UNDERSTANDING Springer

COMPLEX SYSTEMS COMPLEXITY

Bernd Krauskopf
Hinke M. Osinga
Jorge Galn-Vioque
Editors

Numerical
Continuation
Methods
for Dynamical
Systems
Path following
and boundary value problems

 

..

Springer Complexity
Springer Complexity is an interdisciplinary program publishing the best research
and academic-level teaching on both fundamental and applied aspects of complex
systems cutting across all traditional disciplines of the natural and life sciences,
engineering, economics, medicine, neuroscience, social and computer science.
Complex Systems are systems that comprise many interacting parts with the ability to generate a new quality of macroscopic collective behavior the manifestations
of which are the spontaneous formation of distinctive temporal, spatial or functional
structures. Models of such systems can be successfully mapped onto quite diverse
real-life situations like the climate, the coherent emission of light from lasers,
chemical reaction-diffusion systems, biological cellular networks, the dynamics of
stock markets and of the internet, earthquake statistics and prediction, freeway trafc, the human brain, or the formation of opinions in social systems, to name just
some of the popular applications.
Although their scope and methodologies overlap somewhat, one can distinguish
the following main concepts and tools: self-organization, nonlinear dynamics, synergetics, turbulence, dynamical systems, catastrophes, instabilities, stochastic processes, chaos, graphs and networks, cellular automata, adaptive systems, genetic algorithms and computational intelligence.
The two major book publication platforms of the Springer Complexity program
are the monograph series Understanding Complex Systems focusing on the various applications of complexity, and the Springer Series in Synergetics, which is
devoted to the quantitative theoretical and methodological foundations. In addition
to the books in these two core series, the program also incorporates individual titles
ranging from textbooks to major reference works.

Understanding Complex Systems

Founding Editor: J.A. Scott Kelso

Future scientic and technological developments in many elds will necessarily depend upon coming to grips with complex systems. Such systems are complex in both their composition typically
many different kinds of components interacting simultaneously and nonlinearly with each other
and their environments on multiple levels and in the rich diversity of behavior of which they are
capable.
The Springer Series in Understanding Complex Systems series (UCS) promotes new strategies
and paradigms for understanding and realizing applications of complex systems research in a wide
variety of elds and endeavors. UCS is explicitly transdisciplinary. It has three main goals: First,
to elaborate the concepts, methods and tools of complex systems at all levels of description and in
all scientic elds, especially newly emerging areas within the life, social, behavioral, economic,
neuro- and cognitive sciences (and derivatives thereof); second, to encourage novel applications
of these ideas in various elds of engineering and computation such as robotics, nano-technology
and informatics; third, to provide a single forum within which commonalities and differences in the
workings of complex systems may be discerned, hence leading to deeper insight and understanding.
UCS will publish monographs, lecture notes and selected edited contributions aimed at communicating new ndings to a large multidisciplinary audience.

Editorial and Programme Advisory Board

Pter rdi
Center for Complex Systems Studies, Kalamazoo College, USA
and Hungarian Academy of Sciences, Budapest, Hungary

Karl Friston
Institute of Cognitive Neuroscience, University College London,
London, UK

Hermann Haken
Center of Synergetics, University of Stuttgart, Stuttgart, Germany

Janusz Kacprzyk
System Research, Polish Academy of Sciences, Warsaw, Poland

Jrgen Kurths
Nonlinear Dynamics Group, University of Potsdam,
Potsdam, Germany

Linda Reichl
Center for Complex Quantum Systems, University of Texas,
Austin, USA

Peter Schuster
Theoretical Chemistry and Structural Biology, University of Vienna, Vienna, Austria

Frank Schweitzer
System Design, ETH Zurich, Zurich, Switzerland

Didier Sornette
Entrepreneurial Risk, ETH Zurich, Zurich, Switzerland

Bernd Krauskopf . Hinke M. Osinga . Jorge

Galn-Vioque (Eds.)

Numerical Continuation
Methods for Dynamical
Systems
Path following and boundary value problems

With 200 Figures

Dedicated to Eusebius J. Doedel for his 60th birthday

ABC

Dr. Bernd Krauskopf

Dr. Hinke M. Osinga

Dept of Engineering Mathematics

University of Bristol
Bristol
BS8 1TR
United Kingdom

Dept of Engineering Mathematics

University of Bristol
Bristol
BS8 1TR
United Kingdom

Dr. Jorge Galn-Vioque

Applied Mathematics II
University of Sevilla
Escuela Superior de Ingenieros
Camino de los Descubrimientos, s/n,
41092 Sevilla
Spain

A C.I.P. Catalogue record for this book is available from the Library of Congress
ISSN 1860-0832
ISBN 978-1-4020-6355-8 (HB)
ISBN 978-1-4020-6356-5 (e-book)
Published by Springer,
P.O. Box 17, 3300 AA Dordrecht, The Netherlands
In association with
Canopus Publishing Limited,
27 Queen Square, Bristol BS1 4ND, UK
www.springer.com and www.canopusbooks.com
All Rights Reserved
Canopus Publishing Limited 2007
No part of this work may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means, electronic, mechanical, photocopying, microlming, recording or otherwise, without written permission from the Publisher, with the exception of
any material supplied specically for the purpose of being entered and executed on a
computer system, for exclusive use by the purchaser of the work.

A Continuing Inuence in Dynamics

A well-established method for studying a given dynamical system is to identify

the compact invariant objects, such as equilibria, periodic orbits and invariant
tori, and to consider the local behavior around them. This local information
then needs to be assembled in a consistent way, frequently with the help of
geometric and topological arguments, to obtain a unied global picture of the
system. The aim is to nd qualitative (and often also quantitative) representations of the dierent types of behavior that the system may exhibit in
dependence of key parameters. The main result of such an eort is a bifurcation diagram, that is, information on the division of parameter space into
regions of topologically dierent behavior together with representative phase
portraits. The list of theoretical tools one may employ is long, well developed
and dates back at least to the 19th century. However, even when one considers seemingly simple systems, theoretical tools need to be supplemented with
numerical calculations.
Mainly for technological reasons, numerical methods have a shorter history
than theoretical tools. The rst and commonly used tool is numerical time
integration, which allows one to explore the dynamics by solving a (possibly
large) number of initial value problems. This approach is very practical for
the representation of chaotic attractors, and especially their ngerprints in
a suitable Poincare section. However, when it comes to the study of how the
behavior changes as a function of parameters the tool of choice is numerical
continuation one also speaks of path following or homotopy methods. The
basic idea is to compute an implicitly dened curve of a suitable system of
equations that denes the dynamical object under consideration. In its basic
form, numerical continuation implements the stability and bifurcation theory
of equilibria of dierential equations. More global objects, such as periodic
and homoclinic orbits, and their bifurcations can be computed by setting up
dening equations in the form of boundary value problems.
Path following in combination with boundary value problem solvers has
emerged as a continuing and strong inuence in the development of dynamical
systems theory and its application in many diverse elds of science. It is widely
acknowledged that the software package Auto developed by Eusebius J.
Doedel about thirty years ago and further expanded and developed ever since
plays a central role in the brief history of numerical continuation. When
we were thinking how best to present the origin and development of Auto,

VI

a copy of the rst edition of the Auto86 manual, with its authentic ochre
Caltech cover, came in very handy. We quote from the preface, dated May
1986:
The Auto package was rst written in 1979. It was based on a
related program written in 1976 while the author was working with H.B
Keller at the California Institute of Technology. A rst publication
referring to the package by its current name appeared in [22].
Applications often revealed some inadequacy in the algorithms and
resulted in changes. The applications also pointed to additional capabilities that would be useful to have integrated in the package, and eort
was spent on making it easy to use. This explains the delay in publication of an extensive account of the algorithm implemented. Indeed, the
dierence in eort between a theoretical analysis of a new method and
its implementation and integration appears to be considerable. We are
condent, however, that the methods and software presented here will
be of some use in the numerical exploration of nonlinear phenomena
in ordinary dierential equations.
This quote not only highlights the intricate interplay at the very earliest stage
between the development of the software and applications, but it also contains
a major understatement: Auto has not just been of some use, but it has
been used by many hundreds of researchers from all around the world! To
give a rough idea of its impact in the general scientic community, ISI Web of
Knowledge reveals that the dierent versions of the Auto manual, which was
never published other than as a Caltech preprint, has more than 700 citations.
Similarly, the seminal reference [22] in the quote, the paper E.J. Doedel, Auto:
A program for the automatic bifurcation analysis of autonomous systems,
Cong. Num. 30 (Proc. 10th Manitoba Conf. Num. Math. and Comp.), 1981,
265284, has more than 400 citations.
This book has been compiled on the occasion of Sebius Doedels 60th
birthday with the aim to illustrate the power and versatility of numerical
continuation techniques. As is demonstrated in the chapters of this book,
many recent developments build on the ideas of Sebius Doedel as implemented
in the package Auto, whose core of path following routine and collocation
boundary value problem solver is essentially still the same as when it was
released in 1986. It lies in the nature of the subject and the versatility of
Sebius Doedels work that we had to make a choice about which topics to
include. The emphasis of this book is on continuation methods for dierent
types of systems and dynamical objects, and on examples of how numerical
bifurcation analysis can be used in concrete applications. While recognizing
that there are other topics that could have been included, we believe that this
choice is in the spirit of the original motivation for the development of Auto
as expressed in the above quote. In this way, we hope to give an impression
of the continuing inuence and future potential of these powerful numerical
methods for the bifurcation analysis of dierent types of dynamical systems.

A Continuing Inuence in Dynamics

VII

The book opens with an extended foreword by Herb Keller, who is widely
recognized as the founding father of numerical continuation. Chapter 1 is an
edited part of lecture notes that Sebius Doedel has been using in his own
courses. It introduces the basic concepts of numerical bifurcation analysis and
forms a basis for the remainder of the book. The other eleven chapters by
leading experts focus on selected topics that have been inuenced strongly by
Sebius Doedels work. In fact, at least half of the chapters discuss research
in which he has been involved as a co-author. Chapter 2 by Willy Govaerts
and Yuri Kuznetsov surveys recent developments of interactive continuation
tools. Chapter 3 by Mike Henderson is concerned with higher-dimensional
continuation, and Chap. 4 by Bernd Krauskopf and Hinke Osinga discusses
the computation of invariant manifolds with a continuation approach. The
next three chapters are devoted to applications. In Chap. 5 Don Aronson and
Hans Othmer consider the dynamics of a SQUID consisting of two Josephson
junctions. Chapter 6 by Sebastian Wieczorek discusses global bifurcations
in laser systems, and Chap. 7 by Emilio Freire and Alejandro RodrguezLuis demonstrates the use of numerical bifurcation analysis for the study of
electronic circuits. The remaining chapters deal with continuation for special types of dynamical systems. Chapter 8 by John Guckenheimer and Drew
LaMar is concerned with slow-fast systems, and Chap. 9 by Jorge Gal
anVioque and Andre Vanderbauwhede with symmetric Hamiltonian systems.
Spatially extended systems are the topic of Chap. 10 by Wolf-J
urgen Beyn
and Vera Th
ummler and of Chap. 11 by Alan Champneys and Bj
orn Sandstede. Finally, in Chap. 12 Dirk Roose and R
obert Szalai survey numerical
continuation techniques for systems with delay.
We are very grateful for the enthusiastic support from all who were involved in this book project. First of all, we thank all authors for their contributions and for making every eort to stay within the limits of a tight
production schedule. We also thank Tom Spicer of Canopus Publishing Ltd
for his support of this project from its conception to the nal production of
the book. Last, but not least, we would like to thank Sebius Doedel for his
support over many years of collaboration, and for agreeing to the publication
of Chap. 1 without knowing exactly what we were up to. Happy birthday,
Sebius!
Bernd Krauskopf, Hinke Osinga and Jorge Gal
an-Vioque
Bristol and Sevilla, March 2007.

Foreword
Herbert B Keller
California Institute of Technology, Pasadena, USA, and University of California,
San Diego, USA.

Sebius (diminutive for Eusebius) Doedel obtained his Ph.D. in Applied Mathematics from the University of British Columbia in 1976. His advisor was my
friend and ex-colleague Jim Varah. As a consequence, I was able to employ
Sebius as a Research Fellow in Applied Mathematics at Caltech in 1975. Over
the next 26 years, he spent 13 of them at Caltech. However, he was also much
appreciated at Concordia University, where he was employed in 1979 and
rapidly rose to Professor of Computer Science, winning many awards which
fortunately included several years on leave with pay!
We cycled together occasionally, even in Holland, where Sebius was born.
I remember one ride in particular, when on a rather warm day we went east
from Pasadena to Claremont, about 28 miles each way. Somewhere along the
way, I became quite thirsty and so we stopped to get a cool drink. We sat
outside, relaxed and, I thought, enjoyed our drinks. I started to wax poetic
about how nice it was enjoying the outdoors and lovely California weather
when Sebius said: Herb, do you know where we are? I said: sure, near
Claremont. He replied: This is Pomona, the drive-by shooting capital of the
world I cant wait to get out of here. I have never again enjoyed going
past that part of our ride.
Early during his rst appointment at Caltech, Sebius became interested
in bifurcation phenomena, two-point boundary value problems and numerical
path-following or continuation methods, perhaps as a result of sitting in on my
course in these areas. In 1976, I was writing the paper [17] in which pseudoarclength continuation was introduced and Sebius was willing to do some calculations to illustrate how these methods worked. Of course, he did a wonderful
job producing all of the results in 7 of that paper, but more importantly, as
a result, he essentially started working on Auto at that time. The rst publication on Auto appeared in 1981 [7]. It has evolved dramatically since then,
culminating in Auto2000 [22], a fully parallel code in C++ with great graphics (available for free via http://sourceforge.net/projects/auto2000/)
that was produced mainly by Randy Paenroth, working at Caltech under
Sebiuss direction.

Herbert B Keller

Auto is without a doubt the most powerful and ecient tool for determining the bifurcation structure of nonlinear parameter-dependent systems of algebraic and ordinary dierential equations. Inuenced by his colleagues at the
University of British Columbia, who developed COLSYS [5], Sebius has employed orthogonal collocation approximations and mesh renement to obtain
extremely high accuracy. The code is able to determine heteroclinic, homoclinic and periodic orbits, both stable and unstable, by means of a two-point
boundary value problem formulation. Using a brilliant elimination procedure,
the relatively sparse Jacobian is reduced to a low-dimensional dense matrix
from which the Floquet multipliers are computed. The bifurcation structure
at singular points is readily determined in this way.
The development of Auto is but one of the main projects that Sebius
has undertaken. In the course of this work powerful theoretical results have
been produced, many with colleagues and his students, in the general areas
of bifurcation theory, dynamical systems, periodic orbits, delay dierential
equations, collocation methods for nonlinear elliptic PDEs, coupled oscillator
theory, control of bifurcation phenomena, continuation theory of manifolds,
and numerous additional topics. However, there is no doubt that the Auto
software has had a tremendous impact on many applied mathematics areas
and is, indeed, one of the leading tools in scientic computing. The code has
been incorporated into many other large software systems that solve nonlinear
problems involving continuation and bifurcation phenomena.
Essentially, all of the authors contributing to this volume have been coauthors with Sebius on papers related to the work presented here. However,
I would like to point out a few of my favorite contributions made through
these collaborations, not all of which have been fully appreciated yet. A brilliant contribution is contained in a paper by Wolf-J
urgen Beyn and Sebius
Doedel [6], in which it is shown that a continuous nonlinear boundary value
problem and the corresponding discretized problem have the same number of
solutions for all suciently ne meshes.
Sebius introduced a very powerful technique to keep computed families of
periodic solutions of autonomous dierential equations in phase. The idea is
simply to minimize the distance between neighboring solutions with respect
to a change in phase. That is, if u(t, ) is the solution over the normalized
period 0 t 1 at parameter value , then the neighboring solution u(t, +)
with phase shift lies at distance
 1
|| u(t + , + ) u(t, ) ||2 dt.
D2 () =
0

We seek to minimize this distance with respect to the phase shift . Standard
calculus of variations near zero leads to the integral condition
 1
u (t, + ) u(t, ) dt = 0.
0

Foreword

XI

This is a generalization of the standard Poincare phase or transversality condition that is applied only at one point on the orbit. However, the above global
condition is much more robust in calculations, as has been shown in many
examples [9, 10, 12] (the Poincare condition remains preferable for analytical
proofs). I am not sure when this global condition rst appeared in the literature, but we have referred to it in [16] as having been introduced by Sebius
in 1981 [7], which also happens to be his rst publication on Auto.
Many of Sebiuss publications have to do with periodic solutions of dynamical systems. These arise in a great variety of applications starting with
chemical reactors [23], then on to systems of oscillators [2, 21], heteroclinic
orbits [9] in which the above phase condition is crucial, resonances in excitable
systems [1] such as forced Fitzhugh-Nagumo systems, current biased and coupled Josephson junctions [3, 4], delay dierential equations [11, 13, 14], modied Van der Pol oscillators [8], conservative and Hamiltonian systems [20],
cardiac pacemakers [19], the circular restricted three-body problem and the
gure-eight orbit of Chenciner and Montgomery [12, 15], and many more. A
large number of these contributions are in the bio-physics area and, thus, it
turns out that Sebius may be a closet biologist.
More recently, Sebius has returned with others to the important problem
of computing higher-dimensional manifolds, either stable or unstable [18].
This brief account of some of Sebiuss publications and obvious collaborations does not do justice to the impact he has had in the eld of scientic
computation. He has had numerous students, extremely well trained, and
now making their own contributions. Furthermore, he has worked with many
outstanding scientists and has invariably enhanced their ability to do signicant scientic computations so much that it would be dicult to measure his
tremendous inuence in our eld. Hopefully, he will continue as he reaches
maturity.
H. B. Keller
Caltech / UCSD
November, 2006

XII

Herbert B Keller

References
1. J. C. Alexander, E. J. Doedel, and H. G. Othmer. On the resonance structure
in a forced excitable system. SIAM J. Appl. Math. 50(5):13731418, 1990.
2. D. G. Aronson, E. J. Doedel, and H. G. Othmer. An analytical and numerical
study of the bifurcations in a system of linearly-coupled oscillators. Physica D
25(1-3):20104, 1987.
3. D. G. Aronson, E. J. Doedel, and H. G. Othmer. The dynamics of coupled
current-biased Josephson junctions. II. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 1(1): 5166, 1991.
4. D. G. Aronson, E. J. Doedel, and D. H. Terman. A codimension-two point
associated with coupled Josephson junctions. Nonlinearity 10(5):12311255,
1997.
5. U. Ascher, J. Christiansen, and R. Russell. A collocation solver for mixed order
systems of boundary value problems. Math. Comp., 33(146):659679, 1979.
6. W.-J. Beyn and E. J. Doedel. Stability and multiplicity of solutions to discretizations of nonlinear ordinary dierential equations. SIAM J. Sci. Comp.,
2(1):107120, 1981.
7. E. J. Doedel. Auto: a program for the automatic bifurcation analysis of autonomous systems. Congr. Numer., 30:265384, 1981.
8. E. J. Doedel, E. Freire, E. Gamero, and A. J Rodriguez-Luis. An analytical
and numerical study of a modied Van der Pol oscillator. J. Sound Vibration,
256(4):755771, 2002.
9. E. J. Doedel and M. Friedman. Numerical computation of heteroclinic orbits.
Continuation techniques and bifurcation problems. J. Comput. Appl. Math.,
26(1-2):155170, 1989.
10. E. J. Doedel, H. B. Keller, and J.-P. Kernevez. Numerical analysis and control
of bifurcation problems II: Bifurcation in innite dimensions. Internat. J. Bifur.
Chaos Appl. Sci. Engrg., 1(4): 745772, 1991.
11. E. J. Doedel and P. C. Leung. Numerical techniques for bifurcation problems
in delay equations. Congr. Numer., 34:225237, 1982.
12. E. J. Doedel, R. C. Paenroth, H. B. Keller, D. J. Dichmann, J. Gal
an-Vioque,
and A. Vanderbauwhede. Computation of periodic solutions of conservative
systems with applications to the 3-body problem. Internat. J. Bifur. Chaos
Appl. Sci. Engrg., 13(6): 13531381, 2003.
13. K. Engelborghs and E. J. Doedel. Convergence of a boundary value dierence
equation for computing periodic solutions of neutral delay dierential equations.
J. Dier. Equations Appl., 7(6):927940, 2001.
14. K. Engelborghs and E. J. Doedel. Stability of piecewise polynomial collocation
for computing periodic solutions of delay dierential equations. Numer. Math.,
91(4):627648, 2002.
15. J. Gal
an-Vioque, F. J. Mu
noz-Almaraz, E. Freire, E. J. Doedel, and A. Vanderbauwhede. Stability and bifurcations of the gure-8 solution of the three-body
problem. Phys. Rev. Lett. 88(24):241101, 2002.
16. A. D. Jepson and H. B. Keller. Steady state and periodic solution paths: their
bifurcations and computations. In T. Kupper and H. D. Mittleman, editors,
Numerical Methods for Bifurcation Problems, pages 219246. (Birk
auser Verlag,
1984).

Foreword

XIII

17. H. B. Keller. Numerical solution of bifurcation and nonlinear eigenvalue problems. In P. H. Rabinowitz, editor, Applications of Bifurcation Theory, pages
359384. (Academic Press, New York, 1978).
18. B. Krauskopf, H. M. Osinga, E. J. Doedel, M. E. Henderson, J. Guckenheimer,
A. Vladimirsky, M. Dellnitz, and O. Junge. A survey of methods for computing (un)stable manifolds of vector elds. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 15(3):763791, 2005.
19. T. Krogh-Madsen, L. Glass, E. J. Doedel, and M. R. Guevara. Apparent discontinuities in the phase-resetting response of cardiac pacemakers. J. Theor.
Biol., 230(4):499519, 2004.
20. F. J. Mu
noz-Almaraz, E. Freire, J. Gal
an, and E. J. Doedel. Continuation of
periodic orbits in conservative and Hamiltonian systems. Physica D, 181(12):138, 2003.
21. H. G. Othmer, D. G. Aronson, and E. J. Doedel. Resonance and bistability in
coupled oscillators. Phys. Lett. A, 113(7):349354, 1986.
22. R. C. Paenroth and E. J. Doedel. The Auto2000 command line user interface.
In Ninth International Python Conference (Long Beach, California, USA 2001),
pages 233-241, 2001. Available via http://cmvl.cs.concordia.ca/.
23. A. B. Poore, E. J. Doedel, and J. E. Cermak. Dynamics of the Iwan-Blevins
Wake Oscillator Model. Internat. J. Non-Linear Mechanics, 21(4):291302,
1986.

Contents

A Continuing Inuence in Dynamics

................................................................ V
Foreword
Herbert B Keller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . IX
1 Lecture Notes on Numerical Analysis of Nonlinear
Equations
Eusebius J Doedel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Interactive Continuation Tools

Willy Govaerts, Yuri A Kuznetsov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3 Higher-Dimensional Continuation
Michael E Henderson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4 Computing Invariant Manifolds via the Continuation of
Orbit Segments
Bernd Krauskopf, Hinke M Osinga . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5 The Dynamics of SQUIDs and Coupled Pendula
Donald G Aronson, Hans G Othmer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6 Global Bifurcation Analysis in Laser Systems
Sebastian M Wieczorek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7 Numerical Bifurcation Analysis of Electronic Circuits
Emilio Freire, Alejandro J Rodrguez-Luis . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
8 Periodic Orbit Continuation in Multiple Time Scale
Systems
John Guckenheimer, M Drew LaMar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

XVI

Contents

9 Continuation of Periodic Orbits in Symmetric Hamiltonian

Systems
Jorge Gal
an-Vioque, Andre Vanderbauwhede . . . . . . . . . . . . . . . . . . . . . . . . . 269
10 Phase Conditions, Symmetries and PDE Continuation
Wolf-J
urgen Beyn, Vera Th
ummler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
11 Numerical Computation of Coherent Structures
Alan R Champneys, Bj
orn Sandstede . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
12 Continuation and Bifurcation Analysis of Delay Dierential
Equations
Dirk Roose, Robert Szalai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

List of Contributors

Don Aronson
School of Mathematics and
Institute for Mathematics
and its Applications
University of Minnesota
Minneapolis, MN 55455
USA
[email protected]
Wolf-J
urgen Beyn
Department of Mathematics
University of Bielefeld
P.O. Box 100131
33501 Bielefeld
Germany
[email protected]
Alan R Champneys
Department of Engineering
Mathematics
University of Bristol
Queens Building
Bristol BS8 1TR
United Kingdom
[email protected]
Eusebius J Doedel
Department of Computer Science
Concordia University
1455 De Maisonneuve Blvd. West
EV 3285

Montreal, QC H3G 1M8

Canada
[email protected]
Emilio Freire
Departamento de Matem
atica
Aplicada II
Escuela Superior de Ingenieros
Universidad de Sevilla
Camino de los Descubrimientos s/n
41092 Sevilla
Spain
[email protected]
Jorge Gal
an-Vioque
Departamento de Matem
atica
Aplicada II
Escuela Superior de Ingenieros
Universidad de Sevilla
Camino de los Descubrimientos s/n
41092 Sevilla
Spain
[email protected]
Willy Govaerts
Deptartment of Applied Mathematics and Computer Science
Ghent University
Krijgslaan 281 -S9
B-9000 Gent
Belgium
[email protected]

XVIII List of Contributors

Michael E Henderson
IBM Watson Research Center
P.O. Box 218
33-215
Yorktown Heights, NY 10598
USA
[email protected]
Herbert B Keller
Applied and Computational
Mathematics
California Institute of Technology
MC 217-50
Pasadena, CA 91125
USA
[email protected]
Bernd Krauskopf
Department of Engineering
Mathematics
University of Bristol
Queens Building
Bristol BS8 1TR
United Kingdom
[email protected]
Yuri A Kuznetsov
Department of Mathematics
Budapestlaan 6
3584 CD Utrecht
The Netherlands
[email protected]
Hinke M Osinga
Department of Engineering
Mathematics
University of Bristol
Queens Building
Bristol BS8 1TR
United Kingdom
[email protected]
Hans Othmer
School of Mathematics and
Digital Technology Center

University of Minnesota
Minneapolis, MN 55455
USA
[email protected]
Alejandro J Rodrguez-Luis
Departamento de Matem
atica
Aplicada II
Escuela Superior de Ingenieros
Universidad de Sevilla
Camino de los Descubrimientos s/n
41092 Sevilla
Spain
[email protected]
Dirk Roose
Department of Computer Science
KU Leuven
Celestijnenlaan 200A
B-3001 Heverlee - Leuven
Belgium
[email protected]
Bj
orn Sandstede
Department of Mathematics
University of Surrey
Guildford, GU2 7XH
United Kingdom
[email protected]
Robert Szalai
Department of Engineering
Mathematics
University of Bristol
Queens Building
Bristol BS8 1TR
United Kingdom
[email protected]
Vera Th
ummler
Department of Mathematics
University of Bielefeld
P.O. Box 100131
33501 Bielefeld
Germany
[email protected]

List of Contributors

Andr
e Vanderbauwhede
Department of Pure Mathematics
and Computer Algebra
Ghent University
Krijgslaan 281
B-9000 Gent
Belgium
[email protected]

Sebastian M Wieczorek
Mathematics Research Institute
University of Exeter
Harrison Building
North Park Road
Exeter EX4 4QF
United Kingdom
[email protected]

XIX

1
Lecture Notes on Numerical Analysis of
Nonlinear Equations
Eusebius J Doedel
Department of Computer Science, Concordia University, Montreal, Canada

Numerical integrators can provide valuable insight into the transient behavior
of a dynamical system. However, when the interest is in stationary and periodic solutions, their stability, and their transition to more complex behavior,
then numerical continuation and bifurcation techniques are very powerful and
ecient.
The objective of these notes is to make the reader familiar with the ideas
behind some basic numerical continuation and bifurcation techniques. This
will be useful, and is at times necessary, for the eective use of the software
Auto and other packages, such as XppAut [17], Content [24], Matcont
[21], and DDE-Biftool [16], which incorporate the same or closely related
algorithms.
These lecture notes are an edited subset of material from graduate courses
given by the author at the universities of Utah and Minnesota [9] and at
Concordia University, and from short courses given at various institutions,
including the Universite Pierre et Marie Curie (Paris VI), the Centre de
Recherches Mathematiques of the Universite de Montreal, the Technische
Universit
at Hamburg-Harburg, and the Benemerita Universidad Aut
onoma
de Puebla.

1.1 The Implicit Function Theorem

Before starting our discussion of numerical continuation of solutions to nonlinear equations, it is important rst to discuss under what conditions a solution
will actually persist when problem parameters are changed. Therefore, we
begin with an overview of the basic theory. The Implicit Function Theorem
(IFT) is central to our analysis and we discuss some examples. The discussion in this section follows the viewpoint of Keller in graduate lectures at the
California Institute of Technology, a subset of which was published in [23].

Eusebius J Doedel

1.1.1 Basic Theory

Let B denote a Banach space, that is, a complete, normed vector space. In
the presentation below it will be implicitly assumed that B is Rn , although
the results apply more generally. For x0 B, we denote by S (x0 ) the closed
ball of radius centered at x0 , that is,
S (x0 ) = {x B | || x x0 || }.
Existence and uniqueness of solutions is obtained by using two theorems.
Theorem 1 (Contraction Theorem). Consider a continuous function F :
B B on a Banach space B and suppose that for some x0 B, > 0, and
some K0 with 0 K0 < 1, we have
|| F (u) F (v) || K0 || u v ||, for all u, v S (x0 ),
|| F (x0 ) x0 || (1 K0 ) .
Then the equation
x = F (x),

x B,

has one and only one solution x S (x0 ), and x is the limit of the sequence
xk+1 = F (xk ),

k = 0, 1, 2, . . . .

Proof. Let x1 = F (x0 ). Then

|| x1 x0 ||=|| F (x0 ) x0 || (1 K0 ) .
Thus, x1 S (x0 ). Now assume inductively that x0 , x1 , , xn S (x0 ).
Then for k n we have
|| xk+1 xk || = || F (xk ) F (xk1 ) || K0 || xk xk1 ||
=

K0k || x1 x0 ||
k
K0 (1 K0 ) .
Thus,
|| xn+1 x0 || || xn+1 xn || + || xn xn1 || + + || x1 x0 ||
(K0n + K0n1 + + 1) (1 K0 )
= (1 K0n+1 )
.
Hence xn+1 S (x0 ), and by induction xk S (x0 ) for all k. We now show
that {xk } is a Cauchy sequence:

1 Numerical Analysis of Nonlinear Equations

|| xk+n xk || || xk+n xk+n1 || + + || xk+1 xk ||

(K0n1 + K0n2 + + 1) K0k (1 K0 )
= (1 K0n ) K0k
K0k .
For given > 0, choose k such that K0k <

1
2

. Then

|| xk+ xk+m |||| xk+ xk || + || xk+m xk || 2K0k < ,

independently of  and m. Hence, {xk } is a Cauchy sequence and, therefore,
converges to a unique limit lim xk = x , where x S (x0 ). Since we assumed
that F is continuous, we have
x = lim xk = lim F (xk1 ) = F (lim xk1 ) = F (lim xk ) = F (x ).
This proves the existence of x . We get uniqueness as follows. Suppose there
are two solutions, say, x, y S (x0 ) with x = F (x) and y = F (y). Then
|| x y ||=|| F (x) F (y) || K0 || x y || .
Since K0 < 1, this is a contradiction.




The second theorem ensures the parameter-dependent existence of a solution.

Theorem 2 (Implicit Function Theorem). Let G : B Rm B satisfy:
G(u0 , 0 ) = 0 for u0 B and 0 Rm ;
Gu (u0 , 0 ) is nonsingular with bounded inverse,
|| Gu (u0 , 0 )1 || M

for some M > 0;

G and Gu are Lipschitz continuous, that is, for all u, v S (u0 ), and for
all , S (0 ) the following inequalities hold for some KL > 0:
|| G(u, ) G(v, ) || KL (|| u v || + || ||),
|| Gu (u, ) Gu (v, ) || KL (|| u v || + || ||).

Then there exists , with 0 < , and a unique function u() that is
continuous on S (0 ), with u(0 ) = u0 , such that
G(u(), ) = 0, for all S (0 ).
If G(u, 0 ) = 0 and if Gu (u0 , 0 ) is invertible with bounded inverse, then
u0 is called an isolated solution of G(u, 0 ) = 0. Hence, the IFT states that
isolation (plus Lipschitz continuity assumptions) implies the existence of a
locally unique solution family (or solution branch) u = u(), with u(0 ) =
u0 .

Eusebius J Doedel

Proof. We use the notation G0u = Gu (u0 , 0 ). Then we rewrite the problem
as
G(u, ) = 0 G0u u = G0u u G(u, )
 1
u = G0u
[G0u u G(u, )] .



F(u,)

Hence, G(u, ) = 0 if and only if u is a xed point of F(, ). (Note that the
corresponding xed point iteration is, in fact, the Chord Method for solving
G(u, ) = 0.) We must verify the conditions of the Contraction Theorem.
Pick u, v S1 (u0 ), and any xed S1 (0 ), where 1 is to be chosen
later. Then
 1 0

F(u, ) F(v, ) = G0u

Gu [u v] [G(u, ) G(v, )] . (1.1)
By the Fundamental Theorem of Calculus, we have
 1
d
G(tu + (1 t)v, ) dt
G(u, ) G(v, ) =
dt
0
 1
Gu (tu + (1 t)v, ) dt [u v]
=
0

u (u, v, ) [u v],
=G
Then (1.1)
where in the last step we used the Mean Value Theorem to get G.
becomes
|| F(u, ) F(v, ) ||
u (u, v, ) || || u v ||
M || G0u G
 1
= M ||
Gu (u0 , 0 ) Gu (tu + (1 t)v, ) dt || || u v ||


0
1

|| Gu (u0 , 0 ) Gu (tu + (1 t)v, ) || dt || u v ||

M
0


M

KL (|| u0 (tu + (1 t)v) || + || 0 ||) dt || u v ||




S1 (u0 )

M KL 21 || u v || .
  
K0

Therefore, if we take
1 <

1
,
2M KL

then K0 < 1. The second condition of the Contraction Theorem is also satised, namely,

1 Numerical Analysis of Nonlinear Equations

|| F(u0 , ) u0 ||
= || F(u0 , ) F(u0 , 0 ) ||
 1
 1
= || G0u
[G0u u0 G(u0 , )] G0u
[G0u u0 G(u0 , 0 )] ||
 1
= || G0u
[G(u0 , 0 ) G(u0 , )] ||
M KL || 0 ||
M KL ,
where (with 0 < 1 ) is to be chosen. We want the above to be less than
or equal to (1 K0 )1 , so we choose

(1 K0 ) 1
.
M KL

Hence, for each S (0 ) we have a unique solution u(). We now show that
u() is continuous in . Let 1 , 2 S (0 ), with corresponding solutions
u(1 ) and u(2 ). Then
|| u(1 ) u(2 ) ||
= || F(u(1 ), 1 ) F(u(2 ), 2 ) ||
|| F(u(1 ), 1 ) F(u(2 ), 1 ) || + || F(u(2 ), 1 ) F(u(1 ), 2 ) ||
K0 || u(1 ) u(2 ) || +
 1
 1
[G0u u(2 ) G(u(2 ), 1 )] G0u
[G0u u(2 ) G(u(2 ), 2 )] ||
|| G0u
K0 || u(1 ) u(2 ) || +M KL || 1 2 || .

<1

Hence,
|| u(1 ) u(2 ) ||

M KL
|| 1 2 ||,
1 K0




which concludes the proof of the IFT.

So far, under mild assumptions, we have shown that there exists a locally
unique solution family u(). If we impose the condition that F(u, ) is continuously dierentiable in , then we can show that u() is also continuously
dierentiable. To this end, the Banach Lemma is very useful.
Lemma 1 (Banach Lemma). Let L : B B be a linear operator with
|| L ||< 1. Then (I + L)1 exists and
|| (I + L)1 ||

1
.
1 || L ||

Proof. Suppose I + L is not invertible. Then there exists y B, y = 0, such

that
(I + L) y = 0.

Eusebius J Doedel

Thus, y = Ly and

|| y ||=|| Ly |||| L || || y ||<|| y ||,

which is a contradiction. Therefore, (I +L)1 exists. We can bound the inverse

as follows:
(I + L)(I + L)1 = I
(I + L)1 = I L(I + L)1
|| (I + L)1 || 1+ || L || || (I + L)1 ||
1
.
|| (I + L)1 ||
1 || L ||




This proves the Banach Lemma.

The Banach Lemma can be used to show the following.

Lemma 2. Under the conditions of the IFT, there exists M1 > 0 and > 0
such that Gu (u, )1 exists and || Gu (u, )1 || M1 in S (u0 ) S (0 ).
Proof. Using again the notation G0u = Gu (u0 , 0 ), we have
Gu (u, ) = G0u + Gu (u, ) G0u

= G0u [I + (G0u )1 (Gu (u, ) G0u )].




L

Similar to how we veried the second condition of the Contraction Theorem

in the proof of the IFT, we can show that
|| L || M KL (|| u u0 || + || 0 ||) M KL 2.
As for the IFT, we choose
<

1
,
2M KL

and conclude that, therefore, (I + L)1 exists and

|| (I + L)1 ||

1
.
1 M KL 2

Hence, Gu (u, )1 exists and

|| Gu (u, )1 ||=|| (G0u )1 (I + L)1 ||

M
M1 ,
1 M KL 2

as required.




We are now ready to prove dierentiability of the solution branch.

Theorem 3. In addition to the assumptions of the IFT, assume that the
derivative G (u, ) is continuous in S (u0 ) S (0 ). Then the solution
branch u() has a continuous derivative u () on S (u0 ) S (0 ).

1 Numerical Analysis of Nonlinear Equations

Proof. Using the denition of (Frechet) derivative, we are given that there
exists Gu (u, ) such that G(u, )G(v, ) = Gu (u, ) (uv)+R1 (u, v, ),
where R1 (u, v, ) is such that
|| R1 (u, v, ) ||
0 as || u v || 0.
|| u v ||

(1.2)

Similarly, there exists G (u, ) such that G(u, ) G(u, ) = G (u, ) (

) + R2 (u, , ), where R2 (u, , ) satises
|| R2 (u, , ) ||
0 as || || 0.
|| ||

(1.3)

We must show that there exists u () such that

u() u() = u () ( ) + r(, ),
with

|| r(, ) ||
0 as || || 0.
|| ||

Now
0 = G(u(), ) G(u(), )
= G(u(), ) G(u(), ) + G(u(), ) G(u(), )
= Gu (u(), ) (u() u()) + R1 (u(), u(), )
+G (u(), ) ( ) + R2 (u(), , ).
Lemma 2 guarantees the existence of Gu (u(), )1 , and we nd
u() u() = Gu (u(), )1 [G (u(), ) ( ) (R1 + R2 )]
= Gu (u(), )1 [G (u(), ) ( ) r],
where
r = [G (u(), ) G (u(), )] ( ) + R1 + R2 .
Let us, for the moment, ignore the harmless factor Gu (u(), )1 and consider
each term of r. Since u and G are continuous, we have
|| [G (u(), ) G (u(), )] ( ) ||
0 as || || 0.
|| ||

Also, the existence of G implies (1.3)

|| R2 (u(), , ) ||
0 as || || 0.
|| ||

Using (1.2), we have

Eusebius J Doedel

|| R1 (u(), u(), ) || || u() u() ||

|| R1 (u(), u(), ) ||
0,
=
|| ||
|| u() u() ||
|| ||

as || || 0 because the second factor is bounded due to continuity of

u() (see the end of the proof of the IFT). Thus,
u () = Gu (u(), )1 G (u(), ).
To prove that u is continuous it suces to show that Gu (u(), )1 is
continuous. Indeed,
|| Gu (u(), )1 Gu (u(), )1 ||
= || Gu (u(), )1 [Gu (u(), ) Gu (u(), )] Gu (u(), )1 ||
M12 KL (|| u() u() || + || ||),
which concludes the proof of Theorem 3




Remark 1. In fact, if G is Lipschitz continuous then u is Lipschitz continuous (we already assume that Gu is Lipschitz continuous). More generally,
it can be shown that u is C k if G is C k , that is, u inherits the degree of
continuity of G.
We now give some examples where the IFT is used to show that a given
solution persists, at least locally, when a problem parameter is changed. We
also identify some cases where the conditions of the IFT are not satised.
1.1.2 A Predator-Prey Model
Our rst example is that of a predator-prey model dened as
 
u1 = 3u1 (1 u1 ) u1 u2 (1 e5u1 ),
u2 = u2 + 3u1 u2 .

(1.4)

We can think of u1 as sh and u2 as sharks, while the term (1e5u1 ) represents shing, with shing-quota . When = 0 the stationary solutions
are

3u1 (1 u1 ) u1 u2 = 0
(u1 , u2 ) = (0, 0), (1, 0), ( 13 , 2).
=0
u2 + 3u1 u2

The Jacobian matrix is



3 6u1 u2 5e5u1
u1
= J(u1 , u2 ; ).
J=
u2
1 + 3u1
Hence, we have

1 Numerical Analysis of Nonlinear Equations

u1

8
4

lambda

4
u2

Fig. 1.1. Stationary solution branches of the predator-prey model (1.4). Solution 2
and solution 4 are branch points, while solution 8 is a Hopf bifurcation point.

J(0, 0; 0) =

3 0
0 1
3 1
0 2


,

eigenvalues 3, 1 (unstable);

eigenvalues 3, 2 (unstable);

(1 )() + 2 = 0



2 + + 2 = 0
1 13
1

, eigenvalues
J( 3 , 2; 0) =
6 0

= 12 7 ;

Re( ) < 0 (stable).

J(1, 0; 0) =

All three Jacobians at = 0 are nonsingular. Thus, by the IFT, all three
stationary points persist for (small) = 0. In this problem we can explicitly
nd all solutions (see Fig. 1.1):
I:

(u1 , u2 ) = (0, 0).

II:

u2 = 0 and =

III:

u1 =

1
3

and

2
3

3(1 2u1 )
3u1 (1 u1 )
= 35 .)
. (Note that lim = lim
u1 0
u1 0
5e5u1
1 e5u1

13 u2 (1 e5/3 ) = 0 u2 = 2 3(1 e5/3 ).

These solution families intersect at two branch points, one of which is (u1 , u2 , ) =
(0, 0, 3/5).
The stability of Branch I follows from:


3 5 0
J(0, 0; ) =
, eigenvalues 3 5, 1.
0 1

10

Eusebius J Doedel
MAX U1
0.90

0.80

15
0.70

14
12

0.60
0.50

11

13
5

0.40

0.30

8
0.20
0.10

1
0.00
-0.10
0.10

0.30

0.80

0.60

0.40

0.20

0.00

0.50

0.70

1.00
0.90 lambda

Fig. 1.2. Bifurcation diagram of the predator-prey model (1.4). The periodic solution branch is also shown. For stationary solutions the vertical axis is simply u1 , while
for periodic solutions max(u1 ) is plotted. Solid/dashed lines denote stable/unstable
solutions. Open squares are branch points; the solid square is a Hopf bifurcation.

Hence, the trivial solution is unstable if < 3/5, and stable if > 3/5, as
indicated in Fig. 1.2. Branch II has no stable positive solutions. At H
0.67 on Branch III (Solution 8 in Fig. 1.2) the complex eigenvalues cross
the imaginary axis. This crossing is a Hopf bifurcation. Beyond H there are
periodic solutions whose period T increases as increases; see Fig. 1.3 for some
representative periodic orbits. The period becomes innite at = 0.7.
This nal orbit is called a heteroclinic cycle.
From Fig. 1.2 we can deduce the solution behavior for increasing : Branch
III is followed until H ; then the behavior becomes oscillatory due to the
periodic solutions of increasing period until = ; nally, the dynamics
collapses to the trivial solution (Branch I).
1.1.3 The Gelfand-Bratu Problem
The IFT is not only useful in the context of solution branches of equilibria.
The periodic orbits in Sect. 1.1.2 are also computed using the IFT principle.
This section gives an example of a solution branch of a two-point boundary
value problem. The Gelfand-Bratu problem [12] is dened as
 
u (x) + eu(x) = 0, x [0, 1],
(1.5)
u(0) = u(1) = 0.

1 Numerical Analysis of Nonlinear Equations

11

U2
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
0.10

0.20

0.30

0.40

0.50

0.60

0.70
U1

Fig. 1.3. Some periodic solutions of the predator-prey model (1.4). The nal orbits
are very close to a heteroclinic cycle.
MAX U
15.0

5
12.5

10.0

7.5

5.0

4
2.5

3
2

1
0.0
0.0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0
LAMBDA

Fig. 1.4. Bifurcation diagram of the Gelfand-Bratu equation (1.5). Note that there
are two solutions for 0 < < C , where C 3.51. There is one solution for = C
and for 0, and are no solutions for > C .

12

Eusebius J Doedel
U
15.0

12.5

5
10.0

7.5

5.0

4
3

2.5

0.0
0.30

0.10

0.50

1.00

0.80

0.60

0.40

0.20

0.00

0.70

0.90

Fig. 1.5. Some solutions to the Gelfand-Bratu equation (1.5).

If = 0 then u(x) 0 is a solution. We show here that this solution is

isolated, so that there is a continuation u = u
(), for || small. Consider

u (x) eu(x) = 0,
u = u(x; q, ).
u(0) = 0,
u (0) = q,
We want to solve u(1; q, ) = 0, for || small. Here F(0, 0) = 0.
  
F(q,)

We must show (IFT) that Fq (0, 0) uq (1; 0, 0) = 0:


uq (x) 0 eu0 (x) uq = 0,
where u0 0.
uq (0) = 0, uq (0) = 1,
Now uq (x; 0, 0) satises


uq = 0,
uq (0) = 0, uq (0) = 1.

Hence, uq (x; 0, 0) = x, so that uq (1; 0, 0) = 1 = 0.

1.1.4 A Nonlinear Eigenvalue Problem
The equations for column buckling (from nonlinear elasticity theory) [31] are
given by

1 Numerical Analysis of Nonlinear Equations

13

= 0,
u1 (x) + u1 (x)
u3 (x) + 12 u1 (x)2 + = 0,

for x [0, 1], with



u1 (0) = u1 (1) = 0,
u3 (0) = u3 (1) = ( > 0),

where is a stress (to be determined) and is another physical constant; take

= 1. Note that the boundary conditions are overspecied (to determine
). We rewrite the equations as the rst order system

u2 ,
u1 (0) = u1 (1) = 0,
u1 =
u2 = u1 ,
(1.6)

u3 = 12 u22 , u3 (0) = u3 (1) = .

Note that u1 u2 0 implies u3 = , so that u3 (x) = x, with

u3 (0) = and u3 (1) = = . Thus, we must have = 2, so that
u3 (x) = 2x = (1 2x). Hence,
u1 u2 0,

u3 (x) = (1 2x),

= 2,

is a solution for all . Are these solutions isolated? In the formal set-up,
consider

u1 =
u2 ,
u1 (0) = 0,
u2 = u1 ,
u2 (0) = p, u = u(x, p, ; ).

u3 = 12 u22 , u3 (0) = .

We must have
u1 (1, p, ; )
= 0,
u3 (1, p, ; ) + = 0,


F(p, ; ) = 0,

with F : R2 R R2 . So the question is: Is (Fp | F )() nonsingular along

the basic solution branch?
To answer the above question quickly, we omit explicit construction of F.
We linearize (1.6) about u1 ,u2 , u3 , , and , with respect to u1 , u2 , u3 , and
, acting on v1 , v2 , v3 , and , to obtain the linearized homogeneous equations

v2 ,
v1 (0) = v1 (1) = 0,
v1 =
v2 = v1 u1 ,

v3 = u2 v2 , v3 (0) = v3 (1) = 0.
In particular, the linearized homogeneous equations about u1 u2 0,
u3 (x) = (1 2x), and = 2 are

v1 =
v2 , v1 (0) = v1 (1) = 0,
v2 = 2v1 ,
= 0, v3 0.

, v3 (0) = v3 (1) = 0.
v3 =

14

Eusebius J Doedel

Now, if 2 = k 2 2 , k = 1, 2, 3, . . . , then

v1 + 2v1 = 0,
v1 (0) = v1 (1) = 0,
has the unique solution v1 0 and, hence, also v2 0. Thus, if = 12 k 2 2
then the basic solution branch is locally unique. However, if = 12 k 2 2 then
the linearization is singular, and there may be bifurcations. (In fact, there are
buckled states.)

1.1.5 The Pendulum Equation

The equation of a damped pendulum subject to a constant torque is given by

I ,
m R  (t) + m  (t) +m g sin (t) = 
  
torque
damping
that is,
 (t) +

I
g

.
sin (t) =
(t) +
mR
R
R

Scaling time as s = c t we have  =

we obtain

d
dt

d ds
ds dt

= c and, similarly,  = c2 ,

I
g

+ c (s)
,
sin (s) =
+
c2 (s)
mR
R
R
I
g

.
sin =
+
+
2
m R c2
Rc
Rc

Choose c such that Rgc2 = 1, that is, c = g/R, and set = /(R c) and
or, dropping
I = I/(m R c2 ). Then the equation becomes + + sin = I,
the , and using  ,
(1.7)
 +  + sin = I.

We shall consider special solutions, called rotations, that satisfy (t + T ) =

(t) + 2, for all t or, equivalently,
0

  
(T ) = (0) +2
 (T ) =  (0),

where T is the period.

(= 2),

1 Numerical Analysis of Nonlinear Equations

15

The Undamped Pendulum

First consider the undamped unforced pendulum
 + sin = 0,
that is, (1.7) with = I = 0. Suppose the initial data for a rotation are
(0) = 0, and  (0) = p > 0. We have (T ) = 2, and  (T ) =  (0) = p.
Integration gives
 t  
t
dt + 0  sin dt = 0
0
t


1 2 t

cos

 = 0
2
0
0

1 
2
2 (t)

1  2
(t)
2  

kinetic energy

1 2
p 1
2
1
= p2 .
2

cos (t) =

1 cos (t)



potential energy

Thus,
 (t) =

 2
0

dt
d

p2 2 + 2 cos (t)
1
dt

d =
2
p 2 + 2 cos
 2
1

d
d =
2
p 2 + 2 cos
0
 2
1

d.
T =
p2 2 + 2 cos
0

d
dt

We see that
T 0
and

T
0

as

p ,

1
d =
2 + 2 cos

as

p 2.

In fact, rotations exists for all p > 2.

The Forced Damped Pendulum
We now consider the forced damped pendulum (1.7),
 +  + sin = I,
with (0) = 0 (which sets the phase) and  (0) = p. We write the solution
as = (t; p, I, ). Do there exist rotations, i.e., does there exist T such that
(T ; p, I, ) = 2 and  (T ; p, I, ) = p?

16

Eusebius J Doedel

Theorem 4. Let 0 be a rotation of the undamped unforced pendulum:

0 + sin 0 = 0,
0 (0) = 0, 0 (0) = p0 ,
0 (T0 ) = 2, 0 (T0 ) = p0 .
Then there exist (smooth) functions T = T (p, ) and I = I(p, ), with
T (p0 , 0) = T0 and I(p0 , 0) = 0, such that (t; p, I(p, ), ) is a rotation of
period T (p, ) of the damped forced pendulum
0 +  + sin 0 = I,
(0; p, I(p, ), ) = 0,
 (0; p, I(p, ), ) = p,

(T (p, ); p, I(p, ), ) = 2, (T (p, ); p, I(p, ), ) = p.
for all (p, ) suciently close to (p0 , 0).
Proof. The Jacobian matrix with respect to T and I, of the algebraic system

(T ; p, I, ) 2 = 0,
 (T ; p, I, ) p = 0,
evaluated at p = p0 , T = T0 , and I = = 0, is
 0 
0 I

J0 =
(T0 ).
0 0I
We must show that detJ0 = 0. We have
0 + sin 0 = 0 0 (T0 ) = sin (0 (T0 )) = sin (2) = 0,
0 (T0 ) = p0 = 0.


Thus, detJ0 = 0 if 0I (T0 ) = 0. Here, I satises

I + I + I cos = 1,

I (0) = I (0) = 0.

In particular,


0I + 0I cos 0 = 1,
From
and

0I (0) = 0I (0) = 0.



0I 0 + 0I cos 0 0 = 0 ,

0 + sin 0 = 0 
0 + cos 0 0 = 0,

we have
Using integration, we nd




0I 0 0I 
0 = 0 .

1 Numerical Analysis of Nonlinear Equations

T0



0I 0

T0 
0  
I 0 
0

T0

T0

0I 
0 =

T0

17

0 = 2,

T0

 T0



0

0
 
I 0 I 0  +
0I 0 = 2,
 
0

sin 0

0I (T0 )

0


0 (T0 ) 0I (0)

  

  

p0

0 (0) = 2.

Hence,


0I (T0 ) =

2
= 0.
p0



A more general analysis of this type for coupled pendula can be found in [1]
(see also Chap. 5).

1.2 Continuation of Solutions

As mentioned, the IFT plays an important role in the design of algorithms
for computing families of solutions to nonlinear equations. Such continuation
methods are applied in a parameter-dependent setting. Hence, we consider
the equation
G(u, ) = 0,
u, G(, ) Rn , R.
Let x (u, ). Then the equation can be written as
G(x) = 0,

G : Rn+1 Rn .

1.2.1 Regular Solutions

A solution x0 of G(x) = 0 is regular [22] if the n (rows) by n + 1 (columns)
matrix G0x Gx (x0 ) has maximal rank, i.e., if Rank(G0x ) = n.
In the parameter formulation G(u, ) = 0, we have

(i) Gu is nonsingular,

or
Rank(G0x ) = Rank(G0u | G0 ) = n
dim N (G0u ) = 1,

(ii) and

G  R(G0u ).
Here, N (G0u ) denotes the null space of G0u , and R(G0u ) denotes the range of
G0u , i.e., the linear space spanned by the n columns of G0u .

18

Eusebius J Doedel
u

x0

Fig. 1.6. A solution branch of G(u, ) = 0; note the two folds.

Theorem 5. Let x0 (u0 , 0 ) be a regular solution of G(x) = 0. Then, near

x0 , there exists a unique one-dimensional continuum of solutions x(s), called
a solution family or a solution branch, with x(0) = x0 .
Proof. Since Rank(G0x ) = Rank(G0u | G0 ) = n, either G0u is nonsingular and
by the IFT we have u = u() near x0 , or else we can interchange columns
in the Jacobian G0x to see that the solution can locally be parametrized by
one of the components of u. Thus, a unique solution family passes through a
regular solution.


Remark 2. We remark here that the second case in the above proof is that of
a simple fold; see also Fig. 1.6.
1.2.2 Parameter Continuation
In the parameter-dependent setting we assume that the continuation parameter is . Suppose we have a solution (u0 , 0 ) of
G(u, ) = 0,
as well as the direction vector u 0 = du/d, and we want to compute the
solution u1 at 1 = 0 + ; this is illustrated in Fig. 1.7.
To compute the solution u1 we use Newtons method

()
()
()
Gu (u1 , 1 ) u1 = G(u1 , 1 ),
(1.8)
(+1)
()
()
u1
= u1 + u1 ,
= 0, 1, 2, . . . .
As initial approximation, we use
(0)

u1 = u0 + u 0 .
If Gu (u1 , 1 ) is nonsingular and is suciently small, then the convergence
theory for Newtons method guarantees that this iteration will converge.

1 Numerical Analysis of Nonlinear Equations

19

"u"

u (=

du
d

at 0 )

(0)

u1

Fig. 1.7. Graphical interpretation of parameter continuation.

After convergence, the new direction vector u 1 can be computed by solving

Gu (u1 , 1 ) u 1 = G (u1 , 1 ).
This equation follows from dierentiating G(u(), ) = 0 with respect to
at = 1 . Note that, in practice, the calculation of u 1 can be done without
another LU -factorization of Gu (u1 , 1 ). Thus, the extra work to nd u 1 is
negligible.
As an example, consider again the Gelfand-Bratu problem of Sect. 1.1.3
given by
 
u (x) + eu(x) = 0, x [0, 1],
u(0) = u(1) = 0.
If = 0 then u(x) 0 is an isolated solution; see Sect. 1.1.3. We discretize
this problem by introducing a mesh,
0 = x0 < x1 < < xN = 1,
xj xj1 = h, 1 j N,

h = 1/N.

The discrete equations are:

uj+1 2uj + uj1
+ euj = 0,
h2

j = 1, . . . , N 1,

with u0 = uN = 0. (More accurate discretization is discussed in Sect. 1.3.1.)

Let

u1
u2

u
.
uN 1
Then we can write the above as G(u, ) = 0, where G : Rn R Rn , with
n = N 1.

20

Eusebius J Doedel
"u"

u0
u0

( u0, 0 )
s

Fig. 1.8. Graphical interpretation of pseudo-arclength continuation.

For the parameter continuation we suppose that we know 0 , u0 , and u 0 .

(0)
Then we set 1 = 0 + and apply Newtons method (1.8) with u1 =
u0 + u 0 . After convergence nd u 1 from
Gu (u1 , 1 ) u 1 = G (u1 , 1 ),
and repeat the above procedure to nd u2 , u3 , and so on. Here,

2
1
h2 + eu1
h2
1

h22 + eu2 h12

h2

.
.
.
.
Gu (u, ) =

. .
.
2
1
uN 1
+
e

h2
h2

Hence, we must solve a tridiagonal system for each Newton iteration. The
solution branch has a fold where the parameter-continuation method fails; see
Figs. 1.4 and 1.5.
1.2.3 Kellers Pseudo-Arclength Continuation
In order to allow for continuation of a solution branch past a fold, Auto
[8, 11, 12] uses Kellers Pseudo-Arclength Continuation [22]. Suppose we have
a solution (u0 , 0 ) of G(u, ) = 0, as well as the direction vector (u 0 , 0 ) of the
solution branch. Pseudo-arclength continuation solves the following equations
for (u1 , 1 ):

G(u1 , 1 ) = 0,
(1.9)

(u1 u0 ) u0 + (1 0 ) 0 s = 0.
Figure 1.8 shows a graphical interpretation of this continuation method. Newtons method for pseudo-arclength continuation becomes

1 Numerical Analysis of Nonlinear Equations

21

"X-space"
x

x
x

Fig. 1.9. Parameter-independent pseudo-arclength continuation.

(G1u )() (G1 )()

u
0
0

()

u1

()

()

()
()
(u1 u0 ) u 0 + (1 0 ) 0 s

with the new direction vector dened

!
G1u G1
u 0
0

()

G(u1 , 1 )
as
u 1
1

!
=

0
1

!
,

Note that
In practice (u 1 , 1 ) can be computed with one extra back-substitution;
The orientation of the branch is preserved if s is suciently small;
The direction vector must be rescaled, so that indeed || u 1 ||2 + 21 = 1.
Theorem 6. The Jacobian of the pseudo-arclength system is nonsingular at
a regular solution point.
Proof. Let x = (u, ) Rn+1 . Then pseudo-arclength continuation can be
written as
G(x1 ) = 0,

(x1 x0 ) x 0 s = 0,

(|| x 0 ||= 1).

Figure 1.9 shows

0a graphical interpretation. The matrix in Newtons method
Gx
at s = 0 is
. At a regular solution we have N (G0x ) = Span{x 0 }. We
x 0
0
Gx
must show that
is nonsingular at a regular solution. Suppose, on the

0x0
Gx
contrary, that
is singular. Then there exists some vector z = 0 with
x 0
G0x z = 0

and

x 0 z = 0.

Thus, z = cx 0 , for some constant c. But then

0 = x 0 z = cx 0 x 0 = c || x 0 ||2 = c,
so that z = 0, which is a contradiction.




22

Eusebius J Doedel

Consider pseudo-arclength continuation for the discretized Gelfand-Bratu

problem of Sect. 1.1.3. Then the matrix
!
!
G u G
Gx
=
x
u
in Newtons method is a bordered tridiagonal matrix of the form

We now show how to solve such linear systems eciently.
1.2.4 The Bordering Algorithm
The linear systems in Newtons method for pseudo-arclength continuation are
of the form

  
A c
x
f
=
.
b d
y
h
The special structure of this extended system can be exploited; a general
presentation of the numerical linear algebra aspects of extended systems can
be found in [20, 24]. If A is a sparse matrix whose LU -decomposition can be
found relatively cheaply (e.g., if A is tridiagonal), then the following bordered
LU -decomposition [22] will be ecient:




L 0
U
A c
=
.
1
0
b d
After decomposing A = LU (which may require pivoting) we compute , ,
and from

L = c,

U = b,

= d .

The linear system can then be written as

1 Numerical Analysis of Nonlinear Equations

L 0
1

23

  
f
U
x
=
,
h
0
y




f

and we can compute the solution (x, y) through the following steps:

Lf = f ,

= h f ,
h

y = h/,

U x = f y.

Theorem 7. The bordering

algorithm outlined above works if A and the full

A c
are nonsingular.
matrix A
b d
In the proof of Theorem 7 we make use of the Bordering Lemma [22]


A c
. Then
Lemma 3 (Bordering Lemma). Let A
b d
(a) A nonsingular A nonsingular if and only if d = b A1 c;

c  R(A),

(b) dim N (A) = dim N (A ) = 1 A nonsingular if

b  R(A );
(c) If dim N (A) 2 then A is singular.
Proof. (a) (A nonsingular)



I A1 c
A 0
, where e = d b A1 c. Clearly,
In this case A =
e
0
b 1
A is nonsingular if and only if e = 0.
(b) (dim N (A) = 1)
Suppose A is singular in this case. Then there exist z Rn and R,
not both zero, such that
 
 

z
Az + c
0
A c
=
=
.
A=

b z + d
0
b d
We see that c R(A) if = 0, which contradicts the assumptions. On
the other hand, if = 0 and z = 0 then
N (A) = Span{z}

and

b N (A) .

24

Eusebius J Doedel

Since, in general, N (A) = R(A ), it follows that b R(A ), which also

contradicts the assumptions.
(c) (dim N (A) 2)



This case follows from a rank argument.

Proof (Theorem 7). The crucial step in the bordering algorithm is the com
putation of z = h/.
Namely, we must have = 0. Since is determined in
the bordered LU -decomposition, we have
= d (U 1 b) (L1 c)
= d
1 1
= d b U L c = d b (LU )1 c
= d b A1 c,
which is nonzero by Conclusion (a) of the Bordering Lemma.




Remark 3. In pseudo-arclength continuation we have

!
!
G u G
A c
A=
=
b d
u 0 ,
0

that is, A = Gu , which is singular at a fold. Therefore, the bordering algorithm

will fail when it is used exactly at a fold. In practice, the method may still work.
We consider another approach, used in Auto, when discussing collocation
methods in Sect. 1.3.1.

1.3 Boundary Value Problems

Consider the rst-order system of ordinary dierential equations
u (t) f (u(t), , ) = 0,

t [0, 1],

where
u(), f () Rn ,

R,

Rn ,

subject to boundary conditions

b(u(0), u(1), , ) = 0,

b() Rnb ,

and integral constraints

 1
q(u(s), , ) ds = 0,

q() Rnq .

We want to solve this boundary value problem (BVP) for u() and . In order
for this problem to be formally well posed we require that
n = nb + nq n 0.
We can think of as the continuation parameter in which the solution (u, )
may be continued. A simple case is nq = 0, nb = n, for which n = 0.

1 Numerical Analysis of Nonlinear Equations

25

1.3.1 Orthogonal Collocation

Auto solves boundary value problems using the method of orthogonal collocation with piecewise polynomials [2, 7]. This method is very accurate, and
allows adaptive mesh-selection. The set-up is as follows.
First, we introduce a mesh
{0 = t0 < t1 < < tN = 1},
with
hj = tj tj1 ,

(1 j N ).

Dene the space of (vector-valued) piecewise polynomials Phm as

#
"
Phm = ph C[0, 1] | ph |[tj1 ,tj ] P m ,
where P m is the space of (vector-valued) polynomials of degree m. The
orthogonal collocation method with piecewise polynomials [3] consists of nding ph Phm and Rn , such that the following collocation equations are
satised:
ph (zj,i ) = f (ph (zj,i ), , ),

j = 1, . . . , N,

i = 1, . . . , m,

and such that ph satises the boundary and integral conditions. The collocation points zj,i in each subinterval [tj1 , tj ] are the (scaled) roots of the
mth-degree orthogonal polynomial (Gauss points); see Fig. 1.10 for a graphical interpretation. Since each local polynomial is determined by (m + 1)n,
coecients, the total number of degrees of freedom (considering as xed) is
(m + 1)nN + n . This is matched by the total number of equations:
collocation: mnN,
continuity: (N 1)n,
constraints: nb + nq (= n + n ).
If the solution u(t) of the BVP is suciently smooth then the order of accuracy
of the orthogonal collocation method is m, i.e.,
|| ph u || = O(hm ).
At the main meshpoints tj we have superconvergence:
maxj | ph (tj ) u(tj ) |= O(h2m ).
The scalar variables are also superconvergent [7].

26

Eusebius J Doedel
0

1
t

t j-1

z j,1

t j-1

z j,2

z j,3

tj

t j-2/3 t j-1/3
lj,3(t)

tj

lj,1(t)

Fig. 1.10. The mesh {0 = t0 < t1 < < tN = 1}. Collocation points and
extended-mesh points are shown for the case m = 3, in the jth mesh interval. Also
shown are two of the four local Lagrange basis polynomials.

1.3.2 Implementation in Auto

The implementation in Auto [12] is done via the introduction of Lagrange
basis polynomials for each subinterval [tj1 , tj ]. Dene
{j,i (t)},

j = 1, . . . , N,

by
j,i (t) =

i = 0, 1, . . . , m,

m
$

t tj k

k=0,k
=i

tj mi tj k

where

i
hj .
m
The local polynomials can then be written as
tj mi = tj

pj (t) =

m
%

j,i (t)uj mi .

i=0

With the above choice of basis

uj approximates u(tj )

and uj mi approximates u(tj mi ),

where u(t) is the solution of the continuous problem.

Then the collocation equations are
pj (zj,i ) = f (pj (zj,i ), , ),

i = 1, . . . , m,

j = 1, . . . , N,

1 Numerical Analysis of Nonlinear Equations

u0

u1
3

u2
3

u1

u2

uN

27
T

Fig. 1.11. Structure of the Jacobian for the case of n = 2 dierential equations
with the number of mesh intervals N = 3, the number of collocation points per
mesh interval m = 3, the number of boundary conditions nb = 2, and the number
of integral constraints nq = 1. The last row corresponds to the pseudo-arclength
equation, which is not included in the nq = 1 count. From E.J. Doedel, H.B. Keller,
J.P. Kernevez, Numerical analysis and control of bifurcation problems (II): Bifurcation in innite dimensions, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 1(4) (1991)
c
745772 1991
World Scientic Publishing; reproduced with permission.

the discrete boundary conditions are

bi (u0 , uN , , ) = 0,

i = 1, . . . , nb ,

and the integrals constraints can be discretized as

N %
m
%

j,i qk (uj mi , , ) = 0,

k = 1, . . . , nq ,

j=1 i=0

where the j,i are the Lagrange quadrature coecients.

The pseudo-arclength equation is
 1
(u(t) u0 (t)) u 0 (t) dt + ( 0 ) 0 + ( 0 ) 0 s = 0,
0

where (u0 , 0 , 0 ), is the previously computed point on the solution branch,

and (u 0 , 0 , 0 ), is the normalized direction of the branch at that point. The

28

Eusebius J Doedel
u0

u1
3

u2
3

u1

u2

uN

Fig. 1.12. The system after condensation of parameters. The entries have been
eliminated by Gauss elimination.

discretized pseudo-arclength equation is

N %
m
%

j,i [uj mi (u0 )j mi ] (u 0 )j mi

j=1 i=0

+( 0 ) 0 + ( 0 ) 0 s = 0.
The implementation in Auto includes an ecient method to solve these
linear systems [12]; this is illustrated in Figs. 1.121.15. Note that the gures
only illustrate the matrix structure; the indicated operations are also carried
out on the right-hand side, which is not shown in the gures. Figure 1.12
shows the system after condensation of parameters. The entries marked with
have been eliminated by Gauss elimination. These operations can be done
in parallel [34]. The condensation of parameters leads to a system with a
fully decoupled sub-system that can be solved separately. The decoupled subsystem is marked by in Fig. 1.13
1.3.3 Numerical Linear Algebra
The complete discretization consists of
mnN + nb + nq + 1,

1 Numerical Analysis of Nonlinear Equations









u0









u1
3

u2
3







u1







u2

uN







































29
T



























Fig. 1.13. This is the same matrix as in Fig. 1.12, except that some entries are now
marked by a . The  sub-system is fully decoupled from the remaining equations and
can, therefore, be solved separately. From E.J. Doedel, H.B. Keller, J.P. Kernevez,
Numerical analysis and control of bifurcation problems (II): Bifurcation in innite
c
dimensions, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 1(4) (1991) 745772 1991
World Scientic Publishing; reproduced with permission.

nonlinear equations, in the unknowns

{uj mi } RmnN +n ,

Rn ,

R.

These equations can be solved by a Newton-Chord iteration. The structure

of the associated Jacobian is illustrated in Fig. 1.11 for a system of n = 2
dierential equations, with N = 3 mesh intervals, m = 3 collocation points
per mesh interval, nb = 2 boundary conditions, and nq = 1 integral constraint.
In a typical problem N will be larger, say, N = 5 for very easy problems, and
N = 200 for very dicult problems. The standard choice of the number of
collocation points per mesh interval is m = 4.
The decoupled  sub-system can be solved by nested dissection. This procedure eliminates some of the -entries, but also introduces some new nonzero
entries due to ll-in; see Fig. 1.14. However, the structure reveals a new decoupled sub-system that can be solved completely; this subsystem is highlighted
in Fig. 1.15 with +. The + sub-system consists of two sub-matrices A0 and
A1 , as in Fig. 1.15. For periodic solutions, the Floquet multipliers are the
eigenvalues of the matrix A1
1 A0 [18].

30

Eusebius J Doedel




u0
























u1
3

u2
3

u1




u2

uN












































Fig. 1.14. The decoupled  sub-system solved by nested dissection. This procedure
eliminates some of the -entries, but also introduces some new nonzero entries due
to ll-in.

1.4 Computing Periodic Solutions

Periodic solutions can be computed very eectively by using a boundary value
approach. This method also determines the period very accurately. Moreover,
the technique allows asymptotically unstable periodic orbits to be computed
as easily as asymptotically stable ones.
1.4.1 The BVP Approach.
Consider the rst-order system
u (t) = f (u(t), ),

u(), f () Rn ,

Fix the interval of periodicity by the transformation t 

becomes

u (t) = T f (u(t), ),

u(), f () Rn ,

R.
t
T

. Then the equation

T, R,

(1.10)

and we seek solutions of period 1, i.e.,

u(0) = u(1).

(1.11)

1 Numerical Analysis of Nonlinear Equations




u0










+
+
+
+
+
+

A0
+
+
+
+
+
+

u1
3

u2
3

u1




u2

uN




A1

+
+
+
+
+
+

+
+
+
+
+
+

31
T







+
+
+
+
+
+







+
+
+
+
+
+

Fig. 1.15. The same matrix as in Fig. 1.14, except with some entries now marked
by +. Note that the + sub-system is decoupled from the other equations, and can,
therefore, be solved separately.

Note that the period T is one of the unknowns.

Equations (1.10)(1.11) do not uniquely specify u and T . Assume that
we have computed (uk1 (), Tk1 , k1 ) and we want to compute the next
solution (uk (), Tk , k ). Then uk (t) can be translated freely in time: if uk (t)
is a periodic solution then so is uk (t + ) for any . Thus, a phase condition
is needed. An example is the Poincare orthogonality condition
(uk (0) uk1 (0)) uk1 (0) = 0,
where the phase of the next condition is xed such that the dierence at
time t = 0 is perpendicular to the tangent vector of the current solution; this
is illustrated in Fig. 1.16. In the next section we derive a numerically more
suitable phase condition.
1.4.2 Integral Phase Condition
k (t) is a solution then so is u
k (t + ), for any . We want the solution
If u
that minimizes
 1
k (t + ) uk1 (t) ||22 dt.
D() =
|| u
0

32

Eusebius J Doedel

k-1 (0)

u k-1 (0)

u (0)
k

Fig. 1.16. Graphical interpretation of the Poincare phase condition.

k (t+
The optimal solution u
), must satisfy the necessary condition D (
) = 0.
Dierentiation gives the necessary condition
 1
k (t +
(
uk (t +
) uk1 (t)) u
) dt = 0.
0

k (t +
Writing uk (t) u
), gives
 1
(uk (t) uk1 (t)) uk (t) dt = 0.
0

Integration by parts, using periodicity, gives

uk (t) uk1 (t) dt = 0.

(1.12)

This is the integral phase condition [8].

1.4.3 Pseudo-Arclength Continuation
In practice, we use pseudo-arclength continuation to follow a family of periodic
solutions; see Sect. 1.2.3. In particular, this allows calculation past folds along
a family of periodic solutions. It also allows calculation of a vertical family
of periodic solutions, which has important applications to the computation
of periodic solutions to conservative systems [14, 30] (see also Chap. 9). For
periodic solutions the pseudo-arclength equation is


0

(uk (t) uk1 (t)) u k1 (t) dt

(1.13)

+ (Tk Tk1 ) Tk1 + (k k1 ) k1 = s.

Equations (1.10)(1.13) are the equations used in Auto for the continuation
of periodic solutions. In summary, given uk1 , Tk1 , and k1 , we solve the
system

1 Numerical Analysis of Nonlinear Equations

33

MAX U1
1.25

1.00

0.75

0.50

0.25

0.00

-0.25
-1.00

-0.75

-0.50

-0.25

0.00

0.25

0.50

0.75

1.00

LAMBDA

Fig. 1.17. Bifurcation diagram of the stationary solution u(t) 0 of (1.14).

uk (t) = T f (uk (t), k ),




uk (0) = uk (1),
1

uk (t) uk1 (t) dt = 0,

0
1

(uk (t) uk1 (t)) u k1 (t) dt + (Tk Tk1 ) Tk1 + (k k1 ) k1 = s,

where
u(), f () Rn ,

, T R.

1.4.4 A Vertical Family of Periodic Orbits

Consider the system of equations
 
u1 = u1 u2 ,
u2 = u1 (1 u1 ).

(1.14)

Note that u(t) 0 is a stationary solution for all . Another stationary

1
.
solution is u(t)

The bifurcation diagram for u(t) 0 is shown in Fig. 1.17, but we can
also analyze the behavior analytically. The Jacobian along the solution family
u(t) 0 is


1
,
1 0

34

Eusebius J Doedel
U2
0.75

0.50

0.25

0.00

-0.25

-0.50

-0.75
-0.50

-0.25

0.50

0.25

0.00

0.75

1.00

U1

Fig. 1.18. A phase plot of some periodic solutions of (1.14).

U1
1.00

0.75

0.50

0.25

0.00

-0.25

-0.50
0.00

0.40

0.20
0.10

0.30

1.00
0.80
0.60
0.90
0.70
0.50
SCALED TIME

Fig. 1.19. Solution component u1 of (1.14) as a function of the scaled time variable t.

with eigenvalues

2 4

1 Numerical Analysis of Nonlinear Equations

35

U2
0.75

0.50

0.25

0.00

-0.25

-0.50

-0.75
0.00

0.40

0.20
0.10

0.30

1.00
0.80
0.60
0.90
0.70
0.50
SCALED TIME

Fig. 1.20. Solution component u2 of (1.14) as a function of the scaled time variable t.

Hence, the eigenvalues are complex for (2, 2). The eigenvalues cross the
imaginary axis when passes through zero. Thus, there is a Hopf bifurcation
along u(t) 0 at = 0, and a family of periodic solutions bifurcates from
u(t) 0 at = 0. As shown in Fig. 1.17, the emanating family of periodic
solutions is vertical. Some periodic solutions are shown in Fig. 1.18 in the
(u1 , u2 )-plane. These solutions are plotted versus time in Figs. 1.19 and 1.20.
Along this family the period tends to innity. The nal innite-period
orbit is homoclinic to (u1 , u2 ) = (1, 0). The time diagrams in Figs. 1.19 and
1.20 illustrate how the peak in the solution remains in the same location.
This is a result of the integral phase condition (1.12) and very advantageous
for discretization methods.
1.4.5 FitzHugh-Nagumo Equations
The Fitzhugh-Nagumo equations of nerve-conduction are


 
v = c v 13 v 3 + w ,
w = (v a + bw)/c.

(1.15)

Let b = 0.8 and c = 3. Note that there is a stationary solution (v(t), w(t)) =
(0, 0) for a = 0.
We compute the solution family, starting at (v(t), w(t)) = (0, 0) for a = 0,
with Auto. The bifurcation diagram is shown in Fig. 1.21. Note that the
solution is unstable for a small and becomes stable after a Hopf bifurcation at
a 0.4228. Figure 1.21 also shows the emanating family of periodic solutions,

36

Eusebius J Doedel
MAX V
2.5

2.0

1.5

1.0

0.5

0.0
0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

Parameter a

Fig. 1.21. Bifurcation diagram of the Fitzhugh-Nagumo equations (1.15).

V

2.0
1.5
1.0
0.5
0.0
-0.5
-1.0
-1.5
-2.0
0.00

0.40

0.20
0.10

0.30

1.00
0.80
0.60
0.90
0.70
0.50
Scaled Time

Fig. 1.22. The periodic solution of (1.15) at a = 0.

which turns back toward a = 0; the periodic solution at a = 0 is shown in

Fig. 1.22.

1 Numerical Analysis of Nonlinear Equations

37

MAX V
4.0

3.5
3.0

2.5
2.0
1.5
1.0
0.5
0.0
5.0

0.0
2.5

15.0
10.0
12.5
7.5
Forcing Amplitude r

20.0
17.5

Fig. 1.23. Continuation of (1.16) from r = 0 to r = 20.

1.4.6 Periodically Forced and Non-Autonomous Systems

In this section we illustrate computing periodic solutions to non-autonomous
systems. The classical example of a non-autonomous system is a periodically
forced system. In Auto periodic orbits of a periodically forced system can be
computed by adding a nonlinear oscillator with the desired periodic forcing
as one of its solution components. An example of such an oscillator is
 
x =
x + y x(x2 + y 2 ),

y = x + y y(x2 + y 2 ),
which has the asymptotically stable solution
x(t) = sin (t),

y(t) = cos (t).

As an example, consider again the FitzHugh-Nagumo equations of Sect. 1.4.5,

where we assume that the rst component of the equations is periodically
forced by r cos t. Coupling the oscillator to the Fitzhugh-Nagumo equations gives:

x =
x + y x(x2 + y 2 ),


y = x + y y(x2 + y 2 ),
(1.16)
v  = c(v 13 v 3 + w ry),


w = (v a + bw)/c,

where we take b = 0.8, c = 3, and = 10. For a = 0 and r = 0 there exists

the solution

38

Eusebius J Doedel
MAX V
4.

3.
2.
1.
0.

-1.
-2.
-3.
-4.
0.00

1.00
0.80
0.60
0.90
0.70
0.50
Scaled Time

0.40

0.20
0.10

0.30

Fig. 1.24. Solutions along the continuation path of (1.16) from r = 0 to r = 20.

x(t) = sin (t),

y(t) = cos (t), v(t) 0,

w(t) 0.

We continue this solution in the forcing amplitude r, from r = 0 to, say,

r = 20. The result is shown in Fig. 1.23, with some of the solutions along this
family plotted versus time in Fig. 1.24.
If the forcing is not periodic, or dicult to model by an autonomous oscillator, then the equations can be rewritten in autonomous form as follows.
The non-autonomous system
 
u (t) = f (t, u(t)),
u(), f () Rn , t [0, 1],
b(u(0), u(1)) = 0, b() Rn ,
can be transformed into

u (t) = f (v(t), u(t)),


v (t) = 1,
b(u(0), u(1)) = 0,

v(0) = 0,

u(), f () Rn ,
v() R,
b() Rn ,

t [0, 1],

which is autonomous, with n + 1 ODEs and n + 1 boundary conditions.

1.5 Computing Connecting Orbits

Orbits that connect xed points of a vector eld are important in many
applications. A basic algorithm, which can be represented in various forms
[25, 6, 19] consists of continuation of solutions to the equations

1 Numerical Analysis of Nonlinear Equations

u (t) = T f (u(t), ),
u(), f (, ) Rn , Rn ,

f (w0 , ) = 0,
f (w1 , ) = 0,

i = 1, . . . , n0 ,
fu (w0 , )v0i = 0i v0i ,
i = 1, . . . , n1 ,
fu (w1 , )v1i = 1i v1i ,

i = 1, . . . , n0 ,
v0i v0i = 1,

v1i = 1,
i = 1, . . . , n1 ,
v1i
 1
fu (
f (
dt = 0,
(f (u, ) f (
u, ))
u, )
u, )
0

&n0
&n0 2
u(0) = w0 + 0 &i=1
c0i v0i ,
c0i = 1,
&i=1
n1
n1
2
c1i v1i ,
c
u(1) = w1 + 1 i=1
i=1 1i = 1.

39

(1.17)
(1.18)
(1.19)
(1.20)
(1.21)
(1.22)

Equation (1.17) is the ODE with independent variable t scaled to [0, 1]. Equation (1.18) denes two xed points w0 and w1 . We assume in (1.19) that
fu (w0 , ) has n0 distinct real positive eigenvalues 0i with eigenvectors v0i ,
and fu (w1 , ) has n1 distinct real negative eigenvalues 1i with eigenvectors
v1i . Equation (1.20) normalizes the eigenvectors. Equation (1.21) gives the
(t), which is a necessary condition for
phase condition, with reference orbit u


D() =

 (t) ||2 dt
|| u (t + ) u

to be minimized over ; here we use u (t) = fu (u, ) u (t) = fu (u, ) f (u, ).
Finally, (1.22) requires u(0) to lie in the tangent manifold U0 at distance 0
from w0 ; similarly, u(1) must lie in S1 at distance 1 from w1 .
Using (1.22) we can eliminate w0 and w1 , to be left with n coupled differential equations subject to
nc = 2n + (n + 1)(n0 + n1 ) + 3
constraints. In addition to u(t) Rn we have scalar variables
Rn ,
0i , c0i R,
1i , c1i R,

0 , 1 R,
v0i Rn ,

i = 1, . . . , n0 ,

v1i R ,

i = 1, . . . , n1 .

The total number of scalar variables equals

nv = n + (n + 2)(n0 + n1 ) + 2.
Formally, we need nv = nc n for a single heteroclinic connection; this gives
n = n (n0 + n1 ) + 1. For a family of connecting orbits, we must use
n (n0 + n1 ) + 2 free parameters. Note that T is large and xed in this
continuation.

40

Eusebius J Doedel

w1

v1

(a)

w0

v0

u (1)

u(0)

w1

v1

(b)

w0

v0

w1

v1

(c)

w0

v0

Fig. 1.25. Geometric interpretation of the equations for computing heteroclinic

orbits in the special case n = 2, n0 = n1 = 1.

First consider the special case of a heteroclinic connection between two

saddle points in R2 , that is, n = 2, n0 = n1 = 1; a graphical illustration of
this case is shown in Fig. 1.25. Then n + 1 = 2, i.e., a branch of heteroclinic
orbits requires two free problem parameters = (1 , 2 ). Consider 2 as xed
here.
For 1 = 1 we assume the existence of the heteroclinic connection in
Fig. 1.25(b). Generically, perturbation of 1 will produce either Fig. 1.25(a)
or Fig. 1.25(c), depending on the sign of the perturbation. If 0 and 1 are
suciently small, then there exists a 1 close to 1 for which (1.17)(1.20)
(and (1.22)) can be satised; here, this is satised for 1 as in Fig. 1.25(a).
Furthermore, the radii 0 and 1 can be chosen such that the period of the
orbit equals a given large value T , and such that the phase condition (1.21)
is satised.
Some more particular cases are:
1. The connection of a saddle to a node in R2 .
Here n = 2, n0 = 1, n1 = 2, so n = 0. A branch of connections requires
one problem parameter;
2. If n = 3, n0 = 3, n1 = 2,
then n = 1, which means that a two-dimensional manifold of connecting
orbits is already possible for xed problem parameters;

1 Numerical Analysis of Nonlinear Equations

41

NORM U
22.

21.
20.

19.
18.
17.
16.
15.
14.
13.
-2.0

-1.5

-1.0

-0.5

0.0

0.5

1.0

1.5

GAMMA

Fig. 1.26. Bifurcation diagram of the singularly-perturbed BVP (1.23).

3. The homoclinic orbit.

In this case w0 = w1 and n0 + n1 = n, so that n = 1. Such orbits can
also be computed as the limit of periodic orbits as the period T .

1.6 Other Applications of BVP Continuation

We end this chapter with two examples where the boundary value continuation
of Auto is applied in special contexts.
Singularly Perturbed BVP
Auto is well suited for computing solutions in systems with multiple timescales.
The numerical sensitivity caused by the dierence in timescales is dealt with
by the orthogonal collocation solution technique with adaptive meshes. The
pseudo-arclength continuation ensures detection of changes along the solution
family. Consider the singularly perturbed system [26]
u (x) = u(x) u (x) (u(x)2 1) + u(x),
with boundary conditions
u(0) =

3
,
2

u(1) = .

The computational formulation is in the form

42

Eusebius J Doedel
U

1.5
1.0
0.5
0.0
-0.5

-1.0
-1.5
-2.0
-2.5
0.00

0.30

0.50
X

1.00

0.80

0.60

0.40

0.20
0.10

0.70

0.90

Fig. 1.27. Some solutions along the solution branch of (1.23).

u1 = u2


u1 u2 (u21 1) + u1 ,
u2 =

(1.23)

with boundary conditions

u1 (0) =

3
,
2

u1 (1) = .

The parameter is a homotopy parameter to locate a starting solution. In

the rst run varies from 0 to 1 and = 1 is xed. In the second run is
decreased by continuation to the desired value. We use = 103 .
Once a starting solution is obtained, we continue the solution for = 103
in the parameter . This third run takes many continuation steps. Figure 1.26
shows the bifurcation diagram with the solution family obtained by continuation in . A selection of the solutions along the branch is shown in Fig. 1.27.
1.6.1 Orbit Continuation in IVP
One can also use continuation to compute solution families of initial value
problems (IVP). Using continuation instead of integration of a large number
of initial conditions has the advantage that the manifold described by the
orbits is well covered, even in problems with very sensitive dependence on
initial conditions. As an example, we consider the Lorenz equations given by

x = (y x),
y  = x y xz,
(1.24)

z = xy z,

1 Numerical Analysis of Nonlinear Equations

43

MAX X
15.

10.

5.

0.

-5.

-10.
0.

5.

10.

15.

20.

25.

30.
RHO

Fig. 1.28. Bifurcation diagram of the Lorenz equations (1.24).

with = 10 and = 8/3.

Let us rst analyze the stationary solutions of (1.24) as a function of .
A bifurcation diagram is shown in Fig. 1.28. The zero solution is unstable for
> 1. Two nonzero (symmetric) stationary solutions bifurcate at = 1. These
nonzero stationary solutions become unstable for > H 24.7. At H there
are Hopf bifurcations, and a family of unstable periodic solutions emanates
from each of the Hopf bifurcation points; only the maximal x-coordinate is
shown in Fig. 1.28, and Fig. 1.29 shows some of these periodic orbits in the
(x, y)-plane. The families of periodic solutions end in homoclinic orbits (innite period) at 13.9.
Now let = 28. For this parameter value the Lorenz equations have a
strange attractor . Let

x
u = y,
z
and write the Lorenz equations as
u (t) = f (u(t)).
The origin 0 is a saddle point, with eigenvalues 1 2.66, 2 22.8, 3
11.82, and corresponding normalized eigenvectors v1 , v2 , and v3 , respectively.
We want to compute the stable manifold of the origin.
We compute an initial orbit u(t), for t from 0 to T (where T < 0), with
u(0) close to 0 in the eigenspace spanned by v1 and v2 , that is,

44

Eusebius J Doedel
Y

15.

10.

5.

0.

-5.

-10.

-15.
-15.

5.

0.

-5.

-10.

15.

10.

Fig. 1.29. Periodic orbits of the Lorenz equations (1.24).

u(0) = 0 +


sin()
cos()
v2 ,
v1
|2 |
|1 |

for, say, = 0.
The IVP can be solved with Auto as follows. Scale time t 
initial orbit satises

u (t) = T f (u(t)),
and
u(0) =

t
T

. Then the

0 t 1,

v1 .
|1 |

The initial orbit has length



|| f (u(s)) || ds.

L=T
0

Thus the initial orbit is a solution of the equation F(X) = 0, where X =

(u(), , T ) (for given L and ) and

u (t) T f (u(t)),

sin()
cos()
v
v

u(0)

F(X) =
2 ,
1
|2 |
|1 |

1
T 0 || f (u(s)) || ds L .

Once the initial orbit has been integrated up to a suciently long arclength L,
we can use pseudo-arclength continuation to nd a family of solution segments

1 Numerical Analysis of Nonlinear Equations

45

(a)

(b)
Fig. 1.30. The stable manifold of the origin in the Lorenz equations (1.24). Panel
(a) shows the family of orbits that represent part of the manifold. Panel (b) shows
another section of the Lorenz manifold.

that forms an approximation of the Lorenz manifold , the stable manifold of

the origin. The set-up for pseudo-arclength continuation is now:

46

Eusebius J Doedel

F(X1 ) = 0,

0 s = 0,
(X1 X0 ) X

0 ||= 1),
(|| X

with X = (u(), , T ) and L and xed. It is important to note here that we

do not just change the initial point (i.e., the value of ). The continuation
stepsize s measures the change in X. An impression of part of the computed
Lorenz manifold is shown in Fig. 1.30. For more detailed results see [13].

1.7 Outlook
We discussed the set-up in Auto for the numerical continuation of families of
solutions to rst-order systems of ordinary dierential equations. Auto uses
Kellers pseudo-arclength continuation [22], which can equally well be applied
to solution families of algebraic problems, e.g., families of stationary solutions.
When applied to families of orbits, each continuation step involves solving a
boundary value problem. Auto uses piecewise polynomial collocation with
Gauss-Legendre collocation points (orthogonal collocation) [7, 3], similar to
Colsys [2] and Coldae [4], with adaptive mesh selection [32].
The basic objective behind the continuation methods of Auto is the ability
to perform a numerical bifurcation analysis. Such computational results give a
deeper understanding of the solution behavior, stability, multiplicity, and bifurcations, and they often provide direct links to the underlying mathematical
theories. We highlighted only the basic set-up in Auto. For multi-parameter
bifurcation analysis the system that implicitly denes the solution branch
is extended to contain bifurcation conditions; see, for example, [11, 12]. By
monitoring the appropriate bifurcation condition Auto detects, say, a Hopf
bifurcation when continuing a family of stationary solutions in one parameter. This bifurcation point can subsequently be continued by extending the
set-up for pseudo-arclength continuation with extra equations (the bifurcation
condition), and freeing a second parameter. For so-called minimally extended
systems see [20, 24].
There is a need for further renement of existing continuation algorithms
and software for bifurcation analysis, and there is a need for their extension to
new classes of problems. Probably the greatest challenges lie in the development of numerical continuation and bifurcation software for partial dierential
equations. There is such a package for scalar nonlinear elliptic PDEs on general domains in R2 [5], which is based on multigrid solution techniques; see
also [27, 28, 29]. Good results have also been obtained with stabilized simple
iteration schemes for computing stationary PDE solutions with mostly stable
modes; see, for example, [33]. There remains a need for general bifurcation
software for systems of elliptic PDEs, subject to general boundary conditions
and integral constraints. For the case of such systems on simple domains in
R2 , the generalization of the collocation method of Sect. 1.2.3 carries some
promise. To become comparable in performance to current ODE bifurcation

1 Numerical Analysis of Nonlinear Equations

47

software it is necessary to use adaptive meshes. In this case the direct solution
of the linear systems arising in Newtons method remains feasible, so that a
high degree of robustness is possible. For developments in this directions, see
[10, 15].
The chapters in this book also provide a wide range of examples of extensions and renements of the continuation algorithms.

Acknowledgments
Although part of the material in these lecture notes is original, the author is
greatly indebted to many collaborators, and above all to H.B. Keller of the
California Institute of Technology, whose published and unpublished work is
strongly present in these notes, and without whose inspiration and support
since 1975 the algorithms and software described here would not exist in
their current form. The author is grateful to World Scientic Publishing for
permission to reproduce previously published material from [12].

References
1. D. G. Aronson, E. J. Doedel, and H. G. Othmer. The dynamics of coupled
current-biased Josephson junctions II. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 1(1): 5166, 1991.
2. U. M. Ascher, J. Christiansen, and R. D. Russell. Collocation software for
boundary value ODEs. ACM Trans. Math. Software, 7:209222, 1981.
3. U. M. Ascher, R. M. M. Mattheij, and R. D. Russell. Numerical solution of
boundary value problems for ordinary dierential equations. Prentice-Hall, 1988;
SIAM, 1995.
4. U. M. Ascher and R. J. Spiteri. Collocation software for boundary value
dierential-algebraic equations. SIAM J. Sci. Comput., 15:938952, 1995.
5. R. E. Bank. PLTMG, A Software Package for Solving Elliptic Partial Dierential Equations. (SIAM, Philadelphia, 1990).
6. W.-J. Beyn. The numerical computation of connecting orbits in dynamical
systems. IMA J. Num. Anal., 9:379405, 1990.
7. C. de Boor and B. Swartz. Collocation at Gaussian points. SIAM J. Numer.
Anal., 10:582606, 1973.
8. E. J. Doedel. Auto: A program for the automatic bifurcation analysis of autonomous systems, Cong. Num. 30, 1981, 265284. (Proc. 10th Manitoba Conf.
on Num. Math. and Comp., Univ. of Manitoba, Winnipeg, Canada.)
9. E. J. Doedel. Numerical Analysis and Control of Bifurcation Problems. Report
UMSI 89/17. University of Minnesota Supercomputer Institute. February 1989.
10. E. J. Doedel. On the construction of discretizations of elliptic partial dierential
equations. J. Dierence Equations and Applications, 3:389416, 1997.
11. E. J. Doedel, H. B. Keller, and J. P. Kernevez. Numerical analysis and control
of bifurcation problems (I): Bifurcation in nite dimensions. Internat. J. Bifur.
Chaos Appl. Sci. Engrg., 1(3):493520, 1991.

48

Eusebius J Doedel

12. E. J. Doedel, H. B. Keller, and J. P. Kernevez. Numerical analysis and control of

bifurcation problems (II): Bifurcation in innite dimensions. Internat. J. Bifur.
Chaos Appl. Sci. Engrg., 1(4):745772, 1991.
13. E. J. Doedel, B. Krauskopf, H. M. Osinga. Global bifurcations of the Lorenz
manifold. Nonlinearity, 19(12):29472972, 2006.
14. E. J. Doedel, V. Romanov, R. C. Paenroth, H. B. Keller. D. J. Dichmann,
J. Gal
an, and A. Vanderbauwhede. Elemental periodic orbits associated with
the libration points in the Circular Restricted 3-Body Problem. Internat. J.
Bifur. Chaos Appl. Sci. Engrg., 17(8), 2007 (in press).
15. E. J. Doedel and H. Shari. Collocation methods for continuation problems in
nonlinear elliptic PDEs. In D. Henry and A. Bergeon, editors, Issue on Continuation Methods in Fluid Mechanics, Notes on Numer. Fluid. Mech., 74:105-118,
(Vieweg, 2000).
16. K. Engelborghs, T. Luzyanina and D. Roose. Numerical bifurcation analysis
of delay dierential equations using DDE-Biftool. ACM Trans. Math. Software, 28(1): 121, 2002. Available via http://www.cs.kuleuven.ac.be/cwis/
research/twr/research/software/delay/ddebiftools.shtml.
17. B. Ermentrout. Simulating, Analyzing, and Animating Dynamical Systems:
A Guide to XppAut for Researchers and Students, volume 14 of Software,
Environments, and Tools (SIAM, Philadelphia, 2002). Available via http:
//www.math.pitt.edu/bard/xpp/xpp.html.
18. T. F. Fairgrieve and A. D. Jepson. O.K. Floquet multipliers. SIAM J. Numer.
Anal., 28(5):14461462, 1991.
19. M. J. Friedman and E. J. Doedel. Numerical computation and continuation of
invariant manifolds connecting xed points. SIAM J. Numer. Anal., 28:789808,
1991.
20. W. Govaerts. Numerical Methods for Bifurcations of Dynamical Equilibria.
(SIAM, Philadelphia, 2000).
21. W. Govaerts, Yu. A. Kuznetsov and A. Dhooge. Numerical continuation of
bifurcations of limit cycles in Matlab. SIAM J. Sci. Computing, 27(1):231252,
2005.
22. H. B. Keller. Numerical solution of bifurcation and nonlinear eigenvalue problems. In P. H. Rabinowitz, editor, Applications of Bifurcation Theory. (Academic
Press, 1977), pages 359384.
23. H. B. Keller. Lectures on Numerical Methods in Bifurcation Problems. Tata
Institute of Fundamental Research. Bombay, 1987.
24. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory. (Springer-Verlag,
New York, 2004).
25. M. Lentini and H. B. Keller. Boundary value problems over semi-innite intervals and their numerical solution. SIAM J. Numer. Anal., 17:557604, 1980.
26. J. Lorenz. Nonlinear boundary value problems with turning points and properties of dierence schemes. In W. Eckhaus and E. M. de Jager, editors, Singular
Perturbation Theory and Applications. (Springer Verlag, New York, 1982).
27. K. Lust. PDEcont: A timestepper-based continuation code for large-scale systems. Available via http://www.dynamicalsystems.org/sw/sw/detail?item=
8.
28. K. Lust, D. Roose, A. Spence, and A. R. Champneys. An adaptive NewtonPicard algorithm with subspace iteration for computing periodic solutions.
SIAM J. Sci. Computing, 19(4):11881209, 1998.

1 Numerical Analysis of Nonlinear Equations

49

29. K. Lust and D. Roose. Computation and bifurcation analysis of periodic solutions of large-scale systems. In E. J. Doedel and L. S. Tuckerman, editors,
Numerical Methods for Bifurcation Problems and Large-Scale Dynamical Systems, IMA Vol. Math. Appl., volume 119. (Springer-Verlag, New York, 2000),
pages 265301.
30. F. J. Mu
noz-Almaraz, E. Freire, J. Gal
an, E. J. Doedel, and A. Vanderbauwhede.
Continuation of periodic orbits in conservative and Hamiltonian systems. Physica D, 181(1-2), 138, 2003.
31. E. L. Reiss. Column buckling An elementary example of bifurcation. In
J. B. Keller and S. Antman, editors, Bifurcation Theory and Nonlinear Eigenvalue Problems, pages 116. (W. A. Benjamin, Publishers, 1969).
32. R. D. Russell and J. Christiansen. Adaptive mesh selection strategies for solving
boundary value problems. SIAM J. Numer. Anal., 15:5980, 1978.
33. G. M. Shro and H. B. Keller. Stabilization of unstable procedures: The recursive projection method. SIAM J. Numer. Anal., 30(4):10991120, 1993.
34. X.-J. Wang and E. J. Doedel. Auto94P: An experimental parallel version of
Auto. Technical report, Center for Research on Parallel Computing, California
Institute of Technology, Pasadena CA 91125. CRPC-95-3, 1995.

2
Interactive Continuation Tools
Willy Govaerts1 and Yuri A Kuznetsov2
1
2

Department of Applied Mathematics and Computer Science, Ghent University,

Belgium
Department of Mathematics, Utrecht University, The Netherlands

Systematic bifurcation analysis requires the repeated continuation of dierent

phase objects in free parameters, the detection and analysis of their bifurcations, and branch switching. Such computations produce a lot of numerical
data that must be analyzed and, nally, presented in graphical form. Thus,
continuation programs should not only be ecient numerically but should allow for interactive management and have a user-friendly graphics interface.
The development of such programs is progressing rapidly. Here we make an
attempt to survey existing interactive continuation and bifurcation tools and
outline their history and perspectives. This is followed by the presentation
of a framework that organizes the dierent types of objects and bifurcating
branches. We give a brief overview of how such a framework is implemented in
the recent software environment Matcont. In the nal two sections we give
a few examples that illustrate the use of Matcont and indicate directions of
future developments.

2.1 Overview of Existing Software

During the last decades, considerable eorts have been made to develop
general-purpose software tools for bifurcation analysis. One may distinguish
at least three types (generations) of such software:
1. Noninteractive packages and codes,
2. Interactive programs,
3. Software environments.
Since the development of numerical algorithms advances with each generation,
these tools also dier in supported computations. We give here an overview
for each of the above types, with emphasis on the best known packages.

