The IMA Volumes

in Mathematics
and its Applications
Volume 37

Series Editors
Avner Friedman Willard Miller, Jr.
 Institute for Mathematics and
its Applications
IMA
The Institute for Mathematics and its Applications was established by
a grant from the National Science Foundation to the University of Minnesota in
1982. The IMA seeks to encourage the development and study of fresh mathemat-
ical concepts and questions of concern to the other sciences by bringing together
mathematicians and scientists from diverse fields in an atmosphere that will stim-
ulate discussion and collaboration.
The IMA Volumes are intended to involve the broader scientific community in
this process.
Avner Friedman, Director
Willard Miller, Jr., Associate Director
**********
IMA PROGRAMS
1982-1983 Statistical and Continuum Approaches to Phase Transition
1983-1984 Mathematical Models for the Economics of
Decentralized Resource Allocation
1984-1985 Continuum Physics and Partial Differential Equations
1985-1986 Stochastic Differential Equations and Their Applications
1986-1987 Scientific Computation
1987-1988 Applied Combinatorics
1988-1989 Nonlinear Waves
1989-1990 Dynamical Systems and Their Applications
1990-1991 Phase Transitions and Free Boundaries
**********
SPRINGER LECTURE NOTES FROM THE IMA:
The Mathematics and Pl1ysics of Disordered Media
Editors: Barry Hughes and Barry Ninham
(Lecture Notes in Math., Volume 1035, 1983)

Orienting Polymers
Editor: J .L. Ericksen
(Lecture Notes in Math., Volume 1063, 1984)

New Perspectives in Thermodynamics

Edi tor: James Serrin
(Springer-Verlag, 1986)

Models of Economic Dynamics

Editor: Hugo Sonnenschein
(Lecture Notes in Econ., Volume 264, 1986)
Rutherford Aris Donald G. Aronson
Harry L. Swinney
Editors

Patterns and Dynamics in

Reactive Media

With 64 Illustrations

Springer-Verlag
New York Berlin Heidelberg London Paris
Tokyo Hong Kong Barcelona Budapest
Rutherford Aris Donald O. Aronson
Department of Chemical Engineering School of Mathematics
and Material Science University of Minnesota
University of Minnesota Minneapolis, MN 55455
Minneapolis, MN 55455 USA
USA
Harry L. Swinney Series Editors
Physics Department and the Avner Friedman
Center for Nonlinear Dynamics Willard Miller, Jr.
University of Texas Institute for Mathematics
Austin, TX 78712 and its Applications
USA University of Minnesota
Minneapolis, MN 55455
USA

Mathematics Subject Classifications: 80A30, 80A32

Library of Congress Cataloging-in-Publication Data
Patterns and dynamics in reactive media / Rutherford Aris, Donald O.
Aronson, Harry L. Swinney, editors.
p. cm. - (the IMA volumes in mathematics and its
applications; v. 37)
Contains lectures given at a workshop held at the Institute for
Mathematics and its Applications, Minneapolis, Minn., Oct. 16-20,
1989.
Includes bibliographical references.
ISBN-13: 978-1-4612-7832-0 e-lSBN-13: 978-1-4612-3206-3
DOl: 10.1007/978-1-4612-3206-3
(Berlin: alk. paper)
I. Fluid dynamics-Congresses. 2. Combustion-Mathematical
models-Congresses. I. Aris, Rutherford. II. Aronson, Donald O.
III. Swinney, H. L., 1939- IV. Series.
QA911.P38 1991
541.3 '61- dc20 91-28048
Printed on acid-free paper.
© 1991 Springer-Verlag New York Inc.
Softcover reprint of the hardcover 1st edition 1991
All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New
York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis.
Use in connection with any form of information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed is
forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if the
former are not especially identified, is not to be taken as a sign that such names, as understood by
the Trade Marks and Merchandise Marks Act, may accordingly be used freely by anyone.
Permission to photocopy for internal or personal use, or the internal or personal use of specific
clients, is granted by Springer-Verlag New York, Inc., for libraries registered with the Copyright
Clearance Center (Ccq, provided that the base fee of $0.00 per copy, plus $0.20 per page is paid
directly to CCC, 21 Congress St., Salem, MA 01970, USA. Special requests should be addressed
directly to Springer-Verlag New York, 175 Fifth Avenue, New York, NY 10010, USA.
Camera-ready copy prepared by the IMA.
987654321
 The IMA Volumes
in Mathematics and its Applications

Current Volumes:
Volume 1: Homogenization and Effective Moduli of Materials and Media
Editors: Jerry Ericksen; David Kinderlehrer, Robert Kohn, J.-L. Lions

Volume 2: Oscillation Theory, Computation, and Methods of Compensated Compactness

Editors: Constantine Dafermos, Jerry Ericksen,
David Kinderlehrer, Marshall Slemrod

Volume 3: Metastability and Incompletely Posed Problems

Editors: Stuart Antman, Jerry Ericksen, David Kinderlehrer, Ingo Muller

Volume 4: Dynamical Problems in Continuum Physics

Editors: Jerry Bona, Constantine Dafermos, Jerry Ericksen,
David Kinderlehrer

Volume 5: Theory and Applications of Liquid Crystals

Editors: Jerry Ericksen and David Kinderlehrer

Volume 6: Amorphous Polymers and Non-Newtonian Fluids

Editors: Constantine Dafermos, Jerry Ericksen, David Kinderlehrer

Volume 7: Random Media

Editor: George Papanicolaou

Volume 8: Percolation Theory and Ergodic Theory of Infinite Particle Systems

Editor: Harry Kesten

Volume 9: Hydrodynamic Behavior and Interacting Particle Systems

Editor: George Papanicolaou

Volume 10: Stochastic Differential Systems, Stochastic Control Theory and Applications
Editors: Wendell Fleming and Pierre-Louis Lions

Volume 11: Numerical Simulation in Oil Recovery

Editor: Mary Fanett Wheeler

Volume 12: Computational Fluid Dynamics and Reacting Gas Flows

Editors: Bjorn Engquist, M. Luskin, Andrew Majda
Volume 13: Numerical Algorithms for Parallel Computer Architectures
Editor: Martin H. Schultz

Volume 14: Mathematical Aspects of Scientific Software

Editor: J.R. Rice

Volume 15: Mathematical Frontiers in Computational Chemical Physics

Editor: D. Truhlar

Volume 16: Mathematics in Industrial Problems

by Avner Friedman

Volume 17: Applications of Combinatorics and Graph Theory to the Biological

and Social Sciences
Editor: Fred Roberts

Volume 18: q-Series and Partitions

Editor: Dennis Stanton

Volume 19: Invariant Theory and Tableaux

Editor: Dennis Stanton

Volume 20: Coding Theory and Design Theory Part I: Coding Theory
Editor: Dijen Ray-Chaudhuri

Volume 21: Coding Theory and Design Theory Part II: Design Theory
Editor: Dijen Ray-Chaudhuri

Volume 22: Signal Processing: Part I - Signal Processing Theory

Editors: L. Auslander, F.A. Griinbaum, J.W. Helton, T. Kailath,
P. Khargonekar and S. Mitter

Volume 23: Signal Processing: Part II - Control Theory and Applications

of Signal Processing
Editors: L. Auslander, F.A. Griinbaum, J.W. Helton, T. Kailath,
P. Khargonekar and S. Mitter

Volume 24: Mathematics in Industrial Problems, Part 2

by Avner Friedman

Volume 25: Solitons in Physics, Mathematics, and Nonlinear Optics

Editors: Peter J. Olver and David H. Sattinger
Volume 26: Two Phase Flows and Waves
Editors: Daniel D. Joseph and David G. Schaeffer

Volume 27: Nonlinear Evolution Equations that Change Type

Editors: Barbara Lee Keyfitz and Michael Shearer

Volume 28: Computer Aided Proofs in Analysis

Editors: Kenneth R. Meyer and Dieter S. Schmidt

Volume 29: Multidimensional Hyperbolic Problems and Computations

Editors: James Glimm and Andrew Majda

Volume 30: Microlocal Analysis and Nonlinear Waves

Editors: Michael Beals, R. Melrose and J. Rauch

Volume 31: Mathematics in Industrial Problems, Part 3

by Avner Friedman

Volume 32: Radar and Sonar, Part 1

by Richard Blahut, Willard Miller, Jr. and Calvin Wilcox

Volume 33: Directions in Robust Statistics and Diagnostics: Part 1

Editors: Werner A. Stahel and Sanford Weisberg

Volume 34: Directions in Robust Statistics and Diagnostics: Part 2

Editors: Werner A. Stahel and Sanford Weisberg

Volume 35: Dynamical Issues in Combustion Theory

Editors: P. Fife, A. Liiian and F.A. Williams

Volume 36: Computing and Graphics in Statistics

Editors: Andreas Buja and Paul Tukey

Volume 37: Patterns and Dynamics in Reactive Media

Editors: Rutherford Aris, Donald G. Aronson, and Harry L. Swinney,

Forthcoming Volumes:

1989-1990: Dynamical Systems and Their Applications

Twist Mappings and Their Applications
Dynamical Theories of Turbulence in Fluid Flows
 Nonlinear Phenomena in Atmospheric and Oceanic Sciences
Chaotic Processes in the Geological Sciences

Summer Program 1990: Radar/Sonar

Radar and Sonar, Part 2

Summer Program 1990: Time Series in Time Series Analysis

Time Series (2 volumes)

1990-1991: Phase Transitions and Free Boundaries

On the Evolution of Phase Boundaries
Shock Induced Transitions and Phase Structures
Microstructure and Phase Transitions
Statistical Thermodyanimcs and Differential Geometry
of Microstructured Material
Free Boundaries in Viscous Flows
 FOREWORD

This IMA Volume in Mathematics and its Applications

PATTERNS AND DYNAMICS IN REACTIVE MEDIA

is based on the proceedings of a workshop which was an integral part of the 1989-
90 IMA program on "Dynamical Systems and their Applications". The aim of
this workshop was to cross fertilize research groups working in topics of current
interest in combustion dynamics and mathematical methods applicable thereto.
We thank Shui-Nee Chow, Martin Golubitsky, Richard McGehee, George R. Sell,
Rutherford Aris, Donald Aronson, Paul Fife and Harry Swinney for organizing the
meetings. We especially thank Harry Swinney, Rutherford Aris and Donald Aronson
for editing the proceedings.
We also take this opportunity to thank those agencies whose financial support made
the workshop possible: the Army Research Office, the National Science Foundation
and the Office of Naval Research.

Avner Friedman

Willard Miller, Jr.

 PREFACE

This volume contains some of the lectures given at the workshop "Patterns and
Dynamics in Reactive Media" held from October 16-20, 1989 as part of the year on
Dynamical Systems and their Applications at the Institute for Mathematics and its
Applications, Minneapolis, Minnesota.
Ever since the seminal works on traveling waves and on morphogenesis by Fisher,
by Kolmogorov, Petrovski & Piscunov, and by Turing, scientists from many disci-
plines have been fascinated by questions concerning the formation of steady or dy-
namic patterns in reactive media. The contributors to this volume include chemists,
chemical engineers, mathematicians (both pure and applied), and physicists. Their
contributions range from reports of experimental studies, through descriptions of
numerical experiments, to rather abstract theoretical investigations, each exhibit-
ing different aspects of a very diverse field. Although this small volume can hardly
claim to cover the whole range of current research in patterns in reactive media, it
nevertheless presents a representative sample.
It is a pleasure to thank all of the lecturers and participants for a very stim-
ulating workshop. We are also especially grateful to the staff and directors of the
IMA for courteously and efficiently providng a very rich medium in which we could
all react (and form new patterns) . Finally, Patricia Brick, Stephan Skogerboe, and
Kaye Smith deserve both thanks and a medal for their patient work in producing
this volume.

Rutherford Aris

Donald G. Aronson

Harry L. Swinney
 CONTENTS

Foreword ....................................................... IX

Preface......................................................... XI

Simple resonance regions of torus diffeomorphisms 1

Claude Baesens, John Guckenheimer,
Seunghwan Kim and Robert Mackay
A minimal model for spatio-temporal patterns
in thin film flow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
H. S. Brown, 1. G. K evrekidis and M. S. J oily
Localized and extended patterns in reactive media. . . . . . . . . . . . . . . 33
Christian Elphick and Ehud Meron
Some recent results in chemical reaction network theory. . . . . . . . . 43
Martin Feinberg
Genericity, bifurcation and symmetry. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Martin Golubitsky
Dynamics of some electrochemical reactions. . . . . . . . . . . . . . . . . . . . . 89
J.L. Hudson
Construction of the Fitzhugh-Ngumo pulse
using differential forms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
C. Jones, N. Kopell and R. Langer
Kinetic polynomial: A new concept of chemical kinetics. . . . . . . . . . 117
Mark Z. Lazman and Gregory S. Yablonskii
Convergence of travelling waves for phase field equations
to sharp interface models in the singular limit. . . . . . . . .. . . .. . . . . . 151
Yasuma3a Nishiura and Gunduz Caginalp
Standing and propagating temperature waves
on electrically heated catalytic surfaces. . . . . . . . . . . . . . . . . . . . . . . . . . 159
Georgios Philippou and Dan Luss
Mixed-mode oscillations in the nonisothermal autocatalator. . . . . . 171
S.K. Scott and A.S. Tomlin
Bifurcations and global stability in surface
catalyzed reactions using the Monte Carlo method. . . . . . . . . . . . . . . 187
D. G. Vlachos, L.D. Schmidt and R. Aris
SIMPLE RESONANCE REGIONS OF TORUS DIFFEOMORPHISMS
CLAUDE BAESENSt, JOHN GUCKENHEIMERt,
SEUNGHWAN KIM§ AND ROBERT MACKAyt

Abstract. This paper discusses resonance regions for two parameter families of diffeomor-
phisms and vector fields on the two dimensional torus. Resonance regions with at most two
resonant periodic orbits (in the discrete case) or two equilibrium points (for flows) are studied.
We establish global geometric properties of these regions with topological arguments.

Key words. dynamical systems, bifurcation theory

AMS(MOS) subject classifications.

This paper discusses the geometry of the simplest resonance regions displayed
by three weakly coupled oscillators. Much more extensive descriptions of the dy-
namics of three coupled oscillators are given in Baesens, Guckenheimer, Kim and
Mackay [1]. Indeed, the results presented here are a footnote to [1] and the reader
should look to this work for diagrams that illustrate the phenomena discussed here.
Generic two parameter families of three weakly coupled oscillators give vector fields
in which there is a smoothly varying family of invariant three dimensional tori. By
taking cross-sections of the flows on the invariant three dimensional tori, the three
interacting frequencies can be studied via a family of diffcomorphisms of the two
torus. This is the setting in which we shall work.
The theory of diffeomorphisms of the circle is thoroughly developed. It yields
information about resonances and mode lockings of two interacting frequencies.
There ;;,re stringent limits on the dynamical possibilities. The dynamical behavior
of a circle diffeomorphism can be summarized via its rotation number p. If p is
irrational, then the motion is quasiperiodic. Otherwise the rotation number is
rational, and all trajectories tend to periodic orbits of the same period in both
forwards and backwards time.
The theory of diffeomorphisms of the two torus is much more complicated than
the theory of diffeomorphisms of the circle. Phase portraits on a two dimensional
torus can have multiple at tractors as well as chaotic trajectories. Forward and
backward limit sets of different types of different types can coexist. The three types
of limit sets that are seen most frequently in generic diffeomorphisms are periodic
orbits, invariant curves and the whole two torus. Examples with chaotic trajectories
occur, but they are seldom found without a systematic search. A natural setting
for the study of diffeomorphisms of the two torus is the investigation of families in
which the amount of rotation in the two directions around the torus can be varied.
The family of diffeomorphisms In (x) = x + n is called the family of translations.

tMathematics Institute, University of Warwick, Coventry, England CV47 AL

:j:Department of Mathematics, Cornell University, Ithaca, NY 14853. The work of the first
author was supported by the National Science Foundation and the Air Force Office of Scientific
Research.
§Center for Applied Mathematics, Cornell l'niversity, Ithaca, NY 14853
 2

Here the two dimensional parameter n varies the amount of rotation. Most of the
theory that has been developed thus far has been within the context of families of
the form in (x) = x + n + eg( x) with g( x) a nonlinear, doubly periodic function and
e a third parameter which controls the size of nonlinf'arity. For fixed e, the system
goes through regions with many different rotation vectors as n is varied.
One can begin to understand the dynamics of a generic family of diffeomor-
phisms of the two torus by first considering the family of rotations and then pertur-
bations of this family. In a generic two-parameter family that is close to the family
of rotations, KAM theory guarantees that the sets of parameters values for which
the corresponding diffeomorphisms are not smoothly equivalent to irrational rota-
tions occupies a small measure in the parameter spare. The diffeomorphisms with
limit sets that are periodic orbits or invariant curves are the object of our interest.
We recall some properties of the parameter sets giving rise to this mode-locked be-
havior and of the dynamics that are observed on the torus when the mode-locking
occurs.
In the product of the parameter and the phase spaces, there is a two torus of
periodic orbits corresponding to each rational parameter vector (pI/ qI, pz/ qz). The
period q of these orbits is the least common multiple of qI and qz. It is easy to check
that this torus is a nonsingular set of roots of the pair of equations i~ (x) = x on
the product space. It follows that a family of torus diffeomorphisms that are near
rotations will have a torus P of periodic orbits of period q. Unlike the family of
translations, one cannot expect that P will project to a single point in the parameter
space. The image of the projection is called the resonance region of the rotation
vector with (pI/qI,PZ/qZ) associated with P. Basic properties of resonance regions
are described in [3]. As the image of the map of a torus into the plane, there are
topological constraints on the form of the resonance region and its boundary [4]. In
particular, generic families have resonance regions whose only singularities are cusp
points. Moreover, a resonance region cannot have a connected boundary without
cusp points. The presence of cusp points implies there are parameter regions III
which there are four periodic orbits with rotation vector (pI/ qI, pz/ q2)'
There are resonance regions of families of diffeomorphisms for which the indi-
vidual systems contain only one or two periodic orbits of the resonance rotation
vector. These resonance regions appear to arise naturally in two ways. First, if
one starts with a "Mathieu" family in which the nonlinear part eg( x) of the family
in (x) = x + n + eg( x) is given by a pair of trigonometric monomials, then the
expansion of in( x) - x near a resonance region of order n as a power series in e will
normally have dominant and sub dominant terms that are trigonometric monomi-
als. Galkin [2) describes the combinatorics associated with determining which are
the dominant terms and their degrees. The new results of this paper address the
geometry of these "simplest" resonance regions. The global topology of the torus
plays a prominent role in the discussion.
The behavior associated with invariant curves is more complicated. For each ro-
tation vector n = (WI, W2) satisfying only one independent equation PI WI +P2W2 = q
with integer coefficients, there is also a two dimensional torus in the product of the
 3

phase space and the parameter space for the family of translations consisting of or-
bits that have this rotation vector. The trajectories on this torus are quasiperiodic,
but lie on closed curves whose homotopy type is determined by n. For perturba-
tions, these tori mayor not persist: KAM theory provides the techniques proving
persistence for "good" rotation vectors. As n varies over vectors satisfying the re-
lation P1Wl + P2W2 = q, one obtains a line in the parameter space for the family
of translations along which there are invariant curves of the same homotopy type
for the corresponding diffeomorphisms. For perturbations of the family of transla-
tions, the first approximation to the set of parameters with invariant curves of the
specified homotopy type is a strip. As one moves along this strip, there are many
points at which the corresponding diffeomorphism acts like an irrational rotation
on a smooth invariant curve. At other points of the parameter space, there may
be invariant curves with less smoothness that are formed from saddle separatrices.
There may also be holes in the strips in which there is no closed invariant curve at
all. In addition to the rotational invariant curves that have non-trivial homotopy
type, one encounters invariant curves that can be continuously deformed to a point
on the torus.
The basic framework of a set of overlapping resonance regions and strips asso-
ciated with invariant curves of specific homotopy types can be substantially refined
by consideration of bifurcations showing the behavior at the boundaries of the strips
and resonance regions and the transitions that take place inside them. The geome-
try associated with these bifurcations is complex, but there are several basic facts
and principles that can be used to understand this geometry.
First, one has the simple fact that two closed curves of different non-trivial
homotopy types on the two dimensional torus must intersect. This implies that
if rotational invariant curves of different homotopy types exist, they are singular
curves that contain periodic orbits. Consequently, strips associated with different
homotopy types intersect only in resonance regions.
Second, there is an empirical observation that phase portraits of torus diffeo-
morphisms near translations usually look similar to those that come from the time
one map of flows. Thi~ observation is supported by the fact that every torus dif-
feomorphism has finite Taylor expansions that occur as the Taylor expansions of
time 1 maps of flows. Since flows on two dimensional manifolds do not have chaotic
trajectories, examining the dynamics of families of time one maps of flows gives
us a gentler approach to the study of resonance phenomena for diffeomorphisms.
Nonetheless, many of the topological features associated with diffeomorphisms can-
not occur in time one maps of flows, so additional analysis is required once one has
a good intuition for the dynamics of families of flows.
Third, there is a list of co dimension one and two bifurcations that occur m
generic two parameter families of two dimensional flows. This list can be used
to interpret and guide numerical explorations that seek to delineate the dynamics
that occur in specific examples. The interplay between the theoretical analysis
of normal forms and interactive computations of phase portraits has played an
important role in our developing understanding of families of torus maps. The
 4

novel aspect in the analysis of normal forms for our studies has been the necessity
of looking carefully at more complicated patterns of saddle connections than have
been examined previously.
Recall some of the different types of codimension one and two bifurcations for
flows on two dimensional manifolds. In co dimension one, there are saddle-node
bifurcations at which a pair of equilibrium points coalesce and Hopf bifurcations
at which a family of limit cycles ends at an equilibrium. Pairs of limit cycles can
coallesce in bifurcations called either double limit cycles or saddle-nodes of periodic
orbits. The final type of codimension one bifurcation involves a homoclinic or
heteroclinic orbit in which there is a trajectory that terminates at a saddle point as
time tends to ±oo. In co dimension two, the list of bifurcations is longer. The local
bifurcations involving equilibrium points are cusps, degenerate Hopf bifurcations
and the Takens-Bogdanov bifurcations at which an equilibrium has a nilpotent linear
part. In addition, there are cusps of limit cycles and various types of bifurcations
involving homoclinic and heteroclinic trajectories. For understanding the geometry
of the simplest resonance regions occurring in two parameter families, the additional
bifurcations that occur involve homoclinic orbits to saddle points with eigenvalues
of equal magnitude, saddle-node loops in which there is a homo clinic trajectory for
a saddle-node, and double saddle-loops in which a saddle point has two distinct
homoclinic trajectories.
Consider a two parameter family of flows on the two dimensional torus with the
property that each flow contains at most two equilibrium points. We want to prove
as many properties about the bifurcation diagram of such a family as possible.

PROPOSITION. Let f : T2 ~ R2 be a smooth map with generic singularities

and the property that, for each y E R2, f-l(y) is zero, one or two points. Then
the image of f is an annulus with smooth boundaries. The singular set of the map
consists of two curves that lie in the same non-trivial homotopy class.

Proof. The generic singularities of maps between two dimensional manifolds are
folds, cusps and the transversal intersection of two fold curves. Since cusps have
an adjacent region which is triply covered by the map, cusps do not occur. Nor
can there be intersections of two fold curves since each fold has an adjacent region
for which there are at least two preimages of each image point. Thus, adjacent to
a transversal intersection of two folds is a region in which each point has at least
four preimages. We conclude that the singular image of the map is a smooth one
dimensional manifold consisting entirely of folds. Thus, the singular image divides
the plane into a finite number of regions, one of these regions is the regular image of
the map, and the regular image is double covered. Since the Poincare index of the
torus is zero, it follows that there are precisely two boundary components to the
image. These components divide the torus into homeomorphic regions. Hence the
singular set on the torus must consist of two curves that lie in the same non-trivial
homotopy class.

THEOREM. Let x = n + g(x) define a two parameter family of flows on tIle

torus with the property tIlat for each n there are at most two equilibrium points of
the corresponding flow.
 5

(1) The resonance region for the map is an annulus.

(2) The flows in the bounded component of the complement of the resonance
region (the "hole") do not have a cross-section. In the notation of [1 j, the
flow has type Dw for a fixed rational w.
(3) There are at least two curves of Hopf bifurcations and six curves of homo-
clinic bifurcations in the resonance region.
(4) The curves of homoclinic bifurcations meet in at least two necklace points
at which all of the saddle separatrices of the flow are homoclinic.

Proof The first statement is a consequence of the previous proposition since the
resonance region is the image of the map -g : T2 --+ R2. The existence of at least
two Takens-Bogdanov points on each boundary of the resonance region was proved
in an earlier paper [3]. This implies the existence of at least two curves of Hopf
bifurcations and two curves of saddle-node bifurcations. The points on the torus at
which the map -g : T2 --+ R2 has a fold give rise to the saddle-node bifurcations for
appropriate parameter values (that depend on the point). Let C be the curve on
the torus whose image forms the inner component of the boundary of the image of
-g. Then C is a smooth closed curve that is homotopically non-trivial on the torus.
Observe that if n is a parameter value in the hole, then the image of C by the map
-g has winding number ±1 with respect to n. Thus, the map 9 + n : T2 --+ R2 has
image that is an annulus, the origin is in the bounded region of its complement and
the image of C has winding number ±1 with respect to the origin. We assert that
this is incompatible with the presence of a cross-section to the flow.
If a flow has a cross-section, then the winding number of the vector field with
respect to the origin is zero along any closed curve on the torus. This is readily seen
by constructing a homotopy of the vector field to a constant vector field through
vector fields without fixed points. Winding vectors do not change under such ho-
motopies. The construction of the homotopy can be accomplished by an initial
homotopy that makes the vector field have a constant return time. It is then trans-
verse to a one parameter family of cross-sections. A second homotopy of coordinate
changes makes these cross-sections parallel to one another. Finally, a homotopy of
the vector field makes it of unit length and orthogonal to the parallel cross-sections.
It is then constant. We concludp that g(x) + n has no cross-section.
Consider the vector field associated to a parameter value inside the hole of the
resonance region. There are at least two periodic orbits with opposite orientations,
but we can say more as well. The torus is divided into annuli by the periodic orbits.
Each of these annuli has either trajectories whose orientation does not change or
trajectories whose orientation changes by ±7l'. The second type of annulus is called
a Reeb component. The sum of the changes in orientation from all the annuli
is ±27l', corresponding to the index change of the vector field as we traverse the
curve along which folds occur in the vector field. If we collapse the annuli in which
orientation does not change onto closed curves, we obtain a topological flow on
a torus with periodic orbits separated by Reeb components. At least one of the
 6

periodic orbits must be attracting (and hence one repelling) since otherwise the
sum of changes in orientation from the Reeb components would be o.
The flow of vector fields for parameters in the hole have annuli that are at-
tracting their boundaries contained in the Reeb components. The presence of an
attracting annulus with boundary in Reeb components is a persistent phenomenon
under perturbation. Though the size of an attracting annulus may change discon-
tinuously as periodic orbits undergo bifurcation, perturbations of the flow will have
an attracting annulus that intersects the original one. As our parameter varies over
the disk, we can therefore find continuous families of closed curves that lie in the
the Reeb components that contain the boundaries of an attracting annulus.
We next consider the geometry of the vector fields obtained by traversing the
parameter curve , bounding the hole. For each parameter value, there is still a
periodic orbit of the homotopy type of those found in the hole since a saddle-node
bifurcation will not affect periodic orbits of opposite orientations. Observe that
the curve u traversed on the torus by the saddle-node points lies in a different
(non-trivial) homotopy type from the periodic orbits since the winding number of
the vector field along this curve is non-zero for parameter values inside the hole.
We also assert that the saddle-node points cross Reeb components containing the
boundary of an attracting annulus as one traverses ,. If this were not the case,
then a continuous family of closed curves in the Reeb components would undergo a
non-trivial homotopy along u. This is a contradiction since these families of closed
curves in the Reeb components can be extended continuously over the hole, implying
that the homotopy is trivial.
Since the saddle-nodes cross Reeb components as parameters traverse " there
must be parameter values in , for which the saddle-node points cross the boundaries
of Reeb components. In a generic family of vector fields, these must be codimen-
sion two bifurcations which adjoin regions with different numbers of periodic orbits.
Examining the list of co dimension two bifurcations for families of vector fields with
at most one saddle point, the only possibility for these bifurcations are that they
are saddle-node loops. If the saddle-node has a negative eigenvalue, this means
that the unstable separatrix of the saddle-node lies on the boundary of the stable
manifold of the saddle-node. Pursuing this logic further, one finds that there must
be at least four such bifurcation points corresponding to the boundary components
of two distinct Reeb components. i,From each of these co dimension two bifurcations
emerges a curve of homo clinic bifurcations in the parameter space. These bifurca-
tion curves are distinct from those that end at the Takens-Bogdanov points since
the homo clinic orbits are homotopically non-trivial. We conclude that there are at
least six curves of homoclinic bifurcations in our family of vector fields. This ends
the proof of the theorem.

REMARKS. The homoclinic bifurcation curves can end only at saddle-node loops
in generic families with at most one saddle point. The arguments given above can
be extended to show that the homo clinic bifurcation curves that we have identi-
fied must cross the resonance region from one boundary component to the other.
If one tracks periodic orbits along a homotopica.lly non-trivial closed curve in the
 7

resonance region, they follow paths that are homotopic to 0'. Thus the saddle-
ponts cross "Reeb components" as one traverses the parameter curve. The changes
of behavior that occur in the position of saddle points correspond to homoclinic
bifurcations. Since the stability of the two boundary components of a Reeb com-
ponent differ, there must be codimension two bifurcations at which the rotational
homoclinic orbits come from a trace zero saddle point. Generically, there are curves
of saddle-node bifurcations for periodic orbits that emanate from such codimen-
sion two bifurcations. Further analysis of the order with which different homoclinic
bifurcations will be encountered on the two boundary components of a Reeb compo-
nent indicates that the homoclinic bifurcations cross. Thus there are co dimension
two bifurcation points with double homoclinic cycles. These bifurcations are called
necklace points in [1].
When one considers the simplest resonance regions for diffeomorphisms rather
than flows, the structures described above become more complicated. Curves of
homoclinic bifurcations become thickened and the structure of saddle-nodes of pe-
riodic orbits becomes more complicated, with the addition of "Chenciner bubbles"
where a pair of resonant invariant curves try to approach each other. The reader is
referred to [1] for a discussion of what we know about the details of these additional
complications that appear in the bifurcation diagrams of two parameter families of
diffeomorphisms of the two dimensional torus.
 8

FIGURE 1: A numerically computed simple resonance region for the

family of flows

:i; = nx + cos(2rrx) + 0.1 cos(2rry)

if = ny + cos(2rrx) + 0.lcos(2rry)

The entire resonance region with its boundary of saddle-node curves

is shown. Half of each small oval is a curve of Hopf bifurcations and
half is a curve along which there is a saddle point with determinant
one. Curves of parameters with homoclinic bifurcations are shown
on one side of the resonance region.
 9

.......

..........................---~

//
)
..........................., ..../ .
.....
-:........ .

FIGURE 2: An enlarged vIew of one side of the simple resonance

region shown in Figure 1.

REFERENCES

[1] C. BAESENS, J. GUCKENHEIMER, S. KIM AND R.S. MAcKAY, Three Coupled Oscillators:
Mode-locking, Global Bifurcations and Toroidal Chaos, preprint, 0, pp ..
[2] O.G. GALKIN, Resonance regions for Mathieu type dynamical systems on a torus, Physica,
39D (1989), pp. 287-298.
[3] S. KIM, R.S. MACKAY AND J. GUCKENHEIMER, Resonance regions for families of torus maps,
Nonlinearity, 2 (1989), pp. 391-404.
[4] H. LEVINE, Stable maps: an introduction with low dimensional examples, BoJ. Soc. Bras.
Mat., 7 (1976), pp. 145-84.
 A MINIMAL MODEL FOR SPATIO-TEMPORAL
PATTERNS IN THIN FILM FLOW

H.S. BROWN., LG. KEVREKIDIS' AND M.S. JOLLyt

Abstract. We consider the development of spatio-temporal oscillations in the Kuramoto-

Sivashinsky amplitude model of thin film flow. These develop from Ropf bifurcations off of steady
state solutions and are observed to undergo symmetry breaking and period doubling bifurcations.
Oscillatory branches in the parameter regime studied apparently terminate in Silnikov type homo-
clinic connections. A minimal, three mode nonlinear Galerkin discretization, capable of capturing
this bifurcation behavior is constructed. A simple shooting algorithm which exploits this sharp
reduction in dimensionality is used to accurately locate the homoclinic connections.

Key words: Kuramoto-Sivashinsky equation, Silnikov connections, symmetry breaking, iner-

tial manifolds.

1. Introduction. The Kuramoto-Sivashinsky equation (KSE) [27, 36]

(1 )

has been used as an amplitude equation to describe incipient instabilities in several

physical contexts. It can be derived from a long-wave perturbation analysis of the
two dimensional Navier-Stokes Equations describing the flow of a thin, viscous liquid
film with surface tension, flowing down a vertical plane. The equation has been the
subject of numerous theoretical and computational studies during the last fifteen
years (e.g. [21, 9, 26, 19] and references therein) probably because of its ability
to exhibit low-dimensional, spatially coherent but temporally complicated solution
patterns. Numerical simulations have uncovered a wealth of such patterns, ranging
from steady states, periodic oscillations and traveling waves to modulated traveling
waves, symmetry-driven persistent homo clinic loops, and complicated (apparently
chaotic) dynamics as the instability parameter Ct grows. Theoretical work based
largely on symmetry considerations has concentrated on steady state and traveling
wave solutions, as well as heteroclinic and homo clinic connections between them (e.g.
[1 ]).
In this work we concentrate on the computer assisted study of limit cycle (time
periodic) solution branches arising from low Ct Hopf bifurcations. Two types of peri-
odic solutions are observed, one possessing a spatio-temporal symmetry that relates
the first half of the period of the oscillation with the second half. Such a symmetry
is known to suppress period doubling of the oscillations [38]. A symmetry breaking
pitchfork bifurcation follows, with subsequent period doublings of the non-symmetric
limit cycles. All oscillatory branches apparently terminate on Silnikov type infinite
period homoclinic connections (heteroclinic connections for the symmetric limit cy-
cles), involving saddle type steady state solutions. While Silnikov saddle connections
can occur in a phase space of dimension three, the simple three mode Galerkin trun-
cation of the PDE (1) cannot capture these phenomena. However, using the theory

• Department of Chemical Engineering, Princeton University Princeton, NJ 08544

t Department of Mathematics, Indiana University Bloomington, IN 47405
 12

4.0r-------~S~l-.-ad~y-=S~la~l-.-.-------------------------------------.

Limi l Cycles
3.5
Trave ling Waves

3.0

2.5 I

;:j 2.0 , , I
hi
t'l
1.5

1.0
"
,'
", \
I
"
"
.5
~
?
5 10 15 20 25 30 35 40
CI.

Figure 1: Bifurcation diagram for the KSE; 0 marks steady state and traveling wave
bifurcations; 0 marks Hopf bifurcation points.

of approximate inertial manifolds, we construct a minimal (three dimensional) non-

linear Galerkin discretization which qualitatively (and to some extent quantitatively)
captures the Silnikov behavior in the parameter regime studied. This sharp reduction
in the dimensionality of the problem allows for the easy visualization of the stable
and unstable manifolds involved in these interactions. We have exploited this to im-
plement a simple shooting algorithm which converges on the Silnikov loop in (phase
x parameter) space.

2. Numerical Bifurcation Diagrams. Figure 1 is a bifurcation diagram for

the KSE for 0 < a < 40. The diagram shows steady, periodic, and traveling wave
solution branches computed using a traditional Galerkin spectral discretization of the
PDE with nine Fourier modes. The diagram is "converged" in the sense that doubling
the dimension of the discretization does not visibly alter the location or the stability
of the solution branches.

Our calculations were actually performed using the integrated form of the KSE
with periodic boundary conditions on 0 < x < 27r:

(2) Ut + 4u:z:=:z: + a 1
( -(U:z:)2
2
+ U:z::z: - -1
47r
1
0
2".
(u:z:? dx ) = O.
AUTO, a continuation/bifurcation package developed by E. Doedel [12, 13J was used
to perform these calculations. The steady state and limit cycle branches (the latter
are often called "standing waves") were calculated by restricting the dynamics to the
invariant subspace of even functions; all subsequent calculations in this paper will be
similarly restricted. Until recently most of the rigorous results (dissipativity, existence
and dimension estimates for inertial manifolds) for the KSE have been limited to this
invariant subspace (see for example [33]). Il'yashenko [22J has recently presented a
 13

proof of dissipativity for the general case of periodic boundary conditions. The travel-
ing wave branches, computed in full Fourier space with periodic boundary conditions,
are included here for completeness. Another form of solution (modulated traveling
waves) also known to exist in the full Fourier space within this parameter range have
not been included in the diagram, since the discussion will be mostly restricted to
even solutions and their bifurcations. Such bifurcation diagrams, particularly those
including the steady state solutions of the KSE have appeared in the literature (e.g.
[34, 19, 3]). In this paper we focus on the family of limit cycle branches observed
for 30 < a < 37. Certain aspects of these branches and their bifurcations have been
discussed previously in [24, 23, 7J.
The steady state branches bifurcating at a = 4(= 4 * 1 2),16(= 4 * 2 2),36(=
4 * 3 2), ... 4 * k 2 are characterized as "unimodal" (k = 1), "bimodal" (k = 2), ...
"k-modal" respectively, based on their spatial structure close to the respective bi-
furcations. A Hopf bifurcation occurs at a ~ 30.34 off of the bimodal steady state
branch (restricted on the invariant subspace of even functions). This limit cycle with
period T (i.e. u(x, t) = u(x, t + T)) is characterized by a spatio-temporal symmetry:
the second half period of the oscillation (in time) is related to the first half period by:

u(x, t) = u(x + 71", t + T/2).

This symmetry is a version of what Aronson et al. call "Ponies On a Merry-go-round"
for discrete arrays of Josephson junctions [2J. The implications of such a symmetry on
the bifurcation behavior of periodic solutions was discussed by Swift and Wiesenfeld
[38J.
To understand the properties of these symmetric limit cycles, it is helpful to
consider certain replication and invariance properties of the solutions of the KSE. It
is easy to see that the space

C le == span(cos(kx),cos(2kx),cos(3kx) ... )

is invariant under the flow of the KSE and that if u(x, t) is a solution of the KSE for
some parameter value ao, then so is u(kx,k4t) at a = k2ao. This means that steady
state branches of the KSE replicate as discussed in detail in [34J. For every nontrivial
steady state branch, a one-parameter family of steady states exists at every a value,
each one of them a shift (in x) of the others. All steady state branches shown in
figure 1 (unimodal, bimodal, trimodal, as well as a mixed mode branch, the "bi-tri"),
represent solutions which can be shifted to be even (i.e. can be represented by a
cosine series on [0,271"]). Applying the above replication property with k = -1, we
see that two "representatives" of each nontrivial k-modal steady state branch can be
found in CIe , related to each other by a translation by 71"/ k. The bi-tri steady state
branch can be thought of as "unimodal" in this context: two representatives of it
exist in C 1 related to each other by a translation by 71"/1.
We are interested here in the bifurcations of these even solutions, taking place
in the invariant subspace of even functions. The two representatives of odd-modal
(k = 1,3,5, ... ) branches can be shown to have the same stability in C 1; on the other
hand, the two representatives of even-modal (k = 2,4,6, ... ) branches have different
stability properties in C 1 (while of course they have the same stability in C k as well as
in the full Fourier space, where they also both possess a zero eigenvalue corresponding
to the direction of shift invariance). The Hopf bifurcation observed at a ~ 30.34 is
 14

u(x,t) (a) .. - u(x.w/B)

u(J: = '" + 'II/B)
(b)

-..
UJ
-LO
L
f--
-1.5

-2.0

-2.'

-3.0 .....""
o .1 .2 .3
x time

Figure 2: (a) A space-time plot of one period of a symmetric limit cycle (0 < x < 27r)
at ex '" 33.56. (b) Time series of the amplitude at x = 7r /8 and x = 7r + 7r /8 during
one period of the oscillation.

not a replication of a lower ex Hopf bifurcation on the unimodal branch. Since the
eigenvalues (eigenvectors) of the bimodal representatives in C 2 are identical (replicas)
to those of the unimodal representatives in C 1, this Hopf bifurcation does not occur
entirely within C 2 and hence involves components in C 1, in which the stability of the
two bimodal representatives is different. The Hopf bifurcation is therefore expected to
(and indeed observed to) occur for only one of the two representatives of the bimodal
branch.
We now restrict our analysis to the invariant subspace of even functions C 1. The
two bimodal representatives are each invariant under 7r shifts and transformed into
each other by 7r /2 shifts. Since the replication law with k = -1 (the shift by 7r) is a
property of both steady and time dependent solutions, the unique limit cycle resulting
from the single representative Hopf bifurcation must be (as a solution) invariant to a
spatial shift by 7r. In this case, the spatio-temporal symmetry:

u(x) ---7 S ..u(x); t ---7 t + T/2

results, where
S .. u(x) == u(x + 7r).

The symmetry of the limit cycles is illustrated in figure 2. To see how this sym-
metry will appear in Fourier space, consider the ODE system

arising from the Fourier-Galerkin discretization of the KSE:

N N
u(x,t) '" L:(u(x,t), cos(nx)) cos(nx) == L: an(t)cos(nx).
n=l

The notation (', .) refers to the standard inner product on L 2. Figure 3 shows a rep-
resentative phase portrait and time series characteristic of the symmetric oscillations.
This type of symmetry in limit cycles is known to cause suppression of period dou-
 15

- .. 2.5

2.0
r-------,---:-:-:-------,
(u,ell.(.»
(u.cOR(2x»
(b)

-1.0
~ 1.5

-1.2 Iii' 1.0

o
g -1.4 i
u
.~

"~
~
a
(J~ -1.8
E.
-1.8
x -.'
~ -1.0
U
-0.0 i -1.5

-2.2

-2.5 -2.0 -1.5 -1.0 -.5 0 .5 1.0 1.5 2.0 2.5 -2.!l o~-';.1~.';;-2"--';.,:-'-.':-,"--';.5c"--.';;-,"--';.,0-"-.';;-,"--';.'~l. 0
(u.cos{x)) time

Figure 3: (a) An a l-a2 phase space projection of a symmetric limit cycle at a ~ 33.56
and (b) time series of two of its Fourier coefficients (a 1 and a2)'

bling. The proof for the KSE follows directly the arguments of Swift and Wiesenfeld,
substituting u(x) for their x and S ... (the 7r shift) for their -I. It is also easy to
see that the argument carries over for the ODE system resulting from the Galerkin
spectral discretization of the PDE. The symmetry becomes

al -al
a2 a2
a3 -a3 T
-> , t->t+ 2'
a4 a4

In this case, instead of the Swift and Wiesenfeld matrix -I, the diagonal matrix J
with elements j ii = (-1)' is used. The basic argument is that for symmetric orbits,
the state transition matrix DP is the square of another matrix (D P)2 and therefore
has as eigenvalues the squares of the eigenvalues of the matrix D P. The matrices
being real precludes DP from having a single real eigenvalue crossing the unit circle
at -1 as a varies.
In the KSE (as well as in other systems with this property) a symmetry breaking
(pitchfork) bifurcation occurs off of the symmetric limit cycle branch at a ~ 32.85,
giving rise to two "asymmetric" limit cycle branches (symmetric to each other by the
7r shift in x). These are illustrated in figure 4. The asymmetric limit cycles are not
prevented from period doubling, and are indeed observed to subsequently undergo a
period doubling cascade [23]. The first period doubling occurs at a ~ 32.97 and the
second at a ~ 32.99. Figure 5 shows representative phase portraits of limit cycles on
these branches.
These symmetry properties playa significant role in "orchestrating" the compli-
cated structure of limit cycle branches and associated Silnikov connections shown
enlarged in figure 6. Partial observations of this intricate pattern have been pre-
sented and published previously in [25, 24, 23, 7, 8]; in what follows we attempt a
more coherent discussion of the detailed bifurcation picture. Figure 7a shows the fate
 16

-.4r---------------------------------------,
-.8

-.8

-1.0

........... -1.2
><
N -1.'
0'
g -1.8
::i
--1.8

-2.0

-2.2
.'
-2.4

-2.8~--~---L~~----L---~---L--~~--L---~--~
-2.6 -2.0 -1.5 -1.0 -.5 .5 1.0 1.5 2.0 2.5
(U,COS(X))

Figure 4: An al-a2 phase space projection of an asymmetric limit cycle and its 11'
shift at a '" 33.56.

..
-.' r-------------------------~ -.,r----------.-------~

(a) (b)
- -.'
-.. -.'

.
-1.0 -1.0

_-1.2 _ -1.2
X
~-1.4 e-1. 4

"3 -1.11 8" -1.8

,; ,;
- -LB ......... -1.8

..... ....0
-2.2 -2.2

-2.4 -2.4

-2·~2.":.,~--:.": .• .•
.•~-.,.I.~,~-u.I.':'".~_~,....0Lu..~.':'-.~I."".~,u.!"-"'."'.•~•.• -2 ·~2"'
.•~--:.": .•!"-"'."'.•~•.•
.•~-'"1.~5~-U.1.':'".~_~.•,....0Lu..~.':'-.~I.""O~,U.
(u,cos(x)) (u,cos(x))

Figure 5: al-a2 phase space projections of (a) the period 2 and (b) the period 4
asymmetric limit cycles at a '" 33.56.
 17

2.6

2.4

2.2

2.0

1.8

;j 1.6

1.4

1.2

1.0

.8

.6

33.5 34.0 34.5 35.0 35.5 36.0 36.5

0:

Figure 6: Enlarged bifurcation diagram for the KSE. 0 marks steady state bifurca-
tions, 0 marks Hopf bifurcation points, and <> marks period doubling bifurcations.

of both asymmetric limit cycle branches; after the first period doubling, the asymmet-
ric branches undergo a "corkscrew" sequence of turning point bifurcations while their
period (not shown in the figure) approaches infinity. In the figure, the branch is seen
to asymptotically approach the bi-tri steady state solution branch. This b~havior
is typical of a homo clinic connection to a saddle-focus (a Silnikov loop). A detailed
discussion of the bifurcation behavior associated with this phenomenon in three space
dimensions (a one dimensional unstable manifold connecting with a two dimensional
stable manifold corresponding to a complex eigenvalue pair) can be found in [18] (see
also [17]). A space-time plot and a Fourier space projection of a representative limit
cycle close to the "tip" of the corkscrew (figures 7b and c) illustrate the geometry
of this connection: the limit cycle is indeed seen to approach (and spend most of its
period in the neighborhood of) the bi-tri saddle type steady state. The nature of
the eigenvalues of this steady state (confirmed by independent computations) is also
obvious in the figure: one is positive and real, while the two least stable eigenval-
ues of the saddle form a complex conjugate pair (indicated by the spiraling close to
the steady state). The limit cycle branch is expected to corkscrew on its approach
to homoclinicity if the Silnikov condition, 161 < 1, is satisfied; 6 is the ratio of the
real part of the complex eigenvalue pair to the real eigenvalue of the saddle-focus
(see [18]). A number of other features, qualitatively and quantitatively predicted in
the Glendinning and Sparrow analysis have also been computationally observed, and
are partly seen in this figure (for a better illustration see figure 10). These include
the (asymptotically exponential) narrowing of the parameter interval between suc-
cessive turning points on the corkscrew, the existence of period doubling bifurcations
and their relative location with respect to the turning points as well as subsidiary
homoclinicities.
 18

u(x,t) (b)
(a) .

t .•

.,.
"
"

,
::J I.e "

...
(0)
"

.......
.
'.\.
,

,;
'-" -.&

'it;
-\.11 -1.11 -.1> II .11 1.11 1.11 il.1I 11.1>

a (u,cos(x»

Figure 7: (a) Partial bifurcation diagram of the asymmetric limit cycle branch. (b)
Space-time plot and (c), an a l-a3 phase space projection of a limit cycle approaching
homoclinicity (a ~ 34.37, marked with a *).

Figure Sa overlays the fate of the initial period-doubled and the subsequent
period-quadrupled limit cycle branches on figure 7a. We observe that they also
asymptotically approach corkscrews located close to that of the primary asymmetric
oscillations. Figures Sb and c show Fourier space projections of representative limit
cycles close to the tips of the corresponding corkscrews, indicating again an interac-
tion between the stable and unstable manifolds of the same branch of bi-tri saddle
type steady state solutions, Period doubling bifurcations like those on the primary
branch are also expected close to the limit points of these subsidiary corkscrews.
Figure 9 shows the fate of the original symmetric limit cycle branch. Computations
of this branch also display the beginnings of a corkscrew behavior. However, as
the Fourier space projection portraits of representative limit cycles on this branch
indicate (figures 9b and c), the limit cycles appear to asymptote towards a symmetric
heteroclinic connection. This connection apparently involves the unstable manifold of
one saddle type steady state with the stable manifold of its 7r shift and vice-versa. The
computation of this branch presents certain numerical difficulties since the computed
limit cycles are very unstable (typical Floquet multipliers even before the first turning
point become so large -0(10 10) and more- as to be inaccurate). These phase portraits
are strongly reminiscent of the structure of the Lorenz attractor, and the bifurcation
behavior of orbits of this nature has been discussed in Sparrow's book [37].
A much more "textbook" example of Silnikov type behavior is exhibited by the
limit cycle branches emanating from the Hopf bifurcation point at a ~ 34.30 on the
bi-tri steady state branch. The two representatives of this branch have the same
stability in the C 1 space and therefore two Hopf bifurcations occur simultaneously,
 19

'.6

... ( a) l.6.--------------,
(b)

...
1.0

.
'"o
.5

'.0
,;
- -.5

1.6

1.6
-1.~2.'-;.'C-_,,-:::-;;.O~-:f'I.;;:.-:c_,;'o.o~-,';c,-'-t;o--';:
.• --.';;-'--;"-;:-.,.,.....-;'
;g (U,COII(X»
I..
(0)
I..

1.0
":'f,

.'
..
1
.5

.6 ".'
"
" -
d -.5
"
.6

.to":.0"""="'••=""3"',""'0"'=,7
. 3."::'."=..~.70"=..~.•=-'=3.~.0"""'30~.•"',"'.~.0~3,-J.,. -1 '~2''-;',C-_,,:to,o~-:f'l.;:".::;_,;'0,,~_,",_'_t;,-,-';:,--'-;"';;_'__;"_;:_"-;"';....._;'
0. (U,COS(X»

Figure 8: (a) Partial bifurcation diagram of the asymmetric limit cycle branches.
al-a3 phase space projections of the period-2 (b) and period-4 (c) asymmetric limit
cycles close to homoclinicity (a ~ 34.36).

giving rise to two (symmetric to each other) limit cycle branches. Because there are no
symmetry restrictions, these branches can (and indeed are observed to) period double
according to the typical Silnikov scenario. An initial period doubling at a ~ 34.54
results in a limit cycle branch which, as seen in figure lOa, apparently asymptotically
approaches a Silnikov connection of a saddle focus on the "middle" (in the figure)
steady state branch. A second period doubling at a ~ 35.52 gives rise to a branch
with a similar fate. Interestingly the next period doubling on the "primary" corkscrew
(a ~ 34.84) gives rise to a period doubled branch which appears to asymptotically
approach a homoclinic connection of a saddle lying on a different part of the steady
state branch. This is a "subsidiary homoclinicity" and a representative limit cycle
portrait on this branch (figure lOc) clearly illustrates closeness to a "double pass"
Silnikov connection (as opposed to the "single pass" of the primary homoclinicity
(figure lOb).

