THE DEVELOPING B O D Y AND M I N D

Series Editor:
Professor George Butterworth, Department of Psychology, University of Sussex

Designed for a broad readership in the English-speaking world, this major series represents the
best of contemporary research and theory in the cognitive, social, abnormal and biological areas
of development.

Developmental Psychology in the Soviet Union Jaan Valsiner

Language and Representation: A socio-naturalistic approach to human development Chris
Sinha
Causes of Development: Interdisciplinary perspectives George Butterworth and Peter Bryant
(eds)
A n Introduction to the Psychology of Children's Drawings Glyn V. Thomas and Ang2le
M. J. Silk
Child Prodigies and Exceptional Early Achievers John Radford
The Child's Point o f View, second edition Maureen V. Cox
Making Sense of Place: Children's understanding o f large-scale environments M. H.
Matthews
Metalinguistic Development Jean mile Gombert
Ethology and Human Development John Archer
Context and Cognition ~ a u Light
l and George Butterworth (eds)
Foundations of the Mind: Children's understanding of reality Eugene V. Subbotsky
The Child as Psychologist: A n introduction to the development of social cognition Mark
Bennett (ed.)
Developing through Relationships: Origins of communication, self and culture Alan Fogel
Infant Crying, Feeding and Sleeping: Development, problems and treatments Ian St.
James-Roberts, Gillian Harris and David Messer (eds)
Seeing, Reaching, Touching: The relations between vision and touch in infancy Arlette Streri
The Birth of Personality: The development of independent and moral behaviour in preschool
children Eugene V . Subbotsky
Dynamic Systems of Development: Change between complexity and chaos Paul van Geert
 Dynamic Systems of
Development
Change between Complexity and Chaos

I
t
Paul van Geert

HARVESTER
WHEATSHEAF
 First published 1994 by
Harvester Wheatsheaf
C a m ~ u 400.
s Mavlands Avenue
- ~ e m ~he m b s t e i d
Hertfordshire, HP2 7EZ
A division of
Simon & Schuster International Group
O Paul van Geert 1994
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form, or by any
7 means, electronic, mechanical, photocopying, recording or otherwise,
without prior permission, in writing, from the publisher.
Typeset in 10112 pt Ehrhardt
by Vision Typesetting, Manchester
Printed and bound in Great Britain by
T J Press (Padstow) Ltd

British Library Cataloguing in Publication Data

A catalogue record for this book is available from

the British Library

ISBN 0-7450-1 136-5

 Contents

Prologue: Models, concrete and wooden shoes vii

1 The forms of change

Thinking about change and development
Iterates in space
Notes

2 Developmental paths and systems

The paths of change
Cyclical transitory paths
States and transitions
The many faces of dynamic systems
T h e topological forms of dynamic systems
Developmental theories and dynamic systems

3 Observing and measuring dynamic processes

Errors, mistakes and failures of observation
Investigation as communication
T h e science of change and difference
Model and reality
Notes *

I
i
I 4 The dynamic growth model
T h e growth metaphor
T h e basic growth model
Properties of the logistic growth equation
Properties of growth in mental and behavioural development
Modelling real growth
Notes
vi Lontents
5 Transitions a n d developmental fractals
Transitions and sudden changes
Fractals and self-similarity in development
Fractal development and the quantity of transitions
A mathematical model of transitions
Transitions as catastrophes
Notes

6 Oscillations
Oscillations in development
The well model
From credit cards to predators
Notes

7 Connected growers
Development and uncertainty
Some basic mathematics of connected growth and development
T h e unexpected complexity of simple reciprocal dynamics
Notes

8 Building a model: a case history

T h e first steps
Foundations for model building
T h e characteristic data pattern and the dynamic building block
Building the model
Notes

9 T h e dynamics of m u t u a l interaction
Development as a social process
T h e unexpected complexity of a simple mutualist dynamics
Social interaction and the management of the carrying capacity
Vygotsky's dynamic systems

Epilogue: Towards a n experimental theoretical developmental

psychology -
Unsolved problems abound
Towards an experimental theoretical developmental psychology
References

Index
Prologue:
Models, concrete and wooden
shoes

In 1918 my grandfather returned from World War I. He had served in the trenches,
dug deep into the soil of Flanders' fields, where the poppies grew. Under the blasts of
bombs and shellfire, he had lost his hearing, a condition he put to use later when he had
to deal with unsatisfied clients or with my grandmother. He started a small contractor
firm and he became well known for his technical creativity as well as for his great sense
of the absurd. I n the years between the wars, it was often difficult to get the right
building materials, and sometimes his clients did not have enough money to pay for
them. He would then almost invariably come up with the solution to use concrete,
instead of the hardwood, marble or grey stone that was required. Concrete was cheap -
and my grandfather made concrete out of virtually anything - and could be poured into
almost any form and given almost any finish, provided the right sort of technique was
used.
His biggest feat - at least one for which he will be remembered forever - came
during World War 11. Workmen wore wooden shoes in those days. They were cheap,
comfortable and warm. Due to the reduction in supplies caused by the war, the
limewood out of which the wooden shoes were cut became increasingly difficult to
obtain. My grandfather got the brilliant idea to use concrete to make wooden shoes. He
experimented with moulds and different sorts of concrete. Meanwhile, about half the
neighbourhood, under the guidance no doubt of my grandmother, tried to persuade
him that his attempts were futile. Wooden shoes were light, and concrete was heavy;
wooden shoes were warm, and concrete was cold; wooden shoes protected the feet from
getting wet and concrete would suck up water like a thirsty camel. None of these
arguments could change my grandfather's mind. He told them that wooden shoes
made out of wood are great, if there is any. But concrete wooden shoes, he said, are
better than no wooden shoes at all. He continued to experiment until the war was over
and wooden shoes became available again. In the neighbourhood my grandfather was
known as Leon Concrete, which in Flemish sounded more poetically, Leon Beton.
What did my grandfather teach me about scientific work? Nothing in particular, I
am afraid, but I attribute a particular style of work, model building, to his influence.
Scientists are used to make a distinction between the world out there on the one hand,
vii
 ...
VI~I Prologue
with facts and phenomena to be known and explained and on the other hand, our
theories and models. The question to be asked then is: T o what extent do the theories
and models present a recognizable true picture of the facts and phenomena in the
world? The criteria to which the picture is subjected typically involve notions such as
isomorphy, similarity, correspondence or likeness. The building metaphor pictures a
different approach. Facts and models are not independent of one another; they grow
along with the process of model construction. A model is a building, it serves a variety
of functions. It protects you against the storms of criticism, it serves as a haven and
retreat from the precipitation of unordered knowledge. The question whether a
building process has been successful or not is decided on criteria of stability, comfort
and aesthetic pleasure. Does your building collapse if left alone, can it withstand an
earthquake? How easily can it accommodate visitors, bringing in new facts and
unexpected viewpoints? I admit that all this is rather vague, but I never promised I
would go beyond the confines of the metaphorical in this prologue.
This book is about a view on development and a technique of model building that
only recently began to attract the attention of developmental psychologists. That view
and technique are associated with the notion of dynamic systems, of non-linearity,
self-organization and of complexity and chaos. All these terms come from mathemat-
ical and physical disciplines where they have been defined and studied in rigorous
ways. When applied metaphorically to the psychological domain they seem to obtain
an almost magical connotation. They open vast new horizons to the imaginative: chaos
in personality, self-organization in development, bifurcations in behaviour, strange
attractors in thinking. I am afraid, however, that readers of this book will discover that
I am only the grandson of Leon Concrete, a man who was never scared of using his
imagination, as long as he could pour it in a concrete form.
I n this book I have emphasized the constructive or building aspects. I have not really
attempted to show that the dynamic portrait is a picture that resembles the face of
reality better than all those other portraits that have been painted before. The problem
is that we know surprisingly little about that face of reality as far as it relates to the
dynamics of development. Certainly, an incredible number of facts about children and
age have been collected over the years, but it is remarkable how little these facts tell
about dynamics and change.
In this book I construct the rough outlines of a building that is meant to
accommodate more children, guests and friends than are currently around. If they do
not show up in the future, the place will remain depressingly empty. That is to say, I
develop a model that is currently only very moderately supported by available facts and
data. A lot more research is needed of a rather different kind than is generally available.
The constructive or building aspect also shows in a second way. Throughout the
book I describe the process of model construction explicitly. Especially in the second
part of the book, starting with Chapter 4, the chapters contain a lot of 'tutorial' boxes
(models) explaining in as much detail as possible how simple dynamic models can be
built and tested. Instead of presenting a model of development, I present a way of
building models, models of a particular kind, for sure. In accordance with my own view
on the function of building however, I believe, or at least I hope, that presenting a
 Prologue ix
construction procedure will at the same time present a new sort of model and approach
to development.
There is a third sense in which the author of these pages sees himself as the grandson
of Leon Concrete. My grandfather used concrete just about everywhere in his
buildings, where others would have used wood, marble or steel. By doing this, he could
build for almost everybody, where others had to wait and sometimes wait very long, for
the few clients who could afford the expensive materials. My concrete is the
spreadsheet and the type of models that are easily built in the form of spreadsheets.
Spreadsheets are relatively cheap computer programs that will do computations over
columns and rows of cells. Almost everybody has one and in most cases peopIe use
them to calculate their income taxes, or to enter data from an experiment and draw a
diagram. It is amazing, though, how versatile they are, and how much they can teach
you about unexpected aspects of developmental models. No doubt .a variety of
techniques and computer programs exist that are better suited for particular
model-building tasks. For those who have the time, the money and the talent, the
alternative and better-suited models are certainly worthwhile. But my grandfather
would have heartily agreed that for those who are used to carrying the burden on their
shoulders, the use of a wheelbarrow is already a major improvement. On the other
hand, every now and then everyone is likely to run into a problem that even my
grandfather could not solve, namely how to make wooden shoes out of concrete.
The book contains nine major chapters, divided in two parts. Chapters 1 to 3 explain
the notion of dynamic systems as applied to developmental processes. Chapter 1
discusses the concept of development in general and some interesting problems and
insights that are rarely taken into account in developmental theories. I t introduces the
notion of an iterative process, which is central to the dynamic system approach
described in this book. T h e chapter also discusses the notion of an abstract space of
properties and defines development as a particular trajectory through such an abstract
space. Chapter 2 presents various types of trajectories or developmental paths. Its
starting point lies in a generalized notion of a dynamic system and prototypical paths of
change that such systems follow. I t deals with the different ways in which a dynamic
system, including developmental theories, can be represented. It shows how radically
those representations differ from the common view on what a developmental theory or
model describes and how. Chapter 3 presents a general approach to problems of
measurement and empirical methods of investigation. It focuses on the problem of
measurement error and tries to answer some of the problems raised in Chapter 2 by
conceiving psychological measurement as a dynamic process in itself. The chapter is
not aiming at describing and explaining specific statistical methods for checking
dynamic models against available data. The discussion is rather philosophical and I am
afraid not very practical and probably disappointing to those readers who expected to
find something about research methods and statistical techniques.
The second part of the book discusses the dynamic models and the procedures for
building them. Chapter 4 introduces the fundamental notion of growth as a dynamic
process and applies this general model to psychological development. Chapter 5
discusses a second type of change, namely sudden qualitative transitions, and presents
 x Prologue
a model that puts the gradual and transitional change under a general framework of
change and growth. It addresses, among other topics, the eventual fractal nature of
development. Chapter 6 deals with a third type of change, namely one that amounts to
oscillations or cycles. Development is usually seen as a trajectory towards a higher
stable level. Many processes of change, however, amount to cycles visiting similar
sequences of states over and over again. Chapter 7 explains the ecology of development
in terms of connections among 'growers'. Connection patterns account for qualitative-
ly different macroscopic forms of development. The chapter also discusses basic
connections in simple systems consisting of only two growers, such as two positively
interacting skills. Chapter 8 presents a case history of actual model building. It
describes the attempt of Kurt Fischer and myself to build a dynamic growth model of
Fischer's developmental model and data. I have chosen a case history instead of a
systematic account of model-building techniques, in the hope that a case history
manages to give a better idea of how model building actually works. Chapter 9 deals
with social developmental dynamics and invokes the various relationships between a
'tutor' and a 'tutee' involved in a process of co-constructing the tutee's knowledge and
skills.
The tutorials present a detailed account of model-building procedures. Readers who
are not familiar with spreadsheets should first become acquainted with the elementary
spreadsheet techniques. The fastest and easiest way is to ask a more experienced user to
assist in building some of the simple models. I have chosen to explain the models with
one particular spreadsheet program, Lotus 1-2-3 for Windows. Lotus 1-2-3 is
probably the most widespread program of this kind. It is also available for Apple
Macintosh computers. The non-windows and windows versions are highly similar.
Other spreadsheet programs, such as Quattro Pro and Excel differ only in the way the
menus and instructions work. They are similar, however, as far as the form of the
equations is concerned. With a little help from the manuals, it should not be too
difficult to translate the Lotus 1-2-3 into Quattro Pro and Excel procedures. Recently,
Lotus has released a spreadsheet with a radically different setup, called Lotus Improv.
Improv users who have never used the classic version will have to invest more
translation efforts, given the form of the equations and models in the book. During the
, year we worked on Kurt Fischer's model, Kurt changed from Lotus 1-2-3 to Improv.
Improv proved a real improvement as far as the transparency of the equations was
concerned. I t was considerably less time consuming than 1-2-3 to write down the
equations and models, but proved to be rather cumbersome and inflexible in its
graphics. Newer releases will no doubt improve that. Readers who are used to a
programming language such as BASIC or C can easily build the models in their
favourite language. The spreadsheets compute iterative loops of simple equations and
do so for a fixed number of repetitions. Loops of this kind are easily modelled in
standard programming languages. Nevertheless, I would recommend absolute novices
to study the spreadsheet program approach.
Let me also say a few words about the style of the book. Maybe it is not a very good
example of what a scientific book should look like. I introduce many of the problems
and models with little personal stories. For some unexplained reason, these stories and
 Prologue xi
memories came to my mind more or less accidentally when I was preparing the
chapters and sections. Maybe they can serve as anchor points for the many abstract
principles presented in this book. Maybe they make it more easy to swallow the rows of
equations that make so many chapters less agreeable to read. And maybe some readers
will be rather irritated by these little confessions and reminiscences of my past, but I
hope they are prepared to excuse me for such idiosyncrasies as my grandfather's
wooden shoes and my first visits to the local cafi. On the other hand, I felt that
presenting the content matter of this book in the usual rather formal manner would not
have corresponded with the status of the ideas presented here. They are still in the
process of making, often more intuitively than systematically justified. They present a
way of viewing rather than a view itself. I wanted to share the process of thinking and
model building with the reader and hoped that the somewhat impressionist style of
many chapters would enable me to better communicate that intention.
This book is the product of a sabbatical year that I spent at the Center for Advanced
Study in the Behavioral Sciences at Stanford, California, which was made possible by a
grant from the John D. and Catherine T. MacArthur Foundation, and by the support
from the Center and the University of Groningen, The Netherlands. I am grateful to
my Groningen colleagues for giving me the opportunity to take leave for a year and for
taking care of the domestic things.
The Center for Advanced Study is a great place to be. It is based on a sort of mixed
concept of an Italian monastery and a Club Mkditerranie vacation resort, and a
Fellow's only obligation is to be present at lunch. 'Paraphrasing Thorstein Veblen's
classic, some have called it 'the leisure of the theory class', which, of course, is only
based on envy and chagrin. I am grateful to the friendly and helpful staff and to my
fellow Fellows for this great year of scientific productivity and human interest. A
special word of gratitude goes to Kathleen Much for her invaluable editorial help.
A great deal of the work I present in this book grew to its actual form in the context of
a working group called Modeling and Measuring Development, initiated by my
distinguished colleague and friend Kurt Fischer from Harvard University. The group
was very small - it consisted of Kurt and myself - but we had several regular visitors:
Robbie Case, who was actively involved in the model-building endeavour, Kathryn
Nelson, Abe Tesser and Paul Harris. We had regular visitors and consultants, such as
Nira Granott, Bob Thatcher, Peter Molenaar, Han van der Maas and John Willet.
Working with all these people was a great honour and a source of joy (that is, we had a
lot of unprofessional fun together). But most important of all is that I learned so much
from them. I am especially grateful to Kurt Fischer, for the stimulating meetings we
had and for his interest, effort and friendship. Besides the working group, I had a
second opportunity to meet colleagues interested in the problems of mankind. That
group usually met after lunch. While performing some light physical exercises such as
running after balls that would inevitably get stuck under somebpdy's car, the members
discussed the most complicated problems in the most lighthearted way possible. Many
people contributed, and I mention only the most faithful visitors and scholars: Bob
Scott, Lynn Gale, Marlene Scardamalia, Carl Bereiter, Marcello Suarez-Orozco,
Susan Cotts Watkins and Susana Larosa. Although the products of our collaboration
xii Prologue
never lasted longer than the duration of the meetings, they helped me make a lot of
progress.
People from the (very informal) Groningen Dynamic Systems Group read the
manuscript and made a number of valuable editorial suggestions. I am grateful to
Martin Cats and especially to Saskia Kunnen who was very near the edge of a nervous
breakdown when she found out she could not perform a number of model-building
exercises, due to the present writer's negligence and his inability to spot typing errors
in his own equations. Very useful editorial comments also came from Alan Fogel
(University of Utah at Salt Lake City) and Kurt Fischer (Harvard University).
Finally, a word of gratitude to my family, who supported me so much in their own
particular ways: Liesbet, by choosing the arts, the dog, Nathan and the Dutch weather;
David, for a terrific year at Foothill and for his decision to keep out of the back of
pick-up trucks; and Leen, for so many things that I do not know which ones I shall
mention first.

Paul van Geert

The forms of change

at makes a cat look like a cat?

the 1920s, Piaget investigated the child's conception of the world and found that
oung children see the world in an artificialist way (Piaget, 1929). Everything has been
ade by man, on purpose, with a specific goal in mind. Artificialism, however, is not a
ere transient phenomenon typical of children, but deliberately abandoned in adult
inking. On the contrary, in a new disguise artificialism appears to permeate much of
ur adult conceptualization of the world. Of course, adults know the distinction
etween man-made objects and nature. But, if we are asked why the things of nature
e as they are, and especially why they come into being with their proper and
articular form, it is difficult to avoid a style of thinking that is highly reminiscent of
e young child's artificialist explanations. Take for instance the animal kingdom.
hat makes a cat look like a cat (or a Dutchman like a minister')? Educated adults
now about biological conception and about genes containing the information that
akes cats cats. But in what way are genes responsible for the bodily form and
ehaviour we think typical of the feline genus? This is where we have difficulty
oiding a form of artificialism, we are inclined to see genes as building plans. The
enome looks like an agenda, a sequence of instructions describing step by step what
the contractor firm Biology Inc. should do. It starts with an instruction for cell
multiplication, for instance, and it even continues during the cat's adolescence, when it
turns on the gonadal hormones to prepare,the animal for the (partially) pleasant task of
procreating new cats.
Do genes really act like a building plan or agenda? Do feline genes contain a
blueprint for a tail or whiskers and human genes one for a Habsburg chin or an I Q o f
1lo? Genes contain large amounts of information, but all this information is aimed at
the synthesis of proteins, not of large bodily structures. In a sense, genes contain a
building plan, but it relates only to the first step, the building of proteins. The
properties of these proteins lead them to cluster in specific structures. But there is no
 2 The forms of change
overall building plan that a general supervisor takes care of (Oyama, 1989). Each step in
the process of bodily growth creates conditions that specify and constrain the next step.
The idea that a complicated structure such as an organism's body can come about
without an explicit building plan and without a controlling instance external to the
building activities is difficult to accept. What steers the process? The answer that each
step constrains and specifies the next step is hardly acceptable as an explanation. And
. we find such an explanation unacceptable because of our deeply rooted artificialist
tendencies, which make us insensitive to explanations that rely on the intrinsic
ordering activities of time. Oyama (1989) calls this belief in a genetic building plan the
'central dogma' of evolutionary thinking. Developmental psychology offers many
comparable examples. In the sections that follow, I shall discuss a few examples from
other disciplines devoted to the study of change and illustrate the point that structure
over time does not require the working of a master builder or the following of a plan. In
the second part of this chapter, I first introduce a basic notion, iteration, which lies at
the heart of the dynamic approach presented in this book. I shall show how iterative
processes describe development in the form of trajectories through abstract spaces.

Structured pictures and random scribbles

Biological evolution is a prime example of a process of long-term change that has led to
a complicated and orderly structure, namely the world of biological species and
ecosystems. We may safely say that Darwin's legacy has become an intrinsic part of our
world view. There are circles in which creationism rules, but the inspiration behind
that is a religious belief, not a deep belief in artificialism per se.
In biology class, students are taught that biological evolution is not a matter of a
pre-existing building plan. Species change and turn into new species because the fittest
survive and the less fit die before they have been able to transmit their genes -
containing less favourable properties - to new generations. Unexpected random
mutations create new genetic opportunities, which can be tested against the difficulties
of survival. Evolution is a very gradual process of both pruning and forming and it
depends on purely coincidental factors; randomness rules, but order enters because the
pruning is based on fitness and selective survival.
Darwin's principal accomplishment is that he has eradicated all traces of
artificialism in our scientific thinking about biological origins. It is interesting to see,
however, that so many of Darwin's interpreters have not been able to appreciate this
point. The principle of survival of the fittest founded the belief that evolution heads
towards increasing adaptation, complexity or quality of organisms. Evolution is going
somewhere; it has a goal and endpoint. This belief is evident in widely different
interpretations, such as Teilhard de Chardin's or social Darwinism's. In the late
nineteenth century, the survival principle underscored the idea that the adult
European male was the ultimate crown on Mother Nature's head. This notion of
finalism or teleology is related to an artificialist stand: things evolve because there is a
purpose or a goal involved. Nature, on the other hand - and this is clearly stated and
repeated time and again in Origin of Species - has no intrinsic goal, it does not care and
 Thinking about change and development 3
it floats on the stream of randomness and mere coincidence. The fact that even man,
the jewel in the crown of creation, was a mere coincidence has always been difficult to
accept (Gould, 1992).
This difficulty is not only observed in our thinking about evolution. If we reflect on
our personal lives or the lives of others, we find it difficult not to see the actual outcome
as something that was there from the very start. You become a professor of psychology
because you are born with a sufficient amount of intelligence (though not enough to
become a professor in a more advanced science, it seems), because you have a natural
interest in the minds of other people, and so forth. The 'child is father to the man';
there must be potentialities in us that become actual and mature. What we finally
become cannot be a matter of mere coincidence, of mere random drift. Random
factors, we are taught in our first statistics class, will compensate for each other's
effects. Because they occur in such large numbers, random effects will actually average
out. But is this really so? A simple example will show that this conclusion is wrong, at
least as a general rule. The example is based on a simple mathematical model and it will
be the first example in a long list of models and tutorial boxes that will illustrate and I
hope support the major point of this book. (I know that Shortman Webstar English
College Dictionary defines 'tutorial boxes' as 'parts of textbooks marked by a box or
horizontal lines, that are mostly skipped and at best only superficially glanced at'; in
this book, however, those boxes are essential parts of the text, if not the most important
parts of the text and, moreover, if you skip them you will miss a lot of fun). So take your
favourite spreadsheet or programming language, or if you do not have or use one, just
follow me.

Model 1.1 Random walks and random order

The Great Marketplace of the Belgian city of Sint Niklaas is world famous, because it is
said to be the biggest marketplace in the entire universe (well to be honest, it is world
famous only in Belgium . . .). Not only is it very big, it is also surrounded by a really
impressive number of cafes and pubs. On a cold autumn day, one of the local beer
specialists, obviously in a state of moderate drunkenness, leaves a pub at the west side of
the Market and heads for one of the pubs on the east side. The gusty winds blow fiercely
over there, and they change direction almost every second. So our poor walker is now
pushed to the right, then to the left and because of the darkness and his bodily instability,
he becomes the subject of the blind forces of the autumn weather. Where will his random
path lead?
The wind acts like a truly random force, with regard both to its northwest'direction and
to its force. Since the Marketplace is so big, we may expect that the average effect
resembles a straight line. But will this really be the case?
Open a spreadsheet and type the title "Random Walks" in the upper left corner. You
can type any further information you consider necessary in the adjacent cells. Go to cell
A10 and type the following equation specifying a random number varying between 1 +
and - 1:
4 The forms of change

Figure 1.1 The drunkard's path depends on random forces (the gusty winds of Sint
Niklaas) that come from all possible sides. T h e series of random numbers simulating those
forces is represented at the bottom of the diagram; the random walk resulting from those
numbers has a definite overall structure which deviates markedly from a straight line.

(note that most spreadsheets will not allow you to type spaces in the equation; I have just
added them for the sake of clarity).
Copy this equation to the cell range A1 1..A2000 (if you are completely unfamiliar with
spreadsheets, consult your manual. Copying is usually very easy. If you use a
Windows-based version of Lotus 1-2-3, 1 suggest you use the 'Lotus-classic' way of
copying, especially if it requires copying over such large ranges as with the present model;
note that some spreadsheets use a different cell range notation, namely A1 1:A2000, I use
the Lotus and Quattro Pro notation throughout the whole book). Go to cell 610 and enter
the initial value for the random walk, 0 (zero). In cell B11 type the equation that defines the
random walk:

+ 010 + A l l (1.2)

Copy B11 to B12..B2000. This equation adds the value of the random number in the
adjoining A cell to the sum of all the preceding random values. In order to obtain a
 Thinking about change and development
smoothed representation of the random walk path, go to cell C1O and enter the equation
I
for a moving average of fifty consecutive points:
t

and copy this to the range C1 I..C1950.

To see the result of the computations, define a line graph that depicts the range
CIO..C1950 (the manual will explain how to do this if you are not familiar with this
procedure). I suggest you show two patterns in this graph, one of the random walk
(CIO..C1950) and one for the series of random numbers upon which the walk was based
(A1O..AI 950).
Figure 1.1 represents the smoothed paths of the A and C ranges, respectively. Whereas
the random numbers on average resemble the effect of a straight line, the random walk
goes in a specific direction. Of course, each time a random walk is run on the computer, a
different overall pattern will come out. But each of these patterns will have a definite
macrostructure that differs qualitatively from the smoothed random line.

What can we learn from the first model? We might have expected that the net result of
added random factors would boil down to something that wiggles around a straight
line. After all, the mean value of the numbers is always very close to zero. The line of
random numbers actually fulfils our expectation. But the essential difference between
the random line and the random walk is that the latter consists of conditionally coupled
random points. That is, each point in the random walk is the result of all the preceding
randomly driven points. In the true random series of points, each random point is
independent of all the others. Conditionally. coupled randomness leads to an overall
macrostructure, and this, dependence of each later stage on the previous stages, is basic
to development.
Assume that the random walks represent 'lifelines' of individual human beings,
represented in a two-dimensional framework. Figure 1.2 for instance contains
imaginary curves of changing school achievements of four subjects, which in reality are
random walks based on Model 1.1. Although each of them begins at about the same
point, they very soon diverge and lead to characteristic individual patterns with a
definite overall macrostructure. If they represented real school achievement curves, we
would probably tend to explain the macrostructure as the result of major qualitative
changes (e.g. a major change in direction). If one goes up and another down, we might
assume that there is some underlying property, some hidden difference in the initial
state or in the individuals' abilities or innate competences, that drives them in clearly
different directions. A change of direction, for instance, might be attributed to some
underlying change in motivation or interest. Of course, real lifelines are almost
certainly determined by differences in genetic endowment, in educational opportuni-
ties and so forth. But the point of the present illustration is that random
microvariations (random factors varying at each point in time)groduce a macrostruc-
ture if conditionally coupled, for instance in the form of a random walk. Conditional
coupling is an elementary principle of development and it is capable of producing
macrostructure out of nothing more than mere random variation.
I am certainly not suggesting that all major changes over time are just coincidental,
 The forms of change

Figure 1.2 Two curves representing imaginary scholarly achievement. Although they are
based on the drunkard's random walk, their overall structure is such that a casual observer
would be inclined to explain them on the grounds of major life-events or comparable
influences, instead of the small continuous random events.

random-driven events. Maybe none of them are and to decide upon that is a matter of
empirical research. The point is, however, that major event patterns over time do not
necessarily invoke underlying tendencies of change in this or that direction, or major
effects called into existence before a basic change in the nature or direction of a process.
Random variations, if conditionally coupled, can create a form of long-term order.
Observing such order may lead us to believe that it should be the product of some
underlying order. That is true, perhaps, in that the underlying order may be as simple
as a conditionally coupled sequence of random variations.

Smooth causes, abrupt effects

Introductory textbooks on developmental psychology usually describe developmental
stages such as Piaget's four stages of cognitive development. They paint a picture of
relative stability punctuated by sudden qualitative jumps. Many current researchers,
however, see development as more gradual than the great masters of developmental
psychology made us believe and, in a sense, quite similar to the form of biological
evolution. Biological evolution is the accumulation of many small changes and is
basically gradual and smooth. Selection against the odds of a dangerous world and
 Thinking about change and development 7
random mutations goes on with any individual organism and at any time. There are no
moments in history where survival pressures are relaxed, or where adaptation is less
important. Nevertheless, the gradualist stance has recently been challenged (e.g.
Eldredge and Gould, 1972; Eldredge and Tattersall, 1982). Evolution is not gradual at
all, but occurs in the form of rapid changes, spurts in speciation that connect relatively
long periods of species stability. This is the so-called punctuated equilibrium view of
evolution. Evolution is a stair rather than a slide. This view is highly reminiscent of the
stage theory view in psychological development.
If the viewpoint is correct - and it is highly likely to be so in many cases - then the
question is how the punctuations of an equilibrium can be explained. Surely the
Darwinian principle of selective mechanisms that act on a continuous and gradual scale
is not at stake. Better survival chances of better adapted individuals are still the motor
behind evolution. But how can a gradual mechanism that never ceases to operate
explain a long-term process of stability penetrated by sudden changes?
The difficulty is that there is a qualitative distinction between the cause, the gradual,
continuous mechanism of selective survival, and the effect, non-gradual, discontinuous
change. We seem to have conceptual problems dealing with this kind of asymmetry. It
is usually solved by postulating an external contingent factor: species remain stable as
long as the environment remains stable. But if a large comet hits the surface of the earth
and changes the climate for the next thousands of years, it probably sets the process of
rapid change in motion. As far as the vanishing of dinosaurs is concerned, it appears
almost certain that such an external force has had a major impact. But even in this case,
the comet was only a deadly stab to a group of species that was already in decline.
But the point is that sudden changes are not necessarily explained by external forces.
Some, though not many, of the punctuations may result from the internal dynamics of
the processes at issue. Again, this is something that we find difficult to reconcile with
our common notions of cause and effect. Let us proceed to Model 1.2 to see how a
continuous cause can produce discontinuous effects, and how our notions of continuity
and discontinuity depend on the time scales we actually contemplate.

Model 1.2 Time frames and explosions

There is a well-known tale of an Indian king who played chess with a servant. What I
remember is that the king lost, and he seemed so impressed by the victory of his servant,
and by the sheer talent this victory testified to, that he promised to give the servant
whatever he wanted. The servant asked for one grain of wheat on the first checker, two on
the second, four on the third and so forth. The servant was app1yin.g the so-called
doubling scenario. The process, doubling the preceding number of grains, is isochronic,
which means that it is similar over the consecutive points in the sequence (there is no point
where it turns into tripling, for instance).
Open your spreadsheet program, type whatever titles and information you need in the
upper left corner and go to cell A10 where you enter the digit 1. Go to cell A1 1 and type:
8 The forms of change
Copy the content of cell A1 1 to the range A12..A210. This will result in an iterative process,
where each consecutive cell takes the value of the preceding cell, multiplies it by two and
returns that value to its successor.
What would the curve representing the sequence of numbers look like? Since the
process is isochronic-that is, similar over each point-we may expect a steep but gradual
increase. Define five different line graphs: one for the range A10..A12 (a very small time
window indeed), and the others for the ranges A10..A35, A10..A60, AlO..A110 and
A10..A210. Look at the first graph, for A10..A12 (see Figure 1.3). The increase looks quite
linear, although there is a mild, expected increase in its speed. So it seems that we might
extrapolate this form to time windows of any arbitrary size and we would expect to find the
same smooth increase.
Compare this with the other graphs; what they show is very different from the small time
window of only three points. Qualitatively, the doubling scenario results in a long period
with, relatively speaking, no change at all, followed by a sudden explosion at the end of
the time window. There is nothing in particular that actually ignites the explosion at the
end, other than the continuous doubling process applied to each time point. The small
window suggested a smooth increase, but this did not extrapolate to larger windows, i.e.
windows beyond a given scale (in the case of the doubling scenario, the scale threshold is
as low as twenty steps). In all cases there is a sudden jump, which becomes more
pronounced as more points are taken into consideration.
We may argue that a doubling of the preceding magnitude naturally leads to an
explosive increase, but that an increase with a much smaller rate will present a gradually
increasing curve, as expected. So, instead of multiplying by 2, try 0.01. And indeed, 200
consecutive points show a smooth and calm increase. But take 2000 points, and the
seemingly qualitative jump occurs again.

Figure 1.3 A very small time window, consisting of three consecutive points, suggests a
gradual, almost linear, increase; a large time window, however, will reveal the explosive
nature of this kind of increase (see Figure 1.4).
 Thinking about change and development 9
What can we conclude from this simple model, as far as our intuitions on the causes of
sudden changes are concerned? The process we applied was iso~hronic,~ that is, it was
similar for all points along the time axis (or similar for each step in the iterative
process). If we took a small window, we observed a seemingly gradual process. But if
we watched through a sufficientlylarge window, we saw a process remaining almost the
same for a long time, and then suddenly changing its direction, almost as if it had
exploded.
We may draw two conclusions. The first is that sudden changes are not necessarily
caused by external factors appearing just before the change sets in. Of course, many
sudden changes are caused by external factors, or a deus ex machina, such as a comet or a
winning lottery ticket. But conditionally coupled events such as the doubling scenario
naturally show sudden changes just by the way they behave over time.
The second conclusion is that such conditionally coupled processes have a quality
that we call non-linearity. Non-linearity has a technical, mathematical meaning, but I
shall use and define it in an intuitive manner. It means that a linear increase in one
variable corresponds with a non-linear change in another. Take for instance Figure 1.4,
based on the doubling of grains at each checker. The number of iterations (i.e.
applications of the same doubling principle) increases with time along a straight line.
The corresponding number of grains, however, explodes at the end (at this scale of
representation, that is). If the relationship had been linear, we would have expected to
see an increase in the number of grains along another straight line.3 The doubling
scenario is a very simple example of non-linearity. Further in this book I shall give
examples of non- linearity that relate much closer to the process of development (in the

Figure 1.4 Although the number of doubling changes linearly (one at a time), the
number of grains shows an explosive increase (white arrow). This is a simple form of
non-linearity. The curve in the back is based on a non-linear growth model of a type that
will be explained later; the non-linearities are complex.
10 The forms of change
background of Figure 1.4, you will see a complicated curve, based on a simple model of
interacting growers explained in Chapter 7, foreshadowing the interesting patterns our
models will generate).
Isochrony and (non-)linearity are important aspects of developmental processes and
difficult to deal with. We are used to making linear extrapolations and to acting as if
isochrony necessarily means linearity. This habit of thought makes it very difficult to
appreciate the nature and effects of developmental processes, where each new step
builds onto the results of the previous ones.

Meno's problem
Visiting Athens on the First of May, one will find the Parthenon closed for Labour
Day. The Greeks who used that place more than two thousand years ago did not have
to worry too much about manual labour - their slaves took care of that - so they could
enjoy the works of the mind. Slaves could even be the subject of experiments in
educational psychology; Plato, for instance, tells of one in one of his Dialogues
(Hamilton and Cairns, 1963). The story describes a discussion between Socrates and
Meno (and it also features a slave boy and a fourth person called Anytus) taking place,
probably at the Parthenon, but anyway a long time before Labour Day was invented.
The issue is whether virtue can be taught, but the underlying problem is whether
something can be put into the mind if it was not already there. Socrates shows that you
can teach new knowledge. He demonstrates by teaching the slave boy a geometric
principle, in his characteristic Socratic way, namely by letting the boy find out the
principle by himself. But the new knowledge, Socrates concludes, is nothing but a
recollection of knowledge that was already there. Whatever looks like new knowledge
in the mind is nothing but a reordering of existing knowledge.
Meno's problem is still among us. For instance, is adult thinking qualitatively
different and thus structurally new, in comparison with the infant's or the child's
thinking? Or is it the result of a gradual quantitative change of a structure present at
birth?
Piaget for instance considered the age of 6 or 7 a major qualitative turning point. If you
ask a 4-year-old whether the amount of water in a jar, after having been poured from a
container with a different shape, is still the same, the child will most probably say no and
will support this answer, perhaps by referring to the fact that the water is higher than
before. Six-year-olds will say that there is the same amount, for instance because
although the water is higher, it is also 'thinner' (or whatever form of compensation they
will apply). The answers reflect fundamentally different principles of cognitive
organization. Whereas the younger child still dwells in the realm ofaction, the older child
has entered the kingdom of operations, which are mental actions that are reversible.
It remains difficult to imagine, however, how a fundamentally new mental
organization can come about. The difficulty is actually based on a 'Socratic' argument.
How could a structure create a more complicated structure if the more complicated
aspects are not already present, in some form or other? I can teach somebody the
meaning ofpoodle by saying, 'Look, that's a poodle', if a poodle happens to pass by. But
 Thinking about change and development
that person must by necessity already have the meaning of poodle, since, if that is not
the case, my saying 'that is a poodle' will have no meaning at all, except for the trivial
meaning that 'that dark hairy thing over there is a something I don't know'. This is,
approximately, Fodor's argument on the impossibility of a learning theory of meaning
development and it is very close to Socrates' one.
It is probably out of intuitive uneasiness with having to accept that something really
new can come out of something old, that many investigators have been trying to show
that Piaget's view was incorrect. T h e argument is that what he saw as qualitative shifts
can be explained by quantitative changes in some underlying variable. For instance, if
the child's working memory is too small to contain more than one chunk of
information, the child will not understand problems that need two chunks to be solved
correctly. Once the quantitative growth of working memory has widened its scope to
two chunks, the child will actually understand. This is far from a major qualitative
restructuring in the child's cognitive principles, however, although the quantitative
change might have very considerable consequences for a large domain of problems.
What are the major conceptual components or aspects of Meno's problem? T h e first
is concerned with the fact that our ideas on the nature and content of the mind are still
strongly idealistically coloured. Socrates concluded that the boy, to whom he taught a
geometric rule, could come up with the right answer only if the required knowledge
was already present and installed in his mind during a former life. We may see this as an
antique view involving reincarnation, but Socrates' conclusion is closer to modern
thought than it looks at first glance. Modern innatist views would not speak about the
boy's former life, but about his biological ancestors in whom the knowledge has
become installed and transmitted, not in the form of a memory but in the form of
genetic information. But this is basically the same as the Greek argument. Socrates'
conclusion is based on the assumption that the mind cannot create something that it
cannot entertain. Thus, in order to produce a developmentally advanced more
complicated thought or thinking structure, the mind must assemble that thought or
structure and therefore it must be capable of harbouring or containing that structure.
But this is an impossible achievement, for how can you assemble a box that is bigger
than the hall in which the assembly takes place? Consequently, all structures that the
mind can create must be present before that creation occurs, which, therefore, is not a
genuine creation but a form of retrieval.
The problem with this Socratic contention is that it reduces the mind and the
meaning of whatever is in there to the mind's own confinements and enclosures. This
viewpoint is not in itself old-fashioned or obsolete, as Fodor (1980) showed. But the
question is whether it is wise to take it as an axiom for further understanding (since it is
f an axiom, preceding the status of a contention with an empirical truth status). We
1
2 could as well opt for another axiom, that seems more appropriate, namely that the
mind's contents have meaning not in themselves, but in the interplay with their
environment. Thus, to construct a box bigger than the assembling hall, you construct
the part, which must fit into the hall, and assemble them outside. Similarly, a child may
come up with a solution to a problem that the child does not see as qualitatively distinct
from what it usually does. T h e way in which such a solution works out with the
12 T h e forms of change
problems presented in the outside world, however, may turn out to be qualitativel:
different from what the child used to do. For instance, for the child, the application o
the reversibility principle - thinking a problem backwards - may seem similar i~
nature to and only superficially different from employing the familiar forward strategy
But, the new strategy may open up a whole new world of problems and problem,
solving opportunities that mark a qualitative shift in the child's thinking. Thc
qualitative difference, then, did not reside in the child's representation of thc
reversibility strategy, but in the fact that there is a large class of problems in the work
for which such a strategy is necessary, and that is qualitatively different from thc
problems accessible with the old strategy. The contents of the mind could be like key!
suited for opening different doors, giving access to domains that are qualitativel!
different. But the material out of which the new key is made need not be qualitativel!
different from the old one.
A second conceptual aspect of Meno's problem is that there seem to be difficultie!
with accepting the possibility of sudden qualitative shifts. Mechanisms that ac
gradually and isochronically cannot bring about effects that are non-gradual anc
heterochronic, it seems. But this conclusion is based on the false premise tha:
short-term changes extrapolate linearly over long-term changes. Model 1.2 illustratec
clearly that this is not always the case. Where does this false premise come from!
Maybe it comes from the fact that humans, as biological beings, are built to makt
short-term decisions. For this type of decision, anything that approximates a lineal
extrapolation gives the best solution. But in order to see long-term consequences, wt
need mathematics or electronic artifacts such as computers and it appears that their
way of reasoning has not yet entered our human intuitions as to what constitutes ar
'acceptable' model of change.
The world is full of events that boil down to sudden qualitative shifts. For instance.
if walkers increase their speed of walking, they will come to a speed where the walking
pattern suddenly changes into a qualitatively different pattern, that of running. We call
the patterns qualitatively different because the components of both patterns, such as
the typical movement of the legs, trunk and arms, cannot be mixed freely in
intermediate motion patterns. The motion components in each pattern are positively
coupled; once one component occurs it will trigger all the others. But why would a
component of running interfere with a walking pattern? Probably because of a mere
coincidence, a 'random perturbation' of the walking pattern. Because of the physics oi
walking, the chance of such perturbations occurring increases exponentially with
walking speed. For instance, at high speed the walker risks stumbling or losing balance
and the onset of a running pattern will actually restore balance.
In Chapter 5, I shall describe a model of transitions, that is, sudden shifts in a
process, based on interactions between variables that are similar for all the points on
the time-scale. Nevertheless, these interactions produce a sudden shift at one
particular point in time. Figure 1.5 shows a mathematical simulation of growth in three
different cognitive dimensions: conservation, decentration and reversibility. In the
mathematical model, these three dimensions are loosely positively coupled. The child
described in the model would be in a state of non-conservation, perceptual centration
 Thinking about change and development

Figure 1.5 Three cognitive skills - decentration, conservation and reversibility - have
been modelled in the form of a loosely coupled group of variables that positively affect one
another. Their growth curves take the form of a coordinated, sudden jump towards a stable
level.

and non-reversibility for quite a long time.. T h e curves actually represent the
probability that the child would answer a question invoking any of these three
variables. Then a sudden change occurred in which the three cognitive properties
involved jumped to a maximal level. That is, the probability of a child giving a correct
answer to a variety of standard questions, such as the standard conservation question,
is almost 1. This then is the final state, which will remain fixed. Although the jump is
based on an isochronic principle, growth that acts similarly on all points of the curves,
the effect is strongly heterochronic: the jump occurs at one specific moment. It is
difficult to believe that the jump is not caused by some external factor and that
the causes that act upon the system of three variables are similar for the whole time
range.

Understanding development
Meno's problem is of course not solved by these models and examples. What they
demonstrate, however, is that part of the problem rests on our inability to understand
why and how qualitative shifts may come about if the nature of the developmental
mechanism is gradual and quantitative (more of the same). Most of our developmental
models are quite complicated, and involve many variables that are related in all sorts of
ways. But we have no other means of inferring longitudinal processes of change than
simple linear extrapolations which are valid and useful for small time frames. Linear
 The forms of change
extrapolation turns inadequate as soon as the number of repeated interactions or
repeated working of the mechanisms involved exceeds the number that we can still
intuitively grasp, which in general is not more than four or five.
Dynamic model building offers a way (not the only way, maybe and certainly not the
final way) to get out of this impasse and to understand why development is what it is.
Development is not always gradual (but it is not always a matter of qualitative shifts,
either) and development is clearly different between individuals, but also shows
general patterns or prototypical trajectories. Development is not always aiming at a
hierarchically higher steady state, but shows cyclical patterns, transient and lasting
regressions and decay. Development is sometimes highly dependent on small
coincidental factors, or differences among individuals as far as their developmental
conditions are concerned; sometimes development is robust and insensitive to what
would be major threats under different circumstances.
A theory that could explain all these phenomena would certainly considerably
advance developmental psychology. If dynamic systems modelling claims that it can
offer such a theory, it sets itself the formidable task of rebuilding and reconstructing
the whole discipline. But the point is that dynamic systems modelling does not make
such a claim. It simply says that, basically, we already have such a theory, and it is the
good old theory of development that we find in a variety of forms, ranging from Piaget
to Vygotsky, and from information processing to skill theory. These models contain a
limited set of simple, basic explanatory models and those suffice to explain all the
phenomena mentioned before. But so far they have not really explained these
phenomena, because they suffered from the major problem discussed in this chapter,
namely the impossibility of extending the effect of those simple mechanisms to more
than a few points in time. What dynamic systems theory offers is a technique or an
approach, to study what these models do ifwe let them do their work for more than just
a trivial period.

Iterates in space
What is an iteration?
I have always been a great admirer of the Muppets and I especially liked the Pigs in
Space part, featuring, among others, the dazzling beauty Miss Piggy, who would never
become tired of demonstrating that true love reaches even beyond the boundaries of
biological species. For completely trivial and unscientific reasons, the Pigs in Space
thing went through my mind when I was trying to find a short description of a
fundamental formal property of development that I could use as a title for this section.
I only had to replace the pigs by iterates.
If anything like a basic developmental mechanism exists, it must be an iteration. An
iteration is a process that takes its output as its new input, produces new output, which
it takes as input, and so on, ad infiniturn. Originally, the notion of iteration came from
mathematics, where it denoted functions like:
 Iterates in space

we recognize as the doubling scenario from Model 1.2. It starts with an initial
of a (e.g. l), doubles it, then doubles this number and so forth. The function
its output as its new input is the function 'times 2'.
e is a geometric example. Take a straight line of any arbitrary length. Divide it in
quai parts and delete the middle part. Bridge the open part by two lines whose
is one-third of the original line. This will produce a geometric form consisting of
aight lines. Apply the function (dividing, deleting, bridging, etc.) to each
ht line. This will result in a new geometric structure consisting of straight lines. It
t the complexity of the forms increases with the number of

once to a straight line and you will obtain a pattern as in Figure

five times to the resulting straight lines and a pattern results as
(right side). It is impossible to ,carry out such iterations merely
visual imagery. A computer (or lots of patience if you want to draw it by hand)
to reveal the complexity and form of the pattern.
e kinds of designs are often described as fractals, and nowadays they occur in a
of popular books that show what amazing structures such simple iterations
e. Benoit Mandelbrot, a mathematician who cooperated with Piaget for a while,
considered the modern founding father of the science of fractals. In the
ook The Fractal Geometry of Nature (Mandelbrot, 1982), he shows how
ms such as coastlines, clouds, cauliflowers and brains reduce to simple
erations. Biological processes too can be considered iterations. Take for
e sexual reproduction. Sexual reproduction is a biological process that takes two
uals, a male and a female, as its input, and, on average, produces individuals of
te sex as its output. The individuals of opposite sex enter the sexual
ction operation again, produce new individuals of opposite sex, and so on, ad

ns in mental and behavioural development

the iterations of evolutionary biology take place at the time-scale of successive
ns, those of developmental psychology require a time frame ranging from
to several years, but all within a single life span. Let me begin by telling a short
about a developmental process at work and see how the notion of iterative funct-
be applied to it.
 The forms of change

Figure 1.6 An iterative procedure takes its own output as its next input. Applying simple
iterative geometric procedures to an initial form (the triangle) produces a complex pattern
after only five iterations (the snow crystal at the right).

Many children are very interested in computers and some have learned a lot about
how to use them without any formal training. Here is a 10-year-old boy who is playing
with a graphics program that his mother got with a new mouse. He clicks on the pencil
icon and draws the rough outline of a house. He then clicks on the brush icon, goes to
the outline, clicks again, and the house turns yellow, as he expected. He then draws a
window, clicks on the brush, chooses red, clicks on the window. But instead of nicely
filling the polygon, the colour spreads all over the screen. The boy is very puzzled. He
asks his mother what he can do, and she advises him to click on the 'undo' icon. He then
goes to yellow, since the explanation of the disaster may of course be the red colour. But
the result is a yellow screen. He clicks on the pencil, draws a new window, tries to
colour it again; now it is perfect: a bright red window. Is it a testimony to his
competence, or just to his good fortune? He asks his mother why the colour leaks.
'You've probably not closed the line', she says from behind her book. The boy is
puzzled: not closed? What could that mean? 'You should close it, like a fence', his
mother goes on. If you do not close the fence the animals will run away. It is just the
same with the colour. Then the boy looks back at the first window. The beginning and
ending of the line in the second one cross one another, but the first window shows some
open space between the end points. That should be the explanation: forms must be
closed, otherwise the colour will leak out. He then draws a third and fourth window,
and both can be coloured as expected.
The story can be interpreted in various ways, dependent on your theoretical
 Iterates in space 17
background. Let us begin with a Piagetian one (Piaget, 1975). Piaget would view the
scene as an example of the adaptation principle, as it operates in the spontaneous
of children. The boy has an impressive, yet still limited knowledge of the
program. One could say that in his view the whole program boils down to a
few functions, such as the 'draw' and 'colour'. The child's interpretation (actually
his reduction) of the program with all its different possibilities to a limited number
of known functions is what Piaget calls assimilation. But the child encounters
difficulties, such as the leaking of the colour from the first window. The child's
experiencing this as a difficulty depends on his knowledge that the program should
be able to fill forms selectively. The knowledge-dependent nature of this so-called
cognitive conjict shows that it is an example of assimilation. Then his mother tells
him about open and closed forms. Because he knows the difference between them
and understands the metaphor of the animals and the fence, his mother's advice
makes sense to him. He then looks more closely at the difference between the two
windows. One is a closed form the other an open one and he can actually see the
colour flow through the little hole and fill the whole screen. It becomes clear to him,
then, that drawing and colouring is not just a matter of creating a form. It should be
a closed form. In the future, he will have to reckon with that particular property of
forms if he wishes to colour his drawings properly. This adaptation of the child's
knowledge to the properties of experienced reality, experienced though in the form
of a conflict, and mediated by his mother's verbal information, is called accommoda-
tion. Assimilation and accommodation are the polar constituents of adaptation.
After his experience with the open and closed forms, the boy has learned about
the difference between them and about their function in the graphics program. Next
time he might encounter a form that seems to be closed, but still leaks colour. That
would produce a conflict between his expectation and his actual experience, and lead
him to understanding the magnification button on the screen menu. Sometimes
holes are so small you can hardly see them on the screen, and then you have to
magnify the picture. Thus the understanding of why the magnification icon is
necessary is a function of his previous discovery that forms should be closed, and of
the resulting experienced conflict that a form seems closed, but is not, since it leaks
colour .
It is clear that the Piagetian adaptation function (let us call it PA) is an iterative
function. It takes current knowledge and an experience whose meaning is based on that
knowledge as its input. As its output, it produces new knowledge and new experiences
that are a function of the altered knowledge state. The production of new experiences
and not just new knowledge is essential; adaptation takes place in the form of action.
Action entails personal knowledge and experiences that form the interface between the
person's knowledge and the reality touched upon in that action. It may take a while, of
course, before a current experience, such as a cognitive conflict, is transformed into
new consolidated knowledge. This is the amount of time expressed as the parameter 'i'
in the s ~ b s c r i p t : ~

C (knowledge, + , experience,, ,) = PA(knowledge,, experience,)

1
I
18 T h e forms of change
+
The equation says that later knowledge and experience (at time t i) is a Piagetian
adaption function of current knowledge and experience (at time t).
In his book on the mechanisms of development, Piaget (1975) formalized the activity
of assimilation and accommodation in a form very similar to the iterative function
presented here. Piaget could not explore the most important property of iterative
functions, namely the possibility of deductive experimentation. That is, once a
computable form has been given to the iterations, we can study the effect of different
values of the parameters involved, we can investigate whether or not iterations ever
stop producing new outcomes and so forth. By doing so we are performing theoretical,
not empirical research; we investigate the potential developmental trajectories that lie
hidden beneath the seemingly trivial form of the iterative equation. So far,
developmental psychologists, or psychologists in general, for that matter, have not
been used to carrying out this kind of deductive investigation of their models, because
the verbal form itself prevented them from doing so. Only if our models are
transformed into a calculus, however simple its form, will deductive inferences become
possible. Since the nature of the developmental mechanism is iterative and since
iterative mechanisms produce results that are in general not conceivable by linear
extrapolation, a deductive investigation of the possibilities of a model is far from trivial.
The information processing model (e.g. Siegler, 1983) would have interpreted the
scenario differently. Under this model, the child is employing a rule that a class of
computer figures - some broad class of polygons - can be filled and coloured with the
brush icon. Then he encounters an event where the rule does not hold. He comes up
with a hypothesis - with a little help from his mother, to be sure - which he then tests,
and either refutes or accepts, dependent on the empirical outcome of the test. Accepted
hypotheses form new rules, and they are given actual truth values (true or false) in each
experience where they are put to the test. If a rule leads to a negative truth value (it is
false, as in the case above), then the mechanism of hypothesis testing sees to it that a
new rule is suggested and tested and that the old rule is replaced by the new one. If the
truth value is positive (the rule is true), the mechanism takes care of a further
consolidation of the rule. The iterative function of hypothesis testing (HT) is very
similar to the one proposed for the Piagetian model, if 'knowledge' is interpreted as
hypotheses and 'experiences' as 'truth values':
(hypothesis, + ,, truth value, + ,) = HT(hypothesis,, truth value,) (1.8)
(which says that later hypotheses and truth values are a hypothesis- testing function of
earlier ones).
The reader will probably notice that the form of this equation is similar to that of
equation 1.7; at this level of mathematical abstraction, the Piagetian and the
information processing model are indeed similar. Differences occur at the level of the
mechanisms that each theory invokes to explain the working of adaptation and
hypothesis testing.
Let us finally try a Vygotskyan perspective to the drawing story. Vygotsky (1978)
distinguished two major mechanisms of transition at the time level of ontogenesis. The
first is the mechanism of interiorization. It explains how activities that take place at the
 Iterates in space 19
interindividual plane, that is, in cooperation with other people, shift to the
intraindividual realm of mental capacities. The second mechanism is contained in the
concept of the zone of proximal development. This is the domain between the child's
developmental level and the developmental level represented by those activities
the child can carry out with the help of more competent others. If this help is
adequately adapted to the child's actual level, it will be interiorized and will lead to a
higher level of development, enabling the child to do by himself what he could do
earlier with help. By making his computer drawing of the house and colouring it, the
child shows his actual developmental level. The boy's mother can then build on this
level,by telling him to test whether his figures are closed or not. He will then interiorize
this advice and test the closure of figures without being explicitly told to do so. With
this particular boy, the simple verbal advice represented adequate help. His mother
could also have told him to try again, or to read the manual. With adults, the suggestion
'read the manual' could be very helpful, but with 10-year-olds it is probably a wasted
effort. On the other hand, the boy's mother could have inspected the drawing of the
window herself and told him what to do step by step, without giving much of an
explanation. Then the boy would not have learned very much. Now imagine what
would happen next time the boy drew a polygon and the colour leaked out to the entire
screen. His mother could again suggest to him to control for the closure of the figure,
but that would not be of help, since the boy already knows that. Consequently, in order
to promote the boy's further learning, his mother should adapt her help to the higher
level of competence her son has reached, thanks to her previous help. It is clear that this
mechanism is again an iteration, that could be written in a form similar to the Piagetian
form, with 'developmental level' standing for 'knowledge' and 'help' for 'experience'.
It is also possible to write the mechanism in the form of a set of coupled equations,
however, which are not iterations themselves, but which as a couple form an iterative
process.
Let us distinguish the 'tutoring' function T from the interiorization function I.
Tutoring takes as its input the child's actual developmental level, as expressed in the
child's problem-solving actions. As its output, it produces help adapted to the child's
actual developmental level. Interiorization takes help as input and produces a new
developmental level as output:
,
help, + = T(developmenta1 level,)
,
developmental level, + = I(help, + ,)
(help is a tutoring function of developmental level; developmental level is an
interiorization function of help).
This set of coupled equations can also be written in a singular form, namely:
developmental level, + ,= I(T(developmenta1 level,)) (1.10)
The mere fact of concocting these formal expressions is not quite a major contribution
to developmental theory, since they add hardly anything to what we already know
about Vygotsky's model. But, once they are put in a form that enables us to compute a
sequence of developmental and help levels, they will reveal a number of developmental
20 The forms of change
possibilities that would have remained unknown if only the verbal form of the model
had been taken into account. How such computations can be carried out will be
explained in later chapters.

The elementary iterative form of developmental ,processes

Although the nature of the concepts involved differs over different developmental
theories, they all basically agree on an elementary iterative form. It is so elementary
that it appears almost trivial; it states that the next developmental state (ds) is a
developmental function (DF)of the previous developmental state:

,
ds, + = DF(ds,)

'State' does not mean the same thing as 'stage' or 'phase' (see van Geert, 1986). It is a
very general term used to denote whatever developmental level concept you wish to
employ. For instance, for a continuous developmental function such as a continuous
increase in the mastery of a skill, the developmental state could be any point on the
continuous curve that models this increase. In a Piagetian model, 'state' could refer to
one of the four major stages, or to a substage.
How trivial is it to say that a developmental state is a function of its predecessor?
Probably less trivial than it seems at first sight. Imagine a sequence of developmental
states behaving exactly like a sequence of heads and tails resulting from tossing a coin.
Each toss is completely independent of each preceding one. Many people have
difficulty appreciating this principle of independence, however, since it is an essential
aspect of common sense that the future is at least to some extent determined by the
past. For instance, in the unlikely but possible event that you threw 100 heads in a row,
many people would believe that the probability of getting another head for every
further toss would significantly decrease, but that probability is actually always 50 per
cent.
If we were to endorse a strictly maturational theory, the developmental states would
arise independent of one another. Their real dependence would be on the underlying
genetic program. Assume that the genetic program would trigger formal thinking, for
instance, at the age of 11. The previous stage of concrete thinking would not relate in
any sense to the emergence of this new form of thinking, except for the fact that during
a short transition period, thinking might show some mixture of formal and concrete
elements. Similarly, a theory describing developmental states as mere environmentally
driven functions, determined by the developmental or learning tasks set by the
environment, would not subscribe to the iterative model either. For instance, a specific
educational environment may decide to teach Latin to 10-year-olds, and analytic
geometry to 12-year-olds. In this case there is nothing like an obvious developmental
logic behind this sequence of knowledge acquisition. Nevertheless, we may question
whether it is possible to avoid a later stage being determined, in a non-trivial sense, by
the previous stages. Even if we must learn a completely new skill, such as an exotic
language having nothing in common with our mother tongue, the initial states of
 Iterates in space 21
learning will differ from person to person and depend on each person's general learning
&ills, semantic knowledge and so forth. As soon as the learning process has started,
however, each new step in the learning will have to rely on the previous steps. Now,
imagine we have to learn a list of unrelated nonsense words. Each word is a learning
task of itself and the learning of one word would not contribute to the learning of
others. But learning new words may become increasingly demanding as the list of
words to be remembered grows longer, as the person becomes more fatigued or bored.
We may conclude, therefore, that the iterative function expressing each state as being
determined by a preceding state is an almost universal, inescapable property of change.
In developmental processes, the iterative principle probably applies in the most
explicit way possible.

Time series, trajectories and the concept of a developmental space

Development trivially implies change over time. How will the change be represented?
In dynamics, changes are depicted in spaces of various kinds, time being reduced to
one of the spatial variables employed in a model. For instance, a simple population
dynamics model might specify a space consisting of the time dimension and a
dimension for the number of individuals in a population, of foxes for instance. A more
complicated model might add the number of rabbits as a third dimension. At each
moment in time, there is a specific number of rabbits and a specific number of foxes.
That is, the state of the system is represented by a single point in the time-fox-rabbit
space. Since the numbers of foxes and rabbits change, the process of population growth
and decline of rabbits and foxes is represented by a single line in the three-dimensional
space. This line is the population trajectory for the fox and rabbit system.
Let us apply this notion of a descriptive space to development. My current social
knowledge, for instance, differs quite drastically from the social knowledge I had when
I was only 2 years old (well, I sincerely hope it does). It is possible to describe the
distinction between my knowledge now and my knowledge then by using a finite
number of descriptive terms. One of the descriptive terms could refer to my cognitive
representation level, which is now formal operational and used to be sensorimotor
when I was 2 years old. Another descriptive term specifies the number of social roles I
am able to distinguish. A third refers to my egocentricity in social matters, and so on.
Each descriptive term functions as a dimension in a space that covers all possible states
of social knowledge. I can employ whatever collection of descriptive terms I consider
necessary for the task of distinguishing any two developmental states.
The standard way of specifying my development in this descriptive space would
consist of taking a test that covers a specific descriptive dimension. By repeatedly
applying that test, with an interval of, say, a year, a so-called time series of
measurements would be obtained for that dimension. The time series is easily
represented in the form of a curve. T o obtain a complete image bf the developmental
process, time series for each of the descriptive dimensions should be collected. A
collection of time series is far from a coherent image of the overall development,
however. One way of achieving a simple overall picture in the form of a time series
22 The forms of change
works as follows. I can connect any two points in my multidimensional descriptive
space by a straight line of definite length. Figure 1.7 shows how this is done in a space
with no more than three dimensions. The initial state with coordinates (0,0,0) is
connected by a diagonal line with the final state (1,1,1). If the length of each of the
dimensions (or sides of the cube) is set to the unit length 1, the length of the diagonal is
equal to the square root of n, the number of dimensions. The simplest possible way to
phrase the question of developmental dynamics is this: How is the distance bridged
between initial and final state and does the speed and direction eventually change over
time? Put differently, the multidimensional space of descriptive dimensions has now
been reduced to a two-dimensional space consisting of a distance and a time dimension.
It goes without saying that the transformation of an n-dimensional distinction
between developmental states to a one-dimensional, unspecified distance is a
formidable reduction, and we might worry that it is too drastic. But although it is a
severe reduction indeed, it is not irrelevant. The point is that the general dynamics,
that is, the basic characteristics of the way in which the initial state transforms into the
final state, can be specified by using the single distance line.

Figure 1.7 A three-dimensional space consisting of three distinctive features A, B, and C

represents the set of possible paths in a developmental model. The paths are lines
connecting the vertices of the space. The diagonal line represents the distance between the
initial and the final state.
 Iterates in space 23
Figure 1.8 depicts a possible, characteristic trajectory between an initial and a final
ate, in the form of a stepwise growth process. Such a representation is called a time
erie~representation, because it displays a series of data points over time. The data
oints must be at intervals small enough to capture the basic dynamics of the process at
sue. Of course, what is considered basic depends on the scale on which we choose to
dy a particular phenomenon. Note again that the present way of representing
velopment does not boil down to assuming that the development of social
owledge, or any other form of development for that matter, is a one-dimensional
ssue. The single dimension is an abstract entity drawn from a, descriptive space in
which each dimension has a particular meaning. The obvious question that arises as we
specify development in such terms is what this abstract dimension means empirically.
We would certainly have to look for 'local' empirical correlates of this abstract
dimension.
Let us return to the imaginary multidimensional space that represented the
distinctive properties of social knowledge. We do not have to know all these
dimensions, it suffices that we know they constitute a space in which distances can be
measured and axes drawn. Instead of graphing the change in only one dimension at a
time, I now combine the measurement of a dimension or property A with that of a
property B. The developmental state is a point with an A coordinate and a B coordinate
value. When I connect the A-B points in the right temporal order, I obtain a line in the
A-B plane which is called the trajectory of the developmental system (which is my
social knowledge in this particular example). The trajectory runs through a space, the
A-B plane, and connects successive developmental states. The A-B plane is called the

Figure 1.8 If a developmental process takes place in a space of variables as in Figure 1.7,
the resulting developmental process may follow a trajectory of step-wise growth.
24 The forms of change
state space and the trajectory through the space the state space representation. Note that,
unlike the previous pictures, there is no time axis in the state space representation.
Time can be added by imagining the trajectories are paths, or bike tracks, for instance:
the biker needs a certain amount of time to go from the start to the point where the ride
ends (which in path 2 is probably somewhere along the cyclical part of the path).
State spaces can contain any number of descriptive dimensions, but they are usually
confined to only two, for reasons of clarity. Any choice of two dimensions will greatly
reduce the level of detail and specificity of the resulting developmental picture. I can
try to overcome this by selecting any two axes through the multidimensional space that
yield the most characteristic picture. Even if I do not know the meaning of those two
abstract axes, they may be helpful in that they allow me to see the characteristic
developmental trajectory.
Figure 1.9 shows an imaginary state space representation of the development of
social knowledge in two subjects. Each subject is represented by a single line. Their
starting points are very close to one another, but the way they grow towards their final
state point differs. Path 1 shows more progress over the B variable at the beginning,
while path 2 starts to progress over the A variable. The endpoints are different in that
path 1 spirals towards a point attractor, whereas path 2 keeps spiralling over a cyclical
path in the A-B space. Path 2 represents a developmental path which never ends.

Figure 1.9 A state space representation of a developmental model consisting of only two
variables ( A and B). This kind of representation lacks an explicit reference to the time
dimension; developmental trajectories are like bike paths on the plane formed by the
variables A and B; path 1 evolves towards an endpoint, path 2 towards a repetitive cycle.
 Iterates in space 25
ies to trajectories
dents of dynamic systems like to look at things from a variety of different angles.
en they see a process they want to know not only its time series representation, but
e trajectories. Whether you prefer to show the properties of a process
er in a time series or in a trajectory form is a matter of taste, of convenience, but
ty. Some processes show their properties better in the form of a
ave seen that a state space representation requires at least two
what if a process changes over only one dimension? Is it possible to
ransform a time series into a trajectory representation? How do we carry out that

Model 1.3 Economic interest cycles: transforming

representations
David, who is 19, has decided to take a course in macro-economics. You might say that
David is very interested in macro-economics - he has actually thought of taking it as his
major. Neverthless, if we say that David is very interested in macro-economics, we do not
mean that he is equally interested all the time. Sometimes he is more interested in maths, in
playing soccer or in finding the cutest girl at the Halloween party. We could say that his
interest in macro-economics is oscillating with a period equal to a day. His economic
interest is expressed in his doing his economics problem's for homework every day without
being pressed to do so, his willingness to start a discussion on economics and so forth. A
fugitive thing such as a student's oscillating interest in a subject is probably very difficult
to measure. But, we can just pretend we know the function, and can pretend that we have
a mathematical model generating it. This mathematical model is a model for a so-called
exhaustible resource, which will be explained later.
The equation is very simple. It is actually based on the logistic growth equation that will
be explained later, but applies a delay to the resource level. I suggest you define a set of
range names that will enter as parameters in your equations. The spreadsheet manual will
explain how to do this in your particular spreadsheet. Basically, however, it goes like this.
Go to a cell, for instance C100 and type the name of a parameter, r. In the adjacent cell,
D100, enter the value of the parameter r, 0.1. In cell C101 specify the name for the second
parameter, e and in Dl01 its value, 0.2. Via the command Range Names you can specify
the name, rfor instance, and the corresponding range, Dl00 (the cell where you can find
the value of r ) . After defining the range names, go to cell A100 and enter the initial value
0.01. In cell A101, type the equation:

In cell 6101, type a very simple version of the first derivative of the process described in the
A range: . A

Copy range A101 ..B101 to range A102..B500 (or a bigger range if necessary). Define three
graphs: one is a line graph of the range A100..A500 (or whatever A range you have
26 The forms of change
specified), the second is an xy graph, with the A range associated with the y-axis, and the
B range with the x-axis. The third graph is again an xy graph. It takes the range Al00..A500
as its y-axis and the range A108..A500 as its x-axis (note that the x-axis contains the same
data as the y-axis, with a shift of eight cells).

Let us assume that we have been able to measure the oscillatory movement of the
student's interest and that it looks like the time series in Figure 1.10 (bottom); the
student's interest oscillates over time (time is the horizontal axis). The time series can
be transformed into a trajectory representation in two different ways. First, take the
sequence of successive levels of the interest variable and put this on the x-axis. Then
take the first derivative of each point in the sequence. In Figure 1.10, I have feigned the
first derivative by the difference between two consecutive cells. These values are then
put on the y-axis, so that for every value of the interest level there is a value of its first
derivative:

Figure 1.10 Two phase space representations (top) of a time series (oscillation at the
bottom). T h e phase space at the left is based on the time-delay procedure, the one at the
right depicts the relationship between the local level of the process and the acceleration at
that level.
 Iterates in space
(x, y ) = ( 4 , A&)' (Lt + 1, ALt + 1 ), . .
*

is set of x-y points yields a representation of the oscillation function known as the
se space representation. It is a classical representation in physical and technical
ences where many processes amount to oscillations of a variable. The phase portrait.
the imaginary student's 'interest in economics' level is depicted in Figure 1.10 (top
ht). Note that the time axis has been removed from the picture; the x-axis represents
student's interest levels, the y-axis the corresponding changes in that interest level.
e grey dot marked 't' represents the student's interest level at time t and the amount
change in that level at time t . Although the oscillation itself looks quite regular, the
se space shows characteristic increases and decreases in the speed with which the
interest changes and this is the kind of information that is much harder to infer from
mere inspection of the time series. The phase space representation shows a line
piralling inwards, which normally refers to an oscillating process that aims towards a
eady (that is, stable) state. The phase portrait of a regular sine wave, on the other
and, would be a perfect cycle. Many statements about the form of a process, such as
its being cyclical, its aiming towards a steady state or 'point attractor' and so forth, are
based on a phase portrait representation. It is a natural way of representing processes
and it is an intrinsic component of dynamics studies.
There is a second way of representing a phase portrait of a process, one that does not
quire computing the first derivative. The x-axis is the set of points occupied
successively by the variable whose change we want to ,represent (such as the student's
alleged interest in economics function). The second axis, however, portrays the same
sequence of points, starting not with the first but with a later point. For instance, if the
+
first point of the series is taken at time t, begin the sequence at the y-axis at time t n
(Figure 1.10, top left). Thus each point in the sequence is coupled with a later point, for
instance eight points later:

This set of points yields a function that is quite similar to the classical phase
representation and provides roughly the same information (Figure 1.10, top left). The
choice of an optimal delay (for instance, the eight time points employed here) depends
critically on the phase with which the process cycles. With some trial and error, it is in
general possible to find the optimal distance fairly easily.

The many faces of a process

Let us apply the strategy of multiple representations to an empirical example. Kurt
Fischer, whose work will 'be discussed further in this book, has measured the
development of skills, such as arithmetic problem-solving skills, under different testing
conditions (Fischer and Pipp, 1984). If the testing implies feedback from the tester to
the child and information and support given in the testing situation, the curve
representing the skill growth (in terms of the percentage of correct answers to a test)
shows a characteristic shift: a sudden considerable increase around the age of 15.
28 T h e forms of change
Without practice and support, the curve shows a pattern of smooth, constant increase
(see Figure 1.lla). Fischer's point is not that practice and support measurement
captures the 'real' level, whereas the other is 'false' or a measurement error. His point is
that there is no such thing as a real competence level. A skill is actually a bimodal or
bidimensional object and those dimensions can be captured by measuring the skill
under two different conditions.
This particular view on the nature of a skill fits naturally with the notion of
multidimensional representation described in this chapter, except that 'multi' means
no more than two in this case. Let us explore the different ways in which the
development of a skill can be represented. I t should be noted, however, that the
empirical curves used in this example are not longitudinal time series of a single
subject. They are based on cross-sectional research. Since the focus of the present
demonstration is on different representations of processes, I shall pretend that the
curves are indeed individual and longitudinal and therefore real time series.

Figure 1.11 A time series representation of two developmental processes (growth of

arithmetic understanding with and without support during the testing condition, figure a)
can be caught in the form of a combined dimension (figure b), in a state space
representation consisting of the variables 'testing with' and 'testing without support' (c),
and, finally, in the form of a time-delay phase space (figure d). Each form of representation
has its particular advantages and disadvantages.
 Iterates in space

Model 1.4 Mathematical problem solving in two

dimensions
Open a spreadsheet and type the following numbers in column A, beginning with Al:
1,I ,I , I ,15,7,9,90,83,87,100,100.These numbers are percentages correct on a practice and
support test of arithmetic thinking. Each number is the mean score at an age level
expressed in years. In column B, type the numbers corresponding with the scores on a
non-supported arithmetic thinking test: 1, I ,I , I , I , I , I ,5,10,16,21,35.
Let us first define a simple time series representation. Define a graph which has as its A
range the numbers AI..A12, and as its B range BI..B12. Define it as a simple line graph.
The result will look like Figure 1.1l a . The curves show how the distance between a 1 per
cent correct and a 100 per cent correct score is bridged over a period of twelve years.
Next, define a different type of curve, an xy curve, that you can choose from the graph
menu in your spreadsheet. As x-axis, take the range B1..B12, and take A1 ..A12 for the A
range. The result will be as in Figure 1.11~.It is a state space representation of the
development of arithmetic thinking in a state space consisting of the dimensions 'practice
and support testing' and 'unsupported' testing. The state space representation specifies
development as a journey in a space of two measurement dimensions.
The third form of representation reduces the two-dimensional state space again to one
single dimension, namely the distance between an initial state and a final state (which
correspond with the initial and final test scores on both tests). In the spreadsheet, cell A12
contains the last score of the practice and support condition, whereas the last score of the
unsupported condition can be found in cell 812. Go to cell C1 and type the following
equation:

The function of this equation is to compute the distance between any point in the state
space and the endpoint of the growth process of the skill, as measured under two
conditions.
Copy this equation to C2..C12. Define a line graph that takes CI..C12 as its A range. The
result will be as in Figure 1.11b. It is the one-dimensional distance representation of the
trajectory over the two measurement dimensions.
The fourth representation requires a smoothing technique that I will explain later. For
the moment*suffice it to say that it is based on the numbers from column C, the
one-dimensional representation. It is an xy representation and it takes the value of the C
column at time t as its x-axis and the value at some later time as its y-axis. The result is a
time-delay phase space representation.

The four representation forms describe the same data series, one front practice and
support and one from unsupported testing. They are four different ways to look at the
dynamics of change in arithmetic problem-solving. Developmental psychologists are
not really used to working with state and phase space representations, or with abstract
representations such as the one-dimensional distance representation. Their data
usually consist of simple time series representations for separate dimensions. Some
properties, such as cyclical and quasi-cyclical developments, require different
representation forms, such as the state space. In the case of models with many variables
 The forms of change
and dimensions, a distance dimension representation might reveal underlying
characteristics that cannot be seen in the separate curves. One of the advantages of a
dynamic systems approach is that it offers a variety of different representational
formats and a very liberal approach to how they should be used.

1. Before any misunderstanding arises, what makes Belgians look like Belgians is their country,
and moreover, they do not belong to the animal but to the Belgian kingdom.
2. A heterochronic process would be one in which the nature of the mechanisms involved changed
over time, or is time-dependent. For instance, if the process started with a tripling scenario,
then changed to a doubling and ended with a quadrupling, the process would be termed
heterochronic.
3. Strictly speaking, however, the relationship is not non-linear but it is not linear in the
straightforward sense either. It is log-linear, that is, if you plot the logarithm, you find a
straight line again.
4. Later we shall call this the feedback delay and argue that it is of considerable importance in
development.
Developmental paths and
systems

The paths of change

Uniform change
In the previous chapter we saw that developmental mechanisms are iterations, doing
their repetitive labour in spaces consisting of the properties we consider important in a
specific developmental field. Iterations produce paths or trajectories in these abstract
spaces. What are the characteristic trajectories, or 'paths of change' in developmental
processes? In this chapter, I shall first discuss a variety of developmental paths that can
be generated by non-linear dynamic models: I shall then proceed with a second,
theoretical question: What kind of developmental theory is a non-linear dynamics
model? We shall see that the way such models conceptualize and describe development
is very different from standard models of development. Let me begin with an overview
of prototypical ways of bridging the gap between initial and final states.
The simplest possibility is that the bridging of the gap amounts to a gradual and
uniform reduction of the distance. That is, the distance reduction is a linear function of
time. If it takes a week to learn twenty new words, then the child will know forty words
after two weeks, sixty after three and so forth. If the child no longer encounters new
words (which is highly improbable in a culture like ours where the vocabulary fills
literally thousands of dictionary pages), the learning stops.
The iterative function that goes with this sort of change is a simple additive
operation:
x , + ~= x t +a (2.1)
for a an increase (or decrease) parameter. The dependence on time is expressed by the
function:

In Chapter 1, however, we encountered another type of simple linear increase, namely

the doubling scenario. It amounts to the iterative function:
32 Developmental paths and systems

for a a rate of increase. The change is again a function of time, namely:

In Chapter 1, we have seen that it leads to a 'sudden' explosion, however. If I plot the
logarithm of the change, the logplot follows a straight line (see Figure 2.1; a spreadsheet
program usually has an inbuilt facility for logplots).
Stated in the jargon of dynamic systems, these processes have a point attractor. That
is, the process is attracted toward a single point in the state space, namely the point
whose coordinates are the final state values of all the variables that form this state space.
What is the point attractor of the gradual change represented by the 'plus a' and 'times
a' scenarios? It is the mathematical point infinity. Real processes of this particular kind
will never reach infinity (whatever concrete meaning that point might have, if any
exists); they will be stopped by some external factor, such as death.
The simple additive or logarithmic increase serves as a powerful prototype for
short-term change in many developmental investigations. If a researcher repeatedly
tests a child for the same developing variable (such as writing skill or moral judgement)
and finds an irregular, oscillatory or downward path, it is usually assumed that this
must reflect random fluctuations or measurement errors and that the true variable
follows a path of linear or log-linear increase.

