Sie sind auf Seite 1von 360

Chaos Theory in the Social Sciences

Chaos Theory
in the Social Sciences
Foundations and Applications

Edited b y

L. Douglas Kiel and Euel Elliott

Ann Arbor

T H E UNIVERSITY OF MICHIGAN PRESS


First paperback edition 1997
Copyright © by the University of Michigan 1996
All rights reserved
Published in the United States of America by
The University of Michigan Press
Manufactured in the United States of America
© Printed on acid-free paper

2004 7 6 5 4

No part of this publication may be reproduced, stored in a retrieval


system, or transmitted in any form or by any means, electronic,
mechanical, or otherwise, without the written permission of
the publisher.

A CIP catalog record for this book is available from the British Library.

Library of Congress Cataloging-in-Publication Data

Chaos theory in the social sciences : foundations and applications /


edited by L. Douglas Kiel and Euel Elliott,
p. cm.
Includes bibliographical references (p. ).
ISBN 0-472-10638-4 (he : alk. paper)
1. Social sciences—Mathematical models. 2. Chaotic behavior in
systems. I. Kiel, L. Douglas, 1956- . II. Elliott, Euel W.,
1951- .
H61.25.C48 1995
300'. 1'51—dc20 95-35470
CIP

ISBN 0-472-08472-0 (pbk. : alk. paper)


To Our Parents
Contents

Introduction
Euel Elliott and L . Douglas Kiel

Part 1. Chaotic Dynamics in Social Science Data

1. Exploring Nonlinear Dynamics with a Spreadsheet:


A Graphical V i e w o f Chaos for Beginners
L . Douglas Kiel and Euel Elliott
2. Probing the Underlying Structure in Dynamical
Systems: A n Introduction to Spectral Analysis
Michael McBurnett
3. Measuring Chaos Using the Lyapunov Exponent
Thad A. Brown
4. The Prediction Test for Nonlinear Determinism
Ted Jaditz
5. From Individuals to Groups: The Aggregation
o f Votes and Chaotic Dynamics
Diana Richards

Part 2. Chaos Theory and Political Science

6. Nonlinear Politics
Thad A. Brown
7. The Prediction o f Unpredictability: Applications
o f the New Paradigm o f Chaos i n Dynamical
Systems to the O l d Problem o f the Stability
o f a System o f Hostile Nations
Alvin M. Saperstein
8. Complexity in the Evolution o f Public Opinion
Michael McBurnett

Part 3. Chaos Theory and Economics

9. Chaos Theory and Rationality in Economics


J. Barkley Rosser, Jr.
viii Contents

10. L o n g Waves 1790-1990: Intermittency,


Chaos, and Control 215
Brian J. L . Berry and Heja Kim
11. Cities as Spatial Chaotic Attractors 237
Dimitrios S. Dendrinos

Part 4. Implications for Social Systems Management


and Social Science

12. Field-Theoretic Framework for the Interpretation


o f the Evolution, Instability, Structural Change,
and Management o f Complex Systems 273
Kenyon B. De Greene
13. Social Science as the Study o f Complex Systems 295
David L . Harvey and Michael Reed

References 325

Contributors 347
Introduction

Eue I Elliott and L . Douglas Kiel

The social sciences, historically, have emulated both the intellectual and
methodological paradigms o f the natural sciences. From the behavioral revo­
lution, to applications such as cybernetics, to a predominant reliance on the
certainty and stability o f the Newtonian paradigm, the social sciences have
followed the lead o f the natural sciences. This trend continues as new discov­
eries in the natural sciences have led to a reconsideration o f the relevance o f
the Newtonian paradigm to all natural phenomena. One o f these new discov­
eries, represented by the emerging field o f chaos theory, raises questions
about the apparent certainty, linearity, and predictability that were previously
seen as essential elements o f a Newtonian universe. The increasing recogni­
tion by natural scientists o f the uncertainty, nonlinearity, and unpredictability
in the natural realm has piqued the interest o f social scientists in these new
discoveries. Chaos theory represents the most recent effort by social scientists
to incorporate theory and method from the natural sciences. Most importantly,
chaos theory appears to provide a means for understanding and examining
many o f the uncertainties, nonlinearities, and unpredictable aspects o f social
systems behavior (Krasner 1990).
Chaos theory is the result o f natural scientists' discoveries in the field o f
nonlinear dynamics. Nonlinear dynamics is the study o f the temporal evolu­
tion o f nonlinear systems. Nonlinear systems reveal dynamical behavior such
that the relationships between variables are unstable. Furthermore, changes in
these relationships are subject to positive feedback in which changes are
amplified, breaking up existing structures and behavior and creating unex­
pected outcomes in the generation o f new structure and behavior. These
changes may result in new forms o f equilibrium; novel forms o f increasing
complexity; or even temporal behavior that appears random and devoid o f
order, the state o f "chaos" in which uncertainty dominates and predictabil­
ity breaks d o w n . Chaotic systems are often described as exhibiting low-
dimensional or high-dimensional chaos. The former exhibit properties that
may allow for some short-term prediction, while the latter may exhibit such
2 Chaos Theory in the Social Sciences

variation as to preclude any prediction. I n all nonlinear systems, however, the


relationship between cause and effect does not appear proportional and deter­
minate but rather vague and, at best, difficult to discern.
These discoveries have given rise to a new mathematics that belies pre­
vious scientific commitment to prediction and certainty. Natural scientists
have now applied this mathematics to numerous fields o f inquiry. A brief and
partial listing o f the fields includes meteorology (Lorenz 1963), population
biology ( M a y 1976), and human anatomy (West and Goldberger 1987). These
studies consistently show that nonlinearity, instability, and the resulting uncer­
tainty are essential components in the evolutionary processes o f natural sys­
tems. Moreover, these inquiries have given precedence to a greater concern
for the extent o f and challenges o f understanding the inherent complexity o f
natural systems.
The emerging paradigm o f chaos thus has profound implications for the
previously dominant Newtonian view o f a mechanistic and predictable uni­
verse. W h i l e a Newtonian universe was founded on stability and order, chaos
theory teaches that instability and disorder are not only widespread in nature,
but essential to the evolution o f complexity in the universe. Thus, chaos
theory, as relativity theory and quantum theory before i t , presents another
strike against a singular commitment to the determinism o f a Newtonian view
o f the natural realm.
This understanding also suggests that the relative successes in knowledge
acquisition by the natural sciences are the result o f a focus on "simple"
systems that function in an orderly and consistent manner. As natural scien­
tists have shifted their investigative focus to more complex systems, the
previous quest for certainty has given way to a greater appreciation o f uncer­
tainty and the enormity o f potential generated by the uncertainty o f disorder
and disequilibrium.
W i t h the focus o f chaos theory on nonlinearity, instability, and uncer­
tainty, the application o f this theory to the social sciences was perhaps a
predictable eventuality. As Jay W. Forrester (1987, 104) has noted, "We live
in a highly nonlinear w o r l d . " The social realm is clearly nonlinear, where
instability and unpredictability are inherent, and where cause and effect are
often a puzzling maze. The obvious fact that social systems are historical and
temporal systems also stresses the potential value o f chaos theory to the social
sciences. Social systems are typified by the changing relationships between
variables.
The obvious metaphorical value o f applying a theory o f chaos to the
social realm has served as an impetus for the emergence o f the application o f
this theory to social phenomena. Yet chaos theory is founded on the mathe­
matics o f nonlinear systems. Thus, social scientists, in their efforts to match
the mathematical rigor o f the natural sciences, are increasingly applying this
Introduction 3

mathematics to a variety o f social phenomena. Time-series analysis is essen­


tial to these efforts, as researchers strive to examine how nonlinear and cha­
otic behavior occurs and changes over time.
Clearly, the fundamental gap between the clear success o f knowledge
acquisition in the natural sciences versus the rather minimal successes in
understanding the dynamics o f the social realm is the inherent nonlinearity,
instability, and uncertainty o f social systems behavior. The seeming "chaos"
of social phenomena has always been a stumbling block to knowledge acqui­
sition i n the social sciences. Social scientists have long argued that this rela­
tive knowledge gap was due to the relative complexity o f the phenomena
examined by the two scientific cultures. Yet chaos theory teaches that the
"gap" between the t w o sciences may have largely been artificial. As natural
scientists more intensively investigate complex natural phenomena, they too
must contend with the challenges that have long served to keep the social
sciences in the position o f a scientific stepchild. Chaos theory seems to repre­
sent a promising means for a convergence o f the sciences that w i l l serve to
enhance understanding o f both natural and social phenomena.
Chaos theory has now been applied to a wide variety o f social phenom­
ena ranging across the subject matter o f the traditional social science disci­
plines o f economics (Grandmont 1985; Baumol and Benhabib 1989; Arthur
1990) and political science (Saperstein and Mayer-Kress 1989; Huckfeldt
1990; K i e l and Elliott 1992). Economists and political scientists have applied
phaos theory w i t h considerable methodological rigor and success to the tem­
poral dynamics o f a variety o f phenomena in their fields. Chaos theory has
also been applied to sociology. In this field, however, more than in economics
and political science, such efforts have tended toward metaphorical and post­
modernist or poststructuralist usages (Young 1991, 1992). Thus, while this
volume does not include rigorous mathematical assessments o f chaotic dy­
namics in the subject matter o f sociology, the applications in political science
and economics should serve as foundations for the development o f such
research in sociology. W h i l e no specific chapter contends solely w i t h these
postmodernist and poststructuralist issues, David Harvey and Michael Reed's
concluding chapter examines the relevance o f these elements to chaos re­
search and social science inquiry.
The increasingly evident value o f chaos theory in the social sciences is
thus its promise as an emerging means for enhancing both the methodological
and theoretical foundations for exploring the complexity o f social phenomena.
Exploring this emergent and potential value is the purpose o f this book. B y
examining applications o f chaos theory to a range o f social phenomena and by
providing means for exploring chaotic dynamics, the chapters in this book
afford the reader a comprehensive vision o f the promise and pitfalls o f chaos
theory in the social sciences.
4 Chaos Theory in the Social Sciences

This book seeks to provide knowledge to both social scientists new to


this area o f study and the well-informed chaos researcher. Chapters range
from the mathematically and methodologically sophisticated to chapters with
a strictly theoretical emphasis. The book is organized by both disciplinary
area and general methodology. The disciplinary sections examine chaos
theory i n political science and economics. The first section o f the book exam­
ines methods for exploring and examining the existence o f chaotic dynamics
in time-series data, which cut across the disciplines. First, though, an initial
primer on chaos and nonlinear behavior is necessary to provide the basics o f
this theory and an introduction to the unique vocabulary it utilizes.

The M a t h e m a t i c s and Behavior of Chaos

A brief examination o f the mathematics and behavior o f chaotic systems


provides a means for understanding the relevance o f this theory to the com­
plexity o f social phenomena. Distinguishing between linear and nonlinear
equations also reveals both the relevance and the challenge o f contending with
the nonlinear mathematics o f nonlinear systems. Linear equations are typified
by the superposition principle. This principle, simply stated, means that two
solutions o f a linear equation can be combined, or added together, to generate
a new solution. This means that linear equations allow problems to be broken
down into smaller pieces that may generate several separate solutions. In such
linear mathematics, the individual solutions can be added back together to
form a complete solution to the entire problem.
The superposition principle, however, does not hold for nonlinear equa­
tions. A nonlinear equation cannot be broken down into bits and then refor­
mulated to obtain a solution. Nonlinear differential equations, and the phe­
nomena or problems they describe, must be seen as a totality, that is, as
nondecomposable. This further means that nonlinear equations are partic­
ularly intractable for the analyst. The inherent nonlinearity o f many social
phenomena and the intractability o f the relevant mathematics thus must ex­
plain, in part, the challenges social scientists face when attempting to under­
stand the complexity o f social dynamics.
Another element o f the mathematics o f nonlinear equations is the fact
that a simple deterministic equation can generate seemingly random or chaotic
behavior over time. One nonlinear differential equation, the logistic map, is
such an example. The logistic map is described in detail in the chapters in this
edition by K i e l and Elliott (chap. 1), Diana Richards (chap. 5), and A l v i n
Saperstein (chap. 7). These examples o f the logistic map also detail an essen­
tial element o f chaotic behavior. Chaotic behavior occurs within denned pa­
rameters. The logistic map shows that a simple system can create very com­
plex and chaotic behavior. This realization has obvious impact for the social
Introduction 5

sciences. Social systems o f initial relative simplicity may result over time in
very complex behavior.
The varying mathematics o f linear and nonlinear systems also result in
divergent temporal behavior for these types o f systems. Linear systems, char­
acterized by stable relationships between variables, respond to changes in
their parameters, or to external "shocks," in a smooth and proportionate
manner. Consequently, linear systems w i l l exhibit smooth, regular, and w e l l -
behaved motion. Even large waves or pulses in a linear system w i l l be dis­
persed over time, generally resulting in a move back to the typical behavior o f
the system.
Nonlinear systems may be characterized by periods o f both linear and
nonlinear interactions. During some time periods behavior may reveal linear
continuity. However, during other time periods relationships between vari­
ables may change, resulting in dramatic structural or behavioral change. Such
dramatic change from one qualitative behavior to another is referred to as a
"bifurcation." Nonlinear systems are consequently capable o f generating very
complex behavior over time. Studies o f nonlinear systems evidence three
types o f temporal behavior. Nonlinear systems may evidence behavior that
(1) is stable (a mathematical equilibrium or fixed point); (2) oscillates between
mathematical points in a stable, smooth, and periodic manner; or (3) is cha­
otic and seemingly random, devoid o f pattern (nonperiodic behavior). Chapter
one presents graphical images o f these three temporal regimes. These behav­
iors may occur intermittently throughout the " l i f e " o f a nonlinear system. One
regime may dominate for some time periods while other regimes dominate at
other times. It is the potential for a variety o f behaviors that represents the
dynamics o f nonlinear systems.
Chaotic behavior is the behavioral regime in nonlinear systems o f great­
est interest. Chaotic behavior, while occurring within defined mathematical
parameters, appears random and without pattern over time. Chaotic behavior
does not retrace previous points during its temporal evolution. This creates the
appearance o f randomness. Chaotic behavior, however, is not random behav­
ior, since it can be generated with a completely deterministic equation. This
understanding is an essential foundation o f knowledge for chaos researchers.
Even deterministic systems can generate very erratic behavior over time.
Moreover, as noted above, a chaotic system may appear more or less random
depending on its complexity. A system mapped by the logistic equation may
allow for some predictability and is an example o f a low-dimensional chaotic
system.
This point raises another distinctive point regarding nonlinear systems.
Nonlinear systems are historical systems in that they are determined by the
interactions between the deterministic elements in a system's history and
"chance" factors that may alter its evolution. In systems operating in a chaotic
6 Chaos Theory in the Social Sciences

regime, this fact is referred to as sensitive dependence on initial conditions. In


short, the combination o f factors that defines the initial condition o f the phe­
nomenon and the insertion o f chance elements during its " l i f e " may generate
very divergent outcomes from systems that initially appeared quite similar.
This distinguishes chaotic behavior from truly random behavior. In a genu­
inely random system, such a system is insensitive to its initial condition.
Uncertainty is also an important element o f nonlinear systems since the
outcomes o f changing variable interactions cannot be k n o w n . Thus, the com­
plexities o f both internal dynamics and environmental "disturbances" generate
considerable uncertainty during change processes in nonlinear systems. Fur­
thermore, a wide and complex array o f possible outcomes is available to
nonlinear systems. This is particularly true during chaotic regimes. As a
result, any effort at long-term prediction in nonlinear systems is highly sus­
pect (Baumol and Quandt 1985).

Graphical Analysis in Chaos Research

The intractability o f nonlinear equations and the inherent difficulties in under­


standing the dynamics o f complex time series have led chaos researchers to
formulate new methods for analyzing data from nonlinear phenomena. This
point is w e l l stated by Hasslacher (1992, 60), w h o notes in reference to
complex nonlinear systems:

In these systems complexity is usually both emergent and Byzantine.


This means that organized and extended structures evolve and dominate a
system, and the structures themselves are so complex that, when first
seen, they produce a sense o f beauty followed by a deep feeling o f
unease. One instinctively realizes that the analytic tools that worked so
well i n the past are going to be o f little use.

Many o f these new analytical methods are graphical in nature and are based
on researchers' efforts to examine the dynamical motion o f time series gener­
ated by social science data. Chaos researchers have thus focused on examin­
ing the morphology (Abraham and Shaw 1982) o f the graphics generated by
these time series. For example, the chapter in this volume by Brian J. L . Berry
and Heja K i m , on the dynamics o f the economic long wave, relies solely on
this graphical approach to data analysis.
These graphical representations are lagged mapping o f data at adjacent
time periods that result in an amazing array o f geometric structures, resulting
in what Abraham and Shaw (1982) label the "geometry o f behavior." These
mappings reveal that nonlinear systems possess an underlying order k n o w n as
an attractor, where the mathematical points describing the systems' behavior
Introduction 7

create pattern and structure. These geometric formulations are used through­
out the chapters in the text as a means o f examining the underlying structure o f
longitudinal social science data.
Studies o f the attractors o f nonlinear time series reveal that each o f the
three behavioral regimes emanating from nonlinear differential equations cre­
ates a uniquely shaped attractor. A stable equilibrium generates a point extrac-
tor, in which the data are attracted to a single point on the mapping. A stable
periodic oscillation generates a circular mapping, or limit cycle, as the data
revolve back and forth between consistent mathematical points. The chaotic
attractor is represented by a variety o f unique shapes resulting in the labeling
o f such attractors as strange attractors. These attractors are typified by the
creation o f form without retracing previous mappings.
It is an examination o f these attractors that serves as a graphical founda­
tion for the notion o f "order in chaos." Even though the numerical data
describing a chaotic regime appear disorderly, their geometric representation
creates unique shapes o f order. A n d since chaotic regimes function w i t h i n
defined parameters, a stability also exists in chaos. We then begin to see that
chaotic behavior is globally stable, but locally unstable.

O r g a n i z a t i o n of t h e Book

This volume represents research spanning a range o f disciplines, meth­


odologies, and perspectives. We have incorporated many different substantive
areas in order to provide the reader w i t h as balanced a perspective as possible
o f the k i n d o f social science research and writing that is currently being done.
Moreover, while the book is organized into four separate sections, (1) explo­
ration and method, (2) political science, (3) economics, and (4) implications
for social systems management and social science, we would emphasize that
chaos theory is really about not only the interdisciplinary but also the m u l t i -
disciplinary character o f the social sciences. Thus, the reader w i l l occasion­
ally note references and allusions in one section to chaos research being
conducted in other areas o f the social or even natural sciences.
In addition to the desire on our part to incorporate a diverse array o f
substantive areas in this volume, a number o f other considerations were criti­
cal to our thinking about this work. First, we were very concerned that the
subject matter be treated in a fashion that would make the arguments and
concepts as accessible and "user friendly" as possible to the professional
social scientist, as well as to graduate students with an interest in nonlinear
dynamics. W h i l e we recognize that some o f the chapters deal with rather
complex arguments and formulations, the authors o f these chapters have done
an admirable j o b in presenting the material in such a way as to allow anyone
reasonably comfortable w i t h undergraduate mathematics to capture the gist o f
8 Chaos Theory in the Social Sciences

the arguments. M o r e sophisticated readers, however, are not shortchanged.


W h i l e accessible, we have insisted that the integrity o f the material not be
jeopardized. A t the same time, it w i l l be readily apparent that there is substan­
tial variation among the chapters in terms o f methodological rigor.
We were also concerned that the contributions represent both macro- and
micro-level phenomena. Readers w i l l observe that this is particularly the case
in our economics and political science sections. It is crucial, i n our view, to
demonstrate that chaotic processes can occur at the level o f individuals
and small groups, as well as at highly aggregated levels o f analysis. Indeed,
Thad A . B r o w n , in his overview chapter for political science (chap. 6), argues
that an important next step in the research agenda is to attempt a linkage o f the
two perspectives using theoretical approaches drawn from chaos theory.
Finally, we did not want to fall into the trap o f seeing an emerging
intellectual and methodological paradigm as a singular solution to the chal­
lenges o f understanding the complexities o f social systems behavior. As w i t h
all efforts to understand the complexity that constitutes the human and social
realms, a mature and reasoned skepticism is appropriate. The final chapter o f
this book, by David L . Harvey and Michael Reed, attempts to make sense o f
the evolution o f chaos theory in the social sciences and its prospects for
enhancing knowledge in the social sciences.

Chaotic D y n a m i c s in t h e Social Sciences:


Exploration a n d M e t h o d

The first section o f this volume examines the dynamics o f nonlinear and
chaotic systems and focuses on methodological approaches to testing for the
presence o f chaos in a time series. Testing for actual "chaos" in time-series
data is a particularly important issue, due not only to the technical challenge
involved, but also to ensuring that chaotic time series in social science data
emulate chaotic time series discovered in data from the natural sciences.
W h i l e a variety o f techniques exist to test for chaos, we have concentrated our
attention on those approaches that appear to be most often used by chaos
researchers. W h i l e each chapter has been written in what we consider a highly
accessible fashion, many readers may well prefer to start with the substantive
chapters in sections t w o and three and then return later to this section. For
those researchers w h o are either beginning to apply chaos theory to empirical
work or otherwise interested in some o f the more technical methodological
facets o f empirical work in the area, this section should be an invaluable
resource.
The editors o f this volume lead off this first section with a brief explora­
tion o f the time series o f nonlinear and chaotic systems. This chapter is highly
recommended as a starting point for researchers new to chaos theory. It shows
Introduction 9

how an electronic spreadsheet can be used to generate nonlinear and chaotic


time series. These series can be used to create graphs and phase diagrams
essential to investigating nonlinear time series. This chapter is intended to
reveal for social scientists how the dynamics o f a nonlinear differential equa­
tion emulate much o f the temporal dynamics o f social system behavior.
The remaining chapters i n this section are mathematically rigorous ap­
proaches to the statistical analysis o f chaotic dynamics in social science data.
Michael McBurnett's chapter examines the use o f spectral analysis in investi­
gating the dynamics o f a time series. McBurnett begins w i t h a necessary but
tractable mathematical introduction to spectral density and spectral distribu­
tion functions. He then examines the different types o f time series—periodic,
random, and chaotic—and demonstrates with regard to the first two types the
problems i n resolving the nature o f a series when "noise" is introduced into
the analysis. He concludes by examining known chaotic time series as well as
an " u n k n o w n " series. McBurnett's study is an excellent introduction to both
the advantages and the limitations o f spectral analysis in testing for chaotic
dynamics.
A second approach for examining chaotic dynamics relies on the use o f
Lyapunov exponents. The Lyapunov exponent measures the extent to which
"small" changes in initial conditions produce divergence in a system over
time. Thad A . Brown's chapter explores in detail both the advantages and
disadvantages o f such an approach. The Lyapunov is shown to be linked to
the information gained and lost during chaotic episodes, and hence is linked to
the amount o f information available for prediction. This chapter guides the
reader through the formal nature o f unfolding subspaces and the state space
reconstruction needed to estimate the Lyapunov exponent. The mathematics
here are downplayed, i n favor o f words and even some humor.
Ted Jaditz's chapter is concerned with the development o f empirical
techniques for predicting a time series exhibiting deterministic chaos. A s
Jaditz notes in the introduction to chapter 4 , "The Prediction Test for Non­
linear Determinism," standard linear statistical models provide good fits w i t h
data that are taken " i n sample," but out-of-sample predictions do much worse.
For this reason, economic forecasters have been attracted to the possibility o f
nonlinear determinism in economic data. Jaditz discusses the problems inher­
ent in determining whether or not a series truly manifests chaos and demon­
strates some analytical tools for improving forecasting models. Specifically,
by using "near neighbor techniques," Jaditz shows how this new approach
provides dramatic improvements over conventional linear prediction models
typically used by economists, at least with data that are known to be chaotic.
Diana Richards's contribution, "From Individuals to Groups: The Aggre­
gation o f Votes and Chaotic Dynamics," presents another method for testing
for chaotic dynamics. Richards applies Devaney's (1989) three-part test for
10 Chaos Theory in the Social Sciences

chaotic dynamics. A l t h o u g h the application o f chaos theory to the social


choice problem leads to several research questions specific to social choice,
the intent is to introduce chaotic dynamics to a broader social science audi­
ence using the case o f the generic social choice problem. The social choice
problem is only one o f potentially many examples o f interaction among indi­
viduals or groups that is nonlinear, and therefore a potential candidate for the
domain o f chaotic dynamics. Richards's chapter provides a rich understanding
of the translation o f individual preferences into group outcomes and the clas­
sic problem o f intransitivities. The application o f chaos theory is a relatively
small modeling step; it is only a small extension o f existing frameworks into
the nonequilibrium realm. However, chaos theory has major implications, in
terms o f complex outcomes from simple relationships, in terms o f instability
and structural constraints, and in terms o f prospects for prediction for all o f
the social sciences.
This chapter demonstrates that chaotic dynamics are present in many
social choice settings, including some cases o f Arrow-type social choice and
in nearly all cases where t w o or more issues are considered simultaneously.
Since the aggregation o f individual preferences into outcomes is inherently
nonlinear, it is natural to expect chaos theory—the theory o f nonequilibrium
nonlinear dynamics—to apply to social choice. It also therefore becomes
impossible to make long-term predictions concerning group interactions.
However, Richards emphasizes as well the underlying order o f chaotic pro­
cesses. Specifically, she suggests that a complex "fractal structure" exists,
indicating, at a fundamental level, structured stability in the system.

Chaos T h e o r y a n d Political Science

The use o f chaos and related theoretical and methodological constructs in


political science is still in its infancy. Many o f the features that have attracted
economists to chaos theory also exist among political scientists. Like eco­
nomics, much o f political science is concerned w i t h analyzing change and
exploring the evolution o f some phenomenon over time. Studies o f changes in
aggregate-level electoral fortunes and trends in public opinion such as presi­
dential approval or attitudes toward particular issues all fit this genre. Accord­
ingly, such data raise the question o f whether underlying deterministic, and
thus potentially chaotic, processes exist. The methodological advances in
statistical analysis that have been made in recent years, advances that to a
great extent have been borrowed from economics, have made some political
scientists more w i l l i n g , and able, to explore the existence o f complex non­
linear processes. The highly formal game-theoretic and social choice work
has required the application o f mathematical tools that are invaluable in chaos
research.
Introduction 11

The lead chapter in the political science section is Thad A . Brown's


"Nonlinear Politics." B r o w n introduces the reader to the role o f chaos in
understanding political phenomena. B r o w n points out that politics at every
level results from the interactions o f individuals. The difficulty is that, "For­
mally treating interactive political behavior within massively diverse collec­
tives is tricky. Interactive behavior is peculiar in that it can neither be pre­
dicted nor analyzed by observing sets o f individuals cross-sectionally or even
the time series from a given individual or group." B r o w n suggests that this
characteristic, together w i t h the likely existence o f spatial and temporal phase
transitions, calls into serious question traditional methodologies for investi­
gating chaotic phenomena. B r o w n goes on to explain that cellular automata
simulations provide an innovative means o f investigating complex dynamical
systems. He also discusses specific applications o f such an approach including
game theory, electoral behavior, and social choice theory, all o f which are
represented in this volume.
Physicist A l v i n M . Saperstein was among the first natural scientists to
rigorously apply chaos theory to social phenomena. His chapter, "The Predic­
tion o f Unpredictability: Applications o f the New Paradigm o f Chaos i n
Dynamical Systems to the O l d Problem o f the Stability o f a System o f Hostile
Nations," is in much the same spirit as other chapters in this section. L i k e
Richards and McBurnett, Saperstein is concerned with the fundamental prob­
lem o f prediction. Saperstein points out that while the international system
shows considerable stability and hence predictability in an overall sense,
crises represent episodes o f fundamental instability, ergo unpredictability.
Saperstein then points out that an important political "technology" w o u l d be
to know when given national security policies w i l l produce instability. I n
other words, the aim should be to "predict the unpredictable." Saperstein
models several facets o f international interactions, asking fascinating (and
long-standing) questions such as whether bipolar or multipolar international
systems are more likely to produce conflict, and whether democratic or non-
democratic states are more likely to go to war. The conclusions illuminate
w i t h great clarity some o f the most fundamental questions o f the nuclear age.
Saperstein's study has the added advantage o f providing simple algorithms
that can be used by anyone with a desktop computer to generate an evolution
o f national sanity behavior on the part o f nations.
The chapter "Complexity in the Evolution o f Public Opinion" by Michael
McBurnett explores the dynamics o f public opinion in presidential nomination
campaigns. Using data from the 1984 National Election Study's "rolling
thunder" survey, McBurnett demonstrates the series to have properties charac­
teristic o f chaotic behavior. Utilizing techniques discussed in the methodology
section, McBurnett shows how different analytic techniques reveal a complex
nonlinear deterministic chaos pattern to public support for Democratic presi-
12 Chaos Theory in the Social Sciences

dential candidates during the 1984 primary season. McBurnett's analyses and
findings should be o f profound interest to all serious students o f public opin­
ion. I f , indeed, the solution o f public opinion can be described as exhibiting
deterministic chaos, then the question is raised how one can deduce governing
or predictive equations from this time series. Certainly, McBurnett's study
suggests how and why drastic shifts in public opinion may occur, and the
consequences for predicting the evolution o f public opinion.

Chaos Theory and Economics

The third section o f the book examines chaos applications in economics.


A m o n g all the disciplines we cover, chaotic and, more generally, nonlinear
dynamical approaches are most developed in this field. This may be at least
partly explained by the mathematical rigor and statistical sophistication that
have typified economics for the past several decades. However, the interest in
chaos may also have resulted from an increasing dissatisfaction with orthodox
equilibrium-based models o f both micro- and macro-level economics phe­
nomena. Relatedly, the obvious difficulties that economists have encountered
in developing adequate predictive models o f behavior almost certainly have
helped explain the developing interest by many economists in applications o f
chaos theory.
J. Barkley Rosser, Jr.'s "Chaos Theory and Rationality in Economics"
provides an illuminating theoretical overview o f the implications chaos theory
has for orthodox microeconomic theory. Rosser points out that standard neo­
classical theory makes a number o f information assumptions and that eco­
nomic agents, more generally, are in possession o f some basic model o f
reality. The existence o f nonlinearities that are characteristic o f chaotic sys­
tems, however, calls into serious question such assumptions o f neoclassical
economic theory. The "sensitive dependence on initial conditions," especially,
means that the most seemingly trivial initial errors in economic judgment can
produce totally unexpected outcomes. Rosser employs these basic characteris­
tics o f chaotic systems to show how they can produce a variety o f economic
phenomena. Most important, he concludes w i t h a discussion o f the kind o f
decision-making rule that can be employed where chaos exists.
Brian J. L . Berry and Heja K i m ' s chapter is entitled " L o n g Waves 1790¬
1990: Intermittency, Chaos, and C o n t r o l . " Berry, whose earlier research fo­
cused on economic and urban geography, has in recent years devoted his
considerable abilities to demonstrating the existence o f economic long waves,
their origins and impact. This macroeconomic study addresses two questions.
First, the annual fluctuations as well as longer-run fluctuations in prices in the
United States over the 1790-1990 time period, and second, the fluctuations
and swings in the rate o f economic growth over the same period. Using
Introduction 13

graphical analysis, Berry and K i m demonstrate that the inflationary and stag­
flation cycles o f the last 200 years are characterized by a chaotic limit cycle.
Their w o r k shows how chaotic processes can be contained within a larger and
more extensive stable series. Berry and K i m also draw some very important
policy conclusions from their analyses o f the p o s t - W o r l d War I I period relat­
ing to Keynesian macroeconomic management techniques.
The forms o f human settlement in physical space are the subject o f
Dimitrios Dendrinos's "Cities as Spatial Chaotic Attractors." Using an it­
erative process that places a time series o f human activity in this space,
Dendrinos shows how human settlements such as cities can take the form o f
periodic, quasi-periodic, or nonperiodic (or chaotic) attractors. Dendrinos
indicates, for example, that chaotic patterns are the result o f laissez-faire-type
market processes. This analysis distinguishes two distinct forces in locational
choice: (1) those that determine the current location o f a population at particu­
lar points i n time w i t h i n a given space, and that are associated with the
attributes o f location at a distance from where populations are currently
formed, and (2) those locational forces that determine current location o f
populations and are associated with the location o f prior settlement activity.

Chaos Theory I m p l i c a t i o n s f o r Social S y s t e m s


M a n a g e m e n t a n d Social Science

The two chapters in this section serve the purpose o f examining what chaos
theory means for social systems management and public policy and for re­
search and knowledge generation in the social sciences. These chapters
attempt to provide insights to both the practical potentialities and the meth­
odological limitations o f chaos theory as a tool for both altering and under­
standing the dynamics o f social systems. These chapters raise the philosophi­
cal issues o f the relevance o f chaos theory to social systems and social science
investigation that must be considered i f this research paradigm is to remain
robust.
Kenyon B . De Greene's "Field-Theoretic Framework for the Interpreta­
tion o f the Evolution, Instability, Structural Change, and Management o f
Complex Systems" begins by pointing out that theories relating to the man­
agement o f complex systems have tended to lag behind changes in technology
and society. He goes on to point out that an increasing gap exists between
management capabilities and reality. De Greene develops a model for under­
standing organizational dynamics and change by employing a field-theoretic
framework. The author demonstrates the history o f field theory i n the natural
sciences and employs similar approaches to understanding organizational
management. Like many other authors, De Greene demonstrates the linkages
between macro-level phenomena, in this case the " f i e l d , " which is produced
14 Chaos Theory in the Social Sciences

by micro-level events and resulting feedback loops. De Greene's major


theoretical contribution is to apply this particular theoretical approach to
Kondratiev long wave behavior, showing how such waves encompass much
more than just economic waves, but also institutions, technologies, and
the like.
David L . Harvey and Michael Reed's chapter, "Social Science as the
Study o f Complex Systems," provides a capstone for this volume. The authors
begin w i t h a discussion o f some important epistemological issues. A m o n g
them is the question o f the relationship between the natural and social sci­
ences, and the role o f chaos theory as a bridge between t w o scientific and
intellectual traditions. Specifically, the authors "explore the circumstances
under which research strategies employing the deterministic chaos paradigm
can and cannot be deployed in the human sciences." As such, Harvey and
Reed provide a useful antidote to those who would uncritically apply non­
linear and mathematical methods and paradigms originally developed for the
natural sciences. Taking what they consider a "middle course," the work o f
the British philosopher Roy Bhaskar and his modified naturalist epistemology
becomes critical for understanding the form a future science o f society might
take. In elaborating upon this science, Harvey and Reed present a rigorous
demonstration o f how chaos theory fits into various modeling strategies em­
ployed in the social sciences. These authors also provide a vision o f both the
prospects for and limitations o f chaos theory as a means for enhancing our
understanding o f the behavior o f complex social systems.
Finally, this volume brings together a comprehensive bibliography o f
both the chaos literature from the natural sciences and the relevant chaos
literature from the social sciences. This definitive bibliography should serve
as a valuable resource for all chaos researchers, regardless o f the level o f their
mathematical or scientific sophistication and whether or not they are new to
the field or experienced chaos researchers.

Conclusion

The process o f knowledge acquisition in the sciences traditionally follows a


logical flow o f hypothesis development, quantification, testing, and validation
or falsification. Validation and replication then generally lead to theory devel­
opment. Such theory aims at explaining the behavior o f systems and expedites
prediction o f the future state or behavior o f the system. Such an approach to
theory development is founded on assumptions o f global stability and, implic­
itly, o f linearity in the relationships between variables. Stability in such rela­
tionships allows prediction. Thus, the behavior o f nonlinear systems chal­
lenges traditional notions o f theory development. By inhibiting prediction, a
fundamental element o f theory building is restricted. Thus, chaos researchers
Introduction 15

face the compound problem o f dealing with highly intractable data that are not
easily amenable to traditional empirical analysis, as well as w h i c h , by their
nature, may preclude or limit traditional hypothetic-deductive means o f
theory generation.
The dynamics in the relationships between variables over time in non­
linear systems may generate complexities that defy generalization. This diffi­
culty in developing such generalizations underscores the challenge o f building
theories that are relevant to complex social phenomena such as government
budgeting. The fundamental dynamics o f social phenomena clearly exacer­
bate theory building in the social realm. A t the same time, however, chaos
theory suggests a much richer and interesting w o r l d for the social scientist to
explore. For, as Heinz Pagels (1987, 73) has noted, "Life is nonlinear, and so
is just about everything else o f interest." Indeed, as the following chapters
convey, it is this richness and complexity that readers w i l l find most fas­
cinating.
Part 1
Chaotic Dynamics in
Social Science Data
CHAPTER 1

Exploring Nonlinear Dynamics w i t h a


Spreadsheet: A Graphical V i e w
of Chaos for Beginners

L . Douglas Kiel and Euel Elliott

The mathematical foundation o f chaos theory and the unique vernacular o f


this new science can deter some researchers from exploring the dynamics o f
nonlinear systems. Terms such as periodicity, sensitive dependence on initial
conditions, and attractors are not the usual vernacular o f the social sciences.
However, the modern microcomputer and electronic spreadsheet software
provide means for the novice to chaos research to explore the mathematics o f
chaos. The graphics capabilities o f spreadsheet software also provide a visual
means for exploring chaotic dynamics. This is particularly important consid­
ering the reliance o f chaos researchers on graphical analysis. The intrac­
tability o f nonlinear mathematics, while often defying solution, is now ex­
plored via visual analysis. This chapter should thus bring to light the amazing
behavior and visual imagery o f nonlinear dynamical systems and their rele­
vance to social science.
Fortunately, the dynamics o f time-based nonlinear systems can readily be
explored by researchers new to the study o f chaos theory. This exploration is
accomplished here via the use o f a simple algebraic formula and the computa­
tional and graphical powers o f an electronic spreadsheet. The electronic
spreadsheet readily allows the researcher to generate nonlinear time series and
then examine these series graphically. O n l y a minimal knowledge o f spread­
sheets is necessary for the reader to follow the examples in this chapter.
Readers are also urged to examine in greater detail the mathematical formula­
tion and its various dynamics presented here.
By examining the chaotic and, more generally, nonlinear behavior in this
chapter, the reader w i l l understand that a simple deterministic equation can
generate very complex behavior over time. This has considerable value for
social scientists as we learn that systems evolve from the simple to the com­
plex (Prigogine and Stengers 1984). This chapter also reveals the importance
o f history to social systems. The initial starting point o f a social system has
much to do with its eventual structure and behavior.

19
20 Chaos Theory in the Social Sciences

As noted in the introduction to this volume, nonlinear systems can take


on a wide array o f behaviors over time. Scientists have, however, classified
these behaviors into three distinct types o f time-based regimes. These behav­
ioral regimes are (1) convergence to an equilibrium or steady state; (2) peri­
odic behavior or a stable oscillation; and (3) chaos. The most widely used
mathematical formula for exploring these three behavioral regimes is a first-
order nonlinear difference equation, labeled the logistic map. This mapping
takes on the form

x1
t+ — kx,( 1 —
x)
t

The variable to be examined is the value x. The parameter, or boundary value,


o f the formula is a constant, k. Remember, chaotic behavior occurs w i t h i n
denned parameters. The subscript t represents time and is the current value o f
the variable x. The subscript / + 1 represents one time period o f the variable x
f o l l o w i n g the previous x r

Mapping this formula also requires an initial starting value. The starting
point, usually called the initial condition, is represented by the first value o f
x„ x . Once the first value o f x, and the parameter value are determined, a
0

simple "copy" command w i t h the spreadsheet can be used to generate the time
series. The copy command serves the purpose o f recursion or feedback by
using the previous value to generate the current value o f x . t

A couple o f rules must be followed when using the logistic map. First,
the initial condition must be a fractional value such that 0 < x < 1. Second,
0

the parameter value or constant, k, must be greater than 0 and less than 4.
Adventurous readers are welcome to explore higher values o f k. The follow­
ing is the general framework for inputting the logistic map into a spreadsheet.

1. I n cell A l , input a fractional value for x between 0 and 1. This


0

number is the initial condition.


2. I n cell B l , input the k value or constant. Remember, this number
must be greater than 0 and less than 4.
3. I n cell A 2 , input the formula ( $ B $ 1 - A 1 ) · (1 - A l ) . Cell A 2 repre­
sents the value x . t + l

4. Next, copy cell A 2 down to cell A 6 1 . This affords sixty iterations o f


the equation.
5. Then produce a line graph o f the values in cells A 1 - A 6 1 w i t h the
graph function in your spreadsheet. This graph provides a visual
image o f what happens as the system "evolves."
6. To change the dynamics o f the time series, simply change the values
in cells A l (x) and B l (k). To make a longer time series, just copy
more cells down from A 6 1 .
Exploring Nonlinear Dynamics with a Spreadsheet 21

Stable Equilibrium

A fascinating aspect o f the logistic map is that each behavioral regime occurs
w i t h i n defined mathematical boundaries. For example, values o f k between 0
and 3 (Stewart 1989) w i l l converge to an e q u i l i b r i u m . Table 1.1 reveals three
recursions o f the equation w i t h the same initial condition (x ) o f 0.97, but 0

w i t h varying values o f the parameter constant (k). This table reveals that the
iterations rapidly converge to a steady state and do not leave this state, or
mathematical point, once convergence occurs. Note that as the constant ap­
proaches three, convergence to a stability requires more iterations. Figure 1.1
reveals the graph generated for the constant (k) o f 2.827 w i t h 100 iterations.

TABLE 1.1. N u m e r i c a l I t e r a t i o n o f Stable Equilibria f r o m t h e Logistic M a p

k = 1.95 k = 2.35 k = 2.65 k = 1.95 k = 2.35 k = 2.65

0.97 0.97 0.97 0.48717949 0.57446809 0.62264174


0.056745 0.068385 0.077115 0.48717949 0.57446809 0.62264136
0.10437376 0.14971496 0.18859593 0.48717949 0.57446809 0.62264161
0.18228576 0.29915591 0.40552289 0.48717949 0.57446809 0.62264145
0.29066244 0.49270488 0.6388463 0.48717949 0.57446809 0.62264155
0.40204669 0.58737494 0.61141252 0.48717949 0.57446809 0.62264148
0.46879004 0.56955921 0.62960622 0.48717949 0.57446809 0.62264153
0.48560058 0.57612956 0.61798591 0.48717949 0.57446809 0.6226415
0.48709568 0.57388008 0.62561021 0.48717949 0.57446809 0.62264152
0.48717528 0.57467307 0.6206885 0.48717949 0.57446809 0.6226415
0.48717928 0.57439624 0.62390086 0.48717949 0.57446809 0.62264151
0.48717948 0.57449322 0.62181873 0.48717949 0.57446809 0.62264151
0.48717949 0.57445929 0.62317452 0.48717949 0.57446809 0.62264151
0.48717949 0.57447116 0.6222943 0.48717949 0.57446809 0.62264151
0.48717949 0.57446701 0.62286688 0.48717949 0.57446809 0.62264151
0.48717949 0.57446846 0.62249489 0.48717949 0.57446809 0.62264151
0.48717949 0.57446795 0.62273676 0.48717949 0.57446809 0.62264151
0.48717949 0.57446813 0.62257957 0.48717949 0.57446809 0.62264151
0.48717949 0.57446807 0.62268176 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62261534 0.48717949 0.57446809 0.62264151
0.48717949 0.57446808 0.62265852 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62263045 0.48717949 0.57446809 0.62264151
0.48717949 0.57446808 0.62264869 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62263684 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62264455 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62263954 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62264279 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62264068 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62264205 0.48717949 0.57446809 0.62264151
0.48717949 0.57446809 0.62264116 0.48717949 0.57446809 0.62264151
22 Chaos Theory in the Social Sciences

0.9-

0.8¬

0.7-

8 0.6-

I 0.5-

& 0.44

0.3-

0.2-

0.1-

0 I 111 I I I 11 11 11 11 I I I I I III 11 I I M I II I I I I I ! ! I ! I I ! II I I ! I I I II II I I I I I I I I I I I I I II I I ! I I I I I 11 I 11 I I I I I I I I

25 50 75 100
Iterations

x =037 fe = 2.827
o

Fig. 1.1. S t a b l e e q u i l i b r i u m

Periodic Behavior

A second type o f nonlinear behavior that can occur over time is periodic
behavior. Periodic behavior is cyclical or oscillatory behavior that repeats an
identifiable pattern. Such periodic behavior starts to occur when k > 3. This
regime initiates instability into the equation as the data start to oscillate. Such
a change in the qualitative behavior o f the time series is referred to as a
bifurcation, or a branching to a new regime o f behavior. This can be seen in
column one o f table 1.2. This first column represents a two-period cycle in
which the value o f x moves back and forth between t w o values. A t approx­
imately the (k) value o f 3.5 (Stewart 1989) a four-period cycle occurs in which
four numbers alternate in a consistent pattern. This four-period cycle is shown
in column 2 o f table 1.2 and is presented graphically in figure 1.2.
This process o f cycles doubling i n the number o f alternating and contin­
uous patterns o f values is labeled period doubling. It is this continuous bifur­
cation o f period doubling that eventuates in the "road to chaos" (Feigenbaum
1978). B y exploring the range o f (k) between 3.56 and 3.57, this period
Exploring Nonlinear Dynamics with a Spreadsheet 23

TABLE 1.2. N u m e r i c a l I t e r a t i o n of Periodic Behavior f r o m t h e Logistic M a p

k = 3.25 k = 3.5 k = 3.567 k = 3.25 k = 3.5 k = 3.567

0.97 0 97 0 97 0 49526517 0 8269407 0 81057266


0.094575 0 10185 0 1037997 0 81242714 0 50088422 0 54769367
0.27829935 0 32016802 0 33182132 0 49526517 0 87499726 0 8836362
0.65275867 0 76181161 0 79086073 0 81242714 0 38281968 0 3667706
0.73666056 0 63509139 0 58998192 0 49526517 0 82694071 0 82843549
0.63047328 0 81112611 0 86286891 0 81242714 0 50088421 0 50697817
0.75717435 0 5362019 0 4220694 0 49526517 0 87499726 0 89157631
0.5975494 0 87041298 0 87008697 0 81242714 0 38281968 0 34481474
0.78157337 0 39477979 0 4031981 0 49526517 0 82694071 0 80584785
0.55482842 0 83625048 0 85832504 0 81242714 0 50088421 0 55808244
0.80273 0 47927466 0 43375848 0 49526517 0 87499726 0 87971647
0.51465229 0 87349661 0 87609822 0 81242714 0 38281968 0 37744353
0.81180226 0 38675099 0 3871983 0 49526517 0 82694071 0 83817334
0.49653289 0 83011131 0 8463627 0 81242714 0 50088421 0 48382357
0.81246093 0 49359283 0 46382729 0 49526517 0 87499726 0 8908166
0.49519654 0 87485632 0 88708271 0 81242714 0 38281968 0 34693493
0.81242501 0 38318959 0 35729561 0 49526517 0 82694071 0 80817906
0.49526949 0 82724365 0 81910968 0 81242714 0 50088421 0 55297655
0.81242727 0 50019058 0 52851887 0 49526517 0 87499726 0 88173916
0.4952649 0 87499987 0 88884887 0 81242714 0 38281968 0 37194968
0.81242713 0 38281283 0 35240733 0 49526517 0 82694071 0 83326232
0.49526519 0 82693509 0 81404791 0 81242714 0 50088421 0 49558553
0.81242714 0 50089707 0 53995074 0 49526517 0 87499726 0 89168049
0.49526517 0 87499718 0 88605685 0 81242714 0 38281968 0 34452367
0.81242714 0 38281989 0 36012471 0 49526517 0 82694071 0 80552531
0.49526517 0 82694088 0 8219613 0 81242714 0 50088421 0 55878584
0.81242714 0 50088382 0 52199806 0 49526517 0 87499726 0 87942325
0.49526517 0 87499727 0 89002388 0 81242714 0 38281968 0 37823754
0.81242714 0 38281968 0 34914287 0 49526517 0 82694071 0 83886531

doubling can be more closely examined. Table 1.2, column 3 shows the result
of the value 3.567, an eight-period cycle (Stewart 1989). This process o f
period doubling continues, as k increases, to periods o f 32, 64, 128, 256, and
so o n , until the onset o f chaos.

Chaos

Chaotic behavior occurs at the (k) value o f 3.8 to 4 . This mathematical regime
represents another clear bifurcation or qualitative change i n a system's behav­
ior. Three divergent values o f (k) are shown in table 1.3 to show the diverse
forms that chaotic regimes can take.
24 Chaos Theory in the Social Sciences

I f-

0.1 'I I I I I I I I I I I I I I I I I i t I I I I I I I I I I I I I I i I I I I I I I I I I I I I M I I I I I I I I I I I

20 40 60
Iterations

- * = 0 . 9 7 fr = 3.50
o

Fig. 1.2. Periodic behavior, f o u r - p e r i o d cycle

Figure 1.3 reveals the chaotic series when k = 3.98 a n d x = 0.90. What 0

distinguishes chaos from the other regimes o f behavior is the lack o f pattern i n
its longitudinal behavior. Chaotic behavior does not repeat itself and is thus
labeled aperiodic. A close examination o f the decimals in the values o f (k)
evidences this point. The reader w i l l also note that chaotic behavior remains
w i t h i n definable parameters. W h i l e such chaotic behavior appears random, it
is not. Chaos can be generated by a deterministic equation.

S e n s i t i v i t y t o Initial C o n d i t i o n s

In nonlinear dynamical systems, operating in a chaotic regime, small distur­


bances can have explosive and disproportionate (nonlinear) effects. W h i l e
systems operating in a steady state or periodic regime w i l l "damp" such
disturbances, chaotic regimes tend to generate positive feedback and amplify
such disturbances. This phenomenon o f chaotic regimes is referred to as a
sensitivity to initial conditions. I n short, systems operating in chaotic regimes
are very sensitive to small changes. It is this sensitivity that has generated the
"butterfly metaphor" in chaos theory. Can the flapping o f a butterfly's wings in
Exploring Nonlinear Dynamics with a Spreadsheet 25

TABLE 1.3. N u m e r i c a l I t e r a t i o n of Chaotic Behavior f r o m t h e Logistic M a p

k = 3.8 k = 3.89 k = 3.98 k = 3.8 k = 3.89 k = 3.98

0.9 0.9 0.9 0.18600293 0.88477839 0.18185933


0.342 0.3501 0.3582 0.57534218 0.39656834 0.59217033
0.8551368 0.88509166 0.91497318 0.92842951 0.93088436 0.96118843
0.47073584 0.39563016 0.30963308 0.25250299 0.25027741 0.14847484
0.94674571 0.93012599 0.85076653 0.71723187 0.72991427 0.50319163
0.19158941 0.25281746 0.5053121 0.77067919 0.76687238 0.99495946
0.58855506 0.73482408 0.99488769 0.67158454 0.69545082 0.01996024
0.92020041 0.75799626 0.02024297 0.83812323 0.82389802 0.07785607
0.27904015 0.71357355 0.07893611 0.51555619 0.56440038 0.28574213
0.76447163 0.79506285 0.2893667 0.94908042 0.95636658 0.81229239
0.68420808 0.63382848 0.81842177 0.18364175 0.16232792 0.60684438
0.82105605 0.90282986 0.59145814 0.56968635 0.52895274 0.94956543
0.55830744 0.34126233 0.96170892 0.93154649 0.96923916 0.19060589
0.93708092 0.87448115 0.14656298 0.24231699 0.11597882 0.61401563
0.22404903 0.42698145 0.49782745 0.69767797 0.39883289 0.94326173
0.66063403 0.95175965 0.99498121 0.801509 0.93268669 0.21300576
0.8519475 0.17860241 0.01987452 0.60455083 0.24422287 0.66718455
0.47930525 0.57067696 0.07752849 0.90846267 0.71800866 0.88375632
0.94837256 0.95306855 0.28464095 0.31600134 0.78761696 0.40886972
0.18605577 0.17399539 0.81040951 0.82134908 0.65070553 0.96194719
0.57546828 0.55907466 0.61151083 0.55759212 0.88414971 0.1456871
0.92835725 0.95892462 0.94551003 0.93739596 0.39844881 0.49536021
0.25273825 0.15322008 0.20505283 0.22300214 0.93238381 0.99491432
0.71767419 0.50470295 0.64876454 0.6584343 0.24524209 0.02013807
0.7699482 0.97241396 0.90691907 0.85461458 0.7200328 0.07853545
0.67308629 0.10434944 0.33597916 0.47214431 0.78416786 0.28802318
0.83615632 0.36356186 0.88792671 0.94705143 0.65837716 0.81616199
0.52059592 0.90008623 0.39606122 0.19055107 0.87492586 0.59716556
0.94838807 0.34983162 0.95200299 0.85611717 0.42568503 0.95742424

China create a tornado i n Texas? I n a nonlinear system such small occurrences


may have massive results as behavior alters, changes, and perhaps, explodes
over time. Furthermore, systems w i t h very similar starting conditions i n their
evolutions may diverge to very different systems and structure over time. This
point has obvious implications for social scientists as we explore how vir­
tually identical systems generate unique histories.
The phenomenon o f sensitivity to initial conditions can also be examined
using the logistic map. This is best examined by comparing t w o time series
w i t h only slightly different initial conditions. By changing the initial condition
by a mere one m i l l i o n t h (the last decimal place), a system w i t h the same
parameters or boundary values can show very different results over time.
These t w o time series i n i t i a l l y appear quite similar and actually map each
26 Chaos Theory in the Social Sciences

20 40 60
Iterations

x = 0 . 9 0 k = 3.98
o

Fig. 1.3. Chaotic b e h a v i o r

other perfectly. But once the divergence starts, the time series continue to
behave quite differently, as can be seen in figure 1.4.

Attractors

Even researchers and students new to the study o f chaos are familiar w i t h the
notion o f order in chaos. The analyses o f nonlinear time series show that a
deeper underlying order exists in these diverse types o f behavior. This order
was discovered via graphical analysis o f chaotic time series. Yes, to under­
stand nonlinear systems, look at the pictures they generate.
This deeper order is discovered by an investigation o f the attractors o f a
nonlinear system. Attractors provide a qualitative assessment o f dynamic
systems in motion (Mosekilde, A r a c i l , and A l l e n 1988, 21). Baumol and
Benhabib (1989, 91) define an attractor as "a set o f points toward which
complicated time paths starting in its neighborhood are attracted." Pool
(1989, 1292) defines an attractor as "the set o f points in a phase space
corresponding to all the different states o f the system." More simply, the term
attractor is used because the system's temporal evolution appears to be consis-
Exploring Nonlinear Dynamics with a Spreadsheet 27

Fig. 1.4. S e n s i t i v i t y t o i n i t i a l c o n d i t i o n s

tently " p u l l e d " to identifiable mathematical points. The attractor functions as


an abstract representation o f the flow, or m o t i o n , o f a system. I n short, the
attractor stores information about a system's behavior over time. The attractor
is used as a means for examining the structure o f the underlying order w i t h i n a
nonlinear system.
The examination o f an attractor is conducted by a mapping o f the data
onto a phase space (Thompson and Stewart 1986). A phase space represents a
graphic backdrop for presenting the motion o f time-based data. The examina­
tion o f an attractor is conducted i n a (/t + 1 phase space (Baumol and
Benhabib 1989, 91). I n this case, t (time) represents the current value o f x,
while t + 1 (time + 1) represents the next value o f x. The t is plotted on the
horizontal axis and t + 1 is plotted on the vertical axis. This method o f
plotting the data reveals the relationship between a previous period's mea­
sured result relative to the current time period's measured result. When plot­
ting a one equation system such as the logistic map we refer to the phase space
as a phase plane.
The attractors o f nonlinear systems thus represent a dynamic structure
28 Chaos Theory in the Social Sciences

that traces the longitudinal behavior o f a system. Studies o f these attractors,


generated by various nonlinear equations, evidence an enormity o f shapes and
patterns (Gleick 1987). These attractors thus represent the structures that
describe and dominate a nonlinear system during its evolution. Each o f the
three nonlinear regimes described above represent a unique attractor type.
Steady state attractors converge to a point and remain there. Attractors from
periodic regimes are called limit cycles and oscillate around a set o f denned
mathematical points, creating a circular pattern. A chaotic or strange attrac-
tor takes on a multitude o f surprising shapes.
Attractors can be generated using an electronic spreadsheet. This is
accomplished by generating an XY graph as follows:

1. Create a chaotic time series for about sixty iterations in column A .


This series should be the horizontal axis (0·
2. I n cell C I , place a zero. Then recreate the same chaotic series in
column A , under the zero in column C. Make sure you put all values
from column A , including the initial value, in column C. This series
w i l l be the vertical axis (t + 1).
3. Delete the last cell at the end o f the column C (/ + 1) to ensure each
column has the same number o f time periods.
4. Generate the graph to produce the attractor.

When the logistic map is operating i n a chaotic regime it creates a h i l l -


shaped attractor, or parabola. I f the lines are deleted from the graph and only
symbols are used, this underlying structure appears. This structure is the
order in chaos. Given the nonlinearities that exist in social science data,
social scientists may want to explore their data sets using phase space map­
pings. Researchers may find a wide range o f attractors that describe the order
in the apparent chaos o f social science data.

Conclusion

The dynamism o f social systems suggests that each behavioral regime noted
above can appear w i t h i n the long-term behavior o f a nonlinear system. Be­
cause dynamical systems are historical systems they can reveal many types o f
behavior over time. Thus each behavioral type does not reflect permanent
commitment only to that behavioral type, but rather reflects one possible type
that may occur for a period during the life o f a system.
Exploring the behavioral regimes o f nonlinear systems should provide
social scientists w i t h a foundation for discovering such behavior in social
phenomena. For example, notions o f periodicity have always been used in
analyses o f society, as phenomena ranging from fads to vehicular traffic reveal
Exploring Nonlinear Dynamics with a Spreadsheet 29

such oscillatory behavior. Analysts may, however, have a more difficult time
finding social systems that are truly in a steady state. Social systems do seem
to oscillate at varying frequencies. Finally, chaos may provide a means for
understanding the uncertain nature o f social systems as both historical and
evolving systems.
CHAPTER 2

Probing t h e Underlying Structure in Dynamical


Systems: A n Introduction t o Spectral Analysis

Michael McBurnett

This chapter examines one method o f investigating the underlying dynamics


of time-series data. It introduces the reader to spectral analysis, a tool for
evaluating the frequency properties o f a time series. This is distinguished
from the analysis o f the properties o f time series in the time domain, which is
the subject o f most recent research in time-series analysis. Since this discus­
sion serves as an introduction to the topic, I have eschewed mathematical
rigor.
In this chapter, I focus on three basic types o f time series—periodic time
series, chaotic time series, and a random time series—and explore their
frequency properties with numerous examples. I also introduce noise into
some o f the dynamic processes to illustrate how noise affects spectral anal­
ysis. In the final section o f the chapter, I introduce a time series whose
properties are unknown and analyze its frequency properties. This time series
is constructed from survey data collected by the Center for Political Studies.
I also introduce some common problems that research has uncovered in
the spectral analysis o f time series, particularly as these problems relate to the
analysis o f chaotic time series. These problems involve the inability o f spec­
tral analysis to clearly discern cycles, even when they are known to be pre­
sent. Sometimes this is a data problem: the time series is too short or the
signal-to-noise ratio is too high. Other times the signal resembles noise even
though it is not. Unfortunately, the latter case is the more critical problem and
the more likely problem when dealing with chaotic processes.
The chapter is organized as follows. First, I provide the necessary mathe­
matical introduction to the spectral density and spectral distribution functions.
I show how to interpret these measures and explain some o f the problems i n
their use. Then I demonstrate by example how to use spectral analysis and
introduce t w o specific cases: pure noise and periodic. Here, I show how the
presence o f noise can interfere w i t h the resolving power o f the technique.
T h i r d , I introduce several k n o w n chaotic time series and examine their respec-

31
32 Chaos Theory in the Social Sciences

five spectra. This section shows that no two time series have identical spectra
and that there is no single characteristic spectrum for chaotic dynamics.
Fourth, I introduce an empirical time series constructed from surveys con­
ducted during the 1984 Democratic nomination race. These data are subjected
to spectral analysis and compared to some o f the spectra from the prior
section. I determine that spectral analysis is not conclusive in determining the
dynamics o f this time series. I conclude w i t h a discussion o f the use o f
spectral analysis and suggest the use o f some additional measures that can be
used to quantify chaos i f it is present in a time series.

Mathematical Background

Inspection o f a time series may lead one to suppose that it contains a periodic
oscillatory component with a k n o w n wavelength. This can be represented by

Y, = R cos (cot + 9) + E, (1)

where co is the frequency o f the oscillation, R is known as the amplitude o f the


oscillation, 6 is the phase, and E, is a stationary random series. Sometimes the
frequency,/, is expressed a s / = W/2TT, which is a measure o f the number o f
cycles per unit time and is easier to interpret. I use this expression throughout
in the interpretation o f data. The period o f a cycle, called the wavelength, is
given by l / / o r ITTICD. Figure 2.1 shows a graph o f a time series w i t h / = 1/6
and wavelength 6.
Equation 1 is extremely simple and, in practice, the variation in an
observed time series may be caused by variation at several different frequen­
cies. For example, presidential popularity may show variation at yearly, quar­
terly, monthly, and even weekly frequencies. This means that the series shows
variation at high (weekly), medium (monthly or quarterly), and low (yearly)
frequencies. Equation 1 can be generalized to account for the combination o f
variation in the observed series by

Y, = S t f ; COS (ü)jt + Oj) + E„ (2)

where Rj is the amplitude at frequency


In the t w o equations shown thus far, it should be noted that neither is
stationary i f R, 6, {Rj\, and {0,·} are constants since this condition implies that
E{Y ) w i l l shift over time. It is customary to assume that {Rj} are uncorrected
t

random variables uniformly distributed on (0, 2TT), which are fixed for a
particular value o f the process (Chatfield 1992). This assumption allows the
Probing the Underlying Structure in Dynamical Systems 33

I I I I I I I I L

~i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—i—I r
1 2 3 A 5 G 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Iterate

Fig. 2 . 1 . S a m p l e s i n u s o i d a l f u n c t i o n w i t h w a v e l e n g t h 6 a n d
f r e q u e n c y 1/6

treatment o f time series suspected to contain more than one oscillating compo­
nent as a stationary series.
Using the trigonometric identity cos (tot + 6) = cos wt • cos 6 — sin
cot • sin 6, equation 2 can be written as

y, = 2 (oj cos ci)jt + bj sin (o ) + E ,jt t (3)


7= 1

where a- = Rj cos dj and bj = —Rj sin 0,. Clearly, from equations 2 and 3 we
see that only a finite number o f frequencies are represented here (the index j
counts only from 1 to k). W h y are there not more, indeed, an infinite number
of frequencies? Wiener (1949) showed that when k—*«>, any discrete process
measured at unit intervals can be represented as

Y. = I cos (x)t du(a)) + sin (at dv(a>) (4)


Jo Jo
34 Chaos Theory in the Social Sciences

where u(<o) and v(&>) are uncorrelated continuous processes defined for all a> in
the range ( 0 , TT). This equation is called the spectral representation o f the
process. Intuitively it helps to think o f Y, as a linear combination o f orthogo­
nal oscillating terms.
One may also wonder w h y the upper l i m i t o f the integrals in equation 4 is
•nrather than + ° ° . I f the process were continuous, the l i m i t w o u l d be + ° ° . We
1
are concerned w i t h a discrete process measured at unit intervals. Hence,
there is no loss o f generality in restricting co to the range (0, TT), because

cos tat k, t integers where k is even,


cos (TT — o))t k, t integers where k is odd.

Variation at frequencies higher than TT cannot be distinguished from variation


in a corresponding frequency in the interval (0, TT). The frequency <a = IT is
called the Nyquist frequency. I f the change in the unit interval over which
measurements o f the time series are taken is given by A?, then the Nyquist
frequency is given by 7r/At. The introduction o f the spectral representation o f
the time series serves to demonstrate that all frequencies in the interval (0, TT)
may contribute to the variation in the observed series.

The Spectral Density and Spectral


Distribution Functions

The power spectral distribution function, denoted F{co), is the empirical mea­
sure that describes how the oscillations in a time series, i f they are present, are
distributed w i t h respect to the frequency. This measure is the one most fre­
quently consulted when examining a time series for a periodic component.
Suppose we have a stochastic stationary process. Then the spectral distri­
bution function is a continuous, monotone increasing, bounded function over
the interval ( 0 , TT). The function may be differentiated w i t h respect to <o in
(0, TT). The derivative is denoted / ( w ) = dF(u))/da>. This is the (power)
spectral density function, which is simply referred to as the spectrum.
When the derivative exists, and we have a stationary stochastic process
w i t h an autocovariance function y(k), we can express the relationship be­
tween the autocovariance function and the spectral distribution function as

(5)

Set k = 0 and we have


Probing the Underlying Structure in Dynamical Systems 35

"KO) = 0-2 (6)

The physical meaning o f the spectrum is that f(w) du> represents the contribu­
tion to the variance o f components with frequencies in the range (w, <a + dw).
When the spectrum is drawn, the total area under the curve given by equation
6 is equal to the variance o f the process. A peak in the spectrum indicates an
important contribution to variance in the appropriate interval. I n practice one
searches for peaks in the spectrum when it is drawn, as this is an indication
that periodicity is present in the series.
The integral i n equation 6 has another use in analysis. We are actually
dealing w i t h a finite number o f frequencies in the interval [ 0 , IT]. Summing the
contribution to the variance o f each o f these frequencies produces a charac­
teristic shape. For a simple sinusoidal function with a single wavelength, such
as that given by equation 1, one observes a single peak that translates into a
step function. This is seen by using equation 1 to generate some data, say
2,500 iterates, and computing its spectrum. Then i f one computes the running
sum o f the contribution to the variance o f the spectral density function and
plots that sum against the cycles per iteration, the result w i l l have a single
step. More generally, for a series w i t h t w o incommensurate (nonoverlapping
or orthogonal) cycles, the result w i l l have two steps separated by the distance
corresponding to their respective frequencies, each o f which provides an
important contribution to the variance o f the process. Continuing, a process
that exhibits n cycles w i l l show n steps and each step w i l l contribute only a
small amount to the total variance o f the process. Figure 2.2 shows how this
phenomenon translates into problems when the number o f cycles present is
large.
This figure shows five selected spectral density functions displayed as
described by equation 6 above. The equations used here are: a simple periodic
function (equation 1), an equation which contains a four-cycle (equation 7
following), the Lorenz equations (equation 8 following), a chaotic version o f
the logistic equation (equation 7 with its driving parameter set to 3.99), and a
random time series with mean zero.
The variance o f the spectrum is given by the sum over all the frequencies
present between 0 and TT. Since all computer implementations o f spectral
analysis report discrete frequencies, the m a x i m u m number o f which is given
by one-half the length o f the time series, the procedure for computing equa­
tion 6 is to sum the spectral density function over all cycles. A l l the examples
used in figure 2.2 use time series with length 2,500. The vertical axis in this
figure is arbitrary; all the spectral densities have been normalized to match the
scale for the random time series to allow comparison o f their respective
shapes.
36 Chaos Theory in the Social Sciences

n 1 1 1 1 i r
0 .1 .1667 .25 .3 A .5
Cycles

Fig. 2.2. C o m p a r i s o n o f spectral d i s t r i b u t i o n f u n c t i o n s f o r f i v e selected


dynamic equations

The distinctions are quite dramatic. The spectral density for equation 1 is
shown as a step function, which means that all the variance in the spectrum is
accounted for by a single peak. This is also the case for the logistic equation
with the driving parameter at 3.5, which produces a four-cycle. Its spectral
distribution function has a single peak at 0.25 cycles per iteration and a
corresponding step i n this figure at 0.25 cycles. The Lorenz equations produce
a very different spectral density, which is mainly confined to the left region o f
this figure. This figure shows that the Lorenz equations are dominated by a
number o f short (high) frequencies spread over a relatively narrow range,
which accounts for all the variance. The two remaining equations are the
chaotic version o f the logistic equation and a random time series. These
functions have spectral density functions that are indistinguishable from one
another. Both show that all frequencies contribute about the same amount to
the total variance in the time series. The chaotic logistic equation has frequen­
cies o f all cycles present in the series and these cycles are uniformly distrib­
uted across the entire spectrum. This makes it is easy to see why the random
series and this equation look similar in this graph: each frequency contributes
about the same amount to the total variance. This is a serious impediment to
the sole use o f this technique for attempting to identify chaotic dynamics
Probing the Underlying Structure in Dynamical Systems 37

(Glass and Mackey 1988). It shows that deterministic chaos may not be
distinguished from noise in some instances by spectral analysis. Despite this
difficulty we can often learn about the dynamics o f a process from the fre­
quency properties o f a time series. Below I introduce examples o f the spectral
distribution o f a variety o f time series.

I n v e s t i g a t i o n of Selected T i m e Series

The empirical investigation o f the spectrum o f a process involves the exam­


ination o f a graph o f the spectral density function over the range (0, TT). One
first considers the theory underlying the process, and then examines the time
series itself. I f one suspects that a periodic component is present, then the
series is analyzed. Suppose we have a random process represented by a time
series governed by a random process w i t h mean 0 confined to the range
(—3.25, + 3 . 2 5 ) . Such a process can be generated w i t h any random number
generator. For a purely stochastic process, we should suspect that no fre­
quency w i l l contribute more to the variance o f the spectral distribution func­
tion than any other. I n other words, the spectrum w i l l be essentially flat over
its entire range, though some peaks may be present. Such a spectrum is shown
in figure 2.3.
It is common to rescale the vertical axis by taking the logarithm o f the
spectral density estimates and to scale the x-axis by converting the frequency
to cycles per unit time. The latter is accomplished by multiplying the fre­
quency by dlltr, where d is the number o f observations per unit time (d = 1
for discrete series). This follows from the fact that variation at frequencies
higher than TT cannot be distinguished from a corresponding frequency in the
interval (0, TT).
Examination o f figure 2.3 shows that the random time series (N = 1,000)
is featureless. Put another way, we could say that this spectrum has the same
power at all cycles; it could be easily described by a line o f slope zero. Having
the same power at all frequencies means that each component o f the spectrum
contributes about the same amount to the total variance o f the function. This
spectrum is the least interesting case but it provides a backdrop against which
to check other, more interesting time series.
Let us reconsider equation 1. The function Y = R cos (<ot + 6) + E
t r

w i t h R = 2, w = (77/3), 6 = (TT/6), and E, = 0 completes one wavelength for


each six iterations. This is confirmed by inspection o f figure 2 . 1 . Recall that
the Nyquist frequency is a method o f scaling the interval over which the
spectrum is to be estimated. Since the function given by equation 1 completes
one cycle each six iterations, we should anticipate a peak at 1/6 o f the Nyquist
frequency. I n translating the frequency into cycles per unit time, we normalize
the ordinate to unity, so the expectation is that the peak is present at . 16667.
38 Chaos Theory in the Social Sciences

-7 -

-8 -

Cycles per Unit Time

Fig. 2.3. Featureless s p e c t r u m f o r a r a n d o m process. (/V = 1,000.)

Figure 2.4 shows the spectrum for equation 1, and the spike is present pre­
cisely where it is expected. This spectrum shows that the variance in equation
1 is accounted for by a single frequency that corresponds to the precise
wavelength o f the time series. This example demonstrates convincingly the
2
resolving power o f spectral analysis w i t h periodic data.
These two examples use a purely random process and a known periodic
process to illustrate what to look for i n the spectrum o f a time series. In the
next section, I introduce three specific time series and their respective spectra.
First I show how the spectrum responds to the presence o f noise in the time
series. Second, because our interest lies in the examination and characteriza­
tion o f (potentially) chaotic time series I demonstrate the analysis o f a selec­
tion o f time series w i t h k n o w n chaotic properties both w i t h and without noise
present. T h i r d , I analyze the spectrum o f a time series derived from survey
data that may have some chaotic properties.

Spectra o f Selected T i m e Series

In order to illustrate the spectra o f chaotic series, it is useful first to demon­


strate what happens when noise is present in a periodic time series. Social
Probing the Underlying Structure in Dynamical Systems 39

o
<D
Q.
CO
<5
S
o
ÇL
o>
o

~i i r
25 .3
.1 .16667.2
Cycles per Unit Time

Fig. 2.4. S p e c t r u m f o r a periodic f u n c t i o n

scientists are well aware that noise is present in all measurements o f social
processes. The error in measurements can be due to random or systematic
measurement error, the failure to account for a relevant variable, or unknown
other means. I n statistics, measurement errors can be problematic (Achen
1986; Greene 1978; Kenny and McBurnett 1992). I n deterministic dynamics,
the presence o f error can have a catastrophic effect (Peitgen et a l . , 1992). As a
matter o f fact, all models o f dynamical systems w i l l contain some error. One
source o f error is the computer. For many models, the only source o f error is
that due to computer rounding error. I n physical experiments, error can creep
into the measurements from imprecision in the instruments used to take mea­
surements. Social scientists are well aware that many measurements we make
are imprecise. Inferences from surveys rely on sampling theory and the sam­
ple error increases as the sample size decreases. What we need to know is how
stochastic error affects the resolving power o f the spectral distribution func­
tion. I first examine t w o variants o f a well-known equation to illustrate the
resolving power o f the spectral density function in a periodic function, the
periodic function w i t h an additive stochastic component, and a chaotic variant
of the same equation. I then turn to an examination o f t w o additional w e l l -
k n o w n chaotic time series.
40 Chaos Theory in the Social Sciences

A Periodic Function

The logistic equation

Y, = 3.5 · y , _ , ( i - y,_,) (7)

with Y = 0.5 produces a time series that contains a four-cycle. This series
contains 500 elements. The four-cycle is a stable attractor. The series con­
verges to the four-cycle by the third iteration. The points are .38281, .82694,
.50088, and .87499. The series contains no noise except that due to computer
rounding errors.
Figure 2.5 shows the spectrum for the logistic equation above. A four­
cycle translates into .25 cycles per unit time, and the graph shows a very sharp
peak at precisely that value. The remainder o f the graph is flat. What happens
when error is present?
The same time series produced by equation 7 is used again, but 20
percent error has been added to each observation. The error is normal, with
variance equal to .07. This series also contains 500 data points. In this case,

.03

.025

.02

.015 -

.01

.005

.05 .1 .15 .2 .25 .3 .35 .45


Cycles per Unit Time

Fig. 2.5. S p e c t r a l d e n s i t y of t h e l o g i s t i c e q u a t i o n w i t h p a r a m e t e r = 3.5.


