c I•• II C -I

,~ I
- c
-f -:]
c::
I rI
;JJ
r-
~
-1 1 ~
==>
;JJs l
..
C-tt...~
-' I r'
~
 Copyright © Paul R. Krugman, 1996

The right of Paul R. Krugman to be identified as author of this work has been
asserted in accordance with the Copyright, Designs and Patents Act 1988.

First published 1996

Blackwell Publishers, Inc.

238 Main Street
Cambridge, Massachusetts 02142
USA
Contents
Blackwell Publishers, Ltd.
108 Cowley Road
Oxford OX4 1JF
UK
Preface v
Except for the quotation of short passages for the purposes of criti-
All rights reserved.
cism and review, no part may be reproduced, stored in a retrieval system, or transmit- PART I
ted, in any form or by any means, electronic, mechanical, photocopying, recording or Embryos, Earthquakes, and Economics 1
otherwise, without the prior permission of the publisher.
1 Self-Organization in Space 9
Except in the United States of America, this book is sold subject to the condition that The von Thiinen-Mills Model 9
it shall not, by way of trade or otherwise, be lent, resold, hired out, or otherwise cir- 13
Central Place Theory
culated without the publisher's prior consent in any form of binding or cover other 15
Schelling's Segregation Model
than that in which it is published and without a similar condition including this con-
4 Edge City Dynamics 21
dition being imposed on the subsequent purchaser. I_I ....J 25 Complex Landscapes 31
Library if Congress Cataloging-in-Publication Data ./ f' J
..3
(7') The Emergence of Order 36
The self-organizing economy I Paul R. Krugman. v: 3 An Urban Mystery 39
t/ r l.,I (' 43
p. em. •.. t:.,
Meteorites, Earthquakes, and Cities
Includes bibliographical references and index. Simon on Zipf 44
ISBN 1-55786-698-8 (he) <:>
Principles of Self-Organization 47
ISBN 1-55786-699-6 (pb) Order from Instability 47
1. Economics. 2. Self-organizing systems. I. Title . Order from Random Growth 49
HB199.k75 1995 Where We Stand 51
338.9--dc20 95-31593
CIP PART II
Self-Organization in Time and Space 53
British Library Cataloguing in Publication Data
6 Dynamics in Self-Organizing Systems 55
A CIP catalogue record for this book is available from the British Library. 7 Temporal Self-Organization 61
Two Questions about the Business Cycle 61
Composition by Megan H. Zuckerman 63
Nonlinear Business Cycle Theory
Printed in the USA by Book Crafters Percolation Economics 69
Phase Locking and the Global Business Cycle 71
This book is printed on acid-free paper 8 Models of Spatial Self-Organization 75

() \. 9:,(a - \ Cl Cj ("' J\ IY\ I':J.

IV SELF-ORGANIZING ECONOMY

The Edge City Model 76

A Central Place Model 88
Simon's Urban Growth Model 92
9 Concluding Thoughts 99
10 Appendix: The Evolution of Central Places 101
Urban Morphogenesis: The Edge City Model 101
A Central Place Model 106
References 117
Index 119 Preface

Every once in a while the world economy plunges into a severe

recession. Some of these international slumps appear to be caused
by particular noneconomic events: wars, disruptions of oil supply.
Others, however, have no obvious cause - and their global scale is
hard to explain in terms of the conventionally measured linkages
among national economies.
Every once in a (much longer) while, the paleontologists tell us,
the world experiences mass extinctions that wipe out most of the
extant species. Some extinctions appear to have external causes, like
the comet whose impact coincided with the demise of the
dinosaurs. Others, however, have not been tied to any obvious
cause. Some theorists who simulate evolution on their computers
claim that this is as it should be: their models predict occasional,
spontaneous mass extinctions even in the absence of any external
shocks to the system.
Can these two paragraphs have anything to do with each other?
Is there a sense in which a global slump is something like a mass
extinction?
Here is another parallel. When you look at the sizes (however
measured) of many enormously complex physical or biological phe-
nomena, the distribution of those sizes for some reason turns out to
be well described by a very simple power law: the number of objects
(earthquakes, meteorites, species, and - perhaps - extinctions)
whose size exceeds S is proportional to s-a, where a is not only a
mystery parameter but often turns out, weirdly, to be a round num-
ber, like 1 or 2. Among the most spectacular examples of a power
law, however, is one that involves economics rather than physical
science: the size distribution of cities. In the United States, the
"
vi SELF-ORGANIZING ECONOMY

number of cities whose population exceeds S is, simply, propor-

tional to l/S: there are 40 cities with more than a million people, 20
with more than 2 million, and 9 (Houston fell a bit short) with
more than 4 million!
Social scientists are normally suspicious of people who want to PART
import concepts from physical or biological science, and with good
reason: the history of such efforts, from social Darwinism to systems
dynamics, has been little short of disastrous. Nonetheless, this time
-----1-----
things may be different; there is a genuinely interesting interdisci-
plinary movement of which economics ought to be a part. Embryos, Earthquakes, and Economics
In the last few years the concept of self-organizing systems - of com-
plex systems in which randomness and chaos seem spontaneously to
evolve into unexpected order - has become an increasingly influ-
ential idea that links together researchers in many fields, from arti-
ficial intelligence to chemistry, from evolution to geology. For Is an economic slump like a hurricane, or is it more like an earth-
whatever reason, however, this movement has so far largely passed quake? Is a growing city like an embryo, or is it more like a
economic theory by. It is time to see how the new ideas can use meteorite?
fully be applied to that immensely complex, but indisputably self- These questions may sound like children's riddles, to which the
organizing system we call the economy. answer is some kind of bad pun; or maybe they sound like Zen
In this book I try to show how models of self-organization can be koans along the lines of "What is the sound of one hand clapping?"
applied to many economic phenomena - how the principle of in which the absurdity of the question is meant to jolt the listener
"order from instability," which explains the growth of hurricanes into a higher state of awareness. But my aim is neither to tell jokes
and embryos, can also explain the formation of cities and business nor to help you achieve an enlightenment that transcends rationali-
cycles; how the principle of "order from random growth" can ty. On the contrary, in the course of this book I hope to convince
explain the strangely simple rules that describe the sizes of earth- you that these are perfectly reasonable questions to ask, that there
quakes, meteorites, and metropolitan areas. I believe that the ideas may well be a more than poetic sense in which a deepening slump
of self-organization theory can add substantially to our understand- resembles an emerging hurricane, in which a growing city is quite a
ing of the economy; whatever their ultimate usefulness, these ideas lot like a developing embryo.
are very exciting, and playing around with them is tremendous fun. In making these kinds of analogies I am not, of course, being
Finally, a note on style. This book began as the Mitsui lectures, completely original. There is a broad and growing interdisciplinary
which I gave at the University of Birmingham in March 1994. movement in the physical and biological sciences - often referred to
Although I have not tried to maintain the lecture format, I have as the study of complexity - that looks for exactly such parallels
allowed myself to retain some of the license usually granted in such between seemingly disparate phenomena. For example, in the
a lecture series: this book is written in an informal style and contains course of my preparations for this book I ran across an article on
more than a few wild speculations. Nonetheless, I have tried to get "percolation theory" that casually listed 15 areas to which the basic
things right when I can; and I am grateful for discussions with par- approach applied, ranging from atomic nuclei to galaxies. But so far
ticipants in seminars at Birmingham, UCLA, Chicago, and this movement has largely passed economics by.
Stanford that helped me correct some serious errors. In particular, I Actually, let me quality that statement. People who write books
would like to thank Mike Woodford at Chicago for pushing me to or convene conferences on complexity almost invariably assert that
test a pet hypothesis to its well-deserved destruction. the emerging field will make great contributions to the study of
2 SELF-ORGANIZING ECONOMY Embryos, Earthquakes, and Economics 3

that immensely complicated system we call the economy. Indeed, physical scientists at understanding the importance of feedback in
the Santa Fe Institute, one of the hotbeds of this style of research, complex systems.
was initially funded by Citibank largely because John Reed, the A second definition, by Philip Anderson, the Nobel laureate
bank's CEO, hoped that research on complex systems would physicist who may perhaps be regarded as the father of the field, is
improve economic forecasting. For whatever reason, however, the that complexity is the science of "emergence." That is, it is about
authors of articles and books on complexity almost never talk to how large interacting ensembles - where the units may be water
serious economists or read what serious economists write; as a molecules, neurons, magnetic dipoles, or consumers - exhibit col-
result, claims about the applicability of the new ideas to economics lective behavior that is very different from anything you might have
are usually coupled with statements about how economies work expected from simply scaling up the behavior of the individual
(and what economists know) that seem so ill-informed as to make units. (The behavior may also be oddly similar to that of ensembles
any economist who happens to encounter them dismiss the whole of otherwise very different units: collections of neurons may behave
enterprise. a lot like collections of magnetic dipoles.) Anderson's prime exam-
But it does not have to be that way. What I am going to claim, ple of an emergent property is the liquidness of water, which is in
and I hope demonstrate, in these lectures is that some of the ideas no sense an extrapolation of some primordial liquidness of individ-
ual water molecules.
that come out of the interdisciplinary study of complex systems -
the attempt to find common principles that apply across a wide Here again, however, we have a definition that sounds like what
variety of scientific fields, from neuroscience to condensed matter economists already understand pretty well. When Adam Smith
physics - are, in fact, useful in economics as well. That is, you can wrote of the way that markets lead their participants, "as if by an
understand and respect the economic theory we already have and invisible hand," to outcomes that nobody intended, what was he
still find ways both to improve it and to build bridges to other fields describing but an emergent property? And examples of emergence
by taking into account the ideas of these interdisciplinary theorists. abound in economic theory - we need only note the way that com-
Before I go any further, I had better explain a little better what I petitive markets, in which each individual is striving only for his or
am talking about. What is this interdisciplinary effort that I have her own profit, act as if the participants were collectively trying to
been alluding to, and how can it teach economists anything new? maximize the sum of consumer and producer surplus, concepts of
Many of the people who think that earthquakes, embryos, cycles, which they are generally unaware. (Nor is this only theory: experi-
and cities have all got something to do with one another describe mental markets in which the participants are assigned payoffs and
their field as the study of "complexity," based on the insight that then make bids for and offers of units of a notional commodity do,
complicated feedback systems have surprising properties. If that in practice, come very close to maximizing aggregate surplus, even
were all there were to the program, however, economists would be though the participants not only are not trying to achieve that
entitled to change the channel. If economists do understand one objective, they do not even know what each others' payoffs are.)
thing much better than the lay public, it is the sheer complexity of There is, however, a third definition of this field that does not
the economic system and the importance of feedbacks. After all, sound like something that economists already do, or at least not
what is general equilibrium theory but a formalization of the propo- what they do routinely. This is the view that what links the study of
sition that everything in the economy affects everything else, in at embryos and hurricanes, of magnetic materials and collections of
least two ways? If you have ever tried to explain to a roomful of neurons, is that they are all self-organizing systems: systems that, even
engineers why higher manufacturing productivity will probably when they start from an almost homogeneous or almost random
reduce, not increase, manufacturing employment - and why high- state, spontaneously form large-scale patterns. One day the air over
er productivity will not necessarily reduce the trade deficit - you a particular patch of tropical ocean is no different in behavior from
quickly realize that economists are better, not worse, than most the air over any other patch; maybe the pressure is a bit lower, but
4 SELF-ORGANIZING ECONOMY Embryos, Earthquakes, and Economics 5

the difference is nothing dramatic. Over the course of the next few business cycle: the pulses of expansion and contraction around a rel-
days, however, that slight dip in pressure becomes magnified atively stable long-run trend. (No, I don't believe in Kondratieff
through a process of self-reinforcement: rising air pulls water vapor waves.) Some recessions and recoveries are clearly set offby specif-
up to an altitude at which it condenses, releasing heat that reduces ic, essentially exogenous events like oil crises. Over the long sweep
the pressure further and makes more air rise, until that particular of history, however, most booms and slumps have had no obvious
piece of the atmosphere has become a huge, spinning vortex. Early external cause. Most notably, the mother of all economic slumps,
in the process of growth an embryo is a collection of nearly identi- the contraction from 1929 to 1933, came as it were out of a clear
cal cells, but (or at least so many biologists believe) these cells com- blue sky. But then, so do hurricanes.
municate with each other through subtle chemical signals that rein- In fact, let me explain why a slump may be quite a lot like a hur-
force and inhibit each other, leading to the "decision" of some cells ricane. A hurricane is a self-reinforcing process, in which an updraft t
"

to become parts of a wing, others parts of a leg. that pulls water vapor to a level at which it condenses thereby releas-
Is the economy a self-organizing system in this sense? Of course es heat, which in turn reinforces the updraft. A slump is also self-
it is. Think about a metropolitan area - even, or rather especially, a reinforcing: falling output causes firms to slash their investment and
modern metropolitan area like greater Los Angeles, with no clearly consumers to reduce their spending, thereby reducing output still
defined center. Is an urban sprawl like LA a homogeneous, undif- further. But wait, there is more. A hurricane cannot go on forever.
ferentiated mass? No - it is a patchwork of areas of very distinct The fuel for its violence is the water vapor that exists in relative
character, ranging from Koreatown to Hollywood, Watts to Beverly abundance above a warm tropical ocean. But the hurricane itself
Hills. And it contains (according to the recent book Edge City by cools that ocean surface - indeed, in a way the whole point of trop-
Joel Garreau) no fewer than 16 "edge cities," newly emerged busi- ical storms is that they are Nature's way of transferring solar energy
ness centers, each of which includes at least 5 million square feet of from the tropics to higher latitudes. And so hurricanes, even though
office space and 20,000 workers, where the low-rise sprawl sud- they are self-reinforcing in the short run, are self-limiting in the
denly gives way to tall buildings and multistory parking garages. long run. What about an economic slump? Well, it leads to falling
What is so striking about this differentiation is that it is so indepen- prices, or at least disinflation, which gradually increases the real
dent of physical geography: there are no rivers to set boundaries, no money supply. The low or negative net investment in a deep slump
big downtown to define a gradient of accessibility. (OK, the beach- may also lead to a growing backlog of potentially profitable projects.
es and the freeways create a bit of exogenous structure, but grant me And so we may argue that an economic slump is also a self-limiting
a little license.) The strong organization of space within metro Los process, even if there is no deliberate or effective government poli-
Angeles is clearly something that has emerged, not because of any cy to get the economy moving again.
inherent qualities of different sites, but rather through self-reinforc- By the way, Americans who watch the weather news know that
ing processes: Koreans move to Koreatown to be with Koreans, hurricanes sustain themselves by moving, gaining strength when
beautiful people move to Beverly Hills to be with other beautiful they pass over fresh tracts of warm water; if this has any analogy in
people. And when an observer like Garreau traces the development business cycles, I am unaware of it. But I did not claim that the par-
of edge cities, he is immediately drawn to metaphors like "sponta- allel was perfect. And you may argue that my description of what
neous combustion" and "critical mass" - clear indicators that he is goes on in a typical recession is not too accurate - or maybe there is
trying to describe a self-organizing system. no such thing as a "typical" recession, in which case recessions are
I have started with the way that cities spontaneously evolve a pat- more like earthquakes than like hurricanes. But we shall get to that
tern of sharply distinct districts, because the economic self-organi- in Part II.
zation of space is something we can all relate to our immediate While I am making asides, let me also take the opportunity to
experience. But I would assert that there are processes in the econ- make another point: self-organization is not necessarily, or even
omy that produce temporal self-organization as well. I refer to the presumptively, a good thing. I think it is fair to accuse many of the
 6 SELF-ORGANIZING ECONOMY 7
Embryos, Earthquakes, and Economics

writers on complexity, especially but not only the more popular economy, their work has been neglected by the profession (and is
ones, of falling into this fallacy. Book titles like Order out if Chaos (by completely unknown outside of it). Nonlinear business cycle theo-
Nobel laureate Ilya Prigogine and I. Stenger, but with a Foreword ry, for example, was technically far ahead of its time and makes the
by, believe it or not, Alvin Toffler), or Complexity: Life at the Edge if later efforts to shoehorn "catastrophe theory" into economics look
Chaos (by R. Lewin) come perilously close to making self-organi- primitive. But who remembers it? (I may be the only economist in
zation a kind of mystical goal. Even Stuart Kauffman, whose The my generation who has even heard of it.) Thomas Schelling wrote a
Origins if Order is a serious and stimulating tract (if we ignore the remarkable essay on the dynamics of segregation in his wonderful
inevitably careless and ill-informed section on economics), talks far 1978 book Micromotives and Macrobehavior, but the book had little
too casually about coevolutionary systems that maximize "average impact at the time and is still underappreciated.
fitness" - a surely meaningless concept when the fitness of each I think that I can explain why the pioneers of self-organization in
species is defined at least partly in terms of how well it copes with economics have been either neglected or forgotten, and I will talk
competition and predation from others. Luckily, if we are at all seri- about those reasons later in these lectures. For now, however, let us
ous about the economics of self-organization we immediately real- put intellectual history aside and turn to doing some economics.
ize that no value judgment is implied. An economy with a strong Here is my plan of action. In this first part, I am going to do a
business cycle exhibits more temporal self-organization than an quick run-through of stories - I do not want to dignify them by
economy that grows smoothly, but most of us would rather live in calling them models - about ways in which the economy organizes
the latter. A city whose racially integrated communities unravel, itself in space. (I start with spatial self-organization for two main
producing huge segregated domains, becomes more spatially orga- reasons. First, as an empirical matter I am on more solid ground: the
nized, but not better, in the process. Self-organization is something self-organizing spatial character of the economy is obvious to every-
we observe and try to understand, not necessarily something we one, although not usually under that name. Second, I know what I
want. believe about spatial economics, but I am still fairly agnostic about
I have asserted that the study of self-organization is something the macroeconomics of business cycles.) Along the way I shall
that economists do not do, or at least do not do routinely. What I describe two broad, and seemingly paradoxical, principles of self-
have just said about economic slumps is, however, not at all new or organization that are suggested by these stories and that are also
original. On the contrary, I have just given you a loose version of common in much of the literature on self-organization in other
the nonlinear business cycle literature that flourished in the 1940s fields. Just to whet your appetite, let me give these principles names:
and 1950s, with contributors including economists of the stature of "order from instability" and "order from random growth."
John Hicks, Richard Goodwin, and James Tobin. For that matter, In the second part, I will start by showing how these principles
urban economists have hardly been unaware of the self-organization may apply not only to space but to time: in particular, how we
of metropolitan areas, and there have been some notable efforts to might think of the business cycle as temporal self-organization.
model the creation of urban subcenters, as edge cities are colorless- Then we shall come back to the spatial economy. I will describe two
ly known among the professionals. (I have in mind particularly the more or less full-fledged models that illustrate the principle of order
pioneering work of Fujita and Ogawa.) from instability and one that illustrates the principle of order from
And yet I think it is fair to say that few economists have explicit- random growth. Even there I will go light on the equations, stress-
ly realized that they were trying to model self-organization; few ing an intuitive explanation. But there will be a technical appendix,
have tried to draw the parallel between self-organization in space which derives the results in all the gory detail I can manage (which
and self-organization in time; few have tried to use some of the is not much).
techniques for understanding self-organization that have evolved in So let us begin our tour with a look at the way that economies
other fields. Furthermore, although some deeply insightful thinkers organize space, particularly the space within metropolitan areas.
more or less consciously have written about self-organization in the
-----1-----
Self-Organization in Space

THE VON THUNEN-MILLS MODEL

How do economists routinely deal with the question of how the
economy organizes its use of space? The short answer is that most-
ly they do not deal with the question at all. Indeed, there is some-
thing strange about the way that most of our profession neglects
anything having to do with where economic activities happen. For
example, the very popular (and almost 900-page-Iong) economics
principles textbook by William Baumol and Alan Blinder contains
not a single reference to "cities," "location," or "space" in its index.
The rival, 1100-page text by Joseph Stiglitz does contain one refer-
ence to cities, which turns out to occur in a brief discussion of
rural-urban migration in less-developed countries. Considering
how much time most people spend in traffic jams, how many for-
tunes are made and lost in real estate, the way that we turn a blind
eye to spatial economics is little short of eerie. I strongly suspect that
this neglect is closely related to the theme of this book: as a profes-
sion we are implicitly aware that to understand cities and spatial
economics generally we must cope with issues of self-organization
and that rather than face what seem to be intractable issues we sim-
ply avert our gaze.
Anyway, when we do deal with the question of the organization
of space, as urban economists at least must, we generally turn to a
class of models pioneered in the early 19th century by von Thiinen.
Many of you are probably familiar with the von Thiinen model, but
I want to run through it briefly to make a couple of points about
complexity, emergence, and self-organization.