Steady-state 'paths'
A second possible path is one in which nothing happens at all. The development

+
Figure 2.1 The fact that the x ax scenario is called 'linear' follows from the straight
logarithmic plots of the curves based on that scenario. The plots of the actual levels are
curvilinear.
 The paths of change 33
achieved in some variable stays at the same level. The iterative developmental function
produces a stable state, that is, its inputs are identical to its outputs. It may seem
strange to call this a path, but that name is well chosen since path technically means
nothing more than a sequence of states over time. Whether the states differ or remain
the same does not matter. The point attractor of the process is the steady state itself and
the attraction is no change at all as far as that state position is concerned.
The steady-state path occurs, for instance, when the child has reached the level of
linguistic maturity and has appropriated the mother tongue's grammar, or when an
adolescent has reached the level of formal operational thinking.
The steady-state path is very important in classical models of development. Models
such as Piaget's on cognitive development, Erikson's on identity and Kohlberg's on
moral development describe final states or stages, such as the stage of formal
operational thinking (Piaget). As soon as such a state is reached, development takes the
form of a zero transformation; the iterative function that made the process evolve
towards the final state is now continuously producing outputs similar to its inputs.
Of course we could claim that with the advent of the final state, development ceases
to exist, or that the engine that made the process move has come to a standstill. But this
is probably not so. A standstill in nature means death. The mature body, for instance,
renews itself continuously in a metabolic process very similar to growth. But whereas
growth produces change (in length, for instance), the steady state merely maintains the
system. It is highly likely that behavioural development obeys the same law; the
mechanisms that accounted for growth and development during the early years
probably continue to operate to maintain the system and prevent it from deterioration.

Deceleration paths
The third characteristic path form is based on the idea that retrieval from an
exhaustible resource produces a combination of gradual increase ending in a steady
state. As a developmental path, it reduces development to a process of appropriation or
learning. Examples of exhaustible resources are the limited vocabulary of a language,
the limited range of social rules and habits and the current state of accepted knowledge
in a scientific discipline. The iterative function modelling this path is similar to picking
candies from a box. Each time you take some candy out of the box, the amount of
available candy decreases and this decreased supply is then the input for the next
candy-picking operation, until the box is empty (there is one biblical example - with a
different type of supply - in which this general rule did not hold, of course, but for the
rest of history it has remained an iron law of nature that candy boxes, or baskets with
loaves and fishes, for that matter, get empty after a finite amount of time).
There is, however, a much more interesting iterative function than the candy box
scenario, namely:
n,, ,= n, + a(S - n,)
The point attractor of this process is S (for instance, knowledge of all the words in the
vocabulary). Given this particular form of the equation, n will never reach S; it will
34 Developmental paths and systems
only asymptotically approach S. If S is a discrete number (such as the number of
words), there will be a final state for which n = S. The path produced by this equation
is one of continuous deceleration; its rate is maximal at the beginning, and it decreases
as it approaches its asymptote. Note that the iteration keeps working forever, but that it
will produce outputs approximately similar to its inputs if n approximates S closely
enough.
It may come as a surprise that a simple iteration like equation 2.5 produces the
curves typical of classical learning theory, which thrived in psychology roughly until
the early 1960s. An example of such a curve is Figure 2.2. The logarithmic plot is not a
straight line, but a line similar to the curve itself. The nature of the change is therefore
different from the 'times a' scenario, although the resulting curves are superficially
similar and look like each other's inverse.
These curves occurred in classic learning experiments, where subjects had to learn
lists of nonsense words or paired associations. In 'natural' learning processes, however,
taking place over a longer time and involving meaningful and relevant material, we will
very probably find curves that deviate a good bit from the ideal. But again, if
confronted with irregular repeated measurements of a learning function, many
students of development would still be inclined to say that the irregularities are random
errors and that the real underlying learning curve has the ideal form of Figure 2.2.

Rise and decay paths

A very old conception of human life sees it as a process of rise and decay. At birth we are
taken out of the cosmic void, we develop and prosper and finally we experience our final

Figure 2.2 Paths towards a steady state: (a) the 'candy'-scenario; (b) the learning scenario.
Note that the logarithmic plot of the learning scenario is curved, and not straight as in the
case described in Figure 2.1.
 The paths of change 35
cline and return into the same eternal void at death. Of course, this rise .and fall
tion is also applicable to much less romantic phenomena. Take for instance the use of
-called two-word sentences. In general, children younger than 18 months will stick
one-word sentences and show a growth spurt in the number of words they
derstand and use. Then they discover that combining two words in a single
erance greatly enhances their communicadve ability. Some children show a
o-word boom. As they discover the possibility of combining three or more words,
d the use of syntactic and morphological rules, the frequency of genuine two-word
terances drops again to about zero. So the developmental variable 'two-word
ammar' shows a characteristic rise and fall sequence.
Comparable examples hold for any transient developmental phenomenon, such as
instance the temporary use of pre-operational logic by children younger than 6,
ich emerges, grows and blooms, then disappears with the rise of operational
inking. In its ideal form this type of developmental path follows a bell shape (Figure
.3). This path again has a point attractor, namely the zero state.
Partial decay forms are also possible, for instance when pre-operational thinking is
vertaken by operational thinking, but continues to be used in a small domain of
roblems. In this case, the point attractor is the constant, low level of use that this form
hinking retains.

Paths with changing growth rates

he rise and fall scenario and the preceding asymptotic growth scenario are both
amples of processes where the rate of change differs over the process. In the

Figure 2.3 A typical 'rise-and-fall' path of a linguistic strategy (two-word combinations)

whose decay is caused by competition with another, more complex strategy (syntactic
structures).
36 Developmental paths and systems

Figure 2.4 Variations in the rate of change can take the form of a spurt, characterized by
an increase and a decrease in the rate.

asymptotic case, the rate of growth decreases towards the end. In the rise and fall case,
the rate turns into a negative value (decay, loss) at the summit. In fact, the majority of
developmental paths, as we intuitively understand them, show some sort of
combination of changing rates of growth.
A classic example of a process that shows a definite acceleration followed by a
deceleration is the so-called naming explosion that normally takes place between the
ages of 18 and 22 months (see Figure 2.4). It is characterized by a considerable increase
in the rate with which the child learns new words. After a preparatory stage in which
the child builds up a lexicon of, for instance, about fifty words, a lexical growth spurt
will occur. Finally, the rate of lexical growth will.leve1off by the time the child starts to
use two and more word sentences. Every parent will be able to give examples of
development that seems to proceed slowly for quite a time, then suddenly explode and
later slow down again. For instance, there are peaks in children's stubbornness over the
years, or in how easy it is to deal with them in disciplinary situations.
Sometimes temporary regressions may occur, in which children seem to get worse
instead of improving on some specific skill (Figure 2.5). Very often this announces a
leap to a higher level of mastery. For instance, young children use a limited number of
spatial prepositions in a grammatically correct way. Then, as they try to discover the
rules behind the use of the prepositions, they sometimes overgeneralize the
hypothesized rules and make more errors than they did before. For instance, a child
speaks about 'in the bed' and 'in the table', although it formerly distinguished both
verbal contexts and correctly used 'on' in the context of tables. But such a regression
often heralds a spurt in the correct use of a larger variety of prepositions (van Geert,
1983).
 The paths of change 37

igure 2.5 A transient regression or dip may precede a growth spurt.

Figure 2.6 A step-wise increase in the initial-final state distance is characteristic of stage
theories.

Finally, the classic stage theories can best be represented as a stepwise increase
(Figure 2.6). For instance, each Piagetian stage is a single step of a stair with four steps
in all. This type of curve amounts to several asymptotic growthsurves superimposed
on one another. We may of course object that Piaget, and most other structural
theorists, for that matter, have described the stages as qualitative changes, and that a
one-dimensional representation is unsuitable. This is correct, but as the reader will
probably remember, the line represents a distance in the multidimensional space
38 Developmental paths and systems
specifying the qualitative differences in Piaget's model. It specifies how this distance if
bridged without referring to the qualitative nature of the transition.

Non-linearity in point-attrador paths

The trajectories described in the preceding section are of the classic type: developmeni
aims towards a final state (a state of maturity), and stays there. The final state is a singlt
point in the state space formed by the variables that together specify the set of possible
developmental states. So, classically, development is a process with a point attractor
The paths, however, are not characteristic in the empirical sense, in that they have no1
really been found in empirical data, except in very confined situations such as the
classic learning experiment. Rather, they are widely seen as prototypes of develop-
ment, as curves that we should find if measurement problems would stop playing thei~
dirty tricks on us. Onto a few elementary paths, such as the increase and the
steady-state path, we may build more complicated paths, such as the stepwise path
behind stage theories. Remember that, in the theoretical sense, each of these paths iz
not just a line in space, but the product of an iterator. For some elementary paths, the
iterators are simple. For others, such as the stepwise growth or the transient regression,
they are more complicated and we shall postpone discussion of them to Chapters 7 and
8. The iterator generates paths: give it any point of a curve - at the beginning, the end
or somewhere in the middle, at a summit or a dip - and it will return the next point of
that curve, no matter how complicated the curve is.
The more complicated curves differ from the simple ones in an important aspect.
The simple curves obeyed a monotonic increase pattern. Start with any arbitrary point
on the curve and the iterator produces the same change, a uniform increase, for
instance. In the more complex ones, such as the transient regression, the form of
change differs over the length of the curve. That is, take a point at the beginning and
look at its successor. It shows a slow increase. Take a later pair; they show a faster
increase. Still further the increase turns into a decrease, which turns into an increase
again. This phenomenon is called non-linearity. It is an important property of dynamic
systems, which are called non-linear dynamic systems in that case. I shall discuss
non-linearity in Chapter 3. For the moment, suffice it to say that non-linearity involves
non-uniformity, or diversity, of the relation between a cause and its effect.

Cyclical transitory paths

Cycles in classical theories
The previous paths are typical of development in that the) generally aim towards a
higher level, a single point in the state space. Many developmental theories, though,
explicitly reckon with cyclical patterns of change. Although such theories clearly
distinguish an overall process towards a higher, stable state, they often conceptualize
each step as a cyclical process. Piaget, for instance, sees in every major stage an
underlying process of equilibration. Equilibration means that the components of a
 Cyclical transitory paths 39
cture compensate for one another so as to form a coherent and functional whole. In
t sense is development cyclical, according to Piaget? When the child discovers a
structural level of thought, such as concrete operational thinking, the cognitive
m starts with a state of disequilibrium. For instance, the child tends to apply the
Crete operational logic to all kinds of problems, whether it actually fits or not. The
tation mechanism, however, will gradually lead to equilibrating the system, in that
hild learns to use the concrete operational logic to compensate for different kinds
turbances, such as misleading counter-suggestions. Arriving at the summit of the
ete operational competence with its full equilibration of the logical means that go
it, the child discovers a new and higher level, namely formal operational thinking.
th the discovery of this new level, the original equilibrium reached with the previous
m of thinking is set to zero again and the child has to repeat the cycle of finding a new
ilibrium. The process takes the form of an S-shaped curve that jumps to a higher
el of functioning as it approaches its maximum. But this higher level of functioning
igher stage) is automatically the lowest level of equilibrium. If we project such a
cess on a single plane, abstracting from the shifts in cognitive level, we find a
tooth curve, projected on different levels of cognitive functioning (Figure 2.7).

Figure 2.7 Equilibration follows a rise-and-sudden-fall scenario in three dimensions. Each

new cognitive level (e.g. Piagetian stages) shows a new equilibration process, resulting in
the emergence of another level. The three-dimensional process reduces to a sawtooth
pattern if represented in two dimensions.
 Developmental paths and systems

Figure 2.8 Fischer's theory describes a four-step cycle: the fourth step is the first step on
a higher level that emerges as a consequence of the preceding steps. Each developmental
level (called a tier) consists of three sublevels (1, 2, 3).

A second example of a cyclical pattern comes from Kurt Fischer's theory of skill
development (Fischer, 1980). Fischer distinguishes four major tiers of skills: reflex,
sensorimotor, representational and abstract. Each tier develops through the same
cycle. It starts with a collection of single building blocks called sets, for instance, single
representations (of cars, Ninja Turtles, monsters, baseballs, etc.). Then single building
blocks are coupled into relations called mappings (for instance, Volkswagens are a
subclass of cars). In the third stage, relations (mappings) are combined into relations of
relations, called systems (for instance, the child understands that Volkswagen is to car
as apple is to fruit). In the fourth stage, systems are coupled (for instance, the child
understands that there are subordinate relationships mirrored by superordinate
relationships). These coupled systems form the initial component of the next higher
cognitive level (see Figure 2.8). For instance, systems of representational systems form
the singular building blocks or sets of the next higher abstract tier. Fischer's model
describes a cycle of four discrete states and each fourth state jumps to a higher level to
form the first state of the next tier. I n fact, Fischer's model describes a regular
three-state cycle. The phase portrait of the process is based on the assumption that the
distance between any two consecutive stages has the unit length 1. After smoothing,
the trajectory takes the form of a circle flattened into a bean-like shape. It is a
characteristic cyclical trajectory.
 Cyclical transitory paths
uasi-cycles in actual development
In Chapter 1, I explained the notions of phase space and state space and gave the
ample of an oscillating resource, such as a person's interest or effort spent in a specific
bject. Comparable oscillations can be observed in language development. In Roger
own's classic study on first-language development (Brown, 1973)' we can find curves
owing the development of syntactic rules such as the present progressive in
ividual children. The curves of the average percentages of correct use per month
ow a far from steady rise, but go up and down quite considerably. The underlying
end, however, is a continuous increase towards (almost) 100 per cent correct. Figure
9 renders a smoothed version of the monthly averages, clearly showing the curve's
cillation around the rising average. Similar patterns can be observed in the learning
new words. It is not so much the number of words that oscillates- there is hardly any
idence for children's large-scale word forgetting over short periods - but the rate
ith which new words are learned. We might expect this rate to be stable; it depends on
ognitive abilities that do not seem to fluctuate widely, or that are supposed to improve,

Figure 2.9 The steadily increasing oscillatory pattern of acquiring the correct use of the
present progressive has a semi-cyclical phase portrait.
 Developmental paths and systems

Figure 2.10 Irregular oscillations in the rate of word learning: time series and phase
portrait.

such as the child's understanding of new concepts and relationships. Nevertheless, the
rate with which new words are learned varies rather strongly in individual children
(Corrigan, 1983; see Figure 2.10). The data from the literature consist of single data
points, such as a monthly average of percentage of correct uses of a morpho-syntactic
rule. How are they transformed into the cyclical patterns of Figures 2.9 and 2.10?

Model 2.1 A simple smoothing procedure

Drawing the phase space involves filling in intermediate data points and then smoothing
the curve. It is a technique called 'splines', and various statistics software packages
contain several techniques for doing this. There is also a 'quick 'n' dirty' way to do the trick
with a spreadsheet. Let us apply this to Roger Brown's data on the acquisition of the
present progressive (Brown, 1973).
Open a spreadsheet file and type the following series of numbers, beginning with cell
A1: 0.12,0.7,0.2,0.6,0.78,0.43,0.9,0.7,0.98,0.9,0.82,0.97.Represent thesedata points in a
 Cyclical transitory paths 43
line graph; they form a sawtooth-like pattern. Try a time-delay phase diagram: define an
xy graph with range A1 ..A12 as its x-axis and range A2..A12 as y-axis. The result is very
different from the nice and smooth phase portraits of Figures 2.9 and 2.10. The process for
making a smooth diagram is a bit tedious, unfortunately.
Go to cell 01 and type + A1. Then go to cell 621 and type +A2, to cell B41 and type
+ A3 and so on. Go to cell 62 and write the following equation:
+ 01 + ($B$21 - $B$1) 120
Copy it to the range B2..020. Go to cell 822 and type:
+ B1 + ($B$41 - $B$21) I 2 0 (2.7)
copy it to range B22..B40. Go on until all the empty spaces between the data cells have
been filled (do not forget to adapt the numbers).
Next, go to cell C1 and type:
@avg(bl ..blO) (2.8)
nd copy this to the range C2..C240. This range will contain a moving average over the
ell ranges BI..BlO, B2..B11, B3..B12 and so on.
efine a line graph for the range C1..C150: it will show the smoothed form of the line
aph based on the original numbers in range AI..A12. To arrive at a time-delay phase
ace diagram, define an xy graph w ~ t has x-axis range B1..B240, and as y-axis BIO..B240.
e result should be the moving, cyclical path shown in Figure 2.9. You can also try a
erent phase space version, based on the combination of the level and of the change in
level. In cell D2, type:

+ C2 - C1 (2.9)
d copy this to cell range D2..D240. Define an xy graph with range C2..C240 as its x-axis
nd D2..D240 as its y-axis. The result should be very similar to the time-delay phase space.

I have placed the current set of trajectories under the heading 'cycles', and from the
viewpoint of a mathematical model this is not correct. All paths described here actually
grow towards a stable state. That is, they have a point attractor, not a cyclical one. The
cycles are like intermediate paths between an initial state and a point attractor at the
final state. We could also say that they are cycles that move over an oblique axis, as in

The cycles described in the previous section are all transitory, that is, cyclical paths
that lead to a point attractor that is the final'state of development. In principle, this final
state is not expected to change or oscillate over the life span anymore. For instance,
once a child has acquired mastery of a specific grammatical rule, it will not forget the
rule or show noticeable oscillations in the grammatical errors made (real errors, that is,
based on lack of mastery or understanding).
But, should we expect that all final states of development are stable in the sense of a
point attractor? Take intelligence, which might be considered a prime example of a
44 Developmental paths and systems
stable property of intellectual functioning. Whether a person's real intelligence
fluctuates, we do not know, but whatever an IQtest actually measures, its scores are not
stable throughout life. Instead, individual scores fluctuate quite considerably (McCall,
Appelbaum and Hogarty, 1973). You could object, saying that the fluctuations are
error variations and the real score is stable. This objection, however, is based on the
assumption that something exists which we call intelligence and that it must be a stable
attribute of any single person. But of course we do not know whether this assumption
holds. The only measures that we have are IQscores and they seem to fluctuate. The
limited evidence that we have suggests that these fluctuations are far from regular and
that they certainly do not resemble the neat sinusoidal fluctuations produced by
cyclical processes.
Similar irregular oscillations apply to many other non-cognitive functions. Take for
instance self-esteem. Almost nobody feels the same about themselves all the time; the
level of self-esteem fluctuates. It is codetermined by experiences and a bad experience,
such as failing an important exam, will have a degrading effect on almost anybody. But
the effect of an experience on your self-esteem is most probably mediated by the level
of self-esteem during the experience.
In summary, we may expect many cognitive and non-cognitive properties of
individuals to change in a more or less irregular fashion over the life span. Overall
trends exist, such as major cognitive acquisitions between infancy and adolescence, but
they are not neat and smooth increases. Steady states are less steady than they should
be, according to theoretical expectations.
The stability of a person's characteristics over the life span has been investigated in
several longitudinal studies, such as the California Longitudinal Study (Eichorn et al.,
1981), the Fels Longitudinal Study (Kagan and Moss, 1962) and the New York
Longitudinal Study (Thomas and Chess, 1986), and in longitudinal studies of
personality development in adulthood, such as the Kansas City Study (Neugarten,
1964) and the Boston Study (Costa and McCrae, 1976; 1980). As could be expected,
these studies reveal a mixture of stabilities and instabilities, dependent on the
individual person, the property at issue and the age of measurement. As far as the
longitudinal studies allow us to draw conclusions on fluctuations of properties in
individuals, they show that such fluctuations are considerable and also quite irregular.
The classic response to observation of irregularity is to deny it, by reducing it to
'error', 'coincidence', 'randomness7 or whatever other denigrating concept you can
come up with. Irregularity is polished away by taking population means, averages,
prototypical patterns and so forth as images of the hypothesized simplicity underneath.
Whatever the background philosophical inspiration of this epistemological reflex, it
denies the simple fact that irregularity, diversity and randomness are very general and
stable properties of all life. The irregular fluctuations of a person's properties over the
life span are of course constrained and run within specific boundaries, but that makes
the changes no more regular.
 States and transitions

States and transifions

States and stages
Any recent overview of development begins with classical theories, such as Piaget's or
Erikson's and describes the main stages that these theories discerned. I t then usually
explains that there is currently much doubt about the existence of such broad stages
and that developmental trajectories are confined to task-specific domains. Within
domains, however, a problem would arise: Is there stagewise development or not? For
instance, is there a stagewise increase (in the form of steps) in a person's arithmetical
problem-solving skill between elementary school age and late adolescence? Some
authors, such as Kurt Fischer, claim there is, but others would probably favour a more
gradual, eventually oscillating path. Anyway, the problem of stages - whether broad or
content-specific - is an important issue in developmental psychology.
Instead of using the notion of stage, I shall employ the neutral term 'state'. Recall
that I chose to describe development in terms of a multidimensional descriptive space.
At any time, a person is characterized by some value on each of these dimensions, and
the values change with development. A person's developmental state is simply defined
as that person's position in the multidimensional space. Stage theories are character-
ized by spaces which consist of only a few descriptive, basic dimensions. The
dimensions are often dichotomous. For instance, a child's thinking is either operational
or it is not, internally represented or not (for further discussion of this view, see van
Geert, 1986; 1987a,b,c; 1988). In this case, the number of possible states a person can
occupy is highly limited. The states often last f ~a rconsiderable length of time. We may
call such states 'stages'.
Where states and stages exist, there must be transitions between stages, and
therefore transitions are characteristic forms of change wherever development occurs.
What is a transition and what are its characteristics?

Weak transitions and accelerations

First, a transition is an event that is qualitatively different from the states it connects.
Second, it connects prolonged system states that differ in one respect but are similar in
another and, finally, the transition duration should be significantly less than that of the
states it connects. I admit that this description does not appeal to intuition very much,
so let me give an example: conservation acquisition. The two prolonged states are the
not knowing and the knowing of the conservation principle. They are similar in that they
both represent a single knowledge state that lasts for several years. They are different in
that the knowledge is different. The intermediate phase of switching answers lasts for a
relatively short time and it is different in quality from the statesit connects in that it is
not stable at all, but shows unpredictable alternations. Put differently, the states show a
low to non-existent rate of change, but the event that we call the transition shows a
marked growth acceleration.
How should we proceed to find out whether there are transitions in a developmental
 Developmental paths and systems
stage 3

Figure 2.11 Weak transition criterion applied to a developmental curve based on a

mathematical simulation. The developmental curve shows a step-wise increase and a
pre-transitory dip. The transition value is given by the relationship between the lengths of
a, b, c and d. The figure at the right shows an 'ideal' transition in the form-of a sudden
jump.

phenomenon? The first thing we need is a good, sufficiently long time series of data
from one or more individuals. The data represent the change in a variable, such as
conservation knowledge or in a space of variables, such as the space consisting of
conservation, classification and seriation knowledge in a child. We begin by trying to
identify a prolonged state of any sort, starting with the initial value of the
developmental curve.
Take Figure 2.11 as an example. It shows a simulated developmental curve, for
instance of a child's progress in solving fraction problems. Intuitively, there are three
stages in the curve, corresponding with the more or less flat regions at the beginning,
the middle and the end of the curve. They are our prolonged states.
Second, we should agree on a limited 'bandwidth' for defining a state. If the
developmental curve were really flat, the bandwidth could correspond with the single
value of the variable in that flat region. But values will always fluctuate a little bit, or
rise slowly, as in Figure 2.11. The chosen bandwidth should therefore be a small
domain of scores, for instance 2 per cent of the maximum score. Any sequence of
successive values that stays within that band belongs to the first prolonged state or
stage. In Figure 2.1 1, the band has width b and length a.
Third, once we have found the first stage, we look for the second by shifting the band
with width b over the curve in an upward movement. This will lead us to the
intermediate plane, where the values of the measured variable tend to change only a
 States and transitions 47
little. By definition, the width of the state band is b. Its length is most probably
different from the first stage, say a'.
Fourth, we want to know whether the values of the measured variable between stage
1 and stage 2 constitute a real transition. These values can be covered by a rectangle
with dimensions c and d. Given the intuitive criterion for a transition - a sharp rise
between two more or less stable states - it will be clear that the ideal transition is one for
which the value of the fraction:

is as small as possible. The point, then, is to define a value T of the fraction so that it
corresponds with intuitive decisions about what would still be a transition and what
would not. Say that we set the boundary value at 0.2. The values of the transitions in
Figure 2.1 1 would be 0.1 1 and 0.15, respectively. That is, both are transitions
according to the criterion that T should be smaller than 0.2.
Let us have a closer look at the second transition. I t contains a slight regression
before the leap to the higher (and final) level. Such regressions, announcing a marked
leap, occur quite often in cognitive development. The geometric effect of such a
regression is that it enlarges the size of d and therefore decreases the value of the
fraction described above. Put differently, geometrically as well as intuitively, the
existence of a temporary regression just before a considerable leap amplifies the
transitional character of the event.
In reality, however, our longitudinal data will never look like the neat and smooth
curve from Figure 2.11. They will, rather, look like the scatter of data shown in Figure
2.12. The figure simulates a set of measurements of a child's skill in solving fraction
problems. The points could be the scores on tasks administered every two days.
First, we smooth the raw data by applying a method of moving average scores, as
explained earlier. The figure shows three such curves, based on increasing domains
over which the average has been taken.
Second, we take a smoothed data curve and apply exactly the same steps as before.
We notice, however, that the bandwidth we have to choose to obtain a reasonably long
steady state or stage is significantly larger than in the first example, which was based on
an ideal curve, not disturbed by random variation. As a consequence, the height d of
the rectangle covering the score domain between two consecutive stages will become
smaller as b grows larger.
Third, we try to find a bandwidth b such that the resulting values for a, c and d lead
to a transition fraction with minimal value. In Figure 2.12, this leads to values of 0.29
for the first transition and 0.19 for the second (which showed a transient regression that
is now concealed by the randomness of the data).
Finally, why is this form of transition called 'weak'? It is weak because it depends on
a conventional decision about how big T can be. No natural value for T emerges as a
result of empirical or theoretical investigations. Second, it is weak because the states
between which the eventual transitions occur can lie anywhere in the plane formed by
the variable and time. For instance, there is no intrinsic reason why a child's mastery in
solving fraction problems would level off at a score of 10, 20, 30, etc. per cent of a
 Developmental paths and systems
stage

Figure 2.1 2 A simulation of a scatter of measurement points shows that the stages and
transitions become more difficult to observe, but do not disappear. The 'measurement' is
based on a random transformation of the developmental curve in Figure 2.11.

fractions test and not at any other level. If we administered the test on fractions to the
whole adult population, we would probably find a distribution of stable scores covering
all possible values.
If weak transitions exist, we may assume there are also strong transitions. Do they
occur and, if so, what are their properties?

Strong transitions and catastrophes

A weak transition connects two semistable states by travelling through all the
intermediate states. A growth spurt, for instance, connects a height of 80 cm to a height
of 90 cm, but the person will of course at one time be 81 cm, then 82 cm and so on.
A large variety of systems exists where such intermediate states do not really occur,
or are even difficult to conceive. An example is the aggregation state of matter: it is
either gaseous, liquid or solid. Whatever exists between any two of those states is
something highly transient. Does anything like this exist in development? Piaget, for
one, claimed that the major stages of development represent mutually exclusive forms
of organization; there is no conceivable intermediate state between pre-operational and
operational thinking, for instance. A child is either in a pre-operational or in an
 T h e many faces of dynamic systems
erational state of mind. That state of mind can be highly context- and prob-
-dependent (although Piaget would rather see it as a general property of cognition

Recently, van der Maas and Molenaar (1992) claimed that the development of
nservation is a good example of a bimodal developmental state: the child either
*derstands the principle or does not. Sudden switches of a system state into an
rely different one have remarkably similar properties over a wide variety of
ems. They are studied mathematically in catastrophe theory. I shall call a
elopmental transition a strong transition if it amounts to a sudden shift from one
te to another and if it has a number of properties characteristic of a catastrophe (in
e mathematical sense of the word). Strong transitions and catastrophes will be
iscussed in Chapter 5.
In the preceding section, we saw that there is a large variety in the paths taken by
elopment, ranging from growth towards a stable state to chaotic oscillations and
rn uniform increase to sudden transitions in the form of catastrophes. This variety is
essence of development and it should not be seen as a superficial disguise of the real,
derlying simple Platonic forms, whatever they may be. Put differently, the often
predictable oscillations of a process of cognitive learning, for instance, are not a
ter of appearance, or of a random variation that actually conceals the real
erlying smooth increase. The irregularity can (but should not necessarily) be a
damental property of this form of learning.
Does this mean then that there is no underlying regularity in development?Far from
. The underlying regularity lies not in prototypical forms of development, such as
niform or S-shaped growth, but in the simplicity of the dynamics that generates the
credible variety of surface phenomena we observe in reality and in our dynamic
models. So far, however, I have not yet answered a fundamental question: What is a
dynamic system and how does it relate to a theory of development?

The many faces of dynamic systems

,
, Systems and dynamic systems
In the late 1960s and early 1970s, the children of academia were interested either in
Marxism-Leninism or in Eastern philosophy and religions and some (I almost hesitate
to confess I was among this crowd) managed to be interested in both. In our rooms and
dorms, we had little brass statues of Vishnu standing next to very incomplete complete
editions of Marx and Engels. The smell of incense blended with the remarkable odour
of socialist printing ink in irregular swirls that prefigured the later interest in chaos and
the baker's transformation (a very colloquial name for a class of topological operations
that describe kneading and mixing and loss of information typical of dynamic systems).
Both great groups of doctrines had some interesting commonalities. They both saw
the world as a place of becoming, of change, acted on by hidden underlying
mechanisms that revealed themselves in many different forms. But Marxism, like
50 Developmental paths and systems
many other Western philosophies, discerned a basic essence, whereas eastern religions
such as Buddhism saw the world as nothing but appearances that one could take but
also leave.
Dynamic systems have a little bit of all these different flavours and it is probably not
too coincidental that California, one of the main seats of the former hippie kingdom,
was the stage for many basic discoveries in this field. So what is a (dynamic) system?
A system is a set of variables to which an observer relates. Take any object or part of
the world. How does an observer relate to it? Many different viewpoints can be taken,
but there is no privileged access to the essence of that piece of the world, no Camino
Real to reach its heart (as Freud thought the dream was the Royal Highway to the
unconscious). What the world is, is a function of how it is observed and perceived, its
nature a matter of the channels opened to communicate with it. In this sense, all our
scientific encounters with the world we want to study and understand are by necessity
reductionistic, but calling them this does not entail a form of criticism. In fact, we
necessarily reduce the world that we encounter to our own interests and possibilities of
understanding and by doing so we create the possibility of changing that interest and
understanding, reducing our egocentricity. This is the major lesson that the
pipe-smoking Genevan private investigator of the child's mind has taught us and that
we tend to forget as soon as it applies to us academic adults.
A system, however, is more than just a collection of variables or observables we have
isolated from the rest of the world. It is a system because the variables mutually interact.
That is, each variable affects all the other variables, but also affects itself. This is a
property we may call complete connectedness and it is the default property of any system.
The principal distinctive property of a variable is that it changes over time.
Consequently, mutual interaction among variables implies that they influence and
codetermine each other's changes over time. In this sense, a system is, by definition, a
dynamic system and so we define a dynamic system as a set of variables that mutually
affect each other's changes over time.
Let a, b, c, etc., be variables describing a child's cognitive system. For instance, a
could be its logico-mathematical knowledge, b its knowledge of social relations, c its
linguistic knowledge and so forth. If the cognitive system is a dynamic system - and it
is very likely that it is - its time evolution is described by the set of coupled equations:

which means that the present state of the a, b, c, etc., system is based on the state of each
of its separately distinguished variables. The state of a variable at time t is a function of
the state of all the variables at an earlier time (t - i, t -j, etc.). In general, i,j, k, etc.,
are equal, but systems may differ in the delay between causes and effects. Each of the
state equations is an iteration and together they form a system of coupled iterations.
This brings us back to the point of the iterative function mentioned earlier; the
system evolves because, formally speaking, each system state is the input to an iterative
function that produces the next state (or a next state) as its output. What is this
 The many faces of dynamic systems 51
nction? It is the way in which each of the variables involved affects the others. The
est way to explain what a dynamic system of development is, is to construct one. In
e following example, we will travel to the friendly but unfortunately non-existent
erian city of Yeltsingrad and follow the musical development of its juvenile

Yeltsingrad
he people of the city of Yeltsingrad are modest and hardworking folk with one great
sion, singing. They love to sing at work, at home, while taking a cold shower; they
g all the time and everywhere. As you may imagine, becoming a good singer is an
portant achievement over there, not only for economic reasons - you can make a lot
money performing in the opera house or in the local pubs - but even more for the
chological gratification. More precisely, the self-esteem of the Yeltsingradians
ends heavily on the quality of their singing performances and psychotherapists are
ally more concerned with their patients' timbre than with their mental problems.
most all children go to the Academy of Music to enjoy the lessons given by great
usical performers. But deep underneath there is something tragic in this happy
ommunity. We know that self-esteem depends on the approval obtained from other
eople, especially when self-esteem is highly dependent on the quality as a performer.
lmost all the Yeltsingrad children are so busy getting enough approval from others in
n attempt to increase their self-esteem that their quest actually interferes negatively
their learning to sing and slows their progress. Besides, some children who are
y uncritical about themselves adapt their selfyesteemtoo quickly to changes in their
inging competence. For instance, with any applause they get, they become
creasingly certain they are Siberian Pavarottis, ready to conquer the world.
Empirical research carried out by the School for Education of the local university has
shown that there is considerable individual variety among the children, although the
teaching is of uniformly high quality. Could the cause be talent, or some innate
redisposition? Let us try to explain the possible developmental trajectories by
uilding a dynamic systems model and see if there is an ideal combination between
earning to sing properly and acquiring a reasonably fulfilling level of self-esteem (see
Figures 2.13 and 2.14).
Our first assumption is that performance quality improves as a consequence of
teaching and the children's efforts and that it continues to do so until a limit is reached,
dependent on the child's talent. As we shall see later, this type of performance growth
can be modelled by the logistic growth equation. But, we have seen that the growth of
the child's self-esteem, based on the child's improving singing performance, requires a
lot of extra-curricular effort and this interferes negatively with learning to sing. So we
actually decrease the performance growth by a factor dependent on the level of
self-esteem. Second, we know that the child's self-estecm depends on the quality of the
singing performance. Without singing, the self-esteem would shrink and lead the child
into a deplorable state of inferiority feelings. In fact, it requires a lot of public
/ performance to at least maintain a level of self-esteem, let alone to increase it. This
52 Developmental paths and systems

Figure 2.13 The 'psychological model' of singing development in the imaginary city of
Yeltsingrad, Siberia.

Figure 2.14 A standard graphic representation of the dynamic relations between the model
components. Arrows between the variables (self-esteem and singing performance) represent
supportive or competitive relationships. The arrows entering from and returning to the
same variable represent growth relations within the variable at issue: singing performance
increases (plus sign) by itself, but self-esteem decreases (minus sign) if it is not supported
by singing performance.
 The many faces of dynamic systems 53
articular relationship between singing performance level P and self-esteem level s
n be expressed in the form of the following set of coupled equations:
P,, = P,(1 +r - rP,/L - cS,) (2.12)
St+ = S,(l - s +
dPt)
e parameters have the following meaning: r is the rate with which the singing
rformance improves, given the particular teaching and the child's effort; L is the
rformance limit determined by the child's natural talent (it has 1 as default value); c
the factor that determines how much the performance improvement suffers from the
ild's being absorbed in getting the approval of others to increase self-esteem; s is the
ctor by which self-esteem would decrease if the child were not allowed to sing and
w personal qualities to others; d is the factor specifying how much self-esteem
eases as a consequence of an improved singing performance.

me series and state space representation

state space consists of the dimensions 'singing performance' and 'self-esteem'. It
ht be very difficult to measure these variables empirically in a reliable and valid
but for the purpose of model building we have to pretend that we simply know a
's real performance and self-esteem level (the town, the children, and the
evelopmental process are imaginary anyway, so there are no objections to pretending
e have direct knowledge of what goes on in the children's minds). Each child enters
singing class with an initial performance quality (some are quite good at the
inning, others still need much improvement) and an initial self-esteem level (some
ink very highly of themselves, others feel quite inferior). That is, for each child there
s a single point in the state space (P, S) that corresponds with his or her initial state.
et us compute a single time series for a changing performance and self-esteem level,
taking the following parameters: r = 0.03, L = 1, c = 0.03, s = 0.03, d = 0.1,
- 0.1 and Sinit = 0.4. We start with the initial values as mentioned above and
ply the set of equations over and over, until we reach a time series of 1000 data points
r each variable. The result can be seen in Figure 2.15.
We observe that both the singing performance and the self-esteem oscillate until
*

they finally settle into an equilibrium state. We can run this -little simulation with
I different initial states and with different parameter values. We will find the same
overall pattern in the form of oscillations, but with a different amplitude and frequency
and a different final point.
Instead of the time series, we can make a state space representation, plotting each
point of a performance growth sequence against the corresponding point of the
self-esteem sequence and do this for a few different time series, for instance based on
different initial states and different parameters (e.g. the d parameter). The result is
depicted in Figure 2.16. It shows semicircular trajectories, aimihg inwards towards a
stable point (which corresponds with a stable value on the P and on the S variable).
What we can tell from this state space representation is that the initial state of the
performance variable does not seem to matter too much; different initial states will end
54 Developmental paths and systems

Figure 2.15 A time series representation of the growth of singing performance and
self-esteem in a single Yeltsingradian child.

in the same final state (f2and 0,for instance, which have very different initial states i2
and i3). If one of the parameters differs, however, then trajectories starting from the
same initial state end up in very different final states ( f l and R, for i l and i2 equal).
Let us now do some further experiments with different initial state values. Suppose
we had picked a group of children whose initial state of self-esteem and of performance
are perfectly positively correlated. These are the children with realistic judgement of
themselves. Of course they will differ with regard to the level of this state and in the
group we study it will range from low (both low self-esteem and performance) to high.
Let us pretend we have a group of fifty children whose scores neatly divide the initial
state variable into fifty consecutive points and run the dynamics equation for each
initial state point. This gives us fifty final state points (by convention, I take the five
hundredth step in the computation as the final state, which is a reasonable
approximation in this particular system). These fifty points are drawn in the state
space, which means that we exclude the fifty trajectories from being represented in the
state space, and confine ourselves to the endpoints. This results in a curved line as in
Figure 2.17, showing that the relationship between initial and final states is far from
linear.
In the dynamic model discussed earlier, we hypothesized that self-esteem increases
with improving singing performance. Children will differ as to how much their
self-esteem will rise as a consequence of a socially approved rise (or fall) in their
performance. Let us repeat the experiment with the same fifty initial states, but with a
 T h e many faces of dynamic systems

Figure 2.16 A state space representation with different trajectories for different initial
states and parameters. Trajectories with different initial states (i2 and i3) lead to similar
final states (f2 and f3), whereas similar initial states lead to different final states (il and i2).
These differences are explained by different parameter values underlying the growth
processes.

different self-esteem increase parameter. We use in total five different parameter

values, in increasing order. The effect may seem paradoxical, in that the final level of
self-esteem drops as it adapts more strongly to increases or decreases in performance.
This can be seen in Figure 2.18. The lines representing the final states are projected in
lower regions of the state space as the esteem parameter increases. Since the vertical
axis is the self-esteem axis, this lowering actually means a lowering in final self- esteem,
although there remains a considerable variation, which' is represented by the fact that
the lines are curved, not straight.
In reality, Figure 2.18 is a three-dimensional representation, with the'x-axis equal to
the performance level, the y-axis the self-esteem level, and the z-axis (perpendicular to
the pages in this book) corresponding to the self-esteem incredse parameter.
56 Developmental paths and systems

Figure 2.17 The diagonal, white line represents 50 initial states, consisting of maximally
positively correlated initial values for the variables self-esteem and singing performance.
The black, curved line represents the 50 end states that result from the initial states on the
white line. The initial-final state relationship consists of a topological transformation of a
straight into a curved line.

The topological forms of dynamic systems

Our voyage to the imaginary town of Yeltsingrad and the study of the musical
development of its equally imaginary children has, I hope, brought us a bit closer to a
more direct understanding of what a dynamic system is. The whole dynamic system
explaining and describing the potential musical development of all Yeltsingradian
children is represented by the coupled equations, the two simple lines stating how
self-esteem and singing performance affect one another over time. In order for the
equation to produce time series describing possible developmental paths, a number of
variables (parameters and initial state values) have to be specified. These variables
form a multidimensional state space.
The specification of the state space together with the dynamic equation constitutes
the formal definition of the dynamic system. Such a definition, however, is hardly
informative about the possible paths and trajectories it contains. Instead of just
presenting the equation, we can also specify the basic state space - self-esteem and
performance - and draw a representative collection of possible trajectories in this
space.

The dynamic system represented by its state or phase portrait

The basic state space with some prototypical trajectories is what we call the topological
representation of the dynamic system. There are different ways of presenting such a
 The topological forms of dynamic systems 57

Figure 2.18 A representation of five sets of 50 final'states each. The sets have the same
initial state values as the set represented in Figure 2.17, but differ in the value of the
competition parameter (the competition parameter value increases from top to bottom).

topological representation. One is the space and some of its representative trajectories,
as in Figure 2.16. We call this the state portrait or phase portrait of the dynamic model.
It is truly a portrait in that we give a visual impression of the dynamics and do so by
confining ourselves to some characteristic lines, as with a drawing or caricature of a
person's face.
The model of the development of singing in Yeltsingrad, which is after all an
imaginary town on the Siberian steppes, is not a serious ,developmental model, but let
us pretend it is, just for the sake of showing how models are represented in dynamic
systems theory, as opposed to more conventional developmental model building.
Figure 2.16 is the phase portrait of the model. It is very different from the way in which
models are usually presented in developmental psychology. For instance, it does not
describe a 'typical' developmental trajectory, since there is no such thing, but specifies
the set of possible trajectories. By doing so, it explicitly reckons with the fact that
considerable individual differences exist between subjects. Each possible trajectory on
the plane is a potential individual developmental path. In the actual picture, such as
58 Developmental paths and systems
Figure 2.18, we represent only a few trajectories to serve as examples of what is possible
in the state space.

The dynamic system as a topological landscape

Instead of conceiving the state space as a straight plane with trajectories, you could also
see it as a folded object, or a landscape with a specific elevation. Imagine a satellite
equipped with nothing but a water detector; it flies over the earth and collects pictures
of streams and rivers. We could in fact reconstruct a reasonably adequate relief map of
the earth by just looking at the lines of the water. Water goes down and we simply
assume that the bigger the rivers, the lower the position as measured with regard to
sea-level. Lakes mark the presence of confined valleys, straight narrow river lines
correspond with steep flanks and so on. Put differently, the observed dynamics of the
water is transformed into a statics of the earth surface, namely its relief.
We can do a similar thing for the state space. Take Figure 2.17. It represents a curved
line consisting of final states. The initial states actually lay on a straight line, which was
considerably longer than the curve of the final states. From each point on this long
straight line we let a ball roll, simulating a specific trajectory. We observe that the ball
stops at a particular position on the final state curve (Figure 2.19a). Just as we did with
the water, we could imagine the dynamics of the balls specifying the relief of the surface.
They leave at some higher point, roll down and come to rest in the valley specified by the
final state line. This valley is also called a basin of attraction. Figure 2.19b provides a
three-dimensional impression of the basin of attraction; the line on the valley floor is the
set of potential final states of the trajectory followed by the rolling ball.
The topological transformation of a process into a landscape of imaginary hills and
valleys is not just for fun. The landscape types to which many dynamic equations can
be reduced, constitute to a small set of topological forms, such as basins and saddles,
each characterized by a specific pattern of the trajectories involved. Such basic
topological forms are well-studied mathematical objects and allow us to generalize over
superficially different dynamics models and trajectories.
The landscape form is actually quite familiar in developmental psychology, where it
is used as a metaphor. The origin of this way of representing developmental tracks lies
in the work of the geneticist Waddington, who coined the term epigenetic landscape
(Waddington, 1957). The hills and valleys of the imaginary landscape confine the
possible paths of a ball that represents a single developmental trajectory.

The dynamic system as a folding procedure

In addition to the phase portrait and the landscape form of topological representation,
there is a third way of representing a dynamic system. Take Figure 2.18, which
represents five different sets of final states in the form of curved lines. The sets come
from the same collection of initial states, but differ in one parameter, namely the
parameter specifying how much self-esteem progresses as singing improves. The
parameter values differ quite considerably, but imagine that we let the parameter
 The topological forms of dynamic systems

Figure 2.19 The transformation of a set of initial into a set of final states can be modelled
in the form of a folded surface with a valley o r basin. Figure (a) represents the
transformation in the form of a line bending and stretching procedure. Figure (b) provides
an alternative image, based on the notion of an epigenetic landscape.
 Developmental paths and systems

Figure 2.20 The plane of all possible end states of the performance/self-esteem dynamics,
for different values of the self-esteem increase parameter. The plane is a continuous
collection of sets of h a 1 states values (five separate sets were represented in Figure 2.18).

increase by an infinitesimally small step, then compute a final state line for each of the
parameter values. These lines form a plane; not a flat plane, such as a piece of
cardboard, but a bent plane, such as a sheet of paper.
The infinite collection of final state lines would have formed a plane as in Figure
2.20. It is a plane in a three-dimensional space, with as x coordinate singing
performance, as y coordinate self-esteem and as z (or depth) coordinate the self-esteem
increase parameter. Remember that for each value of the self-esteem increase
parameter, we used the same set of initial states, one where the level of the initial
singing performance corresponded with that of self-esteem (low performance, low
esteem, higher performance, higher esteem). The initial state set is therefore a straight
line in a diagonal position with regard to the x and y coordinates. Repeat this line for
each value of the self-esteem parameter and you get a flat plane with a 45" inclination.
What the dynamics equation does, as far as the topology is concerned, is actually fold or
bend the flat plane of the initial states into the curved one of the final states; the
dynamics is a prescription for surface bending and folding (see Figure 2.21).
It should be noted that the present flat plane of initial states, mapping upon the bent
plane, is just one set of initial states, namely the set defined by equal values of the
singing performance and self-esteem variables. In theory, however, the whole space of
potential initial states can be filled with such sheets (for instance, a parallel sheet lying a
tiny distance below the sheet considered here). The initial state space looks like a ream
of paper, with each sheet lying in a specific position. Each of these potential sheets
 Developmental theories and dynamic systems

Figure 2.21 The dynamics equation of the Yeltsingrad model implicitly describes a
procedure for surface bending and folding: the straight plane of initial states (left) is bent
into the curved plane of final states (right).

would map- onto a bent final state sheet. These final state sheets are wrapped into one
another, in a way specific to each of the separate initial state sheets. The whole process
is similar to taking a newspaper, with all the pages neatly ordered and wrapping it all up
in some complicated pattern. The original newspaper is the initial state space; the
wrapped paper is the final state space.
The dynamics model is represented as a particular procedure for folding an initial
into a final state space. How does this all relate to the good old, simple developmental
theories we find in our textbooks, theories that are not used to seeing themselves
compared to a ragged old newspaper?

De velopmenfa1 theories and dynamic systems

The bare necessities of developmental theory building
It took me years to realize that the song of Baloo, the bear in Walt Disney'sJungle Book,
summing up the bare necessities for living in the jungle, contained a pun. This fact
nicely illustrates that the delay between a developmental cause and its actual effect is
sometimes rather considerable, an observation that will play an important role in the
building of dynamic growth models later in this book. It also took me quite a while to
62 Developmental paths and systems
realize that all developmental theories require the same bare necessities to do their
theorizing, irrespective of what comes out of it. What are those necessities?
TO begin with, a developmental theory should specify a domain of change and
development, by describing its range and properties and the way in which such
properties may vary between people and between ages. For instance, the domain could
be acting and thinking and could involve properties such as structure, equilibrium,
formal and concrete contents, reversibility and so forth. Some of those properties can
be binary, such as the property 'formal contents of thinking' whose opposite is
'concrete contents', whereas other properties allow for continuous differences, such as
'degree of equilibration of cognitive structure', which may differ along a gradual scale.
Second, the theory should specify how those properties combine into clusters of
properties describing possible states of affairs in the world. For instance, in Piagetian
theory there are three basic binary properties (see van Geert, 1986b, 1987a and 1988 for
further justifications of this claim). A possible combination is 'actional - internal -
concrete'. It describes a form of thinking that takes place as internal imagery of a real
activity carried out with concrete contents. A combination that is not allowed is
'actional - internal - formal'. The reason is that formal contents cannot, by definition,
be entertained in an action; they require an operation. The possible combinations
specify possible developmental states. Note that state has the neutral meaning
described earlier, that is, a combination of values of variables in a specific theory. If
states last for a sufficiently long time, if there are only a few of them and if they are
irreversible, we may decide to call them 'stages'.
Third, a theory should specify how the potential states can be ordered in potential
developmental sequences. In earlier publications (van Geert, 1986a,b; 1987a,b,c; 1988;
1990; 1991), I have made a distinction between two procedures. One takes the final
state as a conceptual point of departure and views all the other states in light of the final
one. It can be shown that this particular procedure, which I called 'retrospective',
provides a fixed number and order of preceding stages. Piaget's stage theory can be
reconstructed by employing this retrospective procedure. The other procedure takes
the class of potential initial states and tries to infer the set of states potentially following
the initial one. This 'prospective' procedure is not just a logical opposite of the previous
one. It produces developmental sequence descriptions of an open type. Werner's
theory is an example.
In fact, the second and third property coincide: a sufficiently extensive definition of
the predicates or properties employed to distinguish potential developmental states
implicitly specifies the way in which these predicates or properties can be combined
into potential developmental sequences. For instance, it follows from the Piagetian
definition of the properties of advanced logical thinking that the set of properties
actional - external - concrete must be the initial state of development (if any such state
exists in reality, which is a different question). This set of properties specifies the stage
known as 'sensorimotor thinking', which is Piaget's initial stage.
Fourth, a theory should specify a set of mechanisms explaining why the developing
system changes from one state into the other. Piaget, for instance, has used the general
mechanism of adaptation and divided it into the opposite mechanisms of assimilation
 Developmental theories and dynamic systems 63
and accommodation. The mechanism explains developmental change on all levels. ~t
explains, for instance, why a baby acquires a coordinated schema of looking and
grasping, but it also explains why a child progresses from a concrete to a formal
operational mode of cognitive functioning.
Stage theories often rely on a mechanism of conflict or incompatibility to explain
hy a stage transforms itself into a different stage. Piaget has spoken about cognitive
nflict as a motor behind cognitive restructuring, but it is not the major mechanism in
his theory. Erikson has laid considerable emphasis on the notion of 'crisis', which is a
combination of internal and external conflicts, forcing the subject to make a new
commitment or a choice for a new way of functioning.
Finally, a theory should specify how the concepts it employs to distinguish among
developing people, developmental states and mechanisms are mapped onto reality. For
instance, a theory such as Piaget's allows for the possibility of a set of logical groupings
describing potential forms of cognitive organization in the concrete operational child.
Those groupings need not necessarily occur in real children, in that some are actually
found empirically, whereas others are just theoretical possibilities. Of course, the
theory should be able to explain why some theoretically possible states occur and
others do not. This may seem like a suspect form of tolerance; should not theories be
faithful to reality, describe what really is and therefore preclude everything that is not
present in the world outside?
But, the decision as to which empirical possibilities a theory should allow for is not
and should not be decided entirely on the grounds of empirical actualities. Imagine a
meteor striking the earth, and killing all people, except a small family of mentally
retarded Belgians who have never reached the level of operational thinking. Would the
meteorite have blown Piaget's theory into oblivion? It is no longer true, after all, in the
sense that there is no longer any living being capable of operational thought. But the
theory has not lost its validity (provided it was valid before the meteorite struck). For
instance, it can predict that the family, given such-and-such cultural and biological
conditions, will finally grow out into a larger population that will develop the art of
operational thought within a specified number of centuries. Put differently, for a
theory it is as important to describe the real state of affairs in the world as it is to
describe a relevant set of potential states that follow logically from its basic
assumptions.

I Developmental theories as dynamic systems models

The bare necessities of a developmental theory fall into three classes. One could be
called the statics, and refers to the potential states and sequences of states that a
developmental system can be in. Another should of course be called the dynamics of the
theory and specifies why a system moves from one state to another. There is no obvious
candidate term for the third aspect, except for a rather funny neologism. You might call
it the empirics of the theory, and it describes the mapping of the statics and dynamics
onto circumscribed empirical domains. The right word should of course be
'methodology' or something of that kind, but that does not rhyme with dynamics and
64 Developmental paths and systems
statics and that is, of course, a major scientific objection. Moreover, methodology also
refers to the method of theoretical or deductive inference of hypotheses, which is not in
itself an empirical issue but part of the dynamics.
The statics aspect of the theory corresponds with the dynamic state space in a
dynamic model. It is formed by the properties and parameters the dynamic model
employs to distinguish between potentially different states. The relationship between
the statics of a developmental model and the state space of a dynamic model is not
always one to one. The dynamic model requires its own specific parameters and those
are not necessarily distinguished by the developmental model.
The dynamics of a developmental model are of course directly related to the heart of
a dynamic model, the dynamics equations. Again, the relationship is not always simple
and straightforward. The growth models in this book basically work with simple
increments and decrements of the variables discerned, whereas developmental models
describe all sorts of mental processes and intentional activities that do not map onto the
simple pluses and minuses of the dynamic model. The connection between a
developmental model and a dynamic model is usually a matter of a rather complicated
mapping of groups of concepts in one model to groups of concepts in another.
The dynamics is an intrinsic part of any developmental theory. So are not all
developmental theories about dynamics? What is the difference between a dynamics
theory and any other developmental theory?
Actual theories emphasize a particular aspect of the entire developmental process.
This is so because actual theories are limited versions of ideal theories and they are
limited for trivial reasons, such as the lack of time to develop all the aspects of the
theory, the lack of researchers who want to spend their time in elaborating a specific
aspect or the untimely death of a major founder, such as was the case with Vygotsky's
model. As I have noted before, some theories pay considerable attention to the stages or
statics, such as the textbook versions of the classic developmental theories. Other
theories are particularly interested in describing the mechanisms that make develop-
ment work. Werner's classic book on development provides lengthy descriptions of
how the processes of differentiation and integration work in different fields and time
scales (Werner, 1948).
A dynamics theory of the kind explained in the present book is basically interested in
the dynamics, the general forms of change over developmental time. We have seen that
such changes are expressed as coupled increments and decrements in the variables the
theory deals with. A dynamics theory tries to account for the general patterns of change
by providing rather simple interaction models stated in the form of difference or
differential equations. The Yeltsingrad model provided a good example.
Consequently, a dynamics theory pays less attention to the actual psychological
mechanisms that underlie the dynamic relations between the variables involved. This
is not because those mechanisms are considered less important. It is a mere matter of
division of labour. A dynamics theory tries to capture the laws and forms of change
over time, whereas a psychological mechanisms-oriented theory tries to find out the
proximal mechanisms that make an individual change. By 'proximal' I mean the
mechanisms that take place in the form of consciously entertained activities of
 Developmental theories and dynamic systems 65
developing subjects, such as teaching, guiding a child, learning or trying to solve a
Ognitive conflict.
The split between these styles of theorizing is far from absolute, however. After all,
the theory styles result from rather trivial limitations on the human endeavour of
theory building and do not reflect deep empirical and theoretical issues. A theory
emphasizing its own particular aspect should therefore try to justify its claims by
making a link with other styles of explanation. A dynamics theory should explain which
psychological mechanisms could actually account for the general dynamic interactions
it postulates. It need not do so in very much detail, but at least part of the justification
of the theory should lie in its linking with a model of the underlying mechanisms.
Finally, the third aspect, the empirics, poses entirely different problems. Dynamic
models produce time series, state spaces and other particular representations of
processes. Testing them against the available data requires an approach that differs
considerably from the one we are used to in developmental psychology. Ordinary
developmental models usually confine to differences between age groups and to time
series based on those groups and, therefore, are not really dynamic in the true sense of
the word. Let us first see whether the dynamics approach implies any particular
message about the empirics, the methods and assumptions involved in empirically
testing such models.
Observing an.d measuring
dynamic processes

Errors, mistakes and failures of observation

The most aggressive boy at the volleyball court
Human knowledge is, almost by necessity, very incomplete. One of the reasons is that
our access to the world we would like to know and understand is limited. For instance,
if we want to understand a particular child's development, we can observe the child
only for a limited time, and we may see overt behaviour but will never get direct access
to thoughts and covert intentions.
Take for instance John Cooper. He is known to be a very aggressive boy. On the
volleyball court he swears like a retired sailor and kicks the other kids if they try to catch
a ball he thinks was meant for him. The other day he got into a serious fight with a boy
who claimed John had touched the net, which John fiercely denied. That John is
aggressive is beyond doubt, but how do we know it? We simply see it: he swears, fights
and kicks other players. John's aggressivenessis simply his tendency to swear, fight and
kick in situations where other children would probably react more mildly. That it is a
'tendency' or 'trait' of John's is clear from the fact that it is predictable; we are very
often correct when we expect John to act in this way. Of course, John's actual kicking
and swearing is only a subset and by definition a limited subset of all the swearing and
kicking that John could do, if he had more time and more occasions. This is an
ontological limitation on our observations. Second, our actual observation of John's
aggressive behaviour covers just a subset of all the aggressive behaviour that John
displays somewhere or sometimes. This is an observational limitation.
These two limitations - ontological and observational -are a characteristic feature of
aggressiveness, or any other personality trait for that matter. One could claim that our
actual observations of John's aggressiveness produce a distorted image. Since most of
John's current fellow players are peaceful and contemplative, the likelihood that John
will act aggressively is considerably lower than if he were surrounded by equally
aggressive players. But the fact that John will tend to act less aggressively if surrounded
by his current friends is a property typical of his aggressiveness, and not a matter of
 Errors, mistakes and failures of observation 67
istorted observation. Of course, if we want to predict how John will react with a
ifferent team, we should take into account the fact that his present team significantly
Utes his aggressive outbursts.
Some psychologists may think that our occasional observations of John present an
unreliable image of John's real aggressiveness. His real aggressiveness is conceived of
as a personality trait and we should have a reliable and objective measurement of it if
we are going to use it in psychological theory building. A good psychological test of
aggressiveness should do the job. Why? If it is a good test, it will not show the
contextual dependence that our casual observations do. That is, as observers we may be
fooled by the fact that John plays on a very peaceful team now, which definitely lowers
his aggressive outbursts. A good aggressiveness test is standardized, presenting the
same context to all subjects on all occasions.
Second, walking along the volleyball court and watching John play, we sometimes
witness real fights, leading to black eyes and swollen noses, whereas the next day we
will spot a completely changed John Cooper, a boy who is only mildly aggressive and
not very different from the other boys. Thus our casual observations reveal a highly
fluctuating aggressiveness, whereas the test is supposed to show a similar picture over
repeated test sessions.
Again, the point is that these observations are limited, and thus give a distorted
image of John's aggressiveness. The test, however, allegedly succeeds in overcoming
these distortions and presents a more faithful image of John's real aggressiveness. But
does it really? We know that a characteristic featuire of aggressiveness is that it is
context-dependent. We know that another characteristic feature is that people are not
always equally aggressive; they have their good days and their bad days. Thus
aggressiveness is a fluctuating rather than a stable trait. These two characteristic
features are exactly what the test conceals; it acts as if a trait like aggressiveness is
something that is deep inside the person, that it is context-independent and stable
(over a sufficiently short time, that is). It is at least remarkable that in a world of
continuous dynamic fluctuation, a test that produces a static image is believed to be
more 'true' than an observation that reveals the continuous change.
At this point we could make the inference that, if all the previous points are true,
then the whole idea of tests as ways of observing people's traits should be abandoned.
This is not what I intend to defend. What I contest is the idea that casual observations
give only distorted images, and that a technical form of observation, such as a test, is
required to reach the heart of the matter. My point is that the reality of a phenomenon
is revealed by the multitude of observations it allows and not by some 'essential'
observations in particular. Casual psychological observations are important, because
psychological reality reveals itself in casual encounters between people. They allow the
observer to see how context-dependent, fluctuating and 'fuzzy' a trait such as
aggressiveness actually is. Technical observation tools, such as tests, serve different
purposes. For instance, they allow us to predict average levels of aggressive encounters
in a tested population, even if they do not allow us to predict which individuals will act
aggressively or when they will do so. If administered repeatedly, however, most if not
all tests will show the same fluctuation over time and the same context dependence as
68 Observing and measuring dynamic processes
casual observations do. As most textbooks on psychological tests explain, it is the
stability of repeated scores that counts as a measure of the test's reliability, so
fluctuations are interpreted as an indication that the test shows a certain amount of
failure in measuring the real, underlying personality trait.
The topic of this chapter is the relationship between our deeply rooted views on the
nature of change and development and how that view relates to our assumptions about
the meaning and truth of our methodical observations of phenomena that change and
develop. In the first part of this chapter I will argue that our standard ways of
correcting for observation error are based on a model of a basically static or smoothly
and gradually changing reality. The second part of the chapter begins with a discussion
of the historical roots of our scientific and statistical methods and argues that our
emphasis on group comparisons and statistical independence stands in the way of a
thorough understanding of developmental processes. The third part of the chapter
discusses the problem of observation, measurement and error in a somewhat
non-conformist way by assuming that an observer has a completely transparent, almost
divine knowledge of the contents of a child's mind. I shall then invoke the devil - but
rest assured, it is a very friendly devil - to argue that the true, basic pattern of change
and development is irregular, context-dependent, event-based and discrete and far
from smooth, continuous and gradual. Let me begin, however, with the problem of
measurement error and stability of observations over time.

'Erarre hummanum est'

A great part of scientific methodology deals with the error-prone nature of all human
activity, and the Latin proverb in the heading assures us that making errors is human.
In fact, there are also some errors in the heading, and they are examples of what I could
call the class of simple errors. For instance, the correct spelling is not 'hummanum', but
'humanum'. The error is simple in that the conventionally agreed Latin spelling is
'humanum', so any deviation from it, such as 'hummanum', is an error. The example is
clear because the thing itself, the spelling of the word as agreed on in a linguistic com-
munity and its representation, such as its actual appearance in this book, are of the same
kind. Now, let us take a second example. Suppose Abe's IQis 168, but his score on the
Wechsler Intelligence Scale is 179. In what sense is the score incorrect, or an error?
Let us first try to find out what I mean by 'error' if the measurement tool is
isomorphic with the measured dimension. A good example is a person's height. The
yardstick shares an important aspect with the human body, a real physical extension
(just as the correct spelling was of the same kind as the wrong one, namely a series of
letters that can be compared one to one). But even in this very simple case we have to
agree on a number of restrictions. For instance, height is measured with the person
standing upright, shoes and baseball cap off. The point is that even a straightforward
and simple looking property such as height is in fact fuzzy. Is a boy's hair a part of his
height? Does his height begin at the sole of his foot, or at the tip of his big toe? But even
if we come to a full agreement on what it is that we measure, however, we may still
make measurement errors. For instance, the boy is not always standing fully upright,
we sometimes err in reading off the digits on the yardstick and so on.
 Errors, mistakes and failures of observation 69
Various techniques have been developed to deal with this sort of error and they have
been adopted in psychology too. In essence, they amount to making a number of
repeated measurements over a short time period. I t is assumed that since errors are
randomly distributed, such measurements will cluster into a normal distribution. The
average value, the top of the hill, has the greatest probability of representing the 'real'
value of the measured attribute, such as height. Provided it is at all possible to make
repeated measurements, the eventual finding of a normal distribution of
measurements is not in itself a proof of the fact that there is measurement error around
the mean value. If there is measurement error, the measurements cluster in the form of
a normal distribution. But, if there is a normal distribution, it may have been caused by
error as well as by any sort of systematic change in the real scores. But since we assume
that the measured property is stable over the course of the repeated measurements, we
interpret the normal distribution of the measurements as an indication of error.
After Arthur C. Clarke's 2001: A Space Odyssey, most readers will be perfectly
willing to accept the idea of huge black monoliths coming from outer space, sending
signals to their creators. So imagine that a Foreign Legion trooper who got separated
from the rest of his pack finds a grey monolith in the heart of the Sahara. The monolith
sends a continuous stream of sounds, and the legionary, not having completely
forgotten his scientific training, sets out to measure the monolith's dimensions,
equipped only with the most primitive means to carry out the job. Repeated measures
of the size of the monolith, using his gun as a yardstick, clearly vary. The variations are
perceived as measurement errors, due to the very poor conditions under which
measuring must take place, since the legionary assumes that the physical size of the
monolith must remain constant.
Measuring the sound level emitted by the monolith is a bit more difficult, though.
Fortunately the legionary plays the bugle in his military band, and so he is capable of
estimating the pitch. Repeated measures will again produce fluctuations, but here the
legionary assumes that, although contaminated by errors, the fluctuations are real,
since he expects the monolith to send some sort of auditory signal that is assumed to
fluctuate. Moreover, the monolith's whistle is probably not just a single frequency, but
a band of frequencies, whose composition probably changes over time.
The message behind this rather silly example is that the concept of measurement
error depends on the assumptions we make about the nature of the measured
dimensions. Are they stable attributes with single values on a well-defined scale, or do
they fluctuate? Since there is clearly more fluctuation than stability in the world, we
should expect many properties to be of the latter kind. Are psychological variables
stable over (sufficiently short) time periods? If they are not, measured instability over
time is not a sign of measurement error,'but of real fluctuations (or, of course, of a
combination of both).

Errors in psychological measurement

In the previous section, I mentioned the example of an error in establishing a person's
IQ, we gave Abe 179, whereas.he has only 168. What is the nature of our error? We
could of course have made all kinds of trivial m i s t a k ~ ssllrh g c Prrnrc in 9rld;nrn the
70 Observing and measuring dynamic processes
subscores; we could have made a typing error in our report and so on. A well-defined
procedure exists for administering the test, scoring the items, computing the total score
and so on. If we deviate from that procedure, we make an error of the simple kind,
comparable to a writing error. But this is not the kind of error we refer to in the case of
Abe's I Q If you consult a handbook on psychological measurement and methodology,
such as Nunnally's classic Introduction to Psychological Measurement (Nunnally, 1970),
you will find a different interpretation of a measurement error. It is assumed that no
obvious procedural errors have been made and that Abe actually has an IQof 168. The
fact that we have scored 179 instead of 168 is a consequence of the fallible nature of the
test, more precisely of the fact that the test gives a distorted view of the measured I Q
The central issue here is the assumption that Abe actually has an IQof 168. What could
be intended by that statement?
Let me first say that I understand by measurement the association of a number to an
empirical attribute. Unfortunately, reality is not always separated into nicely
distinguishable attributes. Measurement is a procedure for making it isomorphic to the
measure. For instance, if we measure a boy's height, we do not look at how fat he is, or
whether he has blue or brown eyes. We do ask the person to stand upright, to stretch
and press his back against the wall. By so doing we make the body isomorphic to the
vertical, stiff yardstick. The resulting specific match between the body and the
yardstick is the boy's height.
In a similar vein, the measurement of a boy's IQrequires a specific procedure. Just
as there is a widely varying vertical stretch of the human body, dependent on the boy's
posture, there is a widely varying display of intelligence, dependent on the actual
problems a boy has to solve, his current motivation and so forth. An intelligence test is
a procedure for establishing an isomorphism between this highly variable domain of
human conduct on the one hand, and an intelligence measure on the other. The
intelligence measure is based on the paradigm of one-dimensional numerical distance.
Whereas the measurement of height requires the boy to stand upright and stretch, the
measurement of intelligence requires the boy to solve a predefined and ordered set of
problems, often within a limited time span. Just as the boy's height is, by definition, the
match between the body and the yardstick, provided the isomorphic procedure has
been followed correctly, the boy's IQis by definition the match between his intellectual
activity on the one hand and the intelligence test on the other, again provided the
isomorphic procedure has been carried out properly.
Since I Q i s entirely defined by a particular testing procedure, it is meaningless to
claim that a boy has a real IQoutside any testing procedure and that a test error consists
of the deviation between the test score and the real score. In my view, test errors are
errors of procedure and are not related to some sort of ontological question. Of course,
it remains an important issue which test (of IQ, for instance) is best suited for which
sort of goals, such as the prediction of scholarly achievement. Another problem that
remains to be solved is whether the IQtest overlaps sufficiently with other situations
that are assumed to be indicative of a boy's intelligence. Should the validity of a test not
be determined by the amount it overlaps with the normal manifestations of the trait it
pretends to measure? The answer of classic test theory is yes. A test is valid insofar as it
 Errors, mistakes and failures of observation 71
the same information as the totality of natural manifestations of the measured
trait would do. Consequently, a test makes a measurement error to the extent that it
deviates from the collection of natural manifestations. The reason we use tests is that
the natural manifestations are sometimes difficult to observe, because they occur only
infrequently or are expected to take place only in the future.
Just assume, however, that a specific trait like aggressiveness or intelligence is
manifested in a way that is far from consistent over time and that this lack of
consistency is a major, intrinsic property of the trait. There is absolutely no reason why
uaits should be consistent over time and contexts. If we think about traits, we see them
as we see the size of the legionary's monolith. Why should traits not behave like the
monolith's whistle, instead, as something that, by its very nature, changes over time?
There is evidence that a property such as intelligence may-show a dramatic change
over time. The Berkeley Longitudinal Study (Honzik, MacFarlane and Allen, 1948)
showed that the IQs of about half of the children in the sample fluctuated over time,
sometimes up to thirty IQpoints. IQfluctuates in response to environmental events.
For instance, when a child is born in a family, the siblings often show a regress in 1%
lasting for several years. Scores then recover and rise even higher than in siblings who
did not have a younger brother or sister (McCall, 1983; Zajonc and Markus, 1975).
Some mental properties not only change over time, but also are multidimensional.
Intelligence is the classic example, with components such as social or verbal
intelligence. We could argue that they are just different forms of intelligence, not
dimensions of a single trait. But then consider the example of skills discussed
previously (Fischer and Pipp, 1984). A skill is two-dimensional in the sense that it is
characterized by two different types of measurement (one supportive, the other
'neutral'). These two forms of measurement cannot be considered different forms of
skill. They are different sides of the same thing.
But if human traits are fluctuating, inconsistent over time and context and
multidimensional, there is no ground for claiming that a test measures a trait reliably
and validly insofar as the measurement is consistent with the natural manifestation of
that trait. How could a single measurement be consistent with something that has
different levels dependent on time and context? Let us dig somewhat deeper in the
theory of measurement error in an attempt to answer the question of what such
psychological measurement errors really are.

Samples and signals

There are two classical sources of measurement errors. One is concerned with the
relationship between a sample and the population from which it has been drawn, the
other with a signal and superimposed noise. Questions about properties of a
population, would ideally require that each member of the population be studied. This
is impossible for merely practical reasons, such as limited research time and resources.
Instead, the investigator takes a subset of the population (a sample), studies its
properties with all its individuals and then generalizes to the entire population. For
instance, during the 1992 election day a sample of the American voters were asked
72 Observing and measuring dynamic processes
whom they had voted for, in order to make a prognosis of the final results before the
votes were officially counted. The potential measurement errors are obvious; only
rarely is the distribution of properties over the sample identical to that of the
population.
The second type of error is concerned with a signal and the superimposed noise.
Take for instance a noisy family. The television is playing and Dad, preparing a cup of
coffee in the kitchen, wants to hear the weather forecast for next weekend. But his
children, wife and neighbours, all gathered in the living room, make so much noise that
he can hardly understand whether it will rain or shine. The sources of error are obvious
in this case too. There is a circumscribed signal - the weather forecast - and a specific
procedure for receiving it: to listen and watch in a quiet room. The nature of the
problem is almost the reverse of the sampling problem. In sampling we have to expand
incomplete information; the signal problem involves the reduction of too much
information.
An impressive length of bookshelves can be filled with publications on these issues
and a considerable number of problems in the social sciences involve either the
sampling or the signal question. It is less fortunate, however, that the signals and the
sample models have proliferated far beyond their adequate applications. Let us return
to our aggressive young friend John Cooper to illustrate this point.
Our occasional observations of John's aggressive outbursts are only a subset of all his
acts of aggression and they in turn are just a small sample of what could happen. It is
easy to make the comparison with the sampling problem, therefore. The few
observations are a sample drawn from the set of observations that would give a
complete and true image of John's aggressiveness. But the comparison is flawed.
Whereas a population of real subjects has an objective extension (there are that many
persons at this time, and they have specific properties), the set of potential behaviours
has no boundaries whatsoever. Aggressive behaviour is context-dependent (in addition
to being person-dependent, of course) and whether and how a person's aggressiveness
is manifested depends on a multitude of mere coincidences.
There is no direct relationship, therefore, between the person's attribute aggressive-
ness and a circumscribed set of manifestations of it. Any reasonable subset of
observations is as good as any other one and as complete as the total set of actual
aggressive outbursts of the person. The sample model becomes relevant only if we wish
to generalize from a set of actual observations to a set of missed observations. For
instance, John plays volleyball every school-day, but since we watch him only on
Wednesday afternoon, we want to generalize those observations to Monday or
Tuesday.
Instead of following the sampling model, we could have employed the signal and
noise paradigm. The idea is that John has a fixed attribute, a level of aggressiveness.
His behaviour is a signal of that aggressiveness, but the pure signal - which would tell
us exactly how aggressive John is - is covered up by 'noise'. The noise is the effect of
coincidental factors, that is, events independent of the signal. For instance, if John
plays with placid, peaceful and highly tolerant boys, the aggressiveness signal is lower
than it should be, given John's 'real' aggressiveness. If John plays with an aggressive
 Errors, mistakes and failures of observation 73
team, it is likely that he will show more aggressive behaviour than he would normally.
The teammates are of course not causally or logically related to John's aggressiveness,
so they may be considered 'noise', deforming the pure signal. Statistical filtering
techniques should help us to get at the undeformed signal, that is, at a reasonable
of John's real aggressiveness.
The assumption underlying the model is that there exists a single source emitting a
signal independent of the factors that influence our perception of it. Aggressiveness,
however, is clearly a property not of the person alone, but of the combination of a
person and a specific context. The noise is not independent of the 'signal'. In fact, if we
&serve John's aggressive behaviour, we are receiving a pretty clear signal, as far as that
is concerned. But it is a signal about an interaction, not about a pure attribute of a single
person. There is hardly any attribute of a person that is not interactional, that is,
independent of the context in which the attribute is expressed. Moreover, attributes
are not only interactional or contextual, but also time-dependent and variable. How are
we going to characterize a psychological property of a person that does not seem to be a
property of that person at all?