52

Willy Govaerts and Yuri A Kuznetsov

2.1.1 Noninteractive Packages and Codes

Noninteractive packages and codes rst appeared in the beginning of the 1980s
and were written in FORTRAN. They allowed one to continue equilibria and
limit cycles of ODEs, as well as detect and subsequently continue their basic bifurcations: limit point (saddle-node) bifurcation, Hopf bifurcation, and
period-doubling bifurcation. The most widely used packages of this generation are Auto86 [12] and Linlbf [24]. Although these two packages supported a similar level of bifurcation analysis, they employed very dierent
numerical algorithms. For example, Linlbf bases its test functions to locate
Hopf and Neimark-Sacker bifurcation points on Hurwitz determinants, while
Auto86 bases the detection and location of all local bifurcations on monitoring the eigenvalues (multipliers). Moreover, the continuation of equilibrium
(xed point) bifurcations in Auto86 is done using extended augmented systems that include eigenvectors, while in Linlbf minimally augmented systems
are used. The most essential dierence, however, lies in the continuation of
limit cycles and their codimension-one bifurcations. In Linlbf these tasks
are performed via numerically constructed and dierentiated Poincare maps,
while Auto86 employs the discretization of the corresponding boundary value
problems using piecewise-polynomial approximations and orthogonal collocation. The latter proved its superiority for more complex multi-dimensional
ODEs.
Bifurcation theory relies on center manifold reduction, followed by transformation to a normal form. The computation of normal-form coecients is an
important aspect of bifurcation software; for background we refer to [32]. The
explicit computation of normal-form coecients is not supported by Auto86,
and Linlbf only computes normal-form coecients for local codimension-one
bifurcations. There were a few codes available in the 1980s for simple numerical normal-form and branching analysis, e.g. Stuff [5] and Bifor2 [22], but
switching to the computation of dierent bifurcating objects required manual
restarting of the code with new initial data.
Several other noninteractive packages for the continuation of simplest bifurcations in ODEs also appeared around this time, e.g. Bifpack by Seydel [38, 39], and development continues to date, in particular for large-scale
dynamical systems with packages such as Loca3 by Salinger et al. [35, 36].
2.1.2 Interactive Programs
Interactive programs for the bifurcation analysis of ODEs appeared at the end
of the 1980s, when workstations and IBM-PC compatible computers became
widely available at universities and general research institutes. All programs
of this generation have a simple Graphical User Interface (GUI) with buttons,

loca is available via http://www.cs.sandia.gov/loca/.

2 Interactive Continuation Tools

53

Fig. 2.1. Screen snapshot of Auto94/97.

windows, and pull-down menus, and support the on-line input of the righthand side of ODEs (through compilation with a FORTRAN or C compiler).
Computed curves could now be plotted directly in a graphics window.
The continuation code Auto86 mentioned above already comes with a
simple interactive graphics program called plaut that allows for graphical
presentation of computed data. There are versions of plaut for most of the
widespread workstations, as well as a Matlab version mplaut4 , written by
De Feo. There have been several attempts to improve the user interface of
Auto86 and later versions of Auto. The rst interactive version of Auto86
was developed at Princeton University by Taylor and Kevrekidis [41] for SGI
workstations. Another example is XppAut5 by Ermentrout [15] for workstations and PCs, which developed from combining the MS-DOS program
Phaseplane with Auto. XppAut is also capable of simple phase-plane analysis, including the computation of one-dimensional global invariant manifolds
of equilibria, as is Scigma [40]. Note that XppAut is still widely used and
includes tools for the analysis of delay equations, functional equations, and
stochastic equations. Doedel, Wang, and Fairgrieve also designed the interac4
5

mplaut is available via http://www.math.uu.nl/people/kuznet/cm.

XppAut is available via http://www.math.pitt.edu/bard/xpp/xpp.html.

54

Willy Govaerts and Yuri A Kuznetsov

Fig. 2.2. Screen snapshot of LocBif 2.0.

tive version Auto94 for UNIX workstations with X-Windows; see Fig. 2.1.
This version has extended numerical capabilities, including the continuation
of all codimension-one bifurcations of limit cycles and xed points. The software was upgraded in 1997 to support the continuation of homoclinic orbits
using HomCont [9, 13]. This version is called Auto976 and there is also the
C-version Auto2000 which has a new interactive graphics browser.
The major diculty in using all versions of Auto is the analysis of detected
bifurcation points and switching at these points to the continuation of other
bifurcation curves, which requires browsing of several output les and a good
understanding of their formats. However, due to the exceptional numerical
eciency of Auto, attempts to provide a better GUI for it continue to date;
see, for example, Oscill87 .
The rst user-friendly interactive bifurcation program for bifurcation analysis was LocBif8 developed for PCs under MS-DOS by Khibnik, Kuznetsov,
Levitin, and Nikolaev [25]; a screen snapshot of version 2.0 is shown in Fig. 2.2.
The numerical part of the program is based on the non-interactive code
6
7
8

Auto is available via http://cmvl.cs.concordia.ca/.

Oscill8 is available via http://oscill8.sourceforge.net/doc/.
LocBif is available via http://www.math.uu.nl/people/kuznet/LOCBIF;
LocBif works as a DOS-application under MS-Windows, but is no longer
supported.

2 Interactive Continuation Tools

55

Linlbf and allows for continuation of equilibrium, xed-point, and limit cycle
bifurcations up to codimension three. The program allows for easy switching
between the computation of various curves at detected bifurcation points. The
user can manipulate individual computed bifurcation curves, which are stored
separately in an archive. Version 1.0 of LocBif uses an external FORTRAN
compiler and has a very simple keyboard-based interface, but version 2.0 can
be driven by a mouse and has a special built-in compiler for the right-hand
side. Neither version, however, has special tools to output the computed curves
in a graphic format.
The program Candys/QA9 by Feudel and Jansen [16] also belongs to this
generation but is less widely used. All programs mentioned so far have closed
architecture.
2.1.3 Software Environments
The rst software environments for bifurcation analysis were DsTool10 and
Content11 , developed in the 1990s. Both programs support the simulation
of ODEs. The user can dene/modify a dynamical model, perform a rather
complete analysis, and export the results in a graphical form, all without
leaving the program. Though hard, it is possible to extend them. The programs have an elaborate GUI and provide o- or on-line help and extensive
documentation for users and developers.
DsTool [3] runs under UNIX or Linux. It performs simple phase-plane
analysis and includes the computation of equilibria and associated onedimensional stable and unstable manifolds, along with the continuation of
equilibria and their codimension-one bifurcations, which is done by using parts
of the Linlbf code.
The interactive software Content was developed by Kuznetsov and Levitin with contributions by De Feo, Sijnave, Govaerts, Doedel, and Skovoroda;
a screen snapshot of Content 1.5 is shown in Fig. 2.3. The software runs
on most popular workstations under UNIX and on PCs under Linux or MSWindows and supports the continuation of equilibria and their bifurcations of
codimension up to two. Content uses minimal and extended augmented systems, as described in [19, 20], as well as the continuation of limit cycles using
Auto-like algorithms. Moreover, Content supports the normal-form computations for many equilibrium bifurcations, taking advantage of internally
generated symbolic derivatives of order up to three, and allows for branch
switching by using algebraic branching equations. The software provides extensive storage, export and import facilities for computed curves and diagrams, including their numerical and PostScript formats. Switching between
9
10
11

Candys/QA is available via http://www.agnld.uni-potsdam.de/wolfgang/

candys.html.
DsTool is available via http://www.cam.cornell.edu/gucken/dstool.
Contentis available via http://www.math.uu.nl/people/kuznet/CONTENT/.

56

Willy Govaerts and Yuri A Kuznetsov

Fig. 2.3. Screen snapshot of Content 1.5.

various bifurcating objects at special points is very easy and exible in Content. The latest software project Matcont12 is a Matlab interactive toolbox
for the continuation and bifurcation analysis of ODEs [11] that is based on
experience in developing and using Content.
To conclude this section, we mention that numerical bifurcation analysis of smooth iterated maps is also supported by existing software. Location,
analysis, and continuation of xed-point bifurcations are very similar to those
for equilibria of ODEs, and are supported, for example, by Auto, LocBif,
and Content [18]. Other problems, particularly the analysis of global bifurcations, require special algorithms. For example, one needs special algorithms
for the computation of the one-dimensional stable and unstable invariant manifolds of xed points of maps; implementations for one-dimensional manifolds
already exist in DsTool [29, 30] and Dynamics [34, 43, 44]. Such algorithms
are necessary for the continuation of homoclinic orbits and their tangencies [4],
which is also implemented as an Auto-driver by Yagasaki [42]. The computation of two- or higher-dimensional invariant manifolds, for example, global
stable and unstable invariant manifolds and invariant tori, and their bifurcations both for ODEs and maps is much more dicult and only a few algorithms
12

Matcont is available via http://www.matcont.ugent.be/.

2 Interactive Continuation Tools

57

are available; see Chap. 4, and [28, 31] for global manifolds and [6, 14, 37] for
invariant tori.

2.2 Bifurcation Objects and Their Relations

The application of bifurcation theory to multi-parameter dynamical systems
requires a clear continuation strategy that should provide rules on how to
increase and decrease the number of control parameters while studying objects
of dierent codimension. In addition, this strategy should suggest how to
switch between dierent bifurcations of the same codimension, keeping the
number of control parameters constant. As was rst mentioned in [25] from a
theoretical point of view, this strategy must be based on graphs of adjacency [1]
that describe relationships between bifurcations. Below, we present two graphs
of adjacency, describing detection relationships and branching relationships.
The next section identies the bifurcation objects for ODEs; the equivalent
for maps is done in Sect. 2.2.2.
2.2.1 Bifurcation Objects in ODEs
Tables 2.1 and 2.2 list the codimension-zero, -one, and -two objects that can
be found in generic continuous dynamical systems, along with associated labels based on standard terminology [32]. Table 2.1 lists all objects related
to equilibria and limit cycles, while Table 2.2 focusses on objects related to
homoclinic orbits of equilibria. The relationships between these objects are
complicated.
The detection relationships between the objects in Tables 2.1 and 2.2 are
presented in Figs. 2.4 and 2.5, respectively. For example, the arrows from O
to EP and LC mean that it is generically possible for a computed orbit (O)
to converge to a (stable) equilibrium (EP) or to a (stable) limit cycle (LC).
An arrow from an object A dierent from O to an object B means that the
continuation of a one-parameter family of objects of type A can generically
lead to the detection of an object of type B, either because object B is a
special case of object A or because it is a limiting case when the parameter
tends to a special value. An example of the rst situation is a Hopf bifurcation
point (H) on a curve of equilibria (EP); an example of the second situation
is a homoclinic orbit of a hyperbolic saddle (HHS), because it is the limit
of a branch of periodic orbits (LC) when the period tends to innity. We do
not distinguish between the two situations, because the dierence depends
somewhat on the denition of a family of objects and it may depend on the
implementation of the dening system that is used in the computation of the
branch (e.g. an H point on a family of LC objects). Note that any computation
normally starts with a point P as an initial condition to generate an orbit
O; this rst step does not feature in Fig. 2.4. We are interested in generic
detection relationships, which is why the arrows always connect objects from

58

Willy Govaerts and Yuri A Kuznetsov

Table 2.1. Objects and associated labels related to equilibria and limit cycles of
ODEs

Type of object

Label

Point
Orbit
Equilibrium
Limit cycle
Limit Point (fold) bifurcation
Hopf bifurcation
Limit Point bifurcation of cycles
Neimark-Sacker (torus) bifurcation
Period Doubling (ip) bifurcation
Branch Point
Cusp bifurcation
Bogdanov-Takens bifurcation
Zero-Hopf bifurcation
Double Hopf bifurcation
Generalized Hopf (Bautin) bifurcation
Branch Point of Cycles
Cusp bifurcation of Cycles
Generalized Period Doubling
Chenciner (generalized Neimark-Sacker) bifurcation
1:1 Resonance
1:2 Resonance
1:3 Resonance
1:4 Resonance
FoldNeimark-Sacker bifurcation
FlipNeimark-Sacker bifurcation
Fold-ip
Double Neimark-Sacker

P
O
EP
LC
LP
H
LPC
NS
PD
BP
CP
BT
ZH
HH
GH
BPC
CPC
GPD
CH
R1
R2
R3
R4
LPNS
PDNS
LPPD
NSNS

one codimension level down to objects on the next codimension level. The only
two exceptions are the arrows from EP to BP and from LC to BPC, which jump
over two codimension levels. In fact, these situations are non-generic, but they
are so often found in systems with equivariance or invariant subspaces that
most software packages support their detection.
The branching relationships between the objects in Tables 2.1 and 2.2 can
be obtained directly from Figs. 2.4 and 2.5, respectively. In general, if there
is an arrow in Fig. 2.4 or 2.5 from an object A dierent from O to an object
B then for each object of type B there is a unique one-parameter family of
objects of type A that branches o B, provided a total of k + 1 free variables is
chosen, where k is the codimension level of A. There are only four exceptions:
1. The arrows from EP to BP and from LC to BPC: there are generically two
codimension-zero curves emanating from the codimension-two points.

2 Interactive Continuation Tools

59

Table 2.2. Objects and associated labels related to homoclinic orbits of equilibria
of ODEs

Type of object

Label

Limit cycle
Homoclinic orbit of a Hyperbolic Saddle
Homoclinic orbit of a Saddle-Node
Neutral saddle
Neutral saddle-focus
Neutral Bi-Focus
Shilnikov-Hopf
Double Real Stable leading eigenvalue
Double Real Unstable leading eigenvalue
Neutrally-Divergent saddle-focus (Stable)
Neutrally-Divergent saddle-focus (Unstable)
Three Leading eigenvalues (Stable)
Three Leading eigenvalues (Unstable)
Orbit-Flip with respect to the Stable manifold
Orbit-Flip with respect to the Unstable manifold
Inclination-Flip with respect to the Stable manifold
Inclination-Flip with respect to the Unstable manifold
Non-Central Homoclinic to saddle-node

LC
HHS
HSN
NSS
NSF
NFF
SH
DRS
DRU
NDS
NDU
TLS
TLU
OFS
OFU
IFS
IFU
NCH

Codimension
O

EP

LC

LP

BP

CP

BT

ZH

HH

GH

LPC

CPC

BPC

R1

NS

R3

R4

CH

PD

LPNS

PDNS

R2

NSNS

LPPD

GPD

Fig. 2.4. Detection relationships between bifurcations of equilibria and limit cycles of ODEs; the branching relationships are found by following the arrows in the
opposite directions, with four exceptions as discussed in the text.

2. The arrows from H to HH and from NS to NSNS: there are generically two
codimension-one curves emanating from the codimension-two points.
3. The arrow from NS to ZH: the existence of the NS curve rooted in the ZH
point is subject to an inequality constraint.

60

Willy Govaerts and Yuri A Kuznetsov

Codimension

LC

HHS

NSS

NSF

NFF

DR*

ND*

HSN

TL*

SH

OF*

IF*

NCH

Fig. 2.5. Detection relationships between homoclinic bifurcations of ODEs; here *

stands for S or U; the branching relationships are found by following the arrows in
the opposite directions.
Codimension
O

FP

LP

CP

GPD

CH

NS

R1

R2

PD

R3

R4

LPPD

LPNS

PDNS

NSNS

BP

Fig. 2.6. Detection relationships between dynamical objects for maps.

4. The arrow from NS to HH: there are generically two NS curves emanating
from an HH point.
We note that generically a curve of HHS orbits emanates from a Bogdanov
Takens point (BT), as well as two such curves from a Zero-Hopf bifurcation
point (ZH). These are not indicated in Figs. 2.4 and 2.5.
2.2.2 Bifurcation Objects for Cycles of Maps
In this section we present the equivalent objects and relationships for maps.
Table 2.3 lists the codimension-zero, -one, and -two objects that can be found

2 Interactive Continuation Tools

61

Table 2.3. Objects and associated labels related to equilibria and cycles of maps

Type of object

Label

Point
Orbit
Fixed Point
Limit Point of cycle bifurcation
Period Doubling Point of cycles
Neimark-Sacker bifurcation
Branch Point
Cusp bifurcation
Generalized Period Doubling
Chenciner (generalized Neimark-Sacker) bifurcation
1:1 Resonance
1:2 Resonance
1:3 Resonance
1:4 Resonance
FoldNeimark-Sacker bifurcation
FlipNeimark-Sacker bifurcation
Fold-ip
Double Neimark-Sacker

P
O
FP
LP
PD
NS
BP
CP
GPD
CH
R1
R2
R3
R4
LPNS
PDNS
LPPD
NSNS

in generic maps, together with the associated labels [32]. The detection relationships between them are presented in Fig. 2.6. The precise meaning of the
arrows is simpler than in the case of ODEs: if we exclude O then an arrow
from an object A to an object B indicates that object B can generically be
found as a regular point on a branch of objects of type A. The only exception
is the arrow from FP to BP which is again not generic but found in many
examples that exhibit a form of equivariance or have invariant subspaces.
The branching diagram for maps, on the other hand, is far more complicated than for ODEs; this is largely due to the fact that one needs to consider
dierent iterates of the underlying maps, which causes an additional complication. For reasons of clarity we, therefore, present two branching diagrams;
see Figs. 2.7 and 2.8. As before, the arrows indicate the type of object to which
one can generically switch from a given codimension-one or -two bifurcation
point. If the arrow is dashed then this switching is subject to additional constraints. Furthermore, several switches to branches of lower codimension lead
to curves with double, triple or quadruple iteration number, which is indicated
by the symbols 2, 3, and 4, respectively.

2.3 The Implementation in Matcont

The framework described in the previous section has been implemented in
the recent Matlab-based software environment Matcont [11]. At present

62

Willy Govaerts and Yuri A Kuznetsov

Codimension
0

FP

LP

NS

PD

CP

GPD

CH

R1

R2

BP

Fig. 2.7. Partial branching relationships for maps; see also Fig. 2.8. Dashed lines
indicate switching subject to constraints and 2 indicates switching to a curve with
twice the period.
Codimension
1

LP

R3

NS

PD

R4

LPPD

LPNS

PDNS

NSNS

Fig. 2.8. Partial branching relationships for maps; see also Fig. 2.7. Dashed lines
indicate switching subject to constraints and 2 (3, 4) indicates switching to a
curve with twice (three times, four times) the period.

three related Matlab packages are distributed, namely a command-line version Cl Matcont and a GUI version Matcont for ODEs, and a commandline version Cl MatcontM for Maps13 . As in Auto and Content, limit
cycles are computed by an approach based on the discretization via piecewisepolynomial approximation with orthogonal collocation of the corresponding
boundary value problem. However, Matcont uses sparse Matlab solvers
13

Cl Matcont, Matcont, and Cl MatcontM are all available via http://www.

matcont.ugent.be/.

2 Interactive Continuation Tools

63

instead of the original Auto algorithm (a special block elimination; see

Chap. 1). The same approach is applied for homoclinic orbits, in combination
with the continuation of invariant subspaces for the equilibrium end point of
the homoclinic orbit; for details on this method we refer to [10] and for its
implementation in Matcont to [17].
Nearly all functionalities described in Sect. 2.2 are supported. Remaining
functionalities (now under construction) are:

Branch switching at HH (equilibria), NSNS (limit cycles) and NSNS (cycle

of maps) to the secondary branch of type H, NS (limit cycle) or NS (cycle
of maps), respectively.
Branch switching from ZH (equilibria) to NS (limit cycles).
Branch switching from HH (equilibria) to NS (limit cycles).

A computationally more dicult problem is branch switching from ZH to

HHS [8]. It is also planned to have a GUI version for Cl MatcontM and
to introduce automatic dierentiation routines for the computation of the
normal-form coecients, which are now computed either via numerical directional derivatives or using a user-supplied code. Preliminary evidence indicates that nite dierence approximations are not reliable for these computations. Also, for high-order iterates of maps the normal-form computations
are much faster when using automatic dierentiation compared to symbolically generated derivatives. The computation of normal-form coecients for
codimension-two bifurcations of limit cycles is not yet supported in Matcont
and is another topic for further development.
In some cases normal-form coecients are not very informative. For example, for a limit point of equilibria (LP) the quadratic normal-form coecient a
is dened up to a nonzero multiple and the LP point is nondegenerate if a = 0
and degenerate (CP) if a = 0. However, because of truncation and round-o
errors the value computed for a will always be nonzero. Therefore, the value
of a reported at LP points is not very useful. However, provided its continuity
along the LP-branch is ensured, this value is important for the detection of
the CP-points.
In other cases the normal-form coecients are very useful for the user,
because their values determine the number and type of branches of new objects
that emanate from the bifurcation points and whether these objects are stable
or not. In the following two sections we discuss the cases that are of particular
interest.

2.3.1 Normal-Form Coecients for Bifurcations of ODEs as Given

in Matcont
The implementation in Matcont provides normal-form coecients for all
codimension-one and -two bifurcations of equilibria, and periodic normal-form
coecients for all codimension-one bifurcations of limit cycles; see [32] and
[33] for further details and notation used. We discuss here four cases:

64

Willy Govaerts and Yuri A Kuznetsov

1. The Hopf bifurcation H of an equilibrium, where the equilibrium has a

pair of purely imaginary eigenvalues, is determined by the rst Lyapunov
coecient l1 , which is the real part of the third-order coecient in the
complex normal form. If l1 < 0 then the Hopf bifurcation is supercritical,
i.e., unstable xed points coexist with stable periodic orbits on one side
of the bifurcation point in the center manifold. If l1 > 0 then the Hopf
bifurcation is subcritical, i.e., stable xed points coexist with unstable
periodic orbits on one side of the bifurcation point in the center manifold.
2. The Zero-Hopf bifurcation ZH, also called saddle-node Hopf, fold-Hopf
or zero-pair bifurcation [32], is a codimension-two bifurcation where an
equilibrium has one zero eigenvalue together with a pair of purely imaginary eigenvalues. The normal form involves quadratic coecients denoted
s and ; see [32, Lemma 8.11]. An NS curve emanates from the ZH point
only if s < 0. The implementation in Matcont also computes a relatively technical term E(0). If E(0) < 0 then time has to be reversed in the
unfolding analysis in [32], i.e., stable becomes unstable, and vice versa.
The above two cases deal with bifurcation of equilibria. For limit cycles we
have:
3. The period-doubling or ip bifurcation PD, where the limit cycle has one
Floquet multiplier at 1, involves the coecient c in the periodic normal
form that determines the bifurcation. If c < 0 then the ip bifurcation
is supercritical, i.e., unstable periodic orbits coexist with stable doubleperiod orbits on one side of the bifurcation point in the center manifold.
If c > 0 then the ip bifurcation is subcritical, i.e. stable periodic orbits
coexist with unstable double-period orbits on one side of the bifurcation
point in the center manifold.
4. At a Neimark-Sacker bifurcation NS the limit cycle has a pair of complex conjugate Floquet multipliers on the unit circle. The bifurcation is
determined by the cubic coecient Re(d) of the periodic normal form. If
Re(d) < 0 then the NS bifurcation is supercritical, i.e., unstable limit cycles coexist with stable invariant tori on one side of the bifurcation point
in the center manifold. If Re(d) > 0 then the NS bifurcation is subcritical,
i.e., stable limit cycles coexist with unstable invariant tori on one side of
the bifurcation point in the center manifold.
2.3.2 Normal-Form Coecients for Bifurcations of Maps as Given
in MatcontM
The implementation in MatcontM provides normal-form coecients for all
codimension-one and -two bifurcations of xed points; see [32] for details. The
codimension-one cases are very similar to the corresponding bifurcations of
limit cycles listed in the previous section. The user should, in particular, be
aware of:

2 Interactive Continuation Tools

65

1. The period-doubling or ip bifurcation PD, where the xed point has an

eigenvalue 1. The sign of the cubic normal-form coecient b1 determines
whether the bifurcation is supercritical (b> 0) or subcritical (b1 < 0) as
before.
2. The NeimarkSacker bifurcation NS, where the xed point has a pair
of complex conjugate eigenvalues on the unit circle. As before, the NS
bifurcation is supercritical (subcritical) if the cubic normal-form coecient
c1 = Re(d1 ) is negative (positive) and no strong resonances (1:1, 1:2, 1:3,
1:4) are present.
For codimension-two bifurcation points the user should pay particular attention to:
1. At a 1:2 Resonance point R2 the xed point has a pair of complex conjugate eigenvalues on the unit circle that are both at 1. The normal form
contains two cubic coecients C1 and D1 that determine this bifurcation.
If C1 < 0, then an NS curve of double-period cycles emanates from the
R2 point. The MatcontM output is [c, d] = [4C1 , 2D1 6C1 ] .
2. At a 1:4 Resonance point R4 the xed point has eigenvalues i. This
bifurcation is determined by a complex normal-form coecient A0 = a+ib.
If a2 +b2 1 > 0 then two half lines l1,2 of limit points
of quadruple-period
cycles emanate from the R4 point. If |b| > (1 + a2 )/ 1 a2 then there is
a curve of quadruple-period cycles that contains an NS bifurcation point.
3. At a fold-ip bifurcation LPPD the xed point has eigenvalues 1 and 1.
a
and be
MatcontM computes normal-form coecients 2e
2 ; see [32] for
details. If be > 0 then an NS curve of double period emanates from the
LPPD point. In this case, MatcontM also reports an approximation of
the corresponding rst Lyapunov coecient. The NS points of the second
iterate are stable in the center manifold if this coecient is negative; they
are unstable if it is positive.

2.4 Examples and Applications

We end this chapter with two examples that illustrate how Matcont is used
in practice. In the next section we describe the process of a continuation
strategy for a vector eld. We use the model of a Van der PolDung oscillator
that is also used in Chap. 4. Section 2.4.2 illustrates the use of MatcontM
for a discrete model of a production strategy involving two competing rms.
2.4.1 The Koper Model
In [27] Koper introduced the following model to describe a three-dimensional
Van der PolDung oscillator:

66

Willy Govaerts and Yuri A Kuznetsov

Fig. 2.9. Screen snapshot of Matcont with the computed equilibrium and LP
curves of the Van der PolDung oscillator (2.1).

x = (ky x3 + 3x )/1 ,
y = x 2y + z,

z = 2 (y z).

(2.1)

As in [27] we use 1 = 0.1 and 2 = 1. We note that if (x(t), y(t), z(t)) is

a solution of (2.1) for a particular value of , then (x(t), y(t), z(t)) is
a solution for . Therefore, bifurcation diagrams in which is represented
usually have some symmetry.
We begin the analysis of (2.1) by determining the equilibria. Note that
an equilibrium solution (x0 , y0 , z0 ) satises x0 = y0 = z0 , which must be a
solution of
(2.2)
kx x3 + 3x = 0.
In particular, for = 0 and
k = 0.15 the equilibria are (0, 0, 0) and (x0 , y0 , z0 )
with x0 = y0 = z0 = 3.15 1.77482393492988. By selecting one of
these latter two points in Matcont we compute by numerical continuation the solution of (2.2) as a function of ; the cubic solution curve is visualized in the two-dimensional graphics window of the screen snapshot of
Matcont in Fig. 2.9. On the equilibrium curve Matcont detects two limit
points LP at = 2.151860 and reports for both points the critical normalform coecient a = 4.437060. We select one of the LP points, set both k
and as free parameters, and compute a curve of LP points that connects

2 Interactive Continuation Tools

67

0.99

0.98

0.97

0.96

0.95

0.94

0.93
0.93

0.94

0.95

0.96

0.97

0.98
x

0.99

1.01

1.02

1.03

Fig. 2.10. Limit cycles of (2.1) started from a Hopf point and converging to a
homoclinic orbit.

the two LP points for k = 0.15; this curve is also shown in Fig. 2.9. During the computation of the LP curve Matcont detects two BT points at
(k, ) = (0.050000, 1.950209) and (k, ) = (0.050000, 1.950209), with
normal-form coecients (a, b) = (6.870226e + 000, 3.572517e + 001) and
(a, b) = (6.870226e + 000, 3.572517e + 001), respectively; two Zero-Neutral
Saddle points (formally ZH) at (k, ) = (0.300000, 1.707630) and (k, ) =
(0.300000, 1.707630); and a cusp point CP at (k, ) = (3.000000, 0.000000)
with normal-form coecient c = 5.649718e 002. These codimension-two
points are also shown in Fig. 2.9.
Starting from the BT point at = 1.950209 we can compute a Hopf curve
in the two free parameters k and . We stop, fairly arbitrarily, at the Hopf
point with x0 = y0 = z0 = 0.98460576, k = 0.25185549, and = 1.7513143.
Starting from this point we keep k xed and compute a curve of limit cycles
(LC) as a function of ; see Fig. 2.10. It is visually clear that the limit cycles
converge to a homoclinic orbit; this can also be inferred from the fact that
the parameters change very slowly at the end of the continuation, while the
period increases rapidly.
When computing limit cycles, Matcont allows for the computation and
visualization of their phase response curves (PRC) [21] as well as the time
derivatives of these phase response curves (dPRC). The study of such curves
is an important subject in the theory of weakly connected neural networks [23].
In particular, it is well known that they take very specic shapes in the neighborhoods of bifurcations of limit cycles [7]. We demonstrate this by presenting

68

Willy Govaerts and Yuri A Kuznetsov

(a)

1.5

(b)

PRC

x 10

Derivative of response

Response

dPRC

x 10

1.5

0.5

0.5

0.5

0.5

1
0

0.8

0.6

0.4

0.2

0.6

0.4

0.2

Phase

0.8

Phase

Fig. 2.11. The phase response curve PRC (a) and its derivative dPRC (b) of a limit
cycle of (2.1) close to a homoclinic orbit.

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0.6
0.5

0.6

0.7

0.8

0.9

1.1

1.2

1.3

1.4

1.5

Fig. 2.12. Continuation of an orbit of (2.1) that is homoclinic to a hyperbolic

saddle, starting from a limit cycle with large period.

the curves PRC and dPRC in Fig. 2.11 for the limit cycle of the above continuation at = 1.7510571. This limit cycle has period 46.799011, that is, it
is close to a homoclinic orbit.
It is possible in Matcont to start a continuation of homoclinic orbits in
two parameters from a limit cycle close to a homoclinic orbit; see Fig. 2.5.
An example is presented in Fig. 2.12, where we start from the last limit cycle
computed in the previous run, declare it to be of type HHS, and choose k and
as the two free parameters. The time length of the discretized part of the

2 Interactive Continuation Tools

69

0.95

0.9

0.85

0.8

0.75

0.75

0.8

0.85

0.9

0.95
x

1.05

1.1

1.15

1.2

Fig. 2.13. Continuation of an orbit of (2.1) that is homoclinic to a hyperbolic

saddle, starting from a Bogdanov-Takens point.

orbit is kept xed at the period of the original limit cycle while the distances
from the end points in the stable and unstable directions are free.
It is also possible to start the continuation of a curve of homoclinic orbits
from a Bogdanov-Takens point BT; cf. Sect. 2.2.1. An example of such a
continuation is presented in Fig. 2.13. Here we started from the BT point at
= 1.950209 that is shown in Fig. 2.9. In this case the distance from the end
point in the unstable direction was xed.
2.4.2 The Duopoly Model
We demonstrate the use of MatcontM for an example of two competing
rms that decide on annual production quantities in a duopoly environment.
The two rms are homogeneous with regard to forming their expectation and
the action eect on each other. The model that we use is the two-dimensional
map

x1 (t + 1) = (1 )x1 (t) + x2 (t)(1 x2 (t)),
(2.3)
F :
x2 (t + 1) = (1 )x2 (t) + x1 (t)(1 x1 (t)),
described in [2, 26]. The duopoly model assumes that at each discrete time t
the two rms produce the quantities x1 (t) and x2 (t), respectively, and decide
their productions x1 (t + 1) and x2 (t + 1) for the next period. The parameter
> 0 measures the intensity of the eect that one rms actions has on the
other rm. The parameter , which is typically in [0, 1], has an averaging
eect.

70

Willy Govaerts and Yuri A Kuznetsov

1.5
1.4
1.3
1.2

rho

1.1
R4

LPPD

LPPD

0.9
0.8
0.7
0.6
0.5
3.4

3.45

3.55

3.5

3.6

3.65

mu

Fig. 2.14. An R4 point on an NS curve of (2.3) with emanating branches of fold

curves of period-four cycles.

We start with parameter values = 3.5 and = 0.1. It is checked easily

that F has a xed point
!


+ 1 + ( + 1)( 3) + 1 ( + 1)( 3)
.
,
(x1 , x2 ) =
2
2

We now perform a continuation of xed points of F with free parameter

and nd an NS point at (x1 , x2 ) = (0.857143, 0.428571) for = 0.888889.
The normal-form coecient is 6.273434e+001. Since it is negative, the NS
bifurcation is supercritical.
Starting from this NS point we can now compute a curve of NS points
in the two free parameters and . On this curve we nd a 1:4 resonance point R4 at (x1 , x2 ) = (0.849938, 0.439960) for = 1.000000 and
= 3.449490. It is worthwhile to note that this R4 point lies precisely
on the boundary of the region where 1, i.e., the region that is relevant from the applications point of view. The normal-form coecient is
A0 = (3.000000e+000 9.231411e017 i). Since |A0 | > 1, two cycles of period four are born at the R4 point. Furthermore, near the R4 point their region
of existence is bounded by two fold curves of period-four cycles that emanate
from the R4 point. We can start the continuation of these curves from the
codimension-two point R4 in MatcontM. Interestingly, on each of these two
curves an LPPD point (of the fourth iterate) is found. A picture of this situa-

2 Interactive Continuation Tools

71

1.1
LPPD

rho

1.05

R4

0.95
LPPD

0.9

3.4

3.45

3.5
mu

3.55

3.6

3.65

Fig. 2.15. The diamond-shaped region bounded by period-four LP and PD curves

near the R4 point on the NS curve of (2.3) contains stable period-four cycles.

tion is presented in Fig. 2.14. The lower LPPD point is detected at (x1 , x2 ) =
(0.841586, 0.354516) for = 0.935299 and = 3.566686; its normal-form coefcients are a/(2e) = 2.574002e+000 and be/2 = 5.829597e+001. The upper
LPPD point is detected at (x1 , x2 ) = (0.836428, 0.522216) for = 1.071080
and = 3.486079, and has normal-form coecients a/(2e) = 1.733856e+000
and be/2 = 2.471512e+001. We note that the lower LPPD point is in the
region relevant to applications while the upper one is not.
It is further interesting to compute the PD curves that emanate from
the LPPD points and they are presented in Fig. 2.15. The stable period-four
cycles exist in the diamond-shaped region bounded by curves LP and PD of
the fourth iterate F 4 .

2.5 Directions for Future Development

We presented an overview of software tools for bifurcation analysis. At present,
the state of the art is a software environment that provides a clear continuation strategy as implemented in Matcont. We showed two examples of
how to use Matcont for ODEs and maps. Matcont has the advantage that
it is implemented in Matlab, which is standard in many applied elds, particularly in engineering. Furthermore, its numerical capabilities include the
computation of normal-form coecients and automatic branch switching.

72

Willy Govaerts and Yuri A Kuznetsov

The development of Matcont is ongoing. In the near future, we hope

to implement the remaining functionalities listed in Sect. 2.3. Furthermore,
we plan to include algorithms for the computation of invariant manifolds and
develop facilities to analyze global bifurcations. Other directions for further
development would be the provision of higher-codimension bifurcations and
the corresponding detection and branching relationships. Moreover, it would
be of interest to generalize the functionalities to other classes of systems, for
example, systems with symmetry or preserved quantities; see also Chap. 9.

Acknowledgments
The authors thank Reza Khoshsiar Ghaziani (Ghent), Hil Meijer (Utrecht)
and Bart Sautois (Ghent) for help in preparing the gures and several helpful
comments and suggestions.

References
1. D. V. Anosov, S. Kh. Aranson, V. I. Arnold, I. U. Bronshtein, V. Z. Grines,
and Yu. S. Ilyashenko. Ordinary dierential equations and smooth dynamical systems. In Dynamical Systems I, volume 1 of Encyclopaedia Math. Sci.
(Springer-Verlag, Berlin, 1988).
2. H. N. Agiza. On the analysis of stability, bifurcation, chaos and chaos control
of Kopel map. Chaos, Solitons & Fractals, 10(11): 19091916, 1999.
3. A. Back, J. Guckenheimer, M. R. Myers, F. J. Wicklin, and P. A. Worfolk.
DsTool: Computer assisted exploration of dynamical systems. Notices Amer.
Math. Soc., 39(4): 303309, 1992. Available via http://www.cam.cornell.edu/
gucken/dstool.
4. W.-J. Beyn and J.-M. Kleinkauf. The numerical computation of homoclinic
orbits for maps. SIAM J. Numer. Anal., 34(3): 12071236, 1997.
5. R. M. Borisyuk. Stationary solutions of a system of ordinary dierential equations depending upon a parameter. FORTRAN Software Series 6, Research
Computing Centre, USSR Academy of Sciences, Pushchino, Moscow Region,
1981 (In Russian).
6. H. W. Broer, H. M. Osinga, and G. Vegter. Algorithms for computing normally
hyperbolic invariant manifolds. Z. Angew. Math. Phys., 48(3): 480524, 1997.
7. E. Brown, J. Moehlis and P. Holmes. On the phase reduction and response
dynamics of neural oscillator populations. Neural Comput., 16: 673715, 2004.
8. A. R. Champneys and V. Kirk. The entwined wiggling of homoclinic curves
emerging from saddle-node/Hopf instabilities. Physica D, 195: 77105, 2004.
9. A. R. Champneys, Yu. A. Kuznetsov, and B. Sandstede. A numerical toolbox
for homoclinic bifurcation analysis. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
6(5): 867-887, 1996.
10. J. W. Demmel, L. Dieci, and M. J. Friedman. Computing connecting orbits
via an improved algorithm for continuing invariant subspaces. SIAM J. Sci.
Computing, 22(1): 8194, 2001.

2 Interactive Continuation Tools

73

11. A. Dhooge, W. Govaerts, and Yu. A. Kuznetsov. Matcont: A Matlab package

for numerical bifurcation analysis of ODEs. ACM Trans. Math. Software 29(2):
141164, 2003. Available via http://www.matcont.ugent.be/.
12. E. J. Doedel and J.-P. Kernevez. Auto: Software for continuation problems in
ordinary dierential equations with applications, Applied Mathematics, California Institute of Technology, Pasadena, 1986.
13. E. J. Doedel, A. R. Champneys, T. F. Fairgrieve, Yu. A. Kuznetsov, B. Sandstede, and X.-J. Wang. Auto97: Continuation and bifurcation software for
ordinary dierential equations (with HomCont). Computer Science, Concordia University, Montreal, 1997. Available via http://cmvl.cs.concordia.ca/
14. K. D. Edoh, and J. Lorenz. Computation of Lyapunov-type numbers for invariant curves of planar maps. em SIAM J. Sci. Comput., 23(4): 11131134,
2001.
15. B. Ermentrout. Simulating, Analyzing, and Animating Dynamical Systems:
A Guide to XppAut for Researchers and Students, volume 14 of Software,
Environments, and Tools SIAM, Philadelphia, 2002. Available via http://
www.math.pitt.edu/bard/xpp/xpp.html.
16. U. Feudel and W. Jansen. Candys/QA - a software system for the qualitative
analysis of nonlinear dynamical systems. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 2(4): 773-794, 1992. Available via http://www.agnld.uni-potsdam.de/
wolfgang/candys.html.
17. M. Friedman, W. Govaerts, Yu. A. Kuznetsov, and B. Sautois. Continuation of
homoclinic orbits in Matlab. In V. S. Sunderam, G. D. van Albada, P. M. A.
Sloot, and J. J. Dongarra, editors. Computational Science ICCS 2005, Atlanta, LNCS 3514, pages 263270. Springer, 2005.
18. W. Govaerts, Yu. A. Kuznetsov, and B. Sijnave. Bifurcations of maps in
the software package Content. In V. G. Ganzha, E. W. Mayr, and E. V.
Vorozhtsov, editors. Computer Algebra in Scientic ComputingCASC99,
pages 191206. Springer, Berlin, 1999. Available via http://www.math.uu.nl/
people/kuznet/CONTENT/.
19. W. Govaerts, Yu. A. Kuznetsov, and B. Sijnave. Continuation of codimension-2
equilibrium bifurcations in Content. In E. J. Doedel and L. S. Tuckerman, editors. Numerical Methods for Bifurcation Problems and Large-Scale Dynamical
Systems, volume 119 of IMA Vol. Math. Appl., pages 163184. Springer, New
York, 2000. Available via http://www.math.uu.nl/people/kuznet/CONTENT/.
20. W. Govaerts, Yu. A. Kuznetsov, and B. Sijnave. Numerical methods for the
generalized Hopf bifurcation. SIAM J. Numer. Anal., 38(1): 329346, 2000.
21. W. Govaerts and B. Sautois. Computation of the phase response curve: a direct
numerical approach. Neural Comput., 18(4): 817847, 2006.
22. B. D. Hassard, N. D. Kazarino, and Y.-H. Wan. Theory and Applications of
Hopf Bifurcation. In volume 41 of London Mathematical Society Lecture Note
Series. Cambridge University Press, London, 1981.
23. F. C. Hoppensteadt and E. M. Izhikevich. Weakly Connected Neural Networks.
Springer, Berlin Heidelberg New York, 1997.
24. A. I. Khibnik. Linlbf: A program for continuation and bifurcation analysis of
equilibria up to codimension three. In D. Roose, B. De Dier, and A. Spence, editors. Continuation and Bifurcations: Numerical Techniques and Applications,
volume 313 of NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., pages 283296.
Dordrecht, 1990.

74

Willy Govaerts and Yuri A Kuznetsov

25. A. I. Khibnik, Yu. A. Kuznetsov, V. V. Levitin, and E. V. Nikolaev. Continuation techniques and interactive software for bifurcation analysis of ODEs and
iterated maps. Physica D, 62(1-4): 360371, 1993.
26. M. Kopel. Simple and complex adjustment dynamics in Cournot duopoly models. Chaos, Solitons & Fractals, 7(12): 20312048, 1996.
27. M. Koper Bifurcations of mixed-mode oscillations in a three-variable autonomous Van der Pol-Dung model with a cross-shaped phase diagram. Physica D, 80(1-2): 7294, 1995.
28. B. Krauskopf and H. M. Osinga. Globalizing two-dimensional unstable manifolds of maps. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 8(3): 483-503, 1998.
29. B. Krauskopf and H. M. Osinga. Growing 1D and quasi-2D unstable manifolds
of maps. J. Comput. Phys., 146(1): 404419, 1998.
30. B. Krauskopf and H. M. Osinga. Investigating torus bifurcations in the forced
Van der Pol oscillator. In E. J. Doedel and L. S. Tuckerman, editors. Numerical
Methods for Bifurcation Problems and Large-Scale Dynamical Systems, volume
119 of IMA Vol. Math. Appl., pages 199208. Springer, New York, 2000.
31. B. Krauskopf, H. M. Osinga, E. J. Doedel, M. E. Henderson, J. Guckenheimer,
A. Vladimirsky, M. Dellnitz, and O. Junge. A survey of methods for computing (un)stable manifolds of vector elds. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 15(3): 763791, 2005.
32. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory, 3rd edition.
(Springer, Berlin Heidelberg New York, 2004).
33. Yu. A. Kuznetsov, W. Govaerts, E. J. Doedel, and A. Dhooge. Numerical
periodic normalization for codim 1 bifurcations of limit cycles. SIAM J. Numer.
Anal., 43(4): 14071435, 2005.
34. H. E. Nusse and J. A. Yorke. Dynamics: Numerical Explorations, 2nd edition.
(Springer-Verlag, New York, 1998).
35. A. G. Salinger, N. M. Bou-Rabee, E. A. Burroughs, R. P. Pawlowski, R. B.
Lehoucq, L. A. Romero, and E. D. Wilkes. Loca 1.0 library of continuation
algorithms: Theory and implementation manual. Sandia National Laboratories Technical Report, SAND2002-0396, 2002. available via http://www.cs.
sandia.gov/loca/.
36. A. G. Salinger, E. A. Burroughs, R. P. Pawlowski, E. T. Phipps, and L. A.
Romero. Bifurcation tracking algorithms and software for large scale applications.. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 15(3): 10151032, 2005.
37. F. Schilder, H. M. Osinga, and W. Vogt. Continuation of quasi-periodic invariant tori. SIAM J. Appl. Dyn. Sys., 4(3): 459-488, 2005.
38. R. Seydel. From Equilibrium to Chaos: Practical Bifurcation and Stability Analysis. (Elsevier, New York, 1988).
39. R. Seydel. Tutorial on continuation. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
1(1): 311, 1991.
40. M. A. Taylor, M. S. Jolly, and I. G. Kevrekidis. Scigma: Stability computations
and interactive graphics for invariant manifold analysis Technical report, Dept.
of Chem. Eng. Princeton University, 1989.
41. M. A. Taylor and I. G. Kevrekidis. Interactive Auto: A graphical interface for
Auto86. Technical report, Dept. of Chem. Eng. Princeton University, 1990.
42. K. Yagasaki. Numerical detection and continuation of homoclinic points and
their bifurcations for maps and periodically forced systems. Internat. J. Bifur.
Chaos Appl. Sci. Engrg., 8(7): 16171627, 1998.

2 Interactive Continuation Tools

75

43. Y. Zhiping, E. J. Kostelich, and J. A. Yorke. Calculating stable and unstable

manifolds. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 1(3): 605623, 1991.
44. Y. Zhiping, E. J. Kostelich, and J. A. Yorke. Erratum: Calculating stable and
unstable manifolds. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 2(1): 215,
1992.

3
Higher-Dimensional Continuation
Michael E Henderson
IBM T.J. Watson Research Center, Yorktown Heights, NY, USA

Numerical continuation in one dimension produces a sequence of points

(ui , i ), i [0, N ), along a set of arcs that are connected at their endpoints.
Each point is the solution of an equation
F (ui , i ) = 0 ,

F : Rn R Rn .

A regular point is a solution F (u0 , 0 ) = 0 where the Jacobian Fu (u0 , 0 )

is nonsingular. According to the Implicit Function Theorem (IFT) there is a
unique curve (u(s), (s)) of regular points in some small neighborhood |s| <
of a regular point. By extending this small piece of the solution curve, detecting singular points and switching branches at singular points, the continuation
method produces an approximation of the connected component (u0 ,0 ) of solutions of F (u, ) = 0 containing (u0 , 0 ); see Fig. 3.1(a).
Most physical problems depend on more than one parameter. While it can
be useful to study how the solution depends on each parameter separately,
what is really wanted is an understanding of the behavior over some nite
piece of parameter space. This is the higher-dimensional continuation problem.
Instead of solution curves, we wish to nd solution manifolds of
F (u, ) = 0 ,

F : Rn Rk Rn .

Using the IFT again, at a regular point there is a unique manifold of regular
solutions F (u(s), (s)) = 0, which exists in some small neighborhood |s| < ,
s Rk . The connected component (u0 ,0 ) is as dened above, but it is a
branched manifold (the solution manifold), rather than a branched curve; see
Fig. 3.1 (b). Singular points on the manifold where the Jacobian has a rank
deciency of 1 form the boundary of each branch. These singular manifolds
are generically (k 1)-dimensional submanifolds.
Numerical continuation methods use a local analysis, usually a computational version of the IFT, which approximates the solution manifold in a
neighborhood of a regular point Ni (ui ), and aggregate n of these neighborhoods into a global approximation of the solution manifold M . The geometric problem of maintaining the aggregate of neighborhoods, and merging a

Michael E Henderson
arc 1

(a)

(b)
(u 0 ,S 0)

arc 4

78

arc 0

( u 0 ,S0 )

arc
3

arc 2

Fig. 3.1. Panel (a) shows connected components of the solution set of F (u, ) = 0;
each component consists of a set of smooth arcs that meet at singular points. Panel
(b) shows a branched two-dimensional manifold that consists of a set of smooth
manifolds that meet along shared one-dimensional boundary manifolds.

new neighborhood into the aggregate, is roughly equivalent to advancingfront mesh generation on a surface. This is not an easy problem, especially in
higher dimensions n and for manifolds of dimension k > 3. In Sect. 3.1 we give
a brief survey of the methods that can be used to represent and manipulate
complexes, which are generalizations of meshes and are used to represent the
branched manifold that approximates M . In Sect. 3.2 we describe ve algorithms that are in the literature, and attempt to classify them in terms of the
representations of the manifold as discussed in Sect. 3.1. Finally, in Sect. 3.3
we compare the results of the ve algorithms when applied to a sphere.

3.1 Mathematical Setting and Background

In computational and pure mathematics it is necessary to choose a representation or notation for the object being computed or manipulated. The parallel
between notation and computational representation is quite close, although
the computer is able to deal with expressions that are far too complicated
to be dealt with by hand. A good notation makes manipulations easier and
reduces errors, and a good computational representation makes operations
easier to implement, thus reducing coding errors. The choice of representation
is often not made explicitly, especially when the objects are simple.
The choice of representation also serves to distinguish between algorithms.
Indeed, once the choice of a representation is made there are often fewer
choices of how to proceed. We use this approach in Sect. 3.2 to analyze ve
algorithms for higher-dimensional continuation that are in the literature.
For one-dimensional continuation many issues are quite straightforward,
since curves are easily represented as lists of points. However, representing

3 Higher-Dimensional Continuation

79

surfaces and manifolds is more challenging. In Sect. 3.1.1 we discuss simplicial

and cell complexes, which are general meshes that are commonly used in
algebraic topology. Then we dene manifolds and manifolds with boundary,
which are the analogues of curves and arcs in one dimension.
Section 3.1.2 describes two fundamental geometric abstractions, the Voronoi and Delaunay tessellations, and a particular generalization of the Voronoi
tessellation that represents the boundary of a union of spherical balls. In
Sect. 3.1.3 we discuss how these topological and geometrical ideas have been
used to represent manifolds and, nally, how to construct the neighborhood
of a regular point.
3.1.1 Cell and Simplicial Complexes, and Manifolds
There are only a few ways to deal with general surfaces, and most are variations on meshes. In this section we discuss representations of general meshes
in higher dimensions, called cell and simplicial complexes. The Voronoi diagram and Delaunay triangulation are two instances of complexes with special
properties. The Voronoi diagram or tessellation contains information about
neighborhoods of a set of points. There is a variant, the Laguerre-Voronoi
diagram, that provides an ecient means of determining the boundary of a
union of spherical balls. The Delaunay triangulation is related to the Voronoi
diagram, and is frequently used to generate meshes in two and three dimensions.
This subject can be presented in a very opaque, but abstract way. The
author would recommend the books [14, 35] and the paper [16] for reasonably
clear expositions of complexes. Hopefully, the descriptions that follow are as
clear.
A cell complex of dimension k is a set of cells of dimension 0 to k. The 0cells (vertices) may (or may not) be identied with points in an n-dimensional
embedding space. If they are not embedded, the complex represents purely
topological information. For our purposes the vertices will be points on or near
the solution manifold. Each p-cell (except the 0-cells) has a set of faces, which
are (p 1)-cells in the complex. Cells must be compatible with each other:
the intersection of two cells is either empty, or it is a cell in the complex
that both cells have as a common face. This compatibility condition turns out
to create major diculties for some of the algorithms described in Sect. 3.2
below. Figure 3.2 shows some cells that are not compatible.
A complex is a set of lists: of cells of each dimension, the faces of each
cell with orientations (the boundary), and a list of cells of which the cell is
a face (the co-boundary). Some of this information is redundant, and when
storage is an issue it is necessary to avoid storing redundant information.
Guibas and Stol [18] describe a minimal data structure for representing twodimensional cell complexes and show that, with operations which are fairly
simple in this representation, a Delaunay triangulation algorithm can be built;

80

Michael E Henderson

(a)

(b)

(c)

Fig. 3.2. Cells that are not compatible, namely 1-cells embedded in R2 (a), 2-cells
embedded in R2 (b), and 2-cells embedded in R3 (c).

see Sect. 3.1.3. For three-dimensional complexes there are similar representations, e.g., the winged edge; see, for example, [13, 19].
The basic geometrical object in higher-dimensional continuation is a kdimensional branched manifold. (Recall that k is the number of parameters.)
A branched manifold is a set of manifolds with boundaries (the branches),
glued together along common boundaries. The boundaries are sets of singular
points, which are generically (k 1)-dimensional manifolds. Branch switching
is the process of nding the branches that meet at a singular surface.
Convex Polyhedral Cells
Convex polyhedral cells have a particularly nice structure. The polyhedron is
formed by intersecting half-spaces whose planar boundaries contain the faces
of the polyhedron. The cells can be represented by the list of planes that they
lie on. So for example, the set of linear inequalities in three dimensions

0 , 1).(v, 1) 0

( e

(
e

0 , 1).(v, 1) 0

1 , 1).(v, 1) 0
( e
P =
(
e1 , 1).(v, 1) 0

2 , 1).(v, 1) 0
( e

(
e2 , 1).(v, 1) 0
denes a polyhedron (a cube), as is shown in Fig. 3.3(a). Given a vertex v0 (a
0-cell) with representation (0, 2, 4), this indicates that v0 satises the linear
system
(
e0 , 1).(v0 , 1) = 0,
(
e1 , 1).(v0 , 1) = 0,
(
e2 , 1).(v0 , 1) = 0.
The representations of the vertices is sucient to dene the rest of the
cells. In a k-dimensional polyhedron each vertex must lie on k or more planes.

3 Higher-Dimensional Continuation
v2 (0,3,5)

(1,3,5) v6

(a)
5

v1 (0,2,5)

5
3

v3 (0,3,4)

v0 (0,2,4)

(b)

v5 (1,2,5) 3
0

81

0
(1,3,4) v7

(1,2,4) v4

1
2

Fig. 3.3. A cube with the face and vertex labels (a) and its dual (b), which is an
octahedron.

Cells of dimension p > 0 must lie on exactly k p faces, and so have k p

indices. There is an edge between two vertices if the vertices have k 1 indices
in common. For example, for the cube in Fig. 3.3
v0 v1 = (0, 2, 4) (0, 2, 5) = (0, 2) ,
v3 v4 = (0, 3, 4) (1, 2, 4) = (4) .
So there is an edge with endpoints v0 and v1 , but not with endpoints v3 and
v4 . The same holds for cells of higher dimension. A p-cell exists that contains
a set of vertices if, and only if, the intersection of the index sets of the vertices
contains k p indices. This works because each plane corresponds to a single
linear constraint, and a set of indices is just a linear system. For example, in
Rk , k 1 linear equations dene a line. The notation used here is introduced
(in a dierent guise) in [17, Sec. 2.6]; for more information on polyhedra see,
for example, [17] or [12].
Subtracting a Half-Space from a Convex Polyhedron
One operation on convex polyhedra requires mention, namely, that of intersecting the polyhedron with a half-space. This adds a plane to the polyhedron,
and the vertices that are the wrong side of the plane and any cell containing
those vertices must be removed. If the convex polyhedron is represented by a
list of inequalities and a list of vertices, then the algorithm described in [12]
may be used.
This algorithm is simple to describe. First the half-space is added to the
list of inequalities, then the vertices are tested for inclusion in the half-space.
Those that are not in the half-space are removed and a new vertex is added
on each edge containing the deleted vertex. The index of these new points is
the index of the edge plus the index of the new inequality. The performance of
the algorithm can be improved by storing not just the vertices but the edges
as well, as is described in [12].

82

Michael E Henderson
2
1

Simplex
(1,2,7,10)

10

Face 1

Face 2

Face 7

Face 10

10

(2,7,10)

10

(1,7,10)

10

(1,2,10)

(1,2,7)

Fig. 3.4. The simplex (1, 2, 7, 10) and its faces. The faces can be enumerated by
removing from the label of the simplex the index of each vertex in turn. Removing
a vertex index yields the label of the face opposite that vertex. The arrows indicate
the positive ordering of the vertices on the edges.

Simplices
A simplex of dimension k is a set of k + 1 vertices that do not lie in a linear
subspace of Rk . A simplicial complex is a cell complex whose cells are all
simplices. Polyhedra are represented in terms of their faces, but a simplex is
represented in terms of its vertices. The faces of a simplex can be obtained
by simple operations on the list of vertices, so do not have to be tabulated. A
two-dimensional simplex (a triangle) with vertices 1, 4 and 5 is represented as
(1, 4, 5). The natural ordering of the integers determines the order in which
the vertices are listed.
The faces of a k-dimensional simplex are simplices whose representation
is the same as that of the parent simplex, but with one vertex dropped.
The face is the one opposite the vertex that is dropped; see Fig. 3.4. This
enumeration holds for (k 1)-cells down to 0-cells, which have a single index.
Orientation, the Boundary Operator, Duality and the
Co-Boundary Operator
Cells in a complex can have an orientation of 1. For simplicial cells the sign
can be computed directly from the indices. The orientations of the faces of
a cell are chosen so that the boundary of the boundary of a cell is empty;
see Fig. 3.5. The sets of cells of the various dimensions in the complex are
connected by boundary operators: cells of dimension p are connected to cells

3 Higher-Dimensional Continuation

83

2
1

Simplex
(1,2,7,10)

10

Face 1

Face 2

Face 7

Face 10

10

-(2,7,10)

10

-(1,7,10)

10

(1,2,10)

(1,2,7)

Fig. 3.5. The Simplex (1, 2, 7, 10) and its faces, oriented so that the boundary of
its boundary is empty.

of dimension p 1 by the boundary operator dp , and p-cells are connected to

p + 1 cells by the co-boundary operator dp , as dened below.
Exchanging two adjacent vertices changes the orientation of the simplex,
and the simplex with vertices in numerical order is assigned an orientation of
+1. So for an arbitrary list of vertices the sign of the simplex is the parity
of the permutation that orders the vertex list. As a result there is a simple
expression for the boundary operator:
dp (v0 , . . . , vi , . . . , vp ) =

p
%

(1)i (v0 , .., vi1 , vi+1 , . . . vp ) .

i=0

This is illustrated in Fig. 3.5. For example, the face (2, 7, 10) in the simplex
(1, 2, 7, 10) (a tetrahedron) has boundary
d2 (2, 7, 10) = (7, 10) (2, 10) + (2, 7) ,
d1 d2 (2, 7, 10) = {(10) (7)} {(10) (2)} + {(7) (2)} = 0 .
Missing from the representation of a k-dimensional simplicial complex as
a list of the vertices in each simplex is information about how the simplices
are connected. Note, however, that if the complex is the surface of a single
(k + 1)-dimensional simplex then the complex is completely dened by the list
of vertices, as is the case in the complex shown in Figs. 3.4 and 3.5. Figure 3.6
illustrates why the orientation is important. The sum (1, 2, 7, 10) + (1, 2, 9, 10)
does not contain the simplex (1, 2, 10), but (1, 2, 7, 10)+(0, 1, 2, 10) does. Note

84

Michael E Henderson

2
1

10

10

(1,2,7,10)

2
1

10

(1,2,9,10)

(1,2,7,10) + (1,2,9,10)

Fig. 3.6. The sum of the boundaries of two simplices (1, 2, 7, 10) and (1, 2, 9, 10).
(2,7,10)

(1,2,7)
2

(a)

10
(1,2,7,10)

(1,2,10)

(b)

(1,7,10)

Dual

Fig. 3.7. The simplex (1, 2, 7, 10) (a) and its dual (b).

that, if a cell appears twice in a sum it is only counted once, the addition is
like a signed union of the sets of cells.
For polyhedral cells the boundary operator is usually tabulated, because
it is not expressed easily. For the cube in Fig. 3.3(a) the boundary of face 0 is
d2 (0) = (0, 2) + (2, 5) (1, 2) (2, 4),
d1 d2 (0) = (0, 2, 5) (0, 2, 4) + (1, 2, 5) (0, 2, 5) + (1, 2, 4) (1, 2, 5)
+(0, 2, 4) (1, 2, 4) = 0 .
Figure 3.3(b) shows the dual of the cube with respect to the faces of
the cube, and Fig. 3.7(b) shows the dual of the simplex. Duality is a general
relationship, and like the duality between a vector space and its adjoint space,
the dual of a cell is a linear functional. The dual of p [i] maps the sum of pcells onto the values {1, 0, 1}. The dual cell p [i] is the functional that maps
the sum to 1 if p [i] occurs in the sum with a positive orientation, to 1 if it
occurs with a negative orientation, and to 0 if p [i] is not present in the sum.
That is,

3 Higher-Dimensional Continuation
p

0
0
0
0
0
0
0
0

cell

dual

(0,2,4) (0, 2, 4)T

(0,2,5)
(1)
(0,4,5)
(2)
(0,3,4)
(3)
(1,2,4)
(4)
(1,2,5)
(5)
(1,3,5)
(1,3,4)

p cell dual

1
1
1
1
1
1
1
1
1
1
1

(0,1) (0, 2)
(0,3) (0, 4)
(2,3) (0, 3)
(1,2) (0, 5)
(4,5) (1, 2)
(4,7) (1, 4)
(5,6) (1, 5)
(0,4) (2, 4)
(1,5) (2, 5)
(3,7) (3, 4)
(2,6) (3, 5)

85

p cell dual

2
2
2
2
2
2
2
2

(0)
(1)
(2)
(3)
(4)
(5)
(6)
(7)

(0, 2, 4)
(0, 2, 5)
(0, 3, 5)
(0, 3, 4)
(1, 2, 4)
(1, 2, 5)
(1, 3, 5)
(1, 3, 4)

Table 3.1. The cells of the cube and of the dual octahedron from Fig. 3.3(b); all of
the indices refer to the faces of the cube.

p [i]p [j] = ij ,
where ij is the Kronecker delta. There is an obvious identication of p [i]
with p [i], which is what is meant by the dual of a cell.
The co-boundary operator dp is the adjoint of the boundary operator dp ,
dened by the relation
(dp p [i]) p+1 [j] = p [i]dp+1 p+1 [j] .
The boundary operator reduces the dimension of a simplex, and the coboundary operator increases the dimension of a dual cell. The boundary operator on the right-hand side of this equation is dp+1 because p [i] acts on
p-cells and the boundary operator dp+1 produces simplices of dimension p. If
p+1 is a (p + 1)-cell then
%
(dp p [i])p+1 [j] = p [i](dp+1 p+1 [j]) =
(1)lk p [i]p [jk ] .
k

Here the boundary of the jth (p + 1)-dimensional cell is given by the list
of simplices j0 , j1 , with orientations (1)l0 , (1)l1 , . The co-boundary
boundary operator applied to p+1 [j] is zero unless p [i] is a face of p+1 [j],
the co-boundary of p [i] is therefore a signed sum of the duals of the p+1-cells
that are incident on p [i].
For a simplex there is an expression for the boundary operator dp+1 : the
list of simplices containing a face can be found by adding each vertex in turn
to the face. The co-boundary operator for a simplicial complex consisting of
a single simplex can be written explicitly as
dp f p =

p+1
%
i=0

(1)i f p (v0 , .., vi1 , wi , vi , . . . vp ) ,

86

Michael E Henderson

kp+1

0
0
0
0
1
1
1
1
1
1
2
2
2
2

(10)
(7)
(2)
(1)
(7,10)
(2,10)
(2,7)
(1,10)
(1,7)
(1,2)
(2,7,10)
(1,7,10)
(1,2,10)
(1,2,7)

(1, 2, 7)
(1, 2, 10)
(1, 7, 10)
(2, 7, 10)
(1, 2)
(1, 7)
(1, 10)
(2, 7)
(2, 10)
(7, 10)
(1)
(2)
(7)
(10)

dp p

(10)-(7)
(10)-(2)
(7)-(2)
(10)-(1)
(7)-(1)
(2)-(1)
(7,10)-(2,10)+(2,7)
(7,10)-(1,10)+(1,7)
(2,10)-(1,10)+(1,2)
(2,7)-(1,7)+(1,2)

dkp+1 kp+1

dp p

(2, 7) (1, 7) + (1, 2)

(2, 10) (1, 10) + (1, 2)
(7, 10) (1, 10) + (1, 7)
(7, 10) (2, 10) + (2, 7)
(2) (1)
(7) (1)
(10) (1)
(7) (2)
(10) (2)
(10) (7)

(1,10)-(2,10)+(7,10)
(1,7)-(2,7)+(7,10)
(1,2)-(1,7)+(1,10)
(1,2)-(2,7)+(2,10)
(1,7,10)-(2,7,10)
(1,2,10)-(2,7,10)
(1,2,7)-(2,7,10)
(1,2,10)-(1,7,10)
(1,2,7)-(1,7,10)
(1,2,7)-(1,2,10)

Table 3.2. The simplicial complex that forms the surface of the simplex (1, 2, 7, 10);
compare with Figs. 3.4 and 3.5.

where wi is in turn each of the vertices that are missing from the index of the
cell. For cell complexes the boundary and co-boundary need to be tabulated,
although there is redundant information in the lists. Table 3.1 shows the cells
in the cube and their duals. Note that, in general, we would write the dual
of 1-cell (0, 1) as (0, 1)T . The surface of the simplex (1, 2, 7, 10) from Figs. 3.4
and 3.5 is the simplicial complex shown in Table 3.2.
One of the advantages of this notation is that two simplicial complexes
can be added. The example in Fig. 3.6 removes the face they share from the
boundary. The result is still a simplicial complex, but it is the boundary of
an octahedron. This makes it possible to perform such operations as adding
a handle to a cell complex [28]. For simplices there is no need to represent
the faces, since they can be generated from the vertex list. For a simplicial
complex the dual edges to each face of a simplex must be stored (i.e., the pair
of simplices that share the faces). For a cell complex there needs to be a data
structure for storing the lists associated with each cell.
3.1.2 Manifolds
A manifold is the generalization of a surface, but it has a lot of the features
of a cell complex. A k-dimensional manifold is a set of k-dimensional neighborhoods of the origin that are isomorphic to the k-dimensional unit ball
B1 (0) = {x | |x| < 1}, called charts, along with adjacency relations indicating
which charts overlap. Two adjacent charts must agree on some common nonempty sub-region, which means that there is a one-to-one and onto mapping
from the subregion of one chart to the corresponding subregion of the other;
see Fig. 3.8. The neighborhoods are called chart domains, and the collection

3 Higher-Dimensional Continuation

mapping from
chart 0 to chart 1
and
chart 1 to chart 0
where they overlap

1
3

Chart 0

87

Chart 1

Chart 2

0
Chart 3

Fig. 3.8. A manifold is an atlas of overlapping charts, with an identication of

overlapping subregions and one-to-one and onto mappings between them.

1
2

(Boundary) Chart 2

(Boundary) Chart 3

0
Chart 0

Chart 1

Fig. 3.9. A manifold with boundary is a manifold but, in addition to charts with
domains that are full neighborhoods, there are boundary charts with domains that
are the intersection of a full neighborhood with half-spaces containing the origin.

of charts is called an atlas. This is in reference to navigational charts, which

cover a small piece of the globe and are bound together into an atlas. In this
analogy, the overlaps are needed to move from one chart to the next, and the
adjacency is usually indicated by the number of the neighboring chart in the
margin; see Fig. 3.8.
The relation between a manifold and a cell complex is straightforward. A
chart is a k-cell, the overlap between two charts is a k 1 cell and a common
face of each chart, and so on. With this interpretation the manifold shown in
Fig. 3.8 is a tetrahedron.
Manifolds with Boundary
A manifold with boundary has special charts called boundary charts. Instead
of a full neighborhood of the origin, the domain of a boundary chart is iso

1/2
0
morphic to the half-ball B1 (0)
, x1 , . . .) | |x| 1, x
0 0 . The
= x0= (x
restriction to the ball D1 (0) = x = (x , x1 , . . .) | |x| 1, x0 = 0 , a full ball
of one dimension less, is a chart on one of the boundary manifolds. The boundaries themselves can have boundaries (just as the edges of a square have end

88

Michael E Henderson
chart
mapping
1

inverse of
chart
chart mapping
mapping
1
0
chart
mapping
0

chart
mapping
3

2
0

chart
mapping
2

mapping from
chart 0 to chart 1
where they overlap

0
0

Chart 0

Chart 1

3
(Boundary) Chart 2

(Boundary) Chart 3

Fig. 3.10. An embedded manifold with boundary.

points), so instead of a half-ball we really mean a ball that is restricted so

that some set of the coordinates are non-negative; see Fig. 3.9.
This is a second way that manifolds with boundary are like a cell complex. The manifold itself is a cell, and the faces of the cell are the boundary
manifolds.
Embedding a Manifold
The solution manifold M of F (u, ) is a set of points in Rn , but so far the
denition of a manifold has been in terms of neighborhoods of the origin in
Rk . The relation between a k-dimensional manifold and the solution space Rn
is an embedding of the manifold. (Here we assume that n is large enough so
that the manifold does not self-intersect.) Each chart is assigned a one-toone mapping to Rn , called the chart mapping, which must map points in the
overlap of two charts to the same points in the embedding space. With an
embedding the mapping between the overlap of two charts can be expressed
using the chart mapping of one and the inverse of the chart mapping of the
other. For a manifold dened by a system of equations we considered branches
that consist of regular points, and the IFT gives the chart mapping. The
singular boundary manifolds are where this embedding fails; see Fig. 3.10.

3 Higher-Dimensional Continuation
cha
rt 1

89

(b)
chart 5

(a)

ch
ar
t4

chart 2

cha
rt 3

(d)
(c)

Fig. 3.11. The four main ways of representing a manifold. Panel (a) illustrates the
denition for k = 1 and n = 2, namely a list of overlapping chart mappings from Rk
to Rn . Panel (b) shows a covering or containment in solution space for k = 1 and
n = 3, panel (c) a triangulation for k = 2 and n = 3, and panel (d) a polygonal
tiling for k = 2 and n = 3.

Representing a Manifold
The algorithms described in Sect. 3.2 use dierent representations of the solution manifold. Allgower and Schmidts algorithm [3] represents the solution
manifold as a set of simplices in the embedding space Rn , each of which
contains a piece of the manifold. Rheinboldts moving-frame algorithm [31],
Brodziks algorithm [9], and Melville and Mackeys boundary representation
[29] use instead a simplicial complex whose vertices are points on the solution
manifold. The authors algorithm [20] represents the solution manifold as a
complex with convex polyhedral cells. These three dierent representations, as
used by the ve algorithms described in Sect. 3.3, are illustrated in Fig. 3.11.
3.1.3 Basic Computational Geometry
The representations of manifolds that are used for higher-dimensional continuation are all based on cell or simplicial complexes. In computational geometry

90

Michael E Henderson
(a)

(b)

Fig. 3.12. The Voronoi diagram (a) of a set of points in the plane, where each
region is the set of points closest to a particular point, and the dual Voronoi and
Delaunay diagrams (b).

two of the most frequently used complexes are the Delaunay triangulation of
a set of points and the dual Voronoi diagram. The Delaunay triangulation in
higher dimensions is in fact not a triangulation, but it is still referred to as a
triangulation. Generically, the Delaunay triangulation is a simplicial complex,
and the Voronoi diagram is a complex with convex polyhedral cells. There is a
large literature on both, and the reader may wish to consult [7, 8, 11, 15, 34],
or any introductory text on computational geometry.
The Delaunay triangulation has good properties for mesh cells, namely it
creates fat simplices. In two dimensions it has been proved [30] that, over
all triangulations of a xed set of points, the Delaunay triangulation is the
one that maximizes the smallest angle in any of the triangles. The Voronoi
diagram contains information about nearest neighbors, so is used in many
pattern matching applications.
Voronoi Diagrams
Given a set of points ui in Rn , the Voronoi diagram of the points is a decomposition of Rn into n-cells, each associated with one of the points. The
Voronoi cell Vi of ui is
Vi = {u Rn | |u ui | < |u uj | for all j = i} .
Here || is the Euclidean 2-norm. Figure 3.12(a) shows an example of a Voronoi
diagram for a set of points in the plane.
Each Voronoi region is a domain of inuence of the respective point. A
Voronoi region also gives information about the nearest neighbors of each
point through (n 1)-dimensional faces that separate the Voronoi regions of
two nearby points.

3 Higher-Dimensional Continuation

(s0,s1)

Ri

Rj

(s0j ,s1j )

(s0i ,s1i )

(s0,s1)

Ri

(s0i ,s1i )

91

Rj

(s0j ,s1j )

dBi Bj :
(s0 s0i )2 + (s1 )2 = Ri2
(s0 s0j )2 + (s1 )2 Rj2

2(s0i

(s0 s0i )2 + (s1 )2 = Ri2

s0j )s0 + (s0j )2 (s0i )2 Rj2 Ri2

Fig. 3.13. The Laguerre-Voronoi face when |s0i s0j | Ri + Rj (a), and when
|s0i s0j | > Ri + Rj (b).

The Laguerre-Voronoi Diagram and the Boundary of a Union of

Spherical Balls
There are several generalizations of the Voronoi diagram [24]. One of them,
the Laguerre-Voronoi diagram or power diagram [4, 5, 6, 22], can be used to
nd points on the boundary of a union of spherical balls, an operation which
will be referred to in Sect. 3.2. In the Laguerre-Voronoi diagram each point is
given a weight, which is the radius of the spherical ball about the point.
The face between two Laguerre-Voronoi cells is dened by the equation
|s si |2 Ri2 = |s sj |2 Rj2 ,
which is easily solved. We nd that points s on this Voronoi face are solutions
of
2(sj si ).s = Ri2 Rj2 + |si |2 + |sj |2 ,
which is a plane orthogonal to the line connecting the centers of the two cells.
So the Laguerre-Voronoi diagram has cells with planar faces. Substituting
s = si + (sj si ), we nd that the plane intersects the line between the two
points at
=

Ri2 Rj2 + |si sj |2

.
2 |si sj |2

If |si sj | Ri + Rj then the plane contains the intersection of the two

spheres, and it divides the part of the rst sphere that lies outside the second
from the part that lies inside the second (and vice versa); see Fig. 3.13. The

92

Michael E Henderson
(a)

(b)

(c)

Fig. 3.14. Panel (a) shows the Laguerre-Voronoi diagram of a set of points in the
plane, where each region is the set of points closest to a particular point for the
distance |s si |2 Ri2 . Panel (b) shows that the Laguerre-Voronoi diagram contains
information about the boundary of the union of balls, namely, the boundary of the
union of balls is the union of the part of the boundary of each ball that lies within
its Voronoi region. Panel (c) illustrates that only the part of the Laguerre-Voronoi
diagram is needed that corresponds to overlapping balls; it is called the restricted
Laguerre-Voronoi diagram.

n-cell is the intersection of the half-spaces dened by each pair of points,

so it is a convex polyhedron, and the boundary of the union of balls which
dene the diagram is the union of the parts of each sphere that lie inside the
corresponding Laguerre-Voronoi n-cell.
The boundary of a union of spherical balls is the union of the parts of
the spherical boundaries that do not lie inside another ball. Therefore, the
boundary of the union consists of the parts of the spheres that lie inside their
restricted-Laguerre Voronoi diagram; see Fig. 3.14.
The Delaunay Triangulation
The Delaunay triangulation is the dual of the Voronoi diagram in the sense
discussed in Sect. 3.1.1 for complexes; see Fig. 3.15. In two dimensions it has

3 Higher-Dimensional Continuation

(a)

93

(b)

Fig. 3.15. Panel (a) shows the Delaunay tessellation of a set of points in the plane.
Panel (b) illustrates that each simplex in the tessellation satises the in circle test,
that is, no point lies interior to a Euclidean sphere passing through the vertices of
the simplex.

been shown [30] that Delaunay triangulations, among all possible triangulations of a set of points, maximize the smallest angle in any triangle. In three
dimensions there appears to be a similar property, but it is not clear exactly
what maximum principle there might be. Because of this property of having
fat cells, the Delaunay triangulation is widely used for mesh generation. For
a more complete explanation of the properties of Delaunay triangulations, and
how they are constructed, see [7, 15, 18, 33].
The Delaunay tessellation can be dened independently of the Voronoi
diagram by means of the in-circle test. Each n-cell of the triangulation is
such that no point is inside the Euclidean sphere that contains the vertices of
the n-cell; see Fig. 3.15.
The Coxeter-Freudenthal-Kuhn Triangulation
The Coxeter-Freudenthal-Kuhn triangulation (or tessellation) [2] is a simplicial decomposition of Rn . (It is also described in [3], but with quite a few typographic errors.) The tessellation is dened in terms of an initial n-dimensional
simplex with n + 1 vertices labeled vi for i = 0 to n. An (n 1)-dimensional
face separates two n-dimensional simplices that share the face. If one of the
simplices has vertices (v0 , v1 , . . . , vn ) then in this triangulation the simplex
across the face vi is dened as the simplex with vertices
Pvi (v0 , v1 , . . . , vn ) := (v0 , v1 , . . . , vi1 , vi , vi+1 , . . . , vn ) ,
where

v1 + vn v0 ,
vi = vi+1 + vi1 vi ,

vn1 + v0 vn ,

i = 0,
0 < i < n,
i = n.

94

Michael E Henderson
1+(2-0)

1+(0-2)

1+(3-2)

2
3

(a)

0
0

(b)
3-2

1
2
1

0+(2-1)

Fig. 3.16. A pivot in the two-dimensional Kuhn triangulation (a), and in the threedimensional Kuhn triangulation (b).
(0,0)+(0,1)

(0,0)+(0,1)+(1,0)
(0,0)+(1,0)+(0,1)

^ +e
^
(0,0,0) + e^ 0+ e
1
2

(a)

(0,0,0) + e^0+ e^1

(b)

0
(0,0)

(0,0)+(1,0)

(0,0,0)

(0,0,0) + e^0

Fig. 3.17. A path simplex decomposition of a square (a) and of a cube (b). The
simplices for the n-dimensional cube are generated by paths with n segments, each
parallel to a coordinate direction, with no direction repeated.

This operation of moving across a face to an adjacent simplex is called a

pivot. Pivoting works for any initial simplex, but one particular choice has an
explicit representation and is well suited to computations. The initial simplex
is a path simplex. A path simplex has vertices that are dened by the end
points of the segments of a piecewise linear path; see Fig. 3.16. In particular,
if the coordinate directions are used as segments, and no coordinate direction
is allowed to appear twice, then the set of path simplices decomposes the
interior of the unit cube; see Fig. 3.17 and [25] where the idea is attributed to
Tucker [26, Problem 3, p. 140]). It is trivial to tile Rn with cubes and, hence,
this decomposition of the cube also gives a simplicial decomposition of Rn .
This is not true for an arbitrary decomposition of the cube, since translation
in a coordinate direction will generate incompatible faces unless opposite faces
of the cube have the same decomposition.
If the coordinate of the lower left corner of a cube the corner that
is closest to the origin in Fig. 3.17(b) is v0 , and the side of the cube

3 Higher-Dimensional Continuation

95

is length then the simplices in this decomposition can be represented by

a translation vector z Nn , and a permutation Sn from the group
of permutations of the symbols {1, , n}. The permutation identies the
simplex in the decomposition of the cube. The vertices of the simplex with
representation (z, ) are given by the recursion
v0 (z, ) = v0 + z ,
ej+1 .
vj+1 = vj +
The eect of the pivot across face j (Pj ) on the permutation

{0 , . . . , j1 , j+1 , j , j+2 , . . . , n1 } ,
Pj = {1 , . . . , n1 , 0 } ,

{n1 , 0 , . . . , n2 } ,

is
0 < j < n,
j = 0,
j = n,

and the action on z is

z,
0 ,
Pj z = z + e

n1 ,
ze

0 < j < n,
j = 0,
j = n.

3.2 Five Algorithms

We now have the language and tools for describing algorithms for higherdimensional continuation. Cell and simplicial complexes are used to represent
the solution manifold, Delaunay triangulations to decompose the manifold,
the restricted Laguerre-Voronoi tessellation to obtain information about the
boundary of a collection of balls, and the Coxeter-Kuhn-Freudenthal tessellation for nding a set of simplices that encloses the solution manifold.
Specically, we describe ve algorithms:

Allgower and Schmidts pattern algorithm [3];

Rheinboldts moving-frame algorithm [31];
Rheinboldt and Brodziks [10] tiling algorithm for k = 2 and Brodziks [9]
generalization to any dimension;
Melville and Mackeys k = 2 boundary representation algorithm [29];
and the authors covering algorithm [20].
All of these algorithms can be viewed as an iterative application of three basic
steps to a representation Mi of the manifold:
1. nd a point ui on the boundary of Mi .
2. build a neighborhood Ni of ui .
3. merge Ni into Mi to obtain Mi+1 .

96

Michael E Henderson

The algorithms dier in how Mi is represented, and how the three operations
are performed.
Continuation of higher-dimensional manifolds is conceptually dierent
from that of one-dimensional ones. In fact, there are only two ways to represent a 1-manifold.
The rst corresponds to the approach of pseudo-arclength continuation
[23], and it uses a set of polygonal arcs to represent M . Pseudo-arclength
continuation exploits the fact that for 1-manifolds the boundary is a set of
points, so that the merge operation is just a matter of discarding one of the
two intervals that make up the neighborhood of a point on M . However, if a
simplicial approximation of M is used in higher dimensions then the boundary
of Mi is a (k 1)-dimensional simplicial complex. Constructing a simplicial
neighborhood of a boundary point so that the simplices are compatible with
Mi seems tractable, but turns out to be a dicult problem.
The second corresponds to simplicial or piecewise-linear continuation [1],
and it uses n-dimensional simplices in Rn to cover M . For 1-manifolds
in Rn+1 simplicial continuation uses an (n + 1)-dimensional Coxeter-KuhnFreundenthal simplicial complex, and only requires pivots across n-cells. This
is because a boundary point is the face of an (n+1)-dimensional simplex, and
with k = 1, n-dimensional faces intersect M at a point (or not at all). There
are exactly two simplices that contain this boundary face. One is the simplex
on the boundary and the other can be found with the pivot operation. In
higher dimensions the intersection of M with an (n + k)-dimensional simplex
is a k-cell, and there are more than two simplices containing an n-dimensional
simplex that intersects M at a point.
It is easier to understand an algorithm when the manifold is at, so we will
make use of natural parameter continuation to present the ve algorithms.
3.2.1 Natural Parameter Continuation
Let us begin with the case that there is a unique solution at each point in
parameter space, and consider the generalization of natural parameter continuation; see Fig. 3.18. Natural parameter continuation is a fairly obvious
means of adapting an iterative solver for F (u, ) at a xed parameter value
to map out the solution manifold. Iterative methods require an initial guess,
and rather than start with the same initial guess at each new parameter value,
the solution at a nearby parameter value is used.
If a set mesh is used, natural parameter continuation selects a point from
the mesh at which the solution is known, and which has a neighbor at which
the solution is not yet known; see Fig. 3.18(a) and (b). If the points in parameter space are chosen adaptively, the method becomes a generalization
of advancing-front mesh generation, which is a challenging problem in higher
dimensions. The main consideration is that the new point lies on the edge of
the previously computed points, but near enough so that one of the known

3 Higher-Dimensional Continuation

(a)

97

(b)

u
S0 S1 S2 S3 S4

S1
S

(c)

(d)

S1

S0

S1

S0

Fig. 3.18. Panel (a) illustrates natural parameter continuation in one parameter;
the solution at the point to the left is used as an initial guess for an iterative
method at the next point. Panel (b) illustrates natural parameter continuation in
two parameters, making use of a rectangular grid on parameter space; the technique
is the same as for one parameter, but the grid vertices are traveled in a predetermined
order. Panel (c) illustrates natural parameter continuation in two parameters, but
with a triangular grid on the region of interest in parameter space. Panel (d) shows
a case where natural parameter continuation fails.

points provides a good initial guess. This makes the boundary of the meshed
region an important object; see Fig. 3.18(c).
3.2.2 Solution Space Continuation
For many interesting problems the solution manifold cannot be expressed as
a function of the parameter. Fig. 3.18(d) shows such a case for k = 2. Natural
parameter continuation would nd the lower or upper sheet, depending on the
initial guess, and the iteration would fail to converge beyond the fold.
When k = 1 there are two choices of algorithm. If n (the dimension of
the solution space) is small then simplicial continuation can be used. On the
other hand, pseudo-arclength continuation may be used for any n.
Allgower and Georgs k = 1 simplicial continuation begins with a simplex
in Rn+1 that contains a point on the solution manifold, which is a branched
curve. If the simplex is suciently small and the point is a regular point then
the curve enters this simplex through one face and leaves through another.
With these assumptions a linear approximation of F over the simplex results

98

Michael E Henderson
0

S = (S0,S 1 + +S )

u0

S = (S0,S 1 )

S = (S0,S 1 - +S )

Fig. 3.19. The solution of F = 0 for a piecewise-linear interpolant with the second
parameter xed. As the second parameter is allowed to change, the entry and exit
points move and the line segment connecting them sweeps out a ruled surface.

(a)
orthogonal
"corrector"

u0
tangent
"predictor"

tangent
"predictor"

(b)

(u0,S0 )
S

Fig. 3.20. Pseudo arclength continuation (a), and an isola (b) that is repeatedly
traced.
.

in a good approximation to these two points on the (n1)-dimensional faces of

the simplex. It is important that simplices are used, so that there is a unique,
continuous piecewise-linear interpolant. This also produces a piecewise-linear
solution curve. Figure 3.19 shows the eect of a second parameter on the
piecewise-linear solution of F = 0.
Pseudo-arclength continuation (PSALC) [23] uses the tangent of the solution curve to dene a new parameter. This is done by appending a constraint
that the projection of a solution point (u(s), (s)) onto the tangent vector
is s. The curve of solutions is represented as a piecewise-linear arc, and is
extended by using one of the end points and a linear extrapolation for an
initial guess to Newtons method for the system plus the pseudo-arclength

3 Higher-Dimensional Continuation

99

Corrector

Predictor

u0

Fig. 3.21. Pseudo-arclength continuation in higher dimensions where one parameter

is chosen to follow the corresponding curve on the solution manifold M .

constraint; see Fig. 3.20(a) and Chap. 1. When the solution set is a closed
curve, or isola, the PSALC algorithm will repeatedly trace the isola, as shown
in Fig. 3.20(b). Usually an upper limit on the arclength is included in the conditions that terminate the algorithm. In higher dimensions we will call this
the self-intersection problem, and it is related to the problem of incompatible
simplices.
Figure 3.21 shows the eect of adding an additional parameter to PSALC.
The tangent space is no longer one-dimensional, and it is possible to make a
step in more than one direction. The continuation algorithm must choose in
which direction to step while ensuring that the resulting points sample the
solution manifold uniformly.
3.2.3 Local Analysis
To build a neighborhood of M at a point u0 Rn a local analysis is required.
The operations used by the ve algorithms are
Find a basis for the k-dimensional tangent space of M at a point u,
Project a point from the tangent space onto M ,
Estimate the size of a ball in the tangent space so that points within the
ball project uniquely onto M .
Determine if an n-dimensional simplex contains a point on a piecewiselinear approximation to M .

100

Michael E Henderson

M1

u0

M0

(a)

u0+ - s
u

u0
(b)

u0

u0+ - s
uss(s)

(c)

Fig. 3.22. The tangent space of the solution manifold M at u0 spanned by the
orthonormal basis = [0 , 1 ] (a), projecting a point s in the tangent space orthogonally to obtain a point u on M (b), and the curvature uss (s) of M at u (c).

Finding a Basis for the Tangent Space

The manifold M is dened by the equation F (u) = 0. If u is a point on M
then the tangent space of M at u satises
Fu (u) = 0 ,
T = I .
Finding the tangent space means nding a basis for the nullspace of the Jacobian. At a regular point Fji is full rank, which is rank n, so the nullspace
is k-dimensional; see Fig. 3.22(a). If n is moderately small a singular value
decomposition or QR decomposition might be used. If the problem is large,
one approach is to use the tangent space at a nearby point. This replaces the
rst in the normalization, and results in the linear system

Fu (u)
0
= .

I
T0

3 Higher-Dimensional Continuation

101

The columns of 0 are the orthonormal basis for the nearby tangent space.
The columns of will not be orthonormal, but Gram-Schmidt can easily
be used to nd an orthonormal basis from . This linear system aligns the
new tangent space as much as possible with the old one, which is used in
Rheinboldts wrapping algorithm.
Whichever system is used, any structure in Fu should be exploited. For
example, if F is a two point boundary value problem and collocation is used,
then the collocation points can be eliminated to yield a smaller system.
Projecting a Point in the Tangent Space onto M
The projection of a point s Rk in the tangent space of M at u0 (with
orthonormal basis 0 ) orthogonally to the tangent space is the solution of the
system
F (u) = 0 ,
T (u u0 ) = s .
This is illustrated in Fig. 3.22(b). If u0 is a regular point of M and s is small
enough then the Jacobian of this system is nonsingular. Modied Newtons
method for the nonlinear system results in linear systems with the same matrix
as for the tangent space, but with dierent right-hand sides.
Estimating the Size of a Ball in the Tangent Space
The IFT in nite-dimensional spaces is a modied Newtons method, and
there are bounds on the size of the ball in terms of norms of the Jacobian,
its inverse and Lipschitz bounds on the Jacobian. Experience indicates that
local estimates of these quantities are expensive to compute and provide a
very conservative radius. Global bounds can sometimes be found, but then
the estimated radius is even more conservative.
A dierent approach is to impose a maximum number of Newton iterations
required for the projection, and to reduce the radius by a xed factor if the
number exceeds the maximum allowed. This is an estimate that comes after
the fact, and the way we posed the continuation method above requires an
estimate before the projection is performed.
Here we present a method from [20] that chooses the radius of the ball so
that the distance from the tangent space to the manifold is roughly constant;
see Fig. 3.22(c). If the rate of convergence of Newtons method is constant
over the manifold then this method is equivalent to limiting the number of
Newton steps.
Using Taylors remainder theorem, we have
|u(s) u0 s|

1
|uss ()s2 | ,
2

102

Michael E Henderson
(u0 ,S0)
F

Z2

F(u 2,S 2)

Z1
(0,0,1)

(u2 ,S 2)

(0,1,0)

-1

+
(u1,S1)

(0,0)

S1

F(u0 ,S 0 )
S0

(1,0,0)

F(u1 ,S 1 )

Z0

Fig. 3.23. The image of an n-dimensional face in a simplicial complex under F

(with the mapping applied only to the vertices). This is equivalent to a global
piecewise linear approximation of F over the complex. The question of whether
an n-dimensional face in a simplicial decomposition of R n+k contains a zero of the
interpolant is equivalent to whether 0 lies in the image of the face.

where is some point in the ball |s|. If a tolerance is given on the error
then s must lie in a ball of the radius of the ball in which the error is less than
that tolerance, that is
+

.
R(s) =
2|uss |

The second derivative of u is the solution of the system

Fu (u0 )uss = Fuu ,
T uss = 0 .
It is, therefore, possible to nd the second derivative (which is a k k matrix
whose entries are vectors in Rn ), and estimate its norm.
Determining Whether an n-Dimensional Simplex Crosses M
The mapping F (u) takes a simplicial complex with vertices vi Rn+k and
assigns new coordinates F (vi ). In particular, F maps an n-dimensional simplex with vertices in Rn+k into an n-dimensional simplex in Rn (provided the
vertices are regular points and the simplices are small enough); see Fig. 3.23.
If an n-dimensional simplex contains a point on the solution manifold
then it will generically be a single point. Using barycentric coordinates , the
system for the point is
n+k
%

F(

i vi ) = 0 ,

nk
%
0

i = 1 .

3 Higher-Dimensional Continuation

103

This square system for the unknowns i is a mapping from a reference simplex in Rn+1 to the simplex with vertices F (vi ). In practice, a piecewise-linear
approximation to F is used. Again using barycentric coordinates, the approximation within a p-cell is
p
p
%
%
i vi )
F (vi )i .
F(
0

This expression only depends on the vertices of the p-cell, so the approximation is continuous between adjacent simplices. The point where the interpolant
is zero is the solution of the (n + 1) (n + 1) linear system

F 0 (v0 ) . . . F 0 (vn )
0
0

.. ..
..
..

. .
.
.

= .

F n1 (vn ) . . . F n1 (vn ) n1 0
1
...
1
1
n
This is one step of Newtons method when using dierences for the
Jacobian. If, say, 0 is eliminated using the last row then we have that
0 = 1 1 . . . n , and the system for the remaining i is

0
F 0 (v1 ) F 0 (v0 ) . . . F 0 (vn ) F 0 (vn )
F j (v0 )
..

..
..
..
. =
.

.
.
.
F n1 (vn ) F 0 (v0 ) . . . F n1 (vn ) F 0 (vn )

F j (vn )

3.2.4 Allgower and Schmidts Pattern Algorithm

This algorithm described in [3] extends Allgower and Georgs one-dimensional
simplicial continuation [1] to arbitrary dimension. It produces a list of (n+k)dimensional simplices and points at which a piecewise-linear approximation
to F is zero.
The solution space Rn+k is decomposed as a simplicial complex by using the Kuhn-Freudenthal triangulation; cf. Sect. 3.1.3. The initial solution
(u0 , 0 ) is used to choose the origin z so that the simplex (z, ) with the
identity permutation contains the initial solution. For example, the expression for the vertices of the simplex can be
used to choose z so that the point
with barycentric coordinates (1, . . . , 1)/ n + k is the initial point. A list of
the indices (z, ) of the simplices which will be output is kept, starting with
the initial simplex.
Iteratively, the n-dimensional faces of the simplices in the aggregate that
lie on the boundary are tested to nd those that cross M ; cf. Sect. 3.2.3. The
candidate faces are kept in a list, which is updated as simplices are added to
the aggregate. The n-faces of the initial simplex are found by removing all
combinations (without order) of k vertices from the index of the simplex. The
n-dimensional simplices that are found to cross M play the role of points on

104

Michael E Henderson

v3 "u "

v3 "u "
i

Mi

v1

M i+1

Mi

v0

v3

v0

v1

v2

v0

v1

v2

Fig. 3.24. Allgower and Schmidts algorithm: a transverse (n k)-dimensional face

(an edge since n k = 1 ) and the simplices on which it lies. This collection of
simplices is the neighborhood of the transverse face. The merge is easy, since the
simplices are all drawn from a larger simplicial complex.

the boundary. The points on the piecewise-linear interpolant can be stored,

but are not needed by the algorithm.
With a simplex that crosses M , all of the (n + k)-dimensional simplices
that have this simplex as a face (i.e. the co-boundary) are checked. Any that
are not on the list are added. This is a process of adding k vertices to the
index of the face, with vertices that lie across any (n + k 1)-dimensional face
that contains the boundary simplex. This set of simplices is the neighborhood
of the boundary; see Fig. 3.24.
Finally, the list of boundary points is updated, by removing those that
lie on an (n + k 1)-dimensional face for which both cells on either side of
the face are in the list, and by adding those n-dimensional faces of the newly
added simplices that lie on (n + k 1)-dimensional simplices with only one
cell on the list. Finding the index of the simplex on the opposite side of an
(n + k 1)-dimensional face is a pivot, and the action of the pivot on the
index is known.
3.2.5 Rheinboldts Moving-Frame Algorithm
Rheinboldts moving-frame algorithm [32] represents M as a k-dimensional
mesh with vertices on M . Any xed mesh can be used, but there must be an
ordering on the mesh points si and a way to nd a point si with i < i and
close to si . For a rectangular mesh this is straightforward.
The initial point is s0 , which is mapped to the initial point (u0 , 0 ), and
0 is any orthonormal basis for the null space of the Jacobian at the initial
point; see Fig. 3.25.
Each mesh point is computed in sequence. At step i, (ui , i ) is the boundary point.
The neighborhood of the boundary point is not explicitly represented, but
might be dened as the mesh cells with index less than i that include si as a
vertex. The only point without a mapping to M is the point i. The mapped
point (ui , i ) is found by projecting si si orthogonally onto M by using the

3 Higher-Dimensional Continuation

105

u22
(a)

21

22

23

16

17

18

19

20

11

12

13

14

15

10

222

(b)

22

S1

u
S1

S0

S0

Fig. 3.25. Rheinboldts moving-frame algorithm when using a rectangular mesh,

shown as the natural parameter version (a) and as the solution-space version (b).
The natural order (e.g., left to right and bottom to top) is used on the mesh vertices,
and the continuation advances in the same order. Moves from scan line to scan line
use the point at the beginning of the scan as an initial guess.

tangent space i . The new tangent space is found by rst solving the linear
system
i = 0 ,
F( u, )(ui , i )
Ti i = I
and then orthonormalizing i . The merge operation is trivial as it just involves
incrementing i.
3.2.6 Brodziks Tiling Algorithm
Brodzik and Rheinboldts continuation method for 2-manifolds [10] and
Brodziks extension to arbitrary dimension [9] represent M as a k-dimensional
Delaunay triangulation with vertices that lie on M .
The initial simplicial complex is a reference Delaunay triangulation of a
k-dimensional spherical ball. The vertices on M are found by projecting the
set of vertices si onto M by using the tangent space at the initial point.
A boundary point is a vertex on any simplex that has a (k1)-dimensional
face without two simplices on opposite sides of the face.
The neighborhood is the projection of the vertices of the reference decomposition of the spherical ball orthogonally onto M , although not all vertices
are projected.
The dicult step in this algorithm is the merge. The points of the current
Delaunay triangulation which are near the boundary point are projected into
the tangent space at the new point
sj = Ti ((uj , j ) (ui , i )) .

106

Michael E Henderson

Mi
Interior

ui
2i

Triangulation
of Ball

Too Close

M i+1

Incircle!

Fig. 3.26. The natural parameter version of Brodzik and Rheinboldts algorithm
[10] and Brodizks extension to arbitrary dimension [9].

(a)

(b)

Fig. 3.27. The arclength continuation version of Brodzik and Rheinboldts algorithm [10] and Brodizks extension to arbitrary dimension [9].

3 Higher-Dimensional Continuation

107

u0
20

S1

M0
S0

Fig. 3.28. A homotopy in the radius R of the circle about i for obtaining a starting
point on the boundary of the neighborhood of i for Melville and Mackeys boundary
algorithm.

The previously computed vertices carry with them part of the entire Delaunay
triangulation. First, points from the spherical ball that lie inside this projected
triangulation are discarded. Next, points from the sphere that lie too close to a
vertex of the triangulation are also discarded. Finally, the remaining points are
projected orthogonally onto M and the Delaunay triangulation is updated to
include the new vertices. This last step is done in the tangent space. It involves
removing some of the cells from the triangulation and replacing them with
new cells; see Fig. 3.26.
The extension to an arclength continuation is done as follows. The points
that are near the boundary point are projected onto the tangent space at the
boundary point. Then one proceeds as if the tangent space coordinates were
the parameters in the natural parameter continuation; Fig. 3.27.
3.2.7 Melville and Mackeys Tiling Algorithm
Melville and Mackeys algorithm [29] is strictly two-dimensional. M is not
explicitly represented, rather the boundary of the triangulation is represented
as a polygonal curve with vertices on M .
The initial polygon is found using PSALC, starting at the initial solution,
and solving the modied system
F (u, ) = 0 ,
|u u0 |2 + | 0 |2 R2 = 0 .
To obtain an initial solution for this system, a homotopy can be performed
in the radius R. To this end, one starts at R = 0 when (u0 , 0 ) is a solution
and increases R to a nal value that is chosen to be roughly the desired
resolution of M . With a point on the intersection of M and the sphere, PSALC
can be used to trace the intersection of M and the sphere clockwise relative
to 0 . The polygon formed by the computed points is the initial boundary

108

Michael E Henderson

u1

(a)

21

S1

(b)

M1

S1

S0

S0

Fig. 3.29. The natural parameter version of Melville and Mackeys boundary representation algorithm for k = 2. Starting on the boundary of the region of interest,
a neighborhood is constructed by continuing around a circle (a); this neighborhood
is then merged with the region of interest (b).

(a)

(b)

Fig. 3.30. Melville and Mackeys boundary representation algorithm for k = 2 and
n = 3.

polygon. The tangent space i at each point on the boundary is found as in

the moving-frame algorithm, which preserves the orientation of tangent space;
see Fig. 3.28.
With an explicit representation of the boundary, any vertex will serve as
a boundary point.

3 Higher-Dimensional Continuation
M3

M
3

u3

S1

(a)
S0

S1

(b)

S1

S0

M
4

u3
2

109

3
(c)
S0

Fig. 3.31. The natural parameter version of the authors covering algorithm.