9
10 SELF-ORGANIZING ECONOMY Self-Organization in Space 11

Bid rent
Von Thiinen envisaged an isolated town supplied by farmers in
the surrounding countryside. He supposed that crops differ in both
their yield per acre and their transportation costs, as well as allow- Vegetables
ing for the possibility that each crop could be produced with differ-
ent intensities of cultivation. And he asked two questions that might
Wheat
seem to be very different: How should the land around the town be
allocated to minimize the combined costs of producing and trans-
Cattle
porting a given supply of food to the town? How will the land
actually be allocated if there is an unplanned competition among
farmers and landowners, with each individual acting in his or her
perceived self-interest? Distance
We all know the answer to the second question. Competition from
among the farmers will lead to a gradient ofland rents that declines center
from a maximum at the town to zero at the outermost limit of cul-
tivation. Each farmer will be faced with a tradeoff between land
rents and transportation costs; since transportation costs and yields
differ among crops, the result will be a pattern of concentric rings
of production. In equilibrium the land rent gradient must be such
as to induce farmers to grow just enough of each crop to meet the
demand, and it turns out that this condition together with the con-
dition that rents be zero for the outermost farmer suffices to fully
determine the outcome.
Figure 1.1 illustrates schematically the typical outcome of a von
Thiinen model. The upper part of the figure shows the equilibrium
"bid-rent" curves, the rent that farmers would be willing to pay at Figure 1.1 The Von Thiinen Model. Competition for land around a
any given distance from the town, for three crops. The heavy line, town leads to the emergence of concentric rings of production
the envelope of the bid-rent curves, defines the rent gradient.
Along each of the three segments of that line growers of one of the trivial to determine either what should be done or what will happen
crops are willing to pay more for land than the others. Thus one gets in an unplanned market.
concentric rings of cultivation, with a quarter section of the layout Suppose that you instead define the subject as Anderson does, as
shown in the bottom half of the figure. a study of emergence. Then surely the von Thiinen model qualifies.
It is worth thinking about this outcome for a moment in terms of At the most obvious level, the concentric ring pattern is hardly
the various definitions that have been offered of what complexity something that is immanent in the motives of the farmers. Indeed,
theory is all about. If it is really just about complexity, von Thiinen the concentric rings will emerge even if no farmer knows what any-
models are probably complicated enough to qualify. After all, the one else is growing, so that nobody is aware that the rings are there.
problem of which crops to grow where is not that easy: by allocat- Even more striking is the result that economics has trained us to
ing an acre of land near the city to some one crop, you indirectly expect but that remains startling (and implausible) to most
affect the costs of delivering all other crops, because you force them noneconomists: the unplanned outcome is efficient, is indeed the
to be grown further away. Except in the case where there is no pos- same as the optimal plan. More specifically, unplanned competition
sibility of varying the land intensity of cultivation, it is by no means will allocate crops to land in a way that minimizes the total com-
12 SELF-ORGANIZING ECONOMY Self-Organization in Space 13

bined cost of producing and transporting the crops - not including lem: cities ,do not look like that, and they look less like that with
the land rent. This is surely as nice an example of emergence as you every passing decade.
could want. All farmers are trying to maximize their income and are Remember what I said about greater Los Angeles, with its 16
therefore very much concerned with land rents, yet their collective edge cities far overshadowing its two traditional downtowns.
behavior minimizes a function in which land rents do not appear. Sophisticated people used to turn up their noses at California cities,
On two of the criteria that have been used to define this enter- with their lack of clearly defined centers - Gertrude Stein declared
prise called complexity, then, the von Thiinen model qualifies. And of Oakland that "there's no there there." She was wrong, of course:
yet I doubt that many complexity theorists would agree that this was there are lots of quite distinct theres there, just no one big there you
the kind of economics they had in mind. Avon Thiinen model is can call the center. And increasingly that is the way all our cities
not like a hurricane, or an embryo, or a neural network. Why? look. I wrote the draft of this book in Palo Alto, which is part of the
Because it is not self-organizing. The concentric rings of production San Francisco metropolitan area. San Francisco proper is a compact
form around a town whose existence is simply assumed. That does city, which still epitomizes a certain kind of urbanity. But Palo Alto
not make it a bad model, but it does make it a limited one. If your has little to do with that center. One thinks of oneself as being, not
question is not simply how land use is determined given a pre-exist- a San Franciscan, but a resident of Silicon Valley; one reads the San
ing town, but rather how land use is determined when the location Jose Mercury News rather than the San Francisco Chronicle (it is a bet-
of the town or towns - indeed, their number and size - are them- ter paper anyway), and people who live in Palo Alto are much more
selves endogenous, the von Thiinen model offers no help. likely to commute to the edge city in Sunnyvale than to the vicini-
This limitation became painfully clear when the von Thiinen ty of the Golden Gate. The monocentric city model pictures a met-
model was presented in a new form in the 1960s by Edwin Mills, ropolitan area as something like a slice from an onion, with rings
who substituted commuters for farmers and the central business arrayed around a single center. The reality of all large metropolitan
district for the isolated town, to arrive at his now classic model of a areas in the United States today, even those like New York or
monocentric city. Chicago that still have huge, vital downtown office districts, is that
Again, you are probably familiar with the basic idea if not the they are less like an onion slice and more like Jack Horner's plum
details. We imagine a city populated by commuters who want land pudding, in which edge cities correspond to the plums.
to live on but must work in a central business district; there may be The problem of modeling the structure of the modern polycen-
different types of people who vary in the value they place on the tric urban area is in large part one of eXplaining the formation and
time they spend commuting, the amount of land they want to live location of these plums. That is, we cannot avoid the problem of
on, their willingness to substitute fancy houses for large backyards, understanding the urban area's self-organization. Part of the process
and so on. As in the von Thiinen model, competition establishes a by which that self-organization takes place is, of course, a competi-
land rent gradient, which sorts people out into a structure of con- tion for land that establishes a land rent surface across the urban
centric rings; and as in that model, decentralized location choices landscape; in that sense the von Thiinen-Mills approach remains
lead, through nobody's intention, to an efficient outcome. essential. But it is at best half the story - and arguably the less inter-
Mills's 1967 paper introducing the mono centric city model esting half
launched a huge theoretical and empirical literature. And yet in the
end that literature proved rather sterile. Part of the problem was CENTRAL PLACE THEORY
purely aesthetic: a model that simply assumes that there is a central
business district is deeply unsatisfYing if you want to explain cities as Economics as we know it is largely, though not entirely, an Anglo-
opposed to describing them. But urban economists might have Saxon tradition. Location theory, however, was long a German tra-
swallowed their disappointment if it had not been for another prob- dition, containing at least three streams. One stream follows from
14 SELF-ORGANIZING ECONOMY 15
Self-Organization in Space

Thiinen analysis ofland rent and land use, which we have The answer, I believe, is that any economist who thinks hard
lussed. A second stream, associated with Alfred Weber and about central place theory realizes that it does not quite hang
ers, focused on the issue of optimal plant location; that lit- together as an economic model. What do we look for in an eco-
"11play no role in my discussion. But there is a third tradi- nomic model- or to put it differently, what constitutes an "expla-
ich at first sight seems to offer an answer to the problem of nation" from the point of view of economists? I do not think you
If-organization: the central place theory of Christaller and can sum it up better than Thomas Schelling did in the title of his
book, Micromotives and Macrobehavior. We feel that we have really
asic ideas of central place theory seem powerfully intuitive. managed to shed light on a phenomenon when we show how that
a featureless plain, inhabited by an evenly spread popula- phenomenon, the "macrobehavior," emerges (there is one of those
"rmers. Imagine also that there are some activities that serve words again) from the interaction of decisions by individual families
ers but cannot be evenly spread because they are subject to or firms; the most satisfying models are those in which the emer-
o[llies of scale - manufacturing, administration, and so on. gent behavior is most surprising given the "micromotives" of the
seems obvious that the tradeoff between scale economies players. What is therefore deeply disappointing about central place
r~[lllsportation costs will lead to the emergence of a lattice of theory is that it gives no account along these lines. Losch showed
places," each serving the surrounding farmers. that a hexagonal lattice is efficient; he did not show that it would
bvious, but still intuitively persuasive once presented, are tend to emerge out of any decentralized process. Christaller sug-
gested the plausibility of a hierarchical structure; he gave no account
en~ements
",I~ndproducedintroduced
evidenceby inChristaller and central
support, that Losch. places
Christaller
form of how individual actions would produce such a hierarchy (or even
hy: there are a large number of market towns, every group sustain one once it had been somehow created).
t towns is focused on a larger administrative center (which What, then, is central place theory? It is not a causal model. It is
arket town), and so on. Losch pointed out that if a lattice probably best to think of it as a classification scheme, a way of orga-
o minimize transportation costs for a given density of cen- nizing our perceptions and our data. Seen in that light, it is a wor-
s, the market areas must be hexagonal. And thus every thy enterprise indeed. After all, in the physical and biological sci-
on location theory contains a picture of an idealized cen- ences classification schemes have repeatedly served as the basis for
system in which a hierarchy of central places occupies a set great insights - think of the Linnaean classification of species or the
hexagons. periodic table. The point, however, is that classification schemes are
iginal central place theory story applied to towns serving a only a step on the way: they tell you what, but they do not tell you
keto But it is obvious that a similar story can be applied to why. So central place theory is a description but not really an expla-
districts within a metropolitan area. Small neighborhood nation of self-organization.
districts are scattered across the basins that surround larg-
ts with more specialized stores, all eventually centering on SCHELLING'S SEGREGATION MODEL
town with its great department stores and high-end bou-
Illdeed, the hierarchical image is so natural that it is hard to The two approaches to the organization of space just described do
.cribing things that way. not quite make it: the von Thiinen-Mills approach, which is a
\MHyis central place theory not a standard part of the econo- model of spatial organization but not of self-organization, and cen-
lkit? Why do the introductory texts find no room for a tral place theory, which is a useful classification scheme but not a
causal model. Well, enough of frustration: now let us turn to some
Inthat uses
centers, thencentral
uses theplace
von theory to explain
Thiinen-Mills whytowe
model have
explain approaches that do help explain the self-organization of spatial
n ofland use around those centers? econormes.
16 SELF-ORGANIZING ECONOMY Self-Organization in Space 17

I have already expressed my admiration for the work of Thomas residential patterns tend to be unstable in the face of random pertur-
Schelling. He is best known for his famous treatise on non-zero- bations. Second, even if the concerns of individuals are very local-
sum games, The Strategy of Conflict. But I think that his best book is they care about only their immediate neighbors - what emerge are
Micromotives and Macrobehavior, an underappreciated classic that had a large segregated neighborhoods. Thus Schelling derived, without
deep impact on me when I first read it as a wet-behind-the-ears any fanfare, a theme of many writers on complexity: local, short-
assistant professor. The first chapter of the book is surely the best range interactions can create large-scale structure.
essay on what economic analysis is about, on the nature of eco- Characteristically, he made these points with a few simple exam-
nomic reasoning, that has ever been written. And the two chapters ples rather than a fully worked-out mathematical model. It is possi-
on "sorting and mixing" are a wonderful introduction to the idea of ble to set up such a model, and I am sure that one can even prove
self-organization in economics. theorems about it. I shall reserve my limited mathematical firepow-
If there is any flaw in Schelling's work, it is that he is so clear a er, however, for my own models. For our current purposes,
thinker that he can often reach deep conclusions with almost no vis- Schelling's approach will be good enough.
ible technical apparatus and so graceful a writer that he can often Let us, then, imagine a "city" that consists of a number ofloca-
make these conclusions seem intuitively obvious. These virtues, I tions laid out in a square lattice, like a chessboard. For illustration,
believe, have worked against him. As an amateur anthropologist indeed, let us actually use an 8-by-8 chessboard-size lattice,
who has long studied that peculiar culture known as academic eco- although the same principles would apply to a much larger domain.
nomics, it seems to me that an economic idea flourishes best if it is And let us suppose that there are two kinds of people - call them
expressed in a rather technical way, even if the technical difficulty is black and white, but they could represent any kind of racial or cul-
largely spurious.! After all, a teacher wants something to do at the tural groups, or for that matter different types of businesses that tend
blackboard, and a clever student wants something on which to to repel each others' customers, like boutiques and auto supply
demonstrate his or her cleverness. If a deep idea is conveyed with stores.
simple examples and elegant parables, rather than with hard math, We now assume that blacks and whites care about the color of
it tends to get ignored. their immediate neighbors. (That means the abutting squares on the
In any case, however, in Micromotives and Macrobehavior Schelling chessboard.) Their preferences are assumed to be not so much a lik-
presented a simple yet profound model of segregation. The basic ing for neighbors the same color as a fear of being isolated.
idea sounds trivial: segregation results when people prefer not to Specifically, Schelling suggested the following rule: an individual
have too many neighbors who are different from themselves. But with one neighbor will try to move if that neighbor is a different
Schelling made two much less obvious points. First, mild prefer- color; one with two neighbors wants at least one of them to be the
ences about the color or culture of your neighbors - preferences same color; one with three to five neighbors wants at least two to be
that seem on the face of them to be consistent with maintaining an his or her color; and one with six to eight neighbors wants at least
integrated residential pattern - in fact typically lead to a high degree three of them to be like him or her.2
of segregation. Why? Because, even when people have mild prefer- These preferences are consistent with an integrated residential
ences of the form "I don't mind having some neighbors of a differ- pattern. Consider Figure 1.2, in which # and @ signifY the two dif-
ent color, as long as I'm not too much in the minority," integrated ferent groups. Here we have managed to place 60 individuals in a
completely integrated pattern, without violating anyone's con-
straints. That is, complete integration is an equilibrium.
1. You should not, by the way, conclude that I prefer "common sense" to aca-
demic discourse. In the real world of affairs, an economic idea is most likely to
succeed if it is naively wrong - only then can it appeal to the prejudices of impor- 2. An equivalent statement of the rule is that each individual requires that at
tant people. least 37 percent of the neighbors be of his or her own type.
 18 SELF-ORGANIZING ECONOMY 19
Self-Organization in Space

#
@ ##
@
@ #
@ #
@ #
#
@ #@ @
# @

Figure 1.2 An Integrated City. Even if people insist that a minimum Figure 1.3 Perturbing the Equilibrium. If the pattern is given some
fractionof their neighborsresemblethemselves,it is possibleto random scrambling, some individuals are no longer content
create an equilibriumresidentialpattern that is highlyintegrat- with their location.

But are we likely to get a result like this in practice? No, argued It is pretty obvious that in Figure 1.3 some of the people are no
Schelling. Although he did not put it quite this way, an equilibrium longer content with their locations and will move. When they
like the one shown in Figure 1.2 will be unstable with respect to move, however, they will in turn often make those who stay unhap-
some random shuffling and will therefore tend to unravel. py - either by depriving same-color individuals of neighbors or by
To show this, Schelling took the pattern in Figure 1.2 and messed shifting the balance against new neighbors of a different color. So a
it up a bit. (His instinct was unerring. My first inclination would chain reaction begins. To simulate that chain reaction on a comput-
have been to start from a completely random allocation of people to er you would need to specify the order in which people move and
locations. It turns out, however, that self-organizing spatial systems how they pick among available locations. If you are doing it by
yield the greatest order when the initial condition is a small pertur- hand, you can be looser about it - it does not matter much. Like
bation away from the unstable integrated or flat equilibrium. We will Schelling, I did this by hand (using a spreadsheet instead of pennies
see why in the second part.) Specifically, he extracted 20 individuals and dimes on a real chessboard) and just watched the structure
at random, both disrupting the pattern and freeing up some room for evolve. When things settle down, you get Figure 1.4.
discontented individuals to move, then disrupted the pattern a bit Guess what: even though individuals are tolerant enough to
more by filling five empty squares at random with #s or @s. Figure accept an integrated pattern, they end up with more or less total
1.3 shows the result when I and my random number generator do it. segregation. And, even though individuals care about only their
 r··

20 SELF-ORGANIZING ECONOMY Self-Organization in Space 21

As you can tell, I love Schelling's model. It is a perfect example of

# #@
@ ## @ @ how to wear your sophistication lightly; the way that big conclu-
sions are derived from playing with coins on a chessboard has a spe-
@ @ cial charm. As a model of the self-organization of urban areas, how-
@ ever, it has some limitations. Some of them are problems that I will
not try to fix in this book; for example, there is no explicit handling
of the market for land, which surely plays a crucial role in mediat-
ing the choices of households and firms about where to locate. One
problem that I do want to deal with, however, is that Schelling's
model is too one-sided: it tells us why birds of a feather flock
together but offers no reason why there should ever be more than
two flocks.
To see what I mean, ask what would happen if I replaced my
8-by-8 chessboard with a much bigger one, say 1000 by 1000. What
would a Schelling-type model predict as an outcome? Well, a huge
literature in physics deals with systems not too different from
Schelling's model, called spin-glasses, and a number of formal results
have been derived; but I have to admit that I have not yet taken the
time to try to master this literature. Still, a few things seem fairly
obvious. The tendency of Schelling's model is always to divide the
whole city into two vast # and @ territories. That does not mean
1.4 A SegregatedCity. The result is a chain reaction, in which that we will necessarily get there: if we start with a random pattern,
eachmoveprovokesother moves;in the end mildconcernsabout the chain reaction of moving households will typically die out at
beingin a localminorityproducea completelysegregatedcity.
some point, leaving the city "frozen" into # and @ domains of vary-
ing sizes. (If we set the parameters right, the size distribution will
cal neighborhood, the whole chessboard gets organized into surely obey a power law - but more about that kind of thing later.)
e area and a black area. We can, however, "melt" these domains if we add a bit of "temper-
and @ really do represent white and black, it is a nasty outcome. ature" to the story, by giving even contented households some small
That, owever, is beside the point for this book (except as a reminder probability of moving just for the hell of it. In that case we will
that . anized does not mean "good"). What I want to emphasize is indeed eventually find ourselves with a two-neighborhood city.
that t is chessboard city has engaged in a process of self-organization. Now that is too strong a result. Let us go back to Los Angeles.
Larg scale order - not a nice order, but order nonetheless - has There are some ways in which the city has a two-zone structure: all
emer ed from a basically disordered initial condition. the beautiful people live in Beverly Hills, all the nonbeautiful peo-
T .s large-scale order emerges because a disordered state, in ple someplace else. But what is so striking about LA compared with
whicl #'s and @'s are evenly mixed, is unstable: scramble it a bit and a traditional city is precisely its multipolarity. There is no dominant
you s . rt a dynamic process that produces segregation. This is there- downtown office district: instead, there are 16 edge cities, spread in
fore y first illustration of the principle of order from instability - a not-too-irregular fashion across the metropolitan area - and,
but d not worry, there will be a lot more. according to Garreau, 8 more in the process of coalescing.

II
 SELF-ORGANIZING ECONOMY Self-Organization in Space 23

So if we want a story about the self-organization of the urban the ends. That is not simply a formal issue: in reality an office com-
l~ndscape that reflects the polycentric, plum-pudding metropolitan plex near the middle of a metropolitan area will be different from
reas we increasingly inhabit, we need a model that spontaneously one near the point where subdivisions begin to be mixed with farm-
roduces not only order but some kind of more or less regular land. Because I do not want to deal with that issue, I need to make
epetitive pattern. Let us see if we can construct one. a silly assumption: my city is not only one dimensional but circular,
so there are no ends.3
A second issue is the treatment of land and land rents. Modern
DGE CITY DYNAMICS
urban economists are generally willing to play games with geome-
Iina moment I will describe an approach that, it seems to me, sheds try, but they tend to be quite fanatical about explicitly modeling
onsiderable light on how office districts - edge cities - may emerge land rent - when I presented a model without land rents to one
ih a polycentric metropolitan area. I believe and hope that the group, a disgruntled urban theorist told me that as far as he was con-
pproach will provide a useful metaphor for the emergence of spa- cerned, urban economics was essentially about land rent. It is obvi-
tlal structure in a variety of other contexts. But any experienced ous where that attitude comes from: in the von Thiinen-Mills
conomist knows that in presenting a new approach one must begin model land rent is indeed the crucial ingredient. But I am asking a
with some preemptive excuses for the questionable simplifications different question and will ask you to bear with an approach that
ne has made. So before I even start to describe my framework, let recognizes land scarcity at best in an implicit, reduced-form way.
e offer some justifications for the way it is done. A third issue is the treatment of expectations. I will be telling a
The models that I will describe loosely now and present more dynamic story; it is a story about how edge cities evolve. But I do not
lly in the next part will, not surprisingly, represent a kind of ide- want to worry about forward-looking behavior. To many econo-
mists, raised in an environment of rational expectations theory, that
allization
and of place
central reality.theory
This will
beginnotwith
shock
the anyone: both ofvon
idealization Thiinen
a homoge- by itself disqualifies a model. Most of you probably do not care, but
eous agricultural plain, which never bore much resemblance to for those of you who do, I urge you to put your prejudices on hold.
ality and bears even less now than it did when they wrote. What Finally, I have a personal rule that I am about to break. In gener-
ind of idealization seems legitimate is, however, in the eye of the al, I do not like simply assuming the existence of external
eholder. My guess is that there are at least four ways in which the economies and diseconomies. It often seems too close to assuming
a~proach I am about to present will bother people. your conclusions. (I know of one economist who tried to explain
First, there is the issue of geometry. The German tradition in his work to a group of physicists, one of whom sarcastically said, "So
location theory was cheerful about assuming away rivers, roads, and what you're telling us is that firms agglomerate because of agglom-
ariations in land quality but rigorous about facing up to the conse- eration economies.") Most of my work on economic geography has
uences of the fact that the Earth's surface is two dimensional. It is focused on trying to derive external economies out of the interac-
cfmmon in modern urban economic theory, however, to analyze tions among scale economies, transportation costs, and factor
"long, narrow" cities that are effectively one dimensional (as, for mobility - to make external economies an emergent property. I
ebmple, in Fujita 1988), a simplification that seems reasonable to actually arrived at the approach I am getting to in the context of
e but that would probably have horrified many of the Germans. such a model, one that I shall present in the second part. But it turns
Well, it turns out that for my purposes even the one-dimension- out that it is easiest to explain the basic idea simply by assuming
city is not quite simple enough, because I want to focus on self- external economies, and so I will temporarily suspend my rule.

e t differences among locations but from the internal logic of the

s stem. And even in a one-dimensional city locations are distin- 3. Alternatively, one can imagine a city of infinite extent. Los Angeles may
g0rganizat
ished from
.•ion: one
the emergence
another by of
onestructure
crucial aspect:
that arises
theirnotdistance
from inher-
from approximate this condition.
24 SELF-ORGANIZING ECONOMY Self-Organization in Space 25

shopping mall, because they pull in more potential customers;

Now that the excuses are out of the way, let us get to the sub- it does not like it when stores move into a rival mall 10 miles
stance.
away.)
Imagine a metropolitan area, which we can think of as a homo-
geneous expanse of identical housing developments - except that, And that's all that we need. In any model meeting these criteria,
rather oddly, the population is distributed not in a two-dimension- any initial distribution of business across the landscape, nO matter
al sprawl but in a narrow ring, and travel is possible only along that how even (or random), will spontaneously organize itself into a pat-
ring's circumference. And suppose that there is some business activ- tern with multiple, clearly separated business centers.
ity - it could be office work, it could be retailing - that depends On Does this proposition sound obvious? Maybe so (although I did
the spread-out population both as a market and as a source oflabor. not have it clear in my OWnmind until after I started playing with
Let us also, unobjectionably, suppose that the decisions by busi- mathematical models). But there is more. For a wide variety of spe-
nesses about where to locate are interdependent. That is, the desir- cific models in which criteria 1 and 2 are met, any initial distribu-
ability of anyone site as a business location depends On where all of tion of business across the landscape will evolve not merely into a
the other businesses are located. And we may also safely suppose pattern with several business centers but into a pattern in which
that businesses migrate over time from less to more desirable sites. these centers are roughly evenly spaced, with a characteristic dis-
Clearly the dynamics of this process depend On the nature of this tance that depends On the details and parameters of the model but
interdependence. One can imagine two general sorts of inter depen- not On the initial distribution. And the smoother is the initial spa-
dence. On the One hand, businesses might dislike having other tial distribution of the businesses, the more even their eventual
businesses nearby, because they compete for customers, workers, or spacmg.
land. Call these considerations centrifugal forces, forces that promote
dispersion of business. On the other hand, businesses might like to
have other businesses close, because they attract customers to the
I/) _
area or help support a greater variety of local services.4 Call these c:i r___________.

centripetal forces, forces that tend to make businesses clump togeth-

er. If only centrifugal forces existed, businesses would spread them- 't
c:i
selves evenly across the landscape. If there were only centripetal CIl
l.-
forces, they would rush together into One big clump. llS
.J::
"l
"
But what explains the polycentric, plum-pudding pattern of the II>

II>
c:::.

II>

modern metropolis? In general, a model that would explain this pat- CIl
c: '"
tern must meet two criteria: ::l c)
"Vi

cO

Criterion 1. There must be a tension between centripetal and -

c)
centrifugal forces, with neither too strong.