Obviously the details of these diagrams are incomplete -one could go on computing
limit points, period doublings, and subsequent homoclinicities forever. It is sufficient
to summarize this entire scenario as a form of "loss of a limit cycle branch" [18].
Before concluding this descriptive section, we should add that even this complicated
scenario does not contain all essential features of the dynamics of the KSE in C 1 in
the parameter interval studied; Simulations indicate additional global bifurcations for
even earlier values of a than the ones discussed here (e.g. close to a = 32.9000355);
the picture is far from being complete.
 20

3.2

3.0
( a)
-""
2.8

2.6

2.4

2.2

2.0
~ 1.8
=
1.8

1.4

1.2

1.0
:
.8

.6
30 31 37
0:

1.5,--------------,
(b)
-.2
1.0
-.4

.5 -.6

!
"'e.
~ ~

-.8

0 ~ -1.0
0 o
i i-1.2
-.5
-1.4

-1.6
-1.0
-La

-1..!i!.5 -2.0 -1.5 -1.0 -.5 0 .5 1.0 1.5 2.0 2.5 -2·~2.5 -2.0 -1.5 -1.0 -.5 0 .5 1.0 1.5 2.0 2.5
(U.COS(X)) (u.cos(x))

Figure 9: (a) Partial bifurcation diagram of the symmetric limit cycle branch. (b)
al-a3 and (c) al-a2 phase space projections of a representative limit cycle (a'"
36.15).
 21

2.6 ... ... ( a)

2.4

2.2

2.0

;::j 1.8

1.8

1.4

1.2

1.0
34.2 34.4 34.6 34.8 35.0 35.2 35.4 35.6
C(

.80 .•0
(b) {cJ
.75

.B'
.B'
.60
><
e ><
e .56
'"oo g'" .50
i i .4 •

• 40 .40

.30 .3'
.30 .30

.. 1.0 1.5 2.0

(U,COS(X))
2.5 3.0 .2. L.,=",~-''''.O''''''''~I.':-,~-:'":.0""""'-=':--~r,..........J
(U.COS(X»

Figure 10: (a) Partial bifurcation diagram of the limit cycle branches bifurcating from
the bi-tri steady state branch. (b) and (c) show phase space projections of a period
1 (a'" 35.17, marked with a *) and a period 2 (a'" 35.11, marked with a *) limit
cycle approaching homoclinicity.
 22

3. An accurate three mode model. The bifurcation calculations in the

previous section were performed using a nine mode traditional Galerkin spatial dis-
cretization of the PDE. While the computational task involved in this level of analysis
for nine ODEs is tractable, it is conceivable that an alternative discretization tech-
nique could capture the dynamics accurately with a smaller number of modes. A
traditional approach to model reduction, particularly suitable for the qualitative lo-
cal analysis of bifurcating branches, involves studying the flow on a center manifold
in the neighborhood of a particular bifurcating solution. An example of this type
of analysis for the KSE in the recent literature can be found in [1]. An alternative
approach involves Galerkin projection on empirically determined basis functions ob-
tained from statistical processing of computationally or experimentally obtained time
series. This more practically oriented approach (method of empirical eigenfunctions
or Proper Orthogonal Decomposition, POD), introduced by Lumley [30, 31] as a
method of detecting spatially coherent structures in turbulence, has been extensively
used in recent years for the study of transitional flows in various geometries (e.g.
[4,35, 11]) and references therein).
A global (in phase space) model reduction approach is motivated by the theory of
inertial manifolds. This theory can be described for a dissipative PDE written as an
evolutionary equation in a Hilbert space H

du
dt + Au + F(u) = 0, u E H.

For example, for the KSE we let A = ::. and

-8x 2- -411"1 12.. (u"Y dx ) .

(-88xU+ -21 (8U)
2
F(u) = a
2 0

An inertial manifold is a finite dimensional, positively invariant, exponentially at-

tracting Lipschitz manifold embedded in the phase space H. Let P = P m denote
the orthogonal projector onto the span of the first m eigenfunctions of A, and let
Q = Qm = I - Pm be that onto the orthogonal complement. Typically an inertial
manifold is constructed as the graph of a function cJ> : PH -+ Q H. This function is
used to define a finite dimensional ODE (the inertial form)

dp
dt +Ap+PF(p+cJ>(p)) =0, pEPH

which captures all the long time behavior of the PDE.

While inertial manifolds are known to exist for the KSE as well as other dissipative
PDEs (see e.g. Constantin et aI. [10]), they cannot in general be expressed in closed
form. For this reason a number of approximate inertial manifolds (AIMs) have been
introduced in the literature (see [32] and the references therein). The AIM used here
is designed to pass near all steady states of the PDE, which lie on the global attractor,
and consequently on any inertial manifold. This construction is also well motivated
for dissipative PDEs that may not necessarily possess an inertial manifold. Indeed,
this AIM was originally introduced in [39] for the Navier-Stokes equations (NSE).
Consider the Q-component of the condition for a steady state

Aq+QF(p+q)=O, pEPH, qEQH,

 23

which implicitly defines an analytic function il> " : PH -+ QH given by the fixed points
for the family of mappings
T,,(q) = -A -lQF(p + q),
provided m is large enough. The AIM used in computations in this paper

il>2(P) == T,,(T,,(O))
has a distance to il>" (in L2) which is comparable to that between il> " and the global
attractor (see (39) for the NSE, (23) for the KSE).
The implementation of this nonlinear Galerkin approximation is well suited for
computation using spectral methods. The nonlinear Galerkin method used here i.e.
the m dimensional ODE system

(3)

(called an Approximate Inertial Form, AIF) amounts to three successive evaluations

of the nonlinearity F. For the KSE, with its quadratic nonlinearity and periodic
boundary conditions, the graph of il> 2 lives in the span of the first 4m eigenfunctions
of A (Fourier modes, cosines for the even functions here). This means that with
sufficient padding with zeros, spectral codes can be used to compute the vectorfield
of (3) faithfully. In the general case where the nonlinearity is not of polynomial type,
a collocation scheme can be implemented efficiently using the Fast Fourier Transform,
albeit with aliasing errors (7).
Obviously, the minimal number of modes (phase space dimension) necessary to
accommodate the Silnikov loop phenomena described above, is three. However, a
traditional three mode Galerkin discretization does not capture these phenomena (23).
The best a priori estimates for the dimension of an inertial manifold for the KSE do
not suggest that this dimension could be as low as three or, for that matter, even
nine (which we empirically know to be sufficient from our computational results). Our
choice here of the dimension of the approximate inertial manifold (three) is therefore
based on the dynamic phenomena observed rather than on rigorous estimates.
Figure 11 shows the relevant section of the bifurcation diagram for the AIF (3)
with three modes, which does indeed exhibit the same limit cycle bifurcation behav-
ior as the "converged" diagram (figure 6). The original Hopf bifurcation from the
bimodal branch to a symmetric limit cycle, the subsequent symmetry breaking fol-
lowed by period doublings, as well as the ultimate Silnikov loop terminations of the
resulting branches are present (we had some numerical difficulties with the extensive
continuation of the symmetric limit cycle branch). Similarly, the "nonsymmetric"
limit cycle branches bifurcating off of the bi-tri mixed mode steady state branch,
their subsequent period doublings, Silnikov loops, and subsidiary homoclinicities are
correctly (and, to some extent, accurately) reproduced. The actual parameter values
at which the Silnikov connections occur in the reduced system are not identical to
the traditional Galerkin values; however, they are in reasonable semi-quantitative
agreement.

4. Approximating the Silnikov connections. The existence of certain

Silnikov type homoclinic connections "implies" the existence of a complicated pattern
 24

;oj
..........:........

---- -- --- -----: i/J:

33.0 33.5 34.0 34.5 35.0 35.5 36.0 36.5
0:

Figure 11: Partial bifurcation diagram for the KSE obtained using the AIF (3) with
m=3.

of bifurcations and oscillatory branches in the neighborhood of these connections in

(phase x parameter) space. Several aspects of this scenario were illustrated above
for the KSE. On the other hand, this entire bifurcation sequence can be thought of
as a mechanism for the disappearance of the "primary" limit cycle branch as the
bifurcation parameter changes. For example, the collection of limit cycle branches
around a '" 35.17 in figure 10 can be taken to simply denote the "death" of the
nonsymmetric limit cycle branch originating at a '" 34.30.
The continuation of limit cycle branches as they approach homoclinicity involves
considerable computational effort without necessarily producing significant informa-
tion. Due to its infinite complexity, the diagram can never be completed; in ad-
dition, as the limit cycle branches approach homoclinicity their period approaches
infinity, and their accurate calculation poses severe discretization problems. On the
other hand, all these complications occur in a very narrow parameter interval, and it
appears that approximating the Silnikov connection directly would for all practical
purposes contain the essential information regarding the behavior of the limit cycle
branch( es).

Figure 7c is representative of the phase space geometry of a Silnikov connection.

The one dimensional unstable manifold of a saddle steady state connects with the
saddle itself (and therefore lies on the stable manifold). The spiraling approach to
the saddle is indicative of the complex conjugate nature of the two least stable saddle
eigenvalues ("saddle-focus"). This figure is of course only a projection, and the stable
manifold is not two dimensional, as was the case in Glendinning and Sparrow (for
this figure it is actually eight dimensional). The model reduction approach discussed
in section 3 reduces the dimensionality of the stable manifold to two, the minimum
necessary, and allows for the easy visualization of both invariant manifolds and their
interactions. We have used this reduction to implement a geometrically motivated
simple shooting algorithm to converge on the Silnikov loop. Such algorithms (special
cases of boundary value problems on semi-infinite or infinite intervals [29]) have
appeared in the literature for the calculation of saddle connections in low dimensional
 25

Approximation to Stable Manifold

Figure 12: illustration of the geometry and terminology used in the shooting algo-
rithm.

systems [20] (see [28] for the Silnikov case). More recently, results for the general
case with arbitrary (finite) dimensional interacting manifolds have been reported
[14, 15, 16, 6, 5].
We now present an outline of our implementation for the case when the associated
steady state has a one dimensional unstable manifold and a two dimensional stable
manifold. For simplicity we rewrite the approximate inertial form (3) as
(4) p= G(pj 0:).
Let p(o:) denote a fixed point of (4) with associated stable eigenvector v 1 ± iV2 and
unstable eigenvector v 3. Let Or define a cylinder of radius r with its axis in the
direction of V3 passing through the point p. Consider also the plane spanned by
the vectors v 1> V2 passing through p (see figure 12). Given a good guess 0: n for the
parameter value, 0:', at which the Silnikov loop occurs, the computational procedure
is outlined below:

1. Locate the steady state p(o: n) via Newton iteration and compute its eigenvec-
tors.
2. Integrate
p = G(PjO:n)
p(t=O) = P+ev3
until the trajectory intersects the cylinder 0 r as shown in figure 12. Let a
denote the intersection point.
3. Calculate the distance D = D( 0: n, e, r) from a in the direction of v 3 to the plane.
To satisfy D( 0:) = 0:
4. Calculate the derivative ~ and use it to update the current guess for 0: via a
Newton iteration:
 26

e Paracreter Correction
10 '
-----{3---- Distance fran Plane
10 0 m· ---Ill

10- 1

10-2

10-3

10--<
~ 10-6
0
~
~ 10-6
~
10-7
10-6

10-9

10-10

10- 11 r=.449 r=.273

£=.0001 £=.0001
10- 12

10- 13
2 4 6 8 10 12 14
Ilerations
Figure 13: Rate of convergence to the Silnikov connection at a ~ 35.2741195.

Upon convergence of this procedure we obtain an a( E, r) satisfying D( a, E, r) ~ O.

Since Va is tangent to the unstable manifold at p, it is obviously a good approximation
of the unstable manifold for small enough E. Similarly, the plane provides a good
approximation of the stable manifold for small r. Therefore, in the limit, a(r, E) should
approximate a·. In our implementation we have repeated the above procedure by
systematically shrinking rand E until a( E, r) remains constant to within a prescribed
error tolerance (usually < 10 -8).
Calculation of the intersection point a involves a Newton-Raphson iteration en-
suring that it lies on C r, which implicitly defines the "time of flight" of the trajectory,
and can be used to calculate its dependence on the system parameters. The most
involved segment of the computational procedure is the evaluation of the derivative
<lJ:.. D depends on a both explicitly through the vectorfield and implicitly through the
initial condition (the fixed point and its eigenvectors) and the cylinder itself (again
depending on the eigenvectors). The calculation therefore involves repeated applica-
tion of the chain rule, using information from integration of the sensitivity equations
(dependence of the flow on a). It is of course possible to use numerical derivatives
in this calculation. Because of the complexity of the implicit differentiations involved
in computing the "true" derivatives, we preferred to use linear approximations to
the stable and unstable manifolds, and progressively "squeeze" them very close to
the saddle (where they become increasingly accurate), as opposed to higher order
approximations to the local manifolds as in [20, 28]; these approximations are more
accurate further away from the saddle, but evaluating their dependence on the pa-
rameter becomes increasingly complicated.
Figure 13 shows the convergence rate of the Newton iteration for the location of
the Silnikov connection at a ~ 35.2741195. The "elbows" in the graph correspond
to changes in r. The figure has been obtained by letting D(a, E, r) approach zero
to within 10- 6 before shrinking the cylinder. Upon convergence of the iteration for
a prescribed value of r, the new r has been chosen by allowing the last computed
 27

1.2 1.2
(b)

1.0 1.0

.. ..
..
e" ..
.
"
0
"~ ..
0

.i" .2 .i" .2

-.2

-.4 OL.-..~.'~~1.-:-0~"''''''''''2:-'-.0,....,2:'-:
.•,.......,'~.0,.......,,"='
.• ~ •.':-0~'.5 - ., O'-'--C.,:-'-:-'1.0,.......,''"'.':-"-:2":.0:-'-:2"='.,'"-::",.':-0-'-.:",.-:-,-'--:-'
•. :-0'-:"
•. ,
(U,COS(X» (U,COS(X»

Figure 14: Phase space projections of a converged Silnikov connection (a

35.2741195) and a limit cycle near the the tip of the corkscrew (a = 35.27398).

trajectory to further approach the fixed point as close as possible before it starts
moving away from it. The reason for not immediately starting with a very small
(accurate) r is that simple shooting may fail to return to the cylinder if the initial
guess of a is not very accurate. This probably means that convergence for multiple
shooting or collocation discretizations for the boundary value problem would be easier.
Friedman and Doedel, in a series of papers [14, 15, 16], have discussed various issues
associated with computing homoclinic and heteroclinic connections between steady
states in a more general setting. Their implementation of these algorithms in a
version of AUTO will undoubtedly make such calculations routinely possible. Figure
14a shows our approximation of the single pass Silnikov connection (a ~ 35.2741195)
and 14b shows a nearby limit cycle computed by AUTO (in this case the unstable
manifold was actually two dimensional and the stable manifold was one dimensional,
but the Silnikov condition was still obeyed).

Summary and conclusions. A detailed discussion of the spatio-temporal

patterns associated with certain limit cycle solution branches of the KSE has been
presented for the parameter interval 0 < a < 40. A particular space-time symmetry
("Ponies On a Merry-go-round") of one of these branches, its origin and its effect on
subsequent bifurcations has also been discussed. We computationally observe that all
limit cycle branches studied apparently "terminate" in homo clinic (or heteroclinic)
Silnikov connections, involving saddle type steady state solutions. Numerical evidence
for the occurrence of a number of features generically associated with a Silnikov loop
(period doublings, subsidiary homoclinicities, etc.) has also been presented. The
minimal phase space dimension consistent with these observations is three; yet a
traditional three mode Galerkin truncation of the PDE does not capture this behavior.
Exploiting the theory of Approximate Inertial Manifolds (AIMs), we were able to
construct a minimal (three dimensional) model capable of qualitatively (and almost
quantitatively) reproducing the bifurcation structure.
The sharp reduction in dimensionality allows for the easier visualization of the
(approximate) long term dynamics of the PDE; in our three dimensional model the
highest dimension of the interacting stable and unstable manifolds of the saddle-
focus is two. We exploited this tractable geometry in implementing a simple shooting
 28

algorithm to accurately locate the Silnikov connections. More generally, accurate

model reduction for PDE discretizations is important in allowing the approximation
of (low-dimensional) stable manifolds of steady or oscillatory solutions further away
from these solutions in phase space. This should allow for an easier geometric under-
standing of the phase space structures underlying long term dynamics of the PDE,
especially global bifurcations. Our particular example, the KSE, is very well suited
for model reduction using AIMs: dissipation is strong, the first few eigenfunctions of
the linear "dissipative" part, used to parameterize the AIM, also happen to be the
eigendirections in which the first few instabilities occur and carry most of the energy
of the long term solutions. In addition, the spectral implementation of the AIM ben-
efits from the fact that these eigenfunctions are pure Fourier modes. For other PDEs
and other geometries, it is quite possible that alternative basis functions (e.g. based
on some empirical information about the problem, as is the case in POD) can be
much more efficient in parametrizing an appropriate low-dimensional manifold, and
hence give rise to an accurate dynamical system of smaller dimension.

Acknowledgements: We are deeply indebted to Professor Edriss S. Titi for

his invaluable contributions to many aspects of this research. The work of H.S.B.
and LG.K. was supported in part by NSF Grants CTS-8957213, DMS-8906292 and
a David and Lucile Packard Foundation Fellowship; that of M.S.J. was supported in
part by NSF grant DMS-9007802. Part of this work was performed while LG.K. was
enjoying the hospitality (and M.S.J. was a postdoctoral member) of the Institute for
Mathematics and its Applications at the University of Minnesota.

References

[1] D. Armbruster, J. Guckenheimer, and P. Holmes. Kuramoto-Sivashinsky dy-

namics on the center-unstable manifold. SIAM J. Appl. Math., 49(3):676-691,
June 1989.

[2] D. G. Aronson, M. Golubitsky, and J. Mallet-Paret. Ponies on a merry-go-round

in large arrays of Josephson junctions. Nonlinearity (to appear), 1991.

[3] P. J. Aston. Introduction to the numerical solution of symmetry-breaking bifur-

cation problems. In D. Roose, B. DeDier, and A. Spence, editors, Proceedings
on the NATO Advanced Research Workshop on Continuation and Bifurcations:
Numerical Techniques and Applications, pages 139-152. Kluwer Academic Pub-
lishers, 1990.

[4] N. Aubry, P. Holmes, J. Lumley, and E. Stone. The dynamics of coherent struc-
tures in the wall region of a turbulent boundary layer. J. F. M., 192:115-173,
1988.

[5] W. J. Beyn. Global bifurcations and their numerical computation. In D. Roose,

B. DeDier, and A. Spence, editors, Proceedings on the NATO Advanced Research
Workshop on Continuation and Bifurcations: Numerical Techniques and Appli-
cations, pages 169-181. Kluwer Academic Publishers, 1990.
 29

[6] W. J. Beyn. The numerical computation of connecting orbits in dynamical sys-

tems. IMA J. Num. Anal., 9:379-405, 1990.

[7] H. S. Brown, M. S. Jolly, I. G. Kevrekidis, and E. S. Titi. Use of approxi-

mate inertial manifolds in bifurcation calculations. In D. Roose, B. DeDier, and
A. Spence, editors, Proceedings on the NATO Advanced Research Workshop on
Continuation and Bifurcations: Numerical Techniques and Applications, pages
9-23. Kluwer Academic Publishers, 1990.

[8] H. S. Brown and I. G. Kevrekidis. A numerical study of homoclinic orbits to

a saddle-focus and dynamics associated with their existence. Paper 187e, Pre-
sented to the 1990 AIChE annual meeting, Chicago, 1990.

[9] L. H. Chen and H. C. Chang. Nonlinear waves on thin film surfaces. II. bifurcation
analysis of the long-wave equation. Chem. Eng. Sci., 41:2477-2486, 1986.

[10] P. Constantin, C. Foias, B. Nicolaenko, and R. Temam. Integral Manifolds and

Inertial Manifolds for Dissipative Partial Differential Equations. Appl. Math.
Sci., No. 70. Springer-Verlag, 1988.

[11] A. E. Deane, I. G. Kevrekidis, G. E. Karniadakis, and S. A. Orszag. Low dimen-

sional models for complex geometry flows: Application to grooved channels and
circular cylinders. Phys. Fluids A (Submitted), 1990.

[12] E. S. Doedel. AUTO: A program for the automatic bifurcation analysis of au-
tonomous systems. Congo Num., 30:265-284, 1981.

[13] E. S. Doedel. A UTa 86 User Manual, February 1986.

[14] E. S. Doedel and M. J. Friedman. Numerical computation of heteroclinic orbits.

J. Comput. and Appl. Math., 26:159-170, 1989.

[15] M. J. Friedman and E. Doedel. Numerical computation and continuation of

invariant manifolds connecting fixed points with application to computation to
combustion fronts. In Proc. 7th Int. Con! on Finite Element Methods in flow
problems, April 1989.

[16] M. J. Friedman and E. Doedel. Numerical computation and continuation of

invariant manifolds connecting fixed points. SIAM J. Num. Anal., 1991. to
appear.

[17] P. Gaspard, R. Kapral, and G. Nicolis. Bifurcation phenomena near homo clinic
systems: A two-parameter analysis. J. Stat. Phys., 35(5/6):697-727, 1984.

[18] P. Glendinning and C. Sparrow. Local and global behavior near homoclinic
orbits. J. Stat. Phys., 35:645-696, 1984.

[19] J. M. Greene and J. S. Kim. The steady states of the Kuramoto-Sivashinsky

equation. Physica D, 33:99-120, 1988.

[20] B. Hassard. Computation of invariant manifolds. In P. Holmes, editor, New

Approaches to Nonlinear Problems in Dynamics, pages 27-42. SIAM, 1980.
 30

[21] J. M. Hyman, B. Nicolaenko, and S. Zaleski. Order and complexity in the

Kuramoto-Sivashinsky model of weakly turbulent interfaces. Physica D, 23:265-
292, 1986.

[22] J. S. Il'yashenko. Global analysis of the phase portrait for the Kuramoto-
Sivashinsky equation. Technical Report 665, IMA Preprint, 1990.

[23] M. S. Jolly, 1. G. Kevrekidis, and E. S. Titi. Approximate inertial manifolds for

the Kuramoto-Sivashinsky equation: Analysis and computations. Physica D,
44:38-60, 1990.

[24] 1. G. Kevrekidis, H. S. Brown, and M. S. Jolly. Low-dimensional chaos in models

of interfacial instabilities: Global bifurcations and the role of unstable solutions.
Paper 16e, Presented to the 1989 AIChE annual meeting, San Francisco, 1989.

[25] 1. G. Kevrekidis and M. S. Jolly. On the computation of inertial manifolds. Paper

173c, Presented to the 1988 AIChE annual meeting, Washington, 1988.

[26] 1. G. Kevrekidis, B. Nicolaenko, and J. C. Scovel. Back in the saddle again: A

computer assisted study of the Kuramoto-Sivashinsky equation. SIAM J. Appl.
Math., 50(3):760-790, 1990.

[27] Y. Kuramoto and T. Tsuzuki. Persistent propagation of concentration waves in

dissipative media far from thermal equilibrium. Prog. Theor. Phys., 55(2):356-
369, February 1976.

[28] Y. A. Kuznetsov. Computation of invariant manifold bifurcations. In D. Roose,

B. DeDier, and A. Spence, editors, Proceedings on the NATO Advanced Research
Workshop on Continuation and Bifurcations: Numerical Techniques and Appli-
cations, pages 183-195. Kluwer Academic Publishers, 1990.

[29] M. Lentini and H. B. Keller. Boundary value problems on semi-infinite intervals

and their numerical solution. SIAM J. Num. Anal., 17(4):577-604, August 1980.

[30] J. L. Lumley. The structure of inhomogeneous turbulent flows. In A. M. Yaglom

and V. 1. Tatarski, editors, Atmospheric Turbulence and Radio Wave Propaga-
tion, pages 166-178. Moscow: Nauka, 1967.

[31] J. L. Lumley. Stochastic Tools in Turbulence. Academic Press, NY, 1970.

[32] M. Luskin and G. Sell. Approximation theories for inertial manifolds. Technical
Report 88/76, University of Minnesota Supercomputer Institute, July 1988.

[33] B. Nicolaenko, B. Scheurer, and R. Temam. Some global dynamical properties of

the Kuramoto Sivashinsky equation: Nonlinear stability and attractors. Physic a
D, 16:155-183, 1985.

[34] J. C. Scovel, 1. G. Kevrekidis, and B. Nicolaenko. Scaling laws and the prediction
of bifurcations in systems modeling pattern formation. Phys. Letters A, 130:73-
80, 1988.

[35] L. Sirovich and C. Sirovich. Low dimensional description of complicated phe-

nomena. Contemporary Mathematics, 89:277-365, 1989.
 31

[36] G. I. Sivashinsky. Nonlinear analysis of hydrodynamic instability in laminar

flames-I. derivation of basic equations. Acta Astronautica, 4:1177-1206, 1977.

[37] C. Sparrow. The Lorenz Equations: Bifurcations, Chaos and Strange Attractors.
Appl. Math. Sci., No. 41. Springer-Verlag, 1982.

[38] J. W. Swift and K. Wiesenfeld. Suppression of period doubling in symmetric

systems. Phys. Rev. Letters, 52(9):705-708, February 1984.

[39] E. S. Titi. On approximate inertial manifolds to the Navier-Stokes equations.

J. Math. Anal. Appl., 149:540-557, 1990.
 LOCALIZED AND EXTENDED PATTERNS IN
REACTIVE MEDIA

CHRISTIAN ELPHICK* AND EHUD MERONt

Abstract. We study the dynamics of interacting localized structures in homogeneous reactive

media. Equations of motion for solitary waves in excitable media and for vortices in oscillatory
media are derived under the assumption of weak interactions. We show that excitable media with
oscillatory recovery can support a multitude of stable, nonuniform spatial patterns and that phase
field effects in oscillatory media may lead to the formation of bound vortex pairs. The implications
of the latter result on the transition to turbulence in oscillatory media are discussed.

1. Introduction. Reactive media provide important case studies for pattern

formation problems in far from equilibrium systems. Reactive solutions with ex-
citable chemical kinetics can be used as homogeneous and isotropic model systems
in the study of various wave phenomena in biological systems [1,2], whereas oscillat-
ing chemical reactions are useful for studying phase dynamics and weak turbulence
[3].
A significant feature of reactive media is the existence of localized spatial struc-
tures. The simplest of these are the one-dimensional solitary waves of chemical
activity that propagate in quiescent excitable media. These waves constitute the
chemical analog of action potentials in biological membranes [1,2,4]. In two space
dimensions vortex-like structures are known to exist which take the form of rotat-
ing spiral waves [5]. Very often, extended patterns can be viewed as systems of
many interacting localized ones. We will adopt here this approach but consider
only sparse patterns in which collisions among localized structures are avoided.
In addressing the problem of interacting localized structures two questions arise:
what degrees of freedom are relevant for the asymptotic time evolution, and how to
derive their dynamics in the presence of perturbations. We will see in section 2 that
the key to these questions rests in the continuous symmetries of the system. The
potential degrees of freedom of a localized pattern that breaks these symmetries are
the corresponding symmetries parameters (position in the case of translational sym-
metry, etc.). In the presence of symmetry breaking perturbations these parameters
become slow dynamical variables. Evolution equations for these variables appear as
solvability conditions which guarantee that the solution of the unperturbed prob-
lem, but with time dependent symmetry parameters, is an approximate solution of
the perturbed equations. In section 3 we will use these ideas to derive the time
evolution of wavetrains of impulses that propagate along one-dimensional media,
and show how spatially complex patterns may emerge in excitable systems with
oscillatory recovery.
There are situations in which the interactions among localized structures are
mediated by extended fields. An example will be presented in section 4 where the

*Physics Department, Universidad Teenica F. Santa Maria, Casilla llO-V, Valparaiso, Chile.
tDepartment of Chemical Physics, Weizmann Institute of Science, Rehovot i6100, Israel.
 34

dynamics of vortices in oscillating, two-dimensional media will be studied. It will

be shown in this section that phase field effects may lead to the formation of bound
vortex pairs. Possible implications of this result on the transition to turbulence will
be discussed in section 5.

2. Dynamics of Localized Structures. Reactive systems, in the absence of

convection and thermal effects, are normally described by reaction-diffusion equa-
tions (rde's) of the form

(2.1)

where V is a set of concentration fields, R(V) represents the mass-action kinetics

and D is a matrix of diffusion constants. For the sake of illustration we confine
ourselves in this section to patterns in one space dimension. We assume that (2.1)
has an equilibrium solution, V = 0, and solutions which describe localized patterns
such as solitary waves that propagate at constant speed, Vex, t) = H(x - cot).
We also assume that (1.1) describes a homogenous medium and therefore invariant
under translations. The latter assumption implies that the translated form H(X -
Xo), where X := x - cot, is also a solution of (2.1).
Consider now a perturbation, P, that breaks the translational invariance of the
system. The perturbation can be external or due to other localized structures in
the neighborhood. We will consider in this section the former case, postponing the
many body problem of interacting structures to the next section. The main effect of
such a perturbation is to make the symmetry parameter XO a slow dynamical vari-
able, Xo( €i), where c is a small parameter that measures the size of the perturbation.
The rational for that derives from the marginal stability of H to translational per-
turbations. An equation of motion for XO (d) now follows from the requirement that
H(X - Xo(d») is an approximate solution of the perturbed problem

(2.2) DtV = R(V) + DD;V + cPo

To see this we write the solution in the form

(2.3) V = H(X - Xo(d») + cC (X, ct),

insert it in (2.2) and evaluate an equation for the correction term C. To leading
order in € we obtain

(2.4) .cC = X~H' +P

where the prime denotes differentiation with respect to the argument. In order for
H(X - xo(d») to be an approximate solution of (2.2), solutions C of (2.4) should be
finite. We are thus led to the search of zero eigenmodes (or null vectors) of .c. If such
eigenmodes exist, solvability conditions should be employed to eliminate singular
components from C. It is easy to see that .c has indeed a zero eigenmode. The
 35

solitary-wave solution H(X) satisfies the unperturbed equation (2.1). Differentiating

this equation with respect to X = x - cot we find

(2.5) CH' = o.
We may call H' the translation mode since translational perturbations, H(X + OX),
are along that mode. Equation (2.5) suggests the existence of a zero eigenmode of
the adjoint operator as well [6],

(2.6)

where the adjoint property is defined with respect to an appropriately chosen inner
product. Taking the inner product of (2.6) with the correction term C we find

(2.7) Xo = -(G, EP)/(G, H'),

where the overdot denotes differentiation with respect to time. An equation of
motion for XO follows once we realize that G(X - Xo) is localized around X = XO
and P is an explicit function of x. This allows us to express the right hand side
(rhs) of (2.7) as a function of Xo, a function which reflects the form ofthe symmetry
breaking perturbation P.
We have considered a system that has only one continuous symmetry, the sym-
metry of translations. As a result the asymptotic dynamics of a localized pattern
in such a system is governed by a first order, ordinary differential equation (ode)
for the corresponding symmetry parameter Xo. The method introduced here can
be extended to systems possessing a number of continuous symmetries. The time
evolution of structures that break these symmetries will be described by sets of first
order ode's for the pertinent symmetry parameters. An example of this richer case
has recently been studied [7].

3. Extended Patterns in One Space Dimension. In the previous section

we considered the dynamics of a localized structure in the presence of an external
perturbation P. We proposed an approximate solution of the perturbed problem
(2.2) in the form (2.3) where the time evolution of Xo( Et) is determined by (2.7). In
this section we consider the dynamics of widely spaced wavetrains of solitary waves
or impulses. Here, even in the absence of external perturbations, one may expect
for non-trivial dynamics, for impulses are perturbed by the fields of their neighbors.
To describe the dynamics of widely spaced wavetrains we look for solutions of
(2.1) of the form

(3.1) U = LH(X - Xj(Et)) + EC(X,et),

j

where E is a measure of the perturbations that are exerted by nearby impulses. To

be more specific about the small parameter E we need to consider the asymptotic
forms of the solitary wave solution, H(X), as X ~ ±oo. These forms are generically
 36

exponential functions for they follow from a linear analysis of (2.1) around the
equilibrium state U = O. We thus write

(3.2a)

(3.2b) H(X) ~ e~LX COS(VLX + <l>d X -+ -00.

We can define now our small parameter to be € = exp( -1]>") where I] = min(I]R' I]d
and >.. is a typical spacing between successive impulses. Clearly, in the limit of
infinite spacings, € = 0 and the form (3.1) reduces to the solitary solution H.
Equations of motion for the impulse positions Xj (in a frame moving with ve-
locity co) follow from solvability conditions which guarantee the finiteness of the
correction term C. The derivation goes along the lines depicted in the previous
section and the result takes a form similar to that of (2.7):

(3.3)

where P is given by

(3.4) €P = R(LH j ) - LR(H j ),

j

and F j := F(X - Xj). In (2.7) P stands for a symmetry breaking external per-
turbation. Here, it represents spatial nonhomogeneities generated by the impulses
themselves. Notice that only nonlinear terms in R(U) contribute to (3.4) and that
at any point in space the rhs of (3.4) is at most of 0(1'). To evaluate the rhs of (3.4)
we need to define an inner product. We assume the usual inner product of vector
analysis and use integration over the infinite range of X for functional products.
Exploiting the localized nature of Hand G and using the asymptotic forms (3.2)
we find to leading order in E [8,9]

Xj = aRe -~R(X; -xHd COS[VR(Xj - Xj+l) +.,p R]

(3.5)
+aLe-~dx;-l-X;) COS[VL(Xj-l - Xj) + .,pd,

where aR,aL,.,pR and .,pL are parameters that can be calculated for a specific
realization of the reaction-diffusion system (2.1) and Xj > Xi for j < i.
Equations like (3.5) describe the time evolution of widely spaced wavetrains of
impulses. The velocity of a given impulse is determined by the local field-values of
nearby impulses. These values are exponentially small, yet significant because of the
marginal stability of the solitary-wave solution to translations. The present theory
is particularly applicable to excitable media where impulse collisions are naturally
avoided [5]. In such media the rate of growth, I]R, is normally much larger than
the rate of decay, IlL. Consequently, when the distribution of impulse spacings
is sufficiently narrow, the first term on the rhs of (3.5) can be neglected and the
impulse equations simplify to
 37

The asymptotic fonn (3.2b) with I/L # 0 describes a medium with (damped)
oscillatory recovery. We will show now that this manner of recovery gives rise to
a form of spatial complexity. Consider a train of N impulses. The first impulse is
stationary in a frame moving with speed Co, for there is no impulse ahead of it to
perturb its motion. Thus ~) = O. All other impulses (j = 2, ... , N) are affected by
the recovery phase of the medium just ahead of them and satisfy (3.6). Steady state
solutions of (3.6) (Xi = 0) describe trains of impulses that propagate with uniform
speed co. There are many ways to construct such solutions, for any configuration
of impulses with spacings Aj := Xj-l - Xj such that HLP..j) = 0 is a solution.
The criterion for stability, H~ p.. j) > 0 for any spacing in the train, still leaves us
with many solutions. We therefore conclude that a multitude of stable wavetrains,
mostly nonuniform, can exist in the case of oscillatory recovery. As the number
of impulses, N, becomes large the systems develops extreme sensitivity to initial
conditions and to external noise.
When the exponents TJR and TJL are of similar size and the head of the solitary
wave is monotonic (I/R = 0) one can still understand asymptotic solutions of (3.5)
rather easily. Setting Xj = canst. in (3.5) we find a pattern map which relates a
given spacing to the successive one [8,9). Periodic wavetrains and finite wavetrains
correspond, respectively, to periodic and homoclinic orbits of that map. The kind
of spatial complexity discussed above pertains to a pattern map in a chaotic regime.

4. Interacting Vortices in Two Space Dimensions. The localized struc-

tures of one-dimensional systems generalize, in two dimensions, to wavefronts that
are localized in one direction but extend uniformly in the other direction. Expand-
ing ring patterns in excitable media provide an example. There is, however, a new
kind of localized structure in two dimensions. In excitable media it corresponds
to the free end of a terminated wavefront, while in oscillating media it appears as
a point where the amplitude of oscillations vanishes. In both cases the patterns
that surround the structures take the form of spiral waves or vortices [3,5,10]. The
dynamics of a single spiral wave pose a number of interesting questions [1,11-14)
which we will not dwell upon here. We will proceed directly to the many body
problem of interacting vortices and consider primarily the more tractable case of
oscillating media.
Suppose that at a given value of some control parameter the rde's (2.1) undergo
a supercritical Hopf bifurcation. The equilibrium state, U = 0, becomes unstable
and a new, oscillatory state emerges:
(4.1) U(r,t) = Z(r,t)exp(iwot)U o +c.c ..
Here, Z is a complex amplitude that depends weakly on space and time, Wo is the
frequency of a unifonnly oscillating medium and U o is a constant vector. Close
enough to the onset of oscillations the amplitude Z satisfies the equation [3 )
(4.2)
where fl, a and f3 are real constants. Equation (4.2) has spiral-wave solutions of
the form
( 4.3) Zv( r, t) = z( r )e iwt , z( r) = A( r )e in 8+ i.( r) ?
 38

where r and () are polar coordinates in the plane, n is an integer and A( r) and s( r)
have the asymptotic forms [3,15]:

A(r) '" rlnl, s(r) '" r lnl +1 r ----+ 0

(4.4)
A(r) '" Ao + clr, s(r) '" pr + qlnr r ----+ 00.

Here, w = -ap+(a-;3)p2, Ao = .j(p_p2), c = p(1+;32)/[2Ao(a-;3)], q = -(1+

a;3)/[2( a - ;3)], and p = pep, a,;3) is the selected asymptotic wavenumber for all
isolated vortex [15,16]. The integer n in (4.3) is given by the integral (27r)-1 V'P.drli
over a closed loop that contains the spiral core, where 'P = arg( z), and is known a~
the topological charge or the winding number of the spiral wave (vortex). i

In section 3 we used the solitary-wave solution as the basic building block fOI
construction of extended wavetrains. Here, we will use the single vortex solutioIl·
(4.3) to construct multi-vortex states. As before, we utilize the continuous symme.L
tries of the system, this time the symmetries of (4.2): space translations, Z(r) -.
Z(r - ro), and rotations in the complex Z plane, Z ----+ Z exp( i1jJ). Acting with thesE
symmetries on (4.3) we find another solution of (4.2), z(r - ro)exp(iwt + i1jJ). k
multi-vortex solution is now approximated by

(4.5)

where zj(r) is given by (4.3) with winding number nj. If we allow the symmetry
parameters r j to depend on time and 1jJ to depend on space and time we find
solvability conditions, in the form of evolution equations, which guarantee that (4.5 {
is indeed an approximate solution of (4.2). To write down the evolution equatiom
we find it useful to define a "center of mass" coordinate I
i

and a new phase

¢>(r - Rem, i) = 1jJ(r, i) + (N -l)plr - Reml.

In terms of these quantities the evolution equations read

(4.6)

at¢> - Rem· V¢> = 2p(a - ;3)arc~ ¢> + (1 + a;3)V2¢> + (a - ;3)(V¢»2

(4.7) _2p2 A;2(1 + (2)a;c~ ¢> + pA;2( a - ;3)(1 - a;3)V2ar,~ ¢>
1
_pA;2(a + ;3)(1 + a;3)arc~ V2¢> - 2A;2;32(1 + ( 2)V4¢> + h.o.i ..
 39

Here, k is a unit vector perpendicular to the plane and rem == Ir-Reml. In deriving
equations (4.6) and (4.7) we assumed that p is a small parameter (p2 ~ p) and that
'I1,p and Iri - rjl-1 are of O(p). In physical terms, we demand that the inter-vortex
distance and the asymptotic wavelength are both much larger than the size of the
vortex core. Thus (4.7) includes terms to O(p4). For a vortex-free and uniform
medium (p = 0) equation (4.7) reduces to the Kuramoto-Sivashinsky equation [3).
The requirement that the coefficient of o;cm ,p vanishes gives an Eckhaus insta-
bility relation [3,17]

(4.8)

This relation together with the form, p = pep, a, (3), of the asymptotic wavenumber
of the spiral solution (4.3), define a range of parameters in which the truncated
form of (4.7),

is expected to be valid.
Equations (4.6) and (4.9) can be used to derive the interaction between a pair
of vortices. We will consider here the simpler case of n1 = n2 = 1 and look for
axisymetric solutions, ,p = ,p(rem), of (4.9). It is straightforward to show that in
this case Rem = O. Equation (4.9) simplifies then considerably and can be solved
by a Hopf-Cole transformation. We obtain

(4.10)

where
Dt = _p2(oo - (3)(1 - A~), 7rt = p(l - At) and at = pq-1 At
with At = (2€ + 1)-1 (€ E Z+). Pl(x) is a polynomial of degree € that solves the
linear differential equation

(4.11 )

Equation (4.11) can easily be solved for any degree C. For x » 1 the solution
simplifies to Pt(x) :::::: xl. Such a condition is attainable when considering large
inter-vortex distances (compared with p-1) or when q ex (1 + 00(3) becomes small
(i.e. close to the Eckhalls instability curve). Using (4.10) in (4.6) we then find

(4.12a) ~P = -(37rl + [1 - (3q(4C - 1)]~,

p. 1
(4.12b) -{}
4
= 7rt + [(3 + q(4C - 1)]-,
p

where p == Ir2 - r11 is the inter-vortex distance and {} is the angle between r2 - r1
and a fixed direction in. the plane.
 40

Equations (4.12) describe the dynamics of a vortex pair with winding numbers
nl =n2 = 1. Stable steady state solutions of (4.12a) correspond to bound pairs

of vortices. For f3 =1= 0 such solutions do exist. The inter-vortex distances of these
pairs are given by
* 1- f3q( 4£ - 1)
( 4.13)
Pi = f37ri '
while the condition for stability is f3p > O. When f3q is positive only a few bound-pair
states exist. They correspond to inter-vortex distances pi with £ = 1, ... , £max, where
£max is the integer part of (1 + f3q)/4f3q. For f3q negative a whole family of bound-
pair states exists (that is, £ can be any positive integer). Another consequence of
(4.13) is that bound pairs of vortices (with equal winding numbers) rotate with
constant angular velocity, J(pi).
The bound-pair solutions disappear when f3 --+ 0 despite the fact that vortices
with spiral structures still exist in that limit. The parameter f3 is a measure of
the size of phase-field contributions to vortex interactions. For small f3 values the
phase-field effect becomes significant only at large inter-vortex distances where the
1/ p repulsion term is sufficiently small. In that case only vortices whose distances
are large enough (of order (f3p )-1) can form bound pairs. At relatively large f3
values the phase field is already effective at distances of O(p-l).
The fact that vortices in oscillating media can form bound pairs should be at-
tributed to the mediating effect of the spiral phase field; vortices with rectilinear
equiphases either attract or repell each other. A somewhat similar situation applies
to excitable media. Recent numerical simulations [18] indicate that bound vortex-
pairs may exist in this kind of medium as well. For winding numbers nl = n2 = 1
the vortices rotate around a fixed center while for nl = -n2 = 1 they drift in a
direction normal to rl - r2 (drifting vortex pairs have also been found in numerical
simulations on (4.2) [19]). The difference between oscillating and excitable media
seems to lie in the nature of the mediating spiral field. In the former it is an extended
sinusoidal field while in the latter it resembles an array of localized solitary wave-
fronts [14]. Consequently vortices in excitable media which are a few wavelengths
apart appear independent [18]; the interaction between the two cores is screened by
the surrounding spiral wavefronts.

5. Discussion. We have presented here an effort to understand extended pat-

terns in reactive media in terms of their localized constituents: solitary waves in
one dimension, vortices in two dimensions. Evolution equations for these struc-
tures have been derived by exploiting the continuous symmetries of the system.
The equations are valid for sparse patterns where typical distances between nearby
structures are considerably larger than their widths. The approach pursued here
does not allow for annihilation and nucleation of structures and therefore does not
apply, in its present form, to turbulent patterns where such processes are important.
Yet, results obtained with this approach may bear on the transition to turbulence,
as we now discuss.
Recent numerical simulations on (4.2) suggest that the transition to turbulence
in oscillating media may occur by the spontaneous nucleation of vortices [20]. Above
 41

the transition, vortices are continuously being created and annihilated so as to keep
a steady mean number. The result that vortices can form bound pairs may have the
interesting consequence of making the transition hysteretic. Suppose we prepare a
system to contain only one vortex and drive it to the turbulent regime. This can
be done in a number of ways, but we choose to consider here the path in parameter
space, a = -fJ < 0, for which numerical data on the selected asymptotic wavenum-
ber, p = p("" a, fJ), are available [16]. Upon increasing fJ, a critical value fJt is
reached where phase instabilities develop. This value, detennined by (4.8) and the
asymptotic wavenumber p, marks the onset of turbulence since phase instabilities
seem to provide the driving force needed for the nucleation of additional vortices
[20]. On going backward below fJt, vortices which are sufficiently apart from each
other to form bound pairs, will not annihilate each other and a multi-vortex state
will result. On further decreasing fJ, the minimal distance to form bound pairs will
increase because it is proportional to (fJp )-1 and p decreases with fJ = -a [16,19].
As a result more and more annihilation events will take place until a value fJ-; < fJt
is reached where the minimal distance exceeds the size of the system and the initial
one-vortex state is recovered.
Vortex-initiated turbulence appears to be a general phenomenon in two-dimen-
sional nonequilibrium systems; similar transitions have recently been observed in
electroconvecting nematic liquid crystals [21] and in capillary ripples [22]. It would
be of interest to know whether this type of transition may occur in excitable media
as well.

REFERENCES

[1] A. T. Winfree, When Time Breaks Down, Princeton, New Jersey, 1987.
[2] J. J. Tyson and J. Keener, Singular perturbation theory of traveling waves in excitable media
(a review), Physica D, 32 (1988), pp. 327-36l.
[3] Y. Kuramoto, Chemical Oscillations, lVaves, and Turbulence, Springer, Berlin, 1984.
[4] A. L. Hodgkin and A. F. Huxley, A quantitative description of membrane current and its
application to conduction and excitation in nerve, J. Physio!', 117 (1952), pp. 500-544. Springer-
Verlag, Berlin, 1984.
[5] A. T. Winfree, Spiral waves of chemical activity, Science 175 (1972), pp. 634-636.
[6] B. Friedman, Principles and Techniques of Applied Mathematics, Chapman and Hall Ltd.,
London, 1956.
[7] C. Elphick and E. Meron, Localized structures in surface waves, Phys. Rev. A, 40 (1989), pp.
3226-3230
[8] C. Elphick, E. Meron and E. A. Spiegel, Patterns of propagating pulses, to appear in SIAM J.
App!. Math ..
[9] C. Elphick, E. Meron and E. A. Spiegel, Spatiotemporal complexity in traveling patterns, Phys.
Rev. Lett., 61 (1988), pp. 496-499.
[10] The manner by which a rectilinear wavefront with a free end evolves toward a rotating spiral
wave has been studied in E. Meron and P. Pelce, Model for spiral wave formation in excitable
media, Phys. Rev. Lett., 60 (1988), pp. 1880-1883.
[11] W. Jahnke, W. E. Skaggs, and A. T. Winfree, Chemical vortex dynamics in the Belousov-
Zhabotinsky reaction and in the two- variable Oregonator Model, J. Phys. Chern., 93 (1989),
pp. 740-749, and references therein.
[12] S. C. Muller and B. Hess, Nonlinear dynamics in chemical systems, in Cooperative Dynamics
in Complex Physical Systems, Ed. H. Takayama, Springer, Berlin, 1989.
[13] K. I. Agladze, V. A. Davydov, and A. S. Mikhailov, Observation of a helical-wave resonance in
an excitable distributed medium, Pis'ma Zh. Eksp. Teor. Fiz., 45 (1987) pp. 601-603.
[14] E. Meron, Nonlocal effects in spiral waves, Phys. Rev. Lett., 63 (1989), pp. 684-687.
 42

[15] P. S. Hagan, Spiral waves in reaction-diffusion equations, SIAM J. App\. Math., 42 (1982), pp.
762-786.
[16] E. Bodenschatz, Muster und defekte im rahmen der schwach nichtlinearen analyse von anisotropen
structurbildenden systemen, Ph.D. Thesis, University of Bayreuth, 1989.
[17] B. A. Malomed, Nonlinear waves in nonequilibrium systems of the oscillatory type, part 1, Z.
Phys. B, 55 (1984), 241-248.
[18] E. A. Ermakova, A. M. Pertsov, and E. E. Shnol, On the interaction of vortices in two-
dimensional active media, Physica D, 40 (1989), 185-195.
[19] E. Bodenschatz, M. Kaiser, L. Kramer, W. Pesch, A. Weber and W. Zimmermann, Patterns
and defects in liquid crystals, to appear in New Trends in Nonlinear Dynamics and Pattern
Forming Phenomena: The Geometry of Nonequilibrium, Eds. P. Coullet and P. Huerre, Plenum
Press, 1989.
[20] P. Coullet, L. Gil, and J. Lega, Defect-mediated turbulence, Phys. Rev. Lett., 62 (1989), 1619-
1622.
[21] I. Rehberg, Steffen Rasenat, and Victor Steinberg, waves and defect-initiated turbulence in
electro convecting nematics, Phys. Rev. Lett., 62 (1989), pp. 756-759.
[22] N. B. Tufillaro, R. Ramshankar, and J. P. Gollub, Order-disorder transition in capillary ripples,
Phys. Rev. Lett., 62 (1989), pp. 422-425.
 SOME RECENT RESULTS IN CHEMICAL REACTION
NETWORK THEORY

MARTIN FEINBERG*

Abstract. The aim of chemical reaction network theory is to draw connections between re-
action network structure and qualitative properties of the corresponding differential equations.
Some recent results are discussed, in particular those relating to the possibility of multiple steady
states in very complex continuous flow stirred tank reactors, to mechanism discrimination in het-
erogeneous catalysis, and to the possibility of traveling composition waves on isothermal catalyst
surfaces

Key words. chemical reaction networks, ordinary differential equations, reaction-diffusion

equations, traveling waves, Conley index, SCL graphs, catalysis, CFSTR.

1. Introduction. Chemists and engineers are required to confront a bewilder-

ing array of distinct chemical systems, each with its own set of molecular species
and its own network of chemical reactions. In a real reactor, the amounts of the
various species will generally be governed by a large first order system of differential
equations, almost always nonlinear (and usually polynomial). The equations often
contain many parameters (e.g., rate constants) which are known only roughly if
they are known at all. To make matters worse, each new chemical system gives rise
to a completely new system of differential equations.
At first glance, the situation would seem hopeless. Despite great progress in
dynamical systems, it remains quite difficult to study a first order system of poly-
nomial differential equations in, say, seven or eight dependent variables. It is worth
remembering that the Lorenz system, with all the complexity it carries, is a polyno-
mial system in only three variables and that Hilbert's Sixteenth Problem is about
the behavior of polynomial equations in only two variables.
The fad is, however, that the differential equations that arise in the study of
chemical reactors have a very special structure. Given a reactor of specified type-for
example, the classical continuous flow stirred tank reactor-undergraduate students
in engineering know (or should know) how to pass from an indicated network of
chemical reactions to the appropriate system of differential equations. In fact, when
reaction rates are governed by the usual mass action kinetics, the polynomial vector
field is completely determined (up values of parameters such as rate constants)
by the structure of the underlying reaction network. It is this precise connection
between reaction network structure and the shape of the governing equations that
gives the mathematics of chemical reactors its special character.
The central question of chemical reaction network theory can be stated in the
following way: What is the relationship between the structure of a reaction network
and the variety of phase portraits that the corresponding differential equations might

*Department of Chemical Engineering, University of Rochester, Rochester, NY, USA 14627.

Work described in this article was supported by National Science Foundation grants to Martin
Feinberg and (for the research on traveling waves) to David Terman.
 44

exhibit (as parameter values vary)? That is, for a specified network, one would
like to have practical means to determine whether there are rate constant values
such that the corresponding differential equations admit Phenomenon X. Here
Phenomenon X might be the existence of multiple rest points, of periodic orbits,
and so on. By a "practical" theory I mean one that can be implemented (perhaps
with the help of a computer) by a chemist or engineer without much training in
mathematics.
Although this won't be a "from scratch" tutorial on reaction network theory,
it is intended as much for mathematicians as for chemists and engineers. In any
case, I hope I can convey a sense of what the problems are and of what some results
look like. Mathematicians who are encountering chemical reactors for the first time
will be able to get a rough feeling for the way differential equations are formulated,
but they might want to consult [5] for a fuller introduction. The same reference
contains a review of results in chemical reaction network theory as they stood a
decade ago. References [6] and [7] cover a lot of the same ground, but from a
perspective informed by more recent theory. Although they too are meant to be
expository, it is probably better to read [5] first.
Two presumptions will be in force for all the systems considered here. First,
the temperature will be understood to be fixed and spatially uniform. In this case,
complexity in the differential equations will derive entirely from the interplay of the
various reactions and not at all from thermal effects. Second, all reaction rates will
be governed by mass action kinetics. This means, for example, that if CA and CB
are the local molar concentrations of species A and B, then the local occurrence
rate per unit volume of a reaction such as A --> Q + R should be proportional to
CA while the rate of a reaction such as A + B ---+ S + T should be proportional
to CACB. (The proportionality constants are the rate constants for the reactions.)
The rough idea here is that rate of the first reaction should depend simply on the
amount of A present while the occurrence rate of the second reaction should reflect
the likelihood of a collision between A and B.
I'll describe results in three areas: (i) the relationship between reaction network
structure and the possibility of multiple rest points in very complex continuous flow
stirred tank reactors, (ii) the problem of mechanism discrimination in heteroge-
neous catalysis, and, finally, (iii) the extent to which the classical mechanisms for
heterogeneous catalysis induce traveling composition waves on catalytic surfaces.
Without much preparatory work, I can penetrate fairly deeply into the first of these
areas, and so I am going to concentrate my efforts there. Discussion of the two
remaining areas will be more abbreviated.
However, I do want to mention in this introductory section that Conley Index
methods (in the hands of Dave Terman) are crucial to the results on traveling waves.
Charlie Conley was very enthusiastic about chemical reaction network theory, and,
for me, his enthusiasm was a source of encouragement when encouragement was
very much needed. At the time, I did not know that Charlie was enthusiastic about
everything.
 45

2. Multiple rest points in homogeneous continuous flow stirred tank

reactors with complicated chemistry. A continuous flow stirred tank reactor
(CFSTR) is a vessel to which reactants are continuously added (in a feed stream)
and from which the reacting mixture is continuously withdrawn (in an efHuent
stream). I shall suppose that the mixture in the vessel is stirred so zealously that it
is homogeneous at each instant. I'll also suppose that the composition of the feed
stream is fixed, that the volumetric flow rates of the feed and efHuent streams are
constant and identical, and that the volume of the mixture in the vessel remains
constant. The residence time of the reactor, which I'll denote by 8 , is the mixture
volume divided by the volumetric flow rate.
I'll begin by considering a CFSTR in which there are five species, A, B, C, E and
F, and I'll suppose that the reactions among the species are those shown (2.1). The

A+B~E
(2.1 ) B+C -.:. F
C~2A

species concentrations in the feed stream are denoted c~, c1, cb, cf and c~. (Here
the concentrations are in moles per volume.) The species concentrations in the
vessel are, at each instant, identical to those in the efHuent stream, but they vary
with time by virtue of the occurrence of chemical reactions. These concentrations
I'll denote by CA,CB,CC,CE and CF.
The governing differential equations are shown in (2.2). Here k j , k2 and k3
are rate constants for the corresponding reactions in (2.1). Terms in which the
rate constants appear account for contributions of the various chemical reactions
to changes in the species concentrations within the vessel. Terms containing the
residence time account for contributions of the feed and efHuent streams.