(The t i m e series c o n t a i n s a f o u r - c y c l e in t h e u n i t interval.)
Probing the Underlying Structure in Dynamical Systems 41

the spectrum is recovered precisely, with the sharp peak at .25 cycles per
iteration, as before. The relatively large noise component manifests itself as
background noise surrounding the sharp peak. Figure 2.6 shows that noise
does not necessarily compromise the resolving power o f spectral analysis.
What does the spectrum o f a series with multiple cycles look like?
I f the driving parameter in equation 7 is changed to 3.6, the logistic
equation produces a chaotic time signal. This means that the series contains 1,
2, 3, 4 , . . . , oo-cycles. The spectrum for this series should be quite different
from the previous series, but the length o f this series is not sufficient to allow
cycles whose period exceeds 500 to be determined from the data. Figure 2.7
shows the spectrum for this series and it differs dramatically from the previous
two. Here a sharp peak is seen again at .25 cycles, but this peak is straddled
by t w o additional peaks, which are straddled by t w o additional peaks, and
these too are bounded by another t w o peaks, after which the resolution ceases
to allow a clear identification o f the presence o f additional peaks. This chaotic
signal contains multiple cycles as well as broad background noise and the
analysis o f its frequency spectrum allows one to identify the presence o f some
of them. Notice that the spectrum shows that the time series appears to be

.04 -

.02 -

.015 -

Cycles per Unit Time

Fig. 2.6. Spectral d e n s i t y of t h e logistic e q u a t i o n w i t h parameter = 3.5.


(The t i m e series c o n t a i n s a f o u r - c y c l e in t h e u n i t interval.)
42 Chaos Theory in the Social Sciences

.01 -

® .00875 -
o
.9 .0075 -
.22
| .00525 -

S -005 -
>>
c .00375 -
CO
Q
« .0025 -
o
CO
<g- .00125 -

0 -
I I ""I I I I i 1 1 1 n
0 .05 .1 .15 .2 .25 .3 .35 A .45 .5
Cycles per Unit Time

Fig. 2.7. S p e c t r a l d e n s i t y of t h e l o g i s t i c e q u a t i o n w i t h p a r a m e t e r = 3.6.


(The t i m e series c o n t a i n s m u l t i p l e cycles.)

periodic, while it is k n o w n that this series is chaotic. We must recognize that


spectral analysis does not allow the identification o f any cycles that take
longer than the length o f the time series to complete.
Indeed, one prefers a time series o f sufficient duration so that long cycles
have ample time to complete several oscillations. W i t h social science data we
have to recognize that this requirement is almost never met; hence, we run a
constant risk o f failing to characterize correctly the periodicities present in the
data. Note also that the use o f spectral analysis to develop a model specifica­
tion containing periodic components may also fail to characterize correctly the
"true" model because the data fail to span the full range o f oscillations. I f one
is prepared to assume that the length o f the available time series is sufficient to
explore the range o f oscillation present in the true time series, then the
examination o f the spectral distribution function is appropriate.

The Lorenz and Rossler Equations and Their Spectra

The most well-known system o f equations in chaos theory is the Lorenz


equations. These equations comprise an early deterministic model o f the
Probing the Underlying Structure in Dynamical Systems 43

weather formulated by Edward Lorenz (1963). The system o f equations ( i n


dot notation) is

JC] = 10(— x , + x)
2

x 2
=
28x| X2 -^1-^3 (^)
x 3 = —(8/3)x + 3 xx-
x 2

When integrated numerically they appear to possess extremely complicated


dynamics (Hale and Kocak 1991).
The spectrum for the Lorenz equations has been studied extensively.
Here I reproduce the work o f Farmer et al. (1980). Figure 2.8 shows the
spectrum for the Lorenz equations for the parameter values in equation 8
above. The equations were integrated numerically with the Runge-Kutta inte­
grator in Phaser for fifteen iterations with a step size o f 0 . 0 0 1 . The result is a
discrete time series o f length 15,000. The variable J C , _ I ( t ) is used to investi­
gate the spectrum.
This spectrum is essentially featureless, w i t h high power at low frequen-
44 Chaos Theory in the Social Sciences

cies, and it decays rapidly. The peaks that are present are not well defined and
it appears that no periodicities are easily discernible. This result confirms
results reported by Farmer et al. (1980) and is one characterization o f chaotic
time series.
The Rossler equations are another system o f dynamic equations k n o w n
to display chaotic dynamics. The Rossler equations describe the spread o f
disease and have been used effectively to model measles and whooping cough
epidemics in children (Rossler 1976; Schaffer et a l . 1986; Schaffer 1987;
Schaffer et al. 1990). The system o f equations is again written in dot notation.

i , = ~(x 2 + x) 3

x2 = X] + -2x 2 (9)
x 3 = .2 + x xt 3 — 5.7x3

The parameters given in this system o f equations describe what is called "the
simple Rossler attractor" (Farmer et a l . 1980). The equations were integrated
numerically with the Runge-Kutta integrator i n Phaser for 1,000 iterations
with a step size o f 0 . 1 . The result is a discrete time series o f length 10,000.
The spectrum for this system o f equations (using x ), as shown in figure
u

2.9, displays features dramatically different from the Lorenz spectrum. The
Rossler equations contain many identifiable peaks corresponding to cycles o f
particular length, w i t h broad background noise evident between the peaks.
The logistic equation, the Lorenz equations, and the Rossler equations
all display different spectra. Where periodicities are k n o w n to exist a priori,
spectral analysis clearly identifies them as w i t h the logistic equations. The
Lorenz and Rossler systems are k n o w n to contain aperiodic cycles. This is
evident in the spectrum for the Lorenz system, which displays high power at
low frequencies w i t h rapid decay in the spectrum. In contrast, the Rossler
system has broad background noise and clearly identifiable peaks correspond­
ing to periodicity at particular frequencies.
What we observe in these three examples, w i t h and without noise, is the
full range o f possibilities in the spectral distribution functions o f chaotic data.
The spectrum for the logistic equation exhibits period doubling, which is a
well-known route to chaos. The Lorenz equations, in contrast, show no iden­
tifiable structure other than high power at l o w frequency with a decay in
power as the frequency increases. This pattern resembles the spectrum for
many economic time series. (See Granger and Newbold 1986, 66 for a similar
spectrum derived from the Federal Reserve Board index o f industrial produc­
tion.) Finally, the spectrum for the Rossler equations shows some identifiable
peaks at a variety o f frequencies that appear above broad background noise.
W h y do we observe so much variability in these spectra, each o f which is
Probing the Underlying Structure in Dynamical Systems 45

Cycles per Unit Time

Fig. 2.9. S p e c t r u m f o r t h e Rossler e q u a t i o n s . (This s p e c t r u m has d i s -


t i n c t features.)

k n o w n to be chaotic? The problem lies in the difficulty in distinguishing


between signal and noise in chaotic series. As Glass and Mackey describe the
problem:

[b]road-band power spectra, perhaps with superimposed peaks, are often


associated w i t h chaotic dynamics. Unfortunately "noise" is also associ­
ated w i t h broad-band spectra, and consequently the presence o f a broad­
band spectrum is not adequate to establish chaos as opposed to noise
(1988, 48).

The fact is that the spectrum provides us with only one piece o f evidence for
chaotic dynamics and even that may be suspect. The examples used here were
taken from time series whose dynamics are known to be chaotic. I f , for
example, the driving parameter in the logistic equation were to be shifted
downward by only 0.1 (to 3.5), the spectrum would show the single peak
identified in figure 2.4, rather than multiple peaks surrounded by background
noise.
46 Chaos Theory in the Social Sciences

Spectrum for the Public Opinion Series

In 1984, the Center for Political Studies conducted a telephone poll o f the
national electorate, using a probability sample over a very long period. This
survey is k n o w n as the Rolling Cross-Section. Although it has some draw­
backs, such as a fairly large sampling error (Allsop and Weisberg 1988), it is
the only national sample o f its type. This analysis uses that portion o f the
survey containing those respondents interviewed after January 1 1 , 1984, and
before June 12, 1984 (153 days). Those respondents were interviewed during
the period in which all primary and caucus activity took place.
The survey asked respondents to rate the chances o f the various Demo­
cratic candidates winning their party's nomination. As Bartels points out,
most respondents failed to rate candidates' chances as i f they had probability
theory i n m i n d (1988, app. A ) . Though Bartels used this survey for a quite
different purpose, I make use o f his coding scheme, which measures each
individual's perception o f each candidate's chance o f winning the Democratic
nomination (1988, 3 2 2 - 2 3 ) .
The individual responses were averaged according to the date o f the
Probing the Underlying Structure in Dynamical Systems 47

responses. This resulted in a measure o f the perceived mean daily probability


o f each candidate's chance o f winning the Democratic nomination. I n order to
see i f any interesting dynamical patterns were evident, I smoothed the data
using a five-day moving average. Smoothing a time series that contains noise
may allow a cyclical or otherwise interesting pattern to be identified visually,
i f one is present. Smoothing reduces, but does not eliminate, the level o f noise
in the time series.
Figure 2.10 shows the phase portrait for Gary Hart's chances o f winning
the 1984 Democratic nomination. For a review o f how phase portraits are
generated, see chapters 1 and 8 in this volume. The dynamical time path
indicates that the probability Hart w i l l w i n the nomination approaches the
value .25. This matches Bartels's observation o f Hart's chances o f winning
the nomination (1989, 251). Use o f a moving average does not alter the gross
structure o f a time series, it merely serves to accentuate what is already
present. The approach appears to be quite smooth, given the known level o f
noise in the data. The time path progresses upward from the origin, then
appears to approach a limiting value. People's perceptions o f Hart's chance o f

.04 .06 .08 .1 .12 .14 .16 .18

Lag10 Smoothed Probability of Jackson Nomination

Fig. 2 . 1 1 . Phase p o r t r a i t o f p r o b a b i l i t y o f J a c k s o n n o m i n a t i o n time


series
48 Chaos Theory in the Social Sciences

winning the nomination appear to cycle about a particular value. Notice that
Hart is always thought to have a chance o f winning the nomination. He is not
eliminated from the competition; he is considered to be a contender through­
out the campaign. This pattern in public opinion matches the pattern in the
electoral results closely. The expectation is that these data may contain a cycle
w i t h long periodicity, as do the electoral results.
Figure 2.11 shows the dynamics o f the probability o f Jesse Jackson
w i n n i n g the nomination. The dynamics here are anything but smooth and
orderly. The graph contains a shape reminiscent o f a figure eight. The dynami­
cal behavior cycles i n one direction and then reverses, and the two lobes
revolve about separate and distinct values. Though the data are obviously
noisy, the phase portrait shows that the dynamical pattern is bounded, that is,
the dynamics remain near the attractor, and that this does not appear to be an
ordinary cycle.
Figure 2.12 shows the phase portrait o f the probability that Walter
Mondale w i l l w i n the 1984 Democratic nomination. Though it is difficult to see
in the figure, the dynamics indicate that the perception o f Mondale's chances o f
winning increases rapidly at first, cycling about one value (near 0.50), and

.7 -
c
o
'« -65 -
c

CD

S .45 -

k_
0_

.3 .35 .4 .45 .5 .55 .6 .65 .7

Lag5 Smoothed Probability of Mondale Nomination

Fig. 2.12. Phase p o r t r a i t o f p r o b a b i l i t y o f M o n d a l e n o m i n a t i o n time


series
Probing the Underlying Structure in Dynamical Systems 49

then drops to a lower value (near 0.35), about which the series continues to
cycle. Toward the end o f the race the mean perception o f his chances rises
beyond the first level (near 0.60) and continues to cycle about that value. This
graph shows extremely complex behavior but it is hidden by the noise present
in the data.
Taken together, these figures reveal what may be a strange attractor
resembling the Lorenz attractor from climate dynamics (Lorenz 1963; Schuster
1989). Strange attractors are associated with chaotic dynamical systems
(Schuster 1989). The identification o f this type o f structure underlying the
dynamics o f public opinion is important because o f the implications this has
for forecasting the outcome o f nomination races as well as the study o f change
in public opinion over time generally (Huckfeldt 1990). For example, i f the
time series here can be shown to be chaotic, then we w i l l gain a deeper
understanding o f why polls are so variable while election outcomes are rela­
tively easy to forecast (Gelman and K i n g 1993).
The spectral distribution function for the Mondale time series has some
features i n common w i t h the chaotic logistic spectrum. This spectrum is
displayed in figure 2.13. It appears that all frequencies contribute about
equally to the spectral density. The spectrum is dominated by broadband

Fig. 2.13. P o w e r s p e c t r u m f o r t h e M o n d a l e t i m e series ( s p e c t r u m is es-


sentially featureless)
50 Chaos Theory in the Social Sciences

noise. As M o o n (1987) points out, spectral analysis may aid in identifying


chaos i n an experimental time series, but the technique is not conclusive i f the
system has "many hidden degrees o f freedom o f which the observer is un­
aware" ( M o o n 1987, 45).
The Lorenz, Rossler, and chaotic logistic equations have decidedly dif­
ferent spectra. The analysis o f the Mondale series is based on the same
principles, and the result reveals a spectrum that favors the chaotic logistic
spectrum. Evidently there is no fundamental frequency for the Mondale se­
ries, which w o u l d appear as a sharp peak clearly extending above the back­
ground. Admittedly, the level o f noise in this time series is high and that is
probably one source o f the problem in identification o f cycles, i f cycles are
present. Some chaotic time series contain cycles, but in many cases (Lorenz
equations, Rossler equations) these cycles are aperiodic and this fact mani­
fests itself as a broadening o f the spectrum. Spectral analysis o f the Mondale
series does not eliminate a deterministic component in the dynamics.

Conclusions

Spectral analysis serves to illuminate hidden periodicities in data, when they


are present. M a n y chaotic series such as the Lorenz system are aperiodic
while some, such as the logistic map, contain many, indeed an infinite number
of, cycles. The presence o f clear periodic behavior in a time series is easily
discerned. This was demonstrated w i t h the Rossler equations. The Lorenz
equations contain aperiodic cycles and the spectrum fails to sift these out o f
the series, despite its great length. The logistic map, w i t h driving parameter at
3.6, contains infinitely many cycles, o f all lengths ( L i and Yorke 1975).
Spectral analysis cannot discern any cycle whose length exceeds the length o f
the data set (500 elements i n the logistic equation). This may be the problem
for the time series constructed from the CPS survey data. I f the series has a
cycle (or many) whose length exceeds the length o f the observed time series,
we cannot discover its presence. I t seems intuitively reasonable to assume that
the aggregated survey data contain no cycles longer than the entire nomination
period. However, chaotic time series are k n o w n to have aperiodic cycles o f
many lengths, so it seems reasonable to assume that i f they are present in
the candidate time series they should have been observed, though the large
stochastic component here is problematic.
This analysis has highlighted one problem present in many frequency
analyses o f chaotic time series: the results are ambiguous. The ambiguity
arises on one hand from the difficulty in distinguishing between the stochastic
component and the periodic signal when the signal-to-noise ratio is high,
while on the other hand chaotic signals closely resemble pure noise. In some
cases spectral analysis can discern between the two components and in others
Probing the Underlying Structure in Dynamical Systems 51

it cannot. W h i l e it seems that the Mondale (and the Jackson) series contains
identifiable peaks corresponding to several frequencies, the series also con­
tains considerable noise. Noise introduces problems in the resolution o f peri­
odicity, especially when the signal-to-noise ratio is high. Glass and Mackey
(1988) and M o o n (1989) clearly state that spectral analysis is only one tool to
be used in the analysis o f time series that are thought to be chaotic. Here we
have an example o f w h y no particular technique is to be relied upon exclu­
sively in the analysis o f suspected chaotic time series. A set o f tests need to be
applied to the time series. Spectral analysis is a good place to start, but
measures o f the dimension o f the series (the correlation integral) as well as o f
the rate o f divergence o f nearby points on the attractor (the Lyapunov expo­
nent) are required for empirical identification o f chaotic motion.

N O T E S

I would like to thank Thad Brown and Jim Kuklinski for their cogent critiques of
earlier versions of this paper, and Gottfried Mayer-Kress for suggestions regarding the
spectra of a variety of time series.
1. All time series used here are discrete because they are produced by numerical
integration or recursion. Numerical integrators rely on small step sizes, which are a
discrete approximation to an integral solution.
2. A variety of smoothing (or weighting) techniques are commonly used with
spectral analysis to accentuate periodicity (see Chatfield 1991; Wei 1990 for ex­
amples). None of these techniques are used here because smoothing a chaotic time
series destroys the basic property of interest! None of the basic research in the natural
sciences uses smoothing in spectral analysis because the interest lies not in modeling
the process but rather in examining the extant time series for periodicity in its natural
state. Hence smoothing is not used here, nor should it be used in general.
CHAPTER 3

Measuring Chaos Using t h e Lyapunov Exponent

Thad A. Brown

A t some point in time, all scientific things must be measured and calibrated—
even chaos. The Lyapunov characteristic exponent, A, is the clearest measure
to prove the existence and to quantify chaos in a dynamical system or time
series. Chaos exists when A is positive, and this indicates the system under
investigation is sensitive to initial conditions. A dictionary would suggest
chaos is a state o f complete confusion, or one lacking any organization. I n the
field o f nonlinear dynamics, however, the term seems to imply specific prop­
erties o f turbulence i n a nonlinear system. But what is chaos? H o w can we test
for it i n an empirical manner, and precisely, thereby separating it from other
complicated yet fully predictable phenomena? As it turns out, the Lyapunov
exponent, A, is a powerful measurement tool for separating chaotic behavior
from the stable and predictable. Before turning to a discussion o f this mea­
sure, let us review some o f the main features o f chaos.

Chaos

Chaos exists when the long-term prediction o f a system is impossible. Chaos


occurs when the uncertainty o f a system's initial state grows exponentially
fast. I n time-series language, chaos fits the criterion that the autocorrelation
function o f the time signal goes to zero in finite time. Because trajectories are
unstable, errors o f estimation o f initial conditions or parameters, however
small, can later accumulate into substantial errors. Thus, forecasts o f future
behavior based on the past become problematic as current memory o f the past
fails. This has both positive and negative effects on information about the
underlying system.
Nonlinearity is a necessary, but not sufficient, condition for chaotic mo­
tion. The observed chaotic behavior is due neither to external noise nor to an
infinite number o f degrees o f freedom. The source o f irregularity is the non­
linear system's property o f separating initially close trajectories exponentially
fast. Forecasting long-term behavior o f chaotic systems may be more akin to

53
54 Chaos Theory in the Social Sciences

art than to statistics (Peitgen, J ü r g e n s , and Saupe 1992). Does this imply that
it is impossible to predict the long-time behavior o f a political or economic
system (or both)? Such systems are potentially chaotic because their initial
conditions can be fixed only w i t h finite accuracy. Error may explode.
Yet the long-time unpredictability exists only at the individual level (or,
in the language o f dynamics, at the level o f individual trajectories). A t the
level o f statistical properties o f the time evolution (averaged over different
trajectories, say as they evolve from different nearby initial conditions), very
definite predictions are possible. Said differently, when or where chaos exists
can only be k n o w n by statistical prediction. Trajectories w i l l eventually move
only on a small submanifold (chaotic or strange attractor to be described later)
of the entire state space, w i t h predictable visitation frequencies o f the different
parts o f the attractor. Is it possible, then, to estimate the time up to a short-
term prognosis o f an individual trajectory? Yes. It is the coexistence o f pre­
dictable and unpredictable that leads to the phrase order in chaos in the field
of nonlinear dynamic systems.
What are the attributes o f deterministic chaos? Systems o f nonlinear
equations w i t h time dependencies are the easiest place to find chaos. They
range from simple one-dimensional models such as the quadratic equation
x —* ax( 1 — x) to more complex three- and four-dimensional equations (Hale
and Kocak 1991). Chaotic systems exhibit perverse or complex motion i n that
the time series never settles down to simple cycles or equilibrium. Orbits are
inconsistent, wandering, and erratic, m i x i n g the properties o f the time evolu­
tion. The key to chaotic behavior is that the equation or the system o f equa­
tions o f motion exhibits a sensitive dependence on initial conditions. Often a
series o f states can be seen when the value o f a parameter o f the model is
changed and the system moves from one state to another.
In time-series data where the exact equation o f motion can only be
presumed, chaos is harder to find. Even more challenging is to infer the
microscopic process that underlies a time series where chaos has been de­
tected. To know the equations o f motion in politics, economics, and social
dynamics, we require just such discoveries o f microscopic laws. We w i l l turn
to this discussion and point to some interesting approaches at the end o f this
chapter.

An Attractor

The motion o f a dynamical system is represented by an orbit around an


attractor. A n attractor is a closed invariant set that essentially captures all
orbits starting in its domain o f attraction.'As a region o f phase space, an
attractor exerts an appeal to a system, a point, or a subset o f points toward
which any dynamical path w i l l converge. When the behavioral history o f a
system is examined, the nature o f the attractor is the core o f inquiry. When a
Measuring Chaos Using the Lyapunov Exponent 55

rat, voter, or business firm is dropped into a completely different environment,


a period o f erratic behavior is expected as adjustment takes place. This adjust­
ment is transient behavior that gives way to observed behavior in some equi­
librium state.
A n equilibrium state is called an attractor. I f the attractor is a single
point, asymptotically it represents stability because all trajectories that begin
in some neighborhood o f phase space approach the attractor as time passes.
O f course, there are other attractors besides equilibria. L i m i t cycles may
exist. L i m i t cycles occur when the orbit is drawn toward a cyclic path, rather
than toward a point fixed in space. Attractors can also be complex in that they
may be quasi-periodic or strange. A n example o f a quasi-periodic attractor is
an orbit on a torus, a doughnut generated by the cosines o f a pair o f incom­
mensurate frequencies.

Chaos Characteristics

Chaos is measured by the characteristics o f its dynamics: fractal structure o f


the attractor; metric K entropy; and the Lyapunov characteristic exponent, A,.
We w i l l briefly explore each.

Chaotic A t t r a c t o r

A chaotic attractor is characterized by exponential divergence away from any


point w i t h i n the attractor (Ruelle and Takens 1971). The exponential diverg­
ing movement on a compact attractor means that a trajectory is constantly
curling back on the attractor. The strange attractor causes the " f l o w " to con­
tract the volume in some directions and stretch it in other directions. Remain­
ing w i t h i n bounded dominion, the volume and backfolding mixing process
produces a chaotic motion o f the trajectory at the strange attractor. I t is this
folding motion which produces the fractal structure attributed to chaotic mo­
tion.
The properties o f a strange attractor are: (1) the attractor set is such that it
cannot be decomposed into smaller closed invariant sets (Ruelle 1980, 1989;
Guckenheimer and Holmes 1983; StaniSic 1988). (2) There is a noticeable
regularity i n the physical structure or dimension o f the attractor. A point, o f
course, has no dimension; a line, one dimension; a surface, t w o dimensions;
and so o n . B y contrast, a strange attractor has a noninteger dimension. A
strange attractor usually violates Euclidean geometry. After a long series o f
iterations, a fractal geometry stems from long-term stretching and folding o f
trajectories on a manifold. Trajectories on the attractor end up like a Cantor
set (the mathematical equivalent o f a croissant).
The fractal dimension has a fantastic interpretation. It delimits a m i n i ­
mum number o f degrees o f freedom necessary to describe chaotic behavior.
56 Chaos Theory in the Social Sciences

The lower the dimensionality, the fewer independent variables are needed to
describe the chaotic behavior. O f course, i n a chaotic system the dimension o f
the attractor is less than the dimension o f the space w i t h i n which it is embed­
ded (see Feder 1988).
A n d , finally, (3) the attractor is sensitive to initial conditions. Sensitivity
refers to the amplification o f any arbitrarily small interval o f values by itera­
tion. Sensitivity to initial conditions is the nub o f chaos theory, and indeed
may be the point where chaos theory touches science. When nonlinear deter­
ministic models—which are absolutely correct—are sensitive to initial condi­
tions, they cannot be used to predict very far into the future. Lorenz (1963), in
his famous experiments on weather patterns, found that the higher iterations
appeared statistically independent o f earlier values. The smallest error in the
, h
n decimal place becomes magnified and amplified under the numerical left-
shift action o f the iterative function.
We assume that the volume o f an attractor is much smaller than the
volume o f phase space. In a dissipative system the irregular stretching in some
directions and contraction in other directions produces a final motion (once
transients are over) that may be unstable within the attractor. Points that are
arbitrarily near at the initial time step become exponentially separated at the
attractor for sufficiently long times. This leads to positive Kolmogorov en­
tropy, the exponential separation that occurs in the direction o f stretching.
A Fourier analysis o f motion on a strange attractor reveals a continuous
power spectrum. Under conditions o f a linear model, a typical interpretation
w o u l d be that there are an infinite number o f Fourier modes or oscillators.
Under the conditions o f linear theory, such a reasoning has to take place i n an
infinite-dimension phase space. A n interpretation can be either that the system
is searching an infinite number o f dimensions in phase space or that the
system is evolving in a nonlinear fashion on a finite dimensional attractor.
Recognition o f nonlinearities has provoked a fundamental alteration from
thinking about Fourier modes to thinking about the information dimension.
The tool used to measure this information dimension is ergodic theory
(Billingsley 1965).

K Entropy

A second measure o f chaos is K entropy. K entropy is also referred to as the


Kolmogorov-Sinai entropy (Farmer 1982; Schuster 1988). Entropy measures
the rate o f information production and the growth o f uncertainty. I n a system
increasing in entropy, the number o f possible system states that evolve from
some initial distribution over time also increases. In a chaotic system, infor­
mation about the system decreases over time when measured against the
initial state. I f subsequent measurements are not made, indeed, the observer
Measuring Chaos Using the Lyapunov Exponent 57

knows less and less over time. W i t h a second set o f measures, however, we
end up w i t h far more information about the system, even about the initial
state, than w o u l d have been possible at the beginning (Shaw 1981).
Consider that conditions where entropy increases are similar to a party
system where voters were limited to two or three parties but suddenly are
allowed to vote for twice as many. The disorder in the system increases
because voters are no longer limited or confined to their prior political space.
The increase i n disorder is coupled with an increase in our ignorance about the
state o f the system. Before the sudden doubling o f the number o f parties we
knew more about the position o f voters.
Precisely then entropy, S, is proportional to —S, P log P where P is the
t h t

set o f probabilities o f finding the system in state {/} and measures the informa­
tion needed to locate the system in a specific state {i}. Thus S captures our
ignorance about the system.

Lyapunov Characteristic Exponents

The Lyapunov characteristic exponents, A,, o f dynamical systems measure the


average rate by which the distance between close points becomes stretched or
compressed after one iteration. Lyapunovs are generalized eigenvalues over
an entire attractor i n that they give the average rate o f contraction or expansion
o f trajectories on an attractor (Wiggins 1990). The A, allows further resolution
by allowing us to see the (orthogonal) direction o f the growth or compression.
For an attractor to be strange, it must have at least one positive Lyapunov
exponent.
The A, is closely related to the other measures o f chaos. Ruelle (1983,
1989) suggests that K entropy equals the sum o f the positive A,, implying that
positive entropy exists in the presence o f chaos. The A, is linked to the
information lost and gained during chaotic episodes (Ruelle 1980; Shaw
1981, respectively), and it is closely linked to the amount o f information
available for prediction. The A, was "conjectured" to be related to the fractal
character o f the attractor (Kaplan and Yorke 1979), and indeed this was
proven true ( w i t h qualifications) for typical attractors (Russell, Hanson, and
Ott 1980). The A, itself has a fractal, noncontinuous dimension, and that
fractal quality is linked to the information available about the system (Kaplan
and Yorke 1979; Young 1982; Ruelle 1989).
In a one-dimensional system there are three alternative variations o f the
Lyapunov.

A, < 0 = ^ orbit is stable and periodic


A, = 0 ^ orbit is marginally stable
A | > 0 =^ orbit is chaotic
58 Chaos Theory in the Social Sciences

A Lyapunov-positive system is by definition chaotic. The positive value o f the


Lyapunov exponents indicates orbital divergence and a time scale on which
state prediction is impossible. Negative Lyapunov exponents set the time
scale on which transients or perturbations o f the system's state w i l l decay to a
periodic orbit. A zero Lyapunov indicates a marginally or neutrally stable
orbit. This often occurs near a point o f bifurcation.
In a more complex phase space, the Lyapunov spectrum is used to
classify and describe the dynamics. The sign o f the Lyapunov is reported per
h
k' dimension, again with the negative sign indicating convergence and a
positive exponential sign indicating divergence. Hence, for a system denned
in two dimensions, there are t w o Lyapunov exponents for the attractors o f the
system, A, and A . The transformation from the Lyapunov exponent, A, to
2

K
the Lyapunov number, e , can be used to give a more precise interpretation.
A1
The first Lyapunov number, e , indicates the error amplification; the second
n
number, e , illustrates the degree o f contraction (—) or expansion ( + ) . The
A 1 + A 2
sum o f the first and second exponents, e , can be interpreted as a measure
of how much the area is either reduced (—, —) or expanded ( + , + ) . For a
system evolving on a fixed point where k = 3, the signs o f the Lyapunov are
(—, —, — ) . I n such a system there is a continuous convergence on a fixed
value. For a l i m i t cycle the Lyapunov characteristic exponents are (0, —, — ) ,
where convergence occurs on two coordinates, w i t h a marginal limit on the
third. A two-torus attractor is indicated by (0, 0, —) and reflects instances
where t w o dimensions are marginally stable and one converges. Strange
attractors exist i n instances o f at least one positive Lyapunov exponent ( + , 0,
—) and reflect instances o f simultaneous convergence and divergence. A
purely chaotic state ( + , + , + ) represents utter divergence. Purely chaotic
behavior is illustrated by a cantor-like set in a noninteger, fractal dimension.
Intuitively, we can define a purely chaotic state where the three Lyapunov
exponents are all positive. The sum o f the three Lyapunov numbers,
eAi+A2+A3; approximates the maximal average by which the volume changes
per time step.

Since it is increasingly difficult to see the evolution o f a chaotic system as


the divergence o f the trajectories on the attractor becomes more rapid, the
Lyapunov provides the most useful indicator o f chaos. It should be mentioned
that the usual methods o f statistical investigation w i l l not work when chaos
exists. Spectral analysis and traditional time-series model fitting w i l l not work
to reveal the underlying determinism. Running multiple experiments over a
long time and comparing results at the smallest level o f observation run
straight into the consequences o f the "butterfly effect." The slightest measure­
ment error w i l l be magnified, invalidating any comparison or goodness o f fit
measures based on the expected versus observed observations.
Measuring Chaos Using the Lyapunov Exponent 59

Some Mathematical Underpinnings for A


Schuster (1988) nicely summarized some o f the core mathematical properties
o f the Lyapunov exponent. The Lyapunov exponent is a local quantity aver­
2
aged over an entire attractor. It measures exponential separation from some
initial value, x , due to amplified error, e. The following illustration shows i n
0

one dimension these properties:

N x
e ee ^ o>
N iterations N N x
xx+
0 0 e f (x )f ( o
0 + £)'

N x N N
where ee * o> = \f (x 0 + e) —f (x )\. 0 The limits o f the previous expression
can be formally expressed as

N
+ e) -f ix )
0

N—t-x e - » 0 "

which reduces to

. . . lim 1 N
df (x ) 0

dx 0

Schuster showed that is the average distance between closely adjacent


points after one iteration.
The Lyapunov also provides an unambiguous measure o f the presence o f
chaos and gives us a measure o f the average information gain or loss per
iteration. The Kaplan-Yorke conjecture proposed that it was possible to pre­
dict the dimension o f a strange attractor based upon a transformation o f the
Lyapunov exponents o f the selfsame dynamical system (Kaplan and Yorke
1979; Russell, Hanson, and Ott 1980; Frederickson et a l . , 1983). In essence
the units o f A, are bits o f information per iteration from the base o f the initial
measurement. B y an application o f the chain rule applied to the following
equation, we get

Jx f 2 ( x )
U =j f f x
[ { x ) ]
L /'i/<w'c*o)
=

= / ( * i ) / ' ( * o ) where x = x f(x ).


0

Therefore, to evaluate A, we calculate


60 Chaos Theory in the Social Sciences

A. = l i m — los lim - log n / ' U i )

lim ^ E log fix,)


A/-^» 1=0

B y Oseledec's multiplicative ergodic theorem (1968), this limit exists for


almost all x . The average amount o f stretching depends upon the initial error
0

as w e l l as on the initial value o f the system. Oseledec's theorem has been used
in nearly all explications o f Lyapunovs to refine the definition o f their mathe­
matical underpinning (Eckmann and Ruelle 1985).
The implications are as follows: (1) the Lyapunovs are invariant o f the
dynamics; (2) because they are independent o f the position o f the trajectory on
the orbit, the Lyapunov can be used to classify the attractor; (3) they also
define the limits o f predictability for any model. This implies a key interpreta­
tion o f the Lyapunov, since the largest Lyapunov governs the size o f any
perturbation to the orbit around the attractor. When the system is Lyapunov-
positive, the perturbation grows and predictability is lost. We can precisely
k

say now that 2 A, is the Kolmogorov-Sinai entropy, which dictates the rate
i>0

o f loss o f predictability. Hence the computation o f Lyapunovs becomes cru­


cial since the system w i t h them defines its o w n predictability (Eckmann and
Ruelle 1985; Abarbanel 1992).
A map can illustrate how the change in the level o f information may
occur after one iteration. D i v i d i n g the closed interval [ 0 , 1] on x into n equal
intervals, x can occur w i t h a probability o f \ln. B y k n o w i n g w h i c h interval
0

contains x , we gain the following information:


0

/ 0 = - E ^ log ^ = log 2 2 n.
i=l

The linear map, / ( x ) , changes the length o f an interval by a known factor, a


| / ' ( 0 ) j . The loss o f precision leads to a loss o f information:

1 1
i
-\og a
2
- log - 2

n n
Measuring Chaos Using the Lyapunov Exponent 61

Hence,

A/= - l i m i 2 log 2 \f'(x,)\.

The value o f | A / | is proportional to the Lyapunov, in that

\(x )
0 = log 2 · |A/|.

The larger the exponent, the greater the information loss. The gain or loss o f
information reflects the amount o f uncertainty in the dynamical process. W i t h
a loss o f information, the amount o f predictability decreases and the certainty
in the future is diminished. Thus the Lyapunov exponent allows both a calcu­
lation o f a chaotic signal and an interpretation.

Calculating A
The calculation o f the Lyapunov exponents depends upon measuring the rate
by which close or nearby trajectories split in phase space after small changes
in initial conditions. When the equations o f motion are k n o w n , as i n a
computer simulation, estimating the Lyapunov is not difficult. When data are
used, difficulties arise. As a rule, chaotic systems contain at least one positive
Lyapunov exponent. The trick is to calculate the Lyapunov w i t h i n the small­
est possible D-dimensional hypersphere that contains the attractor w i t h i n a
specific number o f time steps. There are three control parameters: the embed­
ded dimension, the number o f time steps over which pairs o f points are
evaluated, and the precision o f the estimate. A fc-dimensional volume segment
grows by the average factor o f A, + A + . . . + \ at each time step. A g a i n ,
2 m

a A > 0 indicates that the attractor is chaotic and that a very small error or
perturbation grows in the short run exponentially fast.
Two algorithms have been used to calculate Lyapunov characteristic
exponents: Wolf's algorithm (1985, 1986) and the Eckmann-Ruelle algorithm
(1985). The W o l f algorithm tracks a pair o f arbitrarily close points over a
trajectory to estimate the accumulated error per time step. The points are
separated i n time by at least one orbit on the attractor. The trajectory is defined
3
by the fiducial and test trajectories. They are tracked for a fixed time period
or until the distance between the two components o f the trajectory exceeds
some specific value. In sequence, another test point near the fiducial trajectory
is selected and estimation proceeds. The end product is that the stretching and
squeezing are averaged.
Figure 3.1 shows a representation o f the Wolf computation o f a
62 Chaos Theory in the Social Sciences

*o('o>

y(t )
0 y(t )
2

Fig. 3 . 1 . S c h e m a t i c r e p r e s e n t a t i o n of t h e W o l f a l g o r i t h m t o c o m p u t e A,.
( F r o m V a s t a n o a n d K o s t e l i c h 1986. R e p r i n t e d w i t h p e r m i s s i o n . )

Lyapunov as presented by Vastano and Kostelich (1986, 102). The initial data
point, y(t ), and its nearest neighbor, z ( / ) , are L units apart. Over A f , a
0 0 0 0

series o f time steps from t to t , the two points y and z evolve until their
0 k

distance, L ' , is greater than some arbitrarily small e. The y value at t be­
0 k

4
comes y ( f , ) and a new nearest neighbor, z , ( / , ) , is selected. This procedure
continues until the fiducial trajectory y reaches the end o f the time series. The
replacement o f the old point by its substitute point and the replacement o f the
error direction by a new directional vector constitutes a renormalization o f
errors along the trajectory.
The largest estimated Lyapunov exponent o f the attractor is then

Â, =
/VAf

where M is the number o f replacement steps (where some arbitrarily small e


was exceeded) and N is the total number o f time steps that the fiducial
trajectory y progressed. W o l f et al. (1985) provide the code for two pro­
grams: one for estimating the Lyapunov spectrum for systems o f differential
equations and a second for estimating the Lyapunov from a time series. The
+ +
code is relatively easy to implement in Fortran or to convert to Turbo C .
The I B M PC computer packages developed by Schaffer (1988) and Sprott and
Rowlands (1992) have implemented the Lyapunov exponent by means o f the
W o l f algorithm. Figure 3.2 illustrates the Lyapunov exponent using the Wolf
algorithm for an iterative logistic map. W i t h i n the areas o f the chaotic region,
the Lyapunovs are positive. Note that the boundary o f the Lyapunov reflects
the erratic nature o f chaos w i t h i n the logistic map. The geometric properties o f
the Lyapunov exponent are known to be chaotic.
There are several folk wisdoms in estimating the Lyapunov exponent on
data using the W o l f algorithm. Schaffer et al. (1988) suggest standardizing the
Measuring Chaos Using the Lyapunov Exponent 63

Fig. 3.2. I t e r a t i o n s of t h e l o g i s t i c m a p (a) a n d a c c o m p a n y i n g L y a p u n o v


e x p o n e n t s (b). (From S c h u s t e r 1988. R e p r i n t e d w i t h p e r m i s s i o n o f t h e
author.)

data into interval [ 0 , 1]. Such transformations allow for comparisons between
different time series and permit a more ready interpretation o f the length
scales. Second, Fraser and Swinney (1986) suggest verifying that the esti­
mated Lyapunov exponent is stable over a range o f parameter choices. I n ­
creasing the m a x i m u m admissible length scale between fiducial and test tra­
jectories eventually induces the value o f the Lyapunov exponent to drop. W o l f
et a l . (1985) suggest that the m a x i m u m length scales should not exceed 10
percent o f the extent o f the attractor.
The Eckmann-Ruelle algorithm offers some marginal gain over the W o l f
approach and is considered by some the "current state o f the art" (Peitgen,
J ü r g e n s , and Saupe 1992). I t uses a least-squares approximation o f the deriva­
tive matrices, an approach that was independently proposed by Sano and
Sawada (1985). Figure 3.3 illustrates the Eckmann-Ruelle algorithm.
64 Chaos Theory in the Social Sciences

Fig. 3.3. S c h e m a t i c r e p r e s e n t a t i o n o f t h e Eckmann-Ruelle m e t h o d .


( F r o m V a s t a n o a n d K o s t e l i c h 1986. R e p r i n t e d w i t h permission.)

The Eckmann-Ruelle method calculates all positive Lyapunov exponents


of the attractor. The fiducial trajectory, y , begins with y(t) and k nearest
neighbors, z,(/) | / = 1 — k, all o f which are w i t h i n e o f y{t). Each point
evolves through a specific period o f time, At, such that z,-(f) and z,.(j + At) are
tracked along w i t h y(t) and y(t + At). A least-squares estimate is possible o f
the spatial derivative Df (y(t)) o f the series using y(t) and the k points. A t each
time step, all points w i t h i n e o f y(t) are used to calculate the best possible fit
5
until all the data are used.
The Eckmann-Ruelle algorithm provides a useful way to calculate the
entire range o f Lyapunov exponent values, while recovering more than one
positive Lyapunov exponent w i t h the W o l f algorithm is substantially more
complicated.

Conclusion

It might be that chaos at some time in the future w i l l be considered a sideshow


of nonlinear dynamics. Issues that allow full investigation o f self-organization
may simply assume that chaotic states and the nature o f prediction, and indeed
theoretical understanding, must take chaos into consideration.
The ideas behind chaos may also become a common heuristic device, a
way o f understanding complexity, criticality, and nuance largely put aside by
the earlier ideas underlying the behavioralism in the social sciences. There is
an intuitive appeal o f nonlinear dynamical ideas in certain areas o f social and
psychological sciences, where difficulty in measurement and systematic ob­
servations often yield qualitative judgment. As a colleague who studies East­
ern Europe suggested, we have to take the countries and people as they are,
w i t h all their warts (complexity) and goofiness (unreliability).
Chaos may, however, find a new home and more precise use and mea­
surement in more positive, prescriptive areas o f social inquiry. A t the end o f a
Measuring Chaos Using the Lyapunov Exponent 65

draft manuscript on infinitely repeated games o f incomplete information,


McKelvey and Palfrey suggest: " . . . the resulting solutions have properties
that are reminiscent o f chaotic dynamical systems, and one wonders whether
such a solution describes in either a normative or positive fashion the behavior
individuals w o u l d or should adapt in such games" (McKelvey and Palfrey
1992). Hubler and Pines (1993) investigate in game-theoretic logic how
players attempt to model, control, and predict future states within chaotic
environments. Indeed, they even suggest that competition between two
players can lead to a chaotic state. In each instance, the ability o f a player to
recognize the presence o f chaos presents a strategic advantage in adapting
behavior to that chaos. In each instance, it w i l l be critical to identify the
degree o f chaos, distinguishing between the "edge o f chaos" and full-blown
chaos.
Dynamical nonlinear systems are thought to predict different things than
linear systems. Included are a set o f invariant measures such as the Lyapunov
exponents or attractor dimensions. Determining their precise interpretation in
the social sciences w i l l provide some interesting dissertations. As M a y (1992,
452) j o k e d : "chaos at last reconciled the Calvinist's God and his foreordained
w o r l d w i t h the illusion o f free-will that we enjoy; sensitivity to initial condi­
tions makes our w o r l d and fate appear unpredictable, even though it is deter­
m i n e d . " I n low-dimensional political, social, and economic Lyapunov-
positive environments, where does this leave choice theory, free w i l l , and
individual responsibility?

NOTES

Figures 3.1 and 3.3 are reprinted from "Comparison of Algorithms for Determining
Lyapunov Exponents from Experimental Data" by J . A . Vastano and E . J. Kostelich, in
Dimensions and Entropies in Chaotic Systems, ed. G . Mayer-Kress (Berlin: Springer¬
Verlag, 1986), by permission of the authors, Elsevier Science, and Springer-Verlag.

1. Formally, a closed invariant set A C R" is an attracting set if there is some


neighborhood U(A) such that <t>(x,t) e (7 for; > Oand <S>(x,t)-^>A a s f —» °° for every x
G U -Ut 0 <P(x,t) is the domain of attraction of A.
2. Some research work concerns local Lyapunov exponents that specifically aim at
short-term forecasts of orbital instabilities (Abarbanel 1992; MaCaffrey et al. 1992;
Nychka et al. 1992; Wolff 1992). These approaches attempt to describe orbital insta-
bilities for a fixed number of time steps ahead rather than over an infinite number of
time steps ahead such as with global Lyapunov exponents. These approaches are
especially interesting in instances where estimating the Lyapunov exponent is from
experimental data. The difficulty with data is that we may not know the exact equation
of motion nor the correct embedding dimension for the phase space. Hence we will not
know the Jacobian matrices necessary to calculate the global Lyapunov exponents.
66 Chaos Theory in the Social Sciences

Local Lyapunov exponents may hence provide a useful approach to detecting chaos in
some cases.
3. B y definition a fiducial distribution makes probability statements about unknown
parameter values.
4. It should be noted that every replacement data point should be in the direction of
the previous point on the test trajectory. The replacement point, z,(f,), should be
located as near as possible to the value y ( / , ) in the direction z (f,). The search for such
0

a value is restricted to a cone of width © and height rj about the fiducial trajectory y.
The angular-width value of 0 is adjusted as needed. Wolf (1985) has suggested that the
resulting errors in the estimation of the Lyapunov are not sensitive to the choice of 0
or to approximate values of e. Usually, 0 = .349, the value of TT/9. See Vastano and
Kostelich 1986 for an excellent discussion of the implementation of both the Wolf and
Eckmann-Ruelle methods.
5. Using a least-squares estimate of the Jacobian at y(i), given a flow x = f(x), the
variation equations u = Df(x)u are integrated using Benettin's algorithm to compute
the Lyapunov.
CHAPTER 4

The Prediction Test for Nonlinear D e t e r m i n i s m

Ted Jaditz

One o f the more interesting problems o f modern empirical economics is what


one might call the forecasting paradox: Standard linear statistical models o f
economic phenomena invariably fit very well in sample. However, results for
out-of-sample prediction are typically much worse than one might expect
given the in-sample fit. This problem has been known in the profession since
the late 1940s; see, for example, the discussion in Malinvaud (1970, chap. 4)
of some early efforts toward predicting postwar consumption patterns.
Given the difficulties that economists have i n forecasting, many in the
profession are quite receptive to the possibility o f nonlinear determinism in
economic data. Some claim to have found evidence o f chaos. (The most
notable examples are Barnett and Chen [1988] and DeCoster and Mitchell
[1991].) We call this the chaos hypothesis. I f true, the hypothesis could help
to explain the forecasting paradox. I f the data-generating mechanisms are
nonlinear, then i t is not surprising that linear models don't forecast very w e l l .
Second, i f the data generator is chaotic, then long-run point prediction is
hopeless anyway, for reasons that we discuss below.
Direct tests o f the chaos hypothesis are problematic. Certain tests, such
as the calculation o f dimension and entropy numbers, are quite useful at
identifying underlying nonlinear determinism in the experimental sciences.
However, these tests are difficult to apply to economic data, and the results to
date are highly controversial. Economic data sets tend to exhibit certain
features, such as regime shifting and volatility clustering, that make it difficult
to sort out whether the underlying data are nonlinear deterministic or non­
linear stochastic. See Jaditz and Sayers (1993a) for a brief overview.
A great deal o f literature is emerging that suggests that an approach
based on out-of-sample prediction may be able to tell us whether it is likely
that the underlying data-generating mechanism is nonlinear. Even i f the
underlying process is chaotic, certain special nonlinear forecasting techniques
ought to do significantly better at short-term forecasting than the linear tech­
niques typically used by economists. We refer to this hunt for forecastable

67
68 Chaos Theory in the Social Sciences

nonlinear structure as the prediction test. This paper w i l l discuss the basis for
the prediction test and evaluate the prospects for its application to economic
data. We illustrate by applying the test to monthly inflation data. We ask
whether nonlinear methods offer statistically significant improvements i n
forecast performance over the standard linear models typically employed by
economists.
To summarize the discussion, the prediction test for nonlinearity is ex­
tremely promising. There are certain difficulties in applying the test and
interpreting the results: It is not at all obvious what measure o f forecast
accuracy should be used, and certain care must be taken in interpreting the
statistical tests o f forecast improvement. Despite these difficulties, out-of-
sample forecast improvement is perhaps the most persuasive evidence that
one can offer in favor o f the hypothesis o f nonlinearity.
Having said all that, the evidence for deterministic nonlinearity in eco­
nomic data is weak. The methods that work very well at predicting even noisy
chaos do not offer substantially improved forecasts o f economic time series.
This suggests that it is rather unlikely that low-dimensional chaos underlies
economic time series.

Preliminaries: W h a t Is t h e Chaos Hypothesis?

The prediction test works because nonparametric forecasting methods can


exploit the subtle dependence that characterizes certain nonlinear determinis­
tic systems. To explain w h y the prediction test works, we must digress some­
what to discuss more precisely the characteristics o f nonlinear systems, and to
point out the features o f these systems that allow us to form accurate short-run
forecasts. Following is a very brief discussion o f what we mean by a chaotic,
deterministic system. We refer the reader to Brock (1986) for a more technical
introduction from an economic perspective, and Eckmann and Ruelle (1985)
for an introduction from the perspective o f physical systems.

D E F I N I T I O N 1. (Brock 1986) The time series {a,}f has a smoothly


=l

deterministic explanation if a system exists (h, F, X j such that h: 3t" —* fft


0

andF: 3?" —» 3t" are "smooth" (i.e., twice differentiable almost everywhere)
and

X, = F(X,_,) (1)

a, = h(X,) (2)

To interpret, the relationship X, = F(X,_ ) t is an unknown law o f motion


involving the unobserved n vector o f state variables. Given an initial starting
The Prediction Test for Nonlinear Determinism 69

value X i n period 0, the state evolves to F(X ) = X , in period 1, then F ( X , ) =


0 0

X in period 2, and so on. The sequence o f values o f the state variable {X , X , ,


2 0

X , • · • } defines an orbit o f the system. We introduce some additional nota­


2

th 2
tion: let F'{x) be denned as the f iterate o f F. For example, F (x) = F(F(x)),
3
F (x) = F(F(F(x))), and so on. Thus, X, - F'(X ). I f we knew the precise
0

initial state o f the process and the transition function, F, then in principle we
could predict the subsequent evolution o f the process perfectly accurately.
Note that the function h can be interpreted as a "measuring device," which
provides information about the current values o f the state variables. I n this
scheme, we need not observe the state variables directly.
O f particular interest is the case where F() is bounded on a closed
subspace o f and admits an attractor: Heuristically, the attractor is a
parsimonious description o f the long-run behavior o f the system. The long-run
behavior o f the system can be envisaged as orbits moving along the attractor.

D E F I N I T I O N 2. A set A is an attractor for the system (h, F, X ) relative to


0

a set U if the following hold.

(a) For all x E. A, F(x) G A . This means that once an orbit enters the
attractor, all subsequent motion of the system is on the attractor.
(Formally, the set A is invariant with respect to the law of motion F.)
(b) The attractor isn't a union of disjoint sets. (The set A is indecompos-
able.)
(c) The orbits generated by F are dense in A. This means that the
trajectory visits all neighborhoods of the attractor, so that a single
realization of a time series will fully describe the system if that
realization has enough observations. (The set A is topologically
transitive.)
(d) Finally, for the law of motion F, A = n j l F'(U). All orbits that start
7

anywhere in U eventually converge to the attractor.

To illustrate these definitions, consider the tent map, whose properties


were first described by Saki and Tokumaru (1980). The tent map is a simple
nonlinear iterative map from [ 0 , 1] to [ 0 , 1]. The transition function F for the
tent map is given by the equations

i f X , < 0.5
(3)
2(1 - X , ) i f X, > 0.5

The phase curve and a sample time path are given in figure 4 . 1 . The attractor
for the tent map is just the phase curve, the kinked line segment in (X,, X , + 1 )
70 Chaos Theory in the Social Sciences

space from (0, 0) to (0.5, 1), to ( 1 , 0). Starting from any point on the attractor
X„ the subsequent value X t+l — F(X,) w i l l also lie on the attractor. Thus, even
though a typical orbit o f the tent map appears haphazard to the eye (and to
many statistical tests!), it is in fact generated by a very simple rule. I f one
knew the rule, one could predict subsequent observations to a high degree o f
accuracy.
The Prediction Test for Nonlinear Determinism 71

Note that attractors need not be as simple as this example. One way to
describe an attractor is to give a measure o f its complexity. One such measure
is the dimension. For the tent map, the attractor is a one-dimensional curve in
a two-dimensional space. Hence we say that the attractor has dimension one.
Attractors can be highly irregular; many interesting systems exhibit attractors
o f fractional dimension. The dimension is generally less than or equal to the
m i n i m u m number o f coordinates necessary to represent the state space o f the
process. See Eckmann and Ruelle 1985 for further discussion o f this impor­
tant concept.
Let us add a trivial complication. Suppose that our observer function is

at = h(X ) = Xf
t (4)

In other words, we do not observe the state o f the system, but we observe
some function o f the state o f the system. O f course, i f we knew h(-), we could
invert it and thus observe the state o f the system directly. However, we
typically do not know enough about h(-) to be able to invert i t . Can we infer
anything about the true state o f the system from the observed time series?
The answer to this question is given by the Takens Embedding Theorem.
Takens's (1985) theorem states that we can learn about the underlying state
variables, X„ from "embeddings" o f the observed a,'s. Given a sequence o f
observations, {a,}, define the embedding or m-history as:

a? = (a„ a, , +l . . . , a l + m ^) (5)

Each m-history consists o f m consecutive observations, so that the first three


points in a sequence o f two-histories are ( a , , a ) , (a , a ), and ( a , a ) . There
2 2 3 3 4

w i l l be a total o f T - m + 1 o f these, or T - 1 histories in the two-history


case. Takens (1985) shows that the mapping from the m-histories o f a se­
quence a, to the state vectors X, given by

op = {h(X ), t h(F(X,)), h(F*(X ), t . . . , h(F> (6)

is a diffeomorphism ( i . e . , a differentiable map w i t h a differentiable inverse),


provided that the embedding dimension, m, is sufficiently greater than dimen­
sion o f the process, F. ( I f the process is n-dimensional, Takens shows that
m £t 2n + 1 w i l l certainly w o r k . ) This means that knowing the m-histories,
af, is essentially equivalent to knowing the current state o f the system, X,.
To illustrate, suppose we observe 250 iterates o f the tent map filtered
through the observer function h(X,) = Xf. Plot the corresponding t w o -
histories, and you get a picture that looks like figure 4.2: the two histories map
out a one-dimensional curve in the plane. The plot o f the two-histories o f
72 Chaos Theory in the Social Sciences

0.8·-

h(X(M))

0-F 1 1 | -.. - • _ !
0 0.2 0.4 0.6 0.8 1
h(X(t))
Fig. 4.2. E m b e d d i n g o f t h e o b s e r v e d values a ,

the observed values, a,, is in many ways similar to the phase diagram for the
unobserved variables, X It is stretched somewhat due to the effects o f the
r

observer function, h, and perhaps it would be stretched or folded differently


for a different observer function. However, Takens theorem states that many
o f the fundamental topological properties o f the attractor i n the unobserved
state space o f the process are preserved in and can be inferred from embed-
dings o f the observed data. Takens theorem is the most important result in
nonlinear science.
One more feature o f the tent map deserves comment. The tent map is
also chaotic. We w i l l say that the deterministic system (h, F, X ) is chaotic i f 0

it exhibits sensitive dependence on initial conditions.

DEFINITION 3. A dynamical system exhibits sensitive dependence on


initial conditions if for all x on the attractor, and for all e, there exists a 8 and
a point y on the attractor such that for some t, we have || x — y |j < e and
|| F'(x) - F'(y) || > 8.

In plainer words, the time paths o f the process started from t w o nearby states
w i l l diverge eventually. Figure 4.3 illustrates two trajectories o f the tent map,
4
using starting values that differ by 1 0 ~ . Even though the t w o starting values
are quite close, by the eleventh iteration, the paths have diverged enough to
lose all subsequent resemblance.
Sensitive dependence to initial conditions makes long-term prediction o f
their future states virtually impossible. Suppose we only know the current
The Prediction Test for Nonlinear Determinism 73

state o f a chaotic process to some fixed degree o f accuracy, due to the impreci­
sion o f our measuring device. Suppose we observe a noisy chaos.

D E F I N I T I O N 4. A noisy chaos is a dynamical system (h*, F, X ) where F


0

and X follow definition 1 above, with the observer function "polluted" by


0

the addition of observer noise:

a, = h*(X )t = h(X ) + e,,


t E , a white noise random variable. (2')

Thus, even i f we knew the transition function, F, forecasting would be com­


plicated by the imprecision w i t h which we observe the current state o f the
system. Suppose that we knew that the current state, X„ was in an interval
[a, b]. Then we would predict that X , would lie in the interval F([a, b]). I f
+1

the process is chaotic, uncertainty about the current state is rapidly magnified,
and thus our long-term forecasts o f a noisy chaotic process can do no better
than to predict that the process w i l l be somewhere on the attractor. On the
other hand, i f the interval [a, b] is small e n o u g h — i f we know the current state
to a fair degree o f precision—then our prediction o f the near future states may
have a reasonable degree o f accuracy.
The chaos hypothesis is just the suggestion that economic data are gener­
ated by a noisy chaos. I f this is true, then there is a nonlinear function, F, that
is the data generator underlying economic time series. To implement the
prediction test, we use embeddings o f the observed data to try and identify
features o f the attractor for this function, F. I f the attractor is simple enough,
74 Chaos Theory in the Social Sciences

i f the observer function does not obscure the underlying structure, and i f the
amount o f observer noise is small, we may be able to learn enough about the
attractor to aid in short-term forecasts.

A n I n t r o d u c t i o n t o t h e Prediction Test

The underlying nonlinear structure o f a deterministic system suggests that


certain specialized nonlinear forecasting techniques may be able to out­
perform linear techniques at short horizons. Casdagli (1992) describes how
near neighbor algorithms can be used to forecast nonlinear processes. Given
the current m-history, a™, we may be able to predict the next observation,
a t + l , by looking at past m-histories, af, s < t, that are sufficiently close to ap.
I f there is a deterministic function underlying the data, and i f the noise-to-
signal ratio is small, then i f af is sufficiently close to af, a t + l ought to be
close to a s + ] . Casdagli shows that i f we have a simple nonlinear process on a
bounded attractor, and i f we have enough observations to get a complete
picture o f the attractor, this method can generate highly accurate forecasts o f
nonlinear systems.
We start w i t h observations on a time series {x^ t=lT which we partition
into a fitting set 3* and a prediction set 9 \ such as

9= = {x, : m + 1 < t < N } f (7)

3> = {x, : N < t < 7"} f

for some N < T. The aim is to use the information in 5F to predict observa­
f

tions i n 9*.
For a given lag length, m , construct for each observation {x},= \j, m+ an
m-history, xf_ x

x x x x
T-\ ~ ( l-\> l-2> · · · > t-m)- (8)

r e a c r i x n tr
Thus we have a set o f ordered pairs, { ( * , , xT-\)} = +i,T- P° t m i < »e
prediction set, we calculate the distance between x™_ and l V J£
Casdagli suggests using the norm to calculate distances,

|| * || = max,-1 |. (9)

We then select the k nearest pairs, (x , JCJL, ) , to estimate the parameters i n the
s

local regression,

x s = a k + x?_ ß x k + e.
s (10)
The Prediction Test for Nonlinear Determinism 75

The estimated parameters &Q and d are used to calculate the prediction,
k k

x = a + xT-ik-
t k (11)

The prediction is then used to calculate the prediction error, x, — x = e . t tk

Citing the Takens theorem, Casdagli suggests setting the lag length, m,
equal to In + I , where n is the dimension o f the process. When the dimension
of the process is not k n o w n , we can estimate i t , using (to pick one possibility
out o f several) the procedure sketched in Brock and Baek 1991.
We should point out how this procedure compares to the global linear
predictors typically used by economists. For the global linear predictor, we
use every observation in the fitting set to estimate the parameters o f the
regression. I f the underlying process is sufficiently nonlinear, the global linear
predictor w i l l do a poor j o b o f approximating the relationship between the
histories and the futures, and the resulting forecasts w i l l be inaccurate.
The essence o f the prediction test is to compare the forecasts generated
by the linear methodologies preferred by economists to the forecasts gener­
ated by the nonlinear, nearest neighbor algorithms preferred by physicists. I f
the near neighbor algorithm gives more accurate forecasts, then we have
evidence that the underlying process is nonlinear. To evaluate forecasting
performance, we use the Root Mean Square Error ( R M S E ) o f the forecast.
Given a time series, { y y , • • • , y } , and a set o f forecasts for that series,
1 ( 2 r

{ y , , y , • • • , y }, we define the R M S E as
2 T

0.5

RMSE = (12)

Many authors further normalize the R M S E by dividing by the standard devia­


tion o f the data that we are trying to predict: Define the normalized R M S E as

NRMSE = RMSE/cr r

This normalized R M S E attains a value o f 1 i f the method o f prediction is no


more accurate than forecasting the unconditional mean o f the prediction set.
This normalization is quite useful as an aid to the interpretation o f the results.
We w i l l follow custom and report only the normalized root mean square error.
Another example is in order. To evaluate this procedure on a short data
set, we generated 600 observations from the tent map, using a starting value
of 0.1234567891. For our example, we set the fitting set to be the first 500
observations, and the prediction set to be the last 100 observations o f the
76 Chaos Theory in the Social Sciences

process. We first attempted to forecast the process w i t h the global linear


predictor. We estimate an A R ( 3 ) model,

X, = a (13)

We use the estimated parameters from the fitting set to generate predictions for
< >
the data in 3 . The results for the linear model are summarized in table 4.1 and
are quite poor. First, the linear model does not fit the data very w e l l . The
2
regression R is only about 0 . 0 1 ; the regression explains only about one-half
of 1 percent o f the total variation o f the process. The only significant coeffi­
cient is the regression intercept. Out-of-sample forecast performance is dis­
mal. The normalized R M S E for the A R ( 3 ) model is 1.024, which is worse
than the result obtained i f we just use the unconditional mean o f the fitting set
as our forecast.
I f we apply the near neighbor methodology, we obtain far superior re­
sults. In figure 4 . 4 , we plot the normalized R M S E as a function o f the number
of near neighbors used in the regression. For a fitting set o f 500 observations,
we can set the number o f near neighbors used in the regression from one to
500. The m i n i m u m o f the N R M S E plot is the best near neighbor predictor.
The plot is minimized at k = 5: our best prediction o f an element in SP is
obtained when we comb through the fitting set and calculate our prediction
based on the five observations in the fitting set that are most similar to the
history o f the observation that we are trying to predict. A regression calcu­
lated using only those five observations w i l l , on average, explain nearly 98

TABLE 4 . 1 . S u m m a r y o f Regression
Results f o r t h e T e n t M a p

Coefficient Value Test Statistic

Constant 0.560 12.99


-0.059 -1.31
-0.019 -0.42
-0.029 -0.65

2
Regression R 0.012
Fitting Set Mean Forecast NRMSE 1.008
Regression Forecast RMSE 1.024
Number of Observations 500

Note: The test statistic is the standard OLS test statis-


tic for the hypothesis that the corresponding coefficient
is significantly different from zero.
The Prediction Test for Nonlinear Determinism 11

1.2 t

0 100 200 300 400 500


Number of Near Neighbors

Fig. 4.4. N R M S E p l o t f o r t h e near n e i g h b o r f o r e c a s t s o f t h e t e n t m a p

percent o f the variation in the process. The forecasts generated by the near
neighbor methodology are stunningly more accurate than the forecasts gener­
ated by the global linear predictor, and this provides very convincing evidence
of the nonlinearity present in the tent map.
It bears mentioning that the N R M S E plot for the tent map is quite typical
of the shape o f the plot when significant nonlinearity is present. Here, we
observe that the plot is minimized for a relatively small number o f near
neighbors, and the plot is smoothly upward sloping from the m i n i m u m out to
the m a x i m u m number o f near neighbors, where we are essentially replicating
the global linear predictor.
For real economic data, the difference in forecast performance between
the linear and nonlinear predictors is seldom so clear cut. The normalized
R M S E for the nonlinear predictor is typically much closer to the R M S E o f the
linear predictor. We are then left to ask whether the forecast improvement
generated by the nonlinear models is sufficient to justify the inference that the
data generator is nonlinear. It turns out that answering this question is sur­
prisingly difficult.

The Problem o f Forecast Evaluation

There are a number o f conceptual problems that must be addressed before one
can sensibly compare and evaluate forecasts generated by different methods.
The first problem is: how should we measure forecast accuracy? Almost
all authors focus on a mechanical measure o f forecast accuracy, such as Root
78 Chaos Theory in the Social Sciences

Mean Square Error ( R M S E ) or Mean Absolute Deviation ( M A D ) . Each


method o f measuring forecast accuracy has its aficionados, and one can make
a purely technical statistical argument for either. However, in economic con­
texts, it is very often useful to measure forecast accuracy in terms o f an
underlying objective function. For example, i f one is forecasting the move­
ments o f stock prices, the forecast R M S E or M A D are poor yardsticks o f
forecast accuracy. There are two reasons for this. The first reason is that one
isn't interested in stock prices per se, but rather in the profits that one could
make by f o l l o w i n g an optimal trading rule based on the forecasts. Measured
in these terms, an accurate forecast is one that allows the trader to make large
(risk-adjusted) rates o f return. Indeed, the empirical finance literature is full o f
anomalies that are i n some sense "statistically significant," but economically
negligible. It is w e l l k n o w n , for example, that there is an element o f mean
reversion in stock prices: a stock that experiences an especially large loss in
one week can be expected to rebound somewhat the next. What is debatable is
whether the effect is large enough to allow one to make profits in excess o f
transactions costs by trading on this information (see, for example, Conrad et
al. 1990). I f one cannot design a trading rule that uses this information to
generate profits, then the anomaly is not economically significant, no matter
how statistically significant it appears to be.

The second reason is that there are types o f forecast improvements that
these measures are poorly suited to identify. R M S E and M A D are global
measures o f forecast accuracy. It is possible to have a method that forecasts
poorly most o f the time, but very well some o f the time. The method may still
be useful, provided that one can tell when it is going to work and when it w i l l
not. Global measures o f forecast accuracy may not be sensitive enough to
identify these occasional successes.
Having said all o f that, we w i l l continue to focus on the R M S E measure
of forecast performance. We w i l l follow the usual practice o f measuring
forecast performance by normalized R M S E .
The second and much more difficult problem in forecast evaluation is:
how much better does the forecast have to be before the improvement is
"significant"? The term significant improvement is thrown about quite fre­
quently i n papers on forecasting, and almost all o f the time the use o f the term
is not justified. One step toward turning a heuristic procedure into a statistical
test is to utilize some recently developed tools for comparing the accuracy o f
competing forecast methods.
A recent paper by Mizrach (1992) has derived a statistic to test the
hypothesis that one forecast has a smaller mean square error than another. The
test is a refinement o f an approach due to Granger and Newbold (1986) and
Meese and Rogoff (1988). To review the test, let { e , , } be the forecast residuals
from method one, and let {e } be the forecast residuals from method t w o .
2t
The Prediction Test for Nonlinear Determinism 79

Granger and Newbold test whether var(e,,) = a] is significantly less than


var(e ,) = a\ by looking at the orthogonalized residuals u, = e — e „ v, =
2 l t 2

e
i,t +
2,t- The correlation between u, and v, is just o-\ — < J \ . Thus, i f the
E

2
correlation between u, and v, is significantly greater (less) than zero, then o~ is
significantly greater (less) than cr . 2

Mizrach's refinement is to generalize the test to biased, heteroscedastic


forecast residuals, using Newey-West (1987) hardware. Mizrach shows that
given that the t w o sequences are a m i x i n g , the statistic is distributed standard
normal asymptotically

( \ 2 «,v,
R = Vf (14)

V X w(j')
i=k
S (i)
uvuv

where

w(0 = 1 -
k + 1

T - i

2 u ,v u,v,,
i+ i+t i < 0
T + i

k(T) 3
k = k(T), w i t h l i m _ ^ , 0, = i n t ( J " ) + 1,
fl/2
r

for example.
The a - m i x i n g assumption is worth explaining. A sequence o f random
variables is independently distributed i f no single observation or group o f
observations contains any information that can be used to forecast any aspect
o f any other observation in the sequence. A sequence o f random variables is a
m i x i n g i f (loosely speaking) the degree o f dependence between observations
declines d o w n the sequence. That is, while information in observation x, (say)
can be useful in predicting some aspect o f x (say), observation x, contains
t+x

very little information about x, for sufficiently large s. This is essentially an


+s

assumption that anomalies or quirks in the data have only a transitory impact.
80 Chaos Theory in the Social Sciences

To use an economic example, it is essentially the assumption that anything


that happened to inflation in (say) 1950 has very little effect on the inflation
rate o f (say) 1990. It is essentially the assumption that there is little long-run
persistence i n our data.
We now have a procedure to test the hypothesis that method one ( R M S E )
generates more accurate forecasts on a given data set than method two
( M A D ) . The test involves applying the Mizrach statistic to test the null hy­
pothesis that the M S E for forecast method number one is equal to the M S E for
forecast method number t w o . I f the test statistic is sufficiently large in abso­
lute value, we reject the hypothesis that the two forecasts are o f equal accu­
racy, and thus we have a winner i n the forecasting competition.
However, there is a conceptual difficulty w i t h this approach that we
should not pass over without comment. The difficulty has to do with certain
inescapable problems w i t h economic data. To explain: suppose that we have
two competing methods o f forecasting, method one and method t w o . H o w
should one decide whether forecast method one is better than forecast method
two? To correctly answer this question, one would like to have access to
repeated sampling from the data generator, and then look at the distribution o f
one's preferred measure o f forecast performance for each method. The answer
would depend on the expected values o f the measures o f forecast performance
for a typical time series from the data generator.
In principle, one could do studies o f the statistical size o f the test under
the null hypothesis o f no nonlinearity, and the statistical power o f the test at
detecting various types o f nonlinearities. This has not yet been done, and I
would venture to guess that one reason for the lack o f results is that the near
neighbor forecasting methodology is very computationally intensive. Calcu­
lating the near neighbor forecast for a prediction set o f 500 observations and a
fitting set o f 1,500 observations at an embedding dimension o f twenty could
easily take a day on a personal computer based on the Intel 80486 chip
running at 33 M H z . A convincing Monte Carlo study could require 1,000
iterations for several different data generators. However, given enough com­
puting power, we could conduct such a study o f the reliability o f the method
on simulated data sets.
On the other hand, the major implication o f the forecasting paradox is
that we do not yet have reliable models o f economic phenomena. Economic
and financial data exhibit exotic dependencies that are often very difficult to
model. There are, for example, controversies over the appropriate model for
volatility clustering in economic data. A number o f approaches have been
tried, including Auto-Regressive Conditional Heteroscedasticity ( A R C H )
models, unconditional mixture models, and regime-shifting models. N o
single model seems to dominate the others for all applications. Worse, even i f
The Prediction Test for Nonlinear Determinism 81

we confine our attention to a single class o f models, one must still demon­
strate that the parameterization chosen for the simulation is sensible.
These difficulties all occur because, unfortunately, economics is not (for
the most part) an experimental science. Economists (particularly macroeco-
nomists) typically observe only a single, short realization o f their data-
generating process. A time series collected monthly since World War I I (like
the Consumer Price Index) has fewer than 600 observations. Data sets col­
1
lected quarterly (like G N P figures) have less than 200 observations. We
cannot do controlled experiments to assess how the economy would respond
to controlled shocks. This means that there is, in general, no way to determine
whether the economic data we observe are typical or representative o f the
underlying process. I n terms o f the prediction test, this means that the best we
can expect is for the test to tell us whether method one outforecast method t w o
on the given sample. We have no way to determine how the methods w o u l d
compare on alternative samples from the underlying data generator. There­
fore, assessing the reliability o f our measure o f forecast performance on real
data is a deep problem. These are important caveats that must be kept i n mind
when assessing our test results.

