What Is Complexity?

Murray Gell-Mann

Abstract. It would require many different concepts to capture all our notions of
the meaning of complexity. The concept that comes closest to what we usually
mean is effective complexity (EC). Roughly speaking, the EC of an entity is the
length of a very concise description of its regularities. A novel is considered
complex if it has a great many scenes, subplots, characters, and so forth. An
elaborate hierarchy can contribute to complexity, as in the case of nested industrial
clusters each composed of a great variety of firms and other institutions. In
general, though, what are regularities? We encounter in many different situations
the interplay between the regular and the random or incidental: music and static on
the radio, specifications and tolerances in manufacturing, etc. But ultimately the
distinction between the regular and the incidental depends on a judgment of what
is important, although the judge need not be human or even alive. For instance, in
the case of songs of a male bird in the nesting season, the identification of
regularities is perhaps best left to the other birds of the same species - what
features are essential in repelling other males from the territory or attracting a
suitable female'? A technical definition of EC involves the quantity called
algorithmic information content (AIC). The description of an entity is converted to
a bit string and a standard universal computer is programmed to print out that
string and then halt. The length of the shortest such program (or, in a
generalization, the shortest that executes within a given time) is the AIC. The AIC
is expressed as the sum of two terms, one (the EC) referring to the regularities and
the other to the random features. The regularities of a real entity are best expressed
by embedding it conceptually in a set of comparable things, the rest of which are
imagined. The EC can then be related to the AIC of the set, the choice of which is
restricted by the conditions imposed by the judge. Theorists like to study highly
simplified models of complex systems, often by computer modelling. What can be
claimed for such models? Overall agreement with observation is hardly to be
expected. However, in many cases simple regularities can be found in both the
observational data and the model, which may then be helpful in understanding
those regularities. Examples are given, involving scaling laws and also
"implicational scales".

It would take a great many different concepts to cover all our intuitive notions of
what is meant by complexity (and its opposite, simplicity), but the concept that
agrees best with the meaning in ordinary conversation and in most scientific
discourse is effective complexity (EC). Roughly, the EC of an entity is the length

A. Q. Curzio et al. (eds.), Complexity and Industrial Clusters

© Physica-Verlag Heidelberg 2002
14 M. Gell-Mann

of a very con..:ise description of its regularities. Thus we would call a novel

complex if it had many different subpluts, scenes, and characters. Likewise, an
international conglomerate firm would be complex if it had many different
branches in different countries with different products, management styles, and so
forth. A hierarchical structure that takes a long time to describe would also
cuntribute to complexity. It is easy to see that an industrial cluster comprising
many different firms and other institutions in diverse locations within a region,
perhaps grouped into sub clusters (which may, in turn, be composed of sub-sub
clusters) is complex in this sense.
An amusing exercise consists of looking at the patterns of neckties to see which
ones are simple (for example those with regimental stripes) and which ones are
complex (for instance most hand-painted ones). Note that we are looking at how
long it takes to describe the regularities, not the features treated as random or
incidental. Of course the description length will typically depend on the "coarse
graining" (level of detail), which depends in turn on how far away the tie is when
viewed with the unaided eye.
We are constantly observing in the world around us the interplay between
regularity and randomness. When we hear music and static on the radio, we
identify the music as regular and the static as random noise, and we try to make
the signal-to-noise ratio as high as we can. In cooking, we may use a recipe but
there are always some variations in the quantities of ingredients, in cooking times
and temperatures, etc. In manufacturing, we encounter specifications and
tolerances.
But the distinction between the regular and the random is not absolute. Many
decades ago scientists at the Bell Telephone Laboratories investigated the origins
of static and found that some of it comes from sources in particular places among
the constellations in the sky. That was the origin of radio astronomy, and hence a
permanent association of some static with important regularities. When we looked
at neckties, we were concerned only with the pattern, and we neglected po~sible
wine or food stains, but to a dry-cleaner the stains might be the significant
regularities, while the patterns are of little concern. We conclude, then, that what
is a regularity depends on a judgment of what is important and what is not. Of
course the judge need not be human, or even alive.
Suppose we are concerned with the regularities in the repeated, but somewhat
variable territorial songs of a male bird of a particular species in the nesting
season. The other birds of the same species would make a suitable judge. Their
behaviour can reveal what features of the song serve to repel other males from the
territory or perhaps to attract a suitable female.
The crude definition of effective complexity given above in terms of minimum
description length needs to be refined and made more technical. First, one should
recognize that the effective complexity is necessarily context-dependent in a
number of ways. To begin with, the entity in question is described at some level of
coarse graining and in some language, and so we have immediately two sources of
context dependence. It is also true that some degree of knowledge and
understanding of the world is assumed. To illustrate the importance of this point,
imagine that you are entering the village of a group of hitherto uncontacted
 What Is Complexity~ IS