Criterion 2. The range of the centripetal forces must be shorter -"-~,

-""
than that of the centrifugal forces: businesses must like to have
other businesses nearby, but dislike having them a little way "oJ"

away. (A specialty store likes it when other stores move into its "l/l'17e

4. Joel Garreau writes: "Five million square feet is a point of spontaneous

combustion. It turns out to be exactly enough to support the building of a luxu- Figure 1.5 The Evolution of Edge Cities. An ititially almost uniform
ry hotel. It causes secondary explosions; businessesbegin to flock to the location distribution of business accross the landscape evolves sponta-
to serve the businessesalready there." neously into a higWy structured metropolis with two concen-
trated business districts.
26 SELF-ORGANIZING ECONOMY Self-Organization in Space 27

At this point I had better show you an illustration. It turns out,

not too surprisingly, that computer simulation - and computer
graphics - are an invaluable aid to thinking about self-organizing o10)
systems. Figure 1.5 is a sample run of a model of a "city" consisting ci
of 24 locations around a circle. The locations are shown on the X II)
'"
axis; bear in mind that location 24 is next to location 1. I started the ci

run with a fairly but not perfectly even allocation of business across Q)O
\.. '"
these locations,5 then let it evolve according to a rule that caused ~ti
III
businesses to move toward locations that were highly desirable. Ill'"
III -
"Desirability" of a location was both positively and negatively ~ti
'iij
affected by the number of businesses at other locations, with both ::::l
d) -ti
<:>

effects declining with distance but with the positive effects declin-
ing faster than the negative effects. The Y axis of the figure shows o
...•

ti
the passage of time; the Z axis shows the share of the businesses in
each location at each point in time. Thus the figure gives you a sort
of frozen portrait of the whole simulated history.
Let us look a bit at Figure 1.5; the picture contains quite a lot of
information. The right edge of the calculated surface, which repre-
sents the initial geographical distribution of business, is almost a
horizontal line. That is, I have started my "city" off with almost no
spatial organization. But eventually the surface rears up into a pair Figure 1.6 Same Story, Different Parameters. With somewhat differ-
ent parameters, the metropolis evolves four business districts.
of dorsal fins:6 all of the businesses end up in locations 8 and 20.
The city has spontaneously developed a strong spatial structure.
This may not be surprising. But now notice which locations get
the businesses. In a 24-location circular city, locations 8 and 20 are parameters. In this case we getfour business districts, equally spaced
exactly opposite one another. That is not an artifact of the particu- around the circle. The general principle, then, is that the city orga-
lar starting position: if you run the model repeatedly with these nizes itself into a structure with a characteristic distance between
parameters, but with a different initial spatial distribution of busi- business districts; if there are only two, this implies that they face
ness each time, you will consistently get two business concentra- each other.
tions opposite each other. Let us go back to Figure 1.5 and look at one more thing. Look at
By the way, I do not want you to get the impression that there is how the almost flat initial surface evolves over time. Although in
something special about cities with two business districts. Figure 1.6 the end only two widely separated locations end up with business
shows a typical run of the same model with somewhat different concentrations, in the early stages of self-organization it does not
look as if the winning locations are growing at the expense of their
neighbors. On the contrary, initially not only the winning locations
5. Each location received a weight 5 + Ui, where Ui was a location-specific ran- but those nearby grow. The surface seems to undulate, with waves
dom variable between 0 and 1, and s was a "smoothing" parameter, set equal to 5 rising up out of the plain. Only after some time has passed do these
for this run. Then each location was assigned a share of the businesses proportional waves gather themselves into dorsal fins, with the eventual centers
to its weight,
cannibalizing their neighbors. Again, this impression is not unique
6. Robert Gordon of Northwestern University has dubbed Figure 1.5 the '59 to a particular run: it is a consistent feature, whatever the initial dis-
Cadillac model.
tribution.
28 SELF-ORGANIZING ECONOMY Self-Organization in Space 29

It turns out that this image of an undulating surface, with the order, the form of that order is dictated by a sort of principle of
waves growing over time, is the key to understanding the process of maximum instability.
self-organization. To fully explain why will take a little time, and I'll What about the last part of the figures, where the wobbles gath-
reserve that for the second part (and the technical appendix). But let er themselves up into spikes? Well, the formal answer is that every-
me offer a preliminary view. thing that I have said is valid only for a linear approximation to the
Remember that I started the simulation in Figure 1.5 with an model, which breaks down when the spatial distribution of business
almost but not exactly flat distribution of businesses around the cir- gets too uneven. Less formally, in the early stages of self-organiza-
cle. The deviation of that distribution from perfect flatness may be tion the most favored locations can grow by pulling businesses away
represented as an irregular, wobbly line. But it is generally true that, from distant locations. Once there are no more businesses in the
to use technical language, an irregular wobble can be thought of as large gaps between business centers, they can continue to grow only
the sum of many regular wobbles at different frequencies. In partic- by eating their neighbors. The important point, however, is that the
ular, an irregular wobble around a circle can be expressed as the sum locations of the winners are determined in the earlier stage: the dis-
of one line that wobbles once as it goes around the circle, another tance between edge cities is determined by the wavelength of the
that wobbles twice, a third that wobbles three times, and so on. most unstable wobble.
(Some of you know that I am talking about a Fourier series.) OK, let us stop and take a deep breath. Even though I have avoid-
Now here is the point: in any model that satisfies my two crite- ed any formal modeling, you may at this point have suddenly
ria, some of these component wobbles - wobbles that go certain realized how seemingly abstract, how unrelated to the details of
particular numbers of times around the circle - will tend to grow freeways and shopping malls, skyscraper construction and fast food
over time (although wobbles at other frequencies7 may tend to die consumption, this has gotten. That is very much the style of com-
out). And, because the decomposition of an irregular wobble will plexity theorists: indeed, the whole rationale of the field is the idea
ordinarily contain wobbles of all possible frequencies, an even dis- that common principles may apply to subjects with very different
tribution of business around the circle is unstable. Some of the com- details. Still, has the abstraction led us into a story that conflicts
ponent wobbles will grow, creating an increasingly uneven spatial badly with reality?
distribution of business. Order from instability! My guess is that many readers will object to the implied regular-
Why, however, is the eventual spacing of business so regular? ity of the result - those equal-size, regularly spaced business con-
Because wobbles that go different numbers of times around the cir- centrations. In reality edge cities are not all the same size or equally
cle will grow at different rates. If the initial distribution of business spaced across the landscape. That does not worry me, however.
is sufficiently smooth, after a while the deviation from smoothness Mter all, the real landscape is not homogeneous. There are high-
will be dominated by whichever frequency wobble grows the fastest ways, whose intersections make particularly favorable sites for busi-
- by the most unstable wobble. And the frequency of this wobble, the ness, variations in the pleasantness or buildability of sites, and for
number of times it goes around the circle, will determine where the that matter real metropolitan areas do have centers and edges and
peak business concentrations are located. The undulations you see thus are prima facie not undifferentiated. Furthermore, the result
in Figures 1.5 and 1.6 are that most unstable wobble taking over the that business districts are evenly spaced is true only for an imaginary
distribution of businesses. For the parameters used to generate history in which business starts out spread almost evenly across loca-
Figure 1.5, the most unstable wobble goes around the circle twice; tions; because the real histories are not like that, we should expect a
for those used to generate Figure 1.6, the most unstable wobble more irregular result.
goes around the circle four times. Not only does instability create But if real cities do not evolve in the way I have just described,
does that mean that the whole approach is irrelevant? I think not,
7. Remember that these are spatial frequencies: these are wobbles in space, not but I need to introduce some more concepts and examples to
time. explain why.
-----2-----
Complex Landscapes

In his wonderful book Nature's Metropolis: Chicago and the Great West,
the historian William Cronon distinguishes between two landscapes
in which urban evolution takes place. One is the natural landscape:
the mountain ranges, rivers, and lakes that are givens of the envi-
ronment. The second is the created landscape of railroad lines,
canals, farming patterns, and cities themselves that results from
human decisions. Cronon argues that in the modern world the cre-
ated landscape, which he refers to as second nature, has become far
more important as a determinant oflocation than the "first nature"
in which it is embedded: Chicago's role as a Great Lakes port was
quickly overshadowed by its role as a rail hub. And second nature is
often self-reinforcing: railroads aimed at Chicago because it was the
economic center of its region and thereby made its centrality all the
greater. That is, the landscape of second nature is inherently
dynamic.
But how does one visualize a landscape that changes over time?
One answer is a historical atlas: a sequence of maps that traces out
the changes. In some cases it may be possible to do a bit better by
drawing a kind of temporal relief map in which the time dimension,
and thus the evolution of the landscape, is more or less imperfectly
represented. Indeed, Figures 1.5 and 1.6 are just that: they show
how some imaginary landscapes change over the course of their
imaginary histories. But even when you can do this, all that you get
is a description of what happened; why it happened is at best implic-
it.
So what do you do? Anyone who has worked on formal models
of dynamic systems knows the answer: you try to draw a picture of
11
32 SELF-ORGANIZING ECONOMY Complex Landscapes 33

yet a third landscape, a "phase space" in which each point summa- states are linked in some way: the energy of a spin-glass, and thus its
rizes the position of the system at a point in time and in which the likely direction of change, depends on whether nearby dipoles are
rules that govern the system's evolution are translated into "laws of oriented in the same direction or not; neurons excite or inhibit each
motion" in that abstract landscape. other; the effect of changing an individual gene on an organism's
Phase space representations of dynamic systems are extremely fitness, and thus the likelihood that a mutation will survive,
common in modern economic analysis. In general, however, we depends on what other genes it has. Such a system, it turns out, will
tend to focus on only a narrow range of possible types oflandscapes produce a complex dynamic landscape as long as two conditions are
- indeed, basically on only two fairly simple forms. Before the mid- met: the responses of the individual units must be discrete (neurons
1970s nearly all dynamic models in economics were globally stable; either firing or not), and there must be a mixture of positive and
that is, the phase landscape was assumed to be like a bowl, a single negative feedback (neurons both excite and inhibit each other).
basin of attraction in which all points drain to a single long-run Now think about spatial economies. They are systems in which
equilibrium. Since about 1975 it has become common also to work many components (firms) are at any particular time in particular
with models in which the phase landscape looks like a saddle - and states (locations) and in which changes in these states are linked
in which some set of forward-looking variables, such as asset prices, (through agglomeration economies and diseconomies). If there are
is determined by the assumption that the economy is always on the significant economies of scale, firms will choose only a few discrete
ridge that is the only path to long-run equilibrium.1 locations; and there will usually be a mixture of positive and nega-
The literature on complexity, however, is largely concerned with tive feedback between these choices. Therefore we might well
systems in which the dynamic landscape looks like neither a bowl expect spatial economies to have complex, rugged dynamic land-
nor a saddle; indeed, it is often concerned with "rugged landscapes" scapes.
(as Stuart Kauffman puts it) that look like the South Dakota bad- And indeed that is what you find when you look at even quite
lands. simple models of spatial economies. Over the last few years, I have
What aspects of a dynamic system lead to a rugged phase land- spent a lot of time trying to understand the behavior of a model
scape? I have no general answer, and I do not know if there is one. almost as minimalist as the edge city model I introduced a little
Many of the models in the complexity literature, however, have a while ago. (In fact, as I shall explain in Part II, the two models
similar setup.2 The modeler represents his or her system as an behave in very similar ways.) In this structure we imagine that there
ensemble of many components, each of which is at any particular are two factors of production: immobile agricultural workers and
time in one of several states: magnetic dipoles that are oriented up mobile manufacturing workers. Manufacturing is a monopolistical-
or down, neurons that are firing or quiescent, genes that have one ly competitive sector characterized by both increasing returns at the
character or the other, and so forth. And future changes in these level of the firm and transport costs. The interaction among factor
mobility, increasing returns, and transport costs generates forces for
1. Strictly speaking, one can represent a two-dimensional dynamic system as a agglomeration: firms tend to concentrate production in locations
three-dimensional relief map only if the system acts as if it were following the gra- with good access to markets, but access to markets is good precisely
dient of some potential function. This is sometimes reasonable, but not always:
the evolution of a single species is in effect maximizing something we can call fit- where other firms are concentrated. Working against these "cen-
ness, but the coevolution of predators and prey is not. Nonetheless, the language tripetal" tendencies, however, is the "centrifugal" pull provided by
commonly used to describe phase space, with its basins of attraction, saddle paths, the geographically dispersed agricultural sector.
and so on, draws heavily on the relief map metaphor; and I at least view the poet- If we add some rudimentary laws of motion, say the assumption
ry as worth the potential confusion.
that manufacturing workers tend to move to locations that offer rel-
2. See, for example, the influential paper by Hopfield (1982) and compare it atively high real wages, we get a dynamic story in a phase space
with the "NK model" ofKauffinan (1993).
 Complex Landscapes 35
34 SELF-ORGANIZING ECONOMY

defined by the allocation of manufacturing workers across locations.

And this dynamic landscape can easily be very complex.3
An easily shown example is the three-region case. Suppose that
there are three equidistant locations, with equal agricultural labor /
/,
o

t ,

forces. The allocation of the manufacturing work force among the / \ I ,

'" - I ••••• "
three locations can be represented as a point on the "unit simplex":
'" "'" ...•. "
- " I - " , ,,
a triangle whose corners are at the points (1, 0, 0), (0, 1, 0), and (0, J \

0, 1) in a space whose axes are the share of manufacturing in each

location but that can conveniently be pasted onto a two-dimen-
sional page. Starting at any given point on that simplex, one can let
the model evolve and see where it ends up. To draw the picture ana- / -
,/
~
~

'"
./'"
~---+-
/"
__/t"
\

- --
.......•.
1
,/
• ........-""""'"
••.....••.. , .•....•
"-

\.
lytically is extremely difficult, but it is straightforward to compute
numerical examples.
/~
•.......-- +-- ..-+- -
......•..•...•• .•.. '"
'" _0'\..
...•..•...•.

Figure 2.1 shows what I get for the most interesting range of para-
meters. (There are only three parameters in the model: the elastici-
ty of substitution among products in the manufacturing sector, set
for this example at 4; the share of manufactures in expenditure, set Figure 2.1 Basins of Attraction. A three-region economic model can end
at 0.2; and the transport cost between any two locations, set at OA.) up with four different locational patterns, depending on initial
conditions.
At each of a number of points on the simplex, representing an ini-
tial allocation of manufacturing workers, I draw an arrow represent-
imentally, by starting with a number of random allocations of man-
ing the direction and speed of "flow". It turns out that there are
ufacturing across locations and seeing how they evolve.
four equilibria: three in which all manufacturing is concentrated in
Figure 2.2 illustrates a typical run. The initial random allocation
one location, one in which there is an equal distribution of manu-
of manufacturing eventually organizes itself into two manufacturing
facturing across the locations. There are correspondingly four basins
concentrations, at locations 6 and 11; that is, 5 apart. In the course
of attraction: a central basin that leads to the equal division out-
of a number of runs with these parameter values, I got two concen-
come, and three flanking basins that lead to concentration.
trations 5 apart about 60 percent of the time, two concentrations 6
The landscape can become far more complex when there are
apart on almost all other occasions. At rare intervals a run would
more locations. Suppose, for example, that we consider an example
in which there are 12 locations, laid out in a circle like the numbers lead to three equally spaced concentrations.
On a circle with 12 locations, there are 12 ways to place two
on a clock. (Twelve is a particularly convenient number because it
markers 5 apart, 6 ways to place them 6 apart, and 4 ways to place 3
is a fairly small number with a large number of divisors.) The
equidistant markers. So it appears that with these parameters the
dynamics once again take place on a unit simplex - but this time an
ll-dimensional one. This is hard for most of us to visualize. We model implies a landscape with 22 basins of attraction - rugged ter-
rain indeed. In such a world the location of economic activity, and
can, however, get a good idea of the properties of the model exper-
to some extent even the structure of the resulting economic geog-
raphy, would depend crucially on initial conditions, which is to say
3. The kind of complex landscape that can arise in models of economic geog- on historical contingency.
raphy can, of course, arise in many other economic contexts as well. Most notably, Now, finally, we get back to the question that led us off on this
the choice among several technologies subject to network externalities will present
a very similar picture, with the landscape complex if none of the technologies has side trip through phase space. The beautifully simple, aesthetically
too strong an inherent advantage and the externalities are sufficiently powerful. pleasing "'59 Cadillac" pictures in Figures 1.5 and 1.6 were based
36 SELF-ORGANIZING ECONOMY Complex Landscapes 37

0.5
Let us look again at what I said about the 12-region model of
economic geography whose outcome is represented in Figure 2.2.
0.4 The model appears to have a dynamic landscape with 22 basins of
c:
t>O
attraction. And yet, looked at a different way, the results are sur-
'C
a prisingly ordered. Starting with a random allocation of manufactur-
~
:::l
0.3 ing across space, the model always organizes itself into a highly
c:
..s ordered structure in which manufacturing is concentrated in two or
E
three equal-size concentrations. Furthermore, although there are
'0 0.2
Q)
L.
many such equilibrium structures, they share strong similarities: all
..s
..r::. involve roughly equidistant city locations, with a fairly narrow range
II)
0.1
of typical distances between cities.
What we saw in the edge city model was these characteristic fea-
tures of equilibrium structures in an extreme form: if we started that
0.0 model from an almost uniform initial distribution of business (cor-
2 3 4 5 6 7 8 9 10 II 12
responding to a starting point near the center of the simplex in
Location
Figure 2.1), we got business concentrations exactly evenly spaced
around the circle, with an invariant distance between concentra-
I_Initial _Final I
tions. The same occurs in this model: if I take the parameters used
Figure 2.2. Exploring a Complex Landscape. Models with more to generate Figure 2.2 but restrict myself to initial positions in
regions can be explored via simulations; here is a typical run for
a 12-region example.
which the distribution of manufacturing is sufficiently flat, I will
consistently get a picture like Figure 1.5 - the economy organizes
on the assumption that the spatial economy started with an almost itself into a structure with two cities, exactly opposite one another.
uniform distribution of business across space. But that is not how What this suggests to me is that the fake history, in which a high-
real economic history works - and we have just seen that spatial ly regular spatial structure emerges from an almost unstructured ini-
economies typically must have complex dynamic landscapes in tial position, can be viewed as a sort of model of the model. That is,
which where you end up depends a lot on where you start. So is the the precise regularities of that special case help us understand the
simple picture basically irrelevant? rough regularities of the more general case.
I think not - because although spatial models easily generate Nor do I think that this insight is merely about modeling. It
complex dynamics, lurking within that complexity we often find seems reasonable to speculate that the immensely more complex
surprising simplicity. landscape that determines the real geography of the world economy
has its own underlying approximate simplicities. That is, there may
THEEMrnRGENCE OF ORDER be many possible outcomes, depending on initial conditions -
Silicon Valley might, given a slightly different sequence of events,
The most provocative claim of the prophets of complexity is that have been in Los Angeles, Massachusetts, or even Oxfordshire. But
complex systems often exhibit spontaneous properties of self-orga- some broader features may be more or less independent of histori-
nization, in at least two senses: starting from disordered initial con- cal contingency.
ditions they tend to move to highly ordered behavior, and at least in And in fact one regularity in spatial economics is so spectacular in
a statistical sense this behavior exhibits surprisingly simple regulari- its exactness and universality that it is positively spooky. That regu-
ties: for example, a power law distribution relating the sizes and fre- larity involves the size distribution of cities, and it leads us to our
quencies of earthquakes. other principle: order from random growth.
,I
-----3-----
An Urban Mystery

Let us start with a picture. Suppose that you take all of the metro-
politan areas listed in the Statistical Abstract if the United States, 130 in
number, and rank them by population. And suppose that you plot
the ranks of these areas against their populations, using a logarith-
mic scale. What you get is Figure 3.1.
What is interesting about this figure? Of course the line relating
rank to population is, by definition, downward sloping. But there
are two things about that line that did not have to be true. First, it
is pretty close to a straight line: there seems to be something close to
a log-linear relationship between rank and population. That in itself
is pretty interesting; surely it suggests that some hidden principle is
at work. But even more interesting is the slope of that approxi-
i I

i
mately linear relationship: it is very close to -1.
There is another way of putting these results, which is also anoth-
er name for what is sometimes called Zipj's law: it is the rank-size
rule. This rule says that the population of a city is inversely propor-
tional to its rank. If the rule held exactly, the number 2 city in a
country would have half the population of the biggest city, the
number 3 city one-third that population, and so on. Obviously the
rule does not hold exactly. If you look at only the largest metropol-
itan areas, you may be unconvinced: Los Angeles is considerably
more than half as populous as New York. But once you get down
the ranking a bit, the fit starts to become almost terrifyingly exact.
11 For example, the 10th ranked metropolitan area in the United
States is Houston, with 3.85 million people. The 100th ranked area
is Spokane, Washington, with 370,000 people - close enough to
I

I 39
40 SELF-ORGANIZING ECONOMY An Urban Mystery 41

3.5 7
0tl()u0 2.0
L..
'"
...J 5.0
3.0 6.0
1.56.5
c ·~
.•....
2.5
5.5
~ 0.5
1.0
0.0
4.5
4.0
6

6c 4~---- -----
~

2~------------------------------------------------------------------------

7.0 7.5 8.0 8.5 9.0 9.5 10

Log of city size 1

3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0

Figure 3.1 Zipt~s Law: The size distribution of metropolitan areasin the In(S)
United Statesis startlinglywelldescribedby a powerlaw,with an
exponent very closeto 1.
Source: Statistical Abstract if the United States, 1993.
I • 1990 + 1940 * 1890 I

Figure 3.2 Zipf's Law over Time. The samepower law has worked rea-
sonablywell for at leasta century.
one-tenth of Houston to be well within reasonable uncertainty of Source: Historical Statistics if the United States.
definition and measurement. If you regress the log of rank on the
log of population, you get a coefficient of -1.003, with a standard Zipfs law is not quite as neat in other countries as it is in the
error of only 0.01 - a slope very close to 1 and very tightly fitted. United States, but it still seems to hold in most places, if you make
We are unused to seeing regularities this exact in economics - it one modification: many countries, for example, France and the
is so exact that I find it spooky. The picture gets even spookier United Kingdom, have a single "primate city" that is much larger
when you find out that the relationship is not something new - than a line drawn through the distribution of other cities would lead
indeed, the rank-size rule seems to have applied to u.s. cities at you to expect. These primate cities are typically political capitals; it
least since 1890! Figure 3.2 shows data taken from Historical Statistics is easy to imagine that they are essentially different creatures from
of the United States, which reports the number of "urban places" in the rest of the urban sample.
specified size ranges; I show the number with more than 100,000, What could explain the existence of something that looks suspi-
more than 250,000, more than 500,000, and more than 1 million ciously like a universal law on city sizes? Bear in mind that there are
for 1890, 1940, and 1990. (These data are not quite comparable to two things we need to explain. We need to understand why the
those in Figure 3.1, because of the difference in definition between rank-size relationship is nearly linear in the logs, and we need to
an urban place and a metropolitan area.) The picture is not perfect, understand why it stays so oddly close to a slope of -1.
but if you had any reason to believe that there was a fundamental When urban economists attempt to explain the rank-size rule,
reason to expect a straight line with a slope of -1, you would find they usually argue that it reflects some sort of hierarchy of central
this picture compelling evidence in favor of your theory! places. Here is how this reasoning goes. First, we point out that an
42 SELF-ORGANIZING ECONOMY An Urban Mystery 43

alternative way to express the linearity of the rank-size relationship Second, the hierarchical story depends on the constancy of a
is to say that the distribution of city sizes follows a power law - the number of parameters that we have no particular reason to think are
number of cities whose population exceeds some size S is propor- constant. Why should the span of control of managers be the same
tional to S to some negative power. when they are supervising very different levels of activity? Why
Next, we note that a power law distribution is consistent with a should the number of people in an office at each level be a constant
simple hierarchy. Suppose, for example, that we think of a manage- multiple of the number at the level below? The assumptions seem a
rial hierarchy, in which each manager has the same span of control; bit too close to the conclusions.
say, each manager oversees 10 subordinates. And suppose also that, And finally, the story fails to explain the full, astonishing fact of
say, three of those subordinates, plus their subordinates (if any) must the rank-size rule. A hierarchical model suggests why a hierarchy of
work in the same location as their superior. Then there will be a cities might more or less obey a power law on sizes; it does not
hierarchy of managerial sites: for every 7 individual field offices, explain why they should almost exactly follow not only a power law,
there will be a central office with 4 people; for every 7 such offices, but the very specific law that size is precisely inversely proportion-
there will be a more central office with 13 people; for every 7 al to rank. Indeed, the example I have just given has size falling
offices there will be yet another level of offices with 40 people each; much too slowly with rank to be consistent with the data.
and so on. If you can imagine this hierarchy going on forever, you But, if we abandon the idea of a central place hierarchy, what pos-
will find that each successive type of office will have approximately sible explanation can we offer for Zipfs law?
three times as many employees as the previous level. Because each
office supervises seven off-site offices at the previous level, this
METEORITES, EARTHQUAKES, AND CITIES
means that the distribution of office sizes will be a power law with
an exponent of -7/3, or -2.33. The mysterious regularity implied by Zipfs law is not unique to
Now imagine that the sizes of central places reflect just such a urban economics. On the contrary, power law regularities appear in
hierarchy: each local central place holds field offices, each member many other fields. The frequency with which meteorites exceeding
of the next level contains subregional offices, and so on - you may any given diameter strike the Earth follows a power law. There is a
have a story about the apparent power law on city sizes. famous log-linear relationship, the Gutenberg-Richter law,
Stories of this kind have come in for some unjustified criticism. between the sizes and frequencies of earthquakes. Plot the number
Taken literally, they seem to imply a set of discrete city sizes, not the of animal species that exceed a given size and you will again find the
smooth distribution of Figure 3.1. The usual answer, which is sure- same kind of relationship. So we are not alone in our mystery.
ly right, is that we need not suppose that the relationships are exact Where do such strangely simple relationships come from?
and that with a little noise the distribution could well appear Loosely speaking - and much of the reasoning in this area is sur-
smooth. prisingly loose - it seems that you get results like this when three
And yet I am convinced that such hierarchical explanations of the criteria are satisfied. First, the objects you are studying are subject to
rank-size rule are wrong, for three quite different reasons. substantial growth over time. Second, the growth rate of any indi-
First, the whole image of central place theory, in which towns vidual object is random, so that over time you get a wide range of dif-
and cities form a structure whose spacing is dictated by a dispersed ferent sizes. Finally, however - and crucially - the expected rate of
agricultural population, seems hard to credit in a modern economy growth (though not necessarily its variance) must be independent of
where land-based activities like farming play such a small role. scale: large objects must grow on average neither faster nor slower
Indeed, a glance at the nighttime satellite photo of the United States than small ones.
that I keep on my wall shows a picture that looks nothing like a reg- Take the case of meteorites. The current best-guess model starts
ular central place pattern. Instead, we see great belts oflight in a few with a bunch of fairly big rocks out there in space. Every once in a
areas, vast stretches of near-darkness in between. while, these rocks collide and break into smaller fragments. (This is
44 SELF-ORGANIZING ECONOMY An Urban Mystery 45

