Agent-Based Economics and Econophysics

/HFWXUH1RWHVLQ(FRQRPLFV
DQG0DWKHPDWLFDO6\VWHPV
)RXQGLQJ(GLWRUV
0%HFNPDQQ
+3.Q]L
0DQDJLQJ(GLWRUV
3URI'U*)DQGHO
)DFKEHUHLFK:LUWVFKDIWVZLVVHQVFKDIWHQ
)HUQXQLYHUVLWlW+DJHQ
)HLWKVWU$9=,,+DJHQ*HUPDQ\
3URI'U:7URFNHO
,QVWLWXWIU0DWKHPDWLVFKH:LUWVFKDIWVIRUVFKXQJ,0:
8QLYHUVLWlW%LHOHIHOG
8QLYHUVLWlWVVWU%LHOHIHOG*HUPDQ\
(GLWRULDO%RDUG
$%DVLOH$'UH[O+'DZLG.,QGHUIXUWK:.UVWHQ86FKLWWNR
$NLUD1DPDWDPH
7DLVHL.DL]RXML
<XXML$UXND(GV
7
KH&RPSOH[1HWZRUNV
RI(FRQRPLF,QWHUDFWLRQV
(VVD\VLQ$JHQW%DVHG(FRQRPLFV
DQG(FRQRSK\VLFV

(GLWRUV
3URIHVVRU$NLUD1DPDWDPH
'HSDUWPHQWRI&RPSXWHU6FLHQFH
1DWLRQDO'HIHQVH$FDGHP\
+DVKLULPL]X<RNRVXND
-DSDQ
HPDLOQDPD#QGDDFMS
7DLVHL.DL]RXML
'LYLVLRQRI6RFLDO6FLHQFHV
,QWHUQDWLRQDO&KULVWLDQ8QLYHUVLW\,&8
2VDZD0LWDND7RN\R
-DSDQ
HPDLONDL]RML#LFXDFMS
3URIHVVRU<XXML$UXND
'HSDUWPHQWRI&RPPHUFH
&KXR8QLYHUVLW\
7DPD7RN\R
-DSDQ
HPDLODUXND#WDPDFFFKXRXDFMS
,661
,6%16SULQJHU%HUOLQ+HLGHOEHUJ1HZ<RUN
,6%16SULQJHU%HUOLQ+HLGHOEHUJ1HZ<RUN
7KLVZRUNLVVXEMHFWWRFRS\ULJKW$OOULJKWVDUHUHVHUYHGZKHWKHUWKHZKROHRUSDUW
RIWKHPDWHULDOLVFRQFHUQHGVSHFLÀFDOO\WKHULJKWVRIWUDQVODWLRQUHSULQWLQJUHXVHRI
LOOXVWUDWLRQVUHFLWDWLRQEURDGFDVWLQJUHSURGXFWLRQRQPLFURÀOPVRULQDQ\RWKHUZD\
DQGVWRUDJHLQGDWDEDQNV'XSOLFDWLRQRIWKLVSXEOLFDWLRQRUSDUWVWKHUHRILVSHUPLWWHG
RQO\XQGHUWKHSURYLVLRQVRIWKH*HUPDQ&RS\ULJKW/DZRI6HSWHPEHULQLWV
FXUUHQWYHUVLRQDQGSHUPLVVLRQIRUXVHPXVWDOZD\VEHREWDLQHGIURP6SULQJHU9HUODJ
9LRODWLRQVDUHOLDEOHIRUSURVHFXWLRQXQGHUWKH*HUPDQ&RS\ULJKW/DZ
6SULQJHULVDSDUWRI6SULQJHU6FLHQFH%XVLQHVV0HGLD
VSULQJHUFRP
6SULQJHU9HUODJ%HUOLQ+HLGHOEHUJ
3ULQWHGLQ*HUPDQ\
7KH XVH RI JHQHUDO GHVFULSWLYH QDPHV UHJLVWHUHG QDPHV WUDGHPDUNV HWF LQ WKLV
SXEOLFDWLRQGRHVQRWLPSO\HYHQLQWKHDEVHQFHRIDVSHFLÀFVWDWHPHQWWKDWVXFK
QDPHVDUHH[HPSWIURPWKHUHOHYDQWSURWHFWLYHODZVDQGUHJXODWLRQVDQGWKHUHIRUH
IUHHIRUJHQHUDOXVH
7\SHVHWWLQJ&DPHUDUHDG\E\DXWKRU
&RYHUGHVLJQ(ULFK.LUFKQHU+HLGHOEHUJ
3ULQWHGRQDFLGIUHHSDSHU'.
Preface
Understanding the mechanism of a socio-economic system requires more than

an understanding of the individuals that comprise the system. It also requires
understanding how individuals interact with each other, and how the aggre-
gated outcome can be more than the sum of individual behaviors. This book
contains the papers fostering the formation of an active multi-disciplinary
community on socio-economic systems with the exciting new fields of agent-
based modeling and econophysics.
We especially intend to increase the awareness of researchers in many
fields with sharing the common view many economic and social activities as
collectives of a large-scale heterogeneous and interacting agents.
Economists seek to understand not only how individuals behave but also
how the interaction of many individuals leads to complex outcomes. Agent-
based modeling is a method for studying socio-economic systems exhibiting
the following two properties: (1) the system is composed of interacting agents,
and (2) the system exhibits emergent properties, that is, properties arising
from the interactions of the agents that cannot be deduced simply by aggre-
gating the properties of the system’s components. When the interaction of
the agents is contingent on past experience, and especially when the agents
continually adapt to that experience, mathematical analysis is typically very
limited in its ability to derive the outcome.
Many physicists have contributed to a better understanding of large-scale
properties of socio-economic systems, and they open the new research field,
“econophysics”. An international scientific development has started to gain
new insight into the dynamics of socio-economic systems by using methods
originally developed in statistical physics and complex theory. This book also
covers the current achievements in this rapidly changing field.
This book contains selected papers presented at the 9-th International
Workshop on Heterogeneous Interacting Agents (WEHIA), which was held at
Kyoto University, Japan, from May 27 to 29, 2004. From the broad spectrum of
activities, leading experts presented important papers and numerous practical
VI Preface
problems appear throughout this book. We also encouraged papers dealing

with applications of agent-based modeling.
WEHIA was initiated as a result of the growing recognition of the im-
portance of agent-based modeling to study large-scale socio-economic sys-
tems at University of Ancona, Italy in 1996. The annual series of WEHIA
serve for sharing the most recent theoretical applications and methodolog-
ical advances on agent-based approaches throughout economists, physicists,
computer scientists, and other scientists in professionals. The main goals of
WEHIA have been to promote interactions and cross-fertilization among dif-
ferent approaches to understanding complex and emergent behaviors and to
mange large-scale socio-economic systems.
WEHIA confers especially to encourage papers at the cutting-edge of other
approaches that are relevant socio-economic systems. By bringing together
three different emerging fields, economics, echonophysics and computer sci-
ence under the same umbrella, WEHIA stresses the expanding importance
of importance close communication and cooperation of the three areas for
the future scientific and technological development. The genuinely interdisci-
plinary approach will enable researchers and students to expand their socio-
economic knowledge and to draw up concepts for future interdisciplinary aca-
demic achievement.
Based on the success of WEHIA for many years, the new association, “The
society for Economic Science with Heterogeneous Interacting Agents” (ES-
HIA) (www.es-hia.org) will be launched in 2006. The official society journal,
“Journal of Economic Interaction and Coordination” (JEIC) will be published
from Springer in 2006. The new society, ESHIA especially features in-depth
coverage of important areas and aims to contribute scientific ally in three di-
rections: (1) To examine theoretical and methodological issues of agent-based
modeling. (2) To discuss multi-agents based simulations and demonstrate ap-
plicability in order to study complex economic behaviors. (3) To contribute
to develop methodological tools of agent-based modeling and apply them to
complex economic and social problems.
We could solicit many high quality papers that reflect the result of the
growing recognition of the importance of the areas. All papers have received a
careful and supportive review, and we selected 22 papers out of 94. The con-
tributions were submitted as a full paper and reviewed by senior researchers
from the program committee. All authors revised their earlier versions pre-
sented at the workshop with reflecting criticisms and comments received at
the workshop. The editors would like to thank the program committee for the
careful review of the papers and the sponsors and volunteers for their valuable
contribution. We hope that as a result of reading the book you will share with
us the intellectual excitement and interest in this emerging discipline.
We are grateful to the many people who have made this symposium pos-
sible. First and foremost, we thank the authors for providing manuscripts on
time and in a standard format. We also thank the many referees who gen-
Preface VII
erously contributed time and Dr. Hiroshi Sato to ensure the quality of the
finished product.
Finally, we would like to acknowledge the support and encouragement of
many peoples in helping us getting this book to be published. Especially the
publication of this book and the 9th WEHIA are financially supported by the
grant from the Commerative Organization for the Japan World Exposition
(’70), Hayasibara Foundation, Kozo Keikaku Engineering Inc. We would like
also thank for the grant-in-aid for Scientific Research (C) No.15201038, Japan
Society for the promotion of Science (JSPS).
October 2005 Akira Namatame

Taisei Kaizoji
Yuji Aruka
Contents
Part I Econophysics
Five Years of Continuous-time Random Walks in Econophysics

Enrico Scalas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Why Macroeconomic Price Indices are Sluggish in Large
Economies ?
Masanao Aoki, Hiroshi Yoshikawa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Growth Volatility Indices

Davide Fiaschi, Andrea Mario Lavezzi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Financial Fragility and Scaling Distributions in the Laboratory

Giovanna Devetag, Edoardo Gaffeo, Mauro Gallegati, Gianfranco
Giulioni . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Part II Complex Economic Network
Heterogeneous Economic Networks

Wataru Souma, Yoshi Fujiwara, Hideaki Aoyama . . . . . . . . . . . . . . . . . . . . 79
The Emergence of Paradigm Setters Through Firms’

Interaction and Network Formation
Rainer Andergassen, Franco Nardini, Massimo Ricottilli . . . . . . . . . . . . . . 93
Part III Economic Dynamics
Statistical Properties of a Heterogeneous Asset Pricing Model

with Time-varying Second Moment
Carl Chiarella, Xue-Zhong He, Duo Wang . . . . . . . . . . . . . . . . . . . . . . . . . . 109
X Contents
Deflationary Recessions in a General Equilibrium Framework

Luca Colombo, Gerd Weinrich . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Concepts of Thermodynamics in Economic Growth
Jürgen Mimkes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Firm Dynamics Simulation Using Game-theoretic Stochastic
Agents
Yuichi Ikeda, Osamu Kubo, Yasuhiro Kobayashi . . . . . . . . . . . . . . . . . . . . . . 153
Part IV Agent-based Modeling
A Functional Modularity Approach to Agent-based Modeling

of the Evolution of Technology
Shu-Heng Chen, Bin-Tzong Chie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Herding Without Following the Herd: The Dynamics of
Case-Based Decisions with Local Interactions
Andreas Krause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Cultural Evolution in a Population of Heterogeneous Agents

Gábor Fáth, Miklos Sarvary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Part V Auction and Two-sided Matching
Simulating Auctions
Konrad Richter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Counterclockwise Behavior Around the Beveridge Curve

Koji Yokota . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
The Waiting-Time Distribution of Trading Activity in a

Double Auction Artificial Financial Market
Silvano Cincotti, Sergio M. Focardi, Linda Ponta, Marco Raberto,
Enrico Scalas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Part VI Minority Games and Collectie Intelligence
Theoretical Analysis of Local Information Transmission in

Competitive Populations
Sehyo Charley Choe, Sean Gourley, Neil F. Johnson, Pak Ming Hui . . . . 251
Analysis of Complexity and Time Restriction in Resources
Allocation Problems
Kiyoshi Izumi, Tomohisa Yamashita, Koichi Kurumatani . . . . . . . . . . . . . 265
Contents XI
How Does Collective Intelligence Emerge in the Standard

Minority Game?
Satoshi Kurihara, Kensuke Fukuda, Toshio Hirotsu, Osamu Akashi,
Shinya Sato, Toshiharu Sugawara . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Part VII Game-theoretic Approach
What Information Theory Says About Bounded Rational

Best Response
David H. Wolpert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Evolution of Reciprocal Cooperation in the Avatamsaka

Game
Eizo Akiyama, Yuji Aruka . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Game Representation - Code Form
Maria Cristina Peixoto Matos, Manuel Alberto M. Ferreira . . . . . . . . . . . . 321
Effect of Mutual Choice Metanorm in Group Dynamics for
Solving Social Dilemmas
Tomohisa Yamashita, Kiyoshi Izumi, Koichi Kurumatani . . . . . . . . . . . . . 335
Part I
Econophysics
Five Years of Continuous-time Random Walks
in Econophysics
Enrico Scalas
DISTA, Università del Piemonte Orientale, Via Bellini 25/G, 15100 Alessandria,
Italy and INFM Unità di Genova, Via Dodecaneso 33, 16146 Genova. Italy
scalas@unipmn.it
Summary. This paper is a short review on the application of continuos-time ran-

dom walks to Econophysics in the last five years.
1 Introduction
Recently, there has been an increasing interest on the statistical properties
of high-frequency financial data related to market microstructural properties
[1, 2, 3, 4, 5, 6]. High-frequency econometrics is now well established after
research on autoregressive conditional duration models [7, 8, 9, 10].
In high-frequency financial data not only returns but also waiting times
between consecutive trades are random variables [11]. This remark is present
in a paper by Lo and McKinlay [12], but it can be traced at least to pa-
pers on the application of compound Poisson processes [13] and subordinated
stochastic processes [14] to finance. Models of tick-by-tick financial data based
on compound Poisson processes can also be found in the following references:
[15, 16, 17].
Compound Poisson processes are an instance of continuous-time random
walks (CTRWs) [18]. The application of CTRW to economical problems dates
back, at least, to the 1980s. In 1984, Rudolf Hilfer discussed the application
of stochastic processes to operational planning, and used CTRWs as tools for
sale forecasts [19]. The revisited and augmented CTRW formalism has been
applied to high-frequency price dynamics in financial markets by our research
group since 2000, in a series of three papers [20, 21, 22]. Other scholars have
recently used this formalism [23, 24, 25]. However, already in 1903, the PhD
thesis of Filip Lundberg presented a model for ruin theory of insurance compa-
nies, which was later developed by Cramér [26, 27]. The underlying stochastic
process of the Lundberg-Cramér model is another example of compound Pois-
son process and thus also of CTRW.
Among other issues, we have studied the independence between log-returns
and waiting times for stocks traded at the New York Stock Exchange in Oc-
4 Enrico Scalas
tober 1999. For instance, according to a contingency-table analysis performed

on General Electric (GE) prices, the null hypothesis of independence can be
rejected with a significance level of 1 % [28]. We have also discussed the
anomalous non-exponential behaviour of the unconditional waiting-time dis-
tribution between tick-by-tick trades both for future markets [21] and for stock
markets [28, 29]. Different waiting-time scales have been investigated in differ-
ent markets by various authors. All these empirical analyses corroborate the
waiting-time anomalous behaviour. A study on the waiting times in a contem-
porary FOREX exchange and in the XIXth century Irish stock market was
presented by Sabatelli et al. [30]. They were able to fit the Irish data by means
of a Mittag-Leffler function as we did before in a paper on the waiting-time
marginal distribution in the German-bund future market [21]. Kyungsik Kim
and Seong-Min Yoon studied the tick dynamical behavior of the bond futures
in Korean Futures Exchange (KOFEX) market and found that the survival
probability displays a stretched-exponential form [31]. Finally, Ivanov et al.
[33] confirmed that a stretched exponential fits well the survival distribution
for NYSE stocks as we suggested in [28]. Moreover, just to stress the relevance
of non-exponential waiting times, a power-law distribution has been recently
detected by T. Kaizoji and M. Kaizoji in analyzing the calm time interval
of price changes in the Japanese market [32]. We have offered a possible ex-
planation of the anomalous waiting-time behaviour in terms of daily variable
activity [29].
The aforementioned empirical results are important as market microstruc-
tural models should be able to reproduce such a non-exponential behaviour
of waiting-time distributions in order to be realistic. However, the rest of this
paper focuses on the theory and is divided as follows: in Section 2, CTRW
theory is presented as applied to finance. Finally, in Sec. 3, a summary of
main results is presented together with a discussion on the direction of future
research.
2 Theory
Random walks have been used in finance since the seminal thesis of Bachelier
[34], a work completed at the end of the XIXth century, more than a hundred
years ago. After a rather long period in which the ideas of Bachelier were
neglected, they were further developed until recent times [35, 36].
Our approach to random walks in finance is related to that of Clark [14]
and to the introductory part of Parkinson’s paper [37]. It is a purely phe-
nomenological approach. There is no assumption on the rationality or the
behaviour of trading agents and it is not necessary to assume the validity of
the efficient market hypothesis [38, 39]. However, as briefly discussed above,
even in the absence of a microfoundation, a phenomenological model can still
be useful to corroborate or falsify the consequences of behavioural or other
Five Years of Continuous-time Random Walks in Econophysics 5
assumptions on markets. The model itself can be corroborated or falsified by

empirical data.
In order to model tick-by-tick data, we use the so-called continuous-time
random walk (CTRW), where time intervals between successive steps are ran-
dom variables, as discussed by Montroll and Weiss [18]. In physics, CTRWs
have been introduced as models of diffusion with instantaneous jumps from
one position to the next. For this reason they can be used as models of price
dynamics as well.
Let S(t) denote the price of an asset at time t. In a real market with a
double-auction mechanism, prices are fixed when buy orders are matched with
sell orders and a transaction (trade) occurs. It is more convenient to refer to
returns rather than prices. For this reason, we shall take into account the
variable x(t) = log S(t): the logarithm of the price. For a small price variation
∆S = S(ti+1 ) − S(ti ), the return r = ∆S/S(ti ) and the logarithmic return
rlog = log[S(ti+1 )/S(ti )] virtually coincide.
CTRWs are essentially point processes with reward [40]. The point process
is characterized by a sequence of independent identically distributed (i.i.d.)
positive random variables τi , which can be interpreted as waiting times be-
tween two consecutive events:

n
tn = t0 + τi ; tn − tn−1 = τn ; n = 1, 2, 3, . . . ; t0 = 0. (1)
i=1
The rewards are (i.i.d.) not necessarily positive random variables: ξi . In the
usual physical intepretation, the ξi s represent the jumps of a diffusing par-
ticle, and they can be n-dimensional vectors. Here, only the 1-dimensional
case is studied, but the extension of many results to the n-dimensional case
is straightforward. The position x of the particle at time t is given by the
following random sum (with N (t) = max{n : tn ≤ t} and x(0) = 0):

N (t)
x(t) = ξi . (2)
i=1
In the financial interpretation outlined above, the ξi ’s have the meaning of log-
returns, whereas the positions or rewards x(t) represent log-prices at time t.
Indeed, the time series {x(ti )} is characterised by ϕ(ξ, τ ), the joint probability
density of log-returns ξi = x(ti+1 ) − x(ti ) and of waiting
times τi = ti+1 − ti .
The joint density satisfies the normalization condition dξdτ ϕ(ξ, τ ) = 1. It
must be again remarked that both ξi and τi are assumed to be independent
and identically distributed (i.i.d.) random variables. This strong assumption
is useful to derive limit theorems for the stochastic processes described by
CTRWs. However, in financial time series, the presence of volatility clustering,
as well as correlations between waiting times do falsify the i.i.d hypothesis.
The reader interested in a review on correlated random variables in finance is
referred to chapter 8 in McCauley’s recent book [41].
6 Enrico Scalas
In general, log-returns and waiting times are not independent from each
other [28]. By probabilistic arguments (see [18, 21, 42]), one can derive the
following integral equation that gives the probability density, p(x, t), for the
particle of being in position x at time t, conditioned by the fact that it was
in position x = 0 at time t = 0:
t +∞
p(x, t) = δ(x) Ψ (t) + ϕ(x − x , t − t ) p(x , t ) dt dx , (3)
0 −∞
where δ(x) is Dirac’s delta function and Ψ (τ ) is the so-called survival function.
Ψ (τ ) is related to the marginal waiting-time probability density ψ(τ ). The two
marginal densities ψ(τ ) and λ(ξ) are:
+∞
ψ(τ ) = ϕ(ξ, τ ) dξ
−∞
∞
λ(ξ) = ϕ(ξ, τ ) dτ, (4)
0
and the survival function Ψ (τ ) is:

τ ∞

Ψ (τ ) = 1 − ψ(τ ) dτ = ψ(τ ) dτ . (5)
0 τ
Both the two marginal densities and the survival function can be empirically
derived from tick-by-tick financial data in a direct way.
The integral equation, eq. (3), can be solved in the Laplace-Fourier domain.
The Laplace transform, g(s) of a (generalized) function g(t) is defined as:
+∞
g(s) = dt e−st g(t) , (6)
0
whereas the Fourier transform of a (generalized) function f (x) is defined as:

+∞

f (κ) = dx eiκx f (x) . (7)
−∞
A generalized function is a distribution (like Dirac’s δ) in the sense of S. L.

Sobolev and L. Schwartz [43].
One gets:
1
p(κ, s) = Ψ(s) , (8)

1 − ϕ(κ,
s)
or, in terms of the density ψ(τ ):

1 − ψ(s) 1

p(κ, s) = , (9)
s s)
1 − ϕ(κ,
as, from eq. (5), one has:

1 − ψ(s)
Ψ (s) = . (10)
s
In order to obtain p(x, t), it is then necessary to invert its Laplace-Fourier
transform p(κ, s). As we shall see in the next subsection, for log-returns in-
dependent from waiting times, it is possible to derive a series solution to the
integral equation (3).
2.1 Limit Theorems: The Uncoupled Case
In a recent paper, Gorenflo, Mainardi and the present author have discussed
the case in which log-returns and waiting times are independent [42]. It is
the so-called uncoupled case, when it is possible to write the joint probability
density of log-returns and waiting times as the product of the two marginal
densities:
ϕ(ξ, τ ) = λ(ξ)ψ(τ ) (11)

with the normalization conditions dξλ(ξ) = 1 and dτ ψ(τ ) = 1.
In this case the integral master equation for p(x, t) becomes:
t +∞

p(x, t) = δ(x) Ψ (t) + ψ(t − t ) λ(x − x ) p(x , t ) dx dt

(12)
0 −∞
This equation has a well known general explicit solution in terms of P (n, t),
the probability of n jumps occurring up to time t, and of the n-fold convolution
of the jump density, λn (x):
+∞ +∞
λn (x) = ... dξn−1 . . . dξ1 λ(x − ξn−1 ) . . . λ(ξ1 ). (13)
−∞ −∞
Indeed, P (n, t) is given by:

t
P (n, t) = ψn (t − τ )Ψ (τ ) dτ (14)
0
where ψn (τ ) is the n-fold convolution of the waiting-time density:

τ τ1
ψn (τ ) = ... dτn−1 . . . dτ1 ψ(τ − τn−1 ) . . . ψ(τ1 ). (15)
0 0
The n-fold convolutions defined above are probability density functions for
the sum of n independent variables.
Using the Laplace-Fourier method and recalling the properties of Laplace-
Fourier transforms of convolutions, one gets the following solution of the in-
tegral equation [44, 42, 45, 46]:
∞

p(x, t) = P (n, t)λn (x) (16)
n=0
8 Enrico Scalas
Eq. (16) can also be used as the starting point to derive eq. (12) via the
transforms of Fourier and Laplace, as it describes a jump process subordinated
to a renewal process [14, 47].
Let us now consider the following pseudodifferential equation, giving rise
to anomalous relaxation and power-law tails in the waiting-time probability:
dβ
Ψ (τ ) = −Ψ (τ ), τ > 0, 0 < β ≤ 1; Ψ (0+ ) = 1, (17)
dτ β
where the operator dβ /dtβ is the Caputo fractional derivative, related to
the Riemann–Liouville fractional derivative. For a sufficiently well-behaved
function f (t), the Caputo derivative is defined by the following equation, for
0 < β < 1:

dβ 1 d t f (τ ) t−β
f (t) = dτ − f (0+ ), (18)
dtβ Γ (1 − β) dt 0 (t − τ )β Γ (1 − β)
and reduces to the ordinary first derivative for β = 1. The Laplace transform
of the Caputo derivative of a function f (t) is:
β
d
L f (t); s = sβ f(s) − sβ−1 f (0+ ). (19)
dtβ
If eq. (19) is applied to the Cauchy problem of eq. (17), one gets:
sβ−1
Ψ(s) = . (20)
1 + sβ
Eq. (20) can be inverted, giving the solution of eq. (17) in terms of the Mittag-
Leffler function of parameter β [48, 49]:
Ψ (τ ) = Eβ (−τ β ), (21)
defined by the following power series in the complex plane:

∞
zn
Eβ (z) := . (22)
n=0
Γ (βn + 1)
The Mittag-Leffler function is a possible model for a fat-tailed survival func-

tion. For β = 1, the Mittag-Leffler function coincides with the ordinary ex-
ponential function. For small τ , the Mittag-Leffler survival function coincides
with the stretched exponential:
τβ
Ψ (τ ) = Eβ (−τ β ) 1 − exp{−τ β /Γ (β + 1)}, 0 ≤ τ 1, (23)
Γ (β + 1)
whereas for large τ , it has the asymptotic representation:

sin(βπ) Γ (β)
Ψ (τ ) ∼ , 0 < β < 1, τ → ∞. (24)
π τβ
Accordingly, for small τ , the probability density function of waiting times
ψ(τ ) = −dΨ (τ )/dτ behaves as:
d τ −(1−β)
ψ(τ ) = − Eβ (−τ β ) , 0 ≤ τ 1, (25)
dτ Γ (β)
and the asymptotic representation is:
sin(βπ) Γ (β + 1)
ψ(τ ) ∼ , 0 < β < 1, τ → ∞. (26)
π τ β+1
The Mittag-Leffler function is important as, without passage to the diffu-
sion limit, it leads to a time-fractional master equation, just by insertion into
the CTRW integral equation. This fact was discovered and made explicit for
the first time in 1995 by Hilfer and Anton [50]. Therefore, this special type
of waiting-time law (with its particular properties of being singular at zero,
completely monotonic and long-tailed) may be best suited for approximate
CTRW Monte Carlo simulations of fractional diffusion.
For processes with survival function given by the Mittag-Leffler function,
the solution of the master equation can be explicitly written:
∞ βn
t (n)
p(x, t) = Eβ (−tβ )λn (x), (27)
n=0
n!
where:
(n) dn
Eβ (z) := Eβ (z).
dz n
The Fourier transform of eq. (27) is the characteristic function of p(x, t)
and is given by:
p̂(κ, t) = Eβ [tβ (λ̂(κ) − 1)]. (28)
If log-returns and waiting times are scaled according to:
xn (h) = hξ1 + hξ2 + . . . + hξn , (29)
and:
tn (r) = rτ1 + rτ2 + . . . + rτn , (30)
the scaled characteristic function becomes:
β
t
p̂h,r (κ, t) = Eβ β (λ̂(hκ) − 1) . (31)
r
Now, if we assume the following asymptotic behaviours for vanishing h and r:
λ̂(hκ) ∼ 1 − hα |κ|α ; 0 < α ≤ 2, (32)

10 Enrico Scalas
and
hα
lim = 1, (33)
h,r→0 r β
we get that:
lim p̂h,r (κ, t) = û(κ, t) = Eβ [−tβ |κ|α ]. (34)
h,r→0
The Laplace transform of eq. (34) is:
sβ−1
(κ, s) =
u . (35)
|κ|α + sβ
Therefore, the well-scaled limit of the CTRW characteristic function coincides
with the Green function of the following pseudodifferential fractional diffusion
equation:

|κ|α u
(κ, s) + sβ u
(κ, s) = sβ−1 , (36)
with u(x, t) given by:
1
x
u(x, t) = β/α
Wα,β , (37)
t tβ/α
where Wα,β (u) is given by:
+∞
1
Wα,β (u) = dκ e−iκu Eβ (−|κ|α ), (38)
2π −∞
the inverse Fourier transform of a Mittag-Leffler function [51, 52, 53, 54, 55].
For β = 1 and α = 2, the fractional diffusion equation reduces to the
ordinary diffusion equation and the function W2,1 (u) becomes the Gaussian
probability density function evolving in time with a variance σ 2 = 2t. In the
general case (0 < β < 1 and 0 < α < 2), the function Wα,β (u) is still a
probability density evolving in time, and it belongs to the class of Fox H-type
functions that can be expressed in terms of a Mellin-Barnes integral as shown
in details in ref. [52].
The scaling equation, eq. (33), can be written in the following form:
h rβ/α . (39)
If β = 1 and α = 2, one recognizes the scaling relation typical of Brownian
motion (or the Wiener process).
In the passage to the limit outlined above,
pr,h (κ, s) and u
(κ, s) are asymp-
totically equivalent in the Laplace-Fourier domain. Then, the asymptotic
equivalence in the space-time domain between the master equation and the
fractional diffusion equation is due to the continuity theorem for sequences
of characteristic functions, after the application of the analogous theorem
for sequences of Laplace transforms [56]. Therefore, there is convergence in
law or weak convergence for the corresponding probability distributions and
densities. Here, weak convergence means that the Laplace transform and/or
Fourier transform (characteristic function) of the probability density function
are pointwise convergent (see ref. [56]).
2.2 Limit Theorems: The Coupled Case
The diffusive limit in the coupled case is discussed by Meerschaert et al. [57].
The coupled case is relevant as, in general, log-returns and waiting times are
not independent [28]. Based on the results summarized in [42] and discussed
in [58, 59], it is possible to prove the following theorem for the coupled case:
Theorem
Let ϕ(ξ, τ ) be the (coupled) joint probability density of a CTRW. If, under the
scaling ξ → hξ and τ → rτ , the Fourier-Laplace transform of ϕ(ξ, τ ) behaves
as follows:
(κ, s) = ϕ(hκ,
ϕ rs) (40)
h,r
and if, for h → 0 and r → 0, the asymptotic relation holds:
ϕ
(κ, s) = ϕ(hκ, rs) ∼ 1 − µ|hκ|α − ν(rs)β , (41)
h,r
with 0 < α ≤ 2 and 0 < β ≤ 1. Then, under the scaling relation µhα = νrβ ,
the solution of the (scaled) coupled CTRW master (integral) equation, eq. (3),
ph,r (x, t), weakly converges to the Green function of the fractional diffusion
equation, u(x, t), for h → 0 and r → 0.
Proof
The Fourier-Laplace transform of the scaled conditional probability density

ph,r (x, t) is given by:

1 − ψ(rs) 1

ph,r (κ, s) = . (42)
s
1 − ϕ(hκ, rs)
s), one asymp-

= ϕ(0,
Replacing eq. (41) in eq. (42) and observing that ψ(s)
totically gets for small h and r:
νrβ sβ−1
ph,r (κ, s) ∼ , (43)
νrβ sβ+ µhα |κ|α
which for vanishing h and r, under the hypotheses of the theorem, converges
to:
sβ−1
p0,0 (κ, s) = u
(κ, s) = β , (44)
s + |κ|α
where u
(κ, s) is the Fourier-Laplace transform of the Green function of the
fractional diffusion equation (see eq. (36)). The asymptotic equivalence in the
space-time domain, between p0,0 (x, t) and u(x, t), the inverse Fourier-Laplace
transform of u
(κ, s), is again ensured by the continuity theorem for sequences
12 Enrico Scalas
of characteristic functions, after the application of the analogous theorem for

sequences of Laplace transforms [56]. There is convergence in law or weak
convergence for the corresponding probability distributions and densities.
An important consequence of the above theorem is the following corollary
showing that in the case of marginal densities with finite first moment of wait-
ing times and finite second moment of log-returns, the limiting density u(x, t)
is the solution of the ordinary diffusion equation (and thus the limiting pro-
cess is the Wiener process). The corollary can be used to justify the popular
Geometric Brownian Motion model of stock prices, here with expected return
set to zero. Again, in order to derive this result, no reference is necessary to
the Efficient Market Hypothesis [38, 39].
Corollary
If the Fourier-Laplace transform of ϕ(ξ, τ ) is regular for κ = 0 and s = 0, and,

moreover, the marginal waiting-time density, ψ(τ ), has finite first moment τ0
and the marginal jump density, λ(ξ), is symmetric with finite second moment
σ 2 , then the limiting solution of the master (integral) equation for the coupled
CTRW is the Green function of the ordinary diffusion equation.
Proof
Due to the hypothesis of regularity in the origin and to the properties of

Fourier and Laplace transforms, we have that:
(κ, s) = ϕ(hκ,
ϕ 0) +
rs) ∼ ϕ(0,
h,r

1 ∂2ϕ
∂ϕ
2 2
+ h κ + rs =
2 ∂κ2 ∂s
(0,0) (0,0)
σ2 2 2
=1− h κ − τ0 rs, (45)
2
and as a consequence of the theorem, under the scaling h2 σ 2 /2 = τ0 r, one
gets, for vanishing h and r:
1
p0,0 (k, s) = u
(k, s) = , (46)
s + k2
corresponding to the Green function (36) for α = 2 and β = 1, that is the
solution of the Cauchy problem for the ordinary diffusion equation.
3 Summary and Outlook

In this paper, a discussion of continuous-time random walks (CTRWs) has
been presented as phenomenological models of tick-by-tick market data.
Continuous-time random walks are rather general and they include compound
Poisson processes as particular instances. Well-scaled limit theorems have been
presented for a rather general class of CTRWs.
It is the hope of this author that this paper will stimulate further research
on high-frequency econometrics based on the concepts outlined above. There
are several possible developments.
First of all, one can abandon the hypothesis of i.i.d. log-returns and waiting
times and consider various forms of dependence. In this case, it is no longer
possible to exploit the nice properties of Laplace and Fourier transforms of
convolutions, but, still, Monte Carlo simulations can provide hints on the
behaviour of these processes in the diffusive limit.
A second possible extension is to include volumes as a third stochastic
variable. This extension is straightforward, starting from a three-valued joint
probability density.
A third desirable extension is to consider a multivariate rather than uni-
variate model that includes correlations between time series.
The present author is currently involved in these extensions and is eager
to know progress in any direction by other independent research groups. He
can be contacted at scalas@unipmn.it
4 Acknowledgements
This work was supported by the Italian M.I.U.R. F.I.S.T. Project “High fre-
quency dynamics in financial markets”. The author wishes to acknowledge
stimulating discussion with P. Buchen, S. Focardi, R. Gorenflo, T. Kaizoji, H.
Luckock, F. Mainardi, M. M. Meerschaert and M. Raberto.
References
1. C. Goodhart and M. O’Hara, High-frequency data in financial markets: Issues
and applications, Journal of Empirical Finance 4, 73–114 (1997).
2. M. O’Hara Making market microstructure matter, Financial Management 28,
83–90 (1999).
3. A. Madhavan Market microstructure: A survey, Journal of Financial Markets
3, 205–258 (2000).
4. M.M. Dacorogna, R. Gençay, U.A. Müller, R.B. Olsen, O.V. Pictet, An Intro-
duction to High Frequency Finance, (Academic Press, 2001).
5. M. Raberto, S. Cincotti, S.M. Focardi, M. Marchesi, Agent-based simulation of
a financial market, Physica A 299, 320–328 (2001).
6. H. Luckock, A steady-state model of the continuous double auction, Quantitative
Finance 3, 385–404 (2003).
7. R. Engle and J. Russel Forecasting the frequency of changes in quoted foreign
exchange prices with the autoregressive conditional duration model, Journal of
Empirical Finance 4, 187–212 (1997).
14 Enrico Scalas
8. R. Engle and J. Russel, Autoregressive conditional duration: A new model for

irregularly spaced transaction data, Econometrica 66, 1127–1162 (1998).
9. L. Bauwens and P. Giot The logarithmic ACD model: An application to the bid-
ask quote process of three NYSE stocks, Annales d’Economie et de Statistique
60, 117–149 (2000).
10. A. Lo, C. MacKinley and J. Zhang, Econometric model of limit-order execu-
tions, Journal of Financial Economics 65, 31–71 (2002).
11. G.O. Zumbach, Considering time as the random variable: the first hitting time,
Neural Network World 8, 243–253 (1998).
12. A. Lo and C. MacKinley, An econometric analysis of nonsynchronous trading,
Journal of Econometrics 45, 181–212 (1990).
13. S.J. Press, A compound events model for security prices, Journal of Business
40, 317–335 (1967).
14. P.K. Clark, A subordinated stochastic process model with finite variance for
speculative prices, Econometrica 41, 135-156 (1973).
15. T.H. Rydberg and N. Shephard, Dynamics of trade-by-trade price movements:
Decomposition and models, Nuffield College, Oxford, working paper series 1998-
W19 (1998).
16. T.H. Rydberg and N. Shephard, Modelling trade-by-trade price movements of
multiple assets using multivariate compound Poisson processes, Nuffield College,
Oxford, working paper series 1999-W23 (1999).
17. T.H. Rydberg and N. Shephard, A modelling framework for the prices and
times of trades made at the New York stock exchange, in W.J. Fitzgerald, R.
Smith, A.T. Walden, P. Young (Editors): Nonlinear and nonstationary signal
processing, (Cambridge University Press, 2000).
18. E.W. Montroll and G.H. Weiss, Random walks on lattices, II, J. Math. Phys.
6, 167–181 (1965).
19. R. Hilfer, Stochastische Modelle für die betriebliche Planung, (GBI-Verlag, Mu-
nich, 1984).
20. E. Scalas, R. Gorenflo and F. Mainardi, Fractional calculus and continuous-time
finance, Physica A 284, 376–384 (2000).
21. F. Mainardi, M. Raberto, R. Gorenflo and E. Scalas, Fractional calculus and
continuous-time finance II: the waiting-time distribution, Physica A 287, 468–
481, (2000).
22. R. Gorenflo, F. Mainardi, E. Scalas and M. Raberto Fractional calculus
and continuous-time finance III: the diffusion limit, in M. Kohlmann and S.
Tang (Editors): Trends in Mathematics - Mathematical Finance, pp. 171–180
(Birkhäuser, Basel, 2001).
23. J. Masoliver, M. Montero, and G.H. Weiss, Continuous-time random-walk model
for financial distributions, Phys. Rev. E 67, 021112/1–9 (2003).
24. J. Masoliver, M. Montero, J. Perello. and G.H. Weiss, The CTRW in finance:
Direct and inverse problem http://xxx.lanl.gov/abs/cond-mat/0308017
25. R. Kutner and F. Switala, Stochastic simulation of time series within
Weierstrass-Mandelbrot walks, Quantitative Finance 3, 201–211 (2003).
26. F. Lundberg,
Approximerad Framställning av Sannolikehetsfunktionen. AAterförsäkering av
Kollektivrisker, (Almqvist & Wiksell, Uppsala, 1903).
27. H. Cramér, On the Mathematical Theory of Risk, (Skandia Jubilee Volume,
Stockholm 1930).
28. M. Raberto, E. Scalas and F. Mainardi, Waiting times and returns in high-
frequency financial data: an empirical study, Physica A 314, 749–755 (2002).
29. E. Scalas, R. Gorenflo, H. Luckock, F. Mainardi, M. Mantelli and M. Raberto,
Anomalous waiting times in high-frequency financial data, preprint.
30. L. Sabatelli, S. Keating, J. Dudley and P. Richmond, Waiting time distribution
in financial markets, Eur. Phys. J. B 27, 273–275 (2002).
31. K. Kim and S.-M. Yoon, Dynamical behavior of continuous tick data in futures
exchange market, Fractals 11, 131–136 (2003).
32. T. Kaizoji and M. Kaizoji, Power law for the calm-time interval of price
changes, Physica A, 336, 563–570 (2004).
33. P.C. Ivanov, A. Yuen, B. Podobnik, Y. Lee, Common scaling patterns in inter-
trade times of US stocks, Phys. Rev. E, 69, 056107 (2004).
34. L.J.B. Bachelier Theorie de la Speculation, (Gauthier-Villars, Paris, 1900,
Reprinted in 1995, Editions Jaques Gabay, Paris, 1995).
35. P.H. Cootner (ed.) The Random Character of Stock Market Prices, (MIT Press,
Cambridge MA, 1964).
36. R.C. Merton, Continuous Time Finance, (Blackwell, Cambridge, MA, 1990).
37. M. Parkinson, Option pricing: The american put, Journal of Business 50 21–36
(1977).
38. E.F. Fama, Efficient capital markets: A review of theory and empirical work,
The Journal of Finance 25, 383–417 (1990) .
39. E.F. Fama, Efficient capital markets: II, The Journal of Finance 46, 1575–1617
(1991).
40. D. Cox and V. Isham Point Processes, (Chapman and Hall, London, 1979).
41. J. McCauley, Dinamics of Markets, (Cambridge University Press, Cambridge
UK, 2004).
42. E. Scalas, R. Gorenflo and F. Mainardi, Uncoupled continuous-time random
walks: Solution and limiting behavior of the master equation, Physical Review
E 69, 011107(1-8) (2004).
43. I. M Gelfand and G. E. Shilov, Generalized Functions, vol. 1. Translated from
the Russian. (Academic Press, New York and London 1964. Russian edition:
1958).
44. G.H. Weiss, Aspects and Applications of Random Walks, (North-Holland, Am-
sterdam, 1994).
45. F. Mainardi, R. Gorenflo and E. Scalas, A Renewal process of Mittag-Leffler
type, in M. Novak (Editor), Thinking in Patterns: Fractals and Related Phe-
nomena in Nature, (World Scientific, Singapore, 2004, pp. 35-46).
46. F. Mainardi, R. Gorenflo and E. Scalas, A fractional generalization of the Pois-
son processes, Vietnam Journal of Mathematics, in press (2004).
47. H. Geman, D. B. Madan, and M. Yor, Time changes for Lévy processes, Math-
ematical Finance 11, 79–96 (2001).
48. F. Mainardi, Fractional relaxation-oscillation and fractional diffusion-wave phe-
nomena, Chaos, Solitons & Fractals 7, 1461–1477 (1996).
49. R. Gorenflo and F. Mainardi, Fractional calculus: integral and differential equa-
tions of fractional order, in A. Carpinteri and F. Mainardi (Editors): Fractals
and Fractional Calculus in Continuum Mechanics, pp. 223–276 (Springer, Wien,
1997).
50. R. Hilfer and L. Anton, Fractional master equations and fractal time random
walks, Phys. Rev. E 51, R848–R851 (1995).
16 Enrico Scalas
51. R. Metzler and J. Klafter, The random walk’s guide to anomalous diffusion: a
fractional dynamics approach, Phys. Rep. 339, 1-77 (2000).
52. F. Mainardi, Yu. Luchko, and G. Pagnini The fundamental solution of the space-
time fractional diffusion equation, Fractional Calculus and Applied Analysis, 4,
153–192 (2001). Downloadable from http://www.fracalmo.org.
53. G.M. Zaslavsky Chaos, fractional kinetics and anomalous transport, Phys. Rep.
371, 461–580 (2002).
54. E. Scalas, R. Gorenflo, F. Mainardi, and M. Raberto, Revisiting the derivation
of the fractional diffusion equation, Fractals 11, 281–289 Suppl. S (2003).
55. R. Metzler, Y. Klafter, The restaurant at the end of the random walk: recent
developments in the description of anomalous transport by fractional dynamics,
Journal of Physics A: Math. Gen. 37, R161–R208 (2004).
56. W. Feller, An Introduction to Probability Theory and its Applications, Vol. 2
(Wiley, New York, 1971).
57. M.M. Meerschaert, D.A. Benson, H.P. Scheffler, and P. Becker-Kern, Governing
equations and solutions of anomalous random walk limits, Phys. Rev. E 66,
060102-1/4 (2002).
58. R. Gorenflo and F. Mainardi, Non-Markovian random walk models, scaling and
diffusion limits, 2-nd Conference on Lévy Processes: Theory and Applications,
21-25 January 2002, MaPhySto Centre, Department of Mathematics, University
of Aarhus, Denmark. Downloadable from http://www.fracalmo.org.
59. R. Gorenflo and E. Abdel-Rehim, From power laws to fractional diffusion, Viet-
nam Journal of Mathematics, in press (2004).
Why Macroeconomic Price Indices are
Sluggish in Large Economies ?
Masanao Aoki1 and Hiroshi Yoshikawa2

1
Department of Economics University of California, Los Angeles, Fax Number
310-825-9528 aoki@econ.ucla.edu
2
Faculty of Economics, University of Tokyo, Fax 81-3-5841-5629
yoshikawa@e.u-tokyo.ac.jp
Summary. Two new reasons are discussed for sluggish behavior of macroeconomic
variables such as price indices.
One is slow spread of the news of microeconomic idiosyncratic shocks in the
economy, when the economy is organized into tree structures of heterogeneous sub-
groups or clusters of agents or goods. Clusters are not symmetrically treated, but
the concept of ultrametric distances measure disparities or similarities of clusters.
Another is the effects of uncertainties that affect decision processes, such as
about the cost surfaces, or about the shapes of cost landscapes which may have
many local minima which are not known precisely. Effectiveness of many search
algorithm is reduced in the face of this kind of uncertainty. Flat cost landscapes,
called entropic barriers, are discussed as an example.3
1 Introduction
The standard approach such as real business cycle theory is based on the
premise that the microeconomic behavior of the optimizing agents mimics
dynamics of the macroeconomy. This premise is wrong, because we see that
macro- and micro-behavior are clearly different in many aspects, such as in
their speeds of responses.
Two of the causes for sluggish responses are familiar: complex organi-
zations of macroeconomy, and effects of uncertainty. On complexity of or-
ganization, the existing economic literature treats the phenomena of sluggish
adjustments or responses of economic variables such as prices, wages, or unem-
ployment rates by treating adjusting variables or agents all on equal footing,
that is, without introducing some notions of similarity or closeness between
various heterogeneous groups of variables or agents.
It is true that some rudimentary notion of social distances between dif-
ferent clusters of agents is found in economic literature, such as agents being
3
For some other aspects of uncertainty, see Aoki, Yoshikawa, and Shimizu (2002).
18 Masanao Aoki and Hiroshi Yoshikawa
placed at sites of lattices. However, no explicit notions of similarity, correla-

tions or distances between various groups of variables or agents are examined.
No models with more formal notion of distances among different groups of
agents apparently exist.
The manner of how news of idiosyncratic disturbances of some types or
price changes spreads through the macroeconomy need be analyzed in more
systematic manner by introducing some notion of distance between clusters
of agents.
In our view, attempts at dealing with groups of agents in the economic
literature do not go far enough. We introduce the notion of hierarchically
structured clusters of goods or producers as an essential ingredient in models
designed to explain sluggish adjustment processes.
This paper analyzes a particular aspect of the macroeconomy, namely
adjustment speeds of macroeconomic price indicies. We use the concept of
ultrametrics as measure of distances between clusters.
Since the publication of Keynes’ General Theory (1936), inflexibility or
ridigity of prices has been always a forcal point of macroeconomics. Many
economists take inflexibility of prices as a sign of agents’ irrationality. They
argue that well organized market forces should make prices flexible. In this
paper we explain that prices are necessarily slow to change in large economies.
To improve on the existing literature, we need more appropriate notion
than correlation to measure relations among agents or variables, since the
notion of correlation is not transitive as has been known, for example, in the
numerical taxonomy literature. Feigelman and Ioffe (1991) have an example
of three patterns: A=(1, 1, 1, 1), B=(1,1,-1,1) and C=(1,1,1,-1). Calculating
correlations by ρ = (1/4) i xi yi where xs and ys are the components of the
patterns above, we see that ρA,B = ρA,C = 1/2 but ρB,C = 0.
To avoid this kind of intransitivity of correaltions, which makes correla-
tion unsuitable as a measure of similarity of patterns, we use the notion of
ultrametrics as a measure of distance between clusters of agents. The con-
cept of ultrametrics has been in the literature of mathematics and physics,
especially in spin glass models. For these, see Schikhof (1984), or Mézard,
Parisi and Virasoro (1986), among others. See Aoki (1996, p. 34) for some
elementary explanation of the ultrametrics, and some simple economic appli-
cations. Aoki (2002), and Aoki and Yoshikawa (2003) have a more complex
example of labor market dynamics and Okun’s law,where unemployed workers
from different sectors of economy or different human capitals or job experi-
ences form separate clusters and ultrametrics are used to measure distances
between clusters. This distance is then used to generate probabilities of un-
employed being recalled by a given industry when job openings are created.
Taylor’s well-known analysis of adjustment of wages is different from ours.
He treats groups of workers with different wage contracts as the source of
slow wage/price dynamics. His model and virtually all multi-sector models
treat sectors symmetrically with equal distances between any two sectors.
These groups are not hierarchically arranged. There is no notion of adjust-
Why Macroeconomic Price Indices are Sluggish in Large Economies ? 19
ment speeds as functions of some similarity measures among clusters, Taylor

(1980).
We place clusters of agents as leaves of a tree. Distances between clus-
ters are measured by counting the number of levels one must travel towards
the root of the tree to find a common node shared by the two leaves. This
is called ultrametric distances.4 News, such as that of the arrival of some
idiosyncratic shocks at some sites, spread throughout the tree as stochastic
processes governed by the backward Chapman-Kolmogorov equation, called
the master equaation in this paper, in which transition rates are functions of
ultrametric distances.
2 Tree Models
We use upside down trees to represent hierarchical structures. At the bottom
of a tree we have leaves, also called sites, where each leaf represents a cluster of
agents or a (price of) goods, as the case may be. Agents in the same cluster are
alike in some sense. They may be producers of some close substitute goods, or
they may have similar reaction or decision making delays given a disturbance
of some kind in signals they use, and so on. A number of the leaves, denoted
by m1 , share a common node of a tree. These leaves are connected to nodes
located on level 2 of the tree. There are m2 of the nodes which branch out from
a node at level three, and so on.5 In general we have K levels in a tree. The
top of the tree is the root consisting of a single cluster with N = m1 m2 · · · mK
number of clusters or leaves.
Without loss of generality we assume that an exogenous idiosyncratic dis-
turbance occurs at site 1 at time zero. Let this disturbance be felt at site
i at time t with probability Pi (t). The initial condition is P1 (0) = 1, and
Pi (0), i = 1. These P s are governed by the master equation where transition
rates are functions of ultrametric distances. See Aoki(1996, 2002) for several
examples.
We also use another measure to gauge the speed with which news or dis-
turbances travel through the tree. We define the expected distance reached
by the disturbance originated at site 1 by time t, i.e.,

< d(t) >= d(i, 1)Pi (t),
i
where d(i, 1) is the ultrametric distance between node i and site 1, which is
the source of news or disturbance. This averaged distance indicates how far,
on the average, the news of disturbance has spread through the model.
4
See for example, Aoki (1996, p. 31).
5
Trees need not have symmetric shape or profiles, that is, the number of branches
from a node can be different from node to node.
3 Two Types of Lags in Tree Dynamics

3.1 Multiplier Lags
Responses of a macroeconomic price index to shocks to one of its component

prices consists of two components. One is the well-known dynamic delays in
multipliers or impluse responses. which are familiar in economics and econo-
metrics. The other is called information lag in this paper. It refers to delays
in the news or effects of exogenous shocks which originated in one sector of an
economy spreading to other sectors stochastically. The multiplier lag is simply
illustrated here by dynamic responses of a second order ordinary differential
equation to a step input changes. Propagations of shocks are treated in this
paper stochastically. This requires solving the master equations for states that
are the leaves of trees.
To illustrate this, it is convenient to use Laplace transforms to relate out-
put responses to input changes as
H(s)Y (s) = U (s),
where s is the Laplace transform variable, Y (s) is the transform of the model
output that is response to change in input, and U (s) the Laplace transform
of input.
We may write this more directly as
U (s)
Y (s) = .
H(s)
The time domain expression is
d2 y dy
+ (a + b) + aby = u.
dt dt
Here 1/H(s) is called the transfer function. This expression shows how the
signal at the input side of the model is transferred to the output side of the
model. If the input signal to the system at time zero at full constant force,
then, without loss of generality, we can think of the input signal u(t) = 1 for t
positive. Its Laplace transform is U (s) = 1/s. On the other hand, if the signal
gradually appear to this system, we may have something like u(t) = 1 − e−µt
with µ > 0,for t > 0 say. This input is initially zero and graudally reach
its full force after about the time elapse of 4/µ. Its Laplace transform is
U (s) = µ/[s(s + µ)].
As a simple illustration of the difference of these two types of input signals
on y(t), we assume that dynamics are described by a second order ordinary
constant coefficient differential equation with zero initial conditions; y(0) = 0,
and dy(0)/dt = 0. To be very concrete suppose that H(s) = (s + a)(s + b)
with some positive a and b. This is the transfer function of a dynamic system
described by a second order differential equation with two stable eigenvalues

−a, and −b.
With the step input, the dynamic response is obtained by taking the in-
verse Laplace transform of
1 C A B
Y (s) = = + + ,
s(s + a)(s + b) s s+a s+b
where A, B, and Care the constants, C = 1/ab, A = −1/[(b − a)a], and

B = 1/[(b − a)b]. The time response of the pair of input and output with this
transfer function is given by
y(t) = C + Ae−at + Be−bt .

This y(t) expression shows the multiplier effects of this block or unit of
dynamics with the indicated transfer function. If a < b, then after the time
span of about 4/a units of time, the output nearly settles to a constant, y(t) ≈
1/ab.6 It takes about this much time for the effect of a sudden application of
a step signal at the input to settle down at the output of the model.
With the other input with a gradually rising magnitude such as u(t) =
1 − e−µt , with µ a positive constant much smaller than a and b, then y(t) is
approximately equal to
µ e−at e−bt µ
y(t) = [− + ]+ .
b−a a−µ b−µ (a − µ)(b − µ)]e−µt
The first two exponential terms are due to the dynamic multiplier effects,
and the third term is due to information transmission delay when u(t) grad-
ually appear at the input terminal of this block or unit with the second-order
dynamics.
This expression is approximately equal to the last term above when µ is
much smaller than a or b. The signal y(t) ≈ (1/ab)(1 − e−λt ), which takes
a long time to reach its steady state value. In this case it is the behavior of
input, not the dynamics, that causes the sluggish output.
3.2 Stochastic Spread of News in Trees: An example
We next turn to the second type of lags that exist in trees with several levels
of nodes.
To illustrate our idea simply, we consider two simple economies with four
sectors which are organized in two different ways. One is organized as a one
level tree, and the other as a two level tree. Two-level trees are generally
more sluggish in response than one-level trees. More generally, the larger the
number of levels, the slower the dynamics.
6
If a is larger than b, then the dynamic lag is about 4/b, that is, with two basic
lag structure, it is 4/min(a, b) is the multiplier lag.
Without loss of generality we assume that exogenous disturnaves are felt

at site 1 at time zero. This disturbance is felt at site i at time t with prob-
ability Pi (t). The initial condition is P1 (0) = 1, and Pi (0) = 0, i = 1. We
pay more attention to the transient behavior than the steady state values of
these probabilities becasue the delays are determined by the transient time
constants.
The probability at site i is changing over time as the difference of the
influx and outflux of probabilities. Denote the transition rate between site i
and j by w(i, j). The master equation which describes the dynamics of the
probabilities is
dPi (t)
= Ii (t) − Oi (t),
dt
where the influx to site i is

Ii (t) = Pj (t)w(j, i),
j=i
and
Oi (t) = Pi (t) w(i, j).
j=i
For the one level tree with four sites
I1 (t) = P2 (t)w(2, 1) + P3 (t)w(3, 1) + P4 (t)w(4, 1),
and
O1 (t) = P1 (t)[w(1, 2) + w(1, 3) + w(1, 4)],
and with similar expressions for the other Is and Os. There are similar ex-
pressions for the outflows and inflows at other sites as well. We also assumne
that w(i, j) = w(j, i) for all i and j.
Next, we posit that the transition rates w(i, j) depends only on the ultra-
metric distance. Exogenous disturbances is felt first by agents or goods in the
same cluster, and then the news or effects will gradually and stochastically
propagate to other leaves, that is, to other clusters of the trees. Therefore we
speak of the expected changes in price indicies as the results of such shocks
to one cluster.
Thus for the one-level tree
w(i, j) = q < 1,
i = j, where q = exp(−γd(i, j)) = exp(−γ), for all i and j between 1 and 4,

because the ultrametric distance between any pair of sites is the same, and
where γ is some constant. Later we identify it with the inverse of economic
temperature.
For the two-level tree we have7
7
This model is similar to the one in Aoki (1996, p. 38). Details of analysis differ
somewhat.
w(1, 2) = w(3, 4) = q,
and
w(1, 3) = w(1, 4) = q 2
because d(1, 3) = 2, hence w(1, 3) = exp(−2γ) = q 2 .
The master equation for the probability vector P(t)of the one level tree
consists of probabilities at the four leaves
dP(t)
= W P(t),
dt
with

W1 W2
W = ,
W2 W1
where
−3q q
W1 = ,
q −3q
W2 = qe2 e2 ,
where e2 = [1 1] .
This matrix W has eigenvalue 0 with eigenvector (1 1 1 1) , and triple
repeated eigenvalue −4q with three independent eigenvectors (1 1 − 1 − 1) ,
(1 − 1 0 0) , and (0 0 1 − 1) .
The probabilities evolve with time according to
P1 (t) = 1/4 + (3/4)e−4qt ,
and
P2 (t) = P3 (t) = P4 (t) = (1/4) − (3/4)e−4qt .
Approximately after time span of 1/q, the probabilities are all about 1/4.8
It takes about this time span for the initial shock to propagate to all the
sectors. Hence this is the time lag for the shock initialted at sector 1 to spread
probabilistically to all the other sectors, i. e., for macroeconomic price index
to fully reflect the price shock to one of its sectors.9
In the other case, the matrix W is given by

W1 W2
W = ,
W2 W1
where
8
Note that e−1 = 0.018.
9
All probabilities eventually stop changing and reach some constants, such as 1/4
here. This means that the news of disturbance at site 1 has reached all four sites
equally by then.

−(q + 2q 2 ) q
W1 = ,
q −(q + 2q 2 )
W2 = q 2 e2 e2 ,
where e2 = [1 1] .
This matrix W has eienvalues 0, with eigenvector (1 1 1 1), eigenvalue
λ1 = −4q 2 , with eigenvector (1 1 − 1 − 1) , and double repeated eigenvalue
λ2 = −2(q + q 2 ), with eigenvectors (1 − 1 0 0) , and (0 0 1 − 1) . Note that
the magnitude of λ1 is less than that of λ2 because q is less than one. The
associated with eigenvalue λ1 is faster than that associated with eigenvalue
λ2 . It represents the escape rate of probability from site 1 to site 2.
The probabilities evolve with time as
P1 (t) = (1/4 + (1/4)e−λ1 t + (1/2)e−λ2 t ,
P2 (t) = (1/4 + (1/4)e−λ1 t − (1/2)e−λ2 t ,
P3 (t) = P4 (t) = 1/4 − (1/4)e−λ1 t .

After time span of 2/q(1 + q), the term e−λ2 t is approximately zero. After
time span of 1/q 2 all probabilies are approximately equal to 1/4. Note however,
that the time span 1/q 2 is much longer than that of 1/q, that is, the two-level
tree is much more sluggish than the one-level tree.
To compare dynamic behavior of this model with the one-level tree, we can
aggregate the tree by defining a two-dimensional state vector with components
S1 (t) = P1 (t) + P2 (t), and S2 (t) = P3 (t) + P4 (t) by defining
Q(t) = SP(t),
where the aggregation matrix S is given by

1100
S= .
0011
The dynamic matrix V for this aggregated vector is given by V =

SW S (SS )−1 which has eigenvalues 0 and −4q 2 .
The vector Q(t) has two components 0.5 + 0.5e−µt , and 0.5 − 0.5e−µt with
µ = 4q 2 .
The dynamics of the one-level tree is much simpler. It has eigenvalues 0
and -4q, the latter with multiplicity 3. We can similarly aggregate the first
two sites and the second two sites separately to produce a two node tree. The
eigenvalue are still 0 and -4q. In other words, after the lapse of time of the
order 1/q, the one-level tree has approximately reached its equilibrium state,
while dynamics of the two-level tree has not.
This fact remains true when one-level trees of K sites is compared with k
level trees with K = 2k . Suppose that we group l of K sites into one cluster,
and the remaining K − l sites into another. The eigenvalues are 0 and −Kq,
repeated K − 1 times, while those of k level tree are 0 and −(2q)k .
We can show that the larger the number of hierarchies the slower the pro-
cess of disturbance propagation, and response of macro-price index to shocks
to one of the sectors. Ogielski and Stein (1985),10 among several others, have
shown that in the limit of the number of hierarchy going to infinity, the re-
sponse becomes power-law, not exponential decay. See also Paladin et al.
(1985)
3.3 Inflexible Macroeconomic Prices: An Example
We present an example of slow adjustments of some macroeconomic price

(index) composed of prices of goods of several sectors of economy. To be
simple suppose that a price index PI is the weighted average of two sectoral
output prices, QA and QB . We outline how Sector 1 price QA is affected by an
exogenous shock to site 1 price, since effects on QB (t) are similarly anlalyzed.
For concreteness suppose that node A is composed of a two level tree with
two more nodes a and b with two branches each. There are thus four more
basic prices at sites 1 through 4, such as factor prices, prices of intermediate
goods and so on. The two-level hierarchical tree traces out the relations among
these prices. As shown in the previous section, the tree generates spill-over
probabilities of an exogenous shock to one of the basic prices.
The Laplace transform of P(t), dropping subscript b, is
1 1 1
P̂(s) = u0 + u2 + u1 ,
4s 2(s + λ2 ) 4(s + λ1 )
where s is the Laplace transform variable, us are the column vectors shown
above.
Consequently we can write down the explicit expression for the expected
values of changes in QA , denoted by E[δ Q̂A (s)]. Assuming that transmission
lags of the transfer function ha (s), hb (s), hi (s), i = 1, . . . 4 are not as large as
1/q 2 , we can extract the slowest decaying term out of this as
1
E(δQA (t)) ≈ ha (−λ1 ){h1 (−λ1 ) + h2 (−λ1 )} − hb (−λ1 ){h3 (−λ1 +
4
h4 (−λ1 )}e−λ1 t + · · · ,
where the slowest term is extracted.
In the case where ha (s) = 1/(s + a), hb (s) = 1/(s + b), hi (s) = 1/(s + αi ,
i = 1, . . . 4), then a sufficient condition that this term is present is
Proposition When q 2 is negligibly small compared with a, b, αi , i = 1, . . . 4,
price QA will exhibit sluggish response to an exogenous price change at site
1 if
10
See Aoki (1996, p. 200) also.
1 1 1 1 1 1
( + ) = ( + ).
a α1 α2 b α3 α4
This condition rules out that effects of exogenous shocks coming from two
subtree branches cancel out.
More generally, structures of interconnections between the basic prices and
QA are conveniently expressed in terms of the Laplace transforms as

4
E[δ Q̂A (s)] = Ĥi (s)E[δ q̂i (s)],
i=1
with Ĥi (s) being the Laplace transform of the transfer function from site i to
the price QA . The symbol E is the expectation operator, that is the expected
values of changes in the basic price qi with the probabilities of spill-over.
For example, Ĥi (s) could be a simple first order transfer function such as
(s + ci ) with some positive constant ci , or more complex second or third order
transfer functions, possibly with complex as well as real roots.
The example of this section illustrates the effects of spill-over delays due
to hierarchical tree structure, in addition to the usual delays due to dynamics
of transmission which are present in the transfer functions.
The macroeconomic price indices thus have two sources of sluggishness;
one is the usual dynamic lags of transfer functions , and the other information
spread or spill-over lags, that is, the lags for the signal to arrive as inputs to
some transfer function that are connected with the price index expressions.
4 Effects of Uncertainty
4.1 Rugged Landscape Problems
Thus far, we have focussed on the tree structure of the economy that is re-
sponsible for producing slow dynamic behavior. This result is generic since it
does not depend on any specific assumptions on the model. We now turn to
another reason for sluggish macroeconomy that has to do uncertainty.
The standard analysis in economics assumes explicitly or implicitly that
agents know the global shape of objective functions, which are smooth and
well-behaved, and constraints. In reality, agents have only local knowledge
at best, and must try to improve their performance by guessing the right
directions to adjust their decision varibles. To do so they face complicated
and often hard optimization problems.
Agents are thus often stuck at some local optimal point or in basins asso-
ciated with local optima, and they may not know of the existence of better
local optima or global ones. ”Rugged or flat landscapes” are the words often
used to indicate that agents do not know which directions they should adjust
their decision variables.
Suppose that agents conduct exploratory moves and evaluate the results
with the Metropolis algorithm, Metropolis et al. (1953). See Ripley (1987)for
exposition of the method, for example. In the Metropolis algorithm, the move
is accepted with probability one if it results in lower cost. In order to move
from a local minimum to a better on, the move is accepted with probability
exp(−β∆c), even when the cost increases by ∆c > 0, where β is some positive
parameter. In summary, a trial move is accepted even if a move is associated
with an immediate increas in cost, in the hope that the direction of the move
is correct in eventually it leads to leaving the current basin to a new basin
with lower cost.
If agents are confident of the directions of move, then they will choose
them with high probability, that is, the value of β is small in economies with
high levels of activities. However, if they are uncertain of the moves, then the
move will be chosen with small probability, that is, β has a large value in case
of uncertainty, i.e., of low level of economic activities.
What is important to recognize is the fact that the parameter β may
become very large, especially wen macroeconomy is in recession. With a very
large values of β any move which does not result in immediate cost reduction
will be rejected with very high probability. This means that agents will not
revise their current position with almost probability one. Tis will prolong
the recessions. Put differently, adjustment paths out of the current states are
rarely taken, or become inaccessible with high values of β. Some such examples
are found in Aoki (2002, p.48, p.53; 1996 p.133, p.179) which illustrate effects
of uncertainty associated with an action or decision.
4.2 Economic Temperature
We often say that ”economy is heating up” or ”cooling down”. We propose

a notion of economic temperature to formalize this to convey the idea that
economic activities increase as agents increase their activities by intensifying
levels of existing economic transactions and establishing or creating new links
between agents. In tree models, more sites are introduced and/or more levels
are added to hierarchical trees. When economic activities slow down, inten-
sities of existing activities reduce and some of the existing links may even
disappear.
One way to capture these effects is to change the transition rates, that
is, probability intensities between between some sites. In the above we used
w(i, j) = exp[−γd(i, j)], where γ is some posiitve constant. We now set γ =
1/Te where Te is the economic temperature. For example, as Te becomes lower,
the slower the transition rate, and < d(t) > will be smaller.
More generally, as degree of uncertainty increases, the economic tempera-
ture becomes cooler because of reduced economic activities.
4.3 Entropic Barriers
This section presents a model that focusses on possibilities of losing adjust-

ment path as β becomes large, and not on the aspect of cost minimization.
Agents’ adjustment problem is thus cast as a random walk problems in one
dimension, that is, as a birth-death process. In this setting we recognize that
∆c = 0. We have a flat cost landscape. See Godréche and Luck (2002) for a
similar model.
Suppose that agents face ’cost’ or ’utility’ surfaces with full of local min-
ima. Agents are not sure whether the minima they find themselves are global
minimum or they are stuck in some local minima. They do not know which
directions, if any, they should move in order to improve their performances.
They must modify steepest descent or some gradient procedure to take ac-
count of a possiblity that they need to overcome some barriers to leave the
current basin of attraction to another one with smaller costs. For this rea-
son we assume that they employ the Metroplis mehtod or some variant of
it, Metropolis (1953). This method gives some probability to go in directions
which increases the cost in the hope that the path leads to a local peak of
the barrier separating the basin from another with smaller local minima. We
show that probabilities with which this type of paths is tried become smaller
as economic activity levels become lower.
We suppose there are K boxes (cluster, categories, or types) into which
agents belong. To be concrete, we may think of the boxes as representing
strategies or directions of descents that agents try. In this model economy-wide
optimum corresponds to a situation in which all boxes are empty. Presence
of an agent in some box indicates that agent has not achieved its optimal
operating condition. At a given time we pick an agent at random uniformly
and place it in another box at random. If there are k agents in total, one agent
is chosen with probability k/N where N is the total number of agents in all
the boxes. The chosen agent exits or depart from the box he is in, and will go
to one of the remaining box with probability 1/(K − 1). Call the box he goes
to as the arrival box. Let S denotes (Boltzmann) entropy of the configuration.
The difference
∆S = S(nd − 1, na + 1) − S(nd , na ),
is the change in entropy when one agent departs box d and goes to box a. We
assume that
∆S = −β, nd = 1, na = 0,
∆S = β, nd = 1, na = 0,
and ∆S is zero in all other cases. In other words, entropy changes only when
the number of empty boxes, that is, successful types of firms is reduced by
one (na = 0, nd = 1), which increases S by β, or the number of empty boxes
increases by one, (na = 0, nd = 1), which increases the number of empty boxes
by one.
To treat the simplest cases we focus on one of the boxes, called box 1, and
let n1 (t) denotes the number of agents in it,
pk (t) = Pr(n1 (t) = k).
This and related models have been analyzed by several physicists. We follow
Codrech and Lux 91997). We write the master equation for it as
dpk
= (k + 1)pk+1 + λ(τ )pk−1 − (λ(τ ) + k)pk , k ≥ 2,
dτ
and similar equations for k = 2 and k = 1. To simplify equations we change
the time from τ to t by defining
dt
= λ(τ ).
dτ
With this change of variable we have
dpk k+1 k
= + pk−1 − [1 + ]pk , k ≥ 2,
dt λ(t) λ(t)
dp1 2
= p2 + µ(t)p0 (t) − 2p1 (t),
dt λ(t)
and
dp0 (t)dt = p1 (t) − µ(t)p0 (t),
where
1/λ(t) = 1 + (e−β − 1)p0 (t),
and
µ(t) = e−β + (1 − e−β p1 (t).
See Appendix for the derivation of the transition rates.
We easily verify that these p s sum to one, and the mean of k, k kpk (t) =
N/K := ρ. We take ρ to be 1 for simplicity.
This set of equations can be used to calculate the generating function

F (z, t) = pk (t)z k ,
k
with the initial condition assumed to be F (z, 0) = z, that is p1 (0) = 1 and all
other ps are zero.
This generating function satisfies the partial differential equation
∂F z − 1 ∂F
= (z − 1)F (z, t) − − (z − 1)Y (t),
∂t λ(t) ∂z
where Y (t) = (1 − e−β p0 (t) = 1 − 1/λ(t). See Aoki (2002, p. 70) for deriving
the partial differential equation for the generating function. See also Aoki
(2002, App. A.1) for solving the partial differential equation by the method
of characteristic curves. The sollution is obtained by solving
dt dz dF
= = .
1 (z − 1)/λ(t) (z − 1)(F − Y )
When β = 0 which corresponds to high level of economic activities and

a well-behaved cost curve, the equations is especially simple, because λ(t) =
µ(t) = 1. We obtain
F (z, t) = {1 + (z − 1)e−t } exp{(z − 1)(1 − e−t },
and from it
1 1
p0 (t) = (e−t − 1) exp(e−t − 1) = − + e−2t + · · · .
e 2e
This shows that p0 (t) approaches the equilibrium value with time constant
1/2.
With βlarge Godrech and Lux show that the time constant becomes
eβ /β 2 , a much larger number, indicating a sluggish approach to the equi-
librium.
5 Concluding Discussion
Elementary examples and discussion of ultrametric trees and spaces are given
in Aoki (1996). See Ogielski and Stein (1985) or Paladin, Mezard, and deDo-
minicis (1985) for additional discussion.
Uncertainty and some macroeconomic policies may affect the adjustment
speed parameters. One simple way to incorporate these into qi is to set
qi = exp(−iωq0 /D), i = 1, . . . R, where β is a parameter to reflect the degree
of uncertainty. Larger degree of uncertainty, slower the speed of adjustment.
This implies that ω gets larger as economy faces more uncertainty. Parameter
D stands for some macroeconomic variables. More positive values of D causes
qi to become larger. The formulation presented in this paper can accommomo-
date these changes by incorporating these Gibbs factors into the adjustment
speed parameters. See Paladin et al, or Aoki (1996, p. 156).
References
1. Aoki M (1996) New Approaches to Macroeconomic Modeling: Evolutionary
stochastic dynamics, multiple equilibria, and externalities as field effects, Cam-
bridge Univ. Press, New York
2. Aoki M (2002) Modeling Aggregate Behavior and Fluctuations in Economics:
Stochastic View of Interacting Agents Cambridge Univ. Press, New York
3. Aoki M, Yoshikawa H (2003) A new model of labor dynamics: ultrametrics,

Okun’s law, and transient dynamics, forthcoming Wehia 2003 proceedings,
Springer-Verlag
4. Aoki M, Yoshikawa H, Shimizu T(2002) Long Stagnation and Monetary Policy
in Japan:A Theoretical Explanation, presented at a conference in honor of
James Tobin, a conference volume forthcoming from Springer-Verlag
5. Feigelman MV, Ioffe CB (1991) Hierarchical Organization of Memory in E.
Domany et al (eds) Models of Neural Network, Springer-Verlag, Berlin
6. Godréche C, Luck JM (2002) Nonequilibrium Dynamics of Urn Models, J. Phy.
Cond. Matt 14:1601–1615
7. Metropolis N, Rosenbluth AW, Rosenbluth N, Teller AH, Teller E (1953) Equa-
tion of state calculation by fast computing machine, it Journal of Chem. Phys
21:1087-92
8. Ogielski AT, Stein DL (1985) Dynamics on ultrametric spaces, Phy. Rev.
55:1634–1637.
9. Paladin O, Mézard J, deDominicis C (1985) Diffusion in an ultrametric space
: a simple case.” J. Phys. Lett. 46 L-985–L-989.
10. Ripley BD (1987) Stochastic Simulation Wiley, New York
11. Schikhof WH (1984) Ultrametric Calculus: An Introduction to p-adic Analysis
Cambridge Univ. Press, London.
12. Taylor, JB (1980) Aggregate Dynamics and Staggered Contracts, Jour. Pol.
Econ. 88:1–23
Appendix
For dynamics of trees see also Paladin, Mézard and deDominicis (1985), and
Schreckenberg in Aoki (1996, Sec.7.2.3).
Power-Law Adjustment Behavior
We now discuss general R level trees with large R. The time constants are
getting larger as you go up the levels towards the root. At level i it is given
by
1/τi = r = iR (qr − qr+1 )Nr ≈ qi Ni .
Then for time t between the two sucessive time constants, τl0 −1 and τl0 ,
and closer to the latter, then t is much larger than the former and we have
an approximate expression for x1 in this time interval

x1 (t) exp(−t/τi )/Ni + 1/N ≈ 1/Nl0 ,
i
where 1/τl0 ≈ 1/t ≈ ql0 Nl0 .

To present simple cases, assume that qi = q −i , and mi = mi , i = 1, . . . R.
Then, it is well known that we observe a power-law behavior for the prices at
level 1
This heuristic argument can be made more precise by evaluating the ex-
pression for the sum by approximating it by an integral. See Paladin, Mezard,
and de Cominicis (1985) or Ogielski and Stein (1985) for detail. In Appendix
we outline the argument. One can use steepest descent argument or change
of variables to approximate the sum by an incomplete Gamma function. The
conclusion that power-law emerges does not change.
x1 (t) ≈ t−θ ,
with θ = lnm/ln(q/m).11
Tree Dynamics
A K-site one level tree has dynamics
dP (t)
= W P (t),
dt
where W = −KIK + qeK eK . This matrix has eigenvalue 0, and K1 repeated
eigenvalues of −Kq, where q is the basic transition rate between any two sites.
Let K = 2k for some positive integer k. For example with k = 2 the
one-level 4 branch dynamics have one 0 and three repeated eigenvalue -4q.
To consider a cluster composed of three sites, define a matrix S with the
first row (1110), and the second row (0011). Define
Q = SW S (ss )−1 ,
as the aggregation matrix. This matrix aggregates four sites into 2 clusters of
2 sites each. The dynamics for the two-dimensiona vector r = Sp is given by
dr
= Rr,
dt
where R has eigenvalues 0 and −4q.
We state two propositions:
Proposition 1 With one level tree, the eigenvalues are invariant with the
manner the sites are aggregated into 2 super-sites of size k and K − k.
Proposition 2 Consider a tree with k levels where K = 2k . This tree
aggregted into one level of two clusters of sizes 2k−1 each, has eigenvalues 0
and −(2q)k .
Example of Dynamics of a Two-Level Tree
This tree has three nodes at level 2, that is, three branches come out of the
root, m2 = 2. Each node at level 2 separates into three branches, that is
11
Note that the relation x(t) = Ctθ , andµx(λt) = x(t) implies that θ = −lnµ/λ.
m1 = 3 . At the end of each branch, there is a leaf. Leaves at the bottom

of the tree, that is at level 1, are the states of a continuous-time Markov
chain. For simpler notation, name the states as a, b, c, d, e, f, g, i. Transition
rate between states a and b is denoted by qa,b . The expression qa,f means the
transition rate between state a and f . These two states belong to a different
clusters. Transition rates are assumed to be symmetric, i.e., qa,i = qi,a , and
depends only on the ultrametric distance between the states.
The ultrametric distance between two states is the levels of trees that
these states need to go up the tree toward the root to reach the common
node, that is node from which both states branch out. If we regard the tree as
a genealogical chart with a node representing an ancester, ultrametric distance
1 means that the states share the same parent, distance 2 grandparent.
Denote the three nodes at level 2 by α, β, and γ. These are superstates or
superclusters. leaves d, e, f branch out from node β.
Therefore, states among the same cluster at level 1 have the same tran-
sition rate, q1 . Transition rates between nodes at level 2 are the same and
denoted by q2 . For example, qa,c = q1 , but qa,i = q2 .
The master equation for state a is
dpa
= Ia − Oa ,
dt
where the infow of probability flux is
Ia = pb qb,a + pc qc,a + pd qd,a + · · · pi qi,a ,
and
Oa = pa [(qa,b + qa,c ) + (qa,d + · · · qa,i )].
From our discussion of the ultrametrics above we can write these more
simply as
Ia = q1 (pb + pc ) + q2 (pd + · · · pi ),
and
Oa = pa (m1 − 1)q1 + (2 −1)m1 q2 .
Noting that pb + pc = (pa + pb + pc ) − pa = pα − pa , and that pd + · · · pi =
pβ + pγ = 1 − pα , we obtain
dpa
= q2 + (q1 − q2 )pα − pa [m1 (q1 − q2 ) + N q2 ,
dt
where N = m1 m2 .
The master equation for pb and pc have the same structure. Recall that pα
is the sum of these three probabilities. Adding the three equations together,
we obtain
dpα 1
= −N (pα − ).
dt m2
The solutions of these equations are
1 1
pα (t) = exp(−t/τ2 )[pα (0) − + ,
m2 m2
where
1/τ2 = m1 m2 q2 ,
and
1 1 1 1
pa (t) = exp(−t/τ1 )[pa (0) − pα ] + exp(−t/τ2 )[ pα (0) − ]+ ,
m1 m1 m1 m2 m1 m2
with
1/τ1 = m1 m2 q2 + m1 (q1 − q2 ).
In a R-level tree with transition rates between two states depending only
on the ultrametric distance between them, that is, the difference in levels of
the lowest nodes common to the two states, the structure of these solutions
remain. Denoting what corresponds to level 2 by pjR , where jR runs from 1 to
mR , which is the number of branches from the root, its differential equation
is
dpjR 1
= −N qR (pjR − .
dt mR
The expression for the leaves at level 1 is

R
p(ji , . . . , jR )(0) p(ji+1 , . . . , jR )(0) 1
p(j1 , . . . , jR ) = exp(−t/τi )[ − ]+ ,
i=1
Ni Ni+1 N
where

R
1/τi = (qr − qr+1 )Nr ,
r=i
where Nr = m1 · · · mr , and qR+1 = 0.
This sum equals

R
1/τi = (m/q)i−1 = (λ−i+1 − λ−R )/(1 − λ),
r=i
where λ = q/m. With lambda less than 1 we obtain the usual exponential
dynamic behavior. With the value of λ larger than 1, by letting t̂ = t/τ1 for
convenience, we obtain

x1 (t) = m−i+1 e−t̂ λ−i+1 + 1/N.
The summand is of the form exp[(1/m)i − c exp(−ilnλ]. Define

lnm
y − ey ].
h(y) = exp[−
lnλ

Then the sum can be approximated by h(lny)dy which can be put into the
form of an incomplete Gamma function.
Ultrametric Dynamics
This section describes dynamics on a two-level ultrametric tree of Fig. 1 as a
simple example. A general R level tree dynamics is sketched in appendix.
In the example, the leaves at level 1 are grouped into three clusters, each
cluster composed of three agents (prices). Denote by d(i, j) the distance be-
tween leave i and j. In the two-level tree, this distance takes values of 1 or 2,
depending on the clusters to which i and j belong. If they belong to the same
cluster, the distance is one, else two.
Associate variable xi with leaf i, i = 1, . . . mi , i = 1, 2. This variable could
be price charged by producer or store i, wage rate or some other economic
varibles. On level two each node is labeled by the averages of xs of leaves in
the cluster associated the node. For simpler notation we use pα = (x1 + · · · +
xm1 )/m1 to label the first node on level 2, and so on.12
We assume that xs adjsuts to reduce differences among them. For example,
x1 adjusts with speed q1 to reduce the gap between it and other xs in the same
cluster, and with speed 2 to reduce differences of the averages of xs in other
clusters, since they are at distance 2 from cluster 1. In other words, x1 (t)
adjusts according to the dynamics of
dx1
= q1 (x2 − x1 ) + q1 (x3 − x1 ) + q2 (x4 − x1 ) + · · · + q2 (x9 − x1 )
dt
= q1 (x2 + x3 ) + q2 (x3 + · · · + x9 ) − x1 (2q1 + 6q2 ).
Define the average (price) of cluster 1 by
m1 pα = x1 + x2 + x3 .
Define pβ as the arithmetic average of x4 , x5 , x6 and pγ the average of

x7 , x8 , x9 .
Noting that x2 + x3 = m1 pα − x1 , we rewrite the adjustment equation for
x1 as
dx1
= q1 (m1 pα − x1 ) + m1 q2 (m2 P̄ − pα ) − x1 (m1 (q1 − q2 ) + N q2 ,
dt
where we denote the total number of agents or good prices by N = +m1 m2 ,
and define the over-all average
P̄ = (pα + pβ + pγ )/m2 = (x1 + . . . x9 )/N,
which is a constant by definition.

