Lecture Notes in Economics

and Mathematical Systems 463

Founding Editors:
M.Beckmann
H. P. Kiinzi

Editorial Board:
H. Albach, M. Beckmann, A. Drexl, G. Feichtinger, W. Giith,
W. Hildenbrand, P. Korhonen, W. Krelle, H. P. Kiinzi, K. Ritter,
U. Schittko, P. SchOnfeld, R. Selten

Managing Editors:
Prof. Dr. G. Fandel
Fachbereich Wirtschaftswissenschaften
Femuniversitat Hagen
Feithstr. 140/AVZ II, D-58084 Hagen, Germany
Prof. Dr. W. Trockel
Institut fUr Mathematische Wirtschaftsforschung (IMW)
Universitat Bielefeld
Universitatsstr. 25, D-33615 Bielefeld, Germany
Springer-Verlag Berlin Heidelberg GmbH
Gilbert Abraham-Frois (Ed.)

Non-Linear Dynamics
and Endogenous Cycles

Springer
Editor
Prof. Gilbert Abraham-Frois
University Paris X Nanterre
Centre National de la Recherche Scientifique
200, Avenue de la Republique
Batiment G
F-9200l Nanterre Cedex, France

Cataloging-in-Publication Data applied for

Die Deutsche Bibliothek - CIP-Einheitsaufnahme

Abraham-Frois, Gilbert:
Non-linear dynamics and endogenous cycles / Gilbert Abraham-
Frois. - Berlin ; Heidelberg ; New York ; Barcelona ; Budapest ;
Hong Kong ; London ; Milan ; Paris ; Santa aara ; Siugapore ;
Tokyo: Springer, 1998
(Lecture Dotes in ecoDomics and mathematical systems ; 463)
ISBN 978-3-540-64321-0 ISBN 978-3-642-58901-0 (eBook)
DOI 10.1007/978-3-642-58901-0

This work is subject to copyright. AII rights are reserved, whether the whole or part
of the material is concemed, specifically the rights of translation, reprinting, re-use
of ilIustrations, recitation, broadcasting, reproduction on microfilms or in any other
way, and storage in data banks. Duplication of this publication or parts thereof is
permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permis sion for use must always be obtained from
Springer-Verlag. Violations are liable for prosecution under the German Copyright
Law.
© Springer-Verlag Berlin Heidelberg 1998
Originally published by Springer-Verlag Berlin Heidelberg in 1998

The use of general descriptive names, registered names, trademarks, etc. in this
publication does not imply, even in the absence of a specific statement, that such
names are exempt from the relevant protective laws and regulations and therefore
free for general use.
Typesetting: Camera ready by author
SPIN: 10649636 43/3142-543210 - Printed on acid-free paper
Table of Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 1
Gilbert Abraham-Frois; Nanterre (France)

Part 1: Economic Growth Models

Infrastructure in an Adaptative Economizing Model of Economic

Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Richard H. Day / Zhigang Wang / Min Zhang; Los Angeles (USA)
Neo-classical Growth and Complex Dynamics: A Note on Day's
(1982) Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 27
Gilles Dufrenot; Creteil (France)

Part 2: Business Cycles

Hysteresis Arising From Individual Inertia and Behavioral Het-

erogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 49
Reiner Franke; Bremen (Germany)
Non Linear Dynamics and Utility Functions in Overlapping Gen-
erations Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 83
Gilles Dufrenot; Creteil (France) / Laurent Mathieu; Nanterre (France)
Pay-as-You-Go System under Permanent Business Cycle .. 111
Michel Botomazava; Antananarivo (Madagascar) / Vincent Touze; Malakoff
(France)

Part 3: Keynesian Models

Relaxation Cycle Chaotic Dynamics and Limit Cycle: a Model

with Keynesian 'Flavour' . . . . . . . . . . . . . . . . . . . . . . . . 127
Gilbert Abraham-Frois / Edmond Berrebi; Nanterre (France)
Credit and Financial Markets in Keynes' Conception of Endoge-
nous Business Cycles: an Interpretation . . . . . . . . . . . . . . 147
Richard Arena / Alain Raybaut; Valbonne (France)
VI

Part 4: Methodological Issues

Business Cycles, Chaos and Predictability . . . . . . . . . . . . 165
Alfredo Medio; Venezio (Italla)
Long-term Memory and Chaos: a Note . . . . . . . . . . . . . . 185
Gilbert Abraham-Frois / Sandrine Lardic / Valerie Mignon; Nanterre (France)

Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Introduction

This book is the result of a workshop organized by MODEM, research

unit supported simultaneously by French "Centre National de la Recherche
Scientifique" (C.N.R.S) and University of Paris-X-Nanterre on May 23-24th
1996.
When dealing with" Non-linear dynamics and endogeneous cycles", it
seems quite difficult, and rather contradictory to give a clear separation
between cycles and growth. However, it may be used as a useful starting
point for presentation of two kinds of papers, a first lot one more concerned
with" economic growth models" , a second mainly with" business cycles" .
Moreover, it has appeared that other papers could be characterized briefly
as dealing with" keynesian models", forming consequently a third part of
this volume, the fourth part being centered about methodological issues,
concerning predictability problems and differences between stochastic and
deterministic systems.

Considerable work has been done on chaotic dynamics in the field of

economic growth and dynamic macroeconomics. The study of chaotic
dynamics in economic growth has a root in a paper dating back from
1982 by Richard Day. It is a great pleasure to open the first part of this
volume with a contribution due to R.Day, Z.Wang and M.Zhang dealing
with" Infrastructure in an Adaptative Economizing Model of Economic
Growth". This study presents an adaptative economizing model of economic
growth based on boundedly rational agents that incorporates infrastructure
in terms of physical and human capital and a utility function based
on a lexicographic preference ordering of present and future potential
consumption. The introduction of infrastructure induces a restriction on
capital labor substitution which, when combined with imperfect foresight,
permits factors to be used inefficiently. In such cases the marginal products
of capital and labor used in production are always positive but the marginal
product of the total work force or that of the total capital stock is negative.
Because of the opportunity cost of resources used in infrastructure, the
usual competitive equilibrium conditions are modified. The authors show
that capital accumulation trajectories are both generically asymptotically
stable and generically unstable, converging to a steady state or fluctuating
around one depending on the weight given by a given generation to it~
heirs.
G.Dufrenot's contribution (" Neo-classical Growth and Complex Dynam·
2

ics: A Note on Day's (1982) model") considers new aspects of Day's orig-
inal (1982) contribution. Two aspects of his paper are under discussion
here. Day's contribution may be viewed as a prototype of a model that
can exhibit chaotic dynamics with an ad hoc modification of the standard
text-book approach to nea-classical growth theory. The first aim of this
paper to provide a plausible explanation which can be used to justify the
presence of the "pollution effect": this effect can be reinterpreted within
the framework of a disaggregate economy, as a consequence of heterogene-
ity amongst entrepreneurs, specially in their expectations behaviour. Such
reformulation of Day's seminal model suggest that volatile expectations
may be a cause of growth irregularity even in a one-sector economy, when
commodity markets are imperfect. Another aspect of complex dynamics is
under discussion in Dufrenot's paper. By using bifurcation analysis, it is
shown that solutions to non-linear difference equations that produce irreg-
ular fluctuations are highly sensitive to a variation of the parameters. This
implies that even when conditions for stable growth are present, transient
chaos may be observed; consequently, the economy may be quite long to
converge towards a sequence of capital accumulation which is monotone
over time.

Three papers centered on business cycles analysis constitute the second

part of this volume. The first one, due to R.Franke, appeals to the concept
of hysteresis which has become widespread in explanations of a variety of
economic phenomena. The two other ones are concerned with overlapping
generation models.
R.Franke's paper (" Hysteresis Arising From Individual Inertia and Be-
havioural Heterogeneity") emphasizes a natural sort of individual inertia
and behavioural heterogeneity as a source of hysteresis in aggregate vari-
ables. It sets up a deterministic dynamical model which, however, cannot
be reduced to a standard system of equations. Studying perturbations of
a growth equilibrium, the economy may converge to different new steady
states or approach different periodic growth cycles. The hysteresis effect
in the first case turn out to be asymmetric. A special feature is that the
economy might enter into persistent fluctuations in response to a medium-
sized shock whereas it finds to a (new or old) equilibrium if a small and
also if a sufficiently large shock occurs. In a scenario of repeated stochastic
shocks, it is indicated that a random walk hypothesis is unlikely to be a
useful approximation to describe the evolution of the (equilibrium) growth
rates.
Two other papers are concerned with cycles in overlapping generation
models. J.M.Grandmont (1985) was among the first to study the possibil-
ity of self-sustaining cycles in the overlapping generations models. In his
paper, the origin of these cycles is the conflict between the substitution
effect and the income effect due to the variations of relative prices. Grand-
mont's paper has been criticized on empirical basis: income effects in the
 3

model are two large compared to their values in real life. G .Dufrenot and
L.Mathieu's paper (" Non linear dynamics and utility functions in over-
lapping generations model") provide some theoretical arguments that may
explain why cyclical and complex paths cannot be ruled out. Following
an early suggestion by Dreze and Modigliani, and using a methodology
initiated by Sato, they show a close connection between the hypotheses
made on the Slutzky equations corresponding to excess demand functions
and classes of utility functions from which cycles and complex dynamics
emerge.
In the same frame of analysis (OLG models), M.Botomazava and
V.Touze examine" Pay-as- You-Go System under Permanent Business Cy-
cle". Relying on Reichlin's (1986) model, they give conditions for the ex-
istence and stability of cycles when a pay-as-you-go system is introduced.
To obtain a permanent cycle, a condition is crucial. Nevertheless, this cycle
around the steady state level can be interpreted as two inter generational
inequality criteria. Some computing results are specially interesting: when
the payroll tax rate rises, the lowest welfare level (Rawlsian criteria) and
the intergenerational inequality (egalitarian criteria) can grow up.

Two papers centered around "keynesian models" constitute the third

part of the book. R.Goodwin (1951) was probably the first to introduce
"relaxation cycles" in macro-economics. Hicks's analysis is indeed anterior
(1950), both authors introducing "non-linear accelerator" in their analysis;
but Goodwin's originality is obvious since he has used relaxation cycle
and Lienard-van der Pol equations. Further work has shown that in many
cases relaxation cycle can be considered as a stable "attractor", a rather
peculiar " limit-cycle" , named in some cases " limiting limit cycle" (Chiarella
1990). Moreover, it is well known that cubic functions can give rise to
chaotic dynamics. In their paper (" Relaxation Cycle, Chaotic Dynamics
and Limit Cycle: a Model with Keynesian 'Flavour"'), G.Abraham-Frois
and E.Berrebi, using previous work by Puu (1991) and themselves (1995)
study the relationship between relaxation cycle and chaotic dynamics.
For certain values of the parameters, there is simultaneous appearance of
chaotic dynamics and relaxation cycle; moreover, relaxation cycles are part
of a cubic which constitute (with its immediate neighborhood) an attraction
basin; chaotic part of the cubic may be inside or outside the relaxation cycle,
which appears as an attractor for the cubic, itself an attractor for the model
and its neighbourhood. In some other cases, one observes relaxation cycles
with "thick frontiers" and even dislocation of the relaxation cycles. At
last, one can observe formal similarities between "classical" (Feigenbaum)
bifurcation diagram and cubic.
R.Arena and A.Raybaut's paper gives us an interpretation of" Credit
and Financial Markets in Keynes' conception of Endogeneous Business
Cycles". Of course, it is well known that J.M.Keynes never developed a
proper and systematic theory of the business cycle. However, the Treatise
4

on Money and the General Theory include, according to R.A & A.R
substantial indications on which it is possible to rely on order to explain the
macroeconomic emergence and persistance of economic fluctuations. The
only obstacle lies in the apparent contradictory elements which are present
in Keynes's works; according to to the post-Wicksellian viewpoint of the
Treatise, the emphasis is put on the divergence between the natural and
the market rates of interest and on its effects upon the demand for credit,
while the General Theory stresses the role of human psychology (including
on the stock exchange) in the determination of investment decisions. The
paper shows how it is possible to build a single framework allowing to cope
with both components of the contradictions and, therefore, to explain why
it was only apparent.

Last part of this volume is concerned with methodological issues, more

specifically with the difference between determinist and stochastic systems,
which seems to meet some difficulties here. The conclusion of A.Medio's pa-
per (" Business Cycles, Chaos and Predictability") is "that the implications
for economists are puzzling". The author consider the case in which a model
of optimal growth gives rise to a dynamic equation of a "logistic" type with
chaotic parameter. When the system is Bernoulli, agents face sequences of
values which have the same probabilistic structure as random processes.
Consequently, for a certain class of concrete dynamical systems, the pos-
sibility exists of repressenting them either as deterministic systems (plus
perhaps some random disturbances) or as stochastic processes. And the
conclusion is quite pessimistic about any hope of providing a general test
for distinguishing deterministic chaos and "true" randomness.
The last paper by G.Abraham-Frois, S.Lardic and V.Mignon (" Long-
term Memory and Chaos") has a a similar preoccupation with a different
point of departure. It is well known that the Hurst exponent H, which allows
to detect the presence of long-term memory, is linked to the fractal dimen-
sion D of the underlying process by the relation D = 2 - H. Thus, it may be
tempting to set a link between long-term memory and chaotic processes by
the mean of D. Moreover, another relation does exist with ARFIMA (Auto-
Regressive Fractionally Integrated Moving Average) processes, which are a
generalization of standard ARIMA (p,d,q) processes with d fractional (and
no more an integer): this particular class of discrete time models, is linked
to fractional Brownian motion defined by Mandelbrot and van Ness and
characterized (also) by Hurst exponent. But the estimation of the value of
fractional differencing parameters is far from being conclusive. At last, one
comes to same kind of perplexity as in precedent paper since same process
might be considered either as deterministic or as stochastic.

This brief introduction would not be complete without mentioning, and

thanking, other contributors and discutants of this wokshop, especially
A. d'Autume (MAD, Paris I), Ch. Bidard (MODEM, Paris-X), J. Blot
 5

(CERSEM, Paris I), P. Cartigny (GREQAM, Aix-Marseille II), G. Caz-

zavillan (Venice), R. Courbis (MODEM, Paris-X), M. Currie (Manchester),
J.P. Drugeon (MAD, Paris I), A. Goergen (MODEM, ENS-Cachan), J.M.
Grandmont (CEPREMAP, Paris), 1. Kubin (Munich), T. Lloyd-Braga (Lis-
bon), P. Pintus (EHESS & CEPREMAP, Paris), W. Nisima-Calmel (MO-
DEM, Paris-X), G. Rotillon (MODEM, Paris-X), D. Sands (Orsay), R.
Topol (OFCE, Paris), R. de Vilder (Amsterdam & CEPREMAP), A. Ven-
ditti (GREQAM, Aix-Marseille II). Financial assistance was provided by
CNRS and University of Paris-X-Nanterre.

Special thanks to V.Touze for (precious) editorial assistance.

G.ABRAHAM-FROIS
 Part 1

Economic Growth Models

Infrastructure in an Adaptative
Economizing Model of
Economic Growth
Richard H. DAY, Zhigang WANG and Min ZHANG

1 Introd uction
This study presents an adaptive economizing model of economic growth
based on boundedly rational agents that incorporates infrastructure in
terms of physical and human capital and a utility function based on a
lexicographic preference ordering of present and future potential consump-
tion. We show that capital accumulation trajectories are both generically
asymptotically stable and generically unstable, converging to a steady state
or fluctuating around one depending on the weight given by a given gener-
ation to its heirs.
Our model is a descendant of the L-D-L growth model originally set
forth by Leontief (1965) and developed by Day (1967), Lin (1987), and Day
and Lin (1991). Although this approach is closely related to the temporary
equilibrium analysis explored in macroeconomic settings by Grandmont
and Laroque (1973), (1986), (1990) it is not equivalent even in the case
of a single generation. This is because the introduction of infrastructure
induces a restriction on capital labor substitution which, when combined
with imperfect foresight, permits factors to be used inefficiently. In such
cases the marginal products of capital and labor used in production are
always positive but the marginal product of the total work force or that
of the total capital stock is negative. Also, because of the opportunity
cost of resources used in infrastructure, the usual competitive equilibrium
distribution conditions are modified.
Section 1 describes preferences. Section 2 reviews the RFS produc-
tion function, and section 3 derives the consumption/savings strategy
when preferences have a specific lexicographic character and adaptive ex-
pectations are naive. Section 4 considers the existence and the stabil-
ity /instability properties of capital accumulation paths.
10

2 Preferences

Distinguish between historical time, represented by a subscript and plan-

ning time, represented by a superscript, which represents the anticipated
future. For example, if x were a decision variable, x~ would be the value to
be acted on i periods in the future but determined "now" where "now" is
at historical time t. Then x~ is the value of the decision variable planned
for the current period.
Let Xt be the value actually acted on in the current period. Then we
also assume the temporary equilibrium condition, Xt = x~, that is, the act
planned for the current period is carried out. Plans for the future, however,
are not always carried out so that it could (and usually will) happen that
x~ =I- Xt+i·
As a further convention, when analyzing the planning problem of the
foreword looking agent, we treat Xi as the decision variable to be determined
and x~ the value that is determined for the historical time t. In this paper we
only consider agents who plan ahead a single generation. Let cO be the level
of current consumption to be determined "now" and c1 be the potential
future standard of living in perpetuity, also to be calculated "now." Since
c~ will be temporary equilibrium in our model, we can drop the superscript
in co, denote by c the current level of consumption to be determined and
by Ct the level actually achieved.
It is convenient to think of the economy as a sequence of overlapping
generations. If Xt is the number of households, then the number of children
per family is 2(1+n)xt and the number of households in the next generation
is Xt+l = (1 + n)xt. Adults of a given generation are manager-worker-
owners to whom all proceeds of production are distributed. They determine
consumption and savings for themselves and their children. The savings are
invested and the capital stock that results, allowing for depreciation in the
meantime, is the bequest which constitutes the endowment for the next
generation of adults. Suppose the adults explicitly consider the trade-off
between their own consumption and the standard of living which their heirs
can potentially enjoy and determine their behavior accordingly. The adults
of the next period then repeat the same economizing decision but on the
basis of the capital stock inherited from their predecessors.
With these conventions the preferences of the current generation of adults
are represented by a utility function,

v (c) ,0:::; c,O :::; c1

v (c) + 1/Jv (c
1) , C :::; c, 0 :::; c 1 :::; C
(1+1/J)v(c+w(c-c)+1/Jv (c 1 -c)) ,c:::; c, c:::; c1
(1)
 11

where v(·) and w(·) are twice differentiable, v' > 0, w' > 0, v" < 0, w" < 0
and v(O) = w(O) = 0. 1

Figure 1: The C* Preferences Between Current Consumption and the

Potential Future Standard of Living

This utility function means that there is a threshold, c, short of which no

utility accrues from considering future consumption. Above that threshold,
the sustainable standard of living that can potentially be enjoyed by the
next generation is given priority over additional current consumption.
Then a second threshold exists (also c) in the level of anticipated (future)
consumption. If these thresholds are exceeded, current consumption and
the anticipated future sustainable standard of living both increase current
utility. It is well known that the preferences thus described cannot be
represented by a continuous utility function. However, the most preferred
choices do depend continuously on income. 2

1 If both e l and e exceed e and if each future generation actually maintained

the capital stock it received, then 'I/J [vee) + w(e l - e)] would be equivalent to the
discounted sum of the utilities of an anticipated constant stream of consumption
el. Thus,

a[v(e) + w(e l - e)l + a 2 [v(e) + w(e l - e)l + ... l~a [v (e) + w(e l - e)l
'I/J[v(e) + w(e l - e)l
where a is the time preference with 0 < a < 1 and 'I/J = a/(l - a). As noted
above, however, in general ei f. et+l so that this equivalence can hold exposte
only at a stationary state. See above.
12

3 Technology

Consider the conventional production function

Y = BF{K,L) (2)
where Y is aggregate output, L is the available supply of labor, K is
production capital, and B represents the "neutral" level of technology. It
is assumed that F{·) is concave, twice differentiable and homogeneous of
the first degree, that

F{O,L) = F{K,O) = ° for all L,K~O,

and that
· of
11m Ii of
K-O u
!3K=OO= L-O
m !3L'
u

In the spirit of Boserup (1981) and North (1981) infrastructure is taken

account of in the following way. Total capital, Z, consists of the component,
K, employed in market production, and the component, k, employed in
the socioeconomic infrastructure,

(3)
The total work force, x, consists of the labor force employed directly in
market production, L, and the infmstructural work force, L. That is,

(4)
We assume that one adult equivalent is used in activities internal to the
family and one adult equivalent is allocated to either the socioeconomic
infrastructure or to market production. According to this treatment, x is
the number of adults equivalents employed outside the family. Equivalently,
it is the number of "families."

2The preferences of this model are an example of the r:o lexicographic

ordering illustrated in a variety of contexts by Encarnacion. An analysis of the
continuity of choice functions based on general Co preferences is contained in
Day (1995). In the present setting Co preferences mean that adults wish to
insure their own survival first (for, of course, the survival of their children depends
upon their own); given this they save so as to insure the survival of their children.
Given that their children can enjoy a standard of living above the given threshold,
parents consider the trade-off between their own and their children's standard of
living.
 13

We assume that no production is possible without the appropriate

infrastructural resources and that the total capital and total work force
cannot be reallocated between production and infrastructure within the
current generational period. Substituting (3) and (4) into (2), we get

,X:::; Lor z:::; f<

y={ ~F[Z-f<,x-L] ,X> L,z > f<
(5)

Suppose that the infrastructural work force required by the technology

is proportional to the total capital stock,
- 1
L= -Z. (6)
II
This implies that for the capital stock in existence to be productive, the
required amount of human capital must be allocated to the infrastructure.
Correspondingly, suppose that the capital requirement for the infrastruc-
ture depends on the total number of families (= the total work force),

f< = /-LX (7)

This implies that for a given population to be productive, the requisite
amount of capital must be allocated to the infrastructure. Rewriting (5)
gives the production function in the (Z, x) space

y_ { 0 ,(Z,X) ~ C
(8)
- BF (Z - /-LX, IIX - Z) ,(Z,x) E C
where the feasibility cone is
1 1
C:= {(Z,x) I /-Lx S Z S IIX} == {(Z, x) I -Z S X S - S Z}.
II /-L
We thus arrive at a Restricted Factor Substitution (RFS) production
function called for by Eisner (1993), Kaldor (1959), and Solow (1959), and
derived by Day and Zou (1994).
Dividing through both sides of (8) by X and taking advantage of the
homogeneity of FO, we get the production function in per capita terms,

y_ { 0 ,Z E ZO:=\(/-L,II)
(9)
- BF (Z - /-LX, IIX - Z) ,Z E Z':=(/-L,II)
where the production capital/work force ratio is z = Z/x and the out-
put/total work force ratio is y = Y/x. Consequently, the infrastructure
determines the range offeasible capital/labor ratios, Z' := (/-L, II).
Given the assumptions stated above, routine calculations show that fO
is twice differentiable, and strictly concave for the feasible capital ratios,
that is,

f"(z) <0, ZEZ', (10)

14

and that

lim f,(z) = 00, lim f,(z) = -00 (11)

Z~IJ. %-+11

Zou (1991, pp. 58-60) showed that for any given output level, the efficient
capital/labor combinations lie in a cone, say C·, within the feasibility
cone C, but the isoquants on which these combinations lie approach the
boundaries of C asymptotically and extend outside C·. In other words,
C \ C* contains capital/labor ratio combinations that are feasible but
inefficient. The rays defining C· are determined by constants, say J-L*, v·
with J-L < J-L* < v* < v which defines the interval of efficient capital/labor
ratios

Z *·.= [*
J-L ,v *] ,

evidently, f' (z) ~ 0 for all z E Z*.

Using the Cobb-Douglas Production Function

Y = BF(K,L) = BKf3(L)I-f3 (12)

the RFS production function is

0 ,(Z,x)~C
y= { I (13)
B (Z - ILx)f3 (vx - Z) -f3 ,(Z,x)EC
where B= B v f3- I . In terms of the capital/labor ratio,

,z ~ ZO
(14)
,z E Zf.
The feasibility cone Zf = (IL, v). The rays define the efficiency cone are
given by

J-L* = {3J-L + r:_

(3)v and v* = {3v + (1 - (3)J-L.

Figure 2 illustrates the isoquants of the RFS production function (13).

 15

Figure 2: Isoquants of the RFS Production Function

Consider the factor combinations shown. Point A lies outside the effi-
ciency cone. The same amount of output could be obtained by a combi-
nation D. Such a point would exhibit excess capacity and unemployment,
but it would necessarily involve free disposal of capital and labor. Points C
involves an increase in output which could be attained by reallocating some
production capital to the infrastructure. This would require fungibility in
the short run and, although excess capacity would be eliminated, some
labor would have to be unemployed. Point B would involve the same in-
creased output level as points A and C but with full employment. It would
require an increase in production capital and the reallocation of part of the
labor force to the infrastructure.
All these of those moves from A to B, C or D would require a complete
knowledge of the production function and the ability to exploit it within
the current time period. They would also require a complete fungibility of
human and production capital in the short run, that is, within the current
production period.
Interpreted literally, this means that if the political and economic admin-
istrators and managers could perceive and understand society's production
structure as a whole. More production and a higher standard of living could
be achieved for any given combination outside of C· by rearranging capital
and labor so as to obtain a combination within C· . It seems to us unrealistic
to suppose that they always manage to do this. In any case, we consider
the situation where they do not. 3

3Infrastructure has always been considered by development economists and is

now receiving increasing emphasis as reflected, for example, in a recent World
Bank (1995) report. That report makes it clear that many countries have in
the past failed to invest sufficiently or effectively in adequate infrastructure to
support growth. There is a growing body of empirical work aimed at estimating
16

4 The Consumption/Savings Strategy

The potential sustainable future standard of living anticipated by the

present generation must account for the maintenance of the capital stock
and an endowment for net additions to the family, so

(15)
where yl is anticipated sustainable income level for the next generation,
Zl, is the planned future capital stock. The capital bequest sufficient to
sustain this standard of living is equal to the amount of capital remaining
at the end of the period plus sufficient investment to allow for population
growth. Therefore,

1
zl = --[y - e + (1- 6)z]. (16)
l+n
Suppose our forward looking agents perceive current output y and the
current rate of return r. Suppose further that they use the latter as a basis
for the consumption/bequest trade-off. If rl is the rate of return anticipated
for the next period, then, in effect,

Our boundedly rational agents then get a rough estimate or ''first-order

approximation" of yl by setting

yl = Y + r(zl - z) (17)
Given these assumptions, the consumer's budget set is

B:= {(c,e1 ) I pe+ e1 ~ (1 + p)[y - (n + 6)z], e ~ O,e 1 ~ O}

where, for notational convenience,
r - (n + 6)
p= l+n .
Obviously, p can be positive, zero, or negative, depending on whether or
not r is greater than, equal to, or less than (n + 6).
The most preferred combination of present consumption and the antic-
ipated future sustainable standard of living, (e, e1 ) must satisfy the maxi-
mizing relationship,

V(y,z,p):= max u(e,e 1 ) (18)

(c,c1)EB

its contributions. Some references are given in Day and Zou (1994).
 17

Clearly, the optimal current consumption is a function of z, y, and p, which

we denote by c(z, y, p).
It should be emphasized that our boundedly rational agents do not
account for the possibility that each succeeding generation will face a
similar problem. Each generation will weigh the opportunity cost and
satisfaction of current savings given their inherited wealth, their current
income and the current rate of return on savings. Each bequest is, therefore,
merely a potential consumption level which could be enjoyed forever if each
succeeding generation chose to maintain the stock of capital at its inherited
level. In effect each generation bestows a flexible asset, the potential income
in perpetuity, leaving it up to its heirs to decide what to do with it when
they take control.
Given the lexicographic utility function (1), we have

Proposition 1 Current consumption based on agents' adaptive behavior

is given by
y if y ~ cor p < 0
c(z,y, p) = { c ifc < y ~ c + (n + 8)c or p >0
min{y, g(z, y, pH ifC + (n + 8) < y or p > 0
Proof It is obvious that c( z, y, p) = y for y ~ c. Moreover, consumption
in the first period should be as big as possible if p ~ 0, i.e., c(z, y, p) = Y
for p < O.
If p > 0, y > c, but c1 = (l+p) [y - (n + 8)z]-pc ~ cor y-(n+8)z ~ c,
then c(z, y, p) = c. If p > 0 and y - (n + 8)z > c and limc-+o w'(c) = +00,
then w'(c - c) - 'ljJpw'(c 1 - c) is a decreasing function of c for c E
[c,(l+e)[y-~n+6)z]-c]. It approaches +00 as c approaches c and approaches
-00 as c approaches (l+p)[y-(n+6)z]-c. Therefore, for each (y, z, p) there is
e
a unique solution to the following first order condition

WI(C - c) = 'ljJpw'(c 1 - c) (19)

which determines a continuous single valued map c = g(y, z, p). Since
g(y, z, p) is not necessary less than y,
c(y,z,p)=min{g(y,z,p),y} if p>O and y-(n+8)z>c. •

Our model's agents are assumed to observe their income wealth and
the marginal product of capital directly and use this information in their
decisions. In theory, however,

y = BJ(z), r = BJI(z) (20)

Consequently, (unknown to our agents) the adaptively optimal consump-
tion depends on the current capital stock,
18

c = h(z) := c (BJ(Z), z, BJI(zi ~ ~n + 6)) (21)

We call this the implicit consumption strategy.

Proposition 2 The Implicit Consumption Strategy. Assume there exits a

value of z such that BJ(z) - (n + c5)z > c. Then
(i) There exist capital/work force ratios, z', Zll, Z"' satisfying J.L < z' <
Zll < Zlll < 1/ and there exists a positive, differentiable, function,
g : jR+ ---+JR+ such that

BJ(z) , z E Z' := [J.L, z')

{ c , z E Z" := [z', Zll)
h (z ) = min{y,g(z)} , z E ZS := [Z", Zlll)
BJ(z) , z E ZIII := [Zlll, 1/]

(ii) there exists a total capital total work force ratio, Z E dZ s such that

~ r- (n + 6) 1 .
(eqUivalently
~
r=
(l+n)
1/J + (n + 6)).
p= l+n =-;j
Proof From Proposition 1 we can give an explicit form for h(z). There
must exist a unique capital/labor ratio Zlll such that J.L < Zlll < 1/ and
B J'(ZIll) = (n+6). If B J(ZIll) - (n+6)zll' > c, clearly there exist z', Z" with
J.L < z' < Z" < Zlll < 1/ such that BJ(z') = c and BJ(ZIl) - (n + 6)zll = c.
See figure 3.

IJf(;; )

i; {

C{
~----'---.
.,11 zl/l

Figure 3: Critical Capital/Labor Ratios

 19

It is easy to see that

p<O for all Z > z'"
y<c for all f.L < Z < z'
y > c, y - (n + 8)z < c and p > 0 for all z' < Z < z"
y - (n + 8) Z > c and p > 0 for all z" < z < Zlll
Given these facts, the results follow . •

5 Capital Accumulation

Given a population growth rate, n, and a depreciation rate, 8, and

given the adaptive optimizing consumption strategy, the behavior of the
capital/labor ratio is governed by
1
Zt+l = O(Zt) = 1 + n [(1 - 8)zt + B!(zt) - h(zt)].
Given the piecewise nature of h(z), the map 0(·) has a piecewise, multiple
phase structure,

,Zt E Z'
,Zt E Z"
, Zt E Z8
,Zt E Z'"
(22)
which is illustrated in figure 4.

:';:" :;'"

Figure 4: The Dynamic Structure

20

Recalling our notational convention, the capital stock planned at time

t for time t + 1, zl = Zt+l. But Yt+1 =F yl except at a steady state.
Consequently, cl =F Ct+1 except at a steady state.

Proposition 3 Existence of Stationary States. There exists a unique

positive interval of (Bo, BO), such that for all B E (Bo, BO) there are two
distinct positive, steady states, zl, zl-' with zl < zl-'.

Proof Let zO be defined by Bf'(zO) - (n + 8) = 0, so that p(zO) = O. Let

B be defined by Bf(zO) - (n + 8)z = c, so that Z" = zoo As B increases,
Zll decreases and f'(ZIl) increases. B -+ 00 implies p(Z") -+ 00. According
to the Intermediate Value Theorem, there exists a Bo such that

Bof'(Z") - (n + 8) 1
(23)
l+n =~.
If B > Bo, then

Bfl(Zll) - (n + 8) > Ij'I/J (24)

l+n
so that z> Z".
Consider z < z < zO, we know that 0 < Bf'<;>;Jn+6l
< IN. As long as
B is sufficiently large, (26) is true for all z E (z, zO). Thus, p(z"') = Ij'I/J
z
and = z"'. Let BO be the smallest value of B such that (26) is true
for all z E (z, zO). Thus, if B < BO, then z < z"'. Therefore, for any
BE (Bo, BO), z E ZS. But we know that Zll is a steady state where c = c
and c1 = C. Thus, there exist two distinct steady states, zl and z/J, where
zl = Zll and zl-' E ZS. When B = Bo, there exists unique z = Z". When
B < Bo, no steady state exists. When B ~ BO, z = z", which is no longer
a steady state since c(z"') = B f(z"') . •

Proposition 4 Boundedness of Trajectories in the Capital/Work Force

Ratio. The smaller steady state is unstable and all trajectories r( z) < zl
converge to zero and all trajectories r(z) such that zl-' > z > zl grow for
at least one period. Given n, 8 > 0, there exist constants t s , zm, zM such
that for all z > zl and for all t > t.,

where zm = O(z"') and zM = argmaxzEZ' O(z).

Proof If the initial capital labor ratio is smaller than zl(= Zll), then c = c
and c1 < C. If z E Z', then Zt+l = [~+!r Zl -+ 0 as t -+ 00.
 21

If z' < Zl < z", then

C
1 (-
C,
Bj()
Zl , Zl,
B!'(Zl) - (n
l+n
+ 6)) < - c,

which implies that S(Zl) < (n+6)zl. Thus,

Similarly, we have Z3 < Z2, Z4 < Z3, ... , Zt+l < Zt. As long as z' < Z < z",
we have s(z) = Bj(z) - c. Thus,

Zl - Z2 =
(n + 6)Zl - Bj(Zl) + c,
l+n

(n + 6)Z2 -Bj(Z2) +c
Z2 - Z3 = .
l+n
Remember that P(Zl) > 0, so that B!'(Zl) > (n + 6), we have

Rearrange the terms in the above inequality, we have

Thus, we have Zl - Z2 < Z2 - Z3. Similarly we can prove Z2 - Z3 < Z3 - Z4 <

.... But

Therefore, for an initial Zl such that z' < Zl < z", there exists a sufficiently
large t such that Zt+l E Z'. This means that for any initial Zl < z", Z will
eventually be reduced to such a level that savings are no longer positive.
Therefore, Z --+ 0 as t --+ 00.
If z/.L > Zt > Zl, we know that s(z) > (n + 6)z. This implies that
Zt+l > Zt . •

Proposition 5 Generic Asymptotic Stability and Generic Instability.

With certain weak restrictions on technology, for all positive n, 6 and
B O > B > B o, there exist open sets (consisting of a union of open
intervals) Fffis, FeB, F/.LS such that (i) for all 'ljJ E ;::rns the steady state is
asymptotically stable and is approached monotonically; Oi) for all 'ljJ E F es
capital/labor ratio trajectories eventually exhibit dampening fluctuations
converging to the steady state; for all 'ljJ E F/.LS almost all trajectories
22

(except those that hit z after a finite number of periods) eventually exhibit
persistent fluctuations around the steady state.

Proof Since g(z) satisfies

w'(g(z) - c) = t/Jpw'(c 1 - c)

where c 1 = (1 + p) [y - (n + 8)z]- pc. This implies

[w"(g(z) - c + t/J p2 W "(c 1 c)]dg(z)

_
dz
= t/Jp'w'(c 1 - c) + t/Jpw"(c1 - c)
x {p'[y - (n + 8)z - c] + p(p + 1)(1 + n)}.
At the steady state g(z) = c1 ===} g(z) = c= fj - (n + 8)z and p= i.
Therefore,

1+n t/J _, w' (c - c)

dz = ~ + 1 + l.P w" (c - c)
t/J ""
_, w'(c - c)
===} O'(x)
1- (l+n)2(I+i)r w"(c-c)"

i
Since r = (n + 8) + ===} lim"".....oo z(t/J) = z* ===} lim""..... oo (}'(z) = -00, it
is easy to see that ()' (z) is a continuous function of t/J and ()' (z) < 1 for all
t/J > O. Let

:rn S = {t/J 10< (}'[z(t/J)] < I}

?S = {t/J 1-1 < (}'[z(t/J)] < O}
]="I-'s = {t/J I (}'[z(t/J)] < -I},
then the proposition holds . •
The fluctuations may be cyclic or chaotic and may be strongly ergodic
if the Miseurwcz condition pertains.

6 Example

Assume the utility function is defined by (2) and

w(c) = A(c - c)", 0 < a < 1 (25)

 23

The first order condition when both present and potential future consump-
tion are both positive is

1
-.
(C-1_-C)1-0 -p
_ (26)
'IjJ c- C
Substituting for c\ using (18) and doing some rearranging, it follows that

c = g(z) = (p'IjJ)~ {(I + p)[y - ~+ 8)z]- c} (27)

1 + p(p'IjJ) ",-1
This function, we recall, is defined on z E ZS. This gives the phase
structure when savings are positive:

Os(z) = _1_ {(I _ 8)z + Y _ (p'IjJ)nl:r {(I + p)[y -

1+n
~n + 8)z]-
1 + p(p'IjJ)~
c} z E ZS

(28)
In figure 5 capital/labor and income trajectories are shown.
24

(a) The Capit"l/l.al)l>r Ibtio in Plta.'IC S)IIICI

-'--------_. ..--.--.---
(I» The Ca"il,ili/I.ahar llatio Trajectory

(c) The Per Fan.ily IJII:ome Traj~tory

Figure 5: Sample Trajectories

 25

References
[1] Boserup, Ester, 1981, Population and Technological Change, A Study
of Long Term Trends, The Chicago University Press, Chicago.
[2] Day, Richard H. and Jean-Luc Walter, 1989, "Economic Growth in
the Very Long Run: On the Multiple-Phase Interaction of Population,
Technology, and Social Infrastructure," Chapter 11 in W. Barnett,
J. Geweke, K. Shell (eds.), Economic Complexity: Chaos, Sunspots,
Bubbles and Nonlinearity, Cambridge University Press, Cambridge.
[3] Day, Richard H. and Gang Zou, 1994, "Infrastructure, Restricted
Factor Substitution and Economic Growth," Journal of Economic
Behavior and Organization, 23, 149-166.
[4] Day, Richard H., 1996, "Satisficing Multiple Preferences In and Out
of Equilibrium," in R. Fabella and E. de Dios (eds.), Choice, Growth
and Development: Emerging and Enduring Issues, University of the
Philippines Press, The Philippines.
[5] Eisner, Robert, 1993, A Presentation paper at the Annual Meeting of
the American Economic Association, Los Angeles.
[6] Grandmont, Jean-Michel and Guy Laroque, 1973, "Money in a Pure
Consumption Loan Model," Journal of Economic Theory, 6, 382-395.
[7] Grandmont, Jean-Michel and Guy Laroque, 1986, "Stability of Cycles
and Expectations," Journal of Economic Theory, 40, 138-151.
[8] Grandmont, Jean-Michel and Guy Laroque, 1990, "Economic Dynam-
ics with Learning," in Equilibrium Theory and Applications, Cam-
bridge University Press, Cambridge.
[9] Kaldor, Nicholas, 1959, "Economic Growth and the Problem of Infla-
tion," Economica, 26, 287-298.
[lOJ King, Robert G., and Sergio Bebelo, 1990, "Government Spending in a
Simple Model of Endogenous Growth," Journal of Political Economy,
98, 126-150.
[l1J North, Douglass C., 1981, Structure and Change in Economic History,
W.W. Norton and Company, New York.
[12J Solow, Robert M., 1956, "A Contribution to the Theory of Economic
Growth," Quarterly Journal of Economics, 70,65-94.
26

[13] Solow, Robert M., 1959, "Is Factor Substitution a Crime? If So, How
Bad? Reply to Professor Eisner," Economic Journal, 69, 597-599.
[14] World Bank, 1994, Infmstructure for Development: World Develop-
ment Report, Oxford University Press, New York.
[15] Zou, Gang, 1991, Growth with Development, Ph.D. Dissertation, Uni-
versity of Southern California, Los Angeles.
Neo-Classical Growth and
Complex Dynamics: A Note on
Day's (1982) Model
Gilles DUFRENOT

1 Introduction

Considerable works have been done on chaotic dynamics in the field

of economic growth and dynamic macroeconomics (see Day and Gang [1],
Day [2], Nishimura and Yano [3], Grandmont et al. [4]. The study of chaotic
dynamics in economic growth has its root in a paper dating back from 1982
by Richard Day. Our purpose is to consider new aspects of this original
contribution. Day's [5] established the existence of a chaotic growth due
to the presence of a "pollution effect" in the capital stock accumulation
process. Two aspects of his paper are under discussion here.
Firstly, the author considered a "pollution effect" to illustrate the pos-
sible existence of a negative externality in the aggregate capital stock ac-
cumulation process. This lead him to assuming a production function for
the representative firm of the economy, having a humped concave shape.
However, as argued by Lorenz [6, pp. 123-125], Day's contribution may be
viewed as a prototype of a model that can exhibit chaotic dynamics with
an ad hoc modification of the standard textbook approach to neoclassical
growth theory. In addition to the ad hoc nature of the "pollution effect",
other authors have commented on the lack of any microeconomic founda-
tions in Day's model. Because of this, one may cast some doubts on all the
models that have been developed in the literature and which use the same
methodology. These criticisms which raise methodological problems, have
received no answers in the literature. It is the aim of this note to provide
one plausible explanation that can be used to justify the presence of the
"polluting effect".
It will be shown that the latter can be interpreted within the framework
of a disaggregate economy, as a consequence of the heterogeneity amongst
entrepreneurs, notably in their expectations behavior .. In this context, the
28

formal content given to the "pollution effect" in Day's model is an aggregate

indicator of firms interactions, when entrepreneurs try to coordinate their
decisions to accumulate capital and have no knowledge of the level of
global demand. This formulation of Day's model suggests that volatile
expectations may be a dominant cause of growth irregularity in a one-
sector economy, when commodity markets are imperfect.
This reinterpretation thus leads to consider the question of the microe-
conomic foundations of the "pollution effect" that is responsible for the
presence of a chaotic dynamics. It will become apparent in this study that
volatile expectations in a decentralized economy may have persistent ef-
fects, in aggregate, causing the instability of economic growth. In other
words, there exists a close connection between the non-linear specification
representing the aggregate capital accumulation and the process that leads
firms to adjust their capital stock according to the decisions taken by their
competitors. The idea of locating the nonlinearity that drives Day's dy-
namic results in expectations interactions at the firm level was the subject
of an earlier paper by Albin [7]. Even thought our discussion may seem
close to that of this author, the approach that we follow is rather"different.
Albin stood exclusively at the disaggregated level and did not compare the
properties of the implications of his model with those at the macroeconomic
level. He rather used a microeconomic cellular automaton arguing that the
process of aggregation eliminates certain dynamic evolutions observed at
the individual level. Our contribution consists in showing that the com-
plex expectations interactions can be aggregated to a single function of the
capital-labor ratio just similar to the one analyzed in Day's paper.
The second element which is under discussion deals with some mathe-
matical points concerning the concept of complex dynamics used in Day's
model. It is now agreed that one essential feature of a complex or chaotic dy-
namics is the sensitivity dependence on initial values with positive Lebesgue
measure (this is ergodic chaos). Besides, following Day [8] and Melese and
Transue [9], two other notions have emerged in the literature, namely "thick
chaos" and "thin chaos" . The first concept corresponds to "true" chaos and
is associated with behavior switches from stable periodic orbits to non pe-
riodic time-paths. The notion of thin chaos leads to consider that chaos is
not truly observed despite the fact that there exists high-order cycles that
are indistinguishable from aperiodic fluctuations.
In this paper, we discuss another aspect of complex dynamics. By
using the bifurcation analysis, it is shown that the solutions of non-linear
difference equations that produces irregular fluctuations are highly sensitive
to a variation of the parameters. These solutions change erratically for
certain ranges of parameter values. What this implies is that even when the
conditions for a stable growth are present, transient chaos can be observed
and then it may take the economy a (very) long time to converge towards
a sequence of capital accumulation which is monotone over time.
 29

The paper is organized as follows. In section 2, the model is presented

and we provide some theoretical justifications of its equations. Our mod-
ified formulation of Day's model mainly concerns the assumptions about
expectations. In section 3, we discuss definitions and procedures for the
study of the model stability. One essential feature of this section is that
the complexity of the dynamics comes from the fact that the solutions are
highly sensitive to a variation of parameters.