Indians in the Brazilian jungle, and that you know their language from experience
with another group speaking the same tongue. Now explain to your hosts the
meaning of a tax-managed mutual fund. It should take quite a while.
Of course as we get more technical we must think in terms of a computer. The
description of the entity is coded into a string of bits (zeroes and ones) according
to some coding scheme, and then a particular standard universal computer is
employed (we have here two more sources of context-dependence). The
algorithmic information content (AIC) of the description is then defined as the
length of the shortest program that will cause the computer to print out the bit
string in question and then stop computing. In fact, we shall generalize that
definition by considering a variable time T and asking for the length of the shortest
program that will perform this task within time T. The AIC is the limit of this
quantity as T goes to infinity.
In discussing K, the AIC of the entity, or the corresponding quantity K(T) for
finite values of T, we are not yet dealing with the effective complexity. The EC
has to do with the length of a very brief description of the regularities of the
entity, not the features treated as random or incidental. Thus we need to exhibit the
AIC (or its generalization to time 7) as the sum of two terms, one applying to the
regularities and the other to the features treated as random.
The best way to represent particular regularities of an entity is to embed that
entity conceptually in a set of comparable things. all the rest of which are
imagined, and to assign probabilities or weights to the members of the set. Such a
set, with probabilities for the members, is called an ensemble, and it embodies the
regularities in question.
The probabilities allow us to define a measure I of "Shannon information" or
ignorance for the ensemble (These two apparent opposites are really two facets of
the same thing: the amount of ignorance about the contents of a letter before one
receives it is the same as the amount of information learned when it is received
and read).
Since the ensemble serves to describe the regularities of our entity, the AIC of
the ensemble (call it Y) is a candidate for the effective complexity. We just have to
find the right ensemble. It turns out that the prescription for that is a
straightforward one:
- First, minimize the "total information" Y+/. The minimum value can easily
be shown to be about equal to K, the AIC of the entity. Thus K is exhibited as
the sum of two terms referring to the regular and random features
respectively. That is just what we needed.
- Second, following a famous principle, maximize the ignorance / or - what is
the same thing - minimize Y, subject to the conditions imposed by the judge.
The imposition of those conditions, based on a judgment of what is
important, is an essential part of the prescription. One can show that
otherwise Y would always be very small and the ignorance / would
correspondingly use up almost all of K. Every entity would be simple.
In the arts, we are accustomed to the use of the imagined alongside the real, as
in fiction and drama, and we are not surprised that light is often thrown in that
way on the regularities of the real.
16 M. Gell-Mann