This equation may be rewritten as
12
More generally in a R-level tree, leaves are labelled by specifying the nodes at all
levels as (j1 , j2 , . . . , jR ), ji = 1, . . . , mi . In this notation x1 = x(j1 = 1, j2 = 1).
Suming x(j1 , j2 ) over j1 = 1, . . . , m1 and dividing the sum by m1 we obtain the
average over a level 1 cluster, (x(j2 = 1), say.
dx1
= N q2 P̄ + m1 (q1 − q2 )pα − x1 [m1 (q1 − q2 ) + N q2 ]. (1)
dt
The time constant of this differential equation is
1/τ1 = m1 (q1 − q2 ) + m1 m2 q2 .
Since the differential equations for x2 and x3 are of the same structure as
that for x1 , we obtain the differential equations for pα by addding them up
and dividing the sum by m1
dpα
= N q2 P̄ − N q2 pα . (2)
dt
The time constant of this average dynamics is
1/τ2 = m1 m2 q2 .
We assume that the tree is such that q1 ≥ q2 . For example information

about prices posted by agents in the same cluster, that is those on level 1
may become available more quickly than information about price averages on
level 2. Then time constant of adjustemt dynamics on level 1 is shorter than
that of level 2. If m1 is large and q1 and q2 differs by an order of magnitude,
then the difference in the two time constants could be very large. This type of
topics is discussed next. In other words, note that 1/τ1 − 1/τ2 = m1 (q1 − q2 )
is positive and large when m1 is large and q1 is much larger than q2 .
Equations (1) and (2) are not the only possible dynamics. We could assume
that a producer reacts to the average price of his cluster, rather than to
individual prices, due to cost of information gathering or some such factors.
Then we have
dx1
= q1 (pα − x1 ) + q2 [pβ − x1 ) + (pγ − x1 )
dt
= m2 q2 P̄ + (q1 − q2 )pα − [q1 + (m2 − 1)q2 ]x1 .
The equation for the average price is
dpα
= m2 q2 P̄ − m2 q2 pα . (3)
dt
In this model, 1/τ1 = q1 + (m2 − 1)q2 , and 1/τ2 = m2 q2 . The difference
in time constants is q1 − q2 . It is still positive, but smaller than that of the
previous model.
Growth Volatility Indices
Davide Fiaschi1 and Andrea Mario Lavezzi2

1
University of Pisa, Dipartimento di Scienze Economiche dfiaschi@ec.unipi.it
2
University of Pisa, Dipartimento di Scienze Economiche lavezzi@ec.unipi.it
Summary. We study the determinants of growth rate volatility in a multisector

economy where sectors are heterogeneous in their individual volatility. We propose
a model where aggregate volatility is explained by structural change and the size of
the economy. We present a first attempt to test these predictions measuring growth
volatility by indices based on Markov transition matrices. Growth volatility appears
to (i) decrease with total GDP and (ii) increase with the share of the agricultural
sector on GDP, although some nonlinearities appear. Trade openness, which we
relate to the size of the economy, also plays a role. In accordance with our model,
the explanatory power of per capita GDP, a relevant variable in other empirical
works, vanishes when we control for these variables.
Key words: growth volatility, Markov transition matrix, structural change,

nonparametric methods.
JEL classification numbers: O11, O40, C14, C21.
1 Introduction
The relationship between growth rate volatility (GRV henceforth), income
levels and other explanatory variables has recently started to receive atten-
tion. Contributions can be divided into two main groups. The first highlights
that development is accompanied by a sharp reduction in GRV (see [1] and
[14]), while the second refers to a negative (power-law) relationship between
the size of an economy and GRV (see [6]). In addition, [2] identify another
possible causal explanation of volatility in the lack of strong “institutions”
(which may, e.g., enforce property rights, reduce corruption and/or political
instability), while [7] focus on the development of the financial sector as a cause
for the reduction in volatility. Therefore, identifying the main determinants

We are very grateful to Carlo Bianchi, Eugene Cleur and seminar participants at
Siena and Guanajuato for helpful comments. The usual disclaimers apply.
38 Davide Fiaschi and Andrea Mario Lavezzi
of growth volatility of a country can be relevant for a better understanding

of the development process, especially in low-income countries, as well as for
the design of economic policies aiming at stabilizing the growth path in un-
derdeveloped economies, as high volatility may be associated to low long-run
growth (see [18]).
In this paper we propose a first attempt to shed light on the determi-
nants of GRV by means of volatility indices. To measure growth volatility we
present a new statistical methodology based on Markov transition matrices,
in contrast to the use of measures of volatility based on the standard devia-
tion of growth rates.3 In particular we propose some growth volatility indices
based the literature on mobility indices (see, e.g. [3]). Specifically, we reinter-
pret a set of indices generally utilized to measure intergenerational mobility
as measures of volatility, and propose two new indices.
Since development is generally proxied by per capita GDP, a first possible
empirical investigation regards the relationship between GRV and per capita
GDP. In this context, we also analyse structural change, a typical phenomenon
associated to development. In fact, a plausible explanation of the reduction
in GRV as development proceeds, resides in the decreasing weight of sectors
with more volatile output, like agriculture and primary sectors, with respect
to sectors with less volatile output, like manufacturing and services.4 Differ-
ently, the increase in the number of sectors (or productive units) associated
to a growing size of the economy is the most common explanation of the rela-
tionship between the size of the economy and GRV. In this case, a reduction in
aggregate GRV may derive from averaging an increasing number of sectoral
growth rates, since idiosyncratic shocks to each sector would tend to cancel
out by the law of large numbers. For example, [6] find that the relation be-
tween growth volatility and the size of the economy takes the form of a power
law.
In this paper we introduce a simple analytical framework and then test
for the existence of these relationships in a large sample of countries from
[12]’s dataset. In particular we focus on the effect of three variables on GRV :
(i) the level of per capita GDP (GDP henceforth) as proxy for the level of
development, (ii) the share of agriculture on GDP (AS henceforth) as proxy
for structural change and (iii) total GDP (TGDP henceforth) as proxy for
the size of the economy. We also consider a measure of trade openness (TR
henceforth), to proxy the effective dimension of an economy which may not
be entirely captured by TGDP alone. Therefore, we limit our analysis to real
factors, without considering monetary factors, e.g. inflation, which can have
explanatory power. In a companion paper (see [9]) we consider a monetary
factor, that is the size of the credit market, and the composition of trade.
3
In a companion paper we measure growth volatility by standard deviations. See
[8].
4
So far the literature on structural change has not paid attention to this issue (see
e.g. [13]).
Growth Volatility Indices 39
Individually, we find evidence of an inverse relationship between GRV and

both GDP and TGDP, and of a positive relationship between GRV and AS
as we expected. However, we show that the relationship between TGDP and
GRV appears more complex than the one found by [6], and that some nonlin-
earities appear in the case of AS. When we consider the interaction among the
variables, the effect of GDP on GRV vanishes when it is considered jointly
with TGDP and AS. The effect of TR, when considered with respect to that
of TGDP, appears also negative. These findings agree with the predictions of
our model, in which GRV is explained by structural change and, especially,
by the size of the economy.
The paper is organized as follows. Section 2 proposes a simple model to
explain the growth volatility of a multisector economy. Section 3 contains a
graphical analysis of GRV based on nonparametric methods; Section 4 dis-
cusses the issue of the measurement of growth volatility and introduces the
GRV indices; Section 5 presents and discusses the results; Section 6 concludes.
2 A Basic Analytical Framework
In this section we present a simple model to highlight the key factors which can
account for GRV in a country. In particular our focus is on the composition
of output and on the size of the economy.
Consider an economy with Nt sectors, where t indexes time. Sector i’s
output grows according to the following rule:

yti = yt−1
i
1 + gti εit ,
where yti is output of sector i in period t, gti is the exogenous growth rate of
sector i, and εti is a random shock.
We assume that shocks are normally distributed with mean 1 and variance
i 2
σ , that is:
2
εit ∼ N 1, σ i .
Let Γt be the Nt × Nt covariance matrix, where γtij is an element. Note that

assuming a nonzero covariance among shocks is a simple way to model sectoral
5
interdependence.
i i We assume that the autocorrelation of the shocks is zero,
that is cov εt , εt−1 = 0, ∀i = 1, ..., Nt and ∀t. Finally, we assume that
σ i−1 ≥ σ i , i = 2, ..., Nt , that is we order sectors on the basis of GRV.6
5
[11] shows that a multisector model with intermediate goods and idiosyncratic
shocks to individual sectors can generate an aggregate dynamics for certain struc-
tures of sectoral outputs’ correlation.
6
[10] endogenize the different volatility of sectors on the basis of rent-seeking the-
ory. Predation may be actually relevant in countries with large primary sectors
(i.e. rich in natural resources). Here, we argue that primary sectors are intrin-
Note that shocks are assumed to be normally distributed for analytical

convenience. In fact, this allows us to measure aggregate GRV by the stan-
dard deviation of the aggregate growth rate. If we relax this assumption, the
measurement of GRV of a country can become complex.
Let Yt be aggregate output in period t, that is:
Nt i
Yt = Σi=1 yt .
Therefore the aggregate growth rate is given by:

Yt − Yt−1 Σ Nt y i 1 + gti εit
µt = = i=1 t−1 Nt i
− 1 = Σi=1
Nt i
αt−1 gti εit , (1)
Yt−1 Σi=1 yt−1
i
i yt−1
where αt−1 = N
t yi
is the share of output of sector i with respect to total
Σi=1 t−1
output, so Nt i
that Σi=1 = 1, ∀t.
αt−1
From definition (1) we have that the expected value and variance of µt are
given by:
Nt i
µ̄t = Et [µt ] = Σi=1 αt−1 gti (2)

2
Nt i
σ̄t2 = Et Σi=1 αt−1 gti ηti , (3)
where ηti = εit −1. Trivially,

η has the same properties of ε, but its mean is equal
i 2
to 0 (that is ηt ∼ N 0, σ
i
). It follows that µt is normally distributed,

that is µt ∼ N µ̄t , σ̄t . From (3) we obtain the following expression for σ̄t2 :
2
Nt
i 2 Nt Nt j i j ij
σ̄t2 = Σi=1 αt−1 gti σ i + Σi=1 Σj=1,j i
=i αt−1 αt−1 gt gt γt , (4)
where γtij − 1 is the covariance between ηti and ηtj .

The functional form of Equation (4) does not allow for a simple identi-
fication of the effects of the elements on the right-hand side on σ̄t2 , except
for gti . An increase of gti , ceteris paribus, increases both µ̄t and σ̄t2 . However,
the effects of the other variables involved,
in particular the number of sectors
1 Nt
Nt and the structure of the economy αt−1 , ..., αt−1 , may not be so easily
identifiable.
sically more subject to random shocks, e.g. changes in terms of trade, climatic
changes and the like. “Good” institutions (or other variables included in the con-
cept of “social capital”) indicated as factors reducing volatility generally emerge
where predation is less important. In this respect, institutions are the endogenous
result of the structure of the economy. In addition, papers such as [19] and [11],
which study the emergence of aggregate fluctuations from sectoral shocks, do not
consider differences in volatility across sectors and thus are not concerned with
structural change.
To proceed, suppose that Yt comes from the agricultural sector A (sector

1), and from the rest of economy R (sectors 2, ..., Nt ), which includes secondary
and tertiary sectors (we will use this distinction in the empirical analysis).
Denoting by αA and αR the shares from the two macrosectors, equation (4)
becomes:
A A A 2 R R R 2
σ̄t2 = αt−1 gt σ + αt−1 gt σt A
+ αt−1 R
αt−1 gtA gtR γtAR . (5)
It is plausible to assume that γtAR = 0 because shocks to A and R are

likely to be of different nature and uncorrelated.7 Therefore we have:
A A A 2 R R R 2
σ̄t2 = αt−1 gt σ + αt−1 gt σt . (6)
A
Generally, a change in αt−1 R
and αt−1 = 1 − αt−1
A
, and/or a change in the
number of sectors Nt have an ambiguous effect on aggregate variance. Let us
analyse first the role of Nt .
Number of Sectors and Growth Volatility Some authors argue that aggregate
GRV is affected by the size of the economy, when measured in terms of num-
ber of sectors or units of production (see e.g. [19]). In our model, the possible
negative correlation between GRV and N can derive from an inverse corre-
lation between σ R and N . We can identify simple conditions under which
dσ R /dN < 0. Assume that gti = g R , αti = α0i = N01−1 , for i, j = 2, ..., Nt .
Thus, from Equation (4) written with respect to R, we have:
2
R 2 gR i 2 i 2 R 2 Nt Nt
σ̄t = Nt
Σi=2 σ + α0 g Σi=2 Σj=2,j=i γtij , (7)
Nt − 1
2
If γtij = 0 for i, j = 2, ..., Nt , then σ̄tR is decreasing in Nt and increasing in
g R , given that σ i−1 ≥ σ i .
Hence, the higher is the number of sectors in R, the lower is the variance
of its growth rate, if the covariance between sectors is negligible (this is an
application of the law of large numbers). If the size of an economy is positively
related to the number of sectors N , then the size of an economy and its growth
volatility are inversely related.
Moreover, higher g R leads to higher GRV but, if the output of some sectors
has a strong positive correlation with the output of others, then GRV can
nonetheless increase if the latter effect is stronger than the effect of the increase
in N . An example can be the emergence of a financial sector, whose output is
correlated to many sectors through the capital market. That is, the appearance
of a global capital market may increase the interdependence among sectors
and affect GRV.8
To conclude, if σ R = σ R (N ), where dσ R /dN < 0, then from Eq. (6) we
obtain:
7
For a discussion of the relationship between σ̄ 2 and Γ see [11].
8
In [9] we study the effects of the financial sector on GRV.
2
∂ σ̄t2 R R 2 d σ R
= αt−1 gt < 0. (8)
∂Nt dNt
We show below that this relationship finds an empirical support when we
proxy for N by the dimension of the economy.9
Composition of Output and Growth Volatility In a typical process of devel-
opment and structural change, primary sectors grow less than industrial and
service sectors. This implies that the share of sectors with higher variance de-
clines over time. The overall result would be a decrease in aggregate GRV, as
the latter is a weighted sum of sectors’ variances, and weights are proportional
to sectors’ shares.
From Eqs. (2) and (6) we have:
A A A 2 2
σ̄t2 = αt−1 gt σ + µ̄t − αt−1
A
gtA σ R .
Calculations lead to:

A 2 −1
∂ σ̄t2 µ̄t σ
A
> 0 ⇔ αt−1 > A 1 +
A
2 = ᾱt . (9)
∂αt−1 gt (σ R )
A A
This means that for αt−1 < ᾱt (αt−1 > ᾱt ) GRV is decreasing (increasing) in
the share of the agricultural sector αA . That is, the relationship between αt−1
A
2 10
and σ̄t is U-shaped. The dependence of the threshold on time introduces
A
some complications, as the threshold itself changes with αt−1 .
Consider two cases: i) gt = gt = g and ii) gt = g < gtR = g R . Case i)
A R A A
refers to the absence of structural change, as the weights of the sectors remain
unchanged. In this case µ̄t = gtA = g and the value of the threshold is the
constant term in square brackets. In case ii), we have αt−1A
→ 0 and µ̄t → g A ,
and A 2 −1
gR σ
ᾱt → A 1 + 2 . (10)
g (σ R )
In case ii) the threshold value is higher than in case i). In general, the threshold
value is lower (and therefore the range of values where GRV is decreasing in
2 2
αA is wider), the higher is σ A with respect to σ R . Moreover, if σ R =
σ R (N ) and dσ R /dN < 0, then the threshold value ᾱ decreases in Nt .
To summarize our results consider the following equation, derived from
Eq. (6):
9
Note that to isolate the effect of Nt , we ruled out the presence of structural change
but, clearly, the two factors interact.
10 A
Note that the U-shaped relation between αt−1 and σ̄t2 resembles the relation
between the variance of a portfolio and the share of the more volatile asset. In
the problem of portfolio choice, the variance of portfolio decreases with the share
of the more volatile asset until a positive threshold value is reached, then increases.

2 A A 2 A 2 R 2 2σ R
σ̄t2 = µ̄t σ R + αt−1 gt σ + σ − A gA
. (11)
αt−1 t
In Eq. (11) aggregate variance depends on two terms: the first term captures
the effect of the variance of the “rest of the economy”, which we argue depends
negatively on the number of sectors N (see Eq. (8)); the second term represents
the effect of the share of agriculture αA , whose sign depends in a non-trivial
way on the relative values of sectoral growth rates and on the interaction with
N , via σ R (see condition (9)). Note finally that GDP does not play any role
in the model.
In our empirical analysis we focus on Eq. (11) and try to shed some light on
the complex relations it highlights. In particular we will consider our variables
in isolation and study their interactions by pairwise relations.11
3 Graphical Analysis
We use data on GDP and TGDP from [12]’s database and data on agriculture
and trade from the World Bank’s World Development Indicators 2002. Our
sample includes 119 countries for the period 1960–1998.12 As noted, we proxy
for the structure of the economy by the share of the agricultural sector in
aggregate value added, AS, and measure the effective dimension of the econ-
omy, related to the number of sectors N in the model, both by the total GDP
(TGDP ) and trade openness (TR), which is the ratio of the sum of imports
and exports on GDP. The latter, jointly with TGDP, provides a more exact
measure of the extent of the overall market for an economy.
We consider both the cross-country and the time-series dimension of
growth volatility. In particular, to evaluate the relation between GRV and
level of development we separate all observations on GDP and TGDP into
151 classes with a similar number of observations (approximately 30), while
to evaluate the relation between GRV and structural change we separate all
observations on AS into 109 classes. Finally, for the relation between GRV
and TR we have 125 classes of observations.13
For every observation on GDP in year t we calculate the growth rate from
t to t + 1.14 Figure 1 reports the standard deviation of growth rates, GRV,
relative to the observations in a class against, respectively, the log of the
11
This choice depends on our methods of empirical investigation. In [8] we estimate
directly Eq. (11) by nonparametric methods.
12
Data on GDP and TGDP are in 1990 international dollars. Not all observations
on agriculture and trade openness were available for each country for all years.
See Appendix A for the country list.
13
In [9] we discuss the methodology for the empirical investigation in more detail.
14
For data on AS, TGDP and TR we consider the corresponding observation on
GDP and calculate the associated growth rate.
0.10
0.10
Log of standard deviation of growth rate

0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
6 7 8 9 10 −4 −3 −2 −1
Log of per capita GDP Log of share of agriculture on GDP
0.10
0.10

0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
12 14 16 18 20 22 −2.0 −1.5 −1.0 −0.5 0.0 0.5
Log of share of total GDP Log of share of trade openness
Fig. 1. GRV estimated by STD vs, respectively, log of GDP, AS, TGDP and TR.
average GDP, AS, TGDP and TR in that class, and run a nonparametric
estimation of these relationships.15
Figure 1 is the counterpart of Figure 1 in [1], where only cross-country
variation in growth volatility is considered. They estimate an OLS regression
and find a decreasing relationship between growth volatility and development,
proxied by the initial level of GDP.
In our case, we see at first glance that GRV tends to fall with GDP. The
high volatility at the lowest and, especially, highest GDP levels is associated
with a much wider variability band, meaning that there the estimate is not
precise. In Figure 1 growth volatility appears to be increasing with AS. This
relation is not monotonic, but the variability band is tighter where the upward
sloping portion is steeper, indicating that the estimation is more precise where
the curve is sharply increasing (we return on this below). In Figure 1 GRV
15
For all the nonparametric estimates we used [17], in particular the statistical
package sm, see [5]. We used the standard settings suggested by the authors (i.e.
optimal normal bandwidth). To test the robustness of this estimate, we ran an
alternative nonparametric regression using the plug-in method to calculate the
kernel bandwidth, and obtained a similar picture. We refer to [4] for more details.
We report the variability bands representing two standard errors above and below
the estimate. They give a measure of the statistical significance of the estimate
(see [4, pp. 29–30] for details on variability bands vs confidence bands). Data sets
and codes used in the empirical analysis are available on the authors’ websites
(http://www-dse.ec.unipi.it/fiaschi and http://www-dse.ec.unipi.it/lavezzi).
clearly decreases with TGDP, as the extreme portions of the estimate have a
wide variability band.
Finally, the relationship between GRV and TR in Figure 1 appears in-
versely U-shaped. In particular, the estimate of both the decreasing parts has
a wide variability band. As noted, the impact of TR on GRV does not interest
us per se, but in conjunction with TGDP when we proxy for the effective size
of an economy. In our view, the effective size of the economy increases if it is
highly integrated with other economies.16
Note that in Figure 1 we have studied the effects on GRV of the variables
taken individually. However, from our model, these variables are expected to
have a joint effect on GRV, that is their effect should be evaluated given the
presence of other variables and of possible interactions among them.17
4 Growth Volatility Indices
In this section we propose a set of synthetic indices to measure GRV.18 In

particular, the measurement of GRV first requires the estimation of a Markov
transition matrix, whose states S = {1, 2, ..., n} represent growth rate classes.
A transition matrix summarizes the information on the dynamics of growth
rates (for more details see [15]), and is the basis to calculate GRV indices.
Heuristically, the indices quantify volatility by the intensity of switches
across growth rate classes. The advantage of the approach based on transition
matrices is that we can keep track of the dynamics of individual countries
in the sample. This is an important aspect which is obscured by measures
of volatility based on standard deviation of growth rates, often found in the
literature (and used in [8]).
For example, consider two countries for which we have observations for,
say, two AS classes and three growth rate classes. In the first AS class both
countries have highly unstable time series, crossing all levels of growth rate
classes. In the second (lower) AS class, both countries have more stable time
series which span only two growth rate classes. However, the first country
has observations in the first and second growth rate class, while the second
has observations in the second and the third growth rate class (in other words
they can have different average growth rates). If we wanted to measure growth
16
In general, as noted by [7, p. 10], the effect of trade openness is ambiguous: trade
may reduce volatility if it represents an increase in the size of the economy, but
it may also increase it by making a country more vulnerable to external shocks.
We remark that a relevant issue is the composition of trade. For instance, a high
level of trade openness associated to a concentration of export in few goods (e.g.
primary or oil) may result in a positive relation between trade openness and
growth volatility. We study the effect of trade composition in [9].
17
In the nonparametric estimation in [8] we explicitly consider the interactions by
introducing interaction terms in the regressions.
18
The statistical properties of the indices are derived in [9].
volatility in the two AS classes, we would not be able to detect the decrease in
volatility if we used the standard deviation, while we would correctly detect
the reduction in volatility if we used a transition matrix. To evaluate the
relationship between GRV and, for instance, AS we calculate the values of
the indices for different classes of AS.
To define indices of GRV we draw on studies of inter- and intragenera-
tional mobility of individuals (see, among others, [3, pp. 24–30] and [20]), and
propose two new indices. Basically, these indices are functions of the elements
of a transition matrix. In a transition matrix high values on the principal
diagonal indicate low mobility, while the values of off-diagonal elements refer
to changes of state and, therefore, high values of the latter are associated to
high mobility.
A simple mobility index is the following, proposed by [20]:
n − trace(P)
I S (P) = , (12)
n−1
where P is a transition matrix of dimension n. The range of the index is
[0, n/ (n − 1)] and a high value means high mobility. However, I S is not well-
suited to measure growth volatility because it is not affected by the value of
off-diagonal elements, a key point for the present analysis, but we refer to it
as a term of comparison with the other indices discussed below.
[3, p. 28], proposes the following index which takes explicitly into account
the distance covered by a transition from i to j, (i, j ∈ S), when the states
correspond to increasing or decreasing values of a variable:
1
n n
B
I (P) = πi pij |i − j| . (13)
n − 1 i=1 j=1
In I B , pij is an element of the transition matrix P, while πi is an element of

the associated ergodic distribution.19 The range of I B is [0, 1]: a higher value
means higher mobility.
In this case only the absolute value of the difference between i and j is
taken into account. It is worth verifying the effect of increasing the weight
attached to “longer” jumps, in order to better appreciate the magnitude of
the fluctuations. Therefore we introduce the following index:
1
n
n
I BM (P) = 2 πi pij (i − j)2 , (14)
(n − 1) i=1 j=1
in which the distance of the transition enters in a quadratic form. As before

I BM ∈ [0, 1] and a higher value means higher mobility/volatility.
19
The ergodic distribution represents the long-run distribution of the Markov pro-
cess. For more details see [3].
Indices I B and I BM weight the transitions from growth rate class i by

the corresponding mass in the long-run equilibrium (i.e. in the ergodic distri-
bution). In other words, considering the elements of the ergodic distribution
as weights amounts to measuring GRV in the long-run equilibrium. How-
ever, also the volatility along the transition path can reveal very interesting
information. The following indices fill this gap:
1
n n
I F L (P) = pij |i − j| ; (15)
A i=1 j=1
1
n n
I F LM (P) = pij (i − j)2 . (16)
A2 i=1 j=1
I F L and I F LM respectively correspond to I B and I BM , except for the ab-

sence of the elements of the ergodic distribution. The constant A normalizes
both indices to the range [0, 1].20 A higher value still means higher mobil-
ity/volatility.
5 Empirical Results
In the following we study the relation between GRV, GDP, AS and TGDP by
calculating the values of the indices described in Section 4. As in Section 3,
we first evaluate the individual effects of our variables, and then study their
interactions, with particular attention to the explanatory power of GDP.
In particular, in a first stage: (i) we separate the observations on GDP, AS
and TGDP in four classes for each variable, from “low” to “high” values; (ii)
we calculate the transition matrix with five growth rate classes for each class,
(iii) we compute indices (12), (13), (14), (15), (16) for every transition matrix
and, finally, (iv) we make inference on these estimates. In a second stage we
evaluate the interactions of the variables in this framework.
First we define the five growth rate classes common to all three variables.
We set the central class to include the average growth rate of the sample,
equal to 2%, and define the other classes symmetrically around this central
class.21 With this criterion we obtain the state space:22
S = {[−∞, −2%) , [−2%, 1%) , [1%, 3%) , [3%, 6%) , [6%, +∞)} . (17)
20
In particular: (
2Σ n−1
n−1 i+ n−1
2
for n odd;
i= 2 +1
A= n−1
2Σi= ni for n even.
2
21
In [9] we propose an alternative method of calculating the growth rate classes
based on the density of observations.
22
Results are not affected by slight changes of the classes’ limits.
Index\GDP class I II III IV

IS 0.8313 0.7831 0.7417 0.7793
(0.0195) (0.0192) (0.0193) (0.023)
IB 0.2583 0.2474 0.2272 0.1983
(0.0108) (0.0093) (0.0086) (0.0077)
BM
I 0.1548 0.1382 0.1206 0.0888
(0.0099) (0.008) (0.0074) (0.0062)
FL
I 0.4178 0.3658 0.3136 0.3253
(0.0149) (0.0126) (0.0115) (0.0162)
F LM
I 0.3199 0.259 0.2083 0.2148
(0.0169) (0.0138) (0.0122) (0.0188)
Table 1. Growth volatility indices. Standard errors in parenthesis. GDP
Alternatively, the state space for the calculation of the transition matrices
could be based not on absolute values of growth rates, but on deviations from
the trend, as in [9].
5.1 Per Capita GDP
We define four GDP classes (in logs) which contain the same number of ob-
servations (≈ 1100), obtaining the following:
I = [0, 6.98) , II = [6.98, 7.9) , III = [7.9, 8.82) , IV = [8.82, +∞) .
For every GDP class we estimate a transition matrix relative to the state
space S.23 Table 1 contains the values of the indices calculated for each of the
four transition matrices.
We observe that in all cases the value of the index is generally decreasing
with respect to the GDP class and that, in particular, the value of the index
in the first GDP class is always higher than in the last. This result broadly
agrees with Figure 1 in which volatility is measured by the standard deviation
of growth rates.
We tested for the joint equality of the estimated value and strongly rejected
the null-hypothesis.24
23
The four transition matrices and the four ergodic distributions (one for each GDP
class) are obtainable with the codes available on the authors’ websites.
24
[8] show that these indices are asymptotically normally distributed. The test
statistic: ş ş ť ť 2
r
nΣi=1 I P̂ i − I (P ) / (r − 1)
Fr−1,r(n−1) = σ̂ 2i
,
r I
Σi=1 n
/r
where r is the number of indices (r = 4), n is the number of observation used
for the computation of every index σ̂I2i /n, i = 1, 2, is the variance of the value
of the index calculated from transition matrix P̂i and I (P ) is the average value
of index, is distributed as a ”‘F”’ distribution with r − 1, r (n − 1) degrees of
freedom.
Index\GDP class I vs II I vs III I vs IV

IS 0.04* 0* 0.04*
IB 0.22 0.01* 0*
I BM 0.10 0* 0*
IF L 0* 0* 0*
I F LM 0* 0* 0*
Table 2. Test of equality between the GRV index of GDP class I versus its value
in the other classes. * means rejection of the null hypothesis of equality at 5%
confidence level.
Table 2 reports the p-values of tests of a null hypothesis of equality between

the value of the index in the first GDP class versus its value in each of the
other GDP classes, for all the indices.
Tests confirm that GDP class I generally has a statistically significant
higher GRV. At a conventional 5% level, the null hypothesis is not rejected
only in the comparison between the value of the index in the first and in the
second GDP class for indices I B and I BM (but it is rejected at 10% for the
latter).
To check if there is a monotonic decreasing relationship among the values
of the indices at different GDP levels we also tested the following hypotheses of
equality (details omitted): (i) GDP class II vs GDP class III; (ii) GDP class
III vs IV . In case (i), the hypothesis is strongly rejected for I F L and I F LM ,
and is rejected at approximately 6% level for the other indices; in case (ii) the
hypothesis is rejected only for I B and I BM . However, in the other cases we do
not reject the null hypothesis that the indices in GDP classes III vs IV are
equal (note that in Table 1 the value of the index in GDP class IV is actually
higher than in GDP class III for I S , I F L and I F LM ). Hence, according to
indices I B and I BM , the decrease is statistically significant when we move
from class II onwards, while with I F L and I F LM the decrease is statistically
significant from class I, but for the two higher GDP classes the relation may
be flat. Also, for I S we do not find evidence of a monotonic decrease as the
value of the index in GDP classes III and IV may be equal.
Overall, the indices indicate the presence of a negative relationship between
GRV and GDP, which may become flat in some GDP ranges. In any case,
a comparison between the first and the last GDP classes always shows a
significantly higher volatility in the former.
5.2 The Dimension of the Economy
In this section we repeat the exercise considering TGDP to proxy for the
dimension of the economy.25 We define four TGDP classes (in logs) with the
same number of observations (≈ 1100):
25
In the next section we examine the interaction of TGDP with TR. In [9] we also
compute the indices for TR classes.
Index\TGDP class I II III IV

IS 0.7721 0.7998 0.7799 0.7339
(0.0185) (0.0182) (0.0186) (0.0193)
IB 0.2904 0.2749 0.2645 0.2155
(0.0104) (0.0085) (0.0088) (0.0072)
BM
I 0.18 0.1524 0.1462 0.1001
(0.01) (0.0076) (0.0081) (0.0055)
FL
I 0.3682 0.3478 0.3449 0.2846
(0.0116) (0.0107) (0.0115) (0.0103)
F LM
I 0.2725 0.2326 0.2375 0.1679
(0.0124) (0.011) (0.0121) (0.01)
Table 3. Growth volatility indices. Standard errors in parenthesis. TGDP
Index\TGDP class I vs II I vs III I vs IV

IS 0.14 0.38 0.08
IB 0.12 0.03* 0*
I BM 0.01* 0* 0*
IF L 0.10 0.07 0*
I F LM 0* 0.02* 0*
Table 4. Test of equality between the GRV index of TGDP class I versus its
value in the other classes. * means rejection of the null hypothesis of equality at 5%
confidence level.
I = [0, 15.45), II = [15.45, 16.6), III = [16.6, 18.2), IV = [18.2, 2.90].
With this class definition, and with the same state space for growth rate
classes, we obtain the volatility indices in Table 3.
In this case we observe a monotonic decrease for indices I B , I BM and I F L
across the TGDP classes, while for I F LM the value is higher in TGDP class
III than in II. Finally, no clear relation emerges from I S .
From Table 4 we see that, at 5% confidence level, the value of the index in
class I is statistically different from the value in class II in two cases, from the
value in class III in three cases, and from the value in class IV in four cases.
However, (i) in neither case we can reject the hypothesis of equality between
the values in TGDP classes II and III (details omitted), (ii) we always reject
the hypothesis of equality between the values of the indices in classes III and
IV .
This is in agreement with Figure 1, in which the relation between GRV
and TGDP is slightly flatter at intermediate TGDP levels, and is nonmono-
tonic (although with a relatively large variability band) at low TGDP levels.
Therefore, the relationship is clearly negative when we compare the extreme
TGDP classes, but it appears overall more complex than the one identified
by [6], who find a linear, decreasing relationship.26
26
They use a smaller dataset and a lower number of TGDP classes, and consider
as an outlier the only observation for low TGDP which would produce nonlinear
behavior.
Index\AS class I II III IV

IS 0.7428 0.8152 0.7834 0.8756
(0.0229) (0.0214) (0.0223) (0.0207)
IB 0.2297 0.2735 0.2576 0.3199
(0.0088) (0.0105) (0.0102) (0.0115)
BM
I 0.1105 0.1501 0.1377 0.1915
(0.007) (0.0094) (0.0092) (0.0114)
FL
I 0.3037 0.3611 0.3364 0.4119
(0.0132) (0.0133) (0.0136) (0.0138)
I F LM 0.1905 0.2478 0.2249 0.2975
(0.0134) (0.014) (0.0142) (0.0151)
Table 5. Growth volatility indices. Standard errors in parenthesis. AS
Therefore, keeping in mind these remarks, we find a broad confirmation of

the existence of a negative relation between GRV and TGDP, in particular
when comparing the extreme classes.
5.3 Structural Change
In this section we address the relationship between GRV and structural change
proxied by AS. We first define four AS classes with the same number of
observations (≈ 784). The resulting classes’ limits are:
I = [0, 0.08), II = [0.08, 0.2), III = [0.2, 0.33), IV = [0.33, 1].
Table 5 contains the volatility indices calculated with this class definition.
Results seem to be in accordance with the pattern in Figure 1. Moving from
high to low levels of AS, that is following a typical development path, volatility
decreases from class IV to class III, then increases in class II and decreases
again in class I. However, tests of equality between the value of the indices
in class III and in class II do not allow to reject the null hypothesis at
conventional 5% level.27 Finally, volatility is significantly higher in class IV
than in class I: the hypothesis of equality between the value of the index in
class I and in class IV is strongly rejected for all indices (we omit the details
of the tests).
At this stage, we take this result as indicating the possible presence of a
more complex behavior at intermediate levels of AS, which is in accordance
with the non–monotonic pattern of GRV found in Figure 1. Note that indices
calculated for classes II and III are not significantly different at 5% level,
but only at about 15%.
5.4 On Conditioning
In [8] we present the results of nonparametric estimations suggesting that

GDP is not informative when TGDP, TR and AS are considered. In other
27
The p-values of the tests for I S , I B , I BM , I F L , I F LM are, respectively, 0.15, 0.14,
0.17, 0.10, 0.13.
words, when the latter explanatory variables are present in a (nonparametric)

regression, GDP does not provide further information on GRV.
Here we address this issue in the approach based on the volatility indices
computed from a Markov transition matrix. First, note that the analysis in
the previous section can be considered as deriving from the estimation of
conditioned transition matrices. In fact, the basis for the calculation of each
index is a transition matrix indicating the probabilities to observe transitions
across growth rate classes, starting from a given growth rate class and a given
e.g. TGDP class.28
The formal definition of a transition probability from growth rate class Si
to growth rate class Sj , given that the observation is in TGDP class I is:
p (gt ∈ Sj |gt−1 ∈ Si , T GDPt−1 ∈ I) =

p (T GDPt−1 ∈ I|gt ∈ Sj , gt−1 ∈ Si )
= p (gt ∈ Sj |gt−1 ∈ Si ) . (18)
p (T GDPt−1 ∈ I|gt−1 ∈ Si )
In Equation (18) the term on the left–hand side is an element of the condi-
tioned transition matrix from which we derived our indices relative to TGDP
class I. The first term on the right-hand side is an element of the transition
matrix for growth rate classes only; the second term is the probability that,
given a transition from growth rate class Si to Sj , the initial growth rate is
associated to a TGDP in class I.
If the conditioning variable TGDP is not relevant, p (gt ∈ Sj |gt−1 ∈ Si ,
T GDPt−1 ∈ I) = p (gt ∈ Sj |gt−1 ∈ Si ): any transition matrix calculated con-
sidering alternative values of TGDP would not be statistically different from
the unconditioned transition matrix.29 Therefore, GRV indices calculated
from the former would not be statistically different from each other, and
from those calculated from the unconditioned transition matrix. In the same
manner, if we condition on two variables, e.g. TGDP and GDP, we have:
p (gt ∈ Sj |gt−1 ∈ Si , T GDPt−1 ∈ I, GDPt−1 ∈ I) =

= p (gt ∈ Sj |gt−1 ∈ Si , T GDPt−1 ∈ I) ∗

p (GDPt−1 ∈ I|gt ∈ Sj , gt−1 ∈ Si , T GDPt−1 ∈ I)
∗ (19)
p (GDPt−1 ∈ I|gt−1 ∈ Si , T GDPt−1 ∈ I)
and the same reasoning for the relevance of GDP, given TGDP, applies.
Here we do not provide a complete discussion of this issue, but only some
evidence on the relevance of TR, AS and GDP in explaining GRV given
TGDP. Namely, we compare the values of two indices, I B and I F L , computed
for TGDP classes only, with the values obtainable when the transition matrix
28
We could have estimated an unconditioned transition matrix for growth rate
classes only, which does not distinguish among the TGDP levels associated to
each transition. An example of conditioned Markov chains is in [16].
29
From the condition p (gt ∈ Sj |gt−1 ∈ Si , T GDPt−1 ∈ I) = p (gt ∈ Sj |gt−1 ∈ Si )
derives p (T GDPt−1 ∈ I|gt ∈ Sj , gt−1 ∈ Si ) = p (T GDPt−1 ∈ I|gt−1 ∈ Si ).
Uncond. AS(I) AS(II) AS(III) AS(IV )

T GDP (I) 0.2904 0.1923∗∗ 0.3079 0.3510 ∗∗ 0.3419∗∗
(0.0104) (0.02921) (0.02720) (0.02908) (0.01852)
T GDP (II) 0.2749 0.3174∗∗ 0.2725 0.2503 0.2958
(0.0085) (0.02348) (0.02304) (0.01817) (0.01962)
∗∗ ∗
T GDP (III) 0.2645 0.2517 0.2989 0.2188 0.2876
(0.0088) (0.02355) (0.01961) (0.01912) (0.03172)
T GDP (IV ) 0.2155 0.2027 0.2204 0.2307 −
(0.0072) (0.01091) (0.01741) (0.01859)
Table 6. Values of I B for TGDP conditioned on AS. ** indicates rejection of null

hypothesis of equality to the index in the first column at 5% level; * at 10% level
Uncond. AS(I) AS(II) AS(III) AS(IV )

T GDP (I) 0.3682 0.2650∗∗ 0.3812 0.4679∗∗ 0.4377∗∗
(0.0116) (0.03750) (0.03108) (0.03331) (0.02014)
T GDP (II) 0.3478 0.3935∗ 0.3641 0.3532 0.4010∗∗
(0.0107) (0.02679) (0.02966) (0.02738) (0.02570)
∗
T GDP (III) 0.3449 0.3216 0.4099 0.3037 0.3932
(0.0115) (0.02692) (0.02415) (0.02582) (0.03876)
∗
T GDP (IV ) 0.2846 0.3330 0.2929 0.2984
(0.0103) (0.02834) (0.02299) (0.02618)
Table 7. Values of I F L for TGDP conditioned on AS. ** indicates rejection of null

is calculated for every TGDP class conditioned to each class of AS, TR and
GDP. The choice of TGDP as the “pivotal” variable is dictated by its high
explanatory power, as assessed in [8].30
Tables 6 and 7 consider the relevance of the information provided by AS
when TGDP is the principal variable, respectively, for index I B and index
I F L.
The first column of the tables contains the volatility index obtained for
TGDP classes (see Table 3). The other columns contain the values of the index
when, for each TGDP class, we condition on each AS class. If AS matters,
then GRV should increase with AS, and the conditioned indices should be
statistically different from the unconditioned indices in the first column.31
What we observe is that: given a TGDP level, an increase in AS is gen-
erally associated to an increase in GRV (at least if we compare AS(I) to
AS(IV)), confirming the insight that a higher agricultural share causes a
30
See in particular Table 1, although present results are not completely comparable
with those of [8]. In that case more variables were considered jointly, while here
we investigate pairwise relations, in which there is one principle variable, TGDP
or AS, and only another one interacting with it.
31
Each conditioned indices is calculated starting from the observations belonging
to a TGDP class and an AS class. We chose to consider only indices calculated
from a minimum number of observations, at least 75. An alternative test can be
conducted, aiming at testing the joint significance of the differences between the
conditioned and unconditioned indices. We leave this issue on a side for future
research.
Uncond. T R(I) T R(II) T R(III) T R(IV )

T GDP (I) 0.2904 0.3494∗ 0.3643∗∗ 0.2838 0.3080
(0.0104) (0.03658) (0.03172) (0.02427) (0.01845)
T GDP (II) 0.2749 0.2875 0.2668 0.2810 0.2900
(0.0085) (0.02191) (0.02064) (0.01773) (0.01771)
∗
T GDP (III) 0.2645 0.2901 0.2581 0.2363 0.2399
(0.0088) (0.02303) (0.01893) (0.01873) (0.01828)
∗∗ ∗ ∗
T GDP (IV ) 0.2155 0.2404 0.2000 0.1936 0.1825
(0.0072) (0.01331) (0.01275) (0.01453) (0.01903)
Table 8. Values of I B for TGDP conditioned on TR. ** indicates rejection of null

Uncond. T R(I) T R(II) T R(III) T R(IV )

T GDP (I) 0.3682 0.4671∗∗ 0.4684∗∗ 0.3691 0.3871
(0.0116) (0.03962) (0.03220) (0.02738) (0.02101)
T GDP (II) 0.3478 0.3817 0.3600 0.4064∗∗ 0.3632
(0.0107) (0.02819) (0.02774) (0.02600) (0.02177)
∗
T GDP (III) 0.3449 0.3678 0.3551 0.3375 0.3120
(0.0115) (0.02748) (0.02418) (0.02755) (0.02231)
∗∗ ∗∗
T GDP (IV ) 0.2846 0.3380 0.2722 0.2578 0.2359
(0.0103) (0.01889) (0.02386) (0.02199) (0.02489)
FL
Table 9. Values of I for TGDP conditioned on TR. ** indicates rejection of null
higher GRV. However, the conditioned indices seem to be statistically dif-

ferent from the unconditioned ones especially for TGDP(I). In fact, for both
I B and I F L , in three out of four cases the difference is statistically signifi-
cant at 5% level. This indicates that structural change may be particularly
important if the economy is “small”. “Large” economies do not seem to be
particularly affected by a reduction in AS.
We have observed that the integration in world markets may reduce GRV,
if the effective size of an economy results increased, or it can increase GRV if
it allows for the “import” of external shocks (recall the remark from [7]). In
general, we need a way to measure the effective size of an open economy, and
therefore something that takes jointly into account TGDP and TR.
Here we suggest to consider TR as a conditioning variable, with respect to
TGDP.32 This is motivated by the fact that, in general, it is possibible to find
economies with all possible situations: high TGDP and high TR, high TGDP
and low TR, etc. so that the definition of a joint measure can be difficult.
Tables 8 and 9 analyse the relation between TGDP and TR.
We find that, comparing the values of the indices for TR(I) and TR(IV),
GRV generally decreases. However, the conditioned indices are statistically
different from the unconditioned ones especially for TGDP(I) and TGDP(IV)
32
In [8] we consider TGDP and TR in interaction terms in a regression.
Uncond. GDP (I) GDP (II) GDP (III) GDP (IV )

T GDP (I) 0.2904 0.3197∗ 0.2956 0.2493∗∗ −
(0.0104) (0.01579) (0.01855) (0.02079)
T GDP (II) 0.2749 0.2763 0.2706 0.2559 0.3272
(0.0085) (0.01673) (0.01468) (0.01424) (0.031413)
∗∗ ∗ ∗
T GDP (III) 0.2645 0.3214 0.2661 0.2414 0.2378
(0.0088) (0.02679) (0.01948) (0.01520) (0.014905)
∗∗
T GDP (IV ) 0.2155 − 0.2183 0.2275 0.1949
(0.0072) (0.02001) (0.01546) (0.008536)
B
Table 10. Values of I for TGDP conditioned on GDP. ** indicates rejection of
null hypothesis of equality to the index in the first column at 5% level; * at 10%
level
Uncond. GDP (I) GDP (II) GDP (III) GDP (IV )

T GDP (I) 0.3682 0.4102∗∗ 0.3770 0.3435 −
(0.0116) (0.01731) (0.02158) (0.02594)
T GDP (II) 0.3478 0.3939∗∗ 0.3584 0.3257 0.4427∗∗
(0.0107) (0.02278) (0.01889) (0.01936) (0.04050)
∗
T GDP (III) 0.3449 0.4263 0.3507 0.3074 0.3173
(0.0115) (0.03413) (0.02367) (0.01851) (0.02052)
T GDP (IV ) 0.2846 − 0.3067 0.2891 0.2900
(0.0103) (0.03041) (0.02013) (0.01776)
Table 11. Values of I F L for TGDP conditioned on GDP. ** indicates rejection of

null hypothesis of equality to the index in the first column at 5% level; * at 10%
level
33
. At 5% or 10% level, the difference is statistically significant for two or three
TR classes.
Conditioning on GDP
We concentrate now on the role of GDP. Our hypothesis is that the informa-
tion on GDP is not relevant when we control for the dimension of the economy
and for structural change, proxied by the share of the agricultural sector. We
first evaluate GDP against TGDP in Tables 10 and 11.
We expect to find a decreasing GRV with GDP, and this tendency can
be partially found in the results.34 On the other hand, there appear to be no
TGDP class for which the inclusion of GDP produces statistically different
values for the GRV indices, with the exception of TGDP(III) for I B , in which
three out of four values are statistically different from the unconditioned value.
However, the most important test for the relevance of GDP is the analysis
of its role in presence of AS, which is directly connected to economic develop-
33
Note that the remark that all possible combinations between the levels of TGDP
and TR are possible is confirmed by the fact that there are enough observations
to compute all values of the conditioned indices.
34
The value of 0.4427 in Table 11, which would in any case contradict the hypothesis
of decrease of GRV with GDP, is based on only 88 observations and is therefore
scarcely relevant.
Non cond. GDP (I) GDP (II) GDP (III) GDP (IV )
AS(I) 0.2297 − − 0.2464 0.2258
(0.0088) (0.01646) (0.01071)
AS(II) 0.2735 − 0.3010 0.2583 0.2346∗∗
(0.0105) (0.02414) (0.01363) (0.02073)
AS(III) 0.2576 0.2965∗ 0.2480 0.2504 −
(0.0102) (0.02742) (0.013) (0.02061)
AS(IV ) 0.3199 0.3260 0.3091 − −
(0.0115) (0.01340) (0.02383)
Table 12. Values of I B for AS conditioned on GDP. ** indicates rejection of null

Non cond. GDP (I) GDP (II) GDP (III) GDP (IV )
AS(I) 0.3037 − − 0.3133 0.3103
(0.0132) (0.02310) (0.01724)
AS(II) 0.3611 − 0.4176∗∗ 0.3349 0.3327
(0.0133) (0.02728) (0.01752) (0.02742)
AS(III) 0.3364 0.3954∗ 0.3213 0.3350 −
(0.0136) (0.03469) (0.02367) (0.01851)
AS(IV ) 0.4119 0.4246 0.3858 − −
(0.0138) (0.01615) (0.02735)
Table 13. Values of I F L for AS conditioned on GDP. ** indicates rejection of null

ment. Tables 12 and 13 contain the results when the principal variable is AS
and we condition on GDP.
Results are in Tables 12 and 13 (values of the unconditioned indices in
the first column are from Table 5). First of all note that many values are
not available for lack of data. This was predictable as it is likely to have
very few observations for, say, AS(I) and GDP(I). This can be a first hint
on the irrelevance of conditioning on GDP in presence of AS. Moreover, the
number of statistically significant differences between the unconditioned and
conditioned values is particularly low, if compared with the previous cases.
Summing up: we have attempted to identify the relative role of our vari-
ables in the explanation of GRV in the approach based on transition matrices.
At this stage, we have found a partial confirmation of the hypotheses formu-
lated from the model in Section 2, on the relevance of the dimension of the
economy and structural change and on the irrelevance of per capita GDP in
the explanation of growth volatility. In addition, in the analysis of conditioned
GRV indices, clearer results appear more often at extreme TGDP classes, in-
dicating that in the transition from low to high TGDP levels, the relations
are more blurred (as resulted also from the preliminary graphical analysis).
6 Conclusions
This paper investigates the relation between growth volatility and the level of
development, structural change and the size of the economy. Growth volatility
is measured by a set of indices inspired by the literature on social mobility.

Growth volatility appears to be negatively related to total GDP, proxy for
the dimension of the economy. In particular it seems appropriate to consider
as an additional control for the dimension of the economy the integration
in the world markets. Moreover, growth volatility appears to be negatively
related to the share of agriculture on GDP, proxy for structural change. Fi-
nally, per capita GDP, proxing for the level of development, does not seem to
add relevant information when the other variables are considered. A direction
for further research may be an assessment of the explanatory power of other
factors related to development and to growth volatility, like the growth of
a financial sector, in relation to structural change. In [9] we provide a first
attempt.
References
1. Acemoglu, D. and F. Zilibotti (1997), “Was Prometheus Unbound by Chance?
Risk, Diversification, and Growth”, Journal of Political Economy, 105, pp.
709-751.
2. Acemoglu, D., S. Johnson, J. Robinson and Y. Thaicharoen (2003), “Insti-
tutional Causes, Macroeconomic Symptoms: Volatility, Crises and Growth”,
Journal of Monetary Economics, 50, pp. 49-123.
3. Bartholomew, D.J. (1982), Stochastic Models for Social Processes, 3rd edition,
John Wiley & Son, New York.
4. Bowman, A.W and A. Azzalini (1997), Applied Smoothing Techniques for Data
Analysis, Oxford, Clarendon Press.
5. Bowman, A.W. and A. Azzalini. Ported to R by B. D. Ripley, (2004). sm: kernel
smoothing methods: Bowman and Azzalini (1997). R package version 2.0-13.
6. Canning, D., L. Amaral, Y. Lee, M. Meyer and H. Stanley (1998) “Scaling the
Volatility of GDP Growth Rate”, Economics Letters, 60, 335-341.
7. Easterly, W., R. Islam and J. E. Stiglitz (2000), “Explaining Growth Volatility”,
mimeo.
8. Fiaschi, D. and A.M. Lavezzi (2003). “On the Determinants of Growth Volatil-
ity: a Nonparametric Approach”. Discussion Paper del Dipartimento di Scienze
Economiche. Nr. 25. University of Pisa.
9. Fiaschi, D. and A.M. Lavezzi (2005). “On the Determinants of Growth Volatil-
ity: a Markov Chain Approach”, mimeo: University of Pisa.
10. Grossman, H. and M. Kim (1996), “Predation and Accumulation”, Journal of
Economic Growth, 1(3), 333-350.
11. Horvarth, M. (1998), “Cyclicality and Sectoral Linkages: Aggregate Fluctua-
tions from Independent Sectoral Shocks”, Review of Economic Dynamics, 1,
781-808.
12. Maddison A. (2001), The World Economy: a Millenium Prospective. Paris:
OECD.
13. Pasinetti, L. (1981), Structural Change and Economic Growth : a Theoretical
Essay on the Dynamics of the Wealth of Nations, Cambridge University Press,
Cambridge.
14. Pritchett, L. (2000), “Understanding Patters of Economic Growth: Searching

for Hills among Plateaus, Mountains, and Plains”, The World Bank Economic
Review, 14, 221-250.
15. Quah D. T. (1993), “Empirical Cross-section Dynamics for Economic Growth”,
European Economic Review, 37, 426-434.
16. Quah D. T. (1996), “Regional Convergence Clusters Across Europe”, European
Economic Review, 40, 951-958.
17. R Development Core Team (2003), R: A language and environment for statisti-
cal computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN
3-900051-00-3, URL http://www.R-project.org.
18. Ramey, G. and Ramey, V. A. (1995), “Cross-country Evidence on the Link
Between Volatility and Growth”, American Economic Review, 85, 1138-1151.
19. Scheinkman J. and M. Woodford (1994), “Self organized criticality and eco-
nomic fluctuations”. American Economic Review, 84, 417-421.
20. Shorrocks, A.F. (1978), “The Measurement of Mobility”, Econometrica, 46,
1013-1024.
Appendix
A Country List
AFRICA
AFRICA 1 Algeria
1 Algeria 2 Angola
2 Angola 3 Benin3 Benin 4 Botswana4 Botswana
5 Cameroon
5 Cameroon 6 Cape
6 Cape VerdeVerde 7 Cent. Afr. Rep.
7 Cent. Afr. Rep. 8 Chad 8 Chad 9 Comoros
9 Comoros
10 Congo
10 Congo 11 Côte d’ Ivoire 12 Djibouti
11 Côte d’ Ivoire 12 Djibouti 13
13 Egypt Egypt 14 Gabon14 Gabon
15 Gambia
15 Gambia 1616 Ghana
Ghana 17 Kenya
17 Kenya 18 Liberia19 Madagascar
18 Liberia 19 Madagascar
20 Mali
20 Mali 2121 Mauritania 22 Mauritius
Mauritania 22 Mauritius 23 Morocco23 Morocco 24 Mozambique
24 Mozambique
25 Namibia
25 Namibia 2626 Niger
Niger 27 Nigeria
27 Nigeria 28 Rwanda
28 Rwanda 29 Senegal
29 Senegal
30 Seychelles
30 Seychelles 31 Sierra Leone 32
31 Sierra Leone 32 Somalia Somalia 33 South Africa 34 Sudan34 Sudan
33 South Africa
35 Swaziland
35 Swaziland 36 Tanzania
36 Tanzania 37
37 Togo Togo 38 Tunisia39 Uganda
38 Tunisia 39 Uganda
40 Zambia
40 Zambia 41 Zimbabwe
41 Zimbabwe LATIN AMERICA
LATIN AMERICA 42 Argentina 42 Argentina
43 Brazil43 Brazil
44 Chile
44 Chile 4545 Colombia 46 Mexico
Colombia 46 Mexico 47 Peru47 Peru 48 Uruguay48 Uruguay
49 Venezuela
49 Venezuela 5050 Bolivia
Bolivia 51 Costa
51 Costa Rica Rica 52 Cuba 52 Cuba 53 Dominican
53 Dominican
Rep. Rep.
54 Ecuador
54 Ecuador 5555
El El Salvador56 Guatemala
Salvador 56 Guatemala 57 Haiti 57 Haiti 58 Honduras
58 Honduras
59 Jamaica
59 Jamaica 6060 Nicaragua 61 Panama
Nicaragua 61 Panama 62 Paraguay 62 Paraguay 63Rico
63 Puerto Puerto Rico
64 Trin. Tobago OFF WESTERN
64 Trin. Tobago OFF WESTERN 65 Australia65 Australia 66 New Zealand 67 Canada Canada
66 New Zealand 67
68 United States WEST
68 United States WEST ASIA ASIA 69 Bahrain
69 Bahrain 70 Iran70 Iran 71 Iraq 71 Iraq
72 Israel
72 Israel 7373 Jordan
Jordan 74 74 Kuwait
Kuwait 75 Lebanon
75 Lebanon 76 Oman76 Oman
77 Qatar
77 Qatar 7878 Saudi
Saudi Arabia
Arabia 79 Syria
79 Syria 80 Turkey81 UAE 81 UAE
80 Turkey
82 Yemen
82 Yemen 8383 W.Bank
W.Bank GazaGaza
EASTEAST
ASIAASIA 84 China 84 China 85 India85 India
86 Indonesia
86 Indonesia 87 Japan
87 Japan 88 Philippines
88 Philippines 89 Korea
89 South South 90
Korea 90 Thailand
Thailand
91 Bangladesh 92 Hong Kong
91 Bangladesh 92 Hong Kong 93 Malaysia 93 Malaysia 94
94 Nepal Nepal 95 Pakistan
95 Pakistan
96 Singapore
96 Singapore 97 Sri Lanka
97 Sri Lanka 98 Afghanistan
98 Afghanistan 99 Cambodia 100 Laos100 Laos
99 Cambodia
101 Mongolia
101 Mongolia 102 102 North
North Korea
Korea 103 103 Vietnam EUROPE
Vietnam EUROPE104 Austria 104 Austria
105 Belgium
105 Belgium 106 106 Denmark 107 107
Denmark Finland 108 France
Finland 108 France 109 Germany
109 Germany
110 Italy
110 Italy 111
111 Netherlands
Netherlands 112 112 Norway 113 Sweden
Norway 113 Sweden 114 Switzerland
114 Switzerland
115 UK
115 UK 116 Ireland
116 Ireland 117 Greece
117 Greece 118 Portugal
118 Portugal 119 Spain
119 Spain
Table
Table 14. 14. Country
Country list list
Financial Fragility and Scaling Distributions in
the Laboratory
Giovanna Devetag1 , Edoardo Gaffeo2 , Mauro Gallegati3 , and Gianfranco

Giulioni4
1
DMCS and CEEL, University of Trento, Italy
2
DE and CEEL, University of Trento, Italy, and SIEC, Politechnic University of
Marche, Ancona, Italy
3
DEA and SIEC, Politechnic University of Marche, Ancona, Italy
4
DMQTE University of Chieti-Pescara, Italy and SIEC, Politechnic University of
Marche, Ancona, Italy
Summary. We present results from human and computer-based experiments aimed

at exploring the role of rationality and financial markets in explaining the emergence
of some well-known stylized facts regarding industrial and aggregate dynamics. We
find that the information conveyed by financial markets helps agents adopt more
rational decision processes. Rationality, in turn, is necessary to observe smooth ag-
gregate behaviors.
1 Introduction
In a recent stream of papers, D. Delli Gatti, M. Gallegati and their coau-

thors (Delli Gatti et al., 2004a, 2004b; Gallegati et al., 2003; hereafter DGG)
have emphasized the relationship between industrial dynamics and financial
fragility as a key issue in understanding macroeconomic fluctuations. The
main insights surfaced from this line of research are that: i) distributions
matter; and ii ) individuals’ level stochasticity and interactions matter. Briefly
stated, the message to be conveyed is that, at odds with the mainstream ap-
proach to macroeconomic modeling, interactions among heterogeneous agents
buffeted with idiosyncratic shocks have a bearing in explaining aggregate fluc-
tuations.
Such a view has emerged after three notable empirical regularities sprung
up in the econophysics literature have started to be employed as stylized facts
against which the predictive ability of business cycle models can be tested.
First, the distribution of firms’ size is right skewed and its upper tail follows
a power law (Axtell, 2001; Gaffeo et al., 2003). Second, the growth rates of
either firms’ output and countries’ GDP follow a Laplace distribution (Stan-
ley et al., 1996; Amaral et al., 1998). Third, the standard deviation of GDP
62 Giovanna Devetag et al.
growth rates and of firms’ growth rates scale down with size with an identi-
cal exponent (Lee et al., 1998). For standard economic theory, which remains
largely rooted on the optimizing behavior of identical agents acting in isola-
tion, these findings are bewildering. In fact, scale invariance and universality
calls for critical behavior in self-organizing economic organizations (Bak, 1996;
Norrelykke and Bak, 2002), or random multiplicative growth processes in hi-
erarchically organized systems with variable number of components (Amaral
et al. 1998; Blank and Solomon, 2000). In other terms, scaling and univer-
sal features are likely to signal the existence of complex interactions among
economic units characterized by heterogeneous sizes, financial positions and
possibly behavioral rules.
The ‘financial fragility’ framework developed in DGG, where heteroge-
neous firms and a banking sector interact giving rise to complex dynamics,
has a case in point. Agent-based simulations have shown this model is capable
to replicate an impressive number of stylized facts regarding both industrial
and aggregate dynamics (see also Delli Gatti et al., 2004). These results are
particularly interesting, given that they are derived in a model which explicitly
attempts to address a well known criticism by e.g. Brock (1999) and Durlauf
(2003), who warn against the practice of providing interpretative frameworks
for scaling laws in economics and finance by simply re-labeling stochastic pro-
cesses known to generate them, without paying any attention to modeling
the interplay of preferences, technology, constraints and beliefs guiding hu-
man economic behavior. In fact, DGG show that scaling phenomena can arise
endogenously in a model where heterogeneous rational agents choose desired
investment in order to maximize a profit function. Individual decision-making
processes are buffeted with idiosyncratic i.i.d.real (i.e., referred to relative
prices) shocks, and the interplay of demand and supply in the credit market
gives rise to mean field interactions among agents.
An issue that deserves further research, however, consists in exploring to
what degree the assumptions on the environment and individual behavior at
the heart of the DGG model can be blamed for its results. In this paper we
present results from a controlled experiment that attempts to provide an an-
swer to this question. In particular, our aim is that of investigating whether
simulations’ findings are confirmed once: i) perfectly-rational agents are sub-
stituted by less-than-perfectly-rational decision makers; and ii ) markets, in
this case the credit market, does not work properly.
Summarizing our results, we find that some celebrated stylized facts of
industrial dynamics, such as the distribution of firms’ size scaling down as
a power law, are consistent with very low degree of rationality on the parts
of economic agents. Interactions on the financial market, in turn, help our
experimental subjects to behave in a risk-averse manner, that is a rational
decision-making process as soon as the possibility of going bankrupt is taken
into account. This last result suggest a possible new role for financial markets,
that is fostering rationality.
Financial Fragility and Scaling Distributions in the Laboratory 63
The remaining of the paper is organized as follows. Section 2 briefly

presents the theoretical framework we brought to the laboratory. Section 3
describes the experimental setting, while Section 4 provides details of our
findings. Finally, Section 5 summarizes the paper and put forwards its main
conclusions.
2 Theoretical Framework
In its basic version, the DGG model consists of only two markets: goods and
credit. In order to simplify the analysis as much as possible, we assume that
output is supply driven, i.e. firms can sell all the output they produce. In
each period, there is a “large” (but, due to the entry-exit process, far from
constant) number of firms N , indexed by i = 1,.., N , which differ according
to their financial conditions. The financial robustness of a firm is proxied by
the so-called equity ratio, i.e. the ratio of its equity base or net worth (Ait )
to the capital stock (Kit ), ait = Ait /Kit . Since firms sell their output at an
uncertain price they may go bankrupt. Bankruptcy occurs if net worth at time
t becomes negative, i.e. if the individual price falls below a critical threshold.
The probability of bankruptcy turns out to be an increasing function of the
interest rate and the capital stock, and a decreasing function of the equity
base inherited from the past.
As shown analytically in Greenwald and Stiglitz (1993), in such a setting
the problem of the firm consists in maximizing its expected profits net of
bankruptcy costs. The optimal capital stock turns out to be a decreasing
function of the interest rate and an increasing function of the equity ratio.
Output follows the evolution over time of the capital stock, which is in turn
determined by investment.
Due to informational imperfections on the equity market, firms can raise
funds only on the credit market. For the sake of simplicity we assume that
many heterogeneous firms interact with the banking sector (as an aggregate).
Investment is financed by means of retained earnings and by new debt. There-
fore the demand for credit of each firm is a function of its financial robustness
and of the interest rate. In order to determine the supply of credit and its
allocation to each firm, we assume that there is a risk coefficient α - i.e. a
threshold ratio of bank’s equity to credit extended - that the banks try to
target, either because of a risk management strategy or as a consequence of
prudential regulation on the part of the monetary authorities. Therefore the
aggregate supply of credit turns out to be approximately a multiple (1/α) of
banking sector equity.
Credit has to be allotted to the heterogeneous firms. We assume that each
firm obtains a portion of total credit equal to its relative size, i.e. the ratio of
the individual capital stock to the aggregate capital stock: highly capitalized
(i.e. collateralized, in economics jargon) borrowers have a higher credit supply,
and vice-versa. The rate of interest charged to each firm is determined in
equilibrium when the demand for credit is equal to the credit extended by
the bank to that firm. The equilibrium interest rate generally decreases with
the firm’s and the bank’s equity. Investment and output therefore generally
increase with the bank’s and the firm’s net worth.
The bank’s equity base, which is the only determinant of total credit sup-
ply, increases with the bank’s profits, which are affected, among other things,
by firms’ bankruptcy. When a firm goes bankrupt, i.e. its net worth becomes
negative because of a huge loss, the bank has a “bad debt”. This is the root of
a potential domino effect: the bankruptcy of a firm today is the source of the
potential bankruptcies of other firms tomorrow via its impact on the bank’s
equity.
The model has been simulated by means of agent-based techniques, and
the initial number of firms has been set equal to 10000. Among the many
stylized facts the model can replicate (Delli Gatti et al., 2004c), it deserves
to be mentioned that the distribution of firms’ size (in terms of capital stock)
is characterized by persistent heterogeneity and is well fitted by a power law
(Figure 1), while the distribution of firms’ growth rates is approximated by a
tent-shaped curve, i.e. a Laplace distribution. To test whether the conditional
distribution of growth rates has a functional form independent of size, data
have been further divided into five bins according to firms’ dimension. In
fact, the five conditional probability distributions collapse onto a single curve
(Figure 2).
Fig. 1. Log-log plot of firms’ size (by capital stock) as derived from simulations of
the financial fragility model. The solid line is the OLS regression line through the
data, with slope 1.11 ± 0.11.
Fig. 2. PDF of simulated firms’ growth rates. Rates are grouped in five bins (indi-
cated with n) according to the rank of firms’ size. Data collapse onto a single curve,
signaling that conditional probabilities are size-independent.
3 Human Subjects and Computational Experiments

Agent-based simulation techniques are powerful methods to derive results, es-
pecially when dealing with complex frameworks as the one described above.
However, in many cases simulation findings are highly dependent on the width
of the parameters’ space chosen on the one hand, and the rationale for what
is truly going on (i.e., who causes what) during simulations remains some-
how hidden on the other one. In other terms, simulations must be confronted
with the issues of robustness and causality. As a matter of example, in the
DGG model the issue of whether scaling in the firms’ size distribution as re-
ported in Figure 1 is merely a result of stochasticity or it is influenced by
rational decision processes and institutions (i.e., financial markets) remains
unexplored.
A possible method to address this question consists in disentangling the
complex interaction of the model’s ingredients by analyzing the behavior of
real people in a simplified but fully comparable environment, that is by re-
curring to experiments.
We run a first treatment in which firms (players) are allowed to operate in-
dependently, the interest rate is fixed and known to subjects, and firms growth
(measured by the equity base dynamics) is therefore only a function of indi-
vidual independent investment decisions. Fully rational players should take
into account the possibility to fail and thus behave in a risk-averse manner,
which in this case means that investment is correlated with available equity.
A first interesting question, therefore, is to explore whether subjects behave

rationally. If the answer is not, it seems interesting to analyze the shape of re-
sulting firms’ size distribution, to assess whether scaling is a mere by-product
of stochasticity.
In a second treatment we add the working of a lending market. Firms
interact by means of a central bank who lends credit on the basis of market
conditions. Each firm faces a different interest rate in every round depending
on its financial soundness (i.e., its equity base) and on the difference between
aggregate supply and demand for credit. The event represented by a firm go-
ing bankrupt (i.e., its equity base becoming negative) impacts on the whole
economy by affecting the banks’ equity base, with consequences on the level
of credit supply and on the interest rate. Subjects have complete information
about the conditions affecting their profitability and earnings; strategic uncer-
tainty however stems from the contemporaneous investment decisions made
by other firms.
3.1 Experimental Design
Subjects were asked to play a simple investment, repeated game which pre-
serves all the main characteristics of the DGG model. In what follows we
describe the two experimental treatments in detail.
Treatment 1: no interaction
In treatment 1 firms are initially endowed with the same amount of equity
(A) and debt (L), which together constitute their initial capital endowment
(K = A + L). Each firm must decide in every period how much to invest in
its production process, i.e., by how much to increase its capital stock K. For
simplicity, we assume that investment can only occur by resorting to debt.
Firms make their investment decisions at the beginning of a round, while
the final output of that round is determined as follows:
Yit = λKit , with Kit = Ai(t−1) + Li(t−1) + Iit (1)
where Iit = ∆Lit is the amount of investment in period t, and λ is the pro-
ductivity parameter, which is the same for all.
The relative price p at which each firm may sell its output is a uniformly
distributed random variable in the [0,2] range. A firm’s profits in real terms
at time t are determined as follows:
πit = λKit pit − rKit (2)
where r is the interest rate, fixed at the same level for all firms. The profit
at time t is added to (subtracted from) the firm’s equity A, so that Ai(t+1) =
Ait + πit .
Treatment 2: interaction
In treatment 2, a central bank (played by the computer program) lends to

firms on the basis of market conditions. The bank decides in every round the
total supply of credit according to the following equation:
E(t−1) Nt −nb(t−1)
LSt = α yt with yt = Nt . (3)
The parameter α is the risk coefficient, and nb(t−1) is the number of firms
that have left the market in the previous round due to bankruptcy. The term
yt is introduced to adjust for the reduction in the number of active firms over-
time, in order to avoid credit oversupply. Each firm faces a different interest
rate in every round depending on its financial soundness (i.e., its equity base)
and on the difference between aggregate supply and demand for credit:
⎛ ⎞
Ki(t−1)
tot
Ai(t−1) ⎜ K̄(t−1) ⎟
rit = r̄ − φ1 ln − φ2 ln ⎝ N Ki(t−1) ⎠+
Ā(t−1) (t−1)
tot
i=1 K̄(t−1)
(4)
N
πi(t−1) t
−φ3 ln + φ4 It − Lt
π̄(t−1) i=1
The log-linear functional form (4) approximates the equilibrium equation

for the interest rate in the DGG model, which is derived from first order condi-
tions. Such an optimization process does not clearly occur in our experimental
setup.
Using an indicator function Bit which equals −Ait if Ait < 0 and 0 if
Ait ≥ 0 we can finally define the bank’s equity base at time t as follows:
N(t−1) N(t−1)

Et = E(t−1) + ri(t−1) Li(t−1) − Bi(t−1) − r̄E(t−1) (5)
i=1 i=1
The bank’s equity base is increased every round by an amount equal to its
net profits (i.e., the total amount of interest gained on credit supplied minus
financial costs) and it is decreased by an amount equal to “bad debt”, i.e.,
the equity base of firms that went bankrupt in the previous round.
3.2 Implementation
Two experimental sessions were conducted for each treatment, with N = 20

subjects participating in each session. Our experimental agents were under-
graduate students in the School of Economics of the University of Trento. Sub-
jects were recruited through ads posted at the Computable and Experimental
Economics Laboratory (CEEL) building and they had never participated in
experiments of this type before. Instructions were distributed at the beginning

of the experiment and read aloud by the experimenters. The experiment itself
was computerized. It was programmed to last for a minimum of 30 periods,
after which the termination was decided on the basis of a random device.
The timing structure was communicated to experimental subjects. Finally,
the interaction was anonymous.
Subjects made their investment decisions at the beginning of every round
and received feedback at the end concerning: 1) their relative price for that
round; 2) the profit earned and, consequently, the new updated values of
their equity base; 3) their debt and their capital stock; 4) in the case of the
interaction treatment, their interest rate. In treatment 2, subjects were also
provided feedback about the average payoff obtained, and the average value
of the equity base in the economy. The variable I (i.e., the amount to invest)
could vary in a range that changed over time in order to calibrate subjects’
possible investment decisions on the level of their capital stock, which tended
to increase over time. The range was set up as shown in table 1.
Table 1.
periods 1-5 6-15 16-30 31-end

range of I 0-100 100-200 200-400 400-600
The initial show up fee was set at 5 euros, and a subject’s final earnings
were determined by the final value of her equity base relative to the maximum
theoretical value that could be achieved.5 In addition, a 3 euros penalty had
to be paid in case of bankruptcy (subtracted from the show up fee). The initial
parameter values were set as follows: A=20, L=40, K=60. The values of the
interest rate parameters were φ1 = φ2 = φ3 = φ4 = 0.001, r̄ = 0.1. The
productivity parameter λ was set equal to 0,20. The parameter α was equal
to 1, and finally, the bank’s initial equity base E0 was set at a value of 10.
4 Results
The main goals of our experiment consisted in testing how fully rational in-
vestment decisions and feedback from financial markets affects the emergence
of some well known regularity in industrial and aggregate dynamics. To ac-
complish this task, we first run an experiment with human subjects, and
subsequently we augmented the number of observations by allowing comput-
erized agents to play strategies identical to those played by real agents. Let
us start our discussion from experimental results.
5
In treatment 2, the maximum value of A was determined under the assumption
of a constant interest rate of 10%.
4.1 Experimental Results
The choice problem that subjects faced in our experimental setting is un-
doubtedly complex, hence we did not expect to observe any kind of optimizing
behavior, but rather the use of largely sub-optimal and “backward looking”
rules of thumb. In fact, several previous experiments have shown that subjects
often tend to use simple heuristics that condition the choice in one period to
the result obtained in the previous one. As a matter of example, Nagel and
Tang (1998) show that subjects playing a centipede game tend to increase the
number chosen in the previous round if that round had been successful, and
vice-versa if the previous round resulted in a loss. Zwick and Rapoport (2003),
in turn, test a simple reinforcement rule in a market entry game to find out
that their experimental subjects tended to choose entry more often in a round
if their entry had been successful the round before. Finally, Bottazzi and De-
vetag (2003) report evidence for an analogous one-lag time-dependence in a
repeated minority game, while dependence of subjects’ choices on more dis-
tant rounds seems weak. Accordingly, we investigated whether our subjects’
choices in the experiment can be explained by similar patterns.
A natural variable to check for is the payoff obtained in the previous round.
Table 2 reports the absolute frequencies (averaged over all groups and treat-
ments) with which a subject’s amount of investment was lower than, higher
than, and equal to, the investment made in the previous round conditional on
the previous round payoff having been positive or negative (a payoff exactly
equal to zero was never observed).
Table 2
Investment choices πi (t − 1) > 0 πi (t − 1) < 0

I(t) > I(t − 1) 0.41 0.30
I(t) = I(t − 1) 0.30 0.28
I(t) < I(t − 1) 0.29 0.42
The data show a modest increase in the probability of raising one’s in-
vestment relative to the previous round if this had been profitable. Overall,
the probability that the investment choice in a round will be at least equal to
that of the previous round is 71% after a profitable round, while it decreases
to 58% after incurring a loss.
A pattern conditioning one’s choice to the previous round payoff is very
myopic as it does not take into account any measure of risk associated with
investment decisions. The risk that subjects should take into account regards
the perceived probability of going bankrupt, which occurs when their equity
base becomes negative. In order to check for some dependence of choices on
parameters signaling “financial solidity”, we investigate whether investment
in a round depended on a player’s equity ratio (i.e., a = A/K) in the previous
period. Table 3 reports the results of the frequency count, taking as a cutoff
level for the equity ratio a value of 0,5:
Table 3
Investment choices πi (t − 1) > 0 πi (t − 1) > 0 πi (t − 1) < 0 πi (t − 1) < 0

ai ≥ 0.5 ai < 0.5 ai ≥ 0.5 ai < 0.5
I(t) > I(t − 1) 0.28 0.41 0.25 0.28
I(t) = I(t − 1) 0.47 0.34 0.37 0.36
I(t) < I(t − 1) 0.24 0.25 0.38 0.36
There seems to be no notable effect of the equity ratio on investment

choices. Frequencies are close to those implied by a uniform probability distri-
bution; it has to be noted that the information about the payoff seems more
salient than the equity ratio. In particular, after a positive payoff round sub-
jects increase or at least not decrease their investment 75% of the times, while
this percentage lowers to 62% after a negative payoff round. The same analysis
has been performed taking into consideration the variation of the equity ratio
from one period to the next, with analogous results. No differences between
groups and treatments were found at this level of analysis.
Next, we tested whether choices in the interaction treatment can be ex-
plained by the interest rate that was applied by the bank to each firm. Such
interest rate could be lower or higher than a benchmark value of 10% de-
pending on whether the firm has been better than the “average” firm in the
economy in terms of profit, equity ratio, and capital stock (see eq. (4)). Sub-
jects were informed of how the interest rate was calculated, and hence they
might have taken it into account as a proxy of their relative standing in the
economy. Therefore, we conditioned choices of investment in the interaction
treatment to both profit and interest rate obtained in the previous period.
Instead of testing for variations in investment from one round to the next,
we considered whether the amount invested was above or below the 50% of
the total amount that subjects could invest. We only considered choices in
the first 15 periods, as in the last 15 periods probability of bankruptcy was
significantly lower. We used a cutoff level of 0.1 for the interest rate. Table 4
reports the results:
Table 4
Investment choices πi (t − 1) > 0 πi (t − 1) > 0 πi (t − 1) < 0 πi (t − 1) < 0

ri ≥ 0.1 ri < 0.1 ri ≥ 0.1 ri < 0.1
I(t) > 0.5 0.28 0.54 0.26 0.38
I(t) = 0.5 0.11 0.17 0.04 0.06
I(t) < 0.5 0.60 0.29 0.70 0.55
The results show a significant effect of the interest rate. In particular, after
a positive payoff and an interest rate below 10%, subjects invest more than
half of their amount 54% of the times, while in the worst case of a negative
payoff and an interest rate above 10%, 70% of the times subjects invest less
than 50% of the total.
From the analysis of individual choice behavior a pair of conclusions can
be drawn. First, in spite of having the opportunity to calculate profit max-
imizing strategies, subjects seem to limit their decisional behavior to rather
naı̈ve rules and to base their choices on very few pieces of information. When
not given explicit information about their risk of bankruptcy, for example,
subjects do not seem to take this variable into account when making their
choices, acting de facto as risk neutral investors. Second, as the introduc-
tion of financial markets augments the information space (e.g., the interest
rate in the interaction treatment), subjects tend to use more rational deci-
sional processes, for example investing more when their financial position is
healthier relative to the average in the economy. These results suggest that
financial markets, besides their well-known efficiency-improving role through
the coordination and saving and investment, could also play a role in helping
economic agents to follow more rational decision processes by disseminating
relevant pieces of information about their own. It seems to us that such an
issue deserves to be the subjects of further research.
Finally, note that the rate of bankruptcy was equal to 27% overall, with
no significant difference between treatments. All bankruptcy events occurred
within the first ten periods of the experiment; hence, if a firm was able to
survive the first ten periods, the probability of subsequently exiting the market
became extremely low. This finding is consistent with empirical evidence on
longitudinal firm-level data, according to which younger firms have a higher
probability of exiting the market (Caves, 1998).
4.2 Simulation results
In order to test the distributional and aggregate features associated with the
behavioral rules adopted by our experimental subjects, we run several agent-
based simulations by allowing a large number of computerized firms to play
the investment game described above over very long horizons. In particular,
firms were instructed to use the rules reported in Table 2 and Table 4 for sim-
ulations based on the no-interaction and the interaction treatments, respec-
tively. Operationally, the growth rate of capital, comprised in the (-0.01,0.02)
interval, has been divided in quantiles of amplitude equal to 0.005. Firms de-
termined their investment rate by choosing a quantile. The effective growth
rate of capital was then a random draw from a uniform distribution over the
chosen quantile. The number of agents were set at 10000, while simulations
spanned over 4200 periods. Firms forced to exit due to bankruptcy were re-
placed by another firm with common starting conditions for the debt/equity
ratio.
The two sets of simulations gave qualitatively different patterns of aggre-

gate behavior. The simple backward-looking decision rule experimental sub-
jects followed when the interest rate was hold constant implies the emergence
of a power law for the firms size distribution, but gives rise to an aggregate
dynamics, expressed in terms of aggregate output, characterized by recurrent
crashes (Figures 3 and 4). In other words, the lack of risk-aversion at an indi-
vidual level implies the winner-takes-all features typical of a stochastic process
composed of random multiplicative shocks combined with an exit threshold
(Blank and Solomon, 2000) on the one hand, but returns an unacceptable high
degree of aggregate instability on the other one. As far as the random demand
allow firms to make profits, they continue to borrow for investment without
any regards to their financial soundness. This implies that their equity ratio
decreases, with the consequence that the probability to draw a negative de-
mand shock forcing them to bankrupt increases as their size grows. Eventually,
the default of very big firms causes huge drops in aggregate output.
Fig. 3. log-log plot of firms’ size distribution in the no interaction treatment.
The introduction of interactions through the credit market translates im-

mediately into interest rates varying with the degree of individual financial
fragility. This piece of information was used by experimental subjects to up-
date their investment decisions, as an increasing interest rate could be imme-
diately recognized as a deterioration of their degree of solvability. As suggested
before, financial markets can be seen as an institution helping agents to cor-
rectly adopt a risk-averse behavior. In fact, simulations for the interaction
scenario shows that the average equity ratio stabilized at 0.20 over the whole
horizon (Figure 5), while the time series of the aggregate output is much
Fig. 4. plot of aggregate output over time in the no interaction treatment.
smoother than before (Figure 6). As regards the firms’ size distribution, now
we observe a breakdown of scaling: while the size of big firms still follows a
power law, the size of small firms seems to obey a log-normal distribution
(Figure 7). This pattern is qualitatively similar to the one observed from sim-
ulations run with the original model (see Figure 1).
Fig. 5. average equity ratio over time in the interaction treatment.

Fig. 6. Plot of aggregate output over time in the interaction treatment.
Fig. 7. log-log plot of firms’ size distribution in the interaction treatment.

