Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)

Nature’s Patterns
This page intentionally left blank

Nature’s
Patterns
A Tapestry in Three Parts
Philip Ball
Nature’s Patterns is a trilogy composed of
Shapes, Flow, and Branches
1
3
Great Clarendon Street, Oxford OX2 6DP
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide in
Oxford New York
Auckland Cape Town Dar es Salaam Hong Kong Karachi
Kuala Lumpur Madrid Melbourne Mexico City Nairobi
New Delhi Shanghai Taipei Toronto
With offices in
Argentina Austria Brazil Chile Czech Republic France Greece
Guatemala Hungary Italy Japan Poland Portugal Singapore
South Korea Switzerland Thailand Turkey Ukraine Vietnam
Oxford is a registered trade mark of Oxford University Press
in the UK and in certain other countries
Published in the United States
by Oxford University Press Inc., New York
# Philip Ball 2009
The moral rights of the author have been asserted
Database right Oxford University Press (maker)
First published 2009
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
without the prior permission in writing of Oxford University Press,
or as expressly permitted by law, or under terms agreed with the appropriate
reprographics rights organization. Enquiries concerning reproduction
outside the scope of the above should be sent to the Rights Department,
Oxford University Press, at the address above
You must not circulate this book in any other binding or cover
and you must impose the same condition on any acquirer
British Library Cataloguing in Publication Data
Data available
Library of Congress Cataloging in Publication Data
Data available
Typeset by SPI Publisher Services, Pondicherry, India
Printed in Great Britain
on acid-free paper by
Clays Ltd., St Ives plc
ISBN 978–0–19–923798–2
1 3 5 7 9 10 8 6 4 2
Branching
S
EEING the splayed, forking channels of rivers, natural philosophers
were reminded of veins and arteries. These in turn speak of
trees—and why not, for all are networks that distribute vital fluids.
What are the rules that make branches? Why does a lightning bolt seek
many routes from heaven to earth, or a crack begin to wander and
divide? Branching forms find a compromise between disorder and
determinism: they hint at a new and peculiar geometry. Yet sometimes
order reasserts itself, as when the arms of snowflakes insist on their
hexagonality. And if branches reunite, the loops form a web that offers
many routes to the same destination. Navigation and dissemination on
such networks then depends on the pattern of connections.
Contents
Preface and acknowledgements ix
1: A Winter’s Tale 1
The Six-Pointed Snowflake
2: Tenuous Monsters 27
Shapes between Dimensions
3: Just For the Crack 71

Clean Breaks and Ragged Ruptures
4: Water Ways 101

Labyrinths in the Landscape
5: Tree and Leaf 131

Branches in Biology
6: Web Worlds 150

Why We’re All in This Together
Epilogue: The Threads of the Tapestry 180

Principles of Pattern
Appendix 210
Bibliography 212
Index 217
Preface and
acknowledgements
A
F T E R my 1999 book The Self-Made Tapestry: Pattern Formation
in Nature went out of print, I’d often be contacted by would-be
readers asking where they could get hold of a copy. That was
how I discovered that copies were changing hands in the used-book
market for considerably more than the original cover price. While that
was gratifying in its way, I would far rather see the material accessible to
anyone who wanted it. So I approached Latha Menon at Oxford Uni-
versity Press to ask about a reprinting. But Latha had something more
substantial in mind, and that is how this new trilogy came into being.
Quite rightly, Latha perceived that the original Tapestry was neither
conceived nor packaged to the best advantage of the material. I hope
this format does it more justice.
The suggestion of partitioning the material between three volumes
sounded challenging at first, but once I saw how it might be done,
I realized that this offered a structure that could bring more thematic
organization to the topic. Each volume is self-contained and does not
depend on one having read the others, although there is inevitably
some cross-referencing. Anyone who has seen The Self-Made Tapestry
will find some familiar things here, but also plenty that is new. In
adding that material, I have benefited from the great generosity of
many scientists who have given images, reprints and suggestions.
I am particularly grateful to Sean Carroll, Iain Couzin, and Andrea
Rinaldo for critical readings of some of the new text. Latha set me
more work than I’d perhaps anticipated, but I remain deeply indebted
to her for her vision of what these books might become, and her
encouragement in making that happen.
Philip Ball
London, October 2007
A Winter’s Tale
The Six-Pointed Snowflake
1
T
HE followers of Pythagoras believed many strange things, among
them that one should not eat beans or break bread, should not
pluck a garland and should not allow swallows to land on one’s
roof. They sound like a bunch of crackpot mystics, but in fact Pytha-
goreanism has, through its influence on Plato, provided a recurrent
theme in Western rationalist thought: the idea that the universe is
fundamentally geometric, so that all natural phenomena display a
harmony based on number and regularity. Pythagoras is said to have
discovered the relationship between proportion and musical harmony,
reflected in the way that a plucked string divided by simple length ratios
produces pleasing musical intervals. The ‘music of the spheres’—
celestial harmonies generated by the heavenly bodies according to
the sizes of their orbits—is ultimately a Pythagorean concept.
‘All things are numbers’, said Pythagoras, but it is not easy now to
comprehend what he meant by this statement. In some fashion, he
believed that integers were building blocks from which the world was
constructed. Bertrand Russell is probably imposing too modern a per-
spective when he interprets the phrase as saying that the world is ‘built
up of molecules composed of atoms arranged in various shapes’, even
if, for Plato, those atoms themselves were geometric: cubes, tetrahedra,
and other regular shapes that, he said, account for the empirical prop-
erties of the corresponding classical elements. All the same, it seems
fair to suppose that a Pythagorean would have been less surprised than
we are to find spontaneous regularity of pattern and form in the
world—five-petalled flowers, faceted crystals—because he would have
envisaged this orderliness to be engraved in the very fabric of creation.
The ancient Greeks were not alone in thinking this way. Chinese
scholars of long ago were as devoted to the study of nature and
2 j NATURE’S PATTERNS: BRANCHES
mathematics as any of their Western counterparts, and by all appear-

ances they were rather more observant. It was not until the European
Middle Ages that the Aristotelian tradition of studying specific natural
phenomena for their own sake began to permeate the Western world,
prompting the thirteenth-century Bavarian proto-scientist Albertus
Magnus to record the ‘star-shaped form’ of snowflakes, which can be
seen with the naked eye. But the Chinese anticipated him by more than
a millennium. Around 135 BC, the philosopher Han Ying wrote in his
treatise Moral Discourses Illustrating the Han Text of the ‘Book of Songs’
that ‘Flowers of plants and trees are generally five-pointed, but those of
snow, which are called ying, are always six-pointed.’ It is a casual
reference, as though he is mentioning something that everyone already
knew.
Chinese poets and writers in the subsequent centuries took this fact
for granted. In the sixth century AD, Hsiao T’ung wrote:
The ruddy clouds float in the four quarters of the caerulean sky
And the white snowflakes show forth their six-petalled flowers.
By the seventeenth century, Chinese scholars had become more sys-

tematic and scientific in their approach. ‘Every year at the end of winter
and the beginning of spring I used to collect snow crystals myself and
carefully examined them’, wrote Hsien Tsai-hang in his Five Assorted
Offering Trays (c.1600). He may have used a magnifying glass for this
work, which led him to conclude that ‘all were six-pointed’.
It was no surprise to the Chinese sages that snow crystals were six-
pointed, because many of them held a view of nature that was every bit
as numerological as that of the Pythagoreans. Still today, numerical
schemes provide a central ordering principle in Chinese thought,
from the Eightfold Way of Daoism to the ‘Four Greats’ of Mao’s person-
ality cult. In a system of ‘correspondences’ analogous to that of the
Western mystical tradition, the elements were deemed to have num-
bers associated with them, and as the great philosopher Chu Hsi wrote
in the twelfth century, ‘Six generated from Earth is the perfected num-
ber of Water.’ Thus, according to the scholar T’ang Chin, ‘when water
congeals into flowers they must be six-pointed’, because ‘six is the true
number of Water’.
The problem with this scheme is that it stifles further enquiry: given
such an ‘explanation’ (which we now see as little more than a tautol-
ogy), there is nothing more to be said. A profound mystery is reduced to
A WINTER’S TALE j 3
a commonplace fact. And so, in the words of sinologist Joseph

Needham, ‘the Chinese, having found the hexagonal symmetry [of
snowflakes], were content to accept it as a fact of nature’.
Here, then, is a rejoinder to the accusation that a scientific attitude is
prone to blunt our wonderment at the world. In the mystic’s teleo-
logical universe, order and pattern are only to be expected: they are part
of the Grand Design. There is nevertheless value in such an outlook,
which can help to bring to our notice the regularities that exist in
nature—we may not see them at all if we do not expect them. In fact,
mysticism in all its guises can lead us to perceive too much order,
making us prone to seeing significance where there is only the play of
chance. The human mind seems to be predisposed to this error, for
pattern recognition is an essential survival tool and it seems we must
resign ourselves to living with its tiresome side-effects, from numer-
ology to ‘faces’ on the surface of Mars.
But although the mystical Platonic vision of a geometric, ordered
universe helped prepare the ground for early Western science, it needed
Fig. 1.1: The snowflake displays an urge for branching growth played
out with exquisite hexagonal symmetry. (Photo: Ken Libbrecht,
California Institute of Technology.)
to be replaced by something more empirical, more discerning and scep-

tical, before we could truly begin to understand how the world works.
The snowflake offers a delightful illustration of that process. For it is only
when we start to regard these ice crystals as things in themselves, and not
as symbols of some deeper principle of nature, that we can truly appre-
ciate how astonishing they are. Their elegance and beauty is, I believe,
unrivalled in the natural world, and even Bach would have been silenced
by the invention with which they play variations on a simple theme,
this interplay of ‘sixness’ and ‘branchingness’ in which symmetry seems
to be taken about as far as it can tolerate (Fig. 1.1 and Plate 1). They are
formed from chaos, from the random swirling of water vapour that
condenses molecule by molecule, with no template to guide them.
Whence this branchingness? Wherefore this sixness?
Kepler’s balls
In the mechanistic worldview that emerged in the West during the wane
of the Renaissance, an appeal to numerology could not suffice to ac-
count for the remarkable symmetry of the snowflake. The spirit of the
age insisted on causative forces that dictated how things happened in
their own terms. One could concede that God set the forces at play while
insisting that, on a day-to-day basis, they were all He had to work with.
Snowflakes interested the Englishman Thomas Hariot, who noted in
his private manuscripts in 1591 that they have six points. Hariot was a
masterful mathematician, noted for his contributions to algebra, but
his enthusiasms showed the characteristic magpie diversity of the
Elizabethan intellectual, among them astronomy, astrology, and lin-
guistics. He tutored Walter Raleigh in mathematics, and when Raleigh
set out on a voyage to the New World in 1585 he employed Hariot as
navigator. Together they sailed to the land that Raleigh named in
honour of his Virgin Queen: Virginia. On the voyage, Raleigh sought
Hariot’s expert advice about the most efficient way to stack cannonballs
on deck.
The question led Hariot to the beginnings of a theory about the close-
packing of spheres. Some time between 1606 and 1608 he communi-
cated his thoughts to a fellow astronomer, the German Johannes Kepler,
who enjoyed the patronage of the Holy Roman Emperor Rudolph II at
his illustrious court in Prague. Most of the correspondence between
Kepler and Hariot concerns the refraction of light and the origin of
rainbows, but they also discussed atomism: what are atoms, and can
empty space come between them? This was an ancient theme,
prompted by the belief that nature abhors a vacuum, but it seemed
then to be as irresolvable as ever. The issue of how atoms sat against
one another brought Hariot back to Raleigh’s cannonballs, and he asked
what Kepler thought about the matter. In 1611 Kepler wrote a short
treatise in which he speculated that the familiar cannonball stacking,
which disports the balls in a hexagonal, honeycomb array, is the dens-
est arrangement there can be. The hexagonal packing ‘will be the
tightest possible’, he wrote, ‘so that in no other arrangement could
more pellets be stuffed into the same container’.* The booklet in
which this assertion was contained was a New Year’s gift from Kepler
to his patron Johann Matthäus Wacker von Wackenfels: seasonably so,
for its title indicates the object towards which Kepler’s thoughts on
close-packing became directed. It was called On the Six-Cornered
Snowflake.
‘There must be a cause why snow has the shape of a six-cornered
starlet’, Kepler says. ‘It cannot be chance. Why always six? The cause is
not to be looked for in the material, for vapour is formless and flows,
but in an agent.’ But Kepler does not claim that he can solve the
mystery; indeed, his booklet is a rather charming study in bafflement,
full of false trails and head-scratching. Nonetheless, it contains the seed
of an important idea. Prompted by his discussions with Hariot, Kepler
began to think about the geometrical shapes that bodies will adopt if
their constituent particles are close-packed like cannonballs. He sug-
gested that the hexagonal symmetry he had seen in snowflakes that he
collected and observed that very winter might stem from the stacking of
‘globules’ of water. These globules are not in themselves atoms; rather,
he said, ‘vapour coagulates into globules of a definite size, as soon as it
begins to feel the onset of cold’. They are like little droplets, and, as
such, are perfectly spherical.
Yet in the end Kepler rejects this idea, for he notes that balls can be
packed into other regular patterns too—notably square arrays—and yet
four-pointed snowflakes are never observed. He remarks that flowers
commonly display five-pointed heads (a notion I explored in Book I),
*Kepler’s conjecture remained just that for nearly four centuries. It was proven to be true by
the American mathematician Thomas Hales in 1998.
which he attributes to a ‘formative faculty’ or plant soul. But ‘to imagine

an individual soul for each and any starlet of snow is utterly absurd’,
Kepler wrote, ‘and therefore the shapes of snowflakes are by no means
to be deduced from the operation of soul in the same way as with
plants.’ So how does water vapour acquire a formative faculty? It
must, in the end, be God’s work—which sounds like a capitulation,
but in fact reflects the semi-mystical belief common among early
seventeenth-century philosophers that nature is imbued with ‘hidden’
forces that shape its forms. Yet what purpose could be served by this
symmetrical expression of a gaseous formative faculty? There is none,
Kepler decides: ‘No purpose can be observed in the shaping of a snow-
flake . . . [the] formative reason does not act only for a purpose, but also
to adorn . . . [it] is in the habit also of playing with the passing moment.’
In this seemingly whimsical conclusion we can discern something valid
and profound—for, as I hope this trilogy will show, nature does indeed
seem to have an intrinsic pattern-forming tendency that it exercises as
though from some irrepressible urge. Kepler even hints inadvertently at
the way this impulse can act in living organisms in apparent defiance of
the strict utilitarianism that Darwinism later seemed to dictate.
Despite its inconclusiveness, Kepler’s treatise on the snowflake es-
tablished the idea that the geometric shapes of crystals are related to
the ordered arrangements of their component units. From this elem-
entary notion came the science of crystallography, beginning in the late
eighteenth century, in which the faceted nature of mineral crystals is
explained in terms of close-packing of their atoms and molecules
(Fig. 1.2). And what is more, his invocation of an almost vitalistic
principle behind the growth of snowflakes, redolent of (if not the
same as) the ‘soul’ that guides the growth of plants, captures something
of the confusion that snowflakes provoke. The sixness, the hexagonal
symmetry, speaks of crystals, of a regularity so perfect that it appears
barren. But the branchingness hints at life and growth, at something
vegetative and vital.
René Descartes, the arch-mechanist of the early Enlightenment,
could not resist the allure of snowflakes. He sketched them in 1637 for
his study of meteorology, Les Météores, where he recorded rarer var-
ieties alongside the six-pointed stars (Fig. 1.3):
After this storm cloud, there came another, which produced only little roses or
wheels with six rounded semicircular teeth . . . which were quite transparent and
Fig. 1.2: Early crystallographers such as René Just Häuy, from whose
book Traité de Minéralogie (1801) this illustration comes, explained
the faceted shapes of crystals in terms of the packing of their
component atoms.
Fig. 1.3: Drawings of snowflakes by René Descartes in 1637.
quite flat . . . and formed as perfectly and as symmetrically as one could possibly
imagine. There followed, after this, a further quantity of such wheels joined two
by two by an axle, or rather, since at the beginning these axles were quite thick,
one could as well have described them as little crystal columns, decorated at
each end with a six-petalled rose a little larger than their base. But after that there
fell more delicate ones, and often the roses or stars at their ends were unequal.
But then there fell shorter and progressively shorter ones until finally these stars
Fig. 1.4: Twelve-pointed snowflakes are formed when two normal

six-pointed varieties fuse together at their centres, rotated in relation
to one another by about 308. (Photo: Ken Libbrecht, California
Institute of Technology.)
completely joined, and fell as double stars with twelve points or rays, rather long
and perfectly symmetrical, in some all equal, in other alternately unequal.
We can recognize in this vivid description some of the unusual forms

that have been found in snowflakes, such as prismatic columns with
end-caps, like elaborate sundials, and twelve-pointed stars in which
two hexagonal flakes have become fused (Fig. 1.4).
The English scientist Robert Hooke had the advantage of a micro-
scope in preparing illustrations of snowflakes for his famous Microgra-
phia (1665), where he shows that the ‘flowers’ are not just six-pointed
but branch repeatedly, in a hierarchical manner (Fig. 1.5a). The organic
associations of these ice crystals are very apparent in the drawings by
the Italian astronomer Giovanni Domenico Cassini in 1692, where they
look almost leafy (Fig. 1.5b). The biologist Thomas Huxley acknow-
ledged this aspect in 1869, when he called snowflakes ‘frosty imitations
of the most complex forms of vegetable foliage’. Huxley’s comments
appeared in an essay on ‘the physical basis of life’, in which he strove
Fig. 1.5: Using an early microscope, Robert Hooke recorded the characteristic ‘Christmas-tree’
branching patterns of snowflakes (a). Giovanni Domenico Cassini’s drawings from 1692 seem to
make reference to their resemblance to plants (b).
like a good positivist to quell any notion of a vital force that animated
organic matter and made it fundamentally different from the inorganic
world. To Huxley, the ‘organic’ forms of snowflakes provided evidence
that the complex shapes of the biological world need not compel the
scientist to invoke some mysterious vitalistic sculpting mechanism,
since something of that nature surely did not operate in the simple
process of the freezing of water:
We do not assume that a something called ‘aquosity’ entered into and took
possession of the oxide of hydrogen as soon as it was formed, and then guided
the aqueous particles to their places in the facets of the crystal, or amongst the
leaflets of the hoar-frost.
It was a reasonable enough assertion, but it surely begs the question: if

there is nothing ‘organic’ about the formation of the snowflake, why
then do they look so tantalizingly as though there is?
Flakes frozen on film
As Descartes hinted, the shapes of snowflakes can evolve and mutate as

the weather changes. Friedrich Martens, on board a ship travelling from
Spitzbergen in Norway to Greenland in 1675, noticed that different
meteorological conditions produce different kinds of flake. It takes an
Arctic chill to condense the best, most symmetrical snowflakes, as the
English explorer William Scoresby noted in his Account of the Arctic
Regions with a History and Description of the Northern Whale-Fishery
in 1820. Scoresby took the observations of snowflakes to a new standard
of detail and accuracy, recording a wide range of different shapes
(Fig. 1.6). One of the most charming of nineteenth-century records
was that produced in 1864 by a minister’s wife in Maine named
Frances Knowlton Chickering, who used, if not invented, the trick
now popular at Christmas of cutting out doily-style snowflakes from
folded paper. Chickering’s paper flakes were masterpieces of dexterity:
from memory of her first-hand observations, she clipped out delicate
frond-like branches and pasted the results into her Cloud Crystals: A
Snow-Flake Album, which implicitly acknowledged the ‘artistry’ of
natural phenomena that the biologist Ernst Haeckel was later to cele-
brate in his drawings of marine life, as we saw in Book I.
The accuracy of all these visual records of snowflakes was limited not
only by the power of the magnifying glass or microscope but also by the
artist’s inevitable tendency to simplify, idealize, and interpolate these
complex geometric forms. That problem was avoided once researchers
Fig. 1.6: In 1820, explorer William Scoresby made accurate drawings of the snowflakes he
observed during a trip to the Arctic.
found a way to marry the new art of photography to the power of the
microscope. Microphotography was already a well-established tech-
nique by the late nineteenth century, and one of its most inventive
practitioners was a Vermont farmer named Wilson Bentley. Between
1885 and 1931, Bentley captured over 5,000 images of snowflakes on
photographic plates, constituting one of the most comprehensive sur-
veys of their astonishing variety and beauty (Fig. 1.7). In the late 1920s
Bentley compiled 2,000 of his photographs into a book entitled Snow
Crystals in collaboration with William J. Humphreys, a physicist work-
ing for the US Weather Bureau. Bentley died only a few weeks after the
book was published in November 1931, allegedly after contracting
pneumonia during one of his forays into the New England winter.
Snow Crystals is rightly regarded as a work of wonder, but it is more
than that. The scientist, gazing at page after page of seemingly infinite
variety on the theme of the six-pointed flower of ice, faces a mystery of
an order not previously encountered in the non-living world. Not only
were the forms indescribably complex, but there was no end to them.
Bentley’s album was pure description, to which Humphreys could
add rather little in the way of hard science. But in the 1930s the book
inspired a Japanese nuclear physicist named Ukichiro Nakaya, working
at the University of Hokkaido, to consider the question of snowflake
growth in a rather more analytical spirit. He made the first systematic
attempt to discover the factors that influenced snowflake growth, lead-
ing to the many different families of shapes that had been seen by
Scoresby and others in the natural environment. Nakaya realized that
snowflakes fall into several distinct categories, and he constructed a
laboratory for exploring the conditions that generated these different
classes of shape.
It was uncomfortable work: Nakaya’s wooden-walled lab could be
cooled to 30 8C, and he worked in padded clothing with a mask to
protect his face. Snowflakes grow slowly as they fall through the atmos-
phere, but Nakaya could not recreate this long descent in the lab, so
instead he decided to reverse the situation: to hold the snowflake fixed
and to let cold, moist air pass over it in a steady stream. The question
was, how do you hold onto a snowflake? Nakaya experimented with
many different kinds of filament for immobilizing a growing crystal of
ice, but most of them simply became coated with frost. He finally found
that the experiment worked best with a strand of rabbit hair, on which
the natural oils suppressed the simultaneous nucleation of many ice
Fig. 1.7: The collection of

snowflake photographs
amassed by Wilson Bentley
in the four decades after
1885 still stands as the
most remarkable record of
their endlessly varied
forms.
Fig. 1.8: Snowflakes made artificially by Ukichiro Nakaya in the 1930s. In the image on the left, the
rabbit’s hair on which the crystals are nucleated is still visible.
crystals at once (Fig. 1.8). Using this equipment, Nakaya and his co-
workers found that the shapes of the individual crystals changed as two
key factors were altered: the temperature and humidity of the air. At low
humidity, the crystals did not develop the six frond-like arms of classic
snowflakes, but took on more compact forms: hexagonal plates and
prisms. These shapes persisted even in moister air if it was very cold
(below about 20 8C). At higher temperatures, however, increasing the
humidity tended to increase the delicacy and complexity of the snow-
flakes, giving rise to the highly branched star forms. In a temperature
range between about 3 and 5 8C, needle-like crystals appeared in-
stead (Fig. 1.9).
Nakaya collected his findings in an album of images clearly indebted
to Bentley and Humphreys, called Snow Crystals: Natural and Artificial
(1954). His studies brought some order to the ice menagerie, but they
did not really bring us any closer to understanding the fundamental
mechanism by which a simple process of crystallization, which typic-
ally generates a compact prismatic or polyhedral shape, in this case
gives us structures that seem to have a life of their own.
As I explained in the previous volumes, the first person to tackle this
sort of question about the genesis of complex form within a modern
scientific framework was the Scottish zoologist D’Arcy Wentworth
Thompson, whose 1917 book On Growth and Form set the scene for
everything I discuss in this series. Thompson included drawings based
on Bentley’s photographs in the 1942 revised edition of his book. ‘The
snow crystal’, he wrote, ‘is a regular hexagonal plate or thin prism.’ But
Fig. 1.9: The ‘morphology diagram’ of snowflakes, showing how their

shape changes for different conditions of temperature and humidity
(supersaturation).
‘ringing her changes on this fundamental form, Nature superadds to

the primary hexagon endless combinations of similar plates or prisms,
all with identical angles but varying lengths of side; and she repeats,
with an exquisite symmetry, about all three axes of the hexagon, what-
soever she may have done for the adornment and elaboration of one.’
In other words, all the arms appear to be identical. ‘The beauty of a
snow-crystal depends on its mathematical regularity and symmetry’,
Thompson observed,
but somehow the association of many variants of a single type, all related but no
two the same, vastly increases our pleasure and admiration. Such is the peculiar
beauty which a Japanese artist sees in a bed of rushes or a clump of bamboos,
especially when the wind’s ablowing; and such is the phase-beauty of a flowering
spray when it shews every gradation from opening bud to fading flower.
Here it is again: flowers and ice. But even Thompson, like Kepler, could
say no more. With all his ideas about forces and equilibria and geometry,
he, too, was forced to take recourse in metaphors from the organic world.
Endless branches
By the time Nakaya’s book appeared, scientists had found a way to

attack the problem. Although ice seems to be unique in forming highly
symmetrical, isolated flakes, many other substances may crystallize as

needle-like protrusions punctuated by regular branches, like a single
snowflake arm. These structures, known as dendrites (from the Greek
for ‘tree’) are found when molten metals freeze (Fig. 1.10a), when salts
precipitate out of a solution, and when metal deposits form on elec-
trically charged electrodes, a process known as electrodeposition and
related to electroplating (Fig. 1.10b) (see page 30). Dendrites typically
have a rounded tip, like the prow of a boat, behind which side-arms
sprout and grow in a Christmas-tree pattern. In general, they appear
when the solidification process happens rapidly, as for example when a
molten metal is quenched (‘undercooled’) by being plunged into cold
surroundings. That’s an important clue. We observed in Book I that
complex pattern and form is often generated in processes that take
place significantly out of thermodynamic equilibrium—which is to say,
when the system is highly unstable. A system in equilibrium does not
change; a system out of equilibrium ‘seeks’ to attain such a stable state
if left alone to do so, but can be driven away from this goal by a constant
influx of energy. We saw in Book II that convection (the flow of a fluid
when heated from below) produces such a non-equilibrium state.
A liquid that is abruptly cooled far below its freezing point is another
non-equilibrium system, being unstable relative to the solid form of
the material. That instability makes change happen rapidly, under
which conditions pattern is apt to appear. In contrast, crystals that
are formed close to equilibrium—very close to their freezing point,
say—grow slowly, and tend instead to develop the familiar compact,
faceted shapes.
In 1947 the Russian mathematician G. P. Ivantsov showed theoretic-
ally that a metal solidifying rapidly from its molten form may develop
needle-like fingers. Ivantsov calculated that the needles have a shape
mathematicians called parabolic, with gently curving sides that con-
verge on a blunt tip. This is the same shape as the trajectory of a stone
thrown through the air and falling under gravity. Ivantsov showed that
in fact all possible types of parabolic needles may be formed, but that
the thinner they are, the more rapidly they grow; so thin, needle-like
tips should shoot rapidly through the molten metal, while fatter bulges
make their way forward at a more ponderous pace.
But in the mid-1970s, Martin Glicksman and co-workers at the
Rensselaer Polytechnic Institute in New York performed careful experi-
ments which showed that, instead of a family of parabolic tips, only one
Fig. 1.10: Dendrites formed by rapid solidification of a molten metal (a) and in the
electrodeposition of a metal (b). (Photos: a, Lynn Boatner, Oak Ridge National
Laboratory, Tennessee. b, Eshel Ben-Jacob, Tel Aviv University.)
single tip shape was seen during rapid solidification of metals. For a
fixed degree of undercooling, a particular tip is privileged over the
others. For some reason, one of Ivantsov’s family of parabolas seems
to be special.
The puzzle was even more profound, however, because in 1963 two
Americans, William Mullins and Robert Sekerka at Carnegie Mellon
University in Pittsburgh, argued that none of Ivantsov’s parabolas
should be stable. They calculated that the slightest disturbance to the
growth of a parabolic tip will be self-amplifying, so that small bulges
that form by chance on the edge of the crystal grow rapidly into thin
fingers. This so-called Mullins–Sekerka instability should cause the tip
to sprout a jumble of random branches.
The instability is an example of a positive feedback process—again,
we have encountered such things already in Books I and II. It works like
this. When a liquid freezes, it releases heat. This is called latent heat,
and it is the key to the difference between a liquid and its frozen, solid
form at the same temperature. Ice and water can both exist at zero
degrees centigrade, but the water can become ice only after it has
becomes less ‘excited’—its molecules cease their vigorous jiggling mo-
tions—by giving up latent heat.
So, in order to freeze, an undercooled liquid has to unload its latent
heat. The rate of freezing depends on how quickly heat can be con-
ducted away from the advancing edge of the solid. This in turn depends
on how steeply the temperature drops from that in the liquid close to
the solidification front to that in the liquid further away: the steeper the
gradient in temperature, the faster heat flows down it. (It may seem odd
that the liquid close to the freezing front is actually warmer than that
further away, but this is simply because the front is where the latent
heat is released. Remember that in these experiments all of the liquid
has been rapidly cooled below its freezing point but has not yet had a
chance to freeze.)
If a bulge develops by chance—because of the random motions of
the atoms and molecules, say—on an otherwise flat solidification front,
the temperature gradient becomes steeper around the bulge than else-
where, because the temperature drops over a shorter distance
(Fig. 1.11). So latent heat is shed around the bulge more rapidly than
it is to either side, and the bulge grows, its apex fastest. This in turn
sharpens the tip and speeds its advance even more.
Fig. 1.11: The Mullins–Sekerka instability makes protrusions at the

surface of a solidifying material unstable. Because the temperature
gradient (shown here as dashed contours of equal temperature) is
steeper at the tip of the protrusion, heat is conducted away faster and
so solidification proceeds more rapidly here.
In principle, this instability will amplify any irregularity on the solid

front into a growing finger, no matter how small it is. But there is
another factor that sets a minimum limit to the width of the fingers.
The interface between the solid and the liquid has a surface tension,
just like that at the surface of water in a glass. As I explained in Book I,
the existence of surface tension means that an interface costs energy:
the bigger the surface area, the higher the energetic cost. Surface
tension thus encourages surfaces to keep their area as small as possible,
and here it tends to ‘pull’ the solidification front flat. Thanks to this
smoothing effect, surface tension suppresses bulges smaller than a
certain limit. This means that the Mullins–Sekerka instability produces
a characteristic branch-tip width, set by the point at which the narrow-
ing of tips caused by positive feedback is counterbalanced by their cost
in surface energy. In other words, the front develops fingers with a
certain wavelength—a regular pattern with a particular size scale to it,
determined by a balance of opposing factors.
In 1977 James Langer at the University of California at Santa Barbara

and Hans Müller-Krumbhaar in Jülich, Germany, suggested that the
Mullins–Sekerka instability might explain Glicksman’s observation of
a single parabolic tip being selected from all of those allowed by
Ivantsov’s theory. The instability will make fat fingers break up into a
mass of smaller ones, they said, while surface tension sets a limit on
how small and narrow they can become. Perhaps, then, there is an
optimal tip width at which these two effects balance, favouring a single
‘marginally stable’ parabolic tip.
But in the early 1980s, Langer and his co-workers showed that surface
tension in fact destroys this neat picture. Its influence makes the tip of
the dendrite become cooler than the regions to either side. So the tip
starts to slow down, and eventually it forks into two new fingers. These
also split subsequently, and so on. This repeated tip-splitting results not
in a dendritic growth shape at all, but instead a dense mass of repeat-
edly forking branches, a pattern known as the dense-branching
morphology.
This turned out to be a persistent problem with theories of dendrite
growth: they seemed prone to instabilities that led to randomly
branched fingering patterns, not the orderly, Christmas-tree shapes of
snowflake arms. What these theories were neglecting, in the mistaken
belief that it was a mere detail, was the most striking aspect of a
snowflake’s shape: its hexagonal symmetry. It had been suspected
ever since Kepler that this was an echo of the underlying symmetry in
the arrangement of constituent particles. But no one had guessed that it
was to this symmetry that the dendrites owed their very existence.
The joy of six
Why hexagons? Remarkably, the Chinese sages were right: there is a

sense in which six is the number of water. In 1922 the English physicist
William Bragg used his new technique of X-ray crystallography to
deduce how water molecules are arranged in an ice crystal. X-rays
bouncing off crystals produce a pattern of bright spots that encode
the positions of the atoms; Bragg deduced how to calculate backwards
from the X-ray pattern to the atomic structure. In this way, he found
that the water molecules in ice are linked by weak chemical bonds into
hexagonal rings, one molecule at each corner (Fig. 1.12). Thus the
Fig. 1.12: Ice has a structure with hexagonal symmetry at the

molecular scale. The water (H2O) molecules are linked together by
weak chemical bonds called hydrogen bonds. In this image the spheres
denote the oxygen atoms at the centres of the molecules, and the
rods denote the hydrogen bonds that bind molecules together.
crystal structure is dictated not by the shapes of the water molecules

themselves but by the way in which they are joined together.
It might seem unlikely that this is the origin of the six-pointed
snowflake, since water molecules are very much smaller than a snow-
flake—how could this sixness become so amplified? But as Kepler and
the early crystallographers realized, geometric packing of a crystal’s
constituent units dictates the geometry of the much larger bodies that
result. In essence, water’s crystal structure imposes an innate hexagon-
ality on the way the ice crystal grows. Or, as D’Arcy Thompson wrote
with his customary elegance, ‘these snow-crystals seem to give visible
proof of the space-lattice on which their structure is framed’.
The presence of the hexagonal ‘space-lattice’ means that not all direc-
tions are the same for the growing crystal. That is why faceted crystals have
the characteristic shapes that they do: the flat facets are simply planes of
stacked atoms or molecules, but the reason why certain planes and not
others define the crystal’s form is that some facets grow faster. This non-
equivalence of directions is called anisotropy; an isotropic substance is
one that looks the same, and behaves in the same way, in all directions.
The anisotropy of crystals means that properties like surface tension

differ in different directions. In 1984 Langer and his co-workers showed
that, for Ivantsov parabolas growing in certain ‘favoured’ directions
picked out by the anisotropy of the material’s crystal structure, surface
tension no longer induces a tip-splitting instability—the parabolic tip
remains stable as it grows. Thus dendritic branches will grow outwards
from an initial crystal ‘seed’ only in these preferred directions: the
snowflake grows six arms. This special role of anisotropy in stabilizing
the growth of a particular needle crystal was identified independently
at the same time by David Kessler, Joel Koplik, and Herbert Levine at the
University of California at San Diego.
Anisotropy also explains why a dendrite develops side branches.
When, by chance, the parabolic tip develops small bulges on its flanks,
these may be amplified by the Mullins–Sekerka instability. But again,
only bulges that grow in certain directions will be stable. And there is
only one kind of dendrite tip, for a given set of growth conditions, which
grows fast enough to avoid being overwhelmed by these side branches.
So a particular dendrite, with side branches sprouting in particular
directions, is uniquely selected from amongst the possible growth
shapes.
How the right arm knows what the left arm is doing
The mind-boggling variety of snowflake forms is therefore the outcome

of a tension between chance and necessity. The mechanics of the
growth process ensures that the arms will sprout in directions that
point to the corners of a hexagon. For any given snowflake, these
arms will all grow at the same rate (because, at such a small scale,
they all experience the same conditions of temperature and humidity),
and so they will have the same length. The side-branches of this six-
pointed star are to a degree at the mercy of fate: they may be triggered
by the random appearance of tiny irregularities or bulges along the
parent arm. Yet they too will always surge outwards in a ‘hexagonal’
direction. Changes in the prevailing conditions that an individual
snowflake experiences as it drifts and falls in the air may trigger simul-
taneous changes in the growth of all the branches, accounting for how,
for example, needle-like arms might develop hexagonal plate-like for-
mations at their tips (Fig. 1.13).
Fig. 1.13: A change in ambient conditions as a snowflake grows in the atmosphere can result in
a change in morphology of the branches. Here needle-like branches have developed plate-like tips.
(Photo: Ken Libbrecht.)
But that doesn’t quite explain it all. If pure chance dictates the side-
branching of the six arms, why are some snowflakes so amazingly
symmetrical even in their fine decorations (Fig. 1.14)? There appears
to be something almost magical at play here—each arm seems some-
how to know what all the others are doing. Nakaya confessed to being
perplexed by this apparent ‘communication’ between the branches.
‘There is apparently no reason’, he said,
why a similar twig must grow, in the course of the growth of the crystal, from one
main branch when a corresponding twig happens to extend from another main
branch . . . In order to explain this phenomenon we must suppose the existence
of some means which informs other branches of the occurrence of a twig on a
point of one branch.
Yet in fact many flakes do not have this perfect symmetry: the six arms
look roughly identical in their general features, but close inspection
reveals differences of detail (Fig. 1.15a). Snowflake expert Ken
Libbrecht of the California Institute of Technology, who has taken up
the mantle of Nakaya and Bentley and Humpreys in cataloguing
the richness and beauty of these crystals by microphotography (his
images adorn these pages) says that he commonly rejects thousands
of flakes for every one he considers beautiful (which is to say, symmet-
rical) enough to record. Physicists Johann Nittmann and Gene Stanley
Fig. 1.14: Why some snowflakes have arms that are essentially identical in all their
intricate detail is still somewhat of a mystery. (Photo: Ken Libbrecht, California
propose that almost perfectly regular snowflakes are in fact the excep-
tion rather than the rule. They say that the apparent perfection is often
illusory: we are fooled into perceiving it simply because each arm has
side branches diverging at the same angle and because the ‘envelope’ of
each arm has the same shape. Nittmann and Stanley showed that a
model in which particles drift about at random but stick when they
make contact* will produce convincing snowflake-like shapes if an
underlying sixfold symmetry is imposed by constraining the particles
to move from site to site on a hexagonal grid, as though hopping
between the cells of a honeycomb (Fig. 1.15b). None of the branches
in this flake is identical, even though at a glance they look similar. But
*We will encounter this model in much more detail in the next chapter, where we will see
that it can give rise to a much more widespread and less orderly branching pattern.
Fig. 1.15: The six branches of some snowflakes can be quite different from one another in their
fine details, even though at a glance they look symmetrical (a). Such flakes can be grown in a
computer model of aggregating particles that assigns at random where a new particle gets
attached, subject to the constraint that the particle positions must lie on a hexagonal grid (b). There
is nothing in the rules of the model to ensure that all branches are the same, and indeed they are not
the same; but our eyes are fooled into seeing more symmetry than there really is by the uniformity of
the branching angles. (Photo and image: a, Ken Libbrecht, California Institute of Technology; b,
Gene Stanley, Boston University.)
the general Christmas-tree shape is preserved in all of them, and their

lengths are more or less the same, simply because both the main
branches and their respective side branches grow at roughly the same
rates. The randomness actually ensures this, because it gives no one
branch any opportunity to grow faster than the others.
Mathematicians Janko Gravner and David Griffeath found something
similar in a more sophisticated model of snowflake growth. Like
Nittmann and Stanley’s model, this assumes that the crystals grow by
accumulating units (much larger than individual water molecules) at
the edges from the surrounding vapour, which are constrained to pack
together in a flat hexagonal array. But the attachment rules are rather
complicated, invoking several parameters whose values can be altered
to explore different accumulation conditions. In effect, these rules
‘build in’ the physical processes involved in crystal growth in a some-
what ad hoc, albeit fairly realistic, way rather than simply letting them
emerge (or not) from much simpler rules. Gravner and Griffeath were
Fig. 1.16: Snowflakes grown in a model developed by Janko Gravner and David
Griffeath are remarkably realistic (a). Here the branches are forced to be identical,
but when that condition is relaxed by an injection of randomness, the branches
still look rather similar (b). (Images: Janko Gravner and David Griffeath, from
Gravner and Griffith, 2008).
able to grow a huge variety of snowflakes in their scheme (Fig. 1.16a). In

the basic model all the arms were forced to be identical, but the
researchers could also subvert this by introducing some randomness
in the density of the vapour surrounding the crystal. In that case, the
snowflakes could still have six branches that looked rather similar at a
Fig. 1.17: A selection of snowflakes grown in the three-dimensional version of Gravner and
Griffeath’s model. These display the kinds of decorations, such as ribs and ridges, seen in real flakes.
(Images: Janko Gravner and David Griffeath, from Gravner and Griffith 2009).
glance, although differing in detail (Fig. 1.16b). Gravner and Griffeath

think that this might be because there are in fact only a relatively small
number of stable side-branching designs: the choices the flake has are
by no means limitless.
They have extended their model so that it can make three-dimensional
crystals, in which case they find not only flakes of startling realism
(Fig. 1.17) but also the needle, prism and columnar forms catalogued
by Nakaya. While questions of detail remain, it now seems that we are
well on the way to decoding the mysteries of ‘snow flowers’.
Snowflakes have provided scientists and natural philosophers with
one of the clearest indications that complexity of form is not a special
attribute of the living world. Within the snowflake they found an echo
of the shapes of trees and flowers, ferns, and starfish: patterns that
hint at an exquisite conciliation between geometric purity and or-
ganic exuberance. These forms have forced us to re-examine notions
of how beauty arises in nature, and how universal processes connect
the living and non-living world. ‘How full of creative genius is the air
in which these are generated!’ exclaimed Henry David Thoreau in
1856. ‘I should hardly admire more if real stars fell and lodged
on my coat.’
Tenuous Monsters
Shapes between Dimensions
2
I
N Mike Leigh’s 1976 film Nuts in May, a pair of gauche campers named
Keith and Candice Marie discover how a kind of modern-day
vitalism colours our preconceptions about complex growth and
form. They take a trip to a local quarry to look for fossils in the ancient
limestone of Dorset in southern England. A quarryman shows the
unsuspecting couple a delicate, plant-like pattern traced out in the
stone. ‘Is that a fossil?’, asks Candice Marie, awestruck. ‘Ar, most people
think that’, the Dorset quarryman tells them. ‘It’s just a mineral.’
It’s easy to see why Keith and Candice Marie jumped to the wrong
conclusion. The structures they saw are called mineral dendrites
(Fig. 2.1), and they look for all the world like the forms we associate
with plants—which is of course precisely why they are named after
them.* But these filigrees contain no fossil material; they are made of
iron or manganese oxides, chemical deposits precipitated when a so-
lution rich in these metal salts was squeezed through cracks in the
rocks in the geological past.
Mineral dendrites scream ‘life’ at us because their branched patterns
are echoed everywhere in the organic world, from corals to leaf veins to
the bronchial structure of the lung. But such branched formations are
not by any means unique to biology: think of river networks, lightning,
cracks in the ceiling. We will encounter each of these in later chapters.
And of course not all branching forms are alike: a snowflake little
resembles a naked oak, and I don’t think we would mistake this mineral
dendrite for a fracture pattern. Botanists and foresters can identify the
*These are not the same ‘dendrites’ as those described in rapidly freezing metals in the
previous chapter, which have a more regular, snowflake-arm appearance. Neither, for that
matter, have they anything to do with what neuroscientists would recognize by the term,
namely the branched extremities of nerve cells in our brain. Metallurgists, geologists and
biologists have all been eager to seize upon the tree metaphor, the Greek dendron, without
much regard for how the word has been applied elsewhere.
Fig. 2.1: A mineral dendrite of manganese oxide found in Bavarian

limestone. (Photo: Tamás Vicsek, Eötvös Loránd University, Budapest.)
species of a tree in winter simply from its silhouette on a hilltop. How

do they do that—what distinguishes the boughs of an elm from those of
a sycamore? I doubt that many botanists could tell you with any great
precision. They might mutter something about angles of divergence in
the branches, but in the end they just seem able to ‘sense’ the charac-
teristics of the pattern. Might we, though, hope to establish a taxonomy
of branching, and perhaps to find within it a boundary that separates
the living from the inorganic?
Organic rocks
Keith and Candice Marie were in venerable company. In the early

1670s, Isaac Newton wrote a short tract that is little known now but
TENUOUS MONSTERS j 29
was probably considered by its author to be the precursor to a (never-

written) magnum opus comparable to his Principia. It was called Of
Natures obvious lawes and processes in vegetation, and it was an at-
tempt to sketch out the framework for a ‘theory of everything’. The work
begins with a description of mineral dendrites, produced artificially—
which is to say, alchemically—in Newton’s laboratory.
Newton doesn’t explain exactly how he made these crystalline
growths, but it seems likely that he followed a prescription similar to
that of the seventeenth-century German chemist Johann Rudolph
Glauber. Glauber described how iron dissolved in ‘spirit of salt’ (hydro-
chloric acid) will generate branched deposits of its red oxide when
added to a tall flask of ‘oil of sand’ or ‘oil of glass’ (potassium silicate).
The dendritic structures that can be made this way from precipitating
metal salts are now known as silica gardens.
Newton never used that term, but he might be expected to have
found it most congenial—for he believed there was something ‘vegeta-
tive’ about metals which led them to grow within the earth as veins that
are entirely analogous to trees. The fact that ‘metal trees’ could be made
in the lab apparently confirmed this view.
There was nothing idiosyncratic about the idea, which was shared by
many scientists in Newton’s day. That metals and salts were formed by
vegetative growth was proposed in the early sixteenth century by the
Swiss alchemist Paracelsus, who stated that, just as trees have their
roots in the soil and grow upwards into the air, so mineral veins have
their roots in subterranean water and grow upwards into the earth.
Mineral veins, wrote Paracelsus’s contemporary Biringuccio, an Italian
authority on mining and metallurgy, ‘show themselves almost like the
veins of blood in the bodies of animals, or the branches of trees spread
out in different directions’.
Taking his lead from Paracelsus, the French ceramicist and natural
philosopher Bernard Palissy suggested around 1580 that minerals grow
from salty ‘seeds’ that germinate in water. But the idea arose in the late
sixteenth century that mineral dendrites might indeed be fossil plants,
perhaps produced by inundation in the biblical flood. This position
was challenged by the Swiss geologist Jean-Jacques Scheuchzer, whose
1709 book Herbarium diluvianum (Herbarium of the Flood) contains
remarkably accurate drawings of the branched forms of mineral den-
drites (Fig. 2.2). Scheuchzer was actually an enthusiastic Diluvianist
himself, and he accepted the idea that mineral dendrites were products
Fig. 2.2: Mineral dendrites drawn by Jean-Jacques Scheuchzer, in his book Herbarium diluvianum
(1709).
of the Deluge—but not, he said, as fossil plants. Instead, he suggested a

mechanism of formation that, as we shall see, was astonishingly pres-
cient. He explained that if a liquid is squeezed between two flat plates
that are then pulled quickly apart, the liquid film is drawn out into a
dense array of branches, just like those seen in mineral dendrites, as air
pushes its way in. Thus, Scheuchzer explained to the Académie des
Sciences in Paris, ‘Each time that one finds small trees in stones that
can be split easily, which seem to be painted artificially and whose
small branches are separated one from the other and never intersect,
then one must attribute this arborescence to the injection of a fluid.’
But how was this connected to the Flood? Scheuchzer proposed that
the seas sloshed violently onto all the lands of the Earth when, in an
instant, God stopped the world turning on its axis. The sudden influx of
water into the cracks and pores of land-borne rock could have
imprinted these delicate mineral fronds.
Flood or no flood, Scheuchzer’s experiment undermined any need to
invoke vegetation as a formative principle for mineral dendrites. By the
mid-eighteenth century, the Swedish mineralogist Johann Gottschalk
Wallerius was able to write in his influential treatise on minerals that
There are naturalists who maintain that minerals have a life like that which
vegetation enjoys; but since no one has yet been able to see, even with the best
microscopes, that these substances have a juice contained in their fibres or
veins, since no one has established this view with some evidence, and moreover
since it is impossible in general to sustain any notion of life without a circulating
juice, one cannot see any basis for ascribing life to minerals, unless one does not
want to call living everything that has the faculty to grow and to increase itself.
Already, life was deemed dispensable in making ‘organic’ forms.