A n Illustrative Example: The Inflation Rate

To illustrate the use o f these techniques, we apply them to a data set generated
from monthly observations on the Consumer Price Index for the period Janu­
ary 1947 to February 1993. I take the first differences o f the logs o f the C P I ,
which results i n the monthly inflation rate.
Setting aside the last 20 percent o f the data, we estimate an A R ( 1 2 )
model on the first 442 observations. For the observations * in our filling set,
we obtain estimates o f the parameters in the regression
12

y, = « + 2 Vs-i ft-i + £,· (15)

The lag length, while plausible, is selected for convenience. I n sample fit o f
2
the model is quite good. The R for the regression is 0.56, indicating that the
regression explains a sizable fraction o f the total variation o f the process. This
2
R would correspond to an in-sample R M S E o f 0.44. In the regression, the
standard measures o f parameter variability indicate that many o f the coeffi­
cients are significantly different from zero at the 95 percent confidence level.
Summary statistics for the regression are given in table 4.2.
M o v i n g to out-of-sample forecasting, we use the estimated coefficients
from the A R ( 1 2 ) to forecast the rest o f the series. For observations o f y, in our
prediction set, and their corresponding histories, we calculate forecasts using
82 Chaos Theory in the Social Sciences

TABLE 4.2. S u m m a r y o f Regression


Results f o r t h e I n f l a t i o n Rate Regression

Coefficient Value Test Statistic

Constant 0.00046 2.13


0.36696 7.50
X.-2 0.13934 2.69
*,-3 0.13432 2.59
X,_4 0.04699 0.93
*.-3 -0.05750 -1.19
X.-6 -0.05846 -1.21
X.-1 0.05797 1.20
Xt-S 0.09294 1.91
%l-9 0.12793 2.66
X-l-10 0.10052 2.19
Xt-l 1 -0.07539 -1.65
X*-I2 -0.02392 -0.54

2
Regression R 0.558
Fitting Set Mean Forecast NRMSE 1.011
Regression Forecast RMSE 1.005
Number of Observations 442

Note: The test statistic is the standard OLS test statis-


tic for the hypothesis that the corresponding coefficient
is significantly different from zero.

the O L S estimates (a, b) o f ( a , /3) from equation 15 above to form the


forecasts

12

£
S, = a + + r (16)
i= l

These results are an illustration o f the forecasting paradox. Even though the
model fits reasonably w e l l in-sample, the out-of-sample forecasts are spec­
tacularly inaccurate.The normalized R M S E for the fitting set is 1.005. To
compare, i f we just use the unconditional mean o f the fitting set as our
predictor, our normalized R M S E is 1.011. The regression estimate offers a
reduction i n R M S E o f less than one-half o f 1 percent out-of-sample, over the
unconditional mean predictor. The conclusion is that the global linear A R ( 1 2 )
model is a very poor representation o f the underlying data-generating process.
One then might cast about for a reason w h y performance is so poor. A
useful start w o u l d be to estimate the dimension o f the inflation data, to look
The Prediction Test for Nonlinear Determinism 83

for evidence that there may be a low-dimensional attractor underlying the


data. Table 4.3 gives the estimated dimension o f the process, calculated at
varying lag lengths, following the procedure given in Brock and Baek 1991.
To remove spurious linear dependence in the data, we calculate the estimates
of dimension from the residuals o f the A R model. Also in table 4 . 3 , we have
two other common measures o f nonlinear structure, the BDS statistic (Brock,
Dechert, and Scheinkman 1986) and the Kolmogorov Entropy (calculated
following the method o f Brock and Baek 1991). A l l three measures o f depen­
dence indicate significant evidence o f nonlinear structure in the residuals o f
the linear model.
To serve as a comparison, we repeat the procedure on simulated data
from noisy chaos. We generated 600 observations from the Henon process
(500 from the fitting set and 100 from the prediction set), which is a simple,
low-dimensional chaos (Henon 1976). We "polluted" the Henon data w i t h
normally distributed pseudorandom numbers with a standard deviation o f 50
percent o f the standard deviation o f the Henon data. This corresponds to a
noise-to-signal ratio o f 50 percent, which is fairly high by the standards o f
physical systems. When we estimate a linear model to the Henon data, we
observe results that are somewhat similar to the results we obtained for the
2
CPI data. In-sample fit is acceptable, w i t h a regression R o f about 0.202.
(This corresponds to an in-sample R M S E o f 0.798.) Out-of-sample forecast
performance is worse than the in-sample fit, with an out-of-sample nor-

TABLE 4.3. Nonlinearity Diagnostics f o r t h e Regression Residuals f r o m Inflation Rate Data

E = 0.81 E = 0.9

Embedding Estimated Test Test Test


Dimension Dimension Statistic BDS Entropy Statistic BDS Entropy Statistic

2 1.569 -2.47 5.71 0.584 -2.13 6.05 0.506 -2.15


3 2.312 -2.51 6.66 0.580 -2.11 7.01 0.502 -2.13
4 3.049 -2.47 6.75 0.569 -2.19 7.08 0.490 -2.24
5 3.794 -2.39 6.88 0.547 -2.39 7.22 0.467 -2.50
6 4.553 -2.27 7.25 0.553 -2.19 7.65 0.476 -2.26
7 5.284 -2.19 7.47 0.550 -2.12 7.85 0.480 -2.09
8 5.945 -2.17 7.72 — - 8.00 — —

Note: The dimension of the process is estimated as the elasticity of the correlation integral with respect to the link
scale e. Under the null of independent and identically distributed data ( I I D ) , the expected value of this statistic is equal
to the embedding dimension. The test statistic should be distributed as a standard normal random variable under the null
of I I D . The entropy statistic is a measure of sensitive dependence to initial conditions. Under the null of I I D , the
expected value of the entropy is minus the log of the correlation integral at embedding dimension 1. The test statistic
should be distributed as a standard normal under the null. The BDS test is a portmanteau test of I I D . Again, the test
statistic should be distributed as a standard normal under the null of I I D .
84 Chaos Theory in the Social Sciences

malized R M S E o f 0.913. This is still a bit better than the R M S E o f the


unconditional mean forecast, which is 0.996. When we estimate the B D S ,
dimension, and entropy numbers for the residuals o f the linear model applied
to the Henon data, we observe results that indicate that, i f anything, the
Henon residuals exhibit less nonlinear dependence than the inflation data.
(These results are summarized in tables 4.4 and 4 . 5 . )
When we forecast the Henon data using the near neighbor method, we
obtain substantial forecast improvements. The R M S E plot for the polluted
Henon data is given in figure 4.5. The R M S E plot is minimized at k = 5 1 .
Regressions based on 51 near neighbors yield normalized R M S E o f 0.78.
Note that this R M S E is about 15 percent lower than the R M S E for forecasts
based on the linear model. When we apply Mizrach's test o f forecast perfor­
mance we observe a test statistic o f 2.13, which is significant at the 95 percent
level o f confidence. I n this case, the nonlinear forecast algorithm generates
predictions that are significantly more accurate than the forecasts generated by
the linear model. This is evidence that there is significant nonlinear determin­
ism underlying these data.

TABLE 4.4. S u m m a r y o f Regression


Results f o r t h e N o i s y Henon
Data Regression

Coefficient Value Test Statistic

Constant 0.399 5.63


-0.120 -2.45
X.-2 0.094 1.92
X.-3 -0.287 -5.84
X.-A -0.059 -1.16
x,-s -0.107 -2.12
X.-6 0.001 0.01
X.-1 -0.006 -0.12
x,-, -0.084 -1.66
X.-9 0.095 1.87
Xi -10 -0.012 -0.25
Xl-I 1 0.042 0.85
X/ -12 0.038 0.78

2
Regression R 0.798
Fitting Set Mean Forecast NRMSE 0.996
Regression Forecast RMSE 0.913
Number of Observations 500

Note: The test statistic is the standard OLS test statis-


tic for the hypothesis that the corresponding coefficient
is significantly different from zero.
The Prediction Test for Nonlinear Determinism 85

TABLE 4.5. N o n l i n e a r i t y D i a g n o s t i c s f o r t h e Regression Residuals


f r o m N o i s y H e n o n M a p Data

e = 0.81 £ = 0.9

Embedding Estimated Test Test Test


Dimension Dimension Statistic BDS Entropy Statistic BDS Entropy Statistic

2 1.728 -1.39 4.52 0.796 -0.60 4.69 0.701 -0.68


3 2.627 -1.22 3.17 0.782 -0.84 3.45 0.690 -0.87
4 3.499 -1.19 2.90 0.787 -0.70 3.14 0.706 -0.53
5 4.269 -1.33 2.57 0.771 -0.96 2.60 0.696 -0.71
6 4.984 -1.48 2.54 0.762 -1.06 2.40 0.688 -0.81
7 5.687 -1.57 2.57 0.705 -1.90 2.32 0.650 -1.40
8 6.204 -1.80 3.07 — — 2.57 — —

Note: The dimension of the process is estimated as the elasticity of the correlation integral with respect to the link
scale e. Under the null of independent and identically distributed data (IID). the expected value of this statistic is equal
to the embedding dimension. The test statistic should be distributed as a standard normal random variable under the null
of I I D . The entropy statistic is a measure of sensitive dependence to initial conditions. Under the null of I I D , the
expected value of the entropy is minus the log of the correlation integral at embedding dimension 1. The test statistic
should be distributed as a standard normal under the null. The BDS test is a portmanteau test of I I D . Again, the test
statistic should be distributed as a standard normal under the null of I I D .

Compared to these results, the results for the inflation data are rather
disappointing. The R M S E plot for the inflation data is given in figure 4.6.
The N R M S E is minimized at k = 85, where N R M S E = 0.975. This is
approximately a 2.5 percent improvement over the N R M S E for the AR(12)
model. We check whether this improvement i n forecast performance is signifi-

1.2 j

1.1 ••

1 ••
NRMSE

0.9¬

0.8 •

0.7 ••

0.6 -I 1 1 1 1 1
0 100 200 300 400 500
Number of Near Neighbors
Fig. 4.5. N R M S E p l o t f o r t h e near n e i g h b o r f o r e c a s t s of t h e noisy Henon
data
86 Chaos Theory in the Social Sciences

1.2-*

0.8··

0.7 ••

0.6 -J 1 1 1 1 1

0 100 200 300 400 500


Number of Near Neighbors
Fig. 4.6. N R M S E p l o t f o r t h e near n e i g h b o r f o r e c a s t s o f t h e i n f l a t i o n
data

cant, using the Mizrach test. The test statistic for this comparison is 1.77. The
improvement in forecast performance is not statistically significant by the
usual criteria.
The lack o f forecast improvement is especially troubling when we con­
sider the computational burden o f the near neighbor approach. The AR(12)
model fits the data w i t h 13 parameters, counting the regression intercept. For
a fitting set o f 500 and a prediction set o f 100 observations, the near neighbor
methodology requires on the order o f 500· 100 = 50,000 regressions, and the
calculation o f 750,000 parameters, all to yield an insignificant forecast i m ­
provement.

Discussion

The prediction test is a very enticing approach to testing for nonlinear deter­
minism i n a data set. For data from systems known to be chaotic, near
neighbor techniques offer spectacular improvements in forecast performance,
relative to the performance o f the global linear predictors favored by econo­
mists. The near neighbor methodology is also useful for predicting noisy
chaos. These developments are well documented in Casdagli (1992) and
references therein.
Numerous authors have applied near neighbor algorithms to economic
data sets, w i t h no significant forecasting success in evidence in the literature.
The Prediction Test for Nonlinear Determinism 87

Recent efforts include Diebold and Nason 1990, LeBaron 1992, Meese and
Rose 1990, Mizrach 1992b, and Jaditz and Sayers 1993b. When the method is
applied to economic data, the forecast improvements are generally not statis­
tically significant. Even when there appears to be ample evidence o f nonlinear
dependence i n a data set, nonlinear methods are unable to exploit the depen­
dence to offer improved forecast performance. The question is, why does this
occur?
In our previous discussions, we have touched upon some possible rea­
sons w h y we may be failing to identify nonlinearities. One could question our
measure o f forecast performance or our method for testing whether our
forecast improvement is significant. Other explanations are also possible. For
example, the data set available may be too short to allow us to conclusively
identify whether nonlinear determinism is present. I t would be very difficult to
detect the presence o f a moderately high-dimensional process in samples as
short as the one at hand. High-dimensional chaos may be observationally
equivalent to random behavior i n time series this short.
However, one could just as easily reject the hypothesis that economic
data generators are chaotic, i n favor o f stochastic alternatives. I n their study o f
near neighbor forecasting o f current exchange rates, Diebold and Nason make
two points on the apparent paradox o f dependence and lack o f forecastability.
The first point is that in economic time series, the conditional variance tends
to change over time. The standard tests for nonlinear determinism are known
to be sensitive enough to pick up this dependence. A discussion o f this point is
carried out in Brock, Hsieh, and LeBaron 1991 in the context o f the BDS
statistic. W h i l e volatility clustering is easily detected, dependence in variance
is not useful for the prediction o f the mean. Thus there may be substantial
dependence in the time series that cannot be exploited to improve level predic­
tion.
A second possible explanation for the poor forecast performance is that
tests for nonlinear dependence may be picking up regime shifts, statistical
outliers, or other underlying structural instability. M a n y economic time series
are subject to regime shifts. For example, in our inflation data there is gener­
ally conceded to be a regime shift occurring about 1980, when the Federal
Reserve Board switched from a policy o f pegging nominal interest rates to a
policy o f pegging money supply growth. Occasional regime shifts or low-
probability outliers may be o f limited utility for short-horizon forecasting by
any pure time-series method.
Hence we reach a paradox. Conventional in-sample measures o f non­
linear dependence suggest that economic data exhibit substantial nonlinear
structure. Forecasting algorithms seem to be unable to utilize any o f that
structure to improve forecasting. Hence, claims o f nonlinear determinism in
economic data are still controversial.
88 Chaos Theory in the Social Sciences

The prediction test does, however, offer hope for a possible resolution o f
the dilemma. I f nonlinear methods generate forecast improvements in eco­
nomic time series on the order o f the improvements observed in the forecasts
o f physical systems, the case for nonlinear determinism w i l l be irrefutable.
Such improvement has not yet been demonstrated. I t is certainly worthwhile
to continue the search.

NOTES

1. It is possible to argue that the problem is less severe in financial economics.


Tick-by-tick price and volume of sales data are available for thousands of stocks,
bonds, and related financial instruments. On the other hand, these many time series are
not independent observations from the same data-generating process, nor are they
independent in cross section. Thus, many of the same problems remain.
CHAPTER 5

From Individuals t o Groups: The Aggregation


of V o t e s and Chaotic Dynamics

Diana Richards

One takes the sum of particular wills . . . [and] . . . take[s] away from these
same wills the pluses and minuses that cancel one another . . . [and] . . . the
general will remains as the sum of the differences.
—Jean Jacques Rousseau

The aggregation o f individual preferences into a group choice is one o f the


most significant questions in political science. H o w citizens combine and
weigh their interests and desires toward a societal agreement is the foundation
of democratic theory. However, the individual-group connection is not as
straightforward as early democratic theorists assumed. The process o f reach­
ing a societal agreement between a group o f individuals, each with their o w n
preferences and abilities to act strategically, is no longer subsumed under a
simple additive relation where the group interest is merely the sum o f individ­
ual interests.
It is now well accepted that the individual-to-group connection is capable
of serious pathologies, regardless o f the voting scheme used (e.g., A r r o w
1963). But establishing the possibility o f pathologies, such as identifying
examples o f voting paradoxes, does not imply that the connection between
individual and group is fully understood. We know some conditions under
which democratic social choice falls short o f our ideals, but we do not know
why it fails or what set o f tools is appropriate to address the "black box" o f the
aggregation from individual to groups.
This chapter demonstrates that the process o f aggregation from individ­
ual to group is in the realm o f chaotic dynamics. In particular, the non-
equilibrium cases o f social choice, where group dynamics fail to reach a
stable agreement point, exhibit chaos. This observation accounts for the wide
diversity o f potential group outcomes and the nonadditive aggregation func­
tion. These findings also suggest that the tools o f chaotic dynamics may shed
light on w h y the instability occurs, where social choice w i l l be most unstable,
and how the instability can be reduced. These are future topics. The first task
is to establish that the micro-macro connection manifested in social choices
exhibits chaos in its nonequilibrium cases. The chapter is organized as f o l -

89
90 Chaos Theory in the Social Sciences

lows. The first section examines the social choice debate and its importance
for democratic theory. The next section introduces chaotic dynamics using the
example o f the logistic function to illustrate the intuition behind examining a
chaotic process using symbolic dynamics, in particular the iterated inverse
image approach. The third section reviews the multidimensional spatial voting
model and the fourth applies the tools illustrated i n the second section to
the case o f multidimensional spatial voting. To suggest the robustness o f
this connection, the fifth section briefly outlines existing results on addi­
tional cases o f individual-group aggregations, including simple voting and
equilibrium-adjusting market mechanisms. A l l these related aggregation
schemes also have connections to chaotic dynamics.
The sixth section explores an implication o f chaotic dynamics in social
aggregation functions: the presence o f an underlying structure created by the
interaction o f individuals. Chaotic processes are unusual in that although they
exhibit complex, seemingly random possibilities, they are in fact highly con­
strained. The final section concludes by returning to the issues raised by
democratic theory and to the implications o f the individual-group connection
being chaotic.

F r o m Individuals t o G r o u p s

Democratic theory, particularly in its populist form, depends on a fundamen­


tal assumption that one can move coherently from individual preferences to a
societal choice (Dahl 1956; Riker 1982a). I f societal choices have little rela­
t i o n , no relation, or an inverse relation to the desires o f its citizens, then the
meaningfulness o f democracy is called into question. Society becomes an
entity that is independent o f the individuals that compose i t . As Riker (1982a)
points out, democracy depends on more than the ends it achieves; it is based
on the validity o f the means by which it achieves its ends. Voting, as the
mechanism by which individual preferences are aggregated into a democratic
choice, is at the center o f the legitimacy o f democratic means.
Early democratic theorists assumed that aggregating individuals' votes
was a coherent process. In some cases, the "general interest" was assumed to
be equivalent to the citizens' interests. Rousseau assumed that the social
contract created a "moral and collective body" that contained the w i l l o f the
people in what he referred to as the "general w i l l . " Similarly, Hegel spoke o f
the "national spirit" as a coherent single entity. Where addressed specifically,
the process o f moving from individual to group was viewed as an additive
relation: an aggregation or sum. Rousseau outlined a scheme where "pluses
and minuses cancel one another." Bentham viewed the connection as a sum o f
1
the utilities o f individuals. Even as recently as 1958, Truman ( 1 9 7 1 , 260)
spoke o f public opinion as "an aggregate o f the more or less rational opinions
held by the individuals who . . . make up the ' p u b l i c . ' "
From Individuals to Groups 91

Even these early theorists had already conceptualized the individual-


group connection in terms o f a mathematical function: a mapping rule from a
set o f individual preferences (or utilities, in the case o f Bentham) into
a societal choice or a societal preference ordering. Rousseau, Bentham,
Truman, and countless others interpreted this "function" in additive terms—
the simplest possible mapping. Yet this function need not be additive. As
A r r o w (1963, 4) points out, there are an infinite number o f possible mappings
from individual utilities to a social utility function: the sum o f individual
utilities, or their product, or the product o f their logarithms, or the sum o f
their products taken two at a time, and so on. A n d even i f one assumes that
interpersonal comparison o f utilities is nonproblematic, there is an implicit
value judgment in choosing the aggregation rule.
The assumption that the connection from individual to group is a
straightforward additive relation became weakened by the discovery o f pre­
sumably isolated voting paradoxes. As democracies spread and interest in
voting was becoming more common, examples were discovered, such as
those by Condorcet, Borda, Dodgson, and Nanson (see Black 1958), where
the aggregation o f votes led to "peculiar" social outcomes. These outcomes
are peculiar in the sense that a different outcome results i f the voting order is
changed, or the addition or withdrawal o f what should be superfluous options
causes unexpected and radical changes in the outcomes, such as reversing the
social ordering by making the winner the new loser (see, e.g., Black 1958;
Farquharson 1969; Fishburn 1973; Ordeshook 1989). The discovery o f these
examples was not consistent w i t h what was assumed to be an additive rela­
t i o n , for additive relations do not show such variability and sensitivity. How­
ever, the implications for democratic theory remained small as long as the
examples were viewed as interesting, but isolated, curiosities.
The discovery o f cases where the aggregation o f individual preferences
was problematic led to debates over which voting rule was the "best" in the
sense that it w o u l d not be subject to a breakdown in the coherence o f the
mapping from individuals to group. As early as the eighteenth century, Borda
and LaPlace were debating and advocating certain voting rules as better in
terms o f avoiding paradoxes (Saari 1985a). However, A r r o w ' s (1963) General
Possibility Theory demonstrated that no procedure exists for passing from a
set o f individual preferences to a pattern o f social outcomes that is consistent
w i t h a set o f minimally desirable conditions o f any democratic process. A l ­
though problems w i t h the individual-group connection did not occur for every
set o f individual preferences (as Black's [1958] single-peakedness condition
demonstrates), the possibility o f unstable group outcomes did occur w i t h more
than a negligible probability (e.g., N i e m i and Weisberg 1968; Riker and
Ordeshook 1973; Ordeshook 1989).
A further blow to the assumption o f a nonproblematic additive relation
between individuals and groups came when voting involved two or more
92 Chaos Theory in the Social Sciences

issues (Davis and Hinich 1966). When the topic to be voted on contains—
either explicitly or i m p l i c i t l y — t w o or more components, then stability in a
group outcome is nearly impossible (Cohen 1979; Cohen and Matthews 1980;
McKelvey 1976, 1979; Schofield 1979, 1980; Plott 1967). I n this case, any
social outcome can arise from a democratic process, since every outcome can
be defeated by some coalition o f voters. I n addition, voters have a disturbing
potential for manipulating the agenda in order to influence societal outcomes
(Gibbard 1973; McKelvey 1976).
We now know that the connection between individuals' preferences and
group outcomes is not as simple as originally thought. What was once as­
sumed to be a trivial exercise in aggregating or summing individual prefer­
ences in an additive way has become a "black box" that has the potential to
transform straightforward individual preferences into outcomes that can be
very complex and sensitive to small changes. The function that maps individ­
ual preferences to group choices is not linear and additive; in fact, as w i l l be
seen, the mapping is nonlinear. A nonlinear relationship implies that an i n ­
crease in one variable does not cause a uniform increase in the other variable:
uniform changes in individual preference orderings do not imply uniform
changes i n the social orderings, as is evident in the winner-turn-loser paradox
and the inverted-order paradox. There is a domain for functions o f this type:
chaotic dynamics. Since the social choice function is nonlinear, it is natural to
expect chaos theory—the theory o f nonequilibrium, nonlinear dynamics—to
apply to the function that aggregates individual preferences into a social
choice (see also Schofield 1993).

Chaotic D y n a m i c s

Before examining the dynamics o f the mapping from individual preferences to


social outcomes using chaos theory, it is worthwhile illustrating the charac­
teristics, methods, and formal definition o f chaos with a simple example from
outside o f choice theory. This section introduces the concepts o f chaos theory
by exploring the w e l l - k n o w n logistic function. The example is important
because it illustrates the definitions and techniques used later in the chapter.
The approach uses symbolic dynamics (Devaney 1989) and, in particular, the
iterated inverse images (Saari 1991). Using this approach makes demonstrat­
ing Devaney's (1989) three requirements o f chaos, sensitive dependence on
initial conditions, topological transitivity, and dense periodic points, rela­
tively straightforward. I n a subsequent section, these techniques w i l l be ap­
plied to the social choice function o f multidimensional spatial voting.
One o f the simplest functions that exhibits chaos is the logistic equation,
discussed in chapter 2. The logistic function illustrates the three important
characteristics o f a function exhibiting chaos: (1) sensitive dependence on
initial conditions, (2) topological transitivity, and (3) dense periodic points.
From Individuals to Groups 93

Devaney (1989) defines the presence o f chaos in a function by the satisfaction


of these three conditions:

D E F I N I T I O N 1. A function exhibits sensitive dependence on initial condi-


tions if two points that are arbitrarily close separate by a distance 8 by the
iteration of that function (Devaney 1989, 49).

D E F I N I T I O N 2. A function exhibits topological transitivity if the mapping


has points that move from one arbitrarily small neighborhood to another
arbitrarily small neighborhood (Devaney 1989, 49).

D E F I N I T I O N 3. A function has dense periodic points if, given a periodic


point, there is another periodic point arbitrarily close by (Devaney 1989, 15,
42).

These conditions combine for the formal definition o f chaos (Devaney


1989, 50):

D E F I N I T I O N 4. A function is chaotic if it satisfies sensitive dependence on


initial conditions, topological transitivity, and dense periodic points.

These three conditions have all been informally illustrated w i t h the example o f
the logistic function i n chapter 2. Sensitive dependence on initial conditions is
evident i n the effect o f a new initial condition. Topological transitivity implies
that one can move from one region of x to another region—i.e., that outcomes
do not merely repeat the same few values but have a diversity o f potential
outcomes. Dense periodic points are evident in the bifurcation diagram. Peri­
odic points are nearby other periodic points; there is a consistency in nearby
trajectories. Although all three conditions are intuitively evident in the ex­
ample o f the logistic function, the proof that these conditions are in fact
satisfied requires more. Establishing these three conditions formally becomes
straightforward when the tools o f symbolic dynamics are used.

Symbolic Dynamics

The intuition behind the presence o f cycles and chaos in a function becomes
clearer w i t h the tools o f symbolic dynamics (see, e.g., Devaney 1989; Saari
1991). This technique permits a cataloging o f the permissible sequences o f the
function. It accomplishes this by abandoning point precision, as the previous
discussion emphasized, and examining instead how regions map into regions.
In particular, this approach reverses the focus to look at how sets o f points
map to other sets, rather than at where a single point maps.
To illustrate this approach, divide x, = [ 0 , 1] into two regions, labeled /?,
94 Chaos Theory in the Social Sciences

and R , as shown in figure 5 . 1 . The image o f these regions is shown on the


2

x
t+) axis by reflecting across the x, = x line. One is interested i n which l + i

regions map to which regions via the function. In this case, the function is the
one-to-one mapping o f the logistic function.
Note that from figure 5.1 it is apparent that region 1 maps to region 2
because for some values o f x, G R , the function results i n some values o f x,
x +i

G R . Similarly, some values o f x, G R map to region 2. B y a similar logic


2 2

we can describe the rules o f mapping under the logistic function for all the
regions, summarized as

i) i f x, G R x then x G R;
t + l 2

ii) i f x, G R 2 then J C , , G R or R .
+ x 2

These "rules" on the sequences o f mappings outline some o f the permis­


sible sequences. Each o f these sequences is called a w o r d , and the set o f all
permissible words under a particular function is called the dictionary. Each
sequence describes the iteration o f a single starting point via that function. Se­
t h
quences that are identical for the first n entries and that differ after the n + 1
entry designate initial points that are nearby. I n other words, the distance
between t w o points is measured by the extent to which their sequences match
(see, e.g., Saari 1991). The possibility o f cyclic sequences arises because o f
the feedback region o f R . B y iterating w i t h i n R cycles o f all lengths are
2 2

possible; the mapping back to /?, allows for the completion o f the cycle. The
rule on the mapping is simply that R must be followed by R \ R can be x 2 2

followed by either /?, or R . 2

For example, a four-cycle sequence is possible: (/?,, R , R , R , R, 2 2 2 t

R , . . . ) . Finding this four-cycle sequence is not obvious w i t h the forward


2

approach used in choice theory. The iterated inverse image approach (Saari
1991) identifies where the cycle must appear. First, find all points in R that x

_ l
map to the first target region o f 7", = R , this is given b y / (T ) = f~ *(R ) H 2 t 2

R . However, the goal is not to map to any point in R , but to the subset o f
{ 2

points in R where the next iterate stays in R . This means we need to refine
2 2

l
the starting region o f /?, to the refined target o f T = f~ (R ) D R C 7",. 2 2 2

- X l
These p o i n t s / ( T ) C f~ (T ) are all the points that start in R , go t o R , and
2 { x 2

on the second iterate remain in R . 2

The same argument ensures that the third iterate is in R . Here the refined 2

target region is T = / - ' ( / - ' ( / ? ) <~l R ) C T , a n d / - ' ( T ) <= f~\T )


3 2 C 2 2 3 2

l
f~ (T ) } C R,. A t each stage, the set o f initial points is refined to include only
those points that hit the refined target—this is the region that ensures that
more o f the steps o f the proposed cycle occur. The final stage, T = 4

_ 1 l
/ ~ ' ( . . . ( / (/?,)) C 7" , defines the set o f inverse iterated points f~ (T )
3 C 4

R . This is the set o f all points in R where the image o f the fourth iterate is i n
x t
From Individuals to Groups 95

X 1 =4gx (1-X )
t+ t t

1.0 1 ~y
: /
: /
/
\: \ '
/

: \ /
:' \
V. \.
/
x
t+1 /
/
/
/
/
/
/
/
JC. . . . . . . ...... .
/ •
/ /
/ /
/ /
1/

0.0 F?i R 2 10
x
t
Fig. 5 . 1 . M a p of t h e l o g i s t i c f u n c t i o n

4
/?,. Trivially, there is at least one point x G R so that f (x) x = x. This is the
period-four orbit.
The possibility o f cyclic points and o f a set o f nonperiodic or chaotic
points arises because o f the presence o f a feedback in the logistic function. It
is the fact that region one maps to region t w o , which maps back to region one,
that allows for the abundance o f permutations o f sequences. A similar feed­
back process occurs in nonequilibrium social choice, evident in the presence
of an intransitivity among three alternatives.
Using symbolic dynamics, Devaney's three conditions for chaos can
easily be shown to be satisfied, verifying that the logistic function is indeed
chaotic. To establish sensitive dependence on initial conditions, one must
demonstrate that two points that are arbitrarily close w i l l separate by a dis­
tance 8 by the iteration o f the function. This requires that one can construct
two sequences o f symbols 1, 2 that are identical for n entries and that eventu­
t h
ally differ i n their n + 1 entry. Let the distance between regions 1 and 2 be 8.
t h l h
Assume that the « entry o f these two sequences is 1. Then the n + l entry
can be 1 for the first sequence and 2 for the other, separating the points by 8
v

and demonstrating sensitive dependence on initial conditions.


Topological transitivity requires that the mapping have points that move
from one arbitrarily small neighborhood to another arbitrarily small neighbor-
96 Chaos Theory in the Social Sciences

hood. Then any point that begins in one and ends in two satisfies this condi­
tion. It is straightforward to come up w i t h a permissible sequence based on the
mapping rules o f the logistic such that the two sequences end in different
symbols.
The condition o f dense periodic points requires that, given a periodic
point, there is another periodic point arbitrarily close by. Recall that distance
between the iteration o f points is evident by the extent to which two sequences
match. Given a sequence o f arbitrary length, it is possible to find another
sequence that matches the beginning n entries o f the first sequence, demon­
strating the existence o f two periodic points arbitrarily nearby. This is obvious
in terms o f the permissible sequences o f 1, 2. For example, the initial condi­
tions leading to (1222222) and (1222221) would be nearby points.
The satisfaction o f these three conditions verifies that the logistic func­
tion is chaotic. Devaney (1989, 50) summarizes these conditions in terms o f
three characteristics that a chaotic function possesses: unpredictability, inde-
composability, and an element o f regularity. The presence o f sensitive depen­
dence on initial conditions makes long-term prediction o f a chaotic process
impossible. Chaotic systems are indecomposable because they cannot be bro­
ken d o w n into two subsystems that do not interact; this arises because o f
topological transitivity. T h i r d , although a chaotic process exhibits complex,
apparently random, behavior, there is an element o f regularity, in that periodic
points are dense and nearby points behave similarly.

M u l t i d i m e n s i o n a l Spatial V o t i n g

These same techniques can be applied to the case o f voting by considering the
social choice function as a mapping rule. This section summarizes some o f the
notation and main results o f multidimensional spatial voting (see Kxehbiel
1988 or Strom 1990 for a complete survey), and can be skipped by those
familiar w i t h the standard notation and approach.
The choice among alternatives is often modeled spatially (Davis and
Hinich 1966), where the alternatives to be voted on are arranged on a contin­
uum, such as the liberal-conservative spectrum or the amount o f a budget
allocation. The one-dimensional spatial approach has led to several interesting
implications, notably Black's (1958) median voter theorem, stating that i f all
voters' preferences are single-peaked then a stable outcome emerges at the
median voter, and Downs's (1957) results on party competition and the ten­
dency for parties to move toward the center o f the distribution o f voters. The
single-dimension continuum can be extended to include voting over several
policy dimensions, such as when an issue implicitly contains two or more
components or when issues are linked and must be decided as a package.
Assume an alternative space o f issues with m issue dimensions. Assume
From Individuals to Groups 97

a set N = { 1 , 2, . . . n) o f voters where each voter has preferences over the


issues in the policy space. Each voter's most-preferred outcome is referred to
as his or her ideal point. Each voter also has preferences over the entire issue
space that are typically assumed to be a monotone decreasing function o f the
distance between a proposed point and the voter's ideal point. Then each
voter's indifference curves are spheres around his or her ideal point. These
utility functions are referred to as Type I utility functions and can be summa­
rized as

*>,yo||*-x,||<||y-*,.||, (1)

or, alternative x is preferred to (or indifferent to) alternative y by voter i f and


only i f the distance between x and /'s ideal point x is less than (or equal to) the
t

distance between proposal y and i's ideal point. Simply put, the rule on
preference among alternatives compared to a voter's ideal point is simply "the
closer the better."
For simplicity, I restrict attention to simple majority rule. A set o f voters,
M, is a majority-rule winning coalition i f

n
\M\ > ^ * i f n is odd and \M\ > " ^ ^ i f n is even.

L e t / d e n o t e the social choice function for simple majority rule. Then f(x) =
> t n e s e t
{y I y *} is o f points in the policy space that are majority-preferred to
x. In other words, f(x) is the set o f points that beats x by some majority
coalition. Iff(x) = 0 , then there are no proposals that beat alternative x and x
is the winner. A n y x £ X for which f(x) = 0 is called a core point. Note that
i f a core point exists then the iteration o f the social choice function w i l l reach
an equilibrium at outcome x.
However, the existence o f an equilibrium under the social choice func­
tion o f majority rule in a multidimensional policy space is extremely rare. In
order to achieve a majority rule equilibrium outcome, severe restrictions on
the voters' ideal points are required (Black 1958; Plott 1967). For nearly all
voter preferences, majority rule does not lead to any stable winning social
outcome, but continues to wander over the policy space. Every issue can be
beaten by some other proposal that is also preferred by a majority o f the
voters. I n addition, the movement o f winning proposals can exhibit intransi-
tivity or cycles, where, by a sequence o f majority votes, the social outcomes
w i l l be alternative b over a, c over b, and then a over c! Moreover, any point
in the policy space can be reached by some sequence o f majority rule deci­
sions. When an equilibrium breaks d o w n , it breaks down completely and
includes the entire policy space (Cohen 1979; Cohen and Matthews 1980;
McKelvey 1976, 1979; Schofield 1979, 1983). These results are typically
98 Chaos Theory in the Social Sciences

referred to as the chaos theorems, because o f the negative interpretation that


all order and stability are lost: agendas can be constructed between any two
points, sequences o f votes do not reach a stable equilibrium consensus, and
any final group choice may have little relationship to the voters' preferences.
However, the use o f the term chaos was semantic. The following section
demonstrates that the connection between multidimensional spatial voting and
chaos is not merely semantic, but theoretic.

Chaotic D y n a m i c s in M u l t i d i m e n s i o n a l Spatial V o t i n g

The previous section briefly summarized some o f the troubling aspects o f


multidimensional voting and pointed out the inherent and pervasive instability
in voting. B y viewing the social choice process as a nonlinear mapping,
symbolic dynamics can be used to establish that the sequences generated by a
group o f voters in a multidimensional policy space do in fact exhibit chaotic
dynamics (Richards 1992, 1994). The chaos theorems o f social choice can be
understood as chaotic nonlinear dynamics; the connection is not merely se­
mantic, but is theoretic. This implies that outcomes and agenda paths are not
completely random and all order is not lost.
This section demonstrates the existence o f chaotic dynamics in m u l t i ­
dimensional spatial voting using an approach analogous to the example o f the
logistic function shown in the second section. However, there is an important
technical distinction. In the case o f the logistic function, the mapping was
one-to-one and continuous. I n the case o f multidimensional social choice, the
"function" is instead a correspondence: each policy point maps to a set o f
points that beat it by majority rule. However, the general logic remains the
same: divide the space into regions and examine how regions map into other
regions. The rules o f the mappings catalog the feasible dynamics and allow
for the establishment o f Devaney's three conditions o f chaos.

The Importance of Cycles

2
Before examining the inverse image o f the social choice function, two pre­
liminary facts must be established. First, every multidimensional spatial vot­
ing context without a core has at least one three-cycle among alternatives.
Second, i f a cycle o f any length greater than three exists, then a three-cycle
also exists. These results are similar to previous findings (e.g., Cohen 1979;
Schofield 1983), but are presented using concepts needed for the subsequent
results. First, the notion o f a cycle must be clarified. A n alternative y G X can
be reached from a point x G X i f it is possible to move from x to y by iterations
of / . I f an x G X can be reached from itself by k iterations o f / a n d x ¥= p(x) for
1 ^ j < k, then a £-cycle exists.
From Individuals to Groups 99

(1) For t w o or more issue dimensions, three or more voters, type I


utility functions, and majority rule, then i f the core is empty, there exists
a sequence o f alternatives (a, b, c) such that b £ f(a), c £ f(b), and a £
f(c).

To see that this is true, suppose that a three-cycle exists among three
alternatives, a, b, and c. Then a, b, and c form a nondegenerate triangle i n
m
R . Label the three perpendicular bisectors o f ab, be, and ca as H , H , and
ab bc

H,
ca which must intersect in a space o f m — 2. Denote H% as the open
b

halfspace that includes b and that by construction contains a majority o f voter


ideal points. Denote / / £ . and H+ similarly, such that each halfspace contains c
a

and a, respectively, and contains a majority o f voters. A cycle exists among


three points a, b, and c i f such a partitioning is possible (fig. 5.2). This
partitioning creates the "pinwheel" o f social choice preferences such that b £
f(a), c £ f(b), and a £ / ( c ) . Therefore, the question o f the existence o f a
three-cycle for a given configuration o f voters is the question o f whether
voters' ideal points can be partitioned into the appropriate pinwheel o f hyper-
planes.
Assume voters' ideal points are located i n the policy space such that a
core exists. Denote the core point as v*. Then v* £ f(x) for all x v* and
/ ( v * ) = 0 . The agreement set o f any winning coalition o f voters must contain
v* or there w o u l d be a winning coalition that could beat v*, violating the
construction o f v* as a core point. Therefore, in partitioning the voters into the
sets H+ , H£ , and H+ , v* must be an element o f H+ , and o f H£ , and o f H+ ,
b c a b c a
100 Chaos Theory in the Social Sciences

i m p l y i n g that H+ f l
b D H+ ¥= 0 . However, for a three-cycle pinwheel
a

partition to exist, the hyperplanes must be such that H£ D H£ ^ 0 and H+


b c h

D f l H+ = 0 (see fig. 5.2). Therefore, i f a core exists, a three-cycle


a

partition is not possible.


Conversely, assume voters' ideal points are located in the policy space
such that a core does not exist. I f the core is empty, then the halfplanes H+ , b

Hg , and H+ must be able to be placed such that a majority o f voters are


c a

contained in each H and such that H£ f l / / £ . D H+ = 0 . I f this were not the


+
b a

case, then there w o u l d be a point that must be included in every winning


coalition's set and this point w o u l d be a core. Given that there is some H ca

such that H+ D H£ Pi / / + = 0 , then there must be an H such that the three


b c ca

lines intersect in a single point. Either the three lines already intersect i n a
single point or H~ includes all o f the cone H^ D H . I n the second case, H
a b bc ca

has the flexibility to move toward the vertex o f the cone (toward H 's "—" ca

region), since this can only result i n more, rather than fewer, voter ideal
points. Therefore, H can be moved until it intersects H and H in a single
ca ab bc

point at the base o f the H+ f l Hfc. cone. Therefore, i f there is no core then a
b

three-cycle partition is possible.

(2) I f there is a &-cycle for a k s 3, then there is some a', V, and c'
such that V = f(a'), c' —f{b'), a n d a ' = f(c'), i . e . , there is a three-cycle.

To see that this is true, note that the existence o f a &-cycle implies that the core
is empty (Cohen 1979; Schofield 1983). B y (1), i f the core is empty then a
three-cycle exists.

The Iterated Inverse Image of the


Social Choice Function

The previous discussion established that when voters' ideal points are such
that no equilibrium exists, then the majority-rule social choice function cre­
ates a mapping from a to b, from b to c, and from c to a. As in the example o f
the logistic function, the existence o f regions that feed back into previous
regions allows for the existence o f cycles o f all lengths and for the demonstra­
tion o f chaos. Therefore, one can characterize the permissible sequences o f
the social choice function in terms o f mappings o f regions. This is accom­
plished by examining the iterated inverse image o f the social choice function.
The policy space is partitioned into regions, and by examining the nested
sequences in terms o f target regions, one can outline the permissible se­
quences under the iteration o f the social choice mapping.
This approach differs from the typical approach to spatial voting in two
From Individuals to Groups 101

ways. First, rather than examining the mapping o f single points—as in the
" w i n set" formed by the indifference curves through a single point—this
approach examines the mapping o f regions into regions. B y abandoning the
point precision o f previous approaches, one can characterize the general dy­
namics o f the social choice function (Saari 1991). Second, the iterated inverse
approach differs from previous analyses o f spatial voting by focusing not on
the social choice function, but on its inverse, / " ' ( * ) · Previous approaches
examined the set o f points that beat an existing status quo point, namely, the
set/(jc). The iterated inverse image approach reverses this logic and examines
the target region o f x—namely, the set o f points that map to x via the social
choice rule. I f y is the target region o f x, then x can follow y in a social choice
sequence. The iterated inverse image approach outlines which sequences o f
outcomes are possible under / b y examining the iterations o f target regions.
This allows for a general cataloging o f everything that can happen under the
social choice rule in terms o f permissible sequences o f target regions.

P R O P O S I T I O N 1. For two or more issue dimensions, three or more voters,


type I utility functions, and majority rule, if a three-cycle among alternatives
exists, then there are cycles among alternatives of all periods greater than or
equal to three.

P R O O F . Let a, b, and c denote three alternatives in a three-cycle, w i t h


H , H , and H the respective perpendicular bisectors (as i n fig. 5.2). Label
ab bc ca

a closed region B bounded by a hyperplane parallel to H and such that b G ab

B. Define a set R(B) as the open set o f all points defined by reflecting across
H ab (see fig. 5.3). By construction a G R(B), since H is the perpendicular
ab

bisector and b G B. Regions C, R(C), A, and R(A) are defined similarly, such
that cEC,bE R(C), a £ A, and c G R(A) (see fig. 5.3).

Assume a sequence {A, B, B, . . . , A, . . . } . First we need to find those


points i n A that map to T, = B in the first iterate. This is given by the cone
x l
f~ (T )x = R(B) f l A. N o w we need to refine/~ (T ) to only those points that
x

remain in B i n the second iterate. That is, we want to find the points in A that
l
map to the refined target region o f T = f~ (B)
2 f l B CT . The open cone B n
X

l
Hbc G f~ (B), since all points i n this cone map to B by approaching the H bc

hyperplane, so T = B D H . The same argument ensures that the third


2 bc

l l
iterate is i n f l . Here the refined target region is T — f~ (f~ (B)
3 D B) C T . A t
2

each iteration, the set o f initial points is refined to include only those points
l l
that remain i n B. Then = f~ . . . (f~ (B) f l B); one can stay in B as
many times as necessary since the boundary H bc is open. The final stage
defines the set o f inverse iterated p o i n t s / ~ '(7*,) C A—i.e., the set o f points i n
102 Chaos Theory in the Social Sciences

R(B) \

ROM

Fig. 5.3. Three-cycles w i t h i n t h r e e a l t e r n a t i v e choices

t h
A where the image o f the i iterate returns to A. This means the final target
x
region must be refined to T = H f l f~ (A).
t This is an open cone and is
bc

nonempty. Therefore, one can construct a sequence that begins in A, iterates


w i t h i n B for any number o f iterations, and then returns to A.
Since sequences o f the type {A, B, B, . . . , A, . . . } are permissible
under this social choice rule, it is straightforward to see that cycles o f all
length greater than three can be constructed: vary the number o f iterates w i t h
B from t w o (a three-cycle) to fl = j (an i + 1-cycle). Note, however, that
unlike the logistic function, the presence o f a three-cycle feedback does not
3
imply cycles o f all lengths in the social choice function. "One-cycles" ( i . e . ,
equilibrium points) are not implied. Although region fl can follow region fl,
in fact, subsequent points in fl cannot return to previous points. As seen in the
above discussion, mappings w i t h i n fl must approach the open boundary o f
H.
bc Therefore, there is no periodic point o f one.
Proposition 1 makes clear that the mapping rules for the social choice
function allow region B to follow A, C to follow fl, and A to follow C. In
addition, note that each region can follow itself, merely by moving closer to
its perpendicular hyperplane. Therefore, the mapping rules on the majority-
rule social choice function can be summarized for x, as the status quo alterna­
tive

i) i f x, G A then x t + l G A or fl or C;
ii) i f x, G fl then x, +x G A or fl or C; (2)
iii) i f x G C then x
t t+l G A or fl or C.
From Individuals to Groups 103

Chaotic Dynamics

Recall from the example o f the logistic function that the presence o f chaotic
dynamics is established by demonstrating satisfaction o f three conditions:
sensitive dependence on initial conditions, topological transitivity, and dense
periodic points. A s in the case o f the logistic function, demonstrating that the
social choice function is chaotic becomes a straightforward exercise when the
permissible sequences in terms o f symbolic dynamics are used.

P R O P O S I T I O N 2. For two or more issue dimensions, three or more voters,


type I utility functions, and majority rule, if an intransitive social choice cycle
exists among any alternatives, then the social choice function exhibits chaotic
dynamics.

Recall that to demonstrate sensitive dependence on initial conditions,


one must demonstrate that two points that are arbitrarily close w i l l separate by
a distance 8 by the iteration o f the function. Let the distance between regions
A and B be 8. Then sensitive dependence on initial conditions implies that one
can construct t w o symbolic sequences that are identical for n entries and that
t h l h
differ on the n + 1 entry. Assume that the n entry o f these two sequences is
t h
A, then the n + 1 entry can be A for one sequence and B for the other,
separating the points by 8 and demonstrating sensitive dependence on initial
conditions.
Recall that topological transitivity requires that the mapping have points
that move from one arbitrarily small neighborhood to another arbitrarily small
neighborhood. Then any point that begins in A and ends in B satisfies this
condition. Clearly such a point does exist, by the permissible sequences
outlined in (2).
Recall that the condition o f dense periodic points requires that, given a
periodic point, there is another periodic point arbitrarily close by. Distance
between the iteration o f points is evident by the extent to which two sequences
match. Given a sequence o f arbitrary length, it is possible to find another
sequence that matches the beginning n entries o f the first sequence, demon­
strating the existence o f two periodic points arbitrarily nearby.
By examining the social choice correspondence o f multidimensional spa­
tial voting i n terms o f the iterated inverse images, the demonstration o f the
existence o f chaotic dynamics becomes straightforward. The existence o f a
three-cycle intransitivity arising from a nonequilibrium preference configura­
tion creates a feedback process among alternatives. It is this nonlinear feed­
back that, as in the case o f the logistic function, creates complex behavior in
the mappings, allowing for countless cyclic outcomes and nonperiodic, cha­
otic sequences.
104 Chaos Theory in the Social Sciences

Chaotic Dynamics i n Other I n d i v i d u a l -


Group Aggregations

Multidimensional spatial voting is one subset o f the larger class o f aggregation


schemes from individuals to groups. The multidimensional case is particularly
interesting however, since virtually all voter preferences lead to chaotic dy­
namics. This section addresses the question o f the general robustness o f
chaotic dynamics i n the aggregation from individual tastes to social outcomes.
To do this, I briefly review two other aggregation schemes where the existence
of chaotic dynamics has been established. These include Saari's (1984,
1985a, 1987, 1989) work on simple voting and the findings o f chaotic dy­
namics in market mechanisms.

S i m p l e Voting

Unlike multidimensional spatial voting, in simple voting individuals rank a


finite set o f discrete one-dimensional alternatives. Each voter has a preference
ordering over the alternatives and the social choice is determined using a
given voting rule, such as pairwise majority votes, two-thirds rule, or the
Borda count.
As in multidimensional voting, one would like the function mapping
individual preferences to a group choice to be coherent in some sense. Arrow
(1963) formalized the notion o f coherence in terms o f five conditions:
(1) every individual preference ordering that is complete and transitive is
4
admissible, (2) i f all voters prefer an outcome x over an outcome y then the
social ordering must be x over y, (3) i f x and y are the only two outcomes that
the group can consider, then the social preference between x and y depends
only on individual preferences over {x, y } , and not on individual preferences
over a larger set, (4) no single voter is decisive for every pair o f outcomes,
and (5) the additional requirement that the social choice ordering must also be
complete and transitive. A r r o w showed that no voting rule exists that satisfies
these conditions. N o matter what voting rule is used, the mapping from
individual to group w i l l not be trouble-free. Paradoxes in individual-to-group
mapping w i l l occur for every voting rule.
The mapping from individual to group in the case o f simple voting also
exhibits chaos, as demonstrated by Saari (1984, 1985a, 1987, 1989, 1990,
1992). Using symbolic dynamics, and for any number o f alternatives and any
positional voting procedure, Saari catalogs all possible social choice rankings
over all possible subsets o f candidates. He shows that any social choice
ordering—no matter how paradoxical (such as the social ordering a > b, c >
a, a > b = c)—can be achieved by some profile o f voters. A n y outcome is
From Individuals to Groups 105

possible, paradoxes are abundant, and it is impossible to construct a voting


5
rule that avoids these paradoxes. This occurs because all sequences are
possible under the individual-group mapping rules. The permissible mappings
in any positional voting procedure create the existence o f chaotic dynamics in
simple voting contexts.

Market Mechanisms

In addition to voting schemes, individual preferences can also be aggregated


using market mechanisms. As A r r o w ( 1 9 6 3 , 1 - 2 ) and Riker and Ordeshook
(1973) point out, voting and the market are special cases o f the general
category o f collective social choice. Both market mechanisms and voting
schemes take the preferences and desires o f many individuals and "amalgam­
ate" them into a single social choice. I n the case o f voting, the choice is a
selection o f a candidate or policy. In the case o f the market, individual
preferences over quantity and prices o f goods translate into a single market
price. I n this way, the market incorporates individual preferences and results
in a social choice in terms o f the allocation o f goods among the individuals
(Riker and Ordeshook 1973). However, i n their acts o f voting, individuals are
consciously choosing and anticipating a social outcome. I n the act o f a market
mechanism, the act is less consciously oriented toward the ultimate social
outcome (Riker and Ordeshook 1973, 82). Nonetheless, the collective out­
come o f a market mechanism, such as the price o f a commodity, is still
achieved by an aggregation rule that takes individual preferences over the
commodities and leads to a social outcome by the combination o f individual
actions.
As in voting, there are various rules by which market mechanisms can
respond to individual preferences. Here I focus only on the most general
form. The basic idea is that one has an exchange economy with a fixed
number o f commodities, a fixed amount o f goods, and a set o f agents w i t h
given initial endowments. Each individual has preferences over each good,
where an individual's demand is given by his or her utility function. The
combined demand o f all the individuals is the excess demand function. As­
sume that total buy is not equal to total sell. The market mechanism is a force
that w i l l move prices to an equilibrium point where total buy equals total sell
and the market clears. Hypothetically, this is modeled as being adjusted by a
Walrasian auctioneer, whose mission is to search for the market clearing price
(see, e.g., Varian 1992). Prices move iteratively in the direction o f excess
, h
demand: i f the i commodity is i n demand, then its price increases until
demand equals supply and it clears. I n general, these models are described as
tâtonnement models or price adjustment mechanisms. Depending on the de-
106 Chaos Theory in the Social Sciences

mand functions, the individuals' preferences w i l l combine through the excess


demand function and the market rule to yield an equilibrium price and thereby
allocate goods among individuals.
However, it is w e l l k n o w n that such market mechanisms need not con­
verge to a stable social outcome. For example, economists have long known
that price adjustment mechanisms are not stable for two or more commodities.
The presence o f chaotic dynamics in iterative market adjustments has also
now been established (see, e.g., Day and Pianigiani 1991). Saari and Simon
(1978) and Saari (1985b, 1992) show that for most prices there need not be
any relationship among the demand functions. The outcome o f price models
can lead to many outcomes in terms o f price and allocation, and the social
outcome is very sensitive to changes in the set o f agents, the set o f commodi­
ties, or the set o f preferences. In addition, no market procedure satisfies three
minimal conditions in moving from aggregate excess demand to a market
price (Saari 1991). Majumdar (1992) and Bala and Majumdar (1994) show
that w i t h t w o commodities, two agents, and the price o f one good fixed,
chaotic dynamics w i l l result, as it w i l l in a competitive economy with two
commodities.

C o n s t r a i n t s o n Instability: The Order


in Chaotic D y n a m i c s

Interpretations o f a chaotic process in terms o f the colloquial use o f the word


chaos—as disorderly, random, or conflictual—ignore the unique aspects o f
chaotic dynamics. A chaotic process is disorderly: it results in an infinite array
o f seemingly random possibilities; it is sensitive to very small shocks; it never
reaches a stable equilibrium; and it is unpredictable in the long term, no
matter how long its behavior is empirically observed. However, the other side
of chaos is the presence o f an underlying order in the dynamics. Although the
outcomes o f a chaotic process are so complex they appear random, in fact,
chaotic processes incorporate specific constraints on these outcomes. It is this
mix o f complexity and order that makes chaos theory an unusual theoretical
framework. In this section we explore some preliminary evidence o f this
structure in the case o f multidimensional spatial voting and discuss its i m p l i ­
cations for understanding the individual-group aggregation process.

The Constraints o n O u t c o m e s and Paths

It is well known that in a multidimensional setting, any outcome in the policy


space—even points that are very far away from the voters' ideal points—can
be reached by a sequence o f majority-rule votes. However, this result is often
From Individuals to Groups 107

interpreted to imply more than it technically does. While it is true that any
point can be reached from any other point, this does not imply that any point
can be reached from any other point by any path. In effect, for any two origin
and destination points, " y o u can get there from here," but the routes from here
to there may be quite constrained.
The symbolic dynamics approach emphasizes the constraints on feasible
sequences. The mapping rules outline what is permissible and what isn't
permissible i n the social choice dynamic. I f majority rule was in fact total
disorder, then every region would map to every other region even as regions
were refined to smaller and smaller sets. Although the construction o f the
fourth section d i d have the property o f the dictionary being the universal set, it
also contained only three partitions. It is obvious that as the regions are
refined, rules on the mapping w i l l emerge and the dictionary w i l l no longer be
the universal set. These constraints were already evident in terms of mappings
within a region. Although, for example, B maps to B, in fact, points must
approach the hyperplane with each iteration: not all points in B map to all
other points in B in one iteration. A refinement o f the region B would high­
light these mapping rules.
Another example o f the order in a chaotic process is evident in the
condition o f dense periodic points. Recall that this condition states that for
any periodic point there is another periodic point arbitrarily close by. I n the
case o f the logistic function, the order in terms o f nearby dynamical behavior
was evident in the bifurcation diagram, where two-cycles transformed to four-
cycles, then a chaotic region, and so on. Bifurcation diagrams can also be
constructed o f the multidimensional spatial voting social choice dynamics.
Like the bifurcation diagram o f the logistic function, the bifurcation diagrams
of the social choice function show complex patterns (fig. 5.4).
The bifurcation diagrams o f figure 5.4 show the outcomes in one o f the
issue dimensions (y on the vertical axis) as a function o f the winning coalition
choosing either an extreme point (near the tip o f the coalition agreement set)
or an incremental point (near the status quo p o i n t — i . e . , the base o f the
coalition agreement set). For many steps over this parameter continuum, a
sequence o f eighty points was generated and the last fifty points were plotted
in a horizontal line. For example, part b o f figure 5.4 shows a three-cycle for
some o f its region. As seen by comparing parts a and b o f figure 5.4, a
different fine structure emerges as a result o f the voting rule and the individual
preferences. Each set o f voters produces a unique bifurcation diagram, con­
stant to transformations that preserve the configuration o f the voters' ideal
points (such as rotating or a linear transformation). Although these diagrams
are o f limited use (because o f the nature o f the parameter choice to make the
function one-to-one), they do illustrate the fine structure to the dynamics o f
the social choice function.
108 Chaos Theory in the Social Sciences

150-r— 1 200-

-200
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

maximum feasible step <-—> minimum incremental step maximum feasible step < — > minimum incremental step

Fig. 5.4. T w o s a m p l e b i f u r c a t i o n d i a g r a m s . D i s t r i b u t i o n o f v o t e r ideal


p o i n t s is: (10. 10) (20, 10) (35, 20) (20, 80) (70, 30) a n d (60, 30) (60, 70)
(70, 50) (90, 7 0 1 ( 1 2 0 , 1 1 0 ) .

The Structure of Outcomes

Given that the social choice function for multidimensional spatial voting is a
chaotic process, we know from the definitions o f chaos that there must be a
strict structure to the feasible outcomes. There are suggestions o f this struc­
ture in a simulation o f iterations o f majority-rule decisions.
Figure 5.5 shows the distribution o f outcomes w i t h five voters deciding
by majority rule over a two-dimensional issue space. Since majority rule is
not one-to-one, for the simulation a single point had to be chosen as the
winner. I t was arbitrarily set as the midpoint o f the largest winning coalition's
agreement set. Note that the choice o f this point is arbitrary—it merely serves
as an anchor on the movement o f winning coalitions. The simulation was
begun w i t h an arbitrary initial point, a . a is beaten by a,, which is the
0 0

alternative that is preferred to a by the largest winning coalition, where a, is


0

the midpoint o f that coalition's agreement set. N o w a, is also not stable and
can be beaten by a new set o f points by a reformation o f coalitions. a is the 2

point that beats a, and is the midpoint o f the largest coalition that can outvote
a,. Continuing this iterative process provides a sample o f the distribution o f
6
alternatives under the majority-rule social choice.
As is apparent in figure 5.5, the iteration o f this process is far from
7
random, as evident i n the complex patterns that emerge. Just as in the case o f
300

V.

A · -\

/4

— ι 1 1 1 • 1 • 1 • —

-200 -100 0 100 200 300


Issue 1

:' ν'
/ ' ·.
' ,Λ S

; if ' «. IV
;* ·
"t
H ' i
- >·(··
I t .*

«#•r·>
<• 3Γ

1
1

10 30 50 70 90
Issue 1

Fig. 5.5. T w o s a m p l e s of t h e d i s t r i b u t i o n of g r o u p choices


110 Chaos Theory in the Social Sciences

the logistic function, there is an order to the dynamic o f the majority-rule


social choice function.
This order has been anticipated by previous studies. For example, there
has been a proliferation o f new solution concepts that seek to capture some o f
the order to voting outcomes. As a sampling, some o f the best k n o w n are the
generalized median set, the minmax set (Kramer 1977), and the uncovered set
(McKelvey 1986). It is not necessary to review or define any o f these con­
cepts, but they do have some relevant common properties. First, these sets are
typically constructed as descriptions o f likely outcomes. For example, Kramer
(1977) notes that over time, sequences roughly converge onto the minmax set.
Ferejohn et a l . (1980, 1984) suggest a probabilistic convergence o f outcomes
toward the generalized median set. Second, these sets are all "centrally
located"—i.e., they tend to be centered around a region in the geometric
center o f the voters' ideal points. For example, Feld et al. (1985) demonstrate
that agendas that move toward the generalized median set are easier to con­
struct than agendas moving away from the center o f the voters. T h i r d , they are
all compact solution sets. For example, the minmax set is a polygon bounded
by voter median lines. The generalized median set and the uncovered set are
circles o f a particular radius.
The order suggested by the theory o f chaotic dynamics has also been
noticed by experimentalists (e.g., Fiorina and Plott 1978; W i l s o n 1986). They
find that outcomes are not only not scattered over the policy space, but tend to
be clustered i n patterns. For example, W i l s o n (1986, 404) concludes that "one
point is clear; the outcomes selected here are not scattered throughout the
alternative space." Rather than covering the entire policy space, most experi­
mental studies find that outcomes cluster w i t h some deviation near the center
of the voters' ideal points. I suspect that the experimentalists are capturing
portions o f the structure o f the chaotic dynamics in their results.
What is happening in the dynamics o f social choice is that the individual-
level preferences i n a context o f interaction ( i . e . , multidimensional majority
rule) are creating their o w n structure, which in turn constrains feasible out­
comes. Although there are no external constraints on individuals' votes (in
terms o f preference orderings, etc.), the voters' actions combine to create a
system where each voter, in effect, constrains the other. Game theorists are
familiar with this intuition in the important role that the rules o f the game have
on structuring feasible outcomes (see, e.g., Schelling 1978). Even i f individ­
uals have complete autonomy over their actions, the context o f their inter­
action creates structural limits on what individuals can do and therefore on the
group outcomes.
Note the parallels between the order implicit in the chaotic dynamics o f
the individual-group connection and the concept o f structure in social science.
In the social sciences, structure is something generated by the interaction o f
From Individuals to Groups 111

unit-level behavior, yet conditions and constrains system-level options. How­


ever, structure in social science is not an external, concrete entity (as an
institution might be). "Structure is not something we see. . . . [SJince struc­
ture is an abstraction, it cannot be defined by enumerating material charac­
teristics o f the system. It must instead be defined by the arrangement o f the
system's parts and by the principle o f that arrangement" (Waltz 1979, 80).
"Structure is a system-wide component that makes it possible to think o f the
system as a w h o l e " (Waltz 1979, 79). It is this structure that makes a chaotic
system indecomposable. The whole is complex (disorder), but it is united by a
common endogenous structure (order), whereby the system cannot be broken
d o w n into additive subsystems.
Chaotic systems are disorderly. They are characterized by an acute sensi­
tive dependence on initial conditions, where slight changes in individual-level
input lead to radically different social outcomes. This characteristic in turn
leads to an impossibility o f long-term prediction based on empirical observa­
tion. The dynamics appear random and, no matter how long the dynamic is
observed, no repetitious, deterministic patterns w i l l be evident. Prediction
must shift from point predictions to boundary predictions, from long-run to
short-run. Chaotic dynamics are also disorderly because o f the vast diversity
o f achievable outcomes. Outcomes are not limited to a few, often-reached,
repeated outcomes. A n y outcome can be reached, from somewhere, via some
route, by some length o f time. As others have already established, group
outcomes are possible anywhere in the policy space.
But chaotic systems are also orderly. They incorporate rigid and u l t i ­
mately describable constraints on mappings and on routes from group choice
to group choice. Previously made choices do constrain future options. Paths
are closed, but paths are also continually opened. Social choices are not
completely unrelated to individual preferences; there are constraints, but these
constraints are "strange." Chaotic systems are also orderly in that they display
a fine structure in the distribution o f points and the paths. I have alluded to this
complex structure in the patterns o f the bifurcation diagrams and o f the
distribution o f outcomes from the dynamic. But the structure is multifaceted.
It is not merely a circle or a polygon. It is a unique endogenous structure
created by the aggregation o f individual-level behavior in a given context.
Chaotic systems are also orderly in that there is a "consistency o f weirdness":
stable regions adjoin stable regions and chaotic ones are near chaotic ones.
There is a smoothness in the dynamical profile o f the social choice function,
as made clear by the condition o f dense periodic points.
We have long known that the aggregation o f individual choices is prob­
lematic. I f we can understand this mapping as a chaotic process then we have
some concepts and tools available that allow for the incorporation o f a m i x o f
order and disorder. The process from individual to group is not a regular,
112 Chaos Theory in the Social Sciences

predictable, determinate process. Yet neither is it an incoherent, completely


random process. I t is this blend that is so unique—both to chaotic dynamics
as a general framework and to the aggregation o f individual preferences in
particular. The individual-group connection has order and constraints that it
creates itself as a result o f the dynamics o f the social choice function. How­
ever, the individual-group connection is also capable o f infinite variety and a
perpetual evolutionary dynamic, where—even assuming deterministic indi-
vidual behavior rules—group outcomes cannot be predicted in the long run.

Chaos, Social Choice, a n d D e m o c r a c y

The presence o f instability in social choice has already been discussed in


terms o f the implications for democracy. These implications fall into two
categories: (1) the view that the instability makes certain forms o f democracy,
such as populism, infeasible (Riker 1982), and (2) the view that the instability
may help the political stability o f pluralist democracy ( M i l l e r 1983). I w i l l
briefly outline some o f the implications for each o f these views in light o f the
individual-group connection being a chaotic process, although I w i l l not at­
tempt to resolve the implicit normative debate.

View 1 : Incoherent Social Choices

Given a social choice function that results i n outcomes that can be anywhere
and that are not consistent w i t h individual preferences, Riker (1982) makes
the case that liberal democracy, rather than democracy in its populist form, is
more effective. Liberal democracy does not attempt to represent its citizens'
preferences, but is democratic in that citizens have the opportunity to remove
their representatives. Populism, on the other hand, seeks to aggregate citi­
zens' preferences on specific issues, achieving a social choice by combining
and weighing citizen's views. However, based on the pervasive presence o f
instability in the social choice function, Riker asserts that liberalism is the
only democracy that works, since it bypasses the instability problems o f the
social choice function. Populist democracy, being dependent on an imperfect
and troublesome aggregation scheme, w i l l only result in arbitrary decisions
and continually changing social outcomes.
What i f these aggregation schemes, including the individual preference-
group choice connection, are chaotic? The instability was previously known.
The new information is that the aggregation function that populism relies on is
a chaotic process. Returning to the technical characteristics leads to several
implications.
First, group outcomes w i l l appear arbitrary. Studying the relationship
From Individuals to Groups 113

between outcomes over time without knowing the mapping rule, the choice
w i l l appear random. N o amount o f empirical observation w i l l allow for pre­
diction. However, as seen in both examples, these outcomes are far from
random. Outcomes are strictly related to individual preferences. They are not
arbitrary, in the sense that the same set o f individual preferences results in
different outcomes at different times. But outcomes are very sensitive to slight
changes in these preferences and the connection between individual and group
is complex. Complex does not mean random or arbitrary, however. Outcomes
in the logistic function were complex, but they were quite "regular" in terms
of resulting from a strict deterministic mapping.
Although the mapping is certainly not random or arbitrary, this may not
be enough for populism to work. Perhaps the very presence o f any instability
must be removed. I f we understand that the aggregation function is chaotic,
what does this tell us about removing the instability? First, it says one can't
solve the problem by breaking it down into smaller problems. I f there is
instability in the whole group, trying to break it down into stable subsystems
w o n ' t w o r k . This tends to imply that i f population w o n ' t work in its purest
form then neither w i l l a type o f "hierarchical populism." Chaotic systems are
indecomposable.
Second, it tells us w h y the instability is occurring. I t is not merely
unstable; i t is unstable because it is in the chaotic regime, which in turn
implies that it is nonlinear and there is some "functional folding" or feedback
occurring. As seen in both the logistic and spatial voting, it is this feedback
process that creates the instability. In theory, i f one could "unfold" the social
8
choice function then the instability would disappear! B y understanding the
aggregation process as chaotic, one can examine why it is occurring. I f one
knows w h y the mapping is chaotic, then one can systematically, rather than
haphazardly, outline ways to reduce the instability, rather than identifying
isolated cases where instability does and does not occur.

View 2 : Instability Fosters Political Stability

The case for the instability o f social choice allowing for a context that sup­
ports the stability o f pluralism is made by M i l l e r (1983). He reasons that the
instability o f the social choice process in fact helps promote pluralism. Plural­
ism views preferences as largely determined by the issue group to which an
individual belongs, w i t h society divided along one or more lines that partition
the individuals into sets. Where the arrangement is pluralistic, these cleavages
are in a crosscutting rather than a reinforcing pattern: individuals that disagree
on one issue are likely to agree on other issues. In addition, intensity over
issues is also dispersed. The argument o f pluralism is that pluralistic prefer-
114 Chaos Theory in the Social Sciences

ences contribute to the stability o f the political system. M i l l e r makes the


argument that the instability o f the social choice function contributes to the
stability o f pluralism. A s M i l l e r (1983, 744) states: "not only does each
competitor ' w i n some and lose some,' but most wins and losses are them­
selves reversible. Thus the competitors can never be confident o f their victo­
ries, nor need they resign themselves to their defeats." It is this indeterminacy
and instability o f winners created by the aggregation process that induces
losers "to continue to play the political game, to continue to w o r k w i t h i n the
system rather than try to overthrow i t " (1983, 744). Pluralism only guarantees
that different people w i l l w i n or lose on different issues. The instability o f the
social choice function guarantees that losers on a given issue can hold out in
hope o f being winners on the same issue.
The presence o f an instability in the social choice mapping may promote
the stability o f a democratic system, but it does not imply any sort o f effi­
ciency. As Rae (1982) points out in his metaphor o f disequilibrium, demo­
cratic decision making becomes an "Escher staircase": leading always up but
only coming back to its o w n foundation. But this instability may be too costly
to justify any advantages in promoting political stability i f the outcomes are
completely arbitrary and unrelated to individual preferences ( M i l l e r 1983,
744). M i l l e r d i d not resolve this issue, but suggested (1983, 745) that "most o f
us w o u l d view political outcomes in the real w o r l d o f pluralism as consider­
ably unpredictable but clearly confined w i t h i n certain bounds o f 'political
feasibility.'" A chaotic aggregation process w o u l d certainly support M i l l e r ' s
conjecture. A social choice mapping in the chaotic domain w o u l d be empiri­
cally unpredictable. This is one o f the characteristics o f chaos and it arises
from the sensitive dependence on initial conditions. Second, the chaotic func­
tion also does contain very definite bounds and restrictions—this is the order
in the complexity o f outcomes. Outcome and paths are highly constrained,
although the variety o f outcomes is endless, as evident in both the logistic
functions and the spatial voting example. The topological transitivity o f chaos
gives a variety o f outcomes that one w o u l d n ' t otherwise get without changing
preferences. Finally, outcomes are not arbitrary or completely unrelated to
public opinion. Just like the logistic function, preferences map to outcomes in
a strict deterministic fashion. However, in a chaotic process, one gets an
indeterminacy from a determinate mapping (Huckfeldt 1990). A deterministic
mapping (majority rule) from individual to group leads to indeterminate,
complex, evolutionary outcomes. This is quite unusual. Even i f one assumes
deterministic choice and behavior rules on the part o f individuals, one gets an
indeterminacy o f social outcomes! B y my interpretation o f M i l l e r , this is the
best combination for democratic political stability: constraints, bounds, and a
nonarbitrariness in the individual-group mapping, combined w i t h an un­
predictability and evolutionary variety o f outcomes over time.
From Individuals to Groups 115

Conclusion

The purpose o f this chapter was to reraise the question o f the "black b o x " o f
the individual-group connection. We know that this connection is not the
simple additive relation that early theorists once assumed. We know it is
"strange," as evidenced by complex outcomes, instability, and sensitivity. I n
this chapter I sought to demonstrate that this black box o f aggregation from
individual to group is i n the realm of chaos theory. The function that maps
preferences to outcomes and that maps outcomes to outcomes through demo­
cratic voting is chaotic—i.e., a unique type o f mapping w i t h a blend o f
disorder—in terms o f sensitivity, unpredictability, and evolutionary instability—
and o f o r d e r — i n terms o f strict mapping rules, regularity in regions o f insta­
bility, and bounds or constraints on outcomes and paths. I showed this w i t h
the example o f multidimensional spatial voting: the voting context w i t h the
most pervasive presence o f instability. However, I suspect this applies to most
aggregations, as I suggested by pointing to similar findings by Saari on voting
and by economists in terms o f aggregation through price mechanisms.
The normative question o f whether this is good or bad for liberalism,
populism, pluralism, or democracy i n general remains open. But it suggests
several future tasks that need to be completed. I have shown that multidimen­
sional spatial voting is a chaotic process. This only sets up the problem o f
understanding what is causing the chaos. M u c h work needs to be done on w h y
the social choice function is chaotic. A l s o , much work remains on specifying
the general conditions o f instability and ways to reduce the instability. But i f
we know what type o f problem we are dealing w i t h , then we can begin to
systematically understand the origins o f the instability. Only then can we
begin to open the black box o f the aggregation functions and to resolve the
normative debates i n democratic theory.

NOTES

Epigraph from The Social Contract and Discourse, Book 2, chapter 3. Quoted in
Frölich and Oppenheimer 1978,15.
1. Obviously Bentham's scheme raises problems of interpersonal comparison of
utility. This is a related issue of such magnitude that it will not be discussed here.
2. Occasionally I will use the term social choice function rather than social choice
correspondence nontechnically, as it is used throughout the voting literature, to desig-
nate a mapping from individual preferences to a social choice.
3. Sarkovskii's theorem (that a three-cycle implies the existence of cycles of all
lengths) applies to the logistic function since it is one-to-one and continuous. These
conditions do not hold for the social choice correspondence.
4. Completeness requires that all alternatives can be ranked.
116 Chaos Theory in the Social Sciences

5. However, Saari (1990) demonstrates that the Borda count reduces the number of
paradoxes.
6. O f course, this distribution is only one of many potential examples, but it
illustrates the presence of order.
7. The connection with fractals is suggested by the complex patterns in the distribu-
tion of outcomes.
8. This is the subject of a future research project.
Part 2
Chaos Theory and
Political Science
CHAPTER 6

Nonlinear Politics

Thad A. Brown

That chaos may be a part o f elementary politics is evidenced by highly


exploratory work i n the fields o f electoral behavior, game theory, axiomatic
choice theory, and conflict analysis. The social dynamical processes that may
induce chaos, methods o f investigating large-scale collective behavior, and
some implications for political science research are outlined in this essay.