The break-up of K - the AIC of the entity - into the effective complexity Yand
the ignorance I can be conceived in terms of a basic program and a lot of data to
be fed into that program. The effective complexity is related to the length of the
basic program, while the ignorance measures the information content of the
additional data needed to specify the entity with which we are dealing.
The famous computer scientist, psychologist, and economist Herbert Simon
used to call attention to the path of an ant, which has a high AIC and appears
complex at first sight. But when we realize that the ant is following a rather simple
program, into which are fed the incidental features of the landscape and the
pheromone trails laid down by the other ants for the transport of food, we
understand that the path is fundamentally not very complex. Herb said, "I got a lot
of mileage out of that ant". And now it is helping me to emphasize the difference
between total AIC and effective complexity.
A significant advantage of generalizing the measures K and Y to K(T) and Y(T)
is that we can then discuss situations such as the following:
When I was a graduate student, more than fifty years ago, we wondered what
kind of dynamics underlay the structure of atomic nuclei. It appeared that the laws
that govern the energy levels of nuclei might be extremely complex. Today,
however, we have the theory of quantum chromodynamics, the field theory of
quarks and gluons, supplemented by quantum electrodynamics, which governs the
electromagnetic interaction. We now believe those simple theories are correct (and
many of their predictions have been verified by experiment). They must yield the
energy levels of nuclei to an excellent approximation, but the calculations are so
elaborate that even today's machines and techniques can't handle them properly.
The standard universal computer would take a very long time to execute the short
program corresponding to these simple theories and thus calculate the energy
levels. If we insist on execution in a much shorter time T, then the program would
have to be exceedingly long. So, the energy levels of nuclei have very little
effective complexity Y in the limit of T going to infinity, but huge values of Y if T
is taken to be small.
There can be no finite procedure for finding all the possible regularities of an
arbitrary entity. We may ask, then, what kinds of things engage in identifying and
using particular sets of regularities. The answer is complex adaptive systems,
including all living organisms on Earth.
A complex adaptive system receives a stream of data about itself and its
surroundings. In that stream, it identifies certain regularities and compresses them
into a concise "schema", one of many possible schemata related by mutation or
substitution. In the presence of further data from the stream, the schema can
supply descriptions of certain aspects of the real world, predictions of events that
are to happen in the real world, and prescriptions for behaviour of the complex
adaptive system in the real world. In all these cases, there are real world
consequences: the descriptions can turn out to be more accurate or less accurate,
the predictions can turn out to be more reliable or less reliable, and the
prescriptions for behaviour can turn out to lead to favourable or unfavourable
outcomes. All these consequences then feed back to exert "selection pressures" on
the competition among various schemata, so that there is a strong tendency for
 What Is Complexity? 17

more successful schemata to survive and for less successful ones to disappear or at
least to be demoted in some sense.
Take the human scientific enterprise as an example. The schemata are theories.
A theory in science compresses into a brief law (say a set of equations) the
regularities in a vast, even indefinitely large body of data. Maxwell's equations,
for instance, yield the electric and magnetic fields in any region of the universe if
the special circumstances there - electric charges and currents and boundary
conditions - are specified (we see how the schema plus additional information
from the data stream leads to a description or prediction).
In biological evolution, the schemata are genotypes. The genotype, together
with all the additional information supplied by the process of development - for
higher animals, from the sperm and egg to the adult organism - determines the
character, the "phenotype" of the individual adult. Survival to adulthood of that
individual, sexual selection, and success or failure in producing surviving progeny
all exert selection pressures on the competition of genotypes, since they affect the
transmission to future generations of genotypes resembling that of the individual
in question.
In the case of societal evolution, the schemata consist of laws, customs, myths,
traditions, and so forth. The pieces of such a schema are often called "memes", a
term introduced by Richard Dawkins by analogy with genes in the case of
biological evolution.
For a business firm, strategies and practices form the schemata. In the presence
of day-to-day events, a schema affects the success of the firm, as measured by
return to the stockholders in the form of dividends and share prices. The results
feed back to affect whether the schema is retained or a different one substituted
(often under a new CEO).
A complex adaptive system (CAS) may be an integral part of another CAS and
it may, in turn, contain smaller complex adaptive systems. In fact, a CAS has a
tendency to give rise to others.
On Earth, all complex adaptive systems seem to have some connection with life.
To begin with, there was the set of prebiotic chemical reactions that gave rise to
the earliest life. Then the process of biological evolution, as we have indicated, is
an example of a CAS. Likewise each living organism, insofar as it can learn, is a
CAS. In a mammal, such as a human being, the immune system is a complex
adaptive system too. Its operation is something like that of biological evolution,
but on a much faster time scale (if it took hundreds of thousands of years for us to
develop antibodies to invading microbes, we would be in serious trouble). The
process of learning and thinking in a human individual is also a complex adaptive
system. In fact, the term "schema" is taken from psychology, where it refers to a
pattern used by the mind to grasp an aspect of reality. Aggregations of human
beings can also be complex adaptive systems, as we have seen: societies, business
firms, the scientific enterprise, and so forth.
Nowadays, we have computer-based complex adaptive systems, such as "neural
nets" and "genetic algorithms". While they may sometimes involve new,
dedicated hardware, they are usually implemented on conventional hardware with
special software. Their only connection with life is that they were developed by
18 M. Gell-Mann