2 The modified model

We will set up the model and then explain its equations.

Notations:
St: volume of saving at time t
lit: investment of firm i at time t
Kit : capital stock of firm i at time t
Lit: quantity of labor used by firm i at time t
Yit : production of firm i at time t
S : saving rate
Ajt : set of firms j whose decisions have an influence upon firm i's choices
git(A jt ): reaction function of the ith firm faced with its j competitors
K t : capital stock of the economy at time t
L t : quantity of labor used at time t
yt: quantity of good produced by the economy at time t
k t : capital-labor ratio
The model is summarized by the following equations:

lit = F(Kit , Lit), 'Vi = 1, ... N (1)

N
Kt+1 = sF(Kt , Lt ) [1 + ntl, nt = (liN) L git(Ajt ) (4)
i=l

(5)

kt+l = Bkf(1- kd, ko > 0, B = sA (6)

30

Theoretical justifications of the equations

(1) This simply means that the economy is composed of N firms. They
produce ¥it units of a good with Kit units of capital and Lit units of labor ..
F is assumed to be degree one homogeneous. This assumption means that
the number of firms is of no consequence on the results. All the firms
produce the same good and the time delay for production is one period.
(2) For simplicity, we assume that there is no physical depreciation of
capital. Kt+l is therefore net investment.
(3) information is imperfect so that no firm can forecast the level of
future demand. We assume that when confronted to an uncertainty on
the future level of demand, entrepreneurs form their expectations in the
following way. Each individual who observes that many of its competitors
are increasing (or decreasing) their capital stock is led to believe that they
are forecasting a high level of activity for the next period. The only way for
him to benefit from this increase (or to avoid over- production) is to produce
more (or less) and consequently to increase (or decrease) the capital stock
that he invests in the process of production. Each firm i is supposed to
react to the decisions of j competitors (j < N). The sign function sgn(x)
has three values depending upon the sign of x. These values are respectively
1, -1, and 0, depending upon whether x is positive, negative or equals O.
This function is used to formalize the idea that expectations are guided
by the collective psychology of the firms. If an entrepreneur observes that
other competitors have decided to produce more (and thus are investing
more capital today to increase the level of production tomorrow), a positive
externality can result from the fact the entrepreneur believes that he will
benefit from a stronger demand, by which higher revenue he is guaranteed
and more profits. So, increasing concentrations of capital can result from
the fact that an entrepreneur is led to invest an amount of capital greater
than the amount he would have invested had he no paid attention to the
decisions of other entrepreneurs.
(4) The function nt is a formal description of the way firms coordinate
their expectations in aggregate. The usual good market equilibrium condi-
tion requires that the capital stock at time t (or net investment) be equal to
the aggregate saving, that is equation (4) with nt = O. The modification we
introduce here is that the capital stock is increased (if nt > 0), or decreased
(if nt < 0) by an amount that depends on the way firms coordinate their
decisions to accumulate capital. This reflects the consequence in aggregate
of the behavioral interactions which are due to the crossed expectations of
agents. The simplest way to obtain equation (4) is to proceed as follows.
The capital stock accumulation at a given period for an individual firm
has two components: firstly, the quantity of good that has not been supply
on the commodity market (which depends on the saving rate of consumers);
secondly a component which can be interpreted as an externality because
 31

firms are supposed to adjust their production according to the decisions

taken by their competitors. We therefore write the following relations:

(7)
or

(8)

Assuming that all firms employ the same quantity of capital and labor,
we write

F(kit,Lit) = (lIN)F(Kt ,Lt ) where Kt = NKit, L = NLit (9)

Then, we sum up the firms on each side of (8) and thus get:

or
N
Kt+1 = sF(Kt ,Lt)[l + (liN) Lgit(Ajt)l (11)
i=l

By examining equation (11), we see that we have to consider the problem

of aggregation of the functions git(A jt ). To start with a simple example
consider an economy composed of four firms. Suppose that in observing
each other, three of them decide to reduce their capital accumulation
because they forecast a decrease in the level of activity for the next period.
Assume further that the fourth firm decides not to modify its rhythm of
accumulation. These hypotheses leads to the following relations:

(12)

(13)

(14)

(15)
32

From the relations above, we have

4 3
(1/4) 'L::git(A jt ) = -(1/4) 'L::ai < 0 (16)
i=l i=l

In the general case, if we note the average aggregate function by nt, this
term will be negative if the entrepreneurs, taken as a whole, are pessimistic
about the level of activity for the forthcoming period (this is the case in
our example). Conversely, a positive value will indicate moods of optimism.
(5) For illustration we choose a Cobb-Douglas production function as
is frequently used in neoclassical one-sector growth models. Since the
functional F(Kit , Lit) is independent of i, we assume that f(kt} is the
reduced production function of the economy. The per-capita production
is a continuous function of the capital-labor ratio and it has continuous
derivatives of all required orders for k t > O. In particular, the marginal
productivity of capital is a decreasing function of k t , while the marginal
productivity oflabor is increasing with kt . This implies that j' (k t ) > 0 and
((Kd < O.
Furthermore, it is noteworthy that nt is dependent on both the value of
k t and its sign. As one can easily understand this variable does fluctuate
following the variation of aggregate expected demand. Indeed, since we
assume that return to scale are constant, the supply of good and the
demand for factors of the firms are a priori undetermined (they vary
between 0 and infinity). In other words, there exists an infinity of capital-
labor ratio corresponding to the optimal profit. Production is therefore
determined by the level of aggregate demand that entrepreneurs forecast.
Those try to find the proof of their expectations in the decisions of all
others and mechanisms of self fulfillment of representations take place. If
a majority of firms are pessimistic, then they will anticipate a decrease
in demand in comparison to its present level. Their expectations will
realize itself on its own: to avoid loss from the predicted decrease, each
entrepreneur will reduce his investment and the number of workers; the
wages they do not pay will effectively decrease the aggregate demand. If
such behaviors are observed, then the sign of will be negative. In order to
facilitate the comparison with Day's model, we assume for illustration that
nt = h(kt} = -kt . This means that during the periods of increasing per-
capita output the economy resources tend to go beyond the level of global
demand, thereby inducing firms to reduce capital accumulation.
(6) This follows by combining equations (4) and (5) and by dividing
each side of equation (4) by . We further assume that the growth rate of
the population is O. The results are unaltered when a positive growth rate
is assumed.
 33

Comparison with Day's (1982) formulation

Equation (6) is a particular case of Day's formulation who supposed
that the dynamics of the capital-labor ratio was described by the following
equation:

kt+l = sf(kt ) = s[Akf(m - ktPl = Bkf(m - ktP (17)

The term (m - kt)'Y reflects the influence of the "pollution effect" on per-
capital output. Increasing concentrations of capital lead to a saturation
level m. This means that some potential resources of the economy have to
be sacrificed if one wonder to avoid such a saturation.
However, when dealing with equation (17), it is not enough to assume
that a "pollution effect" can occur, but one also need to understand
the economic factors that explain such an effect. Day's approach may
appear ad hoc because the author does not answer the second part of the
question. The developments above show how the nonlinearities induced
by "polluting effects" can be achieved by investigating the role played by
crossed expectations. Expectations are crucial in determining the allocation
of production and capital over time: periods of increasing accumulation
may cause firms to decrease their investment in subsequent periods, if they
believe that the level of future demand will be too low in comparison
to the level of production that would be induced by capital increases.
Furthermore, this decision is not taken individually. Each entrepreneur's
problem is to know and decide whether or not to reduce its investment, in
the light of the choice made by the majority of agents, a behavior which
is given a formal content through the use of the function git(A jt ) and its
aggregate counterpart.
One thus observes that even though equations (6) and (17) have identical
mathematical formulation (When m = 'Y = 1), their economic contents
are different. In the formulation above, the mimetic effect due to crossed
expectations of entrepreneurs explain why the capital accumulation is
described by a logistic equation. Moreover, the production function has the
usual Cobb-Douglas form. On the opposite, in Day's approach the evolution
of the capital-labor ratio has its roots in the hypotheses formulated on
the production technology. However, many authors have argued that in
this second case the "pollution effect" lacks some explanations on the
sources of this effect. We have tried here to answer this question using
a modelling strategy that renders Day's approach more keynesian than
was the author's original contribution. As is well known the problem
of expectation coordination has been brought to the fore by Keynes to
explain the functioning of stock markets. We have simply tried here to
extend his arguments to commodity markets. The presence of a chaotic
accumulation path can be interpreted from equation (6) as a consequence of
imperfections that are due to the organizational structure of the commodity
34

markets. And in this context, expectations of individual firms are guided

by their collective psychology, which can be the very motor of instability
of fluctuations.
We now discuss another aspect in Day's approach, that also concerns
all those models which produce chaotic dynamics from logistic family
equations. The reason for such a discussion is the following. There are
plenty of accounts in the literature on the properties of sensitivity to
initial conditions that characterize complex dynamics. Also, the notion
of ergodic chaos has been emphasized. However, little have been said on
the highly sensitivity to parameters that characterize chaotic models. This
point is important since it helps understanding why transient chaos occurs.
Transient chaos is produced whenever a stationary (or a low-order periodic
cycle) exists as a long term stable equilibrium, but the capital stock behaves
like a chaotic series before its trajectory eventually settles down on the
trajectory that leads to the stable stationary state (or to the stable limit-
cycle). This is the case when the stable solutions coexist with a high number
of high-order limit-cycles that are unstable. In this context, the problem of
sensitivity to initial conditions may be reformulated as follows: one wonders
about the best way for the economy to achieve its stable attractor while
not spending most of its time too far from it. Of course, this depends of the
initial point. The closer the initial capital-labor ratio is from the stationary
state, the higher is the likelihood to achieve the stationary state quickly
and the lesser to observe transient erratic fluctuations. These point are now
examined in the following section.

3 Sensitivity to parameters and transient chaos in

logistic family equations

To investigate the property of sensitivity to parameters in logistic family

equations, we present a simple analytical method that enables to show
that the subsequent bifurcation points can be ordered in sets which have a
"boxes within boxes" structure. The procedure involves four stages.
(l)First, we need a simpler formulation of equation (17).
(2) Secondly, we define the limit bounds of the interval within which the
parameters vary.
(3) Thirdly, we show that the existence of chaotic fluctuations is not
incompatible with the existence of a period-halving bifurcation.
(4) Fourthly, we consider the problem of highly sensitivity to parameters,
when chaotic dynamics coexists with odd periods cycles.
 35

3.1 A REFORMULATION OF EQUATION (17)

We focus here on the case m = {3 = 'Y = 1. This case is simpler to analyze

than the more complex formulation m i= {3 i= 'Y i= 1. This general case is
examined in a forthcoming paper.
Proposition 1 Using a linear change of variable, the dynamics of the
capital-labor ratio can be studied by considering the following· equation:

Xt+1=P(Xt,f.L)=X~+f.L, f.L=-B 2/4+B/2 (18)

Proof. Consider the following change of variable:

Xt = ¢kt +0 (19)
where 0 and ¢ are two real parameters. This implies:

Xt - 0
kt = - ¢ - = (1/¢)Xt - (O/¢) (20)

and:

kt+1 = (1/¢)xt+1 - (O/¢) (21)

Equation (17) is thus written:

Under this restriction, equation 17 leads:

Xt+1 = BXt - BO - (B /¢)x~ + 2(OB /¢)Xt - (B02/¢) +0 (23)

Let us assume that ¢ = -B,O = B/2. Equation (23)becomes:

(24)

•
Remark 1 Under the assumptions made above on the parameters ¢ and 0
the equation (24) is much more simpler to study than equation (17). f.L can
thereby be considered as the bifurcation parameter that we vary and from
which the bifurcation points of the saving rate are deduced.

Remark 2 The values of the saving rate are expressed as follows. Prom
equation (17) we know that sA = B. Since f.L = -B 2/4 + (B/2),we have
B2 - 2B + 4f.L = 0, which implies B 1,2 = 1 ± (1 - 4f.L)1/2 with f.L < 1/4. We
deduce that 81,2 = 1/A ± (1 - 4f.L)1/2/A. Given the restrictions above, we
only consider the greatest solution. We shall assume henceforth that A = 4.
36

3.2 DETERMINATION OF THE LIMIT BOUNDS OF THE

BIFURCATION PARAMETER

Proposition 2 The smallest value of the saving rate s is Smin = (1/4).

Proof. Let us consider the solution Bl = 1 + (1 - 4{L)1/2. Since {L < 1/4,
Bl > 1. This inequality implies:

sA = 1 + (1 - 4{L)1/2 > 1 (25)

and s > (I/A) = s. If A = 4, then the smallest value of the saving rate
iss=I/4 . •
To find the highest value of the saving rate, we first need the following
lemma:

Lemma 3 The successive iterates of the map P(Xt, {L) are written in the
form of two polynomials in Xt and {L.
Proof. We first calculate Xt+2 as P2(Xt, {L ) = P[P(Xt, {L)] = (x~ + {L)2 + {L
and develop this expression:

p 2(Xt, {L) = (xt + 2{Lx~) + ({L2 + {L = ~2(Xt) + G2({L) (26)

~2(xd is a polynomial in x of degree 4 and G 2({L) is a polynomial in {L
of degree 2.
Let us see how Xt+3 is written:

Xt+3 = p 3(Xt, {L) = P[P2(Xt, {L)] = [~2(Xt) + G2({LW + {L (27)

~2(Xt) is a 4th degree polynomial and thus the highest degree of p 3(Xt, {L)
is of the same degree as the term [~2(Xt)]2, that is a polynomial of degree
8. On the other hand, the terms in p 3(Xt, {L) which are dependent on {L
alone are given by the 4th degree polynomial:

(28)
Thus P3(Xt, {L) is written as the sum of two polynomials, one in Xt, of
degree 8 and the other in {L of degree 4.
Suppose that the previous relations are satisfied in the order k-l:

Xt+(k-l ) = P k-l( Xt, {L ) = x 2t

k- 1
+ ,.. + Gk-l ()
{L (29)
where Gk-l({L) is a polynomial in {L of degree 2k - 1. Then:
 37

Note that, just as G3 (J.L) = [G 2(J.L»)2 + J.L in equation (30), the term of
the highest degree in Xt is given by xt-1x2
= xt
and the highest degree of
Gk(lL) is given by the highest degree of Gk- 1 (IL), that is 2k- 2 x 2 = 2k - 1 •
Now, we establish the following proposition.

Proposition 4 Under the conditions of the preceding lemma, the greatest

value of the saving rate is Smax = 1.

Proof. To prove the proposition we use the methodology suggested by

Myberg [10). The intuition of the approach is the following. Let us consider
the recursive equation Xt+l = f(xt, a), where x is a real variable, a is a
real parameter and f is a quadratic map. The values of the bifurcation
parameter a can be found by decomposing the successive iterates of the
function f into two parts : a polynomial in x and another one in a. Let us
note hk(a) this second polynomial for any integer k. Then, the bifurcation
points correspond to the roots of hk(a) = 0 (for a proof of this result the
reader is referred to Myberg [10]).
We already know, from the lemma above that the iterates of the map
P(Xt, IL) can be written as the sum of two polynomials, respectively in x
and IL. Thus, the greatest value of J.L is a solution of Gk(lL) = O. From the
proof of lemma 1 we know that this relation can also be written as:

[G k _ 1 (1L»)2 + IL = 0 (31)
from which we deduce that:

(32)
Assuming that all the terms of the sequence (Gk(J.L», k = 1,2, ... are
positive, the expression above reduces to:

Gk-1(1L) = (_J.L)1/2, IL <0 (33)

This last equation can be rewritten:

(34)
from which:

(35)
One finally obtains:

(36)
or:
38

q is the limit of a series antecedents of rank k, defined at Xt = O. Indeed,

we have:

One can easily show that a limit value of this sequence is:

N(p,) = 1/2 + [1/4 - p,]1/2 (39)

Assuming that N(p,) = q, we deduce that the highest value is p, = -2.
The equality B = sA = 1 + [1 - 4p,]1/2 therefore implies that Smax = 4/A.
The assumption A = 4 implies S ::; 1. •
From these propositions, one concludes that the study of the bifurcations
of the map P(Xt, p,) is possible within the range (1/4) ::; s ::; 1.
The approach above can still be used to reformulate some known re-
sults from bifurcation theory. More specifically, the functions Gk(p,), k =
1,2,3, ... are needed to show the existence of a period-doubling bifurcation:
cycles of doubling periods appear as the saving rate parameter is modified.
Further, there also exists a period-halving bifurcation, that is a reduction
by half of the periodicity of cycles. This second phenomenon is impor-
tant because it helps understanding why the fluctuations usually said to
be chaotic are not without a "certain order". This "order" reflects a very
high sensitivity of the model to the values of its parameters: there are as
many attractors as there are possible values for the saving rate when this
parameter varies in the interval within which a period-halving bifurcation
occurs.

3.3 PERIOD-DOUBLING AND PERIOD-HALVING

BIFURCATIONS

In this section we use the approach presented in the proof of lemma 1

to compute the bifurcation points of the saving rate parameter. Following
Myberg's approach, we know that those points are simply the roots of the
polynomials in p, obtained by computing the iterates of the map P(Xt, p,).

Period-doubling bifurcation

General formulation

From the proof of lemma 1 we know that

(40)
 39

This relation holds for any k, especially when k = 2m , m = 1,2,3, ....

We are looking for values of JL which satisfy the equality Gk(JL) = 0 with
G1(JL) = JL.

Example

If, k=2, then the equality G 2(JL) = 0 implies that JL2 + JL = 0 and so
JL = 0 or JL = -1. This equality also implies -JL = (_JL)1/2 and thus both
solutions can be considered as the fixed points of the first-order difference
equation -JLHI = (-JLi)1/2. We only use the first solution, which implies
s = (lIA[l + vis].

Results and comments

Repeating this for all subsequent values of k leads to the usual pitchfork
bifurcation, which is characterized by a limit value s2 towards which all the
successive values of the saving rate converge. Table 1 indicates the order of
the stable limit-cycles corresponding to the values of the saving rate. The
cumulation point is s=s2=0.7246620535.

k 21 2:.1 23 24
S 0.6035533835 0.70019511271 0.7194537996 0.7235359939
k 25 21> 27 21!S
S 0.7244083315 0.7245950748 0.7246350677 0.7246436295
k 29 210 ... 200
s 0.7246454651 0.7246458627 ... 0.7246620535
Table 1. Savmg rate values and penod-doubhng bIfurcatIOn

Period-halving bifurcation

General formulation
The values of the saving rate which bring such a bifurcation are obtained
by seeking the values of JL for which the 2k - 1 maximums and minimums
of the kth iterates of the map P(Xt, JL) become identical to the 2k - 1 fixed
points of its (k - l)th iterates. The procedure requires three steps.
40

First step
One writes down the equations which give the abscissa of the fixed point
of the kth iterate of the map P(Xt,f.l)' These points correspond to the 2k
fixed points of a k-period cycle:

(41)

Second step
It is interesting then to note that the iterate of the map P(Xt, f.l) admits
2k - 1 critical points (that is 2k - 1 points for which the first derivative of
P( Xt, f.l) equals zero). Noting their respective ordinates as Xc, XC" XC2 , ... ,
one assumes that :

or:

Third step
Finally, substituting XCk +2(f.l) = GCk +2(f.l) for Xt in equation (41), one
gets:

(44)
For each k, there is one value amongst all the values of f.l that verify
equation (44), which leads to a stable cycle with an even periodicity.

Example
If, we suppose that Xa is a point of a two-period cycle, then we have
Xa = p2(xa, f.l) (step 1). Assuming that Xa is also a fixed point of P(Xt, f.l),
we write Xa = P( Xa , f.l). Thus, by substituting in the left-hand side of the
first equation Xa with its expression given in the last equation, one gets
P(xa,f.l) = P 2(x a ,f.l)' This equality holds for Xa = 0.866120202. Assuming
that Xa can be written as f.l~ + f.lu (second step), we substitute this value
in the equation P(xa, f.l) = p2(xa, f.l) (third step) and get:
 41

Results and comments

In applying these three steps, one obtains the following values for
the saving rate (see table 2), to which correspond stable cycles whose
periodicity is decreased by half as one increases the parameter s.

k 200 ... 212 211

8 0.7246620535 ... 0.7249952526 0.726289064
k ... 211 27 ...
8 ... 0.7322752378 0.7608547029 ...
k 24 2" 2" 21
8 0.7246350696 0.7246436313 0.7246454656 0.7246458627
Table 2. Savmg rate values and penod-halvmg blfurcatlOn

Theses results enable us to draw the following conclusions.

(1) One notices that the periodicity of cycles associated with the different
values of the saving rate is reduced by half as the value of the parameter
is increased. This is why one speaks of a "period- halving" bifurcation,
by analogy with the "period-doubling" bifurcation where the periodicity
of cycles endlessly doubles. For example, when s=0.7249952526, a stable
limit-cycle of order k = 212 coexists with an infinitely number of unstable
limit-cycles of order higher than 212. When s=0.726289064, then the 212th -
order limit-cycle becomes unstable, and a stable limit-cycle of order 212 /2
is produced. And so on.
(2) S = 0.7246620535 appears as a limit "to the right" of the successive
values of the saving rate (that is 8 < 82), just as it appeared as a limit
"to the left" for the period-doubling bifurcation (that is 82 < 8). It is
noteworthy that the existence of a period-halving bifurcation is no way
incompatible with the fact that aperiodic fluctuations can take place.
Indeed, when s takes the values given in table 2, stable limit-cycles of low-
order exist but they are viewed as being" noisy" . Such stable periodic cycles
may be impossible to determine by visual inspection because they overlap
with high-order limit-cycles that are unstable. Suppose for example that
the saving rate s is 0.7246454656. We know from table 2 that in this case a
limit-cycle of order 24 exists and is stable. But there also exists an infinitely
number of cycles of order higher than 24 which are unstable. So, unless the
initial values of the capital-labor ratio are located on the stable attractor,
the sequence (kt)can switch, during a long time, from trajectories belonging
to unstable high-order limit-cycles before eventually converging towards the
stable 24th order stable limit cycle. this causes fluctuations to be apparently
42

aperiodic. In a word, very complicated dynamics corresponding to very long

transients may not be distinguished from true chaotic motion, in the region
corresponding to a period-halving bifurcation.
(3) The coexistence of both types of bifurcations allows the construction
of an ordering of all the bifurcations points.To show this, we need a few
notations. First, the bifurcation points of the saving rate which lead to
period-doubling cycles are noted (82'), i = 1,2,3, ... Secondly, the saving
rate values for which there exists a period-halving bifurcation are noted
(82')*,i = 1,2,3, ... Thirdly, for each i, n2, will designate the interval
[(82i), (82i)*]. In the light of what has been previously indicated, we have
the following relation:

(46)

(4) The procedure used above can be generalized to any rth iterate
pr(Xt,jL) of the function P(Xt,jL) , where r = m x 2i, m=3,4,5, ... , and
i=1,2,3, ... Suppose that (8;:X2') and (8;:X2')* are bifurcation points of the
saving rates which allow to define both period-doubling and period-halving
bifurcations (just as (82') and (82')* were bifurcation points for P(Xt,jL)).
One interesting point is that we know from bifurcation analysis, that cycles
of period m x 2i exist for values of the bifurcation parameter belonging
to a certain range of values leading to period-halving cycles. Here, this
would implies that there exists values (8;:X2') and (8;:X2')* which lie in
the interval [( 82' ), (82i)*] for each i. These values lead to another ordering
of the saving rate values:

.,. c ... c nmx2'+2 C nmx2 i+l c ... (47)

We therefore write:

(48)
The existence of a period-doubling bifurcation thereby implies that
aperiodic fluctuations when they exist are not without a "certain order".
This "order" reflects a strong sensitivity of the model to the saving rate
values.

3.4 CHAOTIC FLUCTUATIONS AND CYCLES WITH ODD

PERIODS

The Li and Yorke's [11] theorem used by Day in his paper, which has been
generalized by Li et al. [12], says that when odd-period cycles exist chaotic
fluctuations occur. Without going back on what has already been said on
 43

this point in the literature, we just seek to draw the reader's attention on the
following point. One can construct for odd-period cycles the same type of
structure and ordering of the saving rate parameter, as the ones previously
described for cycles with even periodicity. In this view, we consider here the
period-3 cycle, which corresponds to the odd-period cycle with the highest
order in Sarkovskii's [13J relation. Since the methods are similar to those
described above, we avoid technical details and expose directly the main
results.

Period-3 cycle implies a period-doubling bifurcation: main results

Let us define the interval [(8 3), (8 3)*], where (83) and (8 3)* correspond
for p 3(Xt, /1) = P[P[P(Xt, /1)]] to (8 2') and (8 2')* for p2' (Xt, /1) in the
preceding paragraph. We thus decompose the expression of the successive
iterates of the map P3(Xt, /1) into two polynomials in Xt and /1, which are
respectively noted <P~(Xt) and (h(/1). The bifurcation points are solutions
of the following equations (the approach is the same as above to obtain the
period-doubling bifurcation):

v v v 2 v

G k (/1) = 0, G k (/1) = [G k - 1 (/1)J +/1, G k (/1) = /1, k = 3x2m, m = 1,2,3, ...

(49)
We have reported in the table below, the values of the saving rate
corresponding to (49).

k 3 x 2u 3 X 21 3 X22 3 x 2;$
S 0.810 1064046 0.8141127724 0.8153023078 0.8156292668
k 3 x 25 3 X 26 .,. 3 x 200
s 0.8156655041 0.8156854345 ... 0.8156854472
Table 3. Savmg rate values and penod-3 doubhng bifurcatIOn

As one can see, there exists an infinitely number of (3 x 2i )-period cycles

with periodicity endlessly doubled until one reaches the cumulation point
s=0.8156854472.

Period-3 cycle implies a period-halving bifurcation: main results

Similarly, one can apply to the map p 3 (Xt, /1) the arguments held above
to define the period-halving bifurcation associated to P(Xt, /1). We have
reported in the table below the different values of the saving rate for
44

which successive stable limit-cycles exist and belong to a period-halving

bifurcation. As one can see, s=0.8156854472 is again a cumulation point.

k 3x2°O ... 3 x 2° 3 X 25
S 0.8156854472 ... 0.81568544722 0.8156854726
k 3 x 24 3 x 2;5 3 x 2~ 3 X 21
S 0.8156884065 0.8157308581 0.8157556535 0.8161786889
Table 4. Savmg rate values and penod-3 halvmg blfurcatIOn

Just as was previously seen for the cycles with even periodicity, we can
construct a similar ordering here for the period-3 cycle. This ordering leads
to the following relation:

(50)

As a conclusion, it is seen that both period-halving and period-doubling

bifurcations are "universal". Universality expresses that these phenomenon
are independent of the periodicity (either even or odd) of the cycles which
are successive stable attractors as the saving rate parameter is modified.
This finding might be surprising as compared with the results usually
mentioned in the literature, which focus on the property of sensitivity to
initial conditions. However, the mechanisms that leads to chaotic dynamics
exhibit a strong structural instability of the models under study, and this
can be carefully checked by using the procedures described above. Of
course, the arguments exposed can be generalize to all odd-period cycles.

4 Concluding remarks

This paper intended to discuss two aspects of Day's original contribution

on neoclassical one-sector growth and chaotic dynamics. Our arguments can
be summarized in two parts.
First, it is usually argued that Day's model introduces modifications of
the usual production function in one-sector growth models, which are not
economically motivated. Because of this, one may argue that there is no
reason why the monotone behavior of standard neoclassical growth model
should be transformed into erratic fluctuations. Further, this may lead to
cast some doubts on all the models that are based on the same method-
ology. In this note we have tried to show that there may be however a
possibility to rationalize the notion of a "pollution effect" by considering
 45

the macroeconomic consequences of microeconomic interactions. The econ-

omy is composed of many firms, which adjust their production and take
their decisions of accumulating capital, according to the decisions taken
by their competitors. Their mimetic behavior is due to the fact that firms
are individually confronted to imperfect information concerning the level of
future demand. We introduce here a formal expression of the way firms co-
ordinate their decisions, and it has been introduced here to "grasp" the role
of the" pollution effect" which holds such a decisive place in Day's model.
In this context, a chaotic dynamic may reflects the fact that volatile expec-
tations of agents are a dominant cause of the irregularity of macroeconomic
fluctuations.
Secondly, we have tried to show that the monotone evolution of the
capital-labor ratio which is usually observed in standard one-sector growth
models may become complex because Day's model produce a strong
structural instability: slightly modifications of the saving rate lead to a
qualitatively different behavior of the solutions. An important outcome of
this, is that the economy may undergo aperiodic fluctuations corresponding
to long transients, and not to true chaos. We have exposed a simple method,
which helps to highlight the mains causes of this structural instability, that
is the overlapping of period-doubling and period-having bifurcations.

References
[1] Day, R., and Gang Z., (1994), "Infrastructure, Restricted Factor
Substitution and Economic Growth", Journal of Economic Behavior
and Organization, 23, pp.149-166.
[2] Day, R., (1995), "L'existence hors de l'equilibre", Revue Economique,
46, pp.1461-1471.
[3] Nishimura, K. and Yano, M., (1995), "Non-linear Dynamics and Chaos
in Optimal Growth: An Example", Econometrica, 63, pp.981-1001.
[4] Grandmont, J.M., Pintus, P., and, De Vilder R., (1996), "Capital-
Labor substitution and Competitive Nonlinear Endogenous Business
Cycles", Cepremap working paper, forthcoming.
[5] Day, R., (1982), "Irregular Growth Cycles", American Economic Re-
view, 72, pp.406-414.
[6] Lorenz, H.W., (1993), Nonlinear Dynamical Economics and Chaotic
Motion, Springer-Verlag, Heidelberg.
46

[7] Albin, P., (1987), "Microeconomic Foundations of Cyclical Irregulari-

ties or Chaos", Mathematical Social Sciences, 13, 183-214.
[8] Day, R., (1986), "Unscrambling the Concept of Chaos through Thick
and Thin: Reply", Quarterly Journal of Economics, 101, pp.425-426
[9] Melese, F., and Transue, W., (1986), "Unscrambling Chaos Through
Thick and Thin", Quarterly Journal of Economics, 101, pp.419-423.
[10] Myberg, P.J., "sur k'iteration des polynomes reels quadratiques",
Journal de Mathematiques Pures Appliquees, 41, 339-351.
[11] Li, T.Y., and, Yorke, J.A., (1975), "Period Three implies Chaos",
American Mathematical Monthly, 82, pp.985-992.
[12] Li, T.Y., Misiurewicz, M., Pianigiani, G., and Yorke, J.A., (1982),
"Odd Chaos", Physics Letters, 87A, pp.271-273.
[13J Sarkoskii, A.N., (1964), "Coexistence of Cycles of a Continuous Map of
a Line into Itself", Ukranichkii Matematicheskii Zhurnal, 16; pp.61-71
 Part 2

Business Cycles
Hysteresis Arising from
Individual Inertia and
Behavorial Heterogeneity
Reiner FRANKE

1 Introduction

Appeal to the concept of hysteresis has become widespread in explana-

tions of a variety of economic phenomena. Two major fields are industrial
economics, with special reference to international trade, and labour eco-
nomics, with special reference to unemployment.} Generally, hysteresis is a
property of dynamic systems. Hysteresis exists when the long-run or final
value of a variable not only depends on the long-run values of exogenous
variables but also on the system's previous outcomes. This path-dependence
implies that different initial conditions, or different exogenous shocks, may
lead to different long-run or final configurations. A handy formulation is
that "history matters" .
In formal models, and particularly in empirical work, economists have
tended to view hysteresis as a special case of linear systems of equations.
Thus in discrete-time models hysteresis is constituted by the presence of
a unit root. Most convenient are stochastic models that are reducible to
one state variable, so that hysteresis can be identified with a random walk.
These features are usually achieved by specific assumptions about (perhaps
lagged or partial) adjustments of some macroeconomic key variables. 2
Seeking for a more fundamental cause of hysteresis, another direction of
research starts out directly from the micro level. It abandons the concept
of the representative agent and considers heterogeneous agents with dis-
continuous, lumpy adjustments. If there are many small decision units and
agents are sufficiently dissimilar, the aggregate variables will nevertheless

ISee Franz (1990) for a brief overview.

2The influential papers by Blanchard and Summers (1986,1987) on the
persistence of high unemployment in Europe during the 1980s are only one
example of how a random walk is constructed from a few number of simple
macroeconomic relationships.
50

change in a smooth way. Owing to the individual inertia, however, past

decisions and how they were distributed within the population of agents
may still leave their mark on the present. So the whole history of non-
synchronized adjustments could, possibly, affect the paths taken by the
macro variables. This point has recently been worked out by Cross (1994)
in a formal economic model where, on the basis of certain threshold values,
agents choose between two polar types of investment strategies in response
to the motions of an exogenous expected payoff variable. The modelling
device was in fact carried over from the physical sciences and had only to
be translated into economic terms.
The approach provides a useful framework to study the micro-macro
relationship and its hysteretic potential. An imperative extension is the
introduction of a feedback mechanism from the aggregate outcomes of
the agents' decisions back to their expectations about the payoff. This
task is the subject of the present contribution. The endogeneization of the
expected payoff closes the model and gives rise to the notion of equilibrium,
i.e., an infinite number of states of the economy where, if unperturbed,
the dynamic process essentially comes to rest. In our discussion these are
actually steady state positions, with different rates of growth induced by
different compositions of the population. Interesting features of hysteresis
derive from the fact that, in general, the composition depends on the history
of the economy, while, on the other hand, the system's memory is selective
and certain periods of the history may also be forgotten.
On the whole, the model is a deterministic nonlinear dynamic system
expressed in discrete time (with a flexible adjustment period that may be
arbitrarily short). Due to the non-standard description of the population
composition of heterogeneous agents it is no ordinary difference equations
system, so that the usual tools of a mathematical treatment do not apply.
Important points can still be made by means of geometrical arguments.
All statements and conjectures are furthermore substantiated by numeri-
cal simulations. Economically, the model is a highly stylized representation
of the linkages between the real and financial sector. In the present spec-
ification, it incorporates a variety of the multiplier-accelerator mechanism
and, with respect to investment behaviour, might be said to have a weak
Minskian flavour. 3 It should be added that the paper is exclusively con-
cerned with theoretical issues of hysteresis; there is still a long way to go
until this kind of research leads to empirical results.
The remainder of the paper is organized as follows. The next section
puts the model into perspective. Together with a general characterization

3The formulation of the dynamic adjustment equations that close the model
involves more feedback effects than would perhaps be necessary to put forward
our main points. The advantage of the more elaborate model is that it gives rise
to smooth, and in a a sense even realistic time paths - although at the present
stage of research we do not wish to overemphasize such a step towards calibration.
 51

of the structural features, it also anticipates the main results of our

investigation. The formal elaboration begins with the subsequent sections.
Section 3 introduces the basic concept of individual inertia and behavioural
heterogeneity. In particular, it describes how agents react to changes in
the expected payoff and how exogenous variations in the payoff affect
the population composition. This section is a short recapitulation of the
framework developed by Cross (1994), though it is occasionally phrased in
somewhat different terms with a view to the subsequent extension. Section
4 endogenizes the expected payoff and formulates the full model. Sections 5
and 6 study the hysteresis effects that may then emerge. In Section 5, a set
of parameters is assumed which ensures that the economy always converges
directly to some growth equilibrium. Oscillatory motions are dealt with in
Section 6, where both may happen: dampened oscillations (in the relative
magnitudes) or sustained endogenous growth cycles. Section 7 concludes.
A number of finer details are relegated to an appendix.

2 Perspective of the Model

A basic feature of hysteresis in closed models is brought out more clearly

if we neglect repeated random shocks and consider ordinary deterministic
systems of difference equations. They may also be nonlinear. A unit root
occurs if they possess a continuum of stationary states, i.e., one eigenvalue
of the Jacobian matrices evaluated at these points is unity (in continuous-
time systems, the Jacobians have a zero root). If it is supposed that the
system is n-dimensional, if all the matrices have exactly one unit root,
and the (one-dimensional) equilibrium set is attractive, then each of the
equilibria is associated with a stable manifold of dimension n -1, which
together fill out the whole basin of attraction. Conversely, to every point
in this basin, which may be a starting point of the dynamic process, there
corresponds one stable manifold and one state of equilibrium to which the
economy will converge. 4 Stochastic shocks can thus be seen as carrying the
economy from one stable manifold onto another. It goes without saying
that the set of equilibria is predetermined, i.e., independent of any initial
condition or perturbation. In most applications it is also quite easy to
compute.
The situation in the model here advanced is more elaborate. Though
the concept of heterogeneous agents and the component of inertia in their
behaviour is fairly elementary and the population composition admits a

4 An analysis of such deterministic systems can be found in Giavazzi and

Wyplosz (1985) for the linear case or, spelling out the nonlinear subtleties, in
Franke (1987, pp.77-92).
52

unique description by a finite-dimensional vector v, the problem is that

in general the dimension of this vector is not bounded: for each integer
number N a composition C of the population can be constructed such that
the dimension of the associated vector v = v(C) exceeds N. The changes
in the population composition as the dynamic process unfolds may also act
on the dimension of v, which may decrease as well as increase. The fact
that the composition C can be arbitrarily complicated and the dimension
of v( C) arbitrarily large entails that the set of all the growth equilibria that
are a priori possible has a more intricate, even unwieldy structure than the
one-dimensional manifold above. Furthermore, the usual investigations of
local stability do not work any longer. So the equilibrium set continues to
be predetermined, but this fails to be of much use for the analysis.
The concrete analysis of the model is concerned with two topics, where
all statements are illustrated by numerical simulations. In the first stage,
it is assumed that the economy has settled down in a steady state and
that then a one-time shock arrives. We actually concentrate on shocks to
a more or less psychological variable characterizing the financial sector,
which is called financial distress. The possible reactions of the economy fall
into four categories: the economy may find back to the old equilibrium in
a direct way, it may converge to a new equilibrium in a direct way, it may
oscillate towards a new equilibrium position, or the oscillations do not die
out and the economy approaches a periodic orbit, i.e., a growth cycle. An
important finding is that the shocks have asymmetric effects. In the case
of direct convergence, for example, a small or medium negative shock may
increase the growth rate in the resulting new equilibrium, whereas after a
positive shock of the same size the economy returns to its original position.
Another phenomenon, which occurs under cyclical adjustments, is that
both positive and negative shocks induce convergence to new equilibria,
and all of them exhibit either a lower or a higher steady state growth rate.
Which of the two cases prevails depends on the attributes of the population
composition in the initial equilibrium position.
If the system does not approach an equilibrium but enters into persistent
fluctuations, the strength of the shock may affect the amplitude of the
periodic motions. Generalizing the notion of hysteresis discussed above,
which was related to the points of rest of a dynamic system, these effects
may be called hysteresis in closed orbits, or cyclical hysteresis for short.
As regards the stability of steady states, a phenomenon can emerge that,
to our knowledge, has not been observed before in dynamic modelling. The
initial equilibrium may be attractive if the shocks to financial distress are
sufficiently small, while the economy displays sustained cyclical growth if
the shocks become larger. So far, this is a situation which, according to
Leijonhufvud's expression, is known as corridor stability. A new peculiarity
is that when the shocks are further increased in size, there is another switch
in the system's qualitative behaviour and the economy follows a course back
to a steady state position again, though growth takes place at a different
 53

level. Reserving the term stability to indicate convergence to a single state

of equilibrium, it might be said in a pointed manner that medium-sized
shocks are detrimental to stability, whereas large shocks are stabilizing.
In a second stage of the analysis, we revert to the issue of a random walk
as a reduced form in which hysteresis is often seen to manifest itself. To this
end the system is subjected to repeated random shocks at regular intervals.
If the economy always converges to an equilibrium, an old one or a new
one, and the shocks occur sufficiently infrequently, one can focus attention
on the associated growth rates and how they change in a stochastic
manner. Formally, the sequence of these steady state growth rates can
be viewed as a random walk where, however, the random innovations are
no longer independently and identically distributed. We would thus obtain
a 'perturbed' random walk, with induced innovations. Nevertheless, if the
law of motion underlying the time series of the equilibrium growth rates is
not known, the hypothesis of a pure random walk might still be considered
a useful approximation. This is, after all, the basis of the broad literature
on a unit root in GNP.
By studying the distribution and the time series characteristics of the
innovations in the growth rates and comparing them to the original shocks,
it is possible to get some information on the goodness of this approximation,
whether it proves workable or not. We do not provide an extensive Monte
Carlo simulation, but the few sketches of our realizations of the stochastic
experiment are sufficient to indicate that the 'perturbations' of the random
walk may not be negligible.
A typical feature of a time series generated by a pure random walk is
that over some passages of time it hovers around a certain value. Of course,
these levels will shift from one such phase to another. In our context
one may look upon them as different growth 'regimes'. A less rigorous
conception of hysteresis might content itself with a visual inspection of the
time series and an assessment of different growth regimes. This point of view
needs no theoretical background and, in particular, is applicable whether
the economy converges to steady growth paths or produces undampened
oscillations. These growth regimes can also be observed in our experiments
in the case of point-convergence. In the case of sustained endogenous
growth cycles, by contrast, the significance of hysteresis in this sense seems
to vanish. Though the shocks modify the amplitudes of the oscillations,
the effects do not visibly carryover to the time averages. Accordingly,
these economies could even be classified as trend-stationary. We have
therefore a theoretical relationship between hysteresis in growth regimes
and the (paradigmatic) problem whether the observed cyclical behaviour is
basically endogenous or brought about by exogenous shocks to the economy.
54

3 The Basic Description of Changes in the

Population

We distinguish two polar types of strategies adopted by agents. They may

be called prudent and bold, where in the present section the corresponding
actions can be left unspecified. Agents are not stuck in such one type
but decide on the basis of an expected payoff, that is, they change
their strategies if the payoff passes certain threshold values. In detail,
an individual agent is characterized by two critical values A and J.L of
the expected payoff, with A ~ J.L. If this agent is presently prudent, a
payoff 11" ~ J.L is required to induce her/him to switch to the bold type
of behaviour. A switch back from bold to prudent is made when the payoff
falls down to, or below, the threshold value A. If one wishes to invoke
optimizing behaviour, the two strategies might be viewed as corresponding
to the outcome of an optimal control problem for whose solution, in
the presence of certain adjustment costs, the bang-bang principle applies.
Examples are two-sided (8, s) rules (e.g., Miller and Orr 1966) or Dixit's
(1992,1995) argument that, under irreversible investment, waiting for more
information to arrive has a positive value which leads to discontinuities
in the investment decisions. These results may furthermore corroborate
discontinuous adjustment rules as a reasonable policy if agents are faced
with irreducible uncertainty.
The agent in our setting has a specific range of the payoff 11", A ~ 11" ~ J.L,
inside which it is necessary to know the history of the payoff movements in
order to determine which type of strategies is chosen. The switching rules
can be illustrated by the rectangular loop in an input-output diagram; see
Figure 1 (the payoff 11" is the input, bold or prudent behaviour is the output).
As 11" is continuously rising, the ascending branch abcde is followed, while
the descending branch edfba is traced when 11" is continuously decreasing. 5

SIn the language of the physical sciences, these rules define an elementary
hysteresis operator with local memory; cf. Mayergoyz (1991), pp. xiv, 2.
 55

f d e
bold ••

prudent •• 1 t
a b c

Figure 1 Switching rules of an individual agent

Generally, individual agents have different switching values. Reasons may

be different attitudes towards risk, or different technological and institu-
tional characteristics that may limit the flexibility to change strategies.
This is, however, the only dimension of behaviour in which agents are sup-
posed to differ. In particular, they share the same expectations about the
relevant payoff. Accordingly, behavioural heterogeneity is constituted by
the distribution of the pairs of switching values (A, JL) across the popula-
tion of agents. If A and JL are both contained in an interval [Amin, JLmax],
we have a one-to-one correspondence between individual agents and pairs
(A, JL) in the triangle
~ .- {(A,JL): Amin S A S JL S JLmax}
To ease the exposition, let us work with a continuum of agents and assume
that their characteristics (A, JL) are uniformly distributed on ~. The latter
supposition will also make clear that no special clustering of agents is
needed to obtain the hysteresis effects to be investigated below.
As will be seen shortly, at any instant of time the triangle ~ is subdivided
into two regions representing the agents who have presently chosen the bold
and prudent strategies, respectively. The boundary between the two sets
is formed by a staircase line as exemplified in· Figure 2. This line, which is
called the interface, is defined by finitely many vertices, say there are m of
them,
k =0,1, ... ,m
They may be ordered as

(Ao,JLo) = (Amin,JLmax) and Ak > Ak-l, JLk S JLk-l for k = 1, ... ,m

56

There is, however, only one exception, namely for m = 1 and Jl-o = Jl-1 =
Jl-max, where an inequality in the Jl--coordinates might not be strict. The
final stair is given by the vertical line connecting vertex Vm = (Am' Jl-m)
with the hypotenuse of the triangle ~ if Am < Jl-m; otherwise Vm is itself a
point on the hypotenuse. The entire interface, i.e. the set of these vertices,
is denoted by V,

Agents with the lower switching values, (\ Jl-) below the interface, are bold,
while switching values (A, Jl-) in the complementary region above V indicate
that such an agent has opted for the prudent set of strategies. Hence, by
virtue of the hypothesis of a uniform distribution of agents, the proportion
of bold agents in the total population is simply determined by the area
below the staircase line, divided by the total area of the triangle ~.6 In
explicit terms, let Tk = Tk(V) be the area of the trapezoid defined by the
vertices Vk- 1 and Vk. That is, starting from Vk = (Ak, Jl-k) in clockwise
orientation (k=2 in Figure 2), Tk is bounded by the four points (Ak,Jl-k),
(Ak, Ak), (Ak-1, Ak-1), (Ak-1, Jl-k), so that

If the final vertex Vm comes to lie on the hypotenuse of ~, .Am = Jl-m, the
trapezoid Tm degenerates to a triangle. Writing T(~) for the area of the
limiting triangle ~ itself, one has

The share of bold agents in the total population, which we designate b, can
therefore be represented by the mapping B = B(V),

1
L
m
b = B(V):= T(~) Tk(V) (1)
k=l

where the interface V is historically given.

