, e.g., about 90◦ and 30◦ . In the geometrical optics
approximation, the patterns can be explained by the existence of a chaotic saddle
of exceptional light rays that are never transmitted, i.e., they are subject to perma-
nent total internal reflection both forward and backward in time. The form of these
saddles depends sensitively on the shape of the billiard. In addition, the directional-
ity of the light has been shown to be intimately related to the unstable manifold of
32 1 Introduction to Transient Chaos

Fig. 1.21 Light emission intensity in false color (color bar at the sides) as a function of the side-
wall angle Φ and of the camera angle θ for three microcavities, whose shapes are drawn as insets
[699] (copyright 2004, the Optical Society of America)

certain unstable fixed points [699], and more rigorously, of the entire chaotic saddle
[15]. This is thus a further example in which the unstable manifold of chaotic saddle
become related to physical observables.

1.3.9 Maintaining Chaos in a Magnetoelastic Ribbon

Since there are systems that require chaos in order to function properly, it is im-
portant that transient chaos be convertible into permanent chaos. The aim of this
procedure, which is called maintenance of chaos, is to intervene in the dynamics in
such a way as to keep chaotic behavior alive in situations in which it would naturally
be absent. A possible realization is to apply properly chosen perturbations to the
signal in order to keep it always on that side of the stable manifold of a hyperbolic
point from which a return to a nonattracting chaotic set is allowed. The method can
be improved by finding a target point on this side from which a transiently chaotic
trajectory of long lifetime is initiated, and trying to direct the signal to the target
point [697]. In an experiment by In et al. [351], chaos in a magnetoelastic ribbon
was successfully maintained at a parameter set where a fixed-point attractor exists;
see Fig. 1.22. The quantity Xn plotted is proportional to the position of a point on the
ribbon measured once every driving period of an applied external magnetic field.

1.3.10 Turbulence in Pipe Flows

In a pipe flow, a steady laminar solution is linearly stable for all Reynolds num-
bers Re ≡ UD/ν , where U is the average velocity in a pipe of diameter D, and ν is
the kinematic viscosity. The turbulent state can be considered as a high-dimensional
chaotic state associated with a chaotic saddle. In an experiment, Peixinho and Mullin
followed turbulent puffs downstream and measured their positions along the pipe
1.3 Experimental Evidence of Transient Chaos 33

Fig. 1.22 Maintaining chaos in an experiment by In et al. [351]. (a) Projection of the stabilized
chaotic attractor. (b) Blowup of the square shown in (a). The filled circle on the diagonal marks the
hyperbolic point and the filled square marks the target. Small circles 1 –8 show an unperturbed
sequence and small squares 1–8 illustrate the perturbed sequence (Copyright 1998, the American
Physical Society)

Fig. 1.23 For an experiment

of turbulence in a pipe flow,
probability of observing a
turbulent puff as a function of
the dimensionless
downstream distance from the
point where the puff is
generated. The numbers
associated with different
experimental curves denote
the Reynolds number Re. It
can be seen that the escape
rate decreases with Re [589]
(Copyright 2006, the
American Physical Society)

where the puffs are relaminarized [589]. The distributions exhibit exponential de-
cay, as shown in Fig. 1.23. Normalized by length D and time D/U, the dimensionless
distance and the dimensionless time to reach this distance are proportional to each
other, i.e., the dimensionless velocity is of the order of unity. Turbulence in pipe
flows is thus a high-dimensional chaotic transient with finite lifetime. Recent inves-
tigations indicated that the average lifetime tends to grow extremely rapidly with
the Reynolds number [334, 336].
34 1 Introduction to Transient Chaos

1.4 A Brief History of Transient Chaos

The first observation of chaotic transients was part of the prehistory of chaos
science. In the late 1940s, in their early studies of the forced Van der Pol oscillator,
Cartwright and Littlewood [114] and Levinson [482] found signatures of chaotic
transients as the system settles into one of the coexisting attractors (there are in fact
fractal basin boundaries between the basins of attraction). Later, in 1973, Chirikov
and Izraelev identified certain transient features in weakly dissipative systems
[134, 135].
A systematic investigation of transient chaos began in the late 1970s with the
works of Shimizu and Morioka [714], Kaplan and Yorke [386], and Yorke and Yorke
[843] on the dynamics of the Lorenz system in parameter regimes that differ from
the standard one with a chaotic attractor, where the attractors are either limit cy-
cles or fixed points. An important step toward a firm mathematical foundation of
the phenomenon was the introduction of the concept of the conditionally invariant
measure by Pianigiani and Yorke [595, 596]. Subsequently, several theoretical pa-
pers reported this phenomenon in all kinds of nonlinear systems: low-dimensional
maps [141, 184, 500], nonlinear oscillators [337, 338, 512], systems modeled by
time-delayed equations [380], partial differential equations [348, 715], and coupled
oscillators [795].
A comprehensive investigation of transient chaos originated from the discovery
that chaotic transients arise typically in systems passing through a type of global
bifurcation called crisis (Grebogi, Ott, and Yorke in 1983 [293]). The Maryland
Chaos Group has played since then an important role in the understanding of further
transient-chaos-related phenomena, which include fractal basin boundaries, Wada
boundaries [406, 780], and riddled basins [11]. The importance of the natural mea-
sure on nonattracting chaotic sets and of quantities related to it was pointed out by
Kantz and Grassberger in 1985 [380] and later by Hsu, Ott, and Grebogi [73, 341],
generating further interest in the topic (see reviews in [766] and [770]).
Interest in the chaotic aspects of scattering processes dates back to early studies
of classical chemical reactions [56,275,555,622] and point–vortex interactions [25].
A systematic study of the subject began in the late 1980s with the work of Jung
[360], Eckhardt [210, 211], Hénon [326], and Bleher, Ott, and Grebogi [73, 74].
The concept of chaotic advection, coined by Aref [26] and generalized to open
hydrodynamical flows, corresponds to a transiently chaotic spreading of particles
[371]. This is in fact a chaotic scattering process, which has important applications
in pollutant transport. The quantum-mechanical aspects of chaotic scattering were
addressed by Blümel and Smilansky [76], Jung [368], Gaspard and Rice [259], and
Cvitanović and Eckhardt [152].
The work by Crutchfield and Kaneko [146] on transient chaos in spatiotemporal
systems generated a new perspective of research aiming at understanding whether
spatiotemporal complexity, or turbulence, is related in general to attractors or rather
to nonattracting chaotic sets generating long-lived transients.
The first experimental indications of irregular transient behavior were found
in hydrodynamical systems, where transients were followed over days before
1.4 A Brief History of Transient Chaos 35

settling down on a periodic attractor [9]. Although in current terminology these

were spatiotemporal chaotic transients, a number of papers appeared afterward
reporting low-dimensional transiently chaotic behavior in systems exemplified by a
compass forced by a magnetic field [145], lasers [23,154,577], electronic oscillators
[22,333,646], and a parametrically forced pendulum [479]. Besides the convection-
loop [273,274] and the pendulum [479,480] experiments (cf. Sect. 1.3), a spin-wave
experiment [110, 111, 113] seemed to have provided high-quality measurements of
chaotic transients. Other investigations included the dynamics of a bouncing ball
[422] and a driven magnetoelastic ribbon [196]. Many experimental systems in
which transient chaos has been observed are in fact systems with fractal basin
boundaries (e.g., [22, 422, 577]). In spite of these experimental works and the
several experiments carried out in the last 20 years, it is possible that due to the
limited awareness of the phenomena of transient chaos even among researchers in
the nonlinear-dynamics community, transiently chaotic signals were considered to
be uninterpretable and were discarded.
Chapter 2
Transient Chaos in Low-Dimensional Systems

We study low-dimensional dynamical systems, i.e., systems described by one-

dimensional noninvertible or two-dimensional invertible maps. For such systems it
is often possible to obtain analytic understanding of generic properties of transient
chaos that are shared by more realistic physical systems. For example, for a higher-
dimensional system, one-dimensional maps can be used to model the dynamics
along the unstable manifold [220, 564].
For one-dimensional maps, we shall analyze in detail the relation between the
natural and the conditionally invariant measures of repellers, and introduce analytic
tools based on the Frobenius–Perron type of eigenvalue equations. These consid-
erations allow us to derive explicit dimension and entropy formulas. We will also
elaborate the relevance of transient chaos to problems in mathematics, physics, and
engineering, and address the issue of nonhyperbolicity. In particular, we will show
that nonhyperbolic dynamics can lead to a power-law decay, and to the concentra-
tion of the natural measure on a single point, despite the fractal character of the
repeller.
Two-dimensional invertible maps are equivalent to three-dimensional flows and
can be obtained by the standard technique of stroboscopic or Poincaré surface
of section [564]. We shall discuss again the relation between the natural and the
c-measures, derive the corresponding entropy and dimension formulas, and pro-
vide an information-theoretic motivation. A fundamental feature of any chaotic set
is the existence of an infinite number of unstable periodic orbits. We will discuss
how transient chaos is organized about the periodic orbits. While chaotic saddles
are often hyperbolic, meaning that the stable and the unstable manifolds do not be-
come tangent to each other (i.e., there are no homoclinic or heteroclinic tangencies),
nonhyperbolic chaotic saddles can also arise [454].
For completeness, the multifractal formalism, a more systematic characterization
of the natural measure, will be discussed briefly in Appendix A.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 37
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 2,

c Springer Science+Business Media, LLC 2011
38 2 Transient Chaos in Low-Dimensional Systems

2.1 One-Dimensional Maps, Natural Measures, and c-Measures

2.1.1 Basic Properties of One-Dimensional Maps Generating

Transient Chaos

Strong dissipation leads to significant contraction in the phase space and conse-
quently to an approximately one-dimensional discrete map on a stroboscopic or
Poincaré plane. One-dimensional maps are typically noninvertible. They are the
simplest class of dynamical systems in which transient chaos can occur, and in-
sights obtained from studying them can often be useful for exploring transient chaos
in higher-dimensional systems.
Let f (x) be a one-dimensional noninvertible map. Transient chaos occurs if an
interval I is mapped partially outside itself. We shall consider here single-humped
map functions, as illustrated by Fig. 2.1. There is a primary escape interval I0 that is
mapped outside I in a single iterate. The dynamics of the map outside I is irrelevant.
For example, there can be one or more attractors far away, but if there is no feedback
from these regions onto I, the transient chaotic behavior is completely specified by
the function f defined on I.
As an illustrative example, we consider the classical logistic map xn+1 =
f (xn , r) = rxn (1 − xn), which generates transient chaos for r > 4. There is a chaotic
repeller in the unit interval I = (0, 1), and this interval can be chosen as the restrain-
ing region Γ . Figure 2.2 shows the approximate invariant measure on the repeller
for r = 4.03 covered by uniform boxes of size ε = 2 · 10−3, where the box probabil-
ities Pi (ε ) are displayed. Figure 2.2 also contains information about the repeller’s
structure. For example, it can be noticed that the crudest approximation to the

Fig. 2.1 A typical

one-dimensional map
generating transient chaos on
some interval I. Note that
points lying outside the two
(1) (1)
subintervals I1 and I2
escape I after one time unit.
Points that do not exit in n
steps are contained in the
(n − 1)th preimages of the
two subintervals (i.e., in the
nth preimages of I), and
exhibit chaotic behavior on
time scale n due to the global
expansivity of the map
2.1 One-Dimensional Maps, Natural Measures, and c-Measures 39

Fig. 2.2 Natural distribution for the logistic map xn+1 = rxn (1 − xn ) for r = 4.03 obtained by
the ensemble method (cf. Sect. 1.2.2.2) on a grid of size ε = 0.002. The number of initial points
distributed uniformly in I = R = Γ is N0 = 107 , and the first 10 and the last 30 steps of trajectories
are discarded. The truncated trajectories contain about 106 points, so that reasonable statistics can
be obtained. The escape rate can be estimated by comparing the number of trajectories of length 10
and 30 in I, which yields, via (1.6), κ ≈ 0.07. The lower part of the figure illustrates the organization
(n)
of the repeller, where Ii denotes the cylinders at level n, i.e., the nth preimages of I = (0, 1) (cf.
Fig. 2.3), and resembles the construction of a Cantor set

(1) (1)
repeller cover consists of two intervals, the two preimages I1 and I2 of I. At the
next stage, each of them splits into two smaller intervals. Subsequent successive
refinements will then yield a complete hierarchy, the nth level of which contains
all the nth preimages of I. The preimage intervals are called cylinders and are
(n)
denoted by Ii , where the subscript i enumerating them runs, at the nth level, up
n
to 2 . Note that here, the base 2 is due to the two branches that the inverse map
f −1 exhibits. Also note that the cylinders provide coverage of the repeller with
nonuniform boxes that fit the repeller’s structure in a natural manner. An equivalent
way of defining the cylinders is to consider the n-fold iterated map f n , whose graph
contains 2n branches (Fig. 2.3). The intervals mapped exactly onto I by the nth
iterated map are nothing but the cylinders of level n, indicating that points in any
(n)
subinterval Ii do not leave the restraining region I sooner than n + 1 steps. Since
the folds with derivatives smaller than unity are, for sufficiently large n, outside I,
the dynamics are expansive, or in this one-dimensional case, are hyperbolic on the
repeller.
40 2 Transient Chaos in Low-Dimensional Systems

Fig. 2.3 The threefold

iterated logistic map for
r = 4.03, and cylinders
(3)
Ii , i = 1, . . ., 8. The dots
denote points of 3-cycles

2.1.2 Conditionally Invariant Measure

We now consider the conditionally invariant (c-)measure [596] for one-dimensional

maps. This measure is defined on any region Γ containing the repeller, and de-
scribes how trajectories escape this region. For simplicity we take Γ = I. Consider
the conditional probability that a given region is visited by trajectories originated
from random initial conditions in I that do not escape I in m steps. A fraction of
these trajectories will escape at the next time step, and hence their last points in I
are in the gaps among the cylinders. The conditional probability is thus defined on
the entire interval I. The limit to which this conditional probability converges for
m → ∞ is the conditionally invariant measure. As mentioned in Sect. 1.2.3.2, the
c-measure is effectively the time-independent distribution maintained by supplying
new points into the system exactly according to the escape rate.
The density of the c-measure can be constructed from trajectories of minimal
length m in the basic interval. For example, one can select trajectories whose first m
points are disregarded and all the remaining points are kept before escaping I. This
procedure usually converges exponentially fast for rather arbitrary choice of m (e.g.,
m = 10), as exemplified by Fig. 2.4. It can be seen that the measure has a smooth
density everywhere on the interval I, due to the fact that it is in general smooth along
the unstable manifold (cf., e.g., Fig. 1.11).
To connect the conditionally invariant measure with the natural measure, we
restrict the density of the c-measure to cylinders of level n. This requires a renor-
(n)
malization so that the total measure on the cylinders is unity. The c-measure μci
(n)
of intervals Ii characterizes then the motion of trajectories visiting the cylinders
of level n. For sufficiently large n, these are the trajectories exhibiting long-lived
(n)
chaotic transients. Therefore, the limit of the cylinder measures μci for n → ∞ can
be considered as the exact natural measure μ on the repeller:

(n) (n)
μci → μi . (2.1)
2.1 One-Dimensional Maps, Natural Measures, and c-Measures 41

Fig. 2.4 Density ρ (x) of the conditionally invariant measure for the logistic map on I = (0, 1),
obtained by discarding the first 10 steps of trajectories and keeping all points that stay inside the
restraining region I. Parameter, initial conditions, and the box size ε are the same as in Fig. 2.2

Fig. 2.5 Measure obtained by restricting the conditionally invariant measure of Fig. 2.4 to
cylinders of level n = 5 specified by the accuracy ε = 0.002. The cylinder measures in this
(5)
approximation are μi , for i = 1, . . ., 32. Note that the two outermost cylinders are not resolved by
the grid
(n)
For large n we can omit the subscript c and denote the cylinder measure by μi .
We thus have two different approximations to the natural measure: the box proba-
(n)
bilities Pi (ε ) and μi . The equivalence of the two quantities is illustrated in Fig. 2.5,
which displays the c-measure restricted to cylinders of level 5. Alternatively, one can
smooth out the approximate natural measure shown in Fig. 2.2 on the same set of
(n)
cylinders. The resulting distribution is essentially the same as that of μi even for
the relatively low value of n = 5.

2.1.3 The Frobenius–Perron Equation

The Frobenius–Perron equation provides a framework from which analytic insights

into the density ρ (x) of the c-measure can be obtained. In general, the equation
42 2 Transient Chaos in Low-Dimensional Systems

governs the time evolution of a density ρn (x) via an iterative scheme. For a one-
dimensional map exhibiting transient chaos, the equation is

ρn (x)
ρn+1(x ) = R ∑ | f (x) |
, (2.2)
x∈ f −1 (x )

where R is a prefactor and the summation is taken over the preimages of x . For
R = 1, the equation is the Frobenius–Perron equation for attractors [220, 564], cor-
responding to a situation without escape. For transient chaos, escapes can be taken
into account by choosing properly the prefactor R. In particular, by iterating any
smooth, positive initial function ρ0 (x) on I, the series ρn (x) either diverges or tends
to zero unless we choose [596, 764]

R = eκ , (2.3)

which is the compensation factor described in Sect. 1.2.3.2 with κ being the escape
rate. With this choice of R, the series ρn (x) converges to a finite ρ (x):

ρn (x) → ρ (x) (2.4)

independently of the choice of the initial function. The limit ρ (x) is the density of
the c-measure and satisfies the following self-consistent equation:

ρ (x)
ρ (x ) = eκ ∑ | f (x) |
. (2.5)
x∈ f −1 (x )

This equation can be considered as an eigenvalue problem of the Frobenius–

Perron, or transfer, operator ∑x∈ f −1 (x ) | f (x) |−1 . The fact that its largest eigenvalue
exp (−κ ) is less than unity is the mathematical reason for the long-term exponential
decay of the survival probability (1.8).
In numerical experiments where the escape rate κ is not known a priori, different
values for R can be chosen and tested until convergence in ρ (x) is achieved. For
transient chaos in typical one-dimensional maps, the convergence is usually quite
fast and the limiting ρ (x) of reasonable accuracy can be achieved after a few itera-
tions [764], as shown by one example in Fig. 2.6. In this case, both the escape rate
and the density can be found by numerically solving the eigenvalue problem.
In general, the escape rate can be obtained as an integral of density ρ over the
(1) (1)
intervals I1 , I2 (see Fig. 2.1 and (1.14)) not escaping within a single step:
 
e−κ = μc ( f −1 (I)) ≡ (1)
ρ (x)dx + (1)
ρ (x) dx, (2.6)
I1 I2


provided that the c-measure is normalized to unity on I: I ρ (x) dx = 1. As a
consequence, the escape rate can also be expressed by the c-measure of the primary
escape interval I0 ,
2.1 One-Dimensional Maps, Natural Measures, and c-Measures 43

Fig. 2.6 Density ρ (x) of the logistic map on I for r = 4.03, obtained from (2.2) as the eighth
iterate of a constant initial function. The relative error is about 10−3


1 − e− κ = ρ (x) dx. (2.7)
I0

When the escape rate is small so that the approximation exp (−κ ) ≈ 1 − κ holds,
we have 
κ= ρ (x) dx ≈ ρ (xc )L, (2.8)
I0

where L is the size of I0 . The approximate equality expresses that for small κ the
primary escape interval is short, and the integral over I0 can then be approximated
as the density about a typical point xc in I0 multiplied by the length L of the escape
interval.
By iterating (2.2), one can see the presence of a singularity at the maximum
of f (x), but it falls outside I. This supports again the view that the density of the
c-measure is a smooth function on close neighborhoods of hyperbolic repellers.
(n)
Using the definition of the density, the cylinder measure μi can be expressed as

(n) ρ (x) dx
(n) x∈Ii
μi =  . (2.9)
∑j x∈I j
(n) ρ (x) dx

The smoothness of ρ enables us to obtain the actual value of the density from the in-
tegrals, for sufficiently small cylinders. Since the density does not change with the
refinement, the asymptotic scaling with n is governed by the length scales only.
Thus, for large n, the measure of a cylinder is proportional to its length. More
precisely, we have
(n)
(n) ε
μi ∼ i (n) for n  1, (2.10)
∑j εj
(n) (n)
where εi stands for the length of the cylinder Ii . The n-independent proportion-
ality factor not written out here depends on the actual form of ρ . Equation (2.10)
44 2 Transient Chaos in Low-Dimensional Systems

is essential for our subsequent discussions. Note that the total length of cylinders at
level n is proportional to the number of points not yet having escaped I after n steps.
The cylinder measure can therefore be expressed as

(n) (n)
μi ≈ eκ n εi . (2.11)

2.2 General Relations

There exist general relations in simple form among the metric entropy, the Lyapunov
exponent, the information dimension, and the escape rate. The box-counting dimen-
sion and the topological entropy, however, can be related to the escape rate in a more
complicated manner only. We shall derive these relations in this section.

2.2.1 Lyapunov Exponent, Information Dimension, and Metric

Entropy

(n)
With each cylinder Ii , one can associate a unique symbol sequence {S j } ( j =
1, 2, . . . , n) of length n. In a single-humped map, the symbols are binary: S j takes
on the value 0 (1) if a trajectory started in the cylinder is at time step j − 1 in
(1) (1)
the subinterval I1 (I2 ) (see Fig. 2.1). The leftmost and the rightmost cylinders
have the code {0, 0, . . . , 0} and {1, 0, . . . , 0}, respectively. The cylinders can then be
labeled by the corresponding symbol sequences. Moreover, the cylinder measures
are exactly the path probabilities P({S j }) (1.24) for finding a symbolic trajectory
{S j } of length n:
(n)
μi = P({S j }). (2.12)
To obtain the Lyapunov exponent, one observes that the logarithm of the slope of
the n-fold iterated map f n at x is just the stretching factor (see Sect. 1.2.3.3) at this
point. The slope is, however, approximately constant in a cylinder, as illustrated by
Fig. 2.3. Since the length of I can, in general, be chosen to be unity, the stretching
(n) (n) (n)
factor in cylinder Ii of size εi is approximately 1/εi . The stretching exponent
(n)
is then Λ1i (n) = − ln εi . The average Lyapunov exponent (1.18) is given by

1 1
n∑
Λ1i μi = − ∑ εi eκ n ln εi
(n) (n) (n) (n)
λ1 = (2.13)
i n i

for n  1. The metric entropy from (1.24), according to (2.11), and (2.12), is

1  
∑
(n) (n)
K1 = − εi eκ n κ n + ln εi . (2.14)
n i
2.2 General Relations 45

Comparing (2.13) and (2.14), we obtain

K1 = λ1 − κ . (2.15)

This indicates that for a chaotic repeller of a one-dimensional map, the metric en-
tropy K1 is not equal to the Lyapunov exponent (in contrast to the situation of a
chaotic attractor, where the two are equal [220, 564]), with the difference being the
escape rate.
The information dimension of the repeller can be calculated using the probabili-
(n)
ties Pi = μi in (1.22). A straightforward substitution of Pi in (1.22) yields

κ
D1 = 1 − . (2.16)
λ1

We see that the information dimension of the natural distribution on a chaotic re-
peller is always less than unity. The difference is given by the ratio of two rates: the
escape rate and the Lyapunov exponent. In addition, since D1 is nonnegative, we
have
κ ≤ λ1 , (2.17)
where the equality holds only for point repellers whose natural distribution is con-
centrated in a point (D1 = 0). A chaotic repeller is thus globally less repelling than
typical points in it, since the escape rate is smaller than the average Lyapunov expo-
nent. It may be said that the fractal structure tends to “stabilize” the repeller because
a larger dimension implies generally slower escape.
Finally, from (2.15) and (2.16), we obtain

K1 = λ1 D1 . (2.18)

The metric entropy is thus the product of the Lyapunov exponent and the informa-
tion dimension. Equations (2.15), (2.16), and (2.18) are particular instances of the
general relations expressed by (1.29) and (1.30).

2.2.2 Box-Counting Dimension and Topological Entropy

The idea behind the Frobenius–Perron equation (2.5) can be exploited for calculat-
ing the box-counting dimension. In particular, note that in fitting a smooth curve to
the natural distribution on a grid of size ε 1 such as the one in Fig. 2.2, one finds
that its form ρ̃ (x) differs from the density ρ (x) of the c-measure. The reason is that
the cylinders of level n  1 are by far not of equal size. It is possible to find an
equation of the type of the Frobenius–Perron equation (2.5) whose solution is ρ̃ (x).
Consider the coarse-grained chaotic repeller specified on a grid of fine resolu-
tion ε . Take an interval of length Δx 1 within I that is much larger than ε . The
probability of finding a point on the repeller within the interval of length Δx is
46 2 Transient Chaos in Low-Dimensional Systems

N(ε )ε /Δx, where N(ε ) is the number of bins of size ε covering the repeller inside
Δx, and N(ε )ε is the total length of such bins. By the definition of the box-counting
dimension (1.19), this quantity scales with ε as ε 1−D0 . The probability that a point of
the image interval Δx = f (x)Δx belongs to the repeller is N(ε Δx/ | Δx |)ε / | Δx |,
since the map is locally linear over the interval of length Δx, and a longer interval
Δx corresponds to using a finer resolution ε Δx/Δx in N. After one iterate, the prob-
ability of being on the repeller has changed by a factor of | Δx/Δx |1−D0 , which is
typically less than unity. A local escape rate κ (x) can thus be defined as

eκ (x) = | f (x) |1−D0 . (2.19)

The decrease of the probability from a coarse-grained repeller accompanying the

escape process can be compensated by multiplying the probability of being on the
repeller by exp [κ (x)] for every point x. In an equation analogous to (2.2), there is
then no overall correctional factor R, but the exponent of the derivative changes from
unity to D0 . These considerations lead to [753]

ρ̃n (x)
ρ̃n+1 (x ) = ∑ | f (x) |D0
, (2.20)
x∈ f −1 (x )

which we call the dimension equation. By iterating any smooth, positive initial func-
tion ρ̃0 (x) on I, one can find convergence to a finite density if the exponent is chosen
to be the repeller’s box-counting dimension. Equation (2.20) is a kind of eigenvalue
equation for the dimensions, and it provides a fast numerical algorithm for deter-
mining D0 . For our example in Fig. 2.2, the calculation yields D0 = 0.905344. The
series of ρ̃n (x) converges to the density ρ̃ (x), a smooth covering curve for the natu-
ral distribution on the repeller coarse-grained on a uniform grid [753].
Similar types of equations exist for the information dimension and other quan-
tities, as shown in Appendix A (A.12). Using a cumulant expansion of the local
Lyapunov exponents applied to these equations, one finds a relation between the
box-counting dimension and these cumulants in the form of

1
κ = (1 − D0)λ1 + (1 − D0 )2 Q2 + · · · , (2.21)
2
where Q2 is the second cumulant. An analogous relation for the topological entropy
[380] can be obtained, which is

1
K0 = λ1 − κ + Q2 + · · · , (2.22)
2
as also derived in Appendix A. Equations (2.21) and (2.22) show that, in contrast to
D1 and K1 , the box-counting dimension and the topological entropy can be related
to the escape rate only if an infinite series containing the cumulants of the local
Lyapunov exponents is also taken into account.
2.2 General Relations 47

Since all possible binary sequences are allowed to occur in the class of maps in
Fig. 2.1 (there are Ωm = 2m sequences of length m), the topological entropy from
(1.25) is
K0 = ln 2, (2.23)
regardless of the particular form of f (x).

2.2.3 An Analytically Tractable Example: The Tent Map

As an analytic example, we consider the tent map defined on the unit interval I =
(0, 1):
⎧
⎨ax for x < b/(a + b),
f (x) = (2.24)
⎩b(1 − x) for x > b/(a + b),

(n) (n)
where a > 1 and b > a/(a − 1). The lengths of the two subintervals I1 and I2
(cf. Fig. 2.1) are 1/a and 1/b, respectively. For the tent map, the convergence of an
initial distribution ρ0 (x) to the density of the c-measure can be followed explicitly
under the Frobenius–Perron equation (2.2):

ρn (x /a) ρn (1 − x /b)
ρn+1 (x ) = R + . (2.25)
a b

To find a solution to (2.25), we assume that the distribution is linear at any iteration:

ρn = αn x + βn. (2.26)

A direct substitution yields the following mapping for the coefficients:

1 1 1 1 R
αn+1 = R − αn , βn+1 = R + βn + αn . (2.27)
a2 b2 a b b

A finite nonzero limiting value of β exists only if the factor in front of βn is unity,
which gives R(1/a + 1/b) = 1. The escape rate is thus given by

κ = ln R = − ln (1/a + 1/b) = ln a + lnb − ln(a + b). (2.28)

With this value of R, the factor in front of αn is less then unity in modulus, and hence
the series of αn converges to zero, so that the asymptotic distribution is a constant.
Taking into account normalization, we obtain ρ (x) ≡ 1.
48 2 Transient Chaos in Low-Dimensional Systems

Applying (2.20) also yields a constant asymptotic density ρ̃ , provided that the
box-counting dimension satisfies1

a−D0 + b−D0 = 1. (2.29)

Due to the strict self-similarity of the cylinder construction, other dynamical in-
variants can be found from the first level. In particular, since ρ = 1, the first-level
(1) (1)
cylinder measures are μ1 = b/(a + b) and μ2 = a/(a + b) with local slopes a and
b on these cylinders, respectively. From (2.13) we obtain

a ln b + b lna
λ1 = . (2.30)
a+b

For the metric entropy (2.14), we have

a ln a + b lnb
K1 = − + ln (a + b). (2.31)
a+b

From (1.22), the information dimension of the natural measure of the repeller is

−a ln a − b lnb + (a + b) ln(a + b)
D1 = . (2.32)
a ln b + b lna

One can see that indeed, the general relations (2.15), (2.16), and (2.18) are non-
trivially satisfied for the tent map. It is only for the symmetric case a = b that the
dimensions are equal to each other, and so are the entropies. In particular, in this
case we have D0 = D1 (= ln 2/ ln a) and K0 = K1 (= ln 2).

2.3 Examples of Transient Chaos in One Dimension

The following examples, taken from different disciplines including number theory,
engineering, and statistical mechanics, illustrate the ubiquity of transient chaos.

2.3.1 Numbers with Incomplete Continued Fractions

It is known that every positive number x can be expanded in a unique continued

fraction [331, 408], i.e., the number can be written as

1 The same result follows for arbitrary choices of the initial densities, and the constant-valued

distribution is indeed an attractor of the functional recurrences (2.2) and (2.20).

2.3 Examples of Transient Chaos in One Dimension 49

1
x= , (2.33)
1
i1 +
1
i2 +
i3 + · · ·

where the ik are natural numbers. The expansion is obtained by subtracting from
the reciprocal of the number the integer part (i1 ) of the reciprocal, then taking the
resulting reciprocal and subtracting from it its integer part (i2 ), and so on.
The one-dimensional Gauss map2 defined on I = (0, 1] as the difference between
the reciprocal and its integer part is

1 1
xn+1 = f (xn ) ≡ − . (2.34)
xn xn

The map is related to continued fractions. In particular, it contains infinitely many

branches; see Fig. 2.7a. Starting the counting from the rightmost branch as branch
f1 , branch i is of the form fi (x) = 1/x − i and is defined over the interval (1/(i +
1), 1/i). Taking an initial point x0 in this interval, we obtain x1 = 1/x0 − i and hence
x0 = 1/(i + x1 ). The number x1 can thus be considered as the remnant after a sin-
gle step of the continued fraction expansion. The expansion can be continued. For
example, taking an initial point x0 mapped by fi1 first and then by fi2 , we get

1
x0 = . (2.35)
1
i1 +
i2 + x2

A comparison with (2.33) shows that the sequence i1 , i2 , i3 , . . . of branch indices

encountered while iterating map (2.34) with any irrational initial condition x0 corre-
sponds precisely to the infinite continued fraction expansion of x0 . Accordingly,

a b

Fig. 2.7 (a) Gauss map f and (b) truncated Gauss map f m that contains only branches with i ≤ m.
Here m = 3

2Gauss was aware of the fact that this map has an invariant density proportional to 1/(1 + x). The
map is in fact chaotic in modern terminology [328].
50 2 Transient Chaos in Low-Dimensional Systems

the fixed point x∗i belonging to branch i is a number that contains only i in its
continued fraction, where x∗1 is in fact the golden mean. Since any number in (0, 1]
can be expanded if all positive integers can appear in the expansion, the full interval
I = (0, 1] remains invariant under the map (2.34).
One can, however, ask what the numbers whose continued fractions contain cer-
tain integers only are. They form fractal subsets of the unit interval. In particular, we
shall be interested only in numbers with continued fractions containing integers less
than or equal to a certain threshold m. These numbers must be invariant under the
truncated Gauss map fm (xn ) that is of the same form as (2.34) but does not contain
the branches i = m + 1, m + 2, . . ., as shown in Fig. 2.7b. The truncated map fm is
thus of the form of (2.34) but is defined on the interval [1/(m + 1), 1] only.
Points from the interval I0 = (0, 1/(m + 1)) have integers larger than m in their
expansion. As a result, whenever a point under iteration enters this region, it should
be discarded. Map fm is therefore a map from which escape takes place, generat-
ing transient chaos. The numbers with continued fractions containing integers less
than or equal to m form the chaotic repeller of the map fm . The repellers can be
constructed numerically. Several representative repellers are shown in Fig. 2.8.
The box-counting dimension D0m of the repeller of the map fm follows from the
dimension equation (2.20), which now reads
m
1
ρ̃n+1(x ) = ∑ (x + i)−2D0m ρ̃n . (2.36)
i=1 x +i

The values of D0m can be determined from (2.36) numerically, and the results [328]
are summarized in Table 2.1. By now, efficient algorithms exist [330,358] for deter-
mining these dimensions up to many more digits than shown here. It is interesting
to mention that by means of spectral methods, D. Hensley derived [329] an analytic
expression for D0m , valid for large m:

6 72 ln m
D0m = 1 − − 4 2 . (2.37)
π 2m π m

m
0 1

Fig. 2.8 Numbers of the unit interval with continued fraction expansions containing integers less
than or equal to m for m = 2, 3, 5, 8, 20, obtained as invariant chaotic repellers under the map f m .
Each set is similar to a Cantor set
2.3 Examples of Transient Chaos in One Dimension 51

Table 2.1 Box-counting dimension D0m of irrational numbers with continued

fraction expansions containing integers less than or equal to m
m 2 3 4 5 6 7 8
D0m 0.5313 0.7057 0.7889 0.8368 0.8676 0.8890 0.9046

For m > 8, this expression is quite accurate: its predicted values agree with the
numerically obtained dimension values within less than half of 1%.

2.3.2 Shimmying Wheels

The lateral vibrations of towed wheels, the shimmies (called after a dance that was
popular back in the early 1920s), are self-excited nonlinear oscillations of great
engineering relevance. These vibrations can be observed during the motion of the
towed wheels of shopping carts, wheelchairs, trailers, or on the front wheels of air-
planes, bicycles, or motorcycles. In the simplest setting, the vehicle can be modeled
as a rigid body moving along a straight line at a constant speed v. For practical pur-
poses, a stationary rolling of the wheel is desirable so that the wheel’s center of mass
moves parallel to the towing velocity. Shimmy may occur if some part of the wheel
system, either the wheel itself or the point at which it is affixed, is elastic [738].
To gain insight here we consider an idealized model that exhibits shimmying
motion [272]. The wheel is assumed to be rigid, but the connecting assembly can
move laterally, as shown schematically in Fig. 2.9. The vehicle moves at a constant
velocity v along the x-axis. The vertical center of rotation of the wheel, the kingpin,
is attached to the vehicle via a spring that allows the kingpin to oscillate in the y
direction. An important parameter of the system is the caster length l, i.e., the offset
of the wheel’s axis with respect to the kingpin. The deflection angle of the wheel
assembly with respect to the x-axis is ψ ; the rotational angle of the wheel is φ . In a
reference frame co-moving with the vehicle, the degrees of freedom are the angles
ψ and φ , and the instantaneous position y of the kingpin. If the wheel rolls, i.e.,

Fig. 2.9 Schematic diagram

of an idealized wheel model
that exhibits chaotic
shimmying
52 2 Transient Chaos in Low-Dimensional Systems

its contact point has zero velocity relative to the ground, the rotational angle is no
longer an independent variable. The system thus has two degrees of freedom, and the
phase space is four-dimensional. Note, however, that due to the prescribed towing
velocity, the total energy is not conserved; either the system can be dissipative, or
it can absorb energy via the constraining force of the vehicle’s engine that provides
the constant speed.
The state of straight rolling (ψ = 0, y = 0) is unstable for short caster length
for l < lc , where the critical length lc is determined by other parameters such
as the wheel’s radius, the masses, and the towing velocity [738]. Even when the
caster length is sufficiently long that the system is in the regime of stable straight
rolling, the dynamics is nontrivial, since there exists an unstable limit cycle outside
the fixed-point attractor at the origin. For vibrations of amplitude larger than that
of the limit cycle, the corresponding ψ values increase in time. In this process, the
angle of deflection reaches a critical value for which the static friction is no longer
able to provide the necessary constraining force for rolling. That is, the wheel slides.
It is insightful to focus on the dynamics of a sliding wheel for which φ is a rel-
evant variable, and the phase space becomes six-dimensional. Due to the sliding
friction, the dynamics is strongly dissipative, and the velocity of the contact point
starts decreasing after some time and sooner or later it approaches zero asymptoti-
cally, recovering the rolling state. An essential feature of the full dynamics is thus
the temporary change between a four- and a six-dimensional phase space. When a
trajectory enters the former, its destination depends on whether it is inside or out-
side the unstable limit cycle. If it is outside, the amplitude of the ψ -oscillations starts
growing, the condition of rolling cannot be satisfied, and the trajectory will be in the
larger phase space again. Due to such events, the wheel exhibits chaotic “dance”
(motion). If, however, a return occurs to some region inside the limit cycle and into
the basin of attraction of the fixed-point attractor, the straight rolling of the wheel is
approached asymptotically. This actually occurs with a finite probability.
The transiently chaotic shimmying of wheels, occurring for a broad range of pa-
rameters, can be represented by an approximately one-dimensional map due to the
strong dissipation in the underlying dynamics [739]. On the Poincaré section taken
at y = 0, the map ψn+1 = f (ψn ) consists of two branches, as shown in Fig. 2.10.
Rolling is represented by the left branch with an unstable fixed point at ψ ∗ that rep-
resents the unstable limit cycle. The critical value at which sliding starts is denoted
by ψsl . There is a jump in the dynamics because the right branch corresponding to
sliding decreases and stretches somewhat below the level of ψ ∗ . The narrow interval
I0 is thus the escape interval, and it provides a mechanism for trajectories to return
to the attractor at the origin. Insofar as the trajectory is inside the box above ψ ∗ ,
the dynamics is chaotic. Thus, we conclude that the shimmying motion of wheels
can be transiently chaotic for typical parameters, which can indeed be observed,
for example, when trolleys at supermarkets or airports are towed. Control of such
transient chaotic motions in motorcycles and in airplanes is a basic task in vehicle
engineering [738, 754, 755].
A numerical investigation of the transient chaotic dynamics within the inter-
val (ψ ∗ , ψmax ) = (6, 9) degrees in a realistic case leads to the conclusion that the
2.3 Examples of Transient Chaos in One Dimension 53

Fig. 2.10 Schematic diagram yn+1

of the one-dimensional return
map of the deflection angle ψ ymax
of a shimmying wheel. The
square indicates the box
within which chaotic
dynamics occur. Due to the
small gap I0 at its lower right
corner, the map is open, and
shimmying is transiently y*
chaotic
I0

yn
0 y* ysl ymax

max

Fig. 2.11 Chaotic repeller associated with shimmying motion on the angle interval (ψ ∗ , ψmax )
= (6, 9) degrees

underlying chaotic repeller is rather dense, with box-counting dimension D0 ≈ 0.95,

as shown in Fig. 2.11. The escape rate is κ ≈ 0.006. Taking into account the char-
acteristic time of oscillations t0 = 0.63 s [739], one can estimate that the average
lifetime is τ ≈ t0 /κ = 100 s (using (1.7)), which is the typical duration of shimmy-
ing observed, say, when trolleys are towed.

2.3.3 Random-Field Ising Chain

Take a semi-infinite chain of Ising spins {s1 , s2 , . . . , sn , . . . } in an inhomogeneous

external field {h1 , h2 , . . . , hn , . . . } with the interaction Hamiltonian
∞
H= ∑ (Ks j s j+1 + h j s j ), (2.38)
j=1

where K is a coupling constant. For simplicity we use units in which kB T = 1.

The set of local fields {h j } is considered a particular realization of a random field
distribution assuming at each site the values +h and −h with probabilities p and
1 − p, respectively. The thermodynamic properties are determined by the following
partition function:
∞ 
Z = e−F = ∑ exp − Ks1 s2 − h1s1 − ∑ (Ks j s j+1 + h j s j ) (2.39)
{s1 ,s2 ,... } j=2
54 2 Transient Chaos in Low-Dimensional Systems

for a fixed realization of the fields {h j } and then by averaging the free energy over
different realizations. The summation over spins can be obtained recursively [658].
Since the first spin appears in two terms of H only, the partition function can be
evaluated, yielding
 
∞
Z= ∑ 2 cosh (Ks2 + h1) exp − ∑ (Ks j s j+1 + h j s j ) . (2.40)
{s2 ,s3 ,... } j=2

Since s2 can take on the values ±1 only, an exponential representation of the hyper-
bolic cosine function gives

cosh (Ks2 + h1) = exp [A(K, h1 ) + g(K, h1)s2 ], (2.41)

where
1
A(K, x) = ln [cosh (K + x)cosh (K − x)],
2
1
g(K, x) = ln [cosh (K + x)/ cosh (K − x)].
2

This form shows that spin 1 contributes an amount −A(K, h1 ) to the free energy F,
and generates simultaneously also an extra field g(K, h1 ) for spin 2. The partition
sum can thus be rewritten as
 
∞
Z= ∑ exp (A(K, h1 )) exp −Ks2 s3 − x2 s2 − ∑ (Ks j s j+1 + h j s j ) , (2.42)
{s2 ,s3 ,... } j=3

where x2 is an effective field acting on spin 2 and is given by

x2 = h2 + g(K, h1). (2.43)

Note that the partition sum has a form similar to the original one (with x2 replacing
h2 ). The summation over subsequent spins can then be carried out in an analogous
way. After n steps we find the field acting on spin (n + 1) as

xn+1 = hn+1 + g(K, xn ), (2.44)

and the actual contribution to the free energy becomes −A(K, xn ). As a result,
a recurrence can be found that is actually random, since the fields {h j } are ran-
dom variables [302]. According to rule (2.44), the effective field xn+1 takes on
the values h + g(K, xn ) and −h + g(K, xn ) with probabilities p and 1 − p, respec-
tively. Consequently, the recurrence can be written as a two-valued map in which
trajectories stay on the upper and lower branches with probabilities p and 1 − p,
respectively, as shown in Fig. 2.12. The actual form of the map depends on the cou-
pling constant K and the field magnitude h. Although the branches alone are not
2.4 Nonhyperbolic Transient Chaos in One Dimension and Intermittency 55

Fig. 2.12 For a random Ising chain, random map generating the local field x (left) and the inverse
of the map (right). The repeller of this map (K = 1, h = 1.1) shown at the bottom is the attractor
of the random iteration (2.44) for any choice of the probability p

expanding, the random map exhibits chaotic motion on an attractor. The natural in-
variant measure on this attractor is physically important, since the averaged thermal
free energy per spin is just the mean value of −A(K, x) taken with respect to the
natural measure of variable x on the attractor. The averaged magnetization per spin
and other thermodynamic properties can also be expressed by the natural measure
of the attractor [302].
When there is a gap between the branches, as shown in Fig. 2.12, the attractor is a
fractal. This can be seen by noting that the whole interval I on which the dynamics is
(1) (1)
defined is mapped into two smaller subintervals I1 and I2 with a gap in between.
The images of the small subintervals also have gaps at every level. In fact, these
subintervals are exactly the cylinders in the inverted map f (x) shown in the right
panel of Fig. 2.12. It can thus be concluded that the attractor of the random map
is nothing but the repeller of the inverted map, which holds for all values of p.
The natural measure on the attractor, however, depends on the choice of p and is
not related to the natural measure of the repeller [54]. The former can be obtained
by iterating the map f backward with branching probabilities p and 1 − p and is
independent of the choice of the initial point. The attractor geometry (the fractal
pattern in Fig. 2.12) and its box-counting dimension (D0 = 0.8), however, follow
from the dynamics of the inverted map.

2.4 Nonhyperbolic Transient Chaos in One Dimension

and Intermittency

Nonhyperbolicity in a one-dimensional map arises when the local slope of the map
is unity or infinite, which can lead to escaping dynamics differing from those in
hyperbolic systems.
56 2 Transient Chaos in Low-Dimensional Systems

Fig. 2.13 Symmetric

nonhyperbolic map with left
branch f (x) = x + 3x2
generating transient chaos. 1
The construction of the
first-level cylinders is
indicated
f(x)

0 x
0 x1 x2 1

(1) (1)
I1 I0 I2

(2) (2) (2) (2)

I1 I2 I3 I4

In particular, the presence of a single marginally stable orbit implies a slow,

nonexponential decay process and an intermittent type of behavior before escap-
ing. To illustrate these features, we consider the class of maps shown in Fig. 2.13.
About the origin, for x 1 the maps are assumed to have the form

f (x) = x + Axk−1, k ≥ 3, (2.45)

and they are expanding otherwise: | f (x) |> 1. The origin is thus marginally stable,
leading to weak repulsion of trajectories that come close to the origin and rela-
tively long dwelling times there. This region can thus be said to be sticky. Because of
the long dwelling time, iterations of the map can be approximated by a continuous-
time differential equation:
dx
= Axk−1 , (2.46)
dt
solutions to which indicate that the time t needed to reach a finite distance l from
any initial point x0 close to the origin scales as t ∼ x0 −(k−2) . This implies that for
a smooth initial distribution, the number N(n) of trajectories staying inside a small
interval about the origin changes, after a large number n of iterations, as

N(n) ∼ n−1/(k−2). (2.47)

As a consequence, the survival probability (see Sect. 1.2.1) also scales as

P(n) ∼ n−1/(k−2). (2.48)

2.4 Nonhyperbolic Transient Chaos in One Dimension and Intermittency 57

Its decay follows a power law n−σ with an algebraic-decay exponent σ = 1/(k − 2)
for n  1, and hence the escape rate is zero: κ = 0. The cylinder construction shows,
however, that the repeller, in fact a nonhyperbolic repeller, is a fractal set of dimen-
sion D0 > 0 [161].
We now consider the c-measure. Observe first that for maps having a local
form f (x) ∼ 1 − a| x − xi |z (i = 1, 2) in the vicinities of the internal endpoints
(1) (1)
xi of intervals I1 , I2 , the relation between point x ≈ 1 and its preimages x is
x − xi = (−1) [(1 − x )/a]1/z . From the Frobenius–Perron equation (2.5) with κ = 0,
i

we obtain
 1/z−1
ρ (x ) ∼ 1 − x for x → 1. (2.49)

Note that for maps of the type in Fig. 2.13, z = 1 and ρ (x ) tends to a constant for
x → 1. The Frobenius–Perron equation (2.5) for x → 0 becomes

ρ (x) ρ (1)
ρ (x ) = + . (2.50)
1 + (k − 1)Ax k−2 | f (1) |

Since in this regime x ≈ x, the equation can be satisfied only if ρ (x ) x k−2 = const,
i.e.,
−(k−2)
ρ (x ) ∼ x for x → 0. (2.51)
The density of the c-measure is singular at the origin, implying that the c-measure
(n) (n)
μ1 of the leftmost cylinder is much larger than any other μi at level n. Since the
(n)
length ε1 shrinks with n much more slowly than that of the others, in the asymp-
(∞)
totic limit the full measure is concentrated at the origin: μ1 = 1. The natural
measure of this nonhyperbolic repeller can thus be represented by a Dirac delta func-
tion situated at the origin. As a consequence, the information dimension is D1 = 0,
and the average Lyapunov exponent is that of the origin, i.e., λ1 = 0. The repeller is
therefore nonchaotic in the sense of zero Lyapunov exponent, but strange since it is
fractal. It has therefore been called a strange nonchaotic repeller [161] (in analogy
with strange nonchaotic attractors [241]). Trajectories spend long times near the ori-
gin, but burst into other regions of the repeller from time to time, return to the origin
again, and so on. The dynamics is thus intermittent [612] before escape takes place.
In the case of nonhyperbolicity due to infinite derivatives, z < 1 at the endpoints
x1 , x2 of the primary escape interval, so the density vanishes at x = 1, as indicated
by (2.49). The Frobenius–Perron equation about the origin reads

ρ (x)
ρ (x ) = R . (2.52)
| f (x) |

The density can remain constant at the origin, provided that the slope | f (0) | and
R = eκ are equal.
58 2 Transient Chaos in Low-Dimensional Systems

Maps for which the logarithm of the slope at the origin is equal to the escape rate
are infinitely steep at the endpoints of the escape interval. Such systems are therefore
said to be in the border state of transient chaos [400, 541]. Since the density is
finite at the origin, the c-measure of the leftmost cylinder at level n is proportional
to its length, which in turn is proportional to | f (0) |−n . Taking into account the
compensation factor R, we have

(n) −n
μ1 ∼ Rn | f (0) | . (2.53)

(n)
Since the slope is exactly R, μ1 tends to a constant in the large-n limit. In this
case, the natural measure is again concentrated at the origin, and we have D1 = 0.
The slope is, however, larger than unity here, and therefore the average Lyapunov
exponent is finite: λ1 = κ = ln R. The underlying dynamics is intermittent-like, and
in spite of the high concentration of trajectories near the origin and the rare bursts,
the system is sensitive to initial conditions. The escape process is in fact exponential.
This case is called the border state because for the slope | f (0) | larger (smaller)
than eκ , transient chaos is hyperbolic (nonhyperbolic). An additional feature of the
border state is that, besides the c-measure whose density is finite at the origin and
tends to zero at the right end, there exist one or more smooth c-measures that vanish
at both ends and have escape rates different from ln R. The asymptotic character
of the escape process depends then on the exponent of the initial density ρ0 at the
origin [400, 497].

2.5 Analytic Example of Transient Chaos in Two Dimensions

A chaotic saddle arising in an invertible two-dimensional map has the appearance of

the direct product of two Cantor sets, as shown in Fig. 1.8. A pedagogical dynamical
system capable of exhibiting many typical features of a chaotic saddle is the open
baker map [753]. The action of the mapping can be described as follows. Take the
unit square and cut it by a horizontal line into two pieces of surface area c0 and
1 − c0 , respectively. The lower rectangle of width c0 is then stretched in the verti-
cal direction by a factor of a > 1 and simultaneously compressed in the horizontal
direction by a factor c < 1/2, while keeping the lower left corner fixed. The upper
rectangle is transformed by stretching and compression factors b > 1 and d < 1/2,
respectively, during which its upper right corner is fixed. If stretching is sufficiently
strong, i.e., ac0 > 1, b(1 − c0) > 1, the map generates transient chaos. Mathemati-
cally, the map can be written as

xn+1 = cxn , yn+1 = ayn , (2.54)

for yn < c0 , and

2.5 Analytic Example of Transient Chaos in Two Dimensions 59

Fig. 2.14 Open baker map defined on the unit square (the restraining region Γ ). The image (con-
tinuous line) and preimage (dashed line) of the unit square after one and two iterations are shown.
Points in the shaded and cross-shaded regions do not escape the unit square after one and two
iterations (forward or backward), respectively

xn+1 = 1 − d(1 − xn), yn+1 = 1 − b(1 − yn), (2.55)

otherwise. The one and two-step forward and backward images of the unit square
are shown in Fig. 2.14. Note that the inverted map is obtained by interchanging the
roles of x and y and by replacing a, b by 1/c, 1/d, and c, d by 1/a, 1/b, respectively.
Say we distribute N0 initial points on the unit square uniformly. A portion c0 of
them falls on the strip elongated by a factor a after the first step. Since only a ratio
1/ac0 of the total length overlaps with the unit square, the number of trajectories
staying inside the unit square is N0 /a. Analogously, from the other strip there are
N0 /b surviving trajectories. Altogether, a portion of (1/a + 1/b) of the initial points
does not escape the unit square in one time step. The same consideration applies to
future iterations as well. As a result, the escape rate is

κ = − ln(1/a + 1/b), (2.56)

which is the same as that for the tent map (2.28). The positive Lyapunov exponent
is also the same as that of the tent map (2.30).
The chaotic saddle of the baker map can be viewed as the direct product of two
Cantor sets, and the concept of partial dimensions [380, 564] can then be used to
characterize the saddle. Specifically, along the y-axis the contraction rates defining
the Cantor set are 1/a and 1/b, and hence the box-counting dimension along this
direction is the solution to the following equation; cf. (1.20):

(1) (1)
a−D0 + b−D0 = 1, (2.57)

(1)
where D0 is the partial box-counting dimension along the unstable direction.
Along the other axis, the contraction rates are c and d, and the corresponding di-
mension is the solution of
(2) (2)
cD0 + d D0 = 1, (2.58)
60 2 Transient Chaos in Low-Dimensional Systems

(2)
where D0 is the partial box-counting dimension along the stable direction.
The box-counting dimension of the chaotic saddle is the sum of the partial
dimensions along the stable and the unstable directions:

(1) (2)
D0 = D0 + D0 . (2.59)

As the concept of c-measure implies (Sect. 1.2.3.2), escape can be compen-

sated by multiplying the number of trajectory points staying on the unit square by
exp(κ ) at each time step. Starting from a uniform distribution on the unit square
ρ0 ≡ const, the probabilities for the two strips in the first step of the open baker
map are exp(κ )/a and exp(κ )/b. At the nth step there are 2n vertical strips of
different widths given by cm d n−m , m = 0, 1, 2, . . . , n. The probability of finding tra-
jectory points in strip j of width cm d n−m is then exp(κ n)a−m bm−n . Since the factor
appearing after exp(κ n) is just the reciprocal value of the stretching factor (see
Sect. 1.2.3.3) exp [Λ1 j (n)] for all points in a horizontal strip, we can express the
c-measure of strip j at level n by the escape rate and the stretching exponent as

(n)
μc j = eκ n e−Λ1 j (n) . (2.60)

Qualitatively, this implies that unstable regions with relatively large stretching
exponents are less frequently visited. It can also be seen that the limiting c-measure
is concentrated on the unstable manifold, whose branches are parallel to the y-axis,
with constant density.
The natural measure can be obtained in a similar manner. In particular, the com-
mon region between the first preimage and image of the square contains points that
have not escaped after one forward or one backward iteration. Similarly, the union
of the nth image and preimage defines trajectories staying inside the square for at
least n steps under both the direct and the inverted map, as shown in Fig. 2.14. Points
that do not escape thus belong to a double fractal set, the chaotic saddle. Boxes gen-
erated by the overlaps of the nth images and preimages provide a natural partition
of the saddle (the so-called generating partition [220]). To be explicit, we note that
(n)
the width ε1i of horizontal strip i is of the type a−m bm−n , for m = 0, 1, . . . , n. We
can then write
(n)
ε1i = e−Λ1i (n) , (2.61)

and the width of the vertical strips is

(n)
ε2 j = eΛ2 j (n) , (2.62)

where Λ2 j is the contraction exponent, i.e., the stretching exponent of the inverted
map multiplied by (−1). To obtain the natural measure, we note that the smooth
density of the c-measure is restricted to boxes of the partition at level n, and should
be renormalized. Since the weight of a strip has been determined, the renormal-
ization can be done by keeping the measure of the strip constant. This means that
2.5 Analytic Example of Transient Chaos in Two Dimensions 61

Fig. 2.15 Natural measure of a chaotic saddle from the baker map. The parameters are a=3, b=2.6,
c = 0.25, and d = 0.45. (Picture by M. Gruiz and Sz. Hadobás.)

(n)
(2.60) is valid for the natural measure μ j of a full vertical strip as well. For a box
(n)
of vertical size ε1i inside a given vertical strip j, the natural measure is given by

(n) (n) (n)

μi j = μc j eκ n ε1i . (2.63)

The exact natural measure can be obtained by taking the n → ∞ limit, as shown by
an example in Fig. 2.15. It can be seen that the projections of the natural measure
on the y- and x-axes can be regarded as the measures of the stable and the unstable
manifolds, respectively, for n  1. There are thus two different fractal distributions
embedded in the unit square, characterized by the partial information dimensions
(1) (2)
D1 and D1 along the unstable and the stable directions, respectively. The dimen-
sions are given by

(1) −a ln a − b lnb + (a + b) ln(a + b)

D1 = , (2.64)
a ln b + b lna
(2) a ln a + b lnb − (a + b) ln(a + b)
D1 = . (2.65)
a ln d + b ln c

A few remarks are in order.

1. Since the natural measure of a stable strip of order n is also the probability
P({S j }) for a binary symbols sequence {S j } of length n to occur (S j = 0(1)
if y < c0 (> c0 )), the metric entropy is the same as in (2.31). The topological
entropy is K0 = ln 2, since all possible binary sequences can occur.
62 2 Transient Chaos in Low-Dimensional Systems

2. The local Jacobian determinant associated with the first iteration of the map is
ac0 c/c0 for the lower band and b(1 − c0 )d/(1 − c0 ) for the upper band. The
average of the logarithms of these elements is

b a
ln J = ln ac + ln bd = λ1 + λ2, (2.66)
a+b a+b

where λ2 < 0 is the negative Lyapunov exponent of the saddle. The dynamics is
dissipative if it is phase-space contracting on average, i.e., if ln J < 0. The area-
preserving case can be obtained if all the local Jacobians are unity, which is the
case for ac = 1 = bd. This corresponds to an open Hamiltonian system, to be
treated in the chapter on chaotic scattering.
3. Note that the dynamical invariants are all independent of the parameter c0 .
It becomes, however, important in the limit of vanishing escape rates when
c0 = 1/a and c0 = 1 − 1/b hold, i.e., when both stretching rates are determined
by the parameter c0 .
4. General hyperbolic chaotic saddles turn out to be smoothly deformed versions of
the baker saddle.

2.6 General Properties of Chaotic Saddles in Two-Dimensional

Maps

2.6.1 Natural Measure and c-Measure

To gain insight into the properties of hyperbolic chaotic saddles arising in two-
dimensional maps, we take a restraining region Γ of size of the order of unity
containing the chaotic saddle, distribute a large number of initial conditions in Γ ,
and follow the resulting trajectories in the phase space. The map stretches (contracts)
Γ along the unstable (stable) direction and bends it so that a part of the image will
lie outside Γ . After n  1 iterations the overlap of the image with the original re-
gion consists of narrow strips that follow the local unstable directions of the chaotic
saddle, which are unstable strips, as illustrated in Fig. 2.16. For a hyperbolic chaotic
saddle, folds of the unstable manifolds fall outside Γ , and the corresponding unsta-
(n)
ble strips will be only slightly bent. The average width ε2 of an unstable strip is, by
definition, proportional to the local contraction factors exp [Λ2 (n)] of points falling,
after n steps, in the strip. Due to escape, the majority of the points will be outside Γ .
If after each step the density of points inside is compensated by the factor exp(κ ),
then a finite limit is found, the density of the conditionally invariant measure.
The equation describing this density can be obtained as follows. The probabil-
ity of finding points in a small region about x defines a distribution ρn (x) after n
steps. Its dynamics is governed by the requirement that the total probability in a
2.6 General Properties of Chaotic Saddles in Two-Dimensional Maps 63

Fig. 2.16 Schematic diagram

of the phase-space partition
for a chaotic saddle obtained
by taking the nth image and
preimage of region Γ of size
unity that covers the saddle.
Shown are stable strips
(dashed lines), unstable strips
(solid lines), invariant
manifolds, and the directions
along which the partial (1)
dimensions are defined Do
(2)
Do

small region at step n be the same as in the image of that region under the map f,
after taking into account the compensation factor [764]:

ρn (x)
ρn+1 (x ) = eκ , (2.67)
| J(x) | |
x∈f−1 (x )

where J(x) is the Jacobian determinant at the point x. This is the Frobenius–Perron
equation for open invertible two-dimensional maps. Similar to the one-dimensional
case (cf. (2.2)), for any smooth initial density ρ0 (x) the density ρn (x) converges for
n → ∞ to a limit density ρ (x), the density of the c-measure. The resulting c-measure
has a smooth, but not necessarily constant, density along the unstable manifold. The
fact that the largest eigenvalue exp (−κ ) of the Frobenius–Perron operator is less
than unity ensures again the exponential decay.
It is worth pointing out an analogy with the concept of almost invariant sets
introduced by Dellnitz, Froyland, and coworkers [174,250]. Such sets are regions in
the phase space from which a point escapes with a small probability in a period of
time. Chaotic saddles and their manifolds are invariant sets. A neighborhood of the
unstable manifold, e.g., can, however, be considered to be almost invariant, provided
that the escape rate is low. The regions of slow escape determine, in general, a
pattern related to the Frobenius–Perron (or transfer) operator defined by (2.67). The
eigenfunction belonging to the largest eigenvalue below unity provides the backbone
of this pattern. This largest eigenvalue is exp (−κ ) in our case, and the eigenfunction
ρ (x) concentrates on the pattern of the unstable manifold, as Fig. 1.11 illustrates.
We turn now to the determination of the natural measure. Without compensating
for the escaping, the measure of an unstable strip is proportional to the area of its
nth preimage. The nth preimage of Γ consists of strips that are parallel to the local
(n)
stable directions of the chaotic saddle. The average width ε1, j of a stable strip, as
shown in Fig. 2.16, is proportional to the contraction factor of the inverted map, i.e.,
(n)
to the reciprocal of the stretching factor of the forward map: ε1, j ∼ exp [−Λ1, j (n)],
where Λ1, j (n)/n is the positive Lyapunov exponent of points inside the strip.
64 2 Transient Chaos in Low-Dimensional Systems

The nth image of such a strip will then have side lengths of order unity and of order
exp [Λ2,i (n)] along the unstable and the stable directions, respectively. As a result,
a stable strip is nothing but the nth preimage of an unstable one. The measure of
(n)
an unstable strip j without compensating for the escape is thus proportional to ε1, j ,
the width of the nth preimage of the unstable strip. Taking into account the com-
pensation, we find that the c-measure contained in the unstable strip j is given by
(n) (n)
μc j ≈ eκ n ε1 j ∼ eκ n e−Λ1 j (n) . (2.68)
The natural measure is obtained by taking the overlap between the nth image and
preimage of the region Γ , which provides, for large n, a partition of the phase space
and an accurate coverage of the chaotic saddle [45, 220]. The natural measure of
a box in the partition can be obtained by redistributing the c-measures of unstable
(n)
strips according to the length scales ε1 . In particular, in unstable strip j, the natural
measure of box i of length ε1i can be expressed as

(n) (n) (n)

μi j ≈ μc j eκ n ε1i . (2.69)

(n)
It should be emphasized that μc j as given by (2.68) is in fact the natural measure μ
(n) (n) (n)
of strip j: μ j = μc j . Formally, this follows from ∑ ε1i ≈ exp(−κ n), which holds
because escape takes place along the unstable direction only.
The natural measure of stable strips can be obtained in a similar manner.
(n)
By summing (2.69) over j and utilizing ∑ j μc j = 1, one finds the natural mea-
sure of stable strip i to be
(n) (n)
μi ≈ eκ n ε1i , (2.70)

which is the same as for its nth image, an unstable strip. This reflects the fact that
the natural measure is preserved under the map.
Qualitatively speaking, maps generating hyperbolic chaotic saddles can be lo-
cally decomposed into baker’s transformations. The actual form of the local map
can change with the position only smoothly. This is the reason that the general ex-
pressions for the natural measures are similar to those obtained for the chaotic saddle
of the baker’s map.

2.6.2 Entropy and Dimension Formulas

(n)
The stretching factor for points in the stable strip i is given by exp [Λ1i (n)] ∼ 1/ε1i .
The natural measure of this strip is given by (2.70). Replacing Pi (ε ) in (1.18) by
(n)
μi , we obtain the average Lyapunov exponent λ1 > 0 as

1
n∑
(n) (n)
λ1 = − ε1i eκ n ln ε1i , (2.71)
i
2.6 General Properties of Chaotic Saddles in Two-Dimensional Maps 65

where the summation is over all stable strips. In an analogous way, the contracting
average Lyapunov exponent λ2 < 0 is obtained as
1
n∑
(n) (n)
λ2 = ε1 j eκ n ln ε2 j . (2.72)
j

The negative Lyapunov exponent ensures that trajectories from random initial
conditions approach the chaotic saddle. In fact, the time needed for the survival
probability to begin to decay exponentially (the value of n∗ in (1.8)) is approxi-
mately 1/ | λ2 |.
(n)
The metric entropy follows from (1.24) by observing that the measure μi of a
stable strip is simultaneously a path probability for all points within that strip, and
can be written as
1  
K1 = − ∑ ε1i eκ n κ n + ln ε1i .
(n) (n)
(2.73)
n i

Comparing (2.71) and (2.73), we obtain again the relation

K1 = λ1 − κ , (2.74)

which implies that the dynamics along the unstable manifold of an invertible two-
dimensional map is indeed similar to that in a one-dimensional map. This can also
be seen by examining the partial information dimension along the unstable manifold
(cf. Fig. 2.16). In particular, from (1.22), we have

(n) (n)
(1) ∑i μi ln μi
D1 = (n) (n)
, (2.75)
∑i μi ln ε1i

which leads to
(1) κ
D1 = 1 − . (2.76)
λ1
The partial information dimension along the stable manifold is given by

(n) (n)
(2) ∑ j μ j ln μ j
D1 = (n) (n)
. (2.77)
∑ j μ j ln ε2 j

Since the denominator is n times the negative Lyapunov exponent λ2 , we obtain

(2) λ1 − κ λ1 (1)
D1 = = D . (2.78)
|λ2 | |λ2 | 1

Relations (2.76) and (2.78) are the dimension formulas, also called the Kantz–
Grassberger relations [380], for chaotic saddles arising from two-dimensional maps.
66 2 Transient Chaos in Low-Dimensional Systems

The total information dimension is the sum of the partial dimensions:

1 1
D1 = K1 + . (2.79)
λ1 |λ2 |

In the limit of chaotic attractors where κ = 0, K1 = λ1 , (2.79) reduces to the

Kaplan–Yorke formula [386] for two-dimensional maps.
One can see that the metric entropy is the product of the magnitude of the
Lyapunov exponent and the partial information dimension along either the stable
or the unstable direction [564, 773]:

(1) (2)
K1 = λ1 D1 = |λ2 |D1 . (2.80)

For a two-dimensional map, there is only one positive Lyapunov exponent. The
topological entropy K0 is then determined by the exponent and its cumulants, as for
a one-dimensional map. Formula (2.22) thus remains valid. Similarly, the formula
for the escape rate can be generalized by replacing D0 in (2.21) by the partial box-
counting dimension along the unstable direction. We have
  1 
(1) (1) 2
κ = 1 − D0 λ 1 + 1 − D0 Q2 + · · · . (2.81)
2
The box-counting and information dimensions Ds,i and Du,i (i = 0, 1) of the stable
and unstable manifolds are related to the partial dimensions as

(1) (2)
Ds,i = 1 + Di and Du,i = 1 + Di , (2.82)

since the manifolds are locally smooth (one-dimensional) curves. The dimension of
the full chaotic saddle can also be expressed through the manifold dimensions as

Di = Ds,i + Du,i − 2. (2.83)

Note that in the dimension and entropy formalism discussed, the roles of the sets
(n) (n)
{ε1 } and {ε2 } are not equal, since the natural measure is connected with one of
(n) (n)
them only; cf. (2.69). Interchanging {ε1 } and {ε2 } in fact yields dimension and
entropy formulas associated with the natural distribution of the inverted map.

2.6.3 Information-Theoretic Arguments

The dimension and entropy formulas treated in Sect. 2.6.2 play a central role in the
study of transient chaos, which can in fact be derived using an information-theoretic
approach. In particular, the evolution of a chaotic system is unpredictable in long
terms. In communication, it was realized by Shannon in 1948 that a sequence of
2.6 General Properties of Chaotic Saddles in Two-Dimensional Maps 67

events conveys information if the events are not fully predictable [71, 283, 708].
The fundamental unpredictability of chaos thus implies that chaotic systems can be
regarded as sources of information. We have seen that the metric entropy K1 is the
rate at which information flows toward the significant digits, with the mean flow ve-
locity given by the average Lyapunov exponent λ1 (Sect. 1.2.3.3). This observation
led Kantz and Grassberger to argue [380] that the partial information dimension
(1)
D1 is nothing but the density of information per digit on average. As a result,
(2.80) expresses that the flow rate equals the product of velocity and density. The
same amount of information flow is expected along the stable direction.
Escape can take place only along the unstable manifold. If its partial information
dimension were unity, no exponential decay would occur along this direction. The
(1)
difference (1 − D1 ) is thus proportional to the escape rate. Since the velocity of the
(1)
information flow is λ1 , we have κ = λ1 (1 − D1 ), which is equivalent to (2.76).
Similarly, the relation λ1 = K1 + κ (see (2.74)) can be viewed as a consequence of
the fact that only a portion of the mean velocity of the information flow contributes
to information generation, since only the fraction K1 /λ1 of the mean velocity con-
tributes to the unpredictability associated with the dynamics on the nonattracting
set. The remaining fraction κ /λ1 is in fact associated with the process of escape.
Although the information-theoretic arguments are heuristic, assumptions such as
hyperbolicity and invertibility of the dynamics are not necessary. Thus the basic
formulas as given by (2.74), (2.76), (2.78), and (2.80) are expected to be valid for
nonhyperbolic dynamical systems in general.

2.6.4 Organization About Unstable Periodic Orbits

Unstable periodic orbits are the fundamental building blocks of any chaotic set,
attracting or nonattracting. A chaotic set is densely covered by an infinite number
of unstable periodic orbits, and hence they determine the natural distribution on the
set. Chaotic motion can be interpreted as a random walk among periodic orbits, or
cycles. In particular, when a chaotic trajectory approaches a specific periodic orbit,
in a short time interval the behavior of the trajectory is approximately periodic.
Since the orbit is unstable, the trajectory will leave the orbit but then approach a
different periodic orbit, and so on. This suggests that the invariant characteristics of
the chaotic set can be expressed in terms of the properties of the various unstable
periodic orbits. Indeed, natural measures associated with chaotic attractors can be
characterized by unstable periodic orbits [63, 297, 435, 459], and the same can be
done for chaotic saddles [151, 187, 765]. Chaotic saddles are the closures of all the
embedded unstable periodic orbits.
The stability of an unstable periodic orbit of period n is determined by its cy-
cle eigenvalues, which for a map system are the eigenvalues of the linearized nth
iterated map evaluated at the orbit. Chaos can thus be characterized by means of the
cycle eigenvalues, provided that sufficiently many periodic orbits can be determined
and analyzed.
68 2 Transient Chaos in Low-Dimensional Systems

For a one-dimensional map of the type in Fig. 2.1, the various cylinder lengths
can be expressed by the cycle eigenvalues. The fixed points of the nth iterated map,
all period-n points, are the intersecting points of the diagonal line with the graph of
f n . It can be seen from Fig. 2.3 that each cylinder of level n contains one n-cycle
point. The slope of f n at a cycle point is approximately the same for any point of
(n)
the cylinder containing this periodic orbit. As a result, we have εi ∼ 1/ | f n (x∗i ) |,
where x∗i is the n-cycle point belonging to cylinder i. We see that the length scales of
the generating partition for the natural distribution are determined by the stabilities
of various periodic orbits embedded in the underlying chaotic set.
Now consider a two-dimensional map. To be concrete, we can take unstable strip
j and determine its nth preimage. The overlap between them defines points that
return approximately to their initial positions after n steps, and the overlap thus
contains a hyperbolic n-cycle point. The magnitudes of the cycle eigenvalues can be
written as exp (λl∗j n) (l = 1, 2), where λ1∗j > 0 and λ2∗j < 0 are the local Lyapunov
exponents of the periodic orbit. The contracting eigenvalue λ2∗j of the period-n point
in unstable strip j defines the width of this strip:
∗
ε2 j ≈ eλ2 j n .
(n)
(2.84)

Similarly, the width of the nth preimage strip is the reciprocal value of exp (λ1∗j n):
∗
ε1 j ≈ e−λ1 j n .
(n)
(2.85)

With this expression, the measure (2.68) of the unstable strip can be expressed by
the eigenvalue of the period-n orbit in this strip as

(n) eκ n
μj ≈ ∗ . (2.86)
eλ 1 j n
This relation has been shown to hold for nonhyperbolic chaotic saddles in dissipative
dynamical systems as well [186].
We see that strips with more unstable orbits are less probable. Taking into account
(n)
the normalization condition ∑ j μ j = 1, where the summation is taken over all the
strips at level n, i.e., all the periodic orbits of length n, we see that the sum of the
reciprocals of the expanding eigenvalues is not of order unity (as it would be for
attractors), but tends toward zero for large n. The escape rate can be expressed in
terms of the eigenvalues of all period-n orbits as

1
e− κ n = ∑ λ1∗j n
for n  1, (2.87)
j e

where the summation is taken over all period-n points on the chaotic saddle. The
average Lyapunov exponent can be expressed as a similar sum, from (2.71),

eκ n
λ1 = ∑ λ1∗j ∗ for n  1. (2.88)
j eλ 1 j n
2.6 General Properties of Chaotic Saddles in Two-Dimensional Maps 69

The metric entropy (2.73) then appears as the average of λ1∗j − κ , from which (2.74)
follows.
The fractal properties of the natural measure can also be related to the cycle
eigenvalues. To see this, consider a small box of dimensions l1 and l2 along the
unstable and stable directions, respectively, about a point of the ith cycle of length n.
Since the c-measure is smooth along the unstable direction, the c-measure of the box
scales with l1 and l2 as
μc (l1 , l2 ) ∼ l1 l2α2i , (2.89)
where α2i < 1 is a nontrivial exponent, the local crowding index characterizing
the local fractal structure along the stable direction. After n iterations, the box is
stretched (compressed) by factors of exp (λ1i∗ n) and exp (λ2i∗ n) along the unstable
and stable manifolds, respectively, where λ1i∗ and λ2i∗ are the local Lyapunov expo-
nents of the whole cycle. The side lengths of the n-fold image of the box are thus
l1 exp (λ1i∗ n) and l2 exp (λ2i∗ n). Due to escape, the c-measures of the original box and
of its images are different. A stationary distribution can be obtained when the escape
is compensated for by the factor exp(κ n):
∗ ∗
eκ n μc (l1 , l2 ) = μc (l1 eλ1i n , l2 eλ2i n ). (2.90)

Using the scaling from (2.89), we obtain

λ1i∗ − κ + λ2i∗ α2i = 0. (2.91)

The natural measure μ of the same box scales with l1 and l2 as

μ (l1 , l2 ) ∼ l1 α1i l2 α2i , (2.92)

where α1i < 1 is the crowding index along the unstable direction, and α2 is the same
for both the natural and the c-measures. Preservation of the natural measure requires

λ1i∗ α1i + λ2i∗ α2i = 0. (2.93)

Equations (2.91) and (2.93) indicate that the set of local crowding indices are
uniquely determined by the cycle eigenvalues, from which α1 can be obtained as
κ
α1i = 1 − . (2.94)
λ1i∗
Since the crowding indices are a kind of local partial information dimensions, (2.94)
and (2.93) are the local analogues of the dimension formulas (2.76) and (2.78),
respectively.
70 2 Transient Chaos in Low-Dimensional Systems

2.7 Leaked Dynamical Systems and Poincaré Recurrences

2.7.1 Chaotic Saddles Associated with Leaked Systems

When a closed chaotic system is probed experimentally, the window through which
observations are made can induce “leaking” of trajectories. Escape can thus occur,
leading to transient chaos. The invariant sets of a leaked dynamical system are sub-
sets of those in the corresponding closed system. Cutting a hole in a closed system to
generate transient chaos was first suggested by Pianigiani and Yorke [596], and has
since been studied in several contexts (see, e.g., [354, 572, 573], and [19] for a re-
cent review). Since leaked chaotic systems provide a tool for a better understanding
of the closed dynamics, leaking can be regarded as a type of chaotic spectroscopy
[202, 203], or a way of “peeping at chaos” [105, 179].
To define leaking, we take a closed system described by a map fclosed (x) and
choose a subset I of the phase space Ω , which is the leaking region, or the leak for
short. A trajectory is regarded as having escaped the system after entering the region
I. The leaked dynamics can thus be described by the following map:

fclosed (xn ) if xn ∈
/ I,
xn+1 = f(xn ) = (2.95)
escape if xn ∈ I.

Since escape is considered to occur one step after entering I, the map f is defined in I.
There is a chaotic saddle in the leaked system, which is the set of points that do
not escape the complement of the leaking region I for both forward and backward
iterations and is responsible for the exponential decay of the survival probability
of trajectories in the system with some escape rate κ . The saddle is a subset of
the original chaotic set in the corresponding closed system. An example from the
Hénon map [325] is shown in Figs. 2.17a, b, where (a) shows a chaotic attractor
and (b) displays the resulting chaotic saddle when the leaking region is a disk in
the phase space. The chaotic saddle resembles mostly that of the Hénon map in the
transiently chaotic regime (Fig. 1.8) except that the unstable manifold of this saddle
is no longer a continuous curve as in Fig. 1.7. In general, the unstable manifold of a
leaked dynamical system consists of disjoint pieces separated by the leaking region
and its images.
As another example, consider the single-scale, area-preserving baker map. The
phase space Ω is the unit square. Without any leak, a typical trajectory originated
from a random initial condition visits the entire square with uniform probability.
The natural measure μ is thus the Lebesgue measure. Suppose there is a leak in
the system defined as a band extending over the full phase, where the center of the
leaking region I is chosen to be the center of the unit square. The area μ (I) of the
leak and the angle θ of the band relative to the y-axis are the two parameters that
can be changed systematically. Invariant sets of the leaked baker map for leaks of
the same area but with opposite tilt angles are shown in Fig. 2.18. The escape rates
for these sets are different: κ (25◦) = 0.11 and κ (−25◦) = 0.09. For better visibility,
2.7 Leaked Dynamical Systems and Poincaré Recurrences 71

a Leak I b Leak I
1

ρr(x)10 ρr

−1
0
−0.6 x −0.4

−1 0 1 −1 0 1
y y

Fig. 2.17 For the Hénon map xn+1 = 1 − 1.4x2n + yn and yn+1 = 0.3xn , (a) chaotic attractor (thin
line), the leaking region I and its first image f(I) (gray regions). The inset shows the projection
ρr (x) of the first image of the natural measure of the attractor in the leaking region on the x-
axis. (b) Chaotic saddle (crosses) and its unstable manifold (thin line) of the corresponding leaked
system. These invariant sets are generated using the sprinkler method (Sect. 1.2.2.3) with n0 = 40

Fig. 2.18 For the baker map given by (xn+1 , yn+1 ) = (xn /2, 2yn ) for yn ≤ 1/2 and (xn+1 , yn+1 ) =
(1 + 1/2(xn − 1), 1 + 2(yn − 1)) for yn > 1/2 with a tilted leak of area μ (I) = 0.1 at the angle
θ = 25◦ and θ = −25◦ , respectively, the stable manifold (a), (d), the chaotic saddle (b), (e), and
the unstable manifold (c), (f). The difference in the contrast of the pictures is due to the difference
in the escape rates [688] (Copyright 2002, the American Physical Society)

points of the invariant manifolds are not displayed within the leak. The leak is thus
visible in Fig. 2.18 as a white band crossing the center of the square. There are in fact
many more white bands. In particular, in the plot of the stable (unstable) manifold,
these are the preimages (images) of the leak, and in the plot of the chaotic saddle
72 2 Transient Chaos in Low-Dimensional Systems

both the preimages and the images are present. In spite of the uniform density, the
action of the baker map is asymmetric. The preimages of the left- and right-tilted
bands are quite different. As a result, the first preimage of the 25◦ leak does not
overlap with the leak, but there is an overlap in the other case. Thus, the total area
of the leak and its first preimage is larger for the positive-angle case than that for
the negative-angle case. The asymmetry persists in subsequent iterations, leading to
different escape rates.
Figure 2.19 provides an overview of the chaotic saddles for the same leak area
μ (I) but at a different set of angles θ . The textures at angles of opposite signs are
quite different. The dependence of the escape rate on both parameters is summarized
in Fig. 2.20. The escape rates associated with narrow leaks are nearly orientation-
independent. In addition, the values of κ are close to the total area of the leak.
Orientation dependence becomes significant for μ (I) > 0.05, and the amount of
fluctuations about the mean increases as the area is increased. In all cases, the intu-
itive estimate obtained from the assumption that exp (−κ ) equals 1 − μ (I),

κ (I) ≈ − ln [1 − μ (I)], (2.96)

is below the average escape rate over all the angles, but as Fig. 2.20 indicates, it
provides a good approximation for small areas.
Since all periodic orbits in the used baker map have the same Lyapunov expo-
nent, the average positive Lyapunov exponent of the chaotic saddle is λ1 = ln 2. The
topological entropy is equal to the metric entropy. From (2.74), we obtain

K0 = λ1 − κ = ln 2 − κ . (2.97)

Fig. 2.19 Chaotic saddles of the leaked baker map (μ (I) = 0.1) at leak angles θ = 0◦ (a), 45◦
(b), and 75◦ (c), and θ = −45◦ (d), θ = −75◦ (e), and θ = ±90◦ (f) [688] (Copyright 2002, the
American Physical Society)
2.7 Leaked Dynamical Systems and Poincaré Recurrences 73

Fig. 2.20 Dependence of the escape rate on the tilt angle for different leak areas μ (I) = 0.02
(plus), 0.05 (crosses), 0.1(asterisk), and 0.2 (squares). Horizontal lines correspond to the values
− ln(1 − μ (I)) [688] (Copyright 2002, the American Physical Society)

The deviation of K0 from ln 2 indicates that not all possible sequences in the sym-
bolic encoding are allowed to exist. The relatively large difference in the escape
rates for different leaks thus reflects the difference in the topological entropies and
in the rules underlying the symbolic dynamics.
From (1.14), it can be seen that the escape rate can be expressed in terms of the
c-measure of the set of points that do not escape within one iteration. This set is
Ω \ I, where Ω is the phase space of the closed system. Since μc (Ω \ I) = μc (Ω ) −
μc (I) and the c-measure of the full phase space is normalized to be unity, we have
μc (Ω \ I) = 1 − μc (I). This leads to an exact relation for the escape rate of the leaked
system:

κ (I) = − ln [1 − μc(I)]. (2.98)

That is, the escape rate is determined by the c-measure of the leak. The strong de-
pendence of the escape rate on the orientation in Fig. 2.20 reflects that the c-measure
can be drastically different even if the natural measure of the leak is kept constant
[102]. The strong dependence of the c-measure on the orientation is also a conse-
quence of the different kinds of grammatical rules in the symbolic dynamics due to
the different overlaps of the preimages.
In the limit where the leak area becomes infinitesimally small, the escape rate
approaches zero and the c-measure tends thus to the Lebesgue measure [5, 167].
It is only in the limit of μc (I) → μ (I) → 0 that the estimate (2.96) coincides with
the exact result in (2.98). In this case, we have κ = μ (I), so that both the survival
probability and the escape time distribution decay with time as

P(n) ∼ p(n) ∼ e−μ (I)n . (2.99)

74 2 Transient Chaos in Low-Dimensional Systems

The average lifetime is then

1
τ= for μc (I), μ (I) → 0. (2.100)
μ (I)

For larger leaks, the escape rate is given by (2.98), and no general formula can
be obtained for the average lifetime, since this quantity depends also on the initial
conditions.
Relation (2.99) has been used in several problems in dynamical systems ranging
from fractal exit boundaries [72,538,623,668] and the control of chaos [101,571] to
leaked billiards [43, 179–181, 475, 476, 539] and intermittency [836]. An interesting
application is the reinterpretation of Sabine’s law, a central object of architectural
acoustics. The law says that the residual sound intensity in a room decays exponen-
tially with time, and the decay rate is independent of the location of the source and
the details of the room, provided that the room’s shape is sufficiently irregular [532].
The duration to decay below the audible intensity is called the reverberation time.
What is escaping here is not trajectories, but the energy of the sound waves, and the
leak is the union of all energy-absorbing surfaces. In the language of the theory of
dynamical systems, the reverberation time is the reciprocal of the escape rate, which
is proportional to the natural measure of the leak, provided that the escape rate is
small. Sabine’s law, dating back to 1898, appears thus to be the first application of
transient chaos in the history of science.
Finally, we mention the general case of more than one separated leak. For in-
stance, with two leaks I1 and I2 , the escape rate is in general different from the sum
of the single leaked cases:

κ (I1 + I2 ) = κ (I1 ) + κ (I2). (2.101)

The difference is due to the overlap of the preimages of the two leaks [101, 574],
which has been established rigorously by Bunimovich and Dettmann [105]. In the
limit of small leaks, the difference can be expressed in terms of the correlation
function of the temporal dynamics.

2.7.2 Poincaré Recurrences

Poincaré recurrence, the return of trajectories to a specific region in the phase space,
have played an important role in the study of closed dynamical systems because
they constitute the foundation of the kinetic description of nonequilibrium pro-
cesses [201, 851, 852]. In a macroscopic system the average recurrence time is
typically large, but it tends to be smaller in low-dimensional systems. Poincaré re-
currences have also been shown to provide useful insights into the dynamics of
chaotic systems [16, 38, 136, 853]. An important quantity is the distribution pr (k)
of the first Poincaré recurrence times k to a preselected region I in the phase space.
2.7 Leaked Dynamical Systems and Poincaré Recurrences 75

As originally proposed by Chirikov and Shepelyansky [136], the recurrence time

distribution pr (k) is a useful quantity in the analysis of dynamics in the full phase
space. The Poincaré recurrence theorem ensures that for almost all initial conditions
x0 in I, there are infinitely many time instants n = n1 , n2 , . . . such that fn (x0 ) ∈ I. The
first recurrence times are defined as k = ni − ni−1 for i = 0, 1, . . . , and k = 1 if the
point remains in I. The recurrence time distribution pr (k), k ≥ 1, is the probability
of finding recurrence time k from an infinitely long trajectory. Due to the ergodicity
of the natural measure of the closed system, the specific choice of the initial point
x0 ∈ I is irrelevant for pr (k). In the original setting, the recurrence region I is taken
to be arbitrarily small, but in practical applications the size of I is finite, which we
will assume in what follows.
For a generic chaotic system, the distribution pr (k) of recurrence times is found
for any leak size to have the form [16]

fluctuations for 1 < k < kr∗ ,
pr (k) ≈ (2.102)
g r e− γ k for k ≥ kr∗ .

The asymptotic behavior is an exponential decay with a decay rate γ . The short-
time nonexponential behavior occurring before the exponential decay depends on
the choice of the recurrence region I. A consequence of ergodicity of the natural
measure is Kac’s lemma [372], which states that the mean recurrence time τr ≡
∑k kpr (k) is the reciprocal of the natural measure of the recurrence region:
1
τr = . (2.103)
μ (I)
Equation (2.103) is valid for any size of the recurrence region I. The decay rate γ ,
however, cannot be given in terms of the natural measure.
The recurrence problem can be better understood in the context of leaked dynam-
ical systems. In particular, one can choose the recurrence region to be the leak and
examine the interplay between the recurrence and escape times [18]. In this setting,
trajectories contributing to the asymptotic decay of Poincaré recurrences must, after
exiting the recurrence region I, fall into the neighborhood of the stable manifold of
the chaotic saddle of the leaked system. The long-time dynamics preceding the first
arrival back to the recurrence region is governed by the same saddle underlying the
escape process in the leaked system. As a result, the relaxation rate of the recurrence
statistics coincides with the decay rate of the escape statistics:

γ = κ. (2.104)

The slopes of the recurrence time distribution ln pr (n) and of the escape time
distribution ln p(n) are thus identical, as illustrated in Fig. 2.21 for the standard
Hénon attractor with the recurrence/leak region I chosen as in Fig. 2.17. Equation
(2.104) implies that for any Poincaré recurrence problem, there exists a chaotic sad-
dle in the corresponding leaked system and the relaxation rate γ is given by the
escape rate from the saddle. It is remarkable that in view of (2.98), the relaxation
76 2 Transient Chaos in Low-Dimensional Systems

Fig. 2.21 For the Hénon map with leaking region I chosen as in Fig. 2.17, (a) distributions of
recurrence time pr (n) and of escape time p(n). To generate pr (n), a trajectory of length 1011
originated from the center point of the recurrence region is used. The initial conditions for the
escape-time distribution p(n) of the leaked system are chosen according to the natural distribution
of the Hénon attractor. Short-time behaviors of the distributions are shown in the inset; γ = κ =
0.055. The value n∗ (respectively n∗r ) indicates the time after which the decay of p(n) (respectively
pr (n)) is exponential with good accuracy [18] (Copyright 2008, the American Physical Society)

rate of the Poincaré recurrences of a closed system is given by the c-measure (rather
than by the natural measure) of the recurrence region when viewed as a leaked
system.
Despite exhibiting the same exponential decay, the distributions pr (n) and p(n)
are different for typical initial distributions ρ0 . There is, however, a special initial
condition ρ0 = ρr for the leaked problem for which the escape statistics are fully
equivalent to the recurrence statistics. Consider the distribution obtained as the first
iterate of the points x ∈ I distributed according to the natural density ρμ of the
closed system within the leak. For invertible maps f with constant Jacobian, this can
be written as
⎧
⎨ρμ [f−1 (x)]/ μ (I) if x ∈ f(I),
ρr (x) = (2.105)
⎩0 otherwise,

where a proper normalization factor has been included in the denominator. The inset
of Fig. 2.17a shows an example: the projection of the distribution on one of the
dynamical variables of the Hénon map. Due to ergodicity, points of a long trajectory
generating the recurrence-time distribution, one iteration after returning to I, are
distributed precisely according to ρr . As illustrated in Fig. 2.22, due to the definition
(2.95), according to which escape takes place one step after entering the leak, all
escape times n correspond to the recurrence time k = n. Accordingly, we have
pr (n) = p(n) with ρ0 = ρr . (2.106)
We see that the two distributions are identical, including the short-time fluctuations
appearing for n < n∗r = n∗ . The key observation is then that any recurrence problem
2.7 Leaked Dynamical Systems and Poincaré Recurrences 77

Fig. 2.22 Illustration of the escape

equivalence between
recurrence and escape times
I

1 iteration
Ω
f (I)

n-1 iterations

can be considered as a leaked problem with the special initial condition ρ0 = ρr

(2.105).
The average lifetime is then given by Kac’s formula (2.103) τ = 1/ μ (I). For
finite I it deviates considerably from the mean lifetime estimated according to (1.7)
as 1/κ = −1/ ln [1 − μc(I)], which contains the c-measure of the leak. The reason
for the large deviation is that ρr in (2.105) is atypical from the point of view of the
c-measure that concentrates along the unstable manifold of the chaotic saddle, i.e.,
it contains points that never escape under the backward dynamics, while all points
associated with ρr come from the leak.
It has been observed that different recurrence/leak regions of the same natural
measure can lead to quite different exponential decays, particularly if an unstable
periodic orbit of low period falls into I [3, 16, 103, 571]. According to (2.104), there
is a common phenomenon behind these results. Say there is a periodic orbit of period
np in I. We then expect a high recurrence probability pr (k = np ). Since the average
recurrence time is fixed and given by (2.103), the probabilities of other recurrence
times k = np are lower, leading to a reduced decay rate. In the context of leaked
systems, the anomalously low escape rate is due to the significant overlaps between
the leak and its preimages.
Chapter 3
Crises

As a system parameter is varied, sudden and qualitative changes in the chaotic

attractor can occur, the so-called crises [292, 293]. These qualitative changes can
be seen in bifurcation diagrams where one coordinate, say x∗ , of the attractor is
plotted versus a system parameter, as shown in Fig. 3.1. Sudden shrinkage or en-
largements of the set of x∗ values are visible at several parameter values, indicating
the complexity of crisis events in a typical dynamical system.
In generic dynamical systems, there are two major types of crises: boundary
crises and interior crises. A boundary crisis occurs when a chaotic attractor col-
lides with an unstable periodic orbit (or its stable manifold) on the basin boundary,
converting the attractor into a nonattracting chaotic set and generating transient
chaos. At an interior crisis, a chaotic attractor, typically of small size, collides
with a coexisting nonattracting chaotic set (or its stable manifold), resulting in a
larger chaotic attractor. While boundary crisis is a mechanism for chaotic transients,
interior crisis provides a mechanism for sudden enlargement of a chaotic attractor
in the phase space, and intermittency. Crises are quite common in nonlinear dy-
namical systems. For example, in a parameter regime where chaotic attractors arise,
there can be various saddle-node bifurcations, generating stable periodic attractors
and consequently periodic windows [564]. In a periodic window, there is transient
chaos caused by a nonattracting chaotic set, the remnant of the chaotic attractor be-
fore the saddle-node bifurcation. The end of the periodic window is marked by a
sudden enlargement of a small attractor, which is caused by an interior crisis. As
will be discussed in this chapter, such an interior crisis leads to physically observ-
able phenomena such as crisis-induced intermittency. From a different standpoint, a
chaotic attractor can be regarded as consisting of a number of nonattracting chaotic
sets with distinct dynamical properties, which form the backbone of the chaotic
attractor.
The goal of this chapter is to present a comprehensive treatment of the funda-
mental phenomenon of crisis and the associated universal scaling laws.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 79
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 3,

c Springer Science+Business Media, LLC 2011
80 3 Crises

Fig. 3.1 Bifurcation diagram of the Hénon map xn+1 = 1− ax2n + 0.3yn , yn+1 = xn in the parameter
range 1 < a < 1.5. The parameter values of a few crisis events are indicated: an attractor enlarge-
ment point ae , the endpoints of the main period-7 window a1 , a2 , and the endpoint ac of permanent
chaos (the accumulation point of the period-doubling cascade is marked as a∞ )

3.1 Boundary Crises

A boundary crisis [292,293] occurs when a chaotic attractor “touches” its own basin
boundary at a critical parameter value pc . When this happens, certain points of the
attractor belong to the boundary as well, as shown in Fig. 3.2. As the system param-
eter is increased beyond the crisis point, say for p > pc , the boundary “penetrates”
into a region that contains part of the original chaotic attractor for p < pc , as shown
in Fig. 3.3. Since the attractor is enclosed in the unstable manifold of an embedded

Fig. 3.2 Invariant manifolds

for the Lozi map:
xn+1 = 1 − a | xn | +byn ,
yn+1 = xn . Shown are the
unstable manifold of the
upper fixed point H+ and the
stable manifold of the lower
fixed point H− for precisely
the crisis parameter value
a = ac = 2 − b/2 = 1.75,
where b = 0.5. Note that the
stable manifold is in fact the
basin boundary
3.1 Boundary Crises 81

Fig. 3.3 Stable and unstable

manifolds for the Lozi map in
Fig. 3.2 but slightly beyond
the crisis value: a = 1.8 > ac .
Almost all points in the
region now escape to the
attractor at infinity

periodic orbit, for p > pc the boundary effectively converts “pieces” of the unstable
manifold into “pieces” of the basin of a different attractor. If one piece is removed,
so are all its preimages, and hence almost all points of the unstable manifold now
belong to the basin of the different attractor. As a result, the basin of the original
chaotic attractor is destroyed and the attractor is converted into a chaotic saddle.
All these features can be visualized using the Lozi map, a piecewise linear map for
which the exact crisis parameter value can be obtained. In fact, the stable and the
unstable manifolds can be constructed analytically [762]. The manifolds at and be-
yond the crisis are shown in Figs. 3.2 and 3.3, respectively. A numerically obtained
chaotic saddle is shown in Fig. 3.4, which is to be compared with the intersection
point in Fig. 3.3.
While boundary crisis provides a mechanism for transient chaos to arise, it can
also be viewed as a mechanism for generating a chaotic attractor. Specifically, when
the crisis is approached from the side of transient chaos, the opposite situation oc-
curs: the underlying nonattracting chaotic set becomes denser and less repelling in
the phase space. At the crisis all pieces of the set are connected by the unstable
manifold, generating a chaotic attractor. From this point of view, crisis represents a
route to chaotic attractors [292, 293].
For a one-dimensional map, a boundary crisis occurs when the trajectories fill
entirely the interval available for chaos, as can be seen from Fig. 3.5a–c for a
single-humped map f (x), corresponding to the precrisis, the crisis, and the postcri-
sis situations. To make an analogy with the two-dimensional case, we can expand
f (x) into the plane as
xn+1 = f (xn , p), yn+1 = xn . (3.1)
Chaotic sets then appear on the curve x = f (y, p) is the (y, x)-plane. If f (x) is defined
on the unit interval, the fixed point at the origin can be reached along the horizontal
82 3 Crises

Fig. 3.4 Chaotic saddle of

the Lozi map for parameters
a = 1.8 and b = 0.5. A
topologically similar chaotic
saddle can be obtained with
the Hénon map for a = 1.5
and b = 0.3, which is slightly
beyond a crisis at ac ≈ 1.4269

a b c

x=f(y) x=f(y) x=f(y)

I0
y y y

Fig. 3.5 Precrisis (a), crisis (b) and postcrisis (c) configurations of a single-humped one-
dimensional map defined on the unit interval. In (a), chaotic motion is confined to the solid square

lines x = 0 and x = 1 of the plane. In this special case, the stable manifold of the
fixed point consists of these two lines in the (y, x)-plane. Insofar as f (xc ) < 1, where
xc is the point at which f (x) reaches a maximum, the chaotic attractor extends over
an interval (both in x and y) that does not contain either 0 or 1; cf. Fig. 3.5a. At the
crisis, the attractor covers the entire unit interval (both in x and y) and thus touches
the stable manifold of the fixed point, similar to Fig. 3.2.
A typical situation leading to boundary crises in chaotic dynamical systems is
shown schematically in Fig. 3.6. First, a saddle-node bifurcation generates two fixed
points: one stable and another unstable. The former undergoes a period-doubling
cascade to a chaotic attractor, while the latter remains unstable. In a large portion
of the parameter space where there is sustained chaos, the chaotic attractor often
consists of a number of pieces, and the phase-space regions where sustained chaotic
motions occur are called bands, and the regions in between are gaps. A bound-
ary crisis occurs when the stable manifold of the unstable periodic orbit created at
3.1 Boundary Crises 83

Fig. 3.6 A schematic bifurcation diagram leading to transient chaos via a boundary crisis at pc :
black dots mark the region of permanent chaos. In the gray region the chaotic set is nonattracting.
The accumulation point of the period-doubling cascade is marked as p∞

Fig. 3.7 Extended

bifurcation diagram of the
Hénon map as a function
of parameter a (b = 0.3).
A boundary crisis occurs
at ac ≈ 1.4269, as indicated
by the vertical line. Black
dots for a > ac represent
points of the chaotic saddle
projected on the x-axis [447]
(Copyright 1994, the
American Physical Society)

the saddle-node bifurcation collides with the chaotic attractor. After the crisis only
transient chaos is present. An example illustrating the existence of a chaotic saddle
in the postcrisis region is shown in Fig. 3.7.

3.1.1 Nonhyperbolicity of Chaotic Saddles

The chaotic saddles treated in the previous two chapters have the property that
for every point on the saddle, the local stable and unstable manifolds are distinct.
Roughly, this occurs when all the local turning points of the unstable manifold
lie outside the saddle. These are hyperbolic chaotic saddles. However, situations
can arise in which there are points in a chaotic saddle at which the stable and the
unstable manifolds are tangent. Such a saddle is nonhyperbolic, which can often
84 3 Crises

Fig. 3.8 (a) A nonhyperbolic saddle from the Hénon map for a = 1.6 and b = 0.3, and
(b) histogram of angles between local stable and unstable manifolds. The angles can be arbitrarily
close to zero [454] (with kind permission from the Institute of Physics)

occur in the parameter regime slightly beyond a boundary crisis.1 An example of

a nonhyperbolic chaotic saddle is shown in Fig. 3.8a. The main difference between
this saddle and that in, e.g., Fig. 1.8 is that here a seemingly continuous arch of
points along a filament of the unstable manifold (such as, e.g., about x = 0.7, y = 0)
belongs to the saddle and can have tangencies with the stable foliation. Figure 3.8b
presents a distribution of the angles between the local stable and unstable mani-
folds. It can be seen that the angle can be arbitrarily close to zero, which should be
contrasted to hyperbolic chaotic saddles, such as the one shown in Fig. 1.8, where
the angles between the manifolds (see, e.g., Fig. 1.7) are bounded away from zero.
Since for an invertible map, the images and preimages of a tangency point are also
tangency points, they are dense on a nonhyperbolic chaotic saddle. That is, for any
point on the chaotic saddle, there are tangency points arbitrarily nearby. Further-
more, if there is one parameter value p at which the saddle is nonhyperbolic, around
this value there exists a parameter interval of finite length for which the saddles are
nonhyperbolic. Such intervals are called Newhouse intervals [546, 547].
Chaotic saddles arising in naturally open systems, such as systems after a bound-
ary crisis, often have the feature that, beyond a certain parameter value, say ph > pc ,
where pc is the crisis value, they are all hyperbolic.2 Newhouse intervals can thus
exist only in the parameter interval (pc , ph ). The results in [454] suggest, however,
that within this interval, nonhyperbolic chaotic saddles are quite typical, and the

1 It should be noted that chaotic attractors in physical systems are generally nonhyperbolic, due

to the existence of a set of points in the attractor at which the angles between the stable and the
unstable directions are zero.
2 In contrast, chaotic saddles in leaked dynamical systems are generally nonhyperbolic; see

Sect. 2.7.
3.1 Boundary Crises 85

Fig. 3.9 Parameter values of

the plane a, b that belong to
nonhyperbolic Hénon saddles
are marked by gray dots. The
gray region is bounded from
the left by the curve ac of
crisis and from the right by
the curve of hyperbolicity ah .
For b = 0.3, ah (b) = 2.89. All
saddles for a > ah (b) are
hyperbolic [454] (with kind
permission from the Institute
of Physics)

length of the largest Newhouse interval can be several hundredths of the value of
(ph − pc ). The close vicinity of the boundary-crisis point is always covered by
Newhouse intervals. Crisis itself is a phenomenon strongly related to nonhyperbol-
icity, since it is accompanied, by definition, by the formation of tangencies between
manifolds.
Results about the Newhouse intervals obtained for the particular case of the
Hénon map are illustrated in Fig. 3.9 in the parameter space where parameter b is
between zero and unity. The fraction of parameter values for nonhyperbolic chaotic
saddles is for many b values more than 20%. The maximum length of the Newhouse
intervals are of order 0.1 [454].
An important property of hyperbolic dynamical systems is their structural sta-
bilities, where small changes in the parameters cause small changes in the angle
distribution and to the system dynamics [300]. Dynamical invariants characterizing
the underlying chaotic sets thus change smoothly with the parameters. In particular,
one such quantity, the topological entropy, does not change outside the Newhouse
intervals. It is the tangency between the stable and the unstable manifolds that gen-
erates new periodic orbits, leading to an increase in the topological entropy. When
there are no tangencies, the topological entropy remains constant. In this sense,
it can be said that crisis is generally an event that increases the “complexity” of
the system. The subsequent homoclinic and heteroclinic crossings of the stable and
the unstable manifolds beyond pc create an infinite number of new unstable peri-
odic orbits and consequently an increase in the topological entropy of the resulting
chaotic set. An example of the topological entropy of the chaotic saddles from the
Hénon map versus a system parameter is shown in Fig. 3.10. Note that there are
plateau regions where the saddle is hyperbolic (complements of the Newhouse inter-
vals). The function K0 (p) exhibits a devil-staircase character [125, 466]. At ah (b),
K0 reaches its maximum value: all symbolic sequences can occur (the symbolic
dynamics is complete). Beyond ah (b), the topological entropy remains constant.
In spite of nonhyperbolicity, the survival time probability exhibits exponential
decay, and the escape rate is well defined for any p > pc and typically grows with p.
This indicates (as also Fig. 3.8b suggests) the overall dominant role of hyperbolic
86 3 Crises

Fig. 3.10 Dependence of the 1.1

topological entropy K0 in the
Hénon map as a function of
parameter a (b = 0.3). The 0.9
plateau regions correspond to K0/ln2
the white intervals of Fig. 3.9
0.8
within the gray region [125]
(with kind permission from
Elsevier Science) 0.7

0.6

0.5
1.0 1.5 2.0 2.5 3.0 3.5
a

points on the chaotic saddle in dynamics. In cases in which nonhyperbolic points

dominate, the long-term decay is no longer exponential, but a power law, as illus-
trated by the one-dimensional example in Fig. 2.13. Similar long-time decays will
be discussed in the context of nonhyperbolic chaotic scattering (Sect. 6.4).

3.1.2 Critical Exponent of Chaotic Transients

The escape rates for parameters slightly beyond the crisis value are usually small, so
that the corresponding chaotic transients have long lifetimes. This parameter regime
is thus particularly suitable for experimental investigation of transient chaos (see,
e.g., [112, 196, 479, 577, 646]). In general, the escape rate scales with the parameter
difference (p − pc ) as a power law [290, 291], for p slightly above pc :

κ (p) ∼ (p − pc )γ , p  pc , (3.2)

where γ is the critical exponent of the chaotic transients. A remarkable feature of

transient chaos is that the exponent γ is determined by the stability of an unstable
periodic orbit, the mediating periodic orbit. In systems with constant phase-space
contraction (constant Jacobian), two different situations can arise, depending on
whether the crisis is triggered by a homoclinic or a heteroclinic tangency, as shown
in Fig. 3.11, and correspondingly, there are two different expressions for γ . (Scal-
ing rule (3.2) remains valid even if the nonattracting chaotic set is a fat fractal
[321, 359]).

3.1.2.1 Heteroclinic Tangency

A heteroclinic-tangency-induced crisis occurs when the stable manifold of an un-

stable periodic orbit (say H− ) becomes tangent to the unstable manifold of another
3.1 Boundary Crises 87

a b

H+ H+

H− H−

Fig. 3.11 Schematic illustration of a crisis induced by (a) heteroclinic tangencies and (b) ho-
moclinic tangencies. The chaotic attractor is in the closure of the unstable manifold of the fixed
point H+ . In case (b) the attractor is enclosed by the unstable manifold of H−

Fig. 3.12 Configuration of

stable and unstable manifolds
slightly beyond a heteroclinic
crisis, and the characteristic
lengths of the shaded regions
from which escape takes
place, where A B is the nth
preimage of AB

hyperbolic periodic orbit (H+ ), where H+ is embedded in the attractor in the pre-
crisis regime. As p is increased slightly beyond pc , the unstable manifold of H+
crosses the stable manifold of H− at infinitely many locations, and the heights of
the “overshoots” through both manifolds are proportional to r ≡ p − pc , as shown
in Fig. 3.12. The local form of the unstable manifold in the overshoot region is a
parabola of order z, where for generic cases the order is z = 2 (in nonsmooth dy-
namical systems z can assume different values, e.g., z = 1 for the Lozi map). The
width of the parabolic piece (shaded region AB in Fig. 3.12) along the stable man-
ifold is of order r1/z . Trajectories landing in this parabolic piece leave the chaotic
saddle rapidly.
Consider the nth preimage of region AB, denoted by A B . For large n, the defor-
mation of the region is governed by the linearized map about H+ . The side lengths
of A B are of order r exp (−λ2∗ n) and r1/z exp (−λ1∗ n), where λ1∗ and λ2∗ are the
local Lyapunov exponents of the mediating periodic orbit H+ (a fixed point in our
case), so that exp (λ1∗ ) and exp (λ2∗ ) are the moduli of the expanding and contract-
ing eigenvalues at H+ , respectively. Now focus on trajectories that stay near the
chaotic saddle for a finite amount of time before escape. For such a trajectory, when
88 3 Crises

it enters A B , it will subsequently fall in AB. The escape rate is thus approximately
the probability for the trajectory to land in the region A B , which is in fact the
c-measure associated with the region. For small r, we can write
κ (r) ∼ μc (A B ) ∼ rγ . (3.3)
We have seen in (2.89) how the c-measure in a box about a periodic orbit scales with
the lengths of the edges parallel to the unstable and the stable directions. With the
side lengths of region A B shown in Fig. 3.12, we have
∗ ∗ ∗
 ∗
α2
κ (r) ∼ μc (r1/z e−λ1 n , re−λ2 n ) ∼ r1/z e−λ1 n re−λ2 n ∼ r1/z+α2 , (3.4)

where the local crowding index α2 follows from relation (2.91) applied to the fixed
point H+ as
λ∗ −κ
α2 = − 1 ∗ . (3.5)
λ2
Slightly above the crisis, κ is small. We thus have α2 ≈ −λ1c ∗ /λ ∗ , where the
2c
eigenvalues are those evaluated at the crisis. These considerations yield

1 λ∗
γ= + 1c
∗ |. (3.6)
z | λ2c

This formula reveals that the critical exponent is determined by the stability of H+
at the crisis and the singularity properties of the map as characterized by the degree
z of the tangency.

3.1.2.2 Homoclinic Tangency

For a homoclinic-tangency-induced crisis, the stable and the unstable manifolds of

the mediating unstable periodic orbit H− become tangent to each other. This is the
case for Hénon-type maps with positive Jacobian. Since Poincaré or stroboscopic
maps derived from flows can have only positive Jacobians [773], this case is generic
for flows and actual experimental systems.
To obtain the critical exponent, it is useful to first examine the situation at crisis.
Choose the height of the parabolic region (again of order z) of the base AB around
the first homoclinic tangency point to be r 1, as in Fig. 3.13. This value is chosen
so that the unstable manifold of the fixed point pokes over to the other side of the
stable manifold by the same amount r for p > pc . We thus write r ∼ p − pc , and the
natural measure μ (r) of the parabolic region of base AB of the attractor is propor-
tional to the escape rate κ (r). Let A B denote the baseline segment of the parabolic
region, which is a preimage of the tangency region lying close to H− . The height of
this region is also proportional to r.
Consider next a small box around the fixed point of side lengths l1 and l2 along
the unstable and stable directions, respectively. The natural measure of the box
scales with l1 and l2 as, (2.92),
3.1 Boundary Crises 89

r A
r1/z
B

H− A’ B’
r1/z

Fig. 3.13 Schematic illustration of the configuration of the stable and unstable manifolds of the
fixed point H− at a homoclinic crisis. The characteristic lengths and the shaded region whose
measure determines the escape rate are indicated. Close to the fixed point, the attractor contains
several branches that appear parallel to the unstable manifold

μ (l1 , l2 ) ∼ l1α1 l2α2 . (3.7)

In this special situation the measure is not proportional to the box length l1 along the
unstable direction, so we can assume α1 = 1. After n iterations the box is stretched
and compressed by factors of exp (λ1c ∗ n) and exp (λ ∗ n) along the unstable and the
2c
stable manifolds, respectively, where λic∗ (i = 1, 2) are the local Lyapunov expo-
nents of the mediating periodic orbit, H− , at the crisis. The side lengths of the
n-fold image of the box can thus be written as l1 exp(λ1c ∗ n) and l exp (λ ∗ n), and
2 2c
∗ ∗
the preservation of the natural measure μ (l1 , l2 ) = μ (l1 eλ1c n , l2 eλ2c n ) implies
∗ ∗
λ1c α1 + λ2c α2 = 0, (3.8)

which is similar to that found for arbitrary points on a chaotic saddle, (2.93).
Since the images of the tangency point and of the parabolic region AB will, for
small enough r, fall into the box about H− , we expect [291] that the scaling with the
distance along the unstable direction goes with α1 :

μ (r) ∼ rα1 .

Similarly, for measure μ of the parabolic region of base A B the scaling with
vertical size l2 remains unchanged. As Fig. 3.13 indicates, the measure is differen-
tiable along the unstable direction. The measure of the parabolic region is therefore
also proportional to the size r1/z of the base. We have

μ (r) ∼ rα2 +1/z .

Since the parabolic region of base AB is the image of that of base A B , the measures
are equal. We obtain
α1 = α2 + 1/z
90 3 Crises

relating the two crowding indices. Using (3.8), we obtain α1 . The scaling μ ∼ rγ
implies γ = α1 , and thus
∗ |
1 | λ2c
γ= ∗ | −λ ∗ . (3.9)
z | λ2c 1c

This result indicates indeed that α1 = γ = 1, i.e., the natural measure does not grow
linearly along the unstable manifold. The critical exponent is now determined by
the stability properties of the mediating periodic point, H− , on the boundary and the
degree z of tangency.

3.1.2.3 One-Dimensional Maps

For one-dimensional maps, close to crisis, the escape rate is proportional to the size
of the primary escaping interval I0 , i.e., the length of the primary escape interval
for f (x) > 1, as shown in Fig. 3.5c. For maps defined on the unit interval with local
maxima at xc of order z, where

f (x) ≈ f (xc ) − b|x − xc|z , (3.10)

the size is
1/z
f (xc ) − 1
L=2 . (3.11)
b
Since the overshoot f (xc ) − 1 is proportional to p − pc , we have L ∼ |p − pc |1/z .
The critical scaling exponent for one-dimensional maps is thus given by
1
γ= . (3.12)
z

This result can also be obtained from the two-dimensional results for the strong
dissipation case, i.e., when the magnitude of the negative Lyapunov exponent tends
to infinity. We see that the exponent γ is larger for two-dimensional maps, suggesting
that chaotic transients are more persistent in higher-dimensional systems [290].
In higher-dimensional systems the escape rate may be much smaller than that
determined by a power of the parameter difference from the crisis value. In this
case, chaotic transients can be superpersistent, which will be treated in Chap. 8.

3.2 Interior Crises

In the case of an interior crisis, there is a sudden increase in the size of a chaotic
attractor as a parameter p passes through a critical value pc . Because unstable peri-
odic orbits are dense on chaotic attractors, at interior crisis there is a sudden increase
in the number of unstable periodic orbits, which are mostly from the coexisting
nonattracting set in the precrisis regime.
3.2 Interior Crises 91

3.2.1 An Example of Interior Crisis

An enlightening example of interior crisis is provided by the so-called Ikeda map,

an invertible two-dimensional map that is an idealized model of a laser cavity.
The system is expressed in terms of a complex variable z = x + iy as

ip
zn+1 = A + Bzn exp ik − , (3.13)
1 + |zn |2

where the modulus and the phase of z represent the amplitude and the phase of
the electromagnetic field of the laser pulse, respectively, and A, B, k, and p are
parameters. A dramatic example of interior crisis, when A, B, and k are fixed and p
is varied, is shown in Fig. 3.14a,b, where a small four-piece chaotic attractor (a) be-
comes a single-piece but much larger chaotic attractor (b). The parameter difference
between the two cases is of order 10−4 , but the chaotic attractor emerging after the
crisis is large and appears characteristically different from the small attractor be-
fore the crisis. This can be understood by noting that for p < pc , there is a large
chaotic saddle coexisting with the small chaotic attractor, as shown in Fig. 3.15, for
the same parameter values as for Fig. 3.14a. Figure 3.15 indicates the fractal charac-
ter associated with typical chaotic saddles in two-dimensional maps: there are gaps
of different sizes along both the stable and the unstable directions. The similarity
of the overall structure of the saddle to that of the chaotic attractor after the crisis
is remarkable (Fig. 3.14b versus Fig. 3.15). At the interior crisis the chaotic saddle
collides with the small attractor, and together they make up the large attractor. The
unstable periodic orbits of the saddle are then added to the original attractor, and
new periodic orbits are created due to heteroclinic connections.
The rising of the chaotic saddle is in general due to a long sequence of bifur-
cations, in the course of which small-size chaotic attractors are created, typically
via saddle-node bifurcations and period-doubling cascades, and are destroyed by
boundary crises. The large chaotic saddle is thus partly a result of a series of bound-
ary crises prior to the interior crisis. The scenario is schematically illustrated in
Fig. 3.16, a typical bifurcation diagram of a dynamical system exhibiting an interior
crisis. Notice that accumulation of the period-doubling bifurcations occurs before
the crisis, so there is a small multipiece chaotic attractor present right before pc .
The scales indicate how dramatic the increase in the extent of chaos is at the interior
crisis point pc .
A somewhat related phenomenon is attractor-merging crisis. In this case, in the
precrisis regime p < pc , two chaotic attractors coexist, each having its own basin of
attraction. As p is increased, the two attractors enlarge, and at pc they collide with
the basin boundary separating their basins. Merging crisis can happen in systems
possessing some symmetry whereby the precrisis attractors, as well as their basins,
are symmetric images of each other in the phase space.
Another situation in which interior crisis occurs is band-merging crisis [649],
the pairwise merging of chaotic bands (the inverse of the period-doubling cas-
cade taking place in the chaotic regime when approaching the accumulation point
92 3 Crises

Fig. 3.14 For the Ikeda map

(3.13) for A = 0.85, B = 0.9,
and k = 0.4, an interior crisis
occurs at pc ≈ 5.169789.
(a) For p = 5.1697 < pc ,
a small four-piece chaotic
attractor and (b) a much
larger and characteristically
different chaotic attractor for
p = 5.17 > pc

from above). In particular, if the attractor has 2m pieces, these chaotic bands appear
as separate attractors of the 2m -fold iterated map that merge pairwise at the crisis
point. Note that the coexistence of a nonattracting set in this case is not necessary
for the crisis, since at least two chaotic components already exist before the crisis
(although in the form of attractors).
Nonattracting chaotic sets can also undergo crises [384, 641, 642]. If two
chaotic saddles merge [423, 451], the new saddle is not only more extended but
also more dense, as characterized by an increased box-counting dimension.
3.2 Interior Crises 93

Fig. 3.15 For the same

parameter setting as in
Fig. 3.14a, the chaotic saddle
coexisting with the small
four-piece chaotic attractor

Fig. 3.16 A schematic

bifurcation diagram of a x*
nonlinear system with an
interior crisis, beyond which
the extent of chaos is
characteristically increased
(black dots denote attractor
points). The increase can be
attributed to transient chaos
that exists before the crisis

p∞ pc p

When more than one parameter of the system is changed simultaneously, it is

possible to move along a crisis line determined by boundary crises. The chaotic
attractor keeps touching its basin boundary along this line, which turns out to be in-
terrupted by many tiny intervals where the attractor is periodic [563]. It may happen,
however, that the attractor undergoes an interior crisis so that its size is increased.
It can also happen that its basin suddenly increases in extent. In this case, nondif-
ferentiable breakpoints occur along the crisis line [255]. Such a point represents a
double crisis characterized by simultaneous sudden changes in the structure of both
the chaotic attractor and its basin boundary.

3.2.2 Periodic Windows

The parameter regions where chaotic attractors exist are generically interspersed
with regions of periodic windows. The beginning of a periodic window is marked by
94 3 Crises

a saddle-node bifurcation that generates a pair of stable and unstable periodic orbits.
Immediately after the bifurcation point, the stable periodic orbit replaces the original
chaotic attractor, which becomes a nonattracting chaotic set and embeds in it, among
others, the unstable periodic orbit born at the saddle-node bifurcation. The periodic
attractor can go through a cascade of period-doubling bifurcations, resulting in a
chaotic attractor typically of small size in the phase space (as compared with the
original chaotic attractor). The end of the periodic window is caused by the collision
of the small chaotic attractor with the nonattracting chaotic set, mediated by the
unstable periodic orbit created at the original saddle-node bifurcation. A periodic
window has period m if the stable and the unstable periodic orbits generated by the
saddle-node bifurcation have the period m. As examples, a period-3 window from
the logistic map is shown in Fig. 3.17a, and a period-7 window from the Hénon
map is shown in Fig. 3.17b. For any periodic window, a generic feature is that a
nonattracting chaotic set exists throughout the window. For example, for the period-
3 window in Fig. 3.17a, there is a chaotic repeller, whose natural distribution is
shown in Fig. 3.18. Figure 3.19 displays the natural measure of the chaotic saddle
in the period-7 window of the Hénon map displayed in Fig. 3.17b.

Fig. 3.17 (a) Main period-3 window of the logistic map xn+1 = rxn (1 − xn ) in the parameter range
3.828 < r < 3.857. (b) A period-7 window from the Hénon map xn+1 = 1 − ax2n + 0.3yn , yn+1 = xn
for 1.227 < a < 1.2727

Fig. 3.18 Chaotic repeller

and its natural distribution in
the period-3 window of the
logistic map coexisting with
the period-3 attractor
(denoted by large dots) for
r = 3.832
3.2 Interior Crises 95

Fig. 3.19 Chaotic saddle and

its natural distribution in the
period-7 window of the
Hénon map in Fig. 3.17b,
which coexists with a
period-7 attractor (indicated
by large dots). The
parameters are b = 0.3 and
a = 1.25 (Picture by M. Gruiz
and Sz. Hadobás)

It can then be said that periodic windows are parameter regions where transient
chaos occurs. The presence of nonattracting sets in the windows can be understood
by observing that the infinite number of unstable periodic orbits born earlier, for ex-
ample via various period-doubling cascades, cannot simply disappear. Most of these
orbits are embedded in the nonattracting chaotic set [106]. In fact, if a nonattracting
set exists for some parameter value in the window, it must be present for all other
parameter values, since the set is generally hyperbolic and hence robust against pa-
rameter changes. As a result, the topological entropy K0 of the system within an
entire periodic window is approximately constant. Although the small chaotic at-
tractor has embedded within itself infinitely many periodic orbits, they represent
only a negligible fraction of all periodic orbits present in the window. As a result,
the topological entropy of the small chaotic attractor is usually much smaller than
that of the coexisting nonattracting chaotic set. For the period-3 window of the lo-
gistic map, for example, the topological
 entropy of the attractor is K0 ≤ ln(2)/3,
√  
while that of the repeller is K0 = ln 1 + 5 /2 .
For further analysis of transient chaos in periodic windows, we summarize in
Fig. 3.20 the genesis of a period-m window. The saddle-node bifurcation occurs at
parameter value pb . The basin of the periodic attractor is bounded by the stable
manifold of the accompanying unstable period-m orbit. (In one-dimensional maps,
the unstable orbit and its first preimage bound the basin of attraction.) The unstable
period-m orbit is in fact the mediating orbit (denoted by M) of the interior crisis that
destroys the window at the end (at p = pc ). As the parameter changes, the period-m
attractor undergoes period-doubling bifurcations and evolves into a small chaotic
attractor that contains m · 2n pieces. The chaotic attractor gradually grows within
its basin, characterized by a gradual decrease in n toward unity. The small chaotic
96 3 Crises

Fig. 3.20 Genesis of a period-m window, where the bifurcation diagram of one component of
the period-m attractor is shown. The dotted region represents the band region B bounded by the
basin boundary of the m-piece attractor. The unstable period-m orbit, the mediating orbit M, is on
the basin boundary. For p = pb , a large chaotic attractor loses its stability due to a saddle-node
bifurcation, and is converted into a nonattracting chaotic set that exists for pb < p < pb in the
surrounding region S

attractor thus consists of m pieces, and collides with the chaotic saddle at the interior
crisis value pc . Since the basin of attraction of this attractor consists of m pieces in
the phase space (or m bands in the bifurcation diagram), it is convenient to refer to
the basin of attraction as the band region (B). The complement of it can be called the
surrounding region (S). As p passes through the critical value pc , a larger attractor
emerges: the small chaotic attractor in region B collides with the extended chaotic
saddle in S, as mediated by the orbit M on the boundary between the B and S regions.
An example from the Hénon map illustrating the two types of invariant sets in the
phase space immediately before and after pc is shown in Fig. 3.21.
Since transient chaos exists in all periodic windows, it can be said that chaos
exists on continuous parameter intervals, although chaotic attractors occur on a set
of parameters that can be characterized as a fat fractal [234, 564, 773]. In particular,
the Lyapunov exponent of the chaotic set, being nonattracting or attracting, is a
monotonically increasing function of the parameter beyond the accumulation point
of the period-doubling cascade, in contrast to the wild alternation between regions
of positive and negative values when it is evaluated, as usual, with respect to the
actual attractor. That is, transient chaos is robust.
When examining the parameter space of a chaotic system, e.g., the (a, b)-plane
of the Hénon map, periodic windows appear to be organized in the following way,
as described by Gallas and coworkers: they typically have a shrimp-like structure
with sharp boundaries [92, 93, 247, 253, 254, 490]. Every shrimp contains a region
of periodicity k, an infinite succession of adjacent layers of periodicity k × 2n , for
n = 1, 2, 3, . . . , of the period-doubling route toward chaos, and the layers contain-
ing the small-size chaotic attractors resulting from the period-doubling bifurcations.
Most shrimps consist of four main thin and long legs, as can be seen in Fig. 3.22,
3.2 Interior Crises 97

a b

Fig. 3.21 For the Hénon map, at the end of the period-7 window in Fig. 3.17b, (a) chaotic saddle
(light dots) and the coexisting 7-piece chaotic attractor (heavy dots) for a = 1.266 < ac . The dotted
line represents the tangent to the stable manifold of the rightmost point of the mediating period-7
orbit. (b) A single-piece, larger chaotic attractor for a = 1.276 > ac [752] (with kind permission
from Elsevier Science)

Fig. 3.22 Parameter plane of the Hénon map for 1.1 < a < 2.4, | b |< 0.4. The region of
bounded/transient chaos appears in light blue/black. Within the chaotic region, periodic windows
are present in the form of shrimps. The largest one is of period k = 5. The boundary between light
blue and black defines the line of boundary crisis, along which breakpoints, mentioned in Sect. 3.2,
can be seen. Broad color bands represent regions of the main period-doubling cascade starting
from a fixed point (dark blue) [253] (Copyright 1993, the American Physical Society)
98 3 Crises

which leads to a remarkable property: shrimps of all periodicities define a complex

connected network of finite volume within the chaotic region of the parameter space
in which the attractor is periodic. In fact, there exists a nonattracting chaotic set at
any point of this network, further illustrating the ubiquity of transient chaos.