Deterministic Chaos

At the turn o f the century, Poincare, and later Birkoff, noted that fully deter­
ministic dynamics do not necessarily provide explicit predictions on the evo­
lution o f a dynamical system (Poincare 1892; Birkhoff 1927). The subtlety o f
their observations was not fully appreciated until much later (Cartwright and
Littlewood 1945; Levinson 1949; Smale 1967; Ruelle and Takens 1971).
Today, chaos is a component in understanding dynamic macrobehavior in
many sciences.
Chaos exists when the long-term prediction o f a system is impossible
because uncertainty in a system's initial state grows exponentially fast over
time. Chaos fits the criterion that the autocorrelation function o f the time
signal goes to zero in finite time. Because trajectories are unstable, errors o f
estimation o f initial conditions or parameters, however small, can later accu­
mulate into substantial errors. This implies that forecasts o f future behavior
based on the past become problematic as current memory o f the past fails.
Nonlinearity is a necessary but not sufficient condition for the generation o f
chaotic m o t i o n . O f course this implies some feedback mechanism exists. The
observed chaotic behavior is due neither to external noise nor to an infinite
number o f degrees o f freedom. The source o f irregularity is the nonlinear
system's property o f separating initially close trajectories exponentially fast.
Because the time evolution is self-independent from its own past history,
predicting the long-term behavior o f chaotic systems is an interesting exer­
cise. The process does not come to rest i n a stable equilibrium but instead

119
120 Chaos Theory in the Social Sciences

comes to occupy large patches o f state space. Analysis is difficult because


stable and unstable states are strewn together in extremely complicated ways.
The critical feedback process central to the dynamics under investigation
therefore can be docile or unruly, depending solely on the tuning parameter
values o f the equation(s) o f motion. Indeed, for all practical purposes the
distinction between determinism and nondeterminism disappears. That is,
even equations that are solely deterministic become nondeterministic in their
long-term realizations.
One may assume it impossible to predict the long-term behavior o f
political systems that are potentially chaotic because their initial conditions
can be fixed only w i t h finite accuracy and subsequent error increases exponen­
tially fast. For politics, this means political actors w i t h i n a chaotic environ­
ment who are governed by the same laws and are identical in every measur­
able way may evolve differently over a long time. A g a i n , this would occur
despite identical initial configurations.
But this unpredictability o f the long-time behavior exists only at the level
of individual trajectories. A t the level o f statistical properties o f the time
evolution (averaged over different trajectories, say as they evolve from differ­
ent nearby initial conditions), very definite predictions are possible. Trajecto­
ries w i l l eventually move only on a small submanifold (chaotic or strange
attractor) o f the entire state space, w i t h predictable visitation frequencies o f
the different parts o f the attractor. It may be possible to estimate the time up to
a chaotic event, offering a short-term prognosis o f a political actor's trajec­
tory.
So far the theory o f chaotic dynamics has uncovered three major routes
for the onset o f chaos as the external system parameters are changed: the
Ruelle-Takens scenario via quasiperiodicity, Feigenbaum's period doubling,
and Pomeau-Manneville's intermittency route. Each has its o w n characteristic
signature (e.g., in the autocorrelation function). Such signatures may provide
important indicators for incipient changes. Other indicators may come from
the fact that, generically, a system develops characteristic large-amplitude
fluctuations as it approaches a phase transition. W i t h i n measured turbulent
flows (mostly found in the physical and life sciences), chaos is k n o w n to exist.
Thus, for any theory o f dynamical politics to be viable, it must come to grips
with political turbulence and hence incorporate chaos.
Gaps in our understanding o f political dynamics have become partic­
ularly obvious and acute during recent political events in Europe, Russia,
Japan, and the M i d d l e East. These events and rapid macropolitical changes
came as unexpectedly to any political scientist as the 1989 San Francisco
earthquake came unpredicted by any geologist, or the 1993 rains in the M i d ­
west by any meteorologist.
It is not that developments such as those in Russia or Eastern Europe
Nonlinear Politics 121

could have been predicted i n detail i f better models had been available; the
point is that despite the unpredictability o f such developments, there may be
definite patterns i n their occurrence. Furthermore, there may be well-defined
relations between the time scales on which short-term prediction is possible
and the characteristics o f chaotic evolution at long times; there may be indica­
tors o f incipient macropolitical changes; and there may be a wealth o f macro­
scopic variables that are predictable at long times even though the system state
is not. In the field o f nonlinear dynamic systems, such coexistence o f predict­
able and unpredictable properties has led to phrases like deterministic chaos
(Schuster 1988) and order in chaos (Prigogine and Stengers 1984).
A common example is weather forecasting. Here the solutions o f the
relevant (Lorenz) equations o f motion are so sensitive to small changes in
initial conditions that predictions beyond t w o or three days are intrinsically
impossible. In some places, like Los Angeles, the weather is pretty much the
same from day to day. On those days, the utterance o f any local weather
person w i l l be as correct in predicting the weather as forecasts o f the entire
U.S. National Weather Bureau. A n inversion layer produces smog; the end­
less summer continues. Then one day the predictions break d o w n . The mas­
sive computers around the world miss the formation o f a Pacific storm that
flattens the coast o f Baja, California. A t the same time, it is possible to predict
average temperature, amount o f rainfall, and so o n , for the month o f Novem­
ber in the year 2022 at any desired location for any given set o f external
parameters. The equations that govern weather can simultaneously behave
both ways; order shifts to disorder and back through the mathematical pas­
sages o f chaos.

Nonlinear Political Interactions

Politics, whether at the communal, state, national, or international level,


results from the interactions o f individuals. These individuals are members o f
respective political collectives; while some may be equipped with more power
and attain an elite status, even the elites are interdependent. Political leaders
base their decisions on interactions with their advisors; they form impressions
o f their advisors, who form impressions o f them. Relationships change as
these impressions and experiences inform subsequent interactions. Legislators
chat in the halls before key votes and discuss issues with staff, advisors, and
adversaries. A t times they also check with constituents back home. What
legislators say and do influences their standing in a legislative assembly and in
their districts. Their standing in each one influences the other. The process is
reflected down to the level o f the typical voter who decides, say, to support a
Republican or a Democratic candidate in an election, based on interactions
w i t h friends, family, or acquaintances in her local area. Even the individual
122 Chaos Theory in the Social Sciences

voter's thoughts and actions may be the product o f subtle self-interactions, as


in Elster's multiple self.
A useful event that illustrates the impact o f social dynamics might be
Black Monday: October 19, 1987. Might is not a trivial adverb: the most
widely accepted cause o f the market's collapse, programmed trading, has
been shown to have only a weak effect on market volatility (Grossman 1987),
and as a consequence we really don't know what happened.
Brock (1991) interprets the market collapse as being the product o f a
herd behavior induced by the attenuation o f normal information channels
available to traders. His parable goes as follows: In a market functioning
normally, traders regularly upgrade their information by calling on industry
and security analysts. Communication between the traders and industry an­
alysts is open. Prices are a function o f both past price performance and
volume, as well as expectations for future earnings and dividends. When a
negative shock hits the market, prices drop. Traders use the behavior o f other
traders, price, and information from analysts to decide their next move. I f the
negative shock is large enough, the information channels between traders and
analysts j a m — t o o many phone calls. A trader is left w i t h simply following
the herd; the only existing information is the movement o f prices and the
observed behavior o f other traders. As Brock suggests, a sell-off gets started
when a positive feedback loop emerges in an otherwise insipid market. The
individual traders' reward structures are likely to magnify this process into a
serious collapse (LeBaron 1989).
Banerjee (1992) further argues that herd behavior can occur where a
decision maker pays more heed to information others are perceived to have.
The reason is simple: the visible actions o f others may be based on informa­
tion that the decision maker lacks. Private information is overridden by
others' signals and the herd takes off. The event o f an economic landslide
suggests even catastrophic change, let alone small shifts in preference, can
result from group interactive effects.
Unfortunately, attempts to analyze complex political behavior often tend
to attribute observable cooperation or conflict to the individual. Behavior in
politically intricate situations is attributed to higher levels o f political informa­
t i o n , a deeper understanding o f the game's rules, or superior strategies, and o f
course lots o f information (or at least enough information so as not to be
duped). Nearly all published theory on the iterative prisoner's dilemma fits
into this category. What must be remembered and used to calculate strategies
includes: discount values, prior encounters, likely future encounters and asso­
ciated probabilities, w h o might use what piece o f information to figure out
what to do, the cost o f keeping information, and so on. O f course, nobody
thinks this way. A n d even i f anyone d i d , there is no guarantee that a success-
Nonlinear Politics 123

ful strategy at one time w i l l not be utterly unsuccessful at another. I n real


political interactions are we left to laws o f chance?
For decades we have seen glimpses o f the complex individual-group
nexus that defines politics. For instance, we know political beliefs are often
based on socially cued ideologies by proxy. That is, individuals may faithfully
reflect and follow over time the beliefs o f others who possess developed
abstract formulations o f politics (Campbell et a l . , 1960; Converse 1964). That
most voters lack much useful political information or fail to behave rationally
is rarely debated anymore (Ferejohn and K u k l i n s k i 1990). But the end o f such
debates doesn't mean we really understand the mechanism that induces even
the most basic forms o f political behavior. Turnout is one example. Rational
choice theorists still cannot understand why anyone would vote, given the
costs relative to proportional influence. Empirical electoral analysts cannot
figure out precisely w h y aggregate voting rates appear to be going d o w n — o r
are they going up this year? As a good friend in industry says, there is
something wrong w i t h this picture.
Social dynamics result from sets o f local interactions between group
members and their interactions with the environment. Logically, such signal­
ing fits w i t h i n a variety o f classical social psychological perspectives that
suggest individuals adapt to social environment (Asch 1951; Festinger 1957)
or form impressions about others based on behavioral experiences (Heider
1958). I n politics we know that decision makers modify their o w n behavior
and the influence o f social environments through self-selection (Schelling
1978; Finifter 1981), avoidance (MacKuen 1990), migration ( B r o w n 1988), or
a generalized contagion process (Huckfeldt and Sprague 1987, 1988, 1993).
Formally treating interactive political behavior w i t h i n massively diverse
collectives is tricky. Interactive behavior is peculiar in that it can neither be
predicted nor analyzed by observing sets o f individuals cross-sectionally, or
even the time series from a given individual or group. Social dynamics and the
concomitant social behavior cannot be reduced to individual behavior i n the
sense that isolated individuals cannot induce the variety and richness o f global
collective behavior prevalent in any political system. Social and political
behavior is by definition holistic and synergetic (Haken 1978, 1983) and
must be the product o f interacting individuals who can communicate and
modify their behavior as a consequence o f their interactions.
A n y time series is a rough statistical characterization o f a collective
process. A power spectral analysis alone, for instance, cannot decide the
dynamical rules that model the time series. Both a deterministic logistic map
w i t h m a x i m u m parameter values (Mayer-Kress and Haken 1981) and a
stochastic system can generate a white noise signal such that a flat spectrum
cannot tell w h i c h model is correct. Figuring out what drives the show for
124 Chaos Theory in the Social Sciences

interactive social behavior poses some interesting problems. I n political dy­


namics there are likely to be spatial and temporal phase transitions. For
instance, we know that transition from a pattern selection phase to fully
developed turbulence occurs via the intermittency (Kaneko 1989). Nonlinear
interactions w o u l d almost demand that abrupt and widespread events occur
unexpectedly. H o w to proceed is the question.
A m o n g the possible ways to investigate dynamical politics are simula­
tions o f intriguing spatial arrays with cellular automata. Cellular automata are
formal dynamical systems w i t h many discrete degrees o f freedom. The beauty
o f automata is that while the rules o f interaction are surprisingly simple,
complex nonlinearity can be induced by the iterative nature o f interaction. As
nonlinear systems, cellular automata can display the full array o f dynamics o f
any real, l i v i n g system, from fixed points to cyclical behavior to chaos. Can
political life be the product o f simple, primitive models? Cell-space models
have been used i n physics (Herrmann 1992), chemistry, and biology (Gutowitz
1991; Forrest 1991), and to a more limited degree in the economic and
social sciences to investigate the ecological structure o f behavior (von Hayek
1937; Schelling 1971, 1978; Cowen and M i l l e r 1990). In politics, they may
represent the potential to "program" the information available and used by
entities (voters, groups, elites, nations, or whatever), and hence what we can
discern from cellular automata may give us further insight into the informa­
tion dynamics in real politics.
Rules can reflect essential existential states o f being: love and hate. Love
attracts and induces assimilation; hate repels and induces egestion. Rules can
be deterministic or probabilistic. In both cases, cellular automata are formed
on a lattice. Cellular automata can clearly produce patterns in space and time
that can be m o v i n g fractal clusters i f the underlying dynamics are chaotic. We
can determine i f the underlying dynamics are chaotic by watching the spread
o f small changes in the configuration o f the system on a lattice.
A lattice is usually a (/-dimensional array o f discrete cells whose behavior
is governed by local, uniform, and time-independent rules. Boundary condi­
tions may be periodic. The rules determine the behavior o f individual cells as
they interact w i t h others at the next time step. Future states o f a cell are
normally a function o f the values o f neighboring cells and the cell itself. The
radius o f a cellular automata is the number o f neighbors whose state at time t
affects the state o f any cell at t + 1. The state o f the cell is a variable, s „
(where / = 1 , 2 , . . . , N), which can be any measurable political trait. I f s, = 0
we can say the voter is a Democrat; i f s = 1 the voter is a Republican. The state
t

o f the entire system, its configuration, is defined by the collection o f values


(«!, s , • • • , % ) . From these values, the average fraction o f partisans at any
2

time step, t,f(t) = ((\/N) X, .$,-), its variance, and so on, can be easily calcu­
lated, as well as the asymptotic fraction o f partisans based on a long period.
Nonlinear Politics 125

Although individuals remain in fixed physical locations, their partisan­


ship may change as a result o f interactions with other voters on the lattice.
This is a primordial political w o r l d , but well reflects many o f the central ideas
in contextual models o f politics. B y using political rules on a lattice, enough
interactions between decision makers can replicate real political situations.
Finding useful rules that map different forms o f political behavior is the key to
making automata theory viable for the social sciences.
A simple two-dimensional interaction law in which each individual inter­
acts in the same way with its neighbors is the deterministic majority-rule law.
This law makes each site adopt the value prevailing among the cell and its
four nearest neighbors on the lattice. The interaction rule for such electoral
behavior is given by

Gfait + 1), · - · , s (t N + 1» = l l
^ X Sj (t) > 3,

where G, are given functions o f time t and the current values s (t),. . . ,s (t); x N

2 _ 2
\i ~ j \ = V(*,- — xj) + (y, yj) is the Euclidean distance between cells i and
j , w i t h location (x y,) and (x yj), respectively. The summation is over all
t< jt

cells j whose distance from i is less than or equal to a, including s, and its four
nearest neighbors (fig. 6.1). Thus, the voter at cell / w i l l be Republican at time
t + 1 i f , among i and its four nearest neighbors, at least three are Republicans
at time /; otherwise, it w i l l be Democratic. Since the summation condition

Fig. 6 . 1 . A r e p r e s e n t a t i v e lattice site. /', a n d i t s n e i g h b o r s o n a square


lattice w i t h D = 2 a n d lattice c o n s t a n t a. (The c o o r d i n a t e s x , and / ,
d e s i g n a t e t h e physical l o c a t i o n o f site /.)
126 Chaos Theory in the Social Sciences

always examines five cells, no ties can occur and the interaction law is
symmetrical w i t h respect to partisan influence.
To see how the rule works consider i, starting from some given initial
state, 0, its four nearest neighbors, and their respective nearest neighbors,
over t w o time periods, as shown in figure 6.2. In the first time step, the
interaction o f the four neighbors o f i w i t h the peripheral cells leads to a
conversion o f the four neighbors into Republicans, leaving i Democratic. I n
the second time step, the central cell i also becomes Republican. As this
simple illustration shows, over time a voter is affected by voters outside the
range o f the defined radius o f the interaction law. Had we chosen the voter to
the Northwest o f / to be Democratic instead o f Republican at t = 0, then i
w o u l d have remained Democratic at t = 2. A more general statement would
be that the state s,(f) w i l l be influenced by the initial states s-(0) o f all cells j for
which the inequalities

bj - yt + ( J - *,·)! -
x T A

bt - (*,· - •*,·)! '-- ta

hold (r = 0, 1, 2, . . . ). Despite the simplicity o f this political interaction


rule and its explicitly local nature, it shows that the long-time behavior o f the
voter depends on the initial configuration o f the whole system. A more com­
plicated rule, or one w i t h global characteristics, would be at least as sensitive.
B r o w n and McBurnett (1992) have used a stochastic version o f the

f=0 f=1 f= 2

Fig. 6.2. First t w o s t e p s of t h e t i m e e v o l u t i o n of a r e s t r i c t e d s e g m e n t of


t h e l a t t i c e , c e n t e r e d a t s o m e s i t e /', under t h e d e t e r m i n i s t i c m a j o r i t y l a w .
T h e values in t h e circles d i s p l a y t h e p a r t i s a n s h i p o f t h e d i f f e r e n t s i t e s ;
the symbol * means that the state cannot be updated w i t h o u t addi-
t i o n a l k n o w l e d g e o f t h e i n i t i a l c o n f i g u r a t i o n b e y o n d w h a t is s h o w n f o r
t = 0.
Nonlinear Politics 127

above model to simulate multiple political realities. I n the model t w o oppos­


ing groups are arrayed on a 128 x 128 periodic two-dimensional lattice o f
initially randomly distributed sites. From such stochastic rules it is k n o w n that
highly ordered systems evolve from what are essentially random configura­
tions (Clifford and Sudbury 1973), though the specific evolved configurations
are themselves apparently random and are highly sensitive to initial conditions
(Durrett 1988). When a stochastic voting rule is applied to a random initial
configuration, the system behaves in a fashion that is completely dependent on
the initial concentration o f Democrats or Republicans. Small chance events
early in the history o f a political system (or a nested subsystem) may funda­
mentally alter the numerical power balance between or among political actors.
W h i l e the asymptotic state o f nearly all stochastic models such as that
described above is unanimity, the finite time path provides insights into inter­
esting political dynamics (Brown and McBurnett 1992; McBurnett and
B r o w n 1992). For instance, a natural clustering o f like voters emerges as an
outgrowth o f the interaction rule. Over time these clusters become a randomly
built political landscape. ( A time series describing this process is given in
fig. 6.3.)

15,000

J3
Q.
0 1
10,000

5,000
5,000 10,000 15,000 20,000
Time Path

Fig. 6.3. T i m e series of t h e n u m b e r o f Republican v o t e r s g o v e r n e d b y a


s t o c h a s t i c v o t e r r u l e as s i m u l a t e d o n a 124 x 124 square lattice w i t h a
p e r i o d i c b o u n d a r y . Initial c o n d i t i o n s are r a n d o m . T i m e s t e p s 1 t o 20,000
are s h o w n ( B r o w n a n d M c B u r n e t t 1992).
128 Chaos Theory in the Social Sciences

The steady gain o f one group comes at the expense o f the second. The
outcome o f which group w i l l dominate, however, is by no means predictable.
A t any point the fortunes o f a leading group can be reversed. A n equilibrium
is not predictable i n advance, in that the trajectory o f the individual voter
exists w i t h i n a highly irregular political environment. The time series from
these and other experimental data appears to have a fractal, noninteger dimen­
sion, though more experiments are needed to determine i f this is true ( B r o w n ,
Pfeifer, and McBurnett 1992; McBurnett and Krassa 1991).
Wolfram (1986) has offered the most complete classification o f cellular
automata. Cellular automata appear to be classifiable into four general types
of behavior: Class I : cellular automata evolve into a fixed, homogeneous state
and these structures evolve to fixed l i m i t points; Class I I : evolution to a simple
periodic structure represented by l i m i t cycles; Class I I I : chaotic, aperiodic
patterns as found when strange attractors are present; Class I V : complex
localized organizations as represented by systems that exhibit very long tran­
sients.
Distinguishable classes o f cellular automata related to politics can only
begin to tell us what the subtle, underlying structure might be. For instance,
electoral systems may represent an example o f Class I V behavior by exhibit­
ing very long transients, a great deal o f local order, and a global order that
may emerge. Bureaucracies, on the other hand, may be reflected by Class I
dynamics where both local and global scales o f order exist. Yet the statistical
behavior o f such systems would still be unknown. Simple classification o f
dynamics is not sufficient to give us greater insight into political and social
dynamics. While lists o f political systems (and problems) classified by dy­
namical behavior would be an improvement over current methods, these
classifications could not tell us precisely how the reality, process, and order­
ing o f numerous forms o f qualitatively differing systems came about or, more
importantly, how they w i l l evolve.
I f political dynamics are likely to be complex or chaotic, how would we
know? One set o f measures w o u l d be based on the properties o f the aggregate
time series o f the number o f partisans, as well as on their asymptotic fraction.
The correlation integral, the Lyapunov exponents, and spectral analysis are
now commonly used to evaluate the presence o f chaos (Peitgen, J ü r g e n s , and
Saupe 1992). O n a spatial lattice, for the trajectories to be chaotic, slight
perturbations in the initial conditions w i l l have to induce a changed configura­
tion in the values o f cells. H o w two trajectories quickly become different is
not the issue because the exact rule o f interaction is k n o w n . We must instead
estimate how rapidly and how far the trajectories w i l l diverge. The notion o f
chaos in politics may indeed include the speed w i t h which trajectories sepa­
rate. In order for the concept o f closeness o f trajectories to have meaning, a
measure o f just what fraction o f the voters are different for the initial and
Nonlinear Politics 129

perturbed configurations is needed (Herrmann 1992). The Hamming distance


measure provides a useful definition o f distance in phase space

A(0 = S |<x,.(0 - p,.(f)|,

where a (t) and p,(f) are two time dependent configurations in phase space and
t

N, as mentioned above, is the number o f cells indexed by <'. The dynamical


behavior w o u l d be chaotic when in the thermodynamic l i m i t A(t) gets big for
large times i f A(0) —» 0.
W h i l e the w o r k o f B r o w n and McBurnett suggests that even simple
political interaction models show a sensitivity to initial conditions, much
more analysis is needed. To determine an exact dynamical behavior requires
using in a simulation the same sequence o f random numbers applied to t w o
configurations o f a lattice. Meaningful outcomes can only be determined after
averaging over many initial configurations o f equal distance and across differ­
ent sequences o f random numbers. O f course, i n deterministic systems, once
a set o f configurations is identical they w i l l stay identical. Once knowledge o f
the behavior o f cellular automata models is gained, analysts w i l l be i n position
to calibrate lattice rules to match both time signatures and spatial movements
that match some categories o f interesting political mass behavior.

Strategic Interaction

Models o f social interaction investigated by means o f cellular automata may


open up for political science issues as yet unapproachable. We know that
w i t h i n strategies for solving common collective choice problems, cooperation
(or conflict) may appear to be stable or even frozen i n place while it can also
instantly disappear with the slightest nudge or noise. Cellular automata can
model conditions where a large number o f actors make many decisions, as in
any complex social community where individuals interact with imperfect
information. However, knowing the rules that are important to investigate
requires an adept touch at this point, because w i t h i n the structure o f automata
k2r+
theory, the number o f &-state, r-radius rules is rather large (k *). I t is
probably best to look to familiar ground.
Determining how levels o f cooperation or conflict are reached and
spatially diffuse is a core element in iterative prisoner's dilemma research
(Rapaport and Chamma 1965; Axelrod and Hamilton 1981; Axelrod 1984).
As is well k n o w n , the prisoner's dilemma is a two-person, zero-sum game
that can model the political evolution o f either cooperation or conflict. Each
130 Chaos Theory in the Social Sciences

player on every round can choose between cooperation (C) or defection ( D ) ,


and the outcome is determined by a standard payoff matrix, V:

C D
C R,R S,T
D T,S P,P

where T > R > P > S and 2R > T + S. In a one-shot play o f the game, each
player's best strategy is to defect, since D gets the largest payoff, T, while C
w o u l d get 5. The paradox is obvious. I n finitely many repeated plays o f the
game, all Cournot-Nash (noncooperative) equilibria have the property that no
matter how long the sequence, a noncooperative outcome occurs in each
period. I n games o f an indeterminate duration, however, the player's utility is
calculated on the expected long-run payoff o f repeated one-shot plays. In
essence, it is best to figure out how to cooperate, since i f both players use D
both end up w i t h a lower payoff since R > P.
Axelrod explored some spatial properties o f the game w i t h an eye on the
territorial stability o f tit-for-tat (Axelrod 1984). When the iterative prisoner's
dilemma was examined w i t h i n difficult environments, Richards (1990)
showed the clear potential for chaotic behavior. In exploring evolutionary
games played w i t h neighbors on a spatial lattice, the most common approach
is to set each lattice cell to a single strategy. The strategy may be either fixed
(Axelrod 1984) or allowed to evolve based on a genetic algorithm or other
optimization h i l l - c l i m b i n g routine (Axelrod 1987; M i l l e r 1989).
As before, the lattice mechanism is quite simple. A l l cells are updated
simultaneously; the score o f a cell is determined to be the average it obtains
playing the neighbors defined by the given radius; the score o f the cell is then
compared to the scores o f the neighborhood and the cell itself; the cell adopts
the strategy o f the highest score w i t h i n the sites defined by the radius. Ties,
when they occur, are usually broken by adding a touch o f noise to the scores.
As was observed above, the scores o f the defined neighbors are dependent on
the strategies o f their defined neighbors, and so on. Thus, the contextual
influence on any cell extends beyond the defined radius. When the model is
restricted to set strategies, it clearly qualifies as a cellular automata.
To illustrate, let's examine one set o f recent findings on the iterative
prisoner's dilemma. I n a paper, N o w a k and M a y (1992) use fixed strategies o f
cooperation (C) and defection ( D ) to study how cooperative behavior evolves
and is maintained. A s mentioned above, nearly all prior work demanded that
individuals or groups remember past encounters, with some discounting o f
retrospective evaluation for prospective payoffs. Such recollections are neces­
sary because o f the highly complicated strategies employed. Using fixed
strategies means that memory is not a requirement for behavior.
A'onlinear Politics 131

O n a two-dimensional lattice o f varying sizes, interacting players face


the following payoffs: R = 1, T = b (b > 1), S = P = 0. The parameter b
1
identifies the relative advantage o f defection versus cooperation. The entire
lattice is occupied w i t h C's or D's and again the scoring is the sum o f
encounters w i t h neighbors. A t the start o f a next generation (a time-step) each
cell is occupied by the player with the greatest score among the previous
owner and the defined neighbors. The radius and the boundary conditions are
2
varied. The rules among the n players on the n x n lattice are deterministic.
Updating o f sites occurs i n discrete time and in a synchronous fashion.
As reported by Nowak and M a y (1992), extremely complex patterns are
achieved even when strategies are fixed that in essence involve no strategies at
all. What they show is that ecological structure o f cooperation stems from the
magnitude o f the advantage given to defecting, the value o f b. For instance,
the initial configuration starts with 10 percent defectors when b is between
1.75 and 1.8, indicating the prevalence o f cooperation. When the b parameter
is moved to values between 1.8 and 2, the lattice evolves. Nowak and M a y
(1992, 827) interpret their results as indicating that C and D "coexist indefi­
nitely in a chaotically shifting balance, with the frequency o f C being (almost)
completely independent o f the initial conditions." This means that the asymp­
totic C fraction, f , roughly .318, remains constant for nearly all conditions
c

tested. The inference is that o f nonconvergence on defection commonly re­


ported in iterative prisoner's dilemma games.
Despite the assertion o f spatial chaos, the appearance o f complex spatial
patterns does not imply a Class I I I cellular automata. A g a i n , Class I I I implies
that disorder exists on both a local and global scale. In a game-theoretic
situation, as devised by Nowak and May (1992), strategies are fixed. Because
strategies are not available to each player (cell), the rigidity o f the game may
preclude deterministic chaos. Yet the difficulty o f characterizing the patterns
that emerge may well imply that the complexity found reflects the interesting
Class I V behavior.
In a recent paper, Huberman and Glance (1993) examined N o w a k and
May's findings relaxing the assumptions o f synchronous updates o f sites.
They report that the spatial complexity and richness o f evolved actors disap­
pear and the population moves quickly to converge on defection. Bersini and
Detours (1993) also report that asynchrony induces significant stability and
structure in cellular automata based models. Clearly, more analysis is needed.

A x i o m a t i c Theory

Since the seminal work o f A r r o w (1963), the indeterminate characteristic o f


politics in majority-rule settings has caused a search for political equilibrium.
Indeed, democratic theory once could only be understood with the proviso
that neither global nor local equilibrium could exist within even the simplest
132 Chaos Theory in the Social Sciences

multidimensional voting system. A n y outcome o f any majority-rule system


could cycle among alternatives at any time (McKelvey 1976, 1979; Schofield
1978; Cohen 1979). A n y status quo could be overthrown by some other
alternative drawn randomly from the entire issue space (Riker 1982). Even in
political settings packed w i t h elites, such as legislatures, decision making
w o u l d be in constant disarray; policy decisions w o u l d be utterly unpredict­
able. Governments based on populist democratic voting theory should be
without order or structure.
Human subject experimental work on the spatial properties o f voting
systems found outcomes clustered around a central tendency (Fiorina and
Plott 1978). I n experimental voting games where any outcome was feasible,
different groups converged on similar collective choices rather quickly. So
while the presence o f political disequilibrium was well k n o w n , the micro­
scopic causes and dynamics that might induce political order within multi­
dimensional voting environments or disorder were not.
Serious progress has been made. Using simulation techniques, Browne,
James, and M i l l e r ( 1 9 9 1 , 1992) have proven the link between cyclical behav­
ior under conditions o f majority rule. In related w o r k , Richards ( 1 9 9 1 , 1992)
has formally linked the L i and Yorke (1975) period-3 theorem to disequilib­
r i u m in multidimensional choice. Both efforts focus on the dynamical nature
o f decision makers and environments and the interaction that follows pro­
posals to alter the status quo.
In their initial w o r k , Browne, James, and M i l l e r (1991) confirmed that
the McKelvey-Schofield-Cohen models exhibited deterministic chaos. I t was
by no means clear that the "chaos" confirmed by axiomatic theories o f choice
was comparable to dynamical chaos. Using three-dimensional, noncoopera-
tive fixed ideal points in smooth and continuous policy space, they proved that
the position o f the status quo is iterated by decision makers attempting to
maximize expected utility based on their o w n preferred positions. Their re­
sulting analysis shows a Lyapunov-positive dynamical system (see Browne,
James, and M i l l e r 1991).
Richards (1991) employs the Li-Yorke period-3 cycle theorem to prove
the link between chaos and conditions in multidimensional choice under
conditions o f majority rule. The Li-Yorke theorem is an extension o f the
Sarkovskii theorem (1964), which states: I f f(x) has a point x that leads to a
cycle o f period TT, then it also has a point x' that leads to a T-cycle for every
T <— TT. Here, T and TT are elements o f the following sequence:

m
1<—2<—4<—8<— 1 6 . . . 2 . . . « -
m m m m
. . . 2 • 9 *- 2 • 7 <— 2 • 5 <- 2 • 3 . . . <—
2 2 2
. . . 22 · 9 «- 2 · 7 <- 2 · 5 <- 2 · 3 . . . <-
. . . 2 - 9 < - 2 - 7 < - 2 - 5 < - 2 - 3 . . . « -
... 9 ^ 7 ^ 5 « - 3 .
Nonlinear Politics 133

(As before, "<—" means "precedes.") I f / ( x ) has period-3, then it must include
all periods n when n is an arbitrary integer. This inclusion forf(x) must also
contain a chaotic region. B y definition, the emergence o f period-3 implies
1 2 3 3 2 1 2
three f r e q u e n c i e s : / , / , a n d / , wheref = f // (Falconer 1990). Period-3
is one example where chaos was predicted from a mathematical theorem but
could not be seen in earlier computer simulations (Mayer-Kress and Haken
1981).
Richards restates the Li-Yorke Theorem, sketches a proof, and links it to
the three-cycle for the majority-rule function. Written without the work o f
Browne, James, and M i l l e r (1991) at hand, Richards clearly delimits and
urges the analysis they pursue. That is, based on the presence o f a strange
attractor, Richards suggests that within the dynamical decision process, a new
solution set is likely to emerge. In a sequel, Richards (1992) extends the logic
o f the analysis considerably by linking the presence o f period-3 cycles to
multidimensional social choice situations i n which the core is empty. Without
testing numerically for the presence o f chaos, Richards indicates with a series
o f bifurcation diagrams the fractal quality o f the complex decision behavior
over increasingly smaller scales.
Unraveling the dynamics o f axiomatic theories under conditions o f ma­
j o r i t y rule has presented some interesting analytic problems. In their highly
innovative w o r k , James, Browne, and M i l l e r (1993) are reconsidering the
dynamics o f both individual decision makers and resultant system response.
Policy space is allowed to automatically initiate system preserving reactions
as a result o f changes in a decision maker's behavior.
Both proposals, to move the status quo and acceptance o f the change, are
allowed and indeed expected to have an effect on a policy space surface.
When a decision maker proposes or changes a policy position, the system
responds in an effort to automatically reestablish the previous alignment. This
systemic action may cause local disturbances, and the resulting ripple through
the adjacent regions o f policy space is allowed to produce a changed surface.
The decision maker, as a consequence, faces an altered policy environment at
the next iteration. These innovations allow for what we expect to be an
"institutional response o f entrenched interests against significant alterations o f
the system status quo. The expected consequence o f such resistance would be
a reordering o f positions in the vicinity o f a changed status quo that, at the
same time, preserves the general stability o f existing relationships" (Brown et
al. 1992, 13). The inability to preserve such stability could lead to greater
oscillations in support and to behaviors that may cause the system to go
extinct. Because systems do respond, such events are likely to be rare.
Ideal point dynamics are comparably developed. Choice theory assumes
that once a position is taken, it doesn't change. This assumption limits result­
ing cyclical behavior to a subregion o f policy space. James et al. (1993) relax
this constraint and allow the ideal points to move within policy space condi-
134 Chaos Theory in the Social Sciences

tional to the iterated status quo locations. As the status quo iterates in policy
space, the change i n the position o f the status quo indicates the degree o f
stress that must exist on the surface o f the policy space. These forces consti­
tute the sum o f the moments impacting a subregion in policy space; their
limits symbolize the critical values for phase transition.
A t the micro-level, James et al. (1993) provide a decision maker w i t h a
changed context w i t h i n w h i c h to reassess what might maximize his or her
utility. Hence, the movement o f the ideal point is not random in issue space
but dependent upon the change in the immediate local field o f forces, a very
old electoral behavior idea brought to bear on behavior in multidimensional
choice environments.
A t the onset o f cycling behavior, t w o things must occur i f the system is in
a chaotic state: (1) there is a rapid diminishing o f information on where the
next status quo is likely to be and (2) there is a dilemma as to what to do next.
A g a i n , Browne, James, and M i l l e r state: " . . . continuous cycling is likely to
produce strain in the region o f the policy space, and rather than accept this
state, actors (especially those whose ideal points are most distant from the
cycling location) have an incentive to break the cycle by relocating their ideal
points" (Browne et al. 1991b, 13). Implicitly, they are shifting the meaning o f
ideal points away from the strict rational-choice definition to a relative pre­
ferred policy position based on what it is currently possible to realize.
In any phase transition, for instance, any smoothed approximation, or
envelope, o f the volume occupied by area such as a status quo w i l l grow in
volume as it is moved. Unperturbed, the séparatrices divide oscillatory and
rotational trajectories o f the various contours in utility space. Utility space, o f
course, exists as a projection o f indifference functions. As the utility space is
perturbed, a stochastic layer could form in its vicinity. What Browne, James,
and M i l l e r show is that the region occupied by the stochastic layer contains
chaotic motion and holes that a stochastic trajectory cannot enter. This
stochastic layer, moreover, is not uniformly thick, and the topology o f the
augmented space is therefore assumed to be coarse-grained or irregular.

Chaos as Allegory

Maybe the revolution caused by quantum theory during the first half o f this
century swamped P o i n c a r é ' s w o r k . Maybe the nature o f nonlinearities would
have been impossible to see without the development o f high-end computers
and the graphical ability to visualize complex mathematics. But it probably
took serious applications or even potential applications to move developments
like chaos theory o f f the pile o f intellectual toys. Clearly, the characterization
of complex dynamics and chaos w i t h low degrees o f freedom have opened up
a scientific reality in chemistry, physics, and biology. But what is reasonable
Nonlinear Politics 135

for chemistry, physics, or neurobiology is more difficult to accept in the social


sciences.
It is difficult to find crisp illustrations o f chaos in politics. They are rare
precisely because politics is characteristically defined and affected by inter­
actions among individuals, between individuals and groups, among groups,
and so o n . Political reality is often shaped by larger metapolitical environ­
ments. Thus, the dynamics and interactions that are the source o f political
chaos may obscure its presence.
A science o f politics, like other quantitative sciences, may one day soon
begin to reconcile itself to the fact that not everything important to under­
standing politics is accessible to traditional methods o f data collection and
statistical analysis. Some might say it is unfair to expect prediction and
accurate forecasting because politics has been systematically investigated for
too brief a time. Compared to astronomy, in which observations have taken
place for thousands o f years, or even demography, for hundreds o f years, the
politics o f mass publics, elites, or even the behavior o f political business
cycles, have been studied only for decades. Some would suggest we have
been too quick to attempt to generalize based on data over a short period o f
time and a limited range o f political situations and contexts. Maybe. Others
might say that while a time series would be useful, what we really need is a
better understanding o f the formal nature o f political decision making. That
is, we have not reduced the individual and the logic o f the collective decision
to a small enough package o f variables to really understand the most basic
political motives and actions. But as we have already suggested, even formal
areas o f political theory such as collective choice are now under investigation
by nonlinear science.

Systems w i t h strong nonlinearities, which we showed above to exist in


relatively simple polities, and the possible chaos that results do present a
practical problem for continuing conventional techniques o f data analysis. I n
potentially chaotic systems, as several o f the chapters in this book show, the
observed variation is a mixture o f real noise or stochastic error and determin­
istic chaos. M a y (1989) says a problem o f nonlinear relations is separating
density-dependent noise from density-independent noise. I n some patches in a
fluctuating environment, nonlinear relations are sufficient to produce chaos.
The odiousness o f the w o r l d is also able to generate noise by environmental
stochasticity.
When or where chaos lurks in politics probably can only be k n o w n by
statistical prediction. The beauty o f recent developments in the theory o f
chaotic behavior is that they offer a wealth o f predictions, given that we know
the dynamics o f the system (equations o f motion). Research i n areas such as
cellular automata may begin to give us such equations for politics. Develop­
ments also provide powerful methods to extract the characteristics o f chaotic
136 Chaos Theory in the Social Sciences

dynamics (attractors, invariant measures, characteristic exponents) from data


when we do not k n o w the underlying dynamics.
A t a m i n i m u m , the theory o f chaos warns us o f the dangers o f extrapola­
tion from a political time series. But the questions raised (and yet un­
answered) are profound: Can we predict that a political system w i l l display
deterministic chaos? Can we link our current statistical methods and data to
the existence o f deterministic chaos? A n d does the existence o f deterministic
chaos imply that predictability in politics is impossible, or can the analyst
learn from it and refine predictive models to ultimately reflect control over
political processes?
Implicit and explicit assumptions exist about the political link between
the micro- and macro-levels. For instance, some believe that a direct, simple,
additive relationship exists between micro traits and macro manifestations. I n
the earlier work in electoral behavior, the level o f an individual's partisan
strength was assumed, when aggregated, to indicate the strength o f the party
system; the summation o f individual identification was assumed to give a
clear indication o f the standing vote o f the citizenry, and so on. Such simple
micro-macro links were thought to hold for a variety o f measurable variables:
ideology, trust in government, and partisanship. These measures indicated the
state o f the political system and gave the analyst a handle on the immediate
nature o f the electoral system, as well as an ability to assess the long-term
nature o f the entire political system.
Nonlinear theory involves an approach quite different from conventional
methods i n quantitative political science. Conceptually, it shifts the emphasis
in formal theory to investigating dynamical microscopic interactive behavior.
A t the macroscopic level, the emphasis is shifted from short-term forecasting
based on specific, complete initial data to long-term predictions based on
generic, incomplete data. Explicit theory relies upon elusive contextual vari­
ables like political climate or mood, level o f political stimulation, instability,
hegemony, or volatility. These variables are ill-defined as aggregation o f
individual attitudes but may be well defined in terms o f certain statistical
properties o f the system.

Beyond Chaos

The purpose o f theory is to make nature stand still when our backs are turned,
Einstein reportedly said. For political scientists, nature often laughs and
dances around behind us. Is the dance chaotic? In other disciplines, scientists
are finding that chaos is a normal part o f any process. Standing still—order,
stability, and continuity—is merely a special case w i t h i n a more complex
process.
Some political mechanics govern electoral realignment, equilibrium be-
Nonlinear Politics 137

tween parties, the collapse o f the Soviet Union or the return to order in
Lebanon, the occurrence o f a coup d'etat, and so on. While politics is a
complicated process, its complexity may even be understandable and manage­
able once political behavior can be precisely described. Yet despite the wealth
o f contemporary data and information, we really do not know how to pre­
cisely describe political behavior. We do not know the exact degrees o f free­
d o m or the equation(s) o f motion that describe (and maybe predict and con­
trol) political dynamics. We don't know which political phenomena are linear
and hence i f they can be Fourier transformed.
Do equations that describe the cyclical nature o f electoral politics also
describe other political dynamics? D o political dynamical equations contain
universal features similar to other social, economic, or biological dynamics?
To borrow and to twist a line from Feynman's Lectures on Physics, political
scientists cannot say whether something beyond these equations, like free w i l l
or personal liberty, is even needed. Recent discoveries in nonlinear dynamics
may help us to eventually discover political laws and their qualitative content.

N O T E S

1. Mutual cooperators therefore score 1; mutual defectors score 0; D scores


b against C , who receives 0. The reported results are not altered when P = e, when
0 < e « 0. T > R > P > S remains strictly satisfied.
3 2
2. Period-3 means that f (x ) 0 x
0 < f(x )
0 < f (x )-
0
CHAPTER 7

The Prediction of Unpredictability: Applications


of t h e N e w Paradigm o f Chaos i n Dynamical
Systems t o t h e Old Problem of t h e Stability
of a S y s t e m o f Hostile N a t i o n s

Alvin M. Saperstein

The social sciences have long tried to emulate the procedures and results o f
the physical sciences—the laws o f economic determinism o f Marx come
immediately to m i n d . To a large extent, the science o f international relations
has not been successful in this emulation; its capability o f predicting the
outcome o f international competition (war or peace?) has been very limited.
Perhaps this is because too many variables seem to be required, intuitively, to
describe the interactions between nations. Also, nations are composed o f
multitudes o f individually complex people, making it hard to believe that the
internation variables, representing their collective behaviors, can satisfy rela­
tively simple functional relationships. A n y easily expressible theory, relating
such variables, is bound to be incomplete and thus suspect, whether the theory
is expressed verbally or mathematically.
The physical sciences owe much o f their long-term successful hold on the
public imagination and pocketbook to their long record o f successful predic­
tions o f physical events and processes, predictions that have given rise to our
powerful modern technological society. In the process, the sciences have
learned that incomplete descriptions o f the systems being studied are still very
useful. O f the many variables evidently needed to describe a physical system,
some are major, more are minor, and often the latter can be ignored while still
producing very useful predictions about the system. For example, celestial
mechanics can give a very good predictive description o f planetary motion in
our solar system without including the effects o f the motions o f the various
moons about their respective planets. The minor variables only lead to impor­
tant effects in those regions o f the system where the incomplete model,
consisting only o f the major variables, is itself unstable. Successful detailed
predictions have usually been obtained by avoiding these unstable regions,

139
140 Chaos Theory in the Social Sciences

regions i n which small changes i n the (major) variables included in the model,
mirroring the effect o f the excluded (minor) variables, lead to major changes
in the system. Such instability, in a deterministic system, is now commonly
referred to as chaos.
Physics also deals w i t h multitudes o f complex entities, such as mole­
cules. When appropriate—that is, when the interest is in the collection, for
example, a gas, rather than in its individual constituent molecules—physics
successfully uses very few variables. These are often the result o f averaging
over the constituent multitudes, satisfying relatively simple functional rela­
tionships, such as the gas laws and the fluid flow equations (e.g., Kinetic
Theory, Fluid Physics; see Lerner and Trigg 1991). These laws are necessarily
incomplete and hence sometimes incapable o f leading to detailed predictions.
But we have learned how to predict their unpredictability and have found that
this incomplete knowledge is still useful as the model is extended to include
further variables, that is, to more completeness.
Following the lead o f physics, it w i l l be assumed in this chapter that
international relations may be productively modeled by relatively few vari­
ables, simply related; the choice o f variables depends upon the system o f
interest and the questions to be answered. It w i l l further be assumed that the
breakdown o f these incomplete models into chaos may be more useful than
the models themselves, that the qualitative prediction o f model instability may
be more meaningful than the quantitative predictions o f specific system varia­
tion.
Given these assumptions, a procedure is developed for exploring a num­
ber o f questions that are o f major concern to the understanding o f international
relations and the practical application o f this understanding to the creation o f
international security policy. The procedure is the creation o f simple mathe­
matical models for the interaction o f competing states and the examination
of the numerical output o f these models for regions o f stability and chaos
(Saperstein 1990). The general questions to be explored in this chapter are:
" W h i c h is more war-prone—a bipolar or a tripolar world?"; "Are democracies
more or less prone to war than autocracies?"; and " W h i c h is more war-
prone—a system o f shifting alliances or a collection o f go-it-alone states?"
(Further questions suitable for exploration with this procedure should become
apparent to the reader.) For example, it w i l l be shown that the region o f
instability in a tripolar w o r l d is larger than that o f the corresponding bipolar
w o r l d . From this it w i l l be concluded that a tripolar w o r l d is more prone to
war than a bipolar one.
Before applying these assumptions to the development o f models and an
analytical technique for getting useful information from these models, I w i l l
first try to justify them by attempting to formulate a general definition o f
science and prediction and then draw from this requirements for the more
The Prediction of Unpredictability 141

specific science o f international relations. A n important part o f this definition


w i l l be the relation between science and technology, a relation that has been
fundamental to the growth o f the natural sciences and their related technolo­
gies. The technology corresponding to the science o f international relations
should be the making o f successful international security policy. Following
the general discussion o f prediction and nonpredictability, a tool for recogniz­
ing chaotic regions is developed. The chapter concludes with the development
of a simple model for each question and the application o f the chaos tool to the
answering o f that question.

Science ( U n d e r s t a n d i n g ) and Technology (Control)

Many students in introductory courses in the sciences have great difficulty


coming to grips w i t h the notion o f science as a system o f understanding, not a
compilation o f facts or ideas about the system to be understood. Exacerbating
the problem is the difficulty o f compactly setting down what is universally
meant by "understanding" in the physical or social sciences. B y "understand­
i n g , " I mean the development o f a relationship between the phenomenon to be
understood and a previous set o f concepts already understood (even i f only
innately). These prior concepts, axioms for this latest level o f understanding,
are similarly related, in turn, to previously understood axioms, which, in turn,
are related to still more primitive understandings. Thus, science is a tower o f
relationships constantly growing at the top and the bottom.
By "relationship" between two levels o f concepts or phenomena, I mean
that one level can be derived from (caused by) the other level acting as
axioms. " D e r i v e d , " in turn, means that the one follows logically from the
other, arising by necessity from the structure o f the language being used; there
is a "functional relationship" between the t w o levels. For example, i f the prior
level consists o f the two statements: " a l l men are mortal" and "Socrates is a
man," then the derived statement is "Socrates is a mortal." More formally, i f
the system in question can be described by a variable (or set o f variables), x,
and i f x and x
n represent t w o neighboring levels, then we might be able to
n + l

write

*n+\ =/(*„. A
)> (!)

where / designates the functional relation that may depend upon a set o f
parameters A. I n physical science this might be the relationship between the
present configuration o f planets in the solar system and a previous configura­
tion. I n biology, it might be the relation between the behavior o f a cell in a salt
solution to that o f the same or similar cell in plain water. I f there is a time
sequence between the two l e v e l s — i f one occurs at a later real time than the
142 Chaos Theory in the Social Sciences

other—then the implied later level is determined by the former. The resulting
system o f understanding—henceforth also referred to as "the theory"—is
then a deterministic system. ( A n illustration o f a deterministic system w o u l d
be the sequence o f positions along the orbit o f an artillery shell in the absence
of air resistance. A nondeterministic system would be the sequence o f posi­
tions o f an electron in an atom, though the probabilities o f these positions are
well determined by quantum physics.) For example, suppose the theory deter­
mines the value o f some variable, x, at all future times, t, given the starting
value o f the variable, x (this could be the result o f iterating the relationship o f
0

equation 1 for all positive values o f n starting w i t h n = 0 ) ; the result o f the


theory is the curve x(t) shown (see fig. 7 . 1 , part a); x{t) is determined by JC 0

and the theory.


A check on the adequacy o f the understanding system (the theory) is
provided by observation o f the system being understood. Are the axioms
always observed to be accompanied by the derived concepts or phenomena
(logical equivalence)? Or are they always followed by the next level? In this
case, the consequences are predicted by the axioms. Such an empirical check
requires confirmation o f the validity o f the axioms in the system under consid­
eration, as w e l l as the observation o f their outcome. Hence, the prior l e v e l —
the initial state o f the system to be understood—must also be carefully
observed. Actual measurements o f the real world—either physical or social—
are always accompanied by noise, random fluctuations in the values that
would be obtained in a series o f identical measurements upon identical sys­
tems (e.g., "error analysis" in Lerner and Trigg 1991). Hence there is always
some uncertainty (also commonly called experimental error) i n the observed
system parameters or starting values, and hence some lack o f certainty in the
adequacy o f the system o f understanding. The result, therefore, is a range o f
possible starting values, 8x , and a consequent range in determined outcomes,
0

8x(t) (fig. 7 . 1 , part b).


The phenomena being investigated may not refer to definite events but to
the probability o f occurrence o f the events. For example, we may be investi­
gating the probability o f survival o f a bacterium as an antibiotic is added to its
growth medium. The events themselves may be random, nondeterministic,
even though their probability is determined by the relations o f the theory. I n
this case, single observations o f the system, no matter how precise, are not
sufficient. Repeated observations on an ensemble o f identical systems (theo­
retically an infinite number o f identical observations) are required to deter­
mine a probability (see, e.g., Feller 1950). For example, a single toss o f a
coin w i l l not tell you the probability o f a head for that coin. N o r w i l l the
observation o f a single human life span give any actuarial probability informa­
tion.
Thus a scientific theory relates some initial state (definite or a probability,
The Prediction of Unpredictability 143

Fig. 7 . 1 . T i m e e v o l u t i o n o f m o d e l s o l u t i o n s under v a r y i n g d y n a m i c s :
(a) s i n g l e s o l u t i o n o f m a t h e m a t i c a l m o d e l ( d e t e r m i n i s m ) ; (b) f l o w o f
s o l u t i o n s ( p r e d i c t a b i l i t y ) ; (c) f l o w o f s o l u t i o n s (chaos).

but observed somehow) o f the system to be understood to some other state(s)


o f the same system. Since the initial state is uncertain, to the extent o f the
observational noise encountered in empirically determining i t , knowledge o f
the related states may also be uncertain. I n linear theories, small changes in
the parameters describing the initial state lead to correspondingly small
changes in the related states. For example, i f the theory is described by
equation 1, then the small changes are similarly related:

x n + l + 8x„ +l =f(x„ + 8x„, A ) . (2)


144 Chaos Theory in the Social Sciences

Hence, given the theory, the parameters o f the determined states are as
w e l l k n o w n as those o f the initially given states (see fig. 7 . 1 , part b). It
follows that such a deterministic theory is also a predictive theory. The future
state o f the system is as well known as its present initial state. The ability o f
the physical sciences to make predictions, subsequently verified, has given
them a strong hold on the human imagination. A popularly observed eclipse
bolsters the public support o f astronomy, as well as o f all the other sciences,
presumably applying the same methods. (Naturally enough, other intellectual
and pseudointellectual disciplines thus strive to at least appear to be using the
same scientific approaches.)
In nonlinear theories, differences between initial states may not evolve
into allowed differences o f final states; that is, equation 2 is not valid. Hence,
small initial uncertainties about the system may lead to large final uncertain­
ties; little or nothing may be k n o w n about the outcome, no matter how precise
the knowledge o f the initial state. For example, i n figure 7 . 1 , part c the range
in outcomes determined by the theory is equal to the full range o f values
possible to the system; prediction may be impossible. I f a small change i n the
initial configuration o f a deterministic theory leads to large indeterminate
changes in the output configurations, the theory is said to be chaotic (Schuster
1988); it precludes precise prediction. A given theory may be chaotic for
some ranges o f input or system parameters and nonchaotic and predictive for
other possible conditions.
For example, the theory obtained by applying the usual Newtonian dy­
namical laws to the motion o f physical fluids is often referred to as the Navier-
Stokes equations. Given the initial state of, say, a body o f water—the velocity
of every particle o f the water at some given starting time—the solutions o f the
equations w i l l give the velocity o f each o f these particles at every subsequent
time. O f course, noise and measurement errors are associated with the deter­
mination o f the initial and system parameters (e.g., the configurations o f the
rocks i n the creek bed over which the water flows; the random fluctuations o f
the molecules making up the fluid around their mean, fluid, motion). Hence,
there w i l l be fluctuations in the results produced by the equations. I f the
output fluctuations are small, the motion o f the fluid is well predicted: k n o w l ­
edge o f the behavior o f the water at one place, at one time, gives knowledge
of the fluid flow at neighboring places at subsequent times. Such a nonchaotic
flow is termed laminar flow and is characteristic o f liquids flowing slowly in
smoothly varying channels. The same theory w i l l describe the motion o f the
same water flowing rapidly in a boulder-strewn river, a flow that we usually
describe as turbulent (see, e.g., Lerner and Trigg 1991). I n this chaotic
regime, knowledge o f what the water is doing at one time and place provides
almost no insight into what it w i l l be doing later at the same and neighboring
positions.
The Prediction of Unpredictability 145

The difference between the laminar and turbulent regimes o f the theory is
determined by the value o f a parameter, the Reynolds number, calculated
from the fluid speed and the size o f the obstructions to the fluid flow. For
values o f the Reynolds number less than a critical value, determined by the
theory, the flow is laminar; exceeding the critical Reynolds number implies
turbulence. The turbulent regime can further be divided into regimes of soft
chaos and hard chaos. I n the soft regime, the output fluctuations, though
large w i t h respect to the input parameter fluctuations, are small compared to
the extent o f the system. For example, there may be turbulent ripples on the
surface o f a body o f water even though the motion o f the bulk o f the fluid is
well predicted. I n hard chaos the fluctuations dominate the entire system, as
does the flow uncertainty i n a fully turbulent, chaotically tumbling and s w i r l ­
ing river rapids.
Although the flow o f water in a turbulent stream may not be predictable,
the onset o f turbulence itself may be predicted. The transition from laminar
flow to turbulent flow depends upon the Reynolds number, a parameter that
changes as the stream is followed. K n o w i n g how it changes, from a k n o w l ­
edge o f the configuration o f the stream bed and its sources, allows a prediction
as to where and when the stream w i l l change from laminar to turbulent flow
and vice versa. Similarly, the transition to and from turbulent air flow over an
aircraft's wings can be predicted as the shape o f the aircraft and its aerial
maneuvers are varied. Or, since a large storm may be construed as a turbulent
part o f the earth's atmosphere, the possibility o f such a storm may be pre­
dicted even i f the air flow w i t h i n such a storm remains chaotic and unpredict­
able. Hence, a theory o f a system that manifests considerable chaotic, and
therefore unknowable, regimes, but that allows the prediction o f where and
when these chaotic regimes w i l l start, still represents considerable under­
standing o f that system. Such theories may not have the presumably a l l -
powerful predictability o f celestial mechanics, which has caught the public
eye. But all natural systems may not be predictable (as we have learned in this
century), and one o f the characteristics o f successful science is to do the best
you can. The study o f such unknowable systems has become very popular
recently in the physical sciences (see, e.g., chaos in Lerner and Trigg 1991).
The possibility o f prediction implies the possibility o f deliberate control.
Control o f a process means that known intended consequences w i l l follow
from deliberate acts: prior conditions are modified to ensure desired outcomes
or to avoid undesired outcomes. For example, i f the outcome x„ +l in equation
1 is not desired, either the prior condition, x„, or the system parameters, A,
can be changed appropriately. ( I f a car is going in a dangerous direction, the
parameter A—the steering wheel position—can be changed or the system can
be started w i t h the car on a different highway.) I f prediction is not possible,
there is no way o f k n o w i n g the outcome o f a given act or policy, which is
146 Chaos Theory in the Social Sciences

synonymous w i t h saying control doesn't exist. Hence, technology, which is


the controlled manipulation o f some part o f the human environment, requires
predictability. A successful technology, one that produces anticipated ends in
a system from given input policies and means, is thus a confirmation (though
not a proof!) o f the understanding o f that system. Conversely, an unsuccessful
technology, one i n which anticipation is thwarted, is a definitive proof o f the
failure o f the underlying theoretical understanding. The failure o f a bridge or
an agricultural policy is an indication that the appropriate understanding o f the
behavior o f physical materials and environments or o f farmers is lacking. The
success o f the bridge or policy, however, does not guarantee the correctness o f
the theoretical understanding. (There may be alternative theories to the one
that maintains that guardian angels hold up the bridge and protect the crops.)
It follows that there is a symbiotic relationship between the physical and
biological sciences and engineering and medical technologies, a relationship
that has existed ever since these sciences have existed. Scientific advances
have led to new technologies that, in turn, "proved" the correctness o f the
scientific understanding and provided the tools for further measurements and
theory. This symbiotic relationship has led to the success o f both natural
science and technology and has served as a model for all other sciences.

The Science o f I n t e r n a t i o n a l Relations

Turning now to international relations, whose understanding is part o f the


political and social sciences, the corresponding technology is international
policy m a k i n g , w i t h its subset, international security policies. I f we truly
understand the international system, we should be able to make successful
policy, that is, policy whose desired outcome, given the existing system and
starting configurations, actually occurs (see, e.g., Saperstein 1990). I n the
past, a successful international security policy meant that a nation obtained
what it wanted, either without having to resort to war or by winning its wars
without unreasonable costs to itself. That is, the international system, o f
which the nation is a part, either doesn't change to the detriment o f the nation
or, at the very least, the component o f the system encompassing the nation
thrives even i f the rest o f the system is fundamentally changed. In the present
nuclear age, a war between major nuclear powers is unlikely to leave either
the w o r l d system or any o f its subcomponents intact. Hence, a major goal o f
any international policy has to be the avoidance o f such a war. Thus, the
symbiotic science must be able to predict the outbreak o f war.
I f that science could also predict the outcome o f a war, there would be no
need for the war. Nations often strive for overwhelming military strength in
the hope o f ensuring (predicting) their success in the event o f war. They have
often been unsuccessful in this prediction because o f the large uncertainties in
The Prediction of Unpredictability 147

the reactions o f their potential opponents. Considering war as a chaotic pro­


cess (see, e.g., Clausewitz 1982), like a turbulent flow over boulders, it may
be possible to predict its outbreak but not its outcome—that is, which side o f
the boulder gets a given portion o f churning water or which side wins the war.
A science o f international relations that only dealt w i t h probabilities
could lead to no policy technology and hence would be useless. Testing o f
probabilistic arguments requires an ensemble o f identical systems, whereas
we only have one such system—our real w o r l d . Thus, a scientific conclusion
that there is a likelihood o f nuclear war w i l l be treated as a prediction o f
certainty; the corresponding policy w i l l almost certainly be avoided. I n such a
situation, probabilistic predictions such as " l i k e l y " or "unlikely" (where is the
separation?) w i l l be treated by policymakers as "yes" or " n o , " which are the
results o f a definitive or deterministic science. Hence, we seek a science o f
international security with at least some deterministic aspects.
Can we postulate a system o f competing, hostile states to be determinis­
tic? Experience indicates that nations respond to one another's actions in
reasonably determined ways—at least for major actions and responses. I f
these patterns o f action and response didn't exist, there would be no political
science, and perhaps no history, o f international relations—just readings o f
tea leaves or goat's entrails! But why should we expect deterministic behavior
o f a collection o f nations, given that each one is composed o f many, many
individuals, each o f w h o m often acts in an unpredictable, chaotic manner? A n
answer—that deterministic behavior o f entities exists precisely because o f the
random nature o f large numbers o f their subentities—is obtained by analogy
to the physicist's description o f a gas. Such a gas is a collection o f an
enormous number o f randomly moving molecules whose precise description
requires a similarly enormous number o f stochastic variables. Yet its overall
physical behavior is deterministically described by a few gas variables, such
as pressure and temperature, each a complex average over the many molecular
variables. O n a gross scale o f observation these few gas variables usually
change smoothly, though on a very fine scale they exhibit random minor
fluctuations, mirroring the underlying random molecular variables. So we
might expect the behavior o f modern nations to be governed by a relatively
few deterministic variables, each a complex average over the behaviors o f
each o f its multitudinous population. The usual goal is to find rules governing
these variables, admitting that there w i l l be occasional minor fluctuations—a
love affair or assassination at the higher levels o f government or society—not
subject to these rules. M y additional goal is to see how a system, evolving
according to such rules, responds to such fluctuations.

The core o f international relations can be described as follows: Nation A


responds to the actions o f nation B in the context o f the other nations o f the
w o r l d system and its o w n present and past internal situations. Nation B then
148 Chaos Theory in the Social Sciences

responds to nation A's response in a similar context. Each response is deter­


mined by the system's previous responses (Richardson 1960a,b). Hence, the
situation is recursively deterministic. I f A and B designate the appropriate sets
o f variables describing the two nations, the situation might be represented in a
manner similar to that o f equation 1 :

n+l = a(A„, B „ , A)
(3)
B, = b(A , B„, A),
n

where a and b are functional relationships representing the responses o f the


two nations; the parameters À may also represent some average over the other
nations in the w o r l d system. For example, A„ might represent the hostility (or
the exports) o f A to B in year n, w i t h B similarly defined. Then equation 3
n

describes how next year's hostility (or exports) o f A to B w i l l be determined


by this year's hostility (or exports) o f A to B and o f B to A . The role o f the
political scientists (and the need o f the policymakers) is to attempt to predict
the outcome o f the intertwining chain o f responses on the basis o f their
understanding o f the system and its history. Usually this understanding is
verbal, expressed in the words o f historical and policy treatises. ( I n equation
3, the relationships a and b and the parameters À would be verbally ex­
pressed.)
Verbal understanding possesses the vividness, vagueness, redundancy,
and contradictions o f ordinary, daily, spoken and written language. It usually
contains many implicit assumptions and innate biases. The logic taking it from
premise to conclusions is often incomplete and/or fuzzy. It is often difficult to
know what the premises are and how well they are verified, what the conclu­
sions are, and how we got from one to the other (e.g., the functions a and b in
equation 3 are not explicit). Hence, how do you test the model against reality?
In contrast, mathematical formalisms tend to be explicit in their premises,
transparent i n their logic, and concrete in their conclusions. There is thus an
often-expressed desire to describe all aspects o f the macroscopic interactive
behavior o f nations via explicit mathematical recursion relations between
numerical variables representing those behaviors. Equation 3 becomes a set o f
mathematical relations between numerically defined variables: A, B, A, and so
on. I f the resultant equations could be solved, given all the necessary input
representing system parameters and starting conditions, they would produce
concrete predictions that could be tested against observed behavior to confirm
or negate the mathematical model.
However, given the complexity o f the real-world system, we are unlikely
to ever have all the necessary input information, or sufficient understanding,
to create a model system o f equations complete enough to really represent the
w o r l d in all o f its manifestations. A n y practical model would have fewer
The Prediction of Unpredictability 149

variables than are really required. The resultant equations, though far short
of the complexity o f the w o r l d , w o u l d still be complicated and nonlinear i f
they are to represent some o f the essential attributes o f our strongly interacting
real w o r l d , ruling out the possibility o f simple, explicit solutions (see, e.g.,
Prigogine and Stenger 1984). These obvious shortcomings o f any mathemati­
cal model are no different than the difficulties o f any reasonable verbal under­
standing o f the w o r l d . The difference is that the incompleteness o f the mathe­
matical model is usually immediately obvious, whereas that o f verbal models
is concealed in their fuzzy verbosity. In both cases, the resultant predictions
are necessarily incomplete, making theoretical testing and practical policy
formulation difficult.
The more relevant factors left out o f a model or incompletely described
w i t h i n the model, the more gross and incomplete its consequent predictions
w i l l be. Overall system variations may be correctly predicted though the
detailed outcomes may not be trustworthy. Such incomplete predictions may
or may not be useful, depending upon the circumstance o f the policy making
or theory building. It becomes important to determine what kind o f gross
predictions, from necessarily incomplete theoretical models, can be useful in
a given arena o f human experience. I t is the contention o f this chapter that the
prediction o f unpredictability can be very useful i n attempts to understand and
control the advent o f war in the w o r l d system o f nations (Saperstein 1984).

The Prediction o f Chaos a n d t h e O u t b r e a k of War

Crisis instability in the international system usually implies a configuration i n


which small insults can lead to major changes—the loss o f a nail leads to the
loss o f a shoe . . . to the loss o f a kingdom; an assassination o f a minor duke
can lead to the deaths o f millions and the profound transformation o f the
w o r l d system. Such instability represents the loss o f control and the great
potential o f war. This parallels the definition o f chaos (Schuster 1988): small
disturbances o f a deterministic mathematical system lead to dispropor­
tionately large changes in the system and the consequent loss o f control.
Prediction o f unpredictability i n a system is a prediction o f the onset o f
chaos—soft or hard. The range between soft and hard chaos among nations is
the range between minor and major loss o f control i n international relations. I t
is the range between the possibility o f loss o f an unspecified soldier i n a given
battle, the w i n n i n g or losing o f that battle, o f the war, and o f the existent
w o r l d system. We postulate that the presence o f hard chaos, in a theoretical
system modeling the international system o f competing nations, is a represen­
tation o f major crisis instability and o f the extreme likelihood o f the outbreak
of major war in the real w o r l d being modeled. I f the prediction o f hard chaos
is believable, then the international security policies associated with that part
150 Chaos Theory in the Social Sciences

o f the model leading to the onset o f chaos should be avoided. A useful


analogy is to the testing o f a new aircraft to determine its behavior under a
wide variety o f desirable and undesirable circumstances. I f the theoretical
model (either purely mathematical or based upon w i n d tunnel modeling)
indicates that certain maneuvers are likely to lead to loss o f control and
possible loss o f plane and crew, the pilots w i l l be instructed to avoid those
maneuvers—even i f they w o u l d lead to otherwise desirable results. Even i f
the model is not sufficiently complete to indicate how the aircraft w i l l behave
in the danger zone but is believable in predicting the existence o f the loss-of-
control zone, that w o u l d be sufficient to mandate changing policy so as to stay
away from that zone. Thus, i f the prediction o f unpredictability i n an inter­
national system is believable, the ability to make such predictions, even i f
incomplete i n details, can be used to answer political science questions as w e l l
as to help determine practical national security policy.
Thus we must face the question: is the prediction o f unpredictability
believable in an incomplete model? I f the onset o f chaos portends the outbreak
of war, can we believe a model's prediction o f war even though the extent and
outcome o f the war are not predicted and may not even be describable by the
model? Is the prediction o f instability (or stability) o f a model itself stable i n
the face o f expansions o f the model to include more relevant aspects o f the
system? W i l l a model's prediction o f chaos be softened or disappear as more
variables and/or more complicated interactions are added to the model?
A formal answer to the above question requires a working knowledge o f
the complete model, knowledge that doesn't exist. ( I f it d i d , the question
would be irrelevant!) In its place, we must rely on analogy with other
systems—usually from mathematics or physics—where complete models,
and their incomplete component models, exist and are used. Mathematical
experience w i t h dynamical systems indicates that "chaos first appears in the
neighborhood o f non-linear resonances" (Reichl 1992, 14), where new vari­
ables first make their impact upon the system. There is no experience o f such
chaotic regions disappearing as new variables are introduced. Empirical expe­
rience with fluids indicates that chaos (turbulence) appears earlier and stron­
ger when new variables, such as temperature differences and heat flows,
become important in the system. Theoretical experience with specific mathe­
matical models o f real phenomena (the Navier-Stokes equation for fluid flows;
the recursive equations' modeling o f the evolution o f tripolar systems from
bipolar ones [Saperstein 1991], developed later i n this chapter) suggests that
the regions o f stability (areas o f absence o f chaos) decrease in extent when
additional variables come into play.
We thus presume that qualitative (gross) predictions o f the loss o f model
stability are much more believable (more stable) than are the quantitative
(detailed) predictions from the same model. Hence, the prediction o f hard
The Prediction of Unpredictability 151

chaos in a model reasonably representing the international system o f compet­


ing nations is a fair warning to policymakers: embarkation upon the modeled
course o f action is to be done with extreme dread and care. However, a
contrary prediction o f model stability is not an absolute assurance o f system
stability; complacent acceptance o f the corresponding policy is not warranted
since a more faithful, more complete model o f the system in which the policy
is being applied may very well allow chaotic breakdown. Similarly, when
exploring theoretical questions about the international system, the appearance
o f chaos in an appropriate model portends the breakdown o f stability in the
system; however, stability o f the model does not necessarily imply a corre­
sponding stability i n the system.
The stability o f a mathematical system can be tested directly i f sufficient
computing power is available. For example, a large number o f possible differ­
ent but neighboring initial configurations o f the system o f interest may be
specified. Each one o f these starting configurations gives rise, via the mathe­
matical model, to a final configuration (the system at some specified future
time). I f these predicted final states are as similar to each other as the initial
states, the system is stable. I f they are w i l d l y divergent, as compared to the
starting values, the system is chaotic. As an illustration, the S D I system
proposed by President Reagan can be described by variables representing the
number o f offensive missiles available to the opponents, the number o f defen­
sive antimissile weapons, and the number o f anti-anti-missile weapons (also
offensive) (Saperstein and Mayer-Kress 1988). The model consists o f rela­
tions between these variables specifying how each o f the opponent parties
procures weapons in response to the analogous procurements o f the other
nations in the system—a typical arms race. The initial configuration is spe­
cified by the initial numbers o f the different weapons stocks and some numeri­
cal estimates o f their capabilities: neighboring starting configurations imply
slightly different values for these initial numbers. (In the real w o r l d , these
numbers and parameters would not be precisely known to any o f the compet­
ing parties.) The computer then tracks the evolution with time o f the system
emanating from each starting configuration and displays the variations in
the possible outcomes—the numbers o f the different types o f weapons i n the
stocks o f each o f the opponents. I f the final variations are comparable to the
corresponding initial variations, the system is stable. I f the final variations are
large compared to the initial ones, but still small enough that system d o m i ­
nance (superior and inferior weapons stocks) can be distinguished, the system
displays soft chaos. I f the final outcomes vary so widely that they encompass
the entire variation possible to the system and obscure all differences between
the variables corresponding to the different nations, then no prediction o f the
outcome o f the arms race is possible—we have hard chaos.

A more modest approach, suitable for systems with fewer variables, is


152 Chaos Theory in the Social Sciences

the calculation o f Lyapunov coefficients for the model in question. These


coefficients are direct measures o f the rate at which initially neighboring
configurations drift apart as the model system evolves. Each possible starting
point traces out a path toward the future by means o f the model mechanism, as
is illustrated in figure 7 . 1 . I f the paths from closely neighboring starting points
remain close (fig. 7.1b), prediction is possible; i f the paths separate exponen­
tially (fig. 7.1c), so that the final outcomes cover all possibilities allowed in
the system, everything is possible in the future, meaning nothing is k n o w n
about the future, and prediction is impossible—the situation is chaotic. The
Lyapunov coefficient (see, e.g., Schuster 1988) is a useful measure o f such
path separation that is easily calculated given a computer able to follow the
evolution o f neighboring paths. Let X and XQ be any two possible neighbor­
0

ing starting configurations, separated by a small distance 8 . I n the course o f


0

time n these evolve respectively (e.g., via equation 1) to the t w o configura­


tions X and X ' , separated by the distance 8„, where
N N

= 8 0 *"«*o>. (4)

This defines the Lyapunov coefficient Ç(X ), which depends upon the starting
0

configuration as w e l l as the dynamics o f the system. Since the interest is in the


long-time evolution o f the system, it follows that

limit limit 1
log x: (5)
S&o) =
°° fir, 0 n
I f £ < 0, two configurations starting close to each other w i l l remain close to
each other. Thus, predictability is possible (fig. 7.1b), so such a system w i l l
be defined as stable. I f £ > 0, the initially close configurations w i l l drift ever
further apart, making prediction impossible. Thus £ > 0 is the signature o f
chaos or instability.
In this chapter, several aspects o f international arms competitions, nor­
mally modeled qualitatively and verbally, w i l l be put into equivalent algebraic
forms. Only algebra is used! The variable expressing the arms level or pro­
curement level o f one country in one year is expressed in terms o f similar
variables describing the status o f that country and o f its competitor nations in
the prior year (e.g., equation 3). These expressions only involve a small
number o f additions, subtractions, multiplications, and divisions. Hence, they
can be put into standard spreadsheets on a desktop computer and iterated from
year to year to generate the time evolution o f the armaments levels o f the
competing nations. Thus, the future o f these levels, and o f the arms competi­
tion among these states, is determined. Similarly, the corresponding
Lyapunov coefficients, describing how the uncertainty associated w i t h these
levels, given that there are initial uncertainties, evolves with time, are also
The Prediction of Unpredictability 153

computed. O f course, one doesn't reach limits o f infinity or zero, as is


required by the definition in equation 5, on a desktop computer. Instead, the
iterations are run up to large values o f n ( i . e . , 100) for varying small values
of 8 .0

The prediction by a mathematical model is meaningful only i f the results


are stable, that is, i f small changes in the input and/or model parameters lead
to corresponding small changes in the predicted time evolutions. Otherwise,
the model's output is chaotic and indicative that the system being modeled is
crisis-unstable and likely to undergo a phase transition from a competitive,
but nonshooting, cold war to a hot war. Since the models are necessarily
incomplete, their detailed predictions are not really believable, even i f mean­
ingful within the context o f the model itself. But, as we have presumed
before, the prediction o f chaos is itself meaningful and believable. Using the
desktop computer, the models are explored over different ranges o f parame­
ters, or different algebraic forms, to see which regions lead to stable solutions,
which to chaos. Since different parameter ranges or algebraic forms represent
different policy choices, we are thus able to predict which policies w i l l be
dangerous, in that they may lead to crisis-unstable situations.
We have briefly described one approach that can lead to specific policy
choices (e.g., S . D . I , should not be developed i f it w i l l lead to a decrease i n
the range o f stability o f the international offensive nuclear warhead missile
system). I n the remainder o f this chapter the chaos approach w i l l be applied to
three political theory questions: Are bipolar international systems more or less
stable than corresponding tripolar systems? Is a system o f democratic nations
more or less stable than corresponding systems o f autocratic states? Is a
system o f nations that strives for a balance o f power via shifting coalitions o f
states more or less stable than one in which each nation individually seeks to
balance the power o f the others?
There is a great body o f literature exploring the answers to these ques­
tions via qualitative verbal analysis, and no claims are made for new or
surprising answers. However, since all o f these theoretical approaches, quali­
tative and quantitative, are incomplete and hence not completely convincing,
a juxtaposition o f different approaches that lead to similar answers should add
to our understanding o f these questions and their applicability. A disagree­
ment between the results o f the different approaches should be a signal that
understanding is still lacking.