6In the general case, the distribution of agents may be represented by a density
(or weight) function on fl.. Calculating the proportion of bold agents then involves
the evaluation of an integral, similar as in Mayergoyz (1991), p. 2.
 57

f.l

prudent

Figure 2 The interface V = {Va, VI, V2 , V3 }

" I
1f 1f 1fo

Figure 3 A decrease in th payoff down do nil, wiping out vertex V2

58

7
/'

V"4

7ro

, "
7T' 'If

Figure 4 An increase in the payoff up to 7r", wiping out vertex Va

To start the discussion of how variations in the payoff 7r affect the

population composition of bold and prudent agents, consider Figure 3. The
interface V = {Va, Vi, V2, Va} corresponds to the present payoff 7r = 7ra.
Now, let 7r decrease to 7r', which should diminish the share of bold agents.
In fact, this change induces the bold agents with a 'downward' switching
value A between 7ra and 7r' to return to prudent behaviour. Their 'upward'
switching value J.L is less than J.L3' Geometrically, these are the agents with
switching values (A, J.L) contained in the trapezoid bounded by V3 and Va.
In other words, the decline in the payoff shifts the vertical line to the
left which links the vertex V3 and the hypotenuse of 1:1. So the inital
interface V is modified to V' = {Va, VI, '\12, V3}. The (normalized) area
below V' represents the corresponding new proportion of bold agents, i.e.,
b = B(V'}.7
Next, suppose that the payoff continues to decrease to 7r" < 7r', which

7 Of course, the argument equally applies if initially all agents are bold
because of 7r 0 ~ J.t,.,.o.a:. The corresponding interface V in this case is just given
by the top line of the triangle 1:1, the vertices being Vo = (>'min, J.t,.,.o.a:) and
VI = (J.tmaa:, J.tmo.a: ); cf. the remark on the ordering of the vertices.
 59

moves the vertex V; further to the left. At the moment, however, when
the payoff reaches >'2, where >'2 corresponds to the vertex V2, there is a
sudden disproportional increase in the number of switching agents. Here
we also encounter bold agents with an upward switching value JL ~ JL3
(but JL ::; JL2), who now switch in their behaviour back to prudent. The
geometric implication is that as 7[ resumes its fall below >'2, the previous
vertex V2 is wiped out. From then on it is the vertical line connecting
V2 with the hypotenuse that shifts to the left, until 7[ = 7[". Writing
V~' = (A~,JL~) = (7[",JL2), the change in 7[ from 7[' to 7[" shapes the new
three-vertex interface V" = {Vo, Vl,V~'}. The elimination of vertices will
later be an important feature in the analysis of hysteresis effects.
In describing the effects of an increase in the payoff from the initial value
7[0' as sketched in Figure 4, a new vertex emerges. As a first step, define
V4 as the point V4 = (A4, JL4) = (7[0' 7[0)' Then as 7[ starts to rise, V4 moves
upwards along the hypotenuse of ~. At 7[ = 7[', say, prudent agents with an
upward switching value JL between 7[0 and 7[' have converted to boldness.
These are the agents with switching values (A, JL) in the triangle specified by
the corners (7[0' 7l'0), (7[0,7[') and (7[',7['). Setting V~ = (7[',7['), the interface
corresponding to the payoff 7[ = 7[' is given by V' = {VO , Vb V2, V3, VI}.
Continuing the increase in 7[ shifts the horizontal line with end-points
V~ and (7[0,7[') = (A3,7[') upwards (note that V3 = (A3,JL3) = (7[0,JL3)). As
7[ rises beyond JL3 we observe'a similar effect as above, namely, the vertex
% is wiped out. There are also prudent agents with a downward switching
value A less than 7[0 = >'3 (but exceeding A2) who adopt the bold strategies,
if JL3 ::; JL ::; 7[ with respect to their upward switching value JL. The payoff
7l' = 7[" therefore gives rise to the interface V" = {Vo, Vi, V2, V~'}, where
VI' = (7[",7[11). The vertices in the set V" would, of course, be renumbered
before the analysis were taken any further.
The rules determining the modification of the interface as the payoff 7[
changes may be compactly summarized by a mapping cP = cp(7[, V),

cp(7[, V) = interface V' resulting from (the old) interface V (2)

and (the new) payoff 7[

The terms old and new can be seen as referring to the previous and the
present period in a discrete-time setting. It is then understood that the old
payoff 7[0, in the previous period, constituted the A-coordinate of the final
vertex Vm , that is, Am = 7[0' The precise definition of the mapping cP in
algebraic terms is given in the Appendix.
We are now also in a position to reconstruct the motions of the payoff
that may have generated the interface V = {Vo, VI, V2, %} in Figure 2.
Initially, 7[ may have fallen short of Am in and all agents may have been
prudent. A steady rise in 7[ until 7l' = JLI and then down to 7[ = Al will have
produced vertex VI. Vertices V2 and V3 are the outcome of the subsequent
oscillations in 7[ with peaks JL2 and JL3 and the interposed trough A2. The
60

payoff is then presently at 7f = A3. 8 Note that the speed of these payoff
variations has no bearing on the shape of the interface.
To sum up the impact of fluctuations in the payoff 7f on the population
composition, the vertices of the interface V are brought into existence by
the turning points of these motions. Furthermore, each peak in 7f wipes out
the vertices of V whose p-coordinates are below this local maximum of the
time series, while each trough wipes out the vertices whose A-coordinates
are above this local minimum.
After this introduction into the working of the model, a basic hysteresis
effect can be revealed. The aggregate behaviour of agents in macroeconomic
theory is usually modelled in a way that assigns to every vector of inde-
pendent variables a uniquely determined outcome of the individual actions.
For example, the theory assigns to each set of prices a definite quantity of
a good demanded or supplied in the aggregate. If the independent variable
is the expected payoff 7f, then according to this supposition every value of
7f would be associated with a specific value b of the share of bold agents in
the total population (to which might correspond a certain level of aggre-
gate demand for physical capital goods or a financial asset). However, in
the present framework with the discontinuous adjustments of the microe-
conomic units such a functional relationship holds no longer true.
As a case in point, refer to Figure 4 and consider the share of bold agents
b that may be induced by the payoff 7f'. As discussed above for the upward
movement of the payoff, b was given by the (normalized) area below the
interface {Vo, VI, V2, V3, V1}. Now, suppose that the payoff continues to rise
until 7f" and subsequently returns to the previous value 7f'. Owing to the
inertia in the behaviour of the individual agents, not all agents having
switched from prudent to bold in the upturn of 7f have also switched back
to prudent during the subsequent descent. Geometrically, the agents who
have turned bold in response to the rise of 7f from 7f' to 7f" are represented
by the area between the two dashed lines in Figure 4, which is bounded by
the lines connecting the six points VI = (7f',7f'), VI' = (7f",7f"), (A2,7f"),
(A2, P3), V3 = (A3, P3), and (A3,7f'). From these bold agents only those
with switching values (A,p) in the triangle VI, V4 = (7f',7f"), VI' have
resumed the prudent behaviour as the payoff has fallen back to 7f'. This
argument shows that in order to determine the share of agents who, given
the expected payoff 7f, have adopted the set of bold strategies, one must
know the history of the time path of 7f. To be precise, not the history in
every detail but only the turning points of the payoff, and only those that
are not dominated by subsequent turning points, i.e., those that have not
been wiped out in the course of time.

8 Alternatively, V may be reconstructed from a 100% bold population and an

initial payoff 7r :::: J.'maz.
 61

[b. in Yo ]

60.0

50.0

40.0

.
30.0

2.00 4.00 6.00 ( ]

Figure 5 Hysteresis loops arising from a periodic motion of the payoff 7f (b

the share of bold agents)

We illustrate the hysteresis phenomenon in Figure 5. It is another input-

output diagram, but unlike. Figure 1 the output is here a continuous
variable, the share of bold agents b. Underlying are strictly periodic cycles
of the payoff. 9 The loop in Figure 5 is typical for the visualization of a
hysteresis operator and its branch-to-branch transition after input extrema
(see, e.g., the introductory section of Mayergoyz 1991). Additional branches
would be generated if the periodic pattern of the payoff were relaxed and
w exhibited different turning points. It may also be noted that information
about the present output b and all future values of the input 7f are not
sufficient to deduce the future evolution of b in a unique way. A given
pair b = b(t o) and 7f = 7f(to) at time to may be compatible with different
histories of the payoff, so that, depending on the shape of the interface
thus brought about, the same time path of w from to on may generate
quite different time paths of the population composition. 10

9The trough and peak values of w being given by w' and w", respectively, the
interface at the trough value was constituted by {(Ao, /-Lo), (w', w")}, and at the
peak value by {( Ao , /-Lo), err", w")}. With w on the downswing, the interface was
{(Ao, /-Lo), (w, w")}, on the upswing it was {(Ao, /-Lo), (-rr', 7f"), (7f, 7f)}.
lOIn the terms of the physical sciences, we have a hysteresis transducer
with nonlocal memory, constructed as a superposition of elementary hysteresis
nonlinearities with local memories (cf. Mayergoyz 1991, pp. xiv-xvi, 3).
62

4 A Complete Model with Endogenous Payoffs

In the previous section, the evolution of the expected payoff was treated
as exogenously given. We now extend the basic framework and introduce
a number of new variables and relationships such that the payoff 7r is
fully endogenized. The resulting dynamic model, if unperturbed, is able
to produce steady state growth paths as well as oscillatory motions.
The subsequent analysis is concerned with their potential to give rise to
additional hysteresis phenomena.
In the following the microeconomic units are conceived as heterogeneous
firms, which generally expand at different speeds. Since part of their fixed
capital investment must be externally financed, these firms are also faced
with the choice of adopting more or less risky financial postures. Let a
firm's financial practices be represented by a single variable called exposure.
A higher value of a firm's exposure may, for example, be associated with
higher borrowing relative to cash flow, or a greater proportion of short-
term debt in total liabilities. Limiting firms to two types of behaviour
in this respect, which we continue to identify as bold and prudent, the
corresponding levels of exposure are denoted by eb and ep (eb > ep ). They
are coupled with two distinct growth rates of the stock of physical capital,
9b and 9p (9b > 9p), to which the bold and prudent firms respectively
aspire. The switching of individual firms between the two types of policy
is governed by the same principle as above.
A higher value of exposure indicates greater vulnerability to disruptions
in the normal functioning of the financial sector. The financial market
conditions are captured by a measure of distress, d. High levels of distress
can be related to high interest rates and an unwillingness of banks to
refinance positions as the firms' debt payments become due. Low values of
d, on the other hand, are reflective of a state of tranquillity. Certainly, d has
also a psychological dimension. Financial distress will be a key determinant
of the expected payoffs to the two strategies of exposure.
Besides the differential payoff 7r and the interface V in the (A, JL)-plane,
other variables in our economy are e, aggregate exposure; gk, the growth
rate of the aggregate capital stock; gi, the growth rate of aggregate (net)
investment; and g, the growth rate of aggregate output which the firms
expect to prevail on average over the next few months or even years. With
a view to the ensuing computer simulations, the formulation is in discrete
time with a fixed adjustment period of length h,u At the beginning of the
present period t, the variables 7rt-h, Vt-h, gf-h' gt-h, dt - h are historically

llLet us say that time is measured in 'years'. The process of going to the
limit, h -+ 0, will mathematically be well-defined and would result in a piecewise
specified system of differential equations. That is, a 'regime switching' would
occur if a vertex of the interface is wiped out.
 63

given from the previous period. We first state the recursive adjustment
equations to determine, on this basis, the state of the economy in period
t, subsequently a few comments are added (further details can be found in
the appendix). With respect to positive constant adjustment parameters
(3k,{3g,{3d and three functions fy = fy(gi), fd = fd(e,g), f7r = f7r(d,g), the
dynamic system reads as follows:

lit cI>(7rt-h, lit-h) (3)

bt B(lIt) (4)
gf = k
gt-h + h {3k [bt gb + (1 - bt ) gp - gf-hl (5)
k
g; gt-h + (1 + hgth) (gf - gth) / hgth (6)
gt gt-h + h{3g [fy(gn - gt-hl (7)
et bt eb + (1 - bt ) ep (8)
dt dt-h + h{3d [fd(et,gt) - dt-hl (9)
7rt f7r(d t , gt) (10)

Equations (3) and (4), which determine the new interface and the resulting
share of bold firms, are known from Section 4. In the aggregate, firms
intend to expand their capital stocks at the rate btgb + (l-bt}gp. Equation
(5) acknowledges the financial and technical difficulties of adjusting actual
to desired capital stocks and supposes a partial adjustment mechanism
towards this target rate with coefficient {3k. Equation (6) is no more than
an identity for the growth rate of investment, the discrete-time analogue of
the equation i = k + gk in continuous time (which is obtained from the
logarithmic differentiation of the definition gk = k = 1/ K, where I is net
investment, K the capital stock, and the caret designates growth rates).
The term fy in equation (7) stands for the actual growth rate of aggregate
output. Based on the notion of a multiplier process, which moderates
in a depression or boom phase, f y is devised as an increasing S-shaped
function of the growth rate of investment (under 'normal' conditions, i.e.
over a medium range of gi, the two growth rates of output and investment
coincide). As firms are supposed to live in an uncertain environment and 9
is to represent output expectations, not for the next adjustment period, but
over a longer time horizon, we stipulate that 9 is revised only partially in the
direction of the currently observed growth rate fy. The emerging equation
(7) is formally equivalent to the specification of adaptive expectations. 12
Equation (8) defines aggregate exposure. In the determination of the

12 Arguments from the literature which defend the use of an adaptive expec-
tations mechanism in macroeconomic modelling are collected in Franke (1994).
Suffice it here to mention the contribution by Heiner (1988, especially p. 272),
who demonstrated that in an uncertain world with imperfect information, par-
tial adjustments indeed represent a reasonable mechanism of the revision of
expectations.
64

changes in financial distress d, note first that distress on the credit markets
depends on the (expected) cash flows of firms as well as their amount of
contractual payments. These aspects are here captured by the expected
macroeconomic growth rate 9 and aggregate exposure e. The function
fd = fd(e,g) in (9) represents the corresponding self-sustaining level of
distress; the notion is that this level would be reproduced over time if e and
9 remained constant. Clearly, 8fd/8e > 0, 8fd/8g < O. However, we assume
that it takes some time until the changes in e and 9 take full effect in new
loan contracts. Accordingly, equation (9) describes gradual adjustments of
the actual degree of financial distress towards its self-sustaining level fd,
with adjustment speed (3d .
Finally, equation (10) specifies the payoff expectations as being influ-
enced by the conditions firms find on the financial markets, and by the
prospects of future economic growth. The expected differential payoffs 11'
of bold strategies with their higher indebtedness are therefore negatively
affected by financial distress, while the more aggressive policy will yield
higher increases in profits than the prudent strategies when the economy
expands at a faster pace. Hence 8f7r/8d < 0 and 8f7r/8g > O.
Although equations (3-10) constitute a well-defined dynamic process
and the computations involved are straightforward, there is an important
point in which they differ from ordinary systems of difference equations.
The latter exhibit a definite number of state variables, which is the
dimension of the system. Because of the central role of the interface, it
is here impossible to determine a (finite) dimension. The interfaces may be
parametrized by their vertices, so that the dimension of a given interface
would be given by twice the number of its vertices. However, if the process
generates cyclical motions then, as seen in the preceding section, the
number of vertices of the interface may increase. The problem is that we
cannot tell in advance the number of vertices to be shaped in the course of
the process; it might even be infinitely many.
System (3 -10) gives rise to a notion of equilibrium. In a rest point, the
four growth rates will all be equal, 9 = y1< = gi = fy(gi). Expressing their
relation to the share of bold firms as g* = g*(b):= bgb + (1 - b)gp, and
similarly for exposure in (8), e = e(b), the corresponding level of financial
distress is obtained from (9) as d = fd(b):= fd(e(b), g*(b)). Inserting it in
function (10) yields the payoff

11'* = 71'*(b):= f7rIf;(b),g*(b)] (11)

On the other hand, equations (3) and (4) have to be taken into account. If
we now have a given interface V, this means that via the payoff 71'* and the
mapping 4>, the newly generated interface V* must be reproduced in the
equilibrium. That is, 4>(11'* (b), V*) = V* = 4>(1I'*(b), V). V* is reproduced
if (and only if) the payoff does not change or, what amounts to the same, if
the share of bold firms remains constant. A state of rest of the economy is
 65

therefore characterized by a solution of the following fixed-point equation

in b,

b = B[<I>(1I"*(b), V)] (12)

It can be shown that 11"* is a decreasing function of b if the growth rates

effects in the distress and payoff functions in (9) and (10) are of moderate
size (i.e., if the partial derivatives 181d/8gl and 811</8g are not too large),
a condition that will be met for all sets of numerical parameters considered
in this paper. From the discussion of Figures 3 and 4 we know that the area
B = B[<I>(1I", V)] decreases as 11" decreases. b furthermore exceeds (falls short
of) the right-hand side of (12) if b=l (if b=O). Hence there is exactly one
value of b satisfying equation (12). To sum up, for every interface V there
exists another, uniquely determined interface V* supporting an equilibrium
composition of the population of firms. We may succintly write

b* = b*(V) , equilibrium share of bold firms, (13)

wIth respect to an initially given interface V

Since the given interface V may have any shape, this argument demon-
strates that there exists a huge set of steady state positions, much 'larger'
indeed than a one-dimensional manifold. The interface V* = <I>(1I"*(b*), V)
itself can be viewed as constituting a reference equilibrium for the actual
economy in a state with interface V. V* will generally be shifted as V
changes in the course of the dynamic process.

5 Hysteresis Phenomena in the Closed Model

The presence of multiple equilibria has a bearing on the notion of

stability. While in a deterministic system with a unique equilibrium,
(asymptotic) stability means it is this specific state that attracts the
trajectories, it can now only be asked whether the economy will converge to
some state of equilibrium. Here it is the dynamic system as such that will
be called stable, with respect to a given initial condition. If convergence
obtains for all initial conditions, we will call the system universally stable.
Specifying an initial condition, which includes an initial interface V, it would
nevertheless be extremely hard to determine the equilibrium in advance
in which the economy will finally settle down, even more so when the
associated equilibrium share of bold firms, b* = b*(V) in terms of equation
(13), varies over time. Apart from that, it will be seen below that the system
might be stable only for a subset of starting positions.
Given an equilibrium interface with payoff 11"*, which connects at (A, J-L) =
(11"*,11"*) to the hypotenuse of the triangle ~, and an initial value 11"0 of the
66

payoff close to 7r*, a comparison of the trapezoids in Figures 3 and 4 shows

that the calculation of the area below the new interface brought about
by 7ro uses different expressions in equation (1), depending on whether
7ro > 7r* or 7ro < 7r*. Thus locally around every equilibrium interface there
are two distinct regimes of difference equations. This property of the system
precludes an application of the standard tools of stability analysis. Besides
some basic geometrical arguments in the (A, /L)-plane which help to classify
the main cases that may occur, we therefore have to resort to numerical
investigations.
In the following, we confine our interest to two parameters and their
impact on stability. They refer to the distress and the payoff function in
(9) and (10), which we now specify as
e-e
min{l,max{O,--P - Ct.dg + 'Yd}} (14)
eb - ep
O'"·[(I-d)(I+Ct.,,g) - d1/l] + 'Y" (15)

(Ct.d, Ct.", 'Yd, 'Y", 0'",1/1 are all constant parameters.) Function (14) limits
distress to values between zero and one, so that in the discussion below, d
and the shocks to d can be expressed in percentage terms. The intercept
'Yd in (14) serves to shift the initial equilibrium value of d to a convenient
level. Similarly, 0'" and 'Y" scale and shift the payoff function (15) such
that the model's initially given interface can bring about a particularly
nice equilibrium composition of firms, namely, a 50 per cent share of bold
firms, b~. The system always starts from this equilibrium configuration,
which is then disturbed by an exogenous shock to d.
The parameter 1/1 in (15) represents the differential distress losses of bold
firms if d is high. The effects of expected output growth, which reduce
financial distress and raise the payoffs to the bold strategies, are measured
by Ct.d and Ct.", respectively. It is these latter coefficients on which we
concentrate, where three pairs (Ct.", Ct.d) are considered. Somewhat loosely,
they may be characterized as follows.

Scenario 1: no growth rate effects (Ct." = 0, Ct.d = 0); universal stability,

where the (new) equilibrium is approached monotonically.

Scenario 2: moderate growth rate effects (Ct." = 4, Ct.d = 1); universal

stability, where the (new) equilibrium is approached in a cyclical
manner.

Scenario 3: strong growth rate effects (Ct." = 5, Ct.d = 2.5); no convergence,

but persistent growth cycles (for medium-sized shocks at least).

We are mostly concerned with the stable scenarios, as they can be em-
ployed to investigate a possible relationship of system (3 -10) to a random
walk process. The economy is correspondingly subjected to repeated shocks.
 67

In the present framework, the variables most prone to such exogenous per-
turbations are associated with the financial markets and their psychology,
which is our reason for focusing attention on shocks to the level of financial
distress, d. It is furthermore assumed that these shocks arrrive infrequently.
That is, after the economy has been thrown out of a state of rest by a shock
to d, it has sufficient time to converge close to a - possibly new - equilib-
rium until the next shock occurs. It may be conjectured that if the shocks
are i.i.d. with zero mean, then, though perhaps somewhat distorted, this
stochastic process has the basic features of a random walk.
To scrutinize this idea, let the economy start from an initial equilibrium
with a share of bold firms b~ and consider the sequence bt, b2,b3, . .. of
equilibrium shares of bold firms that is generated by a sequence of random
shocks to financial distress. The change from one equilibrium b"k to the next
b"k+I may be summarized by a function F = F(Ok' V*(k)), where Ok denotes
the shocks to d and V*(k) is the equilibrium interface associated with b"k.
Thus,
k = 0, 1,2, ... (16)
The expression F(Ok' V*(k)) being a random variable, equation (16) is
indeed reminiscent of a random walk. The problem to be studied is how the
random shocks Ok and their probability distribution are transformed by the
mapping F. To this end, we consider a given equilibrium interface V* and
inquire into the typical shape of the function 0 f--+ F(o, V*). Subsequently,
it has also to be indicated how this function is affected by a change in the
type of V*.
The initial equilibrium interface V* underlying our numerical examples
is similar to the interface V' = {Yo, VI, "2, V3, Vn in Figure 4. The
important feature to note is that it connects horizontally to the hypotenuse
of the limiting triangle 6.. Denote the corresponding equilibrium share of
bold firms by b'. Then, consider a one-time negative shock to d, which
instantaneously raises the payoff differential from its initial equilibrium
value rr' to rr". The resulting interface is V" = {Yo, VI, V2, V;'} (before
renumbering the vertices; the elimination of V3 is here inessential). It has
already been observed in the discussion of the loop-hysteresis at the end
of Section 3 that a fall of rr from rr" back to rr' does not re-establish the
previous population composition, i.e., rr' and, thus, b' can no longer produce
an equilibrium after the negative shock has materialized. In fact, the new
equilibrium payoff must be smaller, and the new equilibrium share of bold
firms must be larger (this is demonstrated in the Appendix).
Things are different if financial distress experiences a positive shock.
Returning to the initial equilibrium interface V' = {Yo, VI, V2, V3, Vn in
Figure 4, suppose the shock to d is limited in size, such that the resulting
fall in the payoff does not wipe out vertex V3 . For example, let the shock
give rise to V = {Yo, Vb V2, V3, V4 } as the new interface. In this case an
increase in rr back to the original level rr' does set up the previous share b'
68

of bold firms, and the initial interface V' continues to constitute a state of
equilibrium.
However, a new equilibrium comes into being if the positive shock to d
is so strong that the fall in 1f' eliminates a vertex (vertex Vs in Figure 4).
Similarly as before it can be shown that such an equilibrium is associated
with a smaller share of bold firms, while its interface connects horizontally
to the hypotenuse of !:l.. At least for a certain range of shocks, the decrease
in the equilibrium share brought about by a positive shock to d will be less
than the increase resulting from a negative shock of the same size. This
asymmetry in the equilibrium response is illustrated by the solid line in
Figure 6. In terms of the discussion above, this curve is the graph of the
function 61-+ b' +F(6, V') (b' und V' determining the original equilibrium).
The zero slope for extreme values of 6 is due to the fact that all shocks
6 ~ 49% induce the same degenerate interface, the instantaneous effect
being a reduction of the payoff to such a low level that all firms are prudent.
Likewise, all shocks 6 :5 - 29% lead to an increase in 1f' that temporarily
induces all firms to turn bold.
The numerical data underying the solid line in Figure 6 correspond to
Scenario 1 with its monotonic convergence. As a consequence, the new
equilibrium share will be actually reached by the dynamic process after a
one-time shock to d has occurred.

[ b J EQUllfbriu" . . . . . . . . .
70

GO
. -.... -.. _-_ .
..... .--------_.-".'
.-'
~-.-.­

SO

-:10 -20 -10 o 10 ao

Figure 6 New equilibrium share (in %) of bold firms arising from a shock
6 (in %) to financial distress; solid line represents Scenario 1, dashed line
Scenario 2.
 69

The case in which the original equilibrium interface has a vertical con-
nection to the hypotenuse of /). can be dealt with in an analogous way. The
function F(8, V') is again decreasing in 8, but now its constant segment
extends over an interval of negative shocks, and the slope is steeper for
positive 8. The qualitative reactions of the equilibrium share b* of bold
firms are summed up in Table 1.

Connection of Shock to financial distress d

original equil.
interface weakly strongly weakly strongly
to hypotenuse of /). positive positive negative negative
horizontal n.c. b* 1 (H) b* T (V) b* TT (V)
vertical b* 1 (H) b* 11 (H) n.c. b* T (V)
Table 1: Response of the equilibrium share b* of bold firms

Note: n.c. means no change, 1 (ll) indicates a moderate (strong)

reaction, letters H and V signify that the new equilibrium interface
connects horizontally or vertically, respectively, to the hypotenuse
of the limiting triangle fl. Shocks to d are classified as strong (or
weak) if they wipe out (or not) a vertex of the original equilibrium
interface.

With this analysis we are sufficiently equipped to return to the stochastic

process (16) describing the evolution of b*. Since the equilibrium interfaces
V*(k) change in the course of process (16), it is obvious that an assump-
tion of identically and independently distributed shocks 8k does not carry
over to the innovations in b'k. The, also qualitatively, different effects of
the 8k, depending on whether the V*(k) are vertically or horizontally con-
nected to the hypotenuse of /)., do not even allow the transformed shocks
F(8 k , V*(k)) to be approximately i.i.d. Furthermore, their frequency dis-
tribution obtained from a sample simulation of (16) would show little re-
semblance to the frequency distribution of the causal shocks 8k. This is,
in particular, due to the possibility that bounded shocks 8k leave the b'k
unaffected. Hence the frequency of values around zero will be considerably
higher for the induced shocks F(8 k , V*(k)) than for the 8k themselves.
The possibility of zero changes in b'k will also lead to more 'persistence'
in the time series of system (16). A typical feature of a random walk for a
variable Xk, say, are positive probabilities of successive perturbations in the
same direction, so that over certain passages of time x k is steadily increasing
or decreasing. In process (16), the effect of successive shocks 6k of the same
70

sign is much more limited, or there may be no effect at all. For example,
let the equilibrium interface V*(k) connect vertically to the hypotenuse of
~ and let a positive shock 8k occur, which leads to a (moderate) fall in
the share of bold firms, bk+1 < bk. Since according to Table 1, V*(k + 1)
has a horizontal connection to the hypotenuse, a subsequent sequence of
positive shocks, unless they wipe out a vertex, do not alter the equilibrium
composition of the population of firms. The share of bold firms would be
reduced if some of these shocks are stronger, but, as Figure 6 demonstrates,
not proportionately so. A stronger reaction in b* would only occur if the
shocks changed their sign. Besides, this fact suggests that the innovations
in bk are weakly negatively autocorrelated.
To illustrate these remarks, Scenario 1 of process (3-10) has been
simulated under the assumption that a shock to financial distress occurs
every 4 years. 13 This time span was long enough for the system to converge
close to the, old or new, equilibrium position. A sample of the resulting
time series of the aggregate capital growth rate gk is shown in the upper
panel of Figure 7. By virtue of the delayed adjustments of gk in equation
(5), this series is smoother than the series of the actual population share
bt or the stepwise movements of the equilibrium share b~. Three 'growth
regimes' can be clearly distinguished, though the numerical differences in
the growth rates are not too large. In the interval between (roughly) t= 15
and t = 50, the growth rate gk hovers around a trend value of about 3.8% (in
the initial equilibrium, gk=3.5% prevailed), in the time interval [105,130)
trend growth is slightly above 3.6%, and in the interval [140, 160) it is
slightly above 3.9% .

It is interesting to note that the pure random walk that is generated by

this sample of shocks: X4{k+l} = X4k -84k' exhibits a pattern which is rather
distinct from the capital growth rates. 14 In the sample here presented, the
x-series happens to perform a spurious oscillatory motion; from t = 16 until
t = 136 (i.e., k = 4 until k = 34), it is an only mildly perturbed trough-
to-trough movement, which at t = 76 and t = 92 attains two local peak
values of nearly the same height (they are also visible in the gk_series in
Figure 7, but there the two peaks appreciably differ). As a consequence,
the sequence of the shocks 84k with its 42 members displays a positive
first-order serial correlation of PI = 0.219. By contrast, the corresponding
sequence of innovations in the equilibrium share b* of bold firms has a
negative serial correlation, PI = -0.202, where 25 times the shocks 84k

13Each shock independently drawn from a normal distribution with a standard

deviation of 10% .
14The choice of the indices of x makes the time scale of process (3 -10), equation
(16), and the original random walk comparable; 64 1c is the shock to d at time t=4k
(k = 0,1,2, ... ) and the minus sign is employed to let x and b* move in the same
direction.
 71

have no impact at all. These observations may here suffice to point out
that a pure random walk has a significantly different character from the
hysteresis in the dynamics (3 -10), when, being universally stable with
monotonic convergence, this process is subjected to repeated infrequent
shocks.

Tune
Serles
of gk

Figure 7 Time series of the aggregate growth rate gk (in %) of fixed

capital, resulting from the same sequence of repeated random shocks in
these scenarios 81,82 and 83

6 Hysteresis in Cyclical Scenarios

In this section it will first be seen that even if the assumption of universal
stability is maintained, the consequences of a perturbation of the steady
state may be markedly different if convergence is cyclical. We may again
invoke Figure 4 and let a starting equilibrium interface have the shape of
V' = {Vo, VI, V2, V3 , Vn with its horizontal connection to the hypotenuse of
72

L).. Consider then a positive shock to d which drives the payoff from Tr' down
to Tro. The shock-induced interface being given by V = {Yo, VI, \12, V3, V4 },
it has already been noted that this does not change the equilibrium position.
If direct convergence prevails, the process would therefore end up in the
initial interface V'. Under cyclical stability, however, the upward motion of
the payoff will not come to a halt at Tr' but will overshoot this equilibrium
value. 15
The point is that once the payoff has passed Tr', the situation resembles
that of a negative shock occurring to the original steady state, i.e., from
then on V' no longer constitutes an equilibrium. Denoting the interface
arising from the payoff Trt at time t by V(t), we can utilize (essentially)
the same argument as in the previous section to infer an increment
in the corresponding equilibrium share of bold firms: Trt > Tr' implies
b*(V(t)) > b*(V'). Moreover, b*(V(t» increases as long as Trt continues
to rise. When Trt begins to fall, for example after a peak at Tr=Tr" in Figure
4 with the interface V" = {Yo, VI, \12, V~'}, the equilibrium share stays
constant: b*(V(t» = b*(V"). Some of the previous increase in b* is undone
after the next overshooting, when Trt passes the payoff level corresponding
to b*(V"). The equilibrium share starts to rise again after the payoff has
reached a lower turning point, etc., until these oscillations will have died
out. In contrast to direct convergence, where a positive shock to financial
distress has a zero or negative impact on the final level of b* , under cyclical
convergence this effect is positive.
AI; mentioned, this reasoning includes the case of a negative shock to
distress, so that it, too, raises the equilibrium share of bold firms in the
new state of rest. Varying the strength of the shocks to d in the numerical
example of our Scenario 2 and computing the resulting new equilibria yields
the second response function of b* in Figure 6, which is depicted as the
dashed line. The effects in the cyclical scenario depend heavily on the first
peak and trough values after the shock. Hence the fact that the reactions
to the negative shocks are not unambigously stronger or weaker than in
Scenario 1 must be ascribed to the interplay of the variables in the dynamic
process as a whole. Some explanations could be given for the specific case
under consideration, but they seem hard to generalize.
A mirror-image argument to the process above establishes that if the
initial equilibrium interface is vertically connected to the hypotenuse of L).,
then all shocks to d, again positive and negative shocks alike, reduce the
share of bold firms in the new equilibrium. While Figure 6 suggests that b*
is a strictly increasing (or decreasing, respectively) function of the size of
the shocks 8, there are also examples to be found where this relationship is

15 As may be expected from the weak amplifying forces under cyclical stability,
the downward jump in the payoff caused by the initial positive shock actually
initiates a further decrease in Tr for a short while, before the stabilizing mechanism
sets in.
 73

not monotonic; starting from an initial equilibrium with vertical connection

to the hypotenuse, b* may first decrease as 8 rises from zero, and then
increase again as the shocks become stronger. We nevertheless do not wish
to overemphasize this possibility. It is remarkable enough that the steady
state growth rate will react in the same direction, whether there were a
positive or negative shock to financial distress, and that the direction itself
depends on the history of the fluctuations of the economy, i.e., whether the
previous turning points have generated a vertical or horizontal connection
to the hypotenuse of Do. 16
After studying the equilibrium response, the system may be subjected
to the same sequence of financial shocks as in Scenario 1. Generally the
oscillations in Scenario 2 make convergence to an equilibrium slower than
under monotonic convergence. Since within the assumed four years between
two successive shocks the adjustment process still bridges most of the initial
gap between the actual and the equilibrium share of bold firms, this timing
of the shocks was maintained. The resulting time series of the capital growth
rate gk is shown in the middle panel of Figure 7. The three growth regimes
of the first scenario are still discernible. On the other hand, the shocks
produce some amplifying effects, though they are only temporary (note that
the scale is the same for the two panels of Scenario 1 and 2). The respective
average levels of the growth rates in the three regimes are barely affected,
either. In addition, the sequence of the innovations F(8 k , V*(k)) to the bt in
process (16) exhibits nearly the same order of negative serial correlation,
namely, PI = -0.219. To summarize, although the equilibrium response
functions are fairly dissimilar in Scenario 1 and 2, these differences tend to
be washed out in the time series of the variables generated by the repeated
random shocks. The main reason are, of course, the recurrent changes in the
type of the equilibrium interfaces, from a vertical to a horizontal connection
and vice versa.
A difference between the two stochastic processes of Scenario 1 and 2
is obtained in the frequency distribution of the innovations F(8 k , V*(k)).
Comparing the two equilibrium reactions in Figure 6, it is immediate that
in Scenario 2 zero changes, or nearly-zero changes for that matter, will
be observed less often. This implies a greater spread in the frequency
distribution around zero and, in a histogram, makes it look a bit more
like the distribution of the causal shocks 8k .
Another property of the frequency distributions arising from the two
scenarios is Msociated with the slope of the equilibrium response functions.

i6If an equilibrium position with payoff 11"* has been approached by way of
dampened oscillations, there will be many vertices close together in a vicinity of
(A, p.) = (11"*,11"*) (or infinitely many in a mathematical abstraction). Whether
such an interface is classified as having a vertical or horizontal connection to
the hypotenuse of Ll will depend on the shape that remains when the 'nearest'
vertices to (11"*, 11" *) are wiped out by the shock.
74

Figure 6 illustrates that the shocks in the cyclical scenario, in one direction
at least, have a weaker impact on the equilibrium share b* than in Scenario
1. This finds its expression in the fact that the simulations of Scenario 2
yield a distribution of the F(8k , V*(k)) with a lower standard deviation
(the exact values are 1.29% in Scenario 2 versus 2.27% in Scenario 1).
We may now address the problem of instability and ask what happens
if the process does not converge to an equilibrium. It is easily seen that
system (3-10) has built in a global negative feedback mechanism, which
prevents the trajectories from exploding. If, for example, all firms adopt
the set of bold strategies for a longer while, then sooner or later the
corresponding aggregate exposure (equation (8)) dampens the psychology
of financial markets and raises financial distress d to similarly excessive
levels (see (9); at least if, as here assumed, the growth rate coefficient
ad in the distress function (14) is not too large): This rise in d inexorably
reduces the differential payoffs from bold behaviour in (10), and those firms
with the highest A-switching values return to the prudent strategies. An
analogous argument applies should all firms have become prudent. If, on
the other hand, no equilibrium is approached, it follows that the economy
must undergo persistent fluctuations. From a Minskian point of view, which
stresses the influence of financial markets on real activity, these oscillations
might even be seen as a constituent part of a theory of endogenous business
cycles.
It has been noted before that when a vertex is wiped out, the process
gives rise to a discontinuity in the increments of the share of bold firms
from one short period to another. One may surmise that this switching in
the regimes of the ordinary difference equations leads to chaotic dynamics.
It turned out, however, that the sudden changes were not so serious in this
respect. In all our simulation runs the motions soon developed into regular
oscillations which, for all practical purposes, can be described as strictly
periodic. Thus, the economy no more converges to a point of equilibrium,
but to a closed cycle (or a closed orbit, more technically). Similarly as with
the shifting equilibria above, it will be expected that these cycles are not
unique and that different shocks induce convergence to different cycles.
A condition for the equilibria of the process to become unstable is that
the growth effects in the two functions 111: and Id in equations (14) and
(15), as measured by the slope coefficients a11: and ad of the perceived
rate of growth, are sufficiently strong. An example is Scenario 3, which
was made explicit in the preceding section. Let us then perform the same
experiment as with the previous two scenarios and perturb the steady state
position by a shock to financial distress. The first remarkable effect to be
observed is the phenomenon of corridor stability (as Leijonhufvud has called
it). A positive shock does not change the equilibrium, which is already
known. It turns out that its stability is preserved when the shock is small
enough, where convergence is virtually monotonic. The critical value of the
perturbation is about 8 = 0.98%. If 8 exceeds this benchmark, then in its
 75

adjustment back to 71"* the payoff overshoots this initial equilibrium value.
Unlike Scenario 2, the acceleration in the oscillations then setting in is so
vigorous that they are no longer dampened. As indicated, these fluctuations
are attracted by a periodic orbit. Interestingly, negative shocks, which do
affect the equilibrium position, are totally destabilizing, i.e., the periodic
cycles come into being however small in modulus the shocks may be. The
corridor stability arising in this way is therefore one-sided, as we may say.
Under sustained growth cycles, a unique steady state position in ordinary
dynamic systems, though not manifest in the time series, has still a role
to play as a state of reference. Since furthermore the equilibrium values
are often a satisfactory proxy for the time averages of the variables, the
steady state provides a theoretical and, to some extent, even practical
basis of the analysis. In the present dynamics, the steady state which we
used for reference is, and remains, not only virtual, it is also continuously
shifting. Such a system has no more firm anchor. A notion of equilibrium
is maintained, but it is devoid of any deeper meaning.
Besides the time averages of selected variables, another characterization
of the limit cycles are the peak and trough values of the share of bold
firms. Upon systematic variations of the shocks 8, both of them are seen
to increase weakly as the (positive and negative) shocks increase in size;
the troughs somewhat more than the peaks, so that the amplitude slightly
decreases. This is shown by the upper and lower lines in Figure 8. Additional
details of the outcome of the present calibration are given in the Appendix.