The neighborhood of the boundary point is a circular disk of radius R,

represented by its boundary. Two PSALCs are done in the tangent space of
the boundary point. The rst to the circle, and the second around the circle.
The merge operation starts by projecting the part of the existing boundary
whose vertices are near the boundary point into the tangent space at the
boundary point. This leaves a relatively simple two-dimensional problem of
clipping the boundary of the neighborhood against the existing boundary; see
Figs. 3.29 and 3.30.
3.2.8 The Authors Covering Algorithm
This algorithm [20] represents the solution manifold M as a union of the
projection of spherical balls. Each ball lives in the tangent space of a point
ui on M , and is represented by its radius Ri , an orthonormal basis for the
tangent space i , and a restricted Laguerre-Voronoi polyhedral k-cell Pi .
The initial condition is found by nding 0 , estimating the size of the ball
(cf. Sect. 3.2.3), and setting P0 to a cube slightly larger than the spherical
ball.
The boundary point is found by selecting a ball with a polyhedral vertex
that lies outside
its spherical ball. If the ratio of the radii of neighboring balls

is within 2 then the origin (the center of the ball) is inside the polyhedron.
Then a boundary point can be found by nding the intersection of a line
between the origin and the exterior vertex and the sphere.
The neighborhood of the boundary point is the projection of a spherical
ball in the tangent space at the boundary point onto M . The polyhedron for
the boundary point is initialized to a cube centered about the origin.
The merge requires nding neighboring balls and subtracting complementary half-spaces from their polyhedra. The half-spaces are found by projecting
each center into the tangent space of the neighbor and using a ball of the same
radius, but now in the other tangent space; see Figs. 3.31 and 3.32.

110

Michael E Henderson

Fig. 3.32. The authors covering algorithm in solution space. To update the polyhedra Pi the centers of neighboring neighborhoods are projected into the tangent
space at ui .

3.3 Performance of the Algorithms

We have tried to cast each of these algorithms as an iteration of the three basic
steps. Suppose that Mi is the part of the manifold that has been computed
up to step i. Then Mi+1 is found by selecting either a point or transverse
face ui that lies on the boundary of Mi , constructing a neighborhood N (ui )
of that point, and merging the neighborhood into Mi to get Mi+1 . The ve
algorithms dier in how they perform these steps, but there are two common
challenges that they must address.
The rst challenge is that all algorithms use a simplicial or cellular approximation to the manifold. To perform the merge the simplices in the neighborhood must be compatible with those in Mi , that is, they must either be disjoint
or have an intersection which is a face. Allgower and Schmidts pattern algorithm maintains compatibility by selecting simplices from a decomposition
of the entire region of interest, as does Rheinboldts moving-frame algorithm.
Therefore, compatibility is guaranteed for these two algorithms unless the
mapping from simplices to M becomes singular. Melville and Mackey do not
attempt to store a representation of Mi , but instead keep its boundary as a
set of polygons, which is then updated. Their approach is limited to k = 2,
although, in principle, a simplicial approximation could be maintained for the
boundary in higher dimensions. However, in that case the compatibility issue
arises again. Brodziks algorithm maintains a Delaunay triangulation on M

3 Higher-Dimensional Continuation

111

and deals with compatibility by removing points and recomputing the triangulation locally. Finally, the authors algorithm avoids the compatibility issue
by representing Mi as a set of overlapping neighborhoods. However, underneath this covering is a type of Voronoi triangulation of Mi , whose dual is a
Delaunay triangulation, which aids in nding boundary points. So it is possible to view the algorithm as having a merge but by way of a method that
is not easily described without the Voronoi cells.
The second challenge is to avoid computing part of the manifold more than
once. This means that the boundary must be maintained in some form or
other. In Allgower and Schmidts pattern algorithm care is taken not to pivot
to a simplex that is already in Mi . This is done by using a clever integer coding
of the simplices and keeping a coded list of simplices for Mi . Rheinboldts
moving-frame algorithm does not address the issue, but since there are a
nite number of boundary vertices in the reference k-dimensional complex,
the algorithm at least terminates and does not cycle. Melville and Mackey
explicitly represent the boundary and update the boundary of Mi to nd the
boundary of Mi+1 . Brodziks algorithm detects the overlap of simplices, but
simply removes one of the overlapping simplices to leave a gap. The authors
algorithm indirectly represents the boundary in terms of polyhedra in the
tangent space, which ensures that new points are within a given tolerance of
the boundary.
To illustrate how the ve algorithms actually perform we implemented
them all, except Brodziks, and applied them to the simple example of computing a sphere given by
F (u) = |u|2 1
for k = 2 and n = 3.
For these dimensions Allgower and Schmidts algorithm produces a decomposition of the sphere; see Fig. 3.33(a). Even though the simplices in R3 are
fairly uniform, their intersection with the sphere is not, so that the method
can be expected to produce small triangles.
As described above, for a rectangular mesh Rheinboldts moving-frame
algorithm covers parts of the sphere more than once. Although square cells
were used in the k-dimensional reference space, the projection to the sphere
results in clustering of vertices near the poles (the initial point and tangent
space dene an equator); see Fig. 3.33(b).
Melville and Mackeys algorithm is able to cover the sphere, and to avoid
covering it more than once. Each vertex obtained lies on a curve on the sphere,
namely on the intersection of the spherical neighborhood in R3 that is used to
dene a curve about the boundary point. In this case these curves are circular
arcs; see Fig. 3.33(c).
Brodziks algorithm not only leaves gaps, but the triangles on the sphere
are of a range of sizes; see Fig. 3.34.
The authors algorithm successfully covers the entire sphere, and produces
polygonal Voronoi regions on the sphere; see Fig. 3.33(d).

112

Michael E Henderson
(a)

(b)

(c)

(d)

Fig. 3.33. Comparison of dierent methods when applied to a sphere. Panel (a) is
the result of Allgower and Schmidts pattern algorithm with tetrahedra selected from
a decomposition of R3 with 5 20 20 20 = 40000 tetrahedra; the ve-tetrahedral
decomposition of the cube was used, with eight copies reected so that there are
no incompatibilities. Panel (b) is the result of Rheinboldts moving-frame algorithm
with a 50x50 mesh of size 0.1. Panel (c) is the result of Melville and Mackeys
algorithm, which produces a set of boundaries (curves) on the sphere, which in this
case are circular arcs. Panel (c) is the result of the authors algorithm with the
maximum radius of 0.1 and tolerance = 0.01.

The result of these computations is an approximation to a branched manifold, but no mention has been made yet of how to compute the singular
boundary manifolds or how to move from one branch to another. In [21] there
is a discussion of branch switching for manifolds that, although it is cast in
terms of the authors algorithm, could be applied to all of the algorithms except Allgower and Schmidts pattern algorithm. The idea is that moving from
a point on M to a point projected orthogonal to the tangent space is one step
of PSALC. Therefore, singular points may be detected as usual in PSALC.
Branch switching is a little more complicated. For simplicial continuation each

3 Higher-Dimensional Continuation
(a)

113

(b)

xc

Fig. 3.34. Brodziks algorithm applied to a sphere (with a larger simplex edge than
those above). Panel (a) shows the gap that is left when the neighborhoods of points
like xc are dropped; these overlap neighborhoods whose centers, as measured around
the boundary, are not close to xc . Panel (b) shows the same computation with a different tolerance, which allows triangles to be more skewed; this additional exibility
results in a smaller gap, but a lower quality triangulation. From M.L. Brodzik, The
computation of simplicial approximations of implicity dened p-dimensional manc 1998 by Elsevier Science;
ifolds, Computers Math. Applic. 36(6) (1998) 93113 
reprinted with permission.

of the n-dimensional faces of the simplices are checked for points on M , so

unless a singular point lies at a vertex of the complex, there is no diculty.

3.4 Conclusions
How a mathematical object is represented can aect the complexity of an
algorithm. The ve algorithms for higher-dimensional continuation of implicitly dened surfaces discussed here all use dierent representations of either
the solution manifold or the neighborhood of a boundary point. They draw
on techniques from algebraic topology, computational geometry, linear programming (the algorithm for subtracting a half-space from a polyhedron),
and dierential geometry. When the results are placed side by side, it can be
seen that all ve algorithms are a variation of constructing an approximately
spherical neighborhood of a point (or face) on the boundary, and merging the
neighborhood into the whole.
Lastly, a practical note. Some work has been done on visualizing solution
manifolds [27], but it is still not clear how one should visualize a 3-manifold
computed with any of these algorithms.

114

Michael E Henderson

Acknowledgments
The author is grateful to Elsevier Science for permission to reproduce Fig. 3.34.

References
1. E. L. Allgower and K. Georg. Simplicial and continuation methods for approximations, xed points and solutions to systems of equations. SIAM Review,
22:2885, 1980.
2. E. L. Allgower and S. Gnutzmann. An algorithm for piecewise-linear approximation of implicitly dened two-dimensional surfaces. SIAM J. Numer. Analysis,
24(2):452469, 1987.
3. E. L. Allgower and P. H. Schmidt. An algorithm for piecewise-linear approximation of an implicitly dened manifold. SIAM J. Numer. Analysis, 22(2):322346,
1985.
4. F. Aurenhammer. Power diagrams: properties, algorithms and applications.
SIAM J. Comp., 16(1):7896, 1987.
5. F. Aurenhammer. Improved algorithms for discs and balls using power diagrams.
J. Algor., 9:151161, 1988.
6. F. Aurenhammer and H. Edelsbrunner. An optimal algorithm for constructing
the weighted Voronoi diagram in the plane. Pattern Recognition, 17(2):251257,
1984.
7. F. Aurenhammer. Voronoi diagrams a survey of a fundamental geometric data
structure. ACM Comp. Surveys, 23(3):345405, 1991.
8. A. Bowyer. Computing Dirichlet tessellations. The Computer Journal,
24(2):162166, 1981.
9. M. L. Brodzik. The computation of simplicial approximations of implicity dened p-dimensional manifolds. Computers Math. Applic., 36(6):93113, 1998.
10. M. L. Brodzik and W. C. Rheinboldt. On the computation of simplicial approximations of implicitly dened two-dimensional manifolds. Computers Math.
Applic., 28(9):921, 1994.
11. W. Brostow, J.-P. Dussault, and B. L. Fox. Construction of Voronoi polyhedra.
J. Comp. Phys., 29:8192, 1978.
12. P.-C. Chen, P. Hansen, and B. Jaumard. On-line and o-line vertex enumeration
by adjacency lists. Oper. Res. Lett., 10:403409, October 1991.
13. D. P. Dobkin and M. J. Laszlo. Primitives for the manipulation of threedimensional subdivisions. Algorithmica, 4:332, 1989.
14. J. L. Dupont. Curvature and Charateristic Classes. Lecture Notes in Mathematics 640. (SpringerVerlag, Berlin, 1978).
15. S. Fortune. Voronoi diagrams and Delaunay triangulations. In D.-Z. Du and
F. K. Hwang, editors, Computing in Euclidean Geometry, pages 193233. (World
Scientic Publishing, Singapore, 1992).
16. P. W. Gross and P. R. Kotiuga. Data structures for geometric and topological
aspects of nite element algorithms. Progr. Electromagn. Res. PIER 32, 32:151
169, 2001.
17. B. Grunbaum. Convex Polytopes. (Interscience Publishers, London, 1967).

3 Higher-Dimensional Continuation

115

18. L. Guibas and J. Stol. Primitives for the manipulation of general subdivisions
and the computation of Voronoi diagrams. ACM Transactions on Graphics,
4(2):74123, April 1985.
19. P. M. Hanrahan. Creating volume models from edgevertex graphs. ACM
SIGGRAPH Comp. Graph., 16(3):7784, July 1982.
20. M. E. Henderson. Multiple parameter continuation: Computing implicitly dened kmanifolds. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 12(3):451476,
2002.
21. M. E. Henderson. Multiparameter parallel search branch switching. Internat.
J. Bifur. Chaos Appl. Sci. Engrg., 15(3):967974, 2005.
22. H. Imai, M. Iri, and K. Murota. Voronoi diagram in the Laguerre geometry and
its applications. SIAM J. Comp., 14(1):93105, 1985.
23. H. B. Keller. Numerical solutions of bifurcation and nonlinear eigenvalue problems. In Paul Rabinowitz, editor, Applications of Bifurcation Theory, pages
359384. (Academic Press, New York, 1977).
24. R. Klein. Concrete and Abstract Voronoi Diagrams, volume 400 of Lecture Notes
in Computer Science. (Springer-Verlag, Berlin, 1987).
25. H. W. Kuhn. Some combinatorial lemmas in topology. IBM J. Res. and Develop., 45(5):518524, 1960.
26. S. Lefschetz. Introduction to Topology. (Princeton University Press, 1949).
27. J. H. Maddocks, R. S. Manning, R. C. Paenroth, K. A. Rogers, and
J. A. Warner. Interactive computation, parameter continuation, and visualization. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 7:16991715, 1997.
28. W. S. Massey. Algebraic Topology, An Introduction. (Harcourt, Brace & World,
New York, 1967).
29. R. Melville and D. S. Mackey. New algorithm for two-dimensional numerical
continuation. Computers Math. Applic., 30(1):3146, 1995.
30. V. T. Rajan. Optimality of the Delaunay triangulation in R d . In Proceedings
of the seventh annual symposium on Computational geometry, pages 357363.
(ACM Press, New York, 1991).
31. W. C. Rheinboldt. On a moving-frame algorithm and the triangulation of equilibrium manifolds. In T. K
upper, R. Seydel, and H. Troger, editors, ISNM79: Bifurcation: Analysis, Algorithms, Applications, pages 256267. (Birkh
auser Verlag, Basel, 1987).
32. W. C. Rheinboldt. On the computation of multi-dimensional solution manifolds
of parameterized equations. Numer. Math., 53(1/2):165181, July 1988.
33. J. R. Shewchuk. A condition guaranteeing the existence of higher-dimensional
constrained delaunay triangulations. In Proceedings of the Fourteenth Annual
Symposium on Computational Geometry, pages 7685, (ACM Press, New York,
1998).
34. D. F. Watson. Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes. The Computer Journal, 24(2):167172, 1981.
35. H. Whitney. Geometric Integration Theory. (Princeton University Press, 1957).

4
Computing Invariant Manifolds via the
Continuation of Orbit Segments
Bernd Krauskopf and Hinke M Osinga
Department of Engineering Mathematics, University of Bristol, United Kingdom

A key feature of packages such as Auto [12, 15], Content [24] and Matcont
[11] is a collocation solver for two-point boundary value problems (BVPs); see
also Chaps. 1 and 2. In conjunction with pseudo-arclength continuation, it
is possible to nd the solution of a two-point BVP and then continue it in
parameters. This basic idea will be known to most readers as the standard
technique to compute a one-parameter branch of periodic orbits. However, the
continuation of BVPs is a much more versatile tool and the solution need not
be a periodic orbit, but may be any specied orbit segment. For example, the
continuation of a suitable orbit segment is utilized in the HomCont extension
for the computation of connecting orbits; see [7] and also [13, 21, 43].
In this chapter we focus on the idea of representing an invariant global
manifold of a dynamical system as a family of orbit segments, which can
then be computed as a solution family of a suitable BVP. Note that the thus
computed object lies entirely in the phase space, rather than the product
of phase space and parameter space. More specically, we consider an ndimensional autonomous vector eld
x = f (x, )

(4.1)

where x Rn and f : Rn Rm Rn is suciently smooth. The multidimensional parameter Rm remains xed in most methods presented here,
in which case we drop the dependence of f on for notational convenience.
The vector eld (4.1) induces a ow t on Rn that determines the dynamics.
The global dynamics of (4.1) is determined by its equilibria, periodic orbits
and invariant tori, together with their global stable and unstable manifolds.
Equilibria and their stability properties can often be found analytically. However, since t is rarely known explicitly, the task of nding periodic orbits,
invariant tori and global manifolds generally requires the use of numerical
techniques. We focus on the computation of an invariant manifold of (4.1)
that can be viewed as a one-parameter family of orbit segments. In other
words, in terms of the underlying vector eld, we consider two-dimensional

118

Bernd Krauskopf and Hinke M Osinga

manifolds. This allows us to make use of the boundary value solver in Auto
[12, 15] to continue the respective orbit segments in order to build up the
manifold. Note that the same set-up can be used for higher-dimensional manifolds, but then one needs to employ multi-parameter continuation methods;
see [2, 29] and Chap. 3.
Let us assume that (4.1) has an equilibrium x0 , that is, f (x0 ) = 0. We
further suppose that x0 is a hyperbolic saddle, which means that the Jacobian
matrix Df (x0 ) at x0 has eigenvalues in both the open left-half and the open
right-half of the complex plane, but not on the imaginary axis. The stable
manifold W s (x0 ) associated with x0 has the property that all orbits contained
in W s (x0 ) tend to x0 in forward time:

W s (x0 ) = x Rn | t (x) x0 as t .
Similarly, the unstable manifold W u (x0 ) is dened as the set of all orbits that
tend to x0 in backward time:

W u (x0 ) = x Rn | t (x) x0 as t .
The Stable Manifold Theorem [45] guarantees the existence of global stable
and unstable (immersed) manifolds that are as smooth as f . Furthermore,
the dimension of the (un)stable manifold is equal to the number of (un)stable
eigenvalues.
Equivalent notions of stable and unstable manifolds exist if (4.1) has a
saddle periodic orbit with period T . That is, has Floquet multipliers both
inside and outside the unit circle of the complex plane, and only the trivial
Floquet multiplier (associated with the direction tangent to ) on the unit
circle. As before, the Stable Manifold Theorem [45] guarantees the existence
of global stable and unstable (immersed) manifolds W s ( ) and W u ( ) that
are as smooth as f . They are dened as the set of orbits that tend to in
forward and backward time, respectively:

W s ( ) = x Rn | t (x) as t
and

W u ( ) = x Rn | t (x) as t .

Note that the dimensions of W s ( ) and W u ( ) are equal to one plus the
number of stable and unstable Floquet multipliers, respectively.
This chapter is organized as follows. We rst consider in Sect. 4.1 the computation of the one-dimensional intersections of a two-dimensional manifold
with a chosen Poincare section. As we show in Sect. 4.2, our set-up is particularly useful in the context of slow-fast systems and even allows us to compute
slow manifolds near a folded node. We then discuss in Sect. 4.3 how one can
compute two-dimensional global manifolds as a family of orbit segments. In
Sect. 4.3.3 we explain how one can compute two-dimensional invariant tori
with quasiperiodic dynamics in this set-up. We end with conclusions and an
outlook to future research in Sect. 4.4.

4 Computing Invariant Manifolds

119

4.1 One-Dimensional Global Manifolds in a Poincar

e
Section
The Poincare map of an n-dimensional vector eld as given by (4.1) is dened
on an (n 1)-dimensional section Rn . The image P (x) of a point x
under the Poincare map P is then dened as the next (or kth for some integer
k > 0) intersection of the orbit through x with . Considering a Poincare
map is a standard tool to study the properties of a periodic orbit of a
vector eld. The section is chosen transverse to and P is dened such
that an (isolated) intersection point 0 of with is a xed point, that is,
P (0 ) = 0 . Then locally near 0 the map P is a dieomorphism, so that the
dynamics of P on near 0 describes the local dynamics near ; see, for
example, [27, 41, 53].
Suppose now that the vector eld (4.1) has a saddle periodic orbit .
To keep this exposition simple, we assume that the dimension of the phase
space is n = 3. In this case the periodic orbit has two-dimensional stable
and unstable manifolds W s ( ) and W u ( ), respectively. Furthermore, we
consider the standard situation that the Poincare section is a suitable twodimensional plane transverse to , so that there exists a point 0
. When properly dened in a neighborhood of 0 as the kth return to ,
the Poincare map P is a dieomorphism and 0 is a saddle xed point of
P . As a consequence, the Stable Manifold Theorem guarantees the existence
of one-dimensional smooth (un)stable manifolds W s (0 ) and W u (0 ) in the
region where P is a dieomorphism. Moreover, globally we can view W s (0 )
and W u (0 ) as the intersections of the corresponding stable and unstable
manifolds of . That is,
W s (0 ) = W s ( )
and
W u (0 ) = W u ( ) .
Note that this denition has the advantage of being very geometrical. It follows from the properties of the manifolds W s ( ) and W u ( ) in R3 that the
(generalized) global manifolds W s (0 ) and W u (0 ) exist and are a union of
manifolds in .
Indeed, W s (0 ) and W u (0 ) are not necessarily single connected curves,
because P is typically not a global dieomorphism. In general, P is discontinuous at points where the ow is tangent to and we dene the discontinuity
locus C as
(4.2)
C := {x | f (x) nx = 0} ,
where nx is the normal to at x and the dot denotes the inner product in R3 .
(If is a hyperplane then nx = n is independent of x.) The complement of C
consists of open regions of . In the region that contains 0 , we can dene P
in a continuous and smooth manner and W s (0 ) and W u (0 ) are well dened

120

Bernd Krauskopf and Hinke M Osinga

as the global manifolds of the xed point 0 of the map. The intersections of
W s ( ) and W u ( ) with become more interesting as soon as W s (0 ) and
W u (0 ) cross the discontinuity locus C.
Let us consider a simple example, where the three-dimensional vector eld
has a saddle periodic orbit and we can choose a two-dimensional planar
Poincare section that intersects transversely at exactly two points, say,
0 and 1 . Near both 0 and 1 one can dene the Poincare map P as a
local dieomorphism (the second return to ). In the simplest case, the set
C is a single smooth curve that divides the plane into the two regions of
local denition of P . In this situation, the stable and unstable manifolds of
0 and 1 in may cross C. One possibility is that a branch of W s (0 ) coincides with a branch of W s (1 ), eectively connecting 0 and 1 ; see already
Fig. 4.2(c). This corresponds in the three-dimensional phase space to the case
that intersects one side of W s ( ) in a single curve. There are other, more
complicated possibilities of how may intersect W s ( ). For example, the
intersection in may consist of a closed curve that corresponds to a piece of
W s (p0 ) that is not connected to any saddle; see already Fig. 4.4.
4.1.1 The ManBVP Method
From an algorithmic point of view, it is not a good idea to compute W s (0 )
and W u (0 ) by nding W s ( ) or W u ( ) rst and then determine their intersection with . Rather we will nd W s (0 ) and W u (0 ) directly as subsets of
by using what we call the ManBVP method. The key idea is to consider
the manifolds as one-parameter families of solutions to two-point boundary
value problems. Our method as reviewed here was introduced in [16], where
more details can be found.
The main essence of the ManBVP method is that one should think of the
Poincare map not only as assigning P (x), but as assigning the entire orbit
segment
{t (x) | 0 t tx }.
Here tx is the appropriate return time to the section , which depends continuously on x. In particular, for 0 we have P (0 ) = 0 and t0 = T
is the period of . The orbit segment that constitutes the action of P on x
is the solution u of the two-point boundary value problem that solves (4.1)
subject to the boundary conditions u(0) = x and u(tx ) .
The ManBVP method has the key advantage that the locus C is no
longer a discontinuity boundary in the space of boundary value problems. It
is explained here for the computation of a one-dimensional unstable manifold
W u (0 ) of a saddle xed point 0 that is associated with a saddle periodic orbit of (4.1). A one-dimensional stable manifold can be computed
in exactly the same way by reversing time. We assume that the single unstable Floquet multiplier u of is positive, i.e. u > 1, so that the associated
Poincare map P is orientation preserving on W u (0 ); if u < 1, then must
be covered twice to dene the image of 0 under P , that is, t0 = 2 T .

4 Computing Invariant Manifolds

0 = 0
1
2
3

121

k+1
k
k
i1

k1

uk+1 (t)
uk (t)
Fig. 4.1. Graphical illustration of the ManBVP algorithm.

The ManBVP method approximates W u (0 ) as an ordered list of mesh

points M = {0 , 1 , . . . , N } up to a prescribed arclength L. The computation
starts from the saddle periodic orbit , which we nd by continuation with
Auto [12, 15]. The periodic orbit is viewed as an orbit segment that denes
the image of 0 under the Poincare map P . We continue this orbit segment
in Auto [12, 15] not as a periodic orbit but as an orbit with both boundary
points in . The rst boundary point of the orbit segment is initially varied
along the unstable eigenvector of 0 in . In this way, the other boundary
point of the orbit segment begins to trace out an approximation of the initial
piece of W u (0 ). We then allow the rst boundary point to vary along this
initial piece and subsequently computed parts of the manifold, so that the
other boundary point continues to trace out more and more of the unstable
manifold.
Figure 4.1 gives an impression of how the ManBVP method works. The
variation of the rst boundary point along the computed piece is done using
a piecewise-linear approximation of W u (0 ) between mesh points in M . The
distance between consecutive points k M varies depending on the
local curvature of W u (0 ); in this sense, the method can be seen as an adaptation and renement of the method in [36] to the specic context of general
Poincare maps.
Set-Up of the Two-Point Boundary Value Problem
To set up a two-point boundary value problem that has solutions with end
points on W u (0 ) we look for solutions of the system
u (t) = T f (u(t)),
with boundary conditions

(4.3)

122

Bernd Krauskopf and Hinke M Osinga

u(0) ,

(4.4)

u(1) .

(4.5)

Note that (4.3) is in a scaled form so that all orbit segments have total time
1. Hence, the orbit segment {t (x) | 0 t tx } associated with the image
P (x) of a point x is rescaled as {u(t/tx ) | 0 t 1} and the period
T = tx is now a parameter in the system.
We obtain a one-parameter family of solutions to (4.3)(4.5) by letting
u(0) vary along the one-dimensional piecewise-linear approximation of (a rst
piece of) W u (0 ) instead of the two-dimensional space . To this end, we let
Li ( ) = (1 ) i1 + i ,

0 1,

denote the (parametrized) line segment between the already computed mesh
points i1 and i . Then boundary condition (4.4) is replaced by
u(0) Li ( ) = 0,

(4.6)

where the index i varies as the computation progresses. Note that (4.6) automatically ensures that u(0) in case is a hyperplane. For non-planar
this introduces an error that is of the same order as the mesh error on M ; we
refer to [16] for details.
In order to detect during a computation that the end of the line segment
Li ( ) has been reached we introduce the user-dened functions
UZ(0) = ,
UZ(1) = 1.
When such an event is detected the algorithm switches either to the next
or the previous line segment. Indeed, it is necessary to allow switches to the
previous line segment, that is, to let u(0) trace W u (0 ) backwards toward 0 ;
namely, this occurs as soon as W u (0 ) has crossed the discontinuity locus C.
The selection of the mesh points in M is based on the same accuracy conditions as the method for dieomorphisms in [36]. To decide when to generate
the next point k+1 we calculate the distance
k :=|| u(1) k ||
between the end boundary point and the last computed mesh point, and the
angle
k := (k1 , k , u(1))
between the last two computed mesh points and the end boundary point during the continuation. We then monitor the user-dened accuracy conditions
UZ(2) = max k ,
UZ(3) = ()max k k ,
UZ(4) = min k

4 Computing Invariant Manifolds

123

where max , ()max , and min are prespecied bounds set by the user; for
the examples below and in Sect. 4.2.1 we used max = 0.3, ()max = 105 ,
and min = 104 . Note that the above conditions minimize the number of
mesh points required for achieving the accuracy conditions, so that the mesh
selection with respect to the curvature of the manifold is more rened here
than that in [36].
The rst segment L1 ( ) is dened by the line through 0 in in the
eigendirection associated with the unstable Floquet multiplier u ; see [16] for
details on how to nd this direction. We assume that = 1 when a prespecied
maximum distance along this initial segment is reached. Indeed, this initial
distance has an important inuence on the overall error of the approximation
to W u (0 ); see [36]. Starting from the scaled orbit that represents , we
continue the one-parameter family of solutions along L1 ( ) while monitoring
the distance
1 =|| u(1) u(0) ||
and the angle
1 = (0 , u(0), u(1)).
When either of the accuracy conditions are met then the continuation stops
and we set 1 = u(0) and 2 = u(1). We then continue the computation
along the line segment L2 ( ). Hence, the initial line segment between 0 and
1 may, in fact, be shorter than the prespecied maximum distance .
During the computation we also record the orbit segments uk that are
used to nd the points k = uk (1) of M . These orbit segments give a good
impression of the relevant part of the two-dimensional manifold W u ( ) in the
full phase space and provide further insight into the geometry of the manifolds.
Semiconductor Laser with Optical Injection
In this section we illustrate the performance of the ManBVP method by computing the stable and unstable manifolds of a model for an optically injected
semiconductor laser. Manifolds in this system were also considered in [9] and
we show here how the ManBVP method is able to compute a one-dimensional
manifold across the discontinuity boundary C. The laser system is modeled
by the so-called rate equations



E = K + 12 (1 + i) n i E ,
(4.7)


n = 2n (1 + 2Bn) |E|2 1 ,

and is well known to have rich dynamics; see Chap. 6 and [58]. Here,
E = (Ex , Ey ) is the complex electric eld and n is the population inversion (the number of electron-hole pairs). The two main parameters are the
injected eld strength K, and the detuning between the frequency of the
free-running laser and the injected frequency. The material properties of the
laser are described by the values of , B and .

124

Bernd Krauskopf and Hinke M Osinga

We use the parameter values = 2, B = 0.015, = 0.035, and choose

= 0.270 and K = 0.290, as in [9]. We study the Poincare map P on the
section given by the plane = {(E, n) | n = 0}. For this choice of parameters
a saddle periodic orbit intersects at four points denoted 0 , 1 , 2 , and 3
in consecutive order. The discontinuity boundary C, where the ow is tangent
to , is given by n = 0 |E| = 1 and n > 0 for points with |E| < 1, whereas
n < 0 for points with |E| > 1. Figure 4.2 gives an overview of the computations
of W s ( ) (black curves). The intersection points of the periodic orbit are
indicated by crosses. Panel (a) also shows a period-four sink (triangles) and
how W u ( ) intersects (grey curves). Since the manifold W u ( ) does not
interact with C, it is of less interest here.
Figure 4.2(b) shows the branches of stable manifolds that spiral in toward
singular points on . Notice that the two branches starting from 1 and 3
never interact with C. This means that the associated Poincare map is a
dieomorphism along the entire branch, and there is no particular diculty
to calculate these branches. On the other hand, the two branches starting
from 0 and 2 cross C several times. A similar observation can be made
in Fig. 4.2(c) and we discuss here the behavior of the branch of W s (0 ) in
more detail. Figure 4.2(d) shows two branches of the stable manifold that are
disconnected from the four saddle points, which we also discuss in more detail
below.
Let us rst focus on the left branch of W s (0 ) as an example of a branch
that crosses C. This branch of W s (0 ) is shown in Fig. 4.2(c) and the crossing
through C is more clearly illustrated in Fig. 4.3. Each row in Fig. 4.3 shows
the orbit segments used in the computation of an initial piece of W s (0 )
together with the time prole of the last orbit segment computed for this
part. The black dots in the time proles are the mesh points used by Auto
[12, 15] and the green line is . As can be seen in Fig. 4.3(a1), the boundary
points u(0) of the orbit segments all lie near 0 . Since time is reversed to
compute a stable manifold, the orbit segments rst ow downward from .
They next intersect near 3 and so on until they return to for the fourth
time. The boundary points u(1) are added to the mesh M representing the
initial piece of W s (0 ). The time prole in Fig. 4.3(a2) is representative for
all orbit segments in Fig. 4.3(a1) and shows three intersections with (green
line) between u(0) and u(1); note that the end of the orbit segment is owing
downward into the section. Row (b) in Fig. 4.3 shows the computation exactly
up to the point where W s (0 ) intersects C. The time prole in panel (b2)
shows that the last computed orbit segment still has three intersections with
the section between u(0) and u(1), but now the orbit segment is tangent to
at u(1). As we continue to compute W s (0 ) past C, the ends of the added
orbit segments now ow upward into . Row (c) of Fig. 4.3 shows this change
in behavior between orbit segments before and past C. The time prole in
panel (c2) shows that there are now four intersections of the last computed
orbit segment with between u(0) and u(1). This orbit segment is again
representative for all orbit segments that end on the other side of C. As we

4 Computing Invariant Manifolds

Ey 2 (a)

1.5

Wu

0.5

0.5

-0.5

-1

-1

Ws

-2

-2

-2

-1.5

-0.5

-1

0.5

1.5

Ex

-2.5
-2.5

-2

-1.5

1.5

0.5

1.5

Ex

Ey (d)

1.5

-0.5

-1

Ey (c)

0.5

0.5

-0.5

-0.5

-1

-1

-1.5

-1.5

Ws

Ws

-2

-2

-2.5
-2.5

-1.5

Ws

-1.5

-0.5

-2.5
-2.5

Ey (b)

1.5

125

-2

-1.5

-1

-0.5

0.5

1.5

Ex

-2.5
-2.5

-2

-1.5

-1

-0.5

0.5

1.5

Ex

Fig. 4.2. The stable and unstable manifolds, as computed by the ManBVP algorithm, of the four-periodic saddle {0 , 1 , 2 , 3 } (crosses) of (4.7) in = {n = 0}
with = 2, B = 0.015, = 0.035, = 0.270, and K = 0.290. The ow is tangent to along the unit circle labeled C. Panel (a) shows all computed parts of
the stable and unstable manifolds of {0 , 1 , 2 , 3 }. Panel (b) highlights branches
of stable manifolds that spiral into singular points, panel (c) those that join two
of the saddle points, and panel (d) two disjoint pieces of manifold. From J.P. England, B. Krauskopf and H.M. Osinga, Computing one-dimensional global manifolds
of Poincare maps by continuation, SIAM J. Appl. Dyn. Sys. 4(4) (2005) 10081041
c 2005 by the Society for Industrial and Applied Mathematics; reprinted with per
mission.

already observed in Fig. 4.2(c), W s (0 ) continues until it reaches 3 . Indeed,

since 3 is the next intersection of when following backward in time from
0 , the extra intersection picked up once u(1) crosses C makes this possible.
Note that the same branch can also be computed as one of the branches of

126

Bernd Krauskopf and Hinke M Osinga

(a2)

(a1)
n

1
0

2
2

1
2
0

0.2

0.4

0.6

0.8

Ey

1
2

1.5

0.5

0.5

1.5

(b2)

1
0

2
2

1
2
0

0.2

0.4

0.6

0.8

Ey

1
2

1.5

0.5

0.5

1.5

t/T

Ex
(c1)

Ex
(b1)

t/T

(c2)

2
2

Ey

1
2

1.5

0.5

0.5

1.5

0.2

0.4

0.6

0.8

t/T

Ex

Fig. 4.3. A demonstration of how W s (0 ) of system (4.7) is computed across C.

Rows (a)-(c) show the computations before, at, and after the intersection with C,
respectively. Panels in the left column show orbit segments on W s ( ) used to compute W s (0 ). The time proles of the last orbit segment computed in each case are
shown in the right column (blue line), where the black dots indicate the mesh points
and is the green line. From J.P. England, B. Krauskopf and H.M. Osinga, Computing one-dimensional global manifolds of Poincare maps by continuation, SIAM
c 2005 by the Society for Industrial and
J. Appl. Dyn. Sys. 4(4) (2005) 10081041 
Applied Mathematics; reprinted with permission.

4 Computing Invariant Manifolds

(a)

Ey

127

-0.5

-1

-1.5

0.5

-0.5

-1

Ex

1.5

(b)

1
0.75

n
0.5
0.25
0

0.25

C
W s (1 )

0.5
0.75
1
1
0

Ey

Ex

0
2

Fig. 4.4. An enlargement of Fig. 4.2(d) in panel (a) shows a disjoint intersection
of W s ( ) with (blue). The colored segment of W s (1 ) maps to this disjoint
piece. The orbit segments used in the computation are shown in panel (b). From
J.P. England, B. Krauskopf and H.M. Osinga, Computing one-dimensional global
manifolds of Poincare maps by continuation, SIAM J. Appl. Dyn. Sys. 4(4) (2005)
c 2005 by the Society for Industrial and Applied Mathematics; reprinted
10081041 
with permission.

128

Bernd Krauskopf and Hinke M Osinga

W s (3 ), where the computation starts out with orbit segments that intersect
three times in between u(0) and u(1), and which become orbit segments
that intersect only two times in between the boundary points as soon as C
is crossed.
Figure 4.4(a) shows an enlarged view of Fig. 4.2(d). The blue branch that
spirals around C is not connected to any of the intersection points of .
However, all points on this disjoint segment lie on orbits that intersect on
W s (1 ). This part of W s (1 ) is also colored blue in Fig. 4.4(a). Figure 4.4(b)
shows the orbits segments used in the computation of this disjoint part. All
orbit segments have u(0) on W s (1 ) and this part of W s (1 ) is traversed back
and forth several times as u(1) crosses C. The segment traced out by u(1) is
an isolated submanifold that is due to the way W s ( ) intersects . We are
able to compute this disjoint piece by choosing an already computed point on
W s (1 ) and integrating it backward in time until it intersects again. We
then correct the solution such that it satises the boundary conditions and
use the ManBVP method to trace the entire isolated branch.

4.2 Global Manifolds of Slow-Fast Systems

In its simplest form, a slow-fast system can be written as

x = g(x, y, ),
y = f (x, y, ),

(4.8)

where (x, y) Rn (in this section we again consider the case n = 3) and
determines the separation of slow and fast time scales. It is well known that a
slow-fast system of the form (4.8) with  1 typically displays extreme sensitivity to the initial condition [33]. However, the inherent accuracy of solving
two-point boundary problems (for example, with Autos collocation routines
as in our case) allows one to deal eciently with the problem of a possible sensitivity of the initial value problem. As a consequence, the ManBVP method
is able to compute global manifolds reliably also in the context of systems
with multiple time scales. This is demonstrated below in Sect. 4.2.1 where we
compute stable and unstable manifolds in a Poincare section for a Van der
Pol-Dung oscillator modeled by Koper [4, 35]. As is shown in Sect. 4.2.2, a
slight variation of the ManBVP method can be used to compute slow manifolds and their intersections, which are known as maximal canard orbits; see
also Chap. 8.
4.2.1 The Koper Model
We consider the Van der Pol-Dung oscillator [4, 18, 35] as modeled by the
equations

4 Computing Invariant Manifolds

-0.6

129

(a)

-0.65

-0.7

-0.75

Ws

Wu

-0.8

-0.85

-0.9

-0.95

-1.5

-2

-0.5

-1

(b)

0.81

z
0.83

0.85

0.87

0.89
2
1

0
1

0.9

0.8

0.7

0.6

Fig. 4.5. The stable (blue) and unstable (red) manifolds of (4.9) with = 0.1,
k = 22.5, and = 18, computed in the section z = 0.83 using the ManBVP
method. The points 0 and 1 are the two intersections of a saddle periodic orbit
with . The orbit segments used in the computation are shown in panel (b). From
J.P. England, B. Krauskopf and H.M. Osinga, Computing one-dimensional global
manifolds of Poincare maps by continuation, SIAM J. Appl. Dyn. Sys. 4(4) (2005)
c 2005 by the Society for Industrial and Applied Mathematics; reprinted
10081041 
with permission.

130

Bernd Krauskopf and Hinke M Osinga

x = ky x3 + 3x ,
y = x 2y + z,

z = y z,

(4.9)

where determines the dierence between two time scales. System (4.9), which
is often referred to as the Koper model, is also one of the examples in Chap. 2;
note that, while a slow-fast system can be written in dierent ways (on the
slow or the fast time scale) its implementation in Auto is always the same due
to the introduction of the time variable T . We take = 0.1, such that the xvariable evolves on a time scale ten times faster than y and z. A saddle periodic
orbit exists for k = 22.5 and = 18, together with an attracting orbit of
roughly twice the period. We choose the plane = {(x, y, z) | z = 0.83},
which intersects at two points 0 and 1 . The ow of (4.9) is tangent to
along the line C = {y = z = 0.83}; z > 0 for points with y < 0.83 and
z < 0 for points with y > 0.83.
Figure 4.5(a) shows the stable and unstable manifolds of 0 and 1 in ,
along with the computed orbit segments in panel (b). The saddles 0 and 1
are indicated by green crosses. The period-doubled attractor intersects only
twice, at the points indicated by blue squares. One branch of W u (0 ) and one
branch of W s (0 ) cross C and connect to 1 . The other branch of W u (0 )
tends to one of the sinks, as does one branch of W u (1 ). The other branch
of W s (0 ) tends to . Finally, the branch of W s (1 ) that does not connect
to 0 spirals in toward a special point, namely the intersection point of the
one-dimensional stable manifold of a saddle equilibrium located below ; see
the green point below in panel (b).
It is particularly challenging to compute W s (0 ) and W s (1 ). Not only
are the Floquet multipliers u and s of negative, so that we need to use a
double covering, but s is also very strongly contracting. Namely, u 3.03
and s 9.25 104 . Hence, the contraction along the stable manifold is
1/(s )2 = O(106 ). It is precisely this dierence in scales that makes it impossible to compute even the Poincare map by solving an initial value problem.
However, the ManBVP method is able to compute both manifolds of 0 and
1 in ; indeed, u(0) varies only on the order of O(109 ) during the computation of the entire branch of W s (0 ) that connects to 1 .
As Fig. 4.5(b) shows, many orbit segments are needed during the computation. In particular, we are eectively computing the part of the twodimensional manifold W s ( ) that below accumulates on the segment of
the stable manifold of the (green) saddle equilibrium.
4.2.2 Slow Manifolds and Canards Near a Folded Node
In this section we compute slow manifolds associated with a singularity that
is known as a folded node of a slow-fast system in R3 . Suppose that the slow
variable x in (4.8) is two-dimensional (and, hence, y R) and that the critical
manifold S := {(x, y) R3 ; f (x, y, 0) = 0} is a nondegenerate folded surface

4 Computing Invariant Manifolds

131

in a neighborhood of the origin. Hence, the associated slow subsystem is singular along the fold curve. Such a folded critical manifold is the generic case
where Fenichel theory [20, 33] cannot be applied because the required normal
hyperbolicity breaks down along the fold curve. In particular, this situation
creates the interesting phenomenon (especially for = 0) that the two sheets
of the associated slow manifold, one of which is locally attracting or stable and
one locally repelling or unstable, may intersect along special orbits known as
canard solutions. A canard solution has the unusual property that it follows
the unstable sheet for a certain amount of time before being repelled; see also
Chap. 8 and, for example, [5, 26, 54].
A folded node is a singular point on the fold line that corresponds to a
node equilibrium of thedesingularized slow subsystem (where time has been
rescaled by the factor ) [5, 26, 54, 57]. A blow-up procedure, followed by
the restriction to a specic chart on the blown-up locus and desingularization
lead to a normal form near the folded node. Following [57], the normal form
can be written as the vector eld

1
x = 2 y ( + 1)z + O( ),
(4.10)
y = 1,

2
z = x + z + O( ).

Here is the ratio of the strong and the weak eigenvalue of the folded singularity of the reduced ow; the critical manifold of the original system is
represented by the set {x + z 2 = 0}.
We consider here the so-called maximal canards near the folded node,
which correspond to intersections of the stable sheet C and the unstable
sheet C + of the slow manifold for = 0 in (4.10). To this end, we compute
the sheet C as a family of orbit segments with a slight modication of the
ManBVP method. Note that the sheet C + is the image of C under the
symmetry (x, y, z, t)  (x, y, z, t).
We consider orbit segments between the x-section = {x = } and the
y-section = {y = } where and are suitable constants. The manifold C
is a smooth perturbation of the stable sheet of the critical manifold (where
z is negative) away from the fold curve (at z = 0) where Fenichel theory
fails. Hence, for suciently large the intersection curve of the sheet C
with the x-section
is approximated well by the one-dimensional line L =
{(x, z) = (, )}; in our computations we used = 100 throughout. The
idea is now to continue the solution of the family of orbit segments u with
u(0) L and u(1) , while monitoring the accuracy condition in terms
of the curve traced out by u(1) in as is explained in Sect. 4.1.1. In other
words, loosely speaking C is the unstable manifold of the line L. As starting
2
condition we use the explicit solution s (t) = ( 4 t2 + 2 , t, 2 t), which is
known as the strong maximal canard; there is a second explicit solution given
by w (t) = ( 14 t2 + 12 , t, 12 t), which is known as the weak maximal canard; see
[57].

132

Bernd Krauskopf and Hinke M Osinga

.
3

(a)

z (b)

C+

1.5

1.5

0.8

3
8

C
4

C
s

y
(c)

C+

s
2

0.8

w 3 4
1

Fig. 4.6. The slow manifolds C and C + of system (4.10) for = 14.5. Panel (a)
shows C up to the y-section 0.8 , the inset panel (b) shows the intersection curves
of C and C + in the y-section 0.0 , and panel (c) shows how the surfaces C and
C + between the y-sections 0.8 and 0.8 intersect in maximal canards s , w and
1 to 4 .

Figure 4.6 shows the slow manifolds C and C + of (4.10) for = 14.5.
Panel (a) gives a global overview of the two-dimensional slow manifold C
(red surface) computed up to the y-section 0.8 (green). This surface was
rendered from the orbit segments that were used in the computation of the

4 Computing Invariant Manifolds

133

curve C 0.8 . The chosen viewpoint emphasizes that C is close to the

critical manifold {x = z 2 } only for large negative x. Indeed near x = 0
the manifold C starts to spiral and develops several folds. Also shown are
the explicitly known maximal canards s and w . Figure 4.6(b) shows the
intersections with the y-section 0.0 of the slow manifolds C (red curve)
and C + (blue curve); compare with [57]. The intersection points between
both curves are maximal canards, of which the strong maximal canard s is
highlighted. Figure 4.6(c) presents an enlarged view of the geometry of C
(red surface) and C + (blue surface) between the y-sections 0.8 and 0.8 .
The manifolds C and C + intersect along the maximal canards s and w ,
as well as the additional maximal canards 1 to 4 (black curves). Notice how
the maximal canards 1 to 4 spiral around the weak maximal canard w . For
a more detailed study of maximal canards in (4.10) and their dependence on
the parameter we refer to [10].

4.3 Two-Dimensional Global Manifolds of Vector Fields

In the previous sections we computed one-dimensional manifolds in a Poincare
section. Almost as a by-product, the orbit segments that are used in the
computation give an impression of a part of the associated two-dimensional
manifold of the underlying vector eld.
In this section, we show how one can use the idea of continuing a family
of orbit segments to nd enough and suitable orbit segments such that an
entire rst part of a two-dimensional manifold is computed. In contrast to
how we selected a mesh, and corresponding orbit segments, in the previous
section, we are now interested in obtaining a good representation of the entire two-dimensional object, and not just of a one-dimensional intersection
curve. Such a continuation of orbit segments, with a number of choices for
the boundary condition as detailed below, can be readily implemented in the
package Auto [12, 15]. While in general the resulting mesh has quite a lot
of mesh points, the computations are very accurate and remarkably fast; see
[39, Sec. 3] for a general overview of this method. The key property of the
continuation procedure is that the step size measures the change of the entire
computed orbit segment (and various parameters), and not just the change at
one of the end points. This means that, generally, the computation produces
a reasonable distribution of orbit segments along the manifold.
We rst explain and demonstrate the method by computing the unstable
manifolds of the secondary saddle equilibria or saddle periodic orbits in the
Lorenz system [42]. We then discuss briey how the basic idea of solving
suitable boundary value problems plays an important role in our own methods
for growing global (un)stable manifolds as a family of geodesic level sets.
Finally, we show how a suitable family of orbit segments can be used in a
scheme to nd and follow a quasiperiodic invariant torus.

134

Bernd Krauskopf and Hinke M Osinga

4.3.1 A 2D Global Manifold as a Solution Family of BVPs

Let us now assume that the vector eld (4.1) has a saddle equilibrium x0
with a two-dimensional unstable manifold W u (x0 ), meaning that the Jacobian Df (x0 ) has exactly two eigenvalues u1 and u2 with positive real part.
Suppose further that v1 and v2 are the associated (generalized) eigenvectors.
By denition, W u (x0 ) consists of all orbits that converge to x0 in backward
time. The Stable Manifold Theorem [45] implies that W u (x0 ) is tangent at
x0 to the plane spanned by v1 and v2 . Therefore, we approximate W u (x0 ) by
the collection of orbit segments that start on this plane at distance from x0 .
That is, we are looking for solutions of the system
u (t) = T f (u(t)),
u(0) = x0 + (cos() v1 + sin() v2 ),

(4.11)
(4.12)

where and T are free parameters. If the eigenvalues u1 and u2 are real, then
it is advantageous to choose the initial conditions on the ellipse given by the
ratio of the eigenvalues as


v2
v1
(4.13)
u(0) = x0 + cos() u + sin() u .
|2 |
|1 |

In other words, boundary condition (4.12) is replaced by (4.13) during the

continuation.
A rst orbit segment to start the continuation can be generated using
Auto as well. Namely, for arbitrary xed = 0 (0 0 < 2) and T = 0,
the (constant) orbit u(t) = x0 + (cos(0 ) v1 + sin(0 ) v2 ) with 0 t 1 is a
solution for system (4.11)(4.12). An actual trajectory for the specic value of
0 is then obtained using continuation in the free parameter T while keeping
the angle 0 xed. While this may seem like a complicated way of integrating
from an specic initial condition, it has the benet that the output les of this
rst step in Auto are then compatible with subsequent continuation steps.
In the case of computing a stable manifold, one simply generates an orbit for
T < 0, that is, integrating backward in time.
During the continuation of this rst orbit, one can monitor a user-dened
function (a suitable end-point condition) that becomes the second boundary condition in the continuation that generates the two-dimensional manifold. That is, once the rst orbit is generated, we continue this solution as a
boundary value problem where the initial condition on the small circle (4.12),
or ellipse (4.13) varies with the angle , which is now a free parameter in the
continuation. The resulting one-parameter family of orbit segments forms an
approximation of (a part of) W u (x0 ). It is important to note that the angle
, which denes the initial condition, is not the sole continuation parameter
that determines the stepsize for the next orbit segment in the family. Instead
each continuation step is taken in the full product space of the (discretized)
functions u() and the parameters. That is, the continuation stepsize includes

4 Computing Invariant Manifolds

135

variations along the entire orbit, and is not xed a priori, but is one of the
variables solved for in each continuation step.
There are several options for the second boundary condition that denes
the one-parameter solution family. We discuss several here, but our list is
certainly not exhaustive.
1. Fixed integration time
Probably the simplest choice is to compute a rst orbit for xed angle
= 0 up to a suitable integration time T = T0 , after which (4.11)(4.12)
is continued in the angle while the parameter T is kept xed. An example
of a computation with this choice for the second boundary condition can be
found in [14], where the computation of the unstable manifold of one of the
nontrivial equilibria in the Lorenz system [42] is used to nd heteroclinic
connections to the origin 0. Namely, the orbits that (almost) connect to 0
spend a very long time near 0. Hence, when T is xed, these connecting
orbits stand out by their short arclength; see [14, Fig. 9].
2. Fixed arclength
Another obvious choice is to x the total arclength of the orbit segments.
To this end, one imposes the integral constraint
 1
T || f (u(s)) || ds L = 0
(4.14)
0

along the orbit segment, while solving (4.11)(4.12). During the continuation, both and T are free parameters and L is kept at a desired xed
arclength. A computation with a xed arclength was used in [39, Sec. 3] for
the computation of the stable manifold of the origin in the Lorenz system
[42].
3. Fixed product of arclength and integration time
It may be advantageous in certain calculations to x the product L T ,
where L is the total arclength along the orbit segment as dened above.
This is particularly useful if W u (x0 ) contains connecting orbits with a
nite arclength of less than L. Since a connecting orbit is characterized
by the fact that T , keeping the product L T xed allows one to
continue orbit segments on W u (x0 ) in past connecting orbits.
4. Constrained end point
Similar to what was done in Sect. 4.1 one can restrict the end point u(1)
to a codimension-one section of the phase space. This is done by adding
to system (4.11)(4.12) the equation
g(u(1), , T ) = 0.
Here g is an appropriate functional, chosen to control the end point in
a desirable manner, for example, by requiring one coordinate to have a
particular xed value. The free parameters are again the angle and the

136

Bernd Krauskopf and Hinke M Osinga

integration time T , while is kept xed. We discuss an example of this

condition in more detail in the next section.
Indeed there are many other options to restrict the computations to a specic part of W u (x0 ), which makes this approach rather exible. For example,
the integrand of (4.14) in option 2 can be replaced by any other appropriate
functional h = h(u(s), , T ) to control the orbit in a desirable manner. Furthermore, one can choose a combination of end point conditions and integral
constraints, but this is beyond the scope of this chapter.
2D Unstable Manifolds in the Lorenz System
We now give a more detailed example of how the above set-up can be used
to compute (un)stable manifolds even if they have a very complex geometry.
To this end, we consider the well-known Lorenz equations [42], which can be
written as the vector eld

x = (y x),
y = x y xz,
(4.15)

z = xy z.
As is well known, the now classic parameters values of = 10, = 8/3
and  = 28 lead to the Lorenz buttery attractor, arguably the most famous
example of a chaotic attractor.
In [14] a detailed study is presented of how the two-dimensional stable
manifold W s (0) of the origin 0 and the two-dimensional unstable manifolds
W u (p+ ) and W u (p ) of the other two saddle equilibria p+ and p intersect
in a combinatorial structure of heteroclinic orbits. Note that the points p+
and p and the manifolds W u (p+ ) and W u (p ) are each others image under
rotation by about the z-axis, which is a symmetry of (4.15). Hence, it is
sucient to compute W u (p+ ).
While Auto has no problems computing one-parameter families of orbit
segments on W u (p+ ), it is a challenge to render this manifold as a nice surface.
This is because W u (p+ ) accumulates on the Lorenz attractor, which means
that it folds back over itself (and its counterpart W u (p )) innitely often and
exponentially closely. The key ingredient here is to identify suitable families
of orbit segments. To this end, we take the well-established topological point
of view of identifying the dierent sheets to obtain a branched surface, which
is known as the template of the Lorenz system [22, 23].
Specically, we generate an approximate template for W u (p+ ) by considering the two families of orbit segments shown in Fig. 4.7 (a) and (b). Both
start along a xed vector in the unstable eigenspace of p+ (at the diamonds)
and one family of orbit segments winds only around p+ , while the other winds
once around p . The end points of both families (indicated by the black horizontal lines) are constrained to lie in the plane = {z = 1}, which passes

4 Computing Invariant Manifolds

137

50

50

(a1)

(a2)
40
30

40

20
10

30

0
20
50

10

10

20

(a3)

20
40
30

10

20
10

0
20

15

10

10

15

20

Rl

0
20

10

10

20

10

20

10

20

50

50

(b1)

(b2)
40
30

40

20
10

30

0
20
50

10

(b3)

20
40
30

10

20
10

0
20

15

10

10

15

20

0
20

Rl
10

Fig. 4.7. The two families of orbit segments used to compute a template for W u (p+ )
of (4.15). The orbit segments start in the unstable eigenspace (at the ) of W u (p+ )
and end in the section 28 = {z = 27}, either near p+ (a1) or near p (b1). Both
families limit on the two singular orbit segments in panels (a2)/(a3) and (b2)/(b3),
which are composed of a heteroclinic connection to 0 (of type R or Rl) and an
initial piece of W u (0). From E.J. Doedel, B. Krauskopf and H.M. Osinga, Global
c 2006
bifurcations of the Lorenz manifold, Nonlinearity 19(12) (2006) 29472972 
by Institute of Physics Publishing; reprinted with permission.

138

Bernd Krauskopf and Hinke M Osinga

(a)

(b)

(c)

(d)

Fig. 4.8. The unstable manifolds of the nontrivial saddles p and of the saddle
periodic orbit , respectively, as represented by the family of orbits in Fig. 4.7,
for = 28.0 (a), = 23.0 (b), = 19.0 (c), and = 14.0 (d). From E.J. Doedel,
B. Krauskopf and H.M. Osinga, Global bifurcations of the Lorenz manifold, Nonlinc 2006 by Institute of Physics Publishing; reprinted
earity 19(12) (2006) 29472972 
with permission.

through p+ and p . In other words, we use option 4 of the boundary conditions above by imposing the additional boundary condition that u(1) .
This choice was made so that the two families cover most of the right and
left wings of the Lorenz attractor, respectively.
We now explain this set-up and the dierent continuation steps in more
detail. Since the eigenvalues of p+ are complex conjugate, which implies that
orbits near p+ spiral around p+ , it is convenient to use the initial condition
u(0) = x0 + r v1

(4.16)

4 Computing Invariant Manifolds

139

instead of (4.12). Here v1 is the real part of the complex conjugate pair of
eigenvectors associated with the eigenvalues of the linearization at p+ , and r
is the radial distance from p+ . Note that increasing r from one intersection to
the next of the same orbit (a fundamental domain of the return map to this
direction) is equivalent to changing the angle in (4.12) over 2.
To generate an initial orbit segment of the family shown in Fig. 4.7(a)
we choose r = r0 = 5.5 xed and continue the solution u(t) = x0 + r0 v1 ,
T = 0 in the free parameter T . Note that W u (p+ ) is extremely at near p+ so
that a point on the linear approximation at distance r0 = 5.5 from p+ is still
very close to W u (p+ ). This is advantageous, because additional revolutions
around p+ , of which there are more and more as r 0, can only be resolved
accurately if a large number of mesh points is used. By using a quite large
value for r0 it suces to use 800 mesh points. During this rst run, we monitor
the end-point condition
P AR(12) = u(1)z ( 1)

(4.17)

where u(1)z denotes the z-component of u(1). Hence, every time P AR(12) is
zero, which is detected as a user dened point (UZ in Auto), the end point
u(1) lies in . The user can then select any such orbit segment, which is how
we obtained the two families in Fig. 4.7(a) and (b).
The next step of the process then consists of an Auto run where we continue system (4.11) with (4.16) and (4.17) as boundary conditions, and r and
T as free parameters. In the case of the orbit segments in Fig. 4.7(a), in one
direction the end point u(1) moves closer to p+ . The last orbit segment in
this run is shown in Fig. 4.7(a2), where we can observe how the orbit passes
extremely close to 0 before making its loop around p . Indeed, the orbit segments accumulate on a heteroclinic connection concatenated with a rst piece
of the left branch of the one-dimensional unstable manifold W u (0) of 0. The
heteroclinic connection is of type R, that is, it only spirals around p+ and never
around p before converging to 0. When running the continuation in the opposite direction, the end point u(1) moves further away from p+ and the orbit
segments accumulate on the heteroclinic connection of type Rl concatenated
with a rst piece of the right branch of W u (0); the last orbit segment for this
run is shown in Fig. 4.7(a3). Note that the initial point u(0) (indicated by the
diamond) virtually does not move during these computations.
The second part of the template for W u (p+ ) is generated in a similar
way. Here the starting orbit is generated from an initial condition at distance
r0 = 5.7 from p+ so that it spirals more than once around p . The next time
when u(1) passes through after the rst loop around p is where we stop
the continuation in T and use the resulting orbit segment to start the second
run. The resulting family of orbit segments is shown in Fig. 4.7(b1). The continuation in one direction leads to u(1) moving closer to p , where the last
orbit segment is very close to the heteroclinic connection of type R concatenated with a rst piece of the right branch of W u (0); see Fig. 4.7(b2). The
continuation in the opposite direction has the eect that the orbit segments

140

Bernd Krauskopf and Hinke M Osinga

accumulate on the heteroclinic connection Rl concatenated with a rst piece

of the left branch of W u (0); the last orbit in this run is shown in Fig. 4.7(b3).
Again, u(0) (indicated by the diamond) hardly changes in these runs.
The two families shown in Fig. 4.7(a1) and Fig. 4.7(b1) can be used to
represent W u (p+ ). To render W u (p+ ) as a surface we generate a triangulation
by connecting corresponding mesh points on neighboring orbit segments. The
result as generated by the package Geomview [46] is shown in Fig. 4.8(a); this
image also shows the symmetric counterpart W u (p ). The resulting surface
is indeed a good representation of the template of the Lorenz system; its
boundary is given by pieces of the unstable manifold of the origin.
Topologically the same orbit families as in Fig. 4.7 can be used to compute
the unstable manifolds of the saddle equilibria p and the bifurcating saddle
periodic orbits for dierent values of the parameter . Figure 4.8(b)(c)
shows the manifolds W u ( + ) and W u ( ) for  = 23.0,  = 19.0,  = 14.0,
respectively; the periodic orbits are shown as black curves. Notice how for
decreasing  the periodic orbits (black curves) grow. At the same time, the
hole in the respective template around the (now attracting) equilibria grows.
Eventually, for  13.9265 the periodic orbits disappear in a homoclinic
bifurcation with the origin; see [14] for more details of this bifurcation.
One can notice some artifacts in the surface rendering in Fig. 4.8. They
occur because the orbit segments that are used to render the surface are not
evenly spaced everywhere. In fact, in some regions they are very close together.
For example, very many mesh points in Fig. 4.7 are contained in the very rst
part of W u (p+ ) where the orbit segments spiral around p+ ; this part of the
surface could be represented with a lot fewer mesh points. Furthermore, the
distance between mesh points on a given orbit segment is roughly given by
the arclength of the orbit segment divided by the number of mesh points.
Therefore, our straightforward way of generating a triangulated surface from
the Auto data results in quite elongated triangles in some parts of W u (p+ ).
While the actual error is quite small (since all orbit segments lie on W u (p+ ) in
very good approximation), the angles between neighboring triangles may be
quite large. Therefore, the articial lighting scheme of Geomview generates
some color dierences on the surface.
More generally, the construction of a smoother triangulation from Auto
data of orbit segments would require some serious post processing. Eectively,
one would need to decide where and how to thin out the data to have a
uniform bound on the sizes and aspect ratios of the triangles in the overall
triangulation.
4.3.2 A BVP Set-Up to Find Geodesic Level Sets
As we have just seen, the generation of a mesh of a prescribed quality from
a computed one-parameter family of orbit segments is not a simple matter.
While the continuation step size is such that the orbit segments are never

4 Computing Invariant Manifolds

141

far apart, this does not generally guarantee a nice rendering of the twodimensional surface. This problem occurs because the orbit segments are only
required to be good approximations of solutions to the BVP, and there are
no restrictions placed on the position of the mesh points with regard to the
geometry of the computed surface. In fact, all other methods for computing
global manifolds are addressing the mesh generation from a more geometrical point of view; see the recent survey [39] for more details. In particular,
Henderson [30, 39] stays closest to the idea of representing a global manifold
as a one-parameter family of orbit segments. The dierence is that he uses
curvature information to cover the manifold with fat trajectories, so that the
computed mesh points on suitably chosen orbit segments are well distributed
over the manifold.
We now explain how the GlobalizeBVP implementation [17] of our own
method [37, 38] for the computation of two-dimensional global (un)stable manifolds of vector elds makes use of Autos collocation and pseudo-arclength
continuation routines. The main idea behind this method is to build up the
manifold as a collection of geodesic level sets. This is a very natural choice from
a geometrical point of view, as the goal is to generate a circular mesh centered
around the equilibrium x0 (or periodic orbit ) that grows radially in steps
that are dictated by the local curvature. Each circle has the property that the
mesh points on it lie at the same geodesic distance from x0 (or ). Hence, the
GlobalizeBVP method also views the manifold as a one-parameter family
of curves geodesic level circles in the case of a two-dimensional manifold
but these curves are not directly related to the dynamics of the vector eld
(4.1). In contrast to the method discussed in Sect. 4.3.1, each geodesic circle
cannot be expressed as the solution of a family of two-point boundary value
problems. Instead, it is generated one point at a time.
Let us consider again the computation of a two-dimensional unstable manifold W u (x0 ) of a saddle point x0 . The GlobalizeBVP method starts with
N equally-spaced mesh points on the circle (4.12) in the unstable eigenspace
centered at x0 with radius . The piecewise-linear closed curve C through
these mesh points is an approximation of the rst geodesic level set, at distance from x0 . Let us denote the last computed geodesic level set by Cr and
suppose that we wish to nd the next geodesic level set Cb at a distance
from Cr .
For each mesh point rk on Cr we wish to nd the point bk on Cb that lies
closest to rk . To this end, we dene a plane Frk (approximately) perpendicular
to Cr at rk . The (unknown) intersection W u (x0 ) Frk is a one-dimensional
curve that is (locally) well dened. Note that any point in W u (x0 ) Frk lies
on an orbit that passes through Cr since, by denition, orbits on W u (x0 )
come from x0 . The curve W u (x0 ) Frk is parametrized locally near rk by the
end points u(1) Frk of solutions u(t), 0 t 1 to a one-parameter family
of two-point boundary value problems. Namely, we solve system (4.11)
u (t) = T f (u(t))

142

Bernd Krauskopf and Hinke M Osinga

with the boundary conditions

u(0) Cr ,

(4.18)

u(1) Frk .

(4.19)

The continuation starts from the trivial solution u(t) = rk , 0 t 1 with

T = 0. We then continue in the parameter T while we monitor the distance
T =|| u(1) rk ||
between the end point u(1) and the mesh point rk until the required distance
is reached, which denes the point bk on the new geodesic level set.
Once bk is found, it is tested against similar accuracy criteria as those
explained in Sect. 4.1. Namely, we restrict the angle between points on
three successive geodesic level sets and the product . This maintains a
good resolution of the manifold; if bk is not acceptable then the geodesic level
set currently being computed is discarded and is reduced. If all points on
Cb are found and acceptable then the geodesic level set is added to the mesh
representation of W u (x0 ). We refer to [17, 37, 38] for further details.
While earlier implementations [37, 38] used a shooting technique for solving the boundary value problem to nd the new point bk , the GlobalizeBVP
implementation [17] directly calls Autos collocation and pseudo-arclength
continuation routines. In fact, the set-up of the continuation of the onedimensional curve W u (x0 ) Frk is conceptually just as the computation of a
one-dimensional global manifold in a Poincare section discussed in Sect. 4.1.1.
Namely, the start point u(0) is continued along the piecewise-linear representation of the geodesic circle Cr , while the end point u(1) is conned to lie
in the plane Fr ; see [17] for details in terms of the respective user-dened
functions.
Stable Manifold in Chuas Circuit
We now illustrate the GlobalizeBVP method by computing the stable manifold W s (0) of the origin in Chuas circuit [8] with a smooth cubic nonlinearity;
see also [1, 55] and Chap. 7. The system is given by

x = (y a x3 c x),
y = x y + z,
(4.20)

z = y z,
where we take = 10.0, = 15.0, = 0.01, a = 0.1, and c = 0.2 as in [17].
For these parameters the origin has one unstable and two stable eigenvalues,
which are a complex conjugate pair. The strong spiralling dynamics on W s (0)
makes it particularly challenging to nd a good approximation of this global
manifold.

4 Computing Invariant Manifolds

143

Cb

(a)

W s (0)

(b)

Fig. 4.9. Computation of W s (0) in Chuas circuit (4.20) with the GlobalizeBVP
method. Panel (a) shows the orbit segments used in the computation of the geodesic
level set Cb at geodesic distance 0.75 from 0, and panel (b) the mesh with 64 geodesic
level sets when W s (0) is computed up to geodesic distance 2.1875.

We start the computation with 20 mesh points equally distributed on a

circle with radius 0.2 in the stable eigenplane centered around 0. The rst
few geodesic level sets are shown in Fig. 4.9(a). The surface shown is built up
from geodesic level sets at distances 0.25, 0.35, 0.55, and 0.65. Figure 4.9(a)
shows the orbit segments generated by the GlobalizeBVP method to nd
the next geodesic level set Cb at distance 0.75 by solving system (4.11) with
boundary conditions (4.18) and (4.19). Here Cr is the piecewise-linear curve
through the mesh points on the geodesic level set at distance 0.65, that is, the
boundary of the already computed piece of W s (0). The grey dots are the new
mesh points on Cb and the black dots are the other end points of the orbit
segments on Cr .
Notice that about half of the orbit segments move far away from 0 before
ending on Cb , which is a direct consequence of the nature of the dynamics
on W s (0). We remark that the example of Chuas circuit is quite extreme
in this respect. In fact, the required orbit segments become even longer for

144

Bernd Krauskopf and Hinke M Osinga

geodesic level sets at larger distances. This is in contrast to, for example, the
two-dimensional stable manifold of the origin of the Lorenz system, which is
characterized by two real positive eigenvalues. Generally speaking, the need
to nd quite long orbit segments by our BVP approach is the price we pay for
obtaining the parametrization of the global manifold by geodesic level sets.
On the positive side, this parametrization is geometrically optimal in the sense
that it is given by the geometry of the manifold and not by the dynamics on
it. Furthermore, it allows us to derive rigorous error bounds that go to zero
with the user specied accuracy parameters; see [38] for details.
Figure 4.9(b) shows the stable manifold W s (0) in Chuas circuit (4.20)
computed up to geodesic distance 2.1875. Notice how W s (0) folds quite
sharply along the top and bottom edges of the image. The surface is rendered so that the computed mesh is visible. The computed manifold W s (0)
is built up from a total of 64 geodesic level sets, that is, concentric topological circles; the distance between them is determined by the local curvature
as was explained above. The radial curves are approximate geodesics. Notice
how during the course of the computation many new mesh points are added,
which then give rise to new approximate geodesics. Similarly, mesh points
may be removed, which can be identied in the mesh as an end point of an
approximate geodesic; an example can be seen in the top left of Fig. 4.9(b).
For more images of W s (0) we refer to [17].
Stable Manifold of an Optimal Control Problem
The representation of a global manifold by approximate geodesic level sets
can be exploited for visualization purposes. As we demonstrate now with an
example of a two-dimensional stable manifold in a four-dimensional phase
space, this is a particularly useful feature when the phase space is higher
dimensional. Specically, we consider an inverted planar pendulum that is
balanced on a cart subject to a horizontal control force [28, 32, 44]. The
system can be written as

x 1 = x2 ,
g
r
(4.21)
sin(x1 ) 12 mr x22 sin(2x1 ) m
m l cos(x1 ) u

,
x 2 = l
4
2
3 mr cos (x1 )

where x1 is the angle measured from the upright position (not taken modulo
2), x2 is its angular velocity, mr is the mass fraction of the pendulum with
respect to the total mass (of pendulum and cart), l is the length of the pendulum, and g is the Earths gravitational constant. The function u constitutes
a control that is supposed to stabilize the point (x1 , x2 ) = (0, 0), which is the
unstable equilibrium corresponding to the upright position.
A cost is associated with the stabilization via the instantaneous cost function
(4.22)
Q(x1 , x2 , u) = 1 x21 + 2 x22 + 3 u2

4 Computing Invariant Manifolds

145

that penalizes both the state and the control, as long as the origin is not
stabilized. Here 1 , 2 and 3 are positive parameters. Pontryagins maximum
principle [56] ensures that an optimal control u exists that minimizes the
cost function Q over the innite time interval [0, ). The optimal solution is
represented by points on the two-dimensional stable manifold W s (0) of the
four-dimensional vector eld given by the Hamiltonian
H(x1 , x2 , p1 , p2 ) = Q(x1 , x2 , u (x1 , x2 , p1 , p2 )) + p1 x2 +
p2 f (x1 , x2 ) + p2 c(x1 , x2 ) u (x1 , x2 , p1 , p2 ) (4.23)
where u (x1 , x2 , p1 , p2 ) = 21 3 c(x1 , x2 )p2 . Namely, for any given initial condition (x1 , x2 , p1 , p2 ) on W s (0), the projection of the corresponding trajectory
onto the (x1 , x2 )-plane corresponds to a stabilizing solution via the (implicitly
dened) feedback control u = u (x1 , x2 , p1 , p2 ) that locally minimizes (4.22).
Indeed, if in this projection W s (0) covers a point (x1 , x2 ) more than once,
then typically only one of these solutions is optimal and the others are only
suboptimal; see [28, 44] for more details.
Figure 4.10 shows W s (0) for the parameters in [32], namely mr = 0.2,
l = 0.5 m, and cost function parameters 1 = 0.1, 2 = 0.05 and 3 = 0.01.
The two-dimensional manifold W s (0) was computed up to a geodesic distance
of approximately 26.25. It is rendered transparent in Fig. 4.10 and shown as
four projections onto the three-dimensional subspaces that one obtains by
setting one of the coordinates to zero. The transparent rendering allows one
to see how W s (0) sits in each of the three-dimensional projections. Note
that the computed part of W s (0) is a topological disk that is parametrized by
the geodesic level sets. In particular, the boundary of the computed manifold
has the same geodesic distance to the origin, which lies in the center of the
manifold.
To help further with the interpretation of the geometry of W s (0) in R4 , a
single geodesic band can be moved over the manifold to observe how its geometry changes simultaneously in all four projections. As an example, Fig. 4.10
shows the geodesic band covering the range 1920 in a dierent color. Note
that the band is unknotted and that it divides the manifold W s (0) into an
inner disk and an outer annulus, which is not so obvious in Fig. 4.10(d).

4.3.3 A Two-Point Boundary Value Problem Set-Up for the

Computation of Quasiperiodic Invariant Tori
An invariant torus of a given dynamical system of the form (4.1) is a twodimensional compact invariant manifold that can be viewed as the twodimensional analogue of an equilibrium or a periodic orbit. Invariant tori
are born, for example, in a Neimark-Sacker bifurcation, where a pair of complex conjugate Floquet multipliers of a periodic orbit crosses the unit circle
in the complex plane, or they arise from the uncoupled case in a system of
coupled oscillators. The bifurcating torus is initially normally hyperbolic [31],

146

Bernd Krauskopf and Hinke M Osinga

(a)

(b)

(c)

(d)

Fig. 4.10. The two-dimensional stable manifold W s (0) in (x1 , x2 , p1 , p2 )-space of

the controlled inverted pendulum (4.23). The surface is rendered transparent and is
shown simultaneously in the four projections p2 = 0 (a), p1 = 0 (b), x2 = 0 (c), and
x1 = 0 (d); the dierently colored band covers geodesic distances 19.020.0.

which means that it is persistent under small perturbations. Depending on the

rotation number, the dynamics on the torus is either quasiperiodic (equivalent to a parallel ow on the standard torus) or locked, in which case there
is an attracting and a repelling periodic orbit on the torus. Locked dynamics occurs in so-called resonance or Arnold tongues, parameter regions that
are bounded by saddle-node bifurcations of the locked periodic orbits on the
torus. The resonance tongues emerge from points on a Neimark-Sacker curve
where the rotation number is rational, while the quasiperiodic tori exist along
one dimensional curves that start from points where the rotation number is
irrational. The overall picture is truly two-dimensional in parameter space;
see, for example, [3, 41, 52] as entry points to the extensive literature. A main

4 Computing Invariant Manifolds

147

diculty is that invariant tori may lose their normal hyperbolicity in many
dierent ways, some of which are still not fully understood.
It may not be too surprising that the computation of invariant tori is still
considered a very challenging task. In fact, only a few specialized algorithms
exist; see [49] for a recent overview. We review here the algorithm presented
in [50] for the computation of quasiperiodic invariant tori, because it utilizes
the two-point boundary value solver of Auto. Namely, the idea is to view a
quasiperiodic invariant torus T of (4.1) as an invariant circle T in a suitable
Poincare section that is chosen transverse to T. The associated Poincare
map that leaves T invariant can be formulated as a two-point boundary value
problem in the same way as was done in Sect. 4.1.
To explain the method from [50] in detail we need the dependence of (4.1)
on the parameter and, for convenience, we assume that it is of the form
= (, ). We restrict our discussion to the case of a two-dimensional torus
T, but it is straightforward to generalize the method to higher-dimensional
tori. The idea is to construct a two-point BVP in two steps: we start with the
invariance condition for invariant circles of the time-T map of (4.1) and then
replace the time-T map with a BVP.
An invariant circle u() of the time-T map T of (4.1), where S1 is
parametrized over [0, 2], satises the invariance condition
u( + 2) = T (u(), , ),

(4.24)

where  is the rotation number of T restricted to the invariant circle [6,

34]. The invariance condition (4.24) is discretized by approximating u with a
Fourier polynomial
u() = c1 +

N
%

c2k sin + c2k+1 cos ,

k=1

where cj Rn are real coecient vectors, and requiring that (4.24) holds at
the Q = 2N + 1 collocation points j = 2j/Q, j = 1, . . . , Q. Note that the
discretized solution is not unique. First of all, T is still unknown and, secondly,
the solution u() is not isolated because for any phase shift s S1 the shifted
function u( + s) is also a solution.
By identifying xj (0) = u(j ) and xj (1) = u(j + 2) and substituting the
initial value problem {x j = f (xj , , ), xj (0) = u(j )} for the time-T map
T in (4.24), one obtains the BVP

148

Bernd Krauskopf and Hinke M Osinga

n 
%
i=1

x j = T f (xj , , ),
cj = 0,

(4.25)

xj (0) = u(j ),
xj (1) = u(j + 2),
n  2
%
u
i ()ui () d = 0,

(4.27)

i=1
1

(4.26)
(4.28)
(4.29)

(x1,i (t) x
1,i (t))x1,i (t) dt = 0.

(4.30)

Here boundary conditions (4.27) and (4.28) represent the invariance condition
(4.24), while (4.29) and (4.30) are phase conditions that determine a unique
return time T and phase shift s. Note that s does not occur explicitly in (4.29)
and that (4.30) cannot be simplied similarly to (4.29), because x1 is not
periodic in t; see also [47, 48, 50]. Condition (4.26) on the Fourier coecients
can be omitted, but is necessary if one wants to implement (4.25)(4.30) in
Auto.
System (4.25)(4.30) consists of 2Q n-dimensional ODEs and 2Q ndimensional boundary conditions for the functions xj and cj , and two scalar
phase conditions for T and the parameter . Continuation of a solution of
this system with respect to is a way to compute a codimension-one family
of quasiperiodic invariant tori with the xed rotation number .
Tori in an Electronic Circuit Model
To demonstrate the method we consider the model of an electronic circuit that
was investigated in detail in [49, 51]. The system can be written in vector-eld
form as the parametrically forced system

x = y,

y = (/2 )y ( )y 3 (1 + sin 2t)x,

(4.31)

t = 1,
where is the forcing amplitude and determines the nonlinear damping.
For = 0 system (4.31) is autonomous and its ow t (x, y, t) in the phase
space R R S1 is a superposition of the ow in the two-dimensional (x, y)space with the constant ow given by t = 1. For > 0 there is a limit cycle
in (x, y)-space and, hence, system (4.31) has a normally attracting invariant
torus, which persists as an invariant manifold for suciently small forcing
amplitude [19]. For a specied rotation number  a start solution for = 0
is found by continuing the -family of periodic solutions in the (x, y)-plane and
interpolating it as a function of the rotation number  = /T (). Here T ()
is the period of the periodic solutions that for (4.31) increases monotonically
as a function of .

4 Computing Invariant Manifolds

0.6

(d)

(a)

0.5

149

(b)

(c)

0.4
0.3
(b)

0.2
0.1
0
3

(c)

(d)

Fig. 4.11. The bifurcation diagram (a) of (4.31) with selected resonance tongues and
curves of quasiperiodic tori, together with three quasiperiodic tori (b)(d) computed
as solutions to (4.25)(4.30). The torus is shown as a gray surface and the closed
curves at t = 0 and t = represent the same invariant circle. The thin black curves
connecting these circles are orbit segments used in the computation; only every
fourth orbit is shown for clarity.

Figure 4.11 shows the result of computations of quasiperiodic invariant

tori of (4.31) with the above BVP set-up (4.25)(4.30). The computations
were performed with N = 64 Fourier modes and a second-order Gau collocation scheme with 300 mesh points. Panel (a) shows from left to right the
Arnold tongues for the rotation numbers  { 37 , 25 , 38 , 13 } as grey wedges

N
2 with
and the quasiperiodic curves for the rotation numbers  = 140
N {41, 38, 35} as bold black curves. Figure 4.11(b)(d) shows sample solutions of the BVP (4.25)(4.30), that is, representations of quasiperiodic
invariant tori that were obtained at the marked positions in panel (a) during
the continuation along the left-most quasiperiodic curve. The tori are shown
as gray tubes that are bounded by bold black closed curves representing the
invariant circle at t = 0 and t = . The thin black curves are orbits on the
tori that connect the points xj (0) = u(j ) with xj (1) = u(j + 2), which
are part of the solution of (4.25)(4.30); for clarity, only every fourth orbit
is shown. Notice the change to the torus from Fig. 4.11(b) to Fig. 4.11(d)
when is increased towards the top of the locus where the torus loses normal
hyperbolicity and the computation stops.

150

Bernd Krauskopf and Hinke M Osinga

4.4 Summary and Outlook

The basic idea of continuing the solution of a two-point boundary value problem is very versatile. As we have demonstrated in this review, it can also be
used to compute dierent types of global manifolds, including one-dimensional
(un)stable manifolds in Poincare maps, two-dimensional global manifolds of
vector elds, slow manifolds in slow-fast systems, as well as quasiperiodic invariant tori. The common theme here is that one needs to consider a suitable
family of orbit segments. The continuation of this family can then be achieved
by making use of the collocation solver and continuation routines of Auto.
In this chapter we have given a avor of how the BVP solver of a package such as Auto can be used in less obvious ways. We hope that this will
stimulate further research into the computation of invariant manifolds, and
we mention two directions of ongoing research.
As the wide range of applications indicates, there are other types of invariant objects that can be computed within the presented general framework. In
particular, there is an interest in manifolds that are associated with a certain
subset of the stable or unstable eigenvalues or Floquet multipliers [31, 59]. Examples are the computations of a weak unstable manifold in a delay equation
model of a laser with phase conjugate feedback [25], and of a two-dimensional
surface (with the method in Sect. 4.3.1) that separates forward and backward
phase resets in a model of a cardiac pacemaker [40].
Another interesting possibility is the computation of higher-dimensional
global manifolds in a similar fashion, where the manifold is now represented as
a family of orbit segments that is parametrized by more than one continuation
parameter. Combining the general BVP set-up presented here with higherdimensional continuation as discussed in Chap. 3 provides a quite natural
approach for computing, for example, two-dimensional manifolds in Poincare
maps and three-dimensional global manifolds in vector elds. However, there
remain interesting open questions concerning the representation and visualization of the resulting data.

Acknowledgments
The material reviewed in Sects. 4.1.1, 4.2.1 and 4.3.2 is joint work with James
England, who we thank for his contributions during our collaboration. We
also thank Mathieu Desroches and Frank Schilder for their help with text
and gures in Sects. 4.2.2 and 4.3.3, respectively. The material reviewed in
Sect. 4.3.1 is joint work with Sebius Doedel, and we would like to thank
Sebius for his continuing inuence on our research and his invaluable support
throughout our careers. Our visits to Montreal are always productive as well
as enjoyable, and we blame Sebius for our addiction to Montreal croissants!
Finally, we are grateful to the Society for Industrial and Applied Mathematics
and to Institute of Physics Publishing for permission to reproduce previously
published material from [16] and from [14].

4 Computing Invariant Manifolds

151

References
1. A. Algaba, M. Merino, E. Freire, E. Gamero, and A. J. Rodriguez-Luis. Some
results on Chuas equation near a triple-zero linear degeneracy. Internat. J.
Bifur. Chaos Appl. Sci. Engrg., 13(3):583608, 2003.
2. E. L. Allgower and K. Georg. Numerical path following. In P. G. Ciarlet and
J. L. Lions, editors, Handbook of Numerical Analysis, volume 5, pages 3207.
North Holland Publishing, 1996.
3. V. I. Arnold. Geometrical Methods in the Theory of Ordinary Dierential Equations. Springer-Verlag, Berlin, 1983.
4. J. Boissonade and P. DeKepper. Transition from bistability to limit cycle oscillations. J. Phys. Chem., 84:501506, 1980.
5. M. Brns, M. Krupa, and M. Wechselberger. Mixed mode oscillations due to
the generalized canard phenomenon. Fields Institute Communications, 49:39
63, 2006.
` Jorba. On the vertical families of two-dimensional tori near
6. E. Castell`
a and A.
the triangular points of the bicircular problem. Celestial Mech. Dynam. Astronom., 76(1):3554, 2000.
7. A. R. Champneys, Yu. A. Kuznetsov, and B. Sandstede. A numerical toolbox
for homoclinic bifurcation analysis. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
6(5):867887, 1996.
8. L. O. Chua, M. Komuro, and T. Matsumoto. The double scroll family. IEEE
Trans. Circuits and Systems, CAS-33:10731118, 1986.
9. P. Collins and B. Krauskopf. Entropy and bifurcation in a chaotic laser. Phys.
Rev. E, 66:056201, 2002.
10. M. Desroches, B. Krauskopf, and H. M. Osinga. The geometry of slow manifolds
near a folded node. Technical report, University of Bristol, 2007, to appear.
11. A. Dhooge, W. Govaerts, and Yu.A. Kuznetsov. Matcont: A Matlab package
for numerical bifurcation analysis of ODEs. ACM TOMS, 29(2):141164, 2003.
available via http://www.matcont.ugent.be/.
12. E. J. Doedel. Auto, a program for the automatic bifurcation analysis of autonomous systems. Congr. Numer., 30:265384, 1981.
13. E. J. Doedel and M. Friedman. Numerical computation of heteroclinic orbits.
Continuation techniques and bifurcation problems. J. Comput. Appl. Math.,
26(1-2):155170, 1989.
14. E. J. Doedel, B. Krauskopf, and H. M. Osinga. Global bifurcations of the Lorenz
manifold. Nonlinearity, 19(12):29472972, 2006.
15. E. J. Doedel, R. C. Paenroth, A. R. Champneys, T. F. Fairgrieve,
Yu. A. Kuznetsov, B. E. Oldeman, B. Sandstede, and X. J. Wang. Auto2000:
Continuation and bifurcation software for ordinary dierential equations. available via http://cmvl.cs.concordia.ca/.
16. J. P. England, B. Krauskopf, and H. M. Osinga. Computing one-dimensional
global manifolds of Poincare maps by continuation. SIAM J. Appl. Dyn. Sys.,
4(4):10081041, 2005.
17. J. P. England, B. Krauskopf, and H. M. Osinga. Computing two-dimensional
global invariant manifolds in slow-fast systems. Internat. J. Bifur. Chaos Appl.
Sci. Engrg., 16(3), 2007. at press.
18. I. R. Epstein, K. Kustin, P. DeKepper, and M. Orban. Oscillating chemical
reactions. Sci. Am., 3(248):112123, 1983.

152

Bernd Krauskopf and Hinke M Osinga

19. N. Fenichel. Persistence and smoothness of invariant manifolds for ows. Indiana
Univ. Math. J., 21:193226, 1971.
20. N. Fenichel. Geometric singular perturbation theory. J. Di. Eq., 31:5398,
1979.
21. M. Friedman and E. J. Doedel. Numerical computation and continuation of
invariant manifolds connecting xed points. SIAM J. Numer. Anal., 28(3):789
808, 1991.
22. R. Ghrist, P. J. Holmes, and M. C. Sullivan. Knots and Links in ThreeDimensional Flows, volume 1654 of Lecture Notes in Mathematics. SpringerVerlag, Berlin, 1997.
23. R. Gilmore and M. Lefranc. The Topology of Chaos: Alice in Stretch and
Squeezeland. Wiley-Interscience, New York, 2004.
24. W. Govaerts, Yu.A. Kuznetsov, and B. Sijnave. Continuation of codimension-2
equilibrium bifurcations in CONTENT. In E. J. Doedel and L. S. Tuckerman,
editors, Numerical methods for Bifurcation Problems and Large-Scale Dynamical
Systems, pages 163184. Springer-Verlag, New York, 2000. available via http:
//ftp.cwi.nl/CONTENT/.
25. K. Green, B. Krauskopf, and K. Engelborghs. One-dimensional unstable eigenfunction and manifold computations in delay dierential equations. J. Comput.
Phys., 197(1):8698, 2004.
26. J. Guckenheimer, K. Homan, and W. Weckesser. The forced Van der Pol
equation I: the slow ow and its bifurcations. SIAM J. Appl. Dyn. Sys., 2:135,
2003.
27. J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical Systems
and Bifurcations of Vector Fields. Springer-Verlag, New York/Berlin, 2nd edition, 1986.
28. J. Hauser and H. M. Osinga. On the geometry of optimal control: the inverted
pendulum example. Proceedings of the American Control Conference, Arlington,
VA, Vol. 2, pages 17211726, 2001.
29. M. E. Henderson. Multiple parameter continuation: Computing implicitly dened k-manifolds. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 12(3):451476,
2002.
30. M. E. Henderson. Computing invariant manifolds by integrating fattened trajectories. SIAM J. Appl. Dyn. Sys., 4(4):832882, 2005.
31. M. W. Hirsch, C. C. Pugh, and M. Shub. Invariant Manifolds, volume 583 of
Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1977.
32. A. Jadbabaie, J. Yu, and J. Hauser. Unconstrained receding horizon control:
stability and region of attraction results. IEEE Trans. Automat. Control, 1999.
33. C. K. R. T. Jones. Geometric singular perturbation theory. In R. Johnson,
editor, Dynamical Systems (Montecatini Terme, 1994), volume 1609 of Lecture
Notes in Mathematics, pages 44120. Springer-Verlag, New York, 1995.
` Jorba. Numerical computation of the normal behaviour of invariant curves
34. A.
of n-dimensional maps. Nonlinearity, 14(5):943976, 2001.
35. M. T. M. Koper. Bifurcations of mixed-mode oscillations in a three-variable
autonomous Van der Pol-dung model with a cross-shaped phase diagram.
Physica D, 80:7294, 1995.
36. B. Krauskopf and H. M. Osinga. Growing 1D and quasi-2D unstable manifolds
of maps. J. Comput. Phys., 146(1):406419, 1998.
37. B. Krauskopf and H. M. Osinga. Two-dimensional global manifolds of vector
elds. Chaos, 9(3):768774, 1999.

4 Computing Invariant Manifolds

153

38. B. Krauskopf and H. M. Osinga. Computing geodesic level sets on global

(un)stable manifolds of vector elds. SIAM J. Appl. Dyn. Sys., 2(4):546569,
2003.
39. B. Krauskopf, H. M. Osinga, E. J. Doedel, M. E. Henderson, J. Guckenheimer,
A. Vladimirsky, M. Dellnitz, and O. Junge. A survey of methods for computing
(un)stable manifolds of vector elds. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
15(3):763791, 2005.
40. T. Krogh-Madsen, L. Glass, E. J. Doedel, and M. R. Guevara. Apparent discontinuities in the phase-resetting response of cardiac pacemakers. J. Theor.
Biol., 230:499519, 2004.
41. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory. Springer-Verlag,
New York/Berlin, 3rd edition, 2004.
42. E. N. Lorenz. Deterministic nonperiodic ows. J. Atmosph. Sci., 20:130141,
1963.
43. B. E. Oldeman, A. R. Champneys, and B. Krauskopf. Homoclinic branch switching: a numerical implementation of Lins method. Internat. J. Bifur. Chaos
Appl. Sci. Engrg., 13(10):29772999, 2003.
44. H. M. Osinga and J. Hauser. The geometry of the solution set of nonlinear
optimal control problems. J. Dynamics and Dierential Equations, 2006.
45. J. Palis and W. de Melo. Geometric Theory of Dynamical Systems. SpringerVerlag, New York/Berlin, 1982.
46. M. Phillips, S. Levy, and T. Munzner. Geomview: An interactive geometry
viewer. Not. Am. Math. Soc., 40(8):985988, 1993. available via http://www.
geom.uiuc.edu/.
47. F. Schilder. Algorithms for Arnold tongues and quasi-periodic tori: a case study.
In D. H. van Campen, M. D. Lazurko, and W. P.J . M. van den Oever, editors,
Proceedings of the Fifth EUROMECH Nonlinear Dynamics Conference, pages
ID 16100, 2005.
48. F. Schilder. Computing Arnold tongue scenarios: some recent advances. In
Proceedings of the 2006 International Symposium on Nonlinear Theory and its
Applications (NOLTA2006), 11-14 September, Bologna, Italy, 2006.
49. F. Schilder, H. M. Osinga, and W. Vogt. Continuation of quasi-periodic invariant
tori. SIAM J. Appl. Dyn. Sys., 4(3):459488, 2005.
50. F. Schilder and B. B. Peckham. Computing Arnold tongue scenarios. J. Comput.
Phys., 220(2):932951, 2007.
51. F. Schilder, W. Vogt, S. Schreiber, H. M. Osinga. Fourier methods for quasiperiodic oscillations. Internat. J. Numer. Methods Engrg., 67(5):629671, 2006.
52. A. Shilnikov, L. P. Shilnikov, and D. Turaev. On some mathematical topics in
classical synchronization: a tutorial. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
14(7):21432160, 2004.
53. S. H. Strogatz. Nonlinear Dynamics and Chaos. Addison-Wesley Publishing
Company, 1996.
54. P. Szmolyan and M. Wechselberger. Canards in R3 . J. Di. Eq., 177(2):419453,
2001.
55. A. Tsuneda. A gallery of attractors from smooth Chuas equation. Internat. J.
Bifur. Chaos Appl. Sci. Engrg., 15(1):149, 2005.
56. A. J. van der Schaft. L2 -Gain and Passivity Techniques in Nonlinear Control.
Springer-Verlag, Berlin, 1994.
57. M. Wechselberger. Existence and bifurcation of canards in R3 in the case of a
folded node. SIAM. J. Appl. Dyn. Sys., 4(1):101139, 2005.

154

Bernd Krauskopf and Hinke M Osinga

58. S. M. Wieczorek, B. Krauskopf, T. B. Simpson, and D. Lenstra. The dynamical

complexity of optically injected semiconductor lasers. Phys. Reports, 416(1
2):1128, 2005.
59. S. Wiggins. Normally Hyperbolic Invariant Manifolds in Dynamical Systems.
Springer-Verlag, New York, 1994.

5
The Dynamics of SQUIDs and Coupled
Pendula
Donald G Aronson1 and Hans G Othmer2
1
2

School of Mathematics and Institute for Mathematics and Its Applications,

University of Minnesota, Minneapolis, USA
School of Mathematics and Digital Technology Center, University of Minnesota,
Minneapolis, USA

Josephson [8] predicted in 1962 that a DC tunnel current would ow between

two superconductors connected by a thin insulating layer of thickness less
than about 20
A in the absence of a voltage dierence, an eect now called
the DC Josephson eect. The quantum-mechanical current, called the superconducting current, arises from the tunneling of Cooper pairs of electrons of
opposite spin and momenta and is given by
Is = Ic sin ,

(5.1)

where Ic is the critical current and is the dierence of the phases of the wave
functions of the two superconductors. This gives the ideal current through a
junction, but in real circuits there are resistive and capacitive currents as well.
One of the standard models of a more realistic circuit is the so-called StewartMcCumber resistively-shunted-junction (or RSJ) model, which is described
by the following equation for the current [6, 9]:

h d
hC d2
+ Ic sin = I.
+
2
2eR dt
2e dt

(5.2)

Here h is Plancks constant, e is the charge on an electron, h/2e is the ux

quantum, C is the capacitance, R is the resistance, 
and I is the imposed bias
current. To simplify (5.2) dene the frequency = 2eIc /hC and the scaled
time = t; then (5.2) becomes

+ + sin = i,

(5.3)

where = (RC)1 , i = I/Ic , and the dot denotes derivation with respect
to the rescaled time .
A very useful correspondence of this system to a pendulum provides insight
into the dynamics studied later. In fact, the pendulum will serve as the basic

156

Donald G Aronson and Hans G Othmer

Fig. 5.1. Two pendula coupled via a torsion bar. From D.G. Aronson, E.J. Doedel
and H.G. Othmer, The dynamics of coupled current-biased Josephson junctions II,
c
Internat. J. Bifur. Chaos Appl. Sci. Engrg. 1(1) (1991) 5166 1991
by World
Scientic Publishing; reprinted with permission.

physical model; see also [1]. Suppose that a pendulum consists of a bob of
mass m that is attached to a (weightless) rod of length L. Then the equation
of motion is
d
d2
+ mgL sin = T,
(5.4)
2 +
dt
dt

where = mL2 is the moment of inertia of the pendulum, g is the gravitational acceleration, is the damping, is the angle between the bob and
vertical measured from the downward position, and T is the applied torque.
After non-dimensionalization this leads to (5.3).
When a ring of superconducting material contains two Josephson junctions, the result is a superconducting quantum interference device (SQUID),
so called because the wave functions of the Cooper pairs at each junction
interfere. SQUIDS are among the most sensitive devices for detecting magnetic elds a SQUID is capable of detecting magnetic elds of around 2
picotesla, i.e., at the quantum ux level. The coupling between phases across
the junctions is proportional to the dierence of phases, and therefore, the
system of equations governing a SQUID is

1 + 1 + sin 1 = (2 1 ) + I,
(5.5)
2 + 2 + sin 2 = (1 2 ) + I.
Here, is the coupling coecient, and the dimensionless bias current I is
assumed to be the same for both junctions.
An identical pair of equations governs the motion of two pendula coupled
by a linear torsional spring or bar, and forced with an applied torque I; see
Fig. 5.1. We use this system as the paradigm in this chapter and we attempt
to synthesize the results of [3, 5] and the unpublished study [2], which are all
written in collaboration with Eusebius Doedel. The work involves extensive
numerical studies that were carried out using DsTool, MatLab, and primarily, Auto. In the next section we analyze the equilibria of (5.5). Section 5.2

5 The Dynamics of SQUIDs and Coupled Pendula

157

considers the undamped undriven case, which is part of the unpublished results in [2]. We analyze both equilibria and periodic orbits for this case, and
also discuss the computation of heteroclinic connections. Finally, Sect. 5.3
shows the existence of so-called rotations, periodic solutions with a period
that is an integer multiple of the forcing frequency. We discuss their stability
in Sect. 5.4 and draw some conclusions in Sect. 5.5.

5.1 Equilibria and Their Stability

We begin by analyzing the existence and stability of equilibria for the coupled
system. Clearly (5.5) is invariant under the transformation i i + 2
and, thus, denes a ow on the product space {S1 R}2 . In addition, (5.5)
is invariant under the transformations i i+1 (mod 2) and (i , I)
(i , I). Therefore, we assume that I 0 from now on. In order to analyze
(5.5) it is convenient to introduce the variables
r=

1
(1 2 )
2

and

s=

1
(1 + 2 ),
2

where r is (half) the instantaneous phase dierence and s is the average phase
dierence. In these variables, and when written as a rst-order system, (5.5)
becomes

r = u,

s = v,
(5.6)
u
= u sin r cos s 2r,

v = v cos r sin s + I.
If = 0 then the pendula are uncoupled, and if r = 0 then they are in phase or
synchronized. The subspace r = u = 0 is invariant under the ow associated
with this system and we refer to it as the in-phase subspace. The dynamics on
this subspace are well characterized, even when the forcing is time dependent,
because the fourth-order system reduces to a second-order system [4, 13].
The equilibria of (5.6) are given by (R, S, 0, 0), where R and S are solutions
to the system

sin R cos S = 2R,
(5.7)
cos R sin S = I.
Solutions with |R| > 0 are called asynchronous equilibria, and those with
R = 0 are called synchronous equilibria. Clearly the existence of equilibria is
independent of the damping, but the forcing must be small enough (|I| < 1)
to have an equilibrium. In the SQUID context this means that the bias current must be smaller than the superconducting current. In addition, solutions
must satisfy |R| < 1/2, and therefore the asynchronous solutions approach
synchronous solutions as the coupling strength increases.
With a slight abuse of notation we use the abbreviation (R, S) for equilibria, and in this notation we have the following: for = 0 system (5.7) has

158

Donald G Aronson and Hans G Othmer

two countably innite families of equilibria {(Rm , Sm )} and {(Rn , Sn )} that

satisfy
2m + 1

(5.8)
Rm = arccos (1)m I, Sm =
2
and
Rn = n, Sn = (1)n arcsin I,
(5.9)

respectively. If we dene = arcsin I then the complete set of solutions to

(5.8) is generated by 2-translates in S and -translates along the diagonal of
the basic sets
0 1
. /
and (R, S) = (0, {, }) .
(5.10)
,
(R, S) =
2
2

For example, the -translation along the diagonal of the basic set results in
the four equilibria



3

3
3
, (, 2 ) , and (, 2 + ) .
+ ,
,
,
2
2
2
2

When the applied torque vanishes (5.7) reduces to

sin R = 2R.

(5.11)

This equation has innitely many solutions at = 0, and the number of

solutions decreases to zero by a sequence of saddle-node bifurcations as ||
increases. The last solution disappears at the value for which the line y =
2R is rst tangent to the curve y = sin R.
The local stability of any equilibrium of (5.6) is determined by the eigenvalues of the Jacobian of the right-hand side of (5.6), which is

0
0
1 0
3
2

0
0
0 1
0
I2

.
J =
=
L I2
cos R cos S 2
sin R sin S 0
sin R sin S
cos R cos S 0
The eigenvalues of J are solutions to the pair of quadratic equations
2 + = 0.
Here are the eigenvalues of L, that is,
=

1

1.
trace(L) (trace(L))2 4det(L) ,
2

where
trace(L) = 2(cos R cos S + ),
2

det(L) = (cos R cos S) (sin R sin S)2 + 2 cos R sin S.

5 The Dynamics of SQUIDs and Coupled Pendula

159

(a)

(b)

Fig. 5.2. The rst ve sets of equilibria of (5.6) for = 0.15 as a function of the
coupling strength for I = 0 (a) and for I = 0.25 (b). Solid and dashed curves
denote stable and unstable solutions, respectively; saddle-node bifurcations arise at
the values of where there is a vertical tangent. From D.G. Aronson, E.J. Doedel
and H.G. Othmer, The dynamics of coupled current-biased Josephson junctions II,
c
Internat. J. Bifur. Chaos Appl. Sci. Engrg. 1(1) (1991) 5166 1991
by World
Scientic Publishing; reprinted with permission.