a growth process, in which the expected growth rate is negative.) And that is it. If you like, you can easily give the process a mini-
Because collisions and fragmentation are random, you get a wide malist economic interpretation. Think of a "lump" as being the
variety of meteorite sizes; because big rocks and small rocks are just number of people employed directly or indirectly by an entrepre-
as prone to fragmentation, you get a power law. neur. Imagine that new entrepreneurs are usually people inspired by
Earthquakes are a more subtle and disputed case. The Santa Fe (perhaps working for) existing enterprises; so any given city will
theorist Per Bak has proposed a model in which earthquakes are generate new ideas in proportion to its population. And finally
produced by a kind of chain reaction. Think of the land near a fault imagine that entrepreneurs usually, though not always, stay near
line as consisting of many blocks of earth, each under stress coun- their point of original inspiration. Then you have Simon's story,
tered by friction. If the stress exceeds a critical level, a block will slip more or less.
- and in so doing transfer stress to its neighbors, some of which may Now this story feels all wrong. Mter all, it seems to imply that the
in turn be triggered into slipping, and which may trigger their size of a city is essentially irrelevant! In the story, entrepreneurs
neighbors in turn, and so on. Bak has shown that if the average level "give birth" to additional entrepreneurs at a rate that is independent
of stress on the blocks is close to a critical level, such chain reactions of how big the city is. Surely we believe that there are all sort of sys-
constitute a sort of growth process with the right features to pro- tematic differences among cities of different sizes - big cities are
duce a power law, random growth that is independent of scale. And more creative but more costly to live in, small cities typically get the
he argues that earthquake zones tend naturally to evolve precisely to industries that have matured in the dynamic big-city environment,
that critical level of stress, in what he calls self-organized criticality. and so on and so forth. This nihilistic story must be too simple to
Can we tell a story about cities that is anything like these cosmic capture the complexities of reality. Except that the reality of the
accounts? Indeed we can. In fact, Herbert Simon proposed such a city-size distribution is not complex - it is startlingly, spookily sim-
story more than 40 years ago, in one of those papers that should ple. And Simon's story does explain that simplicity.
have been extremely influential but that somehow, whether Consider Figure 3.3. It represents the result of a run of a Simon
because they did not fit the zeitgeist or because they were written story with 1C set equal to 0.2; I started with 10 "seeds," initial cities
in the wrong style, were largely ignored. Perhaps this book will suf- 4.0
0 1.6
1.21:5
2:0
0~c 2.4
....I
fer the same fate, but anyway let me offer my own version of 'v.•..
tlO2.0
1.0
~
Simon's model. ~ 0.8
3.2
0.0
3.6
0.4
2.8

SIMON ON ZIPF
I would describe Simon's story (I call it a story, because it is almost
too nihilistic to call a model) as one of "lumps and clumps." We
imagine an urban system with a growing population. Population
does not, however, arrive smoothly; it arrives in discrete "lumps"
that are neither too small nor too big. I shall explain what that
means in the next part.
Where does a lump land? With some probability 1C the lump lands
in a new place; that is, a new city is formed. With a probability 1 -
2.5 3.0 3.5 4.0 4.5 5.0 5.5
1C the lump attaches itself to an existing "clump," which consists of
Log of city size
one or more lumps. The probability with which any given clump
attracts a new lump is proportional to its population. Figure 3.3 Lumps and Clumps. A model of random urban growth pro-
duces results that are similar to the actual data - but not as neat!
46 SELF-ORGANIZING ECONOMY

each consisting of one lump, then allowed the population to

increase by a factor of 100. The figure shows the rank-size distri
bution for the top 50 cities. It is not quite as neat as the real data -
it is, indeed, a peculiar feature of the whole Zipf issue that the real
data are cleaner than anyone's model! But it does bear a strong
resemblance to the real distribution, both in that it roughly obeys a
power law and that the slope of the relationship is close to -1.
It turns out that Simon's process always does, if you run it long
enough, generate something that looks like a power law distribution
-----4-----
for the sizes of larger clumps. The exponent on this distribution Principles of Self-Organization
depends on the probability of forming new cities, n; numerical
experiments suggest that you get an exponent close to 1 when n is
small. In fact, Simon showed - in a completely impenetrable expo-
sition! - that the exponent of the power law distribution should be The more mystical prophets of" complexity" express the hope that
1/(1 - n). I shall offer some intuition on this in Part II, including we will eventually arrive at universal laws of self-organization that
what I hope is a more user-friendly derivation than the one Simon apply to all complex dynamic systems. A more modest goal would
provided. The point, however, is that, if we accept the simplistic be to identify families of situations, possibly in very different con-
justification for Simon's process just given, there is a natural eco- texts, in which similar principles and behavior arise; it would be dis-
nomic reason not only for the existence of a power law but for this appointing but not crushing if it turned out that no common prin-
power law to have a slope near 1: what the data are telling us is that ciple links the different families and if the families do not cover all
only a small fraction of entrepreneurs move far from the source of possible situations.
their ideas. At least at this point I remain in the camp of modest objectives. I
want to make no grandiose claims about having found a theory of
everything. And yet the principles I have identified apply to a wide
class of phenomena in a startlingly wide array of fields.

ORDER FROM INSTABILITY

Try filling a pail with a mixture of golf balls and Ping-Pong balls. If
you are careful, you can manage to layer them together so that Ping-
Pong balls are as likely to be on the bottom as on the top, golf balls
as likely to be on the top as on the bottom. But shake the pail, and
pretty soon it will be all Ping-Pong above, golfbelow. The initially
disordered structure organizes itself, because a disordered structure
is unstable when subject to random shocks.
The family resemblance of the Ping-Pong/golf example to
Schelling's urban segregation model should be obvious. Indeed,
models in which order emerges from instability are pervasive in the
physical sciences. The classic illustration is convection. Heat the
bottom of a pan of water and cool its top. Once the temperature

47
48 SELF-ORGANIZING ECONOMY Principles of Self-Organization 49

gradient reaches a critical level, a motionless equilibrium becomes greatly in the details: in particular, in the biological analysis it is
unstable. The water then organizes itself into a pattern of hexagonal essential to model cells as being affected directly only by their
convection cells. immediate neighborhood, whereas in an urban model nothing is
Meteorology, which is concerned largely with convection in the disturbing about the idea that a mall's profitability is affected by
wild, is often thought of as the home of chaos. But it is also a sub- conditions 10 miles away. Still, there is a clear family similarity, and
ject preoccupied with the emergence of large-scale if temporary some essential elements of the mathematical formalism, which I
order. In his fascinating book on the subject, Edward Lorenz, who shall discuss in the next part, can be carried over directly. So as I
"discovered" chaos, nominates as the greatest meteorological theo- promised at the beginning, a city is indeed something like an
rist of all time the Norwegian Vilhelm Bjerknes. What was embryo.
Bjerknes's big insight? That "the reason that vortices and other I believe that the principle of order from instability, and in par-
structures of continental size must be present in the atmosphere is ticular the notion that self-organizing systems tend naturally to be
not the dynamic impossibility of a flow pattern without them ... it dominated by the most unstable fluctuations, is a powerful idea. It
is the instability of such a pattern with respect to inevitable distur- has certainly helped me to think about metropolitan structure - or
bances of large horizontal extent but small amplitude." perhaps we should just call it urban morphogenesis. I suspect that it will
Perhaps most surprising, the theme of order from instability often be a way to cut through the complexity of what may seem
appears in biology as well. One of the biggest questions in biology hopelessly difficult models.
is morphogenesis: How do we get from a single fertilized egg to an
organism that not only requires many different kinds of specialized ORDER FROM RANDOM GROWTH
cells but particular types of cells in particular places? An individual
fruit fly cell presumably does not think to itself, "I am part of a Take a ceramic object, say a Grecian urn, and throw it hard against
wing," yet cells collectively seem in effect to decide to become dif- a stone wall, so that it shatters randomly into innumerable pieces.
ferent parts of the organism. Experiments suggest that cells indeed Surely such an act can do nothing but create disorder! And yet (so
behave as if they knew their own polar coordinates! The Economist tells us) a strange hidden order emerges. Carefully
It turns out that there is a classic 1952 theoretical paper on mor- gather up the pieces of that smashed urn and count the number of
phogenesis, which may explain this remarkable fact, by none other pieces larger than 0.1 grams, the number larger than 0.01 grams,
than Alan Turing, of computer fame. Turing pointed out that and so on, and you will find something remarkable: the pieces will
chemical systems - for example, a two-chemical system in which obey a power law. And not just any power law: if it really was a
one chemical catalyzes both its own production and the production Grecian urn, the exponent will take on a particular value; if it was a
of the other, the second inhibits the production of the first, and the ceramic sphere, it will take on another value; and so on. That is, a
two diffuse at different rates - can exhibit spontaneous spatial self- seemingly disorderly and complex process of fragmentation (which
organization. Small fluctuations in the initial distribution of the is random growth with a minus sign) produces the simple order of
chemicals will tend to grow over time; in particular, any such sys- a power law, and the exponent of that power law contains impor-
tem will have a particular "preferred wavelength," a particular size tant information - in this case, incredibly, it turns out to reveal the
of fluctuation that tends to grow fastest and thus to dominate the shape of the original object.
emergent pattern. He then argued that, if one of these chemicals is Like the idea of order from instability, the idea that simple fre-
a "morphogen," a substance that tells a cell what to become, this quency distributions can be the result of random growth - and that
self-organizing chemical system can allow cells to behave as if they the exponent of the power law is telling us something important
know where they are in relation to one another. about the process - is ubiquitous in physical science. Indeed, we
Does this sound familiar? The edge city model described previ- may say that physical scientists now rely on random growth to
ously was inspired by Turing's reaction-diffusion model. It differs explain regularities from the ground up, from Bak's earthquake
50 SELF-ORGANIZING ECONOMY

model to the size distribution of meteorites and beyond. Current

cosmological models now rely on the random growth in an initial
period of "inflation" to generate the uneven distribution of mass
that allowed galaxies to form, giving new meaning to the idea of a
universal principle. And, if it is good enough for the universe, it is
good enough for a trivial thing like the economy!

-----5-----
Where We Stand

I started this lecture with what must have sounded like some very
peculiar questions. I hope that at this point it is clear not only that
the questions make a lot of sense but even what the answers might
be. When I asked whether a city is like an embryo or like a mete-
orite, whether a recession was like a hurricane or an earthquake, I
was asking whether the relevant principle of self-organization was
that of order from instability or order from random growth.
This has been a pretty strange intellectual journey. I hope you
have enjoyed it. Yet many of you may feel that I have been awfully
loose. I have made a lot of assertions about models that I have not
really fully described, let alone worked out. That was deliberate. I
wanted to show you the range of issues, the commonality, and to be
frank, the sheer fun involved in thinking about the self-organizing
economy. Yet I value disciplined thinking as much as anyone; so in
the next part, let us settle down a bit and look at how some real
models of self-organization work.

51
 PART
-----2-----
Self-Organization in Time and Space

In the previous part I tried to suggest, with sweepingly loose expo-

sition, how and why an economy might exhibit principles of self-
organization similar to those exhibited in embryos and neural net-
works, ecologies and hurricanes. Along the way, however, I made
some promises that I never quite fulfilled.
First, I said that I would talk about self-organization both in space
and in time. In the first part, however, all my examples were spatial.
That was deliberate: on just about all counts I am on much more
solid ground when I talk about spatial self-organization than when
I try to tell temporal stories. Partly that's because the business cycle,
which looks as if it might be a self-organizing phenomenon, is also
a subject on which it is all but impossible to speak without offend-
ing either common sense or the equilibrium macro faithful. It is also
true that, if there are striking empirical regularities in business
cycles, I am not aware of them. Still, in this part, I will spend some
time on the subject, describing two models and one crazy specula-
tion.
Second, I alluded to spatial models without quite explaining what
the models were, let alone showing how they worked. So in this
lecture I will also take time to explain the workings of three spatial
models that illustrate the basic principles of self-organization.
Without further ado, then, let us turn to some of the temporal
aspects of self-organization.

c;1
-----6-----
Dynamics in Self-Organizing Systems

In the field of international trade, there is what might seem to the

uninitiated a peculiar distinction between two areas of inquiry: the
solid, well-established theory of "growth and trade" and the much
more controversial subject of "trade and growth."
What is the difference? "Growth and trade" considers the impact
of growth on a trading economy. Suppose that Chinese productiv-
ity rises or Japan trains more engineers: what is the effect on world
trade patterns and U.S. real wages? These are well-understood
issues; what makes them fairly cut and dried is that it is assumed,
tacitly or explicitly, that we may ignore any feedback from trade to
growth, that the forces of change are exogenous. On the other
hand, "trade and growth" is concerned with precisely such feed-
backs: how, for example, does trade liberalization affect the rate of
technological change? Is a liberal trade regime good for develop-
ment, or is sophisticated mercantilism better? These are sexy and
topical questions, but the answers are far more controversial.
I would argue that there is a similar distinction between two con-
cepts of dynamics in self-organizing systems. One subject, which
economists already understand fairly well, involves the time dimen-
sion of a system that is self-organizing in some other dimension,
such as space. The other subject, which is much more controversial,
involves self-organization along the time dimension itself
Let us start with the easy issue. Imagine an economy that, at any
point in time, may be described as a self-organizing system. In par-
ticular, imagine that its dynamics form a complex landscape, in the
sense I described in the preceding part. And now suppose that this
landscape is subjected to exogenous forces of change, forces that
C\"
56 SELF-ORGANIZING ECONOMY Dynamics in Self-Organizing Systems 57

gradually exalt the valleys and bring the hills and mountains low. % of R
Can we say anything interesting about the likely process of change households
in the economy?
H
Yes, we can: in general, a complex dynamic economy will exhib-
it the pattern that in evolutionary theory is known as punctuated equi- ,,,,
librium: long periods of relative quiescence, divided by short periods ,,,,
,,,,
of rapid change.
Let me illustrate this point, not with the spatial models that I use
,,,,
in most of these lectures but with a model of technology choice.
"""'"
This kills two birds with one stone: I can tell a simple story and at ,,,,
the same time I can demonstrate that the concepts I have been ,,,,
advancing apply to technology as well as location. ,,,,
Imagine, then, a situation in which households have a choice ,,,,
,,,,
between two technologies - say, laser discs versus videotape cas- ,,,,
settes. Each technology has some inherent advantages over the ,,,
other, but for many people the most important factor is the avail-
ability of some support service, such as movies to rent in video c
stores. The provision of that service, in turn, depends on the size of
the market. So most people will prefer videotape recorders if there
% of stores
are mostly cassettes in the stores, and stores will mostly stock cas-
settes if that is what most people are prepared to rent.
Figure 6.1 shows what the dynamics of this system might look Figure 6.1 Choosing a Video Technology. If consumers choose VCRs
that can play what stores rent, and stores rent what consumers can
like. On the figure's horizontal axis is the percentage of stores play, the choice of a video technology can be self-reinforcing.
stocking laser discs (as opposed to cassettes); on the vertical axis the
percentage of households with laser (as opposed to videotape) play- nant. So we have the multiple basins of attraction that define a com-
ers. The curve H represents the equilibrium share of consumers plex dynamic landscape. The broken line illustrates schematically
with laser players, given the number of stores offering the discs; the the division between the C basin and the L basin.
curve R represents the equilibrium share oflaser discs available for Now let us introduce change into the picture. Suppose that
rental, given the number of households in a position to use them. I although cassettes currently dominate the scene, the direction of
assume that each of these shares adjusts gradually toward its equilib- technological change is gradually to increase the advantages of laser
rium value. discs. Then other things equal, households will become gradually
Finally, I have drawn both Hand R as S shaped; that is, I assume more willing to buy laser players; as shown in Figure 6.2, the curve
that there is a kind of critical mass of available rentals needed to per- H will shift gradually up toward H'.
suade most consumers to buy laser players and a critical mass of con- But will the actual process of change be gradual? At first it will:
sumers needed to persuade stores to stock the discs. the equilibrium will move along the R curve from C to C'. But at
The arrows on the diagram represent the laws of motion. Clearly, that point something more drastic will happen: the whole cassette-
there are two stable equilibria in this picture. At one equilibrium, dominated equilibrium will collapse. Households will turn to laser
C, there are very few laser players or discs around; in this equilibri- players; as they do so, stores will shift shelf space from cassettes to
um (as in the real world at the time of writing), cassettes are the discs, further reinforcing the attractiveness oflasers, and so on, lead-
dominant medium. At the other equilibrium, L, discs are domi- ing to a discrete shift of the whole system from C' to L'.
58 SELF-ORGANIZING ECONOMY Dynamics in Self-Organizing Systems 59

% of H' ferent locations (as they indeed did in my 1991 book Geography and
households Trade); or for that matter they could represent the survival strategies
of two interacting species, be they competitors, symbiotes, or
predator and prey. 1 In any of these cases, we would expect gradual
changes in the underlying parameters, and therefore of the phase
space, to produce only gradual change in the outcome most of the
time, but big changes every once in a while.
The basic logic of punctuated equilibrium should be clear. In a
complex landscape, there are many locally stable equilibria; sudden
change comes when a previously stable equilibrium becomes unsta-
I ble, setting the system adrift while it searches for a new equilibrium.
I
,, But that change from stability to instability always occurs at a dis-
I crete moment, no matter how glacial the pace of change in the
I
I

,.,'
I, I underlying landscape; and so gradual change in the environment
produces dynamics that resemble the traditional description of war:
long periods of boredom punctuated by brief interludes of terror.
;;
C
---_ .••. ' ;; Does such punctuated equilibrium actually manifest itself in
practice? Certainly some technologies, like cellular phones or
VCRs, seem to lurk in the background for years then suddenly
% of stores explode into mass use. Even more notably, urban history contains
many examples of explosive growth: Chicago in the mid-19th cen-
Figure 6.2 Punctuated Equilibrium. Improvements in the less-used tury, Los Angeles after World War II. And edge cities, when they
technology may have little effect on consumer choices until the system nears emerge, often do so with stunning speed: Joel Garreau's book con-
the critical point at which the current equilibrium collapses; then a small fur- tains "before and after" pictures of Tyson's Corner in Virginia that
ther improvement will trigger large-scale changes.
dramatically make this point.
So the idea that self-organizing systems have a characteristic
The process of change in this story, then, will be one in which rhythm of change seems to apply to the economy as well. But in this
gradual shifts in technological advantage usually lead to equally story the basic sources of change are left unexplained - just as the
gradual changes in behavior, but in which there are occasional sud- sources of growth are left unexplained in the "growth and trade"
den shifts. literature. In an ugly, but rather useful piece of jargon, we can say
I have no idea whether this is a realistic story about the past and that the dynamics are extrinsic rather than intrinsic.
future of video entertainment, although it sounds plausible. The What would constitute a case of intrinsic, self-organizing dynam-
point, however, is that the story is quite generic. The axes in Figure ics? For that we need a model in which the economy organizes itself
6.1 could represent the numbers of workers and firms in two dif- in time rather than (or as well as) in space.

1. The change in the phase space need not be entirely exogenous; both tech-
nological development and evolution are at some level self-generating processes.
What is necessary to create punctuated equilibrium is for some processes to pro-
ceed much faster than others; for example, basic electronic technology moving
gradually, while market shares in videos can be changing much more quickly.
-----7-----
Temporal Self-Organization

TWO QUESTIONS ABOUT THE BUSINESS CYCLE

If you have ever tried to talk about the nature of recessions with an
intelligent friend who does not know economics, you have proba-
bly discovered that the troubles begin with the way you pose the
question. I usually begin by saying something like, "Suppose that
there is a fall in investor confidence - let us try to see how that can
translate into a fall in output and employment." At this point the
friend almost always asks, "But what caused the decline in investor
confidence?" and is not easily persuaded simply to take that decline
as a given. Indeed, most people seem to think that you are cheating
if you start with a decline in confidence, or a monetary contraction,
or whatever - if course, they think, that leads to a recession, the ques-
tion is why it happened.
Well, anyone who has followed debates in macroeconomics
knows that there is no "of course" about it. It is by no means clear
to many economists that a decline in investor confidence will lead
to a fall in aggregate demand; if you are a firm monetarist (or is the
species extinct?) you believe that nominal income is determined by
the monetary aggregates, independent of the state of mind of the
business community. In a way, the whole point of Keynes's General
Theory was to break free of the intellectual straightjacket of the
quantity theory, which asserted that changes in investment demand
were irrelevant; that is why the theory of liquidity preference,
which said that the demand for money depends on the rate of return
on alternative assets, played so crucial a role in his thinking.
Conversely, however, much of the point of monetarism was to put
that straightjacket right back on.