[ b I P• ..tr. and Troughs

70

6~

60

55
~ ~
"------.J
~o

45

-aD -~o 0 ~D aD 30 [ • J

Figure 8 Share of bold firms (in %): peak and though values of the limit
cycles in Scenario 3 induced by shock 8 (in %); convergence to a new
steady state if they coincide.
76

A striking phenomenon emerges when the shocks grow large in the

experiment, that is, if 6 ;::: 24% for positive shocks and 6 ~ -11.5% for
negative shocks. After a shock in this range, the variables still oscillate in
the expected manner for quite some time, even six or seven cycles in some
cases, with a slowly diminishing amplitude. From the experience with the
weaker shocks, one is inclined to think that again the economy is going to
approach a periodic motion. Then, however, although no further vertex is
eliminated, the amplitude is more drastically reduced - and the economy
converges to a single state of equilibrium! The system thus exhibits an
extended kind of corridor stability.
To venture an analogy, it seems as if the economy needs a long run-
up to gain such a momentum and enter so deep into the force field of a
planet that, after a number of narrowing orbits, it is able to escape from
the other forces of attraction in space, so that the planet's forces of gravity
finally succeed. In particular, convergence to an equilibrium is achieved by
extreme shocks which obliterate history and wipe out all previous vertices,
even temporarily causing all firms to adopt the same type of behaviour
(though these features are not a necessary condition for the phenomenon
to occur). The analogy with the gravitation field is also valid insofar as
the larger shocks speed up the process of convergence. As far as the author
knows, such a reswitching of (point-) stability, brought about by sufficiently
strong shocks, has not been obtained before in economic dynamic systems.
The different shock effects produced by the variations of 6 are summa-
rized in Figure 8. Clearly, the intervals where the upper line for the peak
values and the lower line for the trough values of b coincide, signify that
here the economy converges to a single steady state position. This happens
in the small corridor between 6 = 0 and 6 = 0.98%, and for the sizeable
shocks just made explicit. Reading the diagram from the middle to the
outer range of 6, we see the reswitching of steady state stability.
The economy of Scenario 3 was, likewise, subjected to the above finan-
cial random shocks. Maintaining the timing of their arrival every four years
means that the oscillations are perturbed roughly every half cycle, irrespec-
tive of the particular stage in which the economy may presently be (the
fluctuations have a period between 7 and 8 years). Even if a sizeable shock
to d were to produce convergence to a single equilibrium, the adjustments
are cyclical and take some time. So it is quite likely that one of the sub-
sequent shocks neutralizes the gravitation towards the equilibrium point,
and the attraction of a closed orbit again becomes the dominant force. Con-
sidering the latter as the normal case, the single shocks can be viewed as
giving rise to another periodic motion towards which the economy begins
to converge. On the analogy of the shifting steady states, hysteresis in this
framework means that these periodic cycles shift over time. For short, one
may speak of cyclical hysteresis.
It is, however, to be expected that this kind of hysteresis is more difficult
 77

to perceive. One reason is that the cycles do not appear to differ very much
in their characteristics. The variations in the peak and trough values of b
in Figure 8 are relatively small as the shocks 8 are systematically increased
(the scale of b in Figures 6 and 8 is the same). On the other hand, the
shocks in the dynamic economy occur rather frequently in comparison to
the time required to approach the limit cycle. These two features combined
may cause the noise of the shocks to dominate the changes in the true
cycle characteristics, to which the system would converge if it had been
given enough time. Or, if we study the moving time averages of a smooth
cyclical series like gk, there will be only a limited scope for them to evolve
as being governed by a random walk.
This supposition is essentially brought out by the shock-generated time
series of the aggregate capital growth rate gk in the bottom panel of Figure
7. Contrasting it with the other two scenarios, the larger scale in the third
panel should be noticed, which is due to the wider fluctuations in this
scenario. Especially the shocks between t = 70 and t = 90 have a much
stronger effect on the amplitude. The different pattern of the bottom series
makes it also more difficult to recognize the three growth regimes of the
previous scenarios. Referring to the time averages of gk, however, two of
them are still present with approximately the same growth rates as before,
though they stand out less clearly as in the top panel. These are the first
and third regime, i.e., the time intervals [15,50] and [140,160]. On the other
hand, the growth regime over the period [105, 130] with its distinctly lower
growth rate in the upper two panels has disappeared; here average growth
takes place at almost the same level as in the first interval.
Moreover, quite surprisingly when looking at the first hundred 'years',
and knowing that these cycles are basically endogenous, temporarily the
oscillations nearly vanish. This phenomenon can be explained on the basis
of the discussion of Figure 8. A strong positive shock to financial distress
has precipitated a deep fall in gk around t=90, such that, ifleft to itself, the
system would converge to a single equilibrium position. It happens that the
subsequent shocks do not decisively disturb these adjustments. A steady
state growth path is actually reached 20 time units later. Then, around
t = 115, another positive shock to d arrives. Here the Leijonhufvud corridor
stability criterion applies. The shock is so small that the eonomy returns
to a new steady state, with an only slightly higher rate of growth. From
t = 128 on, the shocks are in the 'normal' range and the conomy resumes
its cyclical behaviour.
Apart from hysteresis in the - only mildly - shifting time averages of
gk, it might be said that the process also displays hysteresis in kind: from
sustained oscillations to steady growth, and back to sustained cycles again.
Though, this switching in qualitative behaviour will not be expected to
take place too often.
78

7 Conclusion

Adopting the framework which was discovered by Cross (1994) in the

physical sciences and translated into economic terms, the present paper
has put forward a model of heterogeneous investors in fixed capital with
discontinuous, lumpy adjustments to the changes in the real sector. It was
demonstrated how these individual inertia carry over to the macro level and
lead to path-dependence in the aggregate variables. The dynamic features
of the model are richer than what may be obtained by direct hypotheses
about the macroeconomic relationships, or by employing the representative
agent. Our approach thus widens the scope for the notion of hysteresis. In
particular, the often used random walk processes can here only serve as an
approximation to the model (within a certain setting), and possibly not a
good one.
The individual behaviour of agents is based on a number of simplifying
assumptions. It will, however, not be expected that the hysteresis effects
observed would be weakened if they were relaxed. It may rather be
worthwhile to look for more differentiated or even new effects. Three
extensions of the model present themselves for future research in this
direction. (a) The uniform distribution of the agents' characteristics may
be replaced with other, less regular distributions. This could reinforce
asymmetric reaction patterns. (b) The individual (A, j.t)-switching values
which have been treated as constants may be endogenized, so that their
distribution across agents is shifting over time. (c) Also more than two
polar types of agents' strategies may be examined.
We wish to point out that our analysis may have a bearing on macrody-
namic modelling. Since not every medium- or long-run model can afford to
consider the micro level in similar detail as we did, it may be asked if the
error committed by disregarding the corresponding hysteresis effects is neg-
ligible. According to the discussion in the last two sections an answer will
not be independent of the kind of dynamics the model builder is interested
in. Put briefly, we have found that convergence to the steady state posi-
tions strongly favours persistence in the variables in response to exogenous
shocks, whereas endogenously generated fluctuations tend to absorb these
shocks to the system. A tentative conclusion is that macro models whose
deterministic part gives rise to sustained (nonchaotic) cyclical behaviour
are more robust to hysteresis originating at the micro level than models
with stable equilibria.
 79

8 Appendix
Given V = {Vo,vl, ... ,Vm } = {(Ao,/-Lo),(Ab/-Ll), ... , (Am,/-Lm)} and
a (new) value of the payoff, 1T, distinguish two cases in determining the
corresponding new interface.
Case 1, 1T ~ Am : Determine k = min {i: 1T ~ Ai} and set

<I> (1T, V) = {(Ao, /-Lo),(Al, /-Ll)' ... , (Ak-b /-Lk-d, (max[1T, Amin], /-Lk)}
As for 1T ~ Amin, recall Ao = Amin, so that <I>(1T, V) = {(Ao, /-Lo)}
{( Am in , /-Lmax)} for these extreme payoffs.
Case 2, 1T > Am : Determine
k = { min {i: /-Li ~ 1T} if /-Lm ~ 1T
m+1 else
and set

<I> (1T, V) {(Ao,/-Lo),(Ab/-Ll), ... , (Ak-l,/-Lk-d,

(min[1T, /-Lmax], min[1T, /-Lmax]) }

In particular, for 1T ;::: /-Lmax one has <I> (1T, V) = {(Amin, /-Lmax), (/-Lmax, /-Lmax)}.
Derivation of the Investment Growth Rate g; in Equation (6):
Given the length of the adjustment period h, let hIt be the volume of
aggregate net investment over the time interval [t, t+h) and K t the capital
stock at time t. Specify the (annualized) growth rates gf = (Kt+h-
Kt)/hKt and g; = (It - It-h)/hIt - h . Using It = (Kt+h - Kd/h = gf K t ,
one computes

gf K t - gf-h K t- h (gf - gf_h)Kt + gf_h(Kt - Kt-h)

=
hgf-h K t- h hgf_h K t- h
gf - gf-h Kt Kt - Kt- h

hgf_h K t- h
+
hKt-h
k
gt - gtk_h KKk
= (1+h t - t-h) +
h gf-h hKt - h gt-h

which is equation (6).

Spec~fication of the Output Growth Rate fy = fy(gi) in Equation (7):
With respect to a given value of g and the four parameters Iyl , Iy2, Qyl, Qy2,
where 0 ~ Iyl < Iy2, 0 ~ Q y2 < Qyl ~ 1, define
gi if Igi - gI ~ ,yl
g + Iyl + Qyl (gi - g - ,yd if Iyl ~ gi - g ~ Iy2
g - Iyl + Qyl (gi - g + ,yd if - Iy2 ~ gi - g ~ -,yl
g + Iyl + Qyl (')'y2 - IyI)
+ QY 2(gi - g - Iy2)
g - Iyl - Qyl (')'y2 - Iyl)
+ QY 2(gi - g + Iy2) if gi ~ g - Iy2
80

Thus, fy(gi) has slope lover the interval [g-ry},g+ryI], slope ayl over
the two intervals [g-ry2, g-ryI] and [g+ry}, g+ry2], and slope a y2 else.
Numerical Details of the Simulations:
9 0.03 ryl = 0.03 ry2 0.06
/3g 2.00 ayl 0.667 a y2 0.25
/3k 1.00 gp 0.01 gb 0.06
/3d 1.00 h 0.05
On the basis of Amin = 0 and "'max = 10, the initial equilibrium interface
V*, which accomplishes b* = B(V*) = 50%, is given by
V* = {(O, 10), (1,9), (2,8), (3,7), (n*, n*)} , n* = 6.3166
The special value of n* (the same in the three scenarios) is brought about
by the parameters (11< = 10, 'l/J = 0.30, rd = ad 9 in the payoff and distress
functions (14,15), and
a1< 0 ad 0 2.81662
a1< = 4 ad = 1 2.11662
a1< = 5 ad 2.5 1.94162
in Scenario 1, 2 and 3, respectively.
Discussion of Equation (16) in Scenario 1
With respect to the discussion following equation (16), it has to be
shown that in response to a negative shock to d, the payoff in the new
equilibrium is smaller, and the new equilibrium share of bold firms is
larger. To see this, refer to the value n* = n*(b) in equation (11), which
the payoff must attain in equilibrium. In the initial equilibrium we have
n' = n*(b' ), V' = <I>(n/, V'), and the fixed-point equation (12) reads
b' = B[<I>(n/, V')] = B[<I>(n*(b' ), V')]. With respect to the shock-induced
interface V", the strict inequality B[<I>(n, V")] > b' = B[<I>(n/, V')] holds
for all payoffs n;::: n'P In particular, B[<I>(n*(b' ), V")] > B[ <I>(n*(b' ), V')]
= b'. Since b f---+ B[<I>(n*(b), V")] is a decreasing function (cf. the remark
following equation (12)), the solution b of the new fixed-point equation
B[<I>(n*(b), V")] = b must exceed the original share b'. Since we have
n*(b) < n' < n", i.e., the payoff has to fall from n" to its new equilibrium
level, it follows that the equilibrium interface changes its shape and is now
vertically linked to the hypotenuse of the Do-triangle.
Closed Orbits in Scenario 3:
Besides the peaks and troughs of the share of bold firms mentioned in the
text, the present calibration of the model in Scenario 3 yields the following
features of a typical cycle. The trough values of the perceived rate of growth
g, the aggregate capital growth rate gk, and the growth rate of investment
gi (a second derivative so to speak) are, respectively, -1.42%, 3.46%, and
-3.68% per unit of time; the corresponding peaks are 8.67%, 4.04%, and
11.95% (recall that practically all growth rates are instantaneous). The

I7For example, the interface in Figure 4 that now corresponds to n ' is given by
V= {Vo, VI, V2, V4 }. Clearly, the area below V is larger than the area below V' =
{Vo, VI, V2, V3, Vn, which equals B[cfI(n/, V'}] (apart from the normalization).
 81

period of the cycle is 7.55 time units, where the phases from peak to
trough and from trough to peak are almost symmetric. The fluctuations
in 9 and gi are not too unrealistic compared to ordinary business cycles if
the underlying time unit is taken as a year. A shortcoming in this business
cycle interpretation is the small amplitude in the capital growth rate -
unless the concept of gk is refined and gk is viewed as an aggregate desired
growth rate of fixed capital. Since already the mean g:= gbb+ (1- b)gp was
an aggregate target rate of growth, g and gk would be associated with two
different time horizons. So it has to be admitted that this interpretative
construction is somewhat artificial. If beyond the purely theoretical thrust
of the model, calibration of these cycles is seen as an important task to
validate the model, then a more detailed specification of the adjustments
in the real sector is demanded.

References
[1] Blanchard, O.J. and Summers, L.H. (1986), "Hysteresis and the Eu-
ropean unemployment problem", in S. Fischer (ed.), NBER Macroeco-
nomics Annual 1986. Cambridge: MIT Press; pp.15-78
[2] Blanchard, O.J. and Summers, L.H. (1987), "Hysteresis in unemploy-
ment", European Economic Review, 31, 288-295
[3] Cross, R. (1994), "The macroeconomic consequences of discontinuous
adjustments: selective memory of non-dominated extrema", Scottish
Journal of Political Economy, 41, 212-221

[4] Dixit, A. (1992), "Investment and hysteresis", Journal of Economic

Perspectives, 6, 107-132
[5] Dixit, A. (1995), "Irreversible investment with uncertainty and scale
economics", Journal of Economic Dynamics and Control, 19, 327-350
[6] Franke, R. (1987), Production Prices and Dynamical Processes of the
Gravitation of Market Prices. Frankfurt a.M.: P. Lang

[7] Franke, R. (1994), "A reappraisal of adaptive expectations", mimeo,

University of Bremen
[8] Franz, W. (1990), "Hysteresis in economic relationships: an overview",
Empirical Economics, 15, 109-125

[9] Giavazzi, F. and Wyplosz, C. (1985), "The zero root problem: a note
on the dynamic determination of the stationary equilibrium in linear
models", Review of Economic Studies, 52, 353-357
82

[lOJ HEINER, R.A. (1988), "The necessity of delaying economic adjust-

ment", Journal of Economic Behavior and Organization, 10, 255-286
[11J Mayergoyz, LD. (1991), Mathematical Models of Hysteresis. New York:
Springer

[12J Miller, M.H. and Orr, D. (1966), "A model of the demand for money
by firms", Quarterly Journal of Economics, 80, 413-435
Non Linear Dynamics and
Utility Functions in
Overlapping Generations
Models
Gilles DUFRENOT and Laurent MATHIEU

1 Introduction
Grandmont [7] was among the first to study the possibility of self-sustaining
cycles in the overlapping generations models. In his paper, the origin of
these cycles is the conflict between the substitution effect and the income
effect due to variations of relative prices. Grandmont's approach has been
criticized on empirical basis : income effects in the model are too large
compared to their values in real life. Our paper, conversely, provides some
theoretical arguments that explains why cyclical and complex paths cannot
be ruled out. In this view, we examine the problem of endogenous fluctu-
ations in pure exchange economies from the viewpoint of macroeconomic
theories of consumption and saving.
Many studies explore the idea that in the face of uncertainty concerning
anticipated labor income, consumers can be more or less prudent (Carrol
and Summers [2], Carrol [1]). Also, the role of risk aversion has been
investigated (Shapiro [9], Dynan [4]). The propensity to consume out of
labor income depends on both factors which appear as potential sources of
fluctuations in aggregate consumption and saving. Moreover, recent works
exploring the importance of these factors have reached the conclusion that
the combination of risk aversion and prudent behavior may be a major
cause of the very unstable nature of aggregate saving (Dubois and Bonnet
[3]). These observations motivate our analysis.
Following an early suggestion by Dreze and Modigliani [5], we show that
the degree of curvature of an offer curve (or equivalently, the substitution
and income effects) can be expressed in terms of the first, second, and
third derivatives of the utility functions that describes preferences of the
consumers. Those derivatives are considered for they contain information
on prudent behaviors and attitudes towards risk in the face of uncertainty.
To achieve this result, we start from the methodology initiated by Sato
84

[8J which shows a close connection between the hypotheses made on the
Slutsky equations corresponding to excess demand functions and classes of
utility functions from which cycles and complex dynamics emerge.
The paper is organized as follows. Section 2 presents the model. In
Section 3 we derive classes of utility functions. Section 4 uses some
simulations to illustrate the nonlinear dynamics of consumption associated
to each type of utility functions.

2 The Model
This model, which is based on an overlapping generations model, is com-
posed of two distinct agents each living two periods. In the first period the
agent is in the younger generation; while in the second he is in the older
generation. In this pure exchange economy, each agent receives units of
a good which is neither produced nor stored. The population is constant
and both generations engage in goods exchanges. The market driven by
a system of relative prices. The model is based on the following assump-
tions. There is a representative agent in each generation who determines
how much to consume in the current period cf and in the future period
£1+1. The representatives' utility function U (cf, £1+1) is additively separa-
ble and satisfies the usual assumptions (monotonicity and concavity). The
endowment in the two periods of his life are given as w = (wY, WO).
Based on these assumptions, the structure of the economy is represented
by the following two equations .
• Following Gale [6J, a program is called feasible if the aggregate
consumption equals aggregate endowments in each period:
(1)
• A program is called competitive if a young agent maximises his utility
under the budget constraint:

{ Max U(cf,£1+l)=Udcf)+U2(£1+l) (2)

ptcf + Pt+l t;'+l = PtwY + Pt+l WO
The price of the good in period tis (Pt)t=I,2. The utility functions of
the young and the old agents are respectively Ul (cf) and U2 (£1+1).
These functions satisfy the following conditions:
U; (.) > 0 and U;' (.) < 0 for i = 1,2. (3)

Let us define the interest factor Pt as the rate of exchange between

current and future consumption:
Pt
Pt=--= 1 +tt
. (4)
Pt+l
 85

where it is the real interest rate. In order to simplify the resolution

of the model, we define J.Lt = .l..
p,
Thus, the system (2) is written:

(5)

• We are interested by the consumption time paths which satisfy both

equations (1) and (5). The first order optimality conditions of the
maximisation program implies:

(6)

where U~ and U~ are the first order partial derivatives of U relatively

to c¥ and c~+l' Substituting this last equation into the budget
constraint implies:
0.2' w°- ct0 +1
(7)
U~ c¥ - w Y
Our approach to the study of the dynamics of aggregate consumption,
given the specifications of the utility functions, is based on two steps.
Firstly, we will use the methodology proposed by Sato [8] which consists
of deriving a class of additive utility functions in accordance with the
assumptions of the substitution and income effects of the Slutsky equations.
Secondly, we will use the utility function to study the behavior of the
consumption dynamics.

3 Derivation of the Utility Functions

We first write the Slutsky equations, then in deriving the utility functions
we impose different restrictions on the terms representing the substitution
and income effects: either the price elasticity of the hicksian demand is
constant or the engel curves are linear.

3.1 THE SLUTSKY EQUATIONS

The Lagrangian of the consumer program (5) is:

L (c¥, C~+l'.x) = U1(cn + U2 (c~+1) +.x [w Y - c¥ + J.Lt (W O

- C~+l)] (8)
where .x is the Lagrange multiplier. The first-order conditions are:

(9)
86

The solution of this system is a maximum if the determinant of the

bordered hessian matrix H B is positive. This matrix is written as follows:

u~' o -1 )
o U~' -J.Lt (10)
-1 -J.Lt o
and the determinant is:

(11)

Respectively, U; and U~ are the second order derivatives of UI (cn and

U2(cf+l) relative to c¥ and Cf+I'
The determinant is positive since we have assumed that the utility
function is strictly concave. The differentiation of the system (9) leads
to:
U~' dc¥ - d-X = 0
{ U~' dcf+1 - -XdJ.Lt - J.Ltd-X = 0 (12)
dw Y - dc¥ + J.Ltdw° + W O dJ.Lt - J.Ltdcf+1 - cf+ I dJ.Lt = 0
from which we deduce the expression of the Slutsky equations (see
appendix 1 for the proof):

~ ~ (0
= IHBI + cHI - w IHBI
O)~
{ dJ1.. (13)
dC~±l - >. ( 0 0) J1..u;'
dJ1.. = IHBI + ct +1 - w IHBI

Note that I~% et IN ~I are the slopes of the hicksian demand curves of
c¥ and cf+I'
.
The expreSSlOns (c0
t +1 - w
0) ~
IH BI and (0
ct + 1 - w0) J1..u;'
IH BI d escn'be the
income effects of a relative price variation.

3.2 FIRST RESTRICTION: THE ELASTICITY PRICE OF THE

HICKSIAN DEMANDS ARE CONSTANT
We note Ecf1u=O as the elasticity price of the hicksian demand of a young
agent, which is expressed as:

(14)

using the expressions of IHBI and -X we obtain:

(15)
 87

or (see appendix 2 for the proof)

(16)

This equation is the product of two expressions: the marginal propensity

of the older to consume (see Sato [8] pp. 105 for the proof) and the inverse
of the relative risk aversion coefficient of the younger. We assume first that
the first term is constant. If so, then the elasticity (16) is constant if the
relative risk aversion coefficient is constant. This leads to two expressions
for the utility functions of a young agent:

(17)
Xl> 0, ~l >0
Similarly, the same elasticity for an old agent is given by
-A /-Lt
(18)
EC~+1lu=u = IHBI ~
t+l

and using the expressions of IHBI and A we obtain:

(19)

or (see appendix 3 for the proof):

Eo
Ct + 1
lu= u- = (20)

Once again, this equation is the product of two expressions: the marginal
propensity of the younger to consume (see Sato [8] pp. 105 for the proof)
and the inverse of relative risk aversion coefficient of the older. Assuming
that the first expression is constant, then the elasticity (20) is constant if
the relative risk aversion coefficient is constant. We obtain the following
expressions for the utility functions of the old agent:

(21)
88

3.3 SECOND RESTRICTION: THE ENGEL CURVES ARE

LINEAR
There are two steps to obtain utility functions expressions that satisfy the
hypothesis of linear Engel curves.
The first one consists of obtaining a condition on the derivatives of the
utility function. It can be shown (see appendix 4 for the proof) that linear
Engel curves imply that the marginal propensity to consume is constant.
Which leads to the following relation:

(22)

In the second step we solve this third-order differential equation and

find three expressions for the utility functions. Let us define Xl = cf and
X2 = c~+1'
Then:

k = 1 and Ui = ~ exp (;. Xi + Xi) i = 1,2

where G:i < 0, f3i < 0, Xi E ~
k = 2 and Ui = 7, In ['Yi ( ~Xi + 'Pi)] i = 1,2
where fi < 0,<Pi'> 0, (-Yi,'Pi) E ~2
2-k
k E ~\ {O , 1, 2} and U·1. = _1 2-k [K'' (1- k) (x·'I - ..,.))T=k
1.."1<,, _1_ ."
i = 1,2
where Li > 0, Ki < 0, (1 - k) (Xi -1/i) < 0, 1/i E ~
(23)

4 The Dynamics of the Aggregate Consumption

4.1 CASE 1: THE DYNAMICS WHEN THE ELASTICITY PRICE
OF THE HICKSIAN DEMANDS ARE CONSTANT
As is shown in appendix 2 and 3, it is possible to derive the dynamics of
aggregate consumption by considering the following combination of utility
functions for the first and second periods:

{ Ul (cf) = rlln(Lllcf)
U2 (c~+1) = r2ln (Ll2~+1)
where
where
Xl> 0, Lll >
X2 > 0, Ll2 >
°° (24)
This selection is made according to the hypothesis that the marginal
propensities of the younger and the older to consume are constant.
The consumption dynamics are studied by considering both the opti-
mality conditions of the consumer maximisation program (7) and equation
(1)
 89

After a few transformations this leads to a first-order difference equation

(see appendix 5 for the proof):

wY + (wY + WO) [Gr Cr-yW Y ]

c=
Y
t
f( Y)
Ct = 1+
[rr2 CY_WII]
cr
2

!...l. .:.L-::.-
Ct
where

Proposition 5 When the dynamic of the economy is governed by equa-

tion (25), the timepath of the consumption is described by a monotonic
convergence towards two stationary equilibria (proof: see appendix 5).

For a graphical illustration, the reader is referred to figures 1

(w Y = WO = 0.5, R = 0.5, cg = 0.9) and 2 (w Y = WO = 0.5, R
= 10,
cg = 0.51). The curves represent the timepath of the consumption. The
line in each figure represent the set of equilibria corresponding to the
stationnary state. In figure 1 the dynamics of consumption converge to
the no trade equilibrium (ct)* = 0.5. In figure 2 the convergence is towards
the second equilibrium (Ct)** = 0.909

° 0,2 0,4 0,6 0,8

Figure 1 : No trade equilibrium

90

° 0,2 0,4 0,6 0,8

Figure 2 : Second equilibrium

4.2 CASE 2: THE DYNAMICS OF LINEAR ENGEL CURVES

Proposition 6 Under the assumption of linearity of the Engel curues, only
one specification of the utility functions is kept. Let us define Xl = cf and
X2 = cf+l. We have (for a proof see appendix 4):

u-t(x-)
"
= _1 _1
TiKi 2-k
[K-'I (1 - k) (x-"-. ".,.,_)Jt=k + tn-
2-k
T' (26)
where Ki < 0, Ti > 0, 'T/. E !R and k E !R\ {O, 1, 2}

Proposition 7 Given the utility function represented by equation (26), the

consumption dynamics is given by a first-order difference equation (for a
proof see appendix 6):

cf+l = wY - G:) l-k

2-10 (w Y - cn~== (cf)2~ (27)

This dynamic equation is a kind of logistic equation. Thus the timepath

of the consumption will vary from an oscillatory convergence to a station-
nary equilibrium to periodic cycles and chaos, given a set of values for the
parameters. The figures which illustrate the consumption's timepath for
different values of the parameters are represented next page. We have cho-
sen in particular: wY = 1, k = 0.9 and !2.·1
varying from 1.6 (convergence
to stationnary equilibrium) to 1.99 (chaotic timepath).
 91

1 -.

0,8

0,6

0,4

0,2

o~--~--~~~~~--~

°
0,2 0,4 0,6 0,8

Figure 3 : Phase diagram

0,8

0,6

0,4

0,2

O~----'-----'-----'-----r----.

° 0,2 0,4 0,6 0,8

Figure 4 : Stationnary equilibrium

92

0,8

0,6

0,4

0,2

O~----r----.-----.-----.----~

° 0,2 0,4 0,6 0,8

Figure 5 : Two period cycle

0,8

0,6

0,4
............................ (
0,2

° 0,2 0,4 0,6 0,8

Figure 6 : Four period cycle

 93

0,8

: : I:~;.:i:?].:
0,2
:; ;:

o 0,2 0,4 0,6 0,8

Figure 7 : Eight period cycle

0,6 J::a·:·:·:"·:-:·:·:·:·:·:·:·:·:·:·:·:·:·:·:·:·:···:·:·:·:·:·:·:·:·:·:-:·:·:·:·:·:·:·:-:·:-:·:·:··:·V

0,4

0,2

o
o 0,2 0,4 0,6 0,8

Figure 8 : Chaos
94

Appendix 1 :Derivation of Slutsky equations

From equation (9) pp. 91) we use a second-order differentiation:

(28)

Using a matrix representation we write:

(29)

The Cramer rule implies:

o o -1
>.dJ1-t u.2" -J1-t (30)
-dw Y + cf+l d J1-t -J1-t 0

U~' 0 -1
d ol (31)
ct +1 = IHBI 0 >'dJ1-t -J1-t
-1 -dw Y + Cf+ldJ1-t 0
where IHBI is the determinant of the bordered hessian matrix. In a more
extensive form we have:

dcf = I;BI [J1-t>.dJ1-t - U~' dw Y + U~' C~+l dJ1-t] (32)

dct+l
o 1 [-U1" J1-t dw y + U1" J1-tCt+l
= IHBI 0 d J1-t - 1\'d]
J1-t (33)

We now divide each expression by dJ1-t assuming that w Y is fixed:

(34)

(35)

The first term on the right hand side is the substitution effect of a relative
price variation on the quantity of goods consumed by the young agent and
the old agent.
 95

Appendix 2 : Derivation of utility functions for the first period

when the price elasticity of the hicks ian demand is constant

Let us define Ecr1u=u as the price elasticity of the hicksian demand of

a young agent:
E _ _ dc¥ Ilt (36)
crlu=u - dll t IU = U c¥
Then,

(37)

Substituting IH BI and A by their expression (note that from the first-

order condition of the maximization program A = Ut' = ~),
1',
we have:
, 2
_ Utll t 1
EcY,lu=u - 2" " Y (38)
-Ilt Ut - U2 ct

A s Ilt =~
u" we 0 b
\
'
tam:

-U~ (~r 1
(39)
( ~)2 U" U" c¥
u' t + 2
1

Dividing both the denominator and the numerator by (U~) 2 the equation
above becomes:

(U~)2 -U~
(40)
(U~)2 U~' + (U~)2 U~' c¥
We factor the denominator with U~' U~'

(U~f -U~ (41)

[~ (U~)2] T
U~'U~' u" +7-
\ 2

Of,

(u~r
u~' -U~
(42)
( Cu;,r
u\
+ CU~l) c¥U~'
u2
This last expression is the product of two expressions: the marginal
propension of the older to consume (see Sato [8] pp. 105 for the proof)
and the inverse of the young agent's relative risk aversion coefficient.
96

Conjecture 8 The marginal propension to consume of the old is constant.

Under this assumption, the elasticity (16) is constant if the relative risk
aversion coefficient is constant. This leads to the following proposition.
Proposition 9 .If the marginal propensy to consume of a young is con-
stant, then the preferences are described by the following expressions:
Ul (cn = Al Bl~1 (cn B1 +1 Al > 0, Bl E lR-\ {-1}
{ and (43)
Ul (cn = Xl In (AleI') Xl > 0, Al > 0
Proof
Assume that the· relative risk aversion coefficient is constant. Thus, we
write:

ct 1
~U~
" = kl (44)
where kl is a real number. From this equality, we deduce a second-order
differential equation:
y"
klctU1 - Ul = 0 (45)
I

To solve this equation, we define three cases .

• kl = 1
this implies,
U~' 1
- , =--y (46)
Ul ct
If we integrate this equation, we obtain:

In l,8lU~ I= In lal crl + Xl (47)

Conjecture 10 The constants in the equation above are defined such

that: al,,8l > 0 and Xl = 0

Taking the exponential of both side yields:

(48)

Integrating once more, we obtain:

U1 (cn = ;;1 (cn2+ 1\:1 (49)

Unfortunately, this function is not strictly concave:

U~ = ~cr > 0 since a1,,81 >0

{ and (50)
U1" = F,
'" > O·
smce al,/-,1
(.I >0
We thus eliminate this specification from our study.
 97

• k1 = -1
this implies,
u~' -1
-U' -;yt (51)
1

Integrating this equation leads to:

(52)

Conjecture 11 The constants are assumed to satisfy the following

conditions: Q1, (31 > 0 and Xl = 0

Taking the exponential of both sides yields,

(53)

We integrate once again:

(54)

Conjecture 12 The constants are assumed to satisfy the following

conditions: '0:71 > 0 and 101 = O.

We thus obtain a second specification of the utility function for a

young agent:
(55)

This function is monotone and strictly concave:

U1' = 1 1
o<,{3, ~ >0 since Q1,(31 >0
{
and (56)
-1 1
(cn 2 < 0 >0
/I
U1 = o<,{3, since Q1,(31

• k1 -=1= 1 and k1 -=1= -1

This implies
k U~' _ ~ (57)
1 U' -
1 cYt
We integrate and obtain:

kIln 1'Y1U~ I= In IOl C¥1 + Xl (58)

98

Conjecture 13 The constants are assumed to satisfy the following

conditions: (61, 'Yl) > 0 and Xl = 0

Taking the exponential of both sides yields:

(59)

or,
( I< )l/kl
U1' -_ Ul ( y)l/kl
Ct (60)
1'1
We integrate once more and obtain a third specification of the utility
function of the young agent:

(61)

This function is also monotone and strictly concave:

(62)
if
 99

Appendix 3 : Derivation of utility functions for the second pe-

riod when the price elasticity of the hicks ian demand is constant

Let us define EC~+llu=u as the price elasticity of the hicksian demand of

an old agent:
(63)

Then,
-A f..£t
(64)
EC~+llu=u = IHBI ~
t+l
Substituting IH BI and A by their expression (note that from the first-
order condition of the maximisation program A = u~ = ~),
J.Lt
we have:

(65)

As f..£t = ~, we obtain:
1

u~ 1
(66)
u U"1 + u"2 cf+l
( ~)2
1

We factor the denominator with U~' U~' :

1,)2
(U
Eo -=
ct+1Iu=u [f U "\.2 fu"\"2]
~ (67)
u;' + l::2L
CO
U"U" .L1..L HI
1 2 u;'
or,

Eo
Ct +1
lu = u- = (68)

This last expression is the product of two expressions: the marginal

propensity of the younger to consume (see Sato [8] pp. 105 for the proof)
and the inverse of the old agent's relative risk aversion coefficient.

Conjecture 14 The marginal propensity of the young agent to consume

is constant.
With this assumption, the elasticity (16) is constant if the relative risk
aversion coefficient is constant. This leads to the following proposition:
100

Proposition 15 If the marginal propensity of an old agent to consume is

constant, then the preferences are described by the following expressions:

(69)
X 2 > 0, ~2 >0
Proof: see appendix 2.
 101

Appendix 4 : Derivation of utility functions when the Engel

curves are linear.
Conjecture 16 The Engel Curves are linear.

Proposition 17 Under this assumption, we derive three specifications of

the utility functions. Let us define Xl = c¥ and X2 = Cf+l. We have:

Proof
Two steps are needed to prove this proposition.

1. Linearity condition of the Engel Curves.

The linearity of the Engel Curves implies that the marginal propen-
sities of the young and the old agents' to consume are constants.
Let us define pmCi as the marginal propensity of the young agent
(i = 1) and the old agent (i = 2) to consume. The linearity condition
of the Engel curves implies:

8pm Ci =0 (71)
8(wY + I-ltWo)

This condition is true iff fir); is constant whith any values of c¥ and
Cf+l (see Sato [8], pp. 117 for the proof). Thus:

U~u~1/
(U;')2 - , for i = 1,2, where k
-'-' - k E ~\{O} (72)

Note that the expression above leads to a third-order differential

equation:
(73)

2. Derivation of the utility functions.

Solving this equation yields three specifications depending upon the
value of the parameter k.
102

(a) k = 1
This implies,
(74)
or,
(U:,)2
•
_U:U:"
• •
~.....c....._...,..----O (75)
(Un 2 -

Integrating these equation gives:

(76)

where 0i is a real number. Integrating once again, we obtain:

In l.8iU; I= ~i Xi + Xi (77)

Conjecture 18 We assume that: .8i >0

Under this assumption, we take the exponential of both sides of
the equality and obtain:

U; (Xi) = ~
.8i
exp (-.!:..Xi + Xi)
0i
(78)

A new integration gives a first specification of the utility func-

(1 Xi + Xi ) + Oi
tion:
Ui (Xi) = 0·
.8: exp 0i (79)

where Oi is a constant which we assume to be equal to zero.

*
This function is monotone and strictly concave:

U; (Xi) = exp (;iXi + Xi) > 0 since

{ and (80)
U:'
t
(x·)
t
= .l.l
{3i Q:i exp (.lx.
Q:i'"
X·)
+ 1. < 0 if 0i <0
(b) k = 2
This implies,
''''
UiUi =2 Ui (,,)2 (81)
or,
(82)

We integrate both sides and obtain:

(83)
 103

Conjecture 19 We assume that: €i < 0, <Pi > 0, (i =0

One thus has:
(84)

Taking the exponential of both sides gives:

(85)

or,
(86)

Integrating, we obtain:

(87)

or,
(88)

Integrating once again, we obtain a second specification of the

utility function:

(89)

where we suppose that ei = 0 and "Yi (~/ Xi + 'Pi) > O. This

expression, which is a Stone-Geary utility function, is monotone
and strictly concave:

(90)

(c ) k =I- 1 and k =I- 2

This implies,
U~U~"
-'-'--k (91)
(U;')2-
or,
(92)
104

This yields,

(93)
which can be written as,

( U~,)2
, _ U~u~"
, ,
->---<--(U-;,""ft)2 - = - (k - 1) (94)

We integrate both sides and obtain:

u:'U~, = (1 - k) (Xi - "Ii) (95)

where (1 - k) (-"Ii) is a constant. We.can also write:

,
U~' 1
(96)
U; (1 - k) (Xi - "Ii)
Integrating this expression leads to:

In !LiU;! = (1 ~ k) In I/l:i (1 - k) (Xi - "Ii) I + Vi (97)

where Li, /l:i and Vi are real numbers.

Conjecture 20 We assume: /l:i (1 - k) (Xi - "Ii) > 0, Li > 0,
Vi = 0
and thus write:
(98)
Taking the exponential of both sides we obtain:
, 1 ---.L
Ui = ~ [/l:i (1 - k) (Xi - "Ii)]I="k" (99)

Integrating once more we derive a third specification of the

utility function:
1 1 ~
Ui (Xi) = - - 2 k [/l:i (1- k) (Xi - "Ii)]l-k + Ti (100)
Li/l:i -
where Ti is assumed to be equal to zero.
This function is also monotone and strictly concave:
u; (Xi) = t [/l:i (1- k) (Xi - "Ii)] 6 >0
if Li>O
and (101)
U;' (Xi) = 7:
[/l:i (1 - k) (Xi - "Ii)]~ < 0
if /l:i < 0 and if (1 - k) (Xi - "Ii) < 0
 105

Appendix 5: The dynamics of the aggregate consumption when

the elasticity price of the hicksian demands is constant

From appendix 2 and 3, we use the following specifications of the utility

functions:
Ul (cn = rlln(~lCn where Xl > 0, ~l >0
{ and (102)
U2 (Cf+l) = r2ln (~2cf+1) where X 2 > 0,
The optimality condition can be written as follows:

(c¥ - wY) U~ = (W O
- cf+1) U~ (103)

The equilibrium on the good market implies:

(104)

The derivatives of the utility functions yield:

Ul
, = yr l and U2
'
=
r2
-o- (105)
Ct Ct+l
Combining all these expressions, we obtain:

(106)

Proposition 21 The dynamics of aggregate consumption of the young is

described by a first-order difference equation.
Proof
From the equation (106), we reach:

wY + (wY + WO) [£1. cf"w"]

Y f ( Y) r2 t (107)
1 + [£1. c f -l/w
C

Ct = Ct = ll ]

r2 Ct

Conjecture 22 We assume that: wY +W O = 1 and (w Y , W O ) =I- 1.

Thus, the difference equation is rewritten:

(108)

The first-order derivative of f (cn provides some information on the

timepath of the consumption (monotonic convergence, oscillatory evolu-
tions, chaotic dynamics).
106

1. The dynamic properties of f (en

Let us define l as the first-order derivative of f:

!J. W"_(w")~

r
r2 (cll)2
l (en = t > 0 since (109)
[1 + [R cf:t"]
This derivative is stricly positive thus f is an increasing function of
eYt

2. The domain of definition of f (en

Since we assume that wY + WO ~ 1, e¥ ~ 1. Further, f (en exists for
values of df in the interval 2r:~r2 ~ e¥ ~ 1.
Indeed:

(110)

Thus,
(111)

3. Steady states equilibria

A steady state equilibrium must satisfy the following equation:

(112)

Since e¥ =f. 0 we can multiply both the numerator and denominator

bye¥:

(113)

We obtain,

which leads to a second-order polynomial:

(115)
 107

A first solution is (c¥)* wY,which corresponds to the no-trade

equilibrium.
A second solution is:

(c¥-w Y ) [(l+~~)C¥-~~] =0 (116)

from which we deduce:

( CY)* Y
=wan d (cY)** -_ r1 (117)
t t
r1 +r2
4. Stability of steady states equilibria
To study the stability of the steady states equilibria, we need to
evaluate the first-order derivative when c¥ = (c¥)* and c¥ = (c¥)**.
An equilibrium value x* is stable if jJ' (x*) I< l.
• We first check this condition for the no-trade equilibrium
(c¥)* = w Y

/ ((cn*) = (118)

We deduce:
(119)

thus,
rl1-w Y
-1<---<1 (120)
r 2 wY
The term between the two inequalities is strictly positive. The
stability condition corresponding to the no-trade situation is:
r wY
- 1 < -- (121)
r2 1- w Y
As an illustration, we choose w Y = 0.5 = (cn*. The inequality
above becomes R < 1. The convergence towards the steady
state equilibrium is monotonic / ((c¥n = 0.5> 0 (this is true
under the condition R
= 0.5) .
• We now study the stability of the second steady state equilib-
rium (c¥t* = rl~r2. As f is strictly positive, we have:
r1 wY
- < - -Y (122)
r2 1- w
and one can easily see that the convergence is monotonic.
108

Appendix 6: The dynamics of the aggregate consumption when

the Engel curves are linear.