3.3 Crisis-Induced Intermittency

For parameter values slightly beyond the interior-crisis point, a trajectory originated
in the band region B about the precrisis attractor (cf. Fig. 3.20) spends a long stretch
of time there. It can enter the surrounding region S only when it comes close to
the boundary between B and S. Because of the relatively dense structure of the
stable manifold of the nonattracting chaotic set in B, the remainder of the precrisis
attractor, the trajectory tends to spend relatively short time in S before crossing
the boundary back to B, and so on. The dynamical behavior of the trajectory is
thus characteristic of intermittency, and this phenomenon is known as crisis-induced
intermittency [290], where a sequence of trajectory points falling in the surrounding
region S is called a burst. The dynamical behavior of the bursts is characteristically
different from that of motions in the band region B, as shown in Figs. 3.23 and 3.24.
The intermittent switches of trajectories between the band and the surrounding
regions can be described more accurately in terms of the mediating orbit. In particu-
lar, when a trajectory comes close to the boundary between the two regions, it in fact
does so by approaching the mediating orbit M along its stable manifold and leaves

Fig. 3.23 Crisis-induced intermittency in an electronic-circuit experiment by Kyprianidis et

al. [432]. The bifurcation parameter is a capacitance in the circuit, denoted by C. For this particular
experimental setting, a crisis occurs as C is decreased through the critical value of Cc = 60 nF. That
is, the precrisis and postcrisis regimes correspond to C > Cc and C < Cc , respectively. Shown are
two voltage signals in the precrisis region for C = 60.1 nF (Copyright 1995, the American Physical
Society)
3.3 Crisis-Induced Intermittency 99

Fig. 3.24 For the same

circuit system as in Fig. 3.23,
the respective voltage signals
in the postcrisis regime for
C = 58 nF. The signals appear
intermittent with bursts,
which is characteristic of
crisis-induced intermittency
[432] (Copyright 1995, the
American Physical Society)

along its unstable manifold. The trajectory’s motion in the vicinity of M follows
approximately its dynamics. For example, if M is of period m, then the trajectory
tends to exhibit a periodic motion of the same period at least temporarily when it
is near the boundary. The evolution of a typical trajectory in the postcritical regime
thus follows the following scenario:

approximately
(chaos)1 → → (chaos)2 → (chaos)1 →
periodic

approximately
→ → (chaos)2 → (chaos)1 → · · · (3.14)
periodic

Note that for conventional intermittency in nonlinear dynamical systems [612],

there is only one chaotic phase and a trajectory tends to spend long stretches of
time in some regular phase. For crisis-induced intermittency, however, both phases
are chaotic with typically different positive Lyapunov exponents.
A basic experimentally accessible quantity characterizing crisis-induced inter-
mittency is the average time τ between successive bursts. At the crisis value, there
is no burst, so τ is infinite. Beyond crisis, bursts become more frequent so that τ
decreases. Of particular interest is then how τ scales with the parameter difference
p − pc . For p > pc , the precrisis attractor is no longer an invariant set, but in the
band region, where it is originally situated, there is a nonattracting set. Associated
with this nonattracting set is an escaping problem similar to that with, for example,
a boundary crisis. We thus expect the escape-rate scaling (3.2) to hold. Since τ is
proportional to the reciprocal of the escape rate, we have [290]

τ (p) ∼ (p − pc )−γ , p  pc , (3.15)

where γ is the critical exponent that is determined by the local Lyapunov exponents
of the periodic orbit mediating the crisis (see Sect. 3.1.2). The scaling law (3.15)
100 3 Crises

is quite general: it applies not only to interior crises, but also to band-merging
crises where the chaotic attractors before the crisis are replaced by nonattract-
ing chaotic sets in their respective band regions after the crisis. The validity of
(3.15) has been established extensively both numerically [713] and experimentally
[154, 196, 646, 675].
To better understand the basic dynamical components associated with crisis-
induced intermittency, we reexamine a periodic window, say of period m. Before
the crisis (in the window), there is an attractor in the m-piece band region B and
a nonattracting chaotic set located in the surrounding region S, as exemplified by
Fig. 3.20, which are dynamically separated. That is, every periodic orbit is restricted
exclusively either to the B or to the S region. After the crisis (p > pc ), there is a single
chaotic attractor containing nonattracting sets in both regions. It is thus intuitively
reasonable to regard the chaotic attractor as consisting of two basic components,
the nonattracting chaotic sets in the B and in the S regions, respectively, which are
separated by the stable manifold of the mediating orbit M. The dynamical behav-
iors of trajectories associated with them are distinct in terms of the escape rates, the
Lyapunov exponents, and the fractal dimensions. In fact, slightly beyond the crisis,
the dynamical invariant properties of the basic components are similar to those of
their respective sets before the crisis. The basic components provide a backbone for
the postcrisis attractor, both geometrically and dynamically. In this picture, the av-
erage time τ between bursts in (3.15) is nothing but the average lifetime of chaotic
transients on the basic component in the B region.
Beyond the crisis, there is a single chaotic attractor, so the basic components
are dynamically coupled. This is accomplished via heteroclinic connections, which
contribute a third, coupling component to the dynamics. This component contains
an infinite number of new unstable periodic orbits that do not exist before the crisis,
which are originated from a single periodic orbit at the crisis: the mediating orbit.
The dynamics associated with the coupling component can be seen through long-
time observations when a typical trajectory has visited both components for certain
numbers of times. As we will see below from examples of both one-dimensional and
two-dimensional maps, it is orbits in this component which fill in the gaps along the
unstable foliation of the nonattracting set before the crisis.

3.3.1 Example of Basic Components: One-Dimensional Map

Consider the main period-3 window of the logistic map close to its end. As shown
in Fig. 3.25, the dynamics restricted to the three bands B1 ≡ (xA , xa ), B2 ≡ (xb , xB ),
and B3 ≡ (xC , xc ) have the following properties: B2 is mapped onto B3 , B3 onto B1 ,
and B1 into B2 . The last property follows because the image of the maximum point
of the parabola, f (1/2) = r/4, is less than xB . The three intervals thus constitute the
basins of attraction, each containing a piece of the chaotic attractor. The boundary
consists of points of the unstable period-3 orbit xa , xb , and xc , and their respective
preimages, denoted by xA , xC , and xB , as shown in Fig. 3.25.
3.3 Crisis-Induced Intermittency 101

Fig. 3.25 Dynamical

x/4
structure of the main period-3
window of the logistic map.
The mediating period-3 orbit
(xa , xb , xc ) and the points xA ,
xC , and xB determine the
endpoints of the B and S
regions. The arrows indicate
how they map into each other. f(x)
The broken line shows the
mapping of the maximum
point under forward
dynamics: either into B2
(r < rc ), exactly on xB
(r = rc ), or outside B2 (r > rc )

xC xc xA xa xB xb
x
B3 S1 B1 S2 B2

Focus now on the mapping on the surrounding region S, the two intervals between
the three bands of the B region. Interval S1 ≡ [xc , xA ] is mapped into interval S2 ≡
[xa , xb ], but S2 is mapped onto [xc , xb ] = S1 ∪ B1 ∪ S2 . The map restricted to S ≡
S1 ∪ S2 is thus not closed dynamically: a typical trajectory started from S leaves this
region when it is mapped into B1 . There is, however, a chaotic repeller in S, as shown
in Fig. 3.18. The period-3 points xa , xb , and xc and the boundary points of B belong
also to the repeller. For r < rc there are then two chaotic sets in the phase space
that are dynamically disconnected: the three-piece attractor in B and the repeller in
S. For the crisis value r = rc , the maximum point is mapped to xB . As a result, the
three-piece chaotic attractor covers the whole basin of attraction (B1 , B2 , and B3 ).
The attractor in fact touches the repeller in S at the boundary points, making the
mediating unstable period-3 orbit M shared by both the three-piece attractor and the
repeller. The existence of M as a periodic orbit common to both the chaotic attractor
and the chaotic repeller is a key ingredient of interior crises.
For parameter values beyond crisis, i.e., for r/4 > xB , the maximum point is
mapped outside B2 , as shown in Fig. 3.25. In this case, the map restricted to region
B is no longer closed. The enlarged chaotic attractor now extends to the whole
interval I ≡ [ f (r/4), r/4]. However, the regions B and S remain well defined due
to the existence of the period-3 orbits. There are orbits that never leave the region
S and form a chaotic repeller. Analogously, it is also possible to construct the set
consisting of orbits that never escape from the B region. There is in fact an indepen-
dent chaotic repeller in B. The repeller in S resembles geometrically that before the
crisis: it has wide gaps and a relatively small fractal dimension. However, the gaps
of the repeller in B gradually appear: their width is proportional to (r − rc )1/2 , so the
fractal dimension of this set is close to unity for r slightly larger than rc . In a small
102 3 Crises

parameter interval beyond the crisis, the mediating period-3 orbit {xa , xb , xc },
originally on the boundary between B and S, still belongs to both basic components.
[B]
In this range the topological entropies of the basic components are K0 = ln(2)/3
[S] √
and K0 = ln[(1 + 5)/2]. The constant entropy values are indications of the
structural stability of the chaotic repellers.

3.3.2 Example of Basic Components: Two-Dimensional Map

Consider a period-7 window of the Hénon map, where the basin of attraction of
the seven-piece attractor is bounded by the seven branches of the stable manifold
of a mediating period-7 orbit located at the edge of the chaotic saddle, as shown
in Fig. 3.21a. The distances between the attractor and the mediating-orbit points
decrease gradually as the bifurcation parameter (say a) is increased toward the crisis
value ac , and they vanish for a = ac . Beyond the crisis, the stable manifold of the
mediating period-7 orbit can be used to define the continuations of the band and
surrounding regions (see Fig. 3.20). The remnant of the seven-piece attractor and
the continuation of the precritical saddle become the two basic components after the
crisis, as shown in Fig. 3.26. For a > ac , the chaotic attractor is structurally made up
of two chaotic saddles, one within the continuation of the band region B, the band
chaotic saddle (BCS), and another within the continuation of the surrounding region
S, the surrounding chaotic saddle (SCS). Note that the shape of the SCS is similar
to that of the chaotic saddle in the precrisis region (Fig. 3.21).
Typically, the precrisis attractor is smaller in size and less chaotic than the
coexisting nonattracting set as measured by the Lyapunov exponents. Since their
continuations beyond crisis constitute the basic components of the postcrisis chaotic
attractor, we expect the same features to hold. In particular, the positive Lyapunov

Fig. 3.26 Basic components

of the Hénon map at the same
parameter values as in
Fig. 3.21b. The band chaotic
saddle (BCS) (heavy dots) is
formed by points that never
escape the seven bands of the
B region under either the
forward or the backward
dynamics. The surrounding
chaotic saddle (SCS) in the S
region (light dots) is the
invariant subset restricted to
the surrounding region S. The
dotted line is the tangent to
the stable manifold of the
rightmost period-7 point
3.4 Gap-Filling and Growth of Topological Entropy 103

exponent and the topological entropy of the S-component tend to dominate those of
the postcrisis attractor, respectively. (For an experimental example, see Fig. 1.16.)
This can also be seen from the distribution of the local Lyapunov exponents.3
In particular, each basic component possesses its own distribution of local Lya-
punov exponents, but the distribution associated with the S-component has a larger
maximum and its center is farther away from zero. Slightly beyond crisis, for short
time scales the distribution of the local Lyapunov exponent of the attractor is ap-
proximately the sum of the distributions from the two basic components. With
respect to the chaotic attractor, its Lyapunov distribution thus suddenly widens at
crisis [751, 752].
The methodology based on the basic components has proven to be quite use-
ful in the analysis of crisis-related phenomena in a number of different fields
[94, 127–130, 526, 636, 637], and is applicable to spatially extended systems, as will
be discussed in Sect. 9.4.

3.4 Gap-Filling and Growth of Topological Entropy

Before an interior crisis, there are a chaotic attractor (typically small) and a
nonattracting chaotic set. Associated with the latter are gaps of hierarchical sizes
in the phase space. After the crisis, a larger chaotic attractor arises, which encloses
the continuations of both the attractor and the nonattracting set. As a result, the
gaps are “filled.” Dynamically, gap-filling is accomplished by the creation of a
large number of new unstable periodic orbits that are not present before the crisis.
These coupling periodic orbits provide the support for the dense filling of the gaps
after the crisis. The creation of the coupling orbits thus provides a mechanism for
the structural development of chaotic attractors. Quantitatively, this process can be
characterized by an increase in the topological entropy K0 of the chaotic attractor
after crisis. A scaling theory based on diagrammatic expansions has been developed
to calculate the increase in the topological entropy [749, 750].
To gain insight, we estimate the time l(p) that a trajectory spends in the vicinity
of the mediating orbit, M, before spreading over the surrounding region. This can
be estimated as the time spent around any hyperbolic point. Assume that a trajec-
tory falls initially to a distance proportional to the parameter difference ∼ (p − pc )
(for p > pc ) from the stable manifold of M. The distance grows to order unity in
∗
l(p) steps: (p − pc )eλ1c l ∼ 1, where λ1c
∗
is the positive local Lyapunov exponent of
the mediating orbit at crisis. We have
∗
l(p) ≈ −[ln(p − pc )]/λ1c . (3.16)

For p → pc , l(p) diverges, indicating that close to the crisis, trajectories escaping
from the band region spend a long time in the vicinity of M, in the course of which
the motion is practically indistinguishable from the behavior of the mediating orbit.

3 This distribution can be given in terms of the entropy function S(E) defined in Appendix A.
104 3 Crises

Thus, every burst starts with an approximately periodic motion whose duration l is
much longer than the period m of M. Consequently, since every coupling orbit must
contain at least one burst, (3.16) also represents an asymptotic scaling relation for
the minimum lengths of coupling periodic orbits. As the parameter increases beyond
the crisis value, l(p) decreases, indicating the appearance of new, shorter and shorter
coupling orbits. The creation of the new coupling orbits leads to an increase in the
topological entropy K0 of the enlarged attractor, determined by the growth rate of
the number Nm of points in cycles of length m (see (1.26)).
A difficulty in determining the entropy gain beyond crisis is due to the fact that
the new orbits are rather long. A diagrammatic technique has proven to be efficient
for this task [749, 750]. In particular, let the diagram

  Kn
∼ e 0 =: t n (3.17)

represent the number of unstable periodic orbits of length n in the attractor after the
crisis, i.e., Nn itself. Let

  K [B] n
∼ e 0 =: bn (3.18)

and
  K [S] n
∼ e 0 =: sn (3.19)

denote the number of periodic orbits of length n located entirely within the band and
the surrounding regions, respectively. These two sets of periodic orbits belong to two
nonattracting chaotic sets, both embedded in the enlarged attractor. The respective
[B] [S]
partial topological entropies K0 and K0 of these chaotic saddles determine the
growth rates of the two latter diagrams. The bases b and s in the representations
(3.18) and (3.19), respectively, can be regarded as the propagators for the corre-
sponding diagrams, in the terminology of diagram calculus [505]. These propagators
take into account the contributions to Nn from the periodic orbits in the band and
in the surrounding regions. Similarly, t in (3.17) acts as the propagator representing
the total number of allowed periodic orbits.
To determine t, we also have to take into account the contribution of the coupling
orbits to Nn . In fact, the essence of the gap-filling phenomenon is the growth in the
number of coupling orbits, comprising of various combinations of the orbits already
existing before the crisis. The counting of the total number of orbits of length n,
with 0, 1, 2, . . . bursts during their period, can be expressed by the following dia-
grammatic equation:
     
= +
 
+
 
+ + ··· (3.20)
3.4 Gap-Filling and Growth of Topological Entropy 105

The first two terms correspond to the periodic orbits that never escape from the B
and S regions, as represented by the diagrams (3.18) and (3.19), while the additional
terms correspond to the coupling orbits. The approximately periodic components
of the coupling orbits invoking the bursts give only a constant contribution, because
they closely follow the same mediating orbit, giving rise to a zero topological
entropy. This feature can be incorporated in (3.20) by inserting the dotted “inter-
action” diagram with the corresponding propagator k ≡ 1 at the beginning of each
burst. Thus the number of dotted insertions is equal to the number of escapes from
a close vicinity of the mediating orbit M during n time steps. Note that before each
burst, the trajectory must spend at least one step in the band region, and due to
(3.16), at least l steps to go over the mediating orbit. This implies that the lengths nb
and nl of the double-line and the dotted propagators are at least 1 and l, respectively.
The full length of each diagram term is n.
The number of the simplest coupling orbits with exactly one burst, i.e., the third
diagram on the right-hand side of (3.20), is given by

n−l n−nb 
∑ ∑
(1)
Nn ≈ bnb · knl · sn−nb −nl = C · kl sn−l + Bbn−l + Akn−l (3.21)
nb =1 nl =l

for n > l  1. Here the asymptotic coefficients A, B, and C depend on the values
of the propagators and reflect the effect of short-range correlations between sub-
sequent orbit segments. Equation (3.21) describes the simplest interaction, or
coupling, between the two chaotic saddles in the band and surrounding regions.
The number of possible orbit combinations characterizes the strength of the cou-
pling. Equation (3.20) can be viewed as a perturbation series, with (3.21) being its
first “loop order” term and the subsequent diagrams accounting for the higher “loop
order” terms.
We have seen that before the crisis, the topological entropy of the chaotic set (the
small attractor) in the band region is smaller than that of the coexisting nonattracting
chaotic set in the surrounding region. This behavior persists after the crisis. Thus,
we have for the propagators t > s > b > k = 1, which for large n implies t n > sn 
(1)
bn  1. In the limit n  l the simplest coupling term in (3.21) gives Nn ≈ Csn−l .
Since this is C/sl 1 times the unperturbed result, sn , the coupling is weak. This
fact guarantees the convergence of the perturbation series (3.20).
Since the union of all long diagrams contains, after the single line propagator,
all possible propagator combinations, the entire bold-line propagator (3.20) can be
rewritten in the following self-consistent form:
       
= + + (3.22)

By substituting the propagators, we obtain

n
∑ Nn1 Nn−n1 .
(1)
Nn ≈ t n = sn + bn +
n1 =l
106 3 Crises

With the expression for the first loop order term, (3.21), the solution to the implicit
equation for t is, in the scaling region 1 l n → ∞ [749],

t ≈ s · (1 + Ct −l ). (3.23)

By taking the logarithm, close to the crisis, the topological entropy is obtained as

[S] [S]
K0 ≈ K0 + C · e−K0 l . (3.24)

Using the property that at pc the topological entropy of the enlarged attractor
coincides with that of the nonattracting chaotic set in the surrounding region,
[S]
K0 = K0 (pc ), and combining (3.24) and (3.16), we obtain

K0 (p) − K0 (pc ) ∼ (p − pc)χ , ∗

with χ = K0 (pc )/λ1c . (3.25)

This formula gives the parameter-dependence of the increase in the topological

entropy after the crisis. The exponent χ is the gap-filling exponent, which is simply
the ratio of the topological entropy at the crisis to the Lyapunov exponent of the
mediating orbit.
Numerical verification for the scaling law (3.25) has been obtained [749, 750]
using the example of the period-7 window of the Hénon map. The method of mon-
itoring the growth of the length of an infinitesimal curve, (1.27), is efficient for
numerically calculating the topological entropy K0 (a) of the attractor. The data
K0 (a) − K0 (ac ) versus (a − ac ) on a logarithmic scale can be fitted roughly by
a straight line with slope 1.13. The theoretical value of the exponent χ can be
estimated as follows. The positive eigenvalue of the period-7 mediating orbit is
10.871/7, corresponding to the Lyapunov exponent λ1c ∗ = 0.34. The topological en-

tropy at the crisis can be estimated to be K0 (ac ) = 0.38. The scaling exponent from
the theory is then expected to be χ ≈ 0.38/0.34 ≈ 1.12. We see that the numerical
and theoretical values of the exponent agree.
Chapter 4
Noise and Transient Chaos

In this chapter, we treat transiently chaotic dynamical systems under the influence
of noise, focusing on a number of physical phenomena. Firstly, we will demon-
strate that noise can increase the lifetime of transient chaos and induce dynamical
interactions among different invariant sets of the system. As a result, the stationary
distributions of dynamical variables in a noisy system can be much more extended
in the phase space than those in the corresponding deterministic system. Secondly, if
the system has a nonchaotic (e.g., periodic) attractor but there is transient chaos due
to a coexisting nonattracting chaotic set, noise can cause a trajectory to visit both
the original attractor and the chaotic saddle, leading to an extended chaotic attractor.
This is the phenomenon of noise-induced chaos, which can arise, for instance,
when the dynamical system is in a periodic window. Of particular interest is how
the Lyapunov exponent and other ergodic averages scale with the noise strength.
Thirdly, if the system has a chaotic attractor, noise can cause trajectories on the at-
tractor to move out of its basin of attraction so that either the attractor is enlarged
or the originally attracting motion becomes transient. This is the phenomenon of
noise-induced crisis, dynamically due to noise-induced heteroclinic or homoclinic
tangencies that cause the attractor to collide with its own basin boundary. An issue
of both theoretical and experimental interest is how the average transient lifetime
depends on the noise strength.
We will also discuss physical situations in which some parameters of a dynamical
system fluctuate randomly with time. Mathematically, such systems can be mod-
eled by random maps which, in spite of the random fluctuations, can generate both
persistent and transient chaos with underlying chaotic sets of well-defined fractal
properties. However, due to the randomness, a single trajectory cannot reveal the
fractal patterns. Instead, an ensemble of trajectories is needed. In particular, if the
noisy perturbation to each trajectory is identical at any fixed time, the snapshot pat-
tern of all trajectories in the ensemble is typically fractal if the largest Lyapunov
exponent of the random system is positive. The details of the fractal pattern change
from time to time, but the dimensions remain well defined and time-independent.
When a small amount of phase-space inhomogeneity in the noisy perturbation is
present, the fractal properties of the snapshot pattern can be observed only for a
finite amount of time.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 107
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 4,

c Springer Science+Business Media, LLC 2011
108 4 Noise and Transient Chaos

4.1 Effects of Noise on Lifetime of Transient Chaos

4.1.1 General Setting

In experimental situations, an observed trajectory is always subject to some external

perturbations, e.g., of thermal or of technical origin. In the lack of any specific in-
formation about their own dynamics, other than their time scale being much shorter
than that of the original signal, we can assume that the external perturbations are
random. Their inclusion into the dynamics can be modeled via an additive noise
term, which converts the purely deterministic equation of motion into a stochastic
equation. We are thus interested in the noisy version of the continuous- and discrete-
time dynamics, which in dimensionless forms are described by

dx
= F(x, p) + σ ξ (t) (4.1)
dt
and
xn+1 = f(xn , p) + σ ξ n , (4.2)
respectively, where the parameter σ > 0 represents the noise amplitude, and the
ξ -terms are independent, identically distributed random variables of zero mean and
unit variance. The distribution P(ξ ) is assumed to be known and to be independent
of time, so that the stochastic process generating the noise is stationary. An example
of P(ξ ) is Gaussian distribution,

P(ξ ) ∼ exp (−ξ /2).

2
(4.3)

This form implies that even for small noise strength σ , the random perturbation
can be arbitrarily large, but the probabilities for large perturbations are exponen-
tially small. It is useful to broaden the class of noise by considering distributions of
the type
r
P(ξ ) ∼ exp(−ξ /r), (4.4)
where r is an even integer. The limit r → ∞ is of special importance, since it
corresponds to a uniform distribution of ξ in the interval (−1, 1), which is often
used in numerical experiments. Unless otherwise stated, we shall assume Gaussian
noise (r = 2).
Strong noise can suppress most deterministic features of the underlying system.
The chaotic characteristics of a deterministic dynamical system can be preserved
in the presence of weak noise, which is often the case in experimental situations.
We shall therefore be interested in the effect of weak random perturbations, i.e.,
σ 1. It is a general observation due to Ben-Mizrachi, Procaccia, and Grassberger
[49] that a weak noise usually does not modify the fractal characteristics of the
chaotic dynamics. What noise does is to make the dynamics fuzzy below a cer-
tain threshold scale εc in the phase space, which increases with the noise strength.
4.1 Effects of Noise on Lifetime of Transient Chaos 109

Fig. 4.1 Schematic

illustration of the effective
scaling region, in the presence
of noisy perturbations of two
different strengths σ , of the
number N(ε ) of boxes needed
to cover a fractal invariant set
of dimension Do with boxes
of size ε in the phase space of
dimension d

In an εc -neighborhood of the deterministic fractal set, trajectories fill the phase

space. The fractal nature of the set can be revealed on distance scales larger than
εc in the phase space. The breakdown of the fractal scaling depends on the noise
strength, as shown schematically in Fig. 4.1. The effect of weak noise is thus a re-
duction of the scaling region where a nontrivial fractal dimension can be extracted,
but this in general does not affect the value of the fractal dimension.
In the case of transient chaos, we expect weak noise to affect the escape rate
(Sect. 4.1.2). More generally, noise can have an effect on the probability distribution
on or about the deterministic invariant set (Sect. 4.2). The effect of noise can be
significant for dynamical systems near bifurcations or crises. These issues will be
addressed in the sequel.

4.1.2 Enhancement of Transient Lifetime by Noise

The surprising phenomenon that transient chaos can be stabilized by weak noise in
the sense that the escape rate decreases (or the lifetime increases) was first observed
by Franaszek [243] in two-dimensional maps. Here we give a heuristic explanation,
following the argument of Reimann [625–627] based on one-dimensional maps.
Consider first a weak dichotomous noise, i.e., a random signal that can take on
two values, ±σ0 , with probability 1/2 for each. At any time step the dynamics is
governed with probability 1/2 by either the map f (x) + σ0 or the map f (x) − σ0 .
We assume that σ0 is small enough that there is escape from both maps, which are
of the type of Fig. 2.1. According to (2.8), we can approximate the escape rate as
the product of the density of the c-measure with the length of the primary escape
interval. Denoting their lengths by L1 and L2 for the two maps, we find the overall
escape rate as
L1 + L2
κ = ρ (xc ) , (4.5)
2
where we have used the fact that the density of the c-measure is smooth, and, being
normalized to unity, does not change significantly due to noise. The quantity ρ (xc )
represents thus the conditionally invariant density of the deterministic map f at its
110 4 Noise and Transient Chaos

maximum point xc . Considering maps of local maxima of order z as in (3.10) and

using (3.11), we obtain

ρ (xc ) 
κ (σ0 ) = ( f (x c ) − 1 + σ0 )1/z
+ ( f (x c ) − 1 − σ0 ) 1/z
. (4.6)
b1/z
In the presence of noise, L1 is larger than and L2 is smaller than the size of the escape
window in the deterministic case. Their average is different from the deterministic
size unless the mapping function f is piecewise linear. For z = 1, the decrease in L2
exceeds the increase in L1 for z > 1. The escape rate is then smaller than that for the
deterministic case. This is the basic mechanism leading to an enhancement of the
average lifetime of transient chaos by noise.
For a general noisy perturbation characterized by probability density P(ξ ), the
map is shifted by σ ξ for any noise realization ξ . Using the same approximation,
we obtain
1/z  ∞
f (xc ) − 1 σξ 1/z
κ (σ ) = 2ρ (xc ) 1+ P(ξ )dξ , (4.7)
b −∞ f (xc ) − 1

where we have assumed that the probability for having f (xc ) + σ ξ − 1 < 0, which
does not generate escape, is negligible. For weak noise, a Taylor expansion up to
second order in ξ yields

1 1 σ2
κ (σ ) = κ (0) 1 + −1 , (4.8)
2z z ( f (xc ) − 1)2

where we have used the fact that the second moment of ξ is unity and κ (0) denotes
the deterministic result (2.8). Equation (4.8) shows that the correction is negative
for z > 1. Thus, for maps with local maxima of order larger than unity, an increase
in the transient’s lifetime is expected due to weak noise, but the opposite is true for
maps with a cusp, where z < 1.
These heuristic arguments are based on the behavior observed about the map’s
maximum. Noise can drive trajectories out of the unit interval through its edges,
or can push back trajectories that have already left the interval. By taking into
account all these effects, Reimann [627] showed that the phenomenon as sug-
gested by the qualitative arguments holds for one-dimensional maps. For two- or
higher-dimensional maps, a rigorous argument is not available. However, for the
generic situation of z = 2, the turns of the unstable manifold are locally quadratic.
In this case, stabilization of transient chaos by weak noise is expected to be typi-
cal [20].1 These considerations imply only that the lifetime can increase with the
noise strength. In fact, for stronger noise, a decrease in the lifetime can take place
[20, 243, 244, 625]. There can then be an optimal noise strength for which the es-
cape rate is minimized (or the average transient lifetime is maximized), as shown

1 Although in some experiments [70] and in time-delayed systems [850] no apparent effect of noise

on transient chaos has been found.

4.2 Quasipotentials 111

Fig. 4.2 General dependence

of escape rate on noise
strength

schematically in Fig. 4.2. For the determination of κ (σ ), perturbative methods are

also available [150, 182, 230].

4.2 Quasipotentials

The concept of quasipotentials provides a theoretical paradigm to address the effect

of noise on nonlinear dynamical systems. We shall introduce the concept and focus
on its applications to transient chaos.

4.2.1 Basic Notions

For a nonlinear dynamical system under noise, it is often desirable to know the
probability distribution over the entire phase space. A normalizable distribution can
exist only if there are attractors in the system. Based on a well-established theory in
the weak-noise limit [246, 279, 282], we summarize the results for maps of the type
of (4.2). The steady-state probability distribution W (x) can be written for Gaussian
noise with σ 1 as2
W (x) ∼ Z(x)e−Φ (x)/σ .
2
(4.9)

The proportionality factor not written out is a normalization constant depending

only on σ . The exponential factor is of special importance, since it is similar to
the form describing fluctuations in thermal equilibrium. The function Φ plays a
central role in the theory: it is the analogy of the free energy. The noise intensity σ 2
plays the role of temperature kB T . A difference from equilibrium thermodynamics
is that here, an explicit form of Φ cannot be obtained from the first principles;
Φ is therefore called the quasipotential (or nonequilibrium potential) of the map.
An example of the quasipotential is shown in Fig. 4.3. Note that neither Φ nor the
prefactor Z depends on the noise strength; each depends solely on the underlying
deterministic dynamics.

For noise of order r, W (x) ∼ Z(x)e−Φ (x)/σ .

2 r
112 4 Noise and Transient Chaos

Fig. 4.3 Schematic diagram

of the quasipotential for a
one-dimensional map that has
two coexisting fixed-point
attractors and a chaotic
repeller. The activation
energy ΔΦ in the valley of
the left fixed point is marked

The quasipotential satisfies an extremum principle and can be constructed based

on methods from Hamiltonian mechanics [281,285,311,629]. The basic observation
is that the system can come to a phase-space point x via a large number of noise
realizations. Since, however, noise is weak, there are realizations that are sharply
peaked about a single optimal realization, namely, the most probable path that leads
to x. In N iterations, the optimal path for noise should maximize the probability
   
ξ2
N
(σ ξ n )2
N
P(ξ 0 )P(ξ 1 ) · · · P(ξ N ) ∼ exp − ∑ n ∼ exp − ∑ . (4.10)
n=0 2σ
2
n=0 2

Equivalently, the path corresponds to the minimum of the “noise energy”

∑Nn=0 (σ ξ n )2 /2. The iteration process (4.2) plays the role of a constraint that can be
taken into account by means of Lagrangian multipliers η n that are effectively con-
trol variables. The task of finding the optimal path thus boils down to minimizing
the following “Lagrangian”:

N
1
L(ξ , x) = ∑ 2
(σ ξ n )2 + η n+1 [xn+1 − f(xn, p) − σ ξ n ] . (4.11)
n=0

In the presence of the multipliers, the variables ξ n and xn can be regarded as inde-
pendent. Setting the partial derivatives of L equal to zero yields, in view of (4.2), the
following coupled map between η n and xn for the optimal path3

1/(r−1)
3For noise of order r, the first equation should be replaced by xn+1 = f(xn , p) + η n+1 , for r
even.
4.2 Quasipotentials 113

xn+1 = f(xn , p) + η n , η n+1 = J(xn+1 , p)−1 η n , (4.12)

where J denotes the transpose of the derivative matrix of map f. A comparison

with (4.2) shows that η n+1 /σ is nothing but the optimizing noise process. Another
feature is that the map (4.12) is area-preserving, even if f is dissipative. It can there-
fore be regarded as a kind of Hamiltonian extension of the deterministic dynamics
xn+1 = f(xn , p) through the control variable η n .
From (4.10), the probability for a trajectory to be at point x = xN+1 after the Nth
iteration is
 
1 N 1
P(ξ 0 )P(ξ 1 ) · · · P(ξ N ) ∼ exp − 2 ∑ [xn+1 − f(xn, p)] . 2
(4.13)
σ n=0 2

In order to find a time-independent distribution of x, we take the limit N → ∞. The

quasipotential defined by (4.9) becomes
∞
1
Φ (x) = min ∑ [xn+1 − f(xn , p)]2 ||x∞ ≡x + constant, (4.14)
n=0 2

where the minimum is taken with respect to the value of the control variable at the
initial and end points.4 The quasipotential must be independent of the initial condi-
tions x0 , η 0 within the basin of attraction of an attractor A. This can be realized by
letting (4.12) evolve according to the deterministic dynamics (η n = 0, n = 0, 1, · · · )
until the attractor is reached. This initial evolution does not contribute to Φ , and
hence for all practical purposes, the initial conditions can be taken to be x0 ∈ A, and
η n → 0 for n → 0. The minimization process taken with respect to η ∞ remains to
be carried out.
Since the probabilities of visiting different regions of an attractor cannot differ
exponentially, the quasipotential is constant on the attractor. For a chaotic attractor,
the quasipotential is constant on the entire fractal set. The differences in the proba-
bilities of visiting different regions of the attractor are characterized by the prefactor
Z(x) of (4.9). Since it is σ -independent, the prefactor evaluated on the attractor
must coincide with the density ρ associated with the natural measure. The role of
noise becomes thus important outside the attractor where the essential contribution
to the dynamics is characterized by exp (−Φ /σ 2 ). In particular, the quasipoten-
tial increases with the distance from the attractor, and the “Boltzmann factor”
exp (−Φ (x)/σ 2 ) yields the probability that noise pushes a trajectory point to x,
away from the attractor. In the case of a fractal chaotic attractor, the Boltzmann fac-
tor yields the probability for a trajectory to fall between two branches of the chaotic
attractor (which is the closure of unstable manifolds) as a result of noise.
It should be emphasized that both the map in (4.12) and the quasipotential given
by (4.14) are independent of σ . They characterize the deterministic dynamics.

4 For noise of order r, Φ (x) = min ∑∞ r

n=0 r (xn+1 − f(xn , p)) |x∞ ≡x + constant.
1
114 4 Noise and Transient Chaos

4.2.2 Quasipotential Plateaus Associated with Nonattracting

Chaotic Sets

Analogous to the situation of attractors, quasipotentials are constant on chaotic re-

pellers. For example, in a one-dimensional map, the potential is constant on the
shortest intervals containing the repeller, as shown in Figs. 4.3 and 4.4. In addition,
the prefactor Z(x) is proportional to the density ρ (x) of the c-measure.
For a two-dimensional noninvertible map, the quasipotential is constant over a
two-dimensional region surrounding the repeller. The existence of such quasipoten-
tial plateaus is especially apparent in the case of Julia-type sets of the iterations of
the complex plane, as shown in Fig. 4.5, where the repeller is generated in the pro-
cess of finding the roots of the cubic equation z3 − 1 = 0 by Newton’s method. In
particular, the attractors are the three roots zi = exp (i2π /3) (i = 0, 1, 2). Their com-
mon basin boundaries form a Julia-set type repeller. The potential plateau extends
up to those points of the chaotic repeller that fall closest to the roots zi .
The situation is somewhat different for chaotic saddles arising from invertible
two-dimensional maps. For such a case, noise is more likely to push trajectories
along than across the unstable manifold. In this case, a quasipotential plateau ex-
tends along the unstable manifold of the saddle only, and the prefactor Z coincides
with the density ρ of the c-measure on the unstable manifold. In between branches
of the unstable manifold, the potential assumes larger values. As a result, Φ changes
in the stable direction outside the saddle but remains constant along the unstable di-
rection. An example is shown in Fig. 4.6 for a map whose deterministic version will
be treated in the context of basin boundaries in Chap. 5.

Fig. 4.4 Numerically computed quasipotential Φ (x) for the logistic map in the main period-3
window: f (x) = 1 − ax2 ,√a = 1.752. A chaotic repeller coexists with a period-three attractor
(cf. Fig. 3.17, where r = 1 + 4a + 1 = 3.83). The potential is chosen to be zero on the period-
three attractor. A plateau in the potential occurs on the intervals containing the chaotic repeller
[312] (with kind permission from Elsevier Science)
4.2 Quasipotentials 115

Fig. 4.5 Numerically computed quasipotential Φ (z) arising from the complex iteration zn+1 =
zn − (z3n − 1)/(3z2n ). The attractors are z1 = 1 and zi = exp (i2π /3) (i = 1, 2). In the common fractal
boundaries among the basins of the attractors lies the chaotic repeller, the Julia set, on which the
quasipotential is constant. Only the potential about the valley of root z = 1 is shown [280] (with
kind permission from World Scientific Publishing Co.)

(1)
Fig. 4.6 Quasipotential for the map θn+1 = θn + 1.32 sin 2θn − 0.9 sin 4θn − xn sin θn + σ ξn ,
(2) (i)
xn+1 = −0.9 cos θn + σ ξn , where y ≡ θ /(2π ) and ξn are random Gaussian variables. The system
has two coexisting attracting fixed points at (x, y) = (−0.9, 0) and (0.9, 0.5), respectively, which are
separated by a chaotic saddle whose unstable manifold consists of S-shaped curves, foliations of
which are approximately orthogonal to those of the basin boundary, shown in Fig. 4.10a. Observe
the quasipotential plateau along the unstable manifold [312]. The plateau’s length follows from the
minimum requirement of (4.14) (with kind permission from Elsevier Science)
116 4 Noise and Transient Chaos

4.2.3 Exit Rates from Attractor and Most Probable Exit Paths

The quasipotential, besides being the leading contribution to the stationary

distribution, governs the transient dynamics from the basin of an attractor
[44, 285, 401, 402, 423–426, 628, 630, 645] when an exit exists that allows for
trajectories to leave the basin. Under noise, the probability P(n) that a trajectory
of the system stays within the basin of attraction of an attractor for n steps decays
exponentially with n:
P(n) ∼ exp (−kn), (4.15)
where k is the exit rate from the attractor (not to be confused with the escape rate
κ for nonattracting chaotic sets). This rate k can be expressed as the integral of the
probability current through the basin boundary. Since this current is proportional
to the probability W itself, the exit rate is given, in leading order, by the integral
of exp [−Φ (r)/(σ 2 )] over the boundary. For σ → 0, the main contribution to the
integral comes from the point xe for which the quasipotential is minimal on the
boundary. This exit point is a repeller or a saddle (hyperbolic) point for smooth
boundaries.5 We thus have6

k ∼ e−ΔΦ /σ ,
2
(4.16)

where
ΔΦ ≡ Φ (xe ) − Φ (A) (4.17)
is the difference between the minimum of the potential along the boundary and
its value on the attractor.7 The mean first exit time, te , is then proportional to the
reciprocal of the exit rate
te = τ0 eΔΦ /σ ,
2
(4.18)
which resembles the Arrhenius factor, or Kramers’s relation [314] that characterizes
the escape from a potential well in the presence of thermal noise. This indicates that
ΔΦ can be regarded as a kind of activation energy, and is a basic global measure of
the stability of the attractor.
Relations (4.16) and (4.17) imply that the escape process is dominated in the
weak noise limit by a single noisy trajectory, the most probable exit path. This is a
solution of the optimal path map (4.12) with initial and final conditions

x0 ∈ A, x∞ = xe , η 0,∞ → 0, (4.19)

where the constraint that the control variable vanishes in the long-time limit comes
from the minimum condition given by (4.14). The optimal escape path from an

5 In the case of fractal basin boundaries, the exit point is a point of the nonattracting chaotic set
belonging to the fractal boundary.
6 For noise of order r, k ∼ e−ΔΦ /σ r .
7 For fractal boundaries the exit point x is on the quasipotential plateau, and hence ΔΦ is the
e
difference between its values on the plateau and on the attractor.
4.2 Quasipotentials 117

Fig. 4.7 Most probable exit path for the Ikeda map (3.13) for parameters A = 0.85, B = 0.9,
k = 0.4, and p = 3. The basin of one of the fixed-point attractors is denoted by the blank region,
while that of another fixed-point attractor (lying outside the frame) is shown in gray. Exit takes
place over the hyperbolic point xe , denoted by an asterisk, on a smooth boundary (the stable man-
ifold of xe ). Points on the exit path are connected with lines to guide the eye. This path spirals
out of the attractor and approaches asymptotically the hyperbolic point. The activation energy is
ΔΦ ≈ 1.4 · 10−2 [426] (with kind permission from Elsevier Science)

attractor has been numerically determined in a number of cases [57, 424, 426, 717,
718]. Here we show an example for the Ikeda map (Fig. 4.7).
The equation governing the dynamics of the control variable in (4.12) contains
the inverse of the deterministic derivative matrix. In the vicinity of a fixed point
of the original map, the dynamics governing η is approximately linear and is the
inverse dynamics governing the evolution of x. This enables us to estimate the rate
at which the control variable η vanishes at both ends of the most probable exit
path as

| η n |∼ exp(| λ | n), n = 1, 2, . . . , and | η n |∼ exp (−λ1 n), n → ∞. (4.20)

The control variable initially grows according to the modulus of the contracting
Lyapunov exponent λ < 0 of the attractor, and approaches, after long times, zero
with λ1 > 0, the largest positive Lyapunov exponent of the nonattracting (chaotic or
nonchaotic) set on the boundary.

4.2.4 Enhancement of Exit Rates by Transient Chaos

When a nonattracting chaotic set is present in the basin of attraction of an attrac-

tor, it is possible to reduce the activation energy ΔΦ and consequently to increase
the exit rate as described by Kraut and Feudel [423, 425, 426]. To understand
118 4 Noise and Transient Chaos

this, consider an invertible system with smooth basin boundaries, where the opti-
mal (most probable) exit path passes through an unstable periodic point xe on the
boundary. However, before reaching the boundary, the path also extends through
the chaotic saddle, since from the point of view of energy, motion along the un-
stable manifold of the saddle (or more generally, along any deterministic orbit)
does not contribute to the activation energy. From the quasipotential plateau of the
chaotic saddle, a trajectory can reach xe on the boundary with a relatively low in-
crease in the quasipotential, as exemplified by Fig. 4.8. Numerical simulations reveal
[423,425,426] that although the activation energy is smaller than that in the absence
of the saddle by about fifty percent only, the exit rate can be enhanced by several
orders of magnitude, since the reduction in the quasipotential appears in the expo-
nent of (4.16) divided by the small noise intensity. The origin of the reduction in
the activation energy can be understood by noting that the exit process actually con-
sists of three stages: reaching the quasipotential plateau of the saddle, moving on
the plateau, and leaving the plateau to reach the boundary. The chaotic saddle thus
acts as a “shortcut” for minimizing the quasipotential in the exit process.8

Fig. 4.8 A most probable exit path in the phase space of the Ikeda map (3.13) for parameters
A = 0.85, B = 0.9, k = 0.4, and p = 5. For this parameter setting, there are a fixed-point attractor
and a coexisting chaotic saddle (black dots) in the basin of the attractor. Exit from the basin is
through a hyperbolic periodic point xe (denoted by the asterisk on the smooth boundary). The exit
path in fact wanders through the chaotic saddle before approaching xe . The activation energy is
ΔΦ ≈ 7 · 10−3 [426] (with kind permission from Elsevier Science)

8 The role of a chaotic saddle in enhancing the exit rate suggests that when a dynamical system

undergoes a basin boundary metamorphosis (see Sect. 5.4.1) by which a smooth boundary becomes
fractal so that a nonattracting chaotic set arises on the boundary, the rate of exiting the basin due
to noise can be enhanced significantly. This has indeed been observed [729].
4.3 Noise-Induced Chaos 119

4.3 Noise-Induced Chaos

In a dynamical system with a regular periodic attractor, the addition of weak

noise can lead to a chaotic attractor. This phenomenon of noise-induced chaos
[104, 332, 349] occurs when there is a coexisting nonattracting chaotic set in the
phase space of the noise-free system. A typical setting in which noise-induced chaos
can occur is periodic windows. Specifically, let the system parameter p be chosen
so that the system is in a periodic window where a periodic attractor and a nonat-
tracting chaotic set coexist. Imagine placing the system under noise. As the noise
strength is increased through a critical value, the asymptotic attractor of the system
becomes chaotic, as characterized by the appearance of a positive Lyapunov expo-
nent. Here the notions of attractor and its basin of attraction are still meaningful,
since noise is weak.
Figure 4.9 presents an example of the phenomenon from the one-dimensional
logistic map in a period-8 window, where the noise-induced chaotic attractor in the
(xn−1 , xn )-plane is shown in Fig. 4.9a and a representative noisy chaotic time series
is shown in Fig. 4.9b. An intermittent behavior can be seen, where the noisy trajec-
tory visits the period-8 attractor and a coexisting chaotic repeller in different times.
A two-dimensional example is shown in Fig. 4.10. The deterministic system pos-
sesses two fixed-point attractors and a fractal basin boundary separating the basins
of attraction, as shown in Fig. 4.10a. In the presence of weak noise of strength ex-
ceeding a certain threshold value σc , the attractors merge with the unstable manifold
of the chaotic saddle to form a noisy chaotic attractor, as shown in Fig. 4.10b.

a 1
xn

0.5

0
0 0.2 0.4 0.6 0.8 1
xn−1
b 1
xn

0.5

0
0 1000 2000 3000 4000 5000
n

Fig. 4.9 Nose-induced chaos for the logistic map xn+1 = rxn (1 − xn ) + σ ξn for r = 3.8008 (in a
period-8 window of the deterministic map) and noise strength σ = 10−4.8 ≈ 1.6 × 10−5 : (a) noisy
chaotic attractor and (b) intermittent time series {xn } [458] (Copyright 2003, the American Physi-
cal Society)
120 4 Noise and Transient Chaos

Fig. 4.10 Noise-induced chaos in the map system defined in Fig. 4.6. (a) Deterministic case,
two fixed-point attractors (denoted by large black and blank dots) and their basins of attraction.
(b) Noise-induced chaotic attractor with uniform noise for σ = 0.01. A comparison with Fig. 4.6
indicates that the attractor extends through the plateau of the quasipotential, containing the unsta-
ble manifold of the chaotic saddle. The critical noise strength is σc = 0.009 [717] (Copyright 2010,
the American Physical Society)

While the coexistence of a chaotic saddle and a periodic attractor can result in
a chaotic attractor under noise [66, 171], a situation can arise in dynamical sys-
tems whereby there are some unstable periodic orbits and a periodic attractor in the
phase space. In this case, if homoclinic tangencies between the stable and the un-
stable manifolds of the unstable periodic orbit are imminent, the addition of noise
can induce homoclinic intersections and consequently a chaotic saddle, generating
the conditions for a noise-induced chaotic attractor. Such chaotic saddles are called
stochastic chaotic saddles and have been demonstrated in a class of biological sys-
tems [65, 696].

4.3.1 Critical Noise Strength for Noise-Induced Chaos

The concept of quasipotential provides a convenient way for estimating the crit-
ical noise strength required for noise-induced chaos. The first observation is that
the periodic attractor appears to be fuzzy in the presence of noise. We can de-
fine a noisy attractor as the region in which the probability distribution takes
on large values. Note that any practical observation of the stationary distribution
relies on the existence of a finite threshold resolution, χ (say 10−3 of the max-
imum of the probability density W ). To leading order, we can define, depending
on the threshold, a noisy attractor as the set of phase-space points x that satisfy
Z exp [−(Φ (x) − Φ (P))/σ 2 ] ≥ χ , where P = A denotes the periodic attractor and
Z is a constant. For small σ , the distribution is strongly localized and the exten-
sion of the noisy attractor beyond the deterministic attractor is small but increases
with the noise strength. A noise-induced chaotic attractor appears at a critical noise
strength σc where the noisy attractor touches the edge of the quasipotential plateau.
4.3 Noise-Induced Chaos 121

Denoting the quasipotential difference between the plateau and the attractor by ΔΦ ,
which is the activation energy (4.17) from the attractor, the condition for the critical
noise strength is Z exp (−ΔΦ /σc2 ) = χ . We thus obtain9

σc = ΔΦ / ln (Z/ χ ) ∼ ΔΦ 1/2 . (4.21)

A sudden spreading of the support of distribution W has indeed been observed, as

shown in Fig. 4.11 for the system of Fig. 4.6 at a critical value compatible with the
above estimate.

Fig. 4.11 Stationary probability

√ √ plotted with resolution χ = 0.01
W (x, y) of the map in Fig. 4.6
for noise strength σ = 2 · 10−3 in the upper panel and σ = 20 · 10−2 in the lower panel, where
noise-induced chaos is present [312] (with kind permission from Elsevier Science)

9 For noise of order r, we have σc = [ΔΦ / ln(Z/χ )]1/r .

122 4 Noise and Transient Chaos

For noise strength slightly above σc , the probability distribution observed with
resolution χ extends over the whole chaotic attractor, but the probability about the
original periodic attractor is much larger than that of being farther away, as shown
in Fig. 4.11. The mean first-exit time, te , from the corresponding potential well is
given by (4.18), which is equal to the average lifetime in the noisy system about the
periodic attractor.

4.3.2 Scaling Laws for Critical Noise Strength and for Lifetime
at a Saddle-Node Bifurcation

Close to certain bifurcations, a scaling law of the critical noise strength can be ob-
tained from (4.21). For example, close to the saddle-node bifurcation that initiates
a period-m window, the deterministic dynamical system can effectively be reduced
to a normal form that is one-dimensional [300]. The quasipotential about the fixed-
point attractor, the node, increases quadratically with the distance from the attractor
with a coefficient proportional to (p − pb)1/2 [44, 311]:

ΔΦ (Δx) ∼ (p − pb)1/2 Δx2 ,

where pb denotes the bifurcation point. For a smooth one-dimensional map, the
phase-space distance between the saddle and the node about the bifurcation point is
proportional to (p − pb )1/2 . Since the saddle is part of the nonattracting chaotic set,
the activation energy is

ΔΦ = c(p − pb)3/2 , (4.22)

where c is a constant. From (4.21), at a fixed resolution, the critical noise strength
scales with (p − pb) as

σc ∼ (p − pb)3/4 , (4.23)

which has been verified numerically [44, 401, 402].

The average lifetime te about the node can be calculated using (4.18), which for
a saddle-node bifurcation is
3/2 /σ 2
te (p) ∼ ec(p−pb) (4.24)

for sufficiently weak noise [207, 312]. The scaling of the prefactor with the param-
eter, not written out in (4.24), can be obtained by combining this form with the
deterministic result. As pointed out by Pomeau and Manneville [612], the mean du-
ration time of the laminar phase in intermittency is proportional to the power −1/2
of the parameter difference. This leads to the scaling law [218]

te (p) ∼ (p − pb)−1/2 g[(p − pb)3/4 /σ ], (4.25)

4.3 Noise-Induced Chaos 123

where g is an arbitrary function. A comparison of (4.24) with (4.25) fixes the scaling
function g to be
2
g(z) ∼ ecz , (4.26)
which is valid for Gaussian noise.10 This rule implies that the quantity te (p)
(p − pb)1/2 versus (p − pb)3/2 /σ 2 falls on a straight line on a semilogarithmic plot,
a scaling form that can be verified numerically [312].
It should be emphasized that the quasipotential approach is applicable if the de-
terministic influence dominates the stochastic influence. This implies that given a
fixed finite value of σ , the results presented here are valid only if p − pb exceeds
some minimal value, because noise is dominant for parameter values quite close to
the bifurcation point.

4.3.3 Appearance of a Positive Lyapunov Exponent

An alternative way of defining an attractor under noise to be chaotic is the sensitive

dependence on initial conditions, as characterized by the existence of at least one
positive Lyapunov exponent.11 In particular, in the absence of noise, since the at-
tractor is not chaotic, the largest Lyapunov exponent of the asymptotic attractor is
a negative number for maps (zero for flows). As noise is turned on and its strength
becomes sufficiently large, there is a nonzero probability that a trajectory originally
on the attracting set escapes it and wanders near the coexisting chaotic saddle. In
this case, the largest Lyapunov exponent λ1 becomes positive, indicating that the
asymptotic attractor of the system has become chaotic for trajectories starting from
random initial conditions.
The spectra of the Lyapunov exponents is changed due to the noise. This can
be seen, heuristically, as follows. For maps, let λ1P < 0 and λ1C > 0 be the largest
Lyapunov exponent of the periodic attractor and that of the chaotic saddle (or a
chaotic repeller), respectively, in the absence of noise. The largest Lyapunov ex-
ponent of the noisy system is denoted by λ1 . For σ < σc , the noisy attractor is
only a fattened version of the original periodic attractor. Thus, we have λ1 = λ1P .
For σ > σc , there is an intermittent hopping of the trajectory between regions that
contain the original periodic attractor and the nonattracting chaotic set. Let fP (σ )
and fC (σ ) be the fractions of time that the trajectory spends asymptotically in the
corresponding regions. We have

λ1 (σ ) ≈ fP (σ )λ1P + fC (σ )λ1C , (4.27)

10 For noise of order r, ΔΦ = c(p − pb )r−1/2 , and te (p) ∼ (p − pb )−1/2 g[(p − pb )1−1/(2r) /σ ] with
g(z) ∼ exp (czr ).
11 The Lyapunov exponents are the time-averaged stretching or contracting rates of infinitesimal

vectors along a typical trajectory in the phase space, which can be defined for both deterministic
and stochastic dynamical systems.
124 4 Noise and Transient Chaos

where the Lyapunov exponent changes sign for σ = σc . We thus have fC (σc )λ1C =
fP (σc ) | λ1P |. For σ about σc , the fraction of time spent near the periodic attractor
is relatively large, i.e., fP (σ ) ≈ 1, and we have

λ1 (σ ) ≈ fC (σ )λ1C + λ1P. (4.28)

We see that immediately after the noise strength exceeds the critical value σc , the
noisy attractor is chaotic in the sense that its largest Lyapunov exponent becomes
positive. For σ > σc , the periodic attractor and the chaotic saddle are dynamically
connected, but for σ slightly above σc , a trajectory visits the chaotic saddle only
occasionally. Under this circumstance the sets can be regarded as distinct but only
in an approximate sense. That is, (4.27) is valid only for σ slightly above σc .
The above discussion can be extended to continuous-time dynamical systems.
Consider, for example, a three-dimensional flow. Let λ3P ≤ λ2P < λ1P = 0 and
λ3C < λ2C = 0 < λ1C be the Lyapunov spectra of the periodic attractor and of the
chaotic saddle, respectively, in the absence of noise. Let λ3 < λ2 < λ1 be the
Lyapunov spectrum of the noisy system. For σ < σc , we have λi = λiP (i = 1, 2, 3).
In particular, there is still a null Lyapunov exponent λ1 = 0, despite the presence
of noise, indicating that the topology of the flow is preserved. The critical noise
strength, σc , is set by the condition that an intermittent hopping of the trajectory
between regions that contain the original periodic attractor and the chaotic saddle
becomes observable. In terms of the fraction fC of time that the trajectory spends
asymptotically about the saddle, this implies fC (σ ) > 0 for σ > σc . The Lyapunov
spectrum can then be written as

λ1 (σ ) ≈ fP (σ )λ1P + fC (σ )λ1C = fC (σ )λ1C ,

λ2 (σ ) ≈ fP (σ )λ2P + fC (σ )λ2C = fP (σ )λ2P < 0,
λ3 (σ ) ≈ fP (σ )λ3P + fC (σ )λ3C < 0. (4.29)

What is the way by which λ1 (σ ) varies through the transition? As we will see,
the transition is smooth in the sense that the largest Lyapunov exponent becomes
positive continuously from zero as the noise strength is increased through the critical
value σc .

4.3.4 Scaling Law for the Largest Lyapunov Exponent

Because of the averaging effect of noise, we expect the dependence on noise of the
largest Lyapunov exponent λ1C of the original chaotic set to be weak. For flows, the
largest Lyapunov exponent from (4.29) is λ1 (σ ) = fC (σ )λ1C , which is positive in-
sofar as the measured value of fC (σ ) becomes markedly nonzero. Thus the main
dependence of λ1 on noise comes from fC (σ ), the frequency of visit to the origi-
nally nonattracting chaotic set. We shall establish that for σ slightly above σc , this
4.3 Noise-Induced Chaos 125

probability obeys an algebraic scaling law:

fC (σ ) ∼ (σ − σc )α , (4.30)

where α is a scaling exponent. As a consequence, the largest Lyapunov exponent

scales with the noise variation in the same manner, i.e.,

λ1 (σ ) ∼ (σ − σc )α . (4.31)

To estimate the scaling exponent, we make use of the quasipotential concept

again and note that the chaotic saddle lies on a quasipotential plateau. The probabil-
ity of reaching the plateau is nonzero if the noisy attractor overlaps with the plateau,
i.e., if there are points of the noisy attractor for which Φ (x) − Φ (P) ≥ ΔΦ , where
ΔΦ is the activation energy. The probability of being on the plateau is the integral of
Z exp [−(Φ (x) − Φ (P))/σ 2 ], which is for small σ proportional to the minimum of
the integrand. Due to the threshold value χ introduced in Sect. 4.3.1, any measured
probability should be larger than χ . Thus, for σ slightly above σc , we can express
the probability of being on the plateau as

fC (σ ) = Ze−ΔΦ /σ − χ .
2
(4.32)

Since σc is defined as the noise strength for which Z exp (−ΔΦ /σc2 ) = χ holds
(4.21), we can write
 −2 −2

fC (σ ) = χ e−ΔΦ (σ −σc ) − 1 . (4.33)

The exponent is ΔΦ (σ 2 − σc2 )/(σ 2 σc2 ) ≈ 2ΔΦ (σ − σc )σc−3 . For sufficiently small
σ − σc , the exponential function can be expanded to yield

σ − σc
fC (σ ) ∼ ΔΦ . (4.34)
σc3

Thus, from (4.30), we have

α = 1, (4.35)
which is independent of any system details such as the phase-space dimension.
A similar argument applies to maps. Since λ1P is nonzero, both contributions in
(4.27) are nonzero, and the threshold χ can be neglected.12 Since the problem is
basically a two-state problem with a periodic and a chaotic state, the probability of
being in state i = P,C can be expressed as a ratio of the average lifetimes:

τi (σ )
fi (σ ) = . (4.36)
τP (σ ) + τC (σ )

12 The scaling law and the exponent remain unchanged even if the threshold is not neglected [776].
126 4 Noise and Transient Chaos

The lifetime τP ≡ te about the periodic attractor is given by the Arrhenius factor
(4.18). The lifetime about the nonattracting chaotic set depends nonexponentially
on the noise strength and can be considered to be constant, i.e., τC ≈ 1/κ , where κ
is the escape rate of the chaotic set. Taking this into account in the weak noise limit
ΔΦ /σ 2  1, we have
1 −ΔΦ /σ 2
fP (σ ) ≈ 1, fC (σ ) ≈ e . (4.37)
κτ0
The critical noise strength at which the largest Lyapunov exponent vanishes in maps
thus satisfies, according to (4.28), the following relation:13
| λ1P |
e−ΔΦ /σc = κτ0
2
. (4.38)
λ1C
For σ close to σc we obtain
 −2 −2

λ1 (σ ) = λ1C fC (σ ) + λ1P =| λ1P | e−ΔΦ (σ −σc ) − 1 , (4.39)

which has the same σ -dependence as in (4.33) and leads again to the exponent
α = 1.
We note that the argument presented here applies to any physical quantity Q that
takes on values QP and QC on the original periodic attractor and on the nonattracting
chaotic set, respectively. The noise dependence of the average value Q(σ ) of Q
is then
Q(σ ) ≈ fP (σ )QP + fC (σ )QC . (4.40)
An application of this rule leads to a surprising result in a model of Brownian mo-
tion in a symmetric periodic potential in the presence of bias and periodic driving
[730, 731]. In particular, the effect of driving pushes the system out of thermal equi-
librium even in the presence of temperature fluctuations. For the noiseless system
at positive bias, there is a periodic attractor leading to a negative average velocity
vP < 0 of particles. At the same parameters, there is a coexisting chaotic saddle to
which a positive average velocity vC > 0 belongs. This, however, is not seen with
typical initial conditions in long-term observations. The presence of noise connects
the two original invariant sets and leads to noise-induced chaos. In the context of
transport, the main interest is, however, in the average velocity v(σ ), which follows
from (4.40) with QP = vP and QC = vC . It has the property that v(σ ) changes sign at
a critical value σcv , as has also been demonstrated in an experiment with a Joseph-
son Junction [540]. The behavior of the average velocity about the critical point is
linear: v(σ ) ∼ σ − σcv , analogous to the average Lyapunov exponent.
The scaling of the Lyapunov exponent with the noise strength has been verified
by a number of numerical examples [458], and also experimentally, using a nonlin-
ear electronic circuit [837]. Here we cite one example, the Rössler oscillator with
additive noise:

13 The noise strength σc corresponds to (4.21) with χ /Z = κτ0 | λ1P | /λ1C .

4.3 Noise-Induced Chaos 127

dx/dt = −y − z + σ ξx(t),
dy/dt = x + 0.2y + σ ξy(t),
dz/dt = 0.2 + z(x − c) + σ ξz(t), (4.41)

where c is the bifurcation parameter, σ is the noise strength, and ξx,y,z are indepen-
dent Gaussian random variables of zero mean and unit variance. A period-3 window
exists about c = 5.3. Figure 4.12a shows the projection of the period-3 attractor into
the (x, y)-plane for c = 5.3. Noise-induced chaos arises for σ > σc ≈ 10−2.26 ≈
5.5 × 10−3. For σ slightly above σc , the asymptotic trajectory of the system de-
viates from the original period-3 attractor, but the probability density is still large
around the 3-cycle, as shown in Fig. 4.12b for σ = 0.01. For a larger value of σ ,
the trajectory spends relatively more time in the region where the original chaotic
saddle resides, as shown in Fig. 4.12c. Figure 4.13a shows the first two Lyapunov
exponents of the asymptotic attractor versus the noise strength σ . We see that for
σ < σc , the largest Lyapunov exponent is zero, indicating that the noisy flow is
not chaotic. The presence of the null Lyapunov exponent means that in spite of
noise, there is a neutral direction associated with the flow. For σ > σc , the largest
Lyapunov exponent becomes positive, so the noisy flow is chaotic. In this case, there
is no longer a null Lyapunov exponent, indicating the disappearance of the neutral
direction. Figure 4.13b shows the scaling of the largest Lyapunov exponent of the
noisy chaotic attractor with σ − σc , which is apparently algebraic. A least-squares
fit between log10 λ1 and log10 (σ − σc ) gives the slope 0.94 ± 0.03. There is a rea-
sonable agreement between the theoretical scaling law and numerics.

Fig. 4.12 Asymptotic trajectory of the Rössler system for (a) σ = 0, (b) σ = 0.01 > σc , and (c)
σ = 0.02 [458] (Copyright 2003, the American Physical Society)
128 4 Noise and Transient Chaos

a
0.05

λ1,2
0

σc
−0.05
−3 −2.5 −2 −1.5
log10 σ
b −1
log10λ1

−2

−3

−4 −3.5 −3 −2.5 −2
log10 (σ−σc)

Fig. 4.13 For the Rössler system, (a) the first two Lyapunov exponents versus σ about the tran-
sition, and (b) algebraic scaling of the largest Lyapunov exponent with σ − σc . From (a), we see
that there is no zero Lyapunov exponent for σ > σc , indicating the lack of a neutral direction of
the flow [458] (Copyright 2003, the American Physical Society)

4.4 General Properties of Noise-Induced Chaos

4.4.1 Fractal Properties

A noise-induced chaotic attractor lies in the union of the periodic attractors and
the unstable manifold of the nonattracting set. Since the periodic attractors are
zero-dimensional objects on a Poincaré plane, the dimensions D0 and D1 of the
noise-induced chaotic attractor are the same as those of the unstable manifold of
the nonattracting chaotic set in the absence of noise: D0 = Du,0 and D1 = Du,1 . For
example, for a two-dimensional invertible map, we can use (2.82) and (2.78) to find
the information dimension of the noise-induced attractor as

λ1 − κ
D1 = 1 + . (4.42)
|λ2 |

Note that the dimension is independent of the noise strength σ , a valid property
in the weak-noise limit. In fact, the information dimension of the noise-induced
chaotic attractor is determined uniquely by the parameters of the chaotic saddle in
the underlying deterministic system.
It is the fractal property of the attractor which can be used as a condition to assess
whether noise is weak. As can be seen from Fig. 4.14, which shows the results of
the box-counting algorithm carried out for the example of Fig. 4.10, noise makes
the dynamics space-filling on small phase-space scales, less than εc ≈ e−4 = 0.018
4.4 General Properties of Noise-Induced Chaos 129

12
2

10

4 1.5

1 2 3 4 5 6

Fig. 4.14 Results of the box-counting algorithm for the unstable manifold of the deterministic map
investigated in Fig. 4.10a (dots) and for noisy attractors (b) (filled squares, σ = 0.01 (Fig. 4.10b)
and filled diamonds, σ = 0.03). The slopes of the thick solid lines represent the fractal dimension
Du,0 = 1.5 of the unstable manifold of the chaotic saddle in the deterministic system and of the
phase-space dimension d = 2. The threshold scale beyond which fractality holds is εc ≈ 0.018 for
σ = 0.01. For σ = 0.03, such a threshold value does not exist, indicating that noise dominates
the dynamics. A similar plot can be obtained for the scaling of the information dimension of the
unstable manifold as determined by (4.42), with Du,1 = 1.4 [717] (Copyright 2010, the American
Physical Society)

for σ = 0.01. For weak noise, there is always a scaling region, although short, with
the slope given by the noise-free fractal dimension, which is Du,0 = 1.5 in this case.
This is consistent with the schematic diagram Fig. 4.1. When this scaling region
disappears, it is no longer possible to identify the fractality of the noise-induced
chaos, even on larger phase-space scales, as is the case for σ = 0.03 [777]. In fact,
in this case, noise smears out the dynamics into large, finite bands of the phase
space, indicating that noise begins to dominate the dynamics. When this happens,
the noise can be considered strong.

4.4.2 Noise-Induced Unstable Dimension Variability

For deterministic flows, one of the Lyapunov exponents is zero because the dy-
namics along the flow is neutral, i.e., it is neither expanding nor contracting. An
interesting consequence of noise-induced chaos is that after the transition (σ > σc ),
the topology of the flow is disturbed in a fundamental way: there is no longer a zero
Lyapunov exponent, indicating that for noisy chaos, there exists no neutral direc-
tion along which infinitesimal distances are conserved.14 This is caused by a type of
nonhyperbolicity, unstable dimension variability, associated with the noise-induced
chaotic attractor.

14 This consideration does not apply to nonautonomous systems, for which there is always a neutral
direction along the time axis and therefore always a zero Lyapunov exponent.
130 4 Noise and Transient Chaos

The phenomenon of unstable dimension variability refers to the situation in

which a typical trajectory moves in phase-space regions containing unstable pe-
riodic orbits with distinct numbers of unstable eigendirections, which was first
conceived by Abraham and Smale [1], who constructed a mathematical model of
two unstable fixed points with distinct numbers of local unstable eigendirections.
It was realized later that the phenomenon is quite common in dynamical systems
of high dimensions, and it can have intricate consequences on the shadowability of
numerical trajectories in high-dimensional chaotic systems [41, 159, 416, 449, 676].
We have seen in Chap. 3 that an invariant set is nonhyperbolic if there are tangen-
cies between the stable and unstable manifolds. Unstable dimension variability is
another source of nonhyperbolicity.
To see how unstable dimension variability arises in the context of noise-induced
chaos, we note that for σ above σc , both the chaotic saddle and the periodic at-
tractor belong to a single, connected dynamical invariant set. Since periodic orbits
embedded in the chaotic saddle are all unstable and the originally attracting periodic
orbit is stable, a trajectory moves in regions containing periodic orbits with distinct
unstable dimensions. A unique feature of this type of unstable dimension variability
is that the subsets with different unstable dimensions are located in distinct regions
of the phase space, versus high-dimensional chaotic attractors, such as the kicked
double rotor [41, 159, 416, 449, 676], where unstable periodic orbits in these subsets
tend to mix with each other in the phase space [158, 452].
We can argue that after the onset of chaos, unstable dimension variability in-
duced by noise will destroy the neutral direction of the flow. For a three-dimensional
flow, the original periodic attractor contains no unstable direction, and the chaotic
saddle possesses one unstable direction. The role of noise, when it is sufficiently
large (σ > σc ), is to link these two originally dynamical invariant sets with distinct
unstable dimensions. Now examine the local eigenplanes that contain the neutral
direction of the flow associated with the periodic attractor and with the chaotic sad-
dle, as shown schematically in Fig. 4.15. In the local eigenplane about the periodic
attractor, there are a stable direction and a neutral direction (n). Let v be the eigen-
vector in the neutral direction. In the eigenplane of a point in the chaotic saddle,
there are an unstable direction and a neutral direction (n). When a trajectory is driven
by noise from the periodic attractor to the chaotic saddle, the eigenvector v maps to
v (see Fig. 4.15), which can point in any direction in the local eigenplane of the
corresponding point in the chaotic saddle. After a time, the vector will be aligned in
the unstable direction, due to the expanding dynamics associated with the chaotic
saddle. Distances along the neutral direction of the original periodic attractor can no
longer be preserved. Thus, we see that unstable dimension variability plays a fun-
damental role in shaping the topology of the flow after the onset of noise-induced
chaos. The chaotic attractor is, however, fundamentally different in its flow topol-
ogy from any deterministic chaotic attractors in that it no longer contains a neutral
direction. Readers should keep in mind that this topological disturbance of the flow
exists only for σ > σc .
4.4 General Properties of Noise-Induced Chaos 131

Chaotic
v
v’ saddle
n
Periodic
attractor
s

Fig. 4.15 Schematic illustration of the destruction of the neutral direction of the noisy chaotic
flow due to unstable dimension variability. The local planes about the periodic attractor and a point
in the chaotic saddle do not coincide in general, and these planes are not in a Poincaré surface
of section (Copyright 2003, the American Physical Society)

4.4.3 Ubiquitous Applications to Biological Sciences

The concept of noise-induced chaos can play an important role in the dynamical
evolution of biological systems, since random environmental influences are always
present [177, 221, 683]. Examples for which noise-induced chaotic attractors and
hence the results presented here are relevant are the following:
(1) Epidemiology. The controversy between the unpredictability observed in records
of chickenpox data and the nonchaotic nature of the attractor from the mathematical
models for realistic parameter values was first resolved by Rand and Wilson [621],
who pointed out that weak intrinsic or external noise can convert a chaotic saddle
of the model into a noisy chaotic attractor. Noise-induced chaos has proven to be a
ubiquitous source of unpredictability in epidemics since then [64, 65, 222, 696].
(2) Physiology. It has been suggested that pathological destruction of chaotic be-
havior may induce some types of brain seizures [684] and heart failures [266]. In
vital physiological systems chaotic dynamics can in fact be considered “normal”
[245]. Bifurcations to periodic behavior are viewed as a pathophysiological loss of
the range of adaptive possibilities [840]. In these situations the presence of noise
can be advantageous, since it can help induce or restore chaos.
(3) Ecology. Population-dynamical models sometimes also predict regular behavior
although observations find irregular dynamics. Here we present the model in [221]
to describe the population dynamics of Fennoscandian voles. The time-continuous
equations of motions for the scaled prey (vole) density, n, and predator (weasel)
density, p, are
132 4 Noise and Transient Chaos

Fig. 4.16 A chaotic saddle from the ecological model described by (4.43) and (4.44) for g = 0.12.
It is obtained by evolving N0 = 5 × 105 points uniformly distributed on the rectangle Γ : 0.001 <
n < 1.3 and 0 < p < 0.3. The lifetime of the saddle is τ = 56 years. Trajectories not entering
a circle of size 0.0005 around any of the attractor points (shown by black dots) up to n0 = 100
years are kept, and their points taken at year n = 25 provide a good approximation to the saddle.
The inset shows a magnification of part of the saddle, which exhibits double fractal features [717]
(Copyright 2010, the American Physical Society)

dn gn2 8np
= 4.5n[1 − sin(2π t) − n] − 2 − , (4.43)
dt n + 0.01 n + 0.04
dp  p
= 1.25p 1 − sin(2π t) − , (4.44)
dt n
where the parameters are taken from [791]. The seasonal variation has period
t = 1 year. A stroboscopic section is taken with a sampling of once per year
(at t = 1, 2, . . . ), generating an invertible two-dimensional map. The attractor of the
deterministic problem for g = 0.12 is a 13-cycle [221]. Figure 4.16 demonstrates
a chaotic saddle coexisting with the 13-cycle [777]. It is the chaotic saddle that is
responsible for the appearance of noise-induced chaos described earlier [221].

4.5 Noise-Induced Crisis

In the case of a boundary crisis, in the immediate precritical regime, a chaotic at-
tractor is close to the basin boundary (Sect. 3.1). A small amount of noise can drive
a trajectory out of the basin of attraction and cause it to go to another attractor, ef-
fectively inducing the crisis. For an interior crisis, sudden excursions to a certain
region of the phase space can occur when a system parameter passes through a criti-
cal value, say p > pc , generating crisis-induced intermittency. The average duration
of these excursions is given by (3.15). In the presence of noise, random fluctuations
can cause a similar behavior even in the precrisis region p < pc [24, 757]. What can
happen in this case is that noise can dynamically connect the chaotic attractor with
4.5 Noise-Induced Crisis 133

a coexisting, nonattracting chaotic set before the deterministic-crisis value, causing

a sudden increase in the size of the chaotic attractor. Noise-induced crisis can there-
fore be regarded as a type of noise-induced chaos. Because of this analogy, the
critical noise strength σc above which intermittency can be observed with threshold
resolution χ is given by (4.21), where ΔΦ is the activation energy in the basin of
the small-size attractor. An example of noise-induced crisis is shown in Fig. 4.17,
where a two-piece deterministic chaotic attractor expands into a one-piece attractor
as a result of noise.
A basic quantity of interest for noise-induced crisis is the characteristic time τ
that a trajectory spends on the original deterministic chaotic attractor. This time is
the average time interval between bursts into the newly accessible region of the
phase space. The issue is how the average time τ depends on both the param-
eter difference pc − p and the noise strength σ . Based on the observation that
in the deterministic case (3.15) holds, Sommerer et al. [723, 728] obtained, for
low-dimensional chaotic systems, the following universal scaling law for noise-
induced crisis:

pc − p
τ ∼ σ −γ g , (4.45)
σ

where γ is the critical exponent of the corresponding deterministic crisis, and g is

a function that depends on the system details and on the distribution of the noise.
Note that in contrast to crisis-induced intermittency in the deterministic case, the
parameter range needs not be restricted to p > pc here, as a result of noise.
The particular form of g follows from the exit rate estimates obtained from the
properties of the quasipotential. The basic argument in Sect. 4.3.2 can be applied,
with some modifications. In particular, away from saddle-node bifurcations, the
quasipotential can be assumed to increase quadratically from the attractor, with a
parameter-independent coefficient given by ΔΦ (Δx) ∼ Δx2 . The distance between

a 1.5 b 1.5
x x
1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
−1.5 −1 −0.5 0 0.5 1 y −1.5 −1 −0.5 0 0.5 1 y

Fig. 4.17 Noise-induced crisis and attractor expansion in the Hénon map: xn+1 = 1−ax2n +0.3yn +
(x) (y)
σ ξn and yn+1 = xn + σ ξn for a = 1.13 with a uniform noise. (a) Deterministic attractor (σ = 0),
and (b) attractor in the presence of noise of strength σ = 0.009. The critical noise strength is
σc ≈ 0.008. In the absence of noise, a similar attractor would occur for a = 1.16
134 4 Noise and Transient Chaos

the attractor and the quasipotential plateau then depends linearly on the parameter
difference: Δx ∼ pc − p. We thus obtain ΔΦ ∼ (pc − p)2 . For typical maps with
parabolic tangencies and for Gaussian noise, the average lifetime τ close to the cri-
sis point scales with (pc − p) as
2
/σ 2
τ ∼ ec(pc −p) . (4.46)

Combining this with the scaling law (4.45), we obtain [723] that15
 
−γ c(pc − p)2
τ ∼σ exp . (4.47)
σ2

The scaling function is again of the form of (4.26) as for noise-induced chaos, in-
dicating the similarity between these two phenomena. A particular consequence of
the scaling form (4.45) is that, at the deterministic crisis value p = pc , the average
characteristic time follows a simple algebraic scaling law:

τ ∼ σ −γ . (4.48)

That is, the lifetime on the small-size attractor increases with the noise strength
with the same exponent γ at pc as with the parameter difference for p > pc in the
noise-free problem.
Sommerer et al. considered several examples, ranging from map to flow sys-
tems [728], to verify the scaling laws (4.45) and (4.48). In addition, experimental
evidence was obtained by examining the oscillatory dynamics of a magnetoelastic
ribbon in a time-varying magnetic field [723, 724], where the control parameter p is
the period of the applied magnetic field of amplitude Hac . Controllable noisy fluctu-
ations are introduced by a random magnetic field of strength σ . Deterministically, an
attractor-merging crisis takes place. By plotting ln [τσ γ ] versus [(pc − p)/σ ]2 , data
from different measurements fall on a single straight line, as shown in Fig. 4.18.

4.6 Random Maps and Transient Phenomena

Without noise, a chaotic attractor typically exhibits a fractal structure caused by

the underlying dynamics. Under the influence of small random perturbations, if one
examines a long trajectory produced by the dynamics, one usually observes that
the fractal structure is smeared up to a distance scale proportional to the strength
of the perturbations. In order to observe a clear fractal structure, a remedy is to
examine the snapshot pattern formed by an ensemble of trajectories, subject to the
same random perturbation, as pointed out by Romeiras, Grebogi, and Ott [648]. The

 
c(pc −p)r
15 For noise of order r, we have ΔΦ (Δx) ∼ Δxr and τ ∼ σ −γ exp σr .
4.6 Random Maps and Transient Phenomena 135

Fig. 4.18 For noise-induced crisis in an experimental magnetoelastic system, scaling of the char-
acteristic time, presented according to the scaling law (4.45), where the value of the critical
exponent γ is taken from the deterministic case. Data collapsing onto a straight line confirms the
exponential form of the general scaling (4.47) [723] (with kind permission from Elsevier Science)

details of the fractal structure differ from time to time, but the fractal dimensions
remain invariant [471, 648]. The idea of snapshot attractors [648] has proven to
be useful in the laboratory for visualizing and characterizing fractal patterns arising
in physical situations such as passive particles advected on the surface of a fluid
[726, 844]. Snapshot attractors have been utilized to study the transition to chaos in
quasiperiodically driven dynamical systems as well [446].
A convenient setting for studying the onset of fractal snapshot attractors and vi-
sualizing them is random maps. Here, the mapping rule is not autonomous. Instead,
it depends on the actual instant of time, most often via the time-dependence of a
control parameter, denoted by pn . It is the temporal fluctuations of the parameter
which can be considered to be random. A random map is then defined through the
mapping rule
xn+1 = f(xn , pn ), (4.49)
where f(x, p) is a known function. The actual value pn of the parameter can be
written as
pn = p̄ + δ pn , (4.50)
where the fluctuating part δ pn is taken randomly on each iterate from a stationary
distribution P(δ pn ) of zero mean. Unless otherwise stated, we shall not specify the
form of the distribution. Accordingly, p̄ is the mean value of the parameters over
many realizations. Note that in random map (4.49), there is no additive noise ξ , in
contrast to a noisy map (4.2). Another difference in comparison with the map (4.2)
is that there, the noise realization can be different for different trajectories even
136 4 Noise and Transient Chaos

at the same time. The particular form of (4.49) implies that all the points iterated
by the map are subject to the same noise at a given time. A single trajectory is as
fuzzy as in a noisy map. An ensemble of trajectories, however, behaves in a much
more coherent way, and can exhibit fractal properties. In fact, dynamical properties
averaged over realizations of noise are well defined.
The concepts of random attractors and pullback attractors [29, 261] that have
appeared in recent publications on climate dynamics [122] are practically the same
as that of snapshot attractors. A slight difference is that in the former cases the
random process is typically white noise, and the deterministic dynamics might be
regular. We use and extend the terminology of snapshot attractors in what follows.

4.6.1 Open Random Maps, Snapshot Chaotic Saddles

In the context of transient chaos, the important feature is that nonattracting chaotic
sets can occur in random maps and they also possess fractal patterns. This is the
case if escape is possible from some finite phase-space region, i.e., if the random
map is open.
Similar to the case of sustained chaos in random systems, there cannot be either
periodic orbits or invariant sets in open random maps [352]. The analogues of the
stable and unstable foliations and of the nonattracting chaotic set can nevertheless
be defined. The latter can be called a snapshot chaotic repeller or a snapshot chaotic
saddle. In the case of invertible deterministic dynamics, for the unstable manifold
in a two-dimensional open random map belonging to the time instant n, we can take
an ensemble of initial points at some earlier time ni and let them evolve according
to the random map (4.49). The end points (at time n) of trajectories that do not
leave a restraining region Γ up to time n form, for n − ni  1, a fractal set. In the
spirit of the sprinkler method (Sect. 1.2.2.3), the set of end points can be regarded
as an unstable foliation (unstable manifold) lying within region Γ . In general, the
unstable foliations at time n and n = n are different. Similarly, the stable foliation
(stable manifold) at time instant n can be obtained by letting an ensemble of points
move at time n and keeping all starting points of trajectories that do not leave Γ
up to nf − n  1 steps. The snapshot chaotic saddle can be obtained by plotting the
points at time n of all the trajectories started at some ni that do not escape Γ up to
a final instant nf , provided n − ni and nf − n  1. Examples are given in Figs. 4.19
and 4.20. We see that none of these sets is invariant. In fact, their shapes change
with time due to the random parameter change with time. A comparison between
Figs. 4.20 and 4.21 indicates that the change can be considerable even over a single
time step.
A consequence of a fractal stable foliation is that the lifetimes depend sensitively
on the initial conditions, as shown in Fig. 4.22. The lifetimes become infinite for
the intersecting points between the line of initial conditions and the stable foliation.
The number of trajectories in a region covering the chaotic saddle decays exponen-
tially, as just like in the deterministic case. This defines an escape rate κr for the
4.6 Random Maps and Transient Phenomena 137

Fig. 4.19 Fractal foliations for the random Hénon map defined by xn+1 = 1 − (2 + δ an )x2n + 0.3yn
and yn+1 = xn , where δ an is a random variable chosen from a uniform distribution in (−1, 1)
and represents parameter fluctuations. Shown are a stable foliation (a), a snapshot chaotic saddle
(b), and an unstable foliation (c). These sets are calculated using an ensemble of N = 107 initial
points uniformly distributed in the square region. Panels (a), (b) and (c) correspond to the initial
points, the midpoints (n = 8), and the end points (n = 16), respectively, of trajectories that stay
in the square for at least 16 iterations. The gray lines in panel (b) are copies of the foliations in
(a) and (c) and indicate that the chaotic set is not an intersection of these foliations, since the
stable (unstable) foliation belongs to time n = 0 (n = 16), while the saddle belongs to n = 8. The
corresponding construction of the invariant sets for the deterministic map can be seen in Fig. 1.9

1.5
x a b c
1

0.5

−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 y −1 −0.5 0 0.5 1 y −1 −0.5 0 0.5 1 y 1.5

Fig. 4.20 For the same map as in Fig. 4.19, (a) initial points of trajectories starting at time n = 16
and not escaping the square up to 32 steps, (b) midpoints (n = 16) of trajectories starting at n = 8
and not escaping up to n = 24, and (c) end points (n = 16) of trajectories starting at n = 0 and
not escaping up to n = 16. The gray lines in panel (b) are copies of the foliations in (a) and (c)
and indicate that the snapshot chaotic saddle at time step n = 16 is the set of intersections of the
foliations belonging to the same time instant

random map. Being an asymptotic property, the escape rate does not depend on the
instant that the trajectories are initiated. This is so because over a long period of
time, practically all possible parameter values will have been realized. This means
that since the parameter distribution is stationary, the escape rate becomes indepen-
dent of the initial time when the ensemble of trajectories starts to evolve. Similarly,
the average Lyapunov exponents λi,r of the random map can be determined, and
they are independent of time as well.
138 4 Noise and Transient Chaos

1.5
x a b c
1

0.5

−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 y −1 −0.5 0 0.5 1 y −1 −0.5 0 0.5 1 y 1.5

Fig. 4.21 (a–c) Similar to Fig. 4.20, (a–c), respectively except that all relevant times are shifted
by −1

Fig. 4.22 Survival in the random Hénon map. (a) Number of survivors in the square | xn |, | yn |≤ 1.
The escape rate is determined to be κr ≈ 0.35, which is less than that for the deterministic case
(κ ≈ 0.36). (b) Dependence of the lifetime n on the initial position x (taken at ni = 0), where the
initial conditions are chosen from the interval | x |≤ 1 at y0 = −1.5. The fractal nature of this curve
is as pronounced as for the deterministic case shown in Figs. 1.4 and 1.5

For snapshot chaotic saddles, the basic relations among the information dimen-
sion, the Lyapunov exponents, and the entropies (treated in Sect. 2.6.2) are formally
the same as for deterministic maps [471]. For the random baker map, these relations
are derived in Appendix B. Here we quote the formula that can be used to determine
the partial information dimension along the unstable foliation:

(1) κr
D1,r = 1 − . (4.51)
λ1,r

The fact that this does not depend on the time instant implies that the dimensions
of the chaotic set and its manifolds are independent of time, in spite of the change
in their shape with time in the phase space. Similar conclusions can be drawn for
other formulas. It should be emphasized that the particular values of the escape
rate κr , the average Lyapunov exponent λ1,r , and the partial information dimension
(1) (1)
D1,r are not the same as the respective quantities κ , λ1 , and D1 characterizing the
4.6 Random Maps and Transient Phenomena 139

corresponding deterministic map. Nonetheless, the relationships among them are

the same. Analogously, the formulas for high-dimensional systems (to be discussed
in Chap. 8) are expected to remain valid for open random maps. These formulas
remain valid in the limit κ → 0 of closed maps, and apply then to attractors of
random maps, or snapshot attractors, as well. In the presence of strong noise, noise-
induced chaotic attractors (Sect. 4.3) are fuzzy but can successfully be decomposed
into fractal snapshot attractors [81].
Open area-preserving random maps generating transient chaos arise naturally in
advective fluid dynamics [352, 544]. Whenever the underlying flow is not exactly
periodic (e.g., temporally chaotic), the velocity field can be written as an average
periodic field plus a fluctuating component. As a result, the particle positions ob-
tained by integrating the equations of motion can also be decomposed in a similar
way. Effectively all particles are then subject to the same parameter fluctuations at
all times. If the observational time is longer than some typical correlation time of
the velocity field, the random-map approach becomes applicable to advection prob-
lems, as shall be discussed in Chap. 10 (chaotic advection). There are other physical
situations in which random maps are relevant. For example, all scattering processes
in force fields that depend on time in a complicated manner can be modeled by
random maps. Due to the lack of invariant sets in random maps, KAM tori do not
exist, implying that practically no long-lasting transport barriers can be present. (See
Chap. 6).

4.6.2 Transient Behavior in Fractal Snapshot Attractors

A necessary condition for snapshot attractors to exhibit fractal structures is that at

any instant of time, the influence of the random perturbation on every trajectory in
the ensemble must be identical. The reason is that a difference in the random per-
turbation to different trajectories in the ensemble can be regarded as a phase-space
diffusion that can smear out the fractal structure even in snapshot attractors. To see
this, consider a low-dimensional driven chaotic system described by the following
map:
xn+1 = f(xn , yn ), (4.52)
where x is a state vector and y represents a random or chaotic driving. Now imagine
that we follow the pattern in variable x with time. In order to observe a pronounced
fractal structure, the random perturbations yn must not depend on the phase-space
variable xn . If they do, the fractal structure of the snapshot attractor will be smeared
approximately by an amount proportional to the magnitude of the perturbation. Due
to chaos in the driving system, the fractal pattern will be less and less visible because
the amount of “fuzziness” in the phase space is magnified exponentially with time.
The extent to which fractal snapshot attractors can be observed in systems driven
by random or chaotic perturbations that depend only weakly on the dynamical vari-
ables was addressed in [437]. Such weak dependence can be called phase-space
140 4 Noise and Transient Chaos

inhomogeneity. For the system described by (4.52), it can be assumed conveniently

that the chaotic driving signal comes from the following process:

yn+1 = g(yn , ε xn ), (4.53)

where g is a chaotic map and ε > 0 represents the amount of weak phase-space
inhomogeneity. Note that when ε = 0, the x-dynamics does not influence the
y-dynamics, and hence (4.52) and (4.53) represent a unidirectionally coupled (from
y to x) system. Suppose we choose an ensemble of initial conditions x0 and evolve
them according to (4.52) and (4.53). If ε > 0, the phase-space inhomogeneity will
be amplified exponentially due to the chaotic nature of the driving (4.53), and finite-
scale fractal structures can be seen in the snapshot attractors for only a transient
period of time.
The scaling of the average time τ to observe a snapshot attractor with ε can be
obtained by examining the sub-Lyapunov exponents [581] defined with respect to
(4.52) and (4.53). Let λix (i = 1, . . . , Nx ) and λ jy ( j = 1, . . . , Ny ) be the Lyapunov
exponents of the two respective subsystems. They are

1 N ∂f 
λix = ∑ ln |  · ui |, i = 1, . . . , Nx , (4.54)
N n=1 ∂ x (xn ,yn )

1 N ∂g
λ jy = ∑ ln |  · v j |, j = 1, . . . , Ny ,
N n=1 ∂ y (xn ,yn )

where N  1, and ∂ f/∂ x|(xn,yn ) and ∂ g/∂ y|(xn ,yn ) are the derivative matrices of
(4.52) and (4.53) evaluated along a coupled trajectory. Vectors ui (i = 1, . . . , Nx )
and v j ( j = 1, . . . , Ny ) are unit vectors in the ith eigendirection in the tangent space
of (4.52) and in the jth eigendirection in the tangent space of (4.53), respectively.
x(y) x(y)
The Lyapunov exponents are ordered as λ1 > · · · > 0 > · · · λNx (Ny ) .
Suppose we choose a cloud of initial conditions uniformly distributed in a phase-
space region covering the attractor of the subsystem x. Let δ be the smallest distance
scale to resolve the fractal structure in an observation. The time Tx for the ensemble
to converge to a fractal set of resolution δ can be estimated from exp(−|λNxx |Tx ) ∼ δ
to be Tx ∼ − ln δ /|λNxx |. In order to observe the fractal structure, the amount of phase-
space inhomogeneity ε must be smaller than δ . The time Ty to reach the distance
y
scale δ in the x subsystem due to the y-dynamics satisfies ε eλ1 Ty ∼ δ . We obtain
Ty ∼ (ln δ − ln ε )/λ1y . Thus, the time window for a fractal snapshot attractor to be
observed is  
− ln ε 1 1
τ = Ty − Tx ∼ + ln δ + ,
λ1y λ1y |λNxx |
which gives the following scaling relation:

1
τ ≈ α ln + β , (4.55)
ε
4.6 Random Maps and Transient Phenomena 141

where α is approximately the inverse of the largest Lyapunov exponent of (4.53),

and β is a constant. Demanding τ > 0, we obtain from (4.55) the maximum value
of the phase-space inhomogeneity for a fractal snapshot attractor of resolution δ to
be observed: ε < δ eD , where D ≡ 1 + λ1y /|λNxx |. Thus, in order to observe fractal
snapshot attractors for a long time, the phase-space inhomogeneity must be small
and/or the driving system be only weakly chaotic.
Numerical confirmation of the scaling relation (4.55) has been obtained [437] by
considering a random version of the Ikeda map (3.13) given by

ip
zn+1 = A + Bzn exp ik − + i2πθn , (4.56)
1+ | zn |2

where θn represents a chaotic driving, and A, B, k, p are parameters. The effect

of phase-space inhomogeneity in the driving can be modeled by assuming that θ
comes from the noisy logistic map:

θn+1 = 3.75θn (1 − θn) + ε xn . (4.57)

Under the parameter setting (A, B, k, p) = (0.85, 0.9, 0.4, 5.18), the Ikeda map, in
the absence of perturbation θn , exhibits a chaotic attractor with a fractal structure
(cf. Fig. 3.14b).
Figure 4.23a–j show, for ε = 10−16, snapshot attractors at ten different time
instants from a grid of 128 × 128 initial conditions uniformly distributed in the
region: (−2.0 ≤ x ≤ 4.0, −2.5 ≤ y ≤ 2.5). The attractors are apparently fractal for
20 < n < 80, beyond which time the fractal structure is smeared out. The apparently
nonfractal behavior at very short time, e.g., at n = 10, is due to the fact that it takes
a finite amount of time for the ensemble of trajectories to settle down to the chaotic
attractor. As ε is increased, the time interval for fractal snapshot attractors to be ob-
served decreases. To measure the average transient time interval τ in which snapshot
attractors are apparently fractal, the following box-counting procedure can be used.
The phase-space region from which the initial conditions are chosen is first divided
into a grid of boxes. At each instant of time n, the number of nonempty boxes Nn can
be counted. For small time, since the trajectories have not come close to the chaotic
attractor, we expect to observe a large number of occupied boxes. As the trajecto-
ries begin to settle down in the vicinity of the chaotic attractor, Nn starts to decrease,
reaches a small value, and remains approximately at this value when the snapshot
attractors are apparently fractal. When the effect of phase-space inhomogeneity in
the driving begins to take over so that the fractal structure becomes smeared, we
expect the number of nonempty boxes to increase. The time interval in which Nn
remains approximately constant is taken to be the average time τ . In the τ versus
ln ε plot, one observes the scaling relation (4.55) [437].
The phenomenon of transient fractal snapshot attractors has some implications
to the study of fractal geometry in high-dimensional chaotic systems, i.e., systems
with more than one positive Lyapunov exponent (to be treated in Chap. 8). Consider
the following general class of systems,

xn+1 = f(xn , εx yn ), yn+1 = g(yn , εy xn ), (4.58)

142 4 Noise and Transient Chaos

Fig. 4.23 Random Ikeda map. For (4.56) with phase-space inhomogeneity ε = 10−16 , snap-
shot attractors at times n = 10, 20, . . ., 100; (a)–(j) [437] (Copyright 1999, the American Physical
Society)

where both f and g are chaotic maps, and εx and εy are respectively two parame-
ters characterizing the coupling from x to y and vice versa. The system setting of
(4.58) arises naturally in the context of coupled chaotic oscillators. The maps f and
g can be noninvertible. In order to study the fractal geometry of system (4.58), we
assume that when the two maps are uncoupled, i.e., when εx = εy = 0, both maps
f(x) and g(y) exhibit a chaotic attractor with one positive Lyapunov exponent and
that the attractors have fractal structures in their own phase spaces x and y. When
couplings are present, the coupling terms εx y and εy x can be regarded as two driving
terms to the x and y dynamics, respectively. Since y and x are chaotic variables, the
problem becomes effectively that of studying fractals of randomly driven chaotic
systems. Intuitively we expect snapshot attractors in the x or y space to reveal the
fractal structures in the absence of couplings. Nonetheless, due to coupling, the
influence of driving is not homogeneous in both the x and y subspaces. The phase-
space inhomogeneity of the chaotic driving thus becomes a potential obstacle for
observing low-dimensional fractal structures in high-dimensional chaotic systems.
The fact that fractal snapshot attractors have been observed in laboratory experi-
ments such as passive particles advected on the surface of fluids [726] indicates that
4.6 Random Maps and Transient Phenomena 143

the experimental condition may be such that the amount of phase-space inhomo-
geneity on the fluid surface is near zero (the coupling between the dynamics in the
direction orthogonal to the fluid surface and the dynamics of the passive scalar on
the surface of the fluid is nearly unidirectional) or the dynamics of the driving is
only weakly chaotic with a near-zero positive largest Lyapunov exponent.
 Part II
Physical Manifestations of Transient Chaos
Chapter 5
Fractal Basin Boundaries

Dissipative dynamical systems often possess multiple coexisting attractors. The set
of initial conditions leading to trajectories landing on an attractor is the basin of
attraction of this attractor. Each attractor thus has its own basin, which is invariant
under the dynamics, since images of every point in the basin still belong to the same
basin. The basins of attraction are separated by boundaries. We shall demonstrate
that it is common for nonlinear systems to have fractal basin boundaries, the dy-
namical reason for which is nothing but transient chaos on the boundaries. In fact,
fractal basin boundaries contain one or several nonattracting chaotic sets.
We will describe the basic dynamical properties of basin boundaries and intro-
duce the main types of fractal basin boundaries. In general, a basin boundary can
be characterized by its box-counting dimension and the predictability of the final
state; the latter is quantified by the uncertainty exponent. The issue of how frac-
tal basin boundaries can arise as a system parameter changes will be discussed.
Topics such as Wada basin boundaries (common fractal basin boundaries among at
least three basins of attraction) and sporadically fractal basin boundaries (boundaries
consisting of smooth curves or surfaces and nondifferentiable components) will be
addressed. Attention will also be paid to riddled basins in symmetrical dynami-
cal systems, an extreme type of basin structure that practically defies predictability
of the final state. The consequences of symmetry-breaking perturbations will be
discussed.
A primary goal of science is to make predictions based on a set of physical laws.
A question of natural concern, due to the inevitable error in the specification of the
initial condition, is whether the final state of a trajectory can be predicted from an
initial condition chosen in the vicinity of a basin boundary. The various situations
to be discussed in this chapter illustrate that the prediction of the final state can be
extremely difficult and sometimes practically impossible even for relatively simple
deterministic systems.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 147
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 5,

c Springer Science+Business Media, LLC 2011
148 5 Fractal Basin Boundaries

5.1 Basin Boundaries: Basics

To gain intuition, we consider the following simple physical system in which a

particle of unit mass moves under conservative force determined by a one-
dimensional potential function V (x). In the two-dimensional phase space (x, v ≡
dx/dt), the equations of motion are

dx dv dV
= v, = −γ v − , (5.1)
dt dt dx
where we assume that there is a frictional force proportional to the velocity of the
particle and γ > 0 is a dissipation parameter. To create multiple coexisting attractors,
consider the class of symmetric double-well potentials, as schematically illustrated
in Fig. 5.1a. The two potential wells are located at x± = ±a, and there is a potential
barrier at xb = 0. For a particle in the vicinity of a well, if the initial velocity is small
such that the initial energy of the particle is not large enough for it to overcome the
potential barrier, it will approach asymptotically the bottom of the well, due to the
friction. Each well is thus an attractor and there are two attractors in the phase space,
located at (x, v) = (±a, 0). To understand the structure of the basins of attraction and
the basin boundary, we notice that if a still particle sits precisely on the top of the

a V(x)

−a a x

b v = dx/dt

O x

Fig. 5.1 (a) Double-well

v
potential V (x), (b) c
Hyperbolic point O = (0, 0)
and a linear segment of its
stable manifold belonging to
the boundary between the
x
basins of attraction of the two
attractors located at
(x, v) = (±a, 0). (c)
Schematic illustration of the
basin boundary and the two
basins of attraction
5.2 Types of Fractal Basin Boundaries 149

barrier, it will remain there forever, although an arbitrarily small perturbation, either
in its position or velocity, or a combination of the two, can move the particle to
one of the wells. Thus the point O ≡ (0, 0) is an unstable fixed point on the basin
boundary.
Now consider a particle initially located in the right well but near O. If it is
given a small initial velocity in the −x direction, it may or may not overcome the
potential barrier. There then exists a velocity for which the particle stops precisely
at O. Likewise, for a particle initially in the left well but near O, there exists a
small initial velocity in the +x direction that lands the particle precisely at O. In the
two-dimensional phase space, there then exists a set of initial conditions lying on
a one-dimensional curve that is approximately linear1 near O, which approaches
O asymptotically, as shown schematically in Fig. 5.1b. In the terms of dynamical
systems, the one-dimensional curve is the stable manifold of the saddle point O.
Since points on the curve do not approach any of the two attractors, it is the basin
boundary. The reasoning thus suggests that in situations in which multiple attractors
coexist in invertible systems, the basin boundary is the stable manifold of some
unstable invariant set on the boundary.
While the basin boundary is approximately linear near the unstable fixed point O,
it curves away from O, due to nonlinearity. For the simple example in Fig. 5.1, the
boundary crosses the x-axis an infinite number of times. This can be seen by noticing
that away from an attractor, say from the one on the right well in the +x direction, the
force becomes attractive. There then exists a set of x̄ values, where x̄ > a, for which
the amount of force is just right to place an initially still particle right at the top of
the potential barrier. The points (x̄, 0) are thus on the basin boundary. This leads to
basins consisting of strips near the x-axis for |x| large, as shown schematically in
Fig. 5.1c. A basin of attraction typically possesses an infinite phase-space volume.2
The simple mechanical example in Fig. 5.1 illustrates that when the invariant set
on the basin boundary is simple, e.g., an unstable periodic orbit, the boundary is
smooth. One can imagine the situation that there is a nonattracting chaotic set on
the basin boundary. Since the stable manifold of the chaotic set is a fractal set, the
boundary becomes fractal.

5.2 Types of Fractal Basin Boundaries

In typical dynamical systems, i.e., systems whose behaviors are not due to any
special properties such as symmetry, there are at least three known types of frac-
tal basin boundaries, [294] described in the subsequent subsections.

1 Near O, we have V ≈ −s2 x2 /2. The solution to (5.1) is x(t) = c+ eλ+ t + c− eλ− t with λ± = −γ /2±
(s2 + γ 2 /4) . Thus, for c+ = 0, we have v(t) = λ− x(t) ∼ eλ− t and x(t) → 0, v(t) → 0 as t → ∞,
1/2

along the line v = λ− x

2 Choose a phase-space region R of nonzero volume that encloses an attractor. That the system

is dissipative means that the inverse dynamics is volume-expanding. Since R is completely in the
basin of attraction, all its preimages are in the basin as well. In the limit t → −∞, the volume of
the preimage becomes infinite.
150 5 Fractal Basin Boundaries

Fig. 5.2 Basins of attraction for the forced damped pendulum (5.2) on the stroboscopic surface of
section (θ , θ̇ ) defined by t = 2nπ (n = 0, 1, . . .). (a) For γ = 0.1, f 0 = 1.2, there is a fixed-point
attractor at θ = −2.2055 and θ̇ = 0.3729. Black regions denote initial conditions that go to this
attractor. There is another attractor in the blank region. (b) For f 0 = 2.0, the attractor in the black
basin of attraction is located at θ = −0.8058 and θ̇ = 0.9375 [296] (with kind permission from
Elsevier Science)

5.2.1 Filamentary Fractal Boundaries

Filamentary fractal boundaries are boundaries locally consisting of a Cantor set

of smooth curves or surfaces. This situation typically occurs in invertible dynam-
ical systems, where the asymptotic invariant sets on the boundary contain a chaotic
saddle [491, 794]. Examples are shown in Fig. 5.2a, b for the following periodically
forced damped pendulum [296]:

d2 θ dθ
2
+γ + sin θ = f0 cost, (5.2)
dt dt
where γ is a frictional coefficient and f0 is the forcing amplitude. Such boundaries
may also contain nonfractal parts, e.g., in regions about the attractors. In certain
systems, these two types of boundary behaviors are intertwined on arbitrarily fine
scales. For any area that contains a fractal part of the boundary, there is a subarea
that contains only smooth parts of the boundary. Such fractal boundaries are called
intertwined boundaries.

5.2.2 Continuous Fractal Boundaries

Continuous fractal boundaries are boundaries that are a continuous but nowhere
differentiable curve or surface. An illustrative example is the following class of
noninvertible two-dimensional maps: [511, 564]:
5.2 Types of Fractal Basin Boundaries 151

Fig. 5.3 For the

two-dimensional map (5.3),
the basin boundary between
the y = ±∞ attractors. The
parameters are a = 3 and
λ = 1.5. The boundary is
continuous but nowhere
differentiable, as represented
by a Weierstrass curve [511]
(with kind permission from
Elsevier Science)

xn+1 = axn mod (1), (5.3)

yn+1 = λ yn + cos(2π xn),

where a > λ > 1 and a is an integer. Since λ > 1, almost all initial conditions
lead to trajectories that go to y = ±∞, which can be regarded as two attractors.
The term cos (2π xn ) in the y-equation entails that the basin boundary near y = 0
can be complicated. Indeed, an explicit expression for the boundary curve can be
obtained [511, 564]:
∞
y = g(x) = − ∑ λ − j cos (2π a j−1x). (5.4)
j=1

A direct substitution into (5.3) shows that y = g(x) is an invariant curve, i.e., yn =
g(xn ) and yn+1 = g(xn+1 ). The curve thus contains an invariant set, a chaotic repeller.
The curve y = g(x) is continuous but nowhere differentiable because dy/dx diverges
for every value of x. The curve in (5.4) in fact has the box-counting dimension
D0 = 2 − (ln λ )/(ln a), and is called a Weierstrass curve. An example is shown
in Fig. 5.3.

5.2.3 Sporadically Fractal Boundaries

In a two-dimensional map, basin boundaries of the sporadically fractal type can

be described by a function g(x) that is smooth except for a set of x values of zero
measure (i.e., zero length), but nevertheless has a box-counting dimension larger
than 1. An illustrative example is [346, 651]

xn+1 = f (xn ), (5.5)

yn+1 = λ yn + sin (2π xn),
152 5 Fractal Basin Boundaries

Fig. 5.4 Example of a

sporadically fractal basin
boundary from the
two-dimensional map (5.5)
[346] (Copyright 1999, the
American Physical Society)

where λ > 1 and f (x) is a noninvertible one-dimensional map. Part of the basin
boundary is shown in Fig. 5.4. It can be seen that the boundary consists mostly of
smooth parts but with sporadic “spikes” along the curve. Sporadically fractal basin
boundaries can arise in the context of chaotic phase synchronization in continuous-
time dynamical systems.
If a dynamical system possesses a special property such as simple symmetry, the
topology of the basins of attraction can be quite different from those seen in typical
systems.

5.2.4 Riddled Basins

If the symmetry leads to an invariant subspace in the phase space, where there is a
chaotic attractor, the basin of attraction of this chaotic attractor can be riddled with
holes that belong to the basin of another attractor, provided that such an attractor
exists outside the invariant subspace [11]. A riddled basin thus contains no open
sets (e.g., areas in two dimensions or volumes in three dimensions), in contrast to
fractal basins. Physically, the presence of a riddled basin means that for every initial
condition that goes to the chaotic attractor in the invariant subspace, there are initial
conditions arbitrarily nearby that lead to trajectories to the other coexisting attrac-
tor. Prediction of the asymptotic attractor for a given initial condition thus becomes
practically impossible. An example of a riddled basin has been found experimen-
tally, as shown in Fig. 1.18. A numerically obtained riddled basin can be seen in
Fig. 5.5, in which a particle of unit mass moves in a planar potential given by

V (x) = (1 − x2)2 + (y2 − a2)2 (x − d) + b(y2 − a2)4 ,

5.3 Fractal Basin Boundaries and Predictability 153

Fig. 5.5 Riddled basins. (a) Solution of (5.6) with parameters f 0 = 2.3, γ = 0.05, ω = 3.5, a = 0.8,
b = 0.008, and d = −0.19. Black dots represent points in the basin of the attractor at y = a, vy = 0.
(b) A magnification of part of (a) near the invariant subspace at y = −a and vy = 0. One can see
that arbitrarily close to the attractor at y = −a there are points belonging to the attractor at y = a
[448] (Copyright 1995, the American Physical Society)

where a, b, and d are parameters. The particle is also subject to friction and periodic
forcing. The equation of motion is

d2 x dx
= −γ − ∇V (x) + f0 sin (ω t)ex , (5.6)
dt 2 dt

where x ≡ (x, y), γ is the frictional coefficient, and ex is the unit vector in x. There
are two invariant subspaces determined by y = ±a and vy = 0 in which the dynam-
ics are governed by the forced double-well problem (Duffing’s equation). For proper
choices of the parameters a, b, and d, the basins of the chaotic attractors in the sym-
metric invariant subspaces are both riddled, as shown in Fig. 5.5. When all basins
are riddled by the rest, as is the case here, the basins are said to be intermingled
[448]. Note, however, that riddled basins rely on the symmetry of the system. A
small amount of symmetry-breaking leads to a catastrophic bifurcation whereby a
riddled basin immediately becomes fractal with open areas.
A recent review by Aguirre et al. [7] on fractal basin boundaries gives a
comprehensive treatment of the topic and presents a large number of applications.
Our focus here will be on the interplay between fractal basin boundaries and
transient chaos.

5.3 Fractal Basin Boundaries and Predictability

The box-counting dimension Db0 can be used to characterize the boundary. Let D
be the dimension of the phase space. Since the boundary divides the phase space, we
have D − 1 ≤ Db0 ≤ D. A question of interest in a practical situation is, what are the
154 5 Fractal Basin Boundaries

physical meaning and consequences of having fractal basin boundaries of dimension

Db0 ? In particular, suppose we have two D-dimensional dynamical systems with
(1) (2) (1) (2)
basin boundaries of dimensions Db0 and Db0 , where D − 1 < Db0 < Db0 < D.
What are the physical manifestations of the difference in the dimensions?
The answer to the question concerns the predictability of the asymptotic attractor,
or the final state, given an initial condition and a set of parameters, specifications of
which inevitably contain uncertainties. One is thus interested in how the predictabil-
ity can possibly be improved when the uncertainties are reduced. For concreteness,
consider the situation in which parameters of the system are fixed and the major
uncertainty in specifying the system state occurs in the initial condition. Let ε be
this uncertainty and let f (ε ) be the probability of making an error in the prediction
of the final state, which depends on ε . As ε is reduced, one expects to be able to
predict the final state more accurately, so f (ε ) will decrease. Of interest is thus the
scaling relation between f (ε ) and ε . In general, we have [289, 511, 528]

f (ε ) ∼ ε α , (5.7)

where the scaling exponent α > 0 is called the uncertainty exponent [289, 511].
For fractal boundaries, α satisfies the inequality

α < 1, (5.8)

and α = D − Db0.
For a smooth basin boundary of dimension Db0 = D − 1 in the D-dimensional
phase space, the scaling law (5.7) can be observed straightforwardly, as follows.
Since an initial condition is specified with precision ε , we can associate each initial
condition with a D-dimensional ball of radius ε , centered at the initial condition.
If a ball is located completely in the basin of attraction of an attractor, the fates of
all initial conditions in the ball are certain: they all go to this attractor. Only when
the ball crosses a boundary is a wrong prediction of the final state possible, be-
cause initial conditions contained in the ball can now go to different attractors. The
probability of making an error in prediction is thus proportional to the phase-space
volume contained within ε of the boundary, which is S0 ε ∼ ε , where S0 is the
(D − 1)-dimensional volume of the basin boundary. We thus have

f (ε ) ∼ ε ,

which gives α = 1 = D − Db0.

For a fractal basin boundary Σ of dimension D − 1 ≤ Db0 ≤ D, let V (ε ) be the
volume of all points within distance ε of Σ . To derive a relation between the un-
certainty exponent and the dimension Db0 , one can estimate the upper and lower
bounds of V (ε ) using different covering schemes [511]. Specifically, imagine that
we cover the phase space with a grid of boxes of edge length ε . Each boundary point
x of Σ is in a box that typically has 3D − 1 neighboring boxes, points in which can be
within ε of x. That is, a point y within ε of x can be in one of the 3D boxes (including
the box that contains x itself), as shown schematically in Fig. 5.6. An upper bound
5.3 Fractal Basin Boundaries and Predictability 155

Fig. 5.6 A grid of boxes

of size ε covering basin
boundary Σ . In two
dimensions, each box
Σ
containing a boundary point x ε
has 32 − 1 = 8 neighboring
boxes, which contains points x
that can be within ε of x. Any y
point y within x can be in one
of the 32 boxes. In a
D-dimensional phase space,
the number of boxes
satisfying this requirement
is 3D
ε

to V (ε ) is a coverage using 3D N(ε ) boxes, where N(ε ) is the number of ε -boxes

needed to cover the boundary:

V (ε ) ≤ 3D N(ε )ε D . (5.9)

Now let us choose a smaller grid covering Σ such that any two points in a box
√ by a distance at most ε , which can be achieved using boxes of edge
are separated
length ε√
/ D. The number of such boxes required to cover the entire boundary
is N(ε / D). In this case, every box in the coverage is within distance ε to the
boundary Σ . Therefore, we have
√ √
V (ε ) ≥ (ε / D)D N(ε / D). (5.10)

The number N(ε ) of boxes needed to cover Σ scales (see (1.19)) with ε as N(ε ) ∼
ε −Db0 . We thus have, from (5.9) and (5.10),

V (ε ) ∼ ε D−Db0 . (5.11)

Since f (ε ) ∼ V (ε ), we have f (ε ) ∼ ε α with

α = D − Db0. (5.12)

The uncertainty exponent is the difference between the dimension of the phase space
and that of the boundary.
The physical interpretation of the scaling relation (5.7) is as follows. Suppose
one wishes to reduce the probability of error in the prediction of the final state by
improving the precision in the specification of the initial conditions. If the basin
boundary is smooth so that α = 1, a reduction in ε results in an equal amount
of reduction in f (ε ). For fractal basin boundaries, where α < 1, a more precise
specification of the initial conditions results in a much smaller improvement in the
probability of predicting the final attractor correctly. In the extreme case in which
α ≈ 0, a vast reduction in the uncertainty of specifying the initial conditions will
156 5 Fractal Basin Boundaries

result in almost no improvement in one’s ability to determine the final state, which
can occur, for example, with riddled basins. In this sense, prediction is more difficult
for basin boundaries whose dimension values Db0 are larger [830].
The uncertainty exponent can be expressed in terms of the properties of the
nonattracting chaotic set embedded in the boundary. In invertible systems, fractal
basin boundaries typically contain both smooth parts and the stable manifold of
a chaotic saddle. Since the dimension of the union of two sets is that of the set
with higher dimension, we have Db0 = Ds0 , where Ds0 denotes the box-counting
dimension of the stable manifold. For two-dimensional maps (D = 2), we have
(1) (1)
Ds0 = 1 + D0 , where D0 is the partial box-counting dimension along the unstable
direction. These considerations lead to
(1) κ
α = 1 − D0 ≈ . (5.13)
λ1
(1) (1)
The approximate equality follows from (2.76) and the estimate D0 ≈ D1 , and
states that the uncertainty exponent can be approximated by the ratio of the es-
cape rate and the largest Lyapunov exponent of the chaotic saddle embedded in the
boundary. For the case of sporadic and riddled basins, see (5.22) and (5.29).
Numerically, the uncertainty exponent α can be calculated as follows. Given a
phase-space region containing some basin boundaries, we randomly choose a pair
of initial conditions x0 and x0 + ε , where ε is a small perturbation. We then de-
termine whether these two initial conditions go to the same attractor. If yes, x0
is called certain with respect to the perturbation ε . Otherwise, x0 is uncertain.
The probability f (ε ) of making an error in the prediction of the final attractor can
be estimated by choosing a large number N0 of initial conditions in the phase-space
region according to some smooth probability distribution. For example, if Nu is the
number of uncertain initial conditions with respect to ε , we have f (ε ) ≈ Nu /N0 . The
uncertainty exponent is approximated by the slope of a linear fit in the plot of f (ε )
versus ε on a logarithmic scale.
To illustrate the computation of the uncertainty exponent, we take the two-
dimensional map [511]:

θn+1 = θn + 1.32 sin(2θn ) − 0.9 sin(4θn ) − xn sin (θn ), (5.14)

xn+1 = −J0 cos (θn ),

where x can be regarded as the radial distance from the center of an annulus, θ is
an angle variable such that θ and θ + 2π are equivalent, and J0 is a parameter. (It is
this model whose quasipotential is shown in Fig. 4.6 and which has been used for
the illustration of noise-induced chaos in Fig. 4.10.) The system is invariant under
the symmetry θ → 2π − θ . The determinant of the Jacobian matrix is J0 sin2 (θ ) < 1
(for J0 < 1). There are two attractors, located at x = −J0 , θ = 0 (denoted by A− ) and
x = J0 , θ = π (denoted by A+ ), respectively. The boundaries separating the basins of
attraction of the two attractors are fractal, as shown in Fig. 5.7a for J0 = 0.3, where
black dots represent the basin of attraction of A+ . To compute the fraction f (ε ) of
5.3 Fractal Basin Boundaries and Predictability 157

a 0.5
b 0
A+

α ≈ 0.2

log10f(ε)
−1
0
x

A− −2

−0.5
0 1 2 3 −10 −5 0
θ log10ε

Fig. 5.7 (a) Fractal basin boundaries for the map (5.14) with J0 = 0.3. The two point attractors
are denoted by A− and A+ , respectively. (b) Plot of f (ε ) versus ε on a logarithmic scale. The
uncertainty exponent is estimated to be α ≈ 0.2 [511] (with kind permission from Elsevier Science)

uncertain initial conditions, we fix a line segment θ ∈ [0, π ] at x = 0, choose a pair

of initial conditions at ε -distance apart randomly from this line, and numerically
determine whether the two initial conditions approach different attractors. For every
ε -value the number N0 of the initial-condition pairs is increased until the number
of uncertain initial conditions reaches 1,000, so that f (ε ) ≈ 1,000/N0. Figure 5.7b
shows the algebraic scaling between f (ε ) and ε . A least-squares fit gives a slope of
about 0.2, which is the uncertainty exponent α . The dimension of the basin bound-
ary is thus Db0 = 2 − α ≈ 1.8.
To appreciate the value of the uncertainty exponent, say we make efforts to
reduce the uncertainty in the specification of the initial condition by five orders
of magnitude. Then α ≈ 0.2 means that the probability of making an error in pre-
dicting the final attractor is reduced only by a factor of (10−5 )0.2 = 0.1. Thus the
presence of the fractal basin boundaries makes predicting the final state difficult.
In experimental situations or in high-dimensional systems it is often difficult to
determine the initial conditions. One can then attempt to evaluate the uncertainty
exponent using random variations in the parameter space. A question is whether
the value of the uncertainty exponent so obtained is the same as that obtained us-
ing random perturbations in initial conditions. The answer is affirmative because a
parameter variation can be regarded as being equivalent to a perturbation in the ini-
tial conditions. Specifically, consider a D-dimensional map xn+1 = f(xn , p), where
p denotes a set of parameters. Assume that the system under the initial condition
x0 and parameter value p0 goes to one attractor. The standard approach is to take
a slightly different initial condition x0 = x0 + Δx at fixed parameter p0 . The image
point is then f(x 0 , p0 ). Alternatively, one can take a slightly different set of parame-
ter values p0 = p0 + Δ p, at the same initial condition, which leads to f(x0 , p 0 ). The
image points are identical if ∂ f/∂ p|x0 ,p0 · Δ p = ∂ f/∂ x|x0 ,p0 · Δx. We expect then the
uncertainty exponents computed with respect to perturbations in the parameter and
158 5 Fractal Basin Boundaries

in the phase space to be identical. In fact, one of the first determinations of frac-
tal basin boundaries [528] was done using parameter perturbations. This parameter
and phase-space equivalence was also used to study basin structures in coupled-map
lattice systems [462–464].

5.4 Emergence of Fractal Basin Boundaries

5.4.1 Basin Boundary Metamorphoses and Accessible Orbits

The typical dynamical mechanism that creates fractal basin boundaries from a
smooth boundary is homoclinic or heteroclinic tangencies, as schematically illus-
trated in Fig. 5.8, where p is a bifurcation parameter. Throughout the bifurcation
there is an unstable periodic orbit (e.g., a saddle fixed point), denoted by S. The
stable manifold of S is the basin boundary between an attractor to its right (shown)
and another attractor (not shown). As p is changed through the bifurcation point pc ,
both the basin boundary and the attractor evolve. For p < pc , the basin boundary is
smooth (Fig. 5.8a). Homoclinic tangencies between the stable and the unstable man-
ifolds of S occur at pc , as shown in Fig. 5.8b. For p > pc , the homoclinic crossings
between the stable and the unstable manifolds of S imply a Smale horseshoe-type
dynamics (Sect. 1.2.2.1) in the vicinity of S. As a result, a chaotic saddle is created
that contains the set of intersecting points between the stable and the unstable man-
ifolds. The stable foliation, and equivalently the basin boundary, becomes fractal.
The bifurcation from smooth to fractal basin boundaries is called a smooth-to-fractal
basin boundary metamorphosis [296].
A basin boundary metamorphosis is typically accompanied by a change in the un-
stable periodic orbits on the basin boundary that is accessible to the attractor [296].
A boundary point P is accessible from a region if there is a curve of finite length
that connects P to a point in the interior of the region such that no point on the curve
belongs to the boundary except point P. From Fig. 5.8, we see that the saddle fixed
point S is accessible to the attractor for p < pc . However, for p > pc , the fractal
foliations of the stable manifold entail that it is not possible to connect S to a point
on the attractor through a curve of finite length. The fixed point S is thus inacces-
sible to the attractor for p > pc . Instead, a new unstable periodic orbit, one of the
infinite number of those embedded in the chaotic saddle, becomes accessible to the
attractor.
The change in the accessible unstable periodic orbits can be demonstrated [296]
using the Hénon map with a positive Jacobian J. The map has one attractor at in-
finity. For fixed J, as the bifurcation parameter a is increased through the value
a1 = −(J + 1)2 /4, a saddle-node bifurcation occurs, creating an attracting fixed
point and a saddle point, which separate from each other for a > a1 . For a slightly
above a1 , the map has two attractors: one at the attracting point and another at infin-
ity. The basin boundary is the stable manifold of the saddle, as shown in Fig. 5.9a for
a = 1.150. The saddle point is accessible to the period-1 attractor. As a is increased
5.4 Emergence of Fractal Basin Boundaries 159

Fig. 5.8 Dynamical

mechanism that creates
fractal basin boundaries. a p < pc
(a) Smooth basin boundary attractor
for p < pc , (b) homoclinic
tangencies for p = pc , and S
(c) homoclinic crossings for
p > pc . The stable foliations
and, equivalently, the basin
boundaries, become fractal basin boundary
for p > pc

b p = pc
attractor

basin boundary

c p > pc

attractor

basin boundary

further, a metamorphosis occurs, which converts the smooth boundary into a fractal.
An example of the fractal basin boundary is shown in Fig. 5.9b for a = 1.395. We
see that, because of the fractal foliation of the basin boundary, the original saddle
fixed point is no longer accessible to the attractor, which for this parameter value has
already evolved into a period-2 attractor through a period-doubling bifurcation. The
accessible orbit on the boundary becomes a hyperbolic periodic orbit of period 4.
We can imagine that as the parameter is varied further, this new accessible
unstable periodic orbit can also have homoclinic tangencies, after which it becomes
inaccessible. The subsequent homoclinic intersections mean that the basin bound-
ary must necessarily undergo another metamorphic change to a fractal one that is
distinct from the original boundary. This is a fractal-to-fractal basin boundary meta-
morphosis, after which a different unstable periodic orbit on the boundary becomes
accessible, as illustrated in Fig. 5.9c for a = 1.405. We see that the fractal bound-
ary appears to be quite distinct from that in Fig. 5.9b, and the originally accessible
period-4 orbit in Fig. 5.9b is replaced by a period-3 orbit in Fig. 5.9c.
160 5 Fractal Basin Boundaries

Fig. 5.9 Basin boundary metamorphosis. For the Hénon map (xn+1 , yn+1 = (a − x2n − Jyn , xn ) for
J = 0.3, (a) smooth basin boundary for a = 1.150, where the accessible orbit on the boundary is the
saddle fixed point created at a saddle-node bifurcation. (b) Fractal basin boundaries for a = 1.395
after a smooth-to-fractal basin boundary metamorphosis. The accessible orbit on the boundary is
now a period-4 orbit. (c) Qualitatively different fractal basin boundaries after a fractal-to-fractal
boundary metamorphosis with a new period-3 accessible orbit for a = 1.405. The numerals in
(b) and (c) denote the accessible periodic orbits in the sequences of iterations [296] (with kind
permission from Elsevier Science)

5.4.2 Dimension Changes at Basin Boundary Metamorphoses

As the basin boundary changes characteristically, e.g., from smooth to fractal or

from fractal to fractal, we expect the dimension of the boundary to change abruptly.
This can be seen qualitatively from Fig. 5.9a–c. Let asf and aff denote the parameter
values for the smooth-to-fractal and the fractal-to-fractal boundary metamorphoses
that create the fractal basin boundaries in Fig. 5.9b and c, respectively. For a < asf
(Fig. 5.9a), the boundary is a smooth curve in the two-dimensional phase space, so
its box-counting dimension is Db0 = 1. For asf < a < aff (Fig. 5.9b), the boundary
is locally a Cantor set of smooth curves, so Db0 is between one and two. For a >
aff (Fig. 5.9c), the dimension is also a fractional (but distinct) value between one
and two.
5.4 Emergence of Fractal Basin Boundaries 161

Fig. 5.10 One-dimensional map xn+1 = f (xn ) with a positive peak at x = −1/2, which is always
above 1 ( f (−1/2) > 1), and a negative peak at x = 1/2. As the bifurcation parameter p changes
from p1 to p2 , the negative peak moves from f (1/2) > −1 at p = p1 (dashed lines) to f (1/2) < −1
at p = p2 (solid lines). A smooth-to-fractal basin boundary metamorphosis occurs at p = psf when
f (1/2) = −1

To understand the abrupt dimension change associated with a basin boundary

metamorphosis, an analyzable, piecewise linear, one-dimensional map f (x) was in-
troduced [578], as shown schematically in Fig. 5.10. The map has a positive peak at
x = −1/2, which remains above one, f (−1/2) > 1, and a negative peak at x = 1/2
whose height varies as a bifurcation parameter p changes. In particular, say for
p = p1 the negative peak is above −1 (dashed line). As p is increased from p1 to
p2 , the negative peak moves to below −1 (solid line). The map has three unstable
fixed points: U at x = −1, U1 at x = 1, and U2 in the vicinity of x = 0. As p varies
in the interval [p1 , p2 ], U and U1 are fixed, but the location of U2 can shift about
x = 0. In this parameter interval of interest, the map has two attractors: one at −∞
and another at x = +∞. Since all points in x < −1 map to the attractor at −∞ and
all points in x > +1 go to the attractor at +∞, the basin boundary must lie in the
interval [−1, 1].
Let psf denote the parameter value for which f (1/2) = −1, where p1 < psf < p2 .
For p < psf , since the negative peak at x = 1/2 stays within the square in Fig. 5.10,
the basin of the attractor at −∞ is x < −1. The basin of the attractor at +∞ consists
of the interval x > 1 and almost all points in (−1, 1) except a Cantor set of measure
zero. Let I+ be the primary escape interval in (−1, 1) such that f (x) > 1, which
maps to +∞. All preimages { f −n (I+ )} (n = 1, . . . ) also map to +∞. What is left
in (−1, 1) is a chaotic repeller. Despite the presence of the repeller, for p < psf the
basin boundary between the basins of the ±∞ attractors is a single point: x = −1.
As p increases through psf , the negative peak of f (x) pokes through f = −1.
Let I− be the primary escape interval in x ∈ (−1, 1) such that f (x) < −1, which maps
to the attractor at −∞. In addition, all preimages of I− also map to −∞. The basin of
162 5 Fractal Basin Boundaries

attraction of the −∞ attractor now consists of x < −1 and all preimages of I− . The
preimages of I+ and I− intertwine in a complicated way, and the chaotic repeller in
(−1, 1) that maps neither to −∞ nor to +∞ becomes the new basin boundary, which
is a fractal. We see that there is a smooth-to-fractal basin boundary metamorphosis
at p = psf , after which the basin boundary, which was originally the point x = −1,
jumps far into the interior of the +∞ basin. For p slightly larger than psf , there are
subintervals of the −∞ basin in (−1, 1), which for p < psf were entirely in the +∞
basin. At the metamorphosis, the dimension of the basin boundary changes abruptly
from zero to a fractional value. In particular, as p approaches psf from above, the
box-counting dimension of the basin boundary is the dimension Dsf b0 of the invariant
Cantor set in the limit of p’s approaching psf from below. As p increases further,
the dimension decreases as the Cantor set becomes “thinner” in the interval [−1, 1].
The behavior of the dimension of the basin boundary through the metamorphosis is
schematically illustrated in Fig. 5.11.
The dimension Db0 of the fractal basin boundary for p ≥ psf can be calculated as
follows. Let N(ε ) be the number of intervals of size ε needed to cover the boundary.
As specified in Fig. 5.10, let A, B, and C denote the subintervals [−1, −1/2 − ε2/2],
[−1/2 + ε2/2, 1/2 − ε1/2], and [1/2 + ε1/2, 1], and let NA (ε ), NB (ε ), and NC (ε ) be
the number of boxes of size ε needed to cover the subsets of boundary points in
these subintervals, respectively. We have

N(ε ) = NA (ε ) + NB (ε ) + NC (ε ). (5.15)

Self-similarities stipulate

ε
NA (ε ) = N , (5.16)
1/4 − ε2/4
ε
NC (ε ) = N ,
1/4 − ε1/4
ε
NB (ε ) = N .
(1/4 − ε2/4) + (1/4 − ε1/4)

Fig. 5.11 Schematic

dependence of the
box-counting dimension Db0
of the basin boundary through
a smooth-to-fractal boundary
metamorphosis
5.4 Emergence of Fractal Basin Boundaries 163

Substituting these into (5.15) and making use of the scaling N(ε ) ∼ ε −Db0 , we obtain

1 ε1 Db0
1 ε1 + ε2 Db0
1 ε2 Db0
− + − + − = 1. (5.17)
4 4 2 4 4 4

As p → psf from above, ε1 → 0, and hence the value of the box-counting dimension
b0 (ε2 ) for p = psf + 0 follows from (5.17) with ε1 = 0. For p slightly above psf ,
Dsf
so that ε1 > 0 is small, we can write Db0 = Dsfb0 − η , where η is small. Substituting
this into (5.17) and expanding for small ε1 and η yields

b0 − K(ε2 )ε1 ,
Db0 = Dsf (5.18)

where K(ε2 ) is a smooth function of ε2 [578]. Assuming smooth dependencies

of ε1 and ε2 on the system parameter p for p > psf , we see from (5.18) that
Db0 varies smoothly with p. For instance, suppose ε1 has a power-law depen-
dence on p: ε1 (p) ∼ (p − psf )γ , where γ > 0. Then this dependence is reflected
γ
b0 − Db0 ∼ (p − psf ) , for p > psf .
in Dsf

5.4.3 A Two-Dimensional Model

The structure of fractal basin boundaries and basin boundary metamorphoses in

two dimensions can be understood by constructing invertible-map models based
on the horseshoe dynamics [578], such as the one shown schematically in Fig. 5.12.
Consider the rectangle ABFE, outside which there are two attractors, denoted by
L and R. All initial conditions to the left of the vertical line AB lead to trajecto-
ries approaching the attractor L, and all initial conditions to the right of EF go to
the attractor R. In ABFE, there are three unstable periodic orbits, denoted by S1 , S2 ,
and S3 . The action of the dynamics is that of a double horseshoe, i.e., the rectan-
gle is squeezed vertically, stretched horizontally, and placed back into the original
rectangular region, forming a double S-shape. As a system parameter p changes
from p1 (Fig. 5.12a) to p2 (Fig. 5.12b) and to p3 (Fig. 5.12c), a smooth-to-fractal
basin boundary metamorphosis occurs for psf (p1 < psf < p2 ), and a fractal-to-
fractal basin boundary metamorphosis occurs for pff (p2 < pff < p3 ). Throughout
the parameter range, the stable and the unstable manifolds of S3 cross each other
homoclinically. As can be argued below, the smooth-to-fractal boundary metamor-
phosis at psf is induced by the homoclinic tangencies between the stable and the
unstable manifolds of S1 , while the fractal-to-fractal boundary metamorphosis is
induced by those between the stable manifold of S2 and the unstable manifold of S3 .
In Fig. 5.12a, the unstable manifolds of S2 and S3 cross the stable manifold of S3 ,
so there is a vertical bar UR that maps to the region DR located to the right of the
vertical line EF. As a result, all initial conditions in the rectangle CDFE, except
for a set of measure zero, map asymptotically to the right of EF and approach the
attractor R. We also see that the unstable manifolds of S1 and S2 cross the stable
manifold of S2 , and hence the vertical strip VR maps to the region CR to the right of
164 5 Fractal Basin Boundaries

Fig. 5.12 A two-dimensional horseshoe model, where L and R are two attractors, and S1,2,3 are
saddle points. The double S-shaped band represents the image of rectangle ABFE under the map.
(a) For p = p1 < psf , the vertical line AB is the smooth basin boundary and S1 is then accessible to
the attractor R. (b) For psf < p = p2 < pff , the basin boundary consists of a Cantor set of vertical
lines in the rectangle ABDC and the accessible orbit to R is replaced by S2 . (c) For p = p3 > pff ,
the basin boundary is a Cantor set of vertical lines in the larger rectangle ABFE. In this case, S2 is
no longer accessible to R. The newly accessible orbit on the basin boundary is S3 [578] (with kind
permission from World Scientific Publishing Co.)

the vertical line CD. The consequence is that all initial conditions in ABDC, except
for a set of measure zero, map to the right of CD. Combining the dynamics on the
rectangles CDFE and ABDC so described, we see that all initial conditions in the
larger rectangle ABFE, except for a set of measure zero, result in trajectories that
asymptotically go to the attractor R. In this case p = p1 < psf , and the boundary
between the basins of attraction of L and R is AB, the stable manifold of S1 , which
is smooth. Furthermore, S1 is an unstable periodic orbit on the basin boundary that
is accessible to the attractor R.
For p = psf (not shown), homoclinic tangencies between the stable and the unsta-
ble manifolds of S1 occur, resulting in subsequent homoclinic crossings for p > psf ,
as shown in Fig. 5.12b. In this case, the vertical strip V L maps to the region CL to
the left of AB and goes to the attractor L. However, the vertical strip VR still maps
5.5 Wada Basin Boundaries 165

to the right of CD, which eventually goes to the attractor R. In ABDC, all initial
conditions (except for a set of measure zero) go either to the attractor R or to the
attractor L. The vertical strips that approach asymptotically R and L define two hor-
izontal Cantor sets that intertwine in a fractal manner. The basin boundary is thus
fractal with a dimension between 1 and 2. Because of the fractal boundary, S1 is no
longer accessible to R for p > psf . The newly accessible periodic orbit to R is S2 .
As p increases further, we can imagine that for p2 < p = pff < p3 , the unsta-
ble manifold of S3 becomes heteroclinically tangent to the stable manifold of S2
(the vertical line CD). As shown in Fig. 5.12c, the subsequent heteroclinic cross-
ings stipulate that the vertical bar UL in CDFE maps to the region DL to the left
of CD. As a result, not only a set of vertical strips in ABDC but also such a set in
CDFE map to the attractor L. However, there are vertical strips in ABDC and CDFE
that map to the attractor R. We see that as p increases through pff , the fractal basin
boundary originally confined to the small rectangle ABDC extends suddenly into
the rectangle CDFE, causing a sudden increase in the box-counting dimension of
the basin boundary from one fractional value to another between 1 and 2. After the
heteroclinic tangencies, S2 is no longer accessible to the attractor R. For p > pff , the
accessible orbit to R on the basin boundary is S3 .