C~ = (1/8)(c~ - CA) - kjCACB + 2k3CC

c~ = (1/8)(c1- CB) - kjCACB - k2cBCC
(2.2) c:::: = (1/8)(cb - co) - k2cBCC - k3CC
c~ = (1/8)(c~ - CE) + kjCACB
c'p, = (1/8)( c~ - CF) + k2cBCC

It is already apparent that even simple reaction networks can lead to dynamical
equations that are not so easy to study. Different networks lead to different CFSTR
equations. Our interest is in relationship between qualitative properties of the
equations and the structure of the reaction network from which they came.
There are many questions that can be asked, but in this section I'll be lll-
terested in only one: What is the relationship between reaction network structure
and the possibility of multiple positive rest points in the corresponding homogeneous
CFSTR equations? By a positive rest point I mean one in which all the species con-
centrations are positive. In this section, when I say that a reaction network hac. the
 46

capacity for multiple positive rest points, I'll mean that there is some assignment of
a (positive) residence time, of (positive) rate constants, and of (non-negative) feed
concentrations such that the corresponding homogeneous CFSTR equations admit
at least two positive rest points.
It happens that reaction network (2.1) has the capacity for multiple positive rest
points. On the other hand, networks (2.3) and (2.4) do not, so it is already apparent
that there is some subtlety here. At first glance, it would seem extremely difficult

(2.3)

(2.4)

D --+ 2A

to determine whether a more elaborate network such as (2.5) has the capacity for
multiple rest points. So that the full weight of the problem might be felt, I have
displayed in (2.6) the CFSTR equations that correspond to network (2.5).
1 3
A+B +:t D +:t 2C
2 4
5
C+:tE
6
7
A+F+:tG
8
(2.5) 9
2B +:t H
10

13 "'~ 14 12 /./11

I
15
I+.J+:tI<
16
 47

C~ = (1/8)(c~ - CA) - kICACB + k2cD - k7cACF + k8CG

c~ = (1/8)(c~ - CB) - kIcACB + k2cD - 2(k g + k14)c~ + 2k lO cH + 2k13 C[
C::: = + 2k3cD - 2k4c'b -
(1/8)(c~ - cc) k5CC + k6cE
c'v = (1/8)(cb - CD) + kIcACB + k4c'b - (k2 + k3)CD
C'e = (1/8)(cf - CE) + k5CC - k6cE
(2.6) c'p = (1/8)(c~ - CF) - k7cACF + k8CG
c~ = (1/8)(c~ - cG) + k7cACF - k8CG
c~ = (1/8)( c{{ - CH) + k9C~ + k12 c[ - (kID + 1,,11 )CH
c~ = (1/8)(c{ - Cf) + k14C~ + knCll - (k13 + k12 )Cf - kI5 C[CJ + kI6 CK
c'J = (1/8)(c5 - CJ) - kI5 CfCJ + /"16 CK
c~ = (1/8)( c{ - CK) + kI5 CICJ - k I6 CK

I want to tell you how to determine - within about five minutes! - whether
network (2.5) has the capacity to generate multiple rest points. The procedure is
not only simple but also sufficiently delicate as to indicate why, despite the fact that
network (2.1) can give multiple rest points, the slightly different networks (2.3) and
(2.4) cannot.
The theory I'll describe results from joint work with Paul Schlosser, and I am
happy to say that he was the real hero of effort. Proofs and additional results can be
found in Schlosser's PhD thesis [11]. Reference [12] contains a fuller discussion and
more examples. Some earlier work (leading to narrower results) is described in [10].
It should be kept in mind that discussion in this section is specific to homogeneous
CFSTRs. When I say that such-and-such a network has the capacity for multiple
positive rest points, I mean that we were able to construct rate constants, a residence
time, a feed composition and two positive rest points which, when taken together,
satisfy the appropriate homogeneous CFSTR equations.
I am going to tell you how to take a reaction diagram such as (2.5) and construct
from it another diagram, called the SeL Graph. Then I'll state two theorems that
relate the structure of the SCL Graph to the possibility of multiple positive rest
points.
To begin, I want to introduce a small amount of vocabulary. Network (2.7) will
help illustrate the few terms I'll need. The complexes of a reaction network are the

<=! E
(2.7) ./
2F

objects that live at the heads and tails of the reaction arrows. The set of complexes
for network (2.7) is indicated in (2.8). Note that in (2.7) each complex appears

(2.8) {A + B,C,A + D,E,2F}

 48

precisely once and that arrows are drawn to indicate a "reacts to" relation in the set
of complexes. Such a display is called a standard reaction diagram. The standard
reaction diagram induces a partition of the complex set into linkage classes: Note
that the diagram (2.7) is made up of two distinct connected components, one con-
taining the complexes A + Band C, the other containing the complexes A + D, E
and 2F. The linkage classes of a network are just the subsets of complexes that
reside within the individual connected components that make up the standard re-
action diagram. The linkage classes of network (2.7) are shown in (2.9). The way in
which linkage classes are denoted has no real significance, but my practice will be
to number them from top to bottom as they appear in whatever standard reaction
diagram I am considering.

(2.9) Ll = {A + B, C} L2 = {A + D,E,2F}

Now I can tell you how the Species-Complex-Linkage Graph (SCL Graph) for
a network is constructed. Again, I will use network (2.7) as an example. Its SCL
Graph is shown in Figure 2.1. You begin by writing on paper a symbol for each
species in the network and also a symbol for each linkage class. (After a while
you learn how to arrange these things in a nice way.) Then, for each species, you
do the following: If the species appears in a complex within a particular linkage
class, you draw an arc connecting the species to the symbol for that linkage class,
and you label the arc with the complex name. (If the species appears in several
complexes within a particular linkage class, you draw an arc for each such complex.)
For example, species A appears in linkage class Ll (in complex A + B), so an arc
(labeled with the complex name A + B) is drawn between A and L 1 . Species A also
appears in linkage class L2 (in complex A + D), so an arc (labeled A + D) is drawn
between A and L 2 • After proceeding in this way through the entire species set, you
obtain the SCL Graph shown in Figure 2.1.

B
~C
+B

L1 C
A+B

A
D

E F

Figure 2.1. The SCL Graph for Network 2.7

 49

It will be instructive to consider another example. The SCL Graph for network
(2.10) is shown in Figure 2.2. The linkage classes for the network are

Ll = {A+B,C,A+D}, L2 = {2A,E+G} and L3 = {A+E,F}.

Note that species A appears in two complexes within linkage class L l , and so there
are two arcs (labeled A + B and A + D) that connect A to L l .

A+B-+C+=tA+D

(2.10) 2A +=t E +G

A+E+=tF

~~Y~D
. B(CS):'
yA~
(Q cg
~ ~ ~G
":>-E~
--"

F G

Figure 2.2. The SCL Graph for Network 2.10

There is an important distinction between the SCL Graphs shown in Figures 2.1
and 2.2. The second contains two closed loops (Ll -A-LJ and A-L2 -E-L3 -A),
but the first contains none. The presence of loops and their nature bears heavily
on the capacity for multiple positive rest points. In fact, I am already in a position
to state a theorem.
 50

THEOREM 2.1. Consider a chemical reaction network for which the SCL Graph
contains no loops. If the network is governed by mass action kinetics, then the
corresponding isothermal homogeneous CFSTR equations cannot admit multiple
positive rest points, no matter what the feed composition might be and no matter
what (positive) values the residence time and rate constants take.

I promised that you would be able to determine very quickly whether the intri-
cate network (2.5) has the capacity for multiple positive rest points. (Remember
the daunting system of equations shown in (2.6).) The SCL Graph for network
(2.5) is displayed in Figure 2.3. It has no loops, so the system (2.6) cannot admit
multiple positive rest points. That's that.

Jy+J KK
Fr+F
La
I+J L5

A+F
I
G A
B~H
G

A+B
2B

D
c
D
~E
Figure 2.3. The SCL Graph for Network 2.5

What if the SCL Graph does have loops? In this case, there is still a lot that
can be said, but I will need more vocabulary. In particular, I have to tell you about
three kinds of loops: o-loops, c-Ioops and s-loops. These are not mutually exclusive;
a loop in an SCL Graph can, for example, be both an a-loop and an s-loop.
By c-pair in an SCL Graph I mean a pair of (adjacent) arcs that carry the same
complex label. (The letter "c" is supposed to remind you of the word complex.) In
Figure 2.1, for example, the arcs A - Ll and Ll - B constitute a c-pair because
they both carry the complex label A + B. In Figure 2.2, the arcs A - L3 and L3 - E
both carry the complex label A + E, so they constitute a c-pair.
By an o-loop in an SCL Graph I mean a loop containing an odd number of
c-pairs. (The letter "0" is supposed to remind you of the word odd.) In Figure 2.2
 51

the loop A - L2 - E - L3 - A is an o-loop because it contains precisely one c-pair

(with arcs labeled A + E). The same loop contains only one arc of the c-pair labeled
by complex E + G, so that c-pair is not counted in determining whether the loop
is an o-loop. (Both arcs of the c-pair must lie in the loop.) In Figure 2.2 the loop
A - Ll - A contains no c-pairs, so it is not an o-loop.
By a c-Ioop in an SCL Graph I mean a loop made up entirely of c-pairs. Con-
sider, for example, network (2.11). Its SCL Graph is shown in Figure 2.4. (As I
A+B+=tE
B+C+=tF
(2.11 )

indicated earlier, it will be my practice to number linkage classes from top to bottom
as they appear in the standard reaction diagram under consideration.) The only
loop in the figure is a c-Ioop because its set of arcs is the union of c-pairs. (On the
other hand, the loop is not an o-loop because it contains four c-pairs.)

Figure 2.4. The SCL Graph for Network 2.11

I will use network (2.12) to help explain what an s-loop is. The SCL Graph for
(2.12) is displayed in Figure 2.5.
A +D +=t 2B
(2.12) B+=tC
2C +=t A +E
 52

A1 ~~2B
A+E (1 s-loop 1\B
.,sQ @
E ~CX

Figure 2.5. The SCL Graph for Network 2.12

To determine whether a loop is an s-loop, I shall first have to adorn its arcs with
stoichiometric coefficients. When a species appears within a particular complex,
the stoichiometric coefficient of a species within that complex is just the numerical
multiplier of the species. For example, the stoichiometric coefficient of B in the
complex 2B is 2. The stoichiometric coefficient of A in complex A + D is 1. I adorn
the arcs of a loop with stoichiometric coefficients in the following way: With each
arc there is associated a species (at one end of the arc) and a complex (labeling the
arc). Alongside each arc of the loop, I write the stoichiometric coefficient of the
arc-end species within the labeling complex. (See Figure 2.5.)
Now I am going to do something strange. I am going to calculate a number
for a loop by going around the loop, alternately multiplying and dividing the stoi-
chiometric coefficients I encounter, until I arrive back at my starting point. For
the loop in Figure 2.5, I will start at species A. Traversing the loop in a clockwise
direction, I alternately multiply and divide the stoichiometric coefficients to make
the calculation
1 x (1/2) x 1 x (1/1) x 2 x (1/1) = 1.
I say that a loop is an s-loop if, as in the example, the result of this strange
calculation is one. Thus, the loop shown in Figure 2.5 is an s-loop. (The "s" is
supposed to remind you of stoichiometric coefficients.) To determine if a loop is an
s-loop, the alternate multiplication and division of stoichiometric coefficients can
begin anywhere in the loop. The direction of circumnavigation is immaterial.
Having introduced o-loops, c-Ioops and s-loops, I am almost in a position to
state another theorem. I need just one more piece of terminology: A loop arc in an
SCL Graph is an arc that is contained in a loop.
 53

THEOREM 2.2. Consider a reaction network such that in the SCL Graph

(i) each loop is an a-loop, a c-loop or an s-loop,

(ii) if both arcs of a c-pair are loop arcs, then no loop contains only one of the arcs,

(iii) if a loop is neither a c-loop nor an s-loop (in which case it is an a-loop), then
no linkage class in that loop is adjacent to more than three loop arcs.

If the network is governed by mass action kinetics, then the corresponding

isothermal homogeneous CFSTR equations cannot admit multiple positive rest
points, no matter what the feed composition might be and no matter what (positive)
values the residence time and rate constants take.

REMARK. If there are no loops, all three conditions are satisfied trivially, so
Theorem 2.2 subsumes Theorem 2.1. (Theorem 2.1 was just supposed to get you in
the mood for Theorem 2.2.) Note that the "or" in (i) is not exclusive. That is, (i)
requires only that each loop belong to at least one of the three categories described.
Note also that (iii) is not a restriction on all o-loops. Rather, it is a condition that
must be satisfied only by an a-loop which is neither a c-Ioop nor an s-loop.

We have studied only simple SCL Graphs, so it may be a little difficult to

see how conditions (ii) and (iii) can fail. Let me consider two examples that are
somewhat more complex.
Network (2.13) does have the capacity for multiple positive rest points, so its
SCL Graph (shown in Figure 2.6) must fail to comply with one of the three condi-
tions stated in Theorem 2.2. It is easy to see that every loop in Figure 2.6 is

A+C<=!G
(2.13)

C+E<=!D

an s-loop, whereupon (i) is satisfied and (iii) becomes vacuous. (This will be the
situation for all networks such as (2.13) in which every stoichiometric coefficient is
one.) To see that (ii) is not satisfied, note that the both arcs of the c-pair labeled by
complex C + D are loop arcs. Note also that the loop L1 - A - L2 - C - L3 - B - L1
contains only one of the arcs.
 54

~L
~ A+B
B l~A
B( ,-loop ) A,",
C+D~~ rL;1'
7
__
~ C A+C~
D s-loop) G
~.
C+E E

Figure 2.6. The SCL Graph for Network 2.13

Network (2.14) also has the capacity for multiple positive rest points. Its SCL
Graph is shown in Figure 2.7. Each loop contains just one c-pair, so both loops are
o-loops, whereupon condition (i) is satisfied. It is easy to see that condition (ii) is
satisfied as well. Thus, it must be condition (iii) that fails. In fact, the topmost
loop is neither a c-loop nor an s-loop, and it contains a linkage class, L 2 , that is
adjacent to four loop arcs.

A+B~E

(2.14) C ---7 2A ---7 F +- D ~ B

 55

A
/@~B o-loop

~@~
y-L2~
F
F
C o-loop D

~~--o!.
G

Figure 2.7. The SCL Graph for Network 2.14

Let me give another example. Consider network (2.15). There are 13 species
and 18 reactions, so the corresponding CFSTR equations are even worse than those
shown in (2.6). However horrible these equations might be, Theorem 2.2 tells us

A+B+=!C
B + D +=! E +=! 2A
B+F+=!G
(2.15)
H+I~ J ~2B

H+L ~ K~ 2F
J+K+=!M

very quickly that they cannot admit multiple rest points. The SCL graph for the
network is drawn in Figure 2.8. There are four loops:

A - L1 - B - L2 - A, B - L3 - F - L5 - H - L4 - B,
H - L5 - K - L6 - J - L4 - H, B - L3 - F - L5 - K - L6 - J - L4 - B.

The first is an a-loop, and the others are s-loops, so condition (i) in Theorem
2.2 is satisfied. It is easy to confirm that conditions (ii) and (iii) are satisfied as
well. (In checking condition (iii), it might be noted that the topmost loop is neither
an s-loop nor a c-loop and that the loop contains a species adjacent to four loop
arcs. Condition (iii) requires only that no linkage class be adjacent to four loop
arcs. )
 56

A
c
c

E
G~A+B
B+FB
B+F
L3 I
.-loop
F H
a-loop J
L
.K~M
Figure 2.8. The SCL Graph for Network 2.15

Finally, I want to return to the very simple (but somewhat mysterious) examples
with which we began. Remember that, among the very similar networks (2.1), (2.3)
and (2.4), only (2.1) has the capacity for multiple positive rest points. The SCL
Graphs for the networks are shown in Figures 2.9, 2.10 and 2.11. In each graph
there is only one loop, and from this it easy to see that conditions (ii) and (iii) of
Theorem 2.2 are satisfied for all three networks. It remains to determine for which
of the networks condition (i) is satisfied. The loop in Figure 2.10 is an s-loop, and
the loop in Figure 2.11 is an o-loop. Theorem 2.2 tells us, therefore, that neither
network (2.3) nor (2.4) has the capacity for multiple positive rest points. In Figure
2.9, however, the loop is not an s-loop, an o-loop, or a c-loop, in which case Theorem
2.2 leaves the capacity of network (2.1) for multiple positive rest points undecided.
In fact, there are positive results in Paul Schlosser's thesis [11] which, unlike
Theorems 2.1 and 2.2, indicate when a network doe<. have the capacity for multiple
rest points in a homogeneous CFSTR context. (We expect to make these results
available in an article for Chemical Engineering Science [13].) I won't try to describe
them here, except to say that they are not as wide-ranging as the results I've already
stated.
 57

A
~ B
1 B+C
2A 2

C c B+C
F

Figure 2.9. The SCL Graph for Network 2.1

A
~ B
A 1 s-loop 1 B+C

c c B+C
F

Figure 2.10. The SCL Graph for Network 2.3

 58

~
~A O-l~OP B

~D C
B+C

Figure 2.11. The SCL Graph for Network 2.4

I do, however, want to describe one interesting consequence of the positive

results. Consider the family of networks (parameterized by the integer n) shown
in (2.16). Note that networks (2.1) and (2.3) are special members of the family,

Ao + Al -4 EI
Al + A2 -4 E2
(2.16)

An-I + An -4 En
An -42Ao

corresponding to n = 2 and n = 3. For each n, the induced SCL Graph has just one
loop, just as in Figures 2.9 and 2.11. In each case, the loop is neither a c-Ioop nor
an s-loop. The number of c-pairs in the loop is n, so the loop is an a -loop precisely
when n is odd. Thus, for odd n, Theorem 2.2 precludes the possibility of multiple
positive rest points. On the other hand, for each even n - in particular for network
(2.1) - the positive results indicate that there are multiple positive rest points for
at least some values of the residence time, rate constants and feed concentrations.
In other words, networks in the family (2.16) have the capacity for multiple positive
rest points if and only if n is even.
I want to say a few words about what is seen in nature. There are certainly
experiments that indicate the possibility of multiple rest points in isothermal ho-
mogeneous CFSTRs, but those experiments are fairly recent. In fact, it is almost
doctrine among chemists and chemical engineers that there should be only one rest
 59

point, and it is common to hear reference to "the" steady state, as if uniqueness of

rest points were somehow a consequence of natural law.
This view of things is understandable. It emerged from an acquired experience
that is, I think, explained in part by Theorems 2.1 and 2.2. It is probably the case
that reaction networks selected in some indiscriminate way satisfy the hypotheses
of the theorems far more often than not. Among t.hose net.works that fail to comply
with the hypotheses, there will be some that nevertheless have no capacity for
multiple rest points. And t.he remaining networks will admit multiple rest points
only for certain (perhaps rare) combinations of residence time, rate constants and
feed composition. Even then, the basin of attraction for one of the rest points might
be very large, in which case the other stable rest points would have low probabilities
of det.ection.
Remember that the discussion in this section was tailored very specifically to
homogeneous CFSTRs. I'll turn next to CFSTRs involving heterogeneous catalysis.
For them, the situation is very different..

3. Mechanism discrimination in heterogeneous catalysis. I want to de-

scribe a reactor which, in its outward appearance, looks very much like the CFSTR
studied in the last section. There, the tacit understanding was that reactions take
place homogeneously throughout the fluid occupying the vessel. Here things will be
different. Reactions will not take place in the fluid phase at all but, rather, on a
solid catalytic surface in contact with the fluid phase. This interaction of solid and
fluid phases is the defining characteristic of heterogeneous catalysis.
So that I can introduce the subject, consider once again a vessel to which re-
actants are continuously supplied (in a feed stream) and from which reactants and
product are continuously withdrawn (in an efRuent stream). In this case, think of
the two streams as being gaseous rather than liquid. I'll suppose that, apart from
an inert carrier such as argon, the feed stream contains only two species, A and B,
and that the effiuent stream contains the carrier gas, the reactants A and B, and
also a product C.
I'll also suppose that all steady state experiments suggest the occurrence of the
reaction

(3.1 ) A+ B -> C.

By this I mean that, in all steady state experiments, the loss of a molecule of A
between the feed and effiuent streams is accompanied by the loss of a molecule of B
and by the gain of a molecule of C. It would appear, then, that the chemical events
are as simple as can be: In the gaseous mixture within the chamber, a molecule of
A combines directly with a molecule of B to form a molecule of C.
But things are not always what they seem. Some of the most important chemical
reactions are, in a sense, only apparent. They do not proceed directly, at least not
at appreciable rates. Instead, they proceed by an indirect route mediated by a
catalyst. The cat.alyst is often a metal such as platinum, palladium or rhodium.
 60

I have only told you what is coming out of the vessel and what is going in. I
haven't said very much about what is actually in the vessel, and now I want to
suppose that the apparent reaction (3.1) is being promoted by a solid catalyst.
To keep the picture simple, imagine that the bottom of the vessel is made of a
metallic catalyst and that a gaseous mixture of A, Band C sits above this catalytic
surface. There is no appreciable reaction in the gas phase. Rather, C is produced
from A and B through a mechanism of the following kind: One or both of the
reactants bind from the gas phase onto active catalyst sites, surface-bound species
react with each other (or with species in the adjacent gaseous mixture) to produce
C, and C then desorbs from the catalyst to enter the gas phase. All the whil~, A
and B are being introduced into the gas phase via the feed stream, and gaseous
mixture is being withdrawn continuously in the effluent stream.
For the hypothetical reactor under consideration, I am going to suppose that, to
good approximation, both the gas phase and the catalyst surface remain spatially
homogeneous at all times. That is, I'll regard mixing in the gas phase to be so
thorough that the gas remains uniform in composition (all the way down to the
catalyst surface), and I'll suppose that the distribution of the various adsorbates on
the catalyst surface remains uniform as well.
I have only painted a rough picture of the mechanism whereby A and B come
together to form C. The details are important, and chemists work hard to try to
determine what they are. There are very fine details, such as the way in which the
catalyst actually makes and breaks chemical bonds, but those are not the details I
mean.
When a chemist speaks of a mechanism for the reaction A + n -+ C, he or
she usually means a scenario of the kind depicted in (3.2), (3.3) or (3.4). Here
S denotes a vacant active site on the catalyst surface. (Think of sites as atom-
sized protuberances.) In (3.2), for example, the first two lines represent (reversible)

A+S~ AS
(3.2) B+S~ BS
AS + BS -+ C + 2S

A+S~ AS
(3.3) B+2S~BS2

AS + B S2 -+ C + 3S

A+S~ AS
(3.4) B+S~BS

A+BS -+ C+S

binding of A and B from the gas phase onto single vacant sites. The third line
represents a surface reaction whereby the product C is formed and immediately
enters the gas phase, leaving behind two vacant sites. In (3.3), B binds from the
 61

gas phase onto a pair of vacant sites rather than onto one. In both (3.2) and (3.3),
the product C results from a reaction between two surface-bound species. In (3.4),
on the other hand, C results from a gas phase attack of A on surface-bound B. (In
this case, the adsorption of A on the surface does nothing but block active sites
that might otherwise be occupied by B.)
There are many different mechanistic scenarios that can be written. Which
(if any) fit the experimental facts will depend upon the catalyst and upon the
reactants under study. Thus, for example, one might want to determine, in light
of experimental evidence, a mechanism for the formation of carbon dioxide from
carbon monoxide and oxygen in the presence of platinum.
Elementary steps of the kind shown in (3.2) - (3.3) are molecular events, and
so they cannot be observed directly (at least not today). On the other hand,
different mechanisms have different consequences for reactor behavior in the large,
and so one can try to extract mechanistic information from whatever coarse-grained
measurements are practical.
Mediating between mechanistic conjecture and experimental observations are
the differential equations peculiar to each mechanism. The equations shown in
(3.5), for example, are those we might write for mechanism (3.2). In (3.5) c~ and

C~ = (1/8)(c~ - CA) - kjCACS + k 2 cAS

c~ = (1/8)(c~ - CB) - k 3 cBCS + k 4 cBS

(3.5)
Cc = -(1/8)cc + k 5 cASCBS
c~ = -kjCACS + k 2 cAS - k 3 cBCS + k 4 cBS + 2k5 cASCBS
c~s = kjCACS - k 2 cAS - k 5 cASCBS
c~s = k 3 cBCS - k 4 cBS - k 5 cASCBS

c~ are the molar concentrations of A and B in the feed, C A, C B, Cc are the molar
concentrations of A, Band C in the vessel (and in the effluent stream), CAS, cBS
and Cs are the concentrations of AS, of BS and of vacant sites on the catalyst
surface, k j through k5 are rate constants for the five reactions in (3.2), and 8 is the
residence time for the reactor. Once again, the residence time is the volume of the
mixture filling the vessel divided by the flow rate of the feed and effluent streams.
(I am assuming here that the two streams are made up mostly of inert carrier gas
so that their volumetric flow rates are identical.)
The dynamical equations that correspond to mechanisms (3.3) and (3.4) are
similar in form but differ in detail. The rate of the step in (3.3) whereby B binds
from the gas to the catalyst would, for example, be k3CBC~ rather than k 3 cBCS. The
rate of the step that produces C in (3.4) would be k 5CACBS rather than k 5 cASCBS.
It is the difference in the dynamical equations for the various candidate mech-
anisms that provides a basis for favoring one over another. If gross experimental
observations are incompatible with the equations induced by a particular candidate
mechanism, then that mechanism can only be regarded as suspect.
But what "gross experimental observations" are practical?
 62

Of the six dependent variables appearing in (3.5) (and in the analogous equa-
tions for the other mechanisms), only the gas phase concentrations CA,CB and Cc
can be accessed readily. Concentrations of species adsorbed on the catalyst surface
are much more difficult to determine with any degree of precision. Moreover, static
measurements are much easier to make than transient ones (especially in fast pro-
cesses), and so distinctions that can be drawn on the basis of steady state data are
to be preferred.
Of the parameters appearing in the system (3.5) (and in its analogs for the
other mechanisms) only the residence time and the feed concentrations c~ and c~
are accessible directly. (We cannot expect to know rate constants for reactions
whose very existence is in question.) It is apparent, then, that distinctions between
candidate mechanisms can be made only on the basis questions of the following kind:
For which mechanisms are there rate constant values such that the corresponding
differential equations are consistent with whatever concentration measurements are
available?
My purpose here is to indicate how reaction network theory provides means to
assess the viability of candidate mechanisms, even on the basis surprisingly lim-
ited experimental information. In particular, I am going to discuss ways in which
mechanism discrimination can be based on observations of multiple rest points.
Let me begin with a statement that might seem surprising in view of comments
made at the close of the last section: Multiple rest points (corresponding to different
initial conditions) are very commonly observed in isothermal CFSTRs involving
heterogeneous catalysis-that is, in reactors of the kind we are considering now.
At the end of Section 2, I said that multiple rest points are observed only
rarely in isothermal homogeneous CFSTRs, so I should try to explain why things
are different for catalytic CFSTRs. My claim there was that, for a homogeneous
CFSTR in which the chemistry is randomly selected, multiple rest points would
most likely be precluded by Theorems 2.1 and 2.2. For two reasons, this assertion
is not entirely pertinent to catalytic CFSTRs: First, arguments underlying the two
theorems are based on the supposition that every species in the reactor is also in
the effluent stream. This is not true for catalytic CFSTRs because only the gas
phase species leave the reactor. Second, reaction networks that arise in catalytic
CFSTRs are far from random. Rather, they are heavily skewed toward mechanisms
of the kind displayed in (3.2) - (3.4).
The theory in Section 2 was tailored very specifically to homogeneous CFSTRs.
Now I want to tell you about another theory, called deficiency one theory [6,7],
that was not developed especially for catalytic CFSTRs but which turns out to be
extremely well suited to them. (On the other hand, deficiency one theory is not so
good for homogeneous CFSTRs. The difference in effectiveness of the theory for
the two kinds of reactors is explained in [9]. )
I will not try to indicate, as I did for the SCL Graph theory, how to work with
deficiency one theory. Here I can only hope to tell you something about what the
theory can do, especially in connection with mechanism discrimination problems.
Although deficiency one theory is not as flashy as the SCL Graph theory, it gives
 63

much more detailed information. In rough terms, reaction networks are classified
according to an easily calculated non-negative integer index called the deficiency of
the network. If the deficiency is zero, then the corresponding mass action equations
can admit only very dull, stable behavior. In particular, multiple positive rest points
are impossible. (This result has its roots in work of Horn and Jackson [8).) If the
deficiency turns out to be one, which is the case for catalytic CFSTRs more often
than not, then the theory provides, among other things, a systematic procedure
for determining ~ either affirmatively or negatively ~ whether multiple positive rest
points are possible. When multiple rest points are admitted, deficiency one theory
gives detailed information about relationships that rest points bear to each other.
Now I am going to suppose that two rest points (corresponding to different initial
conditions) have been observed in the reactor under consideration, and I want to ask
how these observations can help in mechanism discrimination. In particular, I will
be interested in the extent to which the experimental facts~taken with deficiency
one theory~can shed light on the viability of mechanisms (3.2) - (3.4).
In order to make the "experimental facts" concrete, I'll assume that the resi-
dence time, () , is 1.0 and that the concentrations of A and B in the feed are c~ = 1.0
and c~ = 0.8. Furthermore, I am going to suppose that the composition of efHuent
gas has been determined in each of the steady states and that the results are those
shown in Table 1. Moreover, I '11 suppose that a sensor indicates the presence of
more A on the catalytic surface in one steady state than in the other. This too is
shown in the table.#

******************************************************************
Table 1

Gas Phase Data

Steady State No. 1 Steady State No. 2

cA = 0.6 cA* = 0.3

Cs = 0.4 cs* = 0.1
Co = 0.4 co' = 0.7

Surface Data

More A on surface in Steady State No.1 than in Steady State No.2.

******************************************************************
# The table is hypothetical, but it captures the essential features of real experiments. Mea-
surements of the kind described were made, for example, by M. - H. Yue [15] in studies of the
formation of ethane from ethylene and hydrogen on rhodium. Electrochemical techniques devel-
oped by my colleague, Howard Saltsburg, permit the assessment of relative amounts of hydrogen
on the rhodium surface in the two steady states.
 64

We might hope that distinctions between (3.2) - (3.4) could be drawn solely
on the basis of the capacity of the corresponding differential equations to admit
multiple rest points.
Consider, for example, mechanism (3.2). Adorned with values specified for the
residence time and feed concentrations, the dynamical equations in (3.5) yield the
steady state equations shown in (3.6). In this case, we would ask if there are rate

0= (1/1.0)(1.0 - CA) - kjCACS + k 2 cAS

o= (1/1.0)(0.8 - CB) - k 3 cBCS + 1.~4CBS
0= -(1/1.0)cc + k 5 cASCBS
(3.6)
0= -k j CACS + k 2 cAS - k 3 cBCS + k 4 cBS + 2k s CASCBS
0= kjCACS - k 2 cAS - k 5 cASCBS

o= k 3 cBCS - k 4 cBS - k 5 cASCBS

constant values such that (3.6) admits two different solutions consistent with the
conservation of catalyst sites. By this I mean that the total concentration of cat-
alyst sites, occupied or not, should be the same in both steady states. In more
precise terms, the two steady states should satisfy the condition

(3.7) Cs + CAS + CBS = CS* + cA*s + CBS'

If no such rate constants exist, then mechanism (3.2) can only be regarded as
suspect.
In a similar way, we can study the steady state equations for mechanisms (3.3)
and (3.4) (along with the appropriate site conservation condition) to determine if
they have the capacity to admit two positive rest points.
Deficiency one theory, however, indicates that mechanisms (3.2) - (3.4) all have
the capacity for multiple rest points in an isothermal CFSTR context. Thus, the
simple qualitative capacity for multiple rest points cannot provide a basis for pre-
ferring one mechanism to another. (In fact, deficiency one theory indicates that
virtually all of the classical catalytic mechanisms have this same capacity. Remem-
ber what I said before: Observations of multiple rest points in isothermal catalytic
CFSTRs are not at all unusual.)
But to say that a mechanism has the capacity to generate two steady states
is not to say that it has the capacity to generate any two steady states. There is
a sharper question we can ask, one related to the seemingly meager quantitative
information that Table 1 carries: For which of the mechanisms (3.2) - (3.4) are
there rate constant values such that the corresponding steady state equations admit
two solutions consistent with site conservation and with the entries in Table I? It
will turn out that this question is surprisingly incisive.
And deficiency one theory gives the answer:
In the case of mechanism (3.2), for example, the theory indicates that there
is no set of (positive) rate constant values for which (3.6) admits two solutions
 65

that are consistent with site conservation and with the information in Table 1. In
particular, no matter how rate constants are assigned, there are no values of the
(unmeasured) surface concentrations CAS, CBS and Cs in the two steady states which
are simultaneously compatible with site conservation, with Table 1, and with (3.6).
The results for mechanism (3.4) are even more striking. In this case, the corre-
sponding steady state equations cannot even admit two solutions that are consistent
with site conservation and with the gas phase data in Table 1. In other words, even
if the surface information in Table 1 were unavailable, mechanism (3.4) would still
be suspect.
The situation for mechanism (3.3) is very different. Deficiency one theory in-
dicates that there are values for the (unknown) rate constants and two surface
compositions, (c'As,cS S2 ,C:s) and (c'A*S,cS*S2'CS*)' such that

(i) c'As > c'A*s

(ii) Site conservation is respected. That is,

(iii) With (c'A, cs, cc,.) and (c'A*,cs*,c'(::*) as in Table 1, both

are solutions of the steady state equations that correspond mechanism (3.3).

Moreover, the theory is constructive. That is, it actually gives sample values of
the rate constants and of the surface concentrations in the two steady states such
that (i) - (iii) are satisfied simultaneously.
The example is, I think, surprising. At first glance, Table 1 does not seem to
carry much information, bllt it turns out to be enough to discriminate between
the very similar mechanisms (3.2) - (3.4). If there is a message here, it is that
even fragmentary macroscopic data may carry powerful dues about the underlying
microscopic machinery.
The story I told was a simple one, but it is a reasonable reflection of the situation
actually encountered by M.- H. Yue in a study of ethylene hydrogenation on rhodium
[15]. In fact, deficiency one theory proved to be even more incisive there than in the
example: Although there is an extensive (and contentious) literature on ethylene
hydrogenation, none of the standard mechanistic proposals could account for the
pairs of steady states Dr. Yue observed (even when allowances were made for gas
phase mass transfer resistance and for Arnall differences between the volumetric flow
rates in the feed and effluent streams). Consistency with the data could be reali7.ed
only after the construction of speculative models involving formation of ethylidine
islands on the catalyst surface.

4. Traveling composition waves on isothermal catalyst surfaces. The

picture of heterogeneous catalysis I painted in the last section was a very simple
 66

one, but it reflects the level of description at which modeling often takes place. Now
I will consider phenomena that are more complicated, involving coupled reaction
and diffusion on the catalyst surface. In particular, I want to discuss work with
David Terman on possibility of traveling composition waves.
We are motivated in part by experiments of Cox, Ertl and Imbihl [4]. During the
oxidation of carbon monoxide on platinum, they observed crystallographic changes
in the catalyst which advanced along the surface i~ a wave-like fashion. These
transformations are thought to be triggered by the presence of carbon monoxide,
and so the observed waves probably indicate wave-like behavior in the surface con-
centrations of carbon monoxide and oxygen as well. On the other hand, changes
in the crystal structure of platinum are almost certainly accompanied by a shift
in catalytic activity. The picture that emerges is a very complex one, involving
an interplay of adsorption from the gas phase, reactions on the catalyst surface,
diffusion along the surface, and transformations of the catalyst itself. Cox, Ertl and
Imbihl constructed a reaction-diffusion model that incorporated these effects, and,
in numerical simulations, wave-like behavior was indicated.
But there remain questions of cause and effect. Although changes in the plat-
inum crystal structure appear to be a real component of carbon monoxide oxidation,
this is not to say that crystallographic shifts are intrinsic to catalysis in general.
Should we, in other systems, regard the presence of composition waves as a signal
that some kind of transfiguration is at work in the catalyst? Or can wave-like behav-
ior emerge as a natural consequence of "normal" catalysis, divorced from ancillary
effects?
Terman and I posed for ourselves these questions: Can traveling compo8ition
waves derive solely from catalytic mechanisms as simple as those shown in (3.2) -
(3.4) (taken only with diffusion along the catalyst surface)? If 80, which mechanisms
are consistent with the existence of composition waves? In particular, is the capacity
for traveling wave8 in some sen8e a generic property, shared by a wide spectrum of
mechanisms?
Our interest is in fundamental occurrences on the catalyst surface itself, un-
complicated by dynamics in the gas phase. With this in mind, we depart from the
CFSTR configuration studied in the last section and consider instead a catalytic
surface immersed in a spatially homogeneous gaseous environment of time-invariant
composition. This is very pristine catalysis.
In this context, we consider the large family of mechanisms shown in (4.1). Here

A+ rnS +:!: kX
(4.1) B + nS +:!: jY
pX + qY --+ (p(rnlk) + q(nfj»S + TC

A, B and C are gas-phase species. The first line represents binding of A from the
gas phase to m vacant catalyst sites and the subsequent formation of k copies of the
surface species X. (Each X occupies rnlk sites.) The situation in the second line is
similar: A molecule of B binds to n sites, and j copies of the surface species Yare
 67

formed. In the last line, p copies of X react with q copies of Y to form r molecules
of the product C, which enter the gas phase, leaving behind p(m/k) + q(nfj) sites.
Note that the mechanisms (3.2) - (3.3) studied in the last section are typical
members of the family (4.1). So are (4.2) and (4.3), which are mechanisms one
might consider as candidates for the oxidation of carbon monoxide. In (4.2) and
(4.3) the gaseous reactants A and B are carbon monoxide and oxygen, and the
product C is carbon dioxide.

CO+S~ CO·S
(4.2) O2 +2S~ 20· S
CO . S +0 .S --+ CO 2 + 2S

CO+S~ CO·S
(4.3) O2 + S ~ O2 , S
2CO . S + O2 . S --+ 2C02 + 3S

The ( constant) concentrations of A and B in the gas phase will be denoted by c A

and c'B, and the five rate constants for the reactions in (4.1) will be denoted (from
top to bottom) by O',(3,/"f, and~. The concentrations of vacant sites and of the
surface species X and Y are the ones we want to study. Were these concentrations
to remain spatially uniform, they would be governed by the system of ordinary
differential equations shown in shown (4.4).

(4.4a) c'x = kO'cA(cS)m - k(3(cxl- p~(cx)P(cy)q

(4.4b) c~ = hc'B(cs t - jf(cy)i - q~(cx )P(cy)q
(4.4c) c~ = -mO'cA(CS)m +m(3(cx)k
- n/,c'B(cst + nf(cy)i
+ (p(m/k) + q(n/j))~(cx )P(cy)q
Note that
(m/k)c'x + (n/j)c~ + c~ = o.
Along solutions of (4.4), therefore, the value of

(m/k)cx + (n/j)c)" + Cs
is invariant. In fact, this sum is just the concentration of catalyst sites, occupied
and unoccupied. (Recall that X and Y occupy (m/k) and (nm sites, respectively.)
With this in mind, we restrict our attention to solutions of (4.4) consistent with the
equation

(4.5) (m/k)cx + (n/j)cy + Cs = K,

 68

where the (positive) number J{ is to be interpreted as the site concentration for the
catalyst surface under study.
In connection with our study of traveling waves, we will want to know which
members of the family (4.1) have the capacity to give multiple positive rest points.
That is, we will want to know for which positive values of the stoichiometric coeffi-
cients m, k, n,j, p and q are there positive values of the parameters ac A, (3, B,e, ,C e,
and J{ such that (4.4) has two or more positive rest points compatible with (4.5).
(Remember that the setting here is different from the CFSTR setting studied in
Section 3.)
The answer is again given by deficiency one theory [14]:

THEOREM 4.1. If m # n, then (regardless of the positive values that the stoi-
chiometric coefficients k, j,p and q take) there are positive values for the parameters
,C
ac A,(3, B,e, e,
and J{ such that (4.4) admits two or more positive rest points com-
patible with (4.5). If m = n, then parameter values of this kind exist if and only if

m =f- p(mlk) + q(n/j)

and either q > j or p > k (or both).

For some special cases of (4.1), all with m = k and n = j, the possibility
of multiple rest points was already studied by Bykov, Chumakov, Elokhin and
Yablonskii [2J.
Theorem 4.1 indicates that, for the special physical situation under consider-
ation here, the capacity for multiple rest points is robust within the mechanistic
family (4.1). What we want to argue is that, within the same mechanistic family,
the capacity for stable traveling waves along the catalyst surface is about as robust
as the capacity for multiple rest points.
In preparation for formulation of reaction-diffusion equations, note that, from
(4.5), Cs can be written in terms of Cx and Cy, and the result can be inserted in
(4.4a) and (4.4b). Then (4.4a) and (4.4b) can be viewed as a self-contained pair
of ordinary differential equations that govern Cx and Cy when those concentrations
remain spatially uniform over the catalyst surface.
Our interest will be in distributions of X and Y that are not necessarily uniform.
If we assume-as Cox, Ertl and Imbihl did-that diffusive transport along the catalyst
surface is of the simple Fickian kind# , then the (one-dimensional) reaction-diffusion

# For the purposes of this exploratory study, it is natural to follow Cox, Ertl and Imbihl in
considering simple Fickian diffusion. Indeed, the model is plausible when the catalyst surface is
not crowded with adsorbed species. When the surface is crowded, however, it seems unreasonable
to expect the local surface diffusion of a species to depend only on the local concentration gradient
of that particular species, independent of the availability of nearby vacant sites. In this case,
models of the kind considered in [1) would seem more apt.
 69

analogs of (4.4a) and (4.4b) are the equations shown in (4.6).

(cx)t = Dx(cx)zz + kac'A(J( - (m/k)cx - (nfj)cy)ffi

- k(3(cX)k - p~(cx)P(cy)q
(4.6)
(cY)t = Dy( Cy )xx + j-ycB(J( - (m/k )cx - (nfj)cy t
-jc(cy)1 -q~(cx)P(cy)q

Now I am going to consider some fixed member of the mechanistic family (4.1)
that is guaranteed by Theorem 4.1 to admit multiple positive rest points in (4.4)
- (4.5). I shall suppose that the parameters ac'A, (3, ,cB' c, ~ , and J( take values
such that multiple positive rest points are, in fact, admitted and such that all the
positive rest points are nondegenerate. The claim I want to make is this: These
parameter values can be supplemented with values ofthe diffusivities Dx and Dy in
such a way that the corresponding reaction-diffusion equations (4.6) admit a stable
traveling wave.
This is summarized in a theorem due to David Terman [14):

THEOREM 4.2. Consider any member of tbe mechanistic family (4.1) for wbich
tbere are values of ac A,(3, ,cB, c,~, and J( tbat give multiple positive rest points in
(4.4) - (4.5). If, for sucb a set of parameter values, all positive rest points of (4.4)
- (4.5) are nondegenerate, tben tbere are values oftbe diffusivities Dx and Dy for
wbicb tbe corresponding reaction-diffusion equations (4.6) admit a stable traveling
wave. In fact, it suffices to take Dx = Dy > O.

As I indicated in the Introduction, the proof rests on Conley Index methods

[3). The waves Terman showed to exist are of the kind that travel from a region
at z = -00 to a region at z = +00, where the two regions are characterized by
compositions corresponding to stable rest points of (4.4) - (4.5).
Terman showed even more: For some members of the family (4.1), studies of
Bykov, Chumakov, Elokhin and Yablonskii [2J indicate the existence parameter
values that give rise to more than two stable rest points of (4.4) - (4.5). In such
instances, Terman proved that the stable rest points can be sequenced in such
a way that there is a stable traveling wave connecting consecutive members of
the sequence, and, moreover, such that wave speeds are nonincreasing along the
sequence.
Taken together, Theorems 4.1 and 4.2 suggest that traveling waves are probably
unexceptional (if not commonplace) in "normal" catalysis, the absence of crystal-
lographic transitions notwithstanding: Theorem 4.1 indicates that, for the mech-
anistic family and the physical setting under consideration here, the capacity for
multiple positive rest points is rather robust. For any mechanism that does have
such a capacity, Theorem 4.2 indicates that, for almost all parameter values for
which multiple positive rest points are admitted by the ordinary differential equa-
tions, there are diffusivities for which stable traveling waves are admitted by the
corresponding reaction-diffusion equations.
 70

REFERENCES

[1) BYKOV, V., A. GORBAN, L. KAMENSCHCHIKOV, AND G.S. YABLONSKII, Inhomogeneous sta-
tionary states in the reaction of carbon monoxide on platinum, Kinetics and Catalysis, 24
(1983) pp. 520-524.
[2) BYKOV, V. 1., G. A. CHUMAKOV, V. 1. ELOKHIN, G. S. YABLONSKII, Dynamicpropertiesofa
heterogeneous catalytic reaction with several steady states, Reaction Kinetics and Catalysis
Letters, 4 (1976), pp. 397-403.
[3) CONLEY, CHARLES, Isolated Invariant Sets and the Morse Index, CBMS Regional Conference
Series in Mathematics, No. 38, American Mathematics Society (1978).
[4) Cox, M.P., G. ERTL, AND R. IMBIHL, Spatial self-organization of surface structure during
an oscillating catalytic reaction, Physical Review Letters, 54 (1985), pp. 1725-1728.
[5) FEINBERG, M., Chemical Oscillations, Multiple Equilibria, and Reaction Network Structure,
in Dynamics and Modelling of Reactive Systems, eds. Warren Stewart, W. Harmon Ray,
and Charles Conley, Academic Press, New York, 1980, pp. 59-130.
[6) FEINBERG, M., Chemical reaction network structure and the stability of complex isothermal
reactors: I. The deficiency zero and deficiency one theorems, Chemical Engineering Science,
42 (1987), pp. 2229-2268.
[7) FEINBERG, M." Chemical reaction network structure and the stability of complex isothermal
reactors: II. Multiple steady states for networks of deficiency one, Chemical Engineering
Science, 43 (1988), pp. 1-25.
[8) HORN, F.J.M. AND ROY JACKSON, General mass action kinetics, Archive for Rational Me-
chanics and Analysis, 47 (1972), pp. 81-116.
[9) LEIB, T.M., D. RUMSCHITZKI AND M. FEINBERG, Multiple steady states in complex isother-
mal CFSTRs: I. General Considerations, Chemical Engineering Science, 43 (1988), pp.
321-328.
[10) RUMSCHITZKI, D. AND M. FEINBERG, Multiple steady states in complex isothermal CFSTRs:
II. Homogeneous reactors, Chemical Engineering Science, 43 (1988), pp. 329-337.
[11] SCHLOSSER, P., A Graphical Determination of the Possibility of Multiple Steady States in
Complex Isothermal CFSTRs, PhD Thesis, Department of Chemical Engineering, Universit.y
of Rochester (1988).
[12) SCHLOSSER, P. AND M. FEINBERG, A graphical determination of the possibility of multiple
steady states in complex isothermal CFSTRs, pp. 102-115 in Complex Chemical Reaction
Syst.ems, eds. J. Warnatz and W. Jager, Springer-Verlag, Berlin-Heidelberg-New York, 1988.
[13) SCHLOSSER, P. AND M. FEINBERG, Multiple steady states in very complex isothermal
CFSTRs, in preparation for Chemical Engineering Science.
[14) TERMAN, D. AND M. FEINBERG, Traveling waves on isothermal catalyst surfaces, in prepa-
ration for Archive for Rational Mechanics and Analysis.
[15) YUE, M.-H., Isothermal Multiple Steady States in Ethylene Hydrogenation over Rhodium:
Mechanistic Screening and Experimental Studies, PhD Thesis, Department of Chemical
Engineering, University of Rochester (1989).
 GENERICITY, BIFURCATION AND SYMMETRY

MARTIN GOLUBITSKY*

In these lectures I would like to discuss how the existence of symmetries alters
the type of bifurcation behavior that one expects to observe. In the first lecture
I will concentrate on the structure and dynamics of steady-state bifurcation from
equilibria. It is here that the influence of symmetries on linearized equations will be
discussed and some facts from elementary representation theory introduced. The
second lecture will be devoted to effects of symmetry on period-doubling in maps
with a short description of an application to large arrays of Josephson junctions. In
the final lecture I will describe how certain standard choices of boundary conditions
(particularly Neumann) can be thought of as symmetry constraints and how this
fact alters notions of genericity. It accord with the style that has developed in the
lectures at this workshop, the lectures are of different length.
Much of the background material for these lectures may be found in [GSS). The
descriptions of the more advanced topics will be brief as the results concerning these
topics have or will appear elsewhere.

Lecture 1: Bifurcation From Equilibria

Consider the system of ODE

dx
(1.1) dt = f(X,A) x E Rn , AE R

with an equilibrium at (xo, Ao), that is,

f(XO,AO) =0.

Let A == (dJ)XO,AO be the n x n Jacobian matrix obtained by differentiation with

respect to x. Then (1.1) becomes:

dx
-=A(x-xo)+'"
dt

Recall that if A is hyperbolic (that is, all eigenvalues have nonzero real part), then
all the dynamics of (1.1) are determined by A near Xo.

*Department of Mathematics, University of Houston, Houston, TX 77204-3476

 72

DEFINITION 1.1. (1.1) has a bifurcation point at (xo, Ao) if some eigenvalue of
A lies on the imaginary axis.

Without loss of generality we may assume Ao = O.

Basic Question: What are the typical transitions in the dynamics of (1.1)?

It is well known that generically the typical transitions are controlled by the
transitions from hyperbolicity in the matrix A as A is varied. Indeed, there are two
possibilities; A has a

(a) simple zero eigenvalue Steady-State Bifurcation

(b) a pair of simple, complex-conjugate, purely imaginary eigenvalues. Hopf
bifurcation

Nonlinear theory then implies that in case (a) the dynamics can be reduced (using
center manifolds) to one dimension and the expected transition is a limit-point or
saddle-node bifurcation with a transition from 0 to 2 equilibria. In case (b) the
dynamics can be reduced to two dimensions and one expects a single branch of
periodic solutions to emanate from this bifurcation. See [GH1].
We now consider how both the linear and the nonlinear transitions change when
(1.1) has a nontrivial group of symmetries.

SYMMETRY
Let r c O( n) be a Lie group of orthogonal matrices.
DEFINITION 1.2. (1.1) has symmetry r if

(1.2) f(iX,A) = if(x,A) for all i E r.

There are two immediate consequences of the commutativity relation (1.2):

(a) The equilibrium Xo has symmetry. Define the isotropy subgroup of rat
Xo to be
~xo == b E r : iXo = xo}

In our discussion we will assume that the equilibrium Xo is a fully symmetric, i.e.
~xo = r. Then, without loss of generality, we may assume Xo = O.

(b) The chain rule implies (df)-YX,Ai = i(df)x,A and hence

that is, the matrix A commutes with r. Thus to understand the dynamics of (1.1)
we must first understand the form of matrices that commute with r. This topic
in representation theory has been well studied and we briefly review the relevant
theory.
 73

ELEMENTARY REPRESENTATION THEORY

DEFINITIONS 1.3. Let W be a subspace of V ~ Rn.

(a) We V is r-invariant if ,(W) = W for all, E r

(b) We V if r-irreducible if the only r-invariant subspaces of Ware {OJ and
W.

It is well known that any representation may be decomposed into a direct sum
of irreducible representations; the simplicity of the proof of this decomposition is,
however, not always appreciated.

THEOREM 1.4 (The Decomposition Theorem). There exist r -irreducible sub-

spaces VI, ... , VB such that

Proof. Since r c O(n), the standard inner product IS r-invariant; that IS,

(Iv, ,w) = (v, w) for all , E r.

Now suppose V has proper r-invariant subspace W. Then define

W.L = {v E V : (v, W) = O}
and observe that W.L is r -invariant. Since

the theorem is proved by induction on the dimension of V. D

We begin our discussion of commuting matrices by first considering commuting
matrices for an irreducible representation U.

THEOREM 1.5. The space of matrices commuting with an irreducible represen-

tation is isomorphic to R, C or H (where H denotes the quaternions).

Proof. Observe that the vector space

'D = {matrices on U commuting with r}

is an algebra over R; that is, we can add, multiply and scalar multiply commuting
matrices. In addition, 'D is a division algebra, that is, every nonzero matrix B in 'D
is invertible. To verify this point note that for B E 'D

ker B is r-invariant

and irreducibility implies

ker B = U or ker B = {O}

Hence, either B = 0 or B is invertible and B- 1 E 'D. The classical Wedderburn

Theorem implies that'D is isomorphic either to R, C or H.
 74

DEFINITION 1.6. U is absolutely irreducible if the only matrices commuting

with r are multiples of the identity ('D ~ R) and nonabsolutely irreducible other-
WIse.

Examples. (a) 50(2) acts nonabsolutely irreducibly on R2.