5 Conclusions
The framework recently offered in Delli Gatti et al. (2004a; 2004b) and Gal-
legati et al. (2003) has been used as the basis for human and computer-based
experiments. Out aim consisted in exploring the role of rationality and finan-
cial markets in explaining the emergence of some well-known stylized facts
regarding industrial and aggregate dynamics. Two main results have emerged.
First, the information conveyed by financial markets seems to help agents to
adopt rational decision processes. Second, rationality, on its part, is a neces-
sary condition to observe smooth aggregate behaviors.
References
1. Amaral L, Buldyrev S, Havlin S, Salinger M, Stanley E (1998) Power Law
Scaling for a System of Interacting Units with Complex Internal Structure,
Physical Review Letters, 80:1385–1388
2. Axtell R (2001) Zipf Distribution of U.S. Firm Sizes, Science, 293:1818–1820
3. Bak P (1997) How Nature Works, Oxford University Press, Oxford
4. Blank A, Solomon S (2000) Power Laws in Cities Population, Financial Markets
and Internet Sites, Physica A, 287:279–288
5. Bottazzi, Devetag G (2003) Coordination and Self-Organization in Minor-
ity Games: Experimental Evidence, in Kirman A, Gallegati M, Marsili M
(Eds.), The Complex Dynamics of Economic Interaction: Essays in Economics
and Econophysics, Lecture Notes in Economics and Mathematical Systems,
Springer-Verlag, Berlin 283-300
6. Brock W (1999) Scaling in Economics: A Reader’s Guide, Industrial and Cor-
porate Change, 8:409–446
7. Caves R (1998) Industrial Organization and New Findings on the Turnover and
Mobility of Firms, Journal of Economic Literature, 36:1947–1982
8. Delli Gatti D, Di Guilmi C, Gallegati M, Giulioni G (2004a), Financial Fragility,
Industrial Dynamics and Business Fluctuations in an Agent-Based Model,
Macroeconomic Dynamics, forthcoming
9. Delli Gatti D, Di Guilmi C, Gaffeo E, Gallegati M, Giulioni G, Palestrini A
(2004b) A New Approach to Business Fluctuations: Heterogeneous Interacting
Agents, Scaling Laws and Financial Fragility, Journal of Economic Behavior
and Organization, forthcoming.
10. Delli Gatti D, Gaffeo E, Gallegati M, Giulioni G, Kirman A, Palestrini A, Russo
A (2004c) Complex Dynamics and Empirical Evidence, Information Sciences,
forthcoming.
11. Durlauf S (2003) Complexity and Empirical Economics, mimeo
12. Gaffeo E, Gallegati M, Palestrini A (2003) On the Size Distribution of Firms:
Additional Evidence from the G7 Countries, Physica A 117-123.
13. Gallegati M, Giulioni G, Kochji N (2003) Complex Dynamics and Financial
Fragility in an Agent-Based Model, Advances in Complex Systems 6
14. Greenwald and Stiglitz (1993) Financial Market Imperfections and Business
Cycles Quarterly Journal of Economics 108:77–113
15. Lee Y, Amaral L, Canning D, Meyer M, Stanley E (1998) Universal Features

in the Growth Dynamics of Complex Organizations, Physical Review Letters,
81:3275–3278
16. Nagel R, Tang FF (1998) Experimental Results on the Centipede Game in
Normal Form: an investigation on learning, Journal of mathematical Psychology
42:356–384
17. Nørrelykke S, Bak P (2002), Self-Organized Criticality in a Transient System,
Physics Review 68
18. Stanley M, Amaral L, Buldyrev S, Havlin S, Leschorn H, Maas P, Salinger M,
Stanley E (1996), Scaling Behavior in the Growth of Nature, 379:804–806
19. Zwick R, Rapoport A (2002) Tacit Coordination in a Decentralized Market
Entry Game with Fixed Capacity, Experimental Economics 5(3):253–272
Part II
Complex Economic Network

Heterogeneous Economic Networks
Wataru Souma1 , Yoshi Fujiwara2 , and Hideaki Aoyama3

1
ATR Network Informatics Laboratories, Kyoto 619-0288, Japan. souma@atr.jp
2
ATR Network Informatics Laboratories, Kyoto 619-0288, Japan.
yfujiwar@atr.jp
3
Department of Physics, Graduate School of Science, Kyoto University, Yoshida,
Kyoto 606-8501, Japan. aoyama@phys.h.kyoto-u.ac.jp
Summary. The Japanese shareholding network at the end of March 2002 is

studied. To understand the characteristics of this network intuitively, we visu-
alize it as a directed graph and an adjacency matrix. Especially detailed fea-
tures of networks concerned with the automobile industry sector are discussed
by using the visualized networks. The shareholding network is also considered
as an undirected graph, because many quantities characterizing networks are
defined for undirected cases. For this undirected shareholding network, we
show that a degree distribution is well fitted by a power law function with
an exponential tail. The exponent in the power law range is γ = 1.8. We also
show that the spectrum of this network follows asymptotically the power law
distribution with the exponent δ = 2.6. By comparison with γ and δ, we find
a scaling relation δ = 2γ − 1. The reason why this relation holds is attributed
to the local tree-like structure of networks. To clarify this structure, the cor-
relation between degrees and clustering coefficients is considered. We show
that this correlation is negative and fitted by the power law function with the
exponent α = 1.1. This guarantees the local tree-like structure of the network
and suggests the existence of a hierarchical structure. We also show that the
degree correlation is negative and follows the power law function with the
exponent ν = 0.8. This indicates a degree-nonassortative network, in which
hubs are not directly connected with each other. To understand these features
of the network from the viewpoint of a company’s growth, we consider the
correlation between the degree and the company’s total assets and age. It is
clarified that the degree and the company’s total assets correlate strongly, but
the degree and the company’s age have no correlation.
Keywords. Shareholding network, Visualization, Network analysis, Power

law, Company’s growth
80 Wataru Souma, Yoshi Fujiwara, and Hideaki Aoyama
1 Introduction
The economy is regarded as a set of activities of heterogeneous agents in
complex networks. However, many traditional studies in economics are for
the activities of homogeneous agents in simple networks, where we call reg-
ular networks and random networks simple networks. To overcome such an
unrealistic situation many efforts have been made, and a viewpoint of hetero-
geneous agents has emerged. However, simple networks are adapted in many
of the studies of heterogeneous agents. Hence it is important to introduce the
true structure of real world networks. Recently, the study of complex networks
has revealed the true structure of real world networks: WWW, the Internet,
social networks, biological networks, etc. [1][4][5][12]. However the true struc-
ture of the economic network is not well known. Hence the purpose of this
study is to reveal it.
As is commonly known, if we intend to discuss networks, we must define
the nodes and edges. Here, edges represent the relationship between nodes.
In business networks, the candidates for the nodes are individuals, compa-
nies, industry categories, countries, etc. In this study we consider companies
as nodes. Hence, in the next step, we must define the relationship between
companies. To define it, we use three viewpoints: ownership, governance, and
activity. The ownership is characterized by the shareholding of companies, and
the networks constructed by this relationship are considered in this article.
The governance is characterized by the interlocking of directors, and networks
of this type are frequently represented by a bipartite graph that is constructed
with corporate boards and directors. The activity networks are characterized
by many relationships: trade, collaboration, etc.
Although we use tree point of view, these have relations with each other.
For example, if the owners of a company change, then the directors of that
company will change. If the directors of the company change, then the ac-
tivities of the company will change. If the activities of the company change,
then the decisions of the owners and directors will change, and sometimes the
owners and the directors will change.
In this article we consider Japanese shareholding networks at the end of
March 2002 (see Ref. [9] for shareholding networks in MIB, NYSE, and NAS-
DAQ). In this study we use data which is published by TOYO KEIZAI INC.
This data provides lists of shareholders for 2,765 companies that are listed on
the stock market or the over-the-counter market. Almost all of the sharehold-
ers are non-listed financial institutions (commercial banks, trust banks, and
insurance companies) and listed non-financial companies. In this article we
ignore shares held by officers and other individuals. The lengths of the share-
holder lists vary with the companies. The most comprehensive lists contain
information on the top 30 shareholders. Based on this data we construct a
shareholding network.
This paper is organized as follows. In Sec. 2, we consider the visualization
of the shareholding network as a directed graph and an adjacency matrix. The
Heterogeneous Economic Networks 81
Companies
Shareholding Companies
Automobile industry Banks
Insurance companies
Fig. 1. A visualization of the Japanese shareholding network at the end of March

2002 (left) and a corresponding adjacency matrix (right).
visualization of networks is a primitive, but powerful tool to intuitively under-

stand the characteristics of networks. As an example, we especially consider
networks concerned with the automobile industry sector. In Sec. 3, we treat
the shareholding network as an undirected network. This is because many
useful quantities characterizing networks are defined for undirected cases. We
consider the degree distribution, spectrum, degree correlation, and the corre-
lation between the degree and the clustering coefficient. In Sec. 4, we discuss
correlations between the degree and the company’s total assets and age. Sec-
tion 5 is devoted to a summary and discussion.
2 Directed Networks
If we draw edges from shareholders to companies, we can obtain shareholding

networks as directed graphs. The primitive way to study networks is to use a
visualization of them. The visualization of the Japanese shareholding network
at the end of March 2002 is shown in the left panel of Fig. 1. This figure
is drawn by Pajek, which is a program for analyzing large networks [13]. In
this figure, the open circles correspond to companies, and arrows are drawn
based on shareholding relationships. The network is constructed from 3,152
nodes and 23,064 arrows. This figure is beautiful, but it is difficult to obtain
characteristics of this network.
Frequently, networks are represented by adjacency matrices. Here the adja-
d
cency matrix for the directed graph, Mij , is chosen based on the shareholding
relation: If the i-th company is a shareholder of the j-th company, we assume
d d
Mij = 1; otherwise, Mij = 0. The size of this matrix (network) is N = 3, 152.
The adjacency matrix is shown in the right panel of Fig. 1. In this figure, the
Fig. 2. A shareholding network constructed from companies in the automobile

industry sector.
rows correspond to companies and the columns correspond to shareholding

companies. The list of companies and the list of shareholding companies are
d
the same. In this figure, the black dots correspond to Mij = 1, and the others
d
correspond to Mij = 0. The number of black dots is 23,063.
To define this adjacency matrix we arranged the order of companies ac-
cording to the company’s code number that is defined based on industry cate-
gories. The solid lines in this figure indicate industry categories. We make two
observations: (i) Dots are distributed in all industry categories in the ranges
where financial institutions (banks and insurance companies) are the share-
holders; (ii) The density of the black dots is relatively high in each box, except
for the financial sector. This indicates that we frequently find companies and
shareholders in the same industry category for non-financial firms. Hence this
network shows ”assortative mixing” on the industry category, except for the fi-
nancial institutions. The concept of (dis)assortativity is explained in Ref. [11].
2.1 Shareholding Network Constructed from Companies in the

Automobile Industry Sector
Here, we consider a network constructed from companies belonging to the au-

tomobile industry sector. This sector corresponds to the range shown in the
right panel in Fig. 1. The visualization of this network is shown in Fig. 2. In

this figure, we include company names only for four major automobile indus-
try companies: Toyota Motor, Nissan Motor, Honda Motor, and Mitsubishi
Motor.
If the direction of the arrows is ignored, then Toyota, Honda, and Mit-
subishi are connected to each other with two edges through the shortest path.
Nissan and Toyota are connected with three edges through the shortest path,
and this is also applicable to the case of Nissan and Mitsubishi. However, we
need five edges to connect Nissan and Honda through the shortest path. In
addition, this path must run through Toyota or Mitsubishi. We presently have
no idea how to explain such a network structure, but we believe that a time
series analysis of networks is important.
This figure shows that these four major companies have many edges. In
the graph theory, the degree of a node is defined by the number of other nodes
to which it is attached. The distribution of the degree is an important quan-
tity to study networks [3]. It is well known that the degree distribution of a
regular network is represented by the δ-function, because each node has the
same degree in the regular network. It is also well known that the degree dis-
tribution of random networks, which are proposed by Erdös and Rény, follows
the Poisson distribution. A node with a large degree is called a hub. Hence
these four major automobile industry companies are hubs in the network. In
Sec. 3.1, the details of a study about the degree distribution of the network
are explained.
In this figure, we can find that almost no hubs are directly connected. These
hubs are mediated by companies, which have a small degree. Networks with
such a characteristic are called uncorrelated networks or degree-nonassortative
networks [11], and are characterized by a negative correlation between degrees.
This is explained in terms of a degree correlation [14] in Sec. 3.2. Intuitively,
this nature of the shareholding network is different from that of human net-
works. This is because, in human networks, for example in the friendship net-
work, the hubs in the network correspond to persons with many friends, and
are also friends with each other with a high probability. Networks with such
characteristics are called correlated networks or degree-assortative networks,
and are characterized by a positive correlation between degrees.
Suppose that a node has k neighbors; then at most k(k − 1)/2 edges can
exist to connect the neighbors with each other. Hence the possible number of
triangles (minimum loops) containing this node is also k(k − 1)/2. The ratio
between the possible number of triangles and that of the actually existing
triangles defines a clustering coefficient [20]. As we can see in this figure,
the clustering coefficient is small for hubs, while it is large for nodes with a
small degree. Hence it is expected that degrees and clustering coefficients are
negatively correlated. The details are explained in Sec. 3.3.
This figure also shows that the network is constructed from subgraphs: tri-
angles, squares, pentagons, etc. However, not all subgraphs occur with equal
frequency. Hence if networks contain some subgraphs as compared to random-
Fig. 3. A shareholding network constructed from edges drawn from the outside of
the automobile industry sector to the inside of it. The open circles correspond to
non-automobile industry companies, and the filled circles correspond to automobile
industry companies. Arrows are drawn from the open circles to the filled circles
ized networks, these subgraphs characterize the networks. These characteristic

subgraphs are the building blocks of networks, and are called network motifs
[10][17]. Although network motifs are not discussed in this article, the spec-
trum of the network is considered. This is a primitive way to study subgraphs,
and the details are discussed in Sec. 3.4.
2.2 Shareholding Network Constructed from Edges Drawn from

the Outside of the Automobile Industry Sector to the Inside of It
Figure 3 is constructed from edges drawn from the outside of the automo-
bile industry sector to the inside of it. To draw this figure, for simplicity, we
ignored the arrows connecting financial institutions and automobile industry
companies. In this figure the open circles correspond to non-automobile in-
dustry companies, and the filled circles correspond to automobile industry
companies. Arrows are drawn from the open circles to the filled circles. The
network is divided into many subgraphs, but there exists one large lump. This
figure contains only Mitsubishi Motors of the four major automobile compa-
nies. This means that Mitsubishi Motor is governed by companies outside the
automobile industry sector.
Fig. 4. A shareholding network constructed from edges drawn from the inside of
the automobile industry sector to the outside of it. The open circles correspond to
non-automobile industry companies, and the filled circles correspond to automobile
industry companies. Arrows are drawn from the filled circles to the open circles.
2.3 Shareholding Network Constructed from Edges Drawn from

the Inside of the Automobile Industry Sector to the Outside of It
Figure 4 is constructed from edges drawn from the inside of the automobile
industry sector to the outside of it. As in the previous case, the open circles
correspond to non-automobile industry companies, and the filled circles corre-
spond to automobile industry companies. In this case arrows are drawn from
the filled circles to the open circles. The network is divided into many sub-
graphs, but a large lump exists. We can see that major automobile industry
companies are also major shareholders of non-automobile industry companies,
except for Nissan Motor. It is expected that such a structure emerged after the
year 1999 when Nissan and Renault announced their strategic alliance. Com-
paring this figure and Figs. 2 and 3 makes us believe that Toyota, Honda and
Nissan are leaders in the automobile industry sector, and especially Toyota
and Honda are also leaders in the Japanese economy.
0
10 0
10
Cumulative Probability
-1 -1
10 10
-2 -2
10 10
-3 -3
10 10
-4
-4 10
10 0 1 2 3 0 00
5 1000 1500
10 10 10 10
k : Degree k : Degree
Fig. 5. A log-log plot (left) and a semi-log plot (right) of the degree distribution .
In these figures, the solid lines correspond to the fitting by the power law function
p(k) ≈ k−γ with the exponent γ = 1.8, and the dashed lines correspond to the
fitting by the exponential function p(k) ≈ exp{−β(k − k0 )}.
3 Undirected Network
As shown in Sec. 2, the visualization of networks is a powerful method to
understand the characteristics of networks. However this method is not always
useful. Hence many quantities have been proposed to obtain the characteristics
of networks . In the previous section, the shareholding network is represented
by a directed graph, but we consider it as an undirected network in this section.
This is because many quantities characterizing networks are for undirected
u
cases. In this case, the adjacency matrix Mij is chosen: If the i-th company
u u
is a shareholder of the j-th company, we assume Mij = Mji = 1; otherwise,
u
Mij = 0. Hence this matrix is symmetrical.
3.1 Degree Distribution
A degree is the number of edges that attach to a node. In terms N of the ad-
jacency matrix, the degree of node i, ki , is defined by ki ≡ j=1 Mij u
. The
log-log plot of the degree distribution is shown in the left panel of Fig. 5,
and the semi-log plot is shown in the right panel of Fig. 5. In this figure, the
horizontal axis corresponds to the degree and the vertical axis corresponds to
the cumulative probability.
In these figures, the solid lines correspond to the linear fitting with the least
square method by the power law function, and the dashed lines correspond
to that by the exponential function. As we can see, the degree distribution
follows the power law distribution with an exponential tail. The probability
density function (PDF) p(k) in the power law range is given by
p(k) ∝ k −γ ,
<knn> : Average Degree of n.n.
10
3
0
Ci : Clustering Coefficient
10
-1
10
2
10
-2
10
1
10 -3
10
0 1 2 3
10
0
10
1
10
2
10
3 10 10 10 10
Fig. 6. A log-log plot of the degree correlation (left) and that of the correlation
between the degree and the clustering coefficient (right). In the left panel, the solid
line corresponds to the fitting by the power law function knn (k) ≈ k−ν with the
exponent ν = 0.8. In the right panel, the solid line corresponds to the fitting by the
power law function C(k) ≈ k−α with the exponent α = 1.1.
where the exponent γ = 1.8, and that in the exponential range is given by
p(k) ∝ e−β(k−k0 ) .
The exponential range is constructed of financial institutions, and the power

law range is mainly constructed of non-financial firms. It has also been shown
that the degree distributions of different networks in the economy also show
power law distributions [18].
3.2 Degree Correlation
To obtain more detailed characteristics of networks, the degree correlation

has been considered [14]. The nearest neighbors’ average degree of nodes with
degree k, knn , is defined by

knn = k pc (k |k),
k

where pc (k |k) is the conditional probability that a link belonging to a node
with degree k points to a node with degree k .
The log-log plot of degree correlation is shown in the left panel of Fig. 6. In
this figure, the horizontal axis corresponds to the degree and the vertical axis
corresponds to knn , i.e., the nearest neighbors’ average degree of nodes with
degree k. We find that the high degree range has a small value of knn . This
means that hubs are not directly connected with each other in this network.
Networks with this characteristic are called uncorrelated networks, which are
also found in biological networks and WWW. On the other hand, networks
with a positive correlation are called correlated networks, and are found in
social and architectural networks.
In Fig. 6, the solid line is the fitting by the power law function in the tail
part with the least square method:
knn (k) ∝ k −ν ,
where the exponent ν = 0.8.
3.3 Clustering Coefficient
Cliques in networks are quantified by a clustering coefficient [20]. Suppose

that a node i has ki edges; then at most ki (ki − 1)/2 edges can exist between
them. The clustering coefficient of node i, Ci , is the fraction of these allowable
edges that actually exist ei :
2ei
Ci = .
ki (ki − 1)
The clustering coefficient is approximately equal to the probability of finding

triangles in the network. The triangle is the minimum loop. Hence if node
i has a small value of Ci , then the probability of finding loops around this
node is low. This means that the network around this node is locally tree-
like. The correlation between ki and Ci is shown in the right panel of Fig. 6.
This figure shows that clustering coefficients have a small value in the high
degree range. This means that the shareholding network has a local tree-like
structure asymptotically.
The solid line in the right panel of Fig. 6 is the linear fitting by the power
law function with the least square method:
C(k) ∝ k −α ,
where the exponent α = 1.1. Such a scaling property of the distribution of

clustering coefficients is also observed in biological networks, and motivates
the concept of hierarchical networks [15][16].
3.4 Spectrum
Here we discuss the spectrum of the network, i.e., the distribution of eigen-
values of the adjacency matrix. The distribution around the origin is shown
in the left panel of Fig. 7. In this figure the horizontal axis is an eigenvalue
λi and the vertical axis is a frequency. As is well known, if the network is
completely random the distribution is explained by Wigner’s semi-circle law.
However, Fig. 7 is apparently different from the semi-circle distribution. We
make four observations (see also Ref. [6]): (i) A δ peak at λi = 0, indicating the
localized eigenstates that are produced by the dead-end vertices ; (ii) δ peaks
100
3
10
80
Frequency
2
10
60
Rank
40 1
10
20 positive
0
10 negative
-1 0 1 2
-4 -2 0 2 4 10 10 10 10
λ : Eigenvalue |λ| : Eigenvalue
Fig. 7. A distribution of eigenvalues around the origin (left) and a log-log plot of
the distribution of the absolute value of eigenvalues (right). In the right panel, the
solid line corresponds to the fitting by the power law function ρ(λ) ≈ |λ|−δ with the
exponent δ = 2.6.
at λi = ±1, indicating the existence of approximately infinite long chains;

(iii) A flat shape in the range −1 < λi < 1 except for λi = 0, indicating the
existence of long chains constructed from weakly connected nodes, i.e., nodes
with small degrees; and (iv) A fat tail.
The log-log plot of the eigenvalue distribution is shown in the right panel
of Fig. 7 (see also Ref. [19]). In this figure the horizontal axis is the absolute
value of eigenvalues |λi |, and the vertical axis is the cumulative probability.
The plus symbols represent the distribution of the negative eigenvalues, and
the cross symbols represent that of the positive eigenvalues. This figure shows
that the shape of distribution in the positive eigenvalue range and that in
the negative eigenvalue range are almost the same. The linear fitting with the
least square method to the tail part is shown by the solid line in the figure.
This fitting suggests that the PDF of the eigenvalue ρ(λ) is asymptotically
given by
−δ
ρ(λ) ∝ |λ| ,
where the exponent δ = 2.6. If we compare the values of γ and δ, we can find
the scaling law: δ = 2γ − 1.
It has recently been shown under an effective medium approximation that
the PDF of eigenvalue ρ(λ) is asymptotically represented by that of the degree
distribution p(k):
ρ(λ) 2 |λ| p(λ2 ),
if the network has a local tree-like structure [6]. Therefore, if p(k) asymptoti-
cally follows the power law distribution, ρ(λ) also asymptotically follows the
power law distribution, and we can obtain the scaling relation δ = 2γ − 1.
In addition, the local tree-like structure is guaranteed by the right panel of
Fig. 6.
0 0
10 10
-1 -1
10 10
-2 -2
10 10
-3 -3
10 10
-4 -4
10 2 3 4 5 6 7 8 9 10 0 500 1000 1500 2000
10 10 10 10 10 10 10 10
Total Assets (millions of yen) Age (months)
Fig. 8. A log-log plot of the distribution of the company’s total assets (left) and a
semi-log plot of the company’s age (right).
4 Correlation Between Degree and Company’s Total

Assets and Age
It is interesting to construct models that can explain the topology of share-
holding networks. However, in this section, we consider the correlation between
the degree and the company’s total assets and age. In many complex networks,
it is difficult to quantitatively characterize the nature of nodes. However, in
the case of economic networks, especially networks constructed of companies,
we can obtain the nature of nodes quantitatively. We consider this to be a
remarkable characteristic of business networks, and this allows us to under-
stand networks in terms of the company’s growth. This is the reason why we
consider the correlation between the degree and the company’s total assets
and age.
The log-log plot of the distribution of the company’s asset is shown in the
left panel of Fig. 8. In this figure, the horizontal axis is the assets with the
unit of millions of yen, and the vertical axis is the cumulative probability. This
figure shows that the distribution in the intermediate range follows the power
law distribution. In this case, the distribution is for companies listed on the
stock market or the over-the-counter market. The completeness of data is a
problem in extracting the true nature of the total asset distribution. However,
it has been shown that data satisfying this completeness shows more clearly
the power law distribution of the total assets [7].
The semi-log plot of the distribution of the company’s age is shown in the
right panel of Fig. 8. In this figure, the horizontal axis is the age with the unit
of months, and the vertical axis is the cumulative probability. This figure shows
that the distribution follows approximately the exponential distribution. It is
expected that the age of companies has a relation with their lifetime, and it is
also clarified that the lifetime of bankrupted companies follows the exponential
distribution [8].
Total Assets (millions of yen) 10
9
8
10 1500
Age (months)
7
10
6
10 1000
5
10
4 500
10
3
10
2 0
10 0 1 2 3 10
0
10
1
10
2
10
3
10 10 10 10
Fig. 9. A log-log plot of the correlation between the degree and the company’s
total assets (left) and a semi-log plot of the correlation between the degree and the
company’s age (right).
The log-log plot of the correlation between degrees and total assets is
shown in the left panel of Fig. 9. In this figure the horizontal axis is the degree,
and the vertical axis is the company’s total assets with the unit of millions of
yen. This figure shows that the degree and the total assets positively correlate.
The semi-log plot of the correlation between degrees and the company’s
age is shown in the right panel of Fig. 9. In this figure the horizontal axis is
the degree, and the vertical axis is the company’s age with the unit of months.
This figure shows that the degree and the company’s age have no correlation.
These two results suggest that the degree of companies has a relation with
their total assets, but no relation with their age. This result means that the
size of the company is an important factor to consider with regard to growing
economic networks, but the age of the company is not. Old companies are
not necessarily big companies. Hence knowing the dynamics of the company’s
growth is a key concept in considering growing economic networks [2].
5 Summary
In this paper we considered the Japanese shareholding network at the end
of March 2002, and found some of the characteristics of this network. How-
ever, there are many unknown facts about the characteristics of shareholding
networks. For example, these include time series changes of networks, the as-
pect of weighted networks, flows in networks, and the centrality of networks.
Together with these studies, it is also important to study the dynamics of
a company’s growth. It is expected that the dynamics of economic networks
cam be explained in terms of the dynamics of the company’s growth.
Acknowledgements
Wataru Souma and Yoshi Fujiwara are supported in part by the National
Institute of Information and Communications Technology. We are also sup-
ported in part by a Grant-in-Aid for Scientific Research (#15201038) from
the Ministry of Education, Culture, Sports, Science and Technology.
References
1. Albert R, Barabási AL (2002) Statistical mechanics of complex networks. Rev.
Mod. Phys. 74: 47–97
2. Aoyama H, Souma W, Fujiwara Y (2003) Growth and fluctuations of personal
and company’s income. Physica A 324: 352–358
3. Barabási AL, Albert R (1999) Emergence of scaling in random networks. Sci-
ence 286: 509–512
4. Barabási AL (2004) Network Biology: Understanding the cell’s functional or-
ganization. Nature Rev. Gen. 5: 101–114
5. Dorogovtsev SN, Mendes JFF (2003) Evolution of Networks: From Biological
Nets to the Internet and WWW. Oxford University Press, New York
6. Dorogovtsev SN, et al. (2003) Spectra of complex networks. Phys. Rev. E68:
046109
7. Fujiwara Y, et al. (2004) Do Pareto-Zipf and Gibrat laws hold true? An analysis
with European firms. Physica A 335: 197–216
8. Fujiwara Y (2004) Zipf law in firms bankruptcy. Physica A 337: 219–230
9. Garlaschelli D, et al. (2003) The scale-free topology of market investments.
arXiv:cond-mat/0310503
10. Milo R, et al (2002) Network Motifs: Simple Building Blocks of Complex Net-
works. Science 298: 824–827
11. Newman MEJ (2002) Assortative Mixing in Networks. Phys. Rev. Lett. 89:
208701
12. Newman MEJ (2003) The Structure and Function of Complex Networks. SIAM
Rev. 45: 167–256
13. Pajek: Program for Analysis and Visualization of Large Networks. URL:
http://vlado.fmf.uni-lj.si/pub/networks/pajek/
14. Pastor Satorras R, Vázquez A, Vespignani A (2001) Dynamical and Correlation
Properties of the Internet. Phys. Rev. Lett. 87: 258701
15. Revasz E, et al. (2002) Hierarchical Organization of Modularity in Metabolic
Networks. Science 297: 1551–1555
16. Revasz E, Barabási AL (2003) Hierarchical organization in complex networks.
Phys. Rev. E 67: 026112
17. Shen-Orr SS, et al. (2002) Network motifs in the transcriptional regulation
network of Escherichia coli. Nature Genet. 31: 64–68
18. Souma W, Fujiwara Y, Aoyama H (2003) Complex networks and economics.
Physica A 324: 396–401
19. Souma W, Fujiwara Y, Aoyama H (2004) Random matrix approach to share-
holding networks. Physica A, in press
20. Watts DJ, Strogatz SH (1998) Collective dynamics of ’small-world’ networks.
Nature 393: 440–442.
The Emergence of Paradigm Setters Through
Firms’ Interaction and Network Formation
Rainer Andergassen1 , Franco Nardini2 , and Massimo Ricottilli3

1
Rainer Andergassen, Department of Economics, University of Bologna (Italy)
anderga@economia.unibo.it
2
Franco Nardini, Department of Mathematics for the Social Sciences, University
of Bologna (Italy) nardini@dm.unibo.it
3
Massimo Ricottilli, Department of Economics, University of Bologna (Italy) and
Centro Interdipartimentale Galvani ricottilli@economia.unibo.it
Summary. Technological innovation requires the gathering of information through

a process of searching and learning. We distinguish two different but definitely com-
plementary and overlapping ways through which searching and learning occur. The
first exploits the spillover potential that lies in a firm’s network and thanks to which
gathering innovation-useful information is actually possible. The second is the au-
tonomous capacity that a firm possesses in order to carry out in-house innovative
search. We build a model where rationally bounded firms try to increase their in-
novative capability through endogenous networking. The paper characterizes the
emergence of technological paradigm setters in terms of network properties as they
result from searching routines, furthermore the corresponding average efficiency of
the system in terms of innovative capability is assessed.
Key words: Networks, Bounded Rationality, Technological Change, Innova-

tive Capability, Paradigm Setters.
1 Introduction
Technological innovation requires the gathering of information through a pro-
cess of searching and learning. In economies at the cutting edge of their fron-
tier the competitive drive compels leading firms to engage in this process
lest their advantage be lost to competitors and imitators. This quest for in-
formation is largely an adaptive, gradual process in which internal, in-house
resources generating innovation-worthy knowledge are woven together with
those obtained through technological spillovers proceeding from other firms.
Firms, and agents within firms, exhibit bounded rationality and their success

We are very grateful to seminar participants of the WEHIA 04 Kyoto meeting
for their very useful comments.
94 Rainer Andergassen, Franco Nardini and Massimo Ricottilli
is largely due to but also constrained by their technological capabilities that

are, nevertheless, magnified by interaction with other firms. The implication
is that the process of search is local and confined within a neighbourhood
that is cognitively reachable and that is, therefore, the medium of effective
information flows. To this effect, the role of a network in which firms nest is
paramount
Because of bounded rationality and radical uncertainty, the process of re-
ciprocal observation and learning occurring through interaction between firms
and making the spreading of technological spillovers possible takes place in
viable networks. Their importance has been highlighted in recent literature
investigating technological and knowledge diffusion, see for example, Cowan
and Jonard (2004, 2005), Silverberg and Verspagen (2002), Arenas et Alii,
(2001, 2002). Networks, however, are subject to change and growth: connec-
tivity between members, linkages between firms,can be viewed as a process
in time. Seminal work in this field has been done by Albert and Barabasi,
(2002) for an exhaustive review, Watts and Strogatz (1998), for the emer-
gence of small worlds. When seen as an interactive system (Kirman 1997a,
1997b), the economy generates capabilities that are mutually acquired ren-
dering firms technologically interdependent. It must, nevertheless, be recog-
nised that capability-building also depends on in-house efforts supported by
investment in formal and informal research and development. In this paper
we treat technological capabilities as a stock of firm-specific knowledge cu-
mulated through time by every single firm in the economy. We, accordingly,
propose a model in which it is posited that each firm is endowed with a mea-
surable index of innovation-enabling knowledge that is subject to change as
a consequence of an active searching policy. We distinguish two different but
definitely complementary and overlapping ways through which searching and
learning occur. The first exploits the spillover potential that lies in a firm’s net-
work and thanks to which gathering innovation-useful information is actually
possible. The second rests with the autonomous capacity that a firm possesses
in order to carry out in-house innovative search. While these two searching
processes not only coexist but are also reciprocally sustaining, we find it expe-
dient to separate them by integrating a knowledge diffusion mechanism that
propagates technological capabilities with an independent stochastic process
capturing innovation arrivals due to internal R.&D. A network’s evolution de-
pends on how firms assess their performance in terms of innovation-enabling
spillovers. In a bounded rationality framework, firms normally explore a lim-
ited part of the firms’ space and require a protocol to target their information
gathering efforts. The paper addresses this issue by designing a routinised be-
haviour according to which firms periodically reshape the neighbourhood that
they observe to glean information by reassessing other firms’ contributions to
their own capability. The way the specific neighbour-choosing routine is ac-
cordingly organised determines in a significant way firms’ average innovative
capability. This feature is modelled by changing the span of network obser-
vation from a very broad setting, the whole economy, to a very narrow one,
The Emergence of Paradigm Setters 95
namely the most proximate neighbourhood membership. As a result of the

structure of the model presented in the next section, there are two distinct
but to some extent overlapping neighbourhoods which are relevant for firms’
interaction. The first is the neighbourhood whose members are observed by
each firm and from which capability contributions are obtained. We term this
neighbourhood inward. The second is the one made up by a firm’s observers,
i.e. by firms observing and learning from it: it evolves as an active search
for new inward members is carried out. We call this neighbourhood outward.
This process of information interaction leads to the emergence of some firms
that are observed by most of the remaining ones. It is they that provide some
or much of the overall technological capability and that we term paradigm
setters. We also assume that the in-house acquired capability is subject to
structural shifts by means of periodic random shocks.
To keep the model mathematically tractable, we formalise the features
stated above by means of a linear system in which technological capabilities are
made to depend on a matrix of interaction with evolving outward neighbours
as well as on a vector of in-house generated knowledge. The model is then
simulated to determine the emergent properties of neighbourhood formation
and stability together with average capability. We aim to identify (i) under
what conditions the emergence of technological paradigm setters occurs, (ii)
the pattern of neighbourhood formation and (iii) the average relative efficiency
in terms of technological capability of the economy as a whole.
The plan of the paper is as follows: section two illustrates the linear model
that is implemented to run simulations; section three describes the simulation
procedures and the index employed to assess results; section four discusses
results obtained and section five draws conclusions and sets an agenda for
further research.
2 Firms’ Technological Capabilities and Spillover

Potential
We view a firm’s technological capability as the upshot of an evolutionary

process owing to learning, searching and gathering of information ultimately
leading to innovation, much of it being explained by the interaction taking
place within the system. These three processes are largely overlapping since
neither can exist without the other two. In this section, we direct our analysis
to investigate firms that are assumed to be technological leaders and whose
major interest lies with innovation. We, accordingly, postulate that they pos-
sess ’in house’ innovative capabilities resulting from past investment and that
we distinguish, in a somewhat artificial manner but useful for modelling, from
those that are entirely due to spillovers. These capabilities can be viewed and
measured in a way akin to the more general category of a firm’s knowledge
base, cognitive potential or set of skills, know-how and competencies: they can
actually be modelled as either a vector arranging different indicators or more

simply as a scalar compounding the whole. We choose the latter approach.
Let Vi (t) be the scalar that at time t designates firm i’s innovative capa-
bility or, to use a term borrowed from biology, its innovative fitness. Then,
V (t) is the vector V (t) = [Vi (t)] , i = 1, 2....J arraying the fitness of all
firms in the economy. By Ci (t) we further designate the in-house capabil-
ity cumulated until time t. As mentioned above, the latter, while embodying
cumulated knowledge, requires investment to be preserved and eventually im-
proved. Considerable efforts are therefore necessary to remain on the forefront
of technological prowess, efforts which need not always prove successful. It
is, accordingly, assumed that Ci (t) be stochastically subject to change and
Ci (t) ∈ (0, 1). C(t) is the corresponding vector.
A significant part of total technological capability is explained by firms’
interaction with other firms. As mentioned above, this is due to searching
activity and to the ease with which each firm is observable by other firms when
broadcasting information. For searching to be significant and transmission
possible, it is necessary that cognitive proximity generate interaction to let
an effective spill over take place. The ensuing ability to broadcast relevant
technological information may be measured by a basic index specific to each
pair ij of firms in the economy. Accordingly, let aij indicate this index in
terms of the part of each firm j’s total innovative fitness that can cognitively
be passed on to firm i should the latter be in a position to observe the former.
The entire web of interfirm technological spillover can then be designated by
a square, JxJ, matrix A. Clearly, the main diagonal of this matrix is made
up by 0’s, aii = 0 , since no firm broadcasts information to itself. Matrix A
is simply an indicator of how well observing firms understand the technology
of other firms and, therefore, it states no more than a spill over potential as
a structural characteristic of the economy.
Bounded rationality restricts the number of neighgbours that a firm can
usefully search to glean technological information and hinders, furthermore,
the optimal choice of new ones when a firm gets the chance to adjust its
neighbourhood.. Actual observation is restricted to a more or less narrow
neighbourhood made up by firms whose informative usefulness has been dis-
covered by a searching process. We accordingly postulate that firms carry out
an active search to single out neighbours best suited to pass on innovative
capability. It is assumed that each firm i searches among its potential infor-
mation suppliers j whose broadcasting capacity is ai = (aij ), j = 1, 2...J,
those that at each point in time it is able to choose as a target for innovative
information and actually observe. This choice can be formalised by introduc-
ing the proximity matrix B(t) = [bij (t)] where each bij (t) = 1 or bij (t) = 0
according to whether neighbour j has been or hasn’t been identified as a use-
ful contributor. This procedure defines matrix M (t) = (aij bij (t)). Thus, the
innovative capability that is determined by interaction can be formalised by
the system M (t)V (t) where actually observed firms are restricted to a limited
number of neighbours. The general equation for firm i’s innovative capability
is4

J
Vi (t) = aij bij (t) Vj (t) + Ci (t)
j=1
and the system for all firms:
V (t) = [I − M (t)]−1 C(t)

where [I − M (t)]−1 plays the role of an endogenous matrix multiplier of
in-house capabilities.
Since firms are bounded in their rationality, different neighbourhood con-
figurations lead to different multipliers.Attempts to improve networks lead to
neighbours’ change and thus to change of matrix M (t).
2.1 Neighbourhood Structure
The structure in which we describe firms’ innovative capability can be repre-

sented by a directed graph of J nodes each of which is connected with other
nodes in two different but overlapping ways. The first is the number of con-
nections that each firm establishes when observing other firms to determine
its own innovative capability. The number of ki,in << J connections defines
for firm i the dimension of its inward neighbourhood. This number is substan-
tially smaller than J since searching is costly and observation bounded . This
neighbourhood can formally be defined as
Γi (t) = {γj ; j : j = 1, 2...J ∧ bij (t) = 1}
This is the set of firms from which at any time t firm i is able to glean
innovative capability through observation and learning.
The second kind of neighbourhood, which we term the outward , is made
up for each firm j by firms that actually observe it. It passively results as a
consequence of their networking activity. Let it be defined by:
Ψj (t) = {i : i = 1, 2...J ∧ bij (t) = 1}
Its size determines the impact of an observed firm’s technological capability

as it propagates throughout the economy contributing to overall performance.
For this purpose, we classify the population of firms according to classes of
their outward neighbourhood size and then define an impact factor by ranking
them.
Definition 1. Technological paradigm setters emerge when the probability of

each impact factor rank is positive.
4
Absorbing the impact of spillovers is clearly a process that requires an adjustment
in time. We simplify this problem by assuming that the time required to complete
adjustment is negligible in relation to the system evolution.
2.2 Evolution
Given this neighbourhood structure, evolution owes to two basic determinants:

search routines and exogenous changes of individual firms’ in-house innova-
tive capabilities. Firms construe their inward neighbourhood by an active
search aimed to single out members that contribute capability to their own.
This search, while bounded by the neighbourhood in which the firm happens
to be nested, may take place according to a variety of algorithms. We have
chosen one that responds to the criteria of bounded rationality and satisfic-
ing. We propose two versions that respectively capture a strong and a weak
form of bounded rationality. In both, we first conjecture that the cardinality
of Γi is |Γi | = ki,in J and generate the choice of neighbours and the evo-
lution of this neighbourhood according to the following routine: each firm i
assesses the fitness contribution of its existing neighbours and picks out the
least contributing one:
γi (t − 1) = arg min [aij bij (t − 1) Vj (t − 1)]

j∈Γi (t)
Secondly, the identified neighbour is substituted with a new one by ran-

domly drawing among the limited number of the latter own neighbours, in
the case of weak bounded rationality, or by randomly drawing among the re-
maining J − ki,in − 1 members of the entire economy, in the case of strong
but bounded rationality. In either case, to generate a new Γi (t) it is necessary
that this simple condition be satisfied:
Vi (t) > Vi (t − 1)
This procedure redefines at each time step M (t) and the system then
generates a new set of solutions.
Next to the dynamics generated by neighbourhood adjustment we intro-
duce in the system the autonomous and independent dynamics involving the
in-house capability C(t). This vector is subject to change by a random draw
of some i ∈ (1, 2...J) and by randomly redefining the ith component by a new
random value Ci (t) uniformly chosen between 0 and 1. These occurrences are
arrivals that take place according to a predetermined mean waiting time µ.
3 Parameters and Benchmarking

The purpose of this model is to study the emergent properties of neighbour-
hood formation their stability and impact on average technological capabili-
ties. In order to check for the impact of randomness in firms’ searching activity
we simplify the model by assuming parameters that insure an even field of
equal starting points. It is, accordingly, conjectured that all firms have the
same skill in broadcasting information and in spilling over their innovative
capability. Matrix A will, therefore, be set at A = (aij = a), ∀i, j. Parameter
a measures the strength of interaction and the degree to which each firm’s in-
novative capability depends on other firms’, given a matrix M (t) = [abij (t)].
For the system to yield positive solutions for any given vector C(t) and pre-
vent explosive behaviour in V (t), a must be kept as a ≤ ki,in
1
, ∀i. We further
assume that the number of inward neighbours is the same for all firms and
constant in time: ki,in = const, ∀i.
In order to make a significant assessment of innovative performances as
simulations are carried out, it is expedient to set benchmarks against which
they can be measured. For this purpose, we propose the following definitions.
Definition 2. Let the vector C̄(t) be the vector of elements of C(t) rear-
ranged in decreasing order and let the matrix B = [bij ] where bij = 1, ∀i =
1, 2....kin , ∀i = j, j ≤ kin + 1 and bij = 1, ∀i > kin + 1......J; bij = 0 otherwise.
J
Furthermore define v (t) = J1 j=1 V j (t), where V (t) = [I − M ]−1 C (t) and
M = aB .

Definition 3. Let matrix B = bij = kJ−1 inn
, ∀i = j, bii = 0∀i . We then define
J
v (t) = J1 j=1 V j (t) , M = aB and V (t) = [I − M ]−1 C(t).
Remark 1. If V̄ (t) can be considered as the capability values resulting from

an optimally planned network, V (t) is the average outcome from a random
choice of neighbours.
J
Definition 4. Given that v (t) = J1 j=1 Vj (t) , an index of average relative
capability gains can be construed as
1 v (t) − v (t)
T
φ (T ) =
T t=t v (t) − v (t)
0
where T is the time period considered and t0 is a sufficiently large transitional

period to render initial conditions irrelevant.
The difference v (t) − v (t) is the system’s average gain resulting from the
actual choice over the potential maximum one v (t) − v (t). Index φ (T ) takes
economically meaningful values between 0 and 1 for T → ∞.
In the following section we run simulations with a population of firms
J = 64, setting the number of inward neighbours kin = 3 and the spillover
coefficient a = 0.25.
4 Simulation Results
The model shown in section two has been simulated by tuning two crucial vari-
ables: (i) the mean waiting time of shock arrivals, µ, determining a change of
C(t) and (ii) the neighbour searching routine that we call τ . The latter variable
is defined in terms of the population of firms’ space spanned when searching
for a new neighbour. At one extreme, if a broad capacity to investigate and
gather information from all firms is assumed, this space is the whole popula-
tion of firms in the economy J; otherwise, any intermediate number up to the
other extreme where firms are limited to explore only their extant neighbour-
hood. Thus, by calling π the probability of exploring across the entire firm
space at each attempt to change a randomly chosen neighbour, we define it
as π ≡ τ1 . It follows that when τ = 1, (π = 1), the search routine is such that
firms always look across the whole economy for a new neighbourhood member
but τ tends to infinity, τ → ∞, when searching is entirely restricted within
their neighbourhood (π → 0, they never look outside). Intermediate values
state the number of times searching occurs inside the firm’s neighbourhood
and just once outside.
4.1 Search Routines and Efficiency
It is in the model’s logic that performance be gradually improved by adjusting

capability-contributing neighbours whatever the choice of the search routine.
This steady improvement, however, is more or less deeply upset by negative
as well as positive shocks arriving randomly according to different waiting
times, µ. In the following diagram we plot the index defined in the previous
section, φ (T ). As it can be seen, it is indeed confirmed that for any average
waiting time, a searching strategy confined to one’s own neighbourhood with
some across-the-board exception brings about a clear relative improvement
in the system’s average innovative capability but up to a certain point. The
differently printed lines correspond to as many shock mean waiting times in
ascending order.
That performance improves as the mean waiting time, µ, rises is obviously
the case since firms have more time to adjust and absorb random perturba-
tions. What is more interesting, however, is the clearly different outcome of
different search routines. It clearly appears that whatever µ turns out to be,
performance is noticeably improved towards the theoretical maximum as firms
cease to randomly scout the entire economy when they get the chance to sub-
stitute the least performing neighbour. As they concentrate on a local rather
than a global search, better results are obtained at least up to approximatly
τ ∗ , where τ ∗ is located in the environs of τ = 4 − 5 in the simulations we have
run: this means that it pays to search within one’s own neighbourhood and yet
to look over the entire economy approximately once every approximately τ ∗
times. Insisting on an even more localised search, however, invariably worsens
performance. The reason why this happens is quite straightforward. Localised
search has both an advantage and a disadvantage over the purely random one
across the whole economy. The former lies in the fact that if a firm has high
performing contributors, their own contributing neighbours are also likely to
be high performing ones; thus, excluding the least performing of these high
Fig. 1. Efficiency index φ (T ) as a function of τ for mean-waiting time µ =8, (dots)

µ =16 (dashed line) and µ =32 (continuous line)
contributors to substitute with one of its own neighbours carries a high prob-
ability of chancing on an even more performing contributor. On the contrary,
if a firm were to carry its search for substitution over the entire firm space, it
would stand an equal probability of finding either a poor or a high perform-
ing new neighbour. An improvement in a neighbour’s contribution could then
come from a comparatively mediocre firm The disadvantage lies in the fact
that if a firm gets locked in poor performers, localised search is likely to per-
petuate mediocre performance. As a consequence, there is a trade off between
localised and across-the-board search. It pays to search locally but to renew
the pool of neighbours by searching randomly across the whole economy every
now and then. To clarify this point, consider the case of a firm that starts its
neighbour adjustment process through local search and by having high per-
formers: it would clearly manage to quickly raise its technological capability.
If, however, one of its neighbours is hit by a negative shock sizably lowering
its performance, the firm would cease to obtain high positive externalities and
its capability growth would consequently plunge. Exclusively localised search
would then leave this firm in a poor performing environment until positive
shocks arrive to lift it out of this lock in: a further drawback is that positive
and negative shocks are equally probable for any given waiting time µ. The
only way to break the lock in is to search outside one’s own neighbourhood.
It follows that a good routine is one that considers local search to exploit the
probability of finding high performers but that periodically searches abroad
for better ones to avoid a poor performing stalemate.
4.2 Outward Neighbourhoods and Paradigm Setters’ Emergence
The evidence shown above indicates that there is an important relationship be-
tween searching routines and their efficiency. It follows that farsighted search
still has an important role to play and that short-sightedness and farsight-
edness must be balanced out to yield the greatest search efficiency. Results
provided by simulations on how technological leadership evolves help to ex-
plain this emergent phenomenon.
The following figures show diagrams in which the x-axis represents classes
of inward neighbours (i.e. the number of firms by which each firm is observed)
and the y-axis represents the average number of firms belonging to each class
within the considered time span. The population of firms is made up by 64
individuals split in 32 classes: thus, the first class in each diagram includes
firms having either 1 or nought neighbours, the last one either 62 or 63 neigh-
bours. We consider as paradigm setters firms lying in the last two classes,
that is by those observed by almost 95% of all firms. We begin our analysis
by discussing the case in which µ = 8 and τ = 1.
Fig. 2. Average frequency distribution of quantiles per time unit, in log − log space;
µ =8, τ =1(dots), τ =2 (dashed line), τ =3 (gray line) and τ =5 (black line).
In Figure 2, the dotted line reveals that the last classes (from 20 to 32)
are empty. This means that when the mean waiting time of shock arrivals
is too short and the search routine randomly spans the whole economy, the
probability of paradigm setters’ emergence is nil. Furthermore, as we have
argued above, lack of adjustment time and broad randomness of searching
hinder many firms from seeking out high performers and, thus, neighbourhood
configurations remain suboptimal. This pattern retains its main features as
long as the search routine remains defined by τ = 1 even if mean waiting times
lengthen to allow sufficient neighbourhood adjustment (see Figure 3 and 4,

dotted lines). Random choice of new neighbours from the entire firm space
implies performance mediocrity but, for better or for worse, also avoidance
of a lock-in no matter what the shock arrival waiting time is. Firms keep
on changing their neighbours so much that no particular firm is allowed to
set its technological capability as a paradigm for the rest of the economy:
as a result, the last classes remain empty and no paradigm setter emerges.
As the routine search becomes more local, see Figure 2, 3 and 4 dashed and
continuous gray and black lines for τ = 2, 3 and 5 respectively, paradigm
setters begin to emerge and as searching approaches τ → ∞, that is when the
choice of new neighbours is restricted exclusively within the narrow bounds
of a firm’s own neighbourhood, the same firms tend to persist as paradigm
setters. The economy ends up by having a distribution featuring very few
paradigm setters and a multitude of firms observed by no one else: firms
with zero outward neighbours. This result occurs independently of τ provided
that µ is sufficiently high, i.e. if shocks hit the economy sufficiently seldom.
The following Figures 2, 3 and 4 illustrates this point for µ = 8, 16 and 32,
respectively.
Fig. 3. Average frequency distribution of quantiles per time unit, in the log − log
space; µ =16, τ =1 (dots), τ =2 (dashed line), τ =3 (gray line) and τ =5 (black
line).
In these cases, the probability that there be firms with 62 or 63 techno-

logical onlookers is significantly positive (see dashed, gray and black lines).
To compound these results, it is important to consider what happens to the
two or three firms that temporarily emerge as leaders and eventually become
paradigm setters as evolution unfolds, shocks arrive and firms’ adjustment
takes place. Simulations show that after some time these two or three firms
Fig. 4. Average frequency distribution of quantiles per time unit, the log − log
space; µ =32, τ =1 (dots), τ =2 (dashed line), τ =3 (gray line) and τ =5 (black
line).
strengthen their acquired status and rise to the role of paradigm setters as
the ones that most other firms elect as their inward neighbours: they possess,
in other words, the highest number of outward neighbours.
It is apparent that the emergence of paradigm setters is related to the
achievement of better performance. Our data suggest that paradigm setters’
emergence precedes the attainment of maximum performance, and as such it
appears to be its prerequisite: the former emerge for τ = 2, while the latter is
achieved for values of τ between 4 and 6. Lock-in and renewal by looking over
the whole economy balance out in the region around τ ∗ . Nevertheless, as τ
rises past the observed peak and search becomes ever more local, the lock-in
effect becomes pervasive and average capability falls. There is an interesting
implication in this statement: past the maximum τ ∗ , while paradigm setters
are the most observed firms in the economy and as such set the technological
standards through their spillovers, they need not be the economy’s most per-
forming members. It is quite likely, in fact, that there may be firms with very
high innovative capability that are left unchosen because never discovered by
a very local search routine. In this sense, the system is likely to lock into a
progressively more inefficient leadership.
5 Conclusions
The foregoing analysis highlights the importance of searching and networking
in fostering the development of technological capabilities to innovate in a con-
text of bounded rationality. Firms obtain information and learn when crucially
placed in a cognitive and information providing neighbourhood. Technological
spillovers flow and give other firms the opportunity to learn only if networks
come into being to give shape to searching and make learning possible. This
paper depicts this process as an effort by firms, which do carry out their own
in-house innovation capability building, to seek out high performers able to
contribute to the latter. Routines differ according to the breadth of this search.
Thanks to a simulated, simple, linear system, it is shown that broad, economy
wide search routines are inefficient. The system’s average innovative capability
rises when searching is more local and firms are constrained to seek out new
neighbours only among their discarded neighbour’s neighbours. Up to a point.
While the said ploy implies a technological lock-in, the probability of finding
better performing firms increases. This result, however, occurs only if firms
still carry out a broad, economy wide search with a frequency that simulations
approximately identify. Thus, tuning short-sightedness and farsightedness im-
proves the system’s innovative efficiency. Past a given combination of the two,
the system slides towards increasing mediocrity but paradigm setters emerge
as a permanent and systematic feature of this hypothetical economy. The more
local search is, the greater is the likelihood that paradigm setters’ emergence
occurs. When searching is highly and in the limit exclusively local, a lock-in
into mediocre paradigm setters takes place and neighbourhoods are no longer
providers of efficient spillovers.
Exogenous shocks set another time scale to the system and upset the
learning and neighbour choosing adjustment. As shock arrival waiting time
increases the system’s technological efficiency rises but there is very little
impact on paradigm setters’ emergence that, therefore, appears to depend
only on the chosen routine pattern.
Further research is required to investigate the effects of a lopsided distribu-
tion of spillover coefficients (aij s) and the possibility that inward neighbour-
hoods may evolve to produce emergent properties of the system’s innovative
capabilities.
References
1. Albert R, Barabasi A-L (2002) Statistical mechanics of complex networks. Re-
views of Modern Physics 74:47–97.
2. Arenas A, Diaz-Guilera A, Pérez CJ, Vega-Redondo F (2002) Self-organized
criticality in evolutionary systems with local interaction. Journal of Economic
Dynamics and Control 26:2115–2142.
3. Arenas A, Diaz-Guilera A, Guardiola X, Llas M, Oron G, P érez CJ, Vega-
Redondo F (2001) New Results in a Self-organized Model of Technological
Evolution. Advances in Complex Systems 1:1–12.
4. Cowan R, Jonard N (2004) Network Structure and the difusion of knowledge.
Journal of Economic Dynamics and Control 28:1557–1575.
5. Cowan R, Jonard N (2005) Innovation on a network. Forthcoming in Structural
Change and Economic Dynamics.
6. Kirman A (1997a) The Economy as an Evolving Network. Journal of Evolu-

tionary Economics 7:339–353.
7. Kirman A (1997b) The Economy as an Interactive System. In Arthur W B,
Durlauf S N, Lane D (eds) The Economy as a Complex Evolving System II.
Santa Fe Institute, Santa Fe and Reading, MA, Addison-Wesley.
8. Watts D J, Strogatz S H (1998) Collective Dynamics of ’small world’ networks.
Nature 393:440–442.
Part III
Economic Dynamics
Statistical Properties of a Heterogeneous Asset
Pricing Model with Time-varying Second
Moment
Carl Chiarella1 , Xue-Zhong He1 , Duo Wang2

1
School of Finance and Economics, University of Technology, Sydney, PO Box
123 Broadway, NSW 2007, Australia
2
LMAM, Department of Financial Mathematics, School of Mathematical
Sciences, Peking University, Beijing 100871, People’s Republic of China
Summary. Stability and bifurcation analysis of deterministic systems has been

widely used in modeling financial markets. However, the impact of such dynamic
phenomena on various statistical properties of the corresponding stochastic model,
including skewness and excess kurtosis, various autocorrelation (AC) patterns of
under and over reactions, and volatility clustering characterised by the long-range
dependence of ACs, is not clear and has been very little studied. This paper aims
to contribute to this issue. Through a simple behavioural asset pricing model with
fundamentalists and chartists, we examine the statistical properties of the model
and their connection to the dynamics of the underlying deterministic model. In
particular, our analysis leads to some insights into various mechanisms that may
generate some of the stylised facts, such as fat tails, skewness, high kurtosis and
long memory, observed in high frequency financial data.
JEL classifications: D83; D84; E21; E32, C60
Key words: Fundamentalists, chartists, stability, bifurcation, investors’ under-

and over-reactions, stylized facts.
1 Introduction
As a result of a growing dissatisfaction with (i) models of asset price dynam-
ics based on the representative agent paradigm, as expressed for example by
Kirman [25], and (ii) the extreme informational assumptions of rational ex-
pectations, research into the dynamics of financial asset prices resulting from

Financial support for Duo Wang by NSFC.10271007 and RFDP.20010001042 is
acknowledged. This paper was prepared when Wang was visiting the University
of Technology, Sydney, whose hospitality is gratefully acknowledged.
110 Carl Chiarella, Xue-Zhong He, and Duo Wang
the interaction of heterogeneous agents has developed, including for example

Arthur et al. [2], Brock and Hommes ([3], [4]), Brock and LeBaron [6], Bullard
and Duffy [7], Chen and Yeh ([8], [9]), Chiarella [10], Chiarella et al. [11],
Chiarella and He ([12], [13], [14]), Dacorogna et al. [16], Day and Huang [17],
Farmer and Joshi [19], Gaunersdorfer [20], Gaunersdorfer et al [21], Hommes
([23], [24]), LeBaron et al. [26], Lux ([27], [28], [29]) and Lux and Marchesi
[30]. In particular, Brock and Hommes ([4], [5]) have introduced the concept of
an adaptively rational equilibrium, where agents adapt their beliefs over time
by choosing from different predictors or expectations functions, based upon
their fitness functions measured by realized profits. The resulting dynamical
system is capable of generating the entire zoo of complex behaviour from lo-
cal stability to high order cycles and chaos as various key parameters of the
model change. Brock and Hommes’s framework has been extended further in
Gaunersdorfer [20] and Chiarella and He ([12], [13], [14]) to incorporate het-
erogeneous variance, risk and learning under both Walrasian auctioneer and
market maker scenarios. It is found that the relative risk attitudes, different
learning mechanisms and different market clearing scenarios affect asset price
dynamics in a very complicated way. It has been shown (e.g. Hommes [24])
that such simple nonlinear adaptive models are capable of explaining impor-
tant stylized facts, including fat tails, clustered volatility and long memory,
of real financial time series.
There are two goals this literature is trying to achieve, first to explain
various market behaviour and, second to replicate the econometric properties
and stylized facts of financial time series. These heterogeneous agent models,
in particular the theoretically oriented ones, have shown great potential in
achieving the first goal. However, we still have some distance in achieving the
second goal. On the one hand, it is well known that most of the stylized facts
can only be observed for high frequency data, such as weekly, daily or intra-
day data. However, most of the heterogeneous agent models have difficulty in
replicating realistic time series at high frequency. On the other hand, stability
and bifurcation analysis have been widely accepted tools in investigating such
models, but it is not clear yet how various types of bifurcations of the deter-
ministic system can be used to characterize various price volatility and time
series properties (such as distribution density and autocorrelations of returns)
of the corresponding stochastic asset pricing models. These issues have been
tackled recently in He [22] through a simple fixed fraction heterogeneous asset
pricing model. This paper is largely motivated by these issues, in particular
by He [22], and seeks to contribute to both of the aforementioned goals.
Gaunersdorfer et al. [21] consider a simple asset pricing model of funda-
mentalists and chartists. The fundamentalists believe that tomorrow’s price
will move in the direction of the fundamental price, while chartists derive their
beliefs from a simple technical trading rule using only the latest observed price
and extrapolation of the latest observed price change. By assuming a constant
conditional variance for both types of traders and imposing a penalty function
in the fitness function of the chartists (to ensure that speculative bubbles can-
Statistical Properties of Heterogeneous Asset Pricing Model 111
not last forever), they show that volatility clustering can be characterized by
a coexistence of a stable steady state and a stable limit cycle, which arises as a
consequence of a so-called Chenciner bifurcation. Simple economic intuition of
this result is also provided. This paper is a significant contribution in relating
asset price volatility to the bifurcation nature of the underlying deterministic
system. However, it is not clear if such a characterization of volatility cluster-
ing still holds for the corresponding stochastic system when the fundamental
price follows a random process, in particular when the trading frequency is
high.
Given the variety of technical trading rules and differing risk aversion of
various investors, this paper introduces a risk adjusted demand function for
the chartists by assuming that they use a weighted process of past prices to
estimate both conditional mean and variance. Therefore their demand func-
tion is a nonlinear function of the conditional mean and variance, instead of
a linear function of the conditional mean only as in Gaunersdorfer et al. [21].
It is found that the mechanism of variance adjusted demand function of the
chartists provides a natural way to prevent the price from getting stuck in
a speculative bubble3 . Similar to Brock and Hommes ([4], [5]), an adaptive
model based on the fitness function is then obtained. We examine how the
price dynamics of the risky asset is affected by the reactions of investors, the
switching intensity of the fitness function, and the weighting process and risk
adjustment of the chartists.
The plan of the paper is as follows. We first develop a simple adaptive as-
set pricing model of fundamentalists, chartists and noise traders. A statistical
analysis is then conducted when parameters are near the bifurcation bound-
aries of the underlying deterministic model and a connection between the
Hopf (flip) bifurcation and under(over)-reaction AC patterns is established.
The paper finishes with a discussion on the capability of the model to generate
the stylised facts and the long memory of high frequency financial time series.
2 The Model
Following the framework of Brock and Hommes [5], consider an asset pricing
model with one risky asset and one risk free asset. It is assumed that the risk
free asset is perfectly elastically supplied at the risk-free rate r (per annum).
Let pt be the price (ex dividend) per share of the risky asset at time t and
{yt } be the stochastic dividend process of the risky asset. Then the wealth of
investor h at t + 1 is given by
Wh,t+1 = RWh,t + Rt+1 zh,t , (1)
where
3
In Gaunersdorfer et al. [21], this is ensured by artificially adding a penalty function
in the fitness function of the chartists, which is unnecessary in our model.
Rt+1 = pt+1 + yt+1 − Rpt (2)

is the excess capital gain/loss, R = 1 + r/K, K is the trading frequency
per annum4 , Wh,t is the wealth at time t and zh,t is the number of shares
of the risky asset purchased at t. Denote by Ft = {pt−1 , · · · ; yt−1 , · · · } the
information set formed at time t. Let Eh,t , Vh,t be the beliefs of investor type
h about the conditional expectation and variance, based on Ft . Then it follows
from (1) and (2) that
2
Eh,t (Wt+1 ) = RWt + Eh,t (Rt+1 )zh,t , Vh,t (Wt+1 ) = zh,t Vh,t (Rt+1 ). (3)
Assume investor has a CARA (constant absolute risk aversion) utility function
u(W ) = −e−ah W with risk aversion coefficient ah > 0, and maximises his/her
expected utility of wealth, leading the optimal demand
Eh,t (Rt+1 )
zh,t = . (4)
ah Vh,t (Rt+1 )
In the following discussion, we adopt the popular fundamentalist/chartist

model by assuming that all investors can be grouped into either fundamen-
talists (type 1) or chartists (type 2).
Fundamentalists—The fundamentalists are assumed to believe that the
expected market price pt is mean reverting to their perceived fundamental
value p∗t and the conditional variance of the market price is constant. That is,

E1,t (pt+1 ) = p∗t−1 + v(pt−1 − p∗t−1 ), 0 ≤ v ≤ 1
(5)
V1,t (pt+1 ) = σ12 ,
where p∗t is the fundamental price of the risky asset estimated by the funda-
mentalists at some cost, v is the speed of mean reversion estimated by the
fundamentalists, and σ1 > 0 is a constant. In particular, E1,t (pt+1 ) = p∗t−1 for
v = 0 and E1,t (pt+1 ) = pt−1 for v = 1. The conditional expectation of the fun-
damentalists (5) can also be written as E1,t (pt+1 ) = (1 − v)p∗t−1 + vpt−1 , 0 ≤
v ≤ 1, which is a weighted average of the fundamental price and latest
price. Hence small (large) values of v indicate that the fundamentalists give
more (less) weight to the fundamental price and less (more) weight to the
latest price, believing that price moves quickly (slowly) towards its funda-
mental value p∗t . For convenience of discussion, we say the fundamentalists
over(under)-react (to the market price) when more (less) weight v is given to
the market price.
Chartists—Unlike the fundamentalists who are able to obtain a good esti-
mate of the fundamental value, chartists base their trading strategy on signals
generated from the costless information contained in recent prices. The sig-
nal may be generated by comparing the latest price pt−1 with some reference
4
Typically, annually, quarterly, monthly, weekly and daily trading periods corre-
spond to K = 1, 4, 12, 52 and 250, respectively.
price trends p̃t−1 , such as a moving average process. Specifically, the chartists
consider the realizations of p̃t−1 as random drawn from some distribution.
The distribution can be conditional on past realized values. For simplicity,
we assume that p̃t−1 is conditionally distributed on prices pt−2 and pt−3 with
weighting probabilities w and 1 − w, respectively. Then the conditional mean
and variance of the trend can be estimated, respectively, as

p̄t−2 ≡ wpt−2 + (1 − w)pt−3 , 0 ≤ w ≤ 1,
(6)
2
σ̄t−2 ≡ w[pt−2 − p̄t−2 ]2 + (1 − w)[pt−3 − p̄t−2 ]2 .
Based on the trading signals pt−1 −p̃t−1 and the conditional mean and variance
estimates (6), we make the following assumptions for the chartists:

E2,t (pt+1 ) = pt−1 + g(pt−1 − p̄t−2 ), g ∈ R,
(7)
V2,t (pt+1 ) = σ12 [1 + bσ̄t−2
2
], b ≥ 0,
where g ∈ R is the estimated extrapolation rate of the chartists. That is,
the chartists’ beliefs are based on the latest price and their extrapolation
of the trading signals generated from the trend. In particular, chartists are
called trend followers when g > 0 and are contrarians when g < 0. For w = 1,
E2,t (pt+1 ) = pt−1 +g(pt−1 −pt−2 ) which is the case discussed in Gaunersdorfer
et al. [21] and, for w = 0, E2,t (pt+1 ) = pt−1 + g(pt−1 − pt−3 ). Similarly, for
convenience of discussion, we say the chartists over(under)-react when they
extrapolate strongly (weakly), that is when |g| is large (small). With regard
to the chartists’ estimate of the variance, they use the historical variance to
scale up the fundamental variance through the parameter b. High b reflecting
a greater sensitivity to variance risk.
For the dividend process, we assume E(yt ) = ȳ, V (yt ) = σy2 . In particular,
when p∗t = p∗ is constant, ȳ = rp∗ . Based on the above assumptions, the
optimal demands for the fundamentalists and chartists are given, respectively,
by
z1,t = [p∗t−1 + v(pt−1 − p∗t−1 ) + ȳ − Rpt ]/A1 ,
(8)
z2,t = [pt−1 + g(pt−1 − p̄t−2 ) + ȳ − Rpt ]/A2,t .
where
A1 = a1 (σ12 + σy2 ), A2,t = a2 [σy2 + σ12 (1 + bσ̄t−2
2
)] (9)
and p̄t and σ̄t2are defined by (6).
Let U1,t and U2,t be the realized profit of the fundamentalists and chartists,
respectively, defined by Ui,t = Rt zi,t−1 − Ci , i = 1, 2, where Ci ≥ 0 measures
the total cost incurred by the agents in generating information to form their
estimates. Let ni,t be the fractions of agents of type i at time t. Assume the
fractions are formed on the basis of discrete choice probability (see Manski
and McFadden [31], Anderson, de Palma and Thisse [1], Brock and Hommes
([4], [5])), namely

2
ni,t = exp[βUi,t−1 ]/Zt−1 , (i = 1, 2), Zt−1 = exp[βUi,t−1 ], (10)
i=1
where β(> 0) is the intensity of choice measuring how fast agents switch
among different prediction strategies. Let mt = n1,t − n2,t . Then n1,t = (1 +
mt )/2, n2,t = (1 − mt )/2 and

β β
mt = tanh Rt−1 (z1,t−2 − z2,t−2 ) − C , (11)
2 2
where C = C1 − C2 . It is assumed that the fundamentalists expend more

on generating relevant information to generate their estimates so that the
constant C ≥ 0 in general.
To clear the market through a Walrasian scenario, we introduce noise
traders whose supply of the risky asset is denoted by zn,t ∼ N (0, σn2 ). The
market is viewed as finding the price pt that equates the sum of these demand
from both fundamentalist and chartist schedules to the supply of the noise
traders via
(1 + mt )z1,t + (1 − mt )z2,t = zn,t . (12)
Substituting (8) and (11) into (12), one obtains the market cleaning price pt
F (pt−1 , p∗t−1 , mt , p̄t−2 , σ̄t−2

2
, zn,t , yt )
pt = ∗ 2 ,z , (13)
G(pt−1 , pt−1 , mt , p̄t−2 , σ̄t−2 n,t , yt )
where
F (pt−1 , p∗t−1 ,mt , p̄t−2 , σ̄t−2

2
, zn,t , yt ) = A2,t [1 + mt ][p∗t−1 + v(pt−1 − p∗t−1 ) + ȳ]
+ A1 [1 − mt ][pt−1 + g(pt−1 − p̄t−2 ) + ȳ] − zn,t A1 A2,t ,
G(pt−1 , p∗t−1 ,mt , p̄t−2 , σ̄t−2
2
, zn,t , yt ) = R[(1 + mt )A2,t + A1 (1 − mt )],
p̄t , A1 and A2,t are defined by (6) and (9). Because of the stochastic nature
of the dividend process yt , the fundamental value p∗t and the noise supply,
the equilibrium price equation (13) is equivalent to a high-order nonlinear
stochastic discrete system in general.
When the stochastic terms in equation (13) take their mean levels, one
obtain the corresponding deterministic system. For the deterministic system,
it is found from Chiarella et al. [15] that the constant fundamental price p∗
is the unique steady state price of the system. In addition, the steady state
price becomes unstable through a Hopf bifurcation only when the chartists
are trend followers (that is when g > 0). However, when the chartists are
contrarians (that is when g < 0), the local stability region is bounded by both
Hopf and flip bifurcations. However, depending on the weighting parameter
w, one of the boundaries may be unbinding. In particular, the steady state
price can become unstable through a Hopf bifurcation when w is small, or a
flip bifurcation when w is large, or both bifurcations when w is near 0.5. We
refer to Chiarella et al. [15] for more detailed analysis on the dynamics of the
deterministic system. In the following discussion, we examine the statistical
properties of the market price of the stochastic model (13). In particular we
are interested in how statistical properties exhibit different behaviour near

different types of bifurcation boundaries of the deterministic system. In so
doing, we obtain some insights into the interaction of the dynamics of the
deterministic system and the noise processes.
In the following, we choose the fundamental price p∗o = $100, annual risk-
free rate r = 5%, annual volatility of the fundamental price σ = 20%. Trading
frequency K = 250, which corresponds to a daily trading period. Correspond-
ingly, the total risk-free return per trading period R = 1 + r/K = 1.0002,
daily price volatility σ12 = (p∗ σ)2 /K = 8/5 and daily dividend volatility
σy2 = r2 σ12 = 1/250. We also choose the risk aversion coefficients for both
types of investor as a1 = a2 = 0.8, the cost difference C = 0, the coefficient
of endogenous variance component b = 1 and the switch intensity β = 0.1.
150 0.050
rt
pt*
Svar2
0.025
100
0.000
−0.025
50
0 750 1500 2250 3000 3750 4500 0 750 1500 2250 3000 3750 4500
Density
30 AC(r t)
0.025
20
0.000
10
−0.025
−0.04 −0.02 0.00 0.02 0.04 0 20 40 60 80 100
2
AC(|r t |) 0.050 AC(r t )
0.025 0.025
0.000 0.000
−0.025
−0.025
0 20 40 60 80 100 0 50 100
Fig. 1. Time series plots of the prices p∗t , returns rt and autocorrelations of rt , |rt |
and rt2 .
We assume that the fundamental price p∗t follows a stationary random

walk process
p∗t = p∗t−1 [1 + t ], t ∼ N (0, σp2 ), p∗0 = $100.00. (14)
Fig.1 displays a realisation of the fundamental prices, p∗t and the correspond-
ing relative return rt = p∗t /p∗t−1 − 1. This fundamental price is used for all
of the following simulations. The simulation contains 5,000 observations, cor-
responding to about 20 years of daily prices. The corresponding statistical
properties, including distribution density of the return rt and the autocorre-

lations (ACs) of rt , |rt | and rt2 , are also included in Fig. 1 One can see that the
ACs of rt , |rt | and rt2 are not significant at the 95% level, which is consistent
with the stationary random walk process (14).
3 Hopf Bifurcations and Under-Reaction AC Patterns
Chiarella et al. [15] have extensively analysed the dynamical properties of the
deterministic skeleton of the price dynamics (13). It is found that the fun-
damental steady state price of the underlying deterministic system becomes
unstable through a Hopf bifurcation only in two situations. The first case
occurs when the chartists are trend followers (with any weighting parameter
w) and the second case occurs when the chartists are contrarians with lower
weighting parameter w. We now examine the stochastic properties of the mar-
ket price when parameters are near the Hopf bifurcation boundary in these
two different situations.
3.1 Case 1
We first consider the case when the chartists are trend followers. It is found
from various simulations that prices become explosive when (i) the parameters
v, g and w are either near or outside the Hopf bifurcation boundary, or (ii) the
switching intensity β is high. Also, an increase in parameter b can make prices
become less explosive. Given the destabilizing effect of the trend following
strategy, this result is very intuitive. To see the impact of the Hopf bifurcation
on the stochastic properties of the stochastic model, we choose w = 0.5, g = 1
and two different values for v = 0.3 and 0.9. It can be verified that the steady
state price of the deterministic model is locally stable with these selections
and, in addition, the parameter moves closer to the Hopf bifurcation as v
increases. In fact, the corresponding Hopf bifurcation values of g for v = 0.3
and 0.9 are given by g ∗ = 1.458 and 1.3, respectively.
To see the impact of the stochastic fundamental price and noise traders
on the price dynamics, we consider two different situations: σn = 0 and σn =
1. For σn = 0, the market price is influenced by the random fundamental
price p∗t only. When σn = 1, the market price is influenced by both the
random fundamental price p∗t and the noisy supply. Fig. 2 illustrates the time
series of the market price pt , the densities of the market fraction mt and the
return rt , and the ACs of rt , |rt | and rt2 for four combinations of (v, σn ) =
(0.3, 0), (0.3, 1), (0.9, 0) and (0.9, 1). Our numerical simulations show that the
features illustrated in Fig. 2 are robust, from which we arrive at the following
observations.
(i) The market prices pt follow closely the random fundamental price p∗t in all
cases. This may partially be due to the local stability of the steady state
150 150 150 150

100
pt
100 100 100

50 50 50 50
0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900
mt
20 10 200 20
0 0 0 0
−0.25 0.00 0.25 −0.25 0.00 0.25 −0.025 0.000 0.025 −0.2 0.0 0.2
50 40 50
100
rt
25 20 50 25
−0.05 0.00 0.05 −0.05 0.00 0.05 −0.01 0.00 0.01 −0.05 0.00 0.05
AC(r t)
1.0 0.50 1.0

0.5 0.25 0.5 0.5
0.0 0.00 0.0
0.0
0 50 100 0 50 100 0 50 100 0 50 100
AC(r t 2) AC(|r t |)
0.50 0.2 1.0 0.50

0.25 0.5
0.00 0.1 0.25
0.0
0 50 100 0 50 100 0 50 100 0 50 100
0.50 1.0 0.50
0.25 0.2
0.1 0.5 0.25
0.00 0.0
0 50 100 0 50 100 0 50 100 0 50 100
(v, σ n )=(0.3, 0) (v, σ n )=(0.3, 1) (v, σ n )=(0.9, 0) (v, σ n )=(0.9,1)
Fig. 2. Time series of the market price pt , distribution density of the population
fraction difference mt and return rt (compared with the normal distribution) and
ACs of rt , |rt | and rt2 for w = 0.5, β = 0.1, a = 0.8, g = 1, b = 1, C = 0, two different
values of v = 0.3 and 0.9 and without, with the noisy supply σn = 0, 1.
price of the underlying deterministic model, partially due to the mean

reverting expectation scheme of the fundamentalists.
(ii) The returns are not normally distributed as indicated by the significant
ACs in all cases. However, including noise traders makes the returns even
less normal, leading to skewness and high kurtosis.
(iii) The ACs of returns show a strong under-reaction pattern, characterised
by the oscillating ACs with positive ACs for small lags, as the parame-
ters move closer to the Hopf bifurcation (the cases (v, σn ) = (0.3, 0) and
(0.9, 0)). Such strong patterns are be washed out by introducing the noise
traders (the case (v, σn ) = (0.3, 1) and (0.9, 1)).
(iv) The noise traders make a significant contribution to the volatility clus-
tering effect characterised by the significant ACs of the |rt | and rt2 in cases
of (v, σn ) = (0.3, 1) and (0.9, 1).
(v) The distribution of the population difference mt shows that the investors
tend to switch toward the fundamental (trend following) strategy as the
parameters move away from (closer to) the Hopf bifurcation boundary,
though the volatility of mt is increased by the introduction of the noise
traders.
Overall, we can see that the Hopf bifurcation can be used to explain the under-
reaction of AC patterns. Also the stylised facts, including skewness, kurtosis,
volatility clustering and significant ACs of |rt | and rt2 , are observed when the
noise traders act as liquidity traders.
3.2 Case 2
We now consider the case when the chartists are contrarians and the weight
parameter w is small. In this case, the flip bifurcation boundary becomes
non-binding. It is found from simulations that, different from the previous
case, prices do not become explosive on both sides of the Hopf bifurcation.
An intuitive explanation is the fact that both fundamental and contrarian
strategies are stabilizing strategies. This may also be due to the unbinding flip
bifurcation. To see how the market price behavior near the Hopf bifurcation,
we choose v = 0.5, w = 0.3 and g = −1 and -3, which are located on both
sides of the Hopf bifurcation value g ∗ = −2.65 of the underlying deterministic
system.
150 150 150 150

pt
100 100 100 100

50 50 50 50
0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900
50 20 5.0
5.0
mt
25 10 2.5 2.5
−0.5 0.0 0.5 −0.5 0.0 0.5 −0.5 0.0 0.5 −0.5 0.0 0.5
100 20
rt
25 20
50 10 10
0
−0.02 0.00 0.02 −0.05 0.00 0.05 −0.1 0.0 0.1 −0.2−0.1 0.0 0.1 0.2
AC(r t)
0.50 0.25 1 1
0.25 0.00 0
0.00 0
−0.25
0 50 100 0 50 100 0 50 100 0 50 100
AC(r t 2) AC(|r t |)
0.2 0.2 1.0 0.50

0.5 0.25
0.0 0.0
0.0
0 50 100 0 50 100 0 50 100
0.50 0 50 100
0.2 1.0 0.4
0.25 0.5
0.00 0.0 0.0 0.2
0 50 100 0 50 100 0 50 100 0 50 100
(g, σ n )=(−1, 0) (g, σ n )=(−1, 1) (g, σ n )=(−3, 0) (g, σ n )=(−3, 1)
ACs of rt , |rt | and rt2 for w = 0.3, β = 0.1, a = 0.8, v = 0.5, b = 1, C = 0, two
different values of g = −1 and -3 and without, with the noise supply σn = 0, 1.
Fig. 3 illustrates the time series of the market price pt , the densities of
the market fraction mt and the return rt , and the ACs of rt , |rt | and rt2 for
four combinations of (g, σn ) = (−1, 0), (−1, 1), (−3, 0) and (−3, 1). Similar
statistical properties are found when the parameters are located inside the
local stability region, but near the Hopf bifurcation boundary, of the under-
lying deterministic system (that is the cases (g, σn ) = (−1, 0) and (−1, 1)).
However, when the parameters are located outside the local stability region,
the steady state price of the deterministic system becomes unstable through
a Hopf bifurcation and the ACs of returns of the stochastic model display
over-reaction patterns characterised by oscillating and decaying ACs which
are negative for odd lags and positive for even lags, though the patterns are
washed out significantly by the presence of the noise traders.
4 Flip Bifurcations and Under/Over-Reaction AC

Patterns
This section examines the impact of the flip bifurcation (of the deterministic
system) on the stochastic properties of the market price. As we know that
the fundamental steady state price of the underlying deterministic system
becomes unstable through a flip bifurcation when the chartists are contrarians
and the weighting parameter w is higher (in this case the Hopf bifurcation
of the underlying deterministic system becomes unbinding). Because of the
stablizing effect of both fundamental and contrarian strategies, prices become
less explosive on both sides of the flip bifurcation boundary (but more sensitive
to the switching intensity β). To see the market price behavior near the flip
bifurcation, we choose v = 0.5, w = 0.8 and g = −1 and -3, which are located
on both sides of the flip bifurcation value g ∗ = −2.18773 of the underlying
deterministic system.
Fig. 4 illustrates the time series of the market price pt , the densities of
the market fraction mt and the return rt , and the ACs of rt , |rt | and rt2 for
four combinations of (g, σn ) = (−1, 0), (−1, 1), (−3, 0) and (−3, 1). Similar
statistical properties are found when the parameters are located inside the
local stability region, but near the flip bifurcation boundary, of the underly-
ing deterministic system (that is the cases (g, σn ) = (−1, 0) and (−1, 1)). In
particular, the under-reaction AC pattern is observed when there are no noise
traders. However, for g = −3, the steady state price of the underlying de-
terministic system becomes unstable through a flip bifurcation and both the
stochastic fundamental price and the noise traders have different impacts on
the stochastic properties of the market price which are summarized as follows:
(vi) Both the population fraction difference mt and the return rt have bi-
modal distributions when there are no noise traders present. Correspond-
ingly, the AC displays a very strong over-reaction pattern characterised
by slowly decaying oscillating AC(rt ) which are negative for odd lags and
150 150 150 150

100 100 100 100
pt
50 50 50 50
0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900 0 1300 2600 3900
50 5.0 5.0
mt
20
25 10 2.5 2.5
0.0 0.1 0.2 −0.2 0.0 0.2 0.00 0.25 0.50 −0.5 0.0 0.5
100 50 20
20
rt
50 25 10 10
−0.01 0.00 0.01 0.02 −0.05 0.00 0.05 −0.05 0.00 0.05 −0.1 0.0 0.1 0.2
AC(r t 2) AC(|r t |) AC(r t)
0.50 1
0.25 0.0 0
0
0.00 −0.5
0 50 100 0 50 100 −1
0 50 100 0 50 100
0.4 0.4 1.0
0.2 0.50
0.2 0.5 0.25
0.0
0 50 100 0 50 100 0 50 100 0 50 100
0.50 0.4 1.0 0.50
0.25 0.5
0.00 0.2 0.25
0 50 100 0 50 100 0 50 100 0 50 100
(g, σ n )=(−1, 0) (g, σ n )=(−1, 1) (g, σ n )=(−3, 0) (g, σ n )=(−3, 1)
ACs of rt , |rt | and rt2 for w = 0.8, β = 0.1, a = 0.8, v = 0.5, b = 1, C = 0, two
different values of g = −1 and -3 and without, with the noisy supply σn = 0, 1.
positive for even lags. This strong AC pattern is also preserved for the ACs
of |rt | and rt2 . All these features are underpinned by the flip bifurcation
of the underlying deterministic system.
(vii) When the noise traders are present, most of the statistical features in
(vi) disappear. The bi-modal distributions of mt and rt disappear, show-
ing skewness and high kurtosis. In addition, the strong over-reaction AC
patterns on returns are washed out and, apart from the first two lags,
become insignificant. However, the ACs of both |rt | and rt2 are significant.
Overall, we can see that, when the parameters cross the flip bifurcation of
the deterministic system, the ACs pattern of the stochastic returns changes
from under-reaction to over-reaction and this is also consistent with trader
behaviour. When the noise traders are present, the model generates most of
the stylised facts that we discuss in the following section.
5 The Stylised Facts and Long Memory

The above analysis has led us to our first goal—to gain some insights into vari-
ous types of market price behaviour and to understand the possible mechanism
generating such behaviour. We have shown that various statistical aspects of
the stochastic system, including distributions and under/over-reaction AC
patterns are closely related to the stability and bifurcation analysis of the
deterministic system.
0.1
1500 S&P500−−−Price Return
0.0
1000
−0.1
500
−0.2
0 750 1500 2250 3000 3750 4500 0 750 1500 2250 3000 3750 4500
75 AC(r t)
Return Density 0.05
50
0.00
25
−0.10 −0.05 0.00 0.05 0.10 0 20 40 60 80 100

0.2
0.2 AC(|r t |) 2
AC(r t )
0.1
0.1
0.0
0 20 40 60 80 100 0 20 40 60 80 100
Fig. 5. Time series of the prices and (relative) returns of S&P500 from 18/11/1983
to 19/11/2003, distribution density of the return (compared with the normal dis-
tribution) and ACs of returns, absolute returns and squared returns for lags up to
100.
We now try to address our second goal, which is the most important and
difficult part of asset price modeling, to replicate the econometric proper-
ties and the stylized facts of financial time series. As a benchmark of such
financial time series, we consider the daily accumulated index of the S&P500
from 18/11/1983 to 19/11/2003. Fig. 5 illustrates the price and (relative) re-
turn time series for the last 20 years, the return distribution and the ACs
of returns, absolute returns and squared returns for lags up to 100. For a
comprehensive discussion of stylized facts and the so-called long memory of
returns of high frequency financial time series, we refer to Ding et al. [18] and
Pagan [32]. The stylised facts include excess volatility, volatility clustering,
skewness and excess kurtosis, while the long memory refers to the insignifi-
cant ACs on returns, but significant and decayed ACs on the absolute and
squared returns, as illustrated by Fig. 5 for the S&P500 daily index. Given
the simplicity of our model, it is not easy to reproduce all those features.
However, based on our analysis in the previous sections, one can see that the
model is indeed able to produce most of those features when the noise traders
act as liquidity traders. In particular, for the cases when the chartists are con-
trarians and the parameters are near either the Hopf or flip bifurcation (but
from inside the local stability region) of the underlying deterministic system
(w, g, σn ) = (0.3, −1, 1), (0.8, −1, 1), the return distributions show skewness
and excess kurtosis, the distributions of the population fraction difference
show some herding behaviour, while the ACs of returns display similar pat-
terns to the S&P500, apart from significant ACs of returns for the first few
lags and persistent ACs of the absolute and squared returns.
References
1. Anderson, S. and A. de Palma, J. Thisse (1993) Discrete Choice Theory of
Product Differentiation, MIT Press, Cambridge, MA
2. Arthur, W., Holland, J., LeBaron, B., Palmer, R., Tayler, P. (1997) Asset pric-
ing under endogeneous expectations in an artifical stock market Economic Notes
26(2):297–330
3. Brock, W., Hommes, C. (1997) Models of Complexity in Economics and Fi-
nance. chapter 1, pp. 3–44 In: Heij, C., Schumacher, J.M., Hanzon, B., Praag-
man, C. (eds) Systems Dynamics in Economic and Finance Models, Wiley, New
York
4. Brock, W.,Hommes, C. (1997) A rational route to randomness, Econometrica
65:1059–1095
5. Brock, W., Hommes, C. (1998) Heterogeneous beliefs and routes to chaos in
a simple asset pricing model, Journal of Economic Dynamics and Control
22:1235–1274
6. Brock, W., LeBaron, B. (1996) A structural model for stock return volatility
and trading volume, Review of Economics and Statistics 78:94–110
7. Bullard, J., Duffy, J. (1999) Using Genetic Algorithms to Model the Evolution
of Heterogeneous Beliefs, Computational Economics 13:41–60
8. Chen, S.-H., Yeh, C.-H. (1997) Toward a computable approach to the efficient
market hypothesis: an application of genetic programming, Journal of Economic
Dynamics and Control 21:1043–1063
9. Chen, S.-H., Yeh, C.-H. (2002) On the emergent properties of artificial stock
markets: the efficient market hypothesis and the rational expectations hypoth-
esis, Journal of Economic Behavior and Organization 49:217–239
10. Chiarella, C. (1992) The dynamics of speculative behaviour, Annals of Opera-
tions Research 37:101–123
11. Chiarella, C., Dieci, R., Gardini, L. (2002) Speculative behaviour and complex
asset price dynamics, Journal of Economic Behavior and Organization 49:173–
197
12. Chiarella, C., He, X. (2001) Asset pricing and wealth dynamics under hetero-
geneous expectations, Quantitative Finance 1:509–526
13. Chiarella, C., He, X. (2002) Heterogeneous beliefs, risk and learning in a simple
asset pricing model, Computational Economics 19:95–132
14. Chiarella, C., He, X. (2003) Heterogeneous beliefs, risk and learning in a simple
asset pricing model with a market maker, Macroeconomic Dynamics 7:503–536
15. Chiarella, C., He, X., Wang, D. (2004) Asset price dynamics with time-varying
second moment, Technical report, University of Technology, Sydney
16. Dacorogna, M., Muller, U., Jost, C., Pictet, O., Olsen, R., Ward, J. (1995)
Heterogeneous real-time trading strategies in the foreign exchange market, Eu-
ropean Journal of Finance 1:383–403
17. Day, R., Huang, W. (1990) Bulls, bears and market sheep, Journal of Economic
Behavior and Organization 14:299–329
18. Ding, Z., Granger, C., Engle, R. (1993) A long memory property of stock market
returns and a new model, Journal of Empirical Finance 1:83–106
19. Farmer, J. D., Joshi, S. (2002) The price dynamics of common trading strate-
gies, Journal of Economic Behavior and Organization 49:149–171
20. Gaunersdorfer, A. (2000) Endogenous fluctuations in a simple asset pricing
model with heterogeneous agents, Journal of Economic Dynamics and Control
24:799–831
21. Gaunersdorfer, A., Hommes, C., Wagener, F. (2003) Bifurcation routes to
volatility clustering under evolutionary learning, Technical Report 03-03, CeN-
DEF, University of Amsterdam, Working Paper
22. He, X. (2003) Asset pricing, volatility and market behaviour-a market fraction
approach, Technical Report 95, Quantitative Finance Research Center, Univer-
sity of Techonology, Sydney
23. Hommes, C. (2001) Financial markets as nonlinear adaptive evolutionary sys-
tems, Quantitative Finance 1:149–167
24. Hommes, C. (2002) Modeling the stylized facts in finance through simple non-
linear adaptive systems, Proceedings of National Academy of Science of the
United States of America 99:7221–7228
25. Kirman, A. (1992) Whom or what does the representative agent represent?,
Journal of Economic Perspectives 6:117–136
26. LeBaron, B., Arthur, W., Palmer, R. (1999) Time series properties of an artifical
stock market, Journal of Economic Dynamics and Control 23:1487–1516
27. Lux, T. (1995) Herd behaviour, bubbles and crashes, Economic Journal
105:881–896
28. Lux, T. (1997) Time variation of second moments from a noise trader/infection
model, Journal of Economic Dynamics and Control 22:1–38
29. Lux, T. (1998) The socio-economic dynamics of speculative markets: Interacting
agents, chaos, and the fat tails of return distributions, Journal of Economic
Behavior and Organization 33:143–165
30. Lux, T., Marchesi, M. (1999) Scaling and criticality in a stochastic multi-agent
model of a financial market, Nature 397(11):498–500
31. Manski, C., McFadden, D. (1981) Structural Analysis of Discrete Data with
Econometric Applications, MIT Press
32. Pagan, A. (1996) The econometrics of financial markets, Journal of Empirical
Finance 3:15–102
Deflationary Recessions in a General
Equilibrium Framework
Luca Colombo1 and Gerd Weinrich2

1
Università Cattolica, Istituto di Economia e Finanza, Largo Gemelli, I-20123
Milano. Tel. +39-02-7234.2637, fax +39-02-7234.2781
lucava.colombo@unicatt.it
2
Università Cattolica, Largo Gemelli 1, I-20123 Milano. Tel. +39-02-7234.2728,
fax +39-02-7234.2671 gerd.weinrich@unicatt.it.
Summary. This paper investigates the role of fiscal and monetary shocks in the
occurrence of deflationary recessions. Our model is based on a temporary equilibrium
approach with stochastic rationing, where inventory dynamics is explicitly taken
into account, amplifying spillover effects between markets. This setting allows us to
study the driving forces behind disequilibrium phenomena, and to investigate the
efficacy of alternative policies in overcoming them. In particular, we provide for an
application of our approach to the study of the Japanese deflationary recession.
JEL Classification: D45, D50, E32, E37.
Key words: Inventories, non-tâtonnement, price adjustment, deflationary reces-

sions
1 Introduction
The motivation for this paper has come from the observation that current
mainstream economic theory has difficulties to convincingly explain defla-
tionary recessions in a general-equilibrium framework. Since in a recession
many factors of production are underemployed, their prices should decrease.
In a standard general-equilibrium model these decreases would be assumed
to occur quickly. However, the recent experience of the Japanese economy as
well as that of the Great Depression indicate that a deflationary recession may
last over a long time. In fact, recession and deflation are a problem in Japan
since the mid-1990s.
To analyze deflationary recessions in a general-equilibrium framework we
choose a non-tâtonnement approach and build a macroeconomic model with
temporary equilibrium with rationing in each period and price adjustment
from period to period. This makes it possible to study the dynamic function-
ing of an economy in which disequilibrium phenomena like underemployment,
126 Luca Colombo and Gerd Weinrich
deflation and excess productive capacities/inventories are allowed to occur. In

addition we put special emphasis on the occurrence of inventories of produced
goods because they reinforce the importance of spillover effects between mar-
kets and therefore favour the persistence of aggregate-demand deficiencies.
We find that restrictive fiscal and monetary policies can cause a perma-
nent recession. Contrary to conventional wisdom, this is favored by sufficient
downward flexibility of the nominal wage. Vice versa, to best avoid or over-
come long-term unemployment, wages should be rigid downwards. We further
obtain that the stationary Walrasian equilibrium is unstable whereas there
exist recessionary Keynesian underemployment equilibria which are quasi-
stationary (i.e. with real variables constant but nominal variables deflating)
and locally stable. Finally, by using our framework to analyze the recent de-
flationary Japanese recession, we argue that a mix of fiscal and monetary
measures should help in overcoming it.
2 The Model
There are overlapping–generations consumers, producers and a government.