61
62 SELF-ORGANIZING ECONOMY 63
Temporal Self-Organization

Moreover, it is if anything even less clear to many economists that slump in world demand occurred in the absence of any obvious
changes in aggregate demand can have any real effects; even if cause.
investor confidence does fall, many academics believe as a matter of So there was a period when economists tried to answer the ques-
principle that any effects of that decline are on the price level rather tion that later fell by the wayside: what causes demand to fluctuate?
than output. Or to put it my way, how and why does the economy organize itself
Now I am not asking you to take sides in these arguments, over time?
though I guess it is pretty obvious that I am a more or less unrepen-
tant Keynesian. The point I want to make is instead one about our
NONLINEAR BUSINESS CYCLE THEORY
profession. For the last 30 years or so macroeconomic debate has
focused very heavily on what is sometimes called the transmission For a period of a decade or so, a number of economists propound-
mechanism, on whether and how a decline in particular types of ed an explanation of economic fluctuations, generally known as
demands like business investment is reflected in a decline in overall nonlinear business cycle theory, that was both intuitively persuasive and
demand, on whether and how a decline in aggregate demand is analytically elegant. Versions of this theory were propounded by
reflected in a decline in output and employment. Do shifts in the IS John Hicks and Richard Goodwin in the United Kingdom; there
curve reduce demand, or is that channel blocked by a near-vertical was also an early contribution along similar lines by James Tobin in
LM curve? Do shifts in aggregate demand affect output, or is that the United States. Unfortunately, the theory rested firmly on a
channel blocked by a near-vertical aggregate supply curve? These strongly Keynesian base. When this fell out of fashion, when the
are the kinds of questions that have preoccupied macro theorists. monetarist controversy was followed by the debate over aggregate
And rightly so, because these are deep issues, and we cannot claim supply, the whole enterprise of trying to explain why and when
to have a real theory of business cycles until we resolve them. fluctuations occur rather than how they are possible at all was large-
And yet something odd has happened as a result of this focus on ly abandoned, and the innovative dynamic theory that had flour-
the transmission: macroeconomics has increasingly ignored the ished in the 1940s was largely forgotten.
engine. I recently looked at Greg Mankiw's excellent textbook in Well, I want to resurrect that theory at least briefly. Even if you
macroeconomics, which gives a wonderfully clear overview of the think that Keynesian ideas are nonsense, bear with me for a little
state of the field (you know what that means: it means that I share while, because the point I want to make is less about the nature of
his views). There is very clear discussion of the reasons why fiscal the business cycle than about the nature of self-organization.
and monetary policy might or might not have real effects. What is What, then, was the essence of nonlinear business cycle theory?
notably lacking, however, is a discussion of the reasons why aggre- In its cleanest version (Goodwin's), it envisaged a modified version
gate demand fluctuates in the first place. Business cycles, in of the Keynesian cross. As I hope most of you know, the tradition-
Mankiw's book and in macroeconomics generally, are treated as if al Keynesian cross asserts that planned spending - investment plus
they arose entirely from exogenous shocks. consumption - is an increasing function of the level of output in the
This was not, however, always the way that macroeconomists economy, like the schedule shown in Figure 7.1. But in a closed
approached their subject. There was a period that we might call the economy, output equals spending. So equilibrium, defined as a sit-
Keynesian spring, from the war until about 1960, when most uation where planned spending equals actual spending, must be at a
macroeconomists accepted Keynes's own view that fluctuations in point where the spending function crosses the 45° line.
investment demand were the cause of booms and recessions. They In Figure 7.1 I have made the conventional assumption that the
were thus in the position of my lay friend: it seemed obvious to marginal propensity to spend is less than 1, so that the equilibrium
them that a fall in investment demand would cause a recession; is unique and it will also be stable under any plausible dynamics.
the interesting question was what caused the fall in demand. To get from here to a theory of the business cycle, we alter two
Furthermore, macroeconomists of that era still had fresh in their elements. First, we assume that over some range the marginal
mind the experience of the Great Depression, in which a massive propensity to spend is more than 1. This can be justified by invoking
 64 SELF-ORGANIZING ECONOMY Temporal Self-Organization 65

Planned Planned
/
spending spending //
/
/
/
/ /,/' H
/
/
/ ,///'
/
/
/
/
/

/
/
/
////
/
/
/ //
/
/
/
/
/
/
/
/
/
?
~
/ /
////
/
/ //
/
/

/
/
/ //
///
/
/
/
/
/
/
///
/
/ //J'/
/
/
/
/ /
/
/ /
/
/ // L
/
/
/
/
/
/
///
Output Output
Figure 7.1 The Keynesian Cross. In the standard Keynesian story, the cir- Figure 7.2 Booms and Slumps. If higher output sometimes leads to a
cular relationship between output and spending leads to a unique equilibrium. more than one-far-one increase in demand, there may be two
equilibria: one with a high level of output, the other with low
investment demand, which could well respond to changes in sales
more than one-to-one. However, this strong response of demand to given level of output, shifting down the expenditure schedule. This
output flattens out at both low and high outputs. If you like, you process is illustrated by the shift from 53 to 52 to 51 in Figure 7.3.
can suppose that at low output many sectors are doing zero invest- And the consequence is obvious: at a certain point the boom will
ment and can cut no further, whereas at high output either invest- collapse. Once the economy is in a slump, the capital stock will
ment or output itself is limited by capacity constraints. The impor- begin to decline and investment will begin slowly to revive; the
tant thing is that we assume an S-shaped expenditure schedule like schedule gradually will shift back up, from 51 to 52 to 53, and the
the one in Figure 7.2. In this case, of course, there are three equi- boom will be on again.
libria, with the middle one unstable. The rise and fall of the capital stock plays a key role in the story,
Second, we suppose that the expenditure function depends not but that role is hidden in Figure 7.3; it can be revealed in a phase
only on the level of output, but also - negatively - on the capital diagram like Figure 7.4, in which the capital stock is on the hori-
stock. And we suppose that at the high-investment equilibrium H zontal axis, the level of output on the vertical. The story about the
in Figure 7.2, the economy is accumulating capital, whereas at the business cycle that I just told implies that the economy ends up fol-
low-investment equilibrium L the economy's capital stock is depre-
ciating faster than it is being replaced.1 1. Or more plausibly,we add an exogenous component of expenditure that
What will happen? When the economy is in a boom, capital will grows with the economy's potential output. What declines over time is then the
accumulate. This will gradually reduce investment demand at any ratio of the capital stock to this trend.
,
66 Temporal Self-Organization 67
SELF-ORGANIZING ECONOMY

Planned y
spending /
/
/
/
53
52

51

/
/
/
/ K
/

Output
Figure 7.4 Another View of the Business Cycle. The same cycle can be
seen in phase space. There is always an equilibrium without a
Figure 7.3 The Business Cycle. The accumulation of capital during a cycle, but it is unstable.
boom eventually triggers a slump, while the depreciation of cap-
ital during a slump sets the stage for the next boom.
they were, understandably, dubious about some of the approach's
economic foundations.
Second, and more important, I want to point out the parallels
lowing a limit cycle like the one shown in the figure, alternating
between the reasons for temporal self-organization in nonlinear
periods of high output and rising capital with periods oflow output
business cycle theory, and those in my edge city story.
and falling capital, each phase of the cycle setting the stage for the
other. The limit cycle in Figure 7.4 is an equilibrium of the dynamic
The limitations of this story are pretty obvious, but I have system. But, is it the only equilibrium? No, there will always be a
described it to make two points. stationary equilibrium, in which the levels of output and the capital
stock are just right to support a level of investment that is just
First, I describe nonlinear business cycle theory partly to uphold
the honor of my profession. If you read books like Mitchell enough to offset depreciation (and trend growth, if you want to
make this a growth-and-cycles story). I have indicated this con-
Waldrop's Complexity: The New Science at the Edge of Order and Chaos,
stant-output, or more likely steady-growth, position as E in the fig-
you get the impression that economists are dull people who never
ure. Given that there exists an equilibrium in which output does
thought of trying to introduce nonlinear dynamics and self-organi-
not fluctuate, then, why do we focus on the fluctuating equilibri-
zation into their analysis until the visionaries at Santa Fe pushed
um? Because the stationary equilibrium is unstable. Jiggle the econ-
them. Well, nonlinear business cycle theory is more than 45 years
old. Ifit failed to catch on at the time, it was not because economists omy, and it will soar into a boom or plunge into a slump, and the
business cycle will commence.
were too unimaginative to realize the possibilities, it was because
68 SELF-ORGANIZING ECONOMY Temporal Self-Organization 69

Nonlinear business cycle theory, then, can be thought of as an PERCOLATION ECONOMICS

approach in which temporal order arises from instability; business
cycles, in this view, are like convection cells, hurricanes, or for that Recently Jose Scheinkman and Michael Woodford applied the
matter edge cities. principle of order from random growth to economic fluctuations.
We can push that last parallel a bit further. I argued in the first Their model was inspired by the work of Per Bak; I am indebted to
part that the existence of multiple urban subcenters depends on a their work for making it clear to me what Bak and his concept of
tension between centripetal and centrifugal forces, with the forces self-organizing criticality is all about.
of attraction fading more rapidly with distance than the forces of The Scheinkman-Woodford model envisages an economy
repulsion. If there were only centripetal forces, everything would whose input-output structure consists of a large number of "lay-
tend to form one big clump. There is a similar logic to the repeti- ers." Firms producing final goods form the top layer; they purchase
tive nature of the nonlinear business cycle. It is possible to imagine intermediate inputs from the next layer, which in turn purchases
a model in which firms want to invest if others do, and that is the inputs from the next layer, and so on. Firms at all levels hold inven-
end of the story; in such a model there would be high and low level tories of their own products, and they follow an "S-s" rule on pro-
equilibria, but no cycle. (The Big Push model of Murphy, Shleifer, duction: when an order for one unit comes in, if they have inven-
and Vishny, 1989, is a good example.) What drives Goodwin's cycle tory, they satisfYthe order by drawing that inventory down. If they
is the way that investment encourages investment in the short run do not have inventory, they produce two units - which requires that
but discourages it in the long run. In his model, firms tend to raise they place orders for inputs from two upstream firms - and put the
or cut their investment at the same time, because the level of invest- extra one into inventory. Orders for final goods arrive randomly.
ment determines the level of aggregate demand, which drives the Obviously an order for a final good will sometimes produce a
level of investment. This is the temporal equivalent of localized chain reaction of orders for intermediate goods. But how far will
agglomeration effects. A prolonged period of high investment, by this chain reaction spread? It depends on the levels of inventory. If
increasing the capital stock, eventually discourages further invest- inventories are high, orders will typically be met out of stock and
ment; this is the temporal equivalent of centrifugal forces like land will not generate additional orders. Ifinventories are low, each order
rent and congestion. will typically give rise to two more orders. Loosely speaking, there
The largely forgotten theory of nonlinear business cycles, then, is should be a critical level of inventories below which the expected
quite close in underlying structure to my story about edge cities; length of chain reactions becomes infinite.
both are stories of self-organization based on the principle of order Now comes the key insight: the level of inventories will always
from instability. tend to be near that critical level. Suppose that the level of invento-
Of course, nonlinear business cycle did not make it into the eco- ries is low. Then an order for a final good will typically set in motion
nomic canon. I have given some reasons for that failure, but let me a long chain reaction of orders for intermediate goods; because each
give another: the approach seems to predict too regular a cycle. The firm that is obliged to produce to satisfYits order also adds a unit to
model literally predicts a cycle that would tick like a macroeconom- its inventory, the level of inventories will tend to rise over time.
ic clock, but we have little difficulty in imagining that gradual Suppose, on the other hand, that the general level of inventories in
changes in structure, plus exogenous shocks, would make the tim- the economy is high. Then most orders will typically be met from
ing of the cycle acceptably irregular. No, the problem is that non- inventory - chain reactions will die out quickly and thus the level of
linear business cycle theory suggests that booms and slumps should inventories will tend to decline. So the economy will tend to evolve
be all more or less the same size - the economy should always surge to the edge of the critical level of inventories.
from the floor to the ceiling, then plunge down again. In fact, how- At a formal level, what I have just described is a familiar story in
ever, booms and slumps seem to come in all sizes. How is this pos- physics: it is a problem in percolation theory. The classic question in
sible? percolation theory, which gave the field its name, was the follow-
70 IELF-<DRGANIZING ECONOMY 71
Temporal Self-Organization

ing: how far wi wate penetrate into a porous rock? It turns out tuation implied by the old nonlinear business cycle theory. The dis-
that a good way model this is to think of the rock as a lattice of advantage is that it seems, if anything, even less realistic, makes less
holes, with so e bl t not all neighboring holes connected. contact with what seems to happen during a boom or slump, than
Percolation theo typ. cally thinks of each potential connection as the old theory. Nor is there, to my knowledge at least, any evidence
being either ope or losed with some probability p, then studies that the size distribution of business fluctuations actually obeys a
the size distribu on 0 connected regions. As one might expect, power law.
there is normally crit cal probability at which the expected size of What does obey a power law remarkably well, as I already pointed
connected regio bec mes infinite. Less obviously (indeed, some- out in the first part, is the size distribution of metropolitan areas;
what mysterious ), ju t about all percolation models have the fea- and we will come back to that example shortly.
ture that when p clos to its critical value, the upper tail of the size Meanwhile, however, back to the business cycle. I am agnostic
distribution of c nnec ed regions follows a power law; that is, the about both the Hicks-Goodwin and Scheinkman-Woodford fluc-
frequency of regi ns larger than any given size S is proportional to tuation theories. I admire them both and find both a little hard to
s-a, with a depen ing n the shape and dimensionality of the lattice. swallow as empirical propositions. At this point I prefer to regard
The point seems 0 be hat, if we think of water penetrating into the them both more as illustrations of how one might approach self-
rock from a parti ular ole, when p is close to its critical value this organization in time than as finished statements of how one actual-
spreading becom sap ocess of random growth whose rate is inde- ly ought to do it.
pendent of the v ume already reached and thus generates a distrib- And yet I cannot resist the temptation to try to use these
ution of "objects," co nected regions filled with water, that obeys approaches to explain something strange about the real world.
a power law.
It may not immediately obvious that Scheinkman and
PHASE LOCKING AND THE GLOBAL BUSINESS CYCLE
Woodford's mod I of t e economy resembles a porous rock, but in
fact it does. Thi of trms as the holes in the rock, with the con- Global interdependence is a favorite phrase these days for politicians
nection between 0 firms closed if the upstream firm has a unit in who want to sound sophisticated. Indeed, the Wall Street Journal
inventory, open i it does not. Then the question of how far a chain once published a whole article on the anthropology of Multilateral
reaction spreads equ valent to the question of how far a drop of Man, that special tribe that seems to like nothing better than to
water penetrates to t e rock. attend G7 meetings, Bank-Fund meetings, economic summits, and
And, if the pr abil ty that two holes will be connected is close so on.2 And yet it is a truism among those who actually work with
to the critical val e, th n the size distribution of connected regions, quantitative models of international macroeconomics that global
of the chain reac ·ons t at result from arrivals of final demand, will interdependence, at least as far as macroeconomic linkages are con-
be described by a ow r law. That is, even an almost steady flow of cerned, is quite modest.
demand will gen ate uctuations in production of all sizes. The arithmetic is straightforward. The United States and the
The physicist er B k has argued that many physical and social European Union each export about 2 percent of their output to
phenomena can e m deled as percolation systems that naturally each other. There is reasonably solid econometric evidence suggest-
tend to move to teed e of criticality. Bak claims, in particular, that ing that the income elasticity of import demand is around 2. So
the concept of' elf- rganized criticality" explains such striking consider the often stated view that economic recovery in the
empirical regular ies a the Gutenberg-Richter power law relating
the sizes of earth ake. to their frequencies.
2. According to the Wall StreetJoumal, the two basic tenets of Multilateral Man
The attractive ss 0 an approach like this for business cycle the- are that cooperation must be improved to facilitate coordination and coordination
ory is obvious: i exp ains how the economy can endogenously improved to facilitate cooperation. (Actually, that sounds a bit like a self-organiz-
generate fluctuat" ns ithout having the kind of typical size offluc- ing system!)
 72 SELF-ORGANIZING ECONOMY Temporal Self-Organization 73

United States helps Europe, by stimulating European exports. Well, nized cycle; a modest linkage would do as long as they were predis-
according to the arithmetic, a 1 percent rise in the U.S. GDP will posed to have cycles in any case and had fairly similar natural peri-
increase Europe's exports to the United States by 2 percent. That is ods.
a stimulus, all right - but it is only 0.04 percent of European out- What about the Scheinkman-Woodford cycle? Imagine two
put, pretty small change considering the rhetoric. economies in which, as in their model, firms purchase inputs from
And yet there is a paradox. If interdependence is so limited, why successive layers of upstream suppliers; but imagine that there is
do we so often have global recessions that strike many of the world's some linkage in the sense that with some modest probability one of
nations at the same time? Why did all of the industrial world share the upstream suppliers for any given firm will be in the other coun-
in not only the Great Depression but in the m~or recessions of try. Following the logic of their story, each of these two economies
1974-1975,1979-1982, and 1990-1992? will tend to evolve to the edge of criticality, so that the arrival of an
One of the luxuries of a format like this one is that I can include order for even a single unit of the final good will sometimes trigger
the kind of loose speculations that I could never write in a journal a large cascade of induced orders for intermediate goods.
and that I can explain, as I am doing now, that I do not necessarily Suppose, however, that there is a very large cascade. This will
believe in the theory I am advancing. So here is a crazy idea about generate a fairly large number of orders in the other economy, even
the global business cycle: it is an example of "phase locking." if the linkage is modest, and will therefore have a pretty good chance
Phase locking, otherwise known as mode locking, frequency pulling, of starting a large cascade in the other country as well. It seems to
or simple synchronization if two oscillators, is one of those phenomena me that even with fairly small "import shares," one might well find
that occur in wildly different contexts and at very different scales. It that large economic fluctuations tend to occur in both economies
was apparently first described by Huygens in the 17th century, who at once.
reported that two pendulum clocks placed back to back on the wall Anyway, all of this is a somewhat wild speculation. I include it
separating two rooms would synchronize their motions. The same because I think it is interesting and also to illustrate the way that
principle turned out to have a more cosmic significance when it was thinking about self-organization stimulates you to think about
learned that some asteroids have orbital periods that are related to familiar issues in novel, if not always sensible, ways.
that of Jupiter by the ratio of two integers; for example, Pallas has an The subject of self-organization of economies in time is a fasci-
orbital period 7/18 that of the giant planet. nating one, and I am sure that I have barely touched its surface.
It seems to me that phase locking might well occur in an interna- Nonetheless, that is all I am going to say on the subject. Instead, at
tional version of either the Hicks-Goodwin or the this point I want to return to space and explain at greater length why
Scheinkman-Woodford business cycle. Start with the and how the models I alluded to in the first part work.
Hicks-Goodwin cycle. Suppose that you envisage two economies
with modest linkages, each of which in isolation would have a cycle
like the one illustrated in Figure 7.4, with the length of the cycle not
too different. Now suppose that one economy plunges into slump
at a time when the other economy has accumulated almost enough
capital to slump as well. One can easily imagine that the large slump
abroad could provide the small adverse shock needed to end the
boom at home. Similarly, a country that is not too far from a spon-
taneous recovery might well be triggered into a somewhat prema-
ture boom by a large surge in foreign markets. Like the two back-
to-back clocks that started ticking in unison, the two economies
would not need to be very strongly linked to develop a synchro-
-----8-----
Models of Spatial Self-Organization

The writer Robert BencWey once explained that there are two
kinds of people in the world: people who divide people into two
kinds, and people who do not. I am one of the people who do; that
is, I believe very strongly in the use of simplified models to help us
think about social science.
In this book, I am following the usual, dishonest practice of sci-
entific exposition: in the first part I presented some intuitive con-
cepts in a breezy, isn't-it-obvious tone; only now am I about to
elaborate on those concepts with a set of more fully worked-out
models. There are good reasons for presenting ideas this way - it is
a much more pleasant experience for the listener or reader than a
direct plunge into the models, and therefore it is a much better way
for someone with an idea to hold the audience's attention. Why,
then, do I call it dishonest? Because it is completely misleading in
the picture it gives of how ideas are formulated or how one tells
good ideas from bad.
The truth is that interesting ideas almost never start with clear
but informal intuition, which is then backed up by a formal model;
intuition that comes that easily usually is not very interesting. The
actual process is far more confused. Here is how I do theory: I start
with a vague, murky intuition, much of which turns out to be
wrong.1 I then try to get a grip on that intuition by building a for-
mal model- a model that, on first pass, is usually an awkward thing

1. Here is an example of what I now believe to be a wrong intuition. It used to

seem to me that, to explain a tendency to form multiple agglomerations at some
distance from each other, one must suppose some mechanism whereby a success-
fullocation casts an "agglomeration shadow" over its neighbors, much as a large
tree prevents smaller nearby trees from getting sunlight. Well, this turns out to be

75
76 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 77

full of pointless complications. Then I begin a sort of process of polycentric structure, in which business is concentrated in several
intellectual self-organization, in which the model suggests a revi- spatially separated clusters. It is my desire to explain the emergence
sion of my intuition, the revised intuition suggests a way to simpli- of a polycentric pattern, rather thanjust a single agglomeration, that
fY the model, and so on. In the end, if I'm lucky, I end up with a gives rise to some subtleties - otherwise it would be just a matter of
story that seems obvious to other people as soon as they hear it. But agglomeration economies giving rise to agglomeration. Here, how-
it was not at all obvious to begin with, and the apparent obviousness ever, I am trying to explain how a tension between centripetal and
is the result of a lot of hard grinding away on the formal model. centrifugal forces creates a spatial pattern.
All of this is by way of explaining that, after the whirlwind tour To simplifY matters, we consider a one-dimensional city, with
of self-organization theory, it is time to settle down and do some households and businesses strung out along a line. We will want to
homework. In the remainder of this part I'm going to try to explain move back and forth between imagining that the line is actually a
the workings of three formal models of spatial self-organization. circle, so that there are no ends, or that it is simply very long, so that
The first is the "edge city" model described in the first part. The the ends are very far away from most places. Realistically, we should
second is a "self-organized" version of central place theory; that is, also think either of the line as consisting of some number of discrete
an attempt to show how a Losch-type lattice of central places could locations or of businesses dotted along the continuous space. In the
arise spontaneously through an invisible hand process. The third is spirit of physics, however, I will do neither and instead think of rep-
Simon's model of the power law on city sizes. Of these, the central resenting the distribution of business along the line as a continuous
place model is by far the most difficult and least realistic. It was also density, with A(X) the density at position x.
the model with which I started - as I said, the course of theorizing At any given time, the desirability oflocations will differ. Because
is not as smooth as later rewriting can make it seem. And as you shall this is supposed to be a model of self-organization, I assume that
see, there are reasons for including this model even though it is nei- there are no inherent differences among locations. Instead, desir-
ther realistic nor easy to understand. ability depends only on the distribution of businesses themselves. I
By the way, those of you who find it hard to sit through equations suppose in particular that a concentration of businesses at some
should not leave. I am going to try to explain how these models location z exerts both positive and negative effects on the desirabil-
work with only a few equations and figures; the heavy stuff is ity oflocation x, with both the positive and negative effects fading
reserved for the appendix. The truth is that I suffered through all of out with the distance Dxz between the two locations:
this in reverse order - mathematical grinding first, loose mathemat-
ical intuition next, grand vision last. But out of the goodness of my P(x) = fz[A(Dxz) - B(Dxz)]A(z)dz (8.1)
heart I will not make you share my pain.
So let us begin with a look at the underlying logic of the edge city In the language of economic geographers, P(x) may be described
model. as a "market potential" function. As you might guess, the results of
the model depend crucially on the way that those positive and neg-
ative effects fall off with distance. The simplest, and it turns out
THE EDGE CITY MODEL
most productive, assumption is that they fall off exponentially with
The purpose of what I am calling the edge city model is to explain, distance, with A falling off faster than B. It will be useful, however,
in as minimalist a way as possible, how the interdependent location also to imagine a situation in which the positive or negative forces
decisions of businesses within a metropolitan area could lead to a between businesses are constant up to some distance, then sudden-
ly disappear. In this case, ranges of positive and negative spillovers
between businesses would be well defined. Of course, this case can
all wrong: as we've already seen (Figure 1.5), widely spaced agglomerations can
emerge from a process in which neighbors initially seem to grow together. The
be viewed as an approximation to the case of gradually declining
actual reason why waves turn into spikes is quite different and explained below. effects, which is actually the way that I will treat it.
 78 SELF-ORGANIZING ECONOMY 79
Models of Spatial Self-Organization