(b) 0(2) acts absolutely irreducibly on R2.

THEOREM 1. 7. Generically, in one-parameter bifurcation, steady-state bifurca-

tion satisfies:

(a) the algebraic multiplicity of the zero eigenvalue equals the geometric mul-
tiplicity, and
(b) r acts absolutely irreducibly on ker A.

Sketch of Proof.

(I) At a bifurcation point, do a center manifold reduction (which can be done

preserving symmetries - Ruelle [RD. Thus, we can assume all eigenvalues of A are
on the imaginary axis.

(II) Suppose A has a zero eigenvalue.

Choose a r-irreducible subspace U C ker A, and define

0 on U
M: V --+ V by {
I on U.l.

Perturb (1.1) to:

dx
dt = f(x, A) + c:Mx == fe(x, A).

For nonzero c:, Ae == (dfe)o,o = A + c:M satisfies:

(a) Geometric multiplicity of eigenvalue zero
= algebraic multiplicity of eigenvalue zero.

(b) r acts irreducibly on ker Ae = U.

Now reduce the bifurcation problem to U (by center manifold).

(III) If dim U = 1, then r acts absolutely irreducibly. So assume dim U 2: 2. We

claim that f(O, A) = 0, that is, x = 0 is a 'trivial' equilibrium.

DEFINITION 1.8. Let L: c r be a subgroup. Define the fixed-point suhspace of

L: to be:
Fix(L:) == {v E V : av = v for all a E L:}

LEMMA 1.9. f: Fix(L:) x R --+ Fix(L:).

Proof. f(V,A) = f(av, A) = af(v,A) for all aEL:.

o
 75

To prove the claim observe that Fix(r) = {OJ since r acts irreducibly and nontriv-
ially. Thus, f(O, A) = O.
Define A>. = (df)o,>. and observe that A>. commutes with r. Hence, A>. is in 'D
and corresponds to a curve d(A) E R, Cor H with d(O) = O.

(IV) Suppose that r acts nonabsolutely irreducibly on U. Then the curve d( A)

is in either C or H. The hyperplane {z E 'D : Re(z) = O} corresponds to matrices
with purely imaginary eigenvalues. Hence the curve d(A) can be perturbed to e(A)
where e(A) crosses Re(z) = 0 at A = 0 with nonzero speed, but NOT THROUGH O.
Since the curve e(A) corresponds to a family of matrices B>., we can perturb (l.1)
to:
dx
dt = f(x, A) + (B>. - A>.)x.

o
The absolute irreducibility noted in Theorem l.9 can be used to transfer the
analytic problem of existence of branches of equilibria to an algebraic one, as the
next theorem shows.

THEOREM 1.10 (Equivariant Branching Lemma). (Vanderbauwhede [V], Ci-

cogna [CD

(a) Let r c O(n) be a Lie group acting absolutely irreducibly on V == Rn.

(b) Assume that (1.1) has a bifurcation at A = 0 and symmetry r.
(c) Let 2: cr be a subgroup such that

(l.3) dim Fix(2:) = l.

then there is a unique branch of equilibria having symmetry 2: if

C'(O) of. 0

where (df)o,>. = c(A)I by (a) and c(O) = 0 by (b).

Proof We know that f : Fix(2:) x R -4 Fix(~). Let Vo be a nonzero vector in

Fix(~),and define 9 : R x R --t R by

g(S,A)VO = f(8Vo,A).
We that g(O, A) = 0 since irreducibility implies that 0 is a 'trivial' solution. Hf'nce
g(8, A) = h(8, A)S by Taylor's Theorem where

h(O, 0) = c(O) = 0 and hs(O, 0) = C'(O) of. o.

Using the Implicity Function Theorem solve

h(s,A)=O for A =1\(8). o

 76

Examples. (a) 0(2) acts on R2 ~ C. Let E = {I, K} ~ Z2 where KZ = z. Then

Fix(E) = R C C has dimension one. Hence, in circularly symmetric bifurcation
problems we expect equilibria with a reflectional symmetry.

(b) The irreducible representations of 50(3) are given by the spherical harmon-
ics of orded denoted by Vi. (Note that dim Vi = 2f!+1.) The Cartan decomposition
of Vi is:

where the action of 50(2) C 50(3) on Vi is given by:

It follows that Fix( 50(2)) = {(x, 0, ... , On has dimension one. Since solutions with
50(2) symmetry have an axis of symmetry, we have proved:

COROLLARY 1.11. In steady-state bifurcations involving spherical symmetry,

generically (in the sense that eigenvalues go through zero with nonzero speed) there
exist a branch of axisymmetric equilibria.

(c) (Nontrivial dynamics in steady-state bifurcation) Let r C 0(3) be the 24

element group generated by:

CT(X,y,Z) = (y,z,x)
c(x,y,z) = (cIX,c2y,c3Z) where Cj = ±1.

Bifurcation with this group action was studied by May and Leonard [ML] in the
context of three competing populations and Busse and Heikes [BH] in the context
of convection in a rotating layer. Later, Guckenheimer & Holmes [GH2] studi('d
this group of symmetries abstractly. They showed that it is possible to have a
structurally stable (in the world of r symmetry), asymptotically stahle, primary
branch of heteroclinic connections. We outline this construction.
Up to conjugacy the isotropy subgroups of rare:

E2 = {cllI, I)} Fix(E 2 ) = {(O,y,z)}

E3 = {(I, CT, 0"2} Fix(E3) = {(x,x,x)}
E4 = {(cl,c2,I)} Fix(E4) = {(O,O,z)}.

Hence, generically, there exist two types of equilibria with isotropy E3 and E 4 ,
respectively.
We determine the dynamics associated with this r-equivariant bifurcation by
describing explicitly the general r-equivariant mapping f. Write f in coordinates
 77

as! = (X, Y, Z). Then

(i) !(uv, >-.) = u!(v, >-.) implies:

Y(x, y, z, >-.) = X(y, z, x, >-.)

Z(x,y,z,>-.) = X(z,x,y,>-.).

((ii) !(EV,>-')=E!(V,>-.) implies

X(x, y, z, >-.) is odd in x and even in y and z.

Thus, we can write X(x,y,z,>-.) = a(x 2,y2,z2,>-.)x. The genericity condition is

a>.(O) # o. We assume:

(HI) a>.(O) > 0

so that the trivial solution losses stability at >-. = O. Rescale >-. so that:

To third order ! has the form:

Hence, computing II Fix(L: 4) x R = 0 yields the equation az 2 )... Thus, if we

assume

(H2) a < 0,

then the L:4 equilibrium A will exist for >-. > 0 and, by exhance of stability, be stable
inside the z-coordinate axis Fix(L: 4 ).
Since (df)A commutes with its isotropy subgroup L: 4 . It follows that (df)A is
diagonal and that, to lowest order, the two eigenvalues outside Fix(L: 4) are:

(in the x-direction)

(in the y-direction)

If we assume:

(H3)
 78

then A will be a sink in the flow-invariant yz-plane Fix(Z:2) and a saddle in the
xz-plane. By considering 11 Fix(Z:2) x R one can show that there are no equilibria
in the yz-plane that lie off the coordinate axes when (H3) is valid.
Thus the unstable manifold leaving A in the xz-plane must either be unbounded
or tend to the equilibrium on the x-axis. Indeed, the saddle-sink connection can be
shown to exist if:
a~O.

Hence the heteroclinic cycle exists.

Finally, we note that this connection is structurally stable, since the coordinate
planes are always flow invariant (being fixed point subspaces) and planar saddle-
sink connections are structurally stable. A calculation shows that this heteroclinic
connection can be asymptotically stable.
Thus, intermittancy is an expected phenomena in symmetric systems. Further
examples of complicated dynamics emanating from a steady-state bifurcation may
be found in Field [F3]. See also [AGH]
We end this lecture by discussing the general form of linear commuting maps
when the representation is not irreducible. This result is useful when computing the
asymptotic stability of nontrivial equilibria, and when considering mode interactions
in multiparameter systems. This material is included mainly for completeness and
may be skipped on a first reading.

DEFINITIONS 1.12. Let r act on a space V.

(a) Let WI and W2 be r-irreducible subspaces of V. Then W l and W 2 are r-

isomorphic if there exists a linear map L : TVI -+ TV2 that commutes with r, that
is, L, = ,L for all'ET.
(b) Let W be a r -irreducible subspace of V. The isotypic component of V corre-
sponding to W is the sum of all r -irreducible subspaces of V are r -isomorphic to
W.

Examples. (a) For each integer £ let 0(2) act on Vi ~ C by:

() . z = eiiO z and K· Z = z.
The actions for £1 and £2 are 0(2)-isomorphic iff £1 = ±£2·

(b) Let the permutation group S3 act on C as symmetries of a equilateral triangle

and on R3 by permuting axes. The second action has a two dimensional S3 Irre-
ducible subspace consisting of points in R3 whose coordinates sum to zero. The
actions of S3 on C and V are isomorphic.

THEOREM 1.13. Let U I , ... , Ut be the distinct r -irreducible representations

appearing in a decomposition of V guaranteed by the Decomposition Theorem.
Then V is the direct sum of isotypic components:

(1.4)
 79

COROLLARY 1.14. Let A : V -4 V be linear and commute with r. Then A can

be block diagonalized by (1.4), that is,

A(VUj) c VUj for j = 1, ... ,t.

Lecture 2: Period-Doubling, Symmetry and Josephson Functions

In this lecture I want to describe how symmetry affects period-doubling bifurca-

tions and apply these ideas to coupled systems of Josephson junctions. The theory
follows closely the discussion of steady-state bifurcations given in the first lecture -
but with an important difference. The period-doubling bifurcation itself introduces
a reflectional symmetry.
Let f: V x R -4 V be a smooth r-equivariant mapping and let fe, Ao) have a
fixed-point at Xo.

DEFINITION 2.1. f has a period-doubling bifurcation at (xo, AO) if -1 IS an

eigenvalue of the Jacobian matrix (df)xo,)\o.

We assume that Xo is r-invariant and hence, without loss of generality, we may

assume that (xo, AO) = (0,0). Our discussion of genericity for steady-state bifur-
cation applies equally well to period-doubling bifurcation. In particular, genericity
implies that the geometric multiplicity of the eigenvalue -1 equals the algebraic
multiplicity and that r acts absolutely irreducibly on the eigenspace V-I corre-
sponding to the eigenvalue -1.
After a center manifold reduction we may assume that V = V-I. Observe that
irreducibility implies that f has a trivial fixed point, i.e. that f(O, A) = 0. Similarly,
absolute irreducibility implies that

(df)o,A = c( A)1

where c(O) = -1. Indeed, genericity implies that c'(O) i 0.

The question we address is: find all branches of period two points of f in the
neighborhood of the period-doubling bifurcation at (0,0). We prove the following
analogue of the Equivariant Branching Lemma.
Define the group
if - 1Er
if - 1 rt. r.

Note that f acts naturally on V.

THEOREM 2.2. Let ~ C f be a subgroup satisfying:

dim (Fix(~)) = 1.
 80

Then there exists a unique branch of period two points for f emanating from the
origin.

A proof of this theorem, based on normal hyperbolicity is given in [ChGj. Here,

however, we present a very simple proof using Liapunov-Schmidt reduction. This
proof was derived independently by Peckham & Kevrikidis, Roberts and Vander-
bauwhede (private communications). The idea of the proof is to convert the problem
of finding period two points of f to one of finding zeroes of a derived mapping F.
The Equivariant Branching Lemma is then used to prove the existence of branches
of zeroes of F. The proof is a discrete analogue of the proof of existence of periodic
solutions given by Liapunov-Schmidt reduction in Hopf bifurcation.
Proof. Observe that finding a point x such that f(J(x)) = x is equivalent to
finding solutions to the system of equations

y = f(x) and x = f(y).

Given this, define F : V x V -> V x V by

F(x,y) = (J(x) - y, f(y) - x).

Then F(x,y) = (0,0) if and only if x and yare period two points of f.
Next use Liapunov-Schmidt reduction to solve F = O. To do this, compute

(dF)o,o = (-I-I -I-I)

and observe that

if == ker (dF)o,o = {(x,-x) E V x V: x E V}

is isomorphic to V. Now use Liapunov-Schmidt reduction to find implicitly a map-

ping g : if -> if whose zeroes near the origin are in one to one correspondence with
the zeroes of F.
Now we consider the equivariance properties of g. Since Liapunov-Schmidt
reduction can be performed in such a way as to preserve symmetries, we need
only consider the equivariance of F. Note that r acts (via the diagonal action) on
V x V and that F commutes with the action of r. In addition F commutes with
the reflection symmetry (x,y) -> (y,x). Since this symmetry acts as -Ion if, it
r
follows that the reduced mapping g commutes with the group acting on if.
Finally, we note that the assumption on Fix(~) is precisely what is needed to
apply the Equivariant Branching Lemma. 0
 81

Arrays of Coupled Oscillators

We will apply this theorem to find period two points emanating from period-
doubling bifurcations in the presence of SN symmetry, where SN is the permutation
group on N letters. To motivate this discussion, we begin by considering arrays of
coupled oscillators.
An array of coupled oscillators is a system of ODE of the form:

(2.1 ) Yj E Rk

lJN =9N(Y1,···,YN).

These oscillators are identical if 91 = ... = 9N =' 9 and identically coupled if

for every permutation on N - 1 letters (1. Observe that systems of identical, iden-
tically coupled, coupled oscillators are precisely those systems (2.1) that have SN
symmetry.
An in-phase solution to (2.1) is one lying in the plane

(2.3) Y1 = ... = YN =' y.

The fact that the plane defined by (2.3) is flow-invariant follows from the fact that
the plane (2.3) is just Fix(SN)' In-phase solutions satisfy the differential equation:

(2.4) if = 9(Y, ... ,y)

An interesting example of a system of identical coupled oscillators is a large array

of Josephson junctions that has been studied by Hadley, Beasley and Wiesenfeld
[HB\Vl, HBW2]. The second order system of ODE for Josephson junctions is

(2.5) (j = 1, ... N)

where

1> j is the difference in phase of the "quasi classical

superconducting" wave functions on the two sides
of the j-th junction

B is the capacitance of each junction

IB is the bias current of the circuit

IL is the load current.

 82

To complete the system, assumptions must be made on how the circuit is loaded.
for example, if the array is capacitative loaded, then

h=LJ j
j=l

while if the array is resistive loaded, then

We make several observations about this system of ODE.

(a) In-phase periodic solutions exist for a large range of the parameters /3, lB.
(b) The Poincare maps for the in-phase periodic solutions can loose stability by
either a fixed-point or a period-doubling bifurcation - but not by a Hopf bifurcation.
Both of these types of bifurcations have been found in numerical computation.
(c) When the in-phase periodic solutions exists, it is asymptotically stable in
the plane (2.3), and hence is unique.

Next we address the question of what types of solutions are expected to emanate
from the bifurcations noted in (b). More detail may be found in [AGE:]'
Fix /3,IB at a point where an in-phase periodic solution yet) exists. Choose a
Poincare section S as follows:

(2.6)

where L is the cross-section to yet) in the plane of in-phase solutions and

Let P : S -+ S be the Poincare map: P(O) = 0 since the in-phase solution is

periodic.
Observe that Sis SN-invariant and that uniqueness of solutions to systems of
ODE forces P to be SN-equivariant. Indeed, we may write

(2.7)

as a direct sum of SN-irreducible subspaces, where

Suppose that the in-phase periodic solution is undergoing a bifurcation; that is,
either of the generalized eigenspaces El or E-l corresponding to the eigenvalues ±1
is nonzero. Invoking genericity, we expect the action of SN on E±l to be absolutely
 83

irreducible. It follows from (2.6, 2.7) that when these eigenspaces are nonzero they
will generally be isomorphic to either L or V. If they happened to be isomorphic
to L, then the bifurcation would produce a new in-phase periodic solution, thus
contradicting (c). Hence E±l ~ V.
We consider first the possibility of a fixed-point bifurcation for the Poincare
map; that is, El ~ V. Field and Richardson [FR] show that generically all fixed
points of P have isotropy with one-dimensional fixed point subspaces. Hence the
Equivariant Branching Lemma (applied to Q( s) = P( s) - s) implies the existence of
all the expected fixed points of P. Up to permutation, the fixed points are classified
as follows. Divide the oscillators into two blocks: one block having k oscillators and
the other having N - k oscillators. The bifurcating fixed points have the first k and
the last N - k coordinates equal. Their symmetry group is:

(2.8)

Unfortunately, [IG] show that if there is a nonzero equivariant quadratic, then

generally all solutions found using the Equivariant Branching Lemma are asymp-
totically unstable. The action of SN on V has such a nonzero equivariant quadratic
mapping.
At a period-doubling bifurcation the local bifurcation results are more interest-
ing. Since -J ~ SN as it acts on V, we use SN and Theorem 2.2 to find period two
solutions. There is another class of periodic solutions obtained in this way. Divide
the oscillators into three blocks, the first two having k elements and the third having
N - 2k elements. The isotropy of such solutions is the group:

Tk generated by Sk x Sk X SN-2k and (x,y,z) -+ -(y,x,z).

Theorem 2.2 implies the existence of period two points having symmetries Ek
and symmetries T k . The interpretation of these properties of these solutions for the
Josephson function model is most interesting. As noted above, the periodic solutions
with isotropy Ek divide the oscillators into two blocks, each block consisting of
in-phase oscillators with period approximately twice that of the original in-phase
solution. The periodic solutions with isotropy Tk divide the junctions into three
blocks, the first two blocks consisting of in-phase oscillation but with a half period
phase shift between the two blocks. The third block consists of junctions with
in-phase oscillation but with a period comparable to the period of the in-phase
periodic solution.
Certain of these solutions can be asymptotically stable (for more details, see
[AGKJ) and have been observed in numerical experiments on the resistive load
Josephson junction model (but not with the capacitive loaded model). [AGK] also
prove that there exists period two solutions to this SN symmetric period-doubling
bifurcation with submaximal isotropy. For these solutions the oscillators divide into
three blocks of unequal size.
 84

Lecture 3: Genericity and Boundary Conditions

The Faraday experiment provides an example where a general qualitative analy-

sis based on period-doubling bifurcations may connect theory with experiment. This
connection highlights the effects that boundary conditions may have on genericity
(see [CGGKS] where the issues raised here are discussed more fully).
In the Faraday experiment a fluid layer is subjected to a vertical oscillation at
frequency wand forcing amplitude A. When A is small the fluid remains essentially
flat and when A is increased the flat surface bifurcates to a standing wave at fre-
quency w/2. What is measured in the experiments of Gollub and coworkers [CG,
GS] is a stroboscopic map S which pictures the surface of the fluid at each period
of the forcing. Since, after the bifurcation, the fluid surface returns to its original
form each second iterate of S, we have a period-doubling bifurcation.
The experiments of Ciliberto and Gollub [CG] focus on fluid layers with circu-
lar cross-section, while the experiments of Gollub and Simonelli [GS] focus on the
square cross-section case. A qualitative analysis of the circular cross-section exper-
iment, along the lines that we describe here for the square cross-section, is given by
Crawford, Knobloch and Riecke [CKR].
The following points are observed in the experiments [GS].

(a) For most values of the forcing frequency w the initial bifurcation from
stability of the flat surface as the amplitude A is increased is by a period-doubling
bifurcation. Spatial modes are detected and described by their wave numbers in
both horizontal directions, such as (3,1), (3,2), (4,0).
(b) For isolated values of w the flat surface loses stability to two modes si-
multaneously.

Given this information we may make several reasonable assumptions concern-

ing the mathematical analysis of any model purporting to describe the Faraday
experiment.

(a) Since The experiment is square symmetric, the loss of stability of the flat
surface to say a (3,1) mode would imply loss of stability to the (1,3) mode as well;
that is, the eigenspace E-l corresponding to the period-doubling -1 eigenvalue is
at least double. Generically, it is precisely double, and hence E-l ~ C.
(b) At the mode interaction point E-l is isomorphic to C 2 •
( c) Assuming that a center manifold reduction is possible, the dynamics of the
(stroboscopic map of the) Faraday experiment near the mode interaction point is
controlled by the dynamics of a D4-equivariant mapping f: C 2 ~ C 2 .

These assumptions, however, lead to a difficulty. The representation of the

symmetry group' of the square, D 4 , on the eigenspace E-l at a generic (non-mode-
interaction point) is either an irreducible two-dimensional representation of D4 or
the sum of two one-dimensional irreducibles.
In the latter case we must then question why a nongeneric situation occurs
in this experiment (since generically eigenspaces are irreducible). In the former
 85

case the representation is irreducible, but a different problem occurs at a point of

mode interaction. Up to isomorphism the two-dimensional irreducible representa-
tion of D4 is unique. Hence, at a codimension two point of mode interaction of two
two-dimensional, isomorphic, irreducible representations, generically we expect the
linearization (of f) to be nilpotent. This nilpotency would imply that there is only
one independent set of eigenfunctions (not two), and hence that the two distinct
modes would, in fact, have to merge together at the co dimension two point (and be
physically indistinguishable).
We are faced with a dilemma: either something nongeneric (the reducibility of
the eigenspace) occurs in models of the Faraday experiment, or something is wrong
with the experimental observation of distinct modes in the square cross-section case.
We present here an alternative explanation based on some subtleties of genericity
and boundary conditions. See [CGGKS].
Fujii, Mimura and Nishiura [FMN] and Armbruster and Dangelmayr [AD] ob-
served that the bifurcation of steady solutions in reaction-diffusion equations on
the line changed from what might have been expected when Neumann boundary
conditions (NBC) were assumed. We abstract part of their reasoning here.
Any solution u to a reaction-diffusion equation on [0,7r] with NBC can be ex-
tended in a solution v to that same equation with periodic boundary conditions
(PBC) on [-7r, 7r] by extending the solution to be even across zero. More precisely,
define:
(3.1) vex, t) = u( -x, t) for all x < o.
Conversely solutions v to the PBC problem that are also even (which, using (3.1)
is a fixed point subspace condition for the symmetry x -+ -x) is a solution to the
NBC model.
What is gained by the extension to PBC is the introduction of 0(2) symmetry
into the problem (translational symmetry of the reaction-diffusion equation modulo
the 27r periodicity of the boundary conditions). The idea for determining genericity
is to look at the generic PBC case (that is, 0(2) symmetric bifurcation) and then
restrict (by fixed-point subspace arguments) to the NBC case.
Similar statements about genericity are valid for Dirichlet boundary conditions
(DB C), although DBC does require an extra reflectional symmetry on the differen-
tial operator to be valid in order to make the extension to PBC. This symmetry is
valid, for example, in the Navier-Stokes equations. Indeed, similar statements hold
for systems and for higher dimensional domains with various mixtures of boundary
conditions.
We now return to the Faraday experiment. In any analytic model of the exper-
iment one must solve for both the surface deformation (x, y) and the fluid velocity
field u( x, y, z). Typically, in models, no-slip or Dirichlet boundary conditions are
valid for u along the lateral boundaries and Neumann boundary conditions are as-
sumed on ( (that is, the fluid surface is assumed to be perpendicular to the side
walls).
 86

As our discussion above indicates these boundary conditions have the effect
of introducing T2 symmetry into the bifurcation problem. The two-torus T2 is
obtained by planar translations modulo the double periodicity of the square. Thus,
the full symmetry group r of the Faraday experiment with square geometry is
generated by D4 and T2.
The consequence of having this enlarged symmetry group is that all of the two-
dimensional eigenspaces noted above are irreducible representations of r, and that,
at mode interaction points, distinct modes have distinct irreducible representations.
Hence, the linearization is diagonal rather than nilpotent and distinct modes need
not merge (thus agreeing with experimental observation).
Period-doubling bifurcations at points of mode interaction in the Faraday ex-
periment is being studied in [CGK] using the group r.
Acknowledgement

The research described in these lectures was supported in part by the Institute
for Mathematics and its Applications, University of Minnesota, and by the following
research grants: NSF/Darpa (DMS-8700897), Texas Advanced Research Program
(ARP-llOO) and NASA-Ames (NAG2-432).
Many of the results discussed in these lectures were the product of unpub-
lished collaborative research. I wish to thank my collaborators on these projects
for permitting me free use of this material: Don Aronson, John David Crawford,
Gabriella Gomes, Edgar Knobloch, Maciej Krupa, and Ian Stewart.
REFERENCES

[AD] D. ARMBRUSTER & G. DANGELMAYR, Coupled stationary bifurcations in non-flux boundary

value problems, Math. Proc. Comb. Phil. Soc. 101 (1987), 167-192.
[AGH] D. ARMBRUSTER, J. GUCKENHEIMER & P. HOLMES, Heteroclinic cycles and modulated trav-
elling waves in systems with 0(2) symmetry, Physica 29 D (1988) 257-282.
[AGK] D.G. ARONSON, M. GOLUBITSKY & M. KRUPA, Coupled arrays of Josephson junctions and
bifurcation of maps with S N symmetry, Nonlinearity. Submitted.
[BH] F.H. BUSSE & K.E. HEIKES, Convection in a rotating layer: a simple case of turbulence,
Science 208 (1980) 173-175.
[ChG] P. CHOSSAT & M. GOLUBITSKY, Symmetry-increasing bifurcation of chaotic at tractors, Phys-
ica 32 D (1988) 423-436.
[C] G. CICOGNA, Symmetry breakdown from bifurcations, Lettere al Nuovo Cimento 31 (1981)
600-602.
[CG] S. CILIBERTO & J. GOLLUB, Chaotic mode competition in parametrically forced surface
waves, J. Fluid Mech. 158 (1985), 381-398.
[CGGKS] J.D. CRAWFORD, M. GOLUBITSKY, M.G.M. GOMES, E. KNOBLOCH & LN. STEWART, Bound-
ary conditions as symmetry constraints, Preprint, University of Warwick (1989).
[CKR] J.D. CRAWFORD, E. KNOBLOCH & H. RIECKE, Competing parametric instabilities with cir-
cular symmetry, Phys. Lett. A 135 (1989) 20-24.
[F1] M. FIELD, Equivariant dynamical systems, Trans. A.M.S. 259, No.1 (1980) 185-205.
[F2] M. FIELD, Equivariant bifurcation theory and symmetry breaking, Dyn. Diff. Eqn. 1, No.4
(1989), 369-421.
[FR1] M. FIELD &R.W. RICHARDSON, Symmetry breaking and the maximal isotropy subgroup
conjecture for reflection groups, Arch. Rational Mech. Anal. 105 (1989) 61-94.
 87

[FR2) M. FIELD & R. W. RICHARDSON, New examples of symmetry breaking and the distribution
of symmetry breaking isotropy types, In preparation.
[FMN) H. FUJII, M. MIMURA & Y. NISHIURA, A picture of the global bifurcation diagram in eco-
logically interacting and diffusing systems, Physica 50 (1982) 1-42.
[GH1] J. GUCKENHEIMER & P. HOLMES, Nonlinear Oscillations, Dynamical Systems, and bifurca-
tion of Vector Fields" App!. Math. Sci. 42, Springer-Verlag, New York, 1983.
[GH2) J. GUCKENHEIMER & P. HOLMES, Structurally stable heteroclinic cycles, Math. Proc. Comb.
Phil. Soc. 103, part 1 (1988), 189-192.
[GSS] M. GOLUBITSKY, LN. STEWART & O.G. SCHAEFFER, Singularities and Groups in Bifurcation
Theory: Vol. II, Applied Math. Sci. 69, Springer-Verlag, New York, 1988.
[I-IBW1) P. HADLEY, M.R. BEASLEY & K. WIESENFELD, Phase locking of Josephson-junction series
arrays, Phys. Rev. B 38, No. 13 (1988) 8712-8719.
[HBW2) P. HADLEY, M.R. BEASLEY & K. WIESENFELD, Phase locking of Josephson-junction arrays,
Applied Phys. Lett. 52, No. 19 (1988), 1619-162l.
[IG) E. IHRIG & M. GOLUBITSKY, Pattern section with 0(3) symmetry, Physica 120 (1984),1-33.
[ML) R.M. MAY & W.J. LEONARD, Nonlinear aspects of competition between three species, SIAM
J. App!. Math. 29 (1975), 243-253.
[M) 1. MELBOURNE, Intermittancy as a co dimension three phenomenon, Dyn. Oiff. Eqn. 1, No.
4 (1989), 347-368.
[MCG] 1. MELBOURNE, P. CHOSSAT & M. GOLUBITSKY, Heteroclinic cycles involving periodic so-
lutions in mode interactions with 0(2) symmetry, Proc. R. Soc. Edinburgh 1I3A (1989),
315-345.
[R) O. RUELLE, Bifurcations in the presence of a symmetry group, Arch. Rational Mech. Ana!'
51 (1973), 136-152.
[SG) F. SIMONELLI & J. GOLLUB, Surface wave mode interactions: effects of symmetry and de-
generacy, J. Fluid Mech. 199 (1989), 471-494.
[V] A. VANDERBAUWHEDE, Local Bifurcation and Symmetry, Habilitation Thesis, Rijksuniver-
siteit Ghent, 1980; Res. Notes in Math. 75, Pitman, Boston, 1982.
 DYNAMICS OF SOME ELECTROCHEMICAL REACTIONS

J. L. HUDSON*

Abstract. Experiments on a few electrochemical reactions are discussed. Time series of ei-
ther current or voltage, obtained under potentiostatic or galvanostatic conditions respectively, are
presented and characterized. We first some some examples of dynamic behavior such as chaos,
quasiperiodicity, and period doubling of tori obtained during the electrodissolution of copper.
Some apparent higher order chaos during electrodissolution of iron is then discussed. Finally, we
treat briefly coupled electrochemical oscillators.

Key words. dynamics, chaos, electrochemical reactions

1. Introduction. Chemically reacting systems are often highly nonlinear and

thus have furnished many examples of interesting dynamic behavior. For exam-
ple, the exothermic, single irreversible reaction (a two variable system) has been
analyzed extensively by chemical engineers; the existence of multiple states and
oscillatory behavior is now well understood [2,3,4,22,40,41]. Two independent reac-
tions, such as consecutive reactions, with heat effects are governed by three ODE's
and can, in addition, produce chaos [16,17,19,20].
There has been a considerable amount of experimental work done on the dy-
namics of isothermal reactions carried out in continuous stirred reactors. In these
cases the nonlinearities are associated with the chemical kinetics themselves rather
than with the Arrhenius temperature dependence of the reaction rate constant as
in the nonisothermal reactions discussed above [11,15]. The best known isothermal
oscillating reaction system is the Belousov-Zhabotinski reaction. Much of the early
experimental work on chemical chaos was done with this reaction [18,31,34].
In this paper we discuss the dynamics, with emphasis on chaotic behavior,
of some electrochemical reactions. Experiments with electrochemical systems have
produced examples of several types of periodic, quasiperiodic, and low order chaotic
behavior. [1,6-10,25,2937,38]. In most cases the dynamics are fast (with frequen-
cies much higher than those obtained in stirred reactors) and reasonably stationary
so that analyses of the time series can be carried out by standard methods such as
attractor reconstruction, dimension calculations, construction of return maps and
Poincare sections, etc.
We concentrate on the dynamics of the reactions (rather than on the physio-
chemical behavior of the electrode surface or the possible causes of instabilities).
Two reactions are discussed: The electro dissolution of copper and that of iron.

*Department of Chemical Engineering, University of Virginia, Charlottesville, VA 22903-2442.

This work was supported in part by the National Science Foundation and by the Center for
Innovative Technology, Commonwealth of Virginia.
 90

2. Experiments. All of the experiments were carried out with a standard

electrochemical apparatus, viz., a rotating disk electrode. This is a rod of the
appropriate metal, exposed at one end, which is rotated in an electrolyte. The
overall rate of reaction is influenced by the kinetics on the electrode surface, by
transport through any film which may exist on the electrode surface, and by mass
transfer in the solution; the latter can be controlled by the rotation rate of the disk.
Experiments are normally carried out either potentiostatically or galvanostat-
ically. In the former case current is measured as a function of time; in the latter
case the potential relative to some reference is measured. Data are taken digitally
at rates fast enough to analyze the time signal. In the experiments discussed here
this is in the range of 60 to 5000 Hz.

3. Copper electrodissolution. Potentiostatic oscillations have been ob-

served using a copper electrode in solutions of sulfuric acid and sodium chloride
[6-9]. Under some conditions the current undergoes temporal oscillations. The ex-
act mechanism of the reaction and the cause of the oscillations have not yet been
determined. Some of the reactions which occur under the conditions of the exper-
iment, however, are the following: The copper is oxidized via the anodic reaction
[23,24].
Cu + CI- --> CuCI + e-.

The CuCI forms a film on the surface of the copper electrode and this film serves
as a resistance to the transport of CI- ions. The film dissolves via the reactions

CuCI + (x -1) CI- --> CuCI;-x (aq)

and [23,24]

The rate of formation of the film is somewhat greater than the rate of dissolution,
and there is thus a slow net growth in the thickness of the CuCI film, and therefore.
also in the resistance to transport of CI- to the metal surface. This resistance
appears to be an important parameter for the system. In some of the results of this
section transitions among states are shown; these transitions are caused by changes
in a slowly varying parameter, viz., the properties and thickness of the layer on the
disk surface.
Evidence for several types of low-dimensional chaotic behavior has been ob-
tained with this reaction system. We show a few examples. In Figure 1 a time
series exhibiting chaotic features is shown [9]. The attractor (Figure 1b), Poincare
section (Figure lc), and return map (Figure Id) indicate that the time series (Fig-
ure la) is simple chaos. A sequence of Poincare sections is shown in Figure Ie to
explain the flow of trajectories around the attractor. The Poincare sections were
made with planes at constant I(t - 27) = 17.5 rnA and I(t - 47) = 17.5 rnA and
were constructed by using all the intersections of the attractor of Figure Ib with
 91

r
•
to

.,
;. ... r'
: ..

1~~I7--------I-(.-.2-4------~2$·C

FIGURE 1.

chaos: (a) time series. (b) Attractor (spline fit). (c) Poincare section
taken at I(t) = 17.5 rnA; increasing I(t - 47). (d) Return map con-
structed from the Poincare section taken at It - 27) = 17.5 rnA with
increasing I(t - 47); using variable I(t - 47) . (e) Series of Poincare
sections showing the flow of trajectories around the attractor. Sec-
tions 1 and 3 were taken at I( t - 27) = 17.5 rnA and sections 2 and
4 w ere taken at I( t - 47) = 17.5 rnA [9].

the above-mentioned planes. In Figure Ie only a few orbits around the attractor are
shown so that the location and orientation of the section can be seen. The arrow on
each section is used to mark the points that started on the edge of the attractor in
section 1. This sequence shows the stretching, folding and mixing, and contraction
processes common to these types of simple chaotic attractors. In going from section
 92

1 to section 2, one observes the stretching of the trajectories. In going from section
2 to 3, the trajectories continue to stretch but the folding process is also observed.
In addition, the points that were on the outside of the attractor are now on the
inside. In going from section 3 to 4, one sees the contraction of the fold onto itself.
This process continues as trajectories proceed around the attractor to give back the
result seen in section 1.
From the shapes of the attractor, Poincare sections, and return map, the chaos
shown in Figure 1 appears to be a simple type that may be embeddable in three-
dimensional space and that has a single positive Lyapunov exponent. For further
verification the correlation dimension [14] and the largest Lyapunov exponent [35]
were calculated. The correlation dimension was found to be 2.2. An approximate
value of 0.6 for the Lyapunov exponent was calculated from the return map.
The chaotic behavior shown in Figure 1 was preceded by a period doubling of
periodic orbits and chaos on a two-band attractor.
Another type of behavior, apparently chaos on a broken toroidal structure, is
shown in Figure 2. The attractor is shown in Figure 2b, and a Poincare section, a
closed curve indicating a toroidal substructure, is shown in Figure 2c. A series of
Poincare sections (Figure 2d) shows the stretching and folding process that indicates
the chaotic nature of flow. The planes used to make the sections are defined by the
equation I(t)+I(t - 27) - 40 = a(I(t - 47) - 20). Changing the value a rotates the
plane about a line that goes through the center of the attract or and is defined by
the intersection of the planes I(t)+I(t - 27) = 40 and I(t - 47) = 20. The first cut
shows a closed cross section. Two areas of the section are marked (A and B) to
point out the folding process. The folding become noticeable in the second section
as A and B are coming together. The folding process is almost complete in the
third section as A and B have almost totally come together. By the fourth section
the folds have become so close together that the separation can no longer be seen.
The fifth section is 180 0 from section one; in the remaining 180 0 the shapes of the
sections undergo no additional significant changes. The behavior shown in Figure
2 was followed by quasiperiodic behavior which is not shown.
 93

104

......
'
<
.!

'or
:,
" i

TIME 17 'xtL--------'---" 1(1)

c

.....-.:~ ... :~~:

, ......:;
¥:'L'
I· ~ ··
i~ ..'
FIGURE 2.

Chaos on a broken torodial structure: (a) Time series. (b) Attrac-

tor (spline fit). (c) Poincare section taken at I(t - 2r) = 20 rnA;
increasing I( t - 2r). (d) Series of Poincare sections showing the fold-
ing process. All planes used to make the sections are defined by the
general equation I(t)+I(t - 2r) - 40 = a(I(t - 4r) - 20). For section
1 and 5, a = 1; for section 2, a = -1; for section 3, a = -5; and for
section 4, a = 7. These planes are marked by lines on the attractor.
Attractor is the same as that shown in Figure 2b [9].

We do show a different example of quasiperiodic behavior which was obtained

during a sequence of period halving of tori [8]. The observed sequence was chaos on a
broken torus, chaos on a two-band broken torus, double torus, torus, limit cycle, and
steady state. Each of these transitions occurred smoothly. For example, the cross
section of the torus decreased slowly until the limit cycle resulted; the amplitude of
the limit cycle decreased resulting on a steady state. The flow on a torus is shown
on Figure 3. The time series (Figure 3a) has quasiperiodic features; there are
oscillations of period 0.76 s, the maxima of which are also periodic, with a period
of 3.57 s. The power spectrum (figure 3b) has two incommensurate frequencies
II = 1.32 Hz and 12 = 0.28 Hz both of which have peaks well above the underlying
 94

it-12.5 s--'l
25.5'--_ _ _ _ _ _ _ _ _ _ _ _----'
-6+--~-~-~-~~-___l
TIME 7.5
FREQUENCY (Hz)

28.6 c 2I1 d

::."
.... .. ~
.
",."
'"
3n
,-
...
'

~l

28.2 ¥----_~-~-~-~-~~ 0
~
.... ,
'
28.2 28.6 0 2I1
6 n _1

e
29

26
26
29
I(t-H) I(t)
26 29

FIGURE 3.

Torus. (a) Time series. (b) Power spectrum; fr = 1.32 Hz; 12 = 0.28
Hz. (c) Next maximum map; (d) Next angle return map of (c);
center point used for calculation was (28.38, 28.38). (e) Attractor
(spline fit); the two projections a~e made at angles differing by 6°.
The right picture is for the left eye and the left for the right. (T =
1/60s) [8).
 95

nOlse. The frequency II is associated with the flow around the main loops of the
torus. The frequency 12(0.28 Hz) is associated with progression in the plane of
the Poincare section or the next maximum map (Figure 3c). This progression is
in a clockwise direction in the next maximum map. The time associated with this
progression, T = 1/ h = 3.57 s, is also indicated on the time series. Both the
Poincare section (not shown) and the next maximum map (Figure 3c) should be
closed curves for quasiperiodic behavior. The next maximum map is shown here
rather than a Poincare section since, as is usually the case with experimental data,
it has less scatter.
A next angle return map was made from Figure 3c and is shown in Figure 3d.
This map was made by first translating the coordinate system of Figure 3c to (28.38,
28.38) which lies in the hole of the map. The coordinate system was then rotated
so that the first point of the next maximum map was at 0 0 • The angle of the n + 1st
point with respect to this new coordinate system was plotted against the angle of
the nth point for all the points in the map. Figure 3d appears to be approximately
one-dimensional and invertible indicating quasiperiodic behavior. A reconstruction
of the attractor is presented in Figure 3e and has the appearance of a torus.
The quasiperiodic behavior of Figure 3 was preceded by the quasiperiodic be-
havior of figure 4. As can be seen from the time series in Figure 4a, the maxima
of the oscillations now form an oscillatory pattern whose period is doubled when
compared to that of Figure 3a. The power spectrum (Figure 4b) shows two in-
commensurate, fundamental frequencies at II = 1.32 Hz and h/2 = 0.14 Hz. The
frequency II corresponds to the main flow around the center hole of the torus. The
frequency h/2 represents progression in the plane of a Poincare section (not shown)
or equivalently in the next maximum map shown in Figure 4c. The frequency dou-
bling which has occurred in going from Figure 3 to Figure 4 is associated with the
frequency of the progression in the plane of the Poincare section or in the next
maxilTIUm map.

The frequency h/2 can be seen in the power spectrum, Figure 4b. Although
the frequency h/2 is not overly strong, it does not appear to be part of the under-
lying noise. This argument is further supported by the following two observations.
First, the frequency h/2 is followed in the power spectrum by the integer multi-
ples 2(h/2), 3(h/2), and 4(h/2). Secondly, we can estimate the frequency h/2
directly from the time series and compare the result favorably to Jd2 = 0.14 Hz.
The time corresponding to frequency 12/2 = 0.14 s, is indicated by the time seg-
ment T = TL + T. on fig. 4a. The time TL represents the transversal of the larger
loop of Figure 4c and T. represents the transversal of the smaller. Thus the wave
forms associated with the maxima of the oscillations as seen in Figure 4a contain a
larger maxima segment for TL and a smaller maxima segment for T., and the total
is commensurate with the frequency 12/2.
The next angle return map is shown in Figure 4d. The coordinates now go up
to 471" since there is a double loop in Figure 4c.. Figure 4d was constructed in the
same manner as the next angle return map in Figure 3d; a point was chosen inside
the small loop of Figure 4c and the angles calculated from it. A reconstruction of
the attractor is shown in Figure 4e.
 96

r,
III
----.-~--

0-
'co"
«
§
Ul
<'
~ '0"
;0
E-<
::;
Po.
:::E
~
iii
~12.S s--l -"
25.5
-5
TIME 0
FREQUENCY (Hz) 7.5

28.7 C 4n d
'.
., .
•*

Z .'
-;
~

'1 en
~
• t

('

. ..
28.2
0 ~

28.2
Imax{N-l) 28.7 0 4n
enol

e
29

!='
~,
..,
~

26
29 26

I(t-BT) I( t)
26 29

FIGURE 4.
Double torus. (a) Time series. (b) Power spectrum; !I = 1.32 Hz;
Jz/2 = 0.14 Hz. (c) Next maximum map. (d) Next angle return
map of (c); center point used for the calculation was (28.38, 28.38).
(e) Attractor (spline fit) [8].

The double torus of Figure 4 is preceded by what appears to be chaos on a

broken double toroidal structure and that itself preceded by chaos on a broken
torus.
 97

4. Iron electrodissolution. Oscillations in current also occur during the

potentiostatic electrodissolution of iron in acidic solutions [10,13, 32]. Those oscil-
lations which occur on the mass transfer limited plateau can be faster and more
complicated than those seen above in the previous section.
An example of a time series for the iron/sulfuric acid system is shown in Figure
5. Note that only one second of data is shown. These data were taken at 5000 Hz.
The detailed nature of the time series depends, of course, on parameters such
as acid concentration, disk rotation rate, applied potential, and the size of the disk.
We consider here only the latter.
The complexity of the time series increases with increasing electrode size. For
example, increasing the disk diameter in small increments from 2.0 mm to 6.25
mm yields a series of behaviors from a limit cycle to low order chaos (correlation
dimension slightly over 2.0) to a higher order chaos (correlation dimension well
above 3.0). Figure 5 shows the time series for the largest electrode.

370,-------------------------------------------------------,

1""<:-----------0.4 s ----.::..1,,1
220~----------------------------------------------------~
TIME

FIGURE 5.
Time series from iron/sulfuric acid system

We do not know why the complexity increases with an increase in the area of
 98

the electrode. One obvious possibility is the increase in the number of reacting sites
on the surface.

5. Coupled electrodes. The final comment of the last section leads perhaps
naturally to the subject of coupled electrochemical reactors. The subject of coupled
chemical oscillators has received considerable, numerical and analytical attention
[5,21,26,27,30,39]. However, there are few experimental studies [36]. It is difficult
to couple chemical reactors without changing the nature of the uncoupled oscillators
in the process.
Electrodes may be suitable for studies of coupled chemical oscillators. We have
carried out some preliminary experiments with two electrodes. One was a rotating
disk electrode similar to those used in the experiments described in the preceding
sections. The second was a similar, but non-rotating, electrode placed opposite the
rotating electrode on the same axis. Two potentiostats and measurement systems
are used so that all the coupling occurs through the electrolyte. The experiments are
run under conditions such that each electrode would oscillate if the other were not
active. In our preliminary experiments only frequency locking has been observed.
We hope to exploit this system (or most likely other geometries) to investigate other
phenomena of coupled chemical oscillators.

6. Final Comments. We have shown some dynamic behavior obtained with

the studies on the electrodissolution of metals. Other types of electrochemical
reactions have received less attention from a dynamic standpoint; however, some
work has been done recently on electrodeposition and electrocatalytic reactions
[12,33].
V>le note that several types of periodic and low dimensional chaotic behavior
have been observed. Although higher order chaos can occur, as discussed briefly on
Section 4, it is interesting that such a degree of order can prevail in such complicated
systems.
REFERENCES

[1] F. N. ALBAHADILY AND M. SCHELL, An experimental Investigation of Periodic and Chaotic

Electrochemical Oscillations in the Anodic Dissolution of Copper in Phosphoric Acid, J.
Chem. Phys. 88 (1988), pp. 4312-4319 ..
[2] R. ARIS, The Mathematical Background of Chemical Reactor Analysis. II. The Stirred Tank,
in Reacting Flows, a.s.s. Luddford (ed.), Lectures in Applied Math, 24 (1986), pp. 75-108.
[3] R. ARIS, AND N. R. AMUNDSON, An Analysis of Chemical Reactor Stability and Control-I,
Chem. Eng. Sci., 7 (1958), pp. 121-13l.
[4] V. BALAKOTAIAH AND D. Luss, Analysis of Multiplicity Patterns of a CSTR, Chem. Eng.
Commun., 13 (1981), pp. 111-132.
[5] K. BAR-ELI, On the Coupling of Chemical Oscillators, J. Phys. Chem., 88 (1984), pp.
3616-3622.
[6] M. R. BASSETT AND J. L. HUDSON, Experimental Evidence for Period Doubling of Tori
During an Electrochemical Reaction, Physica D, 35 (1989), pp. 289-298.
[7] M. R. BASSETT AND J. L. HUDSON, The Dynamics of the Electrodissolution of Copper,
Chern. Eng. Comm., 60 (1987), pp. 145-159.
[8] M. R. BASSETT AND J. L. HUDSON, Shil'nikov Chaos During Copper Electrodissolution,
Journal of Physical Chemistry, 92 (1988), pp. 6963-6966.
 99

[9] M. R. BASSETT AND J. L. HUDSON, Quasiperiodicity and Chaos During An Electrochemical

Reaction, Journal of Physical Chemistry, 93 (1989), pp. 2731-2737.
[10] C. B. DIEM AND J. L. HUDSON, Chaos During the Electrodissolution of Iron, A.I.Ch.E. J.,
33 (1987), pp. 218-224.
[11] M. FEINBERG, Reaction Network Structure, Multiple Steady States, and Sustained Compo-
sition Oscillations: A Review of Some Results, in Modelling of Chemical Reaction Systems"
K. H. Ebert, et al. eds., Springer, Berlin (1981).
[12] N. FETNER AND J. L. HUDSON, Oscillations during the electrocatalytic reduction of hydrogen
peroxide on a platinum electrode, preprint, (1990).
[13] U. F. FRANCK AND R. FITzHUGH, Periodische Elektrodenprozesse und ihre Beschreibung
durch ein mathematisches Modell, Z. Electrochem, 65 (1961), pp. 156-168.
[14] P. GRASSBERGER AND I. PROCACCIA, Characterization of Strange Attractors, Phys. Rev.
Lett., 50 (1983), pp. 346-349.
[15] P. GRAY AND S. K. SCOTT, Autocatalytic Reactions in the Isothermal CSTR: Oscillations
and Instabilities in the system A + 2B ..... 3B, B ..... C, Chern. Eng. Sci., 39 (1984), pp.
1087-1097.
[16] J. L. HUDSON AND I. G. KEYREKlDIS, Dynamic Behavior or Nonisothermal Continuous
Stirred Tank Reactors in Handbook of Heat and Mass Transfer, N. P. Cheremisinoff, ed.,
Gulf, Houston, 3 (1989), pp ..
[17] J. L. HUDSON AND O. E. ROSSLER, Dynamics of Nonlinear Systems, V. Hlavacek, ed., Gordon
and Breach Science: New York (1986), Chapter 6.
[18] J. L. HUDSON, M. HART, AND D. MARINKO, An experimental study of multiple peak periodic
and nonperiodic oscillations in the Belousov-Zhabotinskii reaction, J. Chern. Phys. 71 (1979),
pp. 1601-1606.
[19] D.V. JORGENSEN AND R. ARIS, On the Dynamics of a Stirred Reactor with Consecutive
Reactions, Chern. Eng. Sci., 38 (1983), pp. 45-53.
[20] C. KAHLERT, O. E. ROSSLER, AND A. VARMA, Chaos in a Continuous Stirred Reactor with
Two Consecutive First Order Reactions, one Exo-, One Endothermic, in Modeling Chemical
Reaction Systems, K. Ebert and W. Jaeger, ed., Springer, Heidelberg (1981).
[21] C. R. KENNEDY AND R. ARIS, Bifurcations of a Model Diffusion-Reaction System, in New
Approaches to Nonlinear Problems in Dynamics, P. J. Holmes, ed., SIAM (1980).
[22] V. K. KWONG AND T. T. TSOTSIS, Fine Structure of the CSTR Parameter Space, AIChE
J., 29 (1983), pp. 343-317.
[23] H. P. LEE AND K. NODE, Kinetics and Mechanisms of Cu Electrodissolution in Chloride
Media, J. Electrochem, Soc., 133 (1986), pp. 2035-2043.
[24] H. P. LEE, K. NOBE, AND A. PEARLSTEIN, Film Formation and Current Oscillations in the
Electrodissolution of Cu in Acidic Chloride Media, I. Experimental Studies, J. Electrochem.
Soc., 132 (1985), pp. 1031-1037.
[25] O. LEY, A. WOLFFBERG, M. SllEINTUCH, AND L. M. PISMEN, Bifurcations to Periodic and
Chaotic Motions in Anodic Nickel Dissolution, Chern. Eng. Sci., 43 (1988), pp. 1339-1353.
[26] J. C. MANKIN AND J.L. HUDSON, The dynamics of Coupled Nonisothermal Continuous
Stirred Tank Reactors, Chern. Eng. Sci., 41 (1986), pp. 2651-2661.
[27] H. G. OTHMER, D. G. ARONSON, AND E. J. DOEDEL, RESONANCE AND BISTABILITY IN
COUPLED OSCILLATORS, Phys. Letters, 113A (1986), pp. 349-353.
[28] N. H. PACKARD, J. P. CRUTCHFIELD, J. D. FARMER, AND R. S. SHAW, Geometry from a
Time Series, Phys. Rev. Lett., 45 (1980), p. 712.
[29] A. J. PEARLSTEIN, II. P. LEE, AND K. N OBE, Film Formation and Current Oscillations
in the Electrodissolution of Copper in Acidic Chloride Media, II. Mathematical Model, J.
Electrochem. Soc., 132 (1985), pp. 2159-2165.
[30] V. RAYI KUMAR, B. JAYARAMAN, B. KULKARNI, AND L. DORAISWAMY, Dynamic Behavior of
Coupled CSTRs Operating Under Different Conditions, Chem. Engng. Sci., 38 (1983), pp.
673-686.
 100

[31] J. C. Roux, J. S. TURNER, W. D. MCCORMICK, AND H. L. SWINNEY, Experimental ob-

servations of complex dynamics in a chemical reactor, in Nonlinear Problems Present and
Future, A. R. Bishopp, ed. North Holland (1981).
[32] P. RUSSELL AND J. NEWMAN, Current Oscillations Observed within the Limiting Current
Plateau for Iron in Sulfuric Acid, J. Eledrochem. Soc., 133 (1986), pp. 2093-2097.
[33] M. SCHELL, F. N. ALBAHADlLY, J. SAFlR, AND Y. Yu, Characterization of oscillatory states
in the electrochemical oxidation of formaldehyde and formate/formic acid, J. Phys. Chern.,
83 (1989), pp. 4806-4810.
[34] R. A. SCHMITZ, R. R. GRAZIANI, AND J. L. HUDSON, Experimental evidence of chaotic states
in the Belousov-Zhabotinskii reaction, J. Chern. Phys. 67 (1977), pp. 3040-3044.
[35] H. G. SCHUSTER, Deterministic Chaos, Physik-Verlag: Weinheim (1984).
[36] I. STUCHL AND M. MAREK, Dissipative Structures in Coupled Cells: Experiments, J. Chern.
Phys. 77 (1982), pp. 2956-2963.
[37] J. B. TALBOT AND R. A. ORIANI, Application of Linear Stability Analysis to Passivation
Models, J. Eledrochem. Soc., 132 (1985), pp. 1031-1037.
[38] L. T. TSITSOPOULOS, T. T. TSOTSIS, AND I. A. WEBSTER, An Ellipsometric Investigation
of Reaction Rate Oscillations During the Electrochemical Anodization of Cu in H3 P04
Solutions, Surface Science, 191 (1987), pp. 225-238.
[39] T. T. TSOTSIS, Spatially Patterned States in Systems of Interacting Lumped Reactors,
Chern. Eng. Sci., 38 (1983), pp. 701-717.
[40] A. UPPAL, W. II. RAY, AND A. POORE, On the Dynamic Behavior of Continuous Stirred
Tank Reactors, Chern. Eng. Sci., 29 (1974), pp. 967-985.
[41] A. VARMA AND R. ARIS, Stirred Pots and Empty Tubes, in Chemical Reactor Theory: A
Review, L. Lapidus and N. R. Amundson (ed.), Prentice-Hall, (1977), 79-155 ..
 CONSTRUCTION OF THE FITZHUGH-NAGUMO PULSE
USING DIFFERENTIAL FORMS

C. JONES*, N. KOPELL** AND R. LANGERt

1. Introduction. Systems of singularly perturbed ordinary differential equa-

tions can often be solved approximately by singular solutions. These singular so-
lutions are pieced together from solutions to simpler sets of equations obtained as
limits from the original equations. There is a large body of literature on the ques-
tion of when the existence of a singular solution implies the existence of an actual
solution to the original equations. Techniques that have been used include fixed
point arguments (Conley [1], Carpenter [2], Hastings [3] and Gardner and Smoller
[4]), implicit function theorem and related functional-analytic techniques (Fife [5],
Fujii et al. [6], Hale and Sakamoto [7,8]) differential inequalities (see for instance
Chang and Howes [9]) and nonstandard analysis (Diener and Reeb [10]).
Some of these techniques apply to nerve impulse equations, including the Hodgkin-
Huxley equations and a widely used simplified version known as the FitzHugh-
Nagumo equations. These equations are

au a2 u
-=-+f(u)-w
(Ll) at ax 2
aw
c(u -,w)
at
- =

where c is a small parameter and feu) = u(1 - u)(u - a), with a any constant
°
satisfying < a < 1/2. The solutions of interest are travelling pulses. These are
solutions depending only on (= x+8t, for some wave speed 8 and which, as (-+ 00
or ( -+ -00, approach the unique rest state of (1.1) b is chosen small enough so
that u = ,wand w = f( u) have only (0,0) as a simultaneous solution). These
travelling waves then satisfy

u' =v
(1.2) v'=8v-f(u)+w , = dld(
w' = ':(u -,w)
8

together with the conditions (u, v, w) -+ (0,0,0) as ( -+ ±oo. It was proved inde-
pendently by Carpenter [2] and Hastings [3] that there is a value of 8 at which a
travelling pulse exists, if € is sufficiently small. In fact there are two values of 8 but
we are interested in the faster pulse.