Deposit at your nearest branch
Today it is a rather simple matter to grow ramified crystals such as

mineral dendrites in the laboratory. One way is to use the process of
electrodeposition that we encountered in the previous chapter.
A negatively charged electrode is immersed in a solution of a metal
salt, and the pure metal grows at the electrode surface. Metal ions have
a positive electrical charge, since they are metal atoms that have lost
one or more negatively charged electrons, so they are electrically
attracted to the electrode, where they can pick up their missing elec-
trons and revert back to neutral metal atoms. The atoms stack together
in a crystal that grows outwards from the electrode surface.
In electroplating, this process is conducted at a low electrode voltage,
and the metal film grows slowly, coating the electrode with a smooth,
even veneer. But at higher voltages the deposit grows more quickly—
that is to say, out of equilibrium. In that case the metal deposit becomes
irregular and branches blossom from it (Fig. 2.3a). This no longer looks
like a crystalline entity, although close inspection with a microscope
reveals that the branches are after all composed of tiny crystals fused
together in jumbled disarray (Fig. 2.3b).
In 1984 Robert Brady and Robin Ball from the University of
Cambridge showed that a theoretical model developed three years
earlier by two American physicists, Tom Witten at the Collège de
France in Paris and Len Sander at the University of Michigan, could
account for the shape of these branched electrodeposits. Witten and
Sander hadn’t been thinking about electrodeposition at all, however.
Their model was supposed to describe something quite different: the
way particles of dust form clumps or aggregates as they drift about and
stick together in air. Witten and Sander supposed that each particle
wanders at random (a process called diffusion) until it hits another
particle, whereupon they stick together the moment they touch. This
means that particles are added randomly to a growing cluster from all
directions. The rate at which the cluster grows depends on how quickly
the particles diffuse—how long it takes them to reach the cluster’s
periphery. For this reason, Witten and Sander called their model diffu-
sion-limited aggregation (DLA).
This sounds rather like the way a snowflake grows in the frigid
atmosphere: water molecules diffuse until they collide with a growing
flake, to which they stick. But there’s a crucial difference. In snowflakes
Fig. 2.3: A branching metal formation produced by electrochemical

deposition onto a central electrode (a). Seen close up under a
microscope (b), the branches consist of conglomerates of tiny
crystallites oriented at random. (Photos: a, Mitsugu Matsushita, Chuo
University; b, Vincent Fleury, Laboratory for Condensed Matter Physics,
Palaiseau.)
the water molecules in the ice crystal stack together in an orderly

manner, forming hexagonal rings. This implies that, once they reach a
snowflake’s surface, the molecules can move about until they find the
right slot in the crystal structure. But in DLA the constituent particles of
a cluster have no opportunity for such adjustment—they become im-
mobile on contact, and so the way they pack together in the clump is
irregular. As a result, the surface of the growing cluster soon becomes
very jagged and disorderly. You could say that, because the process is
occurring far from equilibrium, the cluster grows too fast for the par-
ticles to find the most compact way to pack together, and there are lots
of ‘packing errors’ that get frozen in.
Witten and Sander simulated the DLA process on a computer by
introducing particles one by one into a box from random points around
its edges, allowing them to diffuse until they encounter and stick to a
particle at the box’s centre. This generates a cluster that grows in
tenuous branches (Fig. 2.4). It looks very similar to the structures
created in electrodeposition, something that was first recognized by
Mitsugu Matsushita of Chuo University in Japan and co-workers in
1984. Brady and Ball proposed that this is because the mechanism of
non-equilibrium electrochemical growth shares the same broad fea-
tures as the DLA model: random diffusion of ions and their instant
attachment to the electrode deposit. And that’s essentially true,
although the details are more complex.
It is clear to see why the random impacts in the DLA model result in
very rough-edged clusters. But why are they branched? We could per-
haps imagine instead the formation of a dense mass with a highly
irregular border, like a spreading ink blot. Why doesn’t this happen?
The answer is that in DLA, as in snowflake growth, small bumps or
irregularities are amplified by a growth instability that draws them out
Fig. 2.4: A cluster of aggregated particles grown by the diffusion-

limited aggregation model. (Image: Thomas Rage and Paul
Meakin, University of Oslo.)
into long, narrow fingers. When a bump appears by chance on the

cluster surface, it pokes out beyond the rest of the front, so there is a
better chance that a randomly diffusing particle will hit it (Fig. 2.5). As a
result, the bump grows faster than the rest of the surface. And, crucially,
this growth advantage is self-enhancing—the more the bump develops,
the greater the chance of new particles striking and sticking to it; in
other words, there is positive feedback on the growth of a ‘finger’. The
probability of acquiring a new particle is highest at the most exposed
part—the very tip of the bump—so that growth is accompanied by a
sharpening and narrowing of the tip, generating a tendril that itself
sprouts random appendages through the same growth instability.
Notice that the random, diffusive motion of the particles is essential
to all of this. If they were instead all propelled towards the cluster along
straight trajectories, like rain falling on a road, the edge of the cluster
would just advance uniformly.
Not only does the ‘surface’ of the cluster become ramified and dif-
fuse, but the interior never gets ‘filled in’. That is because the sprawling
branches make it very difficult for a particle to reach very far inside.
Given its meandering path, a particle is unlikely to get far down one of
the fjords between branches before hitting something—and once it
hits, it sticks. This means that the clusters attain open, fluffy shapes.
Within the roughly circular envelope of the outermost tips (or spherical
in three dimensions), much of the space is empty. Soot particles, which
grow by the aggregation of tiny flecks of carbon released in combustion,
Fig. 2.5: How the DLA model acquires a growth instability that
promotes branching. Small protrustions on the aggregate surface
accumulate new particles faster than the surrounding flat surface, and
so they become increasingly accentuated (a–c). These protrusions will
themselves have random irregularities that blossom into new fingers,
so the cluster becomes highly branched. Here the dashed lines show the
contours of average density of incoming particles, and the solid lines
show the average flow of particles, which becomes focused towards
branch tips. Each individual particle takes a highly tortuous path—a
random walk. One such is depicted in c.
typically have the same kind of shape, making them as light and airy as
soufflé.
What’s in a branch?
A cluster grown by the DLA mechanism really does resemble a

branched electrodeposit, and this might persuade you that the two
processes are related. But mere visual similarity cannot constitute
scientific proof that there is truly any common mechanism behind
their growth. I cannot stress this point too strongly—it is a fundamental
issue in the study of pattern formation. When we see two things that
look alike, our instinct is to attribute to them the same basic cause—to
infer that at root they represent the same phenomenon. I have shown in
Books I and II that sometimes this intuition is sound, but sometimes it
isn’t. There is reason to think that the stripes on a zebra’s pelt and the
ripples of wind-blown sand stem from the same interplay of activation
and inhibition in the creation of the pattern elements. However, stripes
in an early-stage fruit fly embryo have a rather different origin. Our
facility at recognizing and comparing patterns may help us to identify
connections between systems that seem at face value to have no pro-
spect of being related. But this mental faculty is apt to tempt us into
making false correlations and untenable analogies.
How do we know when to trust these comparisons and when not?
There is no simple answer. I shall show in this book that there are
remarkable links between many of the different types of branching
structure seen in the natural world. But we cannot place much confi-
dence in such claims unless we have, at the very least, some objective,
quantitative way of characterizing the patterns we see. How do you
measure the shape of a tree or a DLA cluster?
It turns out that even forms as apparently irregular as these have at
least one measurable property that is virtually as precise, reproducible,
and characteristic as the number of legs on an insect. It is called the
fractal dimension, and is a measure of how densely packed the
branches are. It is not a foolproof means of telling us whether two
such shapes are ‘the same’ in the sense of being formed according to
the same principles; but it is pretty accurate at letting us know when
they are not. If we measure the fractal dimension of a branched elec-
trodeposit and a DLA cluster and find that these have the same value,
we cannot be sure that this is because electrodeposition is basically a

kind of DLA process. But if the fractal dimensions are different, we may
need to go back to the drawing board to explain how the electrodeposit
arises.
There is not really any way of understanding properly what a fractal
dimension is without using some maths. But I promised in Book I that
any mathematical demands I would make will be minimal, and I do not
need to renege on this promise now. Suppose that, instead of forming a
branching pattern, a DLA cluster grew so densely that there were no
gaps at all between the particles. Then it would be a solid mass whose
outer surface simply expands as the cluster gets larger. For clusters
growing on a flat surface (that is, in two dimensions), this mass would
be roughly circular. In that case, the number of particles in the cluster
(N ) would increase in proportion to the cluster’s area. For a perfect
circle, the area is proportional to the square of the radius: to r 2. If, on
the other hand, the branches were straight-line chains of particles, so that
they formed a many-pointed star like an asterisk, then the number of
particles would increase in proportion to the length of each of the arms: in
other words, proportional to the radius r of the circular envelope defined
by the arm tips. This corresponds to an extremely open, ‘empty’ cluster.
Now, a real DLA cluster like that shown above lies somewhere between
these two extremes: N increases as the cluster ‘size’ r increases, but neither
in proportion to r nor to r 2 (that is, to r raised to the power 1 or 2). Rather,
N grows in proportion to r raised to a power (exponent) between 1 and 2.
This means that the DLA cluster fills up the two-dimensional plane rather
more densely than the asterisk-like cluster but less densely than a com-
pact, circular cluster. The exponent, which is the fractal dimension, can
be calculated by analysing the density of particles in the cluster, and for
two-dimensional DLA clusters it has the value 1.71.
Put this way, the fractal dimension may sound unremarkable—merely
a number that falls out of the model. But it in fact implies that there is
something odd about a DLA cluster: in a sense, it lies ‘between dimen-
sions’. We are used to living in a three-dimensional world, in which
objects have ‘bulk’—they have a volume, enclosed by surfaces. There
are objects in the world that are to all intents one- and two-dimensional
too. Laid out straight, a piece of string is one-dimensional: you could say
that it has ‘length’ but no ‘width’ or ‘height’. Of course, it does have width
and height, but these are negligible in comparison to the length. The
piece of string is strictly speaking a three-dimensional object in which
two of the dimensions are reduced to almost nothing; if the string were
infinitely thin, it would be truly one-dimensional. Likewise, a sheet of
paper extends in two dimensions but has negligible extent in the third
(the thickness)—it is more or less a two-dimensional object. But a DLA
cluster is neither like a piece of string nor like a sheet of paper: it is
neither one-dimensional nor two-dimensional, but 1.71-dimensional.
What does that mean? We will look into this question later, but for
now it will suffice to say that it challenges the notion that the object has
a boundary in any meaningful sense. A piece of paper has edges, and a
piece of string has ends: these are the places where the objects stop. But
for fractal objects, there is no well-defined place where they ‘stop’. If
you’re right at the ‘edge’ of such an object, you can never be sure
whether you are actually standing inside it or outside, because there
is no edge as such. Yes, this is very odd. Just stick with it for a while.
The fractal dimension df is a meaningful and useful property of the
DLA growth process because it is robust and dependable. It stays the
same while the cluster grows bigger and changes shape; and two dif-
ferent DLA clusters, while differing in the precise positions and convo-
lutions of their branches, will have exactly the same value of df. In this
sense, while we can talk about the ‘shape’ of a DLA cluster in vague,
metaphorical terms, perhaps the only way we can be more precise
about this shape is to characterize it by the fractal dimension.
This quantity df reflects the rules according to which the cluster was
grown. If we change these rules, for example by allowing new particles
to make a few short hops around the surface before finally sticking
irrevocably, we will very probably obtain a branched cluster with a
different value of df. Sometimes changes like this will produce very
marked changes in the appearance of the clusters—they might develop
very stout or very wispy branches, for instance. But the effect of other
changes to the growth rules might be rather subtle, so that by visual
inspection we will be unable to say whether the clusters are ‘the same’
or not. The fractal dimension provides a well defined measure by which
we can distinguish such differences.
Here’s an example. In Fig. 2.6 I show another mineral dendrite,
formed from manganese oxide in a fracture plane of a quartz crystal.
Is this the same kind of cluster as that in Fig. 2.1? By eye, you probably
wouldn’t want to place bets. But by calculating its fractal dimension, we
can pronounce confidently that the two are different—the earlier den-
drite has a fractal dimension of 1.78, whereas for the one shown here it
Fig. 2.6: A mineral dendrite inside a quartz crystal. (Photo: Tamás

Vicsek, Eötvös Loránd University, Budapest.)
has a value of about 1.51. You can perhaps see that the smaller the
fractal dimension, the wispier the cluster.
Branched electrodeposits like that in Fig. 2.3a commonly have a
fractal dimension of about 1.7,* and this can give us confidence that
their mechanism of formation does share something in common with
the DLA process. But what about mineral dendrites? You might have
guessed from their shapes alone that the DLA model offers a good
description of their formation too, but we now discover that two
*In three-dimensional growth a DLA cluster has a fractal dimension of about 2.5, while
Brady and Ball showed that electrodeposits grown in three dimensions have a fractal
dimension of around 2.43.
mineral dendrites can have fractal dimensions that differ not only from
that of a DLA cluster but from one another.
The French physicist Bastien Chopard and his colleagues have shown
that the formation of mineral dendrites can in fact be explained by a
more sophisticated version of DLA. In their model, the solution of ions
that form the mineral dendrite diffuses through cracks in the surround-
ing rock, and the ions undergo a chemical reaction when they encoun-
ter each other. The dendrite is formed of metal and oxide ions: crudely
speaking, we can imagine these two dissolved ions diffusing until they
meet and form an insoluble black deposit. This is emulated in the
model by positing two soluble chemical species A and B that move
through the surrounding medium at random and may react to form a
dissolved compound C. If enough C accumulates in a particular region,
the solution becomes super-saturated and C precipitates in the form of
a dark material D, which then stays put. If, meanwhile, a single C
particle encounters a cluster of D, it too will precipitate, becoming
stuck to the cluster. Although couched in different terms, this is actually
Fig. 2.7: Mineral dendrite patterns generated by a computer model in

which particles diffuse by random walks. When particles encounter one
another, they react to form a dark deposit. The model can generate
forms with a range of fractal dimensions: shown here are examples
with a fractal dimension of 1.78 (a), similar to the mineral dendrite in
Fig. 2.1, and 1.58 (b), close to that in Fig. 2.6. (The square boxes show
the regions selected for calculating the fractal dimension.) (From
Chopard et al., 1991.)
a reaction–diffusion model akin to those I described in Book I, where we

saw that they are fertile sources of pattern.
Chopard and colleagues found that their model generates fractal
clusters much like real mineral dendrites (Fig. 2.7). The fractal dimen-
sion of the model clusters varies according to the concentration of
species B: the researchers were able to generate simulated mineral
dendrites with fractal dimensions of 1.75 and 1.58 (close to the values
for the natural samples I have shown) by changing this concentration.
Fractals everywhere?
I can think of no better illustration than fractals of the fact that science,
like any other human activity, is prey to fashion. In the 1980s fractals were
the thing: they were celebrated in popular books, in posters and postcards
and on T-shirts. The most famous fractal object was the abstract structure
discovered in the 1970s by mathematician Benoit Mandelbrot, working at
IBM’s research centre in Yorktown Heights, New York: a bulbous black
kidney shape now called Mandelbrot set (Fig. 2.8). This beast, decorated
with wispy filaments and often sporting furiously (and spuriously)
coloured spirals, erupts across the mathematical plane in response to
a deceptively simple algebraic equation. Mandelbrot pioneered the
recognition of fractals as a new type of geometry that is abundantly
expressed in nature (there were antecedents which had hinted at this
idea). He coined the word ‘fractal’ in 1975, and in his seminal book The
Fractal Geometry of Nature two years later he showed how fractal forms
may be found throughout the natural world, from coastlines to the shapes
of plants and clouds (Plate 2). By the 1980s computer-generated fractal
landscapes and imitations of organic complexity had found their way
into films and commercials. Today these elaborate, convoluted forms no
longer possess the allure of novelty; in scientific research, where once it
was a matter of interest merely to identify a new fractal structure, this now
elicits a weary shrug of the shoulders, for scientists have grown accus-
tomed to the notion that they are ubiquitous.
What are fractals? Mandelbrot called them ‘monsters’, for they have
properties that are strange, puzzling and, to a mathematician, even a
little frightening. Mandelbrot asserted that fractals lie in the middle
ground between forms that display the familiar, regular geometric
order that we associate with Euclidian mathematics—simple shapes
Fig. 2.8: The Mandelbrot set, a mathematical fractal defined by the boundary
between the ‘basins of attraction’ for solutions to an equation. The solutions lie on a
flat plane, with real numbers running vertically and ‘imaginary’ numbers
horizontally.
such as triangles, squares, and circles—and those that seem merely

random and chaotic. They are, he said, a kind of orderly, geometric
chaos: an apt collision of contradictory terms.
If you look closely at the edges of the Mandelbrot set, you will see that
it is bordered with branching tendrils that may make you think of
mineral dendrites and DLA clusters. The link between these structures
can be made more explicit and formal. What both the Mandelbrot set
and DLA clusters, along with all other breeds of fractal structure, have
in common is that they look more or less unchanged as you peer at
them more and more closely, zooming in as though ramping up the
magnifying power of a microscope. Take a patch of a DLA cluster and
magnify it, and you will see a ramified object that looks much the same
as the whole. Do the same again to a patch of the magnified region, and
you find the same thing (Fig. 2.9). What this means is that you cannot
tell the degree of magnification purely from appearance, because there
Fig. 2.9: When a DLA cluster is magnified, it looks more or less the same at each scale. This
property is called self-similarity, or scale invariance. (Image: Thomas Rage and Paul Meakin,
University of Oslo.)
is nothing to guide you. In contrast, you can quite easily assess the scale
of an aerial photo of a town because there will be features, such as cars,
houses, roads, that provide a known measure of length. (Because, as we
shall see, many geological structures are fractal, geologists often leave a
hammer in photographs of rock faces to provide a reference scale.) This
property of keeping the same form under different levels of magnifica-
tion, which is to say, under changes of scale, is called scale invariance, or
more loosely, self-similarity. One aspect of the self-similarity of the
Mandelbrot set is that, as you zoom in on any region of the perimeter,
the ominous black bulb keeps reappearing like a malformed Russian doll.
Because of scale invariance, fractal forms have no boundary. There
are points in the plane of the Mandelbrot set that are unambiguously
inside or outside the black region, but if you are right on the ‘edge’ then
you cannot be sure which side you’re on: each time you zoom in
further, you see more of the convolutions. This can be illustrated
more clearly with reference to a well-known fractal form called the
Koch snowflake, made by repeatedly kinking a boundary line at ever
smaller scales. Take a straight line and introduce an equilateral kink in
the middle third (Fig. 2.10). Now repeat this process for each of the
straight-line segments that result, and go on doing so again and again,
each time at a smaller scale. You end up with a line that zigs and zags
somewhat like the repeatedly and symmetrically branching arm of a
snowflake. At each step, you can say precisely where the boundary line
is. But on the next step, that point in the plane may have been engulfed
by a new kink. If we continue this process an infinite number of times,
Fig. 2.10: A fractal object called the Koch snowflake is produced by introducing identical kinks into
a line at successively smaller scales.
the line is infinitely convoluted and it is impossible to say exactly where

it passes. The ‘Koch snowflake’ remains a finite shape, but its edge has
become fuzzy.
The Mandelbrot set delights visual artists because it is a literally
endless source of baroque patterns that stimulate the eye. It is a kind
of map drawn on a mathematical plane in which each point represents
a number. All the numbers within the black boundary of the set are
associated with one solution of the equation that defines the
Mandelbrot set, while all the numbers outside are associated with
another solution.* The technicoloured swirls commonly depicted in
these maps encode how quickly the numbers they encompass get
converted into one of these solutions. The details don’t matter; what
is remarkable is the way the Mandelbrot set generates patterns and
forms that have a hundred echoes: ferns, vortices, lightning. Fractal
geometry is inherently visual. Mandelbrot says that the nineteenth-
century French mathematicians Lagrange and Laplace:
*Associated with’ is a little vague. What I mean by it is the following. The ‘Mandelbrot
equation’ is essentially a prescription for generating one number (call it z2) from another
(z1): the prescription is z2 ¼ z12 þ c, where c is a constant. You begin with z1 ¼ 0 and calculate
z2, and then use that number as z1 and calculate a new z2, and so on. Eventually you find
that your z2 either heads towards infinity or it doesn’t, depending on the value you choose
for c. The Mandelbrot set is the set of all numbers c that don’t give you an infinite solution as
you keep iterating the equation. The reason the Mandelbrot set is two-dimensional is that c
can be either a real or an imaginary number—real numbers lie along the horizontal axis,
imaginary ones along the vertical axis. Imaginary numbers are multiples of the square root
of minus 1, but if this is unfamiliar, it needn’t detain us a moment longer.
once boasted of the absence of any pictures in their works, and their lead was
followed almost universally. Fractal geometry is a reaction against the tide, and a
first reason to appreciate fractal geometry, because of the ‘characters’ it adds to
the ‘alphabet’ Galileo had inherited from Euclid, [is that they] often happen to be
intrinsically attractive. Many have promptly been accepted as works of a new
form of art. Some are ‘representational’, in fact are surprisingly realistic ‘forger-
ies’ of mountains, clouds or trees, while others are totally unreal and abstract.
Yet all strike almost everyone in forceful, almost sensual, fashion.
It must be said that this is a double-edged virtue. Some mathematicians

sniffily dismissed the fad for fractals as a kind of fancy computer-
graphic doodling, nothing to do with their precise and intricate art of
numbers. And I think it must be said that the ‘art’ that emerges from
fractals generally does not seem very far away from the tiresome kitsch
of rock album covers and psychedelia.*
But there are other reasons to be wary of making fractal geometry the
key to natural form. Once you get the hang of what a fractal structure
looks like, you imagine you see them everywhere. Without doubt the
branched fractals typified by DLA clusters do represent one of nature’s
universal forms, and are splendid examples of how complex, ‘life-like’
shapes can be the product of relatively simple and entirely unbiogenic
processes. But we must remember that not all branched patterns are
fractal; and perhaps more importantly (since researchers had, in the
first flush of fractal frenzy, a tendency to forget it), just saying that a
structure is fractal doesn’t bring you any closer to understanding how it
forms. There is not a unique fractal-forming process, nor a uniquely
fractal kind of pattern. The fractal dimension can be a useful measure
for classifying self-similar structures, but does not necessarily represent
a magic key to deeper understanding.
Furthermore, what we mean by ‘fractal’ in the natural world is by no
means clear. We saw above that the defining property of fractals—the
feature that gives them a fractal dimension—is self-similarity, or invari-
ance under different levels of magnification. For the Mandelbrot set,
this invariance continues however closely we look. The same applies to
the theoretical Koch snowflake. But, of course, for real objects the fine
details have to stop somewhere, if only because eventually we reach the
scale of atoms. In fact, for most ‘real’ fractals the self-similarity stops
well before that. For a branched metal electrodeposit we have seen that
the building blocks are tiny crystals made up of millions of atoms: at the
*I have discussed the aesthetics of ‘fractal landscapes’ in Nature 438 (2005): 915.
scale of these crystallites the surface of the deposit becomes flat, cor-
responding to a crystal facet. And for many of the popular examples of
‘natural fractals’, such as ferns, there are only a few self-similar repeti-
tions of the branching structure. Below the scale of the smallest leaflets,
the leaf structure is no longer fractal, no longer between one- and
two-dimensional, but fills up space entirely: the fern becomes a fully
two-dimensional object. Snowflakes are similar: typically their six arms
have side-branches, and there the branching stops.
Many researchers suggest that, while natural fractals clearly cannot
be truly fractal at all size scales, to qualify for the designation they need
at least to be self-similar over size scales of several factors of ten—say,
for magnifications of up to a thousandfold. But David Avnir of the
Hebrew University of Jerusalem in Israel and his co-workers have
shown that most real objects declared ‘fractal’ tend to lose their self-
similarity at little more than ten times magnification, or a hundred
times at best. This often means that the object in question simply has
a rough surface. Nature’s geometry is fractal, the researchers say, only
because people have become content to call such irregularity ‘fractal’
and not because it corresponds in any meaningful way to Mandelbrot’s
original definition of mathematical fractals in all its recursive infinities.
That’s as may be. But for highly branched objects like mineral dendrites
and DLA clusters, it is quite possible to ascribe a meaningful fractal
dimension so long as we view the objects at scales where the high degree
of branching is evident. In this case, the fractal dimension provides a
good way of making comparisons between patterns that ‘look’ alike. And
I shall continue to use the word ‘fractal’ to describe them.
Squeeze patterns
The Spanish physicist Juan Manuel Garcia-Ruiz was assailed by fractal

branches even as he sat down for a quiet cup of coffee in the Hotel Los
Lebreros in Seville. Across the broad plate-glass windows of the coffee
shop crept three of Mandelbrot’s monsters, like ghostly plants growing
within the glass (Fig. 2.11). Each window pane was made from three
laminated glass sheets, separated from one another by thin plastic
films. The laminates were imperfectly sealed at their edges, so air
could find its way between the sheets. In the Spanish heat, the plastic
had become soft and viscous, and the air had pushed its way through
Fig. 2.11: Viscous fingering patterns in the layered window pane of

the Hotel Los Lobreros in Seville. These patterns are formed as air
penetrates into the plastic film between the glass planes. (Photo: Juan
Manuel Garcia-Ruiz, University of Granada.)
the film in a bubble whose advancing front broke up into a fine tracery
of fingered branches, just like the tendrils of DLA clusters. Recognizing
this characteristic pattern, Garcia-Ruiz took photos and analysed them.
He found that the fractal dimension of the bubble was about 1.7.
The process in which air forces its way under pressure into a viscous
medium as a branching bubble is called viscous fingering. It has been
studied a great deal, because it is relevant to some very practical
problems in engineering. For instance, oil is often extracted from oil
fields by injecting water through a borehole into the oil-saturated
porous rock. The idea is that the water, which does not mix with oil,
should advance in a front that pushes the oil to the wells at the edge of
the field. But if viscous fingering occurs, the water front breaks up into
narrow fingers and the efficiency with which oil is displaced and recov-
ered is very low.
Does this sound familiar? Viscous fingering is essentially the same as
the process described by Jean-Jacques Scheuchzer in the early eight-
eenth century as an analogue of how mineral dendrites form (page 29).
In Scheuchzer’s process the air was not forced into the liquid under
pressure but was pulled in by the vacuum created when two plates
separated by a liquid film are pulled apart. That, however, is basically
the same thing.
It is no coincidence that viscous fingering produces forms similar to
those made in DLA, even though at face value the phenomena seem
quite different—for, like DLA, viscous fingering involves an instability

that makes bulges at the interface prone to elongation. We have seen
that snowflakes and other dendrites are also governed by this kind of
effect, in the form of the Mullins–Sekerka instability. For viscous fin-
gering, the origin of the branching instability was identified in 1958 by
P. G. Saffman and Geoffrey Taylor, who studied viscous fingering using
an apparatus devised by a nineteenth-century British naval engineer
named Henry Hele-Shaw. He was aiming to study how water flows
around a ship’s hull, but his equipment, now called a Hele-Shaw cell,
has now become a standard tool for research into branching patterns. It
consists of two horizontal, parallel plates with a narrow gap between
them. The top plate is made of some transparent material and has a
hole in the centre, through which a relatively inviscid fluid (such as air
or water) can be injected into a more viscous one (such as glycerine or
an oil) held between the plates (see Appendix 1).
The edge of the air bubble (say) moves forward into the oil because
the pressure in the air just behind the interface is greater than that in
the oil just in front of it. The speed at which the interface advances
depends on how steep this pressure gradient is. This is entirely analo-
gous to the role of the temperature gradient in the Mullins–Sekerka
instability; and by the same token, the gradient is steeper at a bulge,
rising to its highest value at the tip. So again there is a self-amplifying
process in which a small bump formed at random at the interface
advances faster and becomes more drawn-out. This is the so-called
Saffman–Taylor instability.*
The analogy between viscous fingering and DLA carries through to
the mathematical level: the equations describing them are equivalent.
Both these and the Mullins–Sekerka instability correspond to a set of
equations derived from the work of the eighteenth-century French
mathematician and scientist Pierre-Simon Laplace, which earns them
all the generic name of Laplacian instabilities. Perhaps if Laplace
had been able to think about maths more pictorially, he would have
discovered the origin of these fractal branching patterns nearly two
centuries earlier.
*In view of Scheuchzer’s early work, and its elaboration by the Abbé de Sauvages in 1745
(who proposed Hele-Shaw’s arrangement of a liquid injected into a second, more viscous
one), the French physicist Vincent Fleury has proposed that this be renamed the Scheuchzer–
Sauvages instability.
However, tenuous fractal patterns resembling those of DLA occur in

viscous fingering only under rather unusual conditions. More commonly
one finds a subtly altered kind of branching structure: the basic pattern or
‘backbone’ of the network has a comparable, disorderly form, but the
branches themselves are fat fingers, not wispy tendrils (you can see this
in Fig. 2.11). And under some conditions the bubbles cease to have the
ragged DLA-like form at all, instead advancing in broad fingers that split
at their tips (Fig. 2.12a). This sort of branching pattern is called the dense-
branching morphology, and it is not really fractal at all: it fills up the
available space almost entirely, and so has a fractal dimension close to 2.
Why, if the same basic tip-growth instability operates in both viscous
fingering and DLA, do these different patterns result?
All viscous-fingering patterns differ from DLA in at least one import-
ant respect: they have a characteristic size scale, defined by the average
width of the fingers. This can be seen most clearly for low injection
pressures, when the bubble front advances slowly with an almost
regular undulation around the perimeter (Fig. 2.12b). In other words,
the pattern now has a particular wavelength or scale, and so it is no
longer a scale-invariant fractal.
This scale arises for the same reason that the branches in snowflake-
like dendritic growth have a particular width: surface tension makes it
costly to form an interface, limiting the minimum size of the crenella-
tions. The fat fingers represent a compromise between the Saffman–
Taylor instability, which favours the growth of branches on all length
scales, and the smoothing effect of surface tension, which washes out
bulges smaller than a certain limit. In the DLA model, on the other
hand, the cluster has essentially no surface tension, and so the
branches may become as thin as they can be, equal to the width of
the aggregating particles themselves. A wispy DLA-like ‘bubble’ can be
produced experimentally in the Hele-Shaw cell by using fluids whose
interface has a very low surface tension. Alternatively, this can be
achieved by making the growth of the bubble more ‘noisy’—more
prone to random fluctuations that encourage the appearance of new
branches at the bubble front. A simple way of doing this is to score
grooves at random into one of the cell plates until it is criss-crossed by a
dense network of disorderly lines (Fig. 2.13a). If, on the other hand,
these grooves are arranged in an orderly grid, that imposes an under-
lying symmetry on the bubble. For a hexagonal, honeycomb grid, the
resulting bubble has a snowflake-like shape (Fig. 2.13b).
Fig. 2.12: The dense-branching morphology in the Hele-Shaw cell (a).

This is only marginally fractal, with all fractal dimension close to 2. At low
injection pressures, the advancing bubble has a rather smoothly
undulating edge with a fairly well-defined wavelength (b). This is no longer
a fractal shape at all (Photo: a, Eshel Ben-Jacob, Tel Aviv University.)
Fig. 2.13: In a Hele-Shaw cell in which the bottom plate is

‘randomized’ with a criss-crossing web of grooves, the bubble has a
shape akin to a cluster grown by diffusion-limited aggregation (a). If the
grooves form a regular honeycomb lattice, on the other hand, the
bubble resembles a snowflake (b). (Photos: Eshel Ben-Jacob, Tel Aviv
University.)
This is the attraction of the Hele-Shaw cell: it provides a tool for

investigating how a diverse family of branching patterns—DLA-like,
viscous fingering, the dense-branching morphology and dendritic
snowflakes—are related. A bit of noise tips the pattern one way; a
dose of symmetry does something else. The branches are the result of
Laplacian growth instabilities that amplify small wobbles of the sur-
face; surface tension can moderate that tendency. The results may be
fractal, but are not necessarily so. These ‘trees’ are a subtle blend of
enthusiasm and restraint operating at the border of order and chaos.
Life in the colonies
In defiance of our long-standing intuition, then, there need be nothing

‘organic’ about branching structures. Yet, in a curious inversion, the
ideas that have been developed to explain branches in minerals and
bubbles are now being exported back into biology, to account for the
complex shapes of communities of simple organisms such as bacteria.
Forms with ruffled, frond-like boundaries that radiate from a central
focus are familiar in the microbiological world: we see them, for in-
stance, in the way lichen spreads across rock (Fig. 2.14a). These fringed
blobs are also found in colonies of malignant cells that grow into
tumours (Figure 2.14b). The shape as a whole is not a fractal, but the
boundary is: whereas the smooth perimeter of a circle is one-dimen-
sional, the tumour growths have roughly 1.5-dimensional edges. This
kind of shape was first described mathematically in 1961 by the math-
ematician M. Eden, in one of the first examples of a computer model of
biological growth. Eden was seeking to model how tumours develop,
but any growth process that leads to such wavy-edged, dense shapes is
now said to be Eden-like. In Eden’s model, particles (cells, say) multi-
plied on a regular grid, with one cell per grid site. New particles were
added at random to sites adjacent to ones already occupied, so that the
cluster was constantly accumulating new particles around its perim-
eter. This is rather similar to Witten and Sander’s DLA model, and in
fact they acknowledged this debt to Eden. But in DLA the particles drift
onto the cluster from the outside, rather than being effectively ‘pushed
out’ from the surface, which makes the clusters more wispy and less
compact.
Fig. 2.14: The growth of lichen colonies (a) and of some colonies
of tumour cells (b) have a compact, roughly circular shape with a
ragged, fractal fringe. This is called Eden growth. (Photos: a, Ottmar
Liebert; b, Antonio Bru, Universidad Complutense, Madrid.)
In the late 1980s, Mitsugu Matsushita and H. Fujikawa at Chuo

University in Japan saw Eden-like shapes in colonies of Bacillus subtilis
cultured in flat, circular Petri dishes containing a water-saturated gel
made of a substance called agar (Fig. 2.15a). They injected a few
bacteria into the centre of the dish, added some of the nutrients needed
for growth, and let nature take its course.
By varying the conditions of growth, they found that they could
obtain colonies with very different shapes. The bacteria cannot pene-
trate into the gel, and so the colony has to push back the gel boundary
as it grows. This is comparable to the injection of a fluid into a
Fig. 2.15: Bacterial colonies in a plate of agar gel can grow into
several different shapes depending on the growth conditions, such
as the availability of nutrients and the hardness of the gel. Here are
several examples in Bacillus subtilis: Eden growth (a), DLA-like
growth (b) and the dense-branching morphology (c). (Photos:
Mitsugu Matsushita, Chuo University.)
Fig. 2.15: (Continued).
Hele-Shaw cell, and the ‘injection pressure’ can be varied by changing

the concentration of agar in the growth medium: the more there is, the
harder the gel, which can vary in consistency from jelly-like to rubbery.
And the harder it is, the greater the resistance to growth of the colony.
The shape of the colony also depends on the amount of nutrient
available to the bacteria.
Below a certain nutrient concentration, the pattern switches from
Eden-like to a now-familiar fractal form (Fig. 2.15b), reminiscent of
DLA clusters and sharing with them the same fractal dimension of
about 1.7. If, meanwhile, the gel is made softer at these low nutrient
levels, the pattern changes from DLA-like to one that more closely
resembles the dense-branching morphology (Fig. 2.15c). If there is
plenty of nutrient and the gel is soft, the colony expands in a dense
mass with a roughly circular perimeter.
The growth of a bacterial colony is clearly much more complex than
the aggregation of inanimate particles or the expansion of a bubble.
The cells eat, they replicate, and they may move around. Yet the growth
patterns that result are the same. It seems that this process, too, is
governed by branching instabilities.
But there are important differences. The Japanese researchers found,

for example, that if the gel gets too hard then the bacteria simply cannot
move. Under a microscope, they could see that bacteria in the dense-
branching colonies were swarming about, while those in the DLA-like
and Eden-like colonies just sat there. If they grew colonies of Bacillus
mutants that lacked the whip-like appendages which enable them to
swim around, only the DLA and Eden patterns were formed no matter
how soft the gel was.
In Book II we saw how models of collective motion can explain the
coherent swarming behaviour of organisms ranging from single cells to
fish and birds. One of the first of these models was devised by the
Hungarian physicists Tamas Vicsek and András Czirók, who showed
that complex behaviour can result when the individual organisms
follow a set of rather simple rules dictating how their actions are
influenced by those of their neighbours. In the mid-1990s they teamed
up with Eshel Ben-Jacob and his co-workers at Tel Aviv University to
apply similar ideas to the growth of bacterial colonies.
They assumed that the most significant facts to include in a model
like this are that the bacteria move, feed and reproduce. So they
adopted the following rules:
1. The bacteria move at random.