Competence and performance

A classic attempt at solving many of the problems above consists of making a
distinction between a person's competence and his or her performance. It is especially
relevant in areas such as cognition and language. Let me tell you a story about the
learned Dr Kurt Harris from the University of Harford who has spent a lifetime doing
research in conservation, but has, unfortunately, never managed to publish anything.
His characteristic experimental set-up goes like this. Once he has found a child,
5-year-old Susan, for instance, who appears to understand the principle of conserva-
tion, he presents her with the standard conservation problem, with the beakers and the
water. In this particular experiment, Susan answers correctly that there is the same
amount after pouring the water into the thinner and higher container. But then Dr
Harris tells her that Roger, whom she greatly admires as a very smart boy, has been
asked the same question just before she came in. Roger, he says, just explained there is
more water after pouring, instead of the equal amount she thinks there is. Dr Harris
then shows her the experiment again, now pouring the water back into the broader,
lower container and repeats the question. After ample reflection, Susan tells him there
is less water than before. Asked why, she says that the water is lower now than it was
before, so there is less.
A week later D r Harris reports this little experiment in a class 'on cognitive
development. He explains to his students that it shows the distinction between Susan's
competence and her performance. Let me give you another example of the distinction
between competence and performance, he says. He first explains the Chomskyan
notion of competence and performance, then claims that every mature language user
has the competence to understand sentences with an arbitrary and (in principle)
infinite amount of embedding. For instance, you have the competence to understand
the sentence 'The mug that the psychologist who the teacher who the headmaster hired
 74 Observing and measuring dynamic processes
introduced used broke', but you would never use the sentence, or you would
understand it incorrectly, because your language performance is hampered by your
limited memory. The students agree, write down the sentence and note that it is an
example of competence and performance, similar to the conservation example under
different testing conditions.
The question is whether the conservation competence and the linguistic competence
are of the same kind. What the learned doctor means by Susan's conservation
competence is the 'real' level of her understanding of conservation, or her purely
cognitive representation of the conservation principle. It is assumed that her actual
answers do not necessarily reflect this real level of purely cognitive understanding.
They are contaminated by contingent and non-cognitive factors, such as memory
limitations, social contagion and so forth. The competence, in this case, is an individual
attribute, whereas the performance is context-dependent and interactional.
It will probably have become clear from the previous sections that I think the
concept of a pure personal attribute is untenable. The child's conservation under-
standing is an interactive and not an individual property; it is a property of a particular
child in a particular situation. It is characteristic of 5-year-olds that their conservation
understanding depends on social information and agreement, whereas 7-year-olds'
conservation understanding is insensitive to such factors. Put differently, there is no
empirically and theoretically meaningful notion of a competence qua psychological
, property of an individual person.
What does competence mean in the case of language as opposed to Susan's
conservation knowledge? The doctor was probably not entirely on the right track when
he told his students that they are the same. Linguistic competence is not a
psychological attribute of an individual person in the same sense that conservation
competence is. If the doctor explains that linguistic competence implies that a sentence
can have an arbitrarily large number of embeddings, he actually makes a claim about
the possibilities of sentence formation that follow from the formal properties of the
grammar. That is, the claim is about the language, not about a language user and his or
her psychological properties. Language and cognition are entities with a specific formal
structure. Similarly numbers and arithmetic operations obey a structure of formal
rules. Under the arithmetic rules, it does not matter whether we add numbers of
billions of digits or numbers of just two. It does matter for our actual addition
operations, however, whether the numbers are extremely long or reasonably
manageable. How is the step made from formal properties of a structure to a statement
about a competence qua psychological attribute?
If we observe a speaker producing and understanding sentences, making errors,
asking for clarification, etc., we can conclude that hislher language use (or
'performance') conforms to the use of a grammar, with specific rules of sentence
formation, irrespective of the errors helshe makes (which are recognized as errors
because we recognize the regularity). That is, what the English speaker's speech
defines and distinguishes from others speaking Danish or Flemish is the grammar of
the English language. It is only in this sense that we can claim that the use of English
grammar is a competence of the English speaker. Because English grammar, or any
 Errors, mistakes and failures of observation 75
other natural grammar for that matter, formally allows the formation of infinitely
embedded sentences, we conclude that it is within the competence of any English
speaker to formulate infinitely embedded sentences. But 'competence' is not
synonymous with 'ability'; that is, the conclusion says nothing about what the actual
speaker can do; it tells only what the speaker's language can do. There is no transitivity
here. It is true that the speaker possesses the grammar. It is also true that one of the
formal possibilities of the grammar is that it allows infinitely embedded sentences. But
this does not imply that the speaker is capable of forming or understanding such
sentences and that it is only by intervention of non-linguistic forces, such as memory
limitations, that the speaker is cut off from the actual possibility of exercising this
competence.
In summary, a competence is the formal definition of what a person is doing when he
or she is exercising that competence. It is not a programme of what goes on in the head
of the person, or a set of rules that generate the behaviour. For instance, when a child is
adding and subtracting numbers, he/she is exercising his/her competence of
elementary numerical arithmetic. That is, what he/she does is not just mere drawing of
lines on paper, or the performance of a trick. What he/she does is calculation and that is
defined by a set of formal rules. They specify the child's action as an instance of a
competence.
As far as measurement is concerned, the notion of competence is an essential
descriptive device, necessary to specify what it is that humans are doing when they are
calculating, thinking logically or forming grammatical sentences. But it does not
present a measurement problem in the sense that the actual observable behaviour gives
a distorted image of that competence, because it is contaminated by contingent
phenomena such as memory limitations.
The problem at stake here could better be described by a metaphor than by a lengthy
philosophical discussion. When the beautiful Narcissus looked into the water of the
pond, he saw a distorted image of his face, caused by the ripples on the water surface.
He had to reconstruct the real image of his face by compensating for the distorting
effects of the ripples. In the standard psychological view, this is about what happens
with the measurement of competence. There is a competence - as there is a face
belonging to Narcissus - and we have no direct access to it, like Narcissus, who has to
look into the water of the pond. The access we have - via behavioural observation -
distorts the image of the underlying competence and so we have to do tricks to
reconstruct the real image. But I do not believe there is anything like a competence in
this sense, that is, an individual psychological property that we see only through the
distorting reflections of transient performances.
Let us consider a different view. What if Narcissus's brother looks at ~arcissus's
face? He sees a face whose actual form is context-dependent; Narcissus laughs or
Narcissus cries. The face changes over time. It changes over a short time, for instance
when Narcissus talks, and over a long time span, for instance between Narcissus as a
baby and as an adult. What his brother sees is not a distorted image of Narcissus's real
face, which he has to subtract from the contingent movements and emotions on it.
Whenever his brother looks, he sees the real face of Narcissus, but it is a face that
 observing and measuring dynamic processes
stantly, though it nevertheless remains the face of Narcissus. The wrinkles
arcissu~'~ laughing eyes are real, dynamic, temporally changing features of
ce. They are fundamentally different from the wrinkles in the reflected face
by the ripples in the water. In a similar vein, the changes and fluctuations in
what we call a person's competence are real aspects of that person's competence (not of
the competence qua formal rule structure), and not distortions we would have to look
through in order to see the real thing. Competence, defined in the abstract formal
sense, reveals itself as a dynamic, ever-changing phenomenon as soon as it takes the
form of activities and behaviour of a real person.

Investigation as communication
Divine observation
From my Roman Catholic boyhood a number of vivid images remain. The teacher in
the first class of my primary school, a placid man with a stocky figure who probably did
not realize what damage he did to the unspoiled souls sitting in front of him in
uncomfortable wooden desks, had just finished a drawing on the blackboard that, in all
its simplicity, was a masterpiece of horror. In chalky but terrifying colours, it displayed
a high mountain with a steep and narrow road leading to Heaven, and a broad and easy
road going straight down to Hell. The choice to be made as well as the difference in
efforts to be spent to reach the heavenly goal were obvious. The teacher then explained
that in trying to reach for the heavenly goal we should not think we could hide anything
from the good Lord, since He was omnipresent and constantly observing not only what
we did but also what we thought and wanted. Needless to say, we boys were quite
impressed by that message. Until then I had neither felt nor realized that somebody
was watching me, that all my deeds and thoughts were observed and remembered, or
that my most personal phone line was constantly tapped. It was actually pretty stupid
of the good Lord to allow this teacher of ours to reveal the terrifying secret, because as
soon as we knew that He was watching we started to try to fool Him, since none of us
wanted to be on the fast lane to Hell. If He wanted to know what we did, He should
have kept to His unobtrusive, unnoticed and non-influencing style of divine
contemplation, and never have allowed our teacher to reveal the dreadful secret.
The ideal of an unobtrusive, undisturbed and direct observation reigns widely in
science, and there are many fields of investigation in which it has been approached
pretty closely. When Tycho Brahe watched the planets moving, through his primitive
telescope, his observation did not prevent the planets from continuing their eternal
ellipses. But, as the nature of the measurement or observation on the one hand, and the
nature of the observed process on the other, come closer to one another - that is, if they
tend to be of the same kind and magnitude - observation becomes a form of
communication. Observation on the level of quantum physical processes is no longer
unobtrusive, but of the same kind as the observed process and affects it in particular
ways.
 Investigation as communication 77
In psychological observation, or measurement for that matter, the observer and
observed are very much of the same kind. They differ in that one is a psychologist and
the other a child, or a student in a social psychology experiment with a stooge; they are
similar in that they both belong to the human community. The fact that psychological
investigation is almost always communication has been observed many times and has
given rise to different conclusions.
The position I would take in the debate is that psychological research is an act of
communication between psychologist and subject and that communication is a very
general, powerful form of event dynamics. The concept of communication involves a
notion of 'communing', that is, of exchange or interchange. Whereas the original
divine observation of our boys' souls was unobtrusive and did not imply an
interchange, all this changed when the teacher told us He was watching and that we had
better watch out if we wanted our final destination not to be the hot furnace of Hell. In
psychology, unobtrusive observation is almost impossible. If a psychologist adminis-
ters a test to a child, there is explicit or implicit information that the psychologist wants
the child to do its best, that this is an important issue, that he/she wants to learn
something important about the child and so forth. In return, most children will do
their best, or maybe feel very nervous about it and spoil the results. When studying a
child's language development, we cannot abstain from communicating with and
guiding the child's ongoing linguistic acquisition process. Of course, you can simply
videotape an interaction between a child and a caregiver, but the principle remains the
same, in that the caregiver plays his or her normal role and interacts with the child in an
educational way.
Another example of the communing aspect is the notorious training effect that
accompanies repeated testing. If you want to know how a child learns about the nature
of objects, you should ideally administer the same object concept test with the same
child on numerous repeated occasions. But the child will of course become used to the
test and learn how to do it. This changes the nature of the test, in that it no longer
measures the intended variable, such as object understanding, but also 'measures' the
increasing familiarity with the test procedure. This is generally seen as a disadvantage,
but it is simply an illustration of a basic fact about human beings: they learn. T o find
out how they learn, for instance about object properties, is the very aim of administer-
ing the object concept test that I gave as an example. Instead of trying to make tests
where no training effect will take place, the investigator should try to develop a logic of
testing and data processing that explicitly takes the training effect of test repetition into
account. But, as long as the basic models about observation and measurement come
I
I
I
from the classic physical sciences, where repeated measurement is not supposed to
'teach' the observed event, investigators in the social sciences will have great difficulty
I
establishing a proper way of dealing with the nature of their phenomena.
I
The conclusion that should be drawn here is simple. Psychological research is an
example of the same dynamic principles that govern the phenomena it studies. I
defined a dynamic system as a system of interacting components, changing each other
in the course of the process, based on an idea of mutual causality. In psychological
research, and developmental research in general, there is a mutuality between
78 Observing and measuring dynamic processes
researcher and subject. It lies, for instance, in the training effect of repeated testing, or
in attempts by the subjects to comply with the demands they assume in the
psychologist's research activity. It has been common practice to get rid of this dynamic
interaction by avoiding all situations in which it could come about. In the next section,
I shall look into the historical roots of this avoidance and try to make clear how it
affected our standard research methodology.

The science of change and difference

Quetelet and the science of human differences
When I defended my PhD thesis in 1975 at the University of Ghent, I was not aware
that I was continuing the tradition that had begun with Lambert Adolphe Jacques
Quetelet, who presented the first dissertation at the University in 1819. I was not
exactly standing where he must have stood so many years ago, since somebody had
made an error in booking the Aula, and my company and I were referred to a storage
room for old chairs. There I stood, behind a reading desk that bore the very
unacademic inscription 'Good Night, Chkrie', nervous but ready to defend a very
unQueteletian study on the language development of a single child.
Quetelet can be considered the founding father of the statistical study of human
properties, and it is remarkable how many of his original ideas are still alive in the
current methodology of the social sciences. Quetelet based his statistical method on
two general principles.The first is that causes are proportional to the effects they
produce. The second is that large numbers are necessary in order to reach valid and
reliable conclusions. Quetelet wanted to know the general laws of the social, and
assumed that statistical measures (the mean, in the first place) could be used to
overcome trivial differences between people and reveal the typical or common aspect.
The mean of a property, as distributed in a large population, was believed to show the
working of the 'constant causes', that is, essential causal influences that shaped the real
underlying properties of humans. Individual differences, fluctuations around the
mean, were caused by 'accidental causes'. Accidental causes must, by natural law,
fluctuate symmetrically around the mean. The essence of humans was revealed by the
means of the distributions of many properties in large populations and the construct to
which these average properties were attached was called Z'homme moyen, the 'average
man'.
Quetelet also applied these principles to what he called moral traits and what we
would now call 'psychological' traits. What a person actually does is caused by hislher
real propensity (such as the propensity to do scientific work) and a variety of accidental
factors. In his younger years, Quetelet was a gifted painter and poet, and he worked as
an apprentice in a painter's studio. He might have stayed there and then someone else
would have had to become the founder of statistics in the social sciences, but it seemed
that for Quetelet the 'accidental' painting did not stand in the way of his real
propensity. The use of large populations and the computation of the mean, Quetelet
 The science of change and difference
believed, can uncover the real propensities of people. Such propensities are related to
the essential causes, which lie in a person's sex, age or social background.
It is indeed remarkable how much of our current research methodology is still
explicitly or just tacitly following these principles. No one would endorse them in the
naked forceful form that Quetelet moulded them into, but they seem to be part of our
methodological intuition to more than an insignificant extent.
For accidental reasons, for which I should not be credited, my PhD thesis was very
different from the methodological pattern that started with Quetelet. It was a study of
the language development of one child, my son; it consisted of daily observations of a
constantly changing knowledge and contained no statistics at all. The committee of
highly learned ladies and gentlemen was eager to know how I thought I was going to
generalize my findings, since one child was indeed a very small sample. I think I then
vied to explain that I was trying to generalize about how a mechanism of language
development worked, and that information about language acquisition in groups is a
sociological and not a psychological issue. I am pretty sure the committee was not
impressed, but since my supervisor had, somewhat reluctantly, agreed to my topic and
method, my torture ended in an academic grumble.
The longitudinal study of a single child taught me more lessons than I could fully
appreciate at the time. For instance, it demonstrated that in a developmental process
causes are not proportional to their effects; the effect is dependent on the context and
time in which the cause occurs. Sometimes you correct a child over and over again, and
he will simply not hear it. And suddenly, at an unexpected moment, after he himself
has found out there might be something wrong with foots, he understands and never
makes the mistake again. Second, the use of great numbers as a way to grasp an
underlying principle is of no use if that principle keeps changing all the time, as
happens with the developing linguistic skills of children between the ages of 1 and 3.
Third, accidental causes often do not average out. The language heard by the child
contains a considerable amount of errors, repairs, incomplete sentences, etc., but they
do not result in some average error level in the child's grammar. He picks up the correct
grammar despite the lousy evidence and it is still not completely understood how
children manage to do that. Fourth, I saw that I myself was not only an observer but
also a coproducer of the phenomena I studied, although I had no idea at all about how I
contributed to that process. Finally, I learned that the study of a single case can teach
you more about a population than the actual study of the population would do. The
single case brings you to the mechanism - an example of the mechanism - whereas the
study of the population would conceal the way in which the mechanism works, because
it would require averaging over many individually different forms in which that
mechanism works.
It is true that many children who have provided the material for language studies are
unrepresentative and precocious. But once an understanding of the longitudinal,
dynamic mechanism is achieved, the population distribution can be reconstructed by
varying the parameters the individual study has revealed.
In fact, many principles of change and development (discussed in Chapter 1) do not
match Quetelet's principles. That would hardly be a serious matter if it were not for the
80 Observing and measuring dynamic processes
fact that Quetelet7s principles still underlie much of the methodology that we use to
study and understand processes of change and development.
Whereas Quetelet was interested in finding the average man, his heirs, such as
Galton, Pearson and Cattell, undertook another quest, aimed at the study of individual
differences. Much of the statistics widely used in psychology today has been developed
to support the study and description of individual variation in and between groups.
Many problems in psychology deal with such individual variation in groups, but an
equally important group of problems addresses the entirely different problem of
processes. Developmental psychology is a case in point. Its objectives are the study and
understanding of developmental processes and mechanisms; it tries to discover what
causes individual subjects to acquire a language, to develop a specific personality, to
solve identity crises and so forth. Let us try to find out how the study of groups is
different from that of processes.

Groups and processes

Whereas the study of groups rejoices in the possession of long-standing, firmly
established statistical techniques, the study of psychological processes has a far poorer
historical background. The great founding fathers of the discipline, such as Freud and
Piaget, put considerable effort into working out a methodology for the study of
processes. Unfortunately, their approaches are entombed in the category of clinical -
read 'unscientific' - methods.
Nowadays it is common practice to consider a developmental problem as a problem
of groups and individual differences within the group. The central criterion for
distinguishing groups is age. Instead of using a group of criminals or French army
conscripts, as Quetelet did, we form groups of 4-year-olds, 5-year-olds and so forth.
We then study means and standard deviations of some property in each group, and
finally tie them all together to form a wreath of years and developmental differences.
Put differently, our empirical model of development is a collection of groups and group
properties, ordered in increasing years, not a model of developmental processes and
mechanisms.
It is interesting to note that the adoption of the Piagetian universe by North
American psychological research in the early 1960sactually consisted of applying these
and comparable group methods to Piagetian stage criteria. The study of developmental
mechanisms is pushed into the same shackles. An example is the investigation of a

whether cooperation among children is an important cognitive development. Lots of

studies have been carried out. In general, the differences between groups in such
experiments are quite small, but doubt is washed away by assuring the reader that the
differences are significant.
I am not condemning this research practice and there are no doubt numerous
problems in which it is a suitable strategy. For instance, if the Dutch Minister of
Education wants to know whether it is wise to invest a considerable sum in a
programme of cooperation at the elementary school level, he had better ask somebody
to carry out a preliminary study on how much improvement in cognitive skill you can
 The science of change and difference 81
expect after a year or two of cooperation teaching. He may then decide whether an
eventual improvement of, say, 3 per cent on a scholastic achievement test is worth the
taxpayers' money.
But it remains to be seen whether this method is suitable for discovering how
cooperation works in cognitive development. We want to know, for instance, under
which circumstances cooperation works and whether it works as a strategy in some
&ildren, or for some forms of cognitive learning or just during specific learning
episodes. We would like to know whether the effect of cooperative learning is delayed
or not - that is, whether the effects can be seen immediately or require a considerable
'incubation time'. We would like to know whether cooperation causes a temporary
regression or fallback in the child's understanding, followed by a large leap, or whether
it works continually and similarly over time. All this (and probably much more) we
would want to learn about cooperation if our question is cooperation as a
developmental mechanism. But the nature of our experimental set-up prevents us
from drawing conclusions on these matters. A chief obstacle here is a notion that is
fundamental to group-based statistics: independence.

Vive 1'IndCpendance
In 1960, the former Belgian Congo became independent. As a 10-year-old I was
intrigued by the meaning of the word 'independence', which, in the colonial French,
sounded a lot more impressive than in Flemish. It 'turned out that the independent
Congo, although no longer ruled by the Belgians (who had done rather well, measured
by the standards of those times; see Gann and Duignan, 1979), was far from
independent of various international influences and pressures, ranging from the USA
to the USSR, that were imposed on her juvenile freedom.
In the statistics that we use currently in our developmental research, a comparable
emphasis is put on the importance of independence and similar doubts can be raised as
to what this actually means as far as the mutuality of influences in a developmental
system is concerned. Let us first try to figure out what independence means in common
statistical practice.
Assume that we take a sample of individuals from a population and study their
personalities. The personality characteristics of each individual are independent from
those of each other individual in the sample. This means that there is no relation of
conditionality or selectivity between any two people. Next assume that we do not take a
sample, but just pick out a single person and ask him or her to ask friends or relatives to
become members of the sample. In this case, the personalities in the sample are likely to
depend on one another, since we may expect the person to select others on the basis of
certain compatibilities of personality. Here is another example. If I measure a
characteristic such as a person's intelligence, I am likely to make a measurement error.
The error I make is believed to be independent of what I measure, or who or when. But
just assume that I am more sensitive to lower intelligences, because I understand them
better. In that case, I will probably make fewer measurement errors than with people
who have higher IQs. From early in the development of statistics, the threats of
82 Observing and measuring dynamic processes
dependence have been noticed and attempts have been undertaken to counter them.
Test situations became standardized, testers were thoroughly trained and techniques
were developed to bypass dependence in those cases where it was unavoidable. If the
experimenter expects a learning effect of one item on another, he or she will
counterbalance the order of the test items over the group of subjects, so as to make an
estimation of order-independent scores possible.
The idea buried deep under all these assumptions and methods is that the world is
made by the collision of forces that are truly independent of one another. This
metaphor is adequate for a variety of phenomena, and the study of individual
differences over groups has proven to be a good example. But for each phenomenon
where the assumption of independence holds, there is one for which it is joltingly
incorrect. One such phenomenon is development.
Take for instance the development of social-role understanding in a child. The
development follows a sequence of knowledge states (not necessarily stages), which are
clearly linked. Later knowledge states are a developmental function of earlier ones; that
is what development is all about. Genetic dependence is far from the statistical
independence of the standard paradigm. It is true that later states depend on earlier
states, but do they not depend also on environmental influences? Are these influences
not dependent on the developmental state? It is obvious they are. The way in which an
environment reacts to a child depends very much on the child's actual developmental
level. Moreover, children tend to select information on the basis of developmentally
determined sensitivities. The environment's dependence on the developmental state
does not imply, however, that it is determined by that state. There may be fluctuations
in the mutual dependence between child and environment, but as' a rule, dependence
and not independence is the default condition. When you engage in a psychological
measurement procedure, you enter into a process of communication, that is, a process
of mutual dynamic dependence.

Independence over time: expecting smooth changes

Scarcely any country in the world has as many political parties for its population size as
The Netherlands. When I moved to Groningen, I learned about the existence of a -
relatively speaking - not so small party that called itself the Anti-Revolutionary Party.
Its members had a very outspoken fear of revolutionary changes, sudden slides in the
social landscape, and irregularity in general. A regulated, smooth transition to a higher
state of well-being was what they were after. Their political ideal came very close to an
assumption that reigns widely in developmental psychology, namely that developmen-
tal change must be smooth and continuous. More precisely, the underlying, 'real'
developmental curve is assumed to be continuous and smooth in the mathematical
sense, which implies, among other things, that even stepwise changes grow smoothly out
of a stable level. Nevertheless, there is hardly any longitudinal study that does not show
irregular fluctuation over time. Take for instance the longitudinal data on the increase of
the number of meaningful units per utterance in the children investigated in Roger
Brown's classic study. On average, the length of the sentences increases, but there is a
 The science of change and difference

Figure 3.1 Empirical growth data mostly show irregularities superimposed on an overall
increase, as with these data on mean number of morphemes per utterance (Brown, 1973).

chaos-like fluctuation imposed on that increase. The same is true for other language
data, or language data from other studies (see Figure 3.1). The standard assumption
most likely to be put forward, then, would be that the real growth of sentence length is a
linear increase, or at any rate a smooth one and that the fluctuations either are caused by
accidental factors independent of the growth of sentence length, or represent errors of
observation. In this double sense, the fluctuations are considered independent of the
growth of sentence length and, therefore, not a real property of that growth process.
You could object, however, that current developmental thinking no longer falls prey
of the independence assumption. It is accepted, almost by definition, that consecutive
points in a developmental series are dependent of one another, in that later states depend
on former ones. That is, each next point of the curve represents a slight increase or
decrease in comparison with the former point. The increases or decreases may alternate
and thus produce curves that show increases and temporary dips, but the changes
remain smooth.
Against the background of the smoothness assumption, the independence adage
takes a different form. It now entails the belief that whatever deviates from the smooth
pattern must be something that is independent of the developmental process that
produces this smooth pattern. But where does our firm belief in the intrinsic
smoothness of processes come from? Most probably, it comes from the way in which
physics and mathematics have historically understood change in general. This is the
way of differential calculus. Its basic idea is that you can take a difference between two
states of a process, say the temperature of a steam engine, over two different periods in
84 Observing and measuring dynamic processes
time. you can then decrease the time difference over which the two temperature probes
are taken, until the time difference approaches the zero limit. At this point, the change
in the process is represented by a linear sloped line, the first derivative. The notion of a
derivative requires a process that is intrinsically smooth and does not show any abrupt
changes.
The idea that such infinitesimally small changes produce smooth processes over
time is deeply rooted in our scientific intuition about change in general. Put differently,
there is a deep, historically founded belief that the inherent dynamics of any process
whatsoever are of a kind that produces smooth change. It follows then that, if the
observed changes are not smooth, the causes for this lack of smoothness must lie
outside the dynamics inherent to that process and therefore be independent of that
dynamics.
In the age before computers, the formulae of calculus were meant to solve the
problem of computing a state of a process without having to go through each separate
step. For instance, in the growth equations of calculus, growth is a function of time,
and you do not need to compute every preceding step to know what the growth level
will be after a time t has elapsed. The point is, however, that not all differential
equations are analytically solvable. That is, many differential equations cannot be
solved by inserting the time parameter. They have to be solved by an approximate
calculation of each step in the process, and if the processes cover very small steps over
long time periods, the calculation becomes laborious. As a result, there has been a
historical emphasis on differentiable, that is, smooth and analytically solvable
equations. They have become the standard images of physical or biological processes.
With the advent of computers, it became relatively simple to experiment
numerically with differential equations that were not analytically solvable; the
computer would simply go through thousands or millions of successive steps and plot
the graph of the equation without any human interference. The computer would go
through these differential equations by estimating their successive values and by using
algorithms that were basically iterative. Not only differential equations, but also
difference equations could be tried out, that is, equations that are intrinsically stepwise.
As a result of all this numerical experimentation - most of which can be done on a
personal computer - we have come to a much better understanding of the nature of
processes. We have discovered, for instance, that deterministic processes can look
pretty random, as is the case with the logistic difference equation (see Chapter 4). We
have learned about non-monotonic and non-linear processes, that is, processes that can
show sudden regressions, fallbacks or jumps to different states as a result of their
inherent dynamics and not of some sort of external influence independent of that
process. We have learned that coupled equations, each describing a simple change can
lead to very unpredictable behaviour over time, if by 'predictable' we mean that they
can be linearly extrapolated. More precisely, we have learned that simple models do
not necessarily produce simple, smooth processes over time. Sudden changes and even
chaotic irregularity may occur, not as the consequence of external, independent
factors, but as a deep intrinsic property of the processes themselves.
 evelopmental curves from a divine viewpoint
n the previous sections I have argued that there is no such thing as a psychological
ttribute in the form of a competence that is a stable trait of a person and expressed only
a distorted way due to accidental factors affecting but independent of the actual
erformance. I claimed that competences are context-dependent and person-depend-
t at the same time, that they fluctuate over time, that there are several different ways
measuring them and that different measurements point to their multidimensional-
. I suggested that irregularity and even chaos can be an intrinsic property of a
process and not something imposed on that process by independent, coincidental

On the other hand, I presented a number of developmental growth curves, for

lnstance of the singing performance of the Yeltsinogradian children. They were
singular curves and often quite smooth and regular. How can the notion of a growth or
developmental curve be reconciled with the previous notion of a rather fuzzy
competence (or skill or knowledge, for that matter)? The mathematically derived
developmental curve ascribes a single point on a single dimension to a specific age and
not a fuzzy cloud of points on a variety of dimensions.
Since the good Lord, as I explained earlier, spiced my boyhood with a considerable
number of nightmares and anxieties, I could ask Him a favour now, namely to lend me
some of His unobtrusive and undisturbed observation of the human soul. And
imagine that my wish comes true. He gives me one of His supernatural videotapes of a
boy's soul covering the age range between 4 and 12 years. Well, actually it is only one
track of the tape, covering everything that relates to the child's acquisition of a
mathematical skill, addition. Not only can I look right into the soul of this child
solving addition problems, I can also replay the child's whole development of the
ability to add. Since I can look into this child's mind and mental capacities, I can
directly see whether or not he really understands a particular problem, how much
memory he uses, how interested he is, whether the correct answer he eventually gives
is just sheer luck and so forth.
While my imaginary subject is solving several addition problems, I observe that his
!
actual difficulty with those problems is caused by a variety of factors. One concerns
how many numbers there are in the addition (two, three or more), another the size of
the numbers. Those are problem-specific factors. If I observe what happens when the
child is given a problem that lies pretty close to his most'advanced level, I see that when
he really concentrates on the problem, and spends a sufficient amount of effort, he has
no difficulty dealing with it, but if he is distracted, bored or tired, he is likely to make a
mistake. Effort, attention, fatigue and so forth are person-specific factors, but they are
clearly transient and fluctuating. I also observe that part of his early difficulty with
addition problems is due to the relatively inefficient way in which he deals with his
memory limitations. This is a person-specific factor of a more stable, slowly changing
sort. Finally, a number of environmental aspects influence his problem-solving. For
86 Observing and measuring dynamic processes
instance, with some problems he was given feedback and support and other problems
were just part of a test.
The factors that influence the degree of difficulty of an actual addition are called
control variables. Each of the control variables can be thought of as a single control
dimension. An addition problem actually administered to the boy can then be
conceived of as a single point in the multidimensional space of control variables. This is
so because each particular problem has a specific problem property (e.g. three numbers
of four digits have to be added), but also hits on specific child- and environ-
ment-dependent control variables (e.g. with this particular problem the child was
highly motivated but was not given any support or feedback at all). Put differently,
each point in the control space represents a potential problem context, e.g. a problem
involving six digits solved by a very distracted but bright child in a quiet and familiar
environment. Note that by problem context I do not mean the question per se, but the act
of problem solving by a particular child at a particular place and time.
In order for me to construct a growth curve of the addition skill in the child, I have to
map the multidimensional control space onto a single variable that I shall identify as
the growth level of addition in this particular child. How do I do that (remembering
that I still have this divine observation power)? At each moment the child is in a
particular state of attention, effort retrieval, fatigue and so forth.
Assume that there are n manipulable control dimensions and that each dimension is
transformed into a scale with m different levels. For instance, the dimension 'problem
complexity' is turned into a ten-point scale. This means that my control space now
contains a total of m to the power n different problem contexts (different combinations
of control variable levels). The child, in his present state, will be able to solve a certain
percentage of this total amount of problems. Given my divine observation powers, I
can determine that percentage exactly. T o put it differently, the child passes a specific
percentage of the total number of possible problem contexts. This total number is, of
course, a single dimension (the product of all control variables). I shall call this
dimension the concatenated dimension ofproblem dzficculty, because it is a concatenation
of all the points on all control variables. The percentage problem contexts the child is
able to solve at any given moment in time is called the pass level. It applies to a specific
type of problem - addition - and to a specific subject - my nameless imaginary victim
of observation.
Now that I have this purely abstract dimension ordering all possible problem
constellations for the addition problem in my subject, it should be easy for me to
discover the growth of the addition skill in the child, as measured against this
concatenated dimension. I can do this by taking time as my second (horizontal)
dimension. For each point in time, my divine observation powers allow me to
determine the child's exact pass level. In the short run, this pass level will go up and
down, since the control variables will fluctuate. For instance, the child will go through
cycles of activity and rest, of interest and boredom, or varying degrees of effort and so
on. In the long run, the growth curve will go up, but it may show transient dips or
stationary states. That is, if I enlarge my time window from a few days to a few years, I
will lose the fine-grained details of the first observation window and find a curve that
probably looks more like the curves from the mathematical models.
 Model and reality 87
1f one is granted such mighty powers as I am at this moment, the Devil is never far
away. "If you can indeed directly observe anything that is of relevance to the child's
answering a specific addition problem," the Devil asks, "why do you bother to look to
such transient control variables as effort or interest? Why don't you take a direct look at
the essential control variable, the child's knowledge of addition? You are not going to
tell me that the child has no knowledge of addition," the Devil whispers in my ear. The
&ild has a knowledge of addition, of course, but that knowledge is the result of the
current state of all the control variables, and not a separate control variable. "But
then," the Devil goes on, "it shouldn't be too difficult to subtract the variation in the
separate control variables from the continuously sampled pass signal. For instance,
why don't you subtract the fluctuating effort level from the pass signal? That would
give you a signal that is independent of the transient effort factor."
I hesitate a little, because the thought is really tempting. My reply is that since
knowledge is not independent of effort, it would be unwise to subtract effort from the
pass signal. For instance, if the child has just mastered adding three numbers in a row,
it takes him a lot of effort to solve such a problem, whereas if the problem is of a very
familiar sort, the child comes to a correct answer with almost no effort at all. So the
relationship between effort and solving the problem says something about the nature of
the child's knowledge at that particular time. If I subtracted the effort signal from the
pass signal, I would lose the information about a distinctive property of the child's
current knowledge, namely its strong context dependency. At this point the Devil
seems satisfied with my answers and decides it is time for him to go, promising me,
however, that he will return.

Back to earth, and a long conversation with the Devil

At this point the good Lord decides that I have had time enough to fiddle around with a
propensity that should be His and His alone, namely pure, unobtrusive and complete
observation of the human mind. He takes the divine eye away from me and reinstalls
me in my poor old human condition. T o what extent do my conditions of earthly
observation present me with a wrong, error-laden, incorrect or distorted image of the
curve that I saw during the happy moments I had this divine insight? What I saw at
that time was a fluctuating and context-dependent growth trajectory, and what I see
now is also very clearly fluctuating and context-dependent. But is the 'earthly' curve
the same as (or sufficiently similar to) the one that I saw earlier? Just look at the
possibilities now. I can manipulate the problem variables, but not the person variables
that actually codetermine how difficult a particular problem situation is; In order to
estimate where this pass level actually lies for a given child on a given occasion, I would
have to present the child with lots of problems. I cannot do that in general without
driving the child, myself or the mother mad. I can only give .thg child a few addition
problems once in a while and observe what answer he is giving. What I observe is only a
rough estimate of the child's pass level. The fact that I have to estimate the real pass
level on the basis of a limited number of problems, test items or whatever, is a main
source of error, which accounts for the mismatch between the curve I observed when I
Wac ~ndnwerlwith divine observation Dowers and the curve I observe back on earth.
88 Observing and measuring dynamic processes
I should have foreseen that the Devil would be back at the most inconvenient time,
and indeed, there he is, eager to SOW the seeds of doubt in my mind. "Have you ever
realized," he asks me, "that those divine observation powers that you borrowed a while
ago, and that made you see the real underlying growth curve have actually fooled you?
The signal of pass levels you observed is a continuous signal," he says. "It is based on
your knowledge of what the child would do if you gave him an addition problem, isn't
it? But in reality you didn't give him such a problem, since, after all, you didn't have to
because you had this divine insight. But then you made a lot of fuss of knowledge being
context-dependent, didn't you? But what is the problem context if the child did not
really solve the problem? Of course, you could have seen what the child would have
done if you gave him a particular problem, but you could have given him any particular
problem with any particular level of distraction or support from your side. But then the
knowledge level is indeterminate, since you did not give the child any actual problem to
solve. So, what is the child's level of problem- or context-dependent ability when there
is no context or when he doesn't solve a problem?"
I must admit that for a while I was quite disconcerted, but then I realized that this
objection is exactly what the notion of pass level accounts for. The pass level - this
one-dimensional representation of the control space - divides the domain of possible
problems into a subset the child would pass, that is, understand and solve, and a subset
the child would fail. The subset of pass problems contains a variety of problem
situations, for instance, difficult problems with support given by an adult, or less
difficult ones the child has to solve alone. In this way, it does not matter which
problem you present to the child; any problem belongs either to the pass or fail class,
and a subset of these problems lies exactly along the boundary between passing and
failing and is therefore particularly suited for diagnosis of the child's knowledge or
skill level.
The Devil scratches his forehead and stares me in the eyes. "This seems like an
acceptable reply, young man," the Devil says, "but you make the mistake of invoking
what could have been the case in explaining what really happens. Especially if you
study development, it is not always possible to generalize from the potential to the
actual, unless you like being immersed in logical entanglements and paradoxes. Let's
recapitulate," he says. "You have been asking about the relationship between the
observed growth curve and the real one. You argued, probably on good grounds, that
the real curve is fluctuating and context-dependent. Then you invented the notion of a
concatenated dimension of control variables which enabled you to compress all those
control variables into one single dimension. Next, you invoked those divine
observation powers so that you could see a real growth curve. For any possible moment
you could see what the pass level would be if an actual problem had been administered
and the trajectory of successive virtual pass levels yielded the real growth curve, is that
right?" I reply that I myself could not have given a better summary.
"But now we have a difficulty," the Devil says. "I don't need to tell you that your
basic idea is that of an iteration, a developmental event that repeats itself and that takes
its previous output as its new input. Let me avoid the jargon and just say that an
iterative process is one in which each earlier state is the cause of each later one, do you
 Model and reality 89
agree?" I nod approbation not expecting the Devil coming with his major point. "But
the points on your allegedly real growth curve are potential points," he says. "How can
a potential point be a cause for another potential point; there is no causal relationship
between potential points. The point is," he goes on, "that any actual working on a
problem has an effect on the growth of the problem solving capacity. That is what you
call learning, isn't it? Sometimes the effect is almost nil, but sometimes it is
quite dramatic. Many of your theories account for that fact. If a child is creating a state
ofcognitive conflict right at the moment you make him solve a problem, that problem
might very much affect the growth of his insight. Or, to mention a different theoretical
angle, if the child has just created a hypothesis on how a specific form of addition
should be carried out, and you give him support while solving that problem, he might
make a real cognitive discovery. The diabolic point is, then, that the actual solving of a
p b l e m that you use as a diagnostic measure is also a potential causal factor in the
further growth of that which you measure. I admit," the Devil goes on, "that the effect
is very often small, but it is big enough for you methodologists to apply the principle of
random order of problems if you administer a test in a population, since you fear the
order and learning effects that the first problem might have on the second, don't you?"
"Have you ever heard about the butterfly effect, the Devil asks?" I say that
everybody who has read Gleick's Chaos - and that is almost anybody - has heard of that
effect. "Well, it was a rhetorical question anyway," the Devil goes on. "Just imagine
what it could mean for learning and development, where one particular problem
situation at the right time and place could have i n enormous effect on the later
development of that problem domain. I am not saying that this happens often, but it
can happen, and it probably happens a number of times in everybody's life."
"And now finally for my conclusion, young man," the Devil says. "What I intended
to show you is that a real cognitive growth curve is a succession of real discrete events,
namely the set of actual expressions of the cognitive aspect it represents. The growth
curve of addition in a particular child is the set of actual additions made by the child,
some of which were in the form of a real educational interaction, some in the form of a
test, others in the form of a spontaneous use of addition in real-life situations. It is a
discrete signal, and probably a rather scattered signal. And it also does not map neatly
onto a pure problem-defined succession of growth levels, since a problem is always
defined by the person- and environment-determined context in which it is asked.
Fortunately, the mathematical curves that your models produce are not real ones, but
sequences of discrete points; each point is a single iterative step. But then you made a
mistake to think that between each two successive points a sheer infinite number of
potential intermediary points exist, points that you thought you could see when you
borrowed the divine observation power. But you fell into a Faustian trap, my friend,"
says the Devil, "since all that was only illusion. All there is in the real curve is the
sequence of real events and in that sense it is not a curve but a-succession of discrete
points. Instead of having a time series of pass levels, you have an event series, for
instance, a series of actual manipulations of numbers or arithmetic problem-solving.
Needless to say, these event series are far less regular and smooth than the
mathematical curves of your growth models. You were tempted to see your curves as
 90 Observing and measuring dynamic processes
representations of an underlying continuous psychological variable, such as a child's
addition competence, but there is no such thing. There's only a series of conditionally
coupled actual events. I wish you good luck in your attempt to reconcile your
mathematical curves with the vagaries and discontinuities of life," the Devil concludes.
He disappears in a cloud of fuzzy numbers, and, as a tribute to Lewis Carroll's
Cheshire Cat, only his grin remains.

Real curves and measurements: the principle of qualitative invariance

and resemblance
Let me try to recapitulate the discussion so far. I started with the idea of a continuous
growth curve. In order to account for the context- and problem-dependency of that
curve, I invoked the notion of a pass level, which was a concatenation of all the control
variables that interfere in a particular problem context. I endowed myself with divine
observation powers in order to see the real growth curve of pass levels of, for instance,
addition skills. I then noticed that my limited human observation powers constituted a
source of error in that the empirically estimated pass level curve would deviate from the
I
real one. Then I introduced my personal Devil who argued that the divine curve was an
illusion, exactly the sort of Faustian thing I, a mortal, get when I think I can act beyond
my power. The real developmental curve of a skill, for instance the child's ability to
solve addition problems, is a sequence of discrete and to a considerable extent also
completely coincidental events relating to how and when that skill has been exercised
in real situations. But if this is what a real developmental curve is supposed to be, the
notion of developmental curves acquires a coincidental character, at least to a certain
extent. It is coincidental in that the occurrence of the events that constitute the curve -
actual addition problems worked on - is not legitimately related to the developmental
mechanism. Whether or not a child is given a specific problem at a specific moment is
to a certain extent determined by mere chance. The classic reply to this observation
would be to filter away the chance elements in the event sequence in order to arrive at
the real underlying growth curve. In the preceding sections I called upon not only my
own meagre authority but also that of the good Lord and the Devil to lend more
substance to the idea that there is no such thing as a real underlying competence
resulting from sieving out the vagaries and unpredictabilities of context and person.
The real curve is the coincidental, contingent, discrete sequence of events. The task of
a model builder is to try to capture characteristic properties of such real curves, for
instance their fluctuating nature, eventual regressions, the general shape of the increase
or decrease and so on.
The starting point of model building is a sequence of empirical measurements, for
instance, a longitudinal measurement of a child's addition skill, measured with a
standard test. Each test session is, by definition, an event in the sequence of events that
forms the real growth curve of addition skill in the child. Since the child has done a lot
more addition problems than the ones administered during the test sessions, the
measurement set is a small subset of the overall sequence of addition skill events. In
general, we may assume that repeated tests or measurements of a skill are only a small
 Model and reality 91
subset of the events that constitute the growth pattern of that skill. How can we
neralize or extrapolate from the limited set of measurement points to the often
set of measurable growth levels that correspond with the child's actual
involvement in addition problems? This question relates to the problem of eventual
scale invariance of growth curves.
The set of points that I achieve with a repeated measurement of a skill or
competence is, most probably, randomly dispersed. Moreover, the intervals between
measurements will be quite long, for instance a school semester. If I could observe a
group of subjects every day or every week, as is often done in early language studies, I
would have an ideal, dense and regular time series. But ideal cases are rarely observed,
as they have the disadvantage of being costly and time consuming. How can one
generalize from a sparse to a dense set of data points?
Let us work the other way around. Assume that I have a very regularly filled curve,
for instance one with daily measurement points, for a child's score on daily arithmetic
work. The score shows a typical fluctuation. I then take measurements at an interval of
a week (note that those measurements are not averages of weekly scores; they are
measures of a single arithmetic work session, but separated by a week). The kind of
fluctuation I observe is probably qualitatively similar to the fluctuation I saw over days.
On average, I see a tendency towards increasing scores,' on which fluctuations of
different magnitude are superposed. If I measure at monthly intervals, the pattern will
probably be quite similar. Of course, it is a qualitative and not a quantitative similarity,
in that, for instance, the absolute increase over months is in general higher than over
weeks or days. Qualitative similarity means that if someone shows me a 'blind' curve
based on scores obtained at weekly intervals and one with an interval of three months, I
will probably not be able to see a qualitative difference between the two curves. That is,
if the absolute scores are not given, I will not be able to tell which curve is the weekly set
and which the bimonthly.
The fact that a pattern of change and fluctuation is qualitatively invariant of the
measurement scale - days, weeks, months, years - is highly characteristic of a great
variety of natural and social phenomena. It is called scale invariance and refers to the
possible fractal nature of phenomena. By fractal I mean that the same principles are
operating irrespective of the time-scale or level of aggregation of the problem (see
Chapter 5). So far it is only an assumption that scale invariance applies to the growth of
a skill like addition. But since it is such a general phenomenon in nature, it is safe to
assume that it also occurs in development and it is at least as plausible as the
assumption that development is a gradual and smooth process.
I started the discussion by asking how we were going to extrapolate intermediate
measurement points, if only points at long intervals are actually known. This led us to
the assumption that a process such as the growth of addition competence was scale
invariant. Now, scale invariance is the key to intrapolation. Assume we had only five
measurements A, B, C, D and E separated by eight-month intervals (Figure 3.2). The
fluctuation can be circumscribed in a parallelogram along the axis A-E. In order to find
extrapolated points between A and E, we assume that the points will fluctuate over the
A-B axis in the same way as they do over the A-D axis. The parallelogram can be
92 Observing and measuring dynamic processes

Figure 3.2 A set of five repeated measurements of a variable consists of values A, B, C, D

and E (figure a). Values for the time interval between measurements A and B are
intrapolated by projecting the variation over the interval A to E onto the interval A to B
(figure b). Further intrapolations can be made by projecting any series of measurements I
onto any interval (figure c).

mapped onto a smaller one covering A-B, retaining the distribution of the points B, C I
I
and D. This breaks the line A-B into four lines, defining the extrapolated
measurements A,, A, and A,, which are separated by intervals of two months. Since
the scale invariance is qualitative, any five consecutive measurement points can be
taken and extrapolated onto any interval between two existing points. You can repeat
this process over and over, until the original eight-month intervals are reduced to
I
intervals of days or weeks.
The extrapolated curve will, of course, not be identical to the real curve based on real
daily measurements. The chance that extrapolated points will coincide with real points I
is small. But if scale invariance applies, as we assumed, the qualitative similarity
between the extrapolated and the real curve will be very high. That is, the dynamics of
change of the real curve, the nature of its fluctuations, its general outlook and so forth, I
will be captured quite reliably by the intrapolation. The real curve and the constructed,
extrapolated curve will be like children from the same parents. They differ in a number
of superficial characteristics, but on the other hand they show a lot of genetic similarity.
You could object, however, that qualitative similarity or family resemblance is a
meagre result; you want to see a reliable image of the real curve. The real curve,
however, with all its peculiarities, is a contingent and coincidental structure, in that it
depends very much on a variety of coincidences, such as the actual doing of the
I
homework or taking the test, holidays and unexpected sickness. What is important in
the real curve, therefore, is the qualitative properties it shares with the extrapolated
 Model and reality 93
curve, such as the nature of its normal fluctuation, the probability of infrequent high
fluctuations, the existence of general trends upwards or downwards and so on. The
uapolation has its time limits, however. For instance, when I continue decreasing the
e interval between any two consecutive additions, I will reach a point where the
cts of boredom and fatigue will become noticeable. Put differently, the intrapla-
tlon procedure works only within limited ranges.

Between theoretical idealization and empirical test

The great Spanish painter Velizquez owes part of his fame to the psychological and
physical realism of his many portraits. His portraits of King Philip IV of Spain,
however, are strongly idealized, and far from realistic in the sense that his other
portraits were. They are idealizations in that they hide certain properties of the thing
they represent, while making others more salient, all in an effort to communicate a
property they consider essential or typical of the represented subject. Neither
Velizquez nor his public ever assumed that the rather ugly and morose face of the King
was actually not his 'real' face and that the face Velizquez painted captured the real
face lying 'underneath' the ugly one. The function of the idealized portraits was to
show a likeness of Philip, but at the same time to show that he was the King and
therefore noble in every respect. Velizquez' first portrait of Philip IV, in the Prado
Museum in Madrid, is highly idealized. But the X-ray image shows that under the
idealization there lies a very realistic portrait, one that his commissioners did not seem
to have liked. The painters of modern nobility have found a different form of
idealization in caricatures. Good caricatures are very realistic in the sense that one can
immediately recognize whom they represent, but on that likeness they impose another
message: criticism or mockery.
The theoretical and mathematical curves based on dynamic systems models
function in ways quite similar to idealized portraits or caricatures. First, they should
show a good likeness with the empirical curves they depict. The models should try to
paint a portrait - or draw a caricature - of the real empirical data. If the portrait or
caricature is adequate, we should be able to recognize the data through the picture,
even if the picture and what it depicts do not overlap in the trivial sense.
Second, in the particular way that curves approach a resemblance to empirical data
they emphasize important qualitative properties of those data. For instance, it is likely
that the error of fit between a smooth curve and a set of mildly irregular data is smaller
than that between the empirical curve and a theoretically inferred irregular curve based
on the principle of scale invariance. The reason is that the average deviation between a
smooth and an irregularly oscillating curve is often smaller than between two
irregularly oscillating curves. If the fluctuations are a distinctive characteristic of the
empirical phenomenon at issue, however, the less well-fitting irrsgular curve should be
preferred over the better-fitting regular one. The model is a bit like a caricature: by
exaggerating characteristic traits, it may gain in recognition and 'realism' at the cost of
decreasing the literal match between the model and the mere physical appearance of
what the model depicts.
 Observing and measuring dynamic processes
e of context-dependence and intrinsic fluctuation in developmental
ight pose a serious threat to model building, even to models of the
non-linear and dynamic type. As can be observed in the examples discussed so far,
many (but not all) of the dynamic models yield smooth curves (and many more will be
seen in Chapters 4 to 9). Are such models, therefore, not missing an important dynamic
aspect of real change? They are, but in the meantime they capture other dynamic
properties of growth and development that many classical models did not capture. For
instance, they show temporary fallbacks, preceding sudden leaps to higher levels, or
they produce qualitatively different patterns based on quantitative changes in the
parameters only. Their eventually smooth and regular nature is a consequence of their
simplicity and typical of the way such theoretically inferred curves idealize that reality.
For instance, the logistic growth and transition model that will be the topic of Chapter
4 produces regular curves (as long as the major parameter stays beneath a bifurcation
value, that is). The model is therefore unrealistic in the smoothness it produces, but
what it loses on this side of the comparison is gained by its emphasis on the S-shaped
form of growth.
In summary, what mathematical curves often do is to abstract from the real
short-term fluctuations and concentrate on an aspect of the growth and developmental
processes considered typical or essential. In general, it should not be assumed that it is
the smoothness and regularity that represents the 'real' underlying growth process and
that the process we actually observe is messed up by contingent factors and errors that
prevent us from seeing it. The fluctuations in the real data are as essential as their
overall shape (although part of the fluctuation could very well be measurement error).
The fact that our models in general capture only the overall shape is a problem of the
models, not of the data. Maybe a different type of model succeeds in explaining the
irregularities, but we may encounter difficulties in constructing models that explain
both the overall, long-term developmental shapes and the short-term turbulences and
fluctuations.

1. It is assumed that I use an absolute scoring system, that is, one that reflects the increasing
complexity of the arithmetic operations the child is able to perform, increasing mastery and so
forth. Normal school marks are relative; an A in the third grade refers to a level of
understanding and mastery of arithmetics that is far below a B in the seventh grade.
4

The dynamic growth model

The growth metaphor

California Spring
In an attempt to find inspiration for this chapter, I was looking through my office
window. The Center for Advanced Study in the Behavioral Sciences where this book
was written lies on a hill and my window overlooks the Stanford University campus.
Among the views that should be intellectually inspiring, no doubt this one ranks among
the highest. But it was not this great place of scholarship that struck my attention, but
the trees and bushes on the hill slope. It was early February and that seems to mean
Spring in this part of California. The buds were swelling and little green leaves
appeared everywhere. This year seemed to be the end of a long drought and there was
abundant growth wherever I looked.
The growth I observed on a sunny February morning carried an old and powerful
metaphor of development. Growth happens all by itself, nourished by environmental
conditions such as sunshine and rain. It has a quiet, almost hesitant beginning, as when
the first green tint covers the trees, or a seedling shows its first tiny leaves. But then it
explodes in a rapid deployment of its inherent forces and the hills and gardens change
almost overnight. Finally, growth settles down into maturity. Growth also bears an
aspect of unfolding. The green bud contains the little leaf tightly folded and wrapped
in its protective cover and as the growth season starts, the leaf literally unfolds. In its
original Old French form, the word 'development' itself meant nothing but 'unfolding'
or 'unwrapping'.
Powerful as the growth metaphor is, ,it does not account for all the phenomena of
development. But it does cover a lot not only in the intuitive, casual use of the concept,
but also in many classic theories of development (Nagel, 1957). The unfolding aspect
relates especially to theories such as Piaget's, or more piecisely to the popular
interpretation of his theory. The course of developmental stages seems prefigured in
their earliest forms, and the stages unfold according to an assumed programme.'
This chapter discusses several basic aspects of the growth model of development:
 The dynamic growth model
conceptual components of the model; the simple non-linear model for growth
contained in the logistic growth equation; the relationship between actual, develop-
mental processes and this simple growth model; and, finally, the modelling of empirical
data on the basis of this extremely simple but powerful growth equation.

A definition of growth
The growth metaphor applies to poetic subjects such as Spring leaves and mimosa
blossoms, but how does it apply to psychological development, for instance to
cognitive development? The definition of (for instance, cognitive) growth I shall
present here is an extension of one given in an earlier paper (van Geert, 1991):
Growth is an autocatalytic quantitative increase in a growth variable, involving
the consumption of resources and following the emergence of a specific
structural possibility in the cognitive system.
Examples are the growth of the lexicon in a child, but also the growth of the child's
understanding of another person's point of view, the growth of skills such as solving
addition problems and so on. The definition implies a number of specifications.
The first is that growth applies to a quantifiable variable and expresses a quantitative
change. For instance, the lexicon counts an increasing number of words, or the child's
social understanding contains an increasing number of elements he/she can simulta-
neously relate to one another. The variable (number of words, elements considered in a
thought process, etc.) must be quantitative in the sense of natural numbers and this
requirement might cause problems with a lot of psychological phenomena.
The second property of growth is that it is autocatalytic. By that I mean it is a process
that sets itself in motion and whose own products keep the process going. Whatever
boils down to a mere addition of elements or magnitude by some external agent cannot
be considered a genuine growth process. If, for instance, we call the child's increasing
understanding of social roles a growth process, the product of that process - a specific
level of understanding - must by itself trigger the further discovery and elaboration of
more complicated social role understanding. The plant metaphor makes this aspect
particularly clear. When a seedling grows, it is not because water and nutrients are
pumped into the plant, as if it were a balloon that enlarges because air is blown into it.
The plant grows because its cells are dividing and multiplying, as a result of internal
mechanisms of cell growth.
The third property is that growth, even if it is internally driven, needs and consumes
resources. In plants, the resources are obvious: sunlight, nutrients, space, etc. In what
sense does an apparently ethereal thing like the learning of new words or the discovery
of a rule of social relationships consume resources? That does not seem to involve a big
physical effort. Cognitive (behavioural, mental, etc.) growth, however, relies on more
or different resources than physical effort alone. I shall explain this in a later section.
Fourth and finally, growth follows a structuralproperty of the cognitive system (or of
whatever developmental system it applies to). With the metaphor of plant growth, the
structural possibility is the presence of a seed, for instance. There is growth only if
 The growth metaphor 97
ere is something that can grow - that is, increase its size, number or magnitude. The
rowth of the lexicon implies the presence of a minimal lexicon of at least one word.
is requirement seems to disable the possibility for the construction of something
w during development. If the growth principle explains development and growth
equires the presence of the grower in at least some minimal form, then everything that
rows must be present at the beginning. Growth, however, is not the only explanatory
echanism in development. Growth is what happens after the construction of a new,
errninal form. The construction of such forms out of existing material involves its own
particular dynamic principles that will be discussed later in this chapter.

First steps towards a mathematical model of growth

The definition provided us with basic constituents and relationships that may lead
towards a mathematical formulation of the growth principle. The first constituent was
the growth variable, which at any moment in time has a specific level or magnitude, L,,
L, + ,, L, + ., etc. For instance, the levels refer to a child's understanding of emotions at
the age of 1, 2 and 3 years. It is understood that this level of understanding can be
expressed in the form of a true quantitative variable.
Since I have assumed that growth is autocatalytic and not forced upon the grower by
some external agent, we may take the relationship between the successive points in the
row of growth levels as a constitutive or characteristic property of the growth process at
issue. The relationship between successive points is expressed in the form of a ratio:

Lt/Lt+rn = Rt+m; Lt+mlLt+n = Rt+n; - - * (4.1)

The ratio R is an important parameter of the growth process, since it expresses its
internal dynamics or autocatalysis.
Next, we know that growth consumes resources. I assume that, all other circum-
stances being equal, the same amount of available resources leads to the same growth
level and that the more resources available, the higher the growth level. We may
therefore equate the growth level to a function of the resources consumed to reach that
level:

I shall use C as a symbol for resources because it refers to the idea of consumption and
to avoid confusion with the R of growth rate. If the resources are infinite, we have no
reason to introduce the resource variable into the growth model, since the specification
of the resources consumed will make no difference. But if the resources are limited,
which is most likely the case, each step in the growth process will be affected by the
amount of resources consumed in the previous steps:

Since the growth rate is supposed to express the intrinsic productivity of the grower
(that is, the autocatalytic process), R cannot be a function of the limited resources
alone, but must also depend on an intrinsic productivity factor (learning or acquisition
98 The dynamic growth model
factor in the case of cognitive development). Thus we change the previous expression
into the following:

for r the learning, acquisition or growth factor.

If we combine the equations described so far, we obtain the following expression:

L,+,, = L, + At- L, .R = L, - (1 -I- At.f (C,/Cma,, r ) ) (4.5)

If we can solve the function f (CtlC, r) to a computable expression, we have found our
mathematical model of growth. Note that equation 4.5 states that the growth level at a
+
later level t At is caused by a growth level at an earlier time t. Cause and consequence
are separated by a time interval At.

The basic growth model

Watership Down on reclaimed land
Richard Adams' famous novel Watership Down describes a bunch of heroic rabbits in
search of a new colony. Let me transfer their story to The Netherlands. The Dutch are
world famous for such cultural contributions as postcards of windmills and tulip fields
and the invention of the wooden shoe, but also for the land they reclaim from the sea.
In my story, they have just created a new island in the IJselmeer. It is still a barren
place, covered mainly with the colonizing vegetation that likes a salt soil, with no
animals to populate it except the seagulls, whose far cries intensify the loneliness of the
place. Our heroic rabbits, in search of a new colony, jump onto the local ferry, and,
after a short trip, find themselves on the shore of the island. It is not a very hospitable
place, but there is enough grass, other edible plants and plenty of space and they soon
start to reproduce. The rabbit population thrives, increasing year after year. As long as
there are not too many rabbits, there are enough edible plants to sustain them. Rabbits
reproduce quickly, but they have their natural enemies to keep their numbers within
bounds. Hawks and other predatory birds will occasionally catch a rabbit. The
migrating geese from the North forage on the same grasslands and they compete with
the rabbits for the scarce resources. A few years after the founders have migrated to the
island, the colony has reached a more or less stable population size. Births and deaths
are clearly in equilibrium, because of the limited food resources available, the
predation and the competition with the geese. But sometimes, if Spring is nice and
warm and diseases are less likely to affect the litters, many more rabbits survive than in
an average year. They survive only to discover that there is not enough food to sustain
them all and that their abundant numbers attract predator birds from all over the place.
What began as a year of prosperity ends in the disaster of a devastating death and at the
end of the year the surviving population is much smaller than it was before. But it will
slowly regenerate to its original level, if circumstances remain within the normal limits.
The story is unrealistic, I must admit, especially the part where the rabbits take the
 The basic growth model 99
ferry.2But it is realistic enough to allow for some important generalizations. First, it
shows that the resources are of various kinds: grass, other edible plants, space to dig
holes, warmth and so on. These are positive resources in the sense that they sustain the
population and contribute to its increase or maintenance. But there are also negative
resources, in the form of the predators that reduce the population and the geese, which
by competing for the grass have the same effect on the population as actual predation.
Technically we have a multidimensional resource space, some of whose dimensions are
positive, others negative. The resource dimensions are to a certain extent interchange-
able or compensatory. An increase in the amount of grass might be compensated by
more geese foraging on it, or by more predators attracted by the increased number of
rabbits. But the compensatory relations are not systematic, in that more grass will not
by definition attract more geese or more predators. Sometimes the positive resources
are more abundant than the negative ones, or vice versa.
Second, we see that although the resources are multiple and various in nature and
they cannot be covered under a single measure, they do correspond with a single
indicator. The population reaches a maximum level, at which births and deaths are in
equilibrium. This is the number of rabbits the available resources can sustain. There
are no resources left for more rabbits and if more rabbits are born than can be sustained
by the available resources, the surplus of rabbits will die and the equilibrium
population will be restored. This equilibrium is the result of all the resources, positive
as well as negative. Put differently, the multidimensional space of resources
corresponds with a one-dimensional value, the $maximal number of rabbits the
resources can sustain. This population size is called the population's equilibrium
density. The structure of dynamically related ,resources that sustains this equilibrium
population is called the carrying capacity of the environment. If the population is below
the equilibrium density level, a fraction of the available resources is still free. This
fraction is the unused capacity for growth.
Third, the growth of the population is a delayed function of the available resources.
During a good year, many more young rabbits survive. They will not only contribute to
a population increase because they mate and have their own litters, but they will also
increase the consumption of the available food resources. By the time the surplus of
young rabbits is born, the surplus of parent rabbits has eaten more than the normal
portion. The young rabbits will be born in a world that has far fewer resources to offer
than necessary to sustain their increased numbers and mass starvation will be the
result.

The formal growth model

In a previous section, I developed a series of equations aimed at specifying growth
mathematically. I ended with an equation containing a function that still had to be
solved, namely the contribution of the remaining resources. lnAthepreceding section, I
explained how the multidimensional resource space corresponded with a one-
dimensional variable, the equilibrium density, the stable population the resources can
sustain. Thus instead of having undefined multidimensional parameters C,, and C,
100 The dynamic growth model
referring to the resource space, we can now use a one-dimensional parameter, the
carrying capacity or equilibrium density of the population. In accordance with the
usage in ecology, we call this parameter K. This parameter is of the same order as the
growth level L (in fact it is the growth level at the point of stability). Let us go back to
equation 4.5:
L,,,, = L,. ( 1 + At. R ) (4.6)
It is a logical step to make the growth rate R proportional to the remaining resources,
that is, the unutilized capacity for growth. The utilized capacity is the ratio between the
current population size and the equilibrium density, that is the size at the carrying
capacity of the environment:

The unutilized capacity, therefore, is what remains:

This equation is the solution for the undefined function of remaining resources from
equation 4.5. Making R proportional to this function means that the growth level
increase over a time period At equals:

If we assume that At equals a time unit 1, the equation can be simplified as follows:

Equation 4.10 is the basic equation for logistic growth. It describes how a later growth
+
level, at time t 1, is caused by an earlier growth level, at time t, and by two additional
parameters r and K. Since t is undefined, it describes how any later growth level
depends on an earlier one. If applied iteratively, starting with t, proceeding to t 1, +
+ +
t 2, t 3, etc., the equation generates a sequence of growth levels that follow one
another and form a growth curve. The numbers 1, 2, 3, etc., added to t can have
different meanings. They refer either to a fixed time unit (a day, for instance) or to a
+
mere ordinal position. In the latter case, t would mark the first event, t 1 the second,
+
t 2 the third and so forth, without specifying the length of the time interval that
separates any two events.
A second way to infer this same equation exists and it shows how the notion of an
equilibrium density follows naturally from the concept of limited resources. The
reasoning goes as follows. Growth is not only the result of a positive term, the growth
rate, but also of a negative term. In the case of biological populations, the positive term
is birth, the negative term is death. If growth applies to psychological phenomena, the
positive term could be learning, the negative one forgetting. If growth is concerned
with the actual behavioural expression of a skill, the positive term enhances the use of
the skill in a problem situation, whereas the negative term suppresses it. Growth is
resource-dependent and the resources are limited. We do not know, however, where
this limitation lies in terms of an equilibrium density or carrying capacity. The only
thing we can say is that the negative term increases as the population or growth level
 The basic growth model 101
increases, because the bigger the population or growth level, the more resources it has
consumed to build itself up, or the more resources it needs to maintain itself. Let r be
the positive and a the negative term and At again be equal to 1. Our equation would
then be:

Under what conditions will this equation lead to an equilibrium density, that is, to a
stable population or growth level? It will be when each successive population or growth
level is equal to the preceding level, when:
= L,; or (L,+, - L,) = 0;
+
or L,(1 r - a.L,) - L, = 0
This condition is fulfilled when:
r - L , - u . L , ~= Oor when L , = r / a (4.13)
By convention, this is the equilibrium density level or carrying capacity and it is
represented by K. It follows therefore that:
r/a = K o r a = r/K (4.14)
Now substitute a for r / K in equation 4.11 and the result is equation 4.10, the equation
that started with the assumption of an upper growth level.
The growth equation has a long history. I t was developed by the Belgian
mathematician Pierre Fran~oisV e r h ~ l s t who
, ~ published his original studies in 1844
and 1847. Verhulst tried to understand the problem of population growth in the spirit
of the Malthusian view on limited resources. Malthus foresaw doomsday, but Verhulst
showed that a population would settle around an equilibrium determined by the
resources available. The Verhulst model is called the logistic growth model and the
equation is known as the logistic growth equation. The term refers to the military term
'logistics', originating from the French word 'logis', which means lodging. Logistics
are the means and infrastructure necessary to sustain the troops. The equation is
traditionally known under its dtferential form, that is, the mathematical form that
solves the equation for the time step approaching the zero limit (DeSapio, 1978). The
differential equation is:

for K the carrying capacity, k the growth rate, t time and c a constant that is equal to:

for No the starting value of the growth level N, and k the growth rate (called r in the
previous equations).
In 1976Robert May published a paper in Nature in which he explained the dzference
form of the equation (May, 1976). Whereas the differential equation involves the time
102 The dynamic growth model
variable, the difference equation simply works with discrete steps (as in equations 4.10
and 4.1 1). I t is a good example of the type of iterative equations and processes
introduced in Chapter 1. What might sound like a mere technical issue - differential or
difference - turned out to mark a fundamental distinction between two classes of
solutions. Whereas the differential form always evolves towards an equilibrium state
(the carrying capacity), the difference form showed a rich spectrum of solutions,
ranging from the simple equilibrium state to chaos. Model 4.1 explains how the
difference form of the logistic growth equation can be implemented in the form of a
spreadsheet. The logistic equation is not only a source of great fun (for whoever thinks
equations can be funny), but in all its simplicity also reveals interesting and deep
properties of natural processes.

Properties of the logistic growth equation

One model, four types of change
'I would therefore urge that people be introduced to the logistic equation early in their
mathematical education', said Robert May in his Nature paper of 1976. 'Such study
would greatly enrich the student's intuition about non-linear systems. Not only in
research but also in the everyday world of politics and economics, we would all be
better off if more people realized that simple non-linear systems do not necessarily
possess simple dynamical properties' (May, 1976).What are the properties that make a
distinguished scholar think it could teach all of us something essential about life?
Decide for yourself with the following tutorial.

Model 4.1 The logistic equation: a source of

qualitative variety and complexity
Type the name of the program in cell A l , for instance 'Model 4.1: Logistic growth
equation'.
Use cells D l ..D3 to specify the values of the main parameters and cells EI..E3 to write
down their names. In D l , type '0.1' and in E l 'r (growth rate)'; in D2 type '1' and in E2, 'K
(carrying capacity)'. Finally, 0 3 takes the initial state value '0.01' and E3 'in (initial state)'.
Define the range names for each of the parameter cells D l to 03. If you use Lotus 1-2-3 for
Windows, the command will be Range Name Createa4Other spreadsheets will have
comparable commands. The name for D l is 'r', for D2 it is 'K' and for D3 'in'. Now your
spreadsheet knows what values to insert when you use those names in your formula.
I suggest you use the cell range AlO..A110 to specify the logistic equation. Go to cell
A10, the starting cell, and type 'in'. The program knows that 'in' refers to the value of the
initial state, entered in cell D3. You will see that the number 0.01 appears in A10. Cell A1 1
contains the first logistic equation. In cell A1 1 type:

Note that the formula should be typed without spaces! Compare this with equation 4.10.
 Properties of the logistic growth equation 103
,lead of L, the equation contalns a reference to the previous cell, A10 The $ slgn
reced~ng the parameter names IS necessary to let the program know that ~tshould keep
reference to r a n d K constant. Copy the content of cell A1 1 to the range A1 2..AI 10. If
use Lotus 1-2-3 for W~ndows,I suggest you use the Lotus Classic Menu to do this klnd
f copying (it IS fast and slmple, especially if you have to copy to big ranges). You will
otlce that each cell contalns a number, w h ~ c hIS the value of the equation In that cell. To
e the equation, for instance, in cell A20, put the cursor on A20 and look at the editlng
indow below the Menu bar. Another way to see the equation is to double-click on the cell
ontent. This will bring you ~ n t othe editing mode, and you can edit the equat~onif you
ant to. You will notice that the equation in cell A20 contains a reference to cell A19. Cell
20 represents the state of the growth level after ten iterative appl~catlonsof the equation.
so far you have seen only a range of numbers. Go to the Chart menu and deflne a line
raph for the data range AIO..AI 10 (different spreadsheets use thelr own ways to def~ne
graphs; if you are not famil~arwith maklng charts, consult your manual). Go back to the
computing window, change the value of the parameters and watch the different graphs.
Try any growth rate value between 0 and 3 (e.g. 0.1, 0.5, 1, 1.9, 2.4, 2.7). Pay special
attention to the threshold values described further in the text.

If you actually performed the exercise above, you have just witnessed the remarkable
non-linear behaviour of an equation whose simplicity is in no way related to the
incredible complexity of its potential outcomes.
For r < 1, the resulting curve has a characteristic S-shape. Its asymptote, the point
it touches at infinity, is the carrying capacity level (equal to 1). I t is called asymptotic
growth. Put differently, the value of the equation, or more precisely the growth level, is
attracted towards a single point, the carrying capacity value. That is why the process is
said to have a point attractor for values of r smaller than 1. This is the scenario for a
rabbit population that is constant from year to year.
For r larger than 1 but smaller than 2, the curve shoots above its carrying capacity,
then drops, increases again, drops, in a decreasing vibration. T h e larger r, the longer it
takes for the process to approach the carrying capacity. This type of growth can be
called approximate growth. I t oscillates until it finally hits its point attractor, the
carrying capacity again. If the growth rate r is between 2 and 2.57, a new pattern
emerges. The growth level fluctuates between a series of states lying above and below
the carrying capacity, but never reaches that carrying capacity. For instance, with
r = 2.2 there are two such states and with r = 2.5 there are four. As r approaches
nearer and nearer 2.57, the number of states visited by the curve doubles more and
more rapidly. Because the curve cycles between 2,4,8, etc., states, we call this growth
orrillatory. Its attractor is not a single point, but a cycle, so it has a cyclical attractor.
Both approximate and oscillatory growth correspond with the scenario in which the
rabbit populations fluctuate between a too small population one year and a too large
population the next.
Above the rate value of 2.57, the process loses its periodicity and moves into chaos. It
is now chaotic growth and has a chaotic attractor. T h e word 'chaos' is actually quite
misleading. T h e chaos reigns only at the surface, so to speak, in that the succession of
growth levels seems truly random. In fact, this equation is sometimes used as a random
104 The dynamic growth model

Figure 4.1 Four qualitatively different types of growth based on different growth rates r
(corresponding values to the left).

number generator. On the other hand, this apparent chaos is caused by a simple and
perfectly deterministic equation. Why do we call this chaotic succession 'growth'?
Because it is a succession produced by a model we agreed to call a growth model, and
which effectively produced canonical growth for most of the growth rate values (see
Figure 4.1).
Above a growth rate of 3, the equation produces negative values. That is the point
where the practical application of the model ends, since there are no negative numbers
of subjects, or negative levels of a skill or ability.

Non-linearity in logistic growth

A major property of the equation is its non-linearity. By this I mean that effects do not
simply add up. For instance, if I increase a growth rate of 0.1 by 0.5, and obtain a
growth rate of 0.6, the result is a marked increase in the speed with which the S-shaped
curve reaches its upper level, but the end result is similar; a process approaching its
asymptote. If I add 0.5 to a growth rate of 1.8, I cross a qualitative boundary, between
approximate growth with a point attractor, and cyclical growth with a cyclical
attractor. Non-linearity thus refers to the fact that the result of an operation (adding
the value of 0.5) critically depends on where along a linear variable (growth rate) it has
been applied. The non-linearity of the equation is most explicitly expressed in the fact
that it yields four qualitatively different types of growth. This brings us to a further
property of the growth equation, its bifurcation behaviour.
By bifurcation I mean a sudden shift in the nature of the outcome of the equation as a
quantitative threshold is crossed. We have seen that if the growth rate crosses the
 Properties of the logistic growth equation 105
of 2, it suddenly jumps from approximate to cyclical growth. In the case of
[he logistic growth equation, the bifurcation is literal. It is usually shown in the form of
_ a bifurcation diagram and Model 4.2 will show how such a diagram can be constructed.

Model 4.2 The bifurcation diagram

Start with your first model, but change the title to 'Model 4.2; Bifurcation diagram'. Add a
parameter called 'r-max' to the list: type its name 'r-max (maximal r)' in cell E4, and its
value, '3' in cell D4. Define the range name 'r-max' for cell D4. Before continuing, set the
method of recalculation to 'manual' (consult your spreadsheet manual if you don't know
how to do this).
In Model 4.1, you defined a column between A10 and A1 10, which contains a row of
equations, starting at A10 with the initial value. Recall that you referred to the growth rate
as '$r'. In this model we shall define 150 growth curves, with different growth rates. The
growth rate of each column of equations will be put in cell 9 (for instance, the growth rate
of the column H10..H110 is in cell H9). Go to cell A9 and type '$rl. Go to cell A l l and
change the contents of the formula as follows:

You will notice that the divider '$Kt has been omitted; this is the same as setting it to the
default value 1, which simplifies the equation. Go to EditJClearSpecial (or any comparable
command in your spreadsheet) and clear the range A12..A110. Next copy the equation
occurring in cell A l l to cells A12..A50. Copy the column from A10 to A50 to the range
BIO..EU50. This will result in 150 columns that take column cell number 9 as their growth
rate. If you run out of memory, try a smaller number of columns.
What we shall do now is give each of the 150 columns of equations a different growth
rate. The growth rate of each column should be a little bit bigger than the rate of its
predecessor (the column to its left). To obtain this result, go to cell B9 (where the growth
rate of the B column is). Type:

(the divisor sl~ouldbe equal to the number of columns you have defined).

Next copy cell B9 to cell range C9..EU9. This will cause the next growth rate to be a bit
higher than its left neighbour and at the end of the row of columns it will have reached the
value 'maximal growth rate' (never make it higher than 8). Set the level of the initial state
high enough, say 0.9.
What you have now are 150 series of growth levels (or less). Their initial state is similar,
namely 0.9. Their growth rate, however, increaies from left to right. Set the value of r(the
growth rate) to 1.8, and the maximal growth rate to 3. Then. press the button for
Recalculation (which is F9 in Lotus).
In designing a bifurcation diagram, we are interested only in the later part of each
growth curve, after the curve has found its steady state, if one exists. The bifurcation
graph is created as follows. Make a new graph, call it 'bifurcation', and define its A
 The dynamic growth model

Figure 4.2 A bifurcation diagram for growth rate values ranging from 2.8 to 2.9. T h e
growth rates are projected onto the horizontal axis. T h e vertical axis specifies the values of
the growth levels corresponding with the growth rates on the horizontal axis; a growth rate
of 2.83, for instance, produces a growth pattern consisting of two alternating levels L1 and
L2.

datarange as 'A30..EU50'.Thus the range covers the 150 growth curves, excluding the first
20 or so steps and presenting only the steady-state outcomes. Make the graph an
ordinary line graph, and set the method of display to symbols only (not to connectors or to
connectors and symbols). Along the x-axis you will find the series of increasing growth
rates, while they- axis will display the steady-state values corresponding with each growth
rate. The result is a beautiful tree structure known as the bifurcation diagram (your screen
should look like that of Figure 4.2; lines and rectangles have been added afterwards).
The bifurcations are literal, in that the process has a single end state until the growth
rate value crosses the boundary value of 2. The single branch of end states splits into a
fork, which again splits at the value 2.57. The higher the growth rate, the more rapidly the
branches bifurcate. A linear increase in growth rate thus leads to a non-linear jump in the
number of end states. It is not entirely true, however, that an increasing growth rate
corresponds with an increasing number of end states. In the chaos of end states, islands
of simplicity occur, as for instance around the growth rate value 2.83. If you set the growth
rate to 2.8 and the maximal growth rate to 2.9, your bifurcation diagram shows a
magnification of the window confined by the values 2.8 and 2.9. The increasing chaos of
end states suddenly switches to a simple three-state cycle, which then bifurcates again
and flows into chaos. In fact, the bifurcation diagram is self-similar, or scale-invariant,
because copies of the diagram reappear at smaller and smaller scales. For more details
(on a reasonably accessible level), see Peitgen, Jiirgens and Saupe, 1992 (Part 11).

The bifurcation diagram provides a particularly good illustration of the principle of

non-linearity that governs an equation as simple as the logistic one. I t makes clear how
 Properties of growth in mental and behavioural development 107
inear increase in a control parameter (the growth rate) corresponds with sudden
fts in the qualitative outcome of the process (number of end states). Another aspect
inearity of the growth equation is its sensitivity to small initial state
ifferences. It is an illustration of the fact that small causes can have big consequences,
in that small differences at the beginning can be magnified afterwards. The
on-linearity extends even to the domain of applicability of this principle, however,
ince it holds within the range of chaotic growth, but not within the range of growth
int attractor. In the latter case, even major differences lead to similar
utcomes. The next tutorial gives an example of how the sensitivity can be tested.

ode1 4.3 Sensitivity to small initial state differences

Start with your original model (Model 4.1) and copy the range A10..A110 to the range
B10..6110. Define the B datarange in your graph as B10..B110. Set the value of r t o 2.7. In
+
cell 810, type ' ~ n 0.0001'. This means that the initial state of the B column is only a tiny
fractlon bigger than that of the A column. Make a line graph for two data ranges, namely
A10..A110 and B10..B110. Watch the graph. What you observe is that the two graphs stay
v~rtuallyslmllar for the first thirty-two points of the iteration. But then the small initial state
difference gets magnified into a major difference in outcomes (see Figure 7.5).

