0 Stimmen dafür0 Stimmen dagegen

405 Aufrufe360 SeitenAug 21, 2010

© Attribution Non-Commercial (BY-NC)

PDF, TXT oder online auf Scribd lesen

Attribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

405 Aufrufe

Attribution Non-Commercial (BY-NC)

Als PDF, TXT **herunterladen** oder online auf Scribd lesen

- Relatório Stern
- Ethiopia
- The Analysis of Mind by Russell, Bertrand Arthur William 3rd, Earl, 1872-1970
- Krugman
- Collective Behaviour and Public Opinion
- The Apache Wars - The Final Resistance
- Selfchao.isr
- Economic Contractions in the United States: A Failure of Government
- Chaos Theory-The Essentials for Military Applications
- Is China a Threat to the U.S. Economy
- Linked Histories
- Eco No Physics and Economic Complexity
- The Return of the Political by Chantal Mouffe
- Theory Of Everything
- The US-Sino Currency Dispute: New Insights from Economics, Politics and Law
- Logic,Hermann Lotze
- Introduction to Chaos - Theoretical and Numerical Methods - Carlo F Barenghi
- ASPECTS OF DIALECTICS AND NONLINEAR DYNAMICS
- Chaos
- EMERGENCE AND COMPLEXITY IN AUSTRIAN ECONOMICS

Sie sind auf Seite 1von 360

Chaos Theory

in the Social Sciences

Foundations and Applications

Edited b y

Ann Arbor

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

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.

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

To Our Parents

Contents

Introduction

Euel Elliott and L . Douglas Kiel

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

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

J. Barkley Rosser, Jr.

viii Contents

Chaos, and Control 215

Brian J. L . Berry and Heja Kim

11. Cities as Spatial Chaotic Attractors 237

Dimitrios S. Dendrinos

and Social Science

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

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

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

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 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.

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

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).

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:

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

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

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.

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

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

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.

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

"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.

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.

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

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

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

Spreadsheet: A Graphical V i e w

of Chaos for Beginners

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

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

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.

0

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

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.

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

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

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

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

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

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

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

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

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:

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.

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

Systems: A n Introduction t o Spectral Analysis

Michael McBurnett

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

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

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

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

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

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.

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)

Probing the Underlying Structure in Dynamical Systems 35

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

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

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

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 -

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.

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

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

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

Cycles per Unit Time

(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 -

(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

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

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.

equations. These equations comprise an early deterministic model o f the

Probing the Underlying Structure in Dynamical Systems 43

dot notation) is

JC] = 10(— x , + x)

2

x 2

=

28x| X2 -^1-^3 (^)

x 3 = —(8/3)x + 3 xx-

x 2

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

t i n c t features.)

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

problem:

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

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

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

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_

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

sentially featureless)

50 Chaos Theory in the Social Sciences

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

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

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

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

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

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

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

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

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.

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 marginally stable

A | > 0 =^ orbit is chaotic

58 Chaos Theory in the Social Sciences

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.

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

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

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

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)

=

0

60 Chaos Theory in the Social Sciences

A/-^» 1=0

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

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

0

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

i=l

| / ' ( 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,

\(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

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

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

( 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.)

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

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

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.

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

Ted Jaditz

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.

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.

=l

0

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

and

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

a, = h(X,) (2)

involving the unobserved n vector o f state variables. Given an initial starting

The Prediction Test for Nonlinear Determinism 69

0 0

2 0

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.

0

(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

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)

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

2 2 3 3 4

case. Takens (1985) shows that the mapping from the m-histories o f a se

quence a, to the state vectors X, given by

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

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

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.

0

0

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

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

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

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)

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)

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

NRMSE = RMSE/cr r

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

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

-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

tic for the hypothesis that the corresponding coefficient

is significantly different from zero.

The Prediction Test for Nonlinear Determinism 11

1.2 t

Number of Near Neighbors

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.

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

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

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

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

+s

assumption that anomalies or quirks in the data have only a transitory impact.

80 Chaos Theory in the Social Sciences

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.

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

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

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

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

tic for the hypothesis that the corresponding coefficient

is significantly different from zero.

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

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-

E = 0.81 E = 0.9

Dimension Dimension Statistic BDS Entropy Statistic BDS Entropy Statistic

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

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.

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

Data Regression

-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

tic for the hypothesis that the corresponding coefficient

is significantly different from zero.

The Prediction Test for Nonlinear Determinism 85

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

e = 0.81 £ = 0.9

Dimension Dimension Statistic BDS Entropy Statistic BDS Entropy Statistic

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

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

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

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

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

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

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

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

of these three conditions:

tions if two points that are arbitrarily close separate by a distance 8 by the

iteration of that function (Devaney 1989, 49).

has points that move from one arbitrarily small neighborhood to another

arbitrarily small neighborhood (Devaney 1989, 49).

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

42).

1989, 50):

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

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

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

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

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

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

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

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

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)

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

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

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.

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

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

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

a

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

b c h

a

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

c a

+

b a

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

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

(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.

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.

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.

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

ab bc ca

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).

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

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

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

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.

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.

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

Group Aggregations

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

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

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

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

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.

in Chaotic D y n a m i c s

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.

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

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 ) .

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

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 • —

Issue 1

:' ν'

/ ' ·.

' ,Λ S

; if ' «. IV

;* ·

"t

H ' i

- >·(··

I t .*

«#•r·>

<• 3Γ

1

1

10 30 50 70 90

Issue 1

110 Chaos Theory in the Social Sciences

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

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

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.

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.

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.