They exchange between them a composite consumption good, labor and
money. Time is divided in periods of equal length and prices are fixed at
the beginning of each period. At these prices trades take place, and the fact
that prices are typically not Walrasian implies that there is some rationing;
in particular there is the possibility of involuntary unemployment. To obtain
a feasible allocation we employ the concept of temporary equilibrium with
quantity rationing. Once goods have been exchanged, the economy moves to
the next period and new prices are quoted. At these prices a new allocation
obtains and so on.
The adjustment of prices at the beginning of a new period is linked to the
size of imbalance on the various markets that materialized in the equilibrium
with rationing of the previous period. In each period t these imbalances will
be measured by the ratios of aggregate demands and supplies
Ytd Ld
γt = s , λt = ts
Yt Lt
on the goods and labor market, respectively.

Consumers and firms hold expectations γte and λet of these values. In an
equilibrium with rationing these expectations will be correct. More precisely,
consumers are subject to stochastic rationing in case they are on the long
side of a market.3 That rationing will be such that the expected transaction
3
Stochastic rationing - as opposed to deterministic rationing - is compatible with
manipulability of the rationing mechanism and therefore provides an incentive for
rationed agents to express demands that exceed their expected trades, as argued
by Green (1980), Svensson (1980), Douglas Gale (1979, 1981) and Weinrich (1982,
Deflationary Recessions in a General Equilibrium Framework 127
is proportional to a consumer’s effective demand (or supply), with the fac-

tors of proportionality on each market being the rationing coefficients γt and
λt . Given also the price pt of the good, the nominal wage, wt , and previous
period’s net aggregate profits, (1 − tax)Πt−1 , that are distributed to young
consumers, the aggregate effective demand of young consumers, after resolv-
ing an optimization problem that takes account of stochastic rationing, is a
function4
Xt = X(γte , λet ; pt , wt , (1 − tax)Πt−1 ).
To this we have to add old consumers’ aggregate demand which is just Mt /pt
- the real value of the money stock held entirely by the old consumers -
and G which denotes real government spending. Thus, setting αt = wt /pt ,
mt = Mt /pt and πt = Πt−1 /pt , the aggregate demand on the goods market is
Ytd = X d (γte , λet ; αt , (1 − tax) πt )+mt +G ≡ Y d (γte , λet ; αt , mt , (1 − tax) πt , G).
Aggregate supply of labor is fixed and denoted by Ls .

Producers derive their effective labor demands dt = d (γte ; αt ) and pro-
duction decisions ytp = f (dt ) from maximizing expected profits subject to a
stochastic rationing mechanism, i.e.
max[min {γte , 1} (f () + st ) − αt ]

where st is the amount of stored goods carried over from the previous period.
In case of demand rationing on the labor market, only the share λt of firms’
aggregate labor demand will be realized. Taking account of this fact, aggre-
gating over all firms yields thus an aggregate labor demand and an aggregate
goods supply function
Ld (γte ; αt ) , Y s (γte , λet ; αt , St ) = Y p (γte , λet ; αt ) + St (1)
where St denotes aggregate inventories held by firms at the beginning of period

t.5
3 Equilibrium
Given αt , mt , (1 − tax) πt , St and G, a temporary equilibrium with rationing is
given by a fixpoint (γt∗ , λ∗t ) of the mapping
1984, 1988). This renders possible to have imbalances of aggregate demand and
supply in an equilibrium with rationing and use the size of these imbalances for
a determination of price adjustment. For a definition of manipulability see for
example Böhm (1989) or Weinrich (1988).
4
For the formal derivation of this function see Colombo and Weinrich (2003).
5
For the precise derivation of these functions see Colombo and Weinrich (2003).
Y d (γt , λt ; αt , mt , (1 − tax) πt , G) Ld (γt ; αt )

χ : (γt , λt ) → ( , )
Y s (γt , λt ; αt , St ) Ls
and by expectations (γte , λet ) = (γt∗ , λ∗t ). In that case all agents have perfect
foresight of the rationing coefficients which they reproduce by their optimal
actions. Thus no agent has an incentive to change something. In Colombo and
Weinrich (2003) we show that such an equilibrium exists and is unique, i.e.
the mapping
(αt , mt , (1 − tax) πt , St , G) → (γt∗ , λ∗t )
is well defined. This yields an aggregate allocation
!
Y t = min Y d (γt∗ , λ∗t ; αt , mt , (1 − tax) πt , G), Y s (γt∗ , λ∗t ; αt , St )
≡ Y (αt , πt , mt , St , G, tax)
!
Lt = min Ld (γt ; αt ) , Ls ≡ L (αt , πt , mt , St , G, tax) .
According to the whether γt∗ and λ∗t are larger or smaller than one there are
different types of equilibrium. More precisely we can summarize the various
possibilities as follows:
γt∗ λ∗t T ype T St U nemployment

< 1 < 1 Keynesian unemployment K + +
>1>1 Repressed inf lation I 0 0
> 1 < 1 Classical unemployment C 0 +
<1>1 U nderconsumption U + 0
Moreover there are intermediate cases the most important of which is the
Walrasian equilibrium which corresponds to γt∗ = λ∗t = 1.
4 Dynamics and Numerical Analysis
The model discussed above gives rise to a four dimensional non-linear dynamic
system, with state variables αt , mt , πt and St , which given by the sequence
∞
{(αt , mt , πt , St )}t=1 where
wt+1 /wt
αt+1 = αt ,
pt+1 /pt
1
mt+1 = [mt + G + (1 − tax) πt ] − πt+1 ,
pt+1 /pt
Πt /pt Y (αt , πt , mt , St ) − αt L (αt , πt , mt , St )
πt+1 = = ,
pt+1 /pt pt+1 /pt
St+1 = Y s (λ∗t , γt∗ ) − Y (αt , πt , mt , St ) + St .
Prices and wages are assumed to rise (fall) whenever an excess of demand
(supply) is observed, i.e.

pt+1 1 − µ1 (1 − γt∗ ) if γt∗ ≤ 1
= ; µ1 , µ2 ∈ [0, 1]
pt 1 − µ2 (1 − γt∗ ) if γt∗ ≥ 1
and
wt+1 1 − ν1 (1 − λ∗t ) if λ∗t ≤ 1
= ; ν1 , ν2 ∈ [0, 1] ,
wt 1 − ν2 (1 − λ∗t ) if λ∗t ≥ 1
where µ1 and µ2 (ν1 and ν2 ) denote downward and upward price (wage)
flexibility, respectively.
As there are four nondegenerate equilibrium regimes, the dynamic system
can be viewed as the union of four subsystems each of which may become ef-
fective through endogenous regime switching. Being of such a complex nature,
we cannot be study it by means of analytical tools only. Therefore we resort
to numerical simulations to investigate the dynamic behavior of the model.6
Note that, under suitable parameter specifications (see Colombo and Wein-
rich, 2003), as a benchmark a stationary Walrasian equilibrium is obtained
for
α∗ = 0.85, m∗ = 46.25, π ∗ = 15, S ∗ = 0, G∗ = 7.5, tax∗ = 0.5, (2)
with trading levels L∗ = Y ∗ = 100. Furthermore, in the following we set

µ1 = µ2 = ν2 = 0.1, whereas ν1 is varied between 0 and 0.1.
4.1 Fiscal and Monetary Shocks
We first consider a change in public expenditure G. Starting from the Wal-

rasian equilibrium, (α0 , m0 , π0 , S0 ) = (α∗ , m∗ , m∗ , S ∗ ), the bifurcation dia-
gram in Figure 1 shows, for ν1 = 0, the stationary values of employment to
which the system converges in dependence of values of G between 0 and 15.
From this it is evident that L < L∗ for G < G∗ and L = L∗ for G = G∗ .
When G < G∗ , aggregate demand Y d is diminished which creates an excess
supply on the goods market. Consequently firms reduce their production and
cut back on employment. The result is an excess supply on the labor mar-
ket, too, and the economy enters in a state of Keynesian unemployment. The
imbalance on the goods market gives rise to a price decrease whereas on the
labor market the nominal wage cannot decrease as ν1 = 0. As a result the real
wage increases. This is illustrated in Figure 2 which shows the time series for
employment L, inventories S, the real money stock m and the real wage α
for the first 200 periods where G = 7. The real wage is rising until approxi-
mately period 30 when it has become large enough so that the system enters
into a regime of Classical unemployment. Here the goods price decreases and
6
See Colombo and Weinrich (2003) for a more detailed analysis characterizing the
dynamic system associated to each regime.
the real wage falls until at around period 50 it settles at a stationary value
α > α∗ . Since the nominal wage rate does not change, the constant real wage
implies that the goods price does not change either beyond period 50, and the
economy has reached a stationary state at the frontier between Keynesian and
Classical unemployment. In that state there is market clearing on the goods
market but excess supply on the labor market.
Fig. 1. Stationary employment values when ν1 = 0.
Next consider, again for G < G∗ , what happens when ν1 > 0. The charts
in Figure 3 show, analogously to Figure 1, stationary values of employment
for various values of G. The two charts in the top row refer to downward wage
flexibilities of ν1 = 0.025 and ν1 = 0.1, respectively. The striking result is that
a little downward wage flexibility has an enormous effect on the impact of fiscal
restraint as is documented by the discontinuity of the graphs at G = G∗ . Note
that this does not happen in the model without inventories, as is shown by
the lower chart in Figure 3 where ν1 = 0.1 but St is exogenously set to zero
at the beginning of each period.
Why inventories have such a dramatic effect is easily explained. When
aggregate demand is diminished due to a decrease in G, inventories become
positive and rise further as excess supply on the goods market builds up. Since
St influences Y s positively (see Equation (1)) and γt = Y t / Y s (λt , γt ; αt , St )
by definition, an increase in St reduces γt which diminishes the labor demand
of firms, and thus increases further the excess supply on the labor market.
Therefore, the downward flexible wage rate decreases more than would be
the case without inventories. If the decrease in the wage rate is larger than
Fig. 2. Time series when ν1 = 0 and G = 7.
Fig. 3. Stationary employment values for ν1 = 0.025, ν1 = 0.1 (top charts), and
ν1 = 0.1 and St ≡ 0.
the decrease in the goods price, the real wage decreases, and it may con-
tinue to decrease permanently approaching a limit level below the Walrasian
real wage. The lower real wage diminishes labor income of workers which
diminishes aggregate goods demand which in turn keeps employment below

full employment. The dynamic system converges to a quasi-stationary Keyne-
sian state with permanent deflation of all nominal variables but constant real
magnitudes. The nominal money stock shrinks because, due to the small gov-
ernment spending, the government is permanently realizing a budget surplus.
These facts are illustrated in Figure 4 which shows time series for ν1 = 0.025.
Fig. 4. Time series when ν1 = 0.025 and G = 7.
It is easy to show that restrictive monetary shocks can have the same effects
as those associated to fiscal shocks investigated above. Consider for instance
a reduction in the initial money stock from m∗ to m0 = 40, keeping all other
parameters and initial values at their Walrasian levels. Since m0 is the demand
of old agents at time t = 0, aggregate demand is reduced. Consequently there
is excess supply on the goods market and, since firms adjust to the reduced
transaction level on the goods market, they reduce their labor demand. Thus
there is excess supply on the labor market, too, and the economy enters in a
state of Keynesian unemployment. What happens next depends on whether
the nominal wage is flexible downwards.7
As shown in Figure 5, where ν1 = 0.025, when downward wage flexibility
is allowed, employment, real wage and real money all converge to values lower
than the respective Walrasian values. The reason is similar to that already
7
With nominal wage rigid downwards, a restrictive money shock has a temporary
but not lasting effect on economic activity, as it is shown in Colombo and Weinrich
(2003)].
discussed in the context of fiscal shocks: the presence of inventories increases

the fall of labor demand by firms which in turn depresses labor income and
aggregate demand. The system tends to a quasi-stationary Keynesian state
with permanent deflation of nominal variables. Note that, in our framework,
money is not neutral in the long run: the restrictive monetary shock has caused
a permanent decrease in employment and output.8
Fig. 5. Time series when ν1 = 0.025 and m0 = 40.
At this point the natural question is which downward wage flexibility is

needed to drive the economy into a permanent recession or even depression.
The answer is given in the bifurcation diagram of Figure 6. From that it can be
seen that approximately until ν1 = 0.02 the economy is capable of returning
to the full employment after the monetary shock, whereas for speeds of wage
adjustment larger than this the economy gets trapped in underemployment.
5 Stability
The fact that a restrictive monetary shock may lead to a quasi-stationary
Keynesian unemployment state as limit of the dynamic system’s trajectory
8
As for fiscal shocks, this is not the case when St ≡ 0. In fact, the real wage
decreases initially but then the decrease in the goods price dominates that in the
nominal wage, and the real wage moves back to its Walrasian level, as do all the
other variables.
Fig. 6. Stationary employment values when m0 = 40.
raises the wider question of the stability of such a state. Analogously, the
stability of the stationary Walrasian state may be investigated. To anticipate
the answer, numerical simulations suggest that the latter is unstable whereas
the Keynesian unemployment state is locally stable.
Let us first look at the stationary Keynesian state. The limit values of
the state variables of the simulation shown in Figure 5, where we have an
initial reduction of the money stock to m0 = 40 at a downward wage flex-
ibility of ν1 = 0.025 and policy parameters G∗ = 7.5 and tax∗ = 0.5, are
(approximately)
α = 0.8281, m = 31.9263, π = 15.7889, S = 6.4060, (3)
with a stationary employment level L = 66.9342. We now take these stationary

values as new initial values, i.e. set (α0 , m0 , π0 , S0 ) = (α, m, π, S), and perform
a bifurcation analysis with respect to each of the state variables around these
initial values and dependent variable the employment level L. The results are
shown in Figure 7. The diagrams show that local deviations in directions of
any of the state variables do not change the fact that the system has this state
as an attracting long-run rest point. Although this is not a strict proof of local
stability, it is highly indicative for the behavior of the dynamic system.
Moreover, in order to return from the stationary recessionary state to
permanent full employment by changing only the money stock, an increase
in that variable to its Walrasian value (46.25) is not sufficient, but a higher
increase, to approximately 51, would be needed. The case is still worse for a
Fig. 7. Local stability of the Keynesian unemployment state.
unilateral variation of real profits which would have to be more than tripled,
to around 52, to get permanently out of unemployment. In both cases the
reason is intuitively clear: only a substantial increase in purchasing power of
consumers, old or young, can succeed to change the situation of insufficient
aggregate demand causing Keynesian unemployment. A detailed analysis of
the respective time series shows that, after an initial shock to m0 > 51 or
π0 > 52, the system enters in the regime of repressed inflation from where it
converges to the Walrasian equilibrium.
The panel of Figure 7 showing the bifurcation diagram over inventories
indicates that no change in the initial value of S is by itself sufficient to over-
come the recessionary state. Even a momentaneous decrease of inventories
to zero is not capable to lead the dynamic system out of the Keynesian un-
employment regime because the insufficient demand very quickly builds up
inventories again and restores the old situation.
A less immediate intuition is available for the effects of a change in the
initial value of the real wage α. An increase of its value from the stationary
state value typically results in a state of classical unemployment, because the
high real wage not only creates excess demand on the goods market but also
excess supply on the labor market. This implies a decrease in the nominal
wage, an increase in the goods price and a decrease in the real wage. This
process continues until the real wage is low enough again to make the system
enter into the Keynesian regime. There inventories are built up and the sys-
tem converges to the quasi-stationary Keynesian state. On the other hand,
a decrease of α0 from α typically leads the dynamic system to enter into
the regime of repressed inflation. From there it may return to the Keynesian
regime (for values of α above 0.425) or converge to the Walrasian state (for
α smaller than 0.425).
By performing bifurcation analysis analogous to the ones above, one can
also show that the Walrasian stationary equilibrium (α∗ , m∗ , π ∗ , S ∗ ) is lo-
cally unstable in three of the four directions defined by the state variables,
namely m, π and S. Only in direction α is the system locally stable. When-
ever m0 < m∗ , π0 < π ∗ or S0 > S ∗ , the dynamic system diverges from the
Walrasian state and converges to the quasi-stationary Keynesian unemploy-
ment state considered above.9 Note that the above results are obtained for
specific parameter values. One particularly significant parameter is the down-
ward wage flexibility ν1 . Here this value has been set to 0.025, but similar
results hold whenever ν1 is above the benchmark value seen in Figure 6 (ap-
proximately 0.018) that separates the stationary values of employment from
full to below full employment.
6 An Application to the Japanese Deflationary Recession

As shown in the previous sections, our framework permits to describe and
to discuss the outcome of alternative economic policies for an economy fac-
ing prolonged deflationary and recessionary situations. The behavior of the
Japanese economy since the Nineties, having been stuck in a quasi-stationary
Keynesian unemployment state for over a decade, provides for a clear example
and allows for natural experiments on the outcome of alternative economic
policies.
In particular, mimicking a proposal to overcome the Japanese recession
by Ben Bernanke (see The Economist, June 21st 2003, p.74), we focus here
on a balanced-budget tax reduction combined with a simultaneous increase
in the money stock. Provided it is of a large enough magnitude this policy
proves indeed a remedy to recession in the long run. Consider, for instance,
a fiscal-monetary policy combination calling for a reduction in the tax rate
(from tax∗ = 0.5 to tax = tax∗ − ∆tax), accompanied by a decrease in G by
∆G = ∆tax · π ∗ (from G∗ = tax∗ · π ∗ to G = tax · π ∗ = G∗ − ∆G) and an
increase in m0 by ∆m = ∆G. Starting from the quasi-stationary Keynesian
state illustrated in (3) and setting government spending more precisely to
G = tax · π ∗ , with π ∗ = 15 the Walrasian value of real profits so that the
government’s budget is balanced at any Walrasian equilibrium, one can show
that a decrease in the tax rate monotonically increase the long-run value of
employment, allowing the economy to reach the full employment stationary
Walrasian equilibrium.10
Note that there are two important ingredients of the policy outlined above.
On the one hand, there is a balanced-budget fiscal measure in the form of a
9
See Colombo and Weinrich (2003) for the bifurcation diagrams.
10
See Colombo and Weinrich (2003) for a more detailed analysis.
reduction in the tax rate, that should reduce both output and employment.
This textbook economics implication, however, impinges on the fact that
the balanced-budget multiplier refers to a static situation with prices and
wages fixed, whereas here we have a dynamic analysis with flexible prices,
wages and money stock. In our case, the fall in the employment is just a
short-run phenomenon. Nonetheless, it is an all but welcome occurrence.11
To avoid this, a simultaneous increase in the money stock m0 can be used.
Specifically, set m0 = m + G∗ − G, together with (α0 , π0 , S0 ) = (α, π, S). If
(G, tax) = (0, 0), then m0 = 31.9263 + 7.5 = 39.4263 and L1 = 73.8248. Sim-
ilarly, for (G, tax) = (2.55, 0.17), m0 = 36.8763 and L1 = 71.4806. In both
cases subsequent employment values increase monotonically to full employ-
ment. This shows that the combined measure of tax reduction and expansive
monetary policy works well in our model economy.
7 Concluding Remarks
Our framework has described consistent allocations in every period, even when
prices are not at their market clearing levels, obeying at the same time a well
defined dynamics. This has rendered possible to study, in a general equilib-
rium framework, the dynamic behavior of an economy in which disequilibrium
phenomena are allowed to occur. In particular, we have shown how, starting
from a Walrasian equilibrium, the economy may end up in a permanent re-
cession (i.e. in a locally stable deflationary quasi-stationary state with unem-
ployment) following fiscal or monetary shocks. Being able to characterize the
driving forces behind disequilibrium phenomena, our setting has allowed us to
investigate the performance of alternative policy measures aimed at solving
them. In this respect, we have provided for an application to the Japanese
deflationary recessions, testing the efficacy of a mix of fiscal and monetary
policies.
References
1. Böhm, V., 1989. Disequilibrium and Macroeconomics, New York, Basil Black-
well.
11
To better understand what is going on, consider the two extreme cases of
(G, tax) = (0, 0) and (G, tax) = (15, 1). Looking at the respective time series,
after one iteration we have L1 = 60.7438 in the first case and L1 = 73.1357
in the second one. Comparing with the initial level L = 66.9342 (stationary at
(G, tax) = (G∗ , tax∗ ) = (7.5, 0.5)) these values fit quite well with the textbook
prediction. Subsequently Lt+1 > Lt in the case (G, tax) = (0, 0) and Lt+1 < Lt
in the case (G, tax) = (15, 1). The first time employment is larger in the first case
than in the second is in period 8 when L8 = 65.9933 in case (G, tax) = (0, 0) and
L8 = 65.9735 in case (G, tax) = (15, 1).
2. Colombo L. and G. Weinrich, 2003. “Unemployment and Inventories in the

Business Cycle”, http://www.unicatt.it/docenti/weinrich/invent.pdf.
3. Gale, Douglas, 1979. “Large Economies with Trading Uncertainty”, The Review
of Economic Studies 46, 319-338. ¡
4. Gale, Douglas, 1981. “Large Economies with Trading Uncertainty: a Correc-
tion”, The Review of Economic Studies 48, 363-364.
5. Green, J., 1980. “On the Theory of Effective Demand”, Economic Journal 90,
341-353.
6. Svensson, L.E.O., “1980. Effective Demand and Stochastic Rationing”, The
Review of Economic Studies 47, 339-355.
7. Weinrich, G., 1982. “On the Theory of Effective Demand”, Economic Journal
92, 174-175.
8. Weinrich, G., 1984. “On the Theory of Effective Demand under Stochastic
Rationing”, Journal of Economic Theory 34, 95-115.
9. Weinrich, G., 1988. “On the Foundations of Stochastic
10. Non-Price Rationing and the Adjustment of prices”, Rivista di matematica per
le scienze economiche e sociali, 107-131.
Concepts of Thermodynamics in Economic
Growth
Jürgen Mimkes
Physics Department, University of Paderborn, Germany mimkes@physik.upb.de
Summary. The Solow-Swan model of economic growth is reviewed on the basis of

natural production. Natural growth is a biochemical process based on the laws of
thermodynamics. Economic production - like work in thermodynamics - is a non ex-
act differential. The production function Y (a, b) as a function of laborers (a) and (b)
depends on the path of integration. The production function may be calculated for
the special processes like constant mean capital per labor (T ), (which corresponds
to the Carnot process in thermodynamics): Y (a, b) = L{ak + bl + T (ln{aa bb })}.
The elasticity coefficients or exponents a, b with a + b = 1 are determined by the
production factors! The production function Y (a, b) has been applied to optimizing
production processes in farming and leads to a Boltzmann distribution of production
factors. The main source of economic growth is entropy, the chance of diversification,
the variety of know how and ideas. The results lead to a new model of economic
growth for interdependent economic systems like Japan and the US, East and West
Germany, North and South America, and agrees well with data for these economies.
Key words: production function, laws of thermodynamics, economic growth, en-

tropy of mixing. Classification codes: C1, E2, I5.
1 Introduction
In the last ten years new interdisciplinary approaches to economics have de-
veloped in social and natural science, especially by W. Weidlich 1972, D. K.
Foley 1994, J. Mimkes 1995, H. G. Stanley 1999, Y. Aruka 2001, and others.
In the present paper the mechanism of economic production is discussed on
the basis of natural production.
Why should concepts of thermodynamics be applied to economic growth? The
appropriate mathematics field for (infinitesimal) growth is calculus of differen-
tial forms. Calculus in one dimension is generally part of mathematical courses
in economics. With two production factors, calculus in two dimensions is the
proper mathematical tool for economic growth, and differs from calculus in
140 Jürgen Mimkes
one dimension: In one dimension there are only exact differentials that can be
integrated independently of the path of integration, in two dimensions there
are also non exact differential that depend on the path of integration. This is
familiar in natural sciences, but, unfortunately, not in general to students of
social and economic sciences. Thermodynamics is the prominent application
of calculus in two dimensions in natural science. Accordingly, we may expect
the concepts of thermodynamics to be valuable to economic growth.
The neoclassical growth theory is based on the Solow-Swan model (Barro,1995).
The production function Y = f (K, L) is determined by the production factors
capital (K) and labor (L). The difference of production (Y ) and consumption
(C) leads to savings, (S). In economic sciences the equilibrium of production
is generally calculated from an exact differential form:
∂f ∂f
dY = dK + dL = 0. (1)
∂K ∂L
The integral of dY in Eq. (1) leads to the production function Y : Solutions
for Y have been given by Gossen (2a) and Cobb Douglas (2b):
Y = C ln(La K b ) (2a)
Y = ALa K b (2b)
with the elasticity constants a + b + c = 1.
However, many questions are left open in the Solow Swan model, e. g.:
1. Does a general production function Y(K, L) in two dimensions always exist?
2. How are the elasticity constants defined?
3. Are production (Y ) and consumption (C) related, Y = Y (C) or C = C(Y )?
Agriculture like many other production processes relies on closed production
cycles. But, if the production of one cycle - like one year - was given by the
closed exact integral of Eq.(1), this would result in zero output:
" "
∂f ∂f
dY = ( dK + dL) = 0 (3)
∂K ∂L
The integral of an exact differential form (dY ) depends on the borders of the
integral, only. A closed cyclic process starts and ends at the same point and
the corresponding integral, Eq. (3), will be zero. This indicates that cyclic
economic production may indeed only be represented by a non exact differ-
ential form.
2 Non Exact Differential Forms

We now look into the calculus of non exact differential forms (Kaplan, 2003).
The integral depends on the border and on the path of integration. A simple
example of a non exact form is given by:
Concepts of Thermodynamics in Economic Growth 141
" "
δW = {a(K, L)dK + b(L, K)dL} = 0 (4a)
Non exact differential forms are marked by ”δ” and the non exact differ-
ential form δW of production, Eq.(4a) replaces Eqs.(1) and (3). The closed
integral of non exact forms will generally not be zero, the output of Eq.(4a)
will depend on the path of integration or the production process. Splitting the
closed path of Eq.(4a) into two parts, one (δW ) from A to B and the other
(δW ) from B to A, we obtain
" " B " A " B " B

δW = δW + δW = δW − δW = Y − C = S = 0 (4b)
A B A A
If we call the first path Y and the returning path C, the output of cyclic
production (4b) leads to non zero profit, which may be invested or saved (S).
The concept of non exact differential forms leads to the proper relation of
neoclassical theory, Y − C = S.
3 Law of Economic Production

According to the laws of calculus (Kaplan, 2003) a non exact form (δW ) may
be turned into an exact form (dσ) by an integrating factor (T ). Two equivalent
non exact forms δW and −T dσ may differ by an exact form (dK), which will
vanish at closed integration:
δW = dK − T dσ. (5)
In thermodynamics this is called the (first and second) law of work (W ). K

is called energy, T is called mean energy per particle or temperature, σ is the
entropy.
In economics we may call Eq.(5) the law of economic production, K is the
capital, T is the mean capital per person, a mean price level or the standard
of living. σ is again the entropy, which is related to probability (P ),
σ = ln P (6)
Probability is determined by combinatorial calculations, as will be discussed,

below. Entropy is not restricted to natural science, but is a function of math-
ematical statistics and has already been introduced to economics by Roegen
(Georgescu-Roegen, 1974). Eq.(5) is the answer to the first question of the
introduction: Does a general production function exist? The answer is: No!
The non exact differential form of production δW may not be integrated in
general, there is no general production function (W ). But for a given produc-
tion process, for a specific path of integration the production function (W )
may be calculated. This will be discussed in the next section.
142 Jürgen Mimkes
4 Production at Constant Price Level T

For one season a number of skilled laborers (L) are working on a farm at
constant wages and prices. At constant price level T = T0 we may integrate
δW in Eq.(5) from 0 to Y to obtain the income of one season:
Y = K − T0 σ (7)
L1 people are specialized in raising cattle, L2 people are specialized in pro-

ducing grain, L3 people are skilled in repairs. During a season each group of
laborers produces a certain amount of capital (K), which may differ for each
group due to different number of laborers (Li ) and different productivity (ki )
in each section:
K= Li ki = L (Li /L)ki (8)
The exact number of possibilities (P ) to put (L) people to work in (i) different
areas of production is given by the number of combinations,
L!
P = . (9)
L1 !L2 !...Ln !
Entropy (σ) is given by
L! Ln Ln
σ = ln P = ln = −L ln( ) (10)
L1 !L2 !...Ln ! L L
(The calculation has been performed by applying Stirling’s formula for
L!). The ratios Li /L in Eqs.(8) and (10) may be renamed by xi = Li /L with
Entropy as production function

0,40
0,35
Production function Y(x)
0,30
0,25
0,20
0,15
0,10
0,05
0,00
0,00 0,05 0,10 0,15 0,20 0,25
Production factor x
Fig. 1. The production function Y (x) = L{kx + T x ln x} according to Eq.(11).


xi = 1. The production output of a farm with L laborers after one season
at constant prices and wages, Eq.(7) now is:

Y (K, L) = L{ xi ki + T xi ln xi }. (11)
The production function (11) is plotted in fig. 1, a growing production
factor x leads to saturation of productivity.
The results are very close to Gossen’s equation (2a) and his law of saturation.
Apparently the Gossen equation corresponds to the entropy of production at
constant T , with only one small modification, which also answers the second
question in the introduction: the constants of elasticity, a, b, c are given by the
production factors xi .
5 Optimization of Production at Constant Price Level T

Production may be optimized by maximizing the production function (11):
1. How many laborers are needed in each field of production? This question
may be answered by the deriving Eq.(11) by x:
Automobile market in Germany 1998

8
(New cars)
Number of cars in mill. units
6
(used cars)
(Boltzmann)
10 20 30 40 50 60 70
Average price per car in k DEM
Fig. 2. Distribution of cars xi in four price classes (i) sold in Germany 1998, data
points according to Statistisches Bundesamt [2000]. Only the complete market (new
and used cars) follows a Boltzmann distribution.
144 Jürgen Mimkes
∂Y
= L{k1 + T (ln xi + 1)} = 0 (12a)
∂xi
ln xi + 1 = −k1 /T (12b)
−ki
xi = Ae T (12c)
The optimal number of laborers (12c) is a Boltzmann distribution and

depends exponentially on the productivity ki of the group (i) in relation to
the standard of living (T ).
2. The Boltzmann distribution has been investigated in a similar problem: How
many cars of four different price categories (Pi ) have been sold in Germany
in 1998 at constant standard of living T ? The answer is again a Boltzmann
distribution,
−Pi
xi = Ae T (12d)
Fig. 2 shows the automobile market in Germany in 1998. The distribution
of sold cars in the four price categories according to the official data follows
a Boltzmann distribution. Similar distributions have been found for wealth
[Willis, 2004], however, some authors have proposed other distributions [Levy,
1996].
6 The Mechanism of Economic Growth
Agriculture is a periodic process, and we have not finished the calculation of

the production cycle in chapter 4. According to Eq.(4b) only the closed cycle
will lead to income (Y ) and costs or consumption (C). Carnot has proposed
an ideal path of integration along constant values of T and constant values of
σ in the T − σ plane, fig. 3. The sign of δW depends on the direction of the
cycle, " "
δW = (dK − T dσ) = ∆T ∆σ = ∆Q = Y − C (13a)
∆Q is the difference between income (Y ) and costs/consumption (C). If the

profit ∆Q is saved, it is generally called savings, (∆Q = S).
A farm - like all production systems - has to gain more than it spends,
Y −C = ∆Q > 0, in order to survive. This is true for all biological, thermody-
namic, human and economic systems. It applies to plants, animals, to electric
generators, to people, workers, business men, to farms, companies, countries.
All these production systems are governed by the first and second law of ther-
modynamics (5) and the Carnot process. In the Carnot process - at constant
T - we may separate the two functions Y and C. The production function Y
of the Carnot process has already been calculated in Eq.(11), consumption C
is obtained, accordingly,
Carnot production cycle

8
6 Y
Integrating factor (T)
4 Profit Q
2 C
0
0 2 4 6 8 10
Entropy
Fig. 3. The Carnot process of a farm in a T − σ diagram. A farm is a generator of

capital and operates like a generator of (electrical) energy. Y : The farmer collects a
high amount of energy/capital (Y ) like grain, cattle etc. from the laborers. C: The
farmer pays a lower amount of energy/capital (C) for food, costs, wages etc. to the
laborers. The profit is ∆Q = Y − C. The efficiency (r) of the system is given by the
ratio of profit ∆Q to invested capital K, r = ∆Q/K.

Y (K, L) = K + T2 ∆σ = L{ xi ln xi } x1 ki + T2 (13b)

C(K, L) = K + T1 ∆σ = L{ x1 ki + Ti xi ln xi } (13c)
The functions Y and C have been calculated for a given process or path
of integration. This answers the third question stated in the introduction: Y
and C are both defined only by the specific production process.
7 Economic Growth of Interdependent Systems

Farms, companies, business firms contain at least two groups of people: farmer
and laborers, owner and workers, capital and labor, first and third world, Y
and C. Both groups together form the economic system. Accordingly, both
groups will discuss, how to divide the profit ∆Q = Y − C of each cycle. This
is negotiated periodically by workers and employers, by unions and industry,
by world trade conferences.
As a result of the negotiations the worker will obtain the percentage ”p” of
the profit, the employer will take the fraction (1 − p). If both groups reinvest
their shares p(Y2 − Y1 ) and (1 − p)(Y2 − Y1 ), they will grow in time (t) and
we obtain a set of two equations, C = Y1 (t) and Y = Y2 (t):
146 Jürgen Mimkes
dY1 (t) = p(Y2 − Y1 )dt (14a)

dY2 (t) = (1 − p)(Y2 − Y1 )dt (14b)
For p = 1
2 the solution of this set of differential equations is given by:
ept − 1
Y1 (t) = Y10 + p(Y20 − Y10 ) (15a)
1 − 2p
ept − 1
Y2 (t) = Y20 + (1 − p)(Y20 − Y10 ) (15b)
1 − 2p
1
For p = 2 the solution is given by
1
Y1 (t) = Y10 + (Y20 − Y10 )t (16a)
2
1
Y2 (t) = Y20 + (Y20 − Y10 )t (16b)
2
8 Results: Economic Growth of Interdependent

Countries
The equations (15) and (16) may be applied to all interdependent systems,
to workers and employers, unions and industry, or interdependent countries.
The results are compared to data of interdependent countries in figs. 4 to 11.
8.1 Growing Economies (0 < p ≤ 0, 5)

p = 0, fig. 4: If all profit goes to the richer party, the income (Y2 ) of group (2)
will grow exponentially, the income of the first party stays constant, (Y10 ).
The efficiency of the system, r = (Y2 − Y1 )/K grows with time.
p = 0, 10, fig. 4: at 10% of the profit for the poorer party (Y1 ) and 90% for
the rich party (Y2 ) both parties grow exponentially. Examples are Japan and
Germany after World War II, both economies were depending on the US and
were growing exponentially, as indicated in fig. 5.
p = 0, 25, fig. 4: at 75% for the rich party Y2 still grows exponentially, but
less than with 90% of the profit. Y1 gets 25% of the profit, but as it is linked
to a less growing Y2 the poorer party Y1 is growing less for longer times than
at 10% of the profit.
This indicates clearly, that a high rise in wages will weaken the economy
and lead to lower wages in the long run. Low increase of wages will lead
to more exponential growth for industry and workers. Workers as well as
their managers will have to be more patient with pay raises, like in Germany
or Japan after world war II, fig. 5. The same is now observed in the rising
economies in East Asia.
Two interdependent economies

7
Relative GDP growth 6
p=0.10
2 p=0.10
p=0.25
p=0.25
1 p=0.50
p=0.50
0
0 2 4 6 8 10
Time (t)
Fig. 4. The development of income of two interdependent economic systems starting

at Y1 = 1 and Y2 = 2. The profit for the poor side varies from p = 0, 10 to p = 0, 50.
In the long run the income of workers (Y1 ) will grow faster with lower pay raise p !!
Interdependent economies
10000
USA
GDP per capita in US $ (1985)
UK
8000
CH
D
J
6000 VRCh
4000
2000
0
1880 1900 1920 1940 1960
Time (t)
Fig. 5. Economic growth of US, UK, Switzerland, Japan, Germany, China between
1870 and 1990 (after Barro and iMartin, 1995). The victorious allies USA and UK
have grown exponentially. Japan and Germany only started to grow exponentially
after World War II by international trade at low wages. China was excluded and did
not take part in economic growth, at that time.
148 Jürgen Mimkes

3
Relative GDP growth (Y)
p=0.75
p=0.75
0
0 2 4 6 8 10
Time (t)
Fig. 6. The development of income of two interdependent economic systems stating

at Y1 = 1 and Y2 = 2. At high values of profit for the poor side, p = 0, 75, economic
growth is reaching a boundary with time.

40
38
GDP per capita in US $ (1995)
36
34
32
30
28
26
Jp
24 US
22
20
1990 1992 1994 1996 1998 2000 2002
Time (t)
Fig. 7. The development of mean income (GDP / person) of the USA and Japan in
quarters between 1980 and 2000. The efficiency of trade between the two countries
is declining with time [World Factbook, USA, 2004].

3,0
Relative GDP growth (Y) 2,5
2,0
1,5
1,0
p=1.25
p=1.25
0,5
0,0
0 2 4 6 8 10
Time (t)
Fig. 8. The development of standard of living of two interdependent economic

systems stating at Y1 = 1 and Y2 = 2. At high values of p for the poor side, p > 1
both economies will converge below Y2 = 2.

45
40
GDP per capita in k DEM
35
30
25
20 West Germany
East Germany
15
88 90 92 94 96 98
Time (t)
Fig. 9. Mean income in West and Germany between 1989 and 1998 due to produc-
tivity and capital transfer. In 1998 East Germany reached about 80% of the mean
income in West Germany (Fründ 2002).
150 Jürgen Mimkes

5
4
Relative GDP growth (Y)
p=-0.25
2 p=-0.25
0
0,0 0,5 1,0 1,5 2,0
Time (t)
Fig. 10. The development of mean income of two interdependent economic systems
stating at Y1 = 1 and Y2 = 2. p = −0, 25 means the poor side is exploited. The
income of the dominating party grows exponentially, the exploited party falters.

35
GDP per capita in k US $ (1985)
30
25
20
US
15 Argentina
10
5
97 98 99 00 01
Time (t)
Fig. 11. The development of mean income of USA and Argentina between 1997 and
2001. Due to the large amounts of capital transferred to the US, the mean income
in Argentina was reduced, considerably.[World Factbook, USA, 2004 : Ministerio de
Economı́a, Argentina, 2004].
p = 0, 50; fig.4: An even split between the two parties seems to be a fair
deal. But this ”fair” deal leads only to a linear growth of both parties. The
efficiency of the system, r = (Y2 − Y1 )/K stays constant with time.
8.2 Stagnation: USA - Japan (0, 5 < p ≤ 1)
p = 0, 75; fig. 6: A high factor p > 0, 5 leads to decreasing efficiency of the sys-
tem, r = (Y2 − Y1 )/K declines with time. Y2 and Y1 are reaching a boundary.
(Y1 ) is trailing a decreasing (Y2 ). After Japan and Germany have acquired
many production plants, the factor p has grown and the efficiency of the ex-
ports to the US started to decrease. The slowly growing economic levels (Y2 )
of Japan and USA (Y1 ) are trailing each other with low efficiency, fig. 7.
p = 1, 00; If all profit goes to the poor side, the income of the poor party soon
reaches the constant income of the rich party, the efficiency of the system is
zero!
8.3 Converging Economies, West and East Germany (p > 1)
p = 1, 25; fig. 8: If more than 100% of the profit goes to the poor party, (Y2 )
will decrease, and (Y1 ) will catch up with (Y2 ). This has been observed in the
relationship of West and East Germany after reunion in 1990, fig. 9.
8.4 Exploitation (p < 0)
p = −0, 25; fig. 10: If the poor side (Y1 ) makes only losses, p < 0, it and will
go bankrupt. The richer party (Y2 ) will grow exponentially.
This has been observed in Argentina in 2000, when many rich people trans-
ferred there assets to the banks in the US, see data in fig. 11. In the last 500
years many European countries have exploited their colonies. The European
countries have grown exponentially, the colonies faltered.
References
1. Aruka, Y, Ed.: Evolutionary Controversies in Economics. Springer, Tokyo
2001.
2. Barro R, and Sala i Martin, X,: Economic Growth. McGraw-Hill (1995).
3. CIA World Factbook: USA, 2004.
4. Foley, D. K.: A Statistical Equilibrium Theory of Markets. Journal of Eco-
nomic Theory Vol. 62, No. 2, April 1994.
5. Fowler, R. and Guggenheim, E. A.: Statistical Thermodynamics. Cam-
bridge University Press (1960).
6. Fründ, Th.: Diploma Thesis. Physics Department, Paderborn 2002.
7. Georgescu-Roegen, N.: The entropy law and the economic process. Cam-
bridge, Mass. Harvard Univ. Press, 1974.
152 Jürgen Mimkes
8. Statistisches Bundesamt: Statistisches Jahrbuch für das Vereinte Deutsch-

land. Wiesbaden (1991) and (2000).
9. Kaplan, W.: Advanced Calculus. Addison-Wesley (2003).
10. Levy, M. and Solomon, S.: Power Laws are Logarithmic Boltzmann Laws.
International Journal of Modern Physics C , Vol. 7, No. 4 (1996) 595.
11. Mimkes, J.: Society as a many-particle System. J. Thermal Analysis 60 (2000)
1055 - 1069.
12. Mimkes, J.: Binary Alloys as a Model for Multicultural Society . J. Thermal
Anal. 43 (1995) 521-537.
13. Ministerio de Economı́a y Obras y Servicios Públicos, Secretarı́a de
Programación Económica, Argentina: 2004.
14. Report Deutsches Institut für Wirtschaftsforschung, DIW, Berlin 1993.
15. Statistisches Bundesamt: Statistisches Jahrbuch 2000. Wiesbaden 2000.
16. Stanley, H. E., Amarala, L. A. N., Canning, D., Gopikrishnana, P.,
Leea Y. and Liua, Y.: Econophysics: Can physicists contribute to the science
of economics?. Physica A, 269 (1999) 156 - 169.
17. Weidlich, W.: Sociodynamics. Amsterdam : Harwood Acad. Publ., 2000.
18. Weidlich, W.: The use of statistical models in sociology. Collective Phenom-
ena 1, (1972) 51.
19. Willis, G., Mimkes, J.: Economics Working Paper Archive. WUSTL, 2004.
Firm Dynamics Simulation Using
Game-theoretic Stochastic Agents
Yuichi Ikeda1 , Osamu Kubo2 , and Yasuhiro Kobayashi3

1
Author to whom correspondence should be made. Hitachi Research Laboratory,
Hitachi Ltd. ,Omika 7-1-1, Hitachi, Ibaraki 319-1292, Japan
yikeda@gm.hrl.hitachi.co.jp
2
Hitachi Research Laboratory, Hitachi Ltd. ,Omika 7-1-1, Hitachi, Ibaraki
319-1292, Japan okubo@gm.hrl.hitachi.co.jp
3
Hitachi Research Laboratory, Hitachi Ltd. ,Omika 7-1-1, Hitachi, Ibaraki
319-1292, Japan kobayash@gm.hrl.hitachi.co.jp
Summary. Decision-making is a crucial task for the business planning of industrial

firms, in order to cope with uncertainties in the business environment. A method
of firm dynamics simulation, i.e. the game-theoretic stochastic agent, was devel-
oped by applying game theory to a stochastic agent model in order to analyze the
uncertain business environment. Each stochastic agent is described by a Langevin-
type equation with an additional term for rational decision-making. In this paper,
the dynamics of firms in computer related industries, which consist of three indus-
trial sectors, i.e. the large scale integrated circuit sector, the personal computer
sector, and the liquid crystal display sector, are simulated using the game-theoretic
stochastic agents model. Then, the importance of the herding behavior of firms is
demonstrated to reproduce the formation and collapse of the bubble in the Japanese
computer related industry markets during the late 90s.
Key words: [83.10.Pp] Particle dynamics, [02.50.Le] Decision theory and

game theory, [05.10.Gg] Stochastic analysis methods (Fokker-Planck, Langevin,
etc.), [05.10.Ln] Monte Carlo methods, [05.20.-y] Classical statistical mechan-
ics, [05.40.Jc] Brownian motion
1 Introduction
Decision-making is a crucial task for the business planning of industrial firms,
in order to cope with uncertainties in the business environment. The valua-
tion result using the net-present value for a given business plan is significantly
changed by three factors, i.e. discount rate, profit volatility, and business plan
feasibility [1]. Business planning and demand forecasting can seldom be esti-
mated reliably and it is almost hopeless to tackle the problem by standard
154 Yuichi Ikeda, Osamu Kubo, and Yasuhiro Kobayashi
economics and financial engineering. In a previous paper [2] applying game

theory to a stochastic agent model develops a method of firm dynamics sim-
ulation, i.e. the game-theoretic stochastic agent. It was shown that the model
can be reduced to the real option valuation method using approximations. A
decision-making process with an enhancement option and an abandonment
option was investigated using the real option version of the model. In this
paper, the dynamics of firms in computer related industries, which consist
of three industrial sectors, i.e. the large scale integrated circuit (LSI) sector,
the personal computer (PC) sector, and the liquid crystal display (LCD) sec-
tor, are simulated using the game-theoretic stochastic agent. In Section 2,
the model of the game-theoretic stochastic agent is described. Each stochastic
agent is described by a Langevin-type equation [3] [4] [5] with an additional
term for rational decision-making [6] and an interaction term between agents.
A transition probability for each vertex, which corresponds to a Nash equi-
librium, is evaluated using backwards induction, after calculating the payoff
for each strategy without volatility. In Section 3, the procedure of parameter
calibration is described in detail. In Section 4, the herding behavior of firms
is investigated and its characteristic features are discussed. The importance
of the herding behavior of firms is demonstrated to reproduce the formation
and collapse of the bubble in the computer related industry market in Japan
during the late 90s.
2 Game-theoretic Stochastic Agent Model

A group of firm agents, in which each agent is characterized by revenue
Ri (t)(i = 1, · · · , N ), is discussed using non-equilibrium statistical mechan-
ics. The revenue Ri (t) for the i-th agent is assumed to be described by the
Langevin-type equation,
dRi ∂U
= Di wik δ(t − tk ) − γi Ri − + σi ξi (t) + η(t)Rc(i) , (1)
dt ∂Ri
k
where γi , U , σi , and ξi are a friction coefficient, an interaction energy, volatil-

ity, and the Gaussian white noise, respectively. The rest of economic system
is considered by the stochastic term . The agent described by Eq. (1) is called
the game-theoretic stochastic agent.
Each agent behaves rationally to maximize its payoff P Vi . Rational decision-
making of the i-th agent is made by the first term of the RHS of Eq. (1).
Di wik = ±Di (k = 1, · · · , K) is the planned revenue of the i-th agent.
Di is the mean absolute difference of revenue, Di = |Ri (t) − Ri (t − ∆t)|.
wik = wik (P Vi ) is a transition probability at a decision-making point in time
k, and depends on the payoff P Vi of the i-th agent. The strategy u of the i-th
agent is ui = (wi1 , · · · , wik , · · · , wiK ) . The total number of strategies summed
Firm Dynamics Simulation Using Game-theoretic Stochastic Agents 155
over all agents is equal to 2N K . A payoff P Vi is the summed discounted cash

flow over l = t/∆t,

P Vi = (1 − T ) {Ri (l∆t) − Ci ((l∆t)}/(1 + r)l , (2)
l
where T is a tax rate, and the cost is assumed to be linear and quadratic
proportional to the revenue,
Ci (t) = αRi (t) + βRi (t)2 . (3)

The second term of the RHS of Eq. (3) corresponds to the fact that larger
firms are less efficient. A transition probability wik , which corresponds to a
Nash equilibrium, is evaluated using backward induction after calculating the
payoff (P V1 , · · · , P Vi , · · · , P VN ) for each strategy without volatility.
The third term of the RHS of Eq. (1) represents an interaction acting on
the i-th agent, which isproportional to the difference of the revenue from the
N
j=i Fi−j (Ri − Rj − Ri − Rj ), where Ri − Rj is the
∂U
other agents, ∂R i
=
average difference of revenue between the i-th agent and the j-th agent.
The last term of the RHS of Eq. (1) represents the irrational herding be-
havior [6]. η(t) and c(i) are the time-varying strength of the herding behavior
and the competitor of the i-th agent, respectively. The time-varying strength
of the herding behavior η(t) is assumed to be externally given in this model.
3 Parameter Calibration
Parameters are calibrated in order to reproduce the market data of the com-
puter related industries in Japan as a Nash equilibrium. The market data of
the computer related industries in Japan are shown in Fig. 1. Revenue for the
LSI, PC, and LCD sectors, shown in Fig. 1 (a), were divided into periods I
to III, according to the stock price of Yahoo Japan, shown in Fig.1 (b). Here,
Yahoo Japan was chosen as a representative of Japanese information technol-
ogy firms, though stock prices showed a similar pattern for many Japanese
industrial firms from 1999 to 2000. Period I which starts at Q3-97 and ends at
Q4-98, is named the normal period, because the stock price of Yahoo Japan is
relatively stable. Period II, which starts at Q1-99 and ends at Q1-00, is named
the bubble period. The stock price of Yahoo Japan rose sharply during 1999
and felt significantly in early 2000. Period III, which starts at Q2-00 and ends
at Q2-01, is named the post-bubble period.
It is regarded that the market data during Period I was formed as a conse-
quence of rational decision-making of firms. Calibration of the parameters was
made using the market data during Period I. The time-varying strength of the
herding behavior was assumed to be small enough to neglect, i.e. for Period
I. Rational behavior of each agent makes the trend of the market . Thus, the
2500
I II III
2000
Stock Price (103 Yen)
1500
1000
500
0
4Q-99
1Q-00
2Q-00
4Q-00
2Q-01
2Q-99
4Q-97
3Q-99
1Q-98
2Q-98
3Q-98
4Q-98
1Q-01
1Q-99
3Q-00
3Q-97
(b) Yahoo Japan
Fig. 1. The market data of the computer related industries in Japan is shown. Mar-
ket data for the LSI, PC, and LCD sectors, shown in Fig. 1 (a), were divided into
Periods I to III, according to the stock price of Yahoo Japan, shown in Fig.1 (b).
Here, Yahoo Japan was chosen as a representative of Japanese information tech-
nology firms, though stock prices showed similar price changes for many Japanese
industrial firms from 1999 to 2000.
first term of the RHS of Eq.(1) is replaced by the trend of the market µi for
Period I,

Di wik δ(t − tk ) → µi . (4)
k
Eq. (1) was applied to the parameter calibration for the three agents system,
RP C (t + ∆t) − RP C (t) = ξP C − γP C RP C (t)
(5)
+ FP C−LSI {RP C (t) − RLSI (t)} + FP C−LCD {RP C (t) − RLCD (t)},
RLSI (t + ∆t) − RLSI (t) = ξLSI − γLSI RLSI (t)

+ FLSI−P C {RLSI (t) − RP C (t)} + FLSI−LCD {RLSI (t) − RLCD (t)}, and
(6)
RLCD (t + ∆t) − RLCD (t) = ξLCD − γLCD RLCD (t)

(7)
+ FLCD−P C {RLCD (t) − RP C (t)} + FLCD−LSI {RLCD (t) − RLSI (t)},
where the first term of the RHS of Eq. (5) to (7) includes the trend of the
market µi and other terms,

N
ξi = µi + Fi−j Ri − Rj . (8)
j=i
With the following constraints on the interaction parameters ,
FP C−LSI = FLSI−P C , (9)
FP C−LCD = FLCD−P C , and (10)
FLSI−LCD = FLCD−LSI , (11)

the parameter calibration was made for the quarterly market data of the
period I by the least square fitting, and the results obtained are summarized
in Table 1.
Table 1. Model Parameters

i ξi γi Fi−j Fi−j σi Di
PC 199243 0.145198 -0.00294063 -0.546054 27357.7 34894.4
(j = LSI) (j = LCD)
LSI 207888 -0.0207392 -0.00294063 -0.345892 22701 59239.6
(j = PC) (j = LCD)
LCD -437974 -0.67084 -0.546054 -0.345892 33007.2 31380.0
(j = PC) (j = LSI)
The parameters α and β in Eq. (3) were determined in order to reproduce

the market data as a Nash equilibrium,
2
(m)
RP C (l∆t) − RP C (l∆t)
2
χ = { +
l
RP C
2 2
(m) (m)
RLSI (l∆t) − RLSI (l∆t) RLCD (l∆t) − RLCD (l∆t)
+ },
RLSI RLCD
(12)
(m)
where Ri (l∆t) was calculated revenue using the game-theoretic stochastic
agent model. The dependencies of χ2 on the parameters α and β are shown
in Fig. 2. Fitted curves are χ2 = 1 × 1015 β 2 − 9 × 107 β + 2.52(R2 = 0.7189)
and χ2 = 1178.5α2 − 2241.4α + 1066.1(R2 = 0.9998) for Fig.2 (a) and (b),
respectively. The optimal value for the parameters were chosen to be α = 0.96
and β = 4 × 10−8 .
1.8 1.6
1.6 =0.96 1.4 =4x10-8
1.4
1.2
1.2
1.0
1.0
2
2
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
0 2x10-8 4x10-8 6x10-8 8x10-8 0.9 0.92 0.94 0.96 0.98 1
(a) The dependence on the parameter (b) The dependence on the parameter
Fig. 2. The dependencies of χ2 on the parameters α and β are show. The parameters
α and β in Eq. (3) were determined in order to reproduce the market data as a Nash
equilibrium. Fitted curves are χ2 = 1 × 1015 β 2 − 9 × 107 β + 2.52(R2 = 0.7189) and
χ2 = 1178.5α2 − 2241.4α + 1066.1(R2 = 0.9998) for Fig.2 (a) and (b), respectively.
The optimal value for the parameters were chosen to be α = 0.96 and β = 4 × 10−8 .
Revenue for Period I was calculated using Eq. (5) to (7) with the obtained
parameters and ηi (t) = 0. The first terms of RHS of Eq. (5) to (7) were
replaced to incorporate rational decision-making,

ξi → ξi − Di /2 + Di wik δ(t − tk ), (13)
k
where the factor 2 of the second term of RHS is due to the absolute value in the
definition of Di . The initial values of revenue were taken at 3Q-97. A transi-
tion probability wik was evaluated using backward induction after calculating
the payoff (P V1 , · · · , P Vi , · · · , P VN ) for each strategy without volatility. The

calculated revenue for Period I is shown in Fig. 3. The agreement of the cal-
culated revenue and the market data of revenue are fairly good, even though
only three agents, i.e. the LSI agent, the PC agent, and the LCD agent, were
used to model the computer related industry.
Fig. 3. The calculated revenues for Period I using Eq. (5) to (7) with the obtained
parameters are shown. The first terms of RHS of Eq. (5) to (7) were replaced to
incorporate rational decision-making, as Eq. (13). The initial values of revenue were
taken at Q3-97. A transition probability was evaluated using backward induction
after calculating the payoff for each strategy without stochastic force. The agreement
of the calculated revenue and the market data of revenue are fairly good, even though
only three agents, i.e. the LSI agent, the PC agent, and the LCD agent, were used
to model the computer related industry.
4 Simulations
Revenue for Periods II and III was calculated using the six agents system,
i.e. the LSI1 agent, the LSI2 agent, the PC1 agent, the PC2 agent, the LCD1
agent, and the LCD2 agent,
RP C1 (t + ∆t) − RP C1 (t)

= pξP C − pDP C /2 + pDP C wP C1k δ(t − tk ) − γP C RP C1 (t)
k (14)
+ FP C−LSI {RP C1 (t) − RLSI1 (t) − RLSI2 (t)}
+ FP C−LCD {RP C1 (t) − RLCD1 (t) − RLCD2 (t)} + η(t)RP C2 (t), and
RP C2 (t + ∆t) − RP C2 (t)

= qξP C − qDP C /2 + qDP C wP C2k δ(t − tk ) − γP C RP C2 (t)
k (15)
+ FP C−LSI {RP C2 (t) − RLSI1 (t) − RLSI2 (t)}
+ FP C−LCD {RP C2 (t) − RLCD1 (t) − RLCD2 (t)} + η(t)RP C1 (t),
where p = RP C1 (0)/(RP C1 (0) + RP C2 (0)) and q = RP C2 (0)/(RP C1 (0) +

RP C2 (0)), and equations for the rest of agents are similar to the above. The
last terms in the RHS of Eq. (14) and (15) represent the herding behavior, and
those magnitudes are proportional to the competitors revenue. Magnitude of
the time-varying strength of the herding behavior is assumed in the following
calculation, because it was not obtained in the parameter fitting in the period
I.
The revenues for Periods II and III were calculated using the six agents sys-
tem without the herding behavior. The time-varying strength of the herding
behavior η(t) was given as follows, η(t) = 0.0 for the period II and η(t) = 0.0
for the period III. The parameters p = 0.4 and q = 0.6 were assumed. The
calculated revenues without the herding behavior for Periods II and III were
summed for each sector and are shown in Fig. 4. The thick lines indicate the
calculated revenues. The rise and fall of revenue observed in the market data
in Periods II and III were not reproduced. For the LSI sector and the PC sec-
tor, the calculated revenues were smaller than the market data during Period
II and the first half of Period III. For all sectors, the calculated revenues were
larger than the market data during the latter half of the period III.
Next, the revenues for Periods II and III were calculated using the six
agents system with the herding behavior. The time-varying strength of the
herding behavior η(t) was given as follows, η(t) = +0.1 for Period II and
η(t) = −0.1 for Period III. The parameters p = 0.4 and q = 0.6 were assumed.
The calculated revenues with the herding behavior for Periods II and III were
summed for each sector and are shown in Fig. 5. The thick lines indicate the
calculated revenues. The rise and fall of revenues observed in the market data
in Periods II and III were reproduced fairly well. Fig. 5 shows that the bubble
formation during Period II and the bubble collapse during Period III were
reproduced for all sectors. It is, however, noted that the calculated revenue
for the LCD sector was larger than the market data during Period III. The
importance of the herding behavior of firms was demonstrated to reproduce
the formation and collapse of the bubble in the computer related industry
market in Japan during the late 90s.
5 Conclusions
Decision-making is a crucial task for the business planning of industrial firms
in order to cope with uncertainties in the business environment. A method
of firm dynamics simulation, i.e. the game-theoretic stochastic agent, was
Fig. 4. The revenues for Periods II and III were calculated using the six agents
system without the herding behavior. The time-varying strength of the herding
behavior η(t) was given as follows, η(t) = 0.0 for the period II and η(t) = 0.0 for
the period III. The calculated revenues without the herding behavior for Periods II
and III were summed for each sector and are shown. The thick lines indicate the
calculated revenues. The rise and fall of the revenues observed in the market data
in the period II and III were not reproduced.
Fig. 5. The revenues for Periods II and III were calculated using the six agents
system with the herding behavior. The time-varying strength of the herding behav-
ior η(t) was given as follows, η(t) = +0.1 for Period II and η(t) = −0.1 for Period
III. The calculated revenues with the herding behavior for Periods II and III were
summed for each sectors and is shown. The thick lines indicate the calculated rev-
enues. The rise and fall of revenues observed in the market data in Periods II and
III were reproduced fairly well. The figure shows that the bubble formation during
Period II and the bubble collapse during Period III were reproduced for all sectors.
developed by applying game theory to a stochastic agent model in order to

analyze the uncertain business environment. Each stochastic agent is described
by a Langevin-type equation with an additional term for rational decision-
making.
The dynamics of firms in computer related industries, which consist of
three industrial sectors, i.e. the LSI, the PC, and the LCD sectors, are sim-
ulated using the game-theoretic stochastic agent model. The calculated rev-
enues with the herding behavior for the bubble period and the post-bubble
period were summed for each sectors and were compared with the market
data. The rise and fall of the revenues observed in the market data in the
bubble period and the post-bubble period were reproduced fairly well. The
importance of the herding behavior of firms was demonstrated to reproduce
the formation and collapse of the bubble in the computer related industry
market in Japan during the late 90s. An appropriate implementation of the
game-theoretic stochastic agent model will be valuable to analyze the uncer-
tain business environment and to provide feasible decision-making for business
planning.
References
1. Y. Ikeda, et al., A Simple Risk Model for Corporate Valuation, submitted to
Asia-Pacific Financial Markets.
2. Y. Ikeda et al., Forecast of Business Performance using an Agent-based Model
and Its Application to a Decision Tree Monte Carlo Business Valuation, Appli-
cations of Physics in Financial Analysis 4 (2003), accepted in Physica A
3. F. Schweitzer, Brownian Agents and Active Particles, Collective Dynamics in
the Natural and Social Sciences, Berlin, Springer-Verlag (2003)
4. P. Richmond, and L. Sabatelli, Langevin processes, agent models and socio-
economic systems, Physica A 336 (2004) 27-38
5. M. Levy, H. Levy, and S. Solomon, Microscopic Simulation of Financial Mar-
kets, From investor Behavior to Market Phenomena, San Diego, Academic Press
(2000)
6. R. Gibbons, Game Theory For Applied Economists, Princeton, NJ, Princeton
University Press (1992)
7. S. Bakhchandani, D. Hirshleifer, and I. Welch, A Theory of Fad, Fashion, Cus-
tom and Cultural Change as Information Cascades, Journal of Political Econ-
omy, 100 (1992) 992-1026
Part IV
Agent-based Modeling
A Functional Modularity Approach to
Agent-based Modeling of the Evolution of
Technology
Shu-Heng Chen1 and Bin-Tzong Chie2

1
AI-ECON Research Center, Department of Economics, National Chengchi
University, Taipei, 116, Taiwan chchen@nccu.edu.tw
2
AI-ECON Research Center, Department of Economics, National Chengchi
University, Taipei, 116, Taiwan chie@aiecon.org
Summary. No matter how commonly the term innovation has been used in eco-
nomics, a concrete analytical or computational model of innovation is not yet avail-
able. This paper argues that a breakthrough can be made with genetic programming,
and proposes a functional-modularity approach to an agent-based computational
economic model of innovation.
Key words: Agent-Based Computational Economics; Innovation; Functional

Modularity; Genetic Programming.
1 Motivation and Introduction
No matter how commonly the term “innovation” or “technological progress”

has been used in economics, or more generally, in the social sciences, a concrete
analytical or computational model of innovation is not yet available. Studies
addressing specific technology advancements in different scientific and engi-
neering fields are, of course, not lacking; however, the general representation
of technology, based on which innovation can be defined and its evolutionary
process studied, does not exist.
While direct modeling of innovation is difficult, economists’ dissatisfaction
with the neo-classical economic research paradigm is increasing, partially due
to its incompetence in terms of producing novelties (or the so-called emergent
property). We cannot assume in advance that we know all new goods and
new technology that will be invented in the future. Therefore, in our model,
we must leave space to anticipate the unexpected. Recently, [2, 3] introduced
Zabell’s notion of unanticipated knowledge to economists [11]. This notion is

NSC research grant No. 92-2415-H-004-005 is gratefully acknowledged.
166 Shu-Heng Chen and Bin-Tzong Chie
motivated by population genetics. In probability and statistics it is referred to

as the law of succession, i.e. how to specify the conditional probability that
the next sample is never seen, given available sets of observations up to now.
However, the Ewens-Pitman-Zabell induction method proposed by Aoki is
still rather limited. Basically, the nature of diversity of species and the nature
of human creativity should not be treated equally [5].
This paper proposes genetic programming as a possible approach leading to
simulating the evolution of technology. Our argument is based on two essential
standpoints. First of all, as regards the innovation process, we consider it
to be a continuous process (evolution), rather than a discontinuous process
(revolution). According to the continuity hypothesis, novel artifacts can only
arise from antecedent artifacts. Second, the evolution can be regarded as a
growing process by combining low-level building blocks or features to achieve
a certain kind of high-level functionality. In plain English, new ideas come
from the use (the combination) of the old ideas (building blocks). New ideas,
once invented, will become building blocks for other more advanced new ideas.
This feature, known as functional modularity, can be demonstrated by GP,
and that will be shown in this paper.
2 Background
The idea of functional modularity is not new to economists. For example, Paul
Romer has already mentioned that “Our physical world presents us with a
relatively small number of building blocks–the elements of the periodic table–
that can be arranged in an inconceivably large number of ways.” (Romer,
1998). That GP can deliver this feature has already been well evidenced in
a series of promising applications to the scientific, engineering, and financial
domains.
A decade ago, financial economists started to apply the functional-modular-
ity approach with GP to discover new trading rules. [9] and [1] took moving
average rules and trading range break-out rules as the building blocks (prim-
itives). GP was employed to grow new trading rules from these primitives.
Hence, GP has already demonstrated the evolution of trading technology:
combining low-level building blocks (MA, filter, or break-out rules) to achieve
a certain kind of high-level functionality (profitable performance).
John Koza’s application of GP to Kepler’s law is another striking example.
Here, not only did GP rediscover the law, but also, as the system climbed
up the fitness scale, one of its interim solutions corresponded to an earlier
conjecture by Kepler, published ten years before the great mathematician
finally perfected the equation [8, 4]. A further application of GP by John Koza
to analog circuits shows that GP-evolved solutions can actually compete with
human ingenuity: the results have closely matched ideas contrived by humans.
Koza’s GP has produced circuit designs that infringe 21 patents in all, and
duplicate the functionality of several others in novel ways [10].
Agent-based Modeling of the Evolution of Technology 167
Fig. 1. A Functional-Modularity Representation of Commodities.
Commodities are associated with their respective production processes which, when
written in LIST programming language, can be depicted as parse trees as shown
here.
3 Commodities and Production

Commodities in economic theory are essentially empty in terms of content.
Little attention has been paid to their size, shape, topology, and inner struc-
ture. A general representation of commodities simply does not exist in current
economic theory. In this paper, each commodity is associated with its produc-
tion process. Each production process is described by a sequence of processors
and the materials employed. In general, each sequence may be further divided
into many parallel subsequences. Different sequences (or subsequences) define
different commodities. The commodity with the associated processor itself is
also a processor whose output (i.e. the commodity) can be taken as a material
used by an even higher level of production. With this structure, we can ascer-
tain the two major elements of GP, namely, the function set and the terminal
set. The former naturally refers to a set of primitive processors, whereas the
latter refers to a set of raw materials. They are denoted respectively by the
following,
F unction Set : Ξ = {F1 , F2 , ..., Fk }, (1)

T erminal Set : Σ = {X1 , X2 , ..., Xκ }. (2)
Each sequence (commodity, processor) can then be represented by a LISP

S-expression or, simply, a parse tree (Fig. 1). The evolution of production
processes (commodities) can then be simulated by using standard GP. The
knowledge capital of the society at a point in time can then be measured by
the complexity and the diversity of its existing production processes.
4 Commodity Space
Before introducing the functional-modularity approach to preferences, let us

start with a brief review of the utility function used in conventional economic
theory. The utility function U (.) is generally a mapping from non-negative

real space to real space R.
U : Rn+ → R (3)
This above mapping is of little help to us when what we evaluate is a sequence
of processors rather than just a quantity. In our economy, what matters to
consumers is not the quantity they consumed, but the quality of what they
consumed. Therefore, the conventional commodity space Rn+ is replaced by
a new commodity space which is a collection of sequences of processors. We
shall call the space Y. The representation of the commodity space Y can be
constructed by using the theory of formal language, for example, the Backus-
Nauer form (BNF) of grammar. So Y is to be seen simply as the set of all
expressions which can be produced from a start symbol Λ under an application
of substitution rules (grammar) and a finite set of primitive processors (Σ) and
materials (Ξ). That is, Y represents the set of all commodities which can be
produced from the symbols Σ and Ξ.
Y = {Y | Λ ⇒ Y } (4)
While, as we saw in Fig. 1, each Y (Y ∈ Y) can be represented by the

language of expression trees (ETs), a more effective representation can be
established by using Gene Expression Programming (GEP) developed by [7].
In GEP, the individuals are encoded as linear strings of fixed length (the
genome or set of chromosomes) which are afterwards expressed as nonlinear
entities of different sizes and shapes, i.e. different expression trees. As [7]
showed, the interplay of chromosomes and expression trees in GEP implies an
unequivocal translation system for translating the language of chromosomes
into the language of ETs. By using GEP, the commodity space can then be
defined as a subset of the Kleene star, namely,
Y = {Yn | Yn ∈ (Σ ∪ Ξ)∗ ∩ GEP }, (5)
where Yn is a string of length n,
Yn = y1 y2 ...yn , yi ∈ (Σ ∪ Ξ), ∀i = 1, ..., n. (6)
We have to emphasize that, in order to satisfy the syntactic validity, Y is

only a subset of the Kleene star (Σ ∪ Ξ)∗ . To make this distinction, the Y
described in (5) is referred to as the strongly-typed Kleene star. Each Yn can
then be translated into the familiar parse tree by using GEP. This ends our
description of the commodity space.
5 Preferences
Unlike a commodity space, a preference space cannot be a collection of finite-
length strings, since they are not satisfied by the non-saturation assumption.

Fig. 2. Preference: The Parse-Tree Representation

What is shown here is only part of the potentially infinitely large parse tree, i.e.
only U l of [U l ].
Economic theory assumes that consumers always prefer more to less, i.e. the
marginal utility can never be negative. Even though we emphasize the quality
dimension instead of the quantity dimension, a similar vein should equally
hold: you will never do enough to satisfy any consumer. If consumers’ prefer-
ences are represented by finite-length strings, then, at a point, they may come
to a state of complete happiness, known as the bliss point in economic theory.
From there, no matter how hard the producers try to upgrade their existing
commodities, it is always impossible to make consumers feel happier. This is
certainly not consistent with our observation of human behavior. As a result,
the idea of a commodity space cannot be directly extended to a preference
space.
To satisfy the non-saturation assumption, a preference must be a string of
infinite length, something like
...u1 u2 ...ul ... = ...U l ... (7)
However, by introducing the symbol ∞, we can regain the finite-length rep-

resentation of the preference, i.e.
∞u1 u2 ...ul ∞ = ∞U l ∞ = [U l ]. (8)
First of all, as we mentioned earlier, consumers may not necessarily know what
their preferences look like, and may not even care to know. However, from
Samuelson’s revealed preference theory, we know that consumers’ preferences
implicitly exist. Equation (8) is just another way of saying that consumers’
preferences are implicit. It would be pointless to write down the consumers’
preferences of the 30th century, even though we may know that these are
much richer than what has been revealed today. To approximate the feedback

Fig. 3. Modular Preference: The LISP Representation
relation between technology advancements and preferences, it would be good

enough to work with local-in-time preferences (temporal preferences).
Secondly, Equation (8) enables us to see the possibility that preference
is adaptive, evolving and growing. What will appear in those ∞ portions
may crucially depend on the commodities available today, the commodities
consumed by the consumer before, the consumption habits of other consumers,
and other social, institutional and scientific considerations.
6 Utility Function
Given the preference [U l ], let U | [U l ] be the utility function derived from

[U l ]. U | [U l ] is a mapping from the strongly-typed Kleene Star to R+ .
U | [U l ] : Y → R+ . (9)
Hereafter, we shall simply use U instead of U | [U l ] as long as it causes no

confusion.
The modular approach to preference regards each preference as a hierarchy
of modular preferences. Each of these modular preferences is characterized by
a parse tree or the so-called building block. For example, the preference shown
in Fig. 2 can be decomposed into modular preferences of different depths. They
are all explicitly indicated in Fig. 3. Consider Si to be the set of all modular
preferences of depth i. Then Table 1 lists all modular preferences by means
of these Si . From both Fig. 3 and Table 1, it is clear that each subtree at a
lower level, say Sj , can always find its parent tree, of which it is a part, at a
higher level, say Si where i > j. This subsequence relation can be represented
as follows:
Si Sj . (10)
A commodity Yn is said to match a modular preference Si of U l if they
are exactly the same, i.e. they share the same the LISP expression and the
same tree representation. Now, we are ready to postulate the first regularity
condition regarding a well-behaved utility function, which is referred to as the
monotonicity condition.
Table 1. Modular Preferences Sorted by Depth

D (d) Subtrees or terminals
1 X2 , X3 , X5 , X8 , X9 , X11 1
2 S2,1 = (F7 X2 X3 ) 2
S2,2 = (F9 X5 X11 )
S2,3 = (F9 X3 X8 )
S2,4 = (F9 X5 X11 )
3 S3,1 = (F12 X3 (F9 X3 X8 )) 4
S3,2 = (F5 X3 (F9 X5 X11 ))
4 S4,1 = (F2 (F9 X5 X11 )(F12 X3 (F9 X3 X8 ))) 8
S4,2 = (F2 X3 (F5 X3 (F9 X5 X11 )))
5 S5 = (F6 X3 (F2 X3 (F5 X3 (F9 X5 X11 )))) 16
6 S6 = (F9 (F2 (F9 X5 X11 )(F12 X3 (F9 X3 32
X8 )))(F6 X3 (F2 X3 (F5 X3 (F9 X5 X11 )))))
7 S7 = (F2 (F7 X2 X3 )(F9 (F2 (F9 X5 X11 )(F12
X3 (F9 X3 X8 )))(F6 X3 (F2 X3 (F5 X3 64
(F9 X5 X11 ))))))
8 S8 = (F4 X3 (F2 (F7 X2 X3 )(F9 F2 (F9 X5 X11 ) 128
(F12 X3 (F9 X3 X8 )))(F6 X3 (F2 X3 (F5 X3
(F 9X5 X11 )))))))
Given a preference [U l ], the associated utility function is said to satisfy

the monotonicity condition iff
U (Yni ) > U (Ynj ) (11)
where Yni and Ynj are the commodities matching the corresponding modular
preferences Si and Sj of U l and Si and Sj satisfy Equation (10).
The monotonicity condition can be restated in a more general way. Given
a preference [U l ] and by letting {h1 , h2 , ...hj } be an increasing subsequence
of N+ , then the associated utility function is said to satisfy the monotonicity
condition iff
U (Ynj ) > U (Ynj−1 ) > ... > U (Yn2 ) > U (Yn1 ) (12)
where Yn1 , ..., Ynj are the commodities matching the corresponding modular
preferences Sh1 , ..., Shj of U l , and
Shi Shi−1 ... Sh2 Sh1 . (13)
If Sk is a subtree of Si as in Equation (10), then Sk is called the largest

subtree of Si if Sk is a branch (descendant) of Si . We shall use “Si Sk ” to
indicate this largest-member relation. Depending on the grammar which we
use, the largest subtree of Si may not be unique. For example, each modular
preference in Fig. 2 has two largest subtrees. In general, let Sh1 , Sh2 , ...Shj be
all the largest subtrees of Si , denoted as follows:
h
Si = hj1 Sk {Sh1 , Sh2 , ...Shj }, (14)
where {h1 , h2 , ...hj } is a non-decreasing subsequence of N+ . Notice these
largest trees may not have sub-relationships (10) among each other. How-
ever, they may have different depths, and the sequence {h1 , h2 , ...hj } ranks
them by depth in an ascending order so that Sh1 is the largest subtree with
minimum depth, and Shj is the one with maximum depth.
The second postulate of the well-behaved utility function is the property
known as synergy. Given a preference [U l ], the associated utility function is
said to satisfy the synergy condition iff

j
U (Yni ) ≥ U (Ynk ), (15)
k=1
where Yni and {Ynk ; k = 1, ..., j} are the commodities matching the corre-
sponding modular preferences Si and {Shk ; k = 1, ..., j} of [U l ], and Si and
{Shk ; k = 1, ..., j} satisfy Equation (14).
For convenience, we shall also use the notation jk=1 Ynk as the synergy of
the set of commodities {Ynk ; k = 1, ..., j}. Based on the New Oxford Dictionary
of English, synergy is defined as “the interaction or cooperation of two or more
organizations, substances, or other agents to produce a combined effect greater
than the sum of their separate effects.” “The whole is greater than the sum
of the parts” is the fundamental source for business value creation. Successful
business value creation depends on two things: modules and the platform to
combine these modules. Consider the consumer characterized by Fig. 2 as an
example. To satisfy him, what is needed are all of the modules listed in Table
1. Even though the technology has already advanced to the level S7 , knowing
the use of processor F4 to combine X3 and S7 can still satisfy the consumer
to a higher degree, and hence create a greater business value.
A modular preference may appear many times in a preference. For exam-
ple, S2,4 in Table 1 appears twice in Fig. 2. In this case, it can simultaneously
be the largest subtree of more than one modular preference. For example, S2,4
is the largest subtree of both S3,2 and S4,1 . Let Sk be the largest subtree of
Sh1 , Sh2 ,..., and Shj . Denote this relation as
Sk = j1 Shi {Sh1 , Sh2 , ...Shj }. (16)

Given a preference [U l ], the associated utility function is said to satisfy
the consistent condition iff
U (Yni | Sk Sh1 ) = ... = U (Yni | Sk Shj ), (17)
where Yni | Sk Sh1 is the commodity which matches the corresponding
modular preference Sk in the designated position, Sk Shi . The consistency
condition reiterates the synergy effect. No matter how intensively the com-
modity Yni may significantly contribute to the value creation of a synergy
commodity, its value will remain identical and lower when it is served alone.
Given a preference [U l ], the associated utility function U is said to be

well-behaved iff it satisfies the monotone, synergy and consistency condition.
It generates a sequence of numbers {U (Yni )}hi=1 where Yni matches the respec-
tive modular preference Sd,j . Sd,j is the jth modular preference with depth
d. The utility assigned in Table 1 is an illustration of a well-behaved utility
function derived from the preference shown in Fig. 2. In fact, this specific
utility function is generated by the following exponential function with base
2.
U (Sd,j ) = 2d−1 (18)
Utility function (18) sheds great light on the synergy effect. Thus, prim-
itive materials or rudimentary commodities may only satisfy the consumer
to a rather limited extent. However, once suitable processing or integration
takes place, their value can become increasingly large to the consumer. The
exponential function with base 2 simply shows how fast the utility may be
scaled up, and hence may provide a great potential incentive for producers to
innovate. Of course, to be a well-behaved utility function, U can have many
different functional forms.
7 Firms and Production
On the production side, the economy is composed of nf producers, each of

which is initially assigned an equal operating capital, K0 .
K1,0 = K2,0 = ... = Knf ,0 = K0 . (19)
With this initial capital, the producers are able to buy materials and pro-
cessors from the input markets up to the amount that they can afford. There
are two types of input markets at the initial stage, namely, the raw-material
market and the rudimentary processor market. For simplicity, we assume that
the supply curves of the two markets are infinitely elastic with a fixed unit
cost (c) for each raw material and for each rudimentary processor:
CX1 = CX2 = ... = CXκ = CF1 = CF2 = ... = CFk = c. (20)

With the materials and the rudimentary processors purchased from the
input market, the producer can produce a variety of commodities, defined by
the associated sequence of processors. The cost of each commodity is then
simply its total amount of materials and the number of processors, or, in
terms of GP, the node complexity of the parse tree. However, to allow for the
scale effect, each additional unit of the same commodity produced by the pro-
ducer should be less costly. This can be done by introducing a monotonically
decreasing function τ (q) (0 ≤ τ (q) ≤ 1), where q is the qth unit of the same
commodity produced. The cost of each additional unit produced is simply
the cost of the previous unit pre-multiplied by τ (q). With this description,
Fig. 4. EvolTech: Preferences Initialization
the capacity constraint for a fully-specialized producer i (i ∈ [1, ...nf ]), i.e. the
producer who supplies only one commodity, should be

q̄
K0 ≥ Cq , (21)
q=1
where Cq = τ (q)C1 is the unit cost of the qth unit and τ (1) = 1. For a fully-
diversified producer, i.e. the producer who produces a variety of commodities
and one for each, the capacity constraint is

m̄
K0 ≥ Cm,1 , (22)
m=1
where Cm,1 is the cost of the first unit of commodity m. In general, the
capacity constraint for the producer i is

m̄
q̄m
K0 ≥ Cm,q , (23)
m=1 q=1
where Cm,q = τm (q)Cm,1 .