human beings, such as that outstanding human John Holland. Once they are put
into operation, they can, for example, invent new strJ.tegies for winning at games
that no person has ever discovered.
It is probably helpful to mention at this point that John Holland and I use
somewhat different terminology. In fact, we both heard at a meeting some years
ago that a scientist would rather use someone else's toothbrush than another
scientist's nomenclature. What I call a schema is called an internal model by John
Holland, and what I call a complex adaptive system is close to what he calls an
adaptive agent. He uses the expression "complex adaptive system" to mean a
loose aggregation of complex adaptive systems adapting to one another's
behaviour, for example investors in a market or organisms in an ecological
system.
According to my system of nomenclature, an industrial district would have to
have a schema of its own to qualify as a complex adaptive system. From what
little I know, I would say that may not generally be the case. Using John Holland's
meaning of complex adaptive system, we would ask whether the component
institutions (such as firms or universities) have schemata, and in general they do.
In my notation, then, an industrial district mayor may not be a complex adaptive
system itself, but it certainly is an aggregation of complex adaptive systems,
adapting to one another and to external int1uences.
Science fiction writers and others may speculate that in the distant future a new
kind of complex adaptive system might be created, a truly composite human
being, by wiring together the brains of a number of people. They would
communicate not through language, which - Voltaire is supposed to have said - is
used by men to conceal their thoughts, but instead through sharing all their mental
processes. My friend Shirley Hufstedler says she would not recommend this
procedure to couples about to be married.
The behaviour of a complex adaptive system, with its variable schemata
undergoing evolution through selection pressures from the real world, may be
contrasted with "simple" or "direct" adaptation, which does not involve a variable
schema, but utilizes instead a fixed pattern of response to external changes. A
good example of direct adaptation is the operation of a thermostat, which simply
turns on the heat when the temperature falls below a fixed value and turns it off
when the temperature rises above the same value. It just keeps mumbling "It's too
hot, it's too hot, it's just right, it's too cold", and so forth.
In the study of a human organization, such as a tribal society or a business firm,
one may encounter at least three different levels of adaptation, on three different
time scales.
1) On a short time scale, we may see a prevailing schema prescribing that the
organization react to particular external changes in specified ways. As long as
that schema is fixed, we are dealing with direct adaptation.
2) On a longer time scale, the real world consequences of a prevailing schema
exert (in the presence of events that occur) selection pressures on the
competition of schemata and may result in the replacement of one schema by
another.
 What Is Complexity') 19

3) On a still longer time scale, we may witness the disappearance of some

organizations and the survival of others, in a Darwinian process. The
evolution of schemata was inadequate in the former cases, but adequate in the
latter cases, to cope with the changes in circumstances.
It is worth making the elementary point about the existence of these levels of
adaptation because they are often confused with one another. As an example of the
three levels, we might consider a prehistoric society in the U.S. Southwest that had
the custom (1) of moving to higher elevations in times of unusual heat and
drought. In the event of failure of this pattern, the society might try alternative
schemata (2) such as planting different crops or constructing an irrigation system
using water from far away. In the event of failure of all the schemata that are tried,
the society may disappear (3), say with some members dying and the rest
dispersed among other societies that survive. We see that in many cases failure to
cope can be viewed in terms of the evolutionary process not being able to keep
pace with change.
Individual human beings in a large organization or society must be treated by
the historical sciences as playing a dual role. To some extent they can be regarded
statistically, as units in a system. But in many cases a particular person must be
treated as an individual, with a personal intluence on history. Those historians
who tolerate discussion of contingent history (meaning counterfactual histories in
addition to the history we experience) have long argued about the extent to which
broad historical forces eventually "heal" many of the changes caused by
individual achievements - including negative ones, such as assassinations.
A history of the U.S. Constitutional Convention of 1787 may make much of the
contlicting interests of small states and large states, slave states and free states,
debtors and creditors, agricultural and urban populations, and so forth. But the
compromises invented by particular individuals and the role that such individuals
played in the eventual ratification of the Constitution would also be stressed. The
outcome could have been different if certain particular people had died in an
epidemic just before the Convention, even though the big issues would have been
the same.
How do we think about alternative histories in science? Is the notion of
alternative histories a fundamental concept?
The fundamental laws of nature are:
(1) the dynamical law of the elementary particles - the building blocks of all
matter - along with their interactions and
(2) the initial condition of the universe near the beginning of its expansion some
ten billion years ago.
Theoretical physicists seem to be approaching a real understanding of the first
of these laws, as well as gaining some inklings about the second one. It may well
be that both are rather simple and knowable, but even if we learn what they are,
that would not permit us, even in principle, to calculate the history of the universe.
The reason is that fundamental theory is probabilistic in character (contrary to
what one might have thought a century ago). The theory, even if perfectly known,
predicts not one history of the universe but probabilities for a huge array of
alternative histories, which we may conceive as forming a branching tree, with
20 M. Gell-Mann