What you witness here is the butterfly effect in action; a tiny difference does not show
up until it suddenly gets magnified and leads to a pattern entirely different from the

Properties of growth in mental and

beha vioural development
Limited resources for mental growth
Mental growth (cognitive, behavioural, etc.) is as dependent on resources as biological
or bodily growth. For the sake of arranging an overview of the different resources,
it could be convenient to make a distinction between two descriptive dimensions.
The first concerns the origin of resources and distinguishes between internal (within
the subject) and external (outside the subject) ones; tbe second dimension deals with
the nature of the resources, namely spatio-temporal,
. . informational, energetic-
motivational and material resources.
The concept of internal spatial resources refers to the limited amount of information
we can deal with simultaneously. What has been called our working memory has a
limited extension, estimated, since Miller's famous paper, to be around 7 units
(whatever 'units' may mean, however; Baddeley, 1983; Kahneman, 1973; Miller,
1956):~hesize of this mental capacity increases with age, either in the form of a literal
increase, or in the form of increasing efficiencyof information processing (Case, 1992a;
 The dynamic growth model
Case ,ot a / s , 1986; Globerson, 1983; Pascual Leone, 1970). It is interesting to note that
the growth of memory span itself follows a pattern that can probably be modelled by
the logistic growth equation (Case, 1992a). Internal temporal resources refer to the
time on task we are able or willing to invest in a specific cognitive activity, relative to the
number of different cognitive activities carried out during a specific period. Internal
informational resources consist of the knowledge and skills already present in the
subject, which act as the internal learning or acquisition context for new skills and
knowledge and which may either ease or impede the acquisition of specific new
knowledge or skills. Internal motivational/energetic resources consist of the amount of
energy, arousal, effort and activation invested in specific acquisition activities (e.g.
Sanders, 1983). Fischer and Lazerson (1984) argue that some significant changes in IQ
over the life span may well be explained by changes in motivational structure, resulting
in a different allocation of resources. If energetic investment is defined as a
content-specific variable, we may call it motivation (see for instance Leontew's theory
in which motivation plays an important developmental role; Leontew, 1973). Internal
material resources amount to the bodily outfit of a developing subject, for instance the
availability of correctly working sensory and nervous systems. A very important
material resource for mental growth is of course the brain. Ample evidence now shows
that brain development occurs in stages (Thatcher, Walker and Giudice, 1987). Spurts
in brain development correlate with behavioural and developmental transitions
(Fischer and Rose, 1993).
External spatio-temporal resources are the spatial and temporal degrees of freedom
given to developing or learning subjects by their controlling environment. Caretakers
and educators explicitly restrict the free-moving spaceltime of children, with the often
implicit intention to structure this limited space in an optimally profitable way for the
child. This principle is also inspired by the educator's need for resource economy in the
environment. Valsiner (1987) has described this principle as the 'Zone of Free
Movement'.
External informational resources primarily amount to the number, availability and
form of the items that could be assimilated by the developing and learning subject (e.g.
the lexicon presented by the speaking environment, or the specific ways in which the
teaching environment makes information available to the learner). The third form of
external resources is the energetic/motivational, for instance, task-specific 'pay offs',
such as the reinforcement provided by the environment for specific activities of the
learner. External material resources are things like food and shelter, or objects such as
books and writing paper.
The availability, nature and relationships of all these resources differ greatly among
individuals and groups and also within individuals (e.g. temporal variations in the
information given to a child, or in the nature and amount of the energetic resources
invested).
However variable the resources may be, they are always limited, although this may
not seem true on so local a scale. For instance, if a child is really interested in some
subject - such as dinosaurs - he or she may easily increase the time and effort spent in
reading or watching television programmes on dinosaurs, go to the library and so forth.
 Properties of growth in mental and behavioural development 109
reallocation of resources, since the time and effort spent on
ent on other learning activities, notwithstanding the fact that in
rning about dinosaurs the child may also improve his or her reading skill or scientific

discrete growth and the principle of feedback

Intuitively, growth seems to be a continuous process. In Chapter 3, however, I argued

that growth curves are not continuous, but discrete. They follow the discreteness of the
events and contexts to which their expression is confined. The fact that the probability
of occurrence of a specific action or behaviour -such as a correct answer to a problem -
has increased from one act of problem-solving to the next, does not necessarily imply
that the line connecting these occurrence probabilities represents an intermediate
continuous growth process. The growth process that gave rise to the increased
probability of giving a correct answer, for instance, could as well have taken place
during the problem-solving activities only, or during actual practice.
Moreover, the actual processes of information exchange or problem-solving that act
as triggers or causes of growth by nature are discrete. A child formulates a hypothesis,
tests it, observes its effect and draws a conclusion to either change or consolidate that
hypothesis. Each part of the process takes its own time and even if some parts may
overlap, they are not really taking place continuously.
The discrete nature of these psychological processes goes hand in hand with a
certain inertia in the emergence of effects relative to the occurrence of the cause. There
seems to be a delay between a causal factor or event and the expression of its effects in
actual behaviour. For instance, if we change a condition of learning, such as the
teaching method in a class, its effectsmay be delayed for quite some time. The reason is
that the individual pupils need some time to adapt to the new approach, because they
have formed their learning habits with the old one. The process of adaptation requires
some time and thereby delays the effect of the environmental change.
In other cases, the delay is caused by the time required for the system to prepare an
adequate response or reaction to a perceived problem or opportunity. For instance, if a
group of children loses its 'natural' leader, for instance because he/she moves to
another town, it may require some time before one of the other children has taken that
place. Time is needed to build a coalition that supports the new leadership.
Delay may occur because the circumstances under which a new skill or behaviour
can be displayed take place much later than the actual learning condition. For instance,
in social imitation learning from watching sex-stereotypic models on television or in
movies, the opportunity for showing the imitated behaviour may have to wait for a
social context in which the imitated behaviour may actually be shown. In some cases
the delay may be extreme, as with parents who once suffered from child abuse. Having
children themselves, they may rely on the abusive behaviours they once witnessed in
their parents, since they have had very little opportunity to learn how to deal with anger
more constructively.
110 The dynamic growth model
In summary, there are various reasons to believe that the effect of an event on a skill,
knowledge or behaviour is delayed. I have chosen the term feedback delay to refer to
this characteristic inertia in processes of mental and behavioural change. Feedback
delay adds a certain degree of coarseness or lack of precision to a process. During the
time a response is prepared, the actual conditions making that response adequate may
have changed. The result can be a continuous mismatch between goal state and actual
state. A good example is the behaviour of the logistic growth equation in the chaotic
domain.
How can the discreteness of mental and behavioural phenomena on the one hand,
and feedback delay on the other, be reconciled? Before answering that question, I have
to go a little deeper into the differences between modelling discrete and continuous
processes. Continuous processes are the traditional domain of calculus. A process is
continuous if it changes at any time-scale, however small that scale is. A continuous
process has a derivative, that is, a parameter of change that applies to an infinitely small
time period. Such a process is described by a differential equation. A good example of a
continuous process is the flow of water through a tap or heat conducting through a
metal object. Discrete processes, on the other hand, jump from value to value. An
example is the increase or decrease of a group of people when a person enters or leaves a
room. If a process is clearly discrete, it can best be modelled by a difference equation,
that is, an equation working with discrete steps. The logistic growth equation in our
spreadsheet is a good example of such an equation.
Feedback delay applies to continuous as well as discrete processes, but the difference
is that it applies optionally to continuous events and obligatorily to discrete ones. For
instance, the movement of my arm towards a target is a continuous process, but it
suffers from feedback delay. The actual hitting of the object takes place after the onset
of the movement, and after receipt of the information that has triggered the movement
(that is why we make anticipatory movements: we aim towards the place where the
target will be after the movement time has elapsed). In a discrete process, there is
always feedback delay, in that a time lag exists between any pair of successive states.
Computers can solve differential as well as difference equations, but the way in
which they do so is basically the same in both cases. That is, even differential equations
describing continuous processes are modelled in successive, iterative steps. Various
numerical algorithms simulate the effect of continuity while using discrete steps.
Difference equations are just straightforwardly iterative processes. The big question
now concerns the type of equation that should be used to model developmental growth:
difference or differential? The answer depends on the specificnature of the processes at
issue. If their discreteness tends to continuity, if the time lag between successive points
of change is very small, then a differential equation should be used. Feedback delay can
then be added to the differential equation by making some of its components depend on
some past state of the process. If the real process tends to a level of discreteness of the
same order as the expected or average feedback delay, then a difference equation type
should be used.
I have chosen a difference model throughout this book for several reasons. First,
given the results obtained with the model, in curve fitting and general model building,
 Properties of growth in mental and behavioural development 111
believe the difference model has proven to be a good choice. Second, it is a very
imple model and can be applied easily in the form of spreadsheets or simple computer
rograms. Third, psychological processes of change suffer from a certain coarseness
d randomness of input. In general, the intrinsic coarseness of the difference equation
as good an approximation of the natural coarseness as a differential model with a
andom factor added.

Minimal structural growth level, growth onset time and the origin

If the initial state in the column of iterative growth equations that you have tried to
build is equal to zero, no growth will occur. Put differently, there must be something
that can grow in order for growth to occur. Since the initial state of growth is in
principle the lowest possible growth level and since this lowest possible level cannot be
zero, it should be some arbitrarily small number (e.g. one word, in the case of lexical
gowth). This arbitrarily small number is the minimal structural growth level of a
developmental variable, such as a skill or a particular knowledge domain.
In many cases, a variable such as pre-operational thinking, shows negative growth
after a period of increase. That is, it tends to fade after it blooms. It will rarely
disappear completely, however, in that it will always retain a minimal structural
growth level. For instance, although the thinking style of adults, we may hope, is
basically rational and logical, some remains of a pre-operational thought form that is
magical and egocentric survive. Another example is a skill learned at school, such as
speaking French. If it is not practised, it will fade away. Although almost completely
invisible, it may survive at a minimal level that could form the basis for the rebirth of
that skill if need be.
The growth onset time is the moment at which a structurally minimal expression of a
cognitive element emerges. For instance, the growth onset time of a child's lexicon is
theoretically the age at which the child has acquired the first real word. The minimal
extension is not only hard to ascertain empirically, however, it is also likely that the
minimal set is actually a few items.
If we do not assume that everything that can grow in mental and behavioural
development is innately present in some minimal, or germinal form, we have to explain
how the step from a nil-state (growth level is zero, i.e. the variable is non-existent) to a
germinal state (growth level is an arbitrarily small positive number) can be made. This
step cannot itself be a growth process. Three possibilities can be discerned. First, the
germinal state is innately given. Second, the germinal state has been inseminated from
outside the developing subject; it has been taught or imitated. Third, the germinal state
has been constructed by the developing individual. Perhaps these logical possibilities
also constitute psychologically relevant distinctions.
With regard to the first possibility, presence in a germinal state actually refers to the
innate nature of the concepts and strategies in question. Basic concepts in particular
have an important genetic component. In a recent paper, Jean Mandler has explained
how many basic concepts, such as object or causality, have a basis in innate
112 T h e dynamic growth model
mechanisms of perceptual analysis (Mandler, 1992). The actual onset of growth of
these innate germinal states is probably timed by the growth of conditional or control
variables. These variables could be neurological, environmental, cognitive or a
combination of them all.
The second possibility for making the step from a nil-state to a germinal state is by
assimilating an externally presented model, through imitation and demonstration or
teaching. This process refers to the main source of intellectual growth as far as the
transmission and appropriation of culture by every new generation is concerned. In
teaching, the germinal form of a new grower is inseminated from outside and its growth
is carefully supported and controlled in a process that is now often termed
'co-construction' (Rogoff, 1990).
The third way in which a new grower can be initiated amounts to an autonomous
construction by the subject. That is, since there is neither an example that can be
imitated, nor any innate inclination, the subject discovers a new cognitive possibility.
This is what probably occurs in true creativity.
The construction of new germinal forms is a major problem of development, and in
Chapter 1, I referred to the original version of this discussion in Plato's Menon. It is
likely that at least some of the processes that lead to new forms in psychological
development are formally similar to those that lead to new species in biological
evolution. Given a specific cognitive (or biological) structure, a limited domain of
degrees of freedom exists for constructing new forms (see Ho and Saunders, 1984;
Saunders, 1984, for examples from biology). The construction of new forms is an
intrinsic possibility of a developing system, in that its reproduction over time and its
maintenance are vulnerable to random perturbation (mutation) and to imported
models (imitation) (Fogel and Thelen, 1987; Siegler, 1984). In some cases, these
unintended mutations of some existing capacity are selected and supported by the
external environment. A good example is the early growth of words, based on meanings
given by the adults to proto-meaningful acoustic productions in a baby (see for
instance Jakobson, 1959, on the growth of 'rnommy-daddy' words).
In general, however, newly emerging forms will have to compete with ones that
already exist and although in the long run the new forms will turn out to be more
powerful than existing ones (e.g. operational as opposed to pre-operational thinking),
they are definitely much less powerful at the time they emerge in a germinal form. In
evolutionary biology, a comparable problem occurs in explaining the emergence of
new species: the problem of cladogenesis (Gottlieb, 1984). It is often solved by using
the concept of allopatric growth or allopatric speciation (Mayr, 1976; Simpson, 1983).
Allopatric speciation is rapid evolutionary change in a geographically separated (i.e.
frontier) part of the original species population. Because the separated part occupies its
own small habitat, relatively isolated from the 'mainland', it can change under
relatively safe circumstances with little or no competition from the main species. Later,
the altered species form, if better adapted to circumstances that might have changed in
the meantime, may take over the habitat of the original main population.
If applied to development, allopatric growth means that a new capacity or skill may
be constructed by random variation, selection or imitation. This may occur in a
 Properties of growth in mental and behavioural development 113
relatively isolated and uncompetitive subfield of the field of application of an already
established capacity or skill. At least in the cognitive domain, allopatric growth is a
natural phenomenon, since almost all fields of application of a rule or production
system break down into subfields. These subfields are characterized by differences in
cognitive complexity, difficulty, specific domain of application and so on.
A particularly clear example is offered in Klausmeier and Allen's (1978) longitudinal
smdy of concept development during the school years. The authors distinguish four
different conceptual rule systems that form a developmental sequence, namely
concrete, identity, classificatory and formal levels. They have observed that conceptual
development is not equal for all concepts at all levels. For instance, there is a natural
delay - or de'calage as Piaget calls it - between object, geometric and abstract concepts
and between concepts within each domain, as far as speed and ease of development is
concerned. It would be very difficult to construct a new conceptual rule system (e.g. a
classificatory level) for the whole domain of concepts at once, but it is relatively easy to
do so for limited domains, such as a particular concept from a particular class (e.g. the
concept of 'cutting tool' from the object class). By the time a conceptual strategy in a
more difficult concept domain has reached the limit set by its carrying capacity, a more
advanced strategy is likely to be ready in a relatively isolated easier subdomain (e.g. a
specific concept). The latter may then be adapted to the requirements of the more
complex conceptual domain with relative ease. DCcalages, as they have been called by
Piaget, are the key to development, in that they create opportunities for allopatric
growth of new developmental forms.

The ecology of the mind

So far my examples have been mostly biological; I told an apocryphal version of
Watership Down, for instance, and compared the emergence of new forms to the
process of speciation. But in the biological example we work with real units, that is,
species, which are separated from other species by reproductive isolation (a
hummingbird cannot and will not mate with a sparrow or a titmouse, for instance, nor
even with related species of hummingbirds).
If we think about a child's mind (the system of skills and knowledge, abilities and so
on), we may metaphorically describe the structural elements that we can discern here,
such as skills, concepts and rules, as species in a mental ecology. Each species occurs with
a specific population (growth level) and relates to other species, that is other structural
elements (Boulding, 1978). For instance, it is likely that fast growth of the species
'words' in a child will affect, positively or negatively, the growth in the species
'grammatical knowledge'. This is so because one growth process may feed on the other.
The onset of grammatical growth may depend on the acquisition of some threshold
number of words, in that skills necessary to learn new words contribute to the learning
of grammatical rules. On the other hand, rapid increases in grammatical knowledge
consume part of the time and effort that might be employed in building up the initial
vocabulary (Dromi, 1986). I shall therefore compare the mental and behavioural
system of a developing person with an evolving ecological system, which is not an
114 The dynamic growth model
ecosystem of animals and plants, but an ecosystem of mental species that take the form
of rules, concepts, skills, abilities, beliefs, activity patterns and so
You may object that the units in the mental ecology are not naturally separated in the
same way as biological species. Why consider 'words' as species, and not the whole of
language? Why not consider abstract words as separate species next to concrete words?
It is indeed true that there is no natural compartmentalization in the mental ecology, in
that the units are not separated by real boundaries. Nevertheless, once we agree on
some level of analysis, that is, on taking either broad categories such as language, or
smaller categories such as abstract words, it is not difficult to reach a working consensus
as to what the units are at that level. For instance, if language is considered a 'species' in
the mind, then we will probably see it next to another broad category such as 'social
knowledge'. If we take 'abstract words' as a 'species' level, we wiil see 'language' as its
overarching family and 'concrete words' as its fellow species. There is no reason,
therefore, to discard the mental ecology and species metaphor only because there are
no such things as real mental species.
The ecological metaphor is very convenient in that it specifies several powerful
heuristic principles (van Geert, 1991, p. 5):

1. Given a specific structural model, the human cognitive system can be

described as an ecosystem of species (i.e., structural elements such as
vocabulary, grammatical rules, problem solving skills, concepts, etc.) that
entertain growth relationships with specific fields of application.
2. The elements engage in various types of functional relationships, which are
either supportive (the growth in one supports the growth in another),
competitive (the growth in one relates to the decline in another), (virtually)
neutral, or conditional (the presence of one is necessary for the emergence of
another).
3. The elements show strongly dissimilar growth rates and growth-onset times.
4. The components compete for limited spatio-temporal, informational, ener-
getic, and material resources.
5. a. There exist more cognitive "species" (skills, knowledge items, rules, etc.)
that can in principle be appropriated than actually will be appropriated by any
particular person. b. In principle, any cognitive "species" may occur with any
possible growth level. c. The set of cognitive species and respective growth levels
characteristic of a person's cognitive system is the dynamic product of cognitive
growth under limited resources.

These heuristic principles are reminiscent of those for biological ecological systems
in general, and evolutionary systems in particular. For instance, the fifth principle is
reminiscent of a principle in Darwinian theory, that the number of offspring exceeds
the number of organisms an environment is able to support, long enough for each
organism to reproduce. From this it follows that the adaptation of organisms to their
environment increases over generations. Likewise, we may heuristically claim that
learning under competition for limited resources favours 'more learnable' more than
'less learnable' mental species. Since 'learnability' is dependent on the set of
supporting cognitive resources that together form a person's cognitive system, more
 Modelling real growth 115
easily learnable cognitive species (rules, skills, concepts, etc., that are more easily
learnable in the person's current cognitive system) tend over time to become more
frequently represented in such a system than do less learnable ones (see for instance
Newport, 1982, for an application to language; van Geert, 1985).
In the framework of this ecological metaphor, we may speak about a 'cognitive
grower' and its 'environment'. A cognitive grower6 can be any of the species in the
mental ecology, any structural element or component of a cognitive system to which
the growth relation applies; by 'grower' I do not mean an individual child, but rather
the child's lexicon, or the child's use of the inversion rule in questions, or any other
mental species. Trivially, a grower is a cognitive species that grows. The environment
is the totality of supporting or competing resources upon which the grower 'feeds'.
Thus as far as the nature of the cause or the magnitude and time of the effect is
concerned, I make no a priori distinction between subject-dependent and external
resources, such as the number of available models or tutorial support (see also Fogel
and Thelen, 1987; Thelen, 1989). A child's mental ecology, therefore, consists of the
contents of his or her own mind in addition to the accessible environment.

Modelling real growth

Keren's early lexicon
In principle, a growth equation models a process of growth in an individual child. The
growth rate and carrying capacity parameters are specific to an individual, in that it is
highly unlikely that they will be similar for all subjects in a population. Even if the
growth rate were biologically determined, which is very improbable, the carrying
capacity would at least in part depend on the input from the environment, which differs
between children. In order to test whether the growth equation provides a good
explanation for a growth process, we need to look for good individual data based on a
longitudinal study with dense measurement points. This type of data can be found
particularly in the field of early language studies, where a majority of researchers have
followed a single-subject design.
In a fascinating study of the early language development of her daughter Keren,
Dromi (1986) presents a number of growth curves, one of which covers the growth of
Keren's lexicon between the ages of 10 and 17 months. Keren was indeed a
linguistically creative and precocious girl, and data from other studies reveal a slower
growth rate (e.g. Nelson, 1985) in other children. Nevertheless, the point is not
whether the data are characteristic of the whole population of early language learners;
they need not be, since there are considerable individual differences in the rate of
almost any developmental process, but rather whether the process can be explained or
modelled by applying the logistic growth equation.
Dromi's study covers the time of one-word sentences, that is, before the onset of true
syntax learning. Syntax is an important resource factor in the development of
language, since it definitely changes the need for specified words (such as words
116 The dynamic growth model
belonging to different syntactic classes) but also allows for the expression of
meaningful content in the form of word combinations instead of single words. Let us
therefore assume that the one-word stage is a period where no significant changes occur
in the carrying capacity for new words and that this capacity will change as soon as
syntax starts to develop. This is of course a rather crude assumption, since we may
expect the carrying capacity to fluctuate. It is assumed only that there will be no major,
systematic change that is comparable to the emergence of syntax. Since Keren's
lexicon counted approximately 350 words by the time she started to use multiword
sentences, the one-word stage carrying capacity can be set to 350.
Because we have no idea how long the assumed feedback delay is in the case of early
word learning, I shall set it to one week, which corresponds with Dromi's sampling
points. That is, the iterations of our equations are conventionally set to one week. This
estimation is as good as any other, since the exact length of the feedback delay is not
really important at this level of modelling.
By applying the equation to any pair of successive sampling points, we arrive at an
estimation of the growth rate r (which is approximately the average of the rs resulting
from comparing each consecutive pair of points). The following tutorial shows how the
data can be transformed into a logistic growth model.

Model 4.4 A logistic model for Keren's data

I reconstructed the data from Keren's cumulative lexicon for successive weeks of
observation by measuring the points in Dromi's diagram, which is probably not the most
reliable method for obtaining somebody's data, but it certainly suffices for the present
model.
Put the data in successive cells in your spreadsheet, for instance in the range BI..B32
and fill a column AI..A32 with the number series 1 to 32, for the weeks of observation. The
number of words in the lexicon are 2,2,2,2,3,5,5,7,10, 15,23,31,34,35,40,43,52,55,63,
68, 88, 104, 118, 140, 184, 220, 250, 272, 293, 319, 333, 340.
Given the equation for logistic growth (4.10), it follows that the growth rate r can be
computed for any two successive data points in the following way:

In cell C2, adjacent to cell B2, you write the spreadsheet equation computing the growth
rate for points B1 and 82 as follows:

and copy this to cells C3..C33.

Compute the average growth rate in cell D l , for instance, by inserting the expression:

The result is 0.31.

Next go to cell G I ,and write 'growth rate r'. Enter the value of the average in cell F1, and
 Modelling real growth
define the range name 'r' for that cell. Enter the following equation in cell C2 and copy it to
the range C3..C32:

Write the value of the first data point (2, for two words) in cell C I . Go to cell D l , and enter
the equation:

which means that you want to see the value of C1 rounded off to the integer level. The
reason is that you will never observe 3.2 words, for instance, but either 3 or 4. Copy this
equation to D2..D32. Define a graph 'Timeseries' with A1 ..A32 as A range, and D l ..D32 as
range. The A range holds the data; the B range is your model outcome. Look at the
result. The similarity between data and model is not very convincing.
How can it be improved? Try several possibilities, for instance, start with a significantly
lower initial state. The value 0.66 for instance will result in a very good fit with the data. But
this value makes no empirical sense, since the child has either 1 or 0 words. Try different
growth rates. Lower growth rates result in a reasonable fit, but this method has two
disadvantages. First, the computed growth rate was 0.31, not a lower value. Second, the
state at iteration point 32 is significantly lower than the empirical data point. The best
solution can be obtained by applying the following assumption. First, assume that the
initial state, the minimal structural growth level, is 1, and not 2 as in the data. Second,
assume that the growth onset time is somewhat later than the onset of the observations.
That is, copy the number 1 to cells CI..C3, thus making the growth onset time equal to
observation week 3. The resulting model curve shows a very good fit with the data, except
for the first few weeks (Figure 4.3).
Finally, write the data series in range AAl..AA16, but skip one data point each time
(thus, use the observation weeks 1,3,5,7, etc.). Compute the growth rate as described in
this tutorial. Since the interval between data points is two weeks, the growth rate is
computed for a corresponding feedback delay (two weeks). Try a growth rate of 0.71 for
the reduced set of 16 data points (Figure 4.4).

The fit between the mathematical curve and the data is very good and it is quite
remarkable that a seemingly uncontrolled and random-based process such as early
word learning follows a mathematical curve so closely. We have seen; however, that the
curve is critically dependent on the initial state value. A change of only one word at the
beginning made a considerable difference in the form of the curve. This seems to lead
to a rather serious objection to the mathematical model: Could a small, probably purely
random-driven difference in a single word at the beginning of the growth process really
be so important that it determines major differences in growth trajectory?,This is a very
unlikely assumption. The point is, however,.that the initial state (of the lexicon, for
instance) is a lot more 'fuzzy' than later states. Are the two or three things that we call
'words' at the very beginning words in the fundamental sense of that concept? This
question is very difficult to answer. The child's first utterakes are seen as words
because the investigator decides they sound sufficiently similar to real words from the
adult lexicon, but what are they from the child's point of view? Are they really
functional in the same way as the later words are? It is clear that this fuzzy beginning
 ,
I

118 The dynamic growth model I

Figure 4.3 A comparison between empirical data on lexical growth in a child (after
Dromi, 1986) and the logistic model with parameters fitted onto those data.

Figure 4.4 Theoretical curve of lexical growth based on a feedback delay of two weeks.
 Modelling real growth 119
carries the roots of the whole lexical growth process, but it is unclear to what extent the
properties of that fuzzy state should be taken in an absolute sense. The
solution I prefer is to manipulate the data from the initial state somewhat and to try to
find a good theoretical fit with minimal manipulation.
Instead of actually manipulating the data at the initial state level, we may conceive of
the whole range of data as affected by some random influence. That is, the number of
observed words is probably not exactly the same as the number of actual words Keren
knows (although it is probably pretty close, in view of the high observation density).
Moreover, the growth process itself may not be as deterministic as the mathematical
model suggests. I t might be more realistic to see each step in the lexical growth as a
randomized function of the previous step, instead of an absolutely deterministic one.
Nevertheless, the basic belief in the dynamic systems model is that simple equations
faithfully model real processes. So, although some randomness may be tolerated, it
should not be too much. Put differently, if I need a high degree of random influences to
explain my data, the model has little to contribute.
Let us try a small random factor, therefore, and investigate what happens if it is
imposed on the theoretical model. One way of studying the effect of random noise is to
multiply the growth level by a random number bigger or smaller than 1. The growth
equation then becomes:
L,,, = L; (1 + r - r - LJK) ( 1 + (rand - rand)/d) (4.25)
for rand a random number between 0 and 1 and d a divider. It should not be too
difficult by now to build a spreadsheet model with this equation. Figure 4.5 shows the
result of applying a maximal + 1 -5 per cent random fluctuation to the growth
process (with d = 20). The initial state is set to 1 word, K to 350 words and r to 0.29.
So far I have manipulated the initial state level in order to obtain a better fit with the
data. I have also tried a model with a small random fluctuation imposed on each
computed growth level point. The feedback delay, however, has not been changed. In
view of the arbitrary setting of that delay (equal to the sampling points), we might as
well try to change it and see how it affects the theoretical curve. By taking a feedback
delay of two weeks (simply double the sampling time), you find a growth rate r of 0.71,
resulting in an even better fit (see Figure 4.4).
On closer inspection, however, Keren's lexical growth curve during the one-word
stage seems to consist of two substages. The first is a stage of growth that seems to level
off during weeks 16-19, at a growth level of about 50 words. I t is immediately
succeeded by a second substage of almost explosive growth, leading to the temporary
ceiling level of around 350 words. The bend in the growth curve actually corresponds
with a change in semantic strategy. This shows that our original assumption, that the
carrying capacity does not change significantly during the one-word stage, was
probably incorrect. In order to model this change more adequately, we should employ
a model of connected growers, for instance a model specifying the relation between
lexical growth and the growth of a new semantic strategy. A connected growers model
can solve the problem of a less than optimal fit with the lexical growth data (see van
Geert, 1994).
120 The dynamic growth model

Figure 4.5 The lexical growth data compared with three curves with 5 per cent random
effect per data point.

The growth of memory and the manipulation of feedback delay

Everybody who has ever asked the way in a strange town, for instance how to get to the
convention centre where the Pan American Society for the Study of Memory holds its
biannual meeting, has experienced human beings' limited attention span. While the
person asked goes on explaining, the listener usually loses track after two blocks right
and one to the left. The amount of new information one can store in short-term
memory is limited to the magical number seven, plus or minus two, as Miller once
remarked (Miller, 1956). Instead of attention span, investigators use the more common
term 'working memory'. There is ample evidence showing that the size of working
memory increases during development. One way of testing a child's working memory
is to let him or her count dots on a card, and ask the child to remember the number for
later retrieval. With only one card, the memory task is trivially simple, but with seven
or eight cards, each of which contains a series of dots that have to be counted and
remembered, the task becomes virtually impossible even for most adults.
In a paper dealing with the relation between brain growth and cognitive growth,
Case has summarized a variety of investigations on children's working memory, using
different types of tasks (Case, 1992a). One series of data shows the increase in working
memory based on a counting and a spatial task, between the ages of 4 and 94 years. I
have turned them into a cumulative graph, assuming that the lower limit is 1 (that is,
the child remembers one number) and that after the age of 10 the growth of working
memory in this task is asymptotic.
 Modelling real growth

Figure 4.6 Data and theoretical curve for the growth of memory span, for r = 0.44 and
f = 6 months.

The problem with building a growth model for these data is that they are group data,
not data on an individual's memory development. I shall assume therefore that the
group curve stays close enough to the growth curve of a significant number of
individuals, and that it can be considered the curve of a 'typical individual'. I am afraid
that this is a rather Queteletian solution, but it is the best we can do at the moment. The
actual raw data should be consulted to see whether the individual variation indeed
allows for such an assumption.
Figure 4.6 shows the result of modelling the data with a growth rate of 0.44, and a
carrying capacity equal to the asymptotic growth level from the data (about 5). The fit
between theoretical model and data is remarkably good. It is possible, however, that
individual growth curves are much less smooth, but that remains to be tested against
the data. A comparable attempt at building a mathematical model has been carried out
with data from another test for working memory, memory for consonants (Bleiker,
1991; Case, 1992a). The data again fit the logistic curve very well. It should be noted,
though, that the best fit requires that either the feedback delay or the initial level have
to be manipulated. Which of both actually corresponds with reality is.something that
cannot be decided except on the grounds of empirical data.
, How can feedback delay be manipulated, given that it is a single iterative step in our
b
spreadsheet model, not a time parameter?
!
I
122 The dynamic growth model
Model 4.5 Manipulating feedback delay
A good and simple algorithm to simulate a growth curve that occurs in arbitrarily small
step sizes is the so-called trapezoidal method. It correctsfor the remaining feedback delay
by averaging over two consecutive steps. Define a parameter name 'step' in addition to
the parameters you have already defined for the normal logistic model. Start with a value
for 'step' of 0.01. The easiest way to write the algorithm in a spreadsheet format is to work
with three columns:

Column A Column B Column C

0.01 +A1 +$step*$r*$Al*(l -$Al) +B1 +$step*$r*$BI*(l -$Bl)
+
(A1 C1)/2 + +
A2 $step * $r * $A2 * ( I - $A2) + +
B2 $step * $r * $82 * (1 - $B2)
(4.26)

Copy range A2..C2 to A3..C1000. Column A contains the values of the logistic equation.
The equations simulate the true differential form of the logistic equation 4.16. The
feedback delay is 1/100th of the delay in the normal spreadsheet version (hence the step
size 0.01). By setting the value of step to any decimal number, you can test the effect of
different feedback delay values. Your test will show that longer feedback delays result in
curves that are less steep at the beginning and steeper at the end than curves based on
smaller feedback delays. Try step sizes that fluctuate around the value 1.

Signs of chaos and cyclical change?

T h e growth of the lexicon and of working memory fits remarkably well with a logistic
growth process. Given the simplicity of the equation and the very few parameters
needed, this is indeed an interesting discovery. Nevertheless, given certain parameter
values, the logistic growth model produces curves that are less regular and smooth than
the nice sygmoid shapes we have seen in the data. As the growth rate increases, a
logistic process enters the domain of cyclical switches and even chaotic change. Is there
any evidence for chaotic change in development? In Chapter 4 I have given examples of
growth models that clearly showed a chaotic aspect. Empirical examples are the growth
of utterance length (Chapter 5, Figure 5.3; Brown, 1973), the growth of the present
progressive (Chapter 3, Figure 3.13; Brown 1973) and the rate of word learning
(Chapter 3, Figure 3.12; Corrigan, 1983).
In fact, something that looks very much like chaos can also be found in the data on
lexical growth from Dromi (1986). The cumulative curve of lexical growth looks quite
smooth, but what do we find if we plot the successive growth rates? Figure 4.7 shows
that they vary strongly over time. This variation could be the result of random noise or
of random external events, but it could as well be the consequence of a chaotic growth
pattern of the growth rate itself. The model does of course not tell which of the two
possibilities is correct. I t just offers a second theoretical possibility - chaotic growth -
in addition to the common random noise explanation of irregularity.
The logistic equation not only produces regular evolutions towards a point attractor
and chaotic evolutions, it also produces patterns of regular switches between two states
 Modelling real growth

Figure 4.7 Although the growth pattern of Keren's lexiion is very close to a smooth
logistic curve, the underlying growth rate r varies in a seemingly chaotic fashion. It appears
there are two substages in the process, each with different magnitudes of change of r over
time.

(and between any multiple of two states). Do we find such patterns in mental or
behavioural growth?
In fact, many such patterns exist in behaviour, but in general we would not be
inclined to put them under the heading of 'development'. For instance, children switch
between effort and relaxation, interest and boredom, sensitivity and habituation,
activity and rest. Another example comes from the research on the phenomenon of
attachment in infancy. It shows, among other things, that children switch between
periods of active, individual exploration of the world and periods of seeking comfort
and protection with the people to whom they are attached. Mike Apter has described
this and similar phenomena in terms of threshold crossings (Apter, 1982). If a girl
toddler is secure she may go out and explore something new, such as the neighbours'
youngest son. But as she discovers that he is not only quite an interesting boy, but also
sometimes very rude, her interest eventually-turns into fear. She runs back home to
seek the comfort of her familiar place and asks her mother to read from her favourite
book. After a while the state of rest and comfort turns into boredom, and she decides to
continue her exploration of the world outside. This so-called reversal shift is clearly a
carrying capacity phenomenon, in that the child can tolerate a certain upward
deviation from the normal level of comfort or security in the form of fear, but that it
 The dynamic growth model
shifts towards a downward deviation from the comfort level (in the form of actually
seeking comfort with familiar persons) if the deviation becomes too great. The process
also goes the other way, in the sense that actual comfort and relaxation-seeking turns
into boredom. These cycles or shifts are not seen as genuine developmental
phenomena, in the sense that they do not lead to a higher or steady-state level. They
are, however, part of the normal dynamics of human behaviour. It is interesting to see
that the logistic growth equation at least qualitatively explains such switches in terms
of a range of growth rate values higher than the ones leading to steady states and lower
than those leading into chaotic oscillations.

1. This is indeed only a popular or superficial interpretation of Piaget's theory, based on the
textbook descriptions that emphasize the universality and necessity of the stages. Piaget's own
point of view was much more subtle.
2. Everyone should know that there are no ferries to uninhabited islands and the Dutch would
have built a bridge anyway.
3. Yes, another Belgian, like Quetelet. I cannot help it that my homeland made such an indelible
impression on my work.
1.. If you use a spreadsheet under Windows, like Lotus 123 or Quattro Pro, there is a simple way
to define all your range names at once. Assuming you have to define four range names, for
parameters named 'r', '8, 'in' and 'I?, write those names in four consecutive cells, for instance
Al..A4. Enter the corresponding values of the parameters in the range Bl..B4 (that is, right to
each parameter name). Select the entire Al..A4 block with the mouse. In Lotus, open the
menu Range, then Name. An interactive window appears. It has a button marked 'Use
Labels' and a small options window, which is by default set to 'To the right'. Just press the
'Use Labels' button, and Lotus knows that the values in cells Bl..B4 are named 'r', '8, 'in' and
'K' respectively. The procedure is very similar in Quattro Pro. Enter the parameter names,
then right to those the values. Activate the Block menu, then names, then Labels, press OK.
The optional direction is set to 'Right'. If you prefer values left to or below the cells with the
names, change the direction option accordingly.
5. The metaphorical term 'cognitive species' is similar to several terms introduced by scholars
who have applied evolutionary analogies to the problem of the cultural transmission of
knowledge and skills. They have proposed several terms to describe the units of transmission;
Dawkins (1976) used the term 'meme' as the cognitive analogon to 'gene'; Boyd and
Richerson (1985) speak about 'culture types'; Lumsden and Wilson (1981) use the term
'culturgen'; see van Geert (1985) for an overview.
6. If anyone ever considers translating this concept into French, he or she will practically be
obliged to use the word 'croissant'. Finally, developmental psychology deals with something
substantial.
Transitions and
developmental fractals

Transitions and sudden changes

Nur ein Aha-erlebnis, gnadige Frau ...
Just a sudden insight, Madam. . . . One can imagine Albert Einstein giving this answer
when asked how he discovered his theory of relativity ('which one, Madam?' . . .) The
German word 'Aha-erlebnis' literally means 'Aha'-experience; after you struggle with
a problem for quite some time the solution may' dawn suddenly and the usual
exclamation then is supposed to be 'Aha!'. Although I doubt whether Einstein got his
relativity theory in the form of a sudden insight, it is a common process in creative
thought.
There is a famous story about the nineteenth-century German chemist Friedrich
August Kekuli. He had been struggling with the problem of the structure of organic
molecules until, one night in 1865,' he had a dream of a snake, swirling around and
biting its own tail. The dream, so the story goes, solved the problem of the six-carbon
benzene ring, which in turn made all the problems of organic molecule structure
Kekuli had been struggling with for years fall into place. It is interesting to note that
the dream sprouted out of fertile soil, since Kekuli had earlier training in architecture,
which made him particularly sensitive to spatial structure.
The emergence of sudden insights or discoveries has been of special interest to
Gestalt psychologists. They attributed it to processes of pattern reorganization. A good
perceptual example is the famous Necker cube, which you see as facing either upward
or downward (Figure 5.1). Sometimes it is difficult to change the perception of
downward orientation, for instance, but once the organization starts changing, it
changes almost instantaneously. Necker cubes are systems with bistable states, and
each state functions as an attractor point (Ta'eed, Ta'eed and Wright, 1988). Compare
this with another bistable system, such as a rowboat in the water. It is bistable in that it
either floats or sinks. Normal perturbations of its balance, such as the movement of the
rower, will shift the balance somewhat, but the position of the boat fortunately
fluctuates around the balance point. If the perturbation is too strong, for instance if the
 Transitions and developmental fractals

Figure 5.1 The Necker cube (top) is an example of two stable points of organization
(perceptual orientation 1 and 2), connected by a sudden-switch-trajectory.

rower stands upright and bends to one side, the boat will topple over and rapidly shift
into its other state, sinking. As soon as the equilibrium point is crossed -like the rower
bending too much to the right - a positive feedback cycle is set to work and the forces
driving the boat towards imbalance will be amplified instead of dampened. Bistable
states abound in perception and thinking - at least that is what personal experience
suggests. Their characteristic dynamics is that of a long period of stability in one state,
then a short intermediate jump to the other state, which is then stable for a long time.
This chapter begins with a discussion of empirical evidence of transitions and
sudden jumps in development, which occur at various levels of generalization,
measurement or description. The distribution of such transitions over all possible
levels of observation leads to a second topic, the fractal nature of developmental
processes. Next, I present a mathematical model of sudden jumps or 'strong'
transitions and show that these strong transitions bear all the marks of an interesting
mathematical object, the cusp catastrophe, which serves as a very general model of
sudden state change. Is there any empirical evidence on sudden changes or jumps in
development?
 Transitions and sudden changes 127

Figure 5.2 Bentler's group data on the growth of conservation as measured by a

conservation test, and two logistic model simulations (aftei- Bentler, 1970).

The logistics of conservation growth

Conservation is a good candidate for a sudden-shift type of developmental model. If we
look at group data of conservation growth, however, the curves themselves look at first
like logistic growth forms (Bee, 1989; Tomlinson-Keasy et al., 1979). The point is that
a group curve of conservation development does not say anything about individual
development. A child's answer is either wrong or right. But is a child a 'conserver'
when giving a correct answer to a conservation question? The conservation answer
could be just an accidental hit. It is customary therefore to ask children a series of
questions and give them a conservation score relative to the number of questions
correct. The Goldschmid and Bentler test of conservation works along this principle.
Bentler (1970) used this test in a cross-sectional study of 560 children between the
ages of 4 and 8. Figure 5.2 displays mean scores on the test, compared with two logistic
model fittings. It should be noted that I linearly extrapolated the data under the age of
4 years to a minimal score at the age of 2$ years and to a maximal score at the age of 9.
The data are again quite in accordance with a logistic growth model. A growth spurt
can be clearly observed, but it is not a sudden transition. This-conclusion, however,
does not do full justice to the available data and is certainly not valid as far as individual
growth of conservation is concerned. This follows from Bentler's finding strong
evidence for a bimodal score distribution. That is, the scores did not follow a normal
 Transitions and developmental fractals
Istribution, with the majority of the scores evenly distributed around some average. If
that would have been the case, development could be expected to follow a gradual
increase in the average score. The bimodal distribution, however, shows that a
majority of children have either a low score or a high score. The test, therefore,
measures the understanding of a principle, namely that of conservation. The children
either understand it or not and apply it to all the conservation problems they can
identify as such. The group curve shows the gradual shift in the number of children
with a high score, resulting in a gradually increasing mean score.
The discovery of the conservation principle appears to take only a little time. This
conclusion can be inferred from the bimodality of the scores; the less time it takes for a
child to shift from a state of ignorance to a state of understanding, the lower the
probability that the investigator will encounter children who are in the transition stage;
or, following the same line of argument, the higher the probability that a child will be
found in a state of either relatively good or relatively poor understanding. The
individual curves, therefore, are much steeper than the group curve. It follows then
that the real growth curve of conservation, the curve as it applies to individuals, is more
like a strong transition than like a growth spurt of the more elongated type.
Convincing evidence for a sudden jump from a state of non-conservation to
conservation comes from the longitudinal study of Han van der Maas (1993). He tested
a group of children with a computerized conservation test. The children made the same
test repeatedly over many trials. The computer kept track of their scores and in this
way a group of individual conservation time series was obtained. The group of
transitional children showed a characteristic jump (Figure 5.3). With this particular
form of testing we do not see children who regress to the non-conservation state, or
who fluctuate between the two states for some time.

States and trajectories of understanding conservation

The question is, however, whether the notion of bistability and the associated sudden
jump implies that there are indeed only two states of conservation development,
namely conservation and non-conservation.
One of my collaborators, Jan Bijstra, designed a conservation test that could
distinguish 'intermediate' conservers from real non-conservers and conservers (the
Goldschmid-Bentler does not make this distinction). He investigated a suggestion of
Flavell (Flavell, Green and Flavell, 1986) that a significant group of children is not
classified as non-conservers if the investigator asks a slightly different question.
Instead of asking 'Is there as much lemonade in this container as in the other one, or is
there more or less?', the question should be rephrased as follows: 'Is there really more
lemonade in this glass, or does it only appear to be so?'. The idea is that children who
are close to conservation understanding, as measured by the standard resting
procedure, already understand that the higher amount of liquid (in the case of a
narrower but higher beaker) is a matter of appearance, not of reality. But the standard
question confuses them, and if they say 'there is more' they actually mean 'there seems
to be more'.
Figure 5.3 Jump in raw conservation scores for 24 transitional subjects (after van der
Maas, 1993). The moment of transition functioned as an anchor point for the scores of the
individual subjects. As a result, the extremes (beginning and end) are based on data from
only a few subjects, and are subject to considerably more noise than the data points around
the jump.

Bijstra (1989) confirmed this hypothesis. He also found that over repeated testing
sessions, separated by a period of a few weeks, some children tend to shift back from a
'higher' to a 'lower' state. Thus a child who is a clear 'conserver' during the first session
is found to be a convinced 'intermediate' during the next session (see also Bijstra, van
Geert and Jackson, 1989). This phenomenon occurred with a significant number of
children (about 17 per cent). The first tentative conclusion that can be drawn from this
observation is that for at least some children, the stable initial state (non-conservation)
switches to a non-stable state, which is actually a fluctuation between two or possibly
more states, for instance, non-conservation and appearance understanding. Finally,
they settle into a stable state, that of conservation understanding.
Piaget (1957; see also Flavell, 1963) actually distinguishes four states, each
considered equilibrium states. The Piagetian experimenter shows the child a ball of
clay, then rolls it into a long thin sausage or into a flat pancake. He then asks if there is as
much clay here as in the ball. In the first knowledge state, the child will concentrate
either on the length or on the thickness i f the sausage during one testing session and
say there is either more or less. In the second state, the child will concentrate on one
dimension, thickness for instance, but switch to the other dimension at some point
during the testing session. In the third knowledge state, the child will simultaneously
look at the thickness and length but will not know how to solve the conflict between the
two. These three knowledge states result in a non-conservation answer. They are
assumed to succeed one another over a period of a few months. Finally, the child
130 Transitions and developmental fractals
understands that the thickness and length stand in a compensatory relationship to one
another and reaches the correct answer.
The bimodality, or multimodality, of the scores on a conservation test suggests that
the shift between one state and another in an individual child is a sudden jump from
one state to another, that is, a strong transition. This is not only for the trivial reason
that there is no meaningful intermediate state (it is either one dimension observed, or
two). The conservation tests present the children with a number of questions, so they
could have any percentage correct. This percentage could gradually grow over time,
but it does not really seem to do so,
On closer scrutiny, however, the sudden jump quality is a matter of the
measurement scale that we use. The Bentler scale divides the group of children into
straightforward conserver and non-conserver groups. Bijstra's scale made a distinction
between three groups, including transitory children, by using an appearance-reality
criterion. Piaget looked at the way children decide about whether there is the same
amount or not and distinguished four different groups. We might be inclined to say
that the more refined the measurement scale, the more the results reflect the real
distinctions and categories. Such a conclusion, however, does not reckon with the fact
that measurement scales are not chosen in function of how closely they reflect the true
distinctions in nature; their choice depends on the specific question asked. It is
legitimate to ask whether a child can conserve or not and for that question the Bentler
test is the right instrument. If you want to know how children come to their answer,
you should employ Piaget's criteria. I t is like the famous example of the coast of
Britain; its measured length depends critically on the smallest length unit employed.
The question which unit that should be depends on what purposes the measurement
should serve.
We can imagine that the sudden jump pattern occurs on any level of measurement.
If we measure the straightforward conserver-non-conserver distinction, it is likely that
the transition from one state to the other takes the form of a sharp transition. If we also
take intermediate conservers into account by asking for the appearance-reality
distinction, it is likely that we will find a sudden jump from non-conserver to
intermediate and from intermediate to conserver.
If the sudden jump indeed occurs between any pair of successive states irrespective
of the actual measurement scale - and there is no reason to doubt this fundamental
assumption - conservation development is highly reminiscent of a mathematical object
that is self-similar over all scales of observation, the fractal.

Fractals and self-similarity in development

Towers and robots
In the previous section, I suggested that conservation transitions occur in a fractal
pattern: the same transition pattern emerges at all levels of measurement. Let me now
turn to a different sort of information about possible self-similarity: microdevelop-
 Fractals and self-similarity in development

Figure 5.4 The tower of Hanoi in its three-disk version. T h e combination of disks at the
bottom is not allowed by the rules of the game, which do not allow the player to put a
larger disk on top of a smaller one.

ment, which involves changes of understanding over the period of a single

problem-solving activity. The question is whether the short-term change follows the
same basic pattern as the long-term macrodevelopment.
Bidell and Fischer (1993) used 'The tower of Hanoi' - not an adventurous Vietnam
movie, but a classic brainteaser game - to study micro-macro similarities. A simple
version of the tower problem involves a block with three pegs and three wooden
doughnut-shaped disks that can be stacked over the pegs. In the starting position, they
form a pyramid on the left peg. The problem is to make a similar pyramid on the right
peg by moving one disk at a time. Putting a larger disk on top of a smaller one is not
allowed (see Figure 5.4).
If a child is presented with the tower problem for the first time, it is likely to
approach it in a way that, on the one hand, depends on its current developmental level,
but that, on the other hand, is considerably below the optimal solution level that it will
eventually reach. The problem-solving process will therefore show a microdevelop-
mental path between the entry level, the opening level of problem-solving and the
optimal level, which is the level reached after working on the problem with eventual
contextual or tutorial support. Children will of course differ in entry and optimal levels
and the difference will no doubt depend on the stage of macrodevelopment they are in.
A striking finding is that the children show a sudden shift from inefficient
problem-solving characterized by wrong movements, hesitations, etc., to highly
efficient acting, resulting in a sequence of correct steps. More precisely, once a child got
two steps in succession right, it had all of them right without further hesitation. A state
132 Transitions and developmental fractals

Figure 5.5 Developmental levels of successive actions and explanations in the robot
experiment (after Granott, 1993); scores are based on discrete levels with one intermediate
score between each two consecutive levels.

of preparation characterized by unsuccessful moves leaps suddenly to a state of correct

understanding and mastery of the problem. It is as if the child jumps from a tower;
after a long and strenuous march up the stairs to the top it reaches the top and jumps
and suddenly the speed of events increases explosively (I admit it is a rather cynical
metaphor, but do not worry, the child has a parachute).
A second finding is that the preparation time, the time before the leap, is markedly
different for children of different ages. The probable explanation is that the older
children start the problem-solving at a higher entry level, with a model or strategy
closer to the optimal level. Solving the tower of Hanoi problem requires cognitive
functioning at the level of representational systems (in Fischer's terminology). For
most children, the entry level is the lower level of representational mappings. During
the problem-solving act, they construct a representational systems approach in much
the same way as they did during the macrodevelopmental process. In this way, micro-
and macrodevelopment are qualitatively similar, but differ in the fact that micro-
development concerns a shift from an entry to a solution level during an experimental
session, whereas macrodevelopment refers to the changes in optimal levels of
functioning over the course of years.
A second example of a microdevelopmental process is Nira Granott's robot
experiment (Granott, 1993). She presented a small group of adults with a little
electronic device on wheels, the robot, and asked them to explain what controls the
robot's behaviour. The robots are tuned to react to light, sound, shade and pressure,
but their reactions are a rather complicated function of these inputs. I have had the
opportunity to play with one of the robots and they are indeed weird little creatures.
During the experiment, which is just a free interaction among the participants, all
 Fractals and self-similarity in development 133
actions and suggestions are videotaped. They are scored in terms of Fischer's thirteen
developmental levels. Because the robots are completely unlike anything else the
subjects have encountered before, the initial actions and suggestions are at the lowest
possible developmental level, sensorimotor actions. The explanations move towards
higher developmental levels during the course of the experiment until the groups reach
an explanation at an optimal cognitive level (Figure 5.5). The pattern of explanation
levels is highly similar to the pattern of macrodevelopment found over the course of
twenty-five years. The same stages occur in the same order. Stage shifts occur in the
form of sudden jumps, every so often leading to temporary regressions to a lower level
of functioning. Granott's experiment shows that we can expect similar patterns at the
micro- and macrolevels only if the microlevel poses a sufficiently strong challenge; with
familiar objects the subjects would have started at a much higher developmental entry
level, possibly their optimal (highest) level. What explains the sudden jumps and
strong transitions found in so many developmental phenomena?

Sandpiles, transitions and subcriticality

If they are not polluted by the effluent of sewers or the oil spilled by the many coasters
scattered over the horizon, the North Sea beaches offer the contemplative sunbather a
beautiful opportunity to carry out experiments with fine, dry white sand. What
experiments, the attentive reader will certainly ask and what have they got to do with
development? Take a handful of fine white sand and let it slowly glide through a little
opening in your fist. The sand will pile up to form a sort of cone. The sand will flow
smoothly from the flanks of the cone, but sometimes a little extra sand takes a lot of
sand with it in a miniature avalanche. You can of course also take your son's plastic
bucket, give him some money to buy himself an ice cream and, after he has disappeared
from sight, cut a tiny little hole in the bucket, fill it with dry sand and repeat the
experiment under more controlled and scientific circumstances. Meanwhile, think of a
good excuse for when the kid comes back.
I admit that the conditions of observation are not really optimal, so it might be better
to leave sand and bucket in their original places and read the paper that Per Bak and
Kan Chen published in Scientzfic American on the topic of self-organized criticality
(Bak and Chen, 1991). In the paper they describe, among others, a comparable
sandpile experiment performed by Glenn Held and his colleagues at the IBM Thomas
J. Watson Research Center. If sand is dropped grain by grain on a pile, each grain will
cause some grains to slide down, but every now and ,then a single grain may cause a
miniature avalanche and it also happens sometimes that part of the pile slides away in a
major avalanche (all relative to the size of the pile, of course). The avalanches are
qualitatively similar, but quantitatively different. The great majority are very small,
involving only a small number of grains. Some are considerably bigger and a few are
big enough to affect the form of the pile.
Put differently, the events that change the form of the pile are fractal (avalanches in
avalanches in avalanches, etc.) and scale invariant; the small ones are qualitatively
similar to the big ones. The causes are quantitatively similar overall, namely the falling
 Transitions and developmental fractals
of a single grain of sand. Bak and Chen describe the sandpile as a self-organizing
system, which evolves towards a state of subcriticality. Subcriticality is the state in
which an event (just adding another grain of sand) that would otherwise have only a
small effect (some grains sliding down) can cause a major shift to occur (an avalanche).
Two additional facts are well worth noting. The first is that the structure we observe
in the sandpile occurs in a large variety of natural events, ranging from earthquakes to
traffic jams to simulated biological evolution. It is, in other words, a widely distributed
event form and applicable much more widely than to the sands of the North Sea or
Long Island Beach. Second, an interesting numerical relationship exists between the
frequency of the events (the avalanches) on different scales. If we plot the logarithm of
the number of events in each size class (number of very very small ones, number of very
small ones, number of small ones, etc.) against the logarithm of the magnitude
(magnitude of very very small events, of very small events, etc.), we obtain a tilted line.
This means that small changes occur in big numbers and that big changes are few in
number. Moreover, the ratio between the numbers is fixed (within statistical
boundaries). Let us now turn to the domain of psychological development and see what
sandpiles and subcriticality have got to do with developmental states and stages.
In the preceding sections, I have argued that development probably follows a fractal
path. That is, the qualitative structure of developments is similar for all possible
time-scales. Developmental time-scales may range from very short to very long
periods. For instance, at the lower end of the scale we may find single perceptions or
thoughts that usually need only a few seconds to come about. Problem-solving events,
such as the tower of Hanoi, require a time-scale of minutes. Learning experiments that
make use of the so-called microgenetic method (Siegler and Crowley, 1991) last for a
few weeks. Major shifts in understanding, such as the skill levels described by Kurt
Fischer, require several months to several years, depending on the time in
development they occur. At each of these levels the same quantitative patterns occur,
be they real transitions, growth spurts or slow linear change. The many little events
that occur daily and involve simple problem-solving activities form an interrelated
structure. A discovery in one little problem may affect a potential discovery in another
little problem, although the relationship between any two arbitrary events is probably
quite weak. Such problems may, however, be connected via links that Case calls central
conceptual structures (Case, 1992b). Domestic problems such as clock reading and
dividing a birthday cake are related to these structures, in that each of them weakly
positively contributes to that structure. In turn, the structure has a strong positive
contribution to each of the separate smaller problem domains, which makes them grow
in the form of a marked spurt.
If problem-solving activities, thoughts, perceptions and the applications of skills
form an interrelated fractal event structure, there is no reason to doubt that this
developmental structure too behaves like interrelated fractal event structures in
general. Similar to the sandpile, the daily activities make the structure move to a state
of subcriticality. This is a state in which an event that would have caused only a small
change under normal circumstances now causes a major shift. The number and
magnitude of those major shifts are statistically fixed, but they depend on the actual
 Fractal development and the quantity of transitions 135
time-scales at which shifts occur. For instance, getting to understand the tower of
Hanoi problem is a major shift compared with the many little shifts during the
preparatory stage in the form of much smaller trials and discoveries. T h e shift from, for
instance, having an abstract understanding of an arithmetic operation to an abstract
understanding of the relationship between different arithmetic operations, is a major
shift in comparison with the smaller ones that occur on a time-scale of weeks or
months. Despite major differences in the magnitude of the shifts, however, the
magnitude of the causes remains the same throughout the process.
Let me try to use the idea of subcriticality in the cognitive system to arrive at a
tentative explanation of why only a limited number of major developmental transitions
exist. Most theories, such as those of Piaget, Fischer, Bruner or Case, agree on three to
four main shifts, dependent on how they are defined. Fischer, however, usually makes
a finer distinction that he calls 'levels' (Fischer, 1980). Thirteen such levels emerge
between birth and the age of approximately 25 years.

Fractal development and the quantity of

transitions
Why are there only a small number of transitions?
The main transitions, such as the shifts between stages or levels of thinking, seem to
occur beyond direct human control or external causation. They emerge by themselves
and not because an external agent such as a teacher or a parent pushes the child into a
new mode of thought. Small transitions, such as the child's understanding of a
particular arithmetic problem, on the other hand, are quite directly manipulable. You
can explain to the child how to solve the problem step by step and see to it that it really
understands what you mean. Given that parents and teachers control a learning
environment in which a certain number of small transitions are supported, how is it
possible that the large-scale developmental process boils down to just a few large
transitions?
I shall assume that cognitive transitions have a certain magnitude. Just pretend that
a certain kind of small transition, such as understanding specific problems, has a
magnitude of 1/10 on the scale of Piaget (people studying earthquakes have the Richter
scale, so why should developmentalists not have their Piaget scale?). T h e number of
events with magnitude M (which is 1/10 in our case), occurring in self-organizing
complex systems of the type discussed here, is given by the equation: a

that is, the number of events of magnitude M is equal to a constant c times 1 divided by
the magnitude raised to a constant power p. Let us, just for the sake of the argument,
assume that c equals 1 and that the power parameter equals 1.72 (which is a value found
in some related systems; the choice, however, is completely arbitrary; see Bak and
Chen, 1991; Schroeder, 1991). If M is 1/10, and p is 1.72, the number of transitions of
136 Transitions and developmental fractals
magnitude 1/10 is about 52 per time unit. Assume that the time unit is a year. That is,
whatever the transitions are, they are of the type that takes approximately a week to be
accomplished. How many transitions per year would there be, on average, that are 10
times as big? We know that:

It follows then that:

that is, each year there is on average 1 transition of a magnitude 10 times bigger than
the transitions that take place about once a week. By the same token, the number of
transitions of magnitude 2 is 1 every 3 years, of magnitude 3 is 1 every 7 years and so
forth (Figure 5.6). The number of major transitions (e.g. of magnitude 3) that we can
expect over a lifetime is therefore very limited. Between the ages of 0 and 21 years, only
3 transitions of magnitude 3 will occur on average. Put differently, the limitation on the
number of major transitions follows directly from the power law that describes their
distribution over time. That power law, in its turn, is a direct consequence of the
eventual fractal nature of the organization of cognitive change.
What is the point of doing an exercise like the previous one? Recall that the
discussion started by asking why there was only a small number of major transitions.
Traditionally, the answer to this question has been based on ontological arguments.
For instance, there are four major Piagetian stages because they represent the four
major steps in the emergence of true scientific thinking. I t is the logic of scientific

Figure 5.6 If we assume that a transition of magnitude 0.1 occurs once every week, a
22-year-old person will have experienced about 1,200 of such minor transitions. The
expected number of transitions decreases drastically with increasing magnitude. That same
22-year-old will have experienced only 2 transitions of magnitude 4 (calculations are based
on the assumption that the power parameter is 1.72; note that the y-axis is logarithmic).
 Fractal development and the quantity of transitions 137
thinking that explains why there are four, not five or three, or approximately four. The
reasoning behind the current exercise was quite different. It started from the idea that
developmental systems have a fractal property. They are transitions in transitions in
transitions. If that is true, they should obey the rules of complex fractal systems that
change as a consequence of some underlying principle, which, in the case of
development, is logistic growth.
A basic rule governing such systems says that the frequency of the changes is a
power function of their magnitude. Given the fact that our daily activities and thought
processes are like small transitions or growth processes, we should expect a nested
structure of bigger and bigger transitions. The number of main transitions - that is,
transitions of maximal magnitude -must therefore be very small. More accurately, it is
small because of the power law that governs its distribution, but it must be big enough
to enable students of development to see a pattern. For some reason or other, the
number four seems an optimal combination of both requirements.

Between-transition intervals: evidence of a fractal process?

The empirical evidence I want to discuss here is indirect. It deals with the length of the
intervals between major transitions, in other words, with the length of developmental
stages. Developmental stages are a tricky issue. Many people believe they are
non-existent or highly dependent on the criteria and scales you employ. Stages are
surprisingly closely related to the eventual fractal nanire of development. I shall take a
very pragmatic stand and define a stage as any time interval during which a measurable
criterion shows a predefined amount of stability. For instance, we might take a
measurement scale for social understanding that applies over the whole life span and
administer it to children of different ages. If we find growth spurts or sudden changes
punctuating longer periods of relative quiet, we may conclude that the scale reveals a
stage-wise development. The claim does not extend beyond the measurement
procedure, context or task domain. Evidence of similar spurts and plateaux in different
domains lends more support to the assumption that the stages are general instead of
local and specific properties of behaviour or understanding. Fischer's work on
developmental stages, called levels, is an example of the procedure I just described. A
standardized scaling procedure which is applicable over the whole life span is
administered to populations of different ages and with different task domains. Fischer
and Rose (1993) present a generalization of findings from different domains that should
be applicable to a broad range of developmental phenomena. The stages (levels) and
age ranges are given in Table 5.1 below.
I simply took the mean of each transition age, which is indeed a very quick-and-dirty
procedure here and transformed them into days. I then graphed the logarithm (to the
base 10) of the transition ages. The result is surprisingly close to a simple linear
logarithmic distribution (see Figure 5.7). What is so special about this? Linear
logarithmic distributions are characteristic of many biological growth phenomena.
They are caused by growth processes that are, in some way or another, determined by a
component that obeys a power law. An example of such a process is the sandpile
138 Transitions and developmental fractals
Table 5.1 Fischer's 13 developmental levels with average ages of emergence

Level Age of emergence

Single reflexes 3-4 weeks

Reflex mappings 7-8 weeks
Reflex systems 10-11 weeks
Single sensorimotor actions 15-17 weeks
Sensorimotor mappings 7-8 months
Sensorimotor systems 11-13 months
Single representations 18-24 months
Representational mappings 3.5-4.5 years
Representational systems 6-7 years
Single abstractions 10-12 years
Abstract mappings 14-16 years
Abstract systems 18-20 years
Principles 23-25 years

Figure 5.7 A linear logarithmic distribution of inter-transition intervals, based on ages of

transition with Fischer7s developmental levels.

discussed earlier. It is interesting to observe that a phenomenon determined by both

biological and cultural variables, cognitive development, for instance, follows a rule of
distribution characteristic of biological growth phenomena in general.
The logarithmic distribution affords for another interesting observation that relates
to the problem of 'fractality'. Let me begin with a simple fractal based on the
Lindenmayer system. Aristid Lindenmayer was a Dutch biologist who conceived of the
growth patterns of plants as fractal, repetitive systems based on simple transformation
rules. In my fractal example, the transformation rule says that you can grow two
branches out of any available branch. The initial state is a single branch. Applying the
 Fractal development and the quantity of transitions 139
rule n successive times produces a tree with 2" branches. Since the tree has to increase
in size each time in order to remain strong enough to support the weight of the
branches, it takes more and more resources as the tree grows more branches. If the
resource supply is a constant (or at least a constant ratio of a variable such as the size of
the trunk), the time needed to grow a new generation of branches is proportional to the
number of branches it has to grow. The tree is a very characteristic fractal pattern and
its growth rule an example of a fractal process. If the time constraint applies, the time
between each new generation of branches increases in accordance with the power
series:
1+2+4+8+16+32-+64+ ...
which is similar to:

The logarithmic representation of this time series is a line, namely:

Any fractal process where the interval between generationsZ is a function of the
magnitude of the fractal pattern (the number of branches, for instance) shows an
interval pattern described by a power distribution. That is, its logarithmic representa-
tion is a straight line.
The fact that the intervals between major transitions lie on a straight logarithmic line
does of course not prove that development is a fractal process. The similarity between
the timing of development and that of fractal processes as described here might be a
mere coincidence. On the other hand, the logarithmic distribution is yet another clue
that increases Miss Marple's suspicion that development and cognitive growth is
indeed a fractal process.

What pushes the subcriticality into a major transition?

If you keep adding sand to a sandpile, a moment will come when just a little bit of sand
will make a major portion of the pile slide down. An event - adding a little bit of sand -
that caused minor changes causes a major avalanche when the sandpile has reached a
subcritical state. The physical parameter that determines the criticality is probably the
angle of the sand relative to the horizontal surface. What makes the critical shift occur
in development, provided development is indeed behaving qualitatively like the
sandpile? The answer should be that the,event that causes minor shifts will cause a
major shift when subcriticality occurs. That is, there is no particular group of
developmental mechanisms that is responsible for pushing the system into a major
transition. In Piaget's numerous works, lots of examples can be found of how the same
information that previously caused children to stick to their old beliefs, suddenly
changes those beliefs when a state of disequilibrium has been reached.
Siegler and Crowley (1991) describe a number of microgenetic experiments aimed at
letting children discover new strategies. The time of discovery and generalization
140 Transitions and developmental fractals
usually takes a few weeks. Such processes are believed to present a model of
developmental processes in a nutshell. The point that Siegler and Crowley make,
however, is that the explosive generalization of a new rule or strategy is usually not
simultaneous with the discovery of that rule or strategy. What are needed to set in
motion the generalization to the whole domain of problems are so-called impasses or
challenge problems. They are the kind of difficult problems that show children that the
new strategy they considered as only one of the many possible strategies, is actually the
only really good one. Put differently, the discovery of the strategy or rule brings the
cognitive system into a subcritical state. Encountering of a challenge problem is the
external event that causes the system to pass the critical threshold. Explosive growth of
the use of the strategy is the result. The point is, however, that a problem becomes a
challenge problem once the subcritical state has been reached, although, of course, not
all problems become challenge problems at that stage.

A mathematical model of transitions

The logistic transition
The standard logistic equation is not suited for modelling sudden, steep jumps that
follow a period where almost no change took place at all. And the latter is what happens
in conservation development or in the tower of Hanoi problem and probably in all
processes where sudden developmental state shifts occur.
In our examples of strong transitions, ranging from perceptual shifts in Necker
cubes to achieving an understanding of the tower of Hanoi to conservation, we have
seen that the initial state is at first very hard to change. It tries to maintain its stability
by strongly reducing the probability of events that will cause the state to shift to a
different value. Note that this is just a figure of speech: the subject in the tower
experiment, for instance, really tries very hard to understand. Nevertheless, helshe
remains unsuccessful for the whole preparatory period, in which no real progress is
actually made. The same is true with conservation; whatever happens with the form of
the clay ball or the liquid column, it confirms whatever belief the child has about the
amount being conserved or not. But once the initial state gets pushed out of its attractor
basin, the change enters into some sort of positive feedback loop. The further one
moves from the stable initial state, the quicker the changes take place, until the process
suddenly halts at its new attractor point, which is the alternative perceptual
organization, the understanding of the tower strategy or conservation, just to name
only two of the possibilities.
What causes such stable states to change? Take for instance conservation.
Conservation training is known to have hardly any effect, unless the child is already at a
subcritical state (see van der Maas, 1993, for a recent demonstration). How do children
ever learn that the amount of liquid does not change in spite of salient form changes?
Children are very often confronted with conservation situations. Any transformation
of the form of a liquid (milk, lemonade or whatever) or of solid matter (sand, clay, your
 A mathematical model of transitions 141
blanket, etc.) potentially involves a conservation problem: does the amount remain the
same, or is it affected by the form? In the majority of cases, the eventually problematic
nature of simple domestic acts, such as the child eating mashed carrots and potatoes or
smearing them around the dish, goes unnoticed. Consequently, there is no conscious-
ness whatsoever of an eventual problem involving increase or decrease of matter upon
&anging its form, and, therefore, also no real causative influences that push the belief
system into the conservation state.
How does conservation develop, then? Whatever the basis of conservation
development, or of many other transitory developments for that matter, it is highly
likely that the prompts, conflicts or experiences that push the initial belief or strategy
out of balance are randomly distributed. That is, there is no systematic educational
effort to teach children conservation around age 5 . Some schools, believing in the
Piagetian idea, present conservation materials in kindergarten, but children discovered
conservation even before Piaget did. More important, the frequency of effective
experiences is inversely related to the degree of consolidation of the belief. I t is well
known in conservation research that 4-year-olds, for instance, can be given quite
strong counterexamples, but they simply do not see them as conflicting with their basic
belief; they may change their answer in accordance with the suggestion, but fall back on
their old answer at the next question. On the other hand, with children who have
started to doubt their non-conservation answers, it suffices just to say they are wrong
and they will immediately switch to conservation, giving perfectly correct logical
justifications (Brainerd, 1973). Conservation training experiments show that the lack of
training effects is proportional to the lack of understanding of conservation.
For the next step, I need to introduce the notion of a learning event. A learning event
is defined as an experience that, for some reason or other, affects the child's under-
standing or skill level. It could do so for a variety of reasons. For instance, a learning
event could be the simple opportunity to experiment with the problem. It could be a
problem solution or a reinforcement given after a successful try. It could be a situation
that puts the child in a cognitive conflict. It could be a combination of a cognitive
conflict arising between two children who solve a common task or problem. In short, a
learning event is something whose actual content and effect vary over a wide range of
learning, promoting variables and properties. An event that would qualify as a learning
event with a child who is close to acquiring an understanding of the conservation
principle, probably has no effect at all on a child who has no idea of conservation
whatsoever. In fact, the further the child moves from the original non-conservation
attitude, the higher the probability that any event involving transformations of quanti-
ties will have a learning effect on conservation understanding. All this can be generaliz-
ed in the principle that the probability of learning (of encountering an effective learn-
ing event) is proportionate to the level of understanding or mastery already achieved. A
low growth level thus corresponds to a low probability that an effective learning event
will take place, that is, with a low chance that anything will b; learned. The difference
with the logistic growth model is that in the latter, the probability of learning is
independent of the growth level. What depends on the growth level is the net effect of
growth, since that is the product of growth level and growth rate.
142 Transitions and developmental fractals
What makes an ordinary event, like drinking your milk or playing with sand and a
bucket, qualify as a conservation learning event? We have seen earlier that problem
events are determined by a wide variety of control parameters. They cover aspects as
diverse as the child's interest and attention to the problem, the salience of the
dimensions of the beakers or forms employed, the support and help of a more
competent child or adult and so forth. With a specific conservation problem, the child
can be highly involved or not, the difference between dimensions can be more or less
salient, the adult can give more or less accessible support and explanation. Each of
these control variables varies randomly. For a non-conserver all the control variables
have to be beneficial for the event to qualify as a learning event (high interest and
motivation, attention, high support, salient distinctions, a simple problem and so on).
Given the random variability of all these dimensions, the chance that they will all have
the required value is fairly low. With children who are closer to an understanding of
conservation, it probably suffices that one or a few of the variables are beneficial, but it
does not matter too much which ones they are.

A mathematical model of the logistic transition

I shall start with the ordinary logistic equation. Recall that each iteration can be
conceived of as a separate learning step. I shall assume that each step now represents an
event that potentially qualifies as a learning event. Examples of such events are the
child's pouring liquids or sand from one container to the other, playing with clay and so
on. The steps in my logistic equation do not represent fixed temporal intervals now, as
they did with the feedback delay, but intervals between events; the feedback delay time
is now determined by the time between consecutive events that potentially qualify as
learning events. From the previous section, I know that there is no growth when the
event is not a learning event. How do I model the occurrence of a learning event that is
proportionate to the growth level already achieved? I have claimed that the control
variables determining a problem event vary randomly. Thus each dimension of the
control space corresponds with a random number. For the sake of simplicity, I map all
these random variables onto one concatenated variable. The random number that
results is needed to determine the probability that an event qualifies as a learning event.
Let me assume that the random number varies between 0 and 1. I know that when K i s
set to 1, the growth level of conservation understanding ranges between 0 and 1. The
simplest way to simulate the occurrence of a learning event, therefore, is by using the
following rule. For each iterative step in the computation of the growth level of
conservation, I pick a random number (between 0 and 1) and compare this with the
growth level (also between 0 and 1). When the random number is smaller than the
growth level, I decide that this particular iterative step corresponds with a learning
event. All these assumptions go into the following simple growth equation:
L,,, = L;(1 + e, - e;L,/K)
for e, = 0 if L, < Rand, K, and e, = r if L, 2 Rand,. K
I shall call this the stochastic transition equation. It is stochastic because it is based on a
 A mathematical model of transitions 143
sequence whose effect depends on the preceding steps in the sequence. ~t is a
transition equation because it is able to produce a sharp, jump-like transition, as we
&all see in Model 5.1.
It is easy to proceed from this stochastic version to a deterministic form of the
logistic transition. From the second part of the equation follows that the growth rate e,
is a linear randomized function of the growth level L, and some fixed parameter r:

What is the simplest form in which the functionf can be cast? Since the growth level L
can vary widely, depending on the value of K, we make e, a simple multiplicative
function of r and of the ratio between L and K, that is:

When I substitute this function in the first part of equation 5.5, I obtain the
deterministic form of the transition equation:

Model 5.1 Jumps: stochastic and deterministic

In this model, we are going to define an equation that enables you to experiment with
either the stochastic or the deterministic version of the transition equation by just
changing one parameter.
First, define the following range values and'range names (the range names are in
boldface, the rest of the text is commentary3):

1 r (growth or learning rate)

0.01 in (initial state of grower)
1 s (if s = 1, stochastic version works)

Go to cell A10 and type in. Cell All takes the basic equation:

Note that the growth rate variable is to be found in the B column. Enter the following
equation in B11:

It says that if the stochasticity parameter is 1, the value of the growth rate is to be found in
cell C11. If it is not 1, the value is to be found in Dl 1. This seems a rather cumbersome way
of specifying the transition equation, but i l will make the steps more transparent.
In cell C11 type:

which states that if a random number is smaller than Al0, the cell takes thevalue of r and if
not, it has the value 0.
Cell Dl I specifies the value of the growth rate variable for the deterministic model:
144 Transitions and developmental fractals

Figure 5.8 Sudden jumps to a higher state level depend critically on the value of the
initial state (all other things being equal).

Copy range A1 I..D11 to A12..D1000 and define a simple line graph for the data range
A1 O..AI 000.
Try different values for r between 0.1 and 2.8; with an r of 1, test the effect of the initial
state value (try for instance 0.001, 0.01, 0.05 and 0.1).

What kind of growth patterns result from applying the transition equations? Let me
take the pattern of change in the tower of Hanoi experiment (Bidell and Fischer, 1993)
as an example of a sharp transition. It turns out that the curves based on a sufficiently
low initial state (around 0.01) and a high growth rate (around 1) reproduce the tower of
Hanoi pattern quite well (Figure 5.8). The initial level remains stable for quite some
time, then suddenly switches to a much higher level of functioning. This level is the
carrying capacity level, which is by definition the optimal level of functioning. The
point in the curve preceding the sudden jump corresponds with the state of
subcriticality from the fractal systems described earlier. Second, the length of the
initial state, corresponding with the preparatory state, is proportional to the initial state
level, corresponding with the entry level in the experiment. If we try a relatively high
initial state, like 0.1, the transition takes place immediately, as it did with the oldest
children in the experimental group. Third, if the initial state becomes too low, the
length of the preparatory state far exceeds any possible problem-solving task,
especially with younger children. It follows then that such entry levels will not lead to a
discovery of the right stacking principle in the tower task.
With lower growth rates and higher initial states, the resulting curves are quite
similar to the ordinary S-shaped curve of the standard logistic equation. With high
 Transitions as catastrophes 145

Figure 5.9 With similar parameter values, the stochastic transition model produces a jump
at dramatically different moments: small, continuous random effects produce major,
non-continuous transition effects.

growth rates qualitative patterns emerge that we have also seen with the standard
equation, namely oscillations and chaos.
The stochastic model produces the same sudden jumps as the deterministic version
(provided the parameter values are the same; see Figure 5.9). The stochastic element
only shows in the moment at which the sudden jump occurs. Whereas the stochastic
element appears in each step of the equation, its effect is confined to a specific small
time range, the moment at which the jump occurs. In order to investigate the extent to
which the moment of the jump is determined by the random element, I ran 1400
simulations of the stochastic model, keeping the parameters constant. The cumulative
curve, showing the number of simulated cases which have reached conservation
criterion is a typical S-shaped curve which resembles the group data reported earlier
(see Figure 5.10). The simulated group data, however, are based on individual growth
curves characterized by a sudden leap.

Transitions as catastrophes
The 'catastrophic' nature of transitions
Catastrophe theory is a branch of mathematics that came into vogue in the social
sciences after the publication of an influential overview paper in Scientific American by
Zeeman (1976). I t presented a mathematical justification for the existence of sudden
changes in, for instance, the behaviour of dogs or people. If a dog is made really
nervous, the animal will either attack or flee from the cause of the disturbance. My dog,
146 Transitions and developmental fractals

Figure 5.10 The cumulative version of the curve from Figure 5.9, representing the
number of 'children' who have reached the criterion: although each individual develops in
the form of a sudden jump, the group curve shows a misleadingly smooth, S-shaped
growth pattern.

who is very clearly misnamed after a peaceful Belgian cartoon character Guust Flater,
who is, I am afraid, not internationally known, has now on several occasions given me
the dubious honour of studying Zeeman's catastrophe 'in the wild'.
It is very easy to make Guust anxious, for instance by putting his leash on in the
wrong way, exerting too much pressure on his neck. He will try to flee from the
threatening feeling, that is, away from the person who holds the leash. This response
will, of course, make matters worse. Suddenly his anxiety turns into aggression and he
attacks whoever is holding the leash. The attack will loosen the pressure on the leash
and reduce the cause of his rage, but this is certainly not enough to calm him down. It
really needs a lot of dog psychology, associated with the right sort of physical
aggression, to get him back to normal. I have never tried to measure the pressure of the
leash around his neck (I am crazy enough to have a dog, but I am not mad), but if I had
done that, and plotted the flight or fight reactions against the pressure, I would very
probably have found that the amount of pressure at which he attacks is higher than the
amount of pressure on his neck at the moment his fear turns into rage.
The sudden shift from fear to rage and attack, or from attack to fear, is an example of
a so-called cusp catastrophe. The cusp catastrophe is a mathematical form that
describes the distribution of equilibrium points. In the case of my ferocious dog, the
equilibrium points are different forms of behaviour associated with points on the
continuum from fear to rage. His behaviour is either a neutral reaction that is neither
flight nor attack, or a choice between flight and attack. Figure 5.11 shows the sheet
upon which the equilibrium points are located, relative to a major control variable,
namely the fearlrage-inducing variable.
The transition model discussed in this chapter pretends to explain sudden jumps in
 Transitions as catastrophes

Figure 5.11 The plane of behavioural equilibrium points - fight, flight and neutral - and
their distribution over a control variable takes the form of a cusp.

a bimodal variable, such as conservation. Does it have the properties specified in the
cusp catastrophe model? What are these properties and how do they apply to
development?