2. While food is available, the bacteria feed at a steady rate.
3. If they eat enough, they reproduce (one cell splits into two); if the
food runs out, they stop moving.
4. The dispersal of food (nutrient) throughout the system takes place
by diffusion.
A single colony might contain as many as ten billion individual ‘par-

ticles’ (cells), which is far too many for a computer to cope with. The
researchers therefore grouped cells together into ‘walkers’, each con-
taining many thousands of cells, and assumed that these walkers
moved around according to the same rules. Thus the walkers, not
individual cells, were the fundamental particles of the model. They
moved about on a regular underlying grid, and the boundary of the
colony could advance onto a new grid point only when that point had
been struck a certain number of times by the moving walkers. By
varying this number, the researchers could simulate the effect of mak-
ing the gel harder.
With nothing more than these elements, they found that their com-
puter model of the growing colony produced the DLA-like and dense-
branching patterns seen in the experiments. The lower the concentra-
tion of food, the more tenuous the branches become (Fig. 2.16).
But in the experiments a curious thing happens when the amount of
nutrient is very low: the colony suddenly becomes denser again. This
does not happen in the model—the branches just go on getting thinner
as food becomes scarcer. The researchers figured that it is at this point,
when things look really desperate, that the starved bacteria start to do
something only living ‘particles’ can do: talk to each other. As we saw in
Book I, the language of bacteria is a chemical one: they communicate
by emitting chemicals which then guide the cells’direction of motion in
a process known as chemotaxis. This enables B. subtilis to aggregate
into clumps, where the hitherto identical cells take on distinct roles,
rather as though they have become a multicellular organism. Some
become spores, which remain in suspended animation until conditions
are more favourable.
So the researchers added to their model a simple description of
chemotaxis. But they made the chemical signal a repellant rather than
an attractant: any walkers that encounter it have a tendency to wander
Fig. 2.16: A computer model of bacterial growth that applies only a few simple rules to the
movement and multiplication of cells generates DLA-like branching patterns that become
increasingly sparse as the gel medium becomes harder (bottom to top) or as the nutrient
concentration decreases (right to left). (Image: Eshel Ben-Jacob.)
Fig. 2.17: There is a change in growth morphology of Bacillus subtilis

from the DLA-type pattern (Fig. 2.15b) to a denser branching pattern (a)
at low nutrient levels. When chemotaxis is included in the computer
model, it reproduces this switch (b: the image on the right is obtained
for a lower nutrient concentration than that on the left). (Photo and
images: Eshel Ben-Jacob.)
away from the source. Cells were assumed to emit the chemo-repellant
if they became immobile due to lack of nutrients. This addition to the
model produced the denser branching patterns at very low nutrient
levels (Fig. 2.17), as seen experimentally. There is, however, no evidence
that Bacillus subtilis really employs repulsive chemotaxis, so one can-
not yet be sure that this apparent success of the model is not just a
happy coincidence.
Invasion of the mutants
Ben-Jacob and his co-workers not only studied bacterial growth theor-
etically; they also learnt the microbiological techniques needed to grow
their own colonies of Bacillus subtilis. While some of their experiments

yielded the same kind of growth patterns as Matsushita and Fujikawa
had seen, they also found some new ones. Occasionally, a colony that
was steadily advancing in one pattern would suddenly sprout an en-
tirely different kind of sub-colony (Fig. 2.18). If cells from such an
outgrowth were extracted and used as the seeds of a new colony, this
would exhibit the new pattern, too. It was a mutant colony.
For unlike aggregating metal atoms or smoke particles, bacteria can
mutate: random errors in duplicating a cell’s DNA when it divides
spawn genetic variations in the progeny. These mutations happen all
the time; some are fatal, others have no observable effect. But just
occasionally a mutation will make the new cell better adapted, more
able to replicate efficiently, than the parent cell. This is exactly how
Darwinian natural selection works.
Ben-Jacob and his colleagues suggested that what they were seeing in
these spontaneous changes of growth pattern was natural selection in a
Fig. 2.18: A mutant colony with a new growth pattern sprouting from a dense cluster of Bacillus
subtilis. (Photo: Eshel Ben-Jacob.)
Petri dish. The explosive growth and dominance of the new pattern
signalled the superior fitness of the mutants—although whether the
new pattern was a cause of this or an incidental side-effect was not
clear.
This process supplied a variety of new forms. When a mutant colony
emerged, the researchers would breed its cells to obtain a new strain of
Bacillus with new pattern-forming behaviour. Some of the mutant
patterns were familiar: dense-branching colonies, for example, which
Ben-Jacob and colleagues called the tip-splitting or T morphotype.
But other mutant patterns were unlike anything seen in non-living
systems. One consisted of elegant hook-like twists that all curved in
the same direction, creating a colony reminiscent of a Chinese dragon
(Fig. 2.19a and Plate 3a). They dubbed this the chiral or C morphotype
(‘chiral’ derives from the Greek word for hand, as these hooks can twist
either in a left- or right-handed direction). Meanwhile, a mutant they
called the vortex or V morphotype advanced as mobile, roughly circu-
lar droplets of cells that left tendrils in their wake (Figure 2.19b and
Plate 3b). Under the microscope, the researchers could see that the
cells in the droplets were all rotating in a spiral vortex—a flow pattern
that, as I describe in Book II, has been seen in groups of other
organisms such as fish. This behaviour is not unprecedented in
bacteria: in 1916 the microbiologist W. W. Ford reported such vortices
in Bacillus colonies that were consequently given the species name
circulans. Apparently the vortex mutants of B. subtilis have acquired a
similar kind of motion.
The researchers have been able to devise models of cell motion that
reproduce these patterns, but it remains very difficult to prove that
such models really capture the right biological ingredients, rather
than just generating a coincidental similarity of form. The biochemist
Jim Cowan has some harsh but reasonable words to say about people
who attempt to develop simple models of complex systems like this:
‘They say ‘‘Look, isn’t this reminiscent of a biological or physical phe-
nomenon!’’ They jump in right away as if it’s a decent model for the
phenomenon, and usually of course it’s just got some accidental fea-
tures that make it look like something.’ It is as well never to lose sight of
this brand of scepticism, which insists that just because we can make a
pattern on a computer or in a theory, that doesn’t mean nature weaves
it using the same rules.
Fig. 2.19: New growth patterns of Bacillus bacteria: the chiral (a) and
vortex (b) modes. (Photos: Eshel Ben-Jacob and Kinneret Ben Knaan,
Tel Aviv University.)
Urban sprawl
Bacteria are not, of course, the only organisms that grow in colonies
which depend on food and other resources and which suffer if the
population becomes too dense. For many of us, this sounds a little bit
like home.
Indeed, the notion of city as organism is an old one. To Lewis
Mumford, whose 1938 book The Culture of Cities was for a long time
the bible of progressive urban planners, ‘the growth of a great city is
amoeboid . . . the big city continues to grow by breaking through the
edges and accepting its sprawl and shapelessness as an inevitable by-
product of its physical immensity’. The American urban theorist Jane
Jacobs insisted in her analysis of the decline of US cities in the 1950s
that they should be considered as living organisms with their own
metabolism and modes of growth, a notion that led her to become
one of the first people to recognize how complex systems of many
interacting parts can display orderly, self-organized behaviour.
It takes an eye receptive to the character of organic form to notice this,
for it is all too easy to regard the city as an amorphous chaos (Fig. 2.20a).
In this fragmented, irregular cluster of little units, it is hard to discern any
sign of the regularity that urban planners might try to impose. Instead,
this structure is reminiscent of that one sees when small particles aggre-
gate at random (Fig. 2.20b), rather like the clumping of dust and soot
or the flocculation of river silt. As we have now seen, however, such
processes do generate characteristic forms, with boundaries that
are more or less branched and ramified, and which may be analysed
mathematically using the techniques of fractal geometry.
Fig. 2.20: The shape of a city like London, shown here as a

map of employment density (a), is highly irregular. Rather
similar shapes can be seen in the aggregation of microscopic
plastic spheres suspended in a liquid (b). (Images: a, Michael
Batty, University College, London; b, Arne Skjeltorp, Institute for
Energy Technology, Kjeller.)
With this in mind, the British geographers Michael Batty and Paul
Longley have asked whether models based on diffusion-limited aggrega-
tion—the source of fractal, branching clusters—can tell us anything
about the growth and form of cities. This was a radical departure from
tradition. Since the major preoccupation of urban planners is with the
design of cities, they have generally attempted to analyse city forms with
those efforts in mind. And so their theories have tended to focus on cities
in whose outlines the guiding hand of human design is clearly discernible.
But hardly any cities are like this. In spite of the efforts of planners to
impose a simplistic order, most large cities present an apparently disor-
dered, irregular scatter of developed space, in which residential neigh-
bourhoods, business districts, and green spaces are mixed haphazardly.
By focusing on centres where planning has created some regularity (like
the US grid-iron street plan), urban theorists have often ignored the fact
that cities tend to grow organically, not through the dictates of planners.
The planned, geometrically ordered city has long been seen as the
ideal. As geometry became a dominant aspect of ancient Greek
thought, its influence extended beyond architecture into the way in
which the buildings themselves were arranged in settlements. The grid
street plan was evident even in the cities of Babylon and Assyria, but is
most apparent in the towns built by imperial Rome: it was a scheme
that allowed these settlements, often beginning as military encamp-
ments, to be erected quickly.
Another favourite scheme of geometrical planners was the radial or
circular city plan, in which main thoroughfares radiate from a central
hub like the spokes of a wheel. This became a popular motif during the
rationalistic climate of the Renaissance and the Enlightenment.
Christopher Wren imagined London rebuilt after the Great Fire of
1666 as a grid connecting radial centres, one of them focused on his
new cathedral of St Paul’s. But he never saw it realized, because cities
that have grown dishevelled will not tolerate an imposed order: the
jumble of streets in the old city reasserted itself faster than Wren or
anyone else could construct new ones.
The fact is that cities are not static objects but growth forms with a
logic that eludes our rectilinear geometric tradition. They are structures
that emerge out of equilibrium. Planners and urban theorists are still
coming to terms with this fact, and they view it with some ambivalence:
is it a good or a bad thing for cities to evolve this way? Should they be
allowed to grow ‘naturally’, or should we try to impose some structure
on it all? Does irregular growth mean that cities will descend into chaos,
spawn slums, and lose control of public services? Or do they grow as
they ‘need’ to, finding their own optimal paths and solutions, so that
attempts to enforce regularity just create inefficiencies and sterile, un-
neighbourly living spaces?
There is surely no universal answer, not least because it is likely to
depend on the social and economic context in which urban growth
occurs. But before we can even assess the issues, we need to have a
description of how cities grow. That is what bothered Batty and
Longley—they felt that this description was lacking. As a result, it was
impossible to predict what a city might look like five years hence,
hampering the ability of planners to anticipate the likely requirements
for transportation, water, power supplies, communications networks,
and so forth. ‘There is a need’, Batty and Longley wrote in their 1994
book Fractal Cities, ‘for a geometry that grapples directly with the
notion that most cities display organic or natural growth, that form
cannot be properly described, let alone explained, using Euclidean
geometry’. The title is a giveaway: Batty believed that the appropriate
new geometry was to be found in the concept of fractals developed by
Benoit Mandelbrot—the ‘geometry of nature’.
For decades, urban theorists have known that the structure of cities can
be described by mathematical relationships called power laws (see page
35). For example, the population density often decreases fairly steadily as
one moves outwards from the city centre, typically by some relationship
in which this density is proportional to the inverse of the distance raised to
some power. A similar relationship describes how the number of settle-
ments (cities, towns, villages, hamlets) in an urbanized area depends on
their size (in population or area, say): there are many more small villages
than there are towns, and still fewer cities, and the power law quantifies
that fact. Planners and geographers could measure these relationships,
but they could not work out why they arose from the underlying economic
and demographic processes that determine the evolution of an urban
area. They didn’t know the natural rules of urban growth.
In the early 1990s, Batty and others used the methods of fractal
analysis to deduce the fractal dimensions of cities. As we can see from
Fig. 2.20a, the boundaries of cities tend to be irregular and fragmented,
and we might imagine that they are indeed fractal objects, extending
over a two-dimensional space but not fully filling it. The studies showed
that in fact the fractal dimensions of cities span a wide range, typically
between about 1.4 and 1.9. London in 1962 has a fractal dimension of
1.77, for Berlin in 1945 it is 1.69 and for Pittsburgh in 1990 it is 1.78.
In general a city’s fractal dimension increases slowly over time, reflect-
ing the fact that more and more of the ‘free’ space between centres
of development tends to get filled in, making them increasingly two-
dimensional. Fractality extends to the characteristic networks of urban-
ization too, such as transport or power lines. The transport networks
of Lyons, Paris and Stuttgart, for example, are branched fractals with
dimensions ranging from only just over 1 (a very sparse network) to
almost 1.9. The Paris metro and suburban rail network, for instance,
has a fractal dimension of 1.47 (Fig. 2.21).
Fig. 2.21: The Paris Metro is a branched network with a fractal form. (Image: M. Daoud, CEN
Saclay.)
These numbers mean little by themselves. The challenge is to under-

stand how they come about: to look for a model of a growth process
that reproduces them. Batty and Longley realized that the mean fractal
dimension of the cities that they and others had analysed (about 1.7) is
rather close to the fractal dimension of clusters formed by diffusion-
limited aggregation (1.71). So, as a start, they decided to mimic urban
growth using the DLA model. Recall that in this model particles execute
random walks until they strike the perimeter of a growing cluster,
whereupon they stick where they hit. Batty and Longley suggested
that something similar happens as cities grow: new development
units, such as business or residential neighbourhoods, are gradually
added to the city with a probability that is greater at the city’s perimeter,
since there is more space there for development. Of course, this highly
simplistic model ignores a great deal that is important for urban devel-
opment, not least all efforts of planners to impose some order on it. Yet
the researchers found that some of the growth laws observed in real
cities are similar to those that apply to DLA clusters.
Even so, DLA clusters and fractal cities do not look much alike: cities
are typically denser, more compact. This shape can be more closely
approximated by relaxing the ‘stick-where-you-hit’ rule for DLA to
allow particles to make a few hops around the cluster’s periphery before
they become immobile. But, ideally, a model of city growth should not
just ape the form: it should make sense in terms of the physical pro-
cesses involved. We know, for example, that urban developments don’t
really bounce from site to site before finally coming to rest. Batty and
Longley explored a variation on the DLA model called the dielectric
breakdown model (DBM), which we will encounter in the next chapter
as a description of cracks and sparks. DBM clusters grow not by accu-
mulating wandering particles at their edges but by pushing their way
out from a central point. That sounds more like city growth: new
development spreads outwards to colonize the surrounding land.
DBM clusters again show growth instabilities that favour extension at
the tips. But their shapes can be ‘tuned’ from highly tenuous and
almost linear (that is, almost one-dimensional) to dense and more
circular (almost two-dimensional) by altering the strength of the in-
stability—which is to say, the likelihood that new growth happens at the
sharpest tips rather than elsewhere on the perimeter. Using this ap-
proach, Batty and Longley attempted to simulate the growth of the city
of Cardiff (Fig. 2.22). They conducted a simulation of DBM growth
Fig. 2.22: A fractal model of the growth of the city of Cardiff (a), which is
constrained by the coastline (shown here in white) and two rivers. Computer
simulations for different strengths of the branching instability (measured by a
model parameter ) produce urban clusters that are more or less dense (b–e). The
best match with reality occurs for a value of of around 0.75 (c). Bridges are
included in the locations of the real ones, to allow the city to spill over the rivers. In
these images, earlier growth is shown as lighter. (Images: Michael Batty, University
College, London.)
constrained by the local geography: by the presence of the coastline to

the south-west and the rivers Taff (to the west of the city centre) and
Rhymney (to the east). The cluster was seeded from a point between
these rivers, which acted as impenetrable barriers except at two points
where there are bridges that the cluster could ‘squeeze’ across. The
results of the model, for different strengths of the branching instability,
show that somewhat realistic approximations to the city shape can be
generated when this parameter is suitably tuned (Fig. 2.22b–e).
One of the problems with this model is that it generates only cities
that form a single large fractal cluster, densest in the centre (around the
central business district) and getting rapidly more tenuous as you move
out. But areas of development at the edges of a big city are not always
part of the main ‘cluster’: there are typically many little satellite towns,
which may be swallowed up as the city boundaries sprawl further. You
can see several clusters of this sort beyond the edges of the main cluster
in the structure of London (Fig. 2.20a). Physicists Hernan Makse, Gene
Stanley and Shlomo Havlin from Boston University felt that a different
growth model was needed to capture this sort of structure. They
allowed for the possibility of new clusters being sparked off close to
but outside of the main body. These nascent centres of development do
not tend to appear entirely at random, but are affected by what is
happening nearby: development attracts further development. If two
small population clusters grow close by one another, for instance, there
is a greater than average chance that development will spring up be-
tween them: shops to serve the new inhabitants, or local businesses keen
to gain a foothold in an up-and-coming area. In short, the growth of new
clusters is interdependent, or what physicists call correlated. Makse and
colleagues borrowed from physics a model that embodies such correl-
ations, which was originally developed to describe fluid permeation in
rocks. Within this model, new particles (representing units of popula-
tion) are added to a growing cluster at random, as in DLA; but in
addition, growth in one region enhances the prospects of growth nearby,
with a probability that falls off quite sharply with distance.
This model generates a jumbled scattering of clusters of all different
sizes. But real cities are firmly rooted to a core, usually the central
business district. So the researchers imposed the condition that the
probability of adding a new unit gets smaller the further it is from a
central point. In effect this means that the model superimposes two
growth processes: one that generates a single compact, roughly circular
Fig. 2.23: Computer

simulations of urban
growth using the
correlated percolation
model. For increasing
degrees of correlation
between the growth
units (top to bottom),
the shape changes from
more or less circularly
symmetrical (a) to
increasingly fragmented
and clumpy (b, c).
(Images: Hernán Makse,
Schlumberger-Doll
Research, Ridgefield,
Connecticut.)
city cluster, and one in which local correlations between units can
create new sub-clusters around the edges. The pattern that emerges
depends on the relative strengths of these two processes, and can vary
from a roughly symmetrical, circular distribution of units to a clumpy
form decorated with sub-clusters and tendrils (Fig. 2.23). It is clear
straight away that these latter shapes, in which the correlations are
strong, look more realistic than those produced by DLA or DBM, at
least for a city like London. The model can also do a pretty good job of
imitating how city shapes evolve over time (Fig. 2.24).
Fig. 2.24: The growth of Berlin from 1875 to 1945 (a, top to bottom) is mimicked fairly well by the
growth of a city in the correlated percolation model (b). (Images: Hernán Makse, Schlumberger-Doll
Research, Ridgefield, Connecticut.)
It seems, then, that randomness (uncoordinated local decisions

about development, akin to the disorderly aggregation of particles in
DLA) can generate something like the clumpy, messy sprawl of cities.
But randomness alone is not enough to explain everything about the
forms of cities. In Batty’s words, it is modified by a ‘deeper order’,
derived from growth instabilities and correlations and manifested in
the power laws that define the structures. This order seems to be of
precisely the same sort that we find in many natural growth patterns.
Batty suggests that rather simple concepts from DLA and DBM do a
fair job of accounting for the shapes of cities that sprung up during the
early industrial era, which tended to be single clusters organized
around the central business district. But the ‘correlated percolation’
model of Makse and his co-workers is better suited to capturing the
growth and form of post-industrial cities in which new communica-
tions technologies have made work in the centre of the city less preva-
lent and the urban landscape less centralized.
The message for planners is salutary, for it appears to question
whether centralized planning for an ‘organism’ as complex as a city
can ever succeed. In the 1960s, planners sought to influence the way in
which London encroached into the surrounding countryside with a
Green Belt policy that would restrict urbanization. Yet there is no sign
that these policies have had any effect on the city’s growth, which has
gone right on expanding to the tune of the same mathematical laws. It
will take more than this, it seems, to undermine the inexorable physics
of cities.
Just for the Crack
Clean Breaks and
3
Ragged Ruptures
F
INGAL’S Cave on the island of Staffa, off Scotland’s west coast, is
the kind of place that left early nineteenth-century Romantics
pondering the Sublime. You can sense as much in the character-
istically storm-ridden depiction by J. M. W. Turner, and also in the
luxurious harmonies of Felix Mendelssohn’s tone-poem The Hebrides,
composed after a trip to Staffa in the 1820s. This geological oddity also
made an awe-inspiring impression on Joseph Banks, president of the
Royal Society, when he sailed there in 1772 during an expedition to
Iceland. ‘Compared to this’, he said,
what are the cathedrals or palaces built by men! Mere models or playthings, as
diminutive as his works will always be when compared with those of nature.
What now is the boast of the architect! Regularity, the only part in which he
fancied himself to exceed his mistress, Nature, is here found in her possession,
and here it has been for ages undescribed.
For Banks saw that the entrance to the cave was flanked by great pillars
of rock with almost perfect hexagonal cross-sections, designs of such
regularity that they could almost have been fashioned by Platonic
masons (Fig. 3.1 and Plate 4).
There is a rather more accessible example of this striking geological
pattern at the Giant’s Causeway in County Antrim, on the coast of
Northern Ireland (Fig. 3.2 and Plate 5). In legend, these two formations
are part of the same structure: a causeway from Ireland to Scotland
built by the legendary third-century Irishman Fionn MacCumhaill
(Finn MacCool, or Fingal), a veritable giant and leader of the band of
warriors called the Fianna who guarded the Kings of Ireland. Finn made
the rocky road across the Irish Sea so that he might do battle with a rival
Fig. 3.1: The natural pillars of Fingal’s Cave on the Scottish island
of Staffa. (Photo: Lucas Goehring, University of Toronto.)
giant of Scotland named Benandonner. When he got there, however,

Finn discovered that Benandonner was a much bigger giant than he,
and so he crept sheepishly back home. But Benandonner crossed over
the causeway to find him, whereupon Finn posed as a baby, attended
by his wife. If this is the size of Finn MacCool’s baby, thought
Benandooner, how big must Finn himself be? And he ran in panic
back to Scotland, demolishing the entire causeway except its beginning
and end.*
This is the way with natural patterns: we must find an explanation for
them even if it need be magical and valorized by legend. We tell
ourselves that these things were constructed with purpose by intelli-
gent agents, for how could wild nature be capable of such craft?
Of course, nothing of that sort appears in D’Arcy Thompson’s de-
scription of the hexagonal columns of Staffa and the Giant’s Causeway.
He explained that these rock formations appear in basaltic lava as it
cools, contracts and cracks: ‘rupture . . . shatters the whole mass into
prismatic fragments’, he said. ‘However quickly and explosively the
cracks succeed one another, each relieves an existing tension, and the
next crack will give relief in a different direction to the first.’ This is all
*Some versions of the legend find Finn’s ruse too timid, and have him leap up and bite
Benandonner’s hand, chasing him all the way back to Scotland while hurling great lumps of
earth, one of which became the Isle of Man.
JUST FOR THE CRACK j 73
Fig. 3.2: The Giant’s Causeway in County Antrim, Northern Ireland (a, b). The polygonal cross-
sectional pattern of the columns contains mostly six-sided shapes, although not in a perfect
honeycomb (c). (Photos: Stephen Morris, University of Toronto.)

very fine, but Thompson does not really account for why the cracks
create such a remarkable pattern: a series of vertical columns, each
with a polygonal cross-section that seems most often to be hexagonal.
What these formations tell us with irrefutable force is that cracks are
pattern-formers. Seldom, however, does fracture produce anything
quite so regular and ordered. When we think of cracks, what comes to
mind are random branching fractures that may intersect in reticulated
networks (Fig. 3.3). These are familiar traceries, often to our dismay—
they are the marks of old age and decay, the webs of failure and regret
and disaster, all of which makes it of paramount importance to know
how they arise and what guides their wandering paths. In this much,
D’Arcy Thompson was able to make little progress. In view of what we
have seen so far concerning branching patterns, can we now do better?
Why things break
It is only recently that scientists have begun to understand why cracks

form. For a long time, the science of fracture and failure of materials
limped along with a ragbag of concepts that could do little to predict
what was seen in the real world. This was much more than an academic
embarrassment, for while scientists remained largely ignorant of what
makes a material tough they had no systematic method for devising
the strong and tough materials that technology demands. They would
earnestly apply what seemed to be sensible criteria, only to end up with
substances ‘about as strong as stale hard cheese’, in the words of British
materials scientist James Gordon. On the other hand, their experience
with new materials that were genuinely strong sometimes flew so much
in the face of what seemed like common sense that they had some
persuading to do. Gordon recalls the response of an Air Marshall during
the Second World War to the idea that Lancaster bombers were to have
glass-fibre domes: ‘Glass!—Glass! I won’t have you putting glass on any
of my bloody aeroplanes, blast you!’
We can understand the Air Marshall’s feeling about glass aeroplanes,
because we know how readily glass shatters. But one can offer appar-
ently sound scientific arguments for why few materials should be better
for making aircraft than glass. It consists of disordered silicon dioxide,
the same stuff as sand and quartz but melted to break up the regular
crystal structure and then cooled quickly so that the atoms become all
Fig. 3.3: Cracks form a variety of branching patterns.

but immobile before they can pack together in an orderly manner. The
chemical bonds between silicon and oxygen atoms are extremely
strong, not far off the strength of those between carbon atoms in
diamond. You need to expend a lot of energy to pull them apart. So
why isn’t glass nearly as strong a diamond?
Well, the Air Marshall was wrong about glass fibres—they are very
strong. But our naive chemical reasoning is wrong about glass too: it
breaks rather easily. What is going on?
A stiff, brittle material like glass is tough so long as cracks cannot get
started. But they can be launched from the tiniest of origins: a mere
scratch may act as the seed for a flaw that shoots through the whole
material. Window glass inevitably contains innumerable little scratches

on its surface, any one of which can give birth to a crack that spreads
with catastrophic speed and vigour. Gordon helped to show in the
1950s that only very minor rubbing of glass against another hard ma-
terial is sufficient to cover the surface with microscopic cracks. The
reason glass fibres are so strong is simply that they have a much smaller
surface area than a plate of window glass, and so have far fewer of these
minute flaws. The thinner they get, the fewer flaws there are. So cracks
have nowhere to start from. In analogous fashion, Galileo (who was
very interested in discovering why things break) found that the ship-
builders of Venice paid more attention to the construction of big ships
than small ones—because, they told him, big ships break more easily.
There are simply more places on a big ship where a crack might start.
But why, once a crack is initiated by a tiny imperfection, does it then
grow with such awesome speed, if the bonds between atoms are really
so strong? Knowing the energy contained in a single chemical bond in
glass, it is a simple matter to calculate what the theoretical strength of
glass ought to be, assuming that fracturing the glass means breaking all
the bonds along the crack’s path. The puzzle was that the observed
strength is typically about a hundred times smaller than this calcula-
tion suggests it should be. In the 1920s the British aeronautical engineer
Alan Arnold Griffith, working at the Royal Aircraft Establishment in
Farnborough, had a critical insight into the problem. He was at the
time laying the foundations of glass-fibre technology by drawing
heated glass rods into thin threads—work that Gordon later developed
at the same institution. Griffith found that the strength of very thin
glass fibres is extremely high, approaching the value implied by the
calculation of chemical-bond strengths. So the question was not so
much why glass fibres are strong but why normal glass in bottles and
windows is so weak. How can a minor scratch confined to the surface
be responsible for disastrous failure?
During the previous decade, the engineer Charles Inglis had investi-
gated why Britain’s iron ships were alarmingly prone to cracking apart.
He showed that if a plate of material is stressed by bending or stretch-
ing, the stresses around a hole can be much greater than those through
the rest of the material: there is ‘stress concentration’ at the flaw.
Griffith realized that the same would apply to microscopic scratches
on the surface of a brittle material. He calculated that a crack just one-
thousandth of a millimetre long and so narrow that it cleaves a single
chemical bond at a time as it advances has a stress at the tip about

200 times that elsewhere in the material. In other words, a stress 200
times smaller than that required to break the chemical bonds in the
flawless material will suffice to set bonds snapping at the crack tip.
What’s more, if the crack lengthens without changing its width, the stress
concentration at the tip gets even greater: the growth is self-amplifying.
What Griffith had demonstrated is that fracture is nonlinear : the effect
does not follow in proportion to the cause. A tiny crack turns a modest
stress into a catastrophic one.
Jagged edge
We have seen already that self-amplifying instabilities tend to have

other consequences too: not only does growth foster more growth,
but the resulting form becomes prey to random perturbations. Small,
chance irregularities generate new appurtenances: wriggles, wobbles,
branches. That is just what we find in cracks. Griffith’s work suggested
that a long, narrow crack initiated at a notch in a brittle material will cut
like a knife straight through the material when it is stressed. This,
however, is the exception rather than the rule. A glazier can make a
clean, straight break through a sheet of glass by first scoring a shallow
scratch along the path of the intended fracture, but without this guid-
ance the crack is more likely to be erratic: ragged or crazy-paved.
A typical brittle crack grows in three stages. During its ‘birth’ from an
initiating notch, it accelerates in less than a millionth of a second to
reach a speed almost equal to the speed of sound. During ‘childhood’
the crack continues to accelerate, but more slowly. All this time the crack
is straight and the fractured surfaces in its wake are smooth and
mirror-like, but once the crack speed rises above a certain threshold, a
mid-life crisis sets in: the velocity starts to fluctuate wildly and unpre-
dictably, and the crack tip veers from side to side. Then the fracture
surfaces become rough, sprouting a forest of small side branches.
In 1951 the Cambridge metallurgist Elizabeth Yoffe uncovered a clue
about why cracks may wander erratically. She found that when a crack
tip moves at virtually the speed of sound, the stress field ahead of the tip
starts to flatten out and develop bulges pointing in different directions
from the tip’s motion. These new stress concentrations could make the
tip deviate from its straight path. John Willis at Cambridge showed
in the 1960s that in fact the largest stresses at the tip of a fast-moving
crack point at right angles to its direction of motion, suggesting that the
tip should be constantly turning corners.
Yoffe’s analysis implies that windows would not shatter into jagged
shards, but would simply split cleanly, if the cracks did not travel so fast.
But that’s not necessarily so, for even cracks that grow very slowly may
take complicated paths. In 1993 the Japanese researchers Akifumi Yuse
and Masaki Sano at Tohoku University developed a method for growing
cracks that move at just a few centimetres per second, which is much
slower than those passing through a brittle material as it shatters. They
sent these slow cracks through flat strips of glass by using heat to
induce stress and lowering the strips slowly through a heater into a
bath of cold water. As anyone knows who has mistakenly put a hot glass
dish from the oven into cold washing-up water, this abrupt cooling can
shatter glass. When hot, the material expands; when cooled, it shrinks.
At the boundary of expanded and shrunken material there are large
stresses which can cause cracking. But the cracks advance only where
there is a sharp change in temperature over a small distance. By varying
the rate at which they lowered the glass strips, the Japanese researchers
were able to control precisely the speed of a crack initiated from a
notch at the bottom.
They found that for very slow speeds (about a millimetre per second)
the cracks were generally perfectly straight (Fig. 3.4a). But if either the
speed or the temperature drop between heater and water bath
exceeded particular thresholds, the crack became unstable and began
to wiggle—not at random, but in a steady oscillation with a well-de-
fined wavelength (Fig. 3.4b). In other words, you can give a glass sheet a
rather beautiful, undulating edge just by cracking it with heat and cold.
Yuse and Sano found that the wider the glass strip, the longer the
crack’s wavelength. If the strip was infinitely wide, the wavelength
would be infinite too, meaning that the crack would head off at a
fixed angle to the vertical and never look back. The researchers couldn’t
lay their hands on an infinitely wide glass strip, but they could find one
without edges: a glass tube. And here indeed the crack simply travelled
around the tube’s axis at a fixed angle, cutting out a perfect helix
(Fig. 3.4f ). Such helical cracks are rumoured (this is not easy to
check) to be found in frozen natural gas pipelines in Alaska, sometimes
winding their way around the pipes for miles.
Fig. 3.4: Growth instabilities in slowly propagating cracks through a glass plate. The crack is
initiated at a notch, and advances owing to the stresses produced as the hot plate is lowered into a
water bath. If the speed is slow enough, the crack is perfectly straight (a). At higher speeds it
becomes oscillatory, with a constant wavelength (b). As the crack speed continues to increase, first
the oscillations increase in amplitude until the sine-wave shape becomes distorted (c,d ). Then the
single crack splits into several branches (e). If a glass cylinder is used in place of the flat plate,
‘oscillatory’ cracks are not wavy but instead thread around the cylinder in a helix (f ). (Photos:
Akifumi Yuse and Masaki Sano, Tohoku University.)
As these wavy cracks get faster, the oscillations become more pro-
nounced, until finally they start to become distorted and kinked
(Fig. 3.4c, d ). And if the temperature drop is large enough, the wavy cracks
grow branches, apparently two at a time: the crack repeatedly bifurcates.
At this stage, the pattern starts to become more disorderly—more like the
classic picture of a crack—although the regular waves can still be seen.
Thin sheets of material often break in wavy patterns, although they
are not usually as regular as this. One of the most familiar is the jagged
tear made by opening a letter with your finger (Fig. 3.5). Animangsu
Fig. 3.5: Opening a sealed envelope with your finger tends to generate an oscillatory tear with
ragged, sawtooth edges.
Fig. 3.6: A carefully controlled ‘letter-opening’ experiment using a

cylindrical rod to tear a flat sheet produces a remarkably regular
sawtooth tear with ‘teeth’ that have a shape described mathematically
as a cycloid (a). This can be explained by considering how the tear
bends and curls away from the advancing rod edge, until there is not
enough energy to sustain this bending (at the tips of the cycloids) and
the sheet begins to rupture instead by stretching (b). (Photo: a, L.
Mahadevan, University of Cambridge; from Ghatak and Mahadevan,
2005.)
Ghatak and Lakshminarayanan Mahadevan of the University of

Cambridge have constructed a physicists’ version of letter-opening,
which reveals where this waviness comes from. They used rigid rods,
ranging in width from a few millimetres to several centimetres, to rip
open stiff plastic film similar to that used as transparent wrappers in
packaging. The researchers pulled the rods through the sheet at a
constant speed of up to 2.5 cm per second. For narrow rods, the tears
were straight and smooth, rather like those made by a paperknife. But
fatter rods produced evenly spaced serrations with sawtooth teeth,
which have a mathematical shape known as a cycloid (Fig. 3.6a).
Ghatak and Mahadevan deduced that this shape stems from a mix-
ture of stretching and bending as the sheet is ripped. The convex slope
of a sawtooth forms as the sheet bends back against the rod in a kind of
curling tongue (Fig. 3.6b). This ‘bending rip’ requires little energy, and
it sends the crack heading off at an increasingly oblique angle to the
direction the rod is travelling in. But the further it goes, the smaller the
region over which bending occurs, and eventually there is just not
enough ‘bend’ left. At that point the rod, still advancing and pushing
against the torn edge of the sheet, begins to stretch it, and the sheet
ruptures by being pulled apart rather than bent—which means that the
rip moves forwards in the same direction as the rod. Very quickly this
‘stretching rip’ runs ahead of the rod and so loses impetus. Then the rod
catches up and begins a new bending rip, which heads away in the
opposite direction from before. So each crest of the cycloid, where the
rip changes direction, corresponds to the switch from bending to
stretching of the sheet. The crack swings constantly from side to side,
at the same time surging ahead and then slowing down like the judder-
ing stick-and-slip of a heavy object being pushed across a floor.
Similar ruptures occur on a much larger scale when polar sea-ice
sheets are pushed past grounded icebergs by ocean currents or winds.
But apparent fractures in ice sheets can take an alternative path: occa-
sionally they are found to run back and forth in rectangular crenella-
tions that interleave like a zipper, a phenomenon called finger rafting
(Fig. 3.7a). These patterns have a rather different origin. They form
when two thin ice sheets, no more than about 10 cm thick, are pushed
against one another. If the ice sheets have ragged edges, then some of
these protuberances on one sheet will ride up over the other, pressing
down on it at the point of overlap. The effect of such an over-riding lip is
to create small corrugations on either side in both sheets, running
Fig. 3.7: Cracks in floating ice sheets can develop a zipper-like shape, a phenomenon called finger
rafting (a). This is caused by the collision of two ice sheets, in which a protrusion on one edge that
rides up over the other creates a wavy deformation to either side, generating ‘zipper teeth’ at
regular intervals (b). (Photo: a, John Wettlaufer, Yale University, New Haven; from Wettlaufer and
Vella, 2007.)
perpendicular to the boundary between them. The troughs of these

ripples in one sheet coincide with the crests in the other, so that a single
lip induces other overlaps regularly spaced to either side. These split
into the ‘fingers’ of the zipper (Fig. 3.7b). John Wettlaufer of Yale
University and Dominic Vella of Cambridge University explained all
this in 2007 and showed that the same effect can be seen in thin layers
of wax pushed together on the surface of water. It is possible that
similar crenellated structures form on immense scales along geological
faults when two tectonic plates converge.
A matter of chance
These are strange cracks, and very different from the jumble of spidery
lines we tend to associate with fracture (Fig. 3.3). The striking thing
about those patterns is that they cover a huge range of size scales: some
are visible only under the microscope, some only from satellite imagery
of geologically vast terrains (Fig. 3.8). And many crack patterns look
much the same over a wide range of scales: as we zoom in at increas-
ingly higher magnification, they don’t appear to change very much.
We merely see ever finer details that we could not make out before.
What this implies is that fracture patterns are scale-invariant fractals.
These patterns, unlike the regular wavy cracks described above,
involve a strong dose of randomness. Some of this may stem from the
nature of the cracked materials themselves: rocks are typically haphaz-
ard compactions of grains of many different sizes and shapes, welded
together at their boundaries; cement and porous rocks like sandstone
are shot through with random networks of pores; hard, brittle plastics
contain a chaotic tangle of polymer chains. But some randomness
seems intrinsic to the way a crack moves: as we saw, the path of a
crack tip is potentially unstable, weaving this way and that or growing
branches at the slightest opportunity. In this way, even materials with
perfectly ordered atomic-scale structures (crystals) can fragment in
erratic, jagged shapes.
A popular model of fracture devised in the 1980s suggests how fractal
branching cracks might thread their ways through an orderly lattice of
particles. The model was intended by its creators, Lutz Niemeyer, Hans
Jurg Wiesmann, and Luciano Pietronero at the Brown Boveri Research
Centre in Baden, Switzerland, to describe a rather specific kind of
Fig. 3.8: Fractal cracks occur on many scales. Here is the network of
fault lines that surrounds the San Andreas fault—the image
encompasses many kilometres, although you could probably just as
easily imagine it to be a diagram of cracks in, say, an old layer of paint
on a window frame.
‘failure’: not fracture in the normal sense, but the passage of a spark
through a material. In electrical devices such as capacitors, an electrical
voltage is applied between metal plates or electrodes separated by a
layer of insulating material called a dielectric. If the voltage is too big,
a spark discharge crackles between the electrodes. This is called dielec-
tric breakdown (Fig. 3.9a), and it usually burns out the device. In some
cases it is accompanied by real fracture: the material is shattered by the
flow of charge. In transparent materials the fracture pattern can leave
visible tracks: a frozen snapshot of the spark’s passage, which is in itself
a thing of considerable beauty (Fig. 3.9b). The structure has a branched,
lightning-like appearance, and indeed atmospheric lightning (Fig. 3.9c)
is itself a closely related phenomenon: air acts as an electrical insulator
between a charged cloud and the ground.
Electrical discharge patterns like these were studied in the eighteenth
century by the German scientist and writer Georg Christoph
Lichtenberg, and they are commonly called Lichtenberg figures.
Lichtenberg, working at Göttingen University, was investigating the
new science of electricity by building up intense charges of static and
Fig. 3.9: Electrical discharges are branched formations that resemble crack patterns, as shown
here in the spark pattern from an electrode on the surface of a glass plate (a). These so-called
dielectric breakdown patterns can be ‘frozen’ and preserved when the passage of current heats up
a solid substance and cracks or vaporizes it (b). The beautiful structures that result are called
Lichtenberg figures. A lightning discharge in the atmosphere is also an example of dielectric
breakdown, with a similarly branched pattern (c). (Photos and images: a, After Niemeyer et al.,
1986; b, Kenneth Brecher, Boston University; c, Michael Mortenson.)
letting it discharge through a block of resin. He found that dust was

attracted to the flow of charge, accumulating on the resin in a way that
revealed the many-fingered routes it took. Lichtenberg’s experiments
led him to invent an early form of electrostatic printing, based on the
way powder arranges itself on a plate of material with an uneven
distribution of electrical charge. He was also the first person to show
conclusively that lightning is an electrical phenomenon by carrying out
a hazardous version of the kite-flying experiment popularly (and prob-
ably apocryphally) attributed to Benjamin Franklin. He constructed a
lightning conductor at Göttingen, and corresponded with Alessandro
Volta, the inventor of the battery. (Lichtenberg noted that Volta had ‘an
expert knowledge of the electricity in girls.’*) But it is for his work as an
essayist and aphorist that Lichtenberg is best remembered—as with
many Enlightenment intellectuals, scientific research could not con-
tain his diverse energies. When Lichtenberg said of his ‘frozen light-
ning’ that it ‘transfer[s] terror into enchantment’, it is a reminder of how
undervalued today is the scientist who, with a mot juste, can distil
wonder from dry experimentation. And when he compared his elec-
trical figures with dendritic ice-fingers on a window pane in winter, we
can now see that his penchant for poetic metaphor put him unwittingly
on the right track.
Lutz Niemeyer and his colleagues chose to model the dielectric
breakdown process by considering a regular checkerboard lattice on
which charge could flow from point to point in straight lines. The
discharge advances one lattice site at a time, and there are generally
several directions it could travel next (Fig. 3.10). Which way does it go?
The researchers assumed that the discharge passes at random to any of
the next accessible points at each time step, but with a bias that
depends on the strength of the electric field at that point; in other
words, there might be (say) a 70 per cent chance that it will move to
an adjacent site where the electric field is high, and a 30 per cent chance
that it will move to a position where the field is lower. This is a reason-
able thing to assume, since the passage of the spark can indeed be
expected to be governed by the electric field it encounters en route.
*There can be little doubt that Lichtenberg and Volta had a good time together. ‘What is the
simplest method to produce a good vacuum in a wineglass without using an air pump?’,
Lichtenberg asked his friend at one point. ‘Just pour in wine! And what is then the best
method to allow the air to come back? Just drink the wine!’ He added that ‘This experiment
will seldom fail!’, suggesting that he had been careful to accumulate good statistics.
Fig. 3.10: In the dielectric breakdown model, an electrical discharge