W h i c h Is M o r e W a r - P r o n e — A Bipolar
or a Tripolar W o r l d ?

Is a bipolar w o r l d more or less stable than a tripolar world? As we leave a


world configuration dangerously dominated by two nuclear superpowers for a
w o r l d perhaps dominated by none, one, or many, it is important to gain
154 Chaos Theory in the Social Sciences

insight into whether w e must be even more cautious or can be a little less
t i m i d in addressing the many nonmilitary security problems o f a twenty-first-
century w o r l d . One approach to this question is to compare the range o f stability
o f a two-power system w i t h that o f a corresponding three-power w o r l d .
The variable chosen to describe national behavior in the interactive
model o f equation 3 is the devotion o f a given nation to war preparation—the
ratio o f arms procurement ( i n the most general sense) i n a given year to
the gross national product o f that nation i n the same year. O f necessity, the
numerical value o f this variable must lie between zero and one. The bipolar
w o r l d is then heuristically modeled by the simple nonlinear Richardson-type
model represented by the coupled equations below

X
n+l = 4ar„(l - Y ) n (6)
Y n+l = 4bX (\ - X„).
n

X is the devotion o f nation X to war in year n. The procurement o f arms


n

by X in the year n + 1 is in response to, and assumed to be proportional to, the


previous year's devotion o f its opponent, nation Y. The nonlinear term (1 —
Y„) expresses the assumption that i f the opponent's resources are stretched to
the breaking point ( i . e . , his previous year's procurements were so great that
they almost preempted his entire G N P , making i t impossible for h i m to
procure any more), there is no need to stretch any further this year, so that this
year's procurements may be correspondingly diminished. This model is so
simple that its region o f stability may be analytically computed (Saperstein
1984) as a function o f the model's proportionality parameters, a and b, which
must also lie between zero and one so that the X , Y„ remain between these
n

bounds. The resultant curve represents the critical relation between a and b;
the region above the curve, in which the two Lyapunov coefficients, £(X ) and 0

£(Y ), each defined as in equation 5, are positive, is the model's chaotic


0

region.
In equation 7 the model is extended to three nations (when e = 0,
equation 7 reduces to the previous two-nation model, equation 6):

X n+1 = 4aY„(l - Y ) + 4 e Z „ ( l - Z„)


n

Y n+l = 4bX„(l - X„) + 4eCZ„(l - Z„) (7)

Z„ + 1 = 4e[X„(l - X„) + CY„(l - /„)].

As e increases from zero, the coupling between X and Y and the third
nation, Z , becomes increasingly significant. Numerical computations o f the
three Lyapunov coefficients pertinent to this model, £(X ), £(Y ), and £(Z ),
0 0 0

made using simple spreadsheets on a small desktop computer (Saperstein


The Prediction of Unpredictability 155

1991), indicate that the stability region decreases in area as e increases, that
is, as the third nation becomes more significant in the w o r l d system. Hence,
the model suggests, in accord w i t h a great deal o f other scholarly evidence,
that a tripolar w o r l d is more dangerous than a bipolar one.

A r e Democracies M o r e or Less Prone t o War?

Given the propensity o f many o f the world's major democracies to aid and
abet less-than-democratic governments elsewhere, it would be useful to know
i f such policies hinder or further the w o r l d peace that these democracies
presumably seek and guard. Required first is a definition o f democracy that is
easy to quantify and fit into a model o f competitive arms procurement, such as
that o f equation 6, suitably generalized.
It seems easiest to define democracy in terms o f diffuseness o f decision
making. The larger the fraction o f the population that has significant input into
matters o f peace and war, the more democratic the nation w i l l be considered
to be. W i t h this definition, the numerical difference between democracy and
autocracy w i l l be based upon population-sampling theory. I assume that every
citizen o f nation 1 has some intrinsic feeling for the totality o f its opponent
nation 2, and conversely, ranging from complete confidence to extreme fear
and loathing. Let a, be the "fear and loathing" coefficient for any one citizen
of nation 1. Then the policy o f nation 1 toward its opponent w i l l be deter­
mined by (a,), the mean o f a, over the class o f citizens o f 1 who are significant
decision makers; (a ) is similarly defined. The fraction o f the citizens belong­
2

ing to this class w i l l be small in an autocracy, large in a democracy.


Assume that there is a natural spread o f values o f the fear and loathing
coefficient whose distribution is characteristic o f humanity, not nationality.
Each large nation w i l l contain many people manifesting small values o f the
coefficient, intermediate values, and so on. Thus the mean, a, o f this coeffi­
cient, taken over an entire large population, is a human characteristic, inde­
pendent o f nation: d , = a = a. (Note that the known examples o f extreme
2

interethnic hostility usually occur in comparatively small groups.)


The decision makers o f a nation are a sample o f the entire nation. I f it is a
very large sample, means over this sample w i l l be close in value to means
over the entire population. Hence, for large democracies, (a) = d. Means over
small samples may vary significantly from the means over the entire popula­
tion. Thus, given a random set o f autocratic nations, there w i l l be some for
which (a) greatly exceeds a and some for which (a) is much less than d . I n a
corresponding set o f democratic nations, all w i l l have intermediate values o f
{a) = d. Therefore, i f is the effective fear and loathing coefficient o f nation
X for nation Y , i f a l l nations i n the set containing X and Y are large democra­
cies, = d , whereas i f the set contains autocracies, can range from very
small values to very large values.
156 Chaos Theory in the Social Sciences

The effective fear and loathing coefficient is taken to be the propor­


tionality constant i n a nonlinear Richardson-type arms procurement recursion
relation. I n a competitive situation between nations X and Y, the annual
procurement o f arms by X , A X , w i l l be proportional to the amount by which
X's arms stocks were exceeded by Y in the previous year: Y — X . (The n n

proportionality constant w i l l be taken to be a^; the similar coefficient for Y's


annual arms buildup is a .) Procurement w i l l also increase proportionally to
yx

the total strength o f Y : Y„; i f Y is very weak, i t w o n ' t matter i f Y is stronger


than X ; the more powerful Y is, the more damaging any discrepancy in power
w i l l be i n any potential conflict. However, arms buildups cannot continue
indefinitely. There w i l l be economic constraints; nations can't procure more
than the total economy w i l l allow. I f C is the m a x i m u m annual expenditure x

for arms possible by X , assume a smoothed economic cutoff function

0 ( 1 - AX/C ) X = (1 - AX/C ) X 0 ( 1 - AX/CJ, (8)

where 0 is the unit step function (0(TJ) = 1 for 17 > 1, 0 = 0 otherwise).


I f X ' s arms stocks exceed those o f Y , then X can decrease his arma­
ments. Assume that the arms build-down in any one year (AX negative) w i l l
be proportional to the amount by which X exceeds Y ' s strength in the previous
year and also proportional to the total power o f X : the more confident a nation
is in its strength, the more it can afford to build d o w n . The proportionality
parameter for build-down is the confidence coefficient, w h i c h , for simplicity,
is taken to be the inverse o f the build-up fear and loathing coefficient. Finally,
there are no economic constraints for build-down. (It is assumed that recon­
version o f the military-industrial complex works!)
The result o f combining the coupled arms buildups and build-downs is
the following set o f recursion relations, relating the total arms stocks o f X and
Y i n any year to the corresponding stocks in the previous year:

x„ +i =x„ + Y„(Y„
axy - x ) 0 (1 - [x
n n+l - x yc ) 0
n x (Y n

~^-X (X n n - Y„)@(X -Y ) n n
u
xy (9)

Y n+l =Y n + a X (X yx n n - Y„) 0 (1 - [Y„+l - Y ]IC )


n y 0 (X n

- Y {Y n n - X„) 0 (Y„ - X ).n

"xy

These relations may be iterated, given arbitrary starting values X and Y 0 0

at n = 0, again using standard spreadsheet routines on a desktop computer.


Going on to large n (n = 100), the t w o Lyapunov coefficients can be calcu­
lated from equation 5. Typical results (Saperstein 1992a) are given in figure
The Prediction of Unpredictability 157

x<0) = y(0) = 0.3


epsilon = 0.001

20

V
0 10 20 30 40
a

Fig. 7.2. T r a n s i t i o n t o e s c a l a t i n g a r m s race

7.2, where the region above the curve is the region o f instability. Note that
small values for both a = a^ and b = a imply stability whereas instability
yx

results when both a^ and a get large or when either gets very large.
yx

Since large values o f the effective fear and loathing coefficient imply
chaos and war, and since any collection o f autocratic states (or large mixed
collection o f autocratic and democratic states) is more likely to have some
pairs o f nations w i t h large values o f this coefficient than a corresponding
collection o f democratic states, it follows that an outbreak o f war is more
likely i n a collection o f autocratic states than in a similar collection o f demo­
cratic states. Even though it is often easier, in foreign affairs, to deal with
autocratic governments, it is evident that, in addition to the usual human
desire to nurture similar systems elsewhere, democracies are more likely to
ensure w o r l d peace by supporting the democratization o f other members o f
the international system.

W h i c h is M o r e War-Prone: A S y s t e m of S h i f t i n g
Alliances o r a Collection o f G o - l t - A l o n e States?

As a last example, consider the question: which is a more stable system o f


competing states: one in which each nation individually procures arms so as to
158 Chaos Theory in the Social Sciences

match any other state in the system or one i n which weaker nations j o i n
coalitions to collectively procure arms against the strong and in which they
shift alliances when a coalition member becomes too strong—independence
security or balance-of-power security? To model the t w o cases simply, con­
sider three competing states, either independent o f each other or in which any
two w i l l be allied against the stronger third.
Considering any pair o f the three, arms buildup or build-down is pre­
sumed to occur just as in the previous case (equation 9), w i t h the effective
fear and loathing coefficient o f X for Y and l/a^ the corresponding confi­
dence coefficient. However, now X may pair against Y or Z , i f all three are
independent, so the resultant arms procurement or build-down is the sum o f
the corresponding pair terms:

AX = X N+L -X„ = [ Y„(YAXY N - X„) 0 (Y N - X„) + a Z (Z XZ N N - X) N

l
e (z„ - x )] © ( - = ^ ) n

- ~- x (x n n - Y ) 0 (x -y„)~
N n ~ x (x n n - Z„) 0 (X - Z„) N

AY = y„ + 1 -Y„ = [a X {XYX N N - Y ) 0 (X
N N - Y„) + a Z (Z YZ N N - Y) N

0 (Z„ - Y )} 0 ( ^ - ^ )
N (10)

- - L Y (Y N N - x„) © (Y -x„)-^- N Y„(Y„ - Z ) 0 (Y - Z„) N N


u u
yx yz

AZ = Z„ + 1 - Z„ = [a X„(X„ ZX - Z„) 0 (X„ - Z„) + a Y (Y ZY N N - Z„)

0 (Y N - Z„)] 0 ( X
-=^ )

- -L z (z n n - x ) © (z„ - x ) - -L z (z
n n n n - Y„) © (z„ - Y„).

In the alliance model, X w i l l j o i n its procurement w i t h Y to match a


superior Z , or j o i n w i t h Z to match a superior Y , and similarly for build-down
(here the appropriate coefficients are written b^, b , and so on): x z

AX = [b^Y,, - X „ - Z„) 0 (Y N ~ X „ - Z„) + b Z (Z XZ N N - X N - Y)


N

0 (Z„ - X n - Y )} 0
N ( )
The Prediction of Unpredictability 159

TT~ V« + z )(X n n + Z n - Y ) 0 (X
n n + Z„ - Y) n

xy

(X n + Y )(X n n + Y - Z„) 0 (X„ + Y -


n n Z„)

AY = [b X {X„ yx n - Y „ - Z„) 0 (X n - Y n - Z„) + b Z (Z„ yz n - Y „ - X)


n

1 - AY
0 (z„ - y„ - x„>] ©

- r - (Y + z„)(Y n n + z„ - © (r„ + z„ - x) n

(ID
- ^ - (y„ + x„)(Y n + x„~ z„) 0 (K„ + x„ - z„)

AZ = [b X (X ZX N N - Z „ - Y„) 0 (X„ - Z„ - K„) + fo /„(K„ - Z„ - X„)


z

0 (K„ - Z„ - X„)] 0 ( )

- j - (z„ + r„)(z„ + Y„ - x ) 0 (z„ + Y„- n x)


n
zx

-T~(Z n + X )(Z„
n + X „ - Y ) 0 (Z„ + X
N N - Y ).
N

The t w o models each consist o f three coupled recursion relations (eqs. 10


and 11). The answer to the question posed is obtained by seeing which o f the
t w o models is more stable.
The analysis is done by assuming that the initial values, X , Y , Z , in 0 0 0

each o f the three nations are almost the same and asking whether these small
differences grow as the recursion progresses. A spreadsheet iteration o f the
recursion relations is again carried out, starting off w i t h XQ, Y' , Z' , represent­ 0 0

ing small random differences about an arbitrarily chosen starting point, X . 0

Lyapunov coefficients are computed (again by iterating out to n = 100 w i t h


small 5 ) , varying the starting value X and the single model parameter a
0 0

(representing a common value for a^, b^, a , and so on). xz

Four possibilities become evident, as the parameters a and X are varied, 0

for the independent nation model as a result o f the numerical computations


(Saperstein 1992b). I n the first case, strong stability, all sequence limits exist
and all Lyapunov coefficients are negative, which implies that all sequence
limits exist and are the same. There is complete predictability and hence no
war. I n the second case, weak stability, all sequences have limits that are
160 Chaos Theory in the Social Sciences

equal to each other, but the Lyapunov coefficients are positive. This occurs
because all limits are less than the starting value, X , except for the singular
0

case when all start off exactly equal to X , in which case the limiting value is
0

X . Hence, the numerator in the definition o f the Lyapunov coefficient doesn't


0

become less than the denominator, implying a positive coefficient. This situa­
tion is an illustration o f bifurcation, in that one point leads to results com­
pletely different from all other points. Since the limits are still well defined,
the sequences are predictable, not chaotic, and so again there is no war. I n the
third case, weak chaos, none o f the sequence limits exist and so there are no
Lyapunov coefficients. However, the variation within any one o f the se­
quences, X' , Y' , Z' , and the differences between the sequences (X' — X )
n n n n n

remain small compared to the arbitrary j o i n t starting value, X . The situation


0

resembles case b i n figure 7 . 1 , rather than case c; hence, the unpredictability


is minor and so, i n the paradigm o f this chapter, does not represent war. The
last possibility is one in which the sequences do not have limits and the
variations w i t h i n each sequence, as w e l l as the differences between them, are
large compared to the starting value. In this case the range o f unpredicta­
bility is the entire range possible, there is no way o f knowing—even
approximately—the outcome o f any policy or action, and hence major fluctu­
ations may result from minor perturbations. This is the criterion for crisis
instability and war.

The numerical results for the independent nation model are displayed in
figure 7.3. N o sharp boundary was evident between the regimes o f weak
stability and weak chaos and so the two regimes are considered to be a single
realm i n w h i c h crisis stability is still valid. Note that there is no upper l i m i t to
the weak chaos regime, no breakdown o f crisis instability, i f the fear and
loathing coefficient is small enough (a ^ 1). I f there is enough confidence in
the system, perturbations to the independent nation international system may
lead to continued fluctuations and uncertainty, but not to war, no matter how
large the initial, equal, stocks o f arms are. On the other hand, with sufficient
fear and loathing (a > I), war w i l l always break out i f the initial arms stocks
are large enough.
The numerical predictions o f the alliance model are quite different. For
large enough symmetric starting values, X = Y = Z , the system imme­
0 0 0

diately and completely disarms, no matter what value is chosen for the fear
and loathing coefficient. This is region I o f figure 7.4; since all sequences tend
immediately to a zero l i m i t i n g value, it is a strong stability region. I n region I I
o f the figure, a l i m i t has not yet been reached numerically at n = 100;
nonetheless, the fluctuations in the monotonically decreasing sequences are
small and hence this region is appropriately characterized as weak chaos: the
course o f events is clear and the eventual complete disarmament is completely
predictable.
a

Fig. 7.4. N u m e r i c a l p r e d i c t i o n f o r t h e alliance m o d e l


162 Chaos Theory in the Social Sciences

Hence, using the possibility o f strong chaos in the model as an indicator


o f the possibility o f war in the corresponding international system, it follows
that a policy o f building (perhaps shifting) alliances is guaranteed to keep an
initially symmetric system peaceful. A policy o f seeking national security by
"going i t alone" allows the possibility o f war in such a system i f the contend­
ing parties start out w i t h large enough levels o f armed force.
I know o f no real-world situation in which a system o f comparable
competing states have balanced themselves d o w n to a complete disarmament
as is predicted by the alliance model o f this chapter. Perhaps this is because a
shifting o f alliances has costs i n the real w o r l d , costs not included in this
model. Also possible is that there are no real balance-of-power situations; real
nations may resort to balance-of-threat instead, a much harder situation to
model (Walt 1987). Or, the presumed anarchy may not have always been
present.

Conclusion

Humans have always developed theories (often mythologies) to explain their


w o r l d . W i t h the advent o f science, theories had to have practical as well as
literary or philosophical implications. A successful theory o f international
relations should be an important component o f competent policy making in
the arena o f international security in a competitive world. Conversely, a
careful analysis o f practical policy—successful or unsuccessful—must be the
foundation for any scientific theory.
Given the complexity o f the real international system, any theory o f it is
bound to be incomplete, describing some aspects o f it more thoroughly than
others. A truly massive modeling enterprise—such as is done i n economics
w i t h national input/output models or in meteorology w i t h large-scale
computer models o f the atmosphere—may at best lead to very rough, short-
term, quantitative agreements between predicted and observed experience.
The quality o f the results is often vastly disproportionate to the effort required
to secure them. It thus seems useful to look for simple models and methods o f
model analysis that can lead to qualitative robust predictions.
One such approach is the prediction o f unpredictability in simple non­
linear models o f the competitive interactions between rival states. It can be an
important approach to policy and theoretical questions, given the fundamental
assumption that such unpredictability may represent crisis instability and war
in the international system. Given that policy questions are usually more
specific and complex than more theoretical questions, we expect to require
more complex models for the former, although they would still be consider­
ably simpler than any intended to address quantitative questions. For example,
very simple theoretical models, easily analyzed w i t h spread sheets on a desk-
The Prediction of Unpredictability 163

top computer, have allowed us to conclude that a tripolar world is less stable
than a corresponding bipolar w o r l d , that a set o f democratic nations is more
likely to be stable than a similar set that includes autocracies, and that a group
o f nations attempting to guard their security via balance-of-power alliance
formation is more stable than a group in which each member makes individual
attempts to cope w i t h other potentially hostile members. A bigger and more
complex computer model indicates that a policy o f introducing S D I into a
situation similar to the recent cold war nuclear confrontation between the
United States and the Soviet Union would likely dangerously destabilize such
a world.
None o f these theoretical or policy results is new. Similar conclusions
have been reached before by many others using conventional verbal analyses.
But the alternative route presented in this chapter—the qualitative analysis o f
simple quantitative models—leads very quickly to the answers without any
hidden implicit assumptions, biases, or faulty logic. Finally, agreement o f
results obtained via traditional means and the methods developed here brings
added credence to both approaches. Science is a fabric: its ability to cover the
w o r l d depends upon the existence o f many different fibers acting together to
give it structure and strength.
CHAPTER 8

C o m p l e x i t y in t h e E v o l u t i o n o f Public O p i n i o n

Michael McBurnett

Most political phenomena are not static. Presidential campaigns and the elec­
torates' response to them; the emergence and outbreak o f international hostili­
ties; the formation o f racial attitudes—all these and more are inherently
dynamic processes.
Be that as it may, we often study such phenomena as though they were
static. Cross-sectional analyses continue to dominate disciplinary research,
for two reasons. First, we typically choose to analyze cross-sectional data.
Exemplary are the many analyses o f year-specific surveys o f the electorate
during presidential and off-year elections (see Huckfeldt and Sprague 1987;
Cohen et al. 1991; and examples based on specific National Election Studies),
w h i c h , it might be noted, represent some o f the most sophisticated empirical
work in political science. Second, our methodological training has largely
concentrated on techniques well suited to the analysis o f cross-sectional data;
ordinary least squares and its cousins continue to hold sway. Although
scholars increasingly are turning to the analysis o f phenomena over time
(Tuma and Hannan 1984; Huckfeldt 1990; B r o w n 1991), the study o f dynamic
processes i n a truly dynamic way remains in its infancy.
One o f the most foreboding challenges is to identify the complexity that
inevitably characterizes dynamic phenomena (Ford 1986). In the pages that
follow, three methods are used to analyze complex dynamical time series.
First I discuss the phase portrait. Second, the correlation integral is introduced
and applied to a variety o f time series. Finally, Lyapunov exponents are used
to illustrate the divergent properties inherent in complex dynamics. I discuss
those methods w i t h i n the context o f the 1984 race for the Democratic nomina­
tion. When candidates do w e l l , estimates o f their chances o f winning
increase. When the candidates fare poorly, citizens report a corresponding
decrease in their estimates o f the candidates' chances o f winning the nomina­
tion. This pattern o f probability estimates does not follow an orderly cycle,
however; the pattern is considerably more complex—so complex that it is best
described as chaos. Chaos implies a deterministic structure in a nonlinear
dynamic process.

165
166 Chaos Theory in the Social Sciences

I first demonstrate that cycles can emerge naturally in competitive nomi­


nation races. I then discuss the phase portrait, a technique designed to reveal
systematic patterns in over-time data. A p p l y i n g this technique to a 1984 NES
study identifies a very complex, yet regular structure. The heart o f the anal­
ysis is a demonstration that this time series is chaotic. After showing that
application o f an alternative linear model fails to account for the complicated
dynamics in the time series, I conclude w i t h some suggestions for extended
application o f this technology to other data series.

The Emergence o f O p i n i o n Cycles i n


Competitive Nominations

The empirical pattern o f voter support for candidates as the presidential nomi­
nation race unfolds provides direct evidence o f how a cyclical pattern in
public opinion may develop during the campaign. Figure 8.1 shows the vote
percentages received by Hart and Mondale in the 1984 Democratic race for
each election date. Where multiple races were held on a single day, as on
Super Tuesday, the results have been averaged over all races held on that day.
The graph presents three pieces o f evidence about the dynamics o f candi-

J . I I I I I L

n 1 1 1 1 1 r
1 5 10 15 20 25 30
Election Sequence

Fig. 8 . 1 . O s c i l l a t i n g p a t t e r n of v o t e percentages in t h e 1984 D e m o c r a t i c


n o m i n a t i o n race b e t w e e n M o n d a l e a n d Hart
Complexity in the Evolution of Public Opinion 167

date competition in presidential nomination races. First, w i n n o w i n g , which


occurs as a result o f competition (Aldrich 1980; Mayer 1987), appears to take
place early in the campaign, as those candidates who are not viable are
eliminated. This is inferred by summing the percentage o f the vote captured
by Hart and Mondale. Second, the graph illustrates the concept o f momen­
t u m . For the three races following the Iowa caucuses (the first point on the
graph) Hart's percentage o f the vote climbs. Following Hart's crushing defeat
o f Mondale i n Vermont's nonbinding primary, the percentage o f the vote Hart
wins begins a three election decline. As Hart's electoral fortunes go into
decline, Mondale's reverse. T h i r d , as the nomination race progresses,
Mondale and Hart trade the lead a number o f times. The predominant pattern
in the race is one o f oscillation, indicative o f the ebb and flow o f electoral
fortunes in the race. This demonstrates that momentum works in two ways: it
can be positive or negative.
The oscillatory pattern can be demonstrated, first, w i t h a simple model o f
the process o f winning votes (see Sprague 1976). Expectations o f future
candidate performance in the nomination race are based on the past perfor­
mance o f the candidate i n electoral contests, weighted by the number o f
contests or the number o f delegates that remain to be won (Aldrich 1980;
Bartels 1988, 3 5 - 3 6 ) .
Two versions o f this model are used to illustrate these dynamics. The first
is unweighted and the second uses the percentage o f the contests that remain
as a weight. This model (for Mondale) has the form

Mondale Percent o f the Vote, = a + (3 • (Mondale Percent o f


( 1 )
the V o t e , . , ) .

The parameters are estimated by O L S . The results establish that the percent­
age o f the vote Mondale wins does oscillate about a limiting value. I n the
unweighted model, the estimate for the lagged variable is —.389 ± .22. The
constant term in this model is .49. The estimate in the weighted model is
— .409 ± . 19, w i t h an intercept o f .52. Both parameter estimates are negative
and the absolute value o f each estimate exceeds zero but is less than one. The
time path for both versions o f this model is an oscillatory trace that converges
rapidly to an equilibrium. Both demonstrate that oscillation in the percentage
o f the vote a candidate wins can occur during the nomination race. However,
the pattern produced by the parameters is short-term, period-two oscillations,
a pattern not found in the original data.
A more elaborate model demonstrates the long-term oscillatory nature o f
the election outcomes more convincingly. The traces for Hart and Mondale in
figure 8-1 show long-term oscillations rather than period-two oscillations or
an oscillatory convergence over time. Competition is an integral part o f the
168 Chaos Theory in the Social Sciences

campaign, so a model should account for the competition for votes between
candidates. Using only Hart and Mondale, a dynamic model o f competition is
given by

AMondale, = j 8 - 0 j8, Mondale, + j8 Hart, 2

(2)
AHart, = B' - Q j3',Hart, + /3 Mondale, 2

where AMondale, refers to the difference in the vote percentage Mondale


received between time t + 1 and time t, B is a constant, j8, reflects the level 0

o f difficulty campaigning plays in attempting to w i n votes by Mondale, and B 2

measures the threat Mondale feels due to Hart's level o f success in repeated
elections. B' , B\, and B' in the second equation are interpreted similarly for
0 2

Hart. This model closely resembles the Richardson model o f arms races.
One equation can be eliminated by substitution (see Huckfeldt et al.
1982). This produces a second-order linear difference equation in one candi­
date in which the interdependent structure o f competition is implicit. The
procedure is symmetric for the other candidate. The result for Mondale is

Mondale, +2 = b 0 + fc,Mondaie, +1 + & Mondale„ 2 (3)

where b represents (B — B B
0 — B B[B ),
0 b represents (1 + B — B —
2 0 2 0 x 2 i

B B[), and b represents (B B[B


2 2 - /3 /3, - B B\ - B ), so that equation 3 is
2 l 2 2 2

isomorphic to a regression equation with three parameters. The parameters are


estimated using O L S , as before. The results for Mondale and Hart are shown
in table 8 . 1 . After estimation, the equation can be put in the homogeneous
form:

Mondale, +2 ~ & Mondale,


t +1 — 6 Mondale, = 0.
2 (4)

Using the quadratic equation, one can solve for the t w o roots o f the equation.
These roots may both be real, repeated roots, or complex conjugates. These

TABLE 8 . 1 . E s t i m a t e s f o r t h e S e c o n d - O r d e r
Dynamic Model

Candidate b 0 b 2

Mondale .57 -0.1020 -0.4188


Hart .35 0.4304 -0.2937

Source: 1984 Rolling Cross Section.


Complexity in the Evolution of Public Opinion 169

are the only alternatives (Goldberg 1958; Huckfeldt et al. 1982). Substituting
the estimates from Mondale's equation into the quadratic equation gives
2
_ - ( . 1 0 2 0 ) ± V ( . 1 0 2 0 ) - 4 · (.4188)
r
ii R
2 2 ' ' '

It is obvious that the sum under the square root radical is less than zero.
Hence, the roots are complex conjugates and the time path described by this
model contains long-term oscillations. The pattern is similar for Hart. For
particular values o f these roots, the time path w i l l show stable long-term
oscillatory motion (Baumol 1958). This implies that the competition between
candidates persists over time during the nomination race, an implication that
can lead to the media's portrayal o f competition between candidates as a horse
race (Brady and Johnston 1987, 184; Popkin 1991, 117-18). Campaign man­
agers for the candidates in the 1984 race agree that this race was one in which
the competition persisted and one in which the outcome was i n doubt for some
time (Moore 1986, 56).
The outcome from this model resembles the time paths shown in figure
8 . 1 , where neither candidate shows a simple oscillatory motion and the lead
changes several times during the primary season. This simple model may
provide some insight into the problem survey respondents have when asked to
estimate the chance each candidate in a large pool o f candidates has o f
winning the party nomination. The problem o f accurately assessing the
chances o f a number o f candidates at a particular point in time is liable to be
quite difficult when the candidates are locked in a race in which the lead
changes a number o f times (Battels 1988, 3 2 0 - 2 3 ) . '
The simple models explored in this section demonstrate that competition
between candidates can lead to fairly complex long-term behavior in the
nomination race. I n this changing environment, where members o f the elec­
torate are liable to switch their candidate preference ordering subject to the
results o f recent electoral outcomes, forecasting the eventual outcome be­
comes extremely difficult. Preferences i n one state are driven by recent elec­
toral results in another state. Issues become less important and electability
increases in importance (Brady and Johnston 1987). The Mondale campaign
discovered that even what is considered to be unshakable support can be
swayed in this fashion. Fifty percent o f Mondale's lead in Maine disappeared
between the New Hampshire primary and the Maine caucuses and this is
attributed to Hart's early successes (Moore 1986, 76). The forecasting prob­
lem for campaigners is difficult, but it is precisely this difficulty that makes the
problem interesting, for campaigners and scholars alike.
I turn now to a method designed to analyze time series graphically and to
discover hidden patterns in the data that may not be revealed by other
methods.
170 Chaos Theory in the Social Sciences

Revealing S y s t e m a t i c Patterns i n Time-Series Data

The visual inspection o f time-series data rarely reveals an underlying regu­


larity in the data. For one thing, the time series may be too short: patterns that
w o u l d be evident in very long series w i l l not appear. For another, the level o f
noise i n the data may be large relative to the systematic movement in the
observations o f the time series.
In this section, I offer a simple example to motivate a strategy that allows
one to discover patterns in time-series data that might not otherwise be visible
due to the length o f the series and to noise in the data. After demonstrating
how the technique works—and that this technique works when noise is
present—I use it to investigate a time series constructed from the 1984 NES
Rolling Cross Section.

A n a l y z i n g D e t e r m i n i s t i c Data

Assume that a candidate's mean thermometer rating among survey respon­


dents follows a regular long-term cycle during the course o f the nomination
race. I f such a cycle were present, it might appear as shown in figure 8.2. This

"I I I I i I I i 1 r 1 1 n
1 30 60 90 120 150 180 210 240 270 300 330 366
Days Since January 1

Fig. 8.2. A h y p o t h e t i c a l c a n d i d a t e ' s average d a i l y t h e r m o m e t e r r a t i n g


o v e r t h e c o u r s e of t h e e l e c t i o n year
Complexity in the Evolution of Public Opinion 171

figure was produced by graphing a two-dimensional linear harmonic function


that is centered on the value 50 percent and that ranges between thermometer
scores o f 60 and 40 percent (see Hale and Kocak 1991). It is deterministic,
clearly periodic, and contains long-term cycles (the oscillations are greater
than period t w o ) . For the purposes o f illustration, the function is graphed over
an interval o f one year. The equation for the function is given by

AY, = aY, + bX t

(6)
A X , = cY, + dX„

where a = 0, b = 1.0, c = —1.0, a n d d = 0, and the time interval is 1, 2, 3,


. . . , 366.
To demonstrate that a cycle is present, a phase portrait is constructed.
The phase portrait is a graph that shows the long-term behavior o f the time
series w i t h time removed. Put another way, a phase portrait displays the
asymptotic behavior o f a dynamic system. It is constructed by plotting the
values generated from equation 6 o f X, against Y . t

This figure contains an ellipse and this ellipse is known as an oscillator.


A n oscillator is an equilibrium solution that arises when a dynamic system
undergoes " w e l l defined cyclic changes in t i m e " (May 1974, 16). A similar
pattern o f survey responses can emerge in a contest like the 1984 Democratic
race, where candidates swap the lead several times. I f the system is perturbed,
as time evolution continues, the trajectory w i l l simply orbit at a new level o f
equilibrium.
This is an explicitly oscillatory phenomenon, and the example here is
based on a linear harmonic oscillator. The simple second-order models devel­
oped in the prior section have nonlinear solutions that produce oscillatory time
paths. The roots o f the dynamic equations describing the percentage o f the
vote Mondale and Hart won in the sequence o f elections were complex conju­
gates. These roots have representations in the real plane as trigonometric
functions that mimic dynamical behavior similar to that shown in figure 8.3.

A n a l y z i n g Noisy Data

What happens when the system contains noise? Noise may enter the data
through measurement error, random response variation, sampling error, or by
unknown other means. Figure 8.4 shows the simulated time series o f mean
thermometer scores w i t h random variation about each point. The initial func­
tion is identical to the first part o f equation 6, but random error has been added
to each observation o f the time series. The random component o f this series is
normal (/it = 0, a = 0.025).
I f the procedure o f examining the phase portrait is adopted to study the
172 Chaos Theory in the Social Sciences

dynamical behavior o f this series, the technique is to plot Y against Yt and t+l

view the result (see fig. 8.5). Though it is k n o w n that an equilibrium is


present, simple random variation in the data masks i t . The result is a hope­
lessly tangled mess.
A solution to the problem is a variation o f a technique suggested by
Packard et a l . (1980). They demonstrate a way to recover a strange attractor
from a time series. A strange attractor is an attractor that is not a cycle or
equilibrium point, but is stable in the sense that the trajectories o f a dynamical
system remain nearby as the system evolves in time ( M o o n 1987, 267). These
attractors do not have ordinary shapes; they are not elliptical or circular i n the
phase space. I n part it is their odd shapes that earned them the name strange.
Packard et al. observe that any time series contains all the information
that went into creating i t , even a series consisting o f several variables. Thus,
one can use lagged values o f one variable in the series to construct a phase
portrait and study the long-term dynamics o f a multivariate time series. This
enables a sufficiently low-dimensional attractor to be revealed i f one is pre­
sent. I f a strange attractor can be revealed, then an equilibrium cycle can also
be revealed.

~l I 1 1 T
.4 .45 .5 .55 .6

Mean Thermometer Rating at Time f for x

Fig. 8.3. Phase p o r t r a i t o f average thermometer scores i n a t w o -


c a n d i d a t e race
Complexity in the Evolution of Public Opinion 173

Technically, this operation amounts to inserting a two-dimensional slice


into a higher-dimensional space and mapping all the points down onto
the plane. The result is the phase portrait. Some orientations o f the t w o -
dimensional plane show more interesting views o f the dynamics than others.
The problem w i t h revealing the attractor is found in the orientation o f the
graph w i t h respect to the viewer. The data cloud in figure 8.5 is oriented from
lower left to upper right, similar to a person holding his or her hand a distance
of one foot in front o f his or her eyes (as i f the person were preparing to issue a
left-handed salute). From this angle, one can see the thumb and first finger,
but little else o f the structure o f the left hand. B y rotating the hand so that the
thumb turns toward the floor, the entire structure o f the back o f the hand is
revealed. H o w can this idea be applied to the tangled structure that emerges
from a plot o f Y, against Y w i t h error? Note that this implies a three-
t+l

dimensional attractor, for obviously a hand is not a two-dimensional object.


Nicolis and Prigogine (1989, 277) display an example o f a technique that
uncovers a strange attractor by using the appropriate lag structure in a three-
dimensional representation. The key is that the lag structure used is that which
is appropriate. Experimentation is required to reveal the appropriate lag. The

1 1 1 1 1 1 1 1 1 I I I I
1 30 60 90 120 150 180 210 240 270 300 330 366

Days Since January 1

Fig. 8.4. S i m u l a t e d t h e r m o m e t e r scores w i t h a r a n d o m component


series f o r C a n d i d a t e Y o n l y
174 Chaos Theory in the Social Sciences

J I I I I 1 I L

n 1 1 1 1 1 1 1 1 i
.3 .35 .4 .45 .5 .55 .6 .65 .7 .75

Mean Daily Thermometer Score at Time f

Fig. 8.5. Phase p o r t r a i t f o r s i m u l a t e d t h e r m o m e t e r scores w i t h r a n d o m


variation. Candidate Y only

identification o f the lag structure proceeds as follows: use successively higher


shifts i n the lag, defined as integer multiples o f a fixed lag, T ( T = mAt), where
m is an integer and A f is the difference interval between successive samplings
(Nicolis and Prigogine 1989, 276). Ar is always 1 for maps such as are used
here (it differs from 1 only when the dynamical system is based on differential
equations and describes a step size for the numerical integrator in those cases).
m is an increasing index, 1 , 2 , . . . , n. This process leads to a set o f discrete
variables Y„ Y , · • • , Y,+ -
t + l n

Nicolis and Prigogine point out that "for a proper choice o f T these
variables are expected to be linearly independent, and this is all we need to
define a phase space" (1989, 277). A l l o f these variables are derived from a
single time series, such as is shown in figure 8.4, and are provided by the
data. This procedure yields an experimental time series and it provides the
data to transcend the one-dimensional space o f the original (univariate) time
series and to "unfold the systems' dynamics in multidimensional phase space"
(Nicolis and Prigogine 1989, 277).
Figure 8.6 provides a graphic revelation o f the structure o f the attractor
k n o w n to be present in the noisy data representing the simulated thermometer
scores. Here the value o f the thermometer score at time t is plotted against the
thermometer score at time t + 10. The tenth lag o f the series has been used to
Complexity in the Evolution of Public Opinion 175

J L

I 1 1 1 I I r~
.35 .4 .45 .5 .55 .6 .65

Mean Thermometer Score at Time t


Fig. 8.6. Phase p o r t r a i t r e c o n s t r u c t i o n o f noisy t h e r m o m e t e r scores,
r e v e a l i n g t h e u n d e r l y i n g l i m i t cycle

reveal the equilibrium cycle. It retains its elliptical shape. This is strong
evidence that this technique produces an insight that might otherwise be
2
masked by random variation or noise in a time series.
Clearly, in order to illuminate fine structure in a dynamical system,
empirical methods may be required. What happens when these techniques are
applied to a time series o f public opinion data? I n the next section, the time
series used in the later analysis is described.

C o n s t r u c t i o n of t h e T i m e Series
f r o m Public O p i n i o n Data

In 1984, the Center for Political Studies conducted a telephone poll o f the
national electorate using a probability sample over a very long period. This
survey is k n o w n as the Rolling Cross Section. Although it has some draw­
backs, such as a fairly large sampling error (Allsop and Weisberg 1988), it is
the only national sample o f its type. This analysis uses that portion o f the
survey containing those respondents interviewed after January 11, 1984, and
before June 12, 1984. Those respondents were interviewed during the period
in which all primary and caucus activity took place.
The survey asked respondents to rate the chances o f the various Demo-
176 Chaos Theory in the Social Sciences

cratic candidates winning their party's nomination. As Bartels points out,


most respondents failed to rate candidates' chances as i f they had probability
theory i n m i n d (1988, Appendix A ) . Though Bartels used this survey for a
quite different purpose, I make use o f his coding scheme, which measures
each individual's perception o f each candidate's chance o f winning the Demo­
cratic nomination (1988, 3 2 2 - 2 3 ) .
The individual responses were averaged according to the date o f the
responses. This results in a measure o f the perceived mean daily probability o f
each candidate's chance o f winning the Democratic nomination.
In order to see i f any interesting dynamical patterns were evident, I
smoothed the data using a five-day moving average. Smoothing a time series
that contains noise may allow a cyclical or otherwise interesting pattern to be
identified visually, i f one is present. Smoothing reduces, but does not e l i m i ­
nate, the level o f noise in the time series.
This time series is shown in figure 8.7. The graph shows the smoothed
mean probabilities that Mondale, Hart, and Jackson w i l l w i n the 1984 Demo­
3
cratic nomination over the span o f the nomination p e r i o d .
The phase portraits suggest that the dynamics o f these time series are not
regular or oscillatory. Instead, the phase portraits reveal what may be a
strange attractor resembling the Lorenz attractor from climate dynamics
(Lorenz 1963; Schuster 1988). Strange attractors have so far been identified
only in chaotic dynamical systems (Schuster 1988). The identification o f this
type o f structure underlying the dynamics o f public opinion is important
because o f the implications this has for forecasting the outcome o f nomination
races as well as the study o f change in public opinion over time generally
(Huckfeldt 1990).
The Lorenz attractor is a strange attractor and the Lorenz equations have
all the identifying characteristics o f chaotic behavior: the motion is highly
erratic, the attractor is contained in a bounded region in the phase space, and
the dynamics display sensitive dependence on initial conditions (Lorenz
1963). H i g h l y erratic motion means it is difficult to tell, a p r i o r i , exactly what
value the next observation in the series w i l l take on, even when the equations
governing the behavior o f the system are k n o w n . The dynamic behavior o f the
series is contained in a restricted region o f the phase space. Being bounded
means that the series does not diverge toward ± ° ° over time. Sensitive depen­
dence on initial conditions is a property o f nonlinear systems. It is defined as
"separating initially close trajectories exponentially fast in a bounded region
of phase space" (Schuster 1988). I n practice, this means that two nearby
points selected as starting values for a dynamic system w i l l separate exponen­
tially quickly over time as the system is iterated.
The dynamics illustrated i n the phase space (chap. 2, figs. 2.10, 2 . 1 1 ,
and 2.12) meet t w o o f these requirements: the attractor is contained in a
Complexity in the Evolution of Public Opinion 177

"1 I I I I I I I I I i I i I I
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140

Days Since January 1,1984

Fig. 8.7. P r o b a b i l i t y o f w i n n i n g t h e 1984 D e m o c r a t i c n o m i n a t i o n (all


respondents)

bounded region in phase space and the motion is erratic. The governing
equations for the system used in this analysis are not k n o w n , so it is impossi­
ble to tell, given the information presented thus far, i f this series displays
sensitive dependence on initial conditions.
However, i f this system can be shown to be sensitively dependent on
initial conditions, then it is established that public opinion measures o f the
mean likelihood o f candidates winning their parties' nomination are very
probably unpredictable far in advance. This is disconcerting, not because
forecasts may be slightly off from time to time, but because forecasts may be
completely wrong between two adjacent polls only narrowly separated i n
time. This means that pollsters may predict the wrong candidate winning
elections. This has occurred in recent nomination races. James Johnson,
Mondale's campaign manager in the 1984 race, relates that

we d i d a poll [one week prior to the New Hampshire primary] o f our o w n


people in the straw p o l l , and it was 50-50 with Hart. These were people
who had signed in blood on October 1 that they were for Mondale until
death, and we were 50-50. (Moore 1986, 73)
178 Chaos Theory in the Social Sciences

Peter Hart was Mondale's pollster in the 1984 race. His forecast o f Mondale's
support i n New Hampshire was off by 50 percent. The electoral result i n the
New Hampshire primary was a boost for Hart and tighter competition for the
Democratic nomination.
The 1976 Democratic race also provides examples. Morris Udall's p o l l ­
sters reported three weeks prior to the Wisconsin primaries that the vote split
was Carter 34 percent, (Henry) Jackson 24 percent, Udall 17 percent, Wallace
15 percent, and Undecided 10 percent. Udall desperately needed to w i n in
Wisconsin in order to stay in the race. This p o l l nearly forced h i m from the
race. Ten days prior to the election, another poll revealed a new mean prefer­
ence ordering between candidates among the public: Carter 34 percent, Udall
30 percent, Henry Jackson 15 percent. This p o l l , as well as the previous one,
was conducted by Peter Hart (Schram 1977, 126). The Wisconsin primary
results (the first primary in the 1976 Democratic race) were Carter 37 percent
and Udall 36 percent, w i t h only five thousand votes separating these two
candidates (Schram 1977, 132). On the Thursday prior to the Wisconsin
primary, Carter's pollster, Pat Caddell, reported that "the weighted results
stood at Carter 38%, and Udall 2 7 % , " w i t h 5 percent no-preference respon­
dents, who were expected to vote for Udall (Schram 1977, 133). This p o l l ,
conducted only days prior to the election, is close to the true outcome.
The forecast error increases as the distance from the actual election date i n ­
creases.
This problem has obvious implications for theoretical political science,
as well as for practical politics. Can this series, or other similar series, be
shown to be chaotic?
The answer turns on the ability to determine i f the phase portrait is a t w o -
dimensional slice o f a higher-dimensional strange attractor with a positive
Lyapunov exponent. This attractor must have a fractional, that is, noninteger
dimension; no strange attractors are k n o w n w i t h integer dimensions (Schuster
1988, 106). I f it can be shown that the phase portrait is a snapshot o f a strange
attractor w i t h a positive Lyapunov exponent, then the dynamics o f the un­
k n o w n equations o f motion must display sensitive dependence on initial con­
4
ditions and the time series is chaotic.
A demonstration o f sensitive dependence on initial conditions in a time
series o f opinion data has ramifications that extend far beyond this essay.
Nonlinear processes define the underlying structure o f opinion formation i f
the series can be shown to be chaotic. Forecasts o f election results based on
opinion data w i l l be unreliable unless the initial levels o f public opinion are
k n o w n quite precisely. This condition is never met in practice. Linear models
are functionally misspecified and results from these models are suspect.
We must probe the dynamics w i t h other tools that provide deeper insights
into the dynamics. One tool in common use is smoothing, which can help
reduce the variation in the observed series. Smoothing has problems o f its
Complexity in the Evolution of Public Opinion 179

own (see Schaffer et al. 1990). Another tool is spectral analysis, which can
uncover regularities in dynamical time series in the presence o f considerable
noise. As this is the subject o f a separate chapter in this volume I refer the
reader to it (see chap. 2). The time series in this chapter are analyzed i n much
greater detail there. Smoothing and spectral analysis are both designed to
probe for periodicity in the time series. What is needed is to attack the
problem from another angle. Two techniques are available to probe the geom­
etry o f the time series: the correlation integral and Lyapunov exponents.
In the next section I explain the construction and use o f the correlation
integral. I also posit an alternative (linear, oscillatory) model and recompute
the correlation integral. I follow w i t h an examination o f the largest Lyapunov
exponent for the Mondale time series.

Measuring t h e Dimension of a Strange Attractor

This analysis uses the Mondale time series rather than data from the entire
5
three-candidate series. It is known that the shape o f attractors in low-
dimensional phase space can be recovered using the phase space representa­
tion from an appropriately lagged univariate time series (Packard et al. 1980;
Nicolis and Prigogine 1989). This result allows the use o f one series to
analyze the dynamic properties o f the entire three-candidate series.
Some vector notation is required for the analytic procedure that follows:
Xj represents a point i n phase space with the coordinates for the point given by
the ordered n-tuple { X , X - ,0 / X ,- ,
0t { • • • , Xot-J-
0 2 Typically, one begins
w i t h pairs o f points, then moves into a three-space, and then continues into
higher-dimensional phase space. The analysis takes place in this subspace o f
the N-dimensional space in which the original attractor lies (Packard et al.
1980, 712). I t is a consequence o f Takens's theorem (1980) that this technique
works. The dimension o f the attractor can be measured i n a number o f ways.
A measure that is computationally efficient for a discrete time series such as
this one is called the correlation dimension.
The correlation dimension is useful when one does not know the exact
dimension i n which the attractor lies. It is an experimental measure that w i l l
determine the dimension o f the attractor numerically.
Based on the ordered n-tuple constructed above, a reference point x is
chosen and all distances |JC, — jt-| from the N — I remaining points in the series
are computed (Nicolis and Prigogine 1989, 278). These distances are arrayed
against various values o f a radius r from the point X in the phase space. This t

process is repeated for all values o f i. This computation calculates a correla­


tion integral. The correlation integral is denned as
N

2
C(r) = l i m (l/N ) .^ 6(r - \ x,•. - X j |), (7)
180 Chaos Theory in the Social Sciences

6
where 6 is the Heaviside function, 6(x) = 0 i f x < 0 and 6(x) = 1 i f x > 0 .
The Heaviside function acts as a dummy variable, allowing one to count
2
points i f they lie inside the radius, r. C(r) varies over the range 1 IN to 1. This
measure can be constructed using Euclidean distances or a non-Euclidean
measure such as the sum o f absolute values o f the component vectors ( M o o n
1987, 216). The former measure is used here. The fact that C ( r ) fails to vanish
measures the extent to which any one data point, X , , affects the positions o f
the other points (Nicolis and Prigogine 1989, 278). Hence the reference to
C(r) as the correlation integral.
This method is also known as the ball-counting method (Barnsley 1988).
Each point has a ball superimposed over it w i t h radius r. A l l the other points
in the phase space inside this radius are counted and stored. The radius is
incremented until all points lie inside the boundary. This radius defines the
upper l i m i t . The lower l i m i t for a particular dimension is UN, for individual
points must occupy a unique address in phase space. This process is repeated
for a number o f dimensions based on the construction o f the ordered n-tuple

Despite the similarity i n names, the correlation dimension is unrelated to


the statistical correlation coefficient commonly used in the social sciences.
The correlation dimension provides a quantitative description o f the rate o f
divergence o f nearby points i n the phase space, as opposed to the direction
and strength o f association between two separate variables in a linear relation­
ship.
For many attractors, the correlation dimension is known to scale as a
power law function. That is,

d
C(r) ~ r (8)

or, i n other words, the dimension o f the attractor is determined by the slope o f
In C ( r ) versus In r in a particular range o f r (Nicolis and Prigogine 1989;
Grassberger 1986). This relation is

In C ( r ) = d · In r. (9)

The procedure then is to calculate C ( r ) for radii varying from zero to that
sufficient to cover all points in the phase space for successively higher dimen­
sions. The relation between In C(r) and In r can be seen graphically in figure
8.8, where the correlation integrals for dimensions four through ten are plot­
ted for the phase portrait o f public opinion data related to Mondale's chances
o f winning the 1984 Democratic presidential nomination.
The slope, d, from equation 9 is estimated using bivariate regression
from the various dimensions over the regions in figure 8.8 that are approx-
Complexity in the Evolution of Public Opinion 181

J I | I I 1 I I I L

Log Radius

Fig. 8.8. C o r r e l a t i o n i n t e g r a l s f o r t h e M o n d a l e t i m e - s e r i e s d i m e n s i o n s
four to ten

imately linear. It is the range o f linearity that allows the inference o f the
dimension o f the attractor. The slope is calculated for each dimension. I f the
slopes reach a saturation point, that is, become parallel as the dimension
increases, then the saturation point is interpreted to be the dimension o f the
attractor (Nicolis and Prigogine 1989, 279; Grassberger and Proccacia 1983,
196; M o o n 1987, 233). Where the slopes fail to saturate and the correlation
integral scales according to the relation

In C(r) = d • In n, (10)

where n is the dimension in which the correlation integral is constructed, the


dynamical process is understood to be Gaussian white noise. This relationship
is represented by a 4 5 ° line in the plane. This clearly differentiates the two
processes that might be expected. Evidence o f the presence o f a strange
attractor is given by a series o f estimates for d that lie under a 45° line.
Figure 8.9 graphs the slopes for dimensions t w o through ten for the
phase portrait for Walter Mondale. The error bars show the range o f standard
errors for the slope estimates. The slopes saturate or become parallel after
dimension five. The horizontal line drawn through the saturated values to the
182 Chaos Theory in the Social Sciences

vertical axis is the mean o f the slopes for dimensions six through ten. This is
the dimension o f the attractor. The value is approximately 3.58. Since this
value is noninteger, we make two inferences: the dynamics are governed by a
nonlinear process and the attractor is a strange attractor. This time series has a
fractal structure and it may be chaotic.
The dimension o f the attractor is between three and four and is non-
integer. The next-largest integer value greater than the dimension o f the
attractor—four—provides us w i t h knowledge o f the maximum number o f
dynamical variables needed to model the behavior o f the system (Schuster
1988, 128; Nicolis and Prigogine 1989, 279). The number o f variables re­
quired to completely explain the dynamics o f the entire system is less than the
number o f candidates that ran for the Democratic nomination. This insight
provides another perspective on candidate viability i n presidential nomination
contests. Unfortunately, no one has as yet devised a way to deduce which
variables are required for modeling the dynamical behavior, though some
imaginative ad hoc forecasting techniques have been applied w i t h limited
success (Crutchfield and McNamara 1987; Casdagli 1989). These efforts have
thus far failed to recover the functional form o f the original equations, even
where it is k n o w n a p r i o r i .
Complexity in the Evolution of Public Opinion 183

The Correlation D i m e n s i o n and t h e Effects o f Noise

Noise can be the bane o f statistical analysis. The survey used in this analysis,
the 1984 Continuous Monitoring Study, is known to contain more sampling
error than other National Election Studies (Allsop and Weisberg 1988). The
original survey contains about 11 percent error. I n the construction o f the time
series used in this analysis, the sampling error increases. The increase in the
error is approximately equal to the square root o f the sampling interval m u l t i ­
plied by the original sampling error. The original time series used a sampling
interval o f seventeen-day "weeks" in which the goal was to obtain two-thirds
of the interviews in the first actual week and use the remaining time to
complete the interview schedule (ICPSR 1984). The error in the collapsed
7
time series I use here is slightly more than . 4 5 . This is a large amount o f error
for any analysis and the question naturally arises as to whether or not the
results in the prior section can be trusted to identify a deterministic component
in the system dynamics.
In this section I report the results o f two Monte Carlo simulations on a
time series w i t h k n o w n dimension, where progressively larger error has been
added to the system dynamics in an attempt to destroy the resolving power o f
the correlation integral. The results show that the method is robust.
The Monte Carlo simulation uses the model constructed in equation 6 as
the baseline, w i t h slight modifications. That model is a linear harmonic os­
cillator in t w o variables. It oscillates in the phase space, tracing a circle. A
circle is a closed line and a line is a one-dimensional geometric figure. Hence,
the dimension o f the experimental time series is k n o w n . Equation 6 was
iterated 1,500 times to obtain a series o f long length and the output was scaled
into the unit interval so as to be able to easily calculate the size o f the error
component. Random normal error with k n o w n variance and mean zero was
added to each observation. Ten computations o f the correlation integral from
8
the X-Y phase space were conducted for each level o f added error. Summary
results o f this experiment are shown in table 8.2.
The results i n the table show that this measure is robust in the presence o f
a normal error component. The subset o f the points used to estimate the
correlation dimension were selected using the method described by Chen
(1988). The linear region o f the correlation integral is identified by plotting
the log o f the radius, r, against the ratio formed by dividing the log o f the
radius into the change i n adjacent points in the correlation integral, that is,
plot [ A C ( r ) / A l o g ( r ) ] · log ( r ) . The linear region is identified as a plateau or flat
segment in the modified plot (see Chen 1988; Abraham et al. 1986; M o o n
9
1987). I n the simulations run with 50 percent error added to each observa­
tion, the estimate for the correlation dimension is, on average, only 2.3
percent larger than the true dimension. This result is consistent with Schaffer
184 Chaos Theory in the Social Sciences

TABLE 8.2. M o n t e Carlo S i m u l a t i o n of t h e Effect o f Error


on Estimate of the Correlation Dimension

Beta Std. Error Error Variance Percent Error

.991 .142 0.02 7%


.9944 .222 0.03 14%
1.004 .258 0.04 18%
1.005 .112 0.05 25%
1.016 .171 0.06 30%
1.016 .247 0.07 42%
1.023 .269 0.08 50%

Source: Author's computations.


Note: N = !,500 for all cases, all entries are means, statistics computed over ten
runs.

(1988). The results i n the table show that the standard error o f the estimates
increases as the amount o f error increases.
The simulation above is based on a time series o f length 1,500. Next I
report the results o f a second simulation where the error components remain
constant and the sample sizes increase. This shows the effect o f noise in small
samples. Two runs are reported. The first has 15 percent error and the second
has 50 percent error. The series lengths are 200, 300, 400, . . . , 1,000, for
each run. Table 8.3 contains the results.
This simulation is based on one run per series. The results show that this
technique allows one to recover the k n o w n dimension o f the linear harmonic
oscillator quite accurately w i t h small sample sizes. These results are similar to

TABLE 8.3. M o n t e Carlo S i m u l a t i o n


of t h e Effect o f Error o n E s t i m a t e
of t h e C o r r e l a t i o n D i m e n s i o n w i t h
F l u c t u a t i o n i n S a m p l e Sizes

N 15% Error 50% Error

200 1.00 ± .011 .994 ± .038


300 1.01 ± .02 1.00 ± .033
400 1.02 ± .022 1.03 ± .024
500 1.01 ± .023 1.02 ± .030
600 1.01 ± .023 1.01 ± .027
700 1.01 ± .024 1.01 ± .027
800 1.004 ± .025 1.01 ± .024
900 1.02 ± .029 1.01 ± .023
1,000 1.01 ± .027 1.01 ± .020

Source: Author's computations.


Complexity in the Evolution of Public Opinion 185

a simulation reported by Schaffer (1988) for the logistic equation with varying
sample sizes (see also Albano et al. 1987). I n both cases the standard errors o f
the estimates are larger w i t h small sample sizes. The larger the sample, as
always, the more accurate the estimate.
The results o f this simulation give added weight to the results from the
aggregate time series. The ability to recover the true dimension is not com­
pletely destroyed by having a short or noisy time series. Can the movement in
opinion in the primary period be explained by a simpler model? I test this
hypothesis i n the next section.

A Test of an A l t e r n a t i v e M o d e l

It can be argued that the results derived in the previous section are the result o f
some unknown nonlinear yet regular process similar to the example used in
equation 6, where the deterministic relation was based on a linear harmonic
equation. A n even stronger argument might begin by positing that the ob­
served relationship is actually explained by a linear model. A test o f an
alternative model is used to demonstrate that the empirical results detailed
above are not due to noise that happens to follow an unusual cyclical pattern.
When coupled w i t h the Monte Carlo results reported above, this provides
further evidence for a low-dimensional deterministic process at work.
Experimental work on physical (Ben-Mizrachi et al. 1984; Ott et al.
1985) and economic time series (Scheinkman and LeBaron 1989) containing
an error component has revealed that the correlation dimension can be recov­
ered when noise is present in the data.
Ott et al. describe how an attractor's volume changes with the addition o f
an error component to a time series. I f e, represents the error component o f the
series, the presence o f noise in a dynamical system "fattens up the attractor by
an amount proportional to e" (1985, 65). This can be seen by referring back
to figures 8.3 and 8.6. The equilibrium in figure 8.3 is "fattened" in figure 8.6
by the addition o f the random error component.
This has the additional effect o f further separating nearby points from
one another. I n the context o f calculating the correlation integral, this addi­
tional separation results in fewer points falling inside small radii, r , which
means that larger and larger radii w i l l be required to discover the upper l i m i t
required to cover the entire set o f points in the phase space. Noise has the
effect o f increasing slightly the calculated dimension o f the attractor using
the correlation integral (Abraham et al. 1986). The estimated slopes over the
correlation integral w i l l , i f the system is dominated by noise, fail to saturate.
Thus, in the presence o f noise, the point at which the slopes o f the various
correlation integrals saturate w i l l be marginally greater than the saturation
point in a system without noise.
186 Chaos Theory in the Social Sciences

However, i f the level o f noise fails to swamp the deterministic structure


in the data, the correlation integral w i l l reach a saturation point and it is
expected that this point, measuring the dimension o f the attractor, w i l l exceed
the original dimension due to the increased proportion o f the series that is
noise.
To illustrate this, an A R 2 model was estimated for the Walter Mondale
series. A n A R 2 model is used because the time path o f such a model can
contain an oscillatory component (Brock 1986). The residuals were computed
and used in a calculation o f a second correlation integral (Brock 1986, 182—
10
84). The A R 2 model explained approximately 45 percent o f the variance. I f
the amount o f noise in the data is at all large and the deterministic component
small, the expectation is that only noise remains to be used in calculating the
correlation integral.
It is k n o w n that the level o f noise in the Rolling Cross Section is large
compared to other NES/SRC surveys o f the electorate. Therefore, it is ex­
pected that saturation, i f it occurs at a l l , should require testing across many
dimensions o f the phase space.
The correlation dimensions were calculated for dimensions two through

Fig. 8.10. S l o p e s a t u r a t i o n i n l o w d i m e n s i o n f o r t h e M o n d a l e series


based o n t h e residuals f r o m an AR2 m o d e l
Complexity in the Evolution of Public Opinion 187

ten, as before. Estimates for the slopes were derived as before, and are shown
in figure 8.10. The results clearly show that the deterministic component in
the data remains after use o f the A R 2 model. The correlation integral saturates
at approximately 3.70, the mean slope over the last five dimensions, and the
standard errors o f these estimates overlap the standard errors o f the slopes
based on the original Mondale time series.
The fact that the linear model fails to account for the deterministic
component in the model coupled w i t h the repeated result o f saturation in the
slopes over the scaling interval i n the correlation dimensions based on the
residuals from the A R 2 model is strong evidence that the time series contains
a nonlinear deterministic component. It remains to be shown that this series is
chaotic. Chaos is confirmed by the presence o f a positive ( > 0 ) Lyapunov
exponent. The next section confronts this problem.

L y a p u n o v Exponents and t h e Presence


of Chaotic D y n a m i c s

The evidence provided to this point indicates that the aggregate time series
constructed from the 1984 Rolling Cross Section may be chaotic. Confirma­
tion o f chaos requires a positive Lyapunov exponent. Lyapunov exponents
measure the rate o f divergence o f nearby points during the evolution o f a
dynamical system. This section discusses the technique o f estimating
Lyapunov exponents from an experimental time series, shows that the tech­
nique is robust in the presence o f large amounts o f noise, and estimates a
positive Lyapunov exponent for the constructed time series.

Estimating Lyapunov Exponents from Experimental


Time Series

The Lyapunov exponent measures the rate o f orbital convergence or d i ­


vergence i n a dynamical system. Positive exponents indicate divergence and
chaos while negative exponents indicate convergence or stability (Wolf 1985,
1986). I f the governing equations for a dynamical system are k n o w n , the
exponents may be derived explicitly and easily; hence, when the equations are
k n o w n , it is simple to determine i f the structure o f the dynamical motion is
chaotic or not. When an observed time series is all that one has, the problem is
much more difficult. The logistic equation is a useful vehicle for illustrating
the properties o f a positive Lyapunov exponent.
The logistic equation w i t h driving parameter r = 3.7 is known to be
chaotic. I f we write the equation in the form

n+\
X r
' n
X
' 0 X
n) (11)
188 Chaos Theory in the Social Sciences

and investigate error propagation using linear stability analysis, it can be


shown that nearby initial conditions diverge after only a few iterations. Fol­
l o w i n g W o l f (1986), define the error in measuring x to be dx„, so that n

x„
+i + dx n+l = f[x„ + dx ] n

~ f[x ] n + dx„ • f'lxj

dx i n+ = dx„- f [x ] n

= dx n · r(l - 2x„).

I f one uses the initial uncertainty, which amounts to a small measurement


error, one can write

n-l

\dx„\ = \dx \ • in= 0 !/'(*,·)!


0

(13)

\dx \ • 0 fi |K1 - 2x . i

1= 0

For any value o f r, the uncertainty grows i n time i f the product o f the local
stretching, | r ( l — 2x,)|, exceeds one. I f equation 13 is consistent w i t h

dx„ = dx 0 • 2*n (14)

the uncertainty grows exponentially fast. Here A is the Lyapunov exponent.


From this equation, one can easily see how convergence and divergence
occur. I f A < 0 then nearby initial conditions converge exponentially fast.
This is the situation i n a stable l i m i t cycle w i t h a perturbation. I f A = 0 then
the dynamics are neutrally stable and nearby initial conditions remain sepa­
rated by the identical distance throughout the evolution o f the dynamics. This
occurs i n equilibrium cycles. I f A > 0 then nearby initial conditions diverge
exponentially quickly and the system w i l l display chaotic motion. Equation
14 requires that

n- 1

A = lim 1 2 log |/'(*,)!·


2 (15)

Note that logarithms are reported in base t w o . I n practice, this means that the
Lyapunov exponent, A, is reported in bits per iteration.
Complexity in the Evolution of Public Opinion 189

Generally it is practical to estimate only the largest Lyapunov exponent


from an experimental time series (see Bryant et al. 1990 for an algorithm for
simultaneously estimating all Lyapunov exponents from a time series). Since
the Lyapunov exponents are simply generalized eigenvalues that give average
rates o f expansion or contraction on an entire attractor, one begins the estima­
tion process by reconstructing the attractor as described above. After recon­
structing the attractor, one searches for points that are close to each other in
the phase space, but separated by at least one orbit. The t w o trajectories, the
fiducial and the test trajectories, are followed for a predetermined time inter­
val or until the distance between them exceeds some specific value (Schaffer
1988; W o l f 1986). Then one searches for another point close to the fiducial
trajectory (a new test trajectory) and repeats the procedure. I n this way the
search process continues until the entire time series is exhausted. Some o f the
paired trajectories w i l l diverge and some w i l l converge toward each other.
The average o f the stretchings and contractions is taken at the end o f the
process. This procedure describes W o l f ' s Fixed Evolution T i m e algorithm and
it has been used effectively to estimate the Lyapunov exponents for a number
of dynamical systems, including the logistic equation, the Rossler equations,
the Lorenz equations, and the Henon map (Wolf et al. 1985; W o l f 1986).

As shown in equation 15 above, A is based theoretically on an infinitely


long time series. N o experimental time series approaches this length, so a
question naturally arises regarding the relationship between the length o f the
experimental series and the accuracy o f the estimated Lyapunov exponent. In
1
table 8.4 I show the result o f an experiment based on the Henon m a p , ' where
the length o f the series from which the Lyapunov exponent is estimated
decreases from 500 to 50 observations. Wolf et al. (1985) report that as few as

TABLE 8.4. C o m p u t a t i o n of t h e Largest


L y a p u n o v E x p o n e n t f o r t h e Henon M a p

N Exponent Time Delay Evolution Time

500 .4661 1 2
450 .4021 1 2
400 .3622 1 2
350 .4092 1 2
300 .4004 1 2
250 .4460 1 2
200 .4737 1 2
150 .3692 1 2
100 .4511 1 2
50 .3852 1 2

Source: Author's computations.


190 Chaos Theory in the Social Sciences

128 points were needed to calculate the largest (positive) exponent for this
attractor to w i t h i n 5 percent o f its value. Hence, the data requirements appear
to be less rigorous here than for the computation o f the correlation integral.
Note that the estimates for the Lyapunov exponents are consistently
greater than zero. I n this simulation I used one time series and drew increas­
ingly short samples from it in order to compute the Lyapunov exponent. W o l f
et a l . (1985) report on similar experiments for a number o f other attractors: the
Rossler attractor, the Lorenz attractor, hyperchaos (a 4 - D variant o f Rossler's
system), the delay differential equation, and others. Their results indicate that
the data requirements for difference equations are far shorter than for systems
of differential equations. The time delay is the lag used to construct the phase
portrait. Evolution time refers to the number o f iterations o f the fiducial and
test trajectories at the end o f which the paired trajectories are measured.
Noise i n the data poses no problem here. W h i l e adding noise to every
observation fattens up the attractor, noise o f the standard normal variety
moves points about in random, equiprobable directions. When computing the
Lyapunov exponents, these variations tend to cancel out and the resulting
estimates remain positive and far from zero. This is demonstrated i n another
simulation by constructing a new time series based on the Henon map but in
which a large random error component is added to each observation. The error
amounts to 50 percent. The results are shown in table 8.5.
These results indicate that error does not interfere greatly with the struc-

TABLE 8.5. C o m p u t a t i o n of t h e Largest


Lyapunov Exponent for the Henon Map
w i t h 50 Percent Error

3 b
N Exponent Time Delay Evolution Time

500 .4057 1 2
450 .4944 1 2
400 .3760 1 2
350 .4236 1 2
300 .3762 1 2
250 .3555 1 2
200 .3280 1 2
150 .4640 1 2
100 .4354 1 2
50 .3371 1 2

Source: Author's computations.


a
The time delay is the lag used to construct the phase portrait.
b
Evolution time refers to the number of iterations of the fiducial
and test trajectories at the end of which the paired trajectories are
measured.
Complexity in the Evolution of Public Opinion 191

ture o f the attractor and i f chaos is present it can be detected. Table 8.5 is
based on an error component that exceeds the error component in the aggre­
gate time series based on the 1984 Rolling Cross Section. It seems reasonable
to proceed on the assumption that i f a positive Lyapunov exponent exists in
the time series o f candidate chances, it w i l l be detected.

A p p l i c a t i o n t o the Constructed T i m e Series

The correlation integral informs us that the dimension o f the attractor for the
Mondale (and the Jackson) time series is between three and four. We make use
of this information in the computation o f the Lyapunov exponent by comput­
ing the magnitude o f the exponent for a range o f evolution times and time
delay reconstructions o f the attractor for dimensions three and four. I n this
case a small amount o f information can dramatically reduce the amount o f
w o r k necessary.
The time series I have constructed is based on daily data. It is a mapping,
like the Henon map used above. Hence the appropriate time delay for recon­
structing the attractor is one, just as for the Henon map or the logistic map.
What must vary is the evolution time for each choice o f dimension. I f the time
series is chaotic, we must expect that all o f the Lyapunov exponents computed
w i t h the fixed evolution time algorithm w i l l be positive. The results are shown
in table 8.6.
A l l estimates exceed zero, hence this time series is chaotic ( M o o n 1987,
12
191-92; W o l f et a l . 1985; Ruelle 1989; Baker and Gollub 1 9 9 0 ) . This
evidence also is an indication that the evolution o f public opinion during the
1984 presidential primary season is sensitively dependent on initial condi-
tions.

TABLE 8.6. L y a p u n o v E x p o n e n t s f o r t h e M o n d a l e T i m e Series,


D i m e n s i o n s T h r e e a n d Four

Evolution Time Lyapunov Exponent(3) Lyapunov Exponent(4)

2 0.1739 0.1255
3 0.1722 0.1350
4 0.1837 0.1202
5 0.1150 0.0871
6 0.1096 0.0833
7 0.0633 0.0826
8 0.0549 0.0493
9 0.0806 0.0445
10 0.0683 0.0862

Source: Author's computations.


192 Chaos Theory in the Social Sciences

In this result we find the reason that shifts in public opinion can occur in
dramatic fashion. The Lyapunov exponent is interpreted as a loss in predictive
power measured in bits per iteration. Large exponents, such as for the logistic
equation w i t h driving parameter equal to four, where the Lyapunov exponent
is one, mean that one bit is lost each iteration. I f one measures w i t h eight-bit
accuracy, all predictive power is lost after eight iterations.
By choosing the exponent for dimension four and a time evolution o f
two, we see that one bit is lost after thirteen iterations (days). Extend this to
six months and the result reported by James Johnson (in Moore 1986, 73),
Mondale's pollster in 1984, is clear: the predictive power has been lost as
people switch their candidate preference ordering. Effectively, such a long
forecast may as well be a roll o f a probability die.
Because the exponents in the Mondale series are relatively small, only a
small amount o f information is lost per iteration. The effective result is to
allow longer forecasts with greater accuracy and this effect is reflected in the
polling done by the Udall campaign in 1976 (Schram 1976, 126, 132). We
know that forecasts made close to an election date are more accurate than
those made long before the election. The presence o f positive Lyapunov
exponents explains w h y that result occurs.

Conclusions

Radical opinion shifts can occur in a short period during primary election
campaigns. This has been demonstrated with the Mondale time series con­
structed from the 1984 Rolling Cross Section NES survey. These shifts in
opinion can be induced by a response to the competition between candidates
for convention delegates.
The concept o f momentum is important here. M o m e n t u m can be positive
or negative (or zero) and the full range o f possibilities is shown in the time
series used here. Mondale w o n in Iowa, but Hart did better than expected and
reaped the electoral benefits in later elections. I n New Hampshire, Maine, and
Vermont, Hart captured ever-greater shares o f the vote. Mondale began to
recover on Super Tuesday and he then captured the higher percentage o f the
vote in a sequence o f elections. This long-term oscillatory pattern persists
throughout the campaign and the electoral results affect public opinion forma­
tion, perhaps by altering individual candidate preferences reported in the
survey. The survey design unfortunately did not ask respondents whether they
had altered their perceptions o f candidates' chances o f winning the nomina­
tion and each individual was only sampled once, so this hypothesis cannot be
confirmed. However, it seems very likely that respondents in this survey
reacted to electoral events in the same way that Mondale's deserters did—as a
reaction to Hart's ability to w i n votes when he was not supposed to.
Complexity in the Evolution of Public Opinion 193

What process is at work here? I show several linear models based on


simple dynamic models o f the primary electoral process (see A l d r i c h 1980).
The first t w o illustrate that simple models cannot capture the rich dynamics
observed i n electoral outcomes during the 1984 primaries. The last was posed
as an alternative to the hypothesis that the dynamics are governed by non­
linear equations. Each o f these failed to correctly characterize the observed
dynamics. The analysis eliminates the linear model as an option when model­
ing the dynamics.
A nonlinear process is at work here. The presence o f a positive
Lyapunov exponent is convincing evidence for chaos in this and any other
time series. As chaotic dynamics are associated w i t h deterministic processes,
I am led to conclude that such a process is at work here. This should not be
surprising. M a n y time series, most notably economic time series, are thought
to consist o f t w o components: a deterministic component and a stochastic
component. What this analysis shows is that the stochastic component may be
much less influential than previously thought.
The combination o f the several tools in this analysis simplifies the
model-building process. We know how many dimensions are required to
model the process. The correlation dimension gives the upper l i m i t on the
number o f equations required to model the dynamics—four. The Lyapunov
exponent confirms that the equations must be nonlinear—no linear process
can produce chaotic behavior. The remaining problem—and it looms large—
involves deducing a plausible mathematical model o f candidate competition
that reproduces the observed dynamics to within the error envelope.
Chaos has been confirmed by research in a number o f disciplines outside
physics (see Holden 1986; B a i - L i n 1990; for a thorough review o f the physical
evidence, see Jackson 1989, 1991). Biologists studying epidemics (Degn et al.
1987; Schaffer et al. 1986; Schaffer 1987; Schaffer et al. 1990) and population
biology ( M a y 1974, 1983, 1987) have developed theoretical models that can
display chaos and have discovered empirical evidence that supports theory.
Economists show a keen interest i n chaos. Studies o f stock market dy­
namics are intriguing for the simple reason that i f the stock market can be
shown to have a deterministic element, then the hope is that it can be con­
trolled or manipulated (see Brock 1986; Anderson et al. 1988; Boldrin 1988;
Brock 1988; Chen 1988; Eckmann et al. 1988; Casti 1989; Scheinkman 1989;
Scheinkman and LeBaron 1989; Brock et al. 1991). The article by Eckmann
et a l . (1988) is particularly interesting, as it estimates a positive Lyapunov
exponent for stock market data on gold futures prices. The magnitude o f their
estimate is comparable to the one reported here for public opinion over time.
Where should political scientists search for chaos? Whenever one has a
time series the possibility exists that the underlying governing equations are
nonlinear. For example, the political business cycle is a good candidate for a
194 Chaos Theory in the Social Sciences

nonlinear dynamical process precisely because no one can find a cycle. I n ­


stead, what is found is a pattern o f aperiodicity very similar to what is found
in economic, biologic, and physical systems exhibiting chaos. The political
business cycle links the president's popularity w i t h a number o f dynamic
variables, each o f which can be described by a separate equation (Hibbs
1987).
Presidential performance, as measured by the number o f bills sent to and
passed by Congress, provides a discrete time series that may be analyzed w i t h
these methods (Hammond and Fraser 1984). Additional work can be done on
the nomination process (see McBurnett 1990, 1991). We need not confine our
efforts to presidential nominations; Congressional as well as lower-level races
should provide fertile ground for research. Perhaps a systematic set o f govern­
ing equations may be discovered that operate across many electoral levels.
The search for the underlying equations o f motion that govern the behav­
ior o f the nonlinear systems is an area o f intense research activity (Crutchfield
and McNamara 1987; Casdagli 1989). Only recently have researchers discov­
ered how to recover the structure o f the governing equations from real data
(Mees 1992). This topic is beyond the scope o f this paper but clearly it is the
next step i n the analysis. We may need to adjust our perception o f the viability
of the linear model as a research tool. Linearity is rare in nature; the norm is
the nonlinear model (Devaney 1986; Casti 1988).
The techniques described and used in this paper are simple to apply,
require little subjective interpretation o f the results, and use no a priori k n o w l ­
edge o f the structure o f the dynamic to be investigated. The results from one
analysis can be used to narrow the range o f approaches taken i n later analysis
o f the same time series. The results are conclusive, whether the result is a
confirmation o f chaos or not. I suspect that many models w i l l be replaced w i t h
a deeper understanding o f the dynamical structure that exists in much o f
political behavior that changes over time.