probabilities at all the branchings. In a short story by the great Argentine writer
Jorge Luis Borges, a character creates a model of these branching histories in the
form of a garden of forking paths.
The particular history we experience is co-determined, then, by the fundamental
laws and by an inconceivably long sequence of chance events, each of which
could turn out in various ways. This fundamental indeterminacy is exacerbated for
any observer - or set of observers, such as the human race - by ignorance of the
outcomes of most of the chance events that have already occurred, since only a
very limited set of observations is available. Any observer sees only an extremely
coarse-grained history.
The phenomenon of chaos in certain non-linear systems is a very sensitive
dependence of the outcome of a process on tiny details of what happened earlier.
When chaos is present, it still further amplifies the indeterminacy we have been
discussing.
Some years ago, at the science museum in Barcelona, I saw an exhibit that
beautifully illustrated chaos. A non-linear version of a pendulum was set up in
such a way that the visitor could hold the bob and start it out in a chosen position
and with a chosen velocity. One could then watch the subsequent motion, which
was also recorded with a pen on a sheet of paper. The visitor was then invited to
seize the bob again and try to imitate exactly the previous initial position and
velocity. No matter how carefully that was done, the subsequent motion was quite
different from what it was the first time. Comparing the records on paper
confirmed the difference in a striking way.
I asked the museum director what the two men were doing who were standing in
a corner watching us. He replied, "Oh, those are two Dutchmen waiting to take
away the chaos". Apparently, the exhibit was about to be dismantled and taken to
Amsterdam. But I have wondered ever since whether the services of those two
Dutchmen would not be in great demand across the globe, by organizations that
wanted their chaos taken away.
Once we view alternative histories as forming a branching tree, with the history
we experience co-determined by the fundamental laws and a huge number of
accidents, we can ponder the accidents that gave rise to the people assembled at
this conference. A fluctuation many billions of years ago produced our galaxy, and
it was followed by the accidents that contributed to the formation of the solar
system, including the planet Earth. Then there were the accidents that led to the
appearance of the first life on this planet, and the very many additional accidents
that, along with natural selection, have shaped the course of biological evolution,
including the characteristics of our own subspecies, which we call, somewhat
optimistically, Homo sapiens sapiens. Finally we may consider the accidents of
genetics and sexual selection that helped to produce the genotypes of all the
individuals here, and the accidents in the womb, in childhood, and since that have
helped to make us what we are today.
Now most accidents in the history of the universe don't make much difference to
the coarse-grained histories with which we are concerned. If two oxygen
molecules in the atmosphere collide and then go off in one pair of opposite
directions or another, it usually makes little difference. But the fluctuation that
 What Is Complexity') 21

produced our galaxy, while it too may have been insignificant on a cosmic scale,
was of enormous importance to anything in our galaxy. Some of us call such a
chance event a "frozen accident". It produces substantial effects, if only in a
limited region of space and time. Once it happens and its outcome is determined,
an accident no longer belongs to the realm of the random. Frozen accidents,
affecting significantly things judged to be important, give rise to regularities. Of
course the fundamental laws also contribute to regularities, but those laws are
thought to be simple, and so effective complexity comes mainly from frozen
accidents.
We can take an example from modern human history to illustrate the idea of a
frozen accident. It was mentioned earlier that a few historians are willing to
discuss "contingent history", in which one asks "What if ..... '1". One of their
favourite incidents occurred when Buffalo Bill's Wild West Show was touring
Europe in 1889. Of course a star attraction of the show was the female
sharpshooter Annie Oakley, who would ask for a male volunteer so that she could
knock the ash off his cigar with a bullet. Normally there were no volunteers, and
her husband, himself a famous marksman, would step forward and stand still
while Annie shot, hitting the ash but not her husband. On this occasion, however,
there was a volunteer, the Kaiser - Wilhelm der Zweite - with an expensive
Havana cigar. Annie was a little worried, having drunk heavily the night before,
but she fired and we know the result. The Kaiser survived, fired Bismarck,
cancelled the reinsurance treaty with Russia, engaged in competition in naval
construction with Great Britain, and in other ways laid the groundwork for the
First World War.
The dominant role played by frozen accidents in generating effective
complexity can help us to understand the tendency for more and more complex
entities to appear as time goes on. Of course there is no rule that everything must
increase in complexity. Any individual entity may increase or decrease in effective
complexity or stay the same. When an organism dies or a civilization dies out, it
suffers a dramatic decrease in complexity. But the envelope of effective
complexity keeps getting pushed out, as more and more complex things arise.
The reason is that as time goes on the results of frozen accidents can keep
accumulating. If that process outstrips the erasure of the consequences of frozen
accidents, then more and more effective complexity can arise. That is so even for
non-adaptive evolution, as in galaxies, stars, planets, rocks, and so forth. It is well
known to be true of biological evolution, where in some cases higher effective
complexity probably confers an advantage. And we see all around us the
appearance of more and more complex regulations, instruments, computer
software packages, and so forth, even though in many cases certain things are
simplified.
The tendency of more and more complex forms to appear in no way contradicts
the famous second law of thermodynamics, which states that for a closed
(isolated) system, the average disorder ("entropy") keeps increasing. There is
nothing in the second law to prevent local order from increasing, through various
mechanisms of self-organization, at the expense of greater disorder elsewhere.
(one simple and widespread mechanism of self-organization on a cosmic scale is
22 M. Gell-Mann