advances between adjacent points on a regular lattice. There are in
general many different points to which the discharge could flow
next—here I show the course of the discharge in black, and the
choices for its next step in white. (The discharge began at the centre of
the image, the circled black grid point.) The next step for the discharge
is selected at random but with a probability that is biased by the
strength of the electric field at each of the candidate points. (After
Niemeyer et al., 1984.)
Here again we see a delicate balance of chance and necessity: in the

evolution of the discharge route, nothing is certain but some things are
more likely than others. There is randomness, but not that alone.
The electric field around the tips of a branching discharge is higher
than that in the valleys and clefts, so advance from the tips is more
likely than advance from the interior of the ‘spark’. That is just how it is
for DLA clusters, too, and for an entirely analogous reason: in that case,
the probability of a new particle striking the tips is higher. It is no
surprise, then, that the dielectric breakdown model produces ramified,
tenuous discharge patterns that look very much like DLA clusters
(Fig. 3.11a). If the probability of the discharge flowing to a new site is
assumed to vary in direct proportion to the electric field at that site,
then the patterns predicted by this model have a fractal dimension of
1.75—almost the same as that of DLA.
Fig. 3.11: The dielectric breakdown model generates branching

fractal patterns very similar to those seen in diffusion-limited
aggregation (a). They have a fractal dimension of about 1.75. This
model can be adapted to describe the propagation of cracks in a brittle
material, represented as a lattice of particles linked by bonds. If the
bonds are considered to be a little elastic, able to stretch and relax in
response to stresses, the branching pattern is more tenuous (b).
(Images: a, Luciano Pietronero, University of Rome ‘La Sapienza’;
b, Paul Meakin, University of Oslo.)
The dielectric breakdown model can be imported wholesale into a

theory of fracture in disordered materials. All this entails is to regard the
discharge as a crack, and the lattice as a network of interconnected
atoms or particles joined by chemical bonds. At each time step, the
crack takes a pace forward as a new bond breaks, and the probability of
that happening at any site adjacent to the crack’s periphery depends on
how big the stress is there. The stress field varies in just the same way as
an electric field: it is greatest at the tips of the crack and is smaller in the
crevices, just as Griffith said.
But this is a very simplistic picture of fracture. For one thing, it insists
that a bond must break with every time step—the only question is
which it will be. In reality, though, there is no reason why that should
happen if the stress simply isn’t large enough. A better model could
allow bonds to stretch a little without breaking, so that they are not like
rigid rods but more like springs. In that case, each time a bond breaks it
will release local stress as the surrounding bonds relax. ‘Elastic’ models
like this generate a range of different fracture patterns, depending on
what is assumed about the bonds’ elasticity. The example in Fig. 3.11b
shows a crack that has a much less dense network of branches than
those generated by the rigid-bond dielectric-breakdown model, and
looks rather more like the kind of pattern you might finds creeping
ominously across the ceiling. The fractal dimension here is 1.16,
confirming that the crack is less like a two-dimensional cluster and
more like a wiggly line.
Patterns in the dry season
All these cracks start at a single point; but that is not always the way it
works. In the hard mud of a dried-up pond, there is no centre to the
network of cracks: the rupturing happens everywhere at once, creating
a reticulated web (Fig. 3.12). As the wet mud of the pond bed becomes
exposed and dried, water withdraws from between the silt particles and
they draw closer together. The whole surface layer contracts. But it
cannot simply shrink like a dried-up piece of fruit, because this dry
layer adheres to the damper mud below, which is still expanded by
water. This means that stresses build up everywhere in the surface
layer. When the stresses get big enough, cracks start to appear, creeping
through the hard mud and intersecting to carve it up into islands.
Fig. 3.12: When a thin layer of material is stressed as it shrinks, it may fragment into a series of
islands. This process is familiar in the mud on the bed of dried-up ponds and lakes. (Photo: Sean
McGee.)
This kind of cracking due to uniform shrinkage (or expansion) of a

thin layer of material ‘pinned’ by adhesion to another surface is very
common. It is apt to afflict a coat of paint as the material on which it sits
expands or contracts: a piece of wood swells or shrinks as the humidity
changes, say, or metal expands as it warms. This is the cause of the
mesh of hairline cracks that weaves across old paintings, known to art
historians as craquelure. The resulting patterns offer a subtle finger-
print of authenticity: they may vary, for example, according to where
and when the picture was painted—there are French, Italian, Dutch
and Flemish ‘styles’ of craquelure. The cracking pattern provides
clues about the artist’s materials and techniques, or the history of the
painting: how it was handled, transported, and changes in its ambient
environment. So art conservators have developed sophisticated
methods for digitally scanning paintings to identify and classify the
craquelure pattern. Forgers, meanwhile, have attempted to mimic it
by baking their works; they have even been known to resort to the crude
method of painting a fine web of craquelure by hand—a deception that
might fool the eye of a careless buyer but which would be immediately
obvious under a magnifying glass.
Cracking of thin layers poses a big challenge for several advanced
technologies. Surface coatings that are deposited ‘wet’ to protect or
modify a material, for example conferring wear-resistance or low re-
flectivity, may shrink as they dry or cool, threatening to leave them
cracked and flaking. Integrated microelectronic devices often incorp-
orate a thin film of one crystalline material (an insulator perhaps) laid
down on top of another (a semiconductor, say) in which the spacing
between the atoms is slightly different, forcing the top layer to expand
or contract. So there are good reasons for wanting to understand what
determines the crack patterns that form.
Arne Skjeltorp from the Institute for Energy Technology in Norway
has explored this type of fracture with what one might call an ‘ideal
mud’: a suspension of microscopic spheres of polystyrene in water. All
the microspheres are the same size, each measuring just a few thou-
sandths of a millimetre in diameter, and Skjeltorp trapped a single layer
of them between two sheets of glass and let the water slowly evaporate.
The particles clump together just like silt particles in pond mud, and
the layer contracts as water disappears. But the identical size and shape
of the spheres means that they pack together more regularly than silt,
forming orderly hexagonal arrays that are also good mimics of atoms
packed in crystalline films.
Skjeltorp found that as the layer dries, it fractures into complex ‘crazy
paving’ patterns like those of dry mud (Fig. 3.13). The web of cracks
looks similar at different scales of magnification: again it is fractal, with
a fractal dimension of about 1.68. But there are signs of orderliness
here: the cracks have preferred directions, tending to intersect at angles
of 1208, particularly in the early stages of drying. This simply reflects the
symmetry of the underlying lattice of particles: the cracks tend to open
up between rows of particles. The particles in mud are packed together
in a much more disorderly fashion, and so the shapes of the final
islands are less regular. The physicist Paul Meakin at the University
of Oslo has adapted the ‘elastic’ dielectric breakdown model to this
situation, and found that it produces crack patterns similar to those
observed (Fig. 3.14).
But many more familiar examples of cracking are rather different
from this. For mud drying on a lake bed, paint and varnish drying on
canvas or wood, or a ceramic glaze hardening and ageing on a piece of
Fig. 3.13: The cracks in a layer of microscopic plastic particles suspended in water—a kind of
‘ideal mud’—as it dries. Because the particles are identical in size and packed in a hexagonal array,
the cracks tend to follow the lines between rows of particles and therefore intersect at angles of
about 1208. This is particularly evident in the early stages of cracking (a). The final crack pattern
(b, c) looks similar at different scales of magnification, at least until we reach the scale at which
individual particles can be seen (c). The region in b is about one millimetre across; that in c is ten
times smaller. (Images: Arne Skjeltorp, Institute for Energy Technology, Kjeller.)
pottery, the cracking layer is fixed to a surface on one side while being
freely exposed to air on the other. In such cases, the crack pattern often
takes on a rather different appearance (Fig. 3.15a). Some cracks seem to
advance in straight or curved lines without splitting or bending sharply
for long distances, while other cracks divide up the space in between.
This breaks the material into fragments that typically have four sides,
many of which are square or rectangular. In the final patterns these
domains have a ‘typical’ size that depends on the thickness of the
cracked layer—which means that this crack pattern is not a fractal.
And despite the right-angled nature of the junctions, the pattern is
not simply a square or rectangular grid, like the grid-iron street plan
Fig. 3.14: A modified form of the dielectric breakdown model can

reproduce the fracture patterns seen in contracting thin films. (Image:
Paul Meakin, University of Oslo.)
of Manhattan: the junctions are staggered, so that domains tend to

share edges not with four but with six neighbours.
Steffen Bohn of the Rockefeller University in New York and his
co-workers have shown that in this case the crack network forms in
stages: first by the appearance of long, smooth cracks, and then by the
successive division of the space in between them as smaller cracks
bridge the gaps (Fig. 3.15b). When an advancing crack intersects with
an existing fracture line, stress is relieved most effectively if the junction
makes a right angle. The new crack may bend in order to satisfy this
condition. Bohn and his colleagues have shown how this pattern of
cracks can be analysed to deduce the order in which the cracks formed,
and thus to reconstruct the history of the fracturing.
The process is not unlike the way new urban roads are built between
existing ones, which is why the resulting crack pattern shows similar-
ities with town road networks. This is seen most clearly in older cities
where the streets were laid down spontaneously, without a planner’s
overarching vision: in Paris, say, but not in New York (Fig. 3.16). Here
the initial long cracks are the oldest roads, leading from the city centre
to surrounding villages or abbeys. Paths and lanes sprang off from
these main arteries, perhaps initially to give carts access to fields.
Fig. 3.15: A crack pattern in the glaze of a ceramic plate forms a web in which the fracture lines
tend to intersect at right angles and the islands have an average of four sides (a). The cracking
happens in a hierarchical manner: first, the longest cracks form, and gradually the spaces between
them become laced with bridging cracks (b). (Photos: Steffen Bohn, The Rockefeller University,
New York.)
These informal routes became streets as the urban centre expanded.

Right-angled T-junctions offer the most direct routes between one
artery and another, and so the road network evolves in a hierarchy of
subdivisions, just like a cracking film. If Bohn’s theory is right, then, a
street map too encodes its own history.
Fig. 3.16: Hierarchical crack patterns with four-sided domains

resemble the street networks seen in old cities that have grown
spontaneously, as shown here for a map of Paris from 1760. (From an
image kindly supplied by Steffen Bohn.)
The devil’s honeycomb
In the polygonal islands of these crack patterns, traced out by cracks

that intersect with near-geometric precision, we might begin to discern
the outlines of the prismatic columns of the Giant’s Causeway. But the
answer is not as simple as that. For one thing, those rock pillars did not
form in a thin layer, but are the result of a hexagonal web of fractures
that passes downwards almost unchanged through many metres of
basalt. And we cannot ascribe these hexagons to any underlying
hexagonal packing of constituent particles, as in the case of the drying
layers of identical microspheres. Any explanation will have to work
a little harder than that.
Whatever the cause of the causeway, it must be a rather unusual one.
Basaltic lava has flowed from many places on the Earth, but few of
these igneous deposits have cracked in such regular array as they cool
(another example is the Devil’s Postpile in California). Yet the process

has been seen to occur other materials too. In 1922 a British optical
engineer named J. W. French studied the cracking of starch—a slurry of
tiny particles, not dissimilar to mud—as it dries. His experiment was
part of a programme designed to mimic the cracking of glass, but
French found that drying starch breaks up into columns: a kind of
mini-Giant’s Causeway, as he recognized. The work was all but forgot-
ten until Gerhard Müller of the J. W. Goethe University in Frankfurt
repeated the experiment in 1998. He found that layers of starch several
centimetres thick fractured into columns a few millimetres wide.*
Stephen Morris and his colleague Lucas Goehring at the University of
Toronto in Canada repeated this experiment in 2006, with the same
result (Fig. 3.17).
The cracking of the Giant’s Causeway began at the top, where the
heat escaped and the molten rock was therefore coolest. The network of
cracks formed there will have gradually advanced downwards into
the solidifying rock bed. But this does not seem to have happened
smoothly: instead, the fracture pattern progressed downwards in
layered steps, each successive layer freezing and cracking before the
next. We can see that this was so because it has left horizontal striations
in the surfaces of the basalt pillars.
This gave physicists Eduardo Jagla of the Centro Atómico Bariloche in
Argentina and Alberto Rojo of the University of Michigan a vital clue to
the way the prismatic columns were created. In 2002 they proposed
that the crack network in the initial layers was much more irregular, but
that it became progressively more orderly and polygonal as it propa-
gated deeper. For a fixed total length of cracks, a polygonal, roughly
hexagonal network is more effective than a random jumble at releasing
the stress that builds up in the cooling and contracting layer of rock.
This is somewhat comparable to the way that a hexagonal honeycomb
network of soap films in a foam provides the least total surface area and
therefore the lowest-energy configuration, as I explained in Book I. The
uppermost layers of the rock cannot ‘find’ this optimal pattern, how-
ever, because the cracks just start anywhere at random and advance
until they meet another. But as each successive layer freezes and rup-
tures, this crack network gets reorganized so that it takes on an
*I know of no other paper with an acknowledgement like this one: ‘I thank my wife for her
patience with my kitchen interferences.’
Fig. 3.17: An artificial Giants’s Causeway made from columns of corn starch left to desiccate in a
thick layer (a). The ruler at the top, marked in millimetres, shows the scale. A columnar structure like
this in basaltic rock can be found in other parts of the world besides Staffa and the Giant’s
Causeway, such as in this formation near Banks Lake in Washington State (b). (Photos: Stephen
Morris, University of Toronto. b, from Goehring et al., 2006.)
increasingly polygonal shape. Once such a network is attained—and

it does not have to be perfectly hexagonal, but only roughly so—it
remains fixed, since further reorganization would not significantly
improve the way stress is relieved. At that point, then, the network
stays identical from one layer to the next, creating vertical-sided
columns.
A process like this in which irregular cracking of the uppermost layers
becomes marshalled into a stable polygonal network as the cracks
descend was in fact suggested in 1983 by Irish physicists Denis Weaire
and Conor O’Carroll in Dublin, who in turn got the idea from the
American metallurgist Cyril Stanley Smith. (Weaire says that the notion
of descending ‘crack layers’ goes back further still.) But they had no
evidence that it could work. Jagla and Rojo showed that it might: they
tested their idea using computer simulations of a layer-wise cracking
process. The first layer had a jumbled web of branching cracks, but by
the eighth layer this had become tidied up into polygons that then
changed hardly at all in lower layers (Fig. 3.18). Most of these polygons
are six-sided (the number ranges from four to eight), and they all have
roughly the same size. This looks very much like the pattern seen in the
Giant’s Causeway (Fig. 3.2c). Better still, the researchers showed that
both the relative proportions of polygonal cross-sections with different
numbers of sides, and the relationship between the number of sides
and the area of the polygons, closely match those observed in the
geological formation.
If this picture is correct, where are the original, more disorderly upper
layers of the Giant’s Causeway? As with many rock formations, they
have most probably been stripped away long ago by wind, rain, and sea.
But Morris and Goehring verified that prismatic pillars may appear in
this manner with their experiments on desiccating starch, where they
found that an early jumble of descending cracks does indeed find its
own pseudo-geometric order (Fig. 3.17a). The final stable pattern de-
pends on several factors, in particular the cooling or drying rate of
the material. Morris and Goehring found that each set of experimental
conditions seems to offer a range of different average pillar widths, and
it is not obvious which of these will be selected in a particular case: this
may depend, for example, on the previous history of the process. The
sizes of all the columns can change suddenly as some descending
cracks run out of steam and the others rearrange themselves. In other
words, specific polygonal crack patterns may be selected from a range of
Fig. 3.18: Computer simulations of the successive ordering of cracks

in a substance that solidifies layer by layer show that they can
evolve from a rather random array to one that has largely hexagonal
domains. (After Jagla and Rojo, 2002.)
possibilities: all may be more or less hexagonal, but the scale varies. You
could say that the propensity for patterning is inherent in the process of
solidification, but the precise pattern that results (and in particular, the
size of its elements) depends partly on the whims of the elements. So
once again, chance and necessity supply the choreography.
Water Ways
Labyrinths in the Landscape
4
R
I V E R S , like all things linked to water, are popular metaphors.
They speak of time and life and journeys, of blood and tran-
quillity and turmoil. The metaphors shift, however, and are liable
to trip us up. If life is a river, what is the Styx? Rivers nurtured the
earliest civilizations, but periodically decimated them.
When the biologist Richard Dawkins compared evolution to a river in
his book River Out of Eden, he had in mind both the notion of time’s
flow and the luxuriant branches of the phylogenetic tree that connects
all species. This is a vivid image, but it’s best not to think too hard about
it—for the river’s source lies in its tips, not in the channel to which all
tributaries converge.*
That is the odd thing about rivers: their networks grow in the oppos-
ite direction to the way the water flows, their headwaters cutting back-
wards into rock. There is a very real sense in which we can regard a river
as a crack, propagating slowly across a range of hills or mountains.** Yet
there is no ‘pressure’ pushing the tips forward. All the same, the result
(Fig. 4.1) is a pattern that looks rather like the branched formations
we have seen already: not only cracks but sooty aggregates, bacterial
colonies, and electrical discharges.
In fact, rivers are arguably the grandfather of branching patterns: the
first that were ever contemplated in terms of their formal shape.
Leonardo da Vinci, who we encountered in Book II as a pioneer of
fluid flow patterns, sketched the most extraordinary topographical
*Moreover, rivers flow in a particular direction, whereas some biologists, such as Stephen
Jay Gould, argue vigorously that there is no direction in evolution: it branches, but is ‘going’
nowhere.
**In the real world it’s actually a little more complex than this, since rivers do not just grow
from the tips. For example, sometimes tributaries of one channel can become captured by
another, causing them to reverse the direction of their flow.
Fig. 4.1: River networks: geomorphological cracks on a grand scale?

Here is the drainage basin of the Dry Tug Fork in California. (Image:
Andrea Rinaldo.)
maps of rivers and their watersheds, in which the mountains are repre-
sented not, in the medieval manner, as so many stylized conical pro-
files, but with shading that shows something like a contour or tree line
(Fig. 4.2). It is as if he really took to the skies in one of his flying
machines to obtain this bird’s eye view, from where the ferny fractal
fronds of the river basin look eerily like those evident in today’s satellite
imagery (Plate 6). Leonardo had good reason to strive for this accuracy,
since these were most probably plans for his hydraulic engineering
schemes, such as the construction of canals on the River Arno in
Tuscany. Yet for all his practicality, Leonardo’s vision was also informed
by his metaphysical convictions: when he called rivers the ‘blood of the
earth’, this allusion to the venous structure of the network was not just
pretty word-play but was rooted in Neoplatonism. He firmly believed
that structures found on a grand scale in nature—the macrocosm—
would be reiterated in the body of man, the microcosm. We will see
WATER WAYS j 103
Fig. 4.2: Leonardo da Vinci’s ‘aerial’ sketch of a river network and the surrounding topography
looks strikingly like the fractal forms seen in today’s satellite imagery (Plate 6 ).
shortly that he was not mistaken in this regard. Not only do these forms
look alike, but they may well have the same ultimate cause.
Scaling up streams
For geomorphologists (those scientists who study the shapes of land-

forms) the branching patterns of rivers were one of the most prominent
features of natural landscapes, and demanded explanation. The first
attempt to formulate one was made by the American hydrologist and
engineer Robert E. Horton in the 1930s, who proposed that there are
universal ‘laws of drainage network composition’. These laws are now
generally cited in the modified form defined in 1952 by geomorpholo-
gist Arthur Strahler. Each branch of the river network is assigning an
‘order’ that signifies its position in the hierarchy. The streams at the tips,
which themselves have no tributaries, are first order. Where two first-
order streams join, the resulting stream is second order; and in general,
the meeting of two streams of a given order signals the beginning of a
stream of next highest order (Fig. 4.3). If a lower-order stream flows into
a higher-order stream, the former terminates but the latter’s order is
unchanged.
Horton claimed that these stream orders obeyed certain mathemat-
ical regularities. His ‘law of stream numbers’ states how the number of
Fig. 4.3: The hierarchy of river network elements in Strahler’s

modification of Horton’s classification scheme. The order assigned to
each branch increases as one progresses from initial tributaries to the
main river channel.
WATER WAYS j 105
streams of a particular order depends in a predictable way on the order.

It is clear from looking at a drainage network that there are fewer
higher-order streams than lower-order. Horton expressed this with
mathematical precision: the number of streams of order n is propor-
tional to the inverse of a constant C raised to the power n. The number
of second-order streams is proportional to C 2, third-order to C 3, and so
on. This law is an example of a power law or scaling law (see page 35).
Another way of expressing this idea is to say that the number of streams
in each order is a constant times the number in the next highest order.
The number of first-order streams in a particular network might, for
example, be four times the number of second-order streams, which is
itself four times the number of third-order, and so on. This would
mean that each stream has, on average, four branching streams of the
preceding order.
Horton also proposed a scaling law for stream lengths: the average
length of a stream of order n is proportional to a (different) constant D
raised to the power n. (Or again: the average length for each order is a
constant times the average length of the preceding order.) Thus,
streams of higher order are longer—again what you would anticipate
intuitively from looking at the drainage map. A third scaling law relates
the downstream slope of a stream to its order. In 1956 the American
geomorphologist Stanley Schumm proposed a fourth law, in the same
spirit: the area of the drainage basin feeding a stream with water
increases with stream order in the same way as stream length: that is,
proportional to a constant raised to the power n. And a year later a
geologist named John Hack added one more scaling relationship,
pointing out that the area of the full drainage basin for a network
increases in proportion to the length of the principal river (that is, the
highest-order element of the network) raised to the power of about 0.6.
Hack’s law seems to be more or less true for drainage networks ranging
in size from those produced in small laboratory experiments to those
almost as big as the Amazon. But there is some dispute about the
precise value of Hack’s exponent: some estimates place it closer to
0.5, while others think it does not really have a universal value at all,
but varies slightly from place to place. It appears, in fact, that the spread
in observed values of Hack’s law is related to the fact that river basins
seem to become elongated (they get narrower) as they get bigger.
These scaling laws are really expressions of the fractal, self-similar
character of drainage networks. They are reflections of the fact that the
structures look the same over a wide range of magnification scales, so

that you might not know from an aerial photo if you are looking at an
area a mile across or a hundred miles. Where does this fractality come
from? And can we explain why the networks follow these particular
scaling laws, and not others?
When Horton first reported his laws they were regarded with awe, as
though uncovering a profound natural regularity. But in 1962 Luna
Leopold and Walter Langbein showed that randomness alone is enough
to ensure that these relationships hold for any branching network. They
proposed a model of stream formation based on that developed by
Horton himself, according to which networks emerge as rain falls
onto a gently undulating surface. Wherever rain delivers more water
than can be removed by filtering down through the rock bed, water
accumulates on the surface and flows down the steepest gradient.
These are not really streams, but so-called ‘rills’—they don’t ‘go’ any-
where in particular, but just cover the land surface with little gullies that
grow bigger and deeper as the water erodes the surface. As rills grow
larger, they begin to merge. In the model developed by Leopold and
Langbein, rills form at random and larger channels arise from the
merging of smaller ones. Because the surface topography is random,
the rills evolve in random wiggles, constrained only by the fact that, like
streams, they cannot re-cross their own tracks (a property called self-
avoidance). This model generates networks that obey Horton’s laws as if
by magic, even though its ingredients reflect only the barest details of
the real geological processes.
In 1966 the geomorphologist Ronald Shreve showed that in fact
Horton’s laws are extremely likely to result from any process that con-
nects at random a given number of stream sources in a drainage basin
into a network. And the geomorphologist James Kirchner demonstrated
in 1993 that even randomness is not essential: almost every kind of
branched network conceivable obeys Horton’s laws, not just those
arising from random processes. In other words, Horton’s laws don’t
really tell us anything at all about the fundamental patterns of stream
networks. They are probably instead an inevitable consequence of
the scheme that Horton (and subsequently Strahler) used to break
down the networks into fundamental units of different order. So it
is no good testing a particular model of drainage network development
by seeing if it generates Horton’s laws, because just about any model
will do that.
WATER WAYS j 107
This is a salutary tale. You see a pattern and you discover that it
conforms to a particular mathematical description, and so you think
that the maths captures the essence of the pattern. But it may be that
lots of other patterns follow that same mathematical law, while differ-
ing in other ways. That is a particular hazard of research on fractals: a
fractal dimension is a useful measure, for example, but it is not a
unique fingerprint of the pattern.
Invasion of the highlands
In any case, it is now clear that drainage networks do not usually form
by the random appearance of rills followed by their merging. Instead, a
network grows from the heads (tips) of the channels, where erosional
processes cut back into the rock. To understand why networks have the
form they do, we must therefore focus on what is happening here at the
stream heads.
Cutting into rock requires energy. In stream networks this comes
from the kinetic energy of rainwater or meltwater flowing downhill.
The energy input is greatest where the water flows fastest and most
abundantly: where steep slopes converge in the funnel-shaped head of
the channels. Downstream, the flow is more sluggish and erosion is less
urgent. So the network grows from the branch tips—just as it does for
cracks, lightning, or viscous fingering, where the branch tips are the
places of steepest gradient in stress, electric field or pressure. And once
again there is an instability that amplifies growth: when a new channel
forms, it becomes a focus for the flow of surface water, producing
further erosion.
But this does not mean that rivers can lengthen, and new branches
form, only at the outermost tips, just as it is not so for electrical
discharges or cracks. As in those cases, chance plays a role. All land-
scapes have random variations: in surface contours, soil type and
permeability, rock type, vegetation cover, and so forth. This is the
equivalent of the randomness of particle trajectories in diffusion-
limited aggregation, say, and it ensures that there is always a chance
that new tributaries may sprout downriver, on the higher-order streams
of Horton’s classification rather than only at the first-order stream
heads. The chance is relatively smaller, because much of the water
that falls on the ground already bounded by channels will be captured
by them instead. But the possibility exists. So river networks grow like
the other randomly branched patterns we have noted already: prey to
randomness everywhere, but biased by preferential growth at the tips.
As Leopold and Langbein observed, drainage networks tend to be
self-avoiding: stream heads hardly ever cut back across other streams to
create islands or loops. This is because, as a stream head advances
towards an existing channel, the area feeding it with water diminishes,
since the other channel claims an increasing proportion of the water
supply. Stream heads therefore generally run out of steam (or more
properly, of water!) before they intersect other streams. Analogously,
the tips of a DLA cluster very rarely intersect other branches because
new particles cannot reach them once they get too close to another
branch.
The connection between drainage network formation and crack for-
mation is made explicit in a network-forming model called invasion
percolation. Percolation is the passage of a fluid through a porous
substance. David Wilkinson and J. F. Willemsen, working at the oil-
mining company Schlumberger-Doll in Connecticut, devised the inva-
sion percolation model in 1983 to describe how one fluid pushes
another through a network of pores: a key process in oil recovery by
injection of water into the oilfield. We have seen how this process can
create branching instabilities that lead to viscous fingering patterns.
But in invasion percolation the fluids are contained in a web of pores
that imposes its own pattern, so that the invading fluid advances in
a densely interweaving network (Fig. 4.4). In effect, the invading fluid
pushes into a surrounding matrix in a way that is governed by the pores.
The probability that the invading fluid displaces the other depends
on the size of the pore through which it passes, since this modifies the
fluid pressure. If the pore network has a random distribution of pore
sizes, then this probability varies more or less randomly through the
system: there is an equal chance of a branch tip advancing at any point.
The invasion percolation model captures this process by considering
the invading fluid to be advancing as a ‘cluster’ through a lattice of
obstacles linked together by bonds whose strength varies randomly
from place to place. The fluid network grows by breaking these bonds
and pushing between the obstacles, like flood waters breaking down
barriers. The next bond to break is always assumed to be whichever is
the weakest one around the perimeter of the cluster. You can now see
that this model is very similar to the dielectric breakdown model
WATER WAYS j 109
Fig. 4.4: Invasion percolation: the displacement of one fluid by another within a porous network.
The ‘invading’ fluid is injected here at the point marked with a circle, and moves forward in a dense,
convoluted system of loops and branches. (Image: Roland Lenormand, Institut Français du Petrole,
Rueil-Malmaison.)
described in the previous chapter, except that there is no ambiguity

about which is the next bond to break: it is always the weakest, whereas
in the dielectric breakdown model weaker bonds simply have a pro-
portionately greater probability of breaking.
The advance of an invasion percolation cluster occurs mostly at the
tips, because the rules of bond breakage mean that the cluster naturally
‘seeks out’ the weakest bonds in its path and leaves behind along its
perimeter those bonds that are stronger. The chance of finding a bond
at the tips that is weaker than those still unbroken deeper inside the
cluster is usually pretty good; only rarely will all the tips happen to
alight on strong bonds, forcing the breakage of one further back down
the cluster’s branches. The cluster grows into fractal form.
Fig. 4.5: The self-avoiding invasion percolation model of river

network formation produces networks resembling those carved out as
rivers cut back into bedrock. (After Stark, 1991.)
The British geomorphologist Colin Stark has proposed that the evo-
lution of drainage networks is rather like invasion percolation. The
breaking of bonds mimics the erosion of bedrock by a steady supply
of surface water from rainfall; and the randomness in bond strengths
reflects the non-uniformity of the landscape. To apply the model to this
situation, he imposed one extra constraint—self-avoidance—so that a
stream head may not intersect an existing channel. Stark showed that
this model produced stream networks that looked rather realistic
(Fig. 4.5). And he found that his model networks obey Hack’s scaling
law with an exponent of about 0.56, matching the value of 0.5–0.6 seen
in nature.
However, like all simple models that have been proposed for explain-
ing the form of river networks, the invasion percolation model only gets
part of the pattern right. For one thing, snapping bonds in a lattice is
not really much like erosion and sediment transport in real rivers. And
WATER WAYS j 111
sometimes three or more tributaries converge in Stark’s networks,

whereas this rarely happens in real river networks. Furthermore, it is a
little unsatisfying that self-avoidance must be imposed rather than
arising naturally in the model. In other words, there is more going on
than mere invasion percolation in the formation of a river network.
Nature tells us this must be so, because, like trees, river patterns do not
all look the same.
Several alternative models offer a similar combination of landscape
randomness and growth instabilities that promote branching, and
almost all of them produce fractal patterns, along with fair agreement
with some of the scaling laws seen in the natural networks. On this
evidence, then, it seems there is no unique way to describe the forma-
tion of river networks. We know the basic principles, but it is not clear
which details are essential and which are incidental. And how well a
model mimics the real world may depend on which rivers we choose
for the comparison. This is all rather unsatisfactory.
The best of all worlds
What is needed is a broader perspective. Instead of a river winding

through the landscape, think for a moment about an apparently
simpler problem: the path of an object falling under the influence of
gravity. It could be a cannonball fired from a cannon, or a pen rolling
off a tabletop, or a raindrop released from a cloud. How can we
calculate the trajectory? Newton’s laws of motion provide the rules,
telling us which forces act on the body. But these laws rarely supply
a very easy prescription for calculating the trajectories that result.
A better way of determining the paths of moving bodies was devised
in the late eighteenth century by the French mathematician Joseph
Louis Lagrange, and modified half a century later by the Irish math-
ematician William Hamilton. Lagrangian and Hamiltonian mechanics
are equivalent to Newtonian mechanics, but instead of formulating
the problem in terms of forces, they consider the energies involved—
for example, the changes in the gravitational potential energy as
an object falls, and its kinetic energy due to motion. This makes the
job easier, and it provides a general criterion for the trajectory of a
falling object, or indeed of any moving object: the path taken is that
which minimizes a quantity called the action, which depends on the
energy changes involved in the motion and the time taken for it to
happen.
The Venezuelan environmental engineer Ignacio Rodriguez-Iturbe,
the Italian physicist Andrea Rinaldo, and their colleagues think that
there is an analogous ‘minimization’ principle guiding a natural river
drainage network into a branched, fractal structure. The network
evolves, they say, in such a way as to minimize the rate at which the
mechanical potential energy of the water flowing through the network
is expended. This claim needs a little unpacking.
As water flows downhill through a river network, it loses potential
energy just as a falling cricket ball does. This is converted mostly into
kinetic energy: the water moves. And the kinetic energy ultimately
drives the process of erosion that leads the network to expand and
rearrange its course. So the potential energy is ultimately dissipated:
some of it goes into kinetic energy of the river water discharging into
the sea (or at least, out of the drainage basin), while some is lost as
frictional heat by the wearing away of rock and soil.
Suppose we had a divine ability to measure everywhere at once the
amount of potential energy that all the water was losing each second.
(We cannot hope to do this in real river systems, but it is something that
can be easily totted up in computer models.) Rodriguez-Iturbe’s mini-
mization principle says that the network will evolve until it acquires the
shape for which the total rate of potential energy dissipation is as small
as possible.
It was not a totally new notion. The analysis of river drainage patterns
conducted by Luna Leopold and his co-workers in the 1960s led him to
conclude that these networks represent an optimal compromise be-
tween two opposing tendencies: for the expenditure of power by water
flow to be as small as possible (which is more or less equivalent to
Rodriguez-Iturbe’s minimal energy dissipation), and for the power of
the flow to be distributed more or less equally throughout the system.
Leopold suggested that the river network evolves into a state where
these things are balanced.*
To investigate what sort of network their model generates,
Rodriguez-Iturbe and his colleagues ‘evolved’ a network created purely
*Leopold insisted only that this optimal compromise be reached at the local scale, how-
ever—that is, for each segment of a network. He did not consider that the balance must also
be achieved throughout the whole river basin. This is rather like saying that the network is
like a collection of little ‘houses of cards’, rather than one delicately poised big one.
WATER WAYS j 113
at random into one that obeyed their minimization principle; that is to

say, they started with a network drawn on a checkerboard grid, in which
water falling uniformly on each grid cell was drained along channels
connected randomly from cell to cell (with the proviso that channels
could not cross) into a network that ultimately flowed out of one corner.
Then they rearranged the channels one cell at a time, calculating at
each step how this affected the total rate of energy dissipation. Each
change was made at random, but was ‘accepted’ only if it produced
a decrease in this total rate—if not, the change was discarded and
another one was tried. This process can generate a wide range of
networks, and each ‘run’ will create a different one. But all of them
turn out to look realistic: they have scaling properties that obey
Horton’s laws, Hack’s law, and several other empirical laws of river
patterns too. Rodriguez-Iturbe and colleagues calls this set of possible
solutions ‘optimal channel networks’.
This seemed to suggest that the form of the network was indeed
governed by the rule that energy dissipation from flow and erosion
would be kept as small as possible: in this sense, the water flow will
search out the ‘best’ solution. But can that really be so? It would be
extraordinary if it was. If you think of a typical river running over a hilly
landscape, the number of possible routes it might take is astronomical.
The chance that the single ‘best’ route will be found is therefore utterly
tiny, and indeed this is not really what happens. Instead, nature finds
river network structures that are simply ‘good enough’: they might not
be the best possible, but they are each the best in the neighbourhood.
In an entirely analogous way, water flowing down a mountain range
sometimes gathers in mountain lakes: the water could lower its energy
still further if it found a way out of the lake and right down to the valley
floor, but it cannot easily do that, and so gets stuck in the ‘local
minimum’ of the lake. There are many alternative networks with
broadly similar shapes that correspond to these local minima of the
energy dissipation rate, and they are all, in a parochial rather than a
global sense, optimal channel networks.
Why should river networks ‘seek’ to minimize their energy expend-
iture in the first place? Initially, Rodriguez-Iturbe and colleagues merely
assumed that they did, and showed that this assumption gave realistic
branching patterns. They did not initially attempt to justify the as-
sumption. Kevin Sinclair and Robin Ball of Cambridge University tried
to explain how the energy-minimization principle arises from the
fundamental flow and erosion processes that govern network evolu-

tion. They deduced that the mathematical equations relating the water
flow rate and the amount of erosion look like those that appear in
Hamilton’s law of least action. In other words, within the flow of a
single stream of water over open ground lies the prescription for the
shape of the Amazon. But of course one could never guess at that
pattern by staking out a lone channel with any number of flow meters,
depth gauges and so forth. The branching pattern is what emerges
when that process is reiterated everywhere at once. And no two pat-
terns are alike: they are all encoded in the physics of flow and erosion,
but not uniquely defined by it.
Rinaldo, Rodriguez-Iturbe and their colleagues went on to make this
connection more concrete by finding a completely different way to
generate optimal channel networks, based on the erosional processes
that happen in real river flow. They considered what would happen as
rain falls uniformly onto a high plain whose roughness is totally ran-
dom. A surface of this sort is more like sandpaper than a mountain
range: it is uneven, but has no big peaks or deep valleys. The resulting
flows of water create erosion that depends, at each point, both on the
rate of flow and the steepness of the gradient down which the water
runs. Again, the researchers divided up this conceptual landscape into
a grid of cells, and assumed that the water would run out of each cell
down the steepest gradient it can find. This flow can in principle
produce erosion of the surface, but in the model it was assumed to
remove and carry away material only when the flow exceeds some
critical threshold value. When this happens, the height of the landscape
at that point is reduced, altering the slope.
Notice that there is nothing in this model to ensure that the flow gets
channelled into a single, connected river network. Yet as the simulation
of landscape erosion proceeds, that is what happens (Fig. 4.6). And
these networks turn out to have the same scaling properties and fractal
dimension as the optimal channel networks produced by the earlier
model, which in turn are excellent mimics of nature. So a set of purely
‘local’ rules that determine how the network and landscape evolve,
describing the kinds of processes that happen in the real world, is
sufficient to ensure that the river network finds its way to the ‘optimal’
shape that minimizes total energy expenditure.
What does the erosional process do to the shape of the landscape
itself ? I’ll come back to this shortly.
WATER WAYS j 115
Fig. 4.6: In this model of river network evolution, water falls on a landscape with a randomly rough
surface and causes erosion as it flows away. The resulting streams organize themselves into a
joined-up river network in which the total rate of energy dissipation is a local minimum: it has the
smallest value of all the networks with similar configurations. These so-called optimal channel
networks have the same scaling laws as those seen in nature. (Image: Andrea Rinaldo, University of
Padova.)
It’s sedimentary
Self-avoidance is the rule for river networks: the channels do not inter-
sect. But when rivers flow across very flat, broad beds, they often break
up into a series of channels that split and rejoin into a series of loops
surrounding islands (Fig. 4.7). These are called braided rivers. Analogous
skeins of water form on a smaller scale when streams run into the sea
across a flat, sandy beach. The dried-up imprint of what appear to be
Fig. 4.7: Braided rivers have channels that loop and converge, creating isolated islands that
form and vanish as the channels change their course. This one is the Waimakariri river in
Canterbury, New Zealand. (Photo: Phillip Capper.)
braided rivers have been seen on Mars. The pattern appears whenever a
broad sheet of water runs over a gently sloping, grainy sediment.
‘The sediments are a sort of epic poem of the Earth’, says the
ecological writer Rachel Carson—to which geologist Chris Paola of the
University of Minnesota adds the frank confession that ‘Unfortunately,
this poem is written in a language we don’t understand.’ But he is one of
those working on decoding it. Paola and his colleague Brad Murray
have proposed, for example, that the transport of suspended sediment
is crucial to the formation of braided rivers. Water scours out sediment
in some parts of the flow and re-deposits it elsewhere to create sand-
bars and islands. If the rate of sediment removal by scouring increases
sharply as the flow gets faster, then this sets up a positive feedback that
makes a dip in the river bed get ever deeper: the dip captures more of
the flow than the surrounding regions, and so proportionately more
sediment is washed away from it. The reverse is true for a bump: the
flow passes around it rather than over it, and so it suffers less erosion
and gets higher than its surroundings. As a result, random small pro-
trusions become islands that divert the flow to either side.
WATER WAYS j 117
Fig. 4.8: A computer model of fluid flow and sediment transport

captures the essential features of braided rivers. The pattern is
constantly shifting, as shown here in three snapshots from a model
run. The images on the right show the topography, and those on the
left show the water discharge—essentially, the river channels
themselves. (Image: Chris Paola, University of Minnesota.)
Murray and Paola devised a model of flow and erosion that captured
these features. Water flowed across a checkerboard lattice of square
cells, whose heights decreased steadily in one direction to produce a
downhill flow. Superimposed on this smooth slope were small, random
variations in height from cell to cell. The amount of water flowing
through each cell depended on its height in relation to its uphill neigh-
bours: the lower the cell, the greater its share of water from those uphill.
The researchers assigned rules that governed how the amount of sedi-
ment either eroded or deposited at a cell depended on the flow across
it. They found that their model results (Fig. 4.8) captured many of the
features of real braided rivers. Channels continually form and reform,
migrate, split and rejoin: the shape of the river is never steady. Although
on average the flow of water and sediment down the river remains
constant, it fluctuates strongly—more so than in non-braided rivers—

because of this constant reorganization of the flow paths.
Any sandy sea shore will show you that shallow water flowing over
grains is a rich source of pattern. When the water’s edge retreats as
waves lap at the beach, the sand can become moulded into diverse
structures; the same patterns may be seen in the sediment at the
bottom of an emptying reservoir. Adrian Daerr at the Université
Pierre et Marie Curie in Paris and his co-workers studied these patterns
Fig. 4.9: Sediment erosion patterns in a thin layer of powder withdrawn at an angle from
immersion in water. The pattern depends on the withdrawal rate and angle. (Photos: Adrian
Daerr, ESPCI, Paris.)
WATER WAYS j 119
in the lab in 2002, using a ‘model shore’ consisting of a plastic plate