We derive three cases.

1. Case 1
Let us define Xl = c¥ and X2 = cf+ I' The expression of the first utility
function is:

Ui (Xi) = /3; exp ( Qi1 Xi + Xi )

Q' where i = 1,2, Qi,/3i < 0 and Xi E ~

(123)
We thus have:
, 1 ( 1
Ui (Xi) = /3i exp Qi Xi + Xi ) (124)

which leads to the following expression of the consumption dynamics:

Unfortunately, it is impossible to find the inverse of the right hand

side of this equality. Therefore, there is not a unique relation between
c¥ and cr.
2. Case 2
We consider the following utility function:

Ui (Xi) = (;:)i 2 ln ['Yi «"'€/)Xi + c,oi) (126)

where Ei < 0, <Pi > 0 and (')'i, c,oi) E ~2

The first-order derivative is written as:

, -1
Ui (Xi) = (",.)2 (127)
€i Xi + c,oi
Conjecture 23 We assume that: c,oi = 0

Proposition 24 The dynamics of the consumption is given by a

first-order difference equation.

Proof
The dynamics of the consumption is given by (let (~/ = ri i =
1,2):
( cY - wY) _1_ - (cY _ wY) 1 (128)
t r I c¥ - t r I (W Y + WO - cn
 109

so,

3. Case 3

We consider the following utility function:

{ u· (x·) =
t.."
-II
-
Tilt, 2-k
[K,." (1- ',...
k:.k
k) (x·." - 'I'I.)]I-k + (Il·
·,.."t (130)
where K,i < 0, Ti > 0, 'TI. E lR et k E lR\ {O, 1, 2}
The first-order derivative is written as:

(131)

Conjecture 25 WO = 0

Conjecture 26 K,i = -1 (I - k) and "li = 0

Proposition 27 Under these assumptions the dynamics of the con-

sumption is given by a non-linear first-order difference equation.

Proof
The consumption dynamics is:
1 -L 1 -L
(ef - wY ) - [(k - 1) ef]l-k = (ef - wY ) - [(k - 1) (w Y - cn]I-1e
£1 £2
(132)
from which we deduce:
Y
-£2 (cf - .J....
wY ) {en I-Ie = (ef - wY ) (w Y -
_1_
ct) I-Ie (133)
£1

or,
(134)

and then,

(135)

This recurrence equation is a kind of logistic equation.

110

References
[1] CARROL C. [1992]' "The Buffer-Stock Theory of Saving: Some Macroe-
conomic Evidence", Brookings Papers on Economic Activity, 2, 61-135.
[2] CAROLL C., SUMMERS 1. [1991], Consumption Growth Parallels
Income Growth: Some New Evidence, University of Chicago Press.
[3] DUBOIS E., BONNET X. [1996], "Peut-on comprendre la hausse
imprevue du taux d'epargne des menages depuis 1990?", Economie et
Prevision, 121, 39-58.
[4] DYNAN K.E. [1993], "How Prudent are Consumers?, Journal of Polit-
ical Economy, 101, 1104-1113.
[5] DREZE J., MODIGLIANI F. [1972], "Consumption Decisions under
Uncertainty", Journal of Economic Theory, 5, 308-335.
[6] GALE D. [1973], "Pure Exchange Equilibrium of Dynamic ~conomic
Models", Journal of Economic Theory, 6, 12-36.
[7] GRANDMONT J.M. [1985], "On Endogenous Competitive Business
Cycles", Econometrica, 53, 995-1045.
[8] SATO K. [1972], "Additive Utility Functions with Double-Log Con-
sumer Demand Functions", Journal of Political Economy, 102-124.
[9] SHAPIRO M.D. [1993], "The Permanent Income Hypothesis and the
Interest Rate: Some Evidence from Panel Data", Economics Letters,
97,305-346.
Pay-as-You-Go System under
Permanent Business Cycle
Michel BOTOMAZAVA and Vincent TOUZE 1

1 Introduction

The literature dealing with overlapping generations economy, so called

because of its assumed demographic structure, is vast. OLG economy was
introduced by Allais [1947] followed by Samuelson [1958]. The three main
application fields are notably pay-as-you-go systems, endogenous business
cycles and intergenerational equity theories. The purpose of this paper is
to combine these three types of applications to study the pay-as-you-go
system effects on the economic cycle in terms of intergenerational justice.

The first field concerns pay-as-you-go system. Two results coexist and
generally deal with a steady state study: the pay-as-you-go system is
or is not welfare improving. When Diamond [1965]'s OLG structure is
used (Breyer and Straub [1993]), the pay-as-you-go system is not Pareto
improving when the capital yield is higher than the social security yield i.e.
population growth rate plus wages growth rate. However, in using Barro
[1974]'s model with bequest motive, one can prove individuals offset the
pay-as-you-go system effects by an offsetting bequest such that the transfers
between generations are unaffected. Moreover, Zhang [1995] showed that
when the population rate is endogenous, a pay-as-you-go program may
stimulate growth by reducing fertility and increasing the ratio of human
capital investment per child.
Second, Gale [1973] was the first to find an example of an OLG economy
with two-period cycles. Benhabib and Day [1982]' Grandmont [1985],
Reichlin [1986] and Jullien [1988] studied the emergence of erratic or
cyclic dynamics in the context of OLG models. More recently, Galor [1992]
investigated OLG model dynamics in a two-sector structure. The main
orientation of this research is to find and to characterize cyclic or chaotic

1 We are grateful to Catherine TOUZE for helpful comments and to Antonio

MELE for a pertinent correction. Of course, any remaining errors are ours.
112

time paths.
Third, generational equity time paths with OLG models were studied by
Phelps and Riley [1978] and Rodriguez [1981]. These authors investigated
the properties of "maximin" growth as meant by Rawls [1971]: so a fair
society would program its taxes and resulting stocks of capital and national
debt so as to maximize the lifetime utility of generations with having the
lowest welfare level. In these studies, the growth is stable and convergent.
A case of generational equity where the growth is cyclic has been studied
by Reichlin [1986]. He searches and finds a fiscal policy rule which implies
equal welfare for all generations when it is announced.

In this paper, we use these three economic concepts to study the effects of
change in a pay-as-you-go system. Our pay-as-you-go system is particular.
No direct pension is distributed but old age public expenditures are
financed by a payroll tax. This public good cannot perfectly substitute
for private good 2 •
In a first part, we present Reichlin [1986]'s model when this pay-as-you-
go system is introduced. We find a steady state capital general explicit form
related to exogenous economic parameters.
Then, in the second part, the conditions for the existence and the stability
of the cycles are given. As we want to obtain a permanent cycle, we must
accept to have a parameter which varies when the payroll tax changes.
This condition is necessary but we propose and then show that it could be
the result of a special maximization behavior. Consequences are interesting
since we can study the pay-as-you-go effects on an efficiency criterium such
as the steady state level and on intergenerational justice criteria such as a
welfare range gap (the egalitarian criterium) or such as the lowest welfare
level (the maximin/Rawlsian criterium).
In the last part, we define an example where functional forms are known.
By using computing simulations, we find there are cases where the lowest
welfare level, the intergenerational inequality and the steady state level
can grow up if the payroll tax rate rises. These results show two interesting
conflicts: a conflict between efficiency (steady state) and equity (welfare
range) criteria and a conflict between two equity criteria (welfare range
and lowest welfare level).

2If the public expenditures consumption can pertfectly substitute for private
good, i.e. if one additional unit of old age public expenditures gives the same
additional welfare as one additional unit of private good consumption, the pay-
as-you-go system is an unfunded pension system. In our numerical example,
old age public good consumption can be perfectly substitute for private good
consumption.
 113

2 Pay-as-you-go system in Reichlin's model

The conditions of existence and stability for a production and overlapping

generations cyclical economy were given by Reichlin [1986]. The economy
consists of two types of agents: two generations of households (young and
old) and a firm. In this paper the new feature is the introduction of a
payroll tax rate denoted T to finance old-age public expenditures.

2.1 THE MODEL

Each generation of households lives for two periods. The population is

constant. During each period t, the younger generation offers It units of
labor and saves its' net wages and the elder consumes its accumulated saving
and old-age public good. Welfare is a sum of private good consumption, old-

° °
age public good consumption and leisure utilities respectively u (c), iL(g)
and v (l) with derivatives satisfying: u' > 0, u" ::; 0, iL' > 0, iL" < and
v' > 0, v" > and where the private good consumption is denoted c, the
old age public expenditures g and the individual labor supply l. Households
have perfect foresights. The wage is denoted Wt and interest factor R t . The
program of each household is:

c~~f. u(ct+l)+iL(gt+l)-v(lt)
{ Ct+l = Rt+1(1- T)Wtlt (1)
gt+ 1 = TWt+1lt+ 1
{=:} Max U(Rt+l(l- T)Wtlt) + iL (gt+d - v(lt) (2)
It

First order condition gives:

Rt+1(1- T)Wt U'(Ct+l) - v'(lt) = ° (3)

Second order condition necessitates:

(Rt+1 (1 - T) Wt)2 u" (Ct+d - v" (It) < ° (4)

Under assumptions on u and v, it's always true.

Technology is Leontiev type Yt = min(olt,J3kt) where a and 13 are re-

spectively capital and labor average productivities. Full capacity utilization
implies: Yt = olt = J3k t . Profit is nil. A unit of output is distributed be-
tween labor and capital as: ~ + 1ft = 1. At equilibrium state, the remaining
output is stored as productive capital:

(5)
114

A little computation gives the following wages and interest factor dynamics:

{
R
i3.(l-T)__
w t ---"'-~
-
k.
1_k'±2
(6)
t+l - 13 (l-T) k.±l

Since Yt+l = f3.kt+l' one has:

(7)

Then, (3) becomes:

(f3. k t+l- 1~Tkt+2) u' (f3kt+l - 1~Tkt+2) = ~ktv' (~kt) (8)

Let U (c) = cu'(c) and V (l) = lv'(l), (8) gives:

f3.kt+l - l~Tkt+2 = U- 1 (V (~kt)) (9)

Let Xt = k t - 1, (9) and balanced budget in each period gives:

kt+l = (1 - T)f3kt - (1 - T)H (~Xt)

{ (10)
Xt+l = kt
gt+l = 1:: T k t+2
with H = U-1oV. The Reichlin's case is T = O. In this case, U (ct+d =
V (~kt).

2.2 STEADY STATE EXISTENCE CONDITIONS:

Proposition 2.1:
Under 13 > l~T [GI] and H(l) - lH'(l) > 0 [G2] or H(l) - lH'(l) < 0
[G2'] , there is a unique positive solution from (10), Ie = ~",,-l (i3g-=:~!1)
with"" (l) = H(l) and Ie = 0 is a trivial solution.

Proof
The fixed points are solutions to (13- l~T)1e = (~Ie) .Positive solution
H
exists if and only if 13 > l~T [GI]. There exists an explicit form for Ie if and
only if"" (l) is monotonic. That is true if H (l) > lH' (l) [G2] , then"" (l) is
increasing or if H (l) < lH' (l) [G2'] , then"" (l) is decreasing. •

Proposition 2.2
Under conditions [GI] and [G2] , the steady state capital increases with
T and under conditions [GI] and [G2'] , the steady state capital decreases
with T.
 115

Proof
ali: _
aT -
1 ./,-11 ( (l-r)!)
(,8(1-T)-1)20/ ,8(1-1'")-1 > 0 ·f . / , - 1 · ·
1 'f"
.
IS Increasmg an
d ali:
Err =
1 ./,-11 ( (l-r)l!)
(,8(1-r)-1)2 0/ ,8(1 r)~l < 0 1·f ./,-1·
0/
d .
IS ecreasmg.
•

3 Permanent business cycle

3.1 CYCLES CONDITIONS

Proposition 3.1
In a neighborhood of the steady state, the dynamics of the economy will
be cyclical if and only if:

(a,{3,r) E {(a,{3,r) E IR _+2 x [0,1[/ (1- r){3 < 2 et H' (~k) =! }[C3].
Proof
In a neighborhood of (k, x), the dynamics of the system are approximated
by:

( ~~::~ ) ~ P. ( :: ) (11)

with dk = k - k, dx =x - x and

There are two conditions on eigenvalues for cycles. The first one is that
eigenvalues module must be equal to 1, (I-r)~H'(~k) = 1 and the second
one is that eigenvalues must be complex (1- r){3 < 2. •

Proposition 3.2
Under [CI], [C2] / [C2'] and [C3] the capital cycle frequency increases
with respect to r.

Proof
Under (C3] , there exists an orthogonal matrix N such as (11) becomes:

( ~~::~ ) ~ N- (~ ~) N (
1
:: ) (12)

where A and Xare the eigenvalues of P.

116

Set () be the argument of the eigenvalues. One has: A = cos () + i sin () and
X = cos () - i sin () with ()= arccos ( (m;T)) . The cycle period T = is 2;
decreasing with respect to () and the frequency () is increasing with respect

V
i!.
to r: ae = 2 ••
aT I -(!3(li-r)r

3.2 STABILITY OF REICHLIN TYPE DYNAMICS

Let fa be a map fa : ( : ) f-+ ( ak - ;a (x) ) defining a Reichlin

type dynamics: (ktH,xtH) = fa (kt,xt). There exists a neighborhood of
a' where IA(a')1 = 1 such that fa is a diffeomorphism mapping an open set
of R+ 2 onto an open set of R+2. Under conditions [Gl] , [G2] / [G2'] ,[G3] ,
-
let A(a) and A(a) the two non-real eigenvalues of P = lOa h' ] . There [a
exists a new basis for dk = k - k and dx = x - x where the system can be
written as:

( ~t+1 ) = ( ReA(a') ImA(a')) ( kt ) _ ( (a(Xt)) (13)

XtH -ImA(a') ReA(a') Xt 0

.th
( ~t
Xt
) = (si!e -~:) (
0 1
dk t
dXt
)

6 ) (si8 e -1::) ( ~a~Xt)

{
WI ( (a(Xt = )

where ~a (Xt) = ~h: (x) (Xt - x)2 - th:' (x) (Xt - x)3 +0 (Xt - x)5) .

Let ~a be the map defined by (13), ~a is one parameter family of

diffeomorphism on R2 satisfying the following conditions for a in near a':
(i) ~a(O, 0) = (0,0)
(ii) d~a(O, 0) has two non real eigenvalues A(a) ,A (a) with IA (a')1 = 1
(iii) ~do
>0
m
(iv) (A (a')) #- 1 for m = 1,2,3,4,5.

Write ~a' = ~ and A(a') = A. By setting z = k + ix, ~ can be written as:

 117

More precisely, for Reichlin type dynamics, we have:

Theorem 3.3 (Lanford-Wan)

<p(z) = N<p(z) + 0 (lzI5), where in polar coordinates, N<p : (r,4» 1-+

( (1 + ii) r - II (a) r 3 , 4> + B (a) +h (a) r2) with (1 + ii) = I). (a) I . Further-
more, II (a) is given by:

Theorem 3.4 (Ruelle-Takens)

If lI(a') > 0, then for all sufficiently small positive ii, <Pa has an
attracting invariant circle.

Corollary 3.5
In the case of a dynamic system (10) the condition for an invariant
2
attractive circle is (Ha) - H::' > 0 [G4], in a neighborhood of a', with
1/

a' = (1 - T) {3.

3.3 PERMANENT BUSINESS CYCLE CONDITION

When T changes, the [G3] cycle condition cannot stay verified. That's
why, we propose to endogenize 0: with respect to T and {3.

Proposition 3.6
Under [G1] , [G2] / [C2'] and y1j;-l1 (y) H" (1j;-1 (y» > -H' (1j;-1 (y»)
[C5] or y1j;-l1 (y) H" (1j;-1 (y») > -H' (1j;-1 (y») [G5'] where y = ~g=:~!l '
if 0: = (1 - T) {3r (s) with r (s) = e-~ (s)' s = ~(l-\)-l and e (y) =

yH' (1j;-1 (y») then [G3] is always true.

Proof:
e (y) must be inversive. If e' (y) > 0 or if e' (y) < 0, that's true . •
How to justify this relation? A simple way is to suppose that it is the
result of a very special maximization behavior. That's why, we add these
new assumptions:
i) there exists a relation between labor productivity and collective effort;
ii) workers prefer to have or not to have a pay-as-you-go system;
118

iii) workers choose effort level and labor supply independently. They
know that minimal effort is not optimal because there is a collective but
not individual positive link between wages and effort.

Proposition 307
Under [CIJ, [C2J / [C2'], [C3J and [C5J / [C5'J there always exists a
functional form of utility Z(e,{3,r) = -~a(e)2 + (1- r){3r (i3(l_lrj_l)
a (e) where a (e) is the relation between effort and labor productivity with
0:' (e) > 0 and a" (e) ~ 0 such as the optimal condition 8Z(~f'T) = 0
implies cycle.

Proof:
One can verify easily that 8Z(~f'T) = 0 gives cycle condition a
(1 - r),8 r (s). Moreover, this solution is optimal because 82Z~:f'T)
_0:'2 (e) < O. •

This condition is restrictive but it is necessary for our purpose. With

this additional utility form, the economic behavior of households offsets
enough changes of r by raising or decreasing their effort to maintain cycle
condition. From now on, as a = (1 - r),8r (s), under [CIJ, [C2J / [C2'J,
[C3J , [C4J and [C5J / [C5'], steady capital level change with respect to r
is given by:
dk 8k + 8k 8Of.
dT 8T 8Of. 8T
1 -11 ( (1-T)~) (l-T);& -11 ( (1-T)~ )
(,6(1-T)_1)21/; i3(l-T)-l - (i3(l-T)-l) 1/; i3(1-T)-l

x (-{3r(s) + (i3g~~l!:1)2r'(S))
(i3(1-~)-1)2 (1 + (1 - r) ~ (,8 (1 - r) - 1)

(,8r(s) - (i3g~~)~1)2r'(S))) x 1/;-11 (i3g~~!l)

4 Example
4.1 THE MODEL

Suppose v(l) = ~l2 + bl; u(c) = c and 3 u(g) = >.g. The household
economic behavior is the solution to Max u(ct+d + u(gtH) - v(lt} +
e,ct+l,lt

3If >. = 1, public good and private good are perfectly interchangeable and the
pay-as-you-go system is equivalent to an traditional no contributive unfunded
pension system. So T can be considerd like an intergenerational transfert rate.
 119

I Z ( e ,T) such that Ct+1

+,-
= Rt+ 1 (1 - T )Wt It, gt+1 = TWt+1lt+1 with
2
Z(+~_, T) e + 2(~~2;:~Ve and
= - 2 Q (e) = De, the relationship between effort
and productivity, The second order condition are always verified because
u" (c) - v" (l) = -a < 0 and z" (e) = -1 < 0,

Hence, by using capital per head dynamic equation (10), we can write:

kt+2 = (1 - T){3kt+1 - (1 - T)~X~+1

{
Xt+2 = kt+1

and e(T) = e= 2("1t :::'td' There exists two steady states A et B:

A= {kiX} = {OiO} et B = {kiX} = {(1-r)[/3-b(~)1-1i (1-r)[/3-b(~)1-1}
a(1-r)( ~) a(1-r)( ~)

In a neighborhood of the steady state B, the Jacobian matrix of this

system is:

IJI = ~(1 - T) [2ax + b] = (1 - T){3 (2 - ~) - 2

The dynamics of the economy will be cyclical if and only if

• Q = 2~~~~j3' This relation allows to obtain eigenvalues module

equal to 1 in absolute value, If Q = De, permanent cycle condition

necessitates a1 = - (3 i bl = 2,62- 3 i a2 = b2 = §j,

• 2(1~r) < {3 < (l~r)' This last condition allows eigenvalues to be

complex and Q > 0, However, there is not (3 verifying this condition
'-'T' 1,5(b+r)/(I-r) < '" < 2(b+r)/(I-r) In thl'S case the cycle stabl'll'ty
v, 2/3(1-r)-3 .... 2/3(1-r)-3' ,
is always verified because: 4~«11~~~2· > 0,

4,2 COMPUTING SIMULATIONS

Figures 1, 2 and 3 illustrate these purposes for the following parameters

values:
120

Utility parameters
1a b or d·IsuttTIty J a = 0.003
b = 0.1
public good utility ). = 1
effort utility 'Y = 100
capital productivity (3 = 2.1
Production parameters ~
labor productivity and effort 8 = 1
Initial values ko = kl = 1000

~~----------------~~~----~

200

-200

-~

-~+--------,----~~,---------,-~

Figure 1 - Capital per head phases diagram for T = 0.26, ... ,0.268

lOCO . , . - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,

7SO

500

250

-250

-500

-750

-lOCO

-1250 +--------,--------....----------r----'
o lOCO

Figure 2 - Private Good Consumption per head phases diagram for

T = 0.26, ... ,0.268
 121

em

100

200

-200

-100

-em
0 1000 1500

Figure 3 - Life Cycle Welfare phases diagram for T = 0.26, ... ,0.268

Figure 3 gives welfare phases diagram. The elliptic forms moving to the
right characterize the T rising. That's also true for figures 1 and 2. Two
sorts of conflict are interesting: a conflict between efficiency (variation of
the steady welfare state level which is identified by the center of a elliptic
form) and equity (variations of the welfare range) criteria and a conflict
between two equity criteria (variations of the welfare range and variations
of the lowest welfare level). Indeed, when T rises, the graphs on figure 3 show
that the steady welfare state level, the lowest welfare level and the welfare
range rise. Therefore, the Rawlsian criterium consisting in maximizing the
lowest welfare level is increasing with respect to T. However, the egalitarian
criterium consisting in reducing the intergenerational welfare range is
decreasing with respect to T. Finally, the efficiency criterium consisting
in maximizing the steady welfare state level is increasing with respect to T.
Time paths are not perfectly periodic because this model is not linear
and chaotic time paths exist. It is why on the graphs, we can see that all
the values of the interval [k min ; kmaxl seem to be on the path. Then, discrete
paths seem to look like to continuous time cycles.

5 Conclusion

In this article, we obtain a permanent business cycle by endogenizing

the labor productivity with respect to payroll tax rate to conserve the
business cycle. The permanent endogenous cycle allows us to contribute to
122

the debate about equity and efficiency of the pay-as-you-go system when
the payroll tax rate is modified in an original fashion. Indeed, by using
computing simulation, an increase in the intergenerational transfer rate i.e.
the payroll tax rate can have two effects on welfare:
i) first on steady welfare state i.e. the efficiency reference level which
rises;
ii) second on welfare range. If the range rises, the increase of T is not
improving according to the pure egalitarian criterium. On the contrary,
if the increase of one half range is inferior to the increase of steady
welfare state, the increase of T is improving according to Rawlsian criterium
because the lowest level of welfare rises.
One of our next perspectives of study consists in specifying analytically
these latter found effects by using computing simulations. Notably, it can
be done by writing economic variables time paths (capital, consumption
and welfare) in the following way: Xt = x + A cos (Bt + <p) where the range
is denoted A , the frequency B, the phase <p and the capital, consumption
or welfare steady state x.
Another perspective consists in suppressing the permanent economic
cycle condition and in investigating all chaotic time paths with respect
to T.

References
[1] ALLAIS M. [1947], Economie et interet, Imprimerie Nationale, Paris.
[2] BARRO R.J. [1974], "Are Gouvernement Bonds Net Wealth?", Jour-
nal of Political Economy, 82, 1095-1117.
[3] BENHABIB J. and R. DAY [1982], "A Characterization of Erratic Dy-
namics in the Overlapping Generations Model", Journal of Economic
Dynamics and Control, 4, 37-55.
[4] BREYER F. and M. STRAUB [1993], "Welfare Effects of Unfunded
Pension Systems when Labor Supply is Endogenous", Journal of
Public Economics, 50 , 77-91.
[5] DIAMOND P. A. [1965], "National Debt in a Neoclassical Growth
Model", American Economic Review, 55 , 1126-1150.
[6] GALE D. [1973], "Pure Exchange Equilibrium of Dynamic Economic
Models", Journal of Economic Theory, 6 , 2-36.
 123

[7] GALOR 0 [1992], " A Two-Sector Overlapping Generations Model: A

Global Characterization of the Dynamical System" , Econometrica, 60,
1351-1386.
[8] GRANDMONT J.M. [1985], "On Endogenous Competitive Business
Cycles", Econometrica, 53,995-1045.
[9] LANFORD O.B. [1972], "Bifurcation of Periodic Solutions into In-
variant Tori: the Work of Ruelle and Tackens", in Nonlinear Problems
in the Physical Sciences and Biology, Lectures Notes in Mathematics,
n0322, Springer-Verlag, 159-192.
[10] PHELPS B.S. and J. C. RILEY [1978], "Rawlsian Growth: Dynamic
Programming of Capital and Wealth for Intergenerational "Maximin"
Justice", Review of Economic Studies, 45, 103-120.
[11] RAWLS J. [1971], A Theory of Justice, Harvard University Press,
Cambridge.
[12] REICHLIN P. [1986], "Equilibrium Cycles in an Overlapping Gener-
ations Economy with Production", Journal of Economic Theory, 40,
89-102.
[13] RODRIGUEZ A. [1981]" "Rawls's Maximin Criterion and Time Con-
sistency", Review of Economic Studies, 48, 599-605.
[14] RUELLE D. and F. TAKENS [1971], "On the Nature of Turbulence",
Communications in Mathematical Physics, 20, 167-192.
[15] SAMUELSON P.A. [1958], "An Exact Consumption-Loan Model of
Interest with or without the Social Contrivance of Money", Journal of
Political Economy, 66, 467-482.
[16] WAN Y.H. [1978], "Computation of the Stability Condition for the
Hopf Bifurcation of Difeomorphisms on ]R2" , SIAM Journal of Applied
Mathematics., 34 , 167-175.
[17] ZHANG J. [1995], "Social Security and Endogenous Growth", Journal
of Publics Economics, 58 , 185-213.
 Part 3

Keynesian Models
Relaxation Cycle, Chaotic
Dynamics and Limit Cycle: a
Model with Keynesian
"Flavour"
Gilbert ABRAHAM-FROIS and Edmond BERREBI

1 Introduction

R. Goodwin (1951) was probably the first to introduce "relaxation

cycles" in macroeconomics; Hicks' analysis is contemporary and indeed
previous (1950), both authors introducing "non-linear accelerator" in their
analysis; but Goodwin' originality is obvious since he has used relaxation
cycle and Lienard-van der Pol equations (as soon as 1933, Ph. Le Corbeiller
had insisted on their interest and relevance). Further work has shown that
in many cases relaxation cycle could be considered as a stable" attractor" ,
a rather peculiar "limit-cycle", named in some cases (Chiarella 1990)
"limiting limit cycle".
In other respects, one knows that cubic function can give rise to chaotic
dynamics; in previous work - Puu (1991) and extension by Abraham-Frois
and Berrebi (1995) - it appears possibility of having - simultaneously!!
- chaotic dynamics and relaxation cycle. Moreover, in Day and Schafer
(1986), IS-LM model can produce chaotic dynamics. The aim of this paper
is to give further extension to this kind of analysis. Some interesting results
appear: first, the relationship between cycle and chaotic dynamics is more
complex than expected; second, the relaxation cycle is also an attractor,
a limit cycle; third, this particular cycle is locally stable, so that one can
contemplate some connection between exogeneous shocks and endogeneous
cycles.
128

2 A macro-model with keynesian flavour

Puu's model is a "keynesian standard" type model and may be situated

as an extension of previous analysis by Samuelson (1939), Hicks (1950) and
Goodwin (1951). Consumption and investment function are rather specific.
More precisely, Puu assumes that investment at time t, It is a non-linear
function of finite differences of incomes at (t - 1) and can be written as:

The graph has the following characteristics: it goes through the origin
with slope v; this slope decrease slowly till stabilization; it turns over
after for higher differences of incomes, which does not seem quite relevant.
Moreover, in case of decreases of income, this function gives rise to negative
investment, which is at least annoying.
To avoid these problems, we suggest to introduce an autonomous invest-
ment so that the investment function now writes:

(la)
There is no loss of generality by writting 10 = 1, since this means a change
of unit in income level; hence the investment function suggested :

It = v (Yt-l - Yt-2) - v (Yt-l - Yt_2)3 + 1 (lb)

By writting Zt = AYt = Yt - Yt-b one obtains:

It = VZt-l - VZ;_l + 1 (2)

Investment function is a cubic, which is essential for following analysis.

Figure 1: Investment with cubic function and autonomous investment

 129

A second assumption, quite central here, concerns the consumption - or

saving - function. s being the propensity to save, one makes the assumption
that consumption in period t depends upon two elements: the first one
is the level of income of (t - 1), the second being the amont of saving
in (t - 2). Part c of saving sYi-2 is consumed in t; part (1 - c) sYi-2 is
never consumed, so that one can define (1 - c) s as the "rate of eternal
saving" (the exact value of this parameter is not very important; what
really matters is to know wheter c is equal to or different from one).
So, consumption at t is the sum of part of income of t - 1, (1 - s) Yi-I
and of part of saving appeared in (t - 2) , csYi-2 hence

C t = (1 - s) Yi-I + csYi-2 (3)

or

Ct = Yi-I - S (Yi-I - Yi-2) - (1 - c) sYi-2 (4)

or

Ct = Yi-I - SZt-1 - (1 - c) SYi-2 (5)

From the equilibrium relation,

Yi = Ct + It (6)
on gets, by replacing Ct and It by their value in 5 or 3 :

Yi = Ct + It (7)
= Yi-I - sZt - (1 - c) sYi-2 + VZt-1 - VZ;_I + 1 (8)
<=>
Yi - Yi-I -sZt - (1 - c) SYi-2 + VZt-l - VZLI +1
<=>
Zt -sZt - (1 - c) sYi-2 + VZt-l - VZ;_l +1
One can write q = v - s, difference between accelerator v and propensity
to consume s. So this equation can now be written as :

Zt = qZt-1 - (q + s) ZLI - (1 - c) sYi-2 +1 (9)

We have now to distinguish the case where rate or eternal savings is

equal to zero and from cases when it is different from zero becomes smaller
and smaller.
130

3 When" eternal rate of saving" is equal to zero:

(l-c)s=O

In this case, on which we will skip rather quickly, one finds a kind of
chaotic dynamics which is rather well known; main characteristics will be
shown as follows. Relation (8) becomes:

Zt = qZt-1 - (q + s) Zl_l + 1 (8')

which is a first order difference equation in Z. In this model, sYi-2 appeared
in period t - 2 is totally consumed in t : as a matter of fact, in each period,
one saves s per cent of current income which is left untouched next period
and entirely consumed two periods later. The evolution of the general term
Zt of the recurrent series (Zt) can be studied from the properties of I
function and of its iteratives p, ... , In, ... , i.e. orbits of function I defined
by:

z --+ I (z) = qz - (q + s) z3 +1 (10)

Particular points of function f graph
I function takes for
• z = 0 value 1(0) = o· q - o· (q + s) + 1 = 1
• z = 1 value 1(1) = 1 . q - 1 . (q + s) + 1 = 1 - s
• z = -1 value I (-1) = -1 . q - ( -1) . (q + s) + 1 = 1 + s
So graphs goes through points (-1,1 + s), (0, 1) and (1,1 - s).
Critical points of I function
Derivative l' (z) = q - 3 (q + s) z2 has two roots,
zi = J 3 (q ~ s) et zi = - J (/+3 s) which are the two critical points

of function f. Since, I (zi) = [q - (q + s) (zi)2] zi + 1 = 2~J q + 1

3v3 q+s
and f"(zi) = -6(q+s)zi < 0, point (zi,l(zi)) is a local maximum;
in the same way, I (zi) = I (-zi) = - [q - (q + s) (zi)2] zi + 1 =

- 2~J
3v3 q + s
q + 1 and since 1" (zi) = 6 (q + s) zi > 0, point (zi, I (zi)) is
a local minimum.
r
All terms of the series {z, I (z) ,12 (z), ... , (z) , ... } where z > 0 will
be positive if local maximum I (zi) < Zl where Zl is the positive real root
of I (z) = qz - (q + s) z3 + 1 = o.

Special case 10 = 0
 131

When in equation (la) one choses 10 = 0, it can be shown that for

r
q < 2.59809, any orbit generated by 1, {z, 1 (z) ,1 2 (z) , ... , (z) , ... } has
all its terms positive if z belongs to interval [0, Zl] and all its terms negative
if z belongs to interval [- Zl, 0]. On the contrary when q > 2,59810
r
mapping {z, 1 (z), P (z), ... , (z) , ... } has positive and negative terms
since some points in the interval [0, Zl] which are such that 1 (Zl) > Zl
and consequently P (z) = 1 [1 (z)] < O.
In this first case, this means that, according to the initial value of z
(which means - let us remind ... - a change in income level), one has either
monotonous increase, either monotonous decrease in the level of income.
For q = 2,59809 ... it appears a bifurcation in the system with alternance
of periods of expansion and recession.
We will return, later on, on the importance of this distinction. It may be
useful to draw the "bifurcation diagram" .

Bifurcation diagram for function 1 with 10 = 0

Let us plot the bifurcation diagram for 1 function; on the x-absciss, we
report the values of parameter q which can vary between 0 to 3 and along
r
the y-absciss the 350 points of {z, 1 (z), P (z), ... , (z), ... } for function
1 associated to the value of different values of q end for values of n going
from 150 to 1500.

Figure 2: Bifurcation diagram for q E [0,3]

Nota: when 10 = 0, the model's dynamics depends only on the value of

q = v - s. On the other hand, when 10 = 1, stability of q does not exclude
a change in the system's dynamics when stakes differents values.
132

4 When" eternal rate of savings" is not equal to

zero: (1 - €) S =I 0 without however being
insignificant ...

The assumption of a non-zero eternal rate of savings modifies deeply the

dynamic of the system, the difference equation now writes:

Zt = [q + (1 - c) sJ Zt-l - (q + s) zl-l - (1 - c) SYt-1 + 1 (11)

The economic interpretation is that is such case saving SYt-2 of (t - 2)

is now divided in two parts:

• cSYt-2 which is consumed in period t

• (1 - c) sYi-2 which is saved until the end of the program.

This entails deep changes in the dynamic evolutionj there will appear
cycles of different kind, which may be associated - or not - with chaotic
evolution; moreover, another difference appears: in the first model, the
general evolution depended only on q = v - Sj now, it depends first on v et
second on (1 - c) s, which obviously means changes in q.
From equations (2) and (11), one gets the following system of two
equations :

(8) { Yi = Yi-l + Zt
Zt = [q + (1 - c) s] Zt-l - (q + s) Zl-l - (1- c) SYi-1 +1
One can represent solution of (8) system in the phase diagram (Yi, Zt)
where Zt = Yi - Yi-l. The phase diagram is here of the discrete type - using
difference equations - when usual presentations uses differential equations.
We shall figure too the time-sequence (Y1 , Y2 , .•• , Yi) for particular values
of accelerator v and to eternal rate of saving (1 - c) s.
In previous workl, we have studied the case where v took different
values for (1 - c) s constant; we shall insist specially on the situation where
(1 - c:) s tends to zero.
For a special value of v, say v = 1.75, one can depict the phase diagram
(Yi, Zt) and the time sequence Yi corresponding to different values of eternal
rate of saving (1 - c:) s.
According to the values of (1 - c) s, it appears for the phase diagram of
system (8) either:

lCf. Abraham-Frois and Berrebi [1995].

 133

• a finite number of points

• a thin continuous curve

• a thick continuous curve
• a laminated surface

So, it appears, for

(1 - c) s Forms Figure
0.08 closed curve 3a
0.25 butterfly 3b
0.45 clouds 3c
0.60 signature 3d
0.70 bracelet 3e
0.85 separate segments 3f

Those phase diagrams presented on fig. 3 are all strange attractors since,
when t goes to infinity, points (Yt, Zt) go back to these attractors without
regularity.

oj •••• . .t:J . . . . . . ······x:

Fig.3a Fig.3b

......•

Fig.3c Fig. 3d

.....

Fig.3e Fig.3f
134

-
~··········-···~·········r···························· .~..............:.ii ' .. :~... "'" . • •.............. "x"

ig.3g i"ig. ::Jh

Fig.3i
Figure 3: Phase diagram for v = 1.75 and (1 - c) s variable

An another hand, for numerous values of (1 - c) s, one has a finite

number of points on the phase diagram which means that one obtains
a periodic time sequence (¥i) .
For instance, it comes for :

(1 - c) s Periodic time sequence

0.20 period 18
0.35 period 12
0.41 period 10
0.55 period 8 (fig. 3h and 3i)

5 When eternal rate is trifling: (1 - c)s turns to 0

Let us look at model (8)

{ ¥i = ¥i-I + Zt
(8) Zt = [q + (1 - c)S]Zt-1 - (q + S)Zl_1 - (1 - c)S¥i-1 + 1
when (1 - €)s turns to O.
There appears a cubic with bifurcations and chaotic zone. This cubic is
symetrical around point ((I!e)s' 0) since image of point ((1!e)s - ¥i-I
, -Zt-d is a point with abscissa and ordinate:

- [q + (1- c)s] Zt-I + (q + s) Zt-I - (1- c) s [(I!e)S - ¥i-I] + 1

= - [(q + (1 - c)s) Zt-I - (q + s) Zt-I - (1 - c)S¥i-1 + 1] = -Zt
 135

So part part corresponding to Zt < 0 (which lies under the ayt axis) is
symetrical of part corresponding to Zt > 0 (which lies above the ayt axis)
around point ((l!£js' 0). Consequently, one shall look at the superior part
of the cubic and afterwards find its symmetry around point ((l!e)s' 0).

5.1 NEW TYPES OF CYCLES APPEAR NOW: RELAXATION

CYCLE

When (l-E)s turns to 0 and for instance has for value 0.0001, there ap-
pear "relaxation cycles" with or without chaotic dynamic. When starting
from point (Yo = (l!e)s = 10000, Zo = 0.5) different cases come to evi-
dence:

• If v = 1.5 system (8) gives rise in phase diagram (yt, Zt) to a

relaxation cycle limited by straight lines Y = 8880 and Y = 11120.

...................
r································8······
~.

Figure 4: Relaxation cycles for v = 1.5

• If v = 1.75, there appears a relaxation cycle and a more interesting

dynamics with appearance of a double bifurcation system. Cycle is
then limited by straight lines Y = 8100 and Y = 11900 .

~
~ ...... -. ................ ~ ................ . ,....;c.

Figure 5: Relaxation cycles and straight lines for v = 1,75

136

• If v = 1.87 and always (1 - e}s = 0.0001, relaxation cycle still exists

with many more bifurcations. This cycle is limited by straight lines
Y = 7700 and Y = 12300.

Figure 6: Relaxation cycles for v = 1.87

• A last, for v = 2.12 and still (1 - e}s = 0.0001, we have again

relaxation cycle and much richer dynamics: there appears double
bifurcation processes with extension to dynamic zones. There are two
limits, straight lines Y = 6855 and Y = 13145 .

.
. ~. .

r- .. - . .
. .

Figure 7: Relaxation cycles for v = 2,12

5.2 WHY RELAXATION CYCLES?

The appearance of relaxation cycle is easy to explain: in the phase

diagram in (yt, Zt), ~~: ~ 00 when (1 - e}s ~ O. As matter of fact,
one has:

and with X t = (1 - e}syt one finds:

Zt - Zt-1 Zt - Zt-1 1 [Q+(1-E)s-ljZt_l-(Q+S)Z: 1 -Xt - 1 +1

= =---
X t - Xt- 1 (1 - e}sZt (1 - e}s Zt
 137

Since term {1!e>s becomes very large when (1 - £)8 -0, ratio ~~! =
;:=~:::.~ -00 except if [q+ (1-£)8 -lJZt-l - (q+8)Z{_1 - X t - l + 1 = 0.
Let us consider the phase diagram (Xt, Zt) or (yt, Zt) with representation
of unity on Y axis proportional to (1 - £)8.

• If [q+(1-£)8-1]Zt-l -(q+8)Z{_1 -Xt-l +1 f; 0, the representation

of the curve is given by a vertical line situated close to the extrema
where IZtl > 1.2 .
• On the other hand, if[q+(1-£)8-1]Zt-I-(q+8)Zt-I-Xt-l +1 = 0,
X t - l = [q + (1 - £)8 - l]Zt-l - (q + 8)Z{_1 + 1 can be represented
by a cubic in the phase diagram (Xt ,Zt).

Cubic function X = [q + (1 - £)8 - l]Z - (q + 8)Z3 has a derivative:

[q + (1 - £)8 - 1J - 3(q + 8)Z2 which cancels for two values of Z :

q + (1 - £)8 - 1 Z- q+(1-£)8-1
Zl= 3(q + 8) et 2- - 3(q + 8)

and since second derivative is -6(q + 8)Z:

• The point (Z}, Xd where

Xl = [q + (1 - £)8 - 1 - (q + 8)Zr] Zl + 1
= [q + (1- £ )8 - 1- q+(I-e)S-I] .
3 V/q+(l-e)s-l
3(q+s)
+ 1

2 q+(I-e)s-13 +1
(q+s)
3y'3

is a local maximum.

X 2 = [q + (1 - £)8 - 1 - (q + 8)Z~] Z2 + 1
=- [q + (1 - £)8 - 1 _ q+{1-;e)s-l] q+(l-e)s-l + 1
3(q+s)
2 ,------
-3y'3

is a local minimum.

Let us notice, by coming back to (Y}, Zl) and (Y2, Z2), and since
X = (1 - £)8Y, that Y values becomes very large and would go out of
the diagram if we had not taken for unity (l~e)s on Y axis.
138

5.3 ATTRACTION BASIN AND LIMIT

As a matter of fact, relaxation cycles of figures 4 to 7 are part of cubic
appearing in figures 8 to 11.

Figure 8: Cubic for v = ~.5

Figure 9: Cubic for v = 1. 75

····x·

Figure 10: Cubic for v = 1.87

Figure 11: Cubic for v = 1.12

 139

There are also the limits of the (yt, Zd wich are solution to (8) system,
where initial value (Yo, Zo) is situated on the cubic or in its immediate
neighbourhood. Cubic of figures 8 to 11 (and their immediate neibourhood)
constitue the attraction basin for (8) model. Let us look, for instance, to
figure 8 where cubic associated to (1 - c)s = 0.0001 and v = 1.5. Starting
from point (Yo, Zo) = (1820, 0.5) situated near the left extremity of the
cubic, we can see that the (yt, Zt) of (8) model goes the right, first rather
slowly (since there are some Zt < a and some Zt > 0, there having larger
influence than the negative ones) then more quickly when all Zt come
to be positive and consequently yt are always increasing. This evolution
continues to the neighbourhood of point (Y}, Zt} where Y1 = 11120 is the
cubic local maximum. In point (Y}, Zt}, ~f tends to infinity and points
(yt, Zt) move to the half-straight line going down from (Yl> Zl)' They are
attracted by the inferior part of the cubic which they join at point C1 .
In CI, Zt < a and points (yt, Zt) decrease on the cubic to point (Y2, Z2)
where Y2 = 8880 and ~f tends again to infinity. From this point (Y2, Z2),
points (yt, Zt) of the model move to the half straight line up to point C2
situated on the cubic and after that move on, keeping always on the cycle
(Y1 , Zl)/CI/(Y2 , Z2)/C2 which stands as a limit cycle for (8) model.
If now v goes throught different values (say 1.75, then 1.87 and at last
2.12), keeping (1 - c) s = 0.0001, one can see a decreasing amplitude of the
cubic; on the other hand, the amplitude of the relaxation cycle -enriched
successively by bifurcations and chaotic zones- goes increasing (cf. figures
9, 10, 11).