*Department of Mathematics, University of Maryland. Research partially supported by NSF

OMS 880 1627 and an award from the Graduate Research Board of the University of Maryland
**Oepartment of Mathematics, Boston University. Research partially supported by NSF OMS
8901913 and AFOSR-90-0017
tSky Computers, Inc.
 102

In [11], one of us (RL) used a geometric approach most similar in spirit to that of
Carpenter. However, the technique of [11] was designed to give a somewhat stronger
result: by constructing the desired solutions as the intersection of stable and un-
stable manifolds, one can obtain local uniqueness as well as existence. Furthermore
the geometry exposed in the proof (specifically the "direction" of transversality)
provides information that was crucial in proving the stability of these solutions (see
Jones [12]).
There is one part of the proof in [11] which stands out as technically the most
difficult; that is the part where the solutions are controlled by the slow flow. In
recent work [13], the other two authors of the current paper have introduced a new
technique that simplifies and considerably clarifies this part of the construction.
This paper is then a new presentation of the proof of the travelling pulse for (1.1)
as given in [11] but with this new method used to control the behavior near the
slow manifold.
To exhibit the desired solution as an intersection of stable and unstable man-
ifolds, it is necessary to follow the latter globally in space. As we show, one can
factor the flow into a series of slow and fast parts; on a time scale in which the fast
flow is order 1, the slow flow takes a time of order lie. Nevertheless it turns out to
be possible to follow an invariant manifold for that length of time. The key idea is
that when the slow flow is dominant, the solution is near a "slow manifold" (to be
defined below). The configuration of the tangent plane to the unstable manifold of
the rest state changes as it passes near the slow manifold. Certain tangent direc-
tions are lost and converted into new ones which reflect the behavior on the slow
manifold. We think of this as an exchange of information that occurs during the
passage near the slow manifold, and thus we call the central result the "exchange
lemma."
Our techniques for tracking an invariant manifold as it passes near a slow man-
ifold are reminiscent of the ".A-lemma" (see [14]), which describes the behavior of
an invariant manifold in the neighborhood of a hyperbolic critical point p. The
latter says that if, at some time, the manifold has the dimension of WU(p) and is
transverse to W8(p), then as t --> 00 the manifold approaches WU(p). Our situation,
however, is considerably more complex: instead of a critical point, we have a slow
manifold of arbitrary dimension. The manifold approaches the product of the un-
stable directions with some center ones. The analogue of the ".A-lemma" would say
that the manifold approaches the unstable directions, but this does not exhaust all
the dimensions and the more subtle part lies in determining which center directions
are picked out.
Our strategy is then to follow the tangent space to an invariant manifold through
studying the induced flow on the space of exterior forms. The coordinates of a tan-
gent vector can be viewed as I-forms; analogously, a tangent plane has coordinates
that are 2-forms. (Each is the area of the projection of the unit square in the tan-
gent space onto a coordinate plane.) These evolve according to an easily computed
equation and we control the tangent planes through estimates on their solutions.
Our goal in this paper is then twofold. First, we wish to make the geometric
 103

proof of [11] widely available. Secondly, we show how the techniques of [13] can be
used in a basic example.

2. FitzHugh-Nagumo Equations. The travelling wave equations for the

FHN system are given by (1.2). We shall show that these equations have an orbit
that is homoclinic to (0,0,0) when € is sufficiently small. The orbit is constructed
close to a certain "singular" solution, which is a union of trajectories for the € = 0
equation.
Setting € = 0 in (1.2), the equation for w is w' = O. Hence w acts as a parameter
in the other equations
u' = V
(2.1 )
v'=(Jv-f(u)+w.
We shall sometimes append to (2.1) one or the other of the auxiliary equations
w' = 0 and (J' = O. In each case, which augmented equation of (1) we are considering
will be made clear by the specific variables written.
For each (J and w, (2.1) has the rest point (u,v,w), with v = 0 and w = feu).
For fixed (J there are two branches of the graph of w = f( u) having negative slope.
We call the associated curves of rest points SL and SR, where Land R denote
the left and right hand branches respectively. See Figure 1. Aside from the points
(u, v, w) at which f' (u) = 0, each such point is hyperbolic in its (u, v) plane, with (J
and w fixed. Thus, each p € SL U SR has a well defined unstable and stable manifold
W"(p) and WS(p).

Figure 1. The rest points of 2.1. Of most interest are the two
curves S Land S R on which the graph of w = f( u) has negative
slope.

For w = 0 and each (J, the rest point (u,v) = (0,0) in SL is a rest point for the
full system (1.2). It is well known that there is a value of (J, which we shall denote by
 104

8., for which there exists a connecting orbit in w == 0 from (u,v,w) = (0,0,0) E SL
to (u,v,w) = (1,0,0) E SR. (See for instance Aronson and Weinberger [15].) The
first piece of the singular solution is this connecting orbit.
By an argument similar to the one that determines 8., it can be shown that for
8 = 8. there is a w. for which (2.1) possesses a connecting orbit from (u_,O,w.) E
SR to (u+,O,w.) E SL, where u_ and u+ satisfy w. = f(u). The second piece of
the singular solution is the subset SR with 8 = 8. and 0 ::; w ::; w •. The third is the
connecting orbit from (u_,O,w.) to (u+,O,w.). The last is the portion of SL with
8 = 8., traversed downward from w = w. to w = O. See Figure 2. The singular
solution is not smooth, since it has corners at the ends of each piece.

Figure 2. The singular orbit associated with 8 = 8*.

We can now state the theorem on the existence and uniqueness of the pulse
solution to (1.2).

THEOREM: If e is sufficiently small, then for some 8 within O(e) of 8., (1.2)
possesses an orbit which is homoclinic to the rest point (0,0,0) of (1.2). Moreover,
this orbit lies within O( e) of the singular solution S, and it is the unique such
homoclinic orbit in such a neighborhood.

The strategy of the proof is to display the homo clinic orbit as the transverse
intersection of the unstable and stable manifolds of (u, v, w) = (0,0,0) for some
8 = 8( e), e i= 0 small. To do this, one shoots forward using the union of the
unstable manifolds over a range of 8. Abusing notation, we shall use (1.2) to also
denote the system with the equation 8' = 0 appended. The union of unstable
manifolds mentioned above is then a portion of the center-unstable manifold of
(1.2). We denote this by WCU = UWU(O, 8), where 0 is the origin in (u, v, w) space
and the union is over 8 E [8. - 6,8* + 6] for some appropriately small 6.
 105

The difficulty of the proof is in following W eu in order to show that, for some
(), WU(O, (}) has a transverse intersection with WS(O, (}). It is fairly straightforward
to follow the manifold over the front, where each trajectory in W eu remains close
to the connecting trajectory of (2.1) from (0,0,0) to (1,0,0). The next and hardest
step is to follow w eu as it is carried close to the manifold SR. This is where
the exchange lemma is used. W eu is then followed over the back. At the end of
this third step, one is in a position to check the transversal intersection with the
center stable manifold. The transversality calculations are carried out in (u, v, w, ())
space, and the locus of intersection determines the value of () for the travelling wave.
The transversality calculations then show that for that value of (), WU(O,(}) has a
transverse intersection with WS(O, (}).
The basic building blocks of the proof are the exchange lemma and two transver-
sality lemmas involving only the behavior of the equations when € = 0. The latter

°
results involve geometric objects built up out of the stable and unstable manifolds
at € = of the points of SL and SR. The first result is transversality along the
front, as the wave speed () varies: For each () near ()., consider (2.1) with w = 0. The
critical point (tt, v) = (0,0) has a one dimensional unstable manifold. The center
unstable manifold Weu(O, 0, (}.) of (tt, v, ()) = (0,0, ().) is a two-dimensional manifold
that is the union of the one-dimensional unstable manifolds. Similarly, at w = the °
rest point (u, v) = (1,0) has a one dimensional stable manifold. The center stable
manifold W es (l,O,(}.) of (u,v,(}) = (1,0,(}.) is a two-dimensional manifold that is
the union of the stable manifolds. These two-dimensional manifolds intersect along
the connecting orbit at () = ().. The first transversality result is that this pair of
two dimensional manifolds are transverse in (u, v, B) space along the front.

LEMMA 1. Weu(O,O,B.) intersects IV es (1,O,B.) transversely in (u,v,(}) space,

w =0.

The other transversality result concerns the back, as () is held fixed at B. and w
is varied. As above, the rest points (u+,O,w.) E SR and (u_,O,w.) E SL have
two-dimensional center unstable and center stable manifolds in (u, v, w) space that
intersect along the connecting orbit in w = w •.

LEMMA 2. Weu(u_,O,w.) intersects Wes(u+,O,w.) transversely in (tt,v,w)

space, B = B•.

The proofs of lemmas 1 and 2 are well known. However, we take the opportunity in
Section 4 to show how these can also be done very simply using differential forms.
We shall soon state the lemmas that will be used to describe the behavior of
W eu as it flows near SR. We first need to define some objects associated with the
€ = °flow. Let SR be a compact portion of SR that includes -8 :S w :S w. + 8
and let B denote a neighborhood in (u,v,w,B) space of the set SR x I., where
I. = [8. - 8, (). + 8). Let WU(SR x I.) and WS(Sn x I.) be the three-dimensional
°
unions of unstable and stable manifold~ of SR x 16 for the € = equation (2.1). See
Figure 3. Let W"(S R) be the two-dimensional restriction of WU(S R x I.) to B = B*;
we think of W"(SR) as a subset of (u,v,w,B) space. (Note that W"(SR) includes
 106

what we previously called WCU(u_,O,w.). We were then focusing on values of w

near w. and we are now interested in a larger portion of SR.) The following two
lemmas are direct consequences of the exchange lemma to be discussed in the next
section. In the proof of the existence of the homoclinic orbit, they will be used to
relate the E =1= 0 manifold W CU that we are trying to follow to the singular object
wu(S R) about which we have transversality information from Lemma 2.

Figure 3. Slice in u, v, w, f) space of the stable and unsta-

ble manifolds of SR and SL for fixed 8. The slices are two-
dimensional; vllS(S R x 16) etc. are all three-dimensional in
u, v, w, f) space.

Let r. be a smooth curve in BB, smoothly parametrized by E, that is transverse

to WS(SR x 16). See Figure 4. Assume the points y. E r. have w coordinates near
w = O. Denote by z. E BB the point at which the trajectory through y. exits B.
The first lemma says that the exit points lie close to the two-dimensional manifold
wu(SR). Moreover, the "entrance curve" r. gets stretched under the flow into an
"exit curve" that is Co close to WU(SR) n BB. See Figure 4.

LEMMA 3. Ify. E r. is sufficiently close to WS then z. is within O(E) ofwu(SR).

Moreover, for each point p E WU(SR) n BB whose w coordinate is bounded away
from 0 by a fixed amount, there exists a y. E r. whose associated z. is within O( E)
ofp.

Let G. be the two-dimensional manifold that r.

sweeps out under the flow
of (1.2). It is an immediate corollary of Lemma 3 that for any z. E G. having
w coordinates bounded away from 0, a neighborhood of z. in G. is CO close to
WU(SR). To use transversality information from Lemma 2, it is necessary to have
G. be Cl close to WU(SR). That result is the following deeper fact, which is a
consequence of the exchange lemma.
 107

SLOW MANIFOLD ' \

G E n FACE OF dB

Figure 4. A curve f. transverse to W 8CSR x 16) in u,v,w,8

space. The 8 direction is suppressed in this figure. The 8 com-
ponent varies strongly along f •. However, it is a consequence
of Lemma 4 that the image curve on G. n 2B has almost no
dependence on 8.

LEMMA 4. lfy. is sufficiently close to W·, then there is a neighborhood N(z.)

so that G. n N(z.) is Cl close to W"(SR).

Note that the entrance curve f. may have points with a range of values of 8.
Nevertheless, the C 1 closeness of G. to l¥"(S R) at the exit set implies that the exit
curve has a tangent vector whose 8 component is almost zero, i.e. the 8 information
has been washed away.

°
The proof of the theorem can now be given. As mentioned above, we follow
around we" for f small. If f is small enough, we" lies close to the set w =
until the trajectory for 8 = 8. enters a neighborhood of SR' It therefore lies close to
°
Weu(O, 0, 8.), the f = singular object in w = 0. Lemma 1 states that weu(o, 0, 8.)
is transverse to the two-dimensional we8(1, 0, 8.) in w = 0. It is an easy corollary
that WeuCO, 0, 8.) is transverse to the three dimensional W 8CS R x 16) in (u, v, w, 8)
space (see Figure 5). Thus we may conclude that weu is transverse to W8(SR x 16)
at a point in f. == W eu naB.
We can now apply Lemmas 3 and 4. It follows from Lemma 3 that for any
value w near w., there are points entering B on f. that exit at height w near the
singular object W eu ( u_, 0, w.). Furthermore, by Lemma 4, at any such exit point
z., the tangent plane to l¥eu is close to that of the singular object W"CSR) or,
 108

Figure 5. A schematic drawing of the intersection of the two-

dimensional manifold Weu(O, 0, e.) = UOEI. WU(O, 0, e) with
the three dimensional manifold WS(SR x [6). The e direc-
tion has been suppressed by projecting onto (u, v, w) space.
For e;:::;j e., the projection is approximately WS(SR x {e.}).

equivalently, to that of weu(u_,O,w.).

We may now use Lemma 2 to follow W eu across the back. By Lemma 2,
Weu(u_,O,w.) has a transverse intersection with TVes(u+,O,w*) in (u,v,e) space;
it follows that
WCU(u_,O,w.) has a transverse intersection in (u,v,w,e) space with the three-
dimensional manifold WS(S L x [6). (The latter is defined analogously to WS(S R x
[6) to be the singular object that is the union over -b S wSw. + b, e E [6, of the
stable manifolds associated to points on SL at € = 0.) The Cl_ closeness of weu to
Weu(u_,O,w.) implies that it too has a transverse intersection with WS(SR x [6).
The final observations concern the center stable manifold W es of the point
(u,v,w) = (0,0,0) for (1.2). If € =J 0, for each e E [6, the center stable mani-
fold of this point is two-dimensional (one fast direction and one slow), so W es is a
°
three-dimensional manifold in (u, v, w, e) space. For € =J small, W es lies close to
WS(SL x [6), at least in a neighborhood ofSL. Thus we may conclude the transver-
sality of We" and weB. Since the transverse intersection of a two-dimensional man-
ifold and a three-dimensional manifold in four-space is a one-dimensional curve,
e
this is the trajectory we seek. is a constant along trajectories, so the intersection
e
defines the wave speed of the arc.

3. The Exchange Lemma. We shall now sketch ideas behind the mam
technical tool of the proof, the exchange lemma. Although this lemma has been
proved in arbitrary dimensions [13], we discuss it here in the lowest interesting
dimensions, which is all that is needed for the Fitzhugh-N agumo equations.
 109

The exchange lemma uses special coordinate systems adapted to a neighborhood

of SL or SR. At e = 0, for fixed 0, it is easy to use the hyperbolic structure of the
rest points to get a coordinate system (a,b) in which the equations have the form

a' = >.(a, b, w)a

(3.1) b' = Il(a, b, w)b
w' =0

where a, b, eRl, >. > 0 and Il < O. w e R2 and represents the (w, B) coordinates. For
any 6 > 0 sufficiently small, there are >'0, Ilo such that

(3.2) >.(a, b, w) ~ >'0 > 0, Il(a, b, w) ::::: Ilo < 0

for all B == {a, b, w : lal < 6,Ibl < 8, w in a given compact region}. For e f= 0, there
are analogues SL,< and SR,< of the manifolds SL and SR. These are constructed
(nonuniquely) as center manifolds of (1.2). On them, the vector fields have size
O(e), so SL,< and SR,< are known as "slow manifolds." It was shown by Fenichel
[16] that in the neighborhood B of SL or SR, there is an analogue of the coordinate
system (3.1) adapted to the hyperbolic structure that persists, with the stable and
unstable coordinates vanishing on S L,< or S R,<' In this coordinate system, which
we shall call "Fenichel coordinates," the equations near SL,< or SR" take the form

a' = ,\(a,b,w,e)a
(3.3) b' = Ilea, b, w, e)b
w' = eg(a, b, w, e)

We may assume that 9 = (gl, g2) satisfies

gl(O,O,Wl,W2,O) ~ c> 0
(3.4)
g2(O,O,Wl,W2,O) =0
That is, the w coordinates are chosen so that W2 vanishes along the limiting
trajectory of the slow flow as e = O. \Ve also assume that the box B has the
form wi"" < WI < wi; IW21, lal, Ibl < 8, and that at e = 0, the critical points
a = 0, b = 0, wi"" < WI < wi, IW21 < 8 form a normally hyperbolic manifold.
These coordinates, also used in [11], simplify the calculations we shall do.
Let M be some two-dimensional invariant manifold that has an intersection with
the manifold {b = 6} at some point q. As we shall show, the trajectory through q,
which lies in M by invariance, enters the interior of B. If the a coordinate of q is
in an appropriate range, the trajectory through q remains in B for a time that is
O(l/e) and leaves at a point 'ij in a face of the form lal = 6. We shall be concerned
with the relationship between the tangent plane to A1 at q and that of the tangent
plane to M at 'ij.
The exchange lemma says that under appropriate restrictions on the tangent
plane at q, including transversali ty with respect to the space {a = O}, the tangent
 110

The exchange lemma uses special coordinate systems adapted to a neighborhood

of SL or SR. At I' = 0, for fixed 8, it is easy to use the hyperbolic structure of the
rest points to get a coordinate system (a,b) in which the equations have the form

a' = A(a, b, w)a

(3.1) b' = Jl( a, b, w)b
w' =0

where a, b, ERl, A > 0 and Jl < O. W ER2 and represents the (w, 8) coordinates. For
any 8 > 0 sufficiently small, there are AO, Jlo such that

(3.2) A( a, b, w) ~ AO > 0, Jl( a, b, w) ::; Jlo <0

for all B == {a, b, w : lal < 8, Ibl < 8, w in a given compact region}. For I' i= 0, there
are analogues SL,E and SR,E of the manifolds SL and SR' These are constructed
(nonuniquely) as center manifolds of (1.2). On them, the vector fields have size
0(1'), so SL,E and SR,E are known as "slow manifolds." It was shown by Fenichel
[16] that in the neighborhood B of SL or SR, there is an analogue of the coordinate
system (3.1) adapted to the hyperbolic structure that persists, with the stable and
unstable coordinates vanishing on S L,E or S R,E' In this coordinate system, which
we shall call "Fenichel coordinates," the equations near SL,E or SR" take the form

a' = A(a,b,w,E)a
(3.3) b' = Jl(a, b, w, €)b
w' = Eg(a,b,w,E)

We may assume that 9 = (gl, g2) satisfies

gl(0,O,Wl,W2,0) c> ~ 0
(3.4)
g2(0,0,Wl,W2,0) = 0

That is, the w coordinates are chosen so that W2 vanishes along the limiting
trajectory of the slow flow as I' = O. We also assume that the box B has the
form w 1 < WI < wt; IW21, lal, Ibl < 5, and that at I' = 0, the critical points
a = 0, b = 0, wi" < WI < wt, IW21 < 5 form a normally hyperbolic manifold.
These coordinates, also used in [11], simplify the calculations we shall do.
Let M be some two-dimensional invariant manifold that has an intersection with
the manifold {b = 5} at some point q. As we shall show, the trajectory through q,
which lies in M by invariance, enters the interior of B. If the a coordinate of q is
in an appropriate range, the trajectory through q remains in B for a time that is
0(1/1') and leaves at a point q in a face of the form lal = 8. We shall be concerned
with the relationship between the tangent plane to M at q and that of the tangent
plane to M at q.
The exchange lemma says that under appropriate restrictions on the tangent
plane at q, including transversality with respect to the space {a = OJ, the tangent
 111

plane at q is C I close to the plane spanned by the a and WI axes. Let ,=Mn{b = o}
be an open arc, and,? a subset of I such that points on,? have trajectories that exit
B through {a = o}. The exchange lemma implies that the map W: '? --4 {a = o}
is smooth, and that the tangent vector to the image W(,?) at a point q is C I close
to the WI direction. In particular, information about transversality with respect
to {a = O} at the entry to B is exchanged for information about transversality
with respect to WI' In some applications, including FitzHugh-Nagumo, knowledge
about transversality with respect to {a = o} at the entry comes from information
about how the limiting unstable spaces WU(p) vary as the other slow variable W2
is changed. Thus information about behavior as W2 is varied is exchanged for
information about the behavior of WI.
It is not hard to establish that, while a trajectory is in B, its b coordinate
decreases exponentially and its a coordinate increases exponentially [13]. Also, the
time from entrance to exit is O(I/E) for any point whose WI coordinate increases by
0(1). This quickly implies that if V is a sufficiently small neighborhood of q in ME,
then the image of V at the exit of B is Co close to {W2 = 0, b = O}.
It is more subtle to establish that the image is C I close to the above two-
dimensional plane, since it is hard to see the behavior of planes by following indi-
vidual vectors. The difficulty is that every trajectory, when exiting from B, does so
approximately tangent to the expanding subspace parameterized by the a variable.
Thus, the study of individual tangent vectors does not pick up the part of the tan-
gent plane to M in the center direction. To do this, one must directly follow how
the flow induces a vector field on tangent planes. This flow is computed from the
variational flow on y == (a, b, WI , W2), which is

oa' = >'(y, E)oa + a(V>'· oy)

ob' = /1(Y, E )ob + b(V /1 . oy)
(3.5)
ow~ = EV91 . oy
ow~ = EV 92 . oy.
We may think of each of the above coordinates oa, ob, OWl, OW2 as one-forms acting
on the vector fields, and the variational equations are evolution equations for the
one-forms. A plane is represented by a two-vector, which is an element of the second
exterior power A2R4 ~ R 6 • The coordinates of this vector in R 6 can be thought
of as two-forms. To follow planes, we derive evolution equations for the two-forms.
The space of two-forms is spanned by the six two-forms Pa(3 == oc< 1\ 0(3, where
c<, (3 = a, b, WI, W2. (There are only six independent such ones since oc< 1\ oc< = 0 and
oc< 1\ 0(3 = -0(31\ oc<.) The evolution equations for the two-forms can be computed
from (3.5) using the product rule:
 112

= )..Paw, + aRI + fR2

P~w,
= )"PaW2 + aR3 + fR4
P~W2
P~b = [).. + I1]Pab + aRs + bR6
(3.6)
Phw , = I1Pbw, + bR7 + fRS
PhW2 = I1Pbw2 + bRg + fRIO
P~'W2 = E(Rll + R 12 ),

where

RI = (V),,· 8y) 1\ 8WI R7 = (VI1' 8y) 1\ 8WI

R2 = 8a 1\ (Vg I .8y) Rs = 8b 1\ (V gl . 8y)
R3 = (V),,· 8y) 1\ 8W2 Rg = (VI1' 8y) 1\ 8W2
R4 = 8a 1\ (Vg2 ·8y) RIO = 8b 1\ (Vg 2 ·8y)
Rs = (V),,· 8y) 1\ 8b Rll = 8WI 1\ (VgI ·8y)
R6 = 8a 1\ (VI1' 8y) Rl2 = (Vg 2 . 8y) 1\ 8W2

It is useful to do the calculations in P 1\2 R4, projective space of two-forms in

R 4 . The reason is that the values of the two-forms acting on a plane are determined
only up to normalization. Also, by (3.6), values of some of the two-forms grow expo-
nentially in time. We can use the fastest growing two-form to normalize. In a region
in which Pa,w, # 0, {POI (3/ Paw,} form local coordinates, with a, {3 = a, b, WI, W2 and
(oo,{3) # (a,wt) or (wI,a). Let POI (3 = POI (3/Paw , and R; = R;jPaw ,. The {R;} can
be written in terms of the {POI (3} and a nonhomogeneous term (time dependent).
In these coordinates, the plane spanned by the directions a and WI is specified by
P01(3 = O. A plane with P01(3 = O( E) is C l close to this plane. We shall sketch the
proof of:

EXCHANGE LEMMA: Let M be a two-dimensional invariant manifold transverse

to {b = 8}. Assume tlJat r = M n {b = 8} is transverse to {a = OJ. Let q ill r be
a point whose trajectOIy exits from {a = 8} after a time T that is O(l/E). Then at
the exit point q, Pa (3 is O( E) for all a, {3.

Sketch: The evolution equations for the {P0I(3} are

P~W2 = Paw2 ( -Q) + aR3 + fR4

P~b = (11 - Q)Pab + aRs + b~
(3.7) Pt,w, = (11 - ).. - Q)Pbw, + bR7 + ERS
Pt w2 = (It - ).. - Q)PbW2 + bR g + fRIO
PW,W2 = (-).. - Q)PW,W2 + f(Rll + R I2 ),
" ' ,

where Q = aR I + ER2 •
One first shows that all POI (3 decay exponentially to size O(f), provided that Paw2
satisfies some a-priori bound. This follows (with some work) from the fact that the
 113

lower R.H. 4x4 matrix of (3.7) is almost diagonal, with negative diagonal elements.
The less intuitive part of the proof is to show that Paw2 must have started small,
and stays smal1. (It therefore satisfies the a-priori estimates.)
Consider the initial conditions for the {po~}, including Paw2' The tangent plane
to M at q is spanned by the vector y' = (a', b' , w~ , w~) at q and the tangent vector
df to M n {b = 8}. In the first vector, (l/c)dwJ/dT ;::: C > O. db/dT is 0(1). By
hypothesis, the trajectory through q remains inside B for a time that is O( 1/ c).
Since a increases exponentially, at q the value of a is exponentially small; hence
so is da/dT. By the transversality hypothesis on f, the first coordinate of df is
bounded away from zero independent of Eo The second is zero by definition of f.
The value of the two-form Po~ is computed up to a normalization factor by taking
the 2x2 determinant of the a, (J components of the above two vectors. (See Section
4 for more details.) From that, we obtain that PaWl is O( E), with its quotient by
c bounded away from zero. Paw2 is exponentially small, and the other two-forms,
evaluated on TqM, have 0(1) bounds. It follows that Paw2 is also exponentially
small in c.
It remains to show that Paw2 stays 0(1). This uses the first equation of (3.7)
and information about R3 and R4 to control the nonhomogeneous term aR 3 + €R 4 .
The technical lemmas and estimates are in [13).

4. Proofs of Transversality. The proofs of lemmas 1 and 2 can also be

achieved through the use offorms. Consider first Lemma 1. We abbreviate vV eu ( 0,0,8.:
and VV es (1,0,8.) by vV eu and w es respectively, warning the reader not to confuse I
°
these with the € 1= objects so denoted in the previous sections. We need to show
that the tangent spaces to lV eu and H1 es intersect transversely in (u, v, 8) space. vVe
calculate the 2-forms which give us the coordinates associated with these tangent
planes; it suffices to show that the vectors made out of these coordinates are linearly
independent.
The variational equations are
8u' = 8v
(4.1) 8v' = 88v - f'(u)8u + v88
88' = 0.

The equations for the 2-forms are calculated as in the previous section with P uv =
8u 1\ 8v etc. The relevant equation is the one for Puv:

(4.2)

vVe are interested in the Puv and Puo associated with the tangent planes to vVcu
and W es , which we shall denote P;;" and Pto ( + for unstable and - for stable). We
shall show that P;;o and p;Jv have the same sign, whereas P;:o and P;;;, have opposite
signs. This will imply that the vectors (Puv , Puo, P vo ) for the tangent planes to lVcS
and vVcu are linearly independent and hence that the planes are transverse. 'IVe
explicitly calculate Pto and then obtain the sign of P;;" from (4.2).
We must first be explicit about how to associate a number to a 2-form acting
on a plane. The plane is specified by giving an orthonormal pair of tangent vectors
 114

whose span determines an oriented unit square in the plane. Consider a 2-form
bal /\ ba2, where the aj vary over u, v, B. The value of bal /\ ba2 on the plane is
given, up to sign, by the area of the projection of the above square on the aj, a2
plane along a3, where a3 f:. a}, a2. The sign is positive if the orientation of the
projected square, induced by the projection, agrees with the orientation of the aj, a2
plane induced from the ordering of aI, a2.
Now we return to wes and weu . Both of these planes have the vector field
(v, ()v - f( u), 0) as a tangent vector. Let (bu±, 8v±, 1) be another tangent vector
for w eu and w es respectively (+ for unstable and - for stable). Note that we can
take 8B = 1; since bB' = 0, this guarantees that the new vector is not a multiple
of the vector field. The above vectors are not orthonormal, but they can be made
orthonormal by normalization and a Gramm-Schmidt process. The quantities Puv ,
etc., for each plane are thus equal, up to a normalization factor, to the 2 x 2 subdeter-
minants of the 2 x 3 matrix whose rows are the above linearly independent vectors in
(u, v, B) space spanning the plane. The normalization factor N = N( 1/., v, b1/.±, 8v±)
is positive and comes from the orthonormalization procedure. Thus

~I =Nv
Since v > 0 along the front, for the tangent planes to both weu and vV es , we then
have that P::U > O. Moreover (4.2) now reads

(4.3)

vv eu and wes each contain a line of critical points whose tangent vector is in the B
direction. For any plane containing such a line, the 2-form Puv vanishes. It follows
e e
that P;;v --t 0 as --t -00 and P;;v --t 0 as --t +00. Equation (4.3) then implies
that for W eu , P;;v > 0 and for W es , P;;v < o. This proves that these vectors
associated with the respective tangent planes are linearly independent, as desired.
D
The proof of Lemma 2 is similar. The equations are now
8u' = 8v
bv' = B.bv - f'(u)bu + 8w
bw' =0
The relevant equation on 2-forms is

P~v = B.Puv + Puw .

Also, Puw = Nv, where N is the analogous normalization factor. Since v is of one
sign (v < 0) along the back, a similar argument holds. D

Acknowledgment.
The authors are grateful to J. Alexander for a number of discussions that clari-
fied the use of forms. We thank the Institute for Mathematics and its Applications
for its hospitality while this paper was shaped.
 115

REFERENCES

[1] CONLEY, C., On travelling wave solutions of nonlinear diffusion equations in dynamical
systems, theory and applications, in Springer Lecture Notes in Physics.
[2] CARPENTER, G., A geometric approach to singular perturbation problems with applications
to nerve impulse equations, JOE, 23 (1977), 335-367.
[3] HASTINGS, S., On travelling wave solutions of the Hodgkin-Huxley equations, Arch. Rat.
Mech. Anal., 60 (1976), 229-257.
[4] GARDNER, R.H. AND SMOLLER, J., Travelling wave solutions of predator-prey systems with
singularly perturbed diffusion, JOE, 47 (1983), 133-16l.
[5] FIFE, P .C., Boundary and interior transition layer phenomena for pairs of second order
differential equations, J. Math. Anal. Appl., 54 (1976), 497-52l.
[6] FUJII, H., NISHIURA, Y., AND HOSONO, Y., On the structure of multiple existence of stable
stationary solutions in systems of reaction-diffusion equations, Studies in Mathematics and
its Applications, 18 (1986), 157-219.
[7] HALE, J. AND SAKAMOTO, K., Existence and stability of transition layers, Japan J. Appl.
Math, 5 (1988), 367-405.
[8] SAKAMOTO, K., Invariant manifolds in singular perturbation problems for ordinary differen-
tial equations, preprint.
[9] CHANG, K.W. AND HOWES, F.A., Nonlinear Singular Perturbation Phenomena: Theory and
Applications, Springer Verlag, N.Y. 1984.
[10] DIENER, F. AND REEB, G., Analyse Non Standard, Hermann, Editeurs des Sciences et des
Arts, PARIS, 1989.
[11] LANGER, R., Existence and uniqueness of pulse solutions to the Fitzhugh-Nagumo equations,
Ph.D. Thesis, Northeastern University (1980).
[12] JONES, C.K.R.T., Stability of the travelling wave solutions of the Fitzhugh-Nagumo system,
Trans. AMS, 286, #2 (1984), 431-469.
[13] JONES, C.K.R.T. AND KOPELL, N., TI'acking invariant manifolds of singularly perturbed
systems using differential forms, in preparation.
[14] PALlS, J. AND DE MELO, W., Geometric Theory of Dynamical Systems, Springer-Verlag,
N.Y. 1980.
[15] ARONSON, D. AND WEINBERGER, H., Nonlinear diffusions in population genetics, combustion
and nerve propagation, in Partial Differential Equations and Related Topics, J. Goldstein,
ed., Lecture Notes in Math #446, Springer Verlag 1975.
[16] FENICHEL, N., Geometric singular perturbation theory for ordinary differential equations, J.
Dilf. Equa., 31 (1979), pp. 53-98.
 KINETIC POLYNOMIAL:
A NEW CONCEPT OF CHEMICAL KINETICS

MARK Z. LAZMAN* AND GREGORY S. YABLONSKII**

Abstract. A system of quasi-steady-state equations for a single pathway mechanism of a

catalytic reaction can always be reduced to a polynomial in terms of the steady state reaction
rate, a kinetic polynomial. The coefficients of this polynomial are polynomials in the parameters
of the elementary reaction rates. The form of the lowest coefficient of the polynomial ensures
the thermodynamic validity of this form of representation of quasi-steady-state equations. The
properties of the kinetic polynomial are discussed in connection with such concepts of chemical
kinetics as "molecularity", "stoichiometric number" .
Possible applications of this form are: asymptotic analysis of steady-state kinetic models as well
as analysis of steady-state multiplicity; description of the steady-state dependences of the reaction
rate, determining relations between kinetic constants when solving the inverse problem.
On the basis of kinetic polynomial explicit equations for the steady-state rate in case when one
of the steps is rate-limiting, and in the neighbourhood of equilibrium have been derived.
Algorithm of computation of the kinetic polynomial and its realisation on the basis of computer
algebra are described.

Key words. steady-state, catalytic reaction.

AMS(MOS) subject classifications.

1. Introduction. The general description of catalytic reaction is given in the

theory of steady reaction (Horiuti-Temkin [1,2]). The mass (surface)-action-Iaw
(MAL) is assumed for the rate of elementary reaction in this theory. According to
the steady reaction theory the system of quasi-steady-state equations is of the form

P
(1.1) LII~P)W(P)=U(s), s=1, ... ,8; p=l, ... ,P,
p=l

where u(s) = Us - u- s , Us, U- s are the rates of s-th stage and its forward and back
reactions; II~P) is the stoichiometric number of the sth step for pth independent
route 1 ; w(p) is the rate along the pth route; 8 is the number of steps; P = 8 - J is
the number of reaction routes (J is the number of independent intermediates).
An important problem concerned with the steady reaction theory was the de-
pendence between the kinetic equations of forward and back reactions or more
generally the connection between thermodynamics and kinetics for complex reac-
tions. The case with a rate-determining step has been studied by Boreskov [4]
and, independently, by Horiuti [1,3]. They have established the dependence of the

*USSR Research and Design Institute of Oil Refining and Petrochemical Industry (VNIP-
INEFT), Moscow 113095, Dimitrova 33/13, USSR.
**Tuva Complex Department of Siberian Branch USSR Academy of Sciences, Kyzyl 667000,
Lenina 30, USSR.
1 This concept was introduced by Horiuti [3]. The set of numbers v~p) provides that adding up
the steps of complex reaction multiplied by respective v~p) results in a net chemical equation free
from intermediates.
 118

steady-state reaction rate w on the rate of forward reaction w+ and thermodynamic

characteristics of the overall reaction:

Here !::.G is free energy variations; Keq is the equilibrium constant of the over-
all reaction; f+(c) and f-(c) are the products of concentrations written according
to the MAL for net stoichiometric equation both forward and backward; c is the
reactant concentration vector; M is the reaction molecularity and I/L is the stoi-
chiometric number of the rate-determining step (M = 1/ L1).
Near equilibrium the rate of a simple reaction is linearly dependent on the free
energy variations [5J. Using this fact Nacamura [6J has obtained for a single-route
reaction the following equation

(1.3)

where 1/. and u. eq are the stoichiometric number and the equilibrium rate of ex-
change for the 8th step.
Formulae of types (1.2) and (1.3) are of importance in the kinetics of chemical
reactions [7J. They were used to study reaction mechanisms for N H3 synthesis,
S02 oxidation, etc. [8J. In many publications the possibility has been discussed for
the extension of their applicability [9-12J.
Further development of steady reaction theory was based on the graph the-
ory that had been applied initially in enzyme kinetics [13J. A structured form of
steady-state kinetic equations based on graph representation of catalytic reactions
with linear mechanism 2 was found and investigated in [14J. This form enables writ-
ing explicitly a steady-state kinetic equation on the basis of a detailed mechanism.
However, the main advantage of such forms is the possibility of obtaining physico-
chemically meaningful corollaries rather than their compactness.
For instance, for a linear single pathway mechanism the steady-state reaction
rate w is represented the following equation [14J:

(1.4)

where b. and L. are the weights of the forward and back reactions, i.e. their rates
at unit concentrations of the intermediates 3 , and Dx is the weight of the framework
(see [14]), i.e. the product of the weights of the component arcs.

2 Only one intermediate reacts in each elementary step of these reaction mechanism, for example,

(Zl, Z2, Z3 are intermediates, A is reactant, and B is product).

3for mechanism in previous footnote the weights are: b1 =
k1 , A, b_ 1 = k_ 1 , ... , b3 = k3 , b_ 3 =
k_ 3 CB; where ki are the rate constants of corresponding steps.
 119

The numerator in (1.4) corresponds to the kinetic equation of the overall reac-
tion assuming that it is a step for which the MAL is valid [14]. One can transform
it to the form k+f+(c) - Lf-(c), thus the equation (1.4) corresponds the equation
(1.2) at M = 1.
From Eq. (1.4) it follows that at w = 0 (equilibrium)

s s
(1.5) II b. = II L. = k+f+(c) - k-f-(c) or Keq = f-(c)/J+(c).
8=1 8=1

Eq. (1.5) is valid obviously for the kinetic equations (1.2) and (1.3).
We'll consider in this paper the general situation of nonlinear reaction mecha-
nism. The first problem here is the form of kinetic equation. The main reason that
there were no structured forms like (1.4) for nonlinear mechanisms in the literature
has apparently been the impossibility of solving explicitly a system of quasi-steady-
state nonlinear equations (1.1). However, it is always possible to apply to this
system of equations a method of elimination of variables and reduce it to a poly-
nomial in one variable, e.g. a polynomial in terms of the steady-state reaction rate
w. The polynomial coefficients are, in turn, polynomials in terms of the reaction
rate constants (k. and k_.) and reactant concentrations (c). Reducing systems of
algebraic equations to polynomials is common practice in higher algebra (e.g. see
the book [15]). In elimination theory [16J this reduction actually constitutes the
solution of the systcm.
We refer to a polynomial in the steady-state reaction rate as a kinetic polynomial
[17-23]. In present work we'll study the properties of the kinetic polynomial for a
nonlinear single pathway mechanism of a catalytic reaction with a single type of
active sites. Note that as a steady-state kinetic equation for a linear single pathway
mechanism and as equations (1.2) and (1.3) for the limiting cases of rate-determining
step and neighbourhood of equilibrium this polynomial must satisfy an "equilibrium
test", i.e. at w = 0 we must have (1.5). On the basis of kinetic polynomial we'll
consider classical concepts in the kinetics of complex reactions: "stoichiometric
number", "molecularity", "limitation", "equilibrium" and refine their contents and
the range of applicability. Then the possibilities of kinetic polynomial application
for the solving of direct and inverse kinetic problems as well as the algorithm of
computation of this polynomial and it realisation via the computer algebra will be
discussed.

2. Theory. The following nonlinear single pathway mechanism of a heteroge-

neous catalytic reaction is considered:

(2.1)

where Zj and C r are jth surface and rth observed substances having concentrations
Zj and C r respectively; (Xij, {Jij, ')'! .. and ')'!~ are their stoichiometric coefficients at

ith step in the forward and back directions; and k i and k-i are the rate constants.
 120

I I I "
The reaction weights are bi = ki TIr=1 c;';r and Li = k-i TIr=1 c;';r. Stoichiometric
coefficients of the surface substances satisfy the following restrictions:
n
(2.2) ~)O!i;-f3ij)=O, i=l, ... ,n
;=1
i.e. the total number of molecules of surface substances remains unchanged at every
step4. Besides, we'll consider mechanisms for only one type of active site. According
to the steady-state reaction theory (see Eq. (1.1», the system of quasi-steady-state
equations for mechanism (2.1) is of the form

II z~lj
n n
(2.3.1) bI II Z Jalj - b- I J = VIW,
;=1 j=1
n n
(2.3.2) b2 II za 2j
J - b-2 II z~2j J = V2W,
;=1 j=1

II
n n
(2.3.n) bn II Zjanj - b-n ZjPnj = VnW,
j=1 j=1
n
(2.3.n+1) LZj = 1;
;=1
where Vi is the stoichiometric number of the ith step (here the set Vi can be chosen
arbitrary, provided that adding up the steps multiplied by respective Vi results in a
net equation free from intermediates), W is the steady-state reaction rate.
Let VI of. O. Let us split system (2.3) into Eq. (2.3.1) and the subsystem (2.3.2)-
(2.3.n+1). Suppose that at a fixed value of W the subsystem has a finite number of
solutions (ZI(r)(W), ... , Zn(r)(w)} with r = 1, ... , M 5 ; the solutions are taken with
their multiplicities. For a further analysis we need the following function

(2.4) R(w) ~ 11 (,' f1 ,!,';,<w) - '-, f1 ,fi;,(w) - v,w) .

R(w) is the resultant (see [24)) of Eq. (2.3.1) with respect to subsystem (2.3.2)-
(2.3.n+1). (Definition of resultant in the classic theory of elimination is given in
[15,16)). The resultant of a system of algebraic equations is a rational function
(i.e. polynomial or a ratio of two polynomials) [24], but in our case function R(w)
defined by formulae (2.4) is namely polynomial (for proof see works [22,23)):

(2.5)
'Restrictions (2.2) correspond to mechanisms (2.1) where each surface substance contains the
same number of active sites. The case most commonly encountered is a surface substance contain-
ing only one active site.
sit was proved that if w = 0 is not a root of system (2.3) and VI f:. 0 then subsystem like
(2.3.2)-(2.3.n+1) has only discrete roots in en and has no roots at infinite hyperplane [22,23].
This assumption means that system (2.3) has no boundary steady states (see below).
 121

The resultant R( w) is zero for these and only these values of w which are solutions
to system (2.3).
The function R( w) is a kinetic polynomial. Its coefficients B o , . .. , B L can be
written in the form of polynomials in bi and L j . We will write Bi ~ a if B;/a =1= 0
when b±l,' .. ,b±n =1= O. Consider now the properties of B i .

THEOREM 2.1.

(2.6)

where stoichiometric numbers 111, .•. ,lin have no greatest common divisor greater
than 1 and p is positive integer.

Proof. It follows from (2.4) and (2.5) that

(2.7)

The solutions zr(O) of subsystem (2.3.2)-(2.3.n+1) for w = 0 may be interior and

boundary value. An interior root is such that all Zj(r) are non zero. Boundary
solutions exist if Z j( r) are zero for certain j. It is easy to show (see [23]) that when
IIk+1,' •• ,lin = 0 and 111 =1= 0, ... ,Ilk =1= 0 we must take into account only such roots
zr(O) that have k or more non-zero coordinates.
Consider at first interior roots. After substitution Zj(O) = pje iNj , with ~2 = -1
we'll get

II p;'j-i32 j
n-1
= (Ldb2)p~2n-a2n
j=l
(2.8.1 )

II p;n j -i3nj = (Ln/bn )p~nn -ann

n-1

j=l

n-1

2) <l:2j - fJ2j )¢j = (fJ2n - <l:2n)¢n + 27rm2

j=l
(2.8.2)
n-1

2)<l:nj - fJnj)¢j = (fJnn - <l:nn)¢n + 27rm n

j=l

n
LPj COS¢j = 1
j=l
(2.8.3) n
2:Pj sin¢j = 0
j=l
 122

with mj = 0, ±1, ±2, ... (the mj must be chosen so as to obtain all possible solutions
(z;(O))). Taking logs in (2.8.1) we get

(2.9) log p, ~ (t, Aj,(log(Lj/bj ) + (Pj. - OJ.) 10gP.)) / '"

where
(a21 ~ ;32d
~1 = det ( :
(anI - ;3nl)
and Aji are the cofactors of the elements in ~l with the corresponding indices.
If ~1 = 0, which is possible, for instance, in the presence of "buffer" steps, an-
other subsystem can be chosen. All ~i cannot be zero, for in our instance the
stoichiometric matrix is of rank n - 1). Expression (2.9) can be transformed thus:

(2.10)

From (2.2) we have

n-l
I: (a21 - ;32,)
1=1

n-l
I: (ani - ;3nl)
1=1

This determinant is obtained from ~1 by replacing the ith column with the sum of
all columns in ~l and is equal to ~1. Then we have
n
(2.11) Pi = pn II(L j /b j )Aj;/oC. 1 •

j=2

Similarly, from (2.8.2)

n
(2.12) tPi = tPn + (21f / ~d L mjA ji .
j=2

From (2.11) and (2.12) for i = 1, ... , n - 1 we find that

(2.13)

Let us return to relation (2.7). It is obvious that

II II zjr':)(O)q,',
Min n

(2.14) Eo ~ q, = b~{"
r=lj=l
 123

where

(2.15)

Here Min is the number of interior solutions of the subsystem. It was proved [23]
on the basis of results [25] that in our case

(2.16)

and these roots are simple. Substituting (2.13) into (2.15), we obtain
(2.17)

\lI' = g(b~1 C~2 )'~2/Al ... Cbnn ) An/A, exp(27r~(m26,2 + ... m n6,n)/6,d -1)
where 6,i, i = 2, ... ,n, are the determinants obtained from 6,1 by replacing row i
with the row ((0'11 - (311) ... (al,n-l - (31,n-d). We consider the sum

(2.18)

On the other hand

(aU ~ (31j) (al,n-l ~ (31,n-d) .

(2.19) S = det ( .
(anj - (3nj) (a n,n-l - (3n,n-d
Then the cofactors of the elements in the first row are 6,1, -6,2, ... ,-6,n. However,
determinant (2.19) is zero as it contains the same columns for join and is the nth-
order determinant of a matrix of rank n -1 for j = n. Consequently, S is identically
zero, and 6,1, -6,2, ... , -6,n can be considered as the stoichiometric numbers of the
corresponding steps in mechanism (2.1). Equation (2.17) contains 6,i divided by
6,1. Reducing (m26,2 + .. . m26,n)/6,1 in (2.17) by the greatest common divisor,
we obtain

(2.20)

The number set {m2, ... , m n } must contain a certain minimum set that yields
different values of exp( -27r~ 2::7=2 m;v;/vd i.e. all different values of remainders
from division - 2::7=2 m;v;/vl. Let VI exceed zero. There exist such set ml, . .. ,mn,
that -(mlvl + ... + mnv n ) = 1 because VI, ... , Vn have no common divisor other
then unity. Then, for obtaining all possible remainders it's sufficient to multiply
m 1, ... ,mn by VI - I, where I = 1, ... , VI.
It follows from (2.17), (2.20) for this way obtained sets {m} that
1

II ((bd L d(bd L 2t,fv, ... (bn/Lnt

Vl-

\lI' "-' n / V, exp(27r~s/vd - 1)

8=0
(2.21)
 124

Thus, we have proved that

(2.22)

(about degree p see below).

Let now Vn = o. Consider the roots (if they are exist) that have one zero
coordinate, for instance, Zl "I- 0, ... , Zn-l "I- 0 and Zn = o. If the product nj=l zi'i
or the product nj=l zt'i includes Zn then these roots doesn't influence the result
(2.22). If both these products doesn't include Zn then assuming Zn = 0 we get
the (n - I)-dimensional system like (2.8) and relation (2.22) is fulfilled in this case
too. 0

COROLLARY 2.1. If VI "I- 0, ... , Vn "I- 0 then the degree pin (2.6) is the greatest
common divisor of 6i : p = 16t!Vll = ... = 16 n /vn l.

Proof. Number of interior roots is 16 1 1 and IVll of these roots gives one multi-
plier (n~l(b;/Liyi -1)6.0

COROLLARY 2.2. If degree p > 1 and VI, ••• , Vn "I- 0 then coefficient B. of R( w)
includes the multiplier (the cyclic characteristic) C = (br' ... b~n - b~l ... b~~) with
degree p - s, s = 0, 1, ... , p - 1; p ::; L, where L is degree of kinetic polynomial.

Proof. It follows from (2.4) and (2.5) that first coefficient of R( w) is

(2.23)
j i;ij

where Ul(Z(j)(O» is the value ofthe rate oflst step in the solution (Zl(j)(O), ... , zn(j)(O);
We can see from Theorem 2.1 and Corollary 2.1 that ILlll interior roots of ).
subsystem can be divided to p groups contained VI elements and each one produces
multiplier C in the product (2.7).
Thus ni;ij Ul (Z(i) (0» ,..., Cp-l. For the second derivative the degree of C is
two units smaller and so on. The degree of kinetic polynomial L ~ maxi ILl i I ~
16;/Vil = p, so multiplier C doesn't contained in B L · 0
Consider now the form of the other coefficients of kinetic polynomial. It follows
from (2.5) that

(2.24)
Bi _ _ 1_di R(w)
- I () .
I , i=I, ... ,L,
Bo i.R 0 dw' w=O

where R(w) is defined by (2.4). The recursive form of (2.24) is

(2.25) Bi =.! ~ .!. Bi-l-k dk+lln R( w) I

Bo i L...J k!
k=O
Bo dWk+l _
w-o
6 If, for instance, lin =0 and there are exist the roots withzn = 0 and n;;l z;'i of. 0 and
n;;l z:'i of. 0 then the number of these roots is I~tl/lann -/3nnl. Thus the degree p increased
by l~d{II1 {ann -/3nn))I (for details see [23]).
 125

It follows, in particular, from (2.25) that

(2.26) E1 = dlnR(w) I '

Eo dw w=o

and

(2.27) E2 = .! ((E1)2 + d2In~(w) I ).

Eo 2 Eo dw w=o

Thus to find the coefficients of (2.5) we must calculate derivatives of In R( w).

Let's prove that

dlnR(w) n M. 1
(2.28) dw = - 2:>k
k=l Jk=l
L
-U-k":"(Z-j-k":"(w-)"")---V-k-W

where Uk(Zjk(W)) is the rate of kth step calculated in the solution of subsystem
obtained from (2.3) by elimination of kth equation (2.3.1-2.3.n), Mk is the number
of roots of kth subsystem7 • Using the formulae for implicit function derivative we
get from (2.4)
(2.29)
dlnR(w)
dw = -V1.
f= 1
U1(Zj1(W)) - V1 W + Vk.
t f= (_1)kJk(Zj1(W))
J 1(Zj1(W))(U1(Zj1(W)) - V1 W)'
11=1 k=2 11=1
where J 1(-), Jk(') are the Jacobians of corresponding subsystems. Let now prove
the identity

(2.30)

Consider the following system, depending on the parameter w

(2.31.1 ) (U1(Z) - V1W)(Uk(Z) - VkW) = 0,

(2.31.2) U2(Z) - V2W = 0,

Uk-1(Z) - Vk-1W = 0,
Uk+1(Z) - Vk+1W = 0,
(2.31.n) un(z) - VnW = 0,
n
(2.31.n+1) L Z i=1

i=l

This system has following properties:

(1) The en into en mapping given by lhs of (2.31) is polynomial mapping,

7It's sufficient to consider in (2.28) the values of w in the neighboUl'hood of zero.