5.5 Wada Basin Boundaries

Our discussion so far has been restricted to situations in which there are two
coexisting basins of attraction. When a dynamical system possesses more than two
coexisting attractors, a type of fractal basin boundary, namely Wada basin bound-
aries [406, 558–561, 613], can arise. For such a case, every boundary point of one
basin of attraction is simultaneously a boundary point of the other basins.
To imagine a Wada basin boundary, take the map of the continental United States
and consider the boundaries between the states. Almost all boundary points are
common to two states, but there are a few dozen of points that are common to
three states, and there exists a single boundary point that is shared by four states
(the Four-State Corner bordering Arizona, Utah, Colorado, and New Mexico). In the
realm of nonlinear dynamical systems, more complicated situations can arise: the set
of boundary points that are common to more than two basins of attraction can be
fractal. The history and the topology of Wada basins are presented in detail in the
review of Aguirre et al. [7].
To give an example, we examine the forced Duffing’s oscillator for parame-
ters where on the stroboscopic map, there are two fixed-point attractors, and an
attractor at infinity. The basins of attraction of these three attractors are shown in
Fig. 5.13 in three different colors. The Wada property of the basin boundaries can be
seen, since successive magnifications of any region containing the boundary exhibit
all three colors. Another example from the forced damped pendulum is shown in
Fig. 5.14.
166 5 Fractal Basin Boundaries

a 0.5

y−0.5

−1
−0.5 0 0.5 1 1.5
x
b −0.45
y

−0.64
0.54 0.7
x
c −0.52
y

−0.55
0.58 0.596
x

Fig. 5.13 Wada basin boundary for the forced Duffing’s oscillator ẍ + 0.1ẋ + x − x2 =
0.06 sin (0.8t) on the stroboscopic section (x, y ≡ ẋ), where (b) is a magnification of the box in
(a), and (c) is the magnification of the box in (b) (Figure by Y. Do.)

For two-dimensional invertible maps, or equivalently, three-dimensional flows,

the mechanism for Wada basin boundaries is well understood, due to the work
of Kennedy, Nusse, and Yorke [406, 558–561]. In particular, Kennedy and Yorke
proved a theorem [406] stating that if p is a periodic point on the basin boundary,
and if the following two conditions are satisfied, (1) its unstable manifold intersects
every basin, and (2) its stable manifold is dense in each of the basin boundaries, then
the basins have the Wada property. This can be intuitively understood by referring to
Fig. 5.15, where there are K coexisting basins, denoted by B1 , B2 , . . . , BK . Suppose
p is a periodic point on the boundary of B1 that is accessible to B1 . Let W s (p)
and W u (p) be the stable and the unstable manifolds of p, where W s (p) is the basin
boundary of B1 . Now arbitrarily choose a point x ∈ W s (p) and imagine a circle
Cε (x) of radius ε centered at x. Since W u (p) intersects every basin, Cε (x) must
contain points of every basin, which can be seen by considering a one-dimensional
curve segment Dk in the basin Bk that intersects W u (p), for k = 1, . . . , K. Under
inverse iterations of the map, the images of the curves will be arbitrarily close to
5.5 Wada Basin Boundaries 167

a 4 b 0.9

y
y

−2−π π −0.5
0
x −0.6 x 1.5

c d
0.5 0.36
y
y

0.2 0.32
0.2 x 0.5 0.27 x 0.3

Fig. 5.14 Wada basin boundary for the forced damped pendulum γ = 0.2, f 0 = 1.66 in (5.2), on
the stroboscopic section (x, y ≡ ẋ). At these parameters, three attracting limit cycles coexist. Panels
(b, c, d) are successive magnifications of boxes in (a, b, c), respectively (Figure by Y. Do)

u
W (p)

Dk
Bk
(−n)
F (Dk)

Ws(p) p B1
x
Cε(x)
Fig. 5.15 Schematic
illustration of the setting for
establishing the Wada
property. See text for details
168 5 Fractal Basin Boundaries

the stable manifold of p and therefore be in Cε (x) [547]. Thus, the boundary of B1
must be the boundaries of all other basins. Since W s (p) is dense in each of the basin
boundaries, all boundaries are common to all basins, and hence the Wada property
is fulfilled.
Computationally, to verify condition (1), one can plot a piece of the unstable
manifold, trace it under the dynamics, and determine whether it intersects all
basins of interest. Condition (2) is more difficult to verify from numerical traces
of the stable manifold. To overcome these difficulties, Nusse and Yorke pro-
posed the construction of basin cells, which leads to numerically verifiable
conditions guaranteeing that the boundary of a basin is a Wada basin boundary
[559, 560].
To explain the idea of basin cells, consider an invertible dissipative map f in the
plane. Traditionally, the basin of attraction of an attractor is defined to be the set
of points that approach the attractor asymptotically. Since there has been no rigor-
ous way to determine whether an attractor is chaotic or whether there are multiple
coexisting attractors, this concept of “basin” is in principle ill defined. Nusse and
Yorke redefined a “basin” as the set of points that enter a trapping region [559,560].
A compact region Q is a trapping region if f(Q) ∈ Q and f(Q) = Q. These two
conditions guarantee that a trajectory entering a trapping region does not leave the
region, and there must then be at least one attractor inside. The basin of the trap-
ping region Q is the set of points that map into the interior of Q. A trapping region
may contain invariant sets such as chaotic saddles; that is, there can be points in
the region whose trajectories do not converge to an attractor. Trapping regions of
practical importance are those having piecewise smooth boundaries that consist of
finitely many smooth curve segments. If a trapping region Q is constructed such that
(1) there is an unstable periodic orbit on its boundary and (2) the boundary consists
of pieces of the stable and the unstable manifolds of the periodic orbit, then Q is
a basin cell. Although there is an infinite number of unstable periodic orbits on a
fractal basin boundary, only a few may be “qualified” to generate a basin cell. Thus,
in order to have a basin cell, the unstable periodic orbit on the cell boundary needs
to be chosen carefully [559, 560].
Figure 5.16 illustrates two types of basin cells that are topologically equivalent to
some basin cells that can be explicitly constructed from the system of forced damped
pendulum in parameter regions with Wada basin boundaries shown in Fig. 5.14a–d.
Let P denote the unstable periodic orbit that generates a basin cell CP . As shown
in Fig. 5.16, for a periodic point p of P, its unstable manifold can form an arc that
starts from p and ends at a corner point, an intersection point between the stable
and the unstable manifolds of p. Such an arc is outside the basin cell (except the
endpoints). The union of all supporting arcs, one for each periodic point p of P,
is called the scaffolding of the basin cell CP [559, 560]. For instance, for the cell
in Fig. 5.16a, the scaffolding consists of the union of two supporting arcs, while
in Fig. 5.16b, the scaffolding is the union of three supporting arcs. With such a
geometric construction, Nusse and Yorke proved the result that if the scaffolding of
CP intersects at least two other basins, the boundary of the basin cell CP is a Wada
basin boundary. This result is remarkable because all the quantities and conditions
5.5 Wada Basin Boundaries 169

Fig. 5.16 Schematic illustration of two types of basin cells that can be constructed from three-
dimensional flows with Wada basin boundaries. In (a), the basin cell is generated by an unstable
periodic orbit of period 2. In (b), the cell-generating periodic orbit has period 3. See [559, 560]

are numerically verifiable. Using this result, one can prove that the basin boundaries
such as those in the forced Duffing’s oscillator (Fig. 5.13) and in the forced damped
pendulum (Fig. 5.14) are Wada basin boundaries [559, 560].
A natural question is how Wada basin boundaries can arise as a system parameter
changes. One route was discovered by Nusse et al. [558], which is by a saddle-node
bifurcation on a fractal basin boundary. Specifically, if the system under consid-
eration already has two coexisting attractors with a fractal basin boundary, then
a saddle-node bifurcation on the boundary can create a third attractor and a third
basin of attraction. In this case, if a basin cell can be constructed that contains the
third attractor and if the scaffolding of an unstable periodic orbit on the boundary of
the cell intersects the original two basins, the fractal basin boundary becomes Wada.
Saddle-node bifurcation on a fractal basin boundary can in fact result in
an extreme form of indeterminacy in dynamical systems, as pointed out by
Thompson [780–782], who asked what happens to an orbit initially placed on a
periodic attractor (node) when it is destroyed via slow variation of a parameter
through a saddle-node bifurcation. He discovered that if there are at least two other
attractors (in addition to the periodic attractor to be destroyed through the saddle-
node bifurcation) with a fractal basin boundary between their basins of attraction,
and if the saddle is located on the boundary (in fact Wada), then the bifurcation
can be indeterminate in the following sense. After the system drifts through the
bifurcation, to which attractor the orbit goes depends sensitively on small effects
such as noise, computer roundoff, and the way the parameter is changed. From our
discussion, we see that this extreme type of indeterminacy is closely related to the
occurrence of a Wada basin boundary. In particular, say μ is the bifurcation param-
eter, and as μ is decreased through the critical value μ0 , a saddle-node bifurcation
occurs. Assume that two other attractors exist in a parameter interval about μ0 with
a fractal basin boundary. Thus for μ0 − ε1 < μ < μ0 there are three attractors with
a Wada basin boundary, where ε1 > 0 is a small constant. Now imagine that μ
increases from this situation through μ0 . For μ0 < μ < μ + ε2 , where ε2 > 0 is
small, there are only two attractors with a fractal basin boundary. Thompson’s result
170 5 Fractal Basin Boundaries

indicates that in a situation in which random perturbations or computational errors

are present, as μ adiabatically increases through μ0 , it is fundamentally impossible
to determine where an orbit placed on the node (attractor) for μ0 − ε1 < μ < μ0
would go.

5.6 Sporadically Fractal Basin Boundaries

Sporadically fractal basin boundaries have the character of a bounded curve, say
y = g(x), such that g(x) is a differentiable (or smooth) function except for a set of x
values of zero measure. Furthermore, the nondifferentiable set of x values is a fractal
set with dimension less than one. The curve thus has a dimension between one
and two. This type of basin boundary was discovered by Rosa et al. [651, 652] and
was subsequently analyzed rigorously by Hunt et al. [346]. It was conjectured that
sporadically fractal basin boundaries exist in typical dynamical systems of phase-
space dimension at least two for noninvertible maps, at least three for invertible
maps (thus at least four for flows).
In order to highlight the relevance of sporadically fractal basin boundaries to
physical situations, in what follows we describe the occurrence of this type of
boundary in the context of chaotic phase synchronization in systems described by
differential equations. We then discuss a mathematical model to understand the dy-
namical origin and properties of these exotic basin boundaries.

5.6.1 Chaotic Phase Synchronization

Chaotic phase synchronization was discovered by Rosenblum, Pikovsky, and Kurths

in 1996 [653] and has since become an active area of research (see [80,602]). If one
examines a chaotic attractor from a three-dimensional flow, such as the Rössler
oscillator, one finds that trajectories on the attractor exhibit rotation-like motions
around the z-axis, as shown in Fig. 5.17. The motions have a well-defined center of
rotation and a unique direction (counterclockwise in Fig. 5.17b). Using the center
of rotation and an arbitrary reference line, a rotational angle φ can be defined, as
shown schematically in Fig. 5.18.
For the Rössler attractor in Fig. 5.17,
 it is convenient to use cylindrical coor-
dinates: (x, y, z) → (r, φ , z), where r = x2 + y2 and φ = tan−1 (y/x) (within one
rotation). For a chaotic trajectory, from a reference point corresponding to the ini-
tial condition, the phase variable φ (t) is a monotonically increasing function of t.
In cylindrical coordinates, the Rössler equations can be conveniently written as
dx/dt = R(x), where x ≡ (r, φ , z). The question Rosa et al. asked [652] was whether
the phase variable can be locked with respect to an external periodic driving. One
can imagine that if the periodic driving is weak, the phase will be chaotic, but under
strong driving, it is likely that the chaotic rotation would follow more or less that of
the external periodic pattern. In this case the phase variable of the chaotic oscillator
5.6 Sporadically Fractal Basin Boundaries 171

a b
15
5

10
0
z

y
5
−5
0
5
0 10 −10
y −5 0
x −5 0 5 10
−10 −10 x

Fig. 5.17 (a) Chaotic attractor from the Rössler oscillator: ẋ = −(y + z), ẏ = x + 0.25y, and ż =
0.9 + z(x − 6.0), in the three-dimensional phase space. (b) Projection of the attractor on the (x, y)-
plane. A chaotic trajectory exhibits a rotation-like motion

Fig. 5.18 Definition of phase y

variable φ for a chaotic
rotation

φ (t)
x

is said to have been synchronized (paced) with the phase of the external periodic
signal. To address this question, consider the following general system:

dx
= S(x, s)R(x) + AP(t), (5.19)
dt

where S(x, s) = 1+s(r2 − r̄2 ), P(t) = [0, sin (2π t/T ), 0], s and r̄ are parameters of the
modulating function S(x, s) (r̄ can be chosen to be the average value of r(t) for s = 0
and A = 0), and A is the amplitude of the external periodic driving. To search for
synchronization, it is convenient to use θ = φ (t) − 2π t/T , the phase difference be-
tween the chaotic oscillator and the external periodic signal. Phase synchronization
is defined by the locking of θ within 2π : −π < θ < π . For a given T , phase synchro-
nization was found to occur for sufficiently large values of A [652]. In fact, there is
a region of finite area in the two-dimensional parameter space (T, A) in which phase
synchronization occurs. To understand the fundamental dynamical mechanism for
the synchronization, Rosa et al. suggested to define the angle variable θ on the
172 5 Fractal Basin Boundaries

real line, −∞ < θ < +∞, rather than on the circle, −π ≤ θ ≤ π . The angle thus
becomes lifted. The phase-synchronized state corresponds to an attractor confined
within −π < θ < π . The attractor is chaotic because the amplitude dynamics re-
mains chaotic even when its phase is locked. Due to the invariance of the system
under the transformation θ → θ ± 2π , there is an infinite array of such attractors
spaced by 2π in θ .
Imagine for a fixed T , as A increases through a critical value Ac , phase syn-
chronization occurs so that for A > Ac , an infinite array of attractors is formed.
For A < Ac , there is no phase locking so that θ cannot be confined in any of the
2π intervals. A trajectory can thus move across the entire θ -axis. However, for A
slightly below Ac , the trajectories will be confined within one of the 2π intervals for
long time before moving to an adjacent 2π interval. There is thus transient chaos
corresponding to the temporal phase locking. The time it takes for a trajectory to
escape an attractor and to move to an adjacent one is typically much smaller than
the time that the trajectory stays on the attractor. What can be expected is thus the
confinements of θ values within 2π for long stretches of time and rapid jumps of
magnitude 2π amid the long confinements. This 2π -jump phenomenon has indeed
been observed numerically and experimentally [80]. The point is that the transition
to chaotic phase synchronization can be regarded as crisis-like transition whereby
isolated chaotic attractors are formed from transient chaos.
Rosa et al. found that after the onset of phase synchronization, the basin bound-
aries between the chaotic attractors in two adjacent 2π cells are sporadically fractal.
The boundaries are in fact similar to those from the two-dimensional map (5.5), as
shown in Fig. 5.4.

5.6.2 Dynamical Mechanism

To understand how sporadically fractal basin boundaries can arise in dynamical

systems, Hunt et al. [346] proposed a class of two-dimensional maps (5.5), where
the x-dynamics is governed by the following one-dimensional map:
⎧
⎪
⎪ 9x/(4 − 5x), for x ≤ 0,
⎪
⎪
⎪
⎪
⎨9x/4, for 0 ≤ x ≤ 4/9,
f (x) = (5.20)
⎪
⎪ (81/4)(x − x2), for 4/9 ≤ x ≤ 5/9,
⎪
⎪
⎪
⎪
⎩(9/4)(1 − x), for x ≥ 5/9,

as shown in Fig. 5.19.

This map has two invariant sets: a stable fixed-point attractor at x = −1 with a
negative Lyapunov exponent and a “middle ninth” Cantor set in 0 ≤ x ≤ 1, the set
of initial conditions in the unit interval that do not approach the attractor. The Can-
tor set is in fact a repeller with a positive Lyapunov exponent ln (9/4) because the
5.6 Sporadically Fractal Basin Boundaries 173

Fig. 5.19 One-dimensional map (5.20) as f (x) in (5.5)

slopes of the map about the Cantor set are ±9/4. For the two-dimensional map (5.5),
at each iterate every vertical line segment is expanded by the factor λ > 1. Thus al-
most all initial conditions approach asymptotically either y = +∞ or y = −∞, which
can be regarded as the two attractors of the system. The boundary between the two
basins of attraction is a function y = g(x), the dynamics on which are determined by
f (x). Since f (x) has two invariant sets and since the y-dynamics is unstable, on the
basin boundary there are two invariant sets as well: a saddle point (at x = −1) with
one positive and one negative Lyapunov exponent, and a chaotic repeller with two
positive Lyapunov exponents. Numerical experiments revealed that g(x) is smooth
for almost all x values but nondifferentiable for a set of x values constituting the
middle-ninth Cantor set in the one-dimensional map f (x). The box-counting di-
mension of the curve y = g(x) turns out for λ = 1.1 to be Db0 ≈ 1.75 [346].
How is it that the basin boundary curve can be smooth at all x except for a set
of measure zero, yet has a box-counting dimension greater than 1? To understand
this property, Hunt et al. [346] considered the Hölder exponent H(x) at x of the
function g(x): |Δy| ∼ |Δx|H(x) , where Δx is infinitesimal and Δy = g(x + Δx) − g(x).
If H(x) < 1, g(x) is not differentiable at x, but if g(x) is differentiable at x, then
H(x) = 1. Points on y = g(x) with H(x) < 1 exhibit a cusplike, spiked behavior.
Now consider two nearby points (x0 , y0 ) and (x0 +Δx0 , y0 +Δy0 ) on the basin bound-
ary and iterate them n times under the map (5.5). Since they are on the basin
boundary, their images (xn , yn ) and (xn + Δxn , yn + Δyn ) must also be on the basin
boundary. For n not too large, Δxn and Δyn can still be regarded as small quantities.
Since the nth iterate of the map (5.5) provides a smooth transformation of the
neighborhood of (x0 , y0 ) to the neighborhood of (xn , yn ), the Hölder exponents are
174 5 Fractal Basin Boundaries

the same for (x0 , y0 ) and (xn , yn ): |Δy0 | ∼ |Δx0 |H and |Δyn | ∼ |Δxn |H . From (5.5),
we have Δxn ∼ exp (λ1 n)Δx0 , where λ1 is the Lyapunov exponent of the one-
dimensional map f (x). If H < 1, |Δy| is much greater than |Δx|, so the effect of
Δx on the y-dynamics is negligible. We have Δyn ∼ exp (λ2 n)Δy0 , where λ2 = ln λ .
The Hölder exponent is thus given by H = λ2 /λ1 if λ1 > λ2 . Otherwise, we have
H = 1 because the assumption H < 1 is contradicted. Since there are two possible
values for the Lyapunov exponent λ1 in the one-dimensional map f (x) (correspond-
ing to the two invariant sets), and particularly λ1 = ln (9/4) for trajectories on the
middle-ninth Cantor set, we see that H < 1 if λ < 9/4. For randomly chosen x on
the basin boundary (with Lebesgue measure one), the trajectory goes to the attractor
at x = −1 that has λ1 < 0. For these points H = 1, and the boundary is smooth.
The relationship between the Hölder exponent and the Lyapunov exponents can
be used to obtain the box-counting dimension of the basin boundary curve y = g(x).
Suppose we cover the (x, y)-plane with square boxes of linear size ε 1. If the
boundary curve contained in a column of width ε is smooth, i.e., no points of the
middle-ninth Cantor set lie in it (H = 1), the number of boxes required to cover
the curve segment is of order 1. If the boundary curve in a column of boxes contains
points of the Cantor set so that H < 1, the variation of the curve in the y-direction is
|Δy| ∼ ε H (because Δx = ε ). The number of boxes required to cover the boundary
curve in this column is thus of order ε H /ε = ε H−1 . Since the total number of boxes
needed to cover the Cantor set (the chaotic repeller of f (x)) of dimension Dx is ε −Dx ,
the number of boxes necessary to cover the spiked parts of the basin boundary curve
is ε −(1+Dx −H) . This implies that the box-counting dimension of the chaotic repeller
embedded in the basin boundary is D0 = 1 + Dx − H. Taking into account the fact
that the number of boxes required to cover the smooth parts of the boundary is of
order ε −1 , we see that if 1 + Dx − H < 1, then ε −1 is much greater than ε −(1+Dx −H) ,
so in this case the number of boxes needed to cover the whole boundary curve is of
order ε −1 . Conversely, this number is of order ε −(1+Dx −H) if 1 + Dx − H > 1. These
estimates yield the box-counting dimension of the boundary curve y = g(x) as

Db0 = max{1, 1 + Dx − H}. (5.21)

An interesting observation is that the basin boundary’s being spiky, i.e., H < 1, is not
sufficient to make Db0 > 1, i.e., to make the boundary sporadically fractal. To have
Db0 > 1 requires H < Dx , i.e., that the dimension D0 of the repeller be larger than 1.
This means that the spiked behavior should be sufficiently intense for sporadically
fractal boundaries to arise. The uncertainty exponent is thus

α = 2 − Db0 = 1 − Dx + H = 2 − D0 < 1. (5.22)

For the model described by (5.5) and (5.20), we have Dx = ln 2/ ln (9/4) and H =
ln λ / ln (9/4), so Db0 = D0 = 2 − α = 1 + ln (2/λ )/ ln (9/4) > 1 for λ < 2 and
Db0 = α = 1, D0 < 1, for λ > 2.
5.7 Riddled Basins 175

5.7 Riddled Basins

We have seen up to now several types of complicated basins. Despite differences, a

common feature among them is open sets (volumes) contained in a basin. In this
section we discuss riddled basins, basins that do not contain any open sets but
nonetheless have a positive Lebesgue measure. Because a riddled basin has no open
sets, for every initial condition that approaches the attractor with a riddled basin
asymptotically, there are initial conditions arbitrarily nearby that go to another
coexisting attractor. Thus, an arbitrarily small uncertainty in the initial condition
can lead to a completely different attractor. Riddled basins are therefore space-
filling. In a D-dimensional phase space, Db0 = D, riddled basins are fat fractals
[234, 564, 773]. In fact, the uncertainty exponent associated with a riddled basin is
close to zero, which means that a vast reduction in the error in specifying the initial
conditions results in hardly any improvement in one’s ability to predict the final at-
tractor. As a consequence, prediction of attractors for specific initial conditions and
parameters becomes practically impossible. Because of this serious consequence,
there has been much effort devoted to riddled basins (for a review and historical
comments, see [7]).
The dynamical conditions for riddling to occur were first described by Alexander
et al. [11]. They offered the following definition for a riddled basin: The basin
of attraction of an attractor is riddled if its complement intersects every disk in
a set of positive measure. Roughly, the term “disk” here refers to D-dimensional
phase-space volumes of all sizes. In order to argue that the basin of a chaotic at-
tractor is riddled, the following two conditions need to be established: (1) a set
of positive measure is attracted to the attractor; and (2) sufficiently many points
near the attractor are repelled from it. In particular, to prove condition (1), one can
compute the transverse Lyapunov exponent (to be defined below) and show that
it is negative [11]. To prove condition (2), it is necessary to show that there ex-
ists an open dense set near the attractor where points approach another coexisting
attractor. In contrast, a fractal basin is open and it is defined with respect to the basin
boundary: a basin is fractal if its boundary is a fractal set. The mathematical feature
that distinguishes a riddled basin from a fractal one is then that the former is a closed
set of positive measure, while the latter is open.
A necessary condition for riddling is the existence of an invariant subspace,
which often results from a symmetry of the system. An example is the following
system of N coupled chaotic oscillators:

dxi
= Fi (xi ) + K ∑ H(xi − x j ), i = 1, . . . , N, (5.23)
dt j

where Fi (xi ) is the velocity of oscillator i when uncoupled, and the coupling is
represented by strength K and the function H(xi − x j ) that satisfies the condi-
tion H(0) = 0. When the individual oscillators are identical, i.e., Fi = F j , the
176 5 Fractal Basin Boundaries

synchronous state xi (t) = x j (t) (i, j = 1, . . . , N) is a solution to (5.23). In this case,

the dynamical equations are identical for each oscillator, so oscillators starting syn-
chronized remain so forever. The subspace defined by xi (t) = x j (t) (i, j = 1, . . . , N)
is therefore invariant. The existence of such an invariant subspace was the start-
ing point for analyzing the dynamics of coupled chaotic oscillators in most existing
works in the area of chaotic synchronization [602, 711]. In fact, the first experimen-
tal evidence for riddled basins was found in this context (see Fig. 1.18).

5.7.1 Riddling Bifurcation

How does a riddling bifurcation occur that creates a riddled basin? The answer
was provided in [455]. In a two-dimensional phase space, the invariant subspace
is a line. In this case, the onset of riddling is determined by a saddle-repeller bifur-
cation [294, 295]. In particular, the chaotic attractor A in the invariant subspace is
one-dimensional. Before the bifurcation, A attracts all points in some neighborhood
of itself, and all the periodic orbits embedded in the chaotic attractor are saddles in
the full phase space. At the riddling bifurcation, one of the periodic orbits, usually
of low period, becomes transversely unstable. Since this periodic orbit is already
unstable in the attractor, it becomes a repeller in the two-dimensional phase space.
To be concrete, let xp be an unstable fixed point embedded in the chaotic attractor in
the invariant subspace. The point is stable transversely to this subspace for p < pc ,
as shown in Fig. 5.20a. Riddling occurs when xp loses its transverse stability as a
parameter p passes through a critical value pc . For such systems, the loss of trans-
verse stability is induced by the collision at p = pc of two point repellers r+ and r− ,
located symmetrically with respect to the invariant subspace, with the saddle at xp
(a saddle-repeller bifurcation). These two repellers exist only for p ≤ pc , as shown
in Fig. 5.20a. For p > pc , the saddle xp becomes a repeller, and the two repellers r+
and r− off the invariant subspace no longer exist.
Due to nonlinearity, a “tongue” opens at xp , allowing trajectories near the
invariant subspace to escape for p > pc , as shown in Fig. 5.20b. Each preimage of
xp also develops a tongue simultaneously. Since preimages of xp are dense in the
invariant subspace, an infinite number of tongues open up simultaneously at p = pc ,
indicating that initial conditions arbitrarily close to the invariant subspace can go to
another attractor.
At the riddling bifurcation a single periodic orbit becomes transversally unstable.
As the parameter p is increased further, more and more periodic orbits become un-
stable until, for another critical parameter p0c (> pc ), the full attractor in the invariant
subspace becomes transversely unstable. This occurs when the average transverse
Lyapunov exponent λT becomes positive. This bifurcation is called the blowout
bifurcation [30, 435, 727].
5.7 Riddled Basins 177

Fig. 5.20 Riddling

bifurcation at pc . (a) Unstable
saddle fixed point in the
invariant subspace and two
repellers off the invariant
subspace for p < pc (before
the saddle-repeller
bifurcation). (b) Tongue
structure formed for p > pc ,
after the onset of riddling.
Trajectories originated from
initial conditions inside the
tongues escape the invariant
subspace to +∞ [455]
(Copyright 1996, the
American Physical Society)

5.7.2 An Example

To make these ideas more concrete, we use the following map [455]:
xn+1 = 4xn (1 − xn), (5.24)

yn+1 = pe−b(xn −x p yn + y3n ,

)2

where y = 0 defines the invariant subspace as a trajectory with y0 = 0 will have

yn = 0 and b > 0 is a parameter. In this system, A is the fully developed chaotic
attractor of the logistic map, and x p = 3/4 denotes the nontrivial unstable fixed point
of the logistic map.
The two eigenvalues of the unstable fixed point xp = (x p = 3/4, y = 0) are
(Λx , Λy ) = (−2, p). Thus, xp is stable in the y direction for p < 1 and unstable for
p > 1. This fixed point is a saddle for√ p < 1. For p < 1, there are two other unstable
fixed points located at r± ≡ (x p , ± 1 − p), which have eigenvalues (−2, 3 − 2p),
both being repellers for p < 1, as shown in Fig. 5.20a. The repellers collide with
each other and with the saddle at p = pc = 1 in a saddle-repeller bifurcation, and
178 5 Fractal Basin Boundaries

they do not exist for p > 1. Thus, beyond the riddling bifurcation (for p > 1), two
tongues, symmetrically located with respect to the invariant subspace, open up at
x = x p , allowing trajectories near y = 0 to escape to |y| = ∞. Observe that the cubic
term in the y-dynamics guarantees that if |yn | > 1, then |yn+1 | > |yn | > 1. Once a tra-
jectory reaches |y| = 1, its y value tends to infinity rapidly. As a result, y = ±∞ can
be regarded as coexisting attractors, A± , of (5.24) with A , fulfilling the condition
for riddling.
The transverse Lyapunov exponent is the average of the logarithms of the stretch-
ing rates of the y-dynamics at y = 0 along a trajectory of the x-dynamics. That is,
  
 dyn+1 
λT = ln    , (5.25)
dyn  |y=0
where the angled brackets denote an average taken with respect to the natural
measure of the attractor A . Since dyn+1/dyn = p exp [−b(xn − x p )2 ] at y = 0, we
have λT = ln p − b(x2  − 2xx p + x2p ). Substituting the averages for the chaotic
attractor of the x-dynamics, we obtain

λT = ln p − 3b/16. (5.26)

The blowout bifurcation takes place at p0c = exp (3b/16) > 1.

5.7.3 Scaling Relation

A quantity characterizing the degree of riddling of a basin is the ratio between the
sizes of the basins of the attractors A+ (or A− ) and A , which can be computed
as follows. Take a line parallel to the invariant subspace at distance y0 1 and
determine the fraction F(y0 ) of the length of this line in the basin of A+ or (A− ).
The fraction typically obeys the following scaling law [567]:

F(y0 ) ∼| y0 |η , (5.27)

where η is a positive exponent. As y0 → 0, the fraction of the basin of A± ap-

proaches zero, but for any finite | y0 | this fraction is nonzero. For p values close to
the blowout bifurcation point p0c , a stochastic model by Ott et al. [567] predicts the
exponent η to be
| λT |
η= , (5.28)
Q
where Q represents the diffusion coefficient characterizing the variance of the finite-
time transverse Lyapunov exponents (the analogue of Q2 introduced in Sect. 2.2.2
and in Appendix A). Close to but below the blowout bifurcation point, λT is negative
and small.
Another measurable quantity is the uncertainty exponent α , defined by
f (ε ) ∼ ε α , where f (ε ) is the probability of finding two points within distance ε
along a line at distance y0 from the invariant plane, which belong to different basins
5.8 Catastrophic Bifurcation of a Riddled Basin 179

(to those of A and of A+ (or A− )). The stochastic theory predicts, for p close to p0c ,
that
λT 2
α= , (5.29)
4Qλ1
where λ1 is the Lyapunov exponent of attractor A . Since λT is small, the uncertainty
exponent is small, signifying a fundamental obstacle to prediction. Due to the fat-
fractal nature of the boundary (5.12) does not hold. In this case, we have Db0 = D
but α = 0.

5.8 Catastrophic Bifurcation of a Riddled Basin

While symmetry and invariance are common in mathematical models of physical

systems, the notion of symmetry and invariance is nongeneric, because in physi-
cal reality, imperfections or perturbations that destroy system symmetry are always
present. In the coupled-oscillator system (5.23), a typical type of imperfection is
parameter mismatches among oscillators. The presence of heterogeneity among the
vector fields Fi , no matter how small, immediately destroys the originally invariant
subspace defined by the synchronous state. A key question is thus, can a riddled
basin be physically observed? Investigation along this line [438, 441] has indicated
that riddling is typically destroyed by symmetry-breaking perturbations and is con-
verted into a fractal basin, no matter how small the perturbations are. This has been
called catastrophe of riddling [438]. However, for small perturbations, the resulting
fractal basin may appear similar to a riddled one.

5.8.1 An Example

We consider the following noninvertible two-dimensional map [441]:

⎧
⎨2xn , 0 ≤ x < 1/2,
xn+1 = T (xn ) = (5.30)
⎩2(1 − x ), 1/2 ≤ x ≤ 1,
n
⎧
⎨ pxn yn + ε , | y |< 1
yn+1 = f (xn , yn ) =
⎩λ y , | y |≥ 1,
n

where T (x) is the tent map, p and λ > 1 are parameters, and ε is the symmetry-
breaking parameter. The phase-space region of interest is {0 ≤ x ≤ 1, −∞ < y < ∞}.
For ε = 0, the system possesses a one-dimensional invariant subspace y = 0, which
is caused by the reflection symmetry y → −y. Because λ > 1, the map has two other
attractors: A± at y = ±∞. The chaotic attractor A of the tent map in y = 0 can be the
180 5 Fractal Basin Boundaries

third attractor of the full system if it is transversely stable. Since dyn+1 /dyn = pxn at
y = 0, we have λT = ln p + ln x = ln p − 1. A blowout bifurcation occurs at p0c = e.
As ε is increased from zero, no matter how little, the chaotic attractor of the tent
map is no longer an attractor of the whole system. A catastrophe of riddling occurs
for p < p0c as |ε | is increased from zero, in which the riddled basin of A for ε = 0
is replaced by the fractal basin either of A+ or of A− , depending on the sign of ε .
For p > p0c , the basins of the y = ±∞ attractors are y > 0 and y < 0, respectively if
ε = 0. In this case, as |ε | is increased from zero, a smooth-to-fractal basin boundary
metamorphosis occurs because the two simple basins (y > 0 and y < 0) are replaced
by fractal ones. Because of the simplicity of (5.30), these bifurcations can in fact be
understood analytically to a certain extent.
The replacement of the riddled basin by a fractal one in the presence of a
symmetry-breaking perturbation can be seen qualitatively as follows. As discussed
above, for ε = 0, the basin of the chaotic attractor A is a closed set with positive
measure, which is the complement of two symmetric open dense sets belonging to
the attractors A± , respectively. While initial conditions with y0 > 0 or y0 < 0 can
go to A , they cannot cross the invariant line y = 0. For ε = 0, the dense set of un-
stable periodic orbits originally embedded in A in y = 0 spread out in the vicinity
of y = 0, converting A into a nonattracting chaotic set. Because of this spread of
unstable periodic orbits, a trajectory initiated in y > 0 can penetrate the originally
invariant line y = 0 and go to the y = −∞ attractor, and vice versa. The basin of the
y = −∞ attractor in y > 0 must be open and therefore is fractal.3 The same holds
for the basin of the y = +∞ attractor in y < 0. Thus, as soon as ε becomes nonzero,
the riddled basin of A is destroyed, and simultaneously, two fractal basins arise.
In what follows we analyze how unstable periodic orbits embedded in the original
chaotic attractor in A are perturbed by the symmetry-breaking, based on which we
can establish the existence of open, but not dense, sets that belong to the basins of
the attractors at infinities.
For concreteness, we consider the map (5.30) with ε < 0 around the blowout
bifurcation, i.e., for p less than but close to p0c . Since unstable periodic orbits are
structurally stable, we expect that they shift to a small neighborhood about the orig-
inally invariant subspace y = 0 for ε = 0. For example, the original fixed point
xp = (x p , 0) (a repeller with an unstable direction in both x and y, where x p = 2/3 is
the nontrivial unstable fixed point of the tent map) is shifted to (x p , y p ), where y p is

−|ε |
yp = . (5.31)
1 − px p

3Consider an open neighborhood B of one of the attractors at infinity. Choose a point p in its
basin and evolve it forward in time. Eventually, the resulting trajectory will approach the attractor,
which means that at some finite time, the trajectory will enter B, say at point p . The point p in B
must then have an open neighborhood. Since p is iterated from p in finite time, p must also have
an open neighborhood in the basin.
5.8 Catastrophic Bifurcation of a Riddled Basin 181

For p ≈ p0c , we have px p ≈ 2e/3 > 1 and hence y p > 0. The eigenvalues of the
perturbed fixed point (x p , y p ) are Λx = −2 and Λy = px p > 1. Thus, under the
symmetry-breaking perturbation, the shifted fixed point is still a repeller. Consider
(2) (2)
now the period-2 orbit of the tent map: (x1 = 2/5, 0) and (x2 = 4/5, 0). The eigen-
(2) (2)
values of the twice iterated map at these points are Λx = −4 and Λy = p2 x1 x2 .
The latter is smaller than unity for p < 1.77. The two-cycle is then a saddle in the
full phase space. For ε = 0, the y-coordinates of the orbit become

(2) (2)
(2) −|ε |(px2 + 1) (2) −|ε |(px1 + 1)
y1 = (2) (2)
and y2 = (2) (2)
.
1 − p 2 x1 x2 1 − p 2 x1 x2

Altogether, we observe that (1) an orbit is shifted upward (downward) from y = 0

if it is a repeller (saddle), and (2) the eigenvalues of the orbit remain unchanged.
In general, this is valid for any periodic orbit. Since all repellers are located in y > 0,
a trajectory starting in y < 0 cannot cross y = 0, but since all saddles are located in
y < 0, a trajectory starting in y > 0 can move across the x-axis and go to the y = −∞
attractor. Thus, due to the symmetry-breaking, the y = −∞ attractor acquires a basin
in y > 0.
The picture depicted above, i.e., saddles shifted downward and repellers upward,
is specific to the system (5.30) for the case of ε < 0. For ε > 0, saddles will shift
upward and repellers downward. In general, in two dimensions we expect to observe
saddles and repellers on both sides of the originally invariant subspace when there
is a symmetry-breaking. As a result, there will be fractal basins both above and be-
low the originally invariant subspace. In higher dimensions, unstable periodic orbits
with different unstable dimensions – a type of nonhyperbolicity known as unstable
dimension variability (see also Sect. 4.4.2), – which are originally all located in the
invariant subspace, will be shifted to its neighborhood under a symmetry-breaking
perturbation.
For ε = 0, the “roots” of the open set, i.e., the fixed point (x p , 0) and all its
preimages, are located in the invariant subspace y = 0 and are dense (see Fig. 5.21a).
As we have seen, for ε = 0, these “roots” are shifted and are distributed in the two-
dimensional phase-space region about y = 0, as shown in Fig. 5.21b. Thus, the open
set is no longer dense. The set of initial conditions in the unit square 0 ≤ (x, y) ≤ 1
that go to the y = −∞ attractor is now open. In fact, it is straightforward to see that
the region bounded by the curve xy < |ε |/p in the unit square maps to y < 0 after
one iteration. The basin of the y = −∞ attractor in 0 ≤ (x, y) ≤ 1 thus consists of this
bounded region and all its preimages. The boundaries separating the basins of the
y = ±∞ attractors are fractal. We remark, however, that in this case, the basin in y > 0
of the y = −∞ attractor may appear indistinguishable from that of a riddled basin
because unstable periodic orbits in the originally invariant subspace are perturbed
only slightly.
We thus see that for p < p0c and ε = 0, persistent chaos in the invariant subspace,
together with its riddled basin for ε = 0, is replaced by a chaotic transient and fractal
basins of the attractors at infinities, respectively.
182 5 Fractal Basin Boundaries

a ε=0 attractor

y=0

attractor

b |ε| ∼ 0 attractor

∼ε

attractor

Fig. 5.21 Schematic illustrations of the dynamics of unstable periodic orbits: (a) for ε = 0, y = 0
is invariant and the roots of the tongues are dense in y = 0, creating a riddled basin; (b) for ε = 0,
y = 0 is no longer invariant, the locations of the periodic orbits are shifted about y = 0, and the roots
of the tongues are no longer dense, leading to fractal basins [438] (Copyright 1999, the American
Physical Society)

5.8.2 Critical Behavior and Scaling Laws

We have seen that the presence of a small amount of symmetry-breaking causes a

spread of unstable periodic orbits in a neighborhood of size about ε in the vicinity
of the originally invariant subspace. The dynamics outside the neighborhood can be
approximately described by that of a random process. To see this, we rewrite (for
yn > 0) the y-equation in (5.30), as follows:
− ln yn+1 = − ln yn − ln (pxn + ε /yn ).
Letting Yn ≡ − ln yn , we obtain
Yn+1 = Yn + νn , (5.32)
where νn = − ln (pxn + ε /yn) is a random variable because xn comes from a chaotic
process. For ε ∼ 0, νn is approximately independent of yn most of the time (except
when yn gets close to the original invariant subspace). Equation (5.32) thus describes
a random walk. If the average drift ν ≡ Yn+1 − Yn  = νn  is small, the random-
walk model can be solved using the diffusion approximation, from which various
scaling relations can be derived. Specifically, since ν is small, the evolution of the
probability as a function of discrete time n can be approximated as an evolution in
continuous time t. Let P(Y,t)dY be the probability of finding the walker in the inter-
val [Y,Y + dY ] at time t. Then P(Y,t) obeys the following diffusion equation [237]:
∂P ∂P ∂ 2P
+ν =Q 2, (5.33)
∂t ∂Y ∂Y
5.8 Catastrophic Bifurcation of a Riddled Basin 183

where Q is the diffusion coefficient, defined as

 
2nQ = (Yn − nν )2 . (5.34)

Adopting the above diffusive picture, we see that ν and Q are the two key parame-
ters that determine the dynamics. In fact, the average drift −ν and Q are analogous
respectively to the transverse Lyapunov exponent λT (which can be defined only
when ε = 0) and the diffusion coefficient Q characterizing the degree of the fluc-
tuations of the finite-time transverse Lyapunov exponent, used in Sect. 5.7.3. In the
simple model (5.30), ν < 0 for p > p0c and ν > 0 for p < p0c . Thus, we have

ν ∼ (p0c − p). (5.35)

When there is a symmetry-breaking so that the notions of invariant subspace and

transverse Lyapunov exponent no longer hold, we can still use ν and Q to character-
ize the dynamics in the vicinity of the original invariant subspace. In particular,
regarding the ε -neighborhood of the original invariant subspace as a pseudoin-
variant manifold under a symmetry-breaking, the stability of this manifold can be
quantified by ν and Q. Defining the pseudotransverse Lyapunov exponent

ΛT ≡ − ν , (5.36)

we see that if ΛT > 0 (ν < 0), the pseudoinvariant manifold is transversely unstable
because a trajectory leaves the pseudoinvariant manifold exponentially rapidly.
If, however, ΛT < 0 (ν > 0), a trajectory can spend a long time near the pseu-
doinvariant manifold, although the trajectory will eventually leave it. In this sense,
the manifold is quasistable with respect to transverse perturbations. Introducing
the pseudotransverse Lyapunov exponent, with the parameter Q characterizing its
finite-time fluctuations, thus enables us to quantify the dynamical property of the
pseudoinvariant manifold [441].
A detailed discussion about the validity of the diffusion approximation near the
transition point to a chaotic attractor with a riddled basin, at which the average drift
(or the transverse Lyapunov exponent) is nearly zero, can be found in [565, 567].
Here, because of the symmetry-breaking, the range for the validity of the diffusion
approximation is limited. In particular, we note that a trajectory cannot enter the ε -
neighborhood of the original invariant subspace y = 0. However, for | y |>| ε |, the
trajectory experiences both repulsion from and attraction toward the ε -neighborhood
of y = 0 due to the existence of periodic orbits with different unstable dimensions,
namely, repellers and saddles. If ν ≈ 0, the amount of repulsion is approximately
equal to that of attraction, and hence we expect the diffusion picture to be valid for
| ε |<| y |< 1. This corresponds to the range Y ∈ (0, ε ), where ε = − ln | ε |  1.
For clarity of the presentation, we consider the case ε < 0, so that the symmetry-
breaking-induced basin of the y = −∞ attractor lies in y > 0.
184 5 Fractal Basin Boundaries

We fix a line segment 0 ≤ x ≤ 1 at y = y0 , 0 < y0 1, and uniformly choose

a large number of initial conditions from it, which leads to the following initial
condition for the diffusion equation (5.33):

P(Y, 0) = δ (Y − Y0 ), (5.37)

where Y0 = − ln y0 . Since a trajectory reaching y = 1 quickly goes to the y = +∞

attractor, we have the following absorbing boundary condition at Y = − ln 1 = 0:

P(0,t) = 0. (5.38)

Roughly, a trajectory entering the |ε |-neighborhood of y = 0 is lost to the basin of

the −∞ attractor. A realistic picture is that the Y -location of the absorbing boundary
depends on x. For instance, from the model (5.30), we see that a trajectory goes to
the y = −∞ attractor whenever pxn yn < |ε |. Insofar as xn is not too small, this hap-
pens when yn < |ε |/pxn ∼ |ε |. Thus, as a crude approximation, we impose another
absorbing boundary at ε :
P(ε ,t) = 0. (5.39)
Let F(|ε |, y0 ) be the fraction of initial conditions from the line segment at y0 that
go to the y = −∞ attractor A− . As |ε | is increased, we expect F(|ε |, y0 ) to increase.
For small |ε |, the diffusion equation (5.33) together with the initial and the boundary
conditions (5.37)–(5.39) can be solved to yield the following scaling law [441]:

ν /Q
y0 − 1
F(|ε |, y0 ) = . (5.40)
|ε |ν /Q − 1

ν /Q
If ν > 0, we have |ε |ν /Q − 1 ≈ −1 for |ε | 1 and hence F ≈ 1 − y0 = constant,
a behavior drastically different from that of the symmetric case (5.27) and (5.28).
For ν < 0, |ε |ν /Q − 1 ≈ |ε |ν /Q , and hence for any fixed y0 , we have

F(|ε |, y0 ) ∼ |ε |−ν /Q = |ε ||ν |/Q for ν < 0. (5.41)

We see that in the parameter regime where ν ≈ 0, the fraction remains roughly
constant, regardless of the amount of symmetry-breaking. This also implies the
catastrophic nature of the symmetry-breaking: riddling is destroyed and a fractal
basin component is immediately induced as the system deviates from the symmetric
one, no matter how small the deviation is.
Consider a trajectory originated from the symmetry-breaking-induced fractal
basin of the y = −∞ attractor in y > 0. After it falls into the negative vicinity of
y = 0, it typically experiences a chaotic transient. In particular, if ν < 0 (ΛT > 0),
the transient time is short. If, however, ν > 0 (ΛT < 0), the time can be extraordi-
narily long [441].
To assess the dimensionality of the boundary between the basin of the y = +∞
attractor and the symmetry-breaking-induced basin, we fix a line segment at y = y0 ,
5.8 Catastrophic Bifurcation of a Riddled Basin 185

where | ε | y0 < 1, and examine the set of intersecting points with it of the basin
boundary. Let d0 be the box-counting dimension of this set. We expect 0 < d0 ≤ 1
and the dimension of the boundary to be Db0 = 1 + d0 in the two-dimensional
phase space. For a riddled basin, Db0 is the phase-space dimension. Here, de-
spite the presence of a small amount of symmetry-breaking, Db0 is still close to 2.
Thus, in a practical sense, the symmetry-breaking-induced fractal basin resembles
a riddled one.4
It can be shown, utilizing the solution to the diffusion equation (5.33) [441, 565,
567], that the uncertainty exponent is independent of the symmetry-breaking param-
eter ε and is given by
ν2
α= , (5.42)
4Qλ1
where λ1 is the Lyapunov exponent on the original attractor A in the invariant
subspace. Thus, in the regime where ν ≈ 0 (but ν = 0), so that the diffusion approx-
imation is valid, we expect α ≈ 0 and hence d0 ≈ 1, leading to

ν2
Db0 = 2 − .
4Qλ1

A fractal basin boundary with dimension close to that of the phase space (or a
near-zero uncertainty exponent) means that the uncertainty probability remains
approximately constant, regardless of how accurately we can specify the initial
condition. Thus, realistically, it is impossible to predict, from a given initial con-
dition, the asymptotic attractor. This fundamental obstacle to prediction is common
for riddled basins and persists even when the riddled basin is replaced by a fractal
one due to symmetry-breaking.

4 Since very close to a boundary arises the chaotic saddles’s stable manifold is nearly space-filling,
the set of initial conditions leading to long transients also exhibits riddled-like behavior [834]
Chapter 6
Chaotic Scattering

Scattering is a fundamental tool for probing many physical and chemical processes.
In a scattering experiment, particles are injected into the system and their charac-
teristics after the scattering are recorded, from which many properties of the system
can be revealed. In a general sense, scattering can be defined as a problem of ob-
taining various relations between some output variables characterizing the particles
after the scattering versus some input variables characterizing the particles before
the scattering. The relations are called scattering functions. In a regular scattering
process, the functions are typically smooth, examples of which can be found in
textbooks of classical mechanics. It has been realized, however, that there can be
situations in which a scattering function may contain an uncountably infinite num-
ber of singularities. Near any of the singularities, an arbitrarily small change in the
input variable can cause a large change in the output variable. This is a sensitive
dependence on initial conditions that signifies the appearance of chaos. Scattering
in this case is chaotic.
Dynamically, chaotic scattering is due to the existence of chaotic saddles in the
Hamiltonian phase space. Incoming particles move toward the scattering region in
which a chaotic saddle resides along its stable manifold. They then wander chaot-
ically near the saddle, but only for a finite amount of time, because the saddle is
nonattracting. Finally, they exit the scattering region along the unstable manifold.
The particle trajectories are thus transiently chaotic, with the sensitivity to initial
conditions originating from the chaotic nature of the saddle. It is generally under-
stood now that chaotic scattering is the physical manifestation of transient chaos in
Hamiltonian systems.
After reviewing a broad range of applications of chaotic scattering, we will
explore a system of three-disk billiard scatterers, which is perhaps the simplest
system exhibiting chaotic scattering. A question of interest concerns how chaotic
scattering arises as a system parameter changes. Two basic routes to chaotic scatter-
ing will be analyzed using a system of three and four “soft” circular potential hills.
In Hamiltonian systems, nonhyperbolicity can arise whereby a chaotic saddle is ad-
jacent to stable regions (also called the Kolmogorov–Arnold–Moser (KAM) tori).
A distinct feature of nonhyperbolic chaotic scattering is that particles can spend a
long time in the vicinity of KAM tori, leading to a long-term algebraic decay of
the survival probability, in contrast to the exponential decay for hyperbolic chaotic

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 187
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 6,

c Springer Science+Business Media, LLC 2011
188 6 Chaotic Scattering

scattering, where all periodic orbits are unstable. We will show that the asymptotic
algebraic decay associated with nonhyperbolic chaotic scattering is often preceded
by an intermediate-time exponential decay. The effect of small dissipation will also
be discussed. As an application, the ray dynamics in deformed optical microcavi-
ties will be studied. The properties of scattering cross section in chaotic cases are
discussed in Appendix D.

6.1 Occurrence of Scattering

Chaotic scattering can occur in applications from many disciplines (for reviews, see
[7, 201, 257, 411, 568]). We shall give a few representative examples.
• Astronomy. The three-body problem provides an example of an open Hamiltonian
system, for which Poincaré [608] predicted the existence of complicated trajec-
tories. For a simplified three-body system, Sitnikov proved that the motion of the
third body is generally chaotic [533] before escape takes place. The structure of
the underlying chaotic saddle, however, has been described only recently [417].
Petit and Hénon considered another class of restricted three-body problems: two
small bodies moving about a massive one [326, 594], which is relevant to the
motions of the moons of large planets or of particles in planetary rings (e.g., the
rings of Saturn). They demonstrated that chaotic scattering can occur in the case
of close encounters of the small bodies. Chaotic scattering has also been identi-
fied in a model of the motion of an incoming star toward a binary [97], during
which gravitational energy is effectively converted into kinetic energy. The es-
cape of stars from galaxies has also been investigated [719, 736], as well as the
problem of light rays around black holes [168, 169], the escape from galactic
halos [169], and chaos in cosmological models [536].
• Chemical reaction dynamics. The observable signature of classical chaotic scat-
tering, namely complicated and discontinuous behavior in the scattering func-
tions, had been noticed by Rankin and Miller [622] in their study of atom–diatom
collisions, before the concept of chaos became widespread. In classical models
of chemical reactions [42, 257, 420], the average lifetime of chaotic trajecto-
ries can often be interpreted as the average lifetime of intermediate complexes.
A central problem is the study of complex barriers separating reacting and non-
reacting trajectories, as determined by the stable and the unstable manifolds of
the underlying chaotic saddle [421, 806, 829].
• Transport processes. Diffusion and other transport phenomena can be regarded
as consequences of chaotic scattering, as pointed out first by Gaspard and
Nicolis [258]. This deterministic way of describing transport phenomena in
a single particle picture is based on the idea of an open but finite Hamil-
tonian system with a large size in a certain direction. The phase space is
low-dimensional but of large linear size. The escape rate from some underlying
chaotic saddle can then be calculated, which can be related to the transport coef-
ficients [201, 257, 258, 399, 802]. Consider, for example, the problem of electric
6.1 Occurrence of Scattering 189

conduction where a diffusive current is also present. Microscopically, random

walks take place in the scattering region, which correspond to a scattering pro-
cess subject to a bias, leading to a nonzero average velocity. In the framework of
the escape-rate formalism one finds that the escape rate can be expressed by the
drift coefficient (to which the electric current is proportional), the diffusion coef-
ficient, and the length of the system [778]. The escape is in fact responsible for
two different types of transport, drift and diffusion, but generalizations to other
transport coefficients also hold [201,257,802]. So far it has been understood that
other characteristics of chaotic scattering such as the Lyapunov exponents cannot
be expressed solely by macroscopic parameters. It is the escape rate alone that
has a well-defined thermodynamic limit [804].
• Optical processes in microcavities. A promising area of application of chaotic
scattering is optical processes in microcavities. This has been a subject of in-
tense study due to its potential implications for the design of novel microlasers
[34,176,509,570]. An interesting feature of optical dielectric cavities is that they
can support whispering-gallery (WG) modes, defined as the situation in which
light rays circulate almost tangent to the surface of the cavity via total internal
reflection, suffering minimal losses caused by evanescent leakage and scattering
due to surface roughness [34, 176, 481, 509, 618]. If there are no deformations
in the cavity geometry from the ideal shape of cylinder or sphere, in a prac-
tical sense light can be trapped in the cavity indefinitely. In order for a cavity
to be useful, the directionality of laser emission is important, which cannot be
achieved if the cavity is perfectly circular. A simple solution is to apply smooth
deformations from the circular symmetry, resulting in so-called asymmetric res-
onant cavities [265, 552, 553]. Due to symmetry-breaking, the condition of total
internal reflection cannot be satisfied indefinitely for classical light rays, the es-
caping dynamics of which can in fact be cast as a chaotic-scattering problem.
This classical approach has proven to be quite useful, particularly considering
the fact that it is difficult to study cavities with large deformations, since the
modes of highly deformed cavities are not perturbatively related to those of the
circular cavities [265, 519, 552–554]. We will discuss the classical approach in
Sect. 6.7.
• Hydrodynamic processes. In two-dimensional incompressible fluid flows, the ad-
vective dynamics are Hamiltonian. If the geometry of the flow is open, such as the
setting where the flow is around an obstacle, the dynamics of particles advected
by the flow can be regarded as those from a scattering problem where particles
coming toward the obstacle are “scattered” away, as illustrated by the experiment
in Fig. 1.19. Indeed, chaotic scattering can occur commonly in the advective dy-
namics of passive particles (Lagrangian chaos). A detailed treatment of this topic
will be presented in Chap. 10.
• Electronic transport in nanostructures. A fundamental structure in modern semi-
conductor devices is mesoscopic junction systems for electronic transport, such
as quantum dots [238]. At the mesoscopic scale, the wavelength of the electrons
is small but cannot be neglected, so both the classical and quantum dynam-
ics are relevant. In the classical picture, electrons can be regarded as point
190 6 Chaotic Scattering

particles moving through the junction, which is equivalent to an open billiard

system in which chaotic scattering can occur generically. At the quantum-
mechanical level, the most studied phenomenon is conductance fluctuations,
which are observed commonly in quantum dots. To explain the fluctuations, it is
necessary to consider the quantum manifestations of classical chaotic scattering.
In particular, the elements of the quantum scattering matrix exhibit fluctuations
in their dependence on the energy, and the origin of the quantum fluctuations
can be attributed to the occurrence of chaotic scattering when the system is in
the classical regime, as the experiment in Fig. 1.20 illustrates. Quantum chaotic
scattering will be treated in Chap. 7.
The problem of chaotic scattering is related to that of chaotic transients in
dissipative dynamical systems. In particular, since the interest is in the different
final outcomes from the scattering process, one can define different exit basins,
each being a region of the phase space where particles exhibit qualitatively sim-
ilar asymptotic behavior. The exit basin associated with exit route n is the set of
initial conditions that lead to an escape from the scattering region via the route. Al-
though attractors do not exist, the exit basins play similar roles to those of basins
of attractions. Analysis tools and methods suitable for basin boundaries can there-
fore be applied to chaotic scattering [164]. For example, the portion of uncertain
initial regions can be determined and the uncertainty exponent (Sect. 5.3) can be
calculated, which determines the box-counting dimensions associated with the un-
derlying chaotic saddle. If the number of exit basins exceeds two, the basins often
exhibit the Wada property [6,164,613,614,668]. In addition, leaking (Sect. 2.7) any
closed Hamiltonian systems leads effectively to a scattering problem.

6.2 A Paradigmatic Example of Chaotic Scattering

A simple model of chaotic scattering is the three-disk system introduced by

Eckhardt, Gaspard, and Rice [210, 260] (for a review see [257]). The system con-
sists of three hard circular disks of identical radii R in the two-dimensional plane,
as shown in Fig. 6.1, and can also be considered a “classical pinball machine.”
The distances
√ between the individual disks
√ are all a, and the disks are located
√ at
x = −a 3/6, y = a/2 (disk 1), x = a 3/3, y = 0 (disk 2), and x = −a 3/6,
y = −a/2 (disk 3). Since the disks are hard, a particle will bounce off a disk upon
collision. The reflection and incident angles with respect to an individual disk are
equal for a bounce, as shown in Fig. 6.1. As such, a particle coming from far away
to the left can experience bounces between the hard disks (the scattering region)
and, after spending a finite amount of time there, exits to infinity with an angle φ ,
where φ is defined counterclockwise with respect to the x-axis, as shown in Fig. 6.2.
Since energy is conserved, the velocity of the particle is constant and chosen to be
unity: v = 1. Let the incident particle trajectory be parallel to the x-axis with an
impact parameter b. The deflection angle with which the particle exits the scattering
region depends on the value of the impact parameter: φ = φ (b). Moreover, the time
6.2 A Paradigmatic Example of Chaotic Scattering 191

Fig. 6.1 The three-disk scattering system [838] (copyright 1995, the American Institute of
Physics)

Fig. 6.2 Two particle trajectories for a = 2.5 and R = 1 with impact parameters (a) b1 = 0.33005
and (b) b2 = 0.33010 [838] (copyright 1995, the American Institute of Physics)

T (b) that the particle spends in a region that contains the scattering region before
exiting, referred to as the delay time, also depends on the impact parameter b. Both
the deflection angle and the delay time are examples of scattering functions.
Figure 6.2a,b show two particle trajectories with close impact parameters:
|b2 − b1| = 5 × 10−4. However, the difference in the deflection angles for the two
trajectories is of order π . Such a large difference is also reflected in the delay time,
the time by which the particles stay within, say, distance d = 10( R, a) of the
origin. Since particle 1 experiences substantially more bounces in the scattering
region than particle 2, we have T (b1 ) > T (b2 ). Thus, about the impact parameter
b ≈ b1 (or b2 ), a small difference in the initial condition causes a large difference in
the outcome of the trajectories, the hallmark of chaos.
192 6 Chaotic Scattering

Fig. 6.3 The deflection function (a) and the delay-time function (b) for a/R = 2.5 [838] (copyright
1995, the American Institute of Physics)

The sensitive dependence on initial conditions occurs commonly at many impact

parameters (in fact, at an infinite number of them). This can be seen in Fig. 6.3a,b,
plots of the deflection function φ (b) and the delay-time function T (b), respectively.
The delay-time function is analogous to the lifetime function in the dissipative cases,
as shown, e.g., in Fig. 1.5. It is apparent that these plots contain both smooth and
wild oscillating parts. The mixture of smooth and wild oscillating parts in φ (b) and
T (b) repeats itself on smaller scales. In fact, there exists a Cantor set of an un-
countably infinite number of b values at which φ (b) and T (b) have singularities. A
b-interval between two neighboring singularities is called the interval of continuity.
An efficient method for computing the box-counting dimension ds of the set of
singularities, which is particularly convenient for chaotic scattering, is the uncer-
tainty algorithm described in Sect. 5.3. For an arbitrary location on the left of the
scattering region (e.g., x0 = −10 in Fig. 6.1), an initial condition (impact parameter)
b is chosen randomly. A small perturbation is applied, yielding a nearby initial con-
dition b + ε . The scattering trajectories originated from these two initial conditions
are computed. If the numbers of bounces in the scattering region experienced by
the two particles are the same, the first initial condition is called “certain” against
small perturbation. This will occur if the initial condition is chosen in the range
where the scattering functions are smooth. If the initial conditions lead to trajecto-
ries with different numbers of bounces in the scattering region, they are “uncertain”
with respect to the perturbation ε . The fraction of uncertain initial conditions f (ε )
can then be computed from many randomly chosen initial-condition pairs. As ε is
decreased, f (ε ) will decrease as ε α (see (5.7)), where α is the uncertainty exponent
and ds = 1 − α is the box-counting dimension of the set of singularities. With the
parameters of Fig. 6.3, the uncertainty exponent is estimated to be α ≈ 0.39, indi-
cating ds ≈ 0.61. The procedure described above is effective when the value of ds
is not close to zero. For situations in which the dimension value is arbitrarily small,
special care needs to be exercised in implementing the uncertainty algorithm [162].
That a chaotic saddle contains an infinite number of unstable periodic orbits can
be seen physically using the three-disk system, where periodic orbits are the particle
trajectories that repeat themselves after a certain number of bounces between the
hard disks, and hence these trajectories are permanently trapped in the scattering
6.2 A Paradigmatic Example of Chaotic Scattering 193

Fig. 6.4 Three unstable

period-2 orbits, one unstable
period-3 orbit, and their
symbolic representations
[838] (copyright 1995, the
American Institute of
Physics)

region. Figure 6.4 shows some examples of periodic orbits. The orbits are all
unstable (hyperbolic): typical perturbations to the orbit, no matter how small, will
result in deviations away from the orbit and exit of the particle from the scattering
region. For the three-disk system, every periodic orbit can be represented by a sym-
bolic sequence. For example, each bounce can be denoted by the label (1, 2, or 3) of
the disk on which it occurs, and a particle trajectory can be encoded as
S1 S2 S3 · · · Sn−1 Sn Sn+1 · · · , (6.1)
where Si = 1, 2, 3 depending on whether the particle bounces off disk 1, 2, or 3 at
the ith bounce, respectively. Since a typical scattering trajectory eventually exits
the scattering region, the corresponding symbolic representation has only a finite
number of symbols. Periodic orbits are represented by a string of infinite symbols
consisting of repetitions of a finite number of symbols. As the period m increases,
the number Nm of periodic orbits increases exponentially, since the number of pos-
sibilities of representing a string of length m increases exponentially as a function
of m. How Nm grows can be assessed by noting that, since two or more identical
symbols in a sequence are forbidden (a particle cannot hit the same disk two or more
times consecutively), the number of nonrepetitive symbolic sequences of length m
is 3 × 2m−1 . The definitions (1.25), (1.26) of the topological entropy K0 lead to the
value K0 = ln 2. The scattering process is thus equivalent to a Bernoulli process on
two symbols.
Besides periodic orbits, there are also trajectories that wander in the scattering
region forever but never repeat themselves, as can be seen again via the symbolic
representation (6.1). There is in fact an uncountably infinite number of such ape-
riodic or chaotic trajectories. These, together with the infinite number of unstable
periodic orbits, make up the chaotic saddle. Since all periodic orbits are hyperbolic,
the chaotic saddle and the scattering process are also hyperbolic.1 To visualize the
chaotic saddle, one can define a discrete-time map from the scattering dynamics.
For example, the incident angle αn (see Fig. 6.1) at bounce n can be taken as

1 In hyperbolic chaotic scattering, all periodic orbits are unstable, while in nonhyperbolic chaotic

scattering, there are stable or neutrally stable periodic orbits and quasiperiodic orbits, which will
be treated in Sects. 6.3 and 6.4.
194 6 Chaotic Scattering

Fig. 6.5 Chaotic saddle in

the phase space of the
three-disk scattering system
for a/R = 2.5

a dynamical variable, in terms of the angle φn of the incident trajectory defined

counterclockwise with respect to the x-axis. (For the last bounce N, φN+1 is identi-
cal to the deflection angle φ of the trajectory plotted in Fig. 6.3.) Figure 6.5 shows
the chaotic saddle in the two-dimensional phase space (φn , sin αn ).
Saddles in Hamiltonian systems have a special property: stable and unstable di-
rections are equivalent, (2.82),
Ds,i = Du,i , (6.2)

and therefore in the two-dimensional phase spaces the partial dimensions coincide:

(1) (2)
Di = Di = Di /2, (6.3)

where i = 0, 1 and D0 and D1 denote the box-counting and the information dimen-
sions, respectively, of the entire saddle. This symmetry is visible in Fig. 6.5. The
scattering singularities correspond to intersections of the stable manifold with a line
of initial conditions. The dimension ds of the set of scattering singularities is then the
same as the partial box-counting dimension along the stable or unstable direction:

(1)
ds = Ds,0 − 1 = D0 , (6.4)

which implies that the box-counting dimension D0 = 2ds of the chaotic saddle can
be expressed by means of the uncertainty exponent α = 1−ds as D0 = 2(1− α ). The
escape rate κ of the saddle can be obtained by launching many particles with dif-
ferent impact parameters toward the scattering region and examining the number of
particles remaining in the region with time. The exponential decay can be followed
both in continuous time and in the discrete time of the Poincaré map. For exam-
ple, for a/R = 2.5, one obtains in continuous time κ ≈ 0.74. The average Lyapunov
6.3 Transitions to Chaotic Scattering 195

exponent can be estimated from the general relation (2.76) and the fact that the
box-counting dimension is close to the information dimension, which leads to
κ κ
λ1 ≈ = . (6.5)
1 − ds α
This relation is valid both in continuous and in discrete time. For example, for
κ ≈ 0.74 and α ≈ 0.39, the dimensionless continuous-time Lyapunov exponent is
λ1 ≈ 1.9.
In the special arrangement of infinitely many aligned hard disks of decreasing
radius, the scattering function was found [364] to be nearly as irregular as in chaotic
cases, although the system is integrable. There is an infinite number of periodic or-
bits, but there are no higher-order cycles and heteroclinic connections. This example
implies that in order to establish that a scattering process is chaotic, it is necessary
to demonstrate the positivity of the topological entropy (or the Lyapunov exponent
of the chaotic set).

6.3 Transitions to Chaotic Scattering

The hard-disk scattering system discussed in Sect. 6.2 exhibits hyperbolic chaotic
scattering. In Hamiltonian systems, hyperbolicity of the chaotic saddle occurs only
if the dynamics can be described by a complete set of symbolic sequences (e.g., in
the Hénon map for a > ah (b); cf. Fig. 3.9). Hyperbolic scattering can therefore be
called in such a case fully developed chaotic scattering [74, 194]. A basic question
is how hyperbolic chaotic scattering arises as a system parameter changes. This
transition issue can be addressed by considering scattering in Hamiltonian systems
of smooth potentials V (r). The equations of motion for a particle of unit mass are
given by
r̈ = −gradV (r). (6.6)
This is a time-independent Hamiltonian problem, and hence the particle’s total en-
ergy E is conserved. The transition scenario is schematically illustrated in Fig. 6.6,
where E is regarded as a bifurcation parameter, and Ec is a critical parameter value.
For E > Ec , the scattering dynamics is either regular or chaotic but nonhyperbolic
(to be discussed in Sect. 6.4). Hyperbolic chaotic scattering occurs for E < Ec . Here
we address possible types of bifurcations to hyperbolic chaotic scattering.
In two-degree-of-freedom time-independent Hamiltonian systems, there are two
known routes to hyperbolic chaotic scattering. The first is the abrupt bifurcation in

Fig. 6.6 Schematic

illustration of bifurcation to
hyperbolic chaotic scattering
196 6 Chaotic Scattering

which the scattering is regular for E > Ec , but as soon as E is decreased through Ec ,
hyperbolic chaotic scattering occurs [74, 436]. Accordingly, the number of unstable
periodic orbits associated with the scattering dynamics is zero or finite for E > Ec
(corresponding to zero topological entropy), but immediately becomes infinite for
E < Ec (K0 > 0). In the second route, the creation of periodic orbits is not abrupt
but through the saddle-center and period-doubling bifurcations.2 The existence of
the center after a saddle-center bifurcation implies that there are surrounding KAM
tori. In this case, when chaotic scattering appears, it is nonhyperbolic. (At a lower
energy, Ec , it might, however, become hyperbolic.)

6.3.1 Scattering from a Single Hill

It is useful to review the dynamics of scattering from a single potential hill. Consider
the localized potential:
⎧ 
⎨E 1 − x R+y
2 2
, x2 + y2 ≤ R 2 ,
m 2
V (x, y) = (6.7)
⎩0, x2 + y2 > R 2 .

The quadratic form is prototypical for any circularly symmetric potential around its
hilltop of height Em . The equation of motion (6.6) within the disk of radius R is
therefore r̈ = s2 r with √
2Em
s= , (6.8)
R
which has the solution
sr0 + v0 st sr0 − v0 −st
r(t) = e + e (6.9)
2s 2s

under the initial condition r(0) = r0 and v(0) = v0 .

Consider now the case in which a particle approaches the scattering region
(circle of radius R) from a large distance along the x-axis. The particle speed
is v0 and the impact parameter is b. The incident angle θ defined by √ Fig. 6.7 is
zero. Upon entering the scattering region, the position is r0 = (− R2 − b2, b).
For sufficiently large R (small b), the particle spends a long time within the scat-
tering region, so that st  1. The second term in (6.9) is then negligibly small
when the particle√ leaves the circle. The velocity components at the time of exit
are vx = [(v0 − s R2 − b2)/2] exp (st) and vy = (sb/2) exp(st). The angle that the
scattered trajectory makes with the x-axis is tan φ = vy /vx . Since E = v20 /2, the
deflection function is, using (6.8),

2 A saddle-center bifurcation in a Hamiltonian system is equivalent to a saddle-node bifurcation

in dissipative systems that is responsible, for instance, for the creation of periodic windows in a
chaotic parameter regime.
6.3 Transitions to Chaotic Scattering 197

Fig. 6.7 Schematic

illustration of a scattering
trajectory. The incoming
particle before the scattering
is characterized by two initial
variables: the impact
parameter b and the incident
angle θ . The outgoing
particle after the scattering is
characterized by the
deflection angle φ

b/R
tan φ (b) =   . (6.10)
E/Em − 1 − (b/R)2

The maximal deflection angle φm occurs at b = R (E − Em )/E and is given by

tan φm = Em /(E − Em ), E ≥ Em . (6.11)

Above the hilltop energy, i.e., for E > Em , the maximum deflection angle is
φm < π /2. As E → Em from above, the maximum angle approaches 90◦ . Below
the hilltop energy, the deflection angle can take on any value: 0 ≤ φ ≤ 180◦. (For
b → 0 backscattering occurs, i.e., φ → 180◦.)

6.3.2 Abrupt Bifurcation to Chaotic Scattering

6.3.2.1 Basic Phenomenon

A prototypical example to illustrate the abrupt route to chaotic scattering is the pla-
nar potential scattering system introduced in [74]. The potential is given by

V (x, y) = x2 y2 exp [−(x2 + y2 )], (6.12)

as shown in Fig. 6.8, where there are four potential hills located at (x, y) = (±1, ±1)
of identical height Em = e−2 ≈ 0.13533. Particle motion is governed by (6.6)
with r = (x, y). Because of the conservation of energy, there are only three
independent first-order differential equations, so the phase-space dimension is
three. Now suppose that particles of energy E and impact parameter b are
launched toward
 the scattering region where the value of V (x, y) is appreciable
(e.g., R = x2 + y2 ≤ 5.0). Since the maximum of the potential hills is locally
quadratic, the results from Sect. 6.3.1 can be applied. For E > Em , particle trajecto-
ries can reach any point of the plane. While the trajectories are typically deflected
198 6 Chaotic Scattering

Fig. 6.8 Surface plot of the

four-hill potential V (x, y)
defined by (6.12) [74] (with
kind permission from
Elsevier Science)

by the potential, the deflection angle from a single potential hill is less than 90◦ ,
(6.11), and hence particles cannot bounce back and forth between any pair of po-
tential hills. In this case, no periodic orbit can be formed. If, however, the particle
energy is smaller than the height of the potential hill, E < Em , trajectories cannot
penetrate the region of the hill where the values of the potential function are larger
than E. As a result, the deflection angle from any single potential hill can reach
180◦, making possible bounces among the hills and consequently periodic motions
in the scattering region. In this case, qualitatively, the system is similar to the
hard-disk scattering system discussed in Sect. 6.2, and we expect hyperbolic chaotic
scattering to occur for E < Em . Indeed, the deflection and the delay-time functions
exhibit features similar to those in Fig. 6.3a,b. Bifurcation to chaotic scattering is
abrupt in the sense that chaos arises immediately as the particle energy is decreased
through the critical value Ec = Em .
For E slightly below Em , trajectories can be confined in the scattering region,
experiencing backscattering, or bouncing among the potential hills. Unstable peri-
odic orbits can then exist, as shown in Fig. 6.9, where several orbits are indicated
schematically. In addition, an uncountably infinite number of aperiodic orbits exist
that traverse among the hills in any order. These periodic and aperiodic orbits can
be determined by utilizing symbolic dynamics, as in the hard-disk scattering system
treated in Sect. 6.2. In particular, four distinct symbols, say Si = 0, 1, 2, and 3, can be
associated with each hill, as shown in Fig. 6.9. Since no symbol can repeat consec-
utively in any sequence, the number Nm of distinct periodic sequences of period m
increases exponentially with m, Nm ∼ 3m , implying that no periodic orbit is missing.
The topological entropy is then K0 = ln 3 > 0, ensuring the existence of a chaotic
saddle. Since every orbit is unstable, scattering is hyperbolic. Explicit numerical
computations using this definition confirm the discontinuous change is the topolog-
ical entropy from 0 to ln 3 when the system exhibits an abrupt bifurcation to chaotic
scattering [191].
6.3 Transitions to Chaotic Scattering 199

Fig. 6.9 Representative

unstable periodic orbits of
low periods in the four-hill 0 3
scattering system

γ4
γ3
γ2
γ1

1 2

γ1

Fig. 6.10 Scattering region crossing orbit

bounded by four distinct
periodic orbits of period 2
and a crossing orbit. The
existence of such an orbit
implies a heteroclinic tangle 0 3
of intersections between the
stable and the unstable
manifolds of the
corresponding pair of R
periodic orbits that are not
located at opposite sides of
the bounding region

1 2

Here we briefly describe the rigorous results [138, 139, 643] that imply the
existence of a chaotic saddle for E < Em . Mathematically, for E < Em , the topol-
ogy of the energy surface is characteristically different from that for E > Em .
Specifically, for E < Em , there are four regions in the (x, y)-plane, determined by
V (x, y) > E, into which classical trajectories are forbidden. There are then four
unstable periodic orbits of period 2 along segments of the straight lines x = ±1
and y = ±1 (the four γ1 orbits in Fig. 6.9), which are the boundaries of the scat-
tering region R, as shown in Fig. 6.10. Orbits leaving R never return. The analyses
in [138, 643] indicate that if there is a trajectory that enters R through one of the
bounding periodic orbits and leaves through another, but nonopposite, bounding
periodic orbit, the corresponding two bounding periodic orbits have a heteroclinic
tangle of intersections of their stable and unstable manifolds. Such a trajectory is
called a “crossing-orbit,” as schematically illustrated in Fig. 6.10. By symmetry, the
200 6 Chaotic Scattering

Fig. 6.11 Stable (left) and unstable (right) manifolds of the chaotic saddle for E = 0.260Em [74]
(with kind permission from Elsevier Science)

existence of a crossing orbit for any pair of nonopposite sides implies the existence
of crossing orbits for all four pairs of nonopposite sides. A heteroclinic tangle of
stable and unstable manifolds implies the existence of a chaotic saddle in R. The
problem of demonstrating hyperbolic chaotic scattering in this case thus reduces to
the straightforward task of finding crossing orbits, which are indeed found in nu-
merical experiments [74].
The saddle and its stable and unstable manifolds can be visualized conveniently
on a proper Poincaré surface of section. Recall that the phase-space dimension is
three, so any three independent variables can be utilized to represent the dynamics.
One can choose, for instance, (x, θ , y) (for the choice of θ , see Fig. 6.7). Choosing
the surface of section defined by y = 0 and computing the intersecting points of
trajectories with the two-dimensional plane (x, θ ) for ẋ > 0, the authors in [74] ob-
tained the stable and unstable manifolds of the chaotic saddle, as shown in Fig. 6.11
for E = 0.26Em . The intersecting points between the stable and unstable manifolds
constitute the chaotic saddle itself, which can also be obtained via the PIM-triple
method (Sect. 1.2.2.4), as shown in Fig. 6.12. Note that the x, y coordinates of points
on the saddle are contained completely in the region R.

6.3.2.2 Scaling of Dynamical Invariants with Energy

While the transition to chaotic scattering is abrupt, the dimension of the set of singu-
larities in the scattering function changes continuously from zero as E is decreased
from Em . To obtain the scaling of the box-counting dimension with the energy dif-
ference (E − Em ), an insightful approach is to construct a model scattering system
with self-similarity in the underlying fractal set. Consider the system of three sym-
metric hills of identical height Em , as shown in Fig. 6.13a. In order for the potential
hills to be equivalent to hard disks, the sizes of the cross sections of the hills should
6.3 Transitions to Chaotic Scattering 201

Fig. 6.12 For E = 0.260Em ,

the chaotic saddle responsible
for the observed hyperbolic
chaotic scattering [74] (with
kind permission from
Elsevier Science)

Fig. 6.13 For a scattering

system of three identical hills a
(a), the corresponding
cantor-set construction (b) 1

s1

s
2 3

s2

s
b
s1 s2

be small. This stipulates that the energy be close to Em . In the “hard-disk” range of
each hill, the deflection angle satisfies | tan φ | ∼ O(1). It can be seen from (6.10) that
this holds in the range b ∼ (Em − E). For the potential configuration in Fig. 6.13a,
deflection angles near 180◦ and 120◦ are relevant because the corresponding trajec-
tories are reflected by the hills and are likely to stay in the scattering region for a
relatively long time. Effectively, a reflection occurs when a particle is within a dis-
tance of order (Em − E) of any peak. The potential hills in Fig. 6.13a can thus be
considered hard disks of radii on the order of (Em − E).
202 6 Chaotic Scattering

Suppose now that particles are launched along the x-direction from a vertical
line segment located between hills 2 and 3. Let the interval s correspond to initial
conditions that reflect from hill 3. The subintervals s1 ⊂ s and s2 ⊂ s correspond to
initial conditions that reflect from hill 3 and then from hills 1 and 2, respectively.
Figure 6.13b shows the first stage of the construction of a Cantor set. The sizes of
interval s and subintervals s1,2 satisfy

s ∼ Em − E, and s1,2 ∼ (Em − E)2 . (6.13)

Since the relation of interest is how the box-counting dimension of the Cantor set
scales with the energy, the Cantor set can be regarded as if it were exactly self-
similar. For such a Cantor set whose first stage of construction is as indicated in
Fig. 6.13b, the dimension ds of the set of singularities is determined by the following
transcendental equation, (1.20):
 s  ds  s  ds
1 2
+ = 1, (6.14)
s s

which yields, as a consequence of (6.13),

(Em − E)ds ∼ constant.

The scaling law of the box-counting dimension near the abrupt bifurcation to chaotic
scattering is then given by

D0 1
ds = ∼ . (6.15)
2 ln [(Em − E)−1 ]

It can be seen that ds → 0 as E → Em , which is consistent with the physical picture

that the scattering is regular for E > Ec = Em .
The dynamics on the interval s is analogous to that of the open one-dimensional
tent map treated in Chap. 2. The escape rate and the Lyapunov exponent are given
by κ = − ln (s1 /s + s2/s) and λ1 = −[s1 ln (s1 /s)+ s2 ln (s2 /s)]/(s1 + s2 ), (2.28) and
(2.30), respectively. Using (6.13), one can find that [768]

κ = − ln (Em − E) + C0 , λ1 = − ln (Em − E) + C1 , (6.16)

where C0 and C1 are constants. The divergence at Em implies that all bounded orbits
are infinitely unstable when they appear at Ec = Em , and consequently, the escape
from the entire saddle is infinitely fast. The orbits become less unstable when the
energy decreases, but the saddle remains ratified as reflected by the scaling of the
dimension (6.15). Deviations from the circular symmetry have also been considered,
such as potential hills with elliptical tops [772]. Abrupt bifurcation thus appears to
be a generic route to chaotic scattering.
6.3 Transitions to Chaotic Scattering 203

6.3.3 Saddle-Center Bifurcation to Chaotic Scattering

Besides the abrupt bifurcation route, in two-degree-of-freedom Hamiltonian

systems there is another generic route to chaotic scattering, which is similar to
the period-doubling route to chaos in dissipative dynamical systems [235, 236]. In
this route [74, 194], when scattering first becomes chaotic, it is typically nonhy-
perbolic in the sense that there are coexisting KAM tori in the phase space. As the
bifurcation parameter is varied further, hyperbolic chaotic scattering arises.
A paradigm for understanding the saddle-center route to chaotic scattering, rela-
tive to the abrupt bifurcation route, is the general scattering configuration consisting
of three asymmetric potential hills [74, 194] of heights Em1 , Em2 , and Em3 , where
Em1 ≤ Em2 ≤ Em3 . Depending on the relative positions of the three hills, there
are two situations of interest, as shown in Fig. 6.14, where the distances between
the centers of the hills are much larger than the effective radii of the hills. The hills
are denoted by dots located at 1, 2, and 3. In case (a), the hill of the lowest maxi-
mal energy (hill 1) is outside the circle whose diameter is the line joining the two
hills with larger maximal energy values (hills 2 and 3). In case (b), hill 1 is inside
the circle.
Consider the first case (a). In order to have a trapped orbit, the minimally required
deflection angle φm∗ by hill 1 for particles coming from hill 2 or hill 3 should be
greater than 90◦ . If the particle energy is greater than Em1 but smaller than Em2 , this
condition cannot be satisfied, and hence there are no bounded orbits reflecting from
hill 1. In this case, the only allowed periodic orbit is the one bouncing back and forth
between hills 2 and 3. Scattering is thus not chaotic for E > Em1 . As E is lowered
through Em1 , the deflection angle at hill 1 can immediately reach 180◦, as seen in
Sect. 6.3.1. Consequently, as E is decreased through Em1 , unstable periodic orbits
of all possible sequences of visits to all three hills can exist, which is analogous to
the three hard-disk scattering system treated in Sect. 6.2. There is then an abrupt
bifurcation to chaotic scattering.
Now consider case (b) with energy in the range Em1 < E < Em2 ≤ Em3 .
The deflection angles by hills 2 and 3 can be as high as 180◦ . Let φm1 (E) be

1 φm*

b
1 φm*

2 3 2 3

Fig. 6.14 Scattering configurations of three nonidentical potential hills: the abrupt-bifurcation
route occurs in case (a), and the saddle-center-bifurcation route can occur in case (b). The heights
of the hills are ordered as Em1 ≤ Em2 ≤ Em3 , and φm∗ denotes the deflection angle required by an
orbit coming from hill 2 to be reflected by hill 1 toward hill 3
204 6 Chaotic Scattering

the maximum deflection angle by hill 1, and let Em∗ denote the energy at which the
maximum deflection angle is φm∗ . From Sect. 6.3.1, we know that φm1 (Em2 ) < 90◦
and φm1 (Em1 ) = 90◦ . Since hill 1 is now inside the circle, the maximal deflection
angle required for an orbit to be trapped is φm∗ < 90◦ . There are then two situ-
ations: φm1 (Em2 ) > φm∗ and φm1 (Em2 ) < φm∗ . As the maximum deflection angle
increases with decreasing energy, (6.11), these situations correspond to Em∗ > Em2
and Em∗ < Em2 , respectively.
In the first case, the maximum deflection angle from hill 1 is greater than φm∗
because for E < Em2 , we have φm1 (E) > φm1 (Em2 ) > φm∗ . Thus, as in case (a) in
Fig. 6.14, periodic orbits of all possible sequences of visits to three hills can exist.
There is then an abrupt bifurcation to chaotic scattering as E is decreased through
Ec = Em2 .
The characteristically distinct route to chaotic scattering occurs when the system
configuration is such that Em∗ < Em2 , i.e., φm1 (Em2 ) < φm∗ . In this case, for E smaller
than but close to Em2 , there is no chaos. As E is decreased, φm1 (E) will increase
until E = Em∗ , where the maximum deflection angle is equal to φm∗ . For smaller
values of E, Em1 < E < Em∗ , there can then be orbits traversing back and forth
between hills 2 and 3 through hill 1 with symbol sequences such as . . . 2131213 . . .
that do not contain subsequences 23 and 32, as shown in Fig. 6.15. This situation
is completely different from abrupt bifurcation, where periodic orbits of all orders
of three symbols are possible. Here, as the energy is decreased from Em∗ , there will
eventually be unstable periodic orbits made up of all possible combinations of the
symbols, leading to hyperbolic chaotic scattering at some energy E = Ec < Em∗ .
As E is decreased from Em∗ to Ec , an infinite number of unstable periodic or-
bits is created. In contrast to the case of abrupt bifurcation, where the infinite
set of unstable periodic orbits is created as a result of a sudden change in the
topology of the energy surface, here the creation of the periodic orbits is a con-
tinuous and smooth process, involving no change in the topology of the energy
surface. In smooth Hamiltonian systems of two degrees of freedom, a generic
mechanism for generating periodic orbits is saddle-center bifurcations.3 For saddle-
center bifurcations, the centers are stable elliptic orbits surrounded by KAM tori.

Fig. 6.15 Two distinct types

of periodic orbits, and the
corresponding symbol
sequences, in the
saddle-center route to chaotic
scattering in the case
Em1 < E ≤ Em∗ < Em2 , where
φm1 (Em∗ ) = φm∗ < 90◦

3 See footnote 2.
6.3 Transitions to Chaotic Scattering 205

Fig. 6.16 Schematic

illustration of the scattering
system consisting of three
localized quadratic potential
hills, as in (6.17), situated at
(xi , yi )

The centers can undergo period-doubling cascades. Typically, there can be KAM
tori coexisting with chaotic saddles, leading to nonhyperbolic chaotic scattering.
When the processes of saddle-center bifurcations followed by period-doubling cas-
cades are completed, all elliptic orbits are destroyed and all periodic orbits are
unstable, giving rise eventually to hyperbolic chaotic scattering.
The particular way that the infinite number of unstable periodic orbits for hy-
perbolic chaotic scattering are created was investigated by Ding et al. [194]. For
numerical convenience, they chose a combination of localized potentials. The po-
tential function of the system is

V (x, y) = V1 (x, y) + V2(x, y) + V3 (x, y), (6.17)

where each Vi is of the form (6.7) with hill height Emi , radius Ri , and centered at
(xi , yi ). Potentials V2 and V3 were chosen to be identical and arranged in a symmetric
fashion, as shown in Fig. 6.16. The particle energy E was fixed, and the height Em1
of the potential hill 1 was chosen to be the bifurcation parameter. The heights of hills
2 and 3 satisfy Em2 = Em3  E. Decreasing E, as we have discussed in the preceding
paragraph, is equivalent to increasing Em1 . Figure 6.17a–c show schematically the
possible periodic orbits of the system together with their stable and unstable mani-
folds respectively before, at, and after the saddle-center bifurcation at Em1 = Em∗ .
Before the saddle-center bifurcation (Fig. 6.17a), there is only one periodic orbit
(denoted by π1 ), the one bouncing back and forth between hills 2 and 3. At the
bifurcation (Fig. 6.17b), a saddle (denoted by π3 ) and a center (denoted by π2 ) are
born, and they coincide at the bifurcation. After the bifurcation, the saddle and the
center are separated, as shown in Fig. 6.17c. Because of the symmetry of the system
with respect to x = 0, it is convenient to choose the Poincaré surface of section
to be x = 0. The dynamical variables on the Poincaré section can be chosen to be
(vy , y). The KAM tori surrounding the center π2 can be obtained numerically [194],
as shown in Fig. 6.18. The period-doubling bifurcation of the center can also be
computed, as shown in Fig. 6.19, where the center π2 is already unstable and a new
center of period 2 exists.
As Em1 is increased further, a cascade of period-doubling bifurcations occurs,
which becomes completed at a large enough value Em1 = Ec . Here, the stable and
206 6 Chaotic Scattering

Fig. 6.17 Schematic illustration of periodic orbits in the configuration space (left) and in the phase
space (right), together with their stable and unstable manifolds, before (a), at (b), and after (c) the
saddle-center bifurcation [194] (copyright 1990 by the American Physical Society)

Fig. 6.18 KAM tori surrounding the center π2 immediately after the saddle-center bifurcation that
occurs at Em1 = Em∗ = 0.187. Here Em1 = 0.195, E = 1, and other parameters are x1 = 0, y1 = 2.2,
x2 = −x3 = 6, y2 = y3 = 0, R1 = 2, R2 = R3 = 3, and Em2 = Em3 = 10 [194] (copyright 1990, the
American Physical Society)

the unstable manifolds of the unstable periodic orbits π1 , π2 , and π3 can have an
intricate tangle of heteroclinic intersections. A horseshoe (Sect. 1.2.2.1) of complete
symbolic dynamics with K0 = ln 2 is then formed. For Em1 > Ec , chaotic scattering
is hyperbolic.
6.3 Transitions to Chaotic Scattering 207

Fig. 6.19 New center

of period 2 and its
accompanying KAM tori,
after the period-doubling
bifurcation of π2 , for
Em1 = 0.207. Other
parameters are the same
as in Fig. 6.18 [194]
(copyright 1990, the
American Physical Society)

6.3.4 Abrupt Bifurcation to Chaotic Scattering

with Discontinuous Change in Dimension

A feature in the abrupt bifurcation discussed in Sect. 6.3.2 is that the dimension ds
of the set of singularities in the scattering functions changes continuously through
the bifurcation. In particular, it can be seen from (6.15) that although the bifurca-
tion is considered abrupt, the box-counting dimension ds is still continuous at the
bifurcation (where it takes on the value ds = 0). There exists, in fact, a different type
of abrupt bifurcation to chaotic scattering in two-degree-of-freedom Hamiltonian
systems for which the box-counting dimension changes discontinuously through
the bifurcation point [436]. To explain this bifurcation, let the particle energy E
be the bifurcation parameter and let E0 be a bifurcation point. Before the bifurca-
tion (E < E0 ), there exists a bounded chaotic region in the phase space surrounded
by forbidden potential barriers. The bounded chaos is typically developed through
the destruction of a hierarchy of KAM tori. Particles from outside cannot enter this
bounded chaotic region, so that the scattering dynamics is regular. At E0 , the barriers
disappear, and the bounded Hamiltonian chaos becomes transient, allowing scatter-
ing particles to access the previously forbidden region. Scattering then becomes
chaotic for E > E0 . The key difference between this scenario to chaotic scattering
and that discussed in Sect. 6.3.2 is that here, the change in the dimension of the set
of singularities in the scattering function is due to the sudden access of scattering
trajectories to an already developed chaotic set.
This type of abrupt bifurcation to chaotic scattering can be demonstrated by
considering the physically realistic situation whereby particles are scattered by
molecules. Assume that there are three molecules located at vertices x j ≡ (x j , y j )
( j = 1, 2, 3) of a regular triangle of unit side length. For each molecule, its interac-
tion with a scattering particle can be modeled by the Morse potential, a paradigm
for studying various phenomena in chemical and atomic physics [323, 495, 496].
The dimensionless potential function of the scattering system can be written as
208 6 Chaotic Scattering

3
V (x) = ∑ VM (x, x j ), (6.18)
j=1

where 1 2
1 − e−α (r j −re ) ,
VM (x, x j ) = (6.19)
 2
and r j = (x − x j )2 + (y − y j )2 ( j = 1, 2, 3) is the distance from vertex j, re is the
effective range of the potential hill, and α determines the steepness of the potential.
Each Morse hill is spherically symmetric and has a repulsive region surrounded
by an attractive region. The potential is highly localized, and the region about
(x, y) = (0, 0), where V (x, y) is appreciable, can be regarded as the scattering re-
gion. The difference between this scattering model and the one in Sect. 6.3.2 is that
the height of the potential is large (Em ≈ VM (x j , x j )  1), so that the potential hills
are classically impenetrable. As a result, the abrupt bifurcation here occurs at low
energies when the energy is increased through a critical value, in contrast to the case
treated in Sect. 6.3.2.
In a physically realistic energy regime, the particle energy is much smaller than
the height of the potential hills: E Em . Figure 6.20a,b show the contours of the
potential for E = 1 and 4, respectively, where for E = 1, the region enclosed between
the inner and outer closed curves is the classically forbidden one, and for E = 4, the
regions enclosed by the three somewhat circular closed curves are the classically
forbidden regions. Thus, for E = 1, particles coming from afar cannot enter the scat-
tering region, and they simply are bounced back from the potential. In this case, the
scattering is regular. However, for E = 4, the potential configuration is similar to that
of three hard disks for which the scattering is chaotic. Figure 6.21 shows the scat-
tering functions for E = 1 and E = 4. For E = 4, scattering becomes chaotic, since

a 1.5 b
1
1
0.5
0.5

0 0
y

−0.5
−0.5
−1
−1
−1.5
−1 0 1 −1 −0.5 0 0.5 1
x x

Fig. 6.20 Contour plots of the Morse potential (6.18) configuration for (a) E = 1 and (b) E = 4.
√ are α = 6 and re = 0.68. The
Other parameters  of the triangle are (x1 , y1 ) =
√ locations of the vertices
(1/2, −1/(2 3), (x2 , y2 ) = (−1/2, −1/(2 3)), and (x3 , y3 ) = (0, 1/3) [436] (copyright 1999,
the American Physical Society)
6.3 Transitions to Chaotic Scattering 209

Fig. 6.21 Scattering functions: (a), (b) Deflection functions φ (b) for E = 1, and E = 4, respec-
tively. (c), (d) Delay-time functions T (b) for E = 1 and E = 4, respectively. Parameters are the
same as in Fig. 6.20 [436] (copyright 1999, the American Physical Society)

there are an infinite number of singularities in the scattering function. Figure 6.21c,d
show the corresponding delay-time plots for E = 1 and E = 4, respectively.
The absence and presence of chaotic scattering at E = 1 and E = 4, respectively,
suggest that there is a bifurcation to chaotic scattering as E is increased from 1 to
4. It can be argued that this bifurcation is abrupt. In particular, for E = 1, there is a
small triangular-like area in the center of the scattering region (Fig. 6.20a) in which
the value of the potential is actually lower than E, and hence this area is allowed
for particle trajectories. This region, however, is inaccessible to scattering particles
from outside, because it is enclosed by a larger forbidden region. Dynamics inside
the triangular-like potential region can be chaotic. For E = 1, the phase space con-
tains both KAM tori and Hamiltonian chaotic seas, as shown in Fig. 6.22a, where the
x-coordinate and the x-component of the particle velocity are plotted on the Poincaré
surface of section defined by y = 0. Chaos in this case is bounded. As E is increased,
the triangular-like area enlarges and the phase-space structure inside also evolves.
In particular, KAM tori are destroyed and the chaotic sea is enlarged, as shown
by the phase-space structure on the Poincaré surface of section in Fig. 6.22b for
E = 2.5. At some critical energy value E0 where the inner allowed region connects
with the outside one, the previously bounded chaotic sea becomes a chaotic saddle
because trajectories can escape through one of the openings (Fig. 6.20b). Now par-
ticles coming from outside can access the transienting chaotic region, so that the
scattering becomes chaotic. The appearance of chaotic scattering is abrupt, because
for E < E0 , the scattering dynamics is smooth, while it is chaotic for E > E0 .
210 6 Chaotic Scattering

Fig. 6.22 Phase-space structure on a Poincaré surface of section of the classically allowed region
that is inaccessible to scattering particles for (a) E = 1, and (b) E = 2.5 < E0 . There are both KAM
tori and bounded chaotic seas for E = 1. For E = 2.5, all KAM tori have disappeared. Parameters
are the same as in Fig. 6.20 [436] (copyright 1999, the American Physical Society)

The critical energy E0 can be estimated as follows. For each individual potential
hill, the radius r > re of the classically forbidden spherical region is

1 √
r = re − ln (1 + 2E).
α
If there is no overlapping between the potentials, two adjacent potentials touch each
other at r = 1/2. Thus, Ec is approximately determined by

1 
re − ln (1 + 2E0 ) ≈ 1/2.
α
For re = 0.68, α = 6, this gives E0 ≈ 1.9. Due to the overlapping among poten-
tials, the actual value of E0 is somewhat larger than the estimated one. It was found
numerically [436] that E0 ≈ 2.55.
In this route to chaotic scattering, depending on whether there are still KAM
tori left for E slightly above E0 , the scattering can be either hyperbolic or non-
hyperbolic. For the scattering configuration described, abrupt bifurcation leads to
hyperbolic chaotic scattering. An alteration in the scattering configuration can lead
to abrupt bifurcation to nonhyperbolic chaotic scattering. For instance, when the
Morse molecule on the y-axis is pulled closer to the pair in the x-direction, say at
y3 = 0.45, there are still KAM tori left in the scattering region after the bifurcation
that connects the inner and outer allowed regions.
The major physically measurable consequence of the abrupt bifurcation to
chaotic scattering described above is that there is a discontinuous change in the di-
mension ds of the set of singularities in the scattering function. From Fig. 6.21a–d,
it is apparent that for E < E0 (before the bifurcation), both the scattering and the
delay-time functions are smooth, and hence, ds = 0. For E > E0 (after the bifur-
cation), the scattering and the delay-time functions suddenly contain a Cantor set
6.4 Nonhyperbolic Chaotic Scattering 211

Fig. 6.23 Box-counting

dimension ds of the set of
singularities in the scattering
function versus E. There is a
jump in ds at E0 , the abrupt
bifurcation point to chaotic
scattering with discontinuous
change [436] (copyright
1999, the American Physical
Society)

of singularities with ds > 0. Figure 6.23 shows ds versus E for 1.0 ≤ E ≤ 4.0.
Theoretically, for E immediately above E0 when the bounded chaotic sea in the
scattering region just becomes transient, the value of ds is unity because by continu-
ity, the box-counting dimension of the chaotic saddle for E slightly above E0 is the
same as that of the Hamiltonian chaotic sea for E slightly below E0 , which is the
phase-space dimension. Numerically, it was found that ds ≈ 0.9 for E = 2.6 > E0 .
This discontinuous change in the box-counting dimension is the key ingredient
that distinguishes this abrupt route to chaotic scattering from the one discussed in
Sect. 6.3.2.

6.4 Nonhyperbolic Chaotic Scattering

6.4.1 Algebraic Decay

In general, Hamiltonian systems arising in situations of physical interest permit

three types of dynamics: regular (integrable), completely chaotic (hyperbolic), and
nonhyperbolic, where regular and hyperbolic dynamics correspond somewhat to the
two oppositely extreme situations, and nonhyperbolic dynamics lies somewhere in
between them. Thus, in open Hamiltonian systems, while hyperbolic scattering can
be expected, a more common situation has chaotic saddles encircling KAM tori in
the phase space, for which the scattering is nonhyperbolic. Indeed, as we have seen,
in the saddle-center route to hyperbolic scattering, there exists a wide parameter
interval for which the scattering is nonhyperbolic.
A fundamental issue in nonhyperbolic chaotic scattering concerns the impact of
the presence of KAM surfaces on particle motions. In this regard, one generally ob-
serves that, if a particle enters a chaotic region near some KAM surface, the particle
wanders close to that surface for a long time. This is called the stickiness effect of
212 6 Chaotic Scattering

KAM tori. The origin of this stickiness effect can be understood by the following
simple observation. Take two nearby points on a given KAM surface and observe
their evolution. What one typically finds is that the distance between the two points
hardly changes with time, because the Lyapunov exponents in the directions along
the KAM surface are zero (i.e., the motion is quasiperiodic). The symplectic nature
of the Hamiltonian dynamics stipulates that the Lyapunov spectrum be organized in
pairs of exponents with equal values but opposite signs. An orbit on a KAM sur-
face there has zero Lyapunov exponents in directions both along and perpendicular
to the surface. Now consider a particle initialized in the chaotic region. Due to er-
godicity, this particle will come arbitrarily close to some KAM surface bordering
the chaotic region. When this occurs, the effective, finite-time Lyapunov exponents
become nearly zero, leading to slow divergence of the particle trajectory from the
KAM surface. The role of the KAM surfaces is similar to that of a marginally sta-
ble fixed point in one-dimensional nonhyperbolic transient chaos, as described in
Sect. 2.4.
The main consequence of the stickiness effect on chaotic scattering is that the
probability (density) p(t) of escaping at time t is asymptotically algebraic, and
hence the long-time behavior of the survival probability P(t) is also algebraic:

p(t) ∼ t −z−1 , P(t) ∼ t −z , for large t, (6.20)

where z is an algebraic-decay exponent [136, 137, 315, 388, 513–515], in contrast to

exponential decay in hyperbolic chaotic scattering or associated with chaotic tran-
sients in dissipative systems.
The algebraic-decay exponent, as defined by (6.20), in fact characterizes scat-
tering situations, i.e., situations in which the initial conditions are chosen far away
from KAM surfaces. There exists a distinct type of initial conditions, e.g., those
chosen in the sticking region, for which the exponent is different. Following the
escape process of these initial conditions is called the transient-chaos approach to
chaotic scattering [603]. The corresponding escape time and survival probability
distributions ptr (t) and Ptr (t), respectively, obey the law

ptr (t) ∼ t −ztr −1 , Ptr (t) ∼ t −ztr , for large t, (6.21)

with a different decay exponent ztr . Pikovsky pointed out [603] that a larger number
of trajectories are held up in the transient case than in the scattering case and the
exponent ztr is smaller than z with unit difference:

ztr = z − 1. (6.22)

A derivation of this rule will be given in Sect. 6.4.5.

The escape-time distribution of the transient case is thus proportional to the sur-
vival probability of the scattering case:

ptr (t) ∼ P(t). (6.23)

6.4 Nonhyperbolic Chaotic Scattering 213

As a consequence of the algebraic decay, the escape rate from regions around
KAM surfaces tends to zero, κ → 0, implying, according to (2.76), that the chaotic
saddle is locally space-filling:

D0 = D1 = 2. (6.24)

A detailed discussion about the dimensionality will be given in Sect. 6.4.3.4

Theoretical models [144,315,514,515] have been proposed to explain the numer-
ically observed algebraic-decay law. A fundamental assumption of these models is
that a particle in the phase space executes a random walk between families of self-
similar chains of islands. Besides yielding the algebraic scaling law, these models
also predict the values of the algebraic-decay exponent based on the number of self-
similar families of KAM islands included in the calculation. It should be pointed out,
however, that chaotic systems with no KAM surfaces may also exhibit an algebraic-
decay behavior, reflecting some long-term correlation among orbits. One example
of such systems is the Bunimovich stadium, in which the dynamics is completely
chaotic. The origin of the long-term correlation in this system is due to the existence
of an infinite class of neutrally stable periodic orbits in the phase space [473, 801].
As a system parameter is changed, if observations are made over finite times,
the resulting algebraic-decay exponent can exhibit large fluctuations [286, 445].
The dynamical mechanism underlying the fluctuations will be treated in Sect. 6.5.
The rest of this section will be devoted to the development of horseshoe structure
in nonhyperbolic chaotic scattering, the box-counting dimension, the intermediate-
time behavior, and the relationship between scattering and leaked dynamics.

6.4.2 Development of Horseshoe Structure in Nonhyperbolic

Chaotic Scattering

As a parameter changes, the saddle underlying nonhyperbolic chaotic scattering can

undergo structural changes before it becomes a fully developed hyperbolic saddle
with a complete grammar. It was pointed out by Jung and coworkers that a param-
eter can be introduced to characterize the development of the underlying horseshoe
structure [362, 657], which measures the penetration depth of the tendrils into the
interior of the fundamental area of the horseshoe.
To be specific, we consider here the case of a binary horseshoe characterized by
the existence of two fixed points on a Poincaré map. One of the fixed points is hyper-
bolic for any parameter values, and it lies inside the scattering region. Its invariant
manifolds trace out the horseshoe. The stability of the other fixed point depends
on the parameter values. Initially, this inner fixed point is stable and surrounded by

4 The algebraic-decay law holds not only for two-degree-of-freedom Hamiltonian systems, but also

for higher-dimensional systems [192, 381].

214 6 Chaotic Scattering

Fig. 6.24 Poincaré section of a scattering system described by a binary horseshoe when the inner
fixed point (black dot) is elliptic. The arrows indicate the flow of trajectories about the hyper-
bolic fixed point and outside the scattering layer [363] (with kind permission from the Institute of
Physics)

KAM surfaces. As a system parameter is changed, the KAM surfaces break into
chains of secondary islands and chaotic layers, which eventually fuse into a fully
developed horseshoe. Figure 6.24 is a schematic representation of the case in which
a large primary island exists and encloses the inner fixed point.
The development parameter α defined in [362, 657] takes on specific values
for certain configurations of the stable and the unstable manifolds of the hyper-
bolic fixed point. To visualize such a situation, consider Fig. 6.25, which shows the
topology of the horseshoe at a somewhat initial stage of its development. The first
intersection of the stable and the unstable manifolds, the primary intersection point,
is denoted by P. This point, together with the hyperbolic point H and the mani-
folds passing through them, defines a rectangular domain, the fundamental area of
the horseshoe. Tendrils of positive labels are defined as arcs of the manifolds lying
within the fundamental area, where labeling starts at P. The tendrils are images or
preimages of each other, a property that allows one to define tendrils with negative
labels, too. The development parameter takes on values α = 2−m if the first tendril
of one of the invariant manifolds is intersected by the tip of the mth tendril of the
other manifold at two points only. Figure 6.25 illustrates the case of m = 4. It can be
seen that the tendrils of levels 2 and 3 also intersect at two points only. In general,
the tips of the primary tendrils of the invariant manifolds penetrate into areas that
are free of manifolds. As a result, homoclinic bifurcations cannot occur under small
changes of the parameters.
When α reaches the values α = 1/8, 1/4, and 1/2, consecutively, the tendrils
become longer, as suggested by Fig. 6.25. The horseshoe becomes more developed
and the stability island shrinks during this process. According to the definition in
6.4 Nonhyperbolic Chaotic Scattering 215

Fig. 6.25 Schematic diagram illustrating the topology of an incomplete binary Hamiltonian horse-
shoe for development parameter α = 1/16, where H and P denote the hyperbolic fixed point and
the primary intersection point, respectively. Tendrils of the stable manifold have been emphasized
(in black). Stars mark the subsequent images of the point labeled by 0 [363] (with kind permission
from the Institute of Physics)

[362,657], α can take on any values r2−m with r = 1, 2, . . . , 2m − 1 at any fixed level
m. For example, α = 3/4 corresponds to the case that tendril 2 crosses the funda-
mental area, turns back into it, and the first tendril of the other manifold becomes
intersected again by this long tendril at two points only. The limit α → 1 corre-
sponds thus to a fully developed binary horseshoe with K0 = ln 2 (α = 0 describes
an integrable homoclinic loop).
The topology of the horseshoe for any α is unique. In fact, there is a one-to-one
correspondence between the topological entropy K0 and the development parameter:

eK0 = g(α ), (6.25)

where g(0) = 1 and g(1) = 2. The explicit from of the function g(α ) is not known,
but its value can be determined [657] for certain values of α , as illustrated in
Table 6.1. Relation (6.25) is universal. The dependence of α on system parame-
ters is, however, not universal. Thus the relation between the topological entropy
and some system parameter depends on system details.
216 6 Chaotic Scattering

Table 6.1 Relation between the development parameter

α and function g = exp (K0 ) at a few points
α 0 0.25 0.50 0.75 1
g(α ) 1 1.54 1.71 1.91 2

The ideas can be generalized straightforwardly to ternary horseshoes [362].

The development parameter is then denoted by β , and its typical values are
β = 3−m . An experiment with ternary horseshoes will be presented in Sect. 7.6
in the quantum-mechanical context.
A careful investigation of the geometry of incomplete horseshoes can lead to
a detailed understanding of their topological features [361, 527]. The rules that de-
scribe how tendrils are mapped onto each other are sometimes called lobe dynamics,
and play an important role in the context of fluid transport (Sect. 10.3.3).

6.4.3 Dimension in Nonhyperbolic Chaotic Scattering

An important result in nonhyperbolic chaotic scattering concerns the box-counting

dimension of the set of singularities in the scattering function. Lau et al. [470]
argued, with numerical support, that the dimension is ds = 1. This unity of the
dimension is a direct consequence of the algebraic-decay law associated with
nonhyperbolic chaotic scattering, which can be seen intuitively by considering a
measure-zero Cantor set that has ds = 1, through the following construction [470].
Start with the unit interval [0, 1], remove the open middle third interval. From each
of the two remaining intervals remove the middle fourth interval. Then from each
of the four remaining intervals remove the middle fifth, and so on. At the nth stage of
the construction, there are N = 2n subintervals, each of length εn = [2/(n + 2)]2−n.
The total length of all subintervals εn N ≈ 2/n goes to zero algebraically for n → ∞.
For n  1, we have εn ≈ 21−n /n, and hence n ≈ ln εn−1 / ln 2. In order to cover the set
with intervals of size εn , the required number of intervals is N(ε ) ≈ 2/(ε n), which
depends on ε as
2 ln 2
N(ε ) ≈ .
ε ln ε −1
The box-counting dimension ds of the set is then obtained as the limit of

ln N(ε ) ln ln ε −1 ln (2 ln 2)
= 1− + .
ln ε −1 ln ε −1 ln ε −1

Note that for ε → 0 this yields ds = 1. The weak logarithmic dependence does not
contribute to the asymptotic dimension value. However, it is the logarithmic term
that is responsible for ensuring that the total length (measure) is zero: ε N(ε ) ∼
(ln ε −1 )−1 → 0 as ε → 0.
6.4 Nonhyperbolic Chaotic Scattering 217

From a more general standpoint, if at each stage a fraction ηn = α /(n + c), where
α and c are constants, is removed from the middle of each of the 2n remaining
intervals, then
1 1 −α
N(ε ) ∼ ln . (6.26)
ε ε
An observation is that the slope of the curve ln N(ε ) versus ln ε −1 , which is
d ln N(ε )/d(ln ε −1 ), is always less than 1 for small ε , but it approaches 1 logarith-
mically as ε → 0. Thus, the result ds = 1 still holds. A practical implication is that
for fractals whose general characters are similar to those for this example, an ac-
curate numerical estimation of the dimension will require going to infinitesimally
small scales, and as such, any numerical estimation of the dimension over a finite
range of scales will be an underestimate. As the scale is decreased, the numerically
determined value of the dimension will increase toward 1.
The relevance of the above construction of the Cantor set to chaotic scattering can
be argued as follows [470]. Consider particles launched from a line segment strad-
dling the stable manifold of the chaotic saddle. There is an interval in the values of
some input variables that lead to trajectories remaining in the scattering region for at
least a duration of time T0 . By time 2T0 a fraction η of these particles will have left.
If the initial conditions of these escaping particles are all located in the middle of the
original interval, there are then two equal-length subintervals of the input variable
that lead to trajectories that remain for at least time 2T0 . By time 3T0 an additional
fraction η of the particles whose initial conditions are located in the middles of the
two subintervals remaining at time 2T0 escape. There are then four subintervals, par-
ticles initiated from which can remain in the scattering region for time at least 3T0 ,
and so on. The resulting set is a Cantor set of measure zero on which particles never
escape. The box-counting dimension of the Cantor set is given by
ln 2
ds = .
ln [2/(1 − η )]

For hyperbolic chaotic scattering, particles escape exponentially from the scattering
region: P(t) ∼ e−κ t , where P(t) is the survival probability and the escape rate is
determined by the fraction η as κ = T0−1 ln (1 − η )−1. For nonhyperbolic dynamics,
because of the algebraic decay P(t) ∼ t −z , the fraction η is no longer a constant: it
varies at each stage of the construction of the Cantor set. At the nth stage, i.e., for
times (n − 1)T0 < t < nT0 (n large), the fraction ηn is approximately given by ηn ≈
−T0 P−1 dP/dt ≈ z/n, which yields (6.26) and hence a Cantor set with dimension 1,
where α in (6.26) is identified to be the algebraic-decay exponent z.
To provide support for the above heuristic argument, Lau et al. [470] conducted
numerical experiments utilizing a two-dimensional scattering map (see Sect. 6.4.4).
Evidence of the box-counting dimension’s approaching unity as the length scale
is reduced is apparent even in the plot of a scattering function: one observes an
apparent increase in the density of the singularities. Computations indicate that the
dimension indeed approaches unity as finer and finer scales are examined [470]. The
signature of the unit box-counting dimension has also been identified in the classical
model of two-electron atoms [313].
218 6 Chaotic Scattering

6.4.4 Intermediate-Time Exponential Decay

In spite of the asymptotic algebraic decay, on intermediate times the decay of the
survival probability can be exponential in nonhyperbolic chaotic scattering. Its oc-
currence depends on the choice of the initial distribution of particles used to obtain
the survival probability. For initial conditions far away from KAM tori the exponen-
tial decay is generically present.
The two-dimensional area-preserving map utilized by Lau et al. to establish the
unity of the box-counting dimension in nonhyperbolic chaotic scattering [470] pro-
vides a convenient numerical model for observing as well the intermediate-time
exponential decay:

x λ [x − (x + y)2/4]
M = , (6.27)
y λ −1 [x + (x + y)2/4]

where λ > 1 is a parameter. Utilizing a discrete-time map (instead of a continuous-

time flow) makes high-precision computation feasible. For the map (6.27), almost
all orbits starting from negative values of y are scattered to infinity. The dynamics is
nonhyperbolic for λ < 6.5 and hyperbolic for λ > 6.5. In the nonhyperbolic regime
there is a major KAM island in the phase space that leads to an asymptotically
algebraic decay.
The survival probability P(n) of a particle in the scattering region can be approx-
imated by the fraction
 of a large number of particles still remaining in the scattering
region (defined by x2 + y2 < r) at time n. In [534], r = 100 is fixed and initial con-
ditions are chosen from the horizontal line at y0 = −2. The decay of P(n) with time
is exponential for intermediate n values, and algebraic for larger times, as shown in
Fig. 6.26. For λ = 4.0, the exponential decay P ∼ e−κ n holds up to n = 250, after
which the algebraic decay law P(n) ∼ n−z takes over, where κ ≈ 0.08 and z ≈ 1.
This suggests that there is a time-scale separation in the system: up to a certain
multiple of 1/κ the behavior is different from that in the long-time limit.
To explain this time-scale separation, the following picture has been proposed
[13, 18, 370]. The entire chaotic saddle can be effectively split into two components:
a hyperbolic one lying outside the KAM surfaces and a nonhyperbolic component

Fig. 6.26 Survival

probability P(n) for the map
(6.27) for λ = 4.0. The inset
illustrates exponential decay
on intermediate time scales
[534] (copyright 2002, the
American Physical Society)
6.4 Nonhyperbolic Chaotic Scattering 219

located around the KAM tori. Particles starting far from the KAM surfaces and
having intermediate lifetimes have no chance to come close to the nonhyperbolic
region, because of the stickiness of the KAM tori. The relaxation to the non-
hyperbolic behavior occurs after most particles have approached the hyperbolic
component. Since the dwelling around the hyperbolic component is sufficiently
long, the intermediate-time dynamics can be characterized by a well-defined escape
rate, a positive Lyapunov exponent, and dimension ds that is strictly less than unity.
The full escape-time distribution can thus be written, after short-time fluctuations
occurring up to time n∗ , as

ae−κ n for n∗ < n < nnh ,
p(n) ≈ (6.28)
ae−κ n + b(κ n)−z−1 for n > nnh ,

where nnh is the time when the contribution of the nonhyperbolic component be-
comes first observable, and therefore ae−κ nnh  b(κ nnh)−z−1 .
It should be emphasized that the development-parameter approach of Sect. 6.4.2
characterizes the hyperbolic component of the saddle only. In a realistic scattering
experiment most particles move along trajectories that stay in the scattering region
for a short time only. They do not have the time necessary for penetrating into the
nonhyperbolic component around the KAM surfaces. In this sense knowledge about
the hyperbolic properties is highly experimentally relevant.