In Equation (23), the strategic parameters are m̄, q̄ and Cm . To survive
well, producers have to learn how to optimize them. m̄ can be be taken as a
measure of the degree of diversification, whereas q̄ can be taken as the degree
of specialization. Cm , i.e. the node complexity of the commodity m, is also a
behavioral variable. Given the capacity constraint, the producer can choose
to supply a large amount of primitive commodities (a quantity-oriented strat-
egy), or a limited amount of highly delicate commodities (a quality-oriented
strategy). Therefore, the choice of Cm can be regarded as a choice of the level
of quality.
8 Demonstration
The idea presented above has been written into a computer program called
EvolTech, which stands for “Evolution of Technology.” In this section, we
Fig. 5. EvolTech: Generation 0
demonstrate a vanilla version of EvolTech. What we mean by vanilla will

become clear as we give the demo.
First, as in all computational economic models, we start with a simple
description of initialization. The initialization in EvolTech includes the gener-
ation of preferences. Based on the formulation given in Sections 5 and 6, the
preferences of three consumers are randomly generated, as shown in Figure
4. The complexity of the preference has been severely restricted to a depth
of 5 and is fixed throughout the entire evolution.3 Notice that these prefer-
ences are characterized by colorful nodes. Each different color is refereed to a
different primitive, sampled from a given primitive set.
In addition to the three consumers, there are three producers in the econ-
omy. Based on the same given primitive set, commodities are also randomly
generated by these producers, as shown in Figure 5. Notice that the three
dimensions of production behavior, i.e. quantity, quality, and diversity, are all
randomly determined as long as they together satisfy the capacity constraint
(23). The initial capital capacity K0 is set to 25, and the unit cost c is set to
0.5. The scale effect is ignored. Each of the commodities is then served to the
consumers whose preferences are displayed in Figure 4.
Without the details about how trade actually proceeds, it is not easy to
describe how the final price and hence the profit is determined. Therefore, in
this vanilla version of EvolTech we simply take the highest reservation price
as the market price.4
3
In other words, we do not consider the general preferences as discussed in Equation
(8), neither the evolution of preferences in this vanilla version.
4
We have proposed an algorithm to compute Equation (9) for a well-behaved utility
function U , as defined in Section 6. This algorithm, called the module-matching
This simplification will facilitate our calculation of profits, π. In Figure 5,

π is shown on the top of each commodity. We can see that most commodities
have negative profits. This is not surprising because at this initial stage all
commodities are randomly designed, and the chance of meeting consumers’
needs to any significant degree is naturally low with the given combinatoric
complexity.
Sophisticated commodities may be even worse than those simple designs
because they induce higher production costs, and can only satisfy consumers
to a very limited degree. So, as evidenced in Figure 5, commodities that suf-
fer great economic losses tend to be those with sophisticated designs, i.e.
trees with a large degree of node complexity. By summing the profits over all
commodities, we get aggregate profits for each firm, which are shown on the
right-hand side of the figure. In this specific case, all three firms make a loss.
This finishes our description of the initial generation.
While moving to the next market period (next generation), firms start to
learn from experience. Their profit profiles provide them with fundamental
clues on how to re-design their products for the next generation. What is
shown in Figure 6 is the result of their adaptation. It is interesting to notice
that these new-generation commodities seem to become simpler compared
with those of the previous generation (Figure 5). This is mainly because so-
phisticated designs do not contribute to profits but losses. Therefore, firms
tend to replace those sophisticated designs with simpler ones. The economy
as a whole can be described as a quantity-based economy, since all firms choose
algorithm, is very intuitive. It looks for the projection of the commodity Yi to [U l ],
i.e. a measure of distance between a commodity Yi and the preference [U l ]. Once
the operational meaning of Equation (9) becomes clear, it is possible to infer the
reservation price which a consumer would like to pay for commodity Yi from [U l ].
to produce a large number of rudimentary commodities (i.e. they repeat doing

simple standard things).
However, this strategy turns out to work well. While each simple commod-
ity can earn a firm a tiny profit, summing them together is still quite notice-
able. So, in the end, the profits of all three firms improve quite significantly.
This process is then further reinforced, and in the coming generations, more
resources are devoted to rudimentary commodities. Sophisticated designs are
almost entirely given up. However, since there are not too many rudimentary
commodities to develop in the market, when all firms concentrate on produc-
ing rudimentary commodities, the limited number of rudimentary-commodity
markets become highly competitive, and the profits from producing these
commodities decline as a result of the keen competition. At this point, the
economy actually moves toward an era of zero profit.
Once producing primitive commodities is no longer profitable, the selection
bias towards it also becomes weaker. Some sophisticated designs occasionally
coming out of the crossover and mutation operators may find it easier to sur-
vive. That improves the chances of satisfying consumers to a higher degree.
When that indeed happens, not only do firms make a breakthrough by suc-
cessfully having a sophisticated (delicate) design, but the lucrative profits also
attract more resources that can be devoted to quality products. While this
does not always happen and the process is not always smooth, the process
may be reinforcing. So, commodities with more delicate designs and higher
profits may come one after the other. In the end, the economy is gradually
transformed into a quality-based economy, as shown in the 10th generation of

our simulation (Figure 7).5
9 Concluding Remarks and Future Work

In this paper, commodities, production and preference, those fundamentals of
economic theory, have been re-formulated in light of functional modularity.
We believe that this re-formulation work is original and productive. It lays the
foundation upon which one can build and simulate the evolution of technology,
more specifically, within the context of agent-based computational economic
(ACE) models. A full picture of this ACE model has not been presented in
this paper, partially due to the limitations of size imposed on the paper. We,
therefore, can only give a sketch of some other essential ingredients, and leave
a more detailed account to a separate paper.
References
1. Allen, F. and R. Karjalainen, (1999). “Using Genetic Algorithms to Find Tech-
nical Trading Rules,” Journal of Financial Economics, 51(2) 245–271.
2. Aoki, M., (2002a). “Open Models of Share Markets with Two Dominant Types
of Participants,” Journal of Economic Behavior and Organization, 49 199–216.
3. Aoki, M., (2002b). “Applications of Ewens–Pitman–Zabell Inductive Methods
in New Economic Dynamics,” Proceedings of the Sixth International Conference
on Complex Systems, 29–35.
4. Banville, J., (1993). Kepler: A Novel (Vintage International). Vintage Books.
5. Basalla, G., (1988). The Evolution of Technology. Cambridge University Press.
6. Chen, S.-H. and B.-T. Chie, (2004) “Functional Modularity in the Test Bed
of Economic Theory– Using Genetic Programming,” in R. Poli, et al. (eds.),
GECCO-2004: Proceedings of the Genetic and Evolutionary Computation Con-
ference, June 26-30, 2004, Seattle, Washington, USA.
7. Ferreira, C., (2001). “Gene Expression Programming: A New Adaptive Algo-
rithm for Solving Problems,” Complex Systems, 13 (2), 87-129.
8. Levy, S., (1992). Artificial Life: A Report from the Frontier Where Computers
Meet Biology, Vintage, New York.
9. Neely, C., P. Weller, and R. Dittmar, (1997). “Is Technical Analysis in the For-
eign Exchange Market Profitable? A Genetic Programming Approach,” Journal
of Financial and Quantitative Analysis, 32(4) 405–426.
10. Willihnganz, A., (1999). “Software that Writes Software,” salon.com, WWW
Article.
11. Zabell, S., (1992). “Predicting the Unpredictable,” Synthese, 90 205–232.
5
The progress may not be smooth. Severe fluctuations can happen. The progress
may not be sustained long enough either. The economy may stagnate after a short
but fast take-off, and consumers are only supplied with some “basic needs.” For
a more detailed demonstration of the complex variety, see Chen and Chie (2004).
Herding Without Following the Herd:
The Dynamics of Case-Based Decisions with
Local Interactions
Andreas Krause
School of Management, University of Bath, Bath BA2 7AY, Great Britain

mnsak@bath.ac.uk
Summary. We use case-based decision theory to evaluate the repeated choices of

individuals who are using the experience of a selected set of other individuals over a
given time horizon as the basis for their own decisions. It is observed that for certain
parameter constellations a large fraction of individuals make identical decisions over
a long period of time, which is not implied by the information. The result of this
behavior is herding, which, however, has its origin in the way individuals process
information rather than a desire to be part of the herd and imitating the behavior
of others.
Key words: Case-based decision theory, local interactions, networks, decision-

making, herding, fads
The most common approach to model decision-making under uncertainty

in economic theory is to choose the alternative that provides the highest ex-
pected utility of all alternatives available to an individual. Numerous empir-
ical works however show that expected utility theories cannot satisfactorily
explain actual behavior in many cases, see e.g. [4, 5, 21] for an overview. A
particular challenge is the modelling of decisions where information on possi-
ble outcomes is limited, e.g. by not knowing all possible outcomes and/or their
probabilities of occurrence. In a series of papers [8, 9, 10, 11, 12, 13, 14, 15]
an alternative to the expected utility framework has been proposed, called
case-based decision theory (CBDT), a special case of the theories on bounded
rationality. Although some applications of CBDT have been proposed in the
literature, see section 4 for a brief overview, dynamical aspects of the so deter-
mined decisions have thus far not received much attention. Only [3] considers
the learning process in CBDT, but does not focus on the aggregate dynamics
of decisions.
Another common observation, particularly in financial markets, is that
individuals show a clear tendency towards making similar decisions, although
an objective evaluation of the information available does not support the
rationality of such a behavior. This so called herding behavior has received
180 Andreas Krause
more attention in recent years and several models have been developed to
explain this observation, see e.g. [23, 24] for models of imitating others or
[1, 2, 6] for models of informational cascades. A common feature of many
models is that they are essentially based on psychological factors and use the
idea that individuals copy the behavior of other individuals out of a desire
to conform with the majority or specific individuals in their peer group. The
obvious result of these models is that individuals neglect information on which
to base their decisions, but rather ”follow the herd”.
In this paper we use CBDT to model the decisions of individuals and show
how they can give rise to a large fraction of individuals making similar de-
cisions, which is not consistent with an objective evaluation of all available
information. Another property is that under fairly general conditions these
decisions are repeated over time, giving rise to prolonged periods of large
fractions of individuals making identical decisions. The main feature of this
model is that the origins of this observation is not the desire to reproduce the
behavior of other individuals but the way information is processed by individ-
uals. This establishes another reason for herding, which in our model is fully
rational. It is well beyond the scope of this paper to provide a full investiga-
tion into the detailed properties of our model, we rather restrict ourselves to
highlight some key properties which merit further investigation.
In the coming section we continue with a very brief introduction to CBDT
and the specific model used. Section 2 then presents the results of the com-
puter simulations conducted to evaluate this model while section 3 provides
explanations for these observations. We discuss some potential applications in
section 4 and finally conclude our findings in section 5.
1 The Decision-Making Process

CBDT is founded on David Hume’s reasoning that similar causes are likely to
produce similar effects. Hence an individual can use the experience from past
decisions, which he or other individuals made to evaluate the current decision
problem. A key to this decision-making problem is the similarity between the
current problem and the past problems he uses. If we let Ξ denote the set of
all possible problems, we define the similarity of two problems as a function
s : Ξ × Ξ → [0; 1], (1)

where for any i, j ∈ Ξ we interpret s (i, j) = 0 as the absence of any similarity
and s (i, j) = 1 as an identical problem; obviously it is s (i, i) = 1. We can
now normalize this similarity to the ”average” similarity in order to avoid any
problems from the scaling of this function later on:
#
s (i,j)
P
s (i,j )ωτ (i,j ) if ∈Ξ s (i, j )ωτ (i, j ) > 0
sτ (i, j) = j ∈Ξ j , (2)
0 if j ∈Ξ s (i, j )ωτ (i, j ) = 0
Herding without Following the Herd 181
where ωτ (i, j) denotes an indicator function, which is one if the decision made
when facing problem j at time τ was the equivalent to the one considered
currently, and zero otherwise. This indicator function thus ensures that only
those outcomes are considered that result from equivalent decisions, e.g. the
decision to buy a product. The outcome of any decision, rτ,i ∈ Θ, at time τ
for problem i ∈ Ξ can be evaluated in the common way by using a utility
function:
u : Θ → . (3)
Let us further assume that when considering past decisions the similarity
of problems reduces by a factor of λ ∈ [0; 1] per time period. This reduction
in similarity can, for example, be the result of a changing environment. With
a total memory of M time periods, we obtain the following function which
can be used to evaluate the decisions of an individual:
⎛ ⎞

M
V (rt,i ) = ⎝λτ sτ (i, j )u(rτ,j )ωτ (i, j )⎠ . (4)
τ =0 j ∈Ξ
Here again the indicator function ωτ (i, j) ensures that we only consider
decisions with equivalent decisions to address the present problem. It is now
possible to compare the values of this function for all rt,i ∈ Θ and choose the
decisions which provides the highest score. In cases where two decisions have
an identical score, the choice is made randomly with all such choices having
equal probability of being selected. This decision-making rule as presented
here is a marginal but in essence unimportant variation of the CBDT devel-
oped in [8, 12, 14], which also provide an axiomatic justification for using this
decision rule.
We will in this paper make a small re-interpretation of this model to suit
our needs. Let us assume that a set of N individuals, called ”agents”, have to
make a repeated decision which can take one of two forms, ”buy” or ”sell”,
hence Θ = {buy; sell}. The outcome of an individual decision is either ”suc-
cess”, corresponding to an outcome of +1, or ”failure”, corresponding to an
outcome of −1, such that = {−1; +1}. Rather than interpreting Ξ as a set
of decision problems, we assume it to be a set of agents, who all have to make
this decision. The similarity function s then merely indicates how similar the
circumstances of these agents making their decisions are.
Let us propose that an agent evaluates the past M decisions of himself
and selected other agents; these agents are selected randomly and do not
change over time. Once an agent i uses the experience of an agent j they are
assigned a similarity of s (i, j) ∈ [0; 1] and if he is not using his experience
we set s (i, j) = 0. We can arrange the s (i, j) in a matrix and interpret
it as the adjacency matrix of a graph which represents a network of agents
interacting with each other in their decision-making. It has to be noticed that
we use a directed graph as in general s (i, j) = s (j, i). We here assume that
182 Andreas Krause
connections between any two nodes of this graph exist with a probability of
p and assign a random similarity taken from a uniform distribution on [0; 1]
to each edge. We now determine the outcome of any decision randomly in
each time period. Hence u is a random variable and we assume that for any
given decision the ex-ante probabilities of success and failure are identical to
one half. The outcomes are serially independent, but we allow for a common
correlation of ρ across agents.
We use this model to evaluate the decision-making process in a simple
Monte-Carlo simulation using a wide range of parameters. The focus of our
attention will be on the properties of the aggregate decision of agents to buy
or sell, in particular the fraction of agents buying.
2 Model Evaluation
From the structure of the above model it is obvious that any agent knowing its
set-up would make random decisions. Thus a situation in which all N agents
make random choices, buying or selling with equal probability, serves as our
benchmark to which we will compare the results of our model.
In order to evaluate the model presented above, we conducted a Monte-
Carlo simulation with all parameters taken from independently and identically
distributed uniform distributions across the following ranges: M ∈ [0; 1000],
λ ∈ [0; 1], ρ ∈ [0; 1], and p ∈ [0; 1]. The number of agents remains fixed
at N = 100; as it is a common feature in most network models that its
properties do not depend on the number of agents but only on the term
N p, we can reasonably infer that the number of agents has no significant
influence on our results. The number of simulations conducted is 1, 000. Each
simulation consists of 11, 000 simulated time periods of which the first 1, 000
are discarded from the further analysis in order to eliminate any effects from
initial conditions. These initial conditions consist only of the memory which
we assume to be empty, i.e. it only consists of zeros.
In addition to the parameters as used in the simulation we also found it
useful to determine the ”average memory”, which we define similarly to the
duration as follows:
1−λ τ
M
D =1+ λ τ. (5)
λ τ =1
We investigate the properties of the fraction of agents buying using the
simulated time series in dependence of the parameters above. The following
variables of the time series and how they are affected by these parameters are
investigated:
1. Volatility (VOLA): average absolute change in the fraction of agents buy-

ing,
2. Volatility upon switch of majority (SWITCH): average absolute change

in the fraction of buyers, given that buyers gain or loose the majority,
3. Deviation from benchmark (BIAS): average absolute deviation of the frac-
tion of buyers from 0.5,
4. Length of deviation from benchmark (LENGTH): average number of time
periods between a switch from the fraction of buyers exceeding (below) to
below (exceeding) 0.5,
5. Autocorrelation (AUTO): first-order autocorrelation of the fraction of
agents buying.
We also determined the appropriate values for our benchmark case of all
agents making random choices as a comparison. It became apparent that the
results from our model are significantly different from that of a random choice.
Secondly, the often strong correlations between the investigated variables sug-
gest a common origin of these relationships, which obviously have to be found
in the parameters.
Some obvious relationships between the variables are easily explained. A
higher serial correlation of the fraction of buyers (AUTO) causes the volatility
(VOLA) to reduce, as subsequent changes are becoming smaller. The same
argument can be used to show that the length of a majority of buyers or
sellers being present (LENGTH) increases with the autocorrelations as the
lower volatility makes it less common for the majority to change once it has
been established. The larger the deviation from the benchmark of 50% buyers
on average is (BIAS), the larger the change in this fraction has to become
in order to cause the majority to switch (SWITCH), as this large deviation
has first to be overcome. If this were to be achieved in many small steps the
average deviation would be reduced, accordingly the volatility of the fraction
of buyers has to be higher. As with many large changes in the fraction of buyers
(VOLA) a majority can easily change, we observe an inverse relationship with
the length of any majority (LENGTH). Finally, as changes in the fraction of
buyers causing the majority to switch (SWITCH) forms part of the volatility
(VOLA), it is only natural to observe a positive relationship between these
two variables.
Of more importance than the relationship between these variables, how-
ever, is how they relate to the parameters. This information could then be
used to determine which constellations are necessary to explain certain effects
in decision-making. We have plotted the five variables against the two most
important parameters for this variable in figure 1.
The correlation of outcomes can be shown in a more detailed analysis to
have a statistically significant negative influence on the volatility (VOLA)
and a positive influence on the length of a deviation from the benchmark
(LENGTH). These influences are however of much less importance than of
those parameters presented in figure 1. The length of the memory itself, M ,
does not influence the variables significantly, but of course shows its effect via
the average memory D, which is an important parameter to consider.
184 Andreas Krause
Fig. 1. Relationships between variables and parameters

We observe that the volatility of the fraction of buyers (VOLA) is increas-

ing in the probability of interaction between two agents (p) and decreasing in
the average memory (D). It is not possible to determine a single parameter as
being dominant. In contrast to this, the volatility upon a switch of the major-
ity (SWITCH) is dominated by the positive relationship with the probability
of interaction (p) with the other parameters not showing an important influ-
ence. A similar effect can be observed on the deviation from the benchmark
(BIAS), which is effectively determined only by this probability.
A very different picture emerges when investigating the length of a devia-
tion from the benchmark (LENGTH). The probability of interaction (p) has
only a minimal influence, but it is instead dominated by the average memory
(D). The same situation is reflected in the autocorrelation of the fraction of
buyers (AUTO), although the influence of the probability of interaction is
larger in this case. We should have expected such a similarity given the very
strong relationship between the two variables as described above.
When conducting an OLS regression of all parameters on the variables,
our above findings are confirmed, all coefficients are highly significant and
the goodness of fit is reasonably high. A more detailed analysis of time se-
ries properties shows the best parameter constellation to generate the typical
herding behavior (a large and prolonged deviation from the benchmark) is
achieved by a moderate interaction of agents (p ≈ 0.5) and a long memory
(D > 10). A higher interaction causes the deviation to become too short and
volatile while a shorter memory reduces the length of the deviation too much
to be reasonably be described as herding.
Having established these findings, we now have to turn our attention to
their explanation, which we address in the coming section.
3 Explaining the Findings
We can explain the above findings quite easily by investigating the effects the
individual parameters have on the variables. We firstly can easily deduct that
with a higher probability of interaction (p) we will on average observe more
actual interactions, hence the decision-making is influenced by the experience
of more agents, which is aggregated. As all agents use more and more similar
experiences given the higher interactions, their decisions are likely to become
more alike; hence any changes of choice are to be conducted by a larger num-
ber of agents simultaneously, causing the volatility (VOLA) to increase and
naturally also the volatility upon a switch of majority (SWITCH). A direct
consequence of the more homogeneous decisions being made is that the de-
viation from the benchmark (BIAS) increases as more agents behave alike.
The increased volatility, on the other hand, makes such changes more likely
and the autocorrelation of the fraction of buyers (AUTO) decreases. For the
same reason also the length of any deviation from the benchmark (LENGTH)
reduces with the increased volatility making such changes more frequent.
186 Andreas Krause
Fig. 2. The effect of memory and interaction on herding
A longer average memory (D) has the obvious consequence that the deci-
sion is based on a longer history and thus the relative importance of recent
events is reduced. As the history does not change over time, decisions are based
on very similar considerations and should therefore not vary that much. With
subsequent decisions of agents being very similar, we see immediately that
the volatility (VOLA) reduces, the length of deviations (LENGTH) increase
as any deviation is unlikely to be reversed quickly and the autocorrelations
(AUTO) are increasing. On the other hand the volatility upon a switch of
the majority (SWITCH) is hardly affected because whether the change in
the choice of the agents is determined predominantly by recent experiences
or experiences in the distant past is of no relevance. Similarly the size of the
deviation from the benchmark (BIAS) is not affected.
We can summarize our findings in figure 2, where we use the term ”herd-
ing” to indicate the deviation from the benchmark (BIAS), ”widespread”
herding is to be interpreted as having a large BIAS and ”isolated” herding
as having a low BIAS; ”short” herding has a low LENGTH and ”prolonged”
herding has a large LENGTH.
We would normally expect a higher probability of interaction (p) and
longer memory (D) to improve the decisions made by agents as they can
base their choice on a broader range of experiences. This should result in a
reduced volatility (VOLA) and deviation from the benchmark (BIAS) and we
should particularly see an improvement over time due to the learning effect in
this stationary environment. Furthermore, the learning should be facilitated
by the simplicity of the choices and outcomes. But in the data simulated we
were not able to detect any learning at all, later choices showed exactly the
same properties as earlier choices.
The reason for this counterintuitive observation can be found in the dif-
ferences between what agents should learn to achieve the benchmark outcome
of random choices and what they actually learn from case-based decision-
making. The Bayesian view is that they should use the experiences to update
their beliefs on the distribution of outcomes for any given choice, thereby
learning the true distribution. In contrast to this, CBDT does not propose
agents learning the true distribution of outcomes, but rather the best choice
directly without reference to the underlying distribution which is determined
by the structure of the economy.
With this discrepancy it becomes apparent why we do not observe any
learning in our model – there does not exist a single best choice. Both alter-
natives are of equal value, an aspect which the CBDT cannot detect given its
decision-making rules. Although Bayesian learning is clearly superior to the
learning implied by CBDT we see that the deviation from the benchmark of
all agents making random choices does neither reduce the welfare of any indi-
vidual agent nor the social welfare; but it also does not increase the welfare.
This is because whatever decision is taken, the ex-ante outcome is always a
success and failure with equal probability, even in cases where all agents make
the same choice and outcomes are perfectly correlated.
It remains an open question how the behavior would change with the
existence of a superior choice. [3] provides first insights into the learning in
such a situation, but does not address the evolution of aggregate decisions over
time. At least in our model without the existence of a best choice, Bayesian
learning is not welfare improving and thus not conducted by individuals. This
absence of learning enabled the properties we investigated above to emerge
and to be maintained over time.
4 Potential Applications
Although CBDT has been developed about a decade ago, with predecessors
being proposed well before that, applications to real economic problems are
quite limited. Most of the literature on CBDT focuses on axiomatic and other
theoretical aspects. Among those isolated applications in the literature are
the works on financial markets [18, 19, 20], voting [17], production theory
[22], and consumer theory [7, 25]. [3] considers explicitly the learning effects
in CBDT using a similar model to ours with specific interaction structures
and a single superior choice. He shows the convergence towards equilibrium
under certain conditions.
The results we obtained in this paper have many potential applications
to explain patterns of behavior that are generally not consistent with the ex-
pected utility hypothesis. As we have seen that with a sufficiently long memory
and high interactions a significant majority of agents make a particular choice
188 Andreas Krause
over a longer period of time, we can use our model to explain various types
of herding, fads and hypes.
Consumers preferring one brand over another for no obvious reason, hence
there is no objectively superior choice, is a situation which is very similar to
the outcome of our model in this paper. We could therefore easily explain
fashion trends by consumers making decisions in line with CBDT as outlined
above. As we furthermore in reality also observe that such trends are reversed
as quickly as they emerged, we have a very good fit of our model with reality.
Of course our model can then be applied to all kind of short-lived fads and
hypes, like the sudden demand for certain computer games, fashion accessories
or Christmas toys, who are usually very quickly reversing.
These short-lived fashions lasting often less than a single season are consis-
tent with a relatively short memory of our model, which causes quick changes.
Proposing a short memory for such fashions seems reasonable given the sig-
nificant changes that occur over time, resulting in a low similarity over time
(λ). Similarly, these fashions are often not that widespread in a society, which
is consistent with a low overall interaction of agents. With fashions usually
spreading in smaller peer groups, e.g. a particular age group, this seems to
fit with the requirements of our model to produce such properties of these
fashions.
We can easily extend the type of choices not only to cover consumer deci-
sions, but also corporate decisions. Examples of fads in the past include the
decisions to merge into conglomerates and the more recent trend for spin-
offs in order to focus on the core business. We may in the same way look at
investments into specific regions, like South America or South-East Asia, or
into certain industrial sectors during particular periods in history, or trends
in production technologies like Just-in Time production, to name but a few.
Given the importance of such decisions for a company, we would expect
those decisions to be based on as much experience as possible, thus corre-
sponding to a high interaction of agents. Furthermore, by adjusting carefully
for any details that may reduce the ability to learn from more distant expe-
riences, we would require agents to have a long memory. Our model would
thus suggest a widespread and long-lasting herding of agents. This proposal
is consistent with observations from such trends in companies, which can last
for years and are widespread.
Another application is primarily relevant in financial markets, but to a
lesser degree also in other areas: the explanation of herding with the emergence
and burst of speculative bubbles. It is common to explain herding behavior of
individuals with psychological factors and a desire by individuals to conform
with the majority or a specific peer group. In essence these models of herding
are in many cases models of imitating other agents, see e.g. [23, 24]. Another
group of models uses informational cascades to obtain similar results, see [1]
for one of the first of such models or [2, 6] for an overview. In our model we
also observe herding as individuals make identical decisions over prolonged
periods of time, which are not based on objective information, but the source
of this behavior is very different. Agents are not trying to imitate the behavior
of other agents but rather seek to learn from their experience by evaluating
their decisions. We thus observe herding not as the result of imitation, but as
the result of information evaluation. In this sense we could argue that agents
are not following the herd, although their behavior resembles that of herding.
The properties of this herding can vary substantially. We can observe herd-
ing for many time periods as e.g. in the recent internet boom, but it can also
be a much shorter event, e.g. in bear market rallies or corrections in a rising
market. In some cases herding is very widespread affecting most investors,
while at other time it is restricted to a small group of investors, e.g. specialist
fund managers. We can replicate these observations by varying the memory
and interactions between agents in our model accordingly; usually the pa-
rameter constellations required are consistent with our understanding of the
underlying reason in financial markets.
Using these considerations we can summarize by stating that our model
would be applicable to situations that involve repeated choices by a large
number of individuals who are in interaction with each other.
There are strong epistemological reasons for supporting CBDT over ex-
pected utility theory as a positive theory of decision-making. If we accept that
agents are applying CBDT in the form presented in this paper, we have found
an alternative explanation of herding, which is not based on imitation but the
way decisions are made by agents and further research in the decision-making
process may be able to decide how well CBDT actually describes this process
and thus our model explains herding behavior.
5 Conclusions
We have presented a model in which agents base their decisions on CBDT,

which uses the past experience of a selected number of other agents over a
given memory length. We used a simple stationary environment in which re-
peated binary choices had to be made by all agents and investigated the aggre-
gate choices of all individuals. We found that agents are not learning the true
structure of the economy and their aggregate choice can deviate substantially
from the benchmark of random choices. We found that these deviations were
primarily influenced by the degree of interaction between agents and their
memory. It was pointed out that the aggregate choices resembled herding be-
havior under certain parameter constellations and that this herding behavior
was not the result of agents imitating each other’s behavior but arise from
the way decisions are made in CBDT, thus agents are not blindly following
the herd. The parameter constellations causing a behavior resembling herding
well used usually a long memory and medium interactions between agents. We
also pointed out that this model might be suitable to explain the emergence
of a multitude of fads like fashions, besides herding.
190 Andreas Krause
The scope of this paper is very limited and the results can only be a
first step into fully understanding its implications. As we allowed for a very
general network structure, it would be very interesting to investigate in as
far the network structure affects the results, indications from [3] show that
they may have an impact on the outcomes. It is also of importance to fully
understand the reasons behind the usually relatively quick change from a
majority of buyers to a majority of sellers. Another obvious strand of future
research would be a much more detailed investigation into the properties this
model, which were outside the scope of this paper. Apart from these more
theoretical considerations we face the need to apply it to specific economic
problems in order to evaluate its capability to explain actual events.
References
1. Abhijit V. Banerjee. A simple model of herd behavior. Quarterly Journal of
Economics, 107:797–818, 1992.
2. Sushil Bikhchandani, David Hirshleifer, and Ivo Welch. Learning from the
behavior of others: Conformity, fads, and informational cascades. Journal of
Economic Perspectives, 12:151–170, 1998.
3. Mattias Blonski. Social learning with case-based decisions. Journal of Economic
Behavior & Organization, 38:59–77, 1999.
4. Colin F. Camerer and M. Weber. Recent developments in modelling preferences
under risk. OR Spektrum, 9:129–151, 1987.
5. Colin F. Camerer and M. Weber. Recent developments in modelling preferences:
Uncertainty and ambiguity. Journal of Risk and Uncertainty, 5:325–370, 1992.
6. Andrea Devenow and Ivo Welch. Rational herding in financial economics. Eu-
ropean Economic Review, 40:603–615, 1996.
7. Itzhak Gilboa and Amit Pazgal. Cumulative discrete choice. Marketing Letters,
12:119–130, 2001.
8. Itzhak Gilboa and David Schmeidler. Case-based decision theory. Quarterly
Journal of Economics, 110:605–639, 1995.
9. Itzhak Gilboa and David Schmeidler. Case-based optimization. Games and
Economic Behavior, 15:1–26, 1996.
10. Itzhak Gilboa and David Schmeidler. Act similarity in case-based decision
theory. Economic Theory, 9:46–61, 1997.
11. Itzhak Gilboa and David Schmeidler. Cumulative utility consumer theory. In-
ternational Economic Review, 38:737–761, 1997.
12. Itzhak Gilboa and David Schmeidler. An overview of case-based decision theory.
In Liuigi Luini, editor, Uncertain Decisions: Bridging Theory and Experiments,
pages 215–235. Kluwer, 1999.
13. Itzhak Gilboa and David Schmeidler. Case-based knowledge and induction.
IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and
Humans, 30:85–95, 2000.
14. Itzhak Gilboa and David Schmeidler. A Theory of Case-Based Decisions. Cam-
bridge University Press, Cambridge, Great Britain, 2001.
15. Itzhak Gilboa and David Schmeidler. Inductive inference: An axiomatic ap-
proach. Econometrica, 71:1–26, 2003.
16. Itzhak Gilboa, David Schmeidler, and Peter P. Wakker. Utility in case-based
decision theory. Journal of Economic Theory, 105:483–502, 2002.
17. Itzhak Gilboa and Nicolas Vieille. Majority vote following a debate. Social
Choice and Welfare, 23:115–125, 2004.
18. Ani Guerdjikova. Case-based decision theory - an application for the financial
markets. Mimeo, University of Heidelberg, 2002.
19. Ani Guerdjikova. Asset pricing in an overlapping generations model with case-
based decision markers. Mimeo, University of Heidelberg, 2003.
20. Ani Guerdjikova. On the definition and existence of an equilibrium in a financial
market with case-based decision makers. Mimeo, University of Heidelberg, 2003.
21. David Harless and Colin F. Camerer. The predictive utility of generalized
expected utility theories. Econometrica, 62:1251–1290, 1994.
22. Hermann Jahnke, Anne Chwolka, and Dirk Simons. Coordinating demand and
capacity by adaptive decision making - an application of the case-based decision
theory. Mimeo, University of Bielefeld, 2003.
23. Andreas Krause. Rational choice, social structure and ’animal spirits’. Mimeo,
University of Bath, 2003.
24. Andreas Krause. Herding behavior of financial analysts: A model of self-
organized criticality. In Mauro Galegatti, Alan Kirman, and Matteo Marsili,
editors, The Complex Dynamics of Economic Interaction: Essays in Economics
and Econophysics, pages 257–268. Springer Verlag, Berlin, 2004.
25. Robert Meyer, Tulin Erdem, Itzhak Gilboa, Wesley Hutchinson, Steven Lipp-
man, Carl Mela, Amit Pazgal, Drazen Prelec, and Joelm Steckel. Dynamic
influences on individual choice behavior. Marketing Letters, 8:349–360, 1997.
Cultural Evolution in a Population of
Heterogeneous Agents
Gábor Fáth1,2 and Miklos Sarvary2

1
Research Institute for Solid State Physics and Optics - P. O. Box 49, H-1525
Budapest, Hungary fath@szfki.hu
2
INSEAD - Boulevard de Constance, 77305, Fontainebleau, France
miklos.sarvary@insead.edu
Summary. A general theory of cultural evolution is formulated using a cognitive

dimension reduction scheme. Rational but cognitively limited agents iteratively in-
vent and redefine abstract concepts in order to best represent their natural and
social environment. These concepts are used for decision making and determine the
agents’ overall behavior. The collection of concepts an agent uses constitutes his/her
cultural profile. As the importance of social interactions increase and/or agents be-
come more intelligent we find a series of dynamical phase transitions by which the
coherence of concepts advances in the society. Our model explains the so-called
“cultural explosion” in human evolution 50,000 years ago as a spontaneous ordering
phenomenon of the individual mental representations.
1 Introduction
Culture is the sum of knowledge, beliefs, values and behavioral patterns
built up by a group of human beings and transmitted from one generation
to the next. Its evolution is investigated in many disciplines using different
paradigms. Economics uses game theory methods to understand the develop-
ment of rules, social norms and other cultural institutions assuming rational
behavior of the constituting individuals (agents) [1]. Cognitive and behavioral
sciences study how culture is represented in the mind [2], and how these mental
representations (concepts, memes, heuristic decision rules, etc.) change under
social interactions [3]. Physics views culture and its evolution as a complex,
dynamical system, and aims to identify its basic universal properties using
simple minimal models [4, 5, 6, 7].
The emergence of culture seems to involve a chicken-and-egg problem: cul-
ture becomes what the constituting agents make it to be, and agents, based on
their capacity of learning, constantly adapt to the culture they happen to live
in. In this sense culture is the dynamical attractor of the complex social dy-
namics, determined by the agent’s actual constraints and social interactions.
194 Gábor Fáth and Miklos Sarvary
The aim of this paper is to investigate the properties of such cultural at-
tractors in an adequately formulated social dynamics model, which integrates
the relevant concepts from the social sciences with those from the physics
perspective.
One of the popular minimal models of culture is that of Axelrod [4, 5],
which takes a stochastic imitation process the basic motif of social interac-
tions, and investigates a population of biologically and economically uniform
agents whose cultural profile is initially random. The interesting question is
under what circumstances global cultural diversity gets preserved despite the
homogenizing interaction [6, 7]. However, the question can also be posed from
the opposite direction: in a population with innate heterogeneity and under
social interactions which are not necessarily imitational but reflect selfish in-
dividual interests how can a coherent culture emerge?
In order to study this question, our model differs from that of Axelrod.
Instead of simply assuming imitative interactions, we focus on individual de-
cision making mechanisms and identify dimension reduction as the governing
heuristics for cognitively limited but otherwise rational agents. Obviously, an
agent’s cultural profile manifests itself in decision making situations where
he has to choose from a set of alternatives. Different biological preferences
and cultural background give rise to different conceivable alternatives and
different evaluation of these. People with unequal views of the world and di-
verse convictions about what is right or wrong, true or false, likely or unlikely,
important or negligible are biased to make different choices.
In this paper we consider agents who are rational but cognitively limited
(bounded rationality [1, 8]). They are constrained to describe and understand
their world along a finite number of abstract concepts (ACs). These ACs con-
stitute the agent’s cultural profile and establish an interface to the objective
world: agents use ACs to build an adequate mental representation of the world,
which they use to evaluate decision alternatives. Obviously, decision making
is successful if evaluation is precise, thus rational agents face the problem of
optimizing their AC sets given the actual state of the world. We think about
the ACs as flexible mental constructs which are continuously adapting in a
learning process for maximizing the agents’ predictive power. As such, our
model treats the evolution of cultural profiles (the set of ACs) as the result
of a rational behavior governed by economic benefit. The ACs are continuous
variables, and agents are assumed heterogeneous in their (biological) prefer-
ences by default.
Our theory is very similar in spirit to the renormalization group schemes
of statistical physics as it is based on the following two key elements: (i) it
combines the original state variables (attributes of decision alternatives) into
important and less important degrees of freedom; agents keep the former (de-
fined to be the ACs), and discard the latter; (ii) this is done in an iterative way
aiming to identify fixed points, which are then associated with culture. The
criterion for truncating the degrees of freedom is local: each agent tries to max-
Cultural Evolution in a Population of Heterogeneous Agents 195
imize his evaluation accuracy, i.e., to minimize the representation/evaluation

error within his cognitive limits.
2 A Cognitive Model
We consider I agents, each restricted to use a number K of ACs only.
Agents divide the world into a number X of contexts in which they evalu-
ate decision alternatives according to their personal preferences. We assume
that alternatives are characterized by their objective (physical) attributes
a = {a1 , . . . , aD }. Agent i’s theoretical payoff from choosing alternative a
in context x is posited to be a linear function
(x) (x)
πi (a) = ω i · a, (1)
(x)
where ω i is the agent’s context vector, reflecting actual circumstances of
the context and the agent’s personal (biological) preferences. For each agent
there are X context vectors each of dimension D, which are assumed fixed
in the model. An agent does not know his context vectors explicitly (as this
would require a detailed understanding of the effect of all attributes on the
payoff), but by collecting experience on choices he has made previously, he
learns to approximate his payoffs using an appropriate mental representation
(this learning process is not modeled explicitly). The mental representation is
built around the world’s K most important degrees of freedom, constituting
the agent’s ACs. Assuming again a linear relationship, the approximate payoff
that the agent can compute directly is
(x) (x)

K
(x)
π̃i (a) = ω̃ i ·a= viµ γ iµ · a (2)
µ=1
(x)
where ω̃ i is the agent’s approximate context vector. This is decomposed using
(x) (x)
mental weights viµ in a reduced subspace of dimension K. The weight viµ
reflects the significance of AC µ in context x. Equation (2) implies that agent i
possesses a number K of AC vectors, {γ iµ }K µ=1 , assumed normalized |γ iµ | = 1,
which the agent uses to evaluate alternatives. We emphasize that the ACs are
defined in a context-independent way, and are used in all occurring contexts.
Note that by Eq. (2) the number of parameters defining the agent (the v’s and
γ’s) is reduced to (X + D)K, which can be much less than XD, the number
(x)
of parameters in ω i , if K << X, D.
The structure of our model is depicted in Fig. 1. We have defined three
layers: the layer of Physical Attributes of the decision alternatives, the layer
of Abstract Concepts (AC vectors) which constitute the renormalized degrees
of freedom, and the layer of context- and agent-specific Payoffs. At the lowest
level agents are homogeneous and have identical information about alterna-
tives. In the highest layer agents are heterogeneous and have individual payoff
π(x)(a)
context 1 … context x … context X Payoffs
vµ(x)
1 2 µ K Abstract Concepts
γµ
ω(x)
a = ( a1, a2, …, aD ) Physical Attributes
Fig. 1. Decision making heuristics using Abstract Concepts (ACs) as renormalized

degrees of freedom.
functions based on their innate preferences. The middle layer shows partial
coherence, whose measure, as we will see, is determined by the strength of
social interactions. This is the layer we intend to monitor across society for
observing the emergence of a (common) culture.
An example may be enlightening at this point. Think about the world of
chess where decision alternatives are the possible moves a player can choose in
a given situation. These moves are characterized by the positions of the pieces
on the board (physical attributes). Calculating the payoff of a move (the prob-
ability of winning the game with that move) is basically a two step procedure
for human players and chess programs alike. First the move is evaluated along
some abstract Concepts like “material advantage”, “positional advantage”,
“pinned piece”, etc. Indeed, research in psychology has demonstrated that a
human grand master possesses tens of thousands of such chess-related con-
cepts [9, 10]. Second, these concept scores are weighted (mental weights) in
an appropriate mental model. The weighting scheme depends on the player’s
personal skills/style (agent heterogeneity) and on the actual adversary (con-
text dependence). There is a theoretically best move (preference vector) for
that player with that adversary in that situation, but that is not known ex-
plicitly to the player. Mastering chess amounts to finding an optimal mental
representation, i.e., optimal Concepts and optimal weights (approximate pref-
erence vectors) which allow a decision close to the theoretical best. In chess
there are different schools (cultures) where the applied Concepts have some-
what different meaning. Of course, decision in chess is very non-linear, and
our simplified linear model can only grasp the big picture.
The hierarchical structure depicted in Fig. 1 formally resembles a linear
two-layer neural network. Note, however, that it is not a microscopic neural
network model of a cognitive function, but a phenomenological (macroscopic)
model based on a psychologically/economically supported, realistic human

decision making strategy.
Due to the reduction of dimensionality, K < D, the approximate payoff
(x) (x)
π̃i deviates from the theoretical payoff πi . The agents’ goal is to find the
best possible set of ACs and mental weights which minimizes the error of this
mental representation under the constraint that only K ACs can be used. The
natural measure of agent i’s representation error is the variance
X $
%
(x) (x)
Ei = (πi − π̃i )2 , (3)
x
x=1
where .x is the average over alternatives in context x. We restrict our at-
tention to the case, where the attributes are delta correlated ad x = 0,
ad ad x = δdd , and nontrivial structure is assumed in the agent-specific
context vectors only. This simplifies the forthcoming analysis without los-
ing essential features. With this proviso, the representation error turns out to
be
2
X
(x)
K
(x)
Ei = ωi − viν γ iν . (4)
x=1 ν=1
The agents learn to minimize the error Ei by adapting their AC vectors

(x) (x)
γ iν and mental weights viν to the ω i structure.
3 Optimal Concepts as a PCA Problem

Theoretically the minimizing parameters can be easily determined. Differenti-
(x)
ating w.r.t. viµ and assuming that the K × K metric tensor Giνµ = γ iν · γ iµ
is invertible (note that the ACs are not necessarily orthogonal), the optimal
mental weights read
(x) (x)
viµ = ω i · γ µi , (5)

where we have introduced the dual AC vectors γ µi = ν=1 [G−1
K
i ]µν γ iν . Writ-
ing this back to Eq. (4) we can write the error as

X
(x)

X
K
(x) (x)
Ei = |ω i |2 − (ω i ·γ iµ )(ω i ·γ µi ). (6)
x=1 x=1 µ=1
The first term is a constant, the second term should be maximized for
{γ iν }K
ν=1 . The problems becomes more transparent, if we write the second
K
term in the form µ=1 γ iµ · W i γ µi , by introducing the positive semi-definite,
D × D dimensional world matrix

X
(x) (x)
Wi = ωi ◦ ωi . (7)
x=1
This matrix encompasses all information about agent i’s personal relation to
the World. This maximization problem is well-known in many domains of
science under the name Principal Component Analysis (PCA). A particular
solution for the optimal AC vectors is given by the K most significant (largest
eigenvalue) eigenvectors of W i . Note, however, that any non-singular basis
transformation yields a different solution with the same error. We can freely
define linear transformations in the subspace of concepts; the mental weights
adapt, and the overall representation error remains unchanged. It is obvious
that if W i is a non-singular, rank-D matrix, then for K < D the error is
strictly positive even with an optimally chosen set of ACs.
The identification of the important degrees of freedom is closely analog
to what occurs in White’s Density Matrix Renormalization Group Method
(DMRG) [11]. In the DMRG the renormalized degrees of freedom are the K
most significant eigenvectors of the reduced density matrix. A formal map-
ping between the DMRG and our method can be established by identifying
the DMRG’s density matrix for block i with the world matrix W i , and the
renormalized (kept) degrees of freedom with the ACs. The DMRG’s truncation
error is analog to our representation error.
4 Social Interactions
As discussed above, by minimizing the representation error each agent learns

to approximate his world matrix in a lower dimensional space. If agents i and
j have different preferences, and thus different world matrices, they would
necessarily invent different optimal ACs. Cultural evolution boils down to the
interaction of these different ACs in a social network. In order to introduce
social interactions we cast contexts into two basic categories: those whose
context vector only depends on a single agent (individual contexts), and those
where the context vector for agent i depends on the (simultaneous) decision of
at least another agent, say j (social contexts). For simplicity only pair interac-
tions will be considered. In social contexts the understanding of other agents’
valuation and decision making processes is an asset. Agent i has strategic
advantage from being able to estimate j’s payoff for her alternatives. Since
j, as any other agent, bases her evaluation on her mental representation us-
ing her actual ACs, the task appears for i as the approximation problem of
(j) (j)
the context vector ω̃ i (the context now being agent j). Notice that ω̃ i is
necessarily within j’s AC subspace. The incentive to approximate agent j’s
context vector accurately introduces a “force” which tries to deform i’s AC
set towards that of j in order to improve i’s predictive power in j-related
social contexts. The interplay between the two competing goals - optimizing
the mental representation rather for individual or for social contexts - is the
ultimate factor which determines the society’s cultural coherence.
In terms of the above classification, agent i’s overall world matrix is a
sum of individual and
Xsocial contributions. As seen above, individual contexts
contribute W 0i = x=1 ω xi ◦ ω xi (x = individual). The contribution of a j-
(j) (j)
related social context is ω̃ i ◦ ω̃ i , which can be expressed explicitly in terms
of j’s AC vectors {γ jµ }Kµ=1 using Eq. (2). We realize that this contribution
(j) (j)
is a rank-K operator, which can also be written as ω̃ i ◦ ω̃ i = B S j , the
K
product of S j = µ=1 γ jµ ◦ γ µj projecting onto agent j’s AC subspace, and
another operator B acting within this subspace. This latter depends on the
actual vjµ weights, but for simplicity in the sequel we assume that it is an
identity operator. This assumption simply states that all directions in j’s AC
subspace have equal importance for i, and is reasonable if i has to predict j’s
behavior in many different j-related contexts which average out the weight
dependence.
Eventually, agent i’s overall world matrix takes the form

W i = W 0i + hij S j , (8)
j∈Ni
the sum being over his social network Ni . The parameters hij measure the rel-
ative strength (importance) of social interaction between agents i and j. The
above “equal importance” assumption assures that our W i remains invariant
w.r.t. nonsingular local linear transformations of the individual AC vectors.
This means that in our model the actual choice of Concepts does not count,
only the subspace they span does.
We view the society as a dynamical system in which agents continuously
act as best-response optimizers, trying to minimize their representation error
given the actual state of nature and other agents. A steady state of this best-
response dynamics [12] is necessarily a Nash equilibrium, where agents have
no incentive to unilaterally change their mental representations any further
because these are already mutually optimal.
In order to monitor cultural coherence we introduce a coherence order pa-
rameter (COP) as the population average of AC subspace projector operators
&K '
µ
O = S j I = γ jµ ◦ γ j . (9)
µ=1 I
Our COP is a tensor order parameter which measures the collective overlap of
the individual AC subspaces. In fact it is a high-dimensional generalization of
de Gennes’s tensor order parameter introduced in the theory of liquid crystals
[13]. The eigenvalue structure of O is useful to characterize the level of coher-
ence in the society and to distinguish different phases. Obviously, if there is
perfect order and the individual AC subspaces are all lined up, the COP has
eigenvalues o1 = o2 = · · · = oK = 1, and oK+1 = oK+2 = · · · = oD = 0. In the
World matrix: Wi (t) = Wi 0 + h O(t − 1)
iteration cycle
PCA
Concepts: γ i1 (t), γ i2 (t), ! , γ iK (t)
Population average
Order parameter: O(t)
Fig. 2. Iteration cycle of the best response dynamics in a mean-field social network.
Fixed points of the dynamics are Nash equilibria.
opposite extreme of complete disorder all directions are equivalent and the
eigenvalues are o1 = o2 = · · · = oD = K/D. The trace of the COP is always
K.
5 Cultural Ordering as a Phase Transition

The equilibrium and dynamical properties of the system depend crucially
on the parameters D, X, K, the statistical properties of W 0i , and the social
interaction matrix hij (connectivity structure). In the following we analyze the
(x)
case when the agents’ individual context vectors ω i are Gaussian random
vectors with zero mean and unit variance, i.e., without interaction agents
choose random AC subspaces (complete disorder). This is a benchmark case
which can demonstrate in its purest form how cultural coherence can emerge
spontaneously in an interacting social system.
Assuming this, the random matrix W 0i , which is quadratic in the context
vectors, has Wishart distribution [14]. We restrict our attention to a mean-
field network (all agents are connected) and set hij = h/I. With this the
world matrix in Eq. (8) simplifies to
W i = W 0i + h O. (10)
We let the system evolve according to the best-response dynamics: we start

from a certain initial condition of individual AC sets. In each step we first
calculate the actual COP, then update each agent’s AC set according to the
K largest principal components of W i . This gives agent i’s best response for
the actual environment (individual plus social). This procedure is repeated
until a fixed point is reached. A fixed point of the dynamics is necessarily a
Nash equilibrium, which we associate with the dynamically emerging culture
of the population. The scheme of the iteration is illustrated in Fig. 2.
An example for the COP spectrum as determined by a numerical simu-
lation is depicted in Fig. 3. For small h there is complete disorder, while for
Fig. 3. Fixed point spectrum of the order parameter matrix O for D = X = 10,
K = 4, I = 2000. Solid lines indicate increasing h, dashed line (only shown for o1 )
decreasing h. The phase transitions are subcritical bifurcations.
h → ∞ the fixed point is completely ordered. In between, however, we ob-

serve a series of dynamical phase transitions with hysteresis. Each transition
can be associated with a first order jump in the eigenvalues. A qualitatively
similar behavior, with a series of first order transitions was found numerically
for other parameter values, too.
Our results suggest that the social dynamics has a unique equilibrium only
for small couplings. Above the critical value hc spontaneous ordering occurs
in the system, and the number of equilibria increases. On the one hand, there
is a discrete number of possible fixed point “families” as represented by the
eigenvalue structure of the COP: an adiabatically slow change in h leads to
bifurcations at some critical values, where the system jumps from one equilib-
rium to another. These transitions are irreversible and give rise to hysteresis,
which assures the coexistence of at least two different equilibrium families
within the hysteresis loop. On the other hand, each family of equilibria is
infinitely degenerate in itself, since the associated eigenvectors of the COP
can point in essentially any direction (respecting orthogonality). Since we as-
sumed that preferences are random, there are no preferred directions on the
social level by default. This symmetry of the D-space is broken spontaneously
above hc . As h → ∞ there is only one possible structure for the eigenvalues
of the COP (there is only one family), meaning that the society is fully coher-
ent. However, this coherent K-dimensional subspace, representing the most
important dimensions of the society is not fixed a priori. The ultimate state
of culture gets selected as a result of the idiosyncratic fluctuations present at
the time of the symmetry breaking bifurcations.
6 The Onset of Ordering

In the mean field model the critical interaction strength hc , where the com-
pletely disordered phase loses stability can be calculated analytically for
I → ∞. Starting from the disordered phase we can write the COP in step l as
O = (K/D)1 + ε δO l , where δO l is an arbitrary perturbation. The dynamics
defines a mapping δO l → δO l+1 . If the disordered state is stable we have
||δO l || → 0, otherwise ||δO l || diverges. We can identify hc as a fixed point of
this mapping δO l+1 = δO l , and determine δO l+1 as a function of δO l and h
using first order perturbation theory in ε. The calculation necessarily involves
the spectral properties of the random W 0i ensemble, and leads to [15]
& D '
D2 1
hc = K , ξν = 2 , (11)
ν=1 ξν
λiν − λim
m=ν I
where λim are eigenvalues (in descending order) of the Wishart matrix W 0i .
The joint probability distribution for real Wishart matrices in the Coulomb
gas representation reads [14]
⎛ ⎞

min(D,X)
(
min(D,X)
P ({λ}) ∼ exp ⎝− V (λn ) + ln |λn − λm |⎠
n=1 n>m
with V (λ) = (λ − (|X − D| − 1) ln λ)/2. For large D and X a saddle-point

approximation can be made to determine the average position of particles
(eigenvalues), giving the alternative expression ξν = 1 + (D − X √
− 1)1/λν .
When ν << D, X we can write 1/λν ≈ 1/λ1 = 1/(D + X + 2 DX) [14].
Finally, in the limit D, X → ∞ with X/D = r > 0 and K << D this leads
[15] to the critical coupling
√
D2 1 + r
hc ≈ . (12)
K 2
Keeping r fixed, hc increases in D (quadratically) and decreases in K. When
the World is complex (large D and X) and the agents are primitive (small
K) the critical coupling is high and the society is likely to stay culturally
disordered.
7 Conclusions
In summary, in this paper we have proposed a rather general theory of cultural
evolution built upon a dimension reduction scheme. We assumed that bounded
rationality forces agents to choose a limited set of abstract concepts in describ-
ing their world. These concepts and the way they are used are continuously
Strength of social interactions h

Ordered
Coherent Culture
hc
Disordered
No Culture
Agent intelligence K
Fig. 4. Schematic phase diagram on the h vs. K plain (D, X = fixed) for the mean-
field model with random preferences. Solid line indicates the principal transition
calculated in the text, dashed lines the additional transitions.
updated and refined in order to find an optimal representation. Individual in-

terest in making the best possible decisions in natural (individual) and social
(collective) contexts act as a driving force for cultural evolution and, at least
in the particular cases studied, drive the system towards a Nash equilibrium
fixed point with measurable level of social (cultural) coherence. In our model
with random preferences this coherence manifested itself as a spontaneous
ordering of the individual concept (sub)spaces, and was quantified by an ad-
equate order parameter. We demonstrated that as the fundamental cognitive
and social parameters change, cultural evolution goes through a number of
phase transitions, which we found first order in the mean-field model. The
critical value of the most relevant transition was calculated analytically.
Transition to a phase with substantially higher cultural coherence can be
triggered in two different ways: social contexts can increase in significance, or
alternatively, agents may develop better cognitive skills (see Fig. 4). A slow
increase in the parameters leads to a bifurcation, beyond which the subspaces
spanned by the agents’ concepts rearrange into a state with higher average
overlap, and we can speak about a boost in cultural coherence. A similar
transition may explain the sudden cultural explosion for humanity approxi-
mately 50,000 years ago [16, 17]. Since in our model the ordering involves a
spontaneous symmetry breaking, the emerging fixed point strongly depends
on the initial perturbation. Such path dependence implies that cultural evolu-
tion is not completely deterministic but depends on a number of idiosyncratic
factors.
Our theory differs conceptually from some traditional thinking on the pos-
sible causes underlying human cultural explosion. Most of these traditional ex-
planations presume a sudden and substantial biological change (e.g., genetic
mutations [17, 18], the integration of cognitive modules in the brain [16], etc.)
as the initial trigger. In contrast to these, the mechanism we have proposed
in this paper suggests a spontaneous ordering of the mental representations,
which emerges when slow changes in some relevant variables (cognitive abili-
ties, social interactions) reach a critical level. In this respect the “cultureless
state” of the society, originally rooted in the heterogeneity of its members
in individual preferences, needs, experiences, and the extreme simplicity of
the mental representations they possess, becomes unstable at a bifurcation
point, beyond which the collective social dynamics drives the system into a
drastically different, culturally ordered fixed point.
It would be interesting to go beyond mean field, and investigate the model’s
dynamics in more realistic social networks, where the fixed point may possess
a more complex internal structure. A non-mean-field-like social structure is
expected to give rise to subcultures or cultural domains, where individuals
are highly coherent within a domain, but the average cultural subspace the
domain represents may be rather different from that of another domain, as
occurs usually in real societies. The type of the occurring dynamical transi-
tions may also change character. An evaluation of how a potential structure
in the attribute space or non-linearity in the world/mental representations
influence the results could also be interesting. Other intriguing problems in-
volve the social dynamics itself, namely how the model should be extended to
cope with intrinsically dynamical aspects of culture such as fads and fashions,
where the social attractor is much rather a limit cycle or a chaotic attractor.
8 Acknowledgments
The authors acknowledge support from Insead Foundation and the Hungarian
Scientific Research Found (OTKA) under grant Nos. T043330 and T047003.
References
1. H. A. Simon. Models of bounded rationality, Vol. 3: Empirically grounded eco-
nomic reason. MIT Press, 1997.
2. A. K. Romney, J. P. Boyd, C. C. Moore, W. H. Batchelder, and T. J. Brazill.
Culture as shared cognitive representations. Proc. Natl. Acad. Sci. USA,
93:4699, 1996.
3. R. Aunger, editor. Darwinizing culture: The status of memetics as a science.
Oxford University Press, 2001.
4. R. Axelrod. The dissemination of culture - a model with local convergence and
global polarization. J. Conflict Res., 41:203–226, 1997.
5. R. Axelrod. The complexity of cooperation. Princeton University Press, 1997.
6. C. Castellano, M. Marsili, and A. Vespignani. Nonequilibrium phase transition
in a model for social influence. Phys. Rev. Lett., 85:3536, 2000.
7. G. Weisbuch. Bounded confidence and social networks. Eur. Phys. J. B,

38(2):339–343, 2004.
8. J. Conlisk. Why bounded rationality? Journal of Economic Literature, 34:669,
1996.
9. L. Mérő. Ways of Thinking: The Limits of Rational Thought and Artificial
Intelligence. World Scientific, 1990.
10. H. A. Simon and W. G. Chase. Perception in chess. Cognitive Psycology,
4:55–81, 1973.
11. S. R. White. Density matrix formulation for quantum renormalization groups.
Phys. Rev. Lett., 69:2863, 1992.
12. J. Hofbauer and K. Sigmund. Evolutionary Games and Population Dynamics.
Cambridge University Press, 1998.
13. P. G. de Gennes. The Physics of Liquid Crystals. Oxford University Press,
1974.
14. A. Edelman. Eigenvalues and condition numbers of random matrices. PhD
thesis, MIT, 1989. downloadable from http://www-math.mit.edu/˜edelman.
15. G. Fáth and M. Sarvary. A renormalization group theory of cultural evolution.
Physica A, 348:611–629, 2005.
16. S. Mithen. The Prehistory of the Mind. Thames and Hudson, London, 1996.
17. R. G. Klein and B. Edgar. The dawn of human culture. John Wiley and Sons,
2002.
18. Y.-C. Ding, H.-C. Chi, D. L. Grady, A. Morishima, J. K. Kidd, K. K. Kidd,
P. Flodman, M. A. Spence, S. Schuck, J. M. Swanson, Y.P. Zhang, and R. K.
Moyzis. Evidence of positive selection acting at the human dopamine receptor
d4 gene locus. Proc. Natl. Acad. Sci. USA, 99:309–14, 2002.
Part V
Auction and Two-sided Matching

Simulating Auctions
Konrad Richter
University of Kiel and McKinsey&Co

konrad richter@mckinsey.com respectively fidibus1972@yahoo.com
Summary. Current auction theory relies crucially on the assumption that all bid-
ders bid homogeneously according to their Nash Equilibrium bidding strategies.
However, it remains silent on whether and how a priori heterogeneous bidders arrive
at the NE. This paper investigates computationally whether evolutionary learning in
repeated auctions could justify the assumption of NE bidding. Simulations show that
Best Response learning of a priori heterogeneous bidders in first-price auctions does
not converge to the NE. Instead, bidders involve in permanent mutual adaptation
that shows non trivial characteristics.
1 Background
The Revenue Equivalence Theorem (see [9, 6, 8]) is one of the most funda-
mental Theorems in auction theory. It basically states that expected seller
revenue and the final allocation of goods in pure-bidder independent private-
value auction formats depends only on the allocation rule but not on the
payment mechanism.1 The Theorem relies crucially on the assumption that
bidders play according to their Nash Equilibrium (NE) strategies. Also most
other auction literature since then has focused on NE bidding. However, the
literature is sparse when it comes to explaining whether and how a priori
heterogeneous bidders arrive at NE play.
The usual justification for NE play is that players in games are perfectly
rational, they know that each of their opponents is perfectly rational, they
know that the opponents know that all the other players are perfectly ratio-
nal and so on. Clearly, in many games, this is not a very realistic assumption.
A common alternative justification for the focus on NE play in repeated games
is given by the theory of evolutionary learning. In particular, Best Response
1
Common payment mechanisms are for instance:
First-price auctions: The winning bidder pays her own bid
Second-price auctions: The winning bidder pays the second highest bid
All-pay auctions: All bidders – also the loosing ones – pay their bid
210 Konrad Richter
(BR) Dynamics – where each bidder plays the strategy that would have max-
imized her payoff in the past – converges to the NE in many games (see, e.g.,
[4, 11, 2]). BR can be interpreted as describing perfectly rational players who
just don’t know whether their opponents are perfectly rational as well. In
that sense the only difference to standard game theory is the disposal of the
assumption of a priori player homogeneity.
In [7] I develop a mathematical formalism to quantify the behavior of BR
bidders in the particular case of first- and second-price private-value auctions
(1PAs resp. 2PAs) with two bidders and values drawn anew in each round from
a uniform distribution. By restricting players to linear (value-independent)
bidding strategies I am able to derive closed-form expressions for the strategies
if one bidder always updates to a BR against a fixed strategy of the opponent.
The restriction to linear bidding functions is motivated by the fact that in
1PAs and 2PAs with uniform value distributions and risk neutral bidders, the
NE is given by linear bidding strategies bi (vi ) = βi vi . In the specific case
of two bidders with uniformly distributed private values, the NE in a 1PA
is given by βi = 0.5. In 2PAs, the NE is bidding the true value (βi = 1),
irrespective of the number of participants in the auction or the underlying
value distribution.
In addition to different auction formats, I also investigate the influence
of memory capabilities on bidders’ behavior. At the one extreme I consider
myopic Best Response (mBR) bidders who in each round play the strategy
that would have maximized the payoff in the last auction. At the other extreme
I consider perfect memory Best Response (pBR) bidders who play the strategy
that would have maximized the cumulated payoff of the whole game up to the
present round. Perfect memory Best Response is in the literature also often
denoted as Fictitious Play.
Both, mBR and pBR, are extreme cases which are not a plausible model
for the behavior of human players. For the interpolation between these two
extremes, I consider truncated Fictitious Play (tFP): Players update to the
strategy that would have maximized their cumulated payoff in the last nRds
rounds. Note that nRds= 1 corresponds to mBR while nRds= L corresponds
to pBR if L denotes the total number of rounds played in the game so far.
The mathematical results show that the Revenue Equivalence Theorem
breaks down under Best Response Dynamics: BR in repeated 2PAs leads
bidders to pursue their NE bidding strategies of bidding their true value as
long as they are not completely myopic. The underlying reason is that bidding
the true value is a (weakly) dominant strategy and therefore is included in
every Best Response set - irrespective of the opponent’s bid.
In contrast, bidders in 1PAs fail in general to find the NE under BR: If
bidders in 1PAs pursue pBR, they converge to the NE from below. However,
it takes them infinitely long to arrive there since the opponent’s below-NE
strategies from early stages of the game push the respondent’s strategy be-
low the NE forever thereafter. If, however, bidders are modeled with limited
memory capabilities via tFP, they can aggregate too little information to accu-
Simulating Auctions 211
rately estimate the opponent’s strategy from her bids. This uncertainty about
their opponent’s true strategy pushes play constantly below the NE resulting
in permanent fluctuations around a mean below the NE. As a consequence
the players’ strategies are not Best Responses to each other any more. There-
fore, under the realistic assumption of finite memory capabilities, bidders in
1PAs lock in permanent mutual adaptation of heterogeneous bidding strate-
gies which in turn leads to a suboptimal allocation of goods.
The positioning of this research is at the boundary of auction theory and

the theory of evolutionary learning. The research differs in two technical as-
pects from the usual focus on matrix games in evolutionary learning theory:
First, bidders can not observe the opponent’s bidding strategy directly but
only her bid bi (vi ) = βi vi that depends on the strategy and the (unobservable)
private value of the opponent. Second, each bidder has a continuum of feasible
strategies as opposed to the finite strategy set that is usually considered in
the literature. On the other hand, the research differs also from most other
papers on auction theory in that it doesn’t investigate how the auction setup
influences NE-bidding2 . Instead, it investigates if and under which assump-
tions initially heterogeneous bidders arrive at the NE and what the dynamics
of repeated auctions looks like.
Since this is a rather unexplored research area, I used two complementary
methodologies for the investigation. On the one hand, I derived in [7] a math-
ematical formalism to analyze the behavior of BR bidders in 1PAs. On the
other hand, I developed a computational model, the Auction Simulator (AS)
that is – among other setups – capable of simulating Best Response strategy
updating with linear bidding strategies. The simulation results are the focus
of this paper.
The correct replication of the mathematical results by the simulation serves
as a double-check for both, the mathematical formalism and the simulation.
Therefore, the AS can be used to reliably investigate auctions at a more
detailed level than would be possible by purely mathematical methods.
In this paper I present the results of simulation runs with the AS which
investigate the internal dynamics of 1PAs. In section 2 I present the basic
structure of the AS. Section 3 compares the simulation results with the mathe-
matical predictions of [7]. Section 4 presents results of simulations of truncated
Fictitious Play and shows how the AS can be used to assess Nash Equilibria
of auctions. Section 5 concludes and gives an outlook for future research.
2
A notable exception in the literature is [3]. This paper focuses on learning in
repeated first-price auctions where bidders’ values are the same in each auction.
Under this assumption the paper shows that evolutionary learning converges to
the NE of the one-shot auction. My research instead investigates auctions where
values are drawn anew in each auction.
212 Konrad Richter
2 The Auction Simulator

The AS is programmed in SWARM. SWARM is basically a library of Objective
C and Java routines that was designed at the Santa Fe Institute for facilitating
agent based modeling. It is distributed freely under the GNU General Public
License and is downloadable from http://wiki.swarm.org. It runs under Unix,
Linux, Windows and Apple operating systems. The source code of the AS can
be obtained from me on request.
2.1 Simulation Flow
The AS simulates an arbitrary number of bidders. Each bidder i has access to

a private set of nStrat strategies. Strategies are real numbers βi . Each bidder
i chooses an active strategy βi∗ that determines the bid bi (vi ) = βi∗ vi that she
is actually placing in a specific auction. All other strategies are evaluated as
well to see how they would have performed if they would have been the active
ones. Changes of the active strategy are possible only every nRds rounds. For
the strategy updating there are two alternative mechanisms available. With
Fixed Strategies (FS), a new strategy is chosen as active while the strategy
set remains fixed. With Genetic Algorithms (GAs), the strategy population
is additionally replaced by a new generation.
A simulation run with nGen generations consists of the following steps:
1. in the beginning, participants initialize their active and nStrat-1 non-
active strategies
2. for nGen generations
a) bidders reset their current payoffs for all strategies
b) for nRds auctions
i. bidders’ values are randomly chosen from the underlying value
distribution
ii. bidders submit their bids bi (vi ) = βi∗ vi according to their active
strategy βi∗
iii. the winner is determined
iv. bidders update the current payoffs of their active and passive
strategies
c) bidders evaluate the payoff, their strategies (would) have generated
during the nRds auctions and choose an active one for the next nRds
auctions (see below)
d) if the GA is used, the strategy population is updated by the genetic
operators crossover and mutation
e) proceed with (a)
2.2 Strategy Updating
A straight-forward implementation of the mathematical model of Best Re-

sponse is implemented by a fixed strategy (FS) grid for players. Under FS
I partition the strategy space (0, 1) into a grid of equidistant strategies, e.g.
nStrat=1000 ranging from 0 to 1.
In each auction the AS calculates for each bidder the payoff that each of the
1000 strategies would have generated if it would have been active in the auc-
tion. Focusing on bidder 0 in a two-bidder auction with bidder 0 and bidder
curr;(i)
13 , the current payoff P O0 in each auction round i for each of bidder 0s
strategies β0 is given by
#
curr;(i) (v0 − β0 v0 )Θ(β0 v0 − β1∗ v1 ) in 1PAs
P O0 (β0 ) =
(v0 − β1∗ v1 )Θ(β0 v0 − β1∗ v1 ) in 2PAs .
The Unitstep-function
#
1 if β0 v0 ≥ β1∗ v1
Θ(β0 v0 − β1∗ v1 ) =
0 if β0 v0 < β1∗ v1
guarantees that a strategy only gets nonzero payoff if it would have won the
auction.
Each nRds auctions, each strategy’s payoffs of the last nRds auctions are
added up and yield the cumulated payoff of that strategy:

nRds
curr;(i)
P O0cum (β0 ) = P O0 (β0 ) .
i=1
After, the strategy that generated the highest payoff is selected as the active
one. If no strategy has positive payoff4 , the currently active strategy remains
active, β0new = β0old . If more than one strategy would have maximized the
cumulated payoff, the AS stays with the previous strategy if it is included in
the Best Response set.
This simulation setup allows the simulation of all kinds of memory capa-
bilities: Setting nRds= 1 simulates myopic Best Response, nRds> 1 allows the
simulation of truncated Fictitious Play. Additionally, the AS allows for the
aggregation of the payoff over all auctions to simulate perfect memory Best
Response.
As an alternative tool for the simulation of learning I employ Genetic Algo-
rithms (GAs). However, simulations with GAs run the danger that the results
are governed by the internal GA dynamics and therefore do not correspond to
the theoretical concept of Best Response anymore. This is especially true if we
consider myopic Best Response because then the fitness landscape fluctuates
too frequently to allow the GA to converge to the optimum. Still, by setting
nRds sufficiently high, GAs are a fast tool to assess the approximate position
of the NE. I will say more on this in section 4.2.
3
Setups with more bidders are straight-forward generalizations
4
The case that no strategy has positive payoff happens for instance in 1PAs under
mBR if v0 < β1 v1 because in this case, no feasible strategy βi ∈ [0, 1] could have
won the auction
214 Konrad Richter
3 Reproduction of Analytic Results

In [7], I derive for a variety of different learning models closed form solutions
if two bidders bid in 1PAs with uniform value distributions. In particular, I
obtain formulas that predict how one bidder adapts under mBR or pBR if
the opponent plays a fixed strategy. For all responses against fixed strate-
gies, the AS replicates the analytic predictions very accurately with FSs. The
simulation results are within a 1%-range of the analytic predictions.
For mutual adaptation, the mathematical results predict that strategies
under pBR converge slowly from below to the NE whereas under mBR they
converge from above to 0.
Table 1 reports the mean strategy for mBR in dependence of the number
of strategies nStrat that are used throughout the simulation.
Table 1. Myopic Best Response
nStrat 9 99 999 9999 49999

β0FS 0.3009 0.1194 0.0608 0.0387 0.0282
The results show that strategies – in contrast to the mathematical pre-

diction – in general do not converge to 0 under mutual adaptation. However,
the simulations for increasing numbers of strategies show that this is a finite
size effect. As the number of strategies goes to infinity, the strategies converge
to 0 under mutual adaptation and are therefore in line with the theoretical
prediction. How can we understand this?
The mathematical analysis predicts that strategies converge to 0 for two mu-
tually adapting bidders who update according to myopic Best Response. In
the simulation, however, there exists in each bidder’s strategy set a lowest
1
strategy β min > 0. In the case of 1.000 strategies this would be β min = 1.000 .
A bidder who would be driven by mBR to play a strategy below this low-
est strategy is forced to overbid because the lowest strategy guarantees her
positive expected payoff whereas bidding 0 gives 0 payoff for sure. This effect
constitutes a barrier against convergence to 0. Now, imagine for instance that
(t) 1
player 1 plays her minimum strategy β1 = 1.000 and that the random values
(t) (t)
in a particular round are given as v1 = 0.5 and v0 = 0.0007. Then, the best
(t) (t)
(t+1) β1 v 1
response of bidder 0 is given by β0 = (t) ∼ 0.715. So, even if the oppo-
v0
nent plays her minimum strategy, there always is a positive probability that
each bidders’ strategy breaks away from 0. This probability doesn’t decrease
over time. As soon as one strategy breaks away from 0, the other strategy
subsequently also breaks away and the strategies involve in mutual oscilla-
tions. These oscillations eventually die out, leading the system near 0 again.
Therefore, the simulation of mutually adapting mBR bidders shows clustered

volatility in the time series of strategies and asset prices5 .
Note that the effects of decimalization quickly vanish as the number of
bidders or nRds increases.
To investigate perfect memory Best Response, I set nRds= 1.000 and cu-
mulate the payoffs over the whole game. I simulate 24.000 generations, that
is a total of 24.000.000 auctions. The following table reports the averages of
the first to the fourth quartile of 6.000 generations each.
Table 2. Perfect Memory Best Response
generation 0-6.000 6.000-12.000 12.000-18.000 18.000-24.000

FS
β¯0 0.4574 0.4771 0.4818 0.4842
In agreement with the mathematical prediction, strategies converge to the

NE of 0.5 but it takes them infinitely long to arrive there.
In summary, the comparison between mathematical predictions and sim-

ulation results shows that the AS is free of bugs. However, in the case of
mutually adapting mBR bidders, the simulation results differ from the math-
ematical predictions due to the quantization of the strategy space. This shows
that bid decimalization can be an important factor in determining the out-
come of repeated auctions.6 The decimalization effects vanish quickly if either
bidders are equipped with nonzero memory capabilities or if more than two
bidders participate in the auctions.
Since the AS is free of bugs, I can use it to investigate also setups where the
mathematical analysis would be too tedious. In the following section I follow
this route by looking at more general auction setups.
4.1 Truncated Fictitious Play
In the following I consider a simulation of 10 bidders who mutually adapt by

truncated fictitious play with nRds=100: Bidders calculate in each auction for
5
The price of the asset, i.e., the seller revenue, is given by the higher of the two
bids.
6
Actually, I model here the effects of strategy quantization, not those of bid quan-
tization. However, the effect that bidders do not converge to 0 is the same in both
cases.
216 Konrad Richter
every strategy the payoff it would have generated if it had been the active one
in the previous 100 rounds. Every 100 rounds they update their strategy to
the strategy that would have maximized their payoff if it had been the active
one. Figure 1 shows typical time series over 200 time steps - i.e. over 20.000
auction rounds - between 18.500 and 18.700.
The upper left plot shows the strategy of bidder 0. Of course, under mutual
Fig. 1. Truncated Fictitious Play with Uniform Value Distribution
adaptation, the strategies of the other bidders look qualitatively the same.
As predicted by the mathematical analysis, strategies fluctuate permanently.
Moreover we see from the plot that the fluctuations are asymmetric: There
are occasional breakouts to extremely low strategies whereas corresponding
breakouts to high strategies are not observed. In consequence, the histogram
of bidder 0’s strategies (not shown) has an asymmetric shape, it is skewed to
the left. Table 3 quantifies the skewness: the mean of β0 is below the median
of β0 . This shows that there are less strategies below the mean than above
the mean. Therefore, the low strategies must show more significant deviations
from the mean than those above the mean - the distribution is skewed to the
left.
The upper right plot shows for each time step the strategy of the bidder who
has obtained the highest payoff in the previous nRds rounds. The plot sup-
ports the theoretical result that the best response against an environment of
mutually adapting bidders is in general not the NE. Again we see visually
from the plot or from Table 3 that the distribution is skewed to the left. Note
that the fluctuations of the best strategies to low strategies are much more
modest than those of the individual strategies. This is intuitively clear: A
player with high values throughout the nRds auctions of a certain time step,
will play a low strategy in the next round. Such a low strategy occurs then in
the time series of the bidder’s strategies. However, the chance that she will be
the most successful among all the other players with this low strategy in these
next nRds auctions – which would put this strategy into the time series of best
strategies – is very low, since in these next auctions new value realizations are
drawn.
The lower two plots in Figure 1 shows the maximum strategy and the average
of the strategies used by all bidders in the corresponding time steps. Visually
and by comparison with Table 3 we see that these strategies are distributed
symmetrically around their mean.7
The maximum strategy is on average above the NE. So, though mathemati-
cally the best response function doesn’t get higher than the NE, the finiteness
of nRds= 100 lets strategies also increase beyond the NE because then the
aggregation of the values over the 100 rounds is subject to stochastic shocks.
The average of all strategies played by the bidders, however, is usually below
the NE. A priori it is not clear, how important the maximum strategy is in
determining the seller revenue when compared to the average strategy: The
bidder with the highest strategy is on average the most likely to win each auc-
tion, therefore the maximum strategy is the most important single strategy
for determining seller revenue. On the other hand, due to the random value
realizations, the highest strategy will not win all of the nRds auctions but all
players will win some of the auctions. Therefore, the average strategy is also
an important factor in determining the expected seller revenue. The question
is, which of the two is more important in determining the seller revenue8 .
To get a feeling for which of the factors is more important for the determina-
β0 β best β max β av
mean 0.858 0.849 0.907 0.858
median 0.864 0.854 0.907 0.859
Table 3. Strategy Means and Medians under Truncated Fictitious Play
7
βav might be slightly skewed to the left. However, this remains to be assessed in
longer simulation runs.
8
If, for instance, the average strategy was 5% below the NE, the maximum strategy
was 1% above the NE and seller revenue was one 1% below the NE level, then the
maximum strategy would be more important in determining the seller revenue.
218 Konrad Richter
tion of seller revenue, I computed the effective strategy. This is the strategy
that would give the same seller revenue as truncated fictitious play if all bid-
ders would use it constantly as their active strategy. To calculate it I use the
seller revenue: Average seller revenue in a first-price auction with 10 bidders
where bidders bid according to the NE of βi = 0.9 is in the simulation ob-
tained as 0.81838. Average seller revenue in the 1PA with truncated fictitious
play is 0.78660. Therefore, the effective strategy of the bidders is given as
0.78660
βieff = 0.9 · = 0.86505 . (1)
0.81838
Comparison with the means in Table 3 show that this is much closer to the
mean of the average strategy than to the mean maximum strategy. Therefore,
the average strategy is the more important factor for seller revenue than the
maximum strategy. The intuition behind this is the following: Bidders’ strate-
gies in a given time step typically are between 0.6 and 0.9 whereas values
come from U (0, 1). Therefore, the more important determinant of the single
bid bi (vi ) = βi vi is the bidder’s value. So, the bidder with the maximum strat-
egy will not make that many more successful bids than an average bidder and
therefore she is not the main driver of seller revenue. An open question is,
what would happen if values come from a sufficiently narrow normal distribu-
tion. Values would be closer together but strategies as well. It remains a topic
for further research to assess wether the effective strategy moves closer to
the maximum strategy or wether its position between average and maximum
strategy is a constant that is independent of the underlying value distribu-
tion. I want to assess this in the future when I have extended the AS for the
simulation of nonlinear strategies.
The simulation result for a 2PA with 10 NE bidders is an average seller
revenue – cumulated over 100 rounds – of 81.835. This is virtually the same
as the 81.838 in the first-price auction in the NE. The equality highlights the
validity of the Revenue Equivalence Theorem in this special setting if bidders
play according to their NE bidding strategies.
The standard deviation of the seller revenue – cumulated over 100 rounds
– in the 1PA where bidders use their NE strategies is 0.74 whereas in the
second-price auction it is 1.112. This is simply a consequence of the fact that
the variance of the second highest of 10 values is higher than the variance
of the highest of 10 values. In the literature this observation has been used
to claim that risk averse sellers should prefer first-price auctions over second-
price auctions.
Under truncated fictitious play, this result doesn’t hold anymore. In 1PAs,
in addition to the random value realizations, the seller revenue is influenced by
the fluctuating strategies. The standard deviation of cumulated seller revenue
in the 1PA under truncated fictitious play with numBidders= 10 and nRds=
100 is 1.417 which is about 25% above the standard deviation in the second-
price auction. In consequence, I would argue that only a sufficiently risk-loving
seller should employ first-price auctions to allocate goods. Risk neutral and
risk averse sellers would prefer second price auctions since they give more
expected revenue and less risk for that revenue.
To assess the dynamics of truncated fictitious play in more detail, I es-
timate an AR(p)-model9 for the average strategy βav . The first step in esti-
mating such a model is to decide on the number of lags of the observable.
Allowing for too few parameters might neglect useful information, allowing
for too many parameters dilutes the focus on the core drivers of the time
evolution. For the model specification I chose p that minimizes the Bayesian
Information Criterion (BIC). The BIC is defined as
ln T
BICp = ln σp2 + p (2)
T
Here, T is the number of time steps in the sample, σp the standard deviation
of the residual error and p the number of considered lags. The first part in
BICp gives lower values, the better the goodness of fit is. The second part
gets the bigger, the more parameters the estimation uses.
Estimation with different settings for p in GAUSS shows that the BIC is
minimized for p = 2 with a value of −136420. Therefore I choose the AR-
Model
(t)
βav (t-1)
= α0 + α1 (βav − α0 ) + α2 (βav
(t-2)
− α0 ) + (t) (3)
The parameter estimates are given by GAUSS as
α0 = 0.858 ± 0
α1 = 0.344 ± 0.007 (0.330, 0.357)
α2 = −0.030 ± 0.009 (−0.047, −.0.014) ,
where the numbers in the bracket give the upper and lower confidence interval
at the 2σ-level. All coefficients are significantly different from 0 at the 2σ- level
and therefore the AR(2)-equation for the average strategy is given as
(t)
βav (t-1)
= 0.858 + 0.344(βav − 0.858) − 0.03(βav
(t-2)
− 0.858) + (t) (4)
That high average strategies promote high average strategies in the next round
is intuitively clear: If each bidder faces at time t an environment of high bid-
ding competitors, she will in the next round also use a high strategy. Con-
versely, by low bidding competitors she is led to use a low strategy in the next
round. Compared to a random walk, this AR(2) process leads to longer con-
secutive periods of the average strategy above the mean respectively below it.
So, in repeated auctions we might expect periods of fiercer competition among
bidders that interchange with periods of lower competition. This process is
purely internally generated and stems only from the mutual adaptation of
bidding strategies.
Analogously I estimated an AR(2)-model of the cumulated seller revenue
over the time steps of 100 rounds and obtained the parameter estimates
9
I do not consider general ARIMA-models since the time series is obviously sta-
tionary. Estimation of an ARM A(1, 1)-model gives the same BIC-criterion as an
AR(2)-model, so alternatively we could also investigate this model.
220 Konrad Richter
α0 = 78.6 ± 0
α1 = 0.285 ± 0.006 (0.273; 0.298)
α2 = −0.009 ± 0.004 (−0.017; −0.001) .
The process is qualitatively the same as for the average strategy: Periods
of relatively high seller revenue change with periods of relatively low seller
revenue. However, the parameter α1 is lower than in the AR(2) model for βav .
The reason is that the random value realizations dilute the effects of the time
evolution of the strategies. I did not estimate AR-models for the individual
strategy β0 and the minimum strategy βmin since they are not symmetrically
distributed around their mean and therefore an AR model with a normal noise
specification is not apt for their investigation.
By setting nRds= 1.000, I obtain the following AR(2)-models for βav
(t)
βav (t-1)
= 0.858 + 0.374(βav − 0.858) − 0.042(βav
(t-2)
− 0.858) + (t) (5)
and for the cumulated seller revenue

(t-1) (t-2)
t
P Osell = 807 + 0.296(P Osell − 807) − 0.005(P Osell − 807) + (t) . (6)
Comparing the coefficients with the coefficient estimates of the simulation

with nRds= 100 shows that the parameter nRds has only little effect on the
(t) (t)
time evolution of βav and P Osell : Increasing nRds leads to a slight increase
of the parameter values since the influence of the random value realizations
decreases.
The coefficients depend only on the number of bidders in the auction. Table 4
shows the coefficients of the AR(2) estimation for different numbers of bidders.
The first-order autocorrelation decreases slowly when we increase the number
numBidders 2 5 10 20
α0 (βav ) 0.423 0.395 0.374 0.34
α1 (βav ) -0.009 -0.032 -0.042 -0.041
α0 (P Osell ) 0.396 0.328 0.295 0.274
α1 (P Osell ) 0.003 -0.006 -0.005 -0.012
Table 4. Coefficient Estimates for Different Number of Bidders
of bidders. It remains a topic for further research to investigate wether the

autocorrelation vanishes for numBidders→ ∞. I want to investigate this when
I have access to faster computer facilities.
4.2 Computational Assessment of Nash Equilibria
[10] proves that the NE in any k th price auction with n bidders and underlying
uniform value distributions is given by the linear bidding strategy
k−2
b(v) = (1 + )v
n−k+1
Since the AS can simulate linear bidding strategies and we have seen that
pBR converges (slowly) to the NE in a 2 bidder 1PA, it is natural to ask if
we can use the AS in general to assess these NEs in general.
As an example I investigate a 4th price auction with 8 participants. The
NE strategy of this auction is b(v) = 1.4v.
Since initially I don’t know anything about the position of the NE, I use the
GA to find the approximate position.10 I put nRds=49.999 to limit the effects
of fictitious play truncation and calculate the average of the last 50 of 150
generations. The result indicates that the NE is at 1.3998 ± 0.0133 which is
in perfect agreement with the theoretical prediction.
In general, for all k th price auctions with n bidders and uniform value
distributions, I was able to get the analytically predicted result for the NE.
For general auction setups like, e.g., asymmetric or normal value distribu-
tions it is often difficult if not impossible to calculate the NEs. Therefore the
AS could provide significant help in identifying NEs of more complex auction
setups. However in these setups, the NE is usually not given by linear bid-
ding strategies. Therefore, I want to extend the AS in the future to allow for
nonlinear bidding functions.
5 Conclusion
In this paper I have used the Auction Simulator to investigate Best Response
learning in repeated first-price auctions. Bidders bid in repeated auctions for
assets. They have different asset valuations which change for every auction.
Over time, each bidder tries to learn her payoff-maximizing bidding strategy.
Bidders can be modeled with different memory capabilities (myopic, perfect
memory or in between with truncated Fictitious Play).
The simulation results confirm the mathematical predictions that were
derived in [7]. Bidders in 1PAs under Best Response dynamics show excess
volatility of strategies and therefore also of seller returns if bidders don’t have
perfect memory. Each bidder’s Best Response depends on all the other bid-
ders’ strategies and the learning process doesn’t converge to the NE because
too much stochasticity is introduced by the random value realizations. The
permanent mutual adaptation of bidding strategies generates a rich internal
dynamics. 2 bidders under myopic play lock in mutual strategy oscillations
that are characterized by clustered volatility. As the memory strength of bid-
ders increases, extended periods of fiercer competition alternate with periods
of lower bidding. In general, bidders in 1PAs pursue heterogeneous bidding
10
For FS I would have to determine a strategy range in the beginning of the simu-
lation.
222 Konrad Richter
strategies. This heterogeneity leads to a suboptimal allocation of goods be-

cause the bidder with the highest value does not necessarily obtain the auc-
tioned good.
In contrast, bidders in repeated 2PAs find the NE easily by following Best Re-
sponse strategy updating with any positive memory strength. Since bidding
strategies are homogeneous in the NE, the 2PA is in the long run maximally
efficient. The underlying reason for the convergence is that the 2PA has a
dominant strategy of bidding the true value. Therefore the Best Response
doesn’t depend on the opponents’ strategies.
In general, auction formats with dominant strategies are maximal efficient
under Best Response. For multi-unit auctions, a dominant strategy auction
format would be the Ausubel auction (see [1]) or the generalized Vickrey-
auction. For single-unit double auctions, McAfee has proposed a dominant
strategy mechanism (see [5]). The development of corresponding formats for
more general auction setups is a topic for further research.
Particularly interesting in this context are double-sided multi-unit affiliated-
value auctions since they correspond to order books in financial markets. By
switching to dominant-strategy order books, the strategic mutual adaptation
of traders’ bidding strategies could be minimized. The resulting reduction of
the volatility of asset prices could benefit the world economy since it leads to
a more efficient capital allocation and to less investment risk.
Further promising fields for auction redesign include order books for elec-
tricity markets, the restructuring of supply chains, IPOs, treasury bill auc-
tions, procurement auctions and many more.
Simulation can be a powerful tool in the search for more efficient auction
formats. In the future I want to extend therefore the Auction Simulator to
allow the simulation of a wider range of auction formats, for non-linear bidding
strategies and for interdependent values.
Table 5. Abbreviations
Abbreviation Meaning
1PA First-price Auction
2PA Second-price Auction
AS Auction Simulator
BR Best Response
mBR Myopic Best Response
pBR Perfect Memory Best Response
tFP truncated Fictitious Play
References
1. L. M. Ausubel. An efficient ascending-bid auction for multiple objects.
forthcoming in American Economic Review, Working Paper Version at
www.ausubel.com/auction-papers/efficient-ascending-auction-r.pdf, 2004.
2. L. E. Blume. The statistical mechanics of best-response strategy revision.
Games and Economic Behavior, 11:111–145, 1995.
3. S. Hon-Snir and A. Sela. A learning approach to auctions. Journal of Economic
Theory, 82:65–88, 1998.
4. M. Kandori, G. J. Mailath, and R. Rob. Learning, mutation and long run
equilibria in games. Econometrica, 61(1):29–56, 1993.
5. P. McAfee. A dominant strategy double auction. Journal of Economic Theory,
6:434–450, 1992.
6. R. B. Myerson. Optimal auction design. Mathematics of Operations Research,
6:58–73, 1981.
7. K. Richter. Revenue equivalence revisited. Workingpaper; mail to kon-
rad richter@mckinsey.com respectively fidibus1972@yahoo.de, 2004.
8. J. G. Riley and W. F. Samuelson. Optimal auctions. American Economic
Review, 71:381–392, 1981.
9. W. Vickrey. Counterspeculation, auctions and competitive selaed tenders. Jour-
nal of Finance, 16(1):8–37, 1961.
10. E. Wolfstetter. Third- and higher-price auctions. Working Paper; www.wiwi.hu-
berlin.de/wt1/papers/1996/third and higher price auctions.pdf, 1996.
11. H. P. Young. The evolution of conventions. Econometrica, 61(1):57–84, 1993.
Counterclockwise Behavior Around the
Beveridge Curve
Koji Yokota
Otaru University of Commerce yokota@res.otaru-uc.ac.jp
It is an established fact that the time-series plot on the uv plane moves around
the Beveridge curve. We show that such a behavior is pertinent to the match-
ing process under imperfect information.
1 Introduction
It is natural to consider that stable matching will arise as an equilibrium when
agents in the matching market have perfect information about other agents.
The original setting of this problem by [6], called stable marriage problem,
is that there are two groups of agents, men and women, each having prefer-
ence over the agents in the other group and being single. The objective of
this problem is to make matches between men and women, one by one, in
a sustainable way. Namely, the matching, i.e. a pattern of matches, should
satisfy such a property that, in equilibrium, there should be no agents who
are not matched as a pair but both of whom have incentive to form a new pair
changing from the current partner. We call such a matching stable matching.
In the current literature, it is assumed that not only agents in the matching
market but also their preferences are common knowledge. [3] introduced a
notion of equilibrium, called bargaining equilibrium, when utility is transfer-
able between agents in a matched pair. This transferable utility case applies
to the matching in the labour market because part of utility of an agent in a
matched pair can be transferred to the other in the form of wage payments. In
the bargaining equilibrium, a pair equally divides the surplus of total payoff
of this pair over the sum of their threats and distribute it on each threat. The
threat is the payoff the agent receives in a new matching once the original pair
becomes unavailable. Since we need to know the threat in the new matching
game, we need to calculate threat recursively until we reach to a trivial match-
ing game in which only one pair is available. The above operation determines
the pattern of utility transfer within all possible pairs, which transforms the
utility profile of each agent when utility is not transferable into another utility
226 Koji Yokota
profile. The bargaining equilibrium is a stable matching equilibrium upon this

transformed utility.
Then, the question arises: can the bargaining equilibrium explain the data
observed in the labour market? The answer is unfavourable. [4] and [12]
showed that, under certain probability distributions that generate evaluation
of the counterpart, such a matching process leads to complete matching when
the market becomes large in size. It implies that there should be no observed
prolonged unemployment. [4] assumed that the utility that an agent feels from
a potential partner is drawn from a special case of extreme value distribution.
The utility to be single is also drawn from the same distribution but multi-
plied by a scaling factor. The great contribution of [4] is that he obtained an
asymptotic but analytical formula for the number of matches. As far as only
number of matches concerns, the assumption on the probability distribution
is harmless because what matters is the preference order of each agent and
the level of utility does not affect the pattern of matching. However, the as-
sumption that utility of being single is random may be arguable. In search
context, the factors which affect the decision to stay single are the expected
utility to match in the next period, unemployment benefits and intertempo-
ral discount rate, which are all not random. However, [12] confirmed that the
same result holds for the cases where the utility of being single is constant. He
also found that the convergence to complete matching occurs fairly quickly
and under empirically conceivable size of matching market, all agents in the
smaller group should be matched.
These stochastic models assume independence of probability distributions
from which evaluation of an agent by another is drawn. Since utility is trans-
ferred within a pair in the form of wages, this is a fictitious story under the
job matching situation. [13] showed that the bilateral utilities within any pair
after the transfer of utility will exhibit positive correlation. It has the effect to
raise the matching probability and it only strengthens the tendency to reject
the hypothesis that job matching is described by a stable matching process.
All the above results suggest that some kind of disruption of perfect in-
formation plays a key role in job matching. Therefore, we want to see what
happens if searching agents cannot find others costlessly. We will find that the
matching process which arises from there brings prolonged unemployment and
unfilled vacancies, constant returns to scale of the aggregate matching function
and counterclockwise trajectory on the uv plane. Traditionally, it is argued
that deterioration of matching efficiency are the cause of the shift of the Bev-
eridge curve after 1970’s. Our result places precautious warning to interpret
the position of estimated Beveridge curve directly as an indicator of efficiency.
The same mechanism which makes the trajectory in the uv plane go around
the Beveridge curve causes apparent “shift” of the Beveridge curve without
any deterioration of matching efficiency. There are empirical studies which
cast doubt on the validity of the Beveridge curve to represent structural or
stable relationship in the labour market. Using the panel data of nine regions
in West Germany, [1] showed that the observed shift of the Beveridge curve oc-
Counterclockwise Behavior Around the Beveridge Curve 227
curred too quickly to be explained by structural changes and that the change
in the composition of unemployment pool, i.e. structural factors on the side
of workers, have little explanatory power of the shifts. He raised a question
on the stability of the Beveridge curve under the presence of business cycles.
He pointed out the possibility that, after a shock, the trajectory on the uv
plane may not converge back to the original Beveridge curve in general, since
parameters change. In the same spirit, we will find that, given the existence
of fluctuation in inflows to the labour market, the apparent shift of the Bev-
eridge curve is inevitable. In our case, what makes the trajectory converge
to a different point is the fact that the market is in a different phase of the
inflow cycle when the market crosses the same vu ratio from above and below.
It makes the trajectory on the uv plane form a cycle, especially in the coun-
terclockwise direction. Counterclockwise behaviour of the trajectory has been
found in a number of countries. [7] pointed out the cyclical behaviour from
the data of United Kingdom from 1953 to 1964. He explained the emergence
of the cyclical behaviour using a model, but failed to show why the direction
of the cycle is counterclockwise, not clockwise. To explain the direction of the
cycle, [2] assumed that vacancies responds more quickly to shocks than un-
employment. [11] explained it from the asymmetry between matching process
and breakup process. He assumed that breakup of a job-worker pair can be
undertaken immediately, whereas matching is a time-consuming process. We
will show in this paper that the counterclockwise behaviour can arise solely
from the matching process regardless the breakup process. It comes from the
time lag until change in flows (inflow of jobs and workers) is reflected in stocks
(u and v).
2 The Model
Now, we consider such a matching process that firms advertise vacant jobs
and workers apply to a job responding to the advertisement. A worker applies
to his best choice among advertising firms at the beginning of a period. At
the end of the period, he receives a notification of appointment or rejection. If
he is appointed, he leaves the labour market and engages in production from
the next period. If he is rejected, he repeats the same application process in
the next period. For simplicity, we put a memoryless assumption. Namely, a
worker who has rejected by a firm can apply to the same job next period if
the vacancy remains unfilled. This simplification assumption is harmless when
the market size is sufficiently large. Or, it can be thought as a case that there
is a possibility that the change of environment by passage of time can change
the evaluation of the worker. A firm will assign one of the applicants to a
job if and only if the evaluation of the best applicant exceeds the threshold
value. Job-seekers will apply if and only if his evaluation of the best job posted
exceeds his threshold value. We express both evaluation in terms of utility,
denoting a firm’s evaluation of a worker by x and a worker’s evaluation of a
228 Koji Yokota
job by y. The threshold value of a firm is denoted by x∗ and that of a worker

is denoted by y ∗ . If the evaluation of the best applicant by a firm is x ≥ x∗ ,
it will employ him. If x ≤ x∗ , it will employ nobody and leave the position
vacant for another period. The same holds for a worker. We assume that each
agent is risk neutral. x and y are random variables that have i.i.d. probability
distributions F (·) and G(·) respectively. We denote the number of unfilled job
vacancies in the labour market by v and that of job-seekers by u.
Proposition 1. There are v vacancies and u job-seekers in the labour market.