Because locations differ in desirability, businesses will have an Moreover, if you change the parameters you will find that the
incentive to move. The easiest way to model that moving decision results change in sensible ways. Make the positive spillovers
is simply to assume that businesses move gradually away from unde- stronger, and there will be fewer, larger business districts; make the
sirable locations and toward desirable ones. If you want to write an negative spillovers stronger, and the districts get smaller and more
equation specifying that movement, you have to be a bit careful, to numerous. Make spillovers fade out faster with distance, and busi-
be sure that the rules imply that the total number of businesses ness districts tend to be closer together; and so on. In a rough way
moving in to better than average locations equals the number mov- the simulation approach does seem to convey an understanding of
ing out of worse than average places. Here is a particular rule that the process of self-organization.
works. First, define the average desirability - actually, the desirabil- And yet after doing a number of simulations one finds oneself
ity of the location occupied by the average business - as driven back to paper and pencil. Partly this is because even those of
us who believe that numerical methods are perfectly legitimate still
P = fxP(x)).,(x)dx (8.2) feel that we want something more - that we want to have models of
our models. But there is more to it than that. When you look at
Now assume that the density of firms at each location changes simulation results, you get a clear sense that some kind of hidden
according to the rule principle of order is lurking in this model about which you ought to
be able to offer more than vague intuition and the evidence of com-
puter runs.
d).,(x2
dt = y[P(x) - P]).,(x) (8.3)
The sense of a hidden principle emerges clearly from Figures 1.5
and 1.6. Each of these figures shows the results of a simulation in a
(What is that extra ).,(x) doing in there? Check the algebra and you model city with 24 locations around a circle, with different para-
will see that I need it to make sure that the total change in the num- meter values. (In fact, the only difference is that spillovers decline
ber of firms equals 0.) twice as rapidly with distance in Figure 1.6.) In each case, the initial
And that is it - a minimalist model indeed. But how do we ana- allocation of business to locations was done under a rule that
lyze it? ensured both that the starting distribution was very flat and that
deviations from that flat distribution were random. And yet, not
Fluctuations and Instability only do the runs show self-organization, they show self-organiza-
One way to analyze this model is simply to put it on the computer tion into perfectly regular structures. In Figure 1.5 the two business
and simulate it. Take some number oflocations around a circle - 12 districts are at locations 8 and 20; that is, exactly 12 apart, which is
or 24 are good numbers, because they are fairly small numbers with to say on opposite sides of the circle. In Figure 1.6 there are four
a fairly large number of divisors - make up some parameters, use business districts, at regular intervals around the circle. Let me now
some rule that includes a random element to create an initial spatial assure you that you can run the model over and over again with
distribution of businesses, and see what happens. That is not at all a these parameters, with a different random distribution of business
bad strategy. Indeed, the willingness to engage in computer-assisted each time, and you will get the same results. That is, for the para-
thinking, to do theory without theorems, is a hallmark of the meters in Figure 1.5 you will always get two business districts exact-
emerging literature on self-organization. ly opposite one another, for the parameters in Figure 1.6 you will
When you simulate the model, you will find that for a wide range always get four equally spaced districts. (Of course, which locations
of parameters our abstract metropolitan area does indeed engage in play those roles changes - the concentrations may be at, say, 1 and
a process of self-organization. Even if you create an initial spatial 13 instead of8 and 20.)
distribution of businesses that differs only imperceptibly from per- Where do these regularities come from? To understand that, we
fect flatness, all businesses eventually end up in just a few locations. need to use a trick very familiar to physical scientists, engineers, and
80 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 81

some econometricians, but that has had very little application in a city with many fluctuations at different frequencies? As long as the
economic theory so far: we need to think of the spatial distribution fluctuations are small, we can approximate our model by a set oflin-
of business as a Fourier series. ear equations; and as long as we can do this, we can, as it were, par-
The basic idea of Fourier series - and the basic idea is all that we tition our metropolis into a set of independently evolving shadow
will need - is that an irregular wiggle can be decomposed into the metropolises - one with a wavelength of 40 miles, one with a wave-
sum of a number of regular, sinusoidal wiggles at different frequen- length of 20, and so on.2 The linear approximation eventually
cies. If we think of A(x) as a wiggle along a line oflength L, then it breaks down, but we'll come to that later.
can be decomposed into nice, regular sine waves at frequencies 0 (a Let me now try to convince you of three propositions. First, very
horizontal line at the average value of A), 2njL, 4njL, 6njL, and so high frequency fluctuations will grow at best very slowly over time.
forth. Second, very low frequency fluctuations will actually fade out over
Now those regular fluctuations at different frequencies are not time - as long as the centrifugal forces are strong enough relative to
really there - they exist only in the mind of the theorist - but the centripetal forces. Finally, fluctuations at some intermediate fre-
nonetheless the regularity of the fluctuations is what explains the quency will grow fastest.
surprising order in the simulation results. To think about this, imagine in each case that the fluctuation has
How can this be? The key point is that fluctuations at different a peak at location 0 (we can relabel locations to make this true).
frequencies will tend to grow at diffrrent rates. More specifically, very Given this, is location 0 a desirable or an undesirable place for busi-
high frequency fluctuations will tend to go nowhere in particular, nesses to locate? Ifit is a desirable location compared with the aver-
and very low frequency fluctuations will tend to fade away. Only in age location, the density of businesses at location 0 will grow over
some middle range will the flucutations grow; and some preferred time, which means that the fluctuation at that frequency will grow.
frequency will grow fastest. What happens is that this preferred fre- If 0 is an undesirable location, the density of businesses there, and
quency comes to dominate the distribution; the regularity of the thus the size of the fluctuation, will shrink.
fluctuation at that frequency creates order in the self-organizing The desirability of a location depends, however, on the density of
city. businesses in other locations. Above-average densities of business
This sounds faintly mysterious; so let us take it step by step. within the range of positive spillovers will tend to raise market
The key insight here is that fluctuations at different frequencies potential of location 0; above-average densities within the larger
have different growth rates and that (subject to certain conditions) range of negative spillovers will tend to lower that potential.
the fastest growth rate is at some intermediate frequency. In the So let us look at some pictures of fluctuations at different fre-
appendix I show that this is true when the falloff of positive and quencies. Figure 8.1 shows a high frequency fluctuation, with a
negative spillovers takes a smooth exponential form. Here I restrict number of peaks and troughs within the ranges of both positive and
myself to showing that it is true given the simpler "range of influ- negative spillovers. Given such a fluctuation, will 0 be an especially
ence" story, in which effects extend without any falloff up to a cer- desirable or an especially undesirable location? If you think about it,
tain distance, then suddenly drop off to nothing. In fact, let us make you will realize that the answer is neither. Because there are so many
life even simpler: let us assume that the range of influence of nega- peaks and troughs within the ranges of spillover, favorable and unfa-
tive spillovers is precisely three times that of positive effects! You vorable effects will roughly cancel each other out. This means that
will see why that is convenient in a second; I hope that you will also
be convinced that the basic insight remains true more generally.
To ask whether and how fast a fluctuation grows, we imagine a 2. To be more precise: any sinusoidal fluctuation around the flat equilibrium
city in which the distribution of business can in fact be described as turns out to be an eigenfunction of this dynamic system; its growth rate is the
associated eigenvalue. And the whole process I am describing is precisely one of
a regular, sine-wave fluctuation around the average density. How representing a dynamic spatial model as a partial differential equation, whose solu-
can we do this when even the Fourier series representation gives us tions are all sine and cosine functions.
 0
~E0 J:
Q.J

.~"
82 -0.5
0.;;:
c::
'iii
Q.J
1.0
0.5
~c:: -10
Q.J
0.0
•

-1.5
SELF-ORGANIZING ECONOMY
Models of Spatial Self-Organization 83
~
-2.0
1.5
2.0 2.0

~ Positive spillovers __ ~ Positive spillovers

'iii 1.5 .••• •
t----------------------------------------------------------------

"
5i

Q.J
1.0

~ 0.5
Q.J

f;
E 0.0
g -----------------------------------------------------
g -0.5
.~
----------------------------------------------------------------
.~ -1.0
o Negative spillovers
----------------------------------------------------------------
-1.5

-2.0
Frequency Frequency

Figure 8.1 A High Frequency Fluctuation. Positive and negative

Figure 8.2 A Low Frequency Fluctuation. Negative spillovers will dom-
spillovers will nearly cancel out, preventing any growth. inate causing the fluctuation to shrink.

there will be little tendency for additional business to move to loca- location of above-average market potential. Consider that, through-
tion 0; this high frequency fluctuation will not tend to grow much out the range of positive spillovers, the above-average density con-
over time.
tributes positively to market potential at O. Moreover, once we go
By contrast, Figure 8.2 shows a very low frequency fluctuation. beyond to the range of negative spillovers, the business density is
Here all the relevant business densities - those that occur within
below average - and this also contributes to the desirability ofloca-
both the range of positive spillover and within the range of negative tion 0.3
spillover - are well above average. But if the negative spillovers are So a fluctuation at this frequency will grow over time. This pic-
strong enough - a condition that is easy to derive in any specific ture actually works only for the special case in which the market
model - then the net effect of these spillovers will be to make 0 an potential function is characterized by limited ranges of effects and in
undesirable location, a location with below-average market potentiaL which the range of negative effects is precisely three times the range
So a low-frequency, long-wavelength fluctuation like this will tend of positive effects, but the intuition holds up more generally.
to die out over time.
Provided that centripetal and centrifugal forces are not too uneven-
Finally, Figure 8.3 shows an intermediate case: a fluctuation that ly balanced, low frequency fluctuations will die out, high frequen-
just happens to be the right frequency so that all of the locations cy fluctuations will go nowhere in particular, but fluctuations at an
within the range of positive spillovers have above-average business intermediate frequency will grow and some particular frequency
density, all those within the range of negative spillovers that extends will have the fastest rate of growth.
beyond this zone have below-average business density. (Remember,
I have assumed for the moment that the range of negative spillovers
extends precisely three times as far as the range of positive effects; 3. Strictly speaking, there is a negative effect from the concentration of bus i-
ness in the middle, which is necessarily offset, two to one, by the effects of the
now you see why.) In this case, it is unambiguously clear that 0 is a troughs on either side.
It
III 84 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 85
Q) c
"0
'p
L..
0~Q)Q)
Ill)
Q)
Cl
1.0
0.5
~E 'in
gc
.~
0.0
-1.0
-0.5
-1.5
-2.0
1.5
Table 8.1
2.0

Growth rate:

~ Positive spillovers ~ Frequency Two-center case Four-center case

1 -.1991 -.3796
2 .0223 -.0215
3 -.0081 -.0228
4 .0096 .0058
5 -.0017 -.0034
6 .0047 .0049
7 -.0007 -.0006
8 .0027 .0033
Negative spillovers 9 -.0004 -.0003
10 .0018 .0023

Predictedgrowth ratesin smallcity model.

Frequency
set of parameters consistently yields two edge cities facing each
Figure 8.3 An Intermediate Frequency Fluctuation. If a fluctuation other, the second consistently yields four equidistant concentra-
has the right frequency,it will grow overtime, eventuallydomi-
nating the economy'sspatialpattern. tions.
But wait a second. Some of you may be saying, "You've been
We can now start to see how a regular pattern may emerge from telling us a story about nice, smooth, sinusoidal fluctuations. What
a disordered initial condition. Suppose that the initial distribution happens in Figures 1.5 and 1.6, however, is that all business con-
of businesses is almost, but not quite, uniform. We can represent centrates in only a few locations; we end up not with wiggles, but
the deviation of that distribution from flatness as the sum of a series with spikes (or, if you look at the whole space-time picture, we get
of regular, sinusoidal wiggles; but if the distribution is almost flat, dorsal fins rather than rolling ridges). How can this actually sharp-
these wiggles will be very small. Over time, some of the wiggles will edged picture emerge from a story about gradual divergence?"
grow; if this period of growth lasts long enough, the distribution
will become increasingly dominated by whichever wiggle grows From Wiggles to Spikes
fastest, no matter how small it may have been to start with. To understand the spikiness of the eventual distribution of business,
Indeed, it is possible to show numerically that this is precisely we need to think a bit more carefully about an issue that I have
what is going on in the examples illustrated by Figures 1.5 and 1.6. slipped past in my exposition so far: there is an adding-up constraint
In both examples, I set up a circular city of circumference 27r. The in this model. I am assuming that the total number of businesses, or
simulations that produced the figures were based on the full, non- equivalently the average density of business, remains unchanged.
linear model. It is easy, however, to take a linear approximation to This constraint can be ignored during the early phase of self-orga-
the model in the vicinity of perfect flatness, and to calculate the nization, in which fluctuations away from average density remain
implied growth rates of fluctuations at different frequencies. Table small, but it starts to playa crucial role at a later stage, when they
8.1 shows the results of that calculation. For the parameters used to have become larger.
generate Figure 1.5, a fluctuation at frequency 2 grows fastest in the Why can we, in effect, ignore the adding-up constraint during
linearized model; for the parameters used to generate Figure 1.6, a the early stage of self-organization? Because a growing fluctuation,
fluctuation at frequency 4 grows fastest. And sure enough, the first
 Models of Spatial Self-Organization 87
86 SELF-ORGANIZING ECONOMY

being a sine wave around the average business density, by definition 4.0
has as much area below the line as above it. Therefore, its growth
3.5
has no effect on the average density. If we think of the distribution
of business as the sum offluctuations, and we then imagine some of
these fluctuations growing while others shrink, the adding-up con- ~
'iij
3.0
c:
straint will be satisfied by each fluctuation and therefore automati-
~ 2.5
cally satisfied by the model city as a whole. I:lO
c:
But this process cannot go on forever, for one reason: the busi- '§ 2.0
•...
ness density at any location cannot become negative. A growing
sine wave will sooner or later imply negative values greater than the
$u:J
c:
'"
average density; at this point, the process that I have described must L 1.0
break down. 4
To see why this means that wiggles become spikes, look at Figure 0.5
8.4. To understand this figure, imagine that the deviation of the
density of business from its average level, here normalized to equal 0.0
1, has become completely dominated by a fluctuation at some one Location
frequency, and this fluctuation keeps on growing. I show what hap-
pens when the sine waves can grow no further without implying Figure 8.4. From Wiggles to Spikes. As the spatial distribution of activi-
negative values. Clearly, the sine wave starts to develop "flat bot- ty gets increasingly uneven, it starts to change from a smooth
sine wave to a pattern of widely separated spikes,
toms" - ranges where the density of business is zero. But these flat
bottoms in effect truncate part of the below-average part of the sine ly even sprawl that many people envisaged a couple of decades ago?
wave; if the curve retained its shape, there would be more distribu- It seems natural to think that a center surrounded by a void must be
tion above average than below, which is possible only for the chil- directly interfering with the development of nearby areas - that, to
dren of Lake Wobegon. Obviously, then, something must give: the use the felicitous phrase proposed by Brian Arthur, it must exert an
growing peaks must grow, not only at the expense of the troughs, "agglomeration shadow." In fact, that is not at all what happens
but by pulling in businesses even from locations that had been here. The areas close to the emergent concentrations have higher
growing up to that point. As a result, the peaks get sharper. And this market potential than places farther away. They just do not have
process continues until the wiggles become spikes. high enough market potential to compete for the limited pool of
I think - I hope - that this image is clear. It may even seem obvi- businesses. The spikiness of the eventual distribution of business is
ous. But there is a fairly subtle feature about the later stages of self- a consequence of the adding-up constraint. That is, it is a general
organization in the edge city model that may be worth emphasiz- equilibrium effect.
mg. This, then, is how what I have called the edge city model works. I
Let us ask the following question: Why does each concentration hope you share at least some of my satisfaction at seeing how a very
of business in this model end up surrounded by a clear area without simple model can yield complex and subtle dynamics - and con-
businesses? Why, in the real world, do our metropolitan areas devel- versely, how dynamics that at first sight appear hopelessly compli-
op reasonably compact business subcenters, rather than the relative- cated turn out to be governed by a surprising principle of order.
Moreover, I believe that the model is basically right as an account of
4. Actually,given the dynamic rule I have assumed, there is no sudden break- how metropolitan areas engage in spatial self-organization.
down; rather, the nonlinearity built into the dynamics becomes more and more Yet there is one big objection to this model: although the con-
important as the distribution of business divergesfrom flatness. clusions do not follow in a trivial way from the assumptions - it is
88 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 89

not simply a matter of agglomeration economies producing development economies, preferred sites will be those with good
agglomerations - the model is still very much in a reduced form. I backward and forward linkages. But where will a firm find good
assume that there are centripetal and centrifugal forces that decline access to markets? Precisely where many other firms also choose to
with distance, but what are these forces? locate. Thus there is a circular process whereby firms tend to con-
It is possible to speak loosely about what they must be. Joel centrate near other firms, producing agglomerations. The agglom-
Garreau, whose book Edge City has been a useful guide throughout eration economies are not assumed; they are derived. Furthermore,
this book, attributes the rapid growth of subcenters once they reach they are a purely emergent property. There is no immanent cen-
the "critical mass" of 5 million square feet of office space essential- tripetal force at the level of the individual firm; things that we can
ly to the growth of local business services; for example, the ability callforward and backward linkages appear when large numbers of firms
to support a local luxury hotel. He also stresses the role ofland costs interact.
and traffic congestion in limiting the growth of edge cities. So these Will a model along these lines always produce a single agglomer-
forces are not mysterious. But so far I have not explicitly modeled ation? No, against the centripetal forces must be set a centrifugal
them. force. Firms provide each other with markets, but they also compete
Let us talk about how we might fill that gap. for markets, in particular the market provided by the dispersed agri-
cultural population. The tension between centripetal and centrifu-
A CENTRAL PLACE MODEL gal forces at least suggests the possibility of an emergent pattern with
multiple centers.
A few years ago I gave a set of lectures on the subject of economic You can guess the next step. Let us simulate it on the computer.
geography. The little book that came out of those lectures, Begin with an almost but not completely smooth distribution of
Geography and Trade, had at its core an approach to location that in manufacturing on a - you guessed it! - circle, and simply let the
effect derived centripetal and centrifugal forces from microfounda- model evolve according to some simple dynamic rule. It turns out
tions; that is, spillovers themselves became an emergent property. that the simulation is much more complex in this case than in the
Here is the basic approach. I envisaged an economy in which edge city model, because you must solve a general equilibrium
there was an immobile, geographically dispersed sector, "agricul- model at each step: given the spatial distribution of manufacturing,
ture," and a mobile sector, "manufacturing." Manufacturing con- you must solve a set of simultaneous nonlinear equations that deter-
sisted of many differentiated products, with economies of scale at mine wage rates and prices at all locations, use the result to deter-
the plant level. (The market structure was therefore monopolistical- mine the incentives for firms to move, and so on. Nonetheless, it
ly competitive.) There were transport costs for manufactured goods, can be done. And, as long as neither the manufacturing share nor
though for convenience I assumed away any transport costs for agri- the degree of scale economies is too large, the result is the emer-
cultural production. There were no pure externalities, no direct gence of a regular pattern of evenly spaced manufacturing centers.
spillovers between firms or households. Figure 8.5 shows the results of one such run, on a circle with 12
How can such a model generate agglomerations? The answer lies locations. The X axis shows the locations; the Y axis shows time;
in the interactions among economies of scale, transportation costs, the Z axis shows the share of manufacturing in each location. It
and factor mobility. Because of economies of scale, each product looks very much like Figure 1.5, which showed a run of the edge
will be manufactured in only a few locations. 5 The favored locations city model. Again, we see an almost flat surface begin to undulate,
will tend to be those with good access to markets and with good then gather itself up into two dorsal fins.
access to the goods produced by other plants - in the language of Given this picture, you will not be surprised to find that the same
approach that allowed us to understand the origins of order in the
5. Strictly speaking, in the monopolistic competition model each individual edge city model works here as well. We can linearize the model
variety is produced in only one location. around the unstable equilibrium in which manufacturing is evenly
90 SELF-ORGANIZING ECONOMY Models of SpatialSelf-Organization 91

ing function of distance, we imagine instead a world in which trans-

port costs are zero up to some distance, then infinite thereafter. Say
this distance is 50 miles. Then the forward and backward linkages,
Il)
ti which are the centripetal forces in the model, will extend only 50
miles.
".
ti
But what about the centrifugal force - the competitive pressure
~ that firms exert on each other? Well, even if manufactures can be
lIS
~ shipped only 50 miles, any firm that is less than 100 miles away will
II> "l

l:>O
c: C still be in a position to compete for some of your customers. So the
or::
::J range of centrifugal forces, of the negative spillovers from the pres-
t: '"
~::J C ence of rival firms, will indeed extend farther than the range of the
c:
lIS positive spillovers from larger markets. And that is all that we need
:E _
c to get multiple agglomerations.
At this point let us pause and ask what we have just developed a
model o£ The answer, I think, is that we have just seen how a cen-
tral place system along the lines envisaged by Losch can emerge via
the invisible hand. The essence of central place theory is the trade-
offbetween economies of scale, which provide an incentive to con-
centrate production in a limited number of sites, and transportation
costs, which provide an incentive to multiply sites so as to stay close
to consumers. Losch argued that to make the best of such a tradeoff
Figure 8.5 Evolution of Central Places. The dynamicsof a centralplace one should have a regular lattice of central places (with hexagonal
model show the same "'59 Cadillac" picture as the edge city market areas in a two-dimensional economy). He was unable, how-
model. ever, to show how such a lattice might emerge through a decentral-
ized, unplanned process. Well, here we have an example of just such
spread across the landscape and represent the deviation of the spatial a process. I have demonstrated the emergence of a regular lattice
distribution from flatness as a Fourier series. It then turns out that only for a one-dimensional economy, but I have no doubt that a
there is a particular frequency of fluctuation that grows more rapid- better mathematician could show that a system of hexagonal market
ly than any other and tends first to dominate the distribution then areas will emerge in two dimensions.
to gather itself up into spikes. Moreover, it seems to me that we can adapt central place theory
What may puzzle you, however - it puzzled me at first - is why the same way that Edwin Mills adapted the von Thiinen model: we
the model so nicely produces multiple agglomerations. In the edge can recast it as a model of the internal structure of urban sprawls,
city model, the key to the result was the assumption that centripetal with suburban families as farmers and office buildings as factories. I
forces falloff more rapidly with distance than centrifugal forces. We do not want to go through the details of such a reinterpretation, but
made no such assumption here. So what about the model creates an I am sure you get the idea. The important point is that we see how
equivalent effect, and is it plausible? the mysterious centripetal and centrifugal forces of the edge city
I can offer a heuristic explanation that suggests that multiple model could arise from pecuniary externalities, alias market-size
agglomerations are indeed a natural outcome in a model in which effects. People like Philip Anderson should be gratified: in this
centripetal forces arise from market linkages. Suppose that instead view, the agglomeration economies that drive the formation of edge
of realistically representing transportation costs as a steadily increas- cities are entirely an emergent property, with not even a hint of such
92 SELF-ORGANIZING ECONOMY 93
Models of Spatial Self-Organization

a property at the level of individual firms. cities of size S reaches a constant ratio to the population. This can-
So I would offer a double reinterpretation of this model that not, of course, ever be exactly true, because the largest city will
started out as being about farmers and manufacturers. First, it always tend to grow larger over time. But let us disregard this for a
implies the formation of Losch-type central places and may thus be moment and suppose that ns/P does in fact tend to approach a con-
taken as a formalization and vindication of central place theory. stant value for all S.
Second, as an empirical matter, it can be interpreted as a model of How does ns/P change? The denominator grows with the gener-
what happens inside metropolitan areas as well as (or indeed in pref- al growth in the population. The numerator can change for two rea-
erence to) a model of the formation of such areas themselves. sons. First, a "lump" can attach itself to a city of size S - 1, turning
it into a city of size S and adding 1 to ns. Second, a lump can attach
SIMON'S URBAN GROWTH MODEL itself to a city of size S, turning it into a city of size S + 1 and reduc-
ing ns by 1. But recall that with a probability n an arriving lump
Finally, let me turn to the very different story of urban growth pro- forms a new city. If it does attach itself to an existing city, then it has
posed by Herbert Simon to explain the stunning empirical success a probability nsS/ P - the combined population of cities of that size,
of a power law in describing city size distributions. I have some seri- divided by the total urban population - of attaching itself to a city of
ous problems with this model; but it is a thing of beauty and also the size S. Writing the expected rates of change and fudging a bit over
best explanation anyone has yet proposed, so let me set out the the discontinuous growth of P, we have
model before criticizing it.
How Random Growth Produces a Power Law
E[d(ns/P)J =~[(1-n)nS_1(S-1)-(1-n)nsS-ns] (8.4)
Recall again Simon's story. It is one oflumps and clumps: as each
new "lump" of population arrives, it either starts a new city (with
probability n) or attaches itself to an existing "clump"; the proba- But if the system approaches a steady state, the right-hand side of
bility that each clump attracts the lump is proportional to that (8.4) must approach o. This implies that in the steady state we must
clump's population. Thus we have a process of random growth of have
clumps, with the expected rate of growth independent of clump
SIze. ns _ (l-n)(S-l) (8.5)
To see why this story works, let me introduce some notation. Let nS-1 - (l-n)S + 1
S be city size, measured in "lumps"; let P be the population, also
measured in lumps; let Ns be the number of cities (clumps) of size S This may be rewritten as
or larger; and let ns be the number of cities of precisely size S. If city
sizes obey a power law, then Ns is proportional to s-a, and ns is pro- ns-ns-l n-2 (8.6)
portional to s-a- 1. Or to put it in a different and useful way, we ns-l (l-n)S + 1
want a story that implies an elasticity of n with respect to S of
-a-1. If we restrict our attention to large cities, we may approximate
It is natural, though not crucial, to suppose that the population the discrete distribution of city sizes by a continuous one; in partic-
grows at an expected constant rate r through Poisson addition of ular, it will be approximately true that
lumps.
To analyze the model, we make a crucial guess: that the urban dn
ns- ns-1 = dS
(8.7)
system tends to approach a steady state, in which the number of
94 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 95