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

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

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

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.

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

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

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

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

(«!, 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

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

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

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

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

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

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

measure provides a useful definition o f distance in phase space

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

t

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

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

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

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

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

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

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

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.

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

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

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

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

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.

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

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

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

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).

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:

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

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.

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

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.

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

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

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

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).

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

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

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.

152 Chaos Theory in the Social Sciences

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

0

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

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

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

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 ?

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

n

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

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):

n

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

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.

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

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

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

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

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

yx

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

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)

n y 0 (X n

"xy

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

epsilon = 0.001

20

V

0 10 20 30 40

a

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?

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:

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)

u u

yx yz

0 (Y N - Z„)] 0 ( X

-=^ )

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

n n n n - Y„) © (z„ - Y„).

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

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

n n Z„)

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„)

z

0 (K„ - Z„ - X„)] 0 ( )

n

zx

-T~(Z n + X )(Z„

n + X „ - Y ) 0 (Z„ + X

N N - Y ).

N

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

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

0 0

(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

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

0

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

162 Chaos Theory in the Social Sciences

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

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

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.

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

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

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

( 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

(2)

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

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

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

where b represents (B — B B

0 — B B[B ),

0 b represents (1 + B — B —

2 0 2 0 x 2 i

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

2 l 2 2 2

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:

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

Hart .35 0.4304 -0.2937

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

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

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

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

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„

. . . , 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

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

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

c a n d i d a t e race

Complexity in the Evolution of Public Opinion 173

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

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

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

variation. Candidate Y only

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

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

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

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

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.

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

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

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)

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

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

on Estimate of the Correlation Dimension

.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%

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

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

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

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

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

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.

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.

Time Series

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

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

dx i n+ = dx„- f [x ] n

= dx n · r(l - 2x„).

error, one can write

n-l

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

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

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

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).

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

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

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

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-

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

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.

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.

D i m e n s i o n s T h r e e a n d Four

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

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

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

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

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

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

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

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

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,

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)

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

sumption behavior. Chaotic dynamics appear for values greater than 3.57.

The second process is given by

(4)

xt+l = ffu e

f

a<1

*>> (5)

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 ! "

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

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

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

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

The young seek to maximize utility

a (6)

p(t)c (t)

y + p(t + \)c {t

a + l)=p(t+ l)w .

0 (7)

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)

R = [(c , c ) : c

y a y + c

a = w y + wj, (9)

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

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

7

Chaotic cycles w i l l occur i f R (a ) > R (a ).

0 0 y y

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

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

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.

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.

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)

This implies an optimal time sequence o f prices given by

(p, )n

+l = (1 + - p,). (12)

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

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

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-

prices

Chaos Theory and Rationality in Economics 209

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

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.

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

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:

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

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 .

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 ( 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

Chaos, and Control

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:

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

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

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

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

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.

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

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 ) :

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, = 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

- Annual Rate of Change

•20 -10 10 20 30 40

Annual Rate of Change in Price [t)

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

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

—O— Annual Rate of Change

Annual Rate of Change in Price (r) Annual Rate of Change in Price (f)

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

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

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 . _

2 0

30 40 SO 60 70

—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)

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).

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.

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

" !

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

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 o f Price C h a n g e {t)

£ 20

(D 20 TO

•

20 -10 0 10 20 30 40

a.

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

basins.

From 1895 to 1930 the two-cycle Kuznets pattern is not as sharply

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

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

(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

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 (*)

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

.15 -10 5 0 5 10 15 15 10 -5 0 5 10

Discussion

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

2 5 » *

••t

% ffc * at

1 1

Annual Rate of Change in Growth (f)

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 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

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

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

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.

237

238 Chaos Theory in the Social Sciences

Contributions

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.

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.

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

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

and testing for a truly dynamic social science theory, currently perceived as

insurmountable.

Lack of Data

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

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

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

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.

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

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 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:

244 Chaos Theory in the Social Sciences

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)

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:

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)

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)

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

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)

x + a 22 In x (t)

2 + a 23 In x (t)3 (11)

x + a 32 In x (t)

2 + a 33 In x (t)3 (12)

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

F = 1.

x

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

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 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.

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

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.

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).

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

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

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

2 2

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.

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

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

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

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

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

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

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.

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.

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.

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)

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.

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)

(23)

In x (t 3 + 1) = In JC,(0 - In x (t) 2 + a 33 In x (t) 3 - In F(t)

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

2

2

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

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

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

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

2 3

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

3 3

Cities as Spatial Chaotic Attractors 261

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

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)

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

266 Chaos Theory in the Social Sciences

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

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

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

(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 .

AND ITS ATTRACTION ANTITRIANGLE

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

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)

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

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

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

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

of Occupied Space

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.

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 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

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

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).

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

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.

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

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.

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

ning with the Industrial Production Revolution about 1785. We are now, an

282 Chaos Theory in the Social Sciences

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.

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

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.

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.

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

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

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

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

(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

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).

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

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

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

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.

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

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

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

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

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

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

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.

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

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:

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)

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

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

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

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

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

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

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

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 S y s t e m s

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

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

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

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)

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

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

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

physical universe

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

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

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.

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

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

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:

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 )

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

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:

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)

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 .

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

"This unreasonable effectiveness," writes Wigner, "is a wonderful gift

which we neither understand nor deserve." (Herbert 1985, 2)

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.

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

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

## Viel mehr als nur Dokumente.

Entdecken, was Scribd alles zu bieten hat, inklusive Bücher und Hörbücher von großen Verlagen.

Jederzeit kündbar.