covered in a thin layer of sediment (alumina powder, mimicking sand
or mud) and withdrawn slowly from a tank of water at various angles of
inclination. The researchers found that as the angle was made progres-
sively steeper, there were rather sudden changes in the kind of erosion
pattern formed, from a densely ‘cross-hatched’ texture to branched
channels to a dimpled ‘orange-skin’ surface and then to chevrons that
overlap like fish scales (Fig. 4.9). They could not explain all these
patterns, but suggested that the chevrons came about as a result of
avalanches of grains down the slope of the sediment layer. Each ava-
lanche spreads into a chevron, funnelling the flow of water in a way that
sharpens the V shapes by erosion.
Such studies raise at least as many questions as they answer. But all
these erosion patterns are self-organizing, in the sense that the flow
becomes organized into a stable state with properties that remain
statistically constant even though the details are for ever changing.
This is a hallmark of self-similar growth, which allows an object to
preserve its form while it grows indefinitely.
What’s left
When we think of river patterns, what usually comes to mind is the plan
view: the convergent, branched network as seen from above, which
Leonardo intuited and topographic maps and aerial photographs
have now rendered familiar. But these images do not really reflect our
experience of rivers, for what we see instead from our nose-high view of
the world is the effect that the flow has on the landscape; in other
words, we see the rugged profile that the river carves into the land-
scape: hills and valleys, gorges, ravines, and lone peaks (Plate 7). There
is a characteristic shape and form in what the river leaves behind, just as
there is in the course it takes.
While the river network is traced out as a pattern of wiggly lines, the
topographic profile of the watershed is a surface. And just as the fractal
nature of the channel network makes it more than a one-dimensional
line, so too the rough erosion surface is a fractal that partially fills up
three-dimensional space and so has a fractal dimension greater than 2.
It is not hard to tell when a landscape is fractal, because you will soon
find that it takes a lot more time and effort to travel between two points
Fig. 4.10: A fractal boundary, such as a cross-section through a

fractal surface, has a length that depends on the yardstick used to
measure it. As the measuring stick becomes smaller, the apparent
length seems to increase as we capture more of the details. Here
the measured length increases slightly each time the measuring
stick is halved in length.
separated by a certain distance as the crow flies than it would on a flat

plain. Journeys in fractal land are arduous.
What exactly does it mean, though, for a surface to be fractal? Simply
put, it means that the bumps have no characteristic size scale: they
come in all sizes. Put another way, it means that the apparent area of
the surface depends on the size of the ruler that one uses to measure it.
Consider getting from A to B over such a bumpy surface (Fig. 4.10). How
far do you travel? That depends on how you measure the route: smaller
and smaller ‘yardsticks’ capture more and more of the ups and downs
and so the overall measured length gets longer. Of course, the ‘real’
length doesn’t get any longer—you just ‘see’ more of it. But the spooky
fact is that, on a genuinely fractal landscape, there is no ‘real’ length at
all. Ups and downs exist on all length scales down to the infinitely
small, so the apparent length goes right on increasing as we measure
with ever smaller rulers. As we have seen, true fractals like this are just
mathematical abstractions, since the crenellations of any real bound-
ary cannot get any smaller than the sizes of atoms. But some physical
objects can remain fractal over a wide range of scales of magnification.
WATER WAYS j 121
It was this apparent dependence of a rough object’s perimeter length

on the size of the yardstick that led Benoit Mandelbrot to uncover
fractal geometry. In 1961 he came across the attempts of the English
physicist Lewis Fry Richardson to specify the length of wiggly coastlines
and borders, including the west coast of Britain and the border of Spain
and Portugal. Richardson was interested in the causes of war and
international conflict, and he was drawn to the issue of border length
by thinking about how territorial disputes arise. He found that the
apparent lengths of such boundaries measured from maps depended
on the scale of the map that one used: small-scale maps show more
detail than large-scale ones, and so capture more of the nooks and
crannies, making the total length seem longer. Mandelbrot realized
that ‘length’ is in this case not something that can be meaningfully
specified. Instead, the ‘shape’ of the boundary is better described by
calculating how quickly the apparent length increases as the yardstick
gets shorter. This is what the fractal dimension measures: it is an
unchanging geometrical property of the way the convoluted object
fills up space.
There is an important subtlety here. I explained earlier that a fractal
object such as a crack or a DLA cluster is self-similar: if you look more
closely at any part of it, you seem to see the same overall shape
repeated again. More precisely, self-similar objects are composed of
(approximate) copies of themselves scaled down by a constant ratio.
They have the same fractal dimension in all directions.
Fractal surfaces are not quite like this. Although they have a fractal
dimension of between 2 and 3, indicating that they partially fill up
three-dimensional space, this space-filling is not the same in all direc-
tions. We can see this rather easily by taking cross-sectional slices
through a rugged mountainous landscape. A vertical cut reveals a
rising and plunging, but continuous, transect across valleys and peaks
(Fig. 4.11a). But a horizontal cut reveals something else entirely: isol-
ated ‘islands’ separated by empty space (Fig. 4.11b). Such a fractal
structure is said to be self-affine, which crudely means that the ratio
by which the elements get scaled down at successive levels of magni-
fication is different in different directions.
Mandelbrot realized in the 1970s that the natural topography of the
Earth is typically a self-affine fractal. This was implicit, he said, in a
description of mountain landscapes by the Victorian climber and ex-
plorer Edward Whymper in his book Scrambles Amongst the Alps: ‘It is
Fig. 4.11: A vertical slice through a rugged valley reveals an irregular

profile of peaks and valleys (a). A horizontal cut, meanwhile, isolates
islands—contour lines corresponding to cross-sections of the peaks,
separated by gaps (b).
worthy of remark that . . . fragments of . . . rock . . . often present the char-

acteristic forms of the cliffs from which they have been broken.’ Self-
affine landscapes can be generated rather easily on computers, and
they supply convincing imitations of real mountainscapes (Fig. 4.12)
which have been used in Hollywood movies such as Star Trek II: The
Wrath of Khan. The crucial point here is that these landscapes are not
simply random. If you set the computer to generate an image in which
the size and shape of the hills and valleys are assigned by a random-
number generator, the result is a topography that is certainly uneven
but looks wrong. Fractal landscapes are ‘noisy’ and unpredictable, but
there is more to them than chance alone.
WATER WAYS j 123
Fig. 4.12: Self-affine fractal landscapes generated by computer look very like real mountainous
terrain. (Image: John Beale.)
In view of how important fractal coastlines were to the evolution of

the general notion of fractal geometry, it’s surprising that a good model
for how coasts acquire this form by erosion was not developed until
2004. Partly this is due to the fact that coastal erosion is a complex,
many-facetted process. The pounding of waves can wear away rock,
but it also redistributes sand and stones. Cliff faces are shattered by
frost expansion or worn down by rain. And water induces chemical
attrition called weathering, for example by salt corrosion or via reac-
tions between minerals and water. Rocks weakened by slow chemical
weathering may be broken apart and removed quite suddenly by the
violence of a storm.
Bernard Sapoval at the Ecole Polytechnique in Palaiseau, France, and
his co-workers simplified these details by assuming just two kinds of
erosion: rapid, due to the mechanical effects of battering in storms, and
slow, due to chemical weathering. In their model, a coast was repre-
sented by a grid of cells containing several rock types with different
resistance to erosion, distributed at random. They suggested that the
Fig. 4.13: A coastline carved out by a computer model of erosion (a: the dark area here is the land)
looks similar to a real coast (b: in Sardinia; the boundary is shown more clearly in c), with
peninsulas, bays, and a ragged outline. Both boundaries are fractal, with the same fractal
dimension. (Images: Bernard Sapoval, CNRS Ecole Polytechnique, Palaiseau.)
WATER WAYS j 125
key to the evolution of fractal geometry was a feedback process by

which erosion of the coast by waves is altered as the coastline becomes
more convoluted, damping down the waves. Allowing for this effect, the
researchers found that a smooth coast soon breaks up into a rough
outline, which becomes increasingly pitted into bays, peninsulas, and
islands (Fig. 4.13). The erosion process doesn’t just expose the strongest
rock—this would result in a random, non-fractal shape—but maintains
a fine balance between removing the weaker sections and removing
those that are most exposed. The model ignores several aspects of
coastline formation that are surely important, not least the movement
of sediment, but nonetheless it seems to capture the most obvious
features of real coasts. They ‘look’ right, for sure.
Carried away
The self-affine relief of a river basin must be related to the branched

structure of the river network. But how? Ignacio Rodriguez-Iturbe,
Fig. 4.14: In the optimal channel network model described earlier, a river network such as that in
Fig. 4.6 has an associated topography. The initial landscape is randomly bumpy, but its profile
changes to a fractal form, with hills and valleys of all sizes and heights. (Image: Andrea Rinaldo,
University of Padova.)
Andrea Rinaldo, and their co-workers think that their model of optimal
channel networks provides a link. I showed earlier that in this model,
water carves realistic-looking river networks through a randomly
bumpy surface. In doing so, the flow reshapes the topology of the
landscape, sculpting it into peaks and valleys. The initially sandpaper-
like surface deepens into a rugged range of hills and valleys with a
fractal character (Fig. 4.14), very like that of a real landscape. Once
this landscape has attained its fractal state, erosion continues and the
shape continues to change; but the basic form, as characterized by the
fractal dimension, remains constant.
Tamás Vicsek and his co-workers in Budapest have studied this same
process experimentally, using real mud and water. They mixed sand
and soil to simulate the granular, sticky stuff of hillslopes, from which
WATER WAYS j 127
Fig. 4.15: An experiment on erosion of a bed of sand and clay by water produces a rugged skyline
(a) that resembles those seen in nature at scales thousands of times larger, such as this
mountainscape in the Dolomites (b). (Photos: Tamás Vicsek, Eötvös Loránd University, Budapest.)
they constructed a flat-topped ridge just over half a metre long. They
sprayed it evenly with water to see what kind of surface would be
carved out by erosion.
The running water carries off material in two ways: the granular
substance is worn down gradually, but from time to time landslides
remodel the surface more abruptly. Both of these processes, of course,
occur in real hill and mountain ranges. The result is a rough, bumpy
ridge that one could easily mistake for a rocky hillslope on a scale
thousands of times bigger (Fig. 4.15)—a reflection of the scale-invariant
self-affinity of these erosion surfaces. But you might protest that moun-
tains are made of rock, not a soft mixture of soil and sand. That is true,
but it may not matter. Both of these substances are worn away by
flowing water; it merely happens much faster in the softer medium.

And both have a resistance to erosion that varies rather randomly from
place to place: the sand and soil were only crudely mixed, and rock is
highly non-uniform. Finally, both materials suffer erosion due to the
same two processes: gradual removal of suspended small particles, and
abrupt landslides.
The results, then, are miniature replicas of the world’s most awesome
vistas: demonstrations of how the elemental forces of nature can ex-
hibit a blithe indifference to scale.
Inverted icicles
The snowfields of the Andes experience a very different kind of erosion

process that creates one of nature’s strangest spectacles. The high
glaciers here can become moulded into a forest of ice spires, typically
between 1 and 4 metres high, called penitentes because of their resem-
blance to a throng of white-hooded monks (Fig. 4.16a). Charles Darwin
saw these eerie formations in 1835 en route from Chile to Argentina.
‘In the valleys there were several broad fields of perpetual snow’, he
wrote in The Voyage of the Beagle,
These frozen masses, during the process of thawing, had in some parts been
converted into pinnacles or columns, which, as they were high and close to-
gether, made it difficult for the cargo mules to pass. On one of these columns of
ice, a frozen horse was sticking as on a pedestal, but with its hind legs straight up
in the air. The animal, I suppose, must have fallen with its head downward into a
hole, when the snow was continuous, and afterwards the surrounding parts
must have been removed by the thaw.
Darwin remarked that the locals believed them to be formed by wind

erosion. But the process is more complicated than that, representing a
classic case of pattern formation by self-amplifying feedback. The air at
these great heights is so dry that sunlight falling on the ice transforms it
straight into water vapour rather than melting it into liquid water.
A small dimple that forms in the smooth ice surface by evaporation
acts as a kind of lens that focuses the sun’s rays into the centre, and so it
is excavated more quickly than the surrounding ice. It’s a little like
diffusion-limited aggregation or dendritic growth in reverse: a ‘finger-
ing’ instability penetrates into the ice rather than pushing outwards
from the surface.
WATER WAYS j 129
Fig. 4.16: Ice and snow fields in the Andes can be carved into spikes called penitentes by
the eroding power of sunlight (a). The process has been mimicked in the laboratory, producing
ice spikes just a few centimetres tall (b). (Photos: a, Cristian Ordenes; b, Vance Bergeron, Ecole
Normale Supérieure, Lyons).
The process can be accelerated by a fine coating of dirt on the snow

surface. As the troughs deepen they expose clean snow that is prone to
further evaporation, whereas dirt in the old snow at the peaks covers
the ice crystals like a cap and insulates them. You might expect that, on
the contrary, snow or ice will melt faster when dirty than when clean,
because the darker material will absorb more sunlight. But whether
a layer of dirt acts primarily as an insulator or an absorber depends on
how thick it is.
The physicist Vance Bergeron of the Ecole Normale Supérieure in
Lyons and his co-workers have mimicked this natural process in the
laboratory, making ‘mini-penitentes’ by exposing blocks of snow or ice
to a bright spotlight. After a few hours of illumination, tiny peaks just a
few centimetres tall appeared in the ice (Fig. 4.16b). Structures this
small are found naturally, and are thought to be the precursors of full-
scale penitentes. It is possible that this process might be exploited on
still smaller scales, using fierce lasers in place of sunlight to erode the
surfaces of silicon wafers used in solar cells. A coating of microscopic
penitente-like peaks could make the silicon less reflective, so that it can
capture more of the sunlight that falls on it. That is the wonder of these
pattern-forming processes: they may operate over such an immense
range of scales, from mountains to molehills to the microworld, and
with such indifference to context, that we can find geology inspiring
technology, or biology apeing a snowflake.
Tree and Leaf
Branches in Biology
5
T
HE tree (dendros) metaphor is invoked so regularly in scientific
descriptions of branching patterns that it seems only reasonable
to expect these models and theories to tell us something about
the shapes of real trees. But therein lies a problem of another order.
A tree is a teleological form: it is a form with a ‘purpose’, an example of
Darwinian designer-less design.
There are many challenges that a tree must meet. How can it pump
water from the roots to the leaves? How can it support its own tremen-
dous weight? How to maximize its light-gathering efficiency? How to
grow tall enough to compete for light with its neighbours, without
becoming too massive for the roots to bear? In the face of these
dilemmas, there is little chance that a simple model based in maths
or physics will tell all about the shape of a tree. It is far from clear, for
example, that this shape is dictated by anything as simple as a branch-
ing growth instability.
If you were to be asked to describe the shape of a planet or a grain of
salt, you can do so in a single word: ‘sphere’, or ‘cube’. But such geo-
metric labels do not work for trees. ‘Branched’ is clearly not enough;
indeed there is no generic ‘tree shape’ anyway, since this varies between
species (Fig. 5.1 and Plate 8). To give a precise description you would
need to specify all of the branches and all of their angles and lengths—
to paint in words a picture of the complete tree (and then only of that
particular tree). You end up, in other words, like Sartre’s Antoine
Roquentin in La Nausée, horribly fixated on the arboreal specifics in
front of you.
The most useful kind of mathematical description of a tree doesn’t do
this. Instead, it encapsulates a general shape in a set of rules—an
algorithm—that generates a whole family of characteristic yet non-unique
Fig. 5.1: The branching patterns of trees are a fingerprint of their species. (Photos: a, Henry Brett;
b, Kyle Flood; c, Amanda Slater; d, Andrew Storms).
forms. One might devise a ‘cypress algorithm’, say, or an ‘oak algorithm’.

An algorithmic approach to generic form underlies much of the work
on mathematical fractals: the algorithm states how to ‘grow’ the form
by enacting a series of steps. It is important to remember that there’s no
guarantee that the mathematical algorithm bears any relation to the
growth process that occurs in nature; but on the other hand, it does
seem clear that some natural forms are created by the repeated imple-
mentation of certain simple steps.
TREE AND LEAF j 133
Algorithmic models are capable of mimicking the essential shapes of

trees without paying the slightest regard to the biology or mechanics
that underpins them. These models are not descriptions of the real
growth process, but rather prescriptions for enabling us to generate
something tree-like (or plant-like). Even if that does not tell us much
about the reasons why a tree looks the way it does, it might provide clues
to the primary characteristics of this form, so that we can begin to make
educated guesses about what a true model of growth should look like.
Leonardo da Vinci suspected (although without formulating it in
quite these terms) that there are algorithmic rules governing tree
growth. For example, he suggested that at branching points, the rule is
that the central trunk is bent by some specific angle when a side branch
occurs on its own, but is not bent at all if two side branches are
positioned opposite one another. An artist could regard these simply
as rules of thumb for making his paintings look realistic; but as I
explained in Book II, for Leonardo representing nature was intimately
bound up with understanding it. He felt that the crucial aspect of a
branching junction is that the total cross-sectional area of the branches
stays the same. As a hydraulic engineer, he considered it important that
these ‘pipes’—for the wood of a tree is a mesh of cellular channels, the
phloem and xylem, that carry water and sugar-rich liquid to and from
the extremities—retain their fluid-bearing capacity as they branch.
Are Leonardo’s rules for branching angles true? Yes, to a degree
(Fig. 5.2), but they seem to depend on the size of the side branch: single
small ones cause next to no bending of the trunk. The German
anatomist and embryologist Wilhelm Roux, a pupil of Ernst Haeckel at
Jena, attempted to specify these rules more precisely at the end of the
nineteenth century. He claimed that:
1. When the central stem forks into two branches with equal width,
they both make the same angle with the original stem.
2. If one branch of the fork is of lesser width than the other, then the
thinner branch diverges at a larger angle than the thicker.
3. Side branches small enough that they do not deflect the main stem
appreciably diverge at angles between 708 and 908.
These rules are illustrated in Fig. 5.3a. Roux did not assert them for
trees, however: he was studying arterial networks, and it was only in the
1920s that the biologist Cecil Murray suggested they applied to plant
stems. Murray too was more interested in blood flow, but researchers
Fig. 5.2: Tree branches often tend to follow Leonardo’s rules that a single side-branch deflects the
main trunk (a) while two, on opposite sides of the trunk, do not (b).
like him had started to suspect that the similarity in form between trees
and the blood circulatory system might be no coincidence. Both
are vascular networks: they are formed from hollow tubes that carry
fluids. Veins and arteries are like pipes of varying width, while wood is
like a bundle of narrow tubes that separates into bunches at branching
points.
TREE AND LEAF j 135
Fig. 5.3: Rules for branching. One of the first sets of rules to formalize
the way natural branching occurs was drawn up in the nineteenth
century by Wilhelm Roux, who formulated them for the cardiovascular
network (a). Hisao Honda devised some rules for creating realistic-
looking tree patterns in an algorithmic way (b). The rules illustrated on
the left apply to all branches except those that diverge from the main
trunk, which are governed by the rules on the right. Here r1 and r2 are
the ratios of the lengths of side-branches to that of the main branch or
trunk (L), and the angles a1 and a2 specify the respective divergence
angles. Successive branches off the main trunk diverge from one
another by an angle a.
Murray’s algorithmic rules generate somewhat realistic-looking

‘trees’ when they are used algorithmically to create a randomly
branched network. A more complex algorithm for making tree-like
branching structures was proposed by the Japanese biologist Hisao
Honda in 1971, and runs as follows (see Fig. 5.3b):
1. Every branch forks into two ‘daughter’ branches at a single branch-

ing point.
2. The two daughter branches are shorter than the ‘mother’ branch by
constant ratios r1 and r2.
3. The two daughter branches lie in the same plane as the mother
branch (the branch plane), and diverge from it at constant angles
a1 and a2.
4. The branch plane is always such that a line lying in this plane,
perpendicular to the mother branch, is horizontal. (This is the tricki-
est of the rules to envisage, but is explained in the figure.)
5. An exception to (4) is made for branches diverging from the main
trunk, which observe the length ratios specified in (2) but branch off
individually at a constant angle a2, with a fixed divergence angle
(here denoted a) between consecutive branches.
With a few minor changes, this set of rules can be used to define
algorithms that produce a whole range of branching patterns closely
mimicking those of real trees (Fig. 5.4). Further modifications to
Fig. 5.4: Trees generated from Honda’s rules in Fig. 5.3b. (From Prusinkiewicz and Lindenmayer,
1990.)
TREE AND LEAF j 137
account for the influences that real trees experience (wind, gravity, the
need to arrange leaves for optimal light harvesting) give even more
realism. Honda’s algorithm is deterministic: it prescribes the branching
pattern fully once the ratios and angles are fixed. Other algorithms that
have been used to generate life-like trees in computer art employ
random elements to create more irregular forms. In nature, random-
ness enters into the branching patterns as a consequence of such things
as breakages, collisions between branches, growth stunting due to the
shade of an overlying canopy, and the mechanical influences of wind
and rain. Another class of deterministic algorithms, called L-systems by
Przemyslaw Prusinkiewicz of the University of Regina in Canada, can
be used to produce plant- and fern-like structures (Fig. 5.5). Ultimately,
one might hope that appropriate rules for these tree-growing algo-
rithms can be derived from models of how plants grow, such as those
mentioned in Book I.
Scaling up
When Cecil Murray posited his rules of vascular network shape in the
1920s, he went a crucial step further. Rather than merely describe these
networks, he also wanted to explain why they have the form they do. He
proposed that they embody a parsimonious minimization principle,
entirely analogous to that which we encountered in the previous chap-
ter for river networks. The vascular network, said Murray, has the form
that requires the smallest amount of work to pump the fluid around it.
Fig. 5.5: Plants and ferns generated by ‘deterministic’ branching algorithms, where the pattern is
completely specified by the rules (that is, it contains no random elements). The same motifs recur
again and again at different scales in these structures, but regularity is evident to greater or lesser
degrees. (From Prusinkiewicz and Lindenmayer, 1990.)
He argued that the energy required to drive blood down branching

arteries is smallest if narrow branches diverge at large angles and
wide ones diverge at shallow angles. It seems clear that Leonardo da
Vinci had something of this nature in mind when he formulated his
own tree-branching rules—he too was thinking about what kind of
branching geometries were best suited to efficient fluid flow through
the system.
The analogies in form and function between rivers and biological
branching networks as fluid-distributing systems led the geologist Luna
Leopold in the 1970s to consider whether his conclusions about the
former might be generalized to the latter. We observed in the previous
chapter that Leopold suspected river networks were shaped by a bal-
ance between distributing the power of water flow evenly and ensuring
that as little of this power is expended as possible. Might plants and
trees be governed by similar criteria, he wondered—and if so, what are
they? The equivalent of an even distribution of power, he suggested,
was a uniform spreading-out of a plant’s canopy so that it might best
harness the available sunlight. And instead of minimizing the power
expenditure of the flow, he thought that plants might seek to minimize
the total length of the branches. It was a reasonable thing to suppose,
given that building new wood or stem costs energy, although
Leopold didn’t bother too much about the fact that branches of the
same length but different thickness contain very different amounts of
tissue.
Wolfgang Schreiner of the Institute for Medical Computer Research
in Vienna, Austria, and his colleagues have shown that a wide variety of
different ‘optimal’ branching patterns can be generated by growing
them according to rules that minimize different quantities. They have
considered networks to which new branches are added in succession,
under the condition that each new branch, starting at some specified
point, makes the connection that minimizes some quantity related to
the width and length of all the existing channels. This sounds a little
abstract, perhaps, and indeed it is: the researchers were not trying to
specify what gets minimized, but simply to see what the patterns look
like for different choices of this criterion, whether it be total network
length, energy dissipation, or whatever. They found that different
choices led to markedly different structures, all of them looking more
or less familiar (Fig. 5.6). Some looked like meandering rivers on open
plains, others like the streamlined drainage networks seen in
TREE AND LEAF j 139
Fig. 5.6: Tree-like networks generated in a computer model that seeks to minimize some ‘global’
property of the system, specifically the total value of some aspect of all the branch segments. The
network in a minimizes the total segment length, that in b the total segment area, in c the total
volume, and in d a ‘four-dimensional’ property, the so-called hypervolume. In each case, the result is
a branching pattern that looks familiar: a looks like a river on rather flat terrain, for example, and
c and d like rivers in increasingly mountainous terrain. (From Schreiner et al., in Brown and West
(eds), 2000.)
mountainous terrain, others like the veinous patterns on leaves or the

passageways of lungs. This does not tell us what controls the structure
of any of these particular networks, but it suggests that this combin-
ation of algorithmic growth (repeatedly applying a set of rules) and
minimization (making the ‘best’ connections) could contain the seeds
of an explanation for a wide range of network structures.
But perhaps the most striking and provocative application of ideas
about minimization to branching network shapes has come from a
collaboration between two ecologists, James Brown and Brian Enquist

from the University of New Mexico, and a physicist, Geoffrey West of
the Los Alamos National Laboratory. They have proposed that the
branched and fractal nature of fluid distribution systems in organisms
can explain a long-standing biological puzzle about how life’s processes
depend on body mass.
Small creatures have faster heartbeat tempos than big ones: babies’
hearts beat faster than those of adults (they also breath faster), and the
heartbeats of small creatures such as birds are more rapid still. This
relationship of beat rate to mass can be expressed in precise math-
ematical terms, and it turns out to be a power law, or scaling law (see
page 35). For a wide variety of organisms, the heartbeat rate turns out to
be proportional to the inverse of the body mass raised to the power 1/4.
The life span of an organism, meanwhile, is directly proportional to the
body mass raised to the power 1/4—the bigger they are, the longer they
live. This means, incidentally, that the total number of heartbeats in an
organism’s lifetime (the heart beat rate multiplied by the life span) is a
constant for all organisms, equal to about one and a half billion. Mouse,
human or elephant: all have the same allotted time, when measured in
heartbeats.
Anyone who knows how often children need feeding will also be
familiar with the idea that smaller organisms have a faster metabolism.
An organism’s metabolic rate—the rate at which it consumes energy—
is proportional to the 3/4 power of body mass, which means that small
creatures need more energy pound for pound: the shrew must consume
more than its entire body mass every day. There are several other
examples of these biological scaling laws: the cross-sectional area of
aortal arteries in mammals and of tree trunks also depend on body
mass raised to the power 3/4, for instance. They are all called allometric
scaling laws, and they are obeyed by organisms ranging from microbes
to whales.
Of course, one would expect large creatures to use up more energy
than small ones. But it is not at all obvious that the same scaling law
should be followed over such a huge range of body sizes. Still more
puzzling are the actual values of the powers in the scaling laws: they all
seem to be multiples of 1/4. If the biological parameters (heartbeat and
so forth) were related to how quickly fluids could be distributed in the
body, you would expect the relationship to depend on how far they
have to travel, which would seem to be proportional to the body’s
TREE AND LEAF j 141
dimensions. These in turn are proportional to the 1/3 power of body

mass.* You would therefore think that all these scaling laws should
come with powers that are multiples of 1/3, not 1/4. It seems almost
to imply that bodies are four-dimensional, not three-dimensional.
Enquist and colleagues suspected that the branched structure of the
body’s distribution networks might hold the key to understanding these
allometric scaling laws. They reasoned that ultimately organisms can
survive only so long as their tissues are supplied with the essential
resources that sustain them. Air-breathing creatures like us need a
steady supply of oxygen to our cells. This is delivered to the blood via
the branching channels of our lungs, and distributed around our bodies
by our cardiovascular network. Trees and plants likewise need a vascu-
lar system to provide water and nutrients. But branching itself is not
enough: the key point is that the branching is hierarchical, with big
branches splitting into smaller ones and those into channels smaller
still. This tends to produce fractal structures, which have the advantage
that they exist ‘between dimensions’: they extend throughout all of
the space they occupy without filling it entirely (which would leave
no room for tissues themselves).
The researchers modelled these networks as systems of tubes which
become progressively thinner at each branching point. There are two
governing principles for the networks. First, all of them, regardless of
size, have to end in tubes of the same size. These terminal branches can
be considered to be the analogues of the smallest capillaries in cardio-
vascular systems, whose size is geared to that of the organism’s indi-
vidual cells—which varies little regardless of the total body size.
Second—and here is the crucial point—the network is structured so
that the amount of energy required to transport fluids through it is
minimized.
For plant vascular systems, the passages of the network are in fact
bundles of vessels of identical cross-section. At each branching point,
the bundles split into thinner bundles with fewer vessels in each. For
this situation, the researchers showed that the 3/4 scaling law of
metabolic rate with body mass falls out quite naturally from an
*This might be easier to see the other way around: the body mass is directly proportional to
the body volume, which varies as the cube—the 3rd power—of the body’s linear dimen-
sions. So the time taken to travel at constant speed across a cube-shaped box depends on
the 1/3 power of the box’s volume.
analysis of the geometric properties of the energy-minimizing net-

work. For mammalian distribution networks, on the other hand, the
situation is rather more complex, and a 3/4 scaling law is obtained
only when the model includes the facts that the fluid flow is pulsed
(due to the pumping of the heart) and the tubes are elastic. Most
importantly, these relationships apply only for fractal distribution
networks. Many human-made systems that distribute fluids or energy
for power generation, such as combustion engines or electric motors,
do not conduct this distribution through fractal networks, and they
show power-law scaling with exponents related to 1/3, not 1/4, as their
mass increases.
This cannot be the whole answer to allometric scaling laws—for one
thing, they are obeyed by organisms that do not have branched distri-
bution systems—but it implies an intriguing significance for fractal
networks in the living world. James Brown suggests that it is in fact
the ability of fractal networks to provide an optimal supply system to
bodies of different sizes that enables living organisms to show such a
huge range in body shapes and sizes, extending over 21 levels of mag-
nification by ten from bacteria to whales. The theory also demands
some rethinking of how biology works. The emergence of the science of
genetics has encouraged the view that living systems from cells to
ecosystems are governed primarily by the transfer of information:
genes control body shapes and determine fitness, and populations
grow and diversity owing to the shifting make-up of their genetic
resources. But the work of Enquist and colleagues places the spotlight
instead on energy: they say it is the need to distribute energy efficiently,
with minimal waste (dissipation), which determines the patterns of the
distribution network, with knock-on consequences for the size of crea-
tures and thus for their abundance, their activity and their longevity.
The researchers have argued that their theory can even explain bio-
logical properties at the level of ecosystems, such as how the density of
plants or animals that occupy a patch of ground depends on the masses
of the individual organisms. Such relationships also tend to follow
scaling laws with exponents that can be explained by considering how
much ‘metabolism’ can be sustained by the available energy. Enquist,
Brown, and West believe that their scaling laws, derived from the
branched networks of distribution systems, may provide a unified
view of how the living world works.
TREE AND LEAF j 143
Webs of life
With his preference for answers to pattern and form that invoke engin-
eering rather than ‘black-box’ Darwinism, D’Arcy Thompson would no
doubt have understood and appreciated all of this. Indeed, he dis-
cussed biological scaling laws in the early pages of On Growth and
Form, explaining for example how the mechanics of a tree dictate
that, if it is not to bend under its own weight, the diameter of the
trunk must increase in proportion to the height raised to the power
3/2. ‘Among animals we see how small birds and beasts are quick and
agile, how slower and sedater movements come with larger size’, he
wrote.
Yet there is, as we have seen, another strand to Thompson’s argu-
ments that demands to know not just why some things are possible and
some not, but how the pieces of the puzzle are put in place. In the same
vein, so to speak, it is one thing to say that a fractal branching network
offers optimal distribution of fluids and minimal energy dissipation;
but the growing organism does not know that. It is clearly not the case
that every little passage of the lungs is specified by a genetic blueprint,
since no two lungs are alike, even in the same body, any more than you
can find a sycamore tree that can be exactly superimposed on another.
These networks have to be grown, and clearly that must happen via the
now familiar combination of chance and necessity: the generic form of
the structure, for example in terms of the fractal dimension and the
average angles of junctions, does not vary significantly from one case to
the next, but the details are always different. What are the rules that
make this happen?
Consider, for example, the blood vessels of the cardiovascular sys-
tem. One of the best studied subsystems is the network of blood vessels
in the human retina (Fig. 5.7), which is a particularly dense-branching
network, since the retina has the highest oxygen requirement of any
tissue in the human body. The ophthalmologist Barry Masters at the
University of Bern in Switzerland has collaborated with physicist
Fereydoon Family in Atlanta, Georgia, to calculate the fractal dimen-
sion of this network, and they find that it has a value of around 1.7: the
same as is seen in clusters of particles grown by diffusion-limited
aggregation, and in cracks and electrical discharges formed in the
dielectric breakdown model. These, we saw earlier, are examples of
processes governed by so-called Laplacian growth, where positive
Fig. 5.7: The blood vessels around the retina form a fractal branching
network with a fractal dimension of about 1.7. (Photo: Barry Masters,
University of Bern, kindly provided by Fereydoon Family, Emory
University.)
feedbacks amplify small, random bumps at the growth front and turn
them into new branches.
Does this mean that the growth of the retinal vascular network is
similarly governed by Laplacian growth instabilities? That is possible,
but it doesn’t necessarily follow. For one thing, we have seen that the
fractal dimension is rather a broad-brush characteristic of a branching
structure: there is no reason to believe that only a single mechanism
can give rise to a network with a fractal dimension of 1.7. And in any
event, the retinal vasculature doesn’t look much like a DLA cluster: it is
far less bumpy and convoluted. Moreover, the process by which grow-
ing blood vessels proliferate and develop branches, called angiogen-
esis, is complicated and doesn’t always generate a diverging, randomly
branched structure—often the vessels are interconnected in more com-
plex ways. For example, blood vessels may sometimes intersect and
join to form closed loops: they do not necessarily exhibit the self-
avoidance we saw in river networks (and which is generally a property
of DLA too). This is particularly evident in the vein networks of leaves
and plants (Fig. 5.8), which grow via processes similar to angiogenesis.
The reconnection between two branches in a vascular system is called
TREE AND LEAF j 145
anastomosis, and it means that there is more than one possible route
for getting from one point to another in the network. So these vascular
systems are more like the Paris metro than like a tree: if you want to go
from A to B, you often have a choice of several possible paths.
The first stage in the formation of a vascular network is the appear-
ance of a web of vessels made from cells called angioblasts. This
appears to happen by chemotaxis: the cells emit some chemical sub-
stance that diffuses into the surroundings and attracts others to move
towards them, homing in on the densest concentration of the attract-
ant. As we’ve seen chemotaxis enables bacteria to talk to one another
and coordinate their movements, forming complex patterns and flows.
Angioblasts emit a protein chemo-attractant called vascular endothe-
lial growth factor, the influence of which causes them to gather into
chainlike bunches that intersect in a web and eventually become veins.
The fractal, branched shape of this web emerges in the next step,
which is angiogenesis. The tissues that the vascular network supplies
with vital fluids get steadily bigger as the organism grows: the channels
Fig. 5.8: The branching patterns of plant vascular systems: ivy leaves
(a), a sea fan (b) and a part of the vein network in a leaf of the African
shrub griffonia (c). (Photos: a, Lars Hammar; b, Gary Rinald; c, Peter
Shanks.)

TREE AND LEAF j 147
are transport networks in an expanding landscape. If any part of the

landscape becomes too distant from one of these supply routes, it risks
becoming depleted in life-sustaining ingredients (oxygen or nutrients,
say), and a new route is needed. But growing a new vein, like building a
new road, has a cost in energy and materials, and it cannot be done
lightly. How are cells at risk identified, and how do they stimulate the
formation of a vein? The form of the vascular network must ultimately
derive from the answers to these questions.
Angiogenesis also depends on diffusing chemical signals. Tissue cells
far from existing blood vessels begin to produce and emit proteins
called angiogenic factors: such cells are said to be ischemic. When the
chemical messengers reach a nearby vessel, it sprouts a new limb which
grows in the direction in which the concentration of the chemical signal
increases, that is, towards its source. In effect the distressed cells are
saying ‘throw me a line’. An analogous process produces the character-
istic vein patterns of leaves (Fig. 5.8), where it seems that the chemical
signal is supplied by the plant hormone auxin. I showed in Book I how
auxin controls the appearance of new leaves or florets at the tip of a
growing plant stem, giving rise to the remarkable mathematical regu-
larities in the arrangements of these features on the plant in the process
called phyllotaxis. In a nice echo of the fractal nature of many plants, it
seems that the same substance dominates the formation of the branch-
ing pattern at the smaller scale of individual leaves.
The distribution of auxin throughout a leaf is often spotty: there are
isolated sites that are rich in auxin, surrounded by regions where the
concentration is lower. New veins will tend to grow towards these auxin
spots, which the plant interprets as a plea for more nutrients. It is
tempting to imagine that the auxin spots correspond to locations
where auxin is being produced at a higher rate—in which case the
shape of the network will be dictated by the underlying pattern of
auxin spots. But it seems that this spotty pattern might instead be
self-organized out of a tendency for the leaf cells to produce auxin at
a constant rate everywhere.
Pavel Dimitrov and Steven Zucker at Yale University have shown how
this may come about. They have represented a portion of leaf as a
collection of cells that all produce and release a signalling hormone,
denoted S. Each cell can ‘detect’ the concentration of S inside itself and
in its neighbours. If an adjacent cell holds much less S, the original cell
may alter its surface membrane so as to let out more S to that
neighbour. Cells that increase their permeability this way are on the
road to becoming vein cells, which conduct S more freely than ordinary
cells. The formation of a nascent vein thus channels the hormone away
from hotspots towards regions of low concentration. Since existing
veins carry the hormone away efficiently, those low-concentration re-
gions are typically such veins themselves—and so a channel grows
between a hotspot and the nearest vein. The more distant a patch of
leaf is from its nearest veins, the more S is likely to accumulate and the
stronger is the signal promoting the formation of a new vein leading to
that site. This process is a little like the way river networks form: rain
falls equally everywhere, but the water runs down the steepest gradient
until it reaches a river, gradually carving out a new tributary.
This model generates the kind of vein growth patterns seen in real
leaves; in particular, new veins tend to join existing ones at right angles.
Veins may appear as ‘stand-alone’ branches, or they may form junc-
tions and loops. As the network evolves on a growing grid of cells, the
hotspots of S move around: old ones are ‘relieved’, while new areas
become starved of access channels. The auxin spots of a growing leaf
are similarly ephemeral, suggesting that there is not after all some
underlying blueprint for where the branches should go. The pattern
draws itself, each stage setting the scene for the next.
This is comparable to the way a combination of diffusion and reac-
tion of chemical substances within a uniform matrix can give rise to
spontaneous pattern formation: the process I described in Book I,
where I showed that it can account for patterns like the spots, stripes,
and networks seen on animal skins. Hans Meinhardt, a German biolo-
gist who has pioneered the understanding of these so-called chemical
reaction–diffusion systems and of how they might produce biological
patterns, was one of the first to suggest that venation networks might be
explained this way. Meinhardt proposed a model that involved two
diffusing chemical signals. In this scheme, the signals stimulate
branches to grow and divide, while branch tips have a tendency to
avoid each other. But growing tips are less strongly repelled by fila-
ments that exist already, and so while tips will not meet end to end, a
single tip might intersect and reconnect with an older branch—this is
the process of anastomasis that I mentioned above. Thus, Meinhardt’s
reaction–diffusion model can also generate realistic-looking vascular
networks; but it is hypothetical and suffers from the fact that it must
TREE AND LEAF j 149
postulate two chemical triggers, whereas in plants only auxin has so far
been identified as one such.
Steffen Bohn and his colleagues in Paris have recognized a similarity
between vascular webs and crack networks in thin layers of brittle
material (see page 94). In both cases, for example, branch intersections
at right angles are common. They have proposed that, by analogy with
cracking, vein networks might be controlled by mechanical forces
rather than chemical signals. They point out that leaves contain both
soft and stiff layers of tissue, which creates stresses, and they speculate
that the directions of vein growth and the angles at which they branch
and intersect might be determined by the way veins push and pull on
one another.
If nothing else, this illustrates that there is more than one possible
way to grow a tree. What seems clear, however, is that nature has found
a way of employing some simple principles that achieve a delicate
balance of chance and determinism to generate distribution networks
beautifully adapted to their role of transporting fluids.
Web Worlds
Why We’re All in This Together
6
I
N the summer of 2003 the lights went out in New York City. In fact,
they went out all over the east side of North America, from Detroit
to parts of Canada, affecting one-third of Canada’s population and
one in seven people in the United States. But there is something about a
stricken Manhattan that always plays to the public imagination, and
here was the city in unaccustomed darkness, its offices abandoned, its
streets benighted, its stock exchange frozen. A state of emergency was
declared.
Inevitably, perhaps, many New Yorkers feared at first that this was a
terrorist act. But it wasn’t. The outage was due to a breakdown of the
power grid itself, on a scale rarely witnessed before.
But surely this grid, so vital to the economy, the security and the
safety of the country’s citizens (imagine if this had happened in the
depths of a New York winter), is designed to avoid such catastrophic
failure? After all, it is a complex, interconnected web that offers many
different routes from one part to another. As in a city road network, if
one way is blocked then there is always another. What event could be so
damaging and pervasive as to undermine this wealth of alternatives?
Culpability was attributed and denied all over the affected territory. The
Canadian Department of National Defense blamed a lightning strike in
the Niagara region. The Canadian prime minister’s office said the cause
was a fire at a power plant in New York. The Canadian Defense Minister
pinned it to an alleged breakdown at a nuclear power plant in
Pennsylvania. The governor of New York State would have none of
this, instead placing the blame on the Canadian side of the border.
In early 2004 the final report of a joint US-Canadian task force
assigned the job of investigating the largest power failure in American
history came up with the following explanation: the power company
WEB WORLDS j 151
FirstEnergy had not been diligent enough in trimming trees in part