On each encounter, the evaluation of a worker by a firm x is drawn from an
i.i.d. distribution F and the evaluation of a firm by a worker y is drawn from
an i.i.d. distribution G. The threshold evaluation of a firm and a worker are
x∗ and y ∗ respectively. Then, the probability for a vacancy to match is given
by1 v u
(1 − F (x∗ )) (1 − G (y ∗ ) )
1− 1− . (1)
v
Proof. The probability that the evaluation of a job by a worker is above the
threshold to apply is 1 − G (y ∗ ). When there are v job vacancies, a worker will
apply to one of them with probability 1 − G (y ∗ ) . Therefore, when there are
v
u job-seekers, the probability that exactly m job-seekers apply to one of the

u v m
[1 − G (y ∗ ) ] G (y ∗ )
v(u−m)
vacancies is m =: Pr(A). On the other hand,
each vacancy is homogeneous ex ante. Once having m effective job-seeking
applicants in the labour market (u − m workers do not apply), the probability
n m−n
that n workers apply to a particular vacancy is m n (1/v) (1 − 1/v) =:
Pr(B). The firm assigns one of those n applicants who is ranked top, if his
evaluation exceeds the threshold. The firm approves a worker with probability
1 − F (x∗ ) =: Pr(C).
n
The event that m job-seekers actually apply and n among them apply to
a particular firm is exclusive each other for different (m, n) pairs. Thus, the
probability that a firm assigns a job to a worker is given by

u
m v u
(1 − F (x∗ )) (1 − G (y ∗ ) )
Pr(A) Pr(B) Pr(C) = 1 − 1 − (2)
m=1 n=1
v
∗ n
second line comes from the fact that 1 − F (x ) = 0 when n = 0 and
The
m
n = 1 when m = 0.
Corollary 1. Suppose that vu ratio θ is given. The probability that a firm

∗
receives applications and rejects all of them is e−(1−F (x ))/θ − e−1/θ > 0 when
market size is large.
u m
Proof. The probability is calculated by m=1 n=1 Pr(A) Pr(B) Pr(C̄). Tak-
∗
ing v/u = θ into account, its limit as v → ∞ is e−(1−F (x ))/θ − e−1/θ .
1
The probability when F (x∗ ) = 0 is obtained by F (x∗ ) → +0.
Proposition 2. Suppose u = 0 and v = 0. If G (y ∗ ) = 1, then for any given

vu ratio θ, the probability that there exist workers who voluntarily do not apply
to any jobs converges to zero as u and v go infinity.
v u
Proof. The probability that all workers apply to a vacancy is [1 − G (y ∗ ) ] .
We define a notation r := G (y ∗ ) . As u and v go infinity keeping θ constant,
θ
the probability that all workers apply to a vacancy becomes

) u u
)
∗ v u)
= lim 1 − G (y ∗ )
θ
lim [1 − G (y ) ] ) =1
u,v→+∞ u→+∞
v/u=θ
As the number of vacancies increases, workers have more and more choices,
making it easier to find a satisfactory candidate. However, as the number of
workers increases, there are more and more chances that at least one worker
cannot find an appropriate job. This proposition says that the former effect is
larger. When the labour market is sufficiently large so that it allows a worker
to have sufficiently many choices, a job-seeker will not give up application
voluntarily. This makes sharp contrast to Corollary 1. There always exist
firms which voluntarily reject all applicants.
Theorem 1. When vu ratio θ is given, matching probability of vacancies Pv

is asymptotically given by

1 − F (x∗ )
Pv = 1 − exp − (3)
θ
as u and v go infinity. Using the asymptotic matching probability of vacancy,

the aggregate matching function µ is given by
u
µ (u, v; F (x∗ )) = Pv v = 1 − exp − [1 − F (x∗ )] v (4)
v
Proof. The result is obtained by taking a limit u, v → +∞ keeping θ constant

in Proposition 2.
In Proposition 1, note that Pu and Pv depend on θ, but not on the value

of u or v themselves. The following proposition shows that most of ordinary
presumption about an aggregate matching function are supported but not all.
We can easily show the following result.
Proposition 3. The asymptotic aggregate matching function µ (u, v) has the

following properties:
1. µ is homogeneous of degree one,

2. µ (u, 0) = µ (0, v) = 0
3. m < min (u, v)
230 Koji Yokota
∂µ ∂µ ∂2µ
4. > 0, > 0, <0
∂u ∂v ∂u2
2
∂ µ u 2
5. < 0 if and only if >
∂v 2 v 1 − F (x∗ )
Since most of aggregate matching models are based on the assumption of ho-
mogeneity of degree one of the matching function, property (1) of Proposition
3 may be of special interest. It gives a justification of such a presumption. It
excludes the possibility of multiple equilibria caused by increasing returns to
scale of the matching function.
3 Inflow Cycles
Empirical data show that inflows of new job-seekers and vacancies to the
labour market fluctuate persistently and fairly regularly along with business
cycles. I said “fairly regularly” but in general, it is hopeless for the inflows to
have strictly regular cycles. Nevertheless, it is still useful to focus on a case in
which inflows have regular cycles, since what is important is the recurrence
nature of cycles. We assume inflows η and ξ to fluctuate and to have same
averages over a cycle. They are balanced in long run although the length of
a cycle may not coincide. Then, a trajectory on the uv plane beginning from
any initial point converges to a unique cycle, the period of which is the least
common multiple of the period of vacancy inflow and the period of job-seeker
inflow. We show the following lemma without a proof.
!
Lemma 1. Let T(0) , T(1) , . . . , T(k−1) be a sequence of contraction map-
pings with modulus βκ ∈ (0, 1) (κ = 0, 1, . . . , k − 1) from a Banach space X
to itself. For any initial value x0 ∈ X, we define a dynamics by
xt+1 = T(t mod k) xt . (t = 1, 2, . . . )

!k−1
Then, the sequence T(κ) κ=1 has a unique and globally stable cycle {x̄(0) , x̄(1) ,
. . . , x̄(k−1) } ∈ X k where x̄(κ) (κ = 0, 1, . . . , k − 1) is a fixed point of the map-
ping T(κ−1) ◦ T(κ−2) ◦ · · · T(0) ◦ T(k) ◦ T(k−1) ◦ · · · ◦ T(κ) .
We denote the state of the labour market at the beginning of each period
t = 0, 1, 2, . . . by (ut , vt ) ∈ R2++ where ut is the number of job-seekers and
vt is the number of unfilled vacancies in the labour market. The initial value
(u0 , v0 ) is arbitrarily given. At the beginning of each period, firms advertise
their vacancies and job-seekers apply to a vacancy which they think best.
Successfully matched mt pairs of a job-seeker and a vacancy exit the market
and the rest remains in the market. At the end of the period, new inflow
of job-seekers ξt and new inflow of vacancies ηt are added to the remaining
stocks, forming the stock of job-seekers and vacancies (ut+1 , vt+1 ) in the next
period. We denote by LCM(x, y) the least common multiple of integers x and
y.
Theorem 2. Suppose that the inflow of new vacancies into the labour mar-
ket has a cycle of! period S ∈ N. The inflows with in a cycle are given by
η 0 , η 1 , . . . , η S−1 ∈ RS+ . The inflow of new job-seekers into the labour mar-
ket has a cycle of ! period T ∈ N. The inflows within a cycle are given by
ξ 0 , ξ 1 , . . . , ξ T −1 ∈ RT+ . Then, if the condition
S−1 T −1
i=0 η (i) i=0 ξ (i)
= (5)
S T
is satisfied, a trajectory on the uv plane beginning from any initial value
(u0 , v0 ) ∈ R2+ converges to a cycle of period LCM (S, T ).
Proof. The transition of (u, v) plot is given by
vt+1 = vt + η (t mod S) − mt (6)

ut+1 = ut + ξ (t mod T ) − mt (7)
for any t = 0, 1, . . .. Using the condition (5), we obtain ut+M = ut −vt +vt+M .
Recursive operation gives
uκ+M n = uκ − vκ + vκ+M n . (8)
for any κ = 0, 1, . . . , M − 1 and for any n ∈ N. Namely, the data picked

up from time series with interval M stay on the same line. As obvious from
derivation, it is one of isoinflux lines.
From equation (4), for any t,
*
vt mod M − ut mod M ∗
vt+1 = η (t mod S)
+exp − 1 − 1 − F (x ) vt (9)
vt
where vt mod − ut mod M is a constant obtained by vt mod M − ut mod M =
tMmod M −1 (i)
v0 − u0 + i=0 η − ξ (i) .
Equations (8) and (9) describe the dynamics. Since (8) is continuous, u con-
verges as v does. Defining the right-hand side of (9) by T(t mod M ) vt , we have
dT(t mod M ) vt /dvt ⊂ (0, exp{−[1−F (x∗ )]} ⊂ (0, 1). For any a, b ∈ R++ , there
exists c ∈ (a, b) such that T(t mod M ) a−T(t mod M ) b = (dT(t mod M ) c/dc)(a−b)
by the mean-value theorem. However, 0 < c < exp {−(1 − F (x∗ ))} < 1. Thus,
for any a and b, |T(t mod M ) a − T(t mod M ) b| < β|a − b| where 0 < β < 1,
i.e. T(t mod M ) is a contraction mapping. From Lemma 1, the sequence of
!M −1 !
mappings T(κ) κ=0 has a unique cycle v̄ (0) , v̄ (1) , . . . , v̄ (M −1) ∈ RM+ and
the paths from any initial value (u0 , v0 ) converge to it. Furthermore, v̄ (κ)
(κ = 0, 1, . . . , M − 1) is a fixed point of T(κ) . This proves that the asymptotic
cycle has a period no greater than M . However, the period of the cycle should
be exactly M as shown below.
Suppose that the period of the cycle is M < M . Then, M must be a
divisor of M . Otherwise, it contradicts the above fact that any v (mod M )
share the same fixed point. For such M to exist,
232 Koji Yokota

v̄ (0) , v̄ (1) , . . . , v̄ (M −1) = v̄ (M ) , v̄ (M +1) , . . . , v̄ (2M −1)
and

ū(0) , ū(1) , . . . , ū(M −1) = ū(M ) , ū(M +1) , . . . , ū(2M −1)
where ū(n) := u(n) − v(n) + v̄ (n) for n = 0, 1, . . . , M − 1 must hold. It implies

that u(n) − v(n) =
u(M +n) − v(M +n) for given n. From equation (9), for
(n+1) (n+1)
ū , v̄ = ū(M +n+1) , v̄ (M +n+1) to hold, the condition η (n) =

η (M must be satisfied. On the other hand, u(n) − v(n) = u(M +n) − v(M +n)
+n)
M −1 (i) M −1 (i)
requires i=0 η =
i=0 ξ for alln, which implies ξ (n) = ξ (M +n) . The
(n) (n)
result η , ξ = η (M +n) , ξ (M +n) contradicts the fact that M is the
least common multiple of S and T .
It is an established fact that the trajectory on the uv plane empirically moves
around the Beveridge curve. During a recession, u gradually increases whereas
v decreases. Thus, the trajectory moves to the lower right of the uv plane. On
recovery, opposite pressure carries the trajectory to the upper right. However,
when recovery takes place, u does not recover quickly whereas v begins to in-
crease. Thus, the trajectory during a recovery phase goes above the one during
a recession phase. Often, this asymmetric response of the labour market to
business cycles is attributed to loss of skills of unemployed workers. However,
Theorem 2 shows that cycle in the uv plane can emerge even without the
existence of loss of skills. Only cyclical fluctuations of new vacancies and new
job-seekers causes the emergence of a cycle in the uv plane.
Theorem 2 does not tell anything about the shape of the cycle. Es-
pecially, it does not tell to which direction the cycle circulates. We will
study conditions that bring counterclockwise trajectory on the uv plane
now. The dynamics when inflows periodical fluctuates are given by equa-
tion (9). Note that T(t mod M ) depends only on v(t mod M ) − u(t mod M ) and
η (t mod M ) . v(t mod M ) − u(t mod M ) determines the position of the isoinflux
line
to which sequence u (t!mod M ) , v(t mod M ) , u2(t mod M ) , v2(t mod M ) ,
u3(t mod M ) , v3(t mod M ) , . . . belongs to. The initial point in the sequence
v(t mod M ) − u(t mod M ) is calculated by the equation
M
t mod

v(t mod M ) − u(t mod M ) = v0 − u0 + η (i) − ξ (i) . (10)
i=0
Once the isoinflux line is determined above, T(t mod M ) solely depends on
η (t mod M ) .
Consider a simple example that the inflow of vacancies is fixed at η = 1 and
the inflow of job-seekers follows ξ = 1 + sin t. This is a good example to begin
with since constant η makes mapping T(t mod M ) fixed for all periods. Note
that it has balanced inflows. Suppose that there are four matching sessions
within a cycle of the inflows as in the figure below.
The matching market opens at points A, B, C, D and E. We assume that

initial point is balanced: v = u. We plot the status of the matching market
that corresponds to points A-E in Figure 1 that shows recursive application
of functions . We label corresponding plots in the figure as points A-E as well.
Point A in the uv plane (it represents all periods 0 mod 4) is on the 45 degree
line v = u. At point B (period 1 (mod 4)), job-seeker inflow ξ (1) is larger
than vacancy inflow η (1) by one. Thus, the labour market is on isoinflux line
v = u − 1. At point C (period 2 (mod 4)), inflows are balanced. It does not
change the isoinflux line the point belong to, so point C stays on the same
line as point B. At point D (period 3 (mod 4)), vacancy inflow ξ (3) is less than
job-seeker inflow η (3) by one, thus the isoinflux line shifts up by one returning
back to the 45 degree line. In sum, point A and D and point B and C are on
the same isoinflux line respectively. We use the following small lemma.
Lemma 2. Let X ⊂ R. If T : X → X is a contraction mapping, then Tx − x

is x a strictly decreasing function.
Proof. If Tx − x is not a decreasing function of x, then there exist a, b ∈ X

such that a > b and Ta − a ≥ Tb − b. However, it implies Ta − Tb ≥ a − b.
Since a > b, it means Ta − Tb ≥ a − b. This contradicts the fact that T
is a contraction mapping. Therefore, the lemma holds.
We characterise the limit cycle by the following proposition.

!
Proposition 4. Let T = T(0) , T(1) , . . . , T(k−1) be a sequence of contraction
mappings from X ⊂ R to itself with modulus βκ ∈ (0, 1) (κ = 0, 1, . . . , k − 1).
Suppose each mapping is an increasing function. Consider a dynamics defined
by
xt+1 = T(t mod k) xt , (t = 1, 2, . . . )
for any initial value x0 ∈ X. Let ȳ (κ) (κ = 0, 1, . . . , k − 1) be a fixed point of
T(κ) and at least for one pair of i, j = 0, 1, . . . , k − 1 (i = j), ȳ (i) = ȳ (j) .
! !k−1
Then the unique stable limit cycle x̄(0) , x̄(1) , . . . , x̄(k−1) ∈ X k of T(κ) κ=1
satisfies the condition
min ȳ (i) < x̄(κ) < max ȳ (i)

i=0,1,...,k−1 i=0,1,...,k−1
where κ = 0, 1, . . . , k − 1.
234 Koji Yokota
Proof. If we show mini=0,1,...,k−1 ȳ (i) < x̄(κ) (κ = 0, 1, . . . , k − 1), then

x̄(κ) < maxi=0,1,...,k−1 ȳ (i) can be shown by a similar argument.
We define ȳ (p) := mini=0,1,...,k−1 ȳ (i) and
(p)denote
by T(p) the operator that
has ȳ as a fixed point. Also denote by T T
(p)
a composite operator T(p−1) ◦

T(p−2) ◦· · ·◦T(0) ◦T(k−1) ◦T(k−2) ◦· · ·◦T(p) . Suppose that x̂ ∈ −∞, ȳ (p) ∩X is a
n
fixed point of T T(p) . Then, for arbitrary n ∈ Z+ , T(p) x̂ ∈ x̂, ȳ (p) holds.
If the fixed
(p)point
n ȳ
(p+n)
(mod M ) of T(p+n) (mod M ) satisfy ȳ (p+n) > ȳ (p) ,
T (p+n)
◦ T x̂ > x̂ holds by Lemma 2. Since any operator T(i) where i = q
n ȳ ≥ ȳ , x̃ := T(p−2) ◦· · ·◦T(0) ◦T(k−1) ◦T(k−2) ◦· · ·◦T(p) x̂ ≥
(i) (p)
has a fixed point
(p) (p)
T (p+n)
◦ T x̂ > x̂. For ȳ to be a fixed point of T T
(p)
, there must exist
T̃ ∈ T such that x̂ = T̃x̃ holds. It means T̃x̂ − x̂ = T̃x̂ − T̃x̃ < 0. Then, from
Lemma 2, the fixed point ỹ of T̃ satisfies ỹ < x̂ ≤ ȳ (p). This contradicts to
the definition of ȳ (p) . Therefore, in the region −∞, ȳ (p) ∩ X, the fixed point

of T T(p) cannot exist. By a similar argument, it is shown that it does not
exist in the region max ȳ (i) , +∞ ∩ X.
i=0,1,...,k−1
In our example, there are two contraction mappings. We denote them by T

and T. T operates on points B and C and T operates on points A and D.
In general, T(t mod M ) is a function of η (t mod M ) and v(t mod M ) − u(t mod M ) .
Since η (t mod

M)
is fixed for the moment,
let us write the operator T(t mod M ) as
T(t mod M ) v(t mod M ) − u(t mod M ) to be explicit on the parameter. Consider
two cases T(t mod M ) (K1 ) and T(t mod M ) (K2 ) where K1 > K2 . Suppose vt =
0. From (9), we can show
*
T(t mod M ) (K1 ) vt K1 − K 2
= exp 1 − F (x∗ ) > 1.
T(t mod M ) (K2 ) vt vt
Thus, T(t mod M ) (K1 ) vt > T(t mod M ) (K2 ) vt for fixed η. This fact implies
T > T. It is shown in Figure 1(a).
By Proposition 4, all points on the limit cycle is between y and y. Thus,
points A and D are on the left-hand side of the fixed point of T. Point A is
obtained by operating T twice on v4 in the graph whereas point D is obtained
by operating it only once. Thus, point A is closer to fixed point y. By the
same reason, point C is closer to fixed point y than point B. Therefore, v1 >
v4 and v2 > v3 . The dynamics of v is translated to the dynamics of u by
isoinflux lines. The trajectory in the uv plane is shown in Figure 1(b). Point
A’ corresponds to point A in the previous figure and so on. As easily observed,
the direction of the cycle is opposite to the empirical cycle. This is obvious
from equation 9. If v − u is greater, the slope of equation 9 is larger. Since
η is fixed, it implies that the corresponding mapping T is located above of
other equations that have smaller v − u. If v − u is higher in period t than
period s, it means that the point of the trajectory at period t is on an isoinflux
line which is upper-left of the isoinflux line at period s. When the position of
Fig. 1. Dynamics of limit cycles
(a) Dynamics of limit cy- (b) Trajectory in the uv

cle plane
an isoinflux line is relatively above of other isoinflux lines that the trajectory
visits, then v in next period will be larger than the current period. Thus, in
the uv plane, as the trajectory goes to upper left (i.e. visits an isoinflux line
upper-left) the path tends to shift outward. Note that only relative position
among related isoinflux lines is crucial in determining the direction of the
shift. The absolute position of the isoinflux line, namely the value of v − u,
is not related. This observation suggests that the fluctuation of vacancies η is
necessary to generate an counterclockwise behaviour of the trajectory in the
actual data. Therefore, we take another simple example in which η fluctuates.
Now, suppose that vacancy inflow fluctuates following η = 1 + sin (t + π)
instead of η = 1 in the previous example. This pattern of ξ and η is similar to
the data observed in the actual labour market. Figure 2 shows the limit cycle
of v and the same movement in the uv plane. The cycle in Figure 2 (b) shows
a counterclockwise movement. This result is more general. We can provide a
sufficient condition for the emergence of counterclockwise trajectory in the uv
plane. We use the following lemma, which we show without a proof.
Lemma 3. Let X ⊂ R. Consider a family of contraction mappings T(η) :
X → X which has η ∈ R as a parameter. Suppose that T(η) is continuous on
the parameter and if η < η then T(η) < T(η ). Suppose η is a continuous func-
tion of t ∈ R and there exists T ∈ R such that for all n ∈ Z, η(t + nT ) = η(t)
Pick up s points ti ∈ [0, T ) where i = 0, 1, . . . , s − 1 and consider a dynamics
defined by xk+1 = T (η (tk mod s )) xk . Then, it has a unique stable limit cycle
(0) (1) (s−1)
{x̄s , x̄s , . . . , x̄s } and if s increases such that max |ti+1 − ti | → 0 holds,
(i) (i) (i) (i)
then x̄s → ȳs holds where ȳs = T (η (ti )) ȳs .
Theorem 3. Suppose that η and ξ are time continuous and dη/dt 0 ⇔
dξ/dt 0. Then, if matching sessions are held sufficiently frequently within a
236 Koji Yokota
Fig. 2. More realistic case
(a) Limit cycle in dynam- (b) Beveridge cycle

ics of v
cycle of η and ξ, then the point that the trajectory on the uv plane crosses a
particular isoinflux line from the right comes upper right of the point that the
path crosses it from the left in the uv plane.
Proof. Since η is time continuous, for any ε > 0 and t, if the interval of
the matching session is made
) (t mod ) sufficiently short, then it is possible to make
)η M +1)
− η (t mod M ) ) < ε. That is,
*
v(t mod M ) − u(t mod M )
η − ε + exp − 1 − 1 − F (x∗ ) vt
vt
*
v(t mod M ) − u(t mod M )
< vt+1 < η + ε + exp − 1 − 1 − F (x∗ ) vt .
vt
t mod M (i)
From v(t mod M ) − u(t mod M ) = v0 − u0 + i=1 η − ξ (i) and dη/dt
0 ⇔ dξ/dt 0, the differentiation by η (t mod M ) of the left-most or the right-
most hand side of the above equation gives
*
∗ v(t mod M ) − u(t mod M ) ∗
1+(1 − F (x )) Kt exp − 1 − 1 − F (x ) >0
vt
where Kt is a positive constant. Therefore, taking ε sufficiently small, T can
be ordered by η (t mod M ) . Namely, if η1 > η2 , then T (η1 ) > T (η2 ).
Now, suppose that the uv path passes over the isoinflux lines v − u = k1
and v − u = k2 where k1 < k2 , and it is currently on the line v − u = k
where k1 < k < k2 after passing over v − u = k2 for the first time. Then
η (t mod M ) − ξ (t mod M ) < 0. Since η and ξ cyclically fluctuates, the path must
return to a point on v − u = k2 from Theorem 2. It means that the path
crosses v − u = k with η − ξ > 0. From dη/dt 0 ⇔ dξ/dt 0, the η in the
latter case has larger value. From Proposition 3, it implies that, if matching
session is held frequently enough, then the fixed point of v in the latter case
must be larger than the former case.
The theorem says that when the trajectory forms a simple cycle, namely when
it crosses any isoinflux line at most twice within a cycle, then the direction
of such a cycle is counterclockwise under the conditions provided by the the-
orem. Generally, when the path crosses any isoinflux line from the right, its
crossing point should be upper right of any points where the path crosses it
from the left, and vice versa. Note that we assumed no changes in matching
technology. The cycle is brought by the asymmetry between job-seekers and
vacancies embedded in the matching process. Figure 3 shows an example of
Fig. 3. Trajectory in the uv plane when inflows are symmetric
the trajectory in the uv plane when inflows to the labour market satisfy the
condition of Theorem 3. The case η = 1 + sin t and ξ = 1 + sin(t − π) is drawn
in the figure.
4 Conclusion
In this paper, we studied a matching process which arises when there is a cost
to collect information about agents in the labour market. We found that its
implication is consistent with data in many ways including the behaviour on
the uv plane. The main reason that the behaviour on the uv plane becomes
not trivial is that the it plots two stock variables. It is considered that wage
rate under the existence of labour market friction becomes a function of the
uv ratio. Since a shock does not reflect immediately to the uv ratio, wage rate
responds slowly. It deteriorates the adjustment power of wage rate between
imbalance of labour supply and demand.
References
1. Axel H. Börsch-Supan (1991) Panel data analysis of the beveridge curve: Is
there a macroeconomic relation between the rate of unemployment and the
vacancy rate?, Economica 58:279–97
238 Koji Yokota
2. Robert J. Bowden (1980) On the existence and secular stability of u – v loci,

Economica 47:33–50
3. Vincent P. Crawford, Elsie Marie Knoer (1981) Job matching with heteroge-
neous firms and workers, Econometrica 49:437–450
4. John K. Dagsvik (2000) Aggregation in matching markets, International Eco-
nomic Review 41(1):27–57
5. Fitoussi, Jean-Paul, Jestaz, David and Phelps, Edmund S. and Zoega, Gylfi
(2000) Roots of the Recent Recoveries: Labor Reforms or Private Sector
Forces?, Brookings Papers on Economic Activity 1:237-311
6. David Gale, Lloyd S. Shapley (1962) College admissions and the stability of
marriage, American Mathematics Monthly 69:9–15
7. Bent Hansen (1970) Excess demand, unemployment, vacancies, and wages,
Quarterly Journal of Economics 84(1):1–23
8. Nickell, Stephan J., Layard, Richard (1999) Labor Market Institutions and
Economic Performance, in Ashenfelter, O. and Card, D. (eds) Handbook of
Labor Economics 3029-3084 North-Holland Amsterdam
9. Nickell, Stephan J., van Ours, Jan C. (2000) Why Has Unemployment in The
Netherlands And the United Kingdom Fallen So Much?, Canadian Public Policy
26:201-220
10. Nickel Stephen, J., van Ours, Jan C. (2000) The Netherlands and the United
Kingdom: A European Unemployment Miracle?, Economic Policy 30:137-175
11. Pissarides, Christopher A. (1985) Taxes, Subsidies and Equilibrium Unemploy-
ment Review of Economic Studies 52:121-133
12. Koji Yokota (2001) Competition between markets: Why can matchmakers sur-
vive in the temporary help service industry?
13. Koji Yokota (2004) Aggregate properties of stable matching
The Waiting-Time Distribution of Trading
Activity in a Double Auction Artificial
Financial Market
Silvano Cincotti1 , Sergio M. Focardi2 , Linda Ponta1 , Marco Raberto1 , and

Enrico Scalas3
1
DIBE, Università di Genova, Via Opera Pia 11a, 16145 Genova, Italy
2
The Intertek Group, rue de Javel 94, 75015 Paris, France
3
DISTA, Università del Piemonte Orientale, Piazza Ambrosoli 5, 15100
Alessandria, Italy
Summary. In this paper, the statistical properties of high-frequency data are inves-
tigated by means of computational experiments performed with the Genoa Artificial
Stock Market (Raberto et al. 2001, 2003, 2004). In the market model, heterogeneous
agents trade one risky asset in exchange for cash. Agents have zero intelligence and
issue random limit or market orders depending on their budget constraints. The
price is cleared by means of a limit order book. The order generation is modelled
with a renewal process where the distribution of waiting times between two consec-
utive orders is a Weibull distribution. This hypothesis is based on recent empirical
investigation made on high-frequency financial data (Mainardi et al. 2000, Raberto
et al. 2002, Scalas et al. 2003). We investigate how the statistical properties of prices
and of waiting times between transactions are affected by the particular renewal pro-
cess chosen for orders. Results point out that the mechanism of the limit order book
is able to recover fat tails in the distribution of price returns without ad-hoc behav-
ioral assumptions regarding agents; moreover, the kurtosis of the return distribution
depends also on the renewal process chosen for orders. As regarding the renewal pro-
cess underlying trades, in the case of exponentially distributed order waiting times,
also trade waiting times are exponentially distributed. Conversely, if order waiting
times follow a Weibull, the same does not hold for trade waiting times.
Introduction
In this paper, we present a model of trading in the Genoa Artificial Stock
Market [4, 10, 12, 14, 15] characterized by a double auction clearing mecha-
nism, i.e., the limit order book [3, 13]. The limit order book is a snapshot at
a given instant of the queues of all buy limit orders and sell limit orders, with
their respective price and volume. Limit orders are organized in ascending

Corresponding author. E-mail: cincotti@dibe.unige.it
240 S. Cincotti et al.
order according to their limit prices. All buy limit orders are below the best
buy limit order, i.e., the buy limit order with the highest limit price (the bid
price). The best buy limit order is situated below the best sell limit order, i.e.,
the sell limit order with the lowest limit price (the ask price). All other sell
limit orders are above the best sell limit order. Orders are stored in the book.
A transaction occurs when a trader hits the quote (the bid or the ask price)
on the opposite side of the market. If a trader issues a limit order, say a sell
limit order, the order either adds to the book if its limit price is above the bid
price, or generates a trade at the bid if it is below or equal to the bid price. In
the latter case, the limit order becomes a marketable limit order or more sim-
ply a market order. Conversely, if the order is a buy limit order it becomes a
market order and is executed if its limit price is above the ask price, otherwise
it is stored in the book. Limit orders with the same limit price are prioritized
by time of submission, i.e., the oldest order has the highest priority. Order’s
execution often involves partial fills before it is completed, but partial fills do
not change the time priority.
In recent years, some studies about the statistical properties of the limit
order book have appeared in the literature [1, 2, 7, 8, 11]. An important
empirical variable is the waiting time between two consecutive transactions.
In fact, trading via the order book is asynchronous, i.e., a transaction occurs
only if a trader issues a market order. For liquid stocks, waiting times vary in
a range between some seconds to a few minutes. Raberto et al. [16] analyze the
intra-day trades of General Electric stock prices and find that waiting times
exhibit a 1-day periodicity, corresponding to the daily stock market activity,
and a survival probability distribution which is well fitted by a stretched
exponential. They also find a significative cross-correlation between waiting
times and the absolute value of log-returns.
In this paper, the effect of a more general distribution of order waiting
times is investigated. In particular, we focus our attention on the Weibull
distribution that admits the exponential distribution as a limit case. Results
show that in the case of exponentially distributed order waiting times, also
trade waiting times are exponentially distributed. Conversely, if order waiting
times follow a Weibull, the same does not hold for trade waiting times.
The paper is organized as follows: Section 1 presents the model, Section
2 shows computational experiments and Section 3 provides a discussion of
results and the conclusions of the study.
1 The Model
A model of artificial trading by means of a limit order book is presented in

this section. Agents trade one single stock in exchange for cash. They are
modelled as liquidity traders; as a consequence, the decision making process
is nearly random and depends on the finite amount of financial resources (cash
Trading Waiting Time Distribution in an Artificial Market 241
+ stocks) they own. At the beginning of the simulation, cash and stocks are
uniformly distributed among agents.
1.1 The Order Generation Process
Trading is organized in M daily sections. Each trading day is subdivided in

T elementary time steps, say seconds. During the trading day, at given time
steps th , a trader i is randomly chosen for issuing an order. Order waiting
times τhO = th − th−1 are determined according to a Weibull distribution,
whose probability density function (PDF) p(τ ) is:
β−1
β
β τ τ
p(τ ) = e− η , (1)
η η
where η is the scale parameter and β is the shape parameter, also known as
the slope, because the value of β is equal to the slope of the regressed line in
a probability plot. The expected value τ of a random variable following a
Weibull PDF is given by:

τ = η · Γ 1/β + 1 , (2)
where
∞ Γ is the Gamma function. The survival probability distribution P> (τ ) =
p(τ ) is given by:
τ β
− τη
P> (τ ) = e . (3)
The exponential distribution is a particular case of the Weibull distribution
for β = 1. In the case β < 1, the Weibull distribution assumes the form of
the so-called stretched exponential and great values of τ occur with higher
probability than in the case of β = 1. In our computational experiments, we
considered only values of the shape parameters β less or equal than one.
The order generation process is then described as a general renewal process
where the times between two consecutive orders τhO are independent and iden-
tically distributed random variables following a Weibull distribution. In the
case β = 1, the order generation process is assumed to be a Poisson process
with an exponential waiting-time distribution. For further details on renewal
processes, see Cox and Isham 1980 [5].
1.2 The Trading Decision Making Process
A trader issues a buy or a sell order with probability 50%. Let us denote with
a(th−1 ) and with d(th−1 ) the values of the ask and of the bid prices stored in
the book at time step th−1 . Let us suppose that the order issued at time step
th be a sell order, then we assume that the limit price si (th ) associated to the
sell order is given by:
si (th ) = ni (th ) · a(th−1 ) , (4)

where ni (th ) is a random draw by trader i at time step th from a Gaussian
distribution with constant mean µ = 1 and standard deviation σ. If si (th ) >
di (th−1 ), the limit order is stored in the book and no trades are recorded;
else, the order becomes a market order and a transaction occurs at the price
p(th ) = d(th−1 ). In the latter case, the sell order is partially or totally fulfilled
and the bid price is updated. The quantity of stocks offered for sale is a random
fraction of the quantity of stocks owned by the trader.
If the order is a buy order, we assume that the associated limit price bi (th )
is given by:
bi (th ) = ni (th ) · d(th−1 ) ; (5)
where ni (th ) is determined as in the previous case. If bi (th ) < a(th−1 ), the limit
order is stored in the book and no trades are recorded; otherwise the order
becomes a market order and a transaction occurs at the price p(th ) = a(th−1 ).
The quantity of stocks ordered to buy depends on cash endowment of the
trader and on the value of bi (th ).
It is worth noting that, in our framework, agents compete for the provision
of liquidity. If an agent issues a buy order, its benchmark is the best limit
buy order given by the bid price. Being µ = 1, half times, he offers a more
competitive buy order (if bi (th ) > d(th−1 )), which may result in a trade if
bi (th ) ≥ a(th−1 ). The same applies for sell limit orders.
2 Computational Experiments
The timing parameters of every simulation have been set as follows: M = 50
daily sections, each characterize by a length of T = 25, 200 s (corresponding to
7 hours of trading activity). Each simulation is characterized by a particular
value of the shape parameter β of the Weibull distribution modelling order
waiting times. We have chosen 17 different values of β ranging from 0.2 to 1
with step 0.05. The average order waiting time τ o has been set to 20 s for
every simulation. The scale factor η is so determined by the choice of β and
τ o according to Eq. 2. The orders lifespan has been set to 600 s τ o .
Sell and buy limit prices are computed following Eq. 4 and Eq. 5 respec-
tively. The random number ni (th ) is a random draw by trader i from a Gaus-
sian distribution with constant mean µ = 1 and standard deviation σ = 0.005.
The number of agents is set to 10,000. At the beginning of the simulation,

the stock price is set at 100.00 units of cash, say dollars and each trader is
endowed with an equal amount of cash and of shares of the risky stocks. These
amounts are 100,000 dollars and 1,000 shares, respectively.
Computational experiments produce realistic intraday price paths, see Ref.

[13] for further details. Log-returns r∆t have been computed in homogeneous
60
55
50
45
average of τ {s}
o
τ
40
τT
35
30
25
20
15
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
o
β of Weibull distribution for τ
Fig. 1. Average values of order waiting times and of trade waiting times as a function
of the shape parameter β of the Weibull distribution for orders.
time windows ∆t, according to the previous-tick interpolation technique [6,

pag. 37]. The time window ∆t has been set to 100 s, which is about two times
the average value of trade waiting times. Figure 1 presents the average values
of orders waiting times τ o and trade waiting times τ T , respectively, as a
function of the shape parameter β. τ o is nearly 20 s, as expected by model
construction, whereas τ T , i.e., an output of the GASM model, is around
55 s. τ T appears to be independent from β when β ≥ 0.3; the pattern of
increasing values of τ T when β < 0.3 may be due to numerical effects.
The choice of ∆t > τ T has been made in order to avoid spurious effect in
the statistical properties of returns due to long period of trading inactivity,
especially in the case of small value of β.
Figure 2 presents the values of kurtosis of log-returns as a function of
the shape parameter β. The figure shows that the distribution of log-returns
is characterized by increasing values of kurtosis as β decreases. This is the
main result of the paper. In previous works [3, 13], we already showed that
the mechanism of the limit order book was able to recover fat tails in the
distribution of log-returns without ad-hoc behavioral assumptions regarding
agents. In that cases, β was set to 1, i.e., the order generation process was
modelled as a Poisson process with exponentially distributed waiting times.
Figure 2 generalizes such a result showing that the same conclusion also holds
for a more general renewal process, i.e., Weibull distributed waiting times.
50
45
40
(∆ t = 100 s)
35
30
∆t
kurtosis of r
25
20
15
10
5
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
o
β of Weibull distribution for τ
Fig. 2. Kurtosis of the distribution of log-returns as a function of the shape param-

eter β.
Furthermore, the tails of the distribution of returns become fatter when β de-
creases from 1, i.e., the distribution of order waiting times becomes a stretched
exponential.
Figure 3 and Figure 4 show estimates of the survival probability distribu-
tion of order waiting times (dots) and of the survival probability distribution
of trade waiting times (crosses) in the case β = 0.4 (Fig. 3) and β = 1 (Fig.
4). As expected by the assumptions of the model, survival probability distri-
butions of order waiting times follow the corresponding Weibull distribution,
represented by the continuous line. The survival probability distribution of
trade waiting times is well fitted by a Weibull distribution only in the case
β = 1. Figure 3 shows that, in the case β = 0.4, the survival probability distri-
bution of trade waiting times departs from a Weibull distribution determined
from data by means of the maximum likelihood principle and represented in
the Figure with the dashed curve. In this case, the Kolmogorov-Smirnov test
rejects the null hypothesis of Weibull distribution at the significance level of
5 %. Generally speaking, we find that trade waiting times are Weibull dis-
tributed only in the case β = 1.
In the case β = 1, the process of trading is a Poisson process as the order
arrival process. An identical conclusion follows from theoretical considerations.
In fact, consider that every transaction occurs when a new order that arrives
in the book finds a matching order in the queue of orders of the opposite type.
0
10
τo
τT
−1
10
−2
10
P (τ)
>
−3
10
−4
10
−5
10
0 500 1000 1500 2000 2500 3000
τ {s}
Fig. 3. Survival probability distribution of order waiting times (dots) and of trade
waiting times (crosses) in the case β = 0.4. The two curves represent the corre-
sponding Weibull fits.
Therefore, any new order will be satisfied or not in function of the state of
the book in that moment. The state of the book varies for each moment and
for each simulation path. Given the absence of significant feedbacks in the
market, however, it is reasonable to assume that the average state of the book
is time invariant. Therefore, each incoming order will be satisfied on average
with a constant probability and the trading process can be regarded as a
random extraction from the Poisson process of orders issuing. The procedure
of random extraction from a Poisson process is called “thinning” and it is well
known that a thinning from a Poisson process is a new Poisson process; see
Ref. [5] for further details.
3 Discussion and Conclusion
In this paper, the results of Raberto et al. 2004 [13] are confirmed and
strengthened. Indeed, the return distribution is leptokurtic not only when
the order generation process is a Poisson process but also in the presence of
memory, as in the case of Weibull-distributed waiting times. Moreover, with
memory, return tails are fatter.
0
10
τo
τT
−1
10
−2
10
P (τ)
>
−3
10
−4
10
−5
10
0 100 200 300 400 500 600
τ {s}
Fig. 4. Survival probability distribution of order waiting times (dots) and of trade
waiting times (crosses) in the case β = 1 (exponential distribution). The two lines
represent the corresponding exponential fits.
This result deserves attention because previous empirical analyses [9, 16,
17] have shown that the distribution of trade waiting times is non-exponential.
Conversely, exponentially distributed trade waiting times only result from a
finite thinning of a Poisson order process. Consequently, the distribution of
order waiting times should be a more general distribution, e.g., a Weibull
distribution, than an exponential distribution underlying a simple memoryless
Poisson process.
The hypothesis of non exponentially distributed order waiting times is
empirically accessible and can be directly checked if full book information
were available. Moreover, one can try to solve the inverse problem: given a
trade waiting time distribution, which is the originating order waiting time
distribution? Our future activity will follow these lines.
Acknowledgments
This work has been supported by the University of Genova and by the Italian
Ministry of Education, University and Research (MIUR) under grant FIRB
2001 and under grant FISR ”Ultra-high frequency dynamics of financial mar-
kets”.
References
1. B. Biais, P. Hillion, and C. Spatt, “An empirical analysis of the limit order book
and the order flow in the Paris bourse,” J Financ, vol. 50, no. 5, pp. 1655–1689,
December 1995.
2. J.-P. Bouchaud, M. Mezard, and M. Potters, “Statistical properties of stock
order books: empirical results and models,” Quantitative Finance, vol. 2, no. 4,
pp. 251–256, August 2002.
3. S. Cincotti, C. Dose, S. M. Focardi, M. Marchesi, and M. Raberto, “Analysis
and simulation of a double auction artificial financial market,” 6-10 July 2003,
Istanbul, Turkey, EURO/INFORM 2003.
4. S. Cincotti, S. M. Focardi, M. Marchesi, and M. Raberto, “Who wins? study
of long-run trader survival in an artificial stock market,” Physica A, vol. 324,
no. 1-2, pp. 227–233, June 2003.
5. D. R. Cox and V. Isham, Point processes. Chapman and Hall, 1980.
6. M. M. Dacorogna, R. Gencay, U. Muller, R. B. Olsen, and O. V. Pictet, An
Introduction to High Frequency Finance. Academic Press, 2001.
7. G. Iori, M. G. Daniels, J. D. Farmer, L. Gillemot, S. Krishnamurthy, and
E. Smith, “An analysis of price impact function in order-driven markets,” Phys-
ica A, vol. 324, no. 1-2, pp. 146–151, June 2003.
8. A. W. Lo, A. C. MacKinlay, and J. Zhang, “Econometric models of limit-order
executions,” J Financ Econ, vol. 65, no. 1, pp. 31–71, July 2002.
9. F. Mainardi, M. Raberto, R. Gorenflo, and E. Scalas, “Fractional calculus and
continuous-time finance ii: the waiting-time distribution,” Physica A, vol. 287,
no. 3-4, pp. 468–481, December 2000.
10. M. Marchesi, S. Cincotti, S. M. Focardi, and M. Raberto, The Genoa artificial
stock market: microstructure and simulation, ser. Lecture notes in economics
and mathematical systems. Springer, February 2003, vol. 521, pp. 277–289.
11. S. Maslov and M. Mills, “Price fluctuations from the order book perspective:
empirical facts and a simple model,” Physica A, vol. 299, no. 1-2, pp. 234–246,
October 2001.
12. M. Raberto, “Modelling and implementation of an artificial financial market
using object oriented technology: the Genoa artificial stock market,” Ph.D.
dissertation, University of Genoa, Italy, 2003.
13. M. Raberto, S. Cincotti, C. Dose, S. M. Focardi, and M. Marchesi, “Price
formation in an artificial market: limit order book versus matching of supply
and demand,” in Proceedings WEHIA 2003. Springer, 2004.
14. M. Raberto, S. Cincotti, S. M. Focardi, and M. Marchesi, “Agent-based simu-
lation of a financial market,” Physica A, vol. 219, pp. 319–327, October 2001.
15. M. Raberto, S. Cincotti, S. Focardi, and M. Marchesi, “Traders’ long-run wealth
in an artificial financial market,” Computational Economics, vol. 22, no. 2-3,
pp. 255–272, October-December 2003.
16. M. Raberto, E. Scalas, and F. Mainardi, “Waiting-times and returns in high-
frequency financial data: an empirical study,” Physica A, vol. 314, no. 1-4, pp.
749–755, November 2002.
17. E. Scalas, R. Gorenflo, F. Mainardi, M. Mantelli, and M. Raberto, “Anomalous
waiting times in high-frequency financial data,” 2003, submitted to Quantita-
tive Finance, http://xxx.lanl.gov/abs/cond-mat/0310305.
Part VI
Minority Games and Collectie Intelligence

Theoretical Analysis of Local Information
Transmission in Competitive Populations
Sehyo Charley Choe1 , Sean Gourley1 , Neil F. Johnson1 , and Pak Ming Hui2
1
Clarendon Laboratory, Physics Department, Oxford University, Parks Road,
Oxford OX1 3PU, UK
s.choe1@physics.ox.ac.uk
2
Physics Department, The Chinese University of Hong Kong, Shatin, New
Territories, Hong Kong
Summary. We study Complex Adaptive Systems in which information is shared

locally within a heterogeneous population of adaptive nodes (‘agents’) competing
for a limited global resource. The emerging collective behaviour is found to depend
strongly on the level of available resource, the connectivity between agents and the
accuracy of information transmission.
1 Introduction
The study of Complex Adaptive Systems is expanding the boundaries of
Physics into non-traditional areas, such as economics, sociology and biol-
ogy. Statistical physics and the validity of its associated assumptions such
as ergodicity, self-averaging and history-independence, is based on particles
obeying deterministic equations-of-motion, time-averaged over the behaviour
of the entire collective. On the other hand, in Complex Adaptive Systems
the adaptive objects (or ‘agents’) modify individual behavioural rules (‘strat-
egy’) according to past experience. It is the local-local interactions between
such agents which gives rise to emergent behaviour. The traditional assump-
tions of statistical physics may therefore not hold, creating a rich area of
study for theorists. These adaptive systems are often rich in empirical data,
offering Physics a new testing ground for theories of complex systems and
non-equilibrium statistical mechanics in general [1]. In particular, Arthur’s
El Farol bar problem (EFBP) [2], and its subsequent binary representation,
namely the Minority Game (MG) [3, 4, 5, 6, 7, 8], are recognized [9] to em-
body the interacting, many-body nature of many real-world complex systems,
in which there is often competition between agents for some global resource
[2, 3, 4]. Researchers have only just started considering the effect of networks
in the MG and EFBP [6, 7], despite widespread interest among physicists in
socio-economic, biological, and informational networks [10].
252 Sehyo Charley Choe, Sean Gourley, Neil F. Johnson, and Pak Ming Hui
Most network research assumes that nodes are inert, focusing rather on
the structural and statistical properties of the connections. In contrast, multi-
agent research has focused on adaptive decision-taking agents in the absence
of network connections. Very little has been said about the combination of
the two, and hence the functionality of such a network, despite its practical
importance [11]. Furthermore, it has so far been assumed that any informa-
tion shared between agents is always perfectly accurate. However, real-world
complex systems do not operate at such levels of perfection. Informational
networks might be used by agents to spy and mislead rather than to benefit
others. This raises the interesting question of the effect of networks within
competitive population where the information transmitted is corrupted.
Here we study a class of Complex Adaptive Systems comprising a com-
petitive multi-agent population, known as Binary Agent Resource (B–A–R)
systems. We give a brief overview of the analytical Crowd-Anticrowd theory
[4] which incorporates the many-agent (many-body) correlations or ‘crowd-
ing’ arising in the system’s strategy space as a result of the dynamics in the
history space. We then extend the Crowd-Anticrowd theory to encompass the
micro and/or macroscopic effects of network functionality and the effects of
information corruption on the system behaviour.
2 Binary–Agent–Resource (B–A–R) Systems

2.1 The Model
Our B-A-R model is a networked binary version of Arthur’s El Farol problem

[2, 3, 4]. The model consists of N agents who repeatedly compete for some
limited resource L. At each timestep t, each agent decides whether to access
resource L (i.e. action +1) or not to access this resource (i.e. action −1). The
two global outcomes at each timestep, ‘resource over-used’ and ‘resource not
over-used’, are denoted as ‘0’ and ‘1’. If the number of agents n+1 [t] choosing
action +1 exceeds L (i.e. resource over-used and hence global outcome ‘0’)
then the n−1 [t] abstaining agents win. By contrast if n+1 [t] ≤ L (i.e. resource
not over-used and hence global outcome ‘1’) then these n+1 [t] agents win. Each
agent decides its actions in light of global information which takes the form of
the history of the m most recent global outcomes. Adaptation is introduced
m
by randomly assigning S strategies to each agent. Each of the 22 possible
strategies is a bit-string of length 2m defining an action (+1 or −1) for each
of the 2m possible global outcome histories. For example, m = 2 has 22 = 4
possible global outcome histories: 00, 01, 10 and 11. Consequently there are
m=2
22 = 16 possible strategies. Strategies which predicted the winning (losing)
action at a given timestep, are assigned (deducted) one point.
Theoretical Analysis of Local Information Transmission 253
2.2 Crowd-Anticrowd Formalism
The Crowd-Anticrowd theory [4] incorporates the many-agent correlations

arising in the system’s strategy space as a result of the dynamics in the history
space. To a very good approximation, we can replace the Full Strategy Space
by a Reduced Strategy Space (RSS) [3] which provides a minimal basis set of
strategies for the system [4]. The appropriate choice for the RSS depends on
the relative frequency of visits to the 2m histories. With all histories visited
equally often, the RSS comprises a total of 2P = 2 · 2m strategies [3]. As
emphasized in previous MG works, small fluctuations in D(t) are signatures
of an efficient self-organization within the population [3, 4, 5, 6]. The excess
demand D(t) at timestep t is given by:

K=2P
D(t) = n+1 (t) − n−1 (t) = aK nK (t) (1)
K=1
where n+1 (t) (n−1 (t)) is the number of agents taking action +1 (−1); aK =
±1 is the action predicted by strategy K in response to history µ(t) and
nK (t) is the number of agents using strategy K at time t. K labels the K’th
highest scoring strategy while K = 2P + 1 − K labels the anti-correlated
strategy. In the non-networked MG at low m, D(t) exhibits large, crowd-driven
fluctuations while µ(t) follows a quasi-deterministic Eulerian Trail3 in which
all histories {µ(t)} are visited equally. Hence, the time-averages n−1 (t) =
n+1 (t) yielding D(t) = 0 which is the optimal value for D(t). We will
assume that the combined effect of averaging over t for a given Ψ (where Ψ is
a given realization of the initial strategy allocation matrix [4]), and averaging
over Ψ , will have the same effect as averaging over all histories. This is true
for the non-networked MG, and produces a mean D(t) of zero. Hence, it can
be shown that the fluctuation (i.e. standard deviation) of the excess demand,
σD , is given by:
& P '
2
P
mean 2
2
σD = (nK (t) − nK (t)) ≈ nK − nmeanK
. (2)
K=1 t,Ψ K=1
In an unconnected system, nmean

K = nK , where nK is the intrinsic crowd-size
due to crowding in the global strategy space and can be shown to be given by
[4]:
S S
(K − 1) K
nK = N 1 − m+1 − 1 − m+1 . (3)
2 2
3
The Eulerian Trail corresponds to a particular cycle around the global information
states (de Bruijn graph), of duration 2P = 2·2m timesteps, in which every possible
transition between global information states is visited once. See P. Jefferies, M.L.
Hart and N.F. Johnson, Phys. Rev. E 65, 016105 (2002).
We have used the orthogonality properties of the vectors with elements aK

where K = 1, 2, . . . , 2P [4]. In addition, since nK (t) will generally fluctuate
around some mean value navg avg
K , we have also written nK (t) = nK + K (t)
and assumed that the fluctuation terms {K (t)} are uncorrelated stochastic
processes.
3 Networked B–A–R Systems

Here we examine the functionality of a network within the B–A–R model.
In particular, whether network connections are beneficial or detrimental de-
pends dramatically on the level of inter-nodal competition. An accurate an-
alytic theory can only be derived if one incorporates both the strong corre-
lations within the internal strategy space of the nodes and the global non-
ergodicity. Standard assumptions related to ergodicity, self-averaging and
history-independence, fail in this system.
3.1 Perfect Information Transmission
The basic setup is identical to the B–A–R model. However, at each time step
t each agent decides its actions in light of (i) global information which takes
the form of the history of the m most recent global outcomes, in addition
to (ii) local information obtained via network connections. At each timestep,
each agent compares the score of its own best-scoring strategy (or strategies)
with the highest-scoring strategy (or strategies) among the agents to whom
it is connected. The agent adopts the action of whichever strategy is highest-
scoring overall, using a coin-toss to break any ties. For simplicity we here
assume a random network, where the connection between any two agents
exists with a probability p. We emphasize that any of the model assumptions
can easily be generalized: the numerical results are reasonably robust.
The mean success-rate of the system, which is inversely proportional to
the wastage σD as defined in Eq. 2, can be expressed [4] in terms of the
mean number or Crowd of agents using the K’th highest-scoring strategy (i.e.
nmean
K ) and the mean number or Anticrowd of agents using strategy K which
is anticorrelated to K (i.e. nmean
K
where K = 2m+1 + 1 − K). Explicitly
nmean
K = nK + n→K − nK→ (4)
where nK is the number of agents who would have used strategy K in the
absence of the network because of the initial random strategy allocation and
is given by Eq. 3.
n→K is a sum over all agents who use strategy K as a direct result of a
network connection:

n→K = nJ (1 − p)ΣG<K nG 1 − (1 − p)nK . (5)
J>K
Hence n→K represents the increase in Crowd-size due to the local connec-
tivity. By contrast, nK→ is a sum over all agents who would have used K in
the absence of a network, but who will now use a better strategy as a result
of their network connections:

nK→ = nK 1 − (1 − p) Σ G<K nG
. (6)
Hence nK→ represents the decrease in Crowd-size due to the local con-
nectivity. The resulting analytic expressions for the mean success-rate are in
excellent agreement with numerical results (see Fig. 2).
Fig. 1. Modest resources. Solid line (left axis) shows numerically-obtained mean
success-rate per agent (node) as a function of the probability p that agent i is
connected to agent j. Right axis: typical histograms of agents’ success-rate in a
typical run of 105 timesteps. N = 101, S = 2, m = 1 and L = 50. Inset: spread in
success-rate as a function of p. Error bars obtained from 20 separate runs.
Figure 1 shows the mean success-rate per agent, averaged over many nu-
merical realizations of the underlying connections and strategy distributions
(left axis: solid curve) and the success-rate distribution within the population
for a typical run (right axis: histograms) as a function of the inter-connectivity
p, in a modest-resource population illustrated by L = (N − 1)/2. We focus on
the small m or ‘crowded’ regime [3, 4] since there are relatively few strategies
compared to the number of agents and hence many agents will simultane-
ously be competing to win with the same strategy. The mean success-rate
decreases rapidly as p increases (left axis: solid curve) before saturating at
p = psat ∼ 0.1. As m increases, psat increases. The success-rate distribution
for a typical run (right axis: histogram) at p = 0 exhibits a definite ‘class
structure’ in terms of success. Dramatic changes then arise as p increases.
The spread in the success-rate – shown explicitly in the inset and by vertical
arrows in Fig. 1 – indicates a rapid increase in the population’s heterogeneity
in the range 0 ≤ p ≤ 0.02. The success-rate distribution becomes almost con-
tinuous, washing out the p = 0 class structure. Above p ∼ 0.02, there is a rapid
drop in the proportion of highly-successful agents. Further increasing p leads
to a decrease of the spread in success-rates. High levels of inter-connectivity
therefore provide increased fairness (i.e. small spread in success-rate) but de-
creased efficiency (i.e. small mean success-rate). Similar results are obtained
for networks with different topological properties, e.g. the Barabási-Albert
network driven by preferential attachment [10].
Fig. 2. Solid lines show analytic Crowd-Anticrowd theory for mean success-rate per
agent (node) with modest resources, for S = 1, 2, 3, 4. Data-points are numerical
results. m = 1, N = 101, L = 50. Inset: S = 2, various N values. Right: Eulerian
Trails (i.e. high p attractor) for m = 1, 2, 3.
Figure 2 compares our analytical results using the Crowd-Anticrowd the-

ory [4] to the numerical results for the mean success-rate per agent. The
mean success-rate per agent can be obtained from the standard deviation of
the number of agents making a given choice (e.g. +1), by averaging this quan-
tity over the attractors of the system (i.e. averaging over the Eulerian Trail
or subsets of it).
Fig. 3. Higher resources. Thin solid lines: numerically-obtained mean success-rate

per agent (node) as a function of p, at resource levels L = 80, 90, 95, 100. Results
averaged over 200 runs of 105 timesteps, with N = 101, S = 2, m = 1. Scattered
circles show mean success-rate per agent for separate runs at L = 100 with m = 1,
as an illustration. Thick solid lines: analytic curves (Eqs. 7 – 9) describing the three
dynamical regimes. The analytic curves obscure some of the numerical results. Inset:
histograms of typical success-rate distribution, for L = 100, m = 1.
Figure 3 shows the regime changes which arise as a function of the inter-
connectivity p in higher resource populations. If the global resource level L
exceeds a critical amount Lcrit = 3N/4, the mean success-rate per agent
can exhibit a maximum at small but finite inter-connectivity, for the follow-
ing reason. When L > Lcrit and p = 0, the population inhabits a ‘frozen’
state in the sense that the global outcome is persistently 1, i.e. . . . 111111.
With S = 2 as in Fig. 3, each agent has probability 1/4 of being assigned
two strategies which both define action −1 following a string of m ‘1’s. Thus
approximately 1/4 of the population always lose and 3/4 of the population
always win, leading to Lcrit = 3N/4 ≈ 75. The global resource is therefore
under-used at p = 0 by ∆L ≈ (L − 3N/4) at each timestep. Therefore in-
creasing p away from p = 0 will benefit some of the less successful agents by
connecting them up to successful agents, thereby giving them access to ∆L,
until a p value is reached where the connectivity is sufficient to break the

outcome series of 1’s. These run-averaged results can be better understood by
considering run-specific results for L = 100 as an illustrative case (scattered
circles). Specific runs appear to aggregate into groups having similar temporal
dynamics and success-rate distributions – these groups correspond to differ-
ent dynamical states of the system. The increase in the mean success-rate at
low p, corresponds to the system following an Efficient State E which has an
outcome series of 1’s as discussed above. The success-rate distribution (inset)
for State E is characterized by groups of persistent winners and losers. As
p increases further, the outcome series for high L tends to move through a
set of Intermediate States which combine a low spread in success-rate with a
high mean (e.g. Intermediate States A and B). In Intermediate State A, the
least-successful agents (which in State E had zero success) now have a success-
rate which exceeds the average success-rate in the high connectivity limit. In
the high p limit, the outcome series tends toward the period-4 Eulerian Trail
given by . . . 00110011 . . . . The corresponding number of agents taking action
+1 follows the pattern . . . N, N, N/2, N/2, N, N, N/2, N/2 . . . . The resulting
Fair State F corresponds to all agents having access to the best performing
strategies, either by being assigned that strategy or by being connected to
another agent with that strategy. The resulting system is fair but inefficient,
having a small spread but also a small mean success-rate ≈ 0.25.
Figure 4 shows how these transitions between regimes in high-resource
populations (e.g. L = 95), vary with history bit-string length m. The Inter-
mediate States, characterized by a reasonably high mean success-rate and a
reasonably small spread (see Fig. 3 inset), have an increasingly dominant ef-
fect on the system’s behaviour as m increases. For a particular value of m,
increasing p moves the system toward cycles of increasing period and hence
the fractions of 1’s and 0’s in the output series become more equal. The re-
sulting State F, which represents an increasingly noisy version of the Eulerian
Trail as m increases, is fair but inefficient (see Fig. 3 inset).
We have derived analytic expressions, within the Crowd-Anticrowd frame-
work including non-ergodicity, which describe well the three regimes (thick
solid lines in Figs. 3 and 4). The upper analytic branch at low p, describing
the efficient State E, is given by:

3 1 3N
+ 1 − (1 − p) 4 . (7)
4 4
The middle analytic branch at intermediate p, describing the Intermediate

States, is given by:
3 9 7N 1 15N 1 3N
− (1 − p) 16 − (1 − p) 16 + (1 − p) 4 . (8)
4 32 32 8
The lower analytic branch at high p, describing the fair State F, is given by:
1 9 7N 1 15N 1 3N
+ (1 − p) 16 + (1 − p) 16 + (1 − p) 4 . (9)
4 128 128 16
Fig. 4. High resources. Thin solid lines: numerically-obtained mean success-rate per
agent (node) as a function of p, at various m. Results averaged over 200 runs of 105
timesteps, with N = 101, S = 2, L = 95. Scattered circles show mean success-rate
per agent for separate runs at m = 3. Thick solid lines: analytic curves (Eqs. 7 –
9) describing the three dynamical regimes. The analytic curves obscure some of the
numerical results.
The outcome series of 1’s at low p which yield State E, can persist up
to pcrit ∼ 1 − (1 − 4∆L/N )4/3N . For L = 80, 90, 95 and 100, this yields
pcrit ∼ 0.002, 0.011, 0.019 and 0.042 respectively, which are also all in excel-
lent agreement with the numerical results.
3.2 Imperfect Information Transmission
We now investigate the effects of microscopic connection-driven data error on

the system’s macroscopic dynamics. We examine how the interplay between p,
the connectivity between agents, and q, the probability of an error occurring,
affects the fluctuations in the excess demand D(t). We show that erroneous
data transmission generates an abrupt global transition in the B–A–R model
[8]. This phase-like transition is driven by a ‘temporal symmetry breaking’ in
the global outcome series. The Crowd-Anticrowd theory provides a quantita-
tive yet physically intuitive explanation of this phase transition.
As in the B–A–R model with perfect information transmission, agents use
the connections they have, if any, to gather information from other agents. As
before, each agent takes the action predicted by the highest-scoring strategy
among his own and those of the agents to which he is connected. However,
there now exists a parameter q which represents the probability that an error
arises in the information he gathers from his cluster. Alternatively, q can be

viewed as the weight an agent places on the information gathered from his
cluster. For example, if the action of the best strategy in his cluster is +1, the
agent records this as a −1 with probability q (and vice versa for a best action
−1). The information transmission has been corrupted with probability q. Any
agent with a higher-scoring strategy than those of his neighbours at a given
timestep, is unaffected by this error – the only source of stochasticity which
might affect him is the standard coin-toss used to break any ties between his
own strategies [4]4 .
As in the unconnected case, we continue the focus on small m here, since
we are interested in the effect of q on these crowd-driven fluctuations. In Eq.
2,
nmean
K = nK + nq→K − nqK→ − nqK (10)
and similarly for nmean
K
, where:
(1) nK is the mean number of agents whose own best strategy is actually the
K’th highest scoring strategy in the game [4] and is given by Eq. 3.
(2) nq→K is the mean number of agents who only possess strategies worse (i.e.
lower scoring) than K, but who will use strategy K due to connections they
have to one or more agents who each possess strategy K but no better:

nq→K = 1 − q · n→K + q · n→K (11)
where n→K is given by Eq. 5, with n→K being obtained from Eq. (5) by
setting K → K = 2P + 1 − K.
(3) nqK→ is the mean number of agents who possess strategy K, but who will
nevertheless use a strategy better than K due to connections:

nqK→ = 1 − q · nK→ + q · nK→ (12)
where nK→ is given by Eq. 6 and similarly for nK→ .

(4) nqK accounts for the situation in which an agent is connected to other
agents with the same highest scoring strategy K as him. q therefore gives the
probability that this agent will take the opposite action to strategy K:

n K
nK
q
nK = q · nK 1 − 1 − p − q · nK 1 − 1 − p . (13)
Figure 5 compares the numerical and analytical results for σD , which is

the standard deviation in excess demand. The agreement is remarkable given
the complexity of σD as a function of p and q. As the ‘noise’ level q increases,
the system undergoes a change in regime at a critical connectivity p defined
by the critical boundary Ccrit (q, p). Moving across Ccrit (q, p), the symmetry
4
We note that, in contrast to the agents’ ‘on-site’ stochastic strategy selection
arising in the Thermal Minority Game (TMG) [5, 8], the stochasticity associated
with q in our game depends on the agents’ connectivity.
Fig. 5. Fluctuation in excess demand, σD , as a function of the error-probability

q and the network connectivity p. (a) Numerical results averaged over 300 runs,
each with 105 iterations. (b) Analytic Crowd-Anticrowd theory. At high q, the two
branches in (a) correspond to different dynamical attractor states, while the single
branch in (b) represents an effective average (see text). The dotted line in (b) at
p = 0.25, illustrates the modified analytical results for the upper branch if one
assumes some knowledge of this branch’s global output series. Parameters: m = 1,
S = 2, and N = 101.
in the global outcome string is spontaneously broken in a manner reminiscent

of a phase transition. Specifically, the global outcome series changes from the
low-q phase where it resembles the period-4 Eulerian Trail . . . 00110011 . . . ,
to a high-q phase where it comprises two distinct branches (see Fig. 5(a)).
Figure 6 shows individual runs near the critical noise threshold. The higher
branch corresponds to a period-2 global outcome series . . . 1010 . . . which is
antiferromagnetic if we denote 0 (1) as a spatial spin up (down) as opposed to
a temporal outcome. The lower branch corresponds to the period-1 series of
‘frozen’ outcomes . . . 0000 . . . or . . . 1111 . . . , i.e. ferromagnetic. In this high-
q phase, the system will ‘choose’ one of these two global outcome branches
spontaneously, as a result of the type and number of links each agent has. This
symmetry-breaking of the global outcome series along the channel of minimum
fluctuation in Fig. 5, Ccrit (q, p), originates in the internal coupling between
the history dynamics, the strategy space and the individual agent networks.
Because of the initial strategy allocation and connections, many agents will
have an in-built bias towards one of the two possible actions and hence act in
a deterministic or ‘decided’ way at a given timestep. However, there exist a
few ‘undecided’ agents who need to toss an unbiased coin to decide between
the equally balanced signals they gather from their local network. It is the
40
30
σD
20
10
0.08
0.06
0.45 0.04
0.50 0.02 p
0.55
q 0.60 0.00
Fig. 6. Numerical results for individual runs, showing the fluctuation in excess
demand, σD , around q = 0.5. Parameters as in Fig. 5.
fluctuations of these few ‘undecided’ agents who then push the system onto a
particular branch.
We now discuss two technical details. First, there are many ties in strategy
scores at very small m, and hence many tie-break coin-tosses. This means that
the fluctuation terms {K (t)} can no longer be ignored. The m = 1 surface in
Fig. 5(b) was therefore produced by averaging over the 2 · 2m = 4 timesteps in
the Eulerian Trail. In other words, the double-average in Eq. (2) was evaluated
over the 2 · 2m = 4 timesteps in the Eulerian Trail. When a tie-break between
the strategies K = 1 and K = 2 arises at one of the four timesteps, one
1
replaces nmean mean mean mean
K=1 and nK=2 by 2 (nK=1 + nK=2 ) at that timestep. Likewise, for
tie-breaks between any other K and K . In this way, the average over the
Eulerian Trail is easily evaluated analytically. As m increases, there are more
timesteps over which one must average (i.e. 2P = 2 · 2m timesteps). However,
since ties also become less frequent as m increases, one can simply ignore them
without significant loss of accuracy (see Ref. [4] in which good agreement is
obtained for the non-networked MG for a wide range of m values without
considering ties). Second, the theory has assumed the non-networked MG
result that the dynamics follow the Eulerian Trail. Only one branch therefore
emerges in Fig. 5(b) at high q, appearing like some effective average over the
global output series for all branches in Fig. 5(a). If instead one uses knowledge
of the actual global output series for these separate branches, then results even
closer to Fig. 5(a) can be obtained. This is illustrated at one particular p by
the dotted line in fig. 5(b).
Figure 7 provides a contour plot of σD around the minimum. The black
contour, centred around the critical curve Ccrit (q, p), effectively separates the
two different regimes of behaviour. The low σD values around Ccrit (q, p) can be
easily understood using the physical picture provided by the Crowd-Anticrowd
1.0 1.0
(a) (b)
0.8 0.8 0
14
28
0.6 0.6
42
q
56
0.4 0.4 70
84
98
0.2 0.2 112
0.0 0.0
0.00 0.02 0.04 0.06 0.08 0.10 0.00 0.02 0.04 0.06 0.08 0.10
p p
Fig. 7. Contour map version of fig. 5. Contours correspond to a constant value

of σD as a function of the error-probability q and the network connectivity p. (a)
Numerical results. (b) Analytic Crowd-Anticrowd theory. For clarity, only the upper
branch of the numerical results is shown for the high-q phase.
theory: the stochasticity induced by q (i.e. noise) breaks up the size of the
Crowds using a given strategy K, while simultaneously increasing the size of
the Anticrowds using the opposite strategy K. It is remarkable that a linear in-
crease in the ‘noise’ q gives rise to such a non-linear variation in σD . Using the
analytic expressions in this paper, an equation for Ccrit (q, p) can be obtained
– however we do not include it because it is cumbersome. As noted above, the
theory neglects a full treatment of the dynamical fluctuations around nmean K .
Hence, the theory overestimates the Crowd-Anticrowd cancellation arising in
Eq. (2) and thus slightly underestimates σD in the neighbourhood of Ccrit (q, p)
(compare Figs. 7(a) and 7(b)). As p increases, Ccrit (q, p) becomes less depen-
dent on the connectivity p since more and more agents join the same network
cluster. For p 0.05 the system passes the percolation threshold and hence
is dominated by a giant, common cluster.
4 Conclusion
We have presented a general model of heterogeneous multi-agent systems com-
peting for some limited resource, in which a communication network is used
between agents. The numerical results can be explained quantitatively using
the Crowd-Anticrowd theory. The network gives rise to a rich dynamical be-
haviour as a function of the inter-agent connectivity. We also found that less-
than-perfect local information transmission between agents can lead to the
emergence of a global transition reminiscent of a phase-transition. Our results

raise the interesting possibility whereby imperfect information transmission
could be induced at the local level in order to achieve a desired change in the
macroscopic fluctuations within biological, informational or socio-economic
systems.
References
1. See, for example, A. Soulier and T. Halpin-Healy, Phys. Rev. Lett. 90, 258103
(2003).
2. W.B. Arthur, Science 284, 107 (1999); W. B. Arthur, Am. Econ. Assoc. Papers
Proc. 84, 406 (1994); N. F. Johnson, S. Jarvis, R. Jonson, P. Cheung, Y. R.
Kwong and P. M. Hui, Physica A 258, 230 (1998).
3. D. Challet, Y. C. Zhang, Physica A 246, 407 (1997); D. Challet, M. Marsili,
and R. Zecchina, Phys. Rev. Lett. 84, 1824 (2000); R. Savit, R. Manuca, and R.
Riolo, Phys. Rev. Lett. 82, 2203 (1999); N. F. Johnson, P. M. Hui, R. Jonson,
and T. S. Lo, Phys. Rev. Lett. 82, 3360 (1999); S. Hod and E. Nakar, Phys.
Rev. Lett. 88, 238702 (2002).
4. N.F. Johnson and P.M. Hui, preprint cond-mat/0306516; N.F. Johnson, S.C.
Choe, S. Gourley, T. Jarrett, P.M. Hui, Advances in Solid State Physics, 44,
427-438 (Springer Verlag, Germany, 2004); M. Hart, P. Jefferies, N. F. Johnson,
P. M. Hui, Physica A 298, 537 (2001); N.F.Johnson, P. Jefferies, P.M. Hui,
Financial Market Complexity (Oxford University Press, Oxford, 2003).
5. A. Cavagna, J. P. Garrahan, I. Giardina, and D. Sherrington, Phys. Rev. Lett.
83, 4429 (1999); M. Hart, P. Jefferies, N. F. Johnson, P. M. Hui, Phys. Rev. E
63, 017102 (2000).
6. M. Anghel, Z. Toroczkai, K.E. Bassler, G. Korniss, Phys. Rev. Lett. 92, 058701
(2004); I. Caridi, H. Ceva, cond-mat/0401372; M. Sysi-Aho, A. Chakraborti,
and K. Kaski, Physica A 322, 701 (2003); D. Challet, M. Marsili and G. Ottino,
preprint cond-mat/0306445.
7. S. Gourley, S.C. Choe, N.F. Johnson and P.M. Hui, Europhys. Lett. 67, 867
(2004).
8. S.C. Choe, N.F. Johnson and P.M. Hui, Phys. Rev. E (Rapid Communication),
(2004) in press.
9. J.L. Casti, Would-be Worlds (Wiley, New York, 1997).
10. D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J. Watts, Phys. Rev.
Lett. 85, 5468 (2000); R. Albert and A.L. Barabási, Phys. Rev. Lett. 85, 5234
(2000); M. E. J. Newman, C. Moore, and D. J. Watts, Phys. Rev. Lett. 84,
3201 (2000); R. Albert and A.L. Barabási, Rev. Mod. Phys. 74, 47 (2002).
11. T. Hogg, Nanotechnology 10, 300 (1999); R.V. Sole, B. Goodwin, Signs of
Life (Basic Books, New York, 2000); A.L. Barabási and Z.N. Oltvai, Nature
(Genetics) 5, 101 (2004).
Analysis of Complexity and Time Restriction
in Resources Allocation Problems
Kiyoshi Izumi1 , Tomohisa Yamashita2 , and Koichi Kurumatani3

1
Information Technology Research Institute (ITRI), AIST & CREST, JST
kiyoshi@ni.aist.go.jp
2
ITRI, AIST & CREST, JST tomohisa@carc.aist.go.jp
3
ITRI, AIST & CREST, JST k.kurumatani@aist.go.jp
Summary. In this paper, we constructed three types of agents, which are different
in efficiency and accuracy of learning. They were compared using acquired payoff
in a game-theoretic situation that is called Minority game. As a result, different
types of learning methods got the highest payoff according to the complexity of
environmental change and learning speed.
1 Introduction
1.1 Resource Allocation Problems and Minority Games
Resources allocation problems are seen in all the places in the real world. The
allocation problem of bandwidth in the Internet or information communica-
tion, competition of the market share between companies. the route determi-
nation problem of traffic, consumers’ purchase selection, and so on, are listed
as examples. One of important features common to these problems is that the
utility of the resources with less number to choose is higher. That is, the profit
of players which choose alternatives different from other players are higher.
Minority games, models of the game theory which abstract and simplify this
feature, and have been studied by many researchers recently[3, 6]. The inter-
esting result has been obtained especially about the phase transition of the
relation between the frequency of the player which chooses each resource and
the memory length of in resources selection. However, in the conventional re-
search, all players determined selection using the same information, and have
changed strategy by the same learning method. The heterogeneity among
players was not taken into consideration. In this research, the heterogeneity
of the information used for a player and the learning method was introduced
into the minority game which is the model of a resources allocation prob-
lem. And two conditions, the complexity of the dynamics of the whole game
and the time restrictions about learning were changed, and we investigated
266 Kiyoshi Izumi, Tomohisa Yamashita, and Koichi Kurumatani
which type of agent becomes advantageous. Especially we were interested in

the point whether the learning method which assumed others’ intelligence at
the time of what conditions becomes advantageous.
1.2 Complexity and Time Restriction
What conditions make one assume that the others have capability to process
information, to determine action according to rules, and to update their be-
havioral rules? In this study, it is thought that the following two conditions
are key points.
• The social systems that one belongs to have complexity about those dy-
namic behavior. The complexity originates mainly from the interaction
among members in the social systems, not from external conditions.
• There is time restriction, when one tires to make decisions and to learn a
new decision rules in social situations. He or she, therefore, must find not
only accurate but also efficient decision rules and learning rules.
The point of view we take in this paper is that the above two conditions,
complexity and time restriction, account for the advantage of the assumption
that the others have intelligence.
• Because of the complexity of social systems, members must update their
decision rules by learning, in order to adapt themselves to social situations
and to acquire higher utility. When the members are heterogeneous, that
is, they have various decision rules, they must use the others’ information
in learning.
• Because of the time restriction to learning, the members must find effective
learning methods. They adopt the learning method with an assumption
that the others have intelligence, because it is effective in social situations.
1.3 Hypothesis
The hypothesis of this research is as follows.