Substituting this approximation back into (17), we have an part by his observation that firm sizes, like city sizes, follow a power
expression for the elasticity of n with respect to S: law, and his conclusion that the U-shaped average cost curve of the
textbooks was nonsense. Yet a closer consideration of his own
dn S 1C-2 model suggests that it is not consistent with strictly constant
(8.8) returns.
dS n 1-1C+1jS
To see why, suppose that there were hardly any economies of
However, when S is large - that is, when we look at the upper tail scale. This would imply, in Simon's story, that the "lumps" of pop-
of the city size distribution - this simplifies to ulation would be very small. But suppose that the lumps were very
small relative to the size of most cities. Then the law oflarge num-
dnS 1C-2 bers would apply: all cities would grow at the same rate - we would
-=-=-a-1 (8.9)
dSn 1-1C have growth, but not random growth - and whatever size distribu-
tion we started with would be preserved. In particular, if cities start-
And so in the end we have the implied exponent on the power ed out of roughly equal size, there would be no way to get to the
law: wide range of inequality implied by power laws with small expo-
nents. To get convergence to a power law, then, the lumps must not
1 be too small. Or, to be more precise, the smaller are the lumps, the
a=l-1C (8.10)
greater the percentage population increase needed to produce a
smooth power law. So the fact that the United States has actually
It is a remarkably elegant result. And it seems to suggest a very achieved a quite smooth power law distribution since it stopped
simple explanation not only for the existence of a power law, but for generating lots of new cities - say, since 1860 - is an indication that
the exponent of 1: new units of economic activity almost always the lumps cannot be too small. 6
form in existing clusters. On the other hand, the lumps cannot be too large, or it would be
I believe that this insight must be essentially right - although impossible to get a smooth distribution of city sizes in any case. The
some problems with Simon's story still give me sleepless nights. I rank-size rule works spectacularly well for American cities over a
shall come back to those problems in a moment. First, however, let range of roughly two orders of magnitude, from cities of around
me point out some implications of Simon's model that may not be 200,000 up to metropolitan New York, with almost 20,000,000.
entirely obvious.
6. In the simulation shown in Figure 3.3, I allowed the urban population to
Implications of the Model grow by a factor of 100. Is this a reasonable number? In reality, America stopped
forming major new cities east of the Rockies around 1860, when the so-called
The model that Simon suggested to explain the city size distribu- manufacturing belt locked into place, a development that I described in Geography
tion is in sharp contrast to the models I have suggested to explain and Trade and attributed to railroads, manufacturing scale economies, and the
the formation of central places. Central place theory depends cru- diminishing importance of agriculture. Since then the urban population has
increased by a factor of only 20. The 1860 city size distribution was not, howev-
cially on economies of scale - on external economies in particular, er, a set of equal size "seeds," like the initial condition I used in my simulation.
although these external economies may be an emergent conse- Instead, it was a highly unequal clumping, with big cities at strategic locations like
quence of internal scale economies. But I have suggested that power the mouth of the Hudson. And I am reasonably sure that convergence to a smooth
laws typically arise from processes of random growth that are power law should take place faster when one starts from a suitably uneven initial
condition. I like to imagine that someday, when we are more sure about the ori-
approximately scale independent. How can these views be recon-
gins of order in urban systems, economists will be in a position to ask whether any
ciled? Does the evidence that city sizes follow a power law imply given theory is sufficient to explain the evolution of order in the time available,
that there are no economies of scale in city formation? and even to use that evolution to make inferences - just as cosmologists have used
Herbert Simon seems to have thought so. Indeed, Simon's disen- the observed large-scale structure of the universe to infer the existence and nature
of "dark matter."
chantment with standard economics seems to have been driven in
96 SELF-ORGANIZING ECONOMY Models of Spatial Self-Organization 97

This suggests that the size of a lump - in effect, the minimum effi- out not to be a problem. The reason is that, although the biggest
cient scale for a city - cannot be more than about 200,000 people. cities keep getting bigger, very big cities eventually hold a trivial
(For what it is worth, the biggest city in Figure 3.3, which shows a share of the total population, so that almost all population growth
run of the Simon model with 1C= 0.2, has 141 lumps.) takes place within existing size classes.
There is an enormously tempting possibility here: to identify the Let us be a bit more formal (although this argument is far from
lumps that make up metropolitan areas with Garreau's edge cities. If rigorous, as it is slightly circular). Suppose that the size distribution
we could do this, we could nicely reconcile central place models really were a power law throughout, with the number of cities big-
that rely on increasing returns - and the anecdotal evidence that ger than S proportional to S- a, and with Smin the smallest possible
suggests substantial increasing returns at the level of urban subcen- city. Then the share of population in cities larger than any given S
ters - with the more or less constant returns suggested by the met- would be (SjSmin)l-a. As long as a> 1, really big cities will have a
ropolitan size distribution. I have my doubts about whether we can trivial share of the population and will therefore attract hardly any
actually make this fit - it looks to me as if edge cities are too big to of the growth in population; so it is reasonable to think of cities
do the job. After all, a power law really should work well only for below some large size as approximately reaching a steady state once
the upper tail of the city size distribution; so the smallest cities for the biggest cities have grown big enough.
which it works should actually consist of several lumps. Yet Garreau But what if a equals 1, as it should if no new cities form (and as
suggests that an edge city normally services a population of at least it does in fact in U.S. data)? Then we have an absurd result: if the
250,000, which is a bit bigger than the smallest cities that still obey power law really held, all population increase should occur in cities
the rank-size rule. The fit does not seem right - yet it is close of a size greater than S, no matter how large S is. This is nonsensical-
enough to be tantalizing. what it really means is that the biggest cities keep getting bigger, and
In any case I have a problem with this whole analysis, which I their growth involves so much of the population increase that the
may as well share. distribution never gets to look anything like a power law.
You might say that this is only a limiting case; let us just imagine
An Unresolved Problem
that almost all, rather than actually all, lumps attach themselves to
In the simulation run shown in Figure 3.3, I set the probability of existing clumps. But this is not really an answer, because when a is
forming new cities at 0.2 and got a power law distribution with a very close to 1, the model will not start to approach a steady-state
slope fairly close to -1. But the real city size distribution has a slope distribution until the biggest cities are very big indeed, which
that is almost exactly 1, and Simon's model suggests that this is what would not happen unless the process of random growth has gone on
we should expect with 1C = O. Why didn't I run the model with 1C = a long time. To get the smooth result in Figure 3.3, I had to let the
O?
urban population increase by a factor of 100, which is just barely
Because it doesn't work, that's why. Try it if you are a masochist reasonable for the United States. This, however, was for a simula-
(I wasted quite a few days trying to find the error in my program tion in which the probability of new city formation was 0.2. If! had
before realizing that it was not a computer problem). No matter tried to set that probability any lower, to reproduce the incredible
how long you run a simulation of Simon's model with all lumps exactness with which a seems to equal 1, I would have needed a far
attaching themselves to existing clumps, you will never get a larger population increase.
log-linear relationship between city size and city rank. At the moment, I have no resolution for this problem.
The reason is hidden in the logic of the model. To justify the Nonetheless, I remain strongly attached to two propositions. First,
emergence of a power law, Simon assumes that the size distribution the power law on city sizes is very real and tells us something very
of firms approaches a steady state. As I have already pointed out, this important about our economy. Second, some kind of random
never happens completely, because the biggest city keeps on getting growth process is far and away the most likely explanation. And
bigger. But as long as at least some new cities are formed, this turns Simon's model is so wonderfully elegant an approach that I contin-
ue to regard it as the best game in town.
-----9-----
Concluding Thoughts

The world is full of self-organizing systems, systems that form

structures not merely in response to inputs from outside but also,
indeed primarily, in reponse to their own internal logic. Global
weather is a self-organizing system; so, surely, is the global econo-
my.
Are there universal laws governing self-organizing systems? I do
not know, and I am surely not the person to find them. There are,
however, some characteristic ways in which systems are known to
organize themselves, and scientists have become increasingly aware
that the existence of these characteristic styles of self-organization
leads to parallels between phenomena that might seem completely
distinct. The physicist John Hopfield argued, in a now famous
paper, that the spin-glass models of condensed matter physics could
shed considerable light on how memory and learning take place in
the brain; it should not then surprise us that Schelling's model of
segregation looks a lot like a spin-glass, too. Per Bak has argued that
similar random growth processes can explain the size distributions
of everything from earthquakes and forest fires to species extinc-
tions; it is not unreasonable to suppose that something along the
same lines may apply to city sizes as well.
In this book I have suggested two principles of self-organization
that seem to me to be particularly useful in eXplaining economic
behavior. The first principle is that of order from instability: when a
system is so constituted that a flat or disordered structure is unsta-
ble, order spontaneously emerges. The classic physical example of
this principle is convection; the global atmospheric circulation,
which is primarily a giant convection machine, spontaneously orga-
99
100 SELF-ORGANIZING ECONOMY

nizes itself into weather patterns, essentially because a smooth cir-

culation of air from poles to equator and back again is unstable. The
business cycle theorists of the 1940s and 1950s thought that eco-
nomic fluctuations arose out of a similar logic. I am not sure
whether they were right, but I am convinced that this is in fact the
right way to think about how metropolitan areas evolve their far
from uniform structure.
The second principle is that of orderfrom random growth. Objects of
many kinds, from earthquakes to asteroids, obey a power law size
-----10-----
distribution. The best explanation is that these objects are formed Appendix:
by a growth process in which the expected rate of growth is approx-
imately independent of scale, but the actual rate of growth is ran- The Evolution of Central Places
dom. The size distribution of cities exhibits sustained empirical reg-
ularities that are every bit as striking and consistent as those of the
physical processes.
What good is the idea of a self-organizing economy? Well, any- Discourses are no place for dense algebra. Yet the idea of self-orga-
thing that makes us reconsider the way we think about economics nization is essentially a mathematical one and needs to be backed up
is bound to have policy relevance, but at this point I have no rec- by some reasonably serious algebraic grinding. So here are the mod-
ommendations to offer. I can see vaguely how my spatial models els underlying what I have said about edge cities and central place
could help in urban planning. To be honest, I can see somewhat theory.
more clearly how they could be helpful in real estate speculation! If
something like the Scheinkman-Woodford model turns out to be
URBAN MORPHOGENESIS: THE EDGE CITY MODEL
really useful in looking at economic fluctuations, it could similarly
be useful both for policy and for forecasting. The objective of this model is to illustrate in as simple a fashion as
Luckily for me, a set of speculative discourses like this need not possible how interdependent location decisions by individual busi-
provide anything of immediate use. I hope that you have found the nesses can lead to a self-organization of space. The basic idea is that
idea of the self-organizing economy as interesting and exciting as I there is a tension between "centripetal" forces that pull businesses
do. More to the point, I hope that I have convinced you that the together and "centrifugal" forces that drive them apart. What we
crazy things I have been saying might even be true. want to see is how, and under what conditions, this tension can lead
to the kind of polycentric pattern that characterizes modern metro-
politan areas.
We simplify the problem by assuming a one-dimensional metro-
politan area. It will be useful to move back and forth between
explicitly representing a metropolitan area as a circle, with a cir-
cumference normalized to 2n, and treating the metropolitan area as
if it were of infinite extent - which may simply be seen as an
approximation to self-organization on a very large circle. The ana-
lytics of the infinite metropolis (I guess that is Los Angeles) are
somewhat simpler and more intuitive, but the numerical examples
must focus on the slightly harder finite city case.
101
 102 SELF-ORGANIZING ECONOMY 103
Appendix: The Evolution of Central Places

We begin by specifYing the interdepedence between firm loca- Notice that the Il(x) in equation (A.4) serves two purposes. It has
tions. Let x be some location on the line, and let Il(x) be the density to be there to ensure that the sum of changes add to zero. It also
of firms at that location. We assume that the desirability of any ensures that the density at any point will never fall below zero, so
given location - which, following the tradition of Harris (1954) and that we need not introduce the requirement that there be a non-
Lowry (1965) we can call the location's market potential - depends negative number of businesses at each location as an explicit con-
both positively and negatively on the density of firms at other loca- straint.
tions. Both forces decline with distance, but the positive forces Unfortunately, the necessary presence of this term also implies
decline faster. In particular, we assume that the market potential that the model is nonlinear, as indeed the nonnegativity constraint
equation takes the form also does. So in spite of the simplicity of the model's description, we
seem to be dealing with a nonlinear dynamic model with a contin-
P(x) = t[A exp (-r1Dxz) - B exp (-r2Dxz)]Il(z)dz (A.1) uum of variables. How is such an apparent monster to be analyzed?
One answer is to turn to the computer. It is a simple matter to set
where Dxz is the distance between x and z, and where r1>r2' Here, up a circular model with a fairly large number of discrete locations,
A represents the strength of the centripetal, agglomerative forces; B, start with a random distribution of business across these locations,
the strength of the dispersing, centrifugal forces; and r1 and r2 rep- and simply see what happens for a number of values of the parame-
resent the rates at which these forces dissipate with distance. By ters. Figures 1.5 and 1.6 show sample results for a 24-location
making r1>r2>we introduce the essential difference in range that model. (It is a good idea to let the number oflocations have a large
makes multiple agglomerations possible. number of divisors.) One might then look for "empirical" regular-
It seems reasonable to suppose that agglomeration will take place ities in the results of these "experiments."
only if A is sufficiently large relative to B, but that multiple subcen- However, the experimental mathematics is so successful at yield-
ters are possible only if B is sufficiently large relative to A. We will ing regularities that we are driven back to to pencil and paper! For
see shortly that this is correct; indeed, the interesting range is where any given set of parameters, the model produces not only multiple
centers but a consistent number of centers, more or less evenly
r1 A r2
->->- (A.2) spaced around the circle. This surprising predictability of behavior
r2 B r1 becomes even more striking when we choose a rule for setting ini-
tial business shares that yields a fairly even distribution around the
To complete this minimalist model, we need only add some circle. For Figures 1.5 and 1.6, the share i of business at any location
dynamics. Let us assume that businesses gradually migrate toward i was set equal to
locations with above-average market potential P and away from
those with below-average potential. The average market potential k + Ui
(A.S)
may be defined as Ili = L,.(k
'} + u)
P = fxP(x)ll(x)dx (A. 3) where ui was a drawing from the uniform distribution between 0
and 1, and k was a "smoothing" parameter; the larger is k, the clos-
A simple dynamic rule that obeys the adding-up constraint that er the initial distribution will be to flat. And, once k is set fairly high
the total number of businesses remain constant is (it was 5 for those figures), one consistently gets the same number
of equally spaced, equal size centers for any given parameters.
How is this result to be explained? By focusing not on the whole
dll(x)
dt = r[p(x) - P]Il(x) (AA)
path to long-run equilibrium but on the process of divergence away
from a flat, uniform spatial structure.
104 SELF-ORGANIZING ECONOMY Appendix: The Evolution of Central Places 105

Linear Dynamics: The Large City Case

Substitute this back into (A.7), and we get a rather simple expres-
By the large city case I mean a city sufficiently large that we can think sion for the change in the density:
of it as if it were a line of infinite length. This means ignoring two
annoying details that arise when you deal with finite size circles:
that the distance between two points can never exceed half the cir- (A.ll)
dA'(X)
~ -- r[A ~
rl+tjJ2 - B rJ+tjJ2
~]A(X)
cumference, and that the shortest route to another point may
involve going either clockwise or counterclockwise. In the large city
case we integrate over z as if it ranged from minus to plus infinity That is, if the density could be described by a fluctuation at a fre-
and measure distances by quency tjJ, then that fluctuation would grow at a rate g(tjJ).
This may seem a moot point, but it is not, because if the actual
Dxz = Ix-zl (A. 6) deviation of the density from flatness can be described by a sum of
sinusoidal fluctuations, then the growth of that deviation can be
Suppose, now, that we consider the behavior of the model thought of as the sum of the fluctuations multiplied by their char-
defined by equations (A.1), (A.3), and (AA) not in general, but only acteristic growth rates:
in the neighborhood of a completely uniform spatial distribution of
business - A(X) = A for all x. (It is helpful to choose units so that at dA'(X) = Lig(tjJ;)A'i(x) (A.12)
this uniform density A(X) = 1.) In this case the model can be given a dt
linear approximation and can indeed be collapsed into a single (par-
tial differential) equation: Therefore, we can think of the evolution of this spatial economy
as consisting of the parallel growth of a set of hypothetical spatial
economies, each of which is characterized by a regular distribution
d;;x) = rJ:[Ae-rtIX-ZI- Be-r2 IX-ZI] (A(Z) -l)dz (A. 7) of business with a different wavelength.
Now suppose that the initial distribution of business density was
Let A'(X) = A(X) - 1. Then it is a general principle that this devia- very flat. As the fluctuations grow over time, the deviation of the
tion of the business density from the average can be represented as density from flatness will not only increase but tend increasingly to
the sum of a number of periodic fluctuations at different frequen- be dominated by whatever frequency fluctuation grows fastest. So
Cles: the crucial question is the shape ofg(tjJ).
From inspection of equation (A.11), three things about g(tjJ)
A'(X) = LiA'i(x) (A. 8) should be apparent. First, at very high frequencies, tjJ~oo, g(tjJ)
approaches zero. That is, very high frequency fluctuations go
where nowhere in particular. Second, at very low frequencies, tjJ~O, g(tjJ)
will become negative as long as Alrt <Blr2' Finally, at sufficiently
A'i(X) = aisin(tjJix) + bicos(tjJix) (A. 9) high tjJ , g(tjJ) will be positive as long as Art>Br2' Thus as long as the
criterion (A.2) is satisfied, g( tjJ) should be negative at low frequen-
Now imagine for the moment that the distribution of business cies, approach zero at high frequencies, and reach a positive maxi-
was completely described by a fluctuation at a particular frequency mum somewhere in between.
tjJ; and suppose without loss of generality that this fluctuation had a The Fourier series representation of a distribution along a very
peak at x = 0, so that we could simply write long line will involve many different frequencies; in effect, the spec-
trum will be covered quite closely. So in the infinite city case, we
A'(X) = h(t)cos(tjJx) (A. 10) can simply think of the deviation of the economy from flatness as
 106 SELF-ORGANIZING ECONOMY 107
Appendix: The Evolution of Central Places

being dominated over time by whatever frequency maximizes the tions is assumed to consist of two sectors: a constant returns, geo-
function g(¢J).The wavelength that corresponds to that frequency is graphically immobile sector ("agriculture"), and an increasing
the normal distance between edge cities! returns, monopolistically competitive, geographically mobile sector
("manufacturing"). When one adds transportation costs in the
The Small City Case
manufacturing sector and some simple dynamics, models of this
For simulation analysis, it is necessary to work with a "small city": type exhibit spontaneous spatial self-organization: even if allioca-
a circular array oflocations with a distance small enough that only a tions are identical in resources and technology, manufacturing firms
few concentrations develop. It is convenient, in fact, to normalize have an incentive to concentrate production close to the markets
units so that the circle is of circumference 2Jr. and supplies that other manufacturing firms provide, thus produc-
When one works with the linearized version of such a small city ing a "centripetal" tendency toward agglomeration. Working
model, two complications appear. First, only fluctuations that go an against this centripetal tendency, however, is the "centrifugal" pull
integer number of times around the circle are possible. For a circle of the immobile agricultural sector.
of circumference 2Jr, this means that the allowed frequencies are 1, In a two-location model, the tension between centripetal and
2, 3, ... Second, the maximum distance of any location from any centrifugal forces can be treated analytically; one can derive a crite-
other is Jr. This makes the expression for the growth rate of a regu- rion, depending in an economically meaningful way on the para-
lar fluctuation somewhat more complicated. It now takes the form meters, that determines whether or not manufacturing concentrates
in one location. Beyond this case, however, it becomes very diffi-
cult to derive analytical results. Simulations show that there may be
(A.13) equilibria with multiple manufacturing concentrations; they also
g(¢J) ['1 H1-B
= yrA rz Hz]
indicate that, as the number oflocations grows, there typically start
to be a very large number of equilibria.
where
And yet, underlying this complexity seems to be some order.
r·Z When one starts from a random distribution of manufacturing on a
(A.14) linear landscape, for example, one typically finds that a roughly reg-
Hi = [1- (-lYPe-Tin] riZ:¢Jz
ular spacing of manufacturing concentrations emerges.
Furthermore, the distance between these concentrations is relative-
(You can see from this where some of the odd sign reversals in ly insensitive to the starting position and appears to depend in a sen-
Table 8.1 came from.) sible way on the model's parameters.
The simulations for Figures 1.5 and 1.6 were both done with A = Given the results of the edge city model, these regularities are no
0.2, B = 1.0. For Figure 1.5 and the first column of Table 8.1, the longer surprising. One can immediately guess that they emerge
rates of decline were set at '1 = 1.4, r2 = 0.2; for Figure 1.6 and the from the process of divergence away from a flat equilibrium, and
second column of Table 8.1 they were set at '1 = 2.8, '2 = 0.4. that in a linearized approximation to the model in the vicinity of a
flat distribution, there is a particular frequency of fluctuation that
A CENTRAL PLACE MODEL tends to grow fastest. This turns out to be correct, but the story is a
bit more complicated, because the centripetal and centrifugal forces
In several recent papers (Krugman 1991, 1993a, 1993b) I have are now derived rather than assumed.
explored one particular approach to spatial modeling that, although I begin with a review of the general approach to spatial dynamics
admittedly capturing only some of the reasons why spatial structure used here, then turn to a specific model of a linear economy and
emerges in real economies, has the virtue of being particularly easy show how the evolution of this model near an even distribution of
to work with. In this approach, an economy with two or more loca- manufacturing can be viewed in terms of the growth rates of fluc-
108 SELF-ORGANIZING ECONOMY Appendix: The Evolution of Central Places 109

tuations of different frequencies. Finally, I show why the model transportation cost per unit distance.
economy has a preferred wavelength, and how this wavelength It is a familiar proposition that, if we take the spatial distribution
depends on the parameters. of workers as given, a model of the form just described yields a
monopolistically competitive equilibrium in which all profits are
A Basic Spatial Model competed away. This equilibrium includes an equilibrium level of
Consider an economy in which there are a number of locations, the real wage at each location; differences in these real wage rates are
indexed by j = 1, ... ,j. Let Djk be the distance between any pair of what drive the economy's dynamics.
locations j and k. Workers are assumed to move gradually toward locations that
In this economy there are two factors of production: immobile offer them above average real wages. Let A.i be the fraction of work-
"farmers" and mobile "workers." It will be convenient to choose ers currently in location j. Then the average real wage rate can be
units so that there are a total of 1 - m farmers and m workers. Also, defined as a weighted average of real wage rates at each location,
in this appendix I will restrict attention to economies in which spa-
tial structure is completely endogenous, so the farmers will be (J) = L .A.(J).
'J ') ')
(A.18)
assumed to be equally divided among the locations.
Everyone in this economy shares the same tastes, which may be and the assumed dynamics take the form 1
represented by a two-level structure. At the upper level, there are dA'
Cobb-Douglas preferences between agricultural goods and a man- ')

ufacturing aggregate: at = Y( (J)i - ro)A:;