6.4.5 Relation to Poincaré Recurrences

The behavior of intermediate-time exponential decay is also present in generic

leaked area-preserving maps f, when the leak is far away from any KAM tori.
The arguments provided in Sect. 2.7 to compare leaked with closed systems can
be applied, which lead to the conclusion that the intermediate-time relaxation of
the Poincaré recurrence distribution pr (n) and of the escape-time distribution p(n)
of the corresponding leaked problem are proportional to each other: pr (n) ∼ p(n)
for any n with p(n) given by (6.28). In particular, (2.104) holds, where γ is
the relaxation rate of the Poincaré recurrence-time distribution (on intermediate
time scales). Furthermore, with initial condition (2.105), the entire distributions
coincide. In area-preserving cases, ρr is a uniform distribution over the image
f(I) of the leak. These features can be illustrated by the standard map whose
phase space is shown in Fig. 6.27(a). Chaotic trajectories tend to stay for a long
time near the hierarchical border of the KAM island, shown in the center of
the figure, which constitutes the nonhyperbolic component of the chaotic sad-
dle. The exponential decay is governed by the hyperbolic component, whose un-
stable manifold is shown in Fig. 6.27b for a specific leak. The distributions of
the escape and the recurrence times are shown in Fig. 6.28. As in the case of
dissipative systems (Fig. 2.21), short-time oscillations and an intermediate-time
exponential decay are present. However, in this case an algebraic decay is observed
220 6 Chaotic Scattering

Fig. 6.27 (a) Phase space of the standard map given by yn+1 = yn − 0.52 sin(2π xn ) and xn+1 =
xn + yn+1 , peridoic in |x, y| ≥ 0.5, with a leak defined by I = [0.25 < x < 0.45, −0.5 < y < −0.03].
The nonhyperbolic component of the chaotic saddle about the KAM island is in the center of the
figure. The support of ρr (see (2.105)) at f(I) for which pr (n) = p(n) is marked as the dark-gray
region. (b) The unstable manifold of the hyperbolic component of the saddle

Fig. 6.28 Recurrence pr (n) and escape p(n) time distributions in the leaked standard map, as
shown in Fig. 6.27. Both functions decay exponentially over intermediate times with exponent
γ = κ . For pr (n) (lower curve) a single initial condition is followed over 1011 steps. For p(n) (upper
curve) ρ0 is a homogeneous distribution ρμ for |x| > 0.25. Lower panel: Distributions multiplied
by exp(κ n), where γ = κ ≈ 0.011

for times n > nnh ≈ 300. This provides additional support for the picture that the
initial density ρ0 quickly converges to the hyperbolic component of the saddle, and
that the nonhyperbolic component is approached afterward. The transition of a typ-
ical trajectory from the hyperbolic to the nonhyperbolic components of the saddle
is illustrated in Fig. 6.29. By increasing the number n0 of iterations over which tra-
jectories that have not escaped are monitored when applying the sprinkler method
of Sect. 1.2.2.3, the nonhyperbolic component appears with an increasing weight.
6.4 Nonhyperbolic Chaotic Scattering 221

Fig. 6.29 Saddles characterizing the dynamics at different time n0 obtained by the sprinkler
method. For n0 nnh ≈ 300 in (a), the hyperbolic component is disjoint from the region con-
taining the KAM island. In (b) and (c) the nonhyperbolic component becomes more and more
apparent

Quantitatively, the crossover time nc > nnh between the exponential decay and
the algebraic decay can be defined as
ae−κ nc = b(κ nc )−(z+1) , (6.29)

which is the time when the contributions from the hyperbolic and the nonhyperbolic
components are of equal importance. The crossover time has been found [18] (as
also can be seen from (6.29) under the assumption that b/a depends at most weakly
on κ ) to be proportional to the reciprocal of the escape rate from the leaked system:
nc ∼ 1/κ . (6.30)
This scaling implies that the exponential decay dominates for small recurrence/leak
regions.
The similarity between scattering dynamics and Poincaré recurrences can be used
to explain the difference between the long-term algebraic-decay exponents for scat-
tering and transient chaos as described in Sect. 6.4.1. Here we follow the argument
of Altmann et al. [17]. Consider first initial conditions around the sticking region
(the case of transient chaos) and examine the time a trajectory takes to escape to
a region far away from the sticking region. The distribution Ptr (τ ) of escape times
longer than time τ  1 is proportional to the natural measure μ (τ ) of the region
of the phase space to which the trajectories stick for a time longer than τ . Due to
ergodicity, we have
nτ
Ptr (τ ) ∼ μ (τ ) = , (6.31)
n
where nτ is the total time spent inside the sticking region and n  1 is the total time
of observation.
For the recurrence problem with a single trajectory of length n initialized far
away from the sticking region, the cumulative probability Pr (τ ) to find recurrence
times larger than τ can be expressed as

Nτ
Pr (τ ) = ∼ P(τ ), (6.32)
N
222 6 Chaotic Scattering

where Nτ is the number of recurrences with recurrence times larger than τ and N is
the total number of recurrences observed in the time interval n. The right-hand side
expresses the observation that the recurrence-time statistics are proportional, also
for long times, to the lifetime distribution P(τ ) of the scattering problem for initial
conditions far away from any sticking region. Since the total observation time can
be estimated as N times the mean recurrence time τr , (2.103), we have n ∼ N τr .
Similarly, the total time spent inside the sticking region is approximately the number
Nτ of recurrences with times longer than τ multiplied by τ : nτ ∼ Nτ τ . Since τr is a
constant, independent of τ , these allow us to write

Nτ nτ /τ nτ τr Ptr (τ )
P(τ ) ∼ ∼ = ∼ , (6.33)
N n/τr nτ τ

from which the shift of algebraic decay exponent by one (6.22) immediately follows
from (6.20) and (6.21).

6.5 Fluctuations of the Algebraic-Decay Exponent

in Nonhyperbolic Chaotic Scattering

Consider the two-dimensional setting in which the phase space contains a central
KAM island encircled by some outermost KAM surface. There is a chaotic region
outside the KAM surface, which contains smaller island chains. For particles initial-
ized in the chaotic region, their escape process takes place on a wide range of time
scales. For intermediate times (e.g., t < t1 ), the decay is exponential. For particles
that stay a long but finite time in the region, their corresponding trajectories spend
substantial amounts of time near some accessible KAM surfaces, leading to an al-
gebraic decay in the survival probability. From an observational point of view, the
more dominant the island chains are in the chaotic region, the more time the particles
spend near them, the slower the escape process is. The “slowness” can be quantified
by the algebraic decay exponent z in (6.20), where a smaller value of z corresponds
to a “slower” escape process over the period of observation. This intuition is con-
sistent with the theoretical picture [514, 515] where the value of the decay exponent
decreases as the number of island chains included in the calculation increases.
When some KAM surfaces undergo transformations from being absolute barri-
ers to partial barriers called Cantori [498], an entire new region of the phase space
becomes accessible to the chaotic orbits. The decay exponent z drops drastically
after this point. After the breakup of the outermost KAM surfaces surrounding
the central island, another time scale t2 > t1 becomes important for the system.
For particles that have escaped in the time interval t1 < t < t2 , their trajectories
lie entirely outside the newly created Cantori. Therefore, the exponent z measured
over this time interval is still roughly the same as that before the breakup of the
corresponding KAM surfaces. The particles that stay longer than t2 , however, will
explore the region enclosed by the newly created Cantori, thereby causing the ex-
ponent z to be markedly smaller than the one measured over the interval t1 < t < t2 .
6.5 Fluctuations of the Algebraic-Decay Exponent in Nonhyperbolic Chaotic Scattering 223

Thus the curve P(t) versus t exhibits two scaling regions with different algebraic de-
cay exponents. The time scale at which this crossover takes place coincides with the
typical time for a chaotic trajectory to penetrate the newly created Cantori. This con-
sideration can be repeatedly applied as the parameter further increases, suggesting
fluctuations in the exponent z. As the system approaches hyperbolicity, the interval
0 < t < t1 becomes larger and larger, reflecting the fact that the island structures in
the phase space become less and less significant.
An acknowledged notion is that the algebraic decay in the presence of KAM
surfaces should be a universal phenomenon with a single exponent. This notion is
partly derived from the numerical work of Karney [388], Chirikov and Shepelyansky
[136], and of Cristadoro and Ketzmerick [144]. While this may indeed be the case
given the hierarchical island structure in the phase space, to attain such a univer-
sal exponent one needs a prohibitively long time to perform the needed numerical
calculations. In the case of actual experiments, the required long period of time
of observation becomes even more unrealistic. This notwithstanding, the algebraic-
decay behavior itself is found to occur on relatively short time scales. The value
of the decay exponent measured over a finite time interval accessible to numerical
experiments is influenced by the presence of some dominant island structures. In
this regard the variations of the exponent can, in fact, be utilized to reflect the major
qualitative changes in the system.

6.5.1 Numerical Model

The chaotic-scattering system introduced by Troll and Smilansky [790] is conve-

nient for illustrating the phenomenon of algebraic-decay exponent fluctuations. In
the system, particles are scattered by an infinite array of nonoverlapping elastic scat-
terers, as shown in Fig. 6.30, which are placed in the plane at constant intervals a
along the y-axis. Each scatterer is represented by a circularly symmetric potential
V (r) that vanishes for r > R, where R < a/2 (nonoverlapping). To have chaos, it is
necessary for the potential V (r) to be attractive, because if V (r) is repulsive, only
trivial invariant orbits can be formed in the scattering region.
The nonoverlapping condition entails that particle trajectories be straight lines
in the region between potential wells and suffer a deflection when they enter some
scatterer. Let θ (b) be the deflection angle of a particle trajectory from an individual
scatterer. In general, θ (b) is a monotonically decreasing function of the impact pa-
√ range of each scatterer, θ (b) = 0
rameter b (at least near b = 0). Due to the finite
for | b |> R. For a particle of unit mass, u = 2E is the constant particle velocity
during its free motion. For a particle moving toward any of the scatterers with im-
pact parameter b and velocity u, its new velocity u after the scattering has the same
magnitude (u) as u but assumes a different direction. Let β and β be the angles of
the velocity vectors relative to the −y-axis when the particle is in regions where the
potential is negligible. From Fig. 6.30, we have

β = β − θ (b). (6.34)
224 6 Chaotic Scattering

Fig. 6.30 The

Troll–Smilansky scattering
system of an infinite array
of nonoverlapping circular
potentials. Here b < 0, b > 0
[445] (copyright 1992, the
American Physical Society)

Note that θ (b) < 0 for b > 0. Now the particle may either collide with the scatterer
above (if uy = −u cos β > 0) or with the scatterer below (if uy < 0). In either case,
scattering is determined by the value of the impact parameter b relative to the new
scatterer. A geometric argument gives

b = b − a sgn[cos β ] sin β . (6.35)

The next scattering takes place only if |b | < R. Otherwise, the particle will continue
to move along a straight-line trajectory, leaving the system. Such particles will be
regarded as having escaped the scattering region.
Symbolically, the two-dimensional map (6.34) and (6.35) can be represented as

(β , b ) = M(β , b). (6.36)

The phase space for M is defined by the domain [0, 2π ) × [−R, R], which is a cylin-
der. It can be verified that M is area-preserving.
In the Troll–Smilansky system, chaotic scattering can arise via the saddle-center
bifurcation route. This can be seen as follows. For a given physical potential V (r),
the deflection angle θ (b) is a monotonically decreasing function of b for values of
6.5 Fluctuations of the Algebraic-Decay Exponent in Nonhyperbolic Chaotic Scattering 225

b near b = 0, and θ (0) = 0. Suppose θ (b) = −kb, where k is a positive constant

measuring the local slope of the deflection function. The scattering map becomes

β = β + kb, (6.37)
b = b − a sgn[cos(β )] sin(β ). (6.38)

This map has fixed points at (β , b) = (0, 0) and (β , b) = (π , 0), corresponding to

orbits traveling along the negative and positive y-axis, respectively. Due to the sym-
metry of the system, it is necessary to analyze the linear stability of only one of the
fixed points, say (π , 0). Near (π , 0) the linearized mapping has the eigenvalues

λ± = [(2 − ak) ± a2 k2 − 4ak]/2. (6.39)

For k = 0, which corresponds to V (r) = 0 in the potential, we have λ± = 1. For

0 < k < 4/a, however, the eigenvalues λ± are complex and the fixed point is el-
liptic. For a nonintegrable system, an elliptic fixed point is surrounded by layers of
quasiperiodic orbits, other elliptic points, and hyperbolic points whose stable and
unstable manifolds form intricate homoclinic tangles, signifying the existence of
chaotic invariant sets. In this sense, transition to chaos occurs as soon as k becomes
nonzero. In physical terms, chaotic scattering takes place immediately after an array
of attractive potentials is placed in the plane, regardless of the depth of each indi-
vidual potential. For k slightly above zero, chaotic regions are exponentially small,
and they exist between and outside the KAM tori. Beyond the initial bifurcation to
nonhyperbolic chaotic scattering, the phase space contains a mixture of KAM sur-
faces and chaotic components. Further increase in k leads to complete hyperbolic
scattering dynamics.

6.5.2 Decay-Exponent Fluctuations

To demonstrate the dependence of the decay exponent on system parameters, it is

convenient to choose V (r) to be a localized quadratic potential given by [445]

V (r) = −V0 [1 − (r/R)2 ], for r ≤ R, (6.40)

where V (r) = 0 for r > R, and V0 ≥ 0 measures the depth of the potential well.
A calculation similar to that of Sect. 6.3.1 yields the following expression for the
deflection angle:
√
−2b R2 − b2
tan θ (b) = 2 , (6.41)
R (E + V0)/V0 − 2b2

where E is the particle energy. The coefficient k in (6.37) is then k = 2V0 /(R(E +
V0 )). For V0 = 0.2, E = 0.5, R = 1.0, and a = 3.7, the decay exponent of the survival
226 6 Chaotic Scattering

Fig. 6.31 Algebraic decay

exponent z as a function of
parameter a for the
Troll–Smilansky scattering
map [445] (copyright 1992,
the American Physical
Society)

Fig. 6.32 Phase-space plots for (a) a1 = 3.60 and (b) a2 = 3.62 [445] (copyright 1992, the
American Physical Society)

probability P(t) for particles initiated far away from KAM islands is found to be
z ≈ 1.32. As the parameter a varies, the decay exponent z fluctuates, as shown in
Fig. 6.31. For larger values of a, there are no KAM surfaces, and particles appear to
escape from the scattering region exponentially.
The dynamical mechanism responsible for the fluctuations in the algebraic-decay
exponent z can be seen by examining the phase-space structure. For example, one
can choose two representative pairs of parameter values from Fig. 6.31: a1 = 3.60,
a2 = 3.62 and a3 = 3.70, a4 = 3.72, and explore the topology of the KAM sur-
faces and chaotic sets for each parameter value. Specifically, for the first pair a1 and
a2 , the decay exponent increases, z2 > z1 , while for the second pair the exponent de-
creases, z4 < z3 . The phase-space structures corresponding to a1 and a2 are shown in
Fig. 6.32a,b. For clarity, only trajectories on KAM surfaces are plotted. A noticeable
6.5 Fluctuations of the Algebraic-Decay Exponent in Nonhyperbolic Chaotic Scattering 227

Fig. 6.33 Long trajectories on the nonhyperbolic chaotic saddles for (a) a1 = 3.60 and (b) a2 =
3.62 [445] (copyright 1992, the American Physical Society)

feature common to both parameter values is that the phase space is divided into
two distinct regions by KAM surfaces: region B enclosed within the KAM sur-
faces and region A that lies outside. As a result of this phase-space partitioning,
particles launched from outside visit only region A before exiting the potential.
Region B in this case plays the role of the central island. The particle decay dy-
namics for both a1 and a2 are thus determined by the structures present in region
A. As a is increased from a1 to a2 , the boundary between the two regions deforms
slightly but otherwise remains essentially intact. In contrast, the area occupied by
islands undergoes a reduction. The effect of this reduction can be further demon-
strated by plotting in its entirety the connected chaotic component in region A.
Figure 6.33a,b, obtained using the PIM-triple procedure (Sect. 1.2.2.4), displays
two such long chaotic trajectories. Evidently, the sum of blank pockets in region
A that represent areas occupied by islands is markedly smaller in Fig. 6.33b than in
Fig. 6.33a. The implication in physical terms is that particles exit the potential faster
for a2 = 3.62 than for a1 = 3.60, thereby furnishing an explanation to the observed
relationship z2 > z1 .
For the second parameter pair a3 and a4 , for which the algebraic-decay exponent
actually decreases, the corresponding phase-space topologies are quite different, as
shown in Fig. 6.34a,b. In particular, while the two regions A and B in Fig. 6.34a
are still visibly separated by KAM surfaces, the boundary between the two regions
in Fig. 6.34b has been destroyed. This situation entails that the chaotic component
previously enclosed in region B is now accessible to particles initialized in region
A. In geometric terms, the two chaotic regions lying on different sides of the KAM
surfaces have been combined into a single connected chaotic set. The latter point
can be seen in Fig. 6.35a,b, a long trajectory on the chaotic saddle for a3 and a4 ,
respectively.
It is known that Cantori created immediately after the breakup of KAM surfaces
serve as effective barriers to particle transport [498]. The typical time for a par-
ticle to penetrate the Cantori thus constitutes a new time scale n2 in the system.
228 6 Chaotic Scattering

Fig. 6.34 Phase-space plots for (a) a3 = 3.70 and (b) a4 = 3.72 [445] (copyright 1992, the
American Physical Society)

Fig. 6.35 Long trajectories on the nonhyperbolic chaotic saddles for (a) a3 = 3.70 and (b) a4 =
3.72 [445] (copyright 1992, the American Physical Society)

The presence of this new time scale leads to observable consequences. For parti-
cles that exit the potential in time less than n = n2 , the corresponding trajectories
lie entirely outside the newly created cantori. This implies that the exponent mea-
sured over some time interval n1 < n < n2 should be roughly the same as that before
the breakup of KAM surfaces. Particles that stay longer than n = n2 , however, are
likely to penetrate the Cantori and explore the chaotic component previously en-
closed within the KAM surfaces. Reflected in the decay dynamics, this corresponds
to a slower escape process, thus a smaller decay exponent measured over the inter-
val n > n2 . The numerical result for a4 is shown in Fig. 6.36, where two regions of
distinct slopes can be seen. The exponent measured over the first interval of scaling
behavior n < n2 is z ≈ 2.27, while that measured over the remaining interval n > n2
6.5 Fluctuations of the Algebraic-Decay Exponent in Nonhyperbolic Chaotic Scattering 229

Fig. 6.36 Crossover behavior in the algebraic-decay law for a4 = 3.72 [445] (copyright 1992, the
American Physical Society)

is z ≈ 1.88, confirming the consideration given above. The crossover time n = n2 in

Fig. 6.36 is roughly 1, 200 iterates. As indicated, this time represents the number of
iterates for a typical trajectory to penetrate the Cantori.5
The discussions above indicate that a sudden structural change, or a metamor-
phosis, in the phase space, as exemplified by the destruction of KAM surfaces and
subsequent exposure of a new layer of KAM islands, can manifest itself through
variations in the decay exponent. Conversely, this relationship can be used to in-
terpret unexpected findings in physical systems. One such example is the study
reported in [450] in which a theoretical model of microwave ionization of hydrogen
Rydberg atoms was considered. The results indicate that contrary to naive intuition,
the ionization rate, analogous to the decay exponent z, is not a monotonically in-
creasing function of the field strength. In many instances the increment of the field
strength actually leads to a decline of the ionization rate. The reasons behind the
finding are precisely what has been described above, namely the complicated meta-
morphosis patterns in the phase space.

5 In higher-dimensional scattering systems, the energy surface is no longer isolated into regions
enclosed by KAM surfaces. The chaotic set in this case forms a single integrated component on
which a typical particle can execute Arnol’d diffusion. If the energy surface is unbounded, the par-
ticle decay still obeys the algebraic law [192]. The characteristic behavior of Arnol’d diffusion in
which a particle hops from one well-defined region in the phase space to another closely resembles
the penetration of Cantori in two-dimensional systems. This is expected to lead to multiple decay
exponents measured over short intervals of time.
230 6 Chaotic Scattering

6.6 Effect of Dissipation and Noise on Chaotic Scattering

An aspect of chaotic scattering concerns about the effect of weak dissipation

[534, 702, 704]. The influence of dissipation can be quite different for hyperbolic
and nonhyperbolic chaotic scattering. In particular, weak dissipations have negli-
gible effect on the physical observables in hyperbolic chaotic scattering, such as
scattering functions, since hyperbolic dynamics are structurally stable (Chap. 1). For
nonhyperbolic dynamics, qualitatively different behavior can arise due to weak dis-
sipation. For instance, marginally stable periodic orbits in KAM islands can become
stable attractors, turning their nearby phase-space regions into the respective basins
of attraction [239, 240]. This means that part of phase-space region that contains
the previous chaotic saddle now becomes part of the basins of the attractors. For
scattering dynamics, the converted subset contains orbits of the previous invariant
set in the neighborhood of the KAM islands. Due to the existence of dense orbits
in the original chaotic saddle, the invariant set remains in the basin boundaries of
the periodic attractors. Consequently, the new invariant set is the asymptotic limit
of the boundaries between scattered and captured orbits. Chaos thus occurs on a
dissipative chaotic saddle whose stable manifold becomes the boundary separating
the basins of the attractors and those of the scattering trajectories. This reasoning
suggests that the structure and the meaning of the Cantor set of singularities is fun-
damentally altered: in successive steps, a constant instead of a decreasing fraction
in the middle of each interval is removed (Sect. 6.4.3). As a result, the scattering
dynamics becomes hyperbolic with exponential decay. The dimension of the Can-
tor set immediately decreases from unity as a dissipation parameter is turned on.
The appearance of attractors accompanied by such a metamorphosis of the chaotic
saddle can occur for arbitrarily small dissipation.
The following dissipative version of the map given by (6.27) has been useful for
providing insights into the effect of weak dissipation on chaotic scattering [534]:

x λ [x − (x + y)2/4 − ν (x + y)]
M = , (6.42)
y λ −1 [x + (x + y)2/4]

where ν ≥ 0 is the dissipation parameter. Figure 6.37a shows the stable manifold of
the nonhyperbolic saddle in the scattering region. When weak dissipation is present
(ν > 0), the fixed point in the center of the island becomes an attractor. The basin
of attraction of this attractor “captures” the island itself and orbits close to the
stable manifold of the previously existing chaotic saddle, as shown in Fig. 6.37b.
The intricate character of the basin of attraction with apparent fractal boundaries
comes from the points of the invariant set that are arbitrarily close to the island
for ν = 0. The newly created basin of attraction contains these points and hence
all their preimages as well. These preimages extend in the phase space along the
original stable manifold of the chaotic saddle, leading to the large-scale structure of
the boundary that mimics the original stable manifold (Fig. 6.37a versus Fig. 6.37b).
Because of this similarity, delay-time functions, which are physically measurable,
resemble each other in both the conservative and the weakly dissipative cases, as
6.6 Effect of Dissipation and Noise on Chaotic Scattering 231

Fig. 6.37 For the scattering map (6.42) for λ = 4.0, phase-space structure and delay-time func-
tion: (a) for ν = 0, KAM island (gray), scattered orbits (blank), and the stable manifold of the
nonhyperbolic chaotic saddle (black); (b) for ν = 0.01, captured orbits (black) and scattered orbits
(basin of allocation) (blank). The plus sign represents the fixed-point attractor. (c), (d) Delay-time
function in the conservative
 and dissipative cases (a) and (b), respectively, where n is the time
taken by particles to reach x2 + y2 ≥ 100 [534] (copyright 2002, the American Physical Society)

Fig. 6.38 Survival

probability P(n) for the map
(6.42) for λ = 4.0, ν = 0.001
in the interval [x0 , x0 + 10−7 ],
y0 = −2, x0 = 0.5760006
[534] (copyright 2002, the
American Physical Society)

shown respectively in Fig. 6.37c,d, where the delay times of particles launched from
the horizontal line y = −2 toward the scattering region are plotted against their x-
coordinates on the line [534].
In the presence of a weak dissipation, the decay becomes strictly exponential,
as shown in Fig. 6.38. The escape rate is the same as that of the intermediate-
time decay in the conservative case: κ ≈ 0.08. The original algebraic decay in the
232 6 Chaotic Scattering

conservative case is destroyed by the dissipation because orbits with points close to
the island, which otherwise would be stuck, are captured by the periodic attractor.
The escape rate in general changes under further increases of the dissipation. (In
the hyperbolic region (λ ≥ 6.5) the decay is always exponential. For λ = 8.0, for
example, κ remains essentially constant and is equal to 0.9 in the range 0 ≤ ν ≤
0.01 [534].)
The uncertainty algorithm can again be used to compute the box-counting di-
mension ds of the set of singularities in the scattering functions. In Sect. 6.4.3, we
have argued that ds = 1 when the map (6.42) is nonhyperbolic and conservative.
A technical point about the numerical evaluation of the dimension in this case, as
discussed in Sect. 6.4.3, is that the result converges slowly to unity. When a small
amount of dissipation is present, ds + 1 is the dimension of the boundary between
the scattered and captured dynamics. The numerical convergence of ds is in this case
faster and essentially independent of the resolution. For example, for λ = 4.0 and
ν = 0.01, the dimension is found to be ds ≈ 0.8 [534], a well-convergent value as
the length scale is reduced over six orders of magnitude. As the dissipation param-
eter is increased, the rate of decrease in the dimension is relatively large initially,
but as the parameter passes through a small critical value νc , the rate is reduced
significantly, and becomes nearly zero for ν > νc . This crossover behavior appears
quite general [704]: as the dissipation parameter is increased from zero, attractors
are constantly created as the system becomes more dissipative (for closed systems
see [239]). When most of the attractors that the system is capable of having have
already been created as the critical dissipation value is approached, the decrease in
the dimension slows.
The addition of weak noise has a qualitatively similar effect on nonhyperbolic
scattering to that of dissipation. Very weak noise leads to a slower algebraic de-
cay than in the deterministic case [20]. Stronger noise destroys KAM islands and
stickiness is no longer present. As a result, the asymptotic decay law changes from
algebraic to exponential [705]. The escape rate has been found in such situations to
increase with the noise amplitude [644,703], preceded in some cases by a local min-
imum [20]. All these studies suggest that the algebraic decay law, regarded to hold
universally in nonhyperbolic chaotic scattering, is apparently structurally unstable
against dissipation and noise.

6.7 Application of Nonhyperbolic Chaotic Scattering: Dynamics

in Deformed Optical Microlasing Cavities

Optical processes in microcavities occur in important applications such as microdisk

semiconductor lasers [119] and optical fiber communication [570], in which total
internal reflection of light is exploited to achieve nearly perfect mirror reflectivity.
Dielectric cavities (cylinders or spheres) are a common type of optical microcavity.
In such a situation, ideally the surface of the cavity confines certain modes of the
electromagnetic field, such as the “whispering gallery” (WG) modes [176,509,618],
defined as the state in which light circulates almost tangent to the surface of the
6.7 Dynamics in Deformed Optical Microlasing Cavities 233

cavity via total internal reflection, suffering minimal loss caused by evanescent
leakage and scattering due to surface roughness. The ray dynamics inside dielec-
tric cavities is governed by the laws of geometric optics: upon collision with the
boundary a ray generally splits into a reflected ray and a transmitted ray.
A fundamental quantity characterizing a microcavity is the quality factor, or the
Q-value, defined as Q = ωτ , where ω is the frequency of the resonant mode and
τ is its lifetime in the cavity [265, 519, 552, 789]. If there are no deformations in
the cavity geometry from the ideal shape of cylinder or sphere, in a practical sense
light in WG modes can be trapped in the cavity for arbitrarily long time, making the
Q-value of the cavity high. This is the principle on which the world’s smallest lasers
were fabricated [176, 509].
While a circular symmetry allows WG modes to have high Q-values, it prevents
the laser emission from having a good directionality. Asymmetric resonant cavities
(ARCs) with smooth deformations from the circular symmetry were then suggested
[265, 519, 552]. Such deformations can be quite large, ranging from 1 to 50% with
respect to the corresponding circular geometry. Although WG modes of a spherical
or of a cylindrical cavity can be treated analytically and the effect of small deforma-
tions can be analyzed using the traditional wave-perturbation theory, it is difficult to
study cavities with large deformations, since the modes of highly deformed cavities
are not perturbatively related to those of the circular cavities. A question is then
whether high-Q modes exist in highly deformed cavities. The pioneering works in
[265, 519, 552] showed that for dielectric materials with low index of refraction
(n < 2, such as glass fibers or cylindrical dye jets, assuming that the surrounding
medium has n0 = 1), if the cavity surface remains convex, high-Q WG modes can
still exist. This important result is obtained by studying chaotic dynamics resulting
from classical ray tracing. Specifically, by treating waves propagating in ARCs as
light rays bouncing within the cavity, the problem of ARCs becomes that of classi-
cal billiards. Far-field emission was found to be peaked in specific directions even
in cavities where the reflected rays have uniformly chaotic dynamics; see Fig. 1.21.
It was demonstrated that the Q-value and the directionality of an ARC can be com-
puted directly from properties of open chaotic ray dynamics, such as the decay
law, which are found to be in good agreement with experimental measurements
[265, 519, 552]. It was also demonstrated both experimentally and computationally
that for high-index semiconductor materials (index of refraction n > 2), WG modes
may not be relevant to the lasing properties of the cavities [265]. Instead, resonant
modes of “bow-tie” shapes are found to be responsible for the laser performance in
the presence of large geometric deformations.
The scope of discussion here will be restricted to low-index dielectric lasing cav-
ities, for which the relevant dynamics is that of WG modes. To illustrate the role of
nonhyperbolic chaotic scattering, we shall focus on the following question: in or-
der to achieve both high-Q values and a good directionality, what is the maximally
allowed amount of deformation from circular symmetry? The question is particu-
larly relevant to the practical design of microdisk semiconductor lasers, where it
is desirable to know the upper bound of the allowed deformation. It is convenient
to choose the well-studied example of a two-dimensional (cylindrical) resonator
with quadrupolar deformation from the circular boundary, and investigate the ray
234 6 Chaotic Scattering

dynamics in the resonator from the standpoint of chaotic scattering by focusing on

the decay property of trajectories in the phase space. Here, decay is due to the fact
that light can be refracted, and then it leaves the cavity. The WG modes thus corre-
spond to chaotic-scattering trajectories in the phase space.

6.7.1 Dynamical Criterion for High-Q Operation

Qualitatively, the interplay between the amount of deformation and the algebraic-
decay exponent can be seen as follows. For a given value of the refraction index
n and circular symmetry, the WG-mode operation stipulates that the angle α of
incidence (Fig. 6.39) satisfy sin α < sin αc , where αc is a critical angle. For cavities
with small deformations, the range of the angle α is small as the corresponding
ray circulates near the boundary, and hence it is relatively easy for the condition
sin α < sin αc to be satisfied, leading to a high probability of light rays being trapped
inside the cavity. The decay exponent z is expected to be small in such cases. Large
deformations from the circular symmetry give rise to a large range of the angle α ,
and consequently, it is more likely for the condition of total internal reflection to
be violated, leading to large values of z. Let ε be the parameter characterizing the
amount of deformation. The algebraic-decay exponent z is a nondecreasing function
of ε . A criterion to determine the maximally allowed amount of deformation can be
derived by noting that the average lifetime τ of very long lived trajectories can be
written as
 ∞

t 2−z , if z > 2,
τ∼ t ptr (t)dt ∼ c (6.43)
tc ∞, if z < 2,
where ptr (t) is the escape-time distribution (6.21) for particles initiated around
KAM tori and tc is the time of the onset of the algebraic decay. If z < 2, the average
lifetime diverges asymptotically, indicating that in any practically long time scales, a

Fig. 6.39 Variables for constructing a Poincaré section for tracing ray dynamics in a two-
dimensional cavity [486] (copyright 2002, the American Physical Society)
6.7 Dynamics in Deformed Optical Microlasing Cavities 235

high value of Q can be expected. If in the corresponding range of deformation, a high

directionality can be maintained (to be discussed below in numerical experiments),
then the criterion for determining the upper bound of ε can be set as ε < εc , where
z(εc ) = zc = 2.

6.7.2 A Numerical Example

The following class of two-dimensional cavities, expressed in polar coordinates

(r, θ ) with quadrupolar deformation characterized by ε , was considered in [486]:
1 + ε cos 2θ
r(θ ) =  . (6.44)
1 + ε 2/2
Classical ray tracing can be done using the Poincaré map defined with respect to
the angles [59] ψ , α , and θ , as shown in Fig. 6.39. The map can be written in the
following implicit form, relating the dynamical variables at successive total internal
reflections off the boundary of the cavity:
(1 + ε cos 2θt+1 ) sin θt+1 − (1 + ε cos 2θt ) sin θt
tan (ψt + αt ) = ,
(1 + ε cos 2θt+1 ) cos θt+1 − (1 + ε cos 2θt ) cos θt
2ε sin θt+1 sin 2θt+1 − cos θt+1 (1 + ε cos 2θt+1 )
tan ψt+1 = ,
sin θt+1 (1 + ε cos 2θt+1 ) + 2ε cos θt+1 sin 2θt+1
αt+1 = ψt+1 − ψt − αt , (6.45)

where t is the discrete-time index denoting the event of bounce of a light ray off
the cavity boundary. In numerical experiments, the critical angle of incidence is
fixed to be αc = 47.3 degrees, which corresponds to cavities with refraction index
n ≈ 1.475 (cos αc = 1/n). Figure 6.40a shows, for ε = 0.1, a representative phase-
space structure of the map in (6.45). There are both KAM tori and chaotic regions.
While the map in (6.45) describes the dynamics of particles in a closed billiard, im-
posing the threshold line at αc makes the system effectively open (a leaked billiard).
Figure 6.40b shows, in the two-dimensional physical space (x, y), a typical scat-
tering trajectory in a WG mode and its escape from the cavity after about 1,000
bounces. Figure 6.40c shows the delay time T measured in the number of reflec-
tions of light rays in the cavity as a function of the initial angle of incidence α0 . If
the trajectory lives on a KAM torus, the time is infinite, and if the ray is in a chaotic
region, it will eventually escape but the time it stays in the cavity can be long. Since
WG modes correspond to trajectories in the open chaotic region above the binding
KAM tori, the Q-value of the cavity in the WG modes is determined by motions of
rays in the chaotic region. The emission of light rays in the WG modes apparently
has a high degree of directionality, as shown in Fig. 6.40d, a histogram of the emis-
sion angle θout , where θout is defined to be the angle of the refracted exiting light
ray with respect to the x-axis.
236 6 Chaotic Scattering

Fig. 6.40 For a dielectric cavity of refraction index n ≈ 1.475 with quadrupolar deformation of
ε = 0.1: (a) phase-space structure with perfect internal reflection (the dashed horizontal line
indicates the value of sin αc ; points above this line are considered to have escaped), (b) behavior
of light ray (in a WG mode) and its escape from the cavity, (c) delay time of light rays versus the
initial angle of incidence, and (d) histogram of the emission angle θout . The highly localized pat-
tern in the histogram indicates a high degree of directionality of the emitting light. A set of 1,000
initial conditions chosen uniformly from α0 ∈ [0.05, π /4] were utilized [486] (copyright 2002, the
American Physical Society)

In the presence of deformation, the probability for a light ray to survive in the
scattering region decays algebraically with time, and the value of the decay expo-
nent z of the probability density p(t) increases as the deformation becomes large.
Figure 6.41a shows that the algebraic exponent increases monotonically as the de-
formation parameter ε is increased. For ε < εc ≈ 0.22, the exponent remains below
the critical value zc = 2, indicating that for 0 < ε < εc , the average lifetime of light
rays in the cavity diverges and the Q-value of the cavity is high in a statistical sense.
In contrast, for ε > εc , the algebraic-decay exponent is above 2, implying relatively
low Q-values. Thus, in order to achieve a high-Q operation, the amount of deforma-
tion should not exceed the value of about 0.2.
While high-Q operation of the cavity is desired, an equally important measure
is the directionality of light emission. To quantify this, consider the worst case, in
which light is emitted equally probably in all directions. The probability distribution
P(θout ) of the emission angle of the exiting light ray is thus P(θout ) = 1/(2π ) for
0 ≤ θout ≤ 2π . For the computed distribution as in Fig. 6.40d, the height of the distri-
bution P(θout ) can be normalized to 1/(2π ) and the total area under the distribution
curve is
 2π
A(ε ) = P(θout )dθout ,
0
6.7 Dynamics in Deformed Optical Microlasing Cavities 237

Fig. 6.41 For a cavity of refractive index n ≈ 1.475 with quadrupolar deformation, (a) algebraic-
decay exponent z versus the deformation parameter ε . High-Q lasing operation can be expected for
ε < εc ≈ 0.22, (b) Measure μ of directionality versus ε . Apparently, a high degree of directionality
can be maintained in the range of high-Q operation. (c), (d) position angle θ versus the emis-
sion angle θout for ε = 0.1 and 0.001, respectively [486] (copyright 2002, the American Physical
Society)

where a unit area indicates uniform emission. The following measure of

directionality can then be defined:

μd (ε ) = 1 − A(ε ), (6.46)

where a high value of μd signifies a high degree of directionality (for uniform emis-
sion, μd = 0). Figure 6.41b shows, for the particular cavity in (6.44), μd versus ε .
Apparently, in the range of high-Q operation (ε < εc ), μd ≈ 1, indicating a high
degree of directionality. It was also found [486] that in the range of the deformation
parameter where light emissions possess a high degree of directionality, the rays ap-
pear to exit the cavity at only a few locations on the boundary. At each exiting point,
the range of the emission angle θout is highly localized. This behavior is shown in
Fig. 6.41c, where the angle θ that defines the position of a boundary point (posi-
tion angle) versus the emission angle θout is plotted for the escaping light rays for
ε = 0.1. In contrast, when the deformation is near zero, light rays can exit from al-
most anywhere on the boundary (0 ≤ θ < 2π ), which means that for the WG-mode
operation, light can be emitted in almost every possible direction: 0 ≤ θout < 2π , as
shown in Fig. 6.41d. That high-Q operation and high degrees of directionality can be
realized simultaneously in deformed cavities can be potentially useful for practical
design and fabrication of microlasing cavities [183, 474, 660, 699].
238 6 Chaotic Scattering

In generic situations the phase space is mixed and also contains hyperbolic
regions. If the region where emissions take place, i.e., the leak region from the point
of view of energy loss, happens to fall far from large KAM surfaces, the total en-
ergy E(t) inside the cavity decays exponentially over intermediate time scales. After
some short time t ∗ one can write E(t) ∼ (1 − r)t , where r is the leakage rate, i.e.,
the emitted energy per unit time. The energy decay rate κE governing the law

E(t) ∼ e−κE t (6.47)

can then be expressed as

κE = − ln (1 − r). (6.48)
The leakage rate has recently been shown by Altmann [15] to be the integral of
the emission rate T of the light ray over the leak region taken with respect to the
conditionally invariant measure of the leaked billiard dynamics:

r= T d μc . (6.49)
I

Relations (6.48) and (6.49) are analogous to relation (2.98), which determines the
escape rate of leaked dynamical systems. The crossover to the algebraic decay takes
over at some time tc , which is expected to scale with the reciprocal of the decay
rate κE , in analogy with (6.30). A detailed investigation also shows [15] that the
localization of the far-field emission in specific directions is related to the unstable
manifold of the chaotic saddle in the leaked system. The energy inside the cav-
ity is distributed according to the c-measure, which is nonzero along the unstable
manifold of the chaotic saddle. Strong intensity can therefore be observed only in
points along the perimeter of the cavity that are on the unstable manifold of the
leaked problem. For long times, the exponential decay crosses over to a power-
law decay governed by the nonhyperbolic component of the chaotic saddle, but the
directionality does not change qualitatively. The reason is that rays escape the non-
hyperbolic component through the hyperbolic regions. The unstable manifold of the
nonhyperbolic component of the chaotic saddle therefore follows closely that of the
hyperbolic component [19].
Chapter 7
Quantum Chaotic Scattering and Conductance
Fluctuations in Nanostructures

This chapter is devoted to the manifestation of classical chaotic scattering in the

quantum world. The major characteristic that distinguishes a quantum system from
its classical counterpart is that in quantum mechanics, the system is characterized
by a nonzero value of the Planck constant. Let h̄ denote the Planck constant nondi-
mensionalized by normalizing to characteristic length and momentum values, so
that h̄ → 0 corresponds to the classical limit, h̄ 1 to the semiclassical regime, and
h̄ ∼ 1 to the fully quantum-mechanical regime. To study the quantum manifestation
of classical Hamiltonian chaos, the semiclassical regime is of particular importance
because this is the regime in which both quantum and classical effects are rele-
vant. In particular, we shall be interested in signatures of chaotic scattering when
the same system is treated quantum-mechanically in the semiclassical regime. The
mathematical methods needed to study the semiclassical regime differ from those
used so far. This chapter is therefore of different character than the others. Our aim
is to flesh out the most important phenomena only, where fingerprints of the classi-
cal transient chaos appear at the semiclassical level, motivating the reader to pursue
more detailed studies.
The fundamental quantity characterizing a quantum scattering system is the
scattering matrix, or the S-matrix, whose elements are the transition probabilities
between quantum states of the system before and after the scattering. The formu-
lation of the S-matrix in terms of classical quantities had been of great interest in
chemical physics even before chaos started attracting wide attention. The seminal
contribution by Miller [525], who obtained a formula for S-matrix elements in terms
of purely classical quantities in the semiclassical regime for reactive scattering sys-
tems, becomes the fundamental tool in the study of quantum chaotic scattering.
Given a system that exhibits chaotic scattering in the classical limit, the S-matrix
elements in the semiclassical regime exhibit random fluctuations as some physi-
cal parameters of the system, such as the energy of the scattering particles or the
strength of some externally applied magnetic field, change in a classically small
but quantum-mechanically large range. We shall see that depending on whether the
classical scattering is hyperbolic or nonhyperbolic, the statistical properties of the
fluctuations in the S-matrix elements can be quite distinct.
An important class of systems in which the semiclassical theory of chaotic scat-
tering finds direct application is semiconductor nanostructures, and particularly

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 239
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 7,

c Springer Science+Business Media, LLC 2011
240 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

quantum dots. In such a system, electrons are restricted to a plane near the interface
between two different semiconductors. Applying voltage to contact gates deposited
above the junction allows for the construction of submicron-sized, two-dimensional
cavities in which electrons are scattered. Furthermore, at millikelvin temperatures,
in such a system the mean free path and the coherence length are typically much
larger than the cavity length. For low currents, the transport characteristics are de-
termined by the approximately ballistic and coherent motion of electrons in the
cavity. As a result, one can expect that the classical orbital dynamics of electrons,
whether regular or chaotic, will play a major role in the transport. Indeed, for a two-
dimensional electron-gas quantum dot that exhibits hyperbolic chaotic scattering
classically, important physical properties of the system, such as various conduc-
tances, show random fluctuations that are absent when the underlying classical
dynamics is regular.
Realistic quantum dots have nonhyperbolic classical scattering dynamics with
chaotic sets coexisting with nonescaping KAM islands surrounding stable orbits in
the phase space. For such a system, the conductances show strong regular fluctu-
ations as some external parameter is varied. The origin of the regular conductance
fluctuations is quantum-dynamical tunneling into the KAM islands, and the domi-
nant frequency of the fluctuation pattern can be predicted by the semiclassical theory
of tunneling. Dynamical tunneling has in fact become an active area of research not
only in condensed-matter physics, but also in atomic and optical physics, which is
illustrated by the example of scattering echoes. Leaked quantum systems are also
discussed.
The materials of this chapter require preliminary knowledge about quan-
tum mechanics. The semiclassical treatment of scattering processes is reviewed
in Appendix C, while the concept of scattering cross sections is discussed in
Appendix D.

7.1 Quantum Manifestation of Chaotic Scattering

The S-matrix elements characterize the transition probabilities between two asymp-
totic states. We consider the fluctuations of the S-matrix elements with energy. The
treatment here follows that of Blümel and Smilansky [76,77]. For convenience, let I
and I be a pair of initial and final states (action variables) of the scattering process.
The energy fluctuations of the S-matrix elements can be examined by the following
autocorrelation function:
∗
CII (ε ) = SII (E)SII (E + ε )E , (7.1)
where SII (E) and SII (E + ε ) are the S-matrix elements between the initial and
the final states at energy E and E + ε (ε small compared with E), respectively,
the asterisk represents the complex conjugate, and the average is over a classi-
cally small but quantum-mechanically large energy interval ΔE. Miller’s theory
7.1 Quantum Manifestation of Chaotic Scattering 241

(see Appendix C, (C.1)) expresses the semiclassical S-matrix elements in the action-
angle representation as
   
1 1/2  ∂ I −1/2 iΦ̄ (s) (I , I) π
SII ≡ I |S|I =
2π ∑  ∂ θ  exp h̄
− iνs ,
2
(7.2)
s (s)

where Φ̄ (s) (I , I) is the dimensionless classical action along paths s, and νs is the
Maslov index (Sect. C.2). The action variable I itself is also dimensionless. The
(s)
quantity |2π∂ I /∂ θ |−1 (s)
is the classical transition probability pI,I for trajectories
with initial and final action-angle variables I, θ and I , θ (C.2). Substituting (7.2)
into (7.1), concentrating on the energy-dependence of the action, and approximating
Φ̄ (s) (E + ε ) ≈ Φ̄ (s) (E) + ε∂ Φ (s) /∂ E, (7.3)
we obtain
   
iε ∂ Φ̄ (s)
∑ ∑ [pII (E)pII
(s) (s) (s )
CII (ε ) ∼ pII (E) exp + (E)]1/2
s h̄ ∂ E s=s
E
 
−i (s) iπ (νs − νs )
exp [Φ̄ (E) − Φ̄ (s ) (E + ε )] + . (7.4)
h̄ 2
E

The phase factor in the second term in (7.4) is mainly determined by the differences
in the classical actions Φ̄ (s) from different trajectories s and s . Since the energy
interval ΔE over which the summations are carried out is chosen to be quantum-
mechanically large, in the semiclassical regime (h̄ 1) the double summation in
the second term is essentially zero. The summation in the first term in (7.4) involves
the partial derivative of the classical action with respect to the energy, ∂ Φ̄ (s) /∂ E,
which is the time it takes for a trajectory to travel from the initial to the final state.
This time is in fact the delay time t (s) associated with this trajectory:
∂ Φ̄ (s)
= t (s) . (7.5)
∂E
Thus the summation in the first term of (7.4) can be regarded as being with respect
to this time. For chaotic scattering the distribution of the delay time can essentially
be from zero to infinity. It is thus reasonable to replace the summation by an integral
with respect to the delay time. This gives

iε t
CII (ε ) ≈ dtpII (E,t)E exp , (7.6)
h̄
which is the Fourier transform of the classical probability pII (E,t) that a transition
I → I occurs when the delay time of the trajectory is in the interval [t,t + dt]. This
is, however, the escape-time distribution (Sect. 1.2.1)
pII (E,t) ∼ p(t) (7.7)
at constant energy E. The main features of the autocorrelation function are thus
independent of I, I , and we can suppress these indices in CII (ε ).
242 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

7.2 Hyperbolic Chaotic Scattering

7.2.1 Autocorrelation of the S-Matrix Elements

As discussed in Chap. 6, for hyperbolic chaotic scattering, the probability pII (E,t)
is independent of the initial and the final states I and I , and it decays exponentially
with the escape rate κ at energy E:

pII (E,t) ∼ exp [−κ t].

Since the energy range ΔE is chosen to be classically small, the escape rate κ (E)
can be regarded as constant in ΔE. Substituting pII (E,t) into (7.6) gives

C(0)κ
C(ε ) = . (7.8)
κ − iε /h̄

The magnitude squared of the autocorrelation function is a Lorentzian of width h̄κ :

|C(0)h̄κ |2
|C(ε )|2 = . (7.9)
(h̄κ )2 + ε 2

It is thus the classical escape rate that determines the half-width of the semiclassical
autocorrelation function of the S-matrix elements. As a consequence,

d|C(ε )|2 
= 0, (7.10)
dε ε =0

that is, the autocorrelation function is flat near ε = 0, indicating that the statistical
fluctuations of the S-matrix elements with energy are relatively mild. The Lorentzian
form of the autocorrelation function of the S-matrix elements has been evidenced in
several cases. An example is shown for the microwave experiment of Doron, Smi-
lansky, and Frenkel (see inset of Fig. 1.20 in Chap. 1), where the Lorentzian curve
with the predicted width is displayed. More recent experiments with microwaves
[427, 493, 576] found and even better agreement between the observed data and the
semiclassical theory.

7.2.2 S-Matrix in the Time Domain

Instead of considering the scattering process as a wave phenomenon of fixed energy

E, one can consider a narrow wave packet and follow its motion in time. The two
approaches are related: the S-matrix S(t) in the time domain is the Fourier transform
of S(E).
7.2 Hyperbolic Chaotic Scattering 243

Consider the magnitude squared of the Fourier transform S(t) of S(E):

  2 
 
| S(t) |2 =  dESII (E)e−iEt/h̄  , (7.11)
II

where the average is taken over the matrix elements of the S-matrix. Using (7.2) and
expressing the magnitude squared of the integral in (7.11) as a double integral, we
have
 2 
 
∑[pII (E)pII
 dESII (E)e−iEt/h̄  = dEdE (s) (s )
  (E )]1/2
s,s
 !
Φ̄ (s) (E) − Φ̄ (s ) (E ) νs − νs E −E
exp i −π −t .
h̄ 2 h̄

When averaging over many matrix elements in the semiclassical regime, the double
sum with off-diagonal elements s = s is essentially zero, so we can concentrate on
the diagonal elements


 !
Φ̄ (s) (E) − Φ̄ (s) (E ) E −E
∑
(s)
dEdE pII (E) exp i −t . (7.12)
s h̄ h̄

Since the dominant contribution comes from the regime in which E − E is small,
we use (7.3) and (7.5) again to obtain

∑ pII (E) exp [i(t (s) − t)(E − E )/h̄],
(s)
dEdE (7.13)
s

where t (s) is the delay time for trajectory s. The integral over E − E is proportional
to δ (t − t (s)). We have
| S(t) |2 ∼  pI,I (E,t)II . (7.14)
The magnitude squared of the Fourier transform of S(E) with respect to the energy is
proportional to the escape time distribution, and decays as exp (−κ t), with κ being
the classical escape rate at energy E. This property has also been demonstrated in
the microwave experiment shown in Fig. 1.20.

7.2.3 Relation to Orthogonal Ensemble of Random Matrices

Another correlation function is

∗
FII (η ) = SII (E)SI(I +η ) (E){I} , (7.15)
244 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

where η denotes the action difference in two final states at energy E and ·{I} is
the average over a classically small but quantum-mechanically large domain of the
initial and final states. Utilizing similar reasoning as for the energy autocorrelation
function and making use of the fact that the partial derivative of the classical action
integral Φ̄ (s) with respect to I is the angle variable, Blümel and Smilansky obtained
[76, 77]
   
iη ∂ Φ̄ (s) iηθ
FII (η ) ∼ ∑ pII exp
(s)
∼ dθ pII (E, θ ){I} exp ,
s h̄ ∂ I h̄
{I}
(7.16)
where θ is the classical angle between the incoming and outgoing scattering tra-
jectories, i.e., the deflection angle, and pII (E, θ ) is the probability that the angle
is in a small range [θ , θ + dθ ] for energy E. For hyperbolic chaotic scattering,
pII (E, θ ) typically peaks at some mean angle θ0 and decays exponentially as θ
deviates from θ0 :
pII (E, θ ) ∼ exp (−α |θ − θ0 |),

and it is independent of I and I . These considerations lead to

F(0)
F(η ) = , (7.17)
1 − iη /(h̄α )

indicating that the magnitude squared of F(η ) also has a Lorentzian shape.
Since η is the difference between the action variables of final states after the
scattering, in quantum mechanics it is an integer multiple of the Planck constant h̄.
In a classically small but quantum-mechanically large range, the integer η /h̄ can
be large. Thus, if α is not too large we have |F(η )| ∼ 0, indicating that the S-
matrix elements are uncorrelated with respect to the final state. Since the S-matrix is
unitary and symmetric, its properties are completely determined by those of Dyson’s
orthogonal ensemble of random matrices [76, 77, 208]. Some important properties
of the ensemble are the following:
(a) The nearest-neighbor distribution of the eigenvalues of the S-matrix on the unit
circle is a Wigner distribution.
(b) At a given energy, the distribution of the magnitude squared of the S-matrix
elements is Poissonian.
(c) Because of (b), i.e., the probability of SII being small is large, the energy de-
pendence of the scattering cross section σII (E) (for a given transition) exhibits
a characteristic fluctuation pattern: the Ericson fluctuations [224].
(d) Because of the Lorentzian autocorrelation function of the S-matrix elements,
the autocorrelation function of the scattering cross section σII (E)σII (E + ε )E
should also be a Lorentzian of the same width h̄κ , set by the classical escape
rate.
These properties have been verified in several studies (see, e.g., [75, 76]).
7.3 Nonhyperbolic Chaotic Scattering 245

7.3 Nonhyperbolic Chaotic Scattering

We have seen in Chap. 6, (6.20), that for classical nonhyperbolic chaotic scattering,

pII (E,t) ∼ t −(z+1), for large t, (7.18)

where the constant z is the algebraic-decay exponent, and the form is independent
of I and I . A natural question is how nonhyperbolic chaotic scattering manifests
itself quantum-mechanically. This question was addressed [442] with the result that
the fine-scale semiclassical quantum fluctuations of the S-matrix with energy can be
enhanced in the nonhyperbolic case as compared to the hyperbolic case.
In the semiclassical regime, the energy autocorrelation function and the particle
decay probability are Fourier-transform pairs, (7.6), regardless of the nature of clas-
sical scattering (i.e., hyperbolic versus nonhyperbolic). Since the algebraic decay
law is valid for large t, to obtain the energy autocorrelation function of the S-matrix
elements, we rewrite pII (E,t) as

g(t), 0 ≤ t ≤ T,
p(E,t) = (7.19)
at −(z+1), t > T,

where T is a large fixed value and g(t) is a smooth function. Substituting (7.19) into
(7.6) yields
 T  ∞
iε t iε t
C(ε ) ∼ g(t) exp dt + at −(z+1) exp dt
0 h̄ T h̄
  ∞

x̄ −(z+1)
h̄ xh̄
= g(xh̄/ε )eix dx + a eix dx
ε 0 x̄ ε
= C0 + C1 ε z , (7.20)

where the change of variable ε t/h̄ = x (x̄ ≡ ε T /h̄) and the fact that x̄ is small for
small ε have been used, and C0 = C(0) and C1 are constants. This form of the energy
autocorrelation function is characteristically different from the Lorentzian function
in (7.9). In particular, the derivative of C(ε ) with respect to the energy variation ε is

dC(ε )
∼ ε z−1 . (7.21)
dε
In two-degree-of-freedom Hamiltonian systems, the classical algebraic-decay ex-
ponent z may fall in the range below 1 (see Fig. 6.31). In these cases (7.21) implies
that the derivative dC(ε )/dε can be arbitrarily large for arbitrarily small values of ε .
That is, for nonhyperbolic chaotic scattering, the energy autocorrelation function
can exhibit a cusp for small energy difference ε . As ε is increased, C(ε ) decreases
rapidly from C0 .
246 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

The above argument for the cusp behavior in C(ε ) cannot be valid for
infinitesimal values of ε because this corresponds to the infinite time limit of the
semiclassical approximation. Strictly speaking, the cusp behavior holds only in
the range εmax > ε > εmin , where εmax ∼ h̄/T and T is the classically determined
time in (7.19). The cutoff of the cusp behavior at small energy difference εmin oc-
curs because the semiclassical approximation breaks down for large times t > tmax ,
which corresponds to εmin ∼ h̄/tmax . This can be understood by noting that the alge-
braic decay results from particle orbits that spend long stretches of time near KAM
surfaces. Orbits contributing to longer times penetrate more deeply into the sticky
region near a KAM surface. At finite wavelengths, there comes a time past which the
phase space structures resolved by classical orbits still in the scattering region are
quantum-mechanically not resolved. This determines the crossover time tmax .1 Thus,
the algebraic-decay exponent appearing in (7.21) is the one characterizing long but
finite time intervals. As Fig. 6.31 shows, this exponent can take on different values.
The cusp behavior in the energy autocorrelation function for nonhyperbolic
chaotic scattering can be numerically verified [442] by the Troll–Smilansky model
defined by (6.34) and (6.35), with the potential function V (r) chosen as the Woods–
Saxon potential
V0
V (r) = − , (7.22)
1 + exp[(r − R0 )/c]
where V0 > 0 and R0 and c are constants. For the fixed set of parameters V0 = 10,
R0 = 1, c = 0.1, and a = 4, the nonoverlapping condition for adjacent potentials
is satisfied because V (r = a/2)/V0 ∼ 10−5 . When the particle energy is large, the
phase space consists of KAM islands and chaotic regions for E = 10, and the decay
of particles from the scattering region is well fit by an algebraic dependence. As the
particle energy decreases from E = 10 to E = Ecrit , where 1 < Ecrit < 10, the KAM
surfaces are destroyed. At lower values of E, there are no KAM surfaces, and the
decay of particles from the scattering region becomes exponential.
Quantum-mechanically, the effect of any single scatterer is completely speci-
fied once the phase shifts are given. Because of the discrete translational sym-
metry of the system in the y direction, the wave function satisfies ψ (α ) (x, y) =
exp(iα y)ϕ (α ) (x, y), where ϕ (α ) (x, y) is periodic with period a in y. So we have
ψ (α ) (x, y + na) = exp(inα a)ψ (α ) (x, y). Thus the y component of the momentum qn
is quantized: qn = α + 2nπ /a. By the conservation of energy q2n ≤ k2 , where k is the
wave number (E = h̄2 k2 /2), there exists a maximum integer NB (maximum number
of open channels) above which qn is not allowed, which corresponds to evanescent
waves.
To compare the quantum calculation with the semiclassical predictions, it is nec-
essary to use sufficiently small values of Planck’s constant h̄, so that there is a large
number of waves on the scale of the system’s spatial period a. For the nonhyperbolic

1 It has been shown that this time scales with the Planck constant algebraically [242,460]. The scal-
ing law of the crossover time with the Planck constant is fundamental for semiclassical treatment
of classically chaotic systems.
7.4 Conductance Fluctuations in Quantum Dots 247

Fig. 7.1 Magnitude squared

of S-matrix autocorrelation
functions (diamonds for the
nonhyperbolic case (E = 10)
and triangles for the
hyperbolic case (E = 1)) and
the corresponding
semiclassical predictions
(thick solid curves) [442]
(copyright 1992, the
American Physical Society)

case E = 10, h̄2 /2 = 10−3 was chosen [442], so that the corresponding wave vector
is k0 = 100 and there are 127 open channels (128 > 2a/λ > 127). For the case
of hyperbolic scattering at E = 1, h̄2 /2 = 10−4 was chosen (k0 = 100), so the
number of open channels is the same as that in the nonhyperbolic case. The au-
tocorrelation function C(ε ) was computed using an energy width ΔE/E ≈ 0.1.
As a function of energy, an S-matrix element contains both a smooth part and
a fluctuating part. To compare with the semiclassical theory, the smooth part
was subtracted off and only the fluctuating part was kept. Figure 7.1 shows the
magnitude squared of the quantum autocorrelation functions |C(ε )|2 for both the
E = 10 case (nonhyperbolic, plotted as diamonds) and the E = 1 case (hyper-
bolic, plotted as triangles), where an average was performed with respect to a
small block of matrix elements. The C(ε ) curves are, however, essentially in-
dependent of the particular matrix block chosen in the averaging process [442].
Also shown in Fig. 7.1 are the semiclassical predictions from (7.20); the lower
thick solid curve corresponds to E = 10, and the upper thick solid curve corre-
sponds to E = 1. To calculate the semiclassical correlation functions, the classical
decay data were substituted directly into (7.6). For both the hyperbolic and the
nonhyperbolic cases, Fig. 7.1 suggests that the quantum calculation is consistent
with the semiclassical theory for small ε . The energy autocorrelation function
appears to follow the predicted cusp behavior near ε = 0 for nonhyperbolic scat-
tering. This means that in comparison with the hyperbolic-scattering case, the
fine-scale fluctuations of the matrix elements with energy are greatly enhanced in
the nonhyperbolic case.

7.4 Conductance Fluctuations in Quantum Dots

One important area in which the semiclassical theory of quantum chaotic scat-
tering finds application is electronic transport in semiconductor nanostructures.
In semiconductor nanodevices, a basic component is mesoscopic junctions,
248 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

or quantum dots, which connect a number of quasi-one-dimensional electronic

wave guides and control the electronic transport. The size of the junctions is usually
appreciable on quantum scales but small on classical scales, so electron motions
are coherent and ballistic. A mesoscopic junction system is thus equivalent to an
open billiard system, for which chaotic scattering dynamics can be expected. How-
ever, quantum effects are also important. Predictions from the semiclassical theory
of quantum chaotic scattering are then expected to be useful and experimentally
observable. In particular, because of the fine-scale fluctuations of the S-matrix
elements with energy (or magnetic field), various conductances measured from
the junction system can exhibit fluctuations as well. Conductance fluctuations in
mesoscopic junction systems have been an active area of research (for reviews, see
[39, 40, 46, 68, 501]), which has recently been extended to the relativistic quantum
regime in graphene dots [342, 343].
Here, we give a brief introduction to this vast area of research. In particular, we
introduce the basic physics of quantum dots, describe the Büttiker–Landauer for-
mula, which relates conductances to transmission probabilities (S-matrix elements),
and discuss random conductance fluctuations in dots for which the classical scat-
tering dynamics are hyperbolic. The case of conductance fluctuations in typical,
nonhyperbolic junction systems, for which dynamical tunneling into KAM surfaces
becomes important, will be addressed in Sect. 7.5.

7.4.1 Basic Physics of Quantum Dots

We use the GaAs/AlGaAs heterojunction to illustrate the basic physics associated

with two-dimensional electron gas semiconductor quantum dots. In such a junc-
tion, material GaAs is made as pure as possible, but typically it is a weak p-type of
semiconductor. AlGaAs has a wider band gap, so it is effectively an insulator. Mod-
ern molecular-beam epitaxy techniques enable atomically sharp interfaces between
these two materials to grow [238], which are heterojunctions. The two materials
have approximately the same lattice and dielectric constants. It is possible to have a
GaAs layer followed by an almost perfectly matched layer of AlGaAs. The energy
bands of the two materials, before any interactions occur, are shown in Fig. 7.2.
The AlGaAs is doped as n-type. Electrons in the conduction band migrate to fill
the few holes in the top of the GaAs valence band, but most of them end up in
states near the bottom of the GaAs conduction band, as shown in Fig. 7.3. Positive
charges associated with donor impurities attract these electrons to the interface and
consequently bend the energy bands on both sides of the interface in the interac-
tion process. This is the source of an electrical field in the system. The transfer of
electrons from AlGaAs to GaAs continues until a dipole layer, formed of the posi-
tive donors and the negative inversion layer, is sufficiently strong. The dipole layer
gives rise to a discontinuity in the potential that forces the Fermi levels of GaAs and
AlGaAs to be equal, as shown in Fig. 7.3. Because of this, electrons will be confined
in a two-dimensional sheet within about 100 Å of the interface, and consequently,
electron motions in the junction are effectively two-dimensional.
7.4 Conductance Fluctuations in Quantum Dots 249

Fig. 7.2 Energy-band Conduction band

structure of the EF
bottom
GaAs/AlGaAs interface
0.4eV
before any electronic
interactions
AlGaAs
GaAs

2.2eV 1.5eV

+ + + + EF
0.3eV
Valence band
top

Fig. 7.3 Electron

energy-level diagram of a AlGaAs GaAs
GaAs/AlGaAs heterojunction
device. Donor electrons
conduction
occupy the first conduction
band
subband of the potential well. +
conduction
The two-dimensional electron +
band
sheet is confined to within + EF
about 100 Å of the interface
− valence
− band
−
−

valence
band

The GaAs/AlGaAs heterojunction can be patterned by electron-beam lithography

and low-energy ion exposure into miniature wires and junctions [238]. The junction
can be designed and fabricated to have different geometric shapes in a controllable
manner. The width of the wire is on the order of a few micrometers and the electron
mobility is typically larger than 106 cm2 /V · s at millikelvin temperatures. As a con-
sequence of this high mobility, the mean distance le between elastic collision events
due to crystal imperfections and impurities can exceed most experimentally accessi-
ble channel lengths L, so electrons travel from an entrance of the junction to an exit
with a minimal number of collisions. The electron motion in the junction system can
then be regarded as ballistic with conserved momentum. The quantum-dot systems,
besides their importance in nanoscience and technology, are an ideal test bed for the
semiclassical theory of quantum chaotic scattering.
250 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

7.4.2 Büttiker–Landauer Formula

The Büttiker–Landauer formula relates various transmission probabilities (S-matrix

elements) of electronic transport in a quantum dot to physically measurable con-
ductances [107, 108, 468]. (For a comprehensive treatment, see [46, 155].) Consider
a junction of arbitrary shape with four terminals. Under the assumption that elec-
trons move ballistically through the junction, phase-destroying events do not occur.
For generality, assume that an external magnetic field is present so that there is an
Aharonov–Bohm flux φ through the system, as shown in Fig. 7.4. Thus the con-
ductances of the junction are functions of the magnetic field. Alternatively, one
can imagine that the energy of the electrons can be varied, and in this case, the
conductances are functions of the electron energy. The leads to the junction being
connected to reservoirs at chemical potentials μ1 , . . . , μ4 , which serve as sources
and sinks of electrons and energy. In particular, at zero temperature a reservoir of
chemical potential μi feeds the connected lead with electrons up to kinetic energy
μi (i = 1, . . . , 4). Every electron coming through the junction and reaching another
reservoir is absorbed by it, regardless of the energy and the phase of the electron.
It is convenient to assume that the wires connecting the junction and reservoirs are
perfect, i.e., they are strictly one-dimensional quantum channels with two states at
the Fermi energy: one with positive velocity (taken to be the direction away from a
reservoir and toward the junction) and another with negative velocity. Scattering in
the junction is elastic, and inelastic events occur only in the reservoirs.
The elastic scattering properties of the junction are characterized by the probabil-
ities Tij (φ ) for electrons incident in lead j to be transmitted into lead i, and Rii (φ ),
the probabilities that electrons coming into lead i are reflected back into the same
lead. Current conservation and time-reversal invariance in the presence of a mag-
netic flux imply Rii (φ ) = Rii (−φ ) and Tii (φ ) = Tii (−φ ). The potentials μi are rather
arbitrary within a small range at the Fermi energy. Let μ0 ≤ min{ μi } (i = 1, . . . , 4).
Thus, below the potential μ0 , states with positive and negative velocities are filled
and there is no net current flow in each lead. It is necessary to consider only the en-
ergy range Δμi = μi − μ0 above μ0 . Reservoir i injects a current evi ρi Δμi into lead i,
where vi is the electron velocity at the Fermi energy, and ρi = 1/(hvi) is the density

μ3
μ1

μ4
μ2

Fig. 7.4 A general junction of four terminals connected to four reservoirs at chemical potentials
μ1 , μ2 , μ3 , and μ4 , respectively. An Aharonov–Bohm flux φ is applied through the junction
7.4 Conductance Fluctuations in Quantum Dots 251

of states in lead i. The current injected by reservoir i is thus (e/h)Δμi . To be con-

crete, consider the current in one of the leads, say lead 1. A current (e/h)R11Δμ1 is
reflected back to reservoir 1. Each of the reservoirs 2, 3, and 4 contributes a current
−(e/h)T1 j Δμ j ( j = 2, 3, 4) to the total current in lead 1. We thus have

 
e
I1 = (1 − R11)μ1 − ∑ T1 j μ j . (7.23)
h j=1

Because of the charge conservation

1 = Rii + ∑ Ti j (7.24)
j

for any i, the reference potential μ0 is canceled in (7.23). In general, the net current
in any lead i can be written as
 
e
Ii = (1 − Rii)μi − ∑ Ti j μ j . (7.25)
h j=i

As an example, we consider the Hall resistance in the presence of a strong

magnetic field B for a symmetric four-disk billiard junction, as shown in Fig. 7.5.
Because of the surfaces of negative curvature in the junction, classical dynamics
can be chaotic, so this junction system provides a paradigmatic system for studying
quantum manifestations of chaotic scattering. Because of the symmetry, the trans-
mission probabilities are T41 = T34 = T23 = T12 ≡ T (B) for large B, while all other

1 3

Fig. 7.5 A symmetric,

four-disk junction from which
Hall conductance can be
measured 2
252 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

transmission probabilities are zero. From (7.24) one finds that 1 − Rii = T (B) for
any i. The currents in (7.25) are
e
I1 = T (B)[μ1 − μ2 ],
h
e
I2 = T (B)[μ2 − μ3 ],
h
e
I3 = T (B)[μ3 − μ4 ],
h
e
I4 = T (B)[μ4 − μ1 ]. (7.26)
h
In a typical four-terminal experiment, only two of the potentials are measured.
Suppose current flows from lead 1 to lead 3. The potentials measured are μ2 = eV2
and μ4 = eV4 under the condition that the currents I2 and I4 are zero. Equation (7.26)
then implies that μ1 = eV4 , μ3 = eV2 , and

e2
I1 = T (B)[V4 − V2] = −I3 . (7.27)
h
The Hall conductance is defined as
I1
GH = , (7.28)
V4 − V2

which is obtained in this particular setting as

e2
GH (B) = T (B). (7.29)
h

The Hall conductance is proportional to the single nonzero transmission probability.

7.4.3 Conductance Fluctuations as Quantum Manifestation

of Chaotic Scattering

One of the first studies on conductance fluctuations in quantum dots was carried out
by Jalabert, Baranger, and Stone [355] (see also [656]). They considered the con-
ductance for an open-stadium junction with two leads, and the symmetric four-disk
junction system in Fig. 7.5. For the open-stadium system the conductance is directly
proportional to the transmission probability, since there are only two leads con-
nected to the junction, one for the incoming and another for the outgoing electrons.
The width of each channel is denoted by W . For both junction systems, the classical
scattering dynamics are chaotic and hyperbolic. Quantum-mechanical calculations
7.4 Conductance Fluctuations in Quantum Dots 253

Fig. 7.6 (a) Hall resistance

RH ≡ 1/GH versus the
magnetic field strength B for
a four-disk junction defined
by the geometric parameters
R/W = 4 (inset). (b) For an
open-stadium billiard
junction of width
W = 1,000 Å, the
transmission probability T (B)
versus the magnetic field
strength [355] (copyright
1990, the American Physical
Society)

of the resistance as a function of the external magnetic field showed significant ran-
dom fluctuations, as shown in Fig. 7.6. In Fig. 7.6a, the fluctuating curves are from
two quantum calculations with slightly different Fermi energies EF . In particular, the
solid fluctuating curve is for kFW /π = 4.3, and the dashed curve is for kFW /π = 4.4,
where kF is the wave number at the Fermi energy. The existence of these fluctua-
tions is consistent with observation (c) in Sect. 7.2.3. The dotted curve, which does
not exhibit fluctuations, is from a purely classical calculation. The dashed-dotted
curve is the result of quantum calculation for a square junction, in which the clas-
sical scattering dynamics is regular. In Fig. 7.6b, the solid, fluctuating curve is the
quantum result for the transmission probability T (B) with kFW /π = 4.5, and the
dashed curve is the smoothed average T (B) used for computing the magnetic-
field autocorrelation function C(ΔB). The dotted line is the purely classical result.
Fluctuations are thus absent in a purely classical calculation of the scattering matrix
(as done by Beenakker and van Houten [47]), or for regular cavities where there
is no classical chaotic scattering. These indicate that the origin of the fluctuations
is quantum-mechanical and classical chaos is also relevant. Jalabert, Baranger, and
Stone performed a semiclassical analysis of the S-matrix elements with predictions
that appear to agree with the quantum calculations. The analysis is similar to that de-
scribed in Sect. 7.2, except that the electron energy is replaced by the magnetic field
strength B. An example of the magnetic field correlation function of the transmis-
sion probability T (B) is shown in Fig. 7.7, which corresponds to the open-stadium
junction in Fig. 7.6b. For small magnetic field strength, the correlation is Lorentzian,
as predicted by the semiclassical theory. The half-width is given again by a classical
quantity, similar to the parameter α in (7.17).
254 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

Fig. 7.7 Magnetic field correlation function C(ΔB) (continuous line) from the dashed smoothed
curve of T (B) in Fig. 7.6b. The dashed curve is the Lorentzian semiclassical prediction [355]
(copyright 1990, the American Physical Society)

7.5 Dynamical Tunneling in Nonhyperbolic Quantum Dots

A tacit assumption in the study of conductance fluctuations in quantum dots, as

discussed in Sect. 7.4, is that the underlying classical dynamics is hyperbolic.
In such an idealized case, the random-matrix theory is applicable. We have seen
in Sect. 7.2 that this approach predicts universal conductance fluctuations with a
Lorentzian type of correlation functions. A fundamental difficulty with the random-
matrix-theory-based approach is that many experimental results have suggested that
hyperbolicity is in fact an unusual dynamical feature in quantum dots (see e.g.,
[10, 345, 523].) Thus, for typical dots, the observed properties of the transport sim-
ply cannot be explained by the random-matrix theory. It is of interest to study the
effect of nonhyperbolicity on conductance fluctuations [68, 166].
In a typical nonhyperbolic quantum-dot system, the conductance exhibits strong
regular fluctuations as the gate voltage (or the magnetic field) is varied, as opposed
to random fluctuations [67, 69]. An example of the measured fluctuations is shown
in the lower left inset of Fig. 7.8. It was argued in [166] that in order to explain
these results, it is necessary to take into account quantum-mechanical tunneling of
electrons through the KAM islands (dynamical tunneling), which corresponds to
classically forbidden trajectories that are not taken into account in the usual semi-
classical approach.
For the quantum-dot system considered in [166], the self-consistently computed
potential profile is shown in the upper right inset in Fig. 7.8 (for more details, see
[69]). For the range of voltages studied, the profile remains basically the same.
For higher voltages, the potential is more confining, and the two openings be-
come smaller. The low-temperature electron and density mobility of this dot are
4 × 1015 m−2 and 70 m2 /V · s [69], respectively, with the corresponding Fermi
7.5 Dynamical Tunneling in Nonhyperbolic Quantum Dots 255

Fig. 7.8 A nonhyperbolic quantum dot. The upper right inset shows the electrostatic potential
profile (as a gray-scale plot) of the dot at the gate voltage of Vg = 0.6 V. Main panel: Poincaré
section of the corresponding classical scattering process with E = EF = 14.3 mV, where v is the
x-component of the velocity, given in units of the Fermi velocity. The “chaotic sea” surrounding
the island is the plot of a single trajectory that stays a long time near the island before escaping.
The phase-space area of the island is about h̄. The lower left inset shows the measured conductance
fluctuations observed by varying the gate voltage [69] (copyright 2002, the American Physical
Society)

energy of 14.3 meV. Each gate voltage corresponds to a different shape of the elec-
trostatic potential that can be computed self-consistently on a grid using a Poisson
solver. The transmission of electrons can be described as a scattering process in the
dot. The classical scattering dynamics can be studied by simulating the motion of
electrons at the Fermi energy in the underlying potential. The phase-space structure
can be visualized using a proper Poincaré surface of section. Figure 7.8 indicates
that the classical dynamics is nonhyperbolic, with a large KAM island dominating
the phase space. This island is centered on a period-1 orbit that bounces back and
forth vertically through the center of the dot, the so-called bouncing-ball orbit. It is
apparent that the usual random-matrix approach cannot be applied to this system.
We have seen in Sect. 7.3 that the semiclassical theory of nonhyperbolic chaotic
scattering predicts stochastic conductance fluctuations with different statistical
properties from those in the hyperbolic case. The theory does not predict the
regular fluctuations seen in experiments. This is because the usual semiclassical
theory considers only the interferences between the classically allowed trajectories,
and hence ignores the possibility of electron tunneling into the KAM island (cor-
responding to diffraction in optics). This effect is negligible only if the de Broglie
wavelength λe of the electron is much smaller than the cavity size. For typical dots
of size 1 μm, λe is about one-tenth of the cavity length, and therefore tunneling
256 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

cannot be neglected. Nonetheless, since λe is small enough, semiclassical concepts

such as the Bohr–Sommerfeld quantization of periodic orbits can still be applied.
Conductance measurements reveal that these regular oscillations have only a few
dominant frequencies [69]. It was argued [166] that these peaks are due to dynam-
ical tunneling [701, 783]: there is a probability that an incoming electron tunnels
into the KAM island. If the electron energy and the system parameters are such that
the semiclassical quantization condition (see below) is satisfied for a low-period
stable periodic orbit within the island, there is a resonance with a sharp decrease in
the transmission. As the system’s parameters change, these resonances occur with
the period given by the position of the peaks in the Fourier transform of the conduc-
tance fluctuations. As a matter of fact, dynamical tunneling is of great importance
in many fields, e.g., in cold-atom physics [327, 735] and in atomic optical billiard
[385]. The semiclassical theory of dynamical tunneling can be relevant for such
systems as well.
Because KAM surfaces are classically impenetrable, the classical dynamics
restricted to the island shown in Fig. 7.8 is that of a closed system. For closed sys-
tems, the Gutzwiller trace formula stipulates that each stable orbit generates a series
of delta functions in the density of states, at energies for which the resonance con-
dition holds [301, 525],
Ω 1 ν
Seff = S + m+ + = n, (7.30)
2π 2 4π

where S = 1h p · dq is the action along the periodic orbit in units of Planck’s con-
stant, Seff is an effective action, Ω is a quantity that characterizes the stability of the
orbit, and ν is the Maslov index; m, n = 0, 1, 2, . . . . The quantum-dot system under
consideration is open. Just as an incoming electron can tunnel in, an electron within
the island is in a metastable state, and can “decay” by escaping. This causes the
peaks in the density of states to broaden, and their width is inversely proportional
to the average time it takes for an electron to tunnel out of the island. If this time is
not too short, the peaks will be sharp enough to be resolved, and their positions are
determined by (7.30).
Consider first the case m = 0. Note that n is the longitudinal quantum number,
which counts the number of nodes in the “eigenfunction” along the orbit. Within
the island of Fig. 7.8 there are infinitely many periodic orbits, but only the lowest-
period ones are expected to be resolved.2 The most important orbit in the island
is the period-one bouncing-ball orbit corresponding to the fixed point at the center
of the island in Fig. 7.8. Moura et al. calculated [166] numerically Seff in (7.30) as
a function of the gate voltage Vg for this island, and compared it with the experi-
mental results, as shown in Fig. 7.9. It can be seen that the points fall reasonably
well on a straight line, which corresponds to a periodic recurrence of the resonance.
Since a resonance occurs each time Seff goes through an integer, the absolute value

2 High-period orbits generate peaks that are too closely spaced to be resolved, even more so with

the broadening of the levels caused by the quantum metastability of the system.
7.5 Dynamical Tunneling in Nonhyperbolic Quantum Dots 257

Fig. 7.9 Effective action Seff versus the gate voltage, for the stable (circles) and unstable (squares)
orbits. The inset shows a pair of closely spaced concentrated wave functions corresponding to the
stable bouncing-ball orbit (left and center images), and a scar due to the unstable orbit (right image)
along with its corresponding classical orbit, obtained by a full quantum-mechanical simulation of
the open system [166] (copyright 2002, the American Physical Society)

of the slope of the straight line gives the semiclassical prediction for the frequency
of the conductance oscillations. This frequency was found to be 16.4 V−1, in re-
markable agreement with the measured value of 15 V−1 [69], suggesting that this
peak corresponds to recurring tunneling resonances. This is reinforced by a fully
quantum simulation of the open system [69]. In particular, concentrations of wave
functions (the insets of Fig. 7.9, each corresponding to a different value of the gate
voltage) computed from the technique in [796] show a recurrent pattern. The re-
currence frequency of the concentrated wave function was determined to be [166]
16 V−1 , in agreement with the semiclassical prediction and with the experimental
result. Since this orbit is classically inaccessible from outside, the results represent
direct evidence of dynamical tunneling is nonhyperbolic quantum dots.
In a general case, i.e., when m is any positive integer, the second term in Seff
in (7.30) represents the quantization of the component of the motion transverse to
the periodic orbit. This means that for each n there is actually an infinite set of
resonances, labeled by m, similar to a vibrational band in a molecule. Assuming
that Seff changes linearly with the gate voltage Vg , one can estimate the separation
ΔVg between two resonances with consecutive transverse quantum numbers:

Ω
ΔVg ≈  , (7.31)
2π dSeff /dVg 

where it is assumed that Ω does not change much between two successive reso-
nances. Although Ω depends on n, its values
 were found [166] to lie in the range
between 1 and 2 rads. Using the value of dSeff /dVg derived from Fig. 7.9, one ob-
tains that ΔVg is between 1 × 10−2 and 2 × 10−2 V. Just as n counts the number of
258 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

nodes along the orbit, m counts the number of nodes across it, so we expect that
for each n, there is a set of concentrated wave functions having 0, 1, 2, . . . transverse
nodes, separated by the gate voltage interval ΔVg . Such recurring pairs of concen-
trated wave functions were observed in the quantum-mechanical simulation for the
bouncing-ball orbit, separated by a gate voltage difference of about 2 × 10−2 V,
which agrees well with the prediction. One pair is shown in Fig. 7.9 (left and center
images of the inset). It can be seen that they correspond to m = 0 and m = 1. They
are also found for other values of n, and they recur at the same period, as predicted. It
was emphasized [166] that this phenomenon cannot be explained without taking the
mechanism of tunneling into account, since it requires electrons to access the stable
orbit that is classically forbidden. The concentrated wave functions corresponding
to higher values of m (m = 2, 3, . . . ) were not resolved by the simulation, because of
their short lifetime.
Although the discussion so far has been focused on the stable orbits, unstable
orbits are also present in the system, and they contribute to the density of states. The
concentrated wave function corresponding to the main unstable orbit is displayed in
the upper right inset in Fig. 7.9. A classical trajectory analysis suggests that the or-
bit giving rise to this whispering-gallery scar is guided by the soft walls around the
perimeter of the lower section of the dot, bouncing from the upper wall at two points,
located close to the two lead openings. The resonant condition for unstable periodic
orbits is given by (7.30) without the Ω term [301]. This means that unstable peri-
odic orbits do not give rise to the subband of resonances associated with m. In the
Fourier transform of the conductance oscillations, a peak was observed at Vg ≈ 37 V,
corresponding to an unstable periodic orbit of period 1. Figure 7.9 also shows a plot
of Seff versus Vg for this orbit, and from the slope we get a recurrence frequency of
36.3 V−1, again in good agreement with the experimental result. The concentrated
wave functions related to these resonances were seen in the quantum-mechanical
simulation, but no subband was observed. The other main periodicities found in the
conductance correspond to harmonics of the main stable and unstable resonances
studied above. Note that in [345], isolated resonances were predicted to arise from
the chaotic part of the phase-space outside the islands (see Fig. 7.8), but these can
be detected for smaller values of h̄ only, i.e., for larger quantum dots.3
We thus see that the usual semiclassical approach is not enough to explain the
transport characteristics of typical semiconductor nanostructures, and the quantum-
mechanical tunneling of the electron through KAM islands has to be taken into
account. Tunneling resonances caused by stable and unstable periodic orbits of low
periods within the KAM islands cause regular oscillations of the conductance. These
results are expected to hold for all quantum-dot systems [166]. The general stand-
point is that tunneling plays a fundamental role in mesoscopic transport.

3 In contrast to [345], h̄ is of the size of the island and therefore the hierarchical part of the phase
space surrounding the island is not resolved quantum-mechanically. That is why the resonances
due to the hierarchical states are not observed. Note, however, that these are always broader than
those due to the stable island.
7.6 Dynamical Tunneling and Quantum Echoes in Scattering 259

7.6 Dynamical Tunneling and Quantum Echoes in Scattering

As pointed out by Jung, Mejia-Monasterio, and Seligmann [363, 365], echoes can
occur in classical and quantum chaotic scattering. In the classical context, echoes
occur in nonhyperbolic systems when the chaotic saddle is in such a stage of de-
velopment that the stable island associated with the inner periodic orbit is relatively
large. Quantum-mechanically, a short pulse in the incoming wave can then lead to
periodic pulses in the outgoing wave.
To describe the classical phenomenon, let us recall the form of the Poincaré sec-
tion of a binary horseshoe as illustrated by Fig. 6.24. When incoming scattering
trajectories approach the central region along the stable manifold, they rotate around
the stable island and return to the vicinity of the hyperbolic fixed point. Depending
on the initial conditions, some of the trajectories leave the hyperbolic point along its
unstable manifold. The remaining trajectories continue to rotate to complete a sec-
ond revolution. By reaching the hyperbolic fixed point again, some trajectories exit
the central region around the island, and the process continues. The asymptotic out-
going flux resulting from a narrow packet of incoming trajectories therefore exhibits
intensity oscillations in time. These oscillations are the scattering echoes.
The time between successive maxima of the outgoing flux corresponds to the
mean period of rotation around the island. This time can be expressed by the de-
velopment parameter introduced in Sect. 6.4.2. Based on Fig. 6.25, which illustrates
a binary horseshoe for development parameter α = 2−m with m = 4, we can esti-
mate the mean period of rotation. Consider a point (star marked by 0) lying close
to the unstable tendril 1. Since the unstable tendril of level n is the image of tendril
(n − 1) on the Poincaré section, the nth image of the starting point will be close to
the unstable tendril (n + 1). The revolution is almost completed after the (m + 1)th
iteration, when the point is around the tip of the unstable tendril m + 2 (tendril 6
in Fig. 6.25). After another step, the trajectory completes a bit more than one rev-
olution. The period on the Poincaré map can thus be estimated as T = m + 3/2.
In terms of the development parameter α , we have

T = − log2 α + 3/2. (7.32)

In continuous time the period can be expressed as

T = t0 (− log2 α + 3/2), (7.33)

where t0 is the average return time to the Poincaré section. For ternary horseshoes
one obtains [363]
T = t0 (−2 log3 β + 3/2), (7.34)
where the factor 2 is a consequence of the fact that there are two hyperbolic
fixed points associated with a ternary horseshoe. The development parameter alone
thus determines the classical winding period, which is the pulsation period of the
scattering echoes.
260 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

At the quantum level, the echoes are of a similar origin, and arise if the system
is excited with a short pulse. A new feature is that the quantum wave function can
penetrate the classical KAM surfaces and reach the inner region of the classical
island. This dynamical tunneling effect results in rotation of the quantum probability
around the island at small distances from the center with a small average speed. The
classical result in (7.34) is thus an upper bound for the quantum pulsation period.
In a measurement of the quantum echoes one is able to extract information from
regions of the phase space that are inaccessible to classical scattering trajectories.
Such experiments were performed in Darmstadt by Dembowski and coworkers
[175], who used open microwave billiards both at room temperature and in a su-
perconducting state. The data were taken for continuous input of microwaves of
different frequencies, and the response to a short incoming pulse was obtained by
Fourier transform. The S-matrix elements Si j between antennas i and j (lying out-
side the scattering region) were evaluated in the time domain. The phenomenon
of pulsation in the transmission probabilities can be seen in Fig. 7.10. Numerical
simulations of the classical open billiard used in the experiment reveals that the
horseshoe is ternary, and its development parameter is β = 3−8 . The classical for-
mula (7.34) yields T = 4.67 ns as the pulsation time. The microwave experiments
show that the period T of the echoes gets shorter with time. Initially, T = 4.2 ns; then
T decreases and stabilizes about T = 3.3 ns. This can be explained by the fact that
the wave packet needs time to penetrate into the region of the classical KAM island.
With time, echoes from deeper layers inside the island are received. The initial
response originates from the chaotic layer, and is close indeed to the classical result.

Fig. 7.10 Transmission probabilities between antennas 2 and 3 (top) and between 1 and 2 (bottom)
in the scattering echo experiment by Dembowski and coworkers. Since the classical horseshoe is
at its initial stage of development, periodicity is observed in the decay process. Period T marks the
classical value given by (7.34) [175] (copyright 2004, the American Physical Society)
7.7 Leaked Quantum Systems 261

After T is stabilized, one observes dynamical tunneling from a fixed penetration

depth of the island. The same data reveal an overall exponential decay with time, an
indication of the tunneling effect.
After a metal disk is introduced into the inner part of the billiard, the echoes
disappear. This can be explained by the destruction of the central integrable island
in the phase space, which is the theoretical basis for the echoes.
The phenomenon of scattering echoes provides an example of the inverse chaotic
scattering problem. From a measured quantity one can extract information about
the structure of the underlying chaotic saddle. It is remarkable in the Darmstadt
experiment that it is achieved in a regime far from the semiclassical limit (as shown
by the deep tunneling). Nevertheless, from the initial period T the development
stage of the underlying classical chaotic saddle and its dynamical invariants such as
the topological entropy can be assessed, at least approximately. Recent numerical
studies indicate that the quantum echoes can also be observed in the conductance
fluctuations of semiconductor nanostructures [472].

7.7 Leaked Quantum Systems

Open quantum systems can arise beyond the scattering context. For example,
absorbing regions can play an important role in quantum mechanics correspond-
ing to the existence of a classical leak. The quantum dynamics is governed then by a
nonunitary evolution operator, and the system has properties different from those of
closed quantum systems. Study of the open kicked rotator leads to the observation
that quantum fractal eigenstates exist, which reflect the pattern of the underlying
chaotic saddle [116]. In particular, the long-lived eigenstates concentrate on fractal
trapped sets. The left and right eigenstates of the nonunitary propagator concentrate
in the limit of h̄ → 0 on the stable and unstable manifolds of the chaotic saddle,
respectively [403], as can be seen from the example of the quantized open baker
map. The quantized version of the leaked baker map of Sect. 2.7.1 was studied in
[226, 587].
A related problem is the analysis of the statistical properties of energy levels.
The well-known Weyl law states that the number N(k) of levels with wave number
smaller than k grows with a power of k. For two-dimensional systems, the exponent
is 2. The extension to open systems is that the number N(k) of resonances with wave
numbers of real part smaller than k scales as

N(k) ∼ kα (7.35)

with
D0 (1)
α = 1+ = 1 + D0 , (7.36)
2
where D0 is the box-counting dimension of the underlying classical are preserv-
(1) (2)
ing chaotic saddle on a properly chosen Poincaré map, and D0 = D0 is the partial
262 7 Quantum Chaotic Scattering and Conductance Fluctuations in Nanostructures

box-counting dimension (see Sect. 2.6.2). This is the fractal Weyl law [209,225,414,
494, 556, 620, 712, 732, 825]. These properties apply also to the microlasing cavities
described in Sect. 6.7. The ray dynamics presented there correspond in a first-order
approximation to geometrical optics. The second-order approximation is a semi-
classical wave dynamics, which is analogous to semiclassical quantum mechanics
[695, 825].
The problem of quantum Poincaré recurrences has also been studied [117]. The
short-time behavior is that of the classical problem, and after a crossover time that
depends on the value of h̄, an algebraic decay sets in both in classically hyper-
bolic and nonhyperbolic cases. Quantum effects are thus able to convert the classical
exponential behavior into a power law, and also to change the algebraic-decay ex-
ponent. The integrated return probability Pr (t), i.e., the probability of returning to
the given region after time t, was found to decay universally as Pr (t) ∼ 1/t, which
corresponds to a decay of the recurrence time distribution , defined in Sect. 2.7.2, as
pr (t) ∼ 1/t 2.
Another phenomenon through which the effect of leaking can be observed is
Loschmidt echo. Since there are no trajectories in quantum systems, the chaoticity
of the system can be checked, as suggested by Peres [593], by evolving a wave
packet under two slightly modified Hamiltonians and measuring how they differ in
time. In Loschmidt echoes, the wave packet evolves forward in time with a Hamil-
tonian up to time t, and then evolves backward in time with a slightly modified
Hamiltonian over the same period t. The overlap between the initial and final wave
packets is measured as a function of t. Goussev and Richter [277] considered a bil-
liard, and the difference in the two Hamiltonians is due to a change in the shape of
the billiard in a small region I. The decay of the echo was found to be governed
by the classical escape rate characterizing the billiard with a leak in region I. These
observations indicate that within the field of quantum chaos, there is an increasing
recent interest in open systems. These, in the semiclassical limit, reflect properties
of the underlying classical transient chaos.
 Part III
High-Dimensional Transient Chaos
Chapter 8
Transient Chaos in Higher Dimensions

This chapter is devoted to transient chaos in higher-dimensional dynamical systems.

The defining characteristic of high-dimensional transient chaos is that the underly-
ing chaotic set has unstable dimension more than one, in contrast to most situations
discussed in previous chapters, where chaotic sets have one unstable dimension. We
shall call nonattracting chaotic sets with one unstable dimension low-dimensional,
while those having unstable dimension greater than one high-dimensional. The in-
crease in the unstable dimension from one represents a highly nontrivial extension
in terms of what has been discussed so far about transient chaos. For instance,
the PIM-triple algorithm, which is effective for finding an approximate continuous
trajectory on a low-dimensional chaotic saddle, is generally not applicable to high-
dimensional chaotic saddles. In a scattering experiment in high-dimensional phase
space, the presence of a chaotic saddle cannot guarantee that chaos can be physi-
cally observed. In particular, if the box-counting dimension of the chaotic saddle is
low, its stable manifold may not intersect a set of initial conditions prepared in the
corresponding physical space; only when the dimension is high enough can chaotic
scattering be observed.
First we present a prototypical example, the three-dimensional baker map, for
which a basic property of high-dimensional hyperbolic systems can be seen ex-
plicitly: different numbers of the expanding and the contracting dimensions. Next
we show how the escape rate and the metric entropy can be expressed in terms of
the Lyapunov exponents in high-dimensional maps, present derivations for the di-
mension formulas of the stable and the unstable manifolds for high-dimensional
chaotic saddles, and address their applicability and the concept of typicality using
particular examples. An efficient algorithm, the stagger-and-step method, is then
described for computing high-dimensional chaotic saddles. Chaotic scattering in
three-degree-of-freedom systems is presented, the conditions for the observability
of chaotic scattering are formulated, and new features of the scattering dynamics
are discussed. The phenomenon of superpersistent chaotic transients and applica-
tions are also treated in this chapter.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 265
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 8,

c Springer Science+Business Media, LLC 2011
266 8 Transient Chaos in Higher Dimensions

8.1 Three-Dimensional Open Baker Map

Three-dimensional invertible maps arise on stroboscopic or Poincaré sections of

four-dimensional flows. There are two classes of hyperbolic maps: (1) maps with
one unstable and two stable directions (type I) and (2) maps with two unstable and
one stable direction (type II). The escape dynamics in these two cases are generally
different.
The open three-dimensional baker map is the spatial extension [165] of the pla-
nar baker map treated in Sect. 2.5. Because of the odd dimensionality of the map,
the number of unstable directions is different from that of stable directions. Con-
sequently, such maps, even if they preserve volume, cannot describe Hamiltonian
systems. Nonetheless, in the volume-preserving case they can be used to model the
advection by three-dimensional time-periodic flows (see [165] and Chap. 10). The
map, denoted by M, is defined on the unit cube, and its action is shown in Fig. 8.1a.

Fig. 8.1 (a) Illustration of the action of one iteration of the three-dimensional open baker map
M on the unit cube (the restraining region Γ ) for type-I dynamics. For simplicity the parameters
are chosen to be b = c ≡ λ and a = 1/λ 2 (volume-preserving case). Note that the drawing is not
to scale. (b) For λ = 0.35, surviving points after two iterations of the map with initial conditions
chosen randomly from the cube. (c) The same as (b), but for the inverse map M −1 (type-II dynam-
ics). (d) For λ = 0.35, intersection with a horizontal plane of the set of surviving points after two
iterations of the map [771] (with kind permission from Elsevier Science)
8.1 Three-Dimensional Open Baker Map 267

One iteration of the map consists of two actions. Firstly, the x and y directions are
contracted by factors b and c, respectively, where b, c < 1/2, while the z direction
undergoes an expansion by a factor a > 4. Under the transformation, the cube turns
into a long, thin rectangular slab with its long edge along the z-axis, as shown in
Fig. 8.1a. Secondly, four pieces of unit height of this slab are selected and placed
in the four corners of the cube. The pieces of the slab that are not selected are dis-
carded and are regarded as having escaped (Fig. 8.1a). The map has two contracting
directions and one expanding direction, and is therefore of type I. For the inverse
map M−1 , stable directions turn into unstable ones, and vice versa. As a result, M−1
has one stable and two unstable directions, and is of type II. The two generic types
of three-dimensional hyperbolic maps can thus be conveniently studied using the
baker map. Since hyperbolic systems are structurally stable, generality is not lost by
assuming any particular form for M.
Because the contracting and the expanding directions of M (and also M−1 ) are
aligned with the x-, y-, and z-axes, it is not difficult to visualize the stable and the
unstable manifolds: the stable manifold of M is a Cantor set of planes parallel to
the horizontal (x, y)-plane, and the unstable manifold is a Cantor set of vertical seg-
ments. We can visualize these manifolds by iterating M forward a given number of
times for many initial conditions chosen randomly within the unit cube. The dis-
tribution of points that have not escaped approximate the unstable manifold of M,
which is the stable manifold of M−1 . Conversely, iterating backward (or iterating
M−1 forward) in time gives the stable manifold of M (or the unstable manifold of
M−1 ). These results are shown in Fig. 8.1b, c.
Consider now the unstable manifold of M, as shown in Fig. 8.1b. Since it is made
up of vertical line segments, and since the expansion and the contraction rates are
uniform, we can restrict attention to the intersection of the unstable manifold with
a horizontal plane. This is depicted in Fig. 8.1d, where the intersection of the set of
surviving points after two iterations of M with a horizontal plane is shown. In the
limit of an infinite number of iterations, a double Cantor set in the plane is formed
(2) (3)
with partial box-counting dimensions D0 = ln 2/ ln (1/b) and D0 = ln 2/ ln (1/c)
along the stable manifolds. The unstable manifold is the product of this Cantor set
and a one-dimensional line segment; its dimension is

1 1
Du,0 = 1 + ln2 + . (8.1)
ln (1/b) ln (1/c)

Similar reasoning can be applied to the stable manifold. The partial box-counting
dimension along the unstable direction is set by the stretching factor a. Since in each
(1)
step a factor 1/a of the slab remains in the unit cube, we have D0 = ln 4/ ln a. The
stable manifold is the product of this Cantor set with a plane. We obtain

ln 4
Ds,0 = 2 + . (8.2)
ln a
268 8 Transient Chaos in Higher Dimensions

The manifold dimensions satisfy

1 < Du,0 < 3, 2 < Ds,0 < 3. (8.3)

The chaotic saddle is the intersection of the stable and the unstable manifolds, and
its box-counting dimension is

3
∑ D0
( j)
D0 = = Du,0 + Ds,0 − 3. (8.4)
j=1

The baker map has one positive Lyapunov exponent λ1+ = ln a and two negative ones
of magnitudes λ1− = ln (1/b) and λ2− = ln (1/c). The escape rate is κ = ln (a/4), and
the topological entropy is K0 = ln 4.
The inverted map has two positive Lyapunov exponents, λ1+ = ln (1/b) and
λ2 = ln (1/c), and a negative exponent of magnitude λ1− = ln a. The topology of
+

the escaping process in the inverted map is different from that of M, which is also
reflected by the difference in the values of the escape rate (for the inverted map it is
κ = − ln (4bc)). The manifold dimensions can be obtained from (8.1) and (8.2) by
interchanging the indices u and s.
In general, the structures of the stable and the unstable manifolds for the two
types of generic maps are topologically similar to those shown in Fig. 8.1. In par-
ticular, for type-I maps, the stable manifold is a Cantor set of surfaces, and the
unstable manifold is a Cantor set of one-dimensional curves (vice versa for type-II
maps). Also, the inequalities (8.3) hold in general for type-I maps.
Based on (8.3), we observe that for type-II maps, the dimension Ds,0 of the stable
manifold may be less than 2. In this case, the stable manifold has generically a
null intersection with a one-dimensional curve (see (8.52)), and thus the lifetime
distribution along a line is a smooth function, even though there is a fractal invariant
set. For Ds,0 > 2, however, a typical lifetime function has a fractal set of singularities
that is similar to that in Fig. 1.5. The transition point defined by Ds,0 = 2 is given
for the map M−1 by the condition 1/ ln b + 1/ lnc = −1/ ln 2. If the map is of type I,
however, this transition does not occur. This is a nontrivial difference between the
dynamics of the two types of map. We will consider a similar problem in the context
of high-dimensional chaotic scattering in Sect. 8.5.1.

8.2 Escape Rate, Entropies, and Fractal Dimensions for

Nonattracting Chaotic Sets in Higher Dimensions

8.2.1 Escape Rate and Entropies

Consider an open dynamical system described by an N-dimensional map. For a

general chaotic saddle in the N-dimensional phase space, there are U positive and
8.2 Escape Rate, Entropies, and Dimensions in Higher Dimensions 269

S negative average Lyapunov exponents, where U + S = N. The exponents can be

ordered as follows

λU+ ≥ λU−1
+
≥ · · · ≥ λ1+ > 0 ≥ −λ1− ≥ · · · ≥ −λS−1
−
≥ −λS− . (8.5)

Thus all quantities λ j+,− are positive, and smaller values of the subscripts j corre-
spond to Lyapunov exponents that are closer to zero in magnitude. Analogous to
the low-dimensional cases, one can define partial box-counting and information di-
( j) ( j)
mensions D0 and D1 , respectively, for any direction j along which an average
Lyapunov exponent exists. The total dimension Di (i = 0, 1) of the nonattracting set
is the sum of the partial dimensions:

Di = ∑ Di ,
( j)
i = 0, 1. (8.6)
j

General expressions for the escape rate and the metric entropy follow from the ex-
tension of the information-theoretic arguments in Sect. 2.6.3 for two-dimensional
maps [380].
Along the stable directions particles cannot escape. Along an unstable direction,
( j)
they cannot escape either if the partial information dimension is D1 = 1 (more
precisely, the escape is slower than exponential). Exponential escape is possible only
( j)
along unstable directions for which D1 < 1. Since escapes in different directions j
are independent of each other, and the mean velocity of the information flow is λ j+ ,
the escape rate is the sum of contributions from all unstable directions:
U
∑ λ j+(1 − D1
( j)
κ= ). (8.7)
j=1

The metric entropy, as discussed in Sect. 2.6.3, is the rate at which information
stored in the insignificant digits of the initial condition flows toward the significant
ones. This flow occurs along the unstable directions only. Since the information di-
mension and the Lyapunov exponent characterize the density of information and the
mean velocity of the flow, respectively, we have
U
∑ λ j+ D1
( j)
K1 = . (8.8)
j=1

Using (8.7), one can write K1 as

U
K1 = ∑ λ j+ − κ . (8.9)
j=1

This relation implies that the degree of unpredictability of the dynamics on the
nonattracting set, measured by K1 , is only a fraction of the total flux ∑Uj=1 λ j+ of dig-
its, because of the loss of information due to escape. Formula (8.9) is an extension
270 8 Transient Chaos in Higher Dimensions

of Pesin’s relation [564], according to which the metric entropy is the sum of all
positive Lyapunov exponents for closed N-dimensional maps.
In invertible systems, the same amount of information flows in along the stable
directions as the amount flowing out along the unstable directions, i.e.,

S U
∑ λ j− D1 ∑ λ j+D1
( j) ( j)
= . (8.10)
j=1 j=1

A special case of this rule has been observed in (2.80) for two-dimensional maps.
The relations (8.7), (8.8), and (8.9) are valid for noninvertible and nonhyperbolic
cases as well. For example, two-dimensional maps with two positive Lyapunov
exponents are covered by these formulas, such as the repellers of the complex
quadratic map, the Julia sets shown in Fig. 1.3b, which are topologically a cir-
cle. Due to the symmetry, both Lyapunov exponents and both partial information
dimensions are equal. Equations (8.7) and (8.8) then imply, with the notation
λ1+ = λ2+ ≡ λ , that
κ = λ (2 − D1), K1 = λ D1 . (8.11)

Since the natural measure of the connected Julia set is known to have information
dimension D1 = 1 [84], we have κ = λ = K1 . This is to be contrasted with the prop-
erties of isolated repeller points of the complex quadratic maps for which D1 = 0
and hence, κ = 2λ with K1 = 0.
Obtaining an expression for the topological entropy is more complicated. As
generalizations of (2.22) and (2.81), we have [380]

U
1 U +
K0 = ∑ λ j+ − κ + 2∑
Q2, j,k + · · · (8.12)
j=1 j,k

and
U
1 U
∑ λ j+ (1 − D0 2∑
( j) ( j) (k)
κ= )+ (1 − D0 )(1 − D0 )Q+
2, j,k + · · · , (8.13)
j=1 j,k

where Q+ 2, j,k represents the cumulant of the Lyapunov exponents from expanding
directions j and k [380].

8.2.2 Dimension Formulas for High-Dimensional Chaotic

Saddles

For high-dimensional chaotic saddles, dimension formulas for the entire stable and
unstable manifolds can be derived. Our treatment here follows that of Hunt et al.
[347]. Imagine normalizing the size of the chaotic saddle so that it can be enclosed
8.2 Escape Rate, Entropies, and Dimensions in Higher Dimensions 271

by a cube of unit length. This cube is considered the restraining region. For a
hyperbolic saddle, the cube can be regarded as having edges parallel to directions
of stretching and contraction as defined by the Lyapunov exponents λ j+ and −λi− .
Now uniformly sprinkle a large number of points in the cube and iterate them
forward n times. The resulting trajectory points will be distributed to slabs within
the cube of dimensions
− − −
1 × 1 × · · ·× 1 × e−λ1 n × e−λ2 n × · · · × e−λS n , (8.14)

where there are U slab edges of unit length in the U unstable directions. Let N (sl) (n)
be the number of slabs at time n. Since trajectory points within these slabs remain
in the cube for at least n iterates, the total content of the slabs is proportional to
exp (−κ n), where κ is the escape rate from
 the chaotic saddle.  Since
 the density of
points has increased by a factor of exp ∑Sj=1 λ j− − ∑Uj=1 λ j+ n , we have

 
U
−λU+ n −λU+−1n −λ1+n
N (sl)
(n)e ×e × ···× e =N (sl)
(n) exp − ∑ λ j+ n ∼ exp (−κ n).
j=1
(8.15)
Using (8.9), we see that (8.15) implies

N (sl) (n) ∼ eK1 n . (8.16)

Since at time n, trajectory points that have not left the restraining region are dis-
tributed in the vicinity of the unstable manifold, we need to examine the set of
N (sl) (n) in (8.14). Say we wish to cover them using small N-dimensional cubes.
A natural choice for the edge length of such a cube is that set by the contraction
of the dynamics. Since there are several contracting directions, we have different
choices. Take
−
εi = exp(−λi+1 n) (8.17)

with index i between 0 and S − 1. The required number of cubes is then

 −
 −
  −

(u) 1 U
e− λ 1 n
e− λ 2 n
e− λ i n
N (εi ) = ··· N (sl) (n) (8.18)
εi εi εi εi
U+i
1
∼ exp [−(λ1− + λ2− + · · · + λi− )n + K1 n],
εi

where (8.16) has been used.

The information dimension of any invariant measure can be considered as the
box-counting dimension of regions containing most of the measure, i.e., regions
covering typical sets taken with respect to the measure (see, e.g., [773]). Applying
272 8 Transient Chaos in Higher Dimensions

this to the c-measure of the unstable manifold, represented by the slabs at time n,
we find an approximation to the information dimension of the unstable manifold for
large n:

ln N (u) (εi ) K1 − (λ1− + λ2− + · · · + λi− )

Du,1 (i) = =U +i+ − . (8.19)
ln (1/εi ) λi+1

The covering by the set of εi -cubes may not be optimal, so Du,1 (i) is an upper bound
of Du,1 : Du,1 ≤ Du,1 (i). It is thus necessary to minimize Du,1 (i) over i to obtain the
true dimension. A convenient way to find the minimum of Du (i) is to examine the
quantity Du,1 (i + 1) − Du,1(i), which is
 
1 1
Du,1 (i + 1) − Du,1(i) = − − − × [(λ1− + λ2− + · · · + λi− + λi+1
−
) − K1 ].
λi+1 λi+2

− −
Since λi+1 ≤ λi+2 , we see that Du,1 (i + 1) − Du,1 (i) is positive (negative) if the
term in the square brackets is positive (negative). Thus, if there exists a value I of i
such that

λ1− + λ2− + · · · + λI− + λI+1

−
≥ K1 ≥ λ1− + λ2− + · · · + λI− , (8.20)

then Du,1 (I + 1) − Du,1 (I) is positive or zero but Du,1 (I) − Du,1 (I − 1) is negative
or zero. That is, we have Du,1 (I) ≤ Du,1 (I + 1) and Du,1 (I) ≤ Du,1 (I − 1) simulta-
neously, indicating that the value of I chosen in (8.20) is the optimal choice of the
index i that yields the true dimension Du,1 :

K1 − (λ1− + λ2− + · · · + λI− )

Du,1 = U + I + − , (8.21)
λI+1

where I is the largest index for which the numerator of (8.21) is still positive.
The information dimension of the natural measure of the stable manifold can be
obtained in a similar manner. To see where the slabs of size (8.14) (whose number is
N (sl) (n)) come from within the cube, we iterate them backward n times and obtain
N (sl) (n) slabs of initial conditions, each of dimension
+ + +
e−λU n × e−λU −1n × · · · × e−λ1 n × 1 × · · · × 1, (8.22)

where for each slab there are S edges of unit length. Since initial conditions leading
to trajectories that remain in the restraining region for at least n iterates are found
in the slabs of size given by (8.22), we can cover them by small cubes of properly
+
chosen edge length ε j = exp (−λ j+1 n) and obtain an upper bound Ds,1 ( j) for the
true dimension Ds,1 . Reasoning similar to that in the derivation of Du,1 yields the
following optimal choice of the index J:
8.2 Escape Rate, Entropies, and Dimensions in Higher Dimensions 273

λ1+ + λ2+ + · · · + λJ+ + λJ+1

+
≥ K1 ≥ λ1+ + λ2+ + · · · + λJ+ , (8.23)

which gives
K1 − (λ1+ + λ2+ + · · · + λJ+)
Ds,1 = S + J + + , (8.24)
λJ+1

where J is the largest index for which the numerator of (8.24) is still positive.
The information dimension of the chaotic saddle, which is the intersection of its
stable and unstable manifolds, is
+
K1 − ∑Ii=1 λi− K1 − ∑ j=1 λ j
J
D1 = Du,1 + Ds,1 − N = (I + J) + − + + . (8.25)
λI+1 λJ+1

This is a generalization of (8.4) for the information dimension of the invariant set to
any N-dimensional map.
In the case of a chaotic attractor, we have κ = 0, so K1 is the sum of all positive
Lyapunov exponents. This leads to the information dimension D1 of the attractor,
since D1 = Du,1 with (8.21), which is the Kaplan–Yorke formula in higher dimen-
sions [564]. Note that the index I in Du,1 is then such that (∑Uj=1 λ j+ − ∑Ii=1 λi− ) is
positive but (∑Uj=1 λ j+ − ∑I+1 −
i=1 λi ) is negative. Furthermore, from (8.23) we see that
J = U − 1 and thus Ds,1 = S + U = N, i.e., the stable manifold is space-filling, as it
should for a basin of attraction.
A special case is high-dimensional maps derived from Hamiltonian flows. Due to
the symplectic structure of the dynamics, positive and negative Lyapunov exponents
arise in pairs: λ j+ = λ j− . The manifold dimensions in Hamiltonian systems therefore
coincide:
Du,1 = Ds,1 = (D1 + N)/2. (8.26)

For a chaotic saddle of a two-dimensional map with one positive Lyapunov ex-
ponent λ1 ≡ λ1+ > 0 and one negative exponent λ2 ≡ −λ1− < 0, we have U = 1 and
S = 1. In dissipative or area-preserving systems, we have λ1 + λ2 ≤ 0. As a result,
λ1 + λ2 − κ < 0. Thus K1 ≤ |λ2 |, and we have I = 0 and J = 0, which leads to the
corresponding formulas derived in Sect. 2.6.2.
Readers should keep in mind that the dimension formulas (8.21), (8.24), and
(8.25) are derived heuristically under the assumption that the chaotic saddle is hy-
perbolic. While there is numerical evidence for a class of open systems (see, e.g.,
[745,746]) with nonhyperbolic high-dimensional chaotic saddles, there has been no
systematic numerical study to validate these formulas, although they are conjectured
to apply to typical systems [347]. This can be seen heuristically by noting that for
such a system, small perturbations cannot change its properties and dynamical in-
variants. Atypical systems, on the other hand, are those whose dynamical invariants
change under small perturbations. In the next section we will consider specific ex-
amples to contrast typical versus atypical systems with respect to their dimensions.
274 8 Transient Chaos in Higher Dimensions

8.3 Models Testing Dimension Formulas

Because of the heuristic nature in the derivation of the dimension formulas, it is

insightful to examine representative models for which the quantities involved in the
formulas can be obtained analytically or numerically. The following two models,
due to Sweet and Ott [746], will be used: (1) an analytic two-dimensional, nonin-
vertible expanding map, and (2) a three-dimensional billiard scatterer.

8.3.1 Two-Dimensional Noninvertible Map Model

8.3.1.1 Natural Measure and Lyapunov Exponents

The map is of the form

xn+1 = 2xn mod(1), (8.27)

η
yn+1 = α (xn )yn + sin (2π xn ),
2π

where α (x) > 1 and the map is defined in the region −∞ ≤ y ≤ +∞ and 0 ≤ x ≤ 1.
The variable x can be considered as an angle-like variable, so the map is defined on
a cylinder. The following piecewise constant function was chosen for α (x):

α1 , 0 < x < 1/2,
α (x) = (8.28)
α2 , 1/2 < x < 1,

where 1 < α1 ≤ α2 . Because α (x) > 1, almost all initial conditions go either to
y = +∞ or to y = −∞, which can be regarded as two attractors, and there is a bound-
ary between the two basins of attraction near y = 0. The boundary is an invariant
set, which is ergodic because of the chaotic dynamics in x. In fact, the invariant set
is a chaotic repeller with two positive Lyapunov exponents. The Jacobian matrix of
(8.27) is
2 0
J(x) = , (8.29)
η cos (2π x) α (x)
so the two Lyapunov exponents of the chaotic repeller are

λa = p ln α1 + (1 − p) ln α2 and λb = ln 2, (8.30)

where p is the measure of the region x < 1/2. Note that for the one-dimensional
map xn+1 = 2xn mod(1) alone, we have p = 1/2 because a random initial con-
dition leads to a trajectory that visits the intervals [0, 1/2] and [1/2, 1] with equal
probabilities. However, the presence of the y-dynamics changes the natural measure
8.3 Models Testing Dimension Formulas 275

of the x-intervals. To calculate p, consider the strip −K ≤ y ≤ K on the cylinder

and sprinkle a large number of initial conditions uniformly in the strip with density
ρ0 . A vertical line segment of length 2K at x = x0 and centered at y0 = 0 iter-
ates to x = x1 , and its center will be at y1 = (η /2π ) sin (2π x0). This line segment
will at the same time be stretched vertically by a factor of α (x0 ). The endpoints
of the segment will then be at (η /2π ) sin (2π x0 ) ± α (x0 )K. In order for the seg-
ment to span the initial strip −K ≤ y ≤ K, it is necessary to choose K such that
(η /2π ) sin (2π x0 ) − α (x0 )K < −K (if y1 > 0) or (η /2π ) sin (2π x0) + α (x0 )K > K
(if y1 < 0). In either case, it is necessary to have

K > (η /2π )| sin (2π x0 )|/[α (x0 ) − 1],

which can be satisfied if we choose

K > (η /2π )/(α1 − 1).

Since the map stretches a region uniformly in the x-direction by a factor of two and
in the y-direction by a factor of either α1 or α2 , after one iterate the density will still
be uniform in the initial strip, and it is ρ1 = [(α1−1 + α2−1 )/2]ρ0 . After n iterations,
the density in the strip becomes
n
1 −1
ρn = (α + α2−1 ) ρ0 ,
2 1

which decays exponentially with time (ρn = ρ0 exp(−κ n)) with escape rate

2α1 α2
κ = ln . (8.31)
α1 + α2

Since both Lyapunov exponents are positive, the chaotic repeller formally coincides
with its stable manifold. It suffices thus to calculate the natural measure of the stable
manifold. To do so, note that in the x-direction, an interval of length 2−n maps to the
unit interval after n iterates. It is thus useful to divide the initial strip −K ≤ y ≤ K
(n)
into 2n vertical substrips. In substrip i, we have x ∈ si = [(i − 1)/2n, i/2n ] (i =
1, . . . , 2n ). For a uniform distribution of N0 = 2K ρ0 points in the strip −K ≤ y ≤ K,
(n)
N0 /2n will be in si and we ask how many of those there are whose trajectories do
(n)
not leave the strip at time n. Assume that in n iterates, the substrip si experiences
n1 (i) and n2 (i) vertical stretches by α1 and α2 , respectively, where n1 (i) + n2 (i) = n.
(n)
The initial subregion in si that can survive at least n iterates has vertical height
−n (i) −n (i)
2K α1 1 α2 2 . There are then

−n1 (i) −n2 (i)

2K α1 α2 N0 −n (m) −n (m)
· = 2−n α1 1 α2 2 N0
2K 2n
276 8 Transient Chaos in Higher Dimensions

(n)
such initial conditions. The measure of the stable manifold in si is

−n1 (i) −n2 (i) −n (i) −n (i) n (i) n (i)

(n) 2−n α1 α2 N0 2−n α1 1 α2 2 N0 α1 2 α2 1
μi = = = . (8.32)
ρn 2K [(1/2)(α1−1 + α2−1 )]n N0 (α1 + α2 )n

We have
2n
μ ([0, 1]) = ∑ μi
(n)
= 1.
i=1

The measures of the intervals [0, 1/2] and [1/2, 1] are given by

(1) α2 (1) α1
p = μ1 = , 1 − p = μ2 = , (8.33)
α1 + α2 α1 + α2

which gives
α2 α1
λa = ln α1 + ln α2 . (8.34)
α1 + α2 α1 + α2
It can be checked that
λa ≤ κ , (8.35)
where the equality holds if the vertical stretching is uniform across the unit interval
in x: α1 = α2 .

8.3.1.2 Dimension Formulas

For a general two-dimensional map with two positive Lyapunov exponents, 0 <
λ1+ ≤ λ2+ , we have, from Sect. 8.2.2, U = 2 = N, I = 0, and S = 0, so that Du,1 = 2
and Ds,1 = D1 . Depending on the value of κ relative to those of λ2+ and λ1+ , there
are two cases in which the dimension formula for D1 is different. The first case is
λ1+ < K1 < λ2+ , so J = 1 and we have

K1 − λ1+ κ
D1 = Ds,1 = 1 + + = 2− +. (8.36)
λ2 λ2

The second case is K1 < λ1+ , so J = 0. In this case, the dimension is

K1 λ2+ κ
D1 = Ds,1 = + = 1+ + − +. (8.37)
λ1 λ1 λ1

Because of the inequality (8.35), there are three distinct cases: (i) λb > κ > λa ,
(ii) κ > λb > λa , and (iii) κ > λa > λb , which should be treated separately. For
illustrative purpose, we set α2 = rα1 and calculate how the dimension D1 varies
with the parameter α1 . We have
8.3 Models Testing Dimension Formulas 277

κ = ln 2 + ln α1 − ln (1 + r−1),
λa = ln α1 + (1 + r)−1 ln r,
λb = ln 2. (8.38)

For case (i), the condition λb > κ > λa stipulates that ln α1 < ln (1 + r−1) ≡ ln αa .
The order of the Lyapunov exponents is then λ2+ = λb > λ1+ = λa . We have λ1+ <
K1 < λ2+ so that J = 1. Application of formula (8.36) gives

ln (1 + r−1) − ln α1
D1 = 1 + , for α1 < αa . (8.39)
ln 2

For case (ii), the defining condition κ > λb > λa is equivalent to ln αa < ln α1 <
ln 2 − (1 + r)−1 ln r ≡ ln αb . The order of the Lyapunov exponents is the same as
in case (i). However, we now have K1 = λ2+ + λ1+ − κ < λ1+ so that J = 0. The
corresponding formula (8.37) thus gives

(1 + r)−1 ln r + ln (1 + r−1)
D1 = , for αa < α1 < αb . (8.40)
ln α1 + (1 + r)−1 ln r

For case (iii), we have J = 0 and ln α1 > ln αb . The order of the Lyapunov exponents
is λ2+ = λa > λ1+ = λb , and (8.37) yields

ln (1 + r−1) + (1 + r)−1 ln r
D1 = , for α1 > αb . (8.41)
ln 2
Results (8.39), (8.40), and (8.41) are summarized schematically in Fig. 8.2. We see
that D1 > 1 for α1 < αa but D1 < 1 for α1 > αa . In fact, for α1 < αa , the chaotic re-
peller, which is the basin boundary between the y = ±∞ attractors, is a fractal curve,
as shown in Fig. 8.3. Numerical computation indicates [746] that for the repeller,

Fig. 8.2 For map (8.27), dimension of the chaotic repeller versus parameter α1 . For α1 < αa , the
information dimension D1 and the box-counting dimension D0 of the chaotic repeller are greater
than 1 and are equal. For α1 > αa , D1 < 1, but D0 = 1. Geometrically, there is a transition from a
fractal to a nonfractal behavior in the basin boundary as α1 is increased through αa
278 8 Transient Chaos in Higher Dimensions

Fig. 8.3 For map (8.27), basins of attraction of the y = +∞ (blank) and y = −∞ (black) attractors
for α1 = 1.1 and r = 3. The basin boundary, a chaotic repeller, is apparently a fractal curve. Both its
box-counting and information dimensions are D ≈ 1.28 [746] (with kind permission from Elsevier
Science)

the box-counting dimension D0 and the information dimension D1 are equal. For
α1 > αa , however, the repeller is a smooth curve with D0 = 1. Geometrically, there
is thus a transition from fractal to nonfractal behavior as α1 is increased through
αa . Numerically obtained values of the information dimension of the repeller agree
with those predicted by the formulas (8.40)–(8.41) [746]. For α1 > αa the natural
measure is rather irregular in spite of the fact that the support of the measure is a
smooth curve.

8.3.1.3 The Issue of Typicality

The dimension formulas (8.21)–(8.25) were conjectured to apply for typical systems
[347]. The two-dimensional map (8.27) provides a good example through which
the notions of typicality and atypicality can be understood. In particular, consider
η = 0. In this case, the line y = 0 is invariant in that a trajectory starting from
this line remains on it forever. This invariant subspace, which is the x-axis, is the
basin boundary in which the chaotic repeller resides. The natural measure is thus
distributed on the x-axis. Dividing the x unit intervals into 2n subintervals of width
(n)
2−n , the measure contained in each subinterval is given by μi in (8.32). The infor-
mation dimension of the natural measure for large n is (see (1.22))
n (n)
∑2 μi (n) ln 1/ μi
D1 = i=1 . (8.42)
ln 2n
Utilizing (8.32) and the fact that for large n, the typical (most probable) values of
n1 , n2 are
n1 α2 n2 α1
= , = , (8.43)
n α1 + α2 n α1 + α2
8.3 Models Testing Dimension Formulas 279

we obtain the same expression for D1 as in case (iii), (8.41). Thus for η = 0, the
information dimension is a constant for all α1 and α2 > 1. The dimension formulas
yield, however, different results for different ranges of α1 . In particular, for λa < λb ,
they provide larger values than the exact information dimension (8.41). The situation
η = 0 is thus atypical. However, as soon as we set η = 0, no matter how small, the
dimension formulas are recovered and (8.41) becomes valid for α1 > αb only. This
is so because the value of η = 0 can always be scaled to one by the change of
variable y → y/η .