NOTES

The present version of this chapter is much improved thanks to the insightful
comments and critique of Courtney Brown, Thad Brown, Jim Kuklinski, Mike
M a c K u e n , Steve Seitz, Seymour Sudman, and Dina Zinnes. Remaining errors should
be attributed to the author. The data used in this paper were made available by the
I C P S R . Neither the original collectors of the data nor the I C P S R are responsible for the
interpretations made.
1. This is also a serious problem for campaign managers and candidates. The
candidates rely on accurate professional polls that correctly identify the issues voters
believe are important. These data aid the candidates in the formation of policy state-
ments in their attempt to sway voters into their camps. When, unexpectedly, one
Complexity in the Evolution of Public Opinion 195

candidate begins to do "better than expected" according to election results, voters in


future elections may change their allegiance in favor of a candidate who is overtaking
the front-runner. This occurred in the 1984 Democratic nomination contest, and
Mondale's and Hart's pollsters documented Hart's surge in popularity after the Iowa
caucuses as well as his later decline (Moore 1986, 6 1 - 9 6 ) .
2. Though the attractor is clearly visible using the tenth lag of the time series, it
appears with lower order lags. Experiments indicate, with this simulation, that an
identifiable limit cycle emerges using as few as three lagged observations. The tenth
lag is used here so as to make the structure obvious. In practice, a sequence of graphs
might be used in which the lag structure is advanced by one unit of time per graph.
Thus the attractor, if it is present and of sufficiently low dimension, will quickly
become recognizable.
The elliptical shape of the attractor is partly a consequence of geometry and the
aspect ratio of the graphic display. If one can determine the period of the cycle, use it
(or the closest integer value) to set the lag, and plot K, against K , t p e r J o d , the result will
be a circle. Elliptical distortion, if any, will be due to a problem in the aspect ratio in
the output device. This also suggests a one-step method to identification of low-
dimensional attractors if the governing equations are known a priori.
3. In the analyses to follow, the dynamics of these time series are probed in the
phase space. To see how the phase space provides information on the dynamics of
these particular series, I refer the reader to chapter two of this volume, where I discuss
the spectrum for the time series, and the accompanying figures contained therein (figs.
2.10, 2.11, and 2.12).
4. According to Schuster (1988), "all strange attractors that have been found up to
now in dissipative systems have fractal HausdorfT dimensions." A dissipative system
is one in which the volume of the attractor shrinks, but not necessarily uniformly in all
directions. Further, "the property which makes the attractor strange is the sensitive
dependence on initial conditions," which means that the dynamics have a positive
Lyapunov exponent and this is a quantitative measure of the stretching and shrink-
ing of the volume of the attractor as the motion evolves on the attractor. In short,
Schuster's claim is that no strange attractors exist that do not also display chaotic
dynamics by virtue of the fact that fractal dimension and positive Lyapunov exponent
are inextricably intertwined. However, a pair of mathematical counterexamples have
been constructed of strange attractors that are not chaotic. These examples are contin-
uous time models (Grebogi et al. 1984). No counterexamples are found in discrete
time.
5. It does not matter which of the series is used; the results are the same. This entire
analysis was first conducted with the Jackson series and then replicated with the
Mondale series, on the suggestion of a reviewer. The results are the same.
6. The correlation integral is only defined in the limit. A s no experimental time
series approaches this length, it is customary to assume that the available data are
sufficient to support the necessary computation. 1 make this assumption here. Some
guidance as to the necessary length is available (see Abraham et al. 1986).
7. I thank Seymour Sudman for his help in deriving the nature of the increase in the
error in my construction of the time series.
8. The use of both variables in the computation of the correlation integral speeds up
196 Chaos Theory in the Social Sciences

the process considerably by reducing to one (from six or more) the number of dimen-
sions over which the calculation of the correlation integral takes place. This technique
(see Ben Mizrachi et al. 1984) is exactly equivalent to the procedure described in this
paper and the dimension estimates are the same. For example, using the technique
described in this paper, I estimated the dimension of the linear harmonic oscillator
using the time-delay method in a series of two hundred data points with no error and
the result was a set of six slope estimates of 1.00 with error less than 8.9e - 9 for each.
With 20 percent error added to the original series, the estimate for the correlation
integral converged to 1.008 ± 0.031 at the reconstructed five-dimensional phase
space. Because the multidimensional correlation integral is easier to interpret in the
context of the simulation, its use and the results are reported here. There is no
difference between the two measures. The results reported here are available from the
author.
9. A l l techniques presently available that allow a determination of the linear region
of the correlation dimension integral involve some intuition on the part of the re-
searcher. The main question, and it has yet to be resolved, revolves about identification
of the linear region. This technique tends to maximize the fit of the regression as
2
measured by R.
10. Solution of the resulting difference equation reveals no oscillatory motion. This
does not detract from the results that follow.
11. The Henon map is given by the coupled difference equations

2
Xn+I = 1 - aX + Y
n

Y n+l = bX , m

where a = 1.4 and b — 0.3, and X 0 = Y = 0.5. This map is the two-equation analog
0

to the logistic map.


12. A l l exponents remain positive when this table is extended to include evolution
times up to thirty.
Part 3
Chaos Theory and Economics
CHAPTER 9

Chaos Theory and Rationality in Economics

J. Barkley Rosser, Jr.

A central assumption in economic theory is that o f rational behavior by


economic agents, although this has been widely criticized by many nonecono-
mists. Since the work o f M u t h (1961), this assumption has increasingly taken
the form o f rational expectations, that agents over time on average accurately
predict the future. This somewhat simplistic formulation o f the assumption
depends on agents knowing the underlying "true" model o f reality and fully
utilizing information that, in turn, is assumed to arrive in a random pattern
that reflects a normal distribution. Agents may make errors from time to time,
1
but these errors tend to cancel each other out over t i m e .
M u t h (1961) originally applied the theory to explain behavior in agri­
cultural markets. Its use has since spread throughout economics at both the
microeconomic and macroeconomic levels, as well as into financial theory.
A m o n g the most controversial applications have been in macroeconomics
(Lucas and Sargent 1981), where models using the assumption have been used
to support a "New Classical" view that systematic stabilization policies by
governments w i l l be ineffective and can only cause inflation in the long run.
The essential argument is that economic agents w i l l rationally forecast the
effects o f government policies and w i l l act in ways that nullify those effects.
The assumption has acquired the status o f a Lakatosian hard core axiom
of the economic research program (Rosser 1993), despite some empirical
findings that it does not hold for a variety o f economic agents (Lovell 1986).
The argument for this hard core axiom view has been articulated by Sargent
(1982, 382), who has asserted that it is a hypothesis not amenable to empirical
testing because o f "the logical structure o f rational expectations as a modeling
strategy, the questions that it invites researchers to face, and the standards that
it imposes for acceptable answers to those questions."
A significant complication for the use o f this assumption has arisen in the
form o f the increasing awareness o f the widespread reality o f nonlinearity i n
many dynamical economic systems. Such nonlinearity can lead to chaotic
2
dynamics in economic systems. A number o f theoretical models have been

199
200 Chaos Theory in the Social Sciences

developed in different areas o f economics that assume rational expectations


but can generate chaotic dynamics for certain parameter values.
This is ironic because in a situation o f chaotic dynamics, the condition o f
sensitive dependence on initial conditions holds. This implies that even a
minute error in estimation can lead to serious errors in long-run forecasting.
This raises very serious doubts about the realistic applicability o f the rational
expectations assumption under such circumstances. It is this contradictory
situation that is the central theme o f this chapter.

Chaotic M i c r o e c o n o m i c M o d e l s w i t h
Rational Expectations

Chaotic Preference Cycles

Generally, in microeconomic analysis it is assumed that individual prefer­


ences are constant and that a person's behavior reflects their responses to
changes in relative prices or income. However, it has been recognized since at
least Pareto (1909) that this is a highly unrealistic assumption, and some
economists have occasionally attempted to develop models that dynamicize
preferences. Such an effort that is both consistent with rational expectations
and also allows for chaotic patterns o f individual consumption behavior is due
to Benhabib and Day (1981).
They propose that a person's preferences reflect past consumption pat­
terns and that such relationships may generate cyclical patterns o f preferences
and consumption as between two goods. Thus people may alternate between
ski and beach holidays. Letting the goods be x and y , w i t h m the person's
money, they assume that the person's utility from consuming the goods can be
represented by the standard Cobb-Douglas utility function,

U(x,y) = xayx— a (i)

They hypothesize t w o different dynamic processes that both operate


through changing the value o f a over time. The first one, which is analogous
to the Lorenz (1963) climate model, is given by

(2)

w h i c h , in turn, implies a dynamic demand curve

x t + i = oanx,(jn — x,). (3)

2
The control parameter for this process is am . A t values slightly greater than
one there is a stationary demand curve. As its value increases, period-
Chaos Theory and Rationality in Economics 201

doubling bifurcations emerge w i t h the resulting emergence o f cyclical con­


sumption behavior. Chaotic dynamics appear for values greater than 3.57.
The second process is given by

(4)

which implies a dynamic demand function

xt+l = ffu e
f
a<1
*>> (5)

I f m = 1, this is analogous to a population dynamics model due to M a y and


Oster (1976) for which a is the control parameter. Chaos appears when it
approximately equals 2.6924.
Both o f these processes imply that as a declines, higher values o f m are
required for chaotic dynamics to occur. Given that a is the experience depen-
dent parameter and m is income, Benhabib and Day (1981, 463) conclude
that "the t w o models then characterize experience dependent demand as con­
verging to a stable long run pattern for relatively low incomes, but exhibit
increasing instability and eventually become completely erratic as income
grows reflecting the whimsical, seemingly arbitrary behavior o f the compla­
cent, or the very r i c h ! "

Interpersonal Utility Effects and Chaotic Dynamics

Another realistic but infrequently allowed aspect o f individual preferences in


microeconomic analysis is the sociological observation that a person's prefer­
ences may depend on the behavior o f others, an observation originally made
by Veblen (1899) and later studied by Leibenstein (1950). I n particular the
latter distinguished between bandwagon effects, wherein one wants to buy
something when others are buying i t , and snob (or reverse bandwagon) effects
wherein one does not want to buy something when others are buying i t .
Granovetter and Soong (1986) postulate that a person could exhibit both
effects w i t h respect to a particular good for different levels o f market satura­
tion. Thus someone might be inclined to buy a push-button phone when 20
percent o f their friends do so, for bandwagon reasons, but may incline toward
an old rotary phone when 90 percent o f their friends have push-button ones,
for snob reasons. Such a situation implies a nonlinear dynamic demand pro­
3
cess and i f it is sufficiently nonlinear the dynamics can be chaotic. Figure
9 . 1 , taken from Granovetter and Soong 1986, shows the demand curve in
such a situation. As price, P, declines below />,, period-doubling bifurcations
appear. P* is the threshold o f chaotic dynamics.
Iannacone (1989) has criticized this model for having consumer prefer­
ences depend only on the behavior o f others during the most recent period. He
202 Chaos Theory in the Social Sciences

Quantity

Fig. 9 . 1 . C h a o t i c d y n a m i c d e m a n d process. (From Granovetter a n d


S o o n g 1986. R e p r i n t e d w i t h permission.)

shows that an adaptive expectations strategy drawn from Heiner 1989 wherein
consumers respond to a weighted average o f past behavior by others reduces
the likelihood o f chaotic dynamics. This is an argument we shall consider in
more detail later.

Chaotic C o b w e b s

Originally conceived by Cheysson (1887), cobwebs are oscillatory and possi­


bly explosive dynamics in markets that can arise when there are lags in the
production process, the classic example being i n agriculture. Ironically, the
original motivation behind Muth's (1961) development o f the rational expec­
tations assumption was to rule out such outcomes for agricultural markets. He
argued that farmers would rationally forecast the true long-run equilibrium
and w o u l d not foolishly indulge in alternating increases and decreases in
output in response to short-run price fluctuations.
Chaos Theory and Rationality in Economics 203

However, it has since been shown that oscillatory and even chaotic
dynamics can emerge because o f production lags i f supply or demand curves
are nonlinear and nonmonotonic, even w i t h rational expectations. Artstein
(1983) posited partially downward-sloping supply curves for agricultural
commodities i f farmers (rationally) forecast that very low prices in one year
w i l l trigger government price supports in the next year. Artstein shows that
such a case can generate a three-period cycle, thus implying chaotic dy­
namics. Jensen and Urban (1984) derived similar results for models w i t h
backward-bending supply curves (not uncommon in labor markets and fish­
eries) and double-valued demand curves (much rarer, perhaps Irish potatoes in
the mid-nineteenth century).
Chiarella (1988) showed the possibility o f chaotic cobwebs even with
normally shaped supply and demand curves. However, his results are not
consistent w i t h rational expectations because they depend on the existence o f
a lag in expectations formation as well as a lag in production, realistic as that
4
may b e . This model o f Chiarella's underlies the empirical studies o f the pork
and m i l k cycles by Chavas and Holt (1991, 1993).

Chaotic M a c r o e c o n o m i c M o d e l s
w i t h Rational Expectations

Overlapping Generations Models

The first macroeconomic models to exhibit chaotic dynamics that are consis­
5
tent w i t h rational expectations (Benhabib and Day 1980, 1982), labeled
"Strong New Keynesian" by Rosser (1990), depend on the use o f the over­
lapping generations concept developed by Allais (1947) and Samuelson (1958).
In each period o f time let there be two generations alive, w i t h the young
possessing endowments o f w and the old possessing endowments o f w and a

the o l d possessing fiat money at the beginning o f time equal to M. Let p(t),
c (t), and c {t) be, respectively, prices, consumption by the young, and con­
y a

sumption by the old in time t.


The young seek to maximize utility

max U[c (t), y c (t + 1)],


a (6)

subject to their intertemporal budget constraint

p(t)c (t)
y + p(t + \)c {t
a + l)=p(t+ l)w .
0 (7)

These generate an intergenerational offer curve, O, by the young as p{t) and


p(t + 1) vary. The old in time 1 face the constraint
204 Chaos Theory in the Social Sciences

p(l)c (l)
0 = p(l) Wo + M. (8)

Given a Ricardian intertemporal production possibilities frontier,

R = [(c , c ) : c
y a y + c
a = w y + wj, (9)

a set o f perfect foresight (hence rational expectations) equilibria are given by a


sequence o f prices and consumption levels satisfying equations 6, 7, and 8.
The pattern generated by this sequence w i l l depend on the "humpiness"
of the intergenerational offer curve, O. Sufficient humpiness can lead not
merely to endogenous cycles, but to chaotic cycles such as that shown in
figure 9.2. Benhabib and Day (1982) show that this humpiness w i l l depend
positively on the variability o f the constrained marginal rate o f substitution
between individuals' present consumption and future consumption. The pre­
cise sufficiency conditions for chaos also depend on the endowments o f the
young and on the rate o f population growth. They also show that these chaotic
6
trajectories are efficient, that is, Pareto-optimal.

Cy
Fig. 9.2. Chaotic i n t e r g e n e r a t i o n a l offer curve
Chaos Theory and Rationality in Economics 205

A somewhat different overlapping generations perfect foresight model


that can generate endogenous chaotic dynamics is due to Grandmont (1985).
His model focuses on interest rate changes initiating conflicts between inter­
temporal wealth effects and intertemporal substitution effects. I f older agents
have a marginal propensity to consume leisure sufficiently greater than that o f
young agents, then chaotic dynamics can emerge. More specifically, let a be
real wealth and V the indirect utility function o f either generation. Then, the
Arrow-Pratt relative degree o f risk aversion, which measures the curvature o f
V, is

R(a) = -V"(a)a/V'(a). (10)

7
Chaotic cycles w i l l occur i f R (a ) > R (a ).
0 0 y y

Figure 9.3 illustrates such a case derived from Grandmont's (1985)


model. The axes are real money balances in time t, fx, on the horizontal and
the perfect foresight dynamics o f real money balances in time t + 1, x(fi), on

Fig. 9.3. Perfect f o r e s i g h t m o d e l i n o v e r l a p p i n g g e n e r a t i o n s with


e n d o g e n o u s c h a o t i c d y n a m i c s . (From G r a n d m o n t 1985.)
206 Chaos Theory in the Social Sciences

the vertical. The humpiness o f x(ix) increases w i t h R (a ).


0 0Grandmont argues
that Keynesian-style governmental policies o f proportional intergenerational
money transfers can pin down expectations about the interest rate and e l i m i ­
nate any cyclical outcomes. However, this depends on both the government
being able to make accurate forecasts under conditions o f chaotic dynamics
w i t h the attendant problem o f sensitive dependence on initial conditions and
people having complete confidence in government policies.
Brock (1988a) has dismissed the realism o f such assumptions as amount­
ing to a "nirvana fallacy." Indeed, D w y e r (1992) has developed a model
wherein stabilization policy generates chaotic dynamics in a situation when
otherwise such would not occur. This echoes monetarist critiques o f discre­
tionary Keynesian policies as generating business cycles when otherwise they
w o u l d not occur. Such an argument has been reinforced by DeCoster and
M i t c h e l l (1992), w h o show that a chaotic policy w i l l be magnified into even
more chaotic dynamics, even when the economy is basically nonchaotic and
reflects linear rational expectations.

Discounting and Chaos in Infinite Horizon Models

Strong advocates o f the rational expectations approach object to the kinds o f


models presented in the previous section on the grounds that they either do
not involve optimization over an infinite time horizon (Benhabib and Day
1982; Grandmont 1985) or involve some kind o f incompleteness o f markets
(Woodford 1989). Thus it is argued that they do not represent truly complete
general equilibrium solutions. Defenders o f the above approaches may cite
their apparently greater realism, but the critics note that agents may well have
time horizons beyond personal lifetimes, either for family bequest motives
(Barro 1974) or because important agents may be transpersonal entities such
as corporations or governments w i t h presumably very long time horizons.
Thus the question arises as to whether or not chaotic dynamics can arise in
such models w i t h completely general infinite horizon equilibria. The answer
is yes.
One way that chaos can appear in such models is i f there is a failure to
converge to a steady-state equilibrium, even though one exists. For optimal,
infinite horizon, multisector models, Sutherland (1970) initially showed that a
low discount factor (high real interest rate) could lead to such noncon-
vergence. Boldrin and Montrucchio (1986) established that in such situations
"every possible" kind o f behavior, including chaotic dynamics, is possible.
B o l d r i n and Woodford (1990) noted that in the simplest case the discount
rates involved imply 10,000 percent real interest rates per period, clearly
ridiculous. However, alterations to the model can allow for the phenomenon
w i t h i n more realistic ranges (Neuman, O ' B r i e n , Hoag, and K i m 1988). I n
particular, Boldrin (1989) has shown that increasing the number o f substi-
Chaos Theory and Rationality in Economics 207

tutable factors and the ranges o f their respective substitutabilities can allow
for chaotic dynamics within realistic discount ranges.
Deneckere and Pelikan (1986) have considered further factors beyond the
discounting issue that can also bring about chaotic dynamics in one-good,
two-sector, optimal infinite horizon, complete markets models. These include
the presence o f strongly decreasing returns to scale and the possibility that the
relative capital intensities o f the two sectors reverse order as the level o f
production rises. This latter condition is key to the arguments o f Boldrin
(1989) noted above.
Deneckere and Pelikan take the strong position that i f optimal behavior is
truly chaotic it w i l l never be observed. This is because the learning processes
necessary for perfect foresight (and the use o f rational expectations more
generally) can only operate in relatively regular environments. Sensitive de­
pendence on initial conditions w i l l destroy this possibility and w i l l also render
impossible any Grandmont-type optimal governmental policy intervention as
well.

Nonseparable Utility and Chaos


in Infinite Horizon Models

Yet another case i n which chaotic dynamics can arise i n optimal infinite
horizon models w i t h complete markets is when real money balances enter
directly into agents' utility functions in a manner nonseparable from consump­
tion goods. Matsuyama (1991) has examined this case. Let p be prices, c be
consumption, and m be real money balances, which grow at a rate u,. Let B be
the discount factor and let 8 = /JL/(B - 1) > 0. Let 8 be the elasticity o f
intertemporal substitution o f real balances, w i t h 17 being a parameter 3: —1
such that 8 = (17 4- 2 ) . Then Matsuyama assumes a utility function o f the
_ 1

form

uic, ) m = \; ^\ rr
[ m
l In g(c) + In m,
1(1 +1)1
i f 17 = - 1 ,
(i i)

w i t h g > 0 and g' > 0.


This implies an optimal time sequence o f prices given by

(p, )n
+l = (1 + - p,). (12)

Matsuyama proceeds to demonstrate that the nature o f these sequences


w i l l vary according to combinations o f 8 and 17. A n important variable deter­
mining the nature o f these sequences is

A(TJ) = TJ-^I + i7) 1+1


i - 1. (13)
208 Chaos Theory in the Social Sciences

In particular, he shows that for any TJ > 0, there exists a value A*(-n) satisfy­
ing 2 r] < A*(T/) < A(TJ) such that a period o f cycle three and hence chaotic
dynamics exists i f 8 > A*(TJ). These results and Matsuyama's more general
categorization o f possible outcomes are summarized in figure 9.4.
Matsuyama argues that this model suggests the possibility that even for
comparative statics analysis the set o f possible equilibrium prices may be
"topologically complex." This result is probably even more destructive o f
standard views o f economic theory than is the associated undermining o f the
usefulness o f the rational expectations hypothesis.

Empirical Evidence

M u c h empirical evidence has been gathered that is suggestive o f the possible


existence o f deterministic chaos in various economic time series. However,
there has yet to be a definitive demonstration o f such existence in any case.
W h y is this?
Brock and Sayers (1988) have presented a methodology for searching
empirically for chaos. The first step is to test for dimensionality and associ-

Fig. 9.4. M a t s u y a m a ' s t o p o l o g i c a l c o m p l e x i t y i n possible e q u i l i b r i u m


prices
Chaos Theory and Rationality in Economics 209

ated nonlinearity. Brock, Dechert, and Scheinkman (1987) have developed a


test called the B D S statistic that is based upon the use o f the Grassberger and
Procaccia (1983) correlation dimension. This test has been used on a variety
o f economic time series and has convincingly shown the presence o f serious
nonlinearities in many o f them. Nonlinear dependence has been found in
monetary aggregates (Barnettand Chen 1987), U . S . stockreturns (Scheinkman
and LeBaron 1989), gold and silver markets (Frank and Stengos 1989),
U . S . Treasury b i l l rates (Brock 1988b), work stoppages (Sayers 1988),
employment, unemployment, and pig iron production (Brock and Sayers
1988; Sayers 1990), industrial production (Brock and Sayers 1988; Ashley
and Patterson 1989), Japanese G N P (Frank, Gencay, and Stengos 1988),
foreign exchange rates (Hsieh 1989; Papell and Sayers 1989), pork pro­
duction (Chavas and Holt 1991), and m i l k production (Chavas and Holt
8
1993).
The second step is to estimate Lyapunov exponents to determine i f the
greatest exponent has a positive real part, thus indicating the presence o f
sensitive dependence on initial conditions. Brock and Sayers (1988) argue
that although algorithms have been developed for estimating Lyapunov expo­
nents (Wolf et al. 1985; Eckmann et al. 1986; McCaffrey et ai. 1992), no
statistical inference theory is k n o w n , although this is disputed by Barnett and
H i n i c h (1993). Nevertheless, tentative indications o f the existence o f positive
real parts o f Lyapunov exponents have been found by Barnett and Chen
(1987) for monetary aggregates, by Brock and Sayers (1988) for employment
data, and by Eckmann et al. (1988) for stock returns. W h y then the continuing
empirical skepticism?
The argument has been made that i f a series is really driven by determin­
9
istic chaos then it should be possible to make short-period forecasts that beat
a random w a l k . A l s o , the dimensionality should be maintained, even as the
intervals o f measurement are shortened. Efforts to make such tests on some o f
the most promising series have not been favorable. Brock and Sayers (1988)
could not support the chaos result for any o f the series they studied. Efforts to
forecast using nearest neighbor techniques failed to beat a random walk for
foreign exchange rates (Diebold and Nason 1990) and stock returns (Hsieh
1991). Mayfield and Mizrach (1992) found changing the time interval
changed the dimension estimates for stock returns. However, none o f these
10
tests can be viewed as definitive and the question remains o p e n .
Indeed, the discussion o f the estimation issues involved has been much
longer and more complicated than we have indicated here. Let us note one
argument that may be o f relevance, namely that o f aggregation. Almost all
these tests have been performed on fairly highly aggregated data sets. How­
ever, there is reason to believe that even i f chaos exists at the specific or
micro-level, it may disappear as aggregation to the macro occurs. Sugihara,
210 Chaos Theory in the Social Sciences

Grenfell, and M a y (1990) found evidence o f chaotic dynamics in city-by-city


measles epidemics. However, upon aggregating to the national level the dy­
namics reduced to a noisy two-year cycle. A similar argument has been made
w i t h respect to the dynamics o f spatial economic systems: that the more lo­
cal they are the more likely they are to exhibit unstable cyclical dynamics
(Dendrinos w i t h M u l a l l y 1985). Thus it may well be that the future search for
economic chaos should focus more on highly disaggregated microeconomic
series.

The Problem for Rational Expectations

Even i f full-blown chaos has not been confirmed for any major macro-
economic time series, many o f the problematic aspects o f chaotic dynamics
still adhere to series that are highly nonlinear while not fully chaotic. We
return to our central contradiction. O n the one hand, the possible coincidence
of chaos and rational expectations could mean the possibility o f improved
forecasting. O n the other hand, the problem o f the sensitive dependence on
initial conditions means that introducing any stochasticity at all into such a
system probably ruins any hope o f improving forecasting ability and renders
11
meaningless the very idea o f rational expectations. DeCoster and Mitchell
(1992) have suggested a possible response to this problem in the form o f
running many simulations and using the expected value, although this is very
time-consuming and expensive.
This problem relates to an o l d issue that has plagued rational expecta­
tions models almost from their inception, namely, the possibility o f inde­
terminacy due to an infinity o f possible equilibria (Rosser 1991b). Such a
possibility was first shown by Gale (1973) and numerous efforts by New
Classical economists have been made since to demonstrate the primacy o f the
golden rule steady state that the economy is supposed to tend to in such
models. Generally the non-golden rule steady state equilibria have the qual­
ity o f being self-fulfilling prophecies, also known as "sunspot equilibria"
(Shell 1977).
A recent line o f arguments that has sought to rule out such extra equi­
libria has been to argue that they are not dynamically stable. I f one assumes
certain kinds o f learning processes then the system w i l l converge on the
golden rule steady state. Evans (1985) first presented such arguments and
Marcet and Sargent (1989) have further developed them by arguing for the so-
called least squares learning process that appears to have this property. But
other learning processes lead to different results. Grandmont (1985) has
shown learning processes that converge to cycles. Woodford (1990) has
shown learning that converges to self-fulfilling prophetic equilibria. Benassy
and Blad (1989) argue that learning w i l l not converge on rational expectations
Chaos Theory and Rationality in Economics 211

unless one starts w i t h rational expectations. Evans and Honkapohja (1992)


have found expectationally stable bubble equilibria, and even in his initial
study Evans (1985) noted that for certain parameter values o f his learning
process the result is not convergence but chaotic dynamics.
Is there a way out o f this indeterminacy? One such way is to abandon
rational expectations in general and to adopt a line o f reasoning mentioned
above due to Heiner (1989). He argues that when forecasting errors occur, as
we would expect i n a chaotic environment impinged upon by stochastic
shocks, agents w i l l use an adaptive expectations strategy by stabilizing their
behavior i n the face o f heightened uncertainty. He recognizes that for conver­
gence to the optimal trajectory to occur strong assumptions are necessary. But
under the right conditions such convergence w i l l proceed slowly at first, then
accelerate, and then slow down again as the goal is approached. Rational
expectations w i l l only hold in the equilibrium case. Otherwise, simplicity and
12
caution w i l l create order out o f chaos. The ultimate irony here is that the
deeper concept o f rationality may be preserved in such circumstances by
abandoning any hope o f following rational expectations.

Conclusion

So what is left o f the edifice o f standard neoclassical theory after the possi­
bility o f nonlinear and chaotic dynamics is recognized? As noted above,
M i r o w s k i (1990) has argued that in fact chaos theory actually reinforces
standard theory. However, this is certainly a minority view. As we have seen,
sensitive dependence on initial conditions is profoundly disruptive o f the
ability to develop rational expectations, especially when any stochastic shocks
are present. This is so, or perhaps especially so, when rational expectations
equilibria are chaotic, as argued above by Deneckere and Pelikan (1986).
Nevertheless, it probably remains the case that standard neoclassical
theory remains useful for many situations despite the difficulties raised by
nonlinear and chaotic dynamics. Let us conclude this discussion with a quota­
tion by Smale, one o f the initial developers o f chaos theory in mathematics, as
well as a student o f the structural stability o f Walrasian general equilibria:

How d i d Relativity Theory respect classical mechanics? For one thing


Einstein worked from a very deep understanding o f the Newtonian the­
ory. Another point to remember is that while Relativity Theory lies in
contradiction to Newtonian theory, even after Einstein, classical me­
chanics remains central to physics. I can well imagine that a revolution i n
economic theory could take place over the question o f dynamics, which
would both restructure the foundations o f Walras and leave the classical
theory playing a central role. (Smale 1977, 95)
212 Chaos Theory in the Social Sciences

NOTES

Figure 9.3 is reprinted from J . - M . Grandmont, "On Endogenous Competitive Busi-


ness C y c l e s , " Econometrica 53 (1985): 9 9 5 - 1 0 4 5 , with permission of The Economet-
ric Society, Department of Economics, Northwestern University, Evanston, I L .

1. The formal definition of rational expectations is that the subjective probability


distribution regarding future reality inside the heads of economic actors coincides with
the objective probability distribution operating outside their heads in reality.
2. Reviews of applications of chaos theory in economics are in Kelsey 1988;
Baumol and Benhabib 1989; Lorenz 1989; Boldrin and Woodford 1990; Rosser 1991a;
Brock, Hsieh, and Lebaron 1991; and Bullard and Butler 1993. For formal definitions
of chaotic dynamics, see chapters 2, 3, and 4 in this volume.
3. Such a model is potentially applicable to bandwagon phenomena in asset mar-
kets such as speculative bubbles. Models of speculative bubbles coinciding with
chaotic dynamics have been developed by Day and Hwang (1990); De Grauwe and
Vansanten (1990); Ahmed, Ayogu, and Rosser (1990); and De Grauwe, Dewachter,
and Embrechts (1993). However, none of these models allow for rational expectations
bubbles because of the boundedness conditions associated with chaotic dynamics.
4. Yet another area of microeconomics where chaotic dynamics models have been
developed that are inconsistent with rational expectations is in oligopoly theory, espe-
cially the case of duopoly dynamics. The first explicit application of chaos theory in
economics (Rand 1978) was in such a model.
5. Chaotic macroeconomic models not consistent with rational expectations have
been developed depending on nonlinear population dynamics (Stutzer 1980), nonlinear
productivity of capital in a Solow-type growth model (Day 1982), high exploitation
of labor in the convergence to a Ricardian steady state (Bhadurk and Harris 1987),
Lotka-Volterra employment dynamics with markup wages (Pohjola 1981), nonlinear
multiplier-accelerator relations (Gabisch 1984), nonlinear investment functions (Dana
and Malgrange 1984), Schumpeterian technological long waves (Goodwin 1986), and
long-wave socialist investment cycles (Rosser and Rosser 1994), among other ap-
proaches.
6. Such outcomes in some economics chaos models have led Mirowski (1990) to
criticize them as being sophisticated efforts to preserve neoclassical economics in the
face of actual extreme stochastic uncertainty.
7. Woodford (1989) has shown that similar results can be obtained, even in models
with infinitely lived agents, if they face short-run borrowing constraints.
8. This nonlinear dependence may take the form of nonlinear dependence in the
variance, or conditional heteroskedasticity, rather than in the mean of the series.
9. A very real possibility is a combination of chaos and stochastic shocks. However
such a combination will exactly resemble purely random processes because of the
sensitive dependence on initial conditions. Every shock will push the system onto a
different path.
10. E v e n though chaos has not been confirmed, nonlinearity most certainly has
been. A popular approach to analyzing it in stock and foreign exchange markets has
been to examine variations in volatility over time following a suggestion of Mandelbrot
(1963). Engle (1982) developed the autoregressive conditional heteroskedasticity
Chaos Theory and Rationality in Economics 213

( A R C H ) technique to estimate such effects. This has been expanded to generalized


A R C H ( G A R C H ) by Bollersley (1986), A R C H in mean ( A R C H - M ) by Engle, Lillien,
and Robins (1987), and nonlinear A R C H ( N A R C H ) by Higgins and Bera (1992).
11. This is not necessarily the case in the long run if there is only one equilibrium.
Intermediate run forecasting may be ruined, but the system may still oscillate around
some expected value in the long run, which is consistent with the idea of forecasting
errors being random.
12. Lavoie (1989) has used neo-Austrian theory to argue that out-of-equilibrium
chaotic dynamics will lead to self-organizing order in a manner similar to that de-
scribed in the nonequilibrium phase transition theory of Prigogine and Stengers (1984).
CHAPTER 10

Long Waves 1790-1990: Intermittency,


Chaos, and Control

Brian J. L . Berry and Heja Kim

There are t w o contending views o f economic dynamics. Insights into their


validity may be provided via concepts o f catastrophe and chaos. The d o m i ­
nant paradigm is that optimizing behavior produces an economy that is inher­
ently equilibrating, tending toward steady-state growth in the absence o f
random shocks. Such shocks, in the Slutsky-Frisch-Tinbergen view, are trans­
formed into cyclical oscillations through the filtering properties o f the econ­
omy's growth-propagation mechanisms (Boldrin 1990, 126-27). A con­
tending paradigm associated with Hicks, Kaldor, and Goodwin is that
endogenously driven growth is inherently unstable, with deviations that move
between upper and lower limits to investment and consumption (Boldrin
1990, 126-27). What we w i l l show in this paper is that long-wave rhythms o f
prices and growth (Berry 1991) accompanied by shorter-term oscillations
display unpredictable dynamic system order, or deterministic chaos. Phase
portraits reveal both intermittent crises at long-wave peaks and quasiperiodic
l i m i t cycles that encompass attractor basins between the peaks (for a discus­
sion o f l i m i t cycles see Berge, Pomeau, and Vidal 1984). It is the presence o f
these properties that supports the Hicksian view before 1919. We also show
that institutional changes since the Great Depression have reduced the limits
w i t h i n which prices and growth oscillate, significantly altering long-wave and
short-run dynamics by forcing attractor basins toward stabler equilibria.
Specifically, we document the following:

1. Slow-moving long-wave rhythms o f prices are the outcome o f the


half-century-long logistics by which new techno-economic paradigms
move from innovation to market saturation (Berry, K i m , and K i m
1993). The usual definition o f chaos results in instabilities at the top
of such logistics (Modis and Debecker 1992), and indeed they leave
in their wake the inflationary spirals that define long-wave peaks.
2. Faster-moving annual fluctuations oscillate around the long-run

215
216 Chaos Theory in the Social Sciences

rhythms o f prices, urged toward acceleration on the upwave and


deceleration on the down wave.
The tension between low-frequency rhythms and high-frequency os­
cillations produces chaotic l i m i t cycles in the periods between long­
wave crises.
Instability at the long-wave price peaks gives rise to bifurcations.
Upwave l i m i t cycles may be different from those observed during a
d o w n wave.
Similar bifurcations and l i m i t cycles characterize quarter-century-
long swings and annual oscillations in the rate o f economic growth.
Since World War I I , both institutional changes and policies designed
to fight inflation have dampened the higher-frequency oscillations o f
prices, leaving the long-wave rhythms as the dominant fluctuations:
the first return map o f prices has been reduced to the long-wave
attractor.
But contrariwise, stabilization programs aimed at fluctuations in
economic growth have had the opposite effect. It is the long swings
o f growth that have been dampened, leaving behind the higher-
frequency oscillations o f the business cycle: the first return map o f
rates o f economic growth has been simplified to the characteristic
cobweb o f a business cycle.

The usual notes o f caution are i n order. We have at most two hundred
data points: It is not possible to prove statistically "from a data set o f the
lengths available in economics . . . whether the data are generated by high-
dimensional chaos" (Brock, Hsieh, and LeBaron 1991, 3). " I f long cycles
. . . exist, we w i l l need another 200 years o f data to determine whether they
do exist or are just a statistical figment o f an overactive imagination" (Becker
1988, 7). A t best, therefore, what we have to offer are suggestive graphics.

Long W a v e s o fPrices

The idea that there have been half-century-long waves o f prices in the past two
centuries begins w i t h the observation o f rhythms in the moving averages o f
the annual growth rates o f prices upward and downward on the sides o f
inflationary spirals (fig. 10.1). These movements have occurred with a broad
cross-national consistency (fig. 10.2). Russian economist N i k o l a i Kondratiev
argued that such roughly 55-year trough-to-trough waves (hence, Kondratiev
waves) were a fundamental characteristic o f capitalist economies, each
wave representing a radical regrouping of, and change i n , society's produc­
tive forces (Berry 1991; also Rosser 1991, 138 f f . , for a discussion o f
Schumpeterian discontinuities and historical reswitching). Later investiga­
tions confirm that these regroupings represent the crystallization o f new
3 1810 1830 1850 1870 1890 1910 1930 1950 1970 1990

Ten-year Moving Averages Annual Rate of Change

Fig. 10.1. A n n u a l a n d t e n - y e a r m o v i n g averages of t h e g r o w t h rates o f


U.S. w h o l e s a l e prices 1790-1990. T h e m o v i n g averages have been used
as evidence of h a l f - c e n t u r y - l o n g w a v e r h y t h m s , a r o u n d w h i c h t h e
annual oscillations fluctuate.

15

*
CO
o
Pri

10

a
o>
c
<o
-C
5
O
**-
o
o

«
"5 0
3C
C
<
' ' '
1780 1930 1960

•UK -US

Fig. 10.2. Ten-year m o v i n g averages o f U.S. a n d B r i t i s h w h o l e s a l e


prices, 1 7 9 0 - 1 9 9 0 . N o t e t h e b r o a d c o n s i s t e n c y of t h e l o n g - w a v e
r h y t h m s , as w e l l as t h e u p w a r d t r a n s l a t i o n o f t h e level a r o u n d w h i c h
prices o s c i l l a t e after W o r l d W a r II. A c o m b i n a t i o n of d e p a r t u r e f r o m t h e
d i s c i p l i n e o f g o l d a n d t h e a d o p t i o n of Keynesian m a c r o e c o n o m i c p h i l o s -
o p h i e s p r o d u c e d a rise of 5 p e rc e n ta g e p o i n t s i n t h e annual i n f l a t i o n
rates i n b o t h c o u n t r i e s .
218 Chaos Theory in the Social Sciences

techno-economic paradigms i n the aftermath o f inflationary spirals/stagflation


crises that define the Kondratiev peaks. Such spirals/crises come in the decade
following market saturation by a previous techno-economic paradigm
(" . . . the inflationary pressure generated during the final stages o f an up­
swing continues for about 10 years . . . , " Rasmussen and Mosekilde 1989,
279). Thus, the surge o f U . S . growth after the C i v i l War centered on the steel
industry as the leading sector and on the expansion and completion o f the
railroads, after the canal-building and wind-powered transportation era had
peaked i n the late 1850s (fig. 10.3, from Berry, K i m , and K i m 1993).
Viewed in this way, inflationary spirals/stagflation crises delineate tran­
sition points beyond which the older techno-economic paradigm declines and
a new paradigm takes off, with growth centering on new leading sectors and
new supportive infrastructure networks. Initially, new product prices are high,
but they decline w i t h achievement o f scale economies and w i t h routinization.
Simultaneously, as consumers switch from older to newer products, product
prices associated w i t h the now-declining paradigm also collapse, and along
w i t h them the values o f their raw materials and fixed assets. O n l y after the
new paradigm achieves market dominance in the next Kondratiev trough and
then diffuses toward market saturation do demand pressures once again force
prices upward, ultimately to be driven by speculation into the next spiraling
Kondratiev peak.

Fig. 10.3. R e l a t i o n s h i p b e t w e e n e x p a n s i o n o f a t e c h n o - e c o n o m i c para-


d i g m t o m a r k e t s a t u r a t i o n a n d t h e l o n g w a v e o f prices. ( F r o m Berry,
K i m , a n d K i m 1993.)
Long Waves 1790-1990 219

The regular succession o f inflationary spirals/Kondratiev peaks at 5 0 - 6 0


year intervals (fig. 10.2) suggests a pattern o f transition via intermittency that
may generally be confirmed by the existence o f a channel i n the first return
map, P t+i = f(P,), as discussed in Berge, Pomeau, and Vidal (1984, 245).
Evidence for such channels has been presented elsewhere (Berry 1991, 22ff.);
suffice it to say that Kondratiev peaks represent critical bifurcation points
beyond which clusters o f innovations can move the economy on to alternative
evolutionary logistics (De Greene 1988, 2 9 1 - 9 2 ) . The alternatives are not
ordained: there are usually several competitors, o f which only one becomes
paradigmatic.
The new growth logistic produced by the successful paradigm may be
characterized by the same kind o f equilibrium-seeking behavior as the pre­
vious techno-economic paradigm, or it may result in a different pattern. In
either case, just as the success o f a new paradigm may be taken to be stochas­
tic, the equilibria are chaotic, i n that the dynamic processes that are unleashed
are captured by attractors o f one kind or another.
We seek to document the equilibrium-seeking/shifting behavior o f prices
for the successive Kondratiev waves, 1790-1845, 1845-95, 1895-1930, and
1955-present, plus the transition period 1930-55, when, in the aftermath o f
the Great Depression and the New D e a l / W W I I / B r e t t o n Woods/Korean War
period, the adoption o f macroeconomic policies moved the U . S . economy
into an environment o f permanent inflation (Berry 1991, 13, 3 1 - 3 4 ) . Prior to
the transition, long waves moved with inflationary/deflationary rhythms;
afterward, the successions became inflationary/disinflationary.
Shifts on either side o f Kondratiev peaks are o f interest, but o f course so
are the peaks themselves. A s Friedman and Schwartz note (1963b, 6 2 - 6 4 ) :

The process we have described w i l l tend to overshoot the mark; it w i l l


not simply produce a smooth movement to the new path consistent w i t h
the new rate o f growth o f the money stock assumed to prevail . . . . (I)n
the process o f going from the initial to the new equilibrium, prices must
rise at a faster rate than their ultimate rate. Hence the rate o f price rise
must overshoot . . . . The tendency to overshoot means that the dynamic
process o f transition from one equilibrium path to another involves a
cyclical adjustment process. The final picture that might ultimately de­
velop could be o f a partly self-generating cyclical mechanism. Distur­
bances i n the rate o f change in the money stock set in train a cyclical
adjustment mechanism including a feedback in the rate o f change in
money itself. Additional disturbances from time to time would prevent
the fluctuations from dying out.

For each o f the time periods we plot the fluctuations in the annual growth
rates o f prices (P) in three spaces:
220 Chaos Theory in the Social Sciences

P, = f{t), the conventional time series graph;


P, = fiP,), the growth rates o f prices w i t h respect to price levels; and
Pt+l = fiP,), the Poincare first-return map, o f which t w o versions are
given, the scatter diagram and the temporally connected sequence.

We ask how the price rhythms have been reset after each inflationary
surge, whether there is evidence o f a catastrophic switch in behavior in the
catastrophe theory sense o f the term, and what the underlying conditions
producing such a shift might have been.
We begin w i t h figure 10.4, covering the period 1790-1845. I n the PIP
space (top right-hand graph), oscillations are evident around two price levels.
The first and higher level is that o f the upwave leading toward the 1815
Kondratiev peak. After this stagflation crisis, the short-run fluctuations switch
to oscillate around a lower downwave level. There is thus evidence o f a
bifurcation astride the Kondratiev peak, with oscillation around a higher price
level before the main inflation surge, and around a lower level after the surge.
The Poincare first-return maps ( t w o bottom graphs in fig. 10.4) reveal
something else, however: despite the switching o f annual oscillations after the
Kondratiev peak, there is a single underlying chaotic regime in which the
oscillations are contained by well-defined limits (see Berge, Pomeau, and
V i d a l 1984, 6 6 - 7 1 , 1 0 3 - 4 , 181). The motions are those o f an endogenous
chaotic cycle, as evidenced by successive returns to the outer limits in the
graph (Rosser 1991, 114), w i t h moderate boom and bustiness revealed by the
elongation o f the ellipse o f points along the 45° diagonal (Gleick 1987, 176—
77; Glass and Mackey 1988, 2 6 - 3 4 ; Casti 1989, 2 4 9 - 5 3 ) . Even without the
filtering presented by Berry (1991) in the attempt to isolate the attractor
describing the longer-run pattern o f boom and bust, the movements point to a
strange-like attractor structure that is consistent both w i t h the notion o f a l i m i t
cycle (see also Goldberger and Rigney 1990, 3 2 - 3 3 ) and w i t h that o f an
attractor basin. Normally, in a limit cycle, the expectation is that sequential
movements w i l l be around the edge, w i t h a hollow interior. In the case o f an
attractor basin, the interior o f the ellipse is noisy and there is skateboard
movement across the short axis that gradually climbs the long axis (the princi­
pal attractor) before descending rapidly from inflationary peak to deflationary
depression.
The long wave is better defined in the period 1845-95, w i t h a very sharp
inflationary surge that peaked in 1865 (top left-hand graph in fig. 10.5). I n the
PIP space (top right-hand graph), upwave/downwave levels are discernible,
but they are not as sharply differentiated as in the prior long wave. Likewise,
while there is a set o f limits in the Poincare first-return maps, they too are less
clearly defined, although the oscillations do occur w i t h i n the same limits as in
figure 10.4, once the 1865 boom and bust are accounted for. Certainly, the
Long Waves 1790-1990 221

1790 1793 1SO0 1105 1110 ISIS 1S20 1825 1830 1835 1840 1845

- Ten-year Moving Averages Price (/>,)


- Annual Rate of Change

•20 -10 10 20 30 40

Annual Rate of Change in Price » 1


Annual Rate of Change in Price [t)

Fig. 10.4. P o r t r a i t s o f price m o v e m e n t s 1 7 9 0 - 1 8 4 5 . T h e t o p l e f t - h a n d


g r a p h is t h e c o n v e n t i o n a l t i m e - s e r i e s p l o t a n d s h o w s b o t h a n n u a l rates
of c h a n g e a n d t h e t e n - y e a r m o v i n g averages. The t o p r i g h t - h a n d g r a p h
p l o t s a n n u a l changes against price levels. The t w o b o t t o m phase por-
t r a i t s p l o t t h e a n n u a l price changes, against t h e i r p r e v i o u s year
value, P . T h e l e f t - h a n d g r a p h s i m p l y s h o w s t h e d i s t r i b u t i o n o f p o i n t s ,
t

b u t t h e r i g h t - h a n d g r a p h links successive p o i n t s i n t i m e t o reveal t h e


year-to-year oscillations.

sharpness o f the 1865 peak emphasizes the magnitude o f the b o o m and bust,
whereas the first-return maps point to endogenous chaos.
The 1867-96 downwave phase o f this long wave is characterized thusly
in Friedman and Schwartz's Monetary History of the United States, 1867¬
1960 (1963a):
222 Chaos Theory in the Social Sciences

2o ' ' I
3 0 1 1
1M5 1850 !I55 I860 1865 1870 1875 1880 1885 1890 1895 " ' '
20 30 AO 50 60

—•—Ten-year Moving Averages Price!/')


—O— Annual Rate of Change

Annual Rate of Change in Price (r) Annual Rate of Change in Price (f)

Fig. 10.5. P o r t r a i t s o f price m o v e m e n t s , 1 8 4 5 - 9 5 . T h e g r a p h s appear i n


t h e s a m e o r d e r as i n f i g u r e 10.4.

The years from 1867 to 1879 were dominated by the financial aftermath
o f the C i v i l War. I n 1862, convertibility o f Union currency into specie
was suspended as a result o f money creation i n the North to help
finance the C i v i l War. From then until the resumption o f specie pay­
ments i n 1879, the U . S . was on a fiduciary (the "greenback") stan­
dard, and the dollar was linked to other currencies via exchange rates
that fluctuated freely. Throughout this period, Great Britain main­
tained a gold standard.
Long Waves 1790-1990 223

In 1879 specie payments were resumed at prewar parity, preceded by a


radical decline in the stock o f money 1875-78. After 1879, wide­
spread adoption o f a gold standard, despite new discoveries o f gold
and new financial techniques for building a superstructure o f money on
it, contributed to a steady downward trend in product prices and
mounting political discontent, expressed as "Greenbackism" and "free
silver," sharpened by recurrent banking crises (1873, 1884, 1890,
1893). The period ended with the defeat o f W i l l i a m Jennings Bryan in
1896.

The resumption o f specie payments in 1879, plus the banking crises o f 1873,
1884, 1890, and 1893 are readily discernible in the greater and lesser inflation
spikes in figure 10.5. Yet equally apparent is that each o f these politically
significant events falls w i t h i n the limits o f the chaotic regime. N o clear
difference is discernible between price behavior in the greenback and gold
standard periods, or on the two sides o f the 1865 spiral.
Between 1895 and 1930, there is, in contrast, an exceedingly well de­
fined two-phase relationship between price movements and levels, as behav­
ioral switching occurs after the 1920 spiral (fig. 10.6, top right-hand dia­
gram). Preceding the surge to the 1920 peak there is a progressive upwave
acceleration o f the rate o f price increase w i t h i n a narrowly oscillating range, a
classic example o f intermittency in action (Berge, Pomeau, and V i d a l , 1984,
245). After the Kondratiev peak there is retreat to oscillation around a stable
attractor. Yet these are two states o f the same narrowed set o f chaotic limits.
As is seen i n the P IP,
t+l space, both prior to and after the peak, the oscilla­
tions occur w i t h i n the same restricted range.
What could have caused these switches? The Friedman-Schwartz
(1963a) characterization o f this period is that

1897 to 1914 combined a long peacetime rise in prices with a relatively


high degree o f economic stability.
But discontent w i t h the banking structure came to a head in 1907, fol­
lowed by the temporary Aldrich-Vreeland A c t o f 1908 and the perma­
nent step o f the Federal Reserve A c t o f 1913. During and after World
War I , weakening o f the international gold standard gave the Federal
Reserve System greater freedom o f action than might otherwise have
occurred. During W W I , the system served as the engine for inflation­
ary financing o f government expenditures, which continued for eigh­
teen months after the war and then was reversed by abrupt Federal
Reserve action. After this reversal, the Federal Reserve assumed a
strong directive role, expanding the stock o f money at a steady rate
until the Great Depression hit.
224 Chaos Theory in the Social Sciences

40
30 . _

1895 1900 1905 1910 1915 1920 1925 1930 , n ~


2 0
30 40 SO 60 70

—•—Ten-year Moving Averages Price [P )


—o— Annual Rate of Change

2 0 1 0 0 !0
20 10 0 10 20 30 40 20 30 «0

Annual Rate of Change in Price (f) Annual Rate of Change in Price it)

Fig. 10.6. P o r t r a i t s of price m o v e m e n t s , 1 8 9 5 - 1 9 3 0 . N o t e t h e t w o - p h a s e


s t r u c t u r e a t h w a r t t h e 1920 s t a g f l a t i o n crisis ( t o p r i g h t - h a n d graph) b u t
t h e c o m m o n phase p o r t r a i t s ( l o w e r graphs).

What was at work were institutional changes resulting from dissatisfaction


w i t h the boom and bustiness o f the banking system. The Fed's actions with
respect to the stock o f money had the effect o f changing inflationary creep into
a decade o f relative price stability (although Weiher (1992) says that this was
the inadvertent consequence o f policy designed to help Britain return to the
gold standard, rather than purposeful countercyclical policy). But Friedman
and Schwartz (1963a) continue:
Long Waves 1790-1990 225

The downturn o f 1929 marked a major transition. There were major bank
failures, whose effects were multiplied by tight money policies and by
the Fed's reactions to Britain's departure from the gold standard.
There began the Great Contraction, during which one-third o f the
nation's banks went out o f existence, and there was a catastrophic drop
in aggregate demand, a result o f the Fed's failure to stop an enormous
decline in the money supply.

O n the heels o f the Great Depression, the relative price stability o f 1921-29
vanishes. Rather, in several surges defined by a l i m i t cycle, prices were
ratcheted upward (top right-hand graph i n fig. 10.7) as the Fed followed a
reflation strategy. Note that the Poincare phase portraits again assume the
form o f chaotic l i m i t cycles. The first points are i n the deflationary lower left
quadrant during the depression, but the l i m i t cycling is confined thereafter to
the upper right quadrant as a permanently inflationary environment is created
(bottom graphs in fig. 10.7).
Friedman and Schwartz (1963a) say that with the coming o f the New
Deal:

Like many countries abroad, the United States entered a phase o f cheap
money policy, w i t h a network o f controls that only partially and inef­
fectively combatted inflationary pressures. The proponents o f the New
Deal were strongly in favor o f easy money: "Easy money was the near-
uniform prescription. Inflation was the near-uniform result" (Friedman
and Schwarz 1963a, 700).
In 1934 came federal deposit insurance as one o f the reactions to the
contraction, the dollar was devalued, and subsequently the United
States departed from gold. Most importantly, faith in the potency o f
the Federal Reserve System collapsed as the United States adopted a
fiduciary standard, and the system assumed a passive and reactive
role. Federal Reserve policy was subordinated to Treasury policy dur­
ing W W I I . The stock o f money became a function o f governmental
taxing and spending and rose 2.5 times between 1939 and 1945, and
then once again during the Korean conflict.

This inflationary environment, combined w i t h a reactive Federal Re­


serve, provides the context for the final set o f graphs (fig. 10.8). Prices march
steadily upward, but annual oscillations are restricted to a limited range along
the diagonal in the P /P
t+i tspace. W i t h the compression o f the chaotic l i m i t
cycle, what is left are the longer-term booms and the busts, displaced upward
to prevent deflation. Only a little more than one-half o f a long wave is
226 Chaos Theory in the Social Sciences

40

30 i

A n n u a l R a t e o f C h a n g e i n P r i c e <r> A n n u a l R a t e o f C h a n
9 e i n P r i c e
" !

Fig. 10.7. P o r t r a i t s o f price m o v e m e n t s , 1 9 3 0 - 5 5

depicted; there is thus no real opportunity to observe whether any up-


wave/downwave switching has followed the 1980-81 crisis.
There has been much debate on whether fiscal policy has been a potent
stabilization tool since the 1950s, given the Fed's inconsistent policies o f
money supply growth, slow in the 1950s, higher in the 1960s to keep down
interest rates and fuel g r o w t h , a captive o f political forces in the later 1960s
and the 1970s (giving rise to runaway inflation), w i t h a sharp reduction in M 2
introduced by Paul Volcker that took hold in 1981 (Weiher 1992). I f the lower
panels o f figure 10.8 are compared w i t h the top four graphs in figure 10.9,
Long Waves 1790-1990 227

Fig. 10.8. P o r t r a i t s o f price m o v e m e n t s , 1955-90

some additional insights into the consequences o f stabilization emerge. Figure


10.9 plots the first-return maps for the long-wave movements o f prices in each
of the five time periods. Each traces out a longer-run boom-bust channel along
the 45° principal diagonal o f the graphs—a strange attractor that defines the
underlying long-wave rhythms. The first-return map i n figure 10.8 (bottom
right-hand graph) looks very much like the earlier long-wave movements,
showing clearly enough that since World War I I , it is the annual oscillations o f
prices that have been suppressed: since 1955 price movements have been
constrained to move along the long-run attractor. Thus, whereas short-run
*
a.
j

A n n u a l Rate of Price C h a n g e (f)


A n n u a l Rate o f Price C h a n g e {t)

£ 20
(D 20 TO


20 -10 0 10 20 30 40

A n n u a l Rate o f Price Change (t) A n n u a l Rate o f Price Change (f)

a.

A n n u a l Rate o f Price C h a n g e (f)

Fig. 10.9. F i r s t - r e t u r n m a p s o f prices (/*,+,//>,) p l o t t e d f o r t h e t e n - y e a r


m o v i n g averages, 1 7 9 0 - 1 8 4 5 ( t o p left), 1 8 4 5 - 9 5 ( t o p r i g h t ) , 1 8 9 5 - 1 9 3 0
( m i d d l e left), 1 9 3 0 - 5 5 ( m i d d l e r i g h t ) , a n d 1 9 5 5 - 9 0 ( b o t t o m ) .
Long Waves 1790-1990 229

stabilization has occurred, the long-run boom-bust pattern o f the long wave
remains.

L o n g Cycles of Economic G r o w t h

We now repeat the foregoing for variations in growth rates o f real per capita
G N P for the same time periods. Note (fig. 10.10) that the long swings o f
growth appear to diminish after 1950, and that the short-run oscillations also
are dampened between 1950 and 1990.
Growth behavior from 1790-1845 and 1845-95 is essentially similar
(figs. 10.11 and 10.12). Two long cycles o f growth are arrayed on either side
o f the 1815 and 1865 inflation peaks/stagflation crises (Berry 1991). These
quarter-century-long cycles are named for Simon Kuznets, who first identified
them. I n both PIP spaces the first Kuznets cycle involves a surge o f real
growth and a cyclical drawback in the stagflation crisis. Likewise, the second
Kuznets cycle also reveals growth surges terminated by drawbacks in the
deflationary depressions that mark the Kondratiev troughs. The P /P t+l t

Poincare first-return maps display chaotic limit cycles encircling attractor


basins.
From 1895 to 1930 the two-cycle Kuznets pattern is not as sharply

Fig. 10.10. A n n u a l a n d e i g h t - y e a r m o v i n g averages o f t h e g r o w t h rates


of real a n d p e r c a p i t a GNP i n t h e U n i t e d S t a t e s , 1 7 9 0 - 1 9 9 0 . N o t e t h e
c o m p r e s s i o n of t h e a n n u a l f l u c t u a t i o n s after W o r l d W a r II.
230 Chaos Theory in the Social Sciences

Fig. 10.11. Phase p o r t r a i t s o f g r o w t h , 1 7 9 0 - 1 8 4 5 . D i a g r a m s o r g a n i z e d


in s a m e m a n n e r as f i g s . 1 0 . 4 - 1 0 . 8 .

defined (fig. 10.13), but nonetheless the PIP space reveals that growth pro­
ceeds until the cyclical setback i n W W I and resumes again until the pullback
in 1929-30. A g a i n , the P lP, maps show a chaotic l i m i t cycle.
t+l

In contrast, the 1930-55 transition involves strongly cyclical behavior


(fig. 10.14), but from 1955-90 what is most apparent is the steadiness o f the
growth path, marred by occasional business cycle downturns and by the
1980-81 stagflation setback (fig. 10.15). What emerges is an increasingly
well defined l i m i t cycle (Berge, Pomeau, and Vidal 1984, 27, 67, 178, 181),
Long Waves J 790-J990 231

an absence o f attractor basin skateboarding, and only limited boom and b u s i ­


ness, surely another outcome o f the reactive federal stabilization policy that
has served to suppress sharper swings. Each o f the circuits around the limit
cycle is one business cycle in length.
I f the lower panels o f figure 10.15 are compared with those o f figure
10.16, the important contrast that has resulted between the stabilization o f
growth and prices is brought into focus. Unlike the restriction o f the annual
oscillations o f prices after 1950 (producing an annual first-return map for
1950-90 that looks like the long-term first-return pattern before World War
232 Chaos Theory in the Social Sciences

IS 10 -5 0 S 10 IS " -'"
A n n u a l Rate o f C h a n g e in G r o w t h ( f ) A n n u a l R a t e o f C h a n g e i n G r o w t h (*)

Fig. 10.13. Phase p o r t r a i t s o f g r o w t h , 1 8 9 5 - 1 9 3 0

I I ) , the long-term first-return maps o f growth (fig. 10.16, top panels) are quite
unlike the annual first-return maps after 1950 (bottom panels o f fig. 10.15).
The latter display a perfect l i m i t cycle, each 360° circuit o f w h i c h is a business
cycle; the former shows the longer-term boom and bustiness o f growth. What
thus has been eliminated since 1950 is the longer-term Kuznets cyclicality: the
first-return map o f the moving average o f growth since 1950 appears to be a
stable attractor (bottom panel, fig. 10.16). What has been produced is a
pattern o f business cycles around a steady-state path o f economic growth.
Long Waves 1790-1990 233

- O — A n n u a l Rate of Change S Real Per C a p i t a G N P

.15 -10 5 0 5 10 15 15 10 -5 0 5 10

Annual Rate of Change in G r o w t h ( f ) A n n u a l Rate o f Change in G r o w t h ( f )

Fig. 10.14. Phase p o r t r a i t s o f g r o w t h , 1 9 3 0 - 5 5

Discussion

We see in the foregoing the advantages o f using multiple phase portraits to


analyze dynamic processes. Both prices and growth reveal a tension between
longer-run rhythms and short-term variability. The long-run rhythms are un­
stable near their peaks, leading to intermittent shifts in the locus o f shorter-run
fluctuations. I n the long run the pattern is one o f Kondratiev waves o f prices
and Kuznets cycles o f growth, and with these are associated intermittent
234 Chaos Theory in the Social Sciences

I960 1965 1970 1975 1980 1985

· - Eight-year Moving Averages 3000 3500 4000

a— Annual Rate of Change Real Per Capita GNP

2 5 » *
••t
% ffc * at

1 1

Annual Rate of Change in Growth (f)


Annual Rate of Change in Growth (f)

Fig. 10.15. Phase p o r t r a i t s o f g r o w t h , 1 9 5 5 - 9 0

booms and busts. I n the shorter run, fluctuations were chaotic within limits
until the Great Depression.
The rhythms have not been immutable, however. Institutional changes
and policy interventions since the Great Depression have raised the level
around which prices fluctuate by five percentage points and, simultaneously,
have suppressed short-term oscillations, leaving behind more gradual upward
and sharper downward movements along the long-wave attractor. But simul­
taneously, economic growth has been constrained to fluctuate around a long-
term steady state o f roughly 2 percent per annum while preserving the shorter-
term rhythms o f the seven- to eleven-year business cycle. Thus, since the
ï
î o
o
o
«s

< IS

Annual Rate of Growth Change (f)


Annual Rate of Growth Change it)

o
S
n
o T 5
<5 o

< 15

Annual Rate of Growth Change (f) Annual Rate of Growth Change (fî

O
si
3
o
<5

Annual Rate of Growth Change (f)

Fig. 10.16. F i r s t - r e t u r n m a p s o f l o n g - t e r m g r o w t h , 1 7 9 0 - 1 8 4 5 , 1 8 4 5 - 9 0 ,
1890-1930, 1930-55, and 1955-90
236 Chaos Theory in the Social Sciences

Great Depression, a combination o f institutional changes and public poli­


cies has resulted in more optimizing (Slutsky-Frisch-Tinbergen) economic
growth behavior, but the interventions have had a different consequence for
prices, constraining them to move within a narrower (but still boom-and-
busty) l i m i t cycle.
CHAPTER 1 1

C i t i e s as S p a t i a l C h a o t i c A t t r a c t o r s

Dimitrios S. Dendrinos

Developments in the mathematical theory o f dynamical stability and bifurca­


tion theory have considerably influenced developments throughout the social
sciences. Sources have been either the field o f structural stability emanating
from catastrophe theory (see Thorn 1975; Zeeman 1977), mathematical chaos
(see Lorenz 1963; M a y 1976; Guckenheimer and Holmes 1983; Thompson
and Stewart 1986; Schuster 1988; Devaney 1989), or the theory o f fractals
(see Mandelbrot 1977; Peitgen and Richter 1986; Barnsley 1988). The above
are only a short list o f basic references among many that have appeared over
the past 30 years i n these fields.
Almost without exception, social science fields have been influenced by
these developments, including areas o f study in archaeology, history, soci­
ology, demography, political science, geography, psychology, and economics.
In particular, the fields o f economics (micro- and macroeconomics, including
economic growth theory, the theory o f finance, business cycles theory, tech­
nological progress, international trade, and so on; see Anderson, Arrow, and
Pines 1988; Barnett, Geweke, and Shall 1989; Barnett and Chen 1990; and
Rosser 1991), and spatial analysis (geography, urban and regional analysis,
transportation dynamics and traffic flow analysis, and so on), have experi­
enced considerable impact from these mathematical advances.
A l o n g w i t h the considerable new (basic as well as marginal) insight
gained by such applications, there has also been a revision o f past claims in
social theory. To some, a call for reconfiguring the social sciences has
emerged from insights gained through these newly developed fields. But,
together w i t h this call, the realization has settled in that there are, at least at
present, certain insurmountable practical as well as theoretical obstacles to
such reconfiguration.

This paper was written during spring, 1992.

237
238 Chaos Theory in the Social Sciences

Contributions

A t the outset, a few basic points w i l l be made as to the magnitude o f contribu­


tions to and extent o f impact from applications emanating from nonlinear
dynamics into social sciences. Then, some remarks w i l l follow, addressing
the question o f the considerable drawbacks o f developing a truly dynamic
social science at present.

Basic and Marginal Inroads

This is not the appropriate forum to survey all the areas where chaos theory
has made inroads or to fully assess the impact from or the value o f these
contributions. Some o f the papers in this volume do so in their o w n special
fields. Whereas in some areas, like those in spatial analysis and geography,
the contributions might be basic, in other areas, particularly i n economics,
one may find the contributions made to date to be rather marginal.
In economics, the most credible o f the attempts to exploit chaos (and
catastrophe) theory occurred when existing classical and neoclassical deduc-
tionist (causality-based) theories were modified to marginally change preexist­
ing (and widely accepted among disciples o f current orthodoxies) assump­
tions. Such changes were offered as means o f showing how chaotic dynamics
can arise from currently prevailing theoretical models, and thus enhance these
models' utility.
There is another side to this coin, however. Marginal extensions o f
currently accepted paradigms in economic analysis point directly to one o f
their weaknesses, namely, that under slight but proper modifications in speci­
fication, these models can reproduce almost anything that the analyst wishes
to produce through theoretical deduction. A scientist might find such accom­
modations somewhat troublesome.

New Dynamical Features

The single most important contribution mathematical chaos has made is to


demonstrate the possible presence o f new dynamical features in social sys­
tems that theoreticians had never addressed before. Until very recently,
aperiodic movement was an attribute o f social science models to avoid and
quasi-periodic motion was never even conceived of.
Chaos theory showed that through such dynamics one can detect interest­
ing and novel insights i n the behavior o f social systems. It also supplied a
method to classify the various dynamical trajectories obtainable in phase
space and to catalog a large number o f possible bifurcations.
Cities as Spatial Chaotic Attractors 239

A New Type of Explanation

Mathematical chaos further suggests that periodic, quasi-periodic, and non-


periodic trajectories in the phase portrait o f interdependent socioeconomic
variables can be generated by very simple dynamical equations. Chaos occurs
in dimensionally very low systems, where it most often has been studied.
High-dimensional chaos is still an elusive subject. Efficiently capturing
aperiodic motions can be accomplished w i t h simple dynamical specifications.
Efficient dynamic specifications can be viewed as a new method o f
explanation in the social sciences. Mathematical chaos highlights a topic first
addressed about twenty years ago with the advent o f catastrophe theory and
the theory o f structural stability. One can label this new type o f explanation
reductionist. The example o f this chapter is an exercise in reductionist expla­
nation.
It bears little correspondence to the prevailing causal or deductionist
explanation, or to the inner structure o f the underlying socio-spatial complex­
ity. A social scientist may find reductionist specifications and inferential ex­
planation to have minimal economic interpretation and to lack any strictly and
currently prevailing disciplinarian meaning, for that matter.
For a good, but extreme, example o f the issues involved in this argument
the reader is referred to the work by Mandelbrot (1977) and Barnsley (1988)
on fractals. Simple iterative processes depict in their phase portrait elements
that resemble clouds, plants and animals, shorelines, or exotic landscapes.
However, the iterative maps responsible for generating such images do not
hold explicitly in their formulation either the physics and chemistry o f clouds,
the genetics and biology o f plant and animal species, or the geology o f coastal
lines and terrestrial (or extraterrestrial) landscapes.
These examples are extreme in the sense that the iterative maps respon­
sible for their images are indeed extremely efficient (in terms o f state variables
and parameters they contain), but also totally devoid o f any substantive mean­
ing or (deductionist or reductionist) explanation. They can accommodate,
however, a wealth o f dynamical features. I n the example to be presented i n
this paper, it is demonstrated that an intermediate case exists involving reduc­
tionist explanation coexisting w i t h model efficiency. Some parsimony is
traded off for explanation, while the variety o f dynamical behavior is largely
preserved.
The requirement for efficiency and parsimony, and the need for obtaining
a comprehensive list o f dynamical behaviors may pose some questions to
currently dominant social science theories. Both criteria were brought to the
forefront o f social science discourse with dynamical modeling; the issue is
further enhanced w i t h the advent o f mathematical chaos.
240 Chaos Theory in the Social Sciences

A need arises for a basic and not simply marginal modification to defin­
ing social systems and for establishing a single, all-encompassing, non-
disciplinarian dynamical social science frame o f reference. This is a call for a
new and ambitious research agenda w i t h potentially high rewards in descrip­
tive power, but at the cost o f losing explanatory value o f the deductionist type.
A theory that offers reductionist explanation, being a different type o f
explanation than social scientists have been used to, may be considered a
severe shortcoming and may conjure criticisms by those w i t h an incantation
for policy analysis. Widespread current dissatisfaction w i t h the multiplicity o f
and tenuous base for most policy-sensitive, behaviorally motivated social
science fields, where multiple deductive explanations vie for few followers,
tempers such criticism. Discounting o f such fears may also be enhanced by
the fact that social science orthodoxies are ephemeral and do not commend the
recognition o f relatively durable natural science theories. But this, being a
very involved and demanding subject, w i l l not be elaborated significantly
here, although a bit more w i l l be added later.

Difficulties

Next, a few comments w i l l be made touching on the difficulties in setting up


and testing for a truly dynamic social science theory, currently perceived as
insurmountable.

Lack of Data

Chaotic paths i n phase portraits o f even the lowest dimensionality, such as


those found in the one-dimensional logistical prototype, require an enormous
amount o f data to trace, calibrate, and test. Extremely few variables in the
social sciences are available at frequent enough time intervals to permit one to
do so at any degree beyond a rather rough level o f accuracy. Most o f the
variables w i t h extensive time-series data are found i n stock market or mone­
tary tables, in the field o f finance. It is not surprising that the initial attempts to
find chaos and statistically test for i t (through Liapunov exponent-based tech­
niques) i n the social sciences occurred in that field.

Frequent Perturbations

In addition to the difficulties surrounding lack o f data, social science variables


also suffer from many other ills. One o f these ills has its source in the
restriction that social systems do not produce repetitive measurements under
experimentally controlled conditions. Perturbations o f social science variables
are very frequent, irregular, and unexpected. Further, very rarely, i f ever, are
Cities as Spatial Chaotic Attractors 241

social systems comprised o f identical components with highly homogeneous


behavior. These concerns raise issues o f validity i n statistical analyses in
social studies. Statistics are based on fundamental assumptions in probability
theory that may not be applicable to social science fields. The difficulty is
aggravated when dynamics are involved and temporal heterogeneity is added.
Specifically, the definition o f and criteria for detecting what is referred to
as white noise present an added difficulty to the issue, since to separate white
noise from internal dynamics in socio-spatial variables is at present almost
impossible, even for dimensionally low-level social systems. O n the side o f
chaos, too, there are difficulties at present related to deriving a conclusive test
for detecting i t . These difficulties create considerable impediments to deriving
a statistical theory o f social systems dynamics, and they render testing almost
prohibitive.