Equilibria for which are both negative are stable, those with + < 0
have a one-dimensional stable manifold, and those for which are both
positive have a two-dimensional unstable manifold.
If = 0, the characteristic equation of L is
2

( + cos R cos S) (sin R sin S)2 = 0.

and, therefore, for (R, S) given by (5.10) we have that

160

Donald G Aronson and Hans G Othmer

+ = = cos < 0 for (R, S) = (0, ),

= cos2 < 0 for (R, S) = (( 2 ), 2 ),
+
cos > 0 for (R, S) = (0, ).
+ = =

Thus, of the four points, only (R, S) = (0, ) is stable, and the same pattern
is found for all the translates of these points. In terms of the phase angles of
the individual pendula, the existence and stability of equilibria at = 0 can
be summarized as follows. The three types of solutions are:
1. solutions with 1 , 2 (0, /2); these are asymptotically stable for any
> 0.
2. solutions with 1 (0, /2) and 2 (/2, ); these have a threedimensional stable manifold and a one-dimensional unstable manifold.
3. solutions with 1 , 2 (/2, ); these have a two-dimensional stable manifold and a two-dimensional unstable manifold.
At = 0 there exists an innite number of other equilibria for which R = 0,
each of which can be continued for small || because none of the equilibria
that exist at = 0 is critical in the sense that the Jacobian has one or more
eigenvalues on the imaginary axis. These equilibria and their continuations
are naturally grouped into families of four equilibria, as determined above, by
the various choices of 1 and 2 at = 0. By translation in S, each of these
families determines an equivalence class of families modulo 2. For any = 0
only nitely many of these exist, the remaining ones having disappeared via
saddle-node bifurcations. Each of the families contains four equilibria at = 0,
from which the entire family can be generated by continuation. The resulting
families are shown in Fig. 5.2(a) for zero forcing and in Fig. 5.2(b) for I = 0.25;
here we plotted the phase dierence R versus . For the nth family, n = 0,
the four solutions at = 0 can be denoted as (d , d 2n), (u , d 2n),
(d , u 2n), and (u , u 2n), where d = , and u = d ; see
also [5].

5.2 Hamiltonian Dynamics

In the absence of damping and forcing (5.5) reduces to the Hamiltonian system
1 + sin 1 = (2 1 ),
2 + sin 2 = (1 2 ),

(5.12)

where the energy is given by

H=

1 . 2
1 + 21 (2 + cos 1 + cos 2 ) + (1 2 )2 .
2
2

In this section we summarize a portion of the results on the undamped undriven case from the unpublished work [2].

5 The Dynamics of SQUIDs and Coupled Pendula

161

Fig. 5.3. The branch of antidiagonal solutions and some bifurcating branches for
= 0.01. Where necessary, periods have been doubled to ensure that branches
connect continuously.

The uncoupled system (5.12) with = 0 has equilibria at all points

em,n = (m, n) in the (1 , 2 ) conguration plane. In the basic square
(, ) (, ) there are heteroclinic orbits joining the diagonal points
e1,1 and e1,1 , the antidiagonal points e1,1 and e1,1 , as well as all four
pairs of neighboring corner points. The system is doubly periodic, which generates the entire plane.
For = 0 the doubly periodic structure is destroyed since the energy
surfaces are bounded by the zero-velocity cylinders
(2 + cos 1 + cos 2 ) +

(1 2 )2 = 0.
2

The symmetry that comes from translation by 2 along the diagonal remains.
When (5.12) is written in (r, s) variables, one sees from (5.6) that the
diagonal r 0 and the antidiagonal s 0 are invariant. On the diagonal
there are equilibria at (k, 0) for all integers k. These equilibria are centers
for k even, while they are saddles for k odd. On the antidiagonal the equilibria
are (, 0), where is given by the solution of (5.11) with the negative sign.
Let us suppose that there is a minimal positive solution = 1 . To construct
an antidiagonal solution to (5.6) (with I = = 0) we solve the initial value
problem
r + sin r + 2r = 0, r(0) = 0 and r(0)

= p.
(5.13)
The rst integral for this problem is

1
1 2
r cos r + r2 = p2 1.
2
2

162

Donald G Aronson and Hans G Othmer

(a)

(b)

1
2

1
2

(c)

(d)

Fig. 5.4. The (1 , 2 ) component of selected solutions in Fig. 5.3.

The solution corresponds to a heteroclinic orbit joining the equilibria (1 , 0)

if
4
p(1 ) = p1 = 2 2 cos (1 ) + 221 .

For < 1 the solution p() corresponds to a periodic orbit about the origin.
For these solutions the energy is given by
H(p) = p2 4.
For H = H1 := H(p1 ) the solution to (5.13) has innite period. As H (and
therefore p) is reduced, there is a value H2 (0, H1 ) such that the periodic
orbit is hyperbolic for H (H2 , H1 ) and elliptic for H < H2 . Moreover, a
new solution branch bifurcates from the antidiagonal solution at H = H2 .
The bifurcation diagram is given in Fig. 5.3 and selected solutions on the
bifurcating branch are shown in Fig. 5.4; the bifurcation at H = H2 is label
2 in Fig. 5.3. Solutions on the new branch connect the zero-velocity surfaces
about (, ) and (, ). Note that the period becomes innite as H  0,
and at H = 0 this branch seems to be generated by a concatenation of the

5 The Dynamics of SQUIDs and Coupled Pendula

163

1
Fig. 5.5. The long-time integration of certain initial conditions on the zero-energy
surface results in an intriguing petal structure. Color indicates the velocity of the
solution, where blue/green is slow and red is fast. (Courtesy of John Guckenheimer.)

heteroclinic orbits that connect the zero-velocity surfaces about (, ) to

the equilibrium point at (, ).
With further reduction of H a value H3 (0, H2 ) is reached (label 3
in Fig. 5.3) at which a degenerate period-doubling bifurcation occurs. This
bifurcation generates a branch of buttery-shaped orbits that undergo additional bifurcations and disappear in an innite-period orbit at H = 0; two
of these buttery-shaped orbits are shown in Fig. 5.4(a), and further bifurcated orbits are shown in Fig. 5.4(b). The bifurcation also generates a branch
of horseshoe-shaped orbits that are symmetric about the antidiagonal; see
Fig. 5.4(c) and (d). As we will see, the horseshoes play an important role in
the overall dynamics of the system. The horseshoe branch persists through

164

Donald G Aronson and Hans G Othmer

(a)

(b)

Fig. 5.6. Heteroclinic orbits in the (1 , 2 )-plane. Panel (a) shows an orbit in the
unstable manifold of the horseshoe orbit for = 0.036377. This orbit connects to the
stable manifold of the symmetric partner of the horseshoe orbit. Panel (b) connects
the unstable equilibrium (, 0, , 0) to the stable manifold of the horseshoe orbit
for = 0.03638.
2

(a)

(b)

(c)

Fig. 5.7. Connecting orbits between horseshoe orbits in neighboring cells. Here,
(, H) = (0.05, 0) in panel (a), (, H) = (0.0268297698, 0) in panel (b), and (, H) =
(0.05, 0.22049) in panel (c).

H = 0. There is a value H4 < 0 (label 4 in Fig. 5.3) at which there is a

saddle-node bifurcation leading to a second branch of horseshoes. The second
branch ends in an innite-period orbit at H = 0, where it appears to be the
concatenation of three hetero- or homoclinic orbits; two examples are shown
in Fig. 5.4(c). Note that for H (H4 , 0) there are two distinct branches of
horseshoes.
Figure 5.5 shows the result of a long-time integration in conguration space
of system (5.12) with = 0.01. The trajectory lies on the zero-energy surface
and has initial condition
1 (0) = ,

1 (0) = 0,

2 (0) = 3.141585,

5 The Dynamics of SQUIDs and Coupled Pendula

(a) 2

165

(b)

2
1
0
1
2
3

Fig. 5.8. Solutions to (5.12) on the zero-energy surface in the (1 , 2 )-plane. Panel
(a) shows a heteroclinic orbit for = 0.03638 that connects the unstable equilibrium (, 0, , 0) to its diagonal translate (, 0, , 0). Panel (b) shows a solution integrated for time T [0, 700] with = 0.01; the initial condition is
1 (0) = 2 (0) = 2.690233, and 1 (0) and 2 (0) are chosen so that the energy
is zero.

with 2 (0) determined by the requirement of zero energy. The unbounded

trajectory forms an intriguing petal structure. The equilibria where the two
pendula are in the upright position lie along the diagonal at the base of
the petals. In this motion each pendulum crosses the upright position and
changes its direction many times. The petals facing out from the diagonal of
the vine-like structure are created when one pendulum crosses the upright position while the other stops and reverses direction. The trajectory moves back
and forth along the diagonal in a seemingly erratic fashion. This and similar
trajectories are induced by the rich structure of connecting orbits joining the
horseshoes and the equilibria, as we now describe briey.
Figure 5.6(a) shows a heteroclinic connection between the unstable manifold of the horseshoe orbit for = 0.036377 and the stable manifold of its symmetric partner (i.e., its reection in the main diagonal). Figure 5.6(b) shows
a heteroclinic orbit that connects the unstable equilibrium (, 0, , 0) to
the stable manifold of the horseshoe orbit for = 0.03638. Figure 5.7 shows
connections between horseshoes in neighboring cells. Panels (a) and (b) show
two solutions in the zero-energy surface, while panel (c) shows a solution for
H = 0.22049. Figure 5.8(a) shows a heteroclinic connection between the unstable equilibrium (, 0, , 0) and its diagonal translate (, 0, , 0). These
connections provide the escape routes from cell to cell, but not all trajectories are ejected from their initial cells; many are simply trapped inside the
horseshoes, as shown in Fig. 5.8(b).

166

Donald G Aronson and Hans G Othmer

5.3 Rotations
In addition to equilibria and periodic solutions, system (5.5) may also have
running solutions or k-rotations. These are are solutions for which there exists
a time T > 0 such that j (t + T ) = j (t) + 2k for some integer k 1.
One can anticipate that rotations exist only for the appropriate relationship
between the damping and the applied torque in the dissipative case. If we
map the conguration space onto a cylinder then these solutions are periodic
with period T . To construct a k-rotation we solve the initial value problem
for system (5.6) with
s(0) = 0,

s(0)

= p2 ,

r(0) = p3 ,

and

r(0)

= p4 .

Solutions depend on the three state parameters p = (p2 , p3 , p4 ) and the three
system parameters , , I. For simplicity we regard the coupling strength
> 0 as xed and only deal with the two system parameters q = (, I). A
solution to the initial value problem, written in the form
[s(t; p, q), r(t; p, q)] ,
is a k-rotation if there exists a minimal T > 0 such that

s(T ; p, q) 2k = 0

s(T
; p, q) p2 = 0,
r(T
; p, q) p3 = 0,

r(T
; p, q) p4 = 0.

(5.14)

It is easy to see that in the Hamiltonian case (q = 0) there exist T0 >

0 and state parameters p0 such that [s(t; p0 , 0), r(t; p0 , 0)] is a k-rotation
with period T0 for each k 1. We dene P0 = (T0 , p0 , 0) and say that a
continuation of the solution [s(t; p0 , 0), r(t; p0 , 0)] to a neighborhood of P0 is
regular if there is a distinguished state parameter and a distinguished system
parameter such that the remaining state and system parameters are all smooth
functions of the distinguished ones in a full neighborhood of P0 .
It is an easy consequence of the Implicit Function Theorem that in-phase
rotations (k = 1) always have a regular continuation. Here, we consider the
general case, which is more complicated. The dierential of (5.14) at P0 is the
matrix
= (0 | | I ) ,
where

s sp2
s s p2
0 =
r rp2
r rp2

sp3
s p3
rp3
rp3


sp4 

s p4

rp4 
rp4 

,
P0

s
s

=
r ,
r

sI
s I

I =
rI .
rI

and

Here, 0 is the dierential of the Hamiltonian continuation problem, that is,

the problem of continuing (s, r)(P0 ) to a neighborhood of P0 in the subspace

5 The Dynamics of SQUIDs and Coupled Pendula

167

q = 0. For the Hamiltonian problem, conservation of energy provides a relationship between s, s,

r, and r so that, generically, only three of them are
independent, i.e., generically
rank(0 ) = 3.

(5.15)

When (5.15) holds then there is a regular continuation of (s, r)(P0 ) if there
exists
/ range(0 ),
(5.16)
span{ , I } such that
that is, can be chosen such that
rank(0 | ) = 4.
Let X(t) denote the fundamental matrix solution to the variational system
associated with (5.6) at (s, r)(P0 ). As is shown in [3], the dierential 0 then
becomes
0 = (T | 2 | 3 | 4 ),
where

s 
p02
s 


T =
0
r  =

p4

0
0

r P
2p3 sin p3
0

and the j for j = 2, 3, 4 are the corresponding columns of X(T0 ) Id. In

order to satisfy (5.15) and (5.16) there are two possibilities: either
rank(X(T0 ) Id) = 3

and

T range(X(T0 ) Id)

(5.17)

rank(X(T0 ) Id) = 2

and

T
/ range(X(T0 ) Id).

(5.18)

or
If (5.17) holds there is no distinguished state parameter and hence no regular
continuation. We note that this case was never observed in any of the numerical studies reported in [3]. On the other hand, possibility (5.18) is known to
occur. Suppose (5.18) holds and let {i1 , i2 , i3 } be a permutation of {2, 3, 4}
such that
range(X(T0 ) Id) = span{i1 , i2 }.
Then there is a Hamiltonian continuation of (s, r)(P0 ) with T , pi1 , and pi2
expressed as smooth functions of pi3 in a neighborhood of p0i3 . If, in addition,
(5.16) holds then there is a regular continuation in a neighborhood of P0 .
The choice of distinguished system parameter is arbitrary. Suppose, for instance, that (s, r) is a regular continuation of (s, r)(t; p0 , 0) with distinguished
system parameter . Then we have a relation of the form I = H(, ), where
H is a smooth function. However, k-rotations must satisfy the kinetic energy
relation

168

Donald G Aronson and Hans G Othmer

(a)

(b)

Fig. 5.9. Solutions of (5.12) for = 0.175 as a function of the kinetic energy for
I = 0 (a) and variable I (b). From D.G. Aronson, E.J. Doedel and H.G. Othmer,
The dynamics of coupled current-biased Josephson junctions II, Internat. J. Bifur.
c
Chaos Appl. Sci. Engrg. 1(1) (1991) 5166 1991,
with permission from World
Scientic Publishing; reprinted with permission.

I=
2k

It follows that

(s 2 + r 2 ) dt.
0


H 
> 0.
P0

Therefore, we can invert H to obtain = h(, I) and we get a regular continuation with I as the distinguished system parameter.

5 The Dynamics of SQUIDs and Coupled Pendula

169

Figure 5.9(a) shows some branches of 2-rotations with q = 0 and =

0.175; the plot shows the period T versus the kinetic energy. Figure 5.9(b)
shows the equivalent solutions for q = (0.01, I) with I variable. The bifurcation at label 21 in Fig. 5.9(a) does not persist as is increased since there is
a regular continuation at this point. The bifurcations to the right of label 21
do persist, because condition (5.16) fails in this case.

5.4 Stability and Bifurcations of the In-Phase Rotations

Most of the solution branches shown in Fig. 5.9(a) are 2-rotations for the
Hamiltonian system. The exception is the left-most branch of synchronous or
in-phase rotations. Here we investigate the stability and bifurcation properties of these solutions and their extensions to the damped/driven regime. To
simplify the analysis we scale the damping and the torque I together by
assuming that I = A for some xed A > 0. It is clear that system (5.5)
has a one-parameter family of rotations () dened by 1 (t) = (t) and
2 (t) = (t + ) for each R when = 0. It was shown in [5] that the only
member of this family that can be continued for = 0 is the in-phase rotation
0 (). In this section we discuss the stability of 0 () as the parameters
and are varied.
It is known that for xed A > 4/ and each > 0, the equation

+ sin = (A )

(5.19)

has a unique rotation solution for which > 0 [12]. If we translate time
so that (0) = 0 then there is a unique positive () > 2 such that the

rotation solution satises (0)

= (). As 0, we have () 0 , where
0 = 0 (A) > 2 is the unique solution of

2A =

2 2 + 2 cos d.

Note that 0 can have any value in the interval (2, ) depending on the choice
of A > 4/. We denote the rotation solution by () and its period by T ().
In order to determine the stability of the in-phase rotation () we must
nd the associated Floquet multipliers, which are the eigenvalues at t = T ()
of the fundamental matrix solution to the variational system associated with
(5.6) at (). For this purpose it is convenient to order the variables as
(r, u, s, v). Then we have to solve the system

0
1
0
0
cos () 2
0
0
V
V =
(5.20)

0
0
0
1

0
0 cos ()

170

Donald G Aronson and Hans G Othmer

subject to the initial condition V (0) = Id. System (5.20) decomposes into the
two 2 2 subsystems
2
3
0
1
X =
X,
X(0) = Id
(5.21)
cos () 2
2

and
X =

3
0
1
Y,
cos ()

Y (0) = Id.

(5.22)

Subsystem (5.22) determines stability with respect to the in-phase subspace

and subsystem (5.21) determines stability with respect to the orthogonal com )
T is a T ()-periodic
plement of this subspace. It is easy to see that Y = (,
solution of (5.22). Therefore, the Floquet multipliers associated with subsystem (5.22) are 1 and exp ( T ()), regardless of the value of . Thus, to
determine the stability of () it suces to study the 2 2 system (5.21).
Let (t, , ) denote the fundamental matrix solution to (5.21). Then the
Floquet multipliers are the eigenvalues of (T (), , ), that is, the roots of
2 (, ) + exp ( T ()) = 0,
where

(, ) = trace( (T (), , ).

Therefore, the multipliers are

1

1.
2 4 exp ( T ()) ,
=
2

and it follows that () is stable if || < 1 + exp ( T ()) and unstable if

|| > 1 + exp ( T ()).
When 0, the second-order equation associated with system (5.21)
reduces to a Hill equation
x
+ (2 q(t)) x = 0,

(5.23)

where the potential is given by

q(t) = cos ()(t)|=0 .
Moreover,

(5.24)

.
1
(, ) 0 () = 1 (T0 ) + 2 (T0 ) ,

where 1 and 2 are solutions to (5.17) that satisfy 1 (0) = 2 (0) = 1 and
1 (0) = 2 (0) = 0, and T0 = T (0).
The general theory for Hills equation [11] shows that there exists a sequence of eigenvalues
< 0 < 1 2 < 3 4 <

5 The Dynamics of SQUIDs and Coupled Pendula

0 ()

(a)

171

(b)

Fig. 5.10. The graph of 0 () (a) and the locus of period-doubling bifurcations from
() for A = 5/3 (b). The in-phase rotation is unstable in the hatched region. From
E.J. Doedel, D.G. Aronson and H.G. Othmer, The dynamics of coupled currentc
biased Josephson junctions I, IEEE Trans. Circ. Sys. 35(7) (1988) 810817 1988
by IEEE; reprinted with permission.

with j as j , such that

|0 | < 2 and 0 < 0
|0 | < 2 and 0 > 0
0 > 2
0 < 2

on (0 , 1 ) (4 , 5 ) ,
on (2 , 3 ) (6 , 7 ) ,
on (, 0 ) (3 , 4 ) (7 , 8 ) ,
on (1 , 2 ) (5 , 6 ) .

(5.25)

At = 0 system (5.21) reduces to (5.22), which has a T0 -periodic solution.

Thus, 0 is an eigenvalue of (5.23) with T0 -periodic boundary conditions, and
we know from previous remarks that the associated eigenfunction is strictly
positive. It follows from Sturm-Liouville theory that 0 is the smallest eigenvalue for this problem and, therefore, 0 = 0. Consequently, 0 () > 2 for all
< 0.
According to Goldbergs theorem [7], equation (5.23) has exactly one nite
interval of instability if and only if the potential q is periodic and integrable,
and satises
(5.26)
q = 3q 2 + q +
for some constants and , that is, if and only if q is an elliptic function.
In the present case, since (0) is a rotation it follows from (5.24) that q is
periodic. Moreover, using (5.19) with = 0 and its rst integral, one can
verify that (5.26) is satised. Thus, there is precisely one nite interval of
instability for > 0, and the numerical computations performed in [5] show
that this interval is (1 , 2 ). It follows that 2j1 = 2j for all j > 2 and
that |0 ()| = 2 for = 2j with j > 1. The graph of 0 () is shown in
Fig. 5.10(a).
The second-order equation corresponding to (5.21) is
x
+ x + (2 + cos ())x = 0.

(5.27)

172

Donald G Aronson and Hans G Othmer

Fig. 5.11. The bifurcation diagram for (5.6), represented as the period T versus
for A = 5/3. From E.J. Doedel, D.G. Aronson and H.G. Othmer, The dynamics of
coupled current-biased Josephson junctions I, IEEE Trans. Circ. Sys. 35(7) (1988)
c
810817 1988
by IEEE; reprinted with permission.

(a)

(b)

Fig. 5.12. Details of the bifurcation diagram shown in Fig. 5.11. The points in
panel (a) labeled 1, 4, and 5 are transcritical bifurcations, while those labeled 2
and 3 are period-doubling bifurcations. There are two regions of stability along
the asynchronous branch that bifurcates at label 3. Panel (b) shows a blow-up
of the region around label 5 in panel (a). From E.J. Doedel, D.G. Aronson and
H.G. Othmer, The dynamics of coupled current-biased Josephson junctions I, IEEE
c
Trans. Circ. Sys. 35(7) (1988) 810817 1988
by IEEE; reprinted with permission.

If = 0 then x1 (t) = ()(t) is a strictly positive T ()-periodic solution to

(5.27) for any > 0. Thus, = 0 is an eigenvalue of (5.27) for any > 0, and
it is a simple eigenvalue because

5 The Dynamics of SQUIDs and Coupled Pendula

x2 (t) = ()(t)


0

173

e
d
[ ()( )]2

is a linearly independent non-periodic solution. Using these solutions one can

construct the fundamental matrix solution to (5.21) and show that (0, ) <
0. It follows that the innite instability interval (, 0) remains invariant
for > 0 and that () becomes stable as increases through 0. Numerical
computations show that there is a vertical bifurcation from () at = 0;
the numerical results are described in more detail below.
By continuity, the unique instability interval (1 , 2 ) for = 0 persists for
suciently small > 0. The numerical computations done in [5] strongly
suggest that the remaining eigenvalues 2j disappear for > 0. Consequently, for suciently small > 0 the rotation solution () is unstable
for (, 0) (1 (), 2 ()) and stable otherwise. Furthermore, there ex) = 2 (
), and () is unstable on R and
ists = (A) > 0 such that 1 (
+
asymptotically stable on R whenever > ; see Fig. 5.10(b). Note that the
bifurcations at = j (), j = 1, 2, are period-doubling bifurcations, because
the multiplier passes through 1; see Fig. 5.10(a). In the Hamiltonian case
these two bifurcations project onto the point labeled Period Doubling in
Fig. 5.9(a).
We now discuss the numerically computed bifurcation behavior in the
interval (1 , 2 ) for the value = = 0.15 used in [10]. Then the perioddoubling bifurcations from () are at 1 = 0.1275 and 2 = 0.6132. Solution
branches that bifurcate from ( ) at = 1 and = 2 are shown in
Fig. 5.11 with enlarged views given in Fig. 5.12. Rotations that correspond to
some of the labels in Fig. 5.12 are shown in Fig. 5.13.
All rotations in Figs. 5.11 and 5.12 have winding number 2, so that
i (T ) i (0) = 4, where T is the integration time. The bifurcating branches
then connect continuously to the horizontal branch ( ). The solutions with
labels 2 and 3 in Fig. 5.12(a) denote the two period-doubling bifurcations from
( ). The branch that emanates from label 2 terminates at an orbit of innite period. The same holds for the branch that emanates from label 3.
The solution with label 6 in Fig. 5.12(b) can be thought of as an approximation to the innite-period orbit that terminates the branch; it is shown in
Fig. 5.13(a). The innite-period orbit is a double homoclinic loop, i.e., an
orbit that passes through the same saddle point twice. Note that this branch
contains two regions of stable rotations. The solutions with labels 4 and 5
in Fig. 5.12(a) are secondary transcritical bifurcations, not period-doubling
bifurcations. The bifurcating tertiary branches from labels 4 and 5 also terminate in innite-period orbits past the solutions labeled 7 and 8, respectively;
compare Figs. 5.13(b) and (c). These branches also contain stable portions
and further bifurcations that are not shown in Figs. 5.11 and 5.12.
The oscillating branch of solutions in Fig. 5.11 is shown in a blow-up in
Fig. 5.12(b); it contains the solutions labeled 9, 10, and 11, which are plotted in Figs. 5.13(d), (e), and (f), respectively. This branch terminates in an

174

Donald G Aronson and Hans G Othmer

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 5.13. Some rotations corresponding to the labels in Fig. 5.12. From E.J. Doedel,
D.G. Aronson and H.G. Othmer, The dynamics of coupled current-biased Josephson
c
junctions I, IEEE Trans. Circ. Sys. 35(7) (1988) 810817 1988
by IEEE; reprinted
with permission.

innite-period orbit at both end points of the interval in which it exists. The
solutions with labels 9 and 11 can be considered as approximations to these
orbits. Solution 9 approximates an innite-period orbit containing two distinct
unstable equilibria, each of which has one complex conjugate pair of eigenval-

5 The Dynamics of SQUIDs and Coupled Pendula

175

ues; see Fig. 5.13(d). Solution 11, at the other end of the branch, is a double
homoclinic loop; see Fig. 5.13(f). Along this branch there are small intervals
of stable behavior near the limit points. For example, one such stable interval
is near the solution with label 10 on the lower part of the branch. These stable
regions are so small that they cannot be distinguished in Fig. 5.12(b). They
are bounded by bifurcations that lead to more complicated solution types.
Many of these more complicated, stable solutions can be observed numerically by careful choice of initial data and accurate integration. For example,
a stable rotation such as solution 10 in Fig. 5.13(e) can be obtained by accurately choosing initial data near solution 10 in Fig. 5.12(b), on the small
portion of the branch that consists of stable rotations.

5.5 Conclusions
We presented an overview of the dynamics of a model of a ring of superconducting material that contains two Josephson junctions, which is known as
a SQUID. The resulting system equations are identical (in non-dimensional
form) to the equations modeling two pendula that are coupled by a linear
torsional spring or bar and forced with an applied torque. Our analysis involved intensive use of Auto to construct bifurcation diagrams and rotating
solutions that can only be found explicitly for the Hamiltonian case.
A complex bifurcation structure organizes the existence of rotation solutions. We focused particularly on the case of xed coupling parameter and
varying damping coecient and forcing I. We found the conditions under
which in-phase rotation solutions that exist for = 0 persist; the resulting
branch leads to a series of bifurcating branches. In particular, there is a region
of relatively small values of and in which the in-phase rotation is unstable.
Maginu [10] was the rst to observe the instability for intermediate values of and suitable and I. His numerical studies indicate the presence of
chaos in the unstable range. The results discussed here provide a more detailed, though still incomplete, understanding of the transitions in dynamics
suggested in [10].

Acknowledgments
All of the work described in this chapter was performed in close collaboration
with Sebius Doedel and we thank him for his contributions. We also thank
`
John Guckenheimer, Angel
Jorba, and Bj
orn Sandstede for their contributions
to the material on Hamiltonian Systems in Sect. 5.2. Furthermore, we are
grateful to World Scientic Publishing and IEEE for permission to reproduce
previously published material from [3] and from [5], respectively.

176

Donald G Aronson and Hans G Othmer

References
1. E. Altshuler and R. Garcia. Josephson junctions in a magnetic eld: Insights
from couppled pendula. Am. J. Phys., 71(4): 405408, 2003.
2. D. G. Aronson, E. J. Doedel, J. Guckenheimer, A. Jorba, and B. Sandstede. On
the dynamics of torsion-coupled pendula. Unpublished, 2002.
3. D. G. Aronson, E. J. Doedel, and H. G. Othmer. The dynamics of coupled
current-biased Josephson junctions II. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 1(1): 5166, 1991.
4. V. N. Belykh, N. F. Pedersen, and O. H. Soerensen. Shunted-Josephson-junction
model. II. The nonautonomous case. Physical Review B, 16(11): 486071, 1977.
5. E. J. Doedel, D. G. Aronson, and H. G. Othmer. The dynamics of coupled
current-biased Josephson junctions I. IEEE Trans. Circ. Sys., 35(7): 810817,
1988.
6. L. Finger. The Josephson junction circuit family: Network theory. Int. J. Circ.
Th. Appl., 28(4): 371420, 2000.
7. W. Goldberg. Necessary and sucient conditions for determining a Hills equation from its spectrum. J. Math. Anal. Appl., 55: 549554, 1976.
8. B. D. Josephson. Supercurrents thorough barriers. Phys Letts, 1: 201, 1962.
9. Konstantin K. Likharev. Dynamics of Josephson Junctions and Circuits. Gordon and Breach Publishers, New York, 1986.
10. K. Maginu. Spatially homogeneous and inhomogeneous oscillations and chaotic
motion in the active Josephson junction line. SIAM J. Appl. Math., 43(2):
225243, 1983.
11. W. Magnus and S. Winkler. Hills Equation. Dover Publ., New York, 1979.
12. N. Minorsky. Nonlinear Oscillations. Robert E. Krieger Publishing Co., Malabar, Fla., 1983.
13. F. M. A. Salam and S. S. Sastry. The Complete Dynamics of the Forced Josephson Junction Circuit: The Regions of Chaos. In J. Chandra, editor, Chaos in
Nonlinear Dynamical Systems, pages 4355. SIAM, Philadelphia, 1984.

6
Global Bifurcation Analysis in Laser Systems
Sebastian M Wieczorek
Mathematics Research Institute, University of Exeter, United Kingdom

Nonlinear dynamics of lasers has been a lively theoretical and experimental eld since the invention of the laser in 1960. Its focus in the last two
decades have been instabilities in widely used semiconductor lasers. Nonlinear studies of laser systems contributed to the eld of dynamical systems
with general phenomena including chaos, (chaotic) synchronization of coupled
oscillators, competition, excitability, delay-induced instabilities, unfolding of
high-codimension bifurcations, bifurcation cascades, and spatial patterns; see
[1, 28, 30, 36, 51, 63] for general reading and further references. These studies also deepened the understanding of nonlinear phenomena that are important for technological applications, e.g. external-modulation response of
semiconductor lasers for faster Internet connections [57]. Furthermore, nonlinear analysis of laser systems stimulated and helped to validate the feasibility
of novel, chaos-based applications including secure communication schemes
[4, 50], chaotic radars [34], and instability-based laser sensors [56].
Much of the recent progress in the eld of laser dynamics is owing to the
application of numerical continuation techniques. The study of lasers with
tools from bifurcation theory started already in 1987 with the work of Maloney and coworkers on the nonlinear dynamics of three-level molecular lasers
[37]. By now, there are over one hundred publications where tools from bifurcation theory are used to investigate dynamics of various laser systems.
To explain the strong impact that numerical continuation techniques had and
are still having on the eld of nonlinear laser dynamics we mention here four
key properties that we found to be very inuential in our research. Namely,
numerical continuation techniques:
1. facilitate immensely the systematic search of an extensive and manydimensional parameter space to identify the important contributions. Interesting phenomena may be missed by the traditional approach of simulation of the governing equations by direct time integration. Indeed, numerical continuation supplements numerical simulation by enabling parameter

178

Sebastian M Wieczorek

studies to be performed to the necessary detail and accuracy, and with

relatively modest computational resources;
2. allow the study of global homoclinic and heteroclinic bifurcations, which
are often associated with interesting nonlinear eects but cannot be studied otherwise;
3. supplement and expand analytical bifurcation studies, which are themselves invaluable but generally restricted to small neighborhoods of parameter space, for example, near isolated individual bifurcations. Numerical
bifurcation analysis is the tool of choice for those problems that cannot be
addressed with analytical techniques but are of great importance to physicists, chemists, biologists, and engineers. For example, using the analytical
results on bifurcation curves expected near a codimension-two bifurcation
point, continuation techniques give answers to questions such as where
these bifurcation curves go and to which other codimension-two bifurcation points they connect. Do phenomena occupy regions of the parameter
space that can be experimentally detectable or that are of any practical interest? From a more general bifurcation theory point of view, this
question concerns a better understanding of the organizing properties of
bifurcations;
4. may actually stimulate real laser experiments where bifurcation diagrams
are used as road maps to guide experimentalists through the complexity
and variety of nonlinear laser dynamics.
The aim of this chapter is to give a taste of how continuation techniques can be
used to understand complicated dynamics in laser systems. At the same time,
laser systems emerge as natural candidates to study how global bifurcation
phenomena manifest themselves in a real system. In this sense, this chapter
should also be seen as a contribution to the wider eld of dynamical systems.
Specically, we present here the following two concrete examples of a global
bifurcation analysis in semiconductor laser systems.

In Sect. 6.1 we consider structures of global n-homoclinic bifurcations that

lead to the phenomenon of multi-pulse excitability in semiconductor lasers
with optical injection [58, 62, 64]; and
in Sect. 6.2 we present the backbone of the bifurcation set for two backto-back coupled lasers in which we nd the counter-intuitive appearance
of chaos at practically vanishing coupling [53, 54].

Both examples feature interesting global bifurcation structures that have a

physical meaning and actually stimulated real laser experiments. The word
global in this context refers to objects in phase space that are due to certain
arrangements of stable and unstable invariant manifolds, as well as to associated bifurcation structures in parameter space. In terms of the parameter
space we use the physically motivated approach of calculating k-parameter
bifurcation sets (usually for k = 2) for several xed values of an additional
(k + 1)st parameter. These k-dimensional bifurcation sets, which consist of

6 Global Bifurcation Analysis in Laser Systems

179

various local and global bifurcations, are slices that are inuenced or even
determined by unfoldings of certain codimension-(k + 1) bifurcations or singularities. These so-called organizing centers provide links between various
types of bifurcations that appear at rst glance to be (and are often thought
of as) unrelated. While analytical techniques typically force investigators to
focus on particular bifurcations, which imposes some sense of isolation of the
particular phenomenon under investigation, numerical bifurcation analysis allows one to connect seemingly unrelated pieces. The goal is to get to a deeper
understanding of the dynamics by obtaining a consistent and global bifurcation picture of the given (laser) system; see also Chaps. 2 and 7.

6.1 Multi-Pulse Excitability and n-Homoclinic Orbits in

an Optically Injected Laser
This section is based on [58] and describes intricate structures of n-homoclinic
orbits and their bifurcations in the rate equation model of an injection
laser. The analysis reveals how codimension-two and -three homoclinic bifurcations act as organizing centers of the bifurcation diagram. First, we
nd heteroclinic cycles known as T-point bifurcations; we are dealing here
with the case that both saddles involved have a pair of complex conjugate
eigenvalues. Such T-point bifurcations were found in systems from applications [2, 19, 20, 22, 26, 32, 47, 68] and their unfolding is known to involve
n-homoclinic orbits for any n [9, 10]. Secondly, we nd double-homoclinic orbits to a saddle-focus, where there are two dierent homoclinic connections
to a single saddle-focus. (This should not be confused with a 2-homoclinic orbit.) This codimension-two global bifurcation has been studied in an abstract
setting in [27, 40, 44]. The bifurcations of 1-homoclinic orbits are known,
but the possible unfoldings are not yet fully understood. We present sketches
of relevant bifurcation curves associated with these global bifurcations and
show with numerical bifurcation diagrams how they manifest themselves in
the optically injected laser model. Also, we explain how these n-homoclinic
bifurcations give rise to the phenomenon of multi-pulse excitability.
All curves of global bifurcations and the associated homoclinic and heteroclinic orbits were calculated with the HomCont [11, 12] part of the continuation package Auto [16]; the invariant manifolds and time series illustrating
multi-pulse excitability were computed with the package DsTool [5].
6.1.1 Optically Injected Laser
From the dynamical systems point of view, a free-running class-B laser (the
active-medium polarization decays much faster than the population inversion
and electric eld) is a damped nonlinear oscillator characterized by a stable
equilibrium with two complex-conjugate eigenvalues. This situation can be
changed drastically when the laser is subjected to an external optical signal,

180

Sebastian M Wieczorek

injected plane wave

laser
active medium

laser output ?

Fig. 6.1. A laser that is being injected with an external optical eld.

as discussed here and sketched in Fig. 6.1. With optical injection it becomes
a driven nonlinear oscillator: one of the nicest physical systems to show a
fascinating array of nonlinear dynamics. Several kinds of complex and chaotic
dynamics were discovered; see, for example, [3, 17, 21, 35, 29, 59, 63]. Of
particular importance is the fact that this system is very well described by a
set of three autonomous ordinary dierential equations for the complex electric
eld E = Ex +iEy and the population inversion n (the number of electron-hole
pairs in the case of a semiconductor laser) [48, 59]. These so-called single-mode
rate equations for this system can be written in dimensionless form as



E = K + 12 (1 + i)n i E
(6.1)
n = 2 n (1 + 2Bn)(|E|2 1) .

The two main parameters are the injected eld amplitude K and the detuning
, the frequency dierence between the injected light and the frequency of the
laser without injection. The explicit time dependence in the drive term proportional to K was eliminated thanks to the S 1 symmetry of the system [63].
While K and can easily be changed in an experiment, the parameters B,
and describe material properties of a given laser. Specically, B is the
rescaled lifetime of photons in the laser cavity and is the rescaled damping
rate of the so-called relaxation oscillations, which are an exchange of energy
between the electric eld E and the population n of a characteristic frequency
R in a free-running laser. We use the realistic values B = 0.015 and = 0.035
throughout in our study.
The material constant , called the linewidth enhancement factor, can
be very dierent for dierent lasers, and it is known that changing has a
very large eect on the dynamics of the injected laser [59]. The parameter
describes the coupling between the phase and the amplitude of the electric
eld E, and it is in the range of [1, 10] for typical semiconductor lasers.
On the other hand, (6.1) for = 0 models injected solid-state and CO2 lasers,
which have a negligible phase-amplitude coupling. This is our motivation for
studying how the bifurcation set in the (K, )-plane depends on , that is,
on the main material property of the particular laser under consideration.

6 Global Bifurcation Analysis in Laser Systems

1

1
0
s 0
1

a 0
1

181

4a

0
s1
0
1
1
a0
1
0

s
a

2
4b

het

4b

B1 1
0

ns

0
1

B2

1
0
0
1

h1
4a

2
Sg

1
0

A1

het

1
0
0
1

Sl

A2

Sg

1
Sl
s00
11

Sg

h1

het

00
s11
00
11

11
00
s
00
11
1
0
0a
1

s
a

Fig. 6.2. Sketches of phase portraits near the boundary of a homoclinic tooth in two
unfolding parameters 1 and 2 . From S. Wieczorek and B. Krauskopf, Bifurcations
of n-homoclinic orbits in optically injected lasers, Nonlinearity 18(3) (2005) 1095
c 2005 by Institute of Physics Publishing; reprinted with permission.
1120 

6.1.2 Homoclinic Teeth

The phenomena we are interested in appear in what we call homoclinic teeth.
What we mean by a homoclinic tooth is sketched in Fig. 6.2. It is the region
bounded by the curve h1 of a 1-homoclinic bifurcation and the grey part
of the curve Sl of local saddle-node bifurcations. The two curves meet at
two points A1 and A2 of codimension-two non-central saddle-node homoclinic
bifurcations. This codimension-two bifurcation was identied in [31] as an
organizing center for multi-pulse excitability (single pulse excitability in this
case of a curve of 1-homoclinic bifurcations); see [6, 14, 15] for more details
on its unfolding. Figure 6.2 shows sketches of phase portraits for dierent
locations of parameter space near the homoclinic tooth. Notice, in particular,
that the saddle-node bifurcation takes place on a periodic orbit along the
parts marked Sg (where g stands for global), but this is not the case along Sl
(where l stands for local).
In the part of region 3 that is close to the curves h1 and Sg the laser is 1excitable: a small perturbation to above the stable manifold of the saddle s will
lead to a large excursion, which follows the one-dimensional unstable manifold

182

Sebastian M Wieczorek

of s, before the laser relaxes back to the attractor a. By comparison, in region

2 there is a smooth invariant circle, but the laser is still excitable close to the
curve Sg : a perturbation beyond the stable manifold of s will lead to a large
excursion around the invariant circle and back to a. In fact, phase portraits 2
and 3 are topologically equivalent. However, away from the codimension-two
points A, they may relate to dierent physical phenomena. On the one hand,
phase portrait 2 represents phase locking because the smooth invariant circle
is centered at the origin of the complex E-plane. Hence, an excitable response
associated with phase portrait 2 is mainly of the form of a 2 phase slip with
only slight variations in the electric eld intensity. On the other hand, the
upper branch of the unstable manifold in phase portrait 3 evolves away from
the origin of the complex E-plane so that an excitable response leads to a
short (30 ps in our case) and distinct intensity pulse. Near A, the dierence
in the excitable response for the phase portraits 2 and 3 disappears. The
curve het is not a bifurcation curve, but when crossing it there is a change
of the direction from which the relevant branch of the unstable manifold of
the saddle approaches the attractor. Consequently, the closure of the unstable
manifold of the saddle is a smooth curve in region 2, while this is not the case
in region 3; see [31] for more details.
The homoclinic tooth is shown to intersect with the dashed curve ns where
the saddle is neutral, that is, the absolute values of the real parts of the real
eigenvalue and of the pair of complex conjugate eigenvalues are equal. What
the dynamics looks like inside the tooth crucially depends on whether one
is above or below ns. Along the parts of h1 below ns, often called a simple
Shilnikov case, the homoclinic orbit bifurcates into an attracting periodic orbit [Fig. 6.2(4a)]. On the other hand, along the parts of h1 above ns, often
called a chaotic Shilnikov case, the bifurcating periodic orbit is no longer
stable [Fig. 6.2 (4b)]. Breaking this type of homoclinic orbit leads to the creation of n-homoclinic orbits for any n. While the curve ns is not a bifurcation
curve, each of its intersection points B1 and B2 with h1 is a codimension-two
homoclinic bifurcation, known as a Belyakov point [7, 24]. Belyakov points
mark the transition between the two cases of homoclinic orbits and, hence,
give rise to an intricate structure of n-homoclinic orbits.
How homoclinic teeth arise in (6.1) is shown in Fig. 6.3 with panels of
the (K, )-plane of (6.1) near the locking region for increasing values of as
indicated. It shows the curves S of saddle-node bifurcations and the curves
H of Hopf bifurcations (both gray), the supercritical parts of which bound
the locking region of the injected laser; see [59]. Also shown is the neutral
saddle curve ns. All these curves are given by local conditions at equilibria of
(6.1) and can be found analytically. The curve h1 of 1-homoclinic bifurcations,
on the other hand, cannot be found analytically. It was computed with the
Auto/HomCont. The computations do not distinguish between a generic
(codimension-zero) homoclinic connection along the parts Sg in Fig. 6.2 and
the codimension-one homoclinic bifurcation along h1 . In other words, when the

6 Global Bifurcation Analysis in Laser Systems

183

bold black curve in Fig. 6.3 coincides with S then the saddle-node bifurcation
takes place on a periodic orbit. If it leaves S we nd a homoclinic tooth.
For = 0 (the case of a solid-state or CO2 laser) the (K, )-plane is
symmetric and there are no homoclinic teeth. As is increased, homoclinic
teeth start to grow along the saddle-node bifurcation curve S that forms the
lower boundary of the locking range. (The other boundary is the Hopf bifurcation curve H.) Initially the teeth are quite small [panels (b)(c)] but then
they grow in size with and the bifurcation diagram changes qualitatively,
showing the existence of codimension-three phenomena. At = 1.21 [panel
(c)] the rst tooth starts to intersect the neutral saddle curve ns. What is
more, new teeth start to appear between already present teeth [panel (e)].
All teeth keep growing, and the tooth closest to the saddle-node Hopf point
G1 develops a rather bizarre shape [panels (f)(i)]. On top of this, when
increases neighboring teeth may merge, meaning that the curve h1 detaches
from the curve S. This occurs at codimension-three points when two neighboring non-central saddle-node homoclinic bifurcation points come together
and vanish. Furthermore, one notices the appearance of codimension-two homoclinic bifurcation points (dots along the curve h1 in panels (h) and (i)).
They are created when the section given by xed crosses a minimum in the
respective codimension-two bifurcation curve, which is discussed in detail in
Sect. 6.1.5.
To study how new teeth are born and neighboring teeth merge we continued with HomCont the curve of codimension-two non-central saddle-node
homoclinic bifurcations in (K, , )-space [6, 45]. The projection of this curve
onto the (, )-plane is shown in Fig. 6.4(a), while Fig. 6.4(b) shows a sketch
of a non-central saddle-node homoclinic orbit. The left-hand fold points of
the curve in Fig. 6.4(a) are points where teeth are born, while right-hand fold
points are points where two neighboring teeth merge. This gure clearly shows
that there are no teeth for < 0.5. New teeth are then born one-by-one as
is increased. Secondary teeth appear from about = 2 on. Merging teeth can
be observed from about = 2.2 onward when the rst two teeth (nearest G1 )
merge. Successively teeth for larger negative detuning also merge. In fact
for > 7.5 there appears to be one giant tooth, if one still wants to call it
that. It is already clear that the situation becomes increasingly complicated
with .
6.1.3 Complex Structure of n-Homoclinic Bifurcations
Complex structures of global homoclinic and heteroclinic bifurcations arise
inside the homoclinic teeth as a result of interactions of the curves of 1homoclinic orbits. The fact that the curve ns intersects the rst homoclinic
tooth, for example, for = 2.0 in Fig. 6.3 (d), giving rise to two Belyakov
points, already allows us to conclude from general theory [7, 24] that there
must be further curves of n-homoclinic orbits. We remark that the exact
combinatorics of these n-homoclinic orbits is still not fully understood [24].

184

Sebastian M Wieczorek

1.5

G1
-0.5

H
-1

-0.5

G2

G1

-1.5

(a)
= 0.0
0.25

h1

-1
-1.5

h1

0.5
0

G1

0.5

0.75

h1

(b)
= 1.21

1.25

1.75

1.5

0.25

ns

0.5

0.75

1.25

1.5

1.75

S
H

-1

h1

K
H

G1

h1

-1

ns

S
-2

-2

ns
-3

-3

(c)
= 2.0
0.2

0.4

0.6

0.8

1.2

1.4

(d)
= 3.0
0.2

0.4

0.6

0.8

1.4

1.2

h1

G1
0

h1
0

G1

-1

-1

S
-2

H
-3

ns

(e)
= 4.0
0.2

0.4

0.6

0.8

-2

ns
-3

1.2

(f)
= 6.0
0.2

0.4

0.6

0.8

1.2

Fig. 6.3. Homoclinic teeth in the locking region as a function of . From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically injected
c 2005 by Institute of Physics Publishlasers, Nonlinearity 18(3) (2005) 10951120 
ing; reprinted with permission.

6 Global Bifurcation Analysis in Laser Systems

185

(a)

(b)

Sh

-1

-2

-3

Fig. 6.4. Curve of codimension-two saddle-node homoclinic bifurcations, projected

onto the (, )-plane (a), and a sketch of a codimension-two saddle-node homoclinic
orbit (b). From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits
c 2005 by Institute
in optically injected lasers, Nonlinearity 18(3) (2005) 10951120 
of Physics Publishing; reprinted with permission.

The question is how these n-homoclinic orbits are organized inside the
homoclinic teeth. At the same time, we obtain an impression of a Belyakov
bifurcation in a concrete system. Furthermore, one may ask where the associated n-homoclinic bifurcation curves go and to which other codimension-two
points they connect. In short: what is the bifurcation diagram, as far as one
can assemble it? These questions cannot be addressed by analytical studies
in a neighborhood near codimension-two points but they require the use of
continuation techniques. From a bifurcation theory point of view, this is the
next step towards the understanding of the organizing properties of global
bifurcations. Physically, we reveal structures that stretch over large regions in
the parameter plane and become experimentally accessible, that is, potentially
relevant for real applications of optically injected lasers.
Figure 6.5 (a1) shows curves hn of n-homoclinic orbits for n 4 inside
the rst tooth for = 2.0, while Fig. 6.5 (a2) is an enlargement near the
saddle-node bifurcation curve S. Many of these curves extend from the region
above ns to below ns and in crossing ns have further Belyakov points on them.
The picture that emerges is that of a complicated arrangement of nested nhomoclinic bifurcation curves. Most interestingly, several curves extend to
very near the curve S, and some even attach to S at points of non-central
saddle-node n-homoclinic orbits.
We now focus on what happens to the innite number of hn -tongues when
the Belyakov points are gone, that is, the homoclinic tooth is entirely below
the curve ns. One straightforward scenario would be that all the hn curves
disappear when B1 and B2 merge. However, this is not the case here. Fig-

186

Sebastian M Wieczorek

-0.7

-0.92

ns

-0.8

-0.93

h1
S

-0.94
-0.9

h1
h3

-1

-0.95

h2

0.3

h3
h2

h4

-0.96

h3

-0.97

-1.1

-1.2

h4

(a1)
= 2.0

(a2)
= 2.0
0.7

0.6

0.5

0.4

0.8

0.45

0.425

h1

-0.95
-1

h2

-1.025

h2

-1.05

K 0.55

h3

-1

h2

0.525

ns

-0.975

-0.9

0.5

0.475

h1

-0.95

ns

-0.85

h2

h2
-1.05

-1.1
-1.15

-1.075

(b1)
= 1.21
0.6

0.65

0.7

0.75

0.8

0.85

-0.85

(b2)
= 1.21
0.625

0.65

0.675

0.7

0.725

0.75

0.775

ns

h1

-0.96

-0.9

h3

-0.98

-0.95

h3

h1

-1

-1
-1.05

h2

-1.02

h2

h2

-1.04
-1.1
-1.15

-1.06

(c1)
= 1.0
0.6

(c2)
= 1.0
0.65

0.7

0.75

0.8

0.85

0.7

0.72

0.74

0.76

0.78

Fig. 6.5. Bifurcation curves of n-homoclinic curves inside a homoclinic tooth for
three dierent values of as indicated in the panels. From S. Wieczorek and B.
Krauskopf, Bifurcations of n-homoclinic orbits in optically injected lasers, Nonlinc 2005 by Institute of Physics Publishing; reprinted
earity 18(3) (2005) 10951120 
with permission.

6 Global Bifurcation Analysis in Laser Systems

187

ure 6.5 (b1)(b2) shows the rst tooth for = 1.21, just as it touches the
curve ns. This is a codimension-three phenomenon in (K, , )-space where
two Belyakov points coincide and then disappear when is decreased, as is
shown in Fig. 6.5 (c1)(c2). (The curve B of Belyakov points in (K, , )space has a minimum.) Even though the tooth is well below the curve ns for
= 1.0, there are still curves of n-homoclinic orbits inside it. In particular,
we nd that the curves h2 and h3 are attached to S.
Our numerical investigation suggests that there are only nitely many
curves of n-homoclinic orbits for < 1.21. To illustrate how subsequent curves
hn appear with increasing we marked one of them with a star. For = 1.0
[Fig. 6.5 (c2)] h2 is the last homoclinic curve that just emerged from the
saddle-node bifurcation curve S. As is increased above = 1.0, the curve h2
develops two extra noncentral-homoclinic points on S, forms a sort of bridge,
and provides space for the next homoclinic curve to emerge [Fig. 6.5 (b2)].
This process seems to repeat, such that for > 1.21 there exist innitely
many curves hn .
6.1.4 Multi-Pulse Excitability
The regions bounded by h2 and h3 near S appear to be large enough to be
experimentally accessible [60]. In such a region the laser exhibits multi-pulse
excitability. We remark that our study shows that this phenomenon can be
found even for surprisingly low values of ; see also [31, 62]. An example is
shown in Fig. 6.6 for = 1.0. The phase portrait in Fig. 6.6 (a1) is as that
of region 3 in Fig. 6.2 the laser is 1-pulse excitable. A small perturbation
above the excitability threshold, given by the stable manifold of the saddle
point, results in the laser sending out a single pulse; see Fig. 6.6 (a2). In the
region bounded by h2 , on the other hand, the phase portrait is close to a
2-homoclinic orbit and the laser produces two pulses in reaction to a single
perturbation; see Fig. 6.6 (b1)(b2). Finally, three pulses result in the region
bounded by the curve h3 , as is illustrated in Fig. 6.6 (c1)(c2). Indeed, it
is possible to nd n-pulse excitability for any n, but the regions for n > 4
become impractically small.
It is important to note a key ingredient for multi-pulse excitability to
occur, namely the fact that the respective curve hn extends all the way below
ns. For the parameters above ns the hn -tongues are so narrow that they
become hard to distinguish, even numerically. Furthermore, there exist an
innite number of unstable periodic orbits in the phase space for parameters
outside the tongues. As a result, the excitable response is often irregular and
unpredictable as the trajectory bounces between the unstable orbits before it
decides to return to the stable equilibrium. On the other hand, below ns the
tongues are easily distinguishable and the phase portraits are simpler as there
are no unstable periodic orbits. Consequently, the system can be prepared
to be well within hn (certainly for n 3) where the excitable response is
predictable and consist of a certain number of pulses.

188

Sebastian M Wieczorek

-2-1
10

1
2
2
1
0
-1
-2

-2
0 -1

(a2)

|E|2

(a1)

n
-2-1
10

Ey

1 2

Ex

1
time (units of R
)

100

1
)
time (units of R

100

1
time (units of R
)

100

1
0

-1

(b1)
2

-2

-1 -2
0

Ey

- 2- 1

n
0
Ex

0 1
2

0
(c2)

|E|2

(c1)

(b2)

|E|2

-1

-2

Ey

-1 -2
0

0
Ex

Fig. 6.6. Examples of multi-pulse excitability. The left column shows the phase
portrait and the right column the reaction of the laser to a small perturbation
above the excitability thresholds. Throughout = 1.0 and from (a) to (c) (K,)
takes the values (0.71, 0.95), (0.745. 1) and (0.735, 0.993). From S. Wieczorek
and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically injected lasers,
c 2005 by Institute of Physics Publishing;
Nonlinearity 18(3) (2005) 10951120 
reprinted with permission.

6.1.5 Codimension-Two Homoclinic Bifurcations

We now study in considerable detail the structure and bifurcations associated with the curve h1 that forms the boundary of the homoclinic teeth.
In particular, we show that codimension-two double-homoclinic and T-point
bifurcations play a prominent role in organizing the dynamics.
Figure 6.7 shows an enlargement near the rst homoclinic tooth (or what
is left of it) for = 4.5; compare with Fig. 6.3. Notice the two points D1
and D2 where additional homoclinic bifurcation curves emerge. The phase

6 Global Bifurcation Analysis in Laser Systems

-1 0
1

-1 0

D1

Ey
0

D2

-1

-1

Ey

-1 Ex

189

-1 Ex

1
S

H
h1

0.5

G1

D1

-0.5
D2

-1
S
ns

= 4.5

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Fig. 6.7. The (K, )-plane for = 4.5 near the point G1 (compare with Fig. 6.3) and
the phase portraits at the codimension-two double-homoclinic bifurcation points D1
and D2 . From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in
c 2005 by Institute
optically injected lasers, Nonlinearity 18(3) (2005) 10951120 
of Physics Publishing; reprinted with permission.

portrait at D1 and D2 show that we are dealing with a codimension-two

double-homoclinic orbit [9, 40]: both branches of the unstable manifold spiral back to the saddle point. This means that there are simultaneously two
individual homoclinic orbits associated with the same saddle point.
The two phase portraits at D1 and D2 are topologically equivalent and
both lie on the primary branch of the curve h1 . This can be seen in the further
enlargement of the (K, )-plane in Fig. 6.8, where panels (a)-(e) show the 1homoclinic orbit in phase space at the indicated parameter points along h1 .
As D2 is approached the unstable manifold forming a homoclinic orbit comes
closer and closer (from below) to the saddle and then leaves a neighborhood
of the saddle roughly along the other branch of the unstable manifold. Finally,
at D2 there are two simultaneous homoclinic orbits, one for each branch of the

190

Sebastian M Wieczorek
-1

-1

(a)

(b)

Ey
1

(c)

-1

-1

-1

-1

-1

Ey

Ex

-1

Ex

Ey

-1

-1

Ex

(d)

-0.3

n
0

D2

(c)

(d)

(e)

-1

-0.4

Ey
-0.5

h1

-1

(a)

-1

Ex

(e)

-0.6

D2

(c)

(e)
(d)

-1

(b)

-0.7

Ey

-1

Ex
-1

-0.8

D2

n
0

-0.9

-1

ns
0.2

0.3

= 4.5
0.4

0.5

0.6

Ey
K

-1

Ex

Fig. 6.8. The (K, )-plane for = 4.5 near the point D2 (compare with Fig. 6.7)
and the phase portraits along the homoclinic curve h1 as D2 is approached. From
S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically
c 2005 by Institute of Physics
injected lasers, Nonlinearity 18(3) (2005) 10951120 
Publishing; reprinted with permission.

unstable manifold. Eectively, the original 1-homoclinic orbit along the curves
h1 has split into two homoclinic orbits. Notice that the curve h1 accumulates
back on itself at D2 , as is also sketched in the inset of Fig. 6.8.

6 Global Bifurcation Analysis in Laser Systems

h1a

191

1
0
0
1

1
h
11
00
11
00

1
0
0
1

h1b

h1b

h1a

1
0
0
1

1
h
1
0
0
1

Fig. 6.9. Sketch of phase portraits and the bifurcation diagram (in two unfolding
parameters 1 and 2 ) of a double-homoclinic point D, as described in [9, 40]. From
S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically
c 2005 by Institute of Physics
injected lasers, Nonlinearity 18(3) (2005) 10951120 
Publishing; reprinted with permission.

This scenario agrees with what is known in the literature about the doublehomoclinic bifurcation [9, 40]. Again, not all details of this codimension-three
global bifurcation are known, but key features are as sketched in Fig. 6.9 (for
the case of a saddle focus as we encounter it here). The double-homoclinic
1 of two dierent
orbit D exists at the intersection point of two curves h1a and h
homoclinic orbits to the same saddle that contain each a dierent branch of
the unstable manifold of the saddle. As sketched, there is a third curve h1b of
homoclinic orbits that accumulates on the curve h1a . The accumulation is as
shown when the saddle quantity is larger than one [9, 40], which is the case
we encounter, because all double-homoclinic orbits occur above the curve ns.
Note that the analysis in the literature is in terms of a small tubular
neighborhood around the double-homoclinic orbits as sketched in panel D.
In this neighborhood the curves h1a and h1b are unrelated. However, as can
be seen in Fig. 6.8, they may be one and the same curve accumulating back
on itself. In fact, we nd this to be the typical situation in system (6.1). We
nally stress that the points Di that we encounter here are of codimension-two
because the two simultaneous homoclinic orbits are not related by symmetry.
Unlike in the case of a codimension-one symmetric double-homoclinic orbit, it

192

Sebastian M Wieczorek
1 0
-1

1 0
-1

1 0
-1





1
-1

0.75

-1

1
-1

-1

1
-1

-1

1 0
-1



 



 

0.5



 

 


1
-1

0.25

-1

1 0
-1

 

-0.25

1
-1

-0.5

1 0
-1

-0.75

-1

 


0.2

0.3

0.4

0.5

0.6

0.7

  
0.8 

1
-1

-1

Fig. 6.10. The (K, )-plane for = 4.5 near the points D1 , D2 , D3 , and T1 with
phase portraits as T1 is approached. From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically injected lasers, Nonlinearity 18(3) (2005)
c 2005 by Institute of Physics Publishing; reprinted with permission.
10951120 

is possible to perturb parameters such that one of the homoclinic connections

is broken and the other persists.
In Fig. 6.7 and Fig. 6.8 we found the double-homoclinic points D1 and
D2 as the end points of the curve h1 as it accumulates on itself. However, we
1 of a second homoclinic orbit
know from Fig. 6.9 that there must be a curve h

6 Global Bifurcation Analysis in Laser Systems

1
0
0
1

11
00
00
11

c
1
0
0
1

11
00
00
11

11
00

193

11
00
00
11

h1
1
0
0
1

T
c

1
0
0
1

Fig. 6.11. Sketches of the phase portraits along the homoclinic bifurcation curve
approaching a T-point in the plane of two unfolding parameters 1 and 2 . From
S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically
c 2005 by Institute of Physics
injected lasers, Nonlinearity 18(3) (2005) 10951120 
Publishing; reprinted with permission.

crossing at Di . In order to nd this new homoclinic orbit we split o the new

homoclinic orbit from the data of the approximate double-homoclinic orbit
at Di (given as the end point of the curve h1 ). We then follow this second
codimension-one homoclinic orbit in the (K, )-plane.
The result is shown in Fig. 6.10. The point D2 is indeed the intersection
1
point of two curves of codimension-one homoclinic orbits. The new curve h
also contains D1 and has two end points. One end point is a point D3 of
1 itself. The other end
a double-homoclinic orbit, which lies on the curve h
point is a point denoted by T1 that is reached in a spiraling fashion, as is also
sketched in the inset.
At the point T1 we encounter a bifurcation that is now generally referred
to as a T-point bifurcation. This type of codimension-two heteroclinic cycle
was studied in a general system, that is, one without any symmetry, in [9, 10]
in a tubular neighborhood around the heteroclinic orbits at the T -point. Note
that the T-point bifurcation is often associated with vector elds that have
the Z2 -symmetry of a rotation by around an invariant axis. In this case,
the heteroclinic cycle involves two saddle-foci, which are each others images
under the symmetry, and the origin (more generally, a point in the invariant
subspace of the symmetry), which is also a saddle-focus. This Z2 -symmetric
T-point bifurcation was initially found and studied in the Lorenz system [22],
but also occurs in other systems with rotational symmetry, such as an optically
pumped three-level laser [20], an electronic oscillator [19], and a semiconductor

194

Sebastian M Wieczorek

laser with phase-conjugate feedback [26]. It was recently also discovered in

systems with the Z2 -symmetry of point-reection [2, 32]; see also Chap. 7.
What we nd is a general T-point bifurcation (that is, in a system without
symmetry) for the case that both saddles involved are of saddle foci. The
approach to the point T is illustrated in Fig. 6.10(a)(e) with images of the
1-homoclinic orbit in phase space at the indicated parameter points along the
1.
curve h
The situation (a part of the bifurcation diagram near a T-point) is sketched
in Fig. 6.11. As the point T is approached along h1 , the homoclinic orbit approaches a second saddle focus, passing closer and closer by the saddle. At
the point T1 there are two heteroclinic connections: a codimension-two heteroclinic connection (black) where the one-dimensional unstable manifold of the
rst saddle coincides with the one-dimensional stable manifold of the second
saddle, and a generic (codimension-zero) heteroclinic connection (gray), given
as the intersection curve of the two-dimensional stable manifold of the rst
saddle and the two-dimensional stable manifold of the second saddle.
According to general theory [9, 10] there must exist a second spiraling
curve of homoclinic connection to the other (lower) saddle, leading to another
curve in parameter space that spirals into T1 . Furthermore, it is known that
there are many more curves of n-homoclinic bifurcations, which pass close
to the saddles an arbitrary number of times. We did not attempt to nd
all these bifurcation curves, but instead concentrated on the structure of 1homoclinic bifurcation curves. Nevertheless, the injection laser appears to be
a good model in which to study global bifurcations near T-point bifurcations
in more detail.
1
The bifurcation diagram in Fig. 6.10 is still quite incomplete. The curve h
of homoclinic orbits also accumulates on itself at D3 . So, as we did near the
double-homoclinic point D1 , we nd and follow the second codimension-one
homoclinic that must exist near D3 . This gives the continuation of the curve
h1 shown in Fig. 6.12, which ends at the point D1 . Furthermore, we followed
from near T1 the codimension-one homoclinic orbit of the (upper) saddle point
to lower values of (see already Fig. 6.13), where we discovered a second Tpoint bifurcation T2 . We then followed this T-point back to = 4.5. As can
be seen in Fig. 6.12, the point T2 is the end point of two spirals. In fact both
spirals turn out to belong to one and the same closed curve of codimension-one
2 as is illustrated by the sketch in the inset.
two-homoclinic orbit h
The bifurcation diagram in Fig. 6.12 is quite intricate: it involves several
double-homoclinic and T-point bifurcations. Unraveling it required detailed
numerical continuation with HomCont, guided by theoretical knowledge of
which homoclinic orbits are possible near the dierent codimension-two points.
We nally remark, that Fig. 6.12 shows a skeleton consisting of curves of 1homoclinic bifurcations. Indeed the existence of the T-points suggests that
there are n-homoclinic orbits for arbitrary n.

6 Global Bifurcation Analysis in Laser Systems

195

T1

h1
D3

0.5

T1
T2
D1
h1

0.25

h1

D2

2
h

h1

1
h
D3

-0.25

T2

h1

D1

-0.5

D2
-0.75

2
h

= 4.5
0.2

0.3

0.4

0.5

0.6

0.7

Fig. 6.12. The (K, )-plane for = 4.5 near the points T1 and T2 . From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically injected
c 2005 by Institute of Physics Publishlasers, Nonlinearity 18(3) (2005) 10951120 
ing; reprinted with permission.

6.1.6 Folds of Codimension-Two Homoclinic Bifurcation Curves

We know from Fig. 6.3 that the complicated structure of codimension-two
bifurcations in Fig. 6.12 is not present for smaller values of . The question
arises of how it disappears.
It turns out that an important ingredient in this change of the bifurcation diagram are minima (more generally, a fold) with respect to of certain
curves of codimension-two homoclinic bifurcations in the three-dimensional
(K, , )-space. This phenomenon is of codimension three, where one codimension is due to the fold with respect to . The other two codimensions

196

Sebastian M Wieczorek
1 0
-1
0.75

0.5



0.25

1
-1



-1

1 0
-1

-0.25

-0.5



-0.75


0.2

0.3

0.4

0.5

0.6

0.7

1
-1



-1

Fig. 6.13. The (K, )-plane for = 4.0 near the points T1 and T2 and the respective
phase portraits at T1 and T2 . From S. Wieczorek and B. Krauskopf, Bifurcations of
n-homoclinic orbits in optically injected lasers, Nonlinearity 18(3) (2005) 10951120
c 2005 by Institute of Physics Publishing; reprinted with permission.


are due to the special object in phase space, in this case a codimension-two
homoclinic bifurcation. One might speak of a codimension-two-plus-one event
to distinguish it from codimension-three bifurcations, where all codimensions
are due to a codimension-three object in phase space.
We already encountered this phenomenon in the creation and disappearance of points of codimension-two saddle-node homoclinic bifurcation (see the
folds with respect to in Fig. 6.4) and in the creation, with increasing , of
Belyakov points in the tangency between the curves ns and h1 (see Fig. 6.5).
In this section we consider two other examples, namely a fold of a curve of
T-point bifurcations and a fold of a curve of double-homoclinic bifurcations.
As we will see now, in both these examples the fold of the codimension-two
curve is accumulated by singularities in associated surfaces of codimension-one
global bifurcations.
We rst consider the case of T-point bifurcations. Figure 6.14 shows what
happens to the points T1 and T2 of T-point bifurcations as is decreased. After
the disappearance of the point D3 , the points T1 and T2 move closer and closer
to each other. There are a number of codimension-three events where the spiral
around T1 touches that around T2 . Each such event leads to a new closed curve
surrounding both T1 and T2 and the curve of homoclinic orbits connecting T1

6 Global Bifurcation Analysis in Laser Systems

197

0.75

T1

0.5

0.6

0.2

D3

T2

-0.25

T2

1
h

-0.5

-0.75

T1

0.4

0.25

(a)
= 4.5

-0.4

0.4

0.3

0.5

0.6

1
h

-0.2

(b)
= 4.0
K

0.5

0.4

0.3

0.6

0.6

T1

0.4

0.4

0.2
0.2

T2

1
h

1
h
0

-0.2

(c)
= 3.8
0.3

0.35

(d)
= 3.5
0.4

0.45

0.5

0.55

0.3

0.35

0.4

0.45

0.5

Fig. 6.14. Dependence on of the (K, )-plane near the points T1 and T1 . From
S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically
c 2005 by Institute of Physics
injected lasers, Nonlinearity 18(3) (2005) 10951120 
Publishing; reprinted with permission.

and T2 , as in Fig. 6.14 (c). This process continues until the points T1 and
T2 nally coincide, leaving behind a number of closed concentric curves of
homoclinic orbits, as in Fig. 6.14 (d). These closed curves then disappear one
by one as is decreased further. (We remark that this phenomenon has been
found independently in [2] in the Z2 -symmetric Chuas circuit with a cubic
nonlinearity; see also Chap. 7.) Finding this transition numerically was quite a
challenge because the curves involved are no longer connected. We succeeded
by starting from suitable points and continuing the respective homoclinic orbit
in .
1 are due to a classical
The individual changes in the structure of the curve h
singularity, namely the passage through an -degenerate point. At such a
1 surface in (K, , )-space does not have
point, the tangent space to the h
an -component (the derivative with respect to is zero). There are two

198

Sebastian M Wieczorek

1
h

Fig. 6.15. In the (K, , )-space the curve of T -point bifurcation is surrounded by
1 that spirals onto the T -curve; compare with
a surface of homoclinic bifurcation h
Fig. 6.14. From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits
c 2005 by Institute
in optically injected lasers, Nonlinearity 18(3) (2005) 10951120 
of Physics Publishing; reprinted with permission.

cases depending on the index of the -degenerate point, namely the transition
through a saddle and the transition through an extremum. Note that these
singularities are also called the simple bifurcation and the isola bifurcation;
see, for example, [23] for details.
This explanation in terms of singularity theory is a consequence of the
geometry of bifurcation surfaces and curves in (K, , )-space. In fact, the
whole sequence of events of T1 and T2 coming together and disappearing can
1 of
be nicely explained with the sketch in Fig. 6.15 of how the surface h
homoclinic bifurcations spirals around the curve T of T-point bifurcations.
The curve of T-point bifurcations is a smooth curve with a minimum with
respect to , and it is surrounded by a surface of codimension-one homoclinic
bifurcations that spirals towards this curve. The panels in Fig. 6.14 are twodimensional cross sections for xed through this surface. If is large enough,
the curve T is intersected in two points T1 and T2 and the spiraling near these
two points must be clockwise and counter-clockwise, respectively. The intersection of the surface with the section is a single curve for suciently large
. However, nearer the minimum of the curve T the surface has -degenerate
points where its tangent space does not have an -component. Passing through
each such point constitutes a basic codimension-one singularity of the surface
of homoclinic bifurcations. More precisely, above the minimum of the curve T

6 Global Bifurcation Analysis in Laser Systems

h1

-0.2

h1

-0.4

h1

-0.6

D2

D2

-0.4

-0.5

0.5

0.45

-0.6

1
h

-0.8

(a)
= 4.5
0.55

0.65

0.6

D1

(b)
= 4.3

1
h

-0.7
0.5

0.45

0.55

h1 -0.35

-0.4

h1

1
h

D2

h1
1
h

h1

-0.4

D1

-0.45

(c)
= 4.1

-0.6

0.5

0.65

0.6

-0.3

h1

-0.5

h1

-0.3

D1

-0.3

h1

199

0.525

0.55

0.575

0.6

0.625 K

(d)
= 4.0
0.45

0.5

0.55

0.6

0.65

Fig. 6.16. The (K, )-plane near D1 and D2 for varying . Points D1 and D2
disappear via codimension-three resonant double-homoclinic bifurcation as is decreased. From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits
c 2005 by Institute
in optically injected lasers, Nonlinearity 18(3) (2005) 10951120 
of Physics Publishing; reprinted with permission.

there are innitely many passages through saddles, which accumulate on the
minimum of the curve T . Globally, this creates the closed concentric curves
by connecting the respective homoclinic curves in a dierent way. Below the
minimum of T , on the other hand, each concentric circle disappears by contracting to a single point, which is the passage through an extremum (with
1 . We respect to a parameter, in this case ) in a two-dimensional surface h
nally remark that it would be quite a challenge to produce a numerical picture
of the surface sketched in Fig. 6.15.
Our second example is the merging and disappearance of the points D1
and D2 as is decreased from = 4.5 to = 4.0. Figure 6.16 shows four
numerical bifurcation diagrams in this transition. As the points D1 and D2
are moving closer together we again encounter a passage through a saddle
point. This happens between panels (a) and (b) of Fig. 6.16 and it leads to a

200

Sebastian M Wieczorek

h1

2
1
0
0
1
D1

h1

h1
11
00
00
11
D2

h1

h1
h1

1
0
0
1

1
0

D1

D2

1
h

1
h

(a)

(b)

h1

h1

1
h

1
h

h1
11
00
11
00

11
00
00
11

D1

D2

h1
00
11

h1

1
0
0
1

00
D11
1

D2

(c)

(d)

h1

2
1
h

h1

h1

1
h

h1
00
11
D12

h1

h1

(e)
1

(f )
1

Fig. 6.17. The sketch of how D1 and D2 disappear via codimension-three resonant
double-homoclinic bifurcation in the two unfolding parameters 1 and 2 . Compare
with Fig. 6.16. From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic
c 2005 by
orbits in optically injected lasers, Nonlinearity 18(3) (2005) 10951120 
Institute of Physics Publishing; reprinted with permission.

change in how the curves h1 in the cross section in the (K, )-plane connect.
After this event, the curve h1 in Fig. 6.16(b) connects the two points D1
and D2 . In a further passage through a saddle point the curve h1 pinches

6 Global Bifurcation Analysis in Laser Systems

201

o to create an isola, which is the situation shown in Fig. 6.16(c). In fact,

the isola is very close to the new connection between D1 and D2 . Numerical
continuation suggests that more and more isolas are formed as D1 and D2
come closer together. These isolas then disappear in passages through minima.
Furthermore, passes through the minimum of the D curve in the (K, , ) 1 and h1 in Fig. 6.16(d) no longer intersect
space. As a result, the curves h
and the points D1 and D2 have disappeared.
To clarify the situation, we sketch the transition leading to the disappearance of D1 and D2 in Fig. 6.17. It can again be understood by the geometry
of bifurcation surfaces in (K, , )-space, which in this case are organized
around a minimum (with respect to ) of the curve D of double-homoclinic
bifurcations. Figure 6.17(a)(c) and (f) are topologically as the numerical
bifurcation diagrams in Fig. 6.16(a)(d), respectively. We remark that it becomes more and more dicult to resolve numerically the dierent, small and
disjoint intersection curves of the surface h1 in (K, , )-space. The sketches
in Fig. 6.17(d) and (e) are based on our numerical investigations, and indicate
how the transition appears to take place. However, the exact details, in particular, the order in which isolas are created and shrink to points and disappear
is yet unknown. Our continuation study suggests the basic ingredients of this
transition and can reveal some of the rst steps in the specic transition at
hand. This scenario agrees with what is known about the (local) codimensiontwo bifurcation diagrams near a double-homoclinic bifurcation as sketched in
Fig. 6.9, but a complete study of this codimension-two-plus-one phenomenon
remains a challenge.
The fact that we encounter minima in curves T and D conrms the experience from simulations and experiments that the dynamics and the bifurcation
diagram of the injected laser become more complicated as the line-width enhancement factor is increased [59]. Indeed, when is increased past these
minima extra organizing centers, T-points or double homoclinic bifurcation
points, are born. These events are associated with innitely many transitions
through saddles and extrema in surfaces of global bifurcations. Furthermore,
general theory shows that the emerging T-points or double homoclinic bifurcation points are organizing centers that give rise to n-homoclinic orbits for
any n.
6.1.7 A Self-Similar Cascade Phenomenon
As a nal example of the increase in the complexity with we show in Fig. 6.18
the bifurcation diagram in the (K, )-plane for = 6.0 near the point G1 .
Notice that we only show the dierent parts of the curve h1 of one-homoclinic
orbits, which form what is left from the left most homoclinic tooth near G1 ;
compare with Fig. 6.3(f). Near the points D1 and D2 , that were already found
for = 4.5, we nd two new points D4 and D5 ; compare with Fig. 6.7. The
dierent bifurcation curves are very close together, and the inset shows a
topological sketch of the bifurcation diagram. Notice further that two extra

202

Sebastian M Wieczorek
-1 0 1

-1 0
1

D6

Ey
0

D7

-1

-1

Ey

-1 Ex

-1 Ex

0.5

D5

h1

D1

G1
D5

D4

-0.5

D1

D4

D2
-1

D6

ns

-1.5

(a)
= 6.0

D2

S
D7
0.2

0.4

(b)
0.6

0.8 K

Fig. 6.18. The (K, )-plane for = 6.0 near the point G1 (compare with Fig. 6.3)
and the phase portraits at the codimension-two points D6 and D6 . From S. Wieczorek and B. Krauskopf, Bifurcations of n-homoclinic orbits in optically injected
c 2005 by Institute of Physics Publishlasers, Nonlinearity 18(3) (2005) 10951120 
ing; reprinted with permission.

double-homoclinic bifurcation points D6 and D7 have just been created in the

same way as D1 and D2 previously. This is another example of the passage
through a minimum of a curve D of double-homoclinic orbits; compare with
Fig. 6.7.
Figure 6.18 shows that in the injected laser we are dealing with a type of
cascade phenomenon: complicated bifurcation scenarios found for one tooth
also occur for all the other teeth when is increased.
6.1.8 Concluding Remarks on Injected Lasers
This section presented a detailed study of the bifurcations of n-homoclinic
orbits in the rate equations describing a semiconductor laser with optical in-

6 Global Bifurcation Analysis in Laser Systems

203

jection. The corresponding curves of n-homoclinic bifurcations are organized

in what we call homoclinic teeth, experimentally accessible regions inside the
locking region of the laser. The analysis of the bifurcation diagram from a
global viewpoint proved to provide new insight into the nature of global bifurcations and allowed us to identify a cascade phenomenon where complicated
bifurcation scenarios repeat for subsequent homoclinic teeth.
The injection laser rate equations emerged as a concrete vector eld in
which complicated global bifurcations can be found and studied. Specically,
we found in this three-dimensional vector eld (without any additional symmetries) T-point bifurcations and double-homoclinic orbits. By making extensive use of continuation techniques for homoclinic and heteroclinic orbits, it is
possible to study these codimension-two global bifurcations themselves, and
also to nd out how they organize the corresponding bifurcation diagrams.
When changing a third parameter, we found a new phenomenon, namely
complicated transitions in two-parameter bifurcation diagrams that are due to
folds (in this case, minima) in codimension-two curves of global bifurcations.
These codimension-two-plus-one events come with accumulations of singularity transitions through saddles and extrema, which can be explained by the
geometry of surfaces of global bifurcations in a three-dimensional parameters
space. Our results raise a number of questions of bifurcation theory. In particular, a detailed study of the unfoldings of the codimension-two-plus-one
events remains a challenging task.
From the physical point of view, we presented here how the regions in which
one may nd multi-pulse excitability depend on the linewidth-enhancement
factor . Our results conrm that the dynamics and bifurcations of an injected laser are more complex the larger the linewidth enhancement factor .
Most importantly, they stimulated new laser experiments. The already demonstrated good agreement between theory and experiment on the level of local
bifurcations in the injection laser [65, 66] has now been extended to global
bifurcations such as the ones described here. In fact, the predicted eect of
multi-pulse excitability was recently measured independently by two dierent
groups [8, 25].

6.2 Phase-Locking Anomaly in Two Back-to-Back

Coupled Lasers
This section is based on [53, 54] and concerned with the dynamics and bifurcations of two coupled lasers. We model the system with spatial compositecavity modes that describe the entire coupled-laser structure [13, 43]. This is
in contrast to the more usual approach of modeling the individual uncoupled
lasers and then introducing the coupling via ad-hoc terms in the equations of
motion.
It is generally believed that the transition from weakly-coupled to totally
isolated lasers occurs smoothly. This means that, as the coupling approaches

204

Sebastian M Wieczorek

T
laser A
laser B
active medium active medium
z
L

L+dL

u1
B1

A1

compositecavity mode 1

u2
B2
A2

z
compositecavity mode 2

Fig. 6.19. Two back-to-back coupled lasers (top), and a pair of composite-cavity
modes (bottom).

zero, there is an uninterrupted vanishing of the detuning range over which

phase-locking is achieved. We show here that there is an anomaly in the transition from coupled to totally isolated lasers. Namely, this transition does not
occur smoothly but is interrupted with the occurrence of instabilities and even
chaotic dynamics. Importantly, there exists an open interval of the coupling
strength where phase locking is impossible at any detuning.
To understand these counter-intuitive phenomena we study in detail
two-dimensional bifurcation diagrams in the plane of laser detuning and
coupling strength for dierent values of the linewidth enhancement factor
; see Sect. 6.1.1. The analysis reveals various codimension-two bifurcations including saddle-node-Hopf and generalized Hopf bifurcations, 1:1 and
1:2 resonances, and homoclinic-doubling points. In particular, we identify
codimension-three bifurcations leading to the appearance of a gap in the
phase-locking region with increasing . Furthermore, we perform a detailed
study of bifurcations of periodic orbits, and specically their origins and mutual connections, to show how the gap is gradually occupied by instabilities
and chaos with increasing .

6 Global Bifurcation Analysis in Laser Systems

205

6.2.1 Composite-Cavity Description of Two Back-to-Back Coupled

Lasers
We consider two laser cavities with instantaneous (no time delay) coupling:
cavity A of length L is coupled via a common mirror of transmission T to
cavity B of length L + dL; see Fig. 6.19. For the consistent description of
coupling and optical nonlinearities, we expand the spatiotemporal laser eld
E(z, t) in terms of standing waves un (z) for the entire double-cavity cavity
structure [13, 43, 53]
0
1 %/
Ej (t)eij (t) uj (z) + c.c. .
(6.2)
E(z, t) =
2 j

Such standing waves are called composite-cavity modes. This is in contrast

to the usual approach that neglects spatial eects in the coupling. Notice,
that in the composite-cavity-mode picture, the resulting equations describe
interaction of composite-cavity modes rather than individual lasers [52].
Mathematically, the situation is described by a system of two globally coupled oscillators that can be described by the set of ve autonomous ordinary
dierential equations
% "
A
B
[Ckj
(1 + NA ) + Ckj
(1 + NB )] cos(kj )
E j = Ej + Cjj
k=1,2

#
A
B
[Ckj
(1 + NA ) + Ckj
(1 + NB )] sin(kj ) Ek , (6.3)
j = j + Cjj

% "
A
B
[Ckj
(1 + NA ) + Ckj
(1 + NB )] cos(kj )
k=1,2

#E
k
A
B
,
(1 + NA ) + Ckj
(1 + NB )] sin(kj )
+[Ckj
Ej

N A/B = (NA/B + 1)

A/B

Ckj

(6.4)

(1 + N A/B ) cos(kj ) Ek Ej , (6.5)

k,j=1,2

where j = 1, 2 and kj = k j is the phase dierence between mode

k and mode j. (Note that it is sucient to consider the equation for the
phase dierence 12 between the two electric elds.) Equations (6.4)(6.5)
are coupled to the algebraic constraints

3
3
2
2
3
2
j
j
1T
j
nb (L + dL) , (6.6)
nb L sin
sin
nb (2L + dL) = 2
sin
c
c
T
c

0
/
1
.

j
j
n
(L
+
dL)
sin
2
n
L
sin
2
b
b
c
c
1
(6.7)
+ Bj2 (L + dL) 1
A2j L
j
j
2
2
4 c nb (L + dL)
4 c nb L



j
1T
2c
nb L = L,
sin2
+A2j
c
T
j nb

206

Sebastian M Wieczorek

Aj sin

3

2
j
j
nb (L + dL) ,
nb L = Bj sin
c
c

A
(T, dL) =
Cjk

1
L

B
(T, dL) =
Cjk

1
L

(6.8)

dzuj (z)uk (z),

(6.9)

L
 L+dL

dzuj (z)uk (z).

(6.10)

Dierential equations (6.3)(6.5) describe the time evolution of the real eld
amplitudes E1 and E2 of the composite-cavity modes, their phase dierence
12 , and the population inversion in lasers A (NA ) and B (NB ). The modal
frequencies 1 and 2 are determined from the transcendental equation (6.6),
the modal amplitudes A1/2 and B1/2 are determined from (6.7)(6.8), and the
A
B
and Cjk
are determined from the spatial overlap of
coupling coecients Cjk
the composite-cavity modes; see (6.9)(6.10). More details on the derivation of
the model, algebraic constraints, and dimensionless parameters can be found
in [13, 52, 53].
The aim is to calculate two-dimensional bifurcation diagrams of system
(6.3)(6.5) in the (T, dL)-plane for dierent xed values of the linewidth enhancement factor . For the other parameters we chose the realistic values,
namely for the refractive index nb = 3.4, for the dimensionless gain coecient
= 9.82, for the dimensionless excitation rate = 2 in cavity A and B, and
for the ratio of the composite-cavity and population decay rates = 10. Because of the nature of the model, the bifurcation analysis of the coupled-laser
system is not as straightforward as in the case of the optically injected laser
in Sect. 6.1. The two main issues are:
1. the coupling parameters (which are the main bifurcation parameters)
namely, the coupling-mirror transmission T and the cavity-length mismatch dL, appear in (6.3)(6.5) implicitly through the modal frequencies
A
B
1 and 2 and integrals Cjk
and Cjk
as described by (6.6)(6.10);
2. the system has two types of periodic solutions: those where the phase
dierence 12 is bounded within a 2 interval, and those where 12 is
unbounded, that is, 12 (t) is periodic modulo 2; the latter oscillations
are also called rotations [16].
The rst issue can be overcome by appending the algebraic constraints (6.6)
(6.10) to the system of ODEs (6.3)(6.5) and solving the extended system,
that is, by performing continuation of solutions to (6.3)(6.5) and (6.6)(6.10)
simultaneously. The second issue becomes problematic only near transitions
between periodic solutions with bounded and unbounded phase. (Each individual type of periodic solution is readily continued with Auto.) Such transitions are common near interesting phenomena (e.g., codimension-two saddlenode-Hopf points with re-injection) and may cause technical inconvenience.
This issue can be overcome in the case of two laser modes by appropriate

6 Global Bifurcation Analysis in Laser Systems

207

change of variables [46]. However, it remains an interesting issue for the bifurcation analysis of multi-mode lasers where phase relations between more
than two individual modes need to be taken into account.
6.2.2 Symmetry Properties
It is interesting to discuss symmetries in the presence of two composite-cavity
modes. Because each composite mode has dierent spatial overlap with the
active media there is no perfect symmetry in the system of coupled-cavity
lasers. However, for long (compared to the wavelength = 1m) cavities this
dierence is small enough so that the system appears to have some symmetries.
A
B
If L  (this works well already when L 10), we have that Cjj
 Ckk
,
A
B
and Cjk  Cjk for j = k . Furthermore, if L  and dL we have
A
B
that the symmetry (Cjj
, dL) (Cjj
, dL). One consequence of the above
relations is the (approximate) reection symmetry (kj , N A , N B , , , dL)
(kj , N B , N A , , , dL). Hence, the bifurcation diagram in the (T, dL)
plane can be symmetric with respect to the change dL dL, provided that
both lasers have equal excitation rates and equal linewidth enhancement
factors .
Another consequence is the symmetry in the phase space. Under the assumption of equally pumped lasers, equal losses for both composite-modes,
0
, NA0 , NB0 } is an
and zero linewidth enhancement factor = 0, if {E10 , E20 , 12
0
0
0
0
0
equilibrium, then we notice that {E1 , E2 , 12 , NB , NA } is an equilibrium
too. Each of the two points may sometimes be associated with lasing at a
single composite-cavity mode. Whether both of them are stable at the same
time depends on the competition between the composite-cavity modes. Strong
competition results in bistability between these two equilibria [42].
6.2.3 Chaos in Practically Isolated Microcavity Lasers
In recent studies focusing on coupled lasers little attention has been devoted
to dynamical properties of practically uncoupled lasers, although such lasers
are encountered in a wide range of applications. By practically uncoupled or
isolated lasers, we mean two or more lasers where the desire is for the lasers
to operate totally independent of one another, while in practice, only partial
isolation is possible. Practically isolated lasers are encountered in the modern
technology of micro-optical circuits, where one faces the problem of reducing
cross-talk between laser diodes that are densely integrated onto a single chip.
Figure 6.20 depicts the bifurcation diagram in the (T, dL)-plane for two
microcavities described by (6.3)(6.5) with L = 2.8m and = 2, where supercritical bifurcations are plotted as solid curves, and subcritical bifurcations
as dashed curves. Phase-locking of lasers corresponds to the situation where
both lasers emit light of constant intensities and the same frequency. This
can be achieved in two ways: through phase locking of the composite-cavity

208

Sebastian M Wieczorek
(a)
4

dL / [10 4 ]

dL / [10 2 ]

locked

-1

-2

unlocked

unlocked

unlocked

(b)

0000000000000000
1111111111111111
0000000000000000
1111111111111111
+
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111

0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111

lock
S

H
G+
2

instabilities
and
chaos

-2

G
2

locked

H
S

unlocked

unlocked

-3

unlocked

-4
S

0.2

0.4

0.6

0.8

107

106

105

4
T 10

103

102

Fig. 6.20. Bifurcation diagram of (6.3)(6.5) in the (T, dL/)-plane of coupled

microcavity lasers (a), and an enlargement for very small T (b). From S. Wieczorek
and W.W. Chow, Chaos in practically-isolated microcavity lasers, Phys. Rev. Lett. 92
c 2004 by the American Physical Society; reprinted with permission.
(2004) 213901 

modes or when both lasers operate at a single composite-cavity mode. In the

{E1 , E2 , 12 , NA , NB } phase space, phase locking is represented by an equilibrium. The phase-locking region in Fig. 6.20(a), where the lasers operate with
a single composite-cavity mode, is indicated by the region between the two
supercritical branches of Hopf bifurcation curve H. The locked state is lost
when H is crossed towards higher values of |dL|. No other transition is visible
at this scale and the general features at moderate coupling are similar to what
is generally expected. The transition from weakly-coupled to totally isolated
lasers appears to occur smoothly, i.e., with an uninterrupted vanishing of the
phase locking region, as the coupling approaches zero.
However, the coupled-laser behavior contains an anomaly, whose presence
is only noticeable with signicant magnication of the (T, dL) parameter space
in the vicinity of the origin as shown in Fig. 6.20(b). There, we nd the curves
of saddle-node and Hopf bifurcation, S and H, respectively. These curves are
tangent at four codimension-two saddle-node-Hopf points where they change
from sub- to super-critical. These saddle-node-Hopf points are often origins of
complex bifurcation structures that give rise to chaos [33]. Starting from the

right, the phase-locking region closes near G+

2 and G2 where the two branches
+

of S merge. Phase-locking reappears at G1 and G1 and ends at the origin

of the (T, dL)-plane. In the notation of [33] G1 and G2 belong to dierent
types of saddle-node-Hopf points. Both types are associated with a complex
web of bifurcations roughly indicated by the shaded region in Fig. 6.20(b)
and studied in more detail in the next section. As the coupling approaches
zero, one expects the oscillators to be more independent. Instead, they start
interacting in a most complicated way and exhibit mutually induced chaotic

6 Global Bifurcation Analysis in Laser Systems

209

oscillations. An interesting question arises as to the origin of this intriguing

and counter-intuitive example of coupled nonlinear-oscillator behavior.
6.2.4 Origin of the Interrupted Phase-Locking Region
To explore the dependence of the coupled-laser instabilities on the resonator
length L, and to avoid technical diculties associated with continuation of
bifurcation curves across eight orders of magnitude in T (see Fig. 6.20) we
now consider longer laser cavities with L = 280m.
Bifurcations of Equilibria for = 0
The case of = 0 is highly degenerate and has an uninterrupted phaselocking range as is shown in Fig. 6.21. The two composite-cavity modes are in
strong competition leading to bistable locking range. Saddle-node and Hopf
bifurcations associated with both stationary points are tangent at T 0.027,
at four codimension-two saddle-node-Hopf points. Starting inside the locking
range and increasing |dL|, the locking is lost either via saddle-node bifurcation
(T < 0.027) or via Hopf bifurcation (T > 0.027). Although there are two
dierent bifurcations responsible for the locking-unlocking transition, one can
distinguish three dierent locking-unlocking mechanisms.
For T < 0.027 each of the two bifurcating stable equilibria has contributions from both composite modes; see Fig. 6.21(a). Here, frequency separation of the two composite-cavity modes is small, and the laser lockingunlocking transition is a transition between two composite-cavity modes which
are phase-locked and two composite-cavity modes which are phase-unlocked.
Locking of the lasers arises from phase-locking of the composite-cavity modes.
In the {E1 , E2 , 12 , NA , NB } space this is represented by two (there is bistability) saddle-node bifurcations of equilibria that take place on a single periodic
orbit. Unlocked operation is represented by a single stable periodic orbit.
For T > 0.15, each of the two bifurcating stable equilibria has a large
contribution from one composite mode, and a vanishing contribution from the
other composite mode. The laser locking-unlocking transition is a transition
between a single composite-cavity mode (see Fig. 6.21(a)) and two phaseunlocked composite-cavity modes. Locking of the lasers arises from strong
competition (owing to strong cross-saturation) between the two composite
modes [13, 52]. In the {E1 , E2 , 12 , NA , NB } space, this is represented by two
supercritical Hopf bifurcations leading to two stable periodic orbits. It also
involves saddle-node bifurcation of periodic orbits in which one of these stable
orbits disappears. Consequently, unlocked operation is represented by a single
stable periodic orbit [53] and no instabilities appear with increasing |dL|.
The most interesting region lies in between the two, near the points
,
where neither the composite-mode phase-locking nor the competing
G
j

210

Sebastian M Wieczorek

(a)

(b)

1.5

G2

0.5

G
2

Es2

0.55

1.5

0
1

G+
1

dL/
[102 ]

0.75

0.5
0.25

-1

0.5
0

G+
1, 2
G
1, 2

-0.5

H
G
1

H
2
dL / [10 ]

1.5

Ea2

S
1

-1

H
S

-1.5

0.2

0.4

0.6

0.8

Fig. 6.21. Bifurcations of equilibria of (6.3)(6.5), shown as a three-dimensional

2
bifurcation diagram in the (T, dL/, E1/2
)-space (a), and as a projection onto the
(T, dL/)-plane (b). The supercritical parts (darker solid curves) of the saddlenode bifurcation curve S (blue) and the Hopf bifurcation curve H (red) bound the
phase-locking region. S and H are tangent and change from super- to subcritical at
codimension-two saddle-node-Hopf points G
j . From S. Wieczorek and W.W. Chow,
Global view of nonlinear dynamics in coupled-cavity lasers-a bifurcation study, Opt.
c 2004 by Elsevier Science; reprinted with perComm., 246(46) (2005) 471493 
mission.

composite-cavity-mode description is valid. This is where the beat note frequency 12 comes close to the lasers characteristic relaxation oscillation frequency, which is known to give rise to nonlinear resonances and chaos. For
0.027 < T < 0.1, starting within the phase-locking region, locking of the lasers
is lost via undamping of the relaxation oscillation at a Hopf bifurcation. The
two stable stationary points become unstable and each of them gives rise to
one stable periodic orbit. Outside the phase-locking region and near G
j , these
two periodic orbits (bistability in unlocked operation) encounter instabilities
leading to complicated dynamics and chaos. It is interesting to note that, as T
increases, the transition between the rst and the second locking mechanism
is clear cut, indicated by G
j . On the other hand, the transition between the
second and the third locking mechanism happens continuously and involves
saddle-node bifurcation of periodic-orbits.
As a result of the degeneracy of the case = 0, in the projection of the
bifurcation diagram onto the (T, dL/) plane shown in Fig. 6.21(b) bifurcations of dierent stationary states appear as a single curve or point. How can
this degeneracy be removed?
Inuence of on Bifurcations of Equilibria
To explore the dynamics of dierent types of lasers and to understand how
qualitative dierences in the behavior of dierent lasers come about, we focus

6 Global Bifurcation Analysis in Laser Systems

211

our attention on the evolution of the phase-locking region with increasing .

The two phase-locking regions in Fig. 6.22, associated with the two stable
equilibria, are distinguished by left-inclined and right-inclined patterning, respectively. Let us recall that bifurcation theory predicts four dierent types
of saddle-node Hopf points. All four points G
j from Fig. 6.22(a) are of type
IV in the notation from [33].
Increasing above zero unfolds the otherwise degenerate bifurcation diagram in Fig. 6.22(a) so that for nonzero neither bifurcation curves nor

G
1 and G2 fully overlap any longer [Fig. 6.22(b)]. One phase-locking re
gion, associated with G+
1 and G1 , expands along the dL/ axis and moves in
the direction of lower values of T . The other phase-locking region, associated

with G+
2 and G2 , moves together with G2 and G2 in the direction of higher
values of T . Furthermore, the phase-locking region associated with G+
1 and
is
no
longer
bounded
by
the
entirely
supercritical
Hopf
bifurcation
curve.
G
1
Codimension-two generalized-Hopf bifurcation points Hg appear where the
Hopf curve changes from supercritical to subcritical. Throughout the range

of under consideration, the type of G+

2 and G2 remains unchanged. On the
+

other hand, G1 and G1 change from type IV to type III (in the notation
from [33]) at  0.5. The bifurcation diagram in Fig. 6.22(c) shows that the

curve H has a cusp at G+

1 and G1 which makes this a very special point,
namely a bifurcation of codimension at least three. During the change in the

type of G+
1 and G1 the two associated branches of H, one supercritical and
the other subcritical, locally exchange their order. This has important consequences to where the chaotic dynamics associated with G
1 appear; see the
next subsection for explanation. Increasing further results in less overlap
between the two phase-locking regions to the point where they no longer coalesce [Fig. 6.22(d)]. At = 1 the two phase-locking regions are well separated
and a gap appears where the coupled-cavity lasers never lock. This gap increases with further increase of [Fig. 6.22(e)] so that for = 3 there are
two distinct phase-locking regions [Fig. 6.22(f)], one at low coupling-mirror
transmissions 0 < T < 0.01 and the other at relatively high coupling-mirror
transmissions 0.45 < T < 1. Furthermore, the generalized Hopf points Hg
are gone and both phase-locking regions are again bounded by the entirely
supercritical parts of S and H.
Bifurcations of Periodic Orbits
The next question concerns nonlinear oscillations for parameter settings
within the gap between the two phase-locking regions. In a coupled-cavity
laser periodic orbits emerge along Hopf bifurcation curves H, along curves S
of global saddle-node homoclinic bifurcation, and along homoclinic bifurcation
curves h. In particular, we already identied two types of codimension-two bifurcations, namely saddle-node-Hopf points G
j and generalized Hopf points
Hg . Both are sources of bifurcations of periodic orbits and, hence, starting

212

Sebastian M Wieczorek
0.75

(a) = 0

0.75

(b) = 0.3

2
dL / [10 ]

0.5

0.5
G+
1,2

0.25

Hg

G+
1
G+
2

0.25

0
G
1,2

-0.25

G
2

-0.25
G
1
H

-0.5

H
Hg

-0.5

-0.75

-0.75
0.05

0.1

0.15

0.75

(c) = 0.5

2
dL / [10 ]

0.5

G+
1

0.05

0.1

0.15

G+
1

(d) = 1
Hg

0.5 S

Hg

0.25

G+
2

0.25

G+
2
H

S
G
2

-0.25
-0.5

G
1

Hg

-0.5

-0.75
0.05

0.1

0.5
0

-0.5
-1

Hg
G
1

0.15

0.05

(e) = 2

1.5
1

G
2

-0.25

-0.75

2
dL / [10 ]

0.75

G+
1

3
2

(f) = 3

G+
1

G+
2

H
Hg

0.15

G+
2

Hg

0.1

H
S

G
2

G
2

-1
H

-2

G
1

-1.5

G
1

-3
0.1

0.2

0.3

0.1

0.2

0.3

0.4

0.5 T

Fig. 6.22. Phase-locking region of (6.3)(6.5) in the (T, dL/)-plane for dierent
values of the linewidth enhancement factor . S and H are tangent and change
from supercritical to subcritical at codimension-two saddle-node-Hopf points G
j .
H also changes from supercritical to subcritical at generalized Hopf points Hg . The
color coding is as in Fig. 6.21. From S. Wieczorek and W.W. Chow, Global view
of nonlinear dynamics in coupled-cavity lasers-a bifurcation study, Opt. Comm.,
c 2004 by Elsevier Science; reprinted with permission.
246(46) (2005) 471493 

6 Global Bifurcation Analysis in Laser Systems

213

points for further analysis. Note that in the plots of curves of bifurcations of
periodic orbits we do not distinguish between super- and subcritical parts.
As expected from general theory [33], there is a torus bifurcation curve
T emerging from each point G
j [Fig. 6.23(a)]. Only two of four T curves
are visible (since the case = 0 is degenerate). These torus curves involve
two frequencies, the relaxation oscillation frequency and the inter-mode frequency. They are associated with a resonance tongue structure (not shown
here) and denote the onset of either quasiperiodic (parameter settings between the tongues) or periodic oscillations (parameter settings within a resonance tongue) when the solid black curves T are crossed from the right to
the left [38]. Also, they signal the appearance of chaos via the break-up of a
2-torus when the resonance tongues start to overlap. The curves T terminate
at 1:2 resonance points [33, Sec. 9.5.3] where they connect to period-doubling
curves P D1 . The P D1 curves are the rst steps in an innite period-doubling
cascade to chaos [18]. The secondary period-doubling curves P Dn>1 may be
arranged in nested or unnested islands of period-doublings [61]. In either
case, period-doubling islands are associated with chaotic dynamics. One of
the period-doubling curves from Fig. 6.23(a) does not form a closed loop but
terminates at two homoclinic-doubling bifurcation points B 1 [39]. Furthermore, there is a non-degenerate saddle-node-of-periodic-orbit curve SL where
one of the two stable periodic orbits, born along the degenerate H curve, disappears. The overall dynamical picture for = 0 consists of the uninterrupted
bistable phase-locking region and complicated, sometimes chaotic, dynamics
found outside of the phase-locking region and near the points G
j .
When is increased from zero [Fig. 6.23(b)], the degenerate bifurcation
diagram unfolds and one clearly sees four torus curves connecting to perioddoubling curves at four 1:2 resonance points. Interestingly, regions of complicated dynamics associated with G
2 start to overlap with the phase-locking
.
There,
depending on the initial condition, the
region associated with G
1
coupled lasers may either be phase-locked or exhibit complicated unlocked
oscillations. At the generalized Hopf points Hg , the curve SL attaches to the

supercritical branches of H emerging from G+

1 and G1 , causing the curves H
to change from supercritical to subcritical. Stable periodic orbits born along
these supercritical branches of H disappear at SL.
Increasing further results in no qualitative changes to saddle-node Hopf
points G
2 nor associated torus and period-doubling bifurcations. As the gap
between the two phase-locking regions appears [Fig. 6.23(c)(f)], the points
G
2 move in the direction of higher values of T . Concurrently, the two torus

curves T emerging from G+

2 and G2 are dragged along, causing the two
attached period-doubling cascades to shift into the gap between the two phaselocking regions. On the other hand, a number of qualitative changes takes

place near G
1 . For < 0.5, bifurcations of periodic orbits emerging from G1
+
evolve in the direction of increasing |dL/|. Near = 0.5, the type of G1 and
G
1 changes.