of its Ohio service area. The report said that at 1.30 in the morning on
14 August a generating plant in Eastlake, Ohio, stopped transmitting
power because a computer error left it unable to cope with the high
electrical demand. This put a strain on nearby high-voltage power lines,
and they were tipped over the edge into failure when they came in
contact with overgrown trees.
Overgrown trees? Does the north-eastern seaboard of North America
stand vulnerable to some tall trees in the woods of Ohio? Was it this that
led to the shut-down of 256 power plants by a quarter past four that
morning? You might reasonably ask, what kind of idiot designed this
system?
The answer is that no one, idiotic or otherwise, designed this system.
How could they? A power grid of this magnitude is not built like an
office block, according to the blueprint of an architect. It is a growing
thing, continually expanding and mutating in response to changes in
electricity demand, in demographics, in technology, and the need for
repair and renewal. This is not a product of human planning but is, like
cities themselves, more akin to an organism, sustained by the compli-
cated interconnections between its component parts.
Civilization generates many complex networks of this sort, which
evolve over time according to no master plan. Road networks and
urban streets are like this, and so are the webs of trade and travel that
link global air and sea ports. The telephone network was one of the first
technological artefacts to be considered as a complex network, but it is
the emergence of the Internet that has truly placed a spotlight on the
interconnectedness of communications. Beyond this, however, there
are human networks that are less tangible but arguably even more vital
to the ebb and flow of society: our webs of friends and associates, the
networks of business and commerce, the virtual conduits along which
ideas, money, rumour, culture, and disease are conveyed.
Since none of these systems enjoys much by way of the debatable
benefits of rational design, it is quite proper that they be regarded as
aspects of what was once called natural philosophy. We do not know
what governs their growth, nor what structures result. We must explore
them as we would aspects of the natural world. And indeed, it is now
becoming clear that they share many characteristics with networks that
exist in nature, such a food webs or the communication between genes
and proteins in our cells. It is rather remarkable, then, that until the
past decade or so scientists have had so little to say about complex

networks, their patterns of links and the qualities that these engender.
The 2003 North American power failure provides just one among very
many reasons why it is a matter of some urgency to understand them,
to decode their laws of growth and to map out their structures.
All the world’s a stage
Until recently, the only archetypal network one tended to see in human
society was that of a tree: the family tree. Today, mapping one’s family
tree has become a hobby, if not sometimes an obsession, for thousands
of people. It is a habit acquired from the nobility, who were traditionally
much exercised about the quality of their bloodlines, and also from
theologians, who made great play of biblical genealogies, sometimes
depicted in medieval art with a very literal allusion to the arboreal
analogy (the Tree of Jesse, showing Christ’s ancestry, is a favourite
theme for stained-glass church windows). There is an obvious concep-
tual link here with the tree symbolism favoured by early Darwinists, its
branches an elaboration of the classical Great Chain of Being.
Now, the striking thing about this network topos is that it implies
direction and insists on a uniqueness of path: there is only one way to
get from the tree trunk to any particular branch tip. To row upstream in
a river to the headwaters of any specific tributary, you must select the
correct branch every time the channel divides. In genealogical trees,
loops are rare—although much less so for the inter-marrying nobility
than for typical members of society today (unless they live in a small,
self-contained community).
But to understand the structure of our social worlds, a tree is not the
right kind of metaphor. If we think instead about our network of
friends, we will see straight away that loops are the norm. I know Joe
and I know Mary, but of course Joe and Mary know each other, because
we all work in the same office. I met my friend Wendy at a party of my
friend Dave’s, because they are old school pals. So if we map out
these friendship networks by drawing links between friends, we find
it interconnected in many ways. We might imagine that the more
appropriate image, then, is a net or a spider’s web. In this network,
there are many different ways to get from one intersection (or node)
to another.
WEB WORLDS j 153
Mathematicians call this kind of network a graph, and they study its
properties using graph theory. The beginnings of this branch of maths
can be traced to the great eighteenth-century Swiss geometer Leonhard
Euler, who studied a problem posed by the bridges of the East Prussian
city of Königsberg (now Kaliningrad in Russia). There are seven of these
bridges across the Pregel river, five of them giving access to an island
where the river divides. Can one stroll around the city along a route that
takes you over each bridge only once? In 1735 Euler proved that this was
impossible. He did so by converting the layout of Königsberg into a
graph in which each node represented one of the land areas and the
links between them represented the bridges. This was, in effect, the first
theorem of graph theory.
Graphs are a form of map, but they generally take no heed of geog-
raphy, of the spatial distance between points. Instead, they are con-
cerned with the pattern of connectivity, or what is often called the
topology of the network. Metro maps are somewhat like this: although
the spatial positions of the stations on the map are more or less related
to their geographical locations, the distances between stations on the
map typically have little relation to those in the real world. If you lived
your entire life riding the metro, then you wouldn’t care about any
approximate geographical realism, such as whether a station lay to
the north or the east—all that would matter is the connectivity, or
how to get from one station to another. And when a graph does not
reflect a spatial structure—for example, if it depicts a friendship net-
work—then the actual positions of each node on the diagram have no
significance at all; everything of importance is in the topology.
So what might a real friendship network look like? Sociologists and
anthropologists have had a long-standing interest in that question. One
of the first attempts to give it a thorough mathematical treatment came
in the 1950s, when the mathematician Anatol Rapaport at the
University of Chicago was seeking to understand how infectious dis-
eases propagate through a population. If the disease is passed on
through personal contact, then clearly its rate of transmission depends
on how often infected people encounter others who are not infected (or
immune). An outbreak in an isolated community might devastate that
community but have little impact on the world beyond. By contrast, in
a big city where individuals come across many strangers every day, we
might expect that the disease would quickly reach epidemic propor-
tions. So there is presumably some critical level of average connectivity,
in terms of the number of strangers encountered in a day, above which

the outbreak switches from being contained and localized to uncon-
tained and pervasive. At what point does this happen?
Rapaport didn’t know what a typical ‘encounter network’ looked like
for any society. But he guessed that it would be reasonable to assume
this was pretty random: you have an equal chance of bumping into
everyone in the population. Moreover, it seemed sensible to imagine
that there was a well-defined average number of such encounters—that
is, an average connectivity, denoted K. You might, for example, get close
enough to 30 other people each day, en route to work or in the shops
say, to infect them. Obviously randomness is not the whole story,
because there is a much bigger probability that we will encounter
(and in this setting, infect) our immediate family members and work
colleagues than we will Joe Random. But the assumption of random-
ness seems a good place to start.
The social network on which the disease spreads, then, is represented
here by a series of nodes (people) connected at random to an average of
K other nodes (Fig. 6.1a). Rapaport and his colleagues were able to
show that in this case the fraction of the total population that becomes
infected increases exponentially as the average degree of connectivity K
increases: the more connections to each node, the greater the spread of
disease. That is what you would expect, of course, but Rapaport put that
expectation into precise mathematical terms.
At the end of the 1950s, the random graphs that Rapaport introduced
were studied more rigorously by the Hungarian mathematician Paul
Erdös and his collaborator Alfred Rényi. They wanted to understand
what general properties these graphs have. That is not an easy thing to
assess, because by definition there is an immense number of different
random graphs that connect any given set of nodes: each graph is
constructed by forging links at random between any two nodes, no
matter how close or far apart they are. You can ‘grow’ such a graph by
following an algorithm, repeating it again and again:
1. Choose a node at random.

2. Choose another node at random.
3. Draw a link between them.
This procedure does not guarantee that all nodes will have the same
number of links; in general, they will not. But each graph has an average
number of links per node (K) which increases steadily as you keep
WEB WORLDS j 155
Fig. 6.1: Social

networks can be drawn
as vertices or nodes
(circles), representing
individuals, linked by
lines that could, for
example, represent ties
of friendship, or
‘infectious’ close
contacts made during a
day, or sexual contacts.
There is a distribution of
degrees of connectivity
k—some individuals are
highly connected, others
have few connections
(a). In a random network
or random graph, where
links between
individuals are forged at
random, the distribution
of k values has a
particular mathematical
form, with a well-
defined average value.
Such networks are
disorderly, in contrast to
regular grids (b), where
all vertices have
precisely the same
degree of connectivity
(in this square grid, the
connectivity k is 4).
iterating the steps above. One of the questions Erdös and Rényi asked
was: how well-connected is the graph as K increases? If there are only a
few links between many nodes, then some nodes will remain isolated,
while others will be connected only into little patches. If there are many
more links than nodes—if K is large—then you may have a good chance
of finding a route between any two nodes in the network. How does this
‘connectedness’ of the entire array of nodes alter as K is increased? You
might expect it to increase steadily, but Erdös and Rényi calculated that
in fact there is an abrupt change at a certain critical value of K equal to 1.
This can be seen by looking for the largest interconnected set of nodes
in the system. If each node has, on average, less than one connection,
this largest interconnected fraction stays small—negligible in compari-
son to the size of the entire array. But once K exceeds 1, the largest
interconnected component grows in size very rapidly, quickly ap-
proaching the size of the entire network.
This is critical for problems of epidemiology such as those Rapaport
studied. It tells us that, for random networks, the spread of disease can
be explosive once the connectivity exceeds a certain threshold, because
at that point just about everyone becomes connected to everyone else.
There is another important property of a random graph, which is
related to how quick it is to navigate. Suppose you want to go from any
one node to any other. If the degree of connectivity is well above the
critical threshold, so that essentially all of the nodes are interconnected,
there are in general many alternative routes you can take. Measured in
terms of the number of links traversed, some of these can be very long. But
some are rather short, because the principle of random interconnection
forges many shortcuts between nodes that are ‘distant’ in visual terms.
(Recall that in general these visual distances do not have any physical
meaning, they’re just a way of representing the nodes on a two-dimen-
sional plot. So the ‘length’ of each link has no meaning either—that is why
we can think about path lengths merely in terms of the number of links
they incorporate. It is perhaps better to say that the ‘shortcuts’ ensure that
every part of the array of nodes is likely to be accessible from every other
part by a few hops, rather than by a succession of many links.)
Just as we can define an average connectivity K for a random graph,
so we can define an average path length L: this is the average number of
links you must cross to get from one node to another. For a random
graph, this length can be surprisingly small even for a big network. We
have a familiar expression for this: ‘it’s a small world’.
WEB WORLDS j 157
It is familiar because that is what experience tells us. You meet

someone at a party who you have never seen before, but it turns out
that she went to school with your brother-in-law. In my experience, the
older you get, the more this seems to happen. If social networks are
random networks, however, it is no surprise. The small-world effect
suggests that the random network is a much better description of social
networks than a regular grid is, in which nodes are connected only to
their immediate neighbours (Fig. 6.1b). In that case there are no short-
cuts: to get from one node to a ‘distant’ one, you have no option but to
pass along all of the links in between.
But there is one problem: social networks are not random.
This is obvious the moment you stop to think about it, for the reason
I alluded to above. What the random-network model says is that two of
your good friends have no greater chance of knowing each other than
they have of knowing anyone else in the population. But this clearly
isn’t so. I met Wendy at Dave’s party, because Wendy and Dave are old
friends. We meet our friends’ friends, and they become our friends too.
Or perhaps we all met in a group, at work, or at college, or at the sports
club. To put it technically, if there are links in a social network from A to
B and A to C, then there is a much higher chance of a link between
B and C than between either of these and some other, random node.
Friends tend to form clusters with a strong degree of interconnectivity,
while these clusters are linked to other clusters via a lower density of
links (Fig. 6.2). The same is true of other social networks: businesses
tend to show clustering in their dealings with other companies, as do
the collaborations between scientists or musicians.
Clustering seems to work in opposition to the small-world effect of
short path length. In effect, clustering implies that there is a higher
chance of a node forming a link to another in its vicinity than to one far
off: there is a bias against making shortcuts. Indeed, clustering is
precisely what one would expect from grid-like graphs, where all links
are ‘local’, joining near neighbours. How can a social network simul-
taneously show the high clustering characteristic of a grid and the short
average path length characteristic of a random network?
The physicists Duncan Watts and Steven Strogatz, working at Cornell
University in the late 1990s, showed how that can happen. They came
up with a model that allowed them to convert a network gradually from
a grid-like structure to a random structure. It was called random rewir-
ing, and it was a surprisingly simple idea. You started with a grid, and
Fig. 6.2: Real social networks tend to be clustered: groups of friends

know one another, linking them in a dense web, while there is a
lower density of links to other groups.
‘rewired’ the nodes one at a time. At each step, a node was selected at
random and one of its links was redirected from a near neighbour to a
randomly chosen node, which could be near or far. Watts and Strogatz
began with a circular grid of nodes, since that has no edges (where the
connectivity of the nodes is different from that in the interior). The
circular grid on the left of Fig. 6.3a may not look like a typical grid, but it
is: each node is connected just to its two neighbours and its two next
nearest. As the random rewiring proceeds, the regular structure disap-
pears into a tangled jumble of links.
As this happens, both the amount of clustering and the average path
length decrease, as expected: these are both large for a regular grid, and
small for a random network. But the crucial point is that they don’t
decrease at the same rate. The path length drops abruptly after only a
few rewirings: a handful of shortcuts is enough to bring most of the
nodes ‘close’ to most of the others. But the clustering changes rather
little until more rewiring has been effected. So there is a family of
networks, for rewiring degrees between these two limits, which have
small L but high clustering (Fig. 6.3b). These have properties that seem
to resemble our social ‘small worlds’, and Watts and Strogatz called
them small-world networks.
WEB WORLDS j 159
Fig. 6.3: The rewiring networks of Strogatz and Watts are steadily changed from regular grids to
random networks by breaking and remaking links at random (a). For a certain range of rewiring
between these extremes (0 ¼ regular grid, 1 ¼ fully random network), the networks are ‘small
worlds’, with short average path lengths L between any two points, but significant degrees of
clustering C (b).
Are social networks really like this? It is very difficult to gather data on
people’s friendship networks—I’ll come back to this shortly—but Watts
and Strogatz took a different approach to the question. They looked at
the network that links film actors according to whether they have
appeared together in a film. Each node of this grid is an actor, and there
is a direct link between them if they have been co-stars. The virtue of
this network is that it is precisely defined—either two actors were in the
same film or they were not. Furthermore, the network is already known,
thanks to a game invented in the early 1990s by a group of American
college students. These film buffs had come to the conclusion that
the actor Kevin Bacon was the centre of the film universe. It was not
that Bacon was a particularly wonderful actor (though he is fine), nor
that he is a particularly big star; but in the movies of that period he

seemed to crop up everywhere. He had worked with countless better-
known stars. (It helped that his early successes included the ensemble
pieces Animal House and Diner.) In the Kevin Bacon Game, the aim was
to link any named actor to Kevin Bacon in as few steps as possible. Elvis
Presley? He appeared in Harum Scarum (1965) with Suzanne
Covington, who appeared with Bacon in Beauty Shop (2005). So you
can reach Bacon from Elvis in two hops: Elvis has a Bacon Number of 2.
All this can be deduced from the Internet Movie Database, which lists
just about all the commercial films made since 1898 (about 200,000 of
them). But Watts and Strogatz did not have to comb through it, because
the film-star network had already been mapped out by computer sci-
entists Brett Tjaden and Glenn Wasson at the University of Virginia,
who had decided to automate the Kevin Bacon Game. At the Oracle of
Bacon,* you can find the Bacon Number of any actor in an instant,
along with the shortest path linking that actor to Bacon. (That’s where
I got the Elvis link—I’m afraid I’d never heard of Suzanne Covington.)
Tjaden was more than happy to give Watts and Strogatz access to his
database.
The researchers hoped that the film actor network would serve as a
surrogate for other social networks. Clearly, making a film together is
not the same as striking up a friendship (it can notoriously mean quite
the opposite), but it seemed reasonable to assume that the two might
share some features. And if nothing else, the film-actor net might be a
good model of professional contact webs. In any event, Watts and
Strogatz established that the film network has precisely the character-
istic they identified for a small-world network: it has an average path
length comparable to that of a random network containing the same
number of nodes and links, but much higher clustering.
In fact, the average path length was 3.65, which means that on
average any actor can be connected to any other in between three
and four hops. Given that the database spans a century and many
different countries, that seems remarkable. But what of Kevin Bacon?
The average Bacon Number currently stands at about 2.9, which im-
plies that Kevin Bacon is indeed better connected than an average
actor—it takes less than three jumps, on average, to reach him. Yet
there is nothing so special about that: there are many actors with
*See http://oracleofbacon.org.
WEB WORLDS j 161
comparably small average path lengths, and over a thousand are better
connected than Bacon. The best is currently Rod Steiger: the average
Steiger Number is 2.68. As Duncan Watts has pointed out, in a small-
world network just about everyone seems to be the centre.
The Kevin Bacon Game echoes a familiar trope popularized by John
Guare’s 1990 play Six Degrees of Separation, in which a character claims
that ‘Everyone on this planet is separated by only six other people.’
Where did Guare get that idea from? Why six? In 1967 the Harvard social
scientist Stanley Milgram devised an ingenious experiment to measure
how well-connected social networks are. He sent 196 letters to ran-
domly selected people in Omaha, Nebraska,** asking them to forward it
to a stockbroker from Sharon, Massachusetts who worked in Boston. All
Milgram provided was the man’s name, along with a curious request:
rather than trying to track the stockbroker down, the recipients were to
send the letter to someone else they knew personally, who they felt
might be better placed to know the man. (Perhaps they might have
relatives in Boston, or were stockbrokers themselves, say.) All recipients
were asked to do the same, until the letter reached its destination. And
surprisingly, some of them did. Even more surprising, however, was the
number of journeys those letters took to get there. On average, just six
were required: just five intermediaries between the start of the chain in
Omaha and the Bostonian stockbroker. A small world indeed.
There are many other indications that our social and profess-
ional networks share this property of short average path length.
Mathematicians have long enjoyed their own version of the Kevin
Bacon Game in which they look for the shortest path that links them
with Paul Erdös, the founder of random-graph theory. It is not this
distinction that led to Erdös being singled out, however, but the great
number of collaborators he worked with on his many papers (he pub-
lished over 1,500 of them). The average Erdös Number in the network of
mathematicians and scientists that can be linked to him this way is
about 4.7.* Mark Newman, working at the Santa Fe Institute in New
Mexico, has established that the average path length in the collabor-
ation network of the 44,000 scientists who have placed preprints of
*Mine is 5, so far as I know, but I take heart from the fact that Erwin Schrödinger’s is 8.
**Actually they were not all selected at random, and another group of recipients in this
study weren’t in Nebraska at all—see page 171.
their papers on the archive set up by physicists at Los Alamos National

Laboratory is 5.9: again, six degrees of separation. ‘This small-world
effect is probably a good sign for science’, Newman concludes. ‘It shows
that scientific information—discoveries, experimental results, theor-
ies—will not have far to travel through the network of scientific ac-
quaintance to reach the ears of those who can benefit by them.’
The same small-world property is found in other collaborations. The
physicists Pablo Gleiser and Leon Danon in Barcelona have looked at
the network formed by over a thousand jazz musicians listed in the Red
Hot Jazz Archive who played in bands between 1912 and 1940, and find
that it has an average path length of just 2.79.
But Watts and Strogatz also looked at the properties of two networks
that have nothing to do with social systems: the electrical power grid of
the western United States, and the links between neural cells in the
nematode worm Caenorhabditis elegans. The latter is one of the best
studied multicellular organisms, since it has a relatively small number
of cells: precisely 959 in the female and 1,031 in the male. Of these, 302
form a primitive nervous system, in which the pattern of connectivity is
known exactly. Watts and Strogatz found that both these networks have
the small-world feature of an average path length close to that of the
corresponding random network, coupled to a considerably higher de-
gree of clustering.
The rich get richer
But what do these networks actually look like? Watts and Strogatz’s
random rewiring model generated specific examples of small-world
nets, but these were rather artificial, being pinned to a ring of nodes in
which each node was constrained (by the way the model was set up) to
have precisely three links. Clearly, social networks are not like that. But
Watts and Strogatz assumed that they are not so different: they imagined
that even if we don’t all have precisely the same number of friends, this
number probably doesn’t differ very much. In other words, there will be
a certain average number of friends per person, and progressively fewer
people will have progressively less or more links than that.
But, Watts confesses with some chagrin, they didn’t check that this
was really how things are. It turns out that many of the small-world
networks in the real world do not look like this at all.
WEB WORLDS j 163
The first intimation of that came in 1999, when Albert-László Barabási

and his colleagues Réka Albert and Hawoong Jeong at the University of
Notre Dame in Indiana investigated the connectivity statistics for over
300,000 web pages on the domain of their university. Each web page is
connected to others by hyperlinks, where a click on the mouse will take
you to the other page. The researchers found that this segment of the
World Wide Web (WWW) is a small world: they estimated that, if it is
representative of the Web as a whole, then any web page is connected to
any other by an average of just 19 links—even though the WWW at that
time held over a billion documents. And because of this small-world
structure, that average path length will not increase much even at the
rapid pace with which the WWW is growing. Even if the number of pages
expands tenfold, the average distance between pages would grow by just
two links. This slow growth in average path length as a network expands is
one of the characteristics of a small-world network, and is the result of the
large number of shortcuts that very quickly weave new nodes into the web.
But the pattern of links was quite different from what Watts and
Strogatz had implicitly assumed. Most pages had only a single link,
though some had hundreds, and a very few had several thousand.*
Now, you might intuitively expect there to be many more pages with
rather few connections than with many; but the connectivity statistics
that Barabási and colleagues found are different from what you might
expect if the connections were being made independently and at ran-
dom. In fact, they have a very particular mathematical form: the number
of pages having k connections is proportional to the inverse of k raised
to some power. This is, in other words, yet another example of a power
law. We saw in Chapter 2 that systems whose structure is described by
power laws have the general property that they are scale-free—there is
no characteristic size to them. What this means here is that there is no
‘typical’ number of links to a node of the network.
These connectivity statistics tell us what the probability is of a node
picked at random having a particular number of connections k.
Although this chance gets smaller as k gets bigger, it does so consider-
ably less quickly than it would either for the kind of networks obtained
from random rewiring or for random graphs. In other words, the WWW
*The picture is slightly complicated by the fact that there is a difference between incoming and
outgoing links. A hyperlink only takes you in one direction—you cannot necessarily get back to
your starting page from the one you end up at via a hyperlink. But Barabási and colleagues
found the same statistical pattern for both incoming and outgoing links.
has a disproportionate number of very highly connected pages. This

means that, when the network is plotted out on a page, parts of it seem
to be ‘pinched’ where a large number of links converge on these highly
connected nodes (Fig. 6.4a). In contrast, a random graph looks relatively
uniform throughout (Fig. 6.4b). The physical structure of the Internet
(by which I mean the actual links between computers, not the virtual net
that connects web pages) also has this shape (Fig. 6.4c).* It means that
Fig. 6.4: Scale-free networks look ‘pinched’ at a few highly connected vertices (a), whereas
random graphs are rather uniform (b). This pinched quality is evident in the structure of a part of
the Internet (c ; see http://www.cybergeography.org/atlas/topology.html). (Images: a was prepared
using NetLogo software, available at http://ccl.northwestern.edu/netlogo/; b: courtesy of Paros
Oikonomou and Philippe Cluzel, University of Chicago.)
*An analysis of this structure—actually that of a subset of over 4,000 nodes of the Internet—
was conducted by the brothers Michalis, Petros and Christos Faloutsos, all of them com-
puter scientists, in 1999, at much the same time as Barabási and colleagues were mapping
the WWW. The Faloutsos brothers found precisely the same kind of power-law relation for
the connectivity of nodes.
WEB WORLDS j 165

not all of the nodes are equal—some enjoy much better connections
than others. There are many nonentities, but a few celebrities.
It now seems that many diverse networks have this same topological
structure, with power-law distributions of connectivity. The pattern is
found, for example, in email communications (where a link is estab-
lished between two nodes if an email is sent between them), in the web
of direct flights between airports, in the network of trade links between
countries—and in the network linking movie actors. Outside the
human social sphere, Barabási and his coworkers have found scale-
free networks in the biochemical pathways of living cells. For example,
they looked at how all the molecules involved in the metabolic chem-
ical reactions of Escherichia coli bacteria, and the protein enzymes in
brewer’s yeast, interact with one another, with a link existing between
them if they participate together in a particular chemical process. In
both cases the network was scale-free (Fig. 6.5).
Where does this topological structure come from? Graph theory has
traditionally considered the different ways that a bunch of nodes may be
wired together; in the random graphs of Erdös and Rényi, links are made
between two randomly selected nodes. But Barabási and Albert realized
that many real networks grow almost like plants branching from a
seedling: they approached the origin of scale-free networks as a prob-
lem of growth and form. The World Wide Web is growing every day as
new web sites are created and new pages added. Whenever a new node
is plugged into the network, the question is, to which others should it
connect? A rule that the link is always made to the nearest node, for
example, would tend to generate a grid. A rule that the selection is made
at random will produce a random graph. But Barabási and Albert said
that scale-free networks grow according to another rule: the new node
connects to an existing node at random, but with a bias: the more links a
node already has, the more likely it is to be chosen. A node with two links
is twice as likely to be selected as a node with one.
This means that the more connections a node already has, the more
likely it is to acquire more as the network grows. But that doesn’t mean
that all the best-connected nodes are guaranteed to be awarded all the
new links, because there is an element of chance in the choice. If there
are lots of nodes, then even an extremely well-endowed one will have
only a relatively small chance of being selected compared with the
chance that a new link will go instead to any one of the others. That is
why the most well-connected nodes in a scale-free network are also the
WEB WORLDS j 167
Fig. 6.5: A part of the network formed from molecules involved in yeast metabolism. Each vertex is
a molecule, and the links denote enzymatic reactions that convert one molecule to another. (Image:
Hawoong Jeong, Korea Advanced Institute of Science and Technology.)
most rare. All the same, the implication of this rule of ‘preferential
attachment’ is plain to see: in terms of connectivity of nodes, the rich
get richer. The connection rule guarantees a persistent inequality of
connectedness—and Barabási and Albert showed that the nature of this
inequality is described by power-law statistics.
In retrospect, this is no surprise. The ‘rich get richer’ principle is
precisely what seems to happen in capitalist societies: wealth attracts
yet more wealth. If that is so, we should expect it to lead to a power-law
distribution of wealth, in which a very few individuals are obscenely rich.
In 1897 the Italian economist and sociologist Vilfredo Pareto showed that
this is the case in many societies: at least for the rich end of the income
distribution, the figures follow a power law, which economists now call
the Pareto law. The American sociologist Robert Merton has dubbed this
the Matthew Principle, since the Gospel of Matthew provides one of the
first known descriptions of this particular injustice: ‘For unto every one
that hath shall be given, and he shall have in abundance; but from him
that hath not shall be taken away even that which he hath.’
Why, though, should networks grow according to this principle? There
isn’t really a general argument that works for every scale-free net from
metabolic pathways to film actors, but in social networks it often boils
down to the issue of fame. The better known you are, the more likely
your fame will be boosted even more. Think about the World Wide Web:
you have made a new web page and want to provide a link to some
standard reference source for a particular aspect of what it describes.
The chances are that you will choose the link that you can see others
have chosen for the same purpose. Of course, these days you’re likely to
find that link by a Google search, but that makes the bias even more
strongly deterministic, because the Google page ranking scheme de-
pends on how many links a page has.* And so pages acquire links not
because someone has reviewed all the alternatives and decided that it is
the best reference source, but because it is already ‘famous’. The same is
true for citations in the scientific literature, which also have power-law
ranking statistics: people cite a book or a paper because that is what
others have done, and not because they have read it themselves.
This doesn’t mean that a node’s connectedness in a network like this
bears no relation to any genuine merit. It may be that a web page or a
citation begins to attract more links than others because it really is
good. But as long as the law of preferential attachment operates, the
connectedness of different nodes probably does not reflect their real
differences in merit, because the effect of fame artificially inflates some
nodes over others, perhaps in ways that even invert the real distinctions
of quality between them.
World of webs
Does this mean that all small worlds are formed from scale-free net-
works? Not at all. After all, Watts and Strogatz began the whole story
*The page rank is a little more sophisticated than that, since it also takes into account the
connectedness of the pages from which the links are coming; but the Matthew Principle still
operates.
WEB WORLDS j 169
by constructing small worlds that did not have this property. And it
turns out that power grids are not scale-free either: the electricity grid of
southern California, for instance, doesn’t show the characteristic
power-law relation between the connectivity of its nodes. That might
have something to do with the fact that power grids are physical entities
that exist geographically in two dimensions: this makes it very hard to
establish extremely highly connected nodes, because any node can
only have a limited number of near neighbours, and direct links from
very distant nodes are uncommon. The same is true of road networks,
which are highly interconnected but not scale-free, and indeed are not
even small-world networks at all: they are more like regular grids.
Even scale-free networks generally have a limit: in theory there is no
ceiling to the number of links a node might have, but in practice there
is. For example, no actor, however much in demand, can make a million
films before they retire or die. The capacity of an airport is ultimately
limited by the number of runways, facilities, and so on. What this
means is that the statistics deviate from a power-law relationship for
very high connectivities so that the probabilities are lower than the
power law predicts. Luı́s Amaral at Boston University and his colleagues
showed in 2002 that this is true even for the WWW network, simply
because the web is too vast for anyone choosing to make a hyperlink to
survey the entire system before selecting the target: the information
about the web has to be filtered before it can be processed at all, which
means that only a subset of all the available nodes is ever considered.
What about real social networks of friends: are they scale-free?
This isn’t clear, because it is extremely hard to gather data, and harder
still to know how general it is. A study of an acquaintance network
among 43 Mormons in Utah, conducted in 1988, and another of 417
secondary-school students in Wisconsin in the 1960s, both seem to
show a statistical distribution that is not scale-free but has a well
defined average number of links. Nonetheless, these networks are
small worlds in the sense that any one person is linked to any other
by only a small number of links.
There is another important characteristic of scale-free networks that
gets rather lost if we focus only on the power-law scaling of the con-
nectivity of nodes. If you look at the graphical representation of the
Internet in Fig. 6.4c, one of the things that strikes you first is that it
seems to be built up from a number of clusters: densely radiating
groups of nodes that look like the head of a dandelion in seed, linked
to one another by a more sparse web of links. There is, in other words, a
number of distinct communities within this web. On the Internet, these
modules seem to be derived largely from geography: each module
corresponds to the sub-network of an individual country. There are
also sub-modules that reflect particular professional communities,
such as military sites. It is no surprise that a community structure exists
in many if not most social networks, for that is after all a reflection of
how our lives tend to be organized. In the network of scientific collab-
orations, those people working in the same discipline, and in the same
sub-field of a discipline, are likely to be bound into a community.
Friendship networks might be structured around a neighbourhood or
a workplace. In some sense, modularity is a reflection of the high
degree of clustering characteristic of small worlds.
And yet it is possible to create scale-free networks that do not have
this modular community structure. So the fact that some of them do is
telling us that there is more to the shape of a network than is revealed
simply by its connectivity statistics. But it is not always an easy matter
to find this community structure: to work out what the modules are and
where their boundaries lie. One of the problems is that, because there is
a strong element of randomness in the way these networks grow, two
groups of nodes might have only a small number of links between them
purely by chance, rather than because they genuinely constitute dis-
tinct communities in a meaningful sense. The task is then to work out
not just if two groups are joined by few links, but if they are joined by
fewer links than we would expect purely by chance.
Various techniques have been devised for teasing community struc-
ture out of complex networks. Extracting this ‘buried’ information is
often of great value, helping to make sense of what otherwise might
look like just a mess of ‘wiring’. In a metabolic network, for example, it
could tell us something about how a cell’s biochemistry is organized
into functional modules. Mark Newman has used a community-finding
scheme to reveal the undercurrents in purchases of books on US pol-
itics through the online bookseller Amazon.com. He studied a network
of 105 recent books, representing nodes, which were linked if the
Amazon site indicated that one book was often bought by those who
purchased the other. The analysis showed a clean split into communi-
ties containing only the ‘liberal’ books and only the ‘conservative’ ones,
as well as two small groups that contained a mixture along with
some ‘centrist’ titles. Newman found a similar political split in links
WEB WORLDS j 171
between over a thousand blogs. This clear division, Newman says, ‘is
perhaps testament not only to the widely noted polarization of the
current political landscape in the United States but also to the cohesion
of the two factions’; put another way, it suggests that people only want
to read things that reinforce their own views.
Newman has highlighted another aspect of the deep substructure
of complex networks. In some of them, highly connected nodes show
a greater-than-average propensity to have links with other highly con-
nected nodes, forming what has become dubbed a ‘rich club’. This
phenomenon is known as assortative mixing. Obviously, the existence
of rich clubs in social, economic, and professional networks could have
an enormous impact on the way society functions—it might imply, for
example, that the ‘rich club’ members are able to share privileged
information that percolates only slowly into the rest of the network.
Newman has shown that while neither the random graphs of Erdös and
Rényi nor the scale-free networks grown from Barabási and Albert’s
‘preferential attachment’ model has rich clubs, many real-world social
networks do, including the collaboration networks of scientists, film
stars, and company directors. On the other hand, some natural net-
works are negatively assortative: they show fewer than expected links
between ‘rich’ nodes. This is true of the Internet, and to a small degree
of the WWW; and it also applies to the protein interaction network of
yeast, the neural network of the nematode worm, and the marine food
web. In other words, there is something different about social networks
compared to others, either technological or biological: we humans
seem disposed towards forming rich clubs. Newman has argued that
this is because of the strong tendency of social networks to partition
into communities, which is less prevalent in other webs.
Searching the web
One can begin to see, then, that the shape and form of networks can
have a crucial bearing on how it performs its function. For instance,
how does the topology affect the ease with which a network can be
navigated? Clearly, we might expect to be able to get around rather
quickly on any network that has the small-world property of a short
average path length between nodes, because we can always find a
shortcut. But that could depend on having a map, for otherwise how
do we know which is the best link to follow? Search engines are of

course invented with the explicit aim of conducting that search for us,
but even the best of them cannot always intuit our wishes from a
handful of search criteria, and in any case the WWW is now so vast
that no search engine can index and encompass it all.
The key characteristic of social networks revealed by Milgram’s
epochal experiment is not simply that they are small worlds but that
they are searchable small worlds. It would be of no value that a short
social path exists between two individuals if we were unable to find it.
But somehow—and this is what is truly surprising—we do, at least
some of the time. Why is that?
Actually, there is a lot that is swept under the carpet in this ‘some of
the time’. The truth is that the rate of ‘successful searches’ in a case like
this is rather low. One fact about Milgram’s experiment that is often
overlooked is that the starting points in his chains were not by any
means all randomly selected citizens of Omaha, Nebraska. Of the
roughly 300, about 100 were stock investors (and remember that the
target was a Boston stockbroker), and another 100 lived not in Nebraska
but in Boston! And of the 96 letters that began their journey from
Nebraskans picked genuinely at random, only 18 reached the destin-
ation. In other words, there was a big drop-out rate. Some of this can be
attributed to mere apathy. But Duncan Watts and his colleagues have
found that whether a journey like this is completed seems also to
depend on the perceptions of the participants about their proximity to
the target. In 2003 they staged a re-run of the Milgram experiment using
email, which made it easier to enlist volunteers: they accumulated more
than 60,000 in 166 different countries. And the targets were similarly
varied: 18 in total, in 13 different countries, with a wide variety of
backgrounds. Of the 24,163 chains that were started, only 384 were
completed, again showing that this kind of search has a high attrition
rate. But one particular target had a much lower drop-out rate than the
others: an American Ivy League professor. The total path lengths for
these chains were not significantly greater than the others; it was just
that they did not break halfway. Since over half of the participants were
college-educated Americans, Watts and colleagues figured that these
chains were more likely to run to the end because the intermediaries
believed that they would: they thought it more worthwhile forwarding a
message intended for an American academic than for an Estonian
archivist, because it was more likely to succeed.
WEB WORLDS j 173
All the same, it remains surprising that the small-world nature of

social networks gets exploited at all in these experiments, since there is
no way for any intermediary to know that he or she is really routing the
message in an efficient direction. In 2002 Duncan Watts, Mark Newman
and the mathematician Peter Dodds set out to understand why this
happens. They suggested that the community structure of the network,
and in particular the identification of its members with specific groups,
is an essential part of the searchability. They supposed that each indi-
vidual mentally breaks down the network into hierarchical arrange-
ments of groups. They do this in more than one way, for example
imagining groups based on profession and on geographical location:
these different hierarchical pictures overlap. This gives each individual
a picture of their ‘social distance’ from others. Most members of the
network lie way over this social horizon and are not even considered:
people think only about their local links to those a short ‘social dis-
tance’ away. The researchers explored a wide range of ways in which
individuals might conceptualize the hierarchy of the network and forge
links between neighbours who are more or less similar to themselves.
They found that in most cases the resulting networks were searchable
in the sense that, if individuals route a message only using their local
view of the network, its average path length between origin and destin-
ation is small.
In other words, people do not need to see the whole map; a parochial
view is enough to ensure efficient routing, as long as everyone feels that
they are part of overlapping communities. It is this multiplicity of
communities that is the key to the social small world, the researchers
claimed, because that is what allows ‘unexpected’ shortcuts to be
found. Joe might pass a message to Mary because they work in the
same office; but then Mary passes it to her brother Ray, who lives in
another country. But Joe doesn’t even know that Ray exists—he has
never spoken to Mary about her family. So long as Mary didn’t come
into the conversation, Joe and Ray would say that the social distance
between them is huge. The shortcut via Mary passes through two
different ‘social dimensions’—work and family—and so is not apparent
to either Joe or Ray in the absence of other information, as it would be
if, say, Joe, Mary and Ray were all working in the same profession.
This picture of a ‘multidimensional’ social web is supported by re-
search by mathematician David Liben-Nowell of Carleton College in
Minnesota and his co-workers. They studied the network formed by
around half a million bloggers in the online community LiveJournal who

live in the United States. The attraction of this network is that members
explicitly list other members who they consider to be friends. This
allowed Liben-Nowell and colleagues to deduce how much friendship
depended on geographical proximity. For an online community, one
might expect that geography hardly matters, but in fact the researchers
found that two-thirds of the friendships depend on the physical dis-
tance between the individuals. This of course means that the remaining
one-third of the friendships are independent of geography, and arise
because of some perceived similarity in some other ‘social dimension’.
The researchers then ran another Milgram-style email-routing ex-
periment on this network. They didn’t actually get the bloggers to do
the forwarding, but instead conducted computer simulations of what
would happen if each person were to send the message to the friend
who was geographically nearest to the target. In this way, 13 per cent of
the messages reached the target in just over four steps. This seemed
puzzling, because the computer scientist Jon Kleinberg had shown that
routing a message according to geographical proximity along a grid
does not allow the network to be navigated at all efficiently—there is no
small-world effect.
The reason for that apparent contradiction, said Liben-Nowell and
colleagues, is that the chance of becoming friends does not simply
decrease the further apart they are, but depends on how many other
people are nearer: in other words, it depends on the size of the pool
from which geographically closer friends might be selected. The chance
of two people becoming friends depends on how many others lie in
between—which is a question both of geographical distance and of the
density of others in that gap. A network like this is quite different from
the uniform geographical grid that Kleinberg considered, and has the
virtue of small-world navigability.
When you want to pass something like information around on a
network, a small-world nature is beneficial: it allows the distribution
to reach all corners of the net pretty efficiently. But there are some
things that we would rather not see spreading through our social and
technological networks—for example, diseases and computer viruses.
Biologists have studied the transmission of disease through popula-
tions for over a century, but it is only recently that they have started to
realize how important a role the topology of the network might play in
this. The standard model for studying epidemics supposes that
WEB WORLDS j 175
individuals come in two classes: healthy and infected. If a healthy

person encounters an infected one, he or she has a chance of becoming
infected. But meanwhile, infected individuals can recover and become
healthy again. The disease then spreads at a rate that depends on the
relative probabilities of infection and recovery. This so-called suscep-
tible-infected-susceptible (SIS) model says that if this spreading rate
exceeds some threshold value, the disease becomes an epidemic,
sweeping through the entire population and persistently infecting
some constant proportion. But if the rate is smaller than this threshold,
the disease dies out. The aim, then, is to keep the spreading rate below
the epidemic threshold, for example by vaccinations that make the
infection probability small enough.
The physicists Romualdo Pastor-Satorras and Alessandro Vespignani
looked at how the SIS model plays out if the encounters between people
are described by a scale-free network. This changes the behaviour
significantly: there is no longer an epidemic threshold, and all diseases
can pervade the network no matter how slow-spreading they are. That
fits with what we find for computer viruses, which are transmitted by
messages passing through the scale-free email network: they are frus-
tratingly hard to eradicate completely, remaining active in the system
by persistently infecting a very small fraction of computers.
How about human diseases? The way they spread depends on
whether encounters between people create a scale-free network. In
the age of cheap air travel, there are surely plenty of shortcuts between
distant populations, which is one of the big concerns about the threat
posed by new infectious diseases, such as novel and virulent strains of
influenza that threaten to emerge from the avian flu virus H5N1. In
2005 Dirk Brockmann at the Max Planck Institute for Dynamics and
Self-Organization in Göttingen, Germany, and his colleagues found
some evidence that our freedom to travel might create a scale-free
small world of human contacts. They used data collected by automated
systems for tracking bank notes by their serial number to look at how
nearly half a million dollar notes ‘travelled’ around the contiguous
United States, and found that the relationship between the distance
travelled and the time taken for these notes followed a power law, just
as it would if the notes were being passed around on a scale-free
network. If most of these notes move from place to place in people’s
wallets, they give a fair indication of how humans (or at least,
Americans) get about. All the same, this is not by any means a direct
indication that humans encounter and infect one another in a scale-

free manner.
Some encounters are more intimate than a brushing of elbows on the
train, and may give you rather more than a cold. AIDS is now the third
biggest cause of premature death in the world, and kills two million
people a year in sub-Saharan Africa alone. Networks of sexual contacts
are the fatal grid on which HIV spread, and it is hard to develop effective
strategies for attacking the epidemic unless we know something about
the patterns it traces out. If this, too, is a scale-free net, it would seem
that prospects for eradicating the virus entirely are dim. We simply do
not know, however, if this is the case or not. A study of 3,000 Swedes in
2001 suggested that the distribution in the number of sexual partners
over a twelve-month period followed a power-law; but the result re-
mains controversial.
Communication breakdown
The implication of these findings is that flows—of information, rumour,

and disease, say—on scale-free networks are not easily disrupted. For
human health, that sounds like a disheartening message. But for tech-
nological networks like the Internet, this aspect of scale-free topologies
would appear to be a virtue. It implies that, if a few links get broken, the
network is in no danger of falling apart. The multitude of pathways
privileged with small-world shortcuts means that you can always find
a good alternative for routing your emails. That idea has been confirmed
by Albert, Jeong, and Barabási, who have compared the way scale-free
and other networks (such as random graphs) fragment as more and
more nodes are ‘killed’ (all links to and from them being severed) at
random. In random graphs these ‘connection failures’ soon make the
network break apart into isolated clusters, so that it becomes impossible
to reach any point in the web from any other (Fig. 6.6a). But scale-free
networks break up differently: there remains a large cluster of intercon-
nected nodes even for rather severe amounts of link failure. This central
cluster merely sheds little ‘islands’ of nodes as the damage gets worse
(Fig. 6.6b). The network doesn’t shatter, but gently deflates.
Since breakdowns do happen—servers get jammed or malfunction—
this is surely just the kind of property one would like in a network like
the Internet. But no one designed it that way. Indeed, if they had
WEB WORLDS j 177
Fig. 6.6: How webs fall apart. As increasing numbers of nodes are
deactivated, so that the links to them are effectively severed, at
random, a network breaks up into isolated units. But this happens
much more quickly for random graphs (a) than for scale-free networks
(b): the latter tend to ‘deflate’, shedding small islands but retaining a
large, interconnected core.
designed it, they probably would have chosen some other network
topology that was nothing like as resilient. The Internet acquired this
happy feature simply from the way that it grew.
Yet scale-free robustness has a cost—one might call it an Achilles’
heel. The resilience of the network depends on the presence of a few
highly connected hubs: the richest of the rich, which offer shortcuts
between many different regions. Now, if one were to deactivate nodes
not at random but in a targeted fashion, taking out only the most highly
connected hubs, the story is very different. While the resilience of the
Internet to random breakdowns is extraordinary—it has been esti-
mated that a connected cluster of nodes that reaches right across the
network remains even for almost 100 per cent breakdown of links—it
becomes highly vulnerable to an attack that knocks out the most highly
connected nodes first. With a few well-placed shots, you could scupper
the entire network. Albert, Jeong, and Barabási have calculated that if
such a strategy deactivates just 18 per cent of the nodes, the Internet
would be shattered into many tiny pieces. That is now one of the
concerns of organizations worldwide that have been established to
combat the threat of cyberwarfare: intentional disruption of computer
networks. With ever more aspects of our lives being dependent on these
information pipelines, from health services to power supplies, this is
emerging as a serious concern.
There is a positive side to this vulnerability of scale-free networks.
Since it seems they exist in the biochemical pathways of living cells,
describing the interactions of protein enzymes for example, then tar-
geting drugs at the ‘hub sites’ of pathogenic organisms or rogue cells
might be a good way of killing them off. And while diseases may spread
faster and become harder to eradicate in scale-free social networks,
immunization and vaccination programmes aimed at the key hubs—
the most highly connected individuals, for example those who are most
sexually active—may have an inordinately positive impact. Barabási
and his colleague Zoltán Dezsö have shown that treating the hubs
against a viral infection in fact restores the threshold for the virus to
spread as an epidemic, making it possible to eradicate it entirely. Of
course, in the real world it may be far from easy to identify who the
hubs are, or to reach and treat them. But an immunization strategy that
does even a rather crude job of finding and treating highly connected
nodes preferentially can reintroduce an epidemic threshold to a scale-
free network, making it easier to contain the virus so that it may die out
naturally.
What, finally, might these studies of network structure have to tell us
about power failures such as the one that turned out the lights of
Manhattan? It is not clear whether any power grids are scale-free
networks—possibly some are and some are not. But in any event,
they do in general appear to be small worlds, with many shortcuts
and small average path lengths between nodes. And this seems to
confer the same kind of mixture of robustness and vulnerability.
Random failures usually do not matter, because alternative routes can
be found for the electricity. But such networks do seem prone to a
particular kind of catastrophic breakdown in which a few local failures
create cascades: overloads get passed on down the line quickly, escal-
ating as they go. This seems to be what happened in the 2003 US
blackout, and very probably in the even larger one a month later that
affected the whole of Italy. A local failure means that the electrical load
gets passed to another part of the grid, which in turn becomes over-
loaded and shuts down, and so the problem gets shunted further down
the line, leaving failed links in its wake. It is possible to design networks
that do not have this tendency for cascading breakdown—they are
WEB WORLDS j 179
generally not small worlds, but require rather long average path lengths
between nodes—but for technological networks that grow without any
central planning, such as power grids and computer networks, design is
not really an option.
Yet cascade failures may not be inevitable for small-world networks,
so long as they are understood within the context of the network
topology in which they occur. Once the pattern of the network is
taken into account, it might be possible to tailor an appropriate re-
sponse strategy. Dirk Helbing at the Dresden University of Technology
and his co-workers have proposed that in such cases the best strategy is
to reinforce the most highly connected nodes first against failure. That
makes intuitive sense and is what you might have guessed anyway—
but only once you appreciate the way the network is configured in the
first place. You first have to see the pattern you are dealing with.
Epilogue
The Threads of the