Inaccuracy

When the dynamics o f state variables are extremely sensitive to either initial
conditions or the magnitude o f the model parameters, then serious questions
due to inaccuracy in measurements o f state variables and in estimating param­
eter values arise in the social sciences. H o w inaccuracy in parameter estima­
tion techniques might, in the presence o f quasi-periodic or chaotic dynamics,
hinder statistical methods is a topic still to be fully addressed.
Instruments, techniques, and methods o f measuring social science data
(state variables) are highly imperfect. Fuzziness in measurement and m u l t i ­
plicity in perception o f social data sets are often necessary conditions for
social planning, policy, and action. H i g h resolution or sharpness in data sets is
not always feasible or efficient to obtain. It may even be undesirable.

Unresolved Issues

In the social sciences, theoretical advances in nonlinear dynamics are also


severely hampered by binding practical constraints. One o f these is associated
with the level o f disaggregation one can employ to analyze socio-spatial
systems. Social systems do not lend themselves, for numerous reasons, to
very fine levels o f disaggregate analysis. Consequently, one may be forced to
deal with rather aggregate data. Coarseness in social data sets might wash
away interesting dynamics found at a finer scale, further inhibiting any
chances for advancement in empirical, as well as theoretical, work.
Finally, questions can be raised regarding the issue o f fractal properties
in chaotic attractors associated with socio-spatial forms. Social systems take
up physical space and cannot be broken down to an infinitely small self-
replicating scale. Lower and upper bounds in the scale o f human settle-
242 Chaos Theory in the Social Sciences

ment activity, for instance, are due to the size o f buildings and cities. The
presence o f these bounds poses certain theoretical questions (unanswered until
now) to socio-spatial models producing chaotic outcomes with fractal dimen­
sions.
To conclude however, that social science ought to downplay mathemati­
cal chaos theory and its insights, and instead persist in learning from only
static, sharp, or stable dynamical models would be erroneous, and even per­
nicious. I t w o u l d deprive the analyst o f an ability to obtain intuitively appeal­
ing qualitative insights, simply because one lacks the comfort o f very many
and accurate measurements. Such luxury may never be afforded the social
scientist, while the social agent is always confronted with a need to act. The
agent w i l l always be either right or wrong in its speculative action; conse­
quently, the agent w i l l be either rewarded or penalized. A mathematical
chaos-ignorant social scientist would not be an effective or desirable advisor
to the social agent.

The Universal M a p o f Discrete Relative Dynamics

Nonlinear interdependent dynamics capable o f generating chaotic paths have


made salient contributions to the study o f socio-spatial evolution over the past
fifteen years or so. They have supplied methods to comprehensively and
abstractly model the dynamics o f socio-spatial systems. Even more impor­
tantly, they have provided a platform for some new and at times fundamental
ideas and novel insights to penetrate and subsequently permeate socio-spatial
analysis.
A m o n g the various contributions made during the 1980s a universal map
of discrete relative socio-spatial dynamics was developed (Dendrinos and
Sonis 1990). It set the stage for a broad framework to emerge in describing,
analyzing, and modeling ( i . e . , simulating) abstract spatio-temporal processes,
and it is an experiment in reductionist explanation. The framework involved
abstract interdependencies among any number o f competing stocks and loca­
tions under a nonlinear interdependent dynamics configuration, based on the
comprehensive notion o f spatial relative comparative advantages.
So far, the various contributions made through the application o f the
universal map o f discrete relative dynamics have been confined to only one
type o f formalism built into this map, namely the temporal evolution in the
relative size o f interdependent but heterogeneous stocks as distributed among
a number o f heterogeneous and interdependent locations. Under this configu­
ration, the phase portrait o f the map was that o f an abstract space depicting
dynamics in the relative size o f stocks at different locations.
Physical as well as abstract stocks can be modeled in physical or abstract
phase space. The universal map can take any number o f configurations, from
the simple one-stock, two-location to the multiple-stock, multiple-location
Cities as Spatial Chaotic At tractors 243

specification, and generate their phase portraits. Qualitatively, the dynamical


features o f the various phase portraits differ, depending on the model configu­
ration.
In what follows, the focus is on topics o f interest to spatial analysts,
addressing interurban or intra-urban spatio-temporal allocation processes.
Formation o f human settlements or communities is addressed, each w i t h a
different temporal vintage. Dynamics o f a variety o f socioeconomic stocks
can be accommodated by the various configurations o f the discrete dynamics
map used here. The abstract and general presentation that follows can be
employed by other social scientists, including economists, sociologists, de­
mographers, political scientists, psychologists, and so on, to simulate the
(spatial or nonspatial) dynamics o f interest to their field o f inquiry.
It may also be o f interest to ecologists and biologists, in analyzing stocks
of animal and plant species moving and spreading in space-time. Further, the
method may also be o f potential use to physical geographers and geologists
studying the anisotropy and heterogeneity (possibly spatially chaotic varia­
tion) o f landscapes and geological formations. The method may also supply
insights into the spatio-temporal distribution o f geological events (for i n ­
stance, the location o f epicenters o f seismic activity occurring o v e r t i m e along
fault lines).
Next, a few remarks w i l l be made on the basic features o f selected
configurations. O n l y a simple statement o f the various configurations and
specifications o f the iterative map w i l l be shown, as the detailed analytical
exposition is found elsewhere (Dendrinos and Sonis 1990).

The Single-Stock Case

The central proposition o f the universal map is as follows: at any time period,
/ (t = 0,1,2, . . . ,T), the relative size o f a stock at some location w i t h i n a
prespecified environment (consisting o f a number o f locations) can be simu­
lated (modeled or described) as being directly proportional to a prior level o f
locational comparative advantages enjoyed by the stock in question at the site
in question. Locational comparative advantage is translated into a likelihood
or probability that a particular event w i l l occur at a particular location.
A replication (or reproduction) force occurring at regular time intervals
is assumed to underlie the iterative procedure. For social stocks (as in the case
at hand), the replication procedure may be attributed to internal ( i . e . , among
the various locations within the prespecified environment) temporal realloca-
tion (or redistribution) o f the stock(s).
In specific, for the case o f a one-stock, multiple-location configuration,
the model is given by the following specifications:

Xi (t + 1) = F^jFjit); i,j = 1,2, . . . , / (1)


244 Chaos Theory in the Social Sciences

F,it) = F^itlxM . . . ,x,(t)} > 0. (2)

The discrete iterative map describes the dynamics o f the relative distribu­
tion o f the (homogeneous) stock at a location i (0 < *,·(/) < 1) within an
environment o f / (heterogeneous) locations, so that at all time periods

2,x,(0 = 1 ; / = 0 , 1 , T (3)

over a time horizon, T, which consists o f time periods (or iterations), t,


identical i n length. Quantity x is the probability that an event w i l l occur at any
location, i, i n / . Quantity F is the comparative advantage that any location, /',
in / has in attracting such an event.
The description o f these dynamics is based on the premise that, under a
one-time period delay, a reproduction process is i n effect; the relative size o f
the stock at location i at the next time period, (t + 1), depends on the
normalized current locational comparative advantages given by the function
F,{t). This is only one possible delayed response, as any time delay, n (or
periodic forcing, t + n) could be incorporated into the above dynamical
configuration.
Functions F depict nonlinear current locational comparative advantages.
A m o n g other factors, these advantages are the result o f interdependencies
among the current sizes o f the homogeneous stock at the various locations;
more about the specification o f these functions w i l l be supplied later. A
distinction is noted i n particular; equations 1-3, through their form, identify
interdependencies among stock sizes at the various locations and not inter­
action, that is, flow o f stock(s) among them.
Without any loss o f generality, equations 1-3 can be restated and scaled
in terms o f a certain numeraire (or reference) location, / , as follows:

xff + 1) = Fft)l{\ + 2,-F/r)}

x,{t + 1) = 1/{1 + (4)

Fj(t) = Fj(t)IF,(t);j = 1,2, . . . , / - 1,

as F,(t) is a strictly positive quantity.

The One-Stock, Three-Location Configuration


O f special interest is the case o f the iterative map depicting the discrete
dynamics o f a single stock as distributed among three locations. It turns out
that this is a generic case involving all possible qualitatively different dy­
namics encountered i n the one-stock, multiple-location configuration. De­
tailed specifications o f this model are as follows (see fig. 11.1):
Cities as Spatial Chaotic Attractors 245

x {t + 1) = Ftf/ZjFft);
t i,j = 1,2,3; 2 , · * , « = 1, (5)

In F (t) = In A, + X^a,-, In A:,-,


t (6)

where the vector o f the state variables, *,(**), defines in the three-dimensional
space an equilateral triangle w i t h sizes equal to the square root o f 2.
The vector o f (strictly positive) parameters, A , depicts in absolute terms
and in a descriptive and bundlelike manner the following prevailing condi­
tions at each o f the three locations (among numerous other factors characteriz­
ing each location). First, it identifies the way i n which the environment favors
or handicaps the location. Second, it picks up the presence and level (or
absence) o f natural resources at the location in question. T h i r d , it records a
pattern o f accessibilities to markets (locational access as a place o f origin for
trips) and to input factor sources (access as a place o f destination for trips).
Fourth, it depicts the technological conditions prevailing there. H o w much
each o f these factors contributes to each element in A is unknown, intractable,
and indeed irrelevant in this discussion.
In combination, all these factors affect production, consumption, market
(exchange and inter- as well as intralocational trade), and public decision­
making conditions, which in turn may entail advantages or disadvantages at
that location. These factors indirectly affect interaction and flow o f stocks.
Whereas the matrix o f real ( i . e . , any negative or positive) coefficients [a]
contains entries, each o f which identifies in a bundlelike manner factors
directly associated w i t h and affecting (facilitating or impeding) interaction
(including interlocational trade, and flow o f stocks), these coefficients identify
comprehensive cross-elasticities o f locational comparative advantages w i t h
respect to stock sizes, since

{dFJdxjViFtlxj) = a . 0 (7)

W i t h i n these cross-elasticities, effects o f physical and other (ecological) dis­


tances are discounted. Parameters [a] could also be interpreted as comprehen­
sive (political, environmental, social, and not merely economic) prices o f
interaction among locations i and j , where transportation (and other spatial
transaction) costs are also included. Elasticities exponents weigh the current
relative size o f the stock(s) at different locations in the F function. Obviously,
the magnitude o f each o f these coefficients, as well as their sum in each o f the
F functions, is o f considerable interest, as w i l l be discussed below.
In a hierarchy o f speeds, the state variables, x, are the fastest moving
entities in this model. Following the so-called slaving principle (Haken 1982),
their behavior is thought o f as being slaved by the parameters [a], which are
slow-moving entities, while parameters A vary even at a slower pace.
246 Chaos Theory in the Social Sciences

Scaling the One-Stock, Three-Location Model


Taking location one as the reference (or numeraire) location, one obtains the
following repetitive or reproductive process based on comparative advantages
set in reference to the numeraire location, found by dividing the comparative
advantage functions, F , , by the comparative advantage function o f location
one, F > 0:x

* , ( / + 1) = + F (t)
2 + F (t)]
3 = 1 - x (t 2 + 1) - x (t 3 + 1) (8)

x (t
2 + 1) = F (t)l[\ 2 + F (t) 2 + F (t)]
3 (9)

x (t
3 + 1) = F ( / ) / [ l + F (t)
3 2 + F (t)]
3 (10)

In F (t) 2 = In s& + a 2 2 l In x (t)


x + a 22 In x (t)
2 + a 23 In x (t)3 (11)

In F (t) 3 = In s& + a 3 3 1 In x (t)


x + a 32 In x (t)
2 + a 33 In x (t)3 (12)

s& , sâ 2 3 > 0; -oo < a , a , a , a , a , a


2 1 2 2 2 3 3 1 3 2 3 3 < +oo, (13)

where

a
« y = o ~ a ; i = 2,3 XJ j = 1,2,3 (14)

^ 2 = A /A,, A 2 3 = A /A,.
3 (15)

In addition, and without any loss o f generality, the above can be further
algebraically scaled by assuming that A , = 1 and that all a = 0, that is, by XJ

assuming without imposing any restrictions on the model's dynamics that


F = 1.
x

Configuring the discrete dynamics system o f a single stock and three


interdependent (perfectly competitive in the case o f relative dynamics) loca­
tions i n such a manner supplies all the dynamical features recorded in the
nonlinear dynamics (mathematical) literature: a wide battery o f bifurcations
(see Thompson and Stewart 1986; Devaney 1989) are encountered in transi­
tions from any state to another by crossing critical thresholds o f distinctly
different dynamical behaviors. A l l possible events that can occur in t w o - or
three-dimensional maps (based on the horseshoe map, the hyperbolic toral
automorphisms, and the various forms o f attractors; see Devaney 1989, 159)
do occur in this map as w e l l . However, the topological properties o f the map
w i l l not be addressed here.
A complete menu o f dynamics is offered, inclusive o f fixed-point
(asymptotic or not, stable or unstable, cyclical, nodal, or saddle) attrac-
Cities as Spatial Chaotic Attractors 247

tors/repellers, periodic movement and cascades o f period-doubling cycles,


quasi-periodic (toroidal) motion, and finally aperiodic (chaotic) behavior. A n
extraordinary variety o f quasi-periodic motions, and numerous types o f chaotic
movement (involving fractal and nonfractal, strange and nonstrange attractors
or containers) is present i n this map (Dendrinos 1991a, 1991b, 1992a).
The discrete iterative process configured as a single-stock, three-location
case is universal because, first, the map's menu o f dynamical behaviors and
their list o f bifurcations is comprehensive; and second, because the variety o f
spatial patterns it can generate and thus describe is as extensive as any k n o w n
map. W i t h the one-stock, three-location configuration o f the multiple-stock,
multiple-location map, one has universality within universality in the discrete
iterative process.

The Multiple-Stock, Multiple-Location Case

The configuration in this case, where one wishes to describe the interdepen­
dent evolution o f many stocks allocated into many interdependent and inter­
acting locations, goes as follows:

{t
Xij + 1) = FipyY.fifi); i = 1,2, . . . J;j = 1,2, . . . J
(16)
l (t)
iXij = l;j = 1,2, - · · J.

Under this configuration, index i stands for locations and index j for stock type;
the functions F could be specified equivalently to those specifications given for
the case o f a single stock and multiple locations. H o w this multiple-stock and
multiple-location case could be used w i l l be briefly outlined later in the
chapter. However, since this configuration w i l l not be used extensively i n this
paper (which deals w i t h only a simple case), not much detail w i l l be supplied
here. The interested reader is referred to Dendrinos and Sonis 1990 for more.

The Universal M a p a n d Physical Space

In what follows, the subject o f evolution in the form o f human settlements is


addressed. A l l movements and bifurcations in phase space o f the universal
map o f discrete relative dynamics have been interpreted so far as associated
with changes in the relative distribution o f one (or more) stock(s) as found in
two or more locations. For example, human population could have been the
entity modeled (described) in the case o f a single stock. Population, wealth,
and capital stock could have been the physical entities abstractly and compre­
hensively modeled in the case o f multiple stocks, among two or more loca­
tions. Phase space in these iterative discrete dynamics has been abstract
248 Chaos Theory in the Social Sciences

relative stock size space, and not physical space. Indeed, locations have been
merely represented by an index, i, i = 1,2, . . . ,/.
Next, a different angle to the universal map is taken, whereby its phase
space is thought to be physical space. A n attempt is made to describe the
formation and evolution o f human communities or settlements comprehen­
sively in the form o f spatio-temporal chains, or a series o f points (sites)
inhabited by human population w i t h i n a specific area. Sites have different
temporal vintage, some being inhabited before others. These chains are pro­
duced as either fixed stable and unique points; periodic, attracting, stable foci;
quasi-periodic, cyclically formed rings; or strange, aperiodically formed at-
tractors in physical space.
The settlement formation (chaining) process to be presented next de­
scribes the form o f communities w i t h i n urban areas and produces intra-urban
allocations in the shape o f urban patches. I t also describes the formation o f
urban settlements w i t h i n nations or regions, producing interurban allocations
in the form o f urban strings. Hence, it can depict formation o f microsettle-
ments o f different vintages w i t h i n metropolitan areas, as well as the rise o f a
series o f interconnected urban macrosettlements appearing as linked clusters
o f urban agglomerations in the context o f nations or even continents. Put in
slightly different terms, what is shown next is the formation o f human settle­
ments (cities) w i t h i n spatio-temporal niches. Patterns in the evolution o f these
niches are also shown.
A question not often raised by spatial analysts is whether the formation
o f settlements at particular sites in space is due to forces operating exactly and
solely at these sites or whether the appearance, g r o w t h , and extinction o f
urban settlements at these sites is due to forces operating at a distance from
them. Obviously, there are endogenous as well as exogenous forces at work at
the sites in question, and the framework shown next is helpful i n putting this
issue in a spatial (physical) context.
Spatial allocations o f human population are influenced by the presence o f
topographical features (such as rivers, shores, mountains, valleys, lakes,
etc.), climatic conditions, or man-made landmarks, such as highways. Alter­
natively, transportation arteries and corridors (in the form o f rivers, shores, or
highways), can be looked at as spatial attractors or containers along which
human settlement activity periodically or aperiodically occurs.
One o f the principles embedded in the spatial shaping o f form emerging
from the universal map o f discrete iterative dynamics is the following: spatial
form, which contains components o f different vintage, is neither shaped nor
does it evolve in a locationally contiguous manner. Rather, it rises, transforms
itself, and disappears in a spatially discontinuous way. A broadly based prin­
ciple may underscore this finding: that spatial diffusion processes do not occur
through locational proximity, but rather through spatial ecological propin­
quity, which might involve spatial discontinuities (Dendrinos 1992b).
Cities as Spatial Chaotic Attractors 249

Stock Allocation and Conventional Analysis: Statics

Allocation o f a stock (for example, residential consumption or industrial


production activity) in a two-dimensional physical space (usually confined to
a triangle defined by points identified either as sources from which inputs
flow, or sinks toward which outputs flow) and based on economic agents
incurring some form o f a spatial transaction ( i . e . , transportation) cost and a
set o f spatially defined prices has been a long-standing problem i n spatial
sciences. I n spatial economics and geography, static long-run market (unique)
equilibrium (or welfare optimum) spatial allocations o f economic (production
and/or consumption) activities have preoccupied analysts for more than a
century.
Most o f this work is based on the (interurban) spatial economics o f
Alfred Weber (see Friedrich 1929 and Isard 1956), to the extent that produc­
tion is concerned. Intra-urban, as well as interurban, allocation o f a residential
or farming activity around a single point in homogeneous physical space
under isotropic transportation conditions is another closely related problem,
due to Von Thunen (see Alonso 1964). These classical problems in spatial
economics have preoccupied spatial analysts for decades.

The Alternative: Dynamics

A dynamic, much broader, comprehensive, rich, and abstract albeit descrip­


tive view o f spatial allocations in physical space is supplied next. The context
in which the model can be applied is both the inter- as well as the intra-urban
scene. The method goes far beyond a mere reliance on (slow-changing) trans­
port costs and/or (fast-altered) spatial prices. There is no requirement that
unique, long-term static spatial configurations be generated. There is also no
need to explicitly model the vastly complex interdependencies and ensuing
interactions among a variety o f spatial agents involved in shaping spatial form
through vastly complex consumption and production as well as market, ex­
change (trade), and governmental intervention mechanisms.
Instead, a small number o f variables and parameters are chosen to effec­
tively model or mimic reality, producing a state o f virtual reality, but not
reality itself. Interdependencies among these variables are captured in the
most general manner, attempting to describe a comprehensive menu o f physi­
cal forms o f human settlements. Deductive explanation, in the form o f ex­
plicit behavior by the many agents responsible for these motions, is largely
absent from the analysis.
Many factors are embedded in the bundlelike parameters included in the
configurations o f the universal map. They form the basis for a reductionist
explanation, which in turn is based on the notion o f (fast-moving) locational
comparative advantages. A few comments on this central notion o f compara-
250 Chaos Theory in the Social Sciences

tive advantages seem warranted, in view o f its rather confined definition and
use w i t h i n the theory o f international trade in economics.
Locational comparative advantages is a Darwinian-like notion in the
evolution o f human settlements. A m o n g other things, it calls for the survival
o f the most advantageous o f all spatial settlements, much like Darwin's notion
o f survival o f the fittest among various animal and plant species. Embedded
within this assertion is Darwin's principle o f evolution o f species through a
natural selection process. The process o f selection acquires a broader frame­
work under the mathematical properties o f chaos, as also does the notion o f
Malthusian-type dynamics, a notion that considerably influenced D a r w i n
originally.
Selection involves processes o f optimization and search for optima.
Questions regarding the optimization processes employed by social agents
individually and/or collectively (objective functions and constraints), and
numerous search mechanisms applicable thereon resulting in a variety o f
locally (as opposed to globally) optimum outcomes in socio-spatial evolution
are not the concern here. Mathematical chaos may be capable o f shedding
additional light on such socio-spatial processes, including processes equiva­
lent to those associated w i t h neural nets, cellular automata, and processes
o f evolution in genetics. But the central concern o f this paper is much
more narrow, and such considerably more elaborate discussions are found
elsewhere (Dendrinos ( w i t h M u l l a l l y ) 1985; Dendrinos and Sonis 1990;
Dendrinos 1992b).

Spatial Triangles and Their Antitriangle


Assume the presence o f three interdependent points (nodes) o f interest, defin­
ing an arbitrary triangular area. This triangle could be inside an urban area; it
could be formed by nodes found i n a spatially extensive national landscape
among widely separated urban areas; or it could represent the area o f an
island. Reconstructing a map o f archeological remains on the island o f Sicily
could provide a good forum to test the model, for example.
The problem is to derive a dynamic process that w i l l generate a map o f
human settlement activity, as it serially appears over a time horizon. Forma­
tion o f spatio-temporal chains o f human settlement activity inside the area o f
the triangle is defined partly by the location o f the three nodes and the forces
operating at them. Because the spatio-temporal chains are closely dependent
upon the exact location o f the three nodes delineating the area chosen to
reproduce the locational chains, choice o f nodes is critical.
Where a settlement appears at a point in time is subject to two sets o f
forces. Certain forces locating settlements at various points in the space
defined by the three nodes are exogenous to the specific location o f the cities
(or city) inside the triangle and attributed to its nodes.
Other forces that affect settlement formation inside the triangle are those
Cities as Spatial Chaotic Attractors 251

associated w i t h the previous settlements. W i t h i n its borders an original state


o f some human activity must be asserted. This initial condition affects the
location inside the triangle o f the subsequent settlements; it constitutes
the endogenous force. As it turns out, the original state (initial location) o f the
human stock i n the triangle is critical for the dynamics o f human settlement
formation at certain instances, whereas in others it matters little.
As a result o f spatial attributes, heterogeneity o f space, and the presence
o f the initial stock at some point within the triangle, spatial interdependencies
appear. Interdependencies among the three points (nodes) involve forces lead­
ing to a net pull or push force attracting toward or repelling from these nodes
human activity during subsequent time periods. This net pull/push force is a
result o f a tug-of-war between two opposing component subforces: attraction,
acceptance, and affinity demonstrated toward the (homogeneous) stock on the
one hand at a node or repulsion, expulsion, and hostility toward the stock at
another.
The three points (nodes) that define the spatial extent o f the physical
space under consideration are located in a two-dimensional Euclidean physi­
cal space that is characterized by some form o f spatial heterogeneity (or
anisotropy). These three points might be intersections o f a network involving
transportation corridors o f various forms (rivers, highways, ravines, etc.), or
peaks or valleys i n contour maps o f any kind (topographical, economic, or
any other type). A n y social stock (say, human settlement activity) moving into
the triangle is subjected to normalized net repulsion from or attraction to
forces the three nodes o f the triangle exercise on i t . Before taking a look at the
details o f these forces, the geometry o f the problem w i l l be addressed.
In a three-dimensional space the three normalized forces form a vector
3
lying w i t h i n an equilateral triangle in R (fig. 11.1). This equilateral triangle
can be transformed into an orthogonal triangle with unitary orthogonal sides
2
by a mere projection into R ( f i g . 11.2). The original three points can repre­
2
sent, as already stated, any arbitrary Euclidean triangle in R space (see fig.
1 1 - A l a in app. A ) . Such an arbitrary triangle can be transformed into a
Euclidean right-angle isosceles triangle (see fig. 1 1 - A l b in app. A ) ; under the
transformation, proportionality in distances o f any point inside the triangle to
its nodes is preserved. This problem and its associated proposition are pre­
cisely stated, demonstrated and proved in appendix A .
2
Hence, any triangle in R Euclidean space can be represented by a right-
angle triangle w i t h unitary orthogonal sides ( f i g . 11.2) and by a one-to-one
correspondence, shown in appendix A , the orthogonal triangle can be trans­
formed into an equilateral triangle o f the type shown in figure 11.1. A vector
2
in R identifies the location o f a settlement in the transformed space, where
the point associated w i t h that location is given by the coordinates (normalized
distances) from the point o f interest associated with the origin.
Let us now revisit and broaden the interpretation o f the universal discrete
252 Chaos Theory in the Social Sciences

X 2 X2W

P 2

Fig. 11.1. T h e phase p o r t r a i t o f Fig. 11.2. T h e phase space i n a n


t h e m a p for the one-stock, three- orthogonal triangle
l o c a t i o n case

map o f relative dynamics. Since on the horizontal and vertical axis one
measures the transformed and normalized distance between the origin (an
arbitrary point among the three, say point P in fig. 11.2) and one o f the other
3

points o f interest (P, or P ) , then [1 - *,(<*)] and [1 - x (t)] represent the


2 2

normalized proximity o f the settlement to the other two points o f interest,


correspondingly. Whereas the distance x , ( f ) is proportional to a net repulsion
force from the origin jettisoning the stock toward point P,, [ 1 - x , ( 0 ] is a
distance proportional to a net attraction force pulling the stock toward point
p.-
3
In R , one can form an equilateral triangle based on forces of net repul-
sion, the triangle at the northeastern section o f figure 11. A 2 i n appendix B .
Further, an antitriangle exists that corresponds to the original net repulsion-
based triangle. This antitriangle can be constructed from measures o f prox­
imity based on the equivalent forces of net attraction, and it is the shaded
triangle found at the southwestern part o f figure 11.A2. I n appendix B certain
analytical properties o f these two triangles are given.

Location as a Response to Attractive


and Repelling Forces
A unit o f human settlement is assumed to have already been located at some
3
point w i t h i n the space outlined by the nodes o f the equilateral triangle in R
(fig. 11.1). Thus, it identifies some original state (or initial perturbation) o f
the system, and it is responsible for triggering a sequence o f subsequent
spatial allocations, that is, a spatio-temporal chain o f settlement activity.
Subsequent allocations are the result o f either endogenous growth or net
in-migration, although this is a subject outside the scope o f this chapter. A
temporal sequence o f spatial allocations can be thought o f as a sequence o f
Cities as Spatial Chaotic Attractors 253

points where new activity by the same stock locates. This is a process o f
spatial colonization. The model could also be interpreted as simply represent­
ing a nomadic process, that is, a temporal sequence in the location o f the same
stock unit.
Interpreting the universal map in physical space, one obtains a process
that generates spatial form as a single frame m, m = 1,2, . . . ,M. Each frame
corresponds to a particular set o f parameter values, initial conditions, and
number o f iterations (time periods), all o f them constituent components o f the
phase space. B y consecutive changes in any o f these elements, one obtains a
process that includes a temporal sequence o f spatial frames (changes in the
spatial form, or phase portrait). I n particular, a change in parameter values or
initial conditions can result in a new frame, m, potentially characterized by a
qualitatively different dynamic; that is, a different phase portrait may emerge
as a result o f a bifurcation. This change constitutes the evolution in spatial
form principle.
In carrying out such changes in parameter values, a condition can arise
according to w h i c h the initial state o f the new frame is the end state o f the last
frame. This represents a last-point-out-first-point-in condition. A series o f
frames, M, that does not obey this last-point-out-first-point-in condition is a
series that contains frames where the slate has been cleaned at the end o f each
one and the initial state has been reset.
When each frame, m, corresponds to a particular set o f parameter values,
frames change when parameter values change. H o w frequently these parame­
ters change must be related to the time horizon, T , o f each frame, m. One
m

could suppose that the more frequently on the average a parameter changes,
the shorter the average length o f the r ' s .
m

3 2
By projecting the study's areal triangle from R into R , quantities x (0) t

and x (0) become the physical location coordinates o f the first settlement
2

(initial condition) (fig. 11.2). In cases o f quasi-periodic or nonperiodic spatial


formations the initial location o f the first settlement in the phase portrait is
significant. It loses its significance in case o f periodic or fixed-point spatial
forms i f the average length o f the time horizon, T, is long enough.
The next human settlement in the temporal sequence o f spatial alloca­
tions (chains) is assumed to occupy a particular point in the two-dimensional
physical space as a one-time-period delayed response to the conditions laid
out by the previous allocation, and in proportion to a force o f net attraction
toward points P, and P from point P (the origin), given the original perturba­
2 3

tion. Stated differently, the next settlement locates at a point at some distance
away from the origin proportional to the current net repulsion force along the
t w o axes.
The above is the locational rule underlying the physical configuration o f
the universal map o f relative discrete dynamics. In the example that follows,
254 Chaos Theory in the Social Sciences

only units o f human settlement (say, an individual, a family, or any other unit)
are thought to be allocated on the basis o f this rule. In case one wishes to
allocate at each time period a different quantity o f a single human a c t i v i t y —
for example, a varying quantity o f human population—then stock size must
be configured in the formulation; then the formulation w i l l include two state
variables, one depicting location and the other stock size, as in the case o f two
stocks and three locations mentioned above.
I f one wishes to consider additional stocks—for example, capital and/or
wealth allocation i n these spatial dynamics—then this is easily accomplished
by adopting a configuration capable o f accommodating additional stocks.
Such configurations w i l l not be elaborated, however, in this chapter. Only the
note w i l l be added that such configurations supply the comprehensive treat­
ment o f all land-use models found in the standard literature o f spatial model­
ing. Land-use models involve dynamics, either directly or indirectly, by way
of long-run static (equilibrium or disequilibrium) solutions obtainable through
some form o f an iterative process.
The magnitude o f the net repulsion force from the origin along the
horizontal axis (and equivalently the size o f the net attraction force toward P,)
is given by

F,(r) = AJ {t)f {t)fp);


x 2 t = 0 , 1 , . . . ,T, (17)

where A , is the level o f relative environmental and access (due to the topogra­
phy and relative location w i t h i n the landscape) favoritism toward location P,
and is constant over the time horizon, T, under consideration. Equivalently,
F (t) is also specified.
2

Functions f depict location-specific and relative temporal comparative


advantages due to current interdependencies as a result o f the current loca-
tional choice o f a unit o f human settlement; they are given by the specifica­
tions

Infp) = a XJ \x\xp); j = 1,2,3; (18)

where coefficients depict net repulsion force elasticities o f response to


normalized distance o f the human settlement from three points o f interest
(nodes) in reference to point P,, and also i n reference to a numeraire location.
Exponents (coefficients), [a], act as weights o f (sensitivity to) nor­
malized distance o f the settlement from the origin along the t w o axes. They
identify positive ( i f a > 0) or negative ( i f a < 0) cross- or interspatial
(j y

external effects (externalities) in reference to the numeraire location. Coeffi­


cients [a] differ partly because o f the heterogeneity o f physical space. These
weights (elasticities) do not change w i t h i n the time horizon, T, although they
may change faster than those in A . Parameters a, could be construed as
Cities as Spatial Chaotic Attractors 255

ecological (comprehensive) prices directly responsible for interaction among


sites i and j .
The magnitude o f these coefficients, individually and in their sum, is o f
considerable interest. I f an exponent is in absolute value greater than one,
then this implies considerable sensitivity in the connectance among two nodes
appropriately weighing the distance in question, and in reference to the nu-
meraire node location. In contrast, in the case o f a magnitude less than one,
relative insensitivity must be present.
In each F , i f the sum o f the exponents exceeds one in absolute value then
one might consider the force F to be under increasing returns; whereas i f the
sum is less than one then the force F is under decreasing returns to scale. This
horizontal sum depicts the effect o f all nodes upon a specific node. Moreover,
i f the exponents are summed vertically, then the effect a node has upon all
nodes is captured. Whether the sum is negative or less than one in absolute
value is also o f interest.
I f the sum is negative, then a competitive association must be inferred
between the node in question and its rivals (the other t w o nodes); on the other
hand, i f positive, then a cooperative association must be present. I f the sum is
less than one in absolute value, then one is confronted w i t h a relatively m i l d
interaction among them. It is stressed that these characterizations apply to the
nodes and not to the actual location o f the human settlement.
Further, it is underscored that spatial agglomeration or déconcentration
is not a force under these specifications (that is, a cause for or against concen­
tration o f human settlements in space). Rather, it is the end result or the
impact from these spatial forces o f net repulsion or attraction present w i t h i n
the triangle at any time period, /. Both o f these statements are to an extent at
variance w i t h currently prevailing theories o f settlement formation.
The problem thus is to derive a temporal sequence o f unit human settle­
ment allocations w i t h i n the triangle, due to either migration or endogenous
growth, and subsequently analyze their form. A number o f cases w i l l be
discussed next, indicative o f the variety in form o f spatial agglomerations one
can obtain using this generic algorithm.

Periodic, Quasi-Periodic, a n d Chaotic


Spatial S e t t l e m e n t s

In what follows one must keep in mind that the phase portrait represents
transformed actual space. First, the case o f a very efficient specification o f the
map w i l l be shown. It has only one or two parameters that are not either zero
or one (in absolute value) and is capable o f generating periodic and quasi-
periodic (in the case o f a single parameter), as well as chaotic (in the case o f
two parameters) form. Then, a more complicated and less efficient specifica­
tion o f the map w i l l be shown. Its phase portraits w i l l be derived in a very
256 Chaos Theory in the Social Sciences

small neighborhood o f the parameter space, where various types o f chaos are
present. The case demonstrates the encompassing behavior o f the universal
map in its capacity to generate and replicate human settlement variety ( i n the
shape o f individual frames, or phase portraits) and their evolution.

Efficient Versions of the Map

Next, t w o sets o f examples are presented. I n the first set, the case o f periodic
movement is shown in a specification o f the map involving only one parame­
ter being neither zero or one in absolute value. H o w this periodic map is
transformed into a quasi-periodic one is demonstrated by a very short trip
along the dimension o f a single parameter in parameter space. The second
case presents the shaping o f spatial form as either m u l t i r i n g , quasi-periodic,
or chaotic motions and it involves a two-parameter specification o f the map.

Periodically Derived Forms


First, consider the specification o f the universal map o f discrete relative
2
dynamics configured as a three-location, one-stock model in R and specified
in the most efficient manner. A n example o f such a model is given by the
parameters

0 0 0 " 1"
[a] = -1 -.1111 1 A = 1
. 1 -1 0 . . 1 _

where the initial state is neutral (JC,-(0) = .333 . . . , i = 1,2,3). The iterative
process under this specification is as follows:

In xi(t + 1) = - I n F(t)
In x2{t + 1) = - I n jc,(r) - .1111 In x2(t) + In x3(t) - In F(t)

(20)
In x3(t + 1) = In JC,(0 - In x2(t) - In F(t)
F(t) = 1 + x2(t)' ilux3(t)/xt(t) + Xl(t)/x2(t),

where xx(t) and x2(t) are plotted in the phase portrait. I n this specification there
is only one parameter, a22, that appears to affect the system's dynamics.
Location-node "one" is the reference or numeraire site. The returns to scale i n
all elements o f F ( i . e . , all three corners o f the triangle) are approximately but
not exactly equal to zero, and the environment is neutral to all (as all elements
o f A are equal to one). The dynamics are shown in figure 11.3, panel (a) and
la)

(b)

Fig. 11.3. T w o e x a m p l e s d e m o n s t r a t i n g f o r m a t i o n of spatial h u m a n


s e t t l e m e n t s (see t e x t f o r p a r a m e t e r specification): {a) o n e - p a r a m e t e r
case: 4 - p e r i o d cycle; ib) t w o - p a r a m e t e r case: quasi periodicity.
258 Chaos Theory in the Social Sciences

represent a four-period cycle spiraling from the center o f the triangular space
(the initial state) toward four particular points very close to its sides. This
example may be modeling the formation o f very stable, highly dense, small in
area, fortified frontier cities o f the past.
Then, consider the very close neighborhood case o f specifications, i n ­
volving t w o parameters:

0 0 0 " 1"
a] = -.77.. -.33. 1 A = 1
1 -1 0 . . 1 .

where the initial state is x,(0) = .333 . . . , / = 1,2,3. This example can be
thought o f as an evolution o f the first specification through a change in only
two parameters, namely a 2l and a . 22 In this case the returns to scale are
exactly equal to zero at all locations (nodes), while the environment remains
neutral. The phase portrait is shown in figure 11.3, panel b , and it represents
the formation o f a quasi-periodic motion (a rotating nine-period cycle) ap­
proaching (but never falling on) the three sides o f the triangle including its
three nodes. Such a spatial form may be seen i n , among other cases, traffic-
ways and traffic patterns o f modern day metropolises. It is again noted that the
shape shown in figure 11.3, panel b is in transformed physical space.

Quasi-Periodicity and Chaos in an Efficient Specification


Consider the f o l l o w i n g efficient specification o f the universal map o f discrete
relative dynamics, involving t w o parameters and configured in its three-
location, one-stock case and interpreted as a physical space allocation o f a
2
single stock in R

0 0 0 1
3
[a] = -1 -1 0 A = 10"
. 1 -1 a 33 . . 1 -

where the neighborhood o f varying the parameter a 33 is 1.15 < a 33 < 1.2 and
where the environment is no longer neutral (A 2 < 1). According to the above
3
specification, the vector in R ( f i g . 11.1) is

In (t
Xl + 1) = - I n F(t)

In x (t 2 + 1) = - 6 . 9 0 7 8 - In x (i) x - In x (t) 2 - In F(t)


(23)
In x (t 3 + 1) = In JC,(0 - In x (t) 2 + a 33 In x (t) 3 - In F(t)

F(t) = [(1 + 10- )/(x,(0* (0)]


3
2 + [(x,(/)A: (0) 33 3
a
(x (f))].
2
Cities as Spatial Chaotic Attractors 259

2
The physical space, in R and in figure 11.2, is given by plotting x {i) and x

x (t)- Under the above specifications, the environment is handicapping point-


2

node P , and the comprehensive access (interaction elasticity) depicting pa­


2

rameter a reveals the presence o f internal increasing returns at the o r i g i n ,


33

point P , as a
3 > 1. The returns to scale are negative at P ( F ' s sum o f
33 2 2

elements o f [a] is —2), whereas they are positive (but less than one) at P . 3

Two phase portraits from this example are shown in figure 11.4. In figure
11.4, panel a, the case o f a multiring spatial form o f settlements is shown
away from node P . Such form can be detected in the location and areal size
2

o f a number o f cities concentrated within a subregion o f the triangle under


consideration, thus producing some form o f spatial dualism.
Figure 11.4, panel b , demonstrates a regime o f spatial form generated
out o f a chaotic motion. This can be seen i n , among many other cases, the
suburbanization phenomenon, whereby a number o f initially isolated settle­
ments are gradually j o i n e d by one (or more) spatially expanding core(s).
The t w o cases shown in figure 11.4 can be also construed as a set o f t w o
successive frames involving a transformation o f spatial form through a trip in
parameter space at a very small neighborhood o f a , as the intralocational
33

returns to scale there decrease very slightly. Evolution o f spatial form is


depicted by this transformation. It is noted that the chaotic form in figure
11.4, panel b , has fractal properties.
In both o f the cases shown in figure 11.4, the initial conditions are not
important w i t h regard to the formation o f the final form when vintage is
suppressed; however, they are extremely important to the sequence involved
in generating the final form.

Nonefficient Specifications

The f o l l o w i n g cases are presented as indications o f the variety o f quasi- and


2
nonperiodically derived spatial forms embedded in the universal map i n R ,
as the map loses in efficiency. They are all found in a small neighborhood o f
the parameter space o f a five-parameter specification o f the one-stock, three-
location configuration o f the map at the point

0 0 0 1
la] = .28 -1 «23 A = 1
5
. -1.11835 -2.28 -1 .10" .

In figures 11.5 and 11.6 some spatial forms are shown, as phase portraits o f
the dynamics corresponding to a set o f parameter values for a in the range 2 3

0.6 s a s 1. Figure 11.5 contains four cases o f quasi-periodic motion,


2 3

whereas i n figure 11.6 four cases o f nonperiodic (chaotic) dynamics are


supplied.
Fig. 11.4. T w o examples of spatial h u m a n s e t t l e m e n t f o r m a t i o n (see t e x t
f o r parameter specification): (a) m u l t i p l e ring structure ( a = 1.188 . .. );
3 3

ib) chaos ( a = 1.177 . . . ).


3 3
Cities as Spatial Chaotic Attractors 261

Transformation o f a smooth ring into an edged but continuous line i n


2
R , which in turn is changed into a set o f ringlets and back into a sharply
edged quasi-periodic structure is shown in figure 11.5, as the interaction-
enhancing coefficient a linking points (nodes) P[ and P increases. Spatial
3 1 2

forms i n the shape o f rings are not frequently observed in human settlement
formation processes. These must represent fragile spatial forms, not lasting
long time periods.
Bifurcations involving various types o f chaos are shown in figure 11.6,
as the elasticity, a , increases further. The chaotic spatial allocations form a
3 1

Cantor-type dust cloud, w i t h varying densities i n space. A l l these chaotic


attractors have a fractal Hausdorff dimension (see Schuster 1988, 55) greater
than one. Their mathematical properties are topological rather than simply
algebraic, and they are: (1) sensitivity to initial conditions; (2) topological
transitivity; and (3) dense periodic points (Devaney 1989, 169).
In the transformations shown in figures 11.6, panels (a) through (d), a
number o f chaotic spatial allocations and their changes are depicted as one
travels through parameter space. Specifically, figure 11.6, panel (c) picks up a
short-lived spatial form, whereby spatial consolidation and an increase i n
densities is an intermittent event. Intermittent events are novel aspects o f
dynamical behavior revealed through mathematical chaos; they point to the
possible presence o f many but not long-lasting conditions o f approximate
periodicity found in a sea o f chaotic motions on trips through parameter space.
Chaotic dust clouds look very similar to spatially distributed human
settlements, particularly those found i n early historical and prehistorical sites.
They also seem to mimic land-use patterns o f preindustrial or early industrial
cities. Such forms are the outcome o f laissez-faire-type market processes,
where the influence o f governmental action and urban design is minimal i n
individual social motions and actions.
The final, and maybe the most important, point to be made from these
examples is that recognizable spatial patterns (of human settlements in this
case) are equivalent to detectable order in chaos, as emerging from interde­
pendent nonlinear spatial dynamics. Spatial forms o f later vintage may thus
simply be remnants o f order in chaotically shaped earlier settlements. For
instance, waves and differential settlement densities can be depicted in the
examples found in figures 11.6, panels (a), (b), and (d).

Conclusions

Mathematical chaos has made considerable inroads into social sciences, rang­
ing from marginal effects in economics to possibly basic contributions in
geography. A brief commentary has been presented on the possibilities and
obstacles currently present in further applying dynamical analysis in social
f

\
\

(c)

id)

Fig. 11.5. Four examples i n v o l v i n g m o r e c o m p l e x f o r m s of spatial settle-


m e n t s g e n e r a t e d b y q u a s i - p e r i o d i c m o t i o n (see t e x t f o r p a r a m e t e r spec-
i f i c a t i o n ) : (a) a
3 1 = .61 (regular r i n g ) ; [b) a 3 1 = .671 (edged r i n g ) ; (c) a 3 1

= .68 (13 r i n g l e t s ) ; (d) a 3 1 = .70833 (ring).


266 Chaos Theory in the Social Sciences

fields of inquiry. W i t h i n this context, a universal map o f discrete relative


spatial dynamics, developed in the socio-spatial field, has been briefly dis­
cussed.
A n extension i n the capacity to interpret the universal map o f discrete
spatial relative dynamics has then been presented. Appropriately transformed
physical space is seen as the phase portrait o f the iterative process. A
number o f quasi-periodic, nonperiodic, and periodic dynamics are considered
as the underlying mechanism giving rise to a variety o f physical forms as
spatio-temporal chains o f a time series o f population allocations. A n example
is given, involving human settlements o f different vintages, as allocated
w i t h i n a specified area o f physical space confined by three points o f interest.
Evolution i n the physical form o f settlements can then be studied as a case in
bifurcating behavior o f nonlinear interdependent dynamics.
Statistical testing and verification o f such model forms against actual
human settlement formations at various regions and subregions o f the globe
and at various epochs seems to be a necessary next step. Tests o f significance
involving both locational and temporal proximity o f such chains to actual
sequences o f human habitats (or other events occurring in a confined space,
ranging from epicenters o f earthquakes along fault lines to points o f riots in
urban areas and the allocation o f buildings w i t h different vintage along central
city roads) could considerably enhance this effort. It is pointed out, however,
that one ought not have high expectations i n such statistical attempts for
verification when chaotic regimes prevail, beyond looking for rough qualita­
tive indications. A number o f reasons are given for this.
Single-frame dynamics, corresponding to a unique set o f parameter
values and initial conditions, are discussed as g i v i n g rise to physical form
corresponding to a particular chain o f spatially distributed human settlements
and its evolution. However, superimposition o f frames (simultaneously incor­
porating phase spaces corresponding to a range o f parameter values and/or
initial conditions) could also be considered. Superimposition o f spatio-
temporal chains is an interesting component i n the definition, testing, and
interpretation o f socio-spatial dynamics (see Dendrinos 1991a). I t underlies
issues o f structural stability o f spatio-temporal forms. Future research along
these lines seems quite promising, indeed.

APPENDIX A: T H EG E O M E T R Y OF LOCATION IN R 2

The problem in this appendix is to define a transformation among two triangles,


triangle A B C in figure l l . A l ( a ) and triangle A ' B ' C in figure l l . A l ( b ) , such that
there is a particular one-to-one correspondence between any two points (points D and
D') within them. The particular transformation under consideration is that specific
Cities as Spatial Chaotic Attractors 267

Fig. 11.A1. A n a r b i t r a r y t r i a n g l e (a), a n d its t r a n s f o r m a t i o n (6)

ratios of measures are preserved. Given the audience of this book, simple geometry
will be used to demonstrate the statements, instead of matrix algebra and topology.
For any arbitrary point in triangle A B C , say point D , one can identify the
following two ratios (parentheses stand for measure or length): ( B E ) / ( E C ) = a , and
( A D ) / ( D E ) = b. With these two ratios alone, one can uniquely identify the location of
point D in the triangle A B C . One can now show that: (1) measures ( B D ) and ( D F )
obey a specific ratio that only depends on the quantities a and b, as do the measures
( C D ) and ( D H ) ; and (2) these ratios are independent of the form of the triangle A B C ,
say the sine or cosine of each of its three angles or the length of its sides. In other
words, given any two arbitrary triangles, A B C and A ' B ' C , to any arbitrary point D in
A B C one can assign a unique point D ' in A ' B ' C so that certain ratios are preserved.
268 Chaos Theory in the Social Sciences

T H E O R E M . The ratios (BD)I(DF) and (CD)I(DH) are functions of only a and b.

PROOF. Draw line F G parallel to B C . Then ( G D ) / ( D E ) = (GF)/(BE) =


(GF)/a(EC) = (AG)/a(AE). Further, (AE) = (AD) + (DE) = (AD) +
(AD)[(£> + l ) / b ] , and b = (AD)/(DE) = [(AG) + (GD)]/(DE) = (AG)/(DE)
+ ( A G ) / a ( A E ) . Thus, b = 6 ( A G ) / ( A D ) + ( A G ) / a ( A E ) and ( A G ) = b/[b/(AD)
+ l / a ( A E ) ] . From the above one obtains: ( A G ) = (AD){[a(6 + 1)]/[1 + a(b + 1)]}.
Now, one also has: ( B D ) / ( D F ) = ( D E ) / ( G D ) = ( A D ) / & ( G D ) ; since ( G D ) = ( A D ) -
(AG) = (AD){1 - [a(b + 1)]/[1 + a(b + 1)]} = ( A D ) / [ 1 + a(b + 1)], then the
previous expression becomes ( B D ) / ( D F ) = [1 + a(b + l)]/b. QED.

The proof that the other ratio, ( C D ) / ( D H ) , is also a function of a and b alone is
identical to the above and thus omitted. Hence, for a unique point in A B C specified by
the ratios a and b, one can locate in any other triangle, including the isosceles
orthogonal triangle of figure 1 1 . A 2 , a unique point preserving certain ratios that are
functions of only a and b. Instead of working with any triangle A B C , one can, as a
3
result, work with A ' B ' C , that is, the phase portrait of the universal map in R as
2
projected in R .

APPENDIX B: T H E REPULSION TRIANGLE


AND ITS ATTRACTION ANTITRIANGLE

In the one-stock and three-location configuration of the universal map of discrete


3
relative dynamics, the phase portrait is contained in the positive domain of R and
confined by the shaded triangle at the upper right hand side of figure 11.B1, since

2,.jr,(f) = 1, i = 1,2,3; t = 0 , 1 , . . . , 7 .

Such is also the case for the physical space interpretation of the one-stock, three-
location configuration of the universal map. In this case, the set of state variables, x,
depict normalized distances away from the respective nodes, the end result of local net
repulsion forces operating at the three nodes. Thus, this is a repulsion triangle.
By forming the quantities ~x,{t) = [1 - *,(/)], /' = 1,2,3, one obtains normalized
(but not relative) proximities to the respective nodes, so that their sum is equal to two.
As a result, one can find the iterative dynamics of the normalized proximities on the
shaded triangle located at the lower left-hand side of figure 11.A2. The triangle is
defined by the condition

2,.-jt,.(f) = 2 , i = 1,2,3; t = 0,1, . . . ,T,

3
in R . This is an attraction of the antitriangle to the repulsion triangle shown above. Its
sides are equal to twice the square root of two. In it, the phase portrait of the
corresponding dynamics of attraction is contained; on this phase portrait the original
Cities as Spatial Chaotic Attractors 269

1-X,(t)

Fig. 11.A2. A n equilateral triangle of repulsion forces in phase space


and its attraction forces antitriangle

dynamics from the repulsion triangle are blown up (exactly twice in size) and pro­
jected.
As distance from repulsion is given by the general (universal relative dynamic)
expression x (t + 1) = F^lzZjFp),
t proximity is given by ~ (t + 1) = 1 - x£t + 1)
Xj

= [ 2 , ^ F / r ) ] / [ S , F / / ) ] . It follows that

[-xff + l)V[ (t
Xi + 1)] = 2 ^ , . F / 0 ; i = 1,2,3

2 * , ( ' + 1) = [2,x,-(r + 1 ) ] [ 2 ^ , F / 0 I = 2.
r

+
By specifying x,(t + 1) = 2x,(r + 1) = 2[1 - ~x,(r + 1 ) ] , then one has:

+
2,- x,-(/ + 1) = 2,-*,-(f + 1) = 2.

+
Quantity JC,-(0 may be viewed as a measure of local push equal to twice the magnitude
of the local repulsive force operating at each node, i.
Part 4
Implications for Social Systems
Management and Social Science
CHAPTER 12

Field-Theoretic Framework for t h e Interpretation


of t h e E v o l u t i o n , Instability, S t r u c t u r a l Change,
and Management of Complex Systems

Kenyon B. De Greene

Policy making and decision making and other aspects o f the management o f
complex systems are becoming increasingly difficult. Management philoso­
phies, approaches, and techniques were developed during simpler times.
However, complex systems are dynamic rather than static, evolve or are
driven into domains o f instability, and emerge into new structures. There is
now a growing gap or loss o f fit between our systems-management capa­
bilities and the real w o r l d . Policymakers and decision makers must deal
especially w i t h severely reduced time frames, consequences-of-action uncer­
tainty, and actions that produce diminishing returns.
Fortunately, advances in systems theory provide a means for narrowing
the gap and providing a better fit between systems management and reality.
The theories have multifarious origins and different emphases. This question
therefore arises: how can these theories be interfaced and interrelated? M y
approach is to use the integrating framework o f field theory. As has histori­
cally been the case with the development o f field theory in the several sci­
ences, field theory is not so much a theory in itself as a framework for
theoretical orientation and study.
I have coined the term world system-field ( o f forces), which consists o f
the dynamically interacting world societal field and the world ecosystem. A
field describes the structure and behavior o f space-time and the placement o f
objects in space-time. I n fields, myriad state variables depend continuously on
the space coordinates. Einstein's field equations relate the properties o f space-
time to those o f matter, but thinking has proceeded quite far beyond Einstein's
original rather static formulation that denied irreversibility and evolution
(Nicolis and Prigogine 1989). A field can also describe the interactions be­
tween a system at a given hierarchical level o f the organization o f matter,
energy, and information (e.g., a human institution) and its external environ­
m e n t s ) . A (turbulent) field can induce structural changes in the various parts

273
274 Chaos Theory in the Social Sciences

o f the environment itself (e.g., De Greene 1990a). A n d in the vertical dimen­


sion, a field can describe micro-macro interrelationships. A field can be
viewed as a macrolevel, semiautonomous, collective structure (or order pa­
rameter) that exists over a relatively limited period o f time. Interactions at a
more microlevel generate the field via positive-feedback mechanisms, and the
field, reciprocally, through negative feedback, delimits the realm o f the pos­
sible at the level o f individual interactions (e.g., De Greene 1989; Weidlich
and Haag 1983).
Fields show differential sensitivity over time and space, and a fluctuation
or perturbation that is usually damped may surprisingly trigger a violent and
explosive reconfiguration o f the field. I t is difficult, perhaps almost always
impossible, for human actors to know the present stability state o f the system-
field. Hence, policies and decisions must be developed that preclude to the
extent possible the precipitation o f the system into an unsuspected realm o f
instability and reconfiguration. O n the other hand, these policies and deci­
sions must not result i n rigidity and the stultification o f desirable evolution and
structural change. Field theory provides a framework for dealing with these
complex issues.
This chapter is a presentation, w i t h considerable further thinking and
updating, o f my research in this area over the past several years. In the sense
of fields within fields within fields, a term sometimes used by systems theo­
rists, the top-level organizing approach is the evolution, stability/instability
features, and structural change o f / w i t h i n the w o r l d system-field and the ef­
fects o f these dynamics on constituent systems. W i t h i n this context the next
level o f organizational emphasis is the order parameter as a sociotechnical,
techno-economic, and macropsychological (collective-cognitive) structure.
The third level o f contextual approach is provided by the theory o f the eco­
nomic long wave or Kondratiev cycle/structure. W i t h i n this overall frame­
w o r k , constructs from a number o f systems theories are utilized. Examples
include the critical threshold or bifurcation point, far from equilibrium, order
through fluctuation, and irreversibility. The constructs are general purpose in
physical, biological, and social science, contributing to laws o f nature, but
they receive specific meaning only as properties o f the given field o f forces.
Deterministic chaos is treated w i t h i n these contexts, but the calculational
specifics w i l l not receive attention here because they are discussed more fully
in other chapters o f this book.
A n additional integrating theme o f this chapter is the wearing out and
exhaustion o f given order parameters or paradigms and their replacement by
paradigms that are better fitted to the new evolutionary situation. Under such
conditions we must reexamine our very methods o f inquiry about nature and
society, as w e l l as status quo policies and practices. Paradigm change pro-
Interpretation of Complex Systems 275

vides new, often surprising challenges, opportunities, and choices. The chap­
ter discusses important systems management consequences o f structural
change and paradigm shift. The chapter concludes with some thoughts about
the meaning o f the new theories to the social sciences.
The approach is transdisciplinary. That is, basic constructs are sought
from the physical and biological, as well as the behavioral and social sciences.
Indeed, many o f the constructs had their origins in the physical sciences.
Transfer o f constructs provides an exciting challenge.
The bibliography at the end o f this book provides a rather comprehen­
sive literature on systems theories and models, definitions o f terms, and
relevant examples. The reader should turn to these references for a fuller
understanding o f the basic ideas and o f previous work. Space limitations
preclude our starting from scratch in this chapter.

E v o l u t i o n , Instability, a n d S t r u c t u r e
of C o m p l e x S y s t e m s

A l l systems evolve, although the rates o f evolution may vary over time both
between and w i t h i n systems. The rate o f evolution is a function o f both the
inherent stability o f the system and changing environmental circumstances.
But no system can be stabilized forever. For the universe as a whole, an
isolated system, time's arrow points toward greater and greater breakdown,
leading to complete molecular chaos, m a x i m u m entropy, and heat death. For
open systems, including the l i v i n g systems that are o f major interest to us and
that interchange matter and energy w i t h their external environments, time's
arrow points to evolution toward greater and greater complexity. Thus, the
universe consists o f islands o f increasing order in a sea o f decreasing order.
Open systems evolve and maintain structure by exporting entropy to their
external environments.
The concept o f evolution is finding increasing application in physics,
chemistry, astronomy, and astrophysics, as well as, o f course, in biology,
geology, and paleontology. But it appears that all too few behavioral and
social scientists use the evolutionary framework, and that all too many theo­
ries, hypotheses, and empirical research efforts are directed toward the static,
the cross-sectional, the linear, the equilibrium seeking, the stable, the revers­
ible, and the structurally constant. These efforts, as interpreted here, operate
w i t h i n a prevailing but exhausted paradigm. This chapter and this book repre­
sent attempts to move beyond this paradigm.
In organic evolution, as summarized by Neo-Darwinism, two principal
processes are involved. First, random mutations lead to genetic differences
among organisms. Second, natural selection tends to enhance the likelihood
276 Chaos Theory in the Social Sciences

of genes that produce adaptive characteristics. Some parts o f the genome may
be more susceptible to mutation than are others, and environmental factors
may also affect mutation rates. But any given mutation is considered to be
independent o f any adaptive or survival value imparted to the organism.
Whether Neo-Darwinism provides a complete answer to organic evolution
remains to be seen, and the issue continues to be debated. Alternative expla­
nations are often dismissed as harking back to Lamarckism and teleology. But
it seems intuitively unlikely that the organic evolutionary system would func­
tion open loop; that is, w i t h no feedback between effect and cause.

D i s s i p a t i v e - S t r u c t u r e Theoretical I n t e r p r e t a t i o n s

Following dissipative-structure theory (Nicolis and Prigogine 1977, 1989;


Prigogine and Stengers 1984), the evolution o f open systems can be inter­
preted as follows. Movement o f the system away from equilibrium, associ­
ated w i t h some internal irreversible processes, increases the rate o f dissipation
as measured by the entropy production. Instability, triggered by nonequi-
l i b r i u m environmental conditions, leads to further dissipation and entropy
production; this in turn leads to the appearance o f further instabilities. Farther
from equilibrium, the probability increases that the system, with its internal
processes, is unstable w i t h respect to given fluctuations.
This concept o f evolutionary feedback regards energy dissipation as the
driving force o f evolution. I t is characterized by nonequilibrium conditions
leading to the system's crossing a critical threshold. Beyond this threshold the
system becomes structurally unstable with regard to the fluctuations, which
leads to increased dissipation and, in a positive-feedback loop, change i n the
threshold. Thus, there is an acceleration o f irreversible evolution over time.
Each new stage o f organization has the potential for further evolution.
A n important aspect o f dissipative-structure theory and o f its and other
interpretations o f evolution is the emphasis on the complementarity o f sto-
chasticity and determinism (or, put another way, chance and necessity, nov­
elty and confirmation). Fluctuations (e.g., genetic mutations) arise randomly.
Far from a critical threshold, called a bifurcation point, the larger system
tends to express average behavior; that is, to follow the law o f large numbers,
with the damping o f fluctuations. But near a bifurcation point, the fluctuations
may self-amplify or cross-amplify (autocatalysis and cross-catalysis) via posi­
tive feedback to produce a nucleation. The nucleation may then enter into
conflict w i t h the larger system or external environment. The latter may still try
to squelch the g r o w i n g nucleation, but sometimes the nucleation prevails and
becomes a new self-organized system. Dissipative-structure theory is also
called order through fluctuations. Self-amplification is an example o f non­
linear behavior, a topic widely discussed throughout this book.
Interpretation of Complex Systems 277

Logistic E v o l u t i o n and t h e Constraints


of Occupied Space

A further kind o f evolution, logistic evolution, has been studied by


dissipative-structure theorists (e.g., Prigogine and Stengers 1984), as well as
others. The evolution o f species can be interpreted as the filling or creating o f
ecological niches by successive species, each o f which is better fitted than was
its predecessor. The well-known mathematical model, the logistic differential
equation, which defines exponential growth limited by a carrying capacity or
saturation level, expresses population growth w i t h i n the niche. But the carry­
ing capacity need not be fixed, so it is the sequence or family o f logistic
functions (or better, logistic functions followed by domains o f diminishing
returns) that is o f the greater interest in this chapter.
The question must be raised, however, as to how much more morpho­
logical diversification is possible, once most biospace has been occupied. Put
another way, as a system-field achieves structure, does not this structure now
constrain further differentiation? W i t h regard to the evolution o f arthropods,
most divergence occurred i n the Cambrian period, early in the evolution o f
multicellular animal organisms. A rigid evolutionary constraint has precluded
further diversification o f body plans. A l l possible body plans had become
exhausted by the end o f the Cambrian, and further evolution has involved only
convergences and reversals. B y the end o f the Cambrian, the limits to d i ­
vergence had been reached. The reasons for this constraint are unknown. (See
the Technical Comments, "Cambrian and Recent Morphological Disparity,"
Science, 258, 1816-18, 11 December 1992.)
Societal evolution, as interpreted in this chapter, utilizes the ideas just
summarized. The fluctuations can be scientific discoveries, inventions, tech­
nological and social innovations, and great persons who appear at the right
time and place in history. Logistic evolution with diminishing returns is ex­
pressed as the evolution o f successive Kondratiev cycles/structures. The satu­
ration and constraining impacts o f collective cognitive (macropsychological)
space and the exhaustion o f ideas and practices, at least occasionally and
temporarily, are emphasized.

Order Parameter as Macropsychological S t r u c t u r e

M u c h o f science is concerned w i t h the interrelationship between the micro­


scopic and the macroscopic. Evolving systems show feedback between macro­
scopic structures or collective fields and the events o f individual interactions
at the microlevel. Macrostructures or fields emergent from the microlevel in
turn modify the individual interactions at each stage o f irreversible evolution.
This is a vertical perspective, like that o f the familiar hierarchy/emergence
278 Chaos Theory in the Social Sciences

theory. For systems management purposes, it should be noted here that things
often look normal at the macrolevel o f the established order parameter, but
that at the microlevel things are seething w i t h incipient change. Preceding a
bifurcation o f the macrostructure, the system has been reorganized via long-
range correlations (see below).
A field i n the horizontal sense at any given time reflects the competition
between the stability provided by communication (and information) and the
instability that can follow fluctuation and nucleation. The result dictates the
threshold o f stability. The threshold o f stability reflects the interplay among
forces. I f the fluctuating area lies, say, below some critical point, the area w i l l
regress; i f it lies above, the fluctuation and nucleation can spread across the
entire field or take over the entire system. The field need not be limited to two
dimensions (physicists, for example, provide both t w o - and three-dimensional
models for many systems that show phase transitions.)
A n order parameter (De Greene 1989, 1993; Haken 1983; Nicolis and
Prigogine 1989; Weidlich and Haag 1983; Wilson 1979) is such an outgrowth
of micro-macro interactions. It is a macrostructure or macrovariable that
emerges along with a great reduction in the number o f degrees o f freedom o f a
system. The emergence o f an order parameter can also be viewed from the
horizontal perspective, but the vertical sense receives more emphasis in this
chapter.

The S t u d y o f Phase Transitions

The term order parameter arose i n the study o f the physics o f the equilibrium
phase transitions mentioned above. It was used to describe a number o f
phenomena that are the properties o f collectivities o f subatomic particles,
atoms, or molecules, rather than o f the sometimes randomly ordered particles
themselves. Order typically emerges at a critical threshold. Examples o f
physical order parameters include spontaneous magnetization, correlation
length, the properties o f metal alloys, density differences between the liquid
and gaseous phases o f a fluid, and concentration differences. The strong
fluctuations that characterize these phase transitions take place in the order-
parameter field. It is not necessary to know the precise form o f the strongly
nonlinear interactions. Collective behavior arises at a statistical level, and
only probabilistic predictions can be made as to the behavior o f any given
element.
The system may display, as a function o f fluctuations o f state near the
critical point, pockets o f structure or order embedded in regions o f disorder
embedded in still greater domains o f order, and so o n . Well above a critical
point (say the Curie point o f temperature that is associated w i t h magnetiza­
tion), the system's elements are arranged essentially randomly, and the system
shows only short-range order. A s the critical point is reached from above,
Interpretation of Complex Systems 279

larger-scale order begins to emerge. A t the critical threshold itself, discon­


nected patches o f order expand to infinite size, but fluctuations o f lesser scales
still remain. A critical point is a discontinuity, above and below which the
system is qualitatively different (e.g., paramagnetized and magnetized).
Associated w i t h the description o f the instability o f fields and the emer­
gence o f new structure w i t h i n fields is the concept o f correlation. I f a single
element in the system is perturbed, the disturbance may propagate, neighbor-
to-neighbor, across the entire field. Distant elements, now structured in the
same way, have become correlated. Correlation length indicates the maxi­
mum distance over which correlation can be determined. Near the critical
threshold, the correlation length grows rapidly. Long-range order has emerged
out o f the same short-range forces. A t the critical point itself, the correlation
length becomes infinite. Phase transition has not yet taken place, but the
system is now hypersensitive to very small perturbations, the effects o f which
can now nearly instantaneously explode across and engulf the entire system.

B i f u r c a t i o n a n d Order Parameter

Under nonequilibrium conditions, the amplitudes o f the long-range correla­


tions increase with distance from equilibrium. Even before bifurcation at the
macrolevel, the system has achieved a prestructure via the long-range correla­
tions.
The phase as well as the amplitude o f a bifurcating branch can be an
order parameter. Indeed, the very phenomenon o f bifurcation can be de­
scribed in terms o f the order parameter rather than in terms o f the state
variables originally present. Larger numbers o f equations can describe the
interactions among several order parameters. A s suggested above, order pa­
rameters can characterize the transition o f a system from an uncorrelated state
to a correlated state far from equilibrium. Near to a bifurcation point, a
multivariable system may be describable in terms o f a limited number o f
collective variables or order parameters (Nicolis and Prigogine 1989).
According to synergetics also, the nonlinear, coupled order-parameter
equations may themselves permit bifurcations. The order parameter, the am­
plitude o f a bifurcating macro variable, slaves the individual constituents as
some control parameter is changed. Equations can also show the competition
among order parameters, and that elements obey that order parameter that
wins the competition (Haken 1983).

Societal a n d M a c r o p s y c h o l o g i c a l Order Parameters

Closer to the emphasis o f this chapter, Haken (1983) offers this example o f an
order parameter. Languages are the order parameters that slave the subsystems
that consist o f individual human beings. The language changes very little over
280 Chaos Theory in the Social Sciences

a given person's lifetime. After his birth, a person learns a language ( i . e . , he


is slaved by i t ) , and over his lifetime he contributes to the survival o f that
language. Order parameters are considered to be unstable modes that slave
stable modes. This makes sense in the longer term because Homo sapiens
change less rapidly than do the myriad languages the species speaks and has
spoken. L i k e languages, I have proposed, the great religions, the scientific
and technological paradigms, and the Kondratiev structures are order parame­
ters. M i n d is also an order parameter, emergent out o f a field o f forces in the
brain as the brain interacts w i t h its internal and external environments. Be­
yond the scope o f this chapter is my contention that studying the brain as a
w i r i n g diagram and the mind as a set o f logical rules and procedures is
fundamentally incorrect; beware o f the claims for artificial intelligence and
expert systems (De Greene 1991a, 1991b)!
A societal field emerges out o f the myriad person-person and person-
machine interactions o f a sociotechnical system; the field reciprocally then
constrains the realm o f the possible at the microlevel. There may be a general
principle o f short-range randomness and activation at the microlevel and long-
range order and inhibition at the macrolevel. The randomness provides the
innovation necessary to explore the field or space, and the inhibition main­
tains a collective stability over a considerable amount o f space-time. A n order
parameter is a collective phenomenon. It is emergent beyond the obvious
other structures o f organizations and societies. I n my writings the order pa­
rameter is macropsychological, and it shows such characteristics as collective
m i n d , collective intelligence, collective perception, collective belief struc­
tures, and collective anxiety. Moods o f the time and overall social climates are
expressions o f macropsychological order parameters. Group and organiza­
tional cultures, climates, and cognitive styles are order parameters that may
encourage or impede problem solving, learning, competition, and adaptation.
Like other evolutionary features, order parameters rise and fall as the stability
dynamics o f the situation change. Experience shows that these cultures and
climates cannot be directly managed, although they may be damaged or
destroyed by management. We shall return to this topic at the end o f this
chapter.

The Kondratiev C y c l e / S t r u c t u r e in Recent


Societal E v o l u t i o n

The Kondratiev cycle has been classified as one o f the business cycles. Cycles
of about three-to-four or four-to-seven, and ten, twenty, and fifty-five years
mean duration have been identified. The first is the business cycle per se and
the last is the Kondratiev cycle or economic long wave. Cycles o f 150 or more
years duration have also been identified. A large literature has now been built
up in this area (e.g., Berry 1991; De Greene 1988a, 1988b, 1989, 1990d,
Interpretation of Complex Systems 281

1992, 1993a, 1993b, 1993c, 1993d, 1994; Kondratiev 1984; Schumpeter


1939; Van Duijn 1983; Vasko 1987; Vasko, Ayres, and Fontvieille 1990).
The term cycle, however, is insufficient. I view each cycle as an order
parameter that fits into even longer order parameters, such as the hegemonic
cycle, which describes the rise and fall o f nations ( v i z . , the great powers) and
the life histories o f the Newtonian paradigm, our phase o f w o r l d civilization,
and w o r l d civilization as a whole (for the last, see Tainter 1988). Hence, my
use o f the term Kondratiev cycle I structure (and hegemonic cycle/structure,
etc.). The Kondratiev cycle/structure can be viewed as a link between the
study o f science and the study o f history. It is an important organizing prin­
ciple for the study o f the evolution o f sociotechnical or techno-economic
macrosystems.

The Three-Tiered S t r u c t u r e o f a M o d e r n Economy

In all the sciences it is difficult to explain a system simply on the basis o f


function or behavior. A n understanding o f structure is necessary. A modern
economy can be viewed as three-tiered, w i t h the most basic and constraining
order-parameter field placed for convenience on the bottom. Here we are not
dealing w i t h individuals versus collectivities but rather with collectivities o f
different levels o f profundity. The top tier is the world o f mainstream eco­
nomics, o f prices, wages, money supply, inflation, simple supply and de­
mand, hiring and firing, and so on, and o f behaviors like the business cycle,
per se. This is the w o r l d o f the recessions and recoveries that presently draw
the attention o f policymakers, decision makers, and the media. The middle
tier deals w i t h the means o f production and consumption—factories, equip­
ment, tools, kinds o f skilled workers, and model changes w i t h i n the prevail­
ing paradigm. The intermediary Juglar and Kuznets cycles arise here. This is
no longer the w o r l d o f mostly the economist. Industrial engineers, behavioral
and social scientists, and ergonomists play important roles here. The third and
lowest tier involves the basic capital, energy, and technology structures o f
society. I t is the world o f the eventual exhaustion o f capital and o f past
investments and commitments, the wearing out o f one technology and its
replacement by another (Schumpeter's creative destruction), and the substitu­
tion o f one energy source (and now informational structure) by another. The
way o f thinking in society and the prevailing paradigm may become exhausted
(De Greene 1993a, 1993c). This is the w o r l d o f the Kondratiev cycle/
structure.

The Phases of t h e K o n d r a t i e v s

Four Kondratiev cycles/structures have been identified, w i t h the first begin­


ning with the Industrial Production Revolution about 1785. We are now, an
282 Chaos Theory in the Social Sciences

increasing number o f authors believe, in the depression phase o f Kondratiev


number four, not in the state o f recovery from a recent business-cycle reces­
sion. Kondratievs are conveniently divided into phases o f recovery, prosper­
ity, recession, and depression. The fourth Kondratiev started around 1940 at
the end o f the Great Depression o f the 1930s. The rising leg o f recovery and
prosperity lasted until about 1970, when the system entered the Kondratiev
phase o f recession. The Great Depression o f the 1990s probably began about
1989-90.
Assumptions obviously help structure policies and decisions as well as
the associated intermediary models. Most economists, policymakers, decision
makers, and the media assume the continued truth o f the present para­
d i g m . But i f the w o r l d economy is in reality in a great depression that w i l l
accompany the end o f meaningfulness o f this paradigm, then it is likely that
the continuation o f present policies and the formulation o f new policies based
on the o l d paradigm w i l l lead to counterintuitive results. Problems o f know-
ability and controllability arise. The Kondratiev phenomenon, with its proper­
ties o f nonequilibrium, nonlinearity, instability, and structural change, is far
from transparent and far from conducive to personal psychological comfort. A
business cycle is easy to understand and seemingly receptive to control ac­
tions. We shall return to this challenge to systems management later in this
chapter.