5.4 CUBIC'S FRONTIER

Let us come back to (8) system represented by a cubic with two

symetrical parts around point ((l.!e)S' 0) .
The upper part is limited to the
rightside by point (Y1 , Zl) local maximum of the cubic precedently defined
; on the left such there comes a chaotic zone with a frontier Y for v, sand
8 given. It's obvious (cf. supra) that coefficient of Zt-l in system (8), i.e.
q+ (1 - c) s = v- (1 - c) s must be less then 3 ; if it is more than 3, Zt goes
to infinity. To calculate Y, one draws the bifurcation diagram for the series
Zt = [v - s + (1 - c) s]- VZr-l + b in a diagram with b = -(1 - c)sY + 1
and as ordinates the stable equilibrium points of Zt. b varies from a to 1
and the program stops when it comes to a value Ii < 1 wich is !"eached for Zt
finite and Zt+1 getting out of the diagram. So that Y = (l~e)s' The upper
part of the cubic is defined for Y ::; Y ::; Y1 = (l~~)S and the (symetric)
inferior part is defined for (l~~)S = Y2 ::; Y ::; - (l~e)s - Y.
SO for (1 - c)s = 0.0001 and v = 1.75, one gets the diagram of figure 12
140

and Y = 3400 since Ii = 0.66.

Figure 12: Bifurcation diagram for v = 1.75 S = 0.0011 - e = 0.1.

So the upper part of the cubics goes from Y = 3400 to Yl = 11890. The
lower part (symetrical from the upper part around point ((1!E)8 = 10000)
goes from Y2 = 8110 to 20000 - Y = 16600. The relaxation cycle, which
appears as a limit cycle of model (8) goes from 8110 to 11890.
If the starting point is close to the relaxation cycle, the representation
of the series (¥t, Zt) is given in the phase diagram by the relaxation cycle
(in the paper, one choses Yo = {1!E)8 and Zo = 0.5 when one wants that
the successive points (¥t, Zt) to be situated on the relaxation cycle).
On the other hand, if the starting point of the cubic (Yo, Zo) is located
to the left (resp. to the right) frontier of the relaxation cycle and after that
to the cycle itself on which one travels indefinitely (figures 13 and 14).

!~
..
~';
;~ ~
~ ............ - •••••••••,
. .................... •.............................K..
.
.f
I '
~

Figure 13: Cubic for v = 2.1, (1 - e)S = 0.001 et Yo = 570

1'!Jl;:;;'························x

Figure 14: Cubic for v = 2.1 (1 - e)S = 0.001 et Yo = 1470

 141

5.5 RELAXATION CYCLE WITH THICK FRONTIERS

For (1-C:)8 given and negligible, it has been observed inside the relaxation
cycle a chaotic zone growing in importance with v. If the starting point
(Yo, Zo) is located on the superior part of the cubic, the series (Yi, Zt) is
drawn (when arriving to the local maximum (Yt, Zl) where g
--+ 00), to
the chaotic zone of the inferior part. Let us suppose that in this chaotic
zone, one has some Zt > 0 and some Zt < 0, with relative influence quasi-
equivalent, but slight avantage for Zt > 0; consequently, terms (Yi, Zt)
will stay during a "certain time" in a vertical strip [Y1 Y1] before going
left to the local minimum (Y2, Z2) where ;~ comes to infini~. Successive
terms (Yi, Zt) will then linger, symetricaliy, in a vertical strip [Y2 Y 2 ] before
going right. So there will appear a relaxation cycle with two thick frontiers
[Y1 Y1] and [Y2 Y2 ] (cf. figure 15).

............................•.

Figure 15: Relaxation cycle with thick frontier

We have precedingly given some privilege to diagram representation; now

we shall call attention to time series of amounts and revenue change.

• The revenue level oscillates mainly between two values (" high" and
"low" level) ; around these two levels, there are minor fluctuations
and one could think that they are due to exogenous shocks , the
length of "high" and "low" level stages are quite variable.
• Moreover, the stage of quick changes which appear on these diagrams
have no connection with the "fast dynamics" which appeared in the
cubic analysis. It is just the other way; as is appears in the simulation,
it is the stage of "slow" dynamics (in the phase diagram) which
corresponds to the abrupt change in the time series. The two levels of
revenue (" high" and "low") have variable length because the system
wanders then through the" chaotic" part of the curve.
142

Figure 16: Variation of revenue level yt through time series

The above time-series is specially interesting ; first lenght of different

stages are quite variable , second, there is abrupt change without any
exogenous shock from one situation to the other:

• the chaotic behavior becomes clear when one examines the change in
revenue level Zt = Doyt; all precedent diagrams have been drawn in
(yt, Zt) coordinates; it is the revenue change of Zt which suffers most
important and unexpected changes. Revenue level yt is sul;ljected to
strong vibration with oscillations less important than these appearing
for Zt.

Figure 17: Variation of level-change Zt through time

5.6 WHEN RELAXATION CYCLE DISLOCATES

Decomposition of the cubic: appearance of two limit strips

v continue to increase and the starting point (Yo, Zo) is situated in the
upper hand of the cubic j when the system arrives in the local maximum
(Y1 , Zl) where g.is infinite, it is attracted into a chaotic zone of the inferior
branch where some Zt > 0 and some Zt < 0 get mutually neutralized. If the
starting point (Yo, Zo) is situated on the inferior branch of the cubic, close
to its right frontier (which is a chaotic zone with some Zt > 0 and some
Zt < 0), successive terms (yt, Zt) go left until they enter into the vertical
strips [Y1 }\] where they will remain. This vertical strip [Y1 Y'd appears
 143

as the limit of the series (yt, Zt) solutions of model (8) when the starting
point (Yo, Zo) is either on the upper part of th cubic with Yi > Yo > Y2
either on the lower branch with Yo> YI
If the starting point (Yo, Zo) is situated either on the lower part of the
cubic with Y2 < Yo < YI or on the upper part with Yo < Y2, point (yt, Zt)
solutions of (8) have for limit a vertical strip (y2 Y2 ] symetrical to the strip
[YI YI ] around point [(I!E)8; 0] .
So relaxation cycle has burst out ; limit of series (yt, Zt) depends
on initial conditions. For instance, for (1 - c)s = 0.0001, we have such
situation when 2.136 < v < 2.263; if v = 2.263, [YI , YI ] = [13623,13716]
and [Y2 , Y2 ] = [6284,6377].

Disappearance of relaxation cycle: divergent series (yt, Zt).

If v still increased and with starting point [Yo = (I!E)8; 0.5], on the upper
part of the cubic, series yt still increase, with Zt > 0, to local maximum
(Yt, Zt) where g. tends to infinity.
Two possibility appears:
a) YI < (I:E)8 - Y and system falls down to the chaotic zone of the
inferior part where negative Zt have less influence than positive Zt.
Zt will ultimately overshoot critical value Z and explodes, hence a very
fast oscillating divergence of the series (yt, Zt).
b) YI > (I:E)8 - Y ; points (yt, Zt) are no more drawn towards the cubic
since they are too far from it. In such case, one has too very quick divergent
oscillation (cf. figure 18).
As a matter of fact, we have just some points on straight lines coming
from (YI , ZI) or (Y2, Z2) before getting to a very quick divergent oscillation
in series (yt, Zt) (cf. figures 18 et 19).

Figure 18: v = 2.8 s = 0.01 (1 - c)s = 0.1 and Yo = 1030

144

Figure 19: v = 2.88 = 0.01 (1 - C:)8 = 0.1 and Yo = 960

A central limiting strip for model (8)

With (1 - e)8 = 0.0001 and v > 2.264, relaxation cycle comes to disloca-
tion. The starting point is (Yo, Zo) = (l~e:)B' 0) situated on a straight line
going through the symetrical point of the cubic (l':e:)B' 0). In the close
neighbourhood of the straight line Y = (l~e:)B = 10000, one has a dynamics
quite similar to the one which appears in the bifurcation diagram of the
series :

Zt = [v - 8 + (1 - C:)8] Zt-l - vZt_l

wich appeared in figure 2. So with q = v - 8 + (1 - e )8, one knows that one
obtains

• for Q < 2, a limit point

• for 2 < Q < 2.2357, stable period 2 cycle
• for 2.2358 < Q < 2.287, stable period 4 cycle
• for 2.2871 < Q < 2.3029, stable period 8,16, ... 2n cycle
• for 2.303 < Q < 2.59809, chaotic dynamic with Zt > 0
• for 2.5981 < Q < 3, chaotic dynamic with Zt > 0 and Zt < O.
Obviously, if Q < 2.59809 and (Yo, Zo) = (10000, Zo) with Zo > 0, Zt are
all positive and Yi increases to local maximum of cubic (Yt, Zl) ; according
to the value of q, one has relaxation cycle, limit frontier strip or divergence.

When Q > 2,59809, one has negative Zt > 0 and positive Zt. If positive
Zt have more influence, the system will arive to the local maximum (Y1 , Zl)
and then diverges.
On the other hand, if positive and negative Zt are mutually neutralized,
the system stays in a narrow strip around the straight line Y = (t':e:)s. This
is the situation represented by fig. 20 where (1 - e)8 = 0.0001, v = 2.8 and
(Yo, Zo) = (10020,0.5).
 145

Figure 20: Central limiting strip v = 2.8, (1 - e)S = 0.0001 and Yo = 1020

6 Analogy and difference between bifurcation

diagram and cubic

It appears that the upper branch of the cubic looks like the bifurcation
diagram of fig. 2 ; however, stable cycles and chaotic zone appear when one
goes from right to left in the cubic but from left to right in the bifurcation
diagram. There are more basic differences between the two diagrams. On
a bifurcation diagram, one can observe a limit point, a stable cycle, or a
chaotic evolution for a given value of v - S ; all points of the cycle or of the
chaotic dynamic are situated on the same vertical line drawn from point of
abscisse v - S . On the other hand, stable cycles or chaotic dynamics are
situated on parallel vertical lines distant by Zt since Yi = Zt + Yi-I. But
since all these lines are squeezed one upon another, one has the impression
that these points are on the same straight line (distance between two
straightlines is about one when Yi is about 10000).
A last difference must be noticed ; in the bifurcation diagram one has
on each straight line the whole dynamics of (Zt) for a particular value of
v - S ; in the upper (or lower) part of the cubic, one has the same value of
v - s ; in the upper (or lower) part of the cubic, one has the same value
of v - s + (1 - e)S ; each vertical line corresponds to a particular Yi and
indicates just one point (Yi, Zt) of the series defined by (S) model.

References
[1] G. ABRAHAM-FROIS and E. BERREBI [1995], InstabiliU, cycles,
chaos, Economica

[2] C. CHIARELLA [1990], Elements of a Non-Linear Theory of Economic

Dynamics, Springer-Verlag
146

[3] R. H. DAY and W. SHAFER [1986], "Keynesian Chaos", Journal of

Macro-economics, 7.
[4] R. GOODWIN [1951]' "The Non-Linear Accelerator and the Persis-
tance of Business Cycles", Econometrica
[5] J. R. HICKS [1950], A Contribution to the Theory of the Trade Cycle,
Oxford University Press.
[6] T. PUU [1991]' Non-Linear Economic Dynamics, Springer-Verlag.
Credit and Financial Markets
in Keynes Conception of
Endogenous Business Cycles:
an Interpretation
Richard ARENA and Alain RAYBAUT

1 Introduction

In his writings, John Maynard Keynes never developed a proper and

systematic theory of the business cycle. However, the Treatise on Money
and the General Theory include substantial indications on which it is
possible to rely in order to explain the macroeconomic emergence and
persistence of economic fluctuations. The only obstacle lies in the apparent
contradictory elements which are present in Keynes's works: according to
the post-Wicksellian viewpoint of the Treatise, the emphasis is put on the
divergence between the natural and the market rates of interest and on
its effects upon the demand for credit, while the General Theory stresses
the role of human psychology (including on the stock exchange) in the
determination of investment decisions.
This apparent contradiction is also present within modern Keynesian
approaches. A good example is provided by the internal debate which,
during the eighties, divided Post-Keynesian economists into two groups:
the supporters of the theory of endogenous money supply (Moore 1988;
Lavoie 1985) and the supporters of the liquidity preference theory (Kregel
1980, 87).
To begin with, our contribution will be devoted to the nature and reality
of this contradiction. At a second stage, the use of a model of business
cycles will show how it is possible to build a single framework allowing to
cope with both components of the contradiction and, therefore, to explain
why it was only apparent.
148

2 Banks, financial markets and business cycles

In the Treatise, as well as in the General Theory, Keynes relates the

existence of business cycles to the adjustment mechanism between saving
and investment. However, this relation takes in a different form depending
on what book is considered.
In the Treatise, expansions and depressions correspond to situations
where increases and decreases of the divergence between the natural and
the market rates of interest. Thus, an initial impulse, causes for instance
by a major innovation, increases the natural rate of interest. This increase
induces, in turn, a wave of optimistic expectations and, therefore, an
increase of the production level. This increase implies a rise of the demand
for investment goods and thus of the demand for credit as well as greater
employment and spending levels. Little by little, this expansion propagates
and is generalised to all industries.
At first sight, this expansion might be seen as endless. Theoretically,
banks are indeed able of creating an unlimited supply of credit:
" If we suppose a closed banking system, which has no relations with
the outside world, in a country where all payments are made by cheque
and no cash is used, and if we assume further that banks do not find it
necessary in such circumstances to hold any cash reserves but settle inter-
bank indebtness by the transfer of order assets, it is evident that there is
no limit to the amount of bank money which the banks can safely create
provided that they move forward in step" (Keynes, 1930, Vol I, p. 23).
Now, if this unlimited creation of credit occurs, that means that produc-
tion and demand will increase. This remark shows the predominant role
played by banks within the credit market. Accordingly, if the bank rate is
defined by Keynes as " the effective rate for lending and borrowing which
prevails in the market. " (Keynes 1930, Vol I, p. 179), this does not prevent
however banks to exert their influence on its formation. This fact is essen-
tial since Keynes uses the term market rate of interest " for the complex of
bank rate and bonds rate" (Keynes 1930, op. cit. p. 179) and " assumes
that changes in bank rate affect the market rate of interest in the same
direction." (Ibid. p. 180). This means that the banking system does not
control the supply of money but the market rate of interest. ( Keynes 1930,
Vol Ii, p. 189):
" Apart from international complications which we are putting on one
side for the present, there is no reason to doubt the ability of a central bank
to make its short term rate of interest effective in the market. " (Keynes,
1930, Vol II, p. 325).
Therefore, the conditions of a Wicksell- type cumulative process are
realised since the banking system is able to impose its own rate to the
market and therefore, to regulate it. However, the idea of an unlimited
credit supply is misleading if we consider the practice of the banking
 149

systems. It is indeed well-known that the mere existence of a central bank

provides institutional constraints which commercial banks have to respect.
Each of them is indeed supposed to hold a given volume of reserves in order
to deal with that Keynes calls " the inevitable minor discrepancies which
are bound to occur over short periods. " (Keynes, 1930, vol I, p. 24). Now,
the amount of these reserves
" will depend partly on the habits of the depositors, as governed by
the practices the country and of the period and by the class of business
conducted by the particular bank's clients, and partly on the scale of the
bank's business which for this purpose is generally measured by the amount
of its deposits. " (Keynes, 1930, vol I, p. 25).
These constraints are institutional, behavioural, and also depend on
social conventions, though they exert a macroeconomic effect on the volume
of bank money which it is possible to create at a certain moment of time:
" We now perceive that there exists, not only the check on individual
banks that they must keep step, but also a check on the banks as a whole. "
This is why " it is the aggregate of the reserve resources which determines
the pace which is common to the banking system as a whole. " (Keynes,
1930, vol I, p. 25).
In other words, the amount of reserves determines the extent according
to which the credit supply is limited. It prevents the permanence of a
divergence between the natural and market rates of interest. At a given
moment of an expansion phase, the ability of commercial banks to create
credit can no longer be increased, since this increase will be contradictory
with the necessity of maintaining a sufficient volume of reserves. As soon
as this limit is reached, the supply of credit cannot extend further and the
market rate of interest rises. The speed of reaction of the banking system
therefore strongly depends on the nature and behaviour of the banking
and financial institutions of the economy. It will contribute to determine
the importance and the length of the business cycle.
The scheme provided by the General Theory differs substantially from
the previous one.
The origin of the business cycle lies no longer in the existence of a
divergence between the market and the natural rate of interest which the
credi t system allows to occur. It is now related to one of three " independent
variables " of the General Theory, namely, the marginal efficiency of
capital. This variable corresponds to the expected rate of profitability of
new investments. Now, on one side, the expectations which govern the
determination of this rate are fragile and volatile. On the other side, the
life span of a capital good may not be anticipated with precision, especially
in the industries in which there is technical progress. Last, the general
conditions which determine the volume of investment belong to different
types: " The considerations upon which expectations of prospective yields
are based are partly existing facts which we can assume to be known more
or less for certain, and partly future events which can only be forecasted
150

with more or less confidence. Among the first may be mentioned the existing
stock of various types of capital-assets and of capital-assets in general and
the strength of the existing consumer's demand for goods which require for
their efficient production a relatively larger assistance from capital. Among
the latter are future changes in the type and quantity of the stock of capital-
assets and in the tastes of the consumer, the strength of effective demand
from time to time during the life of the investment under consideration, and
the changes in the wage-unit in terms of money which may occur during
its life. We may sum up the state of psychological expectation which covers
the latter as being the state of long term expectations. " (Keynes, 1936,
pp. 147-48).
Therefore, on the one hand, some determinants of long run expectations
are fairly well known and their inclusion in the new investment decision-
making process seems simple and clear-cut. On the other hand, for the
remaining determinants, knowledge is replaced by ignorance, namely by a
situation of " extreme precariousness of the basis of knowledge. " (Keynes,
1936, p. 149).
" Precariousness" is reinforced by the influence exerted by financial mar-
kets on the behaviour of investors. The division between owners and man-
agers indeed allows financial agents to modify their investment decisions
when they correspond to the acquisition of shares. Now, these acquisitions
are not independent from the state of the stock exchange. The possibility
for investors to reconsider their commitments increases the " precarious-
ness " of investment expectations even more. We indeed know that " it is
of the nature of organised investment markets, under the influence of pur-
chasers largely ignorant of what they are buying and of speculators who are
more concerned with forecasting the next shift of market sentiment than
with a reasonable estimate of the future yield of capital-assets that, when
disillusion falls upon an over-optimistic and over-bought market, it should
fall with sudden and even catastrophic force. " (Keynes, 1936, p. 316).
Hence, in the General Theory, the emphasis is put on the variation of
the marginal efficiency of capital: " The trade cycle is best regarded, I
think, as being occasioned by a cyclical change in the marginal efficiency
of capital, though complicated and often aggravated by associated changes
in the other significant short-period variables of the economic system. "
(Keynes, 1936, p. 313).
What about the long-run rate of interest? For Keynes, this rate is less
unstable than the marginal efficiency of capital and, therefore, it only plays
a complementary role in the emergence and development of the business
cycle. The decrease of the efficiency of capital therefore implies the rise of
the rate of interest and not the reversal. During the depression, as soon as
the marginal efficiency of capital begins to decrease, agents try indeed to
sell their inventories of goods and assets in order to avoid the decrease of
their price. The level of liquidity preference therefore increases and, ceteris
paribus, enhances the rise of the interest rate and hence reinforces the
 151

depression.
AB a result, two different points of view seem to prevail. On the one hand,
the existence of credit allows the possibility of a divergence between the
natural and the market rate in favour of the latter. This divergence in turn
underpins the phase of expansion. On the other hand, long-run expectations
encourage investment decisions and, hence, expansion or depression phases.
Some authors have interpreted the opposition between these two points of
view as the consequence of the respective different roles given by Keynesians
or Post-Keynesians to banks and financial markets. This opposition must
however be questioned.
Accordingly, we propose a formal framework allowing the analysis of the
compatibility of these developments contained respectively in the Treatise
and in the General Theory. We shall try to point out that, far from being
contradictory, these developments can be combined in order to provide a
more satisfactory explanation of the emergence and persistence of business
cycles.

3 A synthetic model of endogenous fluctuations

In this section a synthetic business cycle model is developed. We first

present the general framework of the model and derive its stationary state.
Then we turn to the analysis of its local dynamics.

3.1 DETERMINATION OF THE LEVEL OF ACTIVITY AND OF

THE EQUILIBRIUM ON THE MONEY MARKET

Let us start with a traditional Keynesian aggregated model of a closed

economy where a single good is produced and where money and financial
assets are held. Nominal income Y can be written as :

Y=C+I (1)
where C and I respectively refer to the volume of national aggregated
consumption and investment. In addition :

Y=wN+P (2)
where P refers to the level of gross profits, and N refers to the level of
employment. The nominal wage rate w is, in line with the Cambridge
tradition, determined by a bargaining process that is exogenous to the
model.
The consumption level, C, is given by:
152

C = cwN + c'IJ d (3)

where c and c' refer respectively to the propensities to consume out of
wages and out of distributed profits Pd. With:

lId = (1 - (3)P, with 0 < {3 < l. (4)

The employment level, N, is determined by the level of production Q
and by labour productivity 1.

Q
N=- (5)
I
In addition, let us assume that the price of the produced good is given
by means of an exogenous mark up m:

P = (1 +m)w (6)
I
Thus, with a given a capital/output ratio v, the real sector of our model
is completely described.
The description of the financial sector borrows from the analysis de-
veloped by Gallegati and Gardini (1991) and by Franke and Semmler
(1991,92).1. One main difference can be found in the fact that we con-
sider two distinct interest rates. The first rate Z , is parametrically fixed by
the banks and refers to the " bank rate " of the Treatise. The second one,
i, is detemined by the equilibrium on the assets market and refers to the
long-term rate of the General Theory.
Let us assume for simplicity that the money supply MS is constant and
exogenous. The demand for money has three components: a transactional
and precautionary element T R, a speculative component consistent with
the notion of liquidity preference of the General Theory, SP, and finally a
component related to the demand for finance Exfi. 2 .
Accordingly, assume that:

TR=kY, (7)
where 0 < k < 1 is the reciprocal of the velocity of money

SP = SP(i) (8)

Exfi = b (9)

1 Hence, for a detailed presentation of the financial sector see Gallegatti and
Gardini (1991) and Franke and Semler (1992).
2 As we know, the " finance motive" historically appears after the publication
of the General Theory.
 153

It is now time to indicate how the real and financial sectors are connected.
from a Keynesian perspective, this link is to be found in the financing of
investment activity. From this standpoint, three sources of financing are
taken into account in the model: retained net profits A, debt (bank loans)
iJ , and finally access to financial markets M .Thus we obtain:
(10)
Let us assume that the internal financing constraint is a proxy for the
ability of firms to fulfill their debt commitments. Hence, according to
Kalecki's increasing risk principle, new external finance raised by the firms
is proportional to net profits, i.e. net cash flow A. We have :

iJ = (b-1)A (11)
with b > 1, and where A is given by

A =,BIT - iD (12)
The investment function we have retained is the following :

I = bA + a( r e - i) (13)
where a > 0 and b > 1 are two real positive parameters. Thus, investment
decisions depend both on net cash flow A and on the difference between the
expected profit rate and the current market interest rate. Consequently (h-
1) refers to the weight of financial constraints undergone by firms (Gallegati
et Gardini 1991)3.
Assume, moreover, that the expected profit rate is given by the ensuing
relation:

r e = R(d,X) (14)
where d refers to the debt/capital ratio and X refers to the state of
confidence. Let us suppose that R is a continuous derivable positive function
satisfying R~ > O.
According to Minsky (1980, 82), the evolution of the leverage ratio D/K
during the cycle precedes an increase (respectively a decline) of the financial
fragility of the economy. A small D / K ratio' is first associated with high
expected profits, i.e. R~ > O. Progressively, a rise in D / K signals a future
deterioration of profitability, hence R~ becomes negative. 4 •
Since the real positive variable X refers to the notion of " state of confi-
dence" of the General Theory mentioned in the introduction of the paper,

3 A simple combination of equations (10), (11) and (13) shows that the amount
of investment financed by the financial market is given by At = a(r" - i).
"'Thus, the function R(d,X) can be specified as follows:
R(d,X) = a.d + {3d + oX, with a. < O,{3 > 0 and 0 > O.
154

investment decisions depend positively on X. AB we may recall, fluctua-

tions of the state of confidence also act upon aggregate activity through
the channel of financial markets. However, in this model, fluctuations of in-
vestment decisions do not rely on an hypothesis of volatility of confidence
on financial markets, volatility transmitted to the real sector. Accordingly,
as noticed by Franke and Semmler (1992), "should that have been the case,
the economy might be seen as being too much dependent on the psychol-
ogy of speculators" (p.331). In line with the General Theory, it is possible
to consider that real and financial sectors are simultaneously affected by
changes of the state of confidence, " the impact on investment demand
being the most direct" (ibid.). From this perspective, in our model the in-
terplay of the state of confidence and of the financial sector is understood
in an indirect way through the determination of expected profits and of the
interest rate.
Let us finally consider the condition for equilibrium on the financial and
money markets. The money market is characterized by two elements. On
the one hand, the financial gap, i.e. the demand for external finance, is
a fraction of the demand for money. On the other hand, the amount of
liquidity created by the banking system is, for a given banking interest
rate and without credit rationing, driven by the firms' demand for finance.
Recent endogenous business cycles or cyclical growth models (Gallegati and
Gardini, 1991, Foley, 1986, 87, ABsada and Semmler, 1992, Semmler and
Sieveking, 1993), adopt the same perspective. We obtain:

(15)
Replacing Y and b by their respective values in K and D, we have:

M -CoK+z(b-l)-SP(i) =0 with Co = w(k(1 + m) + {3m(b- 1)) (16)

, lv
Relation (15) can be rewritten as :

i = L(K,D)

Hence, this type of LM curve with L ~ < 0 and Lv > 0 (cf. Annex 1)
gives the equilibrium or market interest rate as a function of the capital
stock and of the stock of debt.
The real and financial sectors of our model having been described, it is
now possible to determine the stationary state of the economy and to turn
to the analysis of its dynamical properties.

3.2 ENDOGENOUS BUSINESS CYCLES

From the equations given in the previous sections, we derive a nonlinear
dynamical system with three state variables: the capital stock K, the stock
 155

of debt D, and the state of confidence X. The dynamics of capital stock

and of debt is given by:

(17)

D=(b-l)A (18)
As we know, variations in the psychological variable representing in-
vestors' confidence are an essential ingredient in Keynes' conception of
macroeconomic instability in the General Theory. We give a somewhat
simplified, but sufficient for our purpose representation of the evolution
of that variable. Let us simply assume that its movement is given by the
following relation:

x = H(X) (19)
where H is a continuous function defined for all X > 0, satisfying
H(X*) = 0 and H'; i=- 0 for X* > 0 5 .
In line with Keynesian business cycles models based on fluctuations of the
state of confidence which usually refer to the General Theory's Notes on the
Trade Cycle and to the 1937 QJE' article (Woodford 1989, Caminati 1989),
the evolution of the state of confidence, which is basically a psychological
variable is autonomous. Function H captures the idea that this variable will
fall (rise) if a deterioration (improvement) of present conditions is expected.
Let us for instance consider the upswing period of a cycle. Investors who
are optimistically minded should expect confidence to increase more or less
steadily. Conversely, after the crisis, during the downswing period, investors
should expect confidence to decrease. However, as noticed by Boyd and
Blatt (1988), as time goes on, " the memory of the last panic becomes less
vivid, and investors become more willing to look farther into the future
with confidence that their predictions and expectations will be met " (p.
65), bringing about a reversal of opinion. 6.
From equations (17)-(18)-(19) a unique stationary state can easily be
derived. We obtain the following result:

Proposition1: existence of a stationary state

5We can for instance specify this function as follows: H(X) = (hX2 + 82 X,
with (h < 0 and (h > O.
6 A complementary approach consists, in line with Kalecki (1937) and Minsky,
in taking into account the role of increasing risk and of financial fragility. As
noticed by Franke and Semmler (1991), this risk embodied in a function of the
debt capital ratio DjK exerts negative effects on the state of confidence. Thus,
in that case we have H~ < O.
156

Assume that M > SP(i*) where i* refers to the market interest rote
evaluated at the stationary state. System (18) (19) then has a unique strictly
positive stationary point E = (K*, D*, X*). We obtain:
K* -- -lv(SP(i*)-M)
kw(l+m) ,
D* -- f*K were
h f -- /3m1"
lvi' X -- X*7

Proof:
- On the one hand, it is clear that iJ = 0 implies A = O. Hence, D* = tjJK,
with
tjJ = (3,m~.
v.
At the stationary point, the debt/capital ratio, D/K, is
constant and equal to tjJ.
- On the other hand, k = 0 implies :b
A = r e -i. Thus, at the stationary
point we have r - i = 0, the expected profit rate is equal to the market
e
rate. Consequently, r e = R(f, X*) = i*. Now, the equilibrium condition in
the money market implies:

M - SP(i*) = C1 K, with C1 = wk(~: m) . (20)

Hence, it can easily be deduced that K* = -lv1~f~~)M), which means

that K* > 0 if and only if M > SP(i*).]
The remainder of the paper is devoted to the analysis of local properties
of the dynamical system (17), (18), (19). Consequently, we consider a
linearization of the model in a neighbourhood of the stationary state
E = (K*, D*, X*). Thus, the characteristic equation can be written as

(J33 - z) (z2 - (Jl1 + J22) z + J l1 J22 - J21 J12) (21)

where the Jij's refer to the elements of the Jacobian matrix given in Annex
2.
It is obvious that (20) has a positive or negative real root z3 = H' X* .
The two other roots are then real or complex. We shall first give a necessary
and sufficient condition for the existence of two complex roots.

Proposition 2: existence of complex roots

The charocteristic equation (20) evaluated at the stationary state, has
two complex roots if R4 < R~ < ~ , where Rid < 0 et ~ > O. In
other words, the model has two complex roots if, in a neigbourhood of the
stationary state, the reaction of the expected profit rote to the debt capital
rotio D / K is not too high.

7With the specification of the function H(X) mentioned above, we get X* =

-~
91
>0
 157

[Proof:
The model admits complex roots if: (Ji1 - J 22 )2 + 4J21 Ji2 < 0
Replacing the element of the Jacobian matrix by their respective values
evaluated at the stationary state given in Annex 2, we obtain : (Ji1 -
J22 )2 = (C2R~* + C3)2 ¢:} 4J21 Ji2 = 4( C4R~* + C5 )
where coefficients C2, C3, C4, C5 are the following:
C2 = IVW~~;~:» < 0
C3 = i (b - 1) (1 - S;'{iO)) + bi </> - ~~:~/ti':))
c -
4 -
(b-1)ak(l+m)¢
Iv(M-SP(i*))
>0
C5 = i(b - l)((b - 1)a - b)</>
Then, complex roots exist if:
C22R~*2 + (2C2C3 + 4C4)R~* + (C32 + 4C5 ) < 0
Assume now that C32 + 4C5 < 0, we obtain:
R' R!.... = -(C2C3+2C4)±v'(C2C3+2C.)-C2(C3+4C5)
£'
4,
d C2
Consequently, the model has two complex roots if R < R~ < ~,.J
We then make two assumptions related to the banking rate, i , and on
the state of confidence, X, in a neighbourhood of the stationary state which
ensure the existence of a Hopf bifurcation from the left.

Assumption 1
We suppose that r > i . The bank rate i charged by banks is lower than
the actual profit rate r:

(22)

Assumption 2
In a neighbourhood of the stationary state, we have H'; # O. We
assume moreover that H'; # O. That around the stationary state, the
state of confidence is sensitive to small perturbations in opinion, but a
feedback effect prevails. Thus, assumption 2 means that fluctuations in the
psychological variable are not the main destabilizing factor in the model. 8
Under assumptions 1 and 2 we obtain the following result.

Proposition 3: existence of a Hopf bifurcation

Choose the parameter "( = (b-1) which measures the weight of the inter-
nal finance constraint, as a bifurcation parameter. Then a Hopf bifurcation
takes place at "( = "(*.

[Proof:

8It is obvious that this condition is true when the function H(X) is specified
as in note 3 above.
158

This result relies on the Andronov-Hopf bifurcation theorem.

- Proposition 1 tells us that (20) has a pair of complex roots Zl, Z2 =
ab) ± ib('Y) and a real negative root z3.(Assumption 2)
- For the critical value g*, (20) has a pair of complex roots crossing the
imaginary axis at non zero speed with: . (1- 'Y*) ~(~(~J -=1= 0 .
The real part of Zl and Z2 can be written :
ab) = 'YB1 + B2
where coefficients Bl and B2 are the following:
B 1 -- -~ + {3mw
A
Iv
(1 - SP(ioO)
a)
B - {3mw a,Bmwk(1+m)R~o awk(1+m)
2 - Iv + i(lv)2(SP(iO)-M) - IvSP'(i o )
It is easy to verify that Bl > 0 if i < br(1 - SP(i*» (Assumption I).
Thus, a Hopf bifurcation from the left takes place for 9 = goO. (N.V. Th
1994 p. 202))
When this bifurcation is subcritical, the stationary state (K*, D*, XoO)
is an unstable focus for 9 > g*, stable for the critical value and an
unstable cycle exists for 9 < g" 9. In that case, the model exhibits, using
Leijonhuvfud's words, a " corridor of stabilty " property" 10 Conversely,
when the bifurcation is supercritical the stationary state is a stable focus
for 9 < g* which becomes unstable for 9 = g*. A stable cycle appears for
9 > goO and the economy experiments endogenous persistant cycles.
Hence, it appears that the financial structure of investment, especially
the weight of the internal finance constraint g, plays a crucial role in the
dynamics and prevents the development of a cumulative process.

4 Concluding remarks

In this contribution, the acknowledged contention of the existence of

two different, even contradictory, approaches of economic fluctuations in
Keynes' writings has been challenged. Indeed, the nonlinear model of
endogenous fluctuations we have proposed has shown that, for Keynes,
business cycles can be simultaneously explained by the credit market and
by expectations (including on the stock market) of investment.
On the one hand, the state of long-term expectations plays an important

9The bifurcation-type, and therefore the stability criterium of the cycle,

depends on the sign of a coefficient S. If S < 0, the bifurcation is supercritical,
while the bifurcation is subcritical when S > O. However, computation of S
composed of seconds and third derivatives orders of the Taylor approximation of
the system could be messy, even in the simpler case.(Guckenheimer and Holmes
1983 p.154-56).
laThe paternity of this remark concerning this type of Hopf bifurcation
dynamics is due to Franke and Semmler (1992) op. cit. p.344.
 159

role in the analysis of persistent economic instability. If the state of con-

fidence actually exerts a direct positive effect on investment decisions, its
mere autonomous evolution is not however the unique cause of instability.
Indeed, the emergence of the cyclical behaviour of the economy is greatly
explained by the financial structure of investment. So, the explanation of
business cycles outlined in the General Theory is enriched. On the other
hand, the interplay of the two interest rates of the Treatise is also taken
into account. The first rate, the assets yield, is determined by equilibrium
conditions on the financial market, while the other is the loan rate fixed by
the banks. The latter is a parameter imposed by the banking system which
constrains the cyclical dynamics.
Our findings connect in many respects to New Keynesian Macroeco-
nomics. In this literature, credit and financial markets do explain a sub-
stantial part of economic fluctuations. Nevertheless, our approach based on
an endogenous business cycles model relates more closely to Keynes's view
on the inherent tendency to instability of market economies.

5 Appendix

Annex 1

Partial derivatives of the equilibrium interests rate are the following:

"K = SP(i)
Z
01
<0
>0
"D - -i(b-1)
Z - SP(i)
Partial derivatives of the expected profit rate are:
re'K = -2~dR
re'D = -kdR
re'X = R'X

Annex 2

The elements of the jacobian matrix J evaluated at the stationary state

are:
J*11 = a(r K*
e , - i'
K*
) + bf3 mw
Iv
Ji2 = a(rg* - i~:h) - bi
Ji3 = ar'k*
J21 = (b - l)f37.: w > 0
J22 = -i(b - 1) < 0
J 23 = 0
J 31 = 0
J32 = 0
J 33 = H~.
160

References
[1] ARENA R. and RAYBAUT A. (1995), "Cycles et croissance, un point
de vue ne o-kaldorien", Revue Economique, Nov-Dec.
[2] ASADA T. and SEMMLER W. (1992), "Growth, Finance and Cy-
cles", Working paper New School for Social Research, New York.
[3] BOYD I. and BLATT J.M. (1988), Investment, Confidence and Busi-
ness Cycles, Springer Verlag.
[4] CAMINATI M. (1989), "Cyclical Growth and Long-Term Prospects",
Political Economy, Vol. 5 N° 2 p.107.
[5] DELLI GATTI D., GARDIN! L. and GALLEGATI M. (1993), "In-
vestment Confidence, Corporate Debt and Income Fluctuations",
Journal of Economic Behavior and Organization, 22, October, p.154-
161.
[6] FRANKE R. (1992), "Stable, Unstable, and Persist ant Cyclical Be-
haviour in a Keynes-Wicksell Monetary Growth Model", Oxford Eco-
nomic Papers, 44, p.242-56.
[7] FRANKE R. and SEMMLER W. (1991), "Expectations, Dynamics,
Financing and Business Cycles. ", in Profits, Deficits and Instability
edited by D.B Papadimitriou MacMillan
[8] GALEGATI M. and GARDIN! L. (1991), "A Non Linear Model of
Business Cycles with Money and Finance", Metroeconomica, Vol 42
N° 1 p.l
[9] GUCKENHEIMER J. and HOLMES P. (1986), Non Linear Oscilla-
tions Dynamical Systems and Bifurcations of Vector Fields, Springer
Verlag
[10] JARSULIC M. (1993), "Complex Dynamics in a Keynesian Growth
Model ", Metroeconomica, Vol. 44 Fev. p. 43.
[11] KALECKI M. (1937a), "A theory of the Business Cycle", Review of
Economic Studies, Feb. p. 77.
[12] KALECKI M. (1937b), " The Principle of Increasing Risk", Econo-
metrica, Nov p. 441.
[13] KEYNES J.M. (1930), A Treatise on Money, Vol 1 et 2, in The
Collected Writings, vol 5 and 6, MacMillan 1971.
 161

[14] KEYNES J.M. (1936), The General Theory of Employment, Interest

and Money, in The Collected Writings, vol 7 MacMillan, 1973.
[15] KEYNES J.M. (1937), " The General Theory of Employment",
Quarterly Journal of Economics, Vol 51 p.209
[16] KREGEL I. (1980), "Expectations and Rationality within a Capitalist
Framework" ,. Paper presented at the American Economic Association
Meeting, Denver.
[17] KREGEL I. (1987), " Irving Fisher, Great-Grand Parent of the
General Theory: Money, Rate of Return, Overcosts and Efficiency of
Capital", mimeo, Jonhn Hopkins University, Bologna Center.
[18] LAVOIE M. (1985), " Credit and Money: the Dynamic Circuit, Over-
draft economies and Post Keynesian Economics", in M. JARSULIC
ed. Money and Macropolicy, Kluwer publisher, Boston.
[19] MOORE B. (1988), Horizontalits and VeTticalits: the Macroeconomics
of Money and Credit. CUP, Cambridge.
[20] MOORE B.(1988), " The Endogenous Money Supply", Journal of
Post-Keynesian Economics, vollO, N° 3.
[21] PUU N.V. (1994), Dynamical Systems, Springer Verlag
[22] SEMMLER W. (1986), "On Nonlinear Theories of Economic Cycles
and Persistence of Business Cycles", Mathematical Social Science,
12(1) p.47-76
[23] SEMMLER W. (1989), ed. Financial Dynamics and Business Cycles:
New Perspectives, M.E. Sharpe, New York.
[24] SEMMLER W. and SIEVEKING M. (1993), "Nonlinear Liquidity
Growth Dynamics", Journal of Economic Behavior and Organization,
22, October. p. 189-208.
[25] WOODFORD M. (1989), "Finance, Instability and Cycles ", in W.
Semmler ed Financial Dynamics and Business Cycles M.E Sharpe
 Part 4

Methodological Issues
Business Cycles, Chaos and
Predictability
Alfredo MEDIO l

1 Business Cycles and Chaos: An Introduction

A scanty observation of the time series of most variables of economic

interest, such as the price of an individual commodity or the exchange
rate between two currencies, shows the presence of bounded and more
or less regular fluctuations, with or without an underlying trend. Even
more interestingly, this oscillating behaviour seems also to characterize
the aggregate activity of industrialized economies, as represented by their
main economic indicators like GNP, global production, consumption and
investment, employment or exports. Economists have long been concerned
with the explanation of this phenomenon. The literature on the subject
is enormous and the number of different theories equally vast. However,
if we restrict ourselves to the mathematical investigation of economic
fluctuations, we observe that two basic, mutually competing approaches
have dominated this area of research in modern times.
The origin of the first approach - which we shall label "econometric
approach to business cycles" - may be traced back to the seminal works of
Slutsky (1927), Yule (1927) and Frisch (1933) and was later developed and
given the status of orthodoxy by the works of the Cowles Commission in
the 1940s and 1950s.
The fundamental idea of the econometric approach is the distinction be-
tween impulse and propagation mechanisms. In the typical version of this
approach, serially uncorrelated shocks (the impulse mechanism) affect the
relevant variables through distributed lags (the propagation mechanism),
leading to serially correlated fluctuations in the variables themselves 2 .