214

Sebastian M Wieczorek
0.75

(a) = 0

P D1

0.6

SL

1:2

0.5

0.4

2
dL / [10 ]

1:2

-0.25

P D1

2
dL / [10 ]

G+
1

0.75

Hg

PD

-0.6

G+
1

2
dL / [10 ]

0.5
0
-0.5

G
2

P D1

0.04

G+
2

1:2

0.08

0.12

(f) = 3
P D1

2 G+ 1:2
1

P D1

1
Hg

1:2

P D1

1:2

P D1
1:2

Hg

-1
P D1

-1
-1.5

(e) = 2

1:2

-0.75

Hg
G
1

0.12

0.08

0.04

1:2

-0.5

-0.25

1:2

1:2

0 SL

H
G
2

SL G

1:2

0.25

0 P D1
-0.2

P D1

T
G+
2

0.06

(d) = 1

G+
1
Hg

0.5

0.2

Hg

0.04

0.02
S

Hg
1:2

P D1

(c) = 0.5

P D1
SL

1:2

SL

0.04

0.03

0.02

1:2

-0.6

0.01

1.5

G
1

P D1

SL

0.6

-0.4

G
2

-0.2

-0.75

0.4

B1

-0.4

-0.5

G+
2

h
H

1:2

0
G
1,2

1:2

B1

G+
1

0.2

G+
1,2

Hg

SL

P D1

P D1

B1

0.25

(b) = 0.3

-2

1:2
G
1

G
1 1:2

P D1

-3
0.05

0.1

0.15

0.05

0.1

0.15

Fig. 6.23. Bifurcation diagram of (6.3)(6.5) in the (T, dL/) plane for dierent
values of the linewidth enhancement factor . Period-doubling bifurcation curves
P D are in green, saddle-node bifurcation of periodic orbit curves SL are in brown,
torus bifurcation curves T are in solid black, and homoclinic bifurcation curves h
are in dotted black. From S. Wieczorek and W.W. Chow, Global view of nonlinear
dynamics in coupled-cavity lasers-a bifurcation study, Opt. Comm., 246(46) (2005)
c 2004 by Elsevier Science; reprinted with permission.
471493 

6 Global Bifurcation Analysis in Laser Systems

(a) = 0.3

215

(b) = 0.5
1:1

0.34

SL

0.415

2
dL / [10 ]

1:1

0.33

Hg

G+
1

SL

0.41

G+
1

0.32

0.31

Hg

0.405

0.3

(d) = 2
G+
1

1.35

0.65

2
dL / [10 ]

0.019

0.017

0.015

0.021

0.02

0.019

0.018

(c) = 1

1.3
T
S

0.6

1.25

Hg

G+
1

1:2

1.2

1:2

0.55

1.15
P D1

P D1

1.1

SL

0.5

S
H

1.05

0.011

0.013

0.015

0.012

0.014

0.016

Fig. 6.24. Enlarged bifurcation diagram of (6.3)(6.5) near saddle-node-Hopf point

G+
1 for dierent values of the linewidth enhancement factor . From S. Wieczorek
and W.W. Chow, Global view of nonlinear dynamics in coupled-cavity lasers-a bic 2004 by Elsevier Science;
furcation study, Opt. Comm., 246(46) (2005) 471493 
reprinted with permission.

The details near G+

1 are shown in Fig. 6.24. Close to  0.5 a new curve
SL appears that connects to H at the two Hg points, and has a cusp near
(T, dL/) = (0.016, 0.4125) [Fig. 6.24(b)]. The torus curve T detaches from
G+
1 and attaches near the cusp of this new SL curve at 1:1 resonance [33,
Sec. 9.5.2]. At = 1 [Fig. 6.24(b)] the extra SL curve is gone. The torus
curve T attaches again to G+
1 but ips from above to below the curve S
[Fig. 6.24(c)(d)]. As a consequence, starting at  1 [Fig. 6.23(d)], the

two torus curves T emerging from G+

1 and G1 , and the attached period1
doubling curves P D , start lling the gap between the two phase-locking
regions [Fig. 6.23(d)(f)]. For clarity, only parts of the P D1 curves associated

with G+
1 and G1 are plotted in Fig. 6.23(b)(f).

216

Sebastian M Wieczorek

6.2.5 Concluding Remarks on Coupled Lasers

The interesting dynamics of coupled-cavity lasers arises from two types of
nonlinearities: those imparted in the composite-mode properties by the optical
coupling (algebraic constraints), and those imparted in the active medium by
the population dynamics (dierential equations). Their interplay results in a
rich display of chaotic oscillations when two conditions are met simultaneously.
(i) The composite-cavity-mode beat note must be close to resonant with the
characteristic (relaxation oscillation) frequency of the active medium, thus
strongly coupling laser elds and active media;
(ii) an appreciable spatial overlap between composite-cavity modes must be
present for a strong coupling of the lasing modes.
Continuation techniques allowed us to uncover a counter-intuitive example
of chaos in ultra-weakly coupled nonlinear-oscillators and to explain how this
unexpected dynamical picture arises with increasing linewidth enhancement
factor . In particular, an uninterrupted and degenerate (bistable) phaselocking region at = 0 unfolds and develops a gap, which is gradually
occupied with instabilities and chaos for = 0. The underlying mechanism is a change in the competition between composite-cavity modes that
causes a change in the type of codimension-two saddle-node Hopf points via
a codimension-three cusp singularity on the Hopf bifurcation curve. Furthermore, several other codimension-two bifurcations, including strong resonances,
are identied as sources of instabilities and chaos in coupled-cavity lasers.
Many of the phenomena mentioned are interesting from a bifurcation theory
point of view and should be studied in more detail.
The dependence of the dynamics on the cavity length reveals eects due to
nonlinear optical coupling. With decreasing cavity length, the two conditions
that are necessary for the complicated dynamics to occur shift towards the
origin of the (T, dL/) parameter space. Consequently, for short cavities these
two conditions may be satised at ultra-low optical coupling (e.g. T < 105
for L 3), where lasers are generally expected to act independently. This
bifurcation analysis provided new insight into an overall understanding of
coupled-laser behavior.

6.3 Outlook
The eld of nonlinear optical/laser systems is expanding in many new directions. Examples of new types of optical systems include nanoscale photoniccrystal lasers [41], optical resonators with quantum coherence [55, 67], and
multimode quantum-dot lasers [49]. Owing to their nanoscale and quantum
coherence, these systems are expected to have strong optical nonlinearities
that are dierent from those found in conventional optical/laser systems. New

6 Global Bifurcation Analysis in Laser Systems

217

nonlinear phenomena are waiting to be uncovered and, based on our experience so far, we believe that continuation techniques are the tool of choice.
The nonlinear analysis of these newly emerging optical systems faces mathematical challenges such as handling high-dimensional multimode systems,
ODEs with algebraic constraints, and multiple time scales. It is, therefore, an
easy prediction that the bifurcation analysis of newly emerging problems in
laser physics and photonics will continue to stimulate and contribute to the
further development of continuation techniques themselves.

Acknowledgements
Individual results in this chapter have appeared in previous publications, and
I thank my co-workers Bernd Krauskopf, Weng. W. Chow, and Daan Lenstra
for their contributions. Text passages and gures have been reproduced with
permission, and I thank the American Physical Society, Elsevier Science and
Institute of Physics Publishing for permission to use material from [53], [54]
and [58], respectively.

References
1. N. B. Abraham, L. A. Lugiato, and L. M. Narducci (Eds.), Special issue on
Instabilities in Active Optical Media. J. Opt. Soc. Am. B, 2, 1985.
2. A. Algaba, M. Merino, F. Fern
andez-S
anchez, and A. Rodriguez-Luis. Closed
curves of global bifurcations in Chuas equations: a mechanism for their formation. Internat. J. Bifurc. Chaos Appl. Sci. Engrg., 13:609616, 2002.
3. V. Annovazzi-Lodi, S. Donati, and M. Manna. Chaos and locking in a semiconductor laser due to external injection. IEEE J. Quantum Electron., 30(7):1537
1541, 1994.
4. A. Argyris, D. Syvridis, L. Larger, V. Annovazzi-Lodi, P. Colet, I. Fischer,
J. Garc`ia-Ojalvo, C. R. Mirasso, L. Pesquera, and K. A. Shore. Chaos-based
communications at high bit rates using commercial ber-optic links. Nature,
438:343-346, 2005.
5. A. Back, J. Guckenheimer, M. R. Myers, F. J. Wicklin, and P. A. Worfolk.
DsTool: Computer assisted exploration of dynamical systems. Not. Amer. Math.
Soc., 39:303309, 1992.
6. F. Bai and A. R. Champneys. Numerical computation of saddle-node homoclinic orbits of co-dimension one and two. J. Dyn. Stab. Syst., 11:325346,
1996.
7. L. Belyakov. Bifurcation of systems with homoclinic curve of a saddle-focus
with saddle quantity zero. Mat. Zam., 36:681689, 1984.
8. P. Besnard, Private communication, 2006.
9. V. V. Bykov. The bifurcations of separatrix contours and chaos. Physica D,
62(1-4): 290299, 1993.
10. V. V. Bykov. Orbit structure in a neighborhood of a separatrix cycle containing
two saddle foci in Methods of qualitative theory if dierential equations and
related topics. American Math. Soc. Transl. Ser 2, 200:8797, 2000.

218

Sebastian M Wieczorek

11. A. R. Champneys and Yu. A. Kuznetsov. Numerical detection and continuation

of codimension-two homoclinic bifurcations. Internat. J. Bifurc. Chaos Appl.
Sci. Engrg., 4:785822, 1994.
12. A. R. Champneys, Yu. A. Kuznetsov, and B. Sandstede. A numerical toolbox
for homoclinic bifurcation analysis. Internat. J. Bifurc. Chaos Appl. Sci. Engrg.,
6:867887, 1996.
13. W. W. Chow. A composite-resonator mode description of coupled lasers. IEEE
J. Quant. Electron., QE-22(8):11741183, 1986.
14. S-N. Chow and X-B Lin. Bifurcation of a homoclinic orbit with a saddle-node
equilibrium. Di. Int. Eqns., 3:435466, 1990.
15. B. Deng. Homoclinic bifurcations with nonhyperbolic equilibria. SIAM J. Math.
Anal., 21:693720, 1990.
16. E. Doedel, A. R. Champneys, T. Fairgrieve, Yu. A. Kuznetsov,
B. Sandstede, and X. Wang. Auto2000: Continuation and bifurcation software for ordinary dierential equations. Available via
http://indy.cs.concordia.ca/auto/main.html.
17. T. Erneux, V. Kovanis, A. Gavrielides, and P. M. Alsing. Mechanism for perioddoubling bifurcation in a semiconductor laser subject to optical injection. Phys.
Rev. A, 53:43724380, 1996.
18. M. Feigenbaum. Quantitative universality for a class of nonlinear transformations. J. Stat. Phys., 19(1):2552, 1978.
19. F. Fern
andez-S
anchez, E. Freire, and A. Rodriguez-Luis. T-points in a Z2 symmetric electronic oscillator:(I) analysis. Nonlin. Dyn., 28:5369, 2002.
20. W. Forysiak, J. V. Moloney, and R.G . Harrison. Bifurcations of an optically
pumped three-level laser model. Physica D, 53(1):162186, 1991.
21. A. Gavrielides, V. Kovanis, P. M. Varangis, T. Erneux, and G. Lythe. Coexisting periodic attractors in injection-locked diode lasers. Quant. Semiclass. Opt.,
9:785786, 1997.
22. P. Glendinning and C. Sparrow. T-points: A codimension two heteroclinic bifurcation. J. Statist. Phys., 43:479488, 1986.
23. M. Golubitsky and D. G. Schaeer, Singularities and Groups in Bifurcation
Theory, Vol. 1 Applied Mathematical Sciences 51 (New York:Springer).
24. S. V. Gonchenko, D. V. Turaev, P. Gaspard, and G. Nicolis. Complexity in
the bifurcation structure of homoclinic loops to a saddle-focus. Nonlinearity,
10:409423, 1997.
25. D. Goulding and G. Huyet, Private communication, 2006.
26. K. Green, B. Krauskopf, and G. Samaey. A two-parameter study of the locking
region of a semiconductor laser subject to phase-conjugate feedback. SIAM J.
Appl. Dyn. Sys., 2(2):254276, 2003.
27. A. J. Homburg. Periodic attractors strange attractors and hyperbolic dynamics
near homoclinic orbits to saddle-focus equilibria. Nonlinearity, 15:10291050,
2002.
28. D. M. Kane and K. A. Shore (Eds.), Unlocking Dynamical Diversity: Optical
Feedback Eects on Semiconductor Lasers, (John Wiley & Sons, 2005).
29. V. Kovanis, A. Gavrielides, T. B. Simpson, and J.-M. Liu. Instabilities and
chaos in optically injected semiconductor lasers. Appl. Phys. Lett., 67:2780
2782, 1985.
30. B. Krauskopf and D. Lenstra (Eds.), Fundamental Issues of Nonlinear Laser
Dynamics, (AIP Conference Proceedings 548, 2000).

6 Global Bifurcation Analysis in Laser Systems

219

31. B. Krauskopf, K. R. Schneider, J. Sieber, S. M. Wieczorek, and M. Wolfrum.

Excitability and self-pulsations near homoclinic bifurcations in semiconductor
laser systems. Opt. Commun., 215(4-6):367379, 2003.
32. B. Krauskopf and J. Sieber. Bifurcation analysis of an inverted pendulum with
delayed feedback control near a triple-zero eigenvalue singularity. Nonlinearity,
17:85103, 2004.
33. Yu. A. Kuznetsov, Elements of Applied Bifurcation Theory, (Springer, New
York, 1995).
34. F.-Y. Lin and J.-M. Liu. Chaotic radar using nonlinear laser dynamics. IEEE
J. Quantum Electron., 40(6):815820, 2004.
35. L. A. Lugiato, L. M. Narducci, D. K. Bandy, and C. A. Pennise. Breathing
spiking and chaos in a laser with injected signal. Opt. Commun., 46:64-68,
1983.
36. P. Mandel, Theoretical Problems in Cavity Nonlinear Optics, (Cambridge studies in modern optics, Cambridge University Press, 1997).
37. J. V. Moloney, J. S. Uppal, and R. G. Harrison. Origin of Chaotic Relaxation Oscillations in an Optically Pumped Molecular Laser. Phys. Rev. Lett.,
59:2868271, 1987.
38. E. Ott, Chaos in Dynamical Systems (Cambridge University Press, Cambridge,
1993).
39. B. Oldeman, B. Krauskopf, and A. R. Champneys. Death of period-doublings:
locating the homoclinic-doubling cascade. Physica D, 146:100120, 2000.
40. I. M. Ovsyannikov and L. P. Shilnikov. Systems with a saddle-focus homoclinic
curve. Math. USSR-Sb., 58(2):557574, 1997.
41. B. Pasenow, M. Reichelt, T. Stroucken, T. Meier, S. W. Koch, A. R. Zakharian, and J. V. Moloney. Enhanced light-matter interaction in semiconductor
heterostructures embedded in one-dimensional photonic crystals. J. Opt. Soc.
Am. B, 22(9):20392048, 2005.
42. M. Sargent III, M. O. Scully, and W. E. Lamb Jr., Laser Physics. (AddisonWesley, 1974).
43. S. A. Shakir and W. W. Chow. Semiclassical theory of coupled lasers. Opt.
Lett., 9:202204, 1984.
44. M. V. Shashkov and D. V. Turaev. On the complex bifurcation set for a system
with simple dynamics. Internat. J. Bif. Chaos Appl. Sci. Engrg., 6:949968,
1996.
45. S. Schecter. Numerical computation of saddle-node homoclinic bifurcation
points. SIAM J. Numer. Anal., 30(4):11551178, 1993.
46. J. Sieber. Numerical bifurcation analysis for multi-section semiconductor lasers.
SIAM J. Appl. Dyn. Syst. 1(2):248270, 2002.
47. J. Sneyd, A. LeBeau, and D. Yule. Traveling waves of calcium in pancreatic
acinar cells: model construction and bifurcation analysis. Physica D, 145(12):158179, 2000.
48. M. B. Spencer and W. E. Lamb Jr. Laser with a transmitting window. Phys.
Rev. A, 5(2):884892, 1972.
49. Y. Tanguy, J. Houlihan, G. Huyet, E. A. Viktorov, and P. Mandel. Synchronization and clustering in a multimode quantum dot laser. Phys. Rev. Lett.,
96(5):053902, 2006.
50. G. D. VanWiggeren and R. Roy. Communication with chaotic lasers. Science,
279(5354):11981200, 1997.

220

Sebastian M Wieczorek

51. C. O. Weiss and R. Vilaseca, Dynamics of Lasers (VCH Verlagsgesellschaft,

Weinheim, Germany, 1991).
52. S. M. Wieczorek and W. W. Chow. Bifurcations and interacting modes in coupled lasers: A strong-coupling theory. Phys. Rev. A, 69:033811, 2004.
53. S. M. Wieczorek and W. W. Chow. Chaos in practically-isolated microcavity
lasers. Phys. Rev. Lett. 92:213901, 2004.
54. S. M. Wieczorek and W. W. Chow. Global view of nonlinear dynamics in
coupled-cavity lasers-a bifurcation study. Opt. Commun., 246(4-6):471493,
2005.
55. S. M. Wieczorek and W. W. Chow. Self-induced chaos in a single-mode inversionless laser. Phys. Rev. Lett., 97:113903, 2006.
56. S. M. Wieczorek and W. W. Chow. On dynamical sensitivity of lasers and
instability-based coherent-signal detection. submitted, 2007.
57. S. M. Wieczorek, W. W. Chow, L. Chrostowski, and C. J. Chang-Hasnain.
Improved semiconductor-laser dynamics from induced population pulsation.
IEEE J. Quant. Electron., QE-42(6):552562, 2006.
58. S. Wieczorek and B. Krauskopf. Bifurcations of n-homoclinic orbits in optically
injected lasers. Nonlinearity, 18:10951120, 2005.
59. S. M. Wieczorek, B. Krauskopf, and D. Lenstra. A unifying view of bifurcations in a semiconductor laser subject to optical injection. Opt. Commun.,
172(1):279295, 1999.
60. S. M. Wieczorek, B. Krauskopf, and D. Lenstra. Mechanisms for multistability
in a semiconductor laser with optical injection. Opt. Commun., 183(1-4):215
226, 2000.
61. S. M. Wieczorek, B. Krauskopf, and D. Lenstra. Unnested islands of perioddublings in an injected semiconductor laser. Phys. Rev. E, 64:056204, 2001.
62. S. M. Wieczorek, B. Krauskopf, and D. Lenstra. Multipulse excitability in a
semiconductor laser with optical injection. Phys. Rev. Lett., 88:063901, 2002.
63. S. M. Wieczorek, B. Krauskopf, T. B. Simpson, and D. Lenstra. The dynamical
complexity of optically injected semiconductor lasers. Phys. Rep., 416(1-2):1
128, 2005.
64. S. M. Wieczorek and D. Lenstra. Spontaneously excited pulses in an optically
driven semiconductor laser. Phys. Rev. E, 69:016218, 2004.
65. S. M. Wieczorek, T.B . Simpson, B. Krauskopf, and D. Lenstra. Global quantitative predictions of complex laser dynamics. Phys. Rev. E, 65:045207R, 2001.
66. S. M. Wieczorek, T. B. Simpson, B. Krauskopf, and D. Lenstra. Bifurcation
transitions in an optically injected diode laser: theory and experiment. Opt.
Commun., 215(1-3):125134, 2003.
67. W. Yang, A. Joshi, and M. Xiao. Chaos in an electromagnetically induced
transparent medium inside an optical cavity. Phys. Rev. Lett., 95:093902, 2005.
68. M. G. Zimmermann, S. Firle, M. A. Natiello, M. Hildebrand, M. Eiswirth,
M. Bar, A. K. Bangia, and I. G. Kevrekidis. Pulse bifurcation and transition
to spatiotemporal chaos in an excitable reaction-diusion model. Physica D,
110(1-2):92104, 1997.

7
Numerical Bifurcation Analysis of Electronic
Circuits
Emilio Freire and Alejandro J Rodrguez-Luis
Departamento de Matem
atica Aplicada II, Escuela Superior de Ingenieros de
Sevilla, Spain

It is well known that the creation of the modern geometrical theory of dynamical systems by Poincare at the end of the 19th century was motivated
by problems arising in celestial mechanics [41]. Perhaps it is not so widely
known that the dynamics of electronic circuits played an important role at
the early stages of the development of this theory. In the 1920s Van der Pol
[47] described the periodic oscillations of self-sustained circuits in terms of the
limit cycles of Poincare. He performed experiments with periodically excited
circuits and measured, for the rst time, complex behavior in a nonlinear system. In the 1930s there was pioneering work of Andronovs Russian school on
the theory of oscillations in electronic, mechanical and control systems [11].
It is important to realize that these rst applications of Poincares qualitative theory to electronic circuits led to new concepts and theoretical results.
Examples of this include Lienards theorems [40, Chap. 3], as motivated by
the works of Van der Pol, the development of bifurcation theory for planar
systems by Andronov and co-workers [9], and the introduction of the concept
of structural stability by Andronov and Pontriaguin [10].
The relationship between the mathematical theory of ordinary dierential
equations (ODEs) and the dynamics of electronic circuits was initially very
close, but this did not continue for very long. In fact, one might speak of a divorce between the two elds in the subsequent development, where electronic
devices and systems became ever more complex and of greater dimension.
Starting with the invention of the transistor in the 1950s, there was a true
explosion in the size of electronic circuits, culminating in the ascent of the
microchip a complex circuit with thousands or even millions of components. In theoretical investigations of such electronic circuits one can hardly
nd a trace of the geometrical theory of dynamical systems. One of the few
exceptions is the almost forgotten work of Hayashis school in Japan on the
global dynamics in experiments with analog computers [46].
On the other hand, in the 1960s and 1970s the mathematical theory of
dynamical systems experienced much development, with the introduction of
new ideas by Smale, Arnold, Lorenz, Yorke, and Feigenbaum, to name just

222

Emilio Freire and Alejandro J Rodrguez-Luis

a few of the contributors. The theory as we know it today (see, for example,
the textbooks [32, 36, 49]) sheds light on the structure of complex dynamic
behavior (by which we mean recurrent, aperiodic and chaotic) that is present
in numerous nonlinear models arising in applications; see, for example, the
recent survey [35] and references therein.
With these theoretical developments there came a renewed interest in the
dynamics of electronic circuits in the early 1980s, when new ideas and methods were introduced to the study of (periodic or non-periodic) oscillations
generated by nonlinear electronic circuits of low dimension. However, the theory usually only provides a framework for dierent phenomena that one may
nd in a given circuit. To perform an eective study of the actual dynamics it
is necessary to resort to numerical methods. In order to obtain a global view
of the dynamics in phase space and of the bifurcations in parameter space,
one needs to employ numerical methods that go beyond mere numerical simulation. Indeed, what is needed is the technique of numerical continuation that
has been developed since about the 1980s and is now available in the form
of several software packages, most notably the package Auto [19]; see also
Chaps. 1 and 2.
In this chapter we demonstrate how complicated dynamical behavior and
bifurcations can be found and identied in ODE models of electronic circuits.
The combination of theoretical methods and numerical techniques allows one
to obtain a deep understanding of a wide range of dynamical phenomena.
In particular, electronic circuits provide concrete examples of unfoldings of
singularities that act as organizing centers of the dynamics.
Specically, we consider in Sect. 7.1 a three-dimensional modied Van der
Pol oscillator as studied in [20]. We show that there are co-existing canard
periodic orbits, which we nd and continue with Auto. The core of the chapter is the bifurcation analysis in Sect. 7.2 of a three-dimensional ODE model
of a modied Van der Pol-Dung electronic circuit; see [28] and references
therein. This system exhibits very complex dynamics and associated bifurcation structures. We concentrate here on an extensive study of (global) dynamics associated with Arnold (or resonance) tongues and on a global bifurcation
known as a T-point bifurcation. In the process we identify a number of global
bifurcations, including homoclinic bifurcations, Shilnikov-Hopf bifurcations,
T-point bifurcations and T-point-Hopf bifurcations.

7.1 Canards in a Modied Van der Pol Circuit

We consider here the electronic circuit shown in Fig. 7.1. It is obtained from
the well-known Van der Pol circuit with a battery by adding a linear RC
parallel branch. This circuit was chosen as a convenient system to study canard
periodic orbits in a three-dimensional phase space.
Canard orbits arise due to the slow-fast nature of the system and have
rst been found in the Van der Pol equation with a battery (that is, without

7 Bifurcation Analysis of Electronic Circuits

R
f(x)

223

C0

Fig. 7.1. Circuit scheme of a three-dimensional modied Van der Pol circuit with
a parallel linear RC branch.

the RC branch in Fig. 7.1). Canards are composed of slow segments that
closely follow parts of the S-shaped slow manifold of the Van der Pol system,
and they exist in exponentially small parameter regions. Their existence was
rst shown with techniques from non-standard analysis [15]. The name canard
(French for duck) was adopted because the shape of these periodic orbits; see
also [18] and Chap. 8.
By applying Kirchhos laws and a suitable rescaling of the state and time
variables [20], the electronic circuit in Fig. 7.1 can be represented by the vector
eld

3
zx
x
y,
x +
x =
R
3
y = x a,
(7.1)
zx
.
z =
R

System (7.1) has exactly one equilibrium, which may undergo a degenerate
Hopf bifurcation. A numerical study in [20] with Auto and DsTool [33]
corroborates the analytical results and provides evidence of new global bifurcation phenomena, including cusp bifurcations of periodic orbits, perioddoubling bifurcations, and the presence of chaotic attractors.
We concentrate here on canard periodic orbits of (7.1). Figure 7.2 shows
the situation for R = 3, = 0.7 and = 0.001. Panel (a) is a plot of the period
of a periodic orbit as a function of the parameter a. The branch of periodic
orbits emerges from a Hopf bifurcation point H and then is almost vertical in
a very narrow interval of the parameter a. Indeed the periodic orbit grows
very fast in size in this interval until it takes the typical shape of relaxation
oscillations, which exist along the horizontal part of the branch. It is known
for the Van der Pol equation that the parameter interval of canard solutions is
exponentially small in . In other words: the life of canards is very short. The
sudden growth of the periodic orbit is also referred to as a canard explosion.
The enlargement in Fig. 7.2(b) of the narrow a-interval where canard orbits
exist shows that the branch of periodic orbits actually has four saddle-node

224

Emilio Freire and Alejandro J Rodrguez-Luis

(a)

(b)

(c)

(d)

Fig. 7.2. Canard orbits of (7.1) for R = 3 and = 0.7. Panel (a) shows the
bifurcation diagram for = 0.001, and panel (b) is an enlargement. Panel (c) shows
four co-existing canard periodic orbits for a = 0.7286 and = 0.001 in projection
onto the (x, y)-plane, and panel (d) the loci of folds in (a, )-plane. From E.J. Doedel,
E. Freire, E. Gamero and A.J. Rodrguez-Luis, An analytical and numerical study
c 2002 by
of a modied Van der Pol oscillator, J. Sound Vibr. 256 (2002) 755771 
Elsevier Science; reprinted with permission.

bifurcations of periodic orbits (folds with respect to a). As a consequence,

there are up to four co-existing canard orbits, for example, those for a = 0.7286
(labeled 1 to 4) that are shown in Fig. 7.2(c). Canard orbit 1 does not have a
head while orbits 24 are canards with a head. Note that the shape of orbit
4 is very close to a relaxation oscillation. The four folds can been continued in
the (a, )-plane, which results in the curves in Fig. 7.2(d); observe the presence
of two cusp points (one of them in the bottom right corner) on the fold curve.
We remark that the bifurcation diagram of Fig. 7.2(a) is very similar to
the one for a canard explosion in the Van der Pol equations; see [27] and
Chap. 8. According to [50, Fig. 4.7], the period of the periodic orbit does
not simply decrease from the value it has when it is a relaxation oscillation.
Instead, it rst increases up to a maximum (corresponding to the largest
canard without a head) and only then decreases very rapidly until the Hopf
bifurcation is reached. This agrees with what we nd for (7.1) in Fig. 7.2(a)
and (b), but note that the situation is richer because of the presence of saddlenode bifurcations.

7 Bifurcation Analysis of Electronic Circuits

225

(a)

(b)

(c)

(d)

(e)

0
-1

-1
-2

(f)

-1

-2

-1

Fig. 7.3. Periodic orbits of (7.1) and their bifurcations for R = 4, = 2 and =
0.25. Panel (a) shows the bifurcation diagram; solid curves indicate stable objects
and dashed curves unstable ones. Panel (b) shows three co-existing periodic orbits
for a = 0.93. Panel (c) is an enlargement of the bifurcation diagram near the upper
fold, where we now show the L2 -norm of the periodic orbits, and panel (d) shows
periodic orbit 7 in projection onto the (x, y)-plane. Panels (e) and (f) show chaotic
attractors for a = 0.939648 and for a = 0.9396485, respectively. From E.J. Doedel,
E. Freire, E. Gamero and A.J. Rodrguez-Luis, An analytical and numerical study
c 2002 by
of a modied Van der Pol oscillator, J. Sound Vibr. 256 (2002) 755771 
Elsevier Science; reprinted with permission.

Figure 7.3 shows the situation for R = 4, = 2 and = 0.25, that is, for a
larger value of . This allows us to study the bifurcation diagram in more detail
with Auto. The bifurcation diagram is shown in panel (a). When decreasing a,

226

Emilio Freire and Alejandro J Rodrguez-Luis

v1

v2

G2
iL

G1

C0

G3

Fig. 7.4. Scheme of the modied Van der Pol-Dung electronic oscillator.

the equilibrium becomes unstable at the supercritical Hopf bifurcation point

H, where a branch of stable periodic orbits emerges. This branch has two
folds, so that there are up to three co-existing periodic orbits (labels 13),
of which two are stable; see Fig. 7.3(b). To study the bifurcation diagram
near the upper saddle-node of periodic orbits, panel (c) shows an enlargement
that also shows period-doubling bifurcations and branches of period-doubled
orbits. Note that the branch of period-two orbits has a fold on the left. Indeed
we nd a cascade of period-doubling bifurcations; orbits 4 and 5 are on the
principal branch, orbit 6 on the branch of period-two orbits, orbit 7 on the
branch of period-four orbits, and orbit 8 on the branch of period-eight orbits.
Figure 7.3(d) shows period-four orbit 7 in projection onto the (x, y)-plane.
Because the periodic orbits undergoing successive period-doubling bifurcations are stable, one expects that there exist chaotic attractors in a narrow
a-interval. That this is indeed the case is shown in Fig. 7.3(e) and (f); where
the latter chaotic attractor is fully developed.

7.2 Bifurcations in a Modied Van der Pol-Dung

Circuit
Several dierent ways have been proposed to modify the classical Van der
Pol circuit to obtain an electronic circuit with three states. Examples of such
circuits can be found in [26, 31, 34, 37, 38, 43]; see also the review of dierent
congurations in [28].
In this section we consider the modied Van der Pol-Dung electronic
circuit sketched in Fig. 7.4. This system has been suggested as a random waveform generator [43], and it consists of a parallel RCL-circuit and an RC-circuit
that are coupled by a nonlinear conductance; see [28] and references therein.
To arrive at an ODE model, we take the voltages at the capacitors and the
current across the inductance as state variables, and model the current-voltage
characteristics of conductances by means of odd third-order polynomials. In
dimensionless variables the circuit is described by the three-dimensional vector

7 Bifurcation Analysis of Electronic Circuits

eld

x =

+
r


x+

227

A3 3 B3

(y x)3 ,
x +
y
r
r
r

y = x ( + )y z B3 (y x)3 C3 y 3 ,

(7.2)

z = y,
where r > 0 represents the ratio between the two capacitances. Note that
system (7.2) has a Z2 -symmetry due to its invariance under the transformation
(x, y, z)  (x, y, z).
We now summarize the local bifurcations of (7.2) in dependence on the
parameters , and . The origin is always an equilibrium. A pitchfork bifurcation of equilibria occurs on the plane { + = 0}, which creates two
symmetry-related equilibria that exist for + < 0. The origin as well as
the nontrivial equilibria undergo Hopf bifurcations. System (7.2) exhibits two
dierent kinds of Takens-Bogdanov bifurcations (double-zero eigenvalue).
The
r, r)}
rst is of homoclinic
type
and
occurs
on
the
straight
line
{(,
)
=
(

the
second
is
of
heteroclinic
type
and
occurs
on
the
straight
where = r;

line {(, ) = ( r, r)} where = r. A detailed analysis of the Hopf and

Takens-Bogdanov bifurcations in system (7.2) and their degeneracies can be
found in [4].
There are also Hopf-pitchfork bifurcations of (7.2), which occur on the
line segment {(, ) = (, )} where 2 < r. Furthermore, there is a
codimension-three singularity corresponding to a degeneracy of the Hopfpitchfork bifurcation, whose normal form and unfolding is the topic of [3].
How this bifurcation, which occurs for 0.4519, organizes the dynamics
of (7.2) is discussed in [5, 6, 7]. Moreover, system (7.2)exhibits a triple-zero
bifurcation at the critical values c = c = c = r; see [29] for more
information.
In this section we focus on two topics, namely the intriguing bifurcation
structure inside weak resonance tongues and global bifurcation phenomena,
especially those associated with T-point bifurcations.

7.2.1 Analysis of Arnold Tongues

We now present some results on the numerical analysis of Arnold tongues
in the system; see [8] and further references therein. Specically, we consider
(7.2) for xed r = 0.6, A3 = 0.3286, B3 = 0.9336 and C3 = 0.
Figure 7.5(a) shows the Floquet multipliers of the periodic orbit that undergoes a torus (or Neimark-Sacker) bifurcation HH for = 0.6; shown is the
the argument (angle between the horizontal axis and the Floquet multiplier
of the complex conjugate pair with positive imaginary part) versus parameter . When the argument fails to vary monotonically we say that the torus
bifurcation has an angular degeneracy [39]; this occurs in Fig. 7.5(a) at the

228

Emilio Freire and Alejandro J Rodrguez-Luis

.
35

( 0.5088 , 35.269 o)

1/25

(a)

25

1/15
1/12

Arg
15

(b)

1/20

0.57

1/11

0.54
0.51

0.48
5

A
0.45

0.50

2/25

HH

1/12

0.45

HP
0.60

0.55

0.56

1/25

-0.597

1/20

1/15

-0.595

-0.593

(d)

HH

1/12

2/23

0.53

-0.591

0.55

(c)
1/11

1/11
1/12

1/11

0.53
2/21

0.51

D
0.50

0.49

2/21
2/25

0.47
-0.594

1/12

HH

2/23

-0.593

1/11

1/12

1/11

-0.592

-0.591

0.47

-0.593

-0.592

-0.591

Fig. 7.5. Arnold tongues of (7.2) for = 0.6. Panel (a) shows the arguments of the
characteristic multipliers versus along the torus bifurcation curve; D is an angular
degeneration point. Panel (b) depicts Arnold tongues close to the torus bifurcation
curve HH of 1:p resonances for p = 11, 12, 15, 20 and 25, close to the HH curve, panel
(c) the Arnold tongues for the 2:21, 1:11, 2:23, 1:12 and 2:25 resonances, and panel
(d) details of the Arnold tongues for the 1:11 (closed resonance zone) and 1:12 (rst
open resonance zone) resonances. From A. Algaba, M. Merino and A.J. RodrguezLuis, Takens-Bogdanov bifurcations of periodic orbits and Arnolds tongues in a
three-dimensional electronic model, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 11
c 2001 by World Scientic Publishing; reprinted with permission.
(2001) 513531 

maximum D of the parabola-like curve (where the Floquet multipliers of the

periodic orbit reverse their direction of movement on the unit circle). The
appearance of this angular degeneracy is a consequence of the existence of the
point A, associated to a double +1 characteristic multiplier of a periodic orbit
(since the argument is zero at both endpoints of HH).
The value of the argument at the angular degeneracy point D (of 35.3
degrees) indicates that 1:p and 2:q resonances will appear on the curve HH
only for p 11 and q 21. Several of these 1:p Arnold tongues are shown in
Fig. 7.5(b) in a neighborhood of HH. In Fig. 7.5(c) we show, for = 0.6, the
angular degeneracy point D on the torus curve HH as well as the rst 1:p and
2:q resonance zones close to it. The numerical computations show that the
rst three (2:21, 1:11, 2:23) resonance zones are closed, which suggests that
the point D is of banana-type [39].

7 Bifurcation Analysis of Electronic Circuits

180
0.7

SN11

HH1
PD

120

SN11

0.6

60

0.5

0.682

SN 11

-0.67

0o
0.45

-0.65

0.470

(c)

HH1

HP
0.50

0.55

0.60

(d)

SN 11

0.670
-0.689

0.70

SN11

0.678

0.674

0.65

HH2

1/11

(b)

HH2

PD
-0.69

TBS1

TBS2

Arg

TBS1
TBS2

0.4
-0.71

(a)

HP

229

PD11

SN 11

PD11

1/11

SN11

HH2

HH1
-0.687

0.466

-0.685

0.462
-0.667

-0.665

-0.663

Fig. 7.6. Arnold tongues of (7.2) for = 0.69217. Panel (a) shows the partial
bifurcation set where the torus curve HH now has two parts, while panel (b) depicts
the arguments of the respective characteristic multipliers with the points TBS1 and
TBS2. Panels (c) and (d) show details of Arnold tongues on HH1 and on HH2, respectively. From A. Algaba, M. Merino and A.J. Rodrguez-Luis, Takens-Bogdanov
bifurcations of periodic orbits and Arnolds tongues in a three-dimensional electronic
c 2001 by
model, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 11 (2001) 513531 
World Scientic Publishing; reprinted with permission.

In the remainder of this section we will concentrate on the rst two lower
1:p resonances, which are the 1:11 and 1:12 resonances. Note that, for =
0.6, they correspond respectively to a closed resonance zone (1:11) and to
the rst open resonance zone (1:12) in the (, )-plane; see Fig. 7.5(d).
We now change to observe the evolution of the 1:11 Arnold tongue. When
is decreased, we detect that for c 0.69205 the torus curve collides at a
point TBS with a period-doubling curve of the asymmetric periodic orbit that
emerges in a Hopf bifurcation of the nontrivial equilibria.
For < c the torus curve appears to be split into two parts, as is shown
in Fig. 7.6(a) for = 0.69217. The rst part of the torus curve, HH1, joins
the Hopf-pitchfork point HP with the point TBS1 (where the periodic orbit has
a non-diagonalizable double Floquet multiplier 1). The second part of the
torus curve, HH2, connects the points TBS2 and A. At these codimension-two
points the periodic orbit has a non-diagonalizable double Floquet multiplier
1 and a diagonalizable double Floquet multiplier +1, respectively. In fact,
TBS1 and TBS2 correspond to cubic homoclinic-type Takens-Bogdanov bifur-

230

Emilio Freire and Alejandro J Rodrguez-Luis

TBS

HP TBS

HH

PD

< c

HH

TBS

HP TBS

PD

= c

HH

HH
HP

> c

PD
HH

A
Hd

DHP

Hd

HP

Fig. 7.7. Qualitative partial bifurcation set in (, , )-space that explains why
the torus curve splits; lled circles are points on the Hopf-pitchfork curve HP, lled
squares correspond to points on the Takens-Bogdanov curve TBS, lled triangles
stand for points on the Hopf-saddle-node of periodic orbits curve A, and inverted
lled triangles indicate points on the angular degeneracy curve D. From A. Algaba,
M. Merino and A.J. Rodrguez-Luis, Takens-Bogdanov bifurcations of periodic orbits
and Arnolds tongues in a three-dimensional electronic model, Internat. J. Bifur.
c 2001 by World Scientic Publishing;
Chaos Appl. Sci. Engrg. 11 (2001) 513531 
reprinted with permission.

cations of periodic orbits [36]. This splitting of the torus curve induces the disappearance of the angular degeneration point D, as can be seen in Fig. 7.6(b).
Figure 7.7 is a qualitative partial bifurcation set in (, , )-space that
explains why the torus curve splits. A degenerate Hopf-pitchfork point DHP
appears on the curve HP when it intersects with the curve Hd of degenerate
Hopf bifurcation of the origin. On one side of DHP a curve of points A appears.
The torus surface HH is bounded initially by the curves HP and A. As the
Floquet multipliers have argument zero on both curves, an angular degeneracy
curve D exists on the torus surface. Since the maximum value of the argument
increases when separating from DHP (decreasing ), there is a point where the
curve D ends (when the maximum is 180 degrees). This situation occurs exactly
when the surface of period-doubling bifurcations PD reaches the torus surface.
From this moment on, the torus surface is also bounded by a parabola-shaped
curve of Takens-Bogdanov bifurcations of periodic orbits TBS. This means
that the torus surface has a parabolic hole: it does not exist between the
two branches of TBS. Therefore, the points TBS1 and TBS2 appear in a slice of
constant .

7 Bifurcation Analysis of Electronic Circuits

231

.
0.54

0.54

(a)

0.52

SN11
SN11

(b)
SN11

0.52

SN11
0.50

0.50

SN11

SN 11
0.48
-0.595

-0.594

-0.593

0.48
-0.595

-0.594

-0.593

Fig. 7.8. Arnold tongue for the 1:11 resonance for = 0.60184 (a), and for
= 0.6018 (b). From A. Algaba, M. Merino and A.J. Rodrguez-Luis, TakensBogdanov bifurcations of periodic orbits and Arnolds tongues in a three-dimensional
c
electronic model, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 11 (2001) 513531 
2001 by World Scientic Publishing; reprinted with permission.

Now that we know how the torus loci change due to the presence of TakensBogdanov bifurcations of periodic orbits, we focus on Arnold tongues and
their evolution. We will see that two dierent types of Takens-Bogdanov bifurcations of periodic orbits will be also present, including a cascade of one of
them.
Figure 7.6(c)-(d) shows how the 1:11 Arnold tongues emerge from HH1
and HH2, respectively. Note that the right branch of folds that emerges from
the 1:11 resonance on HH1 crosses this curve when it moves away from its
starting point on HH1; see Fig. 7.6(c). The same happens for the right curve of
saddle-node bifurcations starting at the 1:11 resonance on HH2; see Fig. 7.7(d).
Such a crossing implies that periodic orbits (of approximately eleven times the
period of the principal periodic orbit) exist on both sides of the curves HH1 and
HH2. This phenomenon is not a consequence of the splitting of the torus curve
(since it also occurs, for example, for = 0.65 when there is only one torus
curve). Rather it is due to the evolution of the curves HH and the boundaries
(saddle-node bifurcations of periodic orbits) of the 1:11 resonance zones with
. (In the (, , )-space the surfaces HH and SN11 intersect independently of
the collision of the surface PD with HH.)
The continuation of these saddle-node curves of periodic orbits shows that
both right branches are connected, whereas the left branches are disconnected;
see Fig. 7.6(a). (Recall that for = 0.6 the 1:11 resonance zone is a closed
region.) Moreover, in the present situation the repelling periodic orbit undergoes a period-doubling bifurcation (which is again dierent from the behavior
for = 0.6); see the curve PD11 in Fig. 7.6(c) and (d), which has not been
included in panel (a) as it would be indistinguishable from the saddle-node
curves that limit the 1:11 resonance zone.
The question arises how the resonance zones evolve from being open to
being closed when changes between = 0.69217 and = 0.6. To see how

232

Emilio Freire and Alejandro J Rodrguez-Luis

.
0.510

(a)
PD 11

0.510

(b)
PD11

HH11

0.505

HH 11

0.505

SN11
0.500
-0.5941

-0.5939

SN11

-0.5937

TBS211

TB2 11

-0.5937

-0.5935

(d)

PD12
TBS112
0.503

-0.5937

TBS2 12

SN12

HH12

SN 11
-0.5939

TB111

HH211

0.500
-0.5941

-0.5939

0.506

(c)

HH111
TBS111

0.505

0.500
-0.5941

-0.5935

0.510

PD11

TB111
TB2 11

-0.5935

0.500
-0.5910

-0.5908

-0.5906

-0.5904

Fig. 7.9. Partial bifurcation set inside the Arnold tongues of the open 1:11 resonance zones for = 0.601285 (a), = 0.60134 (b), and = 0.60135 (c). Panel
(d) shows the partial bifurcation set inside the 1:12 Arnold tongues of the open 1:12
resonance zone for = 0.5973. From A. Algaba, M. Merino and A.J. RodrguezLuis, Takens-Bogdanov bifurcations of periodic orbits and Arnolds tongues in a
three-dimensional electronic model, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 11
c 2001 by World Scientic Publishing; reprinted with permission.
(2001) 513531 

this happens we increase from = 0.69217. At 0.60184 the upper

and lower left branches of the resonance zone touch, so that, for smaller ,
two separate resonance regions are created; see Fig. 7.8(a) and (b). Note that
the saddle-node curves that bound the closed resonance zone emerge from the
curve HH but the curves of the open zone are not related to HH.
We conclude that, for the initial value of = 0.6, there exists not only
the closed 1:11 resonance zone shown in Fig. 7.5(d) but also the open zone
in Fig. 7.8(a). In fact, this open resonance zone exists even before the closed
resonance zone (which appears for 0.5957).
Note that inside the closed 1:11 resonance zone the periodic orbits do
not exhibit any bifurcation. Both are hyperbolic, one of saddle type and the
other repelling. However, the study of the periodic orbits inside the open 1:11
resonance zone for = 0.6 shows that the stability of one of them changes
relative to the periodic orbits inside the closed resonance zone for the same
value of . Now one periodic orbit is a saddle and the other is attracting. The
latter undergoes a period-doubling bifurcation when crossing the curve PD11 .

7 Bifurcation Analysis of Electronic Circuits

233

Obviously, the change in the stability of one of the periodic orbits in the
1:11 resonance zones indicates that, as varies, some additional bifurcation
has to be present on the saddle-node curve that bounds the open resonance
zone. This bifurcation is necessary to make possible the contact between the
open and the closed 1:11 resonance zones. To understand this situation we
investigate the partial bifurcation set inside this region for values of close
to the value where the two regions join.
Moving again the control parameter we see in Fig. 7.9(a) that a closed
torus curve HH11 of the 11-period orbit appears in the parameter plane (this
curve does not exist for = 0.60128). As decreases, this closed curve approaches the saddle-node curve that bounds the open region of the 1:11 resonance. In this way, two quadratic homoclinic-type Takens-Bogdanov points of
periodic orbits TB111 and TB211 are created; see Fig. 7.9(b). For even lower ,
when the torus curve HH11 interacts with the curve of period-doubling PD11 ,
it splits into two curves HH111 and HH211 ; see Fig. 7.9(c). Two new TakensBogdanov points of periodic orbits (symmetric and of homoclinic type), TBS111
and TBS211 , appear. Note that both curves HH111 and HH211 connect two
Takens-Bogdanov points that are on the saddle-node curve SN11 and on the
period-doubling curve PD11 , respectively.
We have checked whether the presence of closed torus curves also occurs
for other resonances. The answer is armative, but a dierence may appear.
In the above case of a 1:11 resonance, the torus curve rst collides with SN11
and later with PD11 , whereas in other resonances the torus curve rst collides
with the period-doubling curve and later with the saddle-node curve. This is
illustrated for the 1:12 resonance in Fig. 7.9(d). The curve HH12 starts and
ends at two symmetric homoclinic-type Takens-Bogdanov points of periodic
orbits, TBS112 and TBS212 .
Several possible scenarios are proposed in [39] where angular degeneracy
points are present, all involve Hopf-Hopf bifurcations and most involve TakensBogdanov points. These authors wonder whether there is some model that
presents such a behavior in relation to the torus curve and its resonance
tongues; see [39, Fig. 5(d)-(e)]. The three-dimensional autonomous model (7.2)
considered here exhibits four of the ve possible situations for the global
continuation of a Hopf-Hopf curve in a two-parameter family, namely:
1. continuation in each direction may terminate at a quadratic TakensBogdanov point (non-diagonalizable double Floquet multiplier +1); see
Fig. 7.9(b);
2. continuation in each direction may terminate at a cubic homoclinic
Takens-Bogdanov point (nondiagonalizable double Floquet multiplier 1);
this was found for the 1:12 resonance in Fig. 7.9(d);
3. continuation in one direction may terminate at a quadratic TakensBogdanov point, while continuation in the other direction terminates at
a cubic homoclinic Takens-Bogdanov point; see Fig. 7.9(c); and
4. continuation may provide a closed curve; see Fig. 7.9(a).

234

Emilio Freire and Alejandro J Rodrguez-Luis

0.55

HH112
0.53

TB112

0.49

PD12
TBS212
HH212

0.47
-0.594

HH112

SN12
HH

SN12

TB112

PD48
TBS112

0.5218

HH124
PD24

TB212
-0.593

(b)

PD12

SN12

TBS112
0.51

0.5228

(a)

-0.592

-0.591

0.5208
-0.5932

-0.5929

SN 12

-0.5926

-0.5923

Fig. 7.10. Arnold tongues for the 1:12 resonance for = 0.6 (a), and detail inside
the 1:12 resonance zone of the upper region (b). From A. Algaba, M. Merino and
A.J. Rodrguez-Luis, Takens-Bogdanov bifurcations of periodic orbits and Arnolds
tongues in a three-dimensional electronic model, Internat. J. Bifur. Chaos Appl.
c 2001 by World Scientic Publishing; reprinted
Sci. Engrg. 11 (2001) 513531 
with permission.

After we discussed the evolution of the Arnold tongues from closed to

open (when is changed) and how Takens-Bogdanov points of periodic orbits
appear, we now describe the bifurcation set in a neighborhood of these TakensBogdanov points. To illustrate that the respective dynamical behavior occurs
generically for all the weak 1:p resonances (not only for the 1:11 resonance)
we consider the 1:12 resonance for = 0.6. As mentioned before, it is the
rst resonance that appears in an open region.
Figure 7.10(a) shows the principal torus curve HH and the Arnold tongues
of the 1:12 resonance SN12 . Notice how the torus curves HH112 and HH212 connect the saddle-node curves SN12 with the period-doubling curve PD12 . To see
clearly what bifurcations are present we show in Fig. 7.10(b) an enlargement
of the region where the curve HH112 exists. We observe a cascade of torus
bifurcations HH112 , HH124 , and so on, that connects SN12 with PD12 , PD12 with
PD24 , and so on. Therefore, we have found a cascade of Takens-Bogdanov bifurcations of periodic orbits. The rst point TB12 is of quadratic type whereas
the other Takens-Bogdanov points (TBS112 , TBS124 , and so on) are of cubic
homoclinic type.
7.2.2 Isolas, Cusps and Global Bifurcations
We start with a regime that is organized by homoclinic connections of
Shilnikov type. Specically, we report on a detailed numerical study of one
type of periodic orbit that exhibits an isola structure in the bifurcation diagram and cusps of saddle-node bifurcations in a two-parameter bifurcation
set; see [23] for details. The mechanism of isola-creation and its relationship
with global connections is also described. To this end we consider (7.2) for
xed = 0, A3 = 0.3286, B3 = 0.9336 and C3 = 0.

7 Bifurcation Analysis of Electronic Circuits

(a)

235

(b)

Fig. 7.11. Projection onto the (x, y)-plane of a periodic orbit for = 0.4, r = 0.6
(a) that gives rise to a gure-8 isola (b) when the period is plotted against . From
F. Fern
andez-S
anchez, E. Freire and A.J. Rodrguez-Luis, Isolas, cusps and global
c 1997
bifurcations in an electronic oscillator, Dynam. Stab. Sys. 12 (1997) 319336 
by Taylor & Francis; reprinted with permission.

F3

Period

S3

F2
Saddle-Node
Period Doubling

S2

stable
Non-Stable

S4

S1
F4

Period T
Period 2T
Period 4T
Period 8T

F1

Fig. 7.12. Schematic partial bifurcation diagram for = 0.4, r = 0.6 of the periodic orbits with periods 1, 2, 4 and 8; solid curves indicate stable orbits, while
dashed curves correspond to orbits of saddle-type. For the sake of clarity, only two
branches of the 8-periodic orbits are drawn. From F. Fern
andez-S
anchez, E. Freire
and A.J. Rodrguez-Luis, Isolas, cusps and global bifurcations in an electronic oscilc 1997 by Taylor & Francis; reprinted
lator, Dynam. Stab. Sys. 12 (1997) 319336 
with permission.

236

Emilio Freire and Alejandro J Rodrguez-Luis

.
2

(a)

-2

-2

(c)

(d)

-2

-2
-2

-2

-2

(b)

-2

Fig. 7.13. Evolution of the oscillation-sliding phenomenon. Projection of the periodic orbits onto the (x, y)-plane for = 1.041162 (a), = 1.793874 (b),
= 2.045187 (c), and = 2.119680 (d). From A. Algaba, F. Fern
andez-S
anchez,
E. Freire, E. Gamero and A.J. Rodrguez-Luis, Oscillation-sliding in a modied van
c 2002 by
der Pol-Dung electronic oscillator, J. Sound Vibr. 249 (2002) 899907 
Elsevier Science; reprinted with permission.

Numerical continuation of the periodic orbit shown in Fig. 7.11(a) shows

that its branch in the bifurcation diagram is a closed curve in the shape of
a gure-of-eight; see Fig. 7.11(b). The stability of the orbits in this isola is
shown in Fig. 7.12. Four saddle-node bifurcations, S1 to S4, and four perioddoubling bifurcations, F1 to F4, appear on the gure-8 isola. Also sketched in
Fig. 7.12 is the intricate arrangement of the branches of the rst sub-harmonic
periodic orbits, some of which form closed loops and others are S-shaped.
7.2.3 Oscillation-Sliding Between Two Periodic Regimes
We now consider a type of periodic behavior exhibited by (7.2) near a degenerate Hopf-pitchfork bifurcation [1], where we x parameters to r = 0.6,
A3 = 0.5, B3 = 0.01, and C3 = 0.1. A degenerate case of the Hopf-pitchfork
bifurcation takes place at 0.2473, and we x close to this special value,
namely at = 0.24.
In this context, the principal periodic orbit undergoes a secondary Hopf
bifurcation that gives rise to the appearance of an invariant torus. Its breakup results in the presence of resonance phenomena with subharmonic periodic

7 Bifurcation Analysis of Electronic Circuits

237

.
30

(a)

(b)
29

Period

Period

21

20

28

27
2.0

1.5

(c)

75

35

1.5

(d)

70

65

33
2.0

2.0

1.0

Period

Period

37

1.6

1.2

2.0

1.6

1.2

Fig. 7.14. Bifurcation diagram of the period versus for 3T-periodic orbits (a),
4T-periodic orbits (b), 5T-periodic orbits (c), and 10T-periodic orbits (d).

orbits. The numerical continuation of a 3T-periodic orbit in Fig. 7.13 shows

that the periodic orbit slides between two dierent oscillation regimes: a
small- and a large-amplitude periodic orbit.
The bifurcation diagram of this 3T-periodic orbit corresponds to an
isola, shown in Fig. 7.14(a), with sixteen saddle-node bifurcations. But this
oscillation-sliding behavior seems to be present as well in all other subharmonic periodic orbits. Figure 7.14(b)(d) shows the corresponding bifurcation
diagrams for the 4T-, 5T- and 10T-periodic orbits, respectively. Observe the
nice isolas obtained and some rules that they seems to follow. For instance,
the saddle-nodes on the right are organized by pairs that occur approximately
at the same parameter value; there are exactly n 2 pairs in the nT-periodic
orbit branch. A more detailed understanding of this intricate and aesthetic
dynamical behavior remains a challenge for future research.
7.2.4 T-Point Bifurcation
In a three-dimensional system with at least two equilibria a T-point occurs
when the one-dimensional unstable manifold of one equilibrium and the onedimensional stable manifold of the other equilibrium coincide. At the same
time, the two-dimensional manifolds of the two equilibria have a transversal

238

Emilio Freire and Alejandro J Rodrguez-Luis

(a)

C-

0.01

(b)

0.00

-0.01
-0.10

-0.05

0.00

0.05

0.10

Fig. 7.15. Panel (a) is a sketch of a heteroclinic T-point cycle in phase space;
also drawn (dashed line) is its symmetric counterpart. Panel (b) shows the primary T-point heteroclinic orbit in projection onto the (x, y)-plane that exists for
= 0.6 and (, ) (0.7605, 0.7548). From F. Fern
andez-S
anchez, E. Freire
and A.J. Rodrguez-Luis, T-points in a Z2 -symmetric electronic oscillator, Nonlin.
c 2002 by Springer; reprinted with permission.
Dynam. 28 (2002) 5369 

intersection that forms a heteroclinic loop between them. This codimensiontwo heteroclinic loop is usually referred to as a T-point bifurcation.
The unfolding in the vicinity of a T-point in a system with one real saddle
(with three real eigenvalues) and a saddle-focus has been analyzed in the
literature. Glendinning and Sparrow [30] consider T-point bifurcations of this
sort in the Lorenz systems (which has an extra symmetry); Bykov considers
in [12] also the case of two real saddles and in [13, 14] the case where both
equilibria are saddle-foci. T-point bifurcations in Z2 -symmetric systems are
considered in [24], where it is shown by means of a Shilnikov-type analysis
that three spiral curves of codimension-one global bifurcations emerge from
the T-point. The rst corresponds to homoclinic orbits to the origin, the
second to homoclinic connections of the nontrivial equilibria, and the third to
heteroclinic orbits between the nontrivial equilibria. Figure 7.15(a) shows a
sketch of the heteroclinic T-point cycle between the equilibrium at the origin
and the two (symmetry-related) nontrivial equilibria, C . The four planes i
are used in the construction of a Poincare return map; the ow is divided into
four parts; the four corresponding maps are composed [24].
To study T-point bifurcations in (7.2) we x parameters at r = 0.6,
A3 = 0.3286, B3 = 0.9336 and C3 = 0 in this section. Then for = 0.6 a primary T-point exists for (, ) (0.7605, 0.7548); it is shown in Fig. 7.15(b).
Figure 7.16 shows the three curves HO, HNT and Het of global bifurcations that
were introduced above. When plotted in the (, )-plane, as in panel (a), the
three spirals are so close that it is almost impossible to distinguish them. To
open up these curves (one by one) one may perform successive changes in the
parameters (one translation, one rotation and one rescaling). Figure 7.16(b)
shows the result of such a rescaling to new parameters  and  for the
homoclinic orbit of the origin HO. Observe how this curve spirals around the
T-point. To open up the other two curves HNT and Het dierent changes of
parameters are needed.

7 Bifurcation Analysis of Electronic Circuits

239

0.3

0.77

(a)

Hom

(b)

Het HNT
0.2

0.76

0.1

0.75

Hom

Hom

0.0
-0.1

0.74

-0.2
Het HNT
0.73

-0.770

-0.765

-0.760

-0.755

-0.750

-0.3
-0.02

-0.01

0.00

0.01

0.02

Fig. 7.16. Curves spiraling around the T-point for = 0.6 of homoclinic orbits
of the origin, HO, homoclinic orbits of the nontrivial equilibria, HNT, and heteroclinic
orbits between the nontrivial equilibria, Het, shown in the (, )-plane (a), and in
the ( ,  )-plane (b). From F. Fern
andez-S
anchez, E. Freire and A.J. RodrguezLuis, T-points in a Z2 -symmetric electronic oscillator, Nonlin. Dynam. 28 (2002)
c 2002 by Springer; reprinted with permission.
5369 

Finally we illustrate in Fig. 7.17 how the homoclinic connection of the

origin approaches the nontrivial equilibrium on its way towards the T-point.
Panel (a) shows the curve of homoclinic connections of the origin for = 0.2
as it spirals into the T-point at (, ) (0.7098, 0.4796). We focus our attention on three points on this curve (marked by dots) that lie to the left
of the T-point and have the same value of . The corresponding homoclinic
connections are shown in Fig. 7.17(b)(d). Notice how the homoclinic orbit
makes, roughly speaking, one more turn around the (left) nontrivial equilibrium when the homoclinic curve turns one more time around the T-point in
the parameter plane. This is in accordance with the theoretical predictions
[24, Fig. 6].
7.2.5 T-Point-Hopf Bifurcation
It is possible to continue a curve of T-points in a three-dimensional parameter space with the package Auto. To perform this calculation we consider
the linear approximations of the one-dimensional manifolds of the equilibria
involved in this heteroclinic loop. Further, we assume that a transversal intersection between the respective two-dimensional manifolds of the equilibria
(the orbit that closes the loop) indeed exist. In our calculation we consider
system (7.2) for xed A3 = 0.3286, B3 = 0.9336 and C3 = 0.
The continuation in (, , )-space of the primary T-points leads to the
curve shown in Fig. 7.18(a). It starts at a triple-zero degeneracy of the origin,
TZ, and ends at the point H when it intersects the surface where the origin exhibits a Hopf bifurcation. Two more degenerate points are detected along this
curve, both when the Shilnikov quotient (also called the saddle-quantity)
of the origin takes the special values = 1 at D1 (neutral saddle-focus) and

240

Emilio Freire and Alejandro J Rodrguez-Luis

0.60

(a)

0.2

(b)

0.55

0.1

0.50

0.0

0.45

-0.1
0.40

-0.2
0.35
-0.725

0.2

-0.720

-0.715

-0.710

-0.6

-0.705

(c)

0.2

0.0

0.0

0.2

0.4

0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

(d)

0.0

-0.1

-0.1

-0.2

-0.2

-0.2

0.1

0.1

-0.4

-0.6

-0.4

-0.2

0.0

0.2

0.4

0.6

-0.6

Fig. 7.17. Curve of homoclinic connections of the origin spiraling around the
T-point for = 0.2 (a). Panel (b)(d) are projections onto the (x, y)-plane
of three homoclinic orbits corresponding to the three points (from left to right)
that are marked by dots in panel (a). From F. Fern
andez-S
anchez, E. Freire and
A.J. Rodrguez-Luis, T-points in a Z 2 -symmetric electronic oscillator, Nonlin. Dyc 2002 by Springer; reprinted with permission.
nam. 28 (2002) 5369 

= 1/2 at D05 (neutrally-divergent saddle-focus). Some features of the complex dynamics originating in these codimension-two homoclinic bifurcation
can be found in [16] and references therein.
As in each plane of constant , the homoclinic curve of the origin spirals
around the T-point. When is added as the third bifurcation parameter, a
surface of homoclinic connections is expected (at least locally) to spiral around
the T-point bifurcation curve TP. On this surface, there will appear curves of
codimension-two homoclinic connections that end at the corresponding point
on the T-point curve. In this way, the curve of degenerate homoclinic bifurcations of the origin with = 1 appears in Fig. 7.18(b) in the vicinity of the
point D1 where it ends on the curve TP. Similarly, in Fig. 7.18(c) the curve of
degenerate homoclinic bifurcations of the origin with = 1/2 is shown spiraling around TP. Finally, the curve of Shilnikov-Hopf homoclinic connections
appears in Fig. 7.18(d); it ends at the T-point-Hopf point H in panel (a).
We now investigate the evolution of the primary homoclinic connection in
the (, )-parameter plane when the curve of T-points approaches its endpoint,
where a T-point-Hopf bifurcation occurs. In fact, the T-point curve TP emerges
from a triple-zero degeneracy that occurs for 0.7746. When increasing

7 Bifurcation Analysis of Electronic Circuits

241

.
0.8

(a)

(b)

TZ
D1

0.71

0.6

D05

0.4

D1

0.70

D1
D1
D1

0.69

0.2

TP

H
0.85

0.56

(c)

0.80

0.75

0.70

0.74

TP

0.4

0.53

D05

0.73

(d)
TP

0.3
0.2

0.50

0.1
TP

0.47
0.72

0.71

0.70

0.0

1.1

1.0

0.9

0.8

0.7

Fig. 7.18. Projections onto the (, )-plane of global bifurcation curves that
have been continued in (, , )-space. Panel (a) shows the curve of primary T-points; the following points are marked on this curve: triple-zero
point TZ at (, , ) (0.77457, 0.77457, 0.77457); = 1 point D1 at
(, , ) (0.73691, 0.70745, 0.45225); = 1/2 point D05 at (, , )
(0.70614, 0.52746, 0.22966); a T-point-Hopf H at (, , ) (0.86103,
0.10899,0.11867). Panel (b) shows the curve of degenerate homoclinic bifurcations
of the origin with = 1, panel (c) the curve of degenerate homoclinic bifurcations
of the origin with = 1/2, and panel (d) the curve of Shilnikov-Hopf homoclinic
bifurcations; the dashed curve TP is the curve of primary T-points.

, for instance to = 0.6, the curve Hom of primary homoclinic bifurcation

of the origin emerges from a Takens-Bogdanov bifurcation point and ends
spiraling into a T-point; see Fig. 7.16(a). For 0.5921 there is a tangency
(at (, ) (0.6331, 0.4598)) between the curve Hom and the curve H of
Hopf bifurcation of the origin, which gives rise to a so-called non-transverse
Shilnikov-Hopf bifurcation point [17]. (See Sect. 7.2.6 below for some more
comments about this bifurcation.) Note that this non-transverse bifurcation
appears as a consequence of the upper limit point (with respect to ) that
the Shilnikov-Hopf curve exhibits; see the upper-right corner of Fig. 7.18(d)
(although it shows the (, )-plane).
For above the non-transverse critical value, two Shilnikov-Hopf bifurcations appear. The situation for = 0.3 is drawn in Fig. 7.19(a). One branch
of the curve joins the Takens-Bogdanov point TB with a Shilnikov-Hopf point

242

Emilio Freire and Alejandro J Rodrguez-Luis

.
TB

SF

(a)

0.7
D1

TB SF

(b)

D1
D05

D05

0.5

D1

D1

0.6

D05

0.4

0.3

SH

0.75

0.7

Hopf

SF

0.8

D1

0.6

Hopf

0.70
TB SF
D1
D05

0.3

0.4
0.2

Hopf

0.1
0.9

SH

0.75

(d)

D05

0.5

SH

SH

0.80

0.70
TB

(c)

0.2 SH

SH

0.8

0.7

0.0

SH

Hopf

1.1

1.0

0.9

0.8

Fig. 7.19. Curve of homoclinic connections to the origin in the (, )-parameter

plane for = 0.3 (a), = 0.2 (b), = 0.12 (c), and = 0 (d). The dashed
curve Hopf corresponds to a Hopf bifurcation of the origin; also marked on the
homoclinic curve are points TB of Takens-Bogdanov bifurcation, transition points SF
of the equilibrium from saddle to saddle-focus, points D1 where = 1, and D05 where
= 0.5, and Shilnikov-Hopf points SH. For sake of clarity, we have not marked all
SH points in panels (c) and (d).

SH, whereas the second branch emerges from another SH point and ends spiraling around the T-point. The codimension-two homoclinic bifurcations that
occur when the origin changes from saddle to saddle-focus (marked as SF),
when = 1 (marked as D1), and when = 1/2 (marked as D05) are also
indicated in the picture.
When is increased to = 0.2, as in Fig. 7.19(b), a new pair of
Shilnikov-Hopf points appear. This is due to the existence of another nontransverse Shilnikov-Hopf bifurcation for 0.2594. Now three dierent
branches of the homoclinic curve exist.
As the T-point-Hopf for 0.11867 is approached, new non-transverse
Shilnikov-Hopf bifurcations appear and the curve of homoclinic connections
splits into more branches. This phenomenon is a direct consequence of the
limit points on the Shilnikov-Hopf curve as it spirals around the T-pointHopf bifurcation point. For example, for = 0.12 eighteen Shilnikov-Hopf
bifurcations occur; see Fig. 7.19(c). Note that innitely many Shilnikov-Hopf
points are predicted by the time the T-point-Hopf occurs [25].

7 Bifurcation Analysis of Electronic Circuits

243

.
(a)

(b)

0.5

0.5

0.0

0.5

0.5
0.3

0.0

0.0

0.3

0E

Fig. 7.20. One of the two homoclinic orbits to a nontrivial equilibrium for = 0.9.
Shown are its projection on to the (x, y)-plane (a), and its time prole of variable z
(b).

Past the T-point-Hopf bifurcation, branches of the homoclinic curve disappear as a consequence of non-transverse Shilnikov-Hopf points (now in the
downward-pointing case [17]). For example, for = 0, as in Fig. 7.19(d),
the homoclinic curve has four branches and only seven Shilnikov-Hopf points
remain. Furthermore, as Fig. 7.20(a) shows, the homoclinic orbit of the nontrivial equilibria winds closely around the periodic orbit that emerged in the
Hopf bifurcation before returning to the equilibrium. In this way, a heteroclinic loop is formed between an equilibrium and the periodic orbit [25]. The
time prole in Fig. 7.20(b) shows that the homoclinic orbit indeed spends a
lot of time near the saddle periodic orbit. Indeed the orbit in Fig. 7.20 is a
nice numerical approximation of the heteroclinic loop. The loop itself consists
of an intersection between the two-dimensional unstable manifold of the nontrivial equilibrium and the stable manifold of the saddle periodic orbit, and
an intersection between the unstable manifold of the periodic orbit and the
one-dimensional stable manifold of the nontrivial equilibrium.
7.2.6 Non-Transverse Shilnikov-Hopf Bifurcation
The theoretical analysis of the non-transverse Shilnikov-Hopf bifurcation
[17, 25] shows that it contains codimension-two non-transversal homoclinic
orbits to equilibria and non-transversal homoclinic tangencies to periodic orbits in its unfolding. Two cases are classied: the downward-pointing and the
upward-pointing case, depending on whether the variation of a third parameter causes either the annihilation of a locus of saddle-focus homoclinic orbits to
equilibria, or the uncoupling of this locus from the locus of Hopf bifurcations.
The downward-pointing case of a non-transverse Shilnikov-Hopf bifurcation
is shown to cause two wiggly curves to coalesce and leave behind nitely many
isolas of periodic orbits. The upward-pointing case, on the other hand, causes
two wiggly curves to coalesce rst into innitely many and then into nitely
many isolas.

244

Emilio Freire and Alejandro J Rodrguez-Luis

(a)

(b)

Fig. 7.21. Bifurcation diagrams for = 0.65 for asymmetric periodic orbits in the
upward-pointing case for = 0.62 (a) and = 0.58 (b). From A.R. Champneys and
A.J. Rodrguez-Luis, The non-transverse Shilnikov-Hopf bifurcation: uncoupling of
c 1999
homoclinic orbits and homoclinic tangencies, Physica D 128 (1999) 130158 
by Elsevier Science; reprinted with permission.

(a)

(b)

(c)

(d)

Fig. 7.22. Bifurcation diagrams for = 0 for asymmetric periodic orbits in the
downward-pointing case for = 0.12 (a), = 0.14 (b), = 0.15 (c), and = 0.16
(d). From A.R. Champneys and A.J. Rodrguez-Luis, The non-transverse ShilnikovHopf bifurcation: uncoupling of homoclinic orbits and homoclinic tangencies, Physica
c 1999 by Elsevier Science; reprinted with permission.
D 128 (1999) 130158 

Numerical evidence of this bifurcation was found in the electronic circuit

(7.2) for xed r = 0.6, A3 = 0.3286, B3 = 0.9336 and C3 = 0.
Figure 7.21 shows two -slices for the upward-pointing case for = 0.65.
Panel (a) for = 0.62 is close to the turning point of the homoclinic locus,
and the two homoclinic orbits on the primary locus are connected by a single

7 Bifurcation Analysis of Electronic Circuits

245

.
0.8

(a)

TZ

(b)

0.6

0.6

0.5
0.4

0.2
0.80

0.75

(c)

0.6

0.30

0.25

0.20

(d)

0.6

0.5

0.5

0.4

0.4

0.70

0.30

0.25

0.20

0.4

0.30

0.25

0.20

Fig. 7.23. Panel (a) shows the projection of the T-point bifurcation curve onto
the (, )-plane. It emerges from a triple-zero degeneracy, TZ, and has a fold for
0.7059. Panels (b)(d) show partial bifurcation sets consisting of curves of
homoclinic orbits to the origin for = 0.7085, = 0.7084 and = 0.7047,
respectively. From A. Algaba, F. Fern
andez-S
anchez, E. Freire, M. Merino and
A.J. Rodrguez-Luis, Nontransversal curves of T-points: a source of closed curves of
c 2002 by Elsevier Science;
global bifurcations, Phys. Lett. A 303 (2002) 204211 
reprinted with permission.

branch of asymmetric periodic orbits. There are two isolas for small period,
and the two wiggly curves approaching the two homoclinic orbits are connected. As is decreased, the two wiggly curves can be seen to annihilate each
other by forming more and more isolas. Figure 7.21(b) shows for = 0.58 the
rst eight isolas in an evident destruction process.
To nd the downward-pointing case it is necessary to increase to beyond
0.1187, that is, to the other side of the T-point-Hopf bifurcation. Figure 7.22 shows bifurcation diagrams in four -slices for = 0 that illustrate
the process in which isolas are created and destroyed.
7.2.7 Non-Transversal T-Point Bifurcation
A model to explain the existence of closed bifurcation curves of homoclinic and
heteroclinic connections in autonomous three-dimensional systems is derived
in [2]. This scenario is related to the failure of transversality in a curve of

246

Emilio Freire and Alejandro J Rodrguez-Luis

T-points. The predictions deduced from this model strongly agree with the
numerical results obtained for system (7.2) for xed r = 0.6, A3 = 0.3286,
B3 = 0.9336 and C3 = 0. This phenomenon was also found in an ODE model
of a laser with optical injection; see [48] and Chap. 6.
The presence of a fold in the curve of T-point bifurcations in the (, , )space is observed in Fig. 7.23(a). This curve emerges from a triple-zero degeneracy of the origin, marked TZ [29]. The evolution of the curves of homoclinic
orbits to the origin in the vicinity of the non-transversal T-point are shown
in the (, )-plane for several values of in Fig. 7.23(b)(d). The rst two
slices, each with two T-points, show how the rst closed curve of homoclinic
connections is formed. On the other side of the critical value of , where the
fold occurs, only closed curves appear; Fig. 7.23(d) shows the last two of them.
We remark that other bifurcation curves in the bifurcation set (for example, saddle-node and period-doubling bifurcations) must be expected to
be inuenced by these changes to the curves of homoclinic connections. Furthermore, some additional degeneracies may be exhibited by the global connections, which would imply an even richer bifurcation scenario; cf. [16] and
Chap. 6.
7.2.8 Bi-Spiraling Curves of Homoclinic Orbits Around a T-Point
In [22] a model was proposed that considers a non-transversal intersection
between the two-dimensional manifolds of the saddle-focus equilibria involved
in a T-point. The study of this model shows the presence of bi-spiraling
curves of homoclinic connections in the parameter plane: the spiral curve
that emerges from a T-point between two saddle-focus equilibria ends at the
same T-point, which it enters via a dierent spiral. The predictions deduced
from this model strongly agree with the numerical results obtained for (7.2)
where we x = 0.6, r = 0.6, A3 = 0.3286, B3 = 0.9336 and C3 = 0.
Due to the Z2 -symmetry, three curves of global connections emerge from
every T-point. However, in the case we consider here, only the curve of homoclinic connections to the origin bi-spirals around the T-point TP, as is shown in
Fig. 7.24(a). The other two curves (of homoclinic and heteroclinic connections
to the nontrivial equilibria, respectively) emerge from TP but do not return to
it. This is why they are not shown in the gure. As the bi-spiraling homoclinic
curve is very close to itself when shown in the plane (, ), we proceed to open
it up with a combination of translations, rotations and scalings as we did in
Fig. 7.16. The result in the rescaled ( , )-plane is shown in Fig. 7.24(b),
where the bi-spiraling is now clearly visible.
Finally, Fig. 7.24(c) and (d) show the two heteroclinic cycles that co-exist
at the T-point TP: they both have the same codimension-two connection between their one-dimensional manifolds (solid lines), but dier in the transversal intersection between the two-dimensional manifolds (dashed lines).

7 Bifurcation Analysis of Electronic Circuits

247

.
(a)

(b)
0.001

TP

0.7

0.000

0.6

0.001
0.75

0.70

(c)

0.03

0.00

0.70

0.65

(d)

0.03

0.00

0.03

0.03

0.06
0.2

0.75

0.65

0.1

0.0

0.1

0.2

0.06
0.2

0.1

0.0

0.1

0.2

Fig. 7.24. A bi-spiraling curve of homoclinic connections to the origin around the
T-point TP, shown in the (, )-parameter plane of (7.2) (a) and in the rescaled
( , )-plane where the bi-spiraling eect is clearly visible (b). Panels (c) and (d)
are two heteroclinic cycles that co-exist at the T-point TP; both cycles have the same
connection (solid lines) between the one-dimensional manifolds but dierent intersections between the two-dimensional manifolds (dashed lines). From F. Fern
andezS
anchez, E. Freire, L. Pizarro and A.J. Rodrguez-Luis, A model for the analysis of
the dynamical consequences of a nontransversal intersections of the two-dimensional
c 2003 by Elsemanifolds involved in a T-point, Phys. Lett. A 320 (2003) 169179 
vier Science; reprinted with permission.

7.3 Conclusions and Outlook

A local analysis of degeneracies of equilibria provides interesting information about a great variety of behaviors, including equilibria, periodic orbits,
global connections, and their bifurcations. The validity of such an analysis is
restricted, in principle, to a neighborhood in the product of phase and parameter space of the degenerate equilibrium in question. The crux of the matter
is that this neighborhood is generally unknown a priori, but it may actually
be of relevant size.
In this survey we have shown with practical examples of electronic circuits
how continuation methods can be used to extend local results and to nd
new phenomena. In an interesting feedback process, new bifurcation phenom-

248

Emilio Freire and Alejandro J Rodrguez-Luis

ena lead to further analytical and numerical study. By contrast, brute-force

simulation usually provides poor results at large computational expense.
Analogue electronic circuits are of obvious importance in many applications. What is more, they are particularly useful devices from the dynamical
systems point of view, because they allow one, in certain situations, to reproduce the behavior of a given dynamical system [21, 42, 44]. Indeed, many
interesting dynamical phenomena may be studied experimentally by assembling inexpensive electronic components that are readily available.
The study of electronic circuits presented here also provides evidence that
new theoretical and numerical tools need to be developed for a better understanding of possible dynamical behavior. Some of the numerical challenges
are:
1. the detection and continuation of bifurcations of periodic orbits of higher
codimension (in the way it is done, for example, in Content and Matcont);
2. the continuation of bifurcations in piecewise linear systems;
3. the continuation of quasiperiodic orbits; and
4. the inclusion of some continuation procedures into the circuit simulator
SPICE [45].

Acknowledgements
We would like to thank A. Algaba, F. Fern
andez-S
anchez, E. Gamero and
M. Merino for helpful suggestions, and the editors of this volume for their
eort and comments, which have been very much appreciated. The material
presented here is the result of a long-term collaborative eort. Apart from
the colleagues above, we thank A.R. Champneys, E.J. Doedel, L. Pizarro and
E. Ponce for their contributions. We especially thank Sebius Doedel for the
friendship he showed during his visits to our research group in Seville. This
work has been partially supported by the Ministerio de Ciencia y Tecnologa,
Plan Nacional I+D+I, in the frame of the projects BFM20012608, BFM2003
00336 and MTM2004-04066, and by the Consejera de Educaci
on y Ciencia
de la Junta de Andaluca (TIC-0130).

References
1. A. Algaba, F. Fern
andez-S
anchez, E. Freire, E. Gamero, and A. J. RodrguezLuis. Oscillation-sliding in a modied van der Pol-Dung electronic oscillator.
J. Sound Vibr., 249:899907, 2002.
2. A. Algaba, F. Fern
andez-S
anchez, E. Freire, M. Merino, and A. J. RodrguezLuis. Nontransversal curves of T-points: a source of closed curves of global
bifurcations. Phys. Lett. A, 303:204211, 2002.

7 Bifurcation Analysis of Electronic Circuits

249

3. A. Algaba, E. Freire, and E. Gamero. Hypernormal form for the Hopf-zero

bifurcation. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 8:18571887, 1998.
4. A. Algaba, E. Freire, E. Gamero, and A. J. Rodrguez-Luis. Analysis of Hopf
and Takens-Bogdanov bifurcations in a modied van der Pol-Dung oscillator.
Nonlin. Dynam., 16:369404, 1998.
5. A. Algaba, E. Freire, E. Gamero, and A. J. Rodrguez-Luis. On a codimensionthree unfolding of the interaction of degenerate Hopf and pitchfork bifurcations.
Internat. J. Bifur. Chaos Appl. Sci. Engrg., 9:13331362, 1999.
6. A. Algaba, E. Freire, E. Gamero, and A. J. Rodrguez-Luis. A three-parameter
study of a degenerate case of the Hopf-pitchfork bifurcation. Nonlinearity,
12:11771206, 1999.
7. A. Algaba, E. Freire, E. Gamero, and A. J. Rodrguez-Luis. A tame degenerate
Hopf-pitchfork bifurcation in a modied van der Pol-Dung oscillator. Nonlin.
Dynam., 22:249269, 2000.
8. A. Algaba, M. Merino, and A. J. Rodrguez-Luis. TakensBogdanov bifurcations
of periodic orbits and Arnolds tongues in a three-dimensional electronic model.
Internat. J. Bifur. Chaos Appl. Sci. Engrg., 11:513531, 2001.
9. A. A. Andronov, E. A. Leontovich, I. I. Gordon, and A. G. Maier. Theory of
Bifurcations of Dynamic Systems on a Plane. (Israel Program for Scientic
Translations, Jerusalem, 1971).
10. A. A. Andronov and L. Pontriaguin. Syst`emes grossiers. Dokl. Akad. Nauk.
SSSR, 14:247251, 1937.
11. A. A. Andronov, A. A. Vitt, and S. E. Khaikin. Theory of Oscillators (transl.
by F. Immerzi, rst edition in Russian, Moscow, 1937). (Pergamon, London,
1966).
12. V. V. Bykov. The bifurcations of separatrix contours and chaos. Physica D,
62:290299, 1993.
13. V. V. Bykov. On systems with separatrix contour containing two saddle-foci.
J. Math. Sci., 95:25132522, 1999.
14. V. V. Bykov. Orbit structure in a neighborhood of a separatrix cycle containig
two saddle-foci. AMS Translations, Series 2, 200:8797, 2000.
15. J. L. Callot, F. Diener, and M. Diener. Le probl`eme de la chasse au canard.
C. R. Acad. Sci. Paris (Ser. I), 23:10591061, 1978.
16. A. R. Champneys and Yu. A. Kuznetsov. Numerical detection and continuation
of codimension-two homoclinic bifurcations. Internat. J. Bifur. Chaos Appl.
Sci. Engrg., 4:785822, 1994.
17. A. R. Champneys and A. J. Rodrguez-Luis. The non-transverse Shilnikov
Hopf bifurcation: uncoupling of homoclinic orbits and homoclinic tangencies.
Physica D, 128:130158, 1999.
18. F. Diener and M. Diener. Chasse au canard. I. Les canards. Collect. Math.,
32:3774, 1981.
19. E. J. Doedel, A. R. Champneys, T. F. Fairgrieve, Yu. A. Kuznetsov, B. Sandstede, and X. Wang. Auto97 Continuation and bifurcation software for ordinary dierential equations, 1997. Available by anonymous ftp from ftp.cs.
concordia.ca directory pub/doedel/auto.
20. E. J. Doedel, E. Freire, E. Gamero, and A. J. Rodrguez-Luis. An analytical and
numerical study of a modied van der Pol oscillator. J. Sound Vibr., 256:755
771, 2002.

250

Emilio Freire and Alejandro J Rodrguez-Luis

21. F. Fern
andez-S
anchez, E. Freire, L. Pizarro, and A. J. Rodrguez-Luis. Analytical and numerical study of a van der PolDung oscillator. In J. L. Huertas
and A. Rodrguez-V
azquez, editors, Fourth International Workshop on Nonlinear Dynamics of Electronic Systems, NDES96, Sevilla, pages 321326, 1996.
22. F. Fern
andez-S
anchez, E. Freire, L. Pizarro, and A. J. Rodrguez-Luis. A model
for the analysis of the dynamical consequences of a nontransversal intersections
of the two-dimensional manifolds involved in a Tpoint. Phys. Lett. A, 320:169
179, 2003.
23. F. Fern
andez-S
anchez, E. Freire, and A. J. Rodrguez-Luis. Isolas, cusps and
global bifurcations in an electronic oscillator. Dynam. Stab. Syst., 12:319336,
1997.
24. F. Fern
andez-S
anchez, E. Freire, and A. J. Rodrguez-Luis. T-points in a Z2 symmetric electronic oscillator. (I) Analysis. Nonlin. Dynam., 28:5369, 2002.
25. F. Fern
andez-S
anchez, E. Freire, and A. J. Rodrguez-Luis. Analysis of the
T-pointHopf bifurcation. Preprint, 2005.
26. E. Freire, L. G. Franquelo, and J. Aracil. Periodicity and chaos in an autonomous
electronic system. IEEE Trans. Circ. Syst., 31:237247, 1984.
27. E. Freire, E. Gamero, and A. J. Rodrguez-Luis. First-order approximation for
canard periodic orbits in a van der Pol electronic oscillator. Appl Math. Lett.,
12:7378, 1999.
28. E. Freire, A. J. Rodrguez-Luis, E. Gamero, and E. Ponce. A case study for
homoclinic chaos in an autonomous electronic oscillator. A trip from Takens
Bogdanov to HopfShilnikov. Physica D, 62:230253, 1993.
29. E. Gamero, E. Freire, A. J. Rodrguez-Luis, E. Ponce, and A. Algaba. Hypernormal form calculation for triplezero degeneracies. Bull. Belg. Math. Soc.,
6:357368, 1999.
30. P. Glendinning, and C. T. Sparrow. T-points: A codimension two heteroclinic
bifurcation. J. Stat. Phys., 43:479488, 1986.
31. M. G. M. Gomes and G. P. King. Bistable chaos II. Bifurcation analysis. Phys.
Rev. A, 46:31003110, 1992.
32. J. Guckenheimer and P. J. Holmes. Nonlinear Oscillations, Dynamical Systems,
and Bifurcations of Vector Fields, volume 42 of Applied Mathematical Sciences.
(Springer-Verlag, New York, 1986).
33. J. Guckenheimer and S. Kim. Dstool: A dynamical system toolkit with an
interactive graphical interface. Applied mathematics report, Center for Applied
Mathematics, Cornell University, Ithaca, New York, 1992.
34. J. J. Healey, D. S. Broomhead, K. A. Clie, R. Jones, and T. Mullin. The origins
of chaos in a modied van der Pol oscillator. Physica D, 48:322339, 1991.
35. P. Holmes. Ninety plus thirty years of nonlinear dynamics: less is more and
more is dierent. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 15:27032716,
2005.
36. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory, volume 112 of Applied Mathematical Sciences. (Springer-Verlag, New York, 2004).
37. R. Madan, editor. Chuas Circuit: a Paradigm for Chaos, volume 1 of World
Scientic Series on Nonlinear Science, Series B. (World Scientic Publishing,
Singapore, 1993).
38. T. Matsumoto, L. O. Chua, and M. Komuro. The double scroll. IEEE Trans.
Circ. Syst., 32:797818, 1985.

7 Bifurcation Analysis of Electronic Circuits

251

39. B. B. Peckman, C. E. Frouzakis, and I. Kevrekidis. Bananas and banana splits:

A parametric degeneracy in the Hopf bifurcation for maps. SIAM J. Math.
Anal., 26:190217, 1995.
40. L. Perko. Dierential Equations and Dynamical Systems, volume 7 of Texts in
Applied Mathematics. (Springer-Verlag, New York, 2001).
41. H. J. Poincare. Les Methodes Nouvelles de la Mecanique Celeste, volumes 13. (Gauthiers-Villars, Paris, 1892-1893-1899). (English translation edited by D.
Goro, published by the American Institute of Physics, New York, 1993.)
42. R. Rocha and R. F. Machado L. S. Martins-Filho. A methodology for the
teaching of dynamical systems using analogous electronic circuits. Internat. J.
Electr. Eng. Educ., 43:334345, 2006.
43. R. Shinriki, M. Yamamoto, and S. Mori. Multimode oscillations in a modied van der Pol oscillator containing a positive nonlinear conductance. IEEE
Proceedings, 69:394395, 1981.
44. M. Suneel. Electronic circuit realization of the logistic map. S
adhan
a, 31:6978,
2006.
45. P. Tuinenga. SPICE: A guide to circuit simulation and analysis using PSpice.
(Prentice-Hall, Englewood Clis, NJ, 1995).
46. Y. Ueda. The Road to Chaos, II. (Aerial Press Inc., Santa Cruz, CA, 2001).
47. B. van der Pol and B. van der Mark. Frequency demultiplication. Nature,
120:363364, 1927.
48. S. Wieczorek and B. Krauskopf. Bifurcations of n-homoclinic orbits in optically
injected lasers. Nonlinearity, 18:10951120, 2005.
49. S. Wiggins. Introduction to Applied Nonlinear Dynamical Systems and Chaos,
volume 2 of Texts in Applied Mathematics. (Springer-Verlag, New York, 2003).
50. A. K. Zvonkin and M. A. Shubin. Non-standard analysis and singular perturbations of ordinary dierential equations. Russian Math. Surveys, 39:69131,
1984.

8
Periodic Orbit Continuation in Multiple Time
Scale Systems
John Guckenheimer and M Drew LaMar
Mathematics Department, Cornell University, USA

Continuation methods utilizing boundary value solvers are an eective tool

for computing unstable periodic orbits of dynamical systems. Auto [7] is the
standard implementation of these procedures. Unfortunately, the collocation
methods used in Auto often require very ne meshes for convergence on
problems with multiple time scales. This inconvenience prompts the search
for alternative methods for computing such periodic orbits; we introduce here
new multiple-shooting algorithms based on geometric singular perturbation
theory.

8.1 Mathematical Setting

The systems that we study are slow-fast systems of the form

x = f (x, y),
y = g(x, y),

(8.1)

where f : Rm Rm and g : Rn Rn are at least C 1 and > 0 is a

small parameter determining the ratio of time scales between the fast variable
x Rm and the slow variable y Rn . The limit = 0 is a system of dierential
algebraic equations in which motion is constrained to the critical manifold C
dened by f = 0. Rescaling time to the slow time = t yields the system
 
x = f (x, y),
(8.2)
y  = g(x, y).
Here, the limit = 0 is the family of layer equations in y, also called the
fast subsystems of (8.1). We make two standing assumptions about (8.1) that
further constrain the context for our analysis:
1. The critical manifold C of (8.1) is indeed a manifold and its projection
onto the space of slow variables is generic in the sense of singularity
theory;

254

John Guckenheimer and M Drew LaMar

2. The limit sets of all trajectories for the layer equations are equilibria, i.e.,
points of C.
At regular points of the projection , the manifold C can be represented
locally as the graph of a function x = h(y). Substitution of this expression
into g yields the slow ow on the regular points of C. We shall denote the set
of singular points of by S.
Trajectories of (8.1) are typically approximated by candidates, concatenations of trajectories of the slow ow and the layer equations that form continuous curves. Periodic orbits that contain both segments close to trajectories
of the slow ow and segments close to trajectories of the layer equations are
called relaxation oscillations. Trajectory segments close to an unstable sheet
of the critical manifold are called canards. Numerical computation of canards
by forward solution of an initial value problem is not feasible when is suciently small due to the instability on the fast time scale [14]. Thus, even stable
periodic orbits containing canards cannot be computed by forward numerical
integration from initial points in the basin of attraction of these orbits. As
mentioned earlier, tracking such orbits with collocation methods also seems to
require very ne meshes. Our goal in this paper is to re-examine the computation of relaxation oscillations, including those with canards. We propose a
multiple-shooting approach, in which dierent segments of a periodic orbit are
computed dierently and then matched with suitably chosen cross-sections.
The next two sections lay out the general framework that we investigate.
Section 8.4 presents two numerical examples, comparing the methods introduced here with Auto computations of the same orbits. Finally, Sect. 8.5
comments on further extension and improvement of these methods.

8.2 Simple Relaxation Oscillations

We consider rst the simplest relaxation oscillations, namely, those with a fastslow decomposition that makes them readily amenable to analysis. We dene
a relaxation oscillation to be simple if it is approximated by a candidate
0 that satises the following conditions:
S1: 0 consists of slow segments i and fast segments i , i = 1 . . . k, in the
order 1 , 1 , . . . , k , k . The initial and terminal points of i are pi and
qi . The initial and terminal points of i are qi and p(i+1) mod k .
S2: Each i lies on a stable sheet of the critical manifold.
S3: The points qi are saddle-node bifurcations of the layer equations and none
of their eigenvalues have positive real parts. This assumption implies that
there are unique solutions of the layer equations emanating from the points
qi and all nearby fold points of C.
S4: The slow ow satises the normal crossing conditions [23] at qi .
S5: Denote by (S) the forward limits of trajectories of the layer equations
emanating from the fold points of S satisfying condition S3. We require
that (S) is transverse to the slow ow on C at the points pi .

8 Periodic Orbit Continuation in Multiple Time Scale Systems

255

Note that we have not required that a simple relaxation oscillation be stable
or even hyperbolic, although eigenvalues of a return map in the fast directions
are stable and, indeed, approach zero as 0.
We want to establish that well-conditioned multiple-shooting methods can
be formulated for the computation of simple relaxation oscillations. Our strategy is to create cross-sections i to each of the fast segments i of a simple
relaxation oscillation , compute the ow maps i from i to (i+1) mod k
and then solve the multiple-shooting equations
z(i+1) mod k = i (zi )
for points zi i .
Theorem 1. Let be a hyperbolic simple relaxation oscillation, and i and
i as described above. For 0 suciently small, the system of equations
z(i+1) mod k = i (zi ),

zi i ,

is regular and has an isolated solution with zi = i .

Proof. Fenichel theory [10] and a theorem of Levinson [20] imply that the ow
maps i from i to (i+1) mod k are smooth maps that converge to smooth
maps 0i of rank n 1 as 0. We remark that the convergence is continuous but not smooth in due to several phenomena; for example, asymptotic
expansions of trajectories near the folds involve fractional powers of [25].
The point 0i (zi ) is obtained by a three-step process:
(1) follow the trajectory of the layer equations with initial condition (zi ) to
its limit on the critical manifold;
(2) follow this limit point to its rst intersection with a fold of the critical
manifold; and
(3) follow the unstable separatrix of the layer equations from this fold point
to its intersection with (i+1) mod k .
Denote the image of 0(i1) mod k by Wi . The dimension of Wi is n 1, the
dimension of folds of the critical manifold. Condition S5 implies that the
restriction of 0i to Wi has full rank n 1. The restriction of the equations
z(i+1) mod k = i (zi ) to Wi has the same structure as the set of equations for a
multiple-shooting method based upon cross-sections to the ow. The Jacobian
of this system has the block-cyclic structure

D01 |W1
I

D02 |W2 I

..
..

.
.
.

D0k1 |Wk1
I
I
D0k |Wk

256

John Guckenheimer and M Drew LaMar

As shown by Guckenheimer and Meloon [15], this matrix has maximal rank
if and only if 1 is not an eigenvalue of diag(D0k |Wk , . . . , D01 |W1 ), that is,
0 is a hyperbolic xed point of the composition 0k |Wk 01 |W1 . On
a complementary set of coordinates to the Wi the equations z(i+1) mod k =
i (zi ) reduce to z(i+1) mod k = 0 because i (zi ) vanishes in these directions
by denition. Thus, the full system of equations on the product of the i is
regular if and only if 0 is hyperbolic. The equations change continuously
in the C 1 topology as 0 [16], so hyperbolicity of 0 implies that the
equations are regular and that is hyperbolic for > 0 suciently small.
This proves the theorem. 

For a hyperbolic simple relaxation oscillation a multiple-shooting algorithm based upon the cross-sections described above yields a regular system
of equations. In many cases, these equations will be well conditioned uniformly
for small . If they are not, additional cross-sections can be inserted. The effectiveness of the multiple shooting algorithm will be largely determined by
the numerical integration method used to compute the i .

8.3 Degenerate Slow-Fast Decompositions

The multiple-shooting algorithm described above for locating simple relaxation oscillations can be implemented within a standard continuation framework. The procedure may break down when a family of periodic orbits encounters degenerate slow-fast decompositions resulting from the failure of one
of the requirements for the orbit to be simple. Here, we examine modications
of the multiple-shooting algorithm that cope with the instability of canards
in the context of two specic examples of degenerate decompositions [11].
8.3.1 Hopf Bifurcation and Canard Explosions
The lowest-dimensional example of a degenerate slow-fast decomposition occurs at Hopf bifurcations of a systems with one slow and one fast direction
(n = m = 1). The canard explosion of the resulting orbits has been studied
extensively, especially in the system

x = y 13 x3 + x,
(8.3)
y = a x

near a = 1; see, for example, [2, 8, 9, 13]. (Note that in several studies coordinates have been used that place the point (1, 2/3) of (8.3) at the origin [8].)
It has been proven that the periodic orbits of this system grow monotonically
as a decreases from 1, exploding in size from O() to O(1) over a range of a
that is O(exp(c/)) for a suitable c > 0. The trajectories in the middle of this
explosion contain canards that follow the unstable branch of the critical manifold given by y = 13 x3 x for an O(1) distance before jumping right or left to a

8 Periodic Orbit Continuation in Multiple Time Scale Systems

257

stable branch of the critical manifold. Trajectories along the canard segments
of these trajectories diverge from one another at a rate exp(t(x2 1)/).
For small values of , this divergence eectively prevents accurate computation of a trajectory for times that are larger than O(). However, backward
integration along these canards is highly stable.
To compute the periodic orbits with canards in this family, we adopt a
shooting strategy that shoots forward and backward from one cross-section of
the ow to another cross-section. The cross-sections are chosen so that forward
trajectories do not contain canard segments and backward trajectories do not
contain segments that track the stable part of the slow manifold. The initial
cross-section depends upon where we are in the family, in particular on the
direction of the jump away from the canard segment in a periodic orbit. Over
part of the family, the jump is to the right and there is a single stable slow
segment in the periodic orbit. Over another part of the family, the jump is to
the left and there are two stable slow segments and two fast segments in the
slow-fast decomposition of the trajectory. The behavior that occurs between
these two possibilities is that there is a maximal canard that extends over
the entire length of the unstable branch of the critical manifold. When the
jumps from the canards are to the right, we choose the initial cross-section
to be the line {x = a} where the vector eld is horizontal. For jumps to
the left, we choose the initial cross-section {x = 1}, which is crossed by all
trajectories that ow left from the unstable branch of the critical manifold
to the stable branch of the critical manifold. In both cases, we take the nal
target cross-section to be {x = a}.