Assuming that the others have capability to process information, to
update an internal state, and to determine action, have advantage
when the two conditions hold; (a) the complexity of environmental
change, and (b) the time restriction to learning.
In this research, as a preliminary experiment for investigating this hypoth-
esis, the computer simulation was performed on the topic of the game theory
situation called a minority game.
Title Suppressed Due to Excessive Length 267
2 Framework of Experiments
2.1 Minority Game as a Model of Resource Allocation Problems
The minority game is a repeated game in which N (odd number) players must
choose one of two alternatives at each step. A payoff is given to minority group,
players who chose the alternatives which fewer people chose between two al-
ternatives. Arthur, who is one of the proposal persons of complexity systems
economy, proposed the idea of a bar problem that people try to drink at the
bar which fewer person chose between two bars[1]. Then, many researchers
pointed out and analysis the feature as a nonlinear phenomenon and made
various extensions[7].
Since the mechanism that a minority group win is seen also in resource
allocation problems such as traffic flows, ecological systems, and trading in
financial markets[4, 8]. In this research, standard minority game[3, 6] was
developed and it considered as the framework of a preliminary experiment of
the following framework interpreted as a model of resource selection problems.
The agent of N (odd number) participates in a game, and time progresses
discretely. Two resources exist in this game and all agents must choose which
resource they use every time step. One time step in this game consists of the
three stages; (1) determination of action, (2) calculation of payoffs, and (3)
learning.
Determination of Action
Each agent i determines its own action hi (t) = {+1, −1}. When an agent uses
one resource, say, resource A, its action at time t, hi (t), is denoted by +1.
When it uses the other, say, resource B, hi (t) = −1. The action is decided
based on information called memory, Im (t − 1), the time series data about
which resource provided higher payoff at the past m time steps.
Im (t − 1) = {I(t − 1), I(t − 2), · · · , I(t − m)} (1)
I(τ ) = {+1, −1} expresses the higher payoff resource at τ . When the resource
A provided the higher payoff at τ , I(τ ) = +1. When the resource B, I(τ ) =
−1.
Each agent i has the rule that determines its resource selection behavior
hi (t) according to each pattern of memory Im (t−1). This rule is called strategy
S i (t) (see Figure 1).
Calculation of Payoff
The resource selection behavior of all N agents are accumulated to calculate

each agent’s payoff at t. In this game, the resource that fewer agents chose can
provide higher payoff to the agents that chose this resource. And it is assumed
Action hi (t)
m
Memory I (t − 1) A or B
0 1
{−1, −1, −1, · · · , −1, −1} → +1 or − 1
{−1, −1, −1, · · · , −1, +1} → B +1 or − 1 C
B C
.. .. B .. C
. . @ . A
{+1, +1, +1, · · · , +1, +1} → +1 or − 1
↑
Strategy S i (t)
Fig. 1. Strategy
that the fewer agents select one resource, the larger payoff agents can acquire.
Payoffi (t) is calculated to each agent i as follows.

N
payoff i (t) = −hi (t) · hj (t) (2)
j=1
N
When more agents choose resource A ( j=1 hj (t) > 0), the agent that chooses
resource B (hi (t) = −1) belongs to a minority group, and the minority group’s
agents acquire
Npositive payoffs. On the other hand, when more agents choose
resource B ( j=1 hj (t) < 0), the agent that chooses resource A (hi (t) = +1)
acquires a positive payoff.
The information I(t) about which resource has the advantage at t is cal-
N
culated similarly. When more agents choose resource B ( j=1 hj (t) < 0), the
resource A has the higher payoff, then, I(t) = +1. When more agents choose
N
resource A ( j=1 hj (t) > 0), then, I(t) = −1.
# N
+1 ( i=1 hi (t) < 0)
I(t) = N
−1 ( i=1 hi (t) > 0)
Learning
The learning in the standard minority game is the selection of the strategy
Si (t) that is used in the determination of an action. Each agent has s strategies
shown in Figure 1 and s is usually from 2 to 36. These strategies are generated
randomly at the beginning of the game, and they are fixed throughout the
game. Each strategy has a specific value called a virtual value. It is a number
of times that the behavior derived from a strategy could acquire a positive
payoff. Each agent chooses one strategy with the highest value from the s
strategies, and uses it when it decides its own behavior at the next time step
t + 1.
In standard studies of the minority game, only the very simple learning was
assumed and they analyzed in many cases about the relationship between the
memory length m and the fluctuation of the number of agents that chose one
resource. However, in a previous study[2], it is suggested that it has essential
significance that all agents are homogeneous, that is, all agents share the same
information and the same learning algorithm. It is shown that the same results
was obtained in both cases that all agents used the real historical data Im (t)
and that all agents used the randomly generated data. In our study, agents
are heterogeneous, that is, the information and the learning methods differ
among agents.
2.2 Agents
We prepared 4 types of agents; one standard type and three extended types.
First, Pattern matcher (PM) was a standard agent. This agent determines
behavior based on the time series data Im (t). It is extended from the player
in the standard minority game, described in section 2.1.
In addition, we prepared three kinds of agent; Hand imitators (HI),
Strategy imitators (SI), and Perfect predictors (PP). They are differ-
ent in terms of kinds of information that they use (see Table 1). Each agent
took part in the minority game against Pattern matchers, and its payoff was
compared with each other. Hand imitators perform simple learning of imi-
Table 1. Agent types

Efficiency HI > SI > PP
Accuracy HI < SI < PP
Others’ payoff Others’ payoff Others’ payoff
Others’ action Others’ action Others’ action
Information
Memory about resources Memory about resources
Game structure
(HI: Hand imitator, SI: Strategy imitator, and PP: Perfect predictor)
tating the behavior of other agents with high payoff. Since it uses the least
information, the speed of learning is quick and the efficiency is good, but the
learning results will be inaccurate. On the other hand, Perfect predictor infers
both other agents’ strategy and the game structure (payoff matrix) using all
kinds of information. Since it uses much information, the obtained result is
exact, but the speed of learning is slow and the efficiency is low. Strategy
imitators are in the middle of these two types. It performs only a inference
of other agents’ strategy, and imitates the strategy of other agents with high
payoff. The efficiency and accuracy of its learning are in the middle of HIs
and PPs.
Pattern Matcher (PM)

Pattern matchers’ behavior rules are the same as stated in section 2.1. It is
extended about learning. Since a agent in standard minority games continues
to have its strategies given first without changing, as described in section2.1.

Thus, it can not search for all of solution spaces. Then, we extended pattern
matcher’s learning method as follows, to enable it to search for all solution
spaces.
Decision of behavior
Pattern matcher has one strategy described in figure 1. The pattern matching
of the historical data of the past m steps, Im (t − 1), to the strategy S i (t)
determines Pattern matchers’ behavior hi (t).
Learning
When Pattern matchers acquires a positive payoff at t, it does not change

its strategy. When it got a negative payoff, its strategy S i (t) is updated at a
certain probability α (learning rate). The bit of the behavior rule (the resource
A or B) to this historical pattern is reversed.
Hand Imitator (HI)
Hand imitator performs simple learning of imitating the behavior which other
agents with a high payoff.
According to a certain probability pA , Hand imitator uses the resource A.

Probability to use the resource B pB is 1 − pA .
Learning
The probability of dealing behavior of other agents with a high payoff is copied.
1. Inference of others’ probability
The estimated probability p̃jA (t) that another agent j chooses the resource
A is updated by the following equation.
p̃jA (t) = (1 − β) · p̃jA (t − 1) + β · actionj (t) (3)
actionj (t) is probability which the agent j chooses the resource A at this
term.
#
j 1 (Agent j chooses the resource A at t)
action (t) =
0 (Agent j chooses the resource B at t)
Parameter 0 ≤ β ≤ 1 expresses the rate which updates the estimated value

of the probability of behavior of other agents, and means the learning
speed of others’ models.
2. Accumulation of payoff
About all agents j including itself, the accumulation value Rj (t) of payoff
is updated by the following equation.
Rj (t) = (1 − γ) · Rj (t − 1) + γ · payoff j (t) (4)
payoff (t) is a parameter showing the payoff of agent j this term. 0 ≤ γ ≤ 1
j
expressed the update rate of the accumulation value of a payoff, and fixed
it to 0.5 in this study.
3. Copy of the behavior according to the payoff
Each agent copies the probability pA from other agents with a certain
probability α (learning rate). One agent j ∗ is chosen by the probability
proportional to accumulation value Rj (t) of each payoff from all the agents
∗
that he also contains first. And the probability pjA (t) is copied to its own
probability pA .
Strategy Imitator (SI)

Strategy imitator performs only an estimation of other agents’ strategy, and
imitates the strategy of the high agent of a payoff.
The pattern matching of the historical data of the past m steps, Im (t − 1), to
the strategy S i (t) determines behavior hi (t). It is the same as that of Pattern
matcher.
Learning
Strategy imitator estimates others’ strategies and imitates the other’s strategy
with higher payoff.
1. Estimation of others’ strategies
Strategy imitator estimates whether the other agent js to choose the re-
source A or B, using the pattern matching of the historical data Im (t − 1)
to estimated others’ strategies S̃ j (t). If the estimated behavior is different
from the actual behavior which agent j performed, by a certain probabil-
ity 0 ≤ β ≤ 1 (learning speed of others’ models), the estimated strategy
S̃ j (t) will be updated. The bit of the agent j’s behavior corresponding to
the historical pattern in S̃ j (t) will be inverted.
2. Accumulation of a payoff
About all agents j including itself, the accumulation value Rj (t) of payoff
is updated by the equation 4. It is the same as that of Hand imitator.
3. Imitation of strategy according to the payoff
A strategy is copied by a certain probability α (learning rate). One agent
j ∗ is chosen by the probability proportional to accumulation value Rj (t) of
∗
each payoff from all the agents. And the agent’s strategy S̃ j (t) is copied
to its strategy.
Perfect Predictor (PP)
Perfect predictor estimates both other agents’ strategies and the game struc-
ture (the payoff matrix) using all information.
Strategy
Perfect predictor estimates others’ behavior and decide its own behavior ac-
cording to the estimated payoff matrix.
1. Estimation of others’ behavior
Perfect predictor estimates whether the other agent js to buy or sell from
the pattern matching of the historical data Im (t − 1) to estimated others’
strategies S̃ j (t).
2. Decision of behavior
Perfect predictor has its own strategy that represents estimated game
structure (the payoff matrix). That is, the strategy shows which behavior
(selection of the resource A or B) can acquire a positive payoff correspond-
ing to the others’ behavior. According to the strategy and the other agent
j’s estimated behavior, Perfect predictor decides its own behavior.
Learning
Perfect predictor learns both others’ strategies and the game structure.
1. Learning of others’ model
The estimated strategies S j (t) about other agent js are updated. Learning
method is the same as that of Strategy imitator.
2. Learning of the game structure
When its payoff is lower than the average of other agents’ payoff, the bit of
behavior corresponding to the estimated others’ behavior is reversed by a
certain probability α (learning rate). This means renewal of the knowledge
about the payoff structure.
3.1 Setting of the Simulation
Setting of the simulation of the preliminary experiment in this study is shown

in Table 2.
25 agents participate in each trial of the minority game. 20 agents are
Pattern matchers as standard agents in all trials. The remaining 5 agents are
Hand imitators, Strategy imitators, or Perfect predictors.
The learning speed α of their strategy and the update rate of a payoff γ
were fixed. The memory length m and the learning speed of others’ models β
of HI, SI, and PP were changed as shown in Table 2.
Table 2. Setting of the simulation
Number of agents N = 25
Agents’ combination (20 PMs and 5 HIs), (20 PMs and 5 SIs),
and (20 PMs and 5 PPs)
Memory length m = {1, 2, · · · , 15}
Learning speed α = 0.8 (fixed)
Learning speed of
β = {0.1, 0.2, · · · , 0.9}
others’ model
Update rate of payoff γ = 0.5 (fixed)
Number of simulation runs 10 times every parameter combination
Improvement rate of the payoff of 5 HIs, SIs, and PPs
Comparison method
when the average of the payoff of 20 PMs is set to 100.
(PM: Pattern matcher, HI : Hand imitator, SI: Strategy imitator, and PP : Perfect
predictor)
The simulation was performed 10 times every parameter set; {Agents’

combination (PM vs. HI, SI, or PP) × Memory length m × Learning speed
of others’ modelsβ}. The agents (HI, SI, and PP) were compared by the
improvement rate of payoff when the average of payoff of the 20 standard
agents (Pattern matchers) wad set to 100.
3.2 Simulation Results
The simulation results are summarized in Figure 2 and examples of simulation

results are shown in Figure 3(a)-(c).
Overall Results
Different types of learning methods got the highest payoff according to the
memory length m and the learning speed of others’ models β.
Our preceding study revealed that the memory length linked to the com-
plexity of dynamics of the whole system[5]. When memory length is short,
the whole system showed relatively simple patters that can be described by
dynamic systems of finite dimensions. As the memory length got longer, the
dynamics patters became more complex, and finally they can not be described
by any finite dimensional dynamic systems. Therefore the memory length m
can be considered as an index of complexity of the whole system.
The learning speed of others’ models β is an index of the time restriction
to learning. All agents continuously change their own strategies at the fixed
learning speed α. When β is small, agents must track the others’ strategies
at low learning speed, so the time restriction to learning is strong. When β is
large, agents can have relatively enough time to track the others’ strategies,
so the time restriction to learning is weak.
-- HIs got the highest payoff. -- PPs got the highest payoff.
-- SIs got the highest payoff. -- Difference between the 1st payoff and 2nd < 3.0
16
16
15
15
14
14
13
13
12
12
11
Memory length m
10
10
99
8
77
66
55
4
33
2
1
0
0 0.1 0.2
0.2 0.3
0.3 0.4
0.4 0.5
0.5 0.6
0.6 0.7
0.7 0.8
0.8 0.9
0.9 11
Learning speed of others’ model
Fig. 2. Summary of results: Each symbol represents which agent type got the highest
payoff. There are 4 distinct areas ( m ≤ 3 & β < 0.3, m ≤ 3 & β > 0.3, 3 < m < 12,
and m ≥ 13 ) according to the winners’ types.
Figure 2 shows that there are 4 distinct areas according to the two condi-
tions, the complexity and time restriction.
1. Low complexity and strong time restriction area: When the memory length
is short (m ≤ 3) and the learning speed of others’ models is slow (β < 0.3),
Hand imitators’ payoff was high. This result was also shown in Figure 3a.
This is because the Perfect predictors and Strategy imitators can not
have enough time to learn the others’ strategies because of the strong
time restriction, and they showed the poor performance. Only Hand imi-
tators could catch up with the fast change of Pattern matchers’ strategies
because of their simple and quick learning method. Moreover, Hand imi-
tators could learn adequately accurate models because the whole system
is comparatively simple.
2. Low complexity and weak time restriction area: When the memory length
is short (m ≤ 3) and the learning speed of others’ models is fast (β > 0.3),
Perfect predictor’s payoff was high (Figure 3a) It is because the Perfect
predictors can have enough time to learn the others’ strategies and the
game structure. Moreover the dynamics of whole systems is relatively
simple in this area. They then could obtain exact learning results using
all information and beat the other two types, the Hand imitators and
Strategy imitators.
3. Medium complexity area: The Strategy imitators’ payoff becomes high as
memory length becomes longer ( 3 < m < 13). Figure 3b shows that the
Strategy imitators could beat the other two type throughout all learn-
ing speed β. When other agents’ strategy is complicated, simple learning
(a) Memory length m = 2

100
HI PP
Improvement rate of payoff (PMs =100)

80
60 SI
40
20
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
(b) Memory length m = 9

100
SI
80 HI PP
60
40
20
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
(c) Memory length m = 13

100
80
60
40
SI
20 PP
HI
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Fig. 3. Simulation results: (a) When memory length is short and the learning speed
is slow, the payoff of HI (Hand imitator) is high. When the learning speed of an
others model is fast, the payoff of PP (Perfect predictor) is high. (b) The payoff of
SI (Strategy imitator) becomes high as memory length becomes long. (c) Finally
the performance of all agent types have small difference.
methods such as the Hand imitators can only acquire inaccurate learning
results. And learning methods using all information such as the Perfect
predictors spend too much time to get learning results at a limited learn-
ing speed. Therefore, payoff of Hand imitator and Perfect predictor had
fallen in this area. Then, the Strategy imitators, that are in the middle of
these two learning types, got high payoff in this area.
4. High complexity area: In this area, the dynamic structure of the whole
system is very complicated and then can not be described by any finite di-
mensional dynamic systems. That is, the system showed chaotic features.
Therefore, the difference of performance became small among all the 4
agent types; Pattern matchers, Hand imitators, Strategy imitators, and
Perfect predictors. Figure 3c revealed that the improvement rate against
the Pattern matchers became small and it becomes hard to distinguish
among all agent types by performance.
4 Conclusions
Our simulation results showed that there are 4 areas according to the two con-
ditions, the complexity and time restriction. The various resource allocation
problems can be classified into these 4 ares (Figure 4).
High
Chaos
Finantcial markets
Imitation of strategies
Complexity Traffic behavior
Pedestrian behavior
Imitation of behavior Rationality

Evacuation process Conventional economic model
Low One-shot games
Strong Time restriction Weak
Fig. 4. Classification of resource allocation problems
Evacuation process is classified into the low complexity and strong time
restriction area. When people try to determine evacuation routes, they may
not use old memories but mainly use knowledge about the present situation.
Therefore, the memory length may be short. Of course, the time restriction
is strong in emergent situations. If one want to model evacuation process, he
or she had better build a model with agents that do not consider the other
agents’ strategies or decision rules like the Hand imitators.
The situations that are described in conventional economic models, where

the time restriction to learning is weak and environmental changes are not so
complicated, rational agents have advantages. They try to find exact solution
using all available information. This situation is so ideal that it is difficult to
find examples in the real world, but the purchase of expensive goods may be
an example.
The medium complexity area includes a wide range of the resource allo-
cation problems such as the route selection of drivers or pedestrians. In this
area, it is reasonable that people determine their actions assuming that the
other also have decision rules or strategies. Then, the learning method based
on the others’ intelligence such as the Strategy imitators have the advantage.
Finally the high complexity area also contains various examples in the
resource allocation problems such as financial markets. In this area, there
is no universally correct answer of learning. Therefore, it is thought that the
procedure of learning itself becomes a key factor not only the result of learning.
5 Discussions
In conventional economic models, when the participant of an economic sys-

tem was not using all information, only some negative causality had been
considered. For example, information cannot be acquired (in-accessibility).
Information has many noises and it cannot be trusted (low reliability). Many
costs are required to acquire information (high cost).
These models have an implicit assumption, ”a forecast would become ex-
act so much if many information is used.” Although its attention had been
directed to some extent by the cost which acquires information, probably,
about the cost which learns using the acquired information, it had not al-
most been taken into consideration at all. It is because the condition that
a participant’s learning speed assumes infinitely the ideal condition of being
fast, in the economic model with this classic, therefore the speed of change
of the environment which is a learning target can be relatively regarded to a
zero is considered. This is shown in figure 2 as an ultimate situation which
is in the right at an infinite distant place. The algorithm which performs an
exact learning in this condition using all information, without considering the
efficiency of a learning is the most advantageous.
In classic economic models, the ”true model” showing the behavior of the
target economic phenomenon exists solemnly on a cloud as a given thing. And
the participants who have an infinite learning speed using all information are
learning toward a true model.
On the other hand, in the model by the agent simulation which was shown
by this study, the behavior of an economic phenomenon is not being fixed
and it has generated internally from the interaction between participants.
Each participant thinks that he will forecast well and is performing various
learnings. The learning behavior carries out an interaction and the behavior
pattern of the whole economic system changes. While one mikoshi is shoul-
dered all together and a motion of everybody collides, it is moving on the
whole. The target to pursue also escapes or approaches according to a motion
of a mikoshi.
As for a learning in the environment where the target of a learning changes
behavior in response to reaction from the learning behavior of oneself, like a
society and an economic system, learning using all information is not neces-
sarily good. At the point, a classic economic model differs from the condition
of having assumed implicitly. Then, what does actual human being do? Prob-
ably, paying attention to what information, heuristics is used about how it
learns. Heuristics is the learning method discovered experientially. Although
it cannot necessarily obtain an exact learning result, it can obtain efficiently
the result which can be satisfied moderately. Strategy imitator is also a kind
of heuristics of imitating the strategy of others who succeeded in fact.
It is also interesting to analyze the heuristics of a learning of actual human
being found by the cognitive psychology and some learning methods in artifi-
cial intelligence in the viewpoint of the efficiency and accuracy of a learning,
and to evaluate as a model of the learning in a society and an economic sys-
tem. And finally it will become one approach to the elucidation of the function
of an intelligence in social / economical situation.
References
1. W. B. Arthur. Inductive reasoning and bounded rationality (the el farol prob-
lem). American Economic Review, 84:406, 1994.
2. A. Cavagna. Irrelevance of memory in the minority game. PHYSICAL RE-
VIEW E, 59(4):R3783–R3786, 1999.
3. D. Challet and Y.-C. Zhang. Emergence of cooperation and organization in an
evolutionary game. Physica A, 246:407–418, 1997.
4. A. Galstyan, S. Kolar, and K .Lerman. Resource allocation games with changing
resource capacities. In Proceedings of the second international joint conference
on Autonomous agents and multiagent systems, pages 145–152. ACM Press,
2003.
5. K. Izumi. Complexity of agents and complexity of markets. In T. Terano,
T. Nishida, A. Namatame, S. Tsumoto, Y. Osawa, and T.Washio, editors, New
Frontiers in Artificial Intelligence, pages 110–120. Springer, 2001.
6. M. Marsili. Market mechanism and expectations in minority and majority
games. Physica A, 299:93–103, 2001.
7. Minority Game’s web page. http://www.unifr.ch/econophysics/.
8. Y.-C. Zhang. Modeling market mechanism with evolutionary games. Europhys.
News, 29:51–54, 1998.
How Does Collective Intelligence Emerge
in the Standard Minority Game?
Satoshi Kurihara1 , Kensuke Fukuda2 , Toshio Hirotsu3 , Osamu Akashi2 ,

Shinya Sato2 , Toshiharu Sugawara4
1
I.S.I.R. Osaka University, 8-1, Mihogaoka, Ibaraki, Osaka, 567-0047 JAPAN
kurihara@ist.osaka-u.ac.jp
http://www.ai.sanken.osaka-u.ac.jp
2
NTT Network Innovation Labs., Tokyo, JAPAN
3
Toyohashi University of Technology, Aichi, JAPAN
4
NTT Communication Science Labs., Nara, JAPAN
Summary. In this paper we analyze a simple adaptive model of competition called

the Minority Game, which is used in analyzing competitive phenomena such as the
operation of the market economy. The Minority Game is played by many simple
autonomous agents, which develop collective self-organization as a result of simple
behavioral rules. Many algorithms that produce the desired behavior in the game
have been proposed. In all work to date, however, the focus has been on the macro-
scopic behavior of the agents as a whole. We focus on the behavior of individual
agents, paying particular attention to the original form of the Minority Game. We
suggest that the core elements responsible for the development of self-organization
are (i) rules that place a good constraint on the behaviors of individual agents and
(ii) the existence of rules that lead to effective indirect coordination. We also show
that when efficient organization is formed, a power-law can be seen among behavior
of individual agents.
1 Introduction
The individual animals in a school of fish or flock of flying birds seem to
be organized in well-ordered overall behaviors by some form of intelligent
coordination. The ‘boid’ was developed as a simple computer model that
replicates such behavior [11]. Each boid runs the boids algorithm, and the
boid collectively exhibit realistic flocking or schooling behavior. Such work
has clearly shown that well-ordered overall behavior can be formed by merely
local and very simple interaction among elements.
An environment consisting of many autonomous entities, every one of
which can get a large profit even though the individuals are behaving in a
selfish way governed only by simple rules that don‘t require knowledge of the
overall system, is very attractive in term of modeling of the market economy
280 Satoshi Kurihara et al.
and other socio-economic structures. Such an environment will be very useful

as a general means for coordinating agents in multi-agent systems.
The Minority Game is a simulation program for analyzing models of adap-
tive competition. The game is like the market in that it is played by multiple
autonomous agents. Each agent acts with the purpose of obtaining its own
profit by using purely local information. The Minority Game has also been a
popular object of study in ‘econophysics’ and has been looked at from various
other angles since it was first proposed more than four years ago [1]. Many
algorithms that produce the desired behavior in this game, i.e. large profits
for nearly all agents, have been proposed [2-7, 9]. In nearly all of these stud-
ies, however, researchers were concentrating on the behavior of the agents as a
whole; to date, we have not seen any papers on studies where the focus was on
the behavior of the individual agents. Therefore, we investigated the behav-
ior of the individual agents with a view to discovering the kind of individual
behavior that leads to efficient overall behavior. We pay particular attention
to the Minority Game in its initial form [1], because in this case the agents
develop a well-ordered overall behavior through purely local and very simple
interaction, even though each is behaving in a selfish way. We show that any
set of rules which makes all of the agents behave in a well-ordered manner
and get large profits produces fractal characteristics in the behaviors of the
individual agents.
2 The Standard Minority Game
Fig. 1. Rules and strategy table
Firstly, we review the rules of the standard Minority Game (see Fig. 1 (a)).
We have N agents, each of which is an autonomous entity that independently
How Does Collective Intelligence Emerge in the Standard Minority Game? 281
chooses between two alternatives (group 0 or group 1) according to its own

behavioral rules. In each round of the game, all of the agents choose one
alternative or the other, and the agents that then finish in the minority group
are considered to be winners. Each winner is awarded one point, and the total
numbers of points awarded to all agents is the profit in this round of the
game. Therefore, the smaller the difference between the numbers of agents in
the majority and minority groups, the better the result.
Each agent makes its selection based on one of multiple strategy tables
that it holds. The entities in the table contain all combinations of m past
winning-group choices along with next decisions that corresponds to each
of the combinations (see Fig. 1 (b)). At the beginning of the game, each
agent randomly prepares s strategy tables, and randomly set the next decision
entries. In the first round of the game, each agent randomly selects one of the
s strategy tables it holds. If the agent wins, that is, ends up in the minority
group, one point is assigned as profit to the selected strategy table. If the
agent loses, one point is deducted. In the second and subsequent rounds of
the game, the strategy table that has the highest number of points is always
selected. This cycle is repeated a predetermined number of times, and the final
result of the game is the total number of points acquired by winning agents
across all rounds.
Fig. 2. (a) Standard deviations of the number of winning agents. (b) Mean numbers
of winning agents
2.1 Coordinated Behaviors
We begin by verifying the collective behavior of the agents in the same way
as was done in earlier studies [1-3]. Firstly, the standard deviation of the
number of winning agents is shown in Fig. 2(a). The game was played for the
number of rounds described below with the agents possessing various numbers
of strategy tables, s = {2, 5, 10, 16, 32, 64} and the strategy tables having
various history depths, m = 3 to 16. One trial for each parameter pair, s, m,
is 10,000 rounds of the game; ten trials were conducted for each pair. Fig.
2(b) shows the mean numbers of winning agents. The horizontal lines in Fig.
2(a) and (b) represent the standard deviation and mean value when all of the
agents made random choices. These graphs show that, for the lower values
of s, the standard deviation became lowest and the mean number of winning
agents became highest when m was from three to five. This means that some
kind of coordinated behavior among the agents was driving the winning-group
ratio closer to 100:101 in these cases. The most interesting characteristic is
that, although we would expect behavior based on longer histories to be more
efficient, m values larger than 10 produce results that are the same as those
of random behavior.
Even though the standard deviations for the lower s values are low in the
range m=3 to 5, fixed sets of agents may always be ending up winners or
losers. So, we investigated how many times each agent would be able to win
in several situations. Figure 3 shows the rankings of the 201 agents by average
score. In the case where each agent randomly selected ”group 0” or ”group 1”,
all of the agents were able to get approximately 4750 points. In contrast to
this, the mean score was high when the standard deviation was small (m=3 to
5) and, although some differences between agents can be seen in the scores, all
or almost all of the agents were able to achieve stable high scores. Interestingly,
the number of losing agents increased with m, and the trend is towards the
levels seen in the random selection case. This was particularly so when m was
greater than 10. So, it can be seen that a phase transition occurred in m=6
to 9.
3 Key Elements in the Formation of Self-Organization

In the standard Minority Game, the following two rules are thought to be
important elements: (i) the winning-group history is universal, so each agent
must see the same entry in its selected strategy tables, and (ii) each strategy
table’s score is changed by every win and loss like the random-walk.
3.1 A Winning-Group History is Unnecessary
In the original algorithm [1], if the current winning-group history is ”010,” all
agents see the entries for ”010” in the strategy tables they hold, and each agent
selects its currently highest-scoring strategy table. However, we hypothesize
that the memory of winning-group history has no effect.
At this point, we raise the following questions.
(1) If each agent randomly selects the entry from its currently highest scoring
strategy table, does self-organization occur?
Fig. 3. Rankings of 201 agents for each value of m
(2) If only one agent, agent-A, selects an entry based on the winning-group
history and the other agents use the same entry, does self-organization oc-
cur? In other words, agents do not obey the rule of winning-group history,
but simply see the same entry as that selected by agent-A.
(3) If we intentionally generate a random winning-group history, does self-
organization occur?
If self-organization develops in any of the above situations, our hypothesis

that the memory of the winning-group history has no effect may be correct.
Fig. 4(a) shows the result: the same self-organization as with the normal
algorithm developed in cases (2) and (3). That is, even when we made a
random winning-group history as in case (3), the agents were able to develop
good overall organization. This result shows that the memory of the winning-
group history may not be important in the development of self-organization.
However, when each agent selects a random entry from its strategy tables, the
behavior is the same as for random group selection.
We can thus infer an important point concerning the strategy tables. That
is, in the normal algorithm, the rule that all agents depend on the winning-
group history simply places a constraint on the agents. In other words, we
may be able to use any rule that places a ‘good’ constraint on the agents,
as in cases (2) and (3). Applying a good constraint to the agents essentially
means decreasing their freedom. In earlier published work [8], we discussed
how other frameworks lead to self-organization, and made the same suggestion
as in this subsection.
Fig. 4. (a) Is memory necessary? (b) The role of indirect coordination: despite
having this, versions I, II, and III led to great disparities between winning and
losing agents.
3.2 The Importance of Indirect Coordination
Indirect interaction between agents is also considered to be important element

in the development of organization in the standard Minority Game. That is,
each agent decides its behavior based on the results of each round, and each
agent’s decision indirectly affects the behaviors of the other agents. In the
normal algorithm, one point is added to the selected strategy table when the
agent wins, and one point is subtracted when it loses. As an example, consider
an agent that has two strategy tables, table-A and table-B . If table-A has
four points and table-B has one point, table-B will not selected until table-A
has lost at least four times in a row. Will self-organization still develop if
we change the rule for selecting the strategy tables? Fig. 4(b) shows the re-
sults of our investigation of this question. We implemented the following rules.
(Version I) Agents select strategy tables sequentially. The interval of exchange

is randomly set.
(Version II) One point is added to the score for a selected strategy table that
wins, but two points are subtracted if it loses.
(Version III ) If the agent loses a single game, the strategy table is exchanged
for the next-scoring table, even when the currently strategy table still has the
most points.
Unfortunately, self-organization did not develop with any of these versions.

The method of strategy-table selection must have a close relationship with
the initial combinations of strategy tables held by each agent. Therefore, we
investigated how the strategy tables are used by individual agents with the
normal algorithm. In Fig. 5, (a) and (b) show the transitions of the scores for
each strategy table held by the agents in 25th and 200th places among the
201 agents, in trials where all agents had m=3 and m=14, respectively. While
Fig. 5. Which strategy table was used? Scores for (a) m=3, 25th place agent
and (b) m=14, 200th place agent in (upper) 10000, (middle) 1000, and (lower) 100
games.
the agent in Fig. 5(a) used both strategy tables, there was no fixed period for
the continuous use of one strategy table. A fractal characteristic is visible in
these results; that is, the usage of strategy tables shows self-similarity. Fig. 6
shows, on log-log scales, per-agent histograms of the periods over which either
of the strategy tables was continuously used. Results for m=3, 7, and 14 are
given for (a) n = 101, and (b) n = 301. Power-law relations can be seen in
both cases of m=3 (graphs were nearly straight lines) and this reflects the
presence of a fractal characteristic.
Returning to Fig. 5, no fractal characteristic is visible in the results for
the agent in 200th place (Fig. 5 (b)). Figure 2 showed that the performance
Fig. 6. No. of steps L over which either strategy table was in continuous use for
(a) n=101 and (b) n=301 with (upper) For m=3 (all 101 agents and 301 agents),
(middle) For m=7 (all 101 agents and 301 agents) (lower) For m=14.
of agents with m=14 is only as good as random selection. We thus do not see
straight lines in the histograms for m=14 in Fig. 6; these results are similar to
the result for random selection. This reflects a fixed period for the continuous
use of each strategy table in these cases (as seen in the lower part of Fig.
5 (b)). The histograms for agents with m=7 show an interesting mix of the
two types of results; some are similar to the graphs for m=3, while others are
similar to those for m=14. It might be the case that the straight lines are
for winning agents, while the curves similar to those for m=14 are for losing
agents.
4 Discussion
At this point, why power-law relations can be seen in the cases of m=3?
We think that the assigning and deducting rule of strategy table may be
important. The rule of when the agent wins one point is assigned and when
it loses one point is deducted is similar to “random-walk.” Indeed, as the
Fig. 4 (b)-VersionII shows, fractal characteristic cannot be seen by the rule
of when the agent wins one point is assigned and when it loses two points
are deducted, and as the Fig. 6 (a)-upper shows, each slope of agent and the
slope of random-walk5 is almost similar. In random-walk, probability that
assigning-selection and deducting-selection are chosen is equal, so, when m
value is small (like m = 3) it can be thought that probability of winning and
losing may be equal. On the other hand, when m value is big (like m = 14),
it can be thought that probability of losing may be higher than winning, so,
fractal characteristic like random-walk cannot be seen.
As for power-law relations, Fig. 7 shows the slopes6 of the histograms in
Fig. 6 for all agents, from that in first place to that in last place ((a) n=101
and (b) n=301). As you can see, slopes are clearly greater for winning than
for losing agents. This result is very interesting, because it indicates that
individual agents have a way to find out whether they are behaving well or
poorly without knowledge of the results for other agents.
Further detailed investigation will be necessary to clarify the essential
mechanisms in the selection of strategy tables, and this is a topic for future
work. However, we appear to have found a way to check whether or not agents
are performing in a well-ordered manner.
5 Conclusion
In this paper we analyzed a simple adaptive-competition model called the
Minority Game, which is used in the analysis of competitive phenomena in
5
We ran random-walk 10000 steps, and calculated the CDF of histograms of the
number of steps where a positive value or a negative value was consecutive.
6
Slope is calculated by the least-squares method
Fig. 7. Slopes of the histograms in Fig. 6.
markets. In particular, we focused on the standard Minority Game. We veri-

fied the validity of the following hypotheses: (i) the rules simply have to place
a ‘good’ constraint on the behaviors of individual agents and (ii) the rules of
the standard game lead to effective indirect coordination. A. Cavagana [10]
has already demonstrated that the memory of the winning-group is irrelevant,
and stated that the important point is the sharing of some data among the
agents. However, we go further in suggesting that the important point is sim-
ply to place a ‘good’ constraint on the behaviors of individual agents, and
that this constraint might take the form of the sharing of data which, in the
normal algorithm, is the winning-group history. On point (ii), further detailed
investigation will be necessary to clarify the essential mechanism in the selec-
tion of strategy tables that leads to effective indirect coordination. We will
continue our analysis, with the aim of establishing a general algorithm that is
applicable to other competition problems.
References
1. D. Challet and Y.-C. Zhang, Emergence of Cooperation and Organization in
an Evolutionary Game, Physica A 246, 40 7 (1997).
2. Y.-C. Zhang, Modeling Market Mechanism with Evolutionary Games, Euro-
phys. News 29, 51 (1998).
3. D. Challet, A. Chessa, M. Marsili, Y.C. Zhang, From Minority Games to Real
Markets, Quantitative Finance (2001 ).
4. R. Savit et al., Adaptive Competition, Market Efficiency, Phase Transitions
and Spin-Glasses (1997); on the major Minority Game web site:
http: //www.unifr.ch/econophysics/minority/
5. D. Challet, M. Marsili and R. Zecchina, Phase Transition in a Toy market
(1999); on the major Minority Game web site:
http://www.unifr.ch/econophysics/minority/
6. Y. Li et al., Evolution in Minority Games I. Games with a Fixed Strategy

Space; on the major Minority Game web site:
7. Y. Li et al., Evolution in Minority Games II. Games with Variable Strategy
Spaces; on the major Minority Game web site:
8. S. Kurihara, R. Onai, and T. Sugawara, Self-organization based on Nonlinear
Non-equilibrium dynamics of Autonomous Agents, Artificial Life and Robotics,
Vol. 2, No. 3, 1998.
9. R. D’hulst and G. J. Rodgers, Three-sided Complex Adaptive Systems (1999);
on the major Minority Game web site:
10. A. Cavagana, Irrelevance of Memory in the Minority Game, Phys. Rev. E, 59,
R3783 (1999).
11. C. W. Reynolds, Flocks, Herds, and Schools: A Distributed Behavior Model,
Computer Graphics 21:4, pp. 25-34, (SIGGRAPH 87), 1987.
This page intentionally blank
Part VII
Game-theoretic Approach
This page intentionally blank
What Information Theory Says About
Bounded Rational Best Response
David H. Wolpert1
NASA Ames Research Center, Moffett Field, CA, 94035, USA

dhw@email.arc.nasa.gov
Summary. Probability Collectives (PC) provides the information-theoretic exten-

sion of conventional full-rationality game theory to bounded rational games. Here
an explicit solution to the equations giving the bounded rationality equilibrium of
a game is presented. Then PC is used to investigate games in which the players use
bounded rational best-response strategies. Next it is shown that in the continuum-
time limit, bounded rational best response games result in a variant of the replicator
dynamics of evolutionary game theory. It is then shown that for team (shared-payoff)
games, this variant of replicator dynamics is identical to Newton-Raphson iterative
optimization of the shared utility function.
1 Introduction
Recent work has used information theory [9, 12] to provide a principled ex-
tension of noncooperate conventional game theory to accommodate bounded
rationality [25, 27]. Intuitively, this extension starts with the observation that
in the real world ascertaining a game’s equilibrium is an exercise in statistical
inference: one is given (or assumes) partial information about the behavior of
the players, and from that infers (!) what the joint mixed strategy is likely to
be. There are many ways to do such statistical inference. The one investigated
in [27] is based on information theory’s version of Occam’s razor: Predict the
joint mixed strategy that has as little extra information as possible beyond
the provided partial knowledge while being consistent with that knowledge.
This version of Occam’s razor is known as the Maximum entropy (Maxent)
principle [9, +
12]. It tells us that the mixed strategy of a game’s equilibrium,
q(x ∈ X) = i qi (xi ), is the solution to a coupled set of Lagrangian functions
that are specified by the game structure and the provided partial knowledge.
Sec. 2 reviews how information theory can be used to derive bounded
rational noncooperative game theory. Some simple examples of the bounded
rational equilibrium solutions of games are then presented. Sec. 3 analyzes
scenarios in which the players use bounded rational versions of best response
294 David H. Wolpert
strategies. Particular attention is paid to team games, in which the players

share the same utility function. The analysis for this case provide insight into
how to optimize the sequence of moves by the players, as far as their shared
utility is concerned. This can be viewed as a formal way to optimize the
organization chart of a corporation.
Best response strategies, even bounded rational ones, are poor models of
real-world computational players that use Reinforcement Learning (RL) [20].
Sec. 4 considers iterated games in which players use a (bounded rational)
variant of best response, a variant that is more realistic for computational
players, and arguably for human players as well. In this variant the conditional
expected utilities used by player i to update her strategy, expected payoff given
move xi , is a decaying average of recent conditional expected utilities. This
decay biases the player to dampen large and sudden changes in her strategy.
This variant is then explored for the case of team games. The continuum
limit of the dynamics of such games is shown to be variant of the replicator
dynamics. It is shown such continuum-limit bounded rational best response
is identical to Newton-Raphson iterative optimization of the shared utility
function of such games.
The formalism presented in this paper is a special case of the field of
Probability Collectives (PC), a case in which the joint distribution over the
variables of interest is a product distribution. This special case is known as
Product Distribution (PD) theory [25, 27, 29, 28, 26, 7]. PC has many appli-
cations beyond those considered in this paper, e.g., distributed optimization
and control [16, 15, 2, 29]. Finally, see [16] for relations to other work in game
theory, optimization, statistical physics, and reinforcement learning.
2 Bounded Rational Noncooperative Game Theory

In this section we motivate PD theory as the information-theoretic formula-
tion of bounded rational game theory. We use the integral sign ( ) with the
associated measure implicit, i.e., it indicates sums if appropriate, Lebesgue
integrals over Rn if appropriate, etc. In addition, the subscript (i) is used to
indicate all index values other than i. Finally, we use P to indicate the set of
all probability distributions over a vector space, and Q to indicate the sub-
set of P consisting of all product distributions (i.e., the associated Cartesian
product of unit simplices).
In noncooperative game theory one has a set of N players. Each player i
has its own set of allowed pure strategies. A mixed strategy is a distri-
bution qi (xi ) over player i’s possible pure strategies. Each player i also has a
private utility function gi that maps the pure strategies adopted by all N of
the players into the real numbers. So givenmixed +strategies of all the players,
the expected utility of player i is E(gi ) = dx j qj (xj )gi (x).
In a Nash equilibrium every player adopts the mixed strategy that maxi-
mizes its expected utility, given the mixed strategies of the other players. More
What Information Theory Says 295

+
formally, ∀i, qi = argmaxqi dx qi j=i qj (xj ) gi (x). Perhaps the major ob-
jection that has been raised to the Nash equilibrium concept is its assumption
of full rationality [10, 6, 18, 4]. This is the assumption that every player i
can both calculate what the strategies qj=i will be and then calculate its as-
sociated optimal distribution. In other words, it is the assumption
+ that every
player will calculate the entire joint distribution q(x) = j qj (xj ).
In the real world, this assumption of full rationality almost never holds,
whether the players are humans, animals, or computational agents [5, 17,
10, 3, 8, 1, 22, 14]. This is due to the cost of computation of that optimal
distribution, if nothing else. This real-world bounded rationality is a major
impediment to applying conventional game theory in the real world.
2.1 Review of the Minimum Information Principle
Shannon was the first person to realize that based on any of several separate
sets of very simple desiderata, there is a unique real-valued quantification of
the amount of syntactic information in a distribution P (y). He showed that
this amount of information is the negative of the Shannon entropy of that
distribution, S(P ) = − dy P (y)ln[ Pµ(y) (y)
]. So for example, the distribution
with minimal information is the one that doesn’t distinguish at all between
the various y, i.e., the uniform distribution. Conversely, the most informative
distribution is the one that specifies a single
+ possible y.Note that for a product
distribution, entropy is additive, i.e., S( i qi (yi )) = i S(qi ).
Say we given some incomplete prior knowledge about a distribution P (y).
How should one estimate P (y) based on that prior knowledge? Shannon’s re-
sult tells us how to do that in the most conservative way: have your estimate
of P (y) contain the minimal amount of extra information beyond that already
contained in the prior knowledge about P (y). Intuitively, this can be viewed
as a version of Occam’s razor: introduce as little extra information beyond
that you are provided in your inferring of P . This minimum information ap-
proach is called the maxent principle. It has proven extremely powerful in
domains ranging from signal processing to supervised learning [12]. In partic-
ular, it is has been successfully used in many statistics applications, includ-
ing econometrics[13]. It has even provided what many consider the cleanest
derivation of the foundations of statistical physics [11].
2.2 Maxent Lagrangians
Much of the work on equilibrium concepts in game theory adopts the per-
spective of an external observer of a game. We are told something concerning
the game, e.g., its cost functions, information sets, etc., and from that wish to
predict what joint strategy will be followed by real-world players of the game.
Say that in addition to such information, we are told the expected utilities
of the players. What is our best estimate of the distribution q that generated
those expected cost values? By the maxent principle, it is the distribution

with maximal entropy, subject to those expectation values.
To formalize this, for simplicity assume a finite number of players and of
possible strategies for each player. To agree with the convention in fields other
than game theory (e.g., optimization, statistical physics, etc.), from now on
we implicitly flip the sign of each gi so that the associated player i wants to
minimize that function rather than maximize it. Intuitively, this flipped gi (x)
is the “cost” to player i when the joint-strategy is x.
With this convention, given prior knowledge that the expected utilities of
the players are given by the set of values {i }, the maxent estimate of the
associated q is given by the minimizer of the Lagrangian

L (q) ≡ βi [Eq (gi ) − i ] − S(q) (1)
i
(
= βi [ dx qj (xj )gi (x) − i ] − S(q) (2)
i j
where the subscript on the expectation value indicates that it evaluated un-
der distribution q. The {βi } are “inverse temperatures” implicitly set by the
constraints on the expected utilities.
Solving, we get the coupled equations
−Eq(i) (G|xi )
qi (xi ) ∝ e (3)
where theoverall proportionality constant for each i is set by normalization,

and G ≡ i βi gi 1 . In Eq. 3 the probability of player i choosing pure strategy
xi depends on the effect of that choice on the utilities of the other players.
This reflects the fact that our prior knowledge concerns all the players equally.
If we wish to focus only on the behavior of player i, it is appropriate to
modify our prior knowledge. First consider the case of maximal prior knowl-
edge, in which we know the actual joint-strategy of the players, and therefore
all of their expected costs. For this case, trivially, the maxent principle says
we should “estimate” q as that joint-strategy (it being the q with maximal
entropy that is consistent with our prior knowledge). The same conclusion
holds if our prior knowledge also includes the expected cost of player i.
Modify this maximal set of prior knowledge by removing from it specifica-
tion of player i’s strategy. So our prior knowledge is the mixed strategies of all
players other than i, together with player i’s expected cost. We can incorpo-
rate prior knowledge of the other players’ mixed strategies directly, without
introducing Lagrange parameters. The resultant maxent Lagrangian is
Li (qi ) ≡ βi [i − Eq (gi )] − Si (qi )
solved by a set of coupled Boltzmann distributions:

1
The subscript q(i)Qon the expectation value indicates that it is evaluated according
the distribution j=i qj .
−βi Eq(i) (gi |xi )
qi (xi ) ∝ e . (4)
Following Nash, we can use Brouwer’s fixed point theorem to establish that
for any non-negative values {β}, there must exist at least one product distri-
bution given by the product of these Boltzmann distributions (one term in
the product for each i).
The first term in Li is minimized by a perfectly rational player. The second
term is minimized by a perfectly irrational player, i.e., by a perfectly uniform
mixed strategy qi . So βi in the maxent Lagrangian explicitly specifies the bal-
ance between the rational and irrational behavior of the player. In particular,
for β → ∞, by minimizing the Lagrangians we recover the Nash equilibria
of the game. More formally, in that limit the set of q that simultaneously
minimize the Lagrangians is the set of mixed strategy equilibria of the game,
together with the set of delta functions about the pure Nash equilibria of the
game. The same is true for Eq. 3.
Note also that independent of information-theoretic considerations, the
Boltzmann distribution is a reasonable (highly abstracted) model of human
behavior. Typically humans do some “exploration” as well as “exploitation”,
trying each move with probability that rises as the expected cost of the move
falls. This is captured in the Boltzmann distribution mixed strategy.
One can formalize the concept of the rationality of a player in a way that
applies to any distribution, not just a Boltzmann distribution. One does this
with a rationality operator which maps a q and a gi to a non-negative
real value measuring the rationality of player i in adopting strategy qi given
private cost function gi and strategies q(i) of the other players. For the solution
in Eq. 4 and private cost gi , the value of that operator is just βi [27].
Eq. 3 is just a special case of Eq. 4, where all player’s share the same
private cost function, G. (Such games are known as team games.) This
relationship reflects the fact that for this case, the difference between the
maxent Lagrangian and the one in Eq. 2 is independent of qi . Due to this
relationship, our guarantee of the existence of a solution to the set of maxent
Lagrangians implies the existence of a solution of the form Eq. 3. Typically
players will be closer to minimizing their expected cost than maximizing it.
For prior knowledge consistent with such a case, the βi are all non-negative.
For each player i define fi (x, qi (xi )) ≡ βi gi (x) + ln[qi (xi )]. Then we can
write the maxent Lagrangian for player i as

Li (q) = dx q(x)fi (x, qi (xi )). (5)
Now in a bounded rational game every player sets its strategy to minimize its
Lagrangian, given the strategies of the other players. In light of Eq. 5, this
means that we can interpret each player in a bounded rational game as being
perfectly rational for a cost function that incorporates its computational cost.
To do so we simply need to expand the domain of “cost functions” to include
(logarithms of) probability values as well as joint moves.
2.3 Examples of Bounded rational Equilibria
It can be difficult to start with a set of cost functions and associated ratio-
nalities βi and then solve for the associated bounded rational equilibrium q.
Solving for q when prior knowledge consists of expected costs i rather than
rationalities can be even more tedious. (In that situation the βi are not spec-
ified upfront but instead are Lagrange parameters that we must solve for.)
However there is an alternative approach to constructing examples of games
and their bounded rational equilibria that is quite simple. In this alternative
one starts with a particular mixed strategy q and then solves for a game for
which q is a bounded rational equilibrium, rather than the other way around.
To illustrate this, consider a 2-player single-stage game. Let each player
have 3 possible moves, indicated by the numerals 0, 1, and 2. Say the (bounded
rational) mixed strategy equilibrium is
q1 (0) = 1/2, q1 (1) = 1/4, q1 (2) = 1/4;

q2 (0) = 2/3, q2 (1) = 1/4, q2 (2) = 1/12 . (6)
Now we know that at the equilibrium, q1 (x1 ) ∝ e−β1 E(g1 |x1 ) , where β1
is player 1’s rationality, and g1 is her cost function (the negative of her cost
function). This means for example that
q1 (0)
e−(β1 [E(g1 |x1 =0)−E(g1 |x1 =1)]) = = 2, i.e.,
q1 (1)
β1 [E(g1 | x1 = 0) − E(g1 | x1 = 1)] = −ln(2). (7)
A similar equation governs the remaining independent difference in expecta-

tion values for player 1. The analogous two equations for player 2 also hold.
Now define the vectors gi;j (.) ≡ gi (xi = j, .). So for example g1;0 =
(g1 (x1 = 0, x2 = 0), g1 (x1 = 0, x2 = 1), g1 (x1 = 0, x2 = 2)). Then we can
express our equations compactly as four dot product equalities:
β1 (g1;0 − g1;1 ) · q2 = −ln(2) ; β1 (g1;0 − g1;2 ) · q2 = −ln(2) ;

β2 (g2;0 − g2;1 ) · q1 = −ln(8/3) ; β2 (g2;0 − g2;2 ) · q1 = −ln(8) . (8)
We can absorb each βi into its associated gi ; all that matters is their product.
We can now plug in for the vectors q1 and q2 from Eq. 6 and simply write
down a set of solutions for the four three-dimensional vectors gi,j . For these
{gi } the bounded ratinal equilibrium is given by the q of Eq. 6. If desired, we
can evaluate the associated expected values of the cost functions for the two
players; our q is the bounded ratinal equilibrium for those expected costs.
Note that the variables in the first pair of equalities in Eq. 8 are inde-
pendent of those in the second pair. In other words, whereas the Boltzmann
equations giving q for a specified set of gi are a set of coupled equations, the
equations giving the gi for a specified q are not coupled. Note also that our
equations for the gi;j are (extremely) underconstrained. This illustrates how
compressive the mapping from the gi to the associated equilibrium q is. Bear
in mind though that that mapping is also multi-valued in general; in general
a single set of cost functions can have more than one equilibrium, just like it
can have more than one Nash equilibrium.
The generalization of this example to arbitrary numbers of players with
arbitrary move spaces is immediate. As before, indicate the moves of every
player by an associated set of integer numerals starting at 0. Recall that the
subscript (i) on a vector indicate all components but the i’th one. Also absorb
the rationalities βi into the associated gi .
Now specify q and the vectors gi (xi = 0, .) (one vector for each i) to be
anything whatsoever. Then for all players i, the only associated constraint on
the i’th cost function concerns certain projections of the vectors gi (xi > 0, .)
(one projection for each value xi > 0). Concretely, ∀i, xi > 0,
( (
qi (0)
dx(i) gi (xi , x(i) ) qj (xj ) = −ln( ) + dx(i) gi (0, x(i) ) qj (xj ),
qi (xi )
j=i j=i
qi (0)
i.e., ∀i, xi > 0, gi (xi , .) · q(i) = −ln( ) + gi (0, .) · q(i) . (9)
qi (xi )
All the terms on the right-hand side are specified, as well as the q(i) term on
the left-hand side. Any gi (xi , .) that obeys the associated equation has the
specified q as a bounded rational equilibrium.
See [27] for discussion of alternative interpretations of this information-
theoretic formulation of bounded rationality. That reference also discusses
kinds of prior knowledge that do not result in the Maxent Lagrangian, in
particular knowledge based on finite data sets (Bayesian inference). A scalar-
valued quantification of the rationality of a player is also presented there.
3 Bounded Rational Versions of Best Response

One crude way to try to find the q given by Eq. 4 would be an iterative pro-
cess akin to the best-response scheme of game theory [10]. Given any current
distribution q, in this scheme all agents i simultaneously replace their current
distributions. In this replacement each agent i replaces qi with the distribu-
tion given in Eq. 4 based on the current q(i) . This scheme is the basis of the
use of Brouwer’s fixed point theorem to prove that a solution to Eq. 4 exists.
Accordingly, it is called parallel Brouwer updating. (This scheme goes by
many names in the literature, from Boltzmann learning in the RL community
to block relaxation in the optimization community.)
Sometimes conditional expected costs for each agent can be calculated ex-
plicitly at each iteration. More generally, they must be estimated. This can
be done via Monte Carlo sampling, iterated across a block of time. During
that block the agents all repeatedly and jointly IID sample their (unchanging)
probability distributions to generate joint moves, and the associated cost val-
ues are recorded. These are then use to estimate all the conditional expected
costs, which then determine the parallel Brouwer update 2 .
This is exactly what is done in RL-based schemes in which each agent
maintains a data-based estimate of its cost for each of its possible moves,
and then chooses its actual move stochastically, by sampling a Boltzmann
distribution of those estimates. (See [25] for ways to get accurate MC estimates
more efficiently than in this simple scheme, e.g., by exploiting the bias-variance
tradeoff of statistics.)
One alternative to parallel Brouwer updating is serial Brouwer updating,
where we only update one qi at a time. This is analogous to a Stackelberg
game, in that one agent makes its move and then the other(s) respond [4,
6]. In a team game, any serial Brouwer updating must reduce the common
Lagrangian, in contrast to the case with parallel Brouwer updating.
There are many versions of serial updating. In cyclic serial Brouwer up-
dating, one cycles through the i in order. In random serial Brouwer updating,
one cycles through them in a random fashion.
In greedy serial Brouwer updating, instead of cycling through all i, at each
iteration we choose what single player to update based on the associated drop
in the common Lagrangian. Those drops can be evaluated without calculating
the associated Boltzmann distributions. To see how, use Ni to indicate the
normalization constant
of Eq. 4. Then define the Lagrangian gap at q for
player i as ln[Ni ] + dxi qi (xi )Eq(i) (gi | xi ) + dxi qi (xi )ln[qi (xi )]. This is how
much L is reduced if only qi undergoes the Brouwer update 3 .
Another obvious variant of these schemes is mixed serial/parallel Brouwer
updating, in which one subset of the players moves in synchrony, followed by
another subset, and so on. Such updating in a team game can be viewed as
a simple model of the organization chart of the players. For example, this is
the case when the players are a corporation, with G being a common cost
function based on the corporation’s performance.
2
Parallel Brouwer updating has minimal memory requirements on the agents. Say
agent i has just made a particular move, getting cost r, and that the most recent
previous time it made that time was T iterations ago. Then the new estimated
T
cost for that move, E , is related to the previous one, E, by E = r+k Ea
1+kT a
, where k
is a constant less than 1, and a is initially set to 1, while itself also being updated
according to a+ = kT . So agent i only needs to keep a running tally of E, a,
and T for each of its possible moves to use data-aging, rather than a tally of all
historical time-cost pairs
3
Proof outline: Write the entropy after the update as a sum of non-i entropies
(which are unchanged by the update) plus i’s new entropy. Then expand i’s new
entropy. This gives the value of the new Lagrangian as -ln[Ni ]. Then do the
subtraction.
Say we observe the functioning of such an organization over time, and view
those observations as Monte Carlo sampling of its behavior. Then we can use
those samples to statistically estimate how best to do serial/parallel Brouwer
updating, for the purpose of minimizing the shared cost function G. This can
be viewed as a way to optimize the organization chart coupling the players.
4 Parallel Brouwer with Data-aging is Nearest Newton

This section considers a variant of best-response that is more realistic (more
accurately modeling RL-based computational players that are actually used
in machine learning, and arguably more accurately modeling human players
as well). In this variant the expected cost used by each player to update her
strategy is a decaying average of recent expected utilities; this decay reflects
a conservative preference for dampening large changes in strategy.
Such a bias is used (implicitly or otherwise) in most multi-player RL algo-
rithms. For example, in the COIN framework each agent i collects a data set
of pairs of what value its private cost function has at timestep t together with
the move it made then. It then estimates its cost for move xi as a weighted
average of all the cost values in its data set for that move. The weights are
exponentially decaying functions of how long ago the associated observation
was made. This data-aging is crucial to reflect the non-stationarity of agent
i’s environment, i.e., that the other agents are changing their strategies with
time. Arguably, humans use similar modifications to best response. Indeed, in
idealized learning rules like ficticious play, such dampening is crucial.
4.1 The Dynamics of Brouwer Updating
Consider a multi-stage game where at the end of iteration t, each player i

updates her distribution qi (., t) to
e−Φi (xi ,t)

qi (xi , t) = . (10)
dxi e−Φi (xi ,t)
This is a generalization of parallel Brouwer updating, where the function being

exponentiated can be Q values (as in Q-learning[24]), single-instant reward
values, distorted versions of these (e.g., to incorporate data-aging), etc.
As an example, for single-instant rewards (i.e., conventional parallel Brouwer),
Φi (xi , t) is player i’s estimate of (βi times) her conditional expected cost for
taking move xi at time t − 1. If that estimate were exact, this would mean

Φi (xi , t) = βE(gi | xi ) = β dx(i) q(i) (x(i) , t − 1)gi (xi , x(i) ). (11)
As another example, for Q-learning, one player is Nature and her distribution
is always a delta function. In this case Φi (xi , t) is the Q-value for player i
taking action xi , when the state of Nature is as specified by the associated

delta function in q(., t − 1).
Note that there’s no Monte Carlo sampling being done here, as there is
in most real-world RL; this is a somewhat abstracted version of such RL.
Alternatively, the analysis here becomes exact when Φi is evaluated closed
form, or (as when Φi is an empirical expectation value) there’s enough samples
in a Monte Carlo block so that empirical averages effectively give us exact
values of expected quantities.
At this point we have to say something about how Φi evolves with time.
Consider the case where Φi is an estimate of some function φi , formed by
exponential aging of the previous φ values. In our case (since everything is
evaluated closed form) assuming there have been an infinite number of pre-
ceding timesteps, this is the same as geometric data-aging:
Φi (xi , t) = αφi (xi , q(t − 1)) + (1 − α)Φi (xi , t − 1) (12)
for some appropriate function φi 4 . For example, in parallel Brouwer updating,

φi (xi , t) = βE(gi | xi , q(i) (t)), while Φi (xi , t) is a geometric average of the
previous values of φ(xi ).
4.2 The Continuum-time Limit
To go to the continuum-time limit, let t be a real variable, and replace the

temporal delay value of 1 in Eq. 12 with δ and α with αδ (we’ll eventually
take δ → 0). In addition differentiate Eq. 10 with respect to t to get

dq(xi , t) dΦi (xi , t) dΦi (xi , t)
= −qi (xi , t)[ − dxi qi (xi , t) ]. (13)
dt dt dt
In the δ → 0 limit, assuming q is a continuous function of t, Eq. 12 becomes
dΦi (xi , q)
= α[φi (xi , q) − Φi (xi , q)]. (14)
dt
where from now on the t variable is being suppressed for clarity.
If we knew the dynamics of φi , we could solve Eq. 14 via integrating factors,
in the usual way. Instead, here we’ll plug that equation for dΦ i
dt into Eq. 13.
Then use Eq. 10 to write Φi (xi , q) = constant −ln(qi (xi )). The result is
dqi (xi )
= αqi (xi ) [φi (xi , q) + ln(qi (xi ))]
dt
dxi αqi (xi )[φi (xi ) + ln(qi (xi ))]. (15)
4
To see this is exponential data-aging with exponent γ set γ = -ln(1 − α)).
4.3 Relation with Nearest Newton Descent and Replicator

Dynamics
As mentioned previously, there are many ways to find equilibria, and in par-
ticular many distributed algorithms for doing so. This is especially so in team
games, where finding such equilibria reduces to descending a single over-
arching Lagrangian. One natural idea for descent in such games is to use
the Newton-Raphson descent algorithm. However that algorithm cannot be
applied directly to search across q in a distributed fashion, due to the need
to invert matrices coupling the agents. As an alternative, one can consider
what new distribution p the Newton algorithm would step to if there was no
restriction that p be a product distribution. One can then ask what product
distribution is closest to p, according to Kullback-Leibler distance[9]. It turns
out that one can solve for that optimal product distribution. The associated
update rule is called the Nearest Newton algorithm[29].
It turns out that when one writes down the Nearest Newton update rule,
it says to replace each component qi (xi ) with the exact quantity appearing
on the right-hand side of Eq. 15, where α is the stepsize of the update, and
φi (xi , t) = βE(G | xi , q(i) (t)), as in parallel Brouwer updating for a team
game 5 . In other words, in team games, the continuum limit of having each
player using (bounded rational) best response is identical to the continuum
limit of the Newton-Raphson algorithm for descending the Lagrangian, with
the data-aging parameter α giving the stepsize.
Eq. 15 arises in other yet other contexts as well. In particular, say Φi is
conditional expected rewards (i.e., φi (xi , t − 1) = E(gi | q(., t − 1))). Then
the β → ∞ limit of Eq. 15 reduces to a simplified form of the replicator
dynamics equation of evolutionary game theory[21, 23]. (If the stepsize α is
an appropriately increasing function of E(G) other versions of that dynamics
arise.) This is because in that limit the ln term disappears, and the righthand
side of Eq. 15 involves only the difference between player i’s expected cost
and the average expected cost of all players. This 3-way connection suggests
using some of the techniques for solving replicator dynamics to expedite either
parallel Brouwer or Nearest Newton.
4.4 Convergence and Equilibria
By Eq. 15, at equilibrium, for each i, qi (xi )[φi (xi , q)+ln(qi (xi ))] must be inde-
pendent of i. One way this can occur is if it equals 0. However qi (xi ) can never
be 0, by Eq. 10. This means we have an equilibrium at qi (xi ) ∝ e−φi (xi ,q) .
Intuitively, this is exactly what we want, according to Eq. 10 and our inter-
pretation of φi (xi , q) as an estimate of φi (xi , q). Note also that this solution
means that φi (xi , q) = Φi (xi , q), so that (according to Eq. 14) Φi (xi , q) has
also reached an equilibrium.
5
More generally, Nearest Newton uses this update rule with φi (xi , t) = βE(gi |
xi , q(i) (t)) where each gi (x) = G(x) − D(x(i) ) for some function D. See [29].
When our equilibrium has qi (xi )[φi (xi , q) + ln(qi (xi ))] = A = 0, we have
qi (xi ) ∝ e−qi (xi )φi (xi ,q) . (16)
In light of Eq. 10, this means that Φi (xi , q) = φi (xi , q). So by Eq. 14, Φi (xi , q)
hasn’t reached an equilibrium in this case:
dΦi (xi , q)
= αφi (xi , q)[1 − qi (xi )]. (17)
dt
If both qi (xi ) and φi (xi , q) were frozen at this point, this solution for Φi (xi , q)
would not obey Eq. 12. So either qi (xi ) and/or φi (xi , q) cannot be frozen. In
fact, if φi (xi , q) varies with time, then we know by Eq. 15 that qi (xi ) varies
as well. So in either case qi (xi ) must vary, i.e., this equilibrium is not stable.
Although the dynamics has the desired fixed point, it may take a long time
to converge there. There are several ways to analyze that: One is to examine
the second derivatives (with respect to time) of the qi and/or the Φi . Another
is to examine the time-dependence of the residual error,
e−Φi (xi ,t) e−φi (xi ,t)

rige (xi , t) ≡ − . (18)
dxi e−Φi (xi ,t) dxi e−φi (xi ,t)
The next subsection includes a convergence analysis involving residual errors,
but for a different variant of Brouwer from the ones considered so far.
4.5 Other Variants of Brouwer Updating
Data-aging can be viewed as moving only part-way from the current Φi to

what it should be (i.e. to φi ). An alternative is to dispense with the Φi and φi
altogether, and instead step part-way from the current q to what it should be,
i.e., partially move to the (bounded rational) best response mixed strategy.
Formally, this means replacing Eq. 10 so that the update is not implicit, in
how Φi (xi , t) depends on the past value of q(t − 1) (Eq. 12), but explicit:
qi (xi , t) = qi (xi , t − 1) + α[hi (xi , q(i) (t − 1)) − qi (xi , t − 1)] (19)
where hi (xi , q(i) (t)) is the Boltzmann distribution of what qi (xi , t) would be,
under ideal circumstances, and we implicitly have small stepsize α.
The only fixed point of this updating rule is where qi = hi ∀i. So just
like with continuum-limit parallel Brouwer, we have the correct equilibrium.
To investigate how fast the update rule of Eq. 19 arrives at that equilibrium,
write its error at time t as the residual
rist (xi , t) = qi (xi , t) − hi (xi , q(i) (t))

= qi (xi , t − 1)[1 − α] + αhi (xi , q(i) (t − 1)) − hi (xi , q(i) (t))
= qi (xi , t − 1)[1 − α] + αhi (xi , q(i) (t − 1))
− hi [xi , q(i) (t − 1) + α[h(i) (q(t − 1)) − q(i) (t − 1)]] (20)
where we have assumed that all all players other than i are updating them-
selves in the same that i does (i.e., via Eq. 19), and h(i) (q(t − 1)) means the
vector of the values of all hj=i (xj ) evaluated for q(t − 1).
With obvious notation, rewrite Eq. 20 as
rist (xi , t) = qi (xi , t − 1)[1 − α]

+ αhi (xi , q(i) (t − 1))
− hi [xi , q(i) (t − 1) − αr(i) (t − 1)]. (21)
Now use the fact that α is small to expand the last hi term on the righthand
side to first order in its second (vector-valued) argument, getting the result
rist (xi , t) ≈ ri (xi , t)[1 − α] + α∇hi · r(i) (t − 1) (22)
where the gradient of hi is with respect to the vector components of its second
argument. Accordingly, if rist (xi ) starts much larger than the other residuals,
it will be pushed down to their values. Conversely, if it starts much smaller
than them, it will rise.
There are other ways one can reduce a stochastic game to a deterministic
continuum-time process. In particular, this can be done in closed form for
ficticious play games and some simple variants of it [19, 10].
Acknowledgements: I would like to thank Stefan Bieniawski, Bill Macready,
George Judge, Chris Henze, and Ilan Kroo for helpful discussion.
References
1. Al-Najjar, N. I., and R. Smorodinsky, ”Large nonanonymous repeated games”,
Game and Economic Behavior 37, 26-39 (2001).
2. Antoine, N., S. Bieniawski, I. Kroo, and D. H. Wolpert, ”Fleet assignment
using collective intelligence”, Proceedings of 42nd Aerospace Sciences Meeting,
(2004), AIAA-2004-0622.
3. Arthur, W. B., ”Complexity in economic theory: Inductive reasoning and
bounded rationality”, American Economic Review 84, 2 (May 1994), 406-411.
4. Aumann, R.J., and S. Hart, Handbook of Game Theory with Economic Appli-
cations, North-Holland Press (1992).
5. Axelrod, R., The Evolution of Cooperation, Basic Books NY (1984).
6. Basar, T., and G.J. Olsder, Dynamic Noncooperative Game Theory, Siam
Philadelphia, PA (1999), Second Edition.
7. Bieniawski, S., and D. H. Wolpert, ”Adaptive, distributed control of constrained
multi-agent systems”, Proceedings of AAMAS 04 , (2004).
8. Boutilier, C., Y. Shoham, and M. P. Wellman, ”Editorial: Economic principles
of multi-agent systems”, Articial Intelligence Journal 94 (1997), 1-6.
9. Cover, T., and J. Thomas, Elements of Information Theory, Wiley-Interscience
New York (1991).
10. Fudenberg, D., and D. K. Levine, The Theory of Learning in Games, MIT Press
Cambridge, MA (1998).
11. Jaynes, E. T., ”Information theory and statistical mechanics”, Physical Review
106 (1957), 620.
12. Jaynes, E. T., and G. Larry Bretthorst, Probability Theory : The Logic of
Science, Cambridge University Press (2003).
13. Judge, G., D. Miller, and W. Cho, ”An information theoretic approach to
ecological estimation and inference”, Ecological Inference:New methodological
Strategies (King, Rosen, and Tanner eds.), Cambridge University Press (2004).
14. Kahneman, D., ”A psychological perspective on economics”, American Eco-
nomic Review (Proceedings) 93:2 (2003), 162-168.
15. Lee, C. Fan, and D. H. Wolpert, ”Product distribution theory for control of
multi-agent systems”, Proceedings of AAMAS 04 , (2004).
16. Macready, William, and David H. Wolpert, ”Distributed constrained optimiza-
tion with semi-coordinate transformations”, submitted (2004).
17. Neyman, A., ”Bounded complexity justifies cooperation in the finitely repeated
prisoner’s dilemma”, Economics Letters 19 (1985), 227-230.
18. Osborne, M., and A. Rubenstein, A Course in Game Theory, MIT Press Cam-
bridge, MA (1994).
19. Shamma, J.S., and G. Arslan, ”Dynamic fictitious play, dynamic gradient play,
and distributed convergence to nash equilibria”, submitted (2004).
20. Sutton, R. S., and A. G. Barto, Reinforcement Learning: An Introduction, MIT
Press Cambridge, MA (1998).
21. Tuyls, K., D. Heytens, A. Nowe, and B. Manderick, ”Extended replicator dy-
namics as a key to reinforcement learning in multi-agent systems”, Lecture
Notes in Artificial Intelligence, LNAI, (ECML 2003), (2003).
22. Tversky, A., and D. Kahneman, ”Advances in prospect theory: Cumulative
representation of uncertainty”, Journal of Risk and Uncertainty 5 (1992), 297-
323.
23. Verbeeck, K., A. Nowe, and K. Tuyls, ”Coordinated exploration in stochas-
tic common interest games”, Proceedings of AAMAS-3. University of Wales,
Aberystwyth, (2003).
24. Watkins, C., and P. Dayan, ”Q-learning”, Machine Learning 8, 3/4 (1992),
279-292.
25. Wolpert, D. H., ”Factoring a canonical ensemble”, cond-mat/0307630.
26. Wolpert, David H., ”Finding bounded rational equilibria part 1: Iterative focus-
ing”, Proceedings of the International Society of Dynamic Games Conference,
2004 , (2004), in press.
27. Wolpert, D. H., ”Information theory — the bridge connecting bounded rational
game theory and statistical physics”, Complex Engineering Systems (A. M. D.
Braha and Y. Bar-Yam eds.), (2004).
28. Wolpert, D. H., and S. Bieniawski, ”Adaptive distributed control: beyond
single-instant categorical variables”, Proceedings of MSRAS04 (A. S. et al ed.),
Springer Verlag (2004).
29. Wolpert, D. H., and S. Bieniawski, ”Distributed control by lagrangian steepest
descent”, Proceedings of CDC 04 , (2004).
Evolution of Reciprocal Cooperation in the
Avatamsaka Game
Eizo Akiyama1 and Yuji Aruka2

1
Institute of Policy and Planning Sciences, University of Tsukuba, Tel
+81-298-54-9455, Fax +81-298-53-3848 eizo@sk.tsukuba.ac.jp
2
Faculty of Commerce, Chuo University, aruka@tamacc.chuo-u.ac.jp
Summary. The Avatamsaka game is investigated both analytically and using com-
puter simulations. The Avatamsaka game is a dependent game in which each agent’s
payoff depends completely not on her own decision but on the other players’. Con-
sequently, any combination of mixed strategies is a Nash equilibrium.
Analysis and evolutionary simulations show that the socially optimal state be-
comes evolutionarily stable by a Pavlovian strategy in the repeated Avatamsaka
game, and also in any kind of dependent game. The mechanism of the evolutionary
process is investigated from the viewpoint of the agent’s memory and mutation of
strategies.
1 The Avatamsaka Game as a Dependent Game
1.1 Robinson Crusoe Economy and Multiple-person Games
In our daily life, we must sometimes make decisions to improve our current
situation. If utilities for the results depend only upon our own decision, the
problem is usually called an optimization problem, which a solitary person,
such as Robinson Crusoe (or a person whose payoff is unrelated to the deci-
sions of other people) meets (“Robinson Crusoe economy” [3]). What matters
first in this kind of situation is the possibility of constructing the function it-
self. Once the function is constructed, solving for the optimal behavior is a
computational problem to determine the strategy that optimizes the func-
tion. As von Neumann and Morgenstern pointed out in their book “Theory
of Games and Economic Behavior,” the difficulty in solving for the optimal
strategy in one-person games, such as the Traveling Salesman Problem, is a
technical rather than conceptual one.
In multiple person problems, however, the result for each one will depend
in general not merely upon his decision but on those of the others as well.
Moreover, all participants desire maxima simultaneously. Consequently, it is
308 Eizo Akiyama and Yuji Aruka
usually not possible for a player to decide his best behavior without consid-
ering others’ decision-making processes. Such a situation in general is called
a game1 .
1.2 Independent / Dependent Game
Let us refer to the Robinson Crusoe problem, in which decisions of others do

not have any effect on his utility, as an independent game; Robinson Crusoe’s
effort can be termed independent optimization. (In the context of the evolu-
tionary model, it will be called an independent adaptation.) In this paper, we
consider a class of games that is opposite of the independent game: dependent
games.
In a dependent game, the result for each agent will depend merely on the
decisions of others, but not at all on his own decision. Every player’s effort
to increase his profit in a dependent game is called a dependent optimization.
(In the context of the evolutionary model, it will be called a dependent adap-
tation.) Any multiple-person problem is representable as a formulation that
lies between the independent game and the dependent game, in principle.
By definition, any change of a player’s action in a dependent game does
not directly improve her utility. However, if the game is repeated there is a
chance her current action might serve as a signal to induce beneficial behaviors
from others in the future.
For an example of dependent games, let us consider internet auction sites.
Many internet auction sites adopt a “user rating system,” with which each
buyer rates his experiences with sellers, in order to prevent dishonesty among
sellers. In this case, rating sellers high or low does not affect the current utility
of the buyers at all, which exemplifies the dependent game situation. Note that
not a few members of auction sites repeatedly participate in auctions there.
As a result, the rating behaviors are repeated for the future, which usually
motivates the buyers to give honest ratings. If their participation were not
repeated, it would not matter for a player, for example, to rate a good seller
“low” without any consideration (and forever leave the auction site).
1.3 The Avatamsaka Game
As the simplest example of dependent games, let us consider the Avatamsaka

Game2 [1]. Its payoff matrix is shown in Table 1. In this game, a player’s
1
In the theory of games, von Neumann and Morgenstern had an insight into the
fact that this problem is not an ordinary maximization problem and that the
formalization of this type of problem should take the form of a matrix. As a
result, they succeeded in formalizing the problem as one in which every individual
“can determine the variables which describe his own actions, but not those of the
others.” In addition, “those alien variables cannot, from his point of view, be
described by statistical assumptions.”
Evolution of Reciprocal Cooperation in the Avatamsaka Game 309
Table 1. Payoff matrix for the Avatamsaka game: Two matrices below show the
points gained by Player 1 (left) and Player 2 (right). The columns indicate the
actions that Player 1 would choose and the rows indicate those for Player 2. Because
we only deal with symmetric games (i.e. the payoff structures from the viewpoints
of Player 1 and of Player 2 are identical), such as the game in this table, we will
show the payoff matrices only for Player 1 in the remainder of this paper.
Payoff for player 1 Payoff for player 2
DC DC
D 1 2 D 1 1
C 1 2 C 2 2
points depend only on the other agent’s behavior. If player 2 chooses behavior
D (defect), the point player 1 can gain is only 1.0, no matter what behavior
player 1 chooses. If player 2 chooses behavior C (cooperation), player 1 is
bound to receive 2.0.
In this game, any combination of both players’ mixed strategies is a Nash
equilibrium because neither player has an incentive to change their strategy. In
addition, any mixed-strategy is a neutrally stable strategy from the viewpoint
of evolutionary game theory.
1.4 Repeated Interactions in a Society, Players’ Cognitive

Abilities, and Mistakes
Let us imagine that we are actually under circumstances similar to the

Avatamsaka game. In this case, various thoughts might come to mind such as
“Should I be satisfied with his defection, even though we are in a Nash equi-
librium state?” “Can I somehow induce his cooperation?” “Can I somehow
use my behavior as a signal to him?” etc. The reason we may not be satisfied
with our current game play is that, if we actually encounter a situation like
2
Avatamsaka is a well-known Mahayana Buddhist Sutra. The late Shigeo Kamata,
Professor Emeritus, University of Tokyo, who was working under the field of
Buddhist philosophy, skillfully illustrated the situation of Heaven and Hell in
view of Avatamsaka. (See Kamata [6], pp.167-168.)
Suppose that two men are seated across from each other at a table. They are
bound with ropes except for one arm. Each is given an overly long spoon. Thereby,
they cannot serve themselves because of the awkwardness of the long spoon, but
there is sufficient food for both of them on the table. If they cooperate in concert
to provide each other with a meal, they can both be happy. This defines the
Heaven, or Paradise. Alternatively, one will be kind enough to provide the other
with a meal, but the other might not have the same feeling of cooperation. This
case pays the other, which must give rise to a feeling of hate for the inconsiderate
opponent. This describes a situation denoted as Hell. (See [1] for the philosophical
and sociological significance of this game.)
this game, something other than the activity of “analyzing the Table 1 in
the way of traditional game theory” appears in our mind, which might cause
mutual cooperation under Avatamsaka game interactions.
For example, we usually interact not only with a certain individual, but
also with other community members in the real world. Therefore, it is neces-
sary for us to consider, comprehensively, relationships with various types of
individuals (various interactions in a society). In addition, we sometimes, or
often, consider not only current, but also future relationships with the per-
son (repeated interactions). In this case, although my current action has no
meaning for me now, but it might have some meaning as a “signal” to derive
cooperation of the opponent in the future. Whether or not repeated interaction
works as a signal to derive mutual cooperation depends on the perceptions and
cognition of both players. Note that, in the course of repeated interactions,
we must also consider the fact that we are beings who cannot be free from
mistakes in the real world. And so, cognitive abilities to deal with players’
mistakes are crucial for formation of cooperation (players’ cognitive abilities
and mistakes). Furthermore, we do not fix our strategy. We sometimes change
our strategy to adapt to our current situation (mutation / modification of the
strategies).
This paper investigates the following points to study the evolution of co-
operation under a dependent game: the Avatamsaka game.
1. How can we derive cooperative behavior of the opponents in Avatamsaka
situations, when we live in a society in which we interact repeatedly with
various players who might sometimes make mistakes? Players surely need
cognitive abilities to induce another’s cooperation. To what extent are
such abilities required for deriving cooperation?
2. How do social phenomena change with players’ cognitive abilities? For
example, an agent who remembers only the previous actions of the oppo-
nents, such as so-called “Tit-For-Tat strategy” in the Prisoner’s Dilemma
game, may bring about completely different social phenomena from agents
who remember the previous behaviors of themselves and of their oppo-
nents.
3. Do capricious changes of strategies (mutation) have a fundamental effect
on the evolutionary phenomena?
2 The Model
This study addresses a multi-stage (or repeated, iterated) Avatamsaka game;
we refer to one iteration in the game as a stage game. Furthermore, we refer
to C or D in a stage game as an action, while a complete plan in a multi-
stage game for a player to decide his action based on the past information as
a strategy. We investigate evolutionary phenomena in the Avatamsaka game
using an evolutionary model whose formulation was presented by [4] for the
study of the Prisoner’s Dilemma game.
2.1 Evolutionary Dynamics with Mutational Effect
Suppose that there are infinitely many agents in a population (a game world).
All agents in the population play a round-robin tournament with each other
in a generation. In each match, an infinitely iterated Avatamsaka game is
played between two agents. Suppose that the total population of the agents
in the game world is 1.0 and that the fraction of the agents having strategy
i (i = 0, 1, . . . , n − 1) in a generation t is xi , where i xi = 1.0.
The fitness fi of strategy i is defined as the expected value of the pay-
offs in all games in which strategy i is involved in the population. Because
the expected payoff of strategy i depends on the probability of meeting with
strategies against whom he can gain high (or low) payoffs, fi depends on the
strategy distribution x = (x0 , x1 , . . . , xn−1 ) ∈ ∆ over the population, where
∆ = {x : ∀i xi ≥ 0 and i xi = 1.0}
Assume that the population share of a strategy either increases or de-
creases according to the fitness of the strategy and that a small fraction of
newborn agents appear as mutants of other strategies. Considering the fact
that the average fitness of all agents is Pxixfiifi , the population share xi of
i
strategy i in the next generation t + 1 would be the following.

fi
xi = xi + u /(1 + nu), i = 0, . . . , n − 1,
i xi fi
where u is the constant term representing the effect of the uniform and
time-independent mutation: the mutation rate. For computer simulations, we
choose two cases u = 0 (no mutation) and u = 0.0004 (with mutation) to il-
lustrate the effects of mutation. The change of u does not substantially affect
the results if 0 < u << 1.
2.2 Iterated Game with Action-noise and the Strategies with Past
Memory
In the repeated Avatamsaka game, each agent uses a strategy to determine

her action in the next stage game based on the memory of actions that she
and the opponent made in previous rounds. We investigate the following three
cases in this paper: (1) Each agent continues the initial action without any
consideration of past information. This condition is denoted herein as the
“m = 0 (no memory)” case. (2) Each agent chooses her action referring to
the previous action of the opponent (“m = 1”) (3) Each agent determines
her action referring to the previous actions of herself and of the opponent
(“m = 2”). As shown below, the number of possible pure strategies in this
repeated game depends on the agents’ memory size. Let us denote the nth
strategy with memory size of m by Snm .
In the case of no memory (m = 0), only two possible pure strategies exist:
“Always D (which we denote as AllD or S00 )” and “Always C (AllC or S10 ).”
In the case of m = 1, we denote a strategy of an agent as p0 p1 (p0 , p1 ∈ [0, 1]),

where p0 (p1 ) is the probability to play “C” when the opponent’s last move
was “C” (“D”). For example, the strategy “00” (i.e. p0 = 0, p1 = 0) always
plays “D.” Four possible pure strategies exist for m = 1. They are 00, 01, 10,
11, which we denote respectively as S01 , S11 , S21 , and S31 . Note that 00=S01 and
11=S31 are equivalent to AllD and AllC. 10=S21 means that the agent acts in
the same manner as the opponent did previously: the so-called Tit-For-Tat
(TFT) strategy. 01=S11 means that the agent chooses the opposite action of
the opponent’s previous action: the Anti-TFT (ATFT) strategy.
The strategy in the case of m = 2 can also be described as p0 p1 p2 p3
(pi ∈ [0, 1]), where p0 , p1 , p2 and p3 represent the probabilities to play “C”
when, in the previous round, (the agent’s move, the opponent’s move) = (C,
C), (C, D), (D, C) and (D, D), respectively. There are 16 possible pure
strategies, such as 0000, . . . , 1111 = S02 S152
. Similarly, the total number of
2m
possible pure strategies is 2 for the memory size of m.
For simplicity, we assume that each game would be repeated ad infini-
tum between two agents chosen from the population. Moreover, agents are
assumed to make mistakes with probability p during the repetition. (e.g. The
agent with the strategy to play “C” may sometimes play “D” by mistake
against his will.) This means that uncertainty (action-noise) exists in agents’
cognitive functioning. Taking into account the fact that the transition of the
probability distribution for the 2m states above is a Markov process, the tran-
sition matrix can be defined uniquely depending on the two agents’ strategies.
The probability distribution in the steady state corresponds to the normalized
eigenvector with eigenvalue 1 of the transition matrix as long as p > 0. The
average payoff in an infinitely iterated game can be given from the probability
distribution of the steady state and the stage game payoff matrix. (See [4].) A
slight change of p does not substantially affect the analysis result introduced
in the next section as long as p << 1. (Basically we assume 0 < p << 1.
Theoretical values found in the next section such as the equilibria of game
dynamics are the values gained at p → 0 limit. For computer simulations, we
use p = 0.01.)
As a short note for strategies with m = 2, S02 =0000, and S15 2
=1111 mean
2 2
AllD, AllC and S10 =1010 correspond to TFT, and S5 =0101 to ATFT. More-
over, S82 =1000 is “GRIM” (that plays “C” only when the previous choices of
the both players were “C”), S92 =1001 is so-called “PAVLOV” [5]. The last
two strategies can be formed only if m = 2. Note that the nomenclature used
in the above is the one usually used in “Prisoner’s Dilemma” studies. For that
reason, it might be sometimes misleading to use the above strategy names in
other games, but we use those names in this paper for simplicity.
3 Analysis and Simulation Results

In this section, we show both the analytical results and the results of computer
simulations of the evolutionary Avatamsaka game. To capture the nature of
evolutionary phenomena, we especially investigate (i) the dynamics of the
population share of the strategies, and (ii) the dynamics of agents’ actions in
each iterated Avatamsaka games.
3.1 Agents Without any Memory of Past Actions (m = 0)