 126

(2) It follows from (2.2) that Uj(z) are homogeneous polynomials and one can
prove [22,23] that in our assumptions the system u;(z) = 0, (i =I- k),
l:~=l Zj = 0 has unique solution Zl = ... = Zn = O. Thus, the higher
degree homogeneous components of equations (2.31) have a single common
zero point 0,
(3) The degree of system (2.31) Jacobian is deg J = l:~=l deg Uj + 1 - n. The
degree of Jacobian h of the system Uj(z) = 0, (i =I- k), l:~=l Zj = 1 is
deg h = l:~=l deg Uj + 1 - n - deg Uk. Thus deg h < deg J.
It follows from the facts (1)-(3) the applicability of Euler-Jacobi formulae [26], i.e.

(2.32) I= h(zit.(w» == 0,
_ l J(Zjl'(W»
.Jlk-

J = (UI(Z(W» - vlw)(-I)k h(z(w» + (Uk(Z(W» - VkW)Jk(Z(W»

(2.33) = {(UI(Z(W»-VIW)(-I)kJ1(Z(W», z(w) =Zjl(W)
(Uk(Z(W» - vkw)h(z(w», z(w) = Zjk(W)
Then from (2.32), (2.33) we have

t
(2.34)
h(Zjl(W» +~ Jk(Zjk(W» = 0
j,=l (-I)kJ1(Zjl(W»(Ul(Zjl(W» - VIW) j.=l (Uk(Zjk(W» - vkw)h(zjl(W» -

The identities (2.34) and (2.30) are identical. Then from (2.29) and (2.30) we get
the relation (2.28).

THEOREM 2.2.

(2.35)

where Mk is the number of solutions (with their multiplicities) of the kth equilib-
rium subsystem obtained by setting W = 0 in Eqs. (2.3.1)-(2.3.n), except the ktll
equation, and Uk(Zj.(O» are the values of the rate of the kth step in the solution
Zjk (0) = (Zljk (0), ... , Znjk (O)) of kth equilibrium subsystem.

Proof. Formulae (2.35) follows directly from (2.26) and (2.28). 0

Differentiation of (2.28) enables to find the relations for higher coefficients 8 .

THEOREM 2.3.

(2.36) B2
Bo
=.!. ((BI)2
2 Bo

8 Another theory based on the results of complex analysis (multidimensional logarithmic residue
theory) is given in [22,23].
 127

Proof. From (2.28) we get

(2.37)

where
n

(2.38) uk = (-l/h) 2)-1)·+kv.J.lz=zj.(w)

.#k

Then from (2.37)

(2.39)

Thus, from relations (2.27) and (2.39) we get (2.36). D

Remark 2.1. The following identity is valid ([27])

(2.40)

Remark 2.2. Uk(Zjk(O)) of. 0 from the assumption that system (2.3) has no
boundary steady states.
Remark 2.3. We can use (2.36) when equilibrium subsystems have multiple
roots. This fact follows from the corollaries of theorem about inverse Jacobian [26].
Thus the coefficients of kinetic polynomial, calculated by formulae (2.25), (2.28),
are the symmetric functions of the equilibrium subsystem solutions Zjk(O). One can
use (2.13) when finding Zjk(O).
Here are some results and physicochemical corollaries.
(1) A system of nonlinear quasi-steady-state equations for a single pathway
mechanism can always be reduced to a polynomial in terms of the steady-
state reaction rate. The polynomial coefficients are polynomials in the ki-
netic parameters. The lowest coefficient is proportional to the cyclic char-
acteristic C = (b~' ... b~n - b~\ ... b~nn)' After substitution of expressions
for b±i into C we obtain

(2.41)
;=1 r=l ;=1 r=l

where r rI = 2:~=1 Vi/:r and r rT = 2:~=1 Vi/:~ are the stoichiometric coef-
ficients of observed species in the net chemical equation.
 128

It follows from (2.5) that at equilibrium (w = O)Bo = 0 and

n n I I
K = IT k~; / IT k~i = IT C~rT IT C~rI
/
i=l ;=1 r=l r=l
(see (1.5)). Thus such a form of the lowest coefficient ensures the thermo-
dynamic validity of kinetic polynomial.
(2) Relation (2.6) contains a set of values Vi that have no common divisor other
than unity. This set is determined from reaction mechanism (2.1). To find
the set Vi it's sufficient to calculate the determinants ~l' -~2"'" -~n and
eliminate their greatest common divisor p. The meaning of these determi-
nants is simple: they are the cofactors N j of an arbitrary column of the
stoichiometric matrix for the intermediates in mechanism (2.1).
Thus the stoichiometric numbers contained in Bo form a minimum set
of integers. The arbitrary nature of sets of stoichiometric numbers has been
underscored more than once in the literature (e.g. see [28]). However,
the set in Bo is not arbitrary: it depends on the detailed mechanism and
defines the form of kinetic low of complex reaction (2.1). The net equation
corresponding to this set of stoichiometric numbers does not necessarily have
a minimum values of stoichiometric coefficients of the reactants: the values
of r rI, r rT in (2.41) can have greatest common divisor other than unity
(it was called "multiplicity" in [14]). For instance, for the simple linear
mechanism

the set vi(l, 1, 1) yields the net equation 2A ;= 2B, which, in contrast to
the equation A ;= B is physically meaningful, namely, it corresponds to the
reaction cyclic characteristic kl k2k3C~ - k_l k_2k_3C1.
(3) It follows from corollaries 2.1 and 2.2 that when p = IN;! Vi I > 1 kinetic
polynomial takes the form
(2.42) BLWL + ... + Bpw P + B~_l Cw p- 1 + ... + B~CP = 0
The value of p defines the dimension of equilibrium neighbourhood (see
below) i.e. we must take into account only coefficients Bp, ... , Bb in this
case.
(4) It follows from (2.35) that for linear mechanism the inverse of the steady-
state reaction rate is equal to the sum of the inverted reaction rates cal-
culated assuming that the 1,2, ... , nth steps are rate- limiting and the re-
maining steps are at equilibrium:
1 n 1
(2.43) ;=2:;;
i=1 I

It is easy to get from (2.43) the equation (1.4).

(5) The solution of the kinetic polynomial cannot generally be represented in
the explicit form like (1.2), (1.3) customary to the chemist. These equations
are valid only in the following cases: in the neighbourhood equilibrium (a),
when there exist the rate-limiting step (b), and, for a linear mechanism (c).
 129

3. Examples. Let's consider the impact mechanism

k,
1) A2 + 2Z L,
T='c 2AZ 111

(3.1) k2
2) AZ +B T='c
k_2
AB + 2Z 121

A2 + 2B T='c 2AB
where A 2, B, and AB are observed species, and Z, AZ are the surface intermediates.
After elementary transformations we'll get kinetic polynomial

where bl = kI cA 2 ,L I = k_ l ,b2 = k2cB,L 2 = k- 2CAB. The form of the free

term corresponds to (2.6) at p = 1. Its exponpnts correspond to the minimal-
integer stoichiometric numbers (1,2). Physically meaningful solution of (3.2) can
be written as:
(3.3)

The form of (3.3) corresponds to eq. (1.4): the numerator contains cyclic charac-
teristic and denominator is the sum of positive reaction weight functions.
For mechanism

k,
1) A2 + 2Z L,
T='c 2AZ 111

(3.4) k,
2) 2AZ T='c
k_,
B2 + 2Z 111

A2 T='c B2

we have NI = N2 = 2,Vl = V2 = 1,p = IN;/vil = 2. In accordance with corollaries

2.1 and 2.2 (see, also eq. (2.42)) we get
(3.5)
(Ll +b2-L2-bl?w2_2(bIb2-L1L2)(LI +b 1 +b_2+b2)W+(bIb2-LIL2)2 = 0,

where bl = k I cA 2 ,b_ 1 = k_ l ,b2 = k2,b_ 2 = k- 2CB,.

Consider now the following three-step adsorption mechanism

k,
1) A2 + 2Z k_,
T='c 2AZ 111
k,
2) B +Z T='c BZ 121
(3.6) L2

+B Z 2Z + AB
ka
3) AZ T='c 121
k_a

A2 + 2B T='c 2AB
 130

where A 2, B, and AB are observed species, and AZ and B Z are the adsorbed
substances, and Z is the active site of the surface. The following system of quasi-
steady-state equations corresponds to mechanism (3.6)

(3.7.1) blzi - LIZ~ = w,

(3.7.2) b2 z 1 - = 2w,
b- 2 Z 3
(3.7.3) b3z2Z3 - b-3zi = 2w
(3.7.4) Zl + Z2 + Z3 = 1,

where Zl = [Z],Z2 = [AZ],Z3 =

[BZ],b l = k CA j 2 ,L I = k_l,b 2 = k2cB, L2 =
k-2' b3 = k3, L3 = k- 3CAB. Obtaining Z3 = (b2z j - 2w)/L 2 and Z2 = 1 - Zj - Z3
from eqs. (3.7.2) and (3.7.4) and substituting them into eqs. (3.7.1) and (3.7.3), we
reduce eqs. (3.7) to the following system:

(3.8.1) dojzi + d ll Zl + d21 = 0,

(3.8.2) d02 zi + dj2 zj + dn = 0,

where

dOl = L I (L 2 + b2? - b: 2b j ,

d ll = -2L j (L 2 + b2 )(L 2 + 2w),

d21 = 4LIW2 + L2(L2 + 4L I )w + L j b: 2 ,
d02 = b2b3(L2 + b2) + b: 2L 3,
d12 = -b3(2(L2 + 2b 2 )w + L 2b2),
d22 = 4b 3w2 + 2L 2 (L 2 + ba)w

To obtain a kinetic polynomial it is sufficient to find the resultant R( w) of poly-

nomials (3.8) for the variable Zl. The result of its computation obtained utilising
REDUCE computer algebra system is 9

(3.9)

where
(3.10)
Bo = Llb:2(Llb:2b:3 - blb~b;)
(3.11 )
B j = 4b: 1 b~2L3 +8b: j b: 2b: 3 +8b: j b: 2 L 3b2 +4b: j b:2L3b~ +2L I b~2b:3+
4L j b~2L3bl +4L 1 b:' 2L ab1 ba +2L 1 b:'2Lab2ba +8L 1 b:' 2bj b2ba +2L 1 b:2L3b~b3+
8LI b: 2bl b~b3 +4L I b: 2bj b2bi +Lj b:2b~bi - 4L j L 2bj b~bi +2Lj L2b~bi+
Llb~b5 - b:2blb~b5

9The factor b:' 2 contained in Bo, ... , B4 was cancelled.

 131

(3.12)
E2 = 4b:' 1b-: 2 + 16b:' 1b~2L3 + 16b:' 1b~2b2 + 24b:' 1b:' 2b:' 3 + 32b:' 1b:' z L3 b2+
24b:' 1b:'zb~ + 16b:' 1LzL3b~ + 16b:' 1Lzb~ + 4b:' 1b~ - 4LI b-: zL 3-
8L1b-: 2bl + 8Llb~2b:'3 + 16Llb~2L3bl - 8Llb~zL3b2 - 8Llb~zL3b3-
16Llb~2blbz - 8Llb~2blb3 - 4Llb~zbzb3 + 24Llb:'2L3blb3 - 4Llb:'zL3b~-
8L1b:'2L3bZb3 - 8Llb:'zblb~ + 16Llb:'zblb2b3 - 4Llb:'2blb~ - 12Llb:'zb~b3-
4Llb:'zb2b~ + 24LILzblb~b3 + 16LILzblb2b~ - 12LILzb~b3 - 8LIL2b~b~-
4LI bl b~b~ - 4LI b~b3 - 4LI b~b~ + b-: 2b:' 3 + 4b-: zL 3bl +
4b-: 2bi + 4b~2L3bl b3 + 2b~zL3b2b3 + 8b~zbib3 + 4b~2bl bzb3+
2b:'zL3b~b3 + 4b:' 2bi b5 + 4b:' 2bl b~b3 + b:'2b~b5 - 4L2bl b~bh
2Lzb~b~ + b~b~
(3.13)
E3 = 4( 4b:' 1b:' 2 L 3 + 8b:' 1L 2b:' 3 + 8b:' 1L2L3b2 + 4b:' 1L3b~ + 2LI b:' 2b:' 3+
4L l b:' 2 L 3bl - 4L 1b:' 2L 3b3 - 4L l b:' 2bl b3 + Llb:'2b~ + 12LIL2L3blb3-
6LIL2L3bZb3 - 4LIL2blb~ + 2LIL2b2b~ - 2LIL3b~b3 + 4Llblb~b3+
4Ll bl b2b; + LI b~bi + 2b:' 2L 3bl b3 + 4b:' 2bi b3 - b:' 2bl bi+
4Lzbib~ - 2Lzb1b2bi - 2blb~bD
(3.14)
E4 = 16( b:' l b:' 3 + 2LI L3bl b3 - LI bl b~ + bibi)

Note that in contrast to the linear mechanism, the cyclic characteristic may "vanish"
when some reaction step is irreversible (the first or second step in mechanism (3.6)).
To compute the coefficients of R( w) by formulae (2.35), (2.36) we must find the
roots of equilibrium subsystems. 1st subsystem has two solutions (MI = 2):

= 0, Z2(1) = 1, Z3(1) = 0;
ZI(I)
(3.15) z z z 2
ZI(2)/L 2b2b3 = Z2(2)/b_ z L 3 = Z3(2)/b 2b3 = 1/(Lz b2b3 + b_ 2L 3 + b2b3)
For 2nd subsystem A12 = 4 and
(3.16)
ZI(I,2) = 0, Z2(I,Z) = 0, Z3,(1,2) = 1;
ZI(3,4)/ ± b3~ = Z2(3,4)/b 1 b3 = Z3(3,4)/L1L 3 = 1/(±b3 y'bI L I + b1 b3 + L 1L 3)
For 3rd subsystem M3 = 2 and
(3.17)
ZI(I,Z)/Lz~ = Z2(1,2)/±L 2 A = Z3(1,2)/b2~ = 1/(Lz~±L2A+b2~)
Note that subsystem 1 has boundary solution and subsystem 2 has multiple one.
The last has asymptotic

(3.18)

From relations (2.35), (2.36) we have

Ed Eo = -(SI + 2S2 + 2S3),
(3.19)
Ezi Eo = (1/2)((Ed Eo)2 - (PI + 4P2 + 4P3 - 4i12 + 4iI3 - 8i23)),
 132

where Si = '£:::1
u;l(.),Pi = '£:::1
u;2(.),f'l = '£J:~1 Jt/Jsu1Iz=z;,(0). After
substitutions we'll get expressions like (3.11), (3.12).10
In [19] some limiting cases were considered (b±i -+ 00). We consider here an
asymptotic behaviour of the reaction rate when b3 tends to infinity. The kinetic
polynomial (3.9) has in this case the form:

One can show (see [19]) that in the region

(3.21) { b2 > L2
b2 /2::::: bI ::::: (b 2 + L2? /8L2
multiple physical steady states exist, while in the other regions a single steady state
exist. From (3.20) the equations for multiple reaction rates are:
(3.22)
w(1) = b;2(yfji - JH _1)2,
W =
{
W(2) = b;2(yfji + JH _1)2, in region (3.21)
_ h bZ±.l.
W(3) - 2 4b, b2 +.l.'

where H = (b~~b2b:)2 and A = Jb~ + 8L 1(2b 1 - b2 ). Steady-state concentration

of AZ corresponding reaction rates w(1) and W(2) is low (~ b;;I) while the concen-
tration of BZ is low at steady state w(3). The branches w(1), W(2), w(3) in the space
[b I , b2 , w] form a cusp-surface [30]. The difference between a cuspoid when b3 -+ 00
and the case of finite ba is that one line of fold is straight (see [19]). The cross sec-
tions of steady-state rate surface, corresponding the dependences of reaction rate
on reactant concentrations, have breaks. The break is singular point (double point,
point of self-intersection). Consider, for example, kinetic polynomial (3.9)-(3.14)
as a function of parameters 8 = b;;I, bI , W : R( w, 8, bI ) = o. When ba -+ 00 we have
8 = 0 and at bI = W = bd2 the following relations are satisfied: R = R'w = R~, = 0
and R~~, - R:;'wR~,b, > o. Thus it's a double point by definition. Therefore the
breaks on kinetic curves result from the high rate constant of the reaction between
the adsorbed substances. Using the model corresponding to mechanism (3.6) in
assumption b3 ~ 00 it has been possible to account for kinetic curve breaks in
the oxidation of hydrogen on Pt(lll) and Pt(llO) [31] as well as on Pd, Ir and on
Pd-MOS system [27].
Thus, on the basis of a kinetic polynomial one can obtain simple relations for the
limiting cases of kinetic behaviour that describe the specific features of mechanism
(3.6). These relations can be used for a preliminary analysis of the experiment
and estimation of the parameters in a kinetic model [27, 31]. Finally, a kinetic
polynomial permits qualitative classification of the specific kinetic characteristics
of reaction mechanism. Most of these features have been studied only numel·ically
(see [29]).

10 To overcome the difficulties when multiple solutions of the equilibrium subsystems are exist
the relations (3.18) have been applied.
 133

4. Classical approximations. There are two classical limiting cases where

explicit solutions of quasi-steady state equations (2.3) exist, namely, limitation of a
step (I) and neighbourhood of equilibrium (II).
I. Existence of rate-determining kth step in the mechanism (2.1) implies that
the values of the weights of the kth and -kth reactions are small compared to those
of the rest of the reactions. In this case one can expand the solution of system (2.3)
on the degrees of small parameter bk (or L k ):

( 4.1)

where Zo is the physical solution of kth equilibrium subsystem. First member of

the series (4.1) is the rate of limiting step, calculated at values of intermediate
concentrations corresponding the equilibrium offast steps, i.e. wkO) = (l/vk)uk(zo).
Exactly the same approximation had been applied usually for the reaction rate
calculation under limitation (see [2]). It follows from (4.1) that necessary condition
of its applicability is

( 4.2)

We can obtain the expression for wiG) from the results of section 2. Suppose that
1st step is rate-determining. Expressions for Zo can be obtained from (2.13) and
(2.3.n+l) after substitution mj = o. Then from (2.3.1) we get

( 4.3)

where K.j = bj/Lj,Pl = I:j=IO'lj = I:j=l f3lj. After some transformations we

can obtain the symmetrical form of equation (4.3) for the case when kth step is
rate-determining

( 4.4)

where t;;.~~k) are the cofactors for the corresponding elements of the matrix,
(O'ij - f3ij), wherein the kth row is substituted by the row (O'kl, . .. , O'k,,), and Nk
is the cofactor for the arbitrary element of the kth row in matrix ,.
Numerator of eq. (4.4) includes the factor
 134

The rest of eq. (4.4) is the rate of the forward reaction w+. Thus eq. (4.4) corre-
sponds to expression (1.2) from Boreskov-Horiuti theory. It is of interest that the
exponent 1/Vk in the numerator of (4.4) is the apparent molecularity according to
Boreskov and in the denominator (Pk) it is the molecularity generally accepted in ki-
netics (the number of molecules of intermediates taking part in the rate-determining
step). Note that in the general case eq. (4.4) is not a power function of the con-
centrations for the observed substances, as has been suggested for the derivation of
eq. (1.2) in [1,4]11. The new fact is that it is possible to obtain for a single-route
mechanism (2.1) both explicit kinetic equation (4.4) and explicit condition of its
applicability (4.2). Expression for zo in (4.2) is

(4.5)

From (4.4) and (4.5) one can obtain the expressions for the observed reaction order
nJL with respect to the substance GJL, and the observed activation energy Eo:

(4.6)

nJL =
Olnlwl
8lnc
JL
,
= IkJL
-1
+Nk
~
L..-ZiO
;=1
( L..-
~
J#k
b,.'(k) ( ,
ji
" ))
IjJL - IjJL +
Vk
r JLI - r JLT
(Il~
J=l
V.;/Vk _
K, J
1) ,

where Ek and E-k are activation energies of rate-determining step reactions (for-
ward and back), Qj = E_ j - E j is the, ~;at of jth reaction, Q = 2::~1 ViQi is the
heat of net reaction, r JLI(T) = 2::;=1 IS~ )V s are the stoichiometric coefficients for
the Ilth substance in the net reaction equation.
Equations (4.6) and (4.7) can be applied to obtain the relation between and n;
n; that are observed reaction orders with respect to the substance GJL, and the
observed activation energies and Ed Eo
far from equilibrium. Assuming that n;
and Etare determined for the same composition of the surface (i.e. at the same
K,j), we obtain

(4.8)
n; -n;
rJLI - rJLT

Formulae (4.8) are similar to those given by Boreskov [4,8]' but for their application
it is not necessary that the kinetic equation should be exponential. They can be
used to determine a stoichiometric number for the rate-determining step according
to the experimental data. It follows from (4.6), (4.7) that observed order and

11 We must note that attempt [32] to obtain the power-form Bores)wv-Horiuti equation (1.2)
seems no correct since these authors do not consider the balance equations like (2.3.n+l).
 135

observed activation energy include three components: limiting step parameter (1),
addendum accounted for detailed reaction mechanism (b), and addendum described
reversibility of the overall reaction (c). For linear mechanism it's possible to obtain
from (4.6) and (4.7) structured formulae like in [14].
Consider, for example, the mechanism (3.6) when second step is the rate-
determining. We have N2 = V2 = 2,P2 = 1,

2 -2

°
,'(2) = (
~2 1

l:!.'(2) = 10 °1 = ° l:!.'(2) = 1-2 °1 = ° l:!.'(2) = _11 °1 = -1 l:!.'(2)

11 1 1 ' 31 ° ° ' 12 -2 1 ' 32

-I~ ~I = 0, l:!.~~) = 1~2 ~I = 1, l:!.~~) = I~ ~21 = 2. Then, by (4.1), (4.4)

and (4.5) we get

b2b3(yo - XO)+L2(2L3Z0+b3YO)+2(2bIL2Z0+2Ll(b2+L2)XO) (b
W = w 2(0) (
1- +0 . ~
4(b 1 b3z 0(xo - Yo) + LI XO( 2L 3Z0 + b3Xo)) ~)) ,

where

(0) L2 K2K3Ft - 1
w2 -- ,
2 K3Ft+KIK3+1
Xo Zo 1
- - = Yo = - - - = .
KI K3 K3Ft K3Ft + KIK3 +1
From expression (4.6)

When CB is large nA 2 = n1, = (1/2)(yo - xo), when

CB is small nA 2 = n A2
(1/2)(yo - Xo - 1). Thus nt - n A2 =
1/2 in accordance with (4.8).
II. In the neighbourhood of equilibrium the condition ]{ ~ f _( c)/ 1+( c) is ful-
filled. Then the value of C in eq. (2.42) is small (i.e. C = E --+ 0). Consider the
possibilities to find the solution of eq. (2.42) in the form of E-power series.
(A) If one express the parameter, say bI , as a function of E and substitute it in
the coefficients of (2.42) then last takes the form

(4.9)

where subscript "eq" corresponds the equilibrium value. It follows from Newton
diagram [33] of equation (2.42) with coefficient~ (4.9) that there are two classes of
solutions when E --+ 0:
(1) E-order solutions, corresponding the descending part of diagram

(4.10) W.=VI.E, s=1, ... ,p

 136

(2) Solutions of order 0(1) corresponding the horizontal part of diagram

(4.11) WI = VOl, 1= P + 1, ... , L

The values of Vls and VOl are the roots of polynomials

(4.12) B peqvpI + B'(P-I)eqv ~I

I + ... + B'l eq V l + B'Oeq = 0,
(4.13) BLeqV~-P + ... + Bpeq = O.

Reaction rate near equilibrium is one of solutions (4.10). Other solutions are non-
physical.
(B) Suppose that L > p. After substitution W = fV into (2.42) and introduction
of parameter t E [0,1] we'll get

Then the solution of (4.14) at t = 1 can be obtained in the form

(4.15) w=v • f+ d
dv I f2 + ... ,
T r=O,v=v.

where T = fl and V· is the root of equation Bpv P + ... + Bti = O.

Approximations (A) and (B) give the same order error, but examples show that
approximation (4.15) can be more precise.
Thus, the first member of steady-state reaction rate approximation in the neigh-
bourhood of equilibrium is the root of polynomial corresponding the first p + 1
coefficients of the kinetic polynomial (2.42). One can obtain the explicit form of
kinetic equation near equilibrium. Let p = 1. From (4.10) we have

(4.16)

On the other hand, in accordance with theorem 2.2

n

BdBo =- LViSi,
i=I

where Si = ~;:;"11/Ui(Zj;(0». Consider one of Si, for instance SI' It follows from
the proof of theorem 2.1 that

( 4.17)

where

(4.18) Y = L
M,
n
M,
UI(-)/
(M..
nUI(') II ul_(-)(-l)//,-1
Mtin )
.
),=1 -oF), ),=1 "n
 137

Here the subscript "b" corresponds the boundary solutions of the equilibrium sub-
system, UI-O is the rate of first step in back direction. The value of Mlin is Ivd.
We have to define Y value in the equilibrium point (i.e. TI7=1 (b;fLi)Vi = 1). Then
the numerator of Y will contain only one addendum, corresponding the product on
all solutions of equilibrium subsystem beside physical one (subscripted as "ph"):

II
Vt

(4.19) Y = (_ltt-Iu=~eq (UHO/UI-O -1),

iin:Fiph

where U-I eq = U+l eq = Ul eq is equilibrium exchange rate of the first step and UI+(')
is the rate of the first step in forward direction. It follows from (2.21) that

II II (exp(27r~1/vd -1) == (-ltt- lvl

"1 111_1

(4.20) (uH(')/UI-O -1) =

iin#jph 1=1

Then from relation (4.19) Y = vI/uleq' The result is exactly the same for other Si.
From eq (2.35) in the neighbourhood of equilibrium we have

(4.21)

This relation is valid at appropriate p values [27].

After substitution of the expression (4.5) into eq. (4.21) we'll get
n "'j - 1
TI j=l IIj

(4.22)
v2b-1 ("n TIn (1/K, .)A;\k) /PkNk)Pkl '
W= ------------~~~~--------------_.--
"n
L...k=l k k L....=l J#k J
eq

where the subscript "eq" means that the denominator in eq. (4.22) is calculated at
the equilibrium point. The numerator of eq. (4.22) is equal to Kf+(c)/f-(c) -1 ~
-tl.G/(RT), and the denominator is the inverse of the exchange rate of overall
reaction
n

l/w eq =L VUUkeq
k=l

and, hence, eq. (4.22) corresponds to equation (1.3). The value l/w eq is propor-
tional to the characteristic time of the quasi-steady-state process of the reaction
in the closed system. Thus kinetic equation (4.22) includes values determined in
various experiments: steady-state (w), relaxation (w eq ) and isotopic (Ukeq). It is
essential that for these values explicit equations have been obtained through the
parameters of reaction mechanism (2.1).
Near equilibrium the observed values nl' and Eo take the form

f l ' l - f,IT
nl' = -tl.G/(RT)
Q
Eo = tl.G/(RT)
 138

and are independent of the detailed reaction mechanism and controlled only by the
thermodynamic characteristics of the overall reaction.
A case is possible when the step limitation takes place in the neighbourhood of
equilibrium. Here eqs. (4.4) and (4.22) lead to the same result

( 4.23)

The equations (4.4), (4.22), (4.23) can be applied for different purposes: to construct
a kinetic model on the basis of the detailed reaction mechanism (a), and to verify
the hypothesis concerning the reaction mechanism (b).

5. Applications. It was shown in section 3 that kinetic polynomial can be

used for asymptotic analysis of steady-state multiplicity. On the other hand, it can
be used for direct computation of the all steady states. After variable elimination
the problem reduced to the finding of the roots of single-variable polynomial (see
[34]). Here we'll consider the application of kinetic polynomial for description the
steady-state dependences of the reaction rate (I) and the study of the parameter
identifiability of given kinetic model (II).
1. The kinetic polynomial can be written as

(5.1) L!J(k)q,j(c,w) = 0, q,j(c,w) = ci j1 ... c;j1w{j,(I+Il,

j=l

where Ci are the concentrations of the species observed, fjCk) are the polynomials in
the reaction rate constants k, m is the number of different monomials <Ph and eji are
integers. The problem of finding implicit kinetic dependences may have either one of
two formulations: the reaction mechanism and the corresponding kinetic polynomial
in the form (5.1) are specified, and the polynomial coefficients fj = fj(k) are to be
determined (A); the reaction mechanism is unknown, and the kinetic dependence
in the form of polynomial (5.1) and its parameters are to be found (B).
In terms of eq. (5.1) the problem (A) formulated as follows: the experimental
data Ci and Wi being known, the !J values optimal in terms of a certain criterion
are to be found. Usually the optimization criterion has the form
N
(5.2) L(w:, - Wi)2gi ---+ min,
i=l

where wi and Wi are the steady-state reaction rates measured experimentally and
calculated by eq. (5.1), gi are the weights of the experimental points, and N is the
number of measurements.
The kinetic polynomial (5.1) admits various approximations of criterion (5.2).
The Ihs of eq. (5.1) may be expanded in a series 12 in the vicinity of wT

12 C11 ... J Cl are assulned to be measured without experimental error.

 139

By truncating expansion (5.3) one obtains approximations of different orders for

criterion (5.2).
The zero order approximation

(5.4) L1i(k)(MCi,wf) = 0
j=l

the first order one being

(5.5)

Usually the set of equations (5.4) is overdetermined (N 2': m,N 2': dimk). A simple
way to solve the equations is to use the linear least squares method (LSM).

(5.6) IIfll = const.

As an example consider mechanism (3.6). The experimental points are the values of
function wee CA2) calculated at the following parameters: L3 = 0.5, b_ 2 = 1.0, Ll =
0.1, kJ = 1.0, b2 = 2.0, b3 = 104 ,0 < CA 2 < 1.3. Not.e that the "experimental" points
were selected in a region with non-unique steady states. The parameters of the
kinetic polynomial (3.9)-(3.14) corresponding to mechanism (3.6) are as follows (see
e . (5.1»: I = 1 m = 13 (C)T = (0 1 0 1 0 1 2 0 1 2 0 1 2).
q , d 0011222333444
The formulas for the coefficients Ii(k) are given in [21].
Table 5.1 lists the values of x j = Ii / h3 determined by solving the set of normal
equations by the Gauss method, (the number of points is 13, the number of points
in the multiple steady-state region is 10) along with the true values of xj. As is seen
from Table 5.1, all the parameters except Xl and Xli were estimated satisfactorily
(Xl and Xli may be taken to be equal to zero since the backward reaction 3 in

mechanism (3.6) may be disregarded for the given set of rate constants). The
computations proved the accuracy of determinat.ion of x to depend on the number
and character of experimental points, it deteriorates with an increase in the density
of the points, because problem (5.6) is ill-conditioned. The problem (5.6) was also
solved using the algorithm of the singular value decomposition SVD [35] which
reduces the general LSM problem to the one with diagonal matrix. The problem
conditionality may he improved by introducing a bound for the minimal singular
value. The calculation results X(1) for such an example together with the singular
values a are presented in Table 5.1 (the number of points is 50, the number of
points in the multiple steady-state region is 36). The estimates X(1) correspond to
the choice of the bound of the minimal singular value which exceeds a12. As follows
from Table 5.1, ad al2 ~ 10 10 , i.e. the problem is very ill-conditioned. Estimates
X( 1) are worse then X j, nevertheless their signs and orders are the same as for the
true values, except for Xl and Xli'
 140

Estimates obtained by using the SVD method may be improved with the aid of
ridge regression [36]. Example of its implementation is given in [21]. Thus:

(1) the zero order approximation produces quite adequate estimates of x j III

some cases;
(2) despite the fact that problem (5.6) is ill-conditioned it may be regularized
by standard methods of the linear LSM;
(3) there exist a possibility of finding quickly an initial approximation for esti-
mates of the kinetic parameters by linear methods.

TABLE 5.1
Solution of the problem (5.6) for the reaction scheme (3.6) and different samples

Xl. Xj X(1) (J'

J

1.5625E - 12 1 -1.2589E - 08 2.989E - 09 6.735E 06

-2.5000E - 02 2 -2.5023E - 02 -4.094E - 02 3.772E 04
2.2500E - 01 3 2.2521E - 01 3.685E - 01 1.845E 02
-2.9996E - 01 4 -3.0016E - 01 -2.887E - 01 1.131E 01
1.7999E 00 5 1.8006E 00 1.125E 00 3.218E 00
-9.2480E - 01 6 -9.2506E - 01 -8.288E - 01 6.464E - 01
2.5005E - 01 7 2.5012E - 01 2.902E - 01 2.752E - 01
2.2497E - 01 8 2.2630E - 01 1.457E 00 1.482E - 01
-3.1499E 00 9 -3.1536E 00 -3.611E 00 1.394E - 02
1.0001E 00 10 1.0014E 00 1.027E 00 6.821E - 03
2.5000E - 11 11 -9.0114E - 04 -5.473E - 01 1.510E - 03
-9.9990E - 02 12 -9.8605E - 02 -4.677E - 02 3.222E - 04
1.0 13 1.0 1.0

Formally, the problem (B) reduces to the finding the form of the monomials
rPj(c,w) and coefficients Ij of eq. (5.1) which fit best a certain crit.erion used for
the description of experiment. In the numerical experiments presented in [21] this
problem was solved by a step-ordering method [37] which permits one to construct
polynomials with a maximal convergence speed. The informative population of
monomials was estimated with the aid of the slip-control. The results presented in
[21] show that in spite of not all of the monomials of the original polynomial appear
in the reconstructed polynomial, the signs and orders of the recovered coefficients
coincide with the true ones. It's important that the deduced polynomial represents
qualitatively the kinetic dependence characterized by multiple steady states: it
incorporates terms with w 3 , w 4 .
The method offered here consists essentially in linearization of the initial prob-
lem. Linearization is commonly used to evaluate kinetic parameters, e.g. Arrhenius
equation parameters and reaction orders; in the case of explicit rational kinetic de-
 141

pendences form like (5.1) has been used, for instance in [38,39]. Transformation of
a kinetic model to form (5.1) offers the following advantages:

(1) It permits solving the problem rapidly by the multiple linear regression
method;
(2) When the linear problem is nonconfluent (or may be regularized) the solu-
tion uniqueness is guaranteed;
(3) The results of solution of problem (5.6) may be utilized to find the initial
approximation when estimating the original kinetic parameters by nonlinear
procedure;
(4) When eq. (5.1) is used experimental data characterized by multiple steady
states are treated in the same way as those with a single steady state;

(5) Form (5.1) enables one to formulate the problem of determining steady-state
relations in the absence of information on the reaction mechanism.

II. The inverse chemical kinetics problem does not necessarily have a unique
solution, in particular it may posses a continuum of solutions. When analyzing this
problem one should take into account possible relations between kinetic parameters.
The techniques for determining these relations was suggested in [40]. However
this technique can involve rather cumbersome calculations, first of all for nonlinear
mechanisms. Criteria for the appearance of relations between parameters stemming
from the structure of the reaction graph were presented in [41] for linear mechanism.
On the basis of kinetic polynomial concept we can suggest a new approach to the
determination of relations between parameters of nonlinear kinetic models.
Under assumption of quasi-steady-state approximation kinetic model can be
reduced after elimination of unknown int.ermf'oiate concf'ntrations to the form

de
(5.7.1) dt = Gw,
m;

(5.7.2) L!;j;(k)<Pij;(e,Wi)=O, i=l, ... ,P,

ii=l

where ware either the rates on reaction routes or their linear combinations; P =
dim w ~ R (R is the number of reaction routes); G is an integer-valued matrix; and
k is the vector of kinetic constants l3 .

Therefore if it is possible to replace variables k -+ P in (5.7.2) so that dim p =

r < dimk = n, then system (5.7) also depends only on p, i.e. singularity takes
place.

13It is assumed that expressions (5.7.2) are normalized to one of fiji (k). Thus mi refers to the
number of independent coefficients of (5.7.2).
 142

It was proved [20) that relations between constants k can follow from matrix

8111 /8k1 8lll/8kn

811 m, /8k1 811m,/8kn

(5.8) B(m x n) =
8fpd8k 1 8fpd8kn

81Pmp /8k1 8lPm p/ 8k n

where m = 2:;=1 mi. There are at least two situations which may cause the appear-
ance of such relations, namely, when the number m of coefficients in forms (5.7.2)
is less than the number n of rate constants and when m :0:: n and rank (B) < n, i.e.
partial derivatives of the functions I are linearly dependent. The former situation
does not exclude relations arising due to the latter cause (see example below).
Thus the procedure of determining the relations and performing a transforma-
tion to new parameters amounts for system (5.7) to the following: reduction of the
system of quasi-steady equations to a single polynomial (or, in the case of multiroute
mechanism, to several polynomials) in w, transformation of the resulting equations
to form (5.7.2), determination of the rank of Jacobian (5.8), and determination of
the relations between the rate constants.
The last can be found in the following way:

(1) r linearly independent functions f from form (5.7.2) are taken as coordinates
p;
(2) formulae for the remaining functions f are written;
(3) the parameters k are eliminated from resulting algebraic system of equa-
tions.
Now we indicate some properties of the algorithm under discussion in which it differs
from the procedure suggested in [40]: explicit functions of the rate constants k are
analysed and therefore relations between them (their functional combinations) can
be found directly without investigation the Jacobian matrix; the factors causing the
appearance of relations between kinetic parameters can be revealed; the algorithm
requires computer calculation of the rank of a matrix whose elements are rational
functions. To calculate the rank of such a matrix use can be made of the effective
analytical computer methods.
Consider for example a liquid-phase hydrocarbon oxidation scheme (a fragment)

RH + O2 ~ R + H0 2 , R + R ~,
R+02~R02' R02+R~,
(5.9)
R0 2 + RH ~ ROOH + R, R0 2 + R02 ~,
ROOH ~ RO+OH
 143

Here RH is a hydrocarbon; R,R0 2,RO, and H0 2 are radicals, and ROOH is the
observed species. It is assumed that [RH] and [0 2 ] are constant, and quasi-steady-
state conditions hold for [R] and [R0 2]. Mechanism (5.13) is not single-routed but
it is possible to derive a polynomial with respect to the rate of formation of the
observed species ROOH from the quasi-steady-state conditions for the radicals R
and R0 2 •
Let P = [ROOH],wo = kO[02] (the value of ko is known), PI = kd02l/WO,
P2 = kl [RH]/wo, P3 = k3/wo, P4 = 2k4/wo, ps = ks/wo, P6 = 2k6/wo, fJ = two, U =
P2[R0 2]. Then we have

(5.10.1) dP/dfJ = u - p 3 P,
(5.10.2) u4 + au 3 + bu 2 + C1J. + d = 0,
where

a = 2PV P6, b = p~(pip6 + P~P4 + 2PIP2PS - p;)/(P6(P4P6 - pm,

e = 2PIP~PS/(P6(p4P6 - p~)), d = -pipV(P6(P4P6 - pm.

Expression (5.10.2) is a polynomial of the fourth order and hence the number of the
variables on which its solution depends is less than the number of rate constants
in mechanism (5.9). Therefore even formula (5.10.2) alone permits drawing a con-
clusion about relations between the parameters. To determine the exact number of
independent parameters we find the rank of the matrix

B ~ (a.?a p,
aajap2
abjap2 ab/ap4
0
Bb/aps
0
aa/a",
ab/ap6
)
(5.11)
ae/apI ae/ap2 aC/ap4 Be/aps aejap6
ad/apI ad/ap2 ad/ap4 ad/Bps Bd/Bp6
The diagonal minor of the third order of B is not idpntically zero. At the same time
the two minors of the fourth order bordering it are identically zero and hence the
rank of matrix (5.11) is three. Thus, taking into account additionally the parameter
P3, we have 4 independent kinetic parameters.
The simplest system of functional combinations is
I
P3 = P3,
and

a = 2p;, b = (p;)2 + ((p; + p~)2 - p;)/(p~ -1),

e=2p;p~j(p~-1), d=-p;(p~)2j(p~-1)

Note that functional combinations can also be found immediately from the form of
a, b, e, and d.
Thus the reduction of the system of quasi-steady-state equations to polynomials
in the rates of formation of the species observed and the investigation of the co-
efficients of these polynomials permits determination of relations between the rate
 144

constants of nonlinear kinetic models. The determination of the relations between

the rate constants must precede the solution of any inverse kinetic problem. To
determine the relations, use is made of information about the reaction mechanism
and the type of experiment (i.e. what parameters are measured and which of the
species observed change and which are maintained constant). In doing this, no
numerical values of the experimental data are employed.

6. Algorithm of computation. Formulas for coefficients of kinetic polyno-

mial as functions of the equilibrium subsystems roots have been obtained in section
2 (see eqs. (2.35) and (2.36)). These expressions are useful for theoretical anal-
ysis and obtaining of chemical corollaries, but non-effective for calculation. New
expressions for kinetic polynomial coefficients have been obtained recently [42].
Let Uj(z) = bj r17=1 Z~;i - L j r17=1 Zf;i (for notation see section 2) and Un+l =
ZI + ... + Zn -
1. It follows from the assumptions of section 2 (there are no roots of
the system (2.3) at w = 0) and Hilbert theorem about zeros [16J that there exists a
set of homogeneous polynomials aj k (z), j, k = 1, ... ,n such that

(6.1) Z.f+l = I>jk(Z)Uk(Z), j = 1, ... ,n

k=1

The degree L in (6.1) can be chosen 1'1 + ... + 1'n - n or less, where l' j = L:?=1 Oi ji =
L:~=1 /3ji (see (2.2)) is the degree of the Uj(z).
Denote as A the matrix (ajk(z))'j,k=1 and as J s the Jacobian of 8th subsystem.
The following facts have been proved in [42J:

THEOREM 6.1. If the system (2.3) has no roots at w = 0 then

(6.2) Bl =d =~(_1)n+s-l -;::;((ZI+"'+zn)T.-l]sdetA)

Bo 1 L.J Vs- zL ... zL '
8=1 1 n

where:=: is a functional which extracts from Lorain polynomial Q = P( ZI, ... , zn)
/(Zf' ... z~n) its free member (i.e. the coefficient corresponding the monomial
Zf' ... z~n in the polynomial P(Zll ... , zn)).
THEOREM 6.2. When conditions of Theorem 6.1 are fulfilled
(6.3)
2B2 _ (Bl)2 = d = ~ ~ (_I)n+s-l -;::; ((ZI + ... + znY·+T,-IJs det Aakl)
Bo B0 2 L.J L.J VsVI_ L 2L+l L
k=1 s,I=1 Z
1
···z k ···z n

In general case we have

THEOREM 6.3. When conditions of Theorem 6.1 are fulfilled and 8 >0
n n
(6.4) d.= L
1=1 1I0011=s
0'1="'=Ol_1=0, a,>O
 145

wl]ere

Here d. = d(s) In R( w )/dw' (see section 2), a = (al,"" al, .. " an), a! = al!' .. an!,
VOl = Vf' ... v~n,
cit ... j._1 = !JI! ... (3n!, where (31,(32 ... are the numbers of "1",
"2", ... in the set ofsubscriptsjl, ... ,j._I'

Thus the algorithm of kinetic polynomial computation includes the steps:

(1) calculation of matrix A,
(2) calculation of d. by formula (6.4), and
(3) calculation of Bk/Bo by formula (2.25)
Possible way to calculate matrix A is the method of undefined coefficients which
reduces the problem to the solving of some linear equation system. It was applied
when analysing the examples in [42]. Explicit expressions for ajk(z) are given in
[43] for the special form systems.
We find however that the method of ajk(z) generation via the Grobner bases
procedure [44,45] is more convenient for realisation.
This method is technique that gives the solutions of such problems as finding of
exact solutions of algebraic equation systems (i), finding of polynomial solutions of
the systems of homogeneous linear equations (ii), and other problems of polynomial
ideal theory.
For our purpose (calculation of matrix A) we applied Grobner bases to solve
problem (ii)14. The algorithm is as follows
(1) Grobner basis calculation for the system UI (z), ... , un(z). We find simulta-
neously a linear representation of Grobner base polynomials gi through the
original polynomials Ui (z):
n
(6.5) gi = LUjXij, i = I, ... ,m,
j=1

where Xij are polynomials in ZI, ... , Zn;

(2) Reduction to zero (in sense of [44]) of the monomials Zf+1 (j = 1, ... , n)
mod (gl, ... , gm). Simultaneously coefficients of the representation

L h;k) gi
m

(6.6) Zf+l =
i=l

are calculated;
(3) Calculation of the ajk(z) by formula
m

(6.7) ajk(z) = I>;k) Xij

i=l

14Nevertheless one can obtain kinetic polynomial by solving problem (i), but our experience
proved that direct application of Grobner bases to kinetic systems may be difficult in computational
aspect.
 146

Second part of the algorithm calculates Bi by formulas (6.4) and (2.25).

The algorithm has been realized in computer algebra system REDUCE [46]15.
The program description is given in [47]. The results of matrix A calculation are
presented below for adsorption mechanism (3.6)

al,l = (b2b5L2Z3)/(blb~b5 - L l b:' 2b:' 3)

al,2 = (z2b2b3LIL2L3Z1 + z2b3Llb:'2L3Z3 + blb~b5zi-
L I b:' 2b:' 3zi) / (b l b~ b5 - b2L I b:' 2b:' 3)
al,3 = (z2b~b3LI L2 + LI b~2L3Z3)/(bl b~b5 - b2L I b:' 2b:' 3)
a2,1 = (-z2blb~b~ + Z2Llb:'2b:'3 + blb3b:'2L3Z3)/(blb~b~LI - b:' l b:' 2b:' 3 )
a2,2 = (z2bib2b~ZI + z2bib~L2Z3)/(bl b~b~LI - b:' l b:' 2b:' 3)
a2,3 = (Z2bILlb:'2L3 + bib3b:'2Z3)/(blb~b~LI - b:' l b:' 2b:' 3)
a3,1 = (bjb5Z3)/(blb~b5b:'2 - L l b-: 2b:' 3)
a3,2 = (z2b~b3LIL3Z1 + z2b~b3LIL2L3Z3 - blb~b~Z3Z1 - blb~b~L2Z5+
+ Llb~2b:'3Z5)/(blb~b~b:'2 -
b2Llb:'2Z3Z1 L l b-: 2b:' 3)
a3,3 = (z2bjb3LI + b~Llb:'2L3Z3)/(blb~b~b:'2 - L l b-: 2b:' 3)

7. Concluding remarks. It is possible to obtain the relations like kinetic

polynomial for multiroute mechanism. After the elimination of unknown interme-
diate concentrations one can reduce the model to the system of polynomials in the
reaction path rates (or the rates of reaction of certain observable species). The
structure of coefficients in this case however may be rather complex than for single-
route mechanism. Consider, for example, the two-route mechanism

k,
1) A2 + 2Z k_l
~ 2AZ 11 11
k2
2) B+Z ~ BZ 12 21
k_>

(7.1 ) ka
3) AZ + B Z ~ 2Z + AB 12 01
k_a
k4
4) AZ+B ~ Z+AB 10 21
k_4

A2 +2B ~ 2AB

that is the combination of adsorption and impact mechanism (3.6) and (3.1). The
polynomial in the rate of formation of reaction product has fourth degree like (3.9).
Its free member is
(7.2)
Bo =4b:'2(Ll(blb~bi - L 1 b:' 2b:' 3 ) + 2L 1 b1 (L 2 + b3 + b4)(b4L2L3 - L 4b2 b3 )+
2L 1 (b 2 + L2 - L4)(b 1 b2b3 b4 - L 1 L 2L 3 L 4) + (blb~ - L 1 b:' 4)
(L 1 (b 2 + L2 - L4? - b1 (L2 + b3 + bd»
15computers IBM PC AT, PS2.
 147

We should note that there are four terms in expression (7.2). The 1st and 4th
correspond the basic routes in the mechanism (7.1). The 2nd and 3rd correspond
their linear combinations. Thus the structure of the free term for multiroute case
is more complex than in single-route one (see Theorem 2.1). Nevertheless the ex-
pression (7.2) is thermodynamically correct. It follows from conditions of equilib-
rium (see [14]) that k4/k_4 = k2k3/k-2L3' Therefore b4L2L3 = L 4b3b3 for
appropriate values of CA 2,CB, and CAB. Thus expression (7.2) can be written as
Bo = Bb(klk?k~CA2C~ - L1E2k=-3C~B) similar to that of single-route mechanism
(3.1) and (3.6).
Kinetic polynomial is the general form of complex reaction kinetic equation.
As a rule it cannot be simplified to the explicit equation like (1.2) suggested by
Boreskov and Horiuti. The last is valid only in limiting cases: linear mechanism (i),
existence of rate-determining step (ii), neighbourhood of equilibrium (iii). These
cases have been analysed on the basis of kinetic polynomial and there were obtained
explicit analytical expressions here. Kinetic polynomial can find various applica-
tions. It simplifies the analysis of various limiting cases, facilitates the solution of
the inverse kinetic problem. It seems advisable to apply it in describing compli-
cated kinetic behaviour (e.g. multiplicity of steady states). It can also be used
as a correct regression model. (The known regression models do not satisfy the
"equilibrium test"). Complex expressions of the coefficients of kinetic polynomial
can be effectively derived by the methods of computer algebra.
We have proved recently that forms like kinetic polynomial can be obtained not
only in the case of ideal kinetic low for the elementary step (MAL), but also in the
case of non-ideal kinetics or for non-isothermal conditions. There are no however the
explicit polynomial expressions for their coefficients as in the case of ideal models.
But these generalisations as well as the problem of variable elimination in the case
of unsteady state kinetic models are out the scope of presented paper.