provided by gravitation, which has caused material to condense into the familiar
structures with which astronomy is concerned, including our own planet).
Here on Earth, once it was formed, systems of increasing complexity have
arisen as a consequence of the physical evolution of the planet over some four and
half billion years. biological evolution over nearly four billion years, and - over a
very short period on a geological time scale - human cultural evolution.
The process has gone so far that we human beings are now confronted with
immensely complex problems and we are in urgent need of better ways of dealing
with them (the role of industrial clusters and their potential constitute, of course,
an example of a complex issue). When we attempt to tackle such difficult
problems, we naturally tend to break them up into more manageable pieces. That
is a useful practice, but it has serious limitations.
When dealing with any non-linear system, especially a complex one, it is not
sufficient to think of the system in terms of parts or aspects identified in advance,
then to analyse those parts or aspects separately, and finally to combine those
analyses in an attempt to describe the entire system. Such an approach is not, by
itself, a successful way to understand the behaviour of the system. In this sense
there is truth in the old adage that the whole is more than the sum of its parts.
Unfortunately, in a great many places in our society, including academia and
most bureaucracies, prestige accrues principally to those who study carefully some
aspect of a problem, while discussion of the big picture is relegated to cocktail
parties. It is of crucial importance that we learn to supplement those specialized
studies with what I call a crude look at the whole.
Now the chief of an organization, say a head of government or a CEO, has to
behave as if he or she is taking into account all the aspects of a situation, including
the interactions among them, which are often strong. It is not so easy, however, for
the chief to take a crude look at the whole if everyone else in the organization is
concerned only with a partial view.
Even if some people are assigned to look at the big picture, it doesn't always
work out. Some time ago, the CEO of a gigantic corporation told me that he had a
strategic planning staff to help him think about the future of the business, but that
the members of that staff suffered from three defects:
1) They seemed largely disconnected from the rest of the company.
2) No one could understand what they said.
3) Everyone else in the company seemed to hate them.
Despite such experiences, it is vitally important that we supplement our
specialized studies with serious attempts to take a crude look at the whole.
I am sure that observation applies to industrial clusters.
At many research institutions, theorists are busily engaged in constructing
simple models of quite complex real systems. Today, those are usually computer
models. How, though, does one rate the success of such a model in helping us
learn about the real world? Normally, a theory is compared to observation and the
value of the theory is judged by the degree of agreement with observed facts,
along with criteria such as elegance, consistency, and compatibility with well-
established principles. However, when we are dealing with a highly simplified
 What Is Complexity? 23