8.3.2 A Chaotic Billiard Scatterer

We consider a three-dimensional billiard scattering system with dynamics similar to

those seen in the map example in the preceding discussion. As shown in Fig. 8.4a,
the system consists of an ellipsoid placed in an infinite tube in the z-direction with
cross section as shown in Fig. 8.4b. A free particle moving in the tube experiences
elastic bounces off the walls of the tube and off the surface of the ellipsoid as well.
Depending on the initial position and the initial velocity of the particle, it approaches
either z = +∞ or z = −∞. The boundaries between these two exit basins are lo-
cated near z = 0. Since the particle motion has three degrees of freedom, which
corresponds to a phase-space dimension of five (due to energy conservation), the
system represents a physical example in which high-dimensional chaotic scattering
can arise and the dimension formulas can be tested in the typical setting whereby
the ellipsoid is tilted slightly with respect to the z-axis. The symmetric system, in
which the major axis of the ellipsoid is the z-axis, represents an atypical situation

Fig. 8.4 (a) Billiard system consisting of an ellipsoid placed at z = 0 in an infinite tube in the
z-direction. (b) Cross sections of the tube and of the billiard at z = 0. The parameters are R = 25,
d = 10, and the radius of the ellipsoid at z = 0 is 5 [745, 746] (with kind permission from Elsevier
Science)
280 8 Transient Chaos in Higher Dimensions

whereby the dimension formulas could fail. This billiard system thus also represents
an example in which the mathematical notions of typicality versus atypicality can
be understood intuitively in physical terms.
To study the scattering dynamics, it is convenient to focus on bounces from the
ellipsoid. Setting particles at the unit speed and utilizing cylindrical coordinates
(z, φ ) and (vz , vφ ), Sweet and Ott derived a four-dimensional map relating these co-
ordinates at a bounce to the previous one from the ellipsoid [745,746]. If the particle
goes over the top (bottom) of the ellipsoid with vz > 0 (vz < 0), it continues toward
z = +∞ (z = −∞). Due to the inward wall curvature of the tube, there is a sensi-
tive dependence on initial conditions in the particle dynamics, signifying chaotic
scattering. This can be seen explicitly by considering the symmetric case in which
there is an invariant manifold Λ defined by z = 0 and vz = 0 in the four-dimensional
phase space, since particles started in Λ never leave it. The dynamics in the invari-
ant manifold Λ is that of a two-dimensional billiard shown in Fig. 8.4b, which is
hyperbolic and ergodic in the sense that almost every orbit comes arbitrarily close
to any point in the phase space. The invariant set Λ is nonattracting in the four-
dimensional phase space because almost all initial conditions in the vicinity of Λ
lead to trajectories that go to z = ±∞. In particular, if a cloud of initial conditions
is sprinkled in a region containing Λ , the fraction of trajectories that remain in this
region up to n bounces decreases exponentially with time as ∼ exp (−κ n), where
κ is the escape rate. For typical trajectories with respect to the natural measure on
Λ , there are two pairs of Lyapunov exponents, ±λφ and ±λz , which characterize
motions on the chaotic set and toward or away from it, respectively. The scattering
dynamics is thus chaotic with two positive Lyapunov exponents, and numerically
the inequality λφ > λz has been found [745, 746]. In this special configuration, the
full manifold Λ is a chaotic saddle.
For the case in which the ellipsoid is slightly tilted, the saddle survives, but its
geometry becomes more complicated. The Lyapunov exponents for typical trajecto-
ries with respect to the natural measure on the saddle are, however, approximately
the same as those in the untilted case.
The stable manifold of the saddle is physically important because it separates
the space of initial conditions into two regions that yield trajectories approaching
z = ±∞, respectively. These regions, the exit basins, can be determined numerically,
as shown in Fig. 8.5a, b for the untilted and slightly tilted cases, respectively, which
represent two-dimensional cross sections in the four-dimensional map. A straight-
forward application of the dimension formulas (8.21) and (8.24) with U = S = 2 in
this typical case yields

κ κ
Ds,1 = Du,1 = 4 − , D1 = 4 − 2 , for λφ > κ , (8.44)
λφ λφ

where J = 1. For λφ < κ so that J = 0, we have

κ − λφ κ − λφ
Ds,1 = Du,1 = 3 − , D1 = 2 − 2 . (8.45)
λz λz
8.3 Models Testing Dimension Formulas 281

Fig. 8.5 Examples of regions of initial conditions (exit basins) that yield trajectories to z → +∞
(white) and z → −∞ (black) in the two-dimensional cross section (x, z) defined by y = 5.1, vx = 0,
and vz = 0.1 for (a) the untilted case and (b) a small tilt of 2π /100 [745,746] (with kind permission
from Elsevier Science)

Note that in this second case the information dimension of the manifold (saddle) is
less than 3 (2). When plotting these dimensions as a function of κ /λφ one would
see a break at 3 (2), similar to that seen in Fig. 8.2 at αa .
For the untilted (atypical) case, a detailed analysis [746] gives that the dimension
is

λz + κ λz + κ
Ds,1 = Du,1 = 4 − , D1 = 4 − 2 , for λφ > λz + κ . (8.46)
λφ λφ

The meanings and relationship between the above two dimension formulas can
be understood as follows. Suppose one uses some algorithm to compute the dimen-
sion with refining accuracy ε . Then for an infinitesimal amount of tilt the true value
of the dimension as given by (8.44) can be obtained only when ε is small, e.g., for
ε < ε∗ . For resolution size greater than ε∗ , the small amount of tilt has no effect, so
that the value of the dimension extracted for ε > ε∗ would agree with that given by
(8.46). Around ε∗ , the scaling with ε is expected to show a crossover from the form
given by (8.46) to that given by (8.44). In a physical experiment with a finite resolu-
tion of distance scales, if the amount of tilt is small, the measured dimension may be
that given by (8.46). The true dimension can be recovered only in the ε → 0 limit.
Note that since the stable manifold of the saddle divides the four-dimensional
phase space, its box-counting dimension is at least three. Numerically, it was found
[746] that in the tilted and untilted configurations the stable manifold’s box-counting
dimensions are close to the information dimensions predicted by formulas (8.44)
and (8.46), respectively, insofar as these dimensions are larger than 3, as shown in
Fig. 8.6. Otherwise, the box-counting dimensions were found to remain at 3. The
two dimensions deviate here drastically, similar to the situation in the α > αa range
of Fig. 8.2.
282 8 Transient Chaos in Higher Dimensions

Ds 0
4.0
*
3.8
*
3.6 **
3.4

3.2

0.5 1
* 1.5 2

Fig. 8.6 Numerical verification of the stable manifold’s dimension for the tilted case (variable
κ /λφ ) and the untilted case (variable (κ + λz )/λφ ). The linear curve corresponds to formulas (8.44)
and (8.46) for Ds,1 > 3. Dots (stars) represent numerically determined values of the box-counting
dimension Ds,0 for the tilted (untilted) case

8.4 Numerical Method for Computing High-Dimensional

Chaotic Saddles: Stagger-and-Step

8.4.1 Basic Idea

In applications involving transient chaos, it is often useful and desirable to detect and
compute chaotic saddles. Several methods have been described in Chap. 1, but they
are applicable to systems with one unstable direction only. To numerically construct
chaotic saddles in higher dimensions with more than one unstable direction, two
methods are presently available. One is the “PIM-simplex” method by Moresco
and Dawson [529] and another is the “stagger-and-step” method by Sweet et al.
[744]. The PIM-simplex method is relatively sophisticated, and its applicability is
somewhat limited [529, 744]. The stagger-and-step method is, however, relatively
straightforward to implement and it is generally applicable to chaotic saddles that
are unstable in several dimensions. Here we focus on this method.
Consider an N-dimensional continuous map f, where N ≥ 2, and assume that a
chaotic saddle exists within a restraining region Γ that does not contain any attractor.
The transient lifetimes for initial conditions in Γ can be defined as follows. For
initial condition x, the escape time T (x) is the minimum n ≥ 0 for which the nth
iterate is in Γ but the (n + 1)th iterate of x is not in Γ . For points x on the stable
manifold of the chaotic saddle in Γ , the escape time is T (x) = ∞. If T (x) is finite but
large, x is close to the stable manifold. That is, all points with escape time at least n,
where n is large, belong to a small neighborhood of the stable set. This observation
is the main idea behind the stagger-and-step method.
A stagger is a perturbation r to a point x that results in a new point x + r such
that T (x + r) > T (x). The stagger method generates sequences {xn} of the form
8.4 Numerical Method for Computing High-Dimensional Chaotic Saddles: Stagger-and-Step 283

xn+1 = xn + rn , where rn is a stagger, such that T (xn+1 ) > T (xn ). Such sequences
are called stagger trajectories. The purpose is to find a point or a small set of points
that are sufficiently close to the stable manifold. To do so, one can specify some rela-
tively large δ > 0. Starting from n = 0, for each n, random perturbations r of magni-
tude less than δ are repeatedly chosen using some specified probability distribution
until one with T (xn +r) > T (xn ) is found. One can then set rn = r. The process stops
as soon as T (xn+1 ) > T ∗ , where T ∗ is a predetermined (large) time. Sometimes δ
may be too small so that no stagger can be found. In this case, one should increase δ .
In order to guarantee that stagger trajectories can be found in an efficient way,
the probability distribution from which r is chosen is important. A uniform distri-
bution, for instance, is not a good choice because the fraction of perturbations that
are staggers goes to zero exponentially fast as the escape time increases. This is
a consequence of the general exponential decay of transiently chaotic systems. To
overcome this difficulty, Sweet et al. suggested using an “exponential stagger distri-
bution” for choosing r, which can be realized as follows. Write 10−a = δ and let s
be a uniformly distributed random variable between a and b, where 10−b is the ac-
curacy of double precision in digital computers (typically b = 15). The choice of r is
thus r = 10−su, where u is a random directional unit vector. In so doing, the fraction
of staggers decreases much more slowly than exponentially, and hence the proba-
bility of finding a stagger can be enhanced significantly as compared with the case
of uniform distribution, thereby reducing the computation time.
After a stagger trajectory is found, a point x0 can be picked up for which T (x0 ) >
T ∗ . One can then generate a trajectory {xn } using the map f. The basic idea is to
apply the map only when xn has escape time T (xn ) > T ∗ . If T (xn ) ≤ T ∗ , then one
finds a nearby stagger point xn + rn with a higher escape time using δ = ε (say
10−10). The trajectory {xn } is of the form
⎧
⎨f(xn ) if T (xn ) > T ∗ (a step),
xn+1 = (8.47)
⎩f(x + r ) if T (xn ) ≤ T ∗ (rn is a stagger),
n n

where |rn | ≤ ε and T (xn + rn ) > T (xn ). (Note that T [f(xn )] = T (xn ) − 1.) Such
a trajectory is called a stagger-and-step trajectory. By construction, any stagger-
and-step trajectory {xn } satisfies |f(xn ) − xn+1 | < ε , so that {xn } is a numerical
trajectory with precision of order ε = 10−10, and it is close to the chaotic saddle
after a few iterates. From a stagger-and-step trajectory, dynamical invariants such as
the Lyapunov exponents of the saddle can be computed.
To give an example, Sweet et al. [744] considered the following four-dimensional
map:

xn+1 = A − x2n + Byn + k(xn − un),

yn+1 = xn ,
un+1 = C − u2n + Dvn + k(un − xn ),
vn+1 = un , (8.48)
284 8 Transient Chaos in Higher Dimensions

Fig. 8.7 Stagger-and-step method. (a) An example of the probability of finding a stagger versus
the escape time, (b, c) projections of a trajectory of 105 points of the chaotic saddle, for the four-
dimensional map (8.48) at the set of parameter values given in the text [744] (copyright 2001, the
American Physical Society)

where A, B, C, D, and k are parameters. For A = 3.0, B = 0.3, C = 5.0, D = 0.3,

and k = 0.4, there is transient chaos. The restraining region Γ was chosen to be
(−4, 4) × (−4, 4) × (−4, 4) × (−4, 4) and δ to be the length of the diagonal of Γ
(δ = 16). To find a stagger required about 50 choices of perturbed points r, the
probability of which is approximately 0.02, at least for a range of the escape time,
as shown in Fig. 8.7a. It can be seen that for escape time T0 between 5 and 29, this
probability is indeed approximately constant. The projections of a stagger-and-step
trajectory of 105 points on the chaotic saddle in the (x, y)-plane and in the (x, u)-
plane are shown in Fig. 8.7b and c, respectively, where ε = 10−10 and T ∗ = 30.
From this trajectory the Lyapunov exponents of the chaotic saddle were computed
to be λ2+ ≈ 1.33, λ1+ ≈ 0.77, λ1− ≈ 1.97, and λ2− ≈ 2.54.
The stagger-and-step method slows down with increasing dimensionality due to
the exponential growth of phase-space volume with dimension. A recent method
due to Bollt [87] replaces the random choice by a deterministic search for larger life-
times. To this end, one determines the lifetime function in the phase space. For any
point one can then find the direction along which the lifetime increases the fastest.
This gradient-search algorithm leads to solving an ordinary differential equation,
which essentially follows the unstable foliation toward the stable manifold of the
chaotic saddle. The method has been shown to be efficient for maps of dimension at
least eight.

8.4.2 Invariant Sets Constrained to Slow Manifolds

In high-dimensional systems one is often faced with the problem of the separa-
tion of time scales. In such multiscale systems the fast, high-frequency components
damp out rapidly due to dissipation, and the dynamics becomes restricted to a lower-
dimensional manifold embedded in the full phase space. This manifold is called the
slow manifold, the dynamics on which may turn out to be transiently chaotic. Mor-
gan, Bollt, and Schwartz [530] worked out a method to determine invariant sets in
slow manifolds.
8.4 Numerical Method for Computing High-Dimensional Chaotic Saddles: Stagger-and-Step 285

A paradigmatic form of multiscale systems is given by the differential equations

ẋ = F(x, y; μ ),
μ ẏ = G(x, y; μ ), (8.49)

where μ 1 is a small parameter characterizing the separation in time scales, and

the vector y denotes the fast variables. In the limit of extreme separation, μ = 0,
(8.49) reduces to an algebraic constraint: G(x, y, 0) = 0. Solving the constraint for y
yields an expression y = H0 (x). The graph of H0 is the slow manifold in this limit,
assumed to be single-valued for simplicity. The dynamics on the slow manifold is
obtained from (8.49) as
ẋ = F[x, H0 (x); 0]. (8.50)

For nonzero but small values of μ , the slow manifold is expected to persist. It is
given by the graph of a function labeled as y = Hμ (x), which can be obtained as an
expansion in powers of μ starting with H0 (x). The dynamics on this slow manifold
can also be obtained from a perturbation expansion starting with (8.50). For a suit-
ably defined map, the full dynamics can be written as (xn+1 , yn+1 ) = f(xn , yn ), and
the form of the slow manifold is yn = Hμ (xn ).
In order to construct the chaotic saddle on the slow manifold, Morgan et al. [530]
applied the stagger-and-step method with the following modifications. The restrain-
ing region Γ is chosen as a neighborhood of the slow manifold. Since orbits can
enter this region, one looks for the first escape time from Γ . The step-and-stagger
iterations are chosen as
⎧
⎨f[xn , Hμ (xn )] (a step),
(xn+1 , yn+1 ) = (8.51)
⎩f[x + r , H (x + r )] (a stagger),
n n μ n n

and the iterate yn+1 is projected back onto the slow manifold, so that the resulting
stagger-and-step trajectory lies near the slow manifold. In addition, since the slow
manifold is determined with finite precision (typically a power of μ ), the parameter
δ that sets the modulus of the stagger perturbation cannot be chosen to be less than
this accuracy.
The method has been successfully applied to a structural mechanical system in
[530] with two slow variables ψ1 , ψ2 . Figure 8.8 shows the chaotic saddle projected
on the plane of the slow variables.
A more complete picture can be obtained by plotting the slow manifold along
with the stable and unstable foliations on it. In the spirit of the sprinkler method
(Sect. 1.2.2.3), an approximation of these manifolds can be obtained by searching
for trajectories that remain near the slow manifold (and do not approach any attrac-
tor) for sufficiently long times. The initial points of such trajectories approximate the
stable manifold. The unstable manifold can be obtained, e.g., by applying the same
procedure to the time-reversed dynamics. The algorithm is called the constrained
invariant-manifold method. Figure 8.9 shows a case in which for simplicity, there is
286 8 Transient Chaos in Higher Dimensions

Fig. 8.8 Result of the modified stagger-and-step algorithm (8.51) applied to the problem of a
pendulum coupled to a viscoelastic rod. The chaotic saddle is shown on a stroboscopic map in the
plane of the slow variables ψ1 , ψ2 . The time-scale parameter is μ = 0.05, and the slow manifold is
specified with an accuracy of μ 2 [530] (copyright 2003, the American Physical Society)

Fig. 8.9 The slow manifold z1,n = Hμ (ψ1,n , ψ2,n ) (blue) in the problem of a pendulum coupled
to a viscoelastic rod as it appears on a stroboscopic map at the parameters of Fig. 8.8. The stable
(unstable) manifold is plotted in green (red) [530] (copyright 2003, the American Physical Society)

a single fast variable z1 . The slow manifold is a smooth surface in the phase space of
one fast and two slow variables. The invariant manifolds are given both in the slow
manifold and in the plane of the slow variables.
8.5 High-Dimensional Chaotic Scattering 287

8.5 High-Dimensional Chaotic Scattering

As a physical manifestation of high-dimensional transient chaos, we discuss a class

of three-degree-of-freedom, time-independent open Hamiltonian systems that ex-
hibit chaotic scattering. Issues to be addressed are the dimension requirement for
chaotic scattering to be physically observed and the topology of high-dimensional
chaotic scattering. While there have been attempts to address chaotic scattering in
higher-dimensional systems [124,366,444,745–748,829], the topic is relatively un-
explored and much research is needed.

8.5.1 Dimension Requirement for Chaotic Saddles

to be Observables

Due to the high dimensionality of the phase space, an issue of concern is whether
chaotic scattering can be observed even if there is a chaotic saddle of low dimen-
sion in the scattering region. In particular, suppose in a scattering experiment, one
measures a scattering function for particles launched from a one-dimensional line
segment. If the dimension of the chaotic saddle is not sufficiently high, its stable
manifold may not have generic intersections with the line. Such intersections, and
consequently a set of singularities in the scattering function, can be observed only
when the dimension of the chaotic saddle is sufficiently high.
To address this observability issue, we recall a basic mathematical statement con-
cerning the dimension of the intersection between two sets. Let S1 and S2 be two
subsets of an N-dimensional manifold with dimensions D(S1 ) and D(S2 ), respec-
tively."The dimension of the set of intersection between S1 and S2 is denoted by
D(S1 S2 ). The question is whether the sets S1 and S2 intersect generically in the
sense that the intersection cannot be removed by small perturbations. The natural
approach is to look at the dimension DI :

DI = D(S1 ) + D(S2) − N.

If DI ≥ 0, the intersection is generic, and the dimension of the set of intersection is

[232] #
D(S1 S2 ) = DI = D(S1 ) + D(S2) − N. (8.52)
If DI is negative, then S1 and S2 do not have a generic intersection. For example,
consider the intersection between two one-dimensional curves in a two-dimensional
plane: D(S1 ) = D(S2 ) = 1 and N = 2. We obtain DI = 0, which means that the
intersecting set consists of points, and the intersections are generic because small
perturbations cannot remove them. If, however, N = 3, then DI < 0, which means
that two one-dimensional curves do not intersect generically in a three-dimensional
space. If they intersect at a point, small perturbations in the positions of the lines
typically remove the intersection. These two cases, together with an additional one
288 8 Transient Chaos in Higher Dimensions

Fig. 8.10 Illustration of generic and nongeneric intersections of simple geometric sets: (a) D1 =
D2 = 1 and N = 2 (generic intersection), (b) D1 = D2 = 1 and N = 3 (nongeneric intersection),
and (c) D1 = 1, D2 = 2, and N = 3 (generic intersection)

(D(S1 ) = 1, D(S2 ) = 2, and N = 3), are illustrated in Fig. 8.10. It can also be said
that the codimension D − N is additive for generically intersecting sets. Formula
(8.52) is valid for both the box-counting and the information dimensions.
To apply these arguments to chaotic scattering, we consider a continuous-time
autonomous scattering system of phase-space dimension N + 2. Due to energy con-
servation, the corresponding continuous-time flow is (N + 1)-dimensional, so the
scattering map is N-dimensional. The box-counting dimensions of the stable and the
unstable manifolds in the map are denoted by Ds,0 and Du,0 . The symplectic nature
of the dynamics stipulates Du,0 = Ds,0 . The box-counting dimension of the chaotic
saddle is
D0 = Ds,0 + Du,0 − N = 2Ds,0 − N, (8.53)
since dynamically, the chaotic saddle is the intersecting set between the stable and
the unstable foliations. Let 0 < ds ≤ 1 be the box-counting dimension of the singu-
larities probed by a scattering function. This is the set of intersecting points between
the stable manifold of dimension Ds,0 and a one-dimensional line segment from
which particles are initiated in the N-dimensional map. Equation (8.52) implies

ds = Ds,0 + 1 − N, (8.54)
8.5 High-Dimensional Chaotic Scattering 289

or Ds,0 = ds + N − 1. Utilizing (8.53) gives the following formula relating the

dimension of the chaotic saddle to ds :

D0 − N + 2
ds = . (8.55)
2
Scattering singularities can be seen if ds > 0, which is fulfilled only if

D0 > N − 2. (8.56)

For a two-degree-of-freedom autonomous Hamiltonian system, one obtains a two-

dimensional map, N = 2, so the condition is always satisfied. In contrast, in a
three-degree-of-freedom system, N = 4, and D0 must be larger than 2 for chaotic
scattering to be observable. In this case, if the box-counting dimension of the saddle
in the map is smaller than 2, the set of singularities will not be observable, and as a
result, any measured scattering function will typically exhibit only smooth features.
This implies that by examining the scattering functions only, no chaotic behavior
can be revealed, even when there is a chaotic saddle in the phase space and the
scattering dynamics is chaotic.

8.5.2 Normally Hyperbolic Invariant Manifolds

in High-Dimensional Chaotic Scattering

For high-dimensional scattering systems, hyperbolic periodic orbits and their stable
and unstable manifolds often do not have the necessary dimensionality to partition
the phase space on the energy shell. In search of higher-dimensional structures with
features of periodic orbits of low-dimensional scattering, Wiggins and coworkers
[808, 811, 829] suggest the geometrical objects of normally hyperbolic invariant
manifolds [828]. On such a manifold, the expansion and contraction rates for the
invariant motion are dominated by those transverse to the manifold. Like a sad-
dle point, a normally hyperbolic invariant manifold has its own stable and unstable
manifolds. In Hamiltonian systems, normally hyperbolic invariant manifolds can
exist about equilibrium points of saddle-center- · · · -center type. In an n-degree-of-
freedom time-continuous system, such an equilibrium point possesses a pair of real
eigenvalues of opposite signs (say ±λ ) and 2n − 2 purely imaginary eigenvalues oc-
curring in complex conjugate pairs (±iω j , j = 2, . . . , n). In the (2n − 1)-dimensional
energy shell, the normally hyperbolic invariant manifold is a (2n − 3)-dimensional
sphere around the saddle-center- · · · -center type of fixed point.
Normally hyperbolic invariant manifolds are typical in systems with internal
degrees of freedom where some basic “transformation” can take place during the
scattering process. Chemical reactions provide a natural example in this context.
The basic transformation is then that reactants form products. The normally hyper-
bolic invariant manifold is the energy surface of an unstable invariant subsystem
290 8 Transient Chaos in Higher Dimensions

with one degree of freedom less than that of the full system. In the terminology
of chemistry, this subsystem is an activated complex, or an unstable supermolecule
[808]. This unstable subsystem is thus a proper generalization of the basic hyper-
bolic periodic orbits in two-degree-of-freedom systems (e.g., the orbits bouncing
between two disks in the three-disk scattering system; see Fig. 6.4). The activated
complex or the normally hyperbolic invariant manifold is located between reactants
and products. More precisely, around the normally hyperbolic invariant manifold
the phase space has the “bottleneck” property that facilitates the construction of a
dividing surface. This surface has the property of “no-recrossing” and therefore sep-
arates the phase-space region of reactants and products. The dividing surface is of
dimension 2n − 2 and contains as an invariant set the intermediate complex.
The stable and the unstable manifolds of the intermediate complex are (2n − 2)-
dimensional, which is one dimension less than that of the energy surface. These
manifolds can therefore act as separatrices, i.e., they enclose volumes of the energy
shell. Their key dynamical significance is that the only way for trajectories to lead
to reactions is for them to be in certain volumes enclosed by the stable and the
unstable manifolds. Note that any Poincaré section of the continuous dynamics in
the (2n − 1)-dimensional energy shell defines an (N = 2n − 2)-dimensional map.
The normally hyperbolic invariant manifold is an (N − 2)-dimensional object in this
map. Its stable and unstable manifolds have dimension N − 1.
The stable and the unstable manifolds of the intermediate complex can cross each
other. The homoclinic and heteroclinic intersections, as well as such intersections
of subsets of the normally hyperbolic invariant manifolds, can then form a high-
dimensional chaotic saddle.
If the chaotic saddle is formed by the intersections of the separatrix manifolds
of the intermediate complex, the scattering functions are similar to those in low-
dimensional systems. The locally (N − 1)-dimensional manifolds have, with finite
probability, intersections with any line of initial conditions: applying (8.52) with
"
D(S1 ) = N − 1, D(S2 ) = 1 to get D(S1 S2 ) = 0, we see that the typical intersec-
tions are points. Furthermore, in such cases the chaotic saddle’s full stable (unstable)
manifold has box-counting dimension > N − 1. As a consequence, the saddle, that
is the intersection of these manifolds, is of dimension D0 > N − 2. Equation (8.55)
gives then ds > 0, implying that the set of singularities is always observable if the
saddle is formed by normally hyperbolic invariant manifolds. The intersections of
the stable manifold with a line yield the endpoints of intervals of continuity in
the scattering function. They can thus be used to define scattering cross sections
(Appendix D).
Besides chemical reactions, normally hyperbolic invariant manifolds have been
applied to problems of celestial mechanics [806, 807] and to escape problems from
multidimensional potential wells [809]. It is likely that the concept can have a broad
range of potential applications in other contexts. A recent review of both the classi-
cal and quantum aspects of such transition-state theories can be found in [810].
8.5 High-Dimensional Chaotic Scattering 291

8.5.3 Metamorphosis in High-Dimensional Chaotic Scattering

A prototype system for investigating high-dimensional chaotic scattering is the

configuration in which scattering centers are located at the vertices of a regular tetra-
hedron [124, 444, 747, 748]. When the centers are hard-wall spheres [58, 124, 415,
535, 817], the problem corresponds to a light beam bouncing back and forth among
reflecting balls that can be found, for instance, in holiday decorations. This type of
chaotic scattering can be readily observed, and experiments have been carried out,
generating fractal images [747, 748]. While the hard-wall system is illuminating for
demonstrating the fractal structure associated with chaotic scattering in high dimen-
sions, the topology of the scattering is fixed, analogous to the planar three-hard-disk
scattering system. Basic issues such as bifurcations cannot be addressed using the
hard-wall systems.
Because of this difficulty, scattering systems consisting of physically realistic soft
potentials were considered [421, 444], which are relevant to the scattering of parti-
cles by molecules in the three-dimensional physical space. It was found [444] that
(1) the chaotic-scattering topology can undergo a sudden change (metamorphosis)
as a system parameter (e.g., energy) changes continuously, (2) at the metamorphosis,
the behavior of the box-counting dimension of the chaotic saddle changes charac-
teristically, and (3) chaotic scattering can occur in energy regimes for which it is not
possible in the corresponding planar scattering system. An exemplar system con-
sists of four potential hills located at the four vertices of a regular tetrahedron of
unit side length, as shown in Fig. 8.11. To mimic physical situations such as particle
scattering by nonrotating diatomic molecules, the Morse potential was chosen for
each hill [444]. The total potential of the scattering system is
4
V (x) = ∑ VM (x, x j ), (8.57)
j=1

Fig. 8.11 A schematic illustration of the scattering system: four Morse potential hills located at
the vertices of a regular tetrahedron [444] (copyright 2000, the American Physical Society)
292 8 Transient Chaos in Higher Dimensions

Fig. 8.12 Surfaces of equal energy for potential (8.57), for energies above and below the critical
energy Ec . (a) Isoenergy surface for E = 4 (E > Ec ); (b) Isoenergy surface for E = 1 (E < Ec )
[444] (copyright 2000, the American Physical Society)

where VM (x, x j ) isgiven by (6.19), x j ( j = 1, . . . , 4) denote the vertices of the tetra-

hedron, and r j = (x − x j )2 + (y − y j )2 + (z − z j )2 is the distance to vertex j. The
particle energy E is a convenient bifurcation parameter.
The region of the three-dimensional physical space classically inaccessible to
particles of energy E is given by V (x) > E. In order to have chaotic scattering,
the particle energy needs to be below the maximum energy Em of the potential
hills. For energy values larger than a critical energy Ec (and lower than Em ), the
inaccessible part of the physical space consists of four disconnected regions, each
surrounding one vertex of the tetrahedron, as shown in Fig. 8.12a. The regions
are approximately spherical in shape, but are not perfectly spherical because their
shapes are distorted by the other hills. As the energy decreases toward Ec , the inac-
cessible regions grow in radius, and for E = Ec , the previously disconnected regions
begin to connect with each other. For E ≤ Ec , the inaccessible regions are thus fully
connected, as shown in Fig. 8.12b. As E is decreased from Ec , a whole family of
orbits in the invariant set is destroyed. These are orbits that bounce back and forth
between each pair of forbidden regions an arbitrary number of times, including the
six unstable periodic orbits that connect each pair of hills existing for E > Ec . As
these orbits are destroyed, however, another family of orbits is created at E = Ec ;
these orbits bounce off the newly created forbidden regions connecting each pair
of hills. This topological change in the dynamics of the system at Ec can cause a
metamorphosis in the scattering dynamics [444].
The basic physics associated with the topological metamorphosis can be under-
stood in terms of the structural change in the “holes” on each side plane of the
tetrahedron potential configuration as the particle energy is decreased. Each side
plane of the tetrahedron potential (see Fig. 8.13) is similar to the two-dimensional
8.5 High-Dimensional Chaotic Scattering 293

1.5

0.5

y 0

−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
x

Fig. 8.13 Energy contours of the Morse potential in the (x, y)-plane at z = 0. The contours belong
to the energy values E = 4 (red), E = 3 (blue), E = 2 (green), and E = 1 (black). Other parameters
are α = 6 and re = 0.68. The value of the critical energy is Ec = 2.25 [444] (copyright 2000, the
American Physical Society)

system treated in Sect. 6.3.4. The critical energy Ec is close to the value of E0 deter-
mined in the two-dimensional case. It was found numerically [444] that Ec ≈ 2.25.
For E < Ec , the forbidden regions are connected, and an incoming particle can pen-
etrate the interior of the tetrahedron only through the holes on the side planes. The
holes in the central regions of the side planes are always present, because of the at-
tractive parts of each Morse hill. This allows particles to enter the scattering region
inside the tetrahedron at low energies. For E slightly below Ec , the holes are rela-
tively large, and hence the range of initial conditions with which particles can enter
the holes are appreciable. The holes, however, become smaller as E is decreased
further from Ec . For E slightly below Ec , the size of the hole can be estimated as

s ≈ s0 − C(Ec − E), (8.58)

√
where s0 = ( 3 − 1)/2 is the size for E = Ec and C is a positive constant. Thus,
to observe chaotic scattering at low energies in an experimental setting, initial con-
ditions have to be prepared carefully so that particles can enter the holes, since the
scattering will not be chaotic if the particles do not enter the holes. In fact, no un-
stable periodic orbit can be formed outside the holes for E < Ec .
To explore the scattering function, we note that, since the physical space is three-
dimensional, there are two angles characterizing the momentum of a scattering
particle: the azimuthal angle φ and the polar angle θ . Figure 8.14a, b show, for E = 4
and E = 1 respectively, φ after the √scattering versus b, where particles are launched
upward with vx0 = vy0 = 0, vz0 = 2E from z0 = −10.0, and the deflection angle
φ (x0 ) is recorded when the particles exit the scattering region. A Cantor set of sin-
gularities in the scattering dynamics implies the presence of a chaotic saddle whose
294 8 Transient Chaos in Higher Dimensions

Fig. 8.14 Deflection function: azimuthal angle φ as a function of the impact parameter for (a)
E = 4, and (b) E = 1 [444] (copyright 2000, the American Physical Society)

box-counting dimension in the underlying four-dimensional map is larger than two.

The dynamical and physical natures of the scattering observed at these energies are,
however, quite different, because of the topological change in the structure of the
invariant set. A computation of the dimension ds of the set of singularities in the
scattering function leads, by (8.55) to the box-counting dimension D0 of the chaotic
saddle. It was found [444] that for E = 4, the dimension is D0 = 3.33 ± 0.02, and
for E = 1, D0 = 2.83 ± 0.02. Figure 8.15 shows the dimension D0 of the saddle ver-
sus E for 1 ≤ E ≤ 4. For E > Ec , the dimension remains roughly constant. This is
due to the structural stability of the chaotic saddle in this energy regime, where the
potential hills remain isolated and hence there is no exponential change in the num-
ber of unstable periodic orbits. For E < Ec , D0 appears to decrease as E is lowered
from Ec .
The reason that the box-counting dimension decreases as the energy is decreased
from Ec can be understood heuristically as follows. Consider initial conditions on
a line that contains a Cantor set of singularities. The Cantor set corresponds to par-
ticle trajectories that can enter the holes in the side planes of the tetrahedron and
stay in the scattering region forever. Those that cannot enter the holes or enter the
holes but escape in finite time correspond to gaps, also called intervals of continuity,
between points in the Cantor set. Decreasing the size of the holes is equivalent to
enlarging these gaps. For a binary self-similar Cantor set of primary gap size Δ , its
box-counting dimension is
8.5 High-Dimensional Chaotic Scattering 295

Fig. 8.15 Box-counting dimension D0 of the chaotic saddle versus the energy E. The dimension
remains practically constant for E > Ec and decreases as E is decreased through Ec ≈ 2.25 [444]
(copyright 2000, the American Physical Society)

ln 2
ds = .
ln 2 − ln(1 − Δ )

Assuming that 1 − Δ is proportional to s, the size of the holes, we obtain, using

(8.58),
ds ∼ [A − B ln(s0 − C(Ec − E))]−1 , (8.59)
for E slightly below Ec , where A and B are positive constants. This scaling rela-
tion indicates that the dimension of the chaotic saddle decreases as the energy E is
decreased from the critical value Ec , due to the shrinkage of the holes in the side
planes of the potential configuration.
Notice that for energies about E < E0 ≈ Ec , chaotic scattering does not occur
if the system has only two degrees of freedom, due to the fact that the inaccessi-
ble regions are connected. Thus, for two-degree-of-freedom Hamiltonian systems
(corresponding to two-dimensional area-preserving maps), no particle coming from
outside the scattering region can enter the bounded, triangular-like region formed
at the center of the potential hills. The dynamics in the bounded triangular-like re-
gion is typically made up of chaotic seas mixed with KAM tori. Although there is
bounded chaos in this case, it is not accessible to particles from outside, and hence
there is no chaotic scattering (Fig. 6.20a). In the case of three-degree-of-freedom
systems (four-dimensional maps), in the same energy range, the corresponding clas-
sically allowed bounded region in the center of the potential hills is accessible to
scattering particles coming from outside. Chaotic scattering is thus possible, which
for this class of scattering systems is uniquely a high-dimensional phenomenon.
296 8 Transient Chaos in Higher Dimensions

8.5.4 Topological Change Accompanying the Metamorphosis

The topology of chaotic scattering can be studied by examining the structure of exit
basins. For the tetrahedron configuration, there are four side planes through which
particles can exit. Suppose a large number of particles is launched toward the scat-
tering region from a two-dimensional area in an (x-y)-plane at some large negative
z position. After the scattering, the particles in the initial plane can be color-coded,
depending on through which side plane they exit the system. Figure 8.16a shows,
for E = 4, the basin structure in the area defined by (−0.4 ≤ x0 , y0 ≤ 0.4) in the

a 0.4

0.3

0.2

0.1
y0

−0.1

−0.2

−0.3

−0.4
−0.4 −0.2 0 0.2 0.4
x0
b
0.2
0.15
0.1
0.05
0
y0

−0.05
−0.1
−0.15
−0.2
−0.25
−0.2 −0.1 0 0.1 0.2
x0

Fig. 8.16 Exit basins of scattering trajectories for (a) E = 4, and (b) E = 1. In (a), the basin
boundaries common to the four colors consist of completely isolated points only. In (b), the part of
the basin boundary common to the four colors is connected and it is Wada. See text for the meaning
of the color coding [444] (copyright 2000, the American Physical Society)
8.5 High-Dimensional Chaotic Scattering 297

plane located at z0 = −10.0, where 500 × 500 particles uniformly distributed in the
initial area are launched toward the scattering region along the +z-direction [444].
If a particle exits through the plane defined by vertices (1, 2, 4) shown in Fig. 8.11
(or (1, 3, 4), or (1, 2, 3), or (2, 3, 4)), its location in the initial two-dimensional area
is marked by red (or yellow, or blue, or green). As can be seen from Fig. 8.16a,
the boundary contains isolated points where the four colors meet, but for almost all
points on the boundary only two colors meet. The basin boundary is the set of in-
tersecting points between the stable manifold of the chaotic saddle with the initial
plane. For this energy, then, the common boundary points with different colors are
isolated. As the energy is lowered, the previously classically forbidden regions be-
come connected, and extended parts of the basin boundary points are now common
to the four colors, as shown in Fig. 8.16b for E = 1. This is the Wada property of
basins (Sect. 5.5). The topology of the basin undergoes a sudden change (metamor-
phosis) from being disconnected to being Wada at the critical energy value Ec .
For E > Ec , when the forbidden regions are disconnected (Fig. 8.12a), a scat-
tering trajectory will typically enter the scattering region, bounce off the forbidden
regions a number of times, and leave. As the particle leaves the scattering region, it
crosses one of the side planes shown in Fig. 8.11. For typical trajectories, one can
continuously change the initial conditions so as to cause a continuous change in the
trajectory (this is not true if the initial condition lies on the stable manifold of the
saddle, but such points have zero measure in the phase space). That is, there are paths
in the space of initial conditions for which the escape parameters (such as escape
angles, lifetime, etc.) change continuously, and these paths contain all initial condi-
tions except for a set of measure zero. Now consider a subspace M in the full space
of initial conditions, which can be chosen to have dimension two or higher. Consider
one such path, denoted by C, that connects two points a and b in M belonging to two
different escape basins, denoted by S1 and S2 . Distinct side planes defining the dif-
ferent escapes are separated by segments that connect two adjacent triangular faces
(Fig. 8.11). For E > Ec , parts of these segments lie outside the forbidden regions.
Therefore, the path C in M can be chosen such that the corresponding trajectories go
from one escape to a neighboring one continuously, without going through any other
escape; in other words, all points in C belong to either S1 or S2 . This corresponds
to a basin boundary that separates only two escapes, and therefore to a non-Wada
basin.
The picture described above completely changes when the energy goes below Ec .
The forbidden regions are now connected into one single region, and the boundaries
between the side planes that define the different escapes lie entirely within it. One
can no longer go smoothly from one escape to another by a continuous change of
initial conditions, due to the presence of forbidden regions separating the escape
routes.
The change in the topological structure of the escape basin described above
is possible only in three-dimensional physical space: it does not happen for two-
degree-of-freedom systems, in which the basin boundaries in systems with three or
more escapes typically have the Wada property (Sect. 5.5). The reason is that the for-
bidden regions separating distinct escapes can never be bypassed from one escape
channel to another.
298 8 Transient Chaos in Higher Dimensions

8.6 Superpersistent Transient Chaos: Basics

The type of transient chaos discussed so far, such as that induced by a boundary cri-
sis, is characterized by the familiar algebraic scaling law (3.2) of its escape rate κ as
a function of parameter variations in p. There exists, however, another distinct class
of transient chaos: superpersistent transient chaos, characterized by the following
scaling law:
κ ∼ exp [−C(p − pc)−χ ], p > pc , (8.60)
where C > 0 and χ > 0 are constants. As p approaches the critical value pc from
above, the transients become superpersistent in the sense that the exponent in the
average transient lifetime

1
τ≈ ∼ exp [C(p − pc)−χ ], p > pc ,
κ
diverges in an exponential-algebraic manner. This type of transient chaos is quite
common in high-dimensional systems. Its origin can, however, be understood in
simple models. For illustrative purposes we therefore shall again use here low-
dimensional maps.
Superpersistent transient chaos was conceived to occur through the dynamical
mechanism of unstable–unstable pair bifurcations. The same mechanism causes
a riddling bifurcation that creates a riddled basin, so superpersistent chaotic tran-
sients can be expected at the onset of riddling [711]. It was shown that noise can
also induce superpersistent chaotic transients under certain conditions. For a recent
review, see [440].

8.6.1 Unstable–Unstable Pair Bifurcation

Unstable–unstable pair bifurcation represents a generic mechanism for superpersis-

tent chaotic transients [294,295,455]. One can imagine two unstable periodic orbits
of the same period, one on the chaotic attractor and another on the basin boundary,
as shown in Fig. 8.17a. As a bifurcation parameter p reaches a critical value pc , the

a b
Basin boundary

Unstable−unstable
pair

Chaotic attractor Chaotic transient

Fig. 8.17 Schematic illustration of an unstable–unstable pair bifurcation. (a) Invariant sets for
p < pc : a chaotic attractor, the basin boundary, and the pair of unstable periodic orbits. (b) For
p > pc , an escaping channel is created by an unstable–unstable pair bifurcation that converts the
originally attracting motion into a chaotic transient
8.6 Superpersistent Transient Chaos: Basics 299

two orbits coalesce and disappear simultaneously, leaving behind a “channel” in the
phase space through which trajectories on the chaotic attractor can escape, as shown
in Fig. 8.17b. The chaotic attractor is thus converted into a nonattracting chaotic set,
but the channel created by this mechanism is typically supernarrow [294, 295, 455].
Suppose that on average, it takes time T (p) for a trajectory to travel through the
channel in the phase space. We expect the tunneling time T (p) to be infinite for
p = pc , but for p > pc , the time becomes finite and decreases as p is increased from
pc . For p above but close to pc , the tunneling time can be long.
From Fig. 8.17a, we see that if the phase space is two-dimensional, the periodic
orbit on the attractor is a saddle and that on the basin boundary is a repeller. This can
arise only if the map is noninvertible. Thus, the unstable–unstable pair bifurcation
can occur in noninvertible maps of at least dimension two, or in invertible maps of
at least dimension three (or in flows of dimension at least four).
Let λ1 > 0 be the largest average Lyapunov exponent of the chaotic attractor.
After an unstable–unstable pair bifurcation the opened channel is locally transverse
to the attractor. A trajectory that spends time T (p) in the channel opened up at an
unstable periodic orbit on the attractor, the mediating orbit involved in the unstable–
unstable pair bifurcation, must come to within distance of about exp [−λ1 T (p)] from
this orbit. The probability for this to occur is proportional to exp[−λ1 T (p)]. The av-
erage time for the trajectory to remain on the earlier attractor, or the average transient
lifetime, can be related to the tunneling time as
τ (p) ∼ exp [λ1 T (p)],
or equivalently,
κ (p) ∼ exp [−λ1 T (p)]. (8.61)
The tunneling time thus determines the scaling of the escape rate with the parameter
variation.
Since the escaping channel is extremely narrow, the dynamics in the channel
is approximately one-dimensional along the direction from the mediating periodic
orbit to the orbit on the basin boundary, as schematically shown in Fig. 8.17. The
basic dynamics can be captured through the following simple one-dimensional map:

xn+1 = xk−1
n + xn + p, (8.62)

where x denotes the dynamical variable in the channel, k ≥ 3 is an odd integer to

generate two real fixed points, and p is a bifurcation parameter with critical point
pc = 0. For p < pc = 0, the map has a stable fixed point xs = −|p|1/(k−1) and an
unstable fixed point xu = |p|1/(k−1) , which collide at pc and disappear for p > pc ,
mimicking an unstable–unstable pair bifurcation. Since for 0 < p 1, T (p) is large
(see Sect. 2.4), the map (8.62) can be approximated by the following differential
equation:
dx
= xk−1 + p. (8.63)
dt
300 8 Transient Chaos in Higher Dimensions

Suppose the root of the channel is x = 0 and its length is l. The tunneling time is
then  l
dx k−2
T (p) ≈ ∼ p− k−1 . (8.64)
0 x
k−1 +p
Substituting (8.64) into (8.61) gives
 k−2

κ (p) ∼ exp −Cp− k−1 , (8.65)

where C > 0 is a constant. We see that as p approaches the critical value pc = 0

from above, the average transient lifetime diverges in an exponential-algebraic way,
giving rise to superpersistent transients. The exponent χ in the scaling law (8.60)
thus assumes the value (k − 2)/(k − 1) < 1.
To give a concrete example, we use the class of noninvertible two-dimensional
maps considered by Grebogi, Ott, and Yorke [294, 295]:

θn+1 = 2θn mod 2π , zn+1 = azn + z2n + β cos θn , (8.66)

where a and β are parameters. Because of the z2n term in the z-equation, for large zn
we have |zn+1 | > |zn |. There is thus an attractor at z = +∞. Near z = 0, depending
on the choice of the parameters, there can be either a chaotic attractor or none.For
instance, for 0 < β 1, there is a chaotic attractor near z = 0 for a < ac = 1 − 2 β ,
and the attractor becomes a chaotic repeller for a > ac [294]. The chaotic attractor,
its basin of attraction, and part of the basin of the infinity attractor are shown in
Fig. 8.18.

Fig. 8.18 Phase space of the two-dimensional map model (8.66): a chaotic attractor near z = 0
(black), its basin of attraction (blank), and the basin of the attraction of the attractor at z = +∞
(black) for a = 0.5 and β = 0.04, before the unstable–unstable pair bifurcation (ac = 0.6). The
fixed points z± are marked [294] (copyright 1983, the American Physical Society)
8.6 Superpersistent Transient Chaos: Basics 301

Following the argument leading to the scaling law (8.65), one can see that
the map (8.66) allows for superpersistent transients for a > ac . In particu-
lar, for a < ac there are
  two fixed points: (0, z+ ) and (0, z− ), where z± =

1 − a ± (1 − a) − 4β /2, on the basin boundary and on the chaotic attrac-
2

tor, respectively. They coalesce at a = ac . For a > ac , a channel is created through

which trajectories on the original attractor can escape to the attractor at infinity. At
the location of the channel where θ = 0, the z-mapping becomes

zn+1 = azn + z2n + β .

Letting δ = z − z∗ , where z∗ = (1 − a)/2, we obtain

δn+1 = δn2 + δn + b, (8.67)

 
with b = β (a − ac)− [(a − ac)/2]2 . For a ≈ ac , we have b ≈ β (a − ac). Equation
(8.67) is identical to (8.62) with k = 3. The integral (8.64) then yields T ≈ π b−1/2/2.
The Lyapunov exponent is determined by the θ -dynamics: λ1 = ln 2. Using (8.61),
we obtain the scaling of the escape rate for a > ac as
−1/2 −1/2
κ (a) ∼ e−T ln 2 ≈ e(−π ln 2/2)b ≈ e−C(a−ac ) , (8.68)

where C = π (ln 2)β −1/4 /2 is a positive constant.

8.6.2 Riddling Bifurcation and Superpersistent

Chaotic Transients

In Sect. 5.7, the phenomenon of riddling, or riddled basins in dynamical systems

with symmetry, is described. The presence of symmetry often leads to an invariant
subspace. In the absence of symmetry-breaking or random perturbations, a trajec-
tory originated in the invariant subspace remains there forever. Situations can also
be expected whereby a chaotic attractor lies in the invariant subspace. As discussed,
one such example is the system of coupled, identical chaotic oscillators. The syn-
chronization manifold is naturally a low-dimensional invariant subspace in the full
phase space that can be high-dimensional if the number of oscillators is large. If an-
other attractor exists outside the invariant subspace, riddling can occur in the sense
that the basin of the chaotic attractor in the invariant subspace is riddled with holes of
all sizes that belong to the basin of the other attractor. Imagine the situation in which
all unstable periodic orbits embedded in the chaotic attractor are stable with respect
to perturbations in the direction transverse to the invariant subspace. In this case, al-
most all initial conditions in the vicinity of the invariant subspace lead to trajectories
that end up asymptotically on the chaotic attractor. Riddling bifurcation refers to
the situation in which when a system parameter changes, an unstable periodic orbit
302 8 Transient Chaos in Higher Dimensions

(usually of low period) embedded in the chaotic attractor becomes transversely

unstable. An immediate physical consequence of the riddling bifurcation is that
when there is a small amount of symmetry-breaking, an extraordinarily low fraction
of the trajectories in the invariant subspace diverge, as shown in Fig. 5.21.
Due to nonlinearity, a “tongue” opens at xp , allowing trajectories near the invari-
ant subspace to escape for p > pc , as shown in Fig. 5.20b. Each preimage of xp also
develops a tongue simultaneously. Since preimages of xp are dense in the invariant
subspace, an infinite number of tongues open up simultaneously at p = pc , indicat-
ing that initial conditions arbitrarily close to the invariant subspace can approach
another attractor. Trajectories in the chaotic attractor remain there even for p > pc ,
since the subspace in which the chaotic attractor lies is invariant and each tongue has
zero width there. But trajectories near the chaotic attractor have a finite probability
of being in the open and dense set of tongues. Trajectories having initial conditions
in the tongues approach asymptotically the other attractor. The basin of attraction
of the chaotic attractor is thus a Cantor set of leaves of positive Lebesgue measure
(a fat fractal), signifying riddling. Physically, since the onset of riddling induces the
creation of these supernarrow tongues near the invariant subspace, it leads to su-
perpersistent chaotic transient behavior in the vicinity of the chaotic attractor. For
points chosen at random at a small distance d from the attractor, the probability of
not being attracted depends on the distance d as

P(d) ∼ exp [−Kd −γ ], (8.69)

where γ > 0 is a positive exponent, and K > 0 is a constant. In the presence of

symmetry-breaking as characterized by the parameter ε , a similar argument leads to
the escape rate
κ (ε ) ∼ exp [−K ε −γ ], (8.70)

for initial conditions in the original invariant subspace.

To make these ideas more concrete, it is convenient to use the following extension
of the noninvertible two-dimensional map (5.24) [455]:

xn+1 = rxn (1 − xn), (8.71)

yn+1 = ε + pe−b(xn −xp yn + y3n ,
)2

where for ε = 0, y = 0 defines the invariant subspace, r, b > 0 are parameters, and
p is the bifurcation parameter. The broken symmetry is with respect to y → −y.
The dynamics in the invariant subspace is described by the logistic map xn+1 =
rxn (1 − xn), for which chaotic attractors can arise.
In the symmetric case (ε = 0), the two eigenvalues of the unstable fixed point xp
(xp = 1 − 1/r, y = 0) are (2 − r, p). Thus, xp is stable in the y direction for p < 1 and
unstable for p > 1. This fixed point is a saddle for r > 3 and√ p < 1. For p < 1, there
are two other unstable fixed points located at r± ≡ (xp , ± 1 − p). These two fixed
points have eigenvalues (2 − r, 3 − 2p), both pure repellers for r > 3 and p < 1, as
8.6 Superpersistent Transient Chaos: Basics 303

shown in Fig. 5.20a. The two repellers collide with each other and with the saddle
at p = pc = 1 in a saddle-repeller bifurcation. They do not exist for p > 1. Thus, for
p > 1, two tongues, symmetrically located with respect to the invariant subspace,
open up at x = xp , allowing trajectories near y = 0 to escape to |y| = ∞, since the
cubic term in the y−dynamics guarantees that if |yn | > 1, then |yn+1 | > |yn | > 1.
Once a trajectory reaches |y| = 1, its y value approaches infinity rapidly. So |y| = ∞
can be regarded as the second attractor of (8.71) besides the chaotic attractor in the
y = 0 plane (invariant subspace), which exists for r > 3.6.
When there is symmetry-breaking (ε > 0), trajectories can leave the original
chaotic attractor at y = 0 (y = 0 is no longer an invariant subspace), and hence
the attractor is converted into a chaotic saddle. Simulation of (8.71) showed that
only an exceedingly small fraction of the points at y = 0 diverges toward the |y| = ∞
attractor. The transient time can easily be longer than, say, 105 iterations even for
ε = 0.014. To obtain the scaling of the escape rate with the symmetry-breaking pa-
rameter, the first step is to estimate, for ε ≥ 0, the size δ of the tongue at y = 0 for a
trajectory of transient time T (ε ). Since the y = 0 attractor is chaotic, its maximum
Lyapunov exponent λ1 is positive. Let Lu = eλ1 > 1, which is the stretching factor
for an infinitesimal vector in the x direction. That the transient time, the time needed
to reach a distance of order unity, is T implies δ LTu ≈ 1, which gives

δ ≈ Lu −T . (8.72)

The next step is to examine the probability that a trajectory falls into the tongue of
size δ at y = 0 for ε ≥ 0, which is proportional to δ . The average time for a trajectory
to fall into the tongue is

τ ∼ δ −1 ≈ Lu T = exp (λ1 T ). (8.73)

The final step is to evaluate T , the time it takes for the trajectory to exit once
it has fallen into the tongue. Near xp , we have exp [−b(xn − xp)2 ] ≈ 1. For initial
conditions chosen at y0 = 0, the trajectory satisfies yn ≥ ε for n ≥ 1. For small ε it
takes many iterations for a trajectory to reach y = 1. Thus, the y-dynamics within
the tongue can be approximated by the differential equation

dy
= ε + y3 .
dt
This gives
 1  ε −1/3
dy dz
T= = ε −2/3 . (8.74)
0 ε + y3 0 1 + z3
For ε → 0 the integral converges to C = π /33/2. Substitution of this expression into
(8.73) and using κ ≈ 1/τ yields the following scaling of the escape rate:

κ ∼ exp[−Cλ1 ε −2/3 ], (8.75)

304 8 Transient Chaos in Higher Dimensions

Fig. 8.19 Mushroom-shaped phase-space regions (tongues) through which trajectories escape the
y = 0 chaotic attractor (r = 3.8) at p = 1.18 > pc = 1 and ε = 0.005 in (8.71) [455] (copyright
1996, the American Physical Society)

which is the scaling law (8.70) with K ≡ Cλ1 . In (8.75), the exponent 2/3 is a
consequence of the y3 term in the y-dynamics. If this term is replaced by, say, a y2
term, the exponent will be 1/2. Thus, the exponent 2/3 in (8.75) is specific to the
two-dimensional map model (8.71). However, the scaling relation (8.70) is general,
with the exponent γ being positive.
The escaping behavior of trajectories, once they have fallen into the tongue, can
be seen by monitoring their traces in the phase space before they reach y = 1. Since
the tongues are supernarrow at p = pc , it is numerically convenient to examine the
case of p > pc , but for p close to pc . Figure 8.19 shows the last 50 points of 600
trajectories before they reach y = 1. There is a “mushroom-shaped” (tongue) crowd
of trajectory points in the phase space located above the fixed point xp ≈ 0.7368. The
solid curves in Fig. 8.19 indicate the envelope of the tongue, which can be derived
analytically by considering the escaping dynamics in the vicinity of xp . Specifically,
after a trajectory falls into the escaping channel located at xp , its dynamics can be
approximated by

(xn+1 − xp ) ≈ (2 − r)(xn − xp ), yn+1 ≈ ε + pyn + y3n .

By introducing zn ≡ |xn − xp |, the x-dynamics becomes zn+1 = |2 − r|zn = (r − 2)zn .

For p close to pc and ε small, it takes an extremely large number of iterations for a
typical trajectory to escape due to the long chaotic transient. We thus have

dz/dt = (r − 3)z, dy/dt = ε + (p − 1)y + y3. (8.76)

8.7 Superpersistent Transient Chaos: Effect of Noise and Applications 305

This implies that for ε ≈ 0, we have dz/dy = (r − 3)z/[(p − 1)y + y3], which leads to
 (r−3)/(p−1)
y
z(y) =  , for p > pc = 1. (8.77)
(p − 1) + y2

The solid curves in Fig. 8.19 are [xp ± z(y)], respectively. They represent the enve-
lope of the tongue reasonably well.

8.7 Superpersistent Transient Chaos: Effect of Noise

and Applications

8.7.1 Noise-Induced Superpersistent Chaotic Transients

In the general setting, in which an unstable–unstable pair bifurcation can occur,

noise can induce superpersistent transients preceding the bifurcation. Consider, in
the noiseless case, a chaotic attractor in its basin of attraction (p < pc ). When noise
is present, there can be a nonzero probability that two periodic orbits, one belonging
to the attractor and the other to the basin boundary, can get close and coalesce tem-
porally, giving rise to a nonzero probability that a trajectory on the chaotic attractor
crosses the basin boundary and moves toward the basin of another attractor. Tran-
sient chaos thus arises even for p < pc . Due to weak noise, the channels through
which trajectory escapes the chaotic attractor open and close intermittently in time.
Escaping through the channel requires the trajectory to stay in a small vicinity of
the opening of the channel for a finite amount of time, which occurs with extremely
small probability. The creation of the channel by noise and the noisy dynamics in
the channel are thus the key ingredients to the noise-induced transient behavior.
For a two-dimensional phase space, the situation described above is schemati-
cally illustrated in Fig. 8.17a, b for the cases in which noise is respectively absent
and present. If the attractor is close to the basin boundary, noise of strength σ can in-
duce an unstable–unstable pair bifurcation, creating a narrow channel through which
trajectories can escape, as shown in Fig. 8.17b even for p < pc . As (8.61) suggests,
the escape rate can be expressed in terms of the tunneling time T (p, σ ) as

κ (p, σ ) ∼ exp [−λ1 T (p, σ )], (8.78)

where λ1 > 0 is the largest Lyapunov exponent of the original chaotic attractor.
Again, since the escaping channel is extremely narrow, for T (p, σ ) large, the dy-
namics is approximately one-dimensional in the channel along which the periodic
orbit on the attractor is stable but the orbit on the basin boundary is unstable for
p < pc (Fig. 8.17a). This feature can thus be captured through the stochastic version
of the map (8.62):
n + x n + p + σ ξn ,
xn+1 = xk−1 (8.79)
306 8 Transient Chaos in Higher Dimensions

Fig. 8.20 Dynamics of the map (8.79), for k = 3 (a) in the subcritical cases (p < 0), and (b) in
the supercritical case (p > 0). Upper graph: deterministic map (σ = 0), lower graph: quasipoten-
tial Φ (x)

where σ 1 is the noise amplitude and ξn is a Gaussian random variable of zero

mean and unit variance. For T  1, (8.79) can be approximated by the Langevin
equation
dx 1 dΦ
= xk−1 + p + σ ξ (t) ≡ − + p + σ ξ (t), (8.80)
dt 2 dx
where the function
Φ (x) = −2(xk /k + px) (8.81)

is the associated quasipotential of the one-dimensional problem (see Sect. 4.2).

For p < 0, the underlying deterministic system for (8.80) has a stable fixed point
xs = −|p|1/(k−1) and an unstable fixed point xu = |p|1/(k−1) . For p > 0, there are
no fixed points, as shown in Fig. 8.20. It is convenient to define the opening xr
of the channel at the stable fixed point xr = xs when it exists, i.e., for p < 0, and
set xr = 0 otherwise. A properly formulated first-passage-time problem for this
one-dimensional stochastic process can yield the scaling of T (p, σ ) [198, 199] as
follows.
8.7 Superpersistent Transient Chaos: Effect of Noise and Applications 307

Let P(x,t) be the probability density function of the stochastic process described
by (8.80) that satisfies the Fokker–Planck equation:

∂ P(x,t) ∂ σ 2 ∂ 2P
= − [(xk−1 + p)P(x,t)] + . (8.82)
∂t ∂x 2 ∂ x2

Let l be the effective length of the channel in the sense that a trajectory with x > l is
considered to have escaped the channel. The tunneling time T (p, σ ) required for a
trajectory to travel through the channel is equivalent to the mean first passage time
from the opening xr of the channel to l. For an escaping trajectory, once it falls
into the channel through xr , it will eventually exit the channel at x = l without even
going back to the original chaotic attractor. This is so because the probability for a
trajectory to fall into the channel and then to escape is already exponentially small,
(8.78), and hence the probability for any “second-order” process to occur, whereby
a trajectory falls into the channel, moves back to the original attractor, and falls
back in the channel again, is negligible. For trajectories in the channel there is thus
a reflecting boundary condition at x = xr ,

σ 2 ∂ P 
(x k−1
+ p)P(x,t) − = 0, (8.83)
2 ∂ x x=xr

and an absorbing boundary condition at x = l,

P(l,t) = 0. (8.84)

Assuming that trajectories initially are near the opening of the channel (but in the
channel), we have the initial condition

P(x, xr ) = δ (x − x+
r ). (8.85)

Under these boundary and initial conditions, the solution to the Fokker–Planck
equation yields the following mean first-passage-time [256, 640] for the stochastic
process (8.80):
 l  y
2 Φ (y) −Φ (y )
T (p, σ ) = dy exp exp dy . (8.86)
σ2 xr σ2 xr σ2

The double integral in (8.86) can be carried out [198,199] for the three distinct cases
critical (p = 0), supercritical (p > 0), and subcritical (p < 0).
In the weak-noise regime (σ σc ∼ |p|k/(2(k−1)) ), the results can be summa-
rized as ⎧
⎪
⎪ −(k−2)/(k−1) , p > 0,
⎨p
T (p, σ ) ∼ σ −(2−4/k) , p = 0, (8.87)
⎪
⎪
⎩|p|−(k−2)/(k−1) exp (|p|k/(k−1) /σ 2 ), p < 0.
308 8 Transient Chaos in Higher Dimensions

These laws imply the following scaling laws for the escape rate of the chaotic
transients in various regimes (substituting the expressions of T (p, σ ) in (8.78)):
⎧
⎪ −(k−2)/(k−1)
⎨exp [−Cp
⎪ ], p > 0,
−(2−4/k)
⎪  σ
κ (p, σ ) ∼ exp [−C ],  p = 0, (8.88)
⎪
⎩exp −C|p|−(k−2)/(k−1) exp[|p|k/(k−1) /σ 2 ] , p < 0.

The general observation is that different behaviors arise depending on the bifur-
cation parameter p: independent of noise for the supercritical regime, normally
superpersistent for the critical case, and extraordinarily superpersistent for the sub-
critical regime in the sense of scaling in (8.88) (for p < 0). Numerical support for
these distinct scaling behaviors was obtained [198, 199].

8.7.2 Application: Advection of Inertial Particles in Open

Chaotic Flows

We will see in Chap. 10 that the inertia of the advective particles alters the advective
dynamics, and the underlying dynamical system becomes dissipative so that attrac-
tors can arise, and hence particles can be trapped permanently in some region in
the physical space. The possibility that toxin particles can be trapped in physical
space is particularly worrisome. It is thus interesting to study the structural stability
of such attractors. In particular, can chaotic attractors so formed be persistent under
small noise? It was found [197] that in general, the attractor is destroyed by weak
noise and replaced by a chaotic transient, which is typically superpersistent. For
weak noise, the extraordinarily long trapping time makes the transient particle mo-
tion practically equivalent to an attracting motion with similar physical or biological
effects.
Noise in the context of particle advection can be due, for example, to the diffusiv-
ity of the advected particles [143]. Diffusion can be generated by, e.g., microscopic
interactions (molecular diffusion). The dimensionless equation of motion of an in-
ertial particle will be detailed in Chap. 10. The noisy version of this equation in a
two-dimensional incompressible flow u(x, y,t) in the horizontal plane is

3 d
r̈(t) = A [u(r(t),t) − ṙ(t)] + R u(r(t),t) + σ ξ (t), (8.89)
2 dt

where ξ (t) = (ξx (t), ξy (t)), and ξx (t) and ξy (t) are independent Gaussian random
variables of zero mean and unit variance, and σ is the noise amplitude. Parameters
A and R are the inertial parameter and the density ratio, respectively, as defined
in (10.29). The noise-free dynamics in a given time-periodic flow, the so-called
8.7 Superpersistent Transient Chaos: Effect of Noise and Applications 309

von Kármán vortex street around a cylinder, will be described in Sect. 10.7. There
are three attractors [53]: two chaotic ones about the cylinder and a regular one at
x = ∞. We present here the main effect due to noise because of its close relation to
superpersistent chaotic transients.
Because of the explicit time dependence in the flow velocities, the attractors
and their basins move oscillatorily around the cylinder. The remarkable feature is
that in the physical space, there are time intervals during which the attractors come
close to the basin boundaries. Thus, under noise, we expect permanently trapped
motion on any one of the two chaotic attractors to become impossible. In particular,
particles can be trapped near the cylinder, switching intermittently on the two origi-
nally chaotic attractors, but this can last only for a finite amount of time: eventually
all trajectories on these attractors escape and approach the x = ∞ attractor. That is,
chaos becomes transient if one takes into account the effect of noise, or equivalently,
of diffusivity.
To understand the nature of this noise-induced transient chaos, one can distribute
a large number of particles in the original basins of the chaotic attractors and ex-
amine the channel(s) through which they escape to the x = ∞ attractor under noise.
Figure 8.21a–c show, for three dimensionless instants of time (1, 1/4, and 1/2 mod

a 2
y

−2
−2 0 2 4 6
x
b 2
y

−2
−2 0 2 4 6
x
c 2
y

−2
−2 0 2 4 6
x

Fig. 8.21 (a)–(c) At three different instants of time, 1/4 time units apart, locations of the tem-
porally trapped and escaping inertial particles in the von Kármán vortex street in the presence of
noise [197] (copyright 2000, the American Physical Society)
310 8 Transient Chaos in Higher Dimensions

(1), respectively), locations of an ensemble of particles in the physical space. While

there are particles still trapped in the original attractors, many others are already
away from the cylinder. Since this is a two-dimensional projection of the full par-
ticle dynamics, some fractal-like features overlap. The escaping channels through
which particles escape form a set of thin openings surrounding the cylinder and
extending to one of the von Kármán vortices in the flow. After wandering near the
vortex, particles go to the x = ∞ attractor. In the physical space the locations of these
openings vary in time, but the feature that they are narrow is common.
For a fixed noise amplitude, the lifetimes of the particles near the cylinder obey
an extremely slow decaying distribution. A least-squares fit gives [197] the escape
rate as a function of the noise amplitude σ as

1
κ≈ ≈ exp (−3.3σ −0.55). (8.90)
τ
Note that for σ = 0, there is an attracting motion, so that κ vanishes. The way that
κ goes to zero follows the superpersistent transient scaling law as σ is decreased.
Theoretically, the observed noise-induced superpersistent chaotic transient be-
havior can be explained using the approach in Sect. 8.7.1. The result implies that it
may be possible to observe superpersistent chaotic transients in physical space. The
flow system used for experimental study of advective chaotic scattering by Som-
merer and coworkers (Fig. 8.19) is a possible candidate.
Chapter 9
Transient Chaos in Spatially Extended Systems

Chaos is not restricted to systems without any spatial extension: it in fact occurs
commonly in spatially extended dynamical systems that are most typically described
by nonlinear partial differential equations (PDEs). If the patterns generated by such
a system change randomly in time, we speak of spatiotemporal chaos, a kind of tem-
porally chaotic pattern-forming process. If, in addition, the patterns are also spatially
irregular, there is fully developed spatiotemporal chaos. In principle, the phase-
space dimension of a spatially extended dynamical system is infinite. However, in
practice, when a spatial discretization scheme is used to solve the PDE, or when
measurements are made in a physical experiment with finite spatial resolution, the
effective dimension of the phase space is not infinite but still high.
Transient chaos is common in dissipative spatiotemporal systems. The basic
reason is that spatial coupling is typically diffusive. The asymptotic attractors are
therefore often temporally periodic, or even time-independent. It is the approach
toward these attractors that is chaotic. In this sense, spatiotemporal chaos often
collapses after some time, and a regular behavior then takes over. If the lifetime
increases rapidly with the system size, the transients are supertransients. An impor-
tant physical context in which supertransients arise is fluid dynamical turbulence
in pipe flows, where the well-known stationary laminar solution is the only asymp-
totic attractor, and the observed turbulent behavior appears to be a kind of transient
chaos only. Motivated by this example, we shall sometimes call the fully developed
chaotic behavior of other spatiotemporal systems “turbulent.”
In systems exhibiting supertransients, a general picture emerges: In a large
system, it is not possible to determine whether the observed “turbulence” is tran-
sient unless an asymptotic time regime is reached. If the transient time is much
longer than any physically realizable time, the system is effectively “turbulent,” re-
gardless of the nature of the asymptotic attractor. The transients mask in this case the
real attractor, and pose a fundamental difficulty for observing the asymptotic state
of the system. In this sense, attractors are irrelevant to “turbulence.” Supertransients
are thus perhaps the most surprising applications of the concept of transient chaos
to high-dimensional dynamical systems.
In this chapter, we first introduce several paradigmatic models of spatially
extended dynamical systems and discuss supertransients in different models.
Scaling laws with the system size are derived. We then address the effect of noise

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 311
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 9,

c Springer Science+Business Media, LLC 2011
312 9 Transient Chaos in Spatially Extended Systems

and of nonlocal couplings on supertransients, discuss various crisis phenomena in

spatially extended systems, and characterize the fractal properties of supertransients.
Finally, turbulence in pipe flows, which represents an important physical situation
in which supertransients can be observed experimentally, is discussed.

9.1 Basic Characteristics of Spatiotemporal Chaos

9.1.1 Paradigmatic Models

There are several classes of models of spatially extended systems that can be used
to study transient spatiotemporal chaos [775].
Coupled map lattices (CML), introduced by Kaneko [378], provide the simplest
models for spatiotemporal dynamics of continuous variables. In a CML, the local
building blocks of the dynamics are in the form of low-dimensional maps, and they
are coupled to their neighbors according to some rule with a coupling of strength ε .
In this model, both time and space are discrete, but the dynamical variables are
continuous. The dynamics does depend on the boundary conditions. Often periodic
boundary conditions are assumed, but absorbing boundary conditions have also been
used. In one dimension, the typical form of a CML defined on N sites with diffusive
local coupling is

ε $ i+1 %
xin+1 = (1 − ε ) f (xin ) + f (xn ) + f (xi−1
n ) , i = 0, . . . , N − 1, (9.1)
2

where x is the dynamical variable, f (x) is a map describing the local dynamics, and n
and i denote discrete time and space, respectively. For periodic boundary conditions
we have x0n = xN−1
n , while for absorbing boundary conditions we have x0n = xN−1
n =0
for any time instant n. The size of the system is N.
In a cellular automaton (CA) [833] even the dynamical variable is discrete.
By coarsening the x variable of a CML, the dynamics is mapped onto that of a CA
[609]. If, for example, the new variable is chosen to be 0 (1) for x smaller (larger)
than a threshold, a two-state CA is obtained from (9.1).
The Kuramoto–Shivashinsky (KS) equation is a simple PDE exhibiting interest-
ing spatiotemporal dynamics. It was derived to describe propagating patterns in
plasmas, in chemistry and in cellular flames [85]. The KS equation governs the dy-
namics of a continuous scalar field u(x,t) according to a nonlinear equation whose
dimensionless form can be written as

∂u ∂ u ∂ 2u ∂ 4u
= −u − 2 − 4 . (9.2)
∂t ∂x ∂x ∂x
It is remarkable that there are no free parameters in the model, and hence the system
size L serves as the only control parameter. Alternatively, one can fix the size and,
after appropriate rescaling, convert (9.2) to
9.1 Basic Characteristics of Spatiotemporal Chaos 313

∂u ∂ u ∂ 2u ∂ 4u
= −u − 2 − ν 4 , (9.3)
∂t ∂x ∂x ∂x

where ν ∼ L−2 is a dimensionless parameter.

The complex Ginzburg–Landau (cGL) equation governs the spatiotemporal evo-
lution of a complex field ψ (r,t). It is the normal form of spatiotemporal systems in
the vicinity of Hopf bifurcations [85]. In its most commonly used two-dimensional
version, the cGL equation is
∂ψ
= μψ − (1 + iα )Δψ + (1 + iβ )|ψ |2ψ , (9.4)
∂t
where Δ denotes the Laplacian. The system has three dimensionless parameters:
μ , α , and β . When the dimensionless system size is fixed, the parameter μ charac-
terizes the size dependence.
Reaction–diffusion (RD) equations describe the coupled dynamics of chemical
concentrations, or interacting populations. Their general form for two concentra-
tions a(r,t) and b(r,t) is
∂a ∂b
= ra (a, b) + Δa, = rb (a, b) + δ Δb, (9.5)
∂t ∂t
where the functions ra , rb govern the nonlinear reaction equations in the homoge-
neous case, and the dimensionless parameter

δ = Db /Da

is the ratio of the diffusion coefficients.

The Navier–Stokes (NS) equation describes the dynamics of the velocity and the
pressure fields, v(r, t) and p(r,t), respectively, of a viscous fluid. For incompressible
flows not subject to any external force the dimensionless form of the NS equation is
∂v 1
+ v∇v = −∇p + Δv, ∇ · v = 0, (9.6)
∂t Re
where Re is the Reynolds number. Note that there is no dynamical equation for the
pressure. It is the incompressibility equation ∇ · v = 0 that provides a condition of
self-consistency to make the pressure unique.
Figure 9.1 presents several typical complex patterns in spatiotemporal systems.

9.1.2 Phase Spaces of Spatiotemporal Systems

The phase space of a spatiotemporal system is high-dimensional. In a CML, it

is spanned by all the variables xi at different sites i = 0, . . . , N − 1. In a system
described by nonlinear PDEs, the infinite-dimensional phase space is spanned by the
set of all possible spatial distributions of the fields, compatible with a given bound-
ary condition. In the KS and cGL equations, these are the functional spaces u(x)
314 9 Transient Chaos in Spatially Extended Systems

Fig. 9.1 Upper left: space-time diagram of a CML. Black (white) dots correspond to sites in a
laminar (chaotic) regime. Horizontal (vertical) axis represents time (space) [797] (with kind per-
mission from Elsevier Science). Upper right: space-time diagram of a solution of the KS equation
(9.2). The distribution of the field variable u(x) is plotted at an instant of time [715] (copyright
1986, the American Physical Society). Lower left: space-time diagram of an RD problem in one
spatial dimension. The third axis represents the difference between the two concentrations [812]
(with kind permission from the Institute of Physics). Lower right: instantaneous spatial pattern of
an excitable medium in two dimensions. Shading corresponds to the concentration of one sub-
stance. The pattern is similar to that from the cGL equation [741] (copyright 1998, the American
Physical Society)

and ψ (r) respectively. In RD problems and fluid dynamics, two functions define the
phase space: the set of all possible concentrations a(r) and b(r) for the former and
the set of all possible velocity and pressure fields v(r) and p(r) for the latter, where
the forms of the functions are determined by the boundary conditions. For example,
for a fluid system described by the NS equation, all velocity fields vanish on walls
at rest but take on the values of the velocities of moving walls.
9.1 Basic Characteristics of Spatiotemporal Chaos 315

A given spatial distribution of the field variable represents a point of the phase
space. Any of them can be a possible initial condition. The time evolution of the
system corresponds to a motion among different phase-space points, and traces out
a continuous curve emanating from the point representing the initial state. The time
evolution is unique, and the phase-space description is thus complete.
A convenient way of representing an infinite-dimensional phase space is to
expand the field variable(s) in terms of a complete set of orthonormal basis
functions. The expansion coefficients ai , i = 1, . . . , can also be considered phase-
space variables. This expansion can be truncated at some index N if variables ai
with i = N + 1, N + 2, . . . are negligible with respect to global dynamics. Thus,
even systems described by PDEs can be represented as high-dimensional systems
with a finite number of degrees of freedom. In a finite-dimensional phase space,
stationary solutions correspond to fixed points of the phase space. Stable stationary
solutions are thus fixed-point attractors. Homogeneous periodic solutions or waves
correspond to limit cycles. Complicated chaotic solutions can be associated with
chaotic attractors or chaotic saddles. The phenomenon of long transients is naturally
related to situations in which the chaotic set is a saddle, and in addition, this saddle
is rather dense (although not entirely space-filling) in the phase space.

9.1.3 Spatiotemporal Intermittency

There is a large literature on the phenomenon of spatiotemporal intermittency (STI)

[120,121,375,376,378,379,404]. Here we briefly review this phenomenon, in order
to distinguish it from concepts appearing later in this chapter. The concept of STI
applies to spatiotemporal systems in which any space-time point can be classified as
either laminar or turbulent. By “laminar” we mean a regular pattern (whose tempo-
ral dynamics might be both regular and chaotic), while “turbulent” regions have no
apparent regularity either in space or in time. Spatiotemporal intermittency implies
that there are intervals in both space and time in which one of the phases dominates.
Domains of a given type of behavior have well-defined boundaries. An example is
provided by the upper left panel of Fig. 9.1. There were many experiments on STI
(for recent examples, see [269, 478, 659]).
Spatiotemporal intermittency is not the only possible manifestation of spatiotem-
poral chaos, but it is certainly a typical one. It can be considered as state of transition
between laminar and fully turbulent phases. This does not imply that STI must
evolve to be more and more complicated. Spatiotemporal intermittency can very
well provide an asymptotic state, a spatiotemporally chaotic attractor. The usual sta-
tistical measures of STI consider long-time averages of spatial characteristics, such
as the distributions of the size of laminar regions and of the “turbulent” regions
[120, 121]. In a spatiotemporally intermittent state both distributions are exponen-
tially decaying. The decay constants are related to the sizes of the average laminar or
turbulent phases. A difference between the characters of these distributions typically
316 9 Transient Chaos in Spatially Extended Systems

appears at the onset of STI. Here the distribution of the laminar domains follows a
power law, indicating the lack of any characteristic sizes. The onset is, therefore,
similar to a phase transition.
We wish to emphasize that STI can also characterize long transients [376, 619,
831]. In fact, many transients in spatiotemporal systems are of this type. We shall
see that the lifetime can be sufficiently long to make statistical properties stationary
(similar to, e.g., the statistics needed to determine the average Lyapunov exponent
on a chaotic saddle in low-dimensional systems).

9.2 Supertransients

9.2.1 Transient Chaos in Coupled Map Lattices

Perhaps the first indication of complex spatiotemporal patterns appearing as long-

lived transients was found in the thermal convection experiments by Ahlers and
Walden, as early as 1980 [9]. For a detailed investigation of these transients, how-
ever, CMLs have proven to be convenient model systems, initiated by the seminal
paper of Crutchfield and Kaneko [146].
The CML (9.1) has been studied extensively for different types of map f . The ini-
tial conditions are most frequently taken as random numbers at each site. When
the map is strictly contracting, the asymptotic behavior is always spatially regu-
lar and temporally periodic (often homogeneous and steady). The transients toward
this state are, however, typically chaotic. For map f that produces transient chaos
on its own with positive topological entropy but possesses periodic attractors, the
asymptotic behavior can often be spatially regular and temporally periodic. The
CML built on the map f with chaotic attractors generates permanent spatiotempo-
ral chaos if the coupling is weak, but transient chaos leading to a simple attractor
is common for intermediate and strong couplings [831]. These results are summa-
rized in Table 9.1. An observation is that diffusive coupling can generate chaos
even from nonchaotic maps, and it often converts permanent local chaos into global
transients.
A question is how the average transient lifetime τ ≈ 1/κ depends on the sys-
tem size L. For weak coupling ε < ε0 , practically no size-dependence is found.
For slightly stronger coupling, however, the lifetime increases rapidly with the

Table 9.1 Dynamics of the map f and of the corresponding CML

Map f CML
Nonchaotic Transiently chaotic [146, 228, 387, 487, 609, 610]
Transiently chaotic Transiently chaotic [377]
Permanently chaotic Transiently chaotic [465, 619, 831]
Permanently chaotic [619, 831]
9.2 Supertransients 317

Fig. 9.2 Typical space-time pattern of type-I supertransients [377] (with kind permission from
Elsevier Science)

system size. These are the supertransients [146]. There are two distinct types of
supertransients. Type-I supertransients are characterized by a power-law scaling

κ (L) ∼ L−β (9.7)

with a positive exponent β . Type-II supertransients are characterized by an

exponential scaling:
κ (L) ∼ exp (−aLγ ), (9.8)

where γ is a positive exponent and the coefficient a in general depends on the system
parameters.
The patterns associated with the two types of supertransients are qualitatively
different. For type-I supertransients, the basic features are defects whose density
decreases gradually with time, as shown in Fig. 9.2. This can also be considered as a
kind of aging process. Correspondingly, dynamical invariants such as the Lyapunov
exponents and entropies also decrease with time.
Type-II supertransients are, in contrast, statistically steady over a long period of
time, i.e., averages are time-independent in the chaotic state, and the transition to
an attractor is rather abrupt, as exemplified by Fig. 9.3. If the maximum Lyapunov
exponent is positive, a chaotic saddle is expected to exist in the high-dimensional
phase space.

9.2.2 Origin of Supertransient Scaling

The different scaling rules can be traced back to the different patterns that are
characteristic of the two classes of supertransients.
Type-I supertransients: The dominant process is that the defects, as indicated in
Fig. 9.2, undergo a kind of random walk, and when they meet, they annihilate. For an
anomalous random walk, the variance of the displacement scales with time as t 1/β ,
where β is a positive number (β = 2 corresponds to normal diffusion). Estimating
318 9 Transient Chaos in Spatially Extended Systems

Fig. 9.3 (a) Typical space-time pattern of type-II supertransients. From [377] (with kind permis-
sion from Elsevier Science). (b) Time-dependence at a single site illustrating that the crossover to
the nonchaotic behavior is abrupt [465] (copyright 1995, the American Physical Society)

the average lifetime τ as the time needed to reach a displacement variance of the
order of the system size, we obtain τ ∼ Lβ , which is equivalent to (9.7).
Type-II supertransients: Let x∗ (i) denote the coordinate corresponding to the
regular spatiotemporal attractor at site i. A basin size r can be found that is much
smaller than the system size in the following sense: if |x0 (i) − x∗ (i)| < r for all sites,
the system reaches the attractor without chaotic excursions, but if the difference is
larger than r, irregular transients appear. This basin size is a measure of the extension
of the attractor’s basin, restricted to a single dimension. The probability P that a
randomly chosen initial condition at some of the sites falls within the basin size is
proportional to this size: P ∼ r 1.
The following intuitive argument can be used to explain the scaling with the
system size [228, 377, 619, 812]. In a spatially extended system there exists a
correlation length ξ , within which neighboring sites move in a coherent manner.
Conversely, only sites farther apart than ξ move independently. The system can thus
be divided into L/ξ subunits that behave independently. For a random initial con-
dition, the probability Π of falling into all the local basins is P raised to the power
of the number of independent units, i.e., Π ∼ PL/ξ . Time needed to reach the hole
is the phase space of the basin size is proportional to 1/Π , and thus the average
lifetime is estimated as

τ (L) ∼ Π −1 ∼ P−L/ξ ∼ r−L/ξ ∼ eaL , (9.9)

9.2 Supertransients 319

where a = (ln 1/r)/ξ is a positive constant. Here we have assumed that the basin
size is independent of the system size. A strong dependence of r on L can modify
the result. If, for example, r = r(L) ∼ exp (−Lγ −1 ), we have τ (L) ∼ exp(Lγ /ξ )
(cf. (9.8)). Numerical computations often support, however, a linear length-
dependence in the exponent, indicating a weak dependence of the basin size on L.
Finally, we note that the problem of supertransients is effectively the high-
dimensional analogue of chaos in well-stirred chemical reactions in closed con-
tainers. In the absence of any material flux, the final state can be only in thermal
equilibrium governed by a fixed-point attractor. With initial conditions far away
from the thermal equilibrium, one brings the system into a regime whereby long
chaotic transients can arise, as pointed out by Scott, Showalter, and coworkers (see
Fig. 1.15). The novel feature in spatiotemporal systems is that nearly all initial con-
ditions are far away from the attractor, since the probability of falling into the basin
of attraction is extremely small.

9.2.3 Supertransients with Exponentially Long Lifetimes

in Other Systems

It is remarkable that supertransients, mainly of type II, appear in a large number of

systems other than CMLs. Typically, the lifetime scales with the system size as

κ (L) ∼ exp (−aL). (9.10)

Evidence for this behavior has been found in a number of systems, as follows.
Kuramoto–Shivashinsky equation. The investigations of Shraiman [715] and of Hy-
man, Nicolaenko, and Zalesky [348] on phase turbulence in the one-dimensional
KS equation (9.2) provided the first examples of supertransients in a PDE system,
discovered earlier than those in CMLs. The upper right panel of Fig. 9.1 shows a
typical transient pattern.
Complex Ginzburg–Landau equation. After a detailed numerical analysis of long-
lasting spatiotemporal turbulence in the two-dimensional cGL equation by Bohr and
coworkers [85, 86, 344], Braun and Feudel [98] and Houghton and coworkers [340]
provided numerical evidence for an exponential scaling of the average transient life-
time with the system size.
Reaction–diffusion systems. The first example of type-II supertransients in RD sys-
tems of the type (9.5) in one spatial dimension was found by Wacker, Bose, and
Schöll [812]. A typical concentration distribution in the transient phase can be seen
in the lower left panel of Fig. 9.1. A decomposition of patterns during the transients
into eigenmodes indicates that there exist no preferred modes [516–518]. Transient
patterns are thus shown to be uncorrelated, a feature underlying the argument lead-
ing to type-II supertransient scaling. The study of RD systems was extended by
Wackerbauer, Showalter, and coworkers [813–815, 842].
320 9 Transient Chaos in Spatially Extended Systems

Two-dimensional excitable medium. The model investigated by Strain and

Greenside [741] differs slightly in structure from (9.5), but exhibits similar dy-
namical behavior, although with different types of patterns in two dimensions
(lower right panel of Fig. 9.1). This is the first PDE model in which the fractal
properties of a high-dimensional chaotic saddle were investigated (Sect. 9.5).
Complex networks. An observation of Zumdieck, Timme, Geisel, and Wolf [858]
was that in a randomly diluted set of coupled oscillators, the transients toward a
limit-cycle attractor are chaotic and exhibit the scaling (9.10), with the number of
oscillators replacing the length L. The average lifetime of the transients depends
also on the network connectivity, and reaches a maximum at as intermediate level of
dilution. Irregular and exponentially long transients were also observed in different
neural network models [178,428,855,856]. Supertransients with exponentially long
lifetimes were also observed in social networks [51, 52].
Turbulent shear flow. The Theoretical work of Eckhardt and coworkers (see, e.g.,
[229, 231]) based on the Navier–Stokes equation (9.6) predicted the long-lived tran-
sient nature of turbulence in pipes. Recent experiments by Hof et al. [334, 336]
provided evidence for a type-II (or even stronger) supertransient scaling, where the
quantity in (9.10) is replaced by the Reynolds number (for more detail see Sect. 9.6).

9.2.4 Stable Chaos

A peculiar feature of type-II supertransients is that the maximum Lyapunov

exponent is in certain cases negative even during the transients, although the
transient patterns are as irregular as otherwise. This phenomenon has been called
stable chaos [609,610] and provides an example whereby linear stability can coexist
with nonlinear instability in the transient phase. Following the definition of Politi
and coworkers [856], stable chaos means transients that (1) have a negative or zero
maximum Lyapunov exponent and (2) appear stationary for long times, the average
of which scales with the system size exponentially. The phenomenon is robust also
in the sense that it can be present in finite regions of the parameter space [609].
Stable chaos was first found in CMLs for which the local map f is piecewise lin-
ear, is discontinuous at certain points, and possesses a simple periodic attractor. The
map f can be contracting [146,228,857] or can have expanding pieces [62,148,609].
The transients are in any case random, illustrated by an exponential decay in both
the temporal and the spatial correlations. These features are not due to the discon-
tinuity in f , because a continuous variant of the map, in which the discontinuity is
replaced by a steep continuous line, has been shown to exhibit the same behavior
[228,609]. The supertransients as such maps are nonchaotic in the sense of sensitive
dependence on initial conditions, but are chaotic in the sense of positive topological
entropy.
The irregular behavior of stable chaos cannot be related to a local production
of information, due to the lack of a positive Lyapunov exponent. Investigations
[118, 264, 610, 784] led to the conclusion that the irregularity associated with stable
9.3 Effect of Noise and Nonlocal Coupling on Supertransients 321

chaos is produced by transport, i.e., by the nonlinear propagation of finite distur-

bances. The so-called damage spreading analysis [833] can therefore be used to
understand the phenomenon of stable chaos. In particular, one is interested in the
effect produced by finite localized perturbations. Indeed, in systems exhibiting sta-
ble chaos [118, 264, 610, 784], initially perturbed regions in space spread with a
constant front velocity v. Disturbances can thus travel through the system without
damping. It is this velocity that plays in some sense the role of a positive Lyapunov
exponent.
Stable chaos often appears in a certain range of a control parameter, e.g., the
coupling constant ε in (9.1). Outside this range, there are no long transients and the
system rapidly reaches a synchronized, periodic state, in which the front velocity v
is zero. The transition is, however, not a single point in the parameter space. It occurs
in an extended interval [118], where ordered and chaotic dynamics characterized by
v = 0 and v = 0, respectively, alternate in a quite irregular manner.
Stable chaos is not restricted to CMLs. Bonaccini and Politi [91] considered
coupled nonchaotic oscillators in continuous time. The oscillators are subject to a
synchronous periodic forcing over a period T , which is suddenly changed to an
unforced state of length T , and this mechanism is repeated periodically. For suf-
ficiently rare active driving where T /T small, the largest Lyapunov exponent of
the coupled-driven system is negative, and the system exhibits properties of stable
chaos. The diluted neural network model [856] mentioned in Sect. 9.2.3 was shown
to follow the scaling (9.10) in a certain range of parameters, where all Lyapunov
exponents are negative during the transients.
In all the continuous-time examples, the dynamics is associated with the presence
of discontinuities, or with being close to such singularities. In the oscillator model
there is a sudden change in the driving mechanism, and in the neural-network model
the discontinuity is connected with changes in the spike ordering. Thus one can
conclude [172,856] that discontinuities or rapid changes in the dynamical equations
are a necessary condition for the onset of stable chaos. A detailed review of stable
chaos can be found in [611], which also presents an additional realistic system with
stable chaos: a diatomic gas of hard-point particles.

9.3 Effect of Noise and Nonlocal Coupling on Supertransients

Numerical results on the effect of noise on supertransients in spatially extended

dynamical systems were obtained in [433] for a CML system. The diffusive cou-
pling constant ε in (9.1) was replaced by a random variable ε → ε + σ ξn , where ξn
is a random number taken at time instant n, and σ represents the noise intensity. This
choice of noise is homogeneous over the full system, i.e., ξn does not depend on the
site index i. The average lifetime was found to depend little on the noise intensity σ
in the weak-noise regime, suggesting that supertransients are robust [850].
A more recent investigation of Wackerbauer and Kobayashi [814] considered the
effect of spatially inhomogeneous noise as well. They studied an RD system (9.5)
322 9 Transient Chaos in Spatially Extended Systems

in one spatial dimension with periodic boundary condition. The continuous space
dependencies in the concentrations a(x,t) and b(x,t) are approximated by a dis-
crete set [a(i) (t), b(i) (t)] of variables for N  1 sites (i = 1, . . . , N). Correspondingly,
the diffusive coupling term is also discretized. This chain of variables can be con-
sidered to be arranged around a circle. Additive noise σ ξ (i) (t) is included in the
chemical kinetic equation governing the concentration b(i) . The chain is divided into
k blocks such that N/k neighboring sites are subject to the same realization of noise.
The noise terms acting on neighboring blocks are chosen to be independent. Any
value k > 1 corresponds to a spatially inhomogeneous noise – the more inhomoge-
neous, the larger the value of k. The results show that spatially inhomogeneous weak
noise tends to decrease the escape rate of supertransients up to a certain strength at
which a minimum of the escape rate is reached, as shown in Fig. 9.4. The effect in-
tensifies with the degree of the inhomogeneity parameter k. The scenario is similar
to what occurs in a class of low-dimensional systems (Fig. 4.2). Qualitatively, weak
inhomogeneous noise makes the system more random and reduces the chance of
finding the small basin of the attractor. In this model, homogeneous noise (k = 1)
has a destructive effect on the transients: it leads to a monotonic increase in the
escape rate for increasing noise strength. However, type-II supertransient scaling
remains valid in that the lifetime increases exponentially with the size even in the
presence of noise.
In search of a method to control the length of supertransients, an approach is to
investigate the effect of nonlocal coupling in the noise-free problem. Yonker and
Wackerbauer [842] studied the consequence of adding a few nonlocal connections
(shortcuts). At sites coupled not only to the nearest neighbors but to a third, more
distant, site, they modified the discrete Laplacian so that all three sites are included
in a way that ensures the same perturbation, the same as in the locally coupled
model. The length s of the shortcuts is a basic parameter, which is defined as the
minimal number of sites between the two end sites of the shortcut divided by the
number N of sites in the ring. The longest shortcut connecting two opposite sites
along the circle corresponds to length s = 1/2. For a single shortcut of small length,

Fig. 9.4 Dependence of the average lifetime τ on the noise intensity σ in an RD system. The de-
gree of spatial noise inhomogeneity k decreases from k = 20 (stars) to k = 4 (diamonds) to spatially
homogeneous noise k = 1 (squares) [814] (copyright 2007, the American Physical Society)
9.4 Crises in Spatiotemporal Dynamical Systems 323

the average lifetime increases, reaches a maximum at about s = 0.05, then decreases,
finally leading to a reduced lifetime compared to that in the locally coupled system.
The overall dependence is similar to that of the upper curves in Fig. 9.4. In any case,
the type-II supertransient scaling remains valid for any s with a slightly s-dependent
prefactor a(s) in (9.10).
Adding more shortcuts can have a drastic effect on the transients. For example,
two can have the local effect of stabilizing spatiotemporal chaos for arbitrarily long
times, effectively preventing its collapse. Whether this can actually happen depends
on the locations of the shortcuts and the initial conditions. For example, in a large
ensemble of cases with randomly chosen shortcut locations, the probability for spa-
tiotemporal chaos to be permanent is about 70%. Three shortcuts can increase the
likelihood of permanent chaos even more. A further increase in the number of short-
cuts, however, seems to weaken the effect, and the likelihood of transient chaos
increases again.
Control of spatiotemporal transients via nonlinear feedback was suggested in
[619], where it was demonstrated for a CML system that proper control can shorten
the lifetime of the transients by several orders of magnitude. These developments
illustrate that adding weak noise, or taking over methods from the physics of net-
works, has the potential to provide some effective ways to harness transient chaos
in spatially extended systems.

9.4 Crises in Spatiotemporal Dynamical Systems

9.4.1 Boundary Crises: Supertransients Preceding Asymptotic

Spatiotemporal Chaos

When there is an asymptotic spatiotemporal chaotic attractor, or asymptotic

“turbulence,” long chaotic transients typically occur in a parameter range pre-
ceding the permanently chaotic regime. Suppose the latter is in the parameter range
p > p1 . Chaotic transients are then present for p < p1 . Intuitively, their average
length should increase on approaching the critical value p1 . For supertransients,
one expects a power-law divergence in the exponent of the average lifetime, i.e., a
decay of the escape rate as

κ (p, L) ∼ exp[−c(L)(p1 − p)−δ ], (9.11)

where δ > 0 and the coefficient c > 0 depends on the system size L. Combining this
with the size dependence of (9.8) or (9.10), we see that the coefficient a changes
with the parameter p as
a(p) ∼ (p1 − p)−δ . (9.12)
A detailed investigation of the two-dimensional cGL equation (9.4) led to the
conclusion [85, 86, 344] that permanent spatiotemporal chaos is present in a region
of the parameter plane (α , β ) (μ fixed). When approaching the boundary of this
324 9 Transient Chaos in Spatially Extended Systems

Fig. 9.5 For the cGL

equation (9.4), the largest
average Lyapunov exponent
λ1 of the chaotic sets as a
function of the parameter α .
Spatiotemporal chaos is
permanent for α > α1 .
For αc < α < α1 , only
chaotic transients are present
[86] (copyright 1990, the
American Physical Society)

region from outside, the scaling relation (9.11) was found with exponent δ = 2,
which is similar to that for low-dimensional supertransient systems (cf. (8.60)), but
here the exponent δ is larger than unity, and a size-dependence is also present.
When the largest average Lyapunov exponent λ1 of the attractor is plotted as a
function of the parameter p ≡ α , it is positive in the range α > α1 . This curve can be
merged smoothly with the curve of the Lyapunov exponent for the transient regime,
as shown in Fig. 9.5, illustrating that the spatiotemporal chaotic saddle is converted
at α1 into a chaotic attractor. The critical parameter value α1 can thus be viewed
as a point of crisis in the cGL system. The Lyapunov exponent vanishes at some
α < αc , so the transients are not chaotic for α < αc . For α slightly larger than αc ,
the exponent scales as λ f (α ) ∼ (α − αc )1/2 [85, 86].

9.4.2 Interior Crises in Spatially Coherent Chaotic Systems

An investigation of different types of crisis phenomena in the KS equation was

carried out by Chian, Rempel, and coworkers [126, 631, 632, 634]. These authors
used the form (9.3) of the equation in which the length is fixed but the parame-
ter ν contains the size of the original system. A parameter range was chosen for
which the dynamics is chaotic in time but remains coherent in space. A Fourier
decomposition of (9.3) with N = 16 modes appeared to be sufficient to illustrate
the crisis phenomenon. In particular, in a parameter range of ν , a periodic window
was found, as seen by plotting the long-time values of the sixth Fourier component
a6 as a function of ν . The window is bounded by an interior crisis and a saddle-
node bifurcation at its two ends (Fig. 9.6a). Inside the window the attractor is a
period-3 orbit or is localized in three narrow bands. In both cases it is surrounded
by an extended chaotic saddle. The concept of basic components, introduced in
Sect. 3.3 to characterize low-dimensional crises, thus becomes applicable to high-
dimensional problems. This surrounding chaotic saddle (SCS) was determined by
the PIM-triple method (Sect. 1.2.2.4), and its projection on the a6 variable is also
9.4 Crises in Spatiotemporal Dynamical Systems 325

Fig. 9.6 Interior spatiotemporal crisis. (a) Bifurcation diagram of mode amplitude a6 as a func-
tion of parameter ν in the KS equation (9.3). Gray dots indicate points on the surrounding chaotic
saddle (SCS). IC and SNB denote interior crisis and saddle-node bifurcation, respectively. (b) A
three-dimensional projection of the SCS for ν = 0.029925 [631] (with kind permission from
Elsevier Science)

Fig. 9.7 Part of the phase

space projected on the
(a5 , a6 )-plane for
ν = 0.0299211, before
interior crisis (IC). CA:
chaotic attractor, SM: stable
manifold of the mediating
period orbit denoted by a
cross. Gray dots mark the
stable manifold of the SCS
[631] (with kind permission
from Elsevier Science)

shown in the bifurcation diagram. In the full phase space the chaotic saddle turns out
to be extended but low-dimensional, as can be seen in a three-dimensional projection
close to the saddle-node bifurcation, where no chaotic attractor exists. The saddle is
practically a single line segment, but gaps are visible along this line (Fig. 9.6b).
In the middle of the window the attractor undergoes a period-doubling bifur-
cation, after which a small-size chaotic attractor (CA), the three-band attractor,
appears. The surrounding chaotic saddle, SCS, coexists now with the chaotic attrac-
tor. In a projection onto the plane of two Fourier components, the stable manifold
of the mediating periodic orbit separating the attractor from the saddle can be seen,
as shown in Fig. 9.7. The saddle’s stable manifold appears to be dense.
At the crisis, the small-size chaotic attractor collides with the mediating orbit
or with its stable manifold, and thus with the chaotic saddle as well (Fig. 9.8a).
326 9 Transient Chaos in Spatially Extended Systems

Fig. 9.8 Phase-space projection on the (a5 , a6 )-plane at IC, ν = 0.02992021 (a), and slightly
beyond crisis (b) [631] (with kind permission from Elsevier Science)

Fig. 9.9 (a) Bifurcation diagram as in Fig. 9.6 containing only the band chaotic attractor, which is
converted into a band chaotic saddle (BCS, plotted in gray) beyond the interior crisis. (b) Chaotic
saddles forming the backbone of the extended chaotic attractor projected on the (a5 , a6 )-plane at
the postcrisis parameter value ν = 0.02992006 [631] (with kind permission from Elsevier Science)

The large gaps present along the surrounding saddle just before the crisis become
filled up by the newly generated orbits (see Sect. 3.4), and the extended chaotic
attractor to appear contains the previous attractor, the saddle, and the filled-up gaps
(Fig. 9.8b). After the crisis, points of the extended attractor that remain forever on
the three bands occupied by the small attractor in the precrisis regime are connected
to a saddle situated in this region, the band chaotic saddle (BCS). This saddle can
be represented both on the bifurcation diagram (Fig. 9.9a) and on a projection of
the plane of two variables (Fig. 9.9b). Similarly, points never leaving the region of
the former surrounding chaotic saddle form a postcrisis chaotic saddle (SCS) that
can be considered the continuation of the precrisis SCS. These two saddles are the
main building blocks of the extended chaotic attractor arising from the interior crisis.
The situation is thus similar to that for low-dimensional maps (Chap. 3).
9.4 Crises in Spatiotemporal Dynamical Systems 327

9.4.3 Crises Leading to Fully Developed Spatiotemporal Chaos

In another series of papers, Rempel, Chian, and coworkers [633, 638] aimed to
understand crises underlying spatiotemporal inhomogeneities [320]. For this pur-
pose they used a one-dimensional PDE model of regularized long waves for a field
φ (x,t) driven sinusoidally both in space and time. With all other parameters fixed,
the main control parameter is the driving amplitude f . A Fourier decomposition of
φ (x,t) into N = 32 spatial modes was used. As f is changed, the dynamics exhibit
three qualitatively different types of behavior. For the lowest value of f the pattern
is regular in space and quasiperiodic in time (Fig. 9.10a). For higher values of f ,
spatial regularity remains but the pattern becomes temporally chaotic (Fig. 9.10b),
as indicated by the appearance of a positive Lyapunov exponent. The correspond-
ing attractor is called a temporally chaotic attractor (TCA). A further increase in f
leads to the appearance of fully developed spatiotemporal chaos (Fig. 9.10c). This
occurs suddenly and is accompanied by an increase in the maximum Lyapunov
exponent to a much larger value. The new attractor is a spatiotemporally chaotic
attractor (STCA) that possesses a larger dimension value than the previous one
(the TCA).
To follow these changes in the phase space, Rempel and Chian projected the
invariant sets on the plane defined by the real parts of the second and the third
Fourier modes, after taking an appropriate Poincaré map. The quasiperiodic torus
attractor appears to be associated with a few closed curves (Fig. 9.11a). The authors

Fig. 9.10 Spatiotemporal patterns of the field φ for different values of the driving amplitude f :
(a) spatially regular, temporally quasiperiodic; (b) spatially regular, temporally chaotic; and
(c) spatially irregular, temporally chaotic [633] (copyright 2007, the American Physical Society)
328 9 Transient Chaos in Spatially Extended Systems

Fig. 9.11 Phase-space projection of various invariant sets on the plane of two modes. (a) A
quasiperiodic attractor (QPA, black) and a spatiotemporally chaotic saddle (STCS, gray dots),
(b) Temporally chaotic attractor (TCA), (c) a spatiotemporally chaotic attractor (STCA) after cri-
sis, which occupies the regions where the former STCS and TCA reside, and (d) decomposition of
STCA into a postcrisis STCS and a temporally chaotic saddle TCS after the crisis [633] (copyright
2007, the American Physical Society)

pointed out that already here an extended chaotic saddle exists surrounding the at-
tractor. The corresponding transients carry irregular spatiotemporal patterns, and
therefore the saddle is called the spatiotemporally chaotic saddle (STCS). When
the spatially regular dynamics becomes chaotic, the torus attractor breaks, but the
new temporally chaotic attractor (TCA) remains localized around the former torus
(Fig. 9.11b). The TCA is area-filling in the projection, but is of small size. The
surrounding saddle, STCS, does not change appreciably. When permanent spa-
tiotemporal chaos occurs, the chaotic attractor suddenly broadens and becomes
a spatiotemporally chaotic attractor (STCA); Fig. 9.11c. It is remarkable that the
extension of the STCA is practically the same as that of the spatiotemporal sad-
dle (STCS) earlier. At this crisis the temporally chaotic attractor collides with the
surrounding saddle, and the latter becomes embedded in the new attractor. In this
postcrisis regime, Rempel and Chian were also able to identify a chaotic saddle in
the region occupied by the temporal attractor earlier. This saddle is called the tem-
porally chaotic saddle (TCS); Fig. 9.11d. In the projection, it fills a slightly smaller
area than the TCA.
If a trajectory on the extended attractor comes to the vicinity of the TCS, a regular
pattern appears in the space, which changes chaotically in time. After some time,
the trajectory deviates from this saddle, and comes close to the chaotic saddle that
exists outside the TCS, a postcrisis STCS that governs the spatiotemporally chaotic
9.5 Fractal Properties of Supertransients 329

dynamics. After escaping from the STCS, the trajectory returns to the vicinity of the
TCS and the pattern becomes regular again, etc. The average lifetime of the spatially
regular phases can, in principle, be estimated as the average lifetime on the TCS. The
full process is intermittent [131], and the situation is the high-dimensional analogue
of crisis-induced intermittency discussed in Sect. 3.3. A similar phenomenon was
observed in the damped KS equation [635].
The spatiotemporal intermittency (STI) mentioned in Sect. 9.1.3 is not the kind
of intermittency treated here, since regular and irregular phases extend in STI over
finite regions of the real space only. It may be useful to study spatiotemporal in-
termittency in terms of the underlying chaotic saddles. At present, little is known
about the scaling with system size of the lifetimes associated with the intermittent
dynamics discussed here.

9.5 Fractal Properties of Supertransients

9.5.1 Dimensions

Supertransients have specific fractal properties, as emphasized in [462–465]. It has

been observed that chaotic saddles underlying long transients typically have a stable
manifold whose dimension is close to that of the phase space. The basin of attraction
of the regular asymptotic attractor can be determined on a plane of initial conditions
of just a few variables (Fig. 9.12a). In a long observation time only a few points con-
verge to the attractor; the others remain away from it. These points represent initial
conditions that stay close to the chaotic saddle’s stable manifold. Alternatively, in a

Fig. 9.12 (a) Stable manifold of a chaotic saddle (black dots) in the plane of two variables of
a CML described by (9.1) [465] (copyright 1995, the American Physical Society). (b) Transient
lifetime as a function of the initial condition taken from a one-dimensional line in the phase space of
the cGL equation in its supertransient state [98] (copyright 1996, the American Physical Society).
The dimension of the set of points with long lifetimes is denoted by ds
330 9 Transient Chaos in Spatially Extended Systems

plot of the lifetime function, the lifetimes needed to reach the attractor as a function
of a single initial coordinate also appear to be dense (Fig. 9.12b). It is useful to in-
troduce, as in scattering problems (Sect. 6.2), the box-counting dimension ds as the
dimension of the set of points where the lifetime is formally infinite along such a
segment. Since infinite lifetime values belong to the stable manifold of the saddle,
this ds is the dimension of the intersection of a line with the stable manifold of the
saddle.
As a quantitative measure of the fractality, the uncertainty exponent (Sect. 5.3)
can be determined. The numerical value of α was found to be as small as 10−3
(cf. Fig. 9.13), indicating that ds is quite close to unity. It was also shown [462–464]
that the largest Lyapunov exponent computed at fixed finite time is extremely sen-
sitive to small changes in the parameters. Supertransients are thus characterized by
riddled structures in the parameter space.
A simple formula for the partial dimension ds was conjectured in [465].
In particular, escape occurs mostly along the direction of the largest positive
Lyapunov exponent λmax . The system is therefore expected to behave effectively
as a two-dimensional system with positive Lyapunov exponent λmax . Utilizing the
Kantz–Grassberger relation (2.76), one obtains the following information dimension
ds,1 of the set of singularities:

κ (L)
ds,1 (L) = 1 − . (9.13)
λmax

Taking into account that the dimension of a set resulting from the intersection of
two sets follows from the rule according to which the codimensions are additive,
(8.52), one finds for an N-dimensional map that the information dimension Ds,1 of
the stable manifold is given by

κ (L)
Ds,1 (L) = N + ds − 1 = N − . (9.14)
λmax

Fig. 9.13 Plot of the fraction

of uncertain initial conditions
f (ε ) versus uncertainty ε .
With the value of the
uncertainty exponent α , the
box-counting dimension ds is
ds = 0.9985, a quantity that is
hardly distinguishable from
unity [465] (copyright 1995,
the American Physical
Society)
9.5 Fractal Properties of Supertransients 331

We see that since κ is small, the dimension of the stable manifold is close, for type-II
supertransients exponentially close, to the dimension of the phase space.
Equation (9.14) in fact follows directly from the general dimension formulas
in Sect. 8.2.2. In particular, consider an (N  1)-dimensional map with a small
escape rate. When κ is nearly zero, the only possibility for the left-hand side of
(8.23) to be larger than K1 is that all the positive Lyapunov exponents appear on the
right-hand side, i.e., J + 1 = U. The numerator in the ratio in (8.24) then contains
λU+ − κ = λmax − κ . Since S + J = S + U − 1 = N − 1, we recover (9.14).
The dimension formula (8.21) for the unstable manifold can also be applied.
For spatiotemporal systems with small escape rate, observe first that the condition
(8.20) requires that the sum of all Lyapunov exponents (with signs taken into ac-
count) up to index I be greater than κ , but up to index I + 1 be smaller than κ . For
near-zero values of κ , the sums should practically be positive and negative, respec-
tively. This is the condition in the Kaplan–Yorke formula (see (8.21) with κ = 0)

λ1+ + · · · + λU+ − (λ1− + · · · + λI− )

D1 = U + I + − (9.15)
λI+1
for chaotic attractors. One can then imagine a chaotic attractor with the same
Lyapunov exponent spectrum as the saddle, and denote its information dimension
by Dattr,1 . Given a discrete set of Lyapunov exponents, a small κ does not change
the value of I, and we can write

κ (L)
Du,1 (L) = Dattr,1 − − . (9.16)
λI+1

For the saddle’s dimension we then obtain, from (9.14),

 
1 1 κ (L)
D1 (L) = Dattr,1 − κ (L) + − = Du,1 (L) − . (9.17)
λmax λI+1 λmax

Equations (9.14), (9.16), and (9.17) illustrate that a supertransient chaotic saddle
is a quasiattractor in the sense that its dimension is close to that of an attractor
(with almost identical Lyapunov spectrum), its stable manifold is nearly space-
filling (close to forming a basin of attraction), and its unstable manifold has nearly
the same dimension as the chaotic saddle (for an attractor, Du,1 and D1 coincide).
These observations indicate that the dimension of supertransient chaotic saddles can
be approximated by the Kaplan–Yorke formula, and explain why statistical averages
are so well defined on supertransient chaotic saddles. The validity of relations (9.14)
and (9.16) was recently illustrated for various high-dimensional reaction–diffusion
systems [734].
It is worth mentioning that although the stable manifold is nearly space-filling,
the unstable manifold’s dimension can take on any value. It is the number U of posi-
tive Lyapunov exponents and the index I that essentially determine the value of Du,1 .
In principle, it can assume a small value even in a high-dimensional phase space.
332 9 Transient Chaos in Spatially Extended Systems

9.5.2 Dimension Densities

In high-dimensional systems it is useful to define dimension densities [85,487,831],

i.e., quantities expressing the dimension falling on a single degree of freedom.
For supertransients, the dimension density δs ≡ Ds /N of the stable manifold is close
to unity. Little is known, however, about the dimension densities δu = Du,1 /N and
δ = D1 /N of the unstable manifold and of the saddle, respectively. The question
so far has been addressed in a few cases only. In particular, for a model of ex-
citable media, Strain and Greenside [741], and for different RD systems, Stahlke
and Wackerbauer [734], found the dimension density of the chaotic saddle to be of
order a few percent. This is also consistent with the observation [516] that the num-
ber of positive Lyapunov exponents is small even in large systems (although the
number increases with the system size).
An important dynamical property in high-dimensional systems is the existence
of a Lyapunov density [85]. It implies that the set of the Lyapunov exponents λ j± ,
defined by (8.5), as a function of x ≡ j/N converges for N → ∞ to a well-defined
function Λ ± (x), as exemplified by Fig. 9.14. In such a case the number U of positive
(or negative) Lyapunov exponents scales with the dimension of the phase space, and
U/N converges to a constant. As a result, the metric entropy (8.9) can be written as
 U/N
K= Λ + (x)dx − κ . (9.18)
0

Similarly, the sums defining the indices J and I, (8.23) and (8.20), more precisely the
ratios J/N and I/N, can also be expressed as integrals, which depend on the value
of the escape rate. For small escape rates, however, the dependencies are weak, and

Fig. 9.14 Spectrum of Lyapunov exponents λ j associated with chaotic transients for a CML of
size N = 50 (crosses) and N = 100 (triangles), as a function of x = j/N. The convergence to a
limiting Lyapunov density can be seen [487] (with kind permission from Elsevier Science)
9.6 Turbulence in Pipe Flows 333

we have J/N = (U − 1)/N → U/N. Since the fraction appearing in the general
expression of Ds,1 and Du,1 is always less than one, it does not contribute to the di-
mension density. Supertransients are thus characterized by the following dimension
densities:
S +U U +I
δs = = 1, δ = δu = . (9.19)
N N
It is worth defining the signed Lyapunov density Λ (x), as shown in Fig. 9.14. Since,
as stipulated by (8.5), Λ (x) = Λ + (U/N − x) for 0 ≤ x ≤ U/N and Λ (x) = −Λ − (x −
U/N) for 1 ≥ x ≥ U/N, the nontrivial dimension density δu = δ also satisfies the
equation
 δu
Λ (x) dx = 0. (9.20)
0

When considering the integral of the signed Lyapunov density between zero and
some value x, the dimension density is the x value for which the integral vanishes.
In fact, (9.20) is valid for spatiotemporal chaotic attractors as well [262]. We con-
clude that the picture based on the Lyapunov and dimension densities suppresses
the role of the finite lifetime of chaos, and emphasizes the quasiattractor character
of supertransients.
The problem of stable chaos (Sect. 9.2.4) deserves special attention. Although
these systems appear to exhibit fractal features, dimension formulas (8.21) and
(8.24) are not applicable. In fact, these relations are valid for generic chaotic saddles,
but that is not the case here. There is a possibility for strange nonchaotic saddles to
arise in analogy with strange nonchaotic attractors [241]. (For strange nonchaotic
repellers of one-dimensional maps, see Sect. 2.4.) A strange nonchaotic spatiotem-
poral saddle might have a box-counting dimension that does not increase linearly
with the system size, i.e., with a density δu = δ = 0.

9.6 Turbulence in Pipe Flows

9.6.1 Turbulence Lifetime

The transition to turbulence in pipe flows has long been a fascinating problem in
fluid dynamics (for reviews, see [216, 217, 299]). Investigations of the phenomenon
began in the second part of the nineteenth century with the milestone experiments of
Reynolds in 1883. He pointed out that in a pipe of fixed length the flow changes from
smooth (laminar) to irregular (turbulent) at sufficiently large flow velocities. A good
dimensionless measure of the flow velocity is the Reynolds number Re = UD/ν ,
with U and D chosen as the mean flow speed across the pipe and the diameter,
respectively. When the flow velocity slowly increases in a given setting, the transi-
tion from laminar flow to turbulence occurs abruptly at a critical Reynolds number
Rec of order 2,000. Early experiments indicated, however, that under controlled con-
ditions the laminar flow can be maintained up to Reynolds numbers much larger
334 9 Transient Chaos in Spatially Extended Systems

than 2,000. It was recognized later that the roughness of the wall’s surface plays an
important role: the rougher the wall, the smaller the critical Reynolds number. More
recent investigations have led to the observation that perturbations to the laminar
flow such as those caused by surface roughness are needed to trigger turbulence, and
the critical Reynolds number Rec depends on the type and the strength of the pertur-
bation. Thus the onset of turbulence is determined not only by the Reynolds number
but also by the perturbation. To trigger turbulence, the flow has to be sufficiently
fast and the perturbation has to be sufficiently strong. The required perturbation is,
however, smaller for larger values of the Reynolds number. Therefore, in any ex-
perimental setting in which small perturbations cannot be avoided, turbulence will
always appear at sufficiently large values of the Reynolds number.
The steady laminar solution, such as the parabola profile in a pipe of circular
cross section, is linearly stable for all Reynolds numbers [299]. In dynamical-system
terms, this implies the existence of a fixed-point attractor in the infinite-dimensional
phase space, with a relatively small basin of attraction. In addition, there is no ev-
idence for the existence of any stable state with simple spatial or temporal pattern,
e.g., traveling waves, which would be the analogues of limit cycle attractors. The
turbulent state can be considered a high-dimensional chaotic state associated with
either a chaotic attractor or a chaotic saddle.
The first indication of the transient character of pipe turbulence appeared about
20 years ago [99,299], based on investigations of the stability of the laminar profile.
There has been increasing experimental evidence since then indicating that even
if the turbulent state is established for not too large Reynolds numbers, this state
can suddenly decay, without any apparent precursor, toward the laminar state. This
implies that the chaotic sets for not too large values of the Reynolds number are
nonattracting. Research has then been concentrated on the average lifetime τ of the
chaotic saddle. The classical experiments suggest that the lifetime is rather large,
for otherwise, the turbulence would not have appeared to be permanent to earlier
investigators. The use of long pipes and efficient numerical methods have made
more detailed investigations possible. Figure 1.23 shows the experimental findings
of Peixinho and Mullin on the exponential decay in time. The value of the escape
rate appears to be independent of the details of the initial perturbation, but depends
on the Reynolds number only. The exponential decay sets in only after some time t0
in any experimental run.
When plotting the actual lifetime as a function of the perturbation amplitude A
to trigger the turbulence in numerical simulations, a more detailed picture can be
obtained. A slight change in the amplitude can lead to drastically different lifetimes
if the amplitude is above a threshold (Fig. 9.15). The irregular part of the lifetime
distribution is fractal. Furthermore, the average Lyapunov exponent during the tur-
bulent phase was shown to be strictly positive [231]. These features indicate that
the high-dimensional saddle underlying the turbulence has all the characteristics of
low-dimensional chaotic saddles and of transient chaos in many other spatiotempo-
ral systems.
A basic question is the dependence of the turbulent escape rate on the Reynolds
number. In the class of functions exhibiting a rapid decrease with the Reynolds
9.6 Turbulence in Pipe Flows 335

Fig. 9.15 Turbulence lifetime versus perturbation amplitude A in a pipe at Reynolds number
Re = 2,000 obtained numerically. The bottom panel is a magnification of the box indicated in
the top panel. At the edge of chaos, at the value marked by two vertical bars, the function turns
from smooth to fractal-like, indicating that chaotic transients are triggered by sufficiently large
amplitudes [231] (with kind permission from Cambridge University Press)

number, a choice is some functions that approach zero for a finite value of Re. This
form can retain one aspect of the original picture, namely that beyond a threshold
Reynolds number, permanent turbulence can be present. The laminar fixed-point
attractor would then coexist with the chaotic attractor of the turbulence. Another
choice is some monotonically decreasing functions of Re with nonzero values for
any Re. Using a pipe of length 30 m, experiments by Hof, Westerweel, Schneider,
and Eckhardt [336] provided a firm answer to the question. In a set of experiments
covering more than two decades of lifetimes, they showed that the escape rate is
nonzero up to large values of Re. This suggests that turbulence remains a transient,
a feature also observed in superfluid turbulence [693] and magneto hydrodynamical
turbulence [639]. By measuring time in units of D/U, the dimensionless escape rate
was found in [160, 217, 336] to scale with Re as

κ (Re) = ae−bRe (9.21)

with parameter b between 0.03 and 0.04 (see Fig. 9.16). According to this rule, any
increase in the Reynolds number by 100 implies a multiplication of the escape rate
by a factor of 1/33. Thus, pipe turbulence is a kind of type-II supertransient (with
system size replaced by the Reynolds number).
336 9 Transient Chaos in Spatially Extended Systems

Fig. 9.16 Escape rate as a function of the Reynolds number in the experiment by Hof et al. The
straight-line fit corresponds to formula (9.21). The inset shows the same data on a linear scale to
illustrate that the escape rate is asymptotic to zero rather than crossing the horizontal axis at a finite
value of Re [217] (with kind permission from Annual Reviews)

A recent experimental study by Hof, de Lozar, Kuik, and Westerweel [334]

extended the range of turbulent lifetimes by six orders of magnitude by collecting
data from four pipe setups with significantly reduced statistical errors. The extended
data set (which contains, as a subset, the points of Fig. 9.16) suggests a superexpo-
nential scaling with the Reynolds number in the form of

κ (Re) = exp[− exp (c1 Re + c2 )] (9.22)

with c1 = 0.0057, c2 = −8.7. This fit is valid in the Reynolds-number range

(1670, 2040), where the dimensionless κ changes between 0.2 and 10−8 . Note that
a further extension of these results is hardly possible, since due to the rapid increase
in the lifetime, the measurement at Re = 2, 100 would already require an estimated
time of 46 years [334, 663].
Due to extended and improved numerical methods, the range of Reynolds
numbers and sample sizes of experiments have become accessible in computer
simulations. The results of [33] were in close quantitative agreement with the
form of (9.22). Similar superexponential scaling of lifetimes was reported in a
Taylor–Couette flow [95].
A possible theoretical explanation of the superexponential scaling was provided
by Goldenfeld and coworkers [267]. The authors argued that the determining factor
for the suppression of a puff is the probability that the largest velocity fluctua-
tions fall below some threshold value. These large-amplitude events follow extremal
statistics. From general results on such statistics the superexponential form of (9.22)
can be derived.
9.6 Turbulence in Pipe Flows 337

9.6.2 Other Aspects of Hydrodynamical Supertransients

An interesting feature of the lifetime distribution versus perturbation amplitude, as

shown in Fig. 9.15, is that slowly varying regions are interwoven with intervals of
rapid change. In the smooth regions the transients are short and nonchaotic. The
transition (indicated by two bars in Fig. 9.15) between the extended smooth region
at small amplitudes and the region with fractal fluctuations is rather abrupt. This
point on the border between laminar and chaotic regions is called the edge of chaos
[691, 720], which separates initial conditions that decay directly to the laminar at-
tractor and those that come close to the chaotic saddle first, i.e., exhibit turbulence.
Trajectories starting from the edge of chaos move in a region intermediate between
laminar and turbulent dynamics. The results of [409,520,691,720,803] suggest that
the edge of chaos lies, for any Reynolds number, in the stable manifold of an invari-
ant object, the edge state that resides in the phase space between the fixed point and
the chaotic saddle. The stable manifold is thus a kind of basin boundary between
the laminar and the turbulent dynamics. The latter, of course, cannot have a real
basin of attraction, but only one that appears to be so in finite-time observations (the
quasiattractor character discussed in Sect. 9.5). The dynamics restricted to the edge
of chaos converges to a chaotic state, the edge state, and numerical simulations in
pipe flows indicated that it corresponds indeed to an irregular wavy motion along the
pipe, which is, however, less energetic than the turbulent dynamics itself [691]. This
attractor is only a relative attractor, since it is unstable with respect to perturbations
perpendicular to the edge of chaos.
Low-dimensional chaotic saddles contain an infinite number of unstable periodic
orbits (Sect. 2.6.4). In an analogous way, the chaotic saddle underlying pipe turbu-
lence is expected to contain coherent structures. Both in theory and experiments,
they were shown to be regular traveling waves [229, 335], all unstable, correspond-
ing to hyperbolic states in the high-dimensional phase space. Currently, there is an
intensive search underway for such coherent structures (spatiotemporal patterns)
[213, 263, 303, 690, 800] about which chaos is organized. Over a long-time obser-
vation of turbulence one expects to see different coherent states in different time
intervals. This kind of approach may eventually lead to a periodic-orbit expansion
[153] of the chaotic saddle, in full analogy with low-dimensional problems (see
Appendix A). There is then hope that the statistical properties of the turbulent flow
can be expressed in terms of the properties of the coherent structures.
Finally, we mention that there are other hydrodynamical situations in which the
onset of turbulence is similar to that in pipe flows. Notable examples are plane
Poiseuille flows (pressure-driven flows between two large parallel plates) [816]
and Couette flows (driven by a moving wall) [215, 686, 692]. The common fea-
ture in these shear flows is that the laminar profile is stable.1 One expects therefore
in these situations that turbulence is not permanent and decays eventually toward

1 Plane Poiseuille flows are linearly unstable, but the critical Reynolds number is much above the

value at which turbulence transition occurs [217].

338 9 Transient Chaos in Spatially Extended Systems

a laminar profile. Shear-flow turbulence is thus a case of its own, and is present in
the form of high-dimensional chaotic transients. We are thus currently witnessing
the appearance of concepts of transient chaos in the study of classical turbulence.

9.7 Closing Remarks

In this chapter we have focused on supertransients. Although they are quite

common, there are cases in which the distribution of transient lifetimes is not
exponential, or if it is, the average lifetime does not grow rapidly with the system
size. It may, e.g., become saturated (for examples, see [252] and [831]). The type
of coupling plays an important role [797]. It is, nevertheless, an open question at
present whether one can decide from first principles if a system exhibits super-
transients. In fact, the question of how to decide whether a spatiotemporal system
possesses a chaotic attractor has not been answered. A systematic application of
nonlinear stability analysis to different possible asymptotic patterns [517] may pro-
vide insights. A recent investigation [813] showed that the master-stability function
[583], a central tool in the theory of synchronization in dynamical systems, can suc-
cessfully be applied as an indicator for transient versus permanent spatiotemporal
chaos.
A somewhat analogous phenomenon to supertransients was found in Hamiltonian
systems with many degrees of freedom. Any isolated macroscopic system should
eventually relax to a state of thermal equilibrium in which any macroscopic variable
is independent of time. Nevertheless, in systems with global (mean field) coupling,
long relaxations were found whose average time diverges with the number of com-
ponents [14, 21, 607, 839]. More recently, a metastable state was discovered [531],
as characterized by periodic or quasiperiodic oscillations of macroscopic variables
about mean values that are different from the respective equilibrium values. The
lifetime of the metastable state was found to increase linearly with the number of
degrees of freedom. The underlying microscopic dynamics is chaotic, but must have
different characters in the metastable and equilibrium states. Both examples can be
considered as type-I supertransients, which last long in the thermodynamic limit.
In some spatiotemporal problems the linear size may not be freely chosen. Long
transients may, nevertheless, be present (see, e.g., [188, 841]), but it is not apparent
whether they scale at all with some parameter of the problem. It is useful to find a
scaling parameter in such cases and check whether the dependence is power-law or
exponential. In the case of time-delayed systems [841], a natural candidate for some
scaling parameter is the delay time.
It is worth pointing out a difference between the shear-turbulence problem and
the supertransient phenomena in spatially extended systems other than pipe flows.
The scaling in turbulence is not with respect to the length of the pipe, but rather with
the diameter D in the Reynolds number. It would be interesting to understand this
difference better.
9.7 Closing Remarks 339

The concept of unstable coherent structures as building blocks for a periodic-orbit

type of expansion of the chaotic saddle, or those of the edge of chaos and
the invariant sets associated with it, can be applied to all systems exhibiting
supertransients. It is quite remarkable that problems ranging from fluid dynamics
and chemistry to population dynamics and biology with quite different underly-
ing mathematical structures all share similar features, dominated by long-lasting
chaotic transients. A unified understanding of the physics underlying this phe-
nomenon deserves further efforts.
 Part IV
Applications of Transient Chaos
Chapter 10
Chaotic Advection in Fluid Flows

The advection of tracer particles in hydrodynamical flows represents one of the

successful applications of chaos theory. The basic observation is that molecular dif-
fusion is negligible on the typical time scale of the flow.1 As a result, in the absence
of any diffusion-enhancing mechanism such as hydrodynamical turbulence, advec-
tion dominates. Indeed, the main physical mechanism for fluid stirring is advection,
whose efficiency can be enhanced greatly by chaotic dynamics. The spreading of
pollutants on large scales is also dominated by advection. Potential applications of
chaotic advection range from laboratory investigations of fluid dynamics to the study
of large-scale environmental flows. From the point of view of dynamical systems, an
appealing feature of the passive advection problem is that its phase space coincides
with the physical space of the fluid, rendering possible direct experimental observa-
tion and characterization of fractal structures associated with chaotic dynamics.
This chapter is devoted to transiently chaotic advective dynamics in hydrodynam-
ical flows. If a fluid system is open, as schematically shown in Fig. 10.1, transient
chaos can arise. More precisely, a flow is open if there is a net current flowing
through the region of observation. A typical example is the flow around a fixed ob-
stacle, such as a cylinder placed in a channel. In environmental science, flows around
an island or a peninsula can generate transiently chaotic advection. In an open flow,
the dynamical trajectories of advective particles are typically unbounded in the sense
that nearly all particles escape the region where strong stirring takes place in a finite
time. The dynamical invariant sets responsible for such chaotic advective dynamics
are chaotic saddles. Because of the coincidence of the phase space with the physical
space, chaotic saddles and their invariant manifolds are effectively hydrodynami-
cal observables. In the context of the spreading of materials in flows, of interest is
not a single-particle trajectory but rather the motions of an ensemble of trajectories

1 In particular, if the length and velocity scales are L and U, respectively, the typical hydrodynami-
cal time scale is L/U and the diffusive time scale is L2 /Ddiff , where Ddiff is the diffusion coefficient,
whose typical value for water and most tracer substances is of order 10−8 m2 /s. Suppose L = 1 m
and U = 0.1 m/s. The hydrodynamical and the diffusive time scales are thus ten seconds and one
thousand days, respectively, rendering physically irrelevant any diffusive processes in the flow.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 343
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 10,

c Springer Science+Business Media, LLC 2011
344 10 Chaotic Advection in Fluid Flows

inflow outflow

stirring
region

Fig. 10.1 Schematic diagram illustrating an open flow that generates transiently chaotic advective
dynamics. The flow can be time-dependent and relatively complicated in a bounded region, the
stirring region, while the incoming and outgoing flows far away from the region can be simple

originated from a dye droplet. Spreading of substances such as pollutants in a flow

can therefore occur along fractal patterns determined by the unstable manifold of
the chaotic saddle associated with the flow.
This chapter is organized, as follows. After establishing the analogy between
advection in open incompressible flows and chaotic scattering, we consider two
paradigmatic time-periodic flows and treat the dynamics of dye droplets, transport
properties (lobe dynamics), and the fractality of dye boundaries. We then consider
aperiodic or chaotically time-dependent flows and address the problem of coherent
structures. The effect of artificial leaking of the advective dynamics of closed flows
will be shown to reveal useful information via the transients caused by the leak.
The advective dynamics of finite-size particles will also be considered. Finally, how
chemical or biological reactions take place in open flows with chaotic advective
dynamics will be considered, which is of interest in fields ranging from chemical
reactions and combustion to atmospheric chemistry and the population dynamics of
plankton.

10.1 General Setting of Passive Advective Dynamics

For an idealized particle of zero size and zero mass, its velocity ṙ is the flow velocity
u(r,t) at any time instant. In the theory of advection, the velocity field is assumed to
be known and the advective dynamics, or the Lagrangian dynamics, are described by
ṙ(t) = u[r(t),t]. (10.1)
The solution to this equation is the path r(t) of the particle. Because of the explicit
time dependence, (10.1) consists of three first-order ordinary differential equations
that can be nonautonomous and nonlinear. Chaos can then arise, leading to chaotic
advection [26]. Note that the main phase-space variables are the components of r
that specify the position of the particle in the actual physical space.
Advection in two-dimensional, incompressible flows specified by r = (x, y)
and u = (ux , uy ) is especially interesting because of its equivalence to motion
in one-degree-of-freedom Hamiltonian systems. In particular, the condition of
10.1 General Setting of Passive Advective Dynamics 345

incompressibility implies the existence of some stream function [429] ψ (x, y,t) that
determines the velocity components of flow as

∂ ψ (x, y,t) ∂ ψ (x, y,t)

ux (x, y,t) = , uy (x, y,t) = − . (10.2)
∂y ∂x

Incompressibility is ensured because

divu = ∂ ux /∂ x + ∂ uy /∂ y = 0.