REFERENCES
1. J. HORIUTI, Theory of reaction rates as based on the stoichiometric number concept, Ann.
New York Acad. Sci. 213 (1973), pp. 5-30.
2. M.1. TEMKIN, The kinetics of some industrial heterogeneous catalytic reactions, Adv. in
Catalysis 28 (1979), pp. 173-291.
3. J. HORIUTI, "Reaction kinetics", Iwanami Book Co., Tokyo (1940).
4. G.K. BORESKOY, The relation between molecularity and activation energies of reaction in
forward and back directions, Z. Fiz. Chim. 19 (1945), pp. 92-95.
5. S. DE GROOT AND P. MAZUR, Nonequilibrium Thermodynamics", Mir, Moscow (1964).
6. T. NACAMURA, Note on chemical kinetics in the neighhourhood of equilihrium, .J. Res. Inst.
Cata!., Hokkaido Univ. 6 (1958), pp. 20-27.
7. M. BOUDART, "Kinetics of Chemical Processes", Englewood Cliffs, New Jersey (1968).
8. S.1. KIPERMAN, "Introduction to the Kinetics of Heterogeneous Catalytic Reactions", Nauka,
Moscow (1964).
9. J. HAPPEL, A rate expression in heterogeneous catalysis, Chern. Eng. Sci. 22 (1967), pp.
479-480.
10. J. HAPPEL, Comments on Huriuti's stoichiometric number concept, J. Res. lnst. Cata!.,
Hokkaido Univ. 28 (1980), pp. 185-188.
11. M. BOUDART, D.G. LOFFLER AND J.C. GOTTIFREDI, Comments on the linear relation be-
tween reaction rate and affinity, Int. J. Chern. Kinetics 17 (1985), pp. 1119-1123.
 148

12. M.S. SPENCER, Thermodynamic constrains on multicomponent catalytic systems. II. Limits
to pseudo-mass-action kinetics, J. Catalysis 94 (1985), pp. 148-154.
13. E. KING AND C. ALTMAN, A schematic method o[ deriving the rate laws [or enzyme-catalyzed
reactions, J. Phys. Chern. 60 (1956), pp. 1375-1381.
14. G.S. YABLONSKII, V.1. BYKOY AND A.N. GORBAN, "Kinetic Models of catalytic Reactions",
Nauka, Novosibirsk (1983).
15. A.G. KUROSH, "Highter Algebra", Mir, Moscow (1972).
16. B.L. YAN DER WAERDEN, "Modern Algebra, Part 2", Ungar, New York (1970).
17. G.S. YABLONSKII, M.Z. LAZMAN AND V.1. BYKOY, Stoichiometric number, molecularity and
multiplicity, React. Kinet. Cata!. Lett. 20 (1982), pp. 73-77.
18. M .Z. LAZMAN, G.S. YABLONSKII, AND V.1. BYKOY, Steady-state Kinetic equation. Non-linear
single pathway mechanism, SOy. J. Chern. Phys. 2 (1985), pp. 404-418.
19. M.Z. LAZMAN, G.S. YABLONSKII, AND V.1. BYKOY, Steady-state Kinetic equation. Adsorp-
tion mechanism of a catalytic reaction, SOy. J. Chern. Phys. 2 (1985), pp. 693-703.
20. M.Z. LAZMAN, S.1. SPIVAK AND G.S. YABLONSKII, Kinetic polynomial and the problem of
determining relations between kinetic constants when solving the inverse problem, Sov. J.
Chern. Phys. 4 (1987), pp. 781-789.
21. M.Z. LAZMAN, G.S. YABLONSKII, G.M. VINOGRADOYA AND L.N. ROMANOY, Application
o[ the kinetic polynomial to describe steady-state dependence of the reaction rate, SOy. J.
Chern. Phys. 4 (1987), pp. 1121-1134.
22. V.1. BYKOY, A.M. KYTMANOY, M.Z. LAZMAN, AND G.S. YABLONSKII, Resultant of quasi-steady
-state equations for a single-route n-stage mechanism, Khim. Fiz. 6 (1987), pp. 1549-1554.
23. V.1. BYKOY, A.M. J(YTMANOY, M.Z. LAZMAN, AND G.S. YABLONSKII, Kinetic polynomial
for one-route n-stage catalytic reaction, In: Mathematical Problems of Chemical Kinetics,
Nauka, Novosibirsk (1989), pp. 125-149.
24. L.A. AIZENBERG AND A.P. YUZHAKOY, "Integral representations and residues in multi-dimen-
sional complex analysis", Nauka, Novosibirsk (1979).
25. D.N. BERNSTAIN, The number of the roots of the system of equations, Fund. Ana!. App!. 9
(1975), pp. 1-4.
26. V.1. ARNOL'D, A.N. VARCHENKO, AND S.M. GUSEIN-ZADE, "Singularities of differentiating
mappings", Nauka, Moscow (1982).
27. M.Z. LAZMAN, "Study of nonlinear kinetic models of heterogeneous catalytic reactions (Dis-
sertation thesis)" , Institute of Catalysis, Novosibirsk (1986).
28. M.1. TEMKIN, The kinetics o[ heterogeneous catalytic reactions, Zh. D.l. Mendeleev Vses.
Khim. 20 (1975), pp. 7-14.
29. G.S. YABLONSKII, V.1. BYKOY AND V.1. ELOKHIN, "Kinetics of model Reactions of Hetero-
geneous Catalysis", Nauka, Novosibirsk (1983).
30. T. POSTON AND I. STEWART, "Catastrophe Theory and Its Applications", Pitman, London
(1978).
31. M.Z. LAZMAN, G.s. YABLONSKII, AND V.A. SOBYANIN, Interpretation of breaks on kinetic
curves, Kinet. Cata!. 27 (1986), pp. 57-63.
32. M. VLAD AND E. SEGAL, On the kinetic model of the rate-determining step. 1, Rev. Roumaine
de Chimie 24 (1979), pp. 799-805.
33. M.M. VEINBERG AND V.A. TRENOGIN, "Theory of branching of the nonlinear equations
solutions", Nauka, Moscow (1978).
34. L.A. AIZENBERG, V.!. BYKOY, A.M. J(YTMANOY AND G.S. YABLONSKII, Search [or all
steady-states of chemical kinetic equations with the modified method of elimination. I. Al-
gorithm, II. Application, Chern. Eng. Sci. 38 (1983), pp. 1555-1568.
35. G.E. FORSITHE, M. MALCOLM AND C.B. MOLER, "Computer Methods for Mathematical
Computations", Englewood Cliffs, New York (1977).
36. C.L. LAWSON AND R.J. HANSON, Solving Least Squares Problems", Englewood Cliffs, New
York (1974).
37. L.N. ROMANOY, On recovery of a functional dependence by steep ordering, Preprint No.
372, Computer Centre USSR Acad. Sci., Novosibirsk (1982).
38. G.P. MATHUR AND G. THODOS, Initial rate approach in the kinetics of heterogeneous cat-
alytic reactions - an experimental investigation on the sulphur dioxide oxidation reaction,
Chern. Eng. Sci. 21 (1966), pp. 1191-1200.
39. R.P.L. AnsIL, J.B. BUTT AND J.S. DRANOFF, On the estimation o[catalytic rate equation
parameters, J. Catalysis 87 (1984), pp. 530-535.
 149

40. S.I. SPIVAK AND V.G. GORSKII, About completeness of available kinetic data when deter-
mining kinetic constants of complex chemical reactions, Khim, Fiz. 1 (1982), pp. 237-243.
41. V.A. EVSTIGNEEV AND G.S. YABLONSKII, Non-identifiability of kinetic model parameters
as a consequence of non-Hamiltonial structure of complex reaction graph, Teor. i Eksper.
Khimia 6 (1982), pp. 688-694.
42. V.I. BYKOV AND A.M. KYTMANOV, An algorithm of construction of kinetic polynomial
coefficients for non-linear single-route mechanism of catalytic reaction, Prcprint No 40M,
Institute of Physics USSR Acad. Sci., Krasnoyrsk (1987).
43. V.I. BYKOV AND A.M. KYTMANOV, About one modification of the method of non-linear
algebraic equation system resultant construction, Preprint No 44M, Institute of Physics
USSR Acad. Sci., Krasnoyrsk (1988).
44. B. BUCHBERGER, Grobner bases: an algorithmic method in polynomial ideal theory, CAMP
-Publ. Nr 83-290 (1983).
45. F. WINKLER, B. BUCHBERGER, F. LICHTENBERGER AND H. ROLLETSCIIK, An algorithm for
constructing canonical base of polynomial ideals, ACM Trans. Math. Software 11 (1985),
pp. 66-78.
46. A.C. HEARN, "REDUCE user's manual. Version 3.2", Rand Corporation, Santa Monica
(1985).
47. V.I. BYKov, A.M. I(YTMANOV AND M.Z. LAZMAN, "Methods of elimination in computer
algebra of polynomials", Nauka, Novosibirsk, (to appear) (1991).
 CONVERGENCE OF TRAVELLING
WAVES FOR PHASE FIELD EQUATIONS TO
SHARP INTERFACE MODELS IN THE
SINGULAR LIMIT

YASUMASA NISHIURA* AND GUNDUZ CAGINALPt

Abstract. We show the convergence of travelling waves for the phase field equations to that
of a .nodified Stefan model in an appropriate singular limit. Finite surface tension effect is crucial
to prove this convergence.

§1. Introduction. It is known in [1], [4] that, by using a formal asymptotic

analysis, the phase field equations approach a sharp interface model with curvature
and dynamic cooling effects in an appropriate singular limit.
One of the most interesting questions from dynamical point of view is that
how is the relation between the solutions of the phase field model and those of the
limiting sharp interface model. For the steady state problems, it has proven in [3]
under various conditions that the steady state phase field equations converge to the
steady state modified Stefan problem.
In this note we present the first rigorous convergence results in the dynamical
setting. Namely, we show the existence of one-dimensional travelling waves for the
phase field equations (1.1), (1.2) which converge to that of the modified Stefan
model (1.6). The detailed proof is given in [5].
The phase field equations can be written as

(1.1)

(1.2)

subject to appropriate initial and boundary conditions, where land K are dimen-
sionless latent heat and diffusivity, respectively. The function 9 is a derivative of a
symmetric double well potential with minima at ±1, for instance, g(cp) = ~(cp_cp3).
e
The parameters lX, and a are related to the relaxation time, strength of microscopic
interactions and the depth of the double well potential, respectively. and a are e
usually taken to be small.
For our purpose, it is convenient to define two parameters

(1.3)

*Mathematics Department, Hiroshima University, Hiroshima 730, Japan

tMathematics Department, University of Pittsburgh, PA 15620 USA
 152

where c: measures the width of the transition layer for "', 0"0 is related to the
surface tension, and m is a positive constant defined by m == 11.p~lli2(R) (see (2.13)
for .pO). The equation (1.1) becomes
2c:m
(1.4) ac:2"'t=c:2~",+g(",)+-u.
0"0

We are concerned with the singular limit where ~ and a tend to zero, but ~a-l/2
remains finite, namely
(1.5) dO 0"0 = fixed > °.
That is, one can allow the interfacial thickness to vanish while the surface tension
remains fixed. In this scaling limit, it was shown in [1] by a formal asymptotic
analysis that the phase field equations approach the following sharp interface model
(called the "modified Stefan model" in [1]):
ut=K~u

(1.6) Iv = K[Vu . n]:;: }

0"0 0"0
on r,
U =- ~S K - a ~S v
where r denotes the interface, v is the normal velocity of the interface, [ ]:;: denotes
the jump in the normal derivative of u from solid to liquid, K is the sum of principal
curvatures and ~S is the entropy difference between the two phases. It is not
difficult to verify [2], [7] that (1.6) has the following travelling wave solutions with
velocity c* and u(t,+oo) = Ueool:
Ucool + I e K
_L (x-e"t) x> cOt
u(t,x) =
{
Ueool + I x ~ cOt
(1.7)

c* = -~ (Ueaal + I) .
aO"o
For travelling waves to the phase field equations, we impose the boundary conditions

u(f, -00) = Us, ",(f, -00) = ",'--00

(1.8)
u(f,oo) = Ueaal < 0, ",(f,oo) = "'~ ,
where ",'--00 is the left most root of g( "') + 2c:m
0"0
Us = 0, and "'~ is the right most
2c:m ..
root of g(",) + -- Ueaal = 0. Note that the above boundary conchhons are not
0"0
independently given, in fact, it turns out in §2 that once Ueaal is given, other data
are uniquely determined as functions of c:. We also assume that
(1.9) Ueaal +I< 0
saying that the supercooling temperature Ueaal is sufficiently low compared with the
latent heat. This ensures the positivity of the travelling velocity.
Our main result is the following.
 153

THEOREM [5]. In the scaling limit (1.5) and under (1.9), the phase field equa-
tions (1.2), (1.4) has an [-family of one-dimensional travelling waves ('P',u') with
velocity c = c([) satisfying (1.8). Moreover, as [1 O,('P',u') converge to the trav-
elling wave (1.7) of the modified Stefan model. More precisely, c( [) is a continuous
function of [ up to [ = 0 satisfying lim c( [) = c', and ('P', u') has the following
,10
limiting behavior for any A > 0:

-I uniformly for z E (-00, -A)

lim 'P' = {
dO 1 uniformly for z E (A, 00)

Ucool +l for z ::; 0

lim u' ={
eto Ucool + le-tc (x-c't) for z >0
where the limit is uniform on R with respect to the travelling coordinate z
x - c( [)t. Here we fix the phase of the travelling waves as

'P'(O) = ~ .
REMARK 1.1. In different scaling regimes where (1.1), (1.2) approach the clas-
sical Stefan model or the alternative modified Stefan model (see [1]), it can he
proved [5J that the phase field equations do "not" have travelling waves under the
condition (1.9).

There are several related works: travelling wave solutions to the phase field
equations in a different scaling regime (i.e., a = constant) has been studied numeri-
cally [9], and a bifurcation from planar to curved travelling fronts has been obtained
in [8J for a different scaling regime (i.e., a = constant, a = O( ~-1)).

§2. Travelling wave solutions to the phase field model and its singular
limit. Introducing the moving coordinate

(2.1 ) z=x-ct,

equations (1.2), (1.4) become

2 2 2[m
(2.2) ['Pzz+ a[ C'Pz+g('P)+-- u=O
(To

(2.3) J{u zz + C(U z + ~ 'Pz) = 0 .

Integrating (2.3) from z = -00 (see (1.8)), we have

(2.4) J(U z + c{(u - us) + ~ ('P - 'P:"cx,)} = 0 .

 154

The associated first order system (2.2), (2.4) is given by

E'Pz =X
2Em
(2.5) EXz = -O'ECX - g('P) - - - U
0"0

The travelling wave for (2.2), (2.4) corresponds to a heteroclinic orbit of (2.5) which
connects two distinct equilibria. The equilibria of (2.5) are given by the intersection
of two curves

2Em 1
(2.6) g('P)+-u=O and u-u!+-2('P-'P~cx,)=0.
0"0

It is evident from (2.6) that once the supercooling temperature Ucool( < 0) is given,
all the other data 'P~, u;, 'P:"oo are uniquely determined as intersecting points of
the cubic curve and the straight line with slope 1/2.
We are interested in a heteroclinic orbit connecting ('P~, 0, Ucool) to ('P:"oo, 0, us).
Note that

(2.7) 'P±oo = ±1 + E1]±(E) , Us = Ucool + 1+ Ep(E) ,

hold where 1]± (E) and p( E) are smooth and bounded up to E = O.
The strategy is as follows: first we solve (2.4) with respect to u, and substitute
the resulting formula into (2.2) to obtain a single equation for 'P. Applying the
alternative method to it, the whole problem is reduced to solving a bifurcation
equation with respect to the velocity C and E. The key is that the temperature field
U does "not" have a limit in 13(R)-space as E 1 0, however the scaled bifurcation
equation remains valid up to E = O.
For our purpose, it is convenient to use the stretched coordinate y( == Z/ E) and
the "inner variables":

(2.8) <I>(y) = 'P(EY) , U(y) == U(EY) .

Equations (2.2), (2.4) become

2Em
(2.9) <I>yy + O'Ec<I>y + g(<I» + - U = 0
0"0

(2.10)
 155

Solving (2.10) under (l.8) and substituting it into (2.9), we have

<1>yy + ac:c<1>y + g(<1»

(2.11 )

(2.12) <1>(+00) = 'P+oo , <1>y(±oo) = o.

Letting c: 1 0 and using (2.7), we have the reduced problem

(2.13) <1>yy + g(<1» = 0, <1>(±00) = ±1

which has a unique (up to translation) strictly increasing solution <1>0 (y) decaying
exponentially at infinity. It is natural to seek a solution of the form

(2.14) <1> = <1>0+'11.

The equation for <1> becomes

2Em
+- 0"0

w(±OO) = C:1J±(C:) , wy(±oo) = 0,

where G(w, <1>0) == g(<1>0 + w) - g(<1>0) - g'(<1>°)w. We seek a solution W of (2.15) in

the 'B-space, i.e.,

'Bk (Ii!) == {f : R -+ R has bounded and

continuous derivatives up to the k-th order}

k
with the norm Ilfllk = LsupIDjf(z)l.
j=oZER

REMARK 2.1. The reason why we do "not" put <1>0 +c:w in (2.14) is that if we
do so, the resulting '11 does "not" have a limit in 'B°-space as c: -+ O.
d2
The principal linear part of (2.15)a is L == -12 +g' (<1>0) which has one-dimensional
(y
kernel in 'B0(Ii!) spanned by {<1>~}. This is a typical situation where we can use the
 156

alternative method (see, for instance, [6, Chap. 2]). Applying the usual procedures
of the alternative method to (2.15), we see that the problem is reduced to solving
the "bifurcation equation" with respect to c and E:

(2.16)

+ 2;~' {u; - ~ -L ,-\tC.-., ~ (<p"(,) H(",) - 'C~)d'} , <J>:) ~ 0,

where W(c, E) E N(L).L is the unique solution of (2.15)a projected to the range space
of L. Unfortunately (2.16) is degenerate in the sense that B(c,O) °
= = ~~ (c,O),
because we put q, = q,0 + W (not q,0 + EW) as in (2.14) which is necessary to obtain
a solution W in :Eo-space. However it can be proved that the scaled bifurcation
equation

(2.17) B(C,E) == B(C,E)/E

is well-defined up to E = °
owing to the exponentially decaying property of q,~.
Namely, although the integral

J
y

(2.18) ~; e- 1f (y-s) ~ (q,O(s) + W(C,E) - <P=-rx,)ds

-00

does not have a limit in 13°(R)-space as E 1 0, the inner product with q,~ has a
definite limit (=: 0), since (2.18) converges to zero uniformly on any "compact" set
of R when E 1 0. In fact, we can prove

PROPOSITION 2.1. There is a positive Eo such that B( c, E) and (8B / 8c)( c, E)

are well defined and continuous for (c, E) E I X [0, Eo], where I is an arbitrary given
bounded interval in R+. Moreover, one has

- 4m
B(c,O) = acm + - (Ucool + 1)
"'0

8B
Tc(c,O) = am > 0,
where m = 11q,~lli2(nl) .
Hence it follows from Proposition 2.1 and the implicit function theorem that
there is a unique continuous function c = c( E) with limiting velocity

c· == c(O) = ---±- (Ucool + I) > 0

00"'0

satisfying B(C(E),E) = °for E E [O,Eo]. This leads to Theorem in §l.

 157

REFERENCES

[1] G. CAGINALP, Stefan and Hele-Shaw type models as asymptotic limits of the phase field
equations, Physical Review A 39 (1989), 5887-5896.
[2] G. CAGINALP AND J. CHADAM, Stability of interfaces with velocity correction term, to appear
in Rocky Mountain J. of Math.
[3] G. CAGINALP AND P.C. FIFE, Elliptic problems involving phase boundaries satisfying a
curvature condition, IMA J. of Applied Math. 38 (1987), 195-217.
[4] G. CAGINALP AND P.C. FIFE, Dynamics of layered interfaces arising from phase boundaries,
SIAM J. Appl. Math 48 (1988),506-518.
[5] G. CAGINALP AND Y. NISHIURA, The existence of travelling waves for phase field equations
and convergence to sharp interface models in the singular limit, To appear in Quarterly of
Appl. Math.
[6] S-N. CHOW AND J. HALE, Methods of Bifurcation Theory, Springer, Berlin (1982).
[7] J.N. DEWYNNE, S.D. HOWISON, J.R. OCKENDON, \V. XIE, Asymptotic behavior of solutions
to the Stefan problem with a kinetic condition at the free boundary, J. Austral. Math. Soc.
Ser. B 31 (1989), 81-96.
[8] H. FUJII, Y. NISHIURA, M. MIMURA, R. KOBAYASHI, Existence of curved fronts for the phase
field model, In preparation.
[9] J.W. WILDER, Travelling wave solutions for interfaces arising from phase boundaries based
on a phase field model, Rensselaer Polytechnics Inst. Preprint.
 STANDING AND PROPAGATING TEMPERATURE WAVES
ON ELECTRICALLY HEATED CATALYTIC SURFACES*

GEORGIOS PHlLIPPOU AND DAN LUSSt

Abstract. Reaction rates are often measured using a catalytic wire or ribbon, the resistance
(average temperature) of which is kept at a preset value via electrical heating. Previous investi-
gators assumed that the wire temperature was uniform. An I R-thermal imager shows that some
temperature profiles have the shape of a stationary standing wave. A bifurcation map describes
the organization of the regions with these standing wave temperature profiles. In some cases, the
high temperature wave moves back and forth on the ribbon, with continuous changes in its shape.
This leads to both oscillatory and chaotic changes in the overall rate of heat generated by the
reaction. The dynamic behavior of the overall reaction rate is different and less regular than that
of local temperatures on the ribbon. The power spectrum of the overall rate of heat generation
decays exponentially, while that of the local temperatures decays as a power law. Ignoring the
nonuniform nature of the temperature of the ribbon may lead to severe pitfalls in the determination
of the kinetic rate expression and/or its parameters

Key words. catalytic surfaces, stationary temperature waves, propagating temperature waves.

1. Introduction. A large number of spatio-temporal patterns has been ob-

served in chemically reacting liquid phase systems, especially with the Belousov-
Zhabotinskii and the glycolysis reactions l . These include uni-dimensional excitable
and train waves, planar target patterns and spiral waves, and three dimensional
movements of scroll-waves. In recent years similar spatio-temporal patterns were
found in some heterogeneous catalytic systems. The group of Ertl has been able
to observe changes in surface structures or surface coverage on single crystals using
scanning low energy electron-diffraction, scanning photoemission and photoemis-
sion electron microscopy. They observed propagating and standing waves, rotating
spirals and chaos during the low pressure oxidation of carbon monoxide 2 ,3. The
groups of Schmitz4, WolfS and Schmidt 6 observed temperature waves on heteroge-
neous catalytic surfaces on which chemical reactions were carried out at atmospheric
conditions. We explain here how electrical heating of catalytic surfaces may gener-
ate both stationary and propagating temperature waves.

2. Standing Temperature Waves on Catalytic Surfaces. Electrically

heating wires and ribbons have been used for many years to study the multiplicity
features of catalytic systems. The electrical heating is used to keep the resistance
and hence the average temperature of the wire at a preset value. It was assumed
in most previous studies that the temperature of these heated wires was uniform.
Hence, bifurcation diagrams of the heat generated by the reaction vs. wire temper-
ature were used to estimate the kinetic rate constants and/or the functional form
of the rate expression.
Figure (1) describes a typical experimental bifurcation diagram of heat gen-
erated vs. average wire temperature obtained during the oxidation of a mixture

*We are thankful to the National Science Foundation and the Welch Foundation for support of
t his research.
tDepartment of Chemical Engineering, University of Houston, Houston, Texas, 77204-4792
 160

containing 1%, ammonia in air. The catalyst was a 14.7 cm long, 0.05 cm wide
by 0.0025 cm thick pure platinum ribbon. A hot wire anemometer (TSI, IFA-100)
maintained the ribbon at a preset total resistance (and therefore constant average
temperature, T). The temperature profiles were measured by an infra-red imager
(AGEMA thermovision 780), mounted on a motorized table and driven parallel to
the ribbon. Additional details of the experimental system can be found in 7 .

1.

.8
1.0%NH 3 •
•
N

-
E .6
u
II)

2-
Cl .4
d

.2

0
100 150 200 250 300
Average Temperature (OC)

Figure 1: Bifurcation diagram of the heat generated by oxida-

tion of a mixture containing 1% ammonia vs. average ribbon
temperature.

The experiments showed that a uniform, extinguished state with a negligible

rate of heat generation existed for all temperatures below 195°C. A fully ignited
branch of solutions existed for all T > 226°C. The temperature profiles of these
states (see Fig. 2, T = 252 and 282°C) were uniform with the exception of a narrow
end region next to the supports.
A branch of nonuniform temperature (NUTR) states existed for all T in (160,
240°C). The temperature profiles of these states were usually non-symmetric about
the center of the ribbon and consisted of two segments of different but rather uni-
form temperatures, separated by a narrow front. Increasing the average ribbon
temperature increased the fraction maintained at the high temperature. Figure 2
describes some profiles obtained by a slow continuous increase in the average wire
temperature. The standing temperature waves remained stationary for very long
periods of time as any movement of the front changed the electrical resistance of the
ribbon. The anemometer immediately changed the electrical current to counteract
this movement and restored the temperature front to its original position. Thus,
 161

-
0 300
T avg =282 °C
1.0 % NH3

-...
0

-...
(1)

:J
as
(1) 200
c.
E
(1)
I-

100
-8 -4 0 4 8
Position ( em )

Figure 2: Temperature profiles of the ribbon obtained as its

average temperature is slowly increased.

the electrical heating generated and stabilized these standing temperature waves.
The experiments showed that either an extinguished or a NUTR existed for T
in (160, 195°C). The two states which can exist for 195°C are shown in Fig. 3.
Similarly, either an ignited or a NUTR existed for T in (226, 240°C). Only a NUTR
existed for Tin (195, 226°C).
Repeated experiments showed that the branch of NUTR did not emanate from
either one of the two uniform temperature branches. A sub-critical bifurcation
occurred at both the ignition and extinction points of the uniform states. The
exact transition from the NUTR to a branch of uniform states was difficult to
locate due to the proximity of the two branches.
A bifurcation map of regions with qualitatively different states (Fig. 4) was
constructed from the bifurcations diagrams for several feed concentrations. The
map shows that five qualitatively different regions exist, in three of which a single
state exists (either ignited, extinguished or nonuniform) while in two regions two
different types of states can be obtained. We were not able to determine how the
boundaries of the various regions coalesce at low ammonia concentrations, due to
experimental inaccuracies in that range of operation.
In modeling this system, we assume that the local surface concentration is in
equilibrium with that of the gas. A species balance gives:

(1.1)
where kc is the mass transfer coefficient, Cb the bulk reactant concentration, and
the reaction rate, r, depends on the local concentration (Cs ) and temperature (T).
 162

_250
0
0

-200

-
(l)
a..
::::J
a:s
a..
(l) 150
Co
E
(l)
1.0 % NH 3 • Tavg = 195°C
t-
100
-8 -4 0 4 8
Position ( em )

Figure 3: A uniform and NUTR temperature profiles having

the same average temperature of 195°C.

This equation may be brought to the form:

(1.2) Cs=g(T,p)

where p is a vector of parameters. Thus, the local reaction rate can be expressed
as a function of the local temperature. The energy balance for the ribbon is:

(1.3)

where

(1.4) f(T) = a[( -~H)r(T) - h(T - Ta)] + [2 R(T)

and A is the cross section of the ribbon, a the perimeter of the ribbon, ). the effective
thermal conductivity, h the heat transfer coefficient, [ the electrical current, and R
the ribbon resistance per unit length. Usually, we can express R as a linear function
of the temperature, i.e.,

(1.5) R(T) = R(Ta)[l + Cl!(T - Ta)]

It is well known that a standing temperature wave may exist in an infinite medium
(one with no end effects) only irs

(1.6) iT,
T.
f(T)dT =0
 163

-
U
'L
....
III
::J
.....
....C'C
III
a.
E
III
I-
III
Cl
....
C'C
III

Figure 4: A bifurcation map describing regions with quali-

tatively different steady states; E refers to low temperature
(extinguished) state, I high temperature (ignited) state, and
N to a nonuniform temperature state.

where Tl and T2 are the uniform temperatures on both sides of the front. The
uniform temperature states satisfy the relation J(T) = 0, which may be written
as 9 ;

(1.7)

where:

(l.8) Qg = (-!:l.H)r(T)

QR = -~ +(h - aR(~a)I2) (T - Ta +~)

Qg, the heat generated by the reaction is usually a sigmoidal function of the ribbon
temperature. Q R, the heat removal term depends also on the electrical heating
and is represented by a family of straight lines originating at T = Ta - 1/ a and
QR = -h/a. The intersections between QR and Qg are the uniform temperature
states (see Fig. 5).
Inspection of Fig. 5 shows that a stable uniform state cannot exist for any
temperature in the range (L, H). An at tempt to get such a state leads to formation
of a NUTR with the temperatures at the two sections of the ribbon being A and
C. The corresponding heat removal is the line ABC, for which the two-hatched
areas ALBA and CHBC are equal, so that Eqn (1.6) is satisfied. The fraction of
 164

Figure 5: Graphical solution of Eqn. (1.7).

the ribbon which is at the high temperature is determined by the present average
temperature.
The above analysis suggests that the current required to maintain a NUTR
is independent of the average wire temperature. Experiments described in Fig. 6
indicate that the required current increases with average temperature. One possible
explanation is that heat losses from the supports increase with increasing average
temperature. The data show that the current needed to maintain a NUTR differs
from that needed to keep a uniform state at the same average temperature.
Failure to account for the presence of NUTR may lead to severe pitfalls in the
analysis of kinetic experiments. Many previous investigators used the multiplicity
features of electrically heated catalytic wires to determine kinetic parameters, as-
suming the wires have a uniform temperature. The experiments suggest that this
is a very dangerous procedure. The bifurcation map (Fig. 4) shows that if one is
not accounting for the existence of the NUTR, he may interprete the data as in-
dicating the existence of two separate regions, each with two uniform temperature
stable steady states. Clearly, this is not the case. Moreover, when the boundaries
of the region with the extinguished and NUTR states have different slopes in the
(T, C) plane, one may erroneously conclude that a pitchfork singularity exists. We
conjecture that the analysis of the multiplicity features of many electrically heated
wires is erroneous as the possible existence of a NUTR was not considered.

3. Propagating Temperature Waves on Catalytic Surfaces. Experiments

indicate that for certain chemical reactions and concentrations stationary temper-
ature fronts are not formed on electrically heated catalytic surfaces. Instead, a
 165

-0
c.
1.05

--
E
<C 1.
cCI):
~
~ .95
:::J
0
c: .9
0
.c
.c
a: .85
100 150 200 250 300
Temperature ( °C )

Figure 6: Electrical current needed to keep the ribbon at vari-

ous average temperatures.

back and forth propagating temperature wave is obtained. Fig. 7 is a bifurcation

diagram of heat generated vs. average wire temperature for the oxidation of a
mixture containing 0.2% propylene in air. In this case, the overall rate of heat
generation oscillated for temperature exceeding 290°C. The amplitude of these os-
cillations decreased with increasing temperature and became negligible at about
400°C. The heat generation is chaotic for average ribbon temperatures between
228°C and 260°C. An intermediate behavior was observed for average tempera-
tures between 260°C and 290°C. Typical time traces of the heat b"neration in these
three temperatures ranges are shown in Fie:. 8.

1. I I I

.6 r- I CiT I 0 ",. ..... -

N

E
.6 I- tt tt -
u

dl
.!!1
~ .4 I- 0
111 T Oscillatory -
'"
i
0 Transition

.2
r II C Chaos -
~
0.2 % Propylene

o. I I I I
200 250 300 350 400 450
Average Temperature (Oe)
Figure 7: A bifurcation diagram of the heat generated by the
oxidation of a mixture containing 0.2% propylene vs. average
ribbon temperature. Bars denote amplitude of oscillations.
 166

0.60r-.-----~----r-----~--_,----~----~

l
0.2 % Propylene, Tavg= 294 • C

C\I 0.48

::::~L--7....L5-0----..L----1-:.L~-~-=-2-6-9'-.-c---12....15-0----...L----1--150 0
'-' 0.33~~ __
__ ____ ____
~___ __ _____
~ _ ___
______
~__ ___~ ~ ~ ~ ~

01 0 500 1000 1500 2000

" 0'24~T.V9=229.C
0.22
0.20
0.18L------..L-----....I-----~-----L----~------I
500 750 1000 1250
Tim e (sec)
Figure 8: Transient rate of heat generated on a ribbon at the
oscillatory, transition and chaotic regions for a mixture con-
taining 0.2% propylene

Temperature profile measurements showed that the variation in the rate of heat
generation were due to a back and forth movement of a high temperature zone and
the temporal changes in its length (Fig. 9), as the rate of heat generation in the
low temperature region is negligible. The movement of the high temperature zone
in the oscillatory state consisted of five stages. Initially, the ignited section was
stationary on the left side of the ribbon (profile a, Fig 91). This zone then moved
to the right, with its back moving somewhat faster than the front, reducing the
length of the ignited zone and the overall rate of heat generation. The front of
the zone stopped its rightwards movement before reaching the support. The wave
then reversed its direction of movement, and its front moved to the left faster than
its back (profile b, Fig. 91). Eventually, the back of the wave moved faster than
its front, reducing the length of the reaction zone and the overall rate. The front
stopped moving leftwards before reaching the support, and the wave then started
moving to the right (profile c, Fig. 91) until it became profile a. The velocity of the
moving fronts was of the order of 0.1 cm/sec.
The transient burst of irregular heat generation in the transition region (Fig.
8, T = 269°C) was caused by a sudden splitting of the ignited region into two
sections, which moved in the same direction at different velocities (profile d, in Fig
9Il). Eventually, the two zones collided and coalesced into one (profile a, Fig gIl),
returning the ribbon to an oscillatory mode.
At low wire temperatures (228°-260°C) the width of the ignited zone was
smaller than that of the low temperature zone, and its movement (Fig gIll) led
to a chaotic variation in the rate of the heat generation. The distinction between
chaotic and oscillatory behavior is due to changes in the nature of the movement of
 167

- 400

300

-
()
0
200
Q) 100

-
~ 400
::l 300 ,,
~ ~"
d\ ........ ~- . . ".... ~ a ,,
CO 200 .......... - .... ,
,,
~

Q) 100
400
Co ~ III
E 300

....
Q) 200

100
-8 -4 0 4 8
Position (em)

Figure 9: Transient temperature profiles observed during the

oscillatory (I), transition (II) and chaotic (III) operation for a
mixture containing 0.2% propylene.

the high temperature zone.

Experiments indicate that the chaotic nature of the behavior tends to become
less intricate (lower correlation dimension) with increasing concentration of the
propylene. At sufficiently high reactant concentrations (such as 0.5%) propylene
the temperature waves become stationary.
The nature of the overall heat generation was analyzed by constructing the
attractor of the dynamic behavior using the time delay method. Inspection of the
attractor for 0.3% propylene (Fig. 10) shows that the chaotic motion is either around
one of two unstable states, or around both states.
The chaotic temperature waves were observed at various 1 cm sections of the
ribbon by keeping the IR camera stationary. Fig. 11 shows temperature profiles
observed at a section located between 3.3 and 4.3 cm from the right hand support,
close to the position at which the temperature wave reversed its direction. The data
shows that different waves reversed the direction of movement at different positions.
The power spectrum of the chaotic heat generation signal decayed exponentially
for high frequencies. Sigeti and Horsthemke 10 suggested that this implies that the
system is deterministic. On the other hand, the power spectra of local temper-
atures decayed as a power law for high frequencies (straight line on logarithmic
scales) implying a stochastic behavior. Moreover, the oscillation frequency of the
 168

0.3 % Propylene, Tav g=225 · C

Figure 10: A reconstruction of the chaotic attractor using a

time delay of 5 seconds.

Figure 11: Temperature waves observed at a 1 cm section of

the ribbon for a mixture containing 0.3% propylene and 'if of
243°C.
 169

chaotic overall rate was higher than that of local temperatures. The motion of the
temperature waves has many similarities to the electrical current waves observed by
Lev et.all l during the electrochemical dissolution of nickel in a galvanostatic mode.
Many previous theoretical studies of oscillatory behavior of catalytic surfaces
assumed that the surface is homogeneous. The experimental results indicate that
this is often not the case, and that intricate wave motion is the cause of the complex
dynamics of the overall reaction rate. It would be very useful to develop some
criteria or diagnostics prediding when the observed rate is due to temporal behavior
and when it is caused by a spatia-temporal process.
Propagating waves may be formed by several different mechanisms such as an
instability of the high temperature state or periodic read ion induced changes in
the catalytic properties of the surface. Sheintuch 12 presented a simple model in
which a slow adivation-deadivation process can generate spatio-temporal behavior
on an electrically heated wire. It is of much interest to determine the adual chem-
ical mechanism which creates the various dynamic features, the types of possible
bifurcations and singular points, and the different qualitative behaviors. From a
pradical point of view, one needs to know how robust or sensitive these patterns
are to slight variations in the operation conditions or catalyst. A key question is
whether, the dynamic features of the ribbon may be exploited to carry out readions,
which do not proceed under stationary conditions.

REFERENCES

[1] R.J. FIELD AND M. BURGER, Oscillations and traveling waves in chemical systems, Wiley,
New York, 1985.
[2] S. LADAS, R. 1MBIHL AND G. ERTL, Kinetic oscillations and faceting during the catalytic
CO oxidation on Pt (110), Sur. Sci., 198 (1988), pp. 42-68.
[3] S. JAKUBITH, H.H. ROTERMUND, W. ENGEL, A. VON OERTZEN AND G. ERTL, Spatiotemporal
concentration patterns in a surface reaction: Propagating and standing waves, rotating
spirals, and turbulence, Phys. Rev. Let, 65 (1990), pp. 3013-3016.
[4] P.C. PAWLICKI AND R.A. SCHMITZ, Spatial effects on supported catalysts, Chern. Eng. Prog.,
83(2) (1987), pp. 40-46.
[5] J .S. KELLOW AND E.E. WOLF, Infrared thermography and effects on surface reaction dy-
namics during CO and ethylene oxidation on Rhj 5i02 catalysts, Chern. Eng. Sci., 45 (1990),
pp. 2597-2602.
[6] G.A. CORDONIER AND L.D. SCHMIDT, Thermal waves in N H3 oxidation on a Pt wire, Chern.
Eng. Sci., 44 (1989), pp. 1983-1993.
[7] L. LOBBAN, G. PHILIPPOU AND D. LUss, Standing temperature waves on electrically heated
catalytic ribbons, J. Phys. Chern., 93 (1989), pp. 733-736.
[8] V. BARELKO, I.I. KUROCHKA, A.G. MERZHANOV AND lCG. SHKADINSKI, Investigation of
travelling waves on catalytic wires, Chern. Eng. Sci., 33 (1978), pp. 805-811.
[9] M. SHEINTUCH AND J. SCHMIDT, Observable multiplicity features of inhomogeneous sollitions
measured by the thermochemic method: Theory and experiments, Chern. Eng. Cornrn., 44
(1986), pp. 33-52.
[10] D. SIGETI AND W. HORSTHEMKE, High frequency power spectra for systems subject to noise,
Phys. Rev. A, 35 (1987), pp. 2276-2282.
[11] O. LEV, A. WOLFBERG, M. SHEINTUCH AND L. PISMEN, Bifurcations to periodic and chaotic
motions in anodic nickel dissolutions, Chern. Eng. Sci., 43 (1988), pp. 1339-1353.
[12] M. SHEINTUCH, Spatio-temporal structural of controlled catalytic wires, Chern. Eng. Sci., 44
(1989), pp. 1081-1089.
 MIXED-MODE OSCILLATIONS IN THE
NONISOTHERMAL AUTOCATALATOR

S.K. SCOTT* AND A.S. TOMLIN*

Abstract. The internal coupling of chemical and thermal feedback processes is investigated
through a non-isothermal autocatalator model. The isothermal oscillatory scheme is augmented to
include an exothermic "termination" step and a temperature-dependent initiation: P -+ A (rate =
ko(T)p); A+ 2B -+ 3B (rate = klab 2 ); A -+ B (rate = k3a); B -+ C+ heat (rate = k 2 b). The pool
chemical approximation is applied to reactant P. Multiple stationary states exist for all parameter
values and complex dynamic behaviour is also found over wide ranges of these parameters. Path
following techniques are used to follow the periodic solutions and to find bifurcations such as
period doubling cascades leading to chaos. Other types of complex dynamics such as mixed mode
oscillations are categorised, although we find no quasiperiodic behaviour in the model. This gives
rise to the belief that the mixed mode oscillations found, do not stem from phase locking on a
torus in this instance.

1. Introduction. Mixed mode oscillations are characteristic of many non-

linear chemical systems, most notably in the Belousov-Zhabotinskii reaction [2-
4,8,13-18,22,23]. Typically, mixed-mode oscillations are complex periodic wave-
forms comprising a certain number of large and small amplitude excurions in some
particular order. As some bifurcation parameter, such as the flow rate, is varied
so the waveform may change to one with fewer or more small or large peaks. The
bifurcation sequences often give rise to concatenations obeying Farey arithmetic
and a plot of a suitably defined "firing number" against the bifurcation parame-
ter may lead to a Devil's staircase [17,18]. In the Belousov-Zhabotinskii reaction,
these responses have been interpreted as motion on a torus [7] in the appropriate
concentration phase-space.
Recently, Albahadily et al. [1] have reported mixed mode oscillations with
Farey sequences etc. but which apparently do not have an underlying torus struc-
ture. Here we report similar observations for a simple (but unrelated) three-variable
model, based on the cubic autocatalator scheme.

2. The model. The cubic autocatalator model [19] in its simplest, robust form
involves the conversion of a pool chemical reactant P to a final product C via two
intermediates A and B. The interconversion between A and B proceeds via two,
simultaneous routes: an uncatalyzed step, and a process involving autoacatalysis.
The model scheme can thus be represented as

(0) P-+A rate = kop

(1) A+2B -+ 3B rate = klab 2
(3) A-+B rate = k3a
(2) B-+C rate = k 2 b

*Department of Physical Chemistry, University of Leeds, Leeds LS2 9JT, U.K.

 172

(the numbering is chosen for consistency with earlier work). Many, if not most,
spontaneous chemical reactions are exothermic. The heat released may cause an
increase in the temperature (self-heating) of the reacting mixture. This offers an-
other mechanism for feedback, as reaction rate constants generally vary with the
local temperature. Rather than going for generality, however, we seek here to in-
troduce self-heating and its effects in a simple form: the reactions with greatest
temperature dependences (the highest activation energies) are typically the initia-
tion steps, such as (0) in the above scheme. Similarly, the most exothermic processes
are usually the chain termination steps in which intermediates are converted to sta-
ble products, such as reaction (2). For our canonical model, therefore, we assume
that the rate constant ko has an Arrhenius temperature dependence

(2.1) ko(T) = Aoe -E/ RT

and that reaction (2) has a positive molar exothermicity, Q = -6.H > o.
The governing equations for this scheme are: for concentrations

(2.2) dp/dt = -ko(T)p

(2.3) da/dt = -ko(T)p - klab 2 - k3a
(2.4) db/dt = -klab 2 + k3a - k 2b

and for temperature

(2.5)

where Ta is the temperature of the surroundings. All other terms are defined in
the List of symbols at the end of this paper. The initial conditions for a given
experiment might be p = po, ao = bo = 0 and T = Ta at t = O.
Equations (2.2-5) may be case in an equivalent dimensionless form
(2.6) dfL/ dT= -ffLeD
(2.7) do:/dT = fleD - 0:;32 - KilO:
(2.8) d;3/dT = 0:;32 + KilO: -;3
(2.9) d8/dT = 8;3 -,8
Here, f = ko/k2 will be supposed to be small compared with unity. The term eD in
equation (2.7) is an approximate, but conveniently simple, representation of the Ar-
rhenius dependence. To reduce this set to a system of just three coupled equations,
we make the traditional pool chemical approximation regarding a slowly decaying,
large concentration of the reactant P. Formally, we require the reduced reactant
concentration fL to remain of order unity as f tends to zero. Then the integrated
J
form of equation (2.6), fL = floe -< e' dT where flo is the dimensionless initial con-
centration of P, is approximated by fL = flo for sufficiently small dimensionless time
T ~ 0(c 1 ).
 173

Thus, we study here the set of autonomous equations

(2.10) dOlldr = J1.oe B - 01.(32 - KuOl

(2.11 ) d(3ldr = 01.(32 + K/LOI - (3
(2.12) dBldr = 0(3 -,B

3. Results. Equations (2.10-12) involve four parameters: 110,0, Ku and,. Of

these, the dimensionless reactant concentration can be thought of as the most easily
varied during a given experiment (the principal bifurcation parameter): indeed, if
we remember the slow exponential decrease with finite £, an experiment would auto-
matically provide this variation, but we will not emphasize that point unduly here.
There remain, then, three unfolding parameters, perhaps to be adjusted between
successive experiments. Although, in principle, modern path following techniques
[9] are able to lighten the burden of searching through such parameter spaces, the
actual application is generally complicated by various operational difficulties - not
least by the wide range of bifurcation phenomena that even such a simple system
can support [12]. Some aspects, such as the stationary-states and the conditions for
Hopf bifurcation, can be treated generally, but otherwise we choose here to report
on a limited number of traverses through the behaviour of this scheme.

3.1 Stationary-states. The stationary-state solutions satisfying dOli dr = d(3 I dr

= dB I dr= 0 are given by

(3.1)

where (3ss is a solution of the equation

(3.2)

The qualitative form of the stationary-state locus is the same for all J1.0,0 and "
and is shown in figure 1. For J1.0 < ,IDe there are two stationary-state solutions,
one with (3ss < ,Id and the other with (3ss > ,Id. These two solutions merge when
oe
J1.0 = ,I and the system shows a thermal explosion for higher initial concentrations
of the reactant.
 174

{3

"t/8

-- - - --
-----------------------------r~M
"t/8 e

FIGURE 1. Steady state profile of {J against p, the bifurcation parameter.

Marked are the Hopf bifurcation points and the dotted line between them de-
notes a region of dynamic instability of the stationary state.

The classic behaviour of the isothermal autocatalator is regained in either limit

8 -+ 0 or 'Y -+ 00: with then the linear relationship /3•• = po showing no maximum
or runaway.

3.2 Hopf bifurcation. The local stability of a given stationary-state /3•• is

determined by the sign and character of the three eigenvalues ).,1-3 given as roots
of the following cubic equation

(3.3)

where

b = 1 + (3;.+ Ku + 'Y - 2(3;.((3;. + Ku)-I

e = ((3;. + Ku)(l + 'Y) + 'Y[1 - 2(3;.((3;. + KU)-I]
d = ((3;. + Kj)(-y - 8(3•• )

The model has a saddle-node bifurcation when d = 0 ((3 •• = 'Y/d as above). Hopf
bifurcations are prescribed by the condition be = d with b, c and d > O. Any Hopf
 175

points lie on the lower branch of solutions. According to the parameter values, there
may be two, one or no such bifurcations. We will be particularly concerned with
the first of these cases. The lower solution is unstable for some range of 11 given by
112 :::; 11 :::; Ili, with 112 > a and Ili < 'Y / be as shown in figure 1.

4. Specific Cases.

4.1 Period-1 oscillations and influence of uncatalyzed reaction. 'Y =

1.0, b = 0.1, Ku = 7 X 10- 3 • The simplest dynamic response for this model is that
of period-l oscillations. Figure 2( a) shows the emergence of period-l solution from
the Hopf bifurcation points for the system with the above parameter values and the
variation of amplitude across the oscillatory range. In all cases studied, the Hopf
bifurcations are supercritical yielding a stable limit cycle. For the present case, the
period-l solution is stable for all values of 11 between the Hopf points. There are
two sections of the curve for which the amplitude varies rapidly with 11: that near
the upper Hopf point will be of interest later.

If the uncatalyzed reaction rate constant is increased, the Hopf bifurcation

points move closer together, decreasing the range of stationary-state instability.
(In the isothermal autocatalator model, Hopf bifurcations exist only if Ku < 1/8.)
Figure 2(b) shows the variation in amplitude of the period-l limit cycle solution for
a system with Ku = 5 x 10- 2 : as well as having small range, the "canard" behaviour
noted above has been lost so the amplitude varies relatively smoothly with 11.

For smaller values of Ku on the other hand the period-l solution may bifurcate.
With Ku = 5.5 X 10- 3 a stable period-2 solution emerges as a Floquet multiplier for
the period-l limit cycle passes through -1. The period-doubling bifurcation here is
supercritical so the period-2 oscillation is stable. There is a second bifurcation at
which the period-2 disappears and the period-l regains stability (figures 2c and d).

If the uncatalyzed rate constant is decreased further, the character of the upper
period-doubling bifurcation (that at larger 11) changes. for Ku = 10- 3 , this becomes
a sub critical bifurcation, figure 2( e) and (f). An unstable period-2 solution emerges,
growing in amplitude as 11 increases. There is a turning point in the period-2 locus
and the limit cycle is stable along the upper branch before disappearing at the lower
period-doubling point. Such a system will thus show hard-excitation from period-l
into period-2 oscillations as 11 is decreased.
 176

lD.O-,-------------, 6.U-r-----------,
......... : (II)

............
~

00
: 0

..
00 0

···:
00 0
o 0
o 0
o 0

..
·: ;=;::::;=.,..-,..-.......,.......,r-I ---...
~

0.0 ~...... 0.0 +<=;;;..,..-r--r-r---.--,-.,----1

0.0 l.0 0.0 2.0
Il Il
lD.O , . - - - - - - - - - -_ _--, 9.0 ~
.......... ........
__~=_--------_.,

..-..- ..:
.............\ (C) (a)

........ ::
. .0

:o
· ..
:
\\ ""\:
···· ...
•

• 0

.... :
".

0.0.
:
:
\
.1. .... __ ............................................-
~
4.0 + __..,..__-,-__..,..__-r°..;:o_--;
-.i..
..
0.0 l.0 0.75 0.79
Il Il

............ .................,
25.0-,---- ------_.,25.0.-------------,
(e) (f)

................. ···········:1::
~

::
\
~ .-.. .i
f ii::
o •

:
: :!
I .... :
0.. :
:
°
f
~

. . . -.. . . .-~~~~. .J
'0 0
••.••• 0··
..........
I

0.0 0.0 ...................

0.0 l.0 0.75 0.R1
Il Il

FIGURE 2. A series of bifurcation diagrams showing variation in amplitude of periodic

solutions with 1-', for various values of reaction rate of constant K". These have been
calculated using Auto. In all cases 7 = 0.1: (a) K" = 7.0 x 10 - 3, stable period one
oscillation. (b) K" = 5.0 x 10 - 2, stable period one oscillation with smaller maximum
amplitude. (c) K" = 5.5 X 10 - 3, the period one oscillation now bifurcates at a supercritical
period doubling point. (d) K" = 5.5 X 10 - 3, close· up of the period 2 region in (c). (e)
K" = 1.0 X 10 - 3, subcritical period doubling bifurcation. (f) close IIp of unstable period

2 oscillation shown in (e).

4.2 Influence of heat transfer rate: chaos. Returning to the system with
/j K" = 5.5 X 10- 3 and "( = 1.0, we see Hopf bifurcation and period-doubling,
= 0.1,
as shown in fignre 2( c) and (d). If the heat transfer parameter "( is decreased,
 177

corresponding to slower heat loss, further complexity is found. Between I = 0.7

and 0.65, a second pair of period-doubling appear, as shown in figures 3(a) and
(b). Now stable period-4 oscillations exist over a narrow range of the precursor
concentration p,. Yet more period-doublings (bubbling) are rapidly introduced by
further decreasing I and by I = 0.5 a full cascade [10] to chaotic responses has been
achieved, figure 3(c).
1 5 . 0 , - - - - -_ _ _ _ _ _ _ _ _-.
10.0,----------------,

·············1·········.
o •
o •
o •
00 •
o •
o •
°°0 :
00 •

............
oOo~

-
.........................................................
0.0 +---,.---,---r--,--.,---i O.0f---r--.----,---r--,--..,...--I
0.65 ~ 0.80 0.68 ~ 0.75

10.0,-_ _ _ _ _ _ _ _ _ _ _ _-,

........ : :,:
••••••••• (c)

•••• : 0

...
••• : 0

:•
00 0

:
o
~,

"
o ••
: ~
:
:
• 0
• 0
• 00

............. --- .. ----- .. ---.------------~--

0.0
0.0 1.0

FIGURE 3. Series of period doubling bifurcations or 'bubbling', leading to chaois as the

heat transfer coefficient 7 is varied. In all cases K" = 5.5 X 10 - 3, 7 = 0.1. (a) 7 = 0.7,
a simple period 1, period 2, period 1 sequence. (b) By 7 = 0.65 a period 4 solution has
appeared. (c) By 7 = 0.5 a full period doubling cascade takes place and region of chaotic
solutions exists.

Representative time series (showing the autocatalyst concentration (3 as a func-

tion of time) are presented in figure 4. Considering a sequence with decreasing p"
following the upper Hopf point a period-l limit cycle emerges. This period-doubles,
and a typical period-2 trace for p, = 0.715 is given in figure 4(j). By p, = 0.710, a
period-4 oscillation (figure 4( i» is found. For p, = 0.708, the time trace (figure 4(h»
is chaotic. Figure 5( a) shows the corresponding at tractor in phase-space: the next
amplitude map, figure 5(b), has a sharp peak. A window of period-3 responses
emerges at smaller p, figure 4( f-g) and 5( c» before another region of aperiodicity.
In this lower chaotic range, figure 4( d-e) and 5( d), the next amplitude maps have
 178

much flatter maxima, as shown for It = 0.690 in figure 5(e). For It = 0.68870,
period-4 has returned, figure 5(£) : later period-halvings occur as It is decreased
further.

200 time 500

100k~liliW(lkkblli
101lLLlhluflLLLLLliJ
100~LLLllig)80~)

6°UWililililro~
FIGURE 4. Time traces for'Y = 0.7, i = 0.1, K." = 5.5 X 10 - 3, corresponding to the
bifurcation diagram in figure 3(c). (a) JA = 0.65, period 2. (b) JA = 0.687, period 4. (c)
JA = 0.6887, period 4. (d) JA = 0.69, chaotic trace. (e) JA = 0.695, chaotic trace with period 3
solution emerging. (f) p. = 0.696, period 3 window. (g) p. = 0.705, period 3. (h) JA = 0.708,
smaller amplitude chaotic trace. (i) JA = 0.71, period 4. (j) JA = 0.715, period 2.

In these cases the attractors are folded almost at right angles, so as to lie approx-
imately on the planes a = 0 and () = 0 respectively. The fast motion corresponds
to the sharp decrease in a and sharp pulse in fl. The decay of the autocatalyst
concentration is accompanied by a rapid increase in temperature. The system then
cools exponentially, on a longer timescale through Newtonian heat transfer: during
this period the concentration of A begins to increase again.
 179

(D)

,.. il
"<-' -------'>

(e)
a

FIGURE 5. Three dimensional diagrams of the attractors and limit cycles corresponding
to the time traces in figure 4. (a)p = 0.69, chaotic attractor showing characteristic folding.
(b) p = 0.696, period 3 limit cycle. (c) p = 0.708, chaotic attractor. (d) p = 0.71, period
4 limit cycle. (e) p = 0.715, period 2 limit cycle.

4.3 Mixed-mode oscillations. The variation of the period-1 oscillation with

j.lfor a system with Kn = 10- 3 , 'Y = 0.05 and 8 = 0.025 is shown in figure 7.
This solution is stable, except for the range 0.5 < j.l < 0.56. Again, we find a
period-doubling sequence as j.l is decreased (figure 8(m and n)). The time series
then becomes chaotic figure 8(1). The corresponding next-amplitude map again
has a sharp maximum as shown in figure 10(a). As j.l is decreased from 0.5587
to 0.5586, figure 8(k), there is a significant change in the waveform, although this
remains aperiodic. for the larger value of Il, the maximum value of fJ attained
is ca. 8: for j.l S; 0.5586 larger amplitude excursions appear, with fJ attaining
instantaneous values of ca. 20. We can now identify small, intermediate and large
oscillatory pulses. This change in waveform sees a further sharpening of the next-
maximum map, figure 10(b). Periodicity returns by j.l = 0.5580 : the time series,
figure 8(j), has a typical mixed-mode form with eight small peaks separating each
large excursion, a 18 pattern. The amplitude of successive small peaks increases
and there is then a period with (3 ~ 0 before the large peak.
 180

8.0 10.0
(0)
f·· .. (b)

•:'•
••

.I·•· .,.•
\

,..........
•
+
c \
+
c ••
•
-
x x
\
'.
.. -
0.0 0.0
0.0 )(n 8.0 0.0 Xn 10.0

FIGURE 6. Net maxima plots of X n +l VB X n , for {J which correspond to the chaotic aUrac-
tors at 7 = 1.0, ~ = 0.1 k,. = 5.5 X 10 - 3, (a) p = 0.69, (b) P = 0.708.