model of a very complex system, detailed agreement with observation could well
be embarrassing rather than encouraging.
At the Santa Fe Institute, we have had a number of discussions of this question:
what can be claimed on the basis of a simplified model? One outcome of those
discussions is the conclusion that one should look for rather simple regularities in
the data and see if the model predicts some or all of them. This semi-empirical
"middle level theory" can be enormously valuable. Often, the regularities in
question will hold along the whole spectrum from the complex real world down
through successive simplifications of that world to the model being examined. If
that is the case, then the reason for the regularity in the simple model may be easy
to understand and it can throw light on how the regularity arises in the real world.
Examples of such simple regularities certainly include scaling laws, in which
one quantity of interest appears as a power of another such quantity. These laws
crop up in all fields of science: physical, biological, and behavioural. A striking
example is the rule, verified with increasing confidence over most of a century,
but unexplained until recently, that the metabolic rate of a mammal species goes
like the three-quarters power of the mass. The mammal in question can range from
a tiny shrew to an elephant or a whale. If one mammal weighs ten thousand times
as much as another, it will process energy at a rate one thousand times higher.
Similar laws (with the same power but different coefficients) apply to birds and to
higher plants.
This law was explained by Geoffrey West, a theoretical physicist at the Los
Alamos National Laboratory, and James Brown, an ecology professor at the
University of New Mexico in Albuquerque, along with his student Brian Enquist.
These researchers met and collaborated at the Santa Fe Institute, in a perfect
example of how that institute is supposed to function.
They modelled (analytically rather than on a computer) the cardiovascular and
respiratory systems of the animals and the vascular system of the plants. They
based their model on three-dimensional fractals, and they used two facts:
1) in the branchings of the tubes, areas are preserved and
2) the thinnest tubes all have about the same diameter, independent of the mass.
The scaling law then followed, with the fraction 3/4 arising from the number of
spatial dimensions divided by that number plus one.

Sometimes scaling laws are laws of distribution, for example the famous Zipf's
law - found in fact by Auerbach in 1913 - for the populations of "cities" (actually
metropolitan areas). It states that if we rank the cities in order of decreasing
population, then the population is roughly inversely proportional to the numerical
rank. The data are a bit ragged for the very largest cities, but they agree quite well
with the law for the next several hundred cities. Thus the 300 lh city has about one-
third the population of the 100th city.
Recently Xavier Gabaix published an explanation of Zipf's law starting from
Gibrat's law, which is known to be true and states that the mean growth rate and
mean square deviation of growth rates for metropolitan areas are independent of
population to a good approximation. Gabaix showed that the steady state
distribution of populations then obeys Zipf's law.
24 M. Gell-Mann

Besides scaling laws, we can look at other examples of simple middle level
semi-empirical laws, for instance implicational ladders. Around 1940, when the
USA was still not a belligerent in the Second World War, some sociologists were
asked to measure the attitudes of Americans toward Great Britain. They
formulated a long list of questions with yes or no answers, and those questions
were posed to a representative sample of American citizens. The results showed a
fascinating regularity. If the questions were arranged in a certain order, then most
of the returned questionnaires showed responses that were affirmative down to a
certain point and then negative after that. Thus most of the sets of answers to the
questions could be characterized by a single parameter, the point in the list where
the answers switched from positive to negative. As far as the questionnaire went,
attitude toward the United Kingdom could be summarized in a single number.
Such a result is often called an implicational scale (not to be mixed up with
scaling laws, as described earlier). To avoid confusion, I like to call it an
implicational ladder.
Another example can be cited from linguistics. There are some countries, such
as Haiti and Jamaica, where a Creole language is spoken that gets its words
mainly from the official language of the country. In Haiti, nearly everyone speaks
a French-based Creole language at home, while the official language is, or was
until recently, French. Likewise, in Jamaica, the official language is English bUl
many people speak an English-based Creole at home (this is in contrast to the
situation on the island of St. Lucia, for example, where the official language is
English but the Creole is French-based).
In 1961, the linguist DeCamp investigated the situation in Jamaica, where,
under the influence of official English in the schools, there is a tendency for
certain people to substitute usages from regular English for those in the broadest
or deepest Creole. He chose six cases of such usages, four involving vocabulary
and two involving pronunciation. He discovered that those usages could be
arranged in a particular order such that some speakers used the standard English
version in all six cases, some in the first five, some in the first four, some in the
first three, and so forth. Other patterns were negligibly rare. Thus, out of all the 64
possible patterns for choices in the six cases between Creole usage and Standard
English usage, only 7 patterns existed with appreciable frequency among the
population, and those seven formed an implicational ladder.
If I had to model industrial clusters, I would search carefully for "middle-level"
semi-empirical rules that might persist all the way from the real world down to the
highly simplified model, and test the model by seeing if it clarifies how the rules
arise.