The streamlines y(x), the tangent lines to the local velocity u at point (x, y), satisfy
the equation
dy uy
= (10.3)
dx ux
at any instant of time. As a consequence, we have uy dx − ux dy = 0, which implies
that dψ (x, y,t) = 0 according to (10.2). The streamlines are thus contour lines of the
stream function. Equations of motion for the advective dynamics are

∂ ψ (x, y,t) ∂ ψ (x, y,t)

ẋ = , ẏ = − , (10.4)
∂y ∂x
which are equivalent to Hamilton’s equations of motion [268] in classical mechanics
under the “Hamiltonian” ψ (x, y,t). In the Hamiltonian formulation, the coordinate
variables x and y in the physical space correspond to the generalized position and
momentum variables. Consequently, in the (x, y) plane, we obtain area-preserving
tracer dynamics.
Depending on whether the flow is steady or unsteady, the tracer dynamics can
be simple or complex. In particular, for steady flows, the stream function is time-
independent: ψ = ψ (x, y). Equations (10.4) are thus a set of two autonomous first-
order differential equations. The total time derivative of ψ along a tracer trajectory is
dψ (r(t)) ∂ ψ ∂ψ
= ẋ + ẏ = 0, (10.5)
dt ∂x ∂y
indicating that the trajectory actually coincides with a streamline. The dynamics is
integrable, since there exists a conserved quantity, ψ , the analogue of the energy
of conservative motion of a point particle in one dimension. However, when the
stream function depends explicitly on time, the advective dynamics is described by
a time-dependent Hamiltonian function that effectively has one and a half degrees
of freedom. In this case, trajectories do not coincide with the streamlines, and in
fact, they can be much more complicated than the streamlines. Even if the stream-
lines are smooth time-dependent curves, which arise when no turbulence is present
in the flow field, the particle trajectories can be quite complex. Indeed, motions
of tracer particles can typically be chaotic even for simple time-periodic flows.
Chaos associated with tracer dynamics is sometimes called Lagrangian turbulence.
346 10 Chaotic Advection in Fluid Flows

In the past three decades, a great amount of knowledge about advective dynamics
has been obtained both for closed [27, 569, 826] and for open flows [394].
For open two-dimensional incompressible flows with asymptotic simplicity, i.e.,
situations in which there is a central stirring region, and the incoming and the out-
going flows far away from the region are simple and stationary (see Fig. 10.1),
time-dependence is important, but only for the stirring region. The asymptotic dy-
namics are then simple, and the tracer motion can be considered a scattering process
with characteristics of chaotic scattering (Chap. 6), where the stirring region plays
the role of the scattering region.

10.2 Passive Advection in von Kármán Vortex Streets

10.2.1 Flow Model

Flow around an obstacle is a classical problem in fluid mechanics [206, 429].

We consider a viscous, incompressible flow around a long cylinder of radius R0 .
Far away from the obstacle, the flow is uniform. For convenience, we can label
the longitudinal flow direction x and the transverse direction y. Let U be the flow
velocity for x → ±∞. The Reynolds number is defined as
Re = 2R0U/ν , (10.6)
where ν is the kinematic viscosity of the fluid. For sufficiently small values of Re,
the flow is stationary. As Re is increased through a critical value Rec ≈ 80 [206,429],
the stationary solution of the Navier–Stokes equation becomes unstable, and the
flow becomes time-periodic of with period T :
u(r,t + T ) = u(r,t). (10.7)
In this case, vortices are created in the wake of the cylinder, detach from it, and drift
downstream. They gradually lose strength because of viscosity and typically die out
after some distance. New vortices are shed from the cylinder surface at intervals of
half a period T /2, alternating above and below the middle of the cylinder, as shown
in Fig. 10.2. This process generates a von Kármán vortex street behind the cylinder.
In the following, for simplicity we assume that at any instant there are at most two
vortices in the flow. That is, the lifetime of each detached vortex is equal to one
period T .
The flow associated with the von Kármán vortex street was studied by direct
numerical simulation of the Navier–Stokes equation for Re = 250 under the usual
no-slip boundary conditions on the surface of the cylinder [371], based on which an
analytical model for the time-periodic flow was proposed [370, 854]. In the analytic
model, the stream function ψ is given by

ψ (x, y,t) = f (x, y)g(x, y,t), (10.8)

10.2 Passive Advection in von Kármán Vortex Streets 347

Fig. 10.2 Streamlines of the von Kármán flow (10.8) around a cylinder at two different instants
(a,b) of time, separated by one quarter of the full period of the flow. The flow is from left to right.
The vortex shedding can be seen, and this process sets the stirring region to be approximately
1 < x < 4, | y |< 1 [771] (with kind permission from Elsevier Science)

where the factor

f (x, y) = 1 − exp{−a[(x2 + y2 )1/2 − 1]2} (10.9)

guarantees the no-slip boundary condition at the surface of the cylinder whose radius
is chosen to be unity R0 = 1 (equivalently, the radius R0 is taken to be the length unit
of the system). The quantity a−1/2 plays the role of the width of the boundary layer.
The time unit is chosen to be the period of the flow: T = 1. The function g(x, y,t) in
(10.8) is

g(x, y,t) = −W h1 (t)g1 (x, y,t) + W h2 (t)g2 (x, y,t) + u0 ys(x, y), (10.10)

where the first two terms describe the alternating birth and the subsequent death
of the vortices, respectively. The maximum vortex amplitude is W , with time-
dependence described by

h1 (t) = |sin(π t)| , h2 (t) = h1 (t − 1/2). (10.11)

348 10 Chaotic Advection in Fluid Flows

The vortices are of Gaussian shape of characteristic linear size r−1/2 , so the
functions g1 (x, y,t) and g2 (x, y,t) are given by
 & '
gi (x, y,t) = exp −r [x − xi (t)]2 + α 2 [y − (−1)i−1y0 ]2 , i = 1, 2, (10.12)

whose centers move downstream at constant velocity:

x1 (t) = 1 + l0 [t mod 1] ; x2 (t) = x1 (t − 1/2), (10.13)

where y0 is the distance of the vortex centers from the x-axis. The last term in (10.10)
arises from the background flow, and the screening factor
$ %
s(x, y) = 1 − exp −(x − 1)2/α 2 − y2 (10.14)

ensures that the effect of the background flow of dimensionless velocity u0 =

UT /R0 is reduced in the wake. The parameters were set as a = 1, α = 2, r = 0.35,
l0 = 2, y0 = 0.3, u0 = 14, and W = 24 to fit the direct numerical solution to that of
the Navier–Stokes equation [370]. In fact, results obtained from the analytic model
are robust with respect to the dynamics under study, so the exact values of the pa-
rameters are not so important. Only in the wake of the cylinder is the flow nontrivial,
being time-dependent because of the vortex shedding. As a result, the stirring region
is the time-dependent part of the flow, situated within an area of finite extension
(of length about l0 along the x-direction) in the wake. The model thus serves as a
paradigm for a large class of open chaotic flows and has been used widely in the
studies of various aspects of transient chaotic advection [61, 664, 665, 669, 670, 682]
A generalization of these ideas led to the conclusion that the chaotic motion of par-
ticles transported by blood can play an important role in the development of certain
circulatory diseases [681]. An extension of the model to more than two vortices in
the wake has recently been worked out [835].

10.2.2 Advection and Droplet Dynamics

The motions of passively advected particles can be determined by integrating (10.4)

with the stream function (10.8). The chaotic nature of particle motions in the flow
can be seen from Fig. 10.3, where trajectories originated from two nearby initial
conditions are shown. The trajectories separate from each other quickly, indicating
a sensitive dependence on initial conditions, the hallmark of chaos. In both the far
upstream and the far downstream regions, particles move along straight lines. How-
ever, in the wake, there is a chaotic saddle (Fig. 10.4), whose invariant manifolds
play a determining role in particle advection.
The role of the unstable manifold of the chaotic saddle can be assessed by placing
a dye droplet of a continuum of particles upstream into the flow and monitoring the
deformation of the shape of this droplet, the droplet dynamics (the role of the stable
10.2 Passive Advection in von Kármán Vortex Streets 349

Fig. 10.3 For the von Kármán flow (10.8), two representative (a,b) trajectories originated from a
pair of nearby initial conditions. The black region is the cylinder that appears to have an elliptical
shape due to the different scales in the x- and y-directions in the plot. Sensitive dependence on
initial conditions can be seen

manifold will be discussed in Sect. 10.4). If the initial droplet is situated about the x-
axis, it overlaps with the stable manifold of the saddle. Particles exactly on the stable
manifold approach the saddle and never leave it. Neighboring particles approach the
saddle but they typically stay in the vicinity of the saddle for a finite amount of time
before exiting the wake along the unstable manifold of the saddle. This suggests that
the unstable manifold is traced out by particles that stay in the region of observation
in the wake for a relatively long time. The unstable manifold of the chaotic saddle
arising in advection in open flows can be regarded as the main transport route of the
tracer dynamics: particles accumulate on it while being advected away.
In numerical simulations with a finite number of particles, the manifold serves
as a periodically moving template that will eventually be emptied. Figure 10.5
shows the evolution of a droplet in the von Kármán flow. The droplet first be-
comes stretched and folded. It then traces out a moving fractal object, the unstable
manifold. The fact that the region is not uniformly emptied in the last panel and
particles are still visible around the cylinder suggests a type of dynamics similar to
nonhyperbolic chaotic scattering dynamics (Sect. 6.4). In fact, the hyperbolic com-
ponent (of structures resembling the direct product of two Cantor sets) lies outside
the boundary layer and is responsible for an exponential decay of particles, and the
nonhyperbolic component is located close to the smooth cylinder surface, which de-
termines the long-time behavior of particle decay from the wake [370]. The different
patterns of these two components can be distinguished in Fig. 10.4.
These numerical investigations led Sommerer and coworkers [725] to an ex-
perimental investigation of passive advection in the wake of a cylinder. They
demonstrated that dye droplets trace out the unstable manifold of the chaotic saddle
in the wake, as shown in Fig. 1.19. They also determined the escape rate and the
Lyapunov exponent of the transiently chaotic dynamics as well as the asymptotic
box-counting dimension of the dye droplet. For a related experiment, see [276].
The von Kármán vortex street is in fact not a particular property of flows with
a cylindrical obstacle. Approximately, most two-dimensional flows past an obstacle
350 10 Chaotic Advection in Fluid Flows

Fig. 10.4 Chaotic saddle in the von Kármán flow obtained by the PIM-triple method. The
set of black points represents initial conditions of tracers belonging to a certain time instant
(t = 0.3mod1), which do not escape the wake either forward or backward in time. In the advection
problem the chaotic saddle appears in the space of the fluid, and it can also be considered the set
of never-escaping fluid elements [592] (copyright 1995, the American Physical Society)

1 t=0 1 t=0.4
0 0
y

−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=0.6 1 t=1.0
0 0
y
y

−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=1.4 1 t=1.6
0 0
y
y

−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=2.0 1 t=4.0
0 0
y
y

−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x

Fig. 10.5 Droplet dynamics: time evolution of a droplet of 20,000 tracers in the von Kármán flow
shown at different dimensionless time instants [774] (copyright 2000, the American Institute of
Physics)
10.3 Point Vortex Problems 351

Fig. 10.6 Patterns of a sea current around the Gran Canaria island. Black lines are numerically
obtained streaklines, i.e., traces of continuously injected dye from localized point sources. Several
streaklines are shown in panel (a), and a single streakline is shown in panel (b). Many of the
streaklines trace out a fractal pattern, the unstable manifold of the chaotic saddle in the wake of the
island [28] (with kind permission from Elsevier Science)

have this property, provided that their Reynolds numbers are in an appropriate range.
As a result, von Kármán vortices are found in many realistic situations. For example,
Fig. 10.6 shows the dye patterns obtained from a simulation of the sea current around
the island of Gran Canaria [28]. Fractal filaments generated by the von Kármán
vortices are visible and pronounced.

10.3 Point Vortex Problems

10.3.1 Vortex Dynamics

The dynamics of point vortices in two-dimensional or of parallel vortex lines in

three-dimensional ideal incompressible fluids is another classical problem in hydro-
dynamics. It has long been known [412, 467, 549, 662] that the equations of motion
of a system of n such vortices can be cast into a canonical form. The Hamiltonian is

1
H({xi , yi }) = − ∑ ΓiΓj ln ri, j ,
π i<
(10.15)
j
352 10 Chaotic Advection in Fluid Flows

where (xi , yi ) stands for the position of vortex i of strength Γi (i = 1, . . . , n) in the

(x, y)-plane, and ri, j is the distance between vortices i and j. The value E of the
Hamiltonian is constant in time and is thus effectively the energy of the vortex
system. The equations of motion are determined by the corresponding Hamilton
equations [412, 467, 549, 662]:

∂H ∂H
Γi ẋi = , Γi ẏi = − . (10.16)
∂ yi ∂ xi

The equations can be made dimensionless by the transformations

π L2 Γ2
(x, y) → (Lx, Ly), t→ t, H→ H, (10.17)
Γ 2π
where Γ is a preselected vortex strength and L denotes a characteristic length scale.
The vortex dynamics possesses conserved quantities. An example is the coordi-
nates of the center of vorticity: x̄ = ∑i Γi xi / ∑i Γi and ȳ = ∑i Γi yi / ∑i Γi . Besides en-
ergy conservation, there are four more conservative constraints. Thus, out of the 2n
coordinates of the n-vortex problem, only 2n − 5 are independent. Nonintegrability
and chaos require at least three independent nonlinear equations, and hence the
dynamics of four or more vortices can be chaotic.
An interesting example is the collision of vortex pairs. When two vortex pairs
encounter each other, they interact and exchange partners for a finite amount of
time. The new couples strongly perturb each other, until a new collision leads to
an exchange again. The two original vortex pairs are recovered, and they separate
along a straight line from each other. The role of chaos in this scattering process was
pointed out by Eckhardt and Aref [214].
An isolated vortex of strength Γ generates at distance r from its center a cir-
culational flow with a velocity field proportional in modulus to Γ /r. The stream
function ψ (x, y) is −(Γ /π ) ln r. In a system of n vortices these contributions are
superimposed, yielding
Γj
ψ (x, y,t) = − ∑ ln r j (t), (10.18)
j π

where r j (t) is the distance of point (x, y) from vortex j. Because the vortices follow
their own dynamics, the distances r j (t), and consequently the stream function, are
time-dependent. For a particle advected by the vortices, the equation of motion is
given by (10.4) with the stream function ψ (x, y,t) of (10.18). It is worth noting that
the advection problem in the field of n vortices can also be considered as a special
(n + 1)-vortex problem in which one of the vortices is of vorticity zero. This special
vortex does not have any feedback on the flow, and hence its motion corresponds
to that of a tracer particle. Thus, the passive advection in the field of three or more
vortices is typically chaotic.
A condition for a vortex flow to be open is that the total vortex strength vanish:

∑ Γi = 0. (10.19)
i
10.3 Point Vortex Problems 353

In this case the center-of-vorticity coordinates are formally infinite, and there
is no constraint on the vortices to remain bounded to a finite fluid domain. What
happens is that the four vortices can remain close to each other but their geometric
center moves ahead in a certain direction. Far from the vortices, the streamlines
are straight lines along which the vortices can be approached (for ∑i Γi = 0, the
streamlines far from the vortices are closed curves).

10.3.2 Advection by Leapfrogging Vortex Pairs

Consider an example of open flow of four vortices whereby two vortex pairs of equal
strengths (Γ1 = Γ2 = −Γ3 = −Γ4 ≡ Γ ) move in the same direction along a symmetry
axis, the x-axis. The Hamiltonian of the system is [590] (for notation see Fig. 10.7)

Γ2
H(x1 , x2 , y1 , y2 ) = (−2 ln r1,2 + 2 lnr2,4 + ln r1,4 + ln r2,3 )
2π
Γ2 (x1 − x2 )2 + (y1 + y2)2
= ln 4y1 y2 = E. (10.20)
2π (x1 − x2 )2 + (y1 − y2)2

Since the geometric center of the x-coordinates, x0 ≡ (x1 + x2 )/2, does not appear
in H, the conjugate variable

2y0 ≡ (y1 + y2) = const (10.21)

is conserved during the motion, where 2y0 can be considered as the average width of
the vortex pairs. It is convenient to choose 2y0 as the characteristic length L = 2y0
and then to rescale the equations of motion according to (10.17). The following
variables, besides the center-of-mass coordinates, can then be used:

xr ≡ x2 − x1 , yr ≡ y2 − y1 . (10.22)

Fig. 10.7 Arrangement = Γ4

of the four vortices in the
= Γ3
leapfrogging problem
354 10 Chaotic Advection in Fluid Flows

Energy conservation enables us to obtain an explicit formula for trajectories in the

relative coordinates:
1 1
− = e−E . (10.23)
1 − yr 1 + x2r
2

Bounded trajectories are present for energy values E > Es ≡ 0, where no real solu-
tion yr exists for xr → ∞. This corresponds to a strictly periodic motion of the vortex
pairs, called leapfrogging [206, 709]. Since the Hamiltonian depends on xr and yr
only, the vortex dynamics is integrable. The equations of motion can be solved by
direct numerical integration, which yields the time-dependence of the vortex-center
coordinates xi (t) and yi (t) (i = 1, 2).
The dimensionless stream function (10.18) of the advection problem takes the
form
r3 (t) r4 (t)
ψ (x, y,t) = ln , (10.24)
r1 (t) r2 (t)

where ri (t) denotes the distance of the advected particle at (x, y) from the center
of vortex i at time t. The mixing region can be conveniently chosen as a circle
containing all vortices in a frame comoving with the geometric center [x0 (t) =
(x1 (t) + x2 (t))/2, y = 0] along the x-axis. Individual trajectories are again typically
transiently chaotic. A representative chaotic saddle responsible for transient chaos
is shown in Fig. 10.8. We note that the saddle is quite dense and can be regarded as
consisting of a hyperbolic component (the direct product of two Cantor sets) and a

Fig. 10.8 The chaotic saddle of the leapfrogging problem for y > 0 in the comoving frame on a
stroboscopic map at t = 0, obtained by means of the PIM-triple algorithm. The full chaotic saddle
is obtained by mirroring this set to the x-axis. Points P1,2 denote fixed points along the x-axis. The
vortex centers are marked by dots. The value of energy is E = ln 2 (with kind permission from the
institute of physics)
10.3 Point Vortex Problems 355

nonhyperbolic one located about the ellipsoidal regions containing the vortices that
are not accessible by scattering trajectories. These are the regions where the effect
of one vortex is more pronounced than that of any other and can therefore be called
the vortex core for the Lagrangian dynamics. Inside this core the effects of other
vortices can be regarded as a weak perturbation, and the conditions of the KAM
theory are thus fulfilled [430, 431], where the boundary of the core is a KAM torus.
In fact, KAM tori surrounding vortices are present in the advection induced by any
number of vortices [35].
For the advection problem defined by (10.24), the unstable manifold of the
chaotic saddle is shown in Fig. 10.9, which is typically traced out by particle
droplets. It is quite remarkable that this unstable manifold is similar to the pat-
tern obtained from an experiment of three-dimensional smoke rings [206, 710], as
shown in Fig. 10.10. In the experiment, the smoke itself plays the role of the dye.
This example shows that experimentally visualizable flow patterns, such as streak-
line patterns, are in fact unstable manifolds of chaotic saddles. The occurrence of
fractal streakline patterns thus provides evidence of transient chaos in the advection
problem.
The leapfrogging vortex rings can be considered as building blocks of turbulent
jets [278]. The particle transport in such a flow thus gives hints about the entrainment
of ambient fluid by jetlike flows such as cumulus clouds [133].
A dye pattern (an unstable manifold) similar to that of Fig. 10.9 was observed in
a recent experiment on open flow advection by Gouillart et al. [276], who used two
rotating rods to stir dye in a channel of slowly moving viscous flow.

Fig. 10.9 For the advective dynamics of the leapfrogging problem determined by (10.24), the un-
stable manifold of the full chaotic saddle [590] (with kind permission from the Institute of Physics)
356 10 Chaotic Advection in Fluid Flows

Fig. 10.10 Photograph of a planar cross section through two leapfrogging smoke rings from the
1978 experiment of Yamada and Matsui [206] (with kind permission from the Parabolic Press)

10.3.3 Lobe Dynamics

Examining the lobe dynamics is a technique for quantifying transport due to chaotic
advection, as pioneered by Rom-Kedar, Leonard, Wiggins, and coworkers [48, 647]
(for a recent review, see [499]). The method is based on the observation that invariant
manifolds govern the motion of fluid regions, called lobes, which are formed by
segments of the stable and the unstable manifolds. Lobes can transport fluid between
regions of qualitatively different flow characters.
The general idea of lobe dynamics can be conveniently illustrated by the example
of leapfrogging vortex pairs. In particular, the advective dynamics has two hyper-
bolic fixed points, P1 and P2 , along the axis of symmetry (x) in the comoving frame,
which are elements of the chaotic saddle (see Fig. 10.8). The stable and the unstable
manifolds emanating from the symmetry axis are denoted by W1s and W2u , respec-
tively. Figure 10.11 illustrates, schematically, some important topological features,
for simplicity on one half-plane only. One can define an interaction region S, a sub-
set of the stirring region, bounded by segment P2 P of W2u and segment P1 P of W1s ,
where P is the primary intersection point of the manifolds. Lobes formed by W2u and
the boundary of the interaction region are denoted by Ei (Di ) with i > 0 if they are
fully or partially inside (fully outside) S. The advective dynamics transforms each
lobe Ei (Di ) into Ei+1 (Di+1 ) after one flow period, where the convention is that the
first (last) lobe that lies inside the interaction region has label i = 0. These rules
define lobes with negative i. Due to the incompressibility of the flow, the areas of
all lobes are equal. Numerically computed manifold branches, however, show that
the actual topological pattern is much more complex than in the schematic diagram,
as shown in Fig. 10.12. For example, due to the strong stretching and folding inside
the interaction region, lobe E0 has a strange shape and intersects with lobe D0 at six
points (in contrast to Fig. 10.11, where the number of intersecting points is two).
The idea of lobe dynamics can be applied to any Hamiltonian (volume-
preserving) problem (see, e.g., Fig. 6.25). In the hydrodynamical context, lobes
are relevant because they connect fluid regions of different characters: outside the
interacting region S the overall flow surrounds the vortices and moves to the left in
the comoving frame, while fluid inside the interaction region tends to remain in the
10.3 Point Vortex Problems 357

P
D1 E−1

E1 D−1

s u
D2 W1 E0 W2 E−2
D0
D3 E−3

P1 P2

Fig. 10.11 Schematic diagram of the stable manifold W1s of the fixed point P1 and the unstable
manifold W2u of the fixed point P2 on a stroboscopic map taken at integer multiples of the period
of the flow, shown above the axis of symmetry x = 0, where P represents the primary intersection
point. Lobes Ei and Di are indicated for a few values of i

Fig. 10.12 Branches of the numerically determined manifolds W1s and W2u and the corresponding
lobes in the leapfrogging problem (with kind permission from the Institute of Physics)

vicinity of the vortices. Lobe E−1 is transformed in one time unit into E0 , i.e., the
fluid area of E−1 is transported from the ambient region into a region close to the
vortices. While there is a tendency for fluid to be trapped by the vortices, permanent
trapping is impossible due to escape. The same area flows out of the interaction
region via the lobe transformation D0 → D1 . Dividing the area of one lobe by the
period of the flow, one obtains the average material flux inside (and outside) the
interaction region. This is the Lagrangian background of the entrainment in fluid
mechanics [133].
The lobe dynamics can be used to trace out the escape-time distribution in the
interacting region. For example, points escaping the interaction region S in one time
unit, regardless of when they entered the region, lie in D0 . Those spending just
358 10 Chaotic Advection in Fluid Flows

Fig. 10.13 Images Ei of lobe E−1 in different colors: i = −1: gray; 0: blue; 1: green; 2 magenta;
3, 4: white. Lobes with i ≥ 5 are marked by light red. They contain points entering the interaction
region more than five time units earlier and approximate the unstable manifold [591] (with kind
permission from World Scientific Publishing Co.)

one time unit inside lie in E0 ∩ D0 . Points from Ei ∩ D j spend i − j + 1 time units in
the interaction region. Assuming exponential decay of particles from the interaction
region, one can estimate the escape rate as the logarithm of the area ratio between
S and S − D0 . This is, however, typically a poor estimate, since the exponential
decay is an asymptotic property that sets in with good accuracy only after some
time (see (1.6)).
Lobes in general possess a convoluted structure, as shown in Fig. 10.13, where
images of lobe E−1 (gray) over several periods are displayed. The first image E0
(blue) is mapped onto E1 (green), which is so strongly elongated that it does not
appear to be connected (although it is). A considerable part of this lobe is already
outside S. Higher-order lobes become more and more convoluted.

10.4 Dye Boundaries

We have seen that dye particles trace out the unstable manifold of an underlying
chaotic saddle in a generic open flow. It is natural to ask how boundaries between
different dyes behave [407, 592, 670, 787], which are defined as the borderlines
between different colors injected into the flow somewhere in the inflow region.
In particular, imagine that particles are injected continuously far upstream at x = xin
so that the colors above and below a critical value yc are different. We can focus on
how the boundary between these two colors evolves over time. On a stroboscopic
map taken with respect to the period of the flow, the boundary pattern will stabilize
after some time. If the stable manifold of the chaotic saddle intersects the region
10.4 Dye Boundaries 359

of both colors at xin , both types of dyes tend to trace out the unstable manifold
asymptotically, implying that the stabilized boundary between the two colors con-
tains a fractal part, the unstable manifold of the chaotic saddle, or a fractal subset
of it. The dye boundary contains a nonfractal part as well. The particular shape of
the boundary depends on the value of xin and yc , but the box-counting dimension is
independent of such details [592] and coincides with the box-counting dimension
Du,0 of the unstable manifold. An example from the von Kármán flow is shown in
Fig. 10.14.
Dye exit boundaries can be obtained by sprinkling particles in a domain contain-
ing the stirring region. Particles are followed until they cross a line xout far away
downstream. Their initial coordinates are colored depending on whether after cross-
ing x = xout on a stroboscopic map they lie above or below a preselected value yc .
These boundaries are analogues of exit basin boundaries in chaotic scattering (and
are also similar to fractal basin boundaries), the fractal part of which contains the
stable manifold of the chaotic saddle, as shown in Fig. 10.15.

Fig. 10.14 Dye boundary in the von Kármán flow. Shown is the boundary between black and
white dyes injected from xin = −6 with yc = 0 on a stroboscopic map (taken at time t = 0.3mod1).
Particles with initial coordinates xin , y ≥ 0, are black. (a) Global pattern and (b) enlargement of a
part of (a). The fractal part contains the unstable manifold [592] (copyright 1995, the American
Physical Society)
360 10 Chaotic Advection in Fluid Flows

Fig. 10.15 Dye exit boundary in the von Kármán flow. Initial conditions are labeled black or
white, depending on whether after crossing the x = xout = 6 line on the stroboscopic map (taken at
time t = 0.3mod1), they lie in the half-plane y ≥ 0 or y < 0, respectively. The fractal part contains
the stable manifold of the chaotic saddle [592] (copyright 1995, the American Physical Society)

To illustrate a general property of dye boundaries, imagine that the flow in the far
upstream region is separated into bands perpendicular to the inflow velocity whose
length is exactly the distance that the fluid traverses over one full period. Color this
infinite set of bands periodically with at least three colors. As the fluid enters the stir-
ring region, some part of each region of color becomes trapped around the chaotic
saddle and flows away slowly from there. So there is time for the next color to come
close to the previous one around the unstable manifold. The dye boundary will then
have smooth and also fractal components, and the latter have the Wada property
(Sect. 5.5). In the neighborhood of any point on the fractal part of the boundary,
particles of all colors are present, as shown in Fig. 10.16.
The Wada property can be understood as follows. Consider the motion of a dye
droplet in the time-reversed dynamics. If it overlaps initially with the unstable mani-
fold, it traces out the stable manifold more and more accurately with time. The stable
manifold stretches up to infinity in the upstream region. Unavoidably, it crosses an
infinite number of colored bands. When following the dynamics of particles very
close to the stable manifold under the again reversed, i.e., forward, dynamics, we
see that they all end up in a small droplet about a point of the chaotic saddle, where
all the colors accumulate. The points of the droplet on the unstable manifold flow
along the manifold much more slowly than those outside the manifold. These points
have thus in their arbitrarily small neighborhoods all the colors. This observation
shows that not only the chaotic saddle, but also the neighborhood of its unstable
manifold is strongly stirred in open chaotic flows. The argument indicates that Wada
dye boundaries can be expected in all flows where the stable manifold of the entire
chaotic saddle stretches over regions of infinitely many different colors. Note that
for dye boundaries, it is the unstable manifold that may exhibit the Wada property,
not the stable manifold as in basin boundaries (see Sect. 5.5). The argument also
10.4 Dye Boundaries 361

Fig. 10.16 Wada dye boundaries in the leapfrogging flow. The dye distribution is shown around
the leapfrogging vortex pairs swimming in a fluid colored bandwise in red, white, and green on a
stroboscopic map. The lower panel is a magnification of a small region in the upper panel, which
illustrates the Wada property [787] (with kind permission from Elsevier Science)

indicates that the boundaries between particle ensembles of different behaviors are
stable or unstable manifolds. The boundaries are material lines, and therefore the
fluid flux across them is zero. In chaotic flows they nevertheless provide a mecha-
nism (e.g., via the lobe dynamics) for material exchange between different regions.
For three-dimensional flows, many of the above phenomena persist. A basic
new feature (see Chap. 8) is that manifolds can be either locally one-dimensional
or locally two-dimensional, depending on the details of the dynamics. For chaotic
dynamics both the stable and the unstable manifolds are fractal, indicating the pres-
ence of a chaotic saddle in the flow.
362 10 Chaotic Advection in Fluid Flows

Fig. 10.17 Concentration distribution of a passive scalar (red: full concentration; dark blue: zero
concentration) in a micromixer. The flow in the main channel is stationary and is manipulated by
time-periodic flows in the secondary channels. Varying the frequencies of these perturbations can
enhance mixing. (Picture by I. Mezić.)

One area in which the observations of this chapter are relevant is microfluid de-
vices. Recent technological advances have made the fabrication of microchannels
of a few hundred micrometers in cross section possible [96, 140, 389, 733]. These
devices are used in printers and biomedical instruments. At this scale, fluid viscos-
ity dominates, turbulence cannot be present, and diffusion is often negligible. As a
result, the only source of effective stirring is chaotic advection. A typical microfluid
arrangement is a channel, along which fluid moves due to a pressure gradient. It
is often perturbed time-periodically by flows from secondary channels to generate
stirring regions around such junctions, as shown in Fig. 10.17. In this case, chaotic
saddles provide the dynamical mechanism for efficient stirring.

10.5 Advection in Aperiodic Flows

In this section we address the question of how passive advective dynamics change in
the realistic situation in which the temporal behavior of the flow field is more com-
plicated than periodic. Periodic flows have the special feature that the asymptotic
time-dependence can be recovered from finite-time velocity samples. This feature
is the basis for the use of stroboscopic maps. Aperiodic flows, however, cannot be
reconstructed for all times from a one-period velocity sample. In such a case, ve-
locity data are available only over finite-time intervals. Chaotic-like dynamics of
advected particles is then unavoidably of a transient character. Stroboscopic maps
are, however, no longer available and the concepts of hyperbolic orbits, and stable
and unstable manifolds, cannot be directly applied. Nonetheless, analogues of these
quantities can be defined, and certain features such as the fractality of tracer patterns
can be robust even in open aperiodic flows.
10.5 Advection in Aperiodic Flows 363

10.5.1 Coherent Structures in Aperiodic Flows

Coherent structures are well known to characterize turbulent velocity fields, or flows
of general time-dependence. They can be defined as fluid regions exhibiting coher-
ent behavior over long periods of time (see, e.g., [504]). Different criteria based
on the instantaneous velocity field, i.e., the Eulerian frame, have been proposed,
but they all have the drawback of being reference-frame-dependent. It has been ac-
cepted that a proper definition of coherent structures should be established in the
particle-based, i.e., the Lagrangian, framework. In particular, coherent structures
are distinguished sets of fluid particles. This change of view was due to a series
of works by Wiggins, Jones, Haller, and coworkers (see, e.g., [307, 310, 499, 524]),
and can be interpreted as the increasing use of dynamical-system concepts in fluid
dynamics.
Examples of coherent structures are material filaments and vortex cores of finite
lifetime. We have seen in Sects. 10.3.3 and 10.4 that such structures in time-periodic
flows are related to the invariant manifolds of hyperbolic orbits and to KAM tori,
respectively. Both types of structures are special material lines that enhance or in-
hibit stirring. The definition of Lagrangian coherent structures in general aperiodic
flows thus requires a proper generalization of these concepts to situations in which
no stroboscopic repetition is present.
An earlier approach to studying coherent structures was based on finding stagna-
tion points (hyperbolic points in the Eulerian frame) of the flow and searching for
such points associated with finite-time Lagrangian stable and unstable manifolds
[307, 524]. In this setting the analogue of the lobe dynamics can be worked out, and
material transport can be quantified (for a recent review see [499]).
A more general approach due to Haller and coworkers [298,304,305,310,707] is
applicable regardless of whether the flow has Eulerian stagnation points. The basic
idea is to examine the stretching and the distribution of future or past stretchings
assigned to any point in the fluid. The stretchings constitute the stretching field.
Haller and coworkers obtained rigorous results on the characterization of stretching
around material lines of aperiodic flows defined over finite intervals. They pointed
out that the analogue of the stable manifold exists in the form of material surfaces
of large future stretchings. These surfaces are also called repelling surfaces, since
particles on both sides of them depart in opposite directions and have qualitatively
different future behaviors. Regions of large past stretchings correspond to the unsta-
ble manifold (called attracting material surfaces). Special intersections between the
repelling and the attracting surfaces or lines form hyperbolic cores, the analogues of
hyperbolic orbits. Due to the finite-time observation of aperiodic flows, the hyper-
bolic manifolds are not unique, but for sufficiently long time intervals they appear
to be locally unique up to the numerical precision [310]. Dominant hyperbolic co-
herent structures turn out to be maximizing surfaces (ridges) of the stretching field.
The maxima of large past stretchings correspond to unstable manifolds of strongly
hyperbolic orbits. Indeed, contours of constant dye concentrations were found in
experiments [805] to align with lines of large past stretchings. This is analogous to
the observation in Sect. 10.4 according to which dye boundaries contain unstable
manifolds. The maxima of large future stretchings separate regions of qualitatively
364 10 Chaotic Advection in Fluid Flows

different tracer dynamics. This is a generalization of the statement in Sect. 10.4,

according to which dye exit boundaries contain stable manifolds. Elliptic coherent
structures, or finite-time vortex cores, are defined as regions in which both the future
and past stretchings are weak.
A convenient measure to characterize the stretching field is Lyapunov exponents.
There are two types of such exponents. The first type is finite-time Lyapunov ex-
ponents, the instantaneous growth rates over a finite interval τ of the separation
distance D between two nearby trajectories starting from some point x, which are
defined as [597]
1 D(τ )
λ (τ , x) = lim ln , (10.25)
D0 →0 τ D0
where the time interval τ is chosen to be shorter than or equal to the interval over
which the aperiodic fluid flow is defined. The values of the finite-time Lyapunov ex-
ponents are distributed about the average Lyapunov exponent λ1 . They also depend
on the time instant t when the two nearby trajectories are initiated.
The second type of measure is finite-size Lyapunov exponents, which charac-
terize the local amplification rates of a separation distance of size δ for a trajectory
starting from point x until the separation reaches an amplified value rδ , where r > 1.
These exponents are defined as [31, 82]

1
λ (δ , r, x) = ln r, (10.26)
τ (δ , r, x)

where τ (δ , r, x) is the time needed for the separation to reach rδ . If the time is
longer than the interval over which the fluid flow is defined, no finite-size Lya-
punov exponent can be assigned to the given point x. The exponents also depend on
the time instant t when the separation distance is set to δ . Since taking the limit
δ → 0 is not necessary, finite-size Lyapunov exponents are easier to determine
in numerical simulations than finite-time Lyapunov exponents. Typically, the two
Lyapunov-exponent fields are not identical, since they also depend on parameters τ ,
δ , and r. Their extrema appear, however, to coincide. It can therefore be said that
hyperbolic Lagrangian coherent structures are maxima (ridges) of both the finite-
time and the finite-size Lyapunov-exponent fields. Elliptic coherent structures, i.e.,
vortex cores, are assigned to near-zero values of these fields.
As an example, Fig. 10.18 shows the finite-size Lyapunov-exponent field in a
square-shaped lake of linear size 2 km. The flow is generated by wind and is ob-
tained as a numerical solution to the shallow-water equations [579]. Initially, a wind
of strength 14 m/s blows from the northwest; then it changes abruptly but continu-
ously to northeast, and after some time it changes back again, etc. The period over
which the wind direction is constant is not fixed. The period in fact changes ran-
domly from an ensemble with a mean value given as 8 h. The maximum deviation is
±2 h. The flow field was determined over the period of about 60 changes in the wind
direction. The largest values of the finite-size Lyapunov exponent (largest stretch-
ing) in forward time are marked by red. They represent the stable manifolds. The
largest values in backward time are marked by blue, and they trace out the unstable
10.5 Advection in Aperiodic Flows 365

Fig. 10.18 Finite-size Lyapunov-exponent field in an aperiodically wind-driven lake at an instant

of time, where the change in the wind direction occurs in an aperiodic manner at the average of 8 h.
Color coding indicates the values of the finite-size Lyapunov exponent in units of 1/h. Length is
measured in meters. Computational parameters are δ = 7 m and r = 50. (Picture by M. Pattantyús-
Ábrahám.)

manifolds. Several intersections can be seen, which correspond to periodic orbits,

homoclinic and heteroclinic intersections. Light gray marks near-zero values of the
Lyapunov field and indicates elliptic coherent structures (vortex cores). The overall
picture changes with time, but the main topological features remain unchanged over
the period of investigation.
An aspect that makes the concept of Lagrangian coherent structures relevant is
that being material lines or surfaces, they prohibit fluid flux across themselves. They
are barriers to transport. Unstable manifolds thus trap fluid, a property important in
the context of spreading of pollutants. Applications of coherent structures are there-
fore particularly important in environmental flows or in geophysical fluid dynamics
[142, 413, 477, 827], ranging from oceanic upwelling regions [655] to jetstreams
[760]. Unstable manifolds are not only convenient observables; they may also have
biological relevance [374]. For example, top marine predators, such as frigate birds,
have been found to track Lagrangian coherent structures in their effort to locate food
patches. For a recent review of Lagrangian coherent structures, see [580].
A different type of problem of engineering importance is the separation profile of
the boundary layer over obstacles merged in uniform flows, a classical phenomenon
in fluid dynamics. Traditional approaches were based on Eulerian concepts [429],
but recent work by Haller and coworkers [306, 743] indicated that the proper view
for understanding this unsteady process should be Lagrangian. The separation pro-
file turns out to be a coherent structure of filamentary shape. Such a profile can
be thought of as a material line that attracts and ejects particles near the (time-
dependent) separation point. The profile behaves thus as an unstable manifold. A
new feature is that, due to the degeneracy of fixed points on no-slip boundaries,
the manifold is nonhyperbolic. It corresponds nevertheless to a ridge in the past
stretching field, as shown in Fig. 10.19. The theory predicts separation points and
366 10 Chaotic Advection in Fluid Flows

Fig. 10.19 Finite-time Lyapunov-exponent field, computed from integrating backward in time, in
an aperiodic flow around an airfoil at an instant of time. The color bar marks the nondimensional-
ized finite-time Lyapunov-exponent values. The separation profile which separates the free stream
flow from the wake appears as a line with the highest value of the finite-time Lyapunov exponent
[707] (with kind permission from Elsevier Science)

separation angles in unsteady flows, and hence it is of use in monitoring and control-
ling the separation process, which has far-reaching consequences for aerodynamic
properties such as lift and drag.
An alternative method to determine coherent structures is based on the concept
of almost invariant sets mentioned briefly in Sect. 2.6. In the context of advection,
these are regions where fluid is trapped, regions that hardly mix with their neigh-
borhood, i.e., coherent structures. One should then consider the Frobenius–Perron
(or transfer) operator, defined by (2.67), and apply it to the advection problem. In an
incompressible flow the advective dynamics is area-preserving (J ≡ 1). For practi-
cal purposes, one can partition a closed fluid region into boxes and determine the
transition probabilities among all the boxes, in order to approximate the operator
by a linear matrix whose entries are the transition probabilities [173, 248–251]. The
largest eigenvalue of the matrix is unity, as it should be in a closed system, and the
eigenfunction is constant in chaotic regions. The second eigenvalue, provided that it
lies close to unity, appears to have an eigenfunction that concentrates on almost in-
variant sets (coherent structures) only. The coherent structures obtained in this way
practically agree with the ridges of the finite-time or finite-size Lyapunov exponent
fields [248]. As pointed out by Pikovsky and coworkers [600, 601], an analogous
method can be applied to the problem of passive scalars in the presence of diffu-
sion, where the “noisy” version of the transfer operator should be considered.

10.5.2 Open Aperiodic Flows

In open flows of aperiodic time-dependence, analogues not only of hyperbolic orbits

but also of chaotic saddles exist. The key observation is that there are orbits that
10.5 Advection in Aperiodic Flows 367

never escape the mixing region of interest. Particles on these bounded orbits are
never able to reach the upstream or the downstream region. In particular, irregular
flows can have some kind of topological stability, meaning that they preserve some
topological properties of the flow, such as the number of dominant vortex cores.
If the flow is defined over a sufficiently long time interval, the flow pattern may
repeat the same qualitative features a large number of times. In such cases a multi-
plicity of nonescaping orbits can be present in the mixing region, and this set appears
to be fractal numerically [352, 544]. Because the orbits are unstable, their union can
be regarded as a chaotic saddle. As an extension of the concept of hyperbolic mate-
rial lines of Sect. 10.5.1, we can then speak of fractal manifolds of this saddle.
To see these features, we present here an example, which is the generaliza-
tion of the leapfrogging problem treated in Sect. 10.3.2. Breaking the symmetry
of the leapfrogging problem leads to a type of vortex dynamics characterized by
the asymptotic formation of two noninteracting vortex pairs. If, however, the vortex
strengths are also changed so that no pairs can be formed but the resulting vorticity
is still zero, i.e., ∑4i=1 Γi = 0, one obtains a long-term aperiodic vortex motion. The
average distance among the vortices remains bounded, but their geometric center
moves in a certain direction. The stirring region can be any region enclosing all four
vortices in a comoving frame. The stable manifold can be studied by monitoring
the delay times (escape times) from the mixing region. Using tracer trajectories, we
can obtain the time spent in the stirring region for each initial condition. The spatial
distribution of delay times for the four vortices is shown in Fig. 10.20a. It can be
seen that, besides the vortex cores, points with large escape times lie on a filamen-
tary structure. Examples of magnifications are shown in Fig. 10.20b, where it can
be seen that the complex patterns are present on all smaller scales. One can also
construct a set corresponding to the time-reversed tracer dynamics starting from the
same set of initial conditions, which is the unstable manifold that exhibits similar
patterns, as shown in Fig. 10.20c. The set of intersections of these two foliations has
the property that trajectories starting from it never leave the mixing region either
under the forward or the backward dynamics. It is thus a natural generalization of a
chaotic saddle, as shown in Fig. 10.20d, although the flow is aperiodic.
An efficient way of visualizing the unstable manifold is to follow a droplet of
dye injected into the mixing region. Snapshots taken at different times are shown
in Fig. 10.21. It can be seen that the ensemble tends to produce a complicated
filamentary structure. In contrast to the periodic generation of identical lobes (cf.
Fig. 10.9), here the emerging patterns continuously change their form and size due
to the aperiodic motion of the vortices driving the flow. By investigating several
tracer ensembles, one finds on average an exponential decay N(t)/N0 ∼ exp(−κ t)
in the dimensionless time t with escape rate κ = 3.0. The average Lyapunov expo-
nent measured along several orbits of long escape times is estimated to be λ1 ≈ 55.
The fractal properties of the stable foliations can be extracted by covering the sin-
gularities in the delay-time function as in any chaotic scattering process, using boxes
of size ε . The slope of the log N(ε ) versus log ε plot yields, after averaging over
(1)
several ensembles, a well-defined value D0 = 0.95 for the partial box-counting
(1)
dimension, in agreement with the formula D0 ≈ 1 − κ /λ1.
368 10 Chaotic Advection in Fluid Flows

Fig. 10.20 Advection in an aperiodic four-vortex problem (Γ1 = Γ2 = Γ3 = 1, Γ4 = −3).

(a) Delay time from the stirring region, represented on a gray scale, as a function of the initial
position of 160,000 tracers distributed uniformly in a frame comoving with the four vortices. Co-
ordinates are measured in nondimensionalized units. Brighter points correspond to larger escape
times, distributed in the interval [0.18, 1.2] in dimensionless time. (b) Magnification of part of (a).
(c) Dots representing initial conditions whose escape times are larger than 1.2 under the backward
dynamics, which form approximately the unstable manifold. (d) Dots representing initial posi-
tions whose escape times are larger than 1.2 under both the forward and the backward dynamics.
These dots thus represent approximately the underlying chaotic saddle [544] (copyright 1998, the
American Physical Society)

While it seems surprising that the tracer dynamics possesses a well-defined

escape rate and average Lyapunov exponent, and exhibits a pronounced fractal struc-
ture although the flow is temporally irregular, theoretically these features can be
understood in the framework of random maps (Sect. 4.6). In particular, one can ob-
serve the motion of a particle advected by an aperiodically time-dependent flow
at integer multiples of a preselected time lag t0 and define a sequence of “stro-
boscopic” maps that connect the coordinates at t = nt0 with those at t = (n + 1)t0 .
In contrast to periodic flows, this map does depend on the time at which the snapshot
is taken. As a result, following a trajectory in discrete times requires the application
of a sequence of different maps. The actual form of these maps is not known a pri-
ori in general. If, however, the flow preserves some qualitative features, e.g., it is
topologically stable, the map can be assumed to have the form (4.49), in which the
parameters are n-dependent. For sufficiently complex flows and not very short time
10.5 Advection in Aperiodic Flows 369

Fig. 10.21 Temporal evolution of an ensemble of 160,000 particles initially placed on a square.
Snapshots are taken at dimensionless times (a) t = 0, (b) t = 0.1, (c) t = 0.2, (d) t = 0.3, (e) t = 0.4,
and (f) t = 0.5 [544] (copyright 1998, the American Physical Society)

lags, the n-dependence might be so irregular that subsequent maps correspond to

independent choices. This means that on each iterate n, the map (more precisely, the
set of its parameters) is randomly chosen with respect to a stationary distribution.
The chaotic saddle of Fig. 10.20d is thus a fluid-dynamical example of a snapshot
chaotic saddle (see Sect. 4.6.1). If the nonrandom version of the map is chaotic, the
random map is expected to be chaotic, too, although with different dynamical invari-
ants for the respective chaotic processes. The requirement of topological stability of
the flow is reflected by the feature that the topological entropy of the map is positive
for each fixed realization of the parameter ensemble, and does not exhibit significant
fluctuations.
For incompressible two-dimensional open flows, their random-map advection
models are area-preserving two-dimensional maps with open phase spaces. As
shown in Appendix B (see also [352, 544]), random maps possess well-defined
370 10 Chaotic Advection in Fluid Flows

dynamical averages. Moreover, the dimension formulas (e.g., (4.51)) remain valid.
These relations also imply that if a tracer pattern (such as one in Fig. 10.21) is found
to have a well-defined fractal or information dimension, it cannot depend on the
time at which the snapshot is taken, although the actual shapes of the pattern can be
quite different at different times. This indicates that the dynamical-system approach
to characterizing and understanding chaotic advection in aperiodic flows can be
effective.

10.6 Advection in Closed Flows with Leaks

Chaotic advection in closed flows is characterized by space-filling stable and un-

stable manifolds. One can get insight into these foliations by leaking a closed flow
in the sense that whenever a particle enters a preselected region, the leak, it is con-
sidered to have escaped. This process will not alter the fluid flow, and the leaked
advective dynamics is transiently chaotic. Points never escaping the complement
of the leak, both forward and backward in time, form a chaotic saddle. The mani-
folds of this saddle are subsets of the closed system’s manifolds, since the advective
dynamics outside the leak are exactly the same as in the closed system.
The idea of leaked advection dates back to Pierrehumbert [598], who proposed
the resetting mechanism as a simple model of certain chemical reactions. Whenever
a tracer enters a preselected region, a given property of the tracer, such as the con-
centration, is reset to a value associated with that region, regardless of its previous
value. This mimics the situation in which dye is introduced by diffusion from a solid
surface and is maintained at the saturation concentration in a diffusive boundary
layer. In the case of two or more resetting regions, tracers of different concentrations
come close to each other along fractal-like boundaries. An example of two concen-
trations is shown in Fig. 10.22. Note that the black (white) dots can be obtained by
iterating trajectories backward in time until they reach the black (white) resetting
band and then coloring the corresponding initial locations. This is exactly the dye
exit boundary problem treated in Sect. 10.4 under the time-reversed dynamics. Since
such boundaries contain the stable manifold of a chaotic saddle, the concentration
boundaries generated by resetting contain the unstable manifold of the leaked advec-
tive dynamics followed in forward time. The resetting pattern is thus a fingerprint of
the chaotic saddle underlying the leaked dynamics, where the leak is the union of the
resetting regions. The resetting patterns are thus the analogues of the dye-boundary
patterns in open flows. The basic difference is that the resetting regions are set at
will, and the advective dynamics is thus artificially opened. General properties of
leaked systems (Sect. 2.7) hold for the resetting problem.
A direct analogue of dye exit boundaries in the context of leaked flows is the
following. Consider a subregion, a region of observation, of the closed flow prob-
lem that is to be partitioned according to what happens to the tracers outside this
region. Define target regions outside the region of observation and color the initial
locations according to the first arrival at these regions. The problem is relevant to
10.6 Advection in Closed Flows with Leaks 371

Fig. 10.22 Tracer

distribution in the presence
of resetting in a closed
time-periodic flow. The upper
(lower) resetting band sets a
concentration value marked
by black (white). These two
concentrations come
arbitrarily close to each other
along a filamentary pattern
[542] (copyright 2000, the
American Institute of
Physics)

Fig. 10.23 Tracer initial

conditions in a central square
of the problem of the
aperiodically wind-driven
lake in Fig. 10.18, colored
according to which of the
target regions, bands along
the shores (also colored), any
specific tracer is advected.
The initial conditions of
tracers that do not reach any
of the bands along the shores
over the period of observation
are left blank. (Picture by
M. Pattantyús-Ábrahám.)

the prevention of environmental pollution. In particular, imagine that a pollutant is

released in an extended region within the flow. By determining to which boundary
region the different pollutant particles are advected, one can learn which parts along
the boundary are most likely to be affected by the pollution release.
Figure 10.23 shows an example from the problem of the wind-driven lake in
Sect. 10.5.1. The region of observation is a large rectangle in the middle of the lake.
Each point is distinguished according to which of the bands (of width 100 m) along
the four shores will be reached by the particle first. The result indicates that the
pollutants released in the rectangle are most dangerous for the southern shore. The
western shore is somewhat less polluted than the eastern one. The northern coast is
hardly affected. The boundaries between different colors contain the stable mani-
fold of the chaotic saddle residing in the region of observation. A comparison with
Fig. 10.18 shows that the ridges of the finite-size Lyapunov-exponent field run paral-
lel to these boundaries in accordance to the observations that large future stretchings
belong to the stable manifold and that the stable manifold of the leaked system is a
subset of the closed system’s stable manifold.
372 10 Chaotic Advection in Fluid Flows

Fig. 10.24 Snapshots of the tracer distribution in a domain within two convection cells in a nu-
merical simulation of Earth’s mantle. The region of observation is the dotted subregion in panel
(a). The actual positions of the nonescaped tracers in white and the initial positions in black after
2, 6, 14, and 17 average overturns are shown in panels (b), (c), (d), and (e), respectively. Color cod-
ing corresponds to dimensionless temperature differences (dark blue: 0, red: 1) [689] (copyright
2007, the American Institute of Physics)

The leaking method has been applied to visualizing the foliations of three-
dimensional flows [793], and it also has applications in geophysics [687,689]. As an
example we present the results for the convective flow in Earth’s mantle [689],
which is strongly aperiodic. The region of observation is the dotted rectangle within
the flow, as shown in Fig. 10.24. A large number of particles are initiated in this
region, and trajectories staying within the region are followed up to dimension-
less time t. The initial (final) points of these trajectories are denoted black (white).
Both the black and the white regions become more and more ramified with time.
The black points converge to a stationary, fractal-like pattern, which is the stable
manifold of the saddle in the region of observation. The white points trace out a
moving foliation, the unstable manifold, at time t. In fact, these manifolds belong
to a snapshot chaotic saddle. The loss of detail is due, for large times, to the escape
10.7 Advection of Finite-Size Particles 373

of tracers. It can be seen, nevertheless, that in certain small regions the particle
distribution remains dense, which can be interpreted as a sign of poor stirring due
to the existence of elliptic Lagrangian coherent structures. Time is measured here
in average overturn times within a convective cell. Due to the high viscosity of
the mantle material, this time is about 4 · 108 years. The full lifetime of Earth is
thus about 10 overturns, and the time instant in panel (d) of the figure corresponds
approximately to the age of Earth. This is consistent with the observation that the
current state of Earth’s mantle is heterogeneous, i.e., not (yet) uniformly stirred.

10.7 Advection of Finite-Size Particles

So far, we have assumed that particles are passively advected by the fluid in that
the particle motion follows exactly that of the flow, which can occur for idealized
particles of zero size and zero inertia. Finite-size particles, however, cannot adjust
their velocities instantaneously to variations of the flow velocity. They are therefore
also called inertial particles. As a result, in general, the particle velocity v = ṙ ≡
dr/dt will differ from the fluid velocity u.
The equation of motion for the dynamics of a small rigid spherical particle of
radius a, density ρ p , and mass m p in a fluid of density ρ f is given by [32, 506]

d 1 d
m p v̇ = m f u(r(t),t) − m f [v − u(r(t),t)]
dt 2 dt
−6π aρ f ν [v(t) − u(r(t),t)] + (m p − m f )g, (10.27)

where m f is the mass of the fluid displaced by the particle, ν is the kinematic viscos-
ity, and g is the gravity vector. The first term on the right-hand side is the acceleration
of the fluid element at position r(t) and at time t, which represents the force exerted
on the particle by the undisturbed fluid. The derivative du/dt = ∂ u/∂ t + (u · ∇)u is
the total hydrodynamical derivative taken along the trajectory of a fluid element. The
second, the third, and the fourth terms represent the added-mass effect, the Stokes
drag, and the buoyancy-reduced weight, respectively. The so-called history term
[506, 522] is neglected here, which models the slow diffusion of vorticity. Equation
(10.27) is valid for low particle Reynolds number Re p = 2a | v − u | /ν when the
drag is proportional to the velocity difference. This requires that the initial velocity
difference be small as well.
Using dimensionless variables defined by
L
r → rL, v → vU, u → uU, t→ t,
U
where L and U are the typical length and velocity scales of the flow, we obtain the
following dimensionless equation of motion:
3R gL 3 d
r̈(t) = A [u(r(t),t) − ṙ(t)] + 1 − n + R u(r(t),t). (10.28)
2 U2 2 dt
374 10 Chaotic Advection in Fluid Flows

In (10.28), n is a vertical unit vector pointing downward, and the dimensionless

parameters are
9 Lν 2ρ f
A= R, R = , (10.29)
2 Ua2 2ρ p + ρ f

where A is the inertial parameter, which can be interpreted as a dimensionless relax-

ation rate toward the flow velocity. The limit A → ∞ corresponds to the case of point
particles of zero inertia, where the passive advection equation (10.1) holds. The pa-
rameter R is the density ratio, where R < 2/3 corresponds to aerosols (heavier than
the fluid), and R > 2/3 to bubbles (lighter than the fluid).
The general inertial dynamics (10.28) possesses a four-dimensional phase-space
(x, y, vx , vy ) even for planar stationary flows. What can be seen in the fluid is a projec-
tion of the phase-space objects onto the configuration space. The inertial dynamics
is dissipative even for incompressible flows, and the phase-space volume contracts
at the positive rate of 2A. The dissipative character of the inertial dynamics implies
that attractors can arise in the phase space. Indeed, the existence of attractors, reg-
ular or chaotic, corresponds to accumulation of particles in the configuration space.
This is the phenomenon of preferential concentration, which occurs often in reality.
In general, the dynamics of inertial particles are relevant to a number of phenom-
ena ranging from sedimentation [845–847] and cloud physics [233] to engineering
[522] and environmental applications like hurricanes [673] (for a review see [115]).
In what follows we shall focus on the question of how the dynamics of advective
particles change in open flows as a result of inertia.
In the flow model of the von Kármán vortex street of Sect. 10.2, attractors can be
formed in the bubble regime 2/3 < R < 2. Light particles can thus become trapped
in the wake forever. For example, for R = 1.47 and A = 30, there are three attrac-
tors [50, 53, 197]: two chaotic and one at x = ∞. The chaotic attractors are located
near the cylinder (but not stuck on it): one in the upper half-plane y > 0 and another
in the lower half-plane y < 0. The basins of these attractors can be calculated by
distributing a large number of particles in a region overlapping with the cylinder,
setting the initial velocities equal to the flow velocity and then computing toward
which attractor every initial particle is attracted. Figure 10.25 shows an example, a
two-dimensional slice of the basin structure in the full phase space. It can be seen
that near the cylinder, the basin boundaries among the three attractors are appar-
ently fractal.2 It is not only the dynamics on the attractor that is chaotic; there are
actually dissipative chaotic saddles on the basin boundaries, and they ensure that the
approaches toward the attractors, including the escape from the wake (the approach
toward the attractor at x = ∞) are transiently chaotic processes.
Chaotic saddles of the inertial dynamics also exist in parameter regions where
attractors are not present. A systematic investigation of the escape rate κ (A) from

2 Note that Fig. 10.25 represents a plot of initial conditions. For both ideal (Hamiltonian) and in-

ertial particles, those with long lifetimes belong to the stable foliation of the nonattracting chaotic
set. It is known that for a general Hamiltonian system, under weak dissipation, the stable foliations
are converted into the basin boundaries between the coexisting attractors [534], which are fractals.
10.7 Advection of Finite-Size Particles 375

y
0

−1

−2 −1 0 x 1

Fig. 10.25 Finite-size particle (bubble) advection in the von Kármán flow for particle parameters
A = 30 and R = 1.47 (g = 0). Basins of attraction of two chaotic attractors (light blue and yellow,
respectively) in the plane of the fluid at t = 0.3mod1 are shown. The blank region denotes the basin
of the attractor at x = ∞ [197] (copyright 2003, the American Physical Society)

Fig. 10.26 Escape rate as a function of the inertial parameter in the bubble regime (R = 1.7). The
horizontal line is the escape rate for passive tracers [50] (copyright 2003, the American Physical
Society)

the hyperbolic components of these saddles shows that the escape rate is below the
escape rate of fluid parcels or passive tracers in the full range A > 12, as shown
in Fig. 10.26. This indicates that bubbles spend much more time in the wake than
fluid particles. In the interval 14 < A < 45, the escape rate vanishes, indicating the
presence of attractors. For A between 33 and 45 these attractors are chaotic. Beyond
45 the escape rate is positive, and it approaches for large A the value of ideal tracers.
The tendency is opposite for aerosol particles. The escape rate is above the escape
rate of fluid particles for any value of A, i.e., heavy particles spend much less time
in the wake than fluid particles (Fig. 10.27).
376 10 Chaotic Advection in Fluid Flows

Fig. 10.27 Escape rate as a function of the inertial parameter in the aerosol regime (R = 0.5). The
horizontal line is the escape rate for passive tracers [50] (copyright 2003, the American Physical
Society)

A qualitative argument to explain why bubbles tend the form attractors, or escape
much more slowly than aerosols, is as follows. Particles advected by the fluid are
typically subject to the influence of local vortices. The centrifugal force on a particle
comoving with a vortex is proportional to the density difference ρ p − ρ f . For heavy
particles this force pushes particles outward, but for light ones it attracts particles
toward the vortex center. The presence of this “anticentrifugal” force is the main
reason for the existence of bubble attractors.
An interesting approach to the advection of finite-size particles was due to Haller
and coworkers [308, 672]. They show that for small particle sizes, i.e., for large in-
ertial parameters A  1, the dynamics of a finite-size particle can be approximated
by the dynamics on a low-dimensional inertial manifold, which can be calculated
explicitly from a given velocity field. After a short transient time the equation of
motion of such small particles can be well approximated by a first-order differen-
tial equation, called the inertial equation. The advantage of this equation becomes
apparent when one traces particles in backward time. Finding a localized source of
particle release is often of central relevance. Such a source-inversion problem ap-
pears, for example, in locating a source of air-transported contaminant particles. The
approach based on the time-reversed integration of (10.28) leads to an unavoidable
numerical instability due to an exponential growth of the type exp(At). In contrast,
the inertial equation is free from such instabilities. It can be solved readily in back-
ward time, too, and this procedure provides with good accuracy the initial spatial
coordinates of inertial particles, as demonstrated the example of bubbles in the von
Kármán flow [308, 309] and by other advection problems [674, 759].
In spite of the repelling centrifugal force for heavy particles, Vilela and Motter
showed that aerosols can also be trapped by open flows [799]. Such aerosol at-
tractors can arise as a result of the interactions between two or more vortices.
10.8 Reactions in Open Flows 377

Fig. 10.28 For finite-size particle advection in the leapfrogging vortex problem for parameters
A = 50 and R = 0 (g = 0), basins of attraction of four regular attractors (red crosses) at time
tmodT = 0.8. Black arrows mark the fluid velocity vectors at this instant. The closed red curves
display the attracting orbits in the comoving frame, projected onto the plane of the fluid [799]
(copyright 2007, the American Physical Society)

An example is shown in Fig. 10.28, where periodic aerosol attractors in the

leapfrogging problem of Sect. 10.3.2 are demonstrated. Note that the structure
of the basin boundary is qualitatively similar to the stable manifold of the chaotic
saddle in the passive tracer problem.
The case of neutrally buoyant particles, R = 2/3, is special [36, 758]. In an
open chaotic flow, the effect of inertia is to cause a dispersion of particles around
the chaotic saddle that exists for perfect tracers [798]. The main result is that iner-
tia causes the fractal structure of the chaotic set to be lost in the configuration space,
so, below a certain scale determined by the inertial parameter, the spatial distribu-
tion becomes smooth. In the slow manifold approach of Haller and coworkers, the
slow dynamics coincides with that of infinitesimally small ideal tracers: ṙ(t) = u.
Therefore the particle dynamics should synchronize with the Lagrangian tracer mo-
tion. It was shown [672], however, that the slow manifold has domains that repel
nearby trajectories, which explains the numerical findings in [36, 798].

10.8 Reactions in Open Flows

Chemical or biological reactions often occur in fluid flows. Such processes are of
interest in fields ranging from chemical reactors and combustion to atmospheric
chemistry and the population dynamics of plankton. We shall present the basic
378 10 Chaotic Advection in Fluid Flows

understanding of reaction outcomes in the theoretical framework of transient chaos.

Without loss of generality, we restrict our discussion to two-dimensional flows
and specifically consider autocatalytic reactions as a pedagogical model, which in
fact occur commonly in nature [223]. They are generic models of infection-like
processes. To be concrete, the model describes the penetration of a stable phase B
into an unstable one A. The reaction scheme is

A + B → 2 B. (10.30)

Consider discrete particles advected by the flow, which upon collision can change
their properties [521, 786] but do not affect the underlying flow. Component A is
assumed to be uniformly distributed, and hence a single seed of B is sufficient to
trigger an extended reaction. In a closed container, B spreads until it eventually
takes over all the available space. In an open flow, however, particles are transported
away by the outflow, and in a fixed region of observation, product (B) particles
can accumulate in a nontrivial pattern. If two particles of different types come
closer than a threshold distance, they then react. This distance is called the reaction
range, denoted by σ . It is convenient to assume that reactions occur at certain
times only, and that they take place instantaneously. The time interval between
two successive reactions is τ . At integer multiples of τ , the autocatalytic activi-
ties of B-particles convert all the A-particles within σ into B particles. During a
period of length τ between two successive reactions, the particles are passively
advected by the flow of given velocity field u. In the limit τ → 0, the reaction
velocity σ /τ = v remains finite. In what follows we shall study the continuous-time
model.

10.8.1 Heuristic Theory

Consider a droplet of B-particles that initially intersects the stable manifold of

the chaotic saddle. As we have seen (e.g., Sects. 10.2 and 10.4), the material of the
droplet will be distributed in the stirring region along filaments, in bands along the
unstable manifold. In the presence of reactions, the loss due to escape is countered
by the production of new B-particles. As a result, these particles will be distributed
in bands of finite width along the branches of the unstable manifold. The widths
are approximately the same along all branches, and we let δ (t) denote the average
B-bandwidth at time t.
There are two mechanisms for δ (t) to change. Firstly, it decreases at the rate
−λ1 δ . This is due to the exponential stretching along the unstable manifold with an
average Lyapunov exponent λ1 of the passive advective dynamics, which is accom-
panied, because of incompressibility, by a contraction of strength −λ1 . Secondly,
the bandwidth increases due to the reaction, and the rate of increase is given by 2v,
since the reaction front propagates on both sides of the band. We thus have

δ̇ = −λ1 δ + 2v. (10.31)

10.8 Reactions in Open Flows 379

On any filamentary segment, infinitely many other segments accumulate, since they
form a fractal set. The total B-distribution, covering the unstable manifold, appears
thus to be a fractal on length scales above δ (t), but it is two-dimensional below this
scale.
Consider a region of observation of linear size L that contains the B-bands of
average width δ (t) about the unstable manifold. The smallest box size where frac-
tal scaling can be observed is the average bandwidth ε = δ . The number of boxes
of size ε needed to cover the populated filaments in the region of observation is
N(δ ) ∼ δ −Du,1 , where Du,1 is the information dimension of the unstable manifold.
The distribution of material B within the band is continuous, and the density is ap-
proximately the same in each band. The total area covered by material B in the
observation region is therefore N(δ ) times the area δ 2 of a single box, δ 2 N(δ ),
which is proportional to the number concentration c of the B-particles in the total
region of observation:
δ 2−Du,1
c∼ . (10.32)
L

The time derivative of the total number of B-particles is

ċ ∼ LDu,1 −2 (2 − Du,1)δ 1−Du,1 δ̇ . (10.33)

By substituting δ̇ from (10.31), we obtain, after using the Kantz–Grassberger for-

mula (2.76),
v −β
ċ = −κ c + qκ c , (10.34)
λ1 L
where q is a dimensionless factor, and the exponent β is a constant:

Du,1 − 1
β≡ . (10.35)
2 − Du,1

The first term on the right-hand side of (10.34) describes the escape of B-particles
due to the outflow. The second term is the production term, that is, the rate of produc-
tion of B due to the autocatalytic reaction. The exponent −β in the production term
is always negative, since the dimension of the unstable manifold in open flows satis-
fies the relation 1 < Du,1 < 2. This implies that the smaller the number of B-particles,
the greater the production, and the productivity diverges as c approaches zero.
The basic rate equation (10.34) contains a single chemical parameter, the reac-
tion velocity. All other parameters of the equation are parameters of the passive
advection problem in open flows. It is striking to see that the quantities λ1 and
Du,1 (or κ ) associated with transient chaos play an essential role in the chemical
dynamics, indicating that reactions in open flows can have different outcomes from
those in well stirred containers.
Equation (10.34) describes the competition of two effects: outflow and produc-
tion. As a result of the balance between these factors, a steady state sets in after
380 10 Chaotic Advection in Fluid Flows

1 t=0 1 t=0.4
0 0
y

y
−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=0.6 1 t=1.0

0 0
y

y
−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=1.4 1 t=1.6
0 0
y

−1 y −1
−2 0 2 4 6 −2 0 2 4 6
x x
1 t=2.0 1 t=10

0 0
y

−1 −1
−2 0 2 4 6 −2 0 2 4 6
x x

Fig. 10.29 Evolution of a chemical steady state of B-particles in an autocatalytic reaction taking
place in the von Kármán vortex flow. After a few time units, the pattern no longer changes on a
stroboscopic map [774] (copyright 2000, the American Institute of Physics)

sufficiently long time in the observational region. This can be seen in numerical
simulations, as shown in Fig. 10.29, where a droplet of B-particles in the sea of
A-particles is stretched and folded as a passive droplet initially, but the B-area
increases subsequently. In fact, an accumulation takes place along the unstable
manifold. However, in contrast to passive advection (e.g., Fig. 10.5), here the popu-
lation is not emptied but approaches a steady state. The distribution is in dynamical
equilibrium since there is a flux of B-particles permanently leaving the region, but
exactly the same amount of new chemicals is produced on average. The process is
thus dissipative, converging toward a chemical attractor. In continuous time, the at-
tractor is a limit cycle in time-periodic flows. In this sense we can say that reactions
become synchronized to the flow.
This heuristic theory has been generalized to cover different types of reac-
tions in open flows, for example, acid–base reactions [392], reactions with inertial
particles [537, 550, 551], flames and other fronts [410, 742], and reactions in three-
dimensional flows [165]. The basic ideas also contribute to a better understanding of
the properties of chemical transients toward equilibrium in closed flows [390, 395,
396]. Further details can be found in the reviews [545, 771].
10.8 Reactions in Open Flows 381

10.8.2 Biological Activities

The temporal evolution of populations can be considered as a reaction between in-

dividuals and the available resources (food). In this context, the singular production
term in the rate equation (10.34) reflects the principle of advantage of rarity. In par-
ticular, a state with a smaller number of organisms will grow faster than another one
with more individuals. This is due to the fact that a smaller population covers the
unstable manifold with a smaller area but with a longer perimeter. One consequence
of the advantage-of-rarity principle is the existence of a nontrivial steady state in
which the population never dies out.
A particularly important example is plankton dynamics [391, 502, 543, 545],
where several competing species can often coexist. Traditional theoretical and
empirical investigations show that the number of coexisting species in a given
ecosystem is limited by the number of different resources for which the species
are competing. In a well-stirred environment, only those species are able to survive
which are best adapted to the use of at least one of the limiting resources [223]. This
result is based on the hypothesis that the competing species are homogeneously
stirred in their habitat. Observation shows, however, that a large number of species
can coexist in a single ecosystem while competing for the same limiting resources.
This contradiction of competitive exclusion with observations is common in the
plankton communities, and is known as the paradox of plankton.
In plankton dynamics, the condition of open flow is often fulfilled, since a fluid
element escaping the region of observation has no chance to return within the bio-
logical time scale. Thus the advantage of rarity principle applies, a mechanism that
favors the survival of all the species that are about to become extinct. Figure 10.30
provides numerical evidence for the coexistence of two species in the presence of a
single resource in the wake of a cylinder. The theory of reactions in open flows thus
provides one possible resolution to the paradox of plankton. The theory can be used
to study other ecosystem-related problems, such as the transport of plankton across
island wakes [664, 666], plankton blooms in vortices [665], prebiotic evolution
[682], and metabolic networks [393].

10.8.3 Reactions in Open Aperiodic Flows

The observations in Sect. 10.5.2 indicate that the particles of a B-droplet are dis-
tributed in bands along the unstable manifold of a chaotic saddle. The key feature
in deriving (10.34) is that the dimension of the saddle is time-independent. It is thus
plausible that the picture would essentially be the same for aperiodic flows. In par-
ticular, the relevant dimension is that of the aperiodic flow’s unstable manifold. It
can indeed be shown [397] that for random maps with weak parameter variations,
(10.34) remains valid with the exponent β determined the same way by the dimen-
sion as in (10.35), except that Du,1 is replaced by the dimension of the unstable
manifold of the snapshot chaotic saddle. Numerical simulations show that chemical
382 10 Chaotic Advection in Fluid Flows

−0.3

−0.4

−0.5
y

−0.6

−0.7

−0.8

0.7 0.8 0.9 1 1.1

x

Fig. 10.30 Coexistence of two species in the von Kármán vortex street flow. Shown is population
distribution close to the cylinder, after a steady state has settled in. The two species are green
and red. Resource material is white [391] (with kind permission from the National Academy of
Sciences of the United States)

Fig. 10.31 Autocatalytic reaction in the aperiodic four-vortex flow. The main figure shows the
B-distribution at dimensionless time 14. Right inset: Number of B-particles versus time. Left inset:
Fractal scaling coinciding with that of the unstable manifold of the snapshot chaotic saddle in the
reaction-free problem [397] (copyright 2004, the American Physical Society)
10.8 Reactions in Open Flows 383

processes can reach an asymptotic state in such flows, too. An example is given in
Fig. 10.31, where the B-distribution of an autocatalytic reaction in the four-vortex
problem treated in Sect. 10.5.2 is shown at dimensionless time t = 14. By this time
the system is in a “steady state.” Because of the flow’s aperiodicity, the number
of B-particles does not stay constant in time, but fluctuates randomly, as shown in
the right inset of Fig. 10.31. It can also be seen that the B-distribution is filamen-
tary. The product particles accumulate along the unstable manifold of the snapshot
chaotic saddle in this aperiodic flow. The box-counting dimension of the product is
therefore the same as that of the unstable manifold in the passive advective dynamics
(left inset of Fig. 10.31).
This observation shows that the appearance of transiently chaotic advective dy-
namics in reactive systems is a robust phenomenon.
Chapter 11
Controlling Transient Chaos and Applications

Besides the occurrence of chaos in a large variety of natural processes, chaos may
also occur because one may wish to design a physical, biological, or chemical ex-
periment, or to project an industrial plant to behave in a chaotic manner. That chaos
may indeed be desirable is further evidenced by the fact that it can be controlled
using small perturbation of some accessible parameter or dynamical variable of the
system.
The key ingredient for the control of chaos is the observation that any chaotic set
has embedded within it a large number of unstable periodic orbits of low periods.
Because of ergodicity, the trajectory visits or accesses the neighborhood of each
one of these periodic orbits. Some of these periodic orbits may correspond to de-
sired system performance according to some criterion. The second ingredient is the
realization that chaos, while signifying sensitive dependence on small changes to
the current state, thereby rendering the system state unpredictable over long times,
implies that the system’s behavior can be altered using small perturbations. The ac-
cessibility of the chaotic system to many different periodic orbits combined with
its sensitivity to small perturbations allows for the control and manipulation of the
chaotic process. Specifically, the Ott–Grebogi–Yorke (OGY) approach [566] is as
follows. One first determines some of the unstable low-period periodic orbits that
are embedded in the chaotic attractor. One then examines the locations and the sta-
bilities of these orbits and chooses one that yields the desired system performance.
Finally, one applies small controls to stabilize this desired periodic orbit. A partic-
ularly appealing feature of the OGY approach is that control can be achieved based
on data using nonlinear time series analysis for the observation and understand-
ing of the system. This is important, since chaotic systems can be complicated and
equations of the process are often unknown.
Since the seminal paper on the OGY paradigm there has been a tremendous
amount of research on controlling chaos. The focus of this chapter is on control-
ling transient chaos. We shall present the basic idea and methodology of controlling
the dynamics on nonattracting chaotic sets. The existence of transient chaos makes
a new type of control possible, i.e., to convert transient chaos into permanent chaos
via small and infrequent perturbations. The methods of maintaining chaos will
be reviewed. We will then consider applications: voltage collapse and prevention,

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 385
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 11,

c Springer Science+Business Media, LLC 2011
386 11 Controlling Transient Chaos and Applications

and how to prevent species extinction. An algorithm for maintaining chaos in the
presence of noise will also be presented, as well as a method of encoding digital
information using transient chaos.

11.1 Controlling Transient Chaos: General Introduction

11.1.1 Basic Idea and Method

It is possible to control motion on a nonattracting chaotic set to convert transiently

chaotic dynamics into periodic dynamics by stabilizing one of the infinite number
of unstable periodic orbits embedded in the set. The feature of this type of control
is that it stabilizes an orbit that is not on the actual attractor of the system. One thus
selects an atypical behavior that cannot be revealed by a long-time observation of
the unperturbed motion. This type of control is effectively stabilizing a metastable
state. To be specific, we shall discuss the control method in the OGY paradigm [566]
because it leads to an algorithm capable of carrying out the finest possible selection
of the target orbit to be stabilized and applying the weakest possible perturbation.
Other methods [79,123,667,694,785], e.g., the delayed feedback-control method of
Pyragas [617], are also applicable. To be concrete, we focus on invertible dynamical
systems.
To achieve control of transient chaos, one has to use an ensemble of trajectories
[767] because any randomly chosen initial point leads to a trajectory that escapes
any neighborhood of the saddle in finite time. This ensemble is typically chosen
to start from a compact region having intersections with the stable manifold of the
chaotic saddle. One also selects a target region I containing a predetermined unsta-
ble periodic orbit on the chaotic saddle. Then the ensemble of trajectories start to
evolve, and one waits until a trajectory enters the target region to activate control.
The control perturbation is adjusted with time so as to stabilize the periodic orbit.
Only small local perturbations are allowed, smaller in size than some value δ , the
maximum allowed perturbation. In general, δ is proportional to the linear extension
of the target region [566].
To illustrate the OGY method, we shall use a two-dimensional map with p as an
externally accessible control parameter [767]. We restrict parameter perturbations
to be small, i.e., |p − p| < δ , where p is some nominal parameter value, and δ 1
defines the range of parameter variation. We wish to program the parameter p so that
a chaotic trajectory is stabilized when it enters an ε -neighborhood of the target peri-
odic orbit. Without loss of generality, we assume that the target orbit is an unstable
fixed point embedded in the chaotic saddle, denoted by xF (p). The location of the
fixed point in the phase space depends on the control parameter p. Upon application
of a small perturbation Δp, we have p = p + Δp. Since Δp is small, we expect xF (p)
to be close to xF (p), and write

xF (p) ≈ xF (p) + gΔp, (11.1)

11.1 Controlling Transient Chaos: General Introduction 387

where g is a vector given by


∂ xF  xF (p) − xF (p)
g≡ ≈ . (11.2)
∂ p  p=p Δp

The vector g needs to be determined before a control law can be applied to stabiliz-
ing the fixed point xF (p).
To formulate a control law, we make use of the fact that the dynamics of any
smooth nonlinear system is approximately linear in a small neighborhood of a fixed
point. Thus, near xF (p), we have

xn+1 − xF (p) ≈ J[xF (p)] · (xn − xF (p)) , (11.3)

where J[xF (p)] is the 2 × 2 derivative matrix of the map f(x, p) evaluated at the fixed
point xF (p), defined as
 
∂ f  ∂ J 
J[xF (p)] = ≈ J[xF (p)] + Δp. (11.4)
∂ x xF (p) ∂ p  p=p

Note that Δp ∼ ε and |xn − xF (p)| ∼ ε , where ε is the size of the small neighborhood
in which the linear approximation (11.3) is valid. Substituting (11.1) and (11.4) into
(11.3), and keeping only terms that are of first order in ε , we obtain

xn+1 − xF (p) ≈ gΔp + J[xF(p)] · [xn − xF (p) − gΔp]. (11.5)

Since xF (p) is embedded in the chaotic saddle, it has one stable and one unstable
direction. Let es and eu be the stable and the unstable unit eigenvectors at xF (p),
respectively, and let fs and fu be two unit vectors that satisfy fs ·es = fu ·eu = 1 and fs ·
eu = fu · es = 0 (relations by which the vectors fs and fu can be determined from the
eigenvectors es and eu ), which are the contravariant basis vectors associated with the
eigenspaces es and eu [566]. The derivative matrix J[xF (p)] can then be written as

J[xF (p)] = λu eu fu + λses fs , (11.6)

where λs and λu are the stable and the unstable eigenvalues in the eigendirections
es and eu , respectively.
When the trajectory point xn falls into the small ε -neighborhood of the desired
fixed point xF (p), (11.3) becomes valid. A small parameter perturbation Δpn can
there be applied at time n to make the fixed point shift slightly so that at the next
iteration (n + 1), xn+1 falls on the stable direction of xF (p):

fu · [xn+1 − xF (p)] = 0. (11.7)

388 11 Controlling Transient Chaos and Applications

For sufficiently small xn − xF (p) we can substitute (11.5) into (11.7) to obtain Δpn =
cn , where cn is given by
λu fu · [xn − xF (p)]
cn = ≡ C · [xn − xF (p)] . (11.8)
(λu − 1)fu · g
We assume in the above that the generic condition g · fu = 0 is satisfied, so cn can be
calculated. Once xn+1 falls on the stable direction of xF (p), we can set the control
perturbation to zero, and the trajectory for subsequent time will approach the fixed
point at the geometrical rate λs .
The considerations above apply only to a local small neighborhood of xF (p).
Globally, one can specify the parameter perturbation Δpn by setting Δpn = 0 if |cn |
is too large, since the range of the parameter perturbation is limited to be small.
Thus, practically, we have

cn , if |cn | < δ ,
Δpn = (11.9)
0, if |cn | ≥ δ .
In this way, in the definition of cn in (11.8), it is unnecessary to restrict the quantity
|xn − xF (p)| to be small.
Figure 11.1 shows an example [767] of controlling a fixed point on the Hénon
chaotic saddle in comparison with the uncontrolled trajectory. We see that the con-
trolled motion is not a part of the asymptotic dynamics [605, 606, 792].

11.1.2 Scaling Laws Associated with Control

There are scaling laws characterizing the ensemble of trajectories in the limit of
a small allowed perturbation δ . Many of the trajectories approach the asymptotic
attractor before entering the target region enclosing the periodic orbit to be stabilized
on the chaotic saddle. Short transients are therefore irrelevant for the controlling
process, but trajectories with lifetimes significantly larger than 1/κ are unprobable.
As a result, the average time τc needed to achieve control is independent of δ and is
limited from above by the chaotic lifetime τ ≈ 1/κ for some values of δ :

τc ≤ 1/κ . (11.10)

Because of the escape, only a small portion of all trajectories can be controlled.
When the target region is a disk, the number of controlled trajectories N(δ ) de-
creases with decreasing δ according to the power law [767]:

N(δ ) ∼ δ γ (κ ) , (11.11)

where the exponent γ (κ ) depends on the escape rate of the saddle. The number of
controlled trajectories is proportional to the c-measure μc (I) of the target region I.
The c-measure is smooth along the unstable direction, so the measure μc of a region
11.1 Controlling Transient Chaos: General Introduction 389

Fig. 11.1 For the Hénon map at parameters a = 1.45, b = 0.2, where the attractor is a period-5
cycle, (a) a transiently chaotic time series. The trajectory ceases to be chaotic at about the 38th
time step, where it enters the neighborhood of the period-5 attractor. (b) Controlled signal started
from the same initial point. The Hénon map has the form given in the caption of Fig. 5.9 with
a = 1.45 + pn , J = −b, where the maximum allowed perturbation is δ = 0.1. The fixed point is at
xF = yF = 0.868858. Control sets in at the 26th step, and the fixed point on the saddle is stabilized
[767] (with kind permission from Institute of Physics)

of size l1 and l2 along the unstable and the stable direction, respectively, scales
according to (2.89). For l1 ∼ l2 ∼ δ , the scaling exponent is given by

γ (κ ) = 1 + α2 (κ ), (11.12)

where α2 (κ ) is the crowding index along the stable direction. In the particular case
in which the target region contains a fixed point, the exponent γ (κ ) is (2.91):

λ1∗ − κ
γ (κ ) = 1 + , (11.13)
| λ2∗ |
390 11 Controlling Transient Chaos and Applications

where λi∗ (i = 1, 2) are the local Lyapunov exponents of the fixed point to be
∗ = ln | λ
stabilized (λ1(2) u(s) |).
When applying the OGY method to controlling permanent chaos, the scaling
properties of an ensemble of N0 trajectories are different. The average time τc
needed to achieve control is a function of the linear size of the target region, which
is proportional to the maximum allowed perturbation δ . It was pointed out [566]
that τc (δ ) increases algebraically as δ is decreased,

τc (δ ) ∼ δ −γ , for δ 1, (11.14)

where γ > 0 is a characteristic exponent. This scaling law shows that the dynam-
ics of reaching the target region is itself a kind of transient chaos. The process of
control can be interpreted as leaking a closed chaotic system at the target region
(Sect. 2.7). The time needed to achieve control is thus the average lifetime of the in-
variant chaotic saddle of the leaked system. For small leak sizes the lifetime scales
as the inverse of the natural measure μ of the leak. We have

γ = 1 + α2 , (11.15)

where α2 = λ1∗ / | λ2∗ | is the crowding index along the stable direction of the
chaotic set in the closed system.1 However, the number N(δ ) of controlled trajec-
tories is fixed, N(δ ) = N0 , since all N0 trajectories of the ensemble are controlled
sooner or later.
The scaling laws in the control of permanent and transient chaos thus appear to
be the two extremes of a general process, where for the former, N(δ ) is constant,
but for the latter, τc (δ ) is constant. There exists a unifying relation between N(δ )
and τc (δ ) that holds in both cases [769]. The key observation is that the number of
controlled trajectories in the entire process is proportional to the average number of
trajectories controlled per unit time multiplied by the average time needed to achieve
control. The average number of trajectories controlled per time step is proportional
to the probability of falling in the target region. For small regions this is proportional
to the c-measure of the target region in the uncontrolled system. Since the latter
scale as δ γ or δ γ (κ ) , the number of trajectories controlled per time step follows the
scaling law
N(δ )
∼ δ γ (κ ) , (11.16)
τc (δ )
which is valid for both permanent (κ = 0) and transient (κ > 0) chaos.

1 For a larger target region, the exponent γ depends strongly on the location of I even if μ (I) is
kept constant [101, 103, 571, 572, 574]. From the general theory of leaked systems (cf. Sect. 2.7),
this can be understood as being due to the complicated overlap of the leak with its preimages.
11.1 Controlling Transient Chaos: General Introduction 391

11.1.3 Remarks

11.1.3.1 Controlling Fractal Basin Boundaries

An immediate application is the control of motions on a fractal basin boundary

[418, 434, 706], which contains a chaotic saddle whose stable manifold constitutes
the boundary. One can then control a desirable periodic orbit in the saddle. By ap-
plying weak control perturbations, a hyperbolic orbit on the basin boundary can be
converted into an attracting orbit [418]. The methodology is potentially important in
applications where periodic driving can result in a catastrophic failure of the system.
A particular example is ship capsizing, where the method of controlling motion on
fractal basin boundaries was computationally tested to prevent chaos-induced ship
capsizing even in cases where the driving due to environmental influences (e.g.,
waves) is not periodic but has a substantially irregular (chaotic) component [195].
An alternative method for steering most trajectories to a desirable attractor is to
build a hierarchy of paths to it and then stabilize trajectories around one of the paths
in the hierarchy [434]. A pronounced improvement in the probability for a random
trajectory to approach a desirable attractor can be achieved when there are fractal
basin boundaries or riddled basins.

11.1.3.2 Controlling Chaotic Scattering

A feature of chaotic saddles in Hamiltonian systems is that they typically contain a

nonhyperbolic component where the local Lyapunov exponents are arbitrarily close
to zero (Sect. 6.4). A problem is to investigate the influence of the nonhyperbolic
component on the control process. If one selects a periodic orbit close to a KAM
surface, the time to achieve control is usually long due to the stickiness effect.
Numerical investigation showed [453] that the average time to achieve control could
be an order of magnitude longer than the average chaotic lifetime on the hyperbolic
component.
In general, controlling a collisional scattering process means stabilizing the inter-
mediate complexes of a reaction that would otherwise be of finite lifetime. Although
KAM surfaces can be important for the controlling process, the qualitative behavior
of the controlled ensemble is similar to that of a fully hyperbolic system.

11.1.3.3 Improved Method of Controlling a Chaotic Saddle

As we have seen, a major difference between stabilizing unstable periodic orbits

embedded in a chaotic attractor and in a chaotic saddle is that for the attractor, the
probability that a chaotic trajectory enters the neighborhood of the desired unstable
periodic orbit is one, while for transient chaos, only a small set of initial conditions
can be controlled, since most trajectories will have already left the chaotic saddle
before entering the neighborhood of the target periodic orbit. An issue is how to
maximize this probability of control of transient chaos. A useful observation is that
there exists a dense chaotic orbit in the saddle that comes arbitrarily close to any
392 11 Controlling Transient Chaos and Applications

target unstable periodic orbit. Such a dense orbit is the complement of the set of
all unstable periodic orbits in the saddle, and can be numerically obtained by the
PIM-triple method (Sect. 1.2.2.4). The probability that a trajectory approaches this
orbit can be significantly larger than the probability that the trajectory enters the
neighborhood of the target unstable periodic orbit, if the reference orbit is long. By
stabilizing a trajectory about the reference orbit first, and then switching to stabilize
it about the target periodic orbit after the trajectory comes close to it, we can increase
substantially the probability that a trajectory can be controlled [447, 461]. This can
indeed be achieved, since there exist stable and unstable directions at each point of
the reference orbit on the chaotic saddle. Hence in principle, controlling a trajectory
near the reference orbit is equivalent to stabilizing a long unstable periodic orbit.
The longer the length of the reference orbit, the larger the probability of controlling
periodic orbits.

11.2 Maintaining Chaos: General Introduction

The conversion of transient chaos into permanent chaos is called chaos maintenance
or preservation, and the basic ideas date back to the work of Yang et al. [840],
Schwartz and Triandaf [697], and Kapitaniak and Brindley [383]. The term partial
control of chaos is also in use [849], since the algorithms do not determine exactly
where the trajectory goes around a nonattracting chaotic set. The practical relevance
of this approach is due to the fact that there are systems that require chaos in order to
function properly. Notable examples are mechanical systems in which the avoidance
of resonance via chaos is desirable [697], advection in fluids where complete stirring
can be achieved only via permanent chaos (cf. Chap. 10), and biological systems in
which the disappearance of chaos may signal pathological phenomena (see point (2)
of Sect. 4.4.3). Under certain conditions, simple regular attractors may appear, and
it is then important to intervene in order to maintain chaos. Later in this chapter we
shall investigate two examples in detail in which maintenance of chaos is useful:
preventing voltage collapse (Sect. 11.3) and species extinction (Sect. 11.4).

11.2.1 Basic Idea

The aim is to intervene the dynamics in such a way as to keep chaotic behavior alive
in situations in which it would naturally be absent. In fact, stabilizing a trajectory
about a reference orbit on a chaotic saddle, which can enhance the probability of
converting a transiently chaotic behavior into a periodic one, as described in the last
subsection, can be considered as an attempt to maintain chaos if the reference orbit is
long. Other types of algorithms are based on the observation that systems exhibiting
transient chaos have special regions in their phase spaces, called loss regions or
escape regions. They are identified by the property that after the orbit enters such
11.2 Maintaining Chaos: General Introduction 393

a region, it immediately ceases its chaotic motion, i.e., it is rapidly drawn to some
simple attractor. Examples of loss regions are the primary escape interval I0 of open
one-dimensional maps (see Fig. 2.1) and the area bounded by the outermost branch
of the chaotic saddle’s unstable manifold and the outermost branch of its stable
manifold (cf. the shaded area AB in Fig. 3.12).
The strategy can be formulated straightforwardly for map f(xn , p), where p de-
notes the parameter whose temporal change will be used to maintain chaos. After
identifying a loss region L, one considers the preimages Lm of this region under the
unperturbed map f(xn , p̄), where p̄ is the nominal parameter value. The set Lm is
thus the set of points mapped onto the loss region in m iterates, and the width of
Lm decreases exponentially along the unstable direction(s) as m increases. Yang et
al. [840] suggested the following approach. Pick a large value M of m and consider
the preimages of the loss region up to level M + 1. If the unperturbed orbit lands in
LM+1 on iterate n, one applies a control parameter pn (different from p̄) in order to
kick the orbit out of LM on the next iterate. Since LM is thin, the required change
Δpn = pn − p̄ is small. After the orbit is kicked out of LM , it is likely to execute a
chaotic motion. Due to the fractal structure of the nonattracting set, the orbit falls
with probability one outside this chaotic set, i.e., in a region Lr with r > M. Long
chaotic sequences are expected if r happens to be much larger than M. After some
time, the orbit falls again in LM+1 when a small control is activated, and so on.
The amount of the control parameter shift Δpn at the nth step can be estimated
by using the sensitivity vector (11.2) evaluated in the loss region and its preimages.
By assuming that this vector is approximately a constant ḡ over these regions, and
using the maximum width dM of region LM , one finds that

dM
Δpn ≈ , (11.17)
| ḡ |

which is a small number for M  1. These ideas were successfully applied to main-
taining chaos in different models [840], and also in an experiment in which the
intermittent signal of a megnetomechanical ribbon [350] was converted into a nonin-
termittent chaotic signal. This means that chaos was maintained on a chaotic saddle
lying outside a marginally stable periodic orbit.

11.2.2 Maintaining Chaos Using a Periodic Saddle Orbit

The method proposed by Schwartz and Triandaf [697] (see also [698]) can be ap-
plied to situations in which the chaotic attractor of an invertible system has been
destroyed in a crisis (Chap. 3). (The parameter whose change leads to crisis is not
necessarily the same as the parameter p that will be used in the control process.)
The system may then exhibit transient chaos until it reaches the periodic saddle
point mediating the crisis. If the trajectory happens to fall on one side of the stable
manifold of this mediating orbit, it directly approaches a periodic attractor. On the
394 11 Controlling Transient Chaos and Applications

other side of the stable manifold, however, it has chances to return to the chaotic
saddle appearing there as the remnant of a former chaotic attractor. We shall call
this side of the manifold the chaotic side. Once the trajectory enters a neighborhood
of the saddle orbit, a small perturbation in parameter p is applied to ensure that the
trajectory falls on the chaotic side in the next step.2 To optimize chaos maintenance,
one can use the distribution of lifetimes to select a target point xtar lying close to the
mediating orbit with a particularly long lifetime, which ensures that perturbations
should be applied only rarely.
In a two-dimensional map, the local dynamics around a saddle point can be ap-
proximated by equations (11.5) and (11.6). Note, however, that the hyperbolic fixed
point xF is now the mediating orbit and not an unstable point inside the saddle as in
Sect. 11.1. The required amount of control Δpn at time instant n when the trajectory
happens to be close to the mediating orbit can be obtained from these equations by
requiring xn+1 = xtar . After a multiplication of (11.5) by fs , one obtains

fs · [λs (xn − xF (p)) − (xtar − xF (p))]

Δpn = . (11.18)
(λs − 1)fs · g

The required control is thus proportional to λs [xn − xF (p) − (xtar − xF (p))].

The method can be extended to higher dimensions and has successfully been
applied by In et al. [351] to maintain chaos in a magnetoelastic ribbon experiment,
as shown in Fig. 1.22. The perturbation leads to permanent chaos in a system in
which the natural attractor would be periodic.

11.2.3 Practical Method of Control

In [185], a practical method was suggested for converting transient chaos into sus-
tained chaos, based on measured time series. In contrast to the situation of chaotic
attractors, these time series consist of short segments of chaotic oscillations ex-
hibiting a number of local maxima and minima. Let en (n = 1, . . . , L) be the set
of extrema (maxima or minima) from one measured segment of one dynamical
variable x(t). In order to detect the underlying dynamics, an ensemble of tran-
sient chaotic trajectories from a large number of random initial conditions can be
used, each yielding a number of points in the en+1 versus en plot. As a crude ap-
proximation, the dynamics of the underlying system can be represented by a map
en+1 = M(en ), where if the underlying dynamics is approximately one-dimensional,
M(e) is a one-dimensional smooth curve. For higher-dimensional dynamics, the plot
M(e) typically exhibits some complicated structure. It is possible to identify regions
of the plane en = (en , en+1 ) in which the chaotic saddle lies, and a loss region where

2 If the perturbation is chosen so that the trajectory falls on the other side, one can speed up the

escape process to the simple attractor and can reduce the average lifetime of chaos [383].
11.3 Voltage Collapse and Prevention 395

escape from the chaotic saddle occurs. Thus, by applying a small perturbation to
an accessible set ({en }) of dynamical variables at a time when the trajectory is in
the escape region, chaotic motion can also be maintained for a finite period of time.
The difference between this approach and that of the previous subsections is that the
region is identified here only in the two-dimensional plane (en , en+1 ), and not in the
full phase space. Because of this, information about target points in this method is
incomplete. The situation can be improved if more dynamical variables are experi-
mentally accessible [185].

11.3 Voltage Collapse and Prevention

We present an example of application of maintaining chaos: voltage collapse in

electrical power systems and prevention. We shall describe a model system and
demonstrate that voltage collapse is typically preceded by transient chaos. A prac-
tical control method will then be discussed to convert transient chaos into sustained
chaos, thereby preventing voltage collapse while at the same time preserving the
natural dynamics of the system.

11.3.1 Modeling Voltage Collapse in Electrical Power Systems

Electrical power systems are essentially nonlinear dynamical systems. Most major
power-system failures in the past were reported to be caused by the dynamic re-
sponse of the system to disturbances [132, 200]. Voltage collapse occurs when the
system is heavily loaded. In such a case, dynamical variables of the system, such as
various voltages, fluctuate randomly for a period of time before collapsing to zero
suddenly, leading to a complete blackout of the system. Due to an ever-increasing
demand for electrical power, there is an interest in operating the power system near
the edge of its stability boundary. As a consequence, the system becomes highly
nonlinear and can exhibit chaotic behaviors. One possible mechanism for voltage
collapse is then as follows. The system operates in a parameter region where there
is a chaotic attractor. A disturbance or a temporal overload causes a shift in a system
parameter so that a boundary crisis occurs, after which the system exhibits transient
chaos, leading to a voltage collapse. To understand the phenomenon of voltage col-
lapse, Dobson and Chiang [132, 200] introduced a model power system consisting
of a generator, an infinite bus, a nonlinear load, and a capacitor in parallel with the
nonlinear load. Subsequently, Wang and Abed pointed out that the presence of the
capacitor could cause an increase in the reactive power demand of the load to almost
practically unreachable values even in normally encountered parameter regimes.
A modified model was proposed [818, 819], which is mathematically described by
the following set of differential equations:
396 11 Controlling Transient Chaos and Applications

δ˙m = ω, (11.19)
0.01464ω̇ = −0.05ω + 1.0 − 5.25V sin (δm − δ ),
−0.03δ̇ = −2.1V 2 + 2.8V + Q(δm , δ ,V ) − 0.3 − Q1,
−0.0765V̇ = 0.84V 2 − 1.204V − 0.03 [P(δm , δ ,V ) − 0.6]
−0.4[Q(δm , δ ,V ) − 0.3 − Q1],

where the dynamical variables δm , ω , δ , and V are from circuit analysis, Q1 , the
load, is a bifurcation parameter, and P(δm , δ ,V ) and Q(δm , δ ,V ) are the real and
reactive powers supplied to the load by the network, which are nonlinear functions
of their variables [818, 819]. A bifurcation analysis indicated [185] that there is a
period-doubling cascade to chaos, and a crisis occurs at Q1c ≈ 2.56037833, after
which the chaotic attractor is converted into a chaotic saddle. The range for the
attractor is relatively small. Suppose the system operates at some value of Q1 before
the crisis. A small change in Q1 can push the system over the crisis where there is
transient chaos. A voltage collapse can then occur. Figure 11.2 shows a time series
V (t) for Q1 = 2.5603784 > Q1c , where V (t) goes to zero suddenly after about 80
time units.
How to prevent voltage collapse? A possible approach is to reduce the load Q1
to bring the system back into the parameter regime where there is an attractor. In
a practical situation, however, it may not be feasible to change the load of an elec-
trical power system in a relatively short time. One viable strategy is then to control
transient chaos.

Fig. 11.2 A typical example of voltage collapse in the power system (11.19) [185] (copyright
1999, the American Physical Society)
11.3 Voltage Collapse and Prevention 397

11.3.2 Example of Control

Figure 11.3a shows the return map obtained from the local minima of V (t) for
Q1 = 2.5603784 > Q1c . There is a primary escape interval below which V (t) goes
to zero quickly, as shown in Fig. 11.3b. The vertical lines denote the regions from
which target points are chosen. The escape interval corresponds to an escape re-
gion on the chaotic saddle. In contrast, before the crisis, there is no such gap in
the return map. To achieve control in the regime of transient chaos, a set of 3,000
target points was selected [185] in the vicinity of the escape interval with long
lifetime. Figure 11.4 shows the lifetime versus the value of local minima. The
plot is not smooth and contains an infinite number of singularities correspond-
ing to points on the stable manifold of the chaotic saddle. This singular structure
renders selection of desired target points possible. Each target point contains the

Fig. 11.3 Return map constructed from the local minima of V (t): (a) after the crisis for Q1 =
2.5603784; and (b) a magnification of part of (a) near the cusp. There is a primary escape interval,
enclosed between lines II and III, through which a trajectory approaches asymptotically the state
with V = 0 (voltage collapse). Two regions to the left (I – II) and to the right (III – IV) of the
gap are the regions from which target points can be chosen for control [185] (copyright 1999, the
American Physical Society)
398 11 Controlling Transient Chaos and Applications

Fig. 11.4 Lifetime versus the local minima of V (t) in the return map. The plot contains an infinite
number of singularities corresponding to points on the stable manifold of the chaotic saddle [185]
(copyright 1999, the American Physical Society)

values of the four dynamical variables in (11.19), although V (t) is always at a

local minimum. The set of target points is then stored for computing the control
perturbation. When an actual trajectory falls into the escape interval, the computer
selects a target point such that the required perturbation to kick the trajectory onto
the target point is minimal. Perturbations can be applied to the dynamical variables
x directly. Or if there is an accessible system parameter p that can be adjusted,
perturbations can be applied to the parameter based on the difference between the
trajectory point in the escaping window and the target point: Δp = (∂ x/∂ p)|target Δx.
In the power-system model (11.19), since all four dynamical variables can be per-
turbed, it is convenient to apply control directly to these variables. An example
of control is shown in Fig. 11.5a, a controlled voltage signal V (t). The required
control perturbations are shown in Fig. 11.5b. In the time interval shown, only
four small perturbations are required to sustain transient chaos. In general, the
average time interval for applying perturbations is approximately the average life-
time of the chaotic saddle. Perturbations are required only when the system drifts
into the regime of transient chaos, since transient chaos is the culprit of voltage
collapse.
A key question in any scheme of controlling transient chaos concerns the proba-
bility of a typical trajectory being controlled.3 Since the system performs normally
before the collapse and since control is activated only when V (t) falls into the escape

3 We address initial conditions only in the original basin of the attractor because, before the col-
lapse, the system performs normally and operates in the precrisis regime. We are not concerned
with initial conditions outside the basin, although they usually yield trajectories leading to V = 0.
A voltage collapse can thus be regarded as a catastrophic event. Our control method is applicable
to preventing this type of catastrophe.
11.4 Maintaining Chaos to Prevent Species Extinction 399

Fig. 11.5 An example of controlling transient chaos to prevent voltage collapse: (a) a controlled
time series V (t); and (b) required control perturbations. Apparently, only infrequent perturbations
are needed to prevent voltage collapse [185] (copyright 1999, the American Physical Society)

interval, almost all trajectories can be controlled [185]. This implies that voltage col-
lapse can be effectively prevented by controlling transient chaos.

11.4 Maintaining Chaos to Prevent Species Extinction

We consider the problem of species extinction in ecological systems, which can

occur as a consequence of deterministic transient chaos even in the absence of
external disturbances. Controlling transient chaos by applying small, ecologically
feasible perturbations to the populations at appropriate but rare times thus provides
a possibility for preventing species extinction.
400 11 Controlling Transient Chaos and Applications

11.4.1 Food-Chain Model

Extinction of species has been a mystery in nature [604]. A common belief about
local extinction is that it is typically caused by external environmental factors such
as sudden changes in climate. For a species of small population size, small random
changes in the population (known as “demographic stochasticity”) can also lead to
extinction. How species extinction occurs is extremely complex, since each species
typically lives in an environment that involves interactions with many other species
(e.g., through competition for common food sources, predator–prey interactions,
etc.) as well as physical factors such as weather and environmental disturbances.
From a mathematical point of view, a dynamical model for the population size of
a species is complex, involving temporal and spatial variations, external driving,
and random perturbations. Such a system should, in general, be modeled by nonlin-
ear partial differential equations with random and/or regular external driving forces.
A difficulty associated with this approach is that the analysis and numerical solution
of stochastic and/or driven nonlinear partial differential equations present a chal-
lenging problem.
Nonetheless, in certain situations the mathematical model for species extinc-
tion can become simpler. For example, it was suggested by McCann and Yodzis
[510] that deterministic chaos in simple but plausible ecosystem models, mathe-
matically described by coupled ordinary differential equations, can provide a hint
as to how local species extinction can arise without the necessity of considering
temporal or spatial variations and external factors. The key observation is that the
population dynamics of a large class of ecosystems can be effectively modeled by
deterministic chaotic systems [317, 339, 507, 508]. It was shown [510] that transient
chaotic behavior responsible for species extinction can indeed occur in a simple
three-species food-chain model that incorporates biologically reasonable assump-
tions about species interactions [316]. The model involves a resource species, a prey
(consumer), and a predator [510], and is given by

dR R xC yCCR
= R 1− − , (11.20)
dt K R + R0
dC yC R xP yP PC
= xCC −1 − ,
dt R + R0 C + C0
dP yPC
= xP P −1 + ,
dt C + C0

where R, C, and P are the population densities of the resource, the consumer, and
the predator, respectively; K is the resource carrying capacity; and xC , yC , xP , yP ,
R0 , and C0 are parameters.
The biological assumptions of the model are as follows: (1) the life history
of each species involves continuous growth and overlapping generations, with no
age structure, permitting the use of differential equations; (2) the resource pop-
ulation R grows logistically; (3) each consumer species (immediate consumer C,
11.4 Maintaining Chaos to Prevent Species Extinction 401

predator P) without food dies off exponentially; (4) each consumer’s feeding rate,
e.g., xC yC R/(R+R0 ), saturates at high food levels. The resource population R, grow-
ing alone, equilibrates at its carrying capacity K. The resource population and the
intermediate consumer, without the predator, either settles to a stable equilibrium
or a stable limit cycle, a kind of “biological oscillator.” The oscillations are gener-
ated by the saturating feeding response, which permits the resource to periodically
“escape” control by the consumer. With the top predator, there are in a sense two
coupled oscillators in the food chain. A system of coupled oscillators can typically
give rise to chaotic dynamics.

11.4.2 Dynamical Mechanism of Species Extinction

How species extinction can occur in the model can be revealed by a bifurcation anal-
ysis [510]. In particular, a chaotic attractor can arise via the period-doubling route
and is then destroyed through boundary crisis, say at K = Kc . None of the popu-
lations corresponding to trajectories on the chaotic attractor is extinct, because the
attractor is located in a phase-space region away from the origin, (R,C, P) = (0, 0, 0).
In this parameter range, however, there is also a limit-cycle attractor, located in the
plane P = 0, which coexists with the chaotic attractor. Trajectories on the limit-cycle
attractor correspond to the situation in which the predator population is extinct. For
K slightly less than Kc , depending on the choice of the initial condition, the system
either approaches the chaotic attractor or the limit cycle with P = 0. For K slightly
below Kc , there is still a finite distance from the tip of the chaotic attractor to the
basin boundary. Thus, for any initial condition chosen in the basin of the chaotic
attractor, the population of the predator P(t) behaves chaotically in time but never
decreases to zero, because the attractor lives in a region where P(t) = 0. In this case,
the predator never becomes extinct.
As the carrying capacity K increases through the critical value Kc , the predator
will eventually become extinct for almost all initial conditions. This is quite coun-
terintuitive, but it can be understood from the dynamics. At K = Kc , the tip of the
chaotic attractor touches the basin boundary (as in Fig. 3.2), creating “holes” on the
basin boundary through which trajectories can now leak and enter the basin of the
limit-cycle attractor with P = 0. Species extinction can thus occur as the result of
transient chaos.

11.4.3 Control to Prevent Species Extinction

One way to prevent extinction is to decrease the resource carrying capacity K so

that the sustained chaotic motion on the attractor is restored. But ecologically, it
may not be feasible to adjust the carrying capacity of an environment, and even
if this can be done, it may take some time to accomplish after detecting that the
402 11 Controlling Transient Chaos and Applications

predator population is in danger. The predator may already have become extinct
before the carrying capacity can be changed. An alternative approach was proposed
to restore sustained chaotic motions without the need to vary the carrying capacity
of the environment but instead, by making use of the idea of maintaining chaos via
small feedback controls (Sect. 11.2).
One can identify the “dangerous” escape regions surrounding the collision points
between the chaotic attractor and the basin boundary by monitoring the populations
of R, C, and P. If it is determined that the populations are close to a dangerous re-
gion, small but judiciously chosen perturbations to the populations are applied to
guarantee that no immediate exit from the hole occurs. By targeting a set of points
in the escape region for which the trajectory maps back to the region of recurrent
chaotic motion, one can compute the required perturbations. Usually the perturba-
tions need to be applied only rarely. This technique may be of practical use: by
applying small but occasional adjustments to the population at appropriate times
estimated from time series, species extinction can be prevented. From an ecologi-
cal point of view, it may be more feasible to make small adjustments to the local
populations than to change the carrying capacity of the environment.
A potential problem in designing the control algorithm based on the map derived
from a Poincaré surface of section is that a substantial fraction of trajectories escape
and approach the limit cycle at P = 0 without even being controlled. The reason is
that it usually takes a long time for a trajectory to return to the surface of section.
In the case of transient chaos, a trajectory, because of its finite lifetime, may never
pierce through the surface of section before exiting. The following approach was
proposed [716] to maintain sustained chaotic motion for almost all transient chaotic
trajectories. A critical two-dimensional plane in the three-dimensional phase space
(R,C, P) is identified: P = Pcrit = constant, which separates the region of recurrent
chaotic motions from the region in which the dynamics is such that the population
P(t) goes directly to zero. This plane need not be the basin boundary, nor is it a
Poincaré surface of section. The criteria for choosing this plane are these: (1) eco-
logically, it is chosen with respect to the population that can become extinct; and
(2) dynamically, it should be sufficiently close to the originally recurrent chaotic
region. The plane P = Pcrit thus represents a critical level of the endangered popu-
lation at which human intervention must be introduced to prevent the extinction of
the species P. The concept of a “threshold population size” may provide a useful
rule of thumb for manipulating the dynamics, and similar ideas were actually used
in conservation theory [270]. That the critical plane is chosen close to the recurrent
chaotic region indicates that arbitrarily close to but above the critical plane, there ex-
ists an infinite number of points in the phase space, trajectories starting from which
can resume recurrent chaotic motions for at least a finite amount of time. These con-
siderations are illustrated [716], for example, in Fig. 11.6, where the lifetime span
is plotted for trajectories resulting from a grid of 500 × 500 points chosen from a
two-dimensional region in the (R, P) plane at C = 0.5. Here, the lifetime is defined
to be the time that the trajectory spends in the phase-space region with P(t) > Pcrit .
For this example, a simple search procedure leads to the choice of a critical plane
at Pcrit = 0.57. In Fig. 11.6, the yellow and red spots represent points with greater
11.4 Maintaining Chaos to Prevent Species Extinction 403

Fig. 11.6 Lifetime plot of trajectories resulting from a grid of 500 × 500 points chosen from a
two-dimensional region in the (R, P) plane at C = 0.5, where the lifetime is defined to be the
time that the trajectory spends in the phase-space region with P(t) > Pcrit = 0.57. Brighter col-
ors indicate longer lifetime. Model parameters are xC = 0.4, yC = 2.009, xP = 0.08, yP = 2.876,
R0 = 0.16129, and C0 = 0.5 [510]. The bifurcation parameter is set to be K = 1.02 > Kc so that
there is transient chaos [716] (with kind permission from Elsevier Science)

lifetimes than the blue spots. It can be seen that the distribution of the lifetime is
highly nonuniform, due to the fractal structure of the natural measure of the chaotic
saddle.
The setting of a critical plane and the fact that there exists an infinite number
of “hot” spots with long chaotic lifetimes provide us with a feasible way to de-
sign intervention or control. Say the population P(t) falls slightly below the critical
level at time t. Let (R− ,C− , P− ) be the values of the state variables at this time,
where P− is slightly less than Pcrit , and let (R+ ,C+ , P+ ) be the values of the state
variables a little before t, where P+ is slightly above Pcrit . At time t, arbitrarily
small random adjustments [δ R(t), δ C(t), δ P(t)] are made to all the populations
in the phase space within a small neighborhood centered at (R+ ,C+ , P+ ), so that
the trajectory collapses to a point. With a nonzero probability, the trajectory will
be close to one of the hot spots contained in the neighborhood so that chaotic
motion can occur for a finite amount of time. Note that it is not meaningful to
kick the trajectory back directly to the point (R+ ,C+ , P+ ), since this point maps
to (R− ,C− , P− ) immediately. Figure 11.7a shows a controlled population P(t) for
K = 1.02, which indicates a sustained, sizable population of the predator through a
 time. Figure 11.7b shows the magnitude of the applied perturbations δ X(t) ≡
long
[δ R(t)]2 + [δ C(t)]2 + [δ P(t)]2 versus time. It can be seen that the required per-
turbations [δ R(t), δ C(t), δ P(t)] are indeed small (δ X(t) < 0.04, compared with the
size of the population, which is about one) and rare (only about 100 perturbations
404 11 Controlling Transient Chaos and Applications

Fig. 11.7 (a) A controlled population P(t) for K = 1.02, which indicates a sustained, sizable
population of the predator through a long time. (b) Magnitude of the applied perturbations δ X(t)
versus time [716] (with kind permission from Elsevier Science)

are applied in a time interval of (0, 10000)). Numerical computations reveal [716]
that the chaotic population P(t) can be maintained practically indefinitely through
the use of occasional and small adjustments to all the populations, for almost all
initial conditions chosen in the original basin of the chaotic attractor.
An issue of practical interest is how often small adjustments need to be applied
so that finite species populations can be maintained. To address this question, we
observe that the time intervals for successive adjustments of the populations are in
fact the times that the trajectory stays in the region where P > Pcrit . Their average is
the average lifetime τ of the chaotic saddle. For the parameter setting in Fig. 11.6,
it was found [716] that τ ≈ 209, which means that roughly 50 adjustments to the
populations need to be made in a time interval of length of 10, 000. This estimate is
consistent with the result in Fig. 11.7b.
The model discussed has incorporated within itself biologically and ecologi-
cally reasonable assumptions [510]. Even then, the neglected degrees of freedom
would show up as small corrections, and there is always random noise present
in any environment. It thus becomes important to assess the influence of ran-
dom noise. The simplicity embedded in the control method makes it evident that
control is robust against the influence of weak noise. The reason is that in the al-
gorithm, deliberate effort is made to avoid the need to utilize detailed and more
accurate information about the dynamics, such as the derivative matrices and the
stable and the unstable eigenvalues associated with target points. As such, if de-
terministic transient chaos is the main culprit for the extinction of a species for
a particular system, it is possible to control transient chaos to effectively prevent
11.5 Maintaining Chaos in the Presence of Noise, Safe Sets 405

extinction even in noisy environments, regardless of the details of the system dy-
namics. This may be of value to the important environmental problem of species
preservation.

11.5 Maintaining Chaos in the Presence of Noise, Safe Sets

The presence of weak environmental noise may drastically decrease the efficiency
of the algorithms used to maintain chaos. The dynamics is then described by the
stochastic map xn+1 = f(xn , pn ) + σ ξ n , where σ is the noise amplitude, and we
assume that pn is chosen according to one of the chaos-maintaining scenarios de-
scribed in the previous sections. The destructive effect of noise can be weakened or
even eliminated by introducing an additional control variable rn , applied under the
influence of noise. The overall dynamics is then described as

xn+1 = f(xn , pn ) + σ ξ n ,
xn+1 = xn+1 + rn . (11.21)

We assume that the noise is bounded, | ξ n |≤ 1, and weak, σ 1. An interesting

question is how the magnitude r of the control variable should be chosen in order to
ensure maintenance of chaos despite the presence of noise.
The presence of noise implies that trajectories fall, in general, a distance of or-
der σ away from points with long-lived chaotic transients. The amount of control
needed to compensate this shift is therefore at least r = σ . This strategy therefore
does not work for a control weaker than the noise amplitude: r < σ . A remark-
able recent observation of Sanjuán, Yorke, and coworkers [4,661,848,849] was that
there is a strategy for maintaining chaos even if the original control parameter p
is unchanged (it is kept at its nominal value p̄) and even if noise is stronger than
control.
The problem can also be considered as a mathematical game between two players
called the “protagonist” and the “adversary.” The adversary chooses the amount ξ n
of noise, knowing xn and the map f(x) ≡ f(x, p̄). The protagonist’s goal is to survive
around the chaotic saddle, and he/she can choose the response to the adversary’s
action, namely the amount rn of control. The initial condition can also be chosen
by the protagonist. This game was also called Yorke’s game of survival [4]. The
probability that the protagonist will survive in the vicinity of the chaotic saddle is
zero, even without noise, because of escape. This fact makes the survival of the
protagonist nontrivial, in particular if the adversary is allowed to act more strongly
than the protagonist: r < σ .
The idea ensuring survival is based not directly on the chaotic saddle, but rather
on a related concept, the existence of a horseshoe map (cf. Sect. 1.2.2.1) around it.
The action of this map implies that there is a particular set of points, the safe set,
that lies outside but close to the saddle, and the strategy ensures that points of the
map (11.21) remain on the safe set forever.
406 11 Controlling Transient Chaos and Applications

Fig. 11.8 Schematic diagram showing the construction of safe sets Sk . The set S0 (thick line)
consists of a vertical segment that divides Q into two halves and lies outside the horseshoe shape
of f−1 (Q), marked by dashed lines. Safe sets S1 (thin black line) and S2 (thin gray line) are the
preimages of S0 and S1 , respectively [849] (copyright 2008, the American Physical Society)

Different safe sets are needed for different values of σ . Therefore, a family of
safe sets {S j } is defined based on the horseshoe construction. Figure 11.8 shows
how these sets are generated on a topological square, denoted by Q, containing
the chaotic saddle. The action of the inverse map f−1 deforms this square into a
horseshoe. The safe set S0 of level 0 is chosen as a vertical line segment that divides
the square Q into two halves. Points of S0 are in the primary escape region of Q, i.e.,
they leave the square in one iteration. The preimage of S0 within Q contains two
vertical segments. They form the safe set S1 of level 1. Following this procedure,
one defines the set Sk for any k > 1 as the preimage of Sk−1 in Q (see Fig. 11.8).
Thus the safe set Sk of level k has the following properties:
• Sk consists of 2k vertical segments.
• Any vertical segment of Sk has two adjacent segments of Sk+1 that are closer to
it than any other segments of Sk .
• The maximum distance, denoted by δk , between any of the 2k segments of Sk and
its adjacent segments of Sk+1 goes to zero as k → ∞.
The safe set is thus always outside the chaotic saddle, but it is close to the saddle
for k  1. The key idea of the strategy for ensuring survival, and thus maintaining
chaos, is to place the initial condition on one segment of an adequate safe set Sk .
Then one just has to apply the control rn to make point xn+1 of (11.21) lie on the
original segment or another segment of Sk .
11.6 Encoding Digital Information Using Transient Chaos 407

The adequate safe set Sk corresponds to a level k for which δk−1 < σ . For small
σ , k is always large. If the initial condition x0 is on such an Sk , its unperturbed image
f(x0 ) lies on a segment of Sk−1 that has two adjacent segments of Sk . The perturba-
tion ξ 0 due to noise leads to a point x1 = f(x0 ) + σ ξ 0 . If this point lies in the region
between the aforementioned two segments of Sk , there exists a control variable r0 ,
smaller than or equal to δk−1 < σ in modulus, which puts the trajectory on a seg-
ment of Sk . If point x1 is outside the region between the two curves of Sk , its distance
from the segment of Sk−1 is at most σ , and a perturbation smaller than σ can put
the point on the closest segment of Sk . Thus, the image point x1 lies on a segment of
Sk . The same strategy can be applied at any iteration step. One thus always finds a
constant r such that with | rn |≤ r < σ the trajectory xn for any n lies on Sk (with the
same k as the initial condition), and the system is maintained to remain close to the
chaotic saddle forever. The lower bound for the ratio r/σ was shown [4, 848, 849]
to be 1/2, which means that in some cases, chaos can be maintained with a control
as weak as half of the strength of noise. Whether this optimal limit can be reached
depends on the noise amplitude and the properties of the original map f.

11.6 Encoding Digital Information Using Transient Chaos

Developments in nonlinear dynamics and chaos have led to the idea of encoding
digital information using chaos [88–90, 318, 319, 354, 650]. In particular, it was
demonstrated both theoretically and experimentally by Hayes et al. [318, 319] that
a chaotic system can be manipulated, via arbitrarily small time-dependent per-
turbations, to generate controlled chaotic orbits whose symbolic representations
correspond to the digital representation of a desirable message. Imagine a chaotic
oscillator that generates a large-amplitude signal consisting of an apparently random
sequence of positive and negative peaks. A possible way to assign a symbolic repre-
sentation to the signal is to associate a positive peak with a one, and a negative peak
with a zero, thereby generating a binary sequence. The use of small perturbations
to an accessible system parameter or variable can then cause the signal to follow
the orbit whose binary sequence encodes a desirable message that one wishes to
transmit. One advantage of this type of message-encoding strategy is that the non-
linear chaotic oscillator that generates the waveform for transmission can remain
simple and efficient, while all the necessary electronics controlling the encoding of
the signal can remain at a low-powered microelectronic level.
The basic principle that makes the above scheme of digital encoding with chaos
possible lies in the link between chaos and information (Sects. 2.6.3, 8.2.1). The
fundamental unpredictability of chaos implies that chaotic systems can be regarded
as sources that naturally generate digital communication signals. By manipulating a
chaotic system in an intelligent way, digital information can be encoded. A central
issue in any digital communication scheme is channel capacity [71, 708], a quantity
that measures the amount of information that can be encoded. For a chaotic system,
408 11 Controlling Transient Chaos and Applications

channel capacity is equivalent to the topological entropy (Sect. 1.2.3.3) because it

defines the rate at which information is generated by the system.
In a digital communication scheme, it is desirable to have the channel capacity
as large as possible to maximize the amount of information that can be encoded. For
nonlinear digital communication, it is generally advantageous to use transient chaos
as information sources from the standpoint of channel capacity. The orbital com-
plexity associated with trajectories on a chaotic saddle can be greater than that of
trajectories on a chaotic attractor, because crisis is generally a complexity-increasing
event (Sect. 3.1.1). For a symbolic dynamics of two symbols, the maximally allowed
value of the topological entropy, ln 2, is often realized in a parameter regime in
which there is transient chaos (see, e.g., Fig. 3.10). Thus, it is desirable to design a
chaotic system operating in a transient chaotic regime for digital encoding.

11.6.1 The Channel Capacity

For illustrative purpose, we demonstrate how transient chaos can be utilized to en-
code digital information using the one-dimensional logistic map xn+1 = f (xn ) =
rxn (1 − xn ). A symbolic dynamics for the logistic map can be defined by setting
the symbolic partition at the critical point xc = 0.5. A trajectory point x bears the
symbol 0 if x < xc and the symbol 1 if x > xc . A trajectory in the phase space thus
corresponds to a sequence in the symbolic space. The topological entropy K0 quan-
tifies how random such a symbol sequence can be. Its value is obtained from the
number Ωm of possible symbol sequences of length m as given by (1.25). In prac-
tice, one can plot ln Ωm versus m for, say, 1 ≤ m ≤ 16. The slope of such a plot is
approximately K0 .
As r is increased toward rc = 4, the topological entropy K0 continuously in-
creases from zero to ln 2 except when r falls in one of the infinite number of periodic
windows. The topological entropy of the chaotic repeller remains constant in the
window, where the constant is the value of K0 at the beginning of the window. Since
ln 2 is the maximally realizable value of the topological entropy for a symbolic dy-
namics of two symbols, and since a crisis occurs at rc , the entropy remains at ln 2 for
r > rc , as shown in Fig. 11.9. This can be advantageous because message encod-
ing becomes quite straightforward for hyperbolic transient chaos, since there are
no forbidden words associated with the symbolic dynamics. In the communication
terminology, such a channel is unconstrained.