+
The shooting problem that we seek to solve is +
a (y) = a (y) where a (y)
is the ow map from the initial to the nal cross-section in the forward time
direction and
a (y) is the ow map from the initial to the nal cross-section
in the backward time direction. There are three remarks that we make about
this problem:

1. The derivatives of both +

a (y) and a (y) are small for members of the
+
canard family, but the derivative of a (y) is much smaller [8, 12], so the
periodic orbits are stable;

2. The derivative of +
a (y) a (y) with respect to a is O(1), so the small
+

derivative of a (y)a (y) with respect to y yields an extreme sensitivity

of the solution y to the shooting problem as a function of the parameter
a. Therefore, in the middle of the canard family, we x the initial point on
the rst cross-section and vary a to locate a periodic orbit passing through
the initial point instead of xing a and trying to locate the solution y of
the shooting equation. Alternatively, we regard the shooting equation as
a continuation problem in the variables (y, a) and use pseudo-arclength or
other continuation strategies to nd the curve of solutions to the equation;
3. If we move beyond the range of parameter values for which there are
canards, we are likely to nd trajectories that do not reach the target
cross-section. This happens when there is a stable equilibrium point for

258

John Guckenheimer and M Drew LaMar

parameters a > 1. When a is large enough so that the vector eld has a
stable node, trajectories that follow the right-hand branch of the critical
manifold accumulate at the stable equilibrium point without reaching the
line {x = a}. Implementations of the shooting algorithms need to test for
this possibility and take appropriate action if the shooting equation is not
dened.
8.3.2 Folded Saddles
The forced Van der Pol system

3
x = y + x x3 ,
y = x + a sin(2),

=

(8.4)

is a slow-fast system with two slow variables and one fast variable (n = 2,
m = 1). Cartwright and Littlewood studied this system in their seminal work
on chaos in dynamical systems [4, 21, 22]. Recently, Haiduc [17] has extended
the classical results of Cartwright and Littlewood by using geometric singular
perturbation theory; these methods have also been used to investigate the
dynamics of this system numerically [3, 14]. Throughout these studies, folded
saddles play a prominent role in the analysis. Folded saddles are points where
the rescaled slow ow equations
 
= (x2 1),
(8.5)
x = x + a sin(2)
have a saddle point. In the original system they are points on the fold curves
where the normal crossing conditions fail. At the folded saddles the slow ow
changes direction from pointing toward the fold to pointing away from the
fold. Canards emanate from the folded saddles along stable manifolds of the
saddles of (8.5) (lifted back to the unstable sheet of the critical manifold of
(8.4)).
Bold et al. [3] used Auto to track families of periodic orbits in the forced
van der Pol system that contain canards emanating from folded saddles. These
computations required ne meshes when applied to the system with = 103
and even more so with = 104 . We develop here modications similar to
the ones described above for Hopf bifurcations to compute these trajectories
with multiple-shooting methods that use a small number of cross-sections. We
place cross-sections to the ow at the beginning of canard segments and in
the middle of jumps that leave the canard segments. Backward integration
between these cross-sections is stable since there is a single fast variable and
the unstable sheet of the critical manifold is stable for the reversed time ow.
At folded saddles of a system with two slow and one fast variables, the ow
is parallel to the fold curve. For the forced Van der Pol system this suggests
that we choose the cross-section dened by = f s , where f s is the value of

8 Periodic Orbit Continuation in Multiple Time Scale Systems

259

at the folded saddle. All trajectories that jump from the unstable sheet of
the critical manifold to a stable sheet intersect one of the planes {x = 1},
so we choose these planes as the cross-sections for the trajectories that jump
from canards. Periodic orbits of dierent periods yield dierent sequences of
intersections with the cross-sections and dierent dening equations. Unlike
the canard explosions discussed in Sect. 8.3.1, many of the periodic orbits are
embedded in chaotic invariant sets and are unstable.

8.4 Numerical Results

We computed families of periodic orbits containing canards in the Van der Pol
and forced Van der Pol systems with Auto [7] and with our shooting methods.
Our calculations follow the approach in [13] to address the complexity of
periodic orbits in systems with multiple time scales. Thus, we use as many as
1000 mesh points in Auto and error tolerances for both Auto and shooting
on the order of 1010 .
For the multiple-shooting algorithms, we used PyCont, a continuation
package developed for PyDSTool [5]. The shooting algorithms were embedded in a Moore-Penrose continuation framework [1], with the dierential
equations numerically integrated using Radau [18], an implicit Runge-Kutta
method. We set the absolute error tolerances in the calculations with Radau to
1012 and the relative error tolerance to 109 . These tolerances produced trajectories with sucient precision so that the truncation errors did not appear
to be signicant for the Newton iterations within the shooting method.
In the next two sections we show results for each of the two examples
mentioned above. After presenting the results, we do some comparisons between Auto and multiple-shooting with the forced Van der Pol system. Since
algorithmic performance is highly dependent on implementation details, e.g.,
programming language used, we assess algorithmic complexity based on four
properties of Newtons method: the size of Jacobians, condition numbers of
Jacobians, the number of Newton iterations, and domains of convergence.
8.4.1 Hopf Bifurcation and Canard Explosions
Figure 8.1 displays the results of our computations with multiple shooting
of the canard explosion in the Van der Pol oscillator (8.3). Panel (a) shows
the bifurcation diagram and panel (b) is a plot of representative limit cycles.
Three regions are indicated on the bifurcation diagram in Fig. 8.1(a), where
open circles denote the boundaries between them. Region A corresponds to
simple relaxation oscillations, where we use the shooting methods described
in Sect. 8.2. Regions B and C correspond to jump-left and jump-right canards, respectively; in these regions we use the shooting methods described
in Sect. 8.3.1. The open circle that separates regions B and C corresponds to
a maximal canard. Finally, the solid black circle denotes the Hopf point.

260

John Guckenheimer and M Drew LaMar

(a)

(b)

Fig. 8.1. The bifurcation diagram of the Van der Pol oscillator (8.3) near the
canard explosion (a) and the corresponding limit cycles with canard segments (b)
as computed with our shooting method.

In order to start the continuation of canards in the Van der Pol system
with the multiple-shooting algorithm, we proceed as follows. We rst nd
a simple stable relaxation oscillation numerically for the parameter values
(a, ) = (0.5, 1.0) using Radau integration. We then continue this orbit in
Auto with as the free parameter. The continuation terminates at the periodic orbit with = 104 . Starting from this periodic orbit, we subsequently
continue the family of orbits in Auto with a as the free parameter. This
continuation terminates at a limit cycle in the middle of region B. This limit
cycle is our initial limit cycle for continuation with our shooting methods. We
perform continuation in the forward and backward direction with a as the free
parameter. When moving in the backward direction, we pass from region B to
region A, where canards cease and simple relaxation oscillations exist. At the
transition from B to A, backward integration fails, indicating that we should
change shooting methods. During the continuation in the forward direction
we pass from region B to region C by encountering a maximal canard. In this
situation backward integration also fails, and we must switch cross-sections
from {x = 1} to {x = 1}. Note that the tangent vector to the bifurcation
curve switches: in region B both a and y are increasing, while in region C a
is increasing and y is decreasing. Automated detection of periodic orbits that
separate these regions (open circles) is an important topic for future work.

8 Periodic Orbit Continuation in Multiple Time Scale Systems

261

Fig. 8.2. Bifurcation diagram of period-three orbits in the forced Van der Pol system
(8.4) computed with shooting. In this example, = 104 and = 1.55.

8.4.2 Folded Saddles

For the continuation of periodic orbits with canards initiated at folded saddles
in the forced Van der Pol system, we start with a period-three simple stable
relaxation oscillation for the parameter values (, , a) = (104 , 1.55, 2.5). A
bifurcation diagram, calculated with shooting and a as the free parameter,
is shown in Fig. 8.2. During the computation of the periodic orbits we take
advantage of the symmetry (, x, y)  ( + 1.5, x, y), which means that
we need only integrate over half the period. We apply the symmetry transformation to the endpoint at the section { = f s + 1.5}, and compare this
point to the endpoint at { = f s }, giving our matching condition. Our initial
cross-sections are {x = 1}, and thus the matching conditions for shooting
occur in a hyperplane with coordinates (y, , a). The sign of x on the initial
cross-section depends on the location of the periodic orbit in the bifurcation diagram. There are four regions. Region A corresponds to no canards, where we
start from the cross-section {x = 1} and integrate forward to the cross-section
{x = 1+ } (the superscript denotes crossing in the increasing x-direction).
Regions B and D correspond to jump-right canards (right relative to the fast
variable x), where we start from the cross-section {x = 1}, shooting forward to
the cross-section { = f s + 1.5} and backward to the cross-section { = f s }.
Finally, region C corresponds to jump-left canards, where we start from the
cross-section {x = 1}, shooting forward to the cross-section { = f s + 1.5}
and backward to the cross-section { = f s }. Transitions between regions are
again denoted by open circles.

262

John Guckenheimer and M Drew LaMar

8.4.3 Comparisons
Let us now discuss the convergence of the Newton iterations for each method.
In Auto, a collocation boundary value method is used to nd periodic orbits.
With this method, the Jacobian is of dimension mnN +b+q+1, where N is the
number of subintervals, m the number of collocation points per subinterval, n
the dimension of the vector eld, b the number of boundary conditions and q
the number of integral conditions. We have already mentioned that, due to the
slow-fast structure of the systems, the number of subintervals must be large
for accuracy and convergence (we use N = 1000). The number of collocation
points used per subinterval is set at m = 4 in our calculations. Thus, we expect
the size of the Jacobian to be approximately 8, 000 8, 000 in the Van der Pol
system and approximately 12, 000 12, 000 in the forced Van der Pol system.
In the multiple-shooting methods, the Jacobian is of dimension (n 1)s + 1,
where again n denotes the dimension of the vector eld and s is the number
of cross-sections that are used. In the forced Van der Pol system the Jacobian
is 3 3. Note that we will require more cross-sections when tracking periodic
orbits with multiple canard segments.
Computationally, most of the eort in the shooting methods is in the
integration, while in collocation most of the computational eort is in solving a
large, sparse matrix. The construction of the Jacobian in the multiple-shooting
method involves numerical integration of the vector eld, which is an ecient
and speedy process. The Jacobian for collocation is much larger because it
is also used to determine a suitable orbit segment (while in shooting this
is done with an integrator whose accuracy must be controlled separately). It
should be noted that, although the Jacobian is considerably larger with Auto,
ecient numerical techniques are used to invert the Jacobian in two stages
by taking advantage of the sparsity structure of the matrix. The rst stage
uses a method known as condensation of parameters to perform independent
eliminations in N blocks of size nm n(m + 1). The second stage produces a
reduced Jacobian of size (n + b + q + 1) (n + b + q + 1); see also Chap. 1. This
inversion still requires much more computation than Gaussian elimination on
the matrix of size ((n 1)s + 1) ((n 1)s + 1) used in our shooting method.
For both Auto and shooting, we performed a full Newtons method with a
maximum of eight iterations and error tolerances on the order of 1010 . During
the calculations, we kept track of the number of iterations in the convergence
of each step of Newtons method, as well as the condition number of the
Jacobian. In the forced Van der Pol system the Jacobians for the shooting
methods had O(10) condition numbers, while for Auto the condition numbers
for the reduced Jacobians were O(106 ).
Domains of convergence for the family F of limit cycles along the
branch B of Fig. 8.2 are displayed in Fig. 8.3; panel (a) shows the domain of
convergence for the multiple-shooting method and panel (b) that for an Auto
computation (with N = 1000 mesh intervals). The thick (approximately) horizontal black line through the origin represents the a-dependent family of limit

8 Periodic Orbit Continuation in Multiple Time Scale Systems

263

(a)

c1

c0
(b)

c1

c0
Fig. 8.3. Domains of convergence for the computation of limit cycles along branch
B of Fig. 8.2 with the shooting method (a) and with Auto for N = 1000 mesh
intervals.

cycles . The gures are shaded according to the number of iterates needed
for convergence, where shades white to black represent convergence after one
to eight iterations, respectively. In fact, black denotes no convergence of the
method using our error tolerances and choice of maximally eight iterates. Figure 8.2(a) contains an additional darkest shade of gray that is used for those
points that do not converge in eight iterations but show signs of converging.
Specically, such points are marked as converging if the function values and
dierences between variable values for the last three iterates are decreasing.
The plots were obtained by starting with a specic limit cycle 0 and
its intersection 0 with the cross-section dened by x = 1. The section
is three dimensional with coordinates (y, , a). We compute orthonormal
vectors (v 0 and v 1 ) in at 0 so that v 0 is tangent to F and v 1 lies
in the plane spanned by v 0 and (0, 0, 1). The coordinates c = (c0 , c1 ) in the
gure correspond to the points pc = 0 + c0 v 0 + c1 v 1 and, thus, the
origin represents 0 . The horizontal black line in the gure is the projection

264

John Guckenheimer and M Drew LaMar

(a)

c1

c0
(b)

c1

c0
(c)

c1

c0
Fig. 8.4. The domains of convergence for the computation of limit cycles along
branch B of Fig. 8.2 with Auto depend on the number of mesh intervals; panel (a)
to (c) are for N = 1000, N = 500 and N = 250 mesh intervals, respectively.

of F onto the plane spanned by (v 0 , v 1 ). Note that F appears

horizontal because v 0 is tangent to F and the domain represented in the
gure is small.

8 Periodic Orbit Continuation in Multiple Time Scale Systems

265

For the shooting algorithm, we use the point pc for each grid point
c = (c0 , c1 ) in the gure as the initial point for the algorithm and shade
the pixel according to the number of iterates required for convergence as described above. Unfortunately, initialization of Auto requires an entire curve.
To obtain a curve from each of the points pc , we computed trajectories forward and backward as in the shooting algorithm and used the concatenation of
these two trajectory segments to initialize Auto. (Auto computes a mesh of
specied size from the trajectories provided.) Except on F , these curves are
not closed: the nal points of the forward and backward trajectory segments
do not match. Since Auto is based upon solving the dierential equations
within the space of closed curves, it might be preferable to initialize Auto
with closed curves. Lacking a natural way to produce closed curves at dierent
distances from F , we did not pursue such a comparison here.
We also tested the dependence of Autos domain of convergence on the
number of mesh intervals N . Figure 8.4 displays the results of three computations. Panel (a) shows an enlargement of Fig. 8.3(b) (with N = 1000). Panels
(b) and (c) are for computations with N = 500 and 250, respectively. The
domain of convergence for Auto decreases with N , and values of N smaller
than 1000 have very small domains of convergence.

8.5 Towards a General Theory

The examples presented above demonstrate that multiple-shooting algorithms
adapted to the slow-fast decomposition of trajectories in multiple time scale
dynamical systems can be eective for computing periodic orbits. These algorithms are able to exploit the advantages of numerical integration methods
for sti systems to compute canards in vector elds with a single fast variable,
where we use integration backward in time. The strategy presented here relies
upon two ingredients. First, one must choose cross-sections for shooting that
isolate the trajectory segments to be computed in forward time and those to
be computed in backward time and, second, the computation must be placed
in a continuation setting in which the periodic orbits vary at a moderate rate
with respect to the continuation parameter. The rapid change of periodic orbits containing canards with respect to system parameters requires that the
root nding procedure performed by the continuation algorithm is able to
vary a system parameter as well as phase space variables.
Development of algorithms that make suitable choices of cross-sections
and continuation strategy is likely to require good methods for automatically
computing aspects of the slow-fast decomposition of trajectories in order to
base locating suitable cross-sections upon this information. In particular, such
methods should determine where degenerate decompositions are encountered
and use information about the types of canards that are associated with these
decompositions.

266

John Guckenheimer and M Drew LaMar

Canards in slow-fast systems with more than one fast variable typically lie
along sheets of the critical manifold that consist of saddle points for the layer
equations. Accurate computation of these canards cannot be done with either forward or backward numerical integration. Instead, two-point boundary
value solvers, methods that are designed for computing normally hyperbolic
manifolds (see Chap. 4) or methods for shadowing of trajectories of vector
elds [6, 24] will need to be incorporated into shooting methods when one
wants to compute relaxation oscillations that contain these canards. These algorithms will require more computation than numerical integration, but they
still are likely to provide a good alternative to collocation methods for these
problems.

Acknowledgments
This research was partially supported by grants from the National Institutes
of Health, the Department of Energy and the National Science Foundation.

References
1. E. Allgower and K. Georg. Continuation and path following. Acta Numerica
1993, pages 164 (Cambridge University Press, Cambridge, 1993)
2. E. Benoit, J. L. Callot, F. Diener, and M. Diener. Chasse au canard. Collect.
Math. 32:37119, 1981.
3. K. Bold, C. Edwards, J. Guckenheimer, S. Guharay, K. Homan, J. Hubbard,
R. Oliva, and W. Weckesser. The forced van der Pol equation II: Canards in the
reduced system. SIAM J. Appl. Dyn. Syst. 2:570608, 2003.
4. M. Cartwright and J. Littlewood On nonlinear dierential equations of the second
order: II the equation y kf (y, y)
y +g(y, k) = p(t) = p1 (t)+kp2 (t), k > 0, f (y)
1. Ann. Math. 48:47294, 1947. [Addendum 50:504505, 1949]
5. R. Clewley, M. D. LaMar, and E. Sherwood. PyDSTool available via http:
//sourceforge.net/projects/pydstool
6. B. Coomes, H. Kocak, and K. Palmer. Rigorous computational shadowing of
orbits of ordinary dierential equations. Numer. Math. 69:401421 1995.
7. E. J. Doedel, R. C. Paenroth, A. R. Champneys, T. F. Fairgrieve,
Yu. A. Kuznetsov, B. E. Oldeman, B. Sandstede, and X. J. Wang. Auto2000:
Continuation and bifurcation software for ordinary dierential equations. available via http://cmvl.cs.concordia.ca/.
8. F. Dumortier and R. Roussarie. Canard cycles and center manifolds. With an
appendix by Cheng Zhi Li. Mem. Amer. Math. Soc. 121, no. 577, x+100 pp,
1996.
9. W. Eckhaus. Relaxation oscillations, including a standard chase on French ducks.
Lecture Notes in Mathematics, Vol. 985, pages 449494 (Springer-Verlag, 1983).
10. N. Fenichel, Geometric singular perturbation theory. J. Di Eq. 31:5398, 1979

8 Periodic Orbit Continuation in Multiple Time Scale Systems

267

11. J. Guckenheimer. Bifurcation and degenerate decomposition in multiple time

scale dynamical systems. In J. Hogan, A. Champneys, B. Krauskopf, M. di
Bernardo, E. Wilson, H. Osinga, and M. Homer, editors, Nonlinear Dynamics and
Chaos: where do we go from here?, pages 121 (Institute of Physics Publishing,
Bristol, 2002).
12. J. Guckenheimer. Bifurcations of relaxation oscillations. In Normal Forms,
Bifurcations and Finiteness Problems in Dierential Equations, pages 295316,
NATO Sci. Ser. II Math. Phys. Chem., Vol. 137 (Kluwer Acad. Publ., Dordrecht,
2004).
13. J. Guckenheimer, K. Homan, and W. Weckesser. Numerical computation of
canards. Internat. J. Bifurc. Chaos Appl. Sci. Engrg., 10:26692687, 2000.
14. J. Guckenheimer, K. Homan, and W. Weckesser. The Forced van der Pol
Equation I: The Slow Flow and its Bifurcations. SIAM J. App. Dyn. Sys. 2:135,
2003.
15. J. Guckenheimer and B. Meloon. Computing periodic orbits and their bifurcations with automatic dierentiation. SIAM J. Sci. Comp., 22:951985, 2000.
16. J. Guckenheimer, M. Wechselberger, and L.-S. Young. Chaotic attractors of
relaxation oscillators. Nonlinearity, 19:701720, 2006.
17. R. Haiduc. Horseshoes in the forced van der Pol equation, PhD dissertation,
Cornell University, 2005.
18. E. Hairer, S. P. Norsett and G. Wanner. Solving Ordinary Dierential Equations
I, 2nd. ed. (Springer-Verlag, 1993).
19. C. K. R. T. Jones. Geometric singular perturbation theory. In Dynamical
Systems, Lecture Notes in Mathathematics, Vol 1609, pages 44120 (SpringerVerlag, 1995).
20. N. Levinson. Perturbations of discontinuous solutions of non-linear systems of
dierential equations. Acta Math., 82:71106, 1950.
21. J. Littlewood. On nonlinear dierential equations of the second order. III. The
equation y k(1 y 2 )y + y = bk cos(t + ) for large k, and its generalizations.
Acta Math., 97:267308, 1957. [Errata at Acta Math., 98:110, 1957]
22. J. Littlewood. On nonlinear dierential equations of the second order. IV. The
general equation y kf (y)y + g(y) = bkp(), = t + . Acta Math., 98:1110,
1957.
23. E. F. Mishchenko, Yu. S. Kolesov, A. Yu. Kolesov and N. Kh. Rhozov. Asymptotic Methods in Singularly Perturbed Systems. Monographs in Contemporary
Mathematics, (Consultants Bureau, New York, A Division of Plenum Publishing
Corporation, 1994).
24. E. Van Vleck. Numerical shadowing near hyperbolic trajectories. SIAM J. Sci.
Comput. 16:11771189, 1995.
25. A. B. Vasileva. Asymptotic behaviour of solutions of certain problems for ordinary non-linear dierential equations with a small parameter multiplying the
highest derivatives. Russian Mathematical Surveys, 18:1384, 1963. (Russian)
Uspehi Mat. Nauk, 18:1586, 1963.

9
Continuation of Periodic Orbits in Symmetric
Hamiltonian Systems
Jorge Gal
an-Vioque1 and Andre Vanderbauwhede2
1
2

Departamento de Matem
atica Aplicada II, Escuela Superior de Ingenieros de
Sevilla, Spain
Department of Pure Mathematics and Computer Algebra, Ghent University
Belgium

The idea of everything returning eventually to its point of departure has a

strong hold on humanity, with many historical, philosophical and religious
implications. Classical examples are the need to construct a calendar and the
subsequent search for orbits in the solar system in which the planets follow a
closed track and repeat their history over and over again.
Nature, at its most basic level, has decided to be Hamiltonian; nonHamiltonian systems come up in physics only as phenomenological models for
the more complicated underlying processes. However, Hamiltonian systems are
nongeneric dynamical systems with remarkable properties, in particular with
respect to periodic orbits. The role of periodic solutions in Hamiltonian systems and their importance in modern physics was rst recognized by Poincare
[26]. Today periodic orbits are at the basis of both classical and quantum mechanics [13]. Poincare conjectured that periodic orbits, that is, solutions that
return to their initial conditions after some nite time, are densely distributed
among all possible bounded classical trajectories; and he suggested that the
study of periodic orbits would provide the clue to the overall behavior of any
mechanical system. Quoting the original work [26]:
It seems at rst that the existence of periodic solutions could not
be of any practical interest whatsoever. Indeed, the probability is zero
for the initial condition to correspond precisely to those of a periodic
solution. But it may happen that they dier by very little. [...] Here
is a fact which I have not been able to demonstrate rigorously, but
which nevertheless seems very plausible to me. Given equations of the
Hamiltonian form and any particular solution of these equations, we
can always nd a periodic solution (whose period may admittedly be
very long) such that the dierence between the two solutions is as small
as we wish during as long a time as we wish. Besides this, what renders
these periodic solutions so precious is that they are, so to speak, the

270

Jorge Gal
an-Vioque and Andre Vanderbauwhede

only opening through which we may try to penetrate into the fortress
which has the reputation of being impregnable.
We know from Arnold [2] that a literal interpretation of Poincares statement is not possible: the completely integrable two-degrees-of-freedom Hamiltonian system corresponding to the Hamiltonian

H(I1 , I2 , 1 , 2 ) = I1 + 2 I2

(in action-angle variables) admits no periodic solutions whatsoever

since the

ratio of its frequencies is irrational. However, replacing 2 by an arbitrarily

close rational number results in a system that has only periodic orbits. This
insight leads to a reformulation that says that Poincares statement is true for
generic Hamiltonian systems; a more precise statement is given by the Closing
Lemma due to Pugh and Robinson [27], which states that, given a Hamiltonian
and a point in phase space, there exists an arbitrarily close Hamiltonian (in
the C 2 -sense) for which the given point generates a periodic orbit.
In this chapter we show how two-point boundary value problem continuation can help us to enter the fortress Poincare was talking about. Specically,
this approach provides us with an ecient tool to compute families of periodic orbits in Hamiltonian systems and allows us to discover numerically how
these families bifurcate and connect. Particular attention will be paid to symmetric Hamiltonian systems, where the symmetries and the associated rst
integrals typically increase the dimensions of these families. Continuation in
this context is introduced with the simple model example of the mathematical
pendulum, justied theoretically, and then applied to a nontrivial case of computing periodic orbits that are associated with the recently discovered gure-8
solution of the three-body problem. Numerical aspects of the scheme, including comparison with a shooting method, are highlighted with the computation
of a periodic solution of the restricted three-body problem.
The chapter is organized as follows. In Sect. 9.1 we discuss several aspects of periodic orbits in Hamiltonian systems, in particular, how they are
organized in families and how one can approach the numerical calculation
of these families. The discussion is illustrated with the mathematical pendulum and the restricted three-body problem. In Sect. 9.2 we give an outline
of theoretical continuation results for periodic orbits and relative equilibria;
these results include a set-up for the equations that can be used for numerical
computations. In Sect. 9.3 we present continuation results for the three-body
problem, where we start from the gure-8 and the Lagrange solutions, respectively. These periodic solutions exist for equal masses, and we use one of the
masses as (an external) continuation parameter. It appears that the starting
solutions can be continuously connected to a periodic solution of the restricted
three-body problem. In Sect. 9.4 we draw some general conclusions and give
a brief discussion of open problems.

9 Continuation in Hamiltonian Systems

271

9.1 Periodic Solutions in Hamiltonian Systems

It is well known that Hamiltonian (or, more generally, conservative) systems
are quite dierent from dissipative systems in terms of periodic orbits, their
continuation and their bifurcations. In dissipative systems periodic orbits are
generically isolated and, therefore, an external parameter is required in order
to continue such periodic orbits. For Hamiltonian systems there is the celebrated Cylinder Theorem [15], which says that periodic orbits appear in oneor more-parameter families and that, under appropriate nondegeneracy conditions, these families are persistent under small Hamiltonian perturbations.
Computationally the problem of nding a periodic orbit can be formulated
as a boundary value problem with the period as an additional parameter. In
order to avoid phase shifts and to ensure uniqueness one has to introduce
an appropriate phase condition, which can be either a boundary condition
(Poincare-type condition) or an integral version of it (see also Chaps. 1, 10
and 11). In dissipative systems this problem is generically well determined:
the periodicity condition together with the phase condition give n + 1 equations for the n components of the initial point, the period and the external
parameter. Generically these equations can be solved by the Implicit Function
Theorem, giving a one-dimensional solution curve that can be parametrized
by the external parameter.
This scheme does not work for Hamiltonian systems or, more generally,
for systems with a rst integral. (Another exceptional case are time-reversible
systems, but we do not consider them here since they require dierent arguments.) In conservative systems periodic orbits typically belong to oneparameter families, parametrized by the value of the rst integral (the energy
in the Hamiltonian case). This internal or natural parameter is not explicitly available in the equations, at least not directly, and this is why the
standard continuation scheme fails. Additional complications arise for Hamiltonian systems with several independent constants of motion (symmetries):
here periodic orbits belong to families whose dimension is the number of independent integrals. In this case, further phase conditions are required in
order to identify a single member of such a family uniquely.
9.1.1 Continuation of Periodic Orbits in Conservative Systems
A straightforward approach to continue periodic orbits in conservative systems is to use the conserved quantity (the energy, in the Hamiltonian case),
to eliminate one of the variables. Then one chooses a suitable Poincare section for the ow and looks for xed points of the corresponding Poincare map.
This scheme can be extended to the case of several constants of motion and
has been extensively used in the literature; see, for example, [30]. This approach requires numerical integration of the dierential equations, which can
give errors in the case of very sti equations or very unstable orbits, and the
section must be adapted at each step in the continuation process to ensure

272

Jorge Gal
an-Vioque and Andre Vanderbauwhede

transversality. Furthermore, it is dicult to use integral phase constraints,

which, as explained in Chap. 10 of this book, often have signicant computational advantages over the classical Poincare phase condition. The underlying
idea of this scheme is that of the reduction of the dimension of the problem by
making use of the conserved quantities and/or the symmetries; it is a direct
translation of the standard theoretical treatment of the problem. A classical
example is the N -body problem (see, e.g., [35]) that will be discussed later in
this chapter.
The approach that we propose here extends the dimension rather than reducing it, to get the problem into a form where boundary value continuation
methods can be applied directly; a similar approach is discussed in Chap. 10.
Our formulation not only allows one to prove and extend some basic continuation results for periodic orbits in Hamiltonian systems such as the Cylinder
Theorem of [17] but it can also be implemented directly for the numerical
calculation of branches of periodic orbits.
Our starting point is a generalization of some continuation results of Sepulchre and MacKay for periodic orbits of systems having a rst integral. In [28]
the concept of a normal periodic orbit is introduced and it is shown that such
normal periodic orbits belong to one-parameter families. The key idea of their
approach is to embed the conservative equation in a one-parameter family
of dissipative systems by adding a small gradient perturbation term to the
vector eld in such a way that a periodic orbit of the perturbed system can
only exist when the perturbation is zero. Under the normality condition one
can invoke the Implicit Function Theorem to obtain a continuation result for
periodic orbits of the extended system. Because the periodic orbits only exist
when the perturbation is zero, one has in fact a continuation result for the
unperturbed conservative system.
The idea of adding a dissipative term, which allows periodic orbits only
when the dissipation is zero, is not new; for example, it is used in one of the
classical proofs of the Lyapunov Center Theorem where this theorem is shown
to correspond to a vertical Hopf bifurcation; see, e.g., [34]. The idea has been
used for numerical calculations, for example, in [1, 37].
To be more precise, consider a smooth n-dimensional vector eld g : Rn
n
R , and assume that the corresponding system
u = g(u)

(9.1)

has a non-trivial rst integral F : Rn R. This means that each orbit of

(9.1) is contained in a level set of F , and consequently F (u) g(u) 0. Let
u0 (t) be a periodic solution of (9.1) with initial point p0 := u0 (0), minimal
period T0 > 0, and monodromy matrix M . This monodromy matrix is given by
M = V (T0 ), where V : R L(Rn ) is the transition matrix for the variational
equation v = Dg(u0 (t)) v; the eigenvalues of M are the multipliers of the
periodic solution u0 (t). Assuming that F (p0 ) = 0 one can show that 1 is an
eigenvalue of M with geometric multiplicity mg 1 and algebraic multiplicity

9 Continuation in Hamiltonian Systems

273

ma 2; full proofs are in [23]. In order to continue the periodic solution u0 (t)
we replace equation (9.1) by the extended equation
u = T [g(u) + F (u)] ,

(9.2)

which depends on two (real) parameters T and . We look for 1-periodic

solutions u(t) of (9.2), with (u(0), T, ) near (p0 , T0 , 0). The basic observation
is that such solutions can only exist for = 0, since
 1
 1
0 = F (u(1)) F (u(0)) =
F (u(t)) u(t)
dt = T
F (u(t))2 dt.
0

The integral on the right-hand side is dierent from zero (namely, F (u(0))
is close to F (p0 ) = 0). We conclude that a 1-periodic solution of (9.2)
corresponds (after an appropriate time rescaling) to a T -periodic solution of
(9.1).
We denote the ow of (9.2) by u
(t; p, T, ), where p Rn is the initial
value. To nd the 1-periodic solutions we are looking for, we must impose
the periodicity condition u
(1; p, T, ) = p; to avoid phase shifts we must also
impose a phase condition. Therefore, we dene a mapping
G : Rn R R Rn R,
G (p, T, ) := (
u(1; p, T, ) p, g(p0 ), p p0 ),
where ,  denes a scalar product on Rn , and we have G(p0 , T0 , 0) = 0 by
assumption. Now we look for zeros (p, T, ) of G near (p0 , T0 , 0). Using the
Implicit Function Theorem in combination with the remark above, one can
then prove the following.
Theorem 1. Let u0 (t) be a periodic solution of the conservative equation (9.1)
with initial point p0 = u0 (0), minimal period T0 > 0 and monodromy matrix
M . Assume that F (p0 ) = 0 and that 1 is an eigenvalue of M with geometric
multiplicity mg = 1. Then the solution set of the equation G(p, T, ) = 0
consists locally near (p0 , T0 , 0) of a unique smooth curve along which 0.
More precisely, this solution curve can be written in the form {(p (T ), T, 0) |
T near T0 } for some smooth p : R Rn such that p (T0 ) = p0 .
This theorem is essentially a re-statement of the Cylinder Theorem for conservative systems [17] in a form that is adapted to numerical implementation. It
forms the simplest case of a more general continuation result for conservative
systems that can be found in [23] and that is described in Sect. 9.2 for the
particular case of Hamiltonian systems. If (9.1) has several independent rst
integrals then the condition mg = 1 of Theorem 1 is not satised. In this case
the more general theory of [23] is required. This arises in the continuation
of the N -body problem, where apart from the Hamiltonian also the components of the total linear momentum and the total angular momentum are rst
integrals; we refer to Sect. 9.3 for examples.

274

Jorge Gal
an-Vioque and Andre Vanderbauwhede

For the numerical implementation of Theorem 1 (or similar results) the

periodicity condition u
(1; p, T, ) = p is usually replaced by a boundary value
problem for the full solution {u(t) | 0 t 1}, and the Poincare-type phase
(t; p0 , T0 , 0) =
condition is replaced by an integral version. Setting u
0 (t) := u
u0 (T0 t), this leads to a boundary value problem of the following form:
(CON-1) Find (u(t), T, ) near (
u0 (t), T0 , 0) such that

u(t)

= T [g(u(t)) + F (u(t))] ,

u(1) = u(0),
(9.3)

 1
u
0 (t)), u(t) u
0 (t) dt = 0.
0

In the following sections we show some results of the numerical implementation of this approach to two particular examples, namely the mathematical
pendulum and the Restricted Three-Body Problem (R3BP). While discussing
these examples we put some emphasis on numerical issues related to the numerical implementation. In particular, in Sect. 9.1.3 we compare the shooting
method with the collocation method from a numerical point of view by computing a benchmark periodic solution of the R3BP taken from exercise 4.12
of [3]. In Sect. 9.1.4 we present partial bifurcation diagrams of the R3BP for
two values of the mass ratio that demonstrate the power and versatility of our
continuation scheme.
9.1.2 The Mathematical Pendulum
As a simple introductory example we consider the dimensionless mathematical
pendulum
x 1 = x2 ,
(9.4)
x 2 = sin(x1 ).
This vector eld is a one-degree-of-freedom Hamiltonian system, corresponding to the Hamiltonian H(x1 , x2 ) = 12 x22 + 1 cos(x1 ). The variable x1
S1 = R/2Z represents the angular displacement from the vertical axis, and
x2 R the angular velocity. The Hamiltonian has been chosen such that the
equilibrium at the origin, corresponding to the stable hanging solution, has
zero energy.
Equation (9.4) has a family of periodic orbits corresponding to librations
of the pendulum; they originate at the origin and terminate at a homoclinic
orbit to the saddle point (, 0). Figure 9.1(d) shows some representative periodic orbits of this family, which can be parametrized either by the energy,
the period (which increases monotonically from 2 to innity), or the maximal angular displacement. However, neither of these quantities is explicitly
available in (9.4). As we observed before, such behavior (which is nongeneric
in dissipative systems) is typical for conservative systems.
In order to calculate the family of periodic solutions we follow the approach
outlined in Sect. 9.1.1 and replace (9.4) by (a time rescaled version of) the
system

9 Continuation in Hamiltonian Systems

3

(a)

1
1

2
0.5

0.5

x1
x2

10

0.5

1
H

1.5

0.06

(c)
stepsize

2
1
0

0.5

1
H

1.5

(e1)
0.2

0.4

0.6

0.8

0.2

0.6
0.4
scaled time

0.8

0.2

0.6
0.4
scaled time

0.8

0
2
0

3
L2 norm

0
5
0
2

0
x1

(b)
Period

(d)

1
x2

L2norm

275

(e2)
(f)

0.04
0.02
0
0

Fig. 9.1. Continuation results for the mathematical pendulum. Panel (a) shows the
bifurcation diagram in with the vertical branch of periodic solutions. Panels (b)
and (c) show the period (plotted in a logarithmic scale) and the L2 -norm of the
solutions versus the energy H, respectively. Ten representative periodic orbits of the
one-parameter family are shown in panel (d). Panels (e1) and (e2) show the time
evolutions of x1 and x2 , respectively, while panel (f) shows the evolution of the time
step along the orbit for two solutions, namely one far from homoclinic and the other
very close to the homoclinic connection.

x 1 = x2 + sin(x1 ),
x 2 = sin(x1 ) + x2 .

(9.5)

This system still has two equilibria, the origin which is a stable or an unstable
focus depending on the sign of , and the saddle (, 0). For all non-equilibrium
solutions (x1 (t), x2 (t)) the function h(t) := H(x1 (t), x2 (t)) is strictly decreasing if < 0 or strictly increasing if > 0; this excludes periodic solutions for
= 0, which agrees with the theoretical results and is also conrmed by a
phase plane analysis of (9.5). So periodic solutions are only possible for = 0
in which case (9.5) coincides with (9.4) and we have the family of periodic
orbits mentioned before. The bifurcation diagram for periodic orbits of (9.5),
therefore, looks as in Fig. 9.1(a); this diagram very much resembles that of a
classical Hopf bifurcation, except that in this case the bifurcating branch is
completely vertical. It is also clear that cannot be used to parametrize the
family of periodic orbits.

276

Jorge Gal
an-Vioque and Andre Vanderbauwhede

Computationally (9.5) has the desired form, namely, with one external
parameter. Starting the computation from, for example, = 1 and with
initial point (x1 , x2 ) = (0, 0), a software package such as Auto [7, 9] will locate
(, x1 , x2 ) = (0, 0, 0) as a Hopf bifurcation point from the trivial solution and,
after switching branches, compute the vertical branch of periodic orbits.
Along this branch the value of , computed as part of the solution for each
continuation step, appears to be zero (up to numerical precision). The results
of such a computation are illustrated in Fig. 9.1; in particular, panels (b)
and (c) illustrate that either the energy H or the period can be used to
parametrize the family of periodic orbits in panel (d). Some further remarks
on the numerical implementation are in order.
1. In the pseudo-arclength continuation technique used by Auto there is no
distinguished parameter. In our particular example is just one of the
quantities that have to be computed at each continuation step. This allows
(for example) the computation along folds and, as is illustrated by the
example considered here, the continuation of vertical solution branches.
2. Orthogonal collocation with adaptive mesh selection is used in Auto to
solve the boundary value problem at each continuation step. Figure 9.1(f)
shows how the time step varies along the orbit: it shrinks at places where
the solution varies rapidly, whereas it remains large at slowly varying
segments of the orbit. This allows one to compute the family up to orbits
with large period, i.e., very close to the homoclinic orbit that terminates
the branch; see Fig. 9.1(b).
3. The integral phase condition keeps the segments with a rapid variation
of the solution component x2 at practically the same location when the
period becomes large; see Fig. 9.1(e). This allows for bigger continuation
steps compared with phase conditions that allow the dip in the prole to
move. For further details on this particular aspect of the computations
see [8] and Chap. 10.
9.1.3 Collocation Versus Shooting
We now analyze from a numerical point of view a slightly more demanding
example taken from exercise 4.12 of [3]. It is very much related to the other
examples of this chapter, namely, it concerns a periodic orbit of the R3BP. The
aim is to compare a standard shooting method with the orthogonal collocation
method of Auto. Both methods are powerful and versatile enough and have a
long tradition in the dynamical system community. In general, both methods
behave similarly in terms of eciency and accuracy.
The R3BP describes the dynamics of a body with negligible mass under the
gravitational inuence of two massive bodies, called the primaries, which move
in circular orbits about their barycenter. Let (x, y, z) denote the position of the
negligible-mass body in a rotating barycentric coordinate system, where the
x-axis points from the larger to the smaller primary, the z-axis is orthogonal to

9 Continuation in Hamiltonian Systems

277

the orbital plane, and the y-axis completes the orthogonal coordinate system.
The units are chosen so that the distance between the primaries, the sum
of the masses of the primaries, and the angular velocity of the primaries are
all equal to one. The problem then depends on a single external parameter,
denoted , which is the ratio of the mass of the smaller primary and the
total mass. The larger and smaller primaries are then located at (, 0, 0)
and (1 , 0, 0), respectively, and the equations of motion are given by
x
= 2y + x (1 )(x + )r13 (x 1 + )r23 ,
y = 2x + y (1 )yr13 yr23 ,
z = (1 )zr13 zr23 ,
where
r1 =

(x + )2 + y 2 + z 2

and r2 =

(9.6)

(x 1 + )2 + y 2 + z 2 .

This dynamical system has one integral of motion, namely, the Jacobi constant
E=

1
1 2
(x + y 2 + z 2 ) U (x, y, z) (1 ),
2
2

where

1
1 2
+ .
(x + y 2 ) +
r2
r1
2
We have used both a shooting and a collocation method to calculate
a particular benchmark solution of (9.6) as described in [3]; it is shown
in Fig. 9.2(a). This solution is periodic and planar (observe that the subspace z = 0 is invariant under (9.6)), and corresponds to the mass ratio
= 0.01277471 (which is very close to the Earth-moon case). For the calculations with the shooting method we have used the routine ode45 of Matlab
which is based on an explicit Runge-Kutta (4,5) formula (the Dormand-Prince
pair) with decreasing absolute and relative tolerances. For the collocation approach we used Auto, that is, piecewise-polynomial collocation with GaussLegendre collocation points; see Chap. 1. This so-called orthogonal collocation
has the desirable property of preserving the symplectic structure of Hamiltonian systems. Furthermore, as implemented in Auto, it determines at negligible cost the characteristic (Floquet) multipliers as a by-product of the
decomposition of the Jacobian of the collocation system. Hence, the stability
and bifurcation properties of the calculated periodic solution are available as
well.
Figure 9.2 shows the results of the calculations with the shooting algorithm; the collocation approach gives essentially identical results. Panel (a1)
and the enlargement (a2) show a near-collision (or yby) of the negligible,
small mass with the small primary at about t = 6. This results in a strong acceleration of the negligible mass; see Fig. 9.2(b1). This clearly demonstrates
the need for an adaptative mesh so that fast changes in velocity at nearcollisions are properly resolved; see panels (c1) and (c2). Indeed xed step
methods should be avoided.
U=

278

Jorge Gal
an-Vioque and Andre Vanderbauwhede
1.5

(a1)

y(t)

0.5

0
0.02

x(t)

0.5

0.5

x(t)

(b1)

x 10

x(t)

(d)

10
2

(b2)
y(t)

1.5
1.5

20

15

(c2)

2
2
5.95

2
y(t)

stepsize

0.99 x(t) 0.995

0.985

0.98
2

0.5

(a2)

0.02
y(t)

10
t

15

20

6.05

6.1

5.95

6.05

6.1

(c1)

0
2
4

10
t

15

20

Fig. 9.2. Benchmark periodic orbit of the R3BP with = 0.01277471. Panel (a1)
shows the orbit in projection onto the (x, y)-plane, and panel (a2) is an enlargement
close to the near collision; the primaries are depicted by a circle (for mass ) and
a square (for mass 1 ). Panel (b1) shows the rapid change of the velocity of the
negligible mass, and panel (b2) is an enlargement. Similarly, panels (c1) and (c2)
show the time evolution of the velocity component. Panel (d) is a plot of how the
adaptive stepsize used by the RK method of Matlab changes along the orbit; the
sudden drop of the stepsize around t = 6 corresponds to the near collision.

In order to compare the two computational schemes we should study the error, the number of time steps and the CPU-time. The lack of an exact solution
and the fact that Auto is a compiled program, while the shooting calculation
was performed with Matlab, only allows us to plot the error in the Cauchy
sense. That is, we plot in Fig. 9.3(a) the measure max(uk (t) uk1 (t)) of
the convergence between successive iterations as a function of the number of
time steps. From iteration (k 1) to k we increase the number of time intervals (NTST in Auto) or decrease the tolerances of the ode45 command. As
expected, the solution converges more rapidly as the number of time steps
is increased. The comparison between the error for the shooting algorithm
(squares) and the collocation approach (circles) in Fig. 9.3(a) demonstrates
that both methods show quite similar behavior. This might be expected since
shooting with Matlab uses a fourth-order Runge-Kutta method and collocation with Auto was performed with degree-four polynomials (i.e., with m = 4

9 Continuation in Hamiltonian Systems

279

10
S Sh
Col

10

Col
Col MN

error

error

10

10

10

(a)

(b)
4

10

10

10

10

10

10

10

min S Sh
max S Sh
min Col
max Col

stepsize

10

10

(c)
6

10

10

10

Fig. 9.3. Evolution of error and stepsize as a function of the number of time steps
for the benchmark periodic orbit of the R3BP from Fig. 9.2. Panel (a) shows the
evolution of the error in the Cauchy sense (max(||uk (t) uk1 (t)||)) for the shooting
algorithm (squares) and for collocation with Auto (circles). Panel (b) displays the
error of the collocation calculation at all mesh points (circles solid line) and only
at the main mesh points (circles dashed line); this illustrates the phenomenon of
superconvergence at the main mesh points. Both methods use either an adaptive step
or an adaptive mesh; panel (c) shows the maximal (solid lines) and minimal (dashed
lines) stepsize for both methods. Note that both methods perform comparably well
in this example.

collocation points per mesh interval). Indeed, if u(t) is suciently smooth

then the global accuracy of collocation is known to be of order m. Note from
Fig. 9.3(b) that convergence for the collocation results is much better at the
main mesh points; this is due to the phenomenon of superconvergence which
guarantees an error of order 2m; see Chap. 1. Finally, panel (c) shows that
the mesh adaptation for both methods is very similar. We remark that an
exhaustive comparison between shooting and collocation should include the
continuation process as well. Furthermore, one should possibly consider multiple shooting for dicult orbits or long time integration.
9.1.4 Families of Lyapunov Solutions in the R3BP
As an illustration of the power and versatility of our computational approach,
as outlined in Sect. 9.1.1, we now briey discuss some numerical results about

280

Jorge Gal
an-Vioque and Andre Vanderbauwhede

Fig. 9.4. Schematic bifurcation diagram for the R3BP (9.6) of the families of periodic orbits that emanate from the Lagrange point L1 for the mass-ratio = 0.01215
of the Earth-moon system. From E.J. Doedel, R. Paenroth, H. Keller, D. Dichmann, J. Gal
an-Vioque and A. Vanderbauwhede, Computation of periodic solutions
of conservative systems with application to the three-body problem, Internat. J.
c 2003 by World Scientic
Bifur. Chaos Appl. Sci. Engrg. 13 (2003) 13531381 
Publishing; reprinted with permission.

families of periodic solutions of the R3BP; these results are from [10, 11] where
also an extensive bibliography can be found.
It is well known that for each value of system (9.6) has ve equilibria,
called the Lagrange points or libration points. Three of the Lagrange points,
denoted L1, L2 and L3, are collinear with the primary bodies; one of them,
L1, lies between the two primaries. Studying the linearization of (9.6) at the
collinear Lagrange points and using the Lyapunov Center Theorem one can
show that from each of these collinear Lagrange points there emanate two
families of periodic orbits: the Lyapunov family containing planar orbits in
the (x, y)-plane, and the family of so-called Vertical orbits that starts o in
the vertical z-direction.
Figure 9.4 shows a schematic bifurcation diagram for the two families
emanating from L1 and for the secondary families of periodic orbits bifurcating
from them; this diagram was calculated for the mass ratio = 0.01215 that
corresponds to the Earth-moon system. The ve libration points are shown
as grey cubes. The red line (L) represents the Lyapunov orbits and the green
curve (V) represents the Vertical orbits. Any solution branch that intersects
the grey plane has a planar periodic solution at the intersection point. Along
the (red) Lyapunov family there are two branch points; at the rst of these the

9 Continuation in Hamiltonian Systems

281

Fig. 9.5. Schematic bifurcation diagram of the families emanating from L1 in the
RTBP for mass-ratio = 0.09. From E.J. Doedel, V.A. Romanov, R. Paenroth,
H. Keller, D. Dichmann, J. Gal
an-Vioque and A. Vanderbauwhede, Elemental periodic orbits associated with the Libration Points in the Circular restricted three-body
c 2007 by
problem, Internat. J. Bifur. Chaos Appl. Sci. Engrg. 17 (2007) at press 
World Scientic Publishing; reprinted with permission.

blue family (H) of so-called Halo orbits bifurcates, at the second bifurcation
point the yellow family (Y) branches o, connecting the Lyapunov family to
the Vertical family. On the blue family of Halo orbits there are two symmetryrelated bifurcation points which give rise to the cyan family (C) of orbits. In
turn, along this cyan family there are two symmetry-related branching points
which give the magenta family (M) that connects the non-collinear Lagrange
points L4 and L5; in fact, the magenta family forms the Vertical family
emanating from L4 and L5. More details on this bifurcation diagram and on
the orbits represented can be found in [10].
The picture becomes more complicated and more interesting if the massratio is allowed to vary. Detailed computational results for families that
emanate from the ve libration points in the R3BP for all values of are
presented in [11]. As a stimulating example of the richness and complexity
of the problem we show in Fig. 9.5 the schematic bifurcation diagram for
= 0.09. The number of families and subfamilies of periodic orbits and their
interconnections in this example are a stark reminder of the statement of
Poincare cited above: indeed, periodic orbits play the guiding role in the study
of complicated Hamiltonian systems such as the R3BP.

282

Jorge Gal
an-Vioque and Andre Vanderbauwhede

9.2 Theory of Continuation in Hamiltonian Systems

In Hamiltonian systems conserved quantities and symmetries are related by
Noethers theorem, and they result in families of non-isolated periodic orbits.
In this section we describe how the Implicit Function Theorem can be used
in this situation to prove continuation results via a set-up that can also be
exploited numerically; for full details and proofs we refer to [23].
We use the classical set-up for Hamiltonian systems with n degrees of
freedom. The state space is R2n = Rn Rn , with elements u = (x, y), and with
&2n
j . For each smooth function
the standard scalar product u, u
 = j=1 uj u
2n
H : R R the Hamiltonian vector eld XH : R2n R2n is dened by
XH (u) := JH(u), where J L(R2n ) is the standard symplectic matrix
given by J(x, y) := (y, x). We call
u = XH (u)

(9.7)

the Hamilton equation with Hamiltonian H, and we denote the ow of (9.7) by

H (t, u) = tH (u) (with t R and u R2n ). A smooth function F : R2n R
is a rst integral (constant of motion) for (9.7) if and only if {F, H} 0,
where the Poisson bracket {F, H} : R2n R of F and H is dened by
{F, H}(u) := DF (u) XH (u). The Poisson bracket is anti-symmetric, that
is, {F, H} = {H, F }, and, as a consequence, H is a rst integral of (9.7).
In case either XH or XF has at least one bounded orbit, the condition that
{F, H} 0 for F is a rst integral of (9.7) is equivalent to the fact that
for all t, s R the symplectic dieomorphisms sF and tH commute. This
is the statement of Noethers theorem: rst integrals generate symmetries,
and conversely, (symplectic and continuous) symmetries are generated by rst
integrals. We denote the space of all rst integrals of (9.7) by
F := {F C (R2n ; R) | {F, H} 0}.

(9.8)

Now assume that p0 R2n generates a nontrivial periodic orbit 0 :=

| t R} of (9.7), with minimal period T0 > 0. In order to continue
this periodic solution we look for solutions (T, p) near (T0 , p0 ) of the equation

{tH (p0 )

G0 (T, p) := H (T, p) p = 0.

(9.9)

This equation has the one-dimensional solution curve {T0 } 0 . If G0 were

submersive at (T0 , p0 ) then this would (locally) give us all solutions. However,
as we will see, this is not the case. To nd out about the submersivity properties of G0 we calculate the image of DG0 (T0 , p0 ) L(R2n+1 ; R2n ), which is
given by
Im DG0 (T0 , p0 ) = Im(M I) + R XH (p0 ),

M := DTH0 L(R2n ). (9.10)

Here M is the monodromy matrix of the periodic orbit 0 , and its eigenvalues
are the multipliers of the periodic orbit 0 . Since XH (p0 ) Ker(M I) there

9 Continuation in Hamiltonian Systems

283

is always the multiplier 1. We denote by mg and ma the geometric and the

algebraic multiplicity of the eigenvalue 1 of M . It is not hard to show that M
is symplectic, i.e., M T JM = J, and hence ma is even. One can also prove the
following.
Proposition 1. Let W := {F (p0 ) | F F}, Z0 := {G F | {G, F }(p0 ) =
0, F F}, and Z0 := {XG (p0 ) | G Z0 }. Let k := dim W . Then
Z0 JW Ker(M I)

and

Im(M I) + Z0 W .

(9.11)

In particular, mg k and ma k + dim Z0 .

Clearly H Z0 and XH (p0 ) Z0 ; therefore, referring back to (9.10), we see
that
(9.12)
Im DG0 (T0 , p0 ) = Im(M I) + RXH (p0 ) W .
Denition 1. We say that the periodic orbit 0 generated by p0 is normal if
in (9.12) we have equality, i.e., if
Im(M I) + RXH (p0 ) = W .

(9.13)

This notion of normality is a generalization of the one introduced in [28];

for a discussion of the relationship between the normality property and the
submersivity properties of the mapping G0 see [24]. The following proposition
gives conditions for 0 to be normal.
Proposition 2. The periodic orbit 0 of (9.7) is normal if and only if either
mg = k or mg = k + 1 and XH (p0 )  Im(M I). In particular, 0 is normal
if ma = k + 1.
We remark, in view of Proposition 1, that the condition ma = k + 1 can only
be satised if k is odd and dim Z0 = 1.
The usefulness of the concept of a normal periodic orbit becomes clear
when one combines the following two observations.
1. If p0 R2n generates a normal periodic orbit 0 with minimal period T0
then, by denition of normality, the subspace W is lacking in the image
of DG0 (T0 , p0 ), and therefore G0 is not submersive at the point (T0 , p0 ).
We can repair this lack of submersivity by adding some articial terms
to (9.7). More precisely, we replace (9.7) by
u = XH (u) +

k
%

j Fj (u),

= (1 , 2 , . . . , k ) Rk ,

(9.14)

j=1

where Fj F (1 j k) are chosen such that {Fj (p0 ) | 1 j k}

forms a basis of W (typically one will take F1 = H). Denoting the ow of
(9.14) by (t, p, ) and dening G : R R2n Rk R2n by G(T, p, ) :=
(T, p, ) p we see that G(T, p, 0) = G0 (T, p), G(T0 , p0 , 0) = 0, and

284

Jorge Gal
an-Vioque and Andre Vanderbauwhede

one can show that G is submersive at (T0 , p0 , 0). Hence, the solution set
of G(T, p, ) = 0 forms, locally near (T0 , p0 , 0), a (k + 1)-dimensional
submanifold of R R2n Rk . Since G(T, p, ) = 0 means that (t, p, ) is
a T -periodic solution of (9.14), this (k + 1)-dimensional solution manifold
is foliated by a k-parameter family of periodic orbits of (9.14).
2. Suppose that u(t) is a T -periodic solution of (9.14), and let F (u) :=
&k
j=1 j Fj (u). Using F F one then calculates that d/dt F (u(t)) =
F (u(t))2 , from which it follows that
 T
F (u(t))2 dt = F (u(T ))F (u(0)) = 0 F (u(t)) = 0, t R.
0

&k
In particular F (u(0)) =
j=1 j Fj (u(0)) = 0. Since the vectors
Fj (p0 ), 1 j k, are linearly independent, the same is true for the
vectors Fj (u(0)), 1 j k, if u(0) is suciently close to p0 . We conclude that (9.14) can only have a periodic orbit near 0 if = 0. In other
words, all periodic orbits of (9.14) near 0 are, in fact, periodic orbits of
(9.7).
By combining the two observations above, it follows that a normal periodic
orbit 0 of (9.7) belongs (locally) to a k-parameter family of normal periodic
orbits of the same equation (normality is preserved locally). The question how
we can parametrize this family has a straightforward answer only in the simplest possible case, namely if ma = k+1. Then the family can be parametrized
by the values of the rst integrals Fj , 1 j k. Such a parametrization may
fail when ma > k + 1.
Complementing the periodicity condition G(T, p, ) = 0 with appropriate
phase conditions also gives an ecient way for actually calculating the periodic
orbits. Starting from the point p0 on 0 we can generate a k-dimensional
submanifold of initial points for periodic orbits of (9.7) by applying the ows
of the Hamiltonian vector elds XFj , 1 j k. So what we really need is a
way to calculate a one-dimensional solution curve in the missing direction;
such a curve is given by the following theorem; see [23] for the proof.
Theorem 2. Let 0 = {H (t, p0 ) | t R} be a normal T0 -periodic solution of
u = XH (u). With the notations introduced before, consider the following set
of equations for (T, p, ) R R2n Rk :
G(T, p, ) = 0,

XFj (p0 ), p p0  = 0,

1 j k.

(9.15)

Then near (T0 , p0 , 0) the solution set of (9.15) consists of a smooth onedimensional curve along which 0. In the case mg = k this curve can
be parametrized by the period T . Projecting the solution curve onto the phase
space R2n and acting on the projection with the ows of the Hamiltonian vector elds XFj , 1 j k, generates a (k + 1)-dimensional manifold that is
invariant under the ow of XH . Furthermore, this manifold is foliated by a
k-parameter family of normal periodic orbits of u = XH (u).

9 Continuation in Hamiltonian Systems

285

The condition mg = k appearing in Theorem 2 is typically satised, so that

the period T can be used as a parameter along the solution curve. For numerical calculations the k phase conditions in (9.15) will usually be replaced
by integral versions; see [8] and Chap. 10 for an extensive discussion of such
integral phase conditions. Also, instead of looking for T -periodic solutions of
(9.14), it is easier to rescale time and to look for 1-periodic solutions of the
rescaled equation. The way to implement the theoretical continuation result
above then takes the form of the following boundary value problem:
(CON-2) Find (u(t), T, ) near (u0 (t), T0 , 0), with u0 (t) := H (T0 t, p0 ), such
that
.
1

&k

u(t)

=
T
X
(u(t))
+

F
(u(t))
,
H
j

j=1 j

(9.16)
u(1) = u(0),

1
XFj (u0 (t)), u(t) u0 (t) dt = 0,
1 j k.
0
Such boundary value problems are very well suited for pseudo-arclength continuation as implemented, for example, in Auto.
In a number of examples (in particular in N -body problems) the Hamiltonian vector eld XH has a scaling property that allows one to obtain new
solutions from given ones by appropriate rescalings of time and phase space
variables. In this case, Theorem 2 only gives a family of rescaled copies of the
starting periodic orbit 0 . To obtain a more meaningful result one can x the
period and perform a continuation in an external parameter.
We now give an example of the type of results one can prove in this direction. Consider a Hamiltonian H (u) depending smoothly on a real parameter
R and suppose that apart from H the Hamiltonian system
u = XH (u)

(9.17)

has some further independent rst integrals Fj , 2 j k, i.e., {H , Fj } 0,

2 j k, for all R. Suppose also that for = 0 equation (9.17) has
a periodic orbit 0 = {tH0 (p0 ) | t R} with minimal period T0 > 0. We
set F0 := {F C (R2n ; R) | {H0 , F } 0} and W0 := {F (p0 ) | F F0 }.
Finally we assume that {H0 (p0 ), F2 (p0 ), . . . , Fk (p0 )} forms a basis of W0
and that mg = k. Denote by (t, p, , ) the ow of the modied system
u = XH (u) + 1 H (u) +

k
%

j Fj (u).

j=2

Theorem 3. Under the above assumptions and for each xed T near T0 the
set of equations
(T, p, , ) = p,

XFj (p0 ), p p0  = 0 (1 j k, F1 := H0 ),

(9.18)

has a unique one-dimensional solution branch near (p, , ) = (p0 , 0, 0). Furthermore, 0 along this branch, and it can be parametrized by .

286

Jorge Gal
an-Vioque and Andre Vanderbauwhede

Theorem 3 lends itself to numerical implementation much like Theorem 2. For

an application of Theorem 3 to the three-body problem see Sect. 9.3.
9.2.1 Continuation of Relative Equilibria
As we have seen, Noethers Theorem states a strong relationship between
symmetries and rst integrals of Hamiltonian systems. We now go a bit deeper
into the symmetry aspect; see also Chaps. 10 and 11.
Consider a nite-dimensional subspace G of C (R2n ; R) with the property
that {F, G} G for all F, G G (i.e., G together with the Poisson bracket {, }
forms a Lie algebra). Denote by G the (possibly noncompact) Lie group of
symplectic dieomorphisms generated by the one-parameter groups {sF | s
R}, with F G. If for some Hamiltonian H : R2n R we have {H, F } 0
for all F G (in our earlier notation this means that G F) then the
Hamiltonian system (9.7) is equivariant with respect to the group G.
Denition 2. An orbit {H (t, p0 ) | t R} of (9.7) is a relative equilibrium
(with respect to the group G) if it is contained in the group orbit G(p0 ) =
{(p0 ) | G}; this means that there is a mapping : R G such that
H (t, p0 ) = (t)(p0 ).
It is not hard to show that p0 R2n generates a relative equilibrium if and only
if there is some F G such that XH (p0 ) = XF (p0 ). This means that p0 is an
equilibrium of the Hamiltonian vector eld XHF and that tH (p0 ) = tF (p0 ).
If Fj G, 1 j m, are such that {XFj (p0 ) | 1 j m} forms a basis
of Y := {XF (p0 ) | F G} then the condition for p0 to generate a relative
equilibrium is that there exist numbers j0 R, 1 j m, such that
XH (p0 ) =

m
%

j0 XFj (p0 ).

(9.19)

j=1

The corresponding
relative equilibrium is given by tH (p0 ) = tF (p0 ), where
&m
0

F
.
Together
with p0 R2n also all the other points on the
F =
j=1 j j
group orbit G(p0 ) generate a relative equilibrium; this (m-dimensional) group
orbit is then foliated by the XH -orbits of its elements.
A simple example of relative equilibria appears in the N -body problem
where for G one can take either the rotation group generated by the components of the total angular momentum or the Euclidean group generated by the
components of the total linear momentum or the total angular momentum.
For the rotation group the relative equilibria correspond to equilibria in a uniformly rotating frame. Well-known examples are the Lagrange points in the
circular restricted three-body problem and the Euler and Lagrange solutions
of the three-body problem with three equal masses; see already Sect. 9.3. In
each of these examples the relative equilibria are also periodic solutions of the
corresponding Hamiltonian system.

9 Continuation in Hamiltonian Systems

287

For the remaining part of this subsection we make the restrictive assumption that G Z F, i.e., {F, G} 0 for all F F and all G G. This
implies, in particular, that G is Abelian. As is shown in forthcoming work of
Wul and Schebesch [36], one can cancel this assumption by using some more
elaborate symplectic geometry. Assuming (9.19) we complement the Fj G,
1 j m, with some further rst integrals Fj F, m + 1 j k, such that
{Fj (p0 ) | 1 j k} forms a basis of W := {F (p0 ) | F F}. We also set
L := DXH (p0 )

m
%

j0 DXFj (p0 ) L(R2n ).

(9.20)

j=1

Proposition 3. Under the conditions above we have

Y JW Ker(L)

and

Im(L) + Y W .

(9.21)

As a consequence, 0 is an eigenvalue of L with geometric multiplicity m

g k
and algebraic multiplicity m
a k + m.
Denition 3. We say that the relative equilibrium generated by p0 is normal
if we have equality in the second inclusion of (9.21), i.e., if
Im(L) + Y = W .

(9.22)

This normality condition is satised when either m

g = k or m
a = k + m. One
can prove the following continuation result for normal relative equilibria [23].
Theorem 4. Let G Z, and suppose that p0 R2n generates a normal
relative equilibrium of XH with respect to G. Then, in the notation introduced
above, the set of equations
XH (p) =

m
%
j=1

j XFj (p) +

k
%

j Fj (p), XFj (p0 ), p p0  = 0, 1 j k,

j=1

(9.23)
has a solution set near (p, , ) = (p0 , 0 , 0) R2n Rm Rk . The solution
set is a smooth m-dimensional submanifold, along which
0. For each
&m
(p, , 0) on this solution manifold we have that XH (p) = j=1 j XFj (p) and
that p generates a normal relative equilibrium of XH with respect to G.
In case m
g = k (which is generically satised) the solution manifold of (9.23)
can be parametrized by = (1 , 2 , . . . , m ). One can then also x at any
value close to 0 and use a similar set-up to continue the relative equilibrium
in external parameters; see the end of the next section for an example.

9.3 Continuation of the Figure-8 Solution

Celestial Mechanics has been at the origin of the theory of dynamical systems
and many of the techniques from that theory were developed to analyze the

288

Jorge Gal
an-Vioque and Andre Vanderbauwhede

fascinating behavior of a group of massive bodies under the inuence of their

mutual gravitational interaction.
The spectacular discovery by Chenciner and Montgomery [5] of the existence of a new solution of the Three-Body Problem (3BP) with equal masses
in which all three bodies follow the same planar curve with the shape of a gure eight has brought great excitement to the dynamical systems community.
The origin of the name of the orbit becomes apparent from a representation
in physical space such as in Fig. 9.7 (a). This solution was rst discovered (numerically) by Moore [20] in the context of a study of possible braid types associated with the planar N -body problem. The method of proof by Chenciner
and Montgomery is based on variational arguments. After some reductions,
the action integral is minimized on a restricted set of symmetric arcs to prove
the existence of a solution where the three bodies of equal masses chase each
other along a single closed trajectory. However, the variational proof is unable
to determine the stability of the solution.
Sim
o [31] computed this remarkable solution numerically with great accuracy and announced elliptic stability; i.e., the non-trivial characteristic multipliers of the periodic orbit are on the unit circle. The precise values of
the non-trivial characteristic multipliers (those which are dierent from one)
are given in [31] as j = exp(2ij ), j = 1, 2, with 1 = 0.00842272 and
2 = 0.29809253. Note that the smallness of 1 indicates that the gure-8
solution is close to a bifurcation. Sim
o [32] also discovered many (hundreds
of) other similar solutions for three equal bodies, as well as for N equal bodies
with 3 < N < 799 [31]. The dening property of these solutions, which are
referred to as choreographies, is that all bodies follow a single closed curve in
phase space, with a xed time between each of the bodies. From the historical
point of view, the solution found by Lagrange in 1772, in which the three
bodies form the vertices of an equilateral triangle that rotates with constant
angular velocity around its midpoint, can be considered as the rst choreography. It has taken more than two hundred years to nd the second one.
In this section we apply our continuation scheme to following the gure-8
solution of the 3BP; see also [10, 12] for more details. The original motivation
to study this problem was a conjecture by Joe and Herb Keller that it would
be possible to connect the two simple Lagrange and gure-8 choreographies in
a continuous way by only following periodic orbits. In fact, Marchal [14] has
found a family of periodic orbits in a rotating frame (relative periodic orbits)
that connects these two highly symmetrical solutions. A further reason for trying to connect the Lagrange and gure-8 choreographies has to do with some
controversy about the stability properties of the solution corresponding to the
absolute minimizer of the action over certain homotopy classes of loops in conguration space. In Hamiltonian systems with two degrees of freedom such
minimizing orbits are always unstable [4]; however, for higher-dimensional
systems there are counterexamples. The stable gure-8 solution was obtained
by minimizing the action over all loops with a particular symmetry; see [5].
Assuming that the property that minimizing orbits are unstable also holds

9 Continuation in Hamiltonian Systems

289

for the equal mass 3BP there must exist some other unstable periodic orbit
in the homotopy class of the gure-8 solution with a lower action. Could this
orbit be the equilateral Lagrange choreography? It is unstable, and its action
value (332/3 19.60436 when the period equals 2) is lower than that of
the gure-8 solution ( 24.37197), but we do not know whether it is in the
homotopy class of the gure-8 solution. If it is not then the minimizer over
this homotopy class has to be some other unstable periodic orbit which will
then probably not be a choreography. Finding some connection between the
gure-8 and the Lagrange solutions would at least give some partial answers
to the above questions.
We have applied the general continuation scheme of Sect. 9.2 where we
used the numerically computed gure-8 solution as the starting solution. The
equations of motion of three bodies with masses m1 , m2 and m3 under mutual
gravitational attraction take the form

x1 x3
x1 x2
,
m3
|x1 x3 |3
|x1 x2 |3
x2 x3
x2 x1
2 = m1
,
m3
x
|x2 x3 |3
|x1 x2 |3
x3 x2
x3 x1
3 = m1
.
m2
x
3
|x3 x2 |3
|x1 x3 |

1 = m2
x

(9.24)

Here xi = (xi , yi , zi ) R3 denotes the position of the i-th body, i = 1, 2, 3,

and the universal gravitational constant has been set to 1. This system can
be rewritten as a rst-order system of dimension eighteen. There are seven
independent conserved quantities: &
the total energy, the three components of
3
i , and the three components of
the total linear momentum P =
i=1 mi x
&
3
i ; they are an immediate
m
the total angular momentum L =
i xi x
i=1
consequence of the invariance of the equations under time shifts, translations
and rotations.
Additionally, the equations are invariant under the transformation (t, x) 
3
2
(c t, c x), for an arbitrary constant c > 0; see [35]. Due to this scaling property there is a trivial continuation of periodic orbits in the period: arbitrarily
close to any periodic orbit there is another one with slightly dierent period,
obtained from the rst by rescaling. Obviously it has the same stability properties. In order to avoid this trivial continuation we x the period (say to 2)
and use instead the mass m1 of the rst body as an external continuation
parameter. Both other masses are kept equal to 1. It is not dicult to check
numerically that the gure-8 solution (corresponding to m1 = 1) is normal
(we have mg = k = 7) and, hence, we can apply Theorem 3 with = m1 .
As already discussed before, when setting up the continuation algorithm the
phase conditions in (9.18) are replaced by integral versions; for the detailed
computational formulation we refer to [10].
The calculations were performed with Auto [7, 9]. While following a oneparameter family of periodic orbits we also monitor the stability of these orbits

290

Jorge Gal
an-Vioque and Andre Vanderbauwhede

2.1364

L2 norm

2.1363

LP

BP
A

2.1362
0.99980

0.99990

1.00000

m1

Fig. 9.6. Local bifurcation diagram near the gure-8 solution under variation of the
mass m1 . Solutions for the solid section of the curve are stable and those for dashed
curves are unstable. Stable solutions appear in a narrow window between a pitchfork
bifurcation (BP) and a limit point (LP). Two orbits for the 3BP with equal masses
(m1 = 1) are indicated: at A one nds the stable Chenciner-Montgomery gure-8 solution shown in Fig. 9.7(a), and at B one nds the unstable satellite gure-8 solution
shown in Fig. 9.7(b). From J. Gal
an, F.J. Mu
noz-Almaraz, E. Freire, E.J. Doedel
and A. Vanderbauwhede, Stability and bifurcations of the gure-8 solution of the
c 2002 by the American
three-body problem, Phys. Rev. Lett. 88 (2002) 241101 
Physical Society; reprinted with permission.

and the appearance of new branches at bifurcation points; at such bifurcation

points Auto allows one to switch branches and start a new continuation
process.
Starting at the gure-8 solution for m1 = 1, the rst output of our continuation algorithm are the nontrivial characteristic multipliers of this gure-8
solution; we obtain j = exp(2ij ) (j = 1, 2), with 1 = 0.0084227 and
o [31] constitutes
2 = 0.2980925. The good agreement with the results of Sim
a successful practical test for our method.
Figure 9.6 shows the results of the continuation of the gure-8 solution
in a small mass-interval around m1 = 1, where we plot the L2 -norm of the
solution as a function of m1 . The solution labeled A is the Moore-ChencinerMontgomery gure-8 solution that is the starting point of our calculation.
This special planar orbit is plotted in Fig. 9.7(a). When decreasing m1 from
m1 = 1 we obtain a single solution branch with a pitchfork bifurcation at
the point BP. Increasing m1 from A results in a solution branch that reaches
a limit point LP and then continues in the direction of decreasing values
of m1 . All solutions along the branch are unstable (hyperbolic), except for
those on the section of the branch between BP and LP where the solutions
are stable (elliptic). This stable part of the branch corresponds to a very

9 Continuation in Hamiltonian Systems

0.4

291

(a)

0.2

0.0
-0.2
-0.4
0.4

(b)

0.2

0.0
-0.2
-0.4
-1.2

-0.2

0.8

Fig. 9.7. Representation in physical space of the gure-8 solution (a) and of the
satellite gure-8 solution (b) of the 3BP with equal masses that correspond to labels
A and B in Fig. 9.6. From J. Gal
an, F.J. Mu
noz-Almaraz, E. Freire, E.J. Doedel
and A. Vanderbauwhede, Stability and bifurcations of the gure-8 solution of the
c 2002 by the American
three-body problem, Phys. Rev. Lett. 88 (2002) 241101 
Physical Society; reprinted with permission.

narrow m1 -interval of the order of 105 . Continuing the branch beyond the
limit point we return to a situation where all three masses are equal, i.e.,
m1 = 1. The corresponding solution, labeled B in Fig. 9.6, is hyperbolic and
by construction in the same homotopy class as the gure-8 solution. However,
it has less symmetry and is no longer a choreography: as is shown in Fig. 9.7
(b), the three bodies follow three slightly dierent gure-8 paths. We refer to
this solution as the satellite gure-8 solution; it was also found numerically
by Sim
o [32].
At the bifurcation point (BP) there is a pitchfork bifurcation at which
the interchange symmetry of the second and third body (which both have
the same mass m2 = m3 = 1) is broken. At this bifurcation two symmetryrelated branches are born that are represented by a single curve in Fig. 9.6.
Along these branches the solutions are hyperbolic and the mass m1 increases.
Also along these branches one nds a special point where m1 = 1, so that,
therefore, all three masses are equal. The corresponding solution is the same
as before, namely the satellite gure-8 solution of Fig. 9.7(b) that occurs at
the point B; however, the labeling of the three bodies is now dierent. The
fact that all three branches intersect at B is due to the chosen representation.
It is clear from (9.24) that by choosing appropriate units one can always
assume that m3 = 1, leaving m1 and m2 as (dimensionless) parameters. Since
the gure-8 solution (corresponding to m1 = m2 = 1) is normal it can be
continued in both these parameters with a multi-parameter version of The-

292

Jorge Gal
an-Vioque and Andre Vanderbauwhede

m2

1.0001

1.0000

BP

0.9999

0.9998

BP

0.9999

1.0000

m1

1.0001

Fig. 9.8. Stability region in the (m1 , m2 )-plane for the continuation of the gure-8
solution. From J. Gal
an, F.J. Mu
noz-Almaraz, E. Freire, E.J. Doedel and A. Vanderbauwhede, Stability and bifurcations of the gure-8 solution of the three-body
c 2002 by the American Physical Soproblem, Phys. Rev. Lett. 88 (2002) 241101 
ciety; reprinted with permission.

orem 3. Figure 9.8 shows the result of a continuation of solutions with a

numerical determination of their stability. For mass values in the shaded triangular stability region in the (m1 , m2 )-plane the continued gure-8 solution
is elliptic; the point labeled A in the center corresponds to the gure-8 solution. Along the border of this stability region (solid curve) the continuation
manifold exhibits a fold (LP), and the two points labeled BP correspond to
branch points. The diagram is obviously symmetric with respect to the diagonal m1 = m2 , which is a consequence of the invariance of the equations under
the symmetry (x1 , x2 , m1 , m2 )  (x2 , x1 , m2 , m1 ).
In principle one could try to continue the gure-8 solution of the 3BP to a
solution of the R3BP by moving along the diagonal and increasing the value
of m1 = m2 . However, in practice one nds that, as the two equal masses
become larger, the bodies collide.
There is one further remarkable point that came out of our calculations.
Chenciner and Montgomery [5] obtained the gure-8 solution A by minimizing
the action over a class of loops with some particular symmetry properties. The
satellite gure-8 solution B minimizes the action over a much larger class of
paths. Therefore, the action corresponding to solution B should be less than
or equal to the action corresponding to solution A. This is indeed the case,
but in a rather unexpected way: within the precision of our calculations both
solutions have the same action. Using the standard denition of the action
and making the necessary normalizations, the value of the action integral for

9 Continuation in Hamiltonian Systems

293

2.3
2.25

(b)

2.2

L2 norm

2.15

(a)
2.1
2.05
2 (d)
1.95
1.9
0

(c)
0.2

0.4

0.6

0.8

Fig. 9.9. Bifurcation diagram showing branches of periodic orbits that connect the
3BP with equal masses to the R3BP with = 1/2. The solutions corresponding to
the marked points (a)(d) along the branches are shown in Fig. 9.10.

both solutions is found to be S = 24.37197. This surprising result reveals a

degeneracy that deserves further analysis.
Up to now we have restricted the description of our continuation and bifurcation results to a neighborhood of the gure-8 solution. However, nothing
prevents us from carrying out the continuation procedure over larger intervals
for varying mass m1 . In Fig. 9.9 we show a (partial) global bifurcation diagram
spanning the interval [0, 1] for m1 . As before, the plot shows the L2 -norm of
some calculated periodic orbits as a function of the mass m1 of the rst body,
while the two other masses m2 and m3 are equal to 1. The left-hand border at
m1 = 0 corresponds to the R3BP with two equal primaries ( = 1/2, known
as the Sitnikov problem), and the right-hand border at m1 = 1 to the 3BP
with three equal masses. Figure 9.10 shows representations in physical space
of the orbits corresponding to the marked points in the bifurcation diagram
of Fig. 9.9.
At the right-hand border of Fig. 9.9 there are two particular solutions of
the 3BP with equal masses, namely the gure-8 solution that is also shown in
Fig. 9.10 (a) and the equilateral Lagrange solution. (The representation of the
Lagrange solution in physical space is simply a circle with radius 31/6 and is
not shown in Fig. 9.10). The branch that emerges from the gure-8 solution
and crosses the upper-right part of the diagram is the continuation of the lower
branch in Fig. 9.6 that contains A and BP. All orbits along this branch are
planar; this is an outcome of the calculations, not an imposed condition. The
branch ends (just outside of the diagram) in a collision orbit where the smaller
body (with mass m1 ) collides with one of the larger bodies. Along the branch
we have located a bifurcation point, marked (b) in Fig. 9.9 and corresponding

Jorge Gal
an-Vioque and Andre Vanderbauwhede

0.4

(a)

0.2

0.5

294

0.2

0.4

0.5

0.5

0
x

0.5

(c)

0.5

0.5

0
x

0.5

0.5

0
x

0.5

(d)

0.5

(b)

0.5

0
0.5

1
1

0
z

0.5 0
x

0.5

1
1

Fig. 9.10. Representation in physical space of solutions (a)(d) that correspond to

the marked points in the global bifurcation diagram of Fig. 9.9.

to the solution depicted in Fig. 9.10(b). For this solution the small mass follows
the 8-shaped curve in the middle, while the two other masses each follow one
of the curves at the top or the bottom. Along the new branch emanating
from this bifurcation point the planar symmetry is broken, i.e., the solutions
along this branch have a non-vanishing z-component. This branch of threedimensional solutions can be continued all the way to m1 = 0. Here we nd a
special solution (marked (c) in Fig. 9.9) of the so-called Sitnikov problem, that
is, of the R3BP with equal masses for the primaries ( = 1/2). Figure 9.10(c)
shows a three-dimensional plot of the solution: the two primaries follow the
same (planar) circular orbit, always staying opposite to each other, while the
negligible mass traces out the other curve that is intertwined with the circle.
We conclude that it is possible, by changing the mass m1 , to connect the
gure-8 solution of the 3BP with equal masses to this particular solution of
the R3BP with = 1/2. This connection, calculated with the implementation
of Theorem 3 as described before, was rst reported in [10]; further details on
the R3BP for = 1/2 can be found in [11] and in references therein.
We now turn to the equilateral Lagrange solution and its continuation
when the mass m1 is changed. We choose as our starting orbit the particular
Lagrange orbit with period 2. This solution of the 3BP with equal masses
is both a (planar) periodic solution as well as a relative equilibrium under

9 Continuation in Hamiltonian Systems

295

the group G of rotations generated by the components of the total angular

momentum. However, when we try to continue this Lagrange solution we face
several diculties. First, the solution is not normal when considered as a
periodic solution. This is mainly due to the fact that it belongs to the socalled homographic family of planar periodic solutions where the three bodies
are located at the vertices of an equilateral triangle, each traveling along an
appropriate elliptic or circular Kepler orbit; see e.g. [18]. Hence, we cannot
apply the continuation results of Sect. 9.2. On the other hand we also cannot
apply Theorem 4 (or a variant of it) since Z contains only the multiples of the
Hamiltonian and, hence, G  Z. Both diculties remain when we consider
only planar solutions xi R2 .
The way out of these diculties lies in a restriction to planar solutions
that have a xed center of mass at the origin. When we set zi = 0, i = 1, 2, 3,
in (9.24) then the equivalent rst order system is Hamiltonian of dimension
twelve and has four (independent) rst integrals, namely: the Hamiltonian,
the two components of the total linear momentum, and the total angular
momentum Q. Along the Lagrange solution (which is a relative equilibrium)
we have H = Q and XH = XQ . Recall that we have xed the period at 2,
so that 0 = 1. Since the total linear momentum is a constant of motion, we
can choose an inertial frame of reference such that the center of mass remains
xed at the origin. For a general choice of the masses, this gives the condition
3
%
i=1

mi xi = 0

3
%

mi x i = 0.

(9.25)

i=1

These equations can be used to eliminate the third body from the equations; see, e.g., [29, 16] where it is shown how this can be done in a
canonical way. The resulting reduced system is Hamiltonian of dimension
eight and with two remaining rst integrals, namely: the Hamiltonian H
and the total angular momentum Q, both of course transformed by the
elimination process. Since {H, Q} 0 we have Z = F and, setting G
equal to the multiples of Q, it follows that G Z. Along the Lagrange
relative equilibrium we still have XH = XQ , so that k = m = 1. In
[29] the characteristic polynomial of the operator L was explicitly calculated and shown to have the form p() = 2 (2 + 1)(4 + 2 + ) with
2
= 94 for m1 = m2 = m3 = 1.
= 27
4 (m1 m2 +m2 m3 +m3 m1 )(m1 +m2 +m3 )
This shows that the algebraic multiplicity m
a of the zero eigenvalue equals
2 = k + m. This was conrmed by a symbolic computation in Mathematica,
which also showed that the geometric multiplicity is equal to one, that is,
m
g = 1 = k. We conclude that for m1 = 1 the Lagrange relative equilibrium
is normal (in the restricted setting) and, hence, we can apply the continuation
results of Sect. 9.2.1.
In order to avoid rescalings we keep, as before, the period xed at 2 and
use a parameter-dependent version of Theorem 4 to do continuation in the
mass m1 of the rst body. Figure 9.9 shows the result of our calculation in

296

Jorge Gal
an-Vioque and Andre Vanderbauwhede

the form of a unique branch of relative equilibria that connects the Lagrange
solution (lower point at m1 = 1) to a solution of the R3BP with equal primaries (point (d) for m1 = 0). The relative equilibria along this continuation
branch are well known: they are homographic solutions where the three bodies are at the vertices of an equilateral triangle, with the two larger bodies
(with masses m2 = m3 = 1) traveling along a circle around the origin, and
the smaller body with mass m1 rotating along a larger circle. The solution of
the R3BP at the end point (d) of the branch is nothing else but one of the
Lagrange libration points L4 or L5 which one depends on the direction of
rotation of the starting equilateral Lagrange orbit. In a rotating frame, such
as used in Sects. 9.1.3 and 9.1.4, this solution is an equilibrium. In a xed
frame the solution looks as shown in Fig. 9.10(d). At each moment the three
masses form an equilateral triangle: the two primaries rotate and are opposite
to each other on the smaller circle, while the negligible mass rotates along the
larger circle.
Along the branch of relative equilibria we nd no bifurcations (of relative
equilibria). Indeed, the analytical results of Siegel and Moser [29] mentioned
before show that the algebraic multiplicity m
a of the zero eigenvalue of the
linearization remains equal to two. Therefore, all of these relative equilibria
are normal and have unique continuations. We also checked numerically that
the geometric multiplicity m
1 remains equal to 1 along the full branch. The
results of Siegel and Moser also allow us to calculate the multipliers of these
relative equilibria when interpreted as periodic orbits of the planar version of
(9.24). On top of the multiplier 1 with algebraic multiplicity ma = 8 there is
1 ), where = exp(2) and is any
a quadruple of multipliers (,
, 1 ,
2
4
2
.
solution of + + (m1 ) = 0 and (m1 ) = 27
4 (2m1 + 1)(m1 + 2)
To conclude this section we observe that, although we have been able
to show continuous connections between both the gure-8 and the Lagrange
solutions, on the one hand, and specic solutions of the R3BP with = 1/2,
on the other hand, we have not found a connection from the gure-8 to the
Lagrange solution. Therefore, some of the issues discussed earlier remain open.

9.4 Conclusions
We have shown how two-point boundary value problem continuation software
can be used to compute families of periodic solutions of symmetric Hamiltonian systems. The theory and the numerical implementations are well developed but not complete. Further work is necessary on the continuation
of, rst, relative equilibria in the non-Abelian case and, second, of relative
periodic orbits; see [36] for some progress in this direction. How to make
use of reversibility properties in combination with the Hamiltonian structure
also needs further attention; see [22, 24]. An approach for the study of nonholonomic systems based on the ideas in this chapter has been proposed in
[19].

9 Continuation in Hamiltonian Systems

297

In the 3BP and R3BP there are plenty of interesting problems to be investigated and our results can be of some help in this adventure. However,
as Poincare announced, and the last 110 years of Celestial Mechanics have
shown, one can spend several lifetimes following periodic orbits in this incredibly complex jungle of trajectories!
Acknowledgements
We are very much indebted to Sebius Doedel, Emilio Freire and Francisco
Javier Mu
noz-Almaraz. Most of the material of this chapter is the result of
our ongoing collaboration over the last 10 years. In particular, we would like to
mention that Sebius Doedel was the driving force behind the massive project
to obtain a complete picture of families of periodic orbits in the R3BP that
was briey discussed in Sect. 9.1.4. We are very thankful to the publishers
of [10], [11] and [12] for the permission to make use of Figs. 9.4, 9.5, 9.6,
9.7 and 9.8. This work has been partially supported by the Spanish Ministry
of Education through the grant BFM2003-00336 and MTM2006-00847. A.V.
also acknowledges support from the same institution for his sabbatical stay
at the University of Seville through grant SAB2005-0188.

References
1. D. G. Aronson, E. J. Doedel, and H. G. Othmer. The dynamics of coupled
current-biased Josephson junctions II. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 1: 166, 1991.
2. V. I. Arnold, V. V. Kozlov, and A. I. Neishtadt. Dynamical Systems III.
Mathematical Aspects of Classical and Celestial Mechanics. (Springer-Verlag,
Berlin, 1993).
3. U. Ascher and L. P. Petzold. Computer Methods for Ordinary Dierential Equations and Dierential-Algebraic Equations, (SIAM, Philadelphia, 1998).
4. G. D. Birkho. Dynamical Systems. (Amer. Math. Soc., Ann Arbor, 1927).
5. A. Chenciner and R. Montgomery. A remarkable periodic solution of the threebody problem in the case of equal masses. Annals Math., 152: 881901, 2000.
6. S. N. Chow and J. K. Hale. Methods of Bifurcation Theory, Grundlehren der
Mathematischen Wissenschaften, 251. (Springer-Verlag, Berlin, 1982).
7. E. J. Doedel. Auto, a program for the automatic bifurcation analysis of autonomous systems. Congr. Numer., 30:265384, 1981.
8. E. J. Doedel, H. Keller and J.P. Kernevez. Numerical analysis and control of
bifurcation problems: II. Internat. J. Bifur. Chaos Appl. Sci. Engrg., 1:745772,
1991.
9. E. J. Doedel, R. Paenroth, A. Champneys, F. Fairgieve, Yu. A. Kuznetsov,
B. Oldeman, B. Sandstede, and X. Wang. AUTO2000: Continuation and bifurcation software for ordinary dierential equations. Department of Computer Science, Concordia University, Montreal, Canada, 2000. Available from
http://sourceforge.net/projects/auto2000/

298

Jorge Gal
an-Vioque and Andre Vanderbauwhede

10. E. J. Doedel, R. Paenroth, H. Keller, D. Dichmann, J. Gal

an-Vioque, and
A. Vanderbauwhede, Computation of periodic solutions of conservative systems
with application to the 3-Body Problem. Internat. J. Bifur. Chaos Appl. Sci.
Engrg., 13:13531381, 2003.
11. E. J. Doedel, V. A. Romanov, R. Paenroth, H. Keller, D. Dichmann, J. Gal
anVioque, and A. Vanderbauwhede. Elemental periodic orbits associated with the
Libration Points in the Circular restricted three body problem. Internat. J.
Bifur. Chaos Appl. Sci. Engrg., in press, 2007.
12. J. Gal
an, F. J. Mu
noz-Almaraz, E. Freire, E. J. Doedel, and A. Vanderbauwhede. Stability and bifurcations of the gure-8 solution of the three-body
problem. Phys. Rev. Lett., 88:241101, 2002.
13. M. C. Gutzwiller. Chaos in Classical and Quantum Mechanics. (Springer
Verlag, New York, 1990).
14. C. Marchal. The family P12 of the three-body problem the simplest family of
periodic orbits, with twelve symmetries per period. Cel. Mech. Dynam. Astron.,
78:279298, 2000.
15. K. R. Meyer. Periodic solutions of the N -body problem. J. Di. Eqns., 39:238,
1981.
16. K. R. Meyer. Continuation of periodic solutions in three dimensions. Physica
D, 112:310318, 1998.
17. K. R. Meyer. Periodic Solutions of the N -Body Problem, Lecture Notes in Mathematics, 1719. (Springer-Verlag, Berlin, 1999).
18. K. R. Meyer and G. R. Hall. Introduction to Hamiltonian Dynamical Systems
and the N -Body Problem. Applied Mathematical Sciences, 90. (Springer-Verlag,
New York, 1992).
19. M. Molina, J. Gal
an-Vioque, and E. Freire. Generalized Hamiltonian equations
of motion for nonholonomic systems. (Preprint, University of Sevilla, 2007).
20. C. Moore. Braids in classical gravity. Phys. Rev. Lett., 70(24):36753679, 1993.
21. F. R. Moulton. Periodic Orbits. (Carnegie Institutiton of Washington, 1920).

22. F. J. Mu
noz-Almaraz. Continuaci
on y Bifurcaciones de Orbitas
Peri
odicas en
Sistemas Hamiltonianos con Simetra, Ph.D. Thesis, (Universidad de Sevilla,
2003).
23. F. J. Mu
noz-Almaraz, E. Freire, J. Gal
an, E. J. Doedel, and A. Vanderbauwhede. Continuation of periodic orbits in conservative and Hamiltonian
systems. Physica D, 181:138, 2003.
24. F. J. Mu
noz-Almaraz, E. Freire, J. Gal
an, and A. Vanderbauwhede. Continuation of normal doubly symmetric orbits in conservative reversible systems. Cel.
Mech. and Dyn. Astr., 97:1747, 2007.
25. D. On. Instability for symmetric periodic solutions of the planar three body
problem. Technical report, 2001.
26. H. Poincare. Les Methodes Nouvelles de la Mecanique Celeste. (GauthierVillars, 1892).
27. C. Pugh and C. Robinson. The C 1 closing lemma, including Hamiltonians.
Ergod. Th. Dynam. Sys., 3:261313, 1983.
28. J. Sepulchre and R. MacKay. Localized oscilations in conservative networks of
weakly coupled autonomous oscillators. Nonlinearity, 10:679713, 1997.
29. C. Siegel and J. K. Moser. Lectures on Celestial Mechanics. (Springer-Verlag,
Berlin, 1971).
30. C. Sim
o. In Cent ans apr`es les methodes nouvelles de H. Poincare, pages 123.
(Societe Mathematique de France, 1996).

9 Continuation in Hamiltonian Systems

299

31. C. Sim
o. New families of solutions in n-body problems. In Proceedings ECM
2000, Barcelona, 2000.
32. C. Sim
o. Dynamical properties of the gure-eight solution of the three-body
problem. Contemp. Math., 292:209228, 2002.
33. E. Str
omgren. Connaissance actuelle des orbites dans le probl`eme des trois
corps. Bull. Astron. Obs. Copenhagen, 9(100):87130, 1935.
34. A. Vanderbauwhede. Families of periodic orbits for autonomous systems. In Dynamical Systems II , A. Bednarek and L. Cesari (Eds.), pages 42744 (Academic
Press, 1982).
35. E. T. Whittaker. A Treatise on the Analytical Dynamics of Particles & Rigid
Bodies. (Cambridge University Press, 1947).
36. C. Wul and A. Schebesch. Numerical continuation of Hamiltonian relative
periodic orbits. Submitted for publication. http://www.maths.surrey.ac.uk/
personal/st/C.Wulff/publicationsframes.html.
37. J. A. Zura. Symmetry breaking of water waves, Ph.D. Thesis, (Applied Mathematics, Cal. Inst. Technology, Pasadena, CA, 1987).

10
Phase Conditions, Symmetries and PDE
Continuation
Wolf-J
urgen Beyn and Vera Th
ummler
Department of Mathematics, Bielefeld University, Germany

There is a long tradition of making use of continuous symmetries for the analysis of dierential equations; see, for example, the monographs [10, 21, 32]. In
general, such symmetries are expressed as the equivariance of the dierential
operator with respect to the action of a Lie group. Solutions of the dierential
equation then come in group orbits, and this has interesting consequences,
for example, it may lead to inherent symmetries of solutions or symmetrybreaking bifurcations. In the theory of equivariant systems one usually tries
to reduce the dierential equation to the so-called orbit space, the elements of
which are equivalence classes created by applying the group action to a single
point in phase space. After factoring out the group action in this way, one applies specic results on existence and uniqueness of solutions, on bifurcations,
or on asymptotic stability.
Contrary to the situation in the theory, the use of continuous equivariances
for ecient numerical computations seems to be rather rare. An early exception is Eusebius Doedels integral phase condition [13] for the computation of
periodic orbits in autonomous ODEs; see also Chap. 1. It is a typical example that shows how a judicious use of symmetry (in this case, equivariance
with respect to time shifts) can enhance rather than hamper the eciency of
a numerical method. Namely, less eort is needed for mesh adaptation and
larger continuation steps are possible. The ODE example also shows another
paradigm of numerical bifurcation analysis. While theory prefers to reduce
problems, e.g., by Lyapunov-Schmidt or center manifold reduction, it seems
advantageous rather to extend the problem for numerical purposes (e.g., by
choosing unfolding parameters) and then add extra constraints (e.g., normalizing conditions for eigenvectors). In this way one can keep as much structure
as possible from the original problem and simultaneously use the normalizing
conditions to optimize the conditioning of the extended problem.
In this chapter we discuss the usefulness of phase conditions for the numerical analysis of nite- and innite-dimensional dynamical systems that have
continuous symmetries. Our main topic is the general approach known as the
freezing method, which was developed in [33] and [7]. It will be presented in

302

Wolf-J
urgen Beyn and Vera Th
ummler

an abstract framework for evolution equations that are equivariant with respect to the action of a (not necessarily compact) Lie group. Specically, we
introduce an extra parameter (an element in the associated Lie algebra) that
determines the position on the group orbit and impose further constraints or
phase conditions such that the point in phase space (e.g., the spatial prole
in case of a PDE) varies as little as possible. We show particular applications
of phase conditions to periodic, heteroclinic and homoclinic orbits in ODEs,
to relative equilibria and relative periodic orbits in PDEs, as well as to time
integration of equivariant PDEs.
After reviewing phase conditions that eliminate the time shift in ODEs
in Sect. 10.1, we set up in Sect. 10.2 the general freezing method within an
abstract framework. We then apply our method to the computation of various
spatio-temporal patterns, such as traveling and modulated waves in one, spiral
waves in two, and scroll waves in three space dimensions. For problems in one
space dimension we also investigate asymptotic stability and discuss the errors
introduced by nite boundary conditions.

10.1 Phase Conditions for Orbits in ODEs

Consider a dynamical system generated by an autonomous n-dimensional ordinary dierential equation
ut = f (u),

u(t) Rn ,

f : Rn Rn smooth.

(10.1)

Due to its autonomous character the nonlinear dierential operator

Lu = ut f (u)
has a simple equivariance with respect to time shifts. That is, for all R
and for all u in some function space we have
[Lu]( ) = L[u( )].

(10.2)

Depending on the application, appropriate function spaces may be chosen, such as the Sobolev space H1 (R, Rn ), the space of bounded uniformly
1
(R, Rn ) or the space of one-periodic functions
continuous C 1 -functions Cunif
1
n
Cper (R, R ).
10.1.1 Periodic Orbits
In order to determine a periodic orbit of (10.1) we should nd a period T > 0
and a solution u(t) of the boundary value problem
ut = f (u), t [0, T ], u(0) = u(T ).

10 Phase Conditions, Symmetries and PDE Continuation

303

Introducing the scaled function v(t) = u(tT ), t [0, 1] the boundary value
problem for (v, T ) C 1 ([0, 1], Rn ) R now reads
vt = T f (v), t [0, 1],

v(0) = v(1).

(10.3)

Due to equivariance (10.2) the solutions of (10.3) are only determined up to a

phase shift and a further condition is needed to make the solution unique. In
the rst publication on the continuation package Auto [13] Doedel suggested
to use an anchor equation (as it was called in [13]) that tries to minimize the
1
(R, Rn ), i.e., tries to minimize
L2 distance to some template function v Cper

(v, ) =
0

||v(t ) v(t)||22 dt = ||v( ) v||2L2 .

(10.4)

By dierentiating with respect to , a necessary condition for a local minimum

is

1

(v(t ) v(t))T vt (t) dt = 0.

(10.5)

A more formal statement is contained in the following lemma; see [4] for a
proof.
1
Lemma 1. Suppose that v Cper
(R, Rn ) is a nonconstant 1-periodic function.
Then there exist neighborhoods U of v in the C 1 -topology and R of 0 such
that for any v U the L2 -distance from (10.4) has a unique minimum at
v ) = 0 and
= (v) where : U V is a C 1 -mapping satisfying (
condition (10.5).

During computations one selects v such that the phase condition (10.5) holds
at = 0, i.e.,
 1
(v(t) v(t))T vt (t) dt = 0.
(10.6)
0

This condition has several advantages over a Poincare-type condition such as

(v(0) v(0))T f (
v (0)) = 0.

(10.7)

If v is a good approximation of v obtained from continuation along a branch,

then condition (10.6) tries to keep a steep front or a peak of the solution in
the same place. Usually this facilitates mesh adaptation and simultaneously
allows for larger step sizes along branches. This phase condition is now built
into standard continuation packages, such as Auto (with HomCont) [17],
Content [25] and Matcont [12]; see also Chap. 2. It has proved to be most
reliable in many applications.
For an illustration we take the model example from [13], namely



(1 )u1 u2
u1
=
.
u2 t
u1 + u21

304

(a)

Wolf-J
urgen Beyn and Vera Th
ummler
0.6

0.6

(b)

0.4

0.4

0.2

0.2

u2

u2

0.2

0.2

0.4

0.4

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

Fig. 10.1. Eect of an integral phase condition (a) and of a point phase condition (b).

This system is, in fact, Hamiltonian at = 1 and a continuous family of

periodic orbits bifurcating from the origin and ending in a homoclinic orbit
exists. For a detailed treatment of the Hamiltonian case we refer to Chap. 9.
Figure 10.1 shows the result of continuing the periodic orbits with both phase
conditions (10.6) and (10.7).
We nally mention that conditions (10.6) and (10.7) are special cases of
the general form (v) = 0, where : C 1 ([0, 1], Rn ) R is a C 1 -mapping.
Following [4, 16] one can characterize the admissible phase conditions that
lead to a regular solution (v, T ) C 1 ([0, 1], Rn ) R of the operator equation
F (v, T ) = (vt T f (v), v(0) v(1), (v)) = 0

(10.8)

as follows. Let u(t) be a T -periodic solution of (10.1) such that v(t) = u(tT )
satises (v) = 0. Then the pair (v, T ) is a regular solution of (10.8) if, and
only if, 1 is a simple Floquet multiplier and D(v)vt = 0 where D denotes
the Frechet derivative of . An easy calculation shows, that the latter condiv (0)) = 0 for (10.7). One
tion requires vt , vt L2 = 0 for (10.6) and vt (0)T f (
may call (10.8) a dening equation for an isolated periodic orbit. The recent
paper [15] provides a considerable extension of this general approach to dening equations for all codimension-one bifurcations of periodic orbits, namely:
fold (saddle-node), ip (period-doubling) and Neimark-Sacker bifurcations;
see also [26] and Chap. 2.
10.1.2 Homoclinic and Heteroclinic Orbits
It is natural to extend the numerical methods for periodic orbits to orbits
that connect stationary points in innite time. Such orbits typically occur in
parametrized systems
ut = f (u, ),

u(t) Rn ,

f : Rn Rp Rn smooth.