When agents have no memory of past actions (m = 0), only two possi-
ble pure strategies exist: AllD (S00 ) and AllC (S10 ). The payoff matrix for
m = 0 case is given in Table 2. Obviously, any combination of both play-
ers’ mixed strategies is a Nash equilibrium. Therefore, the subset of strate-
gies x ∈ ∆ that are in Nash equilibrium with themselves, ∆N E = {x ∈
∆ : (x, x) is a Nash equilibrium}, is equivalent with the whole mixed-strategy
space ∆.
Furthermore, any mixed strategy can be a Neutrally Stable Strategy
(NSS), but there is no Evolutionary Stable Strategy (ESS) in this game. (The
set of NSS, ∆N SS = ∆, and the set of ESS, ∆ESS = ∅.)
Table 2. Payoff matrix of the iterated Avatamsaka game for m = 0 strategies: The
procedure to derive this matrix is given in Section 2.2. Because m = 0 matrix is
simple, we can understand its intuitive meaning. If the opponent is AllD, a player’s
payoff is (1 − p) + 2p = 1 + p because the opponent’s action is “D” with probability
1 − p and “C” with probability p, where p is the probability of agent’s making
mistake.
AllD AllC
AllD (S00 ) 1 + p 2 - p
AllC (S10 ) 1 + p 2 - p
An example of population dynamics in the case where the agents have no

memory (m = 0) and no mutation occurs (u = 0) is shown in Fig. 1(a). Fre-
quencies of the strategies remain exactly the same as in the initial generation.
In the Avatamsaka game, an agent always obtains an identical payoff whether
he chooses C or D unless the other acts differently. Because we cannot change
the future actions of the strategies AllD and AllC (S00 and S10 ) that play
“D” and “C” forever by their definition, the fitness of either AllD or AllC
is determined only by the distribution of AllD and AllC over the population.
Therefore, the respective fitness values of AllD and AllC are identical: f0 = f1 .
Consequently, the population share of AllC or of AllD does not change with
generations. This population dynamics corresponds to the fact that ∆N SS of
this game is identical to ∆.
1 1
0.8 0.8
0.6 0.6
ALLD ALLD
0.4 0.4
0.2 ALLC 0.2 ALLC
500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
(a) (b)
Fig. 1. Population dynamics in the Avatamsaka game world in which agents have
no memory of past actions, m = 0: The fraction of the strategies in the population is
shown as a function of generation. (a) When the mutation rate u = 0, the population
share of each of the two strategies does not change with generations. (b) When
u = 0.0004, both converge to 0.5.
If the mutation process is introduced (u = 0.0004) in this m = 0 Avatam-

saka game world, the observed phenomenon is altered completely. In this case,
the frequencies (x0 , x1 ) of the strategies (AllD, AllC) converge to the equal
population share, (1/2, 1/2), with generations (Fig. 1(b)), and the entropy of
the configuration of the population is maximized.
3.2 Considering the Opponent’s Previous Action (m = 1)
Payoff Structure
When agents have a memory of the opponent’s previous action (m = 1), they
can implement four kinds of pure strategies, AllD, ATFT, TFT and AllC,
which are denoted respectively as S01 , S11 , S21 , and S31 . The payoff matrix for
m = 1 case is given in Table 3.
In this case, ∆N E = {(u(1 − v), v/2, v/2, (1 − u)(1 − v)) : u, v ∈ [0, 1]}
because the expected payoff for any mixed strategy x ∈ ∆ against x∗ ∈ ∆N E
is always 3v/2 + (1 − v)(1 − u − p), which is independent of the value of x.
Furthermore, the set of NSS, ∆N SS = {((1 − v)/2, v/2, v/2, (1 − v)/2) : v ∈
[0, 1]}. However, there is no ESS.
Table 3. Payoff matrix of the iterated Avatamsaka game for m = 1 strategies: The
procedure to derive this matrix is given in Section 2.2.
AllD ATFT TFT AllC

AllD (S01 ) 1 + p 2 − 2p − 2p2 1 + 2p − 2p2 2−p
ATFT (S11 ) 1+p 3/2 3/2 2−p
TFT (S21 ) 1+p 3/2 3/2 2−p
AllC (S31 ) 1 + p 1 + 2p − 2p2 2 − 2p − 2p2 2−p
1 1
0.8 0.8
ALLD ALLD
0.6 ATFT 0.6 ATFT
0.4 TFT 0.4 TFT
0.2 ALLC 0.2 ALLC
(a) 20 40 60 80 100
(b) 20 40 60 80 100
Fig. 2. Avatamsaka game: m = 1, mutation rate = 0
Population Dynamics
The initial strategy distribution has no influence on the future population
dynamics in the case of m = 0, but it does in the case of m = 1 with no
strategic mutation, in most cases.
In this game world, where each agent may have a memory of the previous
move of the other, the population share x = (x0 , x1 , x2 , x3 ) converges either
to (1, 0, 0, 0) or (0, 0, 0, 1) (the dominance by AllD (S01 ) or by AllC (S31 )) with
generations in most cases other than the knife-edge cases, for which x is exactly
in ∆N E in the initial generation. (x ∈ ∆N E , of course, does not change with
generations.) Convergence depends on the initial distributions of ATFT(S11 )
and TFT(S21 ) (Fig. 2). AllC continues to increase when the frequency of TFT
is greater than that of ATFT (x2 > x1 ); the AllD population continues to
increase when x1 > x2 .
Because AllD always chooses D regardless of past actions, the average
payoff of any opponent of AllD is always 1 in the Avatamsaka game, while
that of the opponent of AllC is always 2. Consequently, the effect of the
strategy distribution of AllD/AllC on the fitness of any strategy is always the
same: it does not affect the population dynamics.
It is ATFT or TFT that may affect the population dynamics. When meet-
ing with AllC, TFT always chooses “C” in response to AllC’s previous “C,”
while ATFT always chooses “D” (except for the change of the action by noise.)
That is, TFT benefits AllC and, in the same way, AT F T benefits AllC.
As shown above, the appearance of strategies with m = 1 gives rise to
population dynamics that are dependent on the strategy distribution in the
initial generation. Note that the relation x2 > x1 or x1 > x2 will not change
with generations because the expected payoff of T F T and that of AT F T are
identical, which can be known by Table 3. Consequently, if x2 > x1 in the
initial generation, the fraction of AllC continues to increase while that of AllD
continues to decrease.
Effect of Mutation
The power of the convergence to the AllD/AllC society observed in the above
is not so strong because it results solely from the difference between the rel-
ative fraction of TFT to ATFT, not on the interaction within AllD (AllC)
0.8
ALLD
0.6 ATFT
0.4 TFT
0.2 ALLC
500 1000 1500 2000 2500 3000
Fig. 3. Avatamsaka game: m = 1, mutation rate u = 0.0004
agents. When the mutation is introduced to the evolutionary process, x con-

verges to (1/4, 1/4, 1/4, 1/4) with generations (Fig. 3). What happens here
is the triumph of the mutational effect that increases the entropy of the pop-
ulational configuration over the adaptation power, which is also observed in
the m = 0 case.
The world of m = 1 shares the same nature with that of m = 0 in the sense
that all existing strategies cannot conduct independent optimization because
the increase of the fitness of AllC (AllD) depends completely on TFT(ATFT).
Consequently, the fitness of AllC also approaches that of AllD with generations
because the fraction of TFT, whose fitness is always identical to ATFT’s,
approaches that of ATFT by mutational effect.
The mutation mechanism in principle has the capability of increasing the
entropy of the population structure. On the other hand, there usually exists in
a game the power that comes from players’ adaptation to the payoff matrix.
We can say that the Avatamsaka game at m = 0 or m = 1 is one of the
games where the effect of the adaptation is overshadowed by the mutational
effect that increases the entropy. In other words, dependent adaptation (See
section 1.2) by itself is insufficient to overcome the mutational effect.
3.3 Considering the Player’s and the Opponent’s Actions (m=2)
Let us consider the case of m = 2, in which an agent can refer to both her
own and the opponent’s previous action to decide the next action. The salient
difference is that, while the fitness of each of m = 0, 1 strategies depends
completely on the population shares of the other strategies, several m = 2
strategies exist whose fitness depends on the shares of all strategies including
that of the strategies themselves.
The payoff matrix for m = 2 strategies is a little complicated function of p;
because we have insufficient space for describing the entire matrix (and doing
it does not seem so informative), we visualize the payoff structure in Fig. 4 by
computing payoffs for p = 0.01 and by drawing them as a figure and a graph.
It can be confirmed analytically that the “PAVLOV3 (1001=S92 )” is the ESS
of this game. That fact can also be confirmed with Fig. 4.
3
As in the study of the iterated Prisoner’s Dilemma game [5], PAVLOV is a good
strategy to form and maintain mutual cooperation under the circumstances with
Population dynamics of the m = 2 Avatamsaka game without mutation

are shown in Fig. 5. In this setting, the game world converges to the exclusive
winning of PAVLOV. Note that the result does not substantially differ even
if there is mutation.
Dominance by PAVLOV is realized for any initial condition as long as
the initial population share corresponds to the interior of the mixed strategy
space ∆. In other words, “independent adaptation” through the use of a m = 2
strategy, PAVLOV, overshadows the effect of mutation.
Let us consider why PAVLOV is strong in this game. The PAVLOV policies
are as follows: (1) If I feel good (the opponent’s action is C), keep doing the
current action, whether C or D. (2) Never comply with the situation if I
feel uneasy (the opponent’s action is D) by sending a signal to the opponent
by changing my action4 . With these two features, PAVLOV can maintain a
reciprocal relationship once it is formed because a Pavlovian player continues
C as long as he feels good.
Furthermore, two PAVLOVs interacting with each other can quickly re-
cover from betrayal relationships caused by action noise. Assume that two
PAVLOVs play a repeated Avatamsaka game, as shown below. PAVLOV’s
action should be C if both cooperated (CC) in the previous round.
PAVLOV | C C D* D C C C ...
PAVLOV | C C C D C C C ...
In this figure, the mark “*” denotes the betrayal of the upper player by mistake
(noise) that is supposed to happen with probability p in this study. While the
upper player continues D, the lower Pavlovian player, who feels uneasy, sends
a signal to the upper player by changing her action into D in the next round
(DD), thereby engendering the recovery of mutual cooperation CC two rounds
later.
Analysis of the m = 2 payoff matrix shows that PAVLOV is the ESS for
m ≤ 2. Furthermore, some additional consideration allows us to know that
PAVLOV is also evolutionarily stable in the whole pure-strategy space including
the longer-memory case if w < 1, in which w is a constant probability to
another round of repetition5 .
action noise. However, PAVLOV is not an evolutionarily stable strategy in the

Prisoner’s Dilemma with noise because the best response to PAVLOV is AllD in
the Prisoner’s Dilemma.
4
Choosing C or D functions only as a “signal” in the Avatamsaka game because
the choice does not affect the player’s own payoff.
5
The only thing players in the Avatamsaka game should do is to derive the other’s
C as frequently as possible. In so doing, the best response to PAVLOV’s D that
happens by mistake should be D in the next step, which will stop the PAVLOV’s
continuing D as quickly as possible. Therefore, the best response should have a
code ((a) CD → D ). The best-response strategy that has fallen into mutual
defection with PAVLOV would recover from the tragic state toward the mutual
cooperation in the next step ((b) DD → C). The best response strategy, of course,
0 9 15
0
1.8
1.6
9 1.4
1.2
(a) 15 (b) 2 4 6 8 10 12 14
Fig. 4. (a) An illustration of the payoff matrix of the m = 2 Avatamsaka game: The
vertical axis shows the pure strategy of player 1 and the horizontal axis shows that
of player 2. The figures 0, 9, 15 on the vertical and horizontal axes correspond to
pure strategies S02 , S92 , S15
2
, respectively. The shade on the tiles represents the payoff
for the iterated Avatamsaka game with p = 0.01. The higher the payoff, the whiter
the color on the correspondent tile. We can visually know the best-response relations
among m = 2 pure strategies in this figure. (b) The payoff when facing PAVLOV
(S92 ); p = 0.01 is plotted as a function of the strategy number. Both (b) and the
arrow in (a) indicate that the only best response for PAVLOV is PAVLOV itself.
(The payoff for PAVLOV against PAVLOV is 2 − 3p + 6p2 − 4p3 .)
0.8 0
1
2
3
0.6 4
5
6
7
8
0.4 9
10
11
12
0.2 13
14
15
20 40 60 80 100
Fig. 5. Population dynamics with m = 2 and u = 0.
would retain mutual cooperation ((c) CC → C) If the best-response strategy has

produced a mistake, it would defect again ((d) DC → D) to revert quickly to
mutual cooperation with PAVLOV. (Otherwise, PAVLOV would continue D.)
(a)–(d) as a whole shows that the only strategy that can derive most frequent
cooperation from PAVLOV is PAVLOV itself, provided w < 1. (If w = 1 (infinite
repetition), a strategy that differs only in his first, say, 100 rounds from PAVLOV
can have exactly the same payoff as PAVLOV, on average.) Consequently, if all
players’ strategies are PAVLOV, they are in a strong perfect equilibrium [2], and
so PAVLOV is the ESS.
4 Discussion
This paper shows evolutionary phenomena in a dependent game, the Avatam-
saka game. Let us sum up the results.
4.1 Memory Size, Mutation and Social Phenomena
First, the memory size generally has a large effect on evolutionary dynamics.
For a memory size of m = 0, strategic interactions do not have any effect on
the population dynamics. Consequently, the population shares of the strate-
gies do not change from the initial generation. When m = 1, a strategic
interaction becomes more influential than in the case of m = 0. As a result,
the population converges to either AllC or AllD state, but the convergence
depends strongly on the initial configuration of population. If the population
consists of agents who consider the previous actions of both themselves and
the opponent (m = 2), the cooperative society is self-organized in the popula-
tion by PAVLOV for any initial distribution of strategies that corresponds to a
mixed strategy in the interior of ∆. However, further extension of memory size
(m ≥ 2) has no substantial effect on evolutionary phenomena: PAVLOV or
its analog always dominates the population. Particularly, if w < 1, PAVLOV
alone dominates the population. In this sense, an effective memory-size of 2
exists in the Avatamsaka game.
Second, the effect of mutation is remarkable in cases of the Avatamsaka
game with m = 0, 1, in which only dependent adaptation is possible. However,
in the case of m = 2, the mutational effect is overshadowed by the indepen-
dent adaptation by PAVLOV, so that the introduction of mutation cannot
have a strong effect on evolutionary phenomena. Summing up, investigation
of the Avatamsaka game shows clearly that the entropy of a populational con-
figuration is maximized by mutation if there is only dependent adaptation,
while the entropy can be decreased when there is an independent adaptation
for which a certain size of memory is required for agents in the Avatamsaka
game.
4.2 Efficiency in the Dependent Game, and the Origin of

Inefficiency
Third, PAVLOV is the only ESS in the Avatamsaka game with m ≤ 2. In

addition, PAVLOV is the only ESS in the Avatamsaka game for any m if
w < 1, where w is a constant probability of another repetition.
Note that, in any kind of dependent game with any number of players and
with any number of strategies, a socially optimal state in which everybody
mutually cooperates is achievable as an evolutionarily stable state through
PAVLOV-like strategies. Let us consider a Pavlovian strategy used by an
agent who continually changes his action until all others each chooses the
most beneficial action for him; that is, the Pavlovian strategy keeps sending
signals by changing his action to the others until he feels comfortable. It can be
easily understood that all players’ mutual cooperation is the only stable state.
Furthermore, it is the Pavlovian strategy that derive most frequent coopera-
tion of opponent Pavlovians, which means that the best response to Pavlovian
strategies is the Pavlovian strategy itself, and that Pavlovian strategy is ESS
(See footnote 5 in Section 3.3)
Summing up, the world of the dependent game will eventually engender
a stable state that is socially optimal. The world of the independent game,
in which players each independently confront an optimization problem, will
also lead to a socially optimal state wherein the only thing that agents should
do is to choose their own optimal strategies, if we disregard computational
complexity of the problem. Although players in dependent games should have
memories of previous actions and players in independent games do not, both
the dependent and the independent games should engender socially optimal
states through adaptation.
Seen from another perspective, the origin of social inefficiency that we
observe in reality arises from both technical and conceptual difficulties: Tech-
nical difficulty lies in computational problems for players to calculate optima,
while conceptual difficulty comes from the existence of the entanglement of
dependent and independent game structures (e.g. the Prisoner’s Dilemma)
but neither from the structure of completely dependent games nor from that
of completely independent games.
Acknowledgment
This research was partially supported by the Ministry of Education, Sci-
ence, Sports and Culture, Grant-in-Aid for Scientific Research (B), 2002–2003,
14580486. The first author (EA) is partially supported by a Ministry of Ed-
ucation, Science, Sports and Culture, Grant-in-Aid for Young Scientists (B),
2002–2003, 14780342.
References
1. Aruka, Y. (2001), “Avatamsaka Game Structure and Experiment on the Web,”
in Aruka, Y. (ed.), Evolutionary controversies in Economics, Springer, Tokyo,
115–132.
2. Boyd, R. (1989), “Mistakes allow evolutionary stability in the repeated pris-
oner’s dilemma game,” Journal of Theoretical Biology 136,47–56.
3. von Neumann, J. and O. Morgenstern (1944), Theory of Games and Economic
Behavior, Princeton University Press, NJ, USA
4. Nowak, M. A. and Sigmund, K. (1993), “Chaos and the evolution of coopera-
tion.” Proc. Natl. Acad. Sci. USA 90, 5091–5094.
5. Nowak, M. A. and Sigmund, K. (1993), “A strategy of win-stay lose-shift that
outperforms tit-for-tat in the Prisoner’s Dilemma game.” Nature 364, 56–58.
6. Kamata, S. (1988), Kegon no Shiso (The Thought of Avatamsaka), in Japanese,
Kodan-sha, Tokyo.
Game Representation - Code Form
Maria Cristina Peixoto Matos1 and Manuel Alberto M. Ferreira2

1
Instituto Politécnico de Viseu - Escola Superior de Tecnologia de Viseu
Departamento de Matemática
Campus Politécnico 3504-510 Viseu PORTUGAL
cristinapeixoto@mat.estv.ipv.pt
2
Instituto Superior de Ciências do Trabalho e da Empresa Departamento de
Métodos Quantitativos
Av. das Forças Armadas 1649-026 Lisboa PORTUGAL
vice.presidente@iscte.pt
Summary. When we analyze a game we want a full representation, i.e., we want

to be able to look at a game and perceive all of its details. If it allows us to visualize
the respective solution all the better. This aim was the reason for researching a
new game representation which is described in this paper. The code form of a game
consists mainly in a table containing all of the information with no suppression or
”adulteration” with regards to the game. Initially presented at Fifth Spanish Meeting
on Game Theory & European Voting Games (Seville 2002), based on examples,
code form is put forth in this paper thoroughly. We begin by presenting code form,
exemplifying it with four games. We finish proving the existence of Nash Equilibrium
in code form games.
KEY WORDS: Game theory, Code Form.
1 Introduction
The idea of a new game representation occurred when we analyzed a game

with 3 players in the normal form. When we model a game we want to capture
the most of the relevant details. A game may have a complex temporal and
information structure. This structure could be very significant to understand
the way the game will be played. Some of these details are not considered
either in the normal form or in the extensive form of a game. It seemed
that a representation that gave us a global picture of the game would be
the ideal. As a consequence of this idea, we tried a game representation that
agglutinated the information of the extensive form and of the normal form
and add some additional information. The fundamental of the code form is a
table where the strategies available to any player are codified. We imagined
the code form representation and we verified that this representation has great
advantages when the number of players is greater or equal than , when the
322 Maria Cristina Peixoto Matos and Manuel Alberto M. Ferreira
game has imperfect information, when the game is a repeated one and mainly
when it is sequential. We begin for presenting the mathematical code form
formalization and then we present four examples, a voting game, a market
entry game, a bargaining game and the matching pennies. In the voting game
we have three players that have to approve a pay rise for themselves by voting.
This is a sequential game that is played only once. In the market entry we
have two firms. One of them has a monopoly market and the other decides
to enter or not. This is a game with sequential and simultaneous moves that
is played only once. The bargaining game is a repeated game with sequential
moves where two players are negotiating how they split a given benefit. Each
one hoping to retain the bigger share. Finally in the matching pennies we
have two players. Each player has a coin hidden in a hand. Then they show
simultaneously the coins. One of the players gets the coin of the other if the
both coins show the same face. Otherwise is the other player that gains his
opponent coin. This is a game with simultaneous moves that is played only
once. We finish proving that the code form verifies the Nash Theorem once
the model definition verifies the conditions of this theorem.
2 The Representation of a Game in Code Form
Game theory is a highly developed science, which allows vast and interesting
outcomes in the classification, formulation, and resolution of situations both
in business and in human relations. This scientific area is also useful in the
management of day-to-day conflicts in all domains, whether human or orga-
nizational, and even in all situations that involve strategic interaction. As in
other sciences, game theory is based on models. The models of this discipline
are rigorous abstract presentations of classes of situations taken from real life.
Its abstraction allows a wide range of phenomena to be studied. Obviously
the main objective of the models is to present solutions. ”Solutions” refer to
a systematic description of the consequences resulting from one type of game.
In following section, we present the mathematical formalization of a new game
model.
3 Code Form Game - Mathematical Formalization
Definition 1. A code form game consists of a finite table, with evident exten-
sion in the case of infinite moves and infinite players, where only some cells
are filled. We fill the cells in the same order that the game is played. For that
we need:
• R = {1, 2, ..., R} , a set of rounds.

Game Representation - Code Form 323
• J = {1, 2, ..., J}, a set of moves.
• C = {1, 2, ..., C}, a set of columns.
• L = {1, 2, ..., L}, a set of rows.
• N = {1, 2, ..., N }, a set of players.
• En = {e1 , e2 , ..., eE }, a set of strategies available to each player.
• E = E1 × E2 × ... × EN , a set of all such strategy profiles - space of strate-

gies profiles.
RN : L × {1} → R
• , a function that indicate de round number.
ai1 → r
P N : L × {2} → J
• , a function that indicate de move number.
ai2 → j
JE : L × {C} → N × EN
• ,c = 1, 2, C,a function that indicates who moves
aic → (n, en )
and what action is played.
P J : L × {C} → IRn
• ,a function that gives the payoff of every players
aic → (u1 , u2 , ...uN )
where un : E → IR is a von Neumann-Morgenstern utility function.
We denote when RN (ai1 ) = RN (ai−1,1 ), P N (ai2 ) = P N (ai−1,1 ) and
JE(aic ) = P N (ai−1,c ), the cells are not filled. The row is changed when the
play change. The column is changed when the player change.
4 Examples
4.1 Voting for a Pay Rise
Let us consider a game where three company board members have to approve
a pay rise for themselves by voting. If two of the members vote in favour of
the rise it will pass, otherwise the salary will remain the same. Whatever the
decision taken, it will be conveyed to the company employees. For this reason,
while each of the board members wants to vote against the measure, he wants
it to be approved. This position will guarantee a political advantage in so far
as, at the eyes of the employees, the administrator will be recognized as a
conscientious and unambitious businessman. In this way each player has two
possible strategies: approve the rise or to disapprove the rise. The payoffs are
established taking into account each player’s preference: approval of the rise
without his vote.
Thus:
• Rise approved /Yes: 1

• Rise approved /No: 2
• Rise disapproved /Yes:-1
• Rise disapproved/No: 0
Suppose, as this is a sequential game, that player 1 votes first, player 2
votes second and finally player 3 votes.
The following table illustrates the code form game representation.
Fig. 1. Voting for a Pay Rise
Reading from left to the right, the first table column indicates the period
number and the second column indicates the move number. The following
columns mention who moves when and in what circumstances and what action
is played when somebody is called upon to move. Last column indicates the
payoffs vector in accordance with the strategies chosen by the players. As we
can see, the code form led us to an equilibrium strategy: Player 1 does not
approve the pay rise, Player 2 approves the pay rise and Player 3 approves
the pay rise. This is the game solution.
4.2 Matching Pennies
Let us consider a game with two players, player 1 and player 2. Both players
place simultaneously the coin on the table.
The strategies are:
• If they are both heads or both tails, player 1 takes 1.
• Otherwise player 2 takes 1.
The game is represented by the following table:
What changed in this representation compared to the previous game rep-
resentation? We can verify, in the simultaneous moves, that the options of the
players are at the same level. In fact, we can see the players acting at same
time. The remaining keeps in the same way.
Fig. 2. Matching pennies
4.3 Market Entry
Let us now imagine that we want to analyze games which simultaneously have
sequential moves and simultaneous moves. How to translate this situation into
code form? Knowing that code form is supposed to show the whole game in-
formation, and also to present to the reader the simultaneous moves.
Let us consider the following example:
Suppose that firm 1 has a monopoly in a given market. Firm 2 is considering
to enter this market. If firm 2 decides to enter, then both companies simulta-
neously decide whether to attack or to accept this situation.
Firstly, let us identify the strategies available to each firm:
• Firm 2 can choose to enter - E - or to stay out of the market - O.

• If it decides to enter then it can choose to attack - T - or to accept - A.
• On the other hand, firm 1 has the possibility, if firm 2 enters the market,
to attack - T or to accept - A.
This situation leads us to a game with imperfect information. The following
table illustrates the codified representation of the game where the first number
of the payoff belongs to the company that wants to penetrate the market and
the second number corresponds to the established company.
Fig. 3. Market Entry
Like the above example, we can verify, in the simultaneous moves, the
players acting at same time. The remaining keeps in the same way.
4.4 Sequential Bargaining
The above examples are about one round games. What happens to the code
form when we have a repeated game? How code form translates this situation?
Let us consider a game with two players, player 1 and player 2, which are
negotiating how they will split a given benefit, each one hoping to retain the
bigger share. Suppose that the offers and counter-offers take time, T periods,
and that each player applies a discount factor δ on future benefits, 0 < δ < 1
with . The next table represents this situation considering T = 3. Let be xi (t)
the sum that player i asks for at moment t.
Fig. 4. Sequential bargaining
Again the code form allows visualizing all the game information. For,
through the code form, to identify if is a repeated game, it is enough to
reorganize the information. All columns keep the meaning presented in the
previous examples.
5 Mixed Strategies
In the above model we defined En as the finite set of available strategies for
the player. The cartesian product of these sets was denoted by E . So a typical
element of this set is e = (e1 , e2 , ..., eN ) where each en is a pure strategy 3
for player n, that is, it is an element of En . We call e = (e1 , e2 , ..., eN ) a
pure strategy profile. When a player plays a pure strategy, his actions are
deterministic, that is he does not include uncertainty in his opponent. If he
wants to he must act in a random way and then he will have to evaluate
probabilistically what the other players will do. In other words, when a player
chooses one of his strategies, every other player might be uncertain about
which pure strategy player will choose.
3
Pure strategy in the sense of that when the player chooses one of these strategies
he does not act in a random way.
Definition 2. A mixed strategy of player n is a lottery over the pure strategies

of player n. One of player n s mixed strategies is denoted σn and the set of
all player n s mixed strategies is denoted Σn .
Thus σn = σn (en 1 ), σn (en 2 ), ..., σn (en kn ) where kn is the number of pure
n ) ≥ 0, i =
i
strategies of playern andσn (e 1, 2, .., kn and σn (en i ) = 1. The
cartesian product = 1 × 2 ×... × N is the set of all mixed strategy
profiles.
Formally, the mixed strategy set for each player is the probability distri-
bution set over his pure strategy set.
Definition 3. A n−dimensional simplex defined by the n+1 points x0 , x1 , ...

xn in IRp , p ≥ n, is denoted x0 , x1 , ..., xn and is defined to be the set {IRp |
n
n
x = θj xj , θj = 1, θj ≥ 0} . The simplex is nondegenerate if the n
j=0 j=0

n
vectors x1 −x0 , ..., xn −x0 are linearly independent. If x = θj xj the numbers
j=0
θ0 , θ1 , ..., θn are called the barycenter coordinates of x. The barycenter of the
simplex x0 , x1 , ..., xn is the point having barycenter coordinates θ0 = θ1 =
1
... = θn = .
n+1
Taking the above definition, since En is a finite set and kn is its cardinal,
we can identify Σn with (kn − 1) − dimensional simplex, ∆kn −1 4 . With
mixed strategies the payoff function extends to the mixed strategy profiles set
Σ = Σ1 × Σ2 × ... × ΣN in expected payoffs terms. To simplify we keep the
same notation for the function who represents effective payoffs function and
expected payoffs function, that is, un represents the expected payoff function
of player n as a function of the mixed strategy profile σ = (σ1 , σ2 , ..., σN ).
Thus we have:

un (σ) = ... σ(e1 )σ(e2 )...σ(eN )un (e1 , e2 , ..., eN )
e1 ∈E1 e2 ∈E2 en ∈En
6 Nash Equilibrium
6.1 Definition
As it was said above, the formulation and resolution of a game is very im-
portant to game theory. There are several game solution concepts. Unhappily
many of these concepts are restricted to a certain kind of games. The most
important solution concept was defined by John Nash (1950) and is known by
Nash equilibrium. As we will see the Nash equilibrium existence is guaranteed
for a large set of games.
4
kn − dimensional vectors non negative whose components sum equal to the unit.
Definition 4. A Nash equilibrium of a game is a profile of mixed strategies

σ = (σ1 , σ2 , ..., σN ) such that for each n = 1, 2, ..., N , for each en and en in
En , if σn (en ) > 0 then:

un (σ1 , σ2 , ..., σn−1 , en , σn+1 , ..., σN ≥ un (σ1 , σ2 , ..., σn−1 , en , σn+1 , ..., σN .
¿From the above definition we can see that an equilibrium is a profile of

mixed strategies such that a player knows what strategies the other players
will go to choose, and no player has incentive to deviate from the equilibrium
since that he cannot improve his payoff through an unilateral change of his
strategy. In fact, the Nash equilibrium induces a necessary condition of strate-
gical stability. however does not induce an enough condition for the analysis of
all the games. However, as Ratliff 5 argues ”Current attempts to satisfactorily
justify the application of Nash equilibrium to a wider class of games many
ultimately prove successful, in which the equilibria of these games will be rel-
evant. Furthermore, Nash equilibrium would be the relevant solution concept
if there were some pregame dynamic which assured that players’ beliefs were
in agreement prior to their strategy selection”.
Let’s go back to our two examples and calculate the equilibria.
As we can see, in ”voting for a pay rise” the Nash equilibrium leads us
to an equilibrium strategy: Player 1 does not approve the pay rise, Player 2
approves the pay rise and Player 3 approves the pay rise. This is the game
solution.
In ”matching pennies” we have that there is no equilibrium in pure strate-

gies. The solution is to randomize and play a mixed strategy. Suppose that
player 1 estimates that the player 2 will play H (head). Obviously player 1
prefer to play H. But then, from the definition of Nash equilibrium, player 2
has to set positive weight in his estimation of that player 1 would play only on
H. Thus, player 2 prefers to play T (tail), contradicting the supposition that
player 1 assessed him as playing H. In fact this belief supposing that player
1’s estimation of player 2 set weight strictly greater than a half on the fact
that player 2 would play H. Similarly we can argue by supposing that player
1’s estimation of player 2 set weight strictly less than a half on the fact that
player 2 would play H. Thus player 1’s estimation of player 2’s choices must
1 1
be ( , ). And similarly player 2’s assessment of player 1’s choices must also
2 2
1 1
be ( , ).
2 2
In ”market entry” we are led us to an equilibrium strategy:

((Enter, Accept), (Accept)).
5
Lecture notes from a PhD game-theory course at the University of Arizona during
1992-1997.
Note that we will only look at the Nash equilibrium, the game presents two
credible equilibria:
((Stayout, Attack), Attack)and((Enter, Accept), (Accept))
and a non-credible one:
((Stayout, Accept), Attack).
To the ”sequential bargaining” we can conclude:
• Round 3:The player 2 should say Yes iff δ 2 (1 − x1 (3)) ≥ 0, that is

x1 (3) ≤ 1. The player 1 should choose x1 (3) to maximize δ 2 x1 (3)
such that the player 2 says Yes, that is, choose x1 (3) = 1.
• Round 2: The player 1 should say Yes iff x2 (2) ≥ δx1 (3) = δ ·1 = δ,
that is x2 (2) ≥ δ. Player 2 offer x2 (2) to maximize δ(1 − x2 (2)
subject to player 1 saying Yes. To get the player 1 to say Yes, x2 (2)
must satisfy x2 (2) ≥ δ. But it should be clear that the player’s 2
maximizes δ(1 − x2 (2) by making x2 (2) as small as possible. Thus,
the solution is for player 2 to set x2 (2) = δ .
• Round 1: The Player 2 should say Yes iff 1−x1 (1) ≥ δ(1−x1 (1)) =
δ · (1 − δ) = δ − δ 2 , that is x1 (1) ≤ 1 − δ + δ 2 . The player 1
should offer x1 (1) to maximize x1 (1) subject to the player 2 saying
Yes. It should be clear that the solution is for the player 2 to set
x1 (1) = 1 − δ + δ 2 . Note that, if Player 1 demands a higher price,
the player 2 will say no, counter-offer x2 (2) = δ, which the Player
1 will do best to say Yes too, thus getting the smaller payoff of
δx2 (2) = δ 2 .
6.2 Existence
As we see, a particular selection of mixed strategy choices of the players

forms Nash equilibrium if and only if each player is using a best response to
the strategy choices of the other players, that is, is using a strategy which
maximizes the player’s payoff given the choices of all of the other players. But
in what circumstances is there Nash equilibrium? John Nash answered this
question in 1950-1951. Before starting to discuss the answer witch John Nash
gave we state for reference an important fixed point theorem.
Theorem 1. 6 Let M ⊂ IRn be a compact convex set. Let F →→ M an upper

hemi-continuous convex valued correspondence. Then the correspondence F
has a fixed point.
Proof:
First suppose the case that M is a nondegenerate simplex in IRn .
6
Kakutani 1941
Let M = x0 , x1 , ..., xn , that is, the simplex defined by the n + 1 points
x0 , x1 , ..., xn .
Now form the mth barycentric subdivision of M 7 . We define the continuous
function f m : M → M such that:
If x is the vertex of any cell of the subdivision let f m (x) for some
y inf(x). For any other x we define f m (x) = y by extending the func-
tion in a linear manner inside each cell. That is, if x is in the cell

n
n
x0 , x1 , ..., xn and x = θj xj with x = θj = 1, θj ≥ 0 then
j=0 j=0

n
f m (x) is defined to be x = θj f m (xj ).
j=0
Applying Brouwer’s Theorem 8 to obtain a fixed point of the map f m , say

m
x . If is a vertex of one of the cells in the subdivision then we are done since
xm = f m (xm ) ∈ F (xm ). If xm is not a vertex of one of the cells then let the
cell in which it does lie be x0 m , x1 m , ..., xn m , and let θ0 m , θ1 m , ..., θn m be
the barycentric coordinates of xm relative to that cell. Thus:

n
xm = θ j m xj m ;
j=0
and

n
xm = f m (xm ) = θj m yj m (1)
j=0
where
yj m = f m (xj m ) ∈ F (xj m for all j = 0, 1, ..., n).
Now choose a subsequence of m → ∞ say mk → ∞ such that:
xmk → x∗ , θmk → θj , and yj mk →= yj .
Also, since the cells shrink to points as mk → ∞ each of the vertices of
the cell containing xmk also converges to x∗ , i.e., xmk → x∗ .
Thus from ( 1)
7
A 0 − dimensional simplex is itself its subdivided simplex. A i − dimensional
simplex x0 , x1 is subdivided into two simplexes of the same dimension x0 , x1
and x0 , x1 where y is the barycentre of x0 , x1 .
8
Brouwer’s Theorem (1910): Let f : B n → B n be a continuous function from the
n − ball to itself. Then there is some x∗ ∈ B n such that x∗ = f (x∗ ), i.e., there is
fixed point.

n
x∗ = θj yj .
j=0
Also since F is upper hemi-continuous yj ∈ F (x∗ ).

Since F (x∗ )is convex and x∗ is a convex combination of the yj s this implies
∈ ∗
that x ∗ F (x )as required.
Now suppose M is not a simplex. We take some simplex M containing M
and a retraction ϕ : M → M 9 .

Then F : M →→ M defined by F (x) = F (ϕ(x)) is clearly an upper

hemi-continuous correspondence and a fixed point of F clearly lies in M and
so is also a fixed point of F .
Theorem 2. 10 The mixed extension of every finite game has, at least, one
strategic equilibrium.
Proof 11 :
Consider the set-valued mapping (or correspondence) that maps each strategy
profile, x, to all strategy profiles in which each player’s component strategy
is a best response to x(that is, maximizes the player’s payoff given that the
others are adopting their components of x). If a strategy profile is contained
in the set to which it is mapped (is a fixed point) then it is an equilibrium.
This is so because a strategic equilibrium is, in effect, defined as a profile that
is a best response to itself.
Thus the proof of existence of equilibrium amounts to a demonstration that
the best response correspondence has a fixed point. The fixed-point theorem
of Kakutani asserts the existence of a fixed point for every correspondence
from a convex and compact subset of Euclidean space into itself, provided
two conditions hold. One, the image of every point must be convex. And two,
the graph of the correspondence (the set of pairs (x, y) where is in the image
of x) must be closed.
Now, in the mixed extension of a finite game, the strategy set of each player
consists of all vectors (with as many components as there are pure strategies)
of non-negative numbers that sum to 1; that is, it is a simplex. Thus the
set of all strategy profiles is a product of simplexes. In particular, it is a
convex and compact subset of Euclidean space. Given a particular choice of
strategies by the other players, a player’s best responses consist of all (mixed)
strategies that put positive weight only on those pure strategies that yield
the highest expected payoff among all the pure strategies. Thus the set of
best responses is a sub simplex. In particular, it is convex. Finally, note that
the conditions that must be met for a given strategy to be a best response
to a given profile are all weak polynomial inequalities, so the graph of the
9
Continuous function taking M to M that leaves all points of M fixed.
10
Nash 1950, 1951
11
This proof was given by John Hillas for the intensive Course on the Game Theory
of Strategic Equilibrium, given at SUNY at Sony Brook in the summer of 1990.
best response correspondence is closed. Thus all the conditions of Kakutani’s

theorem hold, and this completes the proof of Nash’s theorem.
7 A Process to Find the Game Solution

As we see, in the above examples, there are many games where players’ moves
are sequential, that is players move in accordance with a given order. These
games have a great applicability in economics, business and political science.
We present an algorithm that helps us to solve these games in the simplest
manner. The process consists of considering the rationality of the players the
way the game should be played and to find its equilibrium. This process has
a key idea which is the fact that if a player can foresee what the other players
will do, he is in fact in a position to determine the strategy that will allow
him to get the best payoff, and have the ability to evaluate the credibility of
his opponents’ strategies.
In this way we can see how selecting the optimum payoff, at each respective
sets of payoffs of the deciding player we will easily be able to find the game
equilibrium. This selection is made in an order contrary to the development of
the game, that is, the process starts off by selecting the payoffs of the player
who has the last word and ends with the selection of the payoffs of whom
initially plays.
7.1 Best Payoff Method
Algorithm:
1. Let N be the player who takes the last decision in the game.
Consider any move that comes just before terminal move. Select
the payoff sets associated with each move. In each of those sets
select the player N s optimum payoff;
2. Consider the subgame that is obtained by excluding the column
which refers to player N ;
3. Let N be the player which takes the decision immediately prior
to player N;
4. Let N = N . Repeat the process from step 1 until the player who
makes the first decision is found;
5. Select the strategy profile induced by the payoff vector obtained.
8 Conclusion
We presented a new game model which provides a global view of the game
more clear and provides more information about the game than the other game
representations. Considering that the importance of a representation rests on
its ease of interpretation, ease of solving a game and truthful information

about the game, code form representation is the only one, which fulfils all
these requirements. We also observed that a Nash equilibrium exists if and
only if there exists a fixed point of a particular best-response correspondence.
Kakutani’s fixed-point theorem guarantees the existence of this fixed point
when the correspondence satisfies certain conditions. We verify that the best-
response correspondence satisfies the Kakutani conditions and hence prove
the existence of a Nash equilibrium. Looking at code form we can see that
game in code form satisfies the Nash Theorem.
References
1. Aubin, J. P.; ”Applied Functional Analysis”; John Wiley & Sons, Inc; New
York, 1979
2. Bicchieri, Cristina; Jeffrey, Richard, Skyrms, Brian, ”The Logic of Strategy”.
Oxford University Press, Inc.. 1999
3. Davis, Morton D., ”Introducción a la Teora de Juegos”, Spanish translation by
José Carlos Gómez Borrero. Ciencia e Tecnologı́a, Alianza Editorial. Madrid.
1986
4. Brezis, H.; ”Analyse Fonctionelle”; Mac Millan Publishing Co. Inc.; New York,
1979
5. Dresher, Melvin; ”Jeux de Stratgie- Théorie et Applications””; Springer-Verlag,
Dunod, Paris; 1965
6. Eberwein, Curtis, ”The Sensitivity of Repeated Bargaining Outcomes to the
Choice of First Proposer”. June. 2000
7. Fudenberg, Drew; Tirole, Jean, ”Game Theory”. Cambridge. Mass: Mit. Press.
1991
8. Gardner, Roy, ”Juegos para Empresarios e Economistas”. Spanish translation
by Paloma Calvo e Xavier Vila. Antoni Bosch, Editor, S.A. Barcelona. March.
1996
9. Gibbons, Robert, ”Game Theory for Applied Economist”. Princeton University
Press. 1992
10. Gibbons, Robert, ”Um Primer Curso de Teora de Juegos”. Spanish transla-
tion by Paloma Calvo and Xavier Vila. Antoni Bosch, Editor, S.A. Barcelona.
October. 1993
11. Kakutani, S: ”A Generalization of Brouwer’s Fixed Point Theorem”, Duke
Mathematics Journal., 8, 1941
12. Kennan, John; Wilson, Robert, ”Bargaining with Private Information”. Journal
of Economic Literature 31. 1993
13. Matos, M.C.; Ferreira, M.A.M., ”Games in Code Form”. Presented at the Fifth
Spanish Meeting on Game Theory. Seville. 1-3 July. 2002
14. Matos, M. C.; Ferreira, M. A. M., ”Games in Code Form versus games in Exten-
sive Form”. Presented at Quantitative Methods in Economics (Multiple Criteria
Decision Making XI). The Slovak Society for Operations Research. Faculty of
Economics and Management /Slovak Agricultural University in Nitra). Nitra
Slovakia. 5-7 December. 2002
15. Matos, M. C.; Ferreira, M.A.M., ”Jogos na Forma Codificada”. Temas em

Métodos Quantitativos 3. Editores: Elizabeth Reis e Maria Manuela Hill.
ISCTE. Edições Sı́labo. Lisbon. 2003
16. Matos, M. C.; Ferreira, M. A. M., ”Best Payoff Method”. Presented at XV
IMGTA. Italian Meeting on Game Theory and Applications. Urbino. Italy. 9-
12 July. 2003
17. Matos, M. C.; Ferreira, M. A. M., ”Impaciência nos Jogos de Negociação”.
Revista Portuguesa e Brasileira de Gestão. N2/2003. Volume VIII. September
2003. FGV. EBAPE. INDEG. ISCTE
18. Morris, Peter; ”Introduction to Game Theory” Springer-Verlag, New York Inc.;
1977
19. Nash. J, ”Non-Cooperative Games”. Annals of Mathematics, 54, 1951
20. Neumann, J. von; Morgenstern, O., ”Theory of Games and Economic Be-
haviour”. John Wiley & Sons, Inc. New York. 1967
21. Osborne, Martin J., ”An introduction to game theory”. Oxford University
Press. 2000
Effect of Mutual Choice Metanorm in
Group Dynamics for Solving Social Dilemmas
Tomohisa Yamashita, Kiyoshi Izumi, Koichi Kurumatani
Investigation of Improvement of Traffic Efficiency by Route Information Sharing

Information Technology Research Institute
National Institute of Advanced Industrial Science and Technology
2-41-6, Aomi, Koto-ku Tokyo 135-0064, Japan
tomohisa.yamashita@aist.go.jp, kiyoshi@ni.aist.go.jp,
k.kurumatani@aist.go.jp
Summary. In this research, our aim is the increase of both each individual utility
and social welfare in road transportation systems by decreasing traffic congestion.
For the attainment of our purpose, we propose a simple route choice mechanism
based on the concept of mass user support. Through multiagent simulation, we
analyze the effect of our proposed mechanism on the improvement of traffic efficiency
of individual drivers and the whole systems.
1 Introduction
Recently, ubiquitous computing environment in road transportation systems

has been developed rapidly; car navigation systems have been spread widely,
accuracy of GPS (Global Positioning System) are advanced increasingly, a
Vehicle Information and Communication System (VICS) started and then the
range which VICS can provide congestion information has been extended [11],
in-vehicle communication devices, sensors, processors have been developed.
In road transportation systems, one of most important information service
is navigation from origin to destination. Many researchers have been trying
to develop navigation systems based various kinds of traffic information [3, 7].
The purpose of navigation systems is to maximize the efficiency of individual
users by providing the routes that satisfying the intension of user, e.g., the
route with shortest travel distance or travel time. In previous research, it has
been revealed optimization of the performance of road transportation systems
is difficult to improve by providing current congestion information [6, 9, 10].
The reason is that congestion is often caused by route choice behaviors for
congestion avoidance. Based on the current state of congestion, navigation
system provides the shortest time route that is not congested currently. If
many drivers using such systems also choose one route, they concentrate one
336 Tomohisa Yamashita, Kiyoshi Izumi, Koichi Kurumatani
route simultaneously. As a result, traffic congestion is caused in this route

where navigation systems provided as the shortest route in distance. To make
matters worse, traffic congestion sometimes becomes heavier in spite of the
increase of drivers using traffic information. Based on current state that car
navigation systems have been diffused rapidly, this is a serious problem. Fur-
thermore, because this kind of behavior is observed not only in road trans-
portation systems but also in large scale theme park and event hall [2, 8],
general solutions for congestion have been expected.
Here, we introduce the concept of mass user support [4, 5] in order to avoid
a cause of unintentional congestion. Mass user support provides information
services to users, or groups, or society as mass beyond conventional personal
services. Although conventional service considers only utilities of individual
users, mass user support considers not only the summation of personal ser-
vices to individual users but also the interaction among users. The purpose of
mass user support is increase of both each individual utility but totally social
welfare, and definitely not the increase of totally social welfare at the sacrifice
of utilities of certain users. That is to say, mass user support requires some
kind of social coordination, which leads mutual concession among users [4] .
Here, the best solution in whole road transportation systems isn’t necessarily
required. The reason is that, if the performance of road transportation sys-
tems can be improved slightly by mass user support, the profit on the society
is immeasurable because of large-scale of road transportation systems. As a
result of improvement, there are following merits of the improvement of the
performance; i) reduction of directly economic loss, and ii) decrease of the
harmful effects of pollutant included in exhaust gas of vehicles, iii) provision
of more comfortable driving environment to drivers.
In this research we challenge to construct a navigation system that im-
proves the traffic efficiency of individual drivers and the whole systems. For
this purpose, based on mass user support, we propose a simple route choice
mechanism that makes a mutual concession among users autonomously. By
our proposed mechanism, it is expected that the increase of both each indi-
vidual utility and totally social welfare can be realize. With multiagent sim-
ulation, we examine the effect of our proposed mechanism on the following
two points. i) As the drivers using our proposed mechanism increases, the effi-
ciency of them, e.g., travel time from an origin to a destination, becomes better
monotonously. ii) The efficiency of the drivers using our proposed mechanism
is always better than the drivers using other kind of mechanisms.
2 Multiagent Modeling
2.1 Traffic Model
In this paper, in order to examine interdependency between traffic congestion
and route choice behaviors in road transportation systems, we construct traf-
fic flow model as simple as possible. Therefore, we don’t consider following
Fig. 1. Direction of movement of vehicles and revision of block
factors; traffic signal (i.e., stop at a red light), waiting for right turn of on-
coming cars in intersections, right turn lane, multi-lane, overtaking, a blind
alley, U-turn on lane except intersections.
In our traffic flow model, the road between intersections is called a link,
which is divided into several blocks. The length of a block is equal to the
distance that a vehicle runs at free flow speed of the link during one simulation
step. And then, an order is assign to each block from downstream to upstream.
Concerning the block assigned i-th, we define Ki as the density of block i, Li
as the length of block i, Ni as the number of the vehicles in block i, and
Vi as the velocity of vehicle in block i. Here, Ki is the division of Ni by
Li . In block i, Vi is revised based on Greenshield’s V-K relationship.This
relationship represents that speed and density on block i are linearly related,
and is described as
Vi = b − aKi , (1)
where a and b are coefficients determined from field observations. In this
simulation, we set these coefficients as a = 99.20 and b = 13.89.
The process of the flow calculation between the two neighboring blocks
i and i + 1 is as follows. At every step, the speed of vehicles in each block
is revised according to the V-K relationship, and then vehicles move forward
based the speed. The move of vehicles is processed from downstream to up-
stream as shown in Figure 1. Based on Vi , vehicle j can move forward. When
vehicle j moves into block i from block i + 1, its speed changes to Vi from
Vi+1 . If the density of block i is over jam density Kjam , no vehicles can move
into block i from block i + 1. After the process of the move of vehicle j1 in
front of vehicle j2 , if vehicle j1 is within the distance that vehicle j2 can move
forward at the speed of vehicles Vi , vehicle j2 approaches vehicle j1 to the
minimum distance between two cars. Although vehicle j2 has enough speed
to move forward more, it has to stay at the back of vehicle j1 . At the next
step in block i, when Vi is revised based on Ki , vehicles can accelerate or
slowdown to Vi immediately regardless of the speed in the last step.
2.2 Route Choice Behavior
We prepare the following three types of drivers in a route choice behavior.

First and second types are well-known and easy-to-understand because they
simply minimize travel distance or travel time. Third type is our proposed
route choice mechanism based on the concept of mass user support.
Shortest Distance Route
The drivers using the Shortest Distance Route (SD) decide a route based
on the knowledge of the map without information of traffic congestion. The
drivers using the SD a route that has the shortest distance from its origin to
destination. They use the only distance of a link and search the shortest route
in distance and don’t consider traffic congestion at all.
Shortest Time Route
The drivers using the Shortest Time Route (ST) decide a route based on
the knowledge of the map with information of current traffic congestion. The
drivers of this type represent the dynamic route choice based on not only
map information but also current congestion information of entire network at
anytime from traffic information center (e.g., VICS Center) through vehicle
equipments. They search the route that has the shortest travel time from
their origins to destinations, and revise their route whenever they face an
intersection based on current congestion information.
The driver of using the ST is assumed to receive current density of all
blocks through vehicle equipments from traffic information center. They con-
siders the expected travel time of a link, which is calculated based on the
current density of each block as follows. At first, the speed on block i is calcu-
lated based on the V-K relationship with density Ki . And then, the passage
time of block i is calculated based on length Li of block i and speed Vi on
block i. Finally, the passage time of a link is calculated by summing the pas-
sage time of all block on the link. We define expected travel time ET Tl as the
summation of the passage time of all blocks on link l. The driver using the
ST searches the shortest route in the expected travel time from its current
position to destination.
Shortest Time Route with Route Information Sharing
The drivers using the Shortest Time Route with Route Information Sharing
(ST-RIS) decide a route based on the knowledge of the map, information of
current traffic congestion, and information of the routes of other drivers using
ST-RIS. The drivers using the ST-RIS also search the route and revise their
route whenever they face the intersections. The difference of the ST is that the
drivers using the ST-RIS share their routes from thier origin to destination.
Fig. 2. Example of calculation of total passage assurance: Driver 1 has the route
through six links 6, 5, 4, 3, 2, 1 from current position on link 6 to its destination on
link 1. Driver 2 has the route through three links 4, 3, 2 from current position on
link 4to its destination on link 2.
This sharing can be realized easily with communication devices, e.g., cellular
phone, through route information server.
At first, the driver using the ST-RIS searches the shortest route from its
origin to destination in the expected travel time. And then, the drivers notify
their routes to route information server. Route information server collects the
routes of all drivers using the ST-RIS, and assigns the passage assurance of
a driver to a link based on their routes, current positions, and destinations
as follows. The passage assurance means the degree of the assurance that a
driver will passes through a link in the future. We define passage assurance
P Al,j of driver j to link l as follows. If a route passes through p links from
its current position to destination, route information server assigns a order
to each link from a destination to current position of the route. And then,
route information server divides the order by p and regard it as a weight of a
link. For example, 1/p is assigned a link including a destination, 1 (=p/p) is
assigned a link including a current position.
Furthermore, we define total passage assurance T P Al as the summation
of the passage assurance of all drivers to link l. Route information server
calculates T P Al as follows,
k
T P Al = k∈ST −RIS P Al,k , (2)
where ST -RIS is the set of the drivers using the ST-RIS. Figure 2 shows an
example of calculating total passage assurance of links.
Route information server provides the total passage assurance of all links
for the drivers using the ST-RIS. Finally, we define prospective traffic volume
P T Vl of link l as the product of ET Tl and (T P Al + 1.0). The driver using
the ST-RIS searches the shortest route in the prospective traffic volume from
its current position to destination.
Fig. 3. Simple road network with four routes
Fig. 4. Radial and ring network
3 Computer Simulation
3.1 Simulation Settings
In order to evaluate effects of our proposed route choice mechanism in var-

ious conditions, i.e., in the different components of route choice behaviors,
we change the share of three route choice types, the Shortest Distance Route
(SD), the Shortest Time Route (ST), and the Shortest Time Route with Route
Information Sharing (ST-RIS). Concretely, we prepare three cases of the share
of three types. In Case 1 and 2, we treat simple situation as follows. In Case
1, we omit the ST-RIS, and change the share of the SD and the ST. On the
other hand, in Case 2, we omit the ST, and change the share of the SD and the
ST-RIS. In the future, traffic information is easier to access for many drivers
because navigation systems have spread widely. Therefore, in Case 3, we fixed
the share of the SD to 20%, and change the share of the ST and the ST-RIS.
Furthermore, in order to evaluate route choice behaviors in various con-
figurations of road networks, we use two kinds of the configurations of road
networks. One network configuration is a simple road network that all drivers
choose one route out of four routes to reach destination. These configurations
are shown in Figure 3 and Table 1. In the case of this simple road network,
Table 1. Conditions of the simple network and the radial and ring network.
simple network radial and ring
Number of links 16 40
Number of nodes 14 24
Number of vehicles 1200 1650
Total length (km) 125.9 173.1
links 1 and 3 have the same length, and links 2 and 4 also have the same
length. The length of links 1and 3 is slightly longer than that of links 2 and
4. An origin of a driver is assigned randomly to any block on link A or B in
Fig. 3. If an origin of a driver is on link A, a destination of it is assigned to
any block on link B. Otherwise, a destination of it is assigned on link A. All
links has the same capacity except links in links 1 to 4. In the case of four
routes, the capacity of links 1, 2, 3, 4 is 1/4 of other links. The other is a
radial and ring network shown in Figure 4 and Table 1. In this road network,
all links have the same capacity. An origin of a driver is assigned randomly
to any block on link A to H in Fig. 4. And then, its destination is assigned to
any block on a symmetric link. For example, if an origin is assigned on link
B, its destination is assigned on link F.
In order to make the vehicle density in the simple road network and the
radial and ring network all the same value, we arrange the number of vehicles
in the network based on each total length. At every simulation step, twenty
vehicles are generated randomly on links A or B in the case of the simple road
network, and on links A to H in the case of the radial and ring networks. After
reaching its destination, a driver returns to its origin. And then, a driver goes
to its destination again. A driver repeats to go to its destination and return
to its origin throughout simulation.
In order to compare our simulation results under the different shares of
route choice type in the different road networks, we define ideal travel distance,
ideal travel time, and travel time efficiency. The ideal travel distance of a driver
is defined as the distance from an origin to destination when a driver passes
through the shortest route in distance. The ideal travel time is defined as
the time required from an origin to destination when a driver passes through
the shortest route in distance at free flow speed. The travel time efficiency is
defined as the ratio of the time required actually from an origin to destination
and the ideal travel time.
3.2 Simulation Results
As a result of simulations under the conditions about the share of drivers’

types and the configurations of road networks, especially we focus on the
average travel time efficiency of each type and all drivers. The results of our
simulation are shown in Figure 5 to 10. The travel time efficiency is calculated
Fig. 5. Average travel time efficiency of Case 1 in simple road network with four
routes
routes
based on the average of 10 trials. In these graphs, the horizontal axis is the
share of the ST or ST-RIS, and the vertical axis is the travel time efficiency.
Figure 5, 6 and 7 show the result in the simple network with four routes.
Fig. 5 shows the average travel time efficiency in Case 1. The ST continued
rising monotonously from beginning to end. Until 50% of the share of the ST,
the SD and all drivers fell steadily, and then increased slightly. After 50%, all
of three are almost the same. Fig. 6 shows the average travel time efficiency
in Case 2. Until 30% of the share of the ST-RIS, the ST-RIS went up slightly.
After that, the ST-RIS went down slightly. Until 50% of the share of the ST,
the SD and all drivers fell steadily from beginning to end. Fig. 7 shows the
routes
average travel time efficiency in Case 3. Up to 60% of the share of the ST-RIS,
the SD, ST, ST-RIS, and all drivers decreased steadily. After 60%, however
the average of the SD went up slightly, the rest remained almost constant.
Furthermore, the ST-RIS is worse than others until 60%, and after that, the
ST-RIS is better than others. Throughout simulation, their values were very
close.
Figure 8, 9 and 10 show the result in the radial and ring network. In the
graphs in Fig. 8 and 9, there are the parts of dashed line. This means that
the travel time efficiency of these parts was not a valid value for comparing
with the other parts. In the conditions that 0 or 10% of all drivers used the
ST and the rest used the SD, 80% of all vehicles were caught in a deadlock.
In our traffic simulation, the deadlock is defined as follows; vehicles in the top
of link could not move into next links because these links are filled with other
vehicles completely. Furthermore, the vehicles in the top of such a link also
could not move into next links on the same reason.
Fig. 8 shows the average travel time efficiency in Case 1. After 20% of the
share of the ST, the ST increased slightly and steadily. At first, the SD and
all drivers dropped dramatically. After 80%, all of them declined marginally.
Fig. 9 shows the average travel time efficiency in Case 2. Although the shapes
of the SD and all drivers were similar to these in Fig. 8, the ST-RIS decreased
slightly and steadily. Fig. 10 shows the average travel time efficiency in Case
3. From beginning to end, all of them continued decreasing steadily. Until
50% of the share of the ST-RIS, all are ranked in descending order as the SD,
ST-RIS, all drivers, and ST. After that, the rank was changed to the ST-RIS,
all drivers, SD, and ST.
Fig. 8. Average travel time efficiency of Case 1 in the radial and ring network
Fig. 9. Average travel time efficiency of Case 2 iin the radial and ring network
4 Discussion
In this section, we discuss the effect of our proposed mechanism to increase

both each individual utility and social welfare.
At first, concerning the results in the simple network with four routes,
there was the improvement of traffic congestion with the increase of the ST-
RIS. In Fig. 5, as the ST increased, the travel time efficiency of the ST became
worse. The reason is that the drivers using the ST often concentrated simulta-
neously one route that traffic information system told that there was no traffic
congestion . In Fig. 6, as the ST-RIS increased, the travel time efficiency of
the ST-RIS and all drivers became better. In Fig. 7, as the share of the ST-
RIS rose, the travel time efficiency of all types was improved monotonously.
Fig. 10. Average travel time efficiency of Case 3 in the radial and ring network
The drivers using the ST-RIS were able to avoid concentrating simultaneously
one certain route where there is no traffic congestion currently. Our proposed
mechanism realized a mutual concession among users autonomously. Further-
more, although the efficiency of the ST-RIS was not always better than the
drivers of the SD and ST, there were little differences between the ST-RIS
and the other types. Therefore, the ST-RIS was roughly satisfied with the two
features that we require. In the simple network with four routes, our proposed
mechanism works efficiently in the improvement of the travel time efficiency.
Next, concerning the results in the radial and ring network, traffic conges-
tion was improved as the ST-RIS increased. In Fig. 8, as the ST increases,
the travel time efficiency of the ST becomes worse. On the other hand, in
Fig. 9, as the ST-RIS increased, that of the ST-RIS became better In Fig.
10, as the share of the ST-RIS rose, the travel time efficiency of all types was
improved monotonously. Similarly to the simple network with four routes, the
drivers using the ST-RIS were able to avoid concentrating simultaneously one
certain route. The difference is that the travel time efficiency of the ST is
always better than that of the ST-RIS, and that of SD was better than that
of the ST-RIS after 50% share of the ST-RIS. These relationships between
the travel time efficiency of the ST and ST-RIS, i.e., i) as the ST-RIS in-
creased, the travel time efficiency of the ST and the ST-RIS was improved
monotonously, and ii) the travel time efficiency of the ST was always better
than that of the ST-RIS, are satisfied the conditions of the social dilemmas
[1].
According to Dawes the social dilemmas are caused in the situation where
individuals may contribute to a group voluntarily but all members share the
benefits. Based on game theoretical analysis, it is evident that the decision to
do nothing is always the best strategy regardless of what the rest does. The
reason is that the individuals can acquire more benefit whether it contributes
or not. However, all individuals would acquire more if all contribute.
In our traffic model, it can be considered that cooperative behavior is the
ST-RIS, and defective behavior is the SD (and the ST) because the travel
time efficiency of all drivers increases as the ST-RIS increases, and the drivers
using the ST always acquire better efficiency. In the radial and ring network,
a mutual concession among the ST-RIS, i.e., avoiding congested routes and
taking the longer route each other, improves traffic congestion on the whole
system. The ST can pass through the shorter route without heavy congestion
in distance than the ST-RIS and can improve only the travel time efficiency of
only them at the expense of a mutual concession among the ST-RIS. Although
the ST-RIS improves the travel time efficiency of itself and all drivers, the ST
and SD free-ride the contribution of the ST-RIS, and then their travel time
efficiency is improved more significantly than the ST-RIS.
Concerning the construction of the effective mechanism based on route in-
formation sharing, there is room of improvement of our proposed mechanism.
Currently, our proposed mechanism is very simple because information server
receives route from individual drivers and return the accumulative information
to them regardless of current traffic congestion, the ratio of drivers using our
proposed mechanism, and the intention and status of drivers. If our mechanism
can adapt dynamic traffic status, it will be able to improve the efficiency of
the driver using it and all drivers. However, we observed that the route choice
behaviors face the problem of the social dilemmas. The drivers using other
types like the SD and ST free-ride the mechanism based on mass user support
making a mutual concession. It is not difficult relatively to develop the mech-
anisms that can continue improving the travel time efficiency monotonously
with the increase of their share. However, it is difficult significantly to de-
velop the mechanisms that can realize the better travel time efficiency than
other types. In order to solve the social dilemmas in route choice behaviors,
we need to construct new types of route navigation mechanisms against the
social dilemmas. Because the social dilemmas have been taken up in various
research fields earnestly, we apply the knowledge to our mechanism and then
will be able to acquire new solutions suited to the features of traffic flow.
5 Conclusion
In this paper, in order to decrease traffic congestion, we proposed the naviga-
tion mechanism with route information sharing based on mass user support.
With multiagent modeling, we construct a simple traffic flow model based on
the V-K relationships. As the route choice behavior, three types, the Shortest
Distance Route (SD), the Shortest Time Route (ST), and the Shortest Time
Route with Route Information Sharing (ST-RIS), are prepared. For examine
the effect of this mechanism in various conditions, we used three kinds of the
shares of the SD, ST, and ST-RIS, and two kinds of road networks, one is
the simple networks with four routes from an origin to a destination, and the
other is the radial and ring network. As a result of simulations, we conformed
that i) in the simple network with four routes, the ST-RIS works efficiently,
ii) in the radial and ring network, the ST-RIS can make a mutual concession,
and can improve the efficiency of the ST-RIS and all drivers. However, the
social dilemma situation is caused, and then the ST and SD free-ride the ST-
RIS. Therefore, it is difficult for the ST-RIS to realize the best travel time
efficiency. Based on such results, we pointed that it is necessary to modify our
proposed mechanism and construct new types of mechanisms to conquer the
social dilemmas in route choice behaviors.
References
1. Dawes, R. M. 1981. Social Dilemmas. Annual Review of Psychology 31: 169-193.
2. Kawamura, H.; Kurumatani, K.; and A. Ohuchi. 2003. Modeling of Theme
Park Problem with Multiagent for Mass User Support, In Working Note of
The International Joint Conference of Artificial Intelligence 2003, Workshop on
Multiagent for Mass User Support, 1-7.
3. Klugl, F.; Bazzan, A.L.C.; and Wahle, J. 2003. Selection of information types
based on personal utility: a testbed for traffic information markets. In Proceed-
ings of the second International Joint Conference on Autonomous Agents and
Multiagent systems, 377- 384.
4. Kurumatani, K. 2003. Mass User Support by Social Coordination among Users.
In Proceedings of the International Joint Conference of Artificial Intelligence
2003, Workshop on Multiagent for Mass User Support MAMUS-03, 58-59.
5. Kurumatani, K. 2003. Social Coordination with Architecture for Ubiquitous
Agents: CONSORTS. In Proceedings of International Conference on Intelligent
Agents, Web Technologies and Internet Commerce 2003 (CD-ROM).
6. Mahmassani, H. S.; Jayakrishnan, R. 1991. System Performance and User Re-
sponse under Real-Time Information in a Congested Traffic Corridor. Trans-
portation Research 25A(5): 293-307.
7. Shiose, T.; Onitsuka, T.; and Taura, T. 2001. Effective Information Provision
for Relieving Traffic Congestion. In Proceedings of 4th International Conference
on Intelligence and Multimedia Applications,138-142.
8. Suzuki, R.; Arita T. 2003. Effects of Information Sharing on Collective Be-
haviors in Competitive Populations. In Proceedings of the Eight International
Symposium on Artificial Life and Robotics, 36-39.
9. Tanahashi, I.; Kitaoka, H.; Baba, M.; H. Mori, H.; Terada, S.; and Teramoto,
E. 2002. NETSTREAM, a Traffic Simulator for Large-scale Road Networks, R
& D Review of Toyota CRDL, 37(2):47-53 (in Japanese).
10. Yoshii, T.; Akahane, H.; and Kuwahara, M. 1996. Impacts of the Accuracy of
Traffic Information in Dynamic Route Guidance Systems. In Proceedings of
The 3rd Annual World Congress on Intelligent Transport Systems (CD-ROM).
11. http://www.vics.or.jp
/HFWXUH1RWHVLQ(FRQRPLFV
DQG0DWKHPDWLFDO6\VWHPV
)RULQIRUPDWLRQDERXW9ROV²
SOHDVHFRQWDFW\RXUERRNVHOOHURU6SULQJHU9HUODJ
9RO 5 'HPHO )LVFDO 3ROLF\ 3XEOLF 'HEW DQG WKH 9RO 0 5HXWWHU$ 0DFURHFRQRPLF 0RGHO RI :HVW
7HUP6WUXFWXUHRI,QWHUHVW5DWHV;SDJHV *HUPDQ8QHPSOR\PHQW;SDJHV
9RO 0 7KpUD 5 7LFKDWVFKNH (GV ,OOSRVHG 9RO$&DVDMXV)RFDO3RLQWVLQ)UDPHG*DPHV;,
9DULDWLRQDO3UREOHPVDQG5HJXODUL]DWLRQ7HFKQLTXHV9,,, SDJHV
SDJHV
9RO ) 1DUGLQL 7HFKQLFDO 3URJUHVV DQG (FRQRPLF
9RO6+DUWPDQQ3URMHFW6FKHGXOLQJXQGHU/LPLWHG *URZWK;9,,SDJHV
5HVRXUFHV;,,SDJHV
9RO 0 )OHLVFKPDQQ 4XDQWLWDWLYH 0RGHOV IRU
9RO / Y 7KDGGHQ 0RQH\ ,QÁDWLRQ DQG &DSLWDO 5HYHUVH/RJLVWLFV;,SDJHV
)RUPDWLRQ,;SDJHV
9RO1+DGMLVDYYDV-(0DUWtQH]/HJD]-33HQRW
9RO 0 *UD]LD 6SHUDQ]D 3 6WlKO\ (GV 1HZ (GV *HQHUDOL]HG &RQYH[LW\ DQG *HQHUDOL]HG 0RQR
7UHQGVLQ'LVWULEXWLRQ/RJLVWLFV;SDJHV WRQLFLW\,;SDJHV
9RO9+1JX\HQ--6WURGLRW37RVVLQJV(GV 9RO$.LUPDQ-%=LPPHUPDQQ(GV(FRQRP
2SWLPDWLRQ,;SDJHV LFVZLWK+HWHURJHQRXV,QWHUDFWLQJ$JHQWV9,,SDJHV
9RO : % =KDQJ$7KHRU\ RI ,QWHUQDWLRQDO7UDGH
;,SDJHV
9RO 3< 0RL[ (G7KH 0HDVXUHPHQW RI 0DUNHW
9RO 0 .|QLJVWHLQ (TXLW\ (IÀFLHQF\ DQG (YROX 5LVN;,SDJHV
WLRQDU\6WDELOLW\LQ%DUJDLQLQJ*DPHVZLWK-RLQW3URGXF
9RO 6 9R - 5 'DGXQD (GV &RPSXWHU$LGHG
WLRQ;,,SDJHV
6FKHGXOLQJRI3XEOLF7UDQVSRUW;,SDJHV
9RO''*DWWL0*DOOHJDWL$.LUPDQ,QWHUDFWLRQ
DQG0DUNHW6WUXFWXUH9,SDJHV 9RO % 3 .HOOHUKDOV )LQDQFLDO 3ULFLQJ 0RGHOV LQ
&RQWLQXRXV7LPHDQG.DOPDQ)LOWHULQJ;,9SDJHV
9RO$*DUQDHY6HDUFK*DPHVDQG2WKHU$SSOLFD
WLRQVRI*DPH7KHRU\9,,,SDJHV
9RO0.RNVDODQ6=LRQWV0XOWLSOH&ULWHULD'HFL
9RO01HXJDUW1RQOLQHDU/DERU0DUNHW'\QDPLFV VLRQ0DNLQJLQWKH1HZ0LOOHQLXP;,,SDJHV
;SDJHV
9RO . 1HXPDQQ & 6FKZLQGW - =LPPHUPDQQ
9RO<<+DLPHV5(6WHXHU(GV5HVHDUFKDQG 3URMHFW 6FKHGXOLQJ ZLWK 7LPH :LQGRZV DQG 6FDUFH
3UDFWLFH LQ 0XOWLSOH &ULWHULD 'HFLVLRQ 0DNLQJ ;9,, 5HVRXUFHV;,SDJHV
SDJHV
9RO ' +RUQXQJ ,QYHVWPHQW 5 ' DQG /RQJ5XQ
9RO % 6FKPROFN 2PPLWWHG 9DULDEOH 7HVWV DQG
*URZWK;9,SDJHV
'\QDPLF6SHFLÀFDWLRQ;SDJHV
9RO $ 6 7DQJLDQ &RQVWUXFWLQJ DQG $SSO\LQJ
9RO76WHJHU7UDQVLWLRQDO'\QDPLFVDQG(FRQRPLF
2EMHFWLYH)XQFWLRQV;,,SDJHV
*URZWKLQ'HYHORSLQJ&RXQWULHV9,,,SDJHV
9RO 6 0LQQHU 6WUDWHJLF 6DIHW\ 6WRFNV LQ 6XSSO\ 9RO 0 .OSPDQQ 6WRFN 0DUNHW 2YHUUHDFWLRQ DQG
&KDLQV;,SDJHV )XQGDPHQWDO9DOXDWLRQ,;SDJHV
9RO0(KUJRWW0XOWLFULWHULD2SWLPL]DWLRQ9,,, 9RO:%=KDQJ$Q(FRQRPLF7KHRU\RI&LWLHV;,
SDJHV SDJHV
9RO73KDQ+X\&RQVWUDLQW3URSDJDWLRQLQ)OH[LEOH 9RO . 0DUWL 6WRFKDVWLF 2SWLPL]DWLRQ 7HFKQLTXHV
0DQXIDFWXULQJ,;SDJHV 9,,,SDJHV
9RO - =KX 0RGXODU 3ULFLQJ RI 2SWLRQV ; 9RO 6 :DQJ < ;LD 3RUWIROLR DQG $VVHW 3ULFLQJ
SDJHV ;,,SDJHV
9RO ' )UDQ]HQ 'HVLJQ RI 0DVWHU$JUHHPHQWV IRU 9RO*+HLVLJ3ODQQLQJ6WDELOLW\LQ0DWHULDO5HTXLUH
27&'HULYDWLYHV9,,,SDJHV PHQWV3ODQQLQJ6\VWHP;,,SDJHV
9RO , .RQQRY &RPELQHG 5HOD[DWLRQ 0HWKRGV IRU 9RO % 6FKPLG 3ULFLQJ &UHGLW /LQNHG )LQDQFLDO
9DULDWLRQDO,QHTXDOLWLHV;,SDJHV ,QVWUXPHQWV;SDJHV
9RO 3 :HL 8QHPSOR\PHQW LQ 2SHQ (FRQRPLHV 9RO+,0HLQKDUGW&RRSHUDWLYH'HFLVLRQ0DNLQJ
;,,SDJHV LQ&RPPRQ3RRO6LWXDWLRQV9,,,SDJHV
9RO-,QNPDQQ&RQGLWLRQDO0RPHQW(VWLPDWLRQRI 9RO61DSHO%LODWHUDO%DUJDLQLQJ9,,,SDJHV
1RQOLQHDU(TXDWLRQ6\VWHPV9,,,SDJHV
9RO $ .ORVH * 6SHUDQ]D / 19DQ:DVVHQKRYH 9RO % )OHLVFKPDQQ$ .ORVH 'LVWULEXWLRQ /RJLV
(GV 4XDQWLWDWLYH$SSURDFKHV WR 'LVWULEXWLRQ /RJLVWLFV WLFV;SDJHV
DQG6XSSO\&KDLQ0DQDJHPHQW;,,,SDJHV 9RO5+DIQHU6WRFKDVWLF,PSOLHG9RODWLOLW\;,
9RO % *ODVHU (IÀFLHQF\ YHUVXV 6XVWDLQDELOLW\ LQ SDJHV
'\QDPLF'HFLVLRQ0DNLQJ,;SDJHV 9RO'4XDGW/RW6L]LQJDQG6FKHGXOLQJIRU)OH[L
9RO 5 &RZDQ 1 -RQDUG (GV +HWHURJHQRXV EOH)ORZ/LQHV;9,,,SDJHV
$JHQWV ,QWHUDFWLRQV DQG (FRQRPLF 3HUIRUPDQFH ;,9 9RO 0 :LOGL 6LJQDO ([WUDFWLRQ ;, SDJHV
SDJHV
9RO & 1HII &RUSRUDWH )LQDQFH ,QQRYDWLRQ DQG 9RO'.XKQ*HQHUDOL]HG%RXQGVIRU&RQYH[0XOWL
6WUDWHJLF&RPSHWLWLRQ,;SDJHV VWDJH6WRFKDVWLF3URJUDPV;,SDJHV
9RO:%=KDQJ$7KHRU\RI,QWHUUHJLRQDO'\QDP 9RO*1.ULHJ.DQEDQ&RQWUROOHG0DQXIDFWXULQJ
LFV;,SDJHV 6\VWHPV,;SDJHV
9RO 7 /X[ 6 5HLW] ( 6DPDQLGRX 1RQOLQHDU
9RO 0 )U|OLFK 3URJUDPPH (YDOXDWLRQ DQG 7UHDW
'\QDPLFV DQG +HWHURJHQHRXV ,QWHUDFWLQJ $JHQWV ;,,,
PHQW&KRLVH9,,,SDJHV
SDJHV
9RO 6 6SLQOHU &DSDFLW\ 5HVHUYDWLRQ IRU &DSLWDO 9RO-/HVNRZ03XFKHW$Q\XO/)3XQ]R1HZ
,QWHQVLYH7HFKQRORJLHV;9,SDJHV 7RROVRI(FRQRPLF'\QDPLFV;,;SDJHV
9RO&)'DJDQ]R$7KHRU\RI6XSSO\&KDLQV9,,, 9RO & 6XHULH 7LPH &RQWLQXLW\ LQ 'LVFUHWH 7LPH
SDJHV 0RGHOV;9,,,SDJHV
9RO & ( 0HW] ,QIRUPDWLRQ 'LVVHPLQDWLRQ LQ 9RO%0|QFK6WUDWHJLF7UDGLQJLQ,OOLTXLG0DUNHWV
&XUUHQF\&ULVHV;,SDJHV ;,,,SDJHV
9RO 5 6WROOHW] 3HUIRUPDQFH $QDO\VLV DQG 2SWL 9RO5)RHOOPL&RQVXPSWLRQ6WUXFWXUHDQG0DFUR
PL]DWLRQRI,QERXQG&DOO&HQWHUV;SDJHV HFRQRPLFV,;SDJHV
9RO:.UDEV6:3LFNO$QDO\VLV&RQWUROODELOLW\ 9RO-:HQ]HOEXUJHU/HDUQLQJLQ(FRQRPLF6\VWHPV
DQG2SWLPL]DWLRQRI7LPH'LVFUHWH6\VWHPVDQG'\QDPL ZLWK([SHFWDWLRQV)HHGEDFNSODQQHG
FDO*DPHV;,,SDJHV 9RO 5 %UDQ]HL ' 'LPLWURY 6 7LMV 0RGHOV LQ
&RRSHUDWLYH*DPH7KHRU\9,,,SDJHV
9RO 5 :DSOHU 8QHPSOR\PHQW 0DUNHW 6WUXFWXUH
DQG*URZWK;;9,,SDJHV 9RO 6 %DUEDUR (TXLW\ DQG (IÀFLHQF\ &RQVLGHUD
WLRQVRI3XEOLF+LJHU(GXFDWLRQ;,,SDJHV
9RO0*DOOHJDWL$.LUPDQ00DUVLOL(GV7KH
&RPSOH[ '\QDPLFV RI (FRQRPLF ,QWHUDFWLRQ ;9 9RO0)DOLYD0*=RLD7RSLFVLQ'\QDPLF0RGHO
$QDO\VLV;SDJHV
SDJHV
9RO06FKXOPHULFK5HDO2SWLRQV9DOXDWLRQ;9,
9RO.0DUWL<(UPROLHY*3ÁXJ(GV'\QDPLF
SDJHV
6WRFKDVWLF2SWLPL]DWLRQ9,,,SDJHV
9RO$YRQ6FKHPGH,QGH[DQG6WDELOLW\LQ%LPDWUL[
9RO * 'XGHN &ROODERUDWLYH 3ODQQLQJ LQ 6XSSO\ *DPHV;SDJHV
&KDLQV;SDJHV
9RO + %RE]LQ 3ULQFLSOHV RI 1HWZRUN (FRQRPLFV
9RO05XQNHO(QYLURQPHQWDODQG5HVRXUFH3ROLF\ ;;SDJHV
IRU&RQVXPHU'XUDEOHV;SDJHV 9RO 7 /DQJHQEHUJ 6WDQGDUGL]DWLRQ DQG ([SHFWD
9RO ; *DQGLEOHX[ 0 6HYDX[ . 6|UHQVHQ WLRQV,;SDJHV
97·NLQGW(GV0HWDKHXULVWLFVIRU0XOWLREMHFWLYH2SWL 9RO$6HHJHU(G5HFHQW$GYDQFHVLQ2SWLPL]D
PLVDWLRQ,;SDJHV WLRQ;,SDJHV
9RO5%UJJHPDQQ0RGHO5HGXFWLRQ0HWKRGVIRU 9RO 3 0DWKLHX % %HDXÀOV 2 %UDQGRX\ (GV
9HFWRU$XWRUHJUHVVLYH3URFHVVHV;SDJHV $UWLÀFLDO(FRQRPLFV;,,,SDJHV
9RO$(VVHU3ULFLQJLQ,Q&RPSOHWH0DUNHWV;, 9RO : /HPNH7HUP 6WUXFWXUH 0RGHOLQJ DQG (VWL
SDJHV PDWLRQLQD6WDWH6SDFH)UDPHZRUN,;SDJHV
9RO6.RNRW7KH(FRQRPHWULFVRI6HTXHQWLDO7UDGH 9RO 0 *HQVHU $ 6WUXFWXUDO )UDPHZRUN IRU WKH
0RGHOV;,SDJHV 3ULFLQJRI&RUSRUDWH6HFXULWLHV;,;SDJHV
9RO 1 +DXWVFK 0RGHOOLQJ ,UUHJXODUO\ 6SDFHG 9RO$1DPDWDPH7.DL]RXML<$UXJD(GV7KH
&RPSOH[ 1HWZRUNV RI (FRQRPLF ,QWHUDFWLRQV ;,
)LQDQFLDO'DWD;,,SDJHV
SDJHV
9RO + .UDIW 2SWLPDO 3RUWIROLRV ZLWK 6WRFKDVWLF
9RO 0 &DOLHQGR 0LFURHFRQRPHWULF (YDOXDWLRQ RI
,QWHUHVW5DWHVDQG'HIDXOWDEOH$VVHWV;SDJHV /DERXU0DUNHW3ROLFLHV;9,,SDJHV
9RO*\&KHQ;+XDQJ;<DQJ9HFWRU2SWLPL 9RO / 1HXEHFNHU 6WUDWHJLF &RPSHWLWLRQ LQ 2OLJR
]DWLRQ;SDJHV SROLHVZLWK)OXFWXDWLQJ'HPDQG,;SDJHV
9RO-/LQJHQV8QLRQ:DJH%DUJDLQLQJDQG(FRQRP 9RO-:RR7KH3ROLWLFDO(FRQRP\RI)LVFDO3ROLF\
LF*URZWK;,,,SDJHV ;SDJHV
9RO & %HQNHUW 'HIDXOW 5LVN LQ %RQG DQG &UHGLW 9RO7+HUZLJ0DUNHW&RQIRUP9DOXDWLRQRI2SWLRQV
'HULYDWLYHV0DUNHWV,;SDJHV 9,,,SDJHV

Agent-Based Economics and Econophysics

Uploaded by

Copyright:

Available Formats

Agent-Based Economics and Econophysics

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Agent-Based Economics and Econophysics

Uploaded by

Copyright:

Available Formats

/HFWXUH1RWHVLQ(FRQRPLFV

Understanding the mechanism of a socio-economic system requires more than

problems appear throughout this book. We also encouraged papers dealing

October 2005 Akira Namatame

Five Years of Continuous-time Random Walks in Econophysics

Growth Volatility Indices

Financial Fragility and Scaling Distributions in the Laboratory

Part II Complex Economic Network

Heterogeneous Economic Networks

The Emergence of Paradigm Setters Through Firms’

Part III Economic Dynamics

Statistical Properties of a Heterogeneous Asset Pricing Model

Deﬂationary Recessions in a General Equilibrium Framework

Part IV Agent-based Modeling

A Functional Modularity Approach to Agent-based Modeling

Cultural Evolution in a Population of Heterogeneous Agents

Part V Auction and Two-sided Matching

Counterclockwise Behavior Around the Beveridge Curve

The Waiting-Time Distribution of Trading Activity in a

Part VI Minority Games and Collectie Intelligence

Theoretical Analysis of Local Information Transmission in

How Does Collective Intelligence Emerge in the Standard

Part VII Game-theoretic Approach

What Information Theory Says About Bounded Rational

Evolution of Reciprocal Cooperation in the Avatamsaka

Summary. This paper is a short review on the application of continuos-time ran-

tober 1999. For instance, according to a contingency-table analysis performed

assumptions on markets. The model itself can be corroborated or falsiﬁed by

and the survival function Ψ (τ ) is:

whereas the Fourier transform of a (generalized) function f (x) is deﬁned as:

A generalized function is a distribution (like Dirac’s δ) in the sense of S. L.

2.1 Limit Theorems: The Uncoupled Case

Indeed, P (n, t) is given by:

where ψn (τ ) is the n-fold convolution of the waiting-time density:

deﬁned by the following power series in the complex plane:

The Mittag-Leﬄer function is a possible model for a fat-tailed survival func-

whereas for large τ , it has the asymptotic representation:

and the asymptotic representation is:

xn (h) = hξ1 + hξ2 + . . . + hξn , (29)

Now, if we assume the following asymptotic behaviours for vanishing h and r:

λ̂(hκ) ∼ 1 − hα |κ|α ; 0 < α ≤ 2, (32)

The Laplace transform of eq. (34) is:

2.2 Limit Theorems: The Coupled Case

and if, for h → 0 and r → 0, the asymptotic relation holds:

The Fourier-Laplace transform of the scaled conditional probability density

 s), one asymp-

of characteristic functions, after the application of the analogous theorem for

If the Fourier-Laplace transform of ϕ(ξ, τ ) is regular for κ = 0 and s = 0, and,

Due to the hypothesis of regularity in the origin and to the properties of

3 Summary and Outlook

8. R. Engle and J. Russel, Autoregressive conditional duration: A new model for

Masanao Aoki1 and Hiroshi Yoshikawa2

placed at sites of lattices. However, no explicit notions of similarity, correla-

ment speeds as functions of some similarity measures among clusters, Taylor

3 Two Types of Lags in Tree Dynamics

Responses of a macroeconomic price index to shocks to one of its component

H(s)Y (s) = U (s),

The time domain expression is

described by a second order diﬀerential equation with two stable eigenvalues

s), one asymp-

i = j, where q = exp(−γd(i, j)) = exp(−γ), for all i and j between 1 and 4,