Tapestry
Principles of Pattern
Nature uses only the longest threads to weave her patterns, so each small piece
of her fabric reveals the organization of the entire tapestry.
Richard Feynman, The Character of Physical Law
Nature is an endless combination and repetition of a very few laws. She hums the
old well-known air through innumerable variations.
Ralph Waldo Emerson, ‘History’, Essays
It is time to take stock. I hope it has become apparent in the course of

these books that there is going to be no Grand Theory of Pattern await-
ing you at the rainbow’s end. The universe is not made that way. Some
physicists have in recent years encouraged an unfortunate aspiration
towards grand unified pictures, but the world that we encounter, the
world of real stuff that we see and touch, is far too messy for such things.
But this does not mean we must capitulate to a welter of detail. I have
shown that there are fundamental patterning processes that recur in
nature, operating identically in many different settings. Laplacian
growth instabilities lead to snowflakes, or soot, or cracks in pavements
and continents. Convection orders clouds, stones, and pans of hot milk.
Reaction–diffusion processes may generate the leopard’s spots and the
graveyards of ants. What nature uses is not a Law of Pattern but a
palette of principles. And there is, I submit, much more wonder in a
world that weaves its own tapestry using countless elegant and subtle
variations, combinations and modifications of a handful of common
processes, than one in which the details become irrelevant and in
which a few recondite equations are supposed to explain everything.
To my mind, the astonishing thing about many natural patterns is not
just that we can implicate a few basic processes that lie at the heart of
them all, but that small changes in the details, in the specific initial or
THE THREADS OF THE TAPESTRY j 181
boundary conditions, produce such fantastic variety—think, for ex-

ample, of butterfly wings. By the same token, the patterns of a river
network and of a retinal nerve are both the same and utterly different.
It is not enough to call them both fractal, or even to calculate a fractal
dimension. To explain a river network fully, we must take into account
the complicated realities of sediment transport, of changing meteoro-
logical conditions, of the specific vagaries of the underlying bedrock
geology—things that have nothing to do with nerve cells.
The late Rolf Landauer, an uncommonly perceptive and broad-
thinking physicist, has expressed very pithily this need to resist over-
enthusiastic aspirations to the universal: ‘A complex system is exactly
that; there are many things going on simultaneously. If you search
carefully, you can find your favorite toy: fractals, chaos, self-organized
criticality, Lotka–Volterra predator–prey oscillations, etc., in some cor-
ner, in a relatively well developed and isolated way. But do not expect
any single simple insight to explain it all.’
Perhaps it is a shame to begin a summary chapter with a caution
against too much summarizing, but that is for the best. For the ideas
that form the backbone of our understanding of spontaneous pattern
formation seem so powerful, so all-encompassing, that they are all too
often paraded as the keys to a theory of everything. Even D’Arcy
Thompson would not have wanted us to believe that.
And yet what is extraordinary and thrilling is that so many pattern-
forming systems have so much in common, to the extent that by under-
standing one we can predict a great deal about the others. This realization
has made a delicious mockery of the traditional, rigid divisions between
scientific disciplines, so that physicist, economist, ecologist, chemical en-
gineer, and geologist can all talk to one another—and in the same language.
When this happens, something very exciting is going on in science.
Yet we have seen that many of the ideas behind pattern formation are
not new. Oscillating chemical reactions were known in 1901; convec-
tion cells showed up around 1900; and Kepler sensed the reason for the
sixfold symmetry of a snowflake in the seventeenth century. But D’Arcy
Thompson was unable to persuade most of his peers of the importance
of form and pattern in the 1920s, and only in the past two decades or so
has pattern formation emerged as anything like an identifiable field of
study in its own right. Why is that?
One reason is the explosion in computer power. Many of the theoretical
ideas about patterning are tough to test experimentally, since there are so
many factors to bring simultaneously under control; but as we have noted

again and again, computers allow researchers to perform ‘ideal’ experi-
ments in which everything can be repeated exactly and complicating
factors included or excluded at will. Many theoretical models are simple
in principle but utterly impossible to test by manual number-crunching—
the calculations would take for ever if done by hand. But although this
increase in computational capacity has provided scientists with perhaps
the most important technological tool currently at their disposal, it also
serves to underline the phenomenal achievements of early researchers
such as Michael Faraday, Lord Kelvin, and Lord Rayleigh, Geoffrey Taylor,
and Andrei Kolmogorov, who had to rely on their exquisite intuition alone
to deduce the essential physics of pattern-forming processes.
I believe there is another reason, little emphasized but equally im-
portant, for the recent development of ideas in pattern formation. This
is the maturation since the mid-1980s of a field of theoretical physics
that provides much of the framework for understanding the features
that frequently characterize spontaneous patterning, such as abrupt,
global changes of state and scaling laws. The field is the study of phase
transitions and critical phenomena, and it is the largely unseen bedrock
of all physics today. I shall say more about this discipline in what
follows.
Let me now try to pull together some of the threads that have run,
more or less perceptibly, through all the previous chapters. They do not,
even collectively, constitute a ‘theory of patterns’. Rather, these ideas
are like stepping stones that lead us through the turbulent ebb and flow
of pattern and form in the physical and natural world.
Competing forces
Spontaneous patterns typically represent a compromise between

forces that impose conflicting demands. The ordered architectures
adopted by certain polymers and soap-like molecules (Fig. 7.1: see
Book I, Chapter 2) are an elegant solution to the requirement of keeping
the structure’s surface area and curvature small while also packing
molecules together efficiently. The expanding waves and the static
spot and stripe patterns of some chemical mixtures (Fig. 7.2: see
Book I, Chapters 3, 4) result from a delicate balance between reaction
and diffusion of the molecular components, and between short-ranged
Fig. 7.1: Ordered structures in co-polymers: a balance between

surface area and curvature of the interfaces of domains and the
efficient packing of molecules. (Photo: Edwin Thomas, Massachusetts
amplification and long-ranged inhibition of their chemical reactions.

The bulbous pseudopodia of viscous fingering (this book, Chapter 2)
are the manifestation of competition between an instability that gen-
erates branches at the interface, and surface tension, which limits its
size scale. Snowflakes are the product of a similar compromise forged
in an environment that is imprinted with an intrinsic microscopic
symmetry. Trains of breaking-wave vortices appear in fluid flows
when an instability that excites waves wins out over the tendency of
viscosity to suppress them (Book II, Chapter 3)—a victory from which
a preferred patterning size emerges.
Competition lies at the heart of the beauty and complexity of natural
pattern formation. If the competition is too one-sided, all form disap-
pears, and one gets either unstructured, shifting randomness, or fea-
tureless homogeneity—bland in either event. Patterns live on the edge,
in a fertile borderland between these extremes where small changes
can have large effects. This is, I suppose, what we are to infer from the
clichéd phrase ‘the edge of chaos’. Pattern appears when competing
forces banish uniformity but cannot quite induce chaos. It sounds like a
dangerous place to be, but it is where we have always lived.
Fig. 7.2: Spirals, stripes, and spots in oscillating chemical reactions. (Photos: a Stefan Müller,
University of Magdeburg; b, Harry Swinney, University of Texas at Austin.)
Symmetry breaking
When spontaneous patterns appear in systems that are initially uni-

form, this lowers or breaks the original symmetry. That is why it is
important not to confuse symmetry and pattern: high symmetry does
not by any means imply the richest patterns, and indeed often those
that look most striking have rather low symmetry, or none at all. On the
other hand, symmetry tends to break a little at a time, and we may find
that the least symmetrical patterns are those that appear when the
driving force responsible for them is greatest, so that the system is
furthest away from an equilibrium state. I shall say something shortly
about why symmetry breaks away from equilibrium. At this point, we
need appreciate only that symmetry breaking is not like laying a tiled
floor. A hexagonal cellular arrangement like that of Rayleigh–Bénard
convection (Fig. 7.3: Book II, Chapter 3) is not a matter of imposing
cells of an arbitrary shape, one by one, in an inert medium. Rather, the
medium becomes everywhere at once imbued with a ‘hexagon-forming
tendency’, once the driving force (here the heating rate) exceeds the
Fig. 7.3: A hexagonal array of convection cells in Rayleigh–Bénard

convection. (Image: David Cannell, University of California at Santa
Barbara.)
pattern-forming threshold. It then takes only the smallest fluctuation to

release this ‘hexagon-ness’ globally. The hexagonal array seems to rise
out of the floor, you might say.
Non-equilibrium
Nearly all the pattern-forming systems in these books are out of equi-
librium—that is to say, they are not in their thermodynamically most
favourable state. Once scientists considered such systems to be un-
approachable, perhaps even unseemly. Thermodynamics, the science
of change that developed initially as an engineering discipline in the
nineteenth century, was intended to describe the equilibrium state of
systems. It told one about the direction of change, and allowed one to
calculate the amount of useful work that could be extracted from that
change; but what actually took place during a change was something
that classical thermodynamics could barely touch. It was a pretty good
tool for chemical and mechanical engineers who wanted to gauge the
performance of their machines. But it offered a rather artificial view of
the world in which everything happens in a series of jumps between
stable states that do not otherwise alter over time. That is not very like
the world we know. Thermodynamics was silent in the face of the
uncomfortable fact that some processes never seem to reach equilib-
rium. A river does not simply empty itself into the sea in one glorious,
ephemeral rush—the water is cycled back into the sky and redeposited
in the highlands for another journey. And so will it always be, while the
sun still shines.
Out of this somewhat restrictive picture, however, emerged the idea
of an arrow of time. Nearly all processes seem to have a preferred
direction: they go one way but not the reverse. Heat flows from hot to
cold, an ink droplet disperses in water. These processes are said to be
irreversible. One-way processes fit with our intuition—ink droplets do
not re-form—but they become somewhat puzzling when we look
closely at the microscopic events behind them. In the mathematical
equations that describe how a particle of ink pigment moves in water,
there is no arrow of time: you could play a film of the particle’s motion
backwards and not notice the difference, nor appear to break any
physical laws. It is only when you look at the behaviour of the
whole ensemble of particles that you would notice anything odd
when time is reversed and the droplet coalesces from a glass of

uniformly dilute ink.
Most scientists agree that irreversibility has its roots the second law
of thermodynamics, which states that in a system isolated from its
surroundings (so that it cannot exchange energy or matter), the direc-
tion of change is always towards greater entropy, which loosely means,
towards greater disorder. This is a law enforced not by physics but by
probability. It emerges when a system has many choices of how to
arrange its components: there are simply many more disordered ar-
rangements than ordered ones. We encountered the second law in Book
I, where I explained that it prompted objections to the idea of oscillat-
ing chemical reaction in the 1950s. As entropy is in some sense a
measure of disorder, the second law seems to pose a big problem for
the spontaneous appearance of pattern.
Themodynamics allows one to predict what a system at equilibrium
will look like: it tells us that such systems adopt states for which the
total energy is as small as it can be.* This is the energy minimization
criterion that we saw in operation in Book I, controlling the shapes of
soap films. Is there an analogous criterion that determines what the
state of a non-equilibrium system is?
The thermodynamics of non-equilibrium systems is concerned not
with some end point in which entropy has increased in relation to the
initial state; rather, it considers the process of becoming, of how change
occurs. Since the second law specifies that an irreversible change leaves
a system with more entropy than it began with, change produces en-
tropy. When the system reaches its new equilibrium, entropy produc-
tion ceases. So the process of change seems to be bound up with the
issue of entropy production.
In the 1930s the Norwegian-born scientist Lars Onsager began to
explore that connection. He considered change driven by only a small
departure from equilibrium. Under these conditions Onsager deduced
universal laws relating the driving forces to the rates of entropy pro-
duction. For this work he was awarded the Nobel Prize for chemistry in
1963. The Russian-born chemist Ilya Prigogine, working in Brussels,
added an important element to the picture by showing that, in this
near-equilibrium case, non-equilibrium systems tend to behave in a
*You may recall from Book I that what is really being minimized here is a particular quantity
called the free energy, or the Gibbs energy.
way that minimizes the rate of entropy production. If prevented from

reaching equilibrium (where the rate of entropy production is zero), the
system will instead settle into a steady but dynamic state in which
entropy is produced at the lowest possible rate. Here, then, was a
criterion for determining the preferred state of a system out of (but
close to) equilibrium.
Nothing in this prescription, however, gives any hint that away from
equilibrium a system may stumble into a patterned state with orderly
structure, like Bénard’s convection cells or Turing’s spots. States like
this tend to emerge rather far from equilibrium, where the equations
formulated by Onsager and others no longer apply and Prigogine’s
minimum entropy production rule can break down. Where do they
come from?
During the 1950s and 1960s, Prigogine and his colleague Paul
Glansdorff attempted to extend the treatment of non-equilibrium
thermodynamics to the more interesting far-from-equilibrium situ-
ation. They were able to show that, as the force driving a system away
from equilibrium increases, the steady state of minimal entropy pro-
duction reaches some crisis point where it breaks down and becomes
transformed to another state. Technically speaking, there is a bifurca-
tion—literally, a branching in two—at which the evolution of the steady
state splits into two branches, presenting a choice of new states that the
system can adopt (Fig. 7.4).
Fig. 7.4: A bifurcation

occurs when a stable
state develops an
instability that offers the
system a choice of two
new states. A ‘pitchfork’
bifurcation like that
shown here commonly
occurs as a system is
driven further out of
equilibrium.
What are these new states? The theory of Prigogine and Glansdorff
said little about that; but it seemed reasonable to suppose that they
might correspond to the self-organized structures and patterns that
were known to appear far from equilibrium. To progress any further,
however, we first need to appreciate how these regular or ordered non-
equilibrium states are fundamentally different from superficially simi-
lar ordered states in equilibrium systems.
Dissipative structures
Regularity is not rare. A swinging pendulum, a bouncing ball, the

‘greengrocer’s stall’ atomic packing of a crystal, the yearly passage of
the Earth around the Sun: all are periodic in space or time. But the
regular hexagonal lattice of Rayleigh–Bénard convection differs from
the hexagonal atomic lattice of a crystal like copper metal. The latter is
an equilibrium structure whose periodicity is determined by some
characteristic dimension of the components—here, the sizes of the
atoms. The former, meanwhile, is maintained away from equilibrium
by a throughflow of energy, which it dissipates in the process, thereby
generating entropy. Stop the input of energy—let the top-to-bottom
temperature gradient equalize—and the pattern goes away. Likewise,
the oscillations of the Belousov–Zhabotinsky (BZ) reaction (Book I,
Chapter 3) persist only while the reaction is fed with fresh reagents
and the products are removed. Patterns that are supported away from
equilibrium by the generation of entropy are called dissipative
structures.
In contrast to most equilibrium structures, the spatial scale of the
pattern features in a dissipative structure bears no relation to the size
of its constituents. The size of convection cells is much, much larger
than the size of the circulating molecules, for example. Furthermore,
this size scale persists in the face of perturbations. A convecting fluid
may lose its pattern temporarily if stirred, but once the disturbance
has passed, the same structure reforms. The system ‘remembers’ how
it is supposed to look. These robust dissipative structures are said to
be governed by an attractor: a place to which they are ‘anchored’ in
the abstract space of parameters describing the system. An example
of such an attractor is the limit cycle of the oscillating BZ reaction.
The opposite of a dissipative structure is a conservative structure,
which possesses no attractors and so can be altered arbitrarily. An

example is the orbit of a planet around the Sun: if the radius of the
orbit is altered (say, by a catastrophic collision with another body), it
stays that way rather than returning to its former value.
If a dissipative structure is ‘kicked’ too hard, however, it may find
itself closer to a different attractor—a quite different state or pattern—
and end up being drawn towards that instead. Attractors are like dips in
a hilly landscape: if you get pushed over a crest, you fall down into a
different basin. The point here, then, is that non-equilibrium systems
tend to change not gradually but in discrete jumps between different
dissipative attractor states. Let’s look at this in more detail.
Instabilities, thresholds, and bifurcations
Most of the patterns that I have described appear suddenly. One mo-
ment there is nothing; then you turn the dial of the driving force up a
notch, and everything is abruptly different. Stripes appear, or dunes, or
pulsations. This seems to be the nature of most symmetry-breaking
processes: they happen all at once. In that respect, they resemble phase
transitions in equilibrium thermodynamics.
Phase transitions are generally abrupt jumps from one equilibrium
state of matter to another: from ice to water, water to vapour, magnet to
non-magnet. These are ‘all-over’ transformations. When water cools
through its freezing point, we don’t find part of it turning to ice and
the rest remaining liquid.* Below zero degrees centigrade all the water
is ready to become ice, and will surely do so given enough time. And
this is an all-or-nothing affair: the temperature need be only a fraction
above freezing point for all the ice to have melted once equilibrium is
reached. A fraction below, and all is frozen.
In other words, there is a threshold that, once crossed, leaves the
entire system prone to a change in state. Just the same is true for many
pattern-forming processes. Convective patterns, as we observed above,
appear above a threshold heating rate, and vortices in fluid flow above a
*In practice we see that combination quite a lot—a layer of ice on a pond, say. But this is
because the water may not all be below freezing point, or because it takes time for the water
to freeze or the ice to melt. In those situations, the pond isn’t in thermodynamic equilib-
rium.
threshold flow rate. The path of a crack goes crazy above a particular
crack speed.
In addition, the change in state during an equilibrium phase transi-
tion may involve a breaking of symmetry. Crystalline ice has an ordered
molecular structure (in fact it has many ordered structures), while
liquid water is disorderly at the molecular scale. Again, you could be
forgiven for thinking that symmetry is therefore broken during melting,
but in fact it is the other way around: symmetry is broken during
freezing, because whereas the liquid state is isotropic (all directions in
space are equivalent) the crystal structure of ice identifies certain
directions as ‘special’.
Thus equilibrium phase transitions, like the abrupt transitions that
characterize much of pattern formation, are spontaneous, global, and
often symmetry-breaking changes of state that happen when a thresh-
old is crossed.
Some of these transitions involve a straightforward rearrangement of
one state into another. But there are classes of both equilibrium and
non-equilibrium transitions that offer a choice of two alternatives for
the new state, which are equivalent but not identical. Think of the
formation of convection roll cells. Adjacent rolls turn over in opposite
directions, but any particular roll could rotate either one way or the
other as long as all the others switch direction too. Above the convec-
tion threshold, there is a choice of two mirror-image states. Which is
selected? Clearly, there is nothing to favour one over the other, and the
issue is decided by pure chance. The same is true of the rotation of
plughole whirlpools, unless some small outside influence tips the
balance.
The equilibrium behaviour of a magnetic material like iron displays a
comparable choice. In iron’s magnetized state, all the atoms act like
little bar magnets with their north and south poles aligned. If you heat a
magnetized piece of iron above 770 degrees centigrade (its so-called
Curie point) this alignment is lost, because it is overwhelmed by the
random, jiggling effect of heat. The magnetic fields from each atom
then cancel out on average, and the piece of iron as a whole is no longer
magnetic. This abrupt change at the Curie point from a magnet to a
non-magnet is an example of a phase transition. It might seem that this
phase transition involves a single choice: either the iron is magnetic or
not. But in fact there are two possibilities as the metal is cooled from a
non-magnetized, randomly oriented state through the Curie point: the
Fig. 7.5: A magnetic

material such as iron
undergoes a bifurcation
when cooled through its
so-called Curie point (at
temperature Tc), where
the material becomes
spontaneously magnetic
(a). This happens
because the atomic
magnetic poles,
represented here as
arrows, all line up in the
same direction; above Tc
that is prevented by the
randomizing influence of
heat. Below Tc the
magnetization can point
in one of two equivalent
directions. This kind of
bifurcation is analogous
to the situation of a ball
perched on top of a hill
between two identical
valleys (b); it must roll
down one way or
another, but the choice
is arbitrary.
atomic magnetic poles can all point either in one direction or the other
(Fig. 7.5a). The states are entirely equivalent,* and again the choice
depends on random fluctuations that tip the balance. (You may wonder
how, or if, this random choice can be made the same way throughout
*In reality the Earth’s magnetic field could supply a bias in favour of one orientation. Indeed,
this is how changes in the state of the geomagnetic field are deduced from the imprint it
leaves on magnetic rocks that cooled from a molten state many millennia ago.
the entire system. I’ll come back to this). The situation is like a ball
perched on top of a perfectly symmetrical hill (Fig. 7.5b): it is unstable
at the top and has to roll down one side or the other, but which way it
goes is unpredictable and at the mercy of imperceptible disturbances.
Freezing and melting of water might seem rather similar to this
magnetic transition, in the sense that they too involve atomic-scale
order being overwhelmed by, or recovered from, thermal randomness.
But there is an important difference, somewhat technical, but import-
ant. Freezing and melting are said to be first-order phase transitions.
Among the distinguishing features of these are the fact that the switch
in state begins at one or more randomly selected points and spreads
from there throughout the whole system. Freezing starts from a little
‘seed’ or nucleus of ice somewhere in the water. And there is a step-like
change in key properties of the system: in this case, in the density
(water is denser than ice). And it is possible for the less stable state to
persist beyond the transition threshold in a precarious state that is said
to be metastable, and which is liable to switch at any time. Water can be
supercooled below freezing point without turning to ice, if it is free from
small particles on which ice crystals might nucleate. This means that in
practice the transition might happen at a different point from where
equilibrium thermodynamics says it should and, specifically, the
threshold might be different as the system passes through the transi-
tion in one direction (freezing) to the other (melting). This is called
hysteresis. Finally, first-order phase transitions may but do not have to
involve symmetry-breaking.
The spontaneous magnetization of iron at the Curie point, on the
other hand, is an example of a second-order or critical phase transition.
The magnetization changes abruptly but continuously as the system
goes through the transition—there is no sudden jump from one value
to another (see Fig. 7.5a). Second-order and other critical phase tran-
sitions always involve symmetry breaking. And there can be no hyster-
esis: the switch to a new state cannot be delayed. It doesn’t depend
on the formation and growth of some nucleus of the new state, but
happens through a kind of global convulsion, driven by the random
fluctuations of the components.
Now, many of the pattern-forming bifurcations that I have discussed
are analogous to critical phase transitions—they are called supercritical
bifurcations, and they lead to symmetry breaking. The onset of con-
vection is like this, as is the switch between hexagonal and striped
Fig. 7.6: Convection patterns in a fluid may switch from targets to

spirals in an abrupt transition. But because this transition is of a
technical type called sub-critical, targets and spirals can coexist.
(Photo: Michel Assenheimer, Weizmann Institute of Science, Rehovot.)
chemical Turing patterns (Fig. 7.2). But a few pattern-forming pro-

cesses are subcritical bifurcations, analogous to first-order phase tran-
sitions, such as the switch between spiral and target patterns in
convecting fluids (Fig. 7.6: see Book II, Chapter 3). This is why both
spirals and targets can coexist in the same pattern. These distinctions
may seem rather esoteric; but I’ll show shortly that the details of what
happens at a critical phase transition provide important clues to how
patterns arise away from equilibrium.
The solutions reveal more than the equations
These analogies with phase transitions are useful, but they don’t pro-
vide any kind of rigorous mathematical description of pattern-forming
bifurcations. Physicists like analogies, but they prefer rigour. Attempts
to develop a more concrete description of what happens at a sym-
metry-breaking bifurcation in non-equilibrium systems began in earn-
est in 1916 when Lord Rayeigh looked for a theory that would explain
Henri Bénard’s convection patterns. In the 1920s, Geoffrey Taylor
attempted much the same for the case of Taylor–Couette flow between
rotating cylinders (Book II, Chapter 6). As I explained in that volume,
a thorough treatment of any problem in fluid flow must start with the
Navier–Stokes equation—Newton’s law of motion applied to fluids,
which relates the changes in fluid velocity at every point to the forces
that act on the fluid. Both Rayleigh and Taylor looked for the solution to
the Navier–Stokes equation appropriate to their respective systems

close to equilibrium—that is, when the driving force for patterning
was very small. They then examined how this ‘base state’ would re-
spond as the system was driven increasingly far from equilibrium.
For the case of convection, the base state is one in which no flow at all
occurs: if the temperature difference between the top and bottom of the
fluid is small enough, heat flows simply by conduction through the
fluid, just as it does when it spreads through a solid. The question is
then, ‘does the base state remains stable if it is disturbed a little?’
Rayleigh considered an all but imperceptible wavy disturbance in the
fluid, with a particular wavelength. Below the critical Rayleigh number
Rac (a measure of the bottom-to-top temperature difference) at which
convection begins, such perturbations die away over time regardless of
the wavelength, and the system returns to the base state. But exactly at
Rac, something stirs: one particular wavelength neither decays nor
grows, although all others decay. And just above Rac the ‘special’ wavy
perturbation grows: a pattern with this wavelength develops. As Ra
increases, the range of possible wavelengths of the convection pattern
broadens steadily. Above Rac the base state is still a solution to the
Navier–Stokes equation for the system—but it is an unstable solution,
since the slightest disturbance will trigger the appearance of a pattern
at one of the allowed wavelengths.
The key point about this approach, which is called linear stability
analysis, is that it can be applied to a wide range of systems that
undergo spontaneous patterning—not just those in fluids that obey
the Navier–Stokes equation. A similar kind of analysis can be carried
out, for example, for the equations that describe reaction–diffusion
systems. Some researchers have attempted to construct simplified
‘model equations’ that describe the generic features of pattern-forming
systems such as these while avoiding the complexity (and thus the
intractability) of, say, the full Navier–Stokes equation. They are gen-
eral-purpose equations for predicting patterns arising from an instabil-
ity in an unpatterned base state, regardless of the detailed nature of the
particular system in question.
Perhaps the most important message to have emerged from these
studies of instabilities is that we do not necessarily have a complete
understanding of a system once we know the equations that govern
it. What we really want to know are the particular solutions to those
equations. The latter need not be obvious from the former. That is a
point worth remembering in all of mathematical science. The English

physicist Freeman Dyson says that ‘to discover the right equations was
all that mattered’ to Albert Einstein and Robert Oppenheimer in their
later years. One might say the same about some physicists working
today to develop a ‘theory of everything’. But if you take this view, then
fluid dynamics was solved once we could write down the Navier–Stokes
equation. Yet if we had stopped there, we would never have guessed at
the rich variety of solutions that it held in store even for relatively
simple experimental situations. In fact, sometimes even knowing the
mathematical solutions is not enough: we need to do experiments to
interpret what they are telling us.
Pattern selection
Linear stability analysis can reveal the point at which a non-equilib-

rium system is driven across the threshold of pattern formation. But
can it tell us anything about the pattern that results? Exactly at the
threshold, there is (at least for the cases considered by Rayleigh and
Taylor) a single ‘marginally unstable’ wavelength that defines the char-
acteristic pattern scale. But how does this length scale manifest itself—
as stripes, spots, or perhaps travelling waves? And once the threshold is
surpassed, an increasing range of patterning wavelengths may become
stable. Which wavelength is selected?
This question is not peculiar to these examples in fluid dynamics.
Just about any system with pattern-forming potential faces choices: it
may have a gallery of designs from which to select. Soap molecules on
the water surface (Book I, Chapter 2), metal deposits grown at elec-
trodes (this book, Chapter 2), bacterial colonies (Book I, Chapter 5 and
this book, Chapter 2), jumping sand grains (Book II, Chapter 4)—they
all confront a catalogue of riches. Which to choose?
There is no universal way to answer this question. Yet there is in this
respect a major distinction between equilibrium and non-equilibrium
systems. The former may also have several choices of pattern and form,
but there is in that case a simple rule (in principle) for deciding the best
choice: as we have seen, at equilibrium a system will always seek to
adopt the configuration that has the lowest free energy. This means that
balls roll down hills, iron rusts in air, water below freezing point turns to
ice. A few of the complex patterns I have discussed do form under
equilibrium conditions: for example, the shapes of soap bubbles and the
self-organization of hybrid polymers (Book I, Chapter 2). If we know the
various contributing factors to the free energy, we can predict the
selected pattern in these cases by finding the shape that minimizes it.
What about non-equilibrium systems? Might there be some analogous
‘minimization principle’ for them? I shall return to this question shortly.
We can first make a few general observations about how patterns form
away from equilibrium. This normally involves symmetry breaking; and
symmetry tends to break in stages, a little at a time, as the system is driven
harder and harder. This alone enables us to understand why two types of
pattern, stripes and hexagons, are particularly common. The simplest
way to break the symmetry of a uniform, ‘flat’ (two-dimensional) system
such as a shallow layer of fluid—that is, the way to break as little sym-
metry as possible—is to impose a periodic, wavy variation in just one
direction (Fig. 7.7a), which produces parallel bands, stripes, or rolls.
Parallel to the stripes, symmetry is not broken: as we travel through the
medium in this direction, we see no change in its character. It is only in
Fig. 7.7: Breaking symmetry by degrees. The simplest way to break

the symmetry of a uniform two-dimensional system is to impose
periodic variations in one direction, creating stripes (a). When
symmetry is broken in both dimensions, square or triangular cells result
(b). In the latter case, the resulting pattern can be a triangular or a
hexagonal lattice of cells.
the perpendicular direction that we can identify the broken symmetry:

travelling in this direction, we see a periodic change from one state to
another and back again. Thus, stripe-like patterns are often the first to
appear from a uniform, flat system. That is what we see for sand ripples
and in the appearance of convection and Taylor–Couette roll cells.
After breaking symmetry periodically in one dimension, the next ‘min-
imal’ pattern in a two-dimensional system involves breaking it in the
other, dividing up the system into compartments or grids. If the state is to
remain ordered and as symmetric as possible, there are only two options:
to impose the periodic variation perpendicular to the rolls, creating
square cells, or to impose two such variations at 608 angles, creating
triangles or hexagons (Fig. 7.7b). So the square, triangular and hexagonal
patterns that we have seen in Turing patterns (Book I, Chapter 4), in
convection (Book II, Chapter 3) and in shaken sand (Book II, Chapter 4)
are no mystery. They arise simply because the geometric properties of
space constrain the ways in which symmetry can be broken.
These hand-waving arguments are not, however, a reliable guide for
determining exactly what kind of broken symmetry will arise in any
particular case. For that, one needs to look into the messy specifics. For
example, the question of whether stripes/rolls or hexagons will be
preferred has no universal answer. And a linear stability analysis of
convection will not help you to answer that: you have to use a more
sophisticated level of theory to discover that rolls are generally
favoured. Qualitatively we can understand this preference on the
grounds that rolls do not distinguish between upflow and downflow
(what goes up on one side of the cell is mirrored by what goes down on
the other), whereas hexagonal cells do: upflow occurs in their centre
and downflow at the edges. So rolls are geometrically symmetrical
about the plane midway between the top and bottom of the fluid, but
hexagonal cells are not. If this midplane symmetry is broken, how-
ever—for example, if the warmer fluid near the bottom is significantly
less viscous than the cooler fluid towards the top (which is quite
possible)—then hexagonal cells might appear instead. In the case of
the Turing instability of reaction–diffusion systems, on the other hand,
hexagonal spots are usually the default option.
The other aspect of pattern selection for these relatively simple cases
is the size of the features: the wavelength of the stripes or the separation
of the hexagonal spots. Again, it is not possible to identify a single
criterion that determines this. For Rayleigh–Bénard convection, we
Fig. 7.8: Bifurcations in many non-equilibrium systems come in a sequential cascade as the system
is driven further from equilibrium (here from left to right). At each bifurcation, the number of states
of the system doubles. In an oscillating system such as the Belousov–Zhabotinsky reaction, each
bifurcation corresponds to a period-doubling: it takes two, then four, then eight cycles for the
system to return to a given state. This cascade structure gets increasingly finely branched and
eventually gives way to non-periodic, chaotic behaviour, seen here as a dense ‘dust’ of dots.
saw that linear stability analysis can be used to calculate the wave-
length of the instability at the onset of patterning. But above that
threshold there is a range of allowed wavelengths, and then the width
of the roll cells becomes dependent on the history of the system: how it
reached the convecting state. This size may in fact vary throughout the
system, or may change over time. In a chemical Turing system, mean-
while, the scale of the pattern is set by how fast the ingredients diffuse
(recall that one component acts as an activator to initiate the formation
of a pattern element, and the other as an inhibitor that suppresses other
elements nearby). And we have seen that these reaction–diffusion
systems may generate moving patterns (travelling waves) rather than
stationary ones. That’s what happens if, above the patterning threshold,
a wavy disturbance to the system doesn’t just grow in strength but also
itself oscillates.
As we have noted, abrupt transitions between different patterns com-
monly occur at threshold values of the driving force. There is a common
tendency (particularly clear for fluid flow) for the patterns to become
more ornate—we might say more complex—as the system is driven
harder and harder. This sequence of increasing complexity is also evi-
dent in the oscillating Belousov–Zhabotinsky reaction conducted in a
continuous-flow stirred-tank reactor (Book I, Chapter 3). As the flow
rate of chemicals through the vessel increases, the oscillations undergo
a series of period-doubling bifurcations so that the cycle repeats with
every oscillation, then with every second oscillation, then with every
fourth and so on. This can be depicted as a cascade of bifurcations
(Fig. 7.8). One might liken this very crudely to the excitation of add-
itional harmonics as a trumpeter blows harder. Eventually the oscilla-
tions become chaotic, as though the system becomes overwhelmed
with options. Then the cascade loses its branched structure and breaks
up into a dense forest of spots—and we lose sight of any order at all.
The race for dominance
When patterns are growing, pattern selection may depend on how, or

how rapidly, the different candidates advance. We saw in Chapter 1 that
the characteristic scale of branches on a dendritic crystal is set by the
condition that the growth speed just about balances the tendency for
the tip to split: there is a unique, ‘marginally stable’ pattern where
growth only just outpaces a propensity to split repeatedly. In other
cases, for example non-equilibrium electrodeposition or the growth
of a bacterial colony, it has been proposed that the pattern selected is
simply that which grows fastest and which therefore outruns the others.
But it is not clear that this is a criterion that applies to branching growth
in general. Pattern selection is also influenced by noise—by which
I mean the inevitable randomness in the environment, such as that
supplied by thermal fluctuations. Noise does not necessarily affect all
patterns equally; it may favour some over others.
Defects and boundaries
Very often the patterns that appear far above the initial instability
threshold are less than perfectly symmetrical; they are laced through
with ‘mistakes’, sometimes to such an extent that all appearance of
symmetry is lost. For example, we saw how the roll cells of convection
or the stripes of Turing structures can merge, and how the hexagonal
cells of Rayleigh–Bénard convection can become grossly imperfect
honeycombs. Through an accumulation of such distortions, parallel
stripes can become bent into more or less disordered wavy patterns,
and a hexagonal lattice of spots can disintegrate into a jumble. This
gives us something akin to the stripes of zebras and the spots of the
leopard. In some patterns of this sort, perfect regularity is only ever a
kind of Platonic dream—the Giant’s Causeway merely hints at its rela-
tion to the honeycomb. Notice, however, that even in cases where
disorder overwhelms all semblance of symmetry, we can still identify
order of a kind: the average distance between spots or stripes, or the
average number of sides of a polygonal pattern element, remains more
or less constant.
Defects have their own logic and taxonomy, enabling us to ‘decode
the mess’ by considering how generic defect structures arise from
characteristic deformations of the underlying pattern (Book II,
Chapter 3). In attempting this, scientists may often draw on a rich
existing theory of defect formation developed from studies of crystals
and related materials, such as liquid crystals.
The principles I have adduced so far apply to ‘infinite’ systems, by
which I mean ones for which we ignore the boundaries. But of course
no real pattern-forming system is infinite—they always have edges.* If
the size of the system is vastly greater than that of the pattern’s char-
acteristic length scale, the effects of edges may be negligible except
close to the edges themselves. Commonly, though, this is not the case:
the pattern may be influenced throughout by the size or shape of the
‘container’. We saw in Book I, for example, how either hoops or spots
could be selected from the same pattern-forming mechanism on ani-
mal tails, depending on the size and shape of the embryonic tail when
the pattern is laid down during development. And, more generally, we
saw how the patterns of different animal pelts—a two-tone division of
the whole body, say, or a few large blotches or a multitude of small
spots—can be determined by the relative size of the embryo at the
*That is not strictly true: there is a wealth of interesting work, for example, on the patterns
that form on spheres and other self-enclosing surfaces, such as toruses. Although these do
not experience edge effects, the patterns are still constrained by the overall size of the
surfaces.
patterning stage. On ladybird wings, both the size and the curvature of
the surfaces may affect the patterns.
The shape of a boundary can occasionally change a pattern to some-
thing qualitatively different from how it would look in an ‘infinite’
container. In long, rectangular trays, convection rolls tend to form
stripes, whereas in circular dishes the rolls curl up into concentric
circles (Fig. 7.9a, b). Moreover, the need for a whole number of pattern
features to fit within the container may determine the wavelength, just
as the wavelength and thus the frequency of an organ note is deter-
mined by the length of the pipe. In some systems, the pattern may
change locally to adapt to the presence of a boundary—in Fig. 7.9c, for
example, concentric roll cells give way to short parallel rolls at the
edges, so that the rolls can meet the boundary at right angles (which
is a more stable configuration).
Correlations and critical points
Many of these self-made patterns seem to acquire a miraculous ability

to measure and mark out space. It is one thing to talk mathematically
about a ‘marginally stable’ wavelength, but how can we understand this
in terms of the interactions between the fundamental components of
the system? The scale of sand dunes, for example, is so far removed
Fig. 7.9: Coping with boundaries. In these convection patterns, the shape of the vessels has
imposed particular shapes on the global arrangements of the roll-like cells. (Images: a, from Cross
and Hohenberg, 1993, after LeGal, 1986; b, David Cannell, University of California at Santa
Barbara; c, from Cross and Hohenberg, 1993.)
from the size of the grains themselves, or of the hops they make when
landing on a surface, that it is hard to imagine where the pattern scale
really comes from. How do the grains ‘know’ where to start and stop
piling up into a new dune? Just the same is true for Turing patterns: the
size of the molecules and atoms in the chemical mixture, and the range
of the interactions between them, is minuscule (about a tenth of a
millionth of a millimetre), yet the patterns have length scales big
enough for us to see with our unaided eyes, perhaps several millimetres
or so. How on earth can interactions on these unimaginably tiny scales
give rise to patterns millions of times larger?
The implication seems to be that the components of the system are
able to ‘communicate’ with each other over distances much longer than
those to which they are accustomed at equilibrium. Think of the rolls
that appear in Rayleigh–Bénard convection. Before the onset of con-
vection, the molecules are moving about throughout the quiescent
fluid in a random, disorderly way; each molecule barely takes heed of
what its immediate neighbours are doing, let alone what is happening
a millimetre or so away, many millions of molecules distant. Yet above
the patterning threshold this independence has been lost, and the
molecular motions have (on average) become correlated over these
vast distances. That is to say, if we were to observe the molecular
motions on the descending edge of one of the roll cells, we would
know that statistically identical motions were being executed by mol-
ecules one wavelength away—and two, and three, and so forth
throughout the vessel. This kind of long-ranged correlation, according
to which molecules behave coherently over distances that far outstrip
the sphere of their own influence, is characteristic of many pattern-
forming systems.
How is it possible? Are the molecules able to relay their individual,
tiny influences from neighbour to neighbour over such scales? That is
quite out of the question: in the frenzied environment of a hot liquid, it
is like trying to play Chinese whispers at a rock concert.
The appearance of long-ranged correlations in systems undergoing
abrupt changes in behaviour is not unique to non-equilibrium systems.
It has been long recognized in equilibrium phase transitions, too. The
key to such behaviour, both at equilibrium and away from it, is that the
system loses all sense of scale. Long-ranged correlations may develop
when a system becomes scale-invariant: the correlations exist over
every range.
Fig. 7.10: The scale-invariance of domain sizes at a critical point, such as the Curie point of a
magnet. This is a fractal structure. (Image: Alastair Bruce, University of Edinburgh.)
This is what happens in the critical phase transition of iron at

its magnetic Curie point, which is an example of a critical point.
I explained in Book II (Chapter 4) that, at a critical point, fluctuations
occur on all length scales. I showed that fluids also have critical points,
where the distinction between gas and vapour states vanishes. We saw
that the distribution of gas and liquid regions at this critical point is
scale-invariant: the domains occur at all scales (Fig. 7.10). This picture
could equally well show domains in a piece of iron where the magnet-
ization points in one direction or the other (see Fig. 7.5). As the iron is
cooled down through its Curie point, the magnetic poles of all the
atoms switch from being randomly oriented to being aligned. But
right at this point, there is no telling which of the two orientations
(shown as black or white here) will prevail. The critical system is
infinitely sensitive to fluctuations: the slightest imbalance suffices to
tip it one way or the other.
As a system like this approaches its critical point, each component
feels the influence of more and more of the others. Far from the critical
point, only the behaviour of a magnetic atom’s nearest neighbours
influence its own alignment. But as the critical point is approached,
each atom’s sphere of influence (called its correlation length) extends
wider and wider. And exactly at the critical point, the correlation length
becomes ‘infinite’—which is to say, as big as the entire system. This

does not imply that the magnetic field set up by each atomic magnet
becomes stronger, or reaches out further; rather, the behaviour of the
atoms becomes more collective, so that progressively larger groups will
behave cooperatively.
In this regard, non-equilibrium pattern formation can resemble a
critical phase transition: it relies on long-ranged correlations between
components to allow them to become organized at large scales. But
there are important differences, particularly in terms of the structures
that result. For in our piece of iron, the ordering that results as we pass
through the phase transition has a characteristic length scale that
reflects the scale of interatomic forces: the magnetized iron adopts a
regular structure in which the periodicity of the magnetic alignment
occurs on the same scale as the periodicity of the atoms. We could have
guessed this length scale even above the critical point, for it is essen-
tially the same as the range of each atom’s magnetic influence. In non-
equilibrium patterns this is not so: the scale of ordering vastly exceeds
the range of interaction of the constituents, and there is no obvious hint
of a length scale of this magnitude in the microscopic physics of the
unpatterned state. This is why we can regard such patterns as global
emergent properties of the system, which are likely to remain hidden to
a reductionistic analysis of the microscopic interactions.
Power laws and scaling
Most of the discussion so far has been concerned with patterns that
form in systems that are essentially deterministic, which is to say that at
least in principle we can write down equations (such as the Navier–
Stokes equation) that describe the behaviour exactly. That does not by
any means imply that we can solve the equations, but it follows that,
once the initial and boundary conditions (the rate of heating, say, and
the size of the container) are specified, we know what all the ingredients
of the process are.
Some of the patterns that I have talked about in these books, and
particularly in this final volume, do not share this deterministic char-
acter. The equations that describe them contain a strong random
element that is unpredictable and impossible to formulate in anything
other than statistical, average terms. Diffusion-limited aggregation
(Chapter 2) is like this: the particles become attached to the growing