(10.9)

10 Phase Conditions, Symmetries and PDE Continuation

305

Connecting orbits have numerous applications, in particular, they appear as

traveling waves of PDEs; see Sect. 10.2.1.
C 1 (R, Rn ) Rp is called a connecting orbit pair
Denition 1. A pair (
u, )
and if the limits
of system (10.9) if u
is a solution at =
lim u
(t) = u
+ ,

lim u
(t) = u

(10.10)

exist. The connecting orbit is called homoclinic if u

+ = u
and heteroclinic
otherwise.
= 0, and
From (10.10) we infer that u
are stationary points, i.e., f (
u , )
1
n
1
that u
Cb (R, R ), i.e., C and bounded on R. Nondegeneracy of a connecting
orbit may be dened as follows; cf. [3].
is called nondegenerate if the
Denition 2. A connecting orbit pair (
u, )
following conditions hold:
Rn,n are hyperbolic with ns eigenvalues of
(i) The matrices fu (
u , )
negative real part and nu = n ns eigenvalues of positive real part;
(ii) p = ns n+s + 1;
(iii) If u Cb1 (R, Rn ), Rp satises the variational equation
+ f (

u, )u
u, ),
then = 0 and u = c
ut for some c R.
ut = fu (
Conditions (i) and (ii) ensure that the dimension nu +p of the center-unstable
in the extended phase space Rn Rp and the dimension
manifold of (
u , )
add up to n+p+1, which is one
u+ , )
n+s +p of the center-stable manifold of (
plus the dimension of the extended phase space Rn Rp . Condition (iii) then
guarantees that these two manifolds intersect transversely in the connecting
: t R}. Similar to the periodic case, one can characterize
orbit {(
u(t), )
connecting orbit pairs as regular solutions of an operator equation
F (u, ) = (ut f (u, ), (u, )) = 0,

(10.11)

where a smooth map : Cb1 (R, Rn ) Rp R denes the phase condition; see
[3] for a proof.
be a connecting orbit pair satisfying (
= 0
Proposition 1. Let (
u, )
u, )

and conditions (i) and (ii) of Denition 2. Then (

u, ) is a regular solution
ut = 0.
of (10.11) if, and only if, the orbit pair is nondegenerate and u (
u, )
The analogue of the functional (10.4) to be minimized is

(u, ) =
||u(t ) u
(t)||22 dt = ||u( ) u
||2L2 = ||u u
( + )||2L2 ,

Cb2 (R, Rn ) is a template

where we take u u
+ H1 and assume that u
function that satises u
t H1 . Then the phase condition is again obtained
from the necessary condition of a minimum u u
, u
t L2 = 0. In applications it

306

Wolf-J
urgen Beyn and Vera Th
ummler

may be unrealistic to assume that such a template function is known, because

such that
this essentially requires one to know u
beforehand and to choose u
will depend on and be determined by
||
u(t)
u || = O(e|t| ). In general, u
f (
u , ) = 0. Therefore, Doedel and Friedman [14, 20] suggested to minimize
t ( )||L2 , which leads to the phase condition
||ut u
ut u
t , u
tt L2 = 0.

(10.12)

There are several ways to solve the boundary value problem (10.11) on the
innite line. One may discretize it by using globally-dened Galerkin functions
or transform the domain R to a bounded interval and then devise methods
that handle the resulting singularities; see [27, 30, 31]. Perhaps the simplest
method that allows one to employ existing boundary value solvers is to approximate (10.11) by a nite boundary value problem on some large interval
J = [T , T+ ]. This approach was proposed and analyzed in [3, 14, 20] and
implemented in the HomCont part of Auto[17].
For u C 1 (J, Rn ), Rp we consider the nite boundary value problem
FJ (u, ) = (ut f (u, ), B(u(T ), u(T+ ), ), J (u, )) = 0,

(10.13)

where the smooth maps B : R2n+p Rn+p1 ,(u , u+ , )  B(u , u+ , )

and J : C 1 (J, Rn ) Rp R determine the boundary condition and the
approximate phase condition, respectively. The error introduced by this approximation can be estimated as follows; see [3, 20, 40].
be a nondegenerate connecting orbit pair of (10.9)
Theorem 1. Let (
u, )
such that
= 0 and the matrix
(i) B(
u , u
+ , )
1
.
B
+u R(n+p1)(n+p1)
s B (
+ , )X
u , u
+ , )X
u+ u , u
u (

is nonsingular, where the columns of Xs Rnns and X+u Rnn+u

and of the unstable subform a basis of the stable subspace of fu (
u , )

space of fu (
u+ , ), respectively;
= 0, J (
0 as J R, the derivatives DJ are equicontin(ii) (
u, )
u|J , )
and |DJ (
u | > 0
uous in a uniform neighborhood of (
u|J , )
u|J , )
|J
for some > 0.
Then there exist constants , K > 0 and an interval J0 R with the following
properties. For all J0 J the boundary value problem (10.13) has a unique
Furthermore,
solution (uJ , J ) in a C 1 -ball of radius and center (
u|J , ).
= u
( J ) satises
there is a unique phase shift J near zero such that u
J (
u|J , J ) = 0 and the following estimate holds
J || C||B(

||
u|J uJ ||C 1 + ||
u(T ), u
(T+ ), )||.

(10.14)

10 Phase Conditions, Symmetries and PDE Continuation

307

In view of (10.12) and (10.13) it is natural to take the phase conditions

J (u, ) = u u
, u
t L2 (J)

or J (u, ) = ut u
t , u
tt L2 (J) .

The most natural choice for boundary conditions are so-called projection
boundary conditions that force the end points u(T ), u(T+ ) to lie in the tan+ .
gent spaces of the unstable manifold at u
and of the stable manifold at u
These conditions may be written as
T

Ys ()(u u ())
,
(10.15)
B(u , u+ , ) =
T
()(u+ u+ ())
Y+u
where f (u (), ) = 0 and the columns of Ys () Rnns and Y+u ()
Rnn+u form a basis of the stable subspace of fuT (u (), ) and of the unstable
subspace of fuT (u+ (), ), respectively. Note that, by Denition 2, (10.15)
imposes ns +n+u = n+p1 boundary conditions. Methods to compute these
matrices such that they depend smoothly on the parameter were proposed
in [3] and, more recently, via a smooth block Schur decomposition in [11].
For numerous computations that apply this approach to specic examples we
refer to [3, 14, 17, 20].
We nally notice that projection boundary conditions imply exponential
decay of the term on the right-hand side of (10.14). More precisely, we have
J || = O(e2 T + e2+ T+ ),
||
u|J uJ ||C 1 + ||
where 0 < < Re() for all eigenvalues of fu (
u , ) with positive real
with negative real
u+ , )
part and Re() < + < 0 for all eigenvalues of fu (
part. For the parameter a superconvergence behavior was observed in [3] and
a corresponding estimate proved in [34], namely:
J || = O(e(2 ++ )T + e(2+ + )T+ ).
||

10.2 Phase Conditions and Equivariant PDEs

In this section we consider time-dependent PDEs that have continuous symmetries in the spatial operator. Therefore, we will be concerned with phase
conditions that act on the spatial variables of the solutions. First, we introduce
the method of freezing that employs phase conditions in order to decompose
a time-dependent solution into a time-dependent group orbit and a spatial
prole that varies as little as possible. Second, this method will be used to
compute relative equilibria, i.e., spatial proles of which the group orbits are
invariant under the PDE ow. The underlying general approach was developed
independently in [33] and in [7].

308

Wolf-J
urgen Beyn and Vera Th
ummler

10.2.1 Traveling Waves

A special class of relative equilibria in one space dimension are traveling wave
of parabolic PDEs
solutions u(x, t) = v(x t)
ut = Auxx + f (u),

u(, 0) = u0 ,

x R, u(x, t) Rm ,

(10.16)

where A Rmm is positive denite, v denotes the prole of the wave and
R its velocity.

These solutions are stationary in the moving coordinate system which is

i.e., v and
solve
obtained via the transformation v(, t) = u(x, t), = x t,
vx .
v ) +
0 = A
vxx + f (

(10.17)

Given a stationary solution v, each shifted version v = v( ) is also a

solution of (10.17). As in Sect. 10.1 we add a phase condition dened by some
functional in order to obtain a well-posed boundary value problem for (v, ),
namely:
0 = Avxx + f (v) + vx ,
0 = (v, vx , ).
The natural choice for stems from the phase condition discussed in Sect. 10.1.
One minimizes the H1 -distance or the L2 -distance to a template function v.
This leads to the functional (v) = 
vx , v vH1 or
(v) = 
vx , v vL2 .

(10.18)

Transforming to a rst-order system, we can apply the results from Sect. 10.1
for studying well-posedness (Proposition 1) and approximation (Theorem 1).
In our next step we are going to use phase conditions also for the nonstationary case. Now we let the transformation into the moving frame depend
on time in the following way
u(x, t) = v(x (t), t),

(10.19)

where (0) = 0 and we dene (t) = (t).

In this setting, (10.16) together with

the phase condition transforms into a partial dierential algebraic equation
(PDAE) for (v, ), namely:
vt = Avxx + f (v) + vx ,
0 = (v, vx , ).

v(, 0) = u0 ,

(10.20)

Note that the initial value (0) is not prescribed but, as usual with DAEs,
is determined by dierentiating the constraint = 0 with respect to time
and using the dierential equation. In Sect. 10.2.3 we will discuss possible
choices for the phase condition that lead to PDAEs of dierent index. System
(10.20) can be completed by the simple ODE = (t), (0) = 0 (called the
now appears as a
reconstruction equation in [33]). The traveling wave (
v , )
stationary solution of system (10.20) and, in case of stability, we expect the
during time evolution; see Sect. 10.3.
solution of (10.20) to converge to (
v , )

10 Phase Conditions, Symmetries and PDE Continuation

(a)

309

(b)

Fig. 10.2. Calculation of a wave in the Nagumo equation (10.21): traveling wave
(a) and frozen wave (b).

(a)

(b)
2.5
2

1.5
1
0.5
0
0

10

20

30

40

Fig. 10.3. Calculation of a wave in the Nagumo equation (10.21): (t, x)-plot of the
frozen wave (a), and time evolution of (b).

Example 1. The standard toy example of a traveling wave is a heteroclinic

orbit between two metastable states in the Nagumo equation [23]
ut = uxx + u(1 u)(u a),

u(x, t) R, x R, t > 0,

(10.21)

where a (0, 12 ). An explicit traveling wave connecting the stationary points

u = 0, u+ = 1 is
11
.

= 2 ( 1 a).
(10.22)
,
v(x) = 1 + e 2
2

In Fig. 10.2 we show the results of a numerical computation for a = 0.25 with
nite dierences in space (x = 0.1) and the implicit Euler method in time
(t = 0.1). Panel (a) is for the non-frozen system (10.21) and panel (b) for
the frozen system (10.20). In both cases the spatial interval is J = [30, 30]
and we use Dirichlet boundary conditions. Similar to Sect. 10.1, the frozen
system has the advantage that steep gradients stay in approximately the same
place and the front does not leave the computational domain in nite time. In

310

Wolf-J
urgen Beyn and Vera Th
ummler

(b)

(a)

(c)

(d)

u
v

0.8
0.6

0.95

u, v

0.9

0.4

0.85

0.2

0.8

0
10

10

20

0.75
0

20

40

60

80

100

Fig. 10.4. Calculation of a wave in the autocatalytic system (10.23): traveling wave
(a), frozen wave (b), u- and v-components at t = 100 (c), and time evolution of
(d).

Fig. 10.3 we show another representation of the frozen wave as a color-coded

(t, x)-plot in panel (a), while panel (b) is the time evolution of the velocity .
This type of gure will be used throughout this chapter.
Example 2. We consider an autocatalytic system [1, 28] as given by
ut = auxx uf (v),
vt = vxx + uf (v),

a > 0, u, v : R R,

(10.23)

where f (v) = v m for v 0 and zero otherwise. This system has traveling
wave solutions if the parameter m 2 is not too large. As in [1, 28] we choose
limit values (u , v ) = (0, 1), (u+ , v+ ) = (1, 0) in order to eliminate a scaling
invariance.
Figure 10.4(a) and (b) show the solution of the original and of the frozen
system, respectively, in an interval of length 100 for the original system and of
length 30 for the frozen system. Here a = 0.1, m = 2, and we use the CrankNicholson method (x = 0.1, t = 0.1) and Dirichlet boundary conditions.
Figure 10.4(c) and (d) show the u- and v-components of the frozen system

10 Phase Conditions, Symmetries and PDE Continuation

311

and the time evolution of . Again, the example shows how the method of
freezing allows one to observe phenomena that become visible only after a
transient phase, while in a direct numerical simulation the solution may leave
the nite domain before the steady prole appears.
10.2.2 Freezing Solutions of Equivariant PDEs
Let M be a manifold modeled over some Banach space X and let N be
a submanifold modeled over some dense subspace Y X [9]. Consider an
evolution equation
ut = F (u), u(0) = u0 ,
(10.24)
for a vector eld F : N T M where T M denotes the tangent bundle of
M . We assume that (10.24) is equivariant with respect to a nite-dimensional
(possibly noncompact) Lie group G acting on M via
a : G M M, (, v)  a(, v),
with the property
a(1 2 , v) = a(1 , a(2 , v)),

a(1, v) = v,

1 = unit element in G.

By equivariance we mean that the following relation holds:

a(, N ) N G,
F (a(, u)) = T a(, u)F (u),

u N, G,

where T a : G T M T M denotes the tangent action of a. We assume that

the linear map
T a(, v) : Tv M Ta(,v) M,

w  T a(, v)w

is a homeomorphism for each v M . (Note that a(g, ) corresponds to g :

N N in [29, 33] and T a(g, ) corresponds to g : T M T M .) Furthermore,
we assume that for any v M the map
a(, v) : G M,

 a(, v)

is continuous and that it is continuously dierentiable for any v N with

derivative denoted by
da(, v) : T G Ta()v M,

 da(, v).

For the construction of some spaces that satisfy this smoothness requirement
we refer to [7]. Finally, we denote by L : G G, g  g the multiplication
by G from the left and by dL (g) : Tg G Tg G its derivative. Then we
dene the exponential exp(t) for in the Lie algebra T1 G as the solution of

312

Wolf-J
urgen Beyn and Vera Th
ummler

= dL (1).
The evolution of (t) describes the motion on the group. Other equivalent
denitions of exp are in common use [9, 10, 29].
Generalizing ansatz (10.19) to u(t) = a((t), v(t)), (10.24) can be transformed into a system for the unknowns v(t) M , (t) G, (t) T1 G as
follows (cf. [7, 33]):
vt = F (v) da(1, v),
= dL (1),

v(0) = u0 ,

(10.25a)

(0) = 1.

(10.25b)

Lemma 2. For some T > 0 let u C 1 ((0, T ], M ) C([0, T ], N ) be a solution

of (10.24) and let C 1 ([0, T ], G) be arbitrary with (0) = 1. Then v(t)
dened by u(t) = a((t), v(t)) and (t) dened by (10.25b) are solutions of
(10.25a). Conversely, assume that v C 1 ((0, T ], M ) C([0, T ], N ) and
C 1 ([0, T ], T1 G) solve (10.25a) and dene C 1 ([0, T ], G) as the solution of
(10.25b). Then u(t) = a((t), v(t)) solves (10.24) on [0, T ].
Proof. Insert the ansatz u(t) = a((t), v(t)) into (10.24) and use equivariance
to obtain
da(, v) + T a(, v)vt = ut = F (u) = F (a(, v)) = T a(, v)F (v).

(10.26)

Dierentiating the relation a(, a(g, v)) = a( g, v), g, G, v N with

respect to g at g = 1 leads to
T a(, v)da(1, v) = da(, v)dL (1),

T1 G.

(10.27)

Finally, dene (t) by (t)

= dL (1) and combine (10.26), (10.27) to nd

T a(, v) [vt F (v) + da(1, v)] = 0
and, hence, (10.25a) by the invertibility of T a(, v). The converse is proved
in a similar way. 

Lemma 2 shows that system (10.25) does not have a unique solution (v, , ).
Rather we have p = dim G additional degrees of freedom that will be xed
by a phase condition : N T1 G Rp . The phase condition together with
(10.25a) yields a PDAE for v and , namely:
vt = F (v) da(1, v),
0 = (v, ).

(10.28)

Equation (10.25b) is called the reconstruction equation in [33]. It is decoupled

from system (10.28) and can be solved by an a-posteriori process.
The traveling waves in Examples 1 and 2 easily t into the abstract framework.

10 Phase Conditions, Symmetries and PDE Continuation

313

Example 3. For the Lie group G = R consider the shift action a(, u)(x) =
u(x ). There are dierent possibilities for the choice of spaces M and N .
2
or M = w + L2 N = w + H2 where
Either take M = Cunif , N = Cunif
w Cb2 (R, R2 ) satises wx , wxx L2 and has the correct limit behavior, e.g.,
w(x) = u
+ O(e|x| ) as x . For the last choice we actually use the
manifold structure of M and N . In both cases we have da(1, v) = vx and
using a template function v N the system (10.28) is given by (10.20) with
given in (10.18).
Example 4. Consider a system (10.16) of dimension m = 2 such that the
nonlinearity is equivariant with respect to rotations, i.e.,
1
.
sin
f (R v) = R f (v) v R2 , R, where R = cos
sin cos .
Equations of this type arise as real versions of complex-valued systems, such
as the Ginzburg-Landau equation. The Lie group is G = S1 R and the action
a : G L2 L2 on u : R R2 at = (, ) is given by
a(, u)(x) = R u(x ).
With M = L2 , N = H2 we obtain da(1, v)( , ) = vx R 2 v and
(10.28) has the form

vt = Avxx + vx + R 2 v + f (v),

0 = 
v  , v vL2 ,

0 = R 2 v, v vL2 .

The reconstruction equations read = , (0) = 0 and = , (0) = 0.

10.2.3 Fixed Versus Minimizing Phase Conditions
In the abstract setting of Sect. 10.2.2 assume that M is a Banach space in
which we have a continuous inner product , 2 with associated norm ||v||2 .
One way to set up a phase condition (in the spirit of Sect. 10.1) is to minimize
the distance of the frozen solution v from the group orbit O(
v ) = {a(, v) :
G} of a template function v; see Fig. 10.5(a). The necessary condition for
a minimum of a(, v) v2 to occur at = 1 is
x (v) = da(1, v), v v2 = 0 T1 G.

(10.29)

In the beginning one may choose as template the initial value v = u0 . Note
that in (10.29) maps into the dual T1 G of the Lie algebra, which is isomorphic to Rp .
Another possibility is to minimize the temporal change vt 2 at each time
instance which leads to the condition
min (v) = da(1, v), vt 2 = 0 T1 G.

(10.30)

Wolf-J
urgen Beyn and Vera Th
ummler
v)
Tv O(

O(
v(
t3 )

O(v (t

2 ))

(b)

(t
2

(a)

vt

314

v(t2 )

vt (t1 )

vt (t3 )
v(
t3 )

O(v (t1 ))

O(
v)

v(t1 )

Fig. 10.5. Conditions x (a) and min (b).

As is illustrated in Fig. 10.5(b) this condition requires the frozen trajectory

v(t) to be orthogonal to the group orbit of v(t) at all times.
For the case of traveling waves we show how to transform solutions of the
PDAE for both phase conditions into each other, i.e., we transform solutions
(v, ) of
vt = vxx + f (v) + vx
(10.31)
0 = 
vx , v vL2
into solutions (w, ) of
wt = wxx + f (w) + wx

(10.32)

0 = wx , wt L2 .

The following Lemma will be used for the stability analysis in Sect. 10.3.3.
Lemma 3. Let (v, ) be a solution of (10.31). Then (w, ) dened by
w(x, t) = v(x (t), t),
and


(t) =
0

= ,

vx (, ), vt (, )L2
d
vx (, )2L2

is a solution of system (10.32).

Proof. We have
wt = vt vx = vxx + f (v) + ( )v
x = wxx + f (w) + wx .
With the shift invariance of , L2 and the denition of we get
wt , wx L2 = vt v
x , vx L2 = 0.




10 Phase Conditions, Symmetries and PDE Continuation

315

10.3 Relative Equilibria and Stability

In this section we study relative equilibria of equivariant evolution equations.
In particular, we consider relative equilibria of parabolic systems in one space
dimension. We use phase conditions to approximate relative equilibria on nite
intervals and study their asymptotic stability in the Lyapunov sense via the
freezing method.
10.3.1 Relative Equilibria
We seek solutions of (10.24) that have the special form u(t) = a((t), v) for
some time-independent function v.
Denition 3. A solution u
of (10.24) is called a relative equilibrium if it
has the form u
(t) = a(
(t), v) for some v N and for some function
C 1 ([0, ), G).
Without loss of generality we can assume that (0) = 1. Usually the whole
group orbit O(
v ) = {a(, v), G} is called a relative equilibrium if it is invariant under the semi-ow; see [10, 29]. We found the equivalent constructive
denition above more convenient from a numerical point of view [7], because
it explicitly includes the orbit (t) on the group. The following lemma shows
the connection between u
, and v.
Lemma 4. Let u
(t) = a(
(t), v) be a relative equilibrium with trivial stabilizer
T1 G such that (
v,
) solve
Sv = { G : a(, v) = v}. Then there exists
0 = F (
v ) da(1, v)

(10.33)

and = dL (1)
, (0) = 1.
Conversely, if (10.33) holds for (
v,
) then u
(t) = a(
(t), v) with =
exp(t
) is a relative equilibrium of (10.24).
v ) = range(da(1, v)), and it is
Proof. The orbit O(
v ) has tangent space Tv O(
v ) = dim Gdim Sv . Hence, the
well known [10, Lemma 4.10.4] that dim Tv O(
stabilizer is trivial if, and only if, da(1, v) is one-to-one. By Lemma 2 we nd
that
(t) = dL (1)1 T1 G is continuous and satises (10.33). Since v is
independent of t and da(1, v) is one-to-one, we obtain that
is independent
of t as well. 

Remark 1. If v has nontrivial stabilizer then one can still write as an exponential in terms of the Lie algebra of the stabilizer and its normalizer; see [10,
Th. 7.2.4].
Choosing&a basis {e1 , . . . , ep } in T1 G we can identify the Lie algebra with Rp
p
via = i=1 i ei . Further, setting S i (v) = da(1, v)ei , we nd from (10.28)
and Lemma 4 the equation to be solved for (
v,
), namely:

316

Wolf-J
urgen Beyn and Vera Th
ummler

0 = F (v) + S(v),

where S(v) =

p
%

i S i (v),

i=1

0 = (v, ).
10.3.2 Approximation of Relative Equilibria on Finite Intervals
We now treat the special case when the evolution equation (10.24) is a
parabolic system of the form (10.16). We assume that the operators S i
are dierential operators S i (v)(x) = S0i v(x) + S1i vx (x) for suitable matrices
S0i , S1i Rmm .
For the numerical computation of relative equilibria of (10.16) we solve a
boundary value problem on a nite interval J = [x , x+ ], namely:
0 = Avxx + S(v) + f (v),
= Bv,
0 = S i (
v )|J , v v|J J ,

x [x , x+ ],

(10.34a)
(10.34b)

i = 1, . . . , p.

(10.34c)

Here v is a template function and B is the two-point boundary operator

Bv = P v(x ) + Q vx (x ) + P+ v(x+ ) + Q+ vx (x+ ),

P , Q R2mm .

The linearization of (10.34a) with respect to v at (

v,
) is given by
u = Auxx +Bux +Cu,

B=

p
%

i S1i , C(x) = f  (
v (x))+

i=1

p
%

i S0i . (10.35)

i=1

If limx v(x) = v and limx vx (x) = 0 then turns for x into

the constant-coecient operator
v = Avxx + Bvx + C v,

C = lim C(x).
x

The main spectral assumptions on are the following:

Hypothesis 1 (spectral condition) The eigenvalue 0 lies in the connected
component of C \ {+ } that contains a right half-plane, where
= {s C : det(2 A + iB + C sI) = 0, for some R}.
Hypothesis 2 (eigenvalue condition) The functions S i (
v ) = da(1, v)ei ,
2
i = 1, . . . , p lie in H , are linearly independent and span the nullspace of
v ), . . . , S p (
v )}. Moreover, the algebraic
: H2 L2 , i.e., ker() = span{S 1 (
and the geometric multiplicity of zero are both equal to p.
Hypothesis 1 guarantees that the quadratic eigenvalue problem associated
with has m stable and m unstable eigenvalues; cf. [6]. In view of condition
(i) of Theorem 1 we consider the determinant (cf. [6])

10 Phase Conditions, Symmetries and PDE Continuation

D = det

P Q

Ys
s
Ys

P+ Q+

Y+u
u
Y+u +

317


,

(10.36)

s
u
, Ys ), (+
, Y+u ) Rmm Rmm solve the quadratic eigenvalue
where (
problems
AY 2 + BY + C Y = 0
s
u
with Re (
) < 0 and Re (+
) > 0. Then we can formulate the determinant condition and a consistency assumption for the boundary conditions.

Hypothesis 3 (boundary conditions) The boundary condition (10.34b)

is satised at the stationary points v , i.e., = P v + P+ v+ and the determinant D dened in (10.36) is nonzero.
As in Sect. 10.1, the boundary conditions have to control the terms that grow
in forward time on the positive axis and in backward time on the negative axis.
These are given by the stable or unstable manifolds of the stationary points.
Note that the determinant condition is satised for Dirichlet, Neumann and
periodic boundary conditions; cf. [6].
For simplicity we rst formulate the theorem for pulses, i.e., we use M =
L2 , N = H2 . In order to generalize this to fronts one needs the additional
condition a(, v) v H2 for all G.
Hypothesis 4 (phase condition) The phase condition is satised by v,
i.e., S(
v ), v vL2 = 0, v v H1 , S(
v ) L2 and the matrix

p
S(
v ), S(
v )L2 =
[S i (
v )](x)T [S j (
v )](x)dx
Rpp
R

i,j=1

is nonsingular.
The following approximation result is an adaptation of Theorem 1 to the
current situation; see [39] for a proof.
Theorem 2 (Approximation of relative equilibria on nite intervals).
Assume Hypotheses 1 4 hold. Then there exist  > 0, T > 0, such that for
min{x , x+ } > T the boundary value problem (10.34) has a unique solution
v|J ,
) = {(v, ) H2 (J, Rm ) Rp : 
v|J vH2 + 

(vJ , J ) in a ball B (
 < }. Further, there exist group elements J G such that v = a(J , v)
satises the following estimate for some > 0
 const e min{x ,x+ } .
vJ v|J H2 + J
A similar version for a full discretization with nite dierences can be found
in [43]. In that case one obtains an error estimate on the grid Jh = {hn, n
n n+ } for the approximate solution (vh , h ), namely:
 const (h2 + eh min{n ,n+ } ),
vh v|Jh H2h + h

(10.37)

where  H2h is the discrete analog of the H2 norm. A similar result holds for
the norm   .

318

Wolf-J
urgen Beyn and Vera Th
ummler

(a)

(b)
2

10

10

err

10

errv

10

10

10

15

10

10
20

10

40

10

60

10

20

10

60

10

80

40

10

80

(a) and errv = vh v|  2 (b).

Fig. 10.6. Approximation errors err = |h |
Jh H
h

Example 5. In the case of the Nagumo equation (10.21) from Example 1 we

can compare the approximation with the exact solution. Figure 10.6 shows the
approximation errors of the traveling wave for Dirichlet boundary conditions.
The grid size h was varied logarithmically from 104 to 101 and the size
of the symmetric interval [T, T ] linearly from 20 to 80. We observe that
to the exact solution v and
given in (10.22)
the convergence of v and
is exponential in T and quadratic in h. This is in good agreement with the
approximation result (10.37).
10.3.3 Stability of Relative Equilibria in One Space Dimension
Stability results for traveling waves on the real line or, more generally, for relative equilibria are well known for parabolic systems [22, 44]. Here the notion
of asymptotic stability with asymptotic phase is used. By the freezing method
this notion is converted into the usual asymptotic (Lyapunov-) stability. We
now present a stability result for relative equilibria in the frozen setting. To
this end, the spectral assumptions 1 and 2 have to be tightened as follows.
Hypothesis 5 The curves + lie in the open left half-plane and zero
is the only eigenvalue with real part greater equal zero.
Theorem 3. Assume Hypotheses 1, 2, 4 and 5 hold. Then there exist , > 0
such that for all u0 v + H1 (R) with ||u0 v||H1 the system
vt = Avxx + f (v) + S(v),

v(, 0) = u0 ,

0 = S(
v ), v vL2
has a unique solution v C 1 ((0, ), v + H1 (R)) C([0, ), v + H1 (R)) and
C([0, ), Rp ). Moreover, this solution satises
||v(, t) v||H1 + (t)
 const et v 0 vH1

t 0.

10 Phase Conditions, Symmetries and PDE Continuation

319

Remark 2. For the case of traveling waves a proof of this theorem can be
found in [41]. The generalization to arbitrary groups is straightforward by the
techniques used for Theorem 2. It is also shown in [41] that one can allow
more general nonlinearities f (v, vx ) of the form
f (u, v) = f1 (u)v + f2 (u),

f1 C 1 (Rm , Rmm ), f2 C 1 (Rm , Rm ),

where f1 , f2 , f1 , f2 are globally Lipschitz. This includes the case of the nonlinearity uux in Burgers equation.
An analogous result for a spatial discretization with nite dierences is
given in [41] for traveling waves and in [42] for general relative equilibria in
one space dimension.
Remark 3. We note that a general stability theorem for nite-dimensional
equivariant systems is given in [10, Th. 7.4.2].
For the special case of stationary solutions of (10.31) the local stability estimate reads
const et v 0 vH1
v(, t) vH1 + |(t) |

t 0.

(10.38)

Using the transformation between the dierent phase conditions x and min
in Lemma 3, we will show how stability transfers to the min -case.
We dene the bilinear form b : H1 H1 R via

b(u, v) =
ux (x)T Avx (x) + u(x)T (Bvx (x) + C(x)v(x)) dx
R

where A, B, C() are the bounded matrix functions dened in (10.35). Via
integration by parts we then get
b(
vx , v) = 
vx , vL2

for v H2

and

|b(
vx , v)| const vH1 .

We dene the projector P onto vx along vx and the projected dierential

operator P through
P v = v vx 
vx , vx 1
vx , vL2 ,
L2 

P = P |range(P ) .

The following lemma gives the main estimate for solutions of the nonautonomous PDAE
vt = v +
vx + g(t, v, ),
0 = 
vx , v.

v(0) = v 0 ,

(10.39)

Lemma 5. Assume that g satises

g(t, v, ) const et (vH1 + ),

> 0.

(10.40)

Then there exist > 0 and (0, ) such that any solution (v, ) of (10.39)
with v 0 H1 < obeys the exponential estimate
v(t)H1 + (t) const et v 0 H1 .

(10.41)

320

Wolf-J
urgen Beyn and Vera Th
ummler

Proof. The proof relies on the estimates for r range(P ) and some > 0
eP t rL2 Ket rL2 ,

eP t rH1 Ket t 2 rL2 ,

(10.42)

which follow from the fact that the eigenvalue 0 has been eliminated from the
spectrum of P ; cf. [41, Lemma 1.24]. By the variation of constants formula
the PDAE (10.39) can be written equivalently as

v(t) = eP t v 0 +
(t) =

eP (ts) P g(s, v(s), (s)) ds,

0
vx , v(t))

vx , vx 1
L2 [b(

(10.43)

+ 
vx , g(t, v(t), (t))L2 ].

Using this form, one rst shows via Gronwall estimates as in [41] a global
bound
(10.44)
v(t)H1 + (t) Cv 0 H1 t 0.
From the second equation in (10.43) we nd with (10.40) that
(t) C[v(t)H1 + et (v(t)H1 + (t))].
Choose T > 0 such that CeT

1
2

and obtain

(t) Cv(t)H1

t T.

(10.45)

Now choose 0 < < min(, ) and use (10.42), (10.44) and (10.45) in the
rst equation of (10.43) to obtain
.
n(t) = v(t)H1 et C e()t v 0 H1
 t ()(ts)
1
e

es es (v(s)H1 + (s)) ds
+
ts
0
 T ()(T s)
.
e

(s) ds
C e()t v 0 H1 + e()(tT )
ts
0
 t ()(ts)
1
e

es v(s)H1 ds
+
ts
0
 t ()(ts)
1
.
e
()t
0

n(s) ds .
C e
v H1 +
ts
0

The Gronwall inequality with weak singularities (cf. [22, Lemma 7.1.1]) yields
the assertion. 

Lemma 6. Let the assumptions of Theorem 3 be satised for a nonconstant
C 2 R of (10.31). Then there exists a shift R such that
solution (
v , )
b

(
v ( + ), ) is an asymptotically stable solution of (10.32).

10 Phase Conditions, Symmetries and PDE Continuation

321

Proof. Note that w = vt solves

wt = wxx + f  (v)w + wx + vx ,
0 = 
vx , w,
and that the rst equation with := is equivalent to
x + (vx vx ).
wt = w + vx + (f  (v) f  (
v ))w + ( )w
Now we apply Lemma 5 for small v 0 H1 with = and
x + (vx (t) vx ).
g(t, w, ) = (f  (v(t)) f  (
v ))w + ((t) )w
Note that the exponential decay (10.40) follows from the stability estimate
(10.38). Since g is linear in w and , we obtain for all v 0 from (10.41) the
estimate
const et v 0 H1 t 0.
vt  + 
By this estimate the integral

=
0

vx (, ), vt (, )
d
vx (, )2

exists and we have



v(, )H1 vt (, )
|vx (, ), vt (, )|
d
d

|(t) |
2
vx (, )2L2
vx (, )
t
t

(
v H1 + C)Ce
d const et .

2
(
v


C)
x
L
t
2

Together with the local stability estimate (10.38) this leads to

w(, t) v( )H1 + |(t) |

|
v( (t), t) v( )H1 + |(t) (t)
+ |(t)|
v v( + (t) )H1 + |(t) |

v(, t) vH1 + 
const et .




10.4 Spiral Waves and Beyond

Embedding spiral waves of parabolic systems in R2 into the abstract framework of Sect. 10.2.2 is a considerable task [36, 46]. Therefore we do not pursue
this in detail here.

322

Wolf-J
urgen Beyn and Vera Th
ummler

10.4.1 Spiral Waves in Two Space Dimensions

Consider a PDE in two space dimensions
ut = u + f (u),
u(x, 0) = u0 (x),

t 0,

(10.46)

x R2 .

This equation is equivariant with respect to the Euclidean group SE(2) = S1 

R2 " (, ) with action a(, v)(x) = v(R (x )) and group multiplication
(1 , 1 ) (2 , 2 ) = (1 + 2 , 1 + R1 2 ),
where R again denotes rotations; cf. Example 4.
2
Take, for example, v Cunif = M and N = Cunif
or the subspace Ceucl
Cunif on which SE(2) acts continuously [46]. Then the innitesimal generators
da(1, v)ei , i = 1, 2, 3 read da(1, v)e1 = x2 vx1 x1 vx2 , da(1, v)e2 = vx1 ,
da(1, v)e3 = vx2 . For a relative equilibrium u(x, t) = v(R (x )) the
wave form v is a solution of
0 = v + f (v) +
1 (x1 vx2 x2 vx1 ) +
2 vx1 + vx2
3
for some
se(2). The motion on the group orbit is given by


1 0
=

,
= dL (1)
0 R
with solution
(t) =

1 t
,
(I R1 t )

where

1
=

3
.

Then we can represent the relative equilibrium as follows:

u
(x, t) = v(R1 t (x + (R1 t I))) = v(R1 t (x ) + ).
A xed reference point x
R2 with value v(
x) traces the curve
x(t) = R1 t (
x ) + ,

(10.47)

which is a circle of radius 

x and center . If one uses a geometric denition
for the tip xtip of a spiral wave (see e.g. [8, 46]) then the tip moves on a circle
of radius rtip = xtip . A special case are rigidly rotating Archimedean
spirals which can be written in polar coordinates as v(x) = w(r, ) with
w(r, ) w (kr + ) as r for some periodic function w .
Example 6. We use a diusive version of Barkleys system [2] as an example,
namely:
ut = u + 1 u(1 u)(u a1 (v + b)),
vt = Dv v + u v.

10 Phase Conditions, Symmetries and PDE Continuation

323

(a)

(b)
2

1
2
3

5
0

50

100

150

Fig. 10.7. Frozen spiral wave of PDE (10.46): time evolution of the u-component
(a) and time evolution of (b).

Fig. 10.8. Prediction of the tip-motion of the spiral wave from Fig. 10.7 via (10.47)
(red circle) and tip-motion of the non-frozen spiral starting from the same initial
condition (white trace).

We solve the corresponding PDAE

ut = u + 1 u(1 u)(u a1 (v + b)) + 1 (yux xuy ) + 2 ux + 3 uy
vt = Dv v + u v + 1 (yvx xvy ) + 2 vx + 3 vy

0 = y u
x x
uy , u u
L2 ,

0 = 
ux , u u
L2 ,

0 = 
uy , u u
L2

0 = y
vx x
vy , v vL2 ,

0 = 
vx , v vL2 ,

0 = 
vy , v vL2

with (
u, v) = (u0 , v 0 ) numerically for the parameters Dv = 0.5, a = 0.5,
1
by using the Finite Element package Comsol MultiphysicsTM .
b = 0.05, # = 50
In Fig. 10.7 the time evolutions of the u-component and the parameter
are displayed. In Fig. 10.8 the prediction of the motion of the tip via (10.47)
(red circle) is compared to the tip motion of the non-frozen spiral starting

324

Wolf-J
urgen Beyn and Vera Th
ummler

(a)

(b)
0

1
2
3
4
5
6

6
0

100

200

300

Fig. 10.9. Three-dimensional scroll wave of (10.48), namely: slices of the development of u (a) and time evolution of (b).

from the same initial conditions (white trace). For the denition of xtip we
used the condition u = 12 , v = a2 b from [2].

10.4.2 A Scroll Wave in Three Space Dimensions

In three space dimensions (10.46) is equivariant with respect to G = SO(3) 
R3 = SE(3) with action a(, v)(x) = v(R1 (x )), = (R, ), =
+ R
). We denote the ro(1 , 2 , 3 ) and group operation = (RR,
tations about the x1 , x2 and x3 axes by Rxi , and nd by dierentiating with
respect to x1 , x2 and x3 the formula
da(1, ) = 1 (vx2 x3 vx3 x2 ) + 2 (vx3 x1 vx1 x3 ) + 3 (vx1 x2 vx2 x1 )
+ 4 vx1 + 5 vx2 + 6 vx3 .
Example 7. We consider the following - system in complex form
ut = u + 1 |u|2 i|u|2 u,

x R3 , u(x, t) C,

(10.48)

for which rigidly rotating waves exist [24].

We use an adapted version of the code ezscroll [18] and start in a box
of length 40 with x = 0.1 from an initial function given in cylindrical coordinates as
iz
r
(cos() + i sin()),
u0 (r, , z) = e 2 40

which ensures that in each z-slice a rotating spiral develops. We use periodic
boundary conditions on the z-faces and Neumann boundary conditions on the
x- and y-faces. Therefore, the initial function initiates a scroll wave twisted
once in the z-direction; see [18, 19] for more information on scroll waves and
scroll rings.
Figure 10.9 shows the real part of the solution of the frozen system at the
nal time instance, as well as the time evolution of . The solution in panel (a)

10 Phase Conditions, Symmetries and PDE Continuation

325

is shown in the form of slices in x-,y-,z-directions through the origin (0, 0, 0)

which have been projected to the boundaries to increase visibility. From panel
(b) one can see that rst the rotation around and the translations along the
z-axis are active. However, after some transient time only the z-translation is
used to freeze the solution.
We expect that this solution has a nontrivial stabilizer since vertical motions in the z-direction and rotations about the z-axis can be exchanged.
Nevertheless, our method seems to work. In fact, system (10.48) is actually
equivariant with respect to the seven-dimensional group S1 SE(3), where
= ei S1 acts as in Example 4. Including this symmetry in the computations leads to ill-conditioned systems when resolving the phase conditions for
the seven parameters.
10.4.3 Relative Periodic Orbits
For relative periodic orbits we have a similar denition as for relative equilibria. We seek solutions of (10.24) that have the special form u(t) = a(
(t), v(t))
for some time periodic function v.
Denition 4. A solution u
of (10.24) is called a relative periodic orbit if it
has the form u
(t) = a(
(t), v(t)) where : R G is a smooth curve satisfying
(0) = 1 and v : R N satises v, vt , F (
v ) C(R, M ), is nonconstant and
time periodic, i.e., v( + T ) = v for some period T > 0.
As for relative equilibria (see Lemma 3) we can relate u
, and v.
Lemma 7. Let u
(t) = a(
(t), v(t)) be a relative periodic orbit with trivial sta C(R, T1 G)
bilizer Sv(t) , t [0, T ]. Then there exists a T -periodic function
such that for all t R
vt = F (
v ) da(1, v)

t = dL (1)
, (0) = 1.

(10.49)
(10.50)

Conversely, if
C(R, T1 G) and v C 1 (R, M ) with v(t) N t R are
T -periodic and solve (10.49) then a(
(t), v) with dened by (10.50) is a
relative periodic orbit of (10.24).
Proof. The proof is similar to that of Lemma 4. For the rst assertion dene

(t) by (10.50) and obtain (10.49) from Lemma 2. Equation (10.49) then
shows that
is T -periodic. The converse follows in a similar manner. 

Scaling with the period T as in Sect. 10.1, we nd that v(t) = v(tT ), (t)
solve
vt = T [F (v) da(1, v)], t [0, 1], v(0) = v(1).

326

Wolf-J
urgen Beyn and Vera Th
ummler

(a)

(c)

(b)

1.5

u
v

0.8
0.6
u, v

(d)
1

0.4

0.5
0.2
0
10

10

20

0
0

200

400

600

800

1000

Fig. 10.10. Calculation of a modulated traveling wave in the autocatalytic system

(10.23) for = 12 : traveling wave (a), frozen wave (b), u- and v-components at
t = 1000 (c), and time evolution of (d).

Example 8. Relative periodic orbits in one space dimension are modulated

t) with v(x, t) =
traveling waves [37]. These have the form u
(x, t) = v(x t,
v(x, t + T ). Such solutions occur, for example, in the autocatalytic system
(10.23) that was already considered in Example 2. We solve the system for
a = 0.1, m = 9 on an interval of length 35 in the frozen case, and of length 300
for the direct simulation with x = 0.1, with Dirichlet boundary conditions
and with a -method with t = 0.1. In Fig. 10.10(a) and (b) the numerical
solutions for = 12 of the PDE (10.23) and of the corresponding PDAE
(dening the frozen system) are shown. Panel (c) shows the solution at the last
time instance, while panel (d) displays the time evolution of . The periodicity
of the wave and the velocity can clearly be seen.
Figure 10.11 shows the modulated traveling wave and the frozen wave when
the implicit Euler method is used, that is, = 1. Note that the oscillations
are strongly damped for the standard simulation in panel (a), whereas they
are clearly visible in the frozen case in panel (b).

Combining the principles from Sects. 10.1 and 10.2 for the computation of
relative periodic orbits, we arrive at a boundary value problem (in space and
time) for v C([0, 1], M ), C([0, 1], T1 G) and T R as follows:

10 Phase Conditions, Symmetries and PDE Continuation

(a)

327

(b)

Fig. 10.11. Calculation of a modulated traveling wave in the autocatalytic system

(10.23) for = 1 (implicit Euler method): traveling wave (a), frozen wave (b).

vt = T [F (v) da(1, v)], v(0) = v(1),

 1
da(1, v), v v dt T1 G,
0 = (v) =


0
1


vt , v v dt.

0 = (v) =
0

Here ,  denotes an inner product on M and v C([0, 1], M ) is a suitable

template function.

10.5 Conclusions and Perspectives

Phase conditions are an eective tool in selecting specic orbits in equivariant
evolution equations. When based on minimization principles they facilitate
mesh adaptation and speed up continuation along branches. In many applications the underlying symmetry is induced by the Euclidean group SE(d)
acting on functions dened on the whole space Rd . For numerical computations one has to truncate to bounded domains and use asymptotic boundary
conditions. Truncation in combination with the method of freezing spatiotemporal patterns in a co-moving frame raises several numerical as well as
theoretical problems. Only a few of them have been tackled in this chapter,
mainly for parabolic systems in one space dimension.
Considerable challenges remain, and we expect the further development of
the eld to address theoretical and numerical issues, including the following:
1. Computation and continuation of relative equilibria and relative periodic orbits in equivariant systems, the detection of bifurcation points and
branch switching at symmetry breaking bifurcations; see, for example, the
recent progress by Wul and Schebesch [45].

328

Wolf-J
urgen Beyn and Vera Th
ummler

2. Adaptation of the freezing method to relative equilibria with nontrivial

stabilizers.
3. Consideration of linear versus nonlinear stability for spatio-temporal patterns in space dimensions 2. There are extensive studies of the spectra
associated with systems linearized about spiral waves and their truncation
to bounded domains (see [35, 36, 38]), but a result on nonlinear stability
still seems to be lacking.
4. Development of (implementable) asymptotic boundary conditions for spiral waves, scroll waves and the like.
5. Application of the freezing method to viscous conservation laws. As for
modulated or spiral waves, this case is dicult because of the fact that
the essential spectrum has a quadratic tangency with the imaginary axis;
see [5].

References
1. N. J. Balmforth, R. V. Craster, and S. J. A. Malham. Unsteady fronts in an
autocatalytic system. Proc. Royal Soc. A, 455(1984):14011433, 1999.
2. D. Barkley. A model for fast computer simulation of waves in excitable media.
Physica D, 49:6170, 1991.
3. W.-J. Beyn. The numerical computation of connecting orbits in dynamical
systems. IMA J. Numer. Anal., 10(3):379405, 1990.
4. W.-J. Beyn. Numerical methods for dynamical systems. In Advances in Numerical Analysis, Vol. I (Lancaster, 1990), Oxford Sci. Publ., pages 175236.
(Oxford University Press, 1991).
5. W.-J. Beyn and J. Lorenz. Stability of viscous proles: proofs via dichotomies.
J. Dynam. Di. Eqns., 18:141195, 2006.
6. W.-J. Beyn and J. Lorenz. Stability of traveling waves: dichotomies and eigenvalue conditions on nite intervals. Numer. Funct. Anal. Optim., 20(3-4):201
244, 1999.
7. W.-J. Beyn and V. Th
ummler. Freezing solutions of equivariant evolution equations. SIAM J. Appl. Dyn. Sys., 3(2):85116, 2004.
8. M. Braune and H. Engel. Compound rotation of spiral waves in a lightsensitive
Belousov-Zhabotinsky medium. Chem. Phys. Lett., 204(3-4):257264, 1993.
9. P. R. Cherno and J. E. Marsden. Properties of Innite Dimensional Hamiltonian Systems. Lecture Notes in Mathematics, 425. (Springer-Verlag, Berlin,
1974).
10. P. Chossat and R. Lauterbach. Methods in Equivariant Bifurcations and Dynamical Systems, Advanced Series in Nonlinear Dynamics 15. (World Scientic
Publishing, Singapore, 2000).
11. J. W. Demmel, L. Dieci, and M. J. Friedman. Computing connecting orbits
via an improved algorithm for continuing invariant subspaces. SIAM J. Sci.
Comput., 22(1):8194 (electronic), 2000.
12. A. Dhooge, W. Govaerts, and Yu. A. Kuznetsov. Matcont: a MATLAB package for numerical bifurcation analysis of ODEs. ACM Trans. Math. Software,
29(2):141164, 2003.

10 Phase Conditions, Symmetries and PDE Continuation

329

13. E. J. Doedel. Auto: a program for the automatic bifurcation analysis of autonomous systems. In Proceedings of the Tenth Manitoba Conference on Numerical Mathematics and Computing, Vol. I (Winnipeg, Man., 1980), volume 30,
pages 265284, 1981.
14. E. J. Doedel and M. J. Friedman. Numerical computation of heteroclinic orbits.
J. Comput. Appl. Math., 26(1-2):155170, 1989.
15. E. J. Doedel, W. Govaerts, and Yu. A. Kuznetsov. Computation of periodic
solution bifurcations in ODEs using bordered systems. SIAM J. Numer. Anal.,
41(2):401435, 2003.
16. E. J. Doedel, A. D. Jepson, and H. B. Keller. Numerical methods for Hopf
bifurcation and continuation of periodic solution paths. In Computing Methods in Applied Sciences and Engineering, VI (Versailles, 1983), pages 127138.
(North-Holland, 1984).
17. E. J. Doedel, R. C. Paenroth, A. R. Champneys, T. F. Fairgrieve,
Yu. A. Kuznetsov, B. E. Oldeman, B. Sandstede, and X. J. Wang. Auto2000:
Continuation and bifurcation software for ordinary dierential equations (with
HomCont). Available via http://cmvl.cs.concordia.ca/.
18. M. Dowle, R. M. Mantel, and D. Barkley. Fast simulations of waves in threedimensional excitable media. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
7(11):25292545, 1997.
19. B. Fiedler and R. M. Mantel. Crossover collision of scroll wave laments. Doc.
Math., 5:695731 (electronic), 2000.
20. M. J. Friedman and E. J. Doedel. Numerical computation and continuation of
invariant manifolds connecting xed points. SIAM J. Numer. Anal., 28(3):789
808, 1991.
21. M. Golubitsky and I. Stewart. The Symmetry Perspective. Progress in Mathematics, 200. (Birkh
auser, 2002).
22. D. Henry. Geometric Theory of Semilinear Parabolic Equations Lecture Notes
in Mathematics, 840. (Springer-Verlag, Berlin, 1981).
23. S. Arimoto J. Nagumo and S. Yoshizawa. An active pulse transmission line
simulating nerve axon. In Proceedings of the IRE, volume 50, pages 20612070,
1962.
24. Y. Kuramoto and S. Koga. Turbulized rotating chemical waves. Progr. Theor.
Phys., 66(3):10811085, 1981.
25. Yu. A. Kuznetsov, V. V. Levitin, and A. R. Skovoroda. Continuation of stationary solutions to evolution problems in Content. Technical Report AM-R961,
Centrum voor Wiskunde en Informatica, Amsterdam, 1996.
26. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory, third edition. Applied Mathematical Sciences, 112. (Springer-Verlag, Berlin, 2004).
27. L. Liu, G. Moore, and R. D. Russell. Computation and continuation of homoclinic and heteroclinic orbits with arclength parameterization. SIAM J. Sci.
Comput., 18(1):6993, 1997.
28. S. J. A. Malham and M. Oliver. Accelerating fronts in autocatalysis. Proc.
Royal Soc. A, 456(1999):16091624, 2000.
29. J. E. Marsden and T. S. Ratiu. Introduction to Mechanics and Symmetry,
volume 17 of Texts in Applied Mathematics. (Springer-Verlag, second edition,
1999).
30. G. Moore. Computation and parametrization of periodic and connecting orbits.
IMA J. Numer. Anal., 15(2):245263, 1995.

330

Wolf-J
urgen Beyn and Vera Th
ummler

31. G. Moore. Laguerre approximation of stable manifolds with application to connecting orbits. Math. Comp., 73(245):211242, 2004.
32. P. J. Olver. Applications of Lie Groups to Dierential Equations, second edition.
Graduate Texts in Mathematics, 107. (Springer-Verlag, Berlin, 1993).
33. C. W. Rowley, I. G. Kevrekidis, J. E. Marsden, and K. Lust. Reduction and reconstruction for self-similar dynamical systems. Nonlinearity, 16(4):12571275,
2003.
34. B. Sandstede. Convergence estimates for the numerical approximation of homoclinic solutions. IMA J. Numer. Anal., 17(3):437462, 1997.
35. B. Sandstede and A. Scheel. Absolute and convective instabilities of waves on
unbounded and large bounded domains. Physica D, 145(3-4):233277, 2000.
36. B. Sandstede and A. Scheel. Absolute versus convective instability of spiral
waves. Phys. Rev. E, 62(6):77087714, 2000.
37. B. Sandstede and A. Scheel. On the structure of spectra of modulated traveling
waves. Math. Nachr., 232:3993, 2001.
38. B. Sandstede and A. Scheel. Curvature eects on spiral spectra: Generation of
point eigenvalues near branch points. Phys. Rev. E, 73(1):016217, 8, 2006.
39. S. Selle. Approximation von relativen Gleichgewichten auf endlichen Intervallen.
Masters thesis, (Dept. of Mathematics, Bielefeld University, 2006).
40. V. Th
ummler. Numerische Stabilit
atsindikatoren f
ur wandernde Wellen. Masters thesis, (Dept. of Mathematics, Bielefeld University, 1998).
41. V. Th
ummler. Numerical Analysis of the Method of Freezing Traveling Waves.
PhD thesis, (Dept. of Mathematics, Bielefeld University, 2005).
42. V. Th
ummler. Asymptotic stability of frozen relative equilibria. Preprint no.
06-30 of the CRC 701, Bielefeld University, 2006.
43. V. Th
ummler. Numerical approximation of relative equilibria for equivariant
PDEs. Preprint no. 05-017 of the CRC 701, Bielefeld University, 2006.
44. A. I. Volpert, V. A. Volpert, and V. A. Volpert. Traveling Wave Solutions of
Parabolic Systems. Translations of mathematical monographs, 140. (American
Mathematical Society, 1994).
45. C. Wul and A. Schebesch. SYMPERCON (continuation for symmetric periodic
orbits). Technical report, University of Surrey, Department of Mathematics and
Statistics, 2006.
46. C. Wul. Transitions from relative equilibria to relative periodic orbits. Doc.
Math., 5:227274, 2000.

11
Numerical Computation of Coherent
Structures
Alan R Champneys1 and Bj
orn Sandstede2
1
2

Department of Engineering Mathematics, University of Bristol, United Kingdom

Department of Mathematics, University of Surrey, United Kingdom

In many applications one is interested in nding solutions to nonlinear evolution equations with a particular spatial and temporal structure. For instance,
solitons in optical bers and wave guides or buckling modes of long structures
can be interpreted as localized traveling or standing waves of an appropriate underlying partial dierential equation (PDE) posed on an unbounded
domain. Spiral waves or other defects in oscillatory media are time-periodic
waves with an asymptotic spatially periodic structure. All of these examples
are referred to as coherent structures. They represent relative equilibria, that
is, their temporal evolution is determined by a symmetry of the underlying
PDE: namely, translational symmetry for traveling waves, rotational symmetry for spiral waves, and phase symmetry for oscillatory structures.
Given the complexity of typical PDE models, these nonlinear waves are
in general accessible only through numerical computations. One possible approach is via direct simulation which is, however, expensive and fails to capture
solutions that are either unstable or may have a small basin of attraction. Simulation therefore often fails to provide valuable information on how branches
of solutions are organized in parameter space.
In this chapter, we give an overview of boundary value problem formulations for coherent structures which provide a robust and less expensive alternative to simulation. Moreover, setting up well-posed boundary value problems
allows us to continue solutions in parameter space, investigate their spectral
stability directly, and continue branches of solutions eciently as parameters
vary.
In the next section we outline how PDEs can be supplemented by phase
conditions that allow us to compute nonlinear waves as regular zeros of the
resulting nonlinear system. In the remaining sections, we treat dierent kinds
of coherent structures, namely traveling waves, time-periodic structures, and
planar localized patterns. In each case we explain how to set up a well-posed

332

Alan R Champneys and Bj

orn Sandstede

boundary value problem and illustrate the theory with the results of an example computation.

11.1 Symmetries and Phase Conditions

It is desirable to formulate the problem of computing nonlinear waves in such
a way that solutions correspond to regular zeros of an appropriate smooth
function. Indeed, Newtons method is then applicable and gives a robust and
ecient way of computing zeros. When computing coherent structures, solutions are often not unique due to the presence of translational or other
symmetries in the underlying partial dierential equation. In other words,
symmetries may give rise to families of solutions. To make Newtons method
work, we need to single out one particular solution. This can be done by adding
extra functionals to the problem and identifying additional parameters that
compensate for the added functional to make the overall problem invertible;
see also Chaps. 9 and 10. In this section, we recall how the extra functionals,
which we refer to as phase conditions, can be constructed and the additional
parameters be identied in systems with symmetries.
Thus, consider the equation
u X,

ut = f (u),

(11.1)

where X is a Banach space. For ease of exposition, we assume that f maps

X into itself and is smooth; we remark that the rst property fails if (11.1)
represents a partial dierential equation, but the results that follow actually
hold for PDEs nevertheless; see [39].
The key assumption is that (11.1) is equivariant under the action of a
nite-dimensional Lie group G on X so that
gf (u) = f (gu),

g G, u X .

We denote by
exp : alg(G) G,

 exp()

the exponential map of the Lie algebra alg(G) of G into G.

Example 1. Take X = C 1 (R, Rn ) and consider the action of G = R dened by
g : X X , u()  u( + g) for all g G = R; then exp = id.
Particularly interesting solutions u(t) of (11.1) are relative equilibria which
are equilibria when the group action is factored out. More precisely, u is a
relative equilibrium if there exists alg(G) so that
u(t) = exp( t)u

t R.

Note that u(t) Gu for all t. We emphasize that u satises

11 Numerical Computation of Coherent Structures

u = f (u ),

333

(11.2)

where we identify an element alg(G) with the generator


d

: X X
exp(t)
=
dt
t=0

of the one-parameter group exp(t).

Example 2 (Example 1 continued). Relative equilibria of the group action g :
X X , u()  u( + g) for g G = R on X = C 1 (R, Rn ) are of the form
u(x, t) = u (x + c0 t) for c0 R. Therefore, they are traveling waves and we
have



d
d



= c0 u ( + c0 t)
= c0 u ,
exp(c0 t)u 
= u ( + c0 t)
u :=
dt
dt
t=0
t=0
t=0

and (11.2) becomes the traveling wave equation c0 u = f (u ).

If u is a relative equilibrium then the entire group orbit Gu consists of
relative equilibria since
gu(t) = g exp( t)g 1 gu = exp(Adg t)gu
is again a relative equilibrium, where
Adg :=


d

alg(G).
g exp(t)g 1 
dt
t=0

If G is Abelian, then Adg is the identity, and the velocity is the same for
all relative equilibria in the group orbit Gu .
Example 3 (Example 1 continued). With u(t) = u (+c0 t), we nd that u (+
g + c0 t) is a traveling wave for each g G = R.
Thus, to compute or continue relative equilibria, we need to solve
f (u) u = 0
for (u, ) X alg(G) and expect to nd a family {(gu , Adg ); g G} of
solutions. To nd (u, ) as a regular zero of an appropriate function, we need
to add phase conditions that single out a relative equilibrium (u , ) among
the family of solutions.
For simplicity, we assume from now on that the Lie group G is Abelian.
The family of relative equilibria is then given by (gu , ) for g G, and we
have f (gu ) gu = 0 for all g G. Choosing g = exp(t) and taking the
derivative with respect to t at t = 0, we obtain
fu (u )u u = 0
for all alg(G). Thus, if we assume that the isotropy group {g G; gu =
u } is discrete, we can conclude that the linearization fu (u ) of (11.2)
has a null space of dimension at least dim G with elements u for alg(G).

334

Alan R Champneys and Bj

orn Sandstede

Gu

u
u + N()
Fig. 11.1. The translated null space u + N() of an admissible phase condition
is transverse to the group orbit Gu at the relative equilibrium u . Imposing the
condition (u u ) = 0 restricts u to u + N() and, therefore, singles out the
relative equilibrium u .

Hypothesis 1 Assume that G is Abelian with dim G = m, and (u , ) is a

relative equilibrium such that fu (u ) is a Fredholm operator with index
zero. Furthermore, assume that fu (u ) has an eigenvalue = 0 with
geometric and algebraic multiplicity m and that its null space is given by
{u }alg(G) .
We choose linear phase conditions of the form
: X Rm ,

u  u

and say that is admissible if the linear operator

alg(G) Rm ,

 u

is invertible. The following theorem is illustrated in Fig. 11.1.

Theorem 1. Assume that Hypothesis 1 holds and that is an admissible
phase condition. Then the relative equilibrium (u , ) is a regular zero of the
map
F : X alg(G) X Rm ,

(u, )  (f (u) u, (u u )).

Proof. Clearly, F(u , ) = 0. The linearization of F about (u , ) is given

by
L : (u, )  ([fu (u ) ]u u , u).
The bordering lemma [3, Lemma 2.3] shows that this operator is Fredholm
with index zero and, therefore, it suces to show that its null space is trivial.
Since the geometric and algebraic multiplicities of the eigenvalue = 0 of
fu (u ) are equal to each other, there is a spectral decomposition
X = X {u ; alg(G)}

11 Numerical Computation of Coherent Structures

335

so that fu (u ) is an isomorphism from X into itself. Writing u = u

+ u ,
we nd
u u , (
u + u )).
L(u, ) = L(
u + u , ) = ([fu (u ) ]
Using again the decomposition of X , we see that u
= = 0, and the remaining
equation u = 0 nally gives = 0 since is admissible. 

Remark 1. Suppose that X is a Hilbert space with scalar product , . Choose
a basis {j }j=1,...,m of alg(G). Then the phase condition
: X Rm ,

u  (j u , u)j=1,...,m

is admissible.
Example 4 (Example 1 continued). We wish to nd a functional : C 1 (R, Rn )
R such that u = 0. The space C 1 (R, Rn ) is not a Hilbert space, but if u
decays exponentially then

u = u (x), u(x) dx
R

is an admissible phase condition.

11.2 Traveling Waves: Pulses and Fronts

We consider systems of ordinary dierential equations (ODEs) of the form
ux = f (u; ),

x R,

u Rn ,

f C 1 (Rn Rp , Rn ),

Rp .

(11.3)

Such problems typically arise as steady state or traveling wave reductions of

PDEs of the form
Ut = DUXX + F (U ; ),

U RN .

Here, for a traveling wave U (X, t) = U (X ct) of speed c, we have x = X ct,

u = (U, V ) = (U, UX ), = (c, ), and f (u; ) = (V, D1 [F (U ; ) + cU ]).
Front solutions q(x) of (11.3) at parameter values correspond to heteroclinic connections between equilibria, and solve the following boundary value
problem on the real line

qx = f (q(x); ),
q(x) u0 as x , f (u0 ; ) = 0,
(11.4)

q(x) u1 as x , f (u1 ; ) = 0.
In the special case that u0 = u1 , the pulse q(x) corresponds to a homoclinic
orbit of the ODE.

336

Alan R Champneys and Bj

orn Sandstede

To set up a well-posed numerical boundary value problem for solutions of

(11.4), we need to deal with the innite interval x (, ) in some way, to
replace the asymptotic boundary conditions by some regular condition, and
to deal with the translation symmetry q(x) q(x+g). Several approaches for
performing this task have been proposed, including shooting methods [25, 37],
computation of large-period periodic orbits [40], and mapping the innite
domain to a nite interval; see, e.g., [26]. We focus here on methods, owing
to Beyn [3] and Friedman and Doedel [15, 18], that involve truncation to a
nite domain [L, L] and use projection boundary conditions.
Suppose that the matrices fu (uj ; ) have nsj eigenvalues (counting multiplicities) with negative real part, ncj eigenvalues with zero real part, and nuj
eigenvalues with positive real part, so that nsj + ncj + nuj = n for j = 1, 2. To
set up a numerical boundary value problem we then look for a solution u(x)
with x [L, L] subject to the boundary conditions
P s (u0 ; )(u(L) u0 ) = 0,

P u (u1 ; )(u(L) u1 ) = 0.

(11.5)

Here P s (u0 , ) is an ns0 n matrix whose rows form a basis for the stable
eigenspace of fuT (u0 ; ). Accordingly, P u (u1 ; ) is an nu1 n matrix, whose
rows form a basis for the unstable eigenspace of fuT (u1 ; ). The boundary conditions (11.5) thus place the boundary points in the center-unstable eigenspace
of fu (u0 ; ) and the center-stable eigenspace of fu (u1 ; ), respectively.
If each of the equilibria uj is hyperbolic, then ncj = 0, and by simply
counting dimensions of the corresponding stable and unstable manifolds under generic hypotheses on transversality, we nd that the codimension of
the parameter set in which there exists a heteroclinic connection is equal
to m = n (nu0 + ns1 ) + 1. If this number is negative then additional internal boundary conditions need to be set up to choose a member of the |m|dimensional continuum of connections. From now on, we assume that m is
non-negative. Hence, we need to free m parameters, say (1 , . . . , m ), to nd
a regular solution to the boundary value problem (11.3) and (11.5). However,
(11.5) represents only
ns0 + nu1 = (n nu0 ) + (n ns1 ) = n + m
conditions for the n+m+1 unknowns u Rn and (1 , . . . , m ), so that we need
an extra condition. At the same time, we see that the original problem respects
a translation symmetry since if u(x) solves (11.4), then so does u(x + g) for
any g R. So, according to the theory of Sect. 11.1, we need a phase condition
that is transverse to the generator u = ux . Such a condition can be posed at
a specic point in the domain, for example, at the left-hand boundary
(u) := v, u(L) = 0

where v is such that

v, ux (L) =
 0.

(11.6)

This is equivalent to choosing the left-hand boundary point to be in the

Poincare section = {u Rn ; v, u = 0}, with the inner product condition

11 Numerical Computation of Coherent Structures

337

implying that the corresponding trajectory of (11.3) crosses transversally.

However, for practical implementation in continuation software, one cannot
guarantee a priori that all solutions along a branch will remain transverse to
for all parameter values . Therefore, it is better to take as phase condition
the integral condition
 L
(u) :=
u (x), u(x) dx
(11.7)
L

where u is a reference solution (for example the previously computed solution

along a continuation branch); see Example 4 and also Chaps. 1, 9 and 10.
Thus we solve the two-point boundary value problem

ux = f (u; )
P s (u0 ; )(u(L) u0 )
= 0.

(11.8)

P u (u1 ; )(u(L) u1 )
L 
u (x), u(x) u (x) dx
L
Under certain nondegeneracy conditions, the existence of a homoclinic solution to the original problem (11.4) on the innite interval implies the existence
of a unique solution to the truncated problem (11.8) provided L is suciently
large; see [3, 18, 38]. Furthermore, the error involved in the truncation scales
exponentially with L for both the parameter and the solution.
There have been a number of implementations of the above algorithm for
computing homoclinic and heteroclinic connections. For example, the routines
HomCont [7] are part of the package Auto [14]. Some extensions to this basic
algorithm allow one to deal with special situations as we now outline:
Reversibility
Many ODEs arise from reductions of even-order PDEs that contain only
even spatial derivatives. In this case, the ODE can be written as an evendimensional reversible system with respect to an involution R so that
f (Ru; ) = Rf (u; ),
with R2 = id and dim Fix(R) = n/2. Standing symmetric pulses, which are
invariant under x  x, correspond to reversible homoclinic orbits q to a
reversible hyperbolic equilibrium u0 , and are of codimension zero in parameter space [6, 12]. For such an orbit we can replace the right-hand boundary
condition of (11.5) by the condition u(0) Fix(R). Note that, provided the
solution q(x) intersects Fix(R) transversally, this boundary condition also
breaks the translation invariance of the system. Therefore, there is no need
for a phase condition. This is good since we require one condition less, which
agrees with having one free parameter less compared to the non-reversible
case. This approach has been used extensively to compute so-called gap solitons in nonlinear optics; see e.g. [11].

338

Alan R Champneys and Bj

orn Sandstede

Hamiltonian Systems
Mu
noz-Almaraz et al. [33] consider how to compute periodic orbits in Hamiltonian systems; see also Chap. 9. Their approach naturally extends to heteroclinic connections under generic hypotheses on the connection. Hamiltonian
systems conserve a rst integral, the Hamiltonian itself. Noethers theorem
says that any other independent (in the sense of a Poisson bracket) conserved
quantity C(u) corresponds to a symmetry of the underlying system. Thus each
conserved quantity (including the Hamiltonian) reduces the codimension of
the solution one is trying to compute. Thus we need to remove one boundary
condition for each additional conserved integral, or introduce an extra articial parameter. The approach used in [33], for a system with k independent
conserved smooth scalar functionals Cj : Rn Rp R for j = 1, . . . , k, is to
introduce articial parameters
Rk and solve
ux = f (u; ) +

k
%

j u Cj (u; ).

(11.9)

j=1

It has been proved rigorously that this leads to well-posed boundary value
problems under appropriate generic hypotheses. We record that
= 0 for
the true solution to the boundary value problem and that u Cj (u; ) is precisely orthogonal to the level set of Cj , which is in keeping with the theory of
Sect. 11.1 above. Note that we still need to retain exactly the same projection boundary conditions (11.5) and (11.7), but now the free parameters are
replaced with the articial parameters
Rk .
Nonhyperbolic Equilibria
If u0 = u1 is a nonhyperbolic equilibrium, then one has to make special
considerations to determine the codimension of the homoclinic connection
and, hence, the number of free parameters that is needed. We consider two
cases separately.
Our rst case of interest is that when u0 = u1 and nc = 2 with a pair of
purely imaginary eigenvalues in the spectrum. In Hamiltonian or reversible
systems such a linearization is generic, and such points are called saddlecenters if nu = ns > 0: in this situation, it is then possible to form homoclinic connections with exponentially decaying tails that lie in the stable and
unstable manifolds of the equilibrium. Such solutions in reversible systems
correspond to so-called embedded solitons of nonlinear wave problems [46]. To
compute these objects, we must replace the boundary condition (11.5) with
P cs (u0 ; )(u(L) u0 ) = 0,

u(0) Fix(R),

where P cs (u0 ; ) is the spectral projection onto the center-stable eigenspace

of u0 whose rows, therefore, form a basis for the center-stable eigenspace of
fuT (u0 ; ).

11 Numerical Computation of Coherent Structures

339

The second case we consider is that u0 = u1 is a saddle node. Then pulse

solutions can exist with algebraic decay as x . Besides the condition
that u0 is a saddle-node equilibrium, which can be taken to be det fu (u0 ; ) =
0, we provide the n 1 boundary conditions (11.5) plus the phase condition
(11.7). In particular, these solutions have codimension one in the parameter
space; see, e.g., [24, 7.1.2].
Saddle-Homoclinic Orbits and Their Orientation
Homoclinic bifurcations of various kinds can be treated by freeing an additional parameter j and adding appropriate conditions whose regular zeros
dene the extra degeneracy along a branch of homoclinic solutions; see [7] for
a number of examples.
We discuss here a particular kind of codimension-two bifurcation that is
known to lie at the heart of the creation of multi-pulse solutions from a branch
of one-pulses, namely the so-called inclination-ip homoclinic bifurcation. It
is caused by a degeneracy in the global twistedness of the stable and unstable
manifolds W s,u (u0 ) of a saddle u0 around its homoclinic orbit q(x); see e.g. [36]
and references therein. At each point q(x) along the homoclinic orbit, the
normal bundle
Z(x) = [Tq(x) W s (u0 ) + Tq(x) W u (u0 )]
of the tangent spaces of stable and unstable manifolds can be dened. Generically, these tangent spaces intersect along the one-dimensional subspace
spanned by q  (x) and, therefore, Z(x) is one dimensional for each x. The
collection of Z(x) generically forms a one-dimensional bundle along the set
S1
= {u0 } {q(x); x R} which can be orientable (homeomorphic to a
cylinder) or non-orientable (homeomorphic to a M
obius band). Inclinationips occur at codimension-two points in parameter space where this bundle
changes from orientable to non-orientable (or vice versa). An ecient way to
detect inclination-ip points is to compute normal vectors w(x) Z(x) as
solutions to the adjoint variational problem
wx = fuT (q(x); )w,

w 0 as x .

(11.10)

If we truncate, then we can approximate w(x) on the same nite interval

[L, L] as q(x) by dening boundary conditions
Qu (u0 ; )w(L) = 0,

Qs (u1 ; )w(L) = 0,

(11.11)

where Qu (u0 ; ) is an nu0 n matrix whose rows form a basis for the unstable
eigenspace of fu (u0 ; ). Similarly, Qs (u1 ; ) is an ns1 n matrix, whose rows
form a basis for the stable eigenspace of fu (u1 ; ). Note that a solution to
(11.10) is only dened up to scalar multiplication w gw. Therefore, we
need a phase condition that xes the amplitude which we may take to be

340

Alan R Champneys and Bj

orn Sandstede



u 

(a)

c








(b)























 


















 










 



 







Fig. 11.2. Panel (a) shows a continuation of the branch of single-pulses to the
FitzHugh-Nagumo system (11.15) for = 0 and = 0.025; the insets show the fast
and the slow pulse for = 0.02 and a two-pulse orbit that is close to the slow pulse.
Panel (b) is an enlargement near the origin of the solution of the adjoint variational
equation around the pulse at the inclination-ip point and at two points that are
parameter perturbations of O(103 ) along the branch of homoclinic solutions.

(w) =
L

w (x), w(x) w (x) dx = 0,

(11.12)

where w (x) is a reference solution (for instance, the solution computed at a

previous continuation step). Note though that (11.12) and (11.11) provides
n + 1 boundary conditions, so we need to free an additional parameter. In this
case, it is convenient to include a parameter 0 that unfolds the orthogonality
of w and ux :
(11.13)
wx = fuT (u(x); )w + 0 f (u(x); ),
with the true solution to (11.13), (11.11), and (11.12) satisfying 0 = 0.
We nish with an example, namely, the FitzHugh-Nagumo system; see also
[14] for a demo on switching branches to multi-pulse orbits in this system,
during continuation of a single pulse, and also for references. It is given by
the PDE
ut = uyy f (u) w,
wt = (u w),

f (u) = u(u )(u 1),

(11.14)

for the functions (u, w)(y, t). Looking for traveling waves by letting x = y +ct,
we obtain the ODE system
ux = v,
vx = cv + f (u) + w,

wx = (u w).
c

(11.15)

Figure 11.2 shows the result of numerical continuation of the single-pulse

solution in the (, c)-plane for = 0 and = 0.025. Note that for most

11 Numerical Computation of Coherent Structures

341

values of there are either none or two wave speeds c for which there exists
a traveling pulse. The faster wave is the one that is generally stable. Note
that both the upper and the lower branch appear to end in mid air as is
reduced. In fact, the branches fold back on themselves and the pulses return
as their own two-pulse orbits at very nearby parameter values to those at
which the one-pulse exists; see the dashed curve in the inset to Fig. 11.2(a).
For suciently large ( > 0.1318124 on the lower branch and >
0.107652 on the upper branch) the tail of the pulse has monotonic decay
because the origin of the ODE system (11.15) is a real saddle. In such circumstances, the orientability of the normal bundle to the stable and unstable
manifolds along the homoclinic orbit is well dened. Indeed we nd that at
(, c) = (0.240314, 0.211443) there is an inclination ip with respect to the
stable manifold; see Fig. 11.2(b). Theoretical and numerical work for inclination ips in the FitzHugh-Nagumo system can also be found in [23].

11.3 Traveling Waves with Spatially Periodic

Asymptotics
Dieci and Rebaza [13] considered a general boundary value approach for computing heteroclinic connections between hyperbolic periodic orbits and from
equilibria to periodic orbits for ODEs. These correspond to traveling waves
whose prole becomes spatially periodic as x and/or x . For the
special case where the periodic solutions in the two tails are the same (up to
a phase shift), a popular approach for computing such generalized pulses is
to study the problem in the setting of a Poincare map and to compute homoclinic connections to hyperbolic xed points of the map; see, for example,
[1, 5, 21] and also the Auto implementation HomMap due to Yagasaki [45].
Instead, we present here the approach from [13] (see also [22]) that relies on
setting up a coupled boundary value problem for the periodic orbits and the
homoclinic orbit.
Suppose a connection q(x) is sought between two separate hyperbolic periodic solutions pj (x; ) with periods Lj for j = 0, 1 which have nsj Floquet
multipliers inside the unit circle and nuj = nnsj +1 outside. Note that ncj = 1
due to the presence of the trivial Floquet multiplier 1. We then seek a solution
to the boundary value problem

u f (x; )


f (v 0,1 ; )
v0,1

cs v0,1 (0) v0,1 (L0,1 )

(11.16)
P (v0 ; )(u(L) v0 (0)) = 0,
cu

P (v1 ; )(u(L) v1 (0))

)
0,1 (v0,1 v0,1
with solution (u, v0 , v1 ) near (q, p0 , p1 ). Here, the projection boundary condition P cs (v0 ; ) is an (ns0 + 1) n matrix whose columns are orthogonal

342

Alan R Champneys and Bj

orn Sandstede






(a)

max uper



 

(b)

























 

L/L0












 





x/L

Fig. 11.3. Homoclinic orbits to periodic orbits of (11.17) computed on the halfinterval [L, 0] (a), and their locus in the phase shift versus tail amplitude plane
(b). From A.R. Champneys and G.J. Lord, Computation of homoclinic solutions to
c
periodic orbits in a reduced water-wave problem, Physica D, 102 (1997) 101124 
1997 by Elsevier Science; reprinted with permission.

to the unstable eigenspace of the monodromy matrix associated with v0 at

x = 0. In other words, the rows of P cs (v0 ; ) are formed by the center-stable
Floquet eigenspace of the adjoint variational problem wx = fuT (v0 (x); )w
for x [0, L0 ]. Note the importance of solving the adjoint problem in order
that the projection represented by P cs is orthogonal to the unstable Floquet
eigenspace, rather than being along the center-stable Floquet directions (this
detail was inadvertently omitted in [13]). Similarly, P cu sets the component
orthogonal to the stable Floquet eigenspace to zero. The phase conditions 0,1
are chosen to factor out the translation symmetry of the periodic orbits, and
may be taken to be

j (v) =

Lj

x vj (x), v(x) dx,

j = 1, 2,

where vj (x) are reference periodic solutions (we may again take vj to be
pj computed at a previous point along a continuation branch). Note that
there is no phase condition on the heteroclinic orbit u(x) since the translation
symmetry is broken by xing the phase of the two periodic orbits. Moreover,
the intervals L0 , L1 and L should all be taken as unknowns. This is because
L0 and L1 are the a priori unknown periods of the periodic orbits p0 and
p1 , and L must be taken to be free in order to nd the unknown phase shift
between the periodic orbits p0 and p1 .
Practical implementation details of the boundary value problem (11.16)
are discussed in [13], which also contains a convergence proof based on the
methods of [4]. In particular, care has to be taken in order to evaluate the
matrices P cs (v0 ; ) and P cu (v1 ; ) which themselves depend on the solutions
vj (x).
We illustrate the approach outlined above by considering an application
to a fourth-order water-wave problem. Champneys and Lord [9] consider the

11 Numerical Computation of Coherent Structures

system

2 uiv (x) + u (x) u(x) + u(x)2 = 0,

x R,

343

(11.17)

where u(x) represents the amplitude of water waves in the presence of surface
tension for a model equation that is only valid for small-amplitude waves close
to the critical point of Bond number 1/3 and Froude number 1. Equation
(11.17) can easily be shown to be equivalent to a fourth-order Hamiltonian
system, with total energy given by
H = 2 u u +

u3
u2
2
[u ]2
+ .
[u ]2
3
2
2
2

The problem is also reversible with respect to the involution

R:

(u, u , u , u )  (u, u , u , u ).

Figure 11.3 shows the results of computations of reversible homoclinic orbits

to reversible periodic orbits, where the right-hand boundary conditions in
(11.16) are replaced with u(0) Fix(R) and v0 (L0 /2) Fix(R).
Similar results were found for the full water-wave problem in the presence
of surface tension using boundary-integral methods [10]; see Fig. 11.4. These
careful computations helped to settle a conjecture (in the negative) of whether
true solitary waves of elevation exist in the classical water wave problem; see
also [41].

11.4 Moving Discrete Breathers in R Z

The discrete nonlinear Schr
odinger equation with a saturable nonlinear term,
representing the eects of a photorefractive crystal lattice, may be written in
the form
un (t)
, n Z,
(11.18)
iu n (t) = [2 u(t)]n +
1 + |un (t)|2

where 2 is the standard second-order spatial dierence operator [2 u]n =

un+1 2un + un1 , and represents the dimensionless coupling strength between each lattice site. Localized time-periodic solutions to such equations
have been given the name discrete breathers.
Standing discrete breathers are relatively easy to compute since, after making a phase transformation un (t) = eit Un , one has to solve for homoclinic
orbits of the discrete map
iUn = [2 U ]n +

Un
,
1 + |Un |2

n Z,

which are structurally stable (and indeed many of which are stable and may
be found as solutions of the initial value problem (11.18)). However, if one

344

Alan R Champneys and Bj

orn Sandstede
0.06

0.06

J
0

-0.06

10

20

30

40

-0.08

50

curvature

0.04

0.6

0.2

0
-0.01

10

20

30

40

0.06

30

50

10

20

30

40

50

-0.2

40

-0.06

50

20

0
0

10

0.4

0.06

0.04

-0.4

-0.6

-0.08

0.226
0

10

20

30

40

0.228

0.23

0.232

0.234

0.236

50

0.06

0.06

0.238

-0.02

10

20

30

50

40

0.04

0.04

0
0
0
0
-0.08

10

20

30

40

50

-0.06

10

20

30

40

50

-0.02

10

20

30

40

50

-0.01

10

20

30

40

50

Fig. 11.4. Equivalent diagram to Fig. 11.3 for the Euler-equation formulation of
water waves in the presence of surface tension plotted as Bond number against
a signed measure of the tail amplitude for xed Froude number F = 1.002 and
for xed domain size L = 98.33. From A.R. Champneys, J.M. Vanden-Broeck and
G.J. Lord, Do true elevation gravity-capillary solitary waves exist? A numerical
c 2002 by Cambridge University
investigation, J. Fluid Mech., 454 (2002) 403417 
Press; reprinted with permission.

tries to make such structures move with wave speed c = 0, they typically shed
radiation and eventually stop or cease to survive as coherent structures.
Moving discrete breathers can be sought by making the substitution
un (t) = (z)eit

with z = n ct R,

which gives the advanced-retarded equation

ic  (z) = (2 )(z) ((z + 1) + (z 1)) +

(z)
,
1 + |(z)|2

(11.19)

for which we seek homoclinic solutions (z) 0 as z . Simple spectral

analysis shows that the spectrum of the problem linearized about = 0 is
symmetric with respect to the imaginary axis, so that the dimension nc of
the center manifold is even; moreover, nc 2 for all c, > 0. Counting the
codimension of the stable and unstable manifolds shows that the best we
can hope for is that nc = 2, and then homoclinic solutions to (11.19) should

11 Numerical Computation of Coherent Structures

(a)



345

(b)




















Fig. 11.5. Continuation of breathers on a periodic background for various values

of 1/ showing three zeros in for c = 0.7, = 0.5, and L = 60 (a); the shaded
region represents the spectral band where any embedded solitons would be of codimension two. Panel (b) shows solutions on the branch with a zero of near = 1.02.
From T.R.O. Melvin, A.R. Champneys, P.G. Kevrekidis and J. Cuevas, Radiationless traveling waves in saturable nonlinear Schr
odinger lattices, Phys. Rev. Lett. 97
c 2006 by the American Physical Society; reprinted with permission.
(2006) 124101 

exist along codimension-one lines in the parameter plane. An ecient way to

compute such solutions is to seek solutions on a large interval z [L, L]
with appropriate boundary conditions. Note that (11.19) is invariant under
the reversing transformation
R:

z  z,

(Re , Im )  ( Re , Im ).

We can nd R-symmetric periodic solutions by seeking solutions with (L)

Fix(R), (0) Fix(R). The use of these boundary conditions eectively xes
the phase symmetry  eig according to the theory of Sect. 11.1.
An ecient way to solve such boundary value problems is to use a pseudospectral method [16] by making the substitution
(z) =

N
%

aj cos(jz/L) + i bj sin(jz/L)

(11.20)

j=1

for the unknown coecients (aj , bj ). Substitution of (11.20) into (11.19) and
jL
evaluating at the collocation points zj = 2(N
+1) leads to a regular system
of 2N algebraic equations that can be solved numerically.
In particular, we can use this method to compute periodic solutions of xed
large period that approximate quasi-solitons made up of an exponentially
localized core core and a nonvanishing oscillatory background tail :

(z) = core (z) + tail (z).

At a sucient distance from the center of the breather core is zero due to
its exponential localization and, therefore, we are left with the tail which, if

346

Alan R Champneys and Bj

orn Sandstede


(a)




(b)





































Fig. 11.6. Continuation of the three zeros of shown in Fig. 11.5, varying and
c with = 0.5 (a); the depth of shading represents the number of pairs nc /2 of
imaginary eigenvalues of the linearized problem in each parameter region: white
nc = 2, lightest gray nc = 4, and so on; circles on the {c = 0} axis indicate the
transparent points at which the energy barrier for steady breathers vanishes. Panel
(b) shows the real parts of solutions on the rst branch, whose amplitude goes to
zero as the upper grey wedge is approached. From T.R.O. Melvin, A.R. Champneys,
P.G. Kevrekidis and J. Cuevas, Radiationless traveling waves in saturable nonlinear
c 2006 by the American
Schr
odinger lattices, Phys. Rev. Lett. 97 (2006) 124101 
Physical Society; reprinted with permission.

it is of small amplitude, is in the center manifold and hence approximately

sinusoidal. Because of the way (z) has been approximated in (11.20) we know
that the imaginary part of tail (z) is odd around z = 0 and the quantity
= Im (L/2)
can be used as a signed measure of the amplitude of the tail. Regular zeros of
can be followed in multiple parameters, and these give the traveling discrete
breathers that we are seeking. These moving structures with zero tails were
found to be remarkably stable as solutions to the initial value problem.
Figure 11.5 shows the results of computations of solutions with periodic
tails for xed c and , but with allowed to vary. Note that the U- and Nshaped nature of the branches is qualitatively the same as in Fig. 11.4, except
here we see the occasional S-shaped dislocation of the Us and Ns. These lead
to values of c at which the measure of the tail amplitude undergoes a regular
zero. Hence, we have found a truly localized solution that can be continued
in two parameters. Three such zeros are found in Fig. 11.5, and they lead to
the three branches labeled I, II and III in Fig. 11.6. Note that these branches
terminate when they reach the boundary of the white region in Fig. 11.6(a),
when nc = 2 is no longer satised. More details can be found in [32].

11 Numerical Computation of Coherent Structures

347

x
Fig. 11.7. A space-time contour plot of a typical heteroclinic cycle of (11.21) is
shown for = 70, = 1, and = 4, where dark color corresponds to smaller
and light color to larger values of the amplitude |u(x, t)|. From D.J.B. Lloyd,
A.R. Champneys and R.E. Wilson, Robust heteroclinic cycles in the one-dimensional
c 2005 by Elcomplex Ginzburg-Landau equation, Physica D 204 (2005) 240268 
sevier Science; reprinted with permission.

11.5 Computing Robust Heteroclinic Cycles

Heteroclinic cycles are formed of trajectories that connect equilibria or periodic orbits. This behavior is typically structurally unstable in generic dynamical systems. However, in systems with symmetry, robust connections between
saddles can occur if each saddle is actually a sink within an invariant subspace
generated by a discrete symmetry.
As an example, we consider the one-dimensional complex Ginzburg
Landau equation (CGL)
ut = (1 + i)uxx + u (1 + i)|u|2 u

(11.21)

on a periodic domain x S1
= [0, 1]/ , where u C and , , R. The
CGL (11.21) is invariant under the action of the S1 phase symmetry
u(x)  ei u(x)
and the O(2)-symmetry
u(x)  u(x),

u(x)  u(x y)

for y S1
= [0, 1]/ .
A typical heteroclinic cycle of (11.21), obtained by a direct integration, is
shown in Fig. 11.7. We observe long dormant behavior followed by bursts of

348

Alan R Champneys and Bj

orn Sandstede

spatio-temporal activity. The dormant states are relative equilibria of (11.21)

of the form
(11.22)
u(x, t) = U (x)eit
that are reection symmetric about x = 0.5 and satisfy U (x + 1/2) = U (x)
for all x. The relative equilibria observed in Fig. 11.7 are all related by a 1/4spatial translation. The bursts are characterized by initial symmetry breaking
with spatial wavenumber k = 2 of the relative equilibria (11.22), and the
corresponding heteroclinic orbits lie in the xed-point space generated by
reection across x = 0, x = 1/4, x = 1/2 or x = 3/4. The trajectory shown
in Fig. 11.7 follows these dierent heteroclinic orbits seemingly randomly.
Since we have periodic boundary conditions, it is natural to use Fourier
modes and, following [31], we, therefore, seek solutions of the form
%
Wn (t)e2inx ,
(11.23)
u(x, t) = ei(t)
nZ

where we choose (t) so that W0 (t) is real-valued, while Wn (t) is complexvalued for n = 0, to factor out the S1 phase symmetry. Substituting (11.23)
into (11.21) yields the innite-dimensional system
%
2

Wk W l Wm , (11.24)
i(t)W
n + Wn = [(2n) (1+i)]Wn (1+i)
kl+m=n

which governs the evolution of Wn , where the phase velocity is expressed

explicitly by
6
5
%
1
+
i
(11.25)
Wk W l Wm .
= Im
W0
kl+m=0

If we wish to compute the heteroclinic orbit in the xed-point space of functions invariant under the reection across x = 0, we would restrict ourselves
to functions that are even in x and set Wn = Wn for all n Z. This would
then allow us to obtain the desired heteroclinic orbit as a codimension-zero
saddle-sink connection using again a boundary value formulation.
Instead, we now describe the results obtained in [28] for the N = 2 Fourier
truncation of (11.24) without assuming any relation between Wn and Wn ;
in particular, (11.24) is ve dimensional for N = 2. The results are shown
in Fig. 11.8. The spatial average mode W0 , shown as a solid curve, exhibits
the standard heteroclinic cycle behavior. The quantities Re(W1 W1 ) and
Re(W2 W2 ) measure how far we are away from the xed-point space of
even functions. We nd that the quantity Re(W2 W2 ) vanishes for all times
even though W2 , W2 = 0. This implies, and can be checked numerically,
that W2 (t) = W2 (t) is a spatial symmetry of the heteroclinic cycles. The
other quantity Re(W1 W1 ) is plotted as the dashed curve in Fig. 11.8
and turns out to vanish for approximately half the heteroclinic connections.
Hence, roughly half of the heteroclinic connections lie in the invariant subspace

11 Numerical Computation of Coherent Structures





(a)










(b)

349









Fig. 11.8. Heteroclinic cycle in Fourier space (a) found by direct simulation of
(11.24) with N = 2 for = 70, = 4, and = 1; the solid curve shows the
evolution of W0 , while the dashed curve gives Re(W1 W1 ). Panel (b) shows an
enlargement of the cycle. From D.J.B. Lloyd, A.R. Champneys and R.E. Wilson,
Robust heteroclinic cycles in the one-dimensional complex Ginzburg-Landau equac 2005 by Elsevier Science; reprinted with
tion, Physica D 204 (2005) 240268 
permission.

{Wn = Wn ; n = 0} corresponding to even functions. If we plotted Re(W1 +

W1 ), then we would see that this quantity is also zero half of the time but
for precisely the other half of heteroclinic connections. This then gives us the
invariant subspace of odd functions. Both subspaces intersect at the equilibria
where W1 = W1 = 0 allowing for structurally stable heteroclinic cycles to
exist to such equilibria within either subspace.

11.6 Traveling Waves on Cylinders R S 1

We now treat the case of PDEs on multi-dimensional domains with a single
unbounded direction. For simplicity we shall restrict to two spatial dimensions
and assume periodic boundary conditions in the other direction. Rather than
presenting the general theory, which can be found in [30], we treat a specic
example which is concerned with localized buckling of cylindrical shells [29].
Consider an innitely long cylinder of radius R and shell thickness t. The
equilibrium for the in-plane stress function and displacement w in the postbuckling regime of the cylinder is governed by the von K
armanDonnell equations:
2 4 w + wxx xx = wxx yy + wyy xx 2wxy xy ,
2

+ wxx = (wxy ) wxx wyy ,

4

(11.26)
(11.27)

where 4 denotes the two-dimensional bi-harmonic operator, x R is the

axial and y [0, 2R) the circumferential coordinate. Furthermore, := 1/R
is the curvature, 2 := t2 /12(1 2 ), is Poissons ratio, and the bifurcation
parameter is := P/Et where P is the compressive axial load (force per unit

350

Alan R Champneys and Bj

orn Sandstede

length), and E is Youngs modulus. The form of solutions sought are localized
in the axial length, but circumferentially periodic, suggesting that equations
(11.26)(11.27) should be supplemented with periodic boundary conditions in
y and asymptotic boundary conditions in the axial direction x. Simplications
can be made by considering two types of symmetry conditions in the mid plane
of the cylinder, namely
R:
S:

w(x, y) = w(x, y), (x, y) = (x, y),

w(x, y) = w(x, y + R/s), (x, y) = (x, y + R/s),

where s is the axial wavenumber of the post-buckling pattern sought.

The numerical approach taken in [29] is to use a spectral decomposition
circumferentially for each value of s of the form
w(x, y) =

N
%

aj (x) cos(jsy),

(x, y) =

j=0

N
%

bj (x) cos(jsy),

j=0

allied to Galerkin projection, to reduce the PDEs (11.26)(11.27) to a system

of 2N fourth-order ODEs, each akin to (11.17), for the amplitudes (aj , bj )(x).
The symmetries R and S provide two dierent reversing transformations of
the ODEs, and so the projection boundary condition at one end can be replaced by using the Fix(R) or Fix(S) boundary conditions depending on which
symmetric solution is sought.
Figure 11.9(a) and (b) shows numerical continuation results for both Rsymmetric and S-symmetric localized modes for a typical experimentally
amenable geometry, but for dierent circumferential wavenumber s. The horizontal axis in these plots is the overall end shortening implied by the deformation; see [29] for the details. The trend is the same for all solutions, namely
that they bifurcate from the trivial solution subcritically (the load reduces
sharply). They then reach a turning point with respect to at which point
the branch would become stable in a controlled end-shortening experiment.
Figure 11.9(c) shows an example of a localized buckle shape near such a fold.

11.7 Defects in Oscillatory Media

In the preceding section we discussed the computation of localized stationary
solutions of PDEs posed on the cylinder R S1 . Here we illustrate that the
computation of (not necessarily localized) time-periodic solutions of reactiondiusion systems
ut = uxx + f (u),

x R, u Rn

(11.28)

proceeds similarly. We denote the temporal frequency of a time-periodic solution by and use the rescaled time variable = t. Time-periodic solutions
of (11.28) are then in one-to-one correspondence with solutions of

11 Numerical Computation of Coherent Structures

351




(a)

104

104



(b)




































(c)

Fig. 11.9. Bifurcation diagrams for (a) R-symmetric and (b) S-symmetric solutions
to (11.26)(11.27) for dierent seed modes s with = 0.01, t = 0.247, = 0.3,
E = 5.56, and N = 6. Panel (c) shows the displacement w(x, y) of the R-symmetric
mode for s = 11 with at its local minimum value. From G.J. Lord, A.R. Champneys
and G.W. Hunt, Computation of homoclinic orbits in partial dierential equations:
an application to cylindrical shell buckling, SIAM J. Sci. Comp 21 (1999) 591
c 1999 by the Society for Industrial and Applied Mathematics; reprinted with
619 
permission.

u = Duxx + f (u),

(x, ) R S1 ,

(11.29)

where is restricted to S1
= [0, 2]/ . We now have two symmetries present
in our system, namely, translations in x and in . Thus, we need to add two
phase conditions, which we choose according to Remark 1, and we also need
to add the generator cx u of the spatial translation to (11.29).

352

Alan R Champneys and Bj

orn Sandstede
 














(a)

         



(b)










Fig. 11.10. Panel (a) is a space-time contour plot of the third component u3 (x, )
of the solution to the R
ossler system (space x is horizontal and rescaled time is
vertical); since the defect is symmetric with respect to the reection x  x, the
solution is plotted only for x 0. Panel (b) shows the dependence of the temporal
frequency on the parameter .

We obtain the system

u + Duxx + cux + f (u)

 2 

 u (x, ), u(x, ) u (x, ) dx d = 0,

02 R
 u (x, ), u(x, ) u (x, ) dx d
0
R x

(11.30)

which we wish to solve for (u, , c), where u is dened on R S1 . In practice,

one truncates the real line R to a large interval (L, L) and imposes additional
boundary conditions such as Neumann at x = L.
As an example, we consider the three-component Rossler system with u
R3 where D = 0.4 and
f (u; ) = (u2 u3 , u1 + 0.2u2 , 0.2 + u1 u3 u3 )T

(11.31)

is the nonlinearity which depends on a parameter . In [35], a defect solution

was computed by implementing the system (11.30) with Neumann boundary
conditions at x = L in Auto [14]. The time direction is discretized using
nite dierences, taking the periodic boundary conditions for into account.
The resulting system of ODEs in the spatial variable x is solved in Auto as
a boundary value problem.
Note that a symmetry condition x x is used at the left-hand boundary
in Fig. 11.10(a). Thus the defect is occurring along this edge of the plot. This
defect solution is robust and can be continued with the algorithm outlined
above. An example of such a numerical computation is shown in Fig. 11.10(b).

11.8 Localized Structures in R2

Localized planar structures arise in many applications. Examples are light
bullets in optical bres [42], localized roll patterns known as convectons in

11 Numerical Computation of Coherent Structures

353

binary-uid convection [34], buckling modes in mechanical structures [19],

localized micro-structures in solidication [20], and oscillating localized patterns known as oscillons in vertically vibrated sand layers [17, 43] and in the
Belousov-Zhabotinsky reaction [44].
Rather than attempting a general discussion of how these patterns could be
computed numerically, we focus on localized stationary hexagonal structures
in the planar SwiftHohenberg equation
[1 + ]2 u + u u2 + u3 = 0.

(11.32)

Three typical localized hexagon structures are shown in Fig. 11.11; note that
solutions (a)(c) are computed for the same values of the parameters and .
There are dierent ways for attempting to compute such patterns. We may,
for instance, discretize in Cartesian coordinates and use three phase conditions to remove translation and rotation symmetries; see, e.g. [27]. However,
a discretization in Cartesian coordinates will give preference to patterns with
D4 -symmetry, thus possibly leading to spurious solutions when computing
hexagonal structures.
An alternative approach is pursued in [8] where the patterns shown in
Fig. 11.11 were computed using polar coordinates. We now describe the latter approach in more detail. Written in polar coordinates (r, ), the Swift
Hohenberg equation for u(r, ) is given by
[1 + ]2 u + u u2 + u3 = 0,

(11.33)

where

1
1
r, u = urr + ur + 2 u .
r
r
1
The computational domain is S = [0, 2]/ and 0 < r < R for some
R  1 with the boundary condition

u(R, ) = ur (R, ) = 0,

S1
= [0, 2]/

(11.34)

at r = R.
Solutions are now sought using the Fourier ansatz
u(r, ) =

N
1
%

an (r) cos(6n),

(11.35)

n=0

which exploits the D6 -symmetry inherent to hexagons. With this ansatz, the
boundary conditions (11.34) become
an (R) = an (R) = 0,

n = 0, . . . , N 1.

(11.36)

We need to supplement these conditions with pole conditions at r = 0 to take

care of the singularity of the Laplace operator at the origin. We choose to
work with the pole conditions

354

Alan R Champneys and Bj

orn Sandstede

(a)

(b)

(c)

Fig. 11.11. Spatial contour plots of three localized hexagon structures (a)(c) of
the planar SwiftHohenberg equation (11.32) with = 0.3 and = 1.6. The computational domain is R = 100, and N = 20 Fourier modes were used; plotted are
the solutions on a 50 50 square.

a0 (0) = a
0 (0) = ak (0) = ak (0) = 0,

n = 1, . . . , N 1,

(11.37)

which enforce that u(r, ) is even in r and make the variational formulation
of (11.33) well dened at r = 0.
Figure 11.11 shows results of a computation where the resulting ODE
boundary value problem has been implemented and solved with Auto. Note
that localized hexagonal structures of arbitrarily wide spatial extent can be
computed this way. In fact panels (a)(c) are three hexagonal patterns on the
same continuation branch; details will appear in [8].