Features of a Kondratiev Phase of Depression

A Kondratiev depression phase has a number o f identifiable features. The


overall picture is one o f overcapacity, overextension, maturation o f industries,
saturation o f markets, misalignment o f subsystems, and wearing out. Indus­
tries, from hotels and office space to aerospace, automobiles, and o i l , now
show overcapacity. Services, from schools to health care, are overextended
and cannot meet the needs o f constituents. Stock markets are bloated with
speculator shares. Trade and internal deficits have mounted. The markets for
the products o f a wide range o f industries, from ships to commercial aircraft
to most computers, are saturated or nearing saturation. Products are cloned
and typically distinguished by packaging rather than by inherent worth. M y r ­
iad factories and most infrastructure are worn out. Maintenance is deferred.
Labor has been replaced by capital, yielding an increased capital intensity.
Unemployment has increased and remains recalcitrant to reduction. M a n y
kinds o f jobs have been eliminated, never to return. Relatively good jobs in
manufacturing have often been replaced by relatively poor, minimum-wage
jobs ("McJobs") in the services; and the proportion o f involuntary part-time
workers has increased. Wealth has generally flowed in the direction o f the
already well-to-do. Just as before the great depressions o f the 1890s and the
Interpretation of Complex Systems 283

1930s, the number o f millionaires has greatly increased. Productivity shows


mixed patterns, but white-collar and professional productivity, including that
in computerized operations, is largely stagnant. Organizational bloat, conse­
quent to increased administrative overhead, is held to account for much o f this
stagnation.
Whereas there are local, regional, and national differences among the
individual symptoms, the overall picture is worldwide. Indeed, differences
are a source o f further instabilities, as is attested to by continued deterioration
in Eastern Europe and the countries o f the old Soviet Union and by large-scale
migration from poorer regions to regions that are better off. See De Greene
(1990c) for a discussion o f factors common to capitalistic as well as to
communistic collapse.

Explanation of the Kondratiev Cycle/Structure

The Kondratiev cycle/structure can, in the cyclic sense, be explained in terms


o f the expansion and contraction o f the basic capital sector (see the three-
tiered economy discussed above). In the phase o f recovery, new investment
opportunities are perceived and investment decisions made, perhaps to a
considerable extent on the basis o f innovations that have "bunched" late in the
phase o f depression. New technologies and energy sources are developed.
Factories are built, new kinds o f jobs are created, and people are hired. The
overall economy, that is, the sociotechnical or techno-economic macro-
system, grows exponentially as these elements and subsystems reinforce one
another via positive-feedback loops. But eventually environmental and
carrying-capacity constraints (social, economic, natural, etc.) and saturation
effects operate as negative-feedback mechanisms to slow the rate o f growth
past some inflection point. Growth then slows to an asymptotic value, and
may become negative in a domain o f diminishing returns. This is a simple
verbal description o f logistic/diminishing-returns growth. A family o f these
curves can depict the four Kondratievs. See the discussion o f logistic evolu­
tion earlier in this chapter.
Considerable evidence has suggested the bunching or clustering o f inno­
vations in the phase o f depression (e.g., Schumpeter 1939). I interpret the
phase o f recession as a time o f very rigid societal and individual thinking,
when great efforts are expended to reconfirm the status quo. But in the phase
o f depression, it appears that society has given up and is ready to try just about
anything to remedy the misery. A technological innovation is defined as the
first commercial application o f a scientific discovery or invention. During the
phase o f depression, entrepreneurs take the latter and develop the former,
helping to stimulate recovery and the beginning o f a new cycle as noted
above. Just as important to true progress, however, are the social and institu-
284 Chaos Theory in the Social Sciences

tional innovations and societal leaders that often appear to lag behind the new
technological capabilities.
As mentioned earlier, scientific discoveries, inventions, innovations, and
great persons are mutations in societal evolution. The dynamics o f evolution,
as interpreted in terms o f dissipative-structure theory and order-parameter
theory, are held to apply here. Each Kondratiev structure is bounded by i n ­
stabilities. Each Kondratiev structure is an order parameter or macrovariable,
initially emergent out o f interactions at a more microlevel and sustained as a
field that offers both opportunities and constraints. The order-parameter field
defines the realm o f the possible. During much o f its fifty-five or so years
duration, the field functions as a mean field and average behavior dominates,
w i t h the damping o f fluctuations.
Finally, in the search for a basic causality, the order parameter is viewed
as primarily macropsychological, and collective perceptions, intelligence,
anxiety, and so on shape the investment decisions and other features discussed
above.

D e t e r m i n i s t i c Chaos a n d Other Nonlinear Behaviors

In deterministic chaos (not to be confused w i t h molecular chaos), the output


wave forms or the phase portraits o f a model look stochastic or unexpectedly
complex even though the causal mechanisms o f the iterated model are com­
pletely deterministic. Deterministic chaos is associated with very strong feed­
back and therefore w i t h strong nonlinearities. I n deterministic chaos the
system is extremely sensitive to initial conditions, and trajectories w i t h m i ­
nuscule initial differences diverge exponentially fast. Small random fluctua­
tions are amplified over several iterations and eventually take over the system.
Deterministic chaos can be observed in discrete-system simulations involving
only one difference equation (e.g., the logistic difference equation) and in
autonomous (not externally forced) continuous-system simulations w i t h three
or more coupled differential equations (e.g., the Lorenz weather model).
Deterministic chaos has been identified in real-world simple systems like
pendulums and lasers. However, its existence in real-world complex social
and societal systems is still questionable and a matter o f debate. To the extent
that models are used for prediction, chaos does place a definite limit on their
usefulness. The identification o f chaos may be useful in distinguishing be­
tween normal and pathological states. Also important from my perspective is
the extent to which deterministic chaos represents a kind o f instability o f fields
that plays a major role i n systems evolution. Chaos arises from self-
amplification and cross amplification, perhaps o f the hypercycle k i n d , o f
subsystems. These subsystems could, for example, be social and techno­
logical or demographic and natural environmental. Indeed, the related dy-
Interpretation of Complex Systems 285

namic behavior o f turbulence has been interpreted as arising at the boundaries


of m o v i n g subsystems.
Consideration o f the routes to chaos such as the Feigenbaum scenario,
use o f Poincare maps, and calculation o f Lyapunov exponents is beyond the
scope o f this chapter. These matters are dealt with in other chapters o f this
book. However, a word o f caution is in order. Critics point out, for example,
that calculation o f the Lyapunov exponents itself is insufficient, in the absence
o f some explanatory theory, for the identification o f chaos. Beware the i m ­
pulses to calculate prematurely and to overcalculate.
In system-dynamic computer simulation models, instability and deter­
ministic chaos are often associated with negative-feedback loops possessed o f
relatively high loop gain together with appreciable delays, which permit the
amplification o f the small, random fluctuations. Alternatively, in simple
models o f the Kondratiev cycle, instability and chaos arise out o f positive-
feedback loops (Mosekilde, A r a c i l , and A l l e n 1988). A n example is the self-
amplification o f the capital sector mentioned above. See also Forrester 1977.
Berry (1991) reports deterministic chaos in Kondratiev series o f annual
growth rates in wholesale prices. He generated a phase space by plotting each
relevant year, t, against the previous year, t — 1. Rosser (1991) and Zeeman
(1977) also discuss nonlinearities and discontinuities in economic behavior,
particularly o f the catastrophe theory k i n d . Gilmore (1981) provides a mathe­
matical treatment o f catastrophe theory, an important topic that space limita­
tions place beyond the scope o f this chapter.
Later in this chapter, we shall discuss the evolution o f information struc­
tures (macropsychological order parameters) in relationship to both the
Kondratiev cycle/structure and the several kinds o f attractors, including cha­
otic attractors, and their behavior in phase space.

On Paradigm Change

A paradigm can be defined as a set o f interrelated Weltanschauungen: theo­


ries, models, practices, findings, explanations, values, beliefs, and feelings
that characterize a given culture at a given time. This definition is much
broader than most. A simpler definition might equate a paradigm to a theory
or a model. Some authors (e.g., Kuhn 1970) describe different paradigms
w i t h i n a scientific field like physics and present seminal interpretations o f the
nature o f scientific revolutions. Because science and society or culture interact
over long periods o f time, and because scientists have all too human limita­
tions, it is necessary to provide the broad definition given above.
The main paradigms dealt with in this chapter are the Newtonian para­
d i g m , which has come under increasing scrutiny and criticism (e.g., Checkland
1981; De Greene 1993b and the many references therein; and Prigogine and
286 Chaos Theory in the Social Sciences

Stengers 1984), and the "new" paradigm (which might be called the paradigm
o f evolution and reconfiguration).
The Newtonian paradigm may be viewed as part o f the macrostructure o f
information and knowledge that emphasizes and usually tolerates only the
following: rationalism; reductionism; parts isolated from wholes; detached
objectivity o f observation and measurement and separation o f the observer
from the observed system; simple causality; logical, steplike but iterative
analysis; deduction o f rules, procedures, and algorithms; m a x i m u m use o f
numbers; emphasis on average behavior; equilibrium; fixed, inviolable laws;
reversibility; denial o f variety and ambiguity; denial o f subjectivity; and con­
vergent focus on the correct answer or solution. The Newtonian paradigm is
best fitted to a static or slowly changing w o r l d o f stability and structural
continuity, not one o f evolution, instability, and structural change.
In contrast, the new paradigm encompasses nonrationality, nonlinearity,
mutual causality, nonequilibrium, irreversibility, stochasticity/determinism,
uncertainty, opportunity and choice seen in fluctuations and apparent noise.
Moreover, the dominance o f exceptions near critical thresholds, the genera­
tion and maintenance o f variety, structural change, divergent thinking, and the
recognition that there can never be eternal truth and reality but only different
perceptions o f such, also represent crucial elements o f the new paradigm.
F o l l o w i n g the broad definition above, paradigms structure the very way
that science is conducted. Karl Popper emphasizes testability over truth
because truth may never be reached. Further, scientific theories can never be
proven or validated through experimental tests; they can only be disproven or
falsified. Science thus cannot be reduced to the formal, logical method o f the
positivists. Popper considers conceptual structures like Marxism and psycho­
analysis to be outside the realm o f science because they cannot be tested and
falsified. I t appears to me that this orientation rules out o f scientific inquiry all
the really interesting systems—the large-scale, complex, unstable, evolving,
self-organizing, structurally changing systems—of concern in this chapter.
Because o f instability, chance, fluctuation, bifurcation, and chaos, each such
complex system is unique. Both organic and societal evolution are summa­
tions o f chance occurrences over time. Even i f testing o f large-scale l i v i n g
systems were possible, the social objections and social and environmental
dangers and costs w o u l d most likely preclude such activity (see, for example,
many o f the criticisms o f the Strategic Defense Initiative). In some cases
small, experimental studies (e.g., the comparison o f small ecosystems w i t h
and without a given species) can be suggestive, but even here the generaliza­
tion o f the results requires caution.
A great deal o f research in the behavioral and social sciences, both in the
laboratory and in the field, has followed the Newtonian paradigm. One o f the
Interpretation of Complex Systems 287

major purposes o f this book can be the rethinking o f the adequacy o f this kind
of scientific method. A t the present time many scientists are taking a new look
at the very purpose, meaning, and practice o f science. This discussion is
pursued further elsewhere (De Greene 1993b).
A t the level o f societal policy making, decision making, and design,
different paradigms or subparadigms are needed, depending on the apparent
linearity, equilibrium state, stability, and predictability o f the system (De
Greene 1987, 1990d, 1991b). These features can also be related to the phases
o f the Kondratiev cycle/structure.

The Challenge t o t h e M a n a g e m e n t
of C o m p l e x S y s t e m s

A sociotechnical or techno-economic macrosystem is a dissipative structure in


the sense that high-quality inputs (energy and matter) are converted to low-
quality outputs like heat and waste, w i t h an increase in disorder and entropy.
W i t h i n this overall process, o f course, low-quality raw materials are con­
verted into high-quality finished products, but these themselves eventually
break d o w n , yielding further entropy. This is one way to look at the product
life cycle.

The G r o w t h o f Exhaustion o f I n f o r m a t i o n / K n o w l e d g e
at t h e End of t h e Kondratiev C y c l e / S t r u c t u r e

To the extent that information is the negative o f entropy (see the Shannon and
Boltzmann formulas), we can also speak o f an information life cycle and a
growth o f exhaustion o f information as the end o f the cycle, identified by this
author (De Greene 1993a) as the recession-early depression phase o f the
Kondratiev cycle. A t the end o f a Kondratiev, entropy is high and information
level is low. The progressive uniformization o f the system is associated w i t h
the loss o f information. In the stages o f recovery and prosperity, information
is increasing as the new cognitive space or field is explored, and disorder and
entropy are decreasing as new designs are implemented. In the phase o f
recession, when the field has been explored to its boundary, energy is dissi­
pated and information decreases as confirmation o f the status quo, w i t h the aid
o f proliferating bureaucracies, becomes the established modus operandi. But
eventually, reinforcement no longer works, real underlying instability arises,
and the system starts oscillating, with the emergence o f fads and hysterias as
society attempts to preserve and regain order. We note here that entropy has a
material meaning as well as informational and strictly energy-based mean­
ings. The untoward consequences o f the degradation o f matter and energy by
288 Chaos Theory in the Social Sciences

the economic process have been beautifully explicated by Georgescu-Roegen


(1971).
Nicolis and Prigogine (1989) conclude that chaotic motions necessarily
imply that phase-space volume is expanding in certain directions and contract­
ing i n others. Variety and choice are continuously generated along the expand­
ing directions o f the motion, and predictability increases along the contracting
directions. Thus, chaotic attractors, which possess asymptotic stability, can be
generators and processors o f information. Lyapunov exponents can be used to
describe the mean rates o f expansion or contraction. It is tempting, therefore,
to divide the Kondratiev cycle/structure into four epochs that are offset some­
what from the four phases, w i t h chaotic attractors associated w i t h innovation
in the late depression and recovery phases, limit-cycle attractors and oscilla­
tory behavior associated w i t h both the early prosperity and late recession
phases, and point attractors and equilibrium-seeking behaviors characterizing
the late prosperity and early recession phases around the inflection point o f
the logistic curve. M o d i s and Debecker (1992) have identified chaotic wave­
forms at both the beginning and the end o f the logistic function.
The overall Kondratiev system w o u l d evolve in the following manner:
from chaotic attractor to limit-cycle attractor to point attractor to limit-cycle
attractor to chaotic attractor, w i t h successive growth o f exhaustion o f informa­
tion and innovation (De Greene 1993a, 1993c), and so on. Evolution w o u l d
show movement from nonequilibrium to equilibrium to nonequilibrium, and
so o n . Because o f the irreversibility o f structural change, o f course, the
specific structures w o u l d not be the same. Moreover, as noted earlier, the
Kondratiev system overlaps or operates within still larger cycle/structures. In
addition, the Kondratievs are bounded by a hyperbolic (or exponential) enve­
lope curve depicting overall sociotechnical evolution (De Greene 1982). Fur­
ther, similar curves depict changes in natural environmental variables, so that
the overall picture is likely one o f humanity moving farther and farther from
equilibrium and closer and closer to the absolute limits o f sustainability and
stability, as they can realistically be perceived today. A n y thinking about a
fifth Kondratiev has to consider the major structural changes in the larger
sociotechnical and natural environments o f the present Kondratiev (De Greene
1993d). Potential exploratory space, in both the physical and abstract senses,
may be much less than humanity might hope.
Let us look more closely at the growth o f exhaustion o f human cognitive
models and artifacts, along w i t h effective information and knowledge, that
reflects the increasing loss o f divergent thinking and the increasing dominance
o f convergent thinking at the macropsychological level o f a nation or an
interlinked group o f dominant nations (De Greene 1993a, 1993c). Instead o f
nation(s), the words the economy, the sociotechnical macrosystem, and the
Interpretation of Complex Systems 289

world system-field, and so on, could o f course be substituted. Information and


knowledge (information integrated and fitted into a context) are more than just
flows in an information-processing system; it is better to think in terms o f
evolving (and devolving) structures. I n the evolution o f a Kondratiev, users
over time tend to select the more stereotyped forms o f information and k n o w l ­
edge. People seek the expected, the riskless, the tried and true. There is a
great deal o f cloning o f designs, processes, products, services, and organiza­
tional practices, and a great deal o f bandwagon activity. Older organizations
may lose their original purposes or, perhaps worse, may continue rigid, stereo­
typed behaviors derived from original purposes when dynamic, evolution­
ary changes in the external environments have long since passed them by.
Information and knowledge embodied in the original designs gradually dissi­
pate as the designs wear out. Nevertheless, in modern societies equilibrium
structures like machines and technology i n general can perpetuate the exis­
tence o f an o l d system long after it w o u l d have expired naturally. A n d the
collective mind or consciousness possesses tremendous inertia, introducing
time lags into attempts at innovative systems management. Such powerful and
well-known organizations as the United Nations, the World Bank, the Inter­
national Monetary Fund, N A T O , and N A S A may fall into this category.
Experience shows that it is quite difficult to j u m p start a renaissance in
ingrained, postmature organizations. In addition, there is the vast morass o f
collapse that is the relic o f the old Communist w o r l d . This suggests a w o r l d o f
worn-out ideas, designs, and practices; a w o r l d awaiting transformation (De
Greene 1993b).

Following Shannon (1948), total certainty is equivalent to no informa­


tion. New information can be interpreted as reconfirming and strengthening or
stabilizing existing structures. Following Thorn (1975), reduction o f informa­
tion to its scalar measure (given in bits) means loss o f almost all the meaning
that might be imparted to a message. Also, no information exchange is pos­
sible between static forms. A message that conveys what is already known
provides no change in knowledge. A known answer provides no uncertainty
and no information. Further, information is lost when a system has diffused
into its environment, and it is no longer possible to distinguish the system
from its environment. For example, an innovation may be viewed in a figure/
ground relationship; but after the innovation has diffused throughout the soci­
ety, figure merges into ground.
Information and knowledge are more than flows in information-
processing systems. Information and knowledge are themselves structures
that show collective or coherent behaviors at a more macrolevel emergent out
o f interactions at a more microlevel. The resulting cognitive field then helps
regulate behavior at the more microlevel. The field can spontaneously and
290 Chaos Theory in the Social Sciences

irreversibly reconfigure into different populations or complexions that display


different information and energy qualities. See again our discussion o f order
parameter.

S y s t e m s M a n a g e m e n t at a T i m e
of T r a n s f o r m a t i o n a l Change

Organizations and societies thus seek continuously to reconfirm themselves,


which leads to a selective use o f information and a dangerous narrowing o f
perception (De Greene 1991a). The evolution o f new knowledge systems can
be stifled in the search for continued equilibrium. The economy, for example,
aims toward an equilibrium state dominated by confirmation. Predictability is
thought by increasingly conservative mainstream policymakers and decision
makers to be enhanced by attempts to restructure and therefore to control the
overall environment and to maintain a static security. Control or apparent
control reduces anxiety. Yet when the information/knowledge structure
evolves into a stage o f saturation or even diminishing returns following diffu­
sion into the surroundings, most people do not perceive the incipient struc­
tural changes. The present epoch o f diffusion saturation o f innovations re­
sponsible for previous economic growth began about 1970, around the
inflection point o f Kondratiev number four. Policymakers, planners, and deci­
sion makers, and their analyst advisors, continue to apply the same tired, old
logic and management methods to the increasingly evident symptoms o f the
reconfiguring field.
As discussed above, an attractor may over time degrade from chaotic
to l i m i t cycle to fixed equilibrium point. This process can yield a cer­
tainty, confirmation, and predictability that no longer possess information/
knowledge. O l d , decrepit systems may be more easily captured by an equi­
l i b r i u m attractor. Contrariwise, the emergence o f a new self-organizing
collective m i n d , which can provide a driving force for the next (now the fifth)
Kondratiev, can follow the smallest o f stimuli or fluctuations. Nucleations can
then spread explosively across the field. Newly formed collectivities can stand
out in a figure/ground sense from established structures and can thereby
encourage a growth in knowledge. The insight learning interpreted by the
Gestalt psychologists as a sudden reconfiguration o f the cognitive field may
represent capture by a chaotic attractor and its associated basin o f attraction.
Referring back to the popularly understood meaning o f information,
massive investment in this technology has yielded no increase in the produc­
tivity o f information workers. From 1980 through 1989, office equipment
increased from 3 percent to 18 percent o f U . S . capital equipment. But the
average productivity o f the workers was about the same as in the 1960s in
spite o f the huge increase i n the amount o f computer power. Lack o f improve-
Interpretation of Complex Systems 291

ment appears to be due to a considerable extent to increases in organizational


bloat. See the discussion and references in De Greene 1990b.
Further, computerized information systems can constrain and channel
thinking, so that innovation is lost. The latest technologies can rigidify manage­
ment functions like communications, coordination, and control (De Greene
1991a, 1991b). Such reinforcements can be part o f a larger evolution toward
greater standardization and reduced variety during the Kondratiev phases o f
recession and depression. The Office o f Technology Assessment (cited i n
De Greene 1990b) comments that organization and management styles appear
to be becoming more and more identical across such sectors and institutions
as banking, insurance, large farms, supermarkets, construction, textiles, hospi­
tals, and automobiles. Kinds o f jobs and skills become increasingly identical,
and responses to forces in the external environment like financial markets and
w o r l d trade are increasingly standardized and stereotyped, as rationalized or
optimized best fits diffuse and are implemented. Moreover, over the next
decade or t w o , most applications o f new technology w i l l involve doing some­
thing familiar in a better way. These trends imply reconfirmation o f the status
quo and growth without innovation.
I n evolution, however, processes, mechanisms, elements, and inter­
actions are added or subtracted, and evolution selects for variety. Optimiza­
tion and improved efficiency make the organization more vulnerable as orga­
nization, environment, and their interrelationships change over time. The
efficiency o f a single perspective or rationality may improve the organization
only for a specific purpose and w i t h i n a limited time. These matters are
persuasively argued further by A l l e n (1993).
Each Kondratiev is characterized by a primary energy source, primary
technologies, and primary institutional forms. However, any o f these features
can spill over to the next Kondratiev. The railroads and the telegraph were
primary technologies o f Kondratiev number t w o , and improved means o f
organizational communication arose out o f the need to coordinate the far-flung
operations o f the railroads. Internal communications, coordination, and con­
trol also changed greatly, and these changes spread to manufacturing around
1880. Yates (1989) discusses the rise o f the American system o f management
(my term, see her subtitle) over the years 1850-1920 based on the philosophy
o f system and efficiency. She states that managerial control is essentially
management as we now know i t . The important emphasis here is that the
transformation o f American firms was essentially complete by the end o f
World War I , the end o f the phase o f prosperity o f the third Kondratiev.
Subsequent advances in computer-communications technologies have rein­
forced the dominant managerial philosophy and organizational design and
have enhanced the efficiency (but not necessarily the effectiveness) o f organi­
zational processes.
292 Chaos Theory in the Social Sciences

The American system has been a dominant institutional form over part o f
the second and essentially all o f the third and fourth Kondratievs. It has, o f
course, spread around the globe. The myriad criticisms o f Tayloristic/Fordian
work-system designs and their deleterious impacts on worker motivation,
morale, j o b satisfaction, and productivity (e.g., De Greene 1982) can be
noted here but are beyond the scope o f this chapter. The Japanese system
appears presently to be at least a partial alternative to the American system,
but this topic, which has also received a tremendous amount o f attention, is
also beyond the scope o f this theoretically based chapter. I n keeping w i t h
other interpretations made herein, it is my belief that the American system o f
management is part o f the exhausted machine model o f humanity, part o f the
overall Newtonian paradigm. New theories, methods, observations, and find­
ings should not be distorted and made to fit the old paradigm and outmoded
designs.
Finally, a dynamic field o f forces cannot be controlled or managed; but it
can be perturbed, leading to surprising nonlinear, unstable, and reconfigura-
tional results. These results could come about because critical thresholds were
passed, the delicate balance among feedback loops disrupted, and the bound­
aries o f stability passed. On the other hand, reconfiguration is natural in
evolution and can provide new choice and new opportunity. The distinction is
management w i t h i n the field, not o f the field (De Greene 1990b). The chal­
lenge for Americans, and for humanity in general, is to try to induce climates
of creativity, small at first, that might generate desirable innovative fluctua­
tions that could trigger the emergence o f new order parameters better fitted to
evolutionary reality. Note the absence o f deterministic language and the ab­
sence o f any implication o f ultimate human understanding and control.

Concluding Remarks and R e c o m m e n d a t i o n s

How valuable are chaos and nonlinear systems theory to the social sciences?
The answer depends on the nature o f the driving force. I f these theories act as
perturbations and fluctuations, driving a restructure o f social science, and i f
they help generate new paradigm thinking, then the future can indeed be
promising. I f , however, the new theories function just as new tools (like a new
form o f regression analysis), then the social sciences may find that the excit­
ing and challenging aspects o f social reality have been usurped by the more
dynamic, the more imaginative, and the more adventuresome, and that tradi­
tional economics, sociology, political science, and so on have become i n ­
creasingly irrelevant.
The same k i n d o f situation pertains for the high-level systems practices
of management (in this chapter policymaking and decision making in particu­
lar) and design. A time o f major transformation o f the existing force field
Interpretation of Complex Systems 293

always provides greater opportunity as well as greater challenge. A time o f


major transformation in societal as well as organic evolution leads to emer­
gence o f new forms as well as extinction or submersion o f older or less
adapted forms. The next few years may provide a window o f opportunity that
may not return for decades, i f ever. This is because each force field is spatially
and temporally unique, and because the magnitude and acceleration o f the
intersecting environmental and societal forces is probably now the greatest in
the experience o f humanity. I f , say, policymakers can see the opportunity o f
chance is today's instability and transformation, then the chaos o f messes and
problems may be, as a whole, amenable to a new and enlightened manage­
ment. I f , on the other hand, energy and resources are directed to reconfirming
the status quo, then it is likely that these efforts w i l l compound present
difficulties via acts o f omission and commission linked to bad timing and bad
placement.
The theory o f dynamic, reconfiguring fields having only temporary sta­
bility is antithetical to concepts o f absolute and everlasting truth and perma­
nency o f methods. The policy community must understand the cognitive/
emotional ties to such recently and presently powerful belief systems as
humanity's place in nature, organized religion, communism, capitalism, de­
mocracy, the market economy, free trade, and so forth. It is, o f course, easy to
argue for understanding. But (selective) understanding and not understanding
can have apparent survival value to individual and society. Most o f us feel
comfortable w i t h the familiar, the k n o w n , the tried and true. Most o f us build
up elaborate systems o f rationalizations and systems o f denial to defend
ourselves against perceived vulnerability and psychic insult. For example,
most mainstream economists deny the limits to growth, deny that free-market
practices can lead to environmental destruction and societal impoverishment,
and indeed deny the existence o f uncontrollable situations. Because o f time
lags associated w i t h growing up, education, and experience, it w i l l probably
be the rare individual who understands and grasps the opportunities now
emerging through chaos in the broadest sense o f the term.
History shows that great leaders arise at times o f crisis. From a systems
theoretic standpoint, the damping and stabilizing effect o f average behavior
(the law o f large numbers) weakens as control parameters (e.g., human popu­
lation g r o w t h , atmospheric change, biodiversity loss, and inequitable distri­
bution o f wealth) approach bifurcation points. This time around, because o f
the intensity o f planetary crisis, the leaders may emerge too late, may not fit
the field, or may be o f the kinds that present thinking would find objection­
able. Therefore, using perhaps overworked terms, we must be anticipatory
and proactive. T h i n k i n g must be radical and revolutionary. But again this is
easier said than done. There is no simpleminded, linear advice that can be
offered like that o f the evolved-away-from, left-behind, equilibrium-centered
294 Chaos Theory in the Social Sciences

w o r l d o f the 1950s and 1960s: use risk analysis, perform a decision analysis,
optimize, reduce the scope o f concern to the immediately manageable. There
is no meaningful list o f steps that can be taken. There are no real-world-
sensitive panaceas that can be offered. A field o f forces is imprecise. As
system dynamics, dissipative-structure theory, and catastrophe theory all
show, and as experience indicates, a field o f forces can surprisingly swallow
policy and decisional perturbations, or it can be destabilized by such perturba­
tions. What happens is fortuitous and depends on the t i m i n g and placing. The
task confronting us now is a highly individualistic and lonely one, but poten­
tially a most creative one. Leaders and changers w i l l be those persons who
can overlap their cognitive/emotional models, themselves internal fields o f
forces, w i t h the w o r l d system-field or a constituent part o f that field. In the
latter case, an energized local field could excite, or even inflame the entire
field.
Yet in spite o f the formidable nature o f evolution and reconfigurational
change in the w o r l d system-field and in spite o f the great difficulty in mapping
external reality onto the m i n d , certain general advice can be offered to those
w i l l i n g and able to accept it: recognize human limitations, do not compart­
mentalize your thinking, avoid linear thinking about cause and effect, respect
and learn from history, expect the system to behave counterintuitively, expect
surprises and anticipate the opportunities that surprises can provide, expect
policy and decision actions to assume new meanings when imbedded in a field
context, think instability, instability, instability. As emphasized here, rare
persons w i l l intuitively understand these matters, but this advice should help
others.
Finally, we can note that there are implications for education. Once again
the questions o f appropriateness in a field and o f direction o f causality arise.
A t times o f instability and chaos, rigidities such as programs, plans, curricula,
learning objectives, fixed assignments, tests, and administrative control are
likely to be indifferent at best and harmful at worst. Educational needs,
expressed as fluid learning experiences, w i l l evolve and emerge from the
pioneering behavior o f those who can sense and move w i t h the dynamics o f
the field. I f and when things quiet d o w n , some further formalization o f
methods and designs w i l l probably be desirable. But formalization should
never again be allowed to reach the scope o f the Newtonian paradigm o f
science, culture, and society.
CHAPTER 13

S o c i a l S c i e n c e as t h e S t u d y o f C o m p l e x S y s t e m s

David L . Harvey and Michael Reed

Despite three decades o f development in the physical sciences, the social


sciences are only now coming to grips with deterministic chaos and its world-
view. Several explanations have been given for this hesitance. Positivists,
appealing to Comtean doctrine, conceptualize the sciences as constituting a
developmental hierarchy. Mathematics, physics, and astronomy, envisioned
as the most mature o f the sciences, are at the top o f this hierarchy, while the
immature social sciences are situated on the bottom rungs. Given their relative
youth, it is only natural, from the positivists' perspective, for the social
sciences to master scientific advances at a slower rate than their elders. After
all, they do not yet possess the mathematical or technical sophistication o f the
older sciences and are not prepared, therefore, to take advantage o f their most
recent innovations.
A second and more practical account o f this scientific lag appeals to the
formidable mathematics o f deterministic chaos theory. It claims that only a
handful o f applied mathematicians working within the social sciences are
capable o f fully grasping the fundaments o f the new science. Since these
scholars stand on the margins o f the human sciences, it is no wonder that the
chaos paradigm has made little headway in these disciplines.
A third and final interpretation o f the slow reception o f chaos theory into
the social sciences underscores the fact that in the last thirty years science
itself has become déclassé in many circles. The cultural revolution o f the
sixties identified science with the repressive and dehumanizing tendencies o f
modernism. Influenced by this radical humanism, many social scientists re­
jected quantification w i t h a vengeance and have opted for a hermeneutic
method in their research. In the process, they have steadily retreated into the
deconstruction o f texts, or into the study o f postmodern and postindustrial
mentalités.
As is often the case, none o f these accounts can be fully refuted. Positiv­
ists defend their position by citing a long history o f the social sciences
borrowing from the physical sciences. They have trouble explaining, how-

295
296 Chaos Theory in the Social Sciences

ever, w h y chaos scholars reject the classical paradigm o f Newtonian science,


and actually seek the company o f humanists when exploring the full implica­
tions o f the new science o f chaos. Similarly, those who cite the intimidation o f
mathematics ignore the significant headway chaos theory has already made in
econometrics and in urban ecology. A n d , finally, those underscoring the re­
cent antimodernist drift in the social sciences fail to realize that chaos theory
has its cultural roots in the revolutionary sixties (Gleick 1987; Loye and Eisler
1987) and that the theory—at least in its more popular renderings—has even
been used to justify the subjectivism and antirationalism o f the postmodernists
(Young 1991).
W h i l e all three positions give credible accounts o f chaos theory's slow
penetration into the social sciences, all ignore a fundamental ideological
problem that often emerges when a new methodenstreit erupts. Most parties to
such debates i m p l i c i t l y assume that there is only one social science and only
one scientific method. This paper challenges this monological assumption and
argues that no single method can fully appropriate the manifold complexity o f
social life. I t suggests, instead, that social scientists have a plurality o f
methods from which to choose when they research a subject, and it is their
responsibility to select the one method that best fits the ontological contours o f
the problem they are studying. Hence, asking why the social sciences are
slower than the physical sciences in adopting the deterministic chaos para­
d i g m , or w h y sociology, anthropology, and the cultural sciences are lagging
behind econometrics and social ecology misses the mark. What should be
asked, first, is whether the social phenomenon a scientist chooses to investi­
gate imposes ontological limits on the methods he or she can employ.
In this essay we w i l l take up this latter question. We w i l l explore the
circumstances under which research strategies employing the deterministic
chaos paradigm can and cannot be deployed in the human sciences. I n particu­
lar, we w i l l look at how the ontological structure o f institutional life restricts
when and where a given method can be applied. In exploring these issues, we
w i l l steer a course midway between those positivists who would use chaos
theory to revivify an exhausted scientism and those postmodernists who reject
quantification "on principle." Our middle course is grounded in the philosoph­
ical work o f the British philosopher Roy Bhaskar (1978, 1989a, 1989b).
Bhaskar's philosophy advocates a realist ontology and a modified scientific
naturalism in the area o f epistemology. As we w i l l show, Bhaskar's "critical
naturalism" is especially sensitive to the differences between the w o r l d o f
nature and the social realm. It is this sensitivity to the structural differences
that exist in the w o r l d that allows us to wed Bhaskar's realist ontology and
epistemology to a dissipative systems approach. Such a wedding, we believe,
can give rise to a new social scientific w o r l d v i e w — o n e grounded in the
axioms o f deterministic chaos, and which is cognizant o f the profound differ-
Social Science as the Study of Complex Systems 297

ences separating the social world from that o f nature. Such a dynamic realism
is capable o f sustaining the particularity and plurality o f the social w o r l d while
preserving rational canons o f scientific understanding.
A n y hope we have o f realizing this new social scientific worldview is
tempered, however, by our realization that the ontological structure o f social
phenomena places objective limits upon the types o f methods that can be
legitimately deployed when studying the social realm. Hence, the naturalism
we advocate demands social scientists employ a methodological pluralism
when studying institutional life from a dissipative systems perspective. I n line
w i t h these reservations, we make two assumptions. First, w i t h the notable
exceptions o f its judicious application in such areas as demography, social
ecology, and urban place theory, the simple incorporation o f the deterministic
chaos paradigm into social scientific research w i l l prove nugatory. Second, i f
the actual mathematical models o f deterministic chaos and the concrete find­
ings o f the physical sciences have limited value in their direct application to
the social sciences, they can still provide a rich heuristic base from w h i c h
social scientists can w o r k .
Our vehicle for exploring these theses is a matrix that matches modeling
strategies used in the social sciences with the different ontological levels that
characterize social systems. The matrix is constructed in three steps. We begin
by identifying six modeling strategies currently used in the social sciences:
(1) predictive models patterned after the methods o f classical mechanics;
(2) statistical models commonly employed in the social sciences; (3) icon-
ological models recently developed by mathematical chaos theoreticians;
(4) structural models popular i n contemporary anthropology and sociology;
(5) ideal type models utilized in comparative economics and sociology; and,
finally, (6) historiographical models that concentrate on the ideographic
description o f temporally ordered concrete events. The second step in con­
structing the matrix focuses upon the development o f a framework that reflects
the ontological stratification o f the social world. Borrowing from Kenneth
Boulding's (1968, 3 - 1 0 ) work on hierarchies o f complexity and Neil J.
Smelser's (1963, 1-46) reading o f Parsonsian systems theory (Parsons 1961),
we develop a model o f nested social system hierarchies that emphasizes their
ontologically layered nature. This conception o f social systems sees them as
consisting o f emergent layers that are ontologically distinct and irreducible.
The third and final step in this process is the construction o f the matrix itself.
After completing the matrix, we illustrate its use by assessing the relative
efficacy o f each o f the six strategies for studying chaos.
We make no exaggerated claims for this matrix or for the resulting
typology. Ours is a first effort, and like all initial formulations it invites
comment and revision. Rather than presenting a finished product, we offer the
matrix as a tentative first statement to be tested and revised in practice.
298 Chaos Theory in the Social Sciences

Moreover, our perspective is grounded in the assumption that a simultaneous


convergence in several scientific fields is occurring around the new chaos
paradigm (Reed and Harvey 1992).
Before presenting the matrix proper, however, we must briefly describe
the philosophical foundation o f our project, beginning w i t h a discussion o f
Bhaskar's critical realist perspective. We then examine the dissipative systems
perspective (Prigogine 1980; Prigogine and A l l e n 1982; Prigogine and Sten¬
gers 1984) that is at the heart o f our modeling/ontology matrix. Finally, we
take up the problem o f social ontology, and in doing so move to the presenta­
tion o f the matrix itself.

Philosophical A s s u m p t i o n s

Bhaskar's critical realism begins with an immanent critique o f scientific


method and scientific findings. That is, in order to philosophically explore the
ontological foundations o f the w o r l d , he assumes that the methods and find­
ings o f science are essentially correct. Working from this presupposition he
then asks what the ontological structure o f the real w o r l d must be in order for
science to sustain its truth claims. I n this way he moves quite logically
from questions o f scientific method, per se—how scientific knowledge is
produced—to philosophical inquiry into the structures that underlie and orga­
nize the visible w o r l d .
What are the results o f Bhaskar's immanent critique? First, science and
its findings, despite the claims o f some o f its apologists, is a human product
and bears the same indelible markings as all else that is produced by human
hands. L i k e other human endeavors, science is a historically open and contin­
ually evolving enterprise. It has an ideological dimension, and its practices
and findings reflect the cultural commitments o f the society that houses and
nourishes i t . To this extent, science is a historically conditioned, transitive
activity. But i f that were all there was to science, it would be indistinguishable
from other ideological systems. Critical realism assumes that there is more to
science than this, and it distances science from simple ideologies by insisting
upon the obdurate nature o f the real w o r l d . For critical realism, science is
anchored in an intransitive domain, in a w o r l d whose autonomous constitution
stands independent o f the knowing subject and his or her parochial interests. It
is the intransitive nature o f science's object that prevents its full incorporation
into the partial ideologies that constitute a society's belief system. Because o f
this anchoring in the external w o r l d , and the constraints that follow from that
anchoring, scientific findings cannot be dismissed out o f hand as mere self-
interested assertions.
The affirmation o f science's intransitive object serves as a grounding for
the second result o f Bhaskar's immanent critique. He then begins a dissection
Social Science as the Study of Complex Systems 299

of the interior o f that object by examining the commonplace observation that


knowledge, especially in the physical sciences, possesses a cumulative, often
progressive, nature. That is, the findings o f the physical sciences seem to
build upon one another and to form a more or less irreversible compilation o f
factual knowledge about the structure and functioning o f the w o r l d . Working
from these mundane observations, Bhaskar deduces a second set o f ontologi-
cal propositions about the w o r l d . Science's cumulative laying up o f k n o w l ­
edge from one generation to the next proceeds by progressively stripping
away layer after nested layer o f reality. W i t h each new advance, science
probes deeper into nature and uncovers ever more fundamental levels o f
physical reality. I f scientific knowledge has this cumulative structure, then
reality itself, Bhaskar argues, must be organized in a similar fashion.
Moreover, this ontological layering forms a hierarchy o f open, nested
structures so that, as science penetrates each new layer, it discovers the
ontological base o f those entities and structures it has previously uncovered.
For example, as science has penetrated deeper into the structure o f matter, the
perspectives of, first, classical physics, and then molecular chemistry, gave
way to quantum mechanics and the study o f the subatomic world. Upon
deeper probing, quantum mechanics yielded to particle physics and the subse­
quent explosion o f a myriad o f even more fundamental structures o f matter. In
the last t w o decades, the study o f these fundamental particles exposed a new
stratum o f the w o r l d — t h e curiously constrained universe o f colored and
flavored quarks. A t each step new levels o f material reality emerged, only to be
displaced by even more fundamental entities and processes. Hence, the his­
tory o f physics and chemistry in this century and the last attests to the hier­
archical and nested structure o f physical reality. The world is composed o f
layers o f matter emergently related and nested one within another. A n d this
ontological deepening o f reality parallels the evolution o f modern science.
The metaphor o f peeling away successive layers o f an onion is appropriate for
describing this process. Moreover, whether we are talking about the w o r l d o f
matter or the w o r l d o f social systems, this ontological nesting o f irreducible
layers seems to apply w i t h equal force.

This hierarchical ordering o f disjunctive layers is held together by


loosely bounded processes o f determination and overdetermination, thus
forming a multilevel system. I f this system did not exist, even in a loosely
formed state, there could be no coherent body o f scientific knowledge, and
certainly no common ground for communication between the various sci­
ences. This point has been made by P. A . Anderson (1972). He argues the
existing scientific division o f labor points to an ontologically layered uni­
verse, not unlike the one Bhaskar assumes. Running the gamut from elemen­
tary particles to social structures, Anderson suggests that each layer o f reality
is structurally unique, operates according to its o w n laws, and demands its
300 Chaos Theory in the Social Sciences

o w n investigative protocols. I n explaining the origin o f these various levels,


Anderson underscores the role played by broken symmetries. Symmetry-
breaking mechanisms provide the ontological foundations for the emergence
o f new levels o f reality from established levels. Symmetry breaking allows
each new ontological level to organize itself around its o w n set o f irreducible
principles. Since different principles regulate the activities o f each level, no
level is reducible to those from which they have emerged. The resultant
pyramiding complexities produced by the geometry o f broken symmetries is
the source, therefore, o f our ever-deepening modern division o f scientific labor.
Thus, Anderson explains and then schematizes this irreducible relation be­
tween the various sciences and their ontologically unique objects as follows:

The behavior o f large and complex aggregates o f elementary particles, it


turns out, is not to be understood in terms o f a simple extrapolation o f the
properties o f a few particles. Instead, at each level o f complexity entirely
new properties appear, and the understanding o f the new behaviors re­
quires research which I think is as fundamental in its nature as any other.
That is, it seems to me that one may array the sciences roughly linearly in
a hierarchy according to the idea: The elementary entities o f science X
obey the laws o f science Y.

X
solid state or many-body physics elementary particle physics
chemistry many-body physics
molecular biology chemistry
cell biology molecular biology

psychology physiology
social sciences psychology

But this hierarchy does not imply that science X is "just applied Y . " A t
each stage entirely new laws, concepts, and generalizations are neces­
sary, requiring inspiration and creativity to just as great a degree as i n the
previous one. Psychology is not applied biology, nor is biology applied
chemistry. (Anderson 1972, 393)

Combining Bhaskar's a priori deductions and Anderson's supporting


remarks, we can sum up the philosophical ontology employed in this paper by
listing five fundamental assumptions concerning the ontological structure o f
the everyday w o r l d : (1) the w o r l d is real in the sense that it exists and operates
independently o f our knowledge o f i t ; (2) the w o r l d is ontologically layered,
Social Science as the Study of Complex Systems 301

that is, it is made up o f structurally irreducible levels that form a loosely


organized hierarchical system; (3) while not specified by Bhaskar, there is
strong scientific evidence that symmetry breaking is central to producing this
stratified complexity; (4) the progressive accumulation o f scientific k n o w l ­
edge points to an irreversible time frame through which the world develops;
and (5) the emergence o f one layer from another suggests that the w o r l d is an
open, evolving entity.

I m p l i c a t i o n s o f Critical Realism
f o r Social Scientific M e t h o d

The above ontology entails certain methodological principles as w e l l . From


critical realism's perspective, there is the real world and what we say about
that w o r l d . Contrary to idealist epistemologies, our statements about the
structure o f the w o r l d are not the world as such, and the two must never be
confused. B y the same token, the empirical appearance o f reality should never
be mistaken for those generative structures and mechanisms whose powers
and limitations actually produce the world o f appearance. While experience
can often be the point o f departure for scientific inquiry, empirical presence
alone cannot give us access to the lawful structures that produce those appear­
ances. To penetrate to that more fundamental substratum o f causal forces,
more than controlled observation is required. Just as Kant rejected Hume's
empiricism, so critical realism rejects the claim that the experimental method,
in and o f itself, can provide logically feasible laws. Causal explanations
derived from empirical evidence are necessary, but never sufficient, for
the identification o f lawlike principles. The same assertion holds equally
for social scientific surveys, demographic analyses, or participant observer
studies. The empirical w o r l d can reflect, but never give us direct access to the
causal mechanisms o f reality.

Bhaskar is by no means among the first to argue against the idea that
classical prediction is the final arbiter for establishing scientific truth claims.
He has, however, contributed significantly to our philosophical understanding
of this problem by providing an ontological framework that establishes the
limits o f predictive validity in open evolving systems. The universal claims o f
prediction in such systems had already been seriously challenged by quantum
theory early in this century (Peitgen, Jiirgens, and Saupe 1992, 11-14).
Heisenberg's uncertainty principle denied that scientists could ascertain a
dynamic system's initial conditions at the subatomic level. It was impossible
to measure simultaneously and w i t h equal precision a particle's position and
its momentum. As a result, classical mechanics' claim that given any system's
position, momentum, and the dynamic laws, it could predict the system's
1
future state was denied. By the midsixties, chaos theory had further limited
302 Chaos Theory in the Social Sciences

the range o f classical causation's explanatory scope. By then, the science o f


nonlinear dynamics had demonstrated that even i f a scientist could pinpoint a
nonlinear system's initial conditions, there were many instances in which
precise forecasting o f future behavior was all but impossible. O n l y under the
constraints o f certain Lyapunov-specific conditions and w i t h the application o f
various renormalizing techniques could chaotic systems be analyzed—and
even then only under the most restrictive o f circumstances (Wolf 1986;
Peitgen, J ü r g e n s , and Saupe 1992, 5 0 9 - 2 0 , 7 0 9 - 2 0 ) .

Dissipative Social S y s t e m s

Following Bhaskar's logic, what is now needed is a scientific worldview


capable o f filling in the philosophical gaps. Such a worldview is already at
hand, we believe, in the science o f nonequilibrium thermodynamics. This
new science studies dissipative systems. Dissipative systems embrace a wide
variety o f chaotically driven systems: cross-catalytic chemical processes that
often seem to imitate life itself; physically and biologically constituted evolu­
tionary systems; and, as we w i l l see in the next section, social systems.
Dissipative systems, first and foremost, are natural systems. They are
materially and thermodynamically constituted entities. This means their inter­
nal structuration and development, as well as the processes by which they are
born, evolve, and die, are regulated by transfers o f energy from their immedi­
ate environment. The most important feature o f the thermodynamic com­
position o f dissipative systems is their negentropic potentiality. Negentropy,
or negative entropy, and its opposite, positive entropy, are irreversible
thermodynamic processes. Positive entropy refers to the universal tendency o f
thermodynamic structures to evolve irreversibly toward a stage o f maximum
disorder called thermodynamic equilibrium. The most effective measure o f
this increasing disorder is the progressive degradation o f the system's internal
structure. Thus, by the time thermal equilibrium is reached, the energy mak­
ing up the system has been smoothly and equiprobably distributed throughout
its boundaries.
Thermal equilibrium was long thought to be the fate o f all living things
and, indeed, o f the universe itself. In fact, the idea o f thermal equilibrium
became a controversial staple o f popular culture i n the late nineteenth century
as speculation arose concerning the eventual "heat death" o f the universe. In
the last few decades, however, research into systems ordered by negentropic
processes has changed our ideas about irreversible thermodynamic processes
and the role they play in a wide range o f evolutionary phenomena. Ne­
gentropy, unlike positive entropy, occurs in a limited set o f circumstances, but
when it does occur its ramifications are profound. As the name implies,
negentropy represents a tendency that runs counter to that o f positive entropy,
Social Science as the Study of Complex Systems 303

in that systems i n which it is produced are capable o f forestalling their descent


toward thermal equilibrium. Negentropic processes are, therefore, the mate­
rial foundations for the growth and evolution o f thermodynamic systems.
They enable dissipative systems, via internal metabolic mechanisms, to trans­
form free environmental energy into increasingly more complex structuration.
But for dissipative systems to sustain their g r o w t h , they must not only
increase their negentropic potential, they must also eliminate the positive
entropy that naturally accumulates over time and that degrades the system's
internal structuring. That is, as dissipative systems grow and become more
complex internally, the price o f their increasing structural complexity is mea­
sured in terms o f their rising levels o f positive entropy. The latter is a natural
waste by-product o f the process by which dissipative systems develop. I f
allowed to accumulate over time, positive entropy w i l l eventually nullify the
system's tendency toward increasing differentiation and w i l l move the system
toward thermal equilibrium. Negentropic systems must, therefore, develop a
means o f eliminating their accumulated positive entropy. This is usually
achieved by the system transporting its waste energy to its immediate environ­
ment.
For this reason, any long-lived dissipative system must fulfill two pre­
requisites i f it is to persist: (1) it must be able to convert free environmental
energy into ever more elaborate forms o f internal structuration; and (2) it must
transport thermal disorder into the environment. Dissipative systems are thus
characterized by a dynamic tension between their ability to accumulate ne-
gentropy and their need to transfer their positive entropy into the environ­
ment. I f they can sustain this tension, then under proper circumstances they
can achieve a state o f net negative entropy and persist. I f they cannot, they
w i l l evolve to a state o f thermodynamic equilibrium. On the basis o f their
ability to achieve or not achieve a state o f net negative entropy, then, we can
analytically partition thermodynamic systems into two sets: those dissipative
systems that are capable o f sustaining their internal growth and those evolving
toward a state o f thermal equilibrium.
Because dissipative systems can increase their negative entropy, they are
capable o f evolution. Two o f their properties, in particular, allow this to occur.
First, dissipative systems are peripatetic in the sense that they constantly seek
new organizational states. Second, dissipative systems are capable o f achiev­
ing a far-from-equilibrium state that allows them to move to reference states
so distant from thermal equilibrium they can transform themselves into more
complex entities. Both properties, internal peripatetic boundary testing and
far-from-equilibrium dynamics, are sources o f instability and, hence, o f po­
tential change. When these t w o instabilities combine, they give dissipative
systems an evolutionary potential that differentiates them from other forms o f
organization. Let us see how this is done.
304 Chaos Theory in the Social Sciences

The internal source o f a dissipative system's instability resides in its


boundary-testing propensities. That is, dissipative systems are constantly try­
ing to transform themselves—moving away from their present point o f equi­
poise to some alternative state. M o r e often than not, however, this boundary
testing goes for naught since the environment fails to provide a sufficient
energetic base to sustain an evolutionary transformation. Periodically, how­
ever, this internal boundary testing coincides with a sudden, sustained i n ­
crease i n environmental energy. Readied to change by internal boundary
testing and suddenly perturbed by an environmental flux, the far-from-
equilibrium potential o f the system can destabilize i t , force it to abandon its
previous reference state, and begin to evolve to a new configuration.
This evolutionary process begins with the appearance o f bifurcational
behavior, that is, a fluctuating behavior that sends the system into an oscillat­
ing movement between t w o or more new points o f possible equilibrium. This
cyclic oscillation indicates the dissipative system in question has destabilized
and entered a chaotic phase. A t this point one o f two things can happen: the
system can remain chaotic, oscillate ever more rapidly, and eventually destroy
itself, or the fluctuation w i l l dampen as the system settles into a new configu­
ration. When this happens, we speak o f the dissipative system as having
evolved: it has reorganized itself around a new reference point from which it
w i l l once again begin its peripatetic boundary-testing behavior. This testing
w i l l continue until a new perturbation o f sufficient force occurs, and a new
conjunction o f internal and external conditions once again pushes the system
into a new evolutionary trajectory.
Dissipative systems are, hence, chaotically structured, far-from-equilibrium
entities. Their evolution proceeds from two crucial preconditions. First, dis­
sipative systems exhibit what is called a sensitive dependence on initial condi­
tions, a property that is a cardinal indicator o f chaotic structures (Ruelle 1990;
Devaney 1989, 4 4 - 5 3 ) . As we have seen above, the evolution o f dissipative
systems requires a coincidence o f internal boundary testing and external per­
turbations to push it into a new evolutionary path. I f this flux were to impinge
on the far-from-equilibrium system either a moment sooner or a moment later,
this lag or acceleration could either suppress a possible evolutionary tendency
or accelerate i t , sending the system into a path it might otherwise not have
followed.
This sensitive dependence on initial conditions is a key indicator o f
deterministic chaos. Its dynamic presence is a major reason why evolving
systems rooted in the cyclical and self-feeding dynamics o f deterministic
chaos are unpredictable. It is noteworthy, therefore, that with the concept o f
sensitive dependence on initial conditions we find a major point o f conjunc­
ture between our philosophical ontology and the scientific ontology o f dissipa­
tive systems.
Social Science as the Study of Complex Systems 305

A similar convergence holds for the second crucial precondition—the


process o f symmetry breaking. To say that far-from-equilibrium systems
evolve, and that they do so through symmetry-breaking processes, is to state
the same fact at t w o different levels o f abstraction. A t the first level, symmetry
breaking indicates the crucial fact that far-from-equilibrium systems are not
conservative entities. To understand what this means, let us imagine that the
normal growth and evolution o f a system takes place in an n-dimensional
space. We w i l l call this space, for simplicity's sake, a phase space. This phase
space can be visualized as being structured by such dimensions as time and
space coordinates, the degree o f its internal complexity, the nature o f the
cycles governing the internal dynamics o f the system, and the total environ­
mental energy available to the system. A system is conservative when any
evolution it undergoes takes place within the existing parameters o f its phase
space. The growth o f conservative systems preserves the dimensionality o f its
phase space.
In that far-from-equilibrium systems are not conservative constellations,
they do not preserve their phase space during their evolution. Instead, dissipa-
tive systems i n a far-from-equilibrium state are predicated upon bifurcational
processes i n which new dimensionalities emerge to foster change. Put another
way, when a system's evolution requires the breaking o f old symmetries as a
conditio sine qua non for pursuing its o w n far-from-equilibrium path, then it
is not enough to speak just o f the evolution o f an individual system; we must
also recognize a second, global process—the evolution o f evolution itself.
This second evolution is the concrete condition that separates the evolution o f
far-from-equilibrium systems from all other forms o f change. The evolution o f
evolutionary space allows us not only to differentiate between the evolution
of the physical universe, molecular evolution, the evolutionary emergence o f
new species, and, finally, the evolution o f dissipative social systems; it also
reminds us that each different form o f evolution represents discrete moments
of the same self-organizing, evolutionary process that categorically unites all
dissipative systems into a meaningful category.

Social O n t o l o g y : The S t r u c t u r e o f Dissipative


Social S y s t e m s

Dissipative social systems are a subset o f dissipative systems. This means


much o f what is said o f natural dissipative systems is also true o f dissipative
social systems. Hence, dissipative social systems have many o f the charac­
teristics o f natural entities, and, like them, they are grounded in the material
w o r l d . Consequently, any paradigm that studies dissipative social systems
must begin with an accounting o f the economic and ecological components
that mediate between societies and their natural milieu. This does not mean,
306 Chaos Theory in the Social Sciences

however, that an ecological approach to the study o f dissipative social systems


is a warrant for ignoring those symbolically constituted elements that com­
pose a society's culture, nor does it entail the reduction o f the cultural domain
to the material processes and interests that gave rise to i t . It can be demon­
strated, in fact, that cultural systems, as linguistically mediated, symbolic
constellations, have many o f the same dissipative traits as the human commu­
nities that produced them. There is nothing in a materialist, dissipative sys­
tems paradigm, then, to prevent us from dealing forthrightly w i t h cultural
phenomena.
In addition to their material constitution, dissipative social systems dis­
play far-from-equilibrium behavior. That is, their evolution is chaotically
driven and is sensitive to initial conditions. Further, the structure o f dissipa­
tive social systems is generated by symmetry-breaking mechanisms, and is,
consequently, ontologically layered. These evolutionary properties establish
the foundations for the historicity o f dissipative social systems. Hence, any
ecological or cultural analysis o f institutional life must take into account both
the origins and the developmental history o f the entities and events under
investigation.
When these attributes are gathered together and applied to the study o f
society, a portrait o f dissipative social systems emerges that presents us w i t h
the image o f an inherently historical entity whose evolution is driven as much
by internal instability as by external perturbation. Moreover, the grounding o f
dissipative social systems in nature and i n the dynamics o f deterministic chaos
demands a materialist interpretation o f dissipative social systems not unlike
that developed by critical M a r x i s m .
Despite their commonalities, however, there are important differences
separating dissipative social systems from their physically constituted counter­
parts. Most o f these differences hinge on the fact that societies and their
institutional activities are constructed by the collective action o f human be­
ings, and, thus, are profoundly influenced by the way in which humans
subjectively define themselves and their actions. This fundamental difference
has already been expressed in Bhaskar's critical naturalist paradigm, for when
he describes society and its functions he underscores the " w i l d card" nature o f
human beings and their innovative abilities. This same exceptionality has long
been recognized in dissipative systems theory, and can be neatly inserted into
the paradigm advocated by Prigogine and the Brussels School (Prigogine and
Stengers 1984; Nicolis and Prigogine 1989).
Thus, as long as the existence and impact o f this w i l d card on institutional
life is kept i n the forefront o f one's analysis, it is possible to assume that
dissipative social systems can be studied in a manner similar to the study o f
natural dissipative systems. Failure to recognize the indeterminate aspect o f
human nature, however, w i l l result in that reification that is so common in
positivist and behaviorist interpretations o f social life (see Bhaskar 1989a, 38).
Social Science as the Study of Complex Systems 307

Strategies o f M o d e l i n g a n d Their O n t o l o g i c a l L i m i t s

Having specified our metatheoretical premises, we can now describe how we


constructed the modeling/ontology matrix. The matrix is displayed i n figure
13.1. Its six columns, labeled "Levels o f Modeling Abstraction," cover a
wide range o f techniques currently in use i n the social sciences. Starting w i t h
predictive models, these strategies are ordered so that as one moves from left
to right there is a decreasing tendency for the strategy in question to produce

Modeling Strategies for Studying Chaotic Social Systems:


Hierarchy of Ontological Complexity Arrayed by Decreasing Determinist Presuppositions (horzontal axis)
in Social S y s t e m a n d by Levels of S y s t e m Specificity (vertical axis)

Process I I : Societal evolution


via historical m o d e s
of production

Process 1: C l a s s struggle.
Conflict over cultural h e g e m o n y

12 Values I I : H e g e m o n i c culture &


subcultural b a s e s of resistance

11 V a l u e s 1: Struoale of
hegemonic vs. s u b t e r r a n e a n
world views

10. N o r m s II: Allocation of relative


p o w e r a m o n g social institutions

9. N o r m s 1: Personal c o n f o r m i t y
to g e n e r a l h e g e m o n i c s t a n d a r d s

8. Roles I I : Intraoraanization
allocation of roles & resources

7. Roles 1: Distribution of
material r e w a r d s & e s t e e m

6 Facilities I I : T e c h n i c a l division
of labor in productive sphere

5. Facilities 1: S o c i o t e c h n i c a l
infrastructure of organization

4. Ecological organization of
institutional t i m e & space

3. Ecological organization
of local biotic c o m m u n i t y

2. Biological evolution a s a
series of assisted bifurcations

1. Determinant regularities of the


physical universe

Predictive Statistical I c o no logical Structural Ideal T y p e Historical


Modeling Modeling Modeling Modeling Modeling Narratives

Levels of M o d e l i n g A b s t r a c t i o n

Fig. 13.1. P a r a d i g m s h o w i n g general range o f f i t b e t w e e n m o d e l i n g


s t r a t e g i e s a n d d i f f e r e n t levels o f o n t o l o g i c a l c o m p l e x i t y i n t h e social
system
308 Chaos Theory in the Social Sciences

lawlike, scientific propositions. Conversely, moving in the opposite direction,


one moves from modeling strategies designed to understand particular people
and events to those that formulate abstract, general propositions. Hence,
nomothetic methods occupy the left half o f the matrix, while ideographic
methods populate the right. I t should be emphasized from the start, however,
that our ability to locate these various modeling strategies on a continuum,
should not obscure the fact that there are radical differences between these
strategies. As we w i l l point out later, we encounter several epistemological
breaks as we move along this abstraction dimension.
The fourteen rows o f the matrix are organized from bottom to top i n
terms o f increasing ontological complexity. As noted i n the introduction to
this chapter, this ontological array has been constructed using the works o f
Kenneth Boulding (1968, 3 - 1 0 ) and Neil J. Smelser (1963, 1-46). Rows one
through four o f the matrix are adaptations o f Boulding's original w o r k , while
rows 5 through 12 are adapted from Smelser and used to elaborate upon the
"undeveloped upper end," so to speak, o f Boulding's grand ontological
scheme. Similarly, Boulding's typology is used to specify more precisely
those ontological levels to which structural-functional analysis has tradi­
tionally given short shift, that is, the sociotechnical and ecological founda­
tions o f the social system. Hence, the natural and biological milieu in which
dissipative social systems operate occupies the lower rungs o f the hierarchy.
The ecological and sociotechnical spheres o f social life mediating between
society and its natural environment come next, and, moving up the hierarchy,
we encounter the institutionalized systems o f resource distribution, social
organization, and moral evaluation. A t the top o f the hierarchy are the general
regulatory processes that structure modern societies and their evolution. In
this context, class conflict and social evolution are not ontological levels, but
dynamic processes that drive the evolution o f the social system as a whole.

As noted earlier, the ontological levels represented in the row structure o f


the matrix are systemically integrated. Lower levels form the loose founda­
tions and conditions by which higher levels emerge and operate. Concomi­
tantly, the higher levels, once established, feed back upon and delimit the
operation o f those levels that undergird their very existence. We w i l l not,
however, dwell for long on the dynamic integration o f this hierarchy, since we
are more interested in them as static configurations and in the way the struc­
ture o f these configurations brackets the scientist's choice o f method.
Finally, the last feature o f the matrix is the shaded cells running diag­
onally from the lower left-hand corner to the upper right-hand corner. These
cells are hypothesized areas o f optimum fit between modeling strategies and
the different ontological levels o f dissipative social systems, while those
projects falling on either side o f the diagonal are seen as analytically problem­
atic. We w i l l take up the interpretation o f these areas in more detail later on.
Social Science as the Study of Complex Systems 309

M o d e l i n g Strategies and Levels of A b s t r a c t i o n

We have been careful in this chapter to distinguish between deterministic


chaos theory and the modeling techniques actually used to study chaotic
phenomena. Models, as opposed to theories, are well-formed metaphors and
analogies. They do not claim to express the truth o f the w o r l d , but merely to
provide heuristic insights. While theories claim to actually explain reality,
models are only partial, fictitious constructions. They speak a language o f "as
i f , " not "what i s . " But i f models can make few explanatory claims, they are
rich in the conceptual materials upon which they can draw and are freer to
organize those materials i n a manifold o f different directions ( L l o y d 1986,
124-37; Bhaskar 1978, 143-63). Unconstrained by a formally operationist
logic or the presuppositions o f a well-articulated paradigm the way theories
are, modeling freely participates in acts o f imagination to produce a wide
range o f alternative insights to old problems.
These are fairly standard observations concerning models and their gen­
eral scientific functioning. Kaplan (1964) has taken these observations a step
further. He claims models have as their basic focus the mapping o f formal
similarities and the establishment o f strict isomorphisms between systems.
According to Kaplan, once these isomorphic similarities are established, the
goal o f modeling is to map specific structural correspondences between t w o or
more systems (Kaplan 1964).
There are several ways to establish these correspondences in the social
sciences. The column structure o f figure 13.1 lists what we believe to be the
major families o f modeling strategies. Predictive models are the most famil­
iar, for they have been procedural paragons in the social sciences for almost
two hundred years. As noted above, this strategy is being eclipsed due to
discoveries that have been made both by quantum theorists and by chaos
researchers. Its insistence that the test o f a theory is its ability to make long-
term predictions is not compatible with the dynamics o f dissipative systems,
and should be handled gingerly by all those engaged in chaos research. The
second family o f models depicted in our matrix is statistical modeling. I t
focuses on fitting data to a set o f mathematical curves. Here the criterion o f
adequacy is not prediction o f a system's future state, but the post hoc close­
ness o f fit a researcher can forge between longitudinal data and a given
mathematical function.
W h i l e predictive and statistical models are well established in the social
sciences, a third modeling strategy, iconological modeling, is only now com­
ing into its o w n . It represents a radical epistemological break in the type o f
knowledge it provides to researchers. Iconological modeling is rooted in a
pictorial method, in visual correspondences rather than in deductive reason­
ing. Iconological modeling is a recent innovation, originating in the iterative
310 Chaos Theory in the Social Sciences

mapping o f complex systems o f equations—such as the nonlinear differential


equations that generate the quadratic iterator, or the so-called strange attrac-
tors. Iconological mapping arose in large part as a necessary response to the
complexity o f the patterns generated by the iterative analysis o f these equa­
tions. Their evolution becomes so complex, so rapidly, that they overwhelm
human comprehension. Thus, graphical techniques assist the researcher in
visually tracing the chaotic trajectories o f these iterative systems. Indeed,
Gregerson and Sailer (1993) suggest this form o f modeling would have been
all but impossible three decades ago, for much o f the computer know-how had
not yet been perfected to achieve such graphics. Thus they note:

M u c h o f the research performed on chaotic systems i n the physical and


biological sciences has benefitted from visual representations o f chaotic
behavior using high-powered computers w i t h advanced graphic capa­
bilities. . . . I n order to better understand chaos, it is useful to have a
way to produce a visual representation (derived from mathematical
models) o f chaotic behavior. (Gregerson and Sailer 1993, 7 8 3 - 8 4 )

Iconological modeling thus represents a break with earlier modeling strategies


and has been made possible by recent technical progress. Its novelty, how­
ever, resides in the priority it gives to the visual over the analytic. In philo­
sophical terms, we could say that in iconological modeling the gaze is more
important than deductive logic in grasping the evolution o f a chaotic structure.
There is something refreshing yet discomfiting about this almost ad hoc exper­
imental approach. W h i l e the iconoclastic style o f those engaged in this form
o f modeling may be a remnant from the sixties, there is something new and
significant here that was not possible i n the past.
The term we have chosen to describe this modeling strategy has been
borrowed from the Renaissance art historian E r w i n Panofsky (1972, 3 - 3 1 ) .
Iconological analysis, according to Panofsky, is preeminently visual and phe-
nomenological in its exercise. It employs various levels o f interpretive an­
alyses that range from the recognition o f simple visual forms to attempts to
capture the Weltanschauung o f an epoch. More precisely, iconology is the
interpretive study o f aesthetic objects—their visual style and presence, their
formal structure, the cultural symbolism employed in their construction, and
the social forces and personalities that produced them. The uniqueness o f the
iconological method lies in its ability to recognize how an aesthetic object's
phenomenal appearance can communicate all aspects o f itself, and the ways in
w h i c h they express those different aspects. In a real sense, Panofsky's aes­
thetic epistemology parallels the phenomenological and analytic process the
scientist uses when trying to interpret the meanings o f the graphic images
whose unfolding maps the chaotic evolution o f a system.
Social Science as the Study of Complex Systems 311

As we move to our fourth family o f research strategies, structural model­


ing, we encounter a second radical break. W i t h structural modeling we meet a
strategy that in articulating and testing its results largely discounts the evi­
dence o f the senses in favor o f purely formal analysis. Structuralism holds that
the truth o f the w o r l d , including that which is presented by deterministic
chaos, can be captured by axiomatic analysis alone. That is, Platonic-like
intuitions are sufficient to grasp the truth o f models constructed using axiom­
atic methods. Such refined intuitions guide the researcher past empirically
based deceptions and pitfalls, and are able to immediately grasp the structural
essence o f the object itself. This ultra-idealism is given its classic statement
by Thorn (1983) when he argues against those statistical analyses o f chaos that
begin and end w i t h empirical reality. Thorn believes that to understand chaos,
one must not reify the empirical patterns themselves, but unearth the "simple
mathematical forms" o f which they are composed. Employing an ontology
that echoes Plato's vision o f the w o r l d as being structured by a concatenation
o f ideal geometric forms, Thorn writes:

Here I am a Platonist: we forever see only the shadows o f things, and, it


is necessary for us to go beyond the cavern's w a l l . . . , hence to imagine
a reality o f that space beyond, projecting itself on observed phenomenol­
ogy. . . . Replace the complicated visible by the simple invisible: . . . it
is also necessary to think o f simple mathematical forms which generate
complicated empirical datum. (Thorn 1983, 80)

This faith in the ability o f mathematics to penetrate to the very core o f


our world is not new. I t has been with us almost as long as the species has
been engaged in speculative reasoning. In modern times, since the Renais­
sance and the age o f Newton, and into the present, there has been a cult whose
professions o f faith in the power o f mathematics have been profound. One o f
the most touching o f these bears a mystical awe that is hard to find in an age o f
opportunism and high-profile cynicism such as ours. Thus, Nick Herbert
quotes the physicist Eugene Wigner in discussing mathematics and its relation
to the real w o r l d .

In the seventeenth century Galileo, Newton, and the other natural


philosophers discovered that an enormous body o f physical facts could
be encompassed in a few mathematical formulas. For instance w i t h only
three mathematical laws Newton could explain all motion in heaven and
on earth. W h y should mathematics, developed primarily to keep track o f
human business transactions, have anything at all to do with the way the
non-human w o r l d operates? Nobel laureate Eugene Wigner refers to the
magical match between human mathematics and nonhuman facts as "the
312 Chaos Theory in the Social Sciences

unreasonable effectiveness o f mathematics in the natural sciences."


"This unreasonable effectiveness," writes Wigner, "is a wonderful gift
which we neither understand nor deserve." (Herbert 1985, 2)

When Wigner speaks o f that "wonderful gift which we neither under­


stand nor deserve," he expresses the deep emotions we have all felt when we
first glimpsed the elegance o f self-similar constructions (Peitgen, Jiirgens, and
Saupe 1992, 6 3 - 1 8 2 ) or the manifold possibilities that Feigenbaum's period-
doubling law, w i t h its suggestive cascading o f bifurcation points, might hold.
The same astonishment accompanies our discovery o f the wholly unexpected
behavior a quadratic iterator displays when we vary its control variable
through the smallest range o f numerical values (Baumol and Benhabib 1989;
Peitgen, Jiirgens, and Saupe 1992, 5 0 7 - 7 6 8 ) . A n d certainly one cannot sit at
a monitor and watch the unfolding orbit o f Lorenz's strange attractor without
being deeply moved.
The social sciences, to be sure, are not immune to such enthusiasms. Posi-
tivists have for years reined mathematics, applauding it as being the pinnacle o f
knowledge, and urging the social sciences to emulate its example. To match
the mysticism and awe o f Wigner in the social sciences, however, we need
go no further than the work o f the French structuralist Claude Lévi-Strauss.
Lévi-Strauss's (1955; see also Badcock 1975, 3 3 - 6 7 ) structural analysis o f
m y t h , for example, resonates w i t h Thorn's Platonic manifesto. Lévi-Strauss
begins his analysis by collecting different versions o f a single myth drawn
from a number o f different groups or societies. He then abstracts from this
manifold o f empirical detail the key elements o f the myth he is studying. I n
that none o f the original myths contain all o f its essential elements, it is the j o b
o f the social scientist to assemble the fully developed true myth. This recon­
structed myth is an idealized, composite configuration containing all o f the
essential components. Working w i t h these ideal elements, the scientist builds
a relational model o f the myth and shows how its various elements are opposi-
tionally j o i n e d to one another. Since the original versions o f the myth are seen
as incomplete or distorted by the accidents o f local history, they are treated by
the scientist as inferior to his or her rationally reconstructed model. This
reconstituted myth is the equivalent o f Thorn's "simple invisible." It is the
intellectually fulfilled object from which the chaos o f the different empirical
myths has emerged. It is this logically precise and coherent myth that is the
proper object o f a theoretical science.

Ideal typical modeling constitutes the mirror image o f structuralist anal­


ysis. Scientists w h o employ ideal typical modeling in their research are not
interested in structural commonalities. Quite the contrary, they are interested
in the singularities that create deviations from homogeneous patterns. The
Social Science as the Study of Complex Systems 313

structuralists, as we have just seen, garner as many different versions o f a


class o f phenomena as possible. From these they construct a consistent core o f
attributes. They then discard or ignore the variations and turn to the logical
essence or generative form o f the materials being studied. Their goal is to
capture the logically pure form, for therein lies reliable, theoretically articu­
lated knowledge.
Those engaged in ideal typical modeling also gather as many different
empirical examples as possible. They also seek commonalities among the
class o f objects surveyed and, like the structuralists, construct from those
commonalities a logically pure type that captures the essence o f the class o f
entities under study. But here the similarities between the two methods end,
for after the ideal type model is constructed, it is "discarded," so to speak, as
the scientist turns his or her attention to the deviant elements that cannot be
subsumed under the logically pure type. Using the ideal type to identify these
singularities, the scientist focuses upon the nonlogical elements that set each
instance off from the logical pattern. In these nonlogical exceptions, the
scientist discovers what is historically and developmentally unique i n each
instance. The ideal typical modeling strategy, in sum, produces a variety o f
ideographic knowledge that can serve as the basis for a comparative history
and social science.
Historical modeling is close in its form and content to ideal type model­
ing. Like the latter, historical narrative seeks to understand the singular truth
o f events and the manner in which they unfold. What separates ideal typical
modeling and historical narrative from one another is the stubborn insistence
o f historical modeling on the singularity o f the events and personages it is
reconstructing. The purpose o f historical narratives is to render as close a
reading as is possible o f concrete events. From this reading, statements o f
empirical causation may well be constructed. What is denied, however, is that
the causes creating one chain o f events can be compared with the causes
giving rise to other chains o f events. From this perspective, narrative history
is possible, but a comparative science o f history is not.
W i t h the consideration o f narrative history, we conclude our survey o f
the construction o f the methods/ontology matrix. This matrix compares mod­
eling strategies w i t h the different ontological levels o f dissipative social sys­
tems. In identifying different ontol