11.6.2 Message Encoding, Control Scheme, and Noise Immunity

To encode an arbitrary binary message into a trajectory that lives on a nonattracting

chaotic set, it is necessary to use small perturbations to an accessible system pa-
rameter or a dynamical variable. For the logistic map we choose to perturb the state
11.6 Encoding Digital Information Using Transient Chaos 409

Fig. 11.9 For the logistic map, the topological entropy K0 versus r for 3.5 < r < 4.1. After the
devil-staircase, for r > 4, K0 remains at ln2, the maximum possible value for a symbolic dynamics
of two symbols [439] (with kind permission from World Scientific Publishing Co)

variable x. Say we wish to apply only small perturbations of order 2−m . A viable
procedure is as follows. First, we convert the message into a binary sequence using
the ASCII code and store the sequence in a symbol register. Next, we choose an ini-
tial condition whose trajectory stays near the chaotic repeller for a certain number
nc (nc > m) of iterations. This is practically feasible, since one can run the system
and predetermine the phase-space regions where initial conditions yield trajectories
whose lifetimes are at least nc . We then determine all m symbols corresponding to
m points on the trajectory starting from x0 and check to see whether the mth symbol
agrees with the first message bit in the symbol register. If yes, we iterate x0 once to
get x1 and determine the mth symbol from x1 (equivalently, the (m + 1)th symbol
from x0 ) to see whether it matches the second message bit in the symbol register.
If not, we apply a small perturbation to x0 so that the mth symbol from it matches
the first message bit. This process continues until all the message bits in the symbol
register are encoded into the chaotic trajectory.
The required parameter perturbation can be computed using the coding function
[318, 319]. First divide the unit interval in x into N bins of size δ x = 1/N, where
δ x 1/2m and 1/2m is the maximally allowed perturbation. We then choose a point
from each bin, iterate it m times, and determine the corresponding symbol sequence
of length m: S1 S2 . . . Sm , where Si can be either zero or one. Any point leaving the
unit interval in fewer than m iterations is disregarded. For those points x for which
a symbol sequence of length m can be defined, the following is computed:
m
R = ∑ Si /2i , (11.22)
i=1
410 11 Controlling Transient Chaos and Applications

Fig. 11.10 Coding function R(x) for the logistic map at r = 4.1, where R can assume any value
between 0 and 1, but there are many gaps on the x-axis, due to the fact that the chaotic repeller is a
fractal Cantor set [439] (with kind permission from World Scientific Publishing Co)

where 0 ≤ R ≤ 1. This gives the value of the coding function R(x) for points on the
chaotic repeller. Since the chaotic repeller has topological entropy ln 2, R can in
principle have any value between 0 and 1. Figure 11.10 shows the coding function
for the logistic map at r = 4.1, where δ x = 2 × 10−4.
With the coding function, the determination of the state perturbations becomes
straightforward. Let the natural m-bit symbol sequence from x0 be a1 a2 . . . am−1 am
(produced by iterating the map directly) and let the first message bit to be encoded
be b1 . We compare the natural symbol sequence a1 a2 . . . am−1 am with the desirable
symbol sequence a1 a2 . . . am−1 b1 and compute δ R = (am − b1 )/2m . From the coding
function R(x), we can then compute the perturbation δ x. This is done by locating
pairs of points with the same value of δ R in the computer representation of the
coding function R(x) and choosing the one that yields the smallest value of δ x. Thus,
by applying δ x to the initial condition x0 , the trajectory point after m iterations is
associated with the symbol that is the first message bit. Note that if am is identical
to the message bit b1 , no perturbation is necessary. To encode the next message bit,
we iterate the perturbed initial condition once to obtain x1 . Let x0 = x1 . The natural
m-bit symbol sequence of x0 is a1 a2 . . . b1 am , where a1 = a2 , a2 = a3 , . . . , and am is
the binary symbol corresponding to the trajectory point f (m) (x0 ). We now compare
am and b2 to determine the next perturbation to be applied to x0 . Continuing this
procedure, we can encode an arbitrary message into the chaotic trajectory {xn }.
An example of encoding a specific piece of information [439] is shown in
Fig. 11.11a, where the English word “TIGER” is encoded into a trajectory on the
chaotic repeller of the logistic map for r = 4.1. The binary (ASCII) representation
of the word is shown at the top of the figure. Assuming that perturbations of magni-
tude 2−8 are to be applied, we generate a set of initial conditions whose lifetimes in
the unit interval under the map are at least 8. Shown in Fig. 11.11a is a time series
for which the first binary bit of the message is encoded into the trajectory at n = 8.
Time-dependent perturbations are applied at subsequent iterations so that the entire
11.6 Encoding Digital Information Using Transient Chaos 411

Fig. 11.11 Encoding the English word “TIGER” into a trajectory on the chaotic repeller at r =
4.1 for the logistic map. The binary representation of the word is shown at the top of the figure.
Shown in (a) is a time series where small control is initiated at n = 1 and the first binary bit of the
message is encoded into the trajectory at n = 8. The dashed–dotted lines represent the endpoints
of the primary escape interval. Time-dependent perturbations are applied at subsequent iterations
so that the entire message “TIGER” can be encoded into the trajectory. The magnitudes of the
control perturbations required are shown in (b) [439] (with kind permission from World Scientific
Publishing Co)

message “TIGER” can be encoded into the trajectory. Figure 11.11b shows the mag-
nitude of the control perturbations applied at different time steps. We see that the
perturbations required are small. No control perturbation is required for the first six
time steps because for this initial condition, the natural symbols corresponding to
the trajectory points from n = 8 to n = 13 happen to coincide with the first six bits
of the message.
Some features of the control scheme are as follows. Since the channel capacity
of the chaotic repeller is ln 2, there are no forbidden symbol sequences. Thus, in
the encoding scheme, any binary sequences can be produced by a typical trajectory
near the chaotic repeller. Since we use the coding function R(x) to compute the
perturbation δ x, once the perturbation is turned on, the trajectory is automatically
confined in the vicinity of the chaotic repeller because the coding function is defined
with respect to trajectories on the repeller. Suppose that small perturbations on the
order of 2−m are to be applied. To encode a message, we need only identify a set of
412 11 Controlling Transient Chaos and Applications

initial conditions that can stay near the chaotic repeller for m iterations. Since the
typical value of m is, say, 10, it is fairly straightforward to identify a large number
of such initial conditions. In practice, before encoding, we can run the system to
produce a set of initial conditions whose lifetimes are greater than m. Together with
the coding function that also needs to be determined beforehand, one can in principle
encode any binary sequence into a dynamical trajectory on the chaotic repeller.
Besides possessing the maximum topological entropy ln 2, the chaotic repellers
of the logistic map for r > 4 also have the property of strong noise immunity. To
see this, we contrast a chaotic repeller with the chaotic attractor at r = √ 4. For the
chaotic repeller, we see that there is a primary escape interval of size ∼ s, where
s = r/4 − 1, about the partition point xc = 1/2. For the chaotic attractor there is no
such gap. A trajectory on the chaotic attractor can then come arbitrarily close to the
partition point. In a noisy environment, this may cause a bit error. Say the trajectory
point is to the immediate right of xc . This point thus has the symbol 1. Due to noise,
the trajectory can be kicked through xc , and it thus assumes the wrong symbol 0.
For a trajectory on the chaotic repeller, this situation is improved. Insofar as the
noise amplitude is smaller than the size of the primary escape interval across the
partition point xc , the symbolic dynamics is immune to noise. This may be of value
to practical implementation of communication with chaos [89, 90].4
Since all chaotic repellers for r > 4 in the logistic map have the same topological
entropy ln 2, it appears that it is more advantageous to use chaotic repellers at large
r because they possess larger gaps across xc , and thus their corresponding symbolic
dynamics are more robust against noise. However, as r increases, the average life-
time of transient chaos decreases. In general, in choosing an optimal chaotic repeller
for digital encoding, there is a trade-off between the ease of generating a trajectory
near the chaotic repeller and the noise immunity [89, 90].
Although our discussion has been focused on one-dimensional maps, similar
ideas apply to transient chaos in two-dimensional maps [443].

4 The stability of transient chaos against noise has been discussed in Chap. 4.
Chapter 12
Transient Chaotic Time-Series Analysis

Parallel to the rapid development of nonlinear dynamics, there has been a

tremendous amount of effort devoted to data analysis. Suppose an experiment
is conducted and some time series are measured. Such a time series can be, for
instance, a voltage signal from a physical or a biological experiment, or the con-
centration of a substance in a chemical reaction, or the amount of instantaneous
traffic at a point in the Internet, and so on. The general question is this: what can
we say about the underlying dynamical system that generates the time series if the
equations governing the time evolution of the system are unknown and the only
available information about the system is a set of measured time series?
The purpose of this chapter is to introduce the delay-coordinate embedding
technique for transient chaotic time-series analysis. The method has proven ef-
fective, particularly for time series from low-dimensional, deterministic dynamical
systems under the influence of weak noise. That is, for situations in which the dy-
namical invariant set responsible for the behavior of the measured time series is
low-dimensional and the noise amplitude is relatively small, the delay-coordinate
embedding method can yield reliable information about the underlying dynamical
system.
The mathematical foundation of the delay-coordinate embedding technique was
laid by Takens in 1980 [756]. He proved that under fairly general conditions, the
underlying dynamical system can be faithfully reconstructed from time series in
the sense that a one-to-one correspondence can be established between the recon-
structed and the true but unknown dynamical systems. Based on the reconstruction,
quantities of importance for understanding the system can be estimated, such as the
relative weights of determinism and stochasticity of the underlying system, its di-
mensionality, the Lyapunov exponents, and unstable periodic orbits that constitute
the skeleton of the invariant set responsible for the observed dynamics.
There exists a large body of literature on the application of the delay-coordinate
embedding technique to dynamical systems with chaotic attractors. Time series ob-
tained from permanently chaotic systems can in principle be arbitrarily long. For
transient chaos, a measured signal exhibits a random behavior only for a finite
amount of time before settling into an asymptotic state. For data analysis, conven-
tional wisdom may be simply to disregard the transient portion of the data and to
concentrate on the final state. By doing this, however, information about the system

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 413
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3 12,

c Springer Science+Business Media, LLC 2011
414 12 Transient Chaotic Time-Series Analysis

may be lost, because the irregular part of the data is likely to contain important
information about the system dynamics. Analyzing transient chaotic time series is
thus necessary.
Pertinent issues in the data analysis of transient chaotic systems include re-
constructing the chaotic saddle, detecting unstable periodic orbits, estimating the
correlation dimension, and computing the Lyapunov exponents. We shall demon-
strate in this chapter that many of the standard algorithms that are used to estimate
dynamical invariants from time series of sustained chaotic processes can be applied
to ensembles of transient chaotic time series. That is, instead of constructing a single
long time series from a set of shorter ones, one can consider a collection of transient
time series, starting from different initial conditions.

12.1 Reconstruction of Phase Space

Let ui (t) (i = 1, . . . , l) be a set of l measurements from an underlying dynamical sys-

tem that evolves the state variable in time according to a set of deterministic rules
represented, e.g., by a set of differential equations. The phase-space dimension can
in general be quite large. However, it often happens that the asymptotic evolution of
the system occurs on a dynamical invariant set of a much smaller dimension only.
A realistic assumption is that the details of the system equations in the phase space
and of the asymptotic invariant set that determines what can be observed through
experimental probes are unknown. The task is to estimate, based solely on one or
a few time series, statistical quantities characterizing the invariant set. The delay-
coordinate embedding technique established by Takens [756] provides a practical
solution to this task [382, 575]. In particular, Takens’s embedding theorem guaran-
tees that a phase space topologically equivalent to that of the unknown dynamical
system can be reconstructed from time series, on the basis of which characteristics
of the dynamical invariant set can be assessed.
Takens’s delay-coordinate embedding method can be described as follows. From
each measured time series ui (t) (i = 1, . . . , l), the following vector quantity of q
components is constructed:

ui (t) = {ui (t), ui (t + τd ), . . . , ui [t + (q − 1)τd]},

where τd is the delay time. Since there are l time series, a vector of m ≡ ql compo-
nents can be constructed as follows:

x(t) = {u1 (t), u2 (t), . . . , ul (t)}

= {u1 (t), u1 (t + τd ), . . . , u1 [t + (q − 1)τd],
u2 (t), u2 (t + τd ), . . . , u2 [t + (q − 1)τd],
. . . , ul (t), ul (t + τd ), . . . , ul [t + (q − 1)τd]}. (12.1)
12.1 Reconstruction of Phase Space 415

The quantity m is called the embedding dimension. The delay time τd and the
embedding dimension m are the two fundamental parameters in Takens’s method.
1. Delay time τd . In order for the time-delayed components ui (t + jτd ) ( j =
1, . . . , q − 1) to serve as independent variables, the delay time τd needs to be cho-
sen properly. If τd is too small, adjacent components ui (t) and ui (t + τd ) will be
too correlated to serve as independent coordinates. If τd is too large, neighboring
components are too uncorrelated for the purpose. Empirically, one can examine the
autocorrelation function of ui (t) and determine a proper delay time [779]. In partic-
ular, for a time series of zero average, ui (t) = 0, one computes

ui (t)ui (t + τd )
c(τd ) ≡ ,
u2i (t)

where · stands for time average. The delay time τd can be chosen to be the value
of τd such that c(τd )/c(0) ≈ e−1 .
There exist various alternative empirical methods for choosing a proper delay
time [109, 405, 484, 485, 654], all of which yield similar results. A firmer theoret-
ical foundation can be established by exploring the statistics for testing continuity
and differentiability from chaotic time series, as proposed by Pecora and coworkers
[271, 582, 584–586].
2. Embedding dimension m. In order to have a faithful representation of the true dy-
namical system, the embedding dimension m should be sufficiently large. Takens’s
theorem [756] provides a lower bound for m. In particular, suppose the dynamical
invariant set lies in a D-dimensional smooth subspace in the phase space (D is an
integer). One can ask whether the dynamical invariant set will intersect itself in the
reconstructed phase space. In order to obtain a one-to-one correspondence between
points on the invariant sets in the actual and reconstructed phase spaces, self-
intersection should not occur. Recall condition (8.52) for the generic intersection
between two sets. Applying this for two identical sets S1 = S2 , D(S1 ) = D(S2 ) = D,
in the phase space of dimension N = m, no self-intersection requires DI < 0, which
implies m > 2D. Then, if m > 2D, the m-dimensional reconstructed vectors x(t)
have a one-to-one correspondence with the vectors in the true dynamical system.
While Takens’s theorem assumes that the relevant dimension D of the set is that
of the manifold in which the set lies, this dimension can be significantly larger than
the dimension of the set itself which is physically more relevant. The work by Sauer,
Yorke, and Casdagli [679] extended Takens’s theorem to relax the dimension re-
quirement: the dimension D can in fact be the box-counting dimension D0 of the
invariant set. For a faithful reconstruction of an invariant set of box-counting di-
mension D0 , the embedding dimension m should be at least 2D0 : m > 2D0 .
Since D0 is not known a priori, a practical criterion is needed for a proper choice
of the embedding dimension m. The method of false neighbors (for a review see the
book by Kantz and Schreiber [382]) provides such a criterion. The aim is to find
points that are neighbors in the m-dimensional embedding space but are not neigh-
bors in reality. Such false neighbors are mapped far away from their real neighbors
416 12 Transient Chaotic Time-Series Analysis

after one time step τd . For m 2D0 the proportion of false neighbors is large, but
for m ≈ 2D0 it becomes small. Numerical algorithms for determining the proportion
of false neighbors are available [324].

12.1.1 Reconstruction of Invariant Sets

For the reconstruction of a chaotic saddle from time series of a single variable, a
method was proposed by Jánosi et al. [356,357]. It is based on the creation of a long
artificial signal and consists of the following steps:
(a) Take an ensemble of time series containing long irregular transients to an attrac-
tor (either simple or chaotic).
(b) Locate the attractor of the dynamics from the dynamical variable of interest and
separate the transient part.
(c) Construct truncated time series by disregarding points that belong to the tran-
sition periods both from the initial point to the saddle and from the vicinity of
the saddle to the attractor. The length distribution of the truncated time series
follows an exponential decay, from which the escape rate can be determined.
(d) By means of some simple matching procedure (e.g., matching the dynamical
variable and its derivative), “glue” the truncated signals together into a long
signal. Apply the delay-coordinate embedding method to determine the chaotic
saddle in the reconstructed phase space.
This method was applied to the NMR laser experiment described in Sect. 1.3. In
particular, the control parameter was kept at a value below an interior-crisis point at
which a sudden attractor enlargement takes place. Long chaotic transients wander in
the region of the postcrisis attractor (Sect. 3.2) before settling into the small attrac-
tor. Transient signals are thus characterized by large oscillations compared to the
signals on the attractor. The gluing procedure was based on the observation that the
laser action between subsequent pulses is nearly zero. It is with respect to this prac-
tically constant laser background level that different transient pieces can be glued
together. The laser output was recorded as a normalized, dimensionless scalar time
series {ui }1024
i=1 (0 < ui < 1) at the sampling rate of 1/2, 500 s.
The particular realization of steps (a)–(d) was as follows:
(a) An ensemble of nearly 104 short records of length 0.4096 s was generated,
which is about four times the average lifetime of transients. This was compared
with a long record on the chaotic attractor. Figure 12.1a, b show a typical seg-
ment of the permanently chaotic signal and the transient signal, respectively.
(The last two-thirds of the transient signal is very close to the attractor.) Note
that between the large peaks of the transient part there are no peaks of interme-
diate height characterizing the motion on the attractor.
(b) The maximum amplitude of oscillations on the attractor can be determined, and
all signals having a peak larger than this were considered to be away from the
attractor. To separate the transient part, the last peak larger than the attractor
maximum was identified and the segment ending with this peak and a short
piece at the laser background level was kept.
12.1 Reconstruction of Phase Space 417

Fig. 12.1 Normalized time series u(t) from the NMR laser experiment of [356]: (a) signal char-
acterizing the chaotic attractor, (b) signal of a typical transient to the attractor, and (c) part of the
long artificial time series representing motion near the chaotic saddle (copyright 1994, the Ameri-
can Physical Society)

(c) It was observed that the transitions to the saddle were short, rendering it suffi-
cient to discard short segments of the time series before the first large peak is
reached. The escape rate was estimated to be κ = 9.3 ± 0.06 s−1 .
(d) To decrease the error related to the gluing procedure, truncated signals shorter
than 1/κ were discarded. The remaining 5600-plus time series were glued to-
gether at the laser background level. A segment of the glued time series is
shown in Fig. 12.1c. The difference between the dynamics on the attractor and
the transient dynamics can be seen. A single peak consists of typically 20–30
data points, indicating that the correlation is high between consecutive data
points, according to which the delay time Δn = 5 (τd = 5/2,500 s) was chosen.
Figure 1.16a and b show the reconstructed chaotic saddle and the coexisting
chaotic attractor, respectively. It can be seen that the saddle is more extensive in
the phase space than the attractor.
418 12 Transient Chaotic Time-Series Analysis

The main advantage of constructing a long, artificial signal from a set of tran-
siently chaotic time series is that the methods applicable to sustained chaotic time
series can be applied straightforwardly. For example, for the laser experiment de-
scribed above, embedding up to dimension 10 leads to a largest Lyapunov exponent
225 (1/s) for the chaotic saddle, which is about twice as large as that of the attractor.
(1)
The partial information dimension D1 along the unstable manifold (2.76) was esti-
mated to be about 0.96. Its closeness to unity explains the lack of large holes along
the reconstructed saddle in Fig. 1.16a.
The stable and unstable manifolds can also be reconstructed from the time series
of a single variable representing an experimentally accessible system. A variant of
the sprinkler method (Sect. 1.2.2.3) was proposed by Triandaf, Bollt, and Schwartz
[788]. They suggested defining a restraining region Γ not containing any attractor
in the m-dimensional delay-coordinate space of a single scalar variable u. One picks
an initial condition within this space and follows the dynamics up to time t = n0 τd ,
n0  1. If the entire trajectory remains within Γ , then record both the initial and
final points in a file. Repeat this for a large number of initial conditions. The set
of all initial (final) points of the trajectories kept reconstruct the stable (unstable)
manifold in the delay-coordinate space. An example with three delay coordinates
I0 , I1 , I2 of a laser system is shown in Fig. 12.2.

Fig. 12.2 Stable (red) and unstable (blue) manifolds of a laser system producing transient chaos.
The embedding space is three-dimensional. Variables I0 , I1 , I2 are intensity delay coordinates, and
the restraining region is the box defined by the axis segments shown. The stable (unstable) mani-
folds are obtained as initial and final points of trajectories remaining in the box for n0 = 200 times
the delay time [788] (copyright 2003, the American Physical Society)
12.1 Reconstruction of Phase Space 419

12.1.2 Reconstructing Invariant Sets of Delay-Differential

Equations

A special class of high-dimensional problems is provided by systems in which the

rate of change of the system’s state depends explicitly on the state at some past time.
Such problems are described by delay-differential equations [227,737], which serve
as prototypical dynamical systems of infinite dimension. For these, both numerical
and analytical methods are intermediate in complexity between ordinary and partial
differential equations. A paradigmatic form of such equations describing a single
variable x(t) is
dx(t)
= F [x(t), x(t − τ )] , (12.2)
dt
where τ > 0 is a time delay. The solution to this equation is unique for all t ≥ 0 if
initial data are given for all times in the interval −τ ≤ t ≤ 0. Since (12.2) is trans-
lationally invariant in time, the continuation of a solution for t > t0 > 0 is uniquely
determined by its history on the interval [t0 − τ ,t0 ]. The delay-differential equation
(12.2) can therefore be regarded as a dynamical system acting on the infinite-
dimensional space of continuous functions defined on the delay interval [−τ , 0].
Transient chaos in delay-differential equations had long been identified in the
form of fractal basin boundaries (see Chap. 5) as in [8,492]. The identification of the
underlying chaotic saddle was much more recent and was due to Taylor and Camp-
bell [761]. They apply the delay-coordinate embedding method to functions defined
on the interval [−τ , 0] by discretizing this interval into an integer number m − 1 of
equal subintervals. In the notation of Sect. 12.1, this corresponds to choosing the
delay time as τd = τ /(m − 1), and m is the embedding dimension. After choosing a
suitable restraining region in the m-dimensional space, a long chaotic trajectory can
be found by means of the stagger-and-step method (Sect. 8.4). A three-dimensional
projection of the saddle can be visualized in the space of x(t), x(t − τ /2), x(t − τ ),
as shown in Fig. 12.3.
A Poincaré section can be defined by intersecting the saddle trajectory of
Fig. 12.3 with a plane x(t) = x∗ . The obtained section through the saddle shown
in Fig. 12.4 has the structure that resembles the product of two Cantor sets, which
is typical in low-dimensional systems. There are a few regions, however, in which
points overlap, indicating that the reconstruction of the saddle in Fig. 12.3 is faithful
outside these regions only. A computation of the correlation dimension yields the
value D2 = 2.2 ± 0.1 [761]. Assuming the box-counting dimension to be close to
this value, the condition for a faithful embedding requires the embedding dimension
to be larger than 4.4 ± 0.2. The three-dimensional embedding of Fig. 12.3 is thus
not yet large enough to fully avoid self-intersections.
420 12 Transient Chaotic Time-Series Analysis

Fig. 12.3 Projection of the saddle of the logistic delay-differential equation (F = −x(t)
−6.16x(t − τ ) [1 − x(t − τ )] in (12.2) with τ = 1) on the three-dimensional space of x(t), x(t
−1/2), x(t − 1). The embedding dimension is m = 250 [761] (copyright 2007, the American Phys-
ical Society)

Fig. 12.4 Poincaré section through the chaotic saddle of the logistic delay-differential equation
obtained by applying the condition x(t) = 1 to the trajectory of Fig. 12.3. The double Cantor struc-
ture can be seen up to a few regions where self-intersection occurs. This indicates that a faithful
reconstruction would require an embedding of dimension somewhat higher than 3 [761] (copyright
2007, the American Physical Society)
12.2 Detection of Unstable Periodic Orbits 421

12.2 Detection of Unstable Periodic Orbits

A fundamental feature distinguishing a deterministic chaotic system from a stochas-

tic one is the existence of an infinite number of unstable periodic orbits that
constitute the skeleton of the chaotic invariant set (see Sect. 1.2.3.3). Determi-
nation of unstable periodic orbits from system equations [63, 156, 157, 685] and
their detection from experimental time series have been an active area of research
[12, 37, 469, 588, 599, 722]. At a fundamental level, unstable periodic orbits embed-
ded in a chaotic invariant set determine its natural measure (Sect. 2.6.4), the basis
for defining physically important quantities such as the box-counting dimension and
Lyapunov exponents. At a practical level, successful detection of unstable periodic
orbits indicates the deterministic origin of the underlying dynamical process.

12.2.1 Extracting Unstable Periodic Orbits from Transient

Chaotic Time Series

An effective algorithm for detecting unstable periodic orbits from chaotic time series
was due to Lathrop and Kostelich (LK) [469]. The method is based on identifying
sets of recurrent points in the reconstructed phase space. To do this, one first recon-
structs a phase-space trajectory x(t) from a measured scalar time series {u(t)} using
the delay-coordinate embedding method described in Sect. 12.1. To identify unsta-
ble periodic orbits, one follows the images of x(t) under the dynamics until a value
t1 > t is found such that | x(t1 ) − x(t) |< ε , where ε is a predetermined small number
that defines the size of the recurrent neighborhood at x(t). In this case, x(t) is called
a (T, ε ) recurrent point, and T = t1 −t is the recurrence time. A recurrent point is not
necessarily a component of a periodic orbit of period T . However, if a particular re-
currence time T appears frequently in the reconstructed phase space, it is likely that
the corresponding recurrent points are close to some periodic orbits of period T . The
idea is then to construct a histogram of the recurrence times and identify peaks in
the histogram. Points that occur frequently are taken to be, approximately, compo-
nents of the periodic orbits. The LK-algorithm was used to detect unstable periodic
orbits, for instance from measurements of a chaotic chemical reaction [469].
The LK algorithm was also adapted to detecting unstable periodic orbits from
short, transiently chaotic series [189]. The reason that the LK algorithm is applicable
to transient time series lies in the statistical nature of the method, since a histogram
of the recurrence times can be obtained even with short time series. Provided that
there is a large number of such time series so that good statistics of the recurrence
times can be obtained, unstable periodic orbits embedded in the underlying chaotic
saddle can be identified. It is not necessary to concatenate many short time series to
form a single long one (as done in Sect. 12.1). Intuitively, since the time series are
short, periodic orbits of short periods can be detected.
422 12 Transient Chaotic Time-Series Analysis

To demonstrate the LK algorithm, here we describe the numerical examples in

[189] with the following chaotic Rössler system in a periodic window of period 3:

dx/dt = −y − z,
dy/dt = x + 0.2y,
dz/dt = 0.2 + (x − 5.3)z. (12.3)

In [189], ten such time series were generated by integrating the Rössler system from
different initial conditions, and the corresponding time series x(t) for 0 ≤ t ≤ 4
are recorded. The average lifetime of the chaotic transients is about 4. These time
series are assumed to be the only available information about the system. For each
time series, a seven-dimensional vector space is reconstructed using the delay time
τd = 0.02. To obtain recurrence times, it is necessary to determine ε , the size of
the recurrent neighborhood. The value of ε cannot be too large so as to avoid false
positives, but if ε is too small, genuine recurrences will be missed. It was found
in numerical experiments that the number of recurrences N(ε ) increases with the
length and the number of the individual transient trajectories, and with ε . It tends
to saturate when ε is too large. The value of ε at which N(ε ) saturates is taken to
be the appropriate size of the recurrent neighborhood. For the Rössler system, using
ε = 2% of the root-mean-square (rms) value of the amplitude of the chaotic signal
is proper [189]. Figure 12.5a shows the histogram of the recurrence times for the
ten transient chaotic time series from the period-3 window. Figure 12.5b–d show,
in the plane of x(t) versus x(t + τd ), three recurrent orbits. The orbit in Fig. 12.5b
has the shortest recurrence time, so it is a “period-1” orbit. Figure 12.5c, d show a

Fig. 12.5 For the Rössler system: (a) histogram N(ε ) of the recurrence time T for ε = 0.02,
(b)–(d) period-1, period-3, and period-8 recurrent orbits extracted from the histogram in (a),
respectively [189] (copyright 2000, the American Physical Society)
12.2 Detection of Unstable Periodic Orbits 423

Fig. 12.6 For a noisy Rössler system: (a) histogram N(ε ) of the recurrence time T for ε = 0.06,
(b)–(d) period-1, period-2, and period-4 recurrent orbits extracted from the histogram in (a), re-
spectively. The rms value of the noise is about 0.5% of that of the chaotic signal [189] (copyright
2007, the American Physical Society)

period-3 and a period-8 orbit, respectively. The orbits are selected from the set of
recurrent points generating the corresponding peaks in the histogram. In general, it
was found [189] that the LK algorithm is capable of yielding many periodic orbits
of low periods.
In an experimental setting, time series are contaminated by dynamical and/or ob-
servational noise. A question is whether periodic orbits can still be extracted from
noisy transient chaotic time series. Figure 12.6 shows the number of recurrent points
(a) and three periodic orbits (b)–(d) extracted from ten transient chaotic time series
with additive weak noise in the form of a normal (Gaussian) distribution centered
at zero and of variance 0.01. This noise level represents an rms value that is ap-
proximately 0.5% of that of the chaotic signal. It can be seen that at this low noise
level, periodic orbits can still be reliably detected. It was found, however, that for the
Rössler system at ε = 2% of the rms value of the chaotic signal with noise beyond
1%, no periodic orbits can be extracted from the histogram of recurrences. One
way to assess the influence of noise is to compute, for different values of ε , how
the number of recurrent points decreases as the noise amplitude (σ ) is increased.
Figure 12.7 shows the result of such computations for ε = 2% (a) and ε = 6% (b) of
the rms value of the signal. It can be seen that the number of recurrent points goes
to zero for σ ≈ ε /2, which can be explained as follows. Under noise of amplitude
σ , both the center and the boundary of the recurrent region are uncertain within σ .
Thus, the effective phase-space volume in d dimensions in which two points can
still be considered within distance ε (recurrent) is proportional to (ε − σ )d − σ d ,
which vanishes at σ = ε /2. Since ε should be small to guarantee recurrence, we see
that the tolerable noise level is also small.
424 12 Transient Chaotic Time-Series Analysis

a 1

ε = 2%

N/N0
0.5

0
0 0.5 1 1.5 2 2.5 3 3.5
noise %
b 1

ε = 6%
N/N0

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5
noise %

Fig. 12.7 For the noisy Rössler system, the relative number N(ε , σ )/N0 (ε ) of recurrent points
versus the amplitude of noise for two values of the size of the recurrent neighborhood: (a) ε = 2%
and (b) ε = 6% of the rms value of the chaotic signal, where N0 (ε ) ≡ N(ε , σ = 0). The vertical line
in (b) denotes the noise level at which periodic orbits in Fig. 12.6 are extracted [189] (copyright
2000, the American Physical Society)

12.2.2 Detectability of Unstable Periodic Orbits from Transient

Chaotic Time Series

Because trajectories about a chaotic saddle have a finite average lifetime τ ≈ 1/κ , it
is difficult for a typical trajectory to contain periodic orbits of period larger than τ .
The detectability of unstable periodic orbits from transient chaotic time series is thus
a relevant issue. In particular, effort may be devoted to connecting short time series
(as in Sect. 12.1) so that the resulting long time series will contain periodic orbits of
larger period. Such a task may be difficult. If one fails to detect periodic orbits of
high periods, the question is whether one should attempt to increase the number of
measurements so that more time series are available. Alternatively, one may attempt
to improve techniques to link these time series, a computationally demanding task.
The main point is that the probability of detecting orbits of higher periods is typi-
cally exponentially small, as we shall point out here. This is an intrinsic dynamical
property of the underlying chaotic set, and hence increasing the number of measure-
ments or improving techniques of detection will not help to enhance the chance of
detecting these orbits.
Let Φ (p) be the probability of detecting an orbit of period p. A scaling relation
for Φ (p) can be derived [189,190] by noting that Φ (p) is effectively the probability
for a trajectory to stay in a small neighborhood of any periodic orbit of period p.
For a trajectory to stay in a ν -neighborhood of all p points of the ith orbit of period
∗
p, the trajectory must come within δ = ν e−λi (p)p of any of the p points when it first
12.2 Detection of Unstable Periodic Orbits 425

encounters the periodic orbit, where λi∗ (p) > 0 is the Lyapunov exponent of this
orbit. The probability of this event is φi (p) ∼ δ αi , where αi is the crowding index
αi ≡ α1i + α2i (Sect. 2.6.4) of the natural measure about any one of the p points of
this periodic orbit. The probability Φ (p) is the cumulative probability of all φi (p):

N(p) N(p)
Φ (p) = ∑ φi (p) ∼ ∑ ν αi exp [−λi∗(p)αi p], (12.4)
i=1 i=1

where N(p) is the total number of periodic points of period p (see (1.26)). Since
λi∗ (p) and αi are the local positive Lyapunov exponent and pointwise dimension of
periodic orbits of period p, for large p, we expect them to obey distributions centered
at λ1 and D1 , respectively, where λ1 and D1 are the positive Lyapunov exponent
and the information dimension of the chaotic saddle. Thus, the main dependence of
Φ (p) on p is

Φ (p) ∼ e−λ1 D1 p N(p) ∼ e(−λ1 D1 +K0 )p ≡ e−γ p , (12.5)

where γ is an exponential scaling exponent and K0 is the topological entropy. Using

the Kantz–Grassberger formulas of two-dimensional maps (2.74) and (2.79) for
chaotic saddles to express D1 in terms of the Lyapunov exponents λ2 < 0 < λ1
and the escape rate κ , one obtains the scaling exponent as

λ12 λ1
γ = λ1 − K0 + −κ 1+ . (12.6)
|λ2 | |λ2 |

Equations (12.5) and (12.6) are applicable to chaotic saddles in two-dimensional in-
vertible maps or in three-dimensional flows. Note that for chaotic attractors (κ → 0),
the scaling exponent becomes γ ≈ λ1 − K0 + λ12 /|λ2 |.
To test (12.5) and (12.6) numerically, Dhamala et al. [189] used the Hénon map
(x, y) → (1 − ax2 + by, x), taking advantage of the fact that unstable periodic orbits
embedded in the chaotic saddles of the map can be computed systematically [63].
For different values of a in the transiently chaotic regime, 106 initial conditions were
chosen in the region [−2, 2] × [−2, 2] containing the chaotic saddle, which yield 106
transient time series. For a given period p, the fractions of times that these 106 time
series get close to every periodic orbit of period p can be computed. These frac-
tions are used to yield an estimated value for the probability Φ (p). The exponent
γ can be read off from the plots ln Φ (p) versus ln p. To compute the theoretical
scaling exponents in (12.6), it is necessary to compute the Lyapunov exponents, the
topological entropy, and the escape rate of the chaotic saddles. The following tech-
niques were used in the computation: (1) the PIM-triple procedure (Sect. 1.2.2.4) to
obtain a long trajectory on the chaotic saddle from which the Lyapunov exponents
can be computed; (2) the method by Chen et al. [125] (Sect. 3.1.1) to compute the
topological entropy; and (3) the sprinkler method (Sect. 1.2.2.3) to compute κ . The
numerical slopes appear to agree reasonably well with the theoretical ones, as shown
in Table 12.1.
426 12 Transient Chaotic Time-Series Analysis

Table 12.1 Theoretical and numerical values of the scaling exponent γ for
three different parameters of the Hénon map
a 1.6 1.8 2.0
λ1 0.58 0.81 0.87
λ2 −1.78 −2.01 −2.07
K0 0.53 0.54 0.53
1/κ 11.2 4.7 5.4
γ (theoretical) 0.12 0.31 0.44
γ (numerical) 0.13 ± 0.04 0.32 ± 0.03 0.47 ± 0.04

12.3 Computation of Dimension

12.3.1 Basics

An often computed dimension in chaotic time-series analysis is the correlation di-

mension D2 . This is the generalized dimension Dq of order q = 2 in the definition
(A.1), which is a lower bound of the box-counting and information dimension:
D2 ≤ D1 ≤ D0 . Grassberger and Procaccia (GP) showed in their seminal contri-
bution [287] that D2 can be evaluated using the correlation integral C(ε ), which
is the probability that a pair of points chosen randomly in the reconstructed phase
space is separated by a distance less than ε . Let N be the number of points in the
reconstructed vector time series x(t) with embedding dimension m. The correlation
integral can be approximated by the following sum:

N N
2
CN (ε , m) =
N(N − 1) ∑ ∑ Θ (ε − |xi − x j |), (12.7)
j=1 i= j+1

where Θ (·) is the Heaviside function and |xi − x j | stands for the distance between
points xi and x j . Grassberger and Procaccia argued that the correlation dimension is
given by [287]
logCN (ε , m)
D2 (m) = (12.8)
log ε
for ε 1, N  1. In practice, for a time series of finite length, the sum in (12.7) also
depends on the embedding dimension m. Due to such dependencies, the correlation
dimension D2 is usually estimated by examining the slope of the linear portion of
the plot of logCN (ε ) versus log ε for a series of increasing values of m. For m < D2 ,
the dimension of the reconstructed phase space is not high enough to resolve the
structure of the dynamical state, and hence the slope approximates the embedding
dimension. As m increases, the resolution of the dynamical state in the reconstructed
phase space is improved. Thus typically, the slope increases with m until it reaches
a plateau, and the plateaued dimension value can be taken as an estimate of D2
[193, 287]. For an infinite and noiseless time series, the value of m at which this
plateau begins satisfies m = Ceil(D2 ), where Ceil(D2 ) is the smallest integer greater
12.3 Computation of Dimension 427

than or equal to D2 [193]. In a realistic situation, short data sets and observational
noise can cause the plateau onset to occur at a value of m larger than Ceil(D2 ).
Even so, the slope at which the plateau is reached still provides a reasonably sharp
upper bound for the true correlation dimension D2 . Dependencies of the length of
the linear scaling region on fundamental parameters such as m, the time delay τd ,
and mτd were analyzed in [456, 457].
These points can be observed by utilizing the Hénon map for the standard param-
eter values a = 1.4 and b = 0.3, where there is a chaotic attractor. The theoretical
value of the correlation dimension of the attractor is D2 ≈ 1.2 [456, 457]. To select
the delay time τd , note that any discrete-time map can be regarded as arising from a
Poincaré surface of section of a continuous-time flow. Thus, one iteration of the map
corresponds to one period of oscillation of the continuous-time signal x(t), which,
for chaotic systems, is approximately the decay time of the autocorrelation of x(t).
As an empirical rule, the delay time can be chosen to be τd = 1.
After the delay time τd is chosen, the correlation integral CN (ε , m) for a set of
systematically increasing values of the embedding dimension m can be computed.
Figure 12.8 shows, for N = 2 × 104, the plots of CN (ε ) versus ε on a logarithmic
scale. The lines are approximately linear, and they are parallel for m ≥ 2. Least-
squares fits give D2 ≈ 1.2 for m ≥ 2, indicating that the correlation dimension can be
estimated reliably. The saturation of the slope occurs at m = 2, which is the smallest
integer above the value of D2 . However, the embedding theorem (Sect. 12.1) re-
quires a minimum embedding dimension of Ceil(2D0 ), which is 3 for the Hénon
problem. This difference exists because the task here is to estimate the dimen-
sion only, while the embedding theorem guarantees a one-to-one correspondence
between the reconstructed and the true phase spaces. A correlation-dimension esti-
mate does not necessarily require such a one-to-one correspondence. For instance,

Fig. 12.8 For the standard Hénon & chaotic attractor, Plots of the correlation integral CN (ε , m) on
a logarithmic scale for m = 2, . . ., 12. Least-squares fits give D2 ≈ 1.2 for m ≥ 2 [457] (with kind
permission from Elsevier Science)
428 12 Transient Chaotic Time-Series Analysis

consider a two-dimensional surface in a three-dimensional space. The projection of

this surface onto a two-dimensional plane is still a two-dimensional region. Thus,
its dimension can be estimated correctly even in a two-dimensional subspace. One
must be aware, however, of the fact that spuriously low dimension estimates can
be obtained if the data are sparse, or if the set is sampled too finely, or if there are
long-range correlations in the data set [284, 489, 562, 616].

12.3.2 Applicability to Transient Chaotic Time Series

The question is whether the GP paradigm ((12.7) and (12.8)) is applicable to tran-
sient time series from chaotic saddles. An argument was provided suggesting an
affirmative answer to this question [190]. The starting point is the dimension spec-
trum (A.1) with Pi ≡ μi as the natural measure of the chaotic saddle contained in
box i. Setting q = 2 gives, for ε 1,
N(ε )
log ∑i=1 μi2 log μi 
D2 = = , (12.9)
log ε log ε
where · denotes the phase-space average over the chaotic saddle. For an ergodic
trajectory on the chaotic saddle, μi  is approximately the probability that the trajec-
tory comes in the ε -neighborhood of a point xi on the saddle in the ith box, which
is given by the correlation sum in (12.7). From measurements, one does not have a
long ergodic trajectory on the chaotic saddle. Instead, say l transient chaotic time
series are available, each of length L. The probability pi that the reconstructed tra-
jectory comes to the neighborhood of xi is
l L
1 1
pi ≈ ∑ ∑ Θ (ε − | xkj − xi |),
l L(L − 1) k=1 j=1

where xkj is the jth trajectory point reconstructed from the kth transient time series.
Since l is in fact the number of initial conditions, the natural measure μi is (see
(2.86))
pi eκ L l L
μi ≈ −κ L ≈
e ∑ ∑ Θ (ε − | xkj − xi |).
lL(L − 1) k=1 j=1
Averaging over all points xi in the reconstructed phase space of dimension m gives

μi  ≈ eκ LCl,L (ε , m), (12.10)

where
l L L
1
Cl,L (ε , m) ≡
lL2 (L − 1) ∑∑ ∑ Θ (ε − | xkj − xki |) (12.11)
k=1 i=1 j=1, j=i

is the correlation integral associated with l observations of transient chaos, each of

L points in the reconstructed phase space. For fixed L, the correlation dimension is
then given by
12.3 Computation of Dimension 429

logCl,L (ε , m)
D2 (m) = (12.12)
log ε

for ε 1, l  1. Equation (12.12) indicates that if one computes the correlation

integral as defined in (12.11), the GP formulation is applicable to transient chaotic
time series as well.
To provide numerical support, transient chaotic time series from the Hénon map
were used [190] for which the correlation dimension can be obtained both from the
GP formulation (12.12) and from a straightforward implementation of the “box-
counting” definition (12.9) by utilizing a long trajectory on the chaotic saddle
generated by the PIM-triple method. For a = 1.5 and b = 0.3, there is a chaotic
saddle of lifetime τ ≈ 30. The “box-counting” definition gives D2 ≈ 1.2. To apply
the GP algorithm, l = 5, 000 transient chaotic time series were used [190]. To guar-
antee that each time series reflects, approximately, the natural measure of the chaotic
saddle, both the initial and final phases were disregarded, and only 20 points from
the middle of the time series were kept. For a given embedding dimension m, the
number of trajectory points corresponding to each time series is then L = 20. Each
time series was normalized to the unit interval and the correlation sum Cl,L (ε , m)
was computed for 100 values of ε for −30 < log2 ε < 0 using embedding dimen-
sions ranging from m = 1 to m = 8, as shown in Fig. 12.9a. For m > 3, the local
slopes of the plots converge to a plateau value, as shown in Fig. 12.9b, which yields
D2 ≈ 1.12. This agrees reasonably well with the value of D2 obtained from the

Fig. 12.9 For the Hénon map, (a) log2 Cl,L (ε , m) versus log2 ε , and (b) log2 Cl,L (ε , m)/ log2 ε ver-
sus log2 ε for a = 1.5, b = 0.3, l = 5, 000, L = 20 for m = 1 (green) and m = 2, . . ., 8 (blue). The
curves with relatively higher slopes correspond to higher embedding dimensions [189] (copyright
2000, the American Physical Society)
430 12 Transient Chaotic Time-Series Analysis

“box-counting” definition. Note that due to the availability of only short time series,
the embedding dimension needs to be larger than the value of D2 itself to yield the
correct plateau value for D2 , in contrast to the case of long time series from chaotic
attractors, where m ≈ D2 usually suffices.

12.4 Computing Lyapunov Exponents from Transient Chaotic

Time Series

The Lyapunov exponents characterize how a set of orthonormal, infinitesimal dis-

tances evolve under the dynamics. For an N-dimensional dynamical system, there
are N Lyapunov exponents, denoted by λi for i = 1, . . . , N, which define the Lya-
punov spectrum. For a chaotic system, values of λi do not depend on the choice of
the initial condition, insofar as it is chosen randomly.
If the system equations are known, the Lyapunov spectrum can be computed
using the standard procedure by Benettin et al. [55]. For chaotic time series, there
exist several methods for computing the Lyapunov spectrum [100, 219, 220, 671,
832]. While the details of these methods differ, they share the same basic principle.
Here we describe the one developed by Eckmann et al. [219]. The algorithm consists
of three steps: (1) reconstructing the dynamics using delay-coordinate embedding
and searching for neighbors of each point in the embedding space, (2) computing
the tangent maps at each point by least-squares fits, and (3) deducing the Lyapunov
exponents from the tangent maps.

12.4.1 Searching for Neighbors in the Embedding Space

Given an m-dimensional reconstructed vector time series (m = ql), in order to deter-

mine the tangent map at xi ≡ (xi , . . . , xi+m−1 ), it is necessary to search for neighbors,
i.e., search for x j such that
x j − xi  ≤ r, (12.13)
where, r is a small number, and  ·  is defined as

x j − xi  = max |x j+α − xi+α |. (12.14)

0≤α ≤m−1

Such a definition of the distance is computationally convenient. If m = 1, the time

series can be sorted to yield

xΠ (1) ≤ xΠ (2) ≤ · · · ≤ xΠ (N) , (12.15)

where Π is the permutation that is stored, together with its inverse Π −1 . The neigh-
bors of xi can then be obtained by examining k = Π −1 (i) and scanning the sorted
12.4 Computing Lyapunov Exponents from Transient Chaotic Time Series 431

time series xΠ (s) for s = k ± 1, k ± 2, . . . until |xΠ (s) − xi | > r. For m > 1, values of s
are first selected for which
|xΠ (s) − xi | ≤ r, (12.16)
as for the case m = 1. The following conditions are further imposed:

|xΠ (s)+α − xi+α | ≤ r, α = 1, 2, . . . , m − 1, (12.17)

resulting in a complete set of neighbors of xi within distance r.

12.4.2 Computing the Tangent Maps

The task is to determine the m × m matrix Ti describing how the dynamics evolves
a small vector about xi to a small vector about xi+1 :

Ti (x j − xi ) ≈ x j+1 − xi+1. (12.18)

A problem is that Ti may not span Rm , because m is usually much larger than
the actual phase-space dimension of the system to guarantee a proper embedding.
Eckmann et al. proposed a strategy that allows Ti to be a dM × dM matrix, where the
matrix dimension dM is less than or equal to m. In such a case, Ti corresponds to the
time evolution from xi to xi+I , where I ≥ 1 is the integer satisfying

m = (dM − 1)I + 1. (12.19)

A new set of embedding vectors of dimension dM can then be constructed:

yi = (xi , xi+I , . . . , xi+(dM −1)I ≡ xi+m−1 ). (12.20)

The new vector yi is obtained by taking every Ith element in the time series, and
hence Ti is defined in the new embedding space as follows:

Ti (y j − yi ) ≈ y j+I − yi+I , (12.21)

or
⎛ ⎞ ⎛ ⎞
x j − xi x j+I − xi+I
⎜ x j+I − xi+I ⎟ ⎜ x j+2I − xi+2I ⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
Ti ⎜ ... ⎟=⎜ ... ⎟. (12.22)
⎜ ⎟ ⎜ ⎟
⎝ x j+(dM −2)I − xi+(dM −2)I ⎠ ⎝ x j+(dM −1)I − xi+(dM −1)I ⎠
x j+(dM −1)I − xi+(dM −1)I x j+dM I − xi+dM I
432 12 Transient Chaotic Time-Series Analysis

That is, Ti can be expressed as

⎛ ⎞
0 1 0 ... 0
⎜ 0 1 ... 0 ⎟
⎜ 0 ⎟
⎜ ⎟
Ti = ⎜ . . . ... ... ... ... ⎟. (12.23)
⎜ ⎟
⎝ 0 0 0 ... 0 ⎠
a1 a2 a3 . . . a dM

The task of finding Ti then reduces to that of finding the set of dM matrix elements
ai (i = 1, 2, . . . , dM ), which can be accomplished using least-squares fits. Let SiE (r)
be the set of indices j of neighbors x j of xi within distance r. The procedure is to
minimize the quantity
 2
dM −1
∑ ∑ ak+1 (x j+kI − xi+kI ) − (x j+dM I − xi+dM I ) . (12.24)
j∈SEi (r) k=0

If SiE (r) is large, the computation required is intensive. If SiE (r) is too small, the
least-squares fit may fail. Generally, it is necessary to choose r sufficiently large
that SiE (r) contains at least dM elements. But r also needs to be small so that the
linear-dynamics approximation about every xi is valid. Eckmann et al. suggested
the following empirical rule for choosing r: Count the number of neighbors of xi
corresponding to increasing values of r from a preselected sequence of possible
values, and stop when the number of neighbors exceeds min(2dM , dM + 4) for the
first time. Increase r further if Ti is singular.

12.4.3 Computing the Exponents

To compute the Lyapunov exponents from the tangent maps is relatively straightfor-
ward. Eckmann and Ruelle [220] suggested the following procedure. Starting from
an identity matrix Q(0) ≡ 1, one carries out the following matrix decomposition
(QR-decomposition):

T1 Q(0) = Q(1) R(1) ,

T1+I Q(1) = Q(2) R(2) ,
···
T1+ jI Q( j) = Q( j+1) R( j+1) ,
··· (12.25)

where the Q( j) are orthogonal matrices and the R( j) are upper triangular matrices
with positive diagonal elements. The above decomposition is robust; an algorithm
for it can be found in [615]. The Lyapunov exponents are then given by
12.4 Computing Lyapunov Exponents from Transient Chaotic Time Series 433

1 K−1
λk = ∑ ln R( j)kk ,
I τd K j=0
(12.26)

where R( j)kk is the kth diagonal element of the matrix R( j) , and K is the number of
available matrices.

12.4.4 A Numerical Example

Here we give an example of computing the Lyapunov exponents from an ensemble

of transient chaotic time series using the procedure outlined above. Again, consider
the Hénon map in a parameter region where the map generates transient chaos. In
[190], chaotic transients were generated from the Hénon map for the parameter pairs
(a, b) = (1.46, 0.3) and (a, b) = (1.50, 0.3), and 21,000 points near each chaotic
saddle were accumulated using 300 random initial conditions in [−2, 2] × [−2, 2]
for the case a = 1.46 and 700 random initial conditions for a = 1.50. (The average
lifetime of the chaotic saddle is about 70 iterates for a = 1.46 and 30 iterates for
a = 1.50.) A two-dimensional embedding (m = 2, dM = 2, I = 1) with the time delay
of τd = 1 was used from each collection of time series. Local linear maps were
computed using least-squares fits for each neighborhood.
For the case a = 1.46, each transient time series consists of about 70 iterates.
Thus, the Lyapunov exponents computed are actually finite-time approximations,
where a suitable product of the 70 or so linear maps associated with points on
the individual transient time series was considered. Similarly, for a = 1.50, it was
necessary to consider products of 30 or so linear maps. Figure 12.10a, b show
the distributions of λ1 and λ2 , respectively. It can be seen that for a = 1.46,
λ1 = 0.44 ± 0.05 and λ2 = −1.72 ± 0.06. Similarly, for a = 1.50 (Fig. 12.10c, d),
the exponents are λ1 = 0.54 ± 0.06 and λ2 = −1.77 ± 0.08. The estimated values of
the exponents agree reasonably well with the true ones [190].

12.4.5 Remarks

The algorithm can in principle compute all the positive Lyapunov exponents reli-
ably [219], although the correct identification of the negative exponents remains
a challenging issue [677, 678]. There are three practical points. First, dM cannot
be too large; otherwise, spurious exponents may arise. Generally, dM should be
larger than the number of positive exponents. Second, the choice of r is critical,
as discussed above. In the presence of noise, it may be useful to replace the ball
{x j : x j − xi  ≤ r} by a shell {x j : rmin < x j − xi  ≤ r} when searching for neigh-
bors. Third, increasing the number of points in the time series at a fixed recording
time is not helpful. In order to improve the computation, the total recording time
should be increased.
434 12 Transient Chaotic Time-Series Analysis

Fig. 12.10 (a)–(d) Distributions of Lyapunov exponents on the chaotic saddle of the Hénon map
at the two parameter values of a = 1.46 and a = 1.50 (b = 0.3), respectively [189] (copyright 2000,
the American Physical Society)

In brief summary, when computing the Lyapunov exponents from time series, the
following rules should be followed:
1. Use long recording time, but not too small a delay time step τd .
2. Use large embedding dimension m.
3. Use a matrix dimension dM somewhat larger than the expected number of positive
Lyapunov exponents.
4. Choose r such that the number of neighbors is greater than the smaller of 2dM
and dM + 4.
5. Otherwise, keep r as small as possible.
6. Take a product of as many matrices as possible to determine the Lyapunov
exponents.
Final Remarks

There are two final remarks that the authors wish to convey to the readers of this
book. First, for researchers in nonlinear dynamics and chaos, the message is that
transient chaos can often be more relevant and fundamental than chaotic attractors.
Second, for researchers from other disciplines or students who are beginning to
study nonlinear dynamics and chaos, the suggestion is to be aware of the ubiquity
of transient chaos and to develop ways of thinking based on transient chaos as early
as possible. Indeed, as we have explored in this book, besides being fundamental to
nonlinear systems, transient chaos arises and finds applications in a wide variety of
disciplines.
At the time of writing, the authors recognize the rapid spreading of transient-
chaos-related concepts in spatiotemporal systems, particularly in shear-flow turbu-
lence. Further interesting results are expected, which may reinforce the view that
chaos, in its transient form, is indeed related to fluid-mechanical turbulence. The
authors speculate that the theory of transient chaos will find immediate applica-
tions in two further areas of significant recent interest: environmental science and
nanoscience. In the former, understanding the spreading of beneficial or toxic mate-
rials in flows is becoming an essential component in environmental protection. The
advection of inertial particles, a topic treated only briefly in this book, will play a
fundamental role in addressing many issues of current concern, such as cloud dy-
namics. In the latter, manifestations of transient chaos in open quantum systems,
for example transport in devices based on graphene, which has been mentioned but
not treated in this book, will be increasingly appreciated. Having said that, the au-
thors wish to emphasize that these speculations are completely their own personal
opinions.

Y.-C. Lai and T. Tél, Transient Chaos: Complex Dynamics on Finite-Time Scales, 435
Applied Mathematical Sciences 173, DOI 10.1007/978-1-4419-6987-3,

c Springer Science+Business Media, LLC 2011
Appendix A
Multifractal Spectra

A.1 Definition of Spectra

The generalized dimensions [45] Dq follow from the scaling form

∑ Pi(ε )q ∼ ε (q−1)Dq , (A.1)

i

valid for ε 1, where Pi (ε ) denotes the box probability defined in (1.13). The
dimension of index q = 0 is the box-counting dimension, and the limit q → 1 defines
the information dimension D1 . The generalized entropies [45] Kq are defined for
large m via the relation
∑ P({S j })q ∼ e(1−q)Kqm , (A.2)
{S j }

where the summation is over all possible symbol sequences of length m, and
P({S j }) are the path probabilities defined in Sect. 1.2.3.3. For q = 0, the left-hand
side of (A.2) represents the number Ωm of symbol sequences of length m. From
(1.25) we see that Kq=0 is the topological entropy. The limit q → 1 defines the metric
entropy. In general hyperbolic cases, the spectra Kq and Dq are related. In general, it
is insightful to explore the Legendre transforms of (q − 1)Dq and (q − 1)Kq with re-
spect to q, the multifractal spectrum f (α ), and the dynamical multifractal spectrum
g(Λ ) [45]. These spectra are infact related to one another.

A.2 Multifractal Spectra for Repellers of One-Dimensional

Maps

The key ingrediet in the so-called thermodynamical formalism [45] that generates all
the spectra is the unique connection between symbol sequences and microstates of
spin chains. In fact, one can interpret symbol 0 (1) of single humped maps (Fig. 2.1)
as a spin pointing downward (upward) and the whole string as a state of a spin
chain of length n. In order to define the interaction between spins, one can consider

437
438 A Multifractal Spectra

the logarithm of a small dimensionless length scale. In one-dimensional maps, the

cylinder size (Sect. 2.1.1) is chosen to be proportional to the internal energy E per
spin in a given microstate:

1 (n)
E({S j }) = − log εi , (A.3)
n

where {S j } is the symbol sequence of length n that belongs to the cylinder of

(n)
size εi . This rule fixes the additive constant in the energy scale. The thermody-
namic limit n → ∞ corresponds to an infinitely refined coverage of the repeller.
Equation (A.3) reflects that the energies associated with the cylinders remain finite
as the partition is refined. In dynamical terms, any value of nE corresponds to a
local stretching exponent (Λ1 of Sect. 1.2.3.3), and the E-values thus to the local
Lyapunov exponent.
The length distribution of the cylinders can be characterized by taking a certain
real power β of the length scales and summing them at level n. The advantage of the
(n)β
statistical-physics analogy just mentioned is that due to (A.3), εi is effectively the
Boltzmann factor exp (−β En) at the inverse temperature 1/β . The corresponding
sum over all configurations is the partition function, and since the free energy is
extensive, an exponential scaling is expected for large n:

(n) β
∑ εi ∼ e−β F(β )n , (A.4)
i

where F(β ) is the free energy per spin and is simply called the free energy in the dy-
namical context.1 The function β F(β ) is monotonically increasing with a negative
second derivative (as in thermodynamics; see also Fig. A.1). The free energy and

Fig. A.1 Free energy for the logistic map f (x) = rx(1 − x) at r = 4.03 in the range | β |< 3. The
escape rate is F(1) and the Lyapunov exponent is the slope of β F(β ) at β = 1

1 In the mathematical literature, −β F(β ) is called the topological pressure [45, 83].
A.2 Multifractal Spectra for Repellers of One-Dimensional Maps 439

the Legendre transform of β F(β ), i.e., the entropy S(E) = β (E − F(β )), provide a
description of the length distribution and, in view of (A.3), of the local Lyapunov
exponents. Recalling that the total length of the cylinders at level n is proportional
to the number N(n) of surviving particles in the restraining region Γ = I, we obtain

F(1) = κ . (A.5)

When calculating the generalized dimensions, we can use an extension of (1.20)

in the form [45]
(n) q
μi
∑ (n) (q−1)Dq
∼ 1, (A.6)
i εi
(n)
where μi is the cylinder measure. Using (2.11) for hyperbolic systems, the left-
(n)
hand side contains a term that includes the lengths εi only, which can thus be
expressed by the free energy. This leads to the following implicit relation:

β F(β ) |β =q−(q−1)Dq = κ q. (A.7)

Since the cylinder measures are simultaneously the path probabilities, (2.12), the
generalized entropies can also be expressed by the free energy based on (A.2), (A.4),
and (2.11):
q(F(q) − κ )
Kq = . (A.8)
q−1
Taking the limit q → 1 in (A.8) leads to K1 = F (1). Thus (2.15) implies that the
Lyapunov exponent λ1 is the derivative of β F(β ) evaluated at β = 1.
The free energy can also be obtained by using an eigenvalue formalism. Consider
the recurrence scheme

(β )
(β ) ψn (x)
ψn+1 (x ) = R(β ) ∑ | f (x) |β , (A.9)
x∈ f −1 (x )

which is the extension of (2.2) for real exponents β and can be called the generalized
Frobenius–Perron equation. Similar to the case of β = 1, the iteration of any smooth
(β )
positive function ψ0 on I leads to a finite limiting ψ (β ) with a special and unique
choice of R(β ) only. It can be shown [763] that this value is

R(β ) = eβ F(β ) . (A.10)

The free energy is related to the leading eigenvalue of the generalized Frobenius–
Perron operator defined by (A.9).
440 A Multifractal Spectra

The relation (A.7) implies that the recurrence scheme

(q)
φn (x)
∑ | f (x) |q−(q−1)Dq
(q)
φn+1 (x ) = eκ q (A.11)
x∈ f −1 (x )

provides an eigenvalue formalism for the order-q dimension. For q = 0, we recover

the dimension equation (2.20). As a special case, the information dimension D1
follows from the scheme

eκ ρ̂n (x) ln | f (x) |

ρ̂n+1 (x ) = (1 − D1) ∑ , (A.12)
κ x∈ f −1 (x )
| f (x) |

which is obtained by taking the limit q → 1 in (A.11).

The set of local Lyapunov exponents E defined by (A.3) can be considered ran-
(n) (n)
dom variables for large n. Since the natural measure of cylinder i is μi ∼ eκ n εi
(see (2.11)), the partition function can be written as

(n) β (n) β −1 −κ n (n)

∑ εi = ∑ εi e μi = e−E(β −1)ne−κ n . (A.13)
i i

The average is with respect to the natural measure. The cumulant expansion of a
random variable u implies [624]
∞
kl
ln eku  = ∑ l! Cl (u), (A.14)
l=1

where k is a real parameter and Cl (u) stands for the lth cumulant of variable u. In our
case, u = nE, k = 1 − β , and the cumulants of nE can be shown [288] in dynamical
systems to be linear in n: Cl (nE) = nQl (with Q1 = λ1 ). Thus, from (A.4), (A.13)
and (A.14), we obtain
∞
(1 − β )l
β F(β ) = κ + λ1 (β − 1) − ∑ Ql . (A.15)
l=2 l!

The cumulants Ql of the local Lyapunov exponent can thus be obtained from the lth
derivative of β F(β ) at β = 1.
Applying this to the general expressions (A.7) and (A.8), we obtain

∞
(1 − q)l−1(1 − Dq)l
κ = (1 − Dq)λ1 + ∑ Ql ,
l=2 l!
∞
(1 − q)l−1
Kq = λ1 − κ + ∑ Ql .
l=2 l!
A.3 Multifractal Spectra of Saddles of Two-Dimensional Maps 441

For q = 0, (2.21) and (2.22) are recovered. This shows that the relation between the
escape rate and the generalized dimensions and entropies always contains the set of
all cumulants Ql . The only exception is the case q = 1, when (2.16) and (2.15) are
recovered.
We note that the relations between the multifractal spectra f (α ) and g(Λ ) and
the entropy S(E) of the local Lyapunov exponent E are explicit [83, 763, 766]:

S(E) 
f (α ) = , g(Λ ) = S(E)|E=Λ +κ . (A.16)
E E=κ /(1−α )

A.3 Multifractal Spectra of Saddles of Two-Dimensional Maps

For an invertible two-dimensional map, the free energy can be defined based on the
(n)
length scales ε1i generated along the unstable direction (Sect. 2.6.1):

(n) β
∑ ε1i ∼ e−β F(β )n , (A.17)
i

where F(1) = κ . The partial generalized dimensions along the unstable direction
(n)
can be calculated via (A.6) with μi as the natural measure inside the stable strips
(n)
of length ε1i according to
(n) q
μi
∑ (1)
∼ 1. (A.18)
i (n) (q−1)Dq
ε1i
The similarity of these relations to (A.4) and (A.6) implies that for the unstable
direction, analogous relations can be obtained as for one-dimensional maps.
An eigenvalue formalism analogous to (A.9) can also be found [763]. The itera-
tive scheme is
(β )
(β ) ψn (x)
ψn+1 (x ) = R(β ) , (A.19)
| J(x)Π (x)β −1 |x∈f−1 (x )
where Π (x) is the local one-step stretching factor along the unstable manifold, and
R(β ) is given by (A.10). The quantity Π (x) yields the ratio of the length Δ1 of the
image of a small interval along the unstable direction to its original length Δ0 . (In the
notation of Sect. 1.2.3.3, in box i, Πi = exp (Λ1i (1)).) The value of R(β ) is set for
(β )
any smooth ψ0 (x) on a restraining region by the existence of a nontrivial limiting
ψ (β ) (x) for n → ∞ whose integral remains finite over Γ . The leading eigenvalue
1/R(β ) and the eigenvalue spectrum of the generalized Frobenius–Perron operator
defined by (A.9) and (A.19), respectively, can also be obtained via a linear approxi-
mant of the operator [419], in a way similar to that in the search for almost invariant
sets (Sect. 10.5.1). For β = 1, the eigenfunctions, however, do not appear to have
any physical meaning.
442 A Multifractal Spectra

Dynamical invariants associated with the stable direction, however, cannot be

calculated from the free energy alone. To this end, a suitably defined partition func-
(n)
tion is needed which contains the length scales ε2 j along the stable manifold [766].

A.4 Zeta Functions

The periodic-orbit property (2.85) stipulates that the length scales of a multifractal
set can be expressed via the Lyapunov exponents of the unstable periodic orbits. In
particular, we can write
∗
∑ e−λ1inβ ∼ e−β F(β )n. (A.20)
i

The sum contains all n-cycle points that are allowed to exist in the system. In gen-
eral, certain cycles are repetitions of shorter ones. All have been included, includes
the so-called primitive cycles that cannot be decomposed into more elementary ones.
The zeta-function approach allows one to concentrate on primitive cycles only.
Consider the free energy, for example. Equation (A.20) can be rewritten as
∗
∑ zn e−λ1inβ ∼ 1, (A.21)
i

under the condition that the smallest z value that makes the sum balanced (the sum
neither decays nor diverges) be just z(β ) = exp [β F(β )]. Examine now the expres-
sion
∗
Ω (z) = ∑ ∑ zn e−λ1i nβ , (A.22)
n i

which, according to (A.21), diverges at z = z(β ). A basic property of periodic orbits

is useful, namely, that the eigenvalue is the same for all possible cyclic elements of
the orbit. Furthermore, each period n can be written as r repetitions of a primitive
cycle of length np : n = rnp . (For primitive cycles, r = 1.) As a result, we have, for all
points i belonging to an n-cycle, λ1i∗ n = rλ1p
∗ n , where λ ∗ is the Lyapunov exponent
p 1p
of a primitive cycle. The term on the right-hand side of (A.22) can be rearranged as
a sum over all primitive cycles and a sum over repetitions
∗
∞ ∗ n β r
np −λ1p np znp e−λ1p np β
Ω (z) = ∑ np ∑ (z e p
) =∑ ∗ n β
−λ1p
, (A.23)
p r=1 p 1 − znp e p

which can be written as z times the derivative of the logarithm of the zeta function
∗
ζβ (z) = ∏(1 − znp e−λ1p np β ). (A.24)
p

Since a divergent term in Ω (z) corresponds to a zero of ζ (z), the free energy follows
from the smallest positive root of the zeta function.
A.4 Zeta Functions 443

The advantage of this approach is that the zeta function is obtained as a partial
summation over arbitrarily long trajectories. It is thus not surprising that accurate
results can be obtained by keeping only a few short primitive cycles in the product
of (A.24). An efficient computational tool, the periodic orbit expansion [149], is
based on the fact that longer orbits can be approximately pieced together from a few
short fundamental primitive cycles. A detailed description of the cycle expansion
and its applications can be found in the e- book [151].
Appendix B
Open Random Baker Maps

B.1 Single Scale Baker Map

One of the simplest two-dimensional maps for transient chaos is the open area-
preserving baker map, in which the two half-squares are stretched by the same factor
a and are compressed by 1/a, which is equivalent to setting a = b and c = d = 1/a
in (2.54) and (2.55), as shown in Fig. 2.14. To generate a random map, the parameter
a is allowed to take on different values at each iterate: an = ā + δ an > 2 [544].
The escape rate of the dynamical process can be obtained as follows. After the
first step, trajectories remaining inside the unit square from a uniform initial distri-
bution fall into two columns, each having unit height and the width 1/a1 . After n
steps, there are 2n identical columns, each of width Πi=1n (1/a ). A finite-time escape
i
rate κ (n) can be defined for a given realization of the parameter fluctuation by iden-
tifying the quantity exp [−κ (n)n] as the area of these columns, which is the same as
the number of nonescaping trajectories up to time n. Thus, we have
n
−κ (n) n = ln Πi=1
n
(2/ai ) = − ∑ ln (ai /2). (B.1)
i=1

By dividing by −n, the right-hand side contains the arithmetic mean of the expres-
sions ln (ai /2), which for large n, converges to the average of the expression taken
over all realizations of the parameter fluctuation. The quantities κ (n) thus converge
to a well-defined number, which is the escape rate κr of the random map

κ (n) → κr = ln (ai /2) = ln ai  − ln 2, (B.2)

where the bracket denotes averaging over all allowed parameter values.
The Lyapunov exponent can be calculated similarly. In particular, any short line
segment is stretched, after n steps, by the factor Πi=1
n a . The stretching exponent
i
defined in (1.18) is thus Λ1 (n) = ∑i=1 ln ai , from which the finite-time Lyapunov
n
(n)
exponent λ1 = Λ1 (n)/n is
1 n
λ1 = ∑ ln ai .
(n)
(B.3)
n i=1

445
446 B Open Random Baker Maps

The average Lyapunov exponent of the snapshot chaotic saddle is thus given by

λ1,r = ln ai . (B.4)

To calculate the box-counting dimension, we cover the set of nonescaping points

along the x-axis by intervals of widths ε = Πi=1n (1/a ). Since there are 2n columns,
i
(2,n)
we can define an order-n approximant D0 to the partial box-counting dimension
along the stable direction as

(2,n) ln 2n ln 2
D0 = = , (B.5)
ln Πi=1 (ai ) (1/n) ∑ni=1 ln ai
n

where the denominator contains the arithmetic mean of ln ai . For large n (fine
resolution), a limit exists, which is the partial fractal dimension. We obtain

(2) ln 2
D0,r = . (B.6)
ln ai 

This expression illustrates that by gradually improving the resolution, all relevant
length scales can be covered. This also implies that the ensemble of nonescap-
ing trajectories traces out a fractal object after sufficiently long times, but this
object changes constantly in time due to the nature of the random map. The box-
counting dimension, however, converges to a constant. Since the random map is
(1) (2)
area-preserving, the two partial dimensions are equal: D0r = D0r . In addition, since
the distribution is identical in each column, the dimensions are in fact the informa-
tion dimensions.
Corrections to the results from the deterministic map can also be worked out. For
small fluctuations we can write

ln(ā + δ ai ) = ln ā + ln (1 + δ ai/ā) ≈ ln ā + δ ai /ā − (δ ai /ā)2 /2.

By taking the average, we see that the variance σ 2 of the parameter fluctuations
appears: ln(ā + δ ai ) = ln ā − (σ /ā)2 /2. Substituting this into (B.2) and (B.4), we
can obtain the relation between the dynamical invariants of the weakly random map
and those of the underlying deterministic map of parameter ā:

κr (σ ) = κ − (σ /ā)2 /2, λ1r (σ ) = λ1 − (σ /ā)2 /2.

It can be seen that these dynamical invariants tend to be smaller for random maps.
The box-counting dimension becomes

(1) (1) 1 σ2
D0r (σ ) = D0 1+ ,
2 ln ā ā2
B.2 General Baker Map 447

which is actually enhanced by randomness. All these features indicate that a

snapshot chaotic saddle is less unstable than its deterministic counterpart.

B.2 General Baker Map

We briefly show how to obtain the dynamical invariants of the random version of
the general baker map defined by (2.54) and (2.55) with parameters an , bn , cn and
dn different at each iterate. Starting from a uniform distribution, after one iteration,
trajectories are in two columns, one of height a1 and width c1 , and the other of height
b1 and width d1 . The portion of particles remaining inside is (1/a1 + 1/b1). In the
next step, the parameters relevant to escape are a2 and b2 , and a fraction (1/a1 +
1/b1)(1/a2 + 1/b2 ) of the number of original trajectories survive. The “finite-time
escape rate” is
1 n 1 1
κ (n) = − ∑ ln ai + bi ,
n i=1
which gives
 
1 1
κr = − ln + .
ai bi
To compute the positive Lyapunov exponent, observe that after two steps,
the stretching exponents for the four columns are ln (a1 a2 ), ln (a1 b2 ), ln (b1 a2 ),
(2)
and ln (b1 b2 ), respectively. The probabilities Pi for falling in these regions
are 1/(a1 a2 ) exp (+κ (2)2), 1/(a1b2 ) exp (+κ (2) 2), etc. The order-2 finite-time
Lyapunov exponent is thus

(2) ln (a1 a2 ) ln (a1 b2 ) ln (b1 a2 ) ln (b1 b2 ) +κ (2) 2

λ1 2 = + + + e .
a1 a2 a1 b2 b1 a2 b1 b2

The order-n expression can be obtained by regrouping terms, yielding

n n
1/ai 1/bi
λ1 n = ∑ ln ai + ∑ ln bi
(n)
.
i=1 1/ai + 1/bi i=1 1/ai + 1/bi

The average Lyapunov exponent of the snapshot chaotic saddle is given by

   
1/ai 1/bi
λ1r = ln ai + ln bi .
1/ai + 1/bi 1/ai + 1/bi

(1)
To calculate the partial information dimension D1 along the unstable manifold,
(2)
we first note that the lengths ε1i along the manifold, generated by the inverted
map after two steps, are (1/a1 )(1/a2 ) and (1/a1 )(1/b2 ), etc. Their logarithms are
448 B Open Random Baker Maps

(2)
exactly the negatives of the stretching exponents. The probabilities Pi for finding
these scales are the same as in the case of calculating the Lyapunov exponent. An
order-n approximation to the dimension can then be defined based on the general
relation (1.22):
(n) (n)
(1,n) ∑ P ln P
D1 = i i(n) i(n) . (B.7)
∑i Pi ln εi
It can be seen that the denominator is −λ (n)n. The numerator is related to the de-
nominator, since P(n) = ε (n) exp(κ (n) n). We thus have

(n)
(1,n) −λ1 n + κ (n)n
D1 = (n)
, (B.8)
−λ1 n

which yields (4.51). The other partial information dimension can be determined in
a similar way. It is given by the analogue of formula (2.78).
To find the partial box-counting dimension along the stable direction, we observe
that after two steps, the widths become (c1 c2 ), (c1 d2 ), (d1 c2 ), and (d1 d2 ) for the
four columns. The dimension D0 of such a four-scale Cantor set can be obtained
from (1.20) through (cD D0 D0 D0 D0 D0
1 + d1 )(c2 + d2 ) = 1, or ∑ ln (ci + di ) = 0. The partial
0
i
(2)
box-counting dimension D0r of the random map is thus determined by the following
implicit relation:
(2) (2)
D0r D0r
ln (ci + di ) = 0.
The other partial dimension follows from a similar equation in which c and d are
replaced by 1/a and 1/b, respectively.
Appendix C
Semiclassical Approximation

C.1 Semiclassical S-Matrix in Action-Angle Representation

For chaotic scattering, there can be infinitely many classical trajectories between
the initial state before the scattering and the final state after the scattering. It is thus
necessary to sum the contributions to the S-matrix elements from all the classically
allowed trajectories.
In the action-angle coordinates the expression of the S-matrix elements in the
action-angle representation is [76]

 −1/2  
1 1/2 ∂I  iΦ̄ (s) (I , I) π
SII ≡ I |S|I =
2π ∑  ∂ θ  exp h̄
− iνs ,
2
(C.1)
s (s)

where the summation is over all classically allowed trajectories s, I, θ and I , θ are
the classical action variables in the initial and the final states, respectively, Φ̄ (s) (I , I)
is the classical action integral along the path s, and νs is the Maslov index (dis-
cussed in detail in Sect. C.2). The preexponential factor in (C.1) can be regarded
as the square root of the contribution of trajectory s to the total classical transition
(s)
probability, denoted by pI→I :

 
(s) 1  ∂ I −1
pII ≡ . (C.2)
2π  ∂ θ (s)

Equations (C.1) and (C.2) constitute the foundation in the study of quantum man-
ifestations of chaotic scattering [75–78, 204, 205, 442], which relate the quantum-
scattering matrix elements to classical quantities in the semiclassical regime 0 <
h̄ 1. The semiclassical sum in (C.1) is valid under the condition that the difference
in the actions between close orbits is larger than the Planck constant h̄. For chaotic
scattering, there are infinitely many trajectories lying closely in the phase space that
contribute to the sum. For long trajectories, the coalescence of the corresponding
actions can be a serious problem. However, the probability of having a long tra-

449
450 C Semiclassical Approximation

jectory decreases significantly with its length, exponentially for hyperbolic chaotic
scattering and algebraically for nonhyperbolic chaotic scattering (cf. Chap. 6). For
average quantities such as the correlation function between the S-matrix elements,
the contributions from the long trajectories are negligible. The semiclassical sum in
(C.1) is thus expected to be valid for chaotic scattering in general.

C.2 Stationary Phase Approximation and the Maslov Index

The stationary phase approximation is key to the semiclassical formulation of quan-

tum mechanics, which enables the wave function in different representations to be
expressed in terms of classical trajectories. An important outcome of this approx-
imation is the π /2 phase shift in the wave function when the underlying classical
trajectory encounters a turning point. The Maslov index in (C.1) is in fact the number
of turning points along a classical path. For pedagogical purpose here we include
a brief introduction to the stationary phase approximation and how it necessitates
the inclusion of the Maslov index in semiclassical quantum mechanics using the
simple system of particle motion in a one-dimensional potential well. The treatment
here follows that in [60, 740].
The Schrödinger equation for a particle in a one-dimensional potential well is

h̄2 d2Ψ
− + V (x)Ψ = EΨ , (C.3)
2 dx2

where the particle is assumed to have unit mass, Ψ (x) is the wave function in the
coordinate representation, and the potential function V (x) has a local minimum at
x = 0 and it increases with both positive and negative x. In the semiclassical approx-
imation, the wave function is written as [60]

Ψ (x) = A(x) exp [iφ (x)], (C.4)

where the amplitude function A(x) is assumed to vary slowly as compared with the
phase φ (x). Substituting this ansatz into the Schrödinger equation and neglecting
the second derivative of A(x), one obtains the following wave function in the semi-
classical approximation:
 x
Ψ0 i
Ψ (x) =  exp p(x)dx , (C.5)
|p(x)| h̄ x0


where Ψ0 is an integration constant and p(x) = ± 2[E − V (x)] is the local mo-
mentum of the particle. Apparently, this approximation breaks down at the classical
turning point where p(x) = 0. One way to overcome this difficulty is to make use
of the momentum-space representation [212, 503]. This is based on the intuition
C.2 Stationary Phase Approximation and the Maslov Index 451

that if an approximation is poor in the coordinate space, it should be good in the

momentum space, and vice versa. The momentum-space wave function Ψ̄ (p) is the
Fourier transform of the coordinate-space wave function,

1 i
Ψ̄ (p) = √ dxΨ (x) exp − xp (C.6)
2π h̄ h̄
  x
Ψ0 dx i
= √  exp p(x)dx − xp .
2π h̄ |p(x)| h̄ x0

Note that in (C.6), p(x) is the local momentum and p is the argument of Ψ̄ (p). The
integral in (C.6) is of the type

IΦ = dxA(x) exp [iΦ (x)] (C.7)

with the phase function

 x
1
Φ (x) = p(x)dx − xp .
h̄ x0

In the semiclassical limit h̄ → 0, the phase Φ (x) oscillates rapidly, so almost all
contributions to the integral cancel each other except for points in the neighborhood
of xs at which the phase is stationary, i.e., Φ (xs ) = 0. Expanding Φ (x) up to (x −
xs )2 , substituting it in (C.7), and making use of the Fresnel integral

 ∞ 
π 1/2
π
eiα x dx =
2
exp i sgn(α ) ,
−∞ |α | 4

one obtains

2π 1/2 . π /
IΦ ≈ A(xs ) exp iΦ (xs ) + i sgn[Φ (xs )] . (C.8)
|Φ (xs )| 4

The first two derivatives of Φ (x) are Φ (x) = [p(x) − p]/h̄ and Φ (x) = p (x)/h̄.
The stationary point is thus determined by p(xs ) = p. Using (C.8) yields
  x
Ψ0 i s iπ
Ψ̄ (p) ≈  exp p(x)dx − pxs + sgn[p (xs )] , (C.9)
|p(xs )p (xs )| h̄ x0 4

where p(xs )p (xs ) = (1/2)dp2 /dx|x=xs = −V (xs ). Thus, in the vicinity of a generic
turning point where V (xs ) = 0, the semiclassical momentum-space wave function
is well defined.
To understand the meaning of the Maslov index, consider a closed orbit in the
classical phase space (x, p), as shown in Fig. C.1. There are two turning points at
which the local momentum changes sign. To obtain the semiclassical wave function
452 C Semiclassical Approximation

p
use Ψ(x)

use
c d Ψ(p)
turning
point
x
* *
turning
use point
Ψ(p) b a

use Ψ(x)

Fig. C.1 For motion in a one-dimensional potential well, a closed orbit in the classical phase
space. There are two turning points at which the local momentum changes sign. Dividing the orbit
into four segments, one uses either the momentum-space or the coordinate-space wave function,
depending on whether the segment contains a turning point or not

for the entire orbit, one can divide the orbit into four segments, with the dividing
points a, . . . , d, and use the coordinate-space wave function for the two segments that
do not contain any turning point (the segments between “a” and “b” and between
“c” and “d”) and the momentum-space wave function for the two segments each
containing one turning point (the segments between “b” and “c” and between “d”
and “a”). For example, for the segment between “a” and “b,” the coordinate-space
wave function is
 x
Ψ0 i
Ψ (x) = exp p(x)dx . (C.10)
|p(x)|1/2 h̄ x0

At point “b,” it is necessary to transform into the momentum space. This can be
done by replacing xs and p (xs ) in (C.9) by x(p) and 1/x (p), respectively. Since
x (p) < 0 at point b, one has
    x(p)
 x (p) 1/2 iπ
Ψ̄ (p) = Ψ0   exp i p(x)dx − px(p) − . (C.11)
p  h̄ x0 4

At point “c” one transforms the wave function back to the position space. This can
be accomplished by the Fourier transform of Ψ̄ (p), which gives

1 i
Ψ (x) = dpΨ̄ (p) exp xp (C.12)
(2π h̄)1/2 h̄
     x(p)
Ψ0  x (p) 1/2 i iπ
= 
dp   exp p(x)dx − px(p) + px − .
(2π h̄)1/2 p  h̄ x0 4
C.2 Stationary Phase Approximation and the Maslov Index 453

A straightforward application of the stationary phase approximation gives

 x
Ψ0 i iπ
Ψ (x) = exp p(x)dx − . (C.13)
|p(x)|1/2 h̄ x0 2

Comparing (C.10) and (C.13), one sees that a phase loss of π /2 has occurred while
passing through the turning point between points “b” and “c.” Thus, after completing
the closed orbit, the phase shift in the wave function is
0 0
1 π 1 π
ΔΦ = p(x)dx − ν = x(p)dp − ν , (C.14)
h̄ 2 h̄ 2

where ν is the number of turning points along the closed orbit, the Maslov index
[60, 503].
In the semiclassical expression of the S-matrix elements in (C.1), it is thus neces-
sary to include the phase accumulation νπ /2, considering that a chaotic scattering
trajectory can typically have many turning points in the phase space.
Appendix D
Scattering Cross Sections

D.1 Scattering Cross Sections in Classical Chaotic Scattering

In a typical scattering experiment, particles are not followed individually. Rather,

a uniform beam of particles approaching the scattering region is examined and the
fraction of the particle ensemble scattered into a given direction is determined. This
process can be characterized by the differential cross section, generally defined as
the number of particles scattered into a given direction per unit time divided by
the incident intensity [268]. For chaotic scattering, the differential cross section is
proportional to the probability that a particle is scattered into the given direction.
We have seen in Chap. 6 that for chaotic scattering, a typical scattering function
(deflection function or delay-time function) contains both smooth and discontinu-
ous parts. In particular, the impact parameter line splits into an infinite number of
intervals in which the scattering function is smooth, and the boundaries among the
smooth intervals are points of discontinuity. The cross section can be obtained by
fixing a deflection angle θ and a small increment dθ about θ and examining the val-
ues of the impact parameter that result in scattering with deflection angles falling in
the interval [θ − dθ /2, θ + dθ /2]. If there is a one-to-one correspondence between
the deflection angle and the impact parameter, as in regular scattering in classical
mechanics [268], the differential cross section σ (θ ) for scattering in direction θ in
a two-dimensional problem is
 −1
 dθ 
c =  (b) . (D.1)
db

For chaotic scattering there are typically infinitely many contributions to the same
θ from the different intervals of continuity. As pointed out by Jung and Pott [368],
it is necessary to determine all values bs of the impact parameter that lead to the
particular scattering angle θ . For each bs the quantity c in (D.1) is measured and all
the contributions are summed to yield the cross section for chaotic scattering:
 −1
 dθ 
σ (θ ) = ∑ cs = ∑  (bs ) . (D.2)
s s db

455
456 D Scattering Cross Sections

The cross section is usually a complicated expression of the derivatives of the

deflection function taken at bs . In fact, the cross section is the total length of
short impact parameter intervals dbs that lead to scattering into the angle interval
[θ − dθ /2, θ + dθ /2].
In each interval of continuity, the deflection function can have an infinite num-
ber of extrema. At such angles, the cross section is singular, implying that σ (θ )
has much more contributions in certain regions than in others. Let θc denote an ex-
tremum value of the deflection function at bc . The neighborhood of a nondegenerate
extremum is quadratic. Locally, we can write

θ (b) = θc + a(b − bc)2 . (D.3)

In this neighborhood two impact parameter values, b1,2 = bc ± [(θ − θc )/a]1/2 ,

belong to a given angle θ . Their contributions to the cross section are c1,2 =
[4a(θ − θc )]−1/2 . The cross section thus contains singularities of the type

σ (θ ) ∼ (θ − θc )−1/2 . (D.4)

These so-called rainbow singularities dominate the entire cross section function
σ (θ ), as illustrated by Fig. D.1 for a representative chaotic scattering system.
Trajectories starting in short intervals of continuity stay for a long time near the
chaotic saddle, and leave the scattering region along the saddle’s unstable mani-
fold. This manifold is a fractal characterized by a partial box-counting dimension
(2)
D0 < 1. As a result, the distribution of the rainbows in any short interval of continu-
ity reflects the fractal pattern of the saddle in the phase space [368]. The conclusion
is then that the box-counting dimension of the rainbow singularities coincides with

Fig. D.1 Differential cross section of a three-hill system in the scattering angle interval [46◦ , 74◦ ]
for a fixed value of the energy. The rainbow singularities form a fractal set of dimension ds [369]
(with kind permission from the Institute of Physics)
D.2 Semiclassical Scattering Cross Sections 457

(2) (1)
the partial box-counting dimension D0 = D0 ≡ ds of the chaotic saddle. The tech-
nical difficulty that there are a large number of weak rainbow singularities, which
can hardly be distinguished from the smooth background, can be overcome by ap-
plying rapidly converging methods for the determination of the dimension [369].
Rainbow singularities are present in nonhyperbolic chaotic scattering as well. As the
scattering progresses from the regular to the chaotic regime, rainbow singularities
are shown to be created in a series of cascades, related to the bifurcation cascades
undergone in the chaotic saddle [680]. There are systems that exhibit chaotic scat-
tering but still possess a smooth cross section without any singularities [164]. Such
situations cannot occur in high-dimensional systems if the chaotic saddle is formed
by normally hyperbolic invariant manifolds (Sect. 8.5.2) [366]. As pointed out in
[367], knowledge about the rainbow singularities in the cross section allows one to
reconstruct the development stage of the chaotic set defined in Sect. 6.4.2.

D.2 Semiclassical Scattering Cross Sections

In a scattering experiment, the incoming momentum (energy) is specified as pre-

cisely as possible, but the impact parameter is unspecified. The differential cross
section is usually measured as a function of the angle and energy. The maximal
amount of information that can be obtained from a scattering experiment on micro
scales is provided by this cross section. When comparing classical and quantum-
mechanical cross sections of the same process, the distinct feature in the quantum
case is the interference oscillations. At the semiclassical level these oscillations con-
tain relevant information about the classical chaotic saddle.
In a semiclassical approximation, the scattering amplitude follows from Miller’s
theory (C.1) as
 
iΦ̄ (s) π
f (θ , E) = ∑ cs 1/2
exp − iνs . (D.5)
s h̄ 2

The summation is over all classical scattering trajectories s at energy E and with
deflection angle θ . The quantity cs is the contribution of trajectory s with impact
parameter bs to the classical cross section, as given by (D.1) evaluated at b = bs ,
and Φ̄ (s) and νs represent the classical action and the Maslov index associated with
trajectory s, respectively (see Appendix C). Assuming that the angle θ is away from
any classical rainbow singularities, we can express the differential cross section by
the scattering amplitude as [301]

σ (θ , E) = | f (θ , E) |2 . (D.6)

The interference effects in the cross section in the limit of small h̄ can then be
examined. Since, away from rainbow singularities, the classical contribution cs is
a slowly varying function, we can evaluate cs at some reference values θ0 , E0 .
458 D Scattering Cross Sections

Away from caustics, νs is approximately constant. The only source of fast varia-
tion is then the phase Φ̄ (s) /h̄, because h̄ is small. For fixed energy, the action can be
expanded about θ0 :

Φ̄ (s) (θ ) = Φ̄ (s) (θ0 ) + (θ − θ0 )Ls , (D.7)

where Ls = ∂ Φ̄ (s) /∂ θ |θ0 is the outgoing angular momentum associated with trajec-
tory s. The cross section is then obtained from (D.6) and (D.5) as

σ (θ , E) = ∑ cs + ∑ 2(cr cs )1/2 cos [φrs − θ (Lr − Ls )/h̄], (D.8)

s r<s

where the φrs are constant. The first term is the classical contribution (D.2), and
the double sum represents the interference oscillations. The values of the outgoing
(1)
angular momentum differences form a fractal set of dimension D0 . The Fourier
transform

θL θ
g(L) = σ (θ ) cos d (D.9)
h̄ h̄
for a fixed E0 over a suitable range of θ about θ0 away from classical rainbow
positions thus contains singular contributions, as shown in Fig. D.2
The angular-momentum differences L belonging to singular values of the semi-
classical cross section’s Fourier transform form a set of box-counting dimension
(1) (2)
D0 = D0 ≡ ds [369]. It is remarkable that even in an angle range where the clas-
sical cross section is smooth, the interference oscillations reflect the fractal structure
of the classical chaotic saddle.

Fig. D.2 For a three-hill chaotic scattering system, logarithm of the Fourier transform g(L) of the
differential cross section as a function of the angular-momentum difference L. Integration is taken
in the angle interval (5.4, 5.402) with h̄ = 2 · 10−7 at a fixed energy. The singularities form a fractal
set of dimension ds [369] (with kind permission from the Institute of Physics)
References

1. R. Abraham and S. Smale. Nongenericity of Ω -stability. Global Analysis, Proc. of Symposia

in Pure Math. (AMS), 14:5–8, 1970.
2. R. L. Adler, A. G. Konheim, and M. H. McAndrew. Topological entropy. Trans. Am. Math.
Soc., 114:309, 1965.
3. V. S. Afraimovich and L. A. Bunimovich. Which hole is leaking the most: a topological
approach to study open systems. Nonlinearity, 23:643–656, 2010.
4. J. Aguirre, F. d’Ovidio, and M. A. F. Sanjuán. Controlling chaotic transients: Yorke’s game
of survival. Phys. Rev. E, 69:016203, 2004.
5. J. Aguirre and M. A. F. Sanjuan. Limit of small exits in open Hamiltonian systems. Phys.
Rev. E, 67:056201, 2003.
6. J. Aguirre, J. C. Vallejo, and M. A. F. Sanjuán. Wada basins and chaotic invariant sets in the
Hénon-Heiles system. Phys. Rev. E, 64:066208, 2001.
7. J. Aguirre, R. L. Viana, and M. A. F. Sanjuán. Fractal structures in nonlinear dynamics. Rev.
Mod. Phys., 81:333–386, 2009.
8. J. Aguirregabiria and J. Etxebarria. Fractal basin boundaries of a delay-differential equation.
Phys. Lett. A, 122:241–244, 1987.
9. G. Ahlers and R. Walden. Turbulence near onset of convection. Phys. Rev. Lett., 44:445–448,
1980.
10. R. Akis, J. P. Bird, D. K. Ferry, and D. Vasileska. Nonuniform energy level broadening in open
quantum dots: the influence of the closed dot eigenstates on transport. Physica E, 7:745–749,
2000.
11. J. C. Alexander, J. A. Yorke, Z. You, and I. Kan. Riddled basins. Int. J. Bifur. Chaos Appl.
Sci. Eng., 2:795–813, 1992.
12. S. Allie and A. Mees. Finding periodic points from short time series. Phys. Rev. E, 56:346–
350, 1997.
13. H. Alt, H.-D. Gräf, H. L. Harney, R. Hofferbert, H. Rehfeld, A. Richter, and P. Schardt. Decay
of classical chaotic systems: the case of the Bunimovich stadium. Phys. Rev. E, 53:2217–
2222, 1996.
14. E. G. Altmann and H. Kantz. Hypotesis of strong chaos and anomalous diffusion in coupled
symplectic maps. EuroPhys. Lett., 78:10008, 2007.
15. E. G. Altmann. Emission from dielectric cavities in terms of invariant sets of the chaotic ray
dynamics. Phys. Rev. E, 79:013830, 2009.
16. E. G. Altmann, E. C. da Silva, and I. L. Caldas. Recurrence time statistics for finite size
intervals. Chaos, 14:975–981, 2004.
17. E. G. Altmann, A. E. Motter, and H. Kantz. Stickiness in Hamiltonian systems: from sharply
divided to hierarchical phase space. Phys. Rev. E, 73:026207, 2006.
18. E. G. Altmann and T. Tél. Poincaré recurrences from the perspective of transient chaos. Phys.
Rev. Lett, 100:174101, 2008.
19. E. G. Altmann and T. Tél. Poincaré recurrences and transient chaos in leaked systems. Phys.
Rev. E, 79:016204, 2009.

459
460 References

20. E. G. Altmann and A. Endler. Noise-enhanced trapping in chaotic scattering. Phys. Rev. Lett,
105:244102, 2010.
21. M. Antoni and S. Ruffo. Clustering and relaxation in Hamiltonian long-range dynamics.
Phys. Rev. E, 52:2361–2374, 1995.
22. F. T. Arecchi and F. Lisi. Hopping mechanism generating 1/F noise in non-linear systems
Arecchi and Lisi Respond. Phys. Rev. Lett., 50:1330–1330, 1983.
23. F. T. Arecchi, R. Meucci, G. Puccioni, and J. Tredicce. Experimental-evidence of sub-
Harmonic bifurcations, multistability, and turbulence in a Q-switched gas-laser. Phys. Rev.
Lett., 49:1217–1220, 1982.
24. F. T. Arecchi, R. Badii, and A. Politi. Low-frequency phenomenon in dynamical systems with
many attractors. Phys. Rev. A, 29:1006–1009, 1984.
25. H. Aref. Integrable, chaotic, and turbulent vortex motion in two-dimensional flows. Annu.
Rev. Fluid. Mech., 15:345–389, 1983.
26. H. Aref. Stirring by chaotic advection. J. Fluid Mech., 143:1–21, 1984.
27. H. Aref and M.S. El Naschie. Chaos Applied to Fluid Mixing. Pergamon, New York, first
edition, 1995.
28. J. Aristegui, P. Tett, A. Hernández-Guerra, G. Basterretxea, M. F. Montero, K. Wild, P. Sangrá,
S. Hernández-Leon, M. Canton, J. A. Garca-Braun, M. Pacheco, and E. D. Barton. The in-
fluence of island-generated eddies on chlorophyll distribution: a study of mesoscale variation
around Gran Canaria. Deep Sea Res., 44:71–96, 1997.
29. L. Arnold. Random Dynamical Systems. Springer, Berlin, first edition, 1998.
30. P. Ashwin and E. Stone. Influence of noise near blowout bifurcation. Phys. Rev. E, 56:1635–
1641, 1996.
31. E. Aurell, G. Boffetta, A. Crisanti, G. Paladin, and A. Vulpiani. Growth of noninfinitesimal
perturbations in turbulence. Phys. Rev. Lett., 77:1262–1265, 1996.
32. T. R. Auton, J. C. R. Hunt, and M. Prud’homme. The force exerted on a body in inviscid
unsteady non-uniform rotational flow. J. Fluid Mech., 197:241–257, 1998.
33. M. Avila, A. P. Willis, and B. Hof. On the transient nature of localized pipe flow turbulence.
J. Fluid. Mech., 646:127–136, 2010.
34. T. Baba. Photonic crystals and microdisk cavities based on GaInAsP-InP system. IEEE Select.
Top. Quantum Electron., 3:808–830, 1997.
35. A. Babiano, G. Boffetta, A. Provenzale, and A. Vulpiani. Chaotic advection in point vortex
models and two-dimensional turbulence. Phys. Fluids A, 6:2465–2474, 1994.
36. A. Babiano, J. H. E. Cartwright, O. Piro, and A. Provenzale. Dynamics of a small neutrally
buoyant sphere in a fluid and targeting in Hamiltonian systems. Phys. Rev. Lett., 84:5764–
5767, 2000.
37. R. Badii, E. Brun, M. Finardi, L. Flepp, R. Holzner, J. Pariso, C. Reyl, and J. Simonet.
Progress in the analysis of experimental chaos through periodic orbits. Rev. Mod. Phys.,
66:1389–1415, 1994.
38. M. S. Baptista, S. Kraut, and C. Grebogi. Poincaré recurrence and measure of hyperbolic and
nonhyperbolic chaotic attractors. Phys. Rev. Lett., 95:094101, 2005.
39. H. U. Baranger. Quantum transport and chaos in semiconductor microstructures. Physica D,
83:30–45, 1995.
40. H. U. Baranger, R. A. Jalabert, and A. Douglas Stone. Quantum-chaotic scattering effects in
semiconductor microstructures. Chaos, 3:665–682, 1993.
41. E. Barreto and P. So. Mechanisms for the development of unstable dimension variability and
the breakdown of shadowing in coupled chaotic systems. Phys. Rev. Lett., 85:2490–2493,
2000.
42. A. M. Barr, K. Na, L. E. Reichl, and C. Jung. Chaotic scattering in a molecular system. Phys.
Rev. E, 79:026215, 2009.
43. W. Bauer and G.F. Bertsch. Decay of ordered and chaotic systems. Phys. Rev. Lett, 65:2213–
2216, 1990.
44. P. D. Beale. Noise-induced escape from attractors in one-dimensional maps. Phys. Rev. A,
40:3998–4003, 1989.
References 461

45. C. Beck and F. Schlögl. Thermodynamics of Chaotic Systems: An Introduction. Cambridge

University Press, Cambridge, UK, first edition, 1993.
46. C. W. J. Beenakker. Random matrix theory of quantum transport. Rev. Mod. Phys., 69:731–
808, 1997.
47. C. W. J. Beenakker and H. van Houten. Billiard model of a ballistic multiprobe conductor.
Phys. Rev. Lett., 63:1857–1860, 1989.
48. D. Beigie, A. Leonard, and S. Wiggins. Invariant manifold templates for chaotic advection.
Chaos Sol. Fract., 4:749–868, 1994.
49. A. Ben-Mizrachi, I. Procaccia, and P. Grassberger. Characterization of experimental (noisy)
strange attractors. Phys. Rev. A, 29:975–977, 1984.
50. I. J. Benczik, T. Toroczkai, and T. Tél. Advection of finite-size particles in open flows. Phys.
Rev. E, 67:036303, 2003.
51. I. J. Benczik, S. Z. Benczik, B. Schmittmann, and R. K. P. Zia. Lack of consensus in social
systems. Europhys. Lett., 82:48006, 2008.
52. I. J. Benczik, S. Z. Benczik, B. Schmittmann, and R. K. P. Zia. Opinion dynamics on an
adaptive random network. Phys. Rev. E, 79:046104, 2009.
53. I. J. Benczik, Z. Toroczkai, and T. Tél. Selective sensitivity of open chaotic flows on inertial
tracer advection: catching particles with a stick. Phys. Rev. Lett., 89:164501, 2002.
54. J. Bene. Multifractal properties of a class of non-natural measures as an Eigenvalue problem.
Phys. Rev. A, 39:2090–2096, 1989.
55. G. Benettin, L. Galgani, A. Giorgilli, and J. M. Strelcyn. Lyapunov characteristic exponents
for smooth dynamical systems and for Hamiltonian systems: a method for computing all of
them. Part 1: Theory. Meccanica, 15:9–20, 1980.
56. M. Berblinger and C. Schlier. The Double-Morse well and topological chaos. Chem. Phys.
Lett., 145:299–304, 1988.
57. S. Beri, R. Mannella, D. G. Luchinsky, A. N. Silchenko, and P. V. E. McClintock. Solution of
the boundary value problem for optimal escape in continuous stochastic systems and maps.
Phys. Rev. E, 72:036131, 2005.
58. M. V. Berry. Reflection on a Christmas-tree baule. Phys. Edu., 7:1–6, 1972.
59. M. V. Berry. Regularity and chaos in classical mechanics, illustrated by three deformations
of a circular “billiard”. Eur. J. Phys., 2:91–102, 1981.
60. M. V. Berry and K. E. Mount. Semiclassical approximations in wave mechanics. Reps. Prog.
Phys., 35:315–397, 1972.
61. J. J. B. Biemond, A. P. S. de Moura, G. Károlyi, C. Grebogi, and H. Nijmeijer. Onset of
chaotic advection in open flows. Phys. Rev. E, 78:016317, 2008.
62. F. A. Bignone. Cells-genes interaction simulation on a coupled map lattice. J. Theor. Biol.,
161:231–249, 1993.
63. O. Biham and W. Wenzel. Characterization of unstable periodic orbits in chaotic attractors
and repellers. Phys. Rev. Lett., 63:819–822, 1989.
64. L. Billings, E. M. Bollt, and I. B. Schwartz. Phase-space transport of stochastic chaos in
population dynamics of virus spread. Phys. Rev. Lett., 88:234101, 2002.
65. L. Billings and I. B. Schwartz. Exciting chaos with noise: unexpected dynamics in epidemic
outbreaks. J. Math. Biol., 44:31–48, 2002.
66. L. Billings, I. B. Schwartz, D. S. Morgan, E. M. Bollt, R. Meucci, and E. Allaria. Stochastic
bifurcation in a driven laser system: Experiment and theory. Phys. Rev. E, 70:026220, 2004.
67. J. P. Bird, R. Akis, and D. K. Ferry. Magnetoprobing of the discrete level spectrum of open
quantum dots. Phys. Rev. B, 60:13676–13681, 1999.
68. J. P. Bird, R. Akis, D. K. Ferry, A. P. S. de Moura, Y.-C. Lai, and K. M. Indlekofer. Interference
and interactions in open quantum dots. Rep. Prog. Phys., 66:583–632, 2003.
69. J. P. Bird, R. Akis, D. K. Ferry, D. Vasileska, J. Cooper, Y. Aoyagi, and T. Sugano. Lead-
orientation-dependent wave function scarring in open quantum dots. Phys. Rev. Lett.,
82:4691–4694, 1999.
70. J. A. Blackburn, N. Gronbech-Jensen, and H. J. T. Smith. Stochastic noise and chaotic tran-
sients. Phys. Rev. Lett., 74:908–911, 1995.
462 References

71. R. E. Blahut. Principles and Practice of Information Theory. Addison-Wesley, Reading, MA,
first edition, 1988.
72. S. Bleher, C. Grebogi, and E. Ott. Fractal boundaries for exit in Hamiltonian dynamics. Phys.
Rev. A, 38:930–938, 108.
73. S. Bleher, E. Ott, and C. Grebogi. Routes to chaotic scattering. Phys. Rev. Lett., 63:919–922,
1989.
74. S. Bleher, C. Grebogi, and E. Ott. Bifurcation to chaotic scattering. Physica D, 46:87–121,
1990.
75. R. Blümel. Quantum chaotic scattering. In B.-L. Hao, D.-H. Feng, and J.-M. Yuan, editors,
Directions in Chaos. World Scientific, Singapore, 1991.
76. R. Blümel and U. Smilansky. Classical irregular scattering and its quantum-mechanical im-
plications. Phys. Rev. Lett., 60:477–480, 1988.
77. R. Blümel and U. Smilansky. A simple model for chaotic scattering II. Quantum mechanical
theory. Physica D, 36:111–136, 1989.
78. R. Blümel and U. Smilansky. Random-matrix description of chaotic scattering: semiclassical
approach. Phys. Rev. Lett., 64:241–244, 1990.
79. S. Boccaletti, C. Grebogi, Y.-C. Lai, H. Mancini, and D. Maza. Control of chaos: theory and
applications. Phys. Rep., 329:103–197, 2000.
80. S. Boccaletti, J. Kurths, G. Osipov, D. L. Valladares, and C.-S. Zhou. The synchronization of
chaotic systems. Phys. Rep., 366:1–101, 2002.
81. T. Bódai, G. Károlyi, and T. Tél. Fractal snapshot components in chaos induced by strong
noise. Phys. Rev. E, 83:in print, 2011.
82. G. Boffetta, G. Lacorata, G. Redaelli, and A. Vulpiani. Detecting barriers to transport: a
review of different techniques. Physica D, 159:58–70, 2001.
83. T. Bohr and D. Rand. The entropy function for characteristic exponents. Phys. D, 25:387–398,
1987.
84. T. Bohr, P. Cvitanović, and M. H. Jensen. Fractal “aggregates” in the complex plane. Euro-
phys. Lett., 6:445–450, 1988.
85. T. Bohr, M. H. Jensen, G. Paladin, and A. Vulpiani. Dynamical Systems Approach to Turbu-
lence. Cambridge University Press, Cambridge, 1998.
86. T. Bohr, A. W. Pedersen, and M. H. Jensen. Transition to turbulence in a discrete Ginzburg-
Landau model. Phys. Rev. A, 42:3626–3629, 1990.
87. E. M. Bollt. The path towards a longer life: on invariant sets and the escape time landscape.
Int. J. Bif. Chaos, 15:1615–1624, 2005.
88. E. M. Bollt and M. Dolnik. Encoding information in chemical chaos by controlling symbolic
dynamics. Phys. Rev. E, 55:6404–6413, 1997.
89. E. M. Bollt and Y.-C. Lai. Dynamics of coding in communicating with chaos. Phys. Rev. E,
58:1724–1736, 1998.
90. E. M. Bollt, Y.-C. Lai, and C. Grebogi. Coding channel capacity and noise resistance in
communicating with chaos. Phys. Rev. Lett., 79:3787–3790, 1997.
91. R. Bonaccini and A. Politi. Chaotic-like behaviour in chains of stable nonlinear oscillators.
Physica D, 103:362–368, 1997.
92. C. Bonatto, J. A. C. Gallas, and Y. Ueda. Chaotic phase similarities and recurrences in a
damped-driven Duffing oscillator. Phys. Rev. E, 77:026217, 2008.
93. C. Bonatto, J. C. Garreau, and J. A. C. Gallas. Self-similarities in the frequency-amplitude
space of a loss-modulated CO2 laser. Phys. Rev. Lett., 95:143905, 2005.
94. F. A. Borotto, A. C. L. Chian, and E. L. Rempel. Alfvén interior crisis. Int. J. Bif. Chaos,
14:2375–2380, 2004.
95. D. Borrero-Echeverry, R. Tagg, and M. F. Schatz. Transient turbulence in Taylor-Couette
flow. Phys. Rev. E, 81:025301(R), 2010.
96. F. Bottausci, I. Mezic, C.D. Meinhart, and C. Cardonne. Mixing in the shear superposition
micromixer: three-dimensional analysis. Phil. Trans. Roy. Soc., 362:1001–1018, 2004.
97. P. T. Boyd and S. L. W. McMillan. Chaotic scattering in the gravitational three-body problem.
Chaos, 4:507–524, 1993.
References 463

98. R. Braun and F. Feudel. Supertransient chaos in the two-dimensional complex Ginzburg-
Landau equation. Phys. Rev. E, 53:6562–6565, 1996.
99. U. Brosa. Turbulence without strange attractor. J. Stat. Phys., 55:1303–1312, 1989.
100. R. Brown, P. Bryant, and H. D. I. Abarbanel. Computing the Lyapunov spectrum of a dynam-
ical system from an observed time series. Phys. Rev. A, 43:2787–2806, 1991.
101. H. Buljan and V. Paar. Many-hole interactions and the average lifetimes of chaotic transients
that precede controlled periodic motion. Phys. Rev. E, 63:066205, 2001.
102. H. Buljan and V. Paar. Naturally invariant measure of chaotic attractors and the conditionally
invariant measure of embedded chaotic repellers. Phys. Rev. E, 65:036218, 2002.
103. H. Buljan and V. Paar. Parry measure and the topological entropy of chaotic repellers embed-
ded within chaotic attractors. Physica D, 172:111–123, 2002.
104. A. R. Bulsara, W. C. Schieve, and E. W. Jacobs. Homoclinic chaos in systems perturbed by
weak Langevin noise. Phys. Rev. A, 41:668–681, 1990.
105. L.A. Bunimovich and C.P. Dettmann. Peeping at chaos: nondestructive monitoring of chaotic
systems by measuring long-time escape rates. Europhys. Lett., 80:40001, 2007.
106. K. Buszko and K. Stefanski. Measuring transient chaos in nonlinear one- and two-
dimensional maps. Chaos Sol. Fract., 27:630–646, 2006.
107. M. Büttiker. Four-terminal phase-coherent conductance. Phys. Rev. Lett., 57:1761–1764,
1986.
108. M. Büttiker. Absence of backscattering in the quantum Hall effect in multiprobe conductors.
Phys. Rev. B, 38:9375–9389, 1988.
109. T. Buzug and G. Pfister. Optimal delay time and embedding dimension for delay-time coor-
dinates by analysis of the glocal static and local dynamic behavior of strange attractors. Phys.
Rev. A, 45:7073–7084, 1992.
110. T. L. Carroll, L. M. Pecora, and F. J. Rachford. Chaotic transients and multiple attractors in
spin-wave experiments. Phys. Rev. Lett., 59:2891–2894, 1987.
111. T. L. Carroll, L. M. Pecora, and F. J. Rachford. Lyapunov exponents near a crisis in a spin-
wave experiment. Phys. Rev. A, 40:4149–4152, 1989.
112. T. L. Carroll, L. M. Pecora, and F. J. Rachford. Chaos and chaotic transients in an Yttrium
iron-garnet sphere. Phys. Rev. A, 40:377–386, 1989.
113. T. L. Carroll, F. J. Rachford, and L. M. Pecora. Occurrence of chaotic transients during
transitions between Quasiperiodic states in Yttrium iron-garnet. Phys. Rev. B, 38:2938–2940,
1988.
114. M.L. Cartwright and J.E. Littlewood. On Non-linear differential equations of the second
order: I. J. Lond. Math. Soc., 20:180–189, 1945.
115. J. H. E. Cartwright et al. Dynamics of finite-size particles in chaotic fluid flows. In M. Thiel
et al., editor, Nonlinear Dynamics and Chaos: Advances and Perspectives, pages 51–87.
Springer, Berlin, Heidelberg, 2010.
116. G. Casati, G. Maspero, and D. L. Shepelyansky. Quantum fractal eigenstates. Physica D,
131:311–316, 1999.
117. G. Casati, G. Maspero, and D. L. Shepelyansky. Quantum Poincaré recurrences. Phys. Rev.
Lett., 82:524–527, 1999.
118. F. Cecconi, R. Livi, and A. Politi. Fuzzy transition region in a one-dimensional coupled-
stable-map lattice. Phys. Rev. E, 57:2703–2712, 1998.
119. R. K. Chang and A. J. Campillo. Optical Processes in Microcavities. World Scientific,
Singapore, first edition, 1996.
120. H. Chaté and P. Manneville. Transition to turbulence via spatiotemporal intermittency. Phys.
Rev. Lett., 58:112–115, 1987.
121. H. Chaté and P. Manneville. Spatio-temporal intermittency in coupled map lattices. Physica
D, 32:409–422, 1988.
122. M. D. Chekroun, E. Simonnet, and M. Ghil. Stochastic climate dynamics: random attractors
and time-dependent invariant measures. Physica D, 2010.
123. G. Chen. Controlling Chaos and Bifurcations in Engineering Systems. CRC Press, Boca
Raton, FL, first edition, 1999.
464 References

124. Q. Chen, M. Ding, and E. Ott. Chaotic scattering in several dimensions. Phys. Lett. A,
145:93–100, 1990.
125. Q. Chen, E. Ott, and L. P. Hurd. Calculating topological entropies of chaotic dynamical
systems. Phys. Lett. A, 156:48–52, 1991.
126. A. C. L. Chian, E. L. Rempel, E. E. Macau, R. R. Rosa, and F. Christiansen. High-dimensional
interior crisis in the Kuramoto-Shivashinsky equation. Phys. Rev. E, 65:035203, 2002.
127. A. C. L. Chian, F. A. Borotto, and E. L. Rempel. Alfvén boundary crisis. Int. J. Bif. Chaos,
12:1653–1658, 2002.
128. A. C. L. Chian, F. A. Borotto, E. L. Rempel, and C. Rogers. Attractor merging crisis in chaotic
business cycles. Chaos Sol. Fract., 24:869–875, 2005.
129. A. C. L. Chian, E. L. Rempel, and C. Rogers. Complex economic dynamics: chaotic saddle,
crisis and intermittency. Chaos Sol. Fract., 29:1194–1218, 2006.
130. A. C. L. Chian, W. M. Santana, E. L. Rempel, F. A. Borotto, T. Hada, and Y. Kamide. Chaos
in driven Alfvén systems: unstable periodic orbits and chaotic saddles. Nolin. Proc. Geophys.,
14:17–19, 2007.
131. A. C.-L. Chian, R. A. Miranda, E. L. Rempel, Y. Saiki, and M. Yamada. Amplitude-phase
synchronization at the onset of permanent spatiotemporal chaos. Phys. Rev. Lett., 104:254102,
2010.
132. H.-D. Chiang, I. Dobson, R. J. Thomas, J. S. Thorp, and L. Fekih-Ahmed. On voltage collapse
in electric power systems. IEEE Trans. Power Syst., 5:601–607, 1990.
133. S. Muralidharan, K. R. Sreenivas and R. Govindarajan. Dynamical systems model of entrain-
ment due to coherent structures. Phys. Rev. E, 72:046308, 2005.
134. B. V. Chirikov and F. M. Izraelev. Degeneration of turbulence in simple systems. Phys. D,
2:30–37, 1981.
135. B. V. Chirikov and F. M. Izraelev. Some numerical experiments with a nonlinear mapping:
stochastic component. Colloques. Int. du CNRS (Toulouse, Sept. 1973), 229:409–428, 1976.
136. B. V. Chirikov and D. L. Shepelyansky. Correlation properties of dynamical chaos in Hamil-
tonian systems. Physica D, 13:395–400, 1984.
137. F. Christiansen and P. Grassberger. Escape and sensitive dependence on initial conditions in
a symplectic repeller. Phys. Lett. A, 181:47–53, 1993.
138. R. C. Churchill, G. Pecelli, and D. L. Rod. Isolated unstable periodic orbits. J. Diff. Equat.,
17:329–348, 1975.
139. C. C. Conley. Invariant sets in a monkey saddle. In W. A. Harris Jr. and Y. Sibuya, editors,
Proc. United States-Japan Seminar on Differential and Functional Equations, pages 443–447.
Benjamin, New York, 1967.
140. L. Cortelezzi and I. Mezić (eds.). Analysis and Control of Mixing with an Application to
Micro and Macro Flow Processes. Springer, New York, 2009.
141. J. Coste. Iterations of transformations on the unit interval – approach to a periodic attractor.
J. Stat. Phys., 23:521–536, 1980.
142. C. Coulliette, F. Lekien, J. D. Paduan, G. Haller, and J.E. Marsden. Optimal pollution mit-
igation in Monterey Bay based on coastal radar data and nonlinear dynamics. Environ. Sci.
Technol., 41:6562–6572, 2007.
143. A. Crisanti, M. Falconi, A. Vulpiani, and G. Paladin. Lagrangian chaos: transport, mixing
and diffusion in fluids. Riv. Nouvo Cimento, 14:1–80, 1991.
144. G. Cristadoro and R. Ketzmerick. Universality of algebraic decays in Hamiltonian systems.
Phys. Rev. Lett., 100:184101, 2008.
145. V. Croquette and C. Poitou. Cascade of period doubling and large-scale stochasticity from
compass movements. J. de Physique Lett., 292:1353–1356, 1981.
146. J. R. Crutchfield and K. Kaneko. Are attractors relevant to turbulence? Phys. Rev. Lett.,
60:2715–2718, 1988.
147. G. Csernák and G. Stépán. Life expentancy of transient microchaotic behaviour. J. Nonlin.
Sci., 15:63–91, 2005.
148. Y. Cuche, R. Livi, and A. Politi. Phase transitions in 2D linearly stable coupled map lattices.
Physica D, 103:369–380, 1997.
References 465

149. P. Cvitanović. Invariant measurements of strange sets in terms of cycles. Phys. Rev. Lett.,
61:2729–2732, 1988.
150. P. Cvitanović, N. Sondergaard, G. Palla, G. Vattay, and C. P. Dettmann. Spectrum of stochas-
tic evolution operators: Local matrix representation approach. Phys. Rev. E, 60:3936–3941,
1999.
151. P. Cvitanović, R. Artuso, P. Dahlquist, R. Mainieri, G. Tanner, and G. Vattay. Chaos: Classical
and Quantum, e-book: http://ChaosBook.org/chapters/ChaosBook.pdf, 2010.
152. P. Cvitanović and B. Eckhardt. Periodic-orbit quantization of chaotic systems. Phys. Rev.
Lett., 63:823–826, 1989.
153. P. Cvitanović and J. F. Gibson. Geometry of turbulence in wall-bounded shear flows: periodic
orbits. Phys. Scr., T142, 014007, 2010.
154. D. Dangoisse, P. Glorieux, and D. Hennequin. Laser chaotic attractors in crisis. Phys. Rev.
Lett., 57:2657–2660, 1986.
155. S. Datta. Electronic Transport in Mesoscopic Systems. Cambridge University Press, Cam-
bridge, first edition, 1995.
156. R. L. Davidchack and Y.-C. Lai. Efficient algorithm for detecting unstable periodic orbits in
chaotic systems. Phys. Rev. E, 60:6172–6175, 1999.
157. R. L. Davidchack, Y.-C. Lai, A. Klebanoff, and E. M. Bollt. Towards complete detection of
unstable periodic orbits in chaotic systems. Phys. Lett. A, 287:99–104, 2001.
158. R. L. Davidchack and Y.-C. Lai. Characterization of transition to chaos with multiple positive
Lyapunov exponents by unstable periodic orbits. Phys. Lett. A, 270:308–313, 2000.
159. S. Dawson, C. Grebogi, T. D. Sauer, and J. A. Yorke. Obstructions to shadowing when a
Lyapunov exponent fluctuates about zero. Phys. Rev. Lett., 73:1927–1930, 1994.
160. A. de Lozar and B. Hof. An experimental study of the decay of turbulent puffs in pipe flow.
Phil. Trans. R. Soc. A, 367:589–599, 2009.
161. A. P. S. de Moura. Strange nonchaotic repellers. Phys. Rev. E, 76:036218, 2007.
162. A. P. S. de Moura and C. Grebogi. Output functions and fractal dimensions in dynamical
systems. Phys. Rev. Lett., 86:2778–2781, 2001.
163. A. P. S. de Moura and C. Grebogi. Rainbow transition in chaotic scattering. Phys. Rev. E,
65:035206R, 2002.
164. A. P. S. de Moura and C. Grebogi. Countable and uncountable boundaries in chaotic scatter-
ing. Phys. Rev. E, 66:046214, 2002.
165. A. P. S. de Moura and C. Grebogi. Chemical and biological activity in three-dimensional
flows. Phys. Rev. E, 70:026218, 2004.
166. A. P. S. de Moura, Y.-C. Lai, J. P. Bird, R. Akis, and D. K. Ferry. Tunneling and nonhyper-
bolicity in quantum dots. Phys. Rev. Lett., 88:236804, 2002.
167. A. P. S. de Moura and P. S. Letelier. Mixmaster chaos. Phys. Lett. A, 285:127–131, 1999.
168. A. P. S. de Moura and P. S. Letelier. Chaos and fractals in geodesic motion around a nonro-
tating black hole with halos. Phys. Rev. E, 61:6506–6516, 2000.
169. A. P. S. de Moura and P. S. Letelier. Scattering map for two black holes. Phys. Rev. E,
62:4784–4791, 2000.
170. A. S. de Paula, M. A. Savi, and F. H. I. Pereira-Pinto. Chaos and transient chaos in an
experimental nonlinear pendulum. J. Sound and Vibration, 294:585–595, 2006.
171. S. L. T. de Souza, I. L. Caldas, R. L. Viana, A. M. Batista, and T. Kapitaniak. Noise-induced
basin hopping in a gearbox model. Chaos Sol. Fract., 26:1523–1531, 2005.
172. L. Delfini, S. Denisov, S. Lepri, R. Livi, P. K. Mohanty, and A. Politi. Energy diffusion in
hard-point systems. Eur. Phys. J. Special Topics, 146:21–35, 2007.
173. M. Dellnitz, G. Froyland, C. Honerkamp, and K. Pradberg. On the approximation of transport
phenomena – a dynamical systems approach. GAMM-Mitt., 32:47–60, 2009.
174. M. Dellnitz and O. Junge. Almost invariant sets in Chua’s circuit. Int. J. Bif. Chaos, 7:2475–
2485, 1997.
175. C. Dembowski, B. Dietz, T. Friedrich, H.-D. Gräf, A. Heine, C. Mejia-Monasterio, M. Miski-
Ogen, A. Richter, and T. H. Seligman. First experimental evidence for quantum echoes in
scattering systems. Phys. Rev. Lett., 93:134102, 2004.
466 References

176. H. Deng, Q. Deng, and D. G. Deppe. Native-oxide laterally confined whispering-gallery mode
laser with vertical emission. Appl. Phys. Lett., 69:3120–3122, 1996.
177. B. Dennis, R. A. Desharnais, J. M. Cushing, S. M. Henson, and R. F. Constantino. Can noise
induce chaos? Oikos, 102:329–339, 2003.
178. A. Destexhe. Self-sustained asynchronous irregular states and up-down states in thalamic,
cortical and thalamocortical networks of nonlinear integrate-and-fire neurons. J. Comput.
Neurosci., 27:493–506, 2009.
179. C. P. Dettmann. Recent advances in open billiards with some open problems. In Z. Elhadj and
J. C. Sprott, editors, Frontiers in the study of chaotic dynamical systems with open problems,
Singapore, 2010. World Scientific.
180. C. P. Dettmann and O. Georgiou. Survival probability for the stadium billiard. Phys. D,
238:2395–2403, 2009.
181. C. P. Dettmann and O. Georgiou. Transmission and reflection in the stadium billiard: time
reversal symmetry breaking. arXiv:1006.131, 2010.
182. C. P. Dettmann and T. B. Howard. Asymptotic expansions for the escape rate of stochastically
perturbed unimodal maps. Phys. D, 238:2404–2408, 2009.
183. C. P. Dettmann, G. V. Morozov, M. Sieber, and H. Waalkens. Unidirectional emission from
circular dielectric microresonators with a point scatterer. Phys. Rev. A, 80:063813, 2009.
184. R. Devaney and Z. Nitecki. Shift automorphisms in the Hénon mapping. Commun. Math.
Phys., 67:137–146, 1979.
185. M. Dhamala and Y.-C. Lai. Controlling transient chaos in deterministic flows with applica-
tions to electrical power systems and ecology. Phys. Rev. E, 59:1646–1655, 1999.
186. M. Dhamala and Y.-C. Lai. Unstable periodic orbits and the natural measure of nonhyperbolic
chaotic saddles. Phys. Rev. E, 60:6176–6179, 1999.
187. M. Dhamala and Y.-C. Lai. The natural measure of nonattracting chaotic sets and its repre-
sentation by unstable periodic orbits. Int. J. Bif. Chaos, 12:2991–3006, 2002.
188. M. Dhamala, Y.-C. Lai, and R. D. Holt. How often are chaotic transients in spatially extended
ecological systems? Phys. Lett., 280:297–302, 2001.
189. M. Dhamala, Y.-C. Lai, and E. J. Kostelich. Detecting unstable periodic orbits from transient
chaotic time series. Phys. Rev. E, 61:6485–6489, 2000.
190. M. Dhamala, Y.-C. Lai, and E. J. Kostelich. Analysis of transient chaotic time series. Phys.
Rev. E, 64:056207, 2001.
191. M. Ding. Topological entropy and bifurcation of chaotic scattering. Phys. Rev. A, 46:6247–
6251, 1992.
192. M. Ding, T. Bountis, and E. Ott. Algebraic escape in higher dimensional Hamiltonian sys-
tems. Phys. Letts. A, 151:395–400, 1990.
193. M. Ding, C. Grebogi, E. Ott, T. D. Sauer, and J. A. Yorke. Plateau onset for correlation
dimension: when does it occur? Phys. Rev. Lett., 70:3872–3875, 1993.
194. M. Ding, C. Grebogi, E. Ott, and J. A. Yorke. Transition to chaotic scattering. Phys. Rev. A,
42:7025–7040, 1990.
195. M. Ding, E. Ott, and C. Grebogi. Crisis control – preventing chaos-induced capsizing of a
ship. Phys. Rev. E, 50:4228–4230, 1994.
196. W. Ditto, S. Rauseo, R. Cawley, C. Grebogi, G. Hsu, E. Kostelich, E. Ott, H. Savage, R. Seg-
nan, M. Spano, and J. Yorke. Experimental observation of crisis-induced intermittency and
its critical exponent. Phys. Rev. Lett., 63:923–926, 1989.
197. Y. Do and Y.-C. Lai. Superpersistent chaotic transients in physical space: advective dynamics
of inertial particles in open chaotic flows under noise. Phys. Rev. Lett., 91:224101, 2003.
198. Y. Do and Y.-C. Lai. Extraordinarily superpersistent chaotic transients. Europhys. Lett.,
67:914–920, 2004.
199. Y. Do and Y.-C. Lai. Scaling laws for noise-induced superpersistent chaotic transients. Phys.
Rev. E, 71:046208, 2005.
200. I. Dobson and H.-D. Chiang. Towards a theory of voltage collapse in electric power systems.
Sys. Contr. Lett., 13:253–262, 1989.
201. J. R. Dorfmann. An Introduction to Chaos in Nonequilibrium Statistical Mechanics. Cam-
bridge University Press, Cambridge, first edition, 1999.
References 467

202. E. Doron and U. Smilansky. Chaotic spectroscopy. Phys. Rev. Lett, 68:1255–1258, 1992.
203. E. Doron and U. Smilansky. Chaotic spectroscopy. Chaos, 2:117–124, 1992.
204. E. Doron, U. Smilansky, and A. Frenkel. Experimental demonstration of chaotic scattering of
microwaves. Phys. Rev. Lett., 65:3072–3075, 1990.
205. E. Doron, U. Smilansky, and A. Frenkel. Chaotic scattering and transmission fluctuations.
Physica D, 50:367–390, 1991.
206. M. V. Dyke. An Album of Fluid Motion. The Parabolic Press, Stanford, 1982.
207. M. I. Dykman, I. B. Schwartz, and M. Shapiro. Scaling in activated escape of underdamped
systems. Phys. Rev. E, 72:021102, 2005.
208. F. J. Dyson. Statistical theory of energy levels of complex systems (I). J. Math. Phys., 3:140,
1962.
209. A. Eberspächer, J. Main, and G. Wunner. Fractal Weyl law for three-dimensional chaotic
hard-sphere scattering systems. Phys. Rev. E, 82:046201, 2010.
210. B. Eckhardt. Fractal properties of scattering singularities. J. Phys. A: Math. Gen., 20:5971–
5979, 1987.
211. B. Eckhardt. Irregular scattering. Physica D, 33:89–98, 1988.
212. B. Eckhardt. Quantum mechanics of classically non-integrable systems. Phys. Rep., 163:205–
297, 1988.
213. B. Eckhardt. Turbulence transition in pipe flow: some open questions. Nonlinearity, 21:T1–
T11, 2008.
214. B. Eckhardt and H. Aref. Integrable and chaotic motions of 4 vortices. 11. Collision dynamics
of vortex pairs. Philos. T. Roy. Soc. A, 326:655–696, 1988.
215. B. Eckhardt, H. Faisst, A. Schmiegel, and T. M. Schneider. Dynamical systems and the
transition to turbulence in linearly stable shear flows. Philo. Trans. Roy. Soc. A, 366:1297–
1315, 2008.
216. B. Eckhardt and T. M. Schneider. How does flow in a pipe become turbulent? Eur. Phys. J,
64:457–462, 2008.
217. B. Eckhardt, T. M. Schneider, B. Hof, and J. Westerweel. Turbulence in pipe flows. Annu.
Rev. Fluid. Mech., 39:447–468, 2007.
218. J.-P. Eckman, L. Thomas, and P. Wittwer. Intermittency in the presence of noise. J. Phys. A:
Math. Gen., 14:3153–3168, 1981.
219. J.-P. Eckmann, S. O. Kamphorst, D. Ruelle, and S. Ciliberto. Lyapunov exponents from time
series. Phys. Rev. A, 34:4971–4979, 1986.
220. J.-P. Eckmann and D. Ruelle. Ergodic theory of chaos and strange attractors. Rev. Mod. Phys.,
57:617–656, 1985.
221. S. P. Ellner and P. Turchin. When can noise induce chaos and why does it matter: a critique.
Oikos, 111:620–631, 2005.
222. R. Engbert and F. R. Drepper. Chance and chaos in population biology - Models of recurrent
epidemics and food chain dynamics. Chaos Sol. Fract., 4:1147–1169, 1994.
223. I. R. Epstein. The consequence of imperfect mixing in chemical and biological systems.
Nature, 374:321–327, 1995.
224. T. Ericson. Fluctuations of nuclear cross sections in the continuum region. Phys. Rev. Lett.,
5:430–431, 1960.
225. L. Ermann and D.L. Shepelyansky. Ulam method and fractal Weyl law for Perron-Frobenius
operators. arXiv:0912.508302, 2010.
226. L. Ermann, G. G. Carlo, and M. Saraceno. Localization of resonance eigenfunctions on
quantum repellers. Phys. Rev. Lett., 103:054102, 2009.
227. T. Erneaux. Applied Delay Differential Equation. Springer, Berlin, first edition, 2009.
228. S. V. Ershov and A. B. Potapov. On the nature of nonchaotic turbulence. Phys. Lett. A,
167:60–64, 1992.
229. H. Faisst and B. Eckhardt. Traveling waves in pipe flow. Phys. Rev. Lett., 91:224502, 2003a.
230. H. Faisst and B. Eckhardt. Lifetimes of noisy repellors. Phys. Rev. E, 68:026215, 2003b.
231. H. Faisst and B. Eckhardt. Sensitive dependence on initial conditions in transition to turbu-
lence in pipe flow. J. Fluid. Mech., 504:343–352, 2004.
468 References