IThe author gratefully acknowledges the financial support of the Italian

Ministry of the University (M.U.R.S.T.) and the National Council of Research
(C.N.R.).
2For completeness's sake, among the impulse-propagation models of the cycle,
one should distinguish between those in which random external events affect
economic "fundamentals" (essentially, tastes and technology), and those in which
166

All Slutsky showed, even simple linear non-oscillatory propagation mecha-

nisms, when excited by random, structureless shocks, can produce output
sequences which are qualitatively similar to certain observed macroeco-
nomic cycles.
The ability of the econometric approach to provide an explanation
of business cycles was called in question largely on the ground that
explaining fluctuations by means of random shocks amounts to a confession
of ignorance. An alternative approach - which we shall label "nonlinear
disequilibrium (NLD)" - was then developed by a school of economists who,
somewhat misleadingly, was associated with the name of Keynes. The basic
idea of these authors was that instability and fluctuations are essentially
due to market failures and consequently they must be primarily explained
by deterministic models, i.e., by models where random variables play no
essential role. Classical examples of such models can be found in the works
of Kaldor (1940), Hicks (1950), Goodwin (1951). Mathematically, these
models were characterized by the presence of nonlinearities in certain basic
functional relationships of the system and lags in its reaction-mechanisms.
The typical result was that, under certain configurations of parameters,
the equilibrium of the system can loose its stability, giving rise to a
stable periodic solution (a "limit cycle"), which was taken as an idealized
description of self-sustained real fluctuations, with each boom containing
the seeds of the following slump and vice versa. The NLD approach to the
analysis of business cycles was very popular in the forties and fifties, but
its appeal to economists seems to have declined rapidly thereafter and a
recent, not hostile textbook of macroeconomics (Blanchard and Fischer,
1987, p. 277) declares it 'largely disappeared'.
The reasons for the crisis of the Keynesian style of theorizing and the
related NLD theories of the cycle are manifold, not all of them perhaps
pertaining to scientific reasoning, and a full investigation of this interesting
issue is out of the question here. However, there exist two fundamental
criticisms, raised against the NLD approach mainly by supporters of the
rational expectations hypothesis, which are relevant to our discussion and
can be briefly summarized thus:
• In the NLD models, agents' expectations, either explicitly modelled, or
implicitly derived from the overall structure of the model, are, under most
circumstances, incompatible with agents' ''rational'' behaviour .
• The NLD approach has been "refuted" by empirical observation, as
time series generated by the relevant models do not agree with available
data. For example, so the argument runs, a time series generated by a model
characterized by a stable limit cycle will have a power spectrum exhibiting
a sharp peak corresponding to the dominant frequency, plus perhaps a few

those events directly change only agents' expectations. The latter case has been
extensively studied in recent years in the economic literature under the label
"sunspots" .
 167

minor peaks corresponding to subharmonics. On the contrary, aggregate

time series of actual variables would typically have a broad band power
spectrum, often with a predominance of low frequencies (see, for example,
Granger and Newbold, 1977).
The first of these criticisms can be best appreciated by making reference
to the original formulation of the rational expectations hypothesis. In
Muth's own words, "the expectations of firms (or, more generally, the
subjective probability distribution of outcomes) tend to be distributed,
for the same information set, about the prediction of the theory (or the
"objective" probability distributions of outcomes)" (see, Muth, 1961, p.
316). If expectations were not rational in the sense defined above - so
Muth's argument continues - 'there would be opportunities for economists
in commodity speculation, running a firm, or selling the information to
the present owners' (see, Muth, loco cit., p. 330). This argument does have
some validity if the outcomes of the dynamical system under consideration
can be accurately predicted once the 'true' model is known, e.g., if the
outcome is periodic. In this case, if agents are 'rational', fluctuations can
be explained only by exogenous random shocks3 .
However, if the theory implies chaotic, unpredictable dynamics of the
system, the rational expectations argument loses much of its strength and
non-optimizing rules of behaviour - such as adaptive reaction mechanisms
of the kind assumed by the NLD models, or 'bounded rationality' d la
Simon - might not be as irrational as they may seem at first sight. At any
rate, the presence of chaos makes the hypothesis of costless information
and the infinitely powerful learning (and calculating) ability of economic
agents, implicit in the perfect foresight-rational expectations hypothesis,
much harder to accept.
Equally strong reservations can be raised in relation to the second
criticism mentioned above. It is well known that deterministic chaotic
systems can generate output qualitatively similar to the actual economic
time series, e.g., they can have a similar power spectrum.
None of these broad considerations can be used as a conclusive argument
that business fluctuations are actually the output of chaotic deterministic
systems. However, they strongly suggest that, in order to describe complex
dynamics mathematically, one does not have necessarily to make recourse to
exogenous, unexplained shocks. The alternative option - the deterministic
description of irregular fluctuations - provides economists with new research
opportunities undoubtedly worth exploiting.

3Even when the outcome is periodic, however, if the periodicity is long and
the time path very complicated, one may question the idea that well-informed
real economic agents would actually forecast it correctly: all the more so if the
outcome is quasi periodic, possibly with a large number of incommensurable
frequencies.
168

2 Stochastic Behaviour of Deterministic Systems

Although a universally accepted and comprehensive characterization of

chaos is still lacking, in this paper we shall adopt the definition suggested
in a recent conference on chaos (Royal Society, London, 1986), namely,
'stochastic behaviour occurring in a deterministic system'. Broadly speak-
ing, a system is said to be deterministic when it comprises no exogenous
random variables. On the other hand, the observable behaviour of a dynam-
ical system is called stochastic when the transition of the system from one
state to another can only be given a probabilistic description as happens for
truly random processes, e.g., the outcome of spinning a roulette wheel. The
meaning of these rather intuitive considerations will be made clearer, we
hope, in the following pages. Indeed, much of this·paper is devoted to giving
a precise characterization to the basic notions of stochastic behaviour of
deterministic systems, chaos, unpredicatbility and other related concepts.
Before carrying further our investigation, however, we would like to pause
and briefly consider some interesting preliminary questions. First of all, is
deterministic chaos - or if you prefer, stochastic behaviour of deterministic
systems - an intriguing but rare event, or is it a general and ubiquitous
occurrence? A broad answer to this question is: deterministic chaos is not
a rare event in the sense that, as is well known, it can be found generi-
cally in "small" and "simple" dynamical systems e.g., non-invertible one-
dimensional maps, or three-dimensional systems of differential equations,
characterized by a simple quadratic nonlinearity.
A more specific, and economically more interesting question is the
following: how relevant is chaos to economic theory and how common is
chaotic behaviour in real economies? To the first part of this question, we
can give a straight and definite answer, as follows. If we consider the two
basic dynamical models of the prevailing economic theory - i.e., optimal
growth and overlapping generations model - we know that, generically,
their output can be chaotic (see, for example, Boldrin and Montrucchio,
1986; Deneckere and Pelikan, 1986; Medio, 1992; Medio and Negroni, 1996).
"Generically" here means that chaos occurs for parameter constellations
that we cannot exclude a priori on the basis of the economic "first
principles". Hence, if one takes those models seriously as most members of
the economic profession do, chaos must be considered a very real possibility.
A similar conclusion is reached if we consider non-equilibrium dynamical
models inspired by the Keynesian tradition.
It is more difficult to give a neat answer to the second part of the question
above. Here we have two possible strategies. On the one hand, we could
make recourse to the analysis of economic data, trying to find empirical
evidence of chaos. There is a vast literature on this subject but we cannot
discuss it here in any detail. In our opinion, no general consensus has yet
been reached, except on the fact that it is unlikely that economic data
 169

may be represented by low-dimensional chaos (say, an orbit on a three-

dimensional chaotic attractor). A different, not necessarily alternative route
is to try and "calibrate" a theoretical model, i.e., estimate the parameter
values deemed to be more realistic and verify if they correspond to chaotic
behaviour. Although this is a reasonable procedure in principle, we are
rather skeptical about its validity in practice. A serious "calibration"
would require theoretical models conceived in such a way that they can
be empirically tested. Unfortunately, economic models - or at least those
describing equilibrium dynamics - have a normative rather then positive
character, as they are constructed to described what rational agents should
be doing rather than what real economic agents actually do.
At any rate, in this paper we shall not develop this point any further but
shall try instead to clarify some general theoretical issues whose proper
understanding is, in our opinion, crucial for a correct assessment of the
relevance of chaos for economics. In particular, we shall investigate the
basic question of predictability. Rather than discussing the issue in full
generality, we shall consider its simplest possible formulation containing all
the interesting features that concern us here. In this way, we shall obtain
a greater transparance with a smaller amount of technicalities.

3 The Simplest Representation of Chaotic

Dynamics

From the mathematical literature on chaos, we know that the simplest

deterministic model capable of stochastic behaviour can be represented by
a map (a difference equation) such as
Xt+l = f(xt) (1)
where x denoted a real scalar variable and f is a "one-hump" function, such
as a quadratic polynomial. From the economic literature (see, for example,
the works quoted in Section 2 above), we know that (1) may represent
a policy function resulting from a problem of optimal growth or optimal
consumption. Similar maps may describe the t~me evolution of consumption
in a model of overlapping gnerations (see, for example, Benhabib and Day
(1981) and (1982); Grandmont (1985)). Alternatively, equation (1) could
describe the dynamics of inventory in a Keynesian model (see, for example,
Lorenz, 1989).
Since for our present purpose the specific form of f is relatively unim-
portant, we select the celebrated "logistic" map, namely we put
f(x) = rx(1 - x) (2)
Equation (2) is known to have chaotic behaviour for many (actually for
infinitely many) values of the parameter r. However, the only exact value
170

of r for which we can prove this claim rigorously is 4. The reason for this
result is that - when r = 4 - map (2) can be trasformed into a even simpler
equivalent map. Introducing the invertible change of co-ordinates

y = (~) arcsin Vx, Y E [0,1]

map (2) is tranformed into

2y for 0 ::; y ::; 1/2

T(y)= { 2-2y for 1/2 ::; y ::; 1

A formidable result originally due to Ulam and von Neumann shows that
this transformation preserves most dynamical properties of the map. In par-
ticular, it preserves certain properties that charaCterizes chaotic behaviour
such as the Lyapunov characteristic exponent and metric entropy4.
Since map (4) is semilinear and therefore easy to study, we shall concen-
trate our discussion on it, knowing that the basic results of the investigation
can be referred back to the more "realistic" map (2).
In our previous intuitive discussion of chaos, we came to the conclusion
that from the economic point of view, the fundamental characterization
of chaotic dynamics is the lack of predictability, even when the rule that
governs the dynamics is known exactly - i.e., when we assume that the
theoretical model is the "true" model.
In order to provide a more rigorous discussion of this point, we must
first of all clarify a preliminary question. If a dynamical model - a map
or a system of ordinary differential equation - is known and the state of
the system at each instant of time can be measured with absolute precision,
there is nothing left to discuss. Well-known mathematical results guarantee
that, under very general conditions, the future time evolution of the system
is also known exactly.
Absolute precision of measurement, however, is a mathematical expres-
sion without empirical content. To be convinced of that, consider the fol-
lowing proposition: "The length of this rod is 7r centimeters." A moment's
reflection will suggest that there is no way to prove that the proposition
is true. All we can prove is a different proposition such as: "The length of
this rod is between 3 and 4 cm. (or between 3.1 and 3.2 cm., or between
3.14 and 3.15", and so on and so forth)", where points on the real line
are replaced by intervals which are the smaller the greater is the (finite)
precision of our measuring device.
Although in economic models we routinely assume that the relevant
quantities are represented by real numbers (i.e., numbers typically iden-
tified by an infinite number of decimals), the quantities about which real

4For a discussion of this point, see, for example, Ruelle, 1989, pp. 40-43.
 171

economic agents make decisions must be approximations, i.e., in the case

under discussion, intervals of the line.
The assumption of absolute precision of measurement in dynamical
models is innocuous for simple systems, e.g. for systems whose asymptotic
behaviours are described by fixed (equilibrium) points, or by periodic
orbits. For these systems, introducing observational errors into the model
does not change the results qualitatively. Most importantly, for simple
systems knowledge of the dynamical equations and of the past history of
the system guamntee predictability of the arbitmrily distant future within
the limits of precision of the mesuring device.
The situation is entirely different if we consider chaotic systems. For
these systems the behaviour of "perfect measurement dynamics" is totally
different from that of "imperfect measurement dynamics". Knowledge
of the "true" model and (finite precision) knowledge of the infinitely
remote past is not sufficient to predict the future. Moreover, the quality
of prediction is not fixed by the distorting effect of the measuring devise
but deteriorates as the time distance over which prediction is performed
increases. (Of course, if the dynamics of the system is bounded, the size of
the error is also bounded.)
The notion of imperfect measurement may be made clearer by looking
at the specific example of the "tent" map. Consider, first of all, that this
map has two non-negative equilibrium points both unstable and that the
interval I = [0,1] is invariant with respect to T. In what follows, we shall
study the orbit of T on I.
Finite precision of observation means that at each instant of time we can
locate the position of the system not on a point but on an sub-interval of
I. A finite collection of N semi-open and disjoint subintervals (II, ... ,IN)
of I whose union is equal to I is called a partition of I and we shall denote
it by Po. A partition can be also viewed as an "observation function"
Po : I - t {h, ... ,IN}, such that for any exact position of the system
x E I c R, Po(x) denotes the element of the partition in which x is
contained.

A1 f.Z 1
Figure 1. The Lebesgue measure m is preserved by the tent map T. Since
T-l(A) = Al U A2 and m(Al) = m(A2) = ~m(A), m(T- 1 (A)) = m(A).
172

4 Chaos and Predictability. Metric Entropy

We shall now show that the dynamics of T on the partitioned space

{h, ... ,IN} is unpredictable in an appropriately defined sense and, more-
over, that it may be undistinguishable from a stochastic process.
There exist of course many ways of partitioning the interval I but in
this case, as we shall see in a moment, the choice of the partition does
not matter. Let us then begin with the assumption that our measuring
device is very rough and can only distinguish from points situated on
the left or on the right of the middle point x = 1/2. Accordingly,
we start from the crudest of all partitions of the [0,1] interval, namely
h = {O ~ x < 1/2}; 12 = {1/2 ~ x ~ I}. For short, let us indicate II by L
(Left of the middle point) 12 by R (Right of the middle point)5. If we now
start from an initial point situated on L or R and iterate the map T, we
shall generate a sequence of two symbols, something looking like

LLRLRRRL ...
Our final goal is to verify the (un)predictability of this (partitioned)
dynamics. To do this, we must preliminarly change our perspective a take
a probabilistic point of view. Let us consider a single iteration of the map
on the partition P, as an experiment whose outcome is uncertain (it could
be L or R) and let us try to evaluate the amount of uncertainty concerning
the outcome (before the experiment), or, equivalently, the amount of
information provided by the experiment (after it). We can do this by means
of a formula known as "Shannon entropy" after the name of one of the
founders of the modern theory of information. If ~ is a random variable
taking a finite number N of values with probabilities PI. ... ,PN, then the
entropy of ~ is
N
H(~) = - LPi log(Pi) (5)
i=1

with the convention that 0 log 0 = O.

The choice of (5) can be justified by making recourse to certain axioms
of probability theory which cannot be discussed here6 . Suffice it to notice
that H is maximum when we have Pi = liN, Vi, i.e., the probability
is uniformely distributed among the different possible outcomes and it is
minimum when there is a j for which Pi = 1 and Pi = 0, for i /=j, i.e.
when the outcome is certain. This of course agrees with our commonsense
notion of uncertainty.

SIn the present context, whether the middle point x = 1/2 belongs to L or R
does not matter.
6These are the so-called "Khinchin axioms", see Khinchin (1957).
 173

To apply this notion to our system, we need to find a measure of

probability J.£ that assigns (nonnegative) probability values to subsets of
I (i.e., that tells us how likely it is that iterates of the map "visit" that
subset.) The measure must "agree" with the outcome-generating map T in
the sense that subsets of I of a given measure must be mapped to subsets
of equal measure. Mathematically, this property is denoted by the equation

(6)

where Ii is a subset of I and T- 1 (Ii ) denotes the pre-image of the set Ii

under T7. In this case, we say that J.£ is preserved by T or, equivalently,
that J.£ is invariant with respect to T.
The problem of finding an invariant measure (or to select among them
when there are many) is a difficult one and we shall discuss it here. In
our case, however, one can promptly verify that the map T preserves the
so-called Lebesgue measure. This a mathematically sophisticated version
of the common notion of length. If we take the interval [0,1] as a unit of
measure of length, the Lebesgue measure (henceforth denoted by m) is the
measure that assigns to subintervals values equal to their lengths. Figure
1 clearly shows that m is preserved by T.
We conclude that the information obtained by making one observation
of the system on the partition P = {L, R} (alternatively: by means of the
observation function P : I ~ {L, R}) is the following:

H(P"m) = -[m(L)log(m(L)) +m(R)log(m(R)]

= -[I/2log(I/2) + 1/ log(I/2)] = log 2
(remember that L = {O ::; x ::; I/2} and R = {I/2 < x ::; I})
When dealing with the predictability of a dynamical system, we are not
mainly interested in the entropy of a partition of the state space, i.e., the
information contained in a single observation, but we wish to know the
entropy of the system, defined as the rate at which replications of the
experiments - i.e., repeated, finite-precision observations of the system as
it evolves in time - produce information, when the number of observations
becomes very large.
For this purpose, let us then consider the partition formed by taking the
pre-image of Po under T, namely T-l(PO) and then performs the operation

This operation consists in taking all the possible intersections of the

elements of Po and T- 1 (po) and is called a "span", or a "joint". A moment's

7In equation (6), we make use of the pre-image of the map T to take into
account the case in which T is non-invertible. If T is invertible, (6) can be
equivalently written as J.£(T(Ii )) = J.£(Ii).
174

reflection will suggest that a cell of Pl corresponds to a sequence of two

states of the system evolving under T, observed in the partitioned state.
In our case, it corresponds to a sequence such as LL, LR, RL, or RR.
Shannon's entropy can again be used to measure the gain of information
obtained by such (double) observation of the system - or, equivalently by
performing the iteration on the map T twice.

112+-_~--+_--\

1/4 112 314

LA Rl

Figure 2. The partition Po = {L, R} and its counter-image.

Observing Figure 2, we can promptly see that the pre-image of L

under T consists of the two sub-intervals LL == {O $ x $ 1/4} and
RR == {3/4 $ x $ I}, whereas the pre-image of R under T consists of
the two sub-intervals LR == {1/4 < x $ 1/2} and RL == {1/2 < x $ 3/4}.
Then we have

T-l(L) n R = RR

and
T-l(R) n L = LR T-l(R) n R = RL
It follows that the partition P l consists of the four cells LL, LR, RL, RR
each of (normalized) length 1/4 and therefore of Lebesgue measure 1/4.
Consequently, we have

H(Pbm) = -4[1/4Iog(I/4)) = log 4 = 2 log 2

It is now easy to generalize our calculations to an arbitrary number of
observations of the system, constructing the partition

(7)

By replicating the procedure just described, we can verify that partition (7)
consists of 2n cells, each of them corresponding to a sub-interval of length
 175

2- n and therefore of Lebesgue measure 2- n . Thus the Shannon entropy of

partition (7) is equal to

Finally let us know consider the limit for n --+ 00 (i.e., when the number
of oservations becomes arbitrarily large) of the average uncertainty (=the
average gain in performing just another observation). We shall have

1 n 1
h(Po,m) = lim -H(V T-ipo) = lim _[_2n)Tn log(Tn)] = log 2
n---l>OC> n n---l>oo n
i=O
(8)
For any given map, the quantity defined by equation (8) depends on the
invariant measure m and on the partition Po. Let us now consider all
possible partitions of our state space, evaluate the entropy of each of them,
and find the largest one. We shall then write

h(m) = supph(P, m) (9)

The quantity (9) is called "metric" or "Kolmogorov-Sinai (KS) entropy".

In the sequel, the term "entropy" without qualifications will refer to the
KS-entropy. Clearly when h = 0 we can conclude that, if we observe the
(approximate) state of the system for a sufficiently long time, then there
will be no uncertainty left, i.e., there will be no additional information to
obtain from an extra observation. In other words, the past of the system
determines its future, i.e., the sytem is predictable. On the contrary, if the
entropy of the system is positive, we conclude that the knowledge of the
governing map and of the infinitely remote past history of the system is
not sufficient to predict its future; the system is unpredictable.
One might reasonably observe that the actual verification of equation (9)
might never be possibile: how could one evaluate all possible partitions of
the state space? Fortunately, there exists a result due to Kolmogorov and
Sinai which guarantees that the value of the entropy calculated for a certain
partition, called "generating partitions" is automatically the supreme value
over all possible partitions. Given a dynamical system defined by a map T
acting on a space M and a partition P = {PI, P2 , ... , PN} of M, the
partition P is said to be generating if an infinite sequence of symbols
Pi, (i = 1,2, ... ,N) uniquely identifies the initial point Xo of an orbit
of the system.
From our discussion of metric entropy, it is easy to see that this is indeed
the case for the tent map and the {L, R} partition. For this map it is also
true the inverse: to a given initial point Xo, there corresponds a unique
infinite sequence of L, R symbols.
Inspection of equation (8) will help us understand why zero (positve)
metric entropy implies predictability (unpredictability). The equation can
be looked at as the limit of a fraction whose denominator is the number of
176

observations, n which of course goes to infinity linearly with n. The numer-

ator of (8) is the Shannon entropy of the "super-partition" V~=o T-lpO,
namely:
n

(10)
i=O
The cells of that "super-partition" can be interpreted as sequences of the
symbols Land R of length n. Its entropy (i.e., its uncertainty) depends on
two factors: (i) the number of cells, namely the number of possible n-orbits8
of the system on the partitioned space L, Rj and (ii) the probability (with
respect to the invariant measure m) of each n-orbit.
It is clear that, for systems characterized by simple dynamics (e.g., orbits
tending to a fixed point, or to a periodic solution), once transients have died
out, the number of orbits with non-zero probability does not grow at all,
and entropy must be zero. Broadly speaking this means that observers of
the system will have no "surprises" and their prediction of the future course
of the system will be correct (within the prescribed approximation).
It is also clear from the definition of Shannon entropy that the limit
of the fraction (8) will be zero whenever the number of possible n-orbits
increases less than exponentially. This happens, for example, in the often
misunderstood case of quasi-periodic dynamics. In this case, the number
of possible (non-zero probability) sequences is not constant but grows at
a less than exponential rate. Consequently, their metric entropy is zero,
uncertainty disappears asymptotically and prediction is possible.
In chaotic systems, on the contrary, entropy (8) is positive since the
number of possible orbits N(n) grows exponentially. This can be easily
seen in the simple case of the "tent" map. For this purpose, consider that,
in this case the probability of each n-orbit (i.e., the probability of each cell
of partition V~=O T-lpO) is the same and equal Nlnj' The Shannon entropy
of the partition is consequently equal to

Therefore, the formula for the metric entropy is the following

log(N(n)
(11)
n
Clearly equation (11) tends to zero with n, unless N(n) grows exponentially
(or more than exponentially). We showed before that for the tent map this
is indeed the case.
Heuristically speaking, the chaotic, positive-entropy case can be de-
scribed as follows. Consider an observer of the system looking at a sequence

8 n-orbits of course means "orbits consisting of sequences of n terms" .

 177

of n, finite precision states of a dynamical system and trying to evaluate

its next state. Let each possible state be labelled by a symbol (such as L
and R). Broadly speaking, uncertainty depends on the possibility of the
system generating new combinations of symbols but, if the number of the
possible combinations is finite, the probability of their occurrence becomes
the smaller the longer the sequence and it goes to zero with lin. Only if
the system has the capability of generating new sequences ("surprises") at
a sufficiently high rate, can uncertainty remains forever. This is precisely
what happens for the "tent" map (as well as for other chaotic maps such
as the "logistic").
Notice that the capability of generating "surprises" is traditionally
attributed to stochastic processes. The most intriguing feature of chaotic
systems is precisely that, whenever observation is not infinitely precise, they
can behave stochastically. There is an important difference, however. The
statistical properties of stochastic processes are typically given exogenously
with respect to the deterministic (theoretical) model of the phenomena
under investigation. On the contrary, the statistical properties of chaotic
dynamics depend fundamentally on the structure of the model representing
those phenomena.

5 Stochastic Processes and Deterministic Chaos:

Bernoulli Systems

But this is not all. In fact, not all unpredictable (positive entropy)
systems are equally unpredictable. Whereas zero entropy implies that the
dynamics of a system is predictable with regards to any possible finite
partition, positive entropy simply means that the system is unpredictable
with regards to at least one partition. As we shall see in a moment, however,
there exist systems which are totally unpredictable in the sense that
they are unpredictable for any possible partition. Among these systems,
furthermore, there exists a special class called Bernoulli, which is the
"most chaotic" , or the "least predictable" of all.
Bernoulli deterministic systems are fundamental in at least two ways.
First of all, they are the essence of chaotic systems in the sense that for
any chaotic (positive entropy) system, there is at least one partition of the
state space on which its dynamics is isomorphic to a Bernoulli shift. This
property is sometimes described by saying that all chaotic (positve entropy)
systems must have a Bernoulli "factor". On the other hand, all factors of
a Bernoulli system are Bernoulli.
Secondly, the output of deterministic Bernoulli systems cannot be dis-
tinguished from that of certain stochastic processes. We shall return to this
point in a moment.
178

Keeping to the general tenor of this paper, we shall not discuss these
issues in their most general and abstract way. We shall instead try to
convey the basic idea of a "Bernoulli system" by means of the simplest
example. In so doing, we shall show that the "tent" (and the logistic map)
are indeed Bernoulli.
Let us consider again the partitioned state space E2 consisting of just
two elements L and R. Let us now construct an auxiliary state space whose
points consist of sequences of those symbols. We can now define a map a
on E2 in the following way: take a given, bilaterally infinite sequence of the
two symbols L, R and shift it one step to the left, thus:
(i) Before the shift:

... LRLLLRRLL.RRRLRLRRR . ..

(ii) Mter the shift:

... RLLLRRLLR.RRLRLRRRL ...

(where the arbitrary "decimal point" is introduced only for the sake of
presentation). The map a acting on the space E2 in the manner indicated
above is called "shift map".
Suppose now that the probability assigned to the elements Land R of
the L, R space is, say, PL = PR = 1/2 and consider the measure J-£u that
assigns to each finite sequence of L, R a probability equal to the product
of the probability of each element of the sequence. Clearly, J-£u is invariant
with respect to a. We now have a dynamical system (E2, a, J-£u), defined by
a state space, a map and a probability measure invariant with respect to
that map.
Analogous definitions are used for one-sided sequences. In this case, the
space i;2 is the space of (one-sided) sequences of L, R and the map a- acts
by shifting the sequence one step to the left and dropping the first element
of the sequence. In this case, the (one-sided) "shift map" will act as follows.
(i) Before the shift:
.RRRLRLRRR . ..
(ii) Mter the shift:

.RRLRLRRR(Lor R) ...

The system (E 2,a,{Lu) or ((i;2,a-,p.u,) is the simplest example of a

Bernoulli system.
Notice that, if we rename our symbols "Head" and "Tail", the system just
defined can be taken as a mathematical representation of the independent
stochastic process consisting of repeated tosses of a (fair) coin.
It can be shown that the "tent" (and thereby the logistic map) is
isomorphic to the system (i;2, a-, P.u), where the measure P.u is the "product
 179

measure" defined above. Isomorphism is a type of equivalence that - in the

case under investigation - can be defined by the following properties:
(i) the following diagram "commutes"

[0,1] T [0,1]
hl 1h
f;2 f;2

where 'commutes' means that hoT = a 0 h; T is the "tent" map; h is the

invertible function mapping the set [0,1] onto the space f;2 of (one-sided)
infinite sequences of symbols L, R as defined above; a is the (one-sided)
shift map.
(ii) the map h preserves the probability structure, i.e., if A and Bare,
respectively, measurable subsets of [0,1] and f;2, then m(A) = J.Lu 0 h(A)
and f.lu(B) = m 0 h-1(B). Here m is the Lebesgue measure preserved by
T and f.lu is the product measure preserved by a as discussed above. The
commuting property (i) is easily verified by making use of the fact that
the partition L, R is a generating partition for the tent map. Property
(ii) can also be ascertained verifying that the probability assigned by the
Lebesgue measure to a given sub-interval of [0,1]' i.e., its length, is equal
to the probability of the set of (one-sided) sequences of the L, R symbols
corresponding to the set, evaluated by means of the a-invariant measure
f.lu (i.e., the measure that assigns to any sequence of n symbols L, R the
product of the probability of each symbol (in this case 1/2 for both L and
R).
Isomorphism preserves certain statistical properties, in particular the
Lyapunov characteristic exponent and the metric entropy. The class of pro-
cesses which are isomorphic to a Bernoulli process are called B-processes.
It includes both deterministic processes - e.g., besides the the tent map, the
logistic map (for r = 4); the Lorenz geometric model; the Silnikov model
- and stochastic processes - e.g., the Li.d. processes; the continuous-time
Markov processes. The B-processes are characterized by an asymptotic form
of independence called "Very Weak Bernoulli". The latter broadly speak-
ing implies that the future of a process conditioned on its past uniformly
converges, in a certain appropriate metric, to the unconditioned past. For
Bernoulli maps, the uncertainty concerning "the future" (as represented
by the partition P), conditional on "the past" (the partition V'::l T-iP)
asymptotically and uniformly tends to that of the unconditioned future.
The stochastic properties of the behaviour of chaotic deterministic is
dramatically revealed by some recent results on B-processes. In particular,
it was showed that for a deterministic system belonging to the class of
B-processes we can define a stochastic k-state Markov chain such that the
dynamics of the former is arbitrarily close - in an appropriately define sense
- to that of the latter.
Since the mathematics behind these results is rather hard, we cannot
180

provide a full discussion of this question here. In line with the strategy of
this paper, we shall instead illustrate the main idea by means of a simple
application to the tent map, referring the reader to the relevant literature
(see Ornstein and Weiss, 1991; Radunskaya, 1992) for a more rigorous and
general discussion and proofs.
For the tent map, and for a given partition of the state space I = [0, 1],
we can define a Markov chain on k states (the number of states depending
on the partition), such that its sample paths are indistinguishable (within
the prescribed resolution, or precision of observation) from the orbits of the
deterministic tent map9. For example, if the partition is {L,R}, i.e., if we
can only tell whether the state of the system is on the left or on the right
of the middle point of the state space [0,1]' then a Markov chain on two
states Land R, with transition matrix

(12)

will generate sample paths indistinguishable (within the prescribed degree

of accuracy) from those of the tent map on the partitioned space {L, R}.
If we now increase the precision of observation so that we can now
locate point in the state space within sub-intervals of length 2- k ), we can
construct a 2k state Markov chain with transition matrix:
1/2 1/2 o 0 o o
o 0 1/2 1/2 o o
o o o o 1/2 1/2 (13)

o 0 1/2 1/2 o o
1/2 1/2 o 0 o o
Again it will not be possible (within the prescribed precision) to distin-
guish between a typical sample path of the Markov chain and an orbit of
the tent map on the 2k -partitioned state space.
The same exercise could be performed in relation to the "logistic" map
with chaotic parameter, although in that case the construction of the
transition matrix would be more difficult.

6 Conclusion

The implications for economics of the results just obtained are puzzling.
For example, consider the case in which a model of optimal growth gives

9I owe this example to a private correspondence with Amy Radunskaya.

 181

rise to a dynamic equation of the "logistic" type with chaotic parameter.

The sequences thus generated are optimal in the sense that they solve a
problem of intertemporal maximization of rational agents, in an economy
satisfying the requirements of competitive equilibrium at each point of
time. In the absence of exogenous, random disturbances, along optimal
trajectories agents' expectations are supposed to be always fulfilled, Le.,
we assume agents' perfect foresight.
While the latter assumption may be acceptable when the dynamics of
the system are simple (e.g., convergence to a steady state or to a periodic
orbit), it makes little sense if the dynamics are chaotic, eminently so
if they are Bernoulli. In our case, the information set on which agents
base their predictions typically consists of the observations of past values
of the relevant variables. But we have just demonstrated that - if the
entropy of the system is positive - knowledge of the infinitely remote past
with arbitrarily (but not infinitely) great precision of measurement is not
sufficient to forecast future values correctly. Moreover, when the system is
Bernoulli, agents face sequences of values which have the same probabilistic
structure as random processes. Intuitively speaking, in this case the past
no longer determines the future and the assumption of perfect foresight
lacks any motivation.
The equivalence result between (Bernoulli) deterministic and (Markov)
stochastic processes discussed above can be given a more or less optimistic
interpretation according to one's point of view and temperament. The re-
sults discussed above indicate that we cannot hope of providing a generally
valid test for distinguishing deterministic chaos and "true" randomness.
This would certainly be impossible for Bernoulli systems and, to the extent
that the conjecture that "most observable chaos is Bernoulli" (see Ornstein
and Weiss op. cit, p. 22.) is correct, it would be generally impossible. Con-
sequently, at least for a certain class of concrete dynamical systems, the
possibility exists of representing them either as deterministic systems (plus
perhaps some random disturbances), or as stochastic processes. The choice
between the two is a matter of expedience rather than theory. In principle,
a deterministic representation is superior for explanation purposes, but this
is only true if we can provide a physical (economic) interpretation of the
state variables and the functional relationships among them.

References
[1] Benhabib, J. and Day, R.H. (1981). Rational choice and erratic be-
haviour. Review of Economic Studies, 48, 459-47l.

[2] Benhabib, J. and Day, R.H. (1982). A characterization of erratic

dynamics in the overlapping generations models. Journal of Economic
182

Dynamics and Control, 4, 37-55.

[3] Blanchard, O.J. and Fischer, S. (1987). Lectures on Mathematics.
Cambridge, Mass.: MIT Press.
[4] Boldrin, M. and Montrucchio, L. (1986). On the indeterminacy of
capital accumulation paths. Journal of Economic Theory, 40, 26-39.
[5] Deneckere, R and Pelikan, S. (1986). Competitive chaos. Journal of
Economic Theory, 40, 13-25.
[6] Farmer, J.D. and Sidorowich, J.J. (1987). Predicting chaotic time
series. Physics Review Letters, 59, 845-848.
[7] Ford, J. (1983). How random is a coin toss? Physics Today, 36, 40-48.
[8] Frisch, R (1933). Propagation problems and impulse problems in
dynamic economics. In Economic Essays in Honour of Gustav Cassel,
171-205. London: Allen and Unwin.
[9] Goodwin, RM. (1951). The nonlinear accelerator and the persistence
of business cycles. Econometrica, 19, 1-17.
[10] Grandmont, J.M. (1985). On endogenous competitive business cycles.
Econometrica, 53, 995-1045.
[11] Granger, C.W.J. and Newbold, P. (1977). Forecasting Economic Time
Series. New York: Academic Press.
[12] Hicks, J.R (1950). A Contribution to the Theory of the Trade Cycle.
Oxford: Oxford University Press.
[13] Kaldor, N. (1940). A model of the trade cycle. Economic Journal, 50,
78-92.
[14] Khinchin, A.I. (1957). Mathematical Foundations of Information The-
ory. New York, Dover Publ.
[15] Lorenz, H.W. (1989). Nonlinear Dynamical Economics and Chaotic
Motion. New York: Springer Verlag.
[16] Mane, R, 1987, Ergodic Theory and Differentiable Dynamics,
(Springer-Verlag, Berlin)
[17] Medio, A., 1992, Chaotic Dynamics. Theory and Applications to Eco-
nomics, Cambridge University Press
[18] Medio, A. and G. Negroni, 1996, Chaotic Dynamics in Overlapping
Generations Models with Production, in: Barnett, W.A., A. Kirman
and M. Salmon (eds.), Nonlinear Dynamics and Economics, Cam-
bridge University Press
 183

[19] Ornstein, D.S. and B. Weiss, 1991, Statistical Properties of Chaotic

Systems, Bulletin (New Series) of the American Mathematical Society
24, N.1, 11-115.
[20] Radunskaya, A., 1992, Alpha-Congruence of Bernoulli Flows and
Markov Processes: Distinguishing Random and Deterministic Chaos,
unpublished Ph.D. Thesis (Stanford University).
[21] Ruelle D., 1989, Chaotic Evolution and Strange Attractors, (Cam-
bridge University Press, Cambridge).
[22] Slutsky, E. (1927). The summation of random causes as the source of
cyclical processes. Econometrica, 5, 105-146.
[23] Yule, G.U. (1927). On a method of investigating periodicities in
disturbed series. Philosophical Transactions, 226A, 267-298.
Long-Term Memory and
Chaos: a Note
Gilbert ABRAHAM-FRO IS , Sandrine LARDIC and Valerie MIGNON

1 Introduction
Is it possible to study jointly long-term memory processes and chaotic pro-
cesses? In a first sight, this interrogation can appear somewhat surprising
because of the properties showed by the two kinds of processes. Effectively,
a long-term memory process, like an ARFIMA process, is a stochastic one,
while a chaotic process is by definition a deterministic one. However, this
question finds its origins in recent works of Peters (1991, 1994) setting
number of relations between the two processes. On the one hand, Peters
showed that long-term memory and chaos concepts can be linked by the
mean of the fractal dimension: "Fractal time series are characterized as
long memory processes. They possess cycles and trends, and are the result
of a nonlinear dynamic system, or deterministic chaos" (Peters, 1991, p.
119). In this way, a long-term memory process would be a process whose
attractor has a fractal dimension. We can thus think that the link between
fractal dimension and chaos might be set by the mean of the concept of
strange attractor. However, we will show that this relation is useless in order
to detect the presence of chaotic dynamics. On the other hand, according
to Peters, long-term memory can be equivalently detected by Rj S analysis
and Lyapunov exponents since there exist nonperiodic cycles for the two
processes: "The long memory effect in equity prices has now been confirmed
by two separate types of nonlinear analysis. RIS analysis on monthly S&P
500 stock returns found a biased random walk with a memory length about
four years. The Lyapunov exponent for monthly inftation-detrended S&P
500 prices found a 42-month cycle" (Peters, 1991, p. 180).
This note proposes a critical study of links established by Peters (1991,
1994) between chaotic and long-term memory processes. To this end, after
having presented definitions of the main concepts concerning long-term
memory and chaos, we study the relations between the two notions. We
show that these links are only superficial and erroneous.
186

2 Long-term memory and chaotic processes:

definitions

In a first time, we define the main long-term memory processes. The

first long-term memory processes have been presented in terms of fractal
processes because of their self-affinity property (fractional Brownian motion
and fractional Gaussian noise). In a second time, we define the main
concepts relative to chaotic dynamics.

2.1 LONG-TERM MEMORY REPRESENTATION: FRACTAL

PROCESSES AND ARFIMA PROCESSES

The first long-term memory process, the so-called fractional Brownian

motion, was defined by Mandelbrot and Van Ness (1968) and constitutes a
generalization of the ordinary Brownian motion for which increments were
independent. H, the key parameter of this process, characterizes the self-
affinity property of fractional Brownian motion. This parameter is called
the Hurst exponent in reference to the English hydrologist H.E. Hurst. The
Hurst exponent is particularly interesting because it allows to classify time
series according to their dependence structure: null or short-term memory,
long-term memory (Joseph effect) and anti-persistence phenomenon. Be-
cause economic series are discrete time series, Mandelbrot and Wallis (1969)
have derived a discrete time analogue - the fractional Gaussian noise - of
the continuous time fractional Brownian motion. This process also exhibits
the persistence phenomenon. Another class of discrete time models linked to
fractional Brownian motion is constituted by ARFIMA (Auto-Regressive
Fractionally Integrated Moving Average) ·processes. These models are a
generalization of standard ARIMA(p, d, q) processes for which the differ-
encing parameter d was an integer. Traditionally, econometricians consider
either non stationary series (i. e. series with a unit root) or stationary series
( i. e. series without such a unit root). In the first case, time series are mod-
eled by ARIMA(p, d, q) processes, with d generally equal to one, and, in
the second case, by ARMA(p,q) processes (i.e. d = 0). ARIMA processes
refer to an infinite memory and ARMA processes to short-term memory.
Hence, the intermediate case of long-term memory, which corresponds to
a fractional d, has been ignored. In order to take into account the persis-
tence phenomenon, Granger and Joyeux (1980) and Hosking (1981) have
developed ARFIMA(p, d, q) processes where d can take real values and not
only integer ones. These processes are linked to the Hurst exponent by the
mean of a relationship between the fractional differencing parameter and
the Hurst exponent, derived from the auto covariance function of fractional
Gaussian noise.
 187

Generalization of Brownian motion: fractional Brownian motion

A Brownian motion B(t, w), is a continuous-time stochastic process given

by:

B(t,w) = J~oo W(s)ds

where W(s) is a Gaussian white noise.
A Brownian motion is characterized by independent Gaussian increments
and a spectral density of w- 2. In other words, B(t,w) is such that:
• B(t 2, w)-B(tI, w) has zero mean and a variance proportional to it2-tli .
• B(t2,W) - B(tI,w) and B(t4,W) - B(t3,W) are independent, if the
intervals (tl' t2) and (t3,t4) do not overlap.
Standard deviation of the increment B(t + T, w) - B(t, w), for T > 0, is
equal to Tt, which is known as the Tt law.
Because of the independence between increments, the process displays
only short-term memory. Mandelbrot and Van Ness (1968) proposed an
extension of this ordinary Brownian motion in order to take into account
the long-term memory phenomenon, i.e. the possibility of a dependence
between distant observations. This generalization gives the so-called frac-
tional Brownian motion.
Fractional Brownian motion, denoted as B H (t, w), is a zero mean Gaus-
sian stationary process, defined as:

BH(t,w)-BH(O, w) 1 {J~oo [(t - s)H-t - (-s)H-t] dB(s, w) }

r(H +~) + J~(t - s)H-tdB(s,w)
where bo = BH(O,W)
if bo = O,we have:

Thus, if H = !,fractional Brownian motion reduces to ordinary Brow-

nian motion and becomes a short-term (or null) memory process.
To simplify notations, let us write BH(t) for BH(t,W). The variance of
fractional Brownian motion is given by t 2H and its covariance:

The key parameter of fractional Brownian motion is H, where <

H < 1, which is called the Hurst exponent. This parameter appears in
°
a relationship linking .6.BH (t) and .6.t:
188

where
t::.BH(t) = BH (t2) - BH (tt) t::.t = t2 - t1
This relationship illustrates the self-affinity property of fractional Brownian
motion. In other words, this kind of process is statistically self-affine in the
sense that {B H (rt), t 2:: o} has the same finite dimensional distribution as
{rHBH(t),t ~ o} for all r 2:: 0 1
From the covariance function of fractional Brownian motion, we can
derive the correlation between observations separated by a very long time
span as:
C = 22H - 1 -1
In this sense, fractional Brownian motion is a long-term memory process.
Actually, we know that, in the case of a short-term memory process, the
correlation goes to zero when the distance between the two points under
consideration goes to infinity. In the case of fractional Brownian motion,
this correlation tends to 22H -1 - 1 1=0 if H 1= ~. From this correlation, it
is possible to classify time series according to their dependence structure:
• when H = !, fractional Brownian motion reduces to ordinary Brownian
motion. The correlation C is null and the process does not exhibit long-term
memory phenomenon.
• when! < H < 1, the correlation C is positive. The process displays
long-term memory and exhibits the persistence phenomenon.
• when 0 < H < !, the correlation C is negative. The process is anti-
persistent.
Economic time series are typically discrete time series. Hence, in order to
model these series by a long-term memory process, it is necessary to derive
a discrete time analogue of continuous time fractional Brownian motion.
This is accomplished with the discrete time fractional Gaussian noise and
ARFIMA processes.