Catastrophe flags on developmental transitions

Amsterdam is, as is widely known, a city rich in distractions of the most various kinds.
Maybe because they wanted to compensate for too tempting outdoor activities, my
distinguished Amsterdam colleagues Han van der Maas and Peter Molenaar have
turned to the study of catastrophes, a most attractive topic, indeed, but of a less sinful
nature than some of the excitements their city has to offer. In a recent paper (van der
Maas and Molenaar, 1992), they describe stagewise development in terms of
catastrophe theory. The catastrophe model they apply is the cusp discussed in the
preceding section. Stagewise development occurs in models like Piaget's, where it is
supposed to cover all domains of thought, or in a model like Fischer's, where the stages
are strongly domain and context determined. Van der Maas and Molenaar claim that
stages succeed one another in the form of real shifts. The shifts are not growth spurts,
as Kurt Fischer or I would see them, but real switches as described by the catastrophe
model. Just as my dog switched suddenly from flight to fight, the sensorimotor child,
148 Transitions and developmental fractals
for instance, would make a sudden shift to pre-operational thinking, if the Piagetian
model were right. The Netherlands is anything but a conservative country, but we do
like conservation, and conservation therefore is the example van der Maas and
Molenaar employ to show that the shift from non-conservation to conservation
understanding is a real catastrophe (in the mathematical sense, of course).
In essence, their argument is as follows. The mathematical study of the cusp
catastrophe shows that it has eight distinctive features, that is, properties invariantly
associated with the occurrence of a cusp. These properties are called 'catastrophe
flags'. If it can be shown that a developmental transition carries these flags too, it is
likely to be a catastrophe (since no other change phenomena are known that carry just
these flags). Thus, instead of the four horsemen of the Apocalypse, we have the eight
flags of the catastrophe to watch for. In his dissertation, Han van der Maas (1993)
describes a longitudinal study of conservation development with a computerized test.
Since the children had free access to the computers in the classroom, van der Maas was
able to collect individual data series with high measurement density.
The first flag is the bimodal score-distribution. Children presented with a set of
conservation questions tend, in general, to have the majority either right or wrong. We
find relatively few children with an intermediate score. Evidence for this comes, among
others, from the Bentler data (Bentler, 1970) described earlier. The second flag, related
to the first, is inaccessibility. A particular score domain, of intermediate scores, is
'inaccessible', that is, it will not appear in the data. The relative lack of real
intermediate scores supports this criterion.
The third flag is the occurrence of sudden jumps. It is not clear whether the
development of conservation shows a real sudden jump, in the sense of a sudden
switch. It is likely, however, that the growth pattern is at least one of a marked spurt.
Van der Maas found strong evidence for each of these three important flags in his
study. Figure 5.3, for instance, shows the sudden jump occurring with the so-called
transitory subjects, children who are caught in the middle of the transition process.
The fourth flag is hysteresis. It implies that the place where the jump occurs depends
critically on the direction of change of the control variable. Van der Maas (1993) found
indirect evidence for hysteresis in some of his transitory subjects. He states, however,
that conclusive evidence for hysteresis can only be found with an experimental set up,
in which an important control variable can be varied continuously (see further, Figure
5.13).
A related, fifth flag is divergence. It implies that the magnitude of the jump, that is,
the effect of a small change in the vicinity of the transition point, differs over the second
control axis (which in Figure 5.11 would be the axis perpendicular to the
fearlrage-inducing axis). In the conservation example, this second dimension could
well be the amount of support and help during the test. We have seen that real
developmental spurts occur only in the case of supportive testing. In unsupported
testing the increase is slow and gradual. The evidence from van der Maas' experiment
is unfortunately again inconclusive.
The next, sixth, flag is divergence of linear response. It implies that perturbations in
the vicinity of a catastrophe point will lead to sudden loss of stability and oscillations.
 Transitions as catastrophes
The perturbation that best models this type of event is conservation training, that is,
presenting the children with a conservation training programme. It has often been
observed that almost all training works very well and very explicitly with children who
are already in the transitional state and that it does not work for children who are true
believers in non-conservation (Bijstra, 1989).
The seventh flag is delayed recovery of equilibrium. It is based on the measurement of
the time needed for the system to return to its original equilibrium if it is artificially
For instance, you can tell the child that another child gave a different
answer to the same question. We expect that a counter-suggestion will hardly affect the
child if it is far from the transition point to conservation. But, as it approaches that
point more closely, the same counter-suggestion will make it think and hesitate longer.
This scenario certainly sounds plausible, although there is no direct evidence available.
A possibly related effect has been observed in Bidell and Fischer's tower of Hanoi
experiment. Just before the jump to complete understanding takes place, many
children pause, tap their fingers and the like. It seems that the information they
rejected in the previous stage is now causing more and more cognitive trouble.
The eighth and final flag is anomalous variance. As the transition point approaches,
the structure of the statistical variance changes markedly. That is, the nature of the
differences between the children changes as they are ready to make the jump to
conservation. Van der Maas and Molenaar (1992) discuss several examples of this
anomaly, such as the occurrence of oscillations in the answers, or discrepancies in the
reasoning. In van der Maas' experiment no evidence was found for flags six to eight,
although there is indirect evidence in his results and the results of others to show that
those flags might apply.
In summary, despite lacking evidence for a number of catastrophe flags, there is very
good evidence for three essential ones, namely bimodality, inaccessibility and sudden
jump and reasonable evidence for hysteresis. I would say that this evidence is strong
enough for us to place a higher bet on the catastrophe than on the non-catastrophe case.
Some catastrophe flags have been shown in empirical research, others sound quite
convincing, yet others are plausible extensions of findings in other domains. The
question now is whether the flags apply also to the transition model presented in this
chapter.

Is the logistic transition model a catastrophe?

If you do not like mathematical models, I am sure your answer will be yes, of course it is
(but I am also sure your reasons for saying so will be based on a different concept of
catastrophe). Let us turn to the officialmeaning of the catastrophe concept, however,
and investigate whether the flags apply to the transition model presented in the form of
equations 5.7 and 5.10. For a specified parameter range those equations produce
typical transition curves. Let me pretend that those curves are simulated data from a
longitudinal conservation study. Do the catastrophe flags apply to them?
Assume that we take an age distribution for the children that lies around the
transition age, for instance between 4 and 6 years; that is, the time window of the
 150 Transitions and developmental fractals
sample is 2 years. Given the steep form of the transition, we may estimate the actual
transition to take place within a much smaller period. For instance, let us assume it
takes about 10 per cent of the age range over which we sample (that is, on average 0.2
years or roughly 2 i months). The probability that we are going to find a child who is
right in the middle of a transition is 10 per cent.
Consequently, the probability that we shall find a child in the state of either
non-conservation or conservation will be 45 per cent. Put differently, we will find a
highly bimodal distribution. The intermediate scores on the test (some items correct,
some wrong) will occur very infrequently in the sample. Children who are right in the
middle of the transition may respond inconsistently over the items, but they may as
well be consistent in either a conservation or a non-conservation strategy. That is, a
large part of the score variation (the intermediate scores) is inaccessible. Third, the
developments take place in the form of a sudden jump. We have seen, however, that if
many individual scores are added the underlying spurt turns into a much more gradual
S-shaped growth curve. The averaged group data will definitely obscure the transitory
character of the change, but we know from each individual simulation that the
transition is a very sharp growth spurt. In summary, three important catastrophe flags
are found in the transition model.
What about hysteresis, the fourth flag?Figure 5.1 1 shows how hysteresis occurs over
the 'fold' of the equilibrium sheet. If we move to the back of the folded sheet, the
hysteresis effect becomes smaller and smaller and finally disappears. In the example
with the dog, the left-to-right variable was the fearlrage-inducing factor. The
front-to-back variable was assumed to be the dog's level of excitement. In the
transition model, the variables are of course very different. From left to right runs the
time variable. From back to front runs the growth rate variable, with the lowest growth
rate in the back. The problem is that, whereas the left-right axis could be varied in
either direction in the case of the dog's attack catastrophe, this is not possible in the
case of time (trivially, you cannot run time backwards). That is, there is no empirical
interpretation of eventual hysteresis phenomena in case the control variable is time,
which it is in developmental transitions. I discussed this matter with Han
'Catastrophe' van der Maas, and he showed me a mathematical technique for
demonstrating eventual hysteresis effects in dynamic models. In essence, it amounts to
letting an external factor affect the critical part of the dynamic equation, for instance,
by adding an increasing value to each outcome of a sequence of iterative steps. You
then let the sequence run backwards while adding an external factor that decreases in
the same way as the first one increases. I t is possible to compute the range of external
variable values for which hysteresis can be found.
The result of this can be seen in Figure 5.12 where one line represents the forward
and the other the backward loop of the external variable. The model simulates what
would happen in a microdevelopmental investigation of conservation. For instance,
when a critical control parameter, the difference between length and height of the
containers involved, is varied from highly different to about the same, the child would
persist in non-conservation until it reached a pair of beakers that were virtually
identical. But if the difference between the dimensions is varied from almost the same
 Transitions as catastrophes 151

Figure 5.12 Adding an increasing or decreasing external effect to the transition model
results in hysteresis: the growth levels either drop back to or escape from a minimal level
at different places, which depend on the direction in which the control variable has been
changed (left-to-right versus right-to-left in the diagram).

to very different, the child would first conserve, then switch to non-conservation with a
pair of beakers that are very different in shape. In summary, although hysteresis does
not apply to the transition model when time is taken into account, it does apply when
you simulate the effect of an external control factor. That is, the fourth catastrophe flag
applies to the model (Figure 5.13).
The next catastrophe flag, divergence, marks the difference in effects of small
perturbations, depending on where in the catastrophe model the perturbations are
applied. Recall that I suggested employing the growth rate variable as the
back-to-front dimension in the transition catastrophe. If the growth rate is small, a
single step forward along the transition curve (which is then actually very similar to a
smooth growth curve) exerts no significant effect. As the growth rate increases,
however, the same small step, if it occurs near the transition point, will magnify into a
sudden leap to the higher equilibrium state.
I t is difficult, however, to show whether or not the additional catastrophe flags,
divergence of linear response, delayed recovery of equilibrium and anomalous variance
apply to the logistic transition model. This model offers no description of the
underlying cognitive processes and these might well be needed to explain the effects of
external perturbations such as counter-suggestions on the child's thought process.
In summary, the transition model seems to fit a significant majority of catastrophe
indicators, the so-called flags. We may conclude, therefore, that it can be subsumed
under the broad class of models that display a cusp catastrophe and therefore describe
switches between bistable states.
Chapters 4 and 5 have been devoted to two different forms of increase, namely the
 .(aweu a % u e B~ u r p u o d s a ~ ~ oayl
a m d nod alayM auo
aq1 jo ~q8y.1a q l o IIa3
~ atp ' a ~ u w s u ~y o j1la3) lualag!p e uy d ~ e m a m w o 3m o d ~ n 'd( p ~ a ~ d e q 3
' p alou aas) laqaea p a q ~ ~ s aampa3oldp 3 n e u r o ~ n eaq3 q ~ saweN
y ~ a B u q auyap nod 31 .£
. , s a q x ~ e ~Mau
q ow s a m p o ~ d
q z ~ u e ~d ~qa n a'a3uelsuy
, .IOJ ' a m 0 aInl u o s e m l o j s u e u ayl Buyd1ddajo qnsaa ayl ST u o , z s ~ ~ uva 8-Z
- a m Bupunaq sdaay u o r ~ ~ a u u vo g~~ a g
atp p g e w e I .uo 8581 UOLJ lqsnv peq a q alaqm (slay10 j u o m e ~ l a ~ a l alnaiw aM aJaqM
'sad 'wn$31ag ur) 3 u a q g j o h y s l a n l u n ayl 3jaI a q l a y e JO alojaq d p ~ o q uaaqs aaeq z s n u 3eqL '1

. .
. U O I J ~ I ~ ~ ayl
~ S Odlaweu
. . p q y e~aq1.13sap IIeys I 'sarSo103a 1e~uamdo1anap3rmeudpjo anssr atpol
'330198urp11nq
..
%urpaa3oldalojaa -uo1~r~admo3 lo uoddns \enlnw jo sdrysuor~e1a~ uraualua saIqerlen
ayl alayM 'sar%o103a1e~uauzdo1aaapaqu3sap 11rM A a y ~-slapow palex~dwo:,alow jo
. . se m a s IIrM Aatp ' ~ a ~ e. ?u ~ ~UOIJISU~JJ
sy~oiq%urp~rnq o .j . . . ayJ pue ~ J M O J %3~srS01
:,IJSI%OI . .

.qed malagyp d ~ a aayl y q .KO~ J e I y q s JsourIe ayl q q ~

l a y p a ~ 3 . ~ ~auo
3 s JatpaqM u o puadap I I ~ M,raMsue u o p e n ~ a s u o 3ayl 03 10 w o y d w n l ayl 3ayl
pamnsse sl 31 . a 3 u a ~ a p pou 1sou11v 03 a3ua.1agyp m a g e s Alan u r o ~ jpalaplo ale s~auyemo:,
I s ~ a u y e m ojo~ slrad malaglp q l r ~suopsanb u o y e ~ ~ a s u joo 3 saylas palaplo u e q q u~a ~ p l ~ q z ~
B u n u a s a ~ dd q palallsuourap a q ue3 durn[ uopen.1asuo3 a q uy S I S ~ J ~ ~ €?'S S ~ Ha m % ~

u
p S U O.I .J I S U ~ ? J ~
sp?33t?gl ~ ~ u a m d o l a h aput? ZST
Oscillations

Oscillations in development
'What goes up, must come down . . .'
My wife and I share a little foible with a couple of our dearest friends: we love to spend
hours singing old 1960s songs. I admit it is a little pathetic, some old guys and girls
immersed in their collective adolescent nostalgia, but who cares - our neighbours live
at a safe distance.' One of our favourites is a song by Blood, Sweat and Tears that starts
with 'What goes up, must come down . . .' and it immediately goes on to offer an
almost Buddhist explanation in the second line, 'spinning wheel got to go 'round'.
If you are a developmental psychologist, you are probably particularly interested in
'what goes up'. If you are in the life-span business, 'must come down' should also
appeal to you. But if you are in dynamic systems, the spinning wheel should tell you
that things go up and down all the time and progress in spinning the wool is a result of
that incessant oscillation. Technically, a spinning wheel is a device in which a rotation
results from an up- and downward oscillation that is equal to the sine of the rotation
point. Sine functions are frequently used in dynamic models to introduce a regular
oscillator to some changing function. If you write a sine function in your spreadsheet
and graph a time-delay phase space (which was explained earlier), you will notice that it
forms a cycle.
I admit it is a bit tricky to take a campfire song as a source of deep wisdom, but there
is still an attractive aspect to the idea that things progress, if they progress at all, by
going up and down. In nature, it seems,,the shortest line between two points is the
wiggle. In this chapter, I shall first present some evidence of oscillations in
development and of the cyclical nature of resource allocation. The rest of the chapter
discusses a few speculative models of resource oscillations and their effect on
developmental and growth processes.
154 Oscillations

Oscillations in development
The research data abound with evidence for a wiggling developmental path, but for
some reason or other, we have tended to overlook that fact. Development is usually
seen as a great march towards a higher way of being. Whatever irregularities occur
along the way are not caused by the marching of the troops but by unintended external
disruptions. Nevertheless, the oscillatory and fluctuating movement could as well tell a
basic fact about the way change in nature is organized. It is possible that processes just
need to go up and down in order to maintain their dynamics over time. Let me first
discuss some empirical evidence for oscillations in various developmental domains.
In Chapter 4 (on the logistic growth equation), I discussed Dromi's study of the
lexical development of her daughter (Dromi, 1986). The cumulative data, referring to
the increasing number of words the child used and understood, showed a nice logistic
S-shape. But, if we graph the growth rate per observation period, the image changes
dramatically. What we see is an irregular fluctuation over time. The first substage of
learning words is characterized by rather large swings, whereas the second substage
shows much smaller oscillations (Figure 4.7). In a comparable study, Corrigan (1983)
found similar oscillations in the rate of word learning. After an initial period, in which
the first elaboration of the lexicon takes place, the increase changes into an irregular
fluctuation (see Figure 2.11). The net result, the additive growth curve, however,
closely resembles a smooth S-shaped curve. The additive nature of word learrking
conceals the fact that the increase process itself oscillates quite markedly.
Oscillations also occur in other domains of language development. Take for instance
the growth in correct use of a syntactic rule over time. Data from studies on
interrogative questions (Labov and Labov, 1978) and on inflections, such as the
present progressive (Brown, 1973), show a similar qualitative pattern. The average
number of structures used correctly varies widely over time, until it settles down at an
equilibrium level that is 100 per cent correct (Figure 2.10).
Kurt Fischer's studies of the growth of understanding in various cognitive domains,
such as arithmetic understanding or reflective judgement, show strong oscillations. For
instance, in a study on reflective judgement, children were given an ill-structured
problem and were asked to reason about their thinking process. Two different test
scores were compared, one score on understanding abstract mappings, another on
abstract systems. In abstract mappings, children justify their thinking by referring to
the context in which the problem occurs. Abstract systems require a coordination of
mappings and require the child to compare and contrast different contexts, or different
sides of an issue. The increase in the scores oscillates over the zero line; that is, there is
real regression (Kitchener et al., 1993; see Figure 6.1). The pattern is highly similar to
that of arithmetic understanding, more precisely the child's understanding of how
operations such as addition and subtraction relate (Fischer, Kenny and Pipp, 1990; see
further, Figure 6.2).
The data are based on cross-sectional research, not on individual case studies, as in
the examples from lexical and syntactic development. Nevertheless, the group data
must be based on individual data that show at least the same sort of variation over time,
 Oscillations in development 155

Figure 6.1 Oscillations in the yearly increase in scores on reflective judgement (after
Kitchener et al., 1993).

Figure 6.2 Oscillations in difference scores of arithmetic understanding (after Fischer et

al., 1983).
156 Oscillations
although individuals may vary as to the amplitude of those oscillations (see Figures 6.1
and 6.2).
Oscillatory increments are also found in brain growth. Matousek and Petersen
(1973) measured changes in the relative power of the alpha frequency band for the
occipital-parietal region and found a series of characteristic oscillations between the
ages of 1 and 20 years (see Figure 6.3). As Fischer and Rose (1993) have noticed, these
changes show a striking correspondence with periods of rapid change in metabolic and
ERP measures. In a recent series of studies, Bob Thatcher has shown that the process
of cortical organization, far from taking place in a linear and continuous manner,
follows a complicated cyclical path, with several reorganizations that repeat the
structure of the preceding cycles (Thatcher, 1992).
In summary, data from various domains show that the process of growth is not a
simple increase, but it takes place in the form of incremental steps that are very clearly
oscillatory. Before trying to explain where these oscillations come from, I shall take a
look at a domain where you would intuitively expect marked oscillations, the energy
and attention resource functions.

Cycles of resource allocation

In Chapter 2, I discussed several forms of long-term cyclical change found in models
such as Piaget's or Fischer's. Next to these long-term cycles, there is a vast range of
shorter cyclical processes, related to the way in which resource functions are allocated
to human activities. The most general and encompassing timing of resource allocation

Figure 6.3 Changes in the relative power of the alpha frequency band for the
occipital-parietal region (after Matousek and Petersen, 1973).
 Oscillations in development 157
is the pattern of waking and sleep, governed by the circadian rhythm of our biological
clock. The study of these biorhythms is a science in itself, and I shall not go into the
complexities of this field (see for instance, Winfree, 1987, for an introduction relating
the experimental study of biorhythms to processes of self-organization in coupled
oscillators).
Let us take a closer look at attention, which, as everyone would intuitively agree,
represents a main resource for learning and developing. Three aspects can be
distinguished (Olson and Sherman, 1983). The first is called arousal and it refers to the
overall level of activation, ranging from being asleep to being very active. It is related to
the amount of 'energy' a person invests in a specific task. The second sense of attention
is selection. Our brains are not capable of taking into account everything that occurs at a
given time, or that could possibly be relevant to a particular problem. People select
specific aspects or domains that they can keep track of, and these are highly limited
chunks of the information available or potentially relevant. The third aspect involves
the notion of efort. Effort is closely related to arousal, but it contains a more
straightforward reference to the amount of mental energy - whatever the nature of that
may be - allocated to a specific task.
By their very nature, the three aspects of attention are oscillatory functions. Our
arousal state goes up and down with alertness and fatigue and is determined by the
interaction between physical events, effort spent on tasks and a biological clock that
takes care of the general pattern. If we focus on the selection of a particular learning
task or content, the oscillation occurs in the form of cycles of sensitization and
habituation. Since we are limited in the intake of information and the range of the
contents we focus on at a time, each specific learning content shows a pattern of
alternations between selection and deselection (when some other content is selected).
For instance, a child's learning of arithmetic is anything but a continuous activity
function (and in some children it is very discontinuous, because they do not like maths
at all). It is an alternation of short periods of selective attention and effort and
comparatively long periods in which a child allocates its interest and effort to other
subjects. It follows then that since the growth rate of arithmetic learning probably
depends on resources that oscillate, parallel oscillations of the resulting growth process
may be expected. Finally, effort too is based on limited energy resources. Spending
time, attention and effort in carrying out a task leads to fatigue or even exhaustion and
then at least some time is required to recover and reload.
There is, as far as I know, hardly any evidence on the natural fluctuations of
attention and interest in individual subjects. The most revealing data I could find come
from a study that is almost one hundred years old'. In 1898, John Perham Hylan,
instructor in psychology in the University of Illinois, published a little monograph
supplement to the Psychological Review. (Hylan, 1898). He presented several
experiments on the natural fluctuation of attention, the fluctuations not determined by
external events and disturbances. The copy I read bears the bookmark of Gerard
Heymans, first professor of psychology in The Netherlands, who founded his
Groningen laboratory only a few years after Wundt. Hylan employed experimental
data as well as diary data and subjective ratings from his subjects. Hylan asked his
 Oscillations
subjects to add series of columns of numbers and observed the time it took them to add
each column and how many errors they made. In another experiment, he showed his
subjects - there were never more than a few - a series of pictures, once a day and over
periods ranging between one and two months. Hylan noted how much time his
subjects spent looking at those pictures. But he also asked a few subjects to select a
topic of personal interest and to write down every day to what extent they felt
interested in that subject, or liked or disliked it. The topics ranged from work, eating
and physical exercise to family members and the progress of a love affair. Just for fun, I
show a diagram of one subject's rating of how much (s)he liked two family members, x
and y, beginning with 29 December and ending 18 January, almost one hundred years
ago (Figure 6.4). It is interesting to note that all the curves are very similar, irrespective
of the domain to which they apply (love, arithmetic, etc.) and the time-scale at which
they figure. The pattern is always an irregular oscillation.
It is of course very easy to discard these and comparable data, because they are
subjective, or just accidental and coincidental variations, best explained by invoking
the notion of randomness. But the subjective nature of - at least some -of these data is
an essential feature and the explanation by reference to randomness is no explanation at
all, since 'random' simply means 'I do not know'. It goes without saying that some of
the variation is clearly coincidental. Hylan reports some accidental disturbances during
the experiments, for instance, that, in his words, exaggerated the fluctuations. The
basic pattern, however, is a genuine, irregular oscillation of the attention, effort and
energy phenomenon itself.
The patterns of resource allocation and the nature of the oscillations are subject to a

Figure 6.4 Changes in a person's affection for two family members (after Hylan, 1898);
the 'affection level' is the person's intuitive estimation of his current affection towards the
family member in question.
 The well model 159
long developmental process. Take for instance the pattern of arousal. Newborns show
a cyclical pattern of alertness states: regular sleep, irregular sleep, drowsiness, alert
inactivity, waking activity and crying (Olson and Sherman, 1983). On average, the
state of alertness in newborn infants does not last longer than about ten minutes. The
younger the infants, the more erratic the cycles of states and the shorter the duration of
alertness. The aspect of selection undergoes developmental changes too, as does effort
allocation. Children have to learn to organize their attentional function (Brown et al.,
1983). A particularly important aspect is the fine-tuning of attention allocation, for
instance how to adapt the level of effort to the degree of difficulty of the task. Another
aspect concerns the allocation of effort and attention to the various learning tasks that
children have to accomplish and to find a good balance between the tasks, dependent
on the degree of difficulty of the task and the child's own level of mastery. Attention
allocation is important because it involves the distribution of a main resource, effort
and attention, that directly affects the quality and speed of learning. In this chapter, I
shall present several rather speculative models that produce oscillations in a dynamic
way. The first is based on the metaphor of a resource well.

The well model

There's a hole in my bucket ...
When my father was a schoolboy, one of the horrors of his day was a maths problem
that involved exactly the kind of bucket described in the old song. It featured a water
tap from which flowed a certain amount of water per time unit in an attempt to fill the
bucket with the hole in it. The hole in the bucket had a certain diameter and, given all
this formidable information, my father and his peers had to calculate how long it would
take for the bucket to fill (or something like that; he must have been so horrified by this
problem that he never told me the exact details). A resource such as attention and effort
is like the leaky bucket. Assume the water is the resource, and the water that flows out
through the hole is the amount of resources needed. The tap filling the bucket with
fresh water can be compared to some sort of recovery function. The question is how the
opposing tendencies - depletion and recovery - behave over time.
A similar comparison invokes the metaphor of a well. The water is again the resource
function, comparable to effort or energy that a person invests in some task. Just assume
that Victor, a farmer in the San Joaquin Valley, needs a certain supply of water, say S,
to do a particular job, such as irrigating his vineyard. h he well on his farm is fed by an
underground stream with a volume R per time unit. If the vines need a lot of water,
Victor will soon reach a point where the well is empty, which obliges him to wait for a
period equal to SIR. A single irrigation round of the land will again empty the well and
our farmer will have to wait until there is again enough water tido his work, and so on,
ad infiniturn, since there is no end to the farmer's labour. Compare the water with effort
or energy, or whatever is needed to perform some task and assume that the task
requires a minimum amount of effort or energy to be performed well. If effort or energy
160 Oscillations
recovers more slowly than it is retrieved, a resource oscillation will occur similar to the
oscillation of the water from the well.
Let us assume, however, that the demand for a psychological resource, such as effort
and attention, is a function of the amount of resources available. That is, if you are alert
and energetic, there is most probably a higher demand than when you are tired or
drowsy. This follows intuitively from the fact that people initiate activities that require
an energy demand that suits their current state. Assume then that the initial amount of
resources available is R, and that the demand is equal to a ratio m of the available
resource:

and so on. For the sake of symmetry, let me assume that the rate of recovery r is also a
function of the available supply, that is:

What would happen if the recovery is not immediate, that is, if some recovery time is
needed for the supply to be refilled? If the recovery time is equal to a delay of d, the
equation then becomes:

Model 6.1 A resource well

Equation 6.3 is easy to model. Go to cell A1 00, for instance, to make sure you have enough
empty cells preceding the initial state point. You will need them to define the length of the
delay. Go to cell D100, write down the range name 'm' and enter the value of m in cell C100.
With Range Name Create (or any other equivalent instruction dependent on the
spreadsheet you use), define range name 'm' as referring to cell C100. Set A100 to 1 and
specify the following equation in cell A101:

Copy it to the range A102..A300. The recovery time is 25 time units, since each cell
represents a time unit. Note that in this model parameters m and r from equation 6.3 are
represented by one value (m). You can of course try models with different values, if you
define a range name r.
For m, try values ranging from 0.1 to 1. Figure 6.5 shows one of the curves, based on a
value of 0.3 for m. Try different delay times, for instance A90 or A60. Change the value in
cell A101 and copy it to the Al02..A300 block each time you make a change in the recovery
time.
Instead of applying the difference form you could also try a differential form of the
equation by employing the trapezoidal method described earlier. You might also want to
try another version of the difference equation. In its present form, the range of cells
preceding the initial state (A100) is empty. Go to cell A l , type in some small number (0.01
for instance) and write a sequence of logistic growth equations that fills the range A2..A99.
Take a growth rate that allows for a gradual increase over the whole range.
 The well model

Figure 6.5 Oscillations resulting from a resource based on the well model.

In summary, the well model of resource depletion and recovery produces an oscillating
pattern. That is, the resource level - take for instance a child's invested effort and
interest in a particular task - goes up and down over time. A peculiar aspect of this
model is that, as the depletion and recovery go on undisturbed, the process finally
moves towards an equilibrium point where recovery and use are in exact balance. What
does resource oscillation mean for learning or development?

Oscillating resources and cognitive growth

All other things being equal, the learning or growth effect of a task depends on the
amount of attention, effort or interest invested. With a child who is more alert and
interested, paying attention to what it sees and hears and willing to invest a lot of effort
in learning tasks, we should certainly expect better school results and more rapid
learning than with a less alert, less interested and l<essindustrious child. It is quite
natural, therefore, to relate the growth rate parameter in a cognitive growth function -
such as the one describing the growth of the lexicon - to the level of attention, effort
and interest spent during the actual learning process.
For the sake of simplicity, just assume that the growth rate r,is linearly proportional
to the resources invested in the growth or learning process. That is, the more resources
invested, the bigger the increase of whatever increases as a result of a particular
learning activity. Take lexical growth as the standard example. What is the form of the
growth curve that results from applying a growth rate that oscillates?
162 Oscillations
Model 6.2 Growth under oscillating resources
Our starting point is the difference form of the oscillating resource function demonstrated
in Model 6.1. It covers 200 consecutive steps (spreadsheet cells). For simplicity, let us
assume that the time range of these 200 steps is also the time range of the growth or
learning process that depends on the oscillating resource.
The resource -which could be attention, effort, interest or a combination of them -
oscillates between a minimum and maximum value. The absolute values have no intrinsic
meaning in terms of growth rate values and should be brought back to a normal form. The
normal form is one in which the minimum is set to 0 and the maximum to 1. A simple way of
normalizing the oscillatory value goes as follows. Let A1 00 be the first cell. In the adjacent
cell BlOO enter the following equation:

Copy it to B101..B300. This is the range in which you will find your normalized oscillation
values which vary between 0 and 1. The functions @max and @min are built-in
spreadsheet functions; they will give you either the maximal or the minimal value in the
range to which they apply (e.g. range A100..A300 in @max(~100..A300)).~
The oscillating function will act as the variable growth rate parameter for a grower. It
does not matter too much what grower this is. Just assume it is lexical growth. In cell C100,
enter the starting value, for instance 0.01 (assuming that the carrying capacity level is 1).
In cell 0101, enter the logistic equation. It should refer to the varying growth rate in the
range B1OO..B300:

Copy this to the range D102..D300. In this equation, the growth rate is the value of the
adjacent B cell, that is, 6101 for cell D101, 8102 for Dl02 and so on.
The growth rate varies between 0 and 1, because it is the normalized oscillation. You
probably want to experiment with different values, such as a growth rate that varies
between 0 and 0.2, or - 0.5 and 2. In order to achieve this, change the equation in the
range B101 as follows:

(do not forget to copy this changed equation to B101..B300). The equation makes no
sense until you have told the spreadsheet what DRand AR mean. Go to the range where
you have put your other parameters and define the range names DR and AR. If DR is 1
and AR is - 0.5, the normalized values of your oscillating resource will vary between
- 0.5 and + 0.5 (if DR = 0.5 and AR = 0.2 they will vary between 0.1 and 0.6 and so forth;
see Figure 6.6).
Why would we want to let the growth rate value vary between a negative and a positive
number? The idea is that if a child invests too little effort and attention in a learning task,
the effect may be the reverse of learning: performance may drop instead of increase.
Try different values of both ARand DR, but begin with an AR value of 0 and a DRof 0.1.
If AR is negative (e.g. - 0.2), the growth curve will show dips. Try different values, for
instance a high DRvalue (which, however, should never be bigger than 3). Try also the
effect of a high DR (such as 2.5) and a negative AR (such as - 0.2).
 From credit cards to predators

Figure 6.6 Growth based on an oscillating resource curve, for AR = - 0.15 and
DR = 0.12; the resource curve is based on the well model.

Is there any evidence for the existence of growth curves that show the irregular pattern
demonstrated in Model 6.2? Well, we have seen several examples already. Take for
instance Keren's lexical growth curve. Overall, it fitted the logistic form very well, but
it showed small irregularities over time. This became particularly clear when we
computed the underlying growth rates. The pattern of growth rates did indeed
resemble an oscillatory pattern; the biggest oscillations occurred at the beginning. Very
similar data can be found in Corrigan's study of the growth of the lexicon and the
growth of the mean length of utterance (Corrigan, 1983). A nice qualitative fit of these
data can be achieved by modelling them with an oscillating resource (van Geert,
unpublished paper). In the next section, I shall explain that different intuitively
plausible models of resource depletion and recovery all lead to a similar pattern,
namely oscillation over time.

From credit cards to predators'

The credit card effect
A basic idea of the logistic growth model is that a delay exists between the first
occurrence of a causal factor or mechanism, and its actual expression in the
psychological variable it determines. Thisfeedback delay can be explained in a number
of ways. For instance, it could be the effect of an intervening learning period, or of a
certain resistance of a conceptual network to change. I t could also be the consequence
 Oscillations
of an intervening planning and preparation stage. Before actually carrying out an
activity in response to environmental conditions, we must plan the action, recruit
means for its execution or prepare ourselves in any way required by the goal. It is
conceivable, therefore, that there is a class of resource-consuming learning activities
that involves this significant delay between the onset of the planning and the actual
execution. The principle is very similar to buying something with a credit card: buy
now and pay later. The activities to which this principle applies are probably of the
kind that requires a lot of information gathering, preparation, the use of different
cognitive and material means to attack the problems at hand and so forth. Let me first
take you to Toad Hall and show you how a model of resource depletion based on the
credit card principle might work.

Model 6.3 Mr Toad's bank account

A week or so ago as I strolled through a Palo Alto flea market, a little brown book caught
my eye. It was Kenneth Grahame's The Windin the Willows, a book I had always wanted to
read just because its title evokes a tranquil feeling of sitting on the marshy borders of a
creek in the shade of old trees. While reading the book, I was especially taken with one of
the characters, the wealthy but also rather pathetic Mr Toad of Toad Hall. The following
story could very well have happened to Mr Toad. It involves his bank account and his more
than unfortunate discovery of the credit card.
Mr Toad is indeed a very rich toad and his capital, well invested and taken care of,
grows in accordance with classic neo-liberal economic laws, exponentially. When old Mr
Toad died he left his son a starting capital of 0.01 Megadollars. Since the bank manager
has a spreadsheet file of Mr Toad's capital, he enters this starting amount in cell A100.
Good investments in the stock market make this capital grow at the memorable rate of 0.2
a day! The happy bank manager writes down:

in cell A101 and copies this to the range Al02..A1000, after having defined a range r that is
given thevalue 0.2. And then, a few days later, the unimaginable happens. Young Mr Toad
discovers the credit card. 'How could I ever have lived without it', he exclaims, 'All those
wasted years that lie behind me! From now on I shall devote my entire life to the use of the
credit card.' The bank manager is a bit concerned and bids Mr Toad be very cautious.
'How, sir,' he asks, 'are you going to pay your debts back?' 'Well my good man,' says Toad,
'I shall spend my entire capital on my credit card every day, and I shall pay you back any
debt I incur after twenty days, but what I shall pay you back will of course always be a
function of what I possess at that moment.'The bank manager grumblingly submits to the
wishes of his best customer and changes the equation in cell A101 as follows:

which he then copies to the range A102..A1000. In cell B101 he enters the following
equation:
 From credit cards to predators 165
which he copies to B102..B1000. He defines a range named dwhich is given the value 1,
after ample and painful negotiations with Mr Toad, who has a much larger figure in mind.
Equation 6.10 is the amount that Toad has agreed to spend on paying back his debts on
the credit card made twenty days earlier. To his surprise, the bank manager sees that
Toad's bank account oscillates (see Figure 6.7). Spikes of high income are separated by
periods where Toad is as poor as the local churchmouse.

How do Toad's financial problems apply to a model of psychological resources?

Equation 6.8 implies that a person's resources grow, because there is an external input.
For instance, people acquire better and faster ways of information processing,
attention span, effort allocation and so forth. They rest and relax to substitute any
resources they may have consumed earlier. T h e second part of the model says that next
to the repletion there is also a depletion of resources. Resources are consumed, because
the person is investing effort and attention. The third part of the model, equation 6.10,
says that the repletion of resources is a function of an earlier decision, action plan, etc.,
on the one hand and of the present state of available resources on the other. T h e model
applies to all forms of resource investment that are based on action plans or decisions
that require a considerable preparatory stage, causing a significant delay between
decision and actual performance.
Let us see what kind of growth curve would result from a growth or learning process
that used this oscillating resource. Go back to the spreadsheet, enter an initial level in
cell ClOO and write the following equation in cell BlOl (to be copied to the range
C102..C500):

Figure 6.7 The evolution of M r Toad's bank account based on his agreement with the
bank manager; the magnitude and distribution of the spikes depend on the parameter
values chosen.
166 Oscillations

and define range names AR and DR which take, as default values, 0 and 0.05,
respectively. The resulting growth curve will be like the one in Figure 6.8. It has the
form of a simple stepwise function consisting of plateaux alternated by rapid growth
spurts. The stepform is a consequence of the resource release in the form of sharp
spikes, typical of this model.

Earning money, paying debts ...

The credit card concept provides a simple model of growth processes that depend on
exhaustible but recoverable resources. For instance, I use energy to support my
activities during daytime, but I have to pay back in the form of inactivity and rest
during the night. I consume interest and attention resources as long as I am interested
in a specific subject or activity, but I pay back in the form of habituation and temporary
loss of interest in that particular subject. In the previous section, I modelled the
trajectory of an exhaustible resource such as energy and attention. I found an
oscillating function, which I connected to the growth rate parameter of a grower,
resulting in a stepwise process. The following model is an extension of the credit card
concept.
It is obvious to assume that the progress we make is proportionate to the amount of
effort and attention spent in that process (although any developmental progress
depends on more factors than attention or effort). In the preceding sections, I assumed

Figure 6.8 An imaginery growth curve of a complex skill (playing volleyball, for instance).
T h e curve is based on the bank model (Figure 6.7) and forms a pattern of sudden
increases followed by temporary stability.
 From credit cards to predators 167
that we must pay back the exhaustible resources consumed, in the form of rest, or a
temporary decrease in learning or progress. I also assumed that there is a certain delay
between the actual consumption of the resource and its replenishment, for instance in
the form of a rest period.
Let us take a simple logistic equation. We know that the increase in the growth level
L, is determined by the function ( r - r - L,). I have assumed, however, that you pay
back for resources consumed earlier, which implies that some of the resources that
would go into the learning or growth process have to be invested in replenishing the
energy or attention resource used to feed the preceding growth stages. Since that
amount is a function of the amount of previous growth, that is, increase in level, we end
up with a logistic growth equation that has the following form:

The function AL,-, is the increase in the growth level L at time t - n, that is, n time
units ago. The delay time n is the average delay in repletion of resources characteristic
of the current system. The parameter d determines to what extent the growth process
suffers from the growth in the past. What is the form of the AL,-,, function? It could
simply be the increase in L, n time units ago, or an average increase, or any weighted
function of the increase. Instead of reckoning with the absolute increase, we can also
work with the relative increase. The idea is, then, that the resources needed to make a
certain amount of progress depend on the level already attained; learning 5 words
requires more effort when you know only 50 words than when the lexicon is, for
instance, 2000 words. I am not sure whether the relative increase assumption holds for
word learning, but it seems reasonable to assume that it is true in at least some forms of
learning or development.

Model 6.4 The fluctuation of attention

This model lets you experiment with four different versions of the fluctuation equation: the
cubic and the quadratic version of the logistic equation and the relative versus absolute
increase model. You will need the following list of parameters: ini, r, d, r h i , div, cub and
re14. The maximal delay your model affords is 100 cells. This means that the first 100 cells
are a sort of take-off for the actual model. In A l , enter:

The parameter divis a divisor determining which fraction of the initial level the take-off
phase starts with. Since the actual starting point of the model is set at cell ~ 1 0 0we
, let the
initial state level grow towards its starting level. In order to do that, enter:

+ A1 * (1 + $r-ini - $r-ini * A1 /$ini) (6.14)

in cell A2, and copy this to A3..A99.

+
In cell A100 write " ini".
You will use the B column to keep track of the increase or decrease in A, defined either
absolutely or relatively. In cell 82, write:
168 Oscillations
+ (A2 - Al)l@if ($re1 = 1, A l , 1)
and copy this to B3..B3000.
The actual model equation should be entered in cell A101:

Cell C101 is where you specify the AL function. One possibility is:

You can change the delays and weights as you wish. Copy C101 to C102..C3000.
Each time you change the AL function, C101 should be recopied to:

For a start, try ini = 0.01; r = 0.05; d = 0.12; cub = 0; rel= 1.

More complicated weighted functions are possible, for instance with five different
values and weights ranging from 1 to 3. The delays and parameter values can be very
different, but the result is almost always fascinating. With low parameter values,
especially low values for the damping parameter d, the resulting growth curves are
regular S-shaped curves. For somewhat higher values, the curves grow in an oscillatory
way, which strongly resembles the growth curves of morpho-syntactic structures
discussed earlier. For still higher values, the curves settle into chaotic oscillations.
They differ from the chaotic oscillations of the unaltered logistic equation. With the
latter, the levels jump from one state to another. With the current altered form, the
chaotic oscillations are smooth. They make beautiful phase space representations
(Figure 6.9). The model is very rich (but hopefully not too rich) and produces a large
variety of different growth patterns, most of which have the kind of irregularity typical
of many empirical growth phenomena.
One characteristic pattern consists of an irregular oscillation which occurs after the
process has reached its carrying capacity level. It often starts with small oscillations
that magnify and then vary between upper and lower boundaries. It is interesting to
note that these oscillations are very similar to the oscillations in attention that Hylan
found in 1898 (Hylan, 1898). This is probably more than just an accidental similarity.
Hylan's model and mine follow the same assumptions: attention and other exhaustible
resources need recuperation and there is often a significant delay between the use and
the actual recuperation of the resource.

Le secret de l'urinoir . ..
When I was a kid, I used to accompany my parents on their Sunday night visits to the
pub. This was just the Flemish way, and nobody saw any harm in letting the children
taste the divine fruit of the brewery, the Trappist beer, made and sold by Trappist
monks and a proof of the fact that God blesses those who love Him. Anyway, the beer
had a pronounced effect on my inexperienced bladder, and it was in the men's toilet
that I was confronted with an interesting technical problem. As anyone knows, the
 F r o m credit cards to predators

Figure 6.9 A logistic model is altered in accordance with the resource recovery concept:
the resulting growth curve shows chaotic oscillations that die out towards the end. T h e coil
in the foreground is the phase representation of the oscillating growth curve.

male anatomy requires two different kinds of plumbing equipment and it was that
aimed at the fulfilment of the smaller needs that caught my attention. Nowadays, these
things are equipped with a flush button or even'an electronic beam, but the older ones
either had no flush at all, or flushed at regular intervals not caused by any human
intervention. I remember that I spent quite some time trying to reconstruct the secret
behind that regular automatic flushing. One of my designs involved a complicated
system of clocks and levers. It was only much later that I learned that the secret behind
the flushing is very simple and that it is technically subsumed under the class of
so-called relaxation waves, an unintended but appropriate name given the context
where I first discovered one.
Relaxation waves apply to resources, such as the water container, that recover at a
constant rate (the tap filling the container). The resource is released when a threshold
level is crossed, which, in the case of the container, is when the overflow pipe gets filled
with water. Since the pipe has a form of a siphon, the water will start flowing and will
continue to flow until the level in the container reaches the opening of the pipe. Then
the container will fill up again and so on;
It is highly probable that at least some forms of attentional resource functions
behave in accordance with the principle of a relaxation wave. Take for instance a
person's interest in a specific subject. Prolonged activity an& involvement in that
subject will finally lead to saturation. As the saturation level reaches a certain
threshold, the person will quit that activity and proceed to something else. Then
motivation and interest in the subject will build up again, until they reach the threshold
170 Oscillations
needed to initiate the activity again. This is just a speculative model, but it fits our
intuition of how attention and interest behave. The model is also quite reminiscent of
the reversal shifts described by Apter and Smith (1982).
An interesting aspect of the model is that the thresholds are most likely stochastically
determined. That is, the actual point of ceasing an activity following saturation is at
least codetermined by environmental or activity-specific events that are to a large
extent random. Similarly, the renewed initiative to pick up the activity again also
depends on coincidental events. The following model describes a version of the logistic
growth equation to which a relaxation wave effect has been added. It shows how adding
a small random perturbation may affect the long-term course of a growth process
mainly because the random factor has a major impact on events that lie at the boundary
of the attractor points and is unable to affect other regions of the dynamics.

Model 6.5 Relaxation waves and learning

The model of the relaxation wave may seem a bit awkward at first, but it is only a disguised
version of the logistic growth equation. It consists of two coupled equations. One
describes the growth of the resource function that is supposed to go up and down; the
other specifies the change in the growth rate of that growth function. Let us begin with the
latter equation.
The growth rate must be positive when the dependent grower- the resource -goes up,
and negative when it goes down. In this way, the resource will keep growing as long as it
grows and it will continue to decrease as long as it decreases. Let A2 be the starting cell of
the growth rate, with value 0.5, and B2 the starting cell for the resource, with value 0.01.
The growth rate changes in accordance with the following equation:

which has to be entered in range A3..A1000.

Range B3..B1000 takes the equations for the growth of the resource. In 83 enter

First, all the range names necessary to make the model run must be named and specified.
The range name d ~ c o n t a i n sa parameter damping the effect of the growth rate values in
A on the resource in B. The parameter d ~ a n modifies
d the extent to which the random
component affects the resource growth. The r ~ p a r a m e t especifies
r the rate of change in
the growth rate. Suitable values for each of them are: r_r= 0.01, d ~ 114 = and
d ~ a n= d 11200. The graph of range B2..B1000, where the resource levels are found,
should look like the oscillations in Figure 6.10. Hit the recalculation button a few times (F9
in Lotus 1-2-3 for Windows) to see how the small random factor changes the macroscopic
outlook of the resource spikes (or humps). The smaller you make d_r(e.g. 0.11, the more
clearly the random effect may be observed.
Assume that the resource is a major component of a growth rate in an ordinary logistic
growth equation. To model such a dependent grower, use the equation:
 From credit cards to predators

Figure 6.10 Resource oscillations based on the principle of the relaxation wave: the length
and distribution of the waves depend on small random factors acting throughout the whole
process.

that starts with an initial value of 0.01 in cell C1. Define the growth rate damping parameter
d-g and give it the value 0.01, or at least a value small enough to smear the growth out
over the 1000 consecutive cells.

This model shows very clearly how a small random factor which ranges between plus
or minus 0.5 per cent of the maximal level of the resource is active only in the vicinity of
the attractor points, that is, at the levels 1 and 0. The random factor is equally active
over the whole range of resource levels, but its effect is sensed in limited regions only.
They are the regions where the changes are of the same magnitude as the random
influences. The random factor actually either prolongs or shortens the length of the
equilibrium stages, the regions where the resource value is close either to 1 or to 0.

Interest contagion
Most of the subjects that people can pay attention to ,and into which they are likely to
invest effort - like a psychology course, another person's habits and character or a
playground - are complicated structures. They contain a large variety of potential
aspects, components and details. We never pay attention to all these aspects at the same
time. That would simply far exceed the space of our humatl working memory. It
follows then, that a model of attention resources, instead of describing a person's
attention investment in the broad subject as a whole, should try to capture changes and
fluctuations in attention and effort spent in dealing with separate aspects of that subject
domain. There is no natural division of a subject range, say a psychology course, into a
172 Oscillations
set of neatly distinguished elements. Nevertheless, any student will be able to
recognize whether a given topic is new or not, given what he or she already knows about
the course. Let me, for the sake of simplicity, just assume that the number of
components or aspects that can be discerned in a broad subject is determined by a
person's subjective decision. I take that number as a person-specific reference point.
There are three assumptions that can be made in this model. The first is that interest
in or effort spent on some aspect of a subject range is contagious. That is, as we get
interested in a particular aspect of a problem, attention will be directed to other slightly
related aspects. Put differently, there is a growth of involvement in some domain of
interest because each element we pay attention to makes it more probable that another
element will become part of the interest domain. The second assumption is nothing but
a version of the habituation or saturation principle. It says that a person is interested in
some specific aspect or subject for only a limited time. After that time, habituation or
saturation will turn attention to another part of the subject domain. Third, once an
aspect has been rejected as a consequence of saturation or habituation, it will require a
certain recovery time before that particular aspect can become a candidate for interest
investment again.
It is easy to construct a model of attention resource investment over time, given
these three assumptions. Let the total subjectively distinguished number of aspects or
elements in a subject range be normalized to 1 and any number of aspects actually paid
attention to, to a fraction of 1. Call this fraction A. Assume that the average
'contagiousness' of attention investment is described by the parameter c. If At is the
initial number of aspects attended to, the increase in number of aspects interested in is
given by:

The variable St is the number of aspects that are susceptible to becoming a new focus of
attention. What is its magnitude?
Recall that a topic remains interesting only for a limited amount of time, say a time
period p. After that topic has been left, it will require another period of time before the
habituation or saturation that occurred as a consequence of the earlier involvement has
been dissipated. Let us call this recovery time q.
+
If t is the time of the initial state occurrence, at time t 1 the number of susceptible
aspects is:

and at any later time it is:

The sum of As within brackets is the sum of aspects attended to for the total time
-,
period of p. The variable At+, is a set of aspects we were interested in and that has
now recovered from the saturation that was a consequence of that earlier involvement
(for a recovery time of q). This set of coupled equations will by itself run into an
oscillatory movement, as can be seen in the following model.
 From credit cards to predators 173
Model 6.6 Contagious attention
You need three columns of data to model this concept of contagious attention. Start with
cell A99 and enter the following titles and equations:

column A column B column C

row 99 new topics potential topics total topics

row 100 @RAND/$E $N - A100 @SUM (A1OO..A93)

row 101 + C100*$A* + B100 -A100 + A10 @SUM(A101..A94)

8100

In order to make the model run, you need to define a set of parameters and values
(between brackets): A(0.3), N(1), D(I), E(100), R(20), Random(0). Copy the cell values in
A1 01..Dl 01 to range A102..DI 000.
If each cell represents a tick of the clock (and such a tick could correspond with a day, a
week, a month, etc.), the time a particular topic stays in the pool of topics in which we are
interested is eight ticks. It requires a long recovery time, however, before a topic that once
belonged to that pool re-enters the pool of potentially interesting topics again, namely
ninety ticks. These values can be found in the form of cell references in cells B101 and
C100, respectively. The model with the parameter values described here produces an
oscillation of sharp spikes of interest in an increasing number of topics all belonging to a
broad subject domain (see Figure 6.1 1). Experiment with different parameter values, for
instance A = 0.15 or D = 10, and with different delay values. For instance, in cell B11
specify the value A50 instead of A10, but do not forget to copy this equation to the entire B
range). Define the graph for range C100..C1000.
You can enter a fourth column, called 'Total constrained', which allows you to enter the
effect of working memory limitation on the number of topics entertained simultaneo~sly,~
or the effect of a random perturbation (if D = 2, the maximal number of topics entertained
simultaneously is 50 per cent of the total number represented by the parameter N). If you
use the D-column, replace 'C' by 'D' in the A-column.
The equation for range D100..D1000 is a bit awkward, unfortunately:

You can also make the number of possibly distinguished topics smaller by round-
ing-off. For instance, if you round-off to one digit behind the decimal point, the total
number of possible separate topics that a person could attend to in a ~ ' ~ e c i fattention
ic
domain is ten. Range DIOO..D1000 can be plotted separately to show the effect of the
constraints you imposed on your model.

Just as with the previous models, you should make a grower dependent on the
oscillating resource by making its growth rate a function of the resource. The result is
 Oscillations

Figure 6.11 Resource oscillations based on the model of 'contagious' attention, saturation
and recovery; the magnitude and distribution of the spikes depend on the parameter values
chosen.

very similar to the growth curve based on the credit card model. It shows characteristic
stepwise changes, very similar to those from Figure 6.7, based on the credit card
metaphor. The spike form of the resource release produces the characteristic stepwise
growth. Many forms of learning, for instance learning in school, follow this temporal
organization: relatively short periods of learning and increase in mastery followed by
long periods of rest or stability. The resources not utilized during the periods of rest
can be used for other processes of learning that follow the same temporal pattern. This
particular form of resource allocation might explain why our natural form of learning
consists of a succession of many small learning episodes, devoted to different topics,
that succeed one another rather abruptly.

The predator-prey relationship between attention and effort

In the morning, when I enter the redwood cabin that serves as my office at the Center
for Advanced Study - which one of my fellow Fellows called the Center for
Ridiculously Advanced Study after having seen what other people were doing - I am in
general eager to start writing. An unfinished chapter is waiting for completion and
while I am biking to the Center a lot of possibilities and ideas come to my mind. As I
start working, I soon become completely immersed in the problems of cognitive
growers and non-linear relationships and other fancy things. But as the day runs to its
end, I begin to feel some saturation; the Lotka-Volterra models that in the morning
 From credit cards to predators 175
completely filled my mental horizon are now gradually losing their glittering attraction
and I start longing for a glass of wine and some bread with Italian salami.
The present scenario - with or without Lotka, Volterra and salami - should sound
familiar to many readers. I have mentioned it because it contains a potential model of
resource oscillations. In the sort of activities described above, motivation and
willingness to spend effort on a task build up. It is because you have the motivation to
carry out a particular activity that effort and energy resources are allocated and actually
invested in doing the task. But, as you go on, the investment of energy and effort lead to
fatigue or saturation, which in turn decreases the further allocation of resources to the
task. That is, the investment of effort and energy in a task decreases as a function of
time spent on that task. Let E be the effort or energy invested in a task. It decreases over
time but increases as a function of attention, motivation and interest. Let me express
this in the form of the following equation:

for E the energy or effort level, f the parameter that specifies the rate of fatigue, A the
level of attention or interest in the task, and m the positive effect of attention on the
allocation of energy resources.
What about the change in A, that is, the attention, motivation or interest level? Let
me assume that it grows towards an asymptote. That is, it increases over time and levels
off as it reaches a maximum (which is probably person- and context-dependent). The
investment of energy or effort eventually has a negative effect on attention or
motivation, for instance, because it leads to fatigue or saturation. Let me write these
assumptions in the form of the following equation:

which means that attention grows in accordance with the principle of logistic growth
(hence the growth parameter r and the carrying capacity K for the asymptote) and that
it decreases meanwhile as a consequence of the amount of effort and energy invested in
a task, mediated by a parameter d.
This model is similar to the predator-prey model that describes the dynamics of a
population of prey (say rabbits or ranunculus) and predators (say foxes or rabbits,
respectively). It has been developed and studied by an Italian and an American
mathematician whose names are Volterra and Lotka (they could have featured as a
comic duo with names like that, but instead they chose to enrich the world with what is
probably the oldest example of a coupled dynamic equation). If you want to see what
this model does, go back to Chapter 2 and the section on singing in Yeltsingrad. The
model describing the evolution of performance in the happy singers is exactly like the
present one describing the relation between attention and effort. As can be seen in the
figures, both variables are locked into oscillations that are slightly out of phase.
Try the equations in spreadsheet form with the following set of parameters. Take 0.9
as initial value for A and 0.1 as first value of E. Define range names r, d, f and m, and
give them the values r = 0.03, d = 0.04,f = 0.03 and m = 0.1. Try different values for
each parameter, but change them in small increments.
176 Oscillations
What about an eventual learning process that depends on the oscillating attention
and effort? Its growth rate probably amounts to some combination of both variables, of
attention and effort. Assume both make a similar contribution to the growth of, for
instance, the child's lexicon. The equation for the growth of a skill or knowledge
domain dependent on attention and effort would then be:

&+I = L;(l + d.(A, + E,) - d . ( A , + E,).L,) (6.29)

for d a damping parameter that mediates the effect of attention and effort on growth.
Figure 6.12 shows the result of a simulation based on an initial value of 0.001 for L
and 0.012 for d. The number of cells is equal to that used for computing the resource
oscillation, namely 1000. Next to the theoretical curve is the empirical curve based on
Keren's lexical growth (Dromi, 1986). I have not attempted to find a better fitting
theoretical curve, since I just wanted to show the qualitative similarity between the
curves.

Summary: Towards connected growers

The models presented in this chapter are quite speculative, to say the least. My aim was
to show that various intuitively plausible models of resource retrieval and repletion all
lead to the same pattern, an oscillation over time. Each of the models contained a
parameter - or a few parameters - that accounted for qualitatively different patterns.

Figure 6.12 A growth curve based on the predator-prey resource model compared with
Keren's lexical growth curve. No effort has been made to estimate parameter values that
would have led to a better fit between simulation and empirical data; the simulated curve
shows that patterns qualitatively similar to the empirical curve result directly from applying
the predator-prey model.
 From credit cards to predators 177
In general, if the value of such a parameter is low, the oscillation is likely to consist at
the beginning of a (few) large wave(s) that soon dampen out. If it is big enough, the
oscillations take the form of waves, ranging from regular wave patterns to sharp spikes
separated by relatively long periods where the resource level is near zero.
The resource models are not developmental in the pure sense of the word. They do
not model the development of alertness between babyhood and childhood, for
instance, or the allocation of attention and effort resources for classroom tasks. They
are just speculative models that describe the oscillation of those resources over time in
the form of restlactivity or attention/boredom cycles. Their main purpose is to provide
a model for changes in the growth rate function of growers that depend on attention
and effort resources. In this chapter, the coupling between the change in resources and
the growth of a dependent variable was asymmetric. The resource changed by its own
internal logic, not because of the change it promoted in the dependent grower. In
reality, however, we should expect the resource to change as a consequence of the effort
and attention spent in the learning or growth process. The relationship, therefore, is
one of reciprocity. Both growers (resource and dependent grower) affect each other
continuously, although the way they do so differs. The notion of connected growers
and mutual determination will be the main theme in the chapters to come.

Notes
1. Well, that did not seem to be the case in Menlo Park, California, where an evening of (in our
perception very quiet singing) was ended by a policeman whose heart no doubt was with Let it
Be, but whose mind was on his profession.
2. It is actually much more convenient to define two range names, min and max. The value of the
min parameter is the equation

and the value of max is

@max (A100..A300)
You can then replace the @min (A100..A300) and @max (A100..A300) expressions in your
equations by $min and $max respectively-
3. By simultaneously I do not literally mean all at the same moment, but something like 'in the
same time period', which could involve days or weeks during which a set of topics alternates.
Connected growers

Development and uncertainty

Butterflies of Jasper Ridge and butterflies of Lorenz
Not very far from the Center where I hope to finish this book' lies a beautiful biological
preserve called Jasper Ridge. It is owned by the nearby Stanford University and the
other day I was invited there on a nature walk. The guide explained a lot of things
about plants and gophers and seeds and also mentioned that this very preserve was the
place where Paul Ehrlich did a great deal of his fundamental study of population
dynamics and evolution. I had just begun to read Ehrlich's book The Machinery of
Nature, which describes, among others, his ecological studies on a population of Bay
checkerspots, a species of butterfly that lives on the Ridge (Ehrlich, 1986). Ecology is a
good example of a discipline that has successfully applied the principle that (almost)
everything depends on (almost) everything else and the studies of the butterfly
populations demonstrate how the same basic principle of ecological codetermination
may work out in a multitude of ways. Even in a small region like Jasper Ridge, the
checkerspot population seems divided into several smaller subpopulations, each of
which follows its own growth dynamics over time. The populations evolve in answer to
an intricate web of local interactions between the animals and their food sources and
hiding plants within different microclimates and on different soils.
It is highly likely that in their abstract form the mechanisms that govern the subject
of developmental psychology are very close to those of ecology. Both are concerned
with intricate webs and networks of mutual interaction and with changes over
considerable time periods. If we must deal with such structures, we are going to need
the right sort of equipment and method to discover the rules that lead their dynamics.
An interesting property of interwoven structures is that their behaviour shows
unexpected combinations of high regularity and unpredictability.
Thanks to Gleick's fascinating book Chaos (Gleick, 1987), many people have now
heard of the incredibly mighty butterfly of Lorenz. If Lorenz's butterfly flaps its wings
in Singapore, a tornado may occur in Texas. This sounds dramatic, but it is only an
 Development and uncertainty 179
evocative literary illustration of a basic principle in non-linear dynamics. It says that
some dynamics are extremely sensitive to initial conditions and amplify them to a scale
where they completely determine the macroscopic events. The butterfly in Singapore
represents a very, very small event in a region where tropical storms are born, and if it
had not moved, the Texan tornado might not have happened. This discovery puts a
terrible moral burden on butterflies, but they do not seem to worry too much,
fortunately.
Lorenz, a meteorologist who started to work with computers in the 1950s, described
his discovery of the butterfly effect as follows. He had worked on a series of equations
that described atmospheric movements and simulated them numerically on what was
still considered an electronic miracle machine in those days. Once he wanted to
examine the results of such a computation in more detail and typed in the starting
numbers of a previously computed simulation, rounding the numbers off to a few
decimal points. In the beginning of the simulation, the resulting curves were of course
practically like those from the first run, since they were based on starting numbers that
were almost the same as the original ones. But soon the curves began to diverge and the
whole system ran into a completely different path. The effect of rounding-off, which
was (with some exaggeration) like a butterfly flapping its wings, had magnified to the
scale of the simulated events, the atmospheric currents and winds.
Lorenz's equations were coupled equations and if they described development we
would have called them connected growers. Was it because of the connection and
mutuality of effects between them that they ran into unpredictable results, given only a
small error at the starting level? If so, are we not going to run into comparable and even
worse problems in developmental psychology, abandoning all hopes for a science that
is capable of making predictions on the basis of good developmental theory? T o answer
this question I shall discuss several examples of coupled growers. They all start from a
similar assumption about how the carrying capacity, defined as the level and form of a
tutor's or adult's help and support, would evolve as a consequence of changes in the
tutee's learning level.
This chapter presents a basic model of connected, non-linear growers and discusses
characteristic patterns in the simplest possible connected model, namely one with two,
reciprocally connected growers. First, however, I shall go somewhat deeper into the
question already raised in this section: Given that very simple models are already so
complex, will connected growth models not be even more complex? Will their
outcomes be predictable at all?

Predictability in non-linear models .

Maintaining a good level of environmental resources is not free. Take for instance a
school. A high teacher-to-pupil ratio, a good library full of recent books and glossy
educational magazines, lots of computers and materials to expeAment with are indeed
costly. Now assume that the pupils' average learning results remain low and that they
do not seem to take much advantage of the lnformation present in the library and the
attractive videotapes about foreign cultures. The rational reaction to this situation
180 Connected growers
would be to cut down on the expensive resources, since in spite of all the investments
they appear to have produced no results.
How are we going to decide on the resource reduction, what information will be used
and how will the process be paced? A simple model for the adaptation process is one in
which K(the carrying capacity, which is linearly related to the resources) decreases as a
function of the distance between K and the learning or growth level achieved. The
relationship between growth and learning, say of a number of scholastic skills
(represented by L) and the available amount of external resources, such as books and
AV media (represented by K ) would be as follows:

I assume most readers are now sufficiently familiar with the spreadsheet format to
model this set of equations on their own. It is worthwhile trying two identical sets of
coupled equations, and examining the difference between different d and r values.
Good starting values for the variables are: dl = 0.01, d, = 0.01, L, = 0.01,
L, = 0.01, r , = 0.1 and r, = 0.2 (the subscripts 1 and 2 refer to the first and second
set of growers, respectively; if you wish, you can compare three or four at the same
time). What you will see is a growth pattern like Figure 7.1. By experimenting with
different values for r and d you can easily grasp the general principles underlying this
dynamics. Small differences in initial state or parameter values will result in only small
differences in the growth patterns. The model, therefore, is pretty close to a simple
linear model and once you have a reasonable guess of the magnitude of the values
involved, it is relatively easy to predict the outcome.
What happens, however, if the growth rate pushes the growth pattern into the
chaotic domain? This is the situation where the growth level alternates between
overconsuming and underconsuming the available resources. The continuous pattern
of loss followed by rebuilding the level, loss again and so on, is very expensive in
resources invested. The answer to this question is somewhat counterintuitive. Whereas
the adaptation parameter consisted of a small negative adaptation ratio in the first case,
it should now be a big positive number. That is, we should simply increase the
resources, dependent on where the growth level is. This leads to an unexpected
situation where the system soon settles into an equilibrium point. This pattern is quite
counterintuitive, since a strong upward adaptation to a growth process that changes
very rapidly and tends to overshoot its carrying capacity anyway would be expected to
lead to an exponential explosion.
Try the following parameters: dl = - 1.06, d2 = 1.4, L, = 0.3, L, = 0.3, r1 = 3
and r, = 2.4 (note that the negative value of d makes its effect positive in equation 7.1,
since d is preceded by a minus sign). Experiment with some different parameter values
and also try different initial states. If they are small (say 0.01), the equilibrium state
occurs at a much higher level than with relatively high initial states, such as 0.3; see
Figure 7.2. Overall, however, this model still behaves linearly, in that small differences
in a parameter value (between one set of equations and another) result in small
differences between the actual outcomes. The model behaves, however, rather
 Development and uncertainty

Figure 7.1 Different growth rates correspond with different equilibrium points when
carrying capacity adapts itself to the speed with which the growth level approaches the
carrying capacity level.

counterintuitively and requires actual numerical experimentation to find out what it

does, given different values. I t therefore exceeds the limits of purely verbal
extrapolation, but it still allows us to make successful predictions even if the starting
conditions are known only approximately.
In the previous models, the carrying capacity adapts to the growth level with a rate,
K - L, which is big at the beginning and decreases as the growth level approaches the
equilibrium level (when K - L becomes an increasingly small number). We would like
to know what happens if the adaptation occurs the other way around. That is,
adaptation is small in the beginning, because the environment providing the resources
might want to wait and see how the learning or developing proceeds before adjusting its
resources. As it becomes clearer where the learner will aim in the end, adaptation of the
resources provided is supposed to occur faster, inversely dependent on the distance
between growth level and carrying capacity. This model call be described in the form
of the following set of equations, which are only slightly different from those in
equation 7.1:

Whereas the first model behaved very nicely, allowing for prediction and rational
182 Connected growers

Figure 7.2 With chaotic growth rate values, the growers settle into an equilibrium state if
the K-compensation is positive instead of negative.

understanding of what happens on an overall scale, this model behaves strangely. In

order to experiment with the model, start with an initial K of 1 and an initial L of, for
instance, 0.99 and an r of 0.3. The d parameter should be a very small number, for
instance 10-15. It is worth comparing two L - K couples that differ from each other
only in the d parameter. That difference should again be very small if you want to see
something comparable to the butterfly effect. For instance, if dA is lo-", make dB
(1 + 10-14) - 10- 15. Instead of a parameter difference, you can make the initial state
of one L variable slightly different from the other, for instance, with a difference of
10-l5 (see Figure 7.3). The model produces patterns of meandering changes.
The meandering change model defeats any claim to prediction for at least two
reasons. The first is that it shows sudden jumps up and down at unexpected places.
That is, there are no observable signs announcing the jumps; they just occur and an
adjacent growth curve with very similar values does not show them. Second, growth
patterns based on ridiculously small parameter or initial state differences run together
for some time, but suddenly become magnified, often to an extent that exceeds the
average fluctuation of the curves. This happens suddenly and is again not announced
by observable indications. There is a third sense in which such models behave
differently from the ones we are used to. If you add a very small error term to each (or
some) of the steps in the computation, you will see the same divergence from the
comparative path as with the small parameter or initial state difference.
 Development and uncertainty

Figure 7.3 In a non-linear system, very small initial state differences become magnified at
an unpredictable moment; the two growers that differ only in a very small initial state
difference grow together for some time, then one of them suddenly jumps to an entirely
different path.

In what sense, however, does this defeat prediction? Is not the model completely
deterministic and therefore completely predictable? Does not each computer run give
you the same numbers, provided you started with the same initial state conditions as
with the previous run? T h e answer to these questions is affirmative, but that does not
affect the major problem. Prediction is about real states of affairs in the world and is
based on observations made about these empirical states. Observation is accurate only
to a limited extent and however accurate it is, there is always a range of error. If a
mechanism is non-linearly affected by influences whose magnitude lies well under that
error limit, its behaviour is unpredictable.
There is no real problem if the mechanism is affected in a linear way. I n that case the
prediction error always lies within a band of equal width around the predictable
pattern (see Figure 7.4) and is in general proportional to the measurement error. This
is the standard situation, which we traditionally expect to occur. Thus, once we have a
good model, the predicted outcomes are as good (or as bad) as the measurements we
have made of the conditions on which we have based those Gedictions. But we have
just seen a simple model of connected growth where the error is not linearly
distributed. T h e differences between the two patterns, based on only tiny initial state
or parameter differences, are sometimes very small and sometimes very big and no
184 Connected growers

Figure 7.4 In a linear model, errors are distributed evenly around an average process line;
in a non-linear model, major bouts of error or noise may occur in unexpected places.

observation-based principle exists that will tell you where and when those differences
will occur. I n summary, despite its simplicity, the non-linear meandering change
model we just tested is by definition not predictable in the empirical sense of the word.
Should we conclude, therefore, that all models consisting of connected growers
(specified by coupled equations) are unpredictable? Of course not. Earlier in this
chapter, I described another model of resource adaptation, which was superficially
similar to the meandering change model but which provided perfectly predictable
patterns although it is non-linear. Should we conclude then that unpredictability is a
property of a subset of non-linear connected growers? Would the unpredictability
increase with increasing numbers of interacting variables involved? We have seen that
this is not the case (Model 4.3). The chaotic regime of the simple logistic growth curve
is extremely sensitive to very small initial state differences. Logistic, chaotic growth
patterns that are similar except for a tiny initial state difference will remain similar for a
considerable amount of time, and then, suddenly, divergence sets in. If we compute the
difference between the corresponding states of the two series, that difference is almost
zero in the first part and suddenly magnifies to a level where the difference becomes
about as big as the varying growth levels themselves (Figure 7.5).
 Development and uncertainty

Figure 7.5 The dots scattered over the diagram represent growth levels of two different
growers; the levels vary chaotically over time. T h e points refer to two different growers
that start with a minute difference in initial state level. The difference between the two
growers is very small at the beginning, but explodes suddenly and becomes as big as the
chaotic variation itself. T h e dots represent the variation of the growth pattern, the blocks
connected by a line show the difference between that pattern and one with a slightly
different initial state level. T h e point where differences are magnified explosively is the
'prediction horizon'.

The prediction horizon and the function of uncertainty

What the previous models illustrate is the existence of a so-called prediction horizon.
The prediction horizon is the period of time over which our predictions are a linear
function of the measurement errors made. That is, for a given time period, small
measurement errors will lead to only small prediction errors. Beyond that time period,
however, the error magnifies exponentially and will rapidly become as big as the signal
itself. Beyond that point, predictions on the basis of a good model and highly reliable
(though very slightly erroneous) measurement will be as good or as bad as prediction
based on Tarot cards or dice.
The existence of a prediction horizon is a theoretically inferred fact. We have seen
that it exists in a variety of non-linear models. A practical illustration of it is the
mathematical models on which the daily weather forecast is based. They too have a
prediction horizon, which is about five days. Is there any evidence for prediction
horizons in developmental psychology? I t is true that we do not-know how to predict
people's life courses very well, but is this because our models are poor, or because the
information on which the predictions are based is flawed? Or is it because development
is a domain haunted by the kind of non-linear models that have very short prediction
horizons?
186 Connected growers
Let me try to tackle this question by repeating the conclusion from the previous
section: not all non-linear processes of connected growth or growth in general are
unpredictable in the sense outlined above. It is quite probable that many forms of
connected growth in development are of a type that enables good prediction, provided
that we have a sufficiently reliable estimation of the parameters. It is also possible that
the prediction horizon lies far enough ahead to enable us to make sufficiently reliable
and still practical predictions of children's developmental careers. Given these
possibilities, however, what would we expect to be the case, in view of what we know
about human development?
People are both very similar to and very different from one another. They are similar
in a number of basic things, such as their bodily form, basic social habits and behav-
iours, basic understanding of the world and so on. But they differ a lot in the specific
ways that build on those basic things, in the skills they master, the knowledge they have
and so on. This means on the one hand that a number of developmental processes must
be of a kind that enables reasonably faithful reproduction of an existing form. Although
the growth of the embryo, for instance, is an extremely complicated pattern of mu-
tually interacting components, involving environmental as well as genetic information,
the process as a whole is robust. In most cases, it will lead to a very predictable outcome
in the structure and functioning of the physical body that results from it. The same
principle holds for the acquisition of basic skills, such as social behaviour, language and
a basic cognitive understanding of the world. They are robust processes in the sense
that they are successful generation after generation, even if they are accompanied by
individual variation and long-term historical and phylogenetic change.
On the other hand, think of the many domains and characteristics over which people
may differ. They differ in personality, expertise, their attitudes towards life,
temperament, profession, interests, social position and so on. Different traditional
explanations exist for these individual variations. We could claim they are based on
different aptitudes or innate predispositions. I do not doubt many of them are. It has
also been claimed that individual differences arise from coincidental environmental
differences. If you are the son of a baker, you are likely to have a child whose father was
a baker too. Or, finally, individual differences could be just random variations, random
ripples in the great pond of society. These explanations are still question-begging in
the sense that they simply assume the existence of variation; genetic variation in the
case of different genetic endowments, or environmental variation in case the individual
is considered a product of the environment. But where does this variation come from?
If every dimension of difference - say a personality dimension, or a cognitive, affective
or social dimension, etc. -evolves according to a robust dynamic mechanism, variation
will be non-existent, since these dynamics tend to drift inexorably towards one
favoured equilibrium state.
If the mechanism of development were a mixture of robust models, leading to
similar end states in spite of highly varying conditions and 'chaotic' models, which are
extremely sensitive to initial state or environmental conditions, we would have the best
of both worlds. The robust mechanisms would keep everything within manageable
bounds, whereas the chaotic mechanisms would produce extensive variation out of
 Some basic mathematics of connected growth and development 187
what, for all practical purposes, are almost identical starting conditions. In the jargon
of dynamic systems, chaotic dynamics is a great means for exploring possibilities in a
vast domain of potential differences.
Variation seems a necessary condition for life. If you get stuck in a highly particular
ecological niche and you have no sources for experimenting with different forms of life
maintenance, you might be quite well off as long as the circumstances remain similar.
Once the world begins to change, however, you are in serious trouble. The same is true
for the discovery of new ideas and principles. Old ones work fine until you encounter an
unsolvable problem. Then you need a mechanism of variation that produces a range of
possibilities for you, which you can then explore and select dependent on whether they
are successful or not. In summary, adaptation to the vagaries of life requires
mechanisms that produce novelty. The cheapest are the chaotic ones, that is,
mechanisms that produce very different patterns given only very small - sometimes
infinitesimally small - variations in the starting or environmental conditions.