232. K. Falconer. The Geometry of Fractal Sets. Cambridge University Press, Cambridge, 1985.
233. G. Falkovich, A. Fouxon, and M. G. Stepanov. Acceleration of rain initiation by cloud turbu-
lence. Nature, 419:151–154, 2002.
234. J. D. Farmer. Sensitive dependence on parameters in nonlinear dynamics. Phys. Rev. Lett.,
55:661–664, 1985.
235. M. J. Feigenbaum. Quantitative universality for a class of nonlinear transformations. J. Stat.
Phys., 19:25–52, 1978.
236. M. J. Feigenbaum. The universal metric properties of nonlinear transformations. J. Stat.
Phys., 21:669–706, 1979.
237. W. Feller. An Introduction to Probability Theory and Its Applications. Wiley, New York, first
edition, 1966.
238. D. K. Ferry and S. M. Goodnick. Transport in Nanostructures. Cambridge University Press,
Cambridge, UK, first edition, 1997.
239. U. Feudel and C. Grebogi. Multistability and the control of complexity. Chaos, 7:301–310,
1997.
240. U. Feudel, C. Grebogi, B. R. Hunt, and J. A. Yorke. Map with more than 100 coexisting
low-periodic atractors. Phys. Rev. E, 54:71–81, 1996.
241. U. Feudel, S. Kuznetsov, and A. Pikovsky. Strange Nonchaotic Attractors. World Scientific,
Singapore, 2006.
242. S. Fishman, D. R. Grempel, and R. E. Prange. Temporal crossover from classical to quantal
behavior near dynamical critical points. Phys. Rev. A, 36:289–305, 1987.
243. M. Franaszek. Influence of noise on the mean lifetime of chaotic transients. Phys. Rev. A,
44:4065–4067, 1991.
244. M. Franaszek and L. Fronzoni. Influence of noise on crisis-induced intermittency. Phys. Rev.
E, 49:3888– 3893, 1994.
245. W. J. Freeman. Neural networks and chaos. J. Theor. Biol., 171:13–18, 1994.
246. M. I. Freidlin and A.D. Wentzell. Random Perturbations of Dynamical Systems. Springer,
New York, 1984.
247. J. F. Freire, C. Bonatto, C. C. DaCamara, and J. A. C. Gallas. Multistability, phase dia-
grams and intransitivity in the Lorenz-84 low-order atmospheric circulation model. Chaos,
18:033121, 2008.
248. G. Froyland and K. Pradberg. Almost invariant sets and invariant manifold – connecting prob-
abilistic and geometric descriptions of coherent structures in flows. Phys. D, 238:1507–1523,
2009.
249. G. Froyland, S. Lloyd, and N. Santitissadeekorn. Coherent sets for nonautonomous dynamical
systems. Phys. D, 239:1527–1541, 2010.
250. G. Froyland, K. Pradberg, M.H. England, and A.M. Treguier. Detection of coherent oceanic
structures via transfer operators. Phys. Rev. Lett., 98:224503, 2007.
251. G. Froyland, N. Santitissadeekorn, and A. Monahan. Transport in time-dependent dynamical
systems: Finite-time coherent sets. Chaos, 20:043116, 2010.
252. H. Fujisaka, K. Egami, and T. Yamada. Glassy dynamics in a spatially distributed dynamical
system. Phys. Lett., 174:103–110, 1993.
253. J. A. C. Gallas. Structure of the parameter space of the Hénon map. Phys. Rev. Lett., 70:2714–
2717, 1993.
254. J. A. C. Gallas. Dissecting shrimps: results for some one-dimensional physical models. Phys-
ica A, 202:196–223, 1994.
255. J. A. C. Gallas, C. Grebogi, and J. Yorke. Vertices in parameter space – double crises which
destroy chaotic attractors. Phys. Rev. Lett., 71:1359–1362, 1993.
256. C. W. Gardiner. Handbook of Stochastic Methods. Springer, New York, first edition, 1997.
257. P. Gaspard. Chaos, Scattering and Statistical Mechanics. Cambridge University Press, Cam-
bridge, UK, first edition, 1998.
258. P. Gaspard and G. Nicolis. Transport properties, Lyapunov exponents, and entropy per unit
time. Phys. Rev. Lett, 65:1693–1696, 1990.
259. P. Gaspard and S. Rice. Semiclassical quantization of the scattering from a classically chaotic
repellor. J. Chem. Phys., 90:2242–2254, 1989.
References 469

260. P. Gaspard and S. A. Rice. Scattering from a classically chaotic repellor. J. Chem. Phys.,
90:2225–2241, 1989.
261. M. Ghil, M. D. Chekroun, and E. Simonnet. Climate dynamics and fluid mechanics: natural
variability and related uncertainties. Physica D, 327:2111–2126, 2008.
262. G. Giacomelli, S. Lepri, and A. Politi. Statistical properties of bidimensional patterns gener-
ated from delayed and extended maps. Phys. Rev. E, 51:3939–3944, 1995.
263. J. F. Gibson, J. Halcrow, and P. Cvitanović. Visualizing the geometry of state space in plane
Couette flow. J. Fluid. Mech, 611:107, 2008.
264. F. Ginelli, R. Livi, and A. Politi. Emergence of chaotic behaviour in linearly stable systems.
J. Phys. A, 35:499–516, 2002.
265. C. Gmachl, F. Capasso, E. E. Narimanov, J. U. Nöckel, A. D. Stone, J. Faist, and D. L.
Sivco. High-power directional emission from microlasers with chaotic resonators. Science,
280:1556–1564, 1998.
266. A. L. Goldberger. Nonlinear dynamics, fractals and chaos – applications to cardiac electro-
physiology. Ann. Biomed. Eng., 18:195–198, 1990.
267. N. Goldenfeld, N. Guttenberg, and G. Gioia. Extreme fluctuations and the finite lifetime of
the turbulent state. Phys. Rev. E, 81:035304(R), 2010.
268. H. Goldstein. Classical Mechanics. Addison-Wesley, Reading, MA, second edition, 1993.
269. J. P. Gollub. Order and disorder in fluid motion. Proc. Natl. Acad. Sci. USA, 92:6705–6711,
1995.
270. R. Gomulkiewicz and R. D. Holt. When does evolution by natural-selection prevent extinc-
tion. Evolution, 49:201–207, 1995.
271. C. L. Goodridge, L. M. Pecora, T. L. Carroll, and F. J. Rachford. Detecting functional rela-
tionships between simultaneous time series. Phys. Rev. E, 64:026221, 2001.
272. B. Goodwine and G. Stépán. Controlling unstable rolling phenomena. J. Vib. Control, 6:137–
158, 2000.
273. M. Gorman, P. Widmann, and K. Robbins. Chaotic flow regimes in a convection loop. Phys.
Rev. Lett., 52:2241–2244, 1984.
274. M. Gorman, P. Widmann, and K. Robbins. Nonlinear dynamics of a convection loop – a
quantitative comparison of experiment with theory. Physica D, 19:255–267, 1986.
275. L. Gottdiener and R. Carmona. Energy-transfer in high-energy atom-oscillator collisions. J.
Chem. Phys., 70:4207–4211, 1979.
276. E. Gouillart, O. Dauchot, J.-L. Thiffeault, and S. Roux. Open-flow mixing: experimental
evidence for strange eigenmodes. Phys. Fluids, 21:023603, 2009.
277. A. Goussev and K. Richter. Loschmidt-echo decay from local boundary perturbations. Phys.
Rev. E, 75:015201(R), 2007.
278. R. Govindarajan. Universal behavior of entrainment due to coherent structures in turbulent
shear flows. Phys. Rev. Lett., 88:134503, 2002.
279. R. Graham. Macroscopic potentials, bifurcations and noise in dissipative systems. In F. Moss
and P. V. E. McClintock, editors, Noise in Nonlinear Dynamical Systems, Vol. 1, pages 225–
278. Cambridge University Press, Cambridge, 1989
280. R. Graham and A. Hamm. Non-equilibrium potentials for dynamical systems weakly per-
turbed by noise. In G. Györgyi, I. Kondor, L. Sasvári, and T. Tél, editors, From Phase
Transitions to Chaos, pages 449–469. World Scientific, Singapore, 1992
281. R. Graham, A. Hamm, and T. Tél. Nonequilibrium potentials for dynamical systems with
fractal attractors and repellers. Phys. Rev. Lett., 66:3089–3092, 1991.
282. R. Graham and T. Tél. Existence of a potential for dissipative dynamical systems. Phys. Rev.
Lett., 52:9–12, 1984.
283. P. Grassberger. Information flow and maximum entropy measures for 1-D maps. Physica D,
14:365–373, 1985.
284. P. Grassberger. Do climatic attractors exist? Nature, 323:609–612, 1986.
285. P. Grassberger. Noise-induced escape from attractors. J. Phys. A, 22:3283–3290, 1989.
286. P. Grassberger and H. Kantz. Universal scaling of long-time tails in Hamiltonian systems.
Phys. Lett. A, 113:167–171, 1985.
470 References

287. P. Grassberger and I. Procaccia. Measuring the strangeness of strange attractors. Physica D,
9:189–208, 1983.
288. P. Grassberger and I. Procaccia. Dimensions and entropies of strange attractors from a fluc-
tuating dynamics approach. Physica D, 13:34–54, 1984.
289. C. Grebogi, S. W. McDonald, E. Ott, and J. A. Yorke. Final state sensitivity: an obstruction
to predictability. Phys. Lett. A, 99:415–418, 1983.
290. C. Grebogi, E. Ott, F. Romeiras, and J. Yorke. Critical exponents for crisis-induced intermit-
tency. Phys. Rev. A, 36:5365–5380, 1987.
291. C. Grebogi, E. Ott, and J. Yorke. Critical exponent of chaotic transients in nonlinear
dynamical systems. Phys. Rev. Lett., 57:1284–1287, 1986.
292. C. Grebogi, E. Ott, and J. A. Yorke. Chaotic attractors in crisis. Phys. Rev. Lett, 48:1507–
1510, 1982.
293. C. Grebogi, E. Ott, and J. A. Yorke. Crises, sudden changes in chaotic attractors and chaotic
transients. Physica D, 7:181–200, 1983.
294. C. Grebogi, E. Ott, and J. A. Yorke. Fractal basin boundaries, long-lived chaotic transients,
and unstable-unstable pair bifurcation. Phys. Rev. Lett., 50:935–938, 1983.
295. C. Grebogi, E. Ott, and J. A. Yorke. Super persistent chaotic transients. Ergod. Th. Dym.
Syst., 5:341–372, 1985.
296. C. Grebogi, E. Ott, and J. A. Yorke. Basin boundary metamorphosis: changes in accessible
boundary orbits. Physica D, 24:243–262, 1987.
297. C. Grebogi, E. Ott, and J. A. Yorke. Unstable periodic orbits and the dimensions of multifrac-
tal chaotic attractors. Phys. Rev. A, 37:1711–1724, 1988.
298. M. A. Green, C. W. Rowley, and G. Haller. Detection of Lagrangian coherent structures in
3D turbulence. J. Fluid Mech., 572:111–120, 2007.
299. S. Grossmann. The onset of shear flow turbulence. Rev. Mod. Phys., 72:603–618, 2000.
300. J. Guckenheimer and P. J. Holmes. Nonlinear Oscillations, Dynamical Systems, and Bifurca-
tions of Vector Fields. Springer, New York, 1983.
301. M. C. Gutzwiller. Chaos in Classical and Quantum Mechanics. Springer, New York, first
edition, 1990.
302. G. Györgyi and P. Ruján. Strange attractors in disordered systems. J. Phys. C, 17:4207–4212,
1984.
303. J. Halcrow, J. F. Gibson, and P. Cvitanović. Heteroclinic connections in plane Couette flow.
J. Fluid. Mech, 621:365, 2009.
304. G. Haller. Distinguished material surfaces and coherent structures in three-dimensional fluid
flows. Physica D, 149:248–277, 2001.
305. G. Haller. Lagrangian structures and the rate of strain in a partition of two-dimensional
turbulence. Phys. Fluids, 13:3365–3385, 2001.
306. G. Haller. Exact theory of unsteady separation for two-dimensional flows. J. Fluid Mech.,
512:257–311, 2004.
307. G. Haller and A. C. Poje. Finite time transport in aperiodic flows. Physica D, 119:352–380,
1998.
308. G. Haller and T. Sapsis. Where do inertial particles go in fluid flows? Physica D, 237:573–
583, 2008.
309. G. Haller and T. Sapsis. Localized instability and attraction along invariant manifolds. SIAM
J. Appl. Dyn. Syst., 9:611–633, 2010.
310. G. Haller and G. Yuan. Lagrangian coherent structures and mixing in two-dimensional turbu-
lence. Physica D, 147:352–370, 2000.
311. A. Hamm and R. Graham. Noise-induced attractor explosions near tangent bifurcations. J.
Stat. Phys., 66:689–725, 1992.
312. A. Hamm, T. Tél, and R. Graham. Noise-induced attractor explosions near tangent bifurca-
tions. Phys. Lett. A, 185:313–320, 1994.
313. G. Handke. Fractal dimensions in the phase space of two-electron atoms. Phys. Rev. A,
50:R3561–R3564, 1994.
314. P. Hänggi, P. Talkner, and M. Borkovec. Reaction-rate theory: fifty years after Kramers. Rev.
Mod. Phys., 62:251–342, 1990.
References 471

315. J. D. Hanson, J. Cary, and J. D. Meiss. Algebraic decay in self-similar Markov chains. J. Stat.
Phys., 39:327–345, 1985.
316. A. Hastings and K. Higgins. Persistence of transients in spatially structured ecological mod-
els. Science, 263:1133–1136, 1994.
317. A. Hastings, C. L. Hom, S. Ellner, P. Turchin, and H. C. J. Godfray. Chaos in ecology – is
mother nature a strange attractor. Ann. Rev. Ecol. Systemat., 24:1–33, 1993.
318. S. Hayes, C. Grebogi, and E. Ott. Communicating with chaos. Phys. Rev. Lett., 70:3031–
3034, 1993.
319. S. Hayes, C. Grebogi, E. Ott, and A. Mark. Experimental control of chaos for communication.
Phys. Rev. Lett., 73:1781–1784, 1994.
320. K. He and A. C.-L. Chian. Critical dynamic events at the crisis of transition to spatiotemporal
chaos. Phys. Rev. E, 69:026207, 2004.
321. Y. He, Y.-M. Jiang, Y. Shen, and D.-R. He. Crisis induced by an escape from a fat strange set.
Phys. Rev. E, 70:056213, 2004.
322. J. F. Heagy, T. L. Carroll, and L. M. Pecora. Experimental and numerical evidence for riddled
basins in coupled chaotic systems. Phys. Rev. Lett., 73:3528–3531, 1994.
323. J. F. Heagy and J. M. Yuan. Dynamics of an impulsively driven Morse oscillator. Phys. Rev.
A, 41:571–581, 1990.
324. R. Hegger, H. Kantz, and R. Schreiber. TISEAN. http://www.mpipks-
dresden.mpg.de/ tisean/TISEAN3.01/ index.html, Dresden, e-book edition, 2007.
325. M. Hénon. A two-dimensional mapping with a strange attractor. Commun. Math. Phys.,
50:69–77, 1976.
326. M. Hénon. Chaotic scattering modeled by an inclined billiard. Physica D, 33:132–156, 1988.
327. W. K. Hensinger, H. Häffer, A. Browaeys, N. R. Heckenberg, K. Helmerson, C. McKenzie,
G. J. Milburn, W. D. Phillips, S. L. Rolston, H. Rubinsztein-Dunlop, and B. Upcroft. Dynam-
ical tunnelling of ultracold atoms. Nature, 412:52–55, 2001.
328. D. Hensley. The Hausdorff dimensions of some continued-fraction cantor sets. J. Number
Theor., 33:182–198, 1989.
329. D. Hensley. Continued-fraction Cantor sets, Hausdorff dimension, and functional-analysis. J.
Number Theor., 40:336–358, 1992.
330. D. Hensley. A polynomial time algorithm for the Hausdorff dimension of continued fraction
Cantor sets. J. Number Theor., 58:9–45, 1996.
331. D. Hensley. Continued Fractions. World Scientific, Singapore, first edition, 2006.
332. H. Herzel, W. Ebeling, and T. Schulmeister. Nonuniform chaotic dynamics and effects of
noise in biochemical systems. Z. Naturforsch., 42a:136–142, 1987.
333. R. Hilborn. Quantitative measurement of the parameter dependence of the onset of a crisis in
a driven nonlinear oscillator. Phys. Rev. A, 31:378–382, 1985.
334. B. Hof, A. de Lozar, D.J. Kuik, and J. Westerweel. Repeller or attractor? Selecting the dy-
namical model for the onset of turbulence in pipe flow. Phys. Rev. Lett., 101:214501, 2009.
335. B. Hof, C. W. H. Dorne, J. Westerweel, F. T. M. Nieuwstadt, and H. Faisst. Experimental
observation of nonlinear traveling wave in turbulent pipe flow. Science, 305:1594–1598,
2004.
336. B. Hof, J. Westerweel, T. M. Schneider, and B. Eckhardt. Finite lifetime of turbulence in
shear flows. Nature, 443:59–62, 2006.
337. P. Holmes. Non-linear oscillator with a strange attractor. Philos. T. Roy. Soc. A, 292:419–448,
1979.
338. P. Holmes and D. Whitley. On the attracting set for Duffing equation.2. A geometrical model
for moderate force and damping. Physica D, 7:111–123, 1983.
339. R. D. Holt and M. A. McPeek. Chaotic population dynamics favors the evolution of dispersal.
Am. Nat., 148:709–718, 1996.
340. S. M. Houghton, E. Knobloch, S. M. Tobias, and M. R. E. Proctor. Transient spatio-temporal
chaos in the complex Ginzburg-Landau equation on long domains. Phys. Lett. A, 374:2030–
2034, 2010.
341. G.-H. Hsu, E. Ott, and C. Grebogi. Strange saddles and the dimensions of their invariant
manifolds. Phys. Lett. A, 127:199–204, 1988.
472 References

342. L. Huang, Y.-C. Lai, D. K. Ferry, R. Akis, S. M. Goodnick, and R. Akis. Transmission and
scarring in graphene quantum dots. J. Phys. Cond. Mat., 21:344203, 2009.
343. L. Huang, Y.-C. Lai, D. K. Ferry, S. M. Goodnick, and R. Akis. Relativistic quantum scars.
Phys. Rev. Lett., 103:054101, 2009.
344. G. Huber, P. Alstrom, and T. Bohr. Nucleation and transients at the onset of vortex turbulence.
Phys. Rev. Lett., 69:2380–2383, 1992.
345. L. Hufnagel, R. Ketzmerick, and M. Weiss. Conductance fluctuations of generic billiards:
fractal or isolated? Europhys. Lett., 54:703–708, 2001.
346. B. R. Hunt, E. Ott, and E. Rosa. Sporadically fractal basin boundaries of chaotic systems.
Phys. Rev. Lett., 82:3597–3600, 1999.
347. B. R. Hunt, E. Ott, and J. A. Yorke. Fractal dimensions of chaotic saddles of dynamical
systems. Phys. Rev. E, 54:4819–4823, 1996.
348. J. Hyman, B. Nicolaenko, and S. Zaleski. Order and complexity in the Kuramoto-
Shivashinsky model of weakly turbulent interfaces. Physica D, 23:265–292, 1986.
349. M. Iansiti, Q. Hu, R. M. Westervelt, and M. Tinkham. Noise and chaos in a fractal basin
boundary regime of a Josephson junction. Phys. Rev. Lett., 55:746–749, 1985.
350. V. In, S.E. Mahan, W.I. Ditto, and M.L. Spano. Experimental maintenance of chaos. Phys.
Rev. Lett., 74:4420–4423, 1995.
351. V. In, M.L. Spano, and M. Ding. Maintaining chaos in high dimensions. Phys. Rev. Lett.,
80:700–703, 1998.
352. J. Jacobs, E. Ott, T. Antonsen, and J. Yorke. Modeling fractal entrainment sets of tracers
advected by chaotic temporally irregular fluid flows using random maps. Physica D, 110:1–
17, 1997.
353. J. Jacobs, E. Ott, and C. Grebogi. Computing the measure of nonattracting chaotic sets.
Physica D, 110:1–11, 1997.
354. J. Jacobs, E. Ott, and B.R. Hunt. Calculating topological entropy for transient chaos with an
application to communicating with chaos. Phys. Rev. E, 57:6577–6588, 1998.
355. R. A. Jalabert, H. U. Baranger, and A. D. Stone. Conductance fluctuations in the ballistic
regime: a probe of quantum chaos. Phys. Rev. Lett., 65:2442–2445, 1990.
356. I. M. Jánosi, L. Flepp, and T. Tél. Exploring transient chaos in an NMR-laser experiment.
Phys. Rev. Lett., 73:529–532, 1994.
357. I. M. Jánosi and T. Tél. Time-series analysis of transient chaos. Phys. Rev. E, 49:2756–2763,
1994.
358. O. Jenkinson and M. Pollicott. Computing the dimension of dynamically defined sets: E-2
and bounded continued fractions. Ergod. Theor. Dyn. Syst., 21:1429–1445, 2001.
359. Y.-M. Jiang, Y.-Q. Lu, X.-G. Chao, and D.-R. He. A crisis of a stochastic web. Eur. Phys. J.
D, 29:285–292, 2004.
360. C. Jung. Poincaré map for scattering states. J. Phys. A-Math. Gen., 19:1345–1353, 1986.
361. C. Jung and A. Emmanouilidou. Construction of natural partition of incomplete horseshoes.
Chaos, 15:023101, 2005.
362. C. Jung, C. Lipp, and T. H. Seligman. The inverse scattering problem for chaotic Hamiltonian
systems. Ann. Phys., 275:151–189, 1999.
363. C. Jung, C. Mejia-Monasterio, O. Merlo, and T. H. Seligman. Self-pulsing effect in chaotic
scattering. New J. Phys., 6:48, 2004.
364. C. Jung, C. Mejia-Monasterio, and T. H. Seligman. Scattering one step from chaos. Phys.
Lett. A, 198:306–314, 1995.
365. C. Jung, C. Mejia-Monasterio, and T. H. Seligman. Quantum and classical echoes in scatter-
ing systems described by simple Smale horseshoes. Europhys. Lett., 55:616–622, 2001.
366. C. Jung, O. Merlo, T. H. Seligman, and W. P. K. Zapfe. The chaotic set and the cross section
for chaotic scattering in three degrees of freedom. New J. Phys., 12:103021, 2010.
367. C. Jung, G. Orellana-Riuadeneyra, and G. A. Luna-Acosta. Reconstruction of the chaotic set
from classical cross section data. J. Phys. A-Math. Gen., 38:567–585, 2005.
368. C. Jung and S. Pott. Classical cross-section for chaotic potential scattering. J. Phys. A-Math.
Gen., 22:2925–2938, 1989.
References 473

369. C. Jung and T. Tél. Dimension and escape rate of chaotic scattering from classical and
semiclassical cross section data. J. Phys. A-Math. Gen., 24:2793–2805, 1991.
370. C. Jung, T. Tél, and E. Ziemniak. Application of scattering chaos to particle transport in a
hydrodynamical flow. Chaos, 3:555–568, 1993.
371. C. Jung and E. Ziemniak. Hamiltonian scattering chaos in a hydrodynamical system. J. Phys.
A Math. Gen., 25:3929–3943, 1992.
372. M. Kac. Probability and Related Topics in Physical Sciences. Interscience, New York, 1959.
373. L. P. Kadanoff and C. Tang. Escape from strange repellers. Proc. Natl. Acad. Sci. USA,
81:1276–1279, 1984.
374. E. T. Kai, V. Rossi, J. Sudre, H. Weimerskirch, C. López E. Hernández-Garcia, F. Marsac,
and V. Garcon. Top marine predators track Lagrangian coherent structures. Proc. Nat. Aca.
Sci. USA, 106:8245–8250, 2009.
375. K. Kaneko. Spatiotemporal intermittency in coupled map lattices. Prog. Theor. Phys.,
74:1033–1044, 1985.
376. K. Kaneko. Pattern dynamics in spatiotemporal chaos. Physica D, 34:1–41, 1989.
377. K. Kaneko. Supertransients, spatiotemporal intermittency amd stability of fully developed
spatiotemporal chaos. Phys. Lett. A, 149:105–112, 1990.
378. K. Kaneko. Theory and Applications of Coupled Map Lattices. Wiley, Chichester, 1993.
379. K. Kaneko and I. Tsuda. Complex Systems: Chaos and Beyond, a Constructive Approach with
Applications in Life Sciences. Springer, Berlin, 2000.
380. H. Kantz and P. Grassberger. Repellers, semi-attractors, and long-lived chaotic transients.
Physica D, 17:75–86, 1985.
381. H. Kantz and P. Grassberger. Chaos in low-dimensional Hamiltonian maps. Phys. Lett. A,
123:437–443, 1987.
382. H. Kantz and T. Schreiber. Nonlinear Time Series Analysis. Cambridge University Press,
Cambridge, UK, first edition, 1997.
383. T. Kapitaniak and J. Brindley. Preserving transient chaos. Phys. Lett. A, 241:41–45, 1998.
384. T. Kapitaniak, Y.-C. Lai, and C. Grebogi. Metamorphosis of chaotic saddle. Phys. Lett. A,
259:445–450, 1999.
385. A. Kaplan, N. Friedman, M. Andersen, and N. Davidson. Observation of islands of stability
in soft wall atom-optics billiards. Phys. Rev. Lett., 87:274101, 2001.
386. J. L. Kaplan and J. A. Yorke. Preturbulence: A regime observed in a fluid flow model of
Lorenz. Comm. Math. Phys., 67:93–108, 1979.
387. R. Kapral, R. Livi, G. L. Oppo, and A. Politi. Dynamics of complex interfaces. Phys. Rev. E,
49:2009–2022, 1994.
388. C. F. F. Karney. Long-time correlations in the stochastic regime. Physica D, 8:360–380, 1983.
389. G. Karniakidis, A. Beskok, and N. Aluru. Microflows and Nanoflows. Springer, Berlin, first
edition, 2005.
390. G. Károlyi. Fractal scaling of microbial colonies affects growth. Phys. Rev. E, 71:031915,
2005.
391. G. Károlyi, Á. Péntek, I. Scheuring, T. Tél, and Z. Toroczkai. Chaotic flow: the physics of
species coexistence. Proc. Nat. Aca. Sci. USA, 97:13661–13665, 2000.
392. G. Károlyi, Á. Péntek, Z. Toroczkai, T. Tél, and C. Grebogi. Chemical or biological activity
in open chaotic flows. Phys. Rev. E, 59:5468–5481, 1999.
393. G. Károlyi, I. Scheuring, and T. Czárán. Metabolic network dynamics in a open chaotic flows.
Chaos, 12:460–469, 2002.
394. G. Károlyi and T. Tél. Chaotic tracer scattering and fractal basin boundaries in a blinking
vortex-sink system. Phys. Rep., 290:125–147, 1997.
395. G. Károlyi and T. Tél. Chemical transients in closed chaotic flows: the role of effective
dimensions. Phys. Rev. Lett., 95:264501, 2005.
396. G. Károlyi and T. Tél. Effective dimension and chemical reactions in fluid flows. Phys. Rev.
E, 76:046315, 2007.
397. G. Károlyi, T. Tél, A. P. S. de Moura, and C. Grebogi. Reactive processes in random flows.
Phys. Rev. Lett., 92:174101, 2004.
474 References

398. Z. Kaufmann and H. Lustfeld. Comparision of averages of flows and maps. Phys. Rev. E,
64:055206(R), 2001.
399. Z. Kaufmann, H. Lustfeld, A. Németh, and P. Szépfalusy. Diffusion in normal and critical
transient chaos. Phys. Rev. Lett., 78:4031–4034, 1997.
400. Z. Kaufmann, A. Németh, and P. Szépfalusy. Critical states of transient chaos. Phys. Rev. E,
61:2543–2550, 2000.
401. R. L. Kautz. Activation energy for thermally induced escape from a basin of attraction. Phys.
Lett. A, 125:315–319, 1987.
402. R. L. Kautz. Global stability of phase lock near a chaotic crisis in the RF-biased Josephson
junction. J. Appl. Phys., 62:198–211, 1987.
403. J.P. Keating, M. Novaes, S.D. Prado, and M. Sieber. Semiclassical structure of chaotic reso-
nance eigenfunctions. Phys. Rev. Lett., 97:150406, 2006.
404. J. D. Keeler and J. D. Farmer. Robust space-time intermittency and 1/ f noise. Physica D,
23:413–435, 1986.
405. G. Kember and A. C. Fowler. A correlation-function for choosing time delays in phase portrait
reconstructions. Phys. Lett. A, 179:72–80, 1993.
406. J. A. Kennedy and J. A. Yorke. Basins of Wada. Physica D, 51:213–225, 1991.
407. J. A. Kennedy and J. A. Yorke. The topology of stirred fluids. Topology Appl., 80:201–238,
1997.
408. A. Y. Khinchin. Continued Fractions. Dover, New York, 1997.
409. L. Kim and J. Moehliss. Characterizing the edge of chaos for a shear flow model. Phys. Rev.
E, 78:036315, 2008.
410. I. Z. Kiss, J. H. Merkin, and Z. Neufeld. Combustion initiation and extinction in a 2D chaotic
flow. Physica D, 183:175–189, 2003.
411. R. Klages. Microscopic Chaos, Fractal and Transport in Nonequilibrium Statistical Mechan-
ics. World Scientific, Singapore, first edition, 2007.
412. N. E. Kochin, I. A. Kibel, and N. V. Rose. Theoretical Hydrodynamics. Interscience, New
York, 1964.
413. T.-Y. Koh and B. Legras. Hyperbolic lines and the stratospheric polar vortex. Chaos, 12:382–
394, 2002.
414. M. Kopp and H. Schomerus. Fractal Weyl laws for quantum decay in dynamical systems with
a mixed phase space. Phys. Rev. E, 81:026208, 2010.
415. H. J. Korsch and A. Wagner. Fractal mirror images and chaotic scattering. Comp. Phys.,
5:497–503, 1991.
416. E. J. Kostelich, I. Kan, C. Grebogi, E. Ott, and J. A. Yorke. Unstable dimension variability: a
source of nonhyperbolicity in chaotic systems. Physica D, 109:81–90, 1997.
417. T. Kovács and B. Érdi. Transient chaos in the Sitnikov problem. Cel. Mech. Dyn. Astron.,
105:289–304, 2009.
418. Z. Kovács, K. G. Szabó, and T. Tél. Controlling chaos on fractal basin boundaries. In
J. M. T. Thompson and S. R. Bishop, editors, Nonlinearity and Chaos in Engineering Dy-
namics, pages 155–162. Wiley, Chichester, England, 1994.
419. Z. Kovács and T. Tél. Scaling in multifractals: discretization of an eigenvalue problem. Phys.
Rev. E, 40:4641–4647, 1989.
420. Z. Kovács and L. Wiesenfeld. Chaotic scattering in reactive collisions – a classical analysis.
Phys. Rev. E, 51:5476–5494, 1995.
421. Z. Kovács and L. Wiesenfeld. Topological aspects of chaotic scattering in higher dimensions.
Phys. Rev. E, 63:056207, 2001.
422. Z. Kowalik, M. Franaszek, and P. Pieranski. Self-reanimating chaos in the bouncing-ball
system. Phys. Rev. A, 37:4016–4022, 1988.
423. S. Kraut and U. Feudel. Multistability, noise, and attractor hopping: the crucial role of chaotic
saddles. Phys. Rev. E, 66:015207(R), 2002.
424. S. Kraut and C. Grebogi. Escaping from nonhyperbolic chaotic attractors. Phys. Rev. Lett.,
92:234101, 2004.
425. S. Kraut and U. Feudel. Enhancement of noise-induced escape through the existence of a
chaotic saddle. Phys. Rev. E, 67:015204(R), 2003.
References 475

426. S. Kraut and U. Feudel. Noise-induced escape through a chaotic saddle: lowering of the
activation energy. Physica D, 181:222–234, 2003.
427. U. Kuhl, H.-J. Stöckmann, and R. Weaver. Classical wave experiments on chaotic scattering.
J. Phys. A, 38:10433–10463, 2005.
428. A. Kumar, S. Schrader, A. Aertsen, and S. Rotter. The high-conductance state of cortical
networks. Neural Comput., 20:1–43, 2008.
429. P. K. Kundu. Fluid Mechanics. Academic Press, San Diego, first edition, 1990.
430. L. Kuznetsov and G.M. Zaslavsky. Regular and chaotic advection in the flow field of a three-
vortex problem. Phys. Rev. E, 58:7330–7349, 1998.
431. L. Kuznetsov and G.M. Zaslavsky. Passive particle transport in three-vortex flow. Phys. Rev.
E, 61:3777–3792, 2000.
432. I. Kyprianidis, M. Petrani, J. Kalomiros, and A. Anagnostopoulos. Crisis-induced intermit-
tency in a 3rd-order electrical circuit. Phys. Rev. E, 52:2268–2273, 1995.
433. Y.-C. Lai. Persistence of supertransients of spatiotemporal chaotic dynamics in noisy envi-
ronment. Phys. Lett. A, 200:418–422, 1995.
434. Y.-C. Lai. Driving trajectories to a desirable attractor by using small control. Phys. Lett. A,
221:375–383, 1996.
435. Y.-C. Lai. Characterization of the natural measure by unstable periodic orbits in nonhyper-
bolic chaotic systems. Phys. Rev. E, 56:6531–6539, 1997.
436. Y.-C. Lai. Abrupt bifurcation to chaotic scattering with discontinuous change in fractal di-
mension. Phys. Rev. E, 60:R6283–R6286, 1999.
437. Y.-C. Lai. Transient fractal behavior in snapshot attractors of driven chaotic systems. Phys.
Rev. E, 60:1558–1562, 1999.
438. Y.-C. Lai. Catastrophe of riddling. Phys. Rev. E, 62:R4505–R4508, 2000.
439. Y.-C. Lai. Encoding digital information using transient chaos. Int. J. Bif. Chaos, 10:787–795,
2000.
440. Y.-C. Lai. Superpersistent chaotic transients. In M. Thiel et al., editor, Nonlinear Dynamics
and Chaos: Advances and Perspectives. Springer, Berlin, Heidelberg, 2010.
441. Y.-C. Lai and V. Andrade. Catastropic bifurcation from riddled to fractal basins. Phys. Rev.
E, 64:056228, 2001.
442. Y.-C. Lai, R. Blümel, E. Ott, and C. Grebogi. Quantum manifestations of chaotic scattering.
Phys. Rev. Lett., 68:3491–3494, 1992.
443. Y.-C. Lai, E. M. Bollt, and C. Grebogi. Communicating with chaos using two-dimensional
symbolic dynamics. Phys. Lett. A, 255:75–81, 1999.
444. Y.-C. Lai, A. P. S. de Moura, and C. Grebogi. Topology of high-dimensional chaotic scatter-
ing. Phys. Rev. E, 62:6421–6428, 2000.
445. Y.-C. Lai, M. Ding, C. Grebogi, and R. Blümel. Algebraic decay and fluctuations of the decay
exponent in Hamiltonian systems. Phys. Rev. A, 45:4661–4669, 1992.
446. Y.-C. Lai, U. Feudel, and C. Grebogi. Scaling behaviors in the transition to chaos in quasiperi-
odically driven dynamical systems. Phys. Rev. E, 54:6070–6073, 1996.
447. Y.-C. Lai and C. Grebogi. Converting transient chaos into sustained chaos by feedback con-
trol. Phys. Rev. E, 49:1094–1098, 1994.
448. Y.-C. Lai and C. Grebogi. Intermingled basins and two-state on-off intermittency. Phys. Rev.
E, 52:R3313–R3316, 1995.
449. Y.-C. Lai and C. Grebogi. Modeling of coupled chaotic oscillators. Phys. Rev. Lett., 82:4803–
4806, 1999.
450. Y.-C. Lai, C. Grebogi, R. Blümel, and M. Ding. Algebraic decay and phase-space metamor-
phoses in microwave ionization of hydrogen Rydberg atoms. Phys. Rev. A, 45:8284–8287,
1992.
451. Y.-C. Lai, C. Grebogi, R. Blümel, and I. Kan. Crisis in chaotic scattering. Phys. Rev. Lett.,
71:2212–2215, 1993.
452. Y.-C. Lai, C. Grebogi, and J. Kurths. Modeling of deterministic chaotic systems. Phys. Rev.
E, 59:2907–2910, 1999.
453. Y.-C. Lai, C. Grebogi, and T. Tél. Stabilizing chaotic scattering trajectories using control.
Phys. Rev. E, 48:709–717, 1993.
476 References

454. Y.-C. Lai, C. Grebogi, J. A. Yorke, and I. Kan. How often are chaotic saddles nonhyperbolic?
Nonlinearity, 6:779–797, 1993.
455. Y.-C. Lai, C. Grebogi, J. A. Yorke, and S. Venkataramani. Riddling bifurcation in chaotic
dynamical systems. Phys. Rev. Lett., 77:55–58, 1996.
456. Y.-C. Lai and D. Lerner. Effective scaling regime for computing the correlation dimension in
chaotic time series analysis. Physica D, 115:1–18, 1998.
457. Y.-C. Lai, D. Lerner, and R. Hayden. An upper bound for the proper delay time in chaotic
time series analysis. Phys. Lett. A, 218:30–34, 1996.
458. Y.-C. Lai, Z. Liu, L. Billings, and I. B. Schwartz. Noise-induced unstable dimension variabil-
ity and transition to chaos in random dynamical systems. Phys. Rev. E, 67:026210, 2003.
459. Y.-C. Lai, Y. Nagai, and C. Grebogi. Characterization of the natural measure by unstable
periodic orbits in chaotic attractors. Phys. Rev. Lett., 79:649–652, 1997.
460. Y.-C. Lai, E. Ott, and C. Grebogi. Temporal crossover from classical to quantum behavior: a
Markov-chain approach. Phys. Lett. A, 173:148–152, 1993.
461. Y.-C. Lai, T. Tél, and C. Grebogi. Controlling transient chaos in dynamical systems. In
M. Yamaguti, editor, Towards the Harnessing of Chaos, pages 153–167. Elsevier, Amsterdam,
1994.
462. Y.-C. Lai and R. L. Winslow. Extreme sensitive dependence on parameters and initial condi-
tions in spatiotemporal chaotic dynamical systems. Physica D, 74:353–371, 1994.
463. Y.-C. Lai and R. L. Winslow. Fractal basin boundaries in coupled map lattices. Phys. Rev. E,
50:3470–3473, 1994.
464. Y.-C. Lai and R. L. Winslow. Riddled parameter space in spatiotemporal chaotic dynamical
systems. Phys. Rev. Lett., 72:1640–1643, 1994.
465. Y.-C. Lai and R. L. Winslow. Geometric properties of the chaotic saddle responsible for
supertransients in spatiotemporal chaotic dynamical systems. Phys. Rev. Lett., 74:5208–5211,
1995.
466. Y.-C. Lai, K. Zyczkowski, and C. Grebogi. Universal behavior in the parametric evolution of
chaotic saddles. Phys. Rev. E, 59:5261–5265, 1999.
467. H. Lamb. Hydrodynamics. Cambridge University Press, Cambridge, 1932.
468. R. Landauer. Electrical resistance of disordered one-dimensional lattices. Philos. Mag.,
21:863–867, 1970.
469. D. P. Lathrop and E. J. Kostelich. Characterization of an experimental strange attractor by
periodic orbits. Phys. Rev. A, 40:4028–4031, 1989.
470. Y.-T. Lau, J. M. Finn, and E. Ott. Fractal dimension in nonhyperbolic chaotic scattering.
Phys. Rev. Lett., 66:978–981, 1991.
471. F. Ledrappier and L.-S. Young. Dimension formula for random transformations. Commun.
Math. Phys., 117:529–548, 1988.
472. K. Lee, C. Jung, and L. E. Reichl. Scattering echoes in a waveguide with a ripple cavity.
Phys. Rev. B, 73:195315, 2006.
473. K.-C. Lee. Long-time tails in a chaotic system. Phys. Rev. Lett., 60:1991–1994, 1988.
474. S.-Y. Lee, S. Rim, J.-W. Ryu, T.-Y. Kwon, M. Choi, and C.-M. Kim. Quasiscarred resonances
in a spiral-shaped microcavity. Phys. Rev. Lett, 93:164102, 2004.
475. O. Legrande and D. Sornette. Coarse-grained properties of the chaotic trajectories in the
stadium. Physica D, 44:229–247, 1990.
476. O. Legrande and D. Sornette. Fractal set of recurrent orbits in billiards. Europhys. Lett.,
11:583–588, 1990.
477. F. Lekien, C. Coulliette, A. J. Mariano, E. H. Ryan, L. K. Shay, G. Haller, and J. Marsden.
Pollution release tied to invariant manifolds: a case study for the coast of Florida. Physica D,
210:1–20, 2005.
478. V. Lepiller, A. Prigent, F. Dumochel, and I. Mutabazi. Transition to turbulence in a tall annulus
submitted to a radial temperature gradient. Rhys. Fluids, 19:054101, 2007.
479. R. Leven, B. Pompe, C. Wilke, and B. Koch. Experiments on periodic and chaotic motions
of a parametrically forced pendulum. Physica D, 16:371–384, 1985.
480. R. Leven and M. Selent. Construction and quantitative characterization of a chaotic saddle
from a pendulum experiment. Chaos. Soliton. Fract., 4:2217–2222, 1994.
References 477

481. A. F. J. Levi, R. E. Slusher, S. L. McCall, S. J. Pearton, and W. S. Hobson. Room-temperature

lasing action in In0.51 Ga0.49 P/In0.2 Ga0.8 As microcylinear laser diodes. Appl. Phys. Lett.,
62:2021–2023, 1993.
482. N. Levinson. A second order differential equation with singular coefficients. Ann. Math.,
50:127–153, 1949.
483. T. Li and J. A. Yorke. Period 3 implies chaos. Am. Math. Mon., 82:985–992, 1975.
484. W. Liebert, K. Pawelzik, and H. G. Schuster. Optimal embeddings of chaotic attractors from
topological considerations. Europhys. Lett., 14:521–526, 1991.
485. W. Liebert and H. G. Schuster. Proper choice of the time-delay for the analysis of chaotic
time-series. Phys. Lett. A, 142:107–111, 1989.
486. Z. Liu and Y.-C. Lai. Chaotic scattering in deformed optical microlasing cavities. Phys. Rev.
E, 65:046204, 2002.
487. R. Livi, G. Martinez-Mekler, and S. Ruffo. Periodic orbits and long transients in coupled map
lattices. Physica D, 45:452–460, 1990.
488. E. N. Lorenz. Deterministic nonperiodic flow. J. Atmos. Sci., 20:130–141, 1963.
489. E. N. Lorenz. Dimension of weather and climate attractors. Nature, 353:241–244, 1991.
490. E. N. Lorenz. Compound windows of the Hénon map. Physica D, 237:1689–1704, 2008.
491. H.-W. Lorenz and H. E. Nusse. Chaotic attractors, chaotic saddles, and fractal basin bound-
aries: Goodwin’s nonlinear accelerator model reconsidered. Chaos Sol. Fract., 13:957–965,
2002.
492. J. Losson, M.C. Mackey, and A. Longtin. Solution multistability in first-order non-linear
differential delay equations. Chaos, 3:167–176, 1993.
493. W. Lu, M. Rose, K. Pance, and S. Sridhar. Quantum resonances and decay of a chaotic fractal
repeller observed using microwaves. Phys. Rev. Lett., 82:5233–5236, 1999.
494. W. Lu, S. Sridhar, and M. Zworski. Fractal Weyl laws for chaotic open systems. Phys. Rev.
Lett., 91:154101, 2003.
495. Z.-M. Lu, J. F. Heagy, M. Calliéres, and J. M. Yuan. Quantum dissociation boundary of a
kicked Morse oscillator. Phys. Rev. A, 43:1118–1121, 1991.
496. Z.-M. Lu, M. Valliéres, J. M. Yuan, and J. F. Heagy. Controlling chaotic scattering: impul-
sively driven Morse potential. Phys. Rev. A, 45:5512–5523, 1992.
497. H. Lustfeld and P. Szépfalusy. Correlation functions on the border lines of transient chaos.
Phys. Rev. E, 53:5882–5889, 1996.
498. R. S. MacKay, J. D. Meiss, and I. C. Percival. Stochasticity and transport in Hamiltonian
systems. Phys. Rev. Lett., 52:697–700, 1984.
499. A. M. Mancho, D. Small, and S. Wiggins. A tutorial on dynamical systems concepts ap-
plied to Lagrangian transport in oceanic flows defined as finite time data sets: theoretical and
computational issues. Phys. Rep., 437:55–124, 2006.
500. P. Manneville. On the statistics of turbulent transients in dissipative dynamical-systems. Phys.
Lett. A, 90:327–328, 1982.
501. C. M. Marcus, S. R. Patel, A. G. Huibers, S. M. Cronenwett, M. Switkes, I. H. Chan, R. M.
Clarke, J. A. Folk, and S. F. Godijn. Quantum chaos in open versus closed quantum dots:
signatures of interacting particles. Chaos Solitons & Fractals, 8:1261–1279, 1997.
502. A. P. Martin. Phytoplankton patchiness: the role of lateral stirring and mixing. Progr.
Oceanogr., 57:125–174, 2003.
503. V. P. Maslov and M. V. Fedoriuk. Semiclassical Approximation in Quantum Mechanics. D.
Reidel, Dordrecht, 1981.
504. M. Mathur, G. Haller, T. Peacock, J. E. Ruppert-Felsot, and H. L. Swinney. Uncovering the
Lagrangian skeleton of turbulence. Phys. Rev. Lett., 98:144502, 2007.
505. R. D. Mattuck. A guide to Feynman diagrams in the many-body problem. Courieer Dover,
Dover, 1992.
506. M. R. Maxey and J. J. Riley. Equation of motion for a small rigid sphere in a nonuniform
flow. Phys. Fluids, 26:883–889, 1983.
507. R. M. May. Simple mathematical models with very complicated dynamics. Nature, 261:459–
467, 1976.
478 References

508. R. M. May. Necessity and chance – deterministic chaos in ecology and evolution. Bull. Am.
Math. Soc., 32:291–308, 1995.
509. S. L. McCall, A. F. J. Levi, R. E. Slusher, S. J. Pearton, and R. A. Logan. Whispering-gallery
mode microdisk lasers. Appl. Phys. Lett., 60:289–291, 1991.
510. K. McCann and P. Yodzis. Nonlinear dynamics and population disappearances. Am. Nat.,
144:873–879, 1994.
511. S. W. McDonald, C. Grebogi, E. Ott, and J. A. Yorke. Fractal basin boundaries. Physica D,
17:125–153, 1985.
512. J. McLaughlin. Stochastic behavior in slightly dissipative systems. Phys. Rev. A, 20:2114–
2119, 1979.
513. J. D. Meiss, J. R. Cary, C. Grebogi, J. D. Crawford, A. N. Kaufman, and H. D. I. Abarbanel.
Correlations of periodic, area-preserving maps. Physica D, 6:375–384, 1983.
514. J. D. Meiss and E. Ott. Markov-tree model of intrinsic transport in Hamiltonian systems.
Phys. Rev. Lett., 55:2741–2744, 1985.
515. J. D. Meiss and E. Ott. Markov tree model of transport in area-preserving maps. Physica D,
20:387–402, 1986.
516. M. Meixner, S. Bose, and E. Schöll. Analysis of complex and chaotic patterns near a
codimension-2 Turing-Hopf point in a reaction-diffusion model. Physica D, 109:128–138,
1997.
517. M. Meixner, A. De Wit, S. Bose, and E. Schöll. Generic spatiotemporal dynamics near
codimension-2 Turing-Hopf bifurcations. Phys. Rev. E, 55:6690–6697, 1997.
518. M. Meixner, S. M. Zoldi, S. Bose, and E. Schöll. Karhunen-Loeve local characterization of
spatiotemporal chaos in a reaction-diffusion system. Phys. Rev. E, 61:1382–1385, 2000.
519. A. Mekis, J. U. Nöckel, G. Chen, A. D. Stone, and R. K. Chang. Ray chaos and Q spoiling in
lasing droplets. Phys. Rev. Lett., 75:2682–2685, 1995.
520. F. Mellibovsky, A. Meseguer, T. M. Schneider, and B. Eckhardt. Transition in localized pipe
flow turbulence. Phys. Rev. Lett., 103:054502, 2009.
521. G. Metcalfe and J. M. Ottino. Autocatalytic processes in mixing flows. Phys. Rev. Lett.,
72:2875–2878, 1994.
522. E. E. Michaelides. Particles, Bubbles and Drops: Their Motion, Heat and Mass Transfer.
World Scientific, Singapore, first edition, 2007.
523. A. P. Micolich, R. P. Taylor, A. G. Davies, J. P. Bird, R. Newbury, T. M. Fromhold, A. Ehlert,
H. Linke, L. D. Macks, W. R. Tribe, E. H. Linfield, D. A. Ritchie, J. Cooper, Y. Aoyagi, and
P. B. Wilkinson. Evolution of fractal patterns during a classical-quantum transition. Phys.
Rev. Lett., 87:036802, 2001.
524. P. D. Miller, C. K. R. T. Jones, A. M. Rogerson, and L. J. Pratt. Quantifying transport in
numerically generated velocity fields. Physica D, 110:105–122, 1997.
525. W. H. Miller. The classical S-matrix in molecular collisions. Adv. Chem. Phys., 30:77–136,
1975.
526. R. A. Miranda, E. L. Rempel, A. C. L. Chian, and F. A. Borotto. Intermittent chaos in nonlin-
ear wave-wave interactions in space plasmas. J. Atm. Sol-Terr. Phys., 67:1852–1858, 2005.
527. K. A. Mitchell. The topology of nested homoclinic and heteroclinic tangles. Physica D,
238:737, 2009.
528. F. C. Moon and G.-X. Li. Fractal basin boundaries and homoclinic orbits for periodic motion
in a two-well potential. Phys. Rev. Lett., 55:1439–1442, 1985.
529. P. Moresco and S. P. Dawson. The PIM-simplex method: an extension of the PIM-triple
method to saddles with an arbitrary number of expanding directions. Physica D, 126:38–48,
1999.
530. D. Morgan, E.M. Bollt, and I. B. Schwartz. Constructing constrained invariant sets in multi-
scale continuum systems. Phys. Rev. E, 68:056210, 2003.
531. H. Morita and K. Kaneko. Collective oscillation in a Hamiltonian system. Phys. Rev. Lett.,
96:050602, 2006.
532. F. Mortessagne, O. Legrande, and D. Sornette. Transient chaos in room acoustics. Chaos,
3:529–542, 1993.
References 479

533. J. Moser. Stable and random motions in dynamical systems. Princeton University Press,
Princeton, NJ, 1973.
534. A. E. Motter and Y.-C. Lai. Dissipative chaotic scattering. Phys. Rev. E, 65:015205, 2002.
535. A. E. Motter and P. S. Letelier. Hausdorff dimension of repellors in low sensitive systems.
Phys. Lett. A, 277:18–24, 2000.
536. A. E. Motter and P. S. Letelier. Mixmaster chaos. Phys. Lett. A, 285:127–131, 2001.
537. A. E. Motter, Y.-C. Lai, and C. Grebogi. Reactive dynamics of inertial particles in nonhyper-
bolic chaotic flows. Phys. Rev. E, 68:056307, 2003.
538. J. Nagler. Crash test for the Copenhagen problem. Phys. Rev. E, 69:066218, 2004.
539. J. Nagler, M. Krieger, M. Linke, J. Schönke, and J. Wiersig. Leaking billiards. Phys. Rev. E,
75:046204, 2007.
540. J. Nagel, D. Speer, T. Gaber, A. Sterck, R. Eichhorn, P. Reimann, K. Ilin, M. Siegel, D. Koelle,
and R. Kleiner. Observation of negative absolute resistance in a Josephson junction. Phys.
Rev. Lett., 100:217001, 2008.
541. A. Németh and P. Szépfalusy. Properties of border states of transient chaos. Phys. Rev. E,
52:1544–1549, 1995.
542. Z. Neufeld, P. Haynes, and G. Picard. The effect of forcing on the spatial structure and spectra
of chaotically advected passive scalars. Phys. Fluids, 12:2506–2513, 2000.
543. Z. Neufeld, P. H. Haynes, V. Garcon, and J. Sudre. Ocean fertilization experiments may
initiate a large scale Phytoplankton bloom. Geophys. Res. Lett., 29:10.1029/2001GL013677,
2002.
544. Z. Neufeld and T. Tél. Advection in chaotically time-dependent open flows. Phys. Rev. E,
57:2832–2842, 1998.
545. Z. Neufeld and E. Hernández-Garcia. Chemical and Biological Processes in Fluid Fows: A
Dynamical Systems Approach. Imperial College Press, Cambridge, first edition, 2009.
546. S. E. Newhouse. The abundance of wild hyperbolic sests and non-smooth stable sets for
diffeomorphisms. Publ. Math. IHES, 51:101–151, 1979.
547. S. E. Newhouse. Lectures on dynamical systems. In J. Guckenheimer, J. Moser, and S. E.
Newhouse, editors, Dynamical Systems, C.I.M.E. Lectures, Birkhauser, Boston, 1980.
548. S. E. Newhouse and T. Pignataro. On the estimation of topological-entropy. J. Stat. Phys.,
72:1331–1351, 1993.
549. P. Newton. The N-vortex Problem. Springer, New York, 2001.
550. T. Nishikawa, Z. Toroczkai, and C. Grebogi. Advective coalescence in chaotic flows. Phys.
Rev. Lett., 87:038301, 2001.
551. T. Nishikawa, Z. Toroczkai, C. Grebogi, and T. Tél. Finite-size effects on active chaotic
advection. Phys. Rev. E, 65:026216, 2002.
552. J. U. Nöckel and A. D. Stone. Ray and wave chaos in asymmetric resonant optical cavities.
Nature, 385:45–47, 1997.
553. J. U. Nöckel, A. D. Stone, and R. K. Chang. Q spoiling and directionality in deformed ring
cavities. Opt. Lett., 19:1693–1695, 1994.
554. J. U. Nöckel, A. D. Stone, G. Chen, H. L. Grossman, and R. K. Chang. Directional emission
from asymmetric resonant cavities. Opt. Lett., 21:1609–1611, 1996.
555. D. W. Noid, S. Gray, and S. A. Rice. Fractal behavior in classical collisional energy transfer.
J. Chem. Phys., 84:2649–2652, 1986.
556. S. Nonnenmacher and M. Zworski. Fractal Weyl laws in discrete models of chaotic scattering.
J. Phys. A, 38:10683–10702, 2005.
557. H. Nusse and J. A. Yorke. A procedure for finding numerical trajectories on chaotic saddles.
Physica D, 36:137–156, 1989.
558. H. E. Nusse, E. Ott, and J. A. Yorke. Saddle-node bifurcations on fractal basin boundaries.
Phys. Rev. Lett., 75:2482–2485, 1995.
559. H. E. Nusse and J. A. Yorke. Basins of attraction. Science, 271:1376–1380, 1996.
560. H. E. Nusse and J. A. Yorke. Wada basin boundaries and basin cells. Physica D, 90:242–261,
1996.
561. H. E. Nusse and J. A. Yorke. Fractal basin boundaries generated by basin cells and the
geometry of mixing chaotic flows. Phys. Rev. Lett., 84:626–629, 2000.
480 References

562. A.R. Osborne and A. Provenzale. Finite correlation dimension for stochastic systems with
power-law spectra. Physica D, 35:357–381, 1989.
563. H. M. Osinga. Locus of boundary crisis: expect infinitely many gaps. Phys. Rev. E,
74:035201(R), 2006.
564. E. Ott. Chaos in Dynamical Systems. Cambridge University Press, Cambridge, UK, second
edition, 2002.
565. E. Ott, J. C. Alexander, I. Kan, J. C. Sommerer, and J. A. Yorke. The transition to chaotic
attractors with riddled basins. Physica D, 76:384–410, 1994.
566. E. Ott, C. Grebogi, and J. A. Yorke. Controlling chaos. Phys. Rev. Lett., 64:1196–1199, 1990.
567. E. Ott, J. C. Sommerer, J. C. Alexander, I. Kan, and J. A. Yorke. Scaling behavior of chaotic
systems with riddled basins. Phys. Rev. Lett., 71:4134–4137, 1993.
568. E. Ott and T. Tél (eds.). Focus issue on chaotic scattering. Chaos, 3:417–706, 1993.
569. J. M. Ottino. The Kinematics of Mixing: Stretching, Chaos and Transport. Cambridge Uni-
versity Press, Cambridge, UK, first edition, 1989.
570. J. F. Owen, P. W. Barber, P. B. Dorain, and R. K. Chang. Enhancement of fluorescence induced
by microstructure resonances of dielectric fiber. Phys. Rev. Lett., 47:1075–1078, 1981.
571. V. Paar and H. Buljan. Bursts in the chaotic trajectory lifetimes preceding controlled periodic
motion. Phys. Rev. E, 62:4869–4872, 2000.
572. V. Paar and N. Pavin. Bursts in average lifetime of transients for chaotic logistic map with a
hole. Phys. Rev. E, 55:4112–4115, 1997.
573. V. Paar and N. Pavin. Discontinuous pattern of cumulative separation between orbits of
chaotic transient. Physics Lett. A, 235:139–144, 1997.
574. V. Paar and N. Pavin. Overlapping of two truncated crisis scenarios: generator of peaks in
mean lifetimes of chaotic transients. Phys. Rev. E, 68:036222, 2003.
575. N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw. Geometry from a time series.
Phys. Rev. Lett., 45:712–716, 1980.
576. K. Pance, W. Lu, and S. Sridhar. Quantum fingerprints of classical Ruelle-Pollicott reso-
nances. Phys. Rev. Lett., 85:2737–2740, 2000.
577. F. Papoff, D. Dangoisse, E. Poite-Hanoteau, and P. Glorieux. Chaotic transients in a CO2 -
laser with modulated parameters – critical slowing-down and crisis-induced intermittency.
Opt. Commun., 67:358–362, 1988.
578. B.-S. Park, C. Grebogi, and Y.-C. Lai. Abrupt dimension changes at basin boundary meta-
morphoses. Int. J. Bifur. Chaos Appl. Sci. Eng., 2:533–541, 1992.
579. M. Pattantyús-Ábrahám, T. Tél, T. Krámer, and J. Józsa. Mixing properties of a shallow basin
due to wind-induced chaotic flow. Adv. Water Res., 30:525–534, 2008.
580. T. Peacock and J. Dabiri (eds.). Lagrangian coherent structures in fluid flows. Chaos, Focus
Issue, 23:017501–019902, 2010.
581. L. M. Pecora and T. L. Carroll. Synchronization in chaotic systems. Phys. Rev. Lett., 64:821–
824, 1990.
582. L. M. Pecora and T. L. Carroll. Discontinuous and nondifferentiable functions and dimension
increase induced by filtering chaotic data. Chaos, 6:432–439, 1996.
583. L. M. Pecora and T. L. Carroll. Master stability functions for synchronized coupled systems.
Phys. Rev. Lett., 80:2109–2112, 1998.
584. L. M. Pecora, T. L. Carroll, and J. F. Heagy. Statistics for mathematical properties of maps
between time series embeddings. Phys. Rev. E, 52:3420–3439, 1995.
585. L. M. Pecora, T. L. Carroll, and J. F. Heagy. Statistics for continuity and differentiability:
an application to attractor reconstruction from time series. Fields Inst. Commun., 11:49–62,
1997.
586. L. M. Pecora, L. Moniz, and J. Nichols. A unified approach to attractor reconstruction. Chaos,
17:013110, 2007.
587. J. M. Pedrosa, G. G. Carlo, D. A. Wisniacki, and L. Ermann. Distribution of resonances in
the quantum open baker map. Phys. Rev. E, 79:016215, 2009.
588. X. Pei and F. Moss. Characterization of low-dimensional dynamics in the crayfish caudal
photoreceptor. Nature, 379:618–621, 1996.
References 481

589. J. Peixinho and T. Mullin. Decay of turbulence in pipe flow. Phys. Rev. Lett., 96:094501,
2006.
590. Á. Péntek, T. Tél, and Z. Toroczkai. Chaotic advection in the velocity field of leapfrogging
vortex pairs. J. Phys. A, 28:2191–2216, 1995.
591. Á. Péntek, T. Tél, and Z. Toroczkai. Fractal tracer patterns in open hydrodynamical flows: the
case of leapfrogging vortex pairs. Fractals, 3:33–53, 1995.
592. Á. Péntek, Z. Toroczkai, T. Tél, C. Grebogi, and J. A. Yorke. Fractal boundaries in open
hydrodynamical flows: signatures of chaotic saddles. Phys. Rev. E, 51:4076–4088, 1995.
593. A. Peres. Stability of quantum motion in chaotic and regular systems. Phys. Rev. A, 30:1610–
1615, 1984.
594. J. M. Petit and M. Hénon. Satellite encounters. Icarus, 66:536–555, 1986.
595. G. Pianigiani. Conditionally invariant-measures and exponential decay. J. Math. Anal. Appl.,
82:75–88, 1981.
596. G. Pianigiani and J. A. Yorke. Expanding maps on sets which are almost invariant – decay
and chaos. Trans. Am. Math. Soc., 252:351–366, 1979.
597. R. T. Pierrehumbert and H. Young. Global chaotic mixing on isentropic surfaces. J. Atm. Sci.,
50:2462–2480, 1993.
598. R. T. Pierrehumbert. Tracer microstructure in the large-eddy dominated regime. Chaos Sol.
Fract., 4:1091–1110, 1994.
599. D. Pierson and F. Moss. Detecting periodic unstable points in noisy chaotic and limit-cycle
attractors with applications to biology. Phys. Rev. Lett., 75:2124–2127, 1995.
600. A. Pikovsky and O. Popovich. Persistent patterns in deterministic mixing flows. Europhys.
Lett., 61:625–631, 2003.
601. A. Pikovsky, O. Popovich, and B. Eckhardt. Abnormal mixing of passive scalar in chaotic
flows. Phys. Rev. E, 75:036308, 2007.
602. A. Pikovsky, M. Rosenblum, and J. Kurths. Synchronization – A Universal Concept in Non-
linear Sciences. Cambridge University Press, Cambridge, UK, first edition, 2001.
603. A. Pikovsky. Escape exponent for transient chaos and chaotic scattering in non-hyperbolic
Hamiltonian systems. J. Phys. A, 25:L477–L481, 1992.
604. S. L. Pimm. The Balance of Nature?: Ecological Issues in the Conservation of Species and
Communities. University of Chicago Press, Chicago, first edition, 1991.
605. C. M. Place and D. K. Arrowsmith. Control of transient chaos in tent maps near crisis. I.
Fixed point targeting. Phys. Rev. E, 61:1357–1368, 2000a.
606. C. M. Place and D. K. Arrowsmith. Control of transient chaos in tent maps near crisis. II.
Periodic orbit targeting. Phys. Rev. E, 61:1369–1381, 2000b.
607. A. Pluchino, V. Latora, and A. Rapisarda. Metastable states, anomalous distributions and
correlations in the HMF model. Physica D, 193:315–328, 2004.
608. H. Poincaré. Les Méthodes Nouvelles de la Mécanique. Gauthier-villars, Paris, 1892.
609. A. Politi, R. Livi, G. L. Oppo, and R. Kapral. Unpredictable behaviour in stable systems.
EuroPhys. Lett., 22:571–576, 1993.
610. A. Politi and A. Torcini. Linear and non-linear mechanism of information propagation. Eu-
roPhys. Lett., 28:545–550, 1994.
611. A. Politi and A. Torcini. Stable chaos. In M. Thiel et al., editor, Nonlinear Dynamics and
Chaos: Advances and Perspectives. Springer, Berlin, Heidelberg, 2010
612. Y. Pomeau and P. Manneville. Intermittent transition to turbulence in dissipative dynamical
systems. Commun. Math. Phys., 74:189–197, 1980.
613. L. Poon, J. Campos, E. Ott, and C. Grebogi. Wada basin boundaries in chaotic scattering. Int.
J. Bifur. Chaos Appl. Sci. Eng., 6:151–165, 1996.
614. J. S. E. Portela, I. L. Caldas, R. L. Viana, and M. A. F. Sanjuán. Fractal and Wada exit basin
boundaries in tokamaks. Int. J. Bif. Chaos, 17:4067, 2007.
615. W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical Recipies.
Cambridge University Press, Cambridge, UK, first edition, 1986.
616. I. Procaccia. Complex or just complicated? Nature, 333:498–499, 1988.
617. K. Pyragas. Continuous control of chaos by self-controlling feedback. Phys. Lett. A, 170:421–
428, 1992.
482 References

618. S.-X. Qian, J. B. Snow, H.-M. Tzeng, and R. K. Chang. Lasing droplets – highlighting the
liquid-air interface by laser emission. Science, 231:486–488, 1986.
619. Z.-L. Qu and G. Hu. Spatiotemporal periodic states, periodic windows, and intermittency in
coupled-map lattices. Phys. Rev. E, 49:1099–1108, 1994.
620. J. A. Ramilowski, S. D. Prado, F. Borondo, and D. Farrelly. Fractal Weyl law behavior in an
open Hamiltonian system. Phys. Rev. E, 80:055201(R), 2009.
621. D. A. Rand and H. B. Wilson. Chaotic stochasticity: a ubiquitous source of unpredictablity
in epidemics. Proc. Roy. Soc., B 246:179–184, 1991.
622. C. C. Rankin and W. H. Miller. Classical S-matrix for linear reactive collisions of H+Cl2 . J.
Chem. Phys., 55:3150, 1971.
623. S. Ree and L. E. Reichl. Fractal analysis of chaotic classical scattering in a cut-circle billiard
with two openings. Phys. Rev. E, 65:055205, 2002.
624. L.E. Reichl. A Modern Course in Statistical Physics. Wiley, New York, second edition, 1998.
625. P. Reimann. Noisy maps near crises. Z. Naturforsch., 49a:1248–1250, 1994.
626. P. Reimann. Noisy one-dimensional maps near a crisis I: weak Gaussian white and colored
noise. J. Stat. Phys., 82:1467–1501, 1996.
627. P. Reimann. Noisy one-dimensional maps near a crisis II: general uncorrelated weak noise.
J. Stat. Phys., 85:403–425, 1996.
628. P. Reimann, R. Müller, and P. Talkner. Decay of metastable states with discrete dynamics.
Phys. Rev. E, 49:3670–3682, 1994.
629. P. Reimann and P. Talkner. Invariant densities for noisy maps. Phys. Rev. A, 44:6348–6393,
1991.
630. P. Reimann and P. Talkner. Escape rates for noisy maps. Phys. Rev. E, 51:4105–4113, 1995.
631. E. L. Rempel and A. C. Chian. High-dimensional chaotic saddles in the Kuramoto-
Shivashinsky equation. Phys. Lett. A, 319:104–109, 2003.
632. E. L. Rempel and A. C. Chian. Intermittency induced by attractor-merging crisis in the
Kuramoto-Shivashinsky equation. Phys. Rev. E, 71:016203, 2005.
633. E. L. Rempel and A. C. Chian. Origin of transient and intermittent dynamics in spatiotemporal
chaotic systems. Phys. Rev. Lett., 98:014101, 2007.
634. E. L. Rempel, A. C. Chian, E. E. Macau, and R. R. Rosa. Analysis of chaotic saddles in high-
dimensional dynamical systems: The Kuramoto-Shivashinsky equation. Chaos, 14:545–556,
2004.
635. E. L. Rempel, A. C. Chian, and R. A. Miranda. Chaotic saddles at the onset of intermittent
spatiotemporal chaos. Phys. Rev. E, 76:056217, 2007.
636. E. L. Rempel and A. C. Chian. Alfvén chaotic saddles. Int. J. Bif. Chaos, 14:4009–4017,
2004.
637. E. L. Rempel, A. C. Chian, E. E. N. Macau, and R. R. Rosa. Analysis of chaotic saddles on
low-dimensional dynamical systems: the derivative nonlinear Schrödinger equation. Physica
D, 199:407–424, 2004.
638. E. L. Rempel, R. A. Miranda, and A. C. Chian. Spatiotemporal intermittency and chaotic
saddles in the regularized long-wave equation. Phys. Fluids, 21:074105, 2009.
639. E. L. Rempel, G. Lesur, and M. R. E. Proctor. Supertransient magnetohydrodynamic turbu-
lence in Keplerian shear flows. Phys. Rev. Lett., 105:044501, 2010.
640. H. Risken. The Fokker-Planck Equation. Springer, Berlin, first edition, 1989.
641. C. Robert, K. T. Alligood, E. Ott, and J. A. Yorke. Outer tangency bifurcations of chaotic
sets. Phys. Rev. Lett., 80:4867–4870, 1998.
642. C. Robert, K. T. Alligood, E. Ott, and J. A. Yorke. Explosions of chaotic sets. Phys. D,
144:44–61, 2000.
643. D. L. Rod. Pathology of invariant sets in the monkey saddle. J. Diff. Equat., 14:129–170,
1973.
644. C. S. Rodrigues, A.P.S. de Moura, and C. Grebogi. Random fluctuation leads to forbidden
escape of particles. Phys. Rev. E, 82:026211, 2010a.
645. C. S. Rodrigues, C. Grebogi, and A.P.S. de Moura. Escape from attracting sets in randomly
perturbed systems. Phys. Rev. E, 82:046217, 2010b.
References 483

646. R. Rollins and E. Hunt. Intermittent transient chaos at interior crises in the diode resonator.
Phys. Rev. A, 29:3327–3334, 1984.
647. V. Rom-Kedar, A. Leonard, and S. Wiggins. An analytical study of transport, mixing and
chaos in an unsteady vortical flow. J. Fluid Mech., 214:347–394, 1990.
648. F. Romeiras, C. Grebogi, and E. Ott. Multifractal properties of snapshot attractors of random
maps. Phys. Rev. A, 41:784–799, 1990.
649. P. Romeiras, C. Grebogi, and E. Ott. Critical exponents for power spectra scaling at merging
of chaotic bands. Phys. Rev. A, 38:463–468, 1988.
650. E. Rosa, S. Hayes, and C. Grebogi. Noise filtering in communicating with chaos. Phys. Rev.
Lett., 78:1247–1250, 1997.
651. E. Rosa and E. Ott. Mixed basin boundary structures of chaotic systems. Phys. Rev. E,
59:343–352, 1999.
652. E. Rosa, E. Ott, and M. H. Hess. Transition to phase synchronization of chaos. Phys. Rev.
Lett., 80:1642–1645, 1998.
653. M. G. Rosenblum, A. S. Pikovsky, and J. Kurths. Phase synchronization of chaotic oscillators.
Phys. Rev. Lett., 76:1804–1807, 1996.
654. M. T. Rosenstein, J. J. Collins, and C. J. De Luca. Reconstruction expansion as a geometry-
based framework for choosing proper delay times. Physica D, 73:82–98, 1994.
655. V. Rossi, C. López, E. Hernández-Garcia, J. Sudre, V. Garcon, and Y. Morel. Surface mix-
ing and biological activity in the four Eastern boundary upwelling systems. Nonlin. Proc.
Geophys., 16:557–568, 2009.
656. M. L. Roukes and O. L. Alerhand. Mesoscopic junctions, random scattering, and strange
repellers. Phys. Rev. Lett., 65:1651–1654, 1990.
657. B. Rückerl and C. Jung. Scaling properties of a scattering system with an incomplete horse-
shoe. J. Phys. A-Math. Gen., 27:55–77, 1994.
658. P. Ruján. Calculation of free-energy of Ising systems by a recursion method. Physica A,
91:549–562, 1978.
659. P. Rupp, R. Richter, and I. Rehberg. Critical exponents of directed percolation in spatiotem-
poral intermittency. Phys. Rev. E, 67:036209, 2003.
660. J.-W. Ryu, S.-Y. Lee, C.-M. Kim, and Y.-J. Park. Survival probability time distribution in
dielectric cavities. Phys. Rev. E, 73:036207, 2006.
661. J. Sabuco, S. Zambrano, and M. A. F. Sanjuán. Partial control of chaotic transients using
escape times. New J. Phys., 12:113038, 2010.
662. P. G. Saffman. Vortex Dynamics. Cambridge University Press, Cambridge, UK, 1992.
663. D. Samanta, A. de Lozar, and B. Hof. Experimental investigation of laminar turbulent inter-
mittency in pipe flow. arXiv:1008.2294v1, 2010.
664. M. Sandulescu, E. Hernández-Garcia, C. E. López, and U. Feudel. Kinematic studies of
transport accross an island wake, with aplication to the Canary islands. Tellus A, 58:605–605,
2006.
665. M. Sandulescu, C. E. López, E. Hernández-Garcia, and U. Feudel. Plankton blooms in
vortices: the role of biological and hydrodynamics timescales. Nonlin. Process. Geophys.,
14:1–12, 2007.
666. M. Sandulescu, C. E. López, E. Hernández-Garcia, and U. Feudel. Biological activity in the
wake of an island close to costal upwelling. Ecol. Complex., 5:228–237, 2008.
667. M. A. F. Sanjuán and C. Grebogi. Recent Progress in Controlling Chaos. World Scientific,
Singapore, first edition, 2010.
668. M. A. F. Sanjuán, T. Horita, and K. Aihara. Opening a closed Hamiltonian map. Chaos,
13:17–24, 2003a.
669. M. A. F. Sanjuán, J. A. Kennedy, E. Ott, and J. A. Yorke. Indecomposable continua and the
characterization of strange sets in nonlinear dynamics. Phys. Rev. Lett., 78:1892, 1997.
670. M. A. F. Sanjuán, J. A. Kennedy, C. Grebogi, and J. A. Yorke. Indecomposable continua in
dynamical systems with noise: fluid flow past an array of cylinders. Chaos, 7:125–138, 1997.
671. M. Sano and Y. Sawada. Measurement of the Lyapunov spectrum from a chaotic time series.
Phys. Rev. Lett., 55:1082–1085, 1985.
484 References

672. T. Sapsis and G. Haller. Instabilities in the dynamics of neutrally bouyant particles. Phys.
Fluids, 20:017102, 2008.
673. T. Sapsis and G. Haller. Inertial particle dynamics in a hurricane. J. Atmos. Sci., 66:2481–
2492, 2009.
674. T. Sapsis and G. Haller. Clustering criterion for inertial particles in two-dimensional time-
periodic and three-dimensional steady flows. Chaos, 20:017515, 2010.
675. J. Sartorelli, W. Goncalves, and R. Pinto. Crisis and intermittency in a Leaky-Faucet experi-
ment. Phys. Rev. E, 49:3963–3975, 1994.
676. T. D. Sauer, C. Grebogi, and J. A. Yorke. How long do numerical chaotic solutions remain
valid? Phys. Rev. Lett., 79:59–62, 1997.
677. T. D. Sauer, J. A. Tempkin, and J. A. Yorke. Spurious Lyapunov exponents in attractor recon-
struction. Phys. Rev. Lett., 81:4341–4344, 1998.
678. T. D. Sauer and J. A. Yorke. Reconstructing the Jacobian from data with observational noise.
Phys. Rev. Lett., 83:1331–1334, 1999.
679. T. D. Sauer, J. A. Yorke, and M. Casdagli. Embedology. J. Stat. Phys., 65:579–616, 1991.
680. A. B. Schelin, A. P. S. de Moura, and C. Grebogi. Transition to chaotic scattering: signatures
in the differential cross section. Phys. Rev. E, 78:046204, 2008.
681. A. B. Schelin, Gy. Károlyi, A. P. S. de Moura, N. A. Booth, and C. Grebogi. Chaotic advection
in blood flow. Phys. Rev. E, 80:016213, 2009.
682. I. Scheuring, T. Czárán, P. Szabó, G. Károlyi, and Z. Toroczkai. Spatial models of prebiotic
evolution: soup before pizza? Orig. Life Evol. Biosph., 33:319–355, 2003.
683. I. Scheuring and G. Domokos. Only noise can induce chaos in discrete populations. Oikos,
116:361–366, 2007.
684. S. J. Schiff, K. Jerger, D. H. Duong, T. Chang, M. L. Spano, and W. L. Ditto. Controlling
chaos in the brain. Nature, 370:615–620, 1994.
685. P. Schmelcher and F. K. Diakonos. Detecting unstable periodic orbits of chaotic dynamical
systems. Phys. Rev. Lett., 78:4733–4736, 1997.
686. A. Schmiegel and B. Eckhardt. Fractal stability border in plane Couette flow. Phys. Rev. Lett.,
79:5250–5253, 1997.
687. J. Schneider, V. Fernández, and E. Hernández-Garcia. Leaking method approach to surface
transport in the Mediterranean sea from a numerical ocean model. J. Mar. Syst., 57:111–126,
2000.
688. J. Schneider, T. Tél, and Z. Neufeld. Dynamics of leaking Hamiltonian systems. Phys. Rev.
E, 66:066218, 2002.
689. J. Schneider, J. Schmalzl, and T. Tél. Lagrangian avenues of transport in the Earth’s mantle.
Chaos, 17:033115, 2007.
690. T. M. Schneider, B. Eckhardt, and J. Vollmer. Statistical analysis of coherent structures in
transitional pipe flows. Phys. Rev. E, 75:066313, 2007.
691. T. M. Schneider, B. Eckhardt, and J. A. Yorke. Turbulence transition and the edge of chaos in
pipe flow. Phys. Rev. Lett., 99:034502, 2007.
692. T. M. Schneider, J. F. Gibson, M. Lagha, F. De Lillo, and B. Eckhardt. Laminar-turbulent
boundary in plane Couette flow. Phys. Rev. E, 78:037301, 2007.
693. W. Schoepe. Fluctuations and stability of superfluid turbulence at mK temperatures. Phys.
Rev. Lett., 92:095301, 2004.
694. E. Schöll and H. G. Schuster (eds.). Handbook of Chaos Control. Wiley-VCH, Weinheim,
second edition, 2007.
695. H. Schomerus, J. Wiersig, and J. Main. Lifetime statistics in chaotic dielectric microres-
onators. Phys. Rev. A, 79:053806, 2009.
696. I. B. Schwartz, L. Billings, and E.M. Bollt. Dynamical epidemic suppression using stochastic
prediction and control. Phys. Rev. E, 70:046220, 2004.
697. I. B. Schwartz and I. Triandaf. Sustaining chaos by using basin boundary saddles. Phys. Rev.
Lett., 77:4740–4743, 1996.
698. I. B. Schwartz, I. Triandaf, R. Meucci, and T. W. Carr. Open-loop sustained chaos and control:
A manifold approach. Phys. Rev. E, 66:026213, 2002.
References 485

699. H. G. L. Schwefel, N. B. Rex, H. E. Tureci, R. K. Chang, A. D. Stone, T. Ben-Messaoud, and

J. Zyss. Dramatic shape sensitivity of directional emission patterns from similarly deformed
cylindrical polymer lasers. J. Opt. Soc. Am., B 21:923–934, 2004.
700. S. Scott, B. Peng, A. Tomlin, and K. Showalter. Transient Chaos in a Closed Chemical-
System. J. Chem. Phys., 94:1134–1140, 1991.
701. P. Seba. Quantum irregular scattering in the presence of a classical stability island. Phys. Rev.
E, 47:3870–3878, 1993.
702. J. M. Seoane, J. Aguirre, M. A. F. Sanjuán, and Y.-C. Lai. Basin topology in dissipative
chaotic scattering. Chaos, 16:023101, 2006.
703. J. M. Seoane, L. Huang, M. A. F. Sanjuán, and Y.-C. Lai. Effect of noise on chaotic scattering.
Phys. Rev. E, 79:047202, 2009.
704. J. M. Seoane, M. A. F. Sanjuán, and Y.-C. Lai. Fractal dimension in dissipative chaotic
scattering. Phys. Rev. E, 76:016208, 2007.
705. J. M. Seoane and M. A. F. Sanjuán. Exponential decay and scaling laws in noisy chaotic
scattering. Phys. Lett. A, 372:110–116, 2008.
706. J. M. Seoane, S. Zambrano, I. P. Marino, and M. A. F. Sanjuán. Basin boundary metamor-
phoses and phase control. Europhys. Lett., 90:30002, 2010.
707. S. C. Shadden, F. Lekien, and J. E. Marsden. Definition and properties of Lagrangian coherent
structures from finite-time Lyapunov exponents in two-dimensional aperiodic flows. Physica
D, 212:271–304, 2005.
708. C. E. Shannon and W. Weaver. The Mathematical Theory of Communication. The University
of Illinois Press, IL, first edition, 1964.
709. K. Shariff and A. Leonard. Vortex rings. Ann. Rev. Fluids Mech., 24:235–279, 1992.
710. K. Shariff, A. Leonard, and J. H. Ferziger. Dynamical systems analysis of fluid transport in
time-periodic vortex ring flows. Phys. Fluids, 18:047106, 2006.
711. J.-L. Shen, H.-W Yin, J.-H. Dai, and H.-J. Zhang. Dynamical behavior, transient chaos, and
riddled basins of two charged particles in a Paul trap. Phys. Rev. A, 55:2159–2164, 2009.
712. D. L. Shepelyansky. Fractal Weyl law for quantum fractal eigenstates. Phys. Rev. E,
77:015202(R), 2008.
713. H. Shibata, H. Fujisaka, and H. Mori. Scaling structures for dynamics of intermittent switch-
ing just after the 3-band crisis. Physica A, 189:554–574, 1992.
714. T. Shimizu and N. Morioka. Transient behavior in periodic regions of the Lorenz model.
Phys. Lett. A, 69:148–150, 1978.
715. B. Shraiman. Order, disorder, and phase turbulence. Phys. Rev. Lett., 57:325–328, 1986.
716. L. Shulenburger, Y.-C. Lai, T. Yalcinkaya, and R. D. Holt. Controlling transient chaos to
prevent species extinction. Phys. Lett. A, 260:156–161, 1999.
717. A. N. Silchenko, S. Beri, D. G. Luchinsky, and P. V. E. McClintock. Fluctuational transitions
through a fractal basin boundary. Phys. Rev. Lett., 91:174104, 2003.
718. A. N. Silchenko, S. Beri, D. G. Luchinsky, and P. V. E. McClintock. Fluctuational transitions
across different kinds of fractal basin boundaries. Phys. Rev. E, 71:046203, 2005.
719. C. Siopis, H. E. Kandrup, G. Contopoulos, and R. Dvorak. Universal properties of escape in
dynamical systems. Cel. Mech. Dyn. Astr., 65:57–68, 1997.
720. J. D. Skufca, J. A. Yorke, and B. Eckhardt. Edge of chaos in a parallel shear flow. Phys. Rev.
Lett., 96:174101, 2006.
721. S. Smale. Differentiable dynamics systems. Bull. Am. Math. Soc., 73:747–817, 1967.
722. P. So, E. Ott, S. J. Schiff, D. T. Kaplan, T. D. Sauer, and C. Grebogi. Detecting unstable
periodic orbits in chaotic experimental data. Phys. Rev. Lett., 76:4705–4708, 1996.
723. J. C. Sommerer. A universal scaling for noise-induced metastability. Phys. Lett. A, 176:85–88,
1993.
724. J. C. Sommerer, W. L. Ditto, C. Grebogi, E. Ott, and M. L. Spano. Experimental confirmation
of the scaling theory for noise-induced crises. Phys. Rev. Lett., 66:1947–1950, 1991.
725. J. C. Sommerer, H. Ku, and H. Gilreath. Experimental evidence for chaotic scattering in a
fluid wake. Phys. Rev. Lett., 77:5055–5058, 1996.
726. J. C. Sommerer and E. Ott. Particles floating on a moving fluid: a dynamically comprehensive
physical fractal. Science, 259:335–339, 1993.
486 References

727. J. C. Sommerer and E. Ott. Blowout bifurcations – the occurrence of riddled basins and on-off
intermittency. Phys. Lett. A, 188:39–47, 1994.
728. J. C. Sommerer, E. Ott, and C. Grebogi. Scaling law for characteristic times of noise-induced
crises. Phys. Rev. A, 43:1754–1769, 1991.
729. S. M. Soskin, R. Mannella, M. Arrayás, and A. N. Silchenko. Strong enhancement of noise-
induced escape by nonadiabatic periodic driving due transient chaos. Phys. Rev. E, 63:051111,
2001.
730. D. Speer, R. Eichhorn, and P. Reimann. Brownian motion: anomalous response due to noisy
chaos. Eur. Phys. Lett, 79:10005, 2007.
731. D. Speer, R. Eichhorn, and P. Reimann. Transient chaos induced anomalous transport prop-
erties of an underdamped Brownian particle. Phys. Rev. E, 76:051110, 2007.
732. M. A. Spina, I. Garcia-Mata, and M. Saraceno. Weyl law for fat fractals. J. Phys. A,
43:392003, 2010.
733. T. M. Squires and S. R. Quake. Microfluidics: fluid physics at the nanoliter scale. Rev. Mod.
Phys., 77:977–1026, 2005.
734. D. Stahlke and R. Wackerbauer. Transient spatiotemporal chaos is extensive in the reaction-
diffusion networks. Phys. Rev. E, 80:056211, 2009.
735. D. A. Steck, W. H. Oskay, and M. G. Raizen. Observation of chaos-assisted tunneling between
islands of stability. Science, 293:274–278, 2001.
736. A. F. Steklain and P. S. Letelier. Newtonian and pseudo-Newtonian Hill problem. Phys. Lett.
A, 352:398–4003, 2006.
737. G. Stépán. Retarded Dynamical Systems: Stability and Characteristic Functions. Wiley, New
York, first edition, 1989.
738. G. Stépán. Chaotic motion of wheels. Vehicle Syst. Dyn., 20:341–351, 1991.
739. G. Stépán. Delay, nonlinear oscillations and shimmying wheels. In F.C. Moon, editor, IUTAM
Symposium on New Applications of Nonlinear and Chaotic Dynamics in Mechanics, pages
373–387. Kluwer, Amsterdam, 1999.
740. H.-J. Stöckmann. Quantum Chaos – An Introduction. Cambridge University Press, Cam-
bridge, first edition, 1999.
741. M. C. Strain and H. S. Greenside. Size-dependent transition to high-dimensional chaotic
dynamics in a two-dimensional excitable medium. Phys. Rev. Lett., 80:2306–2309, 1998.
742. A.V. Straube and A. Pikovsky. Mixing-induced global models in open active flows. Phys.
Rev. Lett., 99:184503, 2007.
743. A. Surana, O. Grunberg, and G. Haller. Exact theory of three-dimensional flow separation.
Part I. Steady separation. J. Fluid Mech., 564:57–103, 2006.
744. D. Sweet, H. E. Nusse, and J. A. Yorke. Stagger-and-step method: detecting and computing
chaotic saddles in higher dimensions. Phys. Rev. Lett., 86:2261–2264, 2001.
745. D. Sweet and E. Ott. Fractal basin boundaries in higher-dimensional chaotic scattering. Phys.
Lett. A, 266:134–139, 2000.
746. D. Sweet and E. Ott. Fractal dimension of higher-dimensional chaotic repellors. Physica D,
139:1–27, 2000.
747. D. Sweet, E. Ott, and J. A. Yorke. Complex topology in chaotic scattering: a laboratory
experiment. Nature, 399:315–317, 1999.
748. D. Sweet, B. W. Zeff, E. Ott, and D. P. Lathrop. Three-dimensional optical billiard chaotic
scattering. Physica D, 154:207–218, 2001.
749. K. G. Szabó, Y.-C. Lai, T. Tél, and C. Grebogi. Critical exponent for gap filling at crisis.
Phys. Rev. Lett., 77:3102–3105, 1996.
750. K. G. Szabó, Y.-C. Lai, T. Tél, and C. Grebogi. Topological scaling and gap filling at crises.
Phys. Rev. E, 61:5019–5032, 2000.
751. K. G. Szabó and T. Tél. Thermodynamics of attractor enlargement. Phys. Rev. E, 50:1070–
1082, 1994.
752. K. G. Szabó and T. Tél. Transient chaos as the backbone of dynamics on strange attractors
beyond crisis. Phys. Lett. A, 196:173–180, 1994.
753. P. Szépfalusy and T. Tél. New approach to the problem of chaotic repellers. Phys. Rev. A,
34:2520–2523, 1986.
References 487

754. D. Takács, G. Orosz, and G. Stépán. Delay effects in shimmy dynamics of wheels with string-
like tyres. Eur. J. Mech. A Solids, 28:516–525, 2009.
755. D. Takács, G. Stépán, and S.J. Hogan. Isolated large amplitude periodic motions of towed
rigid wheels. Nonlin. Dyn., 52:27–34, 2008.
756. F. Takens. Detecting strange attractors in fluid turbulence. In D. Rand and L. S. Young,
editors, Dynamical Systems and Turbulence, Vol. 898 of Lecture Notes in Mathematics, pages
366–381. Springer, Berlin, 1981.
757. S. Takesue and K. Kaneko. Fractal basin structure. Prog. Theor. Phys., 71:35–49, 1984.
758. P. Tallapragada and S. D. Ross. Particle segregation by Stokes number for small neutrally
buoyant spheres in a fluid. Phys. Rev. E, 78:036308, 2008.
759. W. Tang, G. Haller, J.-J. Baik, and Y.H. Ryu. Locating and atmospheric contamination source
using slow manifolds. Phys. Fluids, 21:043302, 2009.
760. W. Tang, M. Mathur, G. Haller, D. C. Hahn, and F. H. Ruggiero. Lagrangian coherent struc-
tures near a subtropical jet stream. J. Atm. Sci., 2010.
761. S. R. Taylor and S.A. Campbell. Approximating chaotic saddles in delay differential equa-
tions. Phys. Rev. E, 75:046215, 2007.
762. T. Tél. Invariant curves, attractors, and phase-diagram of a piecewise linear map with chaos.
J. Stat. Phys., 33:195–221, 1983.
763. T. Tél. Dynamic spectrum and thermodynamic functions of strange sets from an eigenvalue
problem. Phys. Rev. A, 36:2507–2510, 1987.
764. T. Tél. Escape rate from strange sets as an Eigenvalue. Phys. Rev. A, 36:1502–1505, 1987.
765. T. Tél. On the organization of transient chaos: application to irregular scattering. J. Phys.
A-Math. Gen., 22:L691–L697, 1989.
766. T. Tél. Transient chaos. In B.-L. Hao, editor, Directions in Chaos (3): Experimental Study
and Characterization of Chaos, pages 149–211. World Scientific, Singapore, 1990.
767. T. Tél. Controlling transient chaos. J. Phys. A-Math. Gen, 24:L1359–L1368, 1991.
768. T. Tél. Thermodynamics of chaotic scattering at abrupt bifurcation. Phys. Rev. A, 44:1034–
1043, 1991.
769. T. Tél. Crossover between the control of permanent and transient chaos. Int. J. Bif. Chaos,
3:757–764, 1993.
770. T. Tél. Transient chaos: a type of metastable state. In B.-L. Hao, editor, STATPHYS 19:
The Proceedings of the 19th IUPAP Conference on Statistical Physics, pages 346–362. World
Scientific, Singapore, 1996.
771. T. Tél, A. P. S. de Moura, C. Grebogi, and G. Károlyi. Chemical and biological activity in
open flows: a dynamical system approach. Phys. Rep., 413:91–196, 2005.
772. T. Tél, C. Grebogi, and E. Ott. Conditions for the abrupt bifurcation to chaotic scattering.
Chaos, 3:495–503, 1993.
773. T. Tél and M. Gruiz. Chaotic Dynamics: An Introduction Based on Classical Mechanics.
Cambridge University Press, Cambridge, 2006.
774. T. Tél, G. Károlyi, A. Péntek, I. Scheuring Z. Toroczkai, C. Grebogi, and J. Kadtke. Chaotic
advection, diffusion, and reactions in open flows. Chaos, 10:89–98, 2000.
775. T. Tél and Y.-C. Lai. Chaotic transients in spatially extended systems. Phys. Rep., 460:245–
275, 2008.
776. T. Tél and Y.-C. Lai. Quasipotential approach to critical scaling in noise-induced chaos. Phys.
Rev. E, 81:056208, 2010.
777. T. Tél, Y.-C. Lai, and M. Gruiz. Noise-induced chaos: a consequence of long deterministic
transients. Int. J. Chaos Bif., 18:509–520, 2008.
778. T. Tél, J. Vollmer, and W. Breymann. Transient chaos: the origin of transport in driven sys-
tems. Europhys. Lett., 35:659–664, 1996.
779. J. Theiler. Spurious dimension from correlation algorithms applied to limited time series data.
Phys. Rev. A, 34:2427–2432, 1986.
780. J. M. T. Thompson. Chaotic phenomena triggering the escape from a potential well. Proc. R.
Soc. Lond. A, 421:195–225, 1989.
781. J. M. T. Thompson. Global unpredictability in nonlinear dynamics – capture, dispersal and
the indeterminate bifurcations. Physica D, 58:260–272, 1992.
488 References

782. J. M. T. Thompson and H.B. Stewart. Nonlinear Dynamics and Chaos. Wiley, New York,
second edition, 2002.
783. S. Tomsovic. Tunneling and chaos. Phys. Scripta, T90:162–165, 2001.
784. A. Torcini, P. Grassberger, and A. Politi. Error propagation in extended systems. J. Phys. A,
27:4533–4541, 1995.
785. Z. Toroczkai. Geometric method for stabilizing unstable periodic-orbits. Phys. Lett. A,
190:71–78, 1994.
786. Z. Toroczkai, G. Károlyi, Á. Péntek, T. Tél, and C. Grebogi. Advection of active particles in
open chaotic flows. Phys. Rev. Lett., 80:500–503, 1998.
787. Z. Toroczkai, G. Károlyi, Á. Péntek, T. Tél, C. Grebogi, and J. A. Yorke. Wada dye boundaries
in open hydrodynamical flows. Physica A, 239:235–243, 1997.
788. I. Triandaf, E.M. Bollt, and I. B. Schwartz. Approximating stable and unstable manifolds in
experiments. Phys. Rev. E, 67:037201, 2003.
789. I. Triandaf and I. B. Schwartz. Quality factor control in a lasing microcavity model. Phys.
Rev. E, 61:3601–3609, 2000.
790. G. Troll and U. Smilansky. A simple model for chaotic scattering. Physica D, 35:34–64,
1989.
791. P. Turchin and S. P. Ellner. Living on the edge of chaos: population dynamics of fennoscan-
dian voles. Ecology, 81:3099–3116, 2000.
792. L. F. L. Turci, E. E. N. Macau, and T. Yoneyama. Chaotic transient and the improvement of
system flexibility. Phys. Lett. A, 365:328–334, 2007.
793. I. Tuval, J. Schneider, O. Piro, and T. Tél. Opening up fractal structures of three-dimensional
flows via leaking. EuroPhys. Lett., 65:633–639, 2004.
794. E. Tyrkiel. On the role of chaotic saddles in generating chaotic dynamics in nonlinear driven
oscillators. Int. J. Bif. Chaos, 15:1215–1238, 2005.
795. D. Umberger, C. Grebogi, E. Ott, and B. Afeyan. Spatiotemporal dynamics in a dispersively
coupled chain of nonlinear oscillators. Phys. Rev. A, 39:4835–4842, 1989.
796. T. Usuki, M. Saito, M. Takatsu, R. A. Kiehl, and N. Yokoyama. Numerical analysis of
ballistic-electron transport in magnetic fields by using a quantum point contact and a quantum
wire. Phys. Rev. B, 52:8244–8255, 1995.
797. D. B. Vasconcelos, R. L. Viana, S. R. Lopes, and S. E. de S. Pinto. Conversion of local
transient chaos into global laminar states in coupled map lattices with long-range interactions.
Physica A, 367:158–172, 2006.
798. R. D. Vilela, A. P. S. de Moura, and C. Grebogi. Finite-size effects on open chaotic advection.
Phys. Rev. E, 73:026302, 2006.
799. R. D. Vilela and A. E. Motter. Can aerosols be trapped in open flows? Phys. Rev. Lett.,
99:264101, 2007.
800. D. Viswanath and P. Cvitanović. Stable manifolds and the transition to turbulence in pipe
flow. J. Fluid. Mech, 627:215, 2009.
801. F. Vivaldi, A. Casati, and I. Guarneri. Origin of long-time tails in strongly chaotic systems.
Phys. Rev. Lett., 51:727–730, 1983.
802. J. Vollmer. Chaos, spatial extension, transport, and non-equilibrium thermodynamics. Phys.
Rep., 372:131–267, 2002.
803. J. Vollmer, T. M. Schneider, and B. Eckhardt. Basin boundary, edge of chaos, and edge state
in a two-dimensional model. New J. Phys., 11:013040, 2009.
804. J. Vollmer, T. Tél, and W. Breymann. Dyanmical-system models of transport: chaos charac-
teristics, the macroscopic limit, and irreversibility. Physica D, 187:108–127, 2004.
805. G. A. Voth, G. Haller, and J. P. Gollub. Experimental measurements of stretching fields in
fluid mixing. Phys. Rev. Lett., 88:254501, 2002.
806. H. Waalkens, A. Burbancks, and S. Wiggins. A computational procedure to detect a new type
of high-dimensional chaotic saddle and its application to the 3D Hill’s problem. J. Phys. A,
37:L257–L265, 2004.
807. H. Waalkens, A. Burbanks, and S. Wiggins. Escape from planetary neighbourhoods. Mon.
Not. R. Astron. Soc, 361:763–775, 2004.
References 489

808. H. Waalkens, A. Burbanks, and S. Wiggins. Phase space conduits for reaction in multidimen-
sional systems: HCN isomerization in three dimensions. J. Chem. Phys., 121:6207–6225,
2004.
809. H. Waalkens, A. Burbanks, and S. Wiggins. Efficient procedure to compute the microcanon-
ical volume of initial conditions that lead to escape trajectories from a multidimensional
potential well. Phys. Rev. Lett., 95:084301, 2005.
810. H. Waalkens, R. Schubert, and S. Wiggins. Wigner’s dynamical transition state theory in
phase space: classical and quantum. Nonlinearity, 21:R1–R118, 2008.
811. H. Waalkens and S. Wiggins. Direct construction of a dividing surface of minimal flux for
multi-degree-of-freedom systems that cannot be crossed. J. Phys. A-Math. Gen., 37:L435–
L445, 2004.
812. A. Wacker, S. Bose, and E. Schöll. Transient spatio-temporal chaos in a reation-diffusion
model. Europhys. Lett., 31:257–262, 1995.
813. R. Wackerbauer. Master stability analysis in transient spatiotemporal chaos. Phys. Rev. E,
76:056207, 2007.
814. R. Wackerbauer and S. Kobayashi. Noise can delay and advance the collapse of spatiotempo-
ral chaos. Phys. Rev. E, 75:066209, 2007.
815. R. Wackerbauer and K. Showalter. Collapse of spatiotemporal chaos. Phys. Rev. Lett.,
91:174103, 2003.
816. F. Waleffe. Three-dimensional coherent states in plane shear flows. Phys. Rev. Lett., 81:4140–
4143, 1998.
817. J. Walker. The distorted images seen in Christmas-tree ornaments and other reflecting balls.
Sci. Am., 259:84–87, The Amateur Scientist, December, 1988.
818. H. O. Wang and E. H. Abed. Control of nonlinear phenomena at the inception of voltage
collapse. In Proceedings of the 1993 American Control Conference, pages 2071–2075, San
Francisco, 1993.
819. H. O. Wang, E. H. Abed, and A. M. A. Hamdan. Bifurcations, chaos and crises in voltage
collapse of a model power system. IEEE Trans. Circ. Syst. I: Fund. Theor. Appl., 41:294–302,
1994.
820. J. Wang, P. G. Sorensen, and F. Hynne. Transient period doublings, torus oscillations, and
chaos in a closed chemical-system. J. Phys. Chem.-US, 98:725–727, 1994.
821. J. Wang, P. G. Sorensen, and F. Hynne. Transient complex oscillations in the closed Belousov-
Zhabotinsky reaction: Experimental and computational studies. Z. Phys. Chemie, 192:63–76,
1995.
822. Y. Wang, J. Singer, and H. Bau. Controlling chaos in a thermal-convection loop. J. Fluid
Mech., 237:479–498, 1992.
823. P. Widmann, M. Gorman, and K. Robbins. Nonlinear dynamics of a convection loop. 2. Chaos
in laminar and turbulent flows. Physica D, 36:157–166, 1989.
824. M. Widom, D. Bensimon, L. Kadanoff, and S. Shenker. Strange objects in the complex-plane.
J. Stat. Phys., 32:443–454, 1983.
825. J. Wiersig and J. Main. Fractal Weyl law for chaotic microcavities: Fresnel’s laws imply
multifractal scattering. Phys. Rev. E, 77:036205, 2008.
826. S. Wiggins. Chaotic Transport in Dynamical Systems. Springer, Berlin, first edition, 1992.
827. S. Wiggins. The dynamical systems approach to Lagrangian transport in ocean flows. Annu.
Rev. Fluid Mech., 37:295–338, 2005.
828. S. Wiggins, G. Haller, and I. Mezić. Normally Hyperbolic Invariant Manifolds in Dynamical
Systems. Springer, New York, 1994.
829. S. Wiggins, L. Wiesenfeld, C. Jaffé, and T. Uzer. Impenetrable barriers in phase-space. Phys.
Rev. Lett., 86:5478–5481, 2001.
830. J. W. Wilder. Effect of initial condition sensitivity and chaotic transients on predicting future
outbreaks of gypsy moths. Ecological Modelling, 136:49–66, 2001.
831. F. H. Willeboordse. Supertransients and suppressed chaos in the diffusively coupled logistic
lattice. Chaos, 89:89–98, 1994.
832. A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano. Determining Lyapunov exponents
from a time series. Physica D, 16:285–317, 1985.
490 References

833. S. Wolfram. Cellular Automata and Complexity. Addison-Wesley, Reading, MA, first edition,
1994.
834. M. Woltering and M. Markus. Riddled-like basins of transient chaos. Phys. Rev. Lett., 84:630–
633, 2000.
835. Z.-B. Wu. Generalization of the JTZ model to open plane wakes. Chaos, 20:013122, 2010.
836. Y. Xiao, Y. Wang, and Y.-C. Lai. Dependence of intermittency scaling on threshold in chaotic
systems. Phys. Rev. E, 80:057202, 2009.
837. B. Xu, Y.-C. Lai, L. Zhu, and Y. Do. Experimental characterization of transition to chaos in
the presence of noise. Phys. Rev. Lett., 90:164101, 2003.
838. T. Yalcinkaya and Y.-C. Lai. Chaotic scattering. Comp. Phys., 9:511–518, 1995.
839. Y. Y. Yamaguchi, J. Barré, F. Bouchet, T. Dauxois, and S. Ruffo. Stability criteria of the
Vlasov equation and quasi-stationary sates of the HMF model. Physica A, 337:36–66, 2004.
840. W. Yang, M. Ding, A. J. Mandell, and E. Ott. Preserving chaos: control strategies to preserve
complex dynamics with potential relevance to biological disorders. Phys. Rev. E, 51:102–110,
1995.
841. H. W. Yin, J. D. Dai, and H. J. Zhang. Average lifetime and geometric properties for superlong
transients in a hybrid optical bistable system. Phys. Rev. E, 54:371–375, 1996.
842. S. Yonker and R. Wackerbauer. Nonlocal coupling can prevent the collapse of spatiotemporal
chaos. Phys. Rev. E, 73:026218, 2006.
843. J. A. Yorke and E. D. Yorke. Metastable chaos: the transition to sustained chaotic behavior in
the Lorenz model. J. Stat. Phys., 21:263–277, 1979.
844. L. Yu, E. Ott, and Q. Chen. Fractal distribution of floaters on a fluid surface and the transition
to chaos for random maps. Physica D, 53:102–124, 1991.
845. J. C. Zahnow and U. Feudel. Moving finite-size particles in a flow: a physical example of
pitchfork bifurcations of tori. Phys. Rev. E, 77:026215, 2008.
846. J. C. Zahnow, R. D. Vilela, U. Feudel, and T. Tél. Aggregation and fragmentation dynamics
of inertial particles in chaotic flows. Phys. Rev. E, 77:055301, 2008.
847. J. C. Zahnow, R. D. Vilela, U. Feudel, and T. Tél. Coagulation and fragmentation dynamics
of inertial particles. Phys. Rev. E, 80:026311, 2009.
848. S. Zambrano and M. A. F. Sanjuán. Exploring partial control of chaotic systems. Phys. Rev.
E, 79:026217, 2009.
849. S. Zambrano, M. A. F. Sanjuán, and J.A. Yorke. Partial control of chaotic systems. Phys. Rev.
E, 77:055201(R), 2008.
850. J. Zamora-Munt, C. Masoller, and J. Garcı́a-Ojalvo. Transient low-frequency fluctuations in
semiconductor lasers with optical feedback. Phys. Rev. A, 81:033820, 2010.
851. G. Zaslavsky. Chaotic dynamics and the origin of statistical laws. Phys. Today, 52:39–45,
1999.
852. G. Zaslavsky. Chaos, fractional kinetics, and anomalous transport. Phys. Rep., 371:461–580,
2002.
853. G. Zaslavsky and M. K. Tippett. Connection between recurrence-time statistics and anoma-
lous transport. Phys. Rev. Lett., 67:3251–3254, 1991.
854. E. Ziemniak, C. Jung, and T. Tél. Tracer dynamics in open hydrodynamical flows as chaotic
scattering. Physica D, 76:123–146, 1994.
855. R. Zillmer, N. Brunel, and D. Hansel. Very long transients, irregular firing, and chaotic dy-
namics in networks of randomly connected inhibitory integrate-and-fire neurons. Phys. Rev.
E, 79:031909, 2009.
856. R. Zillmer, R. Livi, A. Politi, and A. Torcini. Desynchronization in diluted neural networks.
Phys. Rev. E, 74:036203, 2006.
857. H. Zou, S. Guan, and C.-H. Lai. Dynamical formation of stable irregular transients in discon-
tinuous map systems. Phys. Rev. E, 80:046214, 2009.
858. A. Zumdieck, M. Timme, T. Geisel, and F. Wolf. Long chaotic transients in complex net-
works. Phys. Rev. Lett., 93:244103, 2004.
Index

A band-merging crisis, 91, 100

abrupt bifurcation, 198, 203, 207 basic components, 100, 102, 324, 326
absorbing region, 261 basin boundary, 148, 163, 298, 337
accessible orbit, 158, 160, 165, 166 basin boundary metamorphosis, 158, 159, 163
action, 241 basin cell, 168
action integral, 449 basin size, 318
action-angle representation, 241, 449 beam of particles, 455
activated complex, 290 bifurcation
activation energy, 112, 116, 121, 125, 133 abrupt, 198, 203, 207
advantage of rarity, 381 blowout, 176, 178
advection, 344
riddling, 176, 178, 301
advection experiment, 29, 349, 355
saddle-center, 196, 203, 224
aerosol, 374
saddle-node, 94, 96, 158
algebraic decay, 211, 219, 223, 232, 262
algebraic-decay exponent, 57, 212, 217, saddle-repeller, 176, 303
221–223, 234, 245, 262 unstable-unstable pair, 298, 305
almost invariant sets, 63, 366, 441 bifurcation diagram, 79
aperiodic flows, 363, 366–369, 382 billiard, 31, 190, 233, 248, 253, 260, 262, 279
Arnol’d diffusion, 229 biological reactions, 377, 381
Arrhenius factor, 116, 126 blackout, 395
astronomy, 188 blowout bifurcation, 176, 178
attractor border state of transient chaos, 58
pullback, 136 boundary crisis, 80, 132, 323, 395, 401
random, 136 box-counting dimension, 23, 45, 48, 50, 59, 66,
snapshot, 135 194, 216, 261, 269, 277, 281, 288,
attractor-merging crisis, 91 294, 415, 446, 448
autocatalytic reaction, 377, 379 bubble, 374
autocorrelation function, 240, 245, 247, 253 burst, 98, 99, 104, 133
average lifetime, 3, 11, 110

B C
Büttiker–Landauer formula, 250 c-measure, 19, 21, 40, 42, 60, 62, 238, 388
baker map (open), 58, 61, 70, 72, 261, 266, 445 Cantor set, 39, 50, 58, 59, 150, 160–162, 164,
quantized, 261 172, 174, 192, 202, 216, 217, 230,
random, 445 267, 293, 294, 302, 349, 354, 410,
three-dimensional, 266 419
band chaotic saddle, 102, 326 Cantori, 222, 227
band region, 82, 96, 100, 102, 104 catastrophic bifurcation of a riddled basin, 179

491
492 Index

cellular automaton, 312 D

channel capacity, 407, 408 deflection angle, 244, 293, 455
chaotic advection, 29, 34, 189, 344 deflection function, 191, 192, 209, 294, 455
chaotic repeller, 7, 38, 50, 94, 101, 114, 274, delay-coordinate embedding, 414, 430
277, 410, 411 delay-differential equation, 419
chaotic saddle, 7, 28, 58, 82, 83, 95, 102, 114, delay time, 235, 367, 414, 419
132, 194, 201, 259, 261, 282, 285, delay-time function, 191, 192, 209, 231, 236,
286, 324, 331, 334, 344, 367, 374, 367, 455
388, 394, 397, 403, 416, 419, 421, density ratio, 308, 374
429, 457 development parameter, 214, 216, 259
snapshot, 136, 138, 369, 372, 383, 446 devil-staircase, 85, 409
chaotic scattering, 8, 34, 188, 279, 287, 346, dielectric cavities, 189, 232
367 dielectric cavity experiment, 31
high-dimensional, 287, 291 dimension
chaotic transients, 8, 34, 190, 212 box-counting, 23, 446, 448
correlation, 419, 426, 429
chemical attractor, 380
embedding, 415
chemical reaction, 188, 377
generalized, 437
chemical reaction experiment, 26
information, 23, 128, 447
cloud physics, 374 of singularities, 192, 194, 330
coding function, 409 partial, 59
coherent structures, 337, 363, 365 dimension density, 332, 333
competition, 381 dimension equation, 46, 50, 440
complex Ginzburg–Landau equation, 313, 319, dimension formulas, 45, 64, 270, 273, 331
323, 329 dimension of singularities, 192, 194, 200, 207,
conditionally invariant measure, 19, 34, 40, 47, 210, 217, 232, 288, 290, 294, 330
57, 60, 63, 77 dissipative system, 8, 147
conductance fluctuations, 190, 247, 252, 254, droplet dynamics, 348, 355, 360
255, 261 Duffing oscillator, 153, 165
continued fraction, 48, 50 dye boundaries, 358
continuous fractal boundaries, 150 dye exit boundary, 359, 370
contraction exponent, 60 dynamical tunneling, 254, 255, 257, 259, 260
contraction factor, 63
controlling transient chaos, 386
convection in Earth’s mantle, 372 E
convection loop experiment, 4, 25 Earth’s mantle, 372, 373
correlation dimension, 419, 426, 429 ecology, 131, 404
correlation function, 450 ecosystem, 381
coupled map lattice, 312, 314, 316, 329 edge of chaos, 337
coupling component, 100 eigenvalue equation, 42, 46
coupling orbit, 103–105 embedding dimension, 415
encoding digital information, 407
crisis, 26, 34, 79, 172, 323, 327, 408, 416
ensemble average, 8
attractor merging, 91
ensemble method, 15
band merging, 91, 100
entrainment, 355, 357
interior, 90, 98, 100, 103, 132 entropy, 268
noise-induced, 132 generalized, 437
boundary, 80, 132, 323 metric, 24
crisis-induced intermittency, 98, 132, 329 topological, 24
crisis-induced intermittency experiment, 98 environmental pollution, 371, 435
critical exponent, 86, 133 epidemiology, 131
critical noise strength, 122, 133 escape interval, 38, 52, 90, 109, 161, 393, 397
crowding index, 69, 389, 390, 425 escape rate, 10, 20, 30, 42, 43, 47, 59, 68, 72,
cylinder, 39–41, 43–45, 55, 56, 438 110, 136, 202, 213, 221, 242–244,
Index 493

262, 268, 271, 280, 334, 335, 367, nowhere differentiable, 150
374, 388, 425, 445, 447 riddled, 152
escape region, 392, 397, 402, 406 sporadically fractal, 151
escape time distribution, 12, 31, 73, 75, 212, Wada, 165
219, 234, 241, 243, 357 fractal basin boundary experiment, 28, 152
escaping channel, 298, 305, 310 fractal Weyl law, 261
excitable medium, 314, 320 free energy, 54, 111, 438
exit basin, 190, 274, 281, 296, 359 Frobenius–Perron equation, 41, 57, 63
exit boundary metamorphosis, 296 generalized, 439
exit rate, 116–120 Frobenius–Perron operator, 42, 63, 366
experiment generalized, 439, 441
advection, 29, 349, 355
chemical reaction, 26
convection loop, 4, 25 G
crisis-induced intermittency, 98 gap-filing, 103, 106, 326
dieletric cavity, 31 generalized dimensions, 437
fractal basin boundary, 28, 152 generalized entropies, 437
laser, 26, 416 generalized Frobenius–Perron equation, 439
magnetoelastic ribbon, 32, 134, 394 generalized Frobenius–Perron operator, 439,
maintaining chaos, 32, 394 441
microwave, 30, 242, 260 graphene, 248, 435
noise-induced crisis, 135 Grassberger–Procaccia algorithm, 426, 429
pendulum, 27
quantum echo, 260
quantum scattering, 30 H
reflecting ball, 291 Hall conductance, 251
scattering echo, 260 Hamiltonian system, 8, 211, 262, 273, 289,
smoke ring, 355 338, 345, 351, 353, 356
turbulent pipe, 32 Hénon map, 3, 4, 11, 15, 17, 19, 21, 22, 71, 75,
experiments, 34 80, 83, 85, 94, 97, 102, 106, 137,
extraordinarily superpersistent chaotic 158, 389, 425, 427, 429, 433
transients, 308 heteroclinic tangency, 86
heteroclinic intersection, 163, 199, 206
heteroclinic tangencies, 158
high-dimensional chaotic scattering, 287, 291
F Hölder exponent, 173
false neighbors, 415 homoclinic intersection, 163
fat fractal, 96, 175, 179, 302 homoclinic orbit, 14
filamentary fractal boundaries, 150 homoclinic tangency, 88, 158
finite-size Lyapunov exponent, 364 horseshoe, 14, 163, 164, 206, 213, 259, 405
finite-size particles, 373 hyperbolic chaotic saddle, 21, 62
finite-time Lyapunov exponent, 212, 364 hyperbolic component, 219–220, 238, 349,
Fokker-Planck equation, 307 354, 375
food-chain model, 400 hyperbolic point, 12
forced damped pendulum, 27, 150, 165, 168 hyperbolic scattering, 193, 195, 205, 230, 242,
four-disk junction, 251, 253 252
four-dimensional map, 283 hyperbolic structure, 13, 268
four-hill problem, 198–200
fractal basin boundary, 34, 116, 149, 153, 277,
337, 359, 374, 377, 391, 419 I
continuous, 150 Ikeda map, 91, 117, 141
filamentary, 150 impact parameter, 190, 223
intermingled, 153 incomplete horseshoe, 215, 216
intertwined, 150 indeterminate bifurcation, 169
494 Index

inertial parameter, 308, 374 Lyapunov exponent, 13, 22, 44, 59, 64, 68, 123,
inertial particles, 373, 435 124, 126, 137, 195, 202, 268, 271,
information dimension, 23, 45, 46, 48, 61, 65, 299, 305, 324, 367, 425, 430, 440,
128, 138, 194, 269, 272–281, 379, 445, 447
418, 425, 440, 447
information dimension equation, 46, 440
information flow, 24, 67
M
interference oscillations, 457
interior crisis, 90, 98, 100, 103, 132, 324, 325 magnetoelastic ribbon experiment, 32, 134,
intermediate complex, 188, 391 394
intermediate time exponential decay, 218, 219, maintaining chaos, 32, 394, 395, 403
231, 238 maintaining chaos experiment, 32, 394
intermingled basins, 153 Maslov index, 241, 450, 451, 453
intermittency, 55, 98, 99, 119, 315, 329, 393 material lines, 363
internal degrees of freedom, 289 mean first exit time, 116
intertwined fractal boundaries, 150 mediating orbit, 86–88, 95, 101, 103, 105, 299,
intervals of continuity, 192, 290, 294, 456 325, 393
invertible dynamics, 6 mesoscopic junction, 189, 247
metamorphosis
basin boundary, 158, 159, 163
J exit boundary, 296
Julia set, 7, 114, 270 metric entropy, 24, 45, 48, 65
microfluidics, 362
microlasing cavities, 262
K microwave experiment, 30, 242, 260
KAM tori, 203, 204, 207, 210, 211, 214, 219, momentum representation, 452
220, 222, 226, 227, 234, 246, 255, Morse potential, 207, 291
260, 295, 355 multifractal spectra, 437
Kantz–Grassberger formula, 65, 379, 425 multiscale system, 285
kicked harmonic oscillator map, 7
Kuramoto–Shivashinsky equation, 312, 314,
319, 324
N
nanostructures, 189, 247, 261, 435
L natural measure, 8, 18, 20, 22, 34, 40, 48,
60–62, 75, 425
Lagrangian coherent structures, 363, 365, 373
Lagrangian dynamics, 344, 345, 363, 365 Navier–Stokes equation, 313, 320
laser experiment, 26, 416 network, 320
leaked advection, 370 Newhouse interval, 84, 85
leaked dynamics, 219, 238, 370 noise, 107, 108, 232, 305, 321, 322, 423
leaked quantum systems, 261 noise amplitude, 108, 232, 405
leaked systems, 70, 261, 370, 390 noise-induced chaos, 119, 128, 131
leapfrogging vortex pairs, 353, 356, 377 noise-induced crisis, 132, 133
lifetime, 299 noise-induced crisis experiment, 135
lifetime function, 13, 192, 268, 330, 398 noise-induced transient chaos, 308
lobe dynamics, 216, 356, 361, 363 noise-induced unstable dimension variability,
local Lyapunov exponent, 46, 68, 69, 103, 438 129
logistic map, 39, 43, 94, 100, 114, 119, 177, nonattracting chaotic set, 5, 8, 10, 13, 92, 104
408, 412, 438 nonchaotic transient, 9
Lorenz system, 34 nonhyperbolic component, 218, 220, 238, 349,
Loschmidt echo, 262 354, 391
loss region, 393–394 nonhyperbolic effects, 10, 55, 83, 85, 131, 132,
Lozi map, 80, 87 193, 211
Lyapunov density, 332 nonhyperbolic repeller, 57
Index 495

nonhyperbolic scattering, 193, 195, 203, 205, quantum echo experiment, 260
210, 222, 230, 232, 245, 254, 259, quantum fractal eigenstate, 261
349, 457 quantum mechanics, 240
noninvertible systems, 6 quantum Poincaré recurrences, 262
normally hyperbolic invariant manifolds, 289, quantum scattering experiment, 30
457 quasiattractor, 331, 337
nowhere differentiable fractal boundaries, 150 quasipotential, 111, 113, 118, 125, 306
quasipotential plateau, 114, 125

O
one-dimensional map, 38, 81, 90, 100, 109,
112, 114, 161, 408, 437 R
open aperiodic flows, 366, 369 rainbow singularities, 456
open flows, 344 random attractor, 136
open quantum systems, 261, 262, 435 random baker map, 445
open random maps, 136 random field Ising chain, 53
Ott–Grebogi–Yorke method, 386 random Hénon map, 137
random Ikeda map, 141
random maps, 134, 139, 368, 369, 381, 445
P random matrix, 244, 254
paradox of plankton, 381 random variable, 108
partial box-counting dimension, 59, 66, 194, ray dynamics, 233
269 reaction–diffusion equation, 313, 314, 319
partial dimension, 63, 269, 446 recurrence time, 70, 74, 219, 421, 422
partial information dimension, 61, 65, 269 recurrence time distribution, 75, 262
partition function, 53, 438 reflecting ball experiment, 291
path probability, 24, 44 reflecting balls, 291
pendulum experiment, 27 resetting mechanism, 370
period-3 implies chaos, 25 restraining region, 10, 11, 38, 62, 266, 271,
periodic flows, 346 282, 285, 439
periodic orbit expansion, 337, 443 reverberation time, 74
periodic window, 93, 100, 102, 114, 119, 122, Reynolds number, 32, 333–335, 346
408 riddled basins, 34, 152, 175, 301, 391
phase synchronization, 170 riddling bifurcation, 176, 178, 301
physiology, 131 Rössler model, 127, 170, 422
PIM-triple method, 17, 200, 227, 350, 354,
392, 425, 429
plankton dynamics, 381
Poincaré recurrences, 70, 74, 219, 262 S
population dynamics, 131, 313, 377, 402 S-matrix, 30, 240, 242, 244, 253, 260, 449, 450
power law decay, 10, 57, 193 Sabine’s law, 74
predictability, 153 saddle-center bifurcation, 176, 196, 203, 224,
preferential concentration, 374 303
primary escape interval, 42, 90, 109, 161, 393, saddle-node bifurcation, 94, 96, 158
397, 398, 411, 412 saddle point, 12
primary intersection point, 214, 356 safe set, 405
production term, 379 scaffolding, 168
pullback attractor, 136 scattering cross section, 244, 455–458
scattering echo, 259, 261
scattering echo experiment, 25
Q scattering function, 191, 209, 289
quantized baker map, 261 scattering region, 190, 192–194, 196–199, 201,
quantum chaotic scattering, 30, 34 208–210, 213, 217, 218, 224, 230,
quantum dots, 189, 247, 248, 252, 254–256 236, 346
496 Index

Schrödinger equation, 450 symbolic dynamics, 24, 44, 73, 408, 437
semiclassical approximation, 30, 241, 243, synchronization, 170, 175, 301, 380
245, 248, 253, 450, 457
separation of time scales, 3, 284
separation profile, 365 T
shimmying wheels, 51 tent map, 47, 59
slow manifold, 284, 377 three-body problem, 188
Smale horseshoe, 14, 405 three-dimensional baker map, 266
smoke ring experiment, 355 three-dimensional billiard, 354
snapshot attractor, 135, 139 three-dimensional map, 266
snapshot chaotic saddle, 136, 138, 369, 372, three-dimensional scattering, 354
381, 383, 446, 447 three-disk system, 190
source-inversion problem, 376 three-hill system, 196–206, 456, 457
spatiotemporal chaos, 311, 327, 435 time series analysis, 27, 414
spatiotemporal intermittency, 315, 329 topological entropy, 24, 46, 61, 66, 72, 103,
species extinction, 399 104, 193, 196, 215, 268, 316, 320,
species preservation, 405 369, 408, 410, 425
sporadically fractal boundaries, 151, 170 transfer operator, 42, 63, 366
spreading of pollutants, 365 transient chaos in spatiotemporal systems, 34
sprinkler method, 17, 21, 136, 220, 221, 285, transport processes, 188
418, 425 transverse Lyapunov exponent, 175, 178, 183
stable chaos, 320, 321 trapping region, 168
stable manifold, 7, 14, 21, 87, 136, 158, 163, traveling waves, 337
166, 261, 267, 270, 285, 290, 297, Troll–Smilansky model, 223–225, 246
329, 331, 337, 356, 358, 359, 363, tunneling time, 299, 305
367, 372, 378, 395, 397, 418, 426
turbulence, 311, 320, 333, 338, 435
stadium, 213, 253
turbulence in pipe flows, 333
stagger-and-step method, 282–285, 419
turbulent jet, 355
standard map, 219, 220
turbulent pipe experiment, 32
stationary phase approximation, 450
two-dimensional map, 58, 150, 156, 163, 172,
steady state, 379, 381, 383
179, 274, 300, 302, 441
stickiness effect, 211, 219
type-I supertransients, 317, 338
stirring region, 346, 348, 356, 367, 378
type-II supertransients, 317, 318, 322, 335
stochastic equation, 108
strange nonchaotic
attractor, 57, 333
repeller, 57 U
saddle, 333 uncertainty exponent, 154, 178, 185, 190, 192,
streakline, 351, 355 232, 330
stream function, 345, 346, 352, 354 unstable dimension variability, 130, 181
stretching exponent, 22, 44, 438, 445 unstable manifold, 7, 14, 20, 30, 87, 136, 158,
stretching factor, 22, 44, 60, 63, 64, 267, 303, 163, 166, 238, 261, 267, 270, 285,
441 290, 331, 348, 355, 356, 359, 360,
stretching field, 363 363, 367, 370, 372, 381, 418, 441
structural stability, 85 unstable periodic orbits, 24, 67, 173, 421, 442
structurally stable, 267 unstable-unstable pair bifurcation, 298, 305
superexponential scaling, 336
superpersistent chaotic transients, 90, 298, 310
supertransients, 311, 329 V
type-I, 317, 338 Van der Pol oscillator, 34
type-II, 317–319, 335 voltage collapse, 395
surrounding chaotic saddle, 102, 324–326, 328 von Kármán vortex street, 29, 309, 346, 374,
surrounding region, 96, 100, 102, 104 376, 380
survival probability, 11, 56, 73, 212, 217, 226 vortex dynamics, 351, 367
Index 497

W Y
Wada basin boundaries, 34, 165, 169, 190, 297, Yorke’s game of survival, 405
360
Wada dye boundaries, 361
weak-noise limit, 108, 111, 232, 305, 321, 404,
423 Z
Weierstrass curve, 151 zeta function, 442