Discrete time: fractional Gaussian noise and ARFIMA processes

Definitions
Fmctional Gaussian noise.
The fractional Gaussian noise yt, introduced by Mandelbrot and Wallis
(1969), is the increment of fractional Brownian motion:
yt = BH(t) - BH(t - 1)

INote that self-affinity is different from self-similarity although the two notions
are widely assimilated in the literature. Unlike self-similar curves, a self-affine
process (like the fractional Brownian motion) requires different scaling factors in
the two coordinates: r for t, but rH for BH(t) reflecting the special status of the t
coordinate. Each t correspond to only one value of BH but any specific BH may
occur at multiple t's. Such non-uniform scaling is known as self-affinity rather
than self-similarity.
 189

It is a zero mean, Gaussian stationary process whose auto covariance

function is given by:

and
lim k 2- 2H ,(k) = H(2H -1)
k-+oo

The spectral density is given by:

f(w; H)=112(2H)-2H-2r(2H +1) sin(7rH)4sin 2 (~) f

n=-CX)
In + (~) 1- 2H 1
-

and lim w 2H - 1 f(w; H) = (2112) r(2H + 1) sin(7rH)

w-+O 7r

Thus, by analogy with the fractional Brownian motion, when H = !,

,(k) = 0 and the Yt's are independent. When! < H < 1, the Yt's are
positively correlated and they display long-term memory. Finally, when
0< H < !, the process exhibits anti-persistence phenomenon.
Thus, fractional Gaussian noise is a discrete time analogue of the
continuous time fractional Brownian motion. However, the definition of
this process is essentially based on its autocorrelation function 2 . The
idea of Granger and Joyeux (1980) and Hosking (1981) is to develop
a truly discrete time process by looking for a discrete time version of
fractional Brownian motion rather than a covariance function. This gives
rise to ARFIMA(p, d, q) processes (Auto-Regressive Fractionally Integrated
Moving Average). They generalize the standard ARIMA processes in which
the differencing exponent d was an integer. In the case of ARFIMA
processes, d can take real values and not only integer ones. A fractionally
integrated time series is characterized by a dependence between distant
observations as one can observe it in the auto covariance function or in
the spectral density function. One would remark thus that these types of
processes allow to reduce the constraints on autoregressive and moving
average coefficients of parametric models 3 .
ARFIMA processes.
These processes constitute a generalization of ARIMA(p, d, q) processes
for which the differencing parameter d was an integer. In the case of

2Recall that its autocorrelation function is the same as whose of the fractional
Brownian motion in first differences.
3 A long-term memory process can always be approximated by an ARMA(p, q)
process, but the orders p and q necessary to obtain a relatively good approxima-
tion can be too great and render parameters estimation very difficult.
190

ARFIMA processes, d can take real values and not only integer ones. The
case of a non-integer differencing parameter is important in terms of long-
term memory. In ARMA(p, q) processes, d = 0, and refers to a time series
which has zero (or short-term) memory. In ARIMA(p, d, q) processes, dis
typically equal to one and refers to infinite memory. Thus the intermediate
case, that is d fractional, is ignored. Typically, the class of ARFIMA
processes can be used to model data dependence which is stronger than
allowed in stationary ARMA processes and weaker than implied by unit
root processes.
ARFIMA(O, d, 0) process
The analogous in discrete time of Brownian motion is the random walk
or ARIMA(O, 1,0):
{l-L)X t =Ut
where Ut is an independently and identically distributed (iid) variable. Thus
the first difference of X t is the white noise Ut in discrete time. By analogy,
one defines the fractional white noise of parameter d by:

where Ut is a white noise and (1- L)d is given by the binomial expansion:

where

r(j - d) II (k -1- d) j = 0,1, ...

7rj = r(j + l)r{ -d) = O<k~j k

and
10
00 fI:-le-tdt
if x > °
°
°
r(x) = 00 ifx=
x- 1r(1 + x) if x <
Such a defined process is called ARFIMA{O, d, 0). As the properties of
this model have been widely studied by Hosking (1981), we just synthesized
here his main results.
Let {Xt, t = 0, 1, ... } be an ARFIMA{O, d, 0) process. Then:
(i) When d < ~, {Xt} is stationary
(ii) When d > -~, {Xt} is invertible.
Hence, the {Xt } process is stationary and invertible only if -~ < d < ~.
The autocorrelation function decreases at an hyperbolic rate, thus much
slower than the simple ARMA autocorrelation function (which decreases at
a geometric rate). Because of this property, ARFIMA processes are called
long-term memory processes.
 191

The autocovariance function of ARFIMA(O, d, 0) process is given by:

r(l - 2d)r(k + d)
-y(k) = r(d)r(l _ d)r(k + 1 - d)

When k goes to infinity, we have:

. r(l - 2d) 2d-t

kJ:.~ -y(k) = r(d)r(l _ d) Ikl
The main advantage of ARFIMA(O, d, 0) process over fractional Gaussian
noise is that it has a particularly simple spectral density:

f(w,d) = (2sin~) -2d for 0 < w ~ 71'

and limw-+o f(w,d) = Iwl- 2d
Generalization:ARFIMA(p, d, q) process
The ARFIMA(O, d, 0) process is a special case of ARFIMA(p, d, q) pro-
cess where dE] - ~, ~ [ which can be defined like this:

where:
ft = V'-d Ut

Ut : W N(O, 0'2)
<I>(L) and 8(L) are lag polynomials of p and q degrees respectively.
Hence,

with
d(d + 1) d(d + l)(d + 2)
ft = Ut + dUt-t + 2! Ut-2 + 3! Ut-3 + ...
As in the case of ARFIMA(O, d, 0) processes, ARFIMA(p, d, q) processes
are long-term memory, stationary, and invertible when dE] -~, H and
d FO.
Relationship between Hurst exponent and fractional differencing parameter

If we compare the autocovariance functions of fractional Gaussian noise

and ARFIMA processes, we see that they have the same power decay. Thus,
relating the exponents in these two expressions gives:
1
d=H--
2
192

From this remarkable relationship it is possible to classify time series

according to their dependence structure as in the case of the Hurst
exponent:
• if d = 0, ARFIMA(P, 0, q) process becomes a stationary ARMA process
and exhibit only short-term memory. The spectral density at frequency zero
is finite and positive.
• if 0 < d < ~, the process is persistent: it is a stationary process with
long-term memory. Autocorrelations are positive and decay hyperbolically
to zero as the lag grows. The spectral density is concentrated around low
frequencies: it increases to very high values as the frequency approaches
zero.
• if - ~ < d < 0, the process is anti-persistent. The spectral density
is dominated by high frequency components and declines to zero as the
frequency approaches zero.
Following section describes and defines some properties of chaotic pro-
cesses.

Chaotic processes
Introduction: Non-linearity and determinism

A chaotic process is a non-linear deterministic process. Chaotic pro-

cesses are necessarily nonlinear. It is well known that linear models can
only generate four types of behavior: oscillatory and stable, nonoscillatory
and stable, oscillatory and explosive, nonoscillatory and explosive. On the
contrary, nonlinear models can generate much richer types of dynamics.
Moreover, despite their deterministic characteristic, trajectories generated
by chaotic processes appear completely random by standard linear time
series methods. Deterministic chaos is characterized by self-sustained oscil-
lations whose period and amplitude are nonrepetitive and unpredictable.
Following Brock (1986), a time series {Yt}, t = 1, . .. , T, has a determin-
istic explanation if there exist a system {h, J, Xo} such that h maps 3?n to
3?, J maps 3?n to 3?n and:
• Yt = h(Xt)Vt
• X t = J (X t - 1 )
• Xo is the initial condition at t = o.
The map J is the unknown deterministic law of motion.
The most well known example of chaotic process is given by the logistic
map:

The dynamics of this equation depends on the value of J.I.. If 0 ~ J.I. ~ 3

the system converges to a stationary state. However, as J.I. increases beyond
3, the system undergoes various qualitative changes, known as bifurcations.
 193

For /1- > 3 there are cyclic solutions. The system goes through an infinite
number of period-doublings: stable cycles of length 2, 4, 8, 16, ... , 2n
,... successively appear and become unstable. This sequence continues
until a limit point /1-00 which is approximately equal to 3.57. At this
point the system behaves in a very complex manner. We can illustrate
some properties of this equation by plotting the bifurcation diagram4
(see Abraham-Frois and Berrebi (1995)). For values of /1- between 3.57
and 4, the bifurcation diagram exhibits black bands which correspond to
chaos separated by windows. In these windows, odd-order cycles arise. For
example, for 3.828 ... < /1- < 3.849 ... a stable cycle of period 3 emerges. It
is successively replaced by stable cycles of period 6, 12, ... , 3.2n , ... These
bifurcations accumulate at a limit point where the behavior is similar to
that at /1-00. This is generally referred to as order within chaos: chaotic
systems have an underlying order. Moreover, a further property of the
logistic equation is that if any window is magnified, the diagram displays
an exact copy of the complete diagram. This illustrates the self-similarity
characteristic of the logistic map. For Peters (1991, 1994) this characteristic
illustrates the fact that there exist a link between fractals and chaos. A
further property of chaotic systems concerns the presence of an attractor.

Attractors and sensitive dependence on initial conditions

Intuitively, we can define an attractor as a subset toward which almost

all sufficiently close trajectories are attracted asymptotically. Thus, an
attractor gives a global picture of the long-term behavior of a dynamical
system.
More formally we retain the following definition.
Definition:
Given the map J such that: Xt+l = J(xt}
A compact set A is an attractor if there is an open set U in the
neighborhood of A in phase space such that: P(x) -+ A if
t -+ 00 for x E U. Union of all U-neighborhoods of A is the basin of
attraction of A.
There exist several types of attractors. The most well-known are the fixed
point, the limit cycle and the torus:
• Attracting fixed point: trajectories are attracted toward a fixed point
equilibrium, which reflects a time independent solution. It is a steady state:
Jt(xp) = xp, \It. This fixed point is the typical attractor of linear systems
and constitutes the typical solution of a great number of economic models.
• Attracting limit cycle: attractor of a periodic function, which is
characterized by its amplitude and period. Trajectories converge toward
this closed curve, named limit cycle. Recall that a point x is a periodic

4The bifurcation diagram is the set of possible solutions in the equation.

194

point of period T if fT(x) = x, but ft(x) /=x for 0 < t < T.

<P = {ft(x) : 0 < t < T} is the corresponding periodic orbit.
• The r-dimensional torus Tr , r ~ 2: this is the attractor for quasiperi-
odic motion with r different periods.
These three attractors are described periodically by trajectories of the
considered system. The behavior of these systems is predictable: given
the initial state, it is possible to predict future states. Moreover, systems
which have these attractors are not sensitive upon initial conditions: two
nearby trajectories will not diverge over time. But, the main characteristic
of chaotic process is precisely the sensitive dependence on initial conditions
whose a definition can be state as follows.

Definition:

A map f on a metric space X is sensitive on initial conditions if there is

a number r, r > 0, such that for all x E X and for all
E> 0 there exist:
• a point y E X such that d(x, y) < E
• and an integer k, k ~ 0, such that d (Jk(x), fk(y)) ~ r
where d(x, y) is the distance between x and y and fk is the kth iterate
of f.

Thus, sensitive dependence on initial conditions means that the distance

between two points that were initially separated by an infinitesimal amount
grows exponentially fast on the average. This is the main property of
chaotic processes. This sensitivity is formalized by the condition that the
largest Lyapunov exponent, which measures the rate of spread of nearby
trajectories in phase space, be positive.
In order to illustrate the sensitivity on initial conditions, the concept of
strange attractor was introduced by Ruelle and Takens (1971). However
the term "strange" is somewhat ambiguous. It refers also to the fractal
nature of attractors 5 • In many well-known examples, chaotic attractors
are also strange; e.g., the Henon map exhibits exponential divergence of
neighborhood trajectories and it has a Cantor set structure. However, there
exist attractors which are fractal but not chaotic: for these attractors,
there is no sensitive dependence on initial conditions6 . On the contrary,
there exist attractors which are chaotic but not fractaf (their dimension
is an integer). Thus, we define a strange attractor to be an attractor with
fractal structure. We define a chaotic attractor to be an attractor with

SIn general, a fractal object is an object for which the Hausdorff dimension is
different from the topological dimension, and usually not an integer. Moreover,
fractal objects are self-similar.
BSee Grebogi, Ott, Pelikan and Yorke (1984).
7 Arnold's cat attractor is chaotic, but not fractal, see Eckmann and Ruelle
(1985).
 195

sensitive dependence on initial conditions. Consequently, a system which

has a chaotic attractor is said to be chaotic. A consequence of the sensitivity
upon initial conditions is the long run unpredictability.
Long-term prediction.
Long-term predictions of chaotic systems are virtually impossible, be-
cause errors in measurement of the initial state propagate exponentially
fast. Because we can never measure current conditions to an infinite amount
of precision, we cannot predict where the process will go in the long term.
Thus, the error, even if very small, will increase and affect the macroscopic
state of the system. Consequently, the least error in the initial conditions
render long run prediction impossible. This is another fundamental prop-
erty of chaotic systems: even if these systems are deterministic, their be-
havior is unpredictable, except perhaps in the very short run. Moreover,
this behavior can emerge in a system with a small number of degree of
freedom: only one degree of freedom in discrete time (logistic map)8 and
only three in continuous time (Lorenz model). For this reason, this kind of
chaos is called deterministic chaos or low dimensional chaos. The following
section studies links established by Peters (1991, 1994) between long-term
memory and chaos.

3 Relations between ARFIMA and chaotic

processes

According to Peters (1991, 1994) long-term memory and chaotic pro-

cesses can be jointly analyzed since there exist links between the two pro-
cesses. On the one hand, we will see that a link between long-term memory
and chaotic processes might come from a relationship between fractal di-
mension, Hurst exponent and fractional differencing parameter of ARFIMA
processes. However, we will show that such a relation cannot be used in
order to detect the presence of chaos. On the other hand, according to Pe-
ters (1991, 1994), calculation of Lyapunov exponents allows to detect the
presence of long-term memory in a time series. In this sense, Lyapunov
exponents might be linked with RIS analysis.
Beside these possible similarities, we will insist on the major differences
between long-term memory processes and chaotic processes. On the one
hand, long-term memory processes are predictable on the long run while
chaotic processes are unpredictable because of sensitivity on initial condi-
tions. On the other hand, a chaotic process is a deterministic one, while a

8More precisely, a necessary condition for a discrete system to exhibit chaotic

dynamics is that it has at least one degree of freedom if it is irreversible (logistic
map) and at least two degrees of freedom if it is reversible (Henon map).
196

long-term memory process, such an ARFIMA process, is a stochastic one.

3.1 LONG-TERM MEMORY AND FRACTAL DIMENSION

Can we establish a link between series whose underlying process is

characterized by a long-term memory and series whose underlying process
has a fractal dimension? In other words, does one of these characteristic
imply the second one?
In a first time, it can be tempting to conclude to the existence of such
a relation between long-term memory and fractal dimension. Effectively,
the Hurst exponent H is linked to fractal dimension D according to the
relation D = 2 - H. This dimension is the dimension of the time path of
fractional Brownian motion. Moreover, we know that fractional differencing
parameter d of ARFIMA processes is linked to Hurst exponent by the
relation d = H - !.
Thus, the fractal dimension of time path of ARFIMA
processes is given by: D = ~ - d. It is possible to establish the following
classification:
• if dE] -!; 0 [ then D E H; 2[
• if d = 0 then D = ~2
• if dE ]0; ~ [ then DE] 1; ~ [ .
!
Thus, if H = (d = 0), the dimension is ~ which is effectively a fractal
dimension. But d = 0 corresponds to a series whose underlying process
is not characterized by long-term memory. As a consequence, there is no
equivalence between fractal dimension and long-term memory. In other
words, a time series can be characterized by an underlying process which
has a fractal dimension without being a long-term memory process. This
is the case of Brownian motion which is a short-term memory process
characterized by a fractal dimension of ~.
However, we observe that a long-term memory series is characterized by
a fractal dimension between ~ and 2 in case of anti-persistence and between
1 and ~ in case of persistence. From these relations, it might be concluded
that, on the one hand, a long-term memory series has a fractal dimension
and, on the other hand, a series with fractal dimension is not necessarily a
long-term memory series.
Nevertheless, beyond these relations between long-term memory and
fractal dimension, a remark must be state: the value of fractional differenc-
ing parameter cannot allow us to conclude if a series is really characterized
by a long-term memory. It must be studied together with the standard devi-
ation of d in order to decide if d is significantly different from zero. In other
words, the measure of the persistence degree in a series is fundamentally
based on statistical tools: calculation of d is the result of an estimation pro-
cedure. The estimation obtained for d is a random variable. In fact, when
we use various methods for estimating d, results can be sometimes divergent
(e.g., Lardic and Mignon 1996a,b,c). We have previously run estimations
 197

for various series of exchange rates, stock returns and macroeconomic series
which illustrate the random character of this parameter: these procedures
(Geweke and Porter-Hudak, exact maximum likelihood, R/S analysis ... )
do not always detect long-term memory in same series, and the signs of d
are sometimes opposite (table 1).
This table relates the estimation of d according to R/ S analysis, modified
R/ S analysis (Lo, 1991), Geweke and Porter-Hudak method, and Sowell
procedure (exact maximum likelihood). Real monthly data in logarithmic
first differences, January 1974-November 1994. In brackets: t-value of d.
This table leads us to a second remark: is the relation between d and H
empirically verified? If we accept this relation from a theoretical point of
view, it seems somewhat unreasonable to use it empirically. This conclusion
directly comes from our previous remark: being not able to determine a
unique estimated value of d, how to determine a unique relation between d
and H? Table 1 illustrates this difficulty: the use of the theoretical relation
d = H - ~ gives estimations of d which are very different from these
obtained by using other methods.
Nevertheless, the dimension obtained by the relation D = 2 - H is the
dimension of a stochastic process, such as an ARFIMA process. Thus, it
cannot be the dimension of a chaotic process (or of its attractor) which is, by
definition, a deterministic process. Moreover, it should be pointed out that
the dimension given by the relation D = 2 - H is the dimension of a time
path, i.e. a time series, and not the dimension of an attractor. Effectively,
the dimension of a time series is between 1 et 2 whereas the dimension of an
attractor can be more important. Thus, the use of the relation D = 2 - H
is useless if the object under study is the calculation of a dimension of a
chaotic attractor. Moreover, we have previously mentioned that a chaotic
attractor is not necessarily an attractor with fractal dimension. Thus,
finding a fractional dimension is not enough to conclude in terms of chaotic
dynamics 9 .
As a consequence, if, as suggested by Peters (1991, 1994), there exist
a link between long-term memory and chaotic processes, this link cannot
be set through the relation D = 2 - H. A second relation between long
memory and chaos has been established by Peters by the mean of R/ S
analysis and Lyapunov exponents.

3.2 LONG-TERM MEMORY AND CHAOS: R/ S ANALYSIS AND

LYAPUNOV EXPONENTS

We have previously observed that, because of the sensitivity on initial

conditions, chaotic processes are unpredictable on the long run. On the

9See the previous distinction between strange attractor and chaotic attractor.
198

contrary, a long memory process is predictable on the long run. We have

thus here a fundamental opposition between the two processes. Despite this
fact, Peters (1991, 1994) showed that the link between long-term memory
and chaos can be obtained through R/ S analysis and Hurst exponent. The
object of R/ S analysis is to detect the presence of long-term memory in
a time series. Moreover this statistic is able to detect nonperiodic cycles
whose we can calculate the average length. This length corresponds to time
during which an exogenous chock will affect the system. Peters (1991) found
that the underlying process of American stock index SP 500 is a long-term
memory one, since the estimated Hurst exponent is 0.78. The duration
of this memory is about 48 months. By using tools for detecting chaos,
Peters found that the value of the largest Lyapunov exponent is 0.0241
bit/month. As a consequence, if we could measure initial conditions to one
bit of precision, we would lose all predictive power after 1/0.0241, or 42
months'time.Since the Lyapunov exponent is positive and the average cycle
length obtained by R/S analysis is roughly equal to the length estimated
using Lyapunov exponent, Peters (1991) writes that American stock market
is characterized by chaotic dynamics. In his second book, Peters (1994)
makes the same reasoning for the Dow Jones: the Dow Jones is a long-term
memory series (H = 0.7, chapter 8) and its underlying process is chaotic
(chapter 16)!
These results seem somewhat inconsistent. How can a series be gener-
ated both by a long-term memory process and a chaotic process? Peters
writes then that on the short run (horizon less than four years), markets
are characterized by long-term memory processes, like fractional Gaussian
noise. On the long run (horizon more than four years), markets are charac-
terized by deterministic chaotic dynamics. Once again, these affirmations
seem inconsistent, particularly concerning the long memory characteristic
on the ... short run! Moreover, we think that these interpretations are not
valid because of the nature (stochastic and deterministic) of the two kinds
of processes.

3.3 DETERMINISM AND STOCHASTICITY

Trajectories generated by chaotic processes appear completely random

by standard linear time series methods. However, chaotic processes are
truly deterministic. Fluctuations are thus endogenous, they are not the
result of exogenous shocks. They are generated by the system itself. On
the contrary, long-term memory processes, like ARFIMA processes, are
stochastic: after an exogenous shock, series will diverge during a more or
less long period from its previous time path. For chaotic processes there
is no need to introduce exogenous shock in order to produce divergent
trajectories.
This dichotomy between stochastic world and deterministic world seems
 199

to be crystallized in the relation D = 2 - H. Effectively, recall that the

dimension D obtained by this relation is the dimension of a stochastic
process (e.g., Peitgen, Jurgens et Saupe, 1992, p. 422). As a consequence,
if the object under study is the detection of chaos, this relation cannot be
used in order to calculate the fractal dimension of the underlying process.
In other words:
• Either the underlying process is stochastic: in this case, a link can
be established between fractional differencing parameter d and fractal
dimension D. Of course, on these conditions, D is not the dimension of
a chaotic process since the latter is necessarily deterministic .
• Or the underlying process is deterministic: in this case, the link between
long-term memory and fractal dimension is impossible to be established by
the relation D = 2 - H or D = ~ - d.
Globally, there is no valid relation between long-term memory and
chaotic processes. They constitute however two interesting different alter-
natives for modeling time series.

References
[1] ABRAHAM-FROIS G. and BERREBI E. (1995), Instabilit€, cycles,
chaos, Economica.
[2] BROCK W.A. (1986), "Distinguishing Random and Deterministic
Systems: Abridged Version" , Journal of Economic Theory,Vol. 40, pp.
168-195.
[3] BROCKWELL P.J. and DAVIS R.A. (1991), Time series: Theory and
methods, Springer Verlag.
[4] ECKMANN J-P. and RUELLE D. (1985), "Ergodic Theory of Chaos
and Strange Attractors", Reviews of Modern Physics, Vol. 57(3), pp.
617-656.
[5] GEWEKE J. and PORTER-HUDAK S. (1983), "The Estimation and
Application of Long Memory Time Series Models", Journal of Time
Series Analysis, Vol. 4(4), pp.221-238.
[6] GRANGER C.W.J. and JOYEUX R. (1980), "An Introduction To
Long-Memory Time Series Models and Fractional Differencing", Jour-
nal of Time Series Analysis,Vol.1(1),pp.15-29.
[7] GREBOGI C., OTT E., PELIKAN S. and YORKE J.A. (1984),
"Strange Attractors that are not Chaotic", Physica, 13D, pp. 261-268.
[8] HOSKING J.R.M. (1981), "Fractional Differencing", Biometrika, Vol.
68(1).
200

[9] HURST H.E. (1951), "Long-Term Storage Capacity of Reservoirs",

Transactions of the American Society of Civil Engineers, Vol. 116, pp.
770-799.
[10] LARDIC S. (1997), "L'hysterese en economie: definition et mesure" ,
These, UniversiteParisX-Nanterre.
[11] LARDIC S. and MIGNON V. (1996a), "Les tests de memoire longue
appartiennent-ils au camp du demon 1", Revue Economique, vol. 47(3),
pp. 531-540.
[12] LARDIC S. and MIGNON V. (1996b), ''Vingt ans de tests de memoire
longue au travers des processus ARFIMA" , Proceedings of the Applied
Econometrics Association Symposium, January, Paris.
[13] LARDIC S. and MIGNON V. (1996c), "Essai de mesure du degre de
la memoire longue des series. L'exemple de la modelisation ARFIMA" ,
UniversiteParis X - Nanterre.
[14] LO A.W. (1991), "Long-Term Memory in Stock Market Prices",
Econometrica, Vol.59(5), pp.1279-1313.
[15] MANDELBROT B.B. (1972), "Statistical Methodology for Nonperi-
odic Cycles: From the Covariance to R/S Analysis", Annals of Eco-
nomic and Social Measurement, Vol.1(3),pp.259-290.
[16] MANDELBROT B.B. and VAN NESS J.W. (1968), "Fractional Brow-
nian Motions, Fractional Noises and Applications" , SIAM Review, Vol.
10(4), pp.422-437.
[17] MANDELBROT B.B. and WALLIS J.R. (1968), "Noah, Joseph and
Operational Hydrology", Water Resources Research, 4(5), pp. 909-
918.
[18] MANDELBROT B.B. and WALLIS J.R. (1969a), "Computer Exper-
iments with Fractional Gaussian Noises", Water Resources Research,
5, pp. 228-267.
[19] MANDELBROT B.B. and WALLIS J.R. (1969b), "Some Long Run
Properties of Geophysical Records", Water Resources Research, 5(2),
pp. 321-340.
[20] MANDELBROT B.B. and WALLIS J.R. (1969c), "Robustness of the
Rescaled Range R/S in the Measurement of Noncyclic Long Run
Statistical Dependence", Water Resources Research, 5 (5), pp.967-988.
[21] MIGNON V. (1996), "Les implications de la memoire longue et de
la non linearite sur l'efficience du marche des changes", Journal de la
Societe Statistique de Paris, tome 137 (1), pp.51-72.
 201

[22] PEITGEN R.O., JURGENS R. and SAUPE D. (1992), Fractals for

the Classroom, Springer Verlag.
[23] PETERS E.E. (1991), Chaos and Order in the Capital Markets, John
Wiley & Sons.
[24] PETERS E.E. (1994), Fractal Market Analysis, John Wiley & Sons.
[25] RUELLE D. and TAKENS F. (1971), "On the Nature of Thrbulence" ,
Communications in Mathematical Physics, 20, pp. 167-192.
[26] SOWELL F. (1992a), "Maximum Likelihood Estimation of Stationary
Univariate Fractionally Integrated Time Series Models", Journal of
Econometrics, Vol. 53, pp. 165-188.
[27] SOWELL F. (1992b), "Modeling Long-Run Behavior with the Frac-
tional ARIMA Model", Journal of Monetary Economics, Vol. 29, pp.
277-302.
Contributors

ABRAHAM-FROIS Gilbert, Professor, University of Parix X-Nanterre -

UFR SEGMI - MODEM - 200 Avenue de la Republique - 92001 NAN-
TERRE Cedex, FRANCE.

ARENA Richard, Professor, University of Nice-Sophia Antipolis, Latapses/

CNRS, 250 Rue A. Einstein - 06560 VALBONNE, FRANCE.

BERREBI Edmond, Professor, University of Paris X-Nanterre - UFR

SEGMI- MODEM - 200, Avenue de la Republique - 92001 NANTERRE
Cedex, FRANCE.

BOTOMAZAVA Michel, SPPM (Macroeconomic Planning Department),

Ministry of Economy and Finance, 101 ANTANANARIVO, MADAGAS-
CAR.

DAY H. Richard, Professor, University of Southern California, Department

of Economics, 3620 S. Vermont Avenue KAP300 - Los Angeles, CA 90089-
0253 USA.

DUFRENOT Gilles, Maitre de Conferences, University of Paris XII-Val-

de-Marne, ERUDITE, 61, Avenue du General-de-Gaulle- 94010 CRETEIL
Cedex, FRANCE.

FRANKE Reiner, Professor, University of Bremen, Department of Eco-

nomics, 28334 BREMEN - GERMANY

LARDIC Sendrine, Maitre de Conferences, University of Paris X-Nanter-

re, UFR SEGMI - MODEM, 200, Avenue de la Republique - 92001
NANTERRE Cedex, FRANCE.

MATHIEU Laurent, Maitre de Conferences, University of Paris X-Nanter-

re, UFR SEGMI - MODEM, 200, Avenue de la Republique - 92001
NANTERRE Cedex, FRANCE.

MEDIO Alfredo, Professor, University Ca'Foscari - Dipartimento di Sci-

enze Economiche - Dorsoduro 3246 - 30123 VENEZIO - ITALIE.
204

MIGNON Valerie, Docteur-ATER, University of Paris X-Nanterre - UFR

SEGMI-MODEM - 200, Avenue de la Repuhlique - 92001 NANTERRE
Cedex, FRANCE.

RAYBAUT Alain, Charge de Recherches au CNRS, University of Nice-

Sophia Antipolis - Latapses/CNRS - 250 Rue A. Einstein - 06560 VAL-
BONNE, FRANCE.

TOUZE Vincent, Ph. D. Candidate, CREST, THEMA and University of

Paris X, Timbre J360, 15 boulevard Gabriel Peri, 92245 MALAKOFF,
FRANCE.

WANG Zhigang, Ph. D., University of Southern California, Department of

Economics, 3620 S. Vermont Avenue, KAP 300, Los Angeles CA 90089-
0253 USA.

ZHANG Min, Ph. D. Candidate, University of California, Department of

Economics, 3620 S. Vermont Avenue, KAP 300, Los Angeles, CA 90089-
0253 USA.
Vol. 361: C. F. Daganzo, Logistics Systems Analysis. X, Vol. 386: M. Desrochers, J.-M. Rousseau (Eds.), Compu-
321 pages. 1991. ter-Aided Transit Scheduling. Proceedings, 1990. XIII,432
Vol. 362: F. Gehrels, Essays in Macroeconomics of an pages. 1992.
Open Economy. VII, 183 pages. 1991. Vol. 387: W. Gaertner, M. Klemisch-Ahlert, Social Choice
Vol. 363: C. Puppe, Distorted Probabilities and Choice under and Bargaining Perspectives on Distributive Justice. VIII,
Risk. VIII, 100 pages. 1991 131 pages. 1992.

Vol. 364: B. Horvath, Are Policy Variables Exogenous? XII, Vol. 388: D. Bartmann, M. J. Beckmann, Inventory Control.
162 pages. 1991. XV. 252 pages. 1992.

Vol. 365: G. A. Heuer, U. Leopold-Wildburger. Balanced Vol. 389: B. Dutta. D. Mookherjee. T. Parthasarathy. T.
Silverman Games on General Discrete Sets. V, 140 pages. Raghavan. D. Ray, S. Tijs (Eds.). Game Theory and
1991. Economic Applications. Proceedings. 1990. IX, 454 pages.
1992.
Vol. 366: 1. Gruber (Ed.), Econometric Decision Models.
Proceedings, 1989. VIII, 636 pages. 1991. Vol. 390: G. Sorger, Minimum Impatience Theorem for
Recursive Economic Models. X, 162 pages. 1992.
Vol. 367: M. Grauer, D. B. Pressmar (Eds.), Parallel
Computing and Mathematical Optimization. Proceedings. V, Vol. 391: C. Keser, Experimental Duopoly Markets with
208 pages. 1991. Demand Inertia. X. 150 pages. 1992.

Vol. 368: M. Fedrizzi, J. Kacprzyk, M. Roubens (Eds.), Vol. 392: K. Frauendorfer. Stochastic Two-Stage
Interactive Fuzzy Optimization. VII, 216 pages. 1991. Programming. VIII. 228 pages. 1992.

Vo1.369: R. Koblo, The Visible Hand. VIII, 131 pages. 1991. Vol. 393: B. Lucke. Price Stabilization on World Agricultural
Markets. XI, 274 pages. 1992.
Vol. 370: M. J. Beckmann, M. N. Gopalan, R. Subramanian
(Eds.), Stochastic Processes and their Applications. Vol. 394: Y.-J. Lai, c.-L. Hwang. Fuzzy Mathematical
Proceedings, 1990. XLI, 292 pages. 1991. Programming. XIII. 30 I pages. 1992.

Vol. 371: A. Schmutzler, Flexibility and Adjustment to In- Vol. 395: G. Haag. U. Mueller, K. G. Troitzsch (Eds.).
formation in Sequential Decision Problems. VIII, 198 pages. Economic Evolution and Demographic Change. XVI. 409
1991. pages. 1992.

Vol. 372: J. Esteban, The Social Viability of Money. X, 202 Vol. 396: R. V. V. Vidal (Ed.). Applied Simulated Annealing.
pages. 1991. VIII. 358 pages. 1992.

Vol. 373: A. BilIot, Economic Theory of Fuzzy Equilibria. Vol. 397: J. Wessels. A. P. Wierzbicki (Eds.), User-Oriented
XIII, 164 pages. 1992. Methodology and Techniques of Decision Analysis and Sup-
port. Proceedings. 1991. XII. 295 pages. 1993.
Vol. 374: G. Pflug, U. Dieter (Eds.), Simulation and Optimi-
zation. Proceedings, 1990. X, 162 pages. 1992. Vol. 398: J.-P. Urbain, Exogeneity in Error Correction Mo-
dels. XI. 189 pages. 1993.
Vol. 375: S.-J. Chen, Ch.-L. Hwang, Fuzzy Multiple Attri-
bute Decision Making. XII, 536 pages. 1992. Vol. 399: F. Gori, L. Geronazzo. M. Galeotti (Eds.), Non-
linear Dynamics in Economics and Social Sciences.
Vol. 376: K.-H. Jockel, G. Rothe, W. Sendler (Eds.), Proceedings. 1991. VIII. 367 pages. 1993.
Bootstrapping and Related Techniques. Proceedings, 1990.
VIII, 247 pages. 1992. Vol. 400: H. Tanizaki. Nonlinear Filters. XII, 203 pages.
1993.
Vol. 377: A. Villar, Operator Theorems with Applications
to Distributive Problems and Equilibrium Models. XVI, 160 Vol. 401: K. Mosler, M. Scarsini. Stochastic Orders and
pages. 1992. Applications. V. 379 pages. 1993.

Vol. 378: W. Krabs, J. Zowe (Eds.), Modern Methods of Vol. 402: A. van den Elzen. Adjustment Processes for Ex-
Optimization. Proceedings, 1990. VIII, 348 pages. 1992. change Economies and Noncooperative Games. VII. 146
pages. 1993.
Vol. 379: K. Marti (Ed.), Stochastic Optimization.
Proceedings, 1990. VII, 182 pages. 1992. Vol. 403: G. Brennscheidt, Predictive Behavior. VI, 227
pages. 1993.
Vol. 380: J. Odelstad, Invariance and Structural Dependence.
XII, 245 pages. 1992. Vol. 404: Y.-J. Lai. Ch.-L. Hwang, Fuzzy Multiple Objective
Decision Making. XIV, 475 pages. 1994.
Vol. 381: C. Giannini, Topics in Structural VAR
Econometrics. XI, 131 pages. 1992. Vol. 405: S. Koml6si. T. Rapcsak. S. Schaible (Eds.).
Generalized Convexity. Proceedings. 1992. VIII, 404 pages.
Vol. 382: W. Oettli, D. Pallaschke (Eds.), Advances in 1994.
Optimization. Proceedings, 1991. X, 527 pages. 1992.
Vol. 406: N. M. Hung. N. V. Quyen, Dynamic Timing
Vol. 383: J. Vartiainen, Capital Accumulation in a Decisions Under Uncertainty. X. 194 pages. 1994.
Corporatist Economy. VII, 177 pages. 1992.
Vol. 407: M. Ooms. Empirical Vector Autoregressive
Vol. 384: A. Martina, Lectures on the Economic Theory of Modeling. XIII, 380 pages. 1994.
Taxation. XII, 313 pages. 1992.
Vol. 408: K. Haase. Lotsizing and Scheduling for Production
Vol. 385: J. Gardeazabal, M. Regulez, The Monetary Model Planning. VIII. I 18 pages. 1994.
of Exchange Rates and Cointegration. X, 194 pages. 1992.
Vol. 409: A. Sprecher, Resource-Constrained Project
Scheduling. XII. 142 pages. 1994.
Vol. 410: R. Winkelmann, Count Data Models. XI, 213 Vol. 437: C. Jordan, Batching and Scheduling. IX, 178 pages.
pages. 1994. 1996.
Vol. 411: S. Dauzere-Peres, J.-B. Lasserre, An Integrated Vol. 438: A. Villar, General Equilibrium with Increasing
Approach in Production Planning and Scheduling. XVI, 137 Returns. XIII, 164 pages. 1996.
pages. 1994. Vol. 439: M. Zenner, Learning to Become Rational. VII,
Vol. 412: B. Kuon, Two-Person Bargaining Experiments 201 pages. 1996.
with Incomplete Information. IX, 293 pages. 1994. Vol. 440: W. Ryll, Litigation and Settlement in a Game with
Vol. 413: R. Fiorito (Ed.), Inventory, Business Cycles and Incomplete Information. VIII, 174 pages. 1996.
Monetary Transmission. VI, 287 pages. 1994. Vol. 441: H. Dawid, Adaptive Learning by Genetic
Vol. 414: Y. Crama, A. Oerlemans, F. Spieksma, Production Algorithms. IX, 166 pages.1996.
Planning in Automated Manufacturing. X, 210 pages. 1994. Vol. 442: L. Corch6n, Theories of Imperfectly Competitive
Vol. 415: P. C. Nicola, Imperfect General Equilibrium. XI, Markets. XIII, 163 pages. 1996.
167 pages. 1994. Vol. 443: G. Lang, On Overlapping Generations Models with
Vol. 416: R S. J. Cesar, Control and Game Models of the Productive Capital. X, 98 pages. 1996.
Greenhouse Effect. XI, 225 pages. 1994. Vol. 444: S. J~rgensen, G. Zaccour (Eds.), Dynamic
Vol. 417: B. Ran, D. E. Boyce, Dynamic Urban Transpor- Competitive Analysis in Marketing. X, 285 pages. 1996.
tation Network Models. XV, 391 pages. 1994. Vol. 445: A. H. Christer, S. Osaki, L. C. Thomas (Eds.),
Vol. 418: P. Bogetoft, Non-Cooperative Planning Theory. Stochastic Modelling in Innovative Manufactoring. X, 361
XI, 309 pages. 1994. pages. 1997.
Vol. 419: T. Maruyama, W. Takahashi (Eds.), Nonlinearand Vol. 446: G. Dhaene, Encompassing. X, 160 pages. 1997.
Convex Analysis in Economic Theory. VIII, 306 pages. 1995. Vol. 447: A. Artale, Rings in Auctions. X, 172 pages. 1997.
Vol. 420: M. Peeters, Time-To-Build. Interrelated Invest- Vol. 448: G. Fandel, T. Gal (Eds.), MUltiple Criteria Decision
ment and Labour Demand Modelling. With Applications to Making. XII, 678 pages. 1997.
Six OECD Countries. IX, 204 pages. 1995.
Vol. 449: F. Fang, M. Sanglier (Eds.), Complexity and Self-
Vol. 421: C. Dang, Triangulations and Simplicial Methods. Organization in Social and Economic Systems. IX, 317
IX, 196 pages. 1995. pages, 1997.
Vol. 422: D. S. Bridges, G. B. Mehta, Representations of Vol. 450: P. M. Pardalos, D. W. Hearn, W. W. Hager, (Eds.),
Preference Orderings. X, 165 pages. 1995. Network Optimization. VIII, 485 pages, 1997.
Vol. 423: K. Marti, P. Kall (Eds.), Stochastic Programming. Vol. 451: M. Salge, Rational Bubbles. Theoretical Basis,
Numerical Techniques and Engineering Applications. VIII, Economic Relevance, and Empirical Evidence with a Special
351 pages. 1995. Emphasis on the German Stock Market.lX, 265 pages. 1997.
Vol. 424: G. A. Heuer, U. Leopold-Wildburger, Silverman's Vol. 452: P. Gritzmann, R. Horst, E. Sachs, R. Tichatschke
Game. X, 283 pages. 1995. (Eds.), Recent Advances in Optimization. VIII, 379 pages.
Vol. 425: J. Kohlas, P.-A. Monney, A Mathematical Theory 1997.
of Hints. XIII, 419 pages, 1995. Vol. 453: A. S. Tangian, J. Gruber (Eds.), Constructing
Vol. 426: B. Finkenstadt, Nonlinear Dynamics in Eco- Scalar-Valued Objective Functions. VIII, 298 pages. 1997.
nomics. IX, 156 pages. 1995. Vol. 454: R-M. Krolzig, Markov-Switching Vector Auto-
Vol. 427: F. W. van Tongeren, Microsimulation Modelling regressions. XIV, 358 pages. 1997.
of the Corporate Firm. XVII, 275 pages. 1995. Vol. 455: R. Caballero, F. Ruiz, R. E. Steuer (Eds.), Advances
Vol. 428: A. A. Powell, Ch. W. Murphy, Inside a Modern in Multiple Objective and Goal Programming. VIII, 391
Macroeconometric Model. XVIII, 424 pages. 1995. pages. 1997.
Vol. 429: R. Durier, C. Michelot, Recent Developments in Vol. 456: R. Conte, R. Hegselmann, P. Terna (Eds.), Simu-
Optimization. VIII, 356 pages. 1995. lating Social Phenomena. VIII, 536 pages. 1997.
Vol. 430: J. R. Daduna,l. Branco, J. M. Pinto Paixiio (Eds.), Vol. 457: C. Hsu, Volume and the Nonlinear Dynamics of
Computer-Aided Transit Scheduling. XIV, 374 pages. 1995. Stock Returns. VIII, 133 pages. 1998.
Vol. 431: A. Aulin, Causal and Stochastic Elements in Busi- Vol. 458: K. Marti, P. Kall (Eds.), Stochastic Programming
ness Cycles. XI, 116 pages. 1996. Methods and Technical Applications. X, 437 pages. 1998.
Vol. 432: M. Tamiz (Ed.), Multi-Objective Programming Vol. 459: H. K. Ryu, D. J. Slottje, Measuring Trends in U.S.
and Goal Programming. VI, 359 pages. 1996. Income Inequality. XI, 195 pages. 1998.
Vol. 433: J. Menon, Exchange Rates and Prices. XIV, 313 Vol. 460: B. Fleischmann, J. A. E. E. van Nunen, M. G.
pages. 1996. Speranza, P. Stahly, Advances in Distribution Logistic. XI,
Vol. 434: M. W. J. Blok, Dynamic Models of the Firm. VII, 535 pages. 1998.
193 pages. 1996. Vol. 461: U. Schmidt, Axiomatic Utility Theory under Risk.
Vol. 435: L. Chen, Interest Rate Dynamics, Derivatives XV, 201 pages. 1998.
Pricing, and Risk Management. XII, 149 pages. 1996. Vol. 462: L. von Auer, Dynamic Preferences, Choise
Mechanisms, and Welfare. XII, 226 pages. 1998.
Vol. 436: M. Klemisch-Ahlert, Bargaining in Economic and
Ethical Environments. IX, 155 pages. 1996. Vol. 463: G. Abraham-Frois (Ed.), Non-Linear Dynamics
and Endogenous Cycles. VI, 204 pages. 1998.