11.9 Planar Spiral Waves

As a nal example, let us comment briey on the computation of spiral wave
solutions of a planar reaction-diusion system
ut = Du + f (u),

x R2 , u Rn .

Spiral waves are relative equilibria with respect to the rotation symmetry
of the plane and, therefore, of the form u(r, , t) = u (r, t) in polar
coordinates (r, ). In particular, the spiral-wave prole u (r, ) is a solution
to the elliptic system
Dr, u + u + f (u) = 0.

(11.38)

To factor out the rotation symmetry, we may add the phase condition
 2
 u (R/2, ), u(R/2, ) u (R/2, ) d = 0
(11.39)
0

posed at r = R/2.

11 Numerical Computation of Coherent Structures

355

Fig. 11.12. A contour plot of the third component u3 of a spiral-wave solution to

the planar R
ossler system (11.28) with (11.31) is shown.

To solve (11.38)(11.39), we truncate to the computational domain (r, )

(0, R)[0, 2] with periodic boundary conditions in and Neumann boundary
conditions at r = R plus appropriate pole conditions at r = 0. Barkley [2]
then used nite dierences in r and a spectral method in to compute spiral
waves in the two-dimensional FitzHugh-Nagumo system (11.14).
Figure 11.12 shows the example of one such computation for the R
ossler
system (11.28) with (11.31) in two spatial dimensions, where we used Barkleys
numerical code [2]. Shown is a contour plot of the third component u3 of a
spiral-wave solution at a xed instant in time.

11.10 Outlook
We have attempted to outline a general approach to computing various kinds
of coherent structures on innite or semi-innite domains. This procedure
involves setting up formally well-posed boundary value problems that factor
out the symmetries or degeneracies that arise due to the particular nature of
the coherent structure in question. In essence, most of our methods rely on
reduction to a two-point boundary value problem, even when solving problems
in the plane. Thus, one can harness the full power of software such as Auto
for the continuation of paths of such solutions as parameters vary.
However, to compute more general patterned states, one needs to solve
boundary value problems in appropriate two- or three-dimensional spatial
domains without the articial reduction to one-dimensional spatial problems.
A reliable package for the continuation of solutions to fully two- and threedimensional elliptic problems is clearly a pressing requirement. Such a package
should be easy to use, yet suciently powerful. This would enable the user

356

Alan R Champneys and Bj

orn Sandstede

to set up specic boundary conditions that factor out the symmetries, in the
manner described here, in dependence on the underlying properties of the
structures that are sought.
Another pressing need, and the subject of much ongoing research, is the
derivation of algorithms that are able to compute information on the spectral
stability of the coherent structures while these structures themselves are being
computed.

Acknowledgments
We thank David Lloyd for his helpful comments and suggestions and, especially, for his help with text and gures in Sects. 11.5 and 11.8. The results
reviewed here have been obtained in collaboration over many years, and we
would like to thank Jes
us Cuevas, Giles Hunt, Panayotis Kevrekidis, Yuri
Kuznetsov, David Lloyd, Gabriel Lord, Thomas Melvin, Bart Oldeman, Arnd
Scheel, Jean-Mark Vanden-Broek, and Eddie Wilson for their contributions.
We are grateful to Elsevier Science, Cambridge University Press, the American Physical Society, and the Society for Industrial and Applied Mathematics
for permission to reproduce previously published material.

References
1. F. S. Bai, G. J. Lord, and A. Spence. Numerical computations of connecting
orbits in discrete and continuous dynamical systems. Internat. J. Bifurc. Chaos
Appl. Sci. Engrg., 6:12811293, 1996.
2. D. Barkley. Linear stability analysis of spiral waves in excitable media. Phys.
Rev. Lett., 68:20902093, 1994.
3. W.-J. Beyn. The numerical computation of connecting orbits in dynamical
systems. IMA J. Num. Anal., 9:379405, 1990.
4. W.-J. Beyn. On well-posed problems for connecting orbits in dynamical systems.
Contemp. Math., 172:131168, 1995.
5. W.-J. Beyn and J. M. Kleinkauf. The numerical computation of homoclinic
orbits for maps. SIAM J. Num. Anal., 34:12071236, 1997.
6. A. R. Champneys. Homoclinic orbits in reversible systems and their applications
in mechanics, uids and optics. Physica D, 112:158186, 1998.
7. A. R. Champneys, Yu. A. Kuznetsov, and B. Sandstede. A numerical toolbox
for homoclinic bifurcation analysis. Internat. J. Bifur. Chaos Appl. Sci. Engrg.,
6(5):867887, 1996.
8. A. R. Champneys, D. J. Lloyd, and B. Sandstede. Localised patterns in the
planar Swift-Hohenberg equation. In preparation.
9. A. R. Champneys and G. J. Lord. Computation of homoclinic solutions to
periodic orbits in a reduced water-wave problem. Physica D, 102:101124, 1997.
10. A. R. Champneys, J. M. Vanden-Broeck, and G. J. Lord Do true elevation
gravity-capillary solitary waves exist? A numerical investigation. J. Fluid Mech.,
454:403417, 2002.

11 Numerical Computation of Coherent Structures

357

11. C. M. deSterke and J. E. Sipe. Gap solitons. Progr. Opt., 33, 1994.
12. R. L. Devaney. Reversible dieomorphisms and ows. Trans. Amer. Math. Soc.,
218:89113, 1976.
13. L. Dieci and J. Rebaza. Point-to-point and periodic-to-periodic connections.
BIT Num. Math., 44:4162, 2004.
14. E. J. Doedel, R. C. Paenroth, A. R. Champneys, T. F. Fairgrieve,
Yu. A. Kuznetsov, B. E. Oldeman, B. Sandstede, and X. Wang. Auto2000:
Continuation and bifurcation software for ordinary dierential equations (with
HomCont). Technical report, Concordia University, 2002.
15. E. J. Doedel and M. J. Friedman, Numerical computation of heteroclinic orbits.
J. Comp. Appl. Math., 26:317327, 1989.
16. D. B. Duncan, J. C. Eilbeck, H. Feddersen, and J. A. D. Wattis. Solitons on
lattices. Physica D, 68:111, 1993.
17. J. Fineberg. Physics in a jumping sandbox. Nature, 382:793764, 1996.
18. M. J. Friedman and E. J. Doedel. Numerical computation and continuation of
invariant manifolds connecting xed points. SIAM J. Numer. Analysis, 28:789
808, 1991.
19. G. W. Hunt, M. A. Peletier, A. R. Champneys, P. D. Woods, M. Ahmer Wadee,
C. J. Budd, and G. J. Lord. Cellular buckling in long structures. Nonlin. Dyn.,
21:329, 2000.
20. H. Jamgotchian, N. Bergeon, D. Benielle, P. Voge, B. Billia, and R. Guerin. Localised microstructures induced by uid ow in directional Solidication. Phys.
Rev. Lett., 87(16):166105, 2001.
21. J. M. Bergamin, T. Bountis, and M. N. Vrahatis. Homoclinic orbits of invertible
maps. Nonlinearity, 15:16031619, 2002.
22. B. Krauskopf and T. Riess. Numerical analysis of heteroclinic connections between hyperbolic equilibria and hyperbolic periodic orbits. Unpublished, 2007.
23. M. Krupa, B. Sandstede, and P. Szmolyan. Fast and slow waves in the FitzHugh
Nagumo equation. J. Dier. Eqns., 133(1):4997, 1997.
24. Yu. A. Kuznetsov. Elements of Applied Bifurcation Theory, third edition,
(Springer-Verlag, New York, USA, 1994).
25. Yu. A. Kuznetsov. Computation of invariant manifold bifurcations. In D. Roose,
A. Spence, and B. De Dier, editors, Continuation and Bifurcations: Numerical
Techniques and Applications, pages 183195 (Kluwer, Dordrecht, The Netherlands, 1990).
26. L. X. Liu, G. Moore, and R. D. Russell. Computation and continuation of
homoclinic and heteroclinic orbits with arclength parameterization. SIAM J.
Sci. Comp., 18:6993, 1997.
27. D. J. B. Lloyd and A. R. Champneys. Ecient numerical continuation and
stability analysis of spatiotemporal quadratic optical solitons. SIAM J. Sci.
Comp., 27(3):759773, 2005.
28. D. J. B. Lloyd, A. R. Champneys, and R. E. Wilson. Robust heteroclinic cycles
in the one-dimensional complex Ginzburg-Landau equation. Physica D, 204:240
268, 2005.
29. G.J . Lord, A. R. Champneys, and G. W. Hunt. Computation of homoclinic orbits in partial dierential equations: An application to cylindrical shell buckling.
SIAM J. Sci. Comp, 21:591619, 1999.
30. G. J . Lord, D. Peterhof, B. Sandstede, and A. Scheel. Numerical computation of solitary waves in innite cylindrical domains. SIAM J. Numer. Anal.,
37(5):14201454, 2000.

358

Alan R Champneys and Bj

orn Sandstede

31. B. P. Luce. Homoclinic explosions in the complex Ginzburg-Landau equation.

Physica D, 84(3-4):553581, 1995.
32. T. R .O. Melvin, A. R. Champneys, P. G. Kevrekidis, and J. Cuevas. Radiationless traveling waves in saturable nonlinear Schr
odinger lattices. Phys. Rev.
Lett., 97:124101, 2006.
33. F. J. Mu
noz Almaraz, E. Freire, J. Gal
an, E. Doedel, and A. Vanderbauwhede.
Continuation of periodic orbits in conservative and hamiltonian systems. Physica D, 181:138, 2003.
34. J. J. Niemela, G. Ahlers, and D. S. Cannell. Localized travelling-wave states in
binary-uid convection. Phys. Rev. Lett., 1990.
35. B. E. Oldeman and B. Sandstede. Numerical computation of defects in the
R
ossler model. Unpublished.
36. B. E. Oldeman, B. Krauskopf, and A. R. Champneys. Death of period-doublings:
locating the homoclinic-doubling cascade. Physica D, 146:100120, 2000.
37. A. J. Rodrguez-Luis, E. Freire, and E. Ponce. A method for homoclinic and
heteroclinic continuation in two and three dimensions. In D. Roose, A. Spence,
and B. De Dier, editors, Continuation and Bifurcations: Numerical Techniques
and Applications, pages 197210 (Kluwer, Dordrecht, The Netherlands, 1990).
38. B. Sandstede. Convergence estimates for the numerical approximation of homoclinic solutions. IMA J. Num. Anal., 17:437462, 1997.
39. B. Sandstede, A. Scheel, and C. Wul. Dynamics of spiral waves on unbounded
domains using center-manifold reductions J. Di. Eqns., 141:122149, 1997.
40. C. Sparrow. The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors.
(Springer-Verlag, New York, 1982).
41. S. M. Sun. Non-existence of truly solitary waves in water with small surface
tension. Proc. Royal Soc. A, 455(1986):21912228, 1999.
42. M. Tlidi, A. G. Vladimirov, and P. Mandel. Interaction and stability of periodic
and localized structures in optical bistable systems. IEEE J. of Quant. Electr.,
39(2):216226, 2003.
43. P. B. Umbanhowar, F. Melo, and H. L. Swinney. Localized excitations in a
vertically vibrated granular layer. Nature, 382:793796, 1996.
44. V. K. Vanag and I. R. Epstein. Stationary and oscillatory localized patterns,
and subcritical bifurcations. Phys. Rev. Lett., 92(12):128301, 2004.
45. K. Yagasaki. Numerical detection and continuation of homoclinic points and
their bifurcations for maps and periodically forced systems. Internat. J. Bifurc.
Chaos Appl. Sci. Engrg., 8:16171627, 1998.
46. J. Yang, B. A. Malomed, D. J. Kaup, and A. R. Champneys. Embedded solitons:
a new type of solitary wave. Math. Comp. Simulation, 55:585600, 2001.

12
Continuation and Bifurcation Analysis of
Delay Dierential Equations
Dirk Roose1 and Robert Szalai2
1
2

Department of Computer Science, Katholieke Universiteit Leuven, Belgium

Department of Engineering Mathematics, University of Bristol, United Kingdom

Mathematical modeling with delay dierential equations (DDEs) is widely

used in various application areas of science and engineering (e.g., in semiconductor lasers with delayed feedback, high-speed machining, communication
networks, and control systems) and in the life sciences (e.g., in population dynamics, epidemiology, immunology, and physiology). Delay equations have an
innite-dimensional state space because their solution is unique only when an
initial function is specied on a time interval of length equal to the largest
delay. Consequently, analytical calculations are more dicult than for ordinary dierential equations and numerical methods are generally the only way
to achieve a complete analysis, prediction and control of systems with time
delays.
Delay dierential equations are a special type of functional dierential
equation (FDE). In FDEs the time evolution of the state variable can depend
on the past in an arbitrary way as long as the dependence is a bounded
function of the past. However, DDEs impose a constraint on this dependence,
namely that the evolution depends only on certain past values of the state at
discrete times. (We do not consider here the case of distributed delay.) The
delays can be constant or state dependent. The equations can also involve
delayed values of the derivative of the state, which leads to equations of neutral
type.
In this chapter we mainly discuss the simplest case, namely a nite number
of constant delays. Specically, we consider a nonlinear system of DDEs with
constant delays j  0, j = 1, . . . , m, of the form
x (t) = f (x(t), x(t 1 ), x(t 2 ), . . . , x(t m ), ),

(12.1)

R is a nonlinear smooth function

where x(t) R , and f : R
depending on a number of (time-independent) parameters Rp . We assume
that the delays are in increasing order and denote the maximal delay by
n

(m+1)n+p

= m = max i .
i=1,...,m

360

Dirk Roose and Robert Szalai

A solution segment is denoted by xt = xt () = x(t + ) C, [, 0]. Here

C = C([, 0]; Rn ) is the space of continuous functions mapping the delay
interval into Rn . For a xed value of the parameter , a solution x(t) of (12.1)
on t [0, ) is uniquely dened by specifying a function segment x0 as an
initial condition. A discontinuity in the rst derivative of x(t) generally appears at t = 0 and is propagated in time, even if f and are innitely smooth.
However, the solution operator of (12.1) smooths the solution, meaning that
discontinuities appear in higher and higher derivatives as time increases.
A DDE can be approximated by a system of ordinary dierential equations
(ODEs) and so standard numerical methods for ODEs could be used. However, to obtain an accurate approximation a high-dimensional system of ODEs
is needed, and this leads to expensive numerical procedures. During the last
decade, more ecient and more reliable numerical methods have been developed specically for DDEs. In this chapter we survey those numerical methods
for the continuation and bifurcation analysis of DDEs that are implemented in
the software packages DDE-Biftool [26, 27] and PDDE-Cont [78]. Where
appropriate, we also briey describe alternative numerical methods. Note that
we do not discuss time integration of DDEs; for this topic see, for example,
[3] and [7].
The structure of this chapter is as follows. In Sect. 12.1 we discuss numerical methods to compute the right-most characteristic roots of steady-state
solutions. In Sect. 12.2 we describe collocation methods for computing periodic solutions and their dominant Floquet multipliers. Section 12.3 presents
dening systems for codimension-one bifurcations of periodic solutions that
allow one to compute the location of bifurcation points accurately. Computation of connecting orbits is discussed in Sect. 12.4 and of quasiperiodic
solutions is discussed in Sect. 12.5. In Sect. 12.6 we briey discuss how to deal
with special types of DDEs, specically, equations of neutral type and DDEs
with state-dependent delays. In Sect. 12.7 we discuss specic details of the
software packages DDE-Biftool and PDDE-Cont. Their functionality is
illustrated in Sect. 12.8, where we present the bifurcation analysis of several
DDE models of practical relevance. Finally, conclusions and an outlook can
be found in Sect. 12.9.

12.1 Stability of Steady-State Solutions

In this and the next section we assume that the parameter is xed and we
omit it from the equations. A steady-state solution x(t) x of (12.1) satises
the nonlinear system
(12.2)
f (x , x , x , . . . , x ) = 0.
The (local) stability of x is determined by the stability of (the zero solution of) the linearized equation
&m
y  (t) = A0 y(t) + j=1 Aj y(t j ),
(12.3)

12 Bifurcation Analysis of Delay Dierential Equations

361

where Aj Rnn denotes the partial derivative of f with respect to its (j+1)th
argument, evaluated at x . The linearized equation (12.3) is asymptotically
stable if all its roots of the characteristic equation
det(I A0

m
%

Aj ej ) = 0

(12.4)

j=1

lie in the open left half-plane (i.e., Re() < 0); see, e.g., [40, 62, 75]. Equation
(12.4) has an innite number of roots , known as the characteristic roots.
However, the number of characteristic roots with real part larger than a given
threshold is nite. Hence, to analyze the stability of a steady-state solution,
one must determine reliably all roots satisfying Re()  r, for a given r < 0
close to zero.
Analytical conditions for stability can be found in Step
an [75] and Hassard
[44]. These conditions are deduced by using the argument principle of complex analysis, and they give a practical method for determining stability. In
recent years, numerical methods have been developed to compute approximations to the right-most (stability-determining) characteristic roots of (12.4),
by using a discretization either of the solution operator of (12.3) or of the
innitesimal generator of the semi-group of the solution operator of (12.3).
The solution operator S(t) of the linearized equation (12.3) maps an initial
function segment onto the solution segment at time t, i.e.,
(S(t)y())() = y(t + ),

0, t 0.

(12.5)

This operator has eigenvalues , which are related to the characteristic roots
via the equation = et [67]. To determine the stability, we are interested
in the dominant eigenvalues. If t is large then these eigenvalues are well separated, which can be exploited during the eigenvalue computation, but the
time integration itself may be costly. In Sect. 12.1.1 we describe a reliable
way to compute the dominant eigenvalues of S(h) where h is the time step of
a linear multistep (LMS) method.
Since S(t) is a strongly continuous semi-group [38, 40], one can dene the
corresponding innitesimal generator A by
Ay = lim

t0+

S(t)y y
.
t

(12.6)

For (12.3) the innitesimal generator becomes

Ay() = y  (), y D(A)
&m
D(A) = {y C : y  C and y  (0) = j=0 Aj y(j )}.

(12.7)

Both operators can be discretized by spectral discretizations or time integration methods; this always leads to a representation by some matrix.
Eigenvalues of this matrix yield approximations to the right-most characteristic roots. Hence, for computational eciency it is important that the size of

362

Dirk Roose and Robert Szalai

the resulting matrix eigenvalue problem is small, or at least that the stability determining eigenvalues can be computed eciently by using an iterative
method, such as subspace iteration. Accurate characteristic roots can be found
by using Newton iterations on
&m
(I A0 j=1 Aj ej )v = 0,
(12.8)
v0 T v = 1,
where v Rn and v0 Rn , to obtain accurate characteristic roots (and the
corresponding eigenfunctions vet ). The dierence between the approximate
and the corrected roots gives an indication of the accuracy of the approximations.
Below we describe how the characteristic roots can be computed via an approximation of the solution operator by time integration, which is the method
that is implemented in DDE-Biftool. We also briey comment on other approaches.
12.1.1 Approximation of the Solution Operator by a Time
Integrator
A natural way to approximate the solution operator is to write the numerical
time integration of the linearized equation as a matrix equation. Engelborghs
et al. [28] have proposed and analyzed the use of a linear multistep method
with constant steplength h to approximate the solution operator S(h). The
delay interval [, 0] (slightly extended to the left and the right; see below) is
discretized by using an equidistant mesh with mesh spacing h, and a solution
is represented by a discrete set of points yi := y(ti ) with ti = ih. A k-step
LMS method with steplength h to compute yk can be written as
k
&
i=0

i yi = h

k
&
i=0

1
.
m
&
i A0 yi +
Aj y(ti j ) ,

(12.9)

j=1

where i and i are parameters and (in case ti j does not coincide with a
mesh point) the approximations y(ti j ) are obtained by polynomial interpolation with s and s+ points to the left and the right, respectively.
The discretization of the solution operator is the (linear) map between
[yLmin , . . . , yk1 ]T and [yLmin +1 , . . . , yk ]T where Lmin = s $ /h% and
where the mapping is dened by (12.9) for yk and by a shift for all variables
other than yk . This map is represented by an N N matrix, where
N = n(Lmin + k) n /h.

(12.10)

Since the time step h is small, the eigenvalues of this matrix are not well
separated (most eigenvalues lie close to the unit circle). They can be computed
by e.g. the QR method, with a computational cost of the order N 3 n3 ( /h)3 ,
and so approximations to the characteristic roots can be derived.

12 Bifurcation Analysis of Delay Dierential Equations

363

To guarantee the reliability of the stability computation, the steplength h

in the LMS method (12.9) should be chosen such that all characteristic roots
with Re()  r (r < 0) are approximated accurately. Procedures for such a
safe choice of h are described in [28, 84] and implemented in DDE-Biftool.
They are based on theoretical properties of
(a) the relation between the stability properties of the solution of the linearized
equation (12.3) and the stability of the discretized equation (12.9);
(b) an a-priori estimate of the region in the complex plane that includes all
characteristic roots with Re()  r.
Note that the solution operator can also be discretized by using a RungeKutta time integrator [11].
12.1.2 Other Approaches
Breda et al. [12] have developed numerical methods to determine the stability
of solutions based on a discretization of the innitesimal generator. By discretizing the derivative in (12.7) with a Runge-Kutta or an LMS method, a
matrix approximation of A is obtained. The resulting eigenvalue problem is
large and sparse, as in the case when the solution operator is discretized by a
time integration method. Breda et al. [13] also proposed a pseudo-spectral discretization of the innitesimal generator. In this approach, an eigenfunction of
the innitesimal generator vet , t [, 0], is approximated by a polynomial
P (t) of degree p. Collocation for the eigenvalue problem for the innitesimal
generator leads to an equation of the form
P  (ti ) = P (ti ),

(12.11)

where the collocation points ti , i = 1...p are chosen as the shifted and scaled
roots of an (orthogonal) polynomial of degree p. These equations are augmented with
&m
(12.12)
A0 P (0) + j=1 Aj P (j ) = P (0),
which introduces the system-dependent information. The resulting matrix
eigenvalue problem has size n(p + 1). The rst np rows are the Kronecker
product of a dense p (p + 1) matrix and the identity matrix. The last block
row consists of a linear combination of the matrices Aj , j = 0, ..., m and the
identity. The matrix is full but can be of much smaller size than in the previous
case, due to the spectral accuracy convergence, as is shown in the detailed
analysis presented in [13].
A pseudo-spectral discretization of the solution operator is proposed in
[11, 85]. Here a polynomial approximation P (t) of an eigenfunction, dened
on the interval [, h] has to satisfy p collocation conditions of the form
P (ti + h) = P (ti ),

364

Dirk Roose and Robert Szalai

where = eh . These equations are augmented with a condition obtained from

integrating the linearized equations over a time interval h. When high accuracy is required, a pseudo-spectral discretization will lead to a more ecient
procedure than when a time integration discretization is used, but numerical
experiments indicate that for low accuracy requirements both approaches are
competitive [85].
However, for the pseudo-spectral approaches no strategy is known that
guarantees a priori that all characteristic roots with real part larger than r
are computed accurately; this is in contrast to discretization of the solution
operator with an LMS method.

12.2 Periodic Solutions

A periodic solution x (t) of an autonomous system of the form (12.1) satises
x (t + T ) = x (t), t,
where T is the period. An extensive literature exists on the existence, stability
and parameter dependence of periodic solutions; see, e.g., [40, XI.1-2]. These
results are essentially analytical in nature and the corresponding methods
have dierent rigorous restrictions and cannot be applied to general nonlinear
systems with several delays.
Because of the dependence on the past, periodicity of x(t) at one moment
in time, x(t) = x(t + T ) for some t, does not imply periodicity for the whole
solution. Instead, a complete function segment of length has to be repeated.
Consequently, a periodic solution to (12.1) can be found as the solution of the
following two-point boundary value problem (BVP),

x (t) = f (x(t), x(t 1 ), . . . , x(t m ), ), t [0, T ],
x0 = xT ,
(12.13)

p(x, T ) = 0,
where x0 and xT are function segments on [, 0] and [ +T, T ], respectively,
and the period T is an unknown parameter. Furthermore, p represents a phase
condition that is needed to remove translational invariance. A well-known
example is the classical integral phase condition [21]
 1
x (0) (s)(x(0) (s) x(s)) ds = 0,
(12.14)
0

where x(0) is a reference solution; see also Chap. 1.

Stable periodic solutions of a DDE can be found by numerical time integration; the convergence of the integration depends on the stability properties of
the periodic solution [46]. However, both stable and unstable solutions can be
computed by solving the above boundary value problem by either collocation
or by a shooting approach. Here we only consider collocation methods.

12 Bifurcation Analysis of Delay Dierential Equations

365

12.2.1 Collocation
In collocation a periodic solution is computed by using a discrete representation that satises the dierential equation at a set of collocation points on
[0, T ]. Doedel and Leung [22] have computed periodic solutions of DDEs using collocation based on a truncated Fourier series; see also [14] for a similar
approach. This Fourier approach has the advantage that periodicity is automatically fullled. However, steep gradients in a solution pose problems and
it is not possible to determine the solution stability.
Collocation based on piecewise-polynomial representations is used in Auto
[18] and Content [52] to compute periodic solutions for systems of ordinary
dierential equations; see also Chaps. 1 and 2. We now discuss how piecewisepolynomial collocation can be used for DDEs. We rst rescale time by a factor
1/T such that the period is one in the transformed system

x (t) = T f (x(t), x(t 1 /T ), . . . , x(t m /T ), ) , for t [0, 1],
x( + 1) x() = 0, for [ /T, 0],
(12.15)

p(x, T ) = 0.
A mesh with L + 1 mesh points {0 = t0 < t1 < . . . < tL = 1} is specied.
This mesh is periodically extended to the left with  points to obtain a mesh
on [ /T, 1] with +L intervals. In each interval an approximating polynomial
of degree d is described in terms of the function values at the representation
points (using Lagrange polynomials as basis). These function values are determined by requiring that the approximating collocation solution fullls the
(time-scaled) dierential equations exactly at the collocation points. In each
interval, the collocation points are typically chosen as the (scaled and shifted)
roots of a dth degree orthogonal polynomial.
The approximating polynomial of degree d on each interval [ti , ti+1 ], i =
, . . . , L 1, can be written as
u(t) =

d
%
j=0

u(ti+ j )Pi,j (t), t [ti , ti+1 ],

(12.16)

where Pi,j (t) are the Lagrange polynomials through the representation points

j
ti+ j = ti + (ti+1 ti ),
d
d

j = 0, . . . , d.

Because polynomials on adjacent intervals share the value at the common

mesh point, this representation is automatically continuous (however, it is
not continuously dierentiable at the mesh points).
The approximation u(t) is completely determined by the coecients
ui+ j := u(ti+ j ), i = , . . . , L1, j = 0, . . . , d1 and uL := u(tL ). (12.17)
d

We dene the starting vector us and the nal vector uf , both of length N =
n(d + 1), as

366

Dirk Roose and Robert Szalai

us := [u , . . . , ui+ j , . . . , u0 ]T , uf := [uL , . . . , ui+ j , . . . , uL ]T .

d

(12.18)

The collocation points are obtained as

ci,j = ti + cj (ti+1 ti ), i = 0, . . . , L 1, j = 1, . . . , d,
from a set of collocation parameters cj , j = 1, . . . , d, e.g., the shifted and
scaled roots of the dth degree Gauss-Legendre polynomial.
A periodic solution for a xed value of the parameters is found as the
solution of the following (n(( + L)d + 1) + 1)-dimensional (nonlinear) system
in terms of the unknowns (12.17) and T ,

i,j ) = T f (u(ci,j ), u(ci,j T1 ), . . . , u(ci,j Tm )), ) = 0,

u(c

i = 0, . . . , L 1, j = 1, . . . , d,
(12.19)
u

u
=
0,

f
s

p(u) = 0.

Here, p again represents a phase condition such as (12.14).

The collocation solution fulls the time-scaled dierential equation exactly
at the collocation points. If ci,j coincides with ti then the right derivative is
taken in (12.19), if it coincides with ti+1 then the left derivative is taken.
Taking into account the periodicity conditions, one can reduce system (12.19)
to the following nonlinear system in the unknowns u = [u0 , . . . , uL ]T and T ,

u(c
i,j ) = T f (u(ci,j ), u((ci,j T1 ) mod 1), . . . , u((ci,j Tm ) mod 1), ) = 0,

i = 0, . . . , L 1, j = 1, . . . , d,

u
=
0,
u

0
L

p(u) = 0.
(12.20)
Hence, the dimension of the system and the number of unknowns is reduced
to (n(Ld + 1) + 1).
When using Newtons method to solve (12.20), the matrix of the linear
system to be solved in each iteration is sparse and has a particular structure,
as is shown in Fig. 12.1. The matrix consists of a (large) nLd n(Ld + 1)
matrix lled with two (circular) bands, bordered by one column and n + 1
rows. The extra column contains derivatives with respect to the period; n
extra rows contain the periodicity condition, and one extra row is due to
the phase condition (12.14). The diagonal band is itself a concatenation of
nd n(d + 1) blocks. The o-diagonal bands are a consequence of the delay
terms. When the mesh is equispaced then the o-diagonal band lies at a xed
distance from the diagonal band as is illustrated in Fig. 12.1(a). This is no
longer the case when the mesh is non-equispaced; see Fig. 12.1(b).
In the case of collocation for ODEs, the matrix of the linear system has a
band structure with a band size proportional to n and d but independent of
the number of mesh intervals L; see also Chap. 1. Hence the system can be
solved eciently by a direct band solver. For delay dierential equations this

12 Bifurcation Analysis of Delay Dierential Equations

367

Fig. 12.1. Structure of the matrix arising in the Newton iteration to solve (12.20) for
one delay that is smaller than the period T ; shown is the case L = 7 and collocation
polynomials of degree d = 3. Panel (a) is for an equispaced mesh and panel (b) for
a non-equispaced mesh; each black box represents an n n block.

is not possible. Indeed, the structure of the matrix, described above, cannot
easily be exploited when using a direct solver, especially in case of several
delays and/or a non-equispaced mesh. However, for moderate values of d, n
and L the linear system (12.20) can still be solved with a direct method. The
eciency can be increased by using a Newton-chord method, in which case
the Jacobian is not recomputed (and factored) in every iteration but remains
xed during a number of iterations. In Sect. 12.2.3 we describe an ecient
iterative procedure to solve (12.20).
Furthermore, an adaptive (non-equispaced) mesh can be used to decrease
the required number of intervals L for dicult solutions (with steep gradients).
For the latter, the interval size hi = ti+1 ti is adapted to equidistribute the
(d + 1)th derivative of the solution along the mesh; see [2, 25].
Engelborghs and Doedel [24] have proven that the convergence rate of the
maximal continuous error E = maxt[0,1] u(t) u (t) is O(hd ) in general
and O(hd+1 ) for Gauss-Legendre collocation points on equispaced and nonequispaced meshes with h = maxi hi . Special convergence rates at the mesh
points (so-called superconvergence) that feature for ODEs, are, in general,
lost for DDEs.
Note that, in the case of a nonautonomous (or forced) system, the collocation method is essentially the same as in (12.20), except that the phase
condition is not needed, since the phase of the solution is determined by the
phase of the forcing.

368

Dirk Roose and Robert Szalai

12.2.2 Monodromy Operator and Floquet Multipliers

The stability of a periodic solution is determined by the Floquet multipliers,
which are the eigenvalues of the monodromy operator. In the case of autonomous equations there is always a trivial multiplier +1, which stems from
the fact that the associated linearized equation is always solved by the time
derivative of the solution itself. According to Floquet theory, a periodic solution is asymptotically stable if all the multipliers not counting the trivial
one lie strictly inside the complex unit disk. The main focus of this section
is the computation of the monodromy operator using the previously described
collocation method.
Denote by x (t) a T -periodic solution of (12.1). As in the previous sections
we rescale time by 1/T . The linearized equation about this periodic solution
in rescaled time is
m
%
d
Aj (t)y(t )),
y(t) = T (A0 (t)y(t) +
dt
j=1

(12.21)

where Aj (t) denotes the partial derivative of f with respect to its (j + 1)th
argument, evaluated at x (T t). Also let U (t, s) be the fundamental solution
operator of (12.21), which is dened as
(U (t, s)s )() = y(t + ),

[ /T, 0],

where s is an initial function and y is the corresponding solution of (12.21).

The monodromy operator is dened as
M = U (1, 0),
that is,
M : C([ /T, 0]; Rn ) C([ /T, 0]; Rn ),
 y1 ,
where is the initial function and y1 is the solution segment y1 () = y(1 + ),
[ /T, 0]. The discretized version of M is Md : us uf and its matrix
representation can be obtained by solving (12.21) with a collocation method
similar to (12.19). This method is used in DDE-Biftool [27].
However, when the maximal delay is larger than the period, us and uf
overlap; the computation of Md can be improved by exploiting this property.
For the sake of generality we use the Riesz representation theorem and write
(12.21) in the form
 /T
dy(t)
=T
d (T , t)y(t ),
(12.22)
dt
0

where is a matrix-valued function of bounded variation that, with (12.21),

can be written as

12 Bifurcation Analysis of Delay Dierential Equations

369

0
if 0,
if 0 < < 1 ,
A0 (t)
(T , t) =
..
..

&. m

A0 (t) + j=1 Aj (t) if .

Notice that (12.22) implicitly depends on the initial function. It can be written
explicitly as
dy(t)
T
dt

/T

d (T , t)y(t ) T
0

d (T , t)(t ) = 0.

(12.23)

We introduce K = $ /T % solution segments of y(t) and as

y1 (t) = y(2 K + t), y2 (t) = y(3 K + t), . . . , yK (t) = y(1 + t)
1 (t) = (1 K + t), 2 (t) = (2 K + t), . . . , K (t) = (t),

t [1, 0],

such that i , yi X := C([1, 0]; Rn ), and we also dene operators on X as

obtained from (12.23) as
d()
T
(A)() =
dt

1+

d (T , )( ),

D(A) = C 1 ([1, 0], Rn ),

0
 i+1+

d (T , )(i + ),

(Bi )() = T

1  i  N.

i+

It is clear that yK is the only unknown, because all the other yi can be found
from the initial conditions as yi = i+1 . Hence, the only equation that has to
be solved is
AyK

K
%

Bi i = 0,

yK (1) = K (0).

i=1

In order to eliminate the explicit boundary condition we introduce extended

operators on X = {(, c) X Rn : c = (0)} in the form of






A0
Bi 0
BN 0

A=
, Bi =
for i < N and BN =
,
L0
0 0
0 I
where L = (1). The extended monodromy operator is dened on X =
C([N, 0]; Rn ); this space is isomorphic to the further extended
"
#
= ((1 , c1 ), . . . , (N , cN )) X N : k (0) = ck = k+1 (1), 1  k < N .
X
In order to obtain stability results it is sucient to construct the monodromy
which becomes
operator on X,

370

Dirk Roose and Robert Szalai

M=

0
..
.

0
..
.

..
.

0
0
..
.
I

0
0
A1 B1 A1 B2 A1 BN

(12.24)

is not
Because of the identity matrices above the diagonal, the operator M
k , k  K are compact.
compact, but its powers M
can be computed by collocation and by inverting the
The operator M
is comresulting discretized A operator. In PDDE-Cont the spectrum of M
puted with the iterative Arnoldi-Lanczos method [68], which is implemented
in the ARPACK software package [54]. Note that in this iterative process,
when the discretized operator is multiplied by a vector, only one solution step
with A is necessary.
gives the same accuracy of
Despite the dierences, using either M or M
the multiplier calculation [56]. In particular, it was shown in [56] that the
computations of the multipliers and of the periodic solution itself have the
same accuracy. The exception is the computation of the trivial multiplier +1,
Hence, inferring the
which was found to be more accurate when using M.
accuracy of the computed periodic solution from the accuracy of the trivial
multiplier can be deceiving.
12.2.3 Collocation-Newton-Picard
Verheyden and Lust [83] have developed an iterative procedure to solve the
linear system arising in Newtons method applied to system (12.19). Consider
the unknowns ui+j/d := u(ti+j/d ) dened in (12.17). Recall the denition of
the starting vector us and the nal vector uf given in (12.18)
us := [u , . . . , ui+ j , . . . , u0 ]T , uf := [uL , . . . , ui+ j , . . . , uL ]T ,

(12.25)

and dene the trajectory vector as

ut := [u d1 , . . . , ui+ j . . . , uL ]T ,

(12.26)

= nLd (here
where us and uf are of length N = n(d+1) and ut is of length N
 and L denote the number of mesh points in [ /T, 0] and (0, 1], respectively).
Note that uf consists of the last n(d + 1) components of ut .
The linearization of (12.19) has the following form
Bus + Aut + r1,T T = r1 ,
us + uf = r2 ,
s us + t ut + T T = ,

(12.27)

where r1 , r2 and denote the residuals of system (12.19) and B, A and r1,T
denote the partial derivatives of the collocation conditions with respect to us ,

12 Bifurcation Analysis of Delay Dierential Equations

371

Fig. 12.2. Typical structure of the linearized collocation system for one delay that
is smaller than the period T ; shown is the case of  + L = 3 + 7 mesh intervals
and collocation polynomials of degree d = 3 (the bordering row and column are
omitted). Panel (a) is for an equispaced mesh and panel (b) for a non-equispaced
mesh; each black box represents an n n block.

ut and T . A typical structure for the matrix of the linearized system in the
case of one time delay that is smaller than the period T is shown in Fig. 12.2.
Panel (a) is for an equispaced mesh, while in panel (b) a non-equispaced mesh
is used. In both cases the mesh with L = 7 mesh intervals is extended with
 = 3 additional mesh intervals and the piecewise polynomials have degree
d = 3.
The linear system can be manipulated and condensed to the form
3
2 3
32
2
r
us
M I bc
= c .
(12.28)
s T
T
c
Here M is the discretization of the monodromy operator, which can be derived
from Mt = A1 B. Afterwards, ut can be computed from (12.27). This
manipulation is based on the correspondence between the linearization of the
collocation scheme and the discretization of the linearized boundary value
problem. The condensation is related to the condensation used in Auto; see
Chap. 1. System (12.28) can be solved with the Newton-Picard method [55],
which leads to a substantial reduction in the computational cost, especially
when only a few Floquet multipliers are larger in modulus than a certain
threshold , e.g., = 0.5 [83].

372

Dirk Roose and Robert Szalai

12.3 Dening Systems for Codimension-One

Bifurcations of Periodic Solutions
Periodic solutions of autonomous DDEs allow three generic codimension-one
bifurcations. First, the monodromy operator may have an algebraically double
+1 eigenvalue, which corresponds to a limit point (or fold, or saddle-node)
bifurcation where the solution ceases to exist. Second, there may be a single
1 multiplier, which gives a period-doubling bifurcation. Third, if two critical
complex conjugate multipliers lie on the unit circle of the complex plane, then
there is a Neimark-Sacker or torus bifurcation. In this case an invariant torus
bifurcates from the periodic solution.
Continuing the bifurcations of periodic solutions in DDEs does not dier
substantially from the case of ODEs. In order to compute bifurcations one
has to include additional equations to (12.20), which are satised by a periodic solution if, and only if, the monodromy operator has a certain kind of
singularity. In the period-doubling and the Neimark-Sacker cases, the simplest procedure to construct such a determining system is to require that the
monodromy operator has a singular vector. Short algebraic transformations
of (12.24) reveals that these bifurcations occur if

N
%
A
j Bj v = 0,
j=1

(12.29)
H

v v = 1,
has a unique solution v with the inverse characteristic multiplier = 1 = 1
on the unit circle. (Throughout, a superscript H denotes the (complex conjugate) transpose.) Because of the appearance of higher powers of , this
equation is dierent from the ODE case if the delay is larger than the period.
Adding (12.30) to the dening system of the periodic solution (12.20) doubles
the size of the problem. The size of (12.30) can be reduced to n + 1 by using characteristic matrices that are equivalent to the operator in (12.30) [80].
However, the smallest possible addition would consist of only one additional
scalar equation to (12.20) without introducing new variables. This can be
achieved by using the bordering theorem [9], which states that the bordered
operator

D
H

A b
cH 0

1
,

/ ker AH , c
/
exists if both A and AH have one-dimensional kernels and b
ker A or A is bijective and cH A1 b = 0. Moreover, can be used as a test
functional of the singularity, because it is zero if, and only if, A is singular. In
order to obtain it is sucient to solve the equation

12 Bifurcation Analysis of Delay Dierential Equations

  
A b

0
=
.

1
cH 0

373

(12.30)

&N
Hence, using a discretized version of A j=1 j Bj for the operator A
in (12.30) with appropriate choices of b and cH in the period-doubling and
Neimark-Sacker cases, the equation (x , ) = 0 determines the bifurcation
point. In a continuation context the resulting can be re-used as the value of
c in the next continuation step. Similarly, by solving the adjoint equation
H   
A c

0
=
,

1
bH 0
the resulting can be the new value of b in the next continuation step.
In the case of a fold bifurcation in an autonomous system (12.1), because
of the algebraically double +1 multiplier, the operator A has to be


&N
A j=1 j Bj 0
ALP =
,
Int0
0
where
0 = f (x(t), x(t 1 ), x(t 2 ), . . . , x(t m ), ),
0 =

N
%

j Bj 0 ,

j=1

and

Int0 =

0 ()() d.

ODEs;
Note that ALP is dierent from what one would expect by analogy with
&N j Bj ;
see [20]. Here, 0 is obtained by computing the Jordan chain of A
j=1
see, e.g., [48]. The regularity of obtained from ALP at the bifurcation point
can be proven either by using the equivalence with characteristic matrices [80]
or by standard techniques [20].

12.4 Connecting Orbits

A solution x (t) of (12.1) at some xed value of the parameter is called a
connecting orbit if the limits
lim x (t) = x

and

lim x (t) = x+

t+

(12.31)

exist, where x and x+ are steady states of (12.1). We call the orbit homoclinic when x = x+ , and heteroclinic otherwise. Orbits of this type exist,

374

Dirk Roose and Robert Szalai

for instance, in laser models with optical feedback, which are discussed in
Sect. 12.8.1; see also [35]. They also appear naturally when looking for traveling waves in delay partial dierential equations [70].
A dening condition for a connecting orbit is that it is contained in both
the stable manifold of x+ and the unstable manifold of x . A classical approach in the ODE case is to approximate this condition by truncating the
time domain to an interval of length Tc and to apply (so-called) projection
boundary conditions [8]: one end point of the connecting orbit is required to
lie in the unstable eigenspace of x and the other end point in the stable
eigenspace of x+ . The projection boundary conditions, therefore, replace the
stable and unstable manifolds by their linear approximations near the steady
states.
Here, the boundary conditions need to be written in terms of solution
segments. Furthermore, x+ has innitely many eigenvalues with negative real
parts (see Sect. 12.1) and so it is impossible to write the nal function segment
as a linear combination of all stable eigenfunctions. Instead, it is required that
the end function segment is in the orthogonal complement of all unstable left
eigenfunctions. We will assume for notational convenience that (12.1) only
contains one delay ; however, the method is implemented in DDE-Biftool
for the general case of m xed delays.
The condition for the initial function segment x0 () can be written as

s
s
%
%

x0 () = x +
k vk ek
|k |2 = 1 ,
k=1

k=1

where s is the number of unstable eigenvalues , with corresponding eigenvectors v . The k are unknown coecients, and is a measure for the desired
accuracy. An extra condition is added to ensure continuity at = 0. Since we
cannot write the end conditions for the nal function segment in a similar way,
a special bilinear form [38] is used to express the fact that the nal function
segment is in the complement of the unstable eigenspace of x+ . This leads to
the s+ extra conditions of the form
 0


+
H
H
wk+ ek (+ ) A1 (x+ , ) x(Tc + ) x+ d = 0 ,
wk+ (x(Tc ) x+ ) +

where k = 1, . . . , s+ . Here s+ is the number of unstable eigenvalues of x+ , wk+

are the left eigenvectors corresponding to the eigenvalues +
k , and the matrix
A1 is dened as in (12.3). While this integral condition works well in practice,
one slight drawback is that it does not control the distance of the end function
segment to the steady state.
As for periodic solutions, connecting orbits arise in one-parameter families
and any time-translate is also a connecting orbit. Therefore, a phase condition
such as (12.14) needs to be added to select one of these orbits.
For the case of a one-parameter family of connecting orbits a number of
free parameters are required to obtain a generically isolated solution. One has

12 Bifurcation Analysis of Delay Dierential Equations

375

to solve (12.1) together with the steady-state equations (12.2) for x and x+

+
+
and characteristic equations of the form (12.8) for
k and vk and k and wk ,
i.e., a system of n dierential equations, supplemented with (s + s+ )(n +
1) + 2n + s+ + 2 extra equations, resulting in the need for s = s+ s + 1
free parameters. This leads to a boundary value problem, which is coupled
to a number of algebraic constraints for the equilibria and their stability.
The boundary value problem can be solved by a collocation method as in
Sect. 12.2.1.
Good starting conditions for Newtons method can be obtained as follows. For a homoclinic orbit, one can start from a nearby periodic solution
with a suciently large period. Heteroclinic orbits can be approximated by
using time integration or by using an extension of the method of successive
continuation [19]. Details of the method, including a numerical study of the
convergence, are presented in [69].

12.5 Quasiperiodic Tori

In dynamical systems quasiperiodic solutions reside on invariant tori. In this
section we describe a method to compute two-dimensional tori as periodic
functions on the unit square. In particular we adapt the method of Schilder
et al. [72], which uses a nite dierence method to discretize the dening
equation. Here we use a spectral collocation method that is well suited to
delay equations.
A quasiperiodic solution x(t) of (12.1) has two rationally independent
frequencies 1 , 2 . Hence, there exists a function y : R2 Rn , which is 2periodic in both variables, such that x can be written as x(t) = y(1 t1 , 2 t2 ).
Putting u into (12.1) yields a rst-order delayed partial dierential equation

1
2

f (u(t1 , t2 ), u(t1 1 1 , t2 2 1 ), . . .
u(t1 , t2 ) =
u(t1 , t2 ) +
1
1 t2
t1
. . . , u(t1 1 m , t2 2 m ), ), (12.32)

where 1 , 2 are unknown frequencies. Because there are translational symmetries in both variables of u, two phase conditions have to be imposed on
u in order to x a unique solution and determine the unknown frequencies.
Assuming that we have a reference solution u(0) of (12.32) at 0 , we formulate
a condition that minimizes the distance of u at from u(0) , i.e.,
 2  2
1
u(t1 + 1 , t2 + 2 ) u(0) (t1 , t2 )22 dt1 dt2 .
(1 , 2 ) =
(2)2 0
0

Taking the rst derivative of with respect to 1 and 2 , the phase conditions
become

376

Dirk Roose and Robert Szalai

1
(2)2

1
(2)2

(0)
u (t1 , t2 )u(t1 , t2 ) dt1 dt2 = 0,
t1

(0)
u (t1 , t2 )u(t1 , t2 ) dt1 dt2 = 0.
t2

In the case of time-periodic systems only the second phase condition is necessary, since the phase in t1 is xed by the phase of the forcing. In addition
to the phase conditions, we also need boundary conditions that guarantee the
periodicity of u, that is,
u(0, t2 ) = u(2, t2 ) and
u(t1 , 0) = u(t1 , 2), t1 , t2 [0, 2].
12.5.1 Spectral Collocation
To obtain an approximation of the quasiperiodic solution the dening sets of
equations can be discretized with an appropriate numerical scheme and solved
by Newtons method. There are several dierent spectral collocation methods
for partial dierential equations that could be used to solve (12.32); for an
introduction see Trefethen [81]. Here we use a method that was developed for
computationally challenging hyperbolic equations such as the Navier-Stokes
equation. The method is a multi-domain spectral collocation method called
the staggered grid Chebyshev method, developed by Kopriva and Kolias [50].
The method is similar to the collocation of periodic solutions. It uses
piecewise polynomials that are represented by their values at discrete points
of a mesh, which is dierent from the mesh on which the equation is solved.
We use a very simple domain subdivision of the area [0, 2][0, 2] that splits
it into the rectangles
j j+1
Di,j = [ti1 , ti+1
1 ] [t2 , t2 ],
l
where {0 = t0l < t1l < < tL
l = 2} with l {1, 2}. On each rectangle Di,j
i,p j,q
we use the Lobatto points (t1 , t2 ) = (ti1 + bp1 (ti+1
ti1 ), tj2 + bq2 (tj+1
tj2 ))
2
1
to represent the solution

u(t1 , t2 ) =

d2
d1 %
%

j,q
i,j,p,q
u(ti,p
(t1 , t2 ),
1 , t2 )P

(12.33)

p=0 q=0
j,q
where P i,j,p,q are the Lagrange polynomials through the points (ti,p
1 , t2 ). The
function u is now completely determined by the values
j,q
ui,j,p,q := u(ti,p
1 , t2 ),

which we consider identical if they represent the same point in [0, 2] [0, 2].
We also need to impose the boundary conditions, which are

12 Bifurcation Analysis of Delay Dierential Equations

377

Fig. 12.3. Sparsity structure of the Jacobian of the discretization of (12.32). The
seemingly irregular pattern is due to the patching of the rectangles dened by
(12.34); parameters are L1 = L2 = 4, d1 = d2 = 3 and n = 1.

u0,j,0,q = uL1 1,j,d1 ,q , 0  q < d2 , 0  j < L2 1,

ui,0,p,0 = ui,L2 1,p,d2 , 0  p < d1 , 0  i < L1 1 and
u0,0,0,0 = uL1 1,L2 1,d1 ,d2 .

(12.34)

It is also possible to think of the piecewise polynomials as discontinuous in the

interfaces and dene mortar equations as in spectral penalty methods (see,
e.g., Hesthaven [45]) instead of (12.34).
Equation (12.32) is solved on the grid
q i+1
i
j,q
(ti,p
ti1 ), tj2 + cq2 (tj+1
tj2 )),
1 , t2 ) = (t1 + c1 (t1
2
0  i < L1 , 0  j < L2 , 0  p < d1 , 0  q < d2 ,

where cp1 , cq2 are the Gauss points. Using the polynomial representation (12.33)
j,q
of u in (12.32) and evaluating at (ti,p
1 , t2 ) yields a large algebraic system
that can be solved by Newtons method. The typical sparsity structure of the
Jacobian of this discretized system is shown in Fig. 12.3, but without the
borders accounting for the phase conditions.

378

Dirk Roose and Robert Szalai

12.6 Further Classes of Delay Equations

We now briey review continuation methods for systems that include delays
that depend on the state variables or whose right-hand side includes delayed
derivatives of the state.
12.6.1 State-Dependent Equations
We briey describe how a DDE with state-dependent delay (sd-DDE) can be
handled; we assume for simplicity that only one delay is present. An sd-DDE
can be of the form
 d
dt x(t) = f1 (x(t), x(t (x(t)), ),
(12.35)
(x(t)) = g1 (x(t)),

where g1 : Rn R is a given (explicit) function of the solution x(t), or it can

be of the form
 d
dt x(t) = f2 (x(t), x(t (t)), (t)),
(12.36)
d
dt (t) = g2 (x(t), x(t (t)), (t)),

where g2 : Rn Rn R R, and the delay is determined by a dierential

equation. We assume that all functions in (12.35) and (12.36) are suciently
smooth and that the delay is bounded, i.e., 0 (t) r, t. Note that, using
x1 x and x2 , (12.36) can be considered as a particular case of (12.35)
with the extended state x (x1 , x2 ).
A steady-state solution of an sd-DDE is determined by the state x and the
delay , i.e., the delay should be considered as a part of the solution. A steadystate solution (x , ) of (12.35) or (12.36) can be computed by solving a
(nonlinear) algebraic system. The local stability of steady-state solutions of sdDDEs was studied in [15, 41]. It was shown, under natural assumptions on the
right-hand side of the equation and on the delay function , that generically
the behavior of the state-dependent delay , except for its value , has no
eect on the stability, and that in the local linearization can be treated as
a constant. Hence, to study the local stability of a steady state of (12.35)
or (12.36), these equations can be linearized at x by setting . The
resulting linearized equation is a DDE with constant delay, and the numerical
procedures discussed in Sect. 12.1 can be used without changes [57].
The existence of periodic solutions for particular cases of sd-DDEs has been
studied by several authors, in particular the existence of slowly oscillating
periodic solutions. The theory suggests that a Hopf bifurcation theorem holds;
see, e.g., [61]. The stability of periodic solutions has only recently been studied;
see, e.g., [42] for non-autonomous sd-DDEs. It was proven that the Frechet
derivative of the solution operator of the nonlinear sd-DDE with respect to
initial data equals the solution operator of the linearized equation. Based on
these results (12.35) and (12.36) can be linearized around a (nonconstant)
solution (x (t), (t)) as follows. Let Dj fi denote the derivative of f1 with
respect to its jth argument, then

12 Bifurcation Analysis of Delay Dierential Equations

d
dt y(t)

d
(x (t))y(t)
x (t (x (t))) x
= D1 f1 (s)y(t) D2 f1 (s) dt

+D2 f1 (s)y(t (x (t))),

379

(12.37)

with s = (x (t), x (t (x (t)))), respectively, and

d
d
x (t (t))y2 (t)
y1 (t) = D1 f2 (s)y1 (t)+D2 f2 (s)y1 (t (t))D2 f2 (s) dt

dt
+D3 f2 (s)y2 (t),
d
d
x (t (t))y2 (t)
y
(t)
=
D1 g2 (s)y1 (t)+D2 g2 (s)y1 (t (t))D2 g2 (s) dt

dt
+D3 g2 (s)y2 (t),

with s = (x (t), x (t (t)), (t)). These linearized equations contain a timedependent (no longer state-dependent) delay. If the coecients in the linear
equation are smooth and periodic (with period T ) and the delay function is
smooth, then the solution operator over the period T (over an interval mT if
m > T and mT m , m = maxt[0,T ] (t)) is compact [38].
A periodic solution can be computed by solving a two-point boundary
value problem in time, similar to (12.13), but in the case of (12.36) the additional equation (0) = (T ) must be imposed. The solution of these boundary
value problems by collocation and the computation of the Floquet multipliers
is conceptually equal to the procedure outlined in Sect. 12.2; see [57].
12.6.2 Collocation Schemes for Equations of Neutral Type
We summarize basic results on two collocation schemes that were proposed
in Barton et. al. [6]. Here we consider the simple equation of neutral type
x(t)

= f (x(t), x(t ), x(t

), ).

(12.38)

The collocation scheme of Sect. 12.2.1 discretizes (12.38) by substituting the

collocation polynomials and evaluating at the collocation points. In the Jacobian matrix of this discretized system the second derivatives of the polynomials appear. This reduces the accuracy by an order, which is only O(hm ).
This drop in the order of convergence is apparent in the examples of [6]. To
remedy the situation (12.38) can be transformed into an ODE coupled to a
dierence equation
x(t)

= y(t)
y(t) = f (x(t), x(t ), y(t );

(12.39)
(12.40)

see [6]. Applying the collocation scheme of Sect. 12.2.1 to this system does
not introduce second-order derivatives in the Jacobian matrix and, hence, a
better convergence can be expected. The numerical experiments in [6] show
a convergence rate of O(hm+1 ). In [6] the Gauss-Legendre points were used
in the collocation scheme, together with a periodic boundary condition on
the algebraic part, but other approaches are possible for delay dierential
algebraic equations.

380

Dirk Roose and Robert Szalai

12.7 Software Packages

Several software packages exist for simulation (time integration) of delay differential equations, including ARCHI [66], DKLAG6 [16], RADAR5 [36]
and XppAut [29]. Furthermore, Matlab now contains the solver dde23 [71].
Probably the earliest computer program specically designed for DDEs has
been published by Hassard [43], namely BIFDD which allows a normal-form
analysis of Hopf bifurcation points. XppAut by Ermentrout [29] allows a limited stability analysis of steady-state solutions of DDEs using the approach
described in [58].
By contrast, the software packages DDE-Biftool and PDDE-Cont implement numerical continuation of DDEs as introduced in the previous sections. In this section we describe the functionality of these numerical tools.
12.7.1 DDE-Biftool
The package DDE-Biftool consists of a collection of Matlab-routines for
the numerical continuation and bifurcation analysis of systems of DDEs with
multiple discrete delays, which may be xed or state-dependent; for detailed
instructions we refer to the user manual [27]. This software allows one to
compute branches of steady-state solutions and steady-state fold and Hopf
bifurcations with continuation. Given an equilibrium, it allows one to approximate the right-most, stability-determining roots of the characteristic equation, which can be further corrected with Newtons method. Periodic solutions
and their Floquet multipliers can also be computed by collocation with adaptive mesh selection. Branches of periodic solutions can be continued starting
from a previously computed Hopf point or from an initial guess of a periodic solution prole. For DDEs with constant delays, connecting orbits (both
homoclinic and heteroclinic solutions) can also be computed. The numerical
methods that are used in the software are as detailed in the previous sections.
DDE-Biftool has no graphical user interface, but a number of routines
are provided to plot solution, branch and stability information. Furthermore,
automatic detection of bifurcations is not supported. Instead, the evolution
of the characteristic roots or the Floquet multipliers can be computed along
solution branches, which allows the user to detect and identify bifurcations
using an appropriate visualization. Starting points for branch switching at
bifurcations on branches of steady-state and periodic solutions can be generated, as well as starting solutions for homoclinic solutions close to periodic
solutions.
Several extensions or add-ons have been developed. We mention here
a Mathematica program written by Pieroux that allows the automatic generation of the system denition les with symbolically obtained derivatives,
software written by Green for the computation of one-dimensional unstable
manifolds in DDEs [34], and the extension by Barton for equations of neutral
type [6].

12 Bifurcation Analysis of Delay Dierential Equations

381

12.7.2 PDDE-Cont
PDDE-Cont implements the numerical methods described in Sect. 12.2. It
is written in C++ with the use of linear algebra packages UMFPACK [17],
LAPACK [1] and ARPACK [54]. The software has a command line interface
and a graphical user interface together with a basic plotting facility.
PDDE-Cont can continue periodic solutions of delay equations that are
in the form
y  (t) = f (t, y(t), y(t 1 (t)), y(t 2 (t)), . . . , y(t m (t)), ).
The right-hand side f and the delays j can be either T -periodic or time
independent. The software does not have any algorithms to continue equilibria apart from the obvious fact that an equilibrium can be considered as a
constant periodic solution. Bifurcations of periodic solutions can be continued in two parameters by using test functions as described in Sect. 12.3, but
PDDE-Cont cannot switch branches automatically. For detailed instructions
see the user manual [78]. Note that PDDE-Cont can be used together with
DDE-Biftool by converting the results between the two packages.
Due to the implementation in C++, the performance of PDDE-Cont
is signicantly better than that of DDE-Biftool (which is implemented
in Matlab). Furthermore, PDDE-Cont uses sparse-matrix algorithms that
require less memory, so that problems of relatively high dimension can be
tackled. The resulting large bordered linear systems (see Sect. 12.2.1) are
solved by using bordering techniques from [31, 32]. The large sparse matrix
without borders is factorized by UMFPACK and the whole system is solved
using the BEMW method [32].

12.8 Examples of Numerical Bifurcation Analysis of

DDEs
In this section we illustrate the performance of the numerical techniques described in the previous sections with examples of DDE models of a number
of physical and biological phenomena.
12.8.1 DDE-PDE Model of a Laser with Optical Feedback
A longitudinally single-mode semiconductor laser subject to conventional optical feedback and lateral carrier diusion can be modeled by the hybrid DDEPDE system
dA(t)
= (1 i)A(t)(t) + A(t )ei ibA(t),
dt
2 Z(x, t)
Z(x, t)
Z(x, t) + P (x)
=d
T
x2
t
F (x)(1 + 2Z(x, t))|A(t)|2 .

(12.41)

(12.42)

382

Dirk Roose and Robert Szalai

3.4

Intensity (|A| )

3.35

3.3

3.25

3.2

x 10

Fig. 12.4. Bifurcation diagram in the plane of intensity |A|2 vs. feedback strength
of steady-state solutions of (12.41)(12.42) for = 3, = 0, T = 1000, d =
1.68 102 and = 1000. Stable solutions are drawn as solid curves and unstable
solutions as dashed curves; also shown are saddle-node bifurcations () and Hopf
bifurcations (*). From K. Verheyden, K. Green, and D. Roose, Numerical stability
analysis of a large-scale delay system modeling a lateral semiconductor laser subject
c 2004 by the American
to optical feedback, Phys. Rev. Lett. 69(3) (2004) 036702 
Physical Society; reprinted with permission.

Here the complex scalar variable A(t) represents the amplitude of the electrical eld E(t) = A(t)eibt , and the real variable Z(x, t), x [0.5, 0.5], represents the carrier density [82]. The functions (t), P (x) and F (x) are specied
in [82]. Continuous-wave solutions, called external cavity modes (ECMs) can
be computed as steady-state solutions of (12.41)(12.42), augmented with a
scalar condition for the unknown b and an extra scalar constraint to remove
the S 1 -symmetry. Zero Neumann boundary conditions for Z(x, t) are imposed
at x = 0.5. In the computations the time variable is rescaled by a factor
of 1000 so that most quantities in the computation are of order one. The
symmetry about x = 0 is exploited by considering only the interval [0, 0.5].
Splitting (12.41) into real and imaginary parts and discretizing (12.42) in
space with a second-order central dierence formula with constant stepsize
x = 0.5/128 results in a DDE system of size n = 131.
Figure 12.4 shows the bifurcation diagram of steady-state solutions of
(12.41)(12.42) with = 3, = 0, T = 1000, d = 1.68 102 and = 1000,
obtained by continuation with DDE-Biftool, with the feedback strength
as the parameter. The diagram shows several branches of steady-state solutions arising from saddle-node bifurcations. During continuation the right-

12 Bifurcation Analysis of Delay Dierential Equations

383

Imaginary axis

60
40
20
0
20
40
60
6

Real axis
Fig. 12.5. Characteristic roots at the Hopf point of (12.41)(12.42) 2.5717
103 from Fig. 12.4. Shown are approximations of characteristic roots with real part
larger than r = 1 derived from the eigenvalues of the discretization of the solution
operator by the sixth-order special-purpose LMS method (+), and their corrections
by Newtons method (). From K. Verheyden, T. Luzyanina, and D. Roose, Ecient
and reliable stability analysis of solutions of delay dierential equations, Proceedings
c
of 2006 International Conference on Nonlinear Science and Complexity, 109120 
2007 by World Scientic Publishing; reprinted with permission.

most characteristic roots are computed and monitored, allowing for the detection of Hopf bifurcation points along these branches.
Figure 12.5 shows the characteristic roots at the moment of the rst Hopf
bifurcation on the middle branch at 2.5717 103 . Since the imaginary
part of the right-most pair of characteristic roots is large, the system presents
a challenging test case for characteristic root calculation with DDE-Biftool.
Approximations to the characteristic roots were obtained by computing
the eigenvalues of the matrix approximation to the solution operator with a
sixth-order LMS method, optimized to retain the stability properties of the
linearized equation. The steplength h in the LMS method is automatically
determined to ensure that all characteristic roots with real part larger than
r = 1 (threshold specied by the user) are approximated accurately. This
leads to the discretization of the delay interval with an equidistant mesh
of 27 points. The resulting eigenvalue problem has dimension 131 27 =
3537, which is large but can still be solved by using the QR-method. These
approximations are subsequently corrected by Newtons method applied to
(12.8). The approximate characteristic roots shown in Fig. 12.5 were derived
from the eigenvalues of the discretization of the solution operator, and their
corrections by Newtons method.
For this example a comparison of the computation of the characteristic
roots using the pseudo-spectral discretizations of the innitesimal generator
and of the solution operator is presented in [85]. In both cases, a polynomial
of degree p = 32 is used, so that the linear eigenvalue problems have size

384

Dirk Roose and Robert Szalai

Table 12.1. The computational cost of four algorithms based on a pseudo-spectral

discretization of the innitesimal generator A and the solution operator S(h) when
using polynomials of degree p = 32 to nd the right-most characteristic roots shown
in Fig. 12.5.

CPU time (seconds)

# matrix-vector products

Right-most
106.2
6528

Shift-Invert
94.1
4951

Forward
103.4
6254

S(h)
Backward
55.5
3146

n(p + 1) = 131 33 = 4323. Table 12.1 shows the computational cost of the
four methods.
To solve the linear eigenvalue problem resulting from the pseudo-spectral
discretization of the innitesimal generator, the Matlab function eigs function
is used to compute the right-most 30 eigenvalues with a requested tolerance
of 108 (results indicated with Right-most). Note that eigs uses Arnoldis
method with implicit restart, and this method does not require the explicit
construction of the matrix. For the results indicated with Shift-Invert, eigs is
used in conjunction with the shift-invert technique and returns the eigenvalues
closest to the shift A0  + A1  4528.5, as proposed in [10]. The pseudospectral discretization of the solution operator S(h) leads to two algorithms,
called forward and backward variants in [85]. The steplength h is chosen to
be 104 for the forward and 103 for the backward variant, respectively.
The accuracy of the computed characteristic roots is similar for the four
methods. For example, the roots 0.285 i11.8 are computed by the four
algorithms with a relative error between 6.5 1012 and 2.4 1014 . The accuracy
is lower for the eigenvalues with large imaginary part (the relative error on
the purely imaginary eigenvalues i47.7 is 104 for the backward variant,
and 7 106 for the three other algorithms. The exponential convergence
with respect to the degree p has been conrmed by numerical experiments.
12.8.2 The Mackey-Glass Equation
The equation
x(t)

= ax(t) + b

x(t )
1 + x10 (t )

(12.43)

models the regeneration of white blood cells [59], and it is today widely known
as the Mackey-Glass equation. Although it is a simple equation, not much is
known about its solution structure.

The three equilibria of (12.43), i.e., x1 = 0 and x2,3 = 10 (a + b)/a are
connected to each other at a = b by a supercritical pitchfork bifurcation.
The nonzero solutions can lose their stability in a Hopf bifurcation along the
curves in parameter space given by

12 Bifurcation Analysis of Delay Dierential Equations

385

Fig. 12.6. Bifurcation diagram showing period-two solutions of (12.43) for xed a =
1.2158. Fold bifurcations are denoted by dots while period-doubling bifurcations
are denoted by +; the numbers of unstable characteristic multipliers are indicated
along the dierent branches.

1
,
a = arccos(d1 )
d2 1
10a
,
b=
d9

where |d| 1. Hopf bifurcations for d > 1 are supercritical, so they give rise to
stable periodic solutions. These periodic solutions bifurcate further via several
period doublings, which then leads to chaotic motion. It was demonstrated in
[39] that chaos arises due to the transverse intersection of the two-dimensional
unstable and innite-dimensional stable manifolds of this periodic solution.
We remark that some square-shaped solutions of large period can be obtained
by singularly perturbing a map to give
x(t)

= ax(t) + b

x(t 1)
,
1 + xc (t 1)

where 0 and = 1; see [60] for details.

Here we analyze the period-two solutions bifurcating from the period-one
solutions that in turn can be related to the supercritical Hopf bifurcation of
the equilibrium. These solutions form a complicated branch structure that is
challenging to compute. In Fig. 12.6 the bifurcation diagram for a = 1.2158

386

Dirk Roose and Robert Szalai

Fig. 12.7. The structure of the period-two solutions of (12.43) over the (a, b)plane; shown are solution branches (black curves), their fold bifurcations (green
dots), period-doubling bifurcations of the period-one solution (blue curves) and fold
bifurcations of period-two solutions (red curves).

and = 2 is shown. As b varies the solution undergoes several fold and

period-doubling bifurcations. These solutions are almost all unstable, and so
they cannot be found by simulation; the number of unstable characteristic
multipliers is shown along the branches in Fig. 12.6. Furthermore, the periodone solution branch is included, but with twice the period so that it matches
up with the branch of period-two solutions.
By investigating the fold bifurcations and computing several branches of
solutions we can obtain a fairly complete picture of the structure of periodic
solutions. Figure 12.7 shows this structure for the xed delay = 2, where
we plot the period T of the solutions as a function of the parameters a and b.
12.8.3 Trac Model with Driver Reaction Time
The trac model in Orosz et al. [64, 65] describes the dynamics of N cars
on a circular track. Each car has a velocity vi and an associated headway hi
dened as the distance to the car in front. The headways hi are calculated
from the velocities as
(12.44)
h i (t) = vi+1 (t) vi (t).
Because of the circular track, we assume that vN +1 = v1 and hN = L
&N 1
i=1 hi . Each car tries to reach its optimal velocity, which is a function of
the headway that can be expressed as
v i (t) = (V (hi (t 1) vi (t)),

(12.45)

where is the sensitivity to velocity dierences. Due to the reaction time of

the driver, a delayed value of the headway is used in the model. The optimal

12 Bifurcation Analysis of Delay Dierential Equations

387

Fig. 12.8. Periodic solution branches of (12.44)-(12.45). Unstable solutions are

denoted by dashed lines, continous lines refer to stable solutions and boxes denote
fold bifurcations.

velocity is a function of the headway, and it is modeled by the optimal velocity

function

0
0 h 1,
3
V (h) =
0 (h1)
v 1+(h1)3 h > 1.

By making use of the algebraic condition for hN , one can reduce the dimension
of system (12.44)(12.45) by one to 2N 1.
In this section we consider N = 17 cars, which is the largest number of
cars that was considered and (partially) analyzed in [64]. Our starting point is
the steady-state solution of the model, which corresponds to equal headways
and equal car velocities and so is given by
hi = L/N,

vi = V (hi ).

The steady state undergoes several Hopf bifurcations from which branches
of periodic solutions arise; they are shown in Fig. 12.8 as computed with
PDDE-Cont for typical parameter values of = 1 and = 2 as a function
of the average headway L/N . Note how all branches of periodic solutions
feature folds and connect pairs of subcritical Hopf bifurcations. The outermost branch is stable between the folds, which shows that there is bistability
between stable periodic solutions (indicating a trac jam) and the stable
steady state (uniform ow of cars). The other branches of periodic solutions
remain unstable throughout, but the outer-most of them has all its unstable
Floquet multipliers very close to 1 (for l/N around 2), which means that the

388

Dirk Roose and Robert Szalai

Fig. 12.9. Hopf and fold bifurcation curves of (12.44)(12.45). The gray regions
are bistable, where a stable equilibrium coexists with a stable periodic solution.
Between the gray regions the equilibrium is unstable, while outside the gray regions
the equilibrium is globally stable.

corresponding periodic solution can be observed as long transients. Physically,

these transients give rise to trac jams, which move towards each other and
eventually either merge with the stable trac jam or disperse [64].
Figure 12.9 shows the curves of fold bifurcations for N = 17 cars in
the (L/N, )-plane. The plot also shows some Hopf bifurcation curves (dashdotted lines) and points of degenerate Hopf bifurcations (+); the regions of
bistability are highlighted in gray. In [64] a similar image was computed for
N = 9 cars with DDE-Biftool by performing one-parameter continuations
in L/N for many values of xed to nd the fold bifurcations. (The locus
of Hopf bifurcations is actually known analytically.) As this approach is very
time consuming, we used PDDE-Cont instead, which is able to follow the
fold bifurcation curves directly in two parameters.
12.8.4 Chatter Motion in Milling
Cutting processes are often subject to the so-called regenerative eect [77],
which comes from the fact that a cutting tool always cuts a surface that was
produced by the same tool some time ago. The cutting forces nonlinearly
depend on the chip geometry, which in turn depends on the current and a
delayed tool position. The underlying dynamics of the tool can be considered
to be linear and, hence, the nonlinearity comes from the geometry of the chip
forming and the cutting force only. There is a vast literature on the dynamics
of machining that mainly focuses on the stability of steady cutting; see, e.g.,

12 Bifurcation Analysis of Delay Dierential Equations

389

Fig. 12.10. Stability chart of (12.46) in the cutting speed and chip width parameter
plane (a); the relative damping is = 0.0038 and the tool cuts continuously for a
time of 10.82% of every period. Panel (b) shows a bifurcation diagram for xed
cutting speed; the fold bifurcations of the periodic solutions are due to the nonsmooth dynamics of the system.

[47]. However, there are only a few papers on the nonlinear dynamics and they
employ either analytical methods [76] or simulation [4].
Machining processes are inherently nonsmooth, because there is the possibility of a loss of cutting force when the tool leaves the work piece. This poses
some challenges, although in some cases one can approximate the equations of
motion with a smooth system. In the case of turning, which is an autonomous
process, DDE-Biftool was used in [23]. Here we summarize the results in
[79], where a milling problem was investigated with PDDE-Cont.
The equation of motion of the nonsmooth milling problem reads
x
(t) + 2 x(t)
+ x(t) = g(t)w(cos

2t/T + 0.3 sin 2t/T )

[H(1 + x(t 2T ) x(t T ))Fc ((1 + x(t T ) x(t)) sin 2t/T )
+H(x(t T ) x(t 2T ) 1)Fc ((2 + x(t T ) x(t)) sin 2t/T )] ,

(12.46)

where Fc is a nonlinear cutting force function, usually modeled with the power
3/4

, and H is the Heaviside function. The function g is a

law Fc (x) = 4w/3x
T -periodic windowing function that changes its value once in a period between 0 and 1 depending on whether the tool is cutting the material. The two
important parameters are the period T , which is inversely proportional to the
spindle speed and the dimensionless chip width w.

In order to conduct a numerical bifurcation analysis of the system with

PDDE-Cont, the Heaviside function H(z) is replaced by the smoothed function (1 + tanh(Cz))/2 with a suciently large value of C. Equation (12.46)
has a unique T -periodic solution which represents steady cutting and can lose
its stability either at a Neimark-Sacker or at a period-doubling bifurcation.
Figure 12.10(a) shows the bifurcation diagram where these bifurcation curves

390

Dirk Roose and Robert Szalai

(a)

(b)

Fig. 12.11. A branch of quasiperiodic solutions for a cutting speed of 2/T = 0.4
(a) and an invariant torus (b) for a point on the branch just before the tool leaves
the work piece and (12.46) becomes invalid.

are shown as solid curves. The period doublings may be subcritical or supercritical; see the bifurcation diagram for a xed cutting speed in Fig. 12.10(b).
Fold and Neimark-Sacker bifurcation curves of the period-two solutions have
been continued in two parameters, and they are shown in Fig. 12.10(a) as
dashed curves. These numerical results were compared to experimental data
in [79].
Quasiperiodic solutions arising at a Neimark-Sacker bifurcation can be
computed with the technique described in Sect. 12.5. A branch of invariant
quasiperiodic tori was continued with PDDE-Cont until the model loses
2
its physical validity. During the continuation the rotation number
1 was
kept constant and w
and T served as free parameters. The resulting curve of
quasiperiodic solutions is shown in Fig. 12.11(a). Since T varies only slightly

12 Bifurcation Analysis of Delay Dierential Equations

391

during the continuation, the dependence on the period is not shown in the
bifurcation diagram. One of the invariant tori along the branch (near where
the model loses its validity) is shown in Fig. 12.11(b). The computation of further quasiperiodic solutions reveals that this system has very narrow Arnold
tongues in the region above the Neimark-Sacker curve in Fig. 12.10(a).
12.8.5 A Laser with Filtered Optical Feedback
One main objective for studying laser dynamics is to nd regions of parameter
values where a constant-amplitude coherent light is produced. In many laser
systems delay is an important feature. It arises due to the nite travel time
of light between components of the system and may lead to dierent types of
dynamic behavior including chaos; see, e.g., [49]. The numerical tools introduced in this chapter are very well suited to the study of nonlinear dynamics
in lasers with delayed optical feedback; see also [51].
In this section we summarize some results of Erzgraber et. al. [30], who investigate a DDE model of a semiconductor laser with ltered optical feedback
of the form
dE
= (1 + i)N (t) + F (t),
dt
dN
= P N (t) (1 + 2N (t))|E(t)|2 ,
T
dt
dF
= E(t )eiCp + (i )F (t),
dt

(12.47)

(12.48)

(12.49)

where the variable E is the complex optical eld, N is the (real-valued) population inversion of the laser, and F is the complex optical eld of the lter.
The material properties of the laser are given by the linewidth enhancement
factor , the electron lifetime T and the pump rate P . The coupling of the
laser with the lter is controlled by the parameter , while is the time that
the light spends in the external feedback loop. The dynamics of the lter is
modeled by (12.49). The feedback phase Cp is the phase dierence between
the laser and the lter elds, and is the detuning between the lter center
frequency F and the solitary frequency 0 of the laser. Hence, Cp = 0
and = F 0 .
The laser equations (12.47)(12.49) exhibit a rotational symmetry (rotation of both E and F over any angle) that is important for the types of
solutions that are supported. It also needs to be dealt with in the continuation to ensure that solutions are isolated. The idea is to consider solutions of
the form
(E(t), N (t), F (t)) = (A(t) eibt , N (t), B(t) eibt ).
By putting this ansatz into (12.47)(12.49) we obtain the new system

392

Dirk Roose and Robert Szalai

dA
= (1 + i)N (t)A(t) ibA(t) + B(t),
dt
dN
= P N (t) (1 + 2N (t))|A(t)|2 ,
T
dt
dB
= A(t )ei(Cp +b ) + (i ib)B(t).
dt

(12.50)

(12.51)

(12.52)

Note that the stability of this transformed system does not dier from the
the stability of (12.47)(12.49), because the norm of (E(t), N (t), F (t)) is the
same as the norm of (A(t), N (t), B(t)). System (12.50)(12.52) still has the
same rotational symmetry, but the equations are now in a form that can be
dealt with in continuation.
The primary interest is in the so-called external ltered modes (EFMs),
which are single-frequency periodic solutions of (12.50)(12.52) that are characterized by xed A(t) = As , N (t) = Ns and B(t) = Bs . EFMs were extensively studied analytically [63] and with numerical continuation [30]. In
order to determine an EFM uniquely one needs to x the phase, which can
be done, for example, by setting Re(Es ) = 0 and treating b as a variable.
Figure 12.12(a) shows a bifurcation diagram in the (, Cp )-plane that was
computed with DDE-Biftool. EFMs are stable in the green region; they
are born in saddle-node bifurcations (blue curves) and lose their stability in
Hopf bifurcations (red curves).
At Hopf bifurcations periodic solutions arise whose continuation requires
a new phase condition [37]. Let us introduce the symmetry group

i
i00
e 0 0
dG()
|=0 = 0 0 0 ,
G() = 0 1 0 , G =
d
00i
0 0 ei

which produces a two-parameter family of solutions

u(t; 1 , 2 ) = G(1 )(A(t + 2 ), N (t + 2 ), B(t + 2 ))
from any solution of (12.50)(12.52). In a continuation context, when looking
for the next solution on a branch of solutions, we want to nd the one closest
in norm to the previous solution u(0) . Hence, the new solution u is chosen to
minimize
 1
D(1 , 2 ) =
u(t; 1 , 2 ) u(0) (t))22 dt
0

in 1 , 2 . Dierentiating with respect to both variables and evaluating at 1 =

2 = 0 yields
 1
u (0) (s)(u(0) (s) u(s)) ds = 0,
(12.53)


0
1

Gu(0) (s)(u(0) (s) u(s)) ds = 0,

0

(12.54)

12 Bifurcation Analysis of Delay Dierential Equations

4
Cp (a)

(c)

SL
SL

L 0
T

(b)

(d)
FO

IF 2

SL

0
3

SL

DH

1:1

4
0

0.01

0.02

(c3)

F 0

(c2)

3
4

RO

EFM

(c1)

0
3

1:1

DH

5
IL

T
FO

393
.

3
0
5

0.03

5
IL

(b1)

(c4)

20

40

t[ns]

IL

60

(d1)

0
15

0
3

L 0

(b2) L 0
3
4

(d2)

15
4
IF 2

(b3) IF 2

(d3)

0
3

0
3

F 0

3
. 0

4 t[ns]

(b4) F 0
3
5
0

(d4)

20

40

t[ns]

60

Fig. 12.12. Panel (a) shows the bifurcation diagram in the (, Cp )-plane of (12.47)
(12.49) for = 0, = 5.0, T = 100, P = 3.5, = 500 and = 0.007. EFMs are
stable in the green region, which is bounded by curves of saddle-node bifurcations
(blue) and Hopf bifurcations (red). Panels (b1)(b4) show an example of relaxation
oscillations, while panels (c1)(c4) and panels (d1)(d4) are examples of frequency
oscillations; plotted are the laser intensity IL , its frequency L , the lter intensity
IF , and its frequency F . The stability regions of the dierent oscillations are shown
in panel (a) in orange, purple and light blue, respectively. From H. Erzgr
aber, B.
c 2007 by
Krauskopf and D. Lenstra, SIAM J. Appl. Dyn. Sys. 6(1) (2007) 128 
the Society for Industrial and Applied Mathematics; reprinted with permission.

where G = d/dG()|=0 is the innitesimal generator of the symmetry group.

Note that (12.53) is actually phase condition (12.14), while (12.54) is a new
phase condition that xes the group invariance.
With (12.53) and (12.54) periodic solutions can be continued as isolated
solutions. These phase conditions are implemented in both DDE-Biftool
and PDDE-Cont. In [30] DDE-Biftool was used to compute the periodic
solutions of (12.50)(12.52) and PDDE-Cont was used to determine their
stability boundaries by continuing the Neimark-Sacker bifurcation in two parameters.
The resulting stability regions of the two dierent types of periodic solutions are colored in Fig. 12.12(a), and examples of typical time series are

394

Dirk Roose and Robert Szalai

shown in panels (b)(d). First, there are the typical relaxation oscillations (not
to be confused with relaxation oscillations of slow-fast systems as discussed
in Chap. 8), which are a periodic exchange of energy between the electric
eld E and the inversion N in a semiconductor laser. Relaxation oscillations
are fast (on the order of a few GHz) and eectively do not involve the lter;
see Fig. 12.12(b). The other type of oscillations are the frequency oscillations, which are slower and oscillate on the time scale given by the external
roundtrip time (that is, the delay ); see Fig. 12.12(c) and (d). These oscillations are unusual for semiconductor lasers because they feature practically
constant laser intensity IL but an oscillating frequency L . Notice that the
dynamics of the lter appears to suppress the dynamics of the intensity. Both
types of oscillations lose their stability at Neimark-Sacker bifurcations, which
are shown as black curves in Fig. 12.12(a).

12.9 Conclusions
We discussed numerical continuation methods for the stability and bifurcation
analysis of delay dierential equations with constant delays, concentrating on
techniques concerning steady-state solutions and periodic solutions. We also
described how to compute connecting (homoclinic and heteroclinic) orbits
and quasiperiodic solutions. Furthermore, we briey mentioned how to deal
with state-dependent delays and with equations of neutral type. Compared
with numerical methods for such tasks in ordinary dierential equations the
methods we presented are either similar but with a higher computational
cost (an example is collocation for computing periodic solutions) or much
more complex (as is the case for computing the stability of a steady state or
nding a connected orbit). These additional diculties are due to the innitedimensional nature of DDEs.
Rather than trying to give a complete literature survey, we focused on the
numerical methods implemented in the software packages DDE-Biftool and
PDDE-Cont. Both have about the same functionality as similar packages
for ODEs, but with less exibility and at a higher computational cost. They
make continuation and bifurcation analysis for DDEs readily available for
scientists dealing with concrete problems arising in applications. We have
included results on the continuation and bifurcation analysis of several realistic
models to illustrate the applicability of the methods.
Numerical developments can also help with the solution of some open
theoretical problems. For example, some numerical results on state-dependent
DDEs are ahead of the theory and suggest that certain conditions imposed
in the theory are rather technical and not fundamental. One of the areas
for future work for both theory and numerical methods is that of piecewisesmooth delayed systems, which have important applications, for example, in
control theory [5, 73], hybrid testing [53, 74] and machine tooling [23, 77].

12 Bifurcation Analysis of Delay Dierential Equations

395

Acknowledgements
We thank Bernd Krauskopf, David Barton, Tatyana Luzyanina, Giovanni
Samaey and Koen Verheyden for their useful comments and suggestions. We
are grateful to the American Physical Society, World Scientic Publishing
and the Society for Industrial and Applied Mathematics for permission to
reproduce Fig. 12.4, Fig. 12.5 and Fig. 12.12, respectively.

References
1. E. Anderson, Z. Bai, C. Bischof, S. Blackford, J. Demmel, J. Dongarra,J. Du Croz, A. Greenbaum, S. Hammarling, A. McKenney, and
D. Sorensen. LAPACK Users Guide, third edition. (SIAM, Philadelphia, 1999).
2. U. M. Ascher, R. M. M. Mattheij, and R. D. Russell. Numerical Solution of
Boundary Value Problems for Ordinary Dierential Equations. (Prentice Hall,
1988).
3. C. T. H. Baker, C. A. H. Paul, and D. R. Wille. A bibliography on the numerical
solution of delay dierential equations. Technical Report 269, University of
Manchester, Manchester Centre for Computational Mathematics, 1995.
4. B. Balachandran. Non-linear dynamics of milling process. Trans. Royal Soc.,
359:793820, 2001.
5. D. A. W. Barton, B. Krauskopf, and R. E. Wilson. Explicit periodic solutions in
a model of a relay controller with delay and forcing. Nonlinearity, 18(6):2637
2656, 2005.
6. D. A. W. Barton, B. Krauskopf, and R. E. Wilson. Collocation schemes for
periodic solutions of neutral delay dierential equations. J. Di. Eqns. Appl.,
12(11):1087-1101, 2006.
7. A. Bellen and M. Zennaro. Numerical Methods for Delay Dierential Equations.
(Oxford University Press, 2003).
8. W. J. Beyn. The numerical computation of connecting orbits in dynamical
systems. IMA J. Numer. Analysis, 9:379405, 1990.
9. W. J. Beyn, A. Champneys, E. J. Doedel, W. Govaerts, B. Sandstede, and
Y. A. Kuznetsov. Numerical continuation and computation of normal forms. In
B. Fiedler, editor, Handbook of Dynamical Systems, pages 149219. (Elsevier,
2002).
10. D. Breda. Numerical computation of characteristic roots for delay dierential
equations. PhD thesis (Department of Mathematics and Computer Science,
University of Udine, 2004).
11. D. Breda. Solution operator approximation for delay dierential equation characteristic roots computation via Runge-Kutta methods. Appl. Numer. Math.,
56:305317, 2006.
12. D. Breda, S. Maset, and R. Vermiglio. Computing the characteristic roots for
delay dierential equations. IMA J. Numer. Analysis, 24:119, 2004.
13. D. Breda, S. Maset, and R. Vermiglio. Pseudospectral dierencing methods
for characteristic roots of delay dierential equations. SIAM J. Sci. Comput.,
27(2):482495, 2005.
14. A. M. Castelfranco and H. W. Stech. Periodic solutions in a model of recurrent
neural feedback. SIAM J. Appl. Math., 47(3):573588, 1987.

396

Dirk Roose and Robert Szalai

15. K. L. Cooke and W. Huang. On the problem of linearization for state-dependent

delay dierential equations. Proc. Am. Math. Soc., 124(5), 1996.
16. S. P. Corwin, D. Sarafyan, and S. Thompson. DKLAG6: A code based on
continuously imbedded sixth order Runge-Kutta methods for the solution of
state dependent functional dierential equations. Appl. Numer. Math., 24(2
3):319330, 1997.
17. T. A. Davies. UMFPACK Version 4.1 User Guide. Technical report, Department of Computer and Information Science and Engineering, University
of Florida, Gainesville, FL, USA, 2003. http://www.cise.ufl.edu/research/
sparse/umfpack/.
18. E. J. Doedel, A. R. Champneys, T. F. Fairgrieve, Y. A. Kuznetsov, B. Sandstede, and X. Wang. Auto97: Continuation and bifurcation software for ordinary dierential equations, 1997; available via ftp.cs.concordia.ca in directory
pub/doedel/auto.
19. E. J. Doedel, M. J. Friedman, and B. I. Kunin. Successive continuation for
locating connecting orbits. Num. Alg., 17:103124, 1997.
20. E. J. Doedel, W. Govaerts, and Yu. A. Kuznetsov. Continuation of periodic
solution bifurcations in odes using bordered systems. SIAM J. Numer. Analysis,
42(2):401435, 2003.
21. E. J. Doedel, H. B. Keller, and J. P. Kernevez. Numerical analysis and control
of bifurcation problems (II): bifurcations in innite dimensions. Internat. J.
Bifur. Chaos Appl. Sci. Engrg., 1(4):745772, 1991.
22. E. J. Doedel and P. P. C. Leung. A numerical technique for bifurcation problems
in delay dierential equations. Congr. Numer., 34:225237, 1982. (Proc. 11th
Manitoba Conf. Num. Math. Comput., Univ. Manitoba, Winnipeg, Canada).
23. Z. Domb
ov
ari. Bifurcation analysis of a cutting process. Masters thesis (Budapest University of Technology and Economics, 2006).
24. K. Engelborghs and E. Doedel. Stability of piecewise polynomial collocation
for computing periodic solutions of delay dierential equations. Numer. Math.,
91(4):627648, 2002.
25. K. Engelborghs, T. Luzyanina, K. J. in t Hout, and D. Roose. Collocation
methods for the computation of periodic solutions of delay dierential equations.
SIAM J. Sci. Comput., 22(5):15931609, 2000.
26. K. Engelborghs, T. Luzyanina, and D. Roose. Numerical bifurcation analysis of
delay dierential equations using DDE-Biftool. ACM Trans. Math. Software,
28(1):121, 2002.
27. K. Engelborghs, T. Luzyanina, and G. Samaey. DDE-Biftool v. 2.00: a Matlab
package for bifurcation analysis of delay dierential equations. Technical Report TW-330, Department of Computer Science, K.U.Leuven, Leuven, Belgium,
2001; available at www.cs.kuleuven.ac.be/cwis/research/twr/research/
software/delay/ddebiftool.shtml.
28. K. Engelborghs and D. Roose. On stability of LMS methods and characteristic
roots of delay dierential equations. SIAM J. Numer. Analysis, 40(2):629650,
2002.
29. B. Ermentrout. XppAut3.91 - The dierential equations tool. University of
Pittsburgh, Pittsburgh, 1998; available at http://www.pitt.edu/phase/.
30. H. Erzgr
aber, B. Krauskopf, and D. Lenstra. Bifurcation analysis of a semiconductor laser with ltered optical feedback. SIAM J. Appl. Dyn. Sys., 6(1):128,
2007.

12 Bifurcation Analysis of Delay Dierential Equations

397

31. W. Govaerts. Bordered augmented linear systems in numerical continuation and

bifurcation. Numer. Math., 58:353368, 1990.
32. W. Govaerts and J. D. Pryce. Mixed block elimination for linear systems with
wider borders. IMA J. of Numer. Analysis, 13:161180, 1993.
33. K. Green, B. Krauskopf, and K. Engelborghs. Bistability and torus break-up in
a semiconductor laser with phase-conjugate feedback. Physica D, 173(1-2):114
129, 2002.
34. K. Green, B. Krauskopf, and K. Engelborghs. One-dimensional unstable eigenfunction and manifold computations in delay dierential equations. J. Comp.
Phys., 197(1):8698, 2004.
35. K. Green, B. Krauskopf, and G. Samaey. A two-parameter study of the locking
region of a semiconductor laser subject to phase-conjugate feedback. SIAM J.
Appl. Dyn. Sys., 2(2):254276, 2003.
36. N. Guglielmi and E. Hairer. Implementing Radau IIA methods for sti delay
dierential equations. Computing, 67:112, 2001
37. B. Haegeman, K. Engelborghs, D. Roose, D. Pieroux, and T. Erneux. Stability
and rupture of bifurcation bridges in semiconductor lasers subject to optical
feedback. Phys. Rev. E, 66:046216 111, 2002.
38. J. K. Hale. Theory of Functional Dierential Equations, volume 3 of Applied
Mathematical Sciences. (Springer-Verlag, New York, 1977).
39. J. K. Hale and N. Sternberg. Onset of chaos in dierential delay eqautions. J.
Comp. Phys., 77:221239, 1988.
40. J. K. Hale and S. M. Verduyn Lunel. Introduction to Functional Dierential
Equations, volume 99 of Applied Mathematical Sciences. (Springer-Verlag, New
York, 1993).
41. F. Hartung and J. Turi. Stability in a class of functional dierential equations
with state-dependent delay. In C Corduneanu, editor, Qualitative problems for
dierential equations and control theory, pages 1531. (World Scientic, Singapore, 1995).
42. F. Hartung and J. Turi. On dierentiability of solutions with respect to parameters in state-dependent delay equations. J. Di. Eqns., 135:192237, 1997.
43. B. D. Hassard. A code for Hopf bifurcation analysis of autonomous delaydierential systems. In F. V. Atkinson, W. F. Langford, and A. B. Mingarelli,
editors, Oscillations, Bifurcations and Chaos, volume 8 of Can. Math. Soc. Conference Proceedings, pages 447463 (Amer. Math. Soc., Providence, RI, 1987).
44. B. D. Hassard. Counting roots of the characteristic equation for linear delaydierential systems. J. Di. Eqns., 136:222235, 1997.
45. J. S. Hesthaven. Spectral penalty methods. Appl. Num. Math., 33:2341, 2000.
46. K. J. in t Hout and Ch. Lubich. Periodic orbits of delay dierential equations
under discretization. BIT, 38(1):7291, 1998.
47. T. Insperger, B. P. Mann, G. Step
an, and P. V. Bayly. Stability of up-milling
and down-milling, part 1: alternative analytical methods. Int. J. of Mach. Tool
Manuf., 43(1):2534, 2003.
48. M. A. Kaashoek and S. M. Verduyn Lunel. Characteristic matrices and spectral
properties of evolutionary systems. Trans. Amer. Math. Soc., 334:479517, 1992.
49. D.M. Kane and K.A. Shore (Eds.), Unlocking Dynamical Diversity: Optical Feedback Eects on Semiconductor Lasers, (John Wiley & Sons, 2005).
50. D. A. Kopriva and J. H. Kolias. A conservative staggered-grid chebyshev multidomain method for compressible ows. J. Comp. Phys., 125:244261, 1996.

398

Dirk Roose and Robert Szalai

51. B. Krauskopf. Bifurcation analysis of lasers with delay. In D.M. Kane and
K.A. Shore, editors, Unlocking Dynamical Diversity: Optical Feedback Eects
on Semiconductor Lasers, pages 147183. (John Wiley & Sons, 2005).
52. Yu. A. Kuznetsov and V. V. Levitin. Content: A multiplatform environment for analyzing dynamical systems, 1997. Dynamical Systems Laboratory,
Centrum voor Wiskunde en Informatica; available via ftp.cwi.nl in directory
pub/CONTENT.
53. Y. N. Kyrychko, K. B. Blyuss, A. Gonzalez-Buelga, S. J. Hogan and D. J. Wagg.
Real-time dynamic substructuring in a coupled oscillator-pendulum system.
Proc. R. Soc. A 462:12711294, 2006.
54. R. B. Lehoucq, D. C. Sorensen, and C. Yang. ARPACK Users Guide: Solution
of Large-Scale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods.
(SIAM, Philadelphia, 1998).
55. K. Lust, D. Roose, A. Spence, and A.R. Champneys. An adaptive NewtonPicard algorithm with subspace iteration for computing periodic solutions.
SIAM J. Sci. Comput., 19(4):11881209, 1998.
56. T. Luzyanina and K. Engelborghs. Computing Floquet multipliers for functional
dierential equations. Internat. J. Bif. Chaos Appl. Sci. Engrg., 12(12):2977
2989, 2002.
57. T. Luzyanina, K. Engelborghs, and D. Roose. Numerical bifurcation analysis of
dierential equations with state-dependent delay. Internat. J. Bif. Chaos Appl.
Sci. Engrg., 11(3):737753, 2001.
58. T. Luzyanina and D. Roose. Numerical stability analysis and computation of
Hopf bifurcation points for delay dierential equations. J. Comput. Appl. Math.,
72:379392, 1996.
59. M. C. Mackey and L. Glass. Oscillation and chaos in physiological control
systems. Science, 77:287289, 1977.
60. J. Mallet-Paret and R. D. Nussbaum. A dierential delay equation arising in
optics an physiology. SIAM J. Math. Analysis, 20(2):249292, 1989.
61. J. Mallet-Paret and R.D. Nussbaum. Boundary layer phenomena for dierentialdelay equations with state-dependent time lags 1. Arch. Rational Mech. Analysis, 120:99146, 1992.
62. S. I. Niculescu. Delay Eects on Stability: A Robust Control Approach. Lecture
Notes in Control and Information Science, 269, (Springer-Verlag, New York,
2000).
63. M. Nizette and T. Erneux. Optical frequency dynamics and relaxation oscilations of a semiconductor laser subject to ltered optical feedback. In M. Pessa
D. Lenstra and Ian H White, editors, Semiconductor Lasers and Laser Dynamics
II, Proceedings of the SPIE, pages 618432. 2006.
64. G. Orosz, B. Krauskopf, and R. E. Wilson. Bifurcations and multiple trac jams
in a car following model with reaction delay time. Physica D, 211:277293, 2005.
65. G. Orosz, R. E. Wilson, and B. Krauskopf. Global bifurcation investigation
of an optimal velocity trac model with driver reaction time. Phys. Rev. E,
70(026207):110, 2004.
66. C. A. H. Paul. A user-guide to Archi - an explicit Runge-Kutta code for solving
delay and neutral dierential equations. Technical Report 283 (The University
of Manchester, Manchester Center for Computational Mathematics, 1997).
67. A. Pazy. Semigroups of Linear Operators and Applications to Partial Dierential
Equations. (Springer-Verlag, New York, 1983).

12 Bifurcation Analysis of Delay Dierential Equations

399

68. Y. Saad. Iterative Methods for Sparse Systems. (SIAM, Philadelphia, 2003).
69. G. Samaey, K. Engelborghs, and D. Roose. Numerical computation of connecting orbits in delay dierential equations. Num. Algor., 30(3-4):335352, 2002.
70. G. Samaey and B. Sandstede. Determining stability of pulses for partial dierential equations with time delays. Dyn. Sys., 20(2):201222, 2005.
71. L. F. Shampine and S. Thompson. Solving DDEs in MATLAB. Appl. Numer.
Math., 37:441458, 2001.
72. F. Schilder, H. M. Osinga, and W. Vogt. Continuation of quasiperiodic invariant
tori. SIAM J. Appl. Dyn. Sys., 4(3):244261, 2005.
73. J. Sieber. Dynamics of delayed relay systems. Nonlinearity, 19(11):24892527,
2006.
74. J. Sieber and B. Krauskopf. Control-based bifurcation analysis for experiments.
Nonlin. Dyn. in press.
75. G. Step
an. Retarded Dynamical Systems. (Longman, London, 1989).
76. G. Step
an and T. Kalm
ar-Nagy. Nonlinear regenerative machine tool vibrations.
In Proceedings of the 1997 ASME Design Engineering Technical Conferences,
Sacramento, California, September 1997. Paper number DETC97/VIB-4021.
77. G. Step
an, R. Szalai, and T. Insperger. Nonlinear dynamics of high-speed milling
subjected to regenerative eect. In G. Radons and R. Neugebauer, editors, Nonlinear Dynamics of Production Systems, pages 111128. (Wiley-VCH, Berlin,
2004).
78. R. Szalai. PDDE-Cont: A continuation and bifurcation software for delaydierential equations, 2005. Department of Applied Mechanics, Budapest
University of Technolgy and Economics; available at http://www.mm.bme.hu/
szalai/pdde/.
79. R. Szalai, B. P. Mann, and G. Step
an. Period-two and quasiperiodic vibrations
of high-speed milling. In I. Grabec and E. Govekar, editors, Proceedings of the
9th CIRP International Workshop on Modeling of Machining Operations, pages
107114. CIRP, 2006.
80. R. Szalai, G. Step
an, and S. J. Hogan. Continuation of bifurcations in periodic
delay-dierential equations using characteristic matrices. SIAM J. Sci. Comp.,
28(4):13011317, 2006.
81. L. N. Trefethen. Spectral Methods in MATLAB. (SIAM, Philadelphia, 2000).
82. K. Verheyden, K. Green, and D. Roose. Numerical stability analysis of a largescale delay system modeling a lateral semiconductor laser subject to optical
feedback. Phys. Rev. E, 69(3):036702, 2004.
83. K. Verheyden and K. Lust. A Newton-Picard collocation method for periodic
solutions of delay dierential equations. BIT Num. Math., 45(3):1605625, 2005.
84. K. Verheyden, T. Luzyanina, and D. Roose. Ecient and reliable stability analysis of solutions of delay dierential equations. Proceedings of 2006 International
Conference on Nonlinear Science and Complexity, pages 109120. (World Scientic Publishing, Singapore, 2006).
85. K. Verheyden and D. Roose. Ecient numerical stability analysis of delay
equations: a spectral method. In D. Roose and W. Michiels, editors, Time-delay
systems 2004, IFAC Proceedings Volumes, pages 209214. (Elsevier, Oxford,
UK, 2005).