Some basic mathematics of connected growth

and development
Steps of the mind's staircase
On the cover of Robbie Case's The Mind's Staircase'(l992b) is a picture of the Case
children. They look at you from behind the corner of what must be the staircase of the
Case house (lots of cases here, indeed). The picture lets you look at the children from
above: you are on the upper floor; they still have a lot of steps to go. What Case and his
collaborators mean by the mind's staircase is a structure of levels of increasing
cognitive complexity. The knowledge at each of the levels is domain- and con-
text-specific, but each of the items of knowledge or skill is supported by what is called a
central conceptual structure. This structure is 'a network of semantic nodes and relations
that plays a central role in mediating children's performance across a broad range of
tasks (though not all), and that also plays a central role in their development' (Case,
1992b, p. 352).
If we were to try to apply a dynamic growth model to this notion of development, a
feat Robbie has actually accomplished during our cooperation at the Center for
Advanced Study, we could not take only this central conceptual structure into account.
Growth and change apply to the relatively separate .domain- and concept-specific
knowledge and skill structures, as well as to the central structure that s'upports and
finally underlies them. In Robbie's own model, each domain-specific structure has a
moderate supportive and stimulating effect on the growth of the central conceptual
structure, whereas the central conceptual structure in its turn has a much greater
supportive influence on each of the separate domains.
I shall not concentrate on the Case theory of development here, but use it as an
example of a network of connected growers and show a number of possible ways in
which the connections could be worked out. Before doing so, let me describe some
188 Connected growers
domain-specific forms of knowledge that can be considered as separate growers,
although they all contribute to and benefit from an underlying central conceptual
network. T h e examples come from what Robbie Case calls the dimensional stage of
thinking, which occurs in most children between the ages of 5 and 12 years, and which
is roughly similar to Piaget's stage of concrete operations.
In order to test a child's mastery of dimensional thinking, Case and his co-workers
use a set of different tasks. They could have used more, no doubt, but these seem to
constitute a representative sample. Six have been used in a study about parallels in
development. One is the birthday party task. A child is presented with a story about
two children celebrating their birthday. T h e child has to decide how happy each of
them is, using information about which presents each of the party birds wanted in the
first place and what they actually got. T h e reward distribution task features two
children who volunteer to stay after school to make postcards for children in the
hospital. The volunteers differ as to the number of days they came in to help and the
number of postcards they made. The child is then asked to decide what reward the
teacher should give to each of them. The balance beam task involves a judgement about
which side a balance beam will tilt to, given different weights attached at different
distances from the centre of the balance beam. T h e fourth task concentrates on
children's ability to tell the time and on their knowledge of different durations. Clock
reading is an important aspect of this skill. Similar to time telling, the fifth task, money
counting, is of considerable cultural and practical importance to the children. It
involves knowledge of the structure of numbers based on coins and banknotes of
different values. The sixth task, number knowledge, asks for different forms of
numeric understanding, the decimal number system, ordering of numbers and so on.
I t is a task that depends on extended school training. T h e six tasks represent the child's
understanding in the general numeric domain, which is of course only one of the
several domains that develop in children.
Each of the tasks has its own developmental or growth path, even if the reported
correlations between scores are relatively high (Case, 1992b). Moreover, considerable
individual differences exist between children's understanding of the various task
domains. Each of the tasks has its own specific resources in the form of specific
information, training at school, the individual child's interest in and affinity for the task
at issue, cultural support and valuation of the task and so forth. In addition, each of the
tasks is related to its central conceptual structure in that it contributes a little to the
growth of that structure, but in turn profits a lot from it.
Finally, we can ask about the relationships between different central conceptual
structures. For instance, some children are clearly more 'talented' - whatever that
concept finally amounts to - in the numeric domain, whereas others show closer
affinities to the social and practical domain.
In summary, there are a lot of different things out there that we may consider
independent growers in the sense that they follow their own paths of change and have
their own access to specific resources. But the independence goes hand in hand with an
intricate network of reciprocal dependence. Some growers will benefit from others and
will compete for time and effort resources with still other growers. Some growers, such
 Some basic mathematics of connected growth and development 189
as the central conceptual structures, lie deep under the surface of actual behaviour and
problem solving and address their resources in very indirect ways, whereas others take
the form of observable activities that profit directly from the information and training
given. How shall we build a model out of these many building blocks?

The form and magnitude of the relationships between connected

growers
In Chapter 2, I explained that the dynamic systems approach I employ in this book is
actually a very simple sort of theory. It deals only with increments and decrements in
what are called 'growers', that is, aspects of development that follow their own paths of
development and that have their own specific access to their resources. It follows then
that a relation between two growers is at bottom either a supportive or a competitive
relationship. If the relation is supportive, grower A contributes to the growth of grower
B. If it is competitive, grower B suffers from the growth of A. The relationships also
differ in magnitude. T h e support or competition can be strong or weak. If it is zero, the
relation between the two growers is neutral.
The relation can also be conditional, in which case it is a combination of competition
and support. If A is a condition to the growth of B, A will prevent B from growing, as
long as A is below some threshold level. As soon as that level has been crossed, A will
support the growth of B. T h e conditional relationship covers what is traditionally
called a prerequisite or precursor relationship in development. I t refers to a skill or
knowledge whose presence is conditional to (but not necessarily sufficient for) the
emergence of a dependent grower. In the standard account of prerequisites or
precursors, the relation is defined in a qualitative way; the child must have this or that
form of understanding before a more complex overarching form of understanding may
emerge. In the dynamic systems model that deals with 'mores' and 'lesses' only, the
prerequisite relation is specified in the form of a quantitative threshold level, above
which the growth rate of a dependent grower turns from a zero into a positive value.
Finally, the actual positive or negative effect of a grower A on another grower B is the
result of (at least) three components: first, the magnitude of the relationship, and
second and third, the growth levels of A and of B at the time the effect is exerted. This
follows from the size-related concept of growth, which says that increase or decrease is
a function of the size already achieved and not of some absolute addition or subtraction.
I t applies to development in that development is a process where each next step is a
function of the preceding one, or a previous step, dependent on the time delay of the
effect. T h e state-related nature of the connections between two growers is a sort of
default option of the model. Different options, such as mere addition and subtraction,
or optimality relationships are possible and may be defined in the model you construct
out of the given building blocks.

The matrix of connected growth

Let me, for the sake of simplicity, confine rhe discussion to the six numeric tasks and
their underlying central conceptual structure. Because these tasks and structures are
190 Connected growers
Table 7.1 A table containing four theoretical assumptions about relations between growers in the
form of cell values
--

Rate P I3 D M N T C

only examples of a more general principle, I shall call them by single capitals, namely B
for balance beam, P for birthday party, D for distributive justice, M for money
knowledge, T for time telling and N for number knowledge. The central conceptual
structure for numeric knowledge is called C.
Provided we know for each pair of tasks or structures how they quantitatively relate
to one another, we can set up a matrix of relationships that will specify a whole model of
connected growth. The matrix consists of a set of columns bearing the names of the
growers, and a set of rows equally named after the growers in the model. Each cell,
therefore, is at the crossing of two growers and contains information about how the
grower named in the row affects the grower named in the column (see Table 7.1). Put
differently, a cell specifies a theoretical assumption about how one grower relates to the
change of another.
Table 7.1 contains four numerically specified assumptions about the relations
between a number of growers in the Case model. The cell contents are read as follows.
Take for instance the cell numbered (1). It contains the value - 0.01 and is at the
crossing of the P column and the N row. It means that numerical understanding
negatively affects the understanding of the birthday party problem, although the
magnitude of the effect is small (0.01). A possible theoretical justification behind this
figure is that the birthday party problem involves social and personal understanding,
and reasoning with pure numbers, as is taught during the arithmetic class, is supposed
to interfere with it slightly. Of course this is just an assumption, which is not supported
by any available data, but I am only demonstrating a process of dynamic model
building. Similarly, numerical understanding is supposed to have a positive, that is
supportive, effect on the understanding of money problems (cell 2), and a similar effect
on the central conceptual structure of numeric understanding (3). In turn, the central
conceptual structure of numeric understanding has a large positive effect on the
understanding of the number system (cell 4). Finally, the central conceptual structure
is supposed to have a strong supportive effect on the understanding of the birthday
problem (cell S), although the magnitude of the effect is considerably smaller than with
numeric understanding (cell 4).
The whole matrix should be filled in with numbers specifying the relationship
between any two variables involved. In the first step of matrix building, the numbers
 Some basic mathematics of connected growth and development 191
are only rough estimations. They have no function other than roughly to compare the
magnitude of the relationships. For instance, we could decide to take the number 1 as
the smallest possible relationship figure (except for zero, which represents the case of
neutral, that is non-existent, interaction). All the other figures in the matrix - 2,5, 10,
etc. - represent estimated multiples of that smallest number. The real numbers
depend on the actual equations we will write and should be determined once the
equations are available and put in the form of a dynamic model.
Two aspects of the matrix need some further consideration. The first is the second,
unnamed column, which contains the values of the r (growth rate) parameter of the
variables specified on the left. The other is the diagonal, that is, the series of cells where
a variable determines itself. The diagonal cell contains the value of the logistic a
parameter, which indirectly specifies the carrying capacity value for the corresponding
variable (since K = rla).

The standard mathematical format of connected growers

Any positive or negative effect on a grower (for instance P, for birthday party problem
understanding, for instance) is treated as a separate variable in the logistic or transition
equation for that grower. Let us have a look at grower P in Table 7.2. This table is an
extension from Table 7.1 and covers all the influences on P from any of the additional
growers in the model.
From Table 7.2, we can reconstruct its growth equation as follows:

which, by way-of definition, is the same as: ,

provided that Z(G,) means the sum of the effect of all additional growers at time t on
grower P at time t 1. +
This equation has a natural meaning in terms of the growth rate and carrying
capacity of the affected grower P. In order to explain this, let me go back to an abstract,
canonical form of the logistic growth equation, which is:

Table 7.2 The effect of all additional growers on the grower 'birthday party' P. The growth rate of P
is 0.2 and its default carrying capacity 0.210.2 = 1

Rate P B D M N T C
192 Connected growers
The growth rate r can be the sum of any set of contributing variables, for instance:

In equation 7.4, the effective growth rate, therefore, is the sum:

This means that the effect of connecting growers to the growth of, say, P results in an
increase in the growth rate of P with the factor C(G,), which is the sum of all the
influences of the additional growers on grower P at time t. The additional growers also
affect the carrying capacity of P. Recall that given equation 7.5 the level of the carrying
capacity is:
K = r/a (7.8)
Since the effective growth rate of, say, P is given in equation 7.7, it follows that:

This means that the effect of all the additional, connected growers on P is not only that
P's growth rate is increasing, but also that the carrying capacity or maximum
equilibrium level of P increases.' This result is in perfect accord with the definition of
the carrying capacity, namely the maximum equilibrium level achievable under the
currently available resources. Any supportive grower in the connected network acts as
a resource-enhancing variable and thus increases the maximal level the dependent
grower can achieve.
There could be theoretical reasons for confining the effect of the connected growers
to the carrying capacity of the dependent grower, thus leaving the latter's growth rate
at a constant level. T o do so, a coupled equation should be used, which is as follows:

We could also make the growth rate dependent on the connected growers and leave the
carrying capacity as it is. It goes without saying that whatever equation we choose, it
should be based on a theory of development of, for instance, cognitive growth and not
on mere mathematical preferences. If there is reason to assume that the effect of some
connected growers consists of increasing (or decreasing) the rate of growth in a
dependent grower, then we should write an equation in which this model is specified.

Non-linear effects, prerequisite relationships and combined positive

and negative influences
In Tables 7.1 and 7.2 each cell specifies a simple linear relationship between a
dependent grower (for instance, the birthday party problem) and a connected grower
(for instance the central conceptual structure for numeric under~tanding).~ The linear
relationship takes the form of a parameter that moderates the effect of the connected
grower on the dependent one. Some growers, however, have an effect that depends on
an optimal value. Assume that self-esteem would be a separate grower in the Case
model. Self-esteem is partly dependent on personal success or failure experiences,
 Some basic mathematics of connected growth and development 193
partly on personality and environmental (family) aspects. Let us pretend it is a separate
and measurable part of the child's perception of self. There is no doubt that positive
and realistic self-esteem has a positive effect on learning and achievement (see for
instance Harter, 1978, 1983; Kunnen, 1992). Increasing self-esteem should therefore
have an increasing effect on learning. But what if self-esteem keeps increasing and goes
far beyond a realistic and acceptable level? We would expect its positive effect to
decrease markedly and finally to become negative. Unrealistically low self-esteem
would also have a negative effect on achievement and learning and turn into a positive
effect only beyond some minimal threshold.
Provided you define the required parameters, the following equation specifies the
self-esteem effect function as depicted in Figure 7.6:

effect (esteem) = exp

- (esteem - optimum)2
sqrt( - 1 . width. In (threshold)) I - threshold

Note that exp means 'exponent of', sqrt stands for 'square root' and In for 'natural
(7.11)

logarithm'.
What if the theory describes a prerequisite relationship? For instance, it could claim
that a minimum level or threshold (T) of representational understanding R is

Figure 7.6 A variable such as self-esteem is assumed to have an optimal positive effect
corresponding with a certain optimal value, 'but a negative effect when that value is
exceeded.
194 Connected growers
necessary before abstract thought A can emerge. The cell that specifies the effect of
representational (R) on abstract (A) understanding should specify first to what extent
A benefits from R (or even suffers from R, if the two compete) and second, should also
contain a prerequisite equation. In its simplest form, the prerequisite equation is a
function that keeps the growth rate of abstract thought A zero as long as the threshold
T of representational thought R has not been passed:
if R < T, r , = 0, if not r, = any positive number (7.12)
for r , the growth rate of the grower abstract understanding A. It is also possible to
write a more complicated function for the prerequisite relation, but the simple yes or
no form in equation 7.12 functions very well.
Finally, it is possible that the relationship from one grower to another is both
supportive and competitive at the same time. Just assume a child is learning
simultaneously about time telling and money. Both learning activities require
evanescent resources such as time spent on task, effort, attention and interest.
Whatever time spent on learning one skill cannot be spent on learning the other.
Preferences for one or the other learning task will negatively interfere with effort spent
in doing the other. On the other hand, however, it is easy to defend the claim that they
also support each other. They both require the understanding of a comparable
hierarchical principle of relationships between hours and minutes, or dollars and cents.
Adding up dissimilar units of time or money works in accordance with the same general
strategies. Discoveries made in one domain -say money - may increase the probability
of discovering that same principle in the other domain - time. The tasks contribute
independently to the central conceptual structure from which they, in turn, both
profit. In that sense, a simultaneous supportive and competitive relationship exists
between the two growers (this is of course just theory, not a statement about
empirically established relationships). What if I wanted to enclose this hypothesis of
simultaneous competition and support in my matrix?
One possibility is that the positive and negative effects simply add up. The net result
is either a smaller negative or a smaller positive interaction from one to the other. This
interaction parameter can be specified in the cell that corresponds to the growers
involved.
A second more interesting possibility is that the competition and support relate to
different aspects of the grower at issue. In the simplest case, the competition between
money and time telling is a competition for limited resources such as time, effort or
attention. Time spent learning about money cannot be used to improve skills in time
telling. For simplicity, assume a linear relationship between the 'consumption' of
resources such as time and effort on the one hand and the actual increase in a grower on
the other hand. The supportive effect relates most plausibly to a different type of
resource. It is likely that the understanding of one principle (of time, for instance)
contributes to representations of more general rules or central conceptual structures
that help us understand the other principle (of money).
We may now summarize this discussion in a model that relates the competitive
parameter to the increase (or decrease) in the variable, and the supportive parameter to
 The unexpected complexity of simple reciprocal dynamics 195
the absolute level of that variable. For instance, let money (M) be the dependent
variable, and time (I") the variable on which it depends:

M t + l =LmP(Tt - Tt-1) (7.13)

Mt + 1 = &up(Tt)

We shall see later that this principle is useful in modelling the growth of hierarchically
related growers, as in Kurt Fischer's model.

The unexpected complexity of simple

reciprocal dynamics
Time and money
In the preceding section I used Robbie Case's tasks for numerical understanding
during the dimensional stage of development as an example of a set of connected
growers. Let me now concentrate on two tasks, money counting and time telling.
We have seen that both money counting and time telling rely on the same underlying
conceptual structure, that of hierarchy and embedding. Seconds go into minutes,
minutes into hours. Cents go into dollars, one dollar bills into five dollar bills and so on.
They both require the use of general numeric knowledge and strategies. Units add up
to higher order units. Thus whatever general principle you discover in one task would
certainly benefit growth and learning in the other. Familiarity with one task may
accelerate getting acquainted with the other, if the latter is presented later as a new task.
Once a child gets interested in one kind of problem - such as money counting - the
interest taken in this type of task might easily transfer to other domains, such as time
telling. Put differently, the relationship between the time and the money tasks is one of
reciprocal positive influence, that is, reciprocal support. In the matrix model, their
relationship would take the following form as shown in Table 7.3.
Table 7.3 specifies the magnitude of the reciprocal effects. It says that time benefits
twice as much from knowing about money counting as money counting benefits from
time telling. This is just an example and I have no theoretical justification for it. In
Robbie Case's model reciprocal support exists via a third, underlying structure acting
as an intermediate, the central conceptual structure.
Positive interaction between two growers is highly likely to occur with bright and

Table 7.3 A reciprocal positive (supportive) relationship between time telling and money counting,
with different magnitudes for the estimated effects

r Time Money

Time + 0.1
Money + 0.2
196 Connected growers
interested children working in high quality educational situations. They are able to
discover and understand general principles and strategies underlying both task
domains. Mentally retarded children, on the other hand, have great difficulties
understanding general abstract principles. They have similar difficulties reflecting on
their own thinking processes and how and why they came to a specific result. The
probability that their learning how to count money will positively affect their time
telling skills and vice versa is very low indeed. It is more likely that the opposite will
hold. Time telling and money counting are similar in their underlying principle, but
different in how the principle is applied. Time is based on a divisor of 60; money has a
decimal divisor. The units have different names and relate to one another in different
ways. We may expect, therefore, that the relation between the two tasks is reciprocally
negative. That is, learning about money might negatively affect time telling and vice
versa simply because any progress made in one task has a chance of creating confusion
in the understanding of the other.
The previous forms of reciprocity were symmetric as far as the nature of the effect is
concerned; the relations were either both positive (supportive) or both negative
(competitive). An asymmetric reciprocity, on the other hand, involves a positive
(supportive) relationship from, for instance, money counting to time telling and a
negative or competitive effect of time telling on money counting.
Is this asymmetric relationship conceivable with the tasks I used as examples before,
money counting and time telling? One possibility is that the positively affecting task,
money counting, has been learned first and some generalizations have been made and
understood about how counting with units works, applied to magnitudes such as
dollars, quarters and dimes. This knowledge may then help a child to understand a new
task, time telling. But, time invokes a different sort of unit with a divisor of 60, which
might confuse the child and negatively interfere with the skill in money counting. I
admit that this scenario is not very likely to occur, but I present it as an example of how
we could hypothesize a relationship between growers that relate in this asymmetric

In view of the networks that constitute the ecologies of real developing children and
their environments, the reciprocal relationships look ridiculously simple and straight-
forward. Nevertheless, I have discussed several models where the complexity of
behaviour was unrelated to the simplicity of the underlying equations. The next
section will explore the basic dynamics of reciprocally connected growers.

Reciprocal support
The matrix for the reciprocal support can be generalized as shown in Table 7.4.
The parameter S:A + B should be read as 'the extent to which grower A is
supported by grower B' (and S : B +-- A means 'the extent to which grower B is
supported by grower A'). In the simplest form, which is also what we are going to
explore, this parameter is a positive number, for instance 0.1.
In accordance with the general equation for connected growth explained in this
chapter, the standard logistic equation for reciprocal support takes the following form:
 The unexpected complexity of simple reciprocal dynamics 197
Table 7.4 The generalized matrix of reciprocal support

A,,, = A;(1 + r J 1 - - r A . A , + S : A t BmB,)

B,,, =B;(1 +r3-r_B.B,+S:BcA.A,)
We know that in the logistic transition model, the effective growth rate is dependent on
the level of the grower (divided by the carrying capacity, which is set to 1). The
simplest thing to do is just to substitute r - 4 . A, and r E . B, for r - 4 or r d ,
respectively:

But a second possibility exists. Formally, the effect of a positive supporter is to increase
the growth rate of the dependent grower. By doing so, it also increases the carrying
capacity. Since in the transition model the growth rate is assumed to be a function of
the level already achieved, any factor increasing the growth rate must, by definition,
also be a function of that growth level. The equation expressing this state of affairs
takes the following form:
+ +
A,,, = A, A t 2 . ( r J 1 - r - 4 . ~ S~: , A + B-B,)
B,+, = B , + B , ~ - ( ~ _ B - Y E - B , + S : B + A . A ~ )
Let me call this the strict transition version of reciprocal support and the model in
equation 7.15 the intermediate transition version, since the support term added to the
equation behaves as a logistic growth rate.
The numerical study of the behaviour of the support model does not differ from the
other reciprocal models. In the following tutorial, I shall describe a spreadsheet format
that can be used to test each of the models, without having to change the equations.

Model 7.1 A universal model for reciprocal

interactions
Start your spreadsheet by defining a number of parameter values, in the form of range
names. The range names you will need are shown in Table 7.5.
If the value of P:A + B, for instance. is a negative number, A suffers from B, or, put
differently, B has a negative or competitive effect on A. If it is a positive number, the effect
is supportive. By giving the variable trans the value of 1, you define the model as a strict
transitional one. If intis 1, the model is intermediate transitional, but only if the value of
transis also 1 (if trans = 0, the value of intis always read as 0, even if it says 1 in the value
column).
198 Connected growers
Table 7.5 The parameters needed with the universal equation for reciprocal interaction

Value Range name Description

i d Initial level of grower A

id3 Initial level of grower B
r A Growth rate of A
r_B Growth rate of B
P:AtB Extent to which A benefits or suffers from B
P:BtA Extent to which B benefits or suffers from A
trans If value = 1, the model is transitional
int If value = 1, the model is intermediate transitional

It is possible that the value of the equation will drop below zero and become negative.
This situation has no analog in reality; children never know a negative number of words,
or have a negative skill. We should consider the initial state value of each of the growers
as its minimal value. I have introduced this concept of minimal value earlier and called it
the minimal structural growth level. You can apply a trick to keep the growth levels equal
to or above the minimal structural level. The trick is to spread the grower out over two
columns. In column A enter the basic equation for the first grower; in column B enter the
condition that the growth level should never be smaller than the structural minimum,
which happens to be the initial state level. In cell A l , enter $ i n A , and in cell A2:

Note that the formula does not refer to the previous cell in the Acolumn, but to that in the B
column. This is so because the B column contains the constrained value of the equation.
In cell B1 enter the following formula:

@if (A1 < $ i n A , $ i n A , A l ) (7.18)

This 'if' statement defines a value that is i n 4 if the value of A1 is smaller than zero; if it is
not, the value is just A l .
This equation makes the result of the growth equation in the corresponding cell A1
equal to the initial level value if its value is lower than i n 4 . The first if statement in equation
7.17 turns it into the transition version, given that the value of trans is 1. The second if
statement makes it an intermediary transition model if both trans and int are 1. In cell C1
type in& in cell D l :

and copy this to cell D2. In cell C2, enter:

Dl + D l * @if ($trans = 1, D l , 1) * ($r-B

- $r-B * Dl
+ +- A * BI/@if ($trans = 1 # and # $int = 1, D l , 1))
$P: B -

Copy cell range A2..D2 to A3..D1000.

Before starting the process of model testing, you should define a few different graph
types. Define a graph with the name 'timeseries', and one with the name 'phaseplot'.
Timeseries is a normal line graph that depicts the values Bl..B1000 and Dl..D1000.
Phaseplot is a state space diagram. Define it as an xy graph. The value for the x-axis is
 The unexpected complexity of simple reciprocal dynamics 199
B300..B1000; for the y-axis it is D300..D1000. I suggest you start with cell 300 to cut off most
of the introductory cell values. The phase plot is interesting only if you hit the chaotic
domain with your values. The route to chaos should in general not be longer than the first
300 cell values.
You test the model by first specifying all your parameter values in the value list (do not
forget to define the range names for the correct cells, namely those that contain the
values, not the parameter names). Then change the value of one parameter in small steps.
It is possible to automate this process by writing a macro; if you are not familiar with this,
consult the spreadsheet manual.
What will be the equilibrium level of A and 6, if any (that is, given they do not run into
oscillation or chaos or something else)? It is the point where the difference part of the
equations becomes zero. It can be shown that for the standard logistic form, the equilibria
for the growers A and B are:

Now that the preparatory work has been done, we can start testing the behaviour of the
reciprocal support model. I suggest you begin with values from Table 7.5. Since
trans = 0, the model behaves as a straightforward logistic growth model, with
symmetric support. What you will see with these parameter values is probably what
you have expected. Both growers stabilize at an equilibrium point that lies higher than
the level they would have achieved without the reciprocal support. T o see how the level
depends on the value of, for instance, the support parameter, change both P: A + B
and P: B t A to 0.05. You might find it counterintuitive to see that the grower with
the highest growth rate profits the least from the support (try 0.3 for r - 4 and 0.2 for
r 3 ) (see Figure 7.7, which represents the first 100 steps of the computation).
Increase the value of the P: B t- A parameter by increments of 0.01. Check the
resulting curves for each increment. You will see that the curves start to show the
two-state flip-flop we also saw in the standard logistic growth equation. Then suddenly
chaos sets in. A nice form of chaos can be seen for P: B t- A = 0.2. The phaseplot
(more precisely, the state space plot you have defined) shows two separate oval forms.
With still higher values, the oscillations become bigger and bigger, until the whole
process ends in an exponential explosion. For the series of parameter values 0.01,0.01,
0.05, 0.03, 0.0235, 0.06, 0 and 0 (in order of appearance in the value list), a very
different form of chaos can be seen, consisting of irregular oscillatory jumps (see Figure
7.8).
If you get chaotic or even explosive processes with the standard logistic form of the
equation, try the same parameter values ,with different models. For,instance, with
trans = 1 and int = 0 you simulate a transition model, and with trans = 1 and int = 1
you have the intermediate transition model. Overall, the intermediate transition model
is the most robust of the three models, in that it tolerates much higher parameter
variation (especially of the support parameters) before it runs into chaos or explodes.
The intermediate model is particularly interesting if you start from different initial
state values, for instance i n t - 4 = 0.01 and i n t 3 = 0.001. With sufficiently low
200 Connected growers

Figure 7.7 Reciprocal support normally results in a higher equilibrium level for both
growers.

Figure 7.8 Besides real chaotic oscillation, reciprocal support sometimes falls into a
quasi-chaotic oscillation resembling a relaxation wave (this figure); in other cases its growth
pattern resembles a chaotic oscillation.
 The unexpected complexity of simple reciprocal dynamics
support parameter values, a stepwise growth process results. A good empirical example
of this model is Kurt Fischer's levels and tiers model of development. I will leave this
example for Chapter 8.

Reciprocal competition
The logical counterpart of reciprocal support is of course reciprocal competition; both
growers have a negative effect on one another, which technically means that they have a
reciprocal competition relation.
I suggest you start with growth rate values greater than 0.2, and with the following
set of parameters: i n - 4 = 0.03, i n 3 = 0.03, r_A = 0.2, r 3 = 0.15, P : A t
B = - 0.1, P : B + A = -0.1, trans = 0 , int = 0.
Before studying the behaviour of these growers, you must disable the function that
kept the growth rate equal to or above the initial state level. With this function still
working, the qualitative patterns will be different. Go to the columns B and D of your
spreadsheet, and alter the function as follows:
@if(Al < 0 # and # $min = 1, $ i d , A 1) (7.22)
This equation makes sense only if you first define a parameter with the range name mzn.
If A1 is smaller than 0 and the value of min is 1, equation 7.22 returns the value i n - 4
and the value A1 in any other case. Enter a comparable equation in the D column; copy
it to the entire columns B and D. Thus, by setting mzn to zero you can disable the
function that kept the growth level at a minimum that was equal to its initial state level.
The pattern you will get is characteristic of reciprocal competition with relatively
high growth rates (greater than 0.1; Figure 7.9). You will see that both growers stabilize
at a level well below the carrying capacity they would have reached if it were not for the
negative effect of the other grower. In this particular case, one grower (A) regresses a
bit, in response to the growth of B. This pattern, by the way, is similar to the one
produced by weak mutual support, that is, when mutual support does not increase the
carrying capacity. Try the transition version of this model by entering the value 1 for
the trans parameter. This model produces a more dramatic shape, with the beginning
of a stepwise flattening in grower B. As a next step, give all parameters similar values
for the two growers, namely 0.03, 0.03, 0.15, 0.15, - 0.1, - 0.1, 0, 0 (in the order of
Table 7.5). Then start decreasing the value of P: B t A, that is, the parameter
specifying to what extent B suffers from A. Go with increments of - 0.01. If the
competition value reaches - 0.15, the B grower drops down to about zero level. For all
practical purposes, it disappears as a behavioural alternative in the repertoire.
It is interesting to try compensatory values. For instance, try to increase the initial
state level whose competition factor is bigger than the other. Can initial state level
compensate for higher competition?
For smaller growth rates, for instance r - 4 = r 3 = 0.05,'the situation changes
qualitatively. For the higher growth rates, the system of two growers could have any
attractor point; that is, each grower could stabilize at any level between 1 and 0. For the
smaller growth rates, this is no longer the case. The system has only two attractor
202 Connected growers

Figure 7.9 A typical pattern of mutual competition, based on growth rates bigger than
0.1.

states, either (A = K; B = 0) or (A = 0; B = K ) . For some values, the growers stay

together for some time at what seems to be an equilibrium level, but they will
eventually separate and one will grow towards zero; that is, it will disappear as a
behavioural alternative. Try, for a start, the values 0.03, 0.03, 0.05, 0.05, - 0.1,
- 0.10001,0,0 (in the order of Table 7.5). The tiny disadvantage of the B grower will
finally lead to its disappearance.
The small growth rate values (e.g. 0.05) allow for a number of interesting
compensatory relations. That is, a higher competition can be compensated for by a
higher initial level or higher growth level. Try for instance 0.03,0.05,0.05,0.05, - 0.1,
- 0.135, 0, 0 (see Figure 7.10). The existence of such compensatory relations in
addition to the binary attractor of the system explains why it is full of threshold levels.
That is, under some level of, say, competition, the result of the dynamics is invariably
that A disappears. But beyond that level, A wins and B disappears. In the vicinity of
those thresholds, small differences in other variables may change the outcome
dramatically, in that a little higher initial state (for instance) may compensate for a little
higher competition factor. The situation becomes even more complicated if you add an
extra variable, carrying capacity. In the present model, K is by default 1, but it could
easily be added as a variable.
Just assume that this model would apply to a developmental domain where
compensatory actions are taken to counterbalance the negative effect of children's
environments. For instance, children growing up in very poor economic and social
circumstances have a future that is 1ike-abistable state: it is either finishing school and
 T h e unexpected complexity of simple reciprocal dynamics 203

Figure 7.10 In a low-growth-rate situation, reciprocal competition allows for a variety of

compensatory effects: the initial dominance of the first grower is compensated by a slightly
higher carrying capacity level in the second grower, leading to the final 'victory' of the
latter.

taking a reasonable job, or not finishing school and ending up in asocial or even
criminal activities; I exaggerate, but there are too many youngsters for whom these are
the alternative choices in life.
If the system determining the outcome of the process is in general similar to the
bistable reciprocal competition system - and it could be, given the mutually exclusive
nature of the alternative futures and the number of factors influencing the choice - it
will be notoriously difficult to handle that system and navigate it in the direction society
wants. In view of the complexity of the situation, it will be very difficult to predict
whether one or another compensatory educational measure will work. It will work
under some circumstances, and not under others and it will be very difficult to
generalize. This does not mean that, if this model is right, compensatory education and
action are no longer possible. It is rather the other way around: the model predicts that
compensatory effects will occur in a variety of ways. But it also shows that the effects
and nature of the compensatory action are very context-dependent and only work if it
passes a threshold value.

An empirical example of reciprocal competition: the meaning of

I
'big'
If you are shown two different objects at a time, and you are asked to pick the big one,
you would probably take a number of aspects into account. What sort of object is it, is
an object of that size and dimensions usually called big, or would you use another word,
204 Connected growers
such as tall or long? If it is a geometric object, such as a rectangle, which dimension -
horizontal or vertical - is the more salient of the two? Earlier developmental studies
suggested a curvilinear trend in the development of big. Whereas 3-year-olds interpret
big as referring to overall size, 5-year-olds were inclined to see big as tall. Later,
children will return to the concept of big as an overall size word and this is also the
typical pattern with adults.
Sena and Smith (1990) not only replicated some of the older studies, but also
presented their subjects, ranging from age 3 to adulthood, with a variety of different
objects. They found that the meaning of big depended greatly on the context and the
objects used, both with children and adults. Overall, children between 3 and 7 years
showed a strong tendency to interpret big as tall. Adults sometimes interpret big as tall,
as long, or as referring to overall size. Finally, significant individual differences exist for
some contexts. Sena and Smith (1990) suggest that the actual application of the word to
a particular object is the result of a dynamic process, involving perceptual and
cognitive processes, taking into account information from a variety of contextual
sources. Calling something big is not just the immediate expression of a present
concept, but rather the outcome of a dynamic process over a time-scale I earlier
referred to as microdevelopmental. The macrodevelopmental changes, for instance
from the 5-year-old's preference for tall to the adult's use of a variety of contextual
cues, boils down to a process of changing the parameters and eventually the rules of the
microdevelopmental dynamics. This macrodevelopmental process is likely to follow a
type of contextual dynamics similar - but on a different time-scale - to the
microdevelopmental process.
The outcome of the microdevelopmental process -which is an actual decision about
which object is big - can be described as an attractor state in a state space consisting of
the major decision dimensions for big, namely the length (or width) and height
dimensions. At the beginning of the decision process, you have not decided yet for
either length, height or a dimensional combination of both. The decision amounts to
pinpointing a particular dimension or combination of dimensional values as your
decisive criterion for big in the given context. I shall explain how this process can be
modelled by using the idea of a reciprocal competition dynamics.
The dynamics that we need is one that results in the specification of two values: one
value for the importance attributed to the height dimension, one for the importance of
length. If the outcome of the process is ( H = 1, L = 0), only height would be
considered the distinctive feature for big in the particular problem situation at hand. If
( H = 0.5, L = 0.5), height and length are considered equally important (in evaluating
a surface). The deviation of the values from 1 would say something about how
important other dimensions, aspects, or considerations are with regard to determining
whether something is big. The higher those values, the more exclusive the attribution
of big to the dimensions for which these values hold.
T o see how the dynamics work, take the parameter values that produced the pattern
as in Figure 7.10 (0.03, 0.05, 0.05, 0.05, - 0.1, - 0.135, 0, 0). Assume that the A
grower stands for the height dimension, and the B grower for length. A state space plot
with A as y-axis and B as x-axis is shown in Figure 7.11. It represents a
 The unexpected complexity of simple reciprocal dynamics 205

Figure 7.11 Phase plot of a decision process: after a period where both height and length
were taken into account when the child had to make a decision about how big an object is,
the child finally focuses on one particular dimension to make a decision (the choice of the
dimension depends on the context in which the big question is asked; in this particular case
the preferred dimension is the length of the object).

microdevelopmental process, namely the growth of the decision about which

dimension - height, length or both -should determine the choice for big in a particular
problem situation. In this simulated decision process, both height and length (surface)
are taken into consideration for some time, but then the length dimension gains more
and more support. The final outcome is that length accounts for 100 per cent of the
decision about which of two presented objects is the big one. Change the value of the
growth rates of A and B to 0.2 and you will observe a very different situation. The
outcome of the decision process would be that (L = 0.75; H = 0.50). That is, big is
clearly attributed mainly to the surface (a combination of height and length), but in this
particular case length is considered more important' than height in the decision about
which object is the big one. You can go 'on changing the parameters. and studying the
phase plots of the trajectories.
How do the parameters and variables in the reciprocal. competition relate to the
perceptual, conceptual and developmental processes that ~ e n aand Smith (1990)
described? I see the reciprocal competition model as a general, abstract model for a
microdevelopmental process whose detaiIed format could be modelled after a neural
network or a symbolic information processing program, to mention only two
206 Connected growers
possibilities. The parameters and variables should therefore be given a meaning that
makes sense in terms of the short time period the equations are supposed to model here.
A low growth rate leads to a situation in which only one of the dimensions, height or
length, is chosen. This is, moreover, the response pattern most characteristic of the
youngest age group. The most plausible assumption then is that growth rate in the
equation stands for a general variable that changes during development. That variable
could be information-processing speed. It could also be working memory, associated
with the ability to take into account more than one dimension in finding the solution for
a problem. The magnitude of that processing variable, corresponding with growth rate
in the equations, is supposed to increase with age. We have seen that a higher growth
rate results in an equilibrium state where both variables, height and length, survive and
contribute to the impression of big. This is exactly what has been found with older
subjects, although they sometimes use only one of the dimensions. In order for this
result to obtain, the competition parameter of one dimension should be pretty high,
given that the growth rate is high. As for the other variables, they probably relate to
contextual and object-specific properties of the objects given in the decide-which-one-
is-big experiment. We may assume that the perceptual salience of either the horizontal
or the vertical dimension of the objects corresponds with the initial state of the
dimensions in the reciprocal competition model. The higher that initial state, the
higher the probability that it will become the dominant dimension when the process
has settled into an equilibrium state. Cultural and linguistic conventions as to what is
called 'big' may correspond with different carrying capacity levels for each of the
dimensions in the equation. A higher carrying capacity corresponds with a stronger
attracting force and could counterbalance the negative effect of initial state, that is, of
perceptual salience.

Asymmetric reciprocity: competition and support

The preceding reciprocal models were symmetric in the sense that both were either
positive or negative. The present model is asymmetric and involves a negative effect of
one to another that positively affects the first.
The following list of values provides a good starting point for experimenting with
the asymmetric reciprocity model: i n A = 0.01, in23 = 0.01, r - 4 = 0.1, r 2 3 = 0.1,
P: A c B = - 0.01, P: B +- A = 0.1, trans = 0, int = 0 K A = 1; K 2 3 = 0.4.
Try the outcome for these parameter values by changing P: A + B from - 0.01 to
- 0.25 in steps of - 0.025. You will notice that the qualitative behaviour of the model
is very similar to that of reciprocal competition with sufficiently high growth rates (say
r > 0. l), in that the variables stabilize at some level. If you increase the value of the
competition factor of A (that is, the extent to which A 'suffers' from the growth of B),
you will reach a point where A, after a period of introductory growth, slides back to its
minimal level (which by default is the initial state level) (see Figure 7.12).
There are two differences with the reciprocal competition model. The first is that
this model does not change its attractor-behaviourwith decreasing growth rates. Recall
that the reciprocal model could give you any combination of values for the variables
 Figure 7.12 Grower B replaces grower A, which virtually disappears. The pre-transitory
dip is followed by a post-transitory bump. The competence level is the highest
performance level of either A or B.

involved if the growth rate was high enough. Below a certain growth rate limit, though,
the endpoint of the process was that one variable survived on a maximal value and the
other disappeared, even when both growers had stayed on the same level for a while. In
the asymmetric model there is no such discontinuity. Second, the reciprocal
competition model often produces a 'hump' in the grower that retains the higher value
of the two (see Figure 7.9). In the asymmetric model, both growers show the hump,
which sometimes amounts to a significant rise, preceding the levelling-off towards the
equilibrium points.
The next section explores the asymmetric model further for the transition model.
Recall that in the strict transition model, growth rate and the parameter value
specifying how much an adjacent grower affects the grower in question is a function of
the growth level already achieved. In the weak transition model, growth rate is a
function of the growth level, but the parameter specifying the effect of another grower
is a constant.

An empirical illustration: take-over phenomena

$

Try the following list of parameter values i n 4 = 0.1; i n 2 = 0.01; r 4 = 0.1;

r 3 = 0.1; P:A+ B = -. 0.1 P:B+ A = 0.1; trans = 1, int = 0. Notice that the initial
 Connected growers

Figure 7.13 With the right parameter setting, an asymmetric model results in the survival
of a grower B, at the cost of the disappearance of a grower A (the end result is similar to
that of a reciprocal competition dynamics with low growth rates; Figure 7.10)

states of the growers are quite different: the first grower starts much higher than the one
it supports (0.1 versus 0.01). On the other hand, the first one suffers from the increase in
the supported grower. By replacing the last number, 0, by 1your equation will turn into
the other transition form, namely the intermediate one. Figure 7.13 shows the trajectory
of the growers. It is highly reminiscent of what happens in what I call take-over
phenomena. By take-over, I mean that a new skill, strategy or problem-solving procedure
is taking over the range of problems that were dealt with by an earlier and in general
weaker and more primitive skill, strategy or procedure. As a result of the take-over, the
previous developmentally more primitive form is supposed to disappear.
Language is an example of a domain where lots of such phenomena occur. Beginning
speakers make mistakes that are not just random errors. They reflect genuine rules of
language production and understanding that are indicative of the level of linguistic skill
the child has reached at that moment. Those more primitive rules are replaced by
others that in general more closely reflect mature language use. Once a more developed
rule has been mastered, the old predecessor rule is rarely used any longer and usually it
is abandoned completely.
Similar processes are assumed to occur in domains where development can be
described in structural cognitive terms. Take for instance the stages of moral
 The unexpected complexity of simple reciprocal dynamics 209
development discerned by Kohlberg (1969). If a child is presented with a story
containing a moral dilemma, such as whether you are allowed to steal food if your poor
old mother is starving and she has no money to buy any, children may base their
judgement on moral conventions and rules. This is basically what they do during the
stage that corresponds with Piaget's concrete-operational stage. They may also judge
on the basis of universal ethical principles and decide that life is a more important good
than someone else's money. This is what we could expect during the formal operational
stage.The higher form of moral judgement simply replaces the preceding form in the
moral dilemma situations. Thus, as far as thinking about those dilemmas is concerned,
a take-over process has occurred, substituting the higher form for the 10wer.~
It has often been observed that take-over phenomena are accompanied by
temporary regression in performance. Sydney Strauss has done a lot of research on
transient performance regression and has coined the term U-shaped g~otl7thto describe
it. Strauss (1982) and Bever (1982) both edited books containing a variety of papers on
regression in different domains. An explanation that keeps coming back in almost any
paper that deals with it is that the performance regression is the consequence of
internal reorganization processes. As the child starts to build up a new mode of
representation or thinking, for instance, thinking in terms of ratios instead of
qualitative identities, he or she must reorganize and redefine the rules and strategies
used to solve problems in the domain at issue. This reorganization causes a temporary
fallback in performance, followed by a higher performance level, due to the new and
usually better cognitive structure. It is like a three-star restaurant serving junk food
while its kitchen is under renovation and getting a fourth star as soon as the chef has got
used to the new equipment. Anyway, the general pattern of such take-overs is the
appearance of a performance pattern that disappears as a new performance pattern
takes over. This take-over is usually accompanied by a temporary regression. This
qualitative pattern is exactly what the intermediate transition form of the asymmetric
reciprocity model produces, as can be judged from Figure 7.13. But is this resemblance
supported by a good theoretical basis?
In a take-over, a more mature or more highly developed form substitutes for a more
primitive, earlier one. The earlier form can be assumed to support the later one. For
instance, in one of Strauss' experiments, children have to judge how hot the water in a
jar will be, basing their answer on the number of candles heating it (Strauss and Stavy,
1982). They should understand, for instance, that it does not make any difference when
you add another cup of water if you also add an extra candle. Four- to five-year-old
children typically reason on the basis of qualitative identities. They will say, 'You only
added some water and you put an extra candle under it'. In a significant number of
cases, this simple strategy will result in a significant number of correct answers. More
important, however, is whether it will also help the child discover and understand
better the more advanced, quantitative ratio strategy (x cups divided by x candles is the
same as one cup over one candle, etc.). This will be so if, for instance, the older
reasoning strategy leads to an agreement with the newer strategy when the latter is still
under construction. Bryant (1982) has argued that agreements between existing and
new strategies are of considerable importance in establishing and developing that new
210 Connected growers
strategy. But will the later strategy also support the older one? Probably not. The
quantitative strategy, once established, will lead to a different answer from the
qualitative one in a variety of more complex problems. It will conflict with the older
one, in other words and so contribute to its decay.
In summary, take-over phenomena, at least of the type often described as the
U-shaped form, comply with the asymmetric reciprocity format. Are they also based
on the transition form of growth? There is a good chance they are. The transition form
requires that the chances of making developmental progress in some field are positively
related to the level of understanding already built up. The less you know about ratios
and compensations between dimensions, the less likely it is that an arbitrary encounter
with such a problem, in real life or in a tutorial situation, will teach you anything about
those principles. If it can be shown that this is indeed the case with the examples of
domains in which U-shaped growth has been found, the asymmetric reciprocity model
would be a good candidate for explaining the phenomena.
Finally, notice that the pattern described as a take-over, showing an intermediary
performance dip, is also produced by the reciprocal competition model. Under specific
parameter values, the earliest grower increases, then starts to fall back as the later
grower takes off. It appears, then, that very different models produce the same
predictions and outcomes. They cannot both be correct, since they start with different
assumptions about what is going on. The way out of this conflict is to start with a
theoretical assumption, for instance, one that tries to establish the nature of the
relationship between two growers. Is it a reciprocal competition, or is it asymmetric?
Once the nature of the relationship has been determined, one should try to find out
whether a model based on that type of relationship can indeed produce the empirical
trajectory found in research. The other way around, starting from a model that
produces an empirically found pattern, then deciding that this must be the model that
explains the pattern, will often lead to mistakes and overexplanatory models. The
theoretical model of the mechanism should be the starting point for the building of a
dynamics model.

Simultaneous competition and support

In all the previous models competition and support were a function of the level of the
grower. For instance, the extent to which the time-telling skill suffered from money
counting (or the extent to which it benefited, depending on the model at issue) was a
function of the level of skill in the money task. Earlier we have seen that it is not
necessarily the level per se that determines the growth pattern of another grower. It
could also be the change, that is the increase or decrease per time unit, that affects a
dependent grower. Let us assume that a moderately bright and interested child is
learning about money counting and time telling. I t is likely that the two tasks will not
develop completely simultaneously. That is, there is a good chance that one will be
ahead of the other. We can still assume that the two task domains will benefit from one
another. Maybe the beneficial effect runs via an intermediate grower, such as the
central conceptual structure of numerosity, that Robbie Case sees as the responsible
 T h e unexpected complexity of simple reciprocal dynamics 21 1
factor here. But there is also a reciprocal competition. Let us assume that the time
children spend on learning skills such as money counting and time telling is quite
limited. If there were few other skills and activities the child had to learn, there would
be ample time to spend on both the money and the time tasks. In practice, however,
there will be considerable competition for evanescent resources such as time or
attention. Time, effort and attention are resources needed for the learning of skills. We
I
1 may therefore assume that the amount of resources consumed is a function of the
I1 increase in the growth level, not of the growth levelper se. That is, all other things being
1:
equal, rapid learning and acquisition of more information will require more resources
I
t than slower or smaller increases.
I There is another sense in which a competition can exist between, say, money
!
i
counting and time telling. We have seen that the underlying principles are similar, but
1 also different enough eventually to cause confusion. Thus the switch from money to
"; ime could have a negative effect on the learning of the latter, because that requires a
: reorganization of knowledge. For instance, the word quarter refers to a collection of 25
'
units of money, but 15 units of time. The amount of confusion, however, is most likely
I
not dependent on the absolute level of either the money or time knowledge, but on the
actual increase in that knowledge. The increase involves the establishment of new
1 principles, facts and rules. Once they are firmly established, for instance, in the case of
1

1
money, they will not negatively interfere with an additional grower, such as time, any
1 more. Thus it is the learning itself (that is, the change) that is responsible for the
I eventual negative interference.
If I summarize this model, I end up with a model already presented under equation
7.13, one where there is simultaneous competition and support. The competition is a
function of the increase or, in general, the change in the grower. The support on the
other hand, can be considered a function of the level of that grower. Equation 7.13
presented this model in only the most general form. Here is a more specific one, for M
the money and T the time task:

The interpretation of the parameters is straightforward: rMis the growth rate of M, KM

its carrying capacity, sM+, the parameter for the support that M gets from T and
cM+, the parameter for the competition that M experiences from T (see Figures
8.4-8.6 in Chapter 8 for examples). Note that the competition is a function not of the
absolute increase or decrease in T, but of the relative increase, that is of AMIM. One
could as well argue that the competition is a function of the absolute increase, then skip
the M, divisor. I shall describe the simultaneous support and competition model at
length in Chapter 8, where it will be used as a major component in a model of long-
term development.

How to build your own house

Chapters 4 to 7 have been devoted to expkining basic building blocks for dynamic
systems models of development and growth. We have seen that despite their simple
212 Connected growers
appearance, the building-block equations could create stunning and unexpected
patterns. The ecological principle requires that such building blocks be connected in
models which do justice to the manifold interactions and reciprocities that shape and
explain the course of development. In this chapter, I have reviewed the general
principles of how growers should be connected into networks. I discussed very simple
networks consisting of two growers. It seemed as if even simple networks could explain
local developmental phenomena, such as a single stepwise increase, or a performance
regression preceding a jump towards a higher level.
So far, however, I have not explained how to build a model for a more complex
developmental process, for instance, of cognitive development between birth and
adulthood. How should we proceed, given empirical data on the developmental
trajectories and at least some theoretical assumptions about the underlying mechan-
isms? Can we build our own houses, or should this be left to the contractor, that is, the
expert model builder whose thoughts and actions are virtually inaccessible to all those
who are immersed in the data and the problems of empirical research? I t is my firm
belief that building models is not really difficult, certainly not more difficult than doing
your own statistics, given the wealth of relatively friendly computer programs now
available. In the tutorials presented so far, I have guided the reader step by step and
tried to explain how you can build your own small models. As far as building the big
models is concerned, I do not believe there is a general strategy, at least not one that can
be explained concisely enough to fit the scope of this book. Instead, I have opted for a
case story, namely the building of a model of Kurt Fischer's tiers and levels of
development in the form of a hierarchic network of growers. This is what Chapter 8
will discuss.

1. Well, this puts me into a sort of logical entanglement, since by the time this line will be read,
the book will have been finished and then the line should have been in the past tense, which
however is wrong at the time of writing. I used to be fascinated by these sorts of unsolvable
puzzles.
2. Or decreases, if the value of Z(G,) is negative, which is the case when the negative effects are
bigger than the positive ones.
3. Recall that in this model every grower is at the same time a dependentgrower (dependent on the
values of all the other ones) as an independent grower (contributing to determining the value
of the growers to which it is connected); the model is thus fundamentally different from a
classic dependentlindependent variable model.
4. Note, however, that the take-over probably applies to this particular domain only, namely
explicitly stated, highly cognitively laden moral dilemmas. The person who thinks on the
basis of universal principles here may be very authoritative and punishment-oriented as soon
as the problem concerns a criminal act of which he or she is the victim: 'a conservative is a
liberal who has been mugged'.
Building a model: a case
history

The firsf steps

A day at the farm
In the summer of 1989, Kurt Fischer was teaching a course somewhere in South
Germany. He suggested he might stop at my house to discuss an article of mine on
dynamic growth models in cognitive and language development (see van Geert, 1991).
When ,I told my neighbours - hardworking, traditional Dutch farmers, very nice
people - they almost fainted. South Germany was the end of the world for them, and
just the idea that somebody from the United ,States could say that he was in the
neighbourhood anyway and was going to make a detour via The Netherlands made
them shiver with awe. They were truly impressed by this demonstration of
international scientific cooperation, although they were in general more impressed by
the fact that I grew good vegetables. Kurt arrived and we spent a day at my house.
Nothing had changed since it was a working farm, so we first made the obligatory walk
in the vegetable garden, visited the horse stables and admired the lanes and trees of the
old and rural part of The Netherlands where I live. Kurt spent the night in the sheep
shed, which, I hasten to say, had long since been adapted to cater to the needs of our
species and no longer serves those of the genus Ovzs.
The next day we had a most interesting discussion. Each of us explained the other
more than a single person could swallow in a week. By the time Kurt left, we both felt
that we had much more to learn and understand from each other's work, although we
were convinced there was enough commonality to make eventual cooperation
worthwhile. At that time I was head of the large Psychology department of Groningen
University (it is amazing how many psychologists there are in The Netherlands; I still
wonder what they are all doing to make a decent living) and for the next two years I was
busy organizing and managing things that probably would have organized and
managed themselves if I had not been around. About two years later I heard the
wonderful news that Kurt had planned a group on Modeling and Measuring
Development at the Center for Advanced Study in the Behavioral Sciences at
214 Building a model: a case history
Stanford, and that my name was on the list. After I had fulfilled my duties, my
university sent me on a long and well-deserved vacation to California.
Our working group at the Center consisted finally of Kurt, Robbie Case who was
professor of education at Stanford University at that time and me. Every Wednesday
afternoon, Robbie would climb up to the Mount Olympus of the Center and after they
got me off the volleyball court - not without considerable pressure, usually - we would
discuss modelling development, scribbling on the blackboard and filling it with the
humps, the jumps and the bumps we thought characteristic of development.
In the course of that year, we managed to build a dynamic growth model of Kurt's
theory of skill development (when I say we, I should add that Kurt did the hard and
tedious work). In addition, Robbie worked on a model of his central conceptual
structure and the task domains associated with it. The growth of these models during
the year, especially of the complicated Fischer model, provides a good case history of
model building and I hope that telling how the model came about may throw more
light on the process of dynamic model building than formal instruction would do. I
shall first discuss some basic aspects: what Fischer's theory says, how we decided on
basic building blocks of the model and what kind of empirical data were available. I
shall then describe the steps we took towards the construction of the dynamic growth
model, which, at the time this book was finished, was still in the process of further
elaboration.

Fischer's model of skill development: some basic principles

The Fischer model of skill development is a combination of flavours from very
different origins. First, it is obviously a structural theory with a firm deductive basis.
This puts it clearly in the tradition of Piaget. When Kurt's theory is given a label in the
literature, it is almost invariably called neo-Piagetian. Second, it is concerned with
growth and learning, that is, with the quantitative-temporal aspects of development.
Transitions from one 'stage' to another are seen as growth spurts. Thus the qualitative
changes are clearly correlated with quantitative leaps. Third, it sees a person's
developmental level as a context-dependent property. A person can be at different
developmental levels at the same time, because the levels make sense only in the
context of actual thinking or acting. The second and third aspects place the theory
more in the perspective of learning theory and behaviourism, which were the theories
Fischer worked with at the beginning of his scientific career. Fourth, it places a strong
emphasis on the dynamic aspects of development. This is clear from the contextual
aspect of, for instance, cognition, defining cognitive level as an interplay between the
person's capacities and the context in which those capacities become manifest. But it is
also present in the concept of skills, which are structures of elements interacting in
accordance with the principles governing dynamic systems. Finally, the theory tries to
relate major changes in behaviour and the underlying control systems to changes in
brain capacity and brain growth. The relation is believed to be reciprocal, not from
brain to behaviour alone.
The key to the structural part of the theory is the notion of coordination into a higher
 The first steps
order structure. Take a representation. At the lowest level, we use it in the form of a
single representation, for instance, a child's representation of his or her dad. Given
several representations, you can try to coordinate them in the form of a relationship.
For instance, you can pretend that one doll is Dad and the other doll is Mom and you
let them act out some event. A relation between two representations (or any other unit,
for that matter) is called a mapping and this is also the name of the second level of
development. Once you understand these mappings, you can coordinate two mappings
in a representational system. For instance, you understand that your dad, who owns a
pet shop, is also a salesman and a shop owner and he sells a goldfish to the mother of
your best friend Nancy, so you know that Nancy's mother is a customer. Finally, you
can coordinate mappings, that is, relate them to one another in a structure called an
abstraction. Abstractions are qualitatively different from representations, but are built
out of them. They are, so to speak, fourth-order representations, since the single
representation was a first order, the mapping a second order and the system a third
order representation. For instance, you have obtained an abstract and generalized
notion of a father role, or the role of a shop owner. The whole process will now repeat
with abstractions instead of representations as building blocks.
Representations and abstractions are called tiers, that is, they are basic types of
behaviour control systems that are qualitatively different from one another. The basic
control system at birth is the reflex. Reflexes get coordinated in the way I described for
representations. The fourth-order coordination transforms the reflex into a control
system of a different nature, the sensorimotor system. Four coordinations further the
sensorimotor has turned into the representational system, which in turn shifts towards
the abstract tier. The whole structure of tiers and levels1 - which are the steps of
coordination - looks very deductive and almoit algebraic (Figure 8.1).
Over the years, Fischer and his collaborators have gathered ample evidence for the
assumption that the qualitative changes (the leaps from one level to another) take place
I
I
in the form of growth spurts. These spurts take place in a relatively narrow age range.
The spurt is of course indicated by a quantitative measurement, which usually
amounts to a score on a test aiming at a particular level, for instance the level of
mappings of representations. An ordinary test, however, will not show the spurt at all,
but will demonstrate a pattern of monotonic and gradual growth. Only a test that
captures the child's highest performance level will be able to reveal the increase in rate
that characterizes a growth spurt. This type of test requires that the child is supported,
given information and presented with the right type of prompt. In Chapter 1, I
explained that both growth levels, one measured with an ordinary test and the other
under practice and support, are both real properties 'of the child's skill level.
There is something peculiar with the growth spurt, however. Figure 8.2 shows the
growth curves for arithmetic understanding. (under supportive testing conditions).
I They resemble an S-shape only very roughly (note that the first part of the S is missing
for the level 7 data). The curves seem to consist of an S-shape built from smaller
I

S-shapes that are added on top of one another. In fact, each of the curves is the result of
repetitive growth spurts. Each time a new level emerges, the preceding level receives
extra input from the successor. It sets into a new spurt and stabilizes at a higher level,
 Building a model: a case history

Figure 8.1 A spatial representation of the structure of the four levels in each tier: a level
is built out of 'atoms', which form 0-, I-, 2- and 3-dimensional structures.

until a second new level emerges, which makes it increase again and so on. This growth
pattern is indicative of the connected and coordinated nature of the levels. Finally, the
growth curves are showing not only spurts and plateaux, but also real regressions,
performance dips. This is quite clear from the level 7 data, for instance, but it can also
be found in data on reflective judgement (Figure 8.3).
In summary, Fischer's theory of dynamic skill development combines a qualitative
structural view, offering clear-cut definitions of the components and elements to be
discerned, with a quantitative growth pattern that is varied and complicated in an
interesting way.

Foundations for model building

The structural nature of the model and the definition of 'growers'
The starting point of a dynamic growth model should be a set of theoretically and
empirically justified assumptions about the mechanisms underlying the development
 Foundations for model building 217

Figure 8.2 The growth curves of arithmetic understanding at levels 7 and 8: an alternating
pattern of increases and decreases constitutes a global step-wise process.

Figure 8.3 The relationship between growth in two succeeding levels follows a pattern of
positively and negatively correlated switches in growth direction (incfeases and decreases).
218 Building a model: a case history
empirically studied. In practice however (and this also happened in the project that
Kurt Fischer and I undertook), some theoretical problems remain unanswered or at
best partially answered by the time the computer generates the first curves. This is an
acceptable procedure as long as the model building and the theoretical reflection
maintain a sort of dialectical tension and one continues to inspire the other.
A basic weakness of the growth model is that it cannot explain the dynamics behind
the qualitative emergence of new levels. Why does a coordination of reflex systems lead
to a structure that we recognize as a sensorimotor action? There is no way for the
dynamic growth model we have explored in this book to answer those questions. In
fact, we have to take the qualitative distinctions for granted. The only thing we may
hope to model is the time of their onset and the form of their growth curve.
A great advantage of a structural theory is that it enables the model builder to
determine easily what the growers of the model will be. Recall that a grower is any
component in the model to which a growth equation applies. Fischer's theory makes a
distinction between potential growers at different levels of generality. The most gen-
eral distinction is that between tiers: reflexes, actions, representations and abstractions.
In principle, it is>ossible to specify a dynamic growth model for these four basic
control systems. They form the highest level of aggregation in the model (and to avoid
confusion with yet another meaning of level, let me call them aggregations, instead of
levels of aggregation or levels of generality). The next aggregation is what Fischer calls
levels (later on, we started to call them stages, to avoid confusion with the term quanti-
tative level). They are a combination of a tier and a level of coordination, for instance,
the level of abstract mappings. There are thirteen such levels and they define the
growers in a model that is one aggregation level below an eventual growth model of
tiers. The next aggregation is that of general task domains, such as arithmetic knowl-
edge or reflective judgement. It is further subdivided in different tasks or tests. Below -
or I should say next to - the aggregation of the task level is the aggregation of support
context. In a sense, support context is a dimension that cuts through the other aggrega-
tions because it applies independently to all other aggregations.
Fischer and I decided to take the empirically measured variables as the growers in
the model we wanted to build. These variables were performance measures of tasks in
different domains (arithmetic, reflective judgement), administered under either
'spontaneous' or 'practice and support' conditions. We never asked ourselves whether
the growers we discerned were 'real' in the sense that they were the real independent
components of cognition. The reason is that there are no such things as independent
units that can be separated from one another by an absolute distinction. In biology, you
can distinguish units in the form of species and use the criterion of reproductive
isolation, but not in the domain of psychology. What we define as a unit is determined
by the choice of an aggregation level and by the distinctions that can be justified there.
The identification of a grower is a matter of justifiable theoretical definition, empirical
data and common sense. I believe that we can choose for many different levels of
aggregation or description and build a model either at the level of tiers or at the level of
highly specific tasks and still uncovei- the same basic dynamics and the same basic
principles of change.
 Foundations for model building 219
Empirical data and theoretical time series
A dynamic growth model generates a time series for an eventual individual subject, not
a group. The parameter values it uses are properties of a single person. The data that
Kurt Fischer and I used, however, are based on cross-sectional research designs. We
had no individual data that span the total age range of 25 years. There is very little hope
that such data will ever become available. Moreover, it is plausible that a longitudinal
study of the kind of developments we were interested in is impossible in principle. It is
likely that the procedure used in the longitudinal data collection itself, testing under
help and support conditions, considerably affects the course of the growth pattern it is
supposed to measure. This is no major disadvantage from the point of view of a
dynamic theory, since such a theory would incorporate the testing itself as one of the
beneficial factors. I t is devastating, however, as far as generalization of the data over the
untested population is concerned. I have said earlier in this book that generalization
over a population is not the first concern for developmental psychology. What we
should seek is generalization over time, that is, over how mechanisms operate, not
necessarily what in a majority of cases they produce. I am afraid I have not been able to
convince my working partner of this view, but I have not tried particularly hard either.
How can cross-sectional group data be used in building a model that seeks to
reconstruct the mechanism of change as it applies to individual cases? First, the group
data provide a reasonable idea of developmental successions: What comes first, what
follows what? Second, they can be treated as constraints on possible individual
patterns. For instance, if the data show a dip followed by a leap, it is highly likely that a
dip will occur in individual data, although its magnitude may show considerable
differences. Finally, the variance for each measurement point indicates how much
individual variation we can expect at different ages.
I must admit that, in the course of our discussions, Kurt and I never really bothered
about the fact we were using group instead of longitudinal data. In the Queteletian
spirit, we simply took the group curve as a typical individual curve that we would try to
model. Once a reasonable model had been worked out, we could start varying the
parameters and study their effect on the form of the curve. The theoretically produced
variation could then ideally be checked against the data, for instance in a comparison
between the theoretical variation and the corresponding empirical variance.
What do the available group data mean in terms of a person's potential growth level?
The data were of two different kinds. The first type consisted of percentages of tasks
passed in an age group. For instance, subjects were presented with two different
reflective judgement tasks. Their answers were scored on the 'single abstractions' pass
level. That is, if the subjects used single abstractions in their explanations they received
a score '1' on the criterion single abstractions. The maximum score of a subject was 2,
because there were two problems. The sum of scores over the group - say 12, out of a
maximum of 2 for each of 10 subjects -would then amount to aApercentagescore of 60
per cent (12120). We assumed that the group score could be taken as an indication of
the knowledge or skill level of our imaginary subject. Our imaginary subject, we
reasoned, would have a probability of 60 per cent to answer a reflective judgement
220 Building a model: a case history
problem at the single abstractions level. Whatever the knowledge or competence of our
subject might be like, it would be represented by this probability. I n Chapter 3, I
introduced the notion of a pass level dimension. I t was considered an abstract measure,
concatenating all possible control variables affecting a person's answer to a reflective
judgement problem. The score of 60 per cent means that from the collection of all
possible problems in all possible combinations of control variables, the subject would
answer 60 per cent correctly. Of course, we do not really know at the moment whether
this inference from group data to potential individual characteristics is justified. At
present, however, it is about the only thing we can do, since there are no individual
time series data available. The second sort of growth curve that we used was based on a
mean score. Raters gave stage or level scores to the responses of the subjects. For
instance, in the reflective judgement study they estimated the skill level demonstrated
in each of the questions asked by the experimenter. T h e levels were added over
questions, problems and subjects and resulted in a sort of Guttman-scale score for each
age group in the study. This score represents the highest level that can reliably be
attributed to the subject. The attribution of levels is completely defined by the
available scales and tests. For instance, if a score of 75 per cent correct on a level 7 task
is defined as a a reliable indication of the subject's correct understanding of level 7 rules
and principles, then level 7 is what the subject is.
What we did more or less implicitly here, using a group score as an approximation of
potential individual scores is common practice in developmental psychology, or
psychology in general for that matter. The problem is, however, that we have no
justification for the belief that a sequence of average scores over age ranges resembles a
succession of scores in a potential single subject. Subjects could very well show very
different magnitudes of oscillations over time. We have seen that sudden transitions in
individuals add up to a smooth S-shaped growth curve. Averaged chaotic oscillations
would very closely resemble a stable state. But, since the group curves are all we have,
there is no other choice than to use them, even if there is a chance they might be quite
misleading. As my grandfather, who even you will recall managed to build wooden
shoes out of concrete, would have said, 'If there's no concrete to build with, use wood
and if there's no wood, find yourself some tobacco to chew on'.

The characteristic data pattern and the

dynamic building block
Combined support and competition
One of the first steps in setting up a dynamics model of development is to form an idea
of what is characteristic in the data. Which characteristic pattern should be the focus,
or rather the starting point, of the model building? Since my Belgian culinary
background still colours my imagination, I sometimes feel that the decision about what
is characteristic in the data amounts to deciding whether a glass of beer is half full or
half empty. Which of these two descriptions is true of the actual glass of beer depends
 The characteristic data pattern and the dynamic building block 221
on what has happened before you observe it. If the glass has been with the drinking
customer, then it is already half empty. If it is still in the hands of the bartender, you
had better say it is still only half full. Especially if we look at developmental data, we
should take the time course into account and try to understand what happens at a given
moment in the context of what preceded it and of what follows.
Nevertheless, the theoretical expectation often determines what is viewed as
characteristic in the data. It is customary, for instance, to see a developmental curve as
the distorted representation of a linear and monotonic trend towards a steady-state
level. The fluctuations along the way are consequently attributed to error variance.
Alternatively, the data can be seen as a set of sudden jumps to a new qualitative overall
level and invoke the notion of de'calage to explain away, in a sense, evidence for
smoother and more gradual change. One of the great advantages of a dynamic model is
that the theoretical study of dynamic interaction patterns, for instance of connected
logistic or transitory growers, reveals a rich collection of different growth and change
patterns. When attempting to come up with a model for an empirical curve, we not only
have all these different patterns in mind, but also know the dynamics and parameter
values that characteristically produce them. It is the combination of the empirical
fluctuations on the one hand and on the other the model we expect to operate in the
data domain in question, that guides our intuition in determining what the
characteristic pattern in the data will be.
At the beginning of our cooperation at the Center, Kurt and I started from different
backgrounds that, we knew, had to be reconciled for us to build successful models.
Kurt was familiar with structural theory and the data, knew how they had been
acquired and put together and knew how they related to one another in the many
domains he and his collaborators have been studying over the years. I was acquainted
with the particular brand of model building I had developed, with its possibilities but
also with its lacunae and problems. We both believed in the importance of explaining
growth curves and in the ecological framework of resource-dependent growers that
either competed with or supported one another to different extents.
We believed that the basic pattern in the data was a combination of support and
competition between different levels (or 'stages'). If you take a closer look at the first
half of the growth curves in Figure 8.3, you will notice that the curve for level 7 goes up
first, then down, then up again. Its decrease precedes the first increase of its successor,
the level 8 grower. Then, as level 8 increases more and more, level 7, the predecessor
goes up again. Although the idea grew along with the later model-building efforts, we
soon agreed on the following basic pattern. Although the predecessor (level 7) is clearly
conditional for the emergence of the successor (level 8), the predecessor also profits
from the growth in the successor. For instance, understanding the relationship
between addition and subtraction will, in general, improve your understanding of
addition and subtraction separately, and therefore have a positive effect on the level 7
test score, which measured these separate understandings. But then there was
something else we had to cope with, the persistent and consistent dip in the predecessor
curve announcing a leap in the successor.
If you have two growth curves of growers that you know are in some way connected
 Building a model: a case history
or related to one another and both curves go up, then there is reason to expect a
supportive relationship from one to another. If one goes up while the other goes down,
there must be a competitive relationship, for instance, when both feed on the same
basic resource such as time and the time spent in mastering one skill is, by definition,
subtracted from the time you could have spent in mastering the other. The problem
was that the empirical curves showed both increase and decrease, that is, both
competition and support. If competition and support are added, they behave like the
result of two opposing forces, or like the sum of a negative and a positive number, the
result of which is, trivially, either negative (competition), positive (support) or zero
(neutral).
Previously, I had worked on models in which the competitive or supportive effect
was not related to the actual growth level of the supporter or competitor, but to its
increase or decrease per time unit (see van Geert, 1991). For instance, a parent may
adjust his or her amount of help to a child, not to the actual competence level that the
child has achieved, but rather in relation to how much the child has learned or
improved over a past period of time. I figured that this model could solve our problem.
We assumed that the previous level, say understanding of addition or subtraction
separately, profited from the degree of understanding at the next level, that is, of how
addition and subtraction relate. But the discovery of how addition and subtraction
relate to one another requires time and attention and probably also a reorganization of
the knowledge about addition and subtraction separately. Time, attention and
reorganization of knowledge are related not to the actual level of knowledge acquired,
but to the increase in the knowledge, that is, to the effort and time invested in the actual
learning or mastering. We started with the simple assumption that the negative effect
of level 8 knowledge on the preceding level 7 knowledge was proportional to the
amount of increase in level 8 knowledge. The positive effect of level 8 on level 7, on the
other hand, should be proportional to the actual growth level of level 8. Remember that
I discussed this model in Chapter 7, but at the time we were struggling with this
problem that chapter had still to be written.
The logistic growth equation for a grower Lnat a developmental level n, is defined as
a function of three parameters: its own preceding growth level, a positive parameter
,
operating on the growth level of a succeeding grower Ln+ and a negative parameter
operating on the increase or decrease of the succeeding grower. If the grower is level 7,
for instance, it would depend on level 8, level 8 would depend on level 9 and so forth.
This sort of chainwise connection between growers became our standard model for
hierarchically connected growers, that is, of connections between growers at different
levels of complexity or cognitive hierarchy. We employed the following growth
equation for a connected grower:

Recall that there is a second form of growth, which specifies a much sharper
transition pattern. We started the model building with this equation which we called
the 'ordinary' version. Some time later, when I had worked out the transition version of
 The characteristic data pattern and the dynamic building block 223
the growth equation, I explained its logic to Kurt and showed the different growth
patterns it generated. At the onset, he was not convinced of the value of the transition
version, because he suspected that the ordinary logistic equation already did what I
wanted this new model to do, namely make the growth rate dependent on the growth
level already achieved. My main justification for this new transition model was that I
assumed there existed a class of growers whose actual growth depended on what I
called 'learning encounters', that is, actual encounters with a context that enables the
child to learn. I claimed that the probability of running into such a learning encounter
was a function of the growth level already achieved. Imagine a child of three who
understands only very little about conservation. The child will encounter a lot of
conservation events - playing with water or sand and containers, for instance - but
most of these events will go unnoticed and will not be detected as potentially
problematic. Only very few conservation events are simple and salient enough for the
3-year-old to be detected as potentially problematic and learned from. Thus, even if
this child is a fast learner, able to make quick and correct inferences from the
information given, he/she would have little chance to practise these learning capacities
as long as his/her knowledge about, for instance, conservation, is still very immature.
In the transition model, the growth rate function is defined by a parameter and a
variable. The parameter is the intrinsic speed of learning and corresponds with the
familiar growth rate r; the variable is the probability of encountering a learning
experience, which is a function of the growth level attained (relative to the carrying
capacity). The equation for the transition v e r s i ~ n ~the
i s following:

Fischer and I agreed that the transition version was at least a plausible alternative to the
conventional logistic model. We decided to build two parallel versions of our
developmental model, one based on the standard logistic, the other on the transition
version. We called the transition model the cubic version because it contains a cubic
term, L,(,,~,resulting from multiplying the Ln(,)just preceding the brackets with the
Ln(,)' inside the brackets.

Precursor connections between growers

With equations 8.1 and 8.2, the growers are connected in a way that appears to be
opposite to what you would expect intuitively. The equations make the predecessor
dependent on the successor for further increase and temporary decrease. Put
differently, an earlier level profits from a later one. We thought we had good reasons for
this, from both the available data and probable mechanisms. For instance, when you
understand the relationship between addition and subtraction (a later level), you are
likely to achieve a better understanding of addition and subtraction separately. For
instance, you are capable of checking the result of a subtraction by making an inverse
addition, and by doing so you may correct or avoid errors.
224 Building a model: a case history
In a hierarchical model such as Fischer's, however, the most obvious connection
between any two growers would be between a predecessor and its successor. For
instance, understanding about addition and subtraction separately is a precondition for
acquiring an understanding of how they both relate in a more abstract sense. In fact,
the whole structure of levels is based on the idea that the predecessor is a necessary
building block of the successor. Each successor integrates, in a sense, the structures of
its predecessors into a new unit. How can a growth model deal with this integration
process? The answer is that, in a qualitative sense, it cannot. The growth model
describes only growth levels and changes in them. It cannot account for structural
properties other than the quantitative growth levels of the growers it refers to. On the
other hand, the growth model can deal with predecessor or prerequisite connections
between growers in a relatively simple way.
In a model I had been working on earlier, I had defined a precursor as any grower
that prevented the dependent grower from growing, as long as the precursor threshold
had not been reached. The understanding of addition and subtraction as separate
operations is clearly a precursor to the understanding of how they are related. It is
obvious that a child will make hardly any progress in learning to understand how
addition and subtraction are related if he or she still lacks a minimal understanding of
addition and subtraction as separate operations. In Chapter 7, I discussed the notion of
precursor and suggested the term prerequisite to better capture its meaning in terms of
a growth model. The prerequisite is the minimal level a grower should have reached in
order to release the growth of the dependent grower. If only a little understanding of
the addition and subtraction operations is required to get the understanding of their
relationship off the ground, the prerequisite level can be set to any low value, for
instance 20 per cent of the prerequisite's K or carrying capacity. Higher levels
represent stricter prerequisite demands, and we may even demand, if there are any
theoretical and empirical reasons to do so, a prerequisite level that is near to carrying
capacity.
The simplest way to put a prerequisite relationship into a growth equation is to
multiply the growth parameters of the dependent grower by a parameter p that takes
the value 0 as long as the prerequisite threshold has not been passed and 1 in the other
case. The dependent grower will then stay at its initial level as long as the prerequisite
has not reached the minimal level necessary for the emergence of the dependent
successor. For instance, to equation 8.1, I add a bimodal variable prec, which specifies
the prerequisite relationship. This results in the following equation:

Ln,t + 1) = Ln,t) + (7, - "nLn(t)IKn SUP Ln + l ( t ) - comp(Ln + 1(t) - L n + ~ ( t - ~ ) ) ) ' P t

Quadratic version: if Ln- > prec, p, = 1 - Ln(,,, otherwise p, = 0
Cubic version: if Ln-, > prec, p, = 1 L,(,,~/K,otherwise pt = 0 (8.3)

By the time we started to write down a simple version of the connected growers model,
we discovered that we had to specify whether or not the support and competition were
reciprocal. The understanding of the addition and subtraction operations (level 7)
profits from the understanding of how they are related abstractly. Thus as soon as level
8 starts off, level 7 increases along with it.
 The characteristic data pattern and the dynamic building block 225
Does the abstract understanding, in turn, profit from the increase in understanding
them separately? And, similarly, is the competitive relationship reciprocal, in that an
increase in understanding them separately negatively affects the growth of understand-
ing how they are related? This is an example of the sort of question that typically comes
up in the context of mathematical model building. While thinking about the equation
specifying the growth of the successor, we asked ourselves whether we should make it
symmetric to the equation of the predecessor, except that the prerequisite condition
cannot apply from the successor to the predecessor. I think that I took an option for the
'aesthetic' solution, that is, for symmetry and argued for a reciprocal support and
competition. Kurt and I had, as far as I can recall, no fundamental justifications for
why it should be either one solution or the other.

Testing the behaviour of the model

At this point in the development of our model, we wanted to know how a simple
version consisting of only two growers would behave, given the conditions of
competition and support defined earlier.

Model 8.1 Testing the simple model of connected

growers
Define a spreadsheet model with a grower A in the A column, grower B i n the C column,
and a prerequisite parameter in the D column. For convenience, start with cell A l l and
use the following equation for the A grower:

Since K is, by default, set to 1, it is omitted from the equation.

Cell C11 is the place for the equation of the B grower, whose prerequisite is the A
grower. The equation is:

In cell A10 write 'INPI' and in C10 'IN-6'. I N 4 and IN-B are the initial or starting levels of
grower A and B, respectively. Cell D l 1 should contain the prerequisite formula, which is as
follows:

Then copy A1 1..Dl 1 to A12..D500 and define a graph that shows the values of the ranges
A1O..A500 and C1O..C500, respectively.
The equations do not make sense unless you define a list of parameter values and
names. The list of parameters can be insertedin range F1..F9. Range GI ..G9 contains the
list of parameter names. The following list shows the parameter names (right-hand side)
with three possible sets of parameter values.
 Building a model: a case history
value set 1 value set 2 value set 3 parameters
r-a
r-b
c(B > A)
c(A > B)
s(B > A)
s(A > B)
prec
in 4
in-B

A parameter name such as c(B > A) means 'the competitive effect of B on A'. Note that I
used a different sort of notation earlier, namely c(B -+ A). Unfortunately, the arrow symbol
is not available in the spreadsheet.
Do not forget to define the parameter names for the spreadsheet. When, for instance,
cell F1 contains the value 0.2 and GI contains its name, r 4 , define a parameter name
' r 4 ' whose value can be found in cell F1. Each column of values in the list above should
be seen as a set of values that you can try in a single run of the model. Begin with the
column at the left.
After having tried the three combinations of values, you may try any arbitrary set you
want and see what sort of curves they produce.