25.0....--_ _ _ _ _ _ _ _ _ _ _--.

......... •••••• '0~

...... ~
• 0o
o
o
o
o
o
o
o
o
o
o
o
••
••
0.0 ......--,............. ·=···;..:..;··"T··.:..;.···;,;..··;,;..··r··-···-·--r-·---·--·-...,\----r------1
;-=-.

0.0 0.80
~

FIGURE 7. Bifurcation diagram showing the period 1 limit cycle at 7 = 0.05, ~ =

0.025, k,. = 1.0 X 10-3. The open circles show the unstable region where more complicated
mixed mode oscillations arise. These are shown in figure 8.
 181

:]lllllll['' HI llllll l 'bi

'"'llll fll U'''llllll C

200 500

""Illllll, ' 'lllllll'

".0 ,I~ t LLLL"
.1 .1 " ",

mOlLU L"wfL L L'"

(k) 60 (I)

W"ll~'lhLILiI' Llk'kll 6.0

6.0 (m) (n)

FIGURE 8. A sequence of mixed mode oscillations and small amplitude chaos for 'Y =
0.05, 5 = 0.025, k,. = 1.0 x 10 - 3. (a) p = 0.498, large amplitude period 1. (b)
p = 0.50 1"3" pattern. (c) p = 0.505, 1"1"2" pattern. (d) p = 0.51,1"3" pattern. (e)
p = 0.52,1"1 2 pattern. (f)p = 0.53,1 2 pattern. (g) I' = 0.535, 13 pattern. (f)p = 0.53,1 2
pattern. (g) p = 0.535, 13 pattern. (h) p = 0.54, 13 14 pattern. (i) p = 0.55, 15 pattern.
(j) p = 0.5580, 18 pattern. (k) p = 0.5586, large amplitude chaotic trace. (I) p = 0 ..5587,
small amplitude chaotic trace. (m) p = 0.5595, period 2. (n) p = 0.56 period 1.
 182

a
(0)

a
(b)

(e)

FIGURE 9. Chaotic attractors and mixed mode limit cycle at (a) II = 0.5587, (b) II =
0.5586, (c) II = 0.5580 in figure 8. The view chosen shows the escape from and spiral-
shaped reinjection into the stable manifold.

The development of the attractor in phase-space through these changes of os-

cillatory form can be seen in figure 9(a-c). for f1. = 0.5587, the trajectory winds
around a small, relatively flat attractor appropriate to the small amplitude chaos
seen in the time series. This part of the flow is also evidenced with f1. = 0.5586
in figure 9(b). Now the winding also gives way to an occasional large excursion.
During this, the auto catalyst concentration falls virtually to zero: the temperature
excess then falls, exponentially, whilst the concentration of A incTeases. At some
point there is then a burst of conversion of A to B with a following temperature
rise. The trajectory returns to the original portion of the attractor by spiralling
 183

in from above. The motion on this attractor gives large amplitude chaos. With
J1. = 0.5580, figure 9( c), the original structure has resolved itself into a true limit
cycle with eight turns between the onset of the large amplitude excursion and the
re-injection.
The number of small peaks in the cycle decreases as J1. is made smaller : there
is a 14 response at J1. = 0.550 and a 13 at J1. = 0.535. In between such parent forms
there are higher-order concatenations, e.g. with J1. = 0.5400 we find a 1413 state,
figure 8(h). Similarly, there is a 1211 state at J1. = 0.520 (figure 8( e)) between the 12
at J1. = 0.530 and the 11 at J1. = 0.510 (figures 8(f) and Cd) respectively). Amongst
other waveforms found between the 11 and the simple period-l (10) are a (11)212
at J1. = 0.505 (figure 8(c)) and a 11(1°)2111° (or 3121) at J1. = 0.500 (figure 8(d)).

10.0
20.0
•
(6) (b)
I

.•..,
:,
•
••
+
c: : \ +
c:

,•.
x x
.I "
-" ....
•
Co
. 1\' . .
~
I •
0.0 0.0
10.0 0.0 20.0
0.0 )en )en

FIGURE 10. Next maxima plots of X,,+l vs x .. , for {J which correspond to the chaotic
attractors at 'Y = 0.05, i = 0.25, K" = 1.0 X 10 - 3, (a) I' = 0.5587, (b) I' = 0.5586.

5. Discussion. The model under consideration here comprises the interaction

of chemical and thermal feedback. The exothermicity of the reaction is indicated
through the parameter 8 whilst 1 is a measure of the heat transfer coefficient for
the system. With the particualr parameter values studies in this paper, we are
concerned mainly with reactions having only moderate exothermicities (8 = 0.1 or
0.025). Thus, much of the evolution is virtually isothermal. Only when chemical
autocatalysis produced a high concentration of the species B does the heat release
rate become significant. The sharp rise of the aut.ocatalyst concentration occurs on
a very short timescale : during this short period there can be little heat transfer so
the system departs from isothermal operation.
The actual temperature excursions can be evaluated. If the temperature of
 184

the reacting mixture is T and the ambient temperature Ta , the temperature excess
6.T = T - Ta is given by

Taking the results from figure 7, we see that with {j = 0.025, () < ca. 05. If we
assume an activation energy of 100 K J moll and an ambient temperature of 298K,
we have 6.T < ca. 4K. For the time series in figure 8(1), the temperature excess
will not exceed 2K. Clearly, then large non-isothermal effects are not required to
produce even the highest degrees of complexity in this scheme.
The origins of the internal forcing within the present model can be seen relatively
clearly. The autonomous oscillations in the concentration of B, which occur even
with {j = 0, lead to oscillations in the temperature. This then provides a forcing to
the chemical kinetics through the term /-leo. If the argument () does become large we
could approximate the exponential dependence with the expansion /-l(1 + () + ... ).
As () is an oscillatory function this form provides a similar forcing to that obtained
by, say, a sinusoidal non-autonomous external forcing such as /-l(1 + Asin(wr)).
The evolution of mixed-mode oscillations and their progression to chaotic be-
haviour in section 4.3 above shows some interesting features. If we consider here
the traverse through parameter space with increasing /-l we find that the bifurcation
from period-1 to mixed-mode at /-l = ca .05 is accompanied by a Floquet multiplier
for the limit cycle leaving the unit circle rapidly through -1. We have not found ev-
idence for complex Floquet multipliers in this system so there is no quasiperiodicity
or bifurcation of the limit cycle to a torus. Recent discussions of the relationships
between mixed-mode waveforms and toroidal behaviour have been given by Barkley
[5,6], Schell [21] and others [11]. Maselko & Swinney [16-18] have observed periodic
states similar to that shown in figure 8(b) with /-l = 0.5, which they have interpreted
as frequency-locked motion on a four-torus. In our system there are, however, only
three variables.
The spiralling nature of the mixed-mode attractor can be related to the eigenval-
ues of the stationary-state which are constituted by a weakly attracting eigenvalue
Al and a complex pair with positive real parts (these real parts of higher magnitude
than At). At slightly higher values of the reactant concentration, there appears to
be a homoclinic tangency, but because of the relative magnitudes of the attracting
and repelling eigenvalues, this must be of a different form from the familiar man-
ifestation of Shilnikov homoclinicity [12] such as that found in the Lorenz system
[24]. The sequence of mixed-mode --+ large amplitude chaos --+ small amplitude
chaos --+ small amplitude oscillations does not appear to be characterized by any of
the current scenarios for the development of chaos, but has been reported recently
by Albahadily et al. [1] in a different (experimental) situation.
 185

List of Symbols

A intermediate in reaction scheme

a concentration of species A
Ao pre-exponential factor in Arrhenius temperature dependence
B autocatalytic intermediate
b concentration of species B
cp molar heat capacity of mixture
Co molar density of mixture
C final product
E activation energy
ki rate constant of i-th step
P precursor reactant
p concentration of P
Q exothermicity of reaction
S surface area across which heat is transferred
T temperature of reacting mixture
Ta ambient temperature of surroundings
V reactor volume
0: = (kI/k2)1/2a, dimensionless concentration of A
(3 = (kI/k2)1/2b, dimensionless concentration of B
I = XS/k2cco V, dimensionless heat transfer coefficient
{j = Q( k2/kI)1/2 E / cpcoRT;, dimensionless exothermicity
E: = ko / k2' dimensionless precursor decay rate const ant
() = (T - Ta)E / RT;, dimensionless temperature rise
"u = k3 / k2' dimensionless uncatalyzed reaction rate constant

jl = kok~/2pO/k;/2, dimensionless precursor concentration

X surface heat transfer coefficient
 186

Acknowledgements. We are grateful to Professor K. Showalter and Mr. Do Peng

(University of West Virginia) for helpful discussion; SKS thanks NATO (grant no.
0124189) and the Institute for Mathematics and its Applications, University of
Minnesota for financial support.

REFERENCES

[1] F.N. ALBAHADILY, J. RINGLAND AND M. SCHELL, J. Chern. Phys., 90 (1989), pp. 813-821.
[2] F. ARGOUL, A. ARNEODO, P. RICHETTI AND J.C. Roux, From quasiperiodicity to chaos in
the Belousov-Zhabotinskii reaction I and II, J. Chern. Phys., 86 (1987), 3325-3356.
[3] F. ARGOUL, A. ARNEODO, P. RICHETTI AND J.C. Roux AND H.L. SWINNEY, Chemical chaos
: frorn hints to confirmation, Acc. Chern Res., 20 (1987) 436-442.
[4] K. BAR-ELI AND R.M. NOYES, Computations simulating experimental observations of com-
plex bursting patterns in the Belousov-Zhabotinskii system, J. Chern. Phys., 88 (1988),
3646-3654.
[5] D. BARKLEY, Near critical behavior for one-parameter families of circle maps, Phys. Lett., A
129 (1988), 219-223.
[6] , Slow manifolds and mixed-mode oscillations in the Belousov-Zhabotinskii re-
action, J. Chern. Phys., 89 (1988), 5547-5599.
[7] D. BARKLEY, J. RINGLAND AND J. TURNER, Observations of a torus in a model of the Be-
lousov-Zhabotinskii reaction, J. Chern. Phys., 87 (1987), 3812-3820.
[8] P. BERGE, Y. POMEAU AND C. VIDAL, Order within chaos, Wiley, New York, 1984.
[9] E. DOEDEL, AUTO: continuation and bifurcation problems in ordinary differential equations
(1986).
[10] M.J. FEIGENBAUM, Universal behavior in nonlinear systems, Los Alarnos Sci., 1 (1980), 4-27.
[11] V. FRANCESCHINI, Bifurcations of tori and phase-locking in a dissipative system of differential
equations, Physica, 6D (1983), 285-304.
[12] J. GUCKENHEIMER AND P. HOLMES, Nonlinear oscillations, dynamical systems and bifurca-
tions of vector fields, Springer, New York, 1983.
[13] J.L. HUDSON, M. HART AND D. MARINKO, An experimental study of multiple peak periodic
and nonperiodic oscillations in the Belousov-Zhabotinskii reaction, J. Chern. Phys., 71 (1979),
1601-1606.
[14] J.L. HUDSON AND J.C. MANKIN, Chaos in the Belousov-Zhabotinskii reaction, J. Chern.
Phys., 74 (1981) 6171-6177.
[15] R.D. JANZ, D.J. VANACEK AND R.J. FIELD, Composite double oscillation in a modified version
of the Oregon at or model of the Belousov-Zhabotinskii reaction, J. Chern. Phys., 73 (1980),
3132-3138.
[16] J. MASELKO, Experimental studies of complicated oscillations, Chern. Phys., 51 (1980),
473-480.
[17] J. MASELKO AND H.L. SWINNEY, Complex periodic oscillations and Farey arithmetic in the
Belousov-Zhabotinskii reaction, J. Chern. Phys., 85 (1986), 6430-6441.
[18] , Phys. Lett., A 119 (1987), 403--406.
[19] J.H. MERKIN, D.J. NEEDHAM AND S.K. SCOTT, Proc. R. Soc. Lond., A 406 (1986), 299-323.
[20] O.E. ROSSLER, An equation for continuous chaos, Phys. Lett., A 57 (1976), 397-398.
[21] M. SCHELL, AND F. ALBAHADILY, J. Chern. Phys., 90 (1989) 821-829.
[22] R.A. SCHMITZ, K.R. GRAZIANI AND J.L. HUDSON, Experimental evidence of chaotic states in
the Belousov-Zhabotinskii reaction, J. Chern. Phys., 67 (1977), 3040-3044.
[23] K. SHOWALTER, R.M. NOYES AND K. BAR-ELI, A modified Oregonator model exhibiting
complicated limit cycle behavior in a flow system, J. Chern. Phys., 69 (1978), 2514-2524.
[24] C. SPARROW, The Lorenz equations: bifurcations, chaos and strange at tractors, Springer,
New York, 1982.
 BIFURCATIONS AND GLOBAL STABILITY IN SURFACE
CATALYZED REACTIONS USING THE MONTE CARLO METHOD*

D.G. VLACHOS, L.D. SCHMIDT AND R. ARIS**

Abstract. The vector of state variables characterizing the dynamics of heterogeneous systems
is often spatially nonuniform. Ordinary differential equations (ODEs) or partial differential equa-
tions (PDEs) can describe spatial and temporal evolution of nonuniform field systems only under
severe assumptions. The Monte Carlo method (MCM) is applied to model such heterogeneous
systems. The conditions under which ODEs or PDEs fail are examined by studying generic model
systems using the stochastic method. In the presence of nonlinearities in the governing det.er-
ministic equations, spat.ial inhomogeneities are observed. Multiplicities, cusp points, and periodic
solutions are calculated by systematic investigation in parameter space and the global stability
of solutions is presented. It is demonstrated that the presence of imperfections on surfaces in-
troduces local nonuniformities in the adlayer (nucleation centers) and their interaction on the
synchronization of the surface is discussed.

1. Introduction.
Studies of heterogeneous reactions in many applications such as catalysis, elec-
trochemistry, and chemical vapor deposition often reveal complicated behavior:
multiplicities in reaction rates, phase transitions, oscillations, and propagating or
standing waves [1-7]. The dynamics and interactions occurring at the microscopic
level can result in exotic macroscopic patterns varying from ordered structures, to
disordered structures, to turbulence so far associated only with fluids [1,7]. The
macroscopic structure developed by the microscopic interactions is related to the
catalyst activity and the overall performance of the system. Therefore, an under-
standing of the macroscopic patterns formed at atomic level is required along with
the performance obtained.
Most models consider spatially averaged concentrations of adsorbed species on
the catalyst surface by solving a set of ODEs [2,5,6]. However, the macroscopic
patterns developed during surface phenomena usually corTespond to a nonuniform
distribution of species on the surface [8,9]. Hence, a spatially independent concen-
tration field is in many cases an oversimplification of reality. Spatial inhomogeneity
of the surface controls the bifurcation behavior which therefore cannot be modeled
using ODEs [8]. The role of local fluctuations in concentration is also ignored by
ODE models, and its influence on the dynamics of the system is missing. An elu-
cidation of the mechanism for kinetic oscillations requires an investigation of all
processes affecting the reaction rate at a molecular level. This can be achieved by
using an atomic level model, such as the Monte Carlo method (MCM) [10].
Recently, the MCM was used to study the phase transitions associated with a
bimolecular reaction between A and B2 (one monatomic and one diatomic species)

*This research partly supported by NSF under grant No. CBT 9000117, a Graduate School
Fellowship, the Minnesota Supercomputer Institute, and the Army High Performance Computing
Research Center.
**Department of Chemical Engineering and Materials Science, University of Minnesota, Min-
neapolis, MN 55455.
 188

or A and B (two monatomic species) [9,11-13]. We also reported phase transitions,

multiplicities, and oscillations for a unimolecular reaction in a continuous stirred
flow reactor [8,14]. In our studies, the roles of surface diffusion and defects on phase
transitions and transient path of the system were examined.
In this paper, an outline of the deterministic theory used to describe the dy-
namics of heterogeneous systems is first presented. Then, the application of the
MCM to surface science problems is explained. The advantages and disadvantages
of utilizing the MCM to model the bifurcation behavior and stability of solutions
in heterogeneous systems are discussed. Deviations between the deterministic and
the stochastic model are reported, and the parameters affecting the deviations are
identified. This is achieved by simulation of two model systems: unimolecular
and bimolecular reactions. The role of spatial inhomogeneities introduced by ab-
sorbate interactions and the presence of surface imperfections on the multiplicity
and stability of the solutions is demonstrated. The relation between fluctuations
and oscillations is examined, and regimes of parameter space where fluctuations
become important are discussed.

2. Mean Field Theory for Nonlinear Dynamics in Surface Reactions.

2.1 Spatially homogeneous systems. Since early 1960's, the common ap-
proach in modeling the dynamics of surface reactions is to use a lumped-parameter
system described by a set of ODEs

dy
(1) dt =F(y,p)

which is known as mean field (MF) or Bragg-Williams approximation. Here, F is

the vector field (assumed to be differentiable) describing the kinetics in the adlayer,
y is a vector of n state variables (phase space) such as concentrations of species
involved in the surface reaction, gas pressure, and/or surface temperature, and p
is a vector of parameters in m-dimensional parameter space. Most surface reaction
systems involve many parameters, but usually, only a few of them are varied over
a wide range. Thus, the dimension m of parameter space is usually 2 or 3.
Insight into the importance of the interplay between the spatial structure and
the dynamics of the system as well as comparison of the MF theory with the MCM
can be gained by modeling generic model systems. The first model system examined
(presented here in more detail) is a unimolecular reaction in a continuous stirred
tank reactor (CSTR). The dynamics in that case are governed by the equations
[5,14]

de
(2) dt = ra - rd - rR

for the surface concentration e, and

(3)
 189

for the partial pressure PA of reactant A. Here, TR is the residence time in the
reactor, PA,o is the pressure of reactant at the inlet, p. is a parameter expressing
the ratio of catalyst surface to reactor volume (surface capacity), and r a , rd, rR are
the adsorption, desorption, and reaction rates.
In the presence of interactions of strength w (in our formulation, attractive
interactions imply w > 0), Eq. (2) is written

(4)

where z. is the number of neighbors of each surface site, k is the Boltzmann con-
stant, T is the temperature, and ka, kd' k R are the adsorption, desorption, and
reaction rate constants. In this model system, the effect of the exponential nonlin-
ear desorption term on the inhomogeneity of vector y and the consequences on the
kinetics are investigated.
The second model system represents a bimolecular reaction between two species
A and B [11,12J

dB A
(5) dt = kaA PA(1- BA - BB) - kdABA - kuBABB
dB B
(6) dt = kaB PB(1- BA - BB) - kdBBB - kRBABB

(known as Langmuir-Hinshelwood kinetics) where the subscripts denote the corre-

sponding species. In this system (hereafter called AB model), the partial pressures
in the gas phase are assumed to be uniform in space and constant in time. More
complicated kinetics can be appropriate for desorption (as in the first model) or reac-
tion. However, deviations between the deterministic (outlined above) and stochastic
model can be observed caused by the simple nonlinear reaction term appearing in
Eqs. (5) and (6).

2.2 Bifurcation analysis and stability. As steady state, the right hand side
of Eq. (1) will be equal to zero. The stationary solutions (fixed points) of

(7) F(y,p)=O

are classified based on the eigenvalues of the linearized Jacobian around the fixed
point

(8)

and the stability is determined by the trace and the determinant of the Jacobian
[5J

(9) det J > 0

(10) trJ < O.
 190

In a two-dimensional phase space, there are five kinds of fixed points: turning point
(zero eigenvalue), stable or unstable point (real non-zero eigenvalue), and stable or
unstable foci (complex conjugate eigenvalues). In addition, limit cycles describing
periodic solutions (closed curves in phase space) are frequently found. Complicated
nonstationary solutions are often observed ranging from regular to irregular to chaos
[1].
It is often possible to express the n - 1 elements of vector y as an explicit or
implicit function of the nth element. This provides n -1 relations which can be used
so that the steady state is described by a single equation, f(y, p) = O. The steady
state manifold in (y, p) space can be defined as a set of all points (y, p) satisfying
this equation. A singular point (yO, pO) is said to be of co dimension k if it satisfies

(11) f(yO, pO) = X (yOpO) = ... = ~:{ (yO, pO) = 0

d k +1 f
(12) dyk+l (yO, pO) i= O.

Singular points of codimension 1 and 2, known as fold and cusp points respectively,
will be discussed next. Considering one parameter of the m dimensional vector p
as the bifurcation parameter A, the steady state equation is written

(13) fey, Ai p) = 0,

and a bifurcation diagram of y vs. A can be generated and compared with experi-
mental data.
For example, construction of a single equation for the unimolecular reaction
model (Eqs. (3) and (4)) results in the following equation

In a differential reactor (TR -+ 0), the above equation becomes (hereafter called
isotherm)

(15)
KP = B[kR/kd + exp( -zsBw/kT)]
A 1-B
where K = ka/kd is the equilibrium adsorption-desorption constant.

2.3 Spatially inhomogeneous systems. Application of Eq. (1) requires that

all variables are uniform along the surface, i.e. there exists an infinite diffusion co-
efficient of Fickian type so that perfect "mixing" is accomplished in the adsorbed
layer. This can be achieved in a homogeneous phase, as for example the Belousov-
Zhabotinsky reaction in a liquid phase [15]. However, it seems very difficulty to con-
trol the "stirring" on a catalyst surface (heterogeneous system). As a consequence,
most experimental and industrial heterogeneous catalytic systems are innately dis-
tributed, with concentration and/or temperature gradients. Spatial inhomogeneities
 191

are usually caused by interactions between the adatoms, surface imperfections (dis-
locations, grain boundaries, point defects), nonuniform distribution of impurities,
or different crystallographic planes present (as in supported catalysts). The vector
F is a set of nonlinear functions owing to the cooperative phenomenon of interac-
tions, inhibition, or autocatalysis [2,5,6]. This phenomenon determines the spatial
nonuniformities on the surface.
Despite the nonuniformity of vector field y along the surface, very few studies
have focused on spatial characteristics of oscillations and steady states. More recent
studies include the thermographic imaging performed by Schmitz et al. [16,17], the
in situ scanning LEED and the photoemission electron microscopy [6,18,19]' and the
photodiode monitor of temperature in methylamine decomposition on electrically
heated wires by Cordonier et al. [20].
Distributed systems in surface reaction systems can be described by PDEs of
the form

(16)
ay
at = -v· ( - D . y) + F(y, p)
where -D . Vy is the diffusive flux and D is the diffusion tensor. Eq. (16) is a
generalization of the well known second Fick's law for an open system.
Most models explaining oscillations in catalytic systems assume either explicitly
or implicitly some type of attractive interactions. In the presence of attractive inter-
actions, there is a net flux from a low concentration (single atoms) towards a high
concentration (clusters) caused by minimization of Gibb's free energy. In contrast,
Eq. (16), which uses a Fickian diffusion, predicts a flux from a high to a low concen-
tration. This implies that Fickian diffusion, as that described by Eq. (16), tends to
homogenize the distribution of species and in the limit of an infinite Fickian diffu-
sion, Eq. (1) would be more accurate. Therefore, Eq. (16) is structurally incorrect
in the presence of attractive interactions [21] and cannot predict the spatial and
temporal evolution of catalytic processes. In the following section, the modeling of
spatial and temporal patterns in heterogeneous systems using the MCM is outlined.

3. The Monte Carlo Method.

3.1 The Monte Carlo method as a stochastic simulation rHode!. The

MCM was developed by von Neumann, Ulam, and Metropolis at the end of the
Second World War at Los Alamos to study the diffusion of neutrons in fissionable
material. The name "Monte Carlo", chosen because of the extensive use of random
numbers in the simulation, was coined by Metropolis in 1947 and used in the title
of a paper describing the early work at Los Alamos [22]. However, the history of
MCM is even longer. In 1901, the Italian mathematician Lazzerini performed a
simulation by spinning round and dropping a needle 3407 times. He estimated 7r
to be 3.1415929. In the beginning of the century, statisticians had also used model
sampling experiments to investigate some problems (the correlation coefficients in
the "t" distribution, trajectories to study the elastic collisions with shaped walls).
The novel contribution of von Neumann, Ulam, and Metropolis was to realize that
 192

determinate mathematical problems could be treated by finding a probabilistic ana-

logue which was then solved by a stochastic sampling experiment [10].
The MCM is appropriate for study of phenomena in heterogeneous systems
because these phenomena are discrete and microscopic. MC simulations play a
valuable role in providing essentially exact results for problems in surface science
which would be otherwise only soluble by approximate methods or might be quite
intractable. In this sense, MC computer simulation is a test of deterministic theory.
MC simulations provide a direct route from the microscopic details of a system to
macroscopic properties of experimental interest (equation of state, transport coef-
ficients, structural order parameters, reactivity, and so on). It may be difficult or
impossible to carry out experiments under extremes of temperatures and pressure,
while a computer simulation would be perfectly feasible. Even though quite subtle
details of molecular motion and structure in heterogeneous catalysis are difficult to
probe experimentally, they can be extracted readily from a computer simulation.
Simulations are performed under well defined conditions so that the effect of dif-
ferent microscopic mechanisms on the macroscopic behavior of the system can be
elucidated. Finally, while the speed of molecular events is itself an experimental
difficulty, it presents no hindrance to computer simulation.
In the MCM, various states of the system are generated using random numbers
and weighted with appropriate probabilities. A macroscopic specimen of matter is
simulated by means of a model system containing a relatively small number of atoms
occupying elements of an array. To simulate a large system, periodic boundary
conditions are used. The value of matrix elements indicate the type of species
present at this element, if any. Thus the surface can be thought as a sparce matrix
where the degree of sparcity shows the portion of the surface which is unoccupied.
Each time an element of the matrix is randomly chosen and various probabilities are
assigned to each matrix element based on which processes are simulated. At each
point of parameter space, application of the MCM provides the degree of sparcity
which is the average concentration of each species as well as the structure of the
matrix, i.e. the distribution of species in space. A record of the matrix evolution
gives the spatial and temporal variation of the state variables.
The limitations of the technique include the finite number of configurations
averaged for 'the calculation of ensemble averages, the finite size of simulation cell,
the boundary effects, and the pseudo-randomness of the random numbers used
generated from multiplicative congruential generators. This procedure of random
numbers generation is deterministic and periodic. Since the periodicity is usually
large enough in comparison with the actual Markov chain length, no serious practical
limitation arises. It is an inherent feature of systems simulated that they contain
relaxation modes with very low frequencies. Such slow relaxation occurs in all
models which exhibit conservation laws (Hydrodynamic slowing down), second order
phase transitions (critical slowing down), first order phase transitions (metastable
states), and at low temperatures compared to energetic barriers [10]. Then the
approach towards equilibrium is extremely computer time consuming.
 193

3.2 Application of the Monte Carlo method to surface catalyzed re-

actions. The MCM in statistical physics models equilibrium and nonequilibrium
systems by stochastic computer simulation. Starting from a description of the de-
sired physical system in terms of a model Hamiltonian, one uses random numbers
to construct the appropriate probability with which the various generated states
of the system have to be weighted. Studies involving conservation of particles are
performed in a closed system (canonical ensemble). However, kinetics of heteroge-
neous systems are usually studied in an open system (grand canonical ensemble)
and these will be discussed briefly below.
To set up the kinetics, a probability function p(k) is introduced which gives
the probability that a given microscopic configuration k = (k j , kz , . .. ,kN.) occurs
at time t, where N. is the total number of adsorption sites on the surface (matrix
elements). Since all relevant processes such as adsorption, desorption, diffusion, etc.
are Markovian, the probability function satisfies a master equation

(17) d:~k) = L W(k' ---; k)P(k') - L W(k ---; k')P(k)

k' k

where W(k' ---; k) is the transition probability per unit time that the microscopic
state k' changes into the microscopic configuration k, and the summation is over
all possible states. The first summation is the rate of generation, and the second
summation is the rate of consumption of state k. TV must satisfy the microscopic
reversibility (or detailed balance condition)

(18) W(k' ---; k)Po(k') = W(k ---; k')Po(k)

where Po is the equilibrium probability. Next, adsorption and desorption processes

are considered (grand canonical ensemble). The equilibrium probability is given by
[23]

(19) PoCk) = exp( -H(k)/kT)/'3.

Here, H is the Hamiltonian for a grand canonical ensemble given by

(20) H(k) = E(k) - N(k)p

and '3 is the partition function

(21) '3 = L exp[( -E(k) + N p)/kT]

k,N

where E(k) is the internal energy of state k, N(k) is the number of particles in the
system, and p is the chemical potential of the gas phase which is related to the gas
pressure.
Using Eqs. (19)-(21), Eq. (18) becomes

W(k' ---; k)
(22) W(k ---; k') = exp[-(E(k) - E(k'»/kT] exp[(N(k) - N(k'»p/kT].
 194

Similarly, diffusion is an equilibrium process, and the principle of microscopic re-

versibility has to be invoked.
The final relation (22) is not sufficient to determine all transition probabilities.
A further choice has to be made based on the physical system simulated. This
has no effect as far as equilibrium (stationary solutions) are determined because
of the Onsager axiom that equilibrium is independent of the route approached.
Specifying the transition probabilities suitably, the MC averages converge to the
exact value in the limit of infinite Markov chain length, since the method is ergotic.
The probabilities and the algorithm used in this study are summarized in figure 1.

Choose event

'--------<~ To start

Figure 1. Flow chart of the MC algorithm and the probabilities

of various events used in the unimolecular model system. ni
is the number of nearest neighbors surrounding atom at ith
element and xsla is the mean number of atomic jumps during
the residence time of an atom on the surface. Probabilities in
the bimolecular reaction model system are defined in a similar
way.

Even for the simple adsorption-desorption model (Eq. (22)), there is no analyti-
cal solution except at the critical point [23]. Systems including irreversible chemical
reaction (nonequilibrium systems) are outside classical statistical mechanics, and no
rigorous theories exist to describe them [8,9]. Since microscopic details of surface
and spatial inhomogeneities are inherently included in the Monte Carlo method, it
is a very suitable technique for these applications.
 195

4. Comparison of the Results Obtained by the Monte Carlo Method

and the Ordinary Differential Equations.
In this section, results obtained by the MCM are compared with the predictions
of deterministic equations. In the simulations presented below, surface diffusion is
not included. The pressure in the gas phase is first assumed constant in time (a
differential reactor) and the solution of Eq. (4) is examined in parameter space. In
order to investigate periodic solutions, the system of Eqs. (3), (4) is then modeled.

4.1 Spatial structure and pattern formation. Figure 2a shows a snap-

shot of the surface configuration obtained after 310 8 attempts in the presence of
attractive absorbate-absorbate interactions (w > 0) for a certain point in parame-
ter space. Agglomeration of adatoms takes place and clusters are formed at several
positions of the surface caused by the dissipative structure of the system. Small
clusters usually shrink and new clusters are formed at different locations. The effect
of local fluctuations in adsorption on the formation of clusters and therefore on the
spatially nonuniform solution is very profound. The role of cluster formation and
growth on the stability of solutions and existence of oscillations will be discussed in
the following sections.
In the presence of repulsive interactions (w < 0), atoms repel each other. When
only first nearest neighbors are considered, and the surface is almost half filled,
atoms occupy the diagonals of a square, as shown in figure 2b. Thus ordering on
the surface occurs which results in a lower Gibbs free energy. A disordering-ordering
transition (from a dilute gas to a ordered structure) can take place as a bifurcation
parameter (pressure) changes. More complicated structures and transitions can
occur on the surface if interactions of longer range are taken into account. The
patterns produced depend on the range, strength, and type of interactions.
In addition to the adsorption-desorption mechanism examined, surface diffusion
is expected to be another mechanism which also drives atoms towards the clusters
minimizing the free energy [8]. Thus, in the presence of attractive interactions,
"stirring" of the adlayer by diffusion does not seem very likely. Hence, the above
examples indicate that the uniform distribution of atoms assumed in Eq. (1) is only
a crude approximation, and spatially nonuniform solutions govern the dynamics of
heterogeneous systems in the presence of absorbate interactions.
 196

(a)
..
..-... .-.....
.." . .,

,.•c.l-r: •- . .. .. ....
• 'I •
I .. . ~ ..

.. .
.. . . .... .:.." . .
~.
"
•
l- :I ~
. ." - .."..
I • • • .. - •
" •• 1
•'"
••
L 'I
.• itt
.-. •
•• I
. ,I
• "
"
I
.•.
. I·t ·.- .
-.- -- ---.
••
. .- .. ••.•
~
•
.- . r- . •
- .
• . .,.
.. t .. "

-: .. 1
.- .,.- :~.;.;
... .
(b)

Figure 2. Snapshots of nonuniform solutions obtained by the

MeM. Panel (a) shows a snapshot obtained from a 100 x 100
surface after 3108 trials in the presence of attractive inter-
actions (w/kT = 2,kR = 0,8 0 = O,J{PA = 0.0192). Panel
(b) shows a snapshot obtained from a 100 x 100 surface after
4108 trials in the presence of repulsive interactions (w / kT =
-4,kR = 0,8 0 = l,J{PA = 125).
 197

4.2 M uItiple solutions and stability. In the limit of no reaction (k RI kd =

0), Eq. (4) has multiple solutions (three solutions) ifthe temperature is lower than a
critical value (w I kT = 1) determined by the cusp point condition (Eqs. (11), (12)).
Application of the equal area Maxwell rule determines the thermodynamic equilib-
rium between the two phases (tie line). States on the lower and upper branches
are stable with respect to perturbations, i.e. the system is trapped in one of them
for an infinite time. States in the intermediate branch are unstable with respect to
perturbations, and trajectories lead either to the upper or lower branch depending
on the direction of the perturbation.
Simulations using the MCM show that as the partial pressure of reactant A
increases, the coverage increases (nonuniform solution) until a discontinuous (first
order) phase transition from the low coverage branch to the high coverage branch
is observed. This behavior is depicted in figure 3a. Time series of coverage shown
in figure 3b, indicate that the system is trapped in the low state branch for a long
time (metastable state) before the transition to the upper branch (stable state)
take place. Growth of one or more big clusters developed on the surface by local
fluctuations results in a phase transition and in a minimization of free energy.
This transition occurs above a critical value of pressure, PA,Cl and it is observed
up to a maximum pressure, PA,t. Simulations at all intermediate pressures exhibit
a phase transition from the lower branch to the upper branch provided a sufficient
long time. The closer the pressure is to PA,c, the longer the metastable lives. At the
critical point, PA,c, an infinite time is required for the transition of the system. The
last point at which the transition occurs, PA,t, (hereafter called pseudospinodal
point) corresponds to the turning point of the deterministic MF model. Further
increase of the pressure results in only one stable solution without any metastable
state.
Starting from a high reactant pressure (upper branch) and decreasing the pres-
sure, a unique solution is obtained if the pressure is higher than PA,c' Below this
value, a phase transition from the upper branch (metastable state) to the lower
branch (stable state) is observed after sufficient time as shown in figure 3b. Thus
the isotherm exhibits hysteresis, and multiple solutions are predicted for an isother-
mal system as shown in figure 3a. Metastability and hysteresis are usually char-
acteristic features of first order transitions. The quantitative agreement between
MCM and MF is not good. Furthermore, the correct stability of two branches and
the corresponding phase transitions are missing from the MF theory.
 198

(a)
1.0 ---,-'- -- ._.
, _. _.-
i s
e i

0.5
_._._._ '-'- -.-.-.-._ i~/MFU
°r'-'''' ........ -.-
i;
Me
! s
__ • ________ 1. ___________ _
0.0
0.01 0.02 0.03
KPA
(b)
1.0

e
0.5

0.0
o 1200 2400
tI't.

Figure 3. Panel (a) shows the MF isotherm and the MC data,

along with their corresponding tie lines. Metastable states are
indicated in panel (b) which shows typical reactant coverage
time series, starting from a clean (8 0 = O,KPA = 0.0195) or a
totally covered surface (8 0 = 1, K P = 0.0172). The parameters
used are w/kT = 2 and kR = O.

4.3 Cusp points: determination and dynamics. The effect of the value of
the reaction rate constant (bifurcation parameter) on the multiplicity of solutions
is shown for both models in the bifurcation diagrams figure 4a and 4b. Since first
order kinetics are assumed, the reaction of individual atoms is independent of the
local structure of the surface. Therefore, removal of atoms from the dense parts of
the matrix (clusters) takes also place. This results in destruction of clusters which
are necessary for the phase transition to occur. Consequently, the bistability regime
 199

is reduced. Above a certain value of the parameter kR/ kd (dependent on the bifur-
cation parameter w/kT), a single solution is obtained as shown in figures 4a and 4b.
The cusp point (singularity of co dimension 2) at which the two turning points
coincide can be determined rigorously in MF theory by application of Eq. (11), as it
is shown in the inset panel of figure 4b. Due to the lack of a rigorous condition in the
stochastic model, two alternative ways are proposed to determine cusp points. The
reactant pressure at the pseudospinodal points, J{ PA,I is first plotted as a function
of kR/kd in the bifurcation diagram 4b. As the cusp point is reached, the two
turning points approach each other, and the coincidence of pseudospinodal points
would determine the singular point of co dimension two (cusp point). The exact
location of this point requires extremely long simulations on large surfaces because
exact values of J{ PA,I are a prerequisite for this calculation.
Secondly, a size effect analysis is used in analogy with the equilibrium Ising
model [10]. Figure 4c shows the amplitude of fluctuations in coverage, boO, as a
function of kR/kd at constant reactant pressure. The maximum of the fluctuations
should indicate the location of the cusp point. It is observed that below a cer-
tain reactant pressure, no large fluctuations dominate the dynamics of the system.
This probably indicates that the simulations are performed in the bistability regime
where a relatively large perturbation brings the system from one state (a low cov-
erage branch) to the other (a high coverage branch). Once the system has left its
original state, it does not seem to return to its initial state, at least during the
computational time, if its size is large enough.
Very large fluctuations characterize the dynamics of the system when kR/kd is
slightly greater than the cusp point value, i.e. as the cusp point is reached from
the unique solution subspace. For the calculations shown in figure 4c, we used a
constant number of surface sweeps of approximately 120000. A typical time series
exhibiting fluctuations is shown in figure 4d. The dynamics of the time series
obtained by the Me simulations are examined by using the correlation integral
method [8,24]. We have found that the correlation exponent is proportional to the
embedding dimension as the latter is increased, indicating that the dynamics are
dominated by stochastic noise [8].
 200

(a) 0.023 r(...;.b)_ _ _ _ _ _ _ _ _--.

...or ••••• ••••••••••••••••••••
MC ~ .•..•...... MC
.' .:.0

::sEJ
(, 10.3 ." .•..•. ·",_ . .a-. b '- I solution
e .' 5 10.3 8'
,<T":."
.0,' .0
10-4
.: :':
.: ,"
l ,lJ'
kRlkdO ~ 10.2 e' •
0.5 :' ! r""?"=,--"'" 0.020
·····3 .....

.. "
I
:: t ~ 9 • A,I

iI
/ "Q'
•••• 0.04 3
1
MF

: '
0'---==-_-1 •. ' I 0.00
c1
'A".&.a
• of
o o 0.04 k.Ik.O.OS
0 0.017
0.01 0.04 0.07 o 12
KPA
(e) (d)
0.4 1.0
K PA ~ 0.022
0.023
9 e
\ 0.024
0.2
!\ ~'N.,/<"x0.025~O,026
: +,' ... 0.5
:)( .'~ \

, 1<
~f'k\.1<
/ \.
•
. 0,

•
0.0 L...---'_-'-_-'----'_-'----.J
o 30 o 5103 I 104
U't,

Figure 4. Effect of reaction rate constant, k R , on metastability

for w/kT = 2. Panel (a) shows isotherms obtained by the
MCM for different values of kR . Isotherms obtained by the
MF for the same parameters are shown in the inset. Panel (b)
shows turning points as a function of kR. No multiple solutions
can be found above a certain value of k R. Notice that the values
of kR obtained by the MCM and the MF model differ almost
by two orders of magnitude at the cusp point. Panel (c) depicts
the amplitude of fluctuations, D.O, of a 15 x 15 square lattice
as a function of reaction rate constant kR, for various values of
reactant pressure. Very small fluctuations are found below a
certain pressure. Panel (d) shows a time series obtained from
a 15 x 15 surface close to the cusp point shown in panel (b)
(KPA = 0.022, kR/kd = 9.510- 4 ).
 201

4.4 Degree of deviations between stochastic and deterministic solu-

tions. The existence of nonlinearities in the governing equations implies a depen-
dence of the corresponding nonlinear rates on the local environment of each atom.
Hence, fluctuations occurring in various processes can give rise in the development
of spatially nonuniform solutions. In the presence of attractive interactions, sur-
face diffusion does not probably lead to stirring of adatoms which would reduce the
spatial inhomogeneities and improve the predictions of MF theory. Figure 2a is an
example of this situation.
Removal processes, such as desorption or reaction, when they do not depend
on local environment, can partly destroy the formation of clusters if they are suf-
ficiently fast. This is a kind of "stirring" occurring through the coupling with the
gas phase. This behavior is observed for both model systems, i.e. unimolecular and
bimolecular reactions. When the desorption rate of reactants is sufficiently high,
good agreement between deterministic and stochastic model can be achieved, as for
example in the AB model system. At sufficiently low desorption rates, random-
ization of the surface occurs slowly compared to adsorption (slow "stirring") and
deviations ofthe MF model from the MCM are found [11,12]. However, "stirring" in
heterogeneous systems through a fast Fickian diffusion or a removal process cannot
be experimentally controlled (e.g. by increasing the temperature) without affecting
the dynamics of the entire system (all rates), i.e. "mixing" is an intrinsic property
of the system which varies with experimental conditions and nature of system.
In the following two sections, the problem of Hopf bifurcation is examined by
simulation of Eqs. (3) and (4).

4.5 Limit cycles. Self sustained oscillations can be found in the MC simula-
tions when attractive interactions are sufficiently strong so that multiplicity of the
isotherm occurs, as shown in figure 5. Panels 5a and 5b show time series of coverage
and reactant pressure predicted by MF and MC simulations, and panels 5c and 5d
depict the portrait of the system in phase space along with the isotherm obtained
from constant pressure simulations.
The dynamics of the time series were first examined by using the correlation
integral method which is used to distinguish random time series from time series
which stem from deterministic chaos [8,11,24]. The correlation integral was calcu-
lated from the pressure time series obtained from a 30 x 30 surface which consisted
of 10000 points separated by about 33 Monte Carlo steps. Below a certain length,
the slope of the curves is constant, ~ 0.95, indicating a high temporal uniformity of
the reactant pressure. The Fourier spectrum obtained from the pressure or coverage
time series indicates a dominant characteristic frequency which is the inverse of the
mean period of the oscillations.
The amplitude and frequency of oscillations remains nearly constant as the size
of simulation array increases from 10 x 10 to 40 x 40. The simulations are performed
far from the cusp point where the dynamics are characterized by large and irregular
fluctuations [8]. The coverage and reaction rate oscillations are driven by phase
transitions due to absorbate-absorbate attractive interactions and are synchronized
by oscillations in the gas pressure.
 202

e
o "'---=----=---'--''---'~--'

2.2 r - - - - - - - ,
KP A
X 102

1. 6 :-~----::-'-=-~_-:'.
o 3.5 7
t/~

(c)

Figure 5. Panels (a) and (b) show time series of coverage and
reactant pressure obtained by the MF and the MeM. Trajec-
tories on the f} - K PA plane for the MP model and the M eM
are shown in panels (c) and (d) respectively. The parame-
ters used in the MP model are w/kT = 1.5, KPA,o = 0.1005,
KP* = 0.03, kdTR = 1000, kR/kd = 0.00333 and in the Me
simulation are w/kT = 2, KPA,o = 0.024, KP* = 0.00144,
kdTR = 6.58104, kR/kd = 10-4, and surface size 30 x 30.

Since the oscillations are not exactly periodic, the influence of initial conditions
on the time evolution of the system is also examined to extract the Lyapunov ex-
ponent [25]. Simulations are carried out starting with different initial coverage and
surface configurations (adatom distribution), keeping constant the other parame-
ters. With a different initial coverage, the oscillations are again observed, but they
start from different points of the limit cycle. The trajectories in phase space are
very close to each other for all cases studied. The absolute value of the difference
in coverage traces obtained from different values of initial coverage reveals no ex-
ponential dependence on time. Recalling the randomness of the "tocha"tic model,
 203

the simulations indicate that the oscillations are not strongly affected by the initial
conditions.

4.6 Role of defects on periodic solutions. In the previous section, it was

demonstrated that mass transport through the gas phase can be a sufficient mecha-
nism for spatial self-organization of the system. This communication is believed to
be the dominant mechanism of synchronization on Pi(llO) surfaces [26]. Clusters
formed on the surface by fluctuations are nucleation centers for subsequent growth
and oscillations. Similar nucleation centers exist on real surfaces close to imper-
fections. Hence, interactions between defects are examined as a possible source of
synchronization through the reaction-diffusion phase, i.e. the surface [14]. Reaction-
diffusion coupling across the surface is probably the dominant mechanism for os-
cillations on Pi(100) surfaces [26]. Waves emanating from different locations of
the surface (probably defect sites) interact resulting in non regular macroscopic
oscillations [26].
In our model, it is assumed that atoms adsorbed on defect sites are not removed
by any annihilation mechanism. These atoms interact with other atoms by attrac-
tive forces and since they are not destroyed, they constitute nucleation centers.
However, there is no obvious reason why all of these nucleation centers should be
synchronized. Study of the influence of defects on the phase transition reveals that
defects facilitate nucleation and shift the transition (from a low to a high coverage)
to a lower pressure. Therefore, defects reduce the pressure amplitude [8] which
synchronizes the oscillations.
Figure 6 shows the effect of several different distributions of defects on the os-
cillations, for 4% imperfect sites. From panel a to c, the number of defect sites
(black squares) which can interact with reactant atoms (dotted squares) increases,
i.e. the importance of coupling through the surface becomes stronger. The simu-
lations show that in the presence of defects the oscillations become more chaotic
(for the same parameters used for a perfect surface). Single defected sites assist
the nucleation and the oscillations vanish for the same defect concentration, as it is
shown in figure 6c2,c3. As the oscillations become more irregular, a size dependence
is also observed, figure 6bl-c3. This indicates that identical unit cells are randomly
phased oscillators, and the reaction rate should be a time invariant over a large
number of these oscillators. By decreasing the inlet pressure, oscillations can be
still found for small defect concentration, but the amplitude of pressure is small
and the oscillations are less regular compared to the oscillations obtained from a
perfect surface.
 204

(a) (b) (c)

D
D § 9:J DD £IIi

...
D
IJ Il Il Il EI EI
rP
EI
El

Il

8 D
IlIlEl

,:m ,:m ,:m

,:mm ,:rnJ ,:t:J
o 3 6 0 3' 6 0 3 6
tit. tit. l/'f.

04 048 048

,:rImJ ,:~ ':0

~ ~ ~

0 3 6 0 3 6 0 3 6
1/1. 1/1. til.

Figure 6. Effect of defect distribution on oscillations for a total

defect concentration of 4%. In panel (a) defects are located in
one patch at the center of the surface. In panel (b) the unit cell
consists of 100 atoms with 4 defect sites at the center. In panel
(c) the unit cell (oscillator) contains 25 sites with one defect
site. The surface snapshots (a)-( c) are from a 30 x 30 surface at
t/TR '" 0.017. Oscillations in B are shown in (al)-(c3)' On the
vertical axis the size effect dependence is indicated. (al )-( Cl)
are obtained from a 10 x 10 surface, (a2 )-( C2) from a 20 X 20
surface, and (a3 )-( C3) from a 30 x 30 surface.
 205

5. Conclusions.
The dynamics of heterogeneous systems have been examined by studying generic
model systems with a stochastic method (MCM) and the traditional deterministic
approach (MF). The cooperative phenomenon caused by absorbate interactions and
the role of fluctuations on the dynamics have been investigated. Spatially nonuni-
form solutions can characterize the distribution of atoms on a sUlface whenever
nonlinear terms appear in the ODEs. Phase transitions can be frequently associ-
ated with these nonlinearities and their role on the stability of the system is crucial.
An experimental probe of the local field reveals large fluctuations which stem from
stochastic noise (strong correlations) in many cases, especially close to cusp points.
The above features of complicated dynamics which are often observed in experi-
ments are missing in the solution obtained from ODEs. Nonlinearities in the ODEs
describing the dynamics of the system imply a dependence of rates on the local
vector field and the MC calculations suggest that under such circumstances the MF
theory can be only an approximation. The validity of approximation depends on
the values of the parameters, the importance of migration, and the surface defects,
i.e. the intrinsic "stirring" set in the adlayer. Therefore, utilization of a uniform
field to model these systems with ODEs results in deviations from the exact solution
of the system and lack of knowledge concerning stability and phase transitions.
On the other hand, even though PDEs can give a spatial variation of the vector
field, in the presence of attractive interactions (w > 0) or local disturbances of the
vector field (defects), they are structurally incorrect. In PDEs of the form (16),
surface diffusion tends to destroy the concentration gradients. Therefore, the need
for stochastic modeling which provides an exact solution of model systems is very
striking. Using the MCM, we have demonstrated the efficiency of the technique to
model the complicated dynamics of heterogeneous systems such as multiple solutions
and stability, Hopf bifurcations, and the problem of synchronization through surface
defects.
REFERENCES

[1] L.F. RAZON AND R.A. SCHMITZ, Cata!. Rev. - Sci. Eng., 28 (1986), p. 89.
[2] E. WICKE, P. KUMMANN, W. KEIL, AND J. SCIIIEFLER, Ber. Bunsenges. Phys. Chern., 84
(1980), p. 315.
[3] P. MOLLER, K. WETZL, M. EISWIRTH, AND G. ERTL, J. Chern. Phys., 85 (1986), p. 5328.
[4] M. EHSASI, M. MATLOCH, O. FRANK, J.II. BLOCK, K. CHRISTMANN, F.S. RYs, AND W.
IIIRSCHWALD, J. Chern. Phys., 91 (1989), p. 4949.
[5] J. KEVREKIDIS, L.D. SCHMIDT, AND R. ARIS, Surface Sci., 137 (1984), p. 15I.
[6] M.A. McKARNIN, R. ARIS, AND L.D. SCHMIDT, Proc. R. Soc. Lond., A 415 (1988), p. 363.
[7] S. JAKUBITH, II.H. ROTERMUND, W. ENGEL, A. VON OERTZEN, AND G. ERTL, Phys. Rev.
Lett., 65 (1990), p. 3013.
[8] D. G. VLACHOS, L.D. SCHMIDT, AND R. ARIS, Surf. Science, (to be published).
[9] R.M. ZIFF, E. GULARI, AND Y. BARSHAD, Phys. Rev. Letters, 56 (1986), p. 2553.
[10] Topics in Current Physics, Monte Carlo Methods in Statistical Physics, ed. K. Binder,
(Springer, New York, 1986).
[11] K. FICHTHORN, E. GULARI, AND R. ZIFF, Chern. Eng. Sci., 44 (1989), p. 1403.
[12] K. FICHTHORN, E. GULARI, AND R. ZIFF, Phys. Rev. Lett., 63 (1989), p. 1527.
[13] M. KOLB AND Y. BOUDEVILLE, J. Chern. Phys., 92, (1990), p. 3935 and references therein.
[14] D.G. VLACHOS, L.D. SCHMIDT, AND R. ARIS, J. Chern. Phys., 93 (1990), p. 8306.
 206

[15) See. e.g., A.M. Zhabotinsky, Ber. Bunsenges. Phys. Chern., 84 (1980), p. 303.
[16) P.C. PAWLICKI AND R.A. SCHMITZ, Chern. Eng. Prog., 83 (1987), p. 40.
[17) R.A. SCHMITZ, G.A. D'NETTO, L.F. RAZON, AND J.R. BROWN, NATO ASI Ser., C (1984),
120 (Chemical Instabilities, G. Nicolis and F. Baras (eds.)), p. 33.
[18) R.J. SCHWANKNER, M. EISWIRTH, P. MOLLER, K. WETZL, AND G. ERTL, J., Chern. Phys.,
87 (1987), p. 742.
[19) M.E. KORDESCH, W. ENGEL, G. JOHN LAPEYRE, E. ZEITLER, AND A.M. BRADSHAW, Appl.
Phys., A 49 (1989), p. 399.
[20) G.A. CORDONIER, F. SCHUTH, AND L.D. SCHMIDT, J. Chern. Phys., 91 (1989), p. 5374.
[21) J.W. CAHN, Acta Metall., 9 (1961), p. 795.
[22) N. METROPOLIS AND S. ULAM, J. Am. Stat. Ass. ,44 (1949), p. 335.
[23) T.L. HILL, An Introduction to Statistical Thermodynamics, (Dover Publications, Inc., New
York, 1986).
[24) P. GRASSBERGER AND I. PROCACCIA, Physica, D9 (1983), p. 189.
[25) RUELLE D., Proc. R. Soc. Lond., A 427 (1990), p. 241.
[26) M. EISWIRTH, P. MOLLER, K. WETZL, R. IMBIHL, AND G. ERTL, J. Chern. Phys., 90 (1989),
p.510.