(A.19)

The dynamic behavior of this model can be thought of as a

u= ctc~-.u (A.15)
sequence of general equilibrium problems. For any given distribu-
tion of manufacturing across locations, the economy reaches an
At the lower level, manufacturing is a CES composite of a large equilibrium that determines the real wage at each location. This
number of symmetric differentiated products: vector of real wages then determines, via (A.18) and (A.19), the dis-
tribution of workers a short time later, and the calculation can be
CM= [LiCfJl/p (A.16) repeated until the model economy converges on some long-run
equilibrium geographical pattern.
where (5= 1/(1 - p) is the elasticity of substitution. In Krugman (1992) I show that the equilibrium of this model at
Each factor is specific to the production of one sector. Farmers any point in time can usefully be described as the simultaneous
produce agricultural output with constant returns to scale. Workers solution of four sets of equations. First, the income of any location
produce manufactured goods. There are economies of scale in this is the sum of the earnings of its immobile farmers and the workers
production, specific to both the firm and the particular variety pro- who are currently located there:
duced; these are represented as a linear cost function,
1-).l
(A.20)
LMi = a + /3QMi (A.17) 1j = ] + ).lAjWj

We also introduce transport costs. For the sake of tractability,

there are assumed to be zero transport costs for agricultural goods. 1. In all of my models to date, I have ignored two important aspects of real-
world spatial economics - forward-looking behavior by agents who try to antici-
Transport costs on manufactured goods are of Samuelson's "ice- pate future spatial patterns and large agents, such as shopping mall developers, who
berg" form. If one unit of a manufactured good is shipped from try to influence these patterns. The excuse for these omissions is, of course,
location j to location k, only exp(-rDjk) units arrive, with A the tractability.
110 SELF-ORGANIZING ECONOMY Appendix: The Evolution of Central Places 111

where Wj is the wage rate measured in terms of the agricultural

good. T(x) = 2
['l'(a-l) l:l(z)w(z) I-a er(l-a)lx-zl dz ]1/(1-0) (A.25)
Second, the true price index of manufactures at any given loca-
tion depends on the distribution of manufacturing, transportation
costs, and wage rates: (A.26)
w(x) = ['l'( a-
2 1) looY(z)T(z)a-le-r(a-l)lx-zldz
00 ] 1/ a

~ = [LklkUi,l-O"e-r(a-l)Djkp/(1-o) (A.21)

co(x) = w(x)T(xtJi (A.27)

Third, the equilibrium wage rate at any location depends on
incomes, true price indices, and transportation costs to all other
locations: This set of nonlinear equations is fairly nasty looking. Suppose,
however, we restrict our attention to situations in which l(x) is
Wj = [Lk Y k T~-le-r(a-l)Djk]l/a (A.22) close to 1; that is, where the distribution of manufacturing is fairly
flat. Then we can take linear approximations to the equations. Let a
j
Finally, the real wage rate at location depends on the nominal prime on a variable represent deviation from 1; then the approxi-
wage rate in terms of agricultural goods and the local true price mate linearized model takes the form
index of manufactured goods:
Y'(x) = J1l'(x) + J1w'(x) (A.28)
~ = Wj~-Ji (A.23)

These equations are fairly simple and very easy to solve numeri- , _l_'l'(a-l) U:l'(z)e-1(a-l)lx-zldz +
cally - one simply starts with guesses at the wage and true price vec-
tors and iterates until convergence. Analytical results, however, in T(x) - i-a 2 - l:w'(z)e-1(a-l)lx-zldz] (A.29)
anything larger than a two-region model are another matter. Hence,
explorations of multilocation settings so far have relied on numeri-
cal examples. Using the Fourier series approach, however, we can - ~ 'l'(a-l) [t Y'(z)e-1(a-l)lx-zldz +
show the underlying logic of self-organization. w'(x) - a 2 -00 f:T'(z)e-1(a-l)lx-zldz] (A.30)

Dynamics Near a Flat Spatial Structure

For the formal analysis, we will consider a version of the basic co'(x) = w'(x) - J1T'(x) (A.31)
model in which farmers are distributed evenly along a line of infi-
nite extent. Workers will also, at any point in time, be distributed These equations do not, at first sight, appear any more tractable
along that line; we let l(x) be the density of workers at position x, than the nonlinear version. But let us assume for a moment that the
normalized so that with a flat distribution l
= 1 everywhere. distribution of manufacturing follows a simple periodic distribu-
For this economy, equations (A.6)-(A. 9) may be rewritten in the tion, say,
following form (the constant terms are added so that when the dis-
tribution is flat, Y(x) = w(x) = T(x) = 1 for all x is a solution): l'(x) = 8cos(cpx) (A.32)

Y(x) = 1- J1 + J1l(x)w(x) (A.24) Now let us simply guess that, if the divergence of l(x) from 1 fol-
lows this simple periodic form, the divergences of all of the other
 112 SELF-ORGANIZING ECONOMY Appendix: The Evolution of Central Places 113

variables from 1 will be constant multiples of A(X). (This conclusion H depending on the frequency of the fluctuation. It is immediately
is actually obvious from the spatial symmetry and the linearity.) obvious that for very high frequencies, H approaches 0, whereas for
That is, we guess that there is a solution of the form low frequencies, it approaches 1.
We can now write our equations as
Y'(X) = ayA'(x) (A.33)
ay = /law + /l (A. 40)

T'(X) = aTA'(x) (A.34) -1

aT = a-1 H + Haw (A.41)

w'(X) = awA'(x) (A.35)

1 a-1
am = a Hay + ----a-HaT (A.42)
m'(x) = amA'(x) (A. 36)
am = aw-/laT (A.43)
If this is a valid solution, then we have managed to reduce a gen-
eral equilibrium problem that is, strictly speaking, the solution of an
infinite number of nonlinear equations to the solution of four lin- These equations can be solved to yield the crucial result that
ear equations.
Let us, then, substitute (A.32) into (A.33)-(A.36). When we do a
w
= /l
a-1
H + (1-/lH) /l
H-HZ (A.44)
so, we will see repeatedly a term of the form

Why is this the crucial result? Because the linearized version of

K(z) = 'l"(~-1) l:cos(¢z)e-T(a-l)lx-zldz (A.37)
the dynamic equation (A.19) is

With a little grinding, it is possible to show that (A.45)

dA'(X) = m'(x)r = awrA'(x) = glpA'(X)
dt
K(z) = H(¢,'l",a)cos(¢z) (A.38)
where g is the rate of growth of a fluctuation at that frequency.
where Now we note that a perturbation of the spatial distribution of
manufacturing around A = 1 can be represented as the sum of a
H (¢, a) =. (a-1)Z
number of sine waves of different wavelengths:
'l", (A.39)

A'(X) = A'l (x) + A'z(X) + ... (A.46)

H represents a sort of discount factor: the ratio of the impact of a
fluctuation to what would happen if there were a uniform increase And the growth of the perturbation may be written
in the same variable that raised the level at x by the same amount.
Therefore in the equation for the true price index we know that an dA'(X) = glA'l(X) + gzA'z(X) + ... (A.47)
dt
equal increase in all wage rates would raise the price index at x by
an amount equal to the increase in the wage rate at x; a fluctuation
will raise the price index by H times the increase at x, with the ratio so that we can think of each periodic fluctuation as growing at its

IIHi
114 SELF-ORGANIZING ECONOMY Appendix: The Evolution of Central Places 115

own characteristic rate. The fluctuation that will grow fastest is the It is more painful to derive the impact of changes in the elasticity
one with the largest (positive) response of the real wage rate to man- of substitution and the share of manufacturing. It is, however,
ufacturing concentration and, given sufficient time, that fluctua- straightforward to calculate the preferred frequency numerically for
tion will dominate the spatial pattern. given f-t and o. This is shown in Table A.1; we see that higher elas-
So all we have to do to determine the preferred wavelength is find ticities of substitution, which imply lower equilibrium economies
the maximum of (A.44). It is straightforward to determine three of scale, tend to reduce the preferred wavelength, whereas a higher
results. First, manufacturing share tends to increase the preferred wavelength.

a(j)= 0 whenH= 0; that is, when <P~oo (A.48) Table A.l

That is, fluctuations at very high frequencies - very short wave- a 1.t
.2 .3 .4
lengths will not tend to grow. Second, 4 5.82 4.48 3.60
5 7.76 6.11 5.00
(A.49)
aOJ < 0 when H = 1, provided that f.l < O'~ 1 6 10.00 7.64 6.25

The condition here is a familiar one, appearing also in Krugman Preferredvaluesof l/>/r in centralplacemodel
(1991). It says, in effect, that economies of scale are not so large that
all workers would prefer to be concentrated in the same place no We see, then, that a model in which there are no assumed
matter how high transportation costs are. Given this condition, we agglomeration economies or diseconomies, in which centripetal
find that very low frequency fluctuations, those with very long and centrifugal forces are entirely emergent properties, exhibits
wavelengths, tend to die out. behavior that is very similar to that in a model in which these forces
Finally, at H = 0 we find are simply assumed in a reduced form. And in both models, we find
that the economy undergoes a systematic process of self-organiza-
daw = ~+~ > 0 tion that can produce highly regular structures.
dH 0"-1 0" (A. 50)

Taken together, these observations imply that the relationship

between the growth rate of a fluctuation and H is like an inverted J.
Growth is slow at very short wavelengths, negative at high wave-
lengths, and most rapid at some intermediate wavelength.
The preferred wavelength, the wavelength of most rapid diver-
gence, is a function of the three parameters 1', 0', and f.l. The trans-
port cost l' enters the solution in only one place, in the definition of
H in (A.39). It is therefore obvious that the preferred frequency is
strictly proportional to l' and hence that the preferred wavelength is
inversely proportional. This is obvious with hindsight, because the
wavelength and the transportation cost can both be changed in the
same proportion by redefining the unit of distance, with no real
change in the model.
 References

Arthur, W. B. (1994). Increasing Returns and Path Dependence in the Economy.

Ann Arbor: University of Michigan Press.
Bak, P. (1991). "Self-Organizing Criticality." Scientific American Oanuary).
Christaller, W. (1933). Central Places in Southern Germany. Jena: Fischer.
English translation by C. W. Baskin. London: Prentice-Hall, 1966.
Cronon, W. (1991). Nature's Metropolis. New York: W. W. Norton.
Dicken, P., and Lloyd, P. (1990). Location in Space. New York: Harper and
Row.
Dixit, A., and Stiglitz, J. (1977). "Monopolistic Competition and
Optimum Product Diversity." American Economic Review Oune).
Fujita, M. (1989). Urban Economic Theory. Cambridge: Cambridge
University Press.
Fujita, M., and Ogawa, H. (1982). "Multiple Equilibria and Structural
Transition on Non-Monocentric Urban Configurations." Regional
Science and Urban Economics 12 (May): 161-196.
Garreau,J. (1992). Edge City. New York: Anchor Books.
Goodwin, R. (1951). "The Nonlinear Accelerator and the Persistence of
Business Cycles." Econometrica 19: 1-17.
Harris, C. (1954). "The Market as a Factor in the Localization ofIndustry
in the United States." Annals if the Association if American Geographers 64:
315-348.
Hicks,J. R. (1950). A Contribution to the Theory if the Trade Cycle. Oxford:
Oxford University Press.
Hopfield, J. J. (1982). "Neural Networks and Physical Systems with
Emergent Collective Computational Abilities." Proceedings if the
National Academy if Sciences 79: 2554-2558.
Ijiri, Y., and Simon, H. (1977). Skew Distributions and the Sizes if Business
Firms. Amsterdam: North-Holland.

117
 118 SELF-ORGANIZING ECONOMY

Kauffman, S. (1993). The Origins if Order. New York: Oxford University

Press.
Krugman, P. (1991). "Increasing Returns and Economic Geography."
Journal if Political Economy (June).
Krugman, P. (1993a). "On the Number and Location of Cities." European
Economic Review (May).
Krugman, P. (1993b). "First Nature, Second Nature, and Metropolitan Index
Location." Journal if Regional Science.
Lewin, R. (1992). Complexity: Life at the Edge if Chaos. New York:
Macmillan.
Losch, A. (1940). The Economics if Location. Jena: Fischer. English transla-
tion, New Haven, Conn.: Yale University Press, 1954.
agglomerations, 33, 102, 107, 115; change, process of: in economy,
Lowry, I. S. (1966). Migration and Metropolitan Growth: Two Analytical 56, 59; technology choice,
Models. San Francisco: Chandler. generated by central place
model, 88-9; multiple, 75nl, models of, 57-9. see also
Mankiw, N. G. (1994). Macroeconomics, 2nd ed. New York: Worth. 77,88,90-1; shadow, 87; dynamic systems
Murphy, R., Shleifer, A., and Vishny, R. (1989). "Industrialization and temporal equivalent of, 68 chaos, 48
the Big Push." Journal if Political Economy. Anderson, Philip, 3, 11,91 Chicago, 13, 59
Nicolis, G., and Prigogine, I. (1989). Exploring Complexity. New York: W. Arthur, Brian, 87 Christaller, W., 14-15
H. Freeman. Citibank,2
Prigogine, I., and Stengers, I. (1984). Order out if Chaos. New York: cities. see also edge cities;
Bantam Books. Bak,Per, 44, 49, 69, 70,99
metropolitan areas; economics
Baumol, William, 9
Scheinkman,J. A., and Woodford, M. (1994). "Self-Organized Criticality and, 9; explosive growth of, 59;
Bjerknes, Vilhelm, 48 Mills monocentric model of,
and Economic Fluctations." American Economic Review 84 (May): Blinder, Alan, 9
417-421. 12-13; one-dimensional models
business cycle,S, 6, 53, 61-3; of, 22, 77, 101; rank-
Schelling, T. (1978). Micromotives and Macrobehavior. New York: W. W.
Norton. global, 71-3; nonlinear theory size distributions. See rank-
of, 6, 7, 63-8, 71, 72-3, 100;
Simon, H. (1955). "On a Class of Skew Distribution Functions." size rule, city
order from random growth and,
Biometrika. Cobb-Douglas preferences, 108
69-71; power law for size collective behavior, 3
Tobin, J. (1955). "A Dynamic Aggregative Model." Journal if Political distribution of, 70, 71
Economy 63: 103-115. competition for land, 10, 11-12, 13
complex landscapes, 31-8; of
Turing, A. (1952). "The Chemical Basis of Morphogenesis." Philosophical
central place model, 76, 88-92, economy, 55-6; punctuated
Transactions if the Royal Society if London 237: 37. 106-15 equilibrium and, 59; of
Waldrop, W. M. (1993). Complexity. New York: Basic Books.
central place theory, 13-15,42, 76 technological change, 57
central places: evolution of, lOl- Complexity: Life at the Edge of
lS; hierarchy of, 41-3 Chaos, 6
centripetal and centrifugal forces: Complexity: The New Science
business cycle, 68; central place at the Edge of Order and
model, 33, 88, 89, 90, 91, 107, Chaos, 66
115; edge city model, 77, 81, complexity, theory of, 1-7. see also
83, 101 complex landscapes; economics,

119
120 SELF-ORGANIZING ECONOMY Index 121

application to, 1-7; emergence rate of, 80-3; large city case, land rents, 10, 12, 13, 23 monopolistically competitive
and, 3, 11-12,89,91; feedback 104-6; nonlinear, 103; small city land use patterns: self- equilibrium, 109
systems and, 2-3; Schelling's case, 106 organization model of, 22-9, 25; morphogenesis, 48, 49, 101-6
model of segregation and, 17; emergence, 3, 11-12, 89, 91 von Thiinen-Mills model, 9-12,
self-organization and, 3-7, 36-7; 23. see also central place theory;
Nature's Metropolis: Chicago and
von Thiinen model and, 10-12 edge cities
feedback, 2-3; international the Great West, 31
concentric ring land use pattern, large cities, 93-4 New York, 13, 95
trade, 55
10, 11, 12 Lewin, R., 6
Fourier series, 28, 80-1, 90, nonlinear business cycle theory,
convection, 47-8, 99-100 linear dynamics, 104-6
105-6,110 63-8, 71, 72-3, 100
created landscape, 31 location, market potential
Fujita, M., 6 of, 77, 102
criticality: edge of, 73; self-
organized, 44, 69, 70 location decisions, Ogawa, H., 6
Cronon, William, 31 Garreau,Joel, 4, 21, 24n4, 59, business:independent, 101; order from instability. See
88,96 interdependent, 24, 76-8 instability, order from
General Theory, 61 location theory, 13-14. see also Order out of Chaos, 6
dynamic systems, 55-9, 102-6;
Geography and Trade, 88 central place theory; edge Origins of Order, The, 6
linear dynamics, 104-6; phase
global interdependence, 71-3 city model
space representations of, 32, 34, Lorenz, Edward, 48
Goodwin, Richard, 6, 63, 68, Palo Alto, 13
36; rugged (complex)landscapes,
71, 72-3 Los Angeles: explosive growth of,
32-6; visualization of, 31-8. See percolation economics, 1,69-71
growth and trade, 55, 59 59; multipolarity of, 21; self-
also change, process of phase locking and global business
growth rates, 43, 114. see also rank- organization of, 4, 13
cycle, 72-3
size rule, city L'sch, A., 14-15, 76, 91, 92
phase space, 32, 34, 58nl, 59
earthquakes, 43, 44, 49-50 Gutenberg-Richter power
economic fluctuations. See power laws, 37; business
law, 43, 70 fluctuations size distribution,
Mankiw, Greg, 62
business cycle
market potential oflocation, 70, 71; city rank-size
economic geography, models of, 77, 102 distribution, 42, 43, 46, 71, 76,
33-6,37 Hicks, John, 6, 63, 71, 72-3
historical contingency, 36 market-size effects, 91 100; of fragmentation, 49;
economic slumps, 5, 6, 65
Hopfield, John, 99 maximum instability, principle regularity of, 43-6; Schelling's
economic theory: central place of, 28-9 segregation model, 21; Simon's
theory and, 14-15; complexity meteorites, 43-4 model of city size, 92-7. see also
applied to, 1-7; emergence and, instability, order from, 37, 47-9, instability, order from
meteorology, 48, 50
3; feedback systems and, 2-3; 99-100; central place model, metropolitan areas. see also edge Prigogine, Ilya, 6
self-organizing systems and, 3- 107; edge city model, 28-9, 79- cities; Mills monocentricity primate cities, 41
7. see also business cycle; central 80, 84, 87; nonlinear business
model of, 12-13; polycentric, punctuated equilibrium, 56, 58, 59
place theory; edge city model; cycle theory, 68; punctuated 13,21-2,24-9,68, 77, 101;
spatial economics equilibrium and, 59; Schelling's power law for size distribution
economies of scale, 94-5, 114 segregation model, 17, 18-20 random growth, order from, 49-
of, 42, 43, 46, 71, 76, 100;
edge cities, 6, 96; explosive growth integration. See segregation rank-size rule. See rank-size 50,99, 100; applied to
of, 59; Los Angeles, 4, 13, 21 internal city structure, model economic fluctuations, 69-71;
rule, city; self-organization of,
Edge City, 4, 88 of, 91-2 power law of, 92-7
4,6,13,21-2,100
edge city model, 22-9, 37, 48-9, international trade, 55 rank-size rule, city, 39-46, 71, 76,
Micromotives and Macrobehavior,
76-88, 101-6, 107; adding-up 92-7,99,100
7, 15, 16 reaction-diffusion model, 48-9
constraint, 85-6, 87; compact Mills, Edwin, 12-13
Kauffinan, Stuart, 6, 32 Reed, John, 2
subcenters in, 86-7; fluctuations,
Keynesian theory, 61-3
122 SELF-ORGANIZING ECONOMY

regularity of self-organization, 37- 92-7; VonThiinen-Mills model,

8,43-6,79-80,84,103,115 9-12. See also central place
rugged landscapes,32-6 model; edge city model
spin-glasses,21, 33, 99
Stenger, I., 6
San Francisco, 13
Stiglitz,Joseph, 9
Santa Fe Institute, 2
Scheinkman,Jose,69-71
Scheinkman-Woodford model, technology choice, models of,
69-71, 72-3, 100 56-9
Schelling, Thomas, 7, 15-22, temporal self-organization, 4-5, 6,
47,99 53,61-73; along time
second nature landscape, 31 dimension itself, 55; edge city
segregation, of cities, 6, 7, 15-22, model, 67-8; nonlinear business
47,99 cycle theory, 63-8; of
self-organization, 3-7, 99-100; organization self-
change in, 55-6, 59; of complex organizing in other dimension,
systems,36-7; dynamics in, 55- 55-9
9, 102-6; Fourier se'riesand, 28, '. Tobin, James, 6,63
80-1,90,105-6,110; of Toffler, Alvin,.6
metropolitan areas, 4, 6, 13,21- trade md growth, 55
2, 100; oyer tiWe, 63-73; transmission mechanism, 62
principles of, 47-50; .punctuateotransportation costs, 88, 90, 107,
equilibrium and, 56, 58, 59; 108-9, 110, 114
regularity of, 37-8, 43-6, 79-80, Turing, Alan, 48
84, 103, 115;value of, 5-6. see
also instability, order from;
urban areas. See metropolitan areas
spatialself-organization; urban morphogenesis, 49, 101-6
temporal self-organization urban subcenters. See edge cities
self-organized criticality, 44, 69,
70
Silicon Valley,37 Von Th nen-Mills model,
Simon, Herbert, 44-6, 76, 92-7 9-12,23
spatialeconomics, 9; complex
landscapesand, 33-6; regularity Waldrop, Mitchell, 66
of, 37-8. see also spatialself- Weber, Alfred, 14
organization Woodford, Michael, 69-71
spatialself-organization, 4, 6, 9-
29,53; Schelling's model of .
segregation and, 16, 18, 20,21- Zipf's law, 39-46
2; Simon's urban growth model,
In the last few years the concept of self-organizing systems - of com-
plex systems in which randomness and chaos seem spontaneously to
evolve into unexpected order - has linked together researchers in
many fields, from artificial intelligence to chemistry, from evolution
to geology. Now leading economist Paul Krugman shows how prin-
ciples that explain the growth of hurricanes and embryos can also ex-
plain the formation of cities and business cycles; how the same prin-
ciples of "order from random growth" can explain the strangely
simple rules that describe the sizes of earthquakes, meteorites, and
metropolitan areas. Weaving together strands from many disciplines,
from location theory to biology, The Self- Organizing Economy offers a
surprising new view of how the economy structures itself in space
and time.

faul Krugman is one of the world's leading economists. He has

served on the President's Council of Economic Advisors and is
currendy a professor of economics at Stanford University. In recog-
nition of his research on international trade and finance, in 1991
the American Economic Association gave him its John Bates Clark
medal, an award given every two years to the best American econo-
mist under 40. He has written numerous books and delivers lectures
around the world.

A~~lltrim Krn[lllmI

" ... the most creative economist of his generation."

- The Economist

" ... writes more fluently for laymen than anyone else
.. economICS."
In
- The Boston Globe

ISBN 1-55786-699-6

Cover deisp by Diane Levy

.1.1111.1111.1 111111