The first thing we discovered in our simulations was that with a set of equations like
this, we could obtain a growth pattern that was quite close to the empirical data. The
magnitude of the dip could be manipulated by augmenting the value of the parameter
C(B > A). Figure 8.4 shows the curves resulting from employing the parameters at the
left (for the first 120 data points).
We learned at least two unexpected things from numerically testing the model. The
first is that if competition and support are reciprocal - that is, if A competes with and
supports B and vice versa - the trajectory soon becomes unstable. Try the list of values
from the middle column, where only the competition has been made reciprocal. The
result is an interesting but probably not very realistic chaotic oscillation of one grower
around the other (see Figure 8.5). Maybe there are growth phenomena where the
growers drag each other into chaos, but the empirical data that Fischer and his group
have collected do not appear to show any. Maybe in reality the growth of the successor
is indeed negatively influenced by the growth in the predecessor, but if that is the case
the influence should be small and almost negligible.
The second unexpected effect we observed was that if we set the prerequisite value
close to the carrying capacity level, that is, close to 1, we obtained turbulence in both
the prerequisite grower A and the successor grower B (see Figure 8.6). Although there
is no evidence for such turbulence in the data, there was no reason for us to discard this
particular side-effect of the model. The data are cross-sectional group data and we
treated the resulting age curve as a potential or characteristic individual growth curve
applying to at least a significant number of subjects. But if a real subject showed
anything like a turbulence, that is, a period of instability before taking the leap to a
 The characteristic data pattern and the dynamic building block 227

Figure 8.4 The prerequisite (e.g. stage 7 skills) shows a dip before it increases again, due
to the increase in the successor (e.g. level 8 skills) by which the prerequisite is supported.

Figure 8.5 A reciprocal and symmetric competition-and-support relationship between a

prerequisite and its successor produces instability and chaotic oscillations; in the model,
prerequisite-successor relationships should be strongly asymmetric.
228 Building a model: a case history

Figure 8.6 Turbulence in the prerequisite grower is caused by setting the prerequisite
level of the successor high, for instance, close to the carrying capacity of the prerequisite
B.

higher level, that phenomenon would not show up in our group data. In our discussion
of conservation development, however, we have seen that instabilities may announce
I
transitions to higher levels. A similar effect might occur in a variety of other areas,
including the child's understanding of arithmetic operations that Kurt and I took as
our modelling domain.
Another thing we found while experimenting with the simple model was that it
would explode or crash if the parameters went beyond particular threshold values. Too
much support or competition would result in unmanageable chaos. Crashes can
sometimes be avoided by putting a constraint on the results computed in each cell, by
preventing each cell from becoming smaller than an arbitrarily small number, which
functions as the minimal growth level of the grower described in that

Building fhe model

The first model of connected growers
Even in a monastic place like the Center for Advanced Study, worldly duties keep
knocking on your door and they often prevent you from reaching the spiritual heights
that you dreamt of when you were packing your luggage. The usual writing obligations
and overdue papers competed fiercely with the model-building work for the one
resource that we thought would be abundantly available: time. So, after having worked
 Building the model 229
on the basic model we did not make much progress for some time. That period might
have functioned as a sort of incubation time, since the model building progressed
rapidly once we actually got to the computer and started writing the sets of equations.
The whole process was in a sense an illustration of S-shaped growth in the model
builders themselves, if not of a transition.
One of our first working models had six different growers connected in the way
described by the basic model. Each grower was conceived as a specific level in Kurt's
structural model, applied to a specific domain of application. The first grower, for
instance, could be the level of sensorimotor actions on collections of objects, the second
would then be the level of sensorimotor mappings, the third that of sensorimotor
systems and so on. Each grower (that is, each level in Kurt's model) would have its
immediate predecessor as its prerequisite. For instance, we would set the prerequisite
for the emergence of the second level to 50 per cent of the carrying capacity of the first.
That is, once the first level had reached a score equal to 50 per cent of its carrying
capacity, the second, dependent level was allowed to start growing. We did this for each
set of successive growers in the row.
If you build a model of connected growers, there is a problem with the initial state.
The initial state in the spreadsheet model is the value of the first cell of each column
that represents a grower. In reality, it is the state of knowledge or mastery of each
grower at the time you have chosen as the starting point for your modelling. Assume
you take, for instance, sensorimotor actions as your starting point and want to continue
up to the highest level, principles. At the age the child~iscoping with the problems of
sensorimotor actions, the level of principles is still a non-existent realm. Technically,
the initial state of principles at the age of sensorimotor actions should be equal to zero.
Recall however that a grower with an initial state of zero will always remain zero. I had
been confronted with this initial state problem before and I suggested the following
solution. The initial state of each successor could be artificially set to one-tenth of the
initial state of its predecessor. This low value, in addition to the prerequisite, would
account for the spacing of the growth onset times.
Not only the initial state levels had to be decided on at the beginning, but also the
carrying capacity levels of each grower. The default carrying capacity is set to 1 and this
is what we used in the first model; each level of the developmental scale had the same
equilibrium level, 1. This choice produced an undesirable result. We wanted each level
to make a new leap when its more advanced successor emerged. Since the levels were
connected in a chain of overlapping support relations, a level would make a leap each
time a new level emerged, irrespective of whether it was directly connected with that
new level or not. The model did what we wanted, but it did so in a very rigorous way.
For instance, in the model the emergence of the level of principles, at the age of
approximately 24 years, had a considerable effect on the mastery of sensorimotor
actions, which develop around the age of 4 months. That such an effect occurs in reality
is highly implausible. What we actually wanted was that the of every new and
more advanced level for some earlier skill decreased as the distance to the newly
emerging skill increased. Maybe the discovery of principles has some effect on
sensorimotor actions in young adults, but if so, that effect must be very small and
230 Building a model: a case history
practically unobservable. The effect of emerging sensorimotor mappings on sen-
sorimotor actions, on the other hand, is assumed to be considerable.
We solved this problem more or less by chance. We figured that it was not a good
principle after all to set the carrying capacity of each grower to 1. The level of
problem-solving complexity we can achieve with, for instance, abstractions, is
considerably higher than the level we can achieve with single representations. This, we
reasoned, should be reflected in the carrying capacities; lower levels should be given
lower Ks. It can be doubted, however, whether this reasoning is correct. Each different
grower operates on a different and independent scale, even if we put them together in
one graph. Numbers on each of the grower scales have a meaning only in terms of that
scale and what matters is how the numbers change, not their absolute level. If that
reasoning is correct, we can attach any possible carrying capacity level to each grower
and account for its relations with other growers in terms of the parameter settings.
Anyway, we gave the growers carrying capacities that corresponded with their
ordering in the model - later ones got higher Ks. By so doing, we obtained the result
that we wanted, a decreasing pay-off of later growers on earlier ones. The emergence of
principles did something to increase sensorimotor action levels, but that something
was imperceptibly small.
It is easy to show by analytical means why this increasing pay-off effect occurs in the
case of connected grower models. The effect of a grower E on an initial grower A, for
instance, is equal to:

and if the Ks are smaller than 1, that effect is only a fraction of the effect E has on its
immediate predecessor D, namely:

As is often the case with model building, the analytical proof came after the more or less
accidental discovery of the principle.
I t turned out that there was a second way to avoid the problem of long-lasting effects
of late growers on the earliest ones. This was again discovered by accident. Since we
believed that the so-called cubic version of the model, the one I presented in Chapter 5,
provided a good if not better alternative to the ordinary logistic one, we built a cubic
version of the model of connected growers. One of the interesting side-effects of that
model was that it greatly flattened the effect of later growers on much earlier ones. The
sixth grower in the sequence had a considerable effect on the fifth, but only a very
minor one on the first, for instance. This is an interesting example of the way in which
dynamic models display a variety of emergent properties and effects that have not
explicitly been built into the equations. The reason for adopting the cubic model was
theoretical. As an unintended side-effect, it got rid of the undesirable late growth in the
earliest growers.
Figure 8.7 shows the growth curve for a model of six connected growers, with
different initial states, prerequisite relationships and carrying capacities of increasing
magnitude.
 Building the model

Figure 8.7 A model of six hierarchically connected growers: the overall pattern is one of
step-wise increase, with temporary dips in prerequisite growers corresponding with a leap
in the successor.

Within and between level growth

By this time Kurt and I were happy with the result of our model building. We had not
made any attempts so far to fit the model to the actual empirical curves. What we had
achieved, we thought, was a model that fitted the essential qualitative properties of the
data. The growers would increase in a stepwise fashion, leaps towards a higher level
would be preceded by regressions, the interval between any two successive growers
would increase and the effect of later growers on earlier ones would decrease with
distance between the growers. Kurt started to build a few models that accounted for
connections between and also within levels.
A level, for instance, the level of single abstractions, is not some monolithic cognitive
structure. Each level comprises a variety of different tasks, skills or task domains. For
instance, the child's abstract understanding of arithmetic operations is different for the
various operations - addition, subtraction, multiplication and division. The operations
vary in difficulty and they are seldom if ever acquired at the same time, with the same
speed or at the same level of mastery. Each of the arithmetic operations can be
conceived of as a separate grower. This made the structure of the connections more
complicated. For instance, the child's understanding of addition at the level of single
abstractions would be connected to his later understanding of how addition and
232 Building a model: a case history
subtraction relate, in terms of an abstract mapping. We had modelled this connection
in terms of simultaneous competition and support. But inside the level of single
abstractions, the child's knowledge of addition would no doubt be connected with his
knowledge of subtraction.
In the model of these within and between level connections, we employed the same
connection principles for the within as for the between level links. For instance,
addition would profit from the acquired level of understanding in subtraction, but be
hampered by the actual increase of subtraction understanding. The result was a model
that behaved pretty much like the previous one, in which we modelled only between
level connections and considered the separate overall developmental levels as growers.
This result supported our belief that the model produced a good qualitative fit with the
data, irrespective of the level of aggregation of growers we chose.
Since some of the empirical data were based on a sort of sum scores of tasks on
different developmental levels, Kurt wanted to know if the sum of growth levels would
show the characteristic dips and jumps of the empirical data. For each simulated point
in time, he summed the available levels. The result was a stepwise growth process, but
there were no dips. Whenever the dips occurred, they were compensated by the
increase of a higher level grower that actually caused the dips. We were quite
concerned about this problem, although we would later find out the error was not in the
model but rather in our interpretation of the empirical sum scores, which were not real
sums but rather maximal levels. Nevertheless, this problem made us look out for
several different solutions that would produce temporary fallbacks that would not be
cancelled out by correlated increases in another grower.

Bumps or dips?
Meanwhile Robbie Case, with whom we met every week to discuss model building,
had been working on his model of growth in a central conceptual structure. That
structure was related with several task domains that grew independently of one
another, although each grower affected each other one, via its contribution to the
central conceptual structure. Robbie wanted the supportive effect of a grower on
another one limited to the period of actual growth. He could have done this by making
the support a function of the increase in the supporter, much as Kurt and I had done
with competition. But Robbie chose a different solution. He multiplied the supporter's
level, Ls(,, by the unutilized capacity for growth in that supporter, namely:

Ls(t1- (Ks - LS(t))IKS (8.9)

This is consistent with the way in which growth is kept within limits by the logistic
model, by multiplying the growth level by the unutilized capacity for growth,
(K - L)/K.It is easy to see that as Ls(,,approaches its carrying capacity Ks the value of
the fraction (Ks - LS(,))/Ks approaches zero which then reduces the effect of Ls(t,to
zero. The effect is that sometimes the supported grower temporarily overshoots its
carrying capacity. This overshooting lasts -only as long as the supporting grower goes
on growing. Whether or when this overshooting occurs is a function of initial state and
 Building the model 233

,Figure 8.8 Support from a grower B during the growth period results in a 'bump' on the
supported grower A.

parameter values, but the bump is a genuine effect of the dynamic principle from
equation 8.9 (see Figure 8.8).
With the Case model in mind, the idea arose that what we had conceived of as a
regression, introducing a leap towards a higher level, was not a regression in the sense
of a dip, but a bump, a temporary overshoot above carrying capacity level, levelling off
to carrying capacity as soon as the supporting grower ceased to increase.
In the hierarchical model that we had developed so far, we had made a distinction
between growers on different levels (such as single representations and representa-
tional mappings) and growers at the same level (e.g. a single-representation
understanding of addition and one of subtraction). We assumed, as Robbie Case did in
his model of central conceptual structures, that growers within a level support each
other only for the time they actually grow. Once they achieve their equilibrium level,
the mutual support dies out.
It seems rather arbitrary to make a distinction between'within and between level
support, in that the between level support continues even after the growth of,supporter
and supported have stopped. The justification for this distinction, however,' lies in the
relationship between carrying capacity and supportive factors. Earlier I explained that
if L is a grower and Ls is a grower supporting L, the carrying capacity of an L is
+
K - (r s - Ls)/r, for K the carrying capacity parameter (usually I), r the growth rate,
Ls the level of the supporting grower, and (lower case) s a parameter specifying the
extent to which Ls actually supports L.
234 Building a model: a case history
The difference between a supporter that goes on supporting and one that ceases to
do so as soon as its growth has come to a standstill is that the first supporter contributes
to the carrying capacity of the supported grower, but the second does not. It is
plausible to assume that the emergence of a new form of understanding or a new skill
level, for instance the discovery of the principle of representational mappings,
contributes to the carrying capacity of the existing skills and understanding. This
higher developmental level is a new source of information and understanding and will
contribute to further mastery and understanding of the lower levels. That is, it enables
those to rise to a level higher than the one they would have achieved without the new
knowledge. Compare this with the relationship between different task domains at the
same level. Opening a new task domain, where the same principles reign as in the task
domains already explored and mastered, might certainly have a temporary positive
effect on the already existing tasks of that same developmental level. But it will not have
a lasting effect, since it adds nothing significantly new to the understanding at that
level.
What happens if equation 8.9, causing the temporary bump, is built into the
equations of hierarchically connected growers? It will produce a temporary overshoot
of the carrying capacity of growers within the same developmental level. Growers
within a level were identified as 'tasks', for instance as the task 'addition' and the task
'subtraction'. The overshoot occurs during the growth process itself. It will then
regress to the normal carrying capacity level. Later, a grower at a higher level will
emerge and will cause the earlier grower to increase its level. The whole pattern, from
growth to overshoot to carrying capacity then to growth again, looks like a temporary
regression or dip, announcing a leap. In the latter case, the dip will occur while the
higher developmental level grows. When the two effects are added, the dip is cancelled
out. The temporary overshoot, however, occurs in general before the jump (they are at
least independent of one another). Thus if the level scores are added, the
pseudo-regression is not cancelled out by the leap (Figure 8.9).

Compensating for changing carrying capacity

Recall that there are two meanings of the term 'carrying capacity' in the model. The
first is the fixed carrying capacity parameter, usually set to 1, that figures as the K
parameter in the logistic equation (and is left out if its value is 1). It would be more
correct to call it the carrying capacity parameter, instead of just carrying capacity. The
second meaning of carrying capacity is the equilibrium level of the grower. This
equilibrium level is an expression of the resources available to promote and sustain
growth in a growth environment. In the simple logistic equation, the equilibrium level
is equal to the carrying capacity parameter (usually set to 1). In a connected growth
model, however, growth is affected by additional variables, for instance a supporting
grower, L,:

Since the only variable in the equation is the supporting grower Ls, the variation in the
 Building the model 235

Figure 8.9 A regression in the sum score of three tasks is caused by a temporary
overshoot in one of the tasks.

equilibrium level is entirely determined by that variable. In our model, there was not
just a single Ls and a support parameter s, but a sum of levels multiplied by support or
competition parameters:

It is much more convenient to write the underlined part of equation 8. lob in the form
of a single variable, namely St. Thus, St is a compound variable consisting of the sum of
all supporting and competing growth levels multiplied by their own support or
competition parameter.
When I thought about these two meanings of carrying capacity, I realized that we
were making a mistake using equation 8.9, which subtracted the growth level from K,,
then divided by K, (K, - Ls)/Ks. If Ls's asymptote is equal to Ks, the value of the
equation will approach zero, as we originally expected. We should have realized,
however, that almost any variable in the connected growers model exceeds the level of
its K, parameter because of the effect of supporting growers. Thus as soon as L, grows
higher than its Ks parameter value (by default l), it would turn the value of equation
8.9 into a negative variable. Its effect would no longer be to stop the growth of the
grower that it modifies, but to make that grower go down and eventually disappear.
There is absolutely no reason to suspect that any grower would have a negative effect
on another one as soon as it grew bigger than some predetermined carrying capacity
value.3The solution to this problem is actually fairly simple; instead of using the value
of the K parameter, we should use the value of the K variable, that is, the real carrying
capacity. The carrying capacity variable is determined by the K parameter on the one
 Building a model: a case history
hand and by the values of the supporting (or competing) variables on the other hand.
Thus we changed equation 8.9 (the equation that confined the supportive relationship
to the time span while the supporter was actually growing) by substituting the variable
carrying capacity Ks(,, for the constant carrying capacity parameter K,:

The value of K,(,, was given by adapting equation 8.10a to the set of growers that in
turn support the supporter L,:

for Ks the carrying capacity parameter of supporter L,, rs the growth parameter of
L,, SS.,, the compound variable consisting of all the supporting and competing effects
on Ls and K,(,, the carrying capacity variable at time t of the supporting grower Ls.
For some sets of parameter values, the curves would grow wild at the end of the
trajectory and eventually make the whole system collapse. This happened primarily
with the curves tracing the growth of the earliest levels. They were the levels whose
magnitude increased the most, since they profited from each successive new level that
emerged. The problems we have had with the behaviour of equation 8.9, especially the
fact it related to an absolute carrying capacity, made me wonder whether the danger of
an eventual collapse could not be countered by employing relative instead of absolute
magnitudes. The idea is that it is not so much the absolute magnitude of a grower that
determines to what extent it affects another grower, but its magnitude relative to its
maximum. An easy way to introduce such a maximum value into the equation for
connected growers is to work with the growth levels divided by the corresponding
carrying capacities and not by the mere growth levels. Assume that we write the
equation for a connected grower in an extremely simplified form, namely:

pt is a name for the function that we would put within the brackets in the complete form
of the equation for connected growers. As L, grows higher and higher, the effect of any
constant parameters in p, will increase accordingly and eventually drag the system into
a chaotic oscillation. Instead of using the absolute value of L,, we should use its value
relative to its variable carrying capacity, Kt. Equations 8.10a and b specify the
magnitude of that variable. Instead of the absolute form of equation 8.13, we should
therefore work with relative magnitudes:

Although each of the steps described here seems logical once it has been explained, it
took Kurt and me a lot of trial and error before we arrived at these solutions. I am afraid
that we did just try the model and experiment with it to see what it did. When you find
out that it does funny things you did not anticipate, but that nevertheless resemble
your empirical data or expectations, you are happy with the result. If something
unwanted happens, you try to correct-it by going over the equation and trying to
understand what might have gone wrong. Of course, you should try to arrive at an
 Building the model
analytical understanding of what the equation does irrespective of whether the results
are anticipated or not, or wanted or not. But during the process of model building itself,
often the excitement takes over. The pleasure of doing some sort of combination
between a computer game and handicrafts often stands in the way of applying the
rational analytical methods.
Finally, with all the elements ready, it is time to present our 'master' equation that
we used to model the entire system of connected growers in development.

The equation for hierarchically connected growth

In an earlier stage of the work, Kurt and I had decided we would specify the model at
the aggregation level of tasks. They constitute stronger psychological units than the
levels or 'stages', which are more like abstract categorizations of task-specific skills.
The minimal number of tasks per level we needed in order to make the structure of
relations meaningful is two. In reality each level contains a multitude of tasks of very
different kinds (social tasks, numeric tasks, emotional understanding, motor tasks and
so forth). The strength of the relationship between any pair of tasks in such a set varies
greatly, dependent on the nature of the tasks related. Some tasks bear significantly
stronger relationships to one another, especially those that have a common general
structure, or that can be integrated in an overarching structure at a higher
developmental level. An example is the task addition and the task subtraction; the
operations are logical inversions and they combine intota higher level structure. The
most interesting connections between growers, therefore, exist between strongly
related ones. Two such growers suffice to test the dynamics. We called them Task1 and
Task2 and specified them on ten different levels. Just for terminological convenience,
we called these levels stages and numbered them StA, StB and so forth (see Figure
8.10).
The Fischer developmental model counts thirteen developmental levels, which we
now called stages to avoid confusion with the term growth level, divided into four
different major tiers. We decided to exclude the first one, the reflex tier, from the model
and confine ourselves to the ten levels that begin with the single sensorimotor action
level. The tasks we had in mind when we built the model were things like reflective
judgement or arithmetic operations. We felt that the developments at the reflex level
would relate to those tasks only in a very weak way (although they would be
indispensible as developmental prerequisites). Discarding them would make the
resulting model a little bit more manageable.
In summary, the model we finally built consisted of two related tasks over ten
successive developmental levels. Each task at each level was considered a separate
grower, so that we had a model of twenty connected growers.
For each of the twenty growers we wrote a specific growth equation. The basis of
that equation was the logistic form. We wanted to alternate between the ordinary
(squared) version of the equation and the cubic form (the sharp transitional form) and
used a cubic parameter to do this. In addition, we had to specify a precursor parameter
that would cause the grower to change as soon as the precursor level of the prerequisite
 Building a model: a case history

within level
~~ mdk2

Figure 8.10 The structure of between- and within level relationships in the model of
hierarchically connected growth.

grower was reached. The growth rate variable that caused the change in a grower was a
combination of its own growth rate and of a set of influences from the other connected
growers, which we called joint factors. Joint factors were based on two sources. The
first are the influences from the grower at the next higher level - that is the between
level factors. The second come from the grower at the same developmental level. They
are the within level factors. Finally, we wanted all these growth factors to modify not
the absolute magnitude of a grower, but the magnitude relative to the carrying capacity
variable.
Instead of giving the usual equation with symbols, I shall present it in the form of
the concept and variable names i n ~ o l v e d : ~

+
Level now = LeveLbefore Precursor-value (Leveibeforel
carrying-capacity) . (cubicsetoff-parameter - (growthate LeveLbeforel
carrying-capacity - growthrate - (Level-before/carrying-capacity) ^ 2 +
Jointactors)

rate = any number in normal growth rate range

Jointfactors = jointfactor_rate-parameter . (Withinfactors t

 Building the model 239
Betweenfactors)
joint.Aactor_rate-parameter = any number modzfying the efect of within
and between factors
Withidactors = supportwithinstage-parameter -1eveLofsupporter.
(1 - levelofsupporter/carrying~capacity~variable~ofsupporter)
- competitio~withinstage-parameter .change-ofsupporter
supportwithinstage-parameter = any number
competitio~withinstage-parameter = any number
carrying-capacity-ofsupporter = carrying-capacity-parameter-of-
+
supporter - (1 levelofsupporter~ofsupporter/
growthate-of- supporter)
change-ofsupporting-task = Levelbefore-ofsupporting-task - L e v e l
before-(1eveLbefore)-ofsupporting task
supporter-ofsupporter = grower supported
Betweenfactors = betweenstagesupport-parameter
levelof_nextigher_leveLof-task -
betweenstage-competitio~parameter .
change-of_nextigher_levelof-task
betweenstagesupportparameter =,any number
Precursor-value = 1 zf levelof-prerequisite-grower >
>
precursor-parameter and Sum (allpreceding~precursor~values)
precursorsetoLparameter
in all other cases precursor-value = 0
precursor-parameter = any number in growth range of prerequisite grower
precursorseton_parameter = any number specifying a number of
times the value of the prerequisite-grower > precursor-parameter ( z f
precursorseton_parameter = 1, the precursor-value = 1 when the
prerequisite-grower has crossed the precursor-parameter once)
cubic-setoffparameter = 1 zfcubic-parameter = 1
cubicsetoff-parameter = carrying-capacity/levelbefore
if cubic-parameter = 0
cubic-parameter = 0 or I
~arr~ing-capacity= carrying-capacity-parameter .( g r o w t h a t e +
jointactors)/growtkrate
carrying-capacity-parameter = any number, default = I
 Building a model: a case history
Finally, we experimented with several models. One consisted of seven levels (or
stages) and two growers at each level, another combined ten levels and two growers.
The most complicated model thus far contains five different growers per level.

Patterns of connected growth

The ultimate goal of model building is to obtain a good quantitative fit with the data,
based on a theoretically justified model. Our data had the disadvantage of being group
data, instead of the individual time series that a dynamic growth model ideally
requires. Nevertheless, the group data enabled us to extract a number of qualitative
properties which with sufficient plausibility also applied to eventual individual growth
patterns. Those were the properties we hoped our model would simulate.
The first and most important of those properties is the stagewise character of
growth. Fischer's data suggested stages in the form of separate skills that grow and
change in a loosely connected way. A stage is introduced by a massive growth spurt in
all the growers. Although time delays exist between the spurts, there is a clearly
observable distinction between periods in which the spurts occur and periods of
relative stability. On average, the periods of stability become longer as the (cognitive)
complexity of the developing control systems increases. The simulated growth curves
show this particular stage pattern very clearly. The pattern of spurts and plateaux is a
bit fuzzy, though, in that growers within a stage are not strictly synchronized.
Sometimes the spurts occur at about the same time; sometimes there is a marked delay.
The overall image is one of a succession of periods where spurts and changes dominate
and periods where the majority of the growers remain more or less stable. This fuzzy
image of stages and spurts corresponds with the current view on stages as loose
collections of separate skills or growers instead of the tight structures of the Piagetian
universe (see Figure 8.1 1).
A second important property of the data is the stepwise growth pattern of the
individual growers. Each time a new grower emerges at a higher developmental level,
the existing growers receive a boost in the form of a new growth spurt. As the distance
between a new grower and the benefiting grower increases, the support effect
decreases. This property is salient in the simulated growth curves; each individual
curve takes the form of a stair, the steps of which decrease in size as the distance from
the growth onset time increases.
A third characteristic property is the performance dip that occurs after a spurt. In
the model, those dips are caused by within-stage support on the one hand and
between-stage competition on the other.
In summary, the model of connected growers shows the characteristic patterns that
appeared in the empirical data and that are supported by theoretical considerations.
Furthermore, the model allows for a large variety of patterns within this general stage
scheme, dependent on the exact parameter values. Finally, the model shows that
dynamic processes of connected growth occur only within (broad) parameter
constraints. Once such constraints are violated they turn into 'pathological' deviations,
in that they become unstable, show wild fluctuations or simply collapse. One of the
 Building the model

Figure 8.11 A model of two growers and seven different levels displays basic properties of
stage-wise development: stepwise growth in each individual grower, and an alternation of
periods of relative rest (horizontal growth levels) and periods of rapid change.

advantages of this type of model building is that it allows for a form of experimental
theoretical psychology. That is, you can experiment with parameter values and various
dynamic relationships and investigate the effects on the resulting growth patterns.
Such numeric experiments may serve as a heuristic means to achieve a better
understanding of the conditions and forms of quantitative aspects of development.

Notes
1. I realize that the use of the word level can lead to some confusion here. I usually refer to the
growth level of a grower, for instance the grower 'use of single representations'. In Fischer's
theory, however, level means a stage in a tier. For instance, the representational tier has a level
of single representations, that is, a stage where the child works with single representations.
2. Did you notice I left the B column blank in Model 8. l ?You can now use that column to define
a minimal growth level.
3. Unless, of course, that predetermined value is an optimum. I discussed optimum models
'
earlier and presented the example of self-esteem. If self-esteem grows far beyond an optimum
value, its effect may become harmful and negative instead of supportive.
4. Kurt Fischer used the new Lotus-Improv program to build the model. It is based on the usual
spreadsheet logic but works with equations that look more like verbal descriptions than the
symbol strings you would get with Lotus 1-2-3. The actual form of the equation used in his
program was quite similar to the one presented here.
5. The circumflex symbol means 'to the power of'; 2 therefore means 'to the second power'.
A A
The dynamics of mutual
interaction

Development as a social process

The game of growth
When you are appointed a Professor at the University of Groningen you are also
solemnly installed as a member of the University's Senate, which, in addition to being
very old and dignified, is unfortunately a completely useless and obsolete administra-
tive body whose only function is to install members and keep them busy until it is time
to proceed to the next part of the festivities, the new Professor's inaugural Public
Lecture. In times when the University was paid for by the hardworking town and
country folk and a new Professor was appointed, the public had the right to see for
themselves whether or not the new man was worth their money. Knowing the
Groninger people a bit better by now, I can imagine this must have been a serious test
in former days, particularly as far as the financial investment was concerned.
Traditionally, the Public Lecture was a display of the new Professor's learnedness and
erudition. Since it was meant for the public, however, it had to be accessible to and
understandable by what was broadly referred to as the interested layman. Nowadays,
the Public Lecture is a gathering of colleagues, family, friends and children, eagerly
waiting for the moment the lecture is concluded by a solemn phrase, which is also the
sign for a massive run to the place where the reception is held.
I decided that I would take the requirement of accessibility seriously and based my
lecture on the metaphor of a popular board game. I explained to my audience that
developmental processes were hardly predictable and hardly guidable because, in a
sense, they take care of themselves. Development, I said, is like one of those traditional
Dutch board games that combines the luck of dice with insight and strategy. The
course of the game is a metaphor for the developmental process. The players are, for
instance, a child and a parent. Both players try to win; that is, they try to achieve their
personal and temporary goals, which are usually short-term, since neither has a real
long-term overview of the future. -
The parents, being adult and experienced, may think they have set their long-term
plans with their children, but they will soon find that they have to readjust them
 Development as a social process 243
continuously because the game of development is likely to take a course they had not
really foreseen. The metaphor represented development as a process governed by
short-term goals and plans. They are acted out by players who are on the one hand
autonomous and free to choose their own path and on the other hand dependent on the
other player; they need to adjust their actions to what the other player has done.
Despite the short-term structure, though, the long-term effect of the developmental
game is stable in that it almost always leads to a mature, cultivated and educated being,
even if it is impossible to tell how exactly the maturity, cultivation and education will
take shape. The metaphor combined two important aspects, that of chance on the one
hand, and that of co-construction or co-regulation on the other. In the first part of this
chapter, I shall present a short overview of the idea of interaction dynamics between
people. The rest of the chapter is devoted to three models: an elementary mutual
interaction model, a model in which the carrying capacity is managed and, finally, a
dynamic systems interpretation of Vygotsky's model of the Zone of Proximal
Development.

Development as co-construction
If one looks at development in a longer time frame, that of successive generations, and a
bigger spatial frame, that of social structure instead of the individual, development
becomes the process in which culture is transmitted and transformed. The mature
members are bearers of their culture. They know,how to act within their cultural
institutions; they possess skills and knowledge developed over the course of cultural
history, such as reading and writing and scientific knowledge. There is a lot of division
of labour in the way individuals carry their culture to new generations, but as a whole,
culture reproduces and changes itself in the process of generational succession. A
newborn child has virtually no culture at all. I t is an empty vessel, as far as culture is
concerned, that has to be filled. The metaphor of pouring the contents of a full vessel
into an empty one is an obvious choice to describe the process of cultural transmission.
If you want to continue that metaphor, you could say that the empty vessel has got only
a little hole in it through which culture can be poured and so it takes quite a long time,
on average, to fill it.
The vessel metaphor, though probably not aesthetically pleasing, corresponds with
a standard view in sociology, anthropology and psychology of how culture is
transmitted. Valsiner (199 1) called this the 'unidirectional view': cultural information
carries cultural messages one way, from the outside world into individuals. The
unidirectional view is incompatible with the fact that development is a process of
construction by the developing person itself. Basic concepts and skills,are constructed
by the person, given the information and guidance of the environment. A cultural
message, sent in some encoded form or another to a child, is by necessity decoded and
recoded by that child as a function of the codes and understanding that child has
already developed. The recoded message is again decoded and recoded by the mature,
cultivated person who then replies in the form of another message. Valsiner (1991) calls
this a bidirectional process of cultural transmission.
Valsiner (1991), and several others, such as Barbara Rogoff and Alan Fogel, use the
244 The dynamics of mutual interaction
term 'co-construction' to refer to a process that has both the aspect of construction and
the aspect of communality. The constructing environment and the constructing child
act as a higher order unit.

A floating cat (or: the dynamics of co-regulation)

A couple of summers ago I drove from the Grand Canyon to Salt Lake City, Utah, to
pay a visit to the Fogel family and to discuss the peculiarities of dynamic systems
models with Alan Fogel. We spent a considerable amount of time in the Fogel
swimming pool, in a company consisting of our wives, Brian Hopkins - whom we had
accidentally met in the weirdest desert places and whose interest lies in early motor
development - and finally the Fogel family cat who used to join the swimming party
floating on an airbed. How can you get a cat on an airbed, floating on the lukewarm
water of a swimming pool? The answer is by a process that Alan Fogel calls
co-regulation and that is central to development. We had no definite plans to get the cat
on the inflatable mattress at first, but the animal showed some interest for the people
swimming. I guess Alan slowly pushed the airbed towards the rim of the pool, and out
of typical feline curiosity, the cat approached. Alan pushed the bed a little nearer,
meanwhile addressing the animal in the sort of voice cats find irresistible. And since the
airbed's surface was dry and warm and formed a sort of extension of the ground, the cat
jumped, and found itself soon in the middle of the pool. In his recent book Developing
Through Relationships (Fogel, 1993), Alan Fogel defines co-regulation as a process that
'. . . occurs whenever individuals' joint actions blend together to achieve a unique and
mutually created set of social actions' (p. 6). Alan sees it as a basic process in the
creation of communication, self and culture.
Children are not taught or instructed in the form of unidirectional input processes.
The instructional activities take place within consensual frames, that is, within
co-regulated consensual agreements about the scope and topic of the communicative
process they are engaged in. Co-regulation leads to new patterns of action and
communication via self-organization; there is no causal primacy, as Fogel calls it,
neither in the parent, nor in the child. The self-organizing follows the principles of
dynamic systems, which converge onto attractor states. There is no preliminary plan
for the attractor state. The state just emerges from the way the components interact. I
have given many examples of how different attractor states emerge from similar
mechanisms, provided some of the control parameters have different critical values.
The same holds for the emergence of developmental patterns.
Alan Fogel's own research focuses on the way early parent-infant co-regulated
action plays a role in the emergence of 'elementary' and allegedly culture-free
acquisitions such as grasping or posture control. By studying videotaped interactions
in great detail, we can show how patterns such as holding a baby in our arms, or feeding
a baby, come about by the same process that later on creates a social self and higher
cognitive skills (Fogel, 1992; 1993).
Although co-regulation is a prime example of the working of a dynamic system, the
system of mother-infant cultural environment, for instance, there is at present no way
 Development as a social process 245
in which we can model the emergence of the qualitative patterns of communication,
action and skills (Fogel, 1993). The dynamic systems models I presented in this book
describe only macroscopic, quantitatve aspects of development; they cannot explain
the emergence of a pattern of grasping, for instance, or a child's understanding of
object properties. What I intend to do, however, is to show that even those macroscopic
quantitative aspects of the co-regulation process are worth studying. They can reveal at
least some of the general properties of those processes, for instance, how and why they
converge towards their attractor states.

A simple dynamic model of co-construction and co-regulation

Co-construction and co-regulation are good examples of iterative dynamic processes.
T o illustrate the nature of the dynamics, I shall take a behaviour protocol from one of
Piaget's famous investigations of formal thinking.
In one of his experiments, Piaget gave his subjects four similar flasks containing
colourless and odourless liquids. There was no way to tell the difference perceptually,
so each of the containers bore a number, 1 to 4. He then added a bottle with a different
chemical substance, g, and a dropper. The subjects were shown that adding a drop of
that substance to a particular combination from the four flasks produced a yellow
colour. They did not know the secret combination that produced the colour, and had to
find out for themselves, by making all possible combinations of the four liquids. One of
the protocols, with a 7-year-old subject, shows the following interaction:
+ +
R (the child) tries 4 g, then 2 g and 3 g +
R: I think I did everything, I tried them all:
E (experimenter): What else could you have done?
R: I don't know.
E gives him the glasses again.
R repeats the sequence of adding drops.
E: You took each bottle separately, what else could you have done?
R: Take two bottles at the same time.
+ + + +
R tries 1 4 g, 2 3 g, thus failing to cross over between the two sets of
+ + + +
bottles, for example 1 2 g, 1 3 g, etc.
+
When E suggests that he add others R puts 1 g in the glass already containing
+
2 3 which results in the appearance of the color.
E: Try to make the color again.
+ +
R: Do I put in two or three? R tries 2 4 g, then adds 3, then tries it with
1+4+2+g
R: No, I don't remember any more. (Inhelder and Piaget, 1958, p. 111)
The experimenter could have told the subject how to make a full combinatorial
+ + + +
diagram given four elements, 1, 2, 3, 4, 1 2, 1 3, 1 4, 2 3, 2 4, 3 4, + +
1+2+3,1+2+4,1+3+4,2+3+4,1 +2+3,1+2J;3+4.Thatwould
not have made much difference, since the 7-year-old would not have understood the
logic behind that series anyway. The only way to tell him how to do it right would
probably invoke a sort of learning and teaching sequence very similar to the one in the
 T h e dynamics of mutual interaction

Figure 9.1 A communicative process takes the form of vertically and horizontally
connected events; the vertical axis represents time, the horizontal axis the relationship
between the contributors (e.g. a pupil and a teacher).

example. Assume that every action or communication by one of the interacting

partners is a single event. We number them, beginning with R's first attempt and
explanation and divide them over a column for R's actions and statements and one for
E's. How would each of the events relate to one another? Each event is directly related
to two others. The first is its own predecessor. For instance, if E suggests R to add other
glasses, that suggestion takes the earlier suggestion into account. The new suggestion,
for instance, is a little more explicit, or it extends the possibilities suggested in the
previous suggestion. The second event to which it is related is the preceding action
from the child. The child makes a highly specific combination of liquids and the
experimenter tries to make clear that more combinations should be tried. If I represent
a conditional link between two events by an arrow and an event by a block, the
structure of the interaction could take the form represented in Figure 9.1.
It probably does not require too much imagination to view the entire communicative
event as a couple of connected iterative processes. The iterator at the left-hand side is
the child; the iterator at the right-hand side is the adult. In the dynamic models I
described earlier, we knew exactly what the nature of the iterator was: a growth
equation of some particular type. In this case we do not know exactly what the child's
or adult's mind is doing, but we do know that they take their previous output and the
output of the other party as their input and produce an output, which can then be taken
as input again and so forth for as long as their communication lasts.
 T h e unexpected complexity of a simple mutualist dynamics 247
Formally, a model of co-construction or co-regulation processes does not differ from
the models of connected growers. The challenge, however, is to find a form for the
iterator, that is, the equation that specifies exactly what each of the partners is doing in
the communicative process. When I say exactly, however, I do not mean completely.
We have not the faintest idea of all the thoughts, processes and activities that go on in
the child's and adult's minds when they are involved in a communicative process like
the one described above. What we should do, therefore, is to isolate one relevant factor
from the stream of internal events, then try to specify how that factor behaves and
changes over the successive events.
Since we are primarily interested in the process of developmental growth and
learning, we should take the growth level of a skill or cognitive principle in the child as
our relevant factor. For instance, in the example from Piaget's research protocols, we
could simply pretend that each action of the child represents a level of understanding of
the combinatorial principle. Over the short span of this experimental session, we
should not expect too much change in that level to occur, especially since the child still
has little knowledge of the relevant principle. But if each event is a teaching session,
spread out over a time-scale of, say, a few years, the relevant variable - understanding
of the combinatorial principle by the child - might show a significant increase and
eventual transient drops. The next question is: What is the relevant factor in the other
partner, the adult? Is it the level of formal knowledge of combinatorial principles, since
we may expect the adult to understand them fully. In the coming sections, I shall
discuss a few different ways of conceptualizing what it is that changes dynamically in
the adult partner, provided it is a level of skill or knowledge that changes in the child
partner.

The unexpected complexity of a simple

mutualist dynamics
A short sketch of an utterly simple and too elementary theory of
learning
When you teach, you have a goal in mind that you want a child to reach. That goal is
some level of understanding of mastery of a skill. The child has not yet achieved that
goal and would probably never achieve it on its own. Teaching amounts to confronting
a child with something that is closer to the intended goal than the child's actual level of
understanding or mastery. Learning is a way of profiting from that confrontation by
decreasing the gap between your current level of mastery and understanding and the
goal state. (So far this looks more like an entry for 'Education' in a Martian
encyclopedia than a serious explanation of what learning and teaching is, but in the title
I promised that what followed would be utterly simple and elementary. Allow me to go
on with our utterly simple and too elementary theory of learning.)
It is completely unclear what teaching and learning entail if conceived at the level of
 T h e dynamics of mutual interaction
specific activities and processes. There seem to be so many forms of teaching and
learning that I would not dare to present a generalized account. But whatever they are
in particular, in general they amount to a sequence of more or less separate events. The
teacher does something, the child does something, the teacher again, the child, etc.
Sometimes these events overlap, but that does not seem to be essential to the process.
Let me assume that a child is learning about some property of the world - where
babies come from, the meaning of the square root of a number, the different ways of
combining four liquids - and a teacher is teaching. The teacher, of course, knows
where babies come from and what a square root is. This is the knowledge or the
understanding he/she wants to teach the child. Let me represent this goal state of
knowledge by the letter K. The child's current knowledge level of that property is the
state L. Teaching takes place in the form of encounters where the teacher presents to
the child something that relates to K. The encounter may consist of a story, a task, an
explanation and so forth. With each discrete cognitive encounter with the intended
property K, the child has some experience of K. It is an experience presenting the
property in a form close but not similar to the child's current level of understanding of
that property. Let us call that experience E.
What determines the nature of E? It is, first, determined by the teacher's choice of
how to present K to the child. In the olden days people told their children that babies
were brought by the stork and they only gradually worked their way towards revealing
the real process. Nowadays parents tell their children that babies grow in their
mother's belly. Second, E is determined by the child's own understanding. Even when
the parents tell the mummy and belly story, the child makes his or her own
interpretation that has probably only very little to do with the biological facts that the
adult intended to convey.
The goal K and and current learning state L differ from one another in many
respects. In Chapter 1, I introduced the notion of a descriptive space that would
contain as many descriptive dimensions as needed to specify the distinction between an
initial and a final developmental state. The initial and the final developmental state are
two distinct points in that descriptive state. It is trivial, then, that there is some distance
line between L and K in that space. Since we do not know exactly which descriptive
dimensions are used in the space and what their properties are, we represent this
distance in some general abstract form. For instance, we set the initial level for L at
some arbitrarily small value and give K the value of 1.
I claimed that the experience E is a sort of local temporary representation of K in a
form that the child can grasp. It is based on the child's own understanding and the
teacher's estimation of what he/she should do to bring the child closer to the goal K. E
has all sorts of varying properties, but it has one formal property we should be
interested in, its distance to L and K, respectively. E can be closer to K than L, further
from K, or at the same distance from K as L. I assume that learning proceeds as a
function of the learning and teaching experiences (at least in this simple model, and
probably also in a considerable part of reality). Since learning, in this simple theory,
amounts to reducing the distance between L and K, preferably to a zero distance at the
 The unexpected complexity of a simple mutualist dynamics 249
end, E must have the function of dragging L somewhat closer to K. It is E, not K, that
functions as an attractor point for L, since K is represented in the form of E. Therefore
the position of E is crucial.
The verbal description of the model allows us to infer the following set of equations
that will describe the changes in the respective distances between L, K and E. The
model prescribes that E should be at a small distance from L and ahead of L:

for d' a parameter that determines how far E is from L. For example, if the experience is
a helping instruction from a teacher, and d' is very small, then the help is very close to
what the child already knows. If d' is large, then the help is way ahead of the child's
current understanding.
The model further claims that as a result of the occurrence of E, L will make a jump
towards E. Let us say that this jump is some fraction c of the distance between E and L
(if c = 1, L bridges the distance between L and E in one jump):
jump (L) = (E - L) c (9.2)
Recall that d' is a parameter that causes E to move. Since K is the goal state, E should
gradually move towards K and if it arrives at K, it should stay there; this is just an
alternative way of saying that teaching is a goal-directed activity. In determining the
level of the child's next learning experience, the teacher wants to shorten the gap
between the previous learning experience and the goal state:
(teacher's increase of E) = ( K - E) .d (9.3)
The actual learning experience, however, is a function both of what the teacher does
and how the child understands this, given the current knowledge level. The jump in
experience level is therefore a function of the teacher's choice and the child's level:
jump (E) = L - ( K - E) - d (9.4)
Since the experiences and learning encounters are discrete events, we may write down
the relation between L and E in the form of a set of coupled difference equations over
+
time intervals t, t -t 1, t 2 and so forth. The first is the equation specifying the
progress that L will make, given a teaching experience E:
= L, + jump (L,) = L;(1 + (E, - L,).c)
Now we need a second equation that specifies the level of the next learning experience.
Equation 9.1 says it is the sum of the learning level and an increase variable. This
increase has been specified in equation 9.4, so that our equation describing the level of
the next learning experience is a combination .of equations 9.1 and 9.4:

Model 9.1 describes the spreadsheet format of this simple model, whose far from
simple outcomes we will investigate in the next section.
250 The dynamics of mutual interaction
Model 9.1 A simple learning and teaching model
with complex patterns
The model combines two versions of the learning and teaching model. One is the
standard version, from equations 9.5 and 9.6, the other is an alternative model based on
the idea of a random walk. It will be explained later. Define the following set of range
names and values:

value parameter
9 K
200 in-divisor
0.2145 C
0.18 d
0 rand-L
0 rand-K
0 rand-E

The most important parameters are cand d. The parameters Kand In-divisor are used to
set the initial state level as a ratio of the goal state.
The A column, starting with cell A10 will contain the values of the L variable, B will
contain the values of the Evariable, and column C the value of the goal state K, which will
vary only if the value of the parameter Rand-K differs from zero. In cell A10, type:

and in cell B10, type ' + A10'. It is not necessary for the initial E t o differ from the initial L.
Cell C10 contains the initial value of K, $ K (which remains constant under the
non-randomization condition).
Write the equation for L in cell A l l . Its first part is the basic equation as explained
earlier, the random part is needed if you want to study the effect of a random walk on the
growth process:

+AlO* (1 + (B10 - AlO) *$c) + $rand-L* (@rand - @rand) * (C10 - Al0) (9.8)

Cell B11 has the equation for the E variable, namely:

Note that the random factors take effect only if the value of rand-E is not zero. Finally, cell
C11 describes an eventual random walk for the value of K, which occurs only if rand-K is
not zero:

Copy the range A1 1..C11 to A12..c500. Define four graphs. One, Timeseries-1 is a line
graph of columns A10..A500 and BIO..B500. Timeseries2 plots only the first 100 steps,
namely AlO..A110 and BlO..B110. The third graph, Phaseplot, is an xy graph. Take the
column B100..B500 as its x-axis, and A100:.A500 as the y-axis. The first 100 steps are
omitted from the phaseplot. The fourth graph can be called K-L-plot. It takes A1 O..A500 as
 The unexpected complexity of a simple mutualist dyna.mics 251

Figure 9.2 Three different curves resulting from low c- and d-values and different initial
state values; if the initial state value is very low (in comparison with K), the resulting
growth is a sudden jump (note that the time scales are very different for the curves at the
left on the one hand, and the curve at the right on the other; the sudden jump requires a
much longer time span).

its x-axis and C10..C500 as y-axis and is defined as an xy graph. It is useful for showing the
endpoint of the process in terms of both K a n d L values in the case of the random walk
version of the model. You can experiment with values described further in the text.

The complex properties of a mutualist dynamics

If the ratio between K (arbitrarily set at value 9) and the initial level Liis about
one-tenth of K and the growth parameters c and d are low (0.033 for both), a curve
results that has the classical ogive or S-form (see Figure 9.2, top left-hand side) that we
also found with the standard logistic equation.
With an initial L about 11200 of the K level the S-curve takes a different slope. We
now find a long platform which stays close to the initial level, then a very sharp increase
and a rather abrupt levelling-off at the K level (see Figu,re 9.2, bottom left-hand side).
This is the sort of curve found with the transition version of the logistic equation, the
cubic equation.
Finally, with an initial learning level about one-half of the K level, we find a curve
that resembles classical learning curves. I t is important to note that considerable
random perturbations imposed on c and d do not change the overall qualitative pattern
(e.g. for random variations of 0 < c < 2c and 0 < d < 2 4 .
The fact that this model produces curve; qualitatively similar to the logistic and
transition models can be interpreted in different ways. You could argue that, since
 The dynamics of mutual interaction

Figure 9.3 With increasing parameter values, the mutualist model evolves from oscillatory
to chaotic fluctuations.

those curves are produced by simpler models, the present one should be considered
superfluous. On the other hand, you could defend the position that the present model
has been based on an elementary theory of teaching and learning and as such forms the
dynamic expression of that theory. The finding that an entirely different type of model
based on learning and teaching yields the same type of growth curves could be seen as a
further justification for the claim that such growth forms are universal and robust, in
that they are produced by dynamic models based on different principles.
Given that the mutualist model produces curves similar to the logistic model, we
may expect that it should resemble that model also in its oscillatory and chaotic
patterns. Experimenting with increasing parameter values for c and d shows the
expected resemblance. For instance, with an arbitrarily determined fixed d value of
0.103, increasing c values produce qualitatively different evolutions (see Figure 9.3a).
For c = 0.16, a mismatch between E and L occurs around K, but it is easily
compensated, and E and L quickly settle down at the attractor point K. For c = 0.23,
the oscillation towards the attractor is. more complicated (Figure 9.3b). For
convenience, only the iterated points of L have been drawn. They show a pattern of
 The unexpected complexity of a simple mutualist dynamics 253

'
Figure 9.4 State space plot for a chaotic oscillation: the vertical dimension represents the
level of experiences, the horizontal dimension the growth level of the knowledge variable
whose growth is dependent on those experiences; note that the points are not connected by
a time line (a time line connects a point to its successor in time); strangely enough, the
state-space representation is highly reminiscent of a n important figure in recent French
history.

waves of oscillations towards K. For c = 0.27, E and L no longer evolve towards K but
settle into what seems to be an infinite mutual chase without L and E ever catching each
other in an equilibrium state (Figure 9.3~).I n psychological developmental terms, this
means that for increasing values of c and d a mismatch between learning and experience
may emerge resulting in the impossibility of reaching the goal state. Phenomenally this
may take the form of oscillations between under- and overgeneralizations of a learned
principle or rule, a perpetuating conflict between a parent and an adolescent about
which norms to follow and how, the impossibility of finding a solution for a problem
and so forth. I t is likely, however, that if this were to happen in real life, attempts would
be made to relax whatever psychological properties correspond with the c and d
parameters.
For still higher values of c and d, the model settles into the infinite oscillation already
seen in the previous example. Figure 9.4 represents a state space plot foi- such chaotic
oscillation in the plane consisting of the values of E and L. It is just one of the many
possible plots, showing that the chaotic oscillations are not rand& but have their own
intrinsic order.
Earlier I described a technique for showing the number and magnitude of the
attractor points of a process for different parameter values. This bifurcation diagram is
254 The dynamics of mutual interaction

Figure 9.5 A bifurcation diagram for a range of d-values shows the non-linearity of chaos
in the model: the horizontal axis shows the range of increasing d-parameter values; the
vertical axis specifies the sets of final growth levels that correspond with the d-parameter
values; sets of d-values corresponding with heavy, chaotic oscillation alternate with highly
regular patterns in a complex, unpredictable fashion.

an alternative way of showing the order underlying the chaos of time series. It is
particularly suited for showing the non-linearity of the model; increasing parameter
values correspond with increasing chaos, punctuated by highly orderly intermediate
states. Figure 9.5 shows a bifurcation diagram for a range of parameter values of d,
between 0.174 and 0.193, with a constant value of c, 0.18. Black bands represent chaotic
oscillations occurring with the corresponding d values. The chaotic regions, however,
jump suddenly towards simplicity, for instance, towards a two-state oscillation. I n
turn, the simple states gradually evolve towards chaos again.
The simple model described in this section is but one of a large family. A description
of the alternative versions and their equations can be found elsewhere (van Geert,
1992a).

Random walks in a mutualist dynamics

The model discussed in the previous section incorporates an extremely simplified but,
I hope, basically correct, vision on the nature of a co-construction or co-regulation
process. I t shows how a learning state in a learner evolves towards a goal state via
experiences that are functions both of the goal state and the learning state. If the crucial
parameters c and d stay under a threshold value, the process leads to a point attractor,
K. If not, chaos sets in. In the 'normal' case, the learning and teaching process reaches
its goal state. This somewhat boring conclusion is misleading, though, in that the
 Social interaction and the management of the carrying capacity 255
model shows only the distance between goal and learning state. It is a distance defined
over an arbitrarily complex space of variables and properties. In that space, the goal
state and the learning state may shift randomly (or systematically) during the process of
learning and teaching, for instance, because both learner and teacher redefine their
goals. This wandering process of goal and learning state is not shown in the model; it
says only that the distance between the two is bridged after a while, but it does not
specify to which points in the space the distance measure applies.
In order to introduce the notion of continuous goal and learning state redefinition in
the simple, one-dimensional growth model of the mutualist dynamics, we can add a
random walk aspect to it. T h e assumption behind this is that each step in the process of
the L, K and E variable is the result of a systematic factor on the one hand and a
coincidental factor on the other. The systematic factor is the dynamic model of L and E
adaptation described in the previous section. T h e coincidental factor is simply a
random number added to or subtracted from the number computed on the basis of the
dynamics. This random number varies between an arbitrarily chosen negative lower
and a positive upper boundary. I assume that it should be a function of the distance
between the learning and the goal state; the further away the goal state from the
learning state, the easier it is to take a deviating road, for the learner as well as for the
teacher. This is, of course, just an assumption, but it is based on the idea that a long
distance between goal and learning state corresponds with less information for both the
'
learner and the teacher about what the goal state and learning state entail. The equation
for the goal state K is as follows:
,
Kt + = Kt + rand. (Kt - L,)
for - b < rand < + b
With values for the rand parameters of 0.1, the process shows an interesting variety of
final state points, that is, points where both the wandering K a n d the wandering L lock
each other up into some attractor region. T h e stable end states line up in a somewhat
skewed unimodal distribution (Figure 9.6).

Social interaction and the management of the

, carrying capacity
Carrying capacity as the educational model
The interactions between more experienced and mature members of a culture and the
less experienced, less mature ones can take any of a vast variety of forms. They range
from strictly organized, formal teaching and-learning practices to highly informal
imitative processes where neither the imitator nor the imitated person have any
conscious intention to either teach or learn something. In th; previous section, I
discussed a simple model explaining how educational goal states are achieved in a
process of mutual adaptation.
But how does this variety of cultural transmission and appropriation processes relate
 The dynamics of mutual interaction

Figure 9.6 The distribution of final states after 200 runs of the random walk mutualist
dynamics: the pattern resembles the 'normal' bell shape of the Gaussian distribution of
random events (the picture in the background represents the skyline of the city where this
simulation was run).

to the logistic growth model? It contains only two parameters, the growth rate and the
carrying capacity. The growth rate is more or less confined to the learning person itself
and represents an abstract property relating to the person's speed of learning,
changing, inferencing, etc. The carrying capacity has been defined as the asymptote of
the growth process, or more precisely, as its equilibrium level. A very important step in
the construction of the logistic growth model was the assumption that this equilibrium
level forms the expression in terms of the growth level of all the resources available to
the grower. These resources were of various kinds: spatial, informational, energetic,
temporal, etc. They could come from the subject itself, for instance the subject's
knowledge or skills that support the learning or achieving of some new knowledge or
skill, or from the environment, for instance, the information and help given to a child in
the form of instruction and guidance by a more experienced person.
Whatever we call education, teaching or guidance actually amounts to monitoring
part of the environmental component of the carrying capacity. For instance, when
children are taught addition at school, the curriculum is carefully sequentialized. The
teacher begins with elementary concepts and very simple forms of addition. Gradually,
the tasks become more complex and the requirements increase, following the
increasing competence of the children. It is almost trivial that, if the the teacher did not
adapt the level of teaching to the increasing learning level and remained at the low
initial level, hardly any learning would take place. If you never taught more than just
elementary addition of numbers below ten, most of the children would never get any
further than that. That is, their addition gr-owthlevel would reach an asymptote at the
level of skill embodied in this very rudimentary teaching. As a general rule, we may
 Social interaction and the management of the carrying capacity 257
state that the maximal level of learning a child may achieve is proportional to the level
embodied in the teaching.
For the sake of simplicity, we set the individual's contribution to the carrying
capacity of one of its growers, such as addition, to a constant. The level of that constant
will greatly differ between individuals, but that need not concern us here. The environ-
mental part of the carrying capacity will fluctuate and will do so in function of the
information, tasks, demands and guidance given by a teacher, a parent or any other
mature member of the culture. It is an almost trivial observation that, on average, the
level of information, demands, or task complexity increases during an educational or
teaching process, or, to put it differently, this is at least the intended course of events.
Since that information or demand level is a measure for the carrying capacity at each
moment, it follows that an educational process basically amounts to increasing the
carrying capacity level. The increase will level-off and come to a standstill as soon as
some goal state has been reached. I t also stops when the person for whom the educa-
tional efforts are meant is no longer making any progress (again, this is the ideal case).

The basic model of carrying capacity increase

We can now write an equation that describes the growth of the carrying capacity level
as a function of the increase of mastery in the learner. Let Lt be the learner's level, for
instance of addition and Kt the carrying capacity level embodied in the teaching.
Assume further that the goal state level of the teaching is set to 1; in practice, the
teaching goal could be reaching the end of the handbook, or successfully finishing a
curriculum. The equation for the growth of the carrying capacity is:

(the meaning of the parameter d will be explained later).

Since the growth of the learning level is a function of the carrying capacity, the
equation for that growth is:

In this model, K grows as a function of the absolute increase in the learner's level. We
may expect this when the learner's level is measured with some sort of objective scale.
An example is addition and other arithmetic operations for that matter, which are
described in handbooks and explicitly formulated curricula. But if you deal with less
formal skills or forms of knowledge, there is hardly ,any absolute yardstick. In that
particular case, the instructor has probably an idea of the relative increase. A
mathematically simple way to describe relative increase in the equation for the growth
of K is as follows:

What is the psychological meaning of the parameter d in equation 9.12?

Mathematically, it is a parameter that determines to what extent the increase of the
learner's level contributes to the increase in the teaching level that embodies K. Some
 The dynamics of mutual interaction
instructors will tend to increase the demands much more than others, given they react
to the same increase in a learner's capacity. Some will react to any increase, others only
every now and then. The extensive literature on the developmental effects of
interactions between parents and babies has produced a concept that applies to this
issue, sensitivity.
Sensitivity is a property of a parent that relates to the extent to which the parent's
reactions to the baby's activities are developmentally adequate, that is, contribute to
the child's growth. The concept entails two aspects, namely the adequacy of the
reaction and its contingency, the extent to which the reaction is direct and consequent.
Apply this, for instance, to teaching arithmetic to a single child. A sensitive instructor
will provide an adequate increase of information and demands, in response to the
child's learning progress. The instructor will also be consequent, in that any progress
in the learner is followed by some adaptation of the teaching level. The parameter d is
assumed to specify the instructor's sensitivity, in the broadest sense of that word. In
the following spreadsheet model, I shall divide that parameter into a component that
corresponds with an ordinary parameter and a component that specifies a probability.

Model 9.2 Increasing the carrying capacity in

response to growth
In this model you will have to define the following list of parameters:

value parameter
0.1 Kin
0.01 C in
0.7 K-d
0.2 L R
0.02 d-start
1.1 d-stop
0.02 step
1 Krand
1 rel-model

K i n is the initial state of the changing carrying capacity K. L i n the initial state of the
learning level. K-d is the dparameter in equations9.12 and 9.14 that modifies the effect of
the increase of L on K. L-R is the growth rate of the learning level L. The following three
parameters, d-start, d s t o p and step are needed in a macro that will simulate an
arbitrarily chosen number of runs of the model. K r a n d is the parameter that determines
the probability that Kwill actually react to an increase in L. If K r a n d is I , that probability is
100 per cent, if K r a n d is, for instance, 0.7, the probability is 70 per cent. The last
parameter, rel-model, lets you switch between a model based on equation 9.12 and one
based on equation 9.14 (the relative increase model). If rel-model is 1, the chosen model is
that from equation 9.14.
In cell A10 write K i n ; in cell C10 write L-in. They are the cells containing the initial
 Social interaction and the management of the carrying capacity 259
values for Kand L. Cell B10 contains a random parameter determining whether or not K
will actually change in a cell. Write the following equation in cell B10 and copy it to B11:

It says that if a random number, which is different for each cell, is smaller than the value
chosen for the parameter K-rand, the value of the corresponding B cell is 1, but 0 in any
other case. The function of this cell becomes clear from the formula that goes into cell A1 1:

If B10 is zero, there is no growth in K, that is, the value of A1 I is the same as that of A10.
The function of the @if statement in the equation is to turn the equation 9.14 model on
and off. If the value of the parameter rel-model is 1, (C10 - C9) is divided by C10. That is,
the growth of K is affected by the relative increase in the learning level, found in the C
column. If the parameter is 0, (C10 - C9) is divided by 1.
The C column contains the equations for the growth level. It is a straightforward logistic
equation. Given that you have put L-in in cell C10, cell C11 takes the following equation:

Copy the range A1 I..C11 to A12..C400.

Remember that the A column con'tains the values of the carrying capacity. Try some
values of K-d and observe the effect on the final state level.
Now it is time to write two simple macros that will allow you to vary the value of K-d
systematically and study its effect on the final state value. Go to cell L1 (or any other cell
that you prefer) and write the following:

{for Kd,d-start,d-stop,step,test} (9.18)

This macro will specify a value for the cell you named K-d. The first value will be the value
you have specified in the cell named d-start. It will then perform the action described by
the macro 'test', which you will define in a minute. Then it will add the value specified in the
cell 'step' to K d , and again perform 'test'. It will continue to do so until the value that goes
in cell K d is equal to the value you defined in d-stop. Thus d-start is the minimal value of
K d and d-stop is the maximal value of K d . The number of times the macro 'test' will run
is equal to (d-stop - d-start)/(step). Before you go on, you must specify a name for the
macro under cell L1. This can be done by defining a range name, a procedure you should
now be familiar with. Define the name 'Go' and attach it to the cell L1.
Since spreadsheets differ in their menu structures, it is better to compose the following
macro by executing and recording the commands. Recording can be done via a
Transcript option contained in the spreadsheet menu: Turn it on, and perform the
following actions: Go to K8 (via the GoTo command; F5 in Lotus); press End; press Down
twice; copy the Range Value of the cell name K-R-damp to the current cell'(the first empty
cell under K8); go to J8; press End; press Down twice; calculate (F9 in Lotus and Quattro
Pro); copy the Range Value of cell C400 to the current cell (the firskempty cell under J8).
When you have performed all these actions correctly, you will find a transcript of them in
the Transcript window. Copy the description and paste it into cell L10. Define the range
name 'test' for cell L10.
In cell K9 type 'Parameter Value', in cell J9 'Final State'. Type an 'x' in cells K8 and J8. For
260 T h e dynamics of mutual interaction
each run, the macro will type the corresponding parameter value of K d in column K and
the final state value associated with that value in column J. It will add new values at the
end of the column of values. In order to run the macro, hit Tools/Macro/Run from the main
menu.
You can graph the results of the repeated runs by assigning the column of parameter
values (column K) to the x-axis and the column of final state values to the y-axis.
You can also assess the effect of randomly distributed Kgrowth. Recall that if K-rand is
equal to 1, Kresponds to the increasing growth of L in every cell. That is, the growth of Kis
completely contingent on the growth of L. If C r a n d is smaller than L, say 0.5, Kremains
constant in every cell for which the value of a random number is smaller than K-rand. In
order to study the effect of a random increase pattern, you need to define a macro that
repeats the calculation of K and L with different columns of random numbers. Go to cell
L40 (provided you chose column L as your macro column) and define cell L40 with the
range name 'loop'. The macro Loop consists of the following actions which can be
recorded and pasted into cell L40via theTranscript command: Go to K8; End; Down twice;
copy the range value of cell K-d to the current cell (the first empty cell under K8); go to J8;
End; Down twice; calculate (F9); copy the range value of cell C400 to the current cell (the
first empty cell under J8). Copy the set of instructions from the Transcript window and
paste it into L40. At the end of this set of copied instructions add the string:

The macro will add the value of K d to the K column, then calculate the value of the
resulting end state and copy that to the end of the L column. It will then repeat itself and
continue to do so until you hit Ctrl-Break to stop the loop.

The qualitative behaviour of a K grower

If you run the program from Model 9.2 with a few different parameter values of K d ,
the K adaptation parameter, you will have noticed a characteristic property of the
model. The carrying capacity as well as the dependent growth level evolves towards a
stable level that can be anywhere between the initial state level of K and the maximal
level of K, which is 1. The smaller L d , the closer the equilibrium level is to the initial
level of K, that is, the less growth in L. This is very different from the effect of a low
growth rate on the growth level. With a low growth rate, it takes a longer time to reach
the equilibrium level, but the asymptote of that level is always the maximal value of K.
With a low adaptation rate of K, the system settles into an equilibrium state that can be
much lower than the maximum. Figure 9.7 shows the distribution of final state values
for a L d parameter ranging between 0.02 and 1.42. The increase in the values is an
S-shaped function of the linear increase in I C d . This means, among other things, that
any increase in K d is more effective in the middle range of K d values than in the
extremes.
A second characteristic property of the present dynamics is what happens if the
random parameter affects the growth of K. With b a n d set to 0.5, the probability for
K to change in reaction to a change in L is 50 per cent for each step of the computation
(that is, for each cell). With K_d = 0.7 a n d - h a n d = 0.5, I obtained a distribution of
end states that differs rather considerably from a normal bell-shaped distribution
 Social interaction and the management of the carrying capacity 261

Figure 9.7 The final state values differ with different K d values: the final states
corresponding with the set of linearly increasing K-adaptation parameters follow an
S-shaped pattern (note the difference with the ordinary logistic S-curve, which is a pattern
of levels over time).

Figure 9.8 The distribution of 4,500 endstates with L d = 0.7 and k a n d = 0.5; the
horizontal axis represents the level of the final state of each individual case (growth
process) ranging from 0.1 to 1; the vertical axis specifies the relative number of cases with
final states grouping around similar averages; there is a characteristic maximum in the
number of final states which are around the value 0.25, and another .concentration of values
around 0.6; only very few runs reach the maximal level (value 1).
262 T h e dynamics of mutual interaction
(Figure 9.8). T h e figure looks more like a bimodal distribution with one characteristic
peak around the value 0.25. With different versions of this model (two are described in
van Geert, 1991), I obtained sharper bimodality. What explains this particular
distribution of end states? In a model like this one, the earlier steps have a greater
impact on determining the equilibrium level than the later ones. This is an unintended
consequence of the way carrying capacity and growth level influence each other over
the course of the process. If you have the bad luck that your first five to six steps show
little or no growth because the random numbers just turned out that way, you will have
a considerably greater chance to end up in the low end of the distribution than if the
first steps drew better tickets from the lottery. This effect is reminiscent of a folk
wisdom about development, that your earliest experiences determine much of what
you will become in later life. I make no claims about the potential truth or falsehood of
that belief. What matters here is that there is at least one formal growth model, the one
presented in this section, that shows this kind of early stage effect. Wherever the model
applies in the real world, we may expect that earlier experiences - whatever their
nature may be - will have a bigger impact on development than later ones.
In summary, what this simple model shows is that if growth and K adaptation are
linked, as is most probably the case in a variety of educational practices, the speed with
which the instructor adapts to growth in the learner determines the final state level.
This level can be well below the level achievable, given a better adaptation process. In
the current model, better adaptation simply means a higher K adaptation parameter.
In reality, however, we may expect a more subtle relationship between the learner's
level and the level embodied in the teaching. More precisely, if the teaching level
deviates from an optimal distance with the learning level, either above or below that
optimum, the learner will profit less from the teaching. In the next section I shall
discuss a classic developmental model, Vygotsky's, in which instruction plays a major
role. We shall see that the notion of an optimum distance is of crucial importance.

Vygotsky's dynamic systems

From potential to actual development via help
The model that I shall discuss in this section is the famous proximal development model,
which tries to answer how learning emerges out of the activity of an instructor and a
learner who cooperates in a particular fashion. T h e Zone of Proximal Development is
defined as the distance between the actual development and the developmental level
that can be reached from the latter, given adequate help by more competent adults or
peers. It expresses several thoughts at once: first, that the actual developmental level is
an unstable state, continuously changing and drifting towards a state yet to be
actualized; second, that the change is a function of two parameters, namely the child's
own capacity and the help given by others. But why does the help provided by other
more competent people turn into a competence of the subject helped? The answer lies
in a major mechanism, interiorization: activities carried out at the social intersubjective
 Vygotsky's dynamic systems

Figure 9.9 The growth curve of a developing competence consists of a sequence of levels
of actual development (Zoad); in this figure, the sequence is a highly idealized linear
progression of Zoad-levels.

level will gradually move onto the intersubjective persdnal level and become part of the
person's own competencies.
But the definition of the Zone of Proximal Development immediately poses a
fundamental problem. How much help is the help that gets internalized? It is not
difficult for an experienced painter to 'help' a blind child paint a realistic and colourful
landscape, but this help will of course not be internalized. Let me try to get a better
grasp on this rather elusive concept of 'help' by making a number of simplifications.
I shall start with an arbitrary developing skill, such as the ability to solve addition
problems. The child's actual skill level, i.e. the level at time t, is his/her Zone of Actual
Development (Zoad). Let me pretend I also know the Zoad at some later time t 1 +
(see Figure 9.9). According to the definition of the Zone of Proximal Development
+
(Zoped), we know that this later Zoad at time t 1 is the Zoped at time t (actually, we
+
have defined the time difference between t and t 1 in such a way that it allows for a
given Zoped to become a Zoad). By convention, we project the later Zoad at time t 1 +
into the current Zoped at time t by drawing a horizontal line (Figure 9.10).
We can perform this projection for each point on the time-scale. This results in a
curve that actually parallels the growth curve of the skill itself. This second curve is the
growth curve of the proximal (or potential) development, whereas the original curve
refers to the actual development. Apart from representing the proximal development,
the second curve also represents another important aspect of the developmental
process. According to the definition, the Zoped turns into a Zoad as a function of help
given to the child.
It follows then that from a purely formal point of view, the help given on the one
264 T h e dynamics of mutual interaction

Figure 9.10 A current Zoped represents a later zone of actual development (Zoad).

hand and later developmental effect of that help, in the form of a later level of actual
development, are interchangeable. That is, the curve representing the later Zoad levels
instantiated in the present Zoped levels, is formally similar to the curve representing
the level of help given. Put differently, we can define the level of help as the level of the
1
Zoad that will result out of it after a given amount of time has elapsed.
The curve representing the growing Zoped must be similar to the curve
representing the optimal help - the help that turns the Zoped into the Zoad at time
t+ 1. We do not know what this help is and it will fully depend on the actual
circumstances in which the help is given, but we know that the second curve must
represent that particular form of help, since Zoped and optimal help are equivalent as
far as the transformation into a new Zoad is concerned (see Figure 9.11).
Since we express the help given in terms of the skill level that will be reached later,
we can ask ourselves what would happen if that help lay below or above the ideal point
corresponding with the Zoped. In the first case, the help is too close to the child's actual
level and will add hardly anything significant to what the child already knows or
masters. In the second case, the help is too far from the child's actual level and it will
not be understood or assimilated. In both cases, such help will project onto a later
+
competence level at time t 1 that lies significantly below the level produced by
optimal help (Figure 9.12).

The dynamics of Zoped

How can this model be translated into a mathematical form? First, there is a state the
learner will achieve as an outcome of the learning and instruction process (e.g. mastery
 Vygotsky's dynamic systems

Figure 9.11 The growth curve of the levels instantiated in the help given to the learner
corresponds with the growth curve of the Zopeds, that is, of the sequence of proximal
development levels of the learner who receives that help.

Figure 9.12 Help above or below the optimum will project onto a lower future Zoad,
whereas optimal help produces optimal growth of the actual developmental level.
266 T h e dynamics of mutual interaction
of writing skill). In formal instruction contexts, such as schools, the end state is
explicitly represented, in the form of an educational goal, for instance. But
consciousness of the end state is in no way essential to reaching that state. We have seen
that in the simple logistic model - as in any dynamic model -the end state is a dynamic
process. It is either a stable state that gets maintained over time, or a cyclical attractor,
or a chaotic oscillation and so on. Whatever the end state?it is always a function of the
whole of resources available. The conscious goal representation of an educator may be
part of that complex of resources and in that way play a role in the process, but it is
certainly not necessary.
The learner is put in a situation where the learning of the skill is made possible. The
learner's activities represent some point along the line leading from complete lack to
full mastery of the skill (e.g. a child making scribbles is close to the minimal level). The
'helper' or tutor is able to provide help in that a joint activity is created that is closer to
the goal state than the child's solitary activity would be. In this sense, the helper is
always a little bit ahead of the learner, in the direction of the goal state. If the help is
optimal, the child will be able to interiorize it and so reach a higher level, the
competence level as represented by the current joint activity of helper and learner.
This means that the child is continuously approaching the level instantiated in the help
given. The present model is very similar to the dynamics based on the experience
component (Model 9.1 in this chapter).
The dynamics is highly reminiscent of the apocryphal story of the farmer and the
donkey. The donkey was unwilling to go home, so the farmer tied a carrot to a stick and
held it in front of the donkey. As the donkey went forward to reach out for the carrot, it
made the carrot move forward too. The donkey's fruitless chase continued until they
both reached the farm.
The mathematical simulation of this model is based on two components and two
basic assumptions. The components are the level of the learner on the one hand, and
the level as instantiated by the help given, on the other hand. The model deals only
with the quantitative aspect of the skill or competence growth.
The assumptions are as follows. First, the increase in both the learner's competence
and the helper's level of help given is described by the logistic growth model. The
second assumption is that the growth rate (the speed with which the skill or
competence increases over time) of both the learner's competence and the helper's help
level will depend on an optimum. If the instructor can keep an optimum distance
between the child's actual competence and the help given, the child will maximally
profit from the help and proceed faster than if the help is above or below that optimum.
The optimum not only holds for the learner, but also for the helper. For instance, if
the distance between the actual progress of the child and the help given, as perceived by
the helper, is beyond or below a certain optimum, the helper will tend to stay closer to
the learner's actual competence level than if the perceived distance were optimal. So
the model allows for conflicts between optima that hold for the learner and optima that
hold for the helper. Imagine the case of a very bright child and a tutor who is used to
teaching much slower children. The bright child will profit most from a large gap
between actual competence and help given. The tutor, on the other hand, is used to
 Vygotsky's dynamic systems 267
keeping a much smaller gap between the learner's competence and the help given and
this will eventually be to the disadvantage of the bright fast learner. For instance,
assume that the tutor keeps giving detailed and concrete instructions, forcing the child
to stop and check every single step. A bright child might profit much more from a
teacher explaining general principles of the task. A mentally retarded child, on the
other hand, will profit more from the first form of instruction and learn virtually
nothing from abstract rules that are given verbally.
The mathematical model is very similar to the one from the preceding section. It
contains two coupled equations, one for the learner's level and one for the help (for
more details, see van Geert, 1992b, 1994b; for a somewhat different form of the model,
see van Geert, 1992a). The help is equivalent to the carrying capacity that the level
feeds on. One major difference with the preceding model, however, is that the growth
rates in both equations are multiplied by an optimal distance variable. That variable is
maximal if the distance between level and help is optimal. The optimal distance is
given by an optimality parameter, which is different for the learner and the helper.
This optimality parameter is a function of the so-called unutilized capacity for growth,
which, given that the carrying capacity is 1, amounts to the function (1 - Pi). The
whole model amounts to a combination of four coupled equations: one for the learner's
level, one for the help level and two for the growth rate variable for the learner and
help, respectively:

The variables used in this model are the following: A and Pare the growth levels of the
actual and potential developmental level, respectively. F L A and R4 are their
respective variable growth rates. The growth rates are determined by a constant
growth rate ( r A and 1-4, respectively) and an optimality function, which depends on
optimality parameters ( O A and 0 4 ) and damping parameters ( b A and b 4 ) . The
vertical lines in the fourth equation mean that the absolute value of the enclosed
function has to be taken.
Figure 9.13 shows five different patterns, resulting from different parameter
constellations (see Table 9.1). The ideal pattern - as far as the speed and of growth is
concerned - is the S-shaped curve from Figure 9.13a. It is somewhat flatter in the
beginning than the S-shaped curve that would result from a singular logistic equation.
The S-shaped curve evolves towards a clear stepwise curve or sudden jump, with
increasing 0 9 parameter values (which 'means that the potential level tries to stay
ahead of the actual developmental level, but 'with a bigger distance between A and P
than in the previous case). Figure 9 . 1 3 shows
~ that potential and actual developmental
level sometimes get trapped in a deadlock (note that the scale of representation of
Figures 9 . 1 3 ~and d is different from the other curves; the point where the virtual
standstill occurs is close to the initial state level). Sometimes P and A drag each other
towards a zero level, to a complete disappearance. With still other parameter values,
268 The dynamics of mutual interaction

Figure 9.13 Qualitatively different growth curves result from a continuous change in the
optimal distance parameter; ideal, S-shaped increase (a), a sudden jump in both the help
and the actual developmental level (b); a process where both help and actual developmental
level reach a maximum which is far below the potential maximum (c), one in which both
levels drop back to almost zero after an initial period of increase (d), and an irregular
oscillation followed by a sudden jump to the maximum level (e).

the pattern shows a chaotic fluctuation in the beginning, but sometimes manages to
escape and achieve the maximum attainable level (Figure 9.13e).
As the critical parameter value increases, the qualitative nature of the resulting
learning or growth curves changes, as Figure 9.13 clearly shows. In the vicinity of these
changing points, the resulting curve could be either of the form typical of lower
parameter values or of the form typical of the higher value. Which of these two actually

Table 9.1 Parameter values for growth trajectories in Figure 9.13

 Vygotsky's dynamic systems

Figure 9.14 A final state space for the control variables rate and optimal distance as
applied to the help component in the Vygotsky model; the space is divided in a region with
final states close to K, and final states close to 0; there exists a sharp discontinuity and an
irregular boundary between both regions.

emerges depends on small fluctuations in the starting points. Put differently, in the
vicinity of qualitative switch points, the actual outcome is very sensitive to minor
random influences. Small normally irrelevant differences in starting point, for
instance, could lead to major long-term differences in the resulting developmental
trajectory. Beyond such switch points, the resulting trajectory is not sensitive to minor
perturbations, and thus very predictable (if we know the parameter values, that is).
Finally, it is interesting to note that the model does not always create the ideal
developmental outcome. The parameters can be seen as dimensions in a control space,
and only part of that control space (only part of the possible combinations of parameter
values) leads to an optimal learning or developmental outcome. Take for instance
Figure 9.14.
The vertical axis represents the average final growth level (the last 500 iterations of a
simulation consisting of 2000 iterative steps). The horizontal axis corresponds with a
range of values of the r 4 parameter (see equation 9.20), which refers to the growth rate
of the help level. The depth axis shows a range of values for 0 4 (the parameter
defining the optimal distance between the pupil's level and the &tor's help level). Both
parameters, r 4 and 0 4 , are control parameters for the tutorial aspect in the
Vygotsky model. It is easy to see that eaclipoint on the horizontal plane corresponds
with a combination of a specific r-p with a specific 0 4 value. For each of those pairs of
 T h e dynamics of mutual interaction
values there is a unique value of the resulting average growth level. Figure 9.14 clearly
shows the sudden shifts between values for which the final growth level is about zero,
and values corresponding with a final growth level equal or close to the carrying
capacity level. It also shows that the boundary between the low and the high final states
is quite irregular (for this part of the r 4 / 0 4 control space, that is). This final state
control space provides a good illustration of the non-linearity of the Vygotsky model:
the model shows a sudden jump in the equilibrium level for a specific set of control
parameter values (and, if the jump is also associated with hysteresis and additional
catastrophe flags, the Vygotsky model is an example of a cusp-catastrophe).
I t is important that a developmental model can explain not only success, but also
failure (for instance a zero outcome), since we may assume that the basic
developmental mechanism is not different in both cases (see van der Veer and Valsiner,
1992). The collection of 'successful' parameter values in the control space forms an
irregularly shaped object, clearly demonstrating the non-linearity of the model (van
Geert, 1994b).

Growth under variable instruction conditions

The models I described in this chapter differ in a number of assumptions. They are
similar, however, in the basic assumption that instruction, teaching, help or support by
a more competent person amounts to an increase in the carrying capacity of the skill
that grows with the learning subject. The carrying capacity embodied in the
instructional practice grows in response to growth in the dependent skill. Carrying
capacity and the skill level in the learner are mutually dependent growers.
An interesting aspect of these models is that they show that processes of instruction
or help are not necessarily beneficial to the learner. The dynamics determining the help
level are governed by a set of parameters, whose values may differ over individuals.
Some value ranges could drag the whole process into chaotic oscillation. Others result
in stable state endpoints well below the value that could have been reached if the
parameter values would have been different. Still others result in upward and
downward movements of a learning level that will never reach the goal state. In
general, however, the different models show how a mutualist dynamics can lead to a
predefined goal state as a consequence of a true interaction between an instructor and a
learner. In this type of process, both participants are equally important.
In this chapter, my standard model involved the interaction between a tutee or
novice and an expert, or tutor. This is of course only one of the many forms that a
co-construction process can take. For instance, co-construction and co-regulation may
occur in an interaction between two novices, or between children at different levels of
mastery or knowledge of a co-constructed skill. Interaction between equals involves
the problem of the goal representation in the system. In the cases described in this
chapter, the goal was represented in the form of an expert, adult or teacher, but if two
(or more) novices are concerned, the goal - or should I say, the set of potential goals - is
represented in an entirely different way. Models of interacting equals are the subject of
further research.
Towards an experimental
theoretical developmental
psychology

Unsolved problems abound

Problems with empirical research
The dynamic growth approach that I presented in this book is a new and exciting
approach to development (or, to put it more modestly, I think it is sufficiently new to be
interesting and I am excited each time I try an equation that simulates some
developmental interaction pattern and discover that it'does what I expected; if it does
not, I am sometimes even more excited). In order to really understand why somebody
would spend day after day building mathematical models of development instead of
going to the beach, sunbathing or playing volleyball, it is necessary to do the model
building exercises and see the computer do its conjuring tricks. But all the modelling
exercises explained here and the multitude of dynamic building blocks and models
presented cannot conceal that the unexplained problems largely outnumber the
explained ones.
The first is the empirical problem. In principle, the dynamic growth model pro-
duces curves that describe potential individual growth patterns. An individual is char-
acterized by a set of parameter values and a history of external events that shape his or
her particular growth trajectory. Unfortunately, there is still only little longitudinal
research, especially of the kind that a dynamic model requires, namely data with a high
sampling frequency. I t is possible, though, to use cross-sectional data, especially if they
are based on good developmental scales. In our attempt at building a model of hier-
archically connected growers, Fischer and I used group data we pretended would
characterize at least a significant subset of individuals. In some cases, it will be practi-
cally impossible to obtain the ideal longitudinal data, for instance, when the process of
development lasts much longer than the average research project, Cognitive develop-
ment between birth and 25 years of age is a good example. This does not mean that we
should not try to set up longitudinal research that spans this long period. But in the
meantime, we have to rely on data that cover much less than the time span covered by
the models. I see several partial solutions to the empirical data problem. First, we could
 Epilogue
concentrate more on major developmental processes that take only a short time. A good
example is language development. One child, two years of daily observation and a lot of
little notebooks is all you need to carry out an interesting longitudinal study that
describes the growth of a complex phenomenon, language, between its onset and a
stage that is rather close to maturity on a number of criteria. Infancy is a particularly
good age range, because it comprises many major changes over short periods. Second,
we could perform longitudinal studies of learning processes, for instance, learning in
the classroom or computer-based instruction. Such learning processes can be observed
non-obtrusively, since the testing and observation is usually part of the learning and
teaching itself and does not intervene with that process in a distortive way. Third, we
could focus more on microgenetic studies, which involve processes that recapitulate, in
a sense, long-term changes over a short time span. Nira Granott's studies (Granott,
1993) on how adults try to understand the principles behind the behaviour of small but
complicated electronic robots is a good example of how fruitful such studies can be, but
also a good example of how much is required to arrive at a really interesting measure-
ment. The justification for microgenetic studies lies for a good deal in the expected
scale invariance of developmental processes. That is, we may assume that the really
interesting phenomena repeat themselves over any different time-scale, ranging from a
couple of hours to a lifetime (and even beyond that range). Fourth, as far as long-term
developments are concerned, we could try to focus on the moments where the real
changes take place, provided such changes are indeed unevenly distributed over the
time-scale. The study of transitions requires individual longitudinal designs. We have
seen earlier that transitions are characterized by increased variance, by considerable
changes over short intervals and by fluctuations that approach chaos and turbulence.
As a matter of principle, such processes cannot be studied in group designs. Finally, we
should try to bring our cross-sectional methods more in line with the need for longi-
tudinal data. It is very important, for instance, to have good, theoretically justified
developmental scales that apply to many different contexts. A good example of such
scales are those developed on the basis of Fischer's model of developmental levels.
They combine a strong logical ordering with contextual flexibility. Another way of
improving longitudinal methods is to apply mixed designs, that is, overlapping sets of
repeated measures in different age groups. Instead of focusing on measures that capi-
talize in one way or another on averages or means, we should make more explicit use of
changes in variance and other measures that relate to individual differences.
The basic adaptation, probably, should lie in changing our expectations of how
development takes place. As long as we go on trying to reduce irregularity, turbulence,
dips and bumps, acceleration and deceleration to error variation imposed upon a
basically smooth, gradual and straight underlying process, we will miss the basic
insight that it is the irregularity rather than the smoothness that is essential to the
process we want to understand.

Problems with statistics

Fortunately, psychology can now rely on a truly impressive arsenal of statistical
weapons useful for silencing anyone who dares say that the study of the soul is not real
 Unsolved problems abound 273
science. Unfortunately, however, most of the statistical tools are aimed at a different
goal than the one pursued by developmentalists. They are very well suited for
describing and analyzing differences among populations but considerably less adapted
to the problem of change in individuals. There is a vast literature on time-series
analysis and the growing interest in non-linear dynamics has boosted the development
and distribution of statistical techniques uniquely designed for capturing the
parameters and forms of change. In order to yield interesting results, however, such
techniques usually require a great number of repeated measurements, often much
more than is possible with even the most ideal repeated measurement designs in
psychology.
How can one test whether the growth curves a model has produced really fit the
available data? The simplest test makes use of an old and small instrument that many
people use primarily to watch TV, the eye. In general, you can easily see whether or not
a curve follows the available data, or whether one curve produces a better fit than the
other. If you can see that things are similar, it is not necessary to deploy an army of
statistical forces that comes up with a bunch of numbers which, at best, correspond
with the judgement of the naked eye. In the models I built of early language
development, I relied almost exclusively on the power of the visual sense. It was
possible to make the claims I made more impressive by using a least-squares method or
something else that tells you to what extent, on average, two curves are similar, but that
would not have made the conclusions more reliable or valid. Another easy way to check
data against a mathematically obtained collection of curves is to test for the presence of
distinctive qualitative features. For instance, in building the model of hierarchically
connected growers, Kurt Fischer and I used the data to argue for a number of
qualitative properties, such as growth spurts, temporary regressions and fluctuation.
We then checked the mathematical curves to see whether these distinctive properties
occurred, and if they did, under which combinations of parameters.
This does not mean we should not bother to look for strict and formal statistical
testing procedures. Our quest for such methods should be based on the model of
change that we believe in, instead of the other way around. At this moment, Han van
der Maas and Peter Molenaar are working on statistical procedures testing for
transitions in the strong sense, that is, patterns of change that follow the cusp
catastrophe. When Kurt Fischer and I were working on our dynamic model at the
Stanford Center for Advanced Study, we were fortunate to have some money to pay for
the travel expenses of consultants to the project. One week we brainstormed with John
Willett from Harvard University, who has done a lot of work on statistics for
behavioural growth and change. Although his models and ours were rather different,
John came with a brilliant idea after seeing a time-delay state space diagram (recall that
such diagrams are very easy to draw in a spreadsheet program). He explained that our
growth equations contained variables that were linear in the parameters, that is,
parameters simply added or multiplied values of variables. The oniy thing our growth
equations did was describe a relationship between a previous state and a next state.
That relationship is fully specified by the *equation. Given a sufficient (but not
astronomical) number of repeated measurements, you could take the preceding states
as predictors for succeeding states. The parameters relating pairs of states could be
 Epilogue
estimated by a statistical technique which is usually employed only with group data,
LISREL. Since the nature of the relationships has been fixed by the dynamic equation,
LISREL would be capable of making quite reliable estimations of the parameters.
Once estimated, the parameters could be fed into the equations again, and you could
then test whether or not they produced the required growth curves. At the time of
writing, it is not known to me if this suggestion will be worth the excitement we felt
when we first heard it. But for whatever it is worth, it is a nice illustration of the fact
that with a good deal of creativity you could adapt existing methods to the needs and
requirements of the study of growth and change.

Problems about time and coincidence

When I look at the neat rows of numbers lining up in the spreadsheet as little soldiers
ready to fight the enemy of ignorance, I cannot help feeling that they are all too neat
and regular. The cells in the spreadsheet correspond with single loops of the iterative
equation and each iteration is supposed to take its own time. The iteration time is the
feedback delay of the system. It has no temporal interpretation in the difference model
(and the spreadsheet model for that matter) in that it is not known how much time it
takes to go from one iteration to the next (or from one cell to the next). If I draw a
diagram of the successive points of a growth curve, I usually call the horizontal axis the
time axis and I almost automatically assume that the time intervals between any two
points are equal. But they need not be. If the time between one iteration and the next
depends on some environmental event, such as the child being confronted with a
problem, the length of the intervals will differ greatly over the growth curve. Just
imagine the growth curve is printed on a rubber sheet and the sheet is stretched
unevenly between any two points. The series of cells (or iterations) in themodel is not a
time series, but an event series. Each cell is a learning or growth event, triggered either
by some internal factor or by an environmental and contextual event. A row of cells in
the model represents a single point in time, but given the previous argument, it should
actually represent an event. In the model, a variable changes over each new cell. In
reality, however, a variable changes only if it is involved in some sort of event that
causes it to change. The distribution of such events can be very different for different
growers and the speadsheet model does not account for that; it makes all variables
change in unison, at each tick of its internal clock that makes the system proceed one
row of cells. I t is possible, however, to treat inter-event time as a separate variable
which is determined by the grower to which it applies. The transition model of logistic
growth, for instance, can be simulated by making the interval between successive
learning events depend on the growth level (higher levels, shorter intervals). The
cumulative curve of event times can be employed as the x-axis in an x-y graph, with the
growth level as y-axis.
The problem of variable inter-event time is related to that of coincidental influences
and eventual unreliability of information as it travels through the developmental
dynamics. In the spreadsheet model (and in the mathematical model), each succeeding
cell knows the exact value of the preceding cell and the exact value of the parameters. A
 Unsolved problems abound 275
constant parameter remains literally constant. But how does this work out in reality?
How does the vocabulary system knows how many words it contains? How does the
representational level know how many problems it can solve? In some way or another,
those systems 'know' about their quantitative properties in that a model that mimics
the knowledge can produce quite realistic growth curves. But this does not mean that
the quantitative information in real systems is transmitted without considerable error.
Part of the error is probably due to environmentally triggered coincidental events.
Fortunately, however, it is not too difficult to account for both the random variation in
event intervals and in reliability of the information available to the system. You can add
any form of random variation to each of the steps in the computation of a growth curve
and study if, where and when such random perturbation makes a major difference. In
some dynamics, the random factors add up linearly to the growth pattern. Such is the
case with the logistic growth equation. Just add a random perturbation to each of the
variables and parameters in the cells and you will see that the results are not very
different for different random runs of the model. In other dynamic models, random
perturbations may have non-linear effects. For instance, in Chapter 9 we saw that the
learning model that made use of the carrying capacity adaptation principle suffered
much more if a random perturbation occurred at the beginning than at the end. I am
currently experimenting with growth models which compute growth as a function of a
'fuzzy', randomized representation of an average growth level in the near past. These
models produce interesting oscillatory patterns that are very similar to longitudinal
data sets on the growth of syntax in individual children (Ruhland, unpublished paper).
Models working with 'fuzzy' growth states are probably very appropriate for
describing growth patterns that are (partly) based on the person's mental representa-
tion of his or her skill or knowledge level.

The haunted house of problems

In San JosC, California, there is a place called the Haunted House. In comparison to
Scotland, sunny California is a lousy place for ghosts, but this particular house carries
an interesting metaphor. It belonged to a lady who believed she would never die as long
as she went on building, adding rooms, staircases leading into blind walls and whatever
else her imagination suggested. In the end she died, of course, like everybody else,
when the time has come. In scientific model building, theorists often behave like the
San Jost lady, thinking that they can chase basic problems away by doing more
research and adding more aspects to their models. In the house of the dynamic growth
models, one problem keeps haunting me and even though I think I can conjure it by
continuing to add new parts to the building, I know the day will come when it will
overtake me and that I shall have to face it anyway. The problem is that of novelty, the
construction of new developmental forms out of existing ones. One of the more
fundamental limitations of a growth model is that it does not explain where the growers
come from. What is it that turns a non-verbal act of communication into a word, a
pre-operational concept into an operational one? More than a century ago, Darwin
showed that new species can be created when small random variations on the building
 Epilogue
plan of an existing species are selected by differential survival. Maybe something along
these lines would prove to be the solution for the creation of novelty in development. In
Chapter 5, I have demonstrated how a principle of minor gradual change may
nevertheless produce a jump to a different level, suggesting that gradual local
mechanisms may in principle produce sudden qualitative change. But in order to
develop a theory of novelty creation, a model is needed of a dynamics applying to
detailed structural models of the skills, competences or abilities in question.
A very good example of a field in which the explanation of how novel patterns come
about as a result of non-linear dynamic processes is motor development. Ground-
breaking work in this developmental area has been done by Esther Thelen and
colleagues (see Thelen, 1989, for an overview). The dynamic models apply to a
well-developed structural model, namely that of the physical body: a structure of joint,
stiff levers (the bones) moved and connected by muscles and coordinated neural
components. This particular structure is the domain of study of a discipline called
biodynamics, which explains actual motor patterns, such as walking, reaching and
grasping. My Dutch colleague Peter Beek even managed to describe and explain
juggling in terms of a dynamic systems model of biodynamics (Beek, 1989; I heard he
had to practise juggling himself before he could do this research; psychology indeed
asks for many different talents). In a recent special issue of the journal Child
Development, fourteen papers appeared discussing several aspects of how motor
patterns emerge during development as a consequence of self-organizing principles
that apply to the neural and physical structures of the body. Thelen and colleagues
studied the emergence of reaching in terms of the infants' ongoing intrinsic movement
dynamics (Thelen et al., 1993). A comparable dynamic systems analysis has been
applied to the development of walking (Clark and Phillips, 1993). In their commentary
on the papers, Turvey and Fitzpatrick (1993) note that several of the authors
emphasize the importance of spontaneous movements as exploratory and formative
mechanisms, leading to new motor patterns. The dynamic explanation of qualitatively
new patterns of action has been very successful in the motor domain. Smith and
Thelen (1993) conclude their discussion of the question whether dynamic systems
theory can be usefully applied in areas other than motor development by saying that
'this hammer is well worth trying out'.
Another phantom haunting the house of dynamic models is the problem of different
time-scales, their proper dynamics and how they relate. I have some ideas about how
this problem may be specified further and eventually solved (van Geert, 1992c), but
those ideas have hardly left the cradle. How can a dynamic system determine its own
goal state, or construct its own attractor? This is a problem associated with a model of
co-construction involving two novices. It is, obviously, related but not similar with the
problem of novelty construction.
These are only a few of the many theoretical problems that still hide somewhere in
the many rooms, closets, halls or doorways of the house of dynamic systems of
development. They have managed to escape from being exorcized, so far, and they will
continue to do so as long as we do not find a way to fomulate those problems in explicit
terms.
 Towards an experimental theoretical developmental psychology 277

Towards an experimental theoretical

developmental psychology
This book was primarily about theory, not so much about a new theory of development
as about a new way to reformulate the old ones. There is a lot of dynamic systems
theory behind the models of our founding fathers, Piaget, Vygotsky, Werner and
several others. There is a lot of dynamics behind simple theories of learning and
teaching. The point is, however, that the dynamics remained implicit. It required a
simple method of mathematical model building to peel them out of the verbal,
philosophical form in which they had been caught. Once we managed to translate those
models and theories into a genuine dynamic form, we found out that they contained a
number of unexpected properties. We discovered the regularity and robustness of
some and the turbulence and chaos of others. We saw how a model showed simple
stepwise increase under some circumstances and temporal regressions under others.
The dynamic models defeated the possibility of simple, linear extrapolation. This
notion implies that a local process can be extended and projected into the future and
that the magnitude of the effect is proportional to the magnitude of the time interval.
We are inclined to think that a little progress over a little time interval, corresponds
with roughly five times as much progress over a time interval five times greater. The
, non-linearity of many dynamic processes, however, stands in the way of this simple
extrapolation principle. The only way to know what a dynamic model is doing is to
simply try it out under as many conditions as possible. h he Latin origin of the word
experiment is 'the act of trying', and it is in this sense that the study of the behaviour of
dynamic models, in itself a part of theoretical developmental psychology, is a form of
experimental theoretical psychology. When your models and theories have a dynamic
character, you cannot suffice with a mere verbal form, preceding the inference and
testing of hypotheses. An intermediate stage is required, where a formal procedure is
used to investigate the set of potential hypotheses the model offers. The growth
concept and the dynamic modelling presented in this book is one particular approach
to experimental theorizing that developmental psychologists might apply in their
attempts to come to a better understanding of that most elusive of all phenomena,
change. At the end of the day, we may find out that these models are not any better than
the concrete wooden shoes my grandfather made, but it is better to use concrete
wooden shoes than to wear no shoes at all, because if you hurt your feet, you will get
nowhere.
References

Apter, M. (1982). The Experience of Motivation: A Theory of Psychological Reversals. New York:
Academic Press.
Baddeley, A. (1990). Human Memory: Theory and Practice. Hillsdale, NJ: Erlbaum.
Bak, P. and Chen, K. (1991). Self-organizing criticality. Scientgc American, 265(1), 46-53.
Bee, H. (1989). The Developing Child (5th edn). New York: Harper and Row.
Beek, P. (1989). Timing and Phase Locking in cascade juggling. Ecological Psychology, 1,55-96.
Bentler, P.M. (1970). Evidence regarding stages in the development of conservation. Perceptual
and Motor Skills, 31, 855-69.
Bever, Th.G. (1982). Regressions in Mental Development: Basic Phenomena and Theories.
Hillsdale, NJ: Erlbaum.
Bidell, T.R. and Fischer, K.W. (1993). Developmental transitions in children's early on-line
planning. In M.M. Haith, J.B. Benson, R.J. Roberts Jr. and B.F. Pennington (eds),
Development of Future-oriented Processes. Chicago: University of Chicago Press.
Bijstra, J. (1989). To be or to look . . . that's the guestion. A comparative study on social-cognitive
development. University of Groningen: Doctoral dissertation.
Bleiker, C. (1991). Validating a common-ceiling model of short-term memory growth:
consonant spans in young children. In R. Case (ed.), The Role of Central Conceptual Structures
in the Development of Children's Numerical, Literary and Spatial Thought. Year 2. Report
submitted to-the spencer Foundation.
Boulding, K. (1978). Ecodynamics: A new Theory of Societal Evolution. Beverly Hills: Sage.
Boyd, R. and Richerson, P.J. (1985). Culture and the Evolutionary Process. Chicago: University
of Chicago Press.
Brainerd, C.J. (1973). Neo-Piagetian experiments revisited: is there any support for the
cognitive-developmental stage hypothesis? Cognition, 2/3, 349-70.
Brown, R. (1973). A First Language: The Early Stages. London: Allen and Unwin.
Bryant, P. (1982). Theories about the causes of cognitive development. In P. van Geert (ed.),
Theory Building in Developmental Psychology. Amsterdam: North Holland.
Case, R., Marini, Z., McKeough, A., Dennis, S. and Goldberg, J. (1986). Horizontal structure
in middle childhood: cross-domain parallels in the course of cognitive growth. In I. Levin
(ed.), Stage and Structure. Reopening the Debate. Norwood, NJ: Ablex.
Case, R. (1992a). T h e role of the frontal lobes in the regulation of human development. Brain
and Cognition, 20, 51-73.
Case, R. (1992b). The Mind's Staircase: Exploring the Conceptual Underpinnings of Children's
Thought and Knowledge. Hillsdale, NJ: E r l b a ~ m .
 References 279
Clark, J.E. and Phillips, S.J. (1993). A longitudinal study of intralimb coordination in the first
year of independent walking: a dynamical systems analysis. Child Development, 64, 1143-57.
Corrigan, R. (1983). The development of representational skills. In K.W. Fischer (ed.), Levels
and Transitions in Children's Development (Vol. 21, pp. 51-64). San Francisco: Jossey-Bass.
Costa, P.T. and McCrae, R. (1976). Age differences in personality structure: A cluster analytic
approach. Journal of Gerontology, 31, 564-70.
Costa, P.T., Jr. and McCrae, R. (1980). Still stable after all these years: Personality as a key to
some issues in adulthood and old age. In P. Baltes and 0 . Brim (eds), Lifespan Development
and Behavior (Vol. 3, pp. 65-102). New York: Academic Press.
Dawkins, R. (1976). The Selfish Gene. Oxford: Oxford University Press.
De Sapio, R. (1978). Calculus for the Life Sciences. San Francisco: Freeman.
Dromi, E. (1987). Early Lexical Development. New York: Cambridge University Press.
Dromi, E. (1986). The one word period as a stage in language development: Quantitative and
qualitative accounts. In I. Levin (ed.), Stage and structure. Norwood, NJ: Ablex.
Ehrlich, P.R. (1986). The Machinery of Nature. New York: Simon and Schuster.
Eichorn, D.H., Clausen, J.A., Haan, N., Honzik, M.P. and Mussen, P.H. (1981). Present and
Past in Middle Life. New York: Academic Press.
Eldredge, N. and Gould, S.J. (1972). Punctuated equilibria: an alternative to phyletic
gradualism. In T.J.M. Schopf (ed.), Models in Paleobiology. San Francisco: Freeman.
Eldredge, N. and Tattersall, I. (1982). The Myths of Human Evolution. New York: Columbia
University Press.
Fischer, K.W. (1980). A theory of cognitive development: The control and construction of
hierarchies of skills. Psychological Review, 87, 477-53 1.
Fischer, K.W. and Lazerson, A. (1984). Human Development from Conception to Adolescence.
New York: Freeman.
Fischer, K.W. and Pipp, S.L. (1984). Processes of cognitive development: Optimal level and
skill acquisition. In R.J. Sternberg (ed.), Mechanisms of Cognitive Development (pp. 45-80).
New York: W.H. Freeman.
Fischer, K.W. and Rose, S.P. (1993). Development of coordination of components in brain and
behavior: A framework for theory and research. In G. Dawson and K.W. Fischer (eds),
Human Behavior and the Developing Brain. New York: Guilford.
Fischer, K.W., Kenny, S.L. and Pipp, S.L. (1990). How cognitive processes and environmental
conditions organize discontinuities in the development of abstractions. In C.N. Alexander,
E.J. Langer and R.M. Oetzel (eds), Higher Stages of Development (pp. 162-87). New York:
Oxford University Press.
Flavell, J.H. (1963). The Developmental Psychology ofJean Piaget. New York: Van Nostrand.
Flavell, J.H., Green, F.L. and Flavell, E.R. (1986). Development of Knowledge about the
Appearance-Reality Distinction. Chicago: University of Chicago Press.
Fodor, J.A. (1980). Methodological solipsism considered as a research strategy in cognitive
psychology. Behavioral and Brain Sciences, 3, 63-109.
Fogel, A. and Thelen, E. (1987). Development of early expressive and communicative action:
Reinterpreting the evidence from a dynamic systems perspective. Developmental Psychology,
23(6), 747-6 1.
Fogel, A. (1992). Movement and communication in infancy: The social dynamics of
development. Human Movement Science, 11, 387423.
Fogel, A. (1993). Developing Through Relationships: Origins of Communication, Selfand Culture.
Heme1 Hempstead: Harvester Wheatsheaf. .
Gann, L.H. and Duignan, P. (1979). The Rulers of Belgian Africa, 188G1914. Princeton, NJ:
Princeton University Press.
Gleick, J. (1987). Chaos. New York: Penguin Books.
Globerson, T. (1983). Mental capacity and cognitive functioning: developmental and social class
differences. Developmental Psychology, 19, 225-30.
 References
Gottlieb, G. (1984). Evolutionary trends and evolutionary origins: relevance to theory in
comparative psychology. Psychological Review, 91, 448-56.
Gould, S.J. (1992). Bully for Brontosaurus. New York: Norton.
Granott, N. (1993). Play, puzzles, and a dilemma: Patterns of interaction in the co-construction
of knowledge. In R. Wozniak and K. Fischer (eds), Specific environments: Thinking in contexts.
Hillsdale, NJ: Erlbaum.
Hamilton, E. and Cairns, H . (eds) (1963). The Collected Dialogues of Plato. Princeton, NJ:
Princeton University Press.
Harter, S. (1978). Effectance motivation reconsidered: Toward a developmental model. Human
Development, 21, 34-64.
Harter, S. (1983). Competence as a dimension of self-evaluation: Toward a comprehensive
model of self-worth. I n R. Leahy (eds), The Development of the Self: New York: Academic
Press.
Ho, M.W. and Saunders, P. (1984). Beyond Neo-Darwinism: A n Introduction to the New
Evolutionary Paradigm. London: Academic Press.
Honzik, M.P., MacFarlane, J.W. and Allen, L. (1948). The stability of mental test performance
between two and eighteen years. Journal of Experimental Education, 4, 309-24.
Hylan, J.P. (1898). The fluctuation of attention. Monograph Supplements to the Psychological
Review, 2(6), March 1898.
Inhelder, B. and Piaget, J. (1958). The Growth of Logical Thinkingfrom Childhood to Adolescence.
New York: Basic Books.
Jakobson, R. (1959). Why 'mama' and 'papa'? In B. Kaplan and S. Wagner (eds), Perspectives in
Psychological Theory. New York: BALR.
Kagan, J. and Moss, H.A. (1962). Birth to Maturity. New York: Wiley.
Kahneman, D. (1973). Attention and Effort. Englewood Cliffs, NJ: Prentice Hall.
Kitchener, K.S., Lynch, C.L., Fischer, K.W. and Wood, P.K. (1993). Developmental range of
reflective judgment. The effect of contextual support and practice on developmental stage.
Developmental Psychology.
Klausmeier, H.J. and Allen, P.S. (1978). Cognitive Development of Children and Youth: a
Longitudinal Study. New York: Academic Press.
Kohlberg, L. (1969). Stage and sequence: The cognitive developmental approach to
socialization, In D.A. Goslin (ed.), Handbook of Socialization Theory and Research (pp.
347-480). Chicago: Rand, McNally.
Kunnen, S. (1992). Mastering (with) a Handicap: The Development of Task-attitudes in
Physically Handicapped Children. University of Groningen: Doctoral Dissertation.
Labov, W. and Labov, T. (1978). Learning the syntax of questions. In R.N. Campbell and P.T.
Smith (eds), Recent Advances in the Psychology of Language: Language Development and
Mother-Child Interaction. London: Plenum Press.
Leontew, A.N. (1973). Probleme der Entwicklung des Psychischen. Frankfurt am Main: Athenaum
Fischer.
Lumsden, C.J. and Wilson, E.O. (1981). Genes, Mind and Culture: The Co-evolutionary Process.
Cambridge, MA: Harvard University Press.
Mandelbrot, B.B. (1982). The Fractal Geometry of Nature. New York: Freeman.
Mandler, J. (1992). How to build a baby: 11. Conceptual primitives. Psychological Review, 99(4),
587-604.
Matousek, M. and Peterskn, 1. (1973). Frequency analysis of the EEG in normal children and
adolescents. In P. Kellaway and I. PetersCn (eds), Automation of Clinical Electroencephalogra-
phy (pp. 75-102). New York: Raven Press.
May, R.M. (1976). Simple mathematical models with very complicated dynamics. Nature, 261,
459-67.
Mayr, E. (1976). Evolution and the Diversity of Lzfe: Selected Essays. Cambridge, MA: Harvard
University Press.
 References 28 1
McCall, R.B. (1983). Exploring developmental transitions in mental performance. In K.W.
Fischer (ed.), Levels and Transitions in Children's Development (Vol. 21, pp. 65-80). Sari
Francisco: Jossey-Bass.
McCall, R.B., Appelbaum, M.I. and Hogarty, P.S. (1973). Developmental Changes in Mental
Performance. In Monographs of the Society for Research in Child Development, 38 (3); serial
number 150.
Miller, G.A. (1956). The magical number seven plus or minus two. Some limits on our capacity
for processing information. Psychological Review, 63, 81-97.
Nagel, E. (1957). Determinism and development. In D.B. Harris (ed.), The Concept of
Development. Minneapolis: University of Minnesota Press.
Nelson, K. (1985). Making Sense: The Acquisition of Shared Meaning. New York: Academic
Press.
Neugarten, B.L. (1964). Personality in Middle and Late Life. New York: Atherton Press.
Newport, E.L. (1982). Task specificity in language learning? Evidence from speech perception
and American Sign Language. In E. Wanner and L.R. Gleitman (eds), Language Acquisition:
The State of the Art. Cambridge: Cambridge University Press.
Nunnally, J.C. (1970). Introduction to Psychological Measurement. New York: McGraw-Hill.
Olson, J.M. and Sherman, T. (1983). Attention, learning and memory in infants. In M.M. Haith
and J.J. Campos (eds), Handbook of Child Psychology (Vol. 2: Infancy and Developmental
Psychobiology). New York: Wiley .
Oyama, S. (1989). Ontogeny and-the central dogma: Do we need the concept of genetic
programming in order to have an evolutionary perspective? In M. Gunnar and E. Thelen
(eds), Systems and Development. The Minnesota Symposia in Child Psychology. Hillsdale, NJ:
Erlbaum.
Pascual-Leone, J. (1970). A mathematical model for the transition rule in Piaget7sdevelopmen-
tal stages. Acta Psychologica, 32, 301-45.
Peitgen, H.-O., Jiirgens, H. and Saupe, D. (1992). Fractals for the Classroom.Part two: Complex
systems and the Mandelbrot set. New York: Springer.
Piaget, J. (1929). The Child's Conception of the World. New York: Harcourt, Brace Jovanovich.
Piaget, J. (1957). Logique et iquilibre dans le comportement du sujet. In L. Apostel, B.
Mandelbrot and J. Piaget (eds), Logique et Equilibre. Etudes d'Epistkmologie Ginitique (vol. 2).
Paris: Presses Universitaires de France.
Piaget, J. (1975). L'Equilibration des Structures Cognitives: Problime central du diveloppement.
Paris: Presses Universitaires de France.
Rogoff, B. (1990). Apprenticeship in Thinking: Cognitive Development in Social Context. New
York: Oxford University Press.
Ruhland, R. (unpublished). Transitions in early language development. University of
Groningen.
Sanders, A.F. (1983). Towards a model of stress and human performance. Acta Psychologica, 53,
, 61-97.

Saunders, P.T. (1984).Development and Evolution. In M.W. Ho and P. Saunders (eds), Beyond
Neo-Darwinism: An Introduction to the New Evolutionary Paradigm. London: Academic
Press.
Schroeder, M. (1991). Fractals, Chaos and Power Laws. New York: Freeman.
Sena, R. and Smith, L. (1990). New evidence on the development of the word 'big'. Child
Development, 61, 1034-54.
Siegler, R.S. (1983). Information processing approaches to development. In W. Kessen (ed.),
History, Theory, and Methods (pp. 129-211). New York: Wiley. .
Siegler, R.S. (1984). Mechanisms of cognitive growth: variation and selection. In R.J. Sternberg
(ed.), Mechanisms of cognitive development. New York: Freeman.
Siegler, R.S. and Crowley, K. (1991). The micro-genetic method: A direct means for studying
cognitive development. American Psychologist, 46(6), 606-20.
 References
Simpson, G.G. (1983). Fossils and the History of Life. New York: Freeman.
Smith, L. B. and Thelen, E. (1993). Can dynamic systems theory be usefully applied in areas
other than motor development? In L.B. Smith and E. Thelen (eds), A Dynamic Sj~stems
Approach to Development: Applications. Cambridge, MA: M I T Press.
Strauss, S. and Stavy, R. (1982). U-shaped behavioral growth: Implications for theories of
development. In W.W. Hartup (ed.), Review of Child Development Research. Chicago:
University of Chicago Press.
Strauss, S. (1982). U-shaped Behavioral Growth. New York: Academic Press.
Ta'eed, L.K., Ta7eed,0. and Wright, J.E. (1988). Determinants involved in the perception of
the Necker cube: an application of catastrophe theory. Behavioral Science, 33, 97-1 15.
Thatcher, R.W. (1992). Cyclic cortical reorganization during early childhood development.
Brain and Cognition, 20, 24-50.
Thatcher, R.W., Walker, R.A. and Giudice, S. (1987). Human cerebral hemispheres develop at
different rates and ages. Ycience, 236, 1110-13.
Thelen, E. (1989). Self-organization in developmental processes: Can systems approaches work?
In M. Gunnar and E. Thelen (eds), Systems and Development: The Minnesota Symposia in
Child Psychology. Hillsdale, NJ: Erlbaum.
~helen,-E.,Corbetta, D., Kamm, K, Spencer, J.P., Schneider, K. and Zernicke, R.F. (1993).
The transition to reaching: Mapping intention and intrinsic dynamics. Child Development, 64,
1058-98.
Thomas, A. and Chess, S. (1986). The New York Longitudinal Study: From infancy to early
adult life. In R. Plomin and J. Dunn (eds), The Study of Temperament: Changes, Continuities
and Challenges. Hillsdale, NJ: Erlbaum.
Tomlinson-Keasey, C., Eisert, D.C., Kahle, L.R., Hardy-Brown, K. and Keasey, B. (1978).
The structure of concrete operational thought. Child Development, 50, 1153-63.
Turvey, M.T. and Fitzpatrick, P. (1993). Commentary: Development of perceptionaction
systems and general principles of pattern formation. Child Development, 64, 1175-90.
Valsiner, J. (1987). Culture and the Development of Children's Action: A Cultural-Historical
Theory of Developmental Psychology. New York: Wiley.
Valsiner, J. (1991) Introduction: Social co-construction of psychological development from a
comparative-cultural perspective. In J. Valsiner (ed.j, Child Developnzent within Culturally
Structured Environments. Volume III: Comparative-cultural and Constructivist Perspectives.
Norwood, NJ: Ablex.
van der Maas, H. and Molenaar, P. (1992). A catastrophe-theoretical approach to cognitive
development. Psychological Review, 99, 395417.
van der Maas, H , (1993). Catastrophe Analysis of Stage-wise Cognitive Development: Model,
Method and Applications. University of Amsterdam: Doctoral dissertation.
van der Veer, R. and Valsiner, J. (1992). Voices at play: Understanding van der Veer and
Valsiner. Comenius, 48, 423-9.
van Geert, P. (1985). Sociocultural reproduction and individual mental development. Cognitive
Systems, 1(3), 207-3 1.
van Geert, P. (1986). The structure of developmental theories. In P. van Geert (ed.), Theory
Building in Developmental Psychology. Amsterdam: North Holland.
van Geert, P. (1986a). The concept of development. In P. van Geert (ed.), Theory Building in
Developmental Psychology. Amsterdam: North Holland.
van Geert, P. (1986b). The structure of developmental theories. In P. van Geert (ed.), Theory
Building in Developmental Psychology. Amsterdam: North Holland.
van Geert, P. (1987a). The structure of developmental theories: A generative approach. Human
Development, 30(3), 160-77.
van Geert, P. (1987b). The structure of Erikson7smodel of the Eight Ages of Man: A generative
approach. Human Development, 30(5), 236-54..
van Geert, P. (1987~).The structure of Gal'perin's theory of the formation of mental acts: A
 References
generative approach. Human Development, 30(6), 355-81.
van Geert, P. (1988). A graph theoretical approach to the structure of developmental models.
Human Development, 3 1(2), 107-35.
van Geert, P. (1990). Theoretical problems in developmental psychology. In P. van Geert and L.
Mos (eds), Annals of Theoretical Psychology: Developmental Psychology. New York: Plenum.
van Geert, P. (1991). A dynamic systems model of cognitive and language growth. Psyclrologicnl
Review, 98, 3-53.
van Geert, P. (1992a). The complicated dynamics of a simple interaction model. Unpublished
paper.
van Geert, P. (1992b). Vygotsky's dynamic systems. Comenius, 48, 383-40 1.
van Geert, P. (1992~).The dynamics of developmental sequences: Towards a calculus for
developmental theory. Unpublished paper.
van Geert, P. (1993). A dynamic systems model of cognitive growth: competition and support
under limited resource conditions. In L.B. Smith and E. Thelen (eds), A Dynamic Systems
Approach to Development: Applications. Cambridge, MA: M I T Press.
van Geert, P. (1994a). Dynamics of development. In R. Port and T. van Gelder (ed.), Mind as
Motion. Cambridge, MA: M I T Press.
van Geert, P. (1994b). Vygotskyan Dynamics of Development. Human Development, 37.
Vygotsky, L.S. (1958). Mind in Society: The Development of Higher Psj~chologicalProcesses.
Cambridge, MA: Harvard University Press.
Waddington, C.H. (1957). The Strategy of the Genes: A Discussion of Some Aspects of Theoretical
Biology. London: Allen and Unwin.
Werner, H. (1948). Comparative Psychology of Mental Development. New York: Science
Editions.
Winfree, A.T. (1987). The Timing of Biological Clocks. New York: Scientific American Library.
Zajonc, R.B. and Markus, G.B. (1975). Birth order and intellectual development. Psychological
Review, 82, 74-88.
Zeeman, E.C. (1976). Catastrophe theory. Scientgc American, 234, 65-83.
Index

abstract mapping 138, 154, 218, 232 historical emphasis 84

principle 196 anomalous variance 149, 151
system 138, 154 Anytus 10
thought 194 appearance-reality distinction 128, 130
tier 40, 215, 218 appropriation 33
understanding 194 of culture 112
abstraction approximate growth 103, 104
Fischer model 138, 215, 230 Apter 123, 170
acceleration 272 arithmetic 218
accidental cause 78, 79 growth curve 215
accommodation 17, 18, 63 learning 157
action tier 218 operations 231, 237
see also tier understanding 154, 228
activation 108, 157 arousal 108, 157
actual development 2624, 267 oscillatory function 157
adaptation 17, 39, 262 pattern of 159
adaptation artificialism 2
equation 18 assimilation 17-1 8, 62
iterative function 17 asymmetric reciprocity 206, 209, 210
Piagetian principle 17 asymptotic growth 103
addition 90, 154, 221-4, 231, 232, 234, 237, 256, scenario 35
257, 263 attachment 123
control variables 86 attention 85, 157, 159-61, 165, 166, 168, 171,
growth curve 89 172, 175-7, 194
problem 85 allocation 159, 177
adequacy of reaction 258 aspects of 157
aggressiveness 66, 71, 72 domain 173
context-dependent 67 fluctuation 157, 171
fluctuating 67 resource 171
aha-erlebnis 125 resource function 169
alertness state in newborns 159 resource investment model 172
Allen 113 span 120
allopatric attractor 276
growth applied to development 112 basin 58, 140
speciation 112 behaviour 206
alpha frequency 156 binary 202
analytically solvable equations chaotic 103, 122
and computers 84 cyclical 24, 43, 103-4, 266
attractor (continued) Brainerd 141
point 24, 27, 32-3, 35, 38, 43, 103, 122, 125, Bransford 159
140, 171, 252 Brown 15, 41, 42, 82, 122, 154
state 204, 244, 245 Bruner 135
autocatalysis Bryant 209
internal dynamics 97 building block
autocatalytic increase 96 of dynamic model 220
average man 78 bump 232, 234, 272
equation 234
Baddeley 107 butterfly
Bak 133-5 effect 89, 107, 179, 182
Balance Beam task 188 of Lorenz 178
bandwidth population 178
defining a state 46
transition criterion 47 C programming language x
BASIC x calculus 18
basin of attraction see attractor before computers 84
Bay checkerspots 178 Cairns, 10
Bee 127 California Longitudinal Study 44
Beek 276 Campione 159
behaviour control system 215 candy box scenario 33
Beiker 121 carrying capacity 99-103, 115, 119, 121, 144,
Belgian Connection 152 168, 180, 181, 191, 192, 197, 201, 206, 211,
Bentler 127, 130, 148 224, 229, 230, 232-5, 238, 243, 256-60, 267,
test 130 270
benzene ring 125 adaptation 275
Berkeley Longitudinal Study 71 and reversal shift 123
Between-transition interval 137 as educational model 255
Bever 209 compensation 234
bi-modality 149 for new words 116
bi-stability 125, 126, 128 increase 257
competition system 203 parameter 236
state 151 two meanings 234
Bidell 131, 144 variable 234, 236
bifurcation 104, 106 Case 107, 108, 120, 121, 134, 135, 187, 188, 195,
behaviour 104 210, 214, 232, 233
diagram 105, 106, 254 model 190, 192, 195, 233
big theory of development 187
development of 204 catastrophe 147, 149, 151
meaning of 203 flag 147-51, 149, 270
Bijstra 128-30, 149 model 147, 151
bimodal theory 49, 145
distribution 128, 150, 262 see also cusp
score-distribution 148 Cattell 80
scores on a conservation test 130 cause 79
variable 224 abrupt effects 6
biodynamics 276 accidental 78, 79
biorhythm 157 constant 78
Birch 44 smooth 6
Birthday Party task 188 qualitative distinction from effect 7
Bleiker 121 celirange 4
Boston Study 44 Center for Advanced Study xi, 95, 174, 187, 213,
Boulding 113 228, 273
Boyd 124 central conceptual structure 134, 187-90, 194-5,
brain 210, 214, 233
and cognitive growth 120 model of growth 232
development 108 central dogma of evolutionary thinking 2
growth 156 change 1
 Index 287
change (continued) reciprocal 201, 204, 206, 207
historical and phylogenetic 186 competition and support 194, 220-1, 225-6,
paths 31 232
uniform 31 simultaneous 210, 21 1
chaos viii, 85, 89, 102, 103, 106, 122, 178, 184, model 191
199, 226, 228, 252-4, 272, 277 see also reciprocal, support
chaotic competitive relationship 222, 225
domain 180 complete connectedness 50
dynamics 187 complexity viii
fluctuation 268 concatenated dimension
growth 103 of control variables 88
growth pattern 184 of problem difficulty 86
models 186 concept
oscillation 168, 220, 236, 266, 270 of causality 111
see also attractor of object 111
Chen 133-5 concept development 113
Chess 44 concrete, identity, classificatory, and formal
Chomsky 73 levels 113
circadian rhythm 157 conception of the world 1
circumflex symbol 241 artificialist I
cladogenesis 112 conceptual structure 189
Clark 276 concrete operational
Clarke 69 cognitive functioning 63
classical theory cognitive organization 63
Erikson 45 concrete operational thinking 39
learning theory 34 stage 288, 209
Piaget 45 conditional
classification 46 coupling 5, 9
Clausen 44 relation 189
co-construction 112, 243, 244, 254, 270, 276 conflict mechanism 63
model 247 Congo 81
co-regulation 243-5, 247, 254, 270 connected grower 119, 176, 179, 189, 192, 222,
action 244 236-7, 247
dynamic model 245 master equation 238
cognitive model 224, 230
conflict 17, 63, 89, 141 standard mathematical format 191
development 212, 271 connected growth 186, 187
grower 115 basic mathematics 187
growth 192 matrix 189
level 214 patterns 240
principle 247 connectors 105
species 124 consensual frame 244
collapse 240 conservation 12, 45, 46, 73, 127-9, 140, 148-50,
combinatorial principle 247 223, 228
communication 245 competence and performance 74
and investigation 76 development 141
communicative event 246 four Piagetian states 129
compensatory education 202, 203 group data 127
competence 73, 74, 85 problem 141-2
definition 75 self-similar over all scales 130
dynamic phenomenon 76 testing procedure 128
in language 74 training 140, 141,149
measurement 75 transition 130
vs ability 75 understanding 141 '

competence and performance conserver

Chomskyan notion 73 real intermediary and non- 128-30, 142
competition 114, 189, 194, 201, 240 constant causes 78
asymmetric reciprocity 206 constraint 228
 Index
construction of new developmental forms 275 trajectory 40
contagiousness of attention 172 transitory paths 38
context-dependence 94 see also attractor, cyclical
contingency 258
continuity and discontinuity 7 damping parameter 171, 176
continuous process 110 Darwin 2, 275
control dimension 86 Darwinian theory 114
control parameter 107, 142, 150, 244, 270 Dawkins 124
control space 88, 142, 269 dicalage 113, 221
mapping onto a single variable 86 deceleration 272
control system 218 path 33
control variable 86, 88, 142, 146, 148, 150 decentration 12
essential 87 delay 160, 167, 168, 240
transient 87 delayed recovery of equilibrium 149, 151
Cooper 66, 72 dependence between growers 188, 210, 224
cooperation 81 dependentlindependent variable model
in cognitive development 8 1 212
teaching 81 derivative 84, 110
Corrigan 42, 122, 163 DeSapio 101
cortical organization 156 deterministic process
Costa 44 and randomness 84
coupled and uncertainty 178
difference equations 249 assumption of independence 82
equations 50, 52, 84, 170, 172, 175, 179, 192, in co-construction 243
249, 267 in development 1
oscillator 157 in social process 242
credit card development
effect 163 board game metaphor 243
model 166, 174 descriptive space 21
crisis 63 ecology of x
croissant 124 fractal nature x
cross-sectional from potential to actual 262
data 219, 226, 271 gradual change 6
method 272 intuitive concept 95
research 28, 154, 219 patterns 14
Crowley 134, 139 stage theory view 7
cubic model 223, 230 stagewise 45
cultural history 243 variety 49
cultural transmission 243, 255 developmental change
bi-directional process 243 smooth and continuous 82
of knowledge and skills 124 developmental curve 38, 85
transformation 243 coincidental character 90
types 124 developmental cycles 124
unidirectional view 243 dynamics 22
vessel metaphor 243 failure 270
culturgen 124 developmental irregularity 49
cusp catastrophe 126, 146-8, 151, 270 developmental level 237
cycle x, 43 and problem solving 131
in classical theories 38 context-dependency 2 14
of restlactivity or attention/boredom 177 developmental mechanism 14
of sensitization and habituation 157 path 31
quasi- 41 problem 80
transitory 43 developmental process
cyclical elementary iterative form 20
change 122, 156 developmental psychology and ecology 178
growth 104 developmental
path of cortical organization 156 . space 21
process 38 spurt 148
 Index 289
developmental stage randomness 118
Piaget 6 represented by its state or phase portrait 56
developmental state 20-1 time evolution 50
developmental systems 3 1 topological representation 56
developmental theories ix Vygotsky's model 262
and dynamic systems 61 dynamic system of development 64
as dynamic systems models 63 difference with developmental theories 61, 64
empirics 63 example 51
necessities 63 laws of change 64
statics 63 dynamic systems modelling in developmental
developmental trajectory 212 psychology 14, 63, 244
deductive investigation 18 dynamic systems model 244
developmental transition 150 and developmental theories 63
catastrophe flags 147 basic building blocks 211
dichotomous descriptive dimension 45 theory 276
difference dynamics
equation 84, 110 and dynamics equations 54, 64
model 110 and statics 58
differential calculus 83 and surface bending and folding 60
differential equation 110 developmental viii, 22
analytically solvable 84 dynamics model 210
differentiation 64 and developmental theory 61, 63
dimensional stage 188, 195
dip 232, 234, 240, 272 early language development 115, 273
discrete process 110 early word learning 117
disequilibrium 39, 139 ecological co-determination 178
distance ecological metaphor
dimension in developmental space 22 heuristic principles 114
representation 29 ecological '
uniform reduction 3 1 niche 187
divergence of linear response 148, 151 principle 2 12
division 231 system 114
division of labour 243 ecology 178
domain-specific structure 187 mental 113-1 4
doubling scenario 7, 8, 15, 31 ecosystem
explosion 8 human cognitive system 114
Dromi 113, 115, 116, 122, 154, 176 of mental 'species' 114
Duignan 81 effort 85, 108, 123, 157, 159-61, 165, 166, 172,
dynamic equation 274 175-7, 194
hysteresis 150 fluctuation 171
dynamic growth model 95, 187, 213, 214, 216, interest 172
218, 219, 240, 255, 271 level 175
. dynamic model 191, 246, 255, 271, 277 Ehrlich 178
building 190 Eichorn 44
dynamic properties of growth 94 Einstein 125
of the physical body 275 Eldredge 7
dynamic process 204 elementary iterative form
measuring 66 developmental 20
observing 66 elementary theory of learning 247
dynamic system 38, 56, 77, 153, 187, 189, 214, emergence of new levels 218
276 empirical problem 27 1
and developmental processes ix, 64 empirics 65
and systems in general 49 of a developmental theory 63
as a folding procedure 58 entry level 133, 144
as a topological landscape 58 environment 115
definition 49, 50 epigenetic landscape 58
formal definition of 56 equation
idealized portrait 93 difference and differential 84, 110
 Index
equation (continued) learning theory of meaning development 1I
for hierarchically connected growth 237 Fogel 112, 115, 243-5
equilibration 38 forgetting 100
equilibrium 39, 99, 149, 151, 180, 181, 192, 199, formal
202, 206, 229, 233, 234, 262, 270 growth model 99
density 99-101 instruction 266
point 146 formal operational
sheet 150 cognitive functioning 63
stage 171 thinking 33, 39
state 102 formal statistical testing 273
Erikson 33, 45, 63 formal thinking 245
error 68 fractal 15, 91, 125, 130, 138
correction techniques 69 and self-similarity in development 130
in psychological measurement 69 development and the quantity of
procedural 70 transitions 135
randomly distributed 69 developmental 125
simple 68 events in sandpile 133
source of 71, 90 nature of developmental processes 126
term 182 path of development 134
variance 221 process and between-transition intervals
variation 44, 272 137
event series 89, 274 process 139
evolution 178 process of cognitive growth 139
endpoint 2 system 137, 144
punctuated equilibrium view 7 tree 139
evolutionary system 114 fraction problem 46, 47
Excel1 x Freud 80
exhaustible resource 166 fizzy property 68
experience 248, 249
experimental theoretical developmental Galton 80
psychology 241, 271, 276 Gann 81
extrapolated curve 92 gene 124
extrapolation as building plan 1
linear 10, 13, 84, 277 general task domain 218
of measurement points 91 generalization
over a population 219
feedback delay 30, 109, 110, 117, 119, 121, 122, over time 219
142, 163, 274 under- and over- 253
Fels Longitudinal Study 44 generation 139, 152
Ferrara 159 generational succession 243
final state 33, 38, 58, 259 genetic information 186
space 61 genome 1
Fischer 27, 28, 40, 45, 71, 108, 132-5, 137, 144, germinal form 97, 111, 112
147, 154, 156, 214, 215, 218, 219, 224, 226, germinal state 111, 112
240, 241, 271-3 Gestalt psychology 125
Fischer model 40, 194, 201, 212, 214, 237, Giudice 108
272 Gleick 89, 178
levels and tiers 135 Globerson 107
of skill development 214 goal 266
phase portrait 40 shift 255
Fitzpatrick 276 state 248, 254, 255, 257, 270, 276
Flavell 128, 129 Goldschmid and Bentler test 127-8
fluctuation 69, 91, 240, 272, 273 Gottlieb 112
equation 167 Gould 3, 7
of attention 157 gradual process 272
qualitatively similarity 91 gradualism 7
random 32 Grahame 164
Fodor 11 grammar 74, 79
 Index
grammatical of the use of a strategy 140
growth 113 of understanding 154
rules 113, 114 of utterance length 122
Granott 132, 133, 272 of words 112
robot experiment 132 onset time 111, 114, 117
grasping 276 oscillatory 103, 156
Green 128 parameter 236
group pattern 180, 221, 226, 241
and processes 80 power law 137
data 219, 240 process 118, 166, 170, 177
design 272 qualitatively different types 104
grower 97, 173, 177, 188-91, 194, 196-7, 199, quantitative change 96
208, 210, 212, 218, 230, 231, 237, 240, 241, resource dependency 96
256 smooth 8
and its environment 115 stage-wise character 240
as task 234 stepwise 38, 166, 240, 277, 321-3
cognitive 115 structural property 96
competitive 194 unused capacity for 99
connected 192 U-shaped 209
dependent 192 variable 97
Fischer model 216 Verhulst model 101
hierarchically related 195 within- and between-level 23 1
supportive 194 growth curve 85, 90, 100, 115, 122, 166, 168,
growth 112, 247, 270 174, 221, 240, 273-5
acceleration 45 connected growers 230
approximate 103, 105 context- and problem-dependency 90
as a dynamic process ix intrinsic fluctuation 94
asymptotic 37, 103 irregular pattern 163
autocatalytic 96 of addition 89
basic model 98 of potential development 263
biological 137 scale invariance 91
chaotic 103 succession of discrete events 89
concatenated dimension 86 growth dynamics 178
continuous versus discrete 109 growth equation 101, 115, 170, 191, 198, 218,
cyclical 105 246, 273
definition 96 bifurcation behaviour 104
delayed function 99 non-linear behaviour 103
formal model 99 random noise 118
fractal 138 growth model 218, 224, 262
function of time 84 basic properties 95, 98
general properties 95 growth rate 97, 100, 102, 103, 105, 114-17, 119,
grammatical 113 121, 141, 143, 144, 151, 154, 161, 163, 166,
in mental and behavioural development 170, 171, 173, 176, 180, 191, 192, 194, 197,
107 201, 206, 207, 223, 238, 258, 266, 267
' isochronic principle 13 oscillatory pattern 161, 163
level 98, 111, 113, 141, 143, 167, 180, 189, positive and negative 100
24 1 proportional to the resource 161
lexical 36 growth spurt 127, 134, 137, 147, 166, 214-15,
logistic 101, 127, 152 240, 273
logistic equation 25 Gumnan-scale 220
mathematical model 97 Guust Flater 146
metaphor 95, 96
minimal structural level 111 Haan 44
non-linearity 104 habituation 123, 157,' 1 6
of addition 86 and saturation 172
of memory span 108 Hamilton 10
of sentence length 83 Harter 193
of the child's self-esteem 51 Held 133
292 Index
help 262-4, 266, 269, 270 intermediate transition model 199
optimal 264, 266 interrogative 154
heterochronic process 30 interval pattern 139
heterochrony 12 intrapolation 91-3
Heymans 157 invariance
hierarchical model 224, 233 qualitative 90
hierarchically connected growers 195, 222, 271, investigation as communication 76
273 I Q 69-71, 81, 108
master equation 237-8 score 44
higher order structure 214 testing procedure 70
Ho 112 irregularity 44, 272
homme moyen 78 model explaining 94
Hoznik 44, 71 response to 44
human differences 78 isochronic
Hylan 157-8, 168 growth 13
hysteresis 148, 150, 151, 270 mechanism 12
process 9
idealization 93 Isochrony and (non-)linearity 10
and empirical test 93 isomorphy in measurement 70
imitation 112, 255 iteration 2, 14, 19, 31, 34, 88, 110, 116, 117, 269,
Improv x, 241 274
inaccessibility 148, 149 basic developmental mechanism 14
increase biological processes 15
additive 32 hnctions 14
logarithmic 32 geometric 15
independence in mental and behavioural development 15
and smoothness assumption 82-3 iterative
statistical 68, 81 equation 102, 274
individual differences and statistics 80 function 31, 33
individual variation 219 hypothesis testing 18
information processing model 18 iterative processes ix, 2, 245-6
speed 206 in learning 20
theory 14 in maturation 20
Inhelder 245 in spreadsheet 8
initial level 121, 144, 167, 197 iterative step 89
initial state 53, 58, 102, 117, 119, 129, 138, 144, iterator 38, 246
180, 182, 207, 229, 232, 250, 258, 260
plane 60 Jackson 129
set 60 Jakobson 112
space 60 joint factors 238
innate concepts 111 juggling 276
innatist view 11 jump 126, 128
instruction 270 Jiirgens 106
integration 64
intelligence 71, 81 K 100
change over time 71 K-grower 260
fluctuation 44 Kagan 44
multidimensionality 7 1 Kahneman 107
intelligence test 70 Kansas City Study 44
interaction Kekult 125
between equals 270 Kenny 154
dynamics 243 Kitchener 154
mutual 242 Klausmeier 113
interest 166 Kohlberg 209
contagion 171 Kumen 193, 199
domain 172
oscillation 25-6 L' 100
interiorization 18, 262 Labov 154
 Index
language logistic form 163
competence and performance 74 logistic growth 122, 127, 175, 266
development 77, 79, 154, 272 basic equation 100
Lazerson 108 model 94, 101, 122, 127, 163, 199, 232, 252,
leap to a higher level of mastery 36 256, 266
learnability 114 non-linearity 104
learning 33, 89, 100, 248, 264, 266 logistic transition 140, 15 1, 152
classical theory 34 as catastrophe 149
encounter 223 mathematical model 142, 197
event 141, 142, 274 long-term development 272
goal shift 255 longitudinal
hypothesis testing 18 data 47, 272
level 258 design 272
process 34 research 271
state 248 study 44, 79, 82, 113, 115, 128, 219, 272
theory 214 study of conservation development 148
least squares method 273 time series 28
Leontew 108 Lorenz 178-9
level 7-8 equation 179
attention 175 Lotka 175
entry 132 Lotka-Volterra model 174
Fischer model 135, 137, 215, 218, 220-2, 224, Lotus x, 102-3, 105, 170, 241, 259
231, 234, 237 Lumsden 124
information and demand 257 Lynch 154
level of single representations 241
opening and optimal 13 1 macro 259, 260
: Lewis Carroll 90 macrodevelopment 131
lexical development 154 Mandelbrot 15
substage 119, 154 Mandler 111, 112
lexical growth 36, 122, 163, 176 mapping
lexicon 36, 108, 115-17, 122, 163, 167, 176 level 138, 154
fuzzy initial state 117 in' Fischer model 40, 215
life span 153 McCall 44, 71
Lindenmayer 138 McCrae 44
system 138 Marcus 71
linear extrapolation see extrapolation Matousek 156
linear relationship 194 matrix 190-1
LISREL 274 maximum level 99
log-linear relationship 30 May 101, 102
logarithm of the change 32 Mayr 112
logarithmic distribution 138 mean of a property 78
of transition intervals 137 mean length of utterance 163
logarithmic plot 134 meandering change model 182, 184
, representation of power series 139 measurement 77
logical thinking as a dynamic process ix
Piagetian definition 62 concept 69
logistic curve 121 delinition 70
logistic equation 25, 51, 102, 108, 110, 115-16, error ix, 32, 69-71
122, 124, 142, 144, 154, 167, 168, 170, 191, generalization 91
199, 222, 237, 251, 259, 275 in intelligence 81
cubic and the quadratic version 167 qualitative invariance 90
difference form 84, 101 scale 130
differential form approximation 122 sources 71
feedback delay 110 tool isomorphy 68
for reciprocal support 196 mechanism
properties 102 proximal 64
standard 251 single case study 79
transition 25 1 meme 124
 Index
memory dilemma 212
for consonants 121 trait 78
limitation 85 more-word sentence 36
short-term 120 morpho-syntactic structure 168
working 11 Moss 44
see also working memory motivation 108, 175
memory growth and the manipulation of motor development 275
feedback delay 120 pattern 276
Meno 10 moving average 5, 47
problem 10-11 multiplication 231
Menon 112 Muppets 14
mental action Mussen 44
reversible 10 mutation 2, 112
mental content mutual causality 77
meaning 11 interaction 178, 242
mental ecology 115 mutualist dynamics 247, 27
mental growth complex properties 25 1
limited resources 107 random walk 254
mentally retarded child 196, 267 mutuality between researcher and subject
micro-development 131-2, 150 77
and self-similarity 130 mutually dependent growers 270
process 132, 204
scale 204 Nagel 95
micro-genetic naming explosion 36
experiment 139 Narcissus 75
method 134 Necker cube 125, 140
studies 272 Nelson 115
Miller 107 neo-Piagetian model 214
mind's staircase 187 network of growers 212
minimal structural growth level 111, 117, 198, Neugarten 44
228 neural network 205
Miss Piggy 14 New York Longitudinal Study 44
mixed design 272 Newport 115
model noise 72, 73
and reality 85 non-gradual effects 12
as caricature 93 non-linear behaviour 103
Fischer x, 40 conditionally coupled processes 9
non-linear dynamic 3 1 doubling scenario 9
Piaget 38 dynamic model 31
spreadsheet ix dynamics 38, 179, 273
standard model of development 3 1 grower 179
model building viii, 90, 218, 237 model 185
criteria viii non-linear effect 192
developmental vii predictability 179
dynamic 14 system 102
and model collapse 236 non-linearity viii, 38, 104, 106, 270
model of connected growers 228-30, 240 in Vygotsky model 270
cubic version 230 intuitive definition 9
models of development normal distribution 69
Erikson 33 novelty 275-6
Kohlberg 33 problem 10
Piaget 33 numeric
Molenaar 49, 147-9, 273 experiment 226, 241
mommy-daddy-words 112 knowledge 195
monotonic task 189
increase 38 testing 226
trend 221 understanding 190
moral Nunnally 70
 Index
observation 76 steady state 32
as communication 76 stepwise 38
distortion 67 with changing growth rates 35
limit 66-7 Pearson 80
unobtrusive 76, 85, 87 Peitgen 106
observation error 66 perceptual salience 206
correcting for 68 performance 73
of the present progressive 122 dip 216, 240
of the rate of word learning 122 in language 74
ogive 251 measure 218
see also S-shape periods of stability 240
Olson 157, 159, 198 perturbation 125, 149, 151, 173, 251, 269,
one-word 275
sentence 115 random 170
stage 116 Petersen 156
operational thinking 48, 63, 112 phase
see also formal operational developmental 20
optimal plot 199
cognitive level 133 phase portrait 27, 57
distance 267, 269 first derivative method 26
help 264 of a regular sine wave 27
level 144 of Fischer's theory 40
optimum 241, 266 time delay method 27
Origin of Species 2 phase space 27, 41
oscillating representation 168
function 162 Philip IV 93
process 27 Phillips 276
resource 165, 173 Piaget 1, 14-15, 38, 48, 49, 62, 63, 80, 95, 113,
oscillation x, 152-4, 156-8, 168, 175-7, 252-3 124, 129, 130, 135, 139, 141, 147, 156, 209,
chaotic 168 214, 245, 247, 276
in cognitive learning 49 Piagetian
in development 154 adaptation mechanism 17
in growth 103 stage criteria 80
in language development 41 Pipp 27, 71, 154
in grammatical errors 43 Plato 10, 112
irregular 168 point attractor see attractor point
normalized 162 population
of attention, effort and energy 158 dynamics 178
of singing and self-esteem dynamics 53 growth 101
of the alpha frequency band 156 model 21
overshoot 232-3, 234 trajectory 21
Oyama 2 positive feedback loop 140
possibility 96
parameter plane 60 potential development 262, 267
Pascual Leone 107 potentialities 3
pass level 86, 87, 220 power
as continuous signal 88 distribution 139
event series 89 function 137
path law 137
chaotic 43 series 139
cyclical 38 practice and support testing method 28-9, 215,
deceleration 33 218-19
developmental 3 1 pre-operational
non-linearity in point-attractor 38 logic 35
pathological deviation 240 thinking 48, 111, 112, 148
patterns of communication 245 precursor 189, 223, 224
of change 31 model 224
rise-and-decay 34 parameter 237
 Index
predator-prey order principle 89
model 175 perturbation 12, 112, 170, 173
model of attention and effort 174 variations conditionally coupled 6
predecessor 223, 224 random walk 250, 254, 255
prediction 182, 183 'lifelines' 5
horizon 185 equation 4
linear error 183, 185 macrostructure 5
preparation time 132 tutorial 3
preposition 36 randomness 158
prerequisite 189, 191-4, 224-6, 229, 237 range name 25, 102, 117, 143, 165, 170, 197, 199
equation 194, 225 rate 8, 32
in a growth equation 224 recovery 160
model 224 word learning 41
present progressive 41, 42, 154 ratio 97
principle level in Fischer model 138, 229, 230 reaching 276
procedural errors 70 Recalculation 105
process reciprocal competition 195, 201, 205-7
non-linear 84 and support 194, 220-1, 225-6
non-monotonic 84 dynamics 204
of connected growth 240 model 210
smooth 84 reciprocal relationship 196
vs group studies 80 reciprocal support 195, 196, 199
prospective model 62 and competition 225-6
protein synthesis 1 intermediary transition model 197-9
prototypes of development 38 matrix 196
proximal strict transition version 197
development 262-3 reciprocity (a)symmetry 196
mechanism 64 recoverable resource 166
pseudo-regression 234 recovery
psychological research as act of function 159
communication 77 time 172
psychological test 67 reductionism and understanding 50
punctuated equilibrium 7 reflective judgement 154, 216, 218, 220, 237
oscillation 154
quadratic vs cubic model 224 reflex
qualitative changes Fischer developmental level 138
and random walks 5 mappings 138
features 273 system 138, 218
identity reasoning 209 tier 40, 215, 218, 237
invariance 90
jump 6
properties of connected growth 240 before' the leap to higher level 47
properties of data 93 in cognitive development 47
shift 12 temporary or transient 36, 38, 47, 81, 133,
similarity 92 209, 273, 277
qualitative similarity 92 regulation 245
quantitative ratio strategy 209 relationship
quasi-cycles 41 between supportive, competitive, conditional
Quattro Pro x, 105, 259 growers 189
Quetelet 78, 80, 124, 152 in Fischer model 214, 215
general principles 78-9 relaxation wave 169
reorganization
R 100, 101 of knowledge 21 1
random process 209
drift 3 repeated measure 69
fluctuation 119 representation 2 15
influence 118 = Fischer developmental level 138, 215
noise 122 mapping 138, 234
 Index
representation (continued) error 72
system 132 frequency 27 1
tier 40, 215, 218, 241 model 72
reproductive isolation 113, 218 Sanders 108
research method sandpile experiment 133
Queteletian principles 79 saturation 172, 175
resource 96, 97, 99-101, 114, 115, 139, 159, 165, Saunders 112
167, 169, 170, 177, 179, 180, 194, 211, 222, Saupe 106
234, 256, 266 scale invariance 91, 92, 272
adaptation 181, 184 qualitative 92
cycles 156 scenario
descriptive dimensions 107 asymptotic 35
energy and attention 156 rise-and-fall 35
exhaustible 33 Schroeder 135
external energeticlmotivational 108 selection 159
external informational 108 of attention 157
external material 108 self-esteem 51, 193, 241
external spatio-temporal 108 dynamic model 51
for mental growth 107 equation 193
internal informational 108 self-organized criticality 133
internal material 108 self-organization viii, 244
internal motivational/energetic 108 in motor patterns 276
internal spatial 107 system 134
internal temporal 108 self-similarity 130
internal vs external 107 semantic strategy 119
limitation 97, 108, 157 Sena 204, 205
model 25 sensitivity 258
oscillation 160-3, 165, 173, 175-6 test 107
positive and negative 99 to initial condition 107, 179, 184, 186
reduction 180 to random influences 269
resource allocation 156, 158, 174 sensitization 157
retrieval and repletion 176 sensorimotor action 147, 218, 229, 230
space 99 mapping 138, 229, 230
spatio-temporal, informational, system 138 229
energetic-motivational, and material 107 thinking 62
spike 170, 174 tier 215
well 159 sentence length 82
retrospective model 62 seriation 46
reversal shift 123, 170 set
reversibility 12 in Fischer's theory 40
in mental actions 10 sexual reproduction as iteration 15
Reward Distribution task 188 Sherman 157, 159
Richerson 124 short-time development 272
Richter scale 135 Siegler 112-13, 134, 139
rise-and-decay path 34 signal 71, 73
rise-and-fall scenario 35 and noise model 72
robot experiment 132 discrete 89
robust model 186 signal error 72
Rogoff 112, 243 similarity '
Rose 108, 137, 156 qualitative and quantitative 91
rule simple error 68
morphological 35 Simpson 112
syntactic 35 Sine function 153
singing
S-shaped curve 94, 103-4, 144, 145, 150, 154, and self-esteem 51
168, 215, 220, 229, 251, 260 dynamic model 51
sample 71 representations 138
sampling sensorimotor actions 138
skill 71, 90, 100, 111, 114, 240, 247, 264 stable 33, 129
arithmetical problem-solving 45 steady 33, 47
as bimodal object 28 sub-critical 139, 140
Fischer's theory 40 system 21
level 134, 270 state space 24, 32, 38, 41
measurement 71 developmental representation 29
multiple representations 27 diagram 198
theory 14 dynamic equation 56
Smith 204, 205, 276 of the development of arithmetic thinking
smooth change 83 29
smoothing technique 29 of singing and self-esteem 52-3
smoothness assumption 83 plot 199
social representation 24
Darwinism 2 trajectories 25
intelligence 71 statics 63-4
interaction and the management of the statistical
carrying capacity 255 filtering technique 73
role understanding 82 independence 68, 81
Socrates 10-1 1 measure 78
space problems 272
developmental 2 1 Queteletian principles 78
multi-dimensional descriptive 45 study of groups 80
speciation 112 statistics for behavioural growth 273
spurts 7 Stavy 209
species 218, 275 steady state see state
cognitive 113-1 5 step-wise growth see growth
evolution 2 Sternberg 113
spinning wheel 153 stochastical threshold 170
spline 42 stochastic transition equation 142
spontaneous condition 218 strategy, new discovery 139
movement 276 sudden jumps 139
spreadsheet ix-x, 3, 7, 29, 42, 102, 111, 116, Strauss 209
119, 121, 153, 160, 162, 164, 165, 175, strong transition 48, 128, 130, 133, 140
and catastrophe 48
definition 49
stability 240 structural
of a person's characteristics 44 model 229
stable state 33 possibility 96
stage 45, 62, 95, 124, 214, 240, 241 theory 214, 218, 221
and conflict mechanism 63 structure of interaction 246
character of growth 240 sub-criticality 133-5
developmental 20, 48, 137 subtraction 154, 221-4, 231, 232, 234, 237
fuzzy image 240 sudden
in Fischer model 218, 237 change 9, 137
Piagetian 37, 80, 136 insight 125
theory 37 sudden jump 126, 128, 130, 133, 140, 145, 150,
standard deviation 80 151, 270
state 45, 62 and level of measurement 130
and stage 45 flag 148-9
and transition 45 in Tower of Hanoi problem 131
bimodal 49 sudden shift 146, 270
developmental 20 support 114, 189, 194, 240
equation 50 and competition 194, 210-11, 220-1, 225
final 31, 33, 38, 43, 58 reciprocal 195, 196, 199, 206
initial 31, 53, 58, 129 supportive
of understanding conservation 128 grower 192
Piagetian model 20 relationship 222
portrait 57 survival-of-the-fittest 2
 Index
switch representation 23, 29
patterns 123 transformation into trajectory 25-6
point 269 time-delay state space diagram 273
sygrnoid shape 122 time-on-task 108
see also S-shape Tomlinson-Keasy 127
symbolic information processing 205 topological representation 56
syntactic of dynamic equations 58
development 154 tower of Hanoi 131, 132, 134, 135, 140, 144
developmental curve 41 training effect in testing 77
rule 154 trait 67, 71, 85
Syntax 115, 116 and context 71
system and observation set 72
definition 50 context-dependent 67
developmental 31 fluctuating 67, 71
dynamic viii moral 78
in Fischer's theory 138, 140, 154 property 71
psychological 78
Ta'eed 125 trajectory 23, 25, 31, 57, 210, 271
take-over phenomenon 207, 209, 210, 212 conservation 128
in language 208 cyclical 40, 43
in moral development 208 developmental ix, 18, 23, 58, 212
task semi-circular 53
domain 188, 232 of population 21
level 237 Transcript 259, 260
teaching 247, 256 transient regression see regression
and learning 247 transition ix, 45, 125-7, 130, 133-5, 139, 144,
as goal-directed activity 249 148, 151, 220, 222, 229, 230, 251, 272, 273,
Teilhard de Chardin 2 275
test 89 age 149
conservation 127-8 age of Fischer's level 137
continuous change 67 and sudden change 125
purpose 67 as. catastrophe 48, 145
reliability and stability 68 conservation 130
theory 70 description 45
training effect 77 developmental 148, 150
validity 70 fraction 47
Thatcher 108, 156 intermediate states 48
Thelen 112, 115, 275, 276 logistic 140, 149
theory building magnitude of cognitive 135
developmental 61-2 mathematical model 140
theory model 12, 94, 150-1, 207, 223
of dynamic skill development 2 14, 2 16 number of 135-6
of teaching and learning 252 point 148-9, 151
thinking quantity of 135
concrete operational 39 weak 45, 47-8
formal operational 39
sensorimotor 62 transition equation 142, 143, 191
Thomas 44 deterministic 143
threshold 169, 193, 202, 203, 224 growth pattern 144
tier 40, 212, 215, 218, 237, 241 stochastic 142
abstract 40 transitory children 128, 130, 148
representational 40 trapezoidal method 122, 160
sensorimotor 40 Trappist beer 168
time 274 turbulence 226, 272, '277
evolution 50 Turvey 276
scale 15, 135, 276 tutor 266, 264,
time series 21, 25, 28, 117, 219, 240 ' Tutorial Box
analysis 273 bifurcation diagram 105
 Index
Tutorial Box (continued) Velazquez 93
contagious attention 173 verbal intelligence 71
definition 3 Verhulst 101
deterministic 143 vessel metaphor 243
economic interest cycles 25 vocabulary 113, 114, 274
fluctuation of attention 167 Volterra 175
growth under oscillating resources 162 Vygotsky 14, 18, 64, 243, 262, 269, 276
increasing the carrying capacity in response to model 270
growth 258
logistic equation 102 Waddington 58
logistic model for Keren's data 116 wake and sleep pattern 157
manipulating feedback delay 122 Walker 108
mathematical problem solving in two walking 12, 276
dimensions 29 weak transition see transition
Mr Toad's Bank account 164 well model 159, 161
relaxation waves and learning 170 Werner 62, 64, 276
resource well 160 Willett 273
sensitivity to small initial state differences 107 Wilson 124
simple learning and teaching model 250 Winfree 157
simple model of connected growers 225 within- and between-level
simple smoothing procedure 42 connections 232
time frames and explosions 7 within- and between-level support 233
universal model for reciprocal interactions 197 Wood 154
two-word sentence 35, 36 word
Tycho Brahe 76 first utterances 117
learning of new 41
U-shaped growth 209-10 see also lexicon
uncertainty 185 working memory 11, 107, 122, 171, 206
underlying order 6 building a growth model 121
understanding of arithmetic operations 23 1 development 120
unidirectional input 244 limitation 173
unobtrusive observation 77 Wundt 157
unpredictability 184
unsupported testing 29 Zajonc 71
unutilized capacity for growth 99, 100, 232, 267 Zeeman 145, 146
zero-transformation 33
Valsiner 108, 243, 270 Zoad 263, 264
van der Maas 49, 128, 140, 147-50, 273 Zone
van der Veer 270 of Actual Development 263
van Geert 20, 45, 62, 96, 104, 115, 120, 124, of Free Movement 108
129, 163, 222, 241, 254, 262, 267, 276 Zone of Proximal Development 19, 243, 262,
variable 263, 267
continuous nature 90 growth patterns 267
person 87 mathematical simulation 266
problem 87 model assumptions 266
Veblen xi Zoped 263, 264