DLA aggregate only after a random walk through space, jostled by air
molecules say, so that their precise trajectories are indeterminate in
advance. Noise is an essential ingredient of the patterning process in
these cases, and the forms that result look rather disorderly.
The characteristic, invariant forms of systems like this are commonly
‘hidden’—they are mathematical rather than visual. We saw in the case
of sand-pile avalanches (Book II, Chapter 4) how the apparently ran-
dom sizes of slope collapses are governed by the ‘hidden regularity’
characteristic of self-organized criticality, distinguished by power-law
statistics. And the most robust feature of disordered fractals like DLA
clusters or city shapes is their power-law scaling behaviour or fractal
dimension, which enables us to classify and compare structures that
bear no particular visible features in common. Some researchers be-
lieve that power laws are the key to understanding much of the complex
behaviour exhibited by non-equilibrium systems that are strongly in-
fluenced by noise. But it remains unclear to what extent self-organized
criticality offers a general framework for describing such systems.
Nonetheless, it is clear that noise, power-law behaviour, scale invari-
ance, avalanche behaviour and fractal forms are intimately connected
in some deep way that remains to be fully explored and unravelled.
The role of entropy production
No doubt this all looks like a rather piecemeal approach to the issue of
pattern selection in non-equilibrium systems. During the 1960s and
1970s, Ilya Prigogine’s group at Brussels held out the hope of finding
a more general criterion: a ‘minimization principle’ analogous to the
minimization of free energy at equilibrium. In other words, the selected
pattern in each case would be one that minimizes some quantity. As we
saw earlier, Prigogine showed that systems only slightly out of equilib-
rium observed the principle of minimum entropy production. But this
principle did not seem to hold in general for systems further from
equilibrium, which is where spontaneous patterns form. It now appears
that there is probably no such minimization principle that can be
applied in general to all non-equilibrium systems. To address the prob-
lem of pattern selection, we are forced to consider the specific details of
each system. In many cases the only option is to resort to experiment,
to characterize and categorize the taxonomy of dissipative structures

that appear.
Yet there is at least one candidate for a ‘general’ principle of non-
equilibrium pattern selection that applies very widely, if not universally.
It stems from work begun in the 1950s by the American mathematical
physicist Edwin Thompson Jaynes, who attempted to reformulate the
microscopic, molecular-scale foundations of thermodynamics. In the
late nineteenth century, James Clerk Maxwell, Ludwig Boltzmann, and
others showed how the laws of thermodynamics governing the proper-
ties of heat and matter could be understood by considering the behav-
iour of individual atoms and molecules as they jiggle and collide. This
discipline became known as statistical mechanics, since it drew on the
average behaviours in the molecular melée. Boltzmann showed what
entropy really means at this microscopic scale: it is a measure of the
different ways molecules can be arranged. In the 1940s an American
engineer named Claude Shannon showed that the concept of entropy
could be applied not just to molecules but to information: it could
quantify the ways in which units of information may be arranged.
Jaynes then attempted to unite Shannon’s ‘information theory’ with
the statistical mechanics of Boltzmann and his successors. This en-
abled him to start developing a kind of statistical mechanics that
applied also to non-equilibrium systems too—something that Lars
Onsager had imagined but only glimpsed.
Out of this marriage, Jaynes extracted a principle for determining
‘how things happen’ which, he argued, applied equally to equilibrium
and non-equilibrium systems: entropy tends to get maximized. In the
latter case, what this means is that a system tends to adopt the state in
which entropy is produced at the greatest rate: that is, it does not
minimize entropy production, but rather, maximizes it. This idea has
not yet gained general acceptance, but there is growing support for it.
One of the attractive justifications of the theory of maximal entropy
production is that it offers a rationalization of why ordered patterns
may appear far from equilibrium. This, it is worth reminding ourselves
once more, is a highly counter-intuitive phenomenon: we might expect
systems driven out of equilibrium to dissolve into chaos. What is more,
it appears to (although in fact does not) challenge the second law of
thermodynamics, which insists that entropy and thus disorder must
increase. And indeed, if we are now insisting that not only does entropy
increase but it tends to do so at the maximum rate, why should that be a
prescription for order rather than its opposite? The answer, according
to Jaynes’s theory of entropy maximization, is that ordered states are
more effective than disordered ones at producing entropy. To put it
another way: suppose a system has accumulated a lot of energy and
‘needs’ to discharge it. A rather literal expression of that situation is the
build-up of electrical charge in a thundercloud, which may be released
by passing electrical current to the ground. One way this could happen
is for the charge to hop out onto droplets of moisture or dust in the air,
and for these to gradually diffuse down to the ground. That is a slow
process. What often happens instead, of course, is that the charge
grounds itself all at once in a lightning bolt, creating one of the branch-
ing patterns we encountered in this book. Lightning, the dielectric
breakdown of air (page 85), provides a ‘structured channel’ for the
release of the electrical energy at the maximal rate of entropy produc-
tion. As the physicist Roderick Dewar puts it, ‘far from equilibrium, the
coexistence of ordered and dissipative regions produces and exports
more entropy to the environment than a purely dissipate soup’.* And so,
according to Rod Swenson of the University of Connecticut ‘the world
can be expected to produce order whenever it gets the chance’.
The implications of this idea are extraordinary. It is one thing to
explain convective roll cells and river networks as structures that arise
because they offer ‘channels’ for relieving energy stress and producing
entropy efficiently. But some researchers have gone much further than
this. Harold Morowitz of George Mason University in Fairfax, Virginia,
and Eric Smith of the Santa Fe Institute in New Mexico point out that
life itself is an example of non-equilibrium regularity and structure, and
that perhaps it is one of Swenson’s inevitable ordered forms, waiting to
burst forth as soon as the universe gets the chance. Morowitz and Smith
argue that the early Earth was a storehouse of energy ‘needing’ to be
dissipated. In particular, there may have been plentiful hydrogen and
carbon dioxide: two molecules that release energy when they react, but
which do so only very slowly on their own. Primitive living organisms
would have supplied a way for this to happen, ‘fixing’ carbon dioxide
into organic matter through reactions that use electrons extracted from
hydrogen. Similarly, some geological environments generate molecules
rich in electrons and others hungry for them; but only living cells would
*‘Dissipative’ here might be confusing in view of the discussion of ‘dissipative structures’

earlier. Here what it implies is a process of random rather than focused dispersal.
let this transfer proceed at an appreciable rate. In other words, life may
have appeared on the early Earth as a kind of lightning conductor, using
order to speed up entropy production. In that picture, say Morowitz
and Smith, ‘a state of the geosphere which includes life [was] more
likely than a purely abiotic state’.
Life itself
This is a very different view of life from the one scientists have long
wrestled with. They have tended to think that, because even the most
primitive organisms are hideously complicated, and because their in-
gredients seem to be rather rare in an abiotic (inorganic) environment,
life on Earth was a remarkable stroke of luck. That, however, sits uneasily
with geological evidence suggesting that life probably began on our
planet the instant this became geologically feasible—that is, once the
surface was no longer molten, and water had condensed from the
atmosphere as oceans. Moreover, the fact that life seems to thumb its
nose at the second law of thermodynamics, creating order rather than
succumbing to randomness, has long left scientists uneasy: the physi-
cist Erwin Schrödinger felt forced to talk in uncomfortable terms about
life as a source of ‘negative entropy’. The notion of maximum entropy
production and the concomitant drive towards non-equilibrium order
potentially removes such paradoxes. It implies that life itself is a result of
the seemingly irrepressible tendency for order to crystallize away from
equilibrium.
If that is right, we have little reason to fear that we might be alone in
the universe.
Appendix 1
The Hele-Shaw Cell
T
HE cell is basically two clear, rigid plates separated by a small gap.
It’s simplest to make these plates in the form of trays with raised
edges, which keeps the liquid confined to the lower one. Glass is
recommended, but clear plastic (perspex/plexiglass) works fine and is
easier to use. The top tray shown here measures 27 27 cm, and the lower
one 34 34 cm. The perspex is 4 mm thick, and is glued with epoxy resin.
The top plate is separated from the lower one by flat spacers at
each corner—British pennies give about the right separation, as will
American nickels. The viscous liquid is glycerine, bought from a pharma-
cist. For a readily visible and attractive pattern, you can add food colour-
ing. (Using glycerine rather than oil makes the assembly easier to clean.)
Air is injected through a small hole in the top plate; I simply drilled a hole
to fit the empty ink tube from a ball-point pen (about 2 mm internal
diameter). This was glued in place. It is simplest to inject the air through a
plastic syringe connected by rubber tubing; but you can just blow
APPENDIX j 211
through the tube instead. Remember that the viscous fingering pattern is
a non-equilibrium shape, which means that you need to create a sub-
stantial disequilibrium: in plain words, blow hard and sharp!
I have taken this design from:
T. Vicsek, ‘Construction of a radial Hele-Shaw cell’, in Random
Fluctuations and Pattern Growth, ed. H. E. Stanley and N. Ostrowsky
(Dordrecht: Kluwer Academic, 1988), p. 82.
Bibliography
Assenheimer, M., and Steinberg, V., ‘Transition between spiral and target states
in Rayleigh-Bénard convection’, Nature 367 (1994): 345.
Audoly, B., Ries, P. M., and Roman, B., ‘Cracks in thin sheets: when geometry
rules the fracture path, preprint <www.lmm.jussieu.fr/platefracture/preprint_
geometry_fracture.pdf>.
Avnir, D., Biham, O., Lidar D., and Malcai, O., ‘Is the geometry of nature fractal?’
Science 279 (1998): 39–40.
Ball, P., Critical Mass (London: Heinemann, 2004).
Barabási, A.-L., Linked (Cambridge, MA: Perseus, 2002).
Batty, M., and Longley, P., Fractal Cities (London: Academic Press, 1994).
Ben-Jacob, E., Goldenfeld, N., Langer, J. S., and Schön, G., ‘Dynamics of inter-
facial pattern formation’, Physical Review Letters 51 (1983): 1930.
Ben-Jacob, E., ‘From snowflake formation to growth of bacterial colonies. Part I:
diffusive patterning in azoic systems’, Contemporary Physics 34 (1993): 247.
Ben-Jacob, E., ‘From snowflake formation to growth of bacterial colonies. Part II:
cooperative formation of complex colonial patterns’, Contemporary Physics 38
(1997): 205.
Ben-Jacob, E., and Garik P., ‘The formation of patterns in non-equilibrium
growth’, Nature 343 (1990): 523.
Ben-Jacob, E., Shochet, O., Cohen, I., Tenenbaum, A., Czirók, A., and Vicsek, T.,
‘Cooperative strategies in formation of complex bacterial patterns’, Fractals 3
(1995): 849.
Ben-Jacob, E., Shochet, O., Tenenbaum, A., Cohen, I., Czirók, A., and Vicsek, T.,
‘Generic modelling of cooperative growth patterns in bacterial colonies’,
Nature 368 (1994): 46.
Bentley, W. A., and Humphreys, W. J., Snow Crystals (New York: Dover, 1962).
Bergeron V., Berger C., and Betterton, M. D., ‘Controlled irradiative formation of
penitentes’, Physical Review Letters 96 (2006): 098502.
Betterton, M. D., ‘Theory of structure formation in snowfields motivated by
penitentes, suncups, and dirt cones’, Physical Review E 63 (2001): 056129.
Bohn, S., Douady, S., and Couder, Y., ‘Four sided domains in hierarchical space
dividing patterns’, Physical Review Letters 94 (2005): 054503.
Bohn, S., Pauchard, L., and Couder, Y., ‘Hierarchical crack pattern as formed by
successive domain divisions. I. Temporal and geometrical hierarchy’, Physical
Review E 71 (2005): 046214.
Bohn, S., Platkiewicz, J., Andreotti, B., Adda-Bedia, M., and Couder, Y.,
‘Hierarchical crack pattern as formed by successive domain divisions. II.
BIBLIOGRAPHY j 213
From disordered to deterministic behavior’, Physical Review E 71 (2005):

046215.
Bohn, S., Andreotti, B., Douady, S., Munzinger, J., and Couder, Y., ‘Constitutive
property of the local organization of leaf venation networks’, Physical Review E
65 (2002): 061914.
Bowman, C., and Newell, A. C., ‘Natural patterns and wavelets’, Reviews of
Modern Physics 70 (1998): 289.
Brady, R. M., and Ball, R. C., ‘Fractal growth of copper electrodeposits’, Nature
309 (1984): 225.
Brockmann, D., Hufnagel, L., and Geisel, T., ‘The scaling laws of human travel’,
Nature 439 (2006): 462–465.
Brown, J. H., and West, G. B. (eds), Scaling in Biology (New York: Oxford
University Press, 2000).
Buchanan, M., Small World (London: Weidenfeld & Nicolson, 2002).
Buzna, L., Peters, K., Ammoser, H., Kühnert, C., and Helbing, D., ‘Efficient
response to cascading disaster spreading’, Physical Review E 75 (2007): 056107.
Chopard, B., Herrmann, H. J., and Vicsek, T., ‘Structure and growth mechanism
of mineral dendrites’, Nature 353 (1991): 409.
Cohen, R., Erez, K., ben-Avraham, D., and Havlin, S., ‘Resilience of the Internet to
random breakdowns’, Physical Review Letters 85 (2000): 4626–4628.
Couder, Y., Pauchard, L., Allain, C., Adda-Bedia, M., and Douady, S., European
Physical Journal B 28 (2002): 135–138.
Cowie P., ‘Cracks in the Earth’s surface’, Physics World (February 1997): 31.
Cross, M. C., and Hohenberg, P., ‘Pattern formation outside of equilibrium’,
Reviews of Modern Physics 65 (1993): 851.
Czirók, A., Somfai, E., and Vicsek, T., ‘Experimental evidence for self-affine
roughening in a micromodel of geomorphological evolution’, Physical Review
Letters 71 (1993): 2154.
Czirók, A., Somfai, E., and Vicsek, T., ‘Self-affine roughening in a model experi-
ment in geomorphology’, Physica A 205 (1994): 355.
Daerr, A., Lee, P., Lanuza, J., and Clément, É., ‘Erosion patterns in a sediment
layer’, Physical Review E 67 (2003): 065201.
Dawkins, R., River Out of Eden (London: Weidenfeld & Nicolson, 1996).
Dewar, R. C., ‘Maximum entropy production and non-equilibrium statistical
mechanics’, in Lorenz, R. D., and Kleidon, A. (eds), Non-Equilibrium
Thermodynamics and the Production of Entropy: Life, Earth, and Beyond 41
(Berlin and Heidelberg: Springer, 2005).
Dezsö, Z., and Barabási, A.-L., ‘Halting viruses in scale-free networks’, Physical
Review E 65 (2002): 055103(R).
Dimitrov, P., and Zucker, S. W., ‘A constant production hypothesis guides leaf venation
patterning’, Proceedings of the National Academy of Sciences USA 103 (2006): 9363.
Dodds, P. S., Muhamad, R,, and Watts, D. J., ‘An experimental study of search in
global social networks’, Science 301 (2003): 827.
Family, F., Masters, B. R., and Platt, D. E., ‘Fractal pattern formation in human
retinal vessels’, Physica D 38 (1989): 98.
Fleury, V., Arbres de Pierre (Paris: Flammarion, 1998).
Fleury, V., ‘Branched fractal patterns in non-equilibrium electrochemical depos-
ition from oscillatory nucleation and growth’, Nature 390 (1997): 145.
214 j BIBLIOGRAPHY
Garcia-Ruiz, J. M., Louis, E., Meakin, P., and Sander, L. M. (eds), Growth Patterns
in the Physical Sciences and Biology (New York: Plenum Press, 1993).
Ghatak, A., and Mahadevan, L., ‘Crack street: the cycloidal wake of a cylinder
tearing through a thin sheet’, Physical Review Letters 91 (2003): 215507.
Goehring, L., and Morris, S. W., ‘Order and disorder in columnar joints’,
Europhysics Letters 69 (2005): 739–745.
Goehring, L., Morris, S. W., and Lin, Z., ‘An experimental investigation of the
scaling of columnar joints’, Physical Review E 74 (2006): 036115.
Goehring, L., and Morris, S. W., ‘Scaling of columnar joints in basalt’, Journal of
Geophysical Research 113 (2008): B10203.
Gordon, J. E., The New Science of Strong Materials (London: Penguin, 1991).
Gravner, J., and Griffeath, D., ‘Modeling snow crystal growth II: a mesoscopic
lattice map with plausible dynamics’, Physica D 237 (2008): 385.
Gravner, J., and Griffeath, D., ‘Modeling snow-crystal growth: a three-
dimensional mesoscopic approach‘, Physical Review E 79 (2009): 011601.
Hurd, A. J. (ed.), Fractals. Selected Reprints (College Park: American Association
of Physics Teachers, 1989).
Ijjazs-Vasquez, E., Bras, R. L., and Rodriguez-Iturbe, I., ‘Hack’s relation and
optimal channel networks: the elongation of river basins as a consequence
of energy minimization’, Geophysical Research Letters 20 (1993): 1583.
Jacobs, J., The Death and Life of Great American Cities (New York: Vintage, 1961).
Jagla, E. A., and Rojo, A. G., ‘Sequential fragmentation: the origin of columnar
quasihexagonal patterns’, Physical Review E 65 (2002): 026203.
Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N., and Barabási, A.-L., ‘The large-
scale organization of metabolic networks’, Nature 407 (2000): 651.
Kauffman, S., At Home in the Universe (Oxford: Oxford University Press, 1995).
Kessler, D., Koplik, J., and Levine, H., ‘Pattern selection in fingered growth
phenomena’, Advances in Physics 37 (1988): 255.
Kirchner, J. W., ‘Statistical inevitability of Horton’s laws and the apparent ran-
domness of stream channel networks’, Geology 21 (1993): 591.
Landauer, R., ‘Stability in the dissipative steady state’, Physics Today 23
(November 1978).
Landauer, R., ‘Inadequacy of entropy and entropy derivatives in characterizing
the steady state’, Physical Review A 12 (1975): 636.
Libbrecht, K., and Rasmussen, P., The Snowflake: Winter’s Secret Beauty
(Stillwater, MN: Voyageur Press, 2003).
Libbrecht, K., ‘The enigmatic snowflake’, Physics World, January 2008: 19.
Libbrecht, K., ‘The formation of snow crystals’, American Scientist 95(1) (2007): 52.
Liljeros, F., Edling, C. R., Nunes Amaral, L. A., Stanley, H. E., and Åberg, Y., ‘The
web of human sexual contacts’, Nature 411 (2001): 907–908.
Makse, H. A., Havlin, S., and Stanley, H. E., ‘Modelling urban growth patterns’,
Nature 377 (1995): 608.
Mandelbrot, B., The Fractal Geometry of Nature (New York: W. H. Freeman, 1984).
Mandelbrot, B., ‘Fractal geometry: what is it, and what does it do?’, Proceedings of
the Royal Society of London, Series A 423 (1989): 3.
Marder, M., ‘Cracks take a new turn’, Nature 362 (1993): 295.
Marder, M., and Fineberg, J., ‘How things break’, Physics Today 24 (September
1996).
BIBLIOGRAPHY j 215
Maritan, A., Rinaldo, A., Rigon, R., Giacometti, A., and Rodriguez-Iturbe, I.,
‘Scaling laws for river networks’, Physical Review E 53 (1996).
Masters, B. R., ‘Fractal analysis of the vascular tree in the human retina’, Annual
Reviews of Biomedical Engineering 6 (2004): 427–452.
Matsushita, M., and Fukiwara, H., ‘Fractal growth and morphological change in
bacterial colony formation’, in Garcia-Ruiz, J. M., Louis, E., Meakin, P., and
Sander, L. M. (eds), Growth Patterns in Physical Sciences and Biology (New
York: Plenum Press, 1993).
Meakin P., ‘Simple models for colloidal aggregation, dielectric breakdown and
mechanical breakdown patterns’, in Stanley, H. E., and Ostrowsky, N. (eds),
Random Fluctuations and Pattern Growth (Dordrecht: Kluwer, 1988).
Milgram, S., ‘The small world problem’, Psychology Today 2 (1967): 60.
Morowitz, H., and Smith, E., ‘Energy flow and the organization of life. Santa Fe
Institute Working Papers’, available at <http://www.santafe.edu/research/
publications/workingpapers/06-08-029.pdf>.
Müller, G., ‘Starch columns: analog model for basalt columns’, Journal of
Geophysical Research 103, B7 (1998): 15239–15253.
Mullins, W. W., and Sekerka, R. F., ‘Stability of a planar interface during solidifi-
cation of a dilute binary alloy’, Journal of Applied Physics 35 (1964): 444.
Mumford, L., The Culture of Cities (London: Secker and Warburg, 1938).
Murray, A. B., and Paola, C., ‘A cellular model of braided rivers’, Nature 371
(1994): 54.
Nicolis, G. ‘Physics of far-from-equilibrium systems and self-organization’, in
Davies, P. (ed.), The New Physics (Cambridge: Cambridge University Press, 1989).
Niemeyer, L., Pietronero, L., and Wiesmann, H. J., ‘Fractal dimensions of dielec-
tric breakdown’, Physical Review Letters 52 (1984): 1033.
Nittmann, J., and Stanley, H. E., ‘Tip splitting without interfacial tension and den-
dritic growth patterns arising from molecular anisotropy’, Nature 321 (1986): 663.
Nittmann, J., and Stanley, H. E., ‘Non-deterministic approach to anisotropic
growth patterns with continuously tunable morphology: the fractal properties
of some real snowflakes’, Journal of Physics A 20 (1987): L1185.
Oikonomou, P., and Cluzel, P., ‘Effects of topology on network evolution’, Nature
Physics 2 (2006): 532.
Pastor-Satorras, R., and Vespignani, A., ‘Epidemic spreading in scale-free net-
works’, Physical Review Letters 86 (2001): 3200.
Pastor-Satorras, R., and Vespignani, A., ‘Optimal immunisation of complex net-
works’, preprint <http://www.arxiv.org/abs/cond-mat/0107066 (2001)>.
Perrin, B., and Tabeling, P., ‘Les dendrites’, La Recherche 656 (May 1991).
Prigogine, I., From Being to Becoming (San Francisco: W. H. Freeman, 1980).
Prusinkiewicz, P., and Lindenmayer, A., The Algorithmic Beauty of Plants (New
York: Springer, 1990).
Rigon, R., Rinaldo, A., and Rodriguez-Iturbe, I., ‘On landscape self-organization’,
Journal of Geophysical Research 99 (B6) (1994): 11971.
Rinaldo, A., Banavar, J. R., and Maritan, A., ‘Trees, networks, and hydrology’,
Water Resources Research 42 (2006): W06D07.
Rodriguez-Iturbe, I., and Rinaldo, A., Fractal River Basins. Chance and Self-
Organization (Cambridge: Cambridge University Press, 1997).
Rodriguez-Iturbe, I., Rinaldo, A., Rigon, R., Bras, R. L., Ijjasz-Vasquez, E., and
Marani, A., ‘Fractal structures as least energy patterns: the case of river net-
works’, Geophysical Research Letters 19 (1992): 889.
216 j BIBLIOGRAPHY
Ryan, M. P., and Sammis, C. G., ‘Cyclic fracture mechanisms in cooling basalt’,
Geological Society of America Bulletin 89 (1978): 1295.
Sapoval, B., Baldassarri, A., and Gabrielli, A., ‘Self-stabilized fractality of sea-
coasts through damped erosion’, Physical Review Letters 93 (2004): 098501.
Sapoval, B., Universalités et Fractales (Paris: Flammarion, 1997).
Sander, L. M., ‘Fractal growth’, Scientific American 256(1) (1987): 94.
Shorling, K. A., Bruyn, J. R. de, Graham, M., and Morris, S. W., ‘Development and
geometry of isotropic and directional shrinkage crack patterns’, Physical
Review E 61, 6950 (2000).
Skjeltorp, A., ‘Fracture experiments on monolayers of microspheres’, in Stanley,
H. E., and Ostrowsky, N. (eds), Random Fluctuations and Pattern Growth
(Dordrecht: Kluwer, 1988).
Sinclair, K., and Ball, R. C., ‘Mechanism for global optimization of river networks
from local erosion rules’, Physical Review Letters 76 (1996): 3360.
Stanley, H. E., and Ostrowsky, N. (eds), On Growth and Form (Dordrecht:
Martinus Nijhoff, 1986).
Stanley, H. E., and Ostrowsky, N. (eds), Random Fluctuations and Pattern Growth
(Dordrecht: Kluwer, 1988).
Stark, C., ‘An invasion percolation model of drainage network evolution’, Nature
352 (1991): 423.
Stewart, I., What Shape is a Snowflake? Magical Numbers in Nature (London:
Weidenfeld & Nicolson, 2001).
Swenson, R., ‘Autocatalysis, evolution, and the law of maximum entropy
production: a principled foundation towards the study of human ecology’,
Advances in Human Ecology 6 (1997): 1–47.
Swinney, H., ‘Emergence and the evolution of patterns’, in Fitch, V. L., Marlow,
D. R., and Dementi, M. A. E. (eds), Critical Problems in Physics (Princeton:
Princeton University Press, 1997).
Temple, R. K. G., The Genius of China (London: Prion Books, 1998).
Thompson, D’A. W., On Growth and Form (New York: Dover, 1992).
Van Damme, H., and Lemaire, E., ‘From flow to fracture and fragmentation in
colloidal media’, in Charmet, J. C., Roux, S., and Guyon, E. (eds), Disorder and
Fracture (New York: Plenum Press, 1990).
Vella, D., and Wettlaufer, J. S., ‘Finger rafting: a generic instability of floating ice
sheets’, Physical Review Letters 98 (2007): 088303.
Watts, D. J., and Strogatz, S. H., ‘Collective dynamics of ‘‘small-world’’ networks’,
Nature 393 (1998): 440–442.
Watts, D. J., Small Worlds (Princeton: Princeton University Press, 1999).
Watts, D. J., Dodds, P. S., and Newman, M., ‘Identity and search in local networks’,
Science 296 (2002): 1302–1305.
Watts, D. J., Six Degrees (New York: W. W. Norton, 2004).
Weaire, D., and O’Carroll, C., ‘A new model for the Giant’s Causeway’, Nature 302
(1983): 240–241.
West, G. B., Brown, J. H., and Enquist, B. J., ‘A general model for the origin of
allometric scaling laws in biology’, Science 276 (1997): 122.
Whitfield, J., In the Beat of a Heart (Washington: Joseph Henry Press, 2006).
Yuse, A., and Sano, M., ‘Transitions between crack patterns in quenched glass
plates’, Nature 362 (1993): 329.
Index
action, in mechanics 110, 111 braided rivers 114–117

activator-inhibitor systems 181, 183 breakdowns, of networks 149–151,
Albert, Réka 162, 165, 166, 175, 176 175–178
Albertus Magnus 2 Brockmann, Dirk 174
algorithms 130, 132–136 Brown, James 138–142
allometric scaling 138–142
Amaral, Luis 168 Carson, Rachel 115
angiogenesis 144–146 Cassini, Giovanni Domenico 8, 9
anisotropy 20, 21 chemical waves 181, 183
attractors 188 chemotaxis 55, 56, 144, 145
auxin 147, 148 Chickering, Frances Knowlton 10
Avnir, David 44 Chinese scholarship, on snowflakes 2, 3
Chopard, Bastien 38
Bacon, Kevin 159 Chu Hsi 2
Bacon Number 159, 160 citys, see urban growth
bacterial growth morphologies 50–59, clouds, fractal properties 39
201 coastlines 119–124
Ball, Robin 30, 32, 112 collaboration networks 160, 161
Banks, Joseph 70 columnar cracking 70–73, 95–99
Barabási, Albert-László 162, 163, 165, convection 179, 180, 184, 187–190, 193,
166, 175–177 194, 197, 198, 201–204
Batty, Michael 60, 62, 64–66, 69 correlated percolation model 66–69
Belousov-Zhabotinsky reaction 183, correlations 66, 202–206
188, 198, 199 Cowan, Jim 58, 59
Ben-Jacob, Eshel 54–59 cracks 70–99, 148
Bentley, William 11, 12 see also fracture
Bergeron, Vance 129 craquelure 90–92
bifurcations 80, 187, 189–193, 198, 199 critical points and transitions 181, 192,
biochemical networks 150, 165, 166, 193, 204–206
169, 170, 177 crystal shapes 6, 7
Biringuccio 28 cyberwarfare 176, 177
blood, circulatory networks 134, Czirók, András 54
136–148
Bohn, Steffen 93–95, 148 Daerr, Adrian 116, 118
Boltzmann, Ludwig 208 Darwin, Charles 127
Brady, Robert 30, 32 Dawkins, Richard 100
Bragg, William 19 defects 201, 202
218 j INDEX
dendritic growth 15–19 Galileo 77

see also mineral dendrites Garcia-Ruiz, Juan Manuel 44, 45
dense-branching morphology 19, 47, 48, genealogies 151
53 geology 70–73, 84, 95–99
Descartes, René 6–8 Ghatak, Animangsu 80, 81
Devil’s Postpile 96 Giant’s Causeway 70–73, 95–99, 201
Dewar, Roderick 209 Goehring, Lucas 96, 98, 99
Dezsö, Zoltán 177 Gordon, James 76, 77
dielectric breakdown model 64, 69, Gottschalk, Johann 29
85–89, 91, 107, 108 Gould, Stephen Jay 100
diffusion-limited aggregation 30–41, 43, Glansdorff, Paul 187, 188
45–47, 49, 50, 52, 60, 61, 64, 65, 69, glass, fracture 76–80
89, 206, 207 Glauber, Johann 28
Dimitrov, Pavel 147 glaze, cracking 92–94
dissipative structures 188, 189, 207 Glicksman, Martin 15, 19
Dodds, Peter 172 graph theory 152–170
Dyson, Freeman 195 Gravner, Janko 24, 25
Griffeath, David 24, 25
Eden, M. 50 Griffith, Alan Arnold 77, 78
Eden growth 50–53 Guare, John 160
electrodeposition 15, 16, 30–32, 43, 44, 201
email networks 171, 173 Hack, John 104
Enquist, Brian 138–142 Hack’s law of river networks 104, 109, 112
entropy 186, 208 Hamilton, William 110
rate of production 186, 187, 207–210 Hamiltonian mechanics 110
epidemiology 152, 153, 173–175, 177 Han Ying 2
Erdös, Paul 153, 155 Hariot, Thomas 4, 5
Erdös Number 160 Häuy, René Just 7
erosion 112–118, 121–129 Havlin, Shlomo 66
Euler, Leonhard 152 Helbing, Dirk 178
Hele-Shaw, Henry 46
Faloutsos, Michalis, Petros and Hele-Shaw cell 46–50, 53, 211
Christos 163 Honda, Hisao 134–136
Fingal’s Cave 70, 71 Hooke, Robert 8, 9
finger rafting, ice 82, 83 Horton, Robert E. 103, 104
Finn MacCool (Fingal) 70, 71 Horton’s laws of river networks 103–105,
Fleury, Vincent 46 112
Ford, W. W. 58 Hsiao T’ung 2
fractals 34–45, 47, 50, 62–64, 88, 89, 91, Hsien Tsai-hang 2
118–125, 141–144, 180 Humphreys, William J. 11, 12
fractal dimension 34–37, 53, 62, 63, 88, Huxley, Thomas 8, 9
89, 91, 106, 118, 142–144, 180
The Fractal Geometry of Nature ice, crystal structure 19, 20
(Mandelbrot) 39 information theory 208
fracture 76–80, 83–95 Inglis, Charles 77
see also cracks Internet, topology 150, 163, 164,
freezing 189, 190, 192 168–170, 175, 176
French, J. W. 96 invasion percolation 107–110
Fujikawa, H. 51 Ivantsov, G. P. 15, 19
INDEX j 219
Jacobs, Jane 60 minimization principles 111, 112,

Jagla, Eduardo 96, 98 136–142, 186, 187, 196, 207–210
Jaynes, Edwin Thompson 208 Morowitz, Harold 209, 210
Jeong, Hawoong 162, 175, 176 Morris, Stephen 96, 98, 99
mountains, fractal 120–122
Kevin Bacon Game 159, 160 movie-star networks 158–160, 168
Kepler, Johann 4–6, 20, 180 Müller, Gerhard 96
Kessler, David 21 Müller-Khrumbhaar, Hans 19
Kirchner, James 105 Mullins, William 17
Kleinberg, Jon 173 Mullins-Sekerka instability 17–19, 21, 46
Koch snowflake 41–43 Mumford, Lewis 60
Königsberg bridges problem 152 Murray, Brad 115–117
Koplik, Joel 21 Murray, Cecil 134–137
mutation, and bacterial growth
Lagrangian mechanics 110 patterns 57–59
Lagrange, Joseph Louis 42, 110
Landauer, Rolf 180 Nakaya, Ukichiro 11, 13, 22, 25
Langbein, Walter 105, 107 natural selection 58
Langer, James 19, 21 Navier-Stokes equation 194, 195
Laplace, Pierre-Simon 42, 46 Needham, Joseph 3
Laplacian instabilities 46, 50, 143, 179 Neoplatonism 101
leaves, vasculatory networks 144–148 networks 84–114, 130–178
Leonardo da Vinci 100–102, 118, 132, 133 neural networks 161
Leopold, Luna 105, 107, 111, 137 Newman, Mark 160, 161, 169, 170, 172
Levine, Herbert 21 Newton, Isaac 27, 28
Libbrecht, Kenneth 22 Niemeyer, Lutz 85, 86
Liben-Nowell, David 172, 173 Nittmann, Johann 22, 23
lichen 50, 51 non-equilibrium growth 15, 185–189
Lichtenberg figures 85, 86 Nuts in May (Leigh) 26
Lichtenberg, Georg Cristoph 86
lightning 85 O’Carroll, Conor 98
Longley, Paul 60, 62, 64–66 On Growth and Form (Thompson) 13, 142
On the Six-Cornered Snowflake
magnetism 189–191, 205, 206 (Kepler) 5, 6
Mahadevan, Lakshminarayanan 81 Onsager, Lars 186, 187
Makse, Hernán 66, 69 optimal channel networks 111–114, 124,
Mandelbrot, Benoit 39, 42, 119, 120 125
Mandelbrot set 40–42 origin of life 209, 210
Martens, Friedrich 10 oscillating chemical reactions 180, 186
Matthew Principle 167
Matsushita, Mitsugu 32, 51–53 packing, of spheres 4, 5
Maxwell, James Clerk 208 Palissy, Bernard 28
Meakin, Paul 91 Paola, Chris 115–117
Meinhardt, Hans 148 Paracelsus 28
Merton, Robert 166 Pareto, Vilfredo 166
metabolic rate 139, 142 Pareto law 166
Milgram, Stanley 160, 171 Pastor-Satorras, Romualdo 174
mineral dendrites 26–30, 36–39, 44 penitentes 127–129
220 j INDEX
phase transitions 181, 189–193 self-affinity 120

Pietronero, Luciano 85 self-organized criticality 207
Plato 1 self-similarity 41
Platonism 1, 3 sexual-contact networks 175
power laws 34, 104, 138–142, 162, 174, Shannon, Claude 208
207 Shreve, Ronald 105
power networks 149–151, 161, 168, 177, silica gardens 28
178 Sinclair, Kevin 112
Prigogine, Ilya 186–188, 207 Six Degrees of Separation (Guare) 160
Prusinkiewicz, Przemyslaw 135–137 Skjeltorp, Arne 91, 92
Pythagoras 1 small-world networks 156–178
Smith, Cyril Stanley 98
Raleigh, Walter 4 Smith, Eric 209, 210
random graphs 153–158, 162–164, 170 Snow Crystals (Bentley and
random rewiring networks 157, 158, 162 Humphreys) 11, 12
Rapaport, Anatol 152, 153, 155 snow, erosion 127–129
Rayleigh, Lord 181, 193, 194 snowflakes 1–25, 30, 31, 44, 47, 49, 182
reaction-diffusion systems 39, 148, 179, soap films 181
181, 183, 198 social networks 151–175
Rényi, Alfred 153, 155 Stanley, Gene 22, 23, 66
rich-club networks 170 Stark, Colin 109
Richardson, Lewis Fry 120 Strahler, Arthur 103
Rinaldo, Andrea 111–114, 124 stream networks, see river networks
river networks 100–117 street patterns 93–95
topography 118, 124–127 see also road networks
road networks 150, 168 stress concentration, in fracture 77, 78
Rodriguez-Iturbe, Ignacio 111–114, 124 Strogatz, Steven 156–159, 161, 167
Rojo, Alberto 96, 98 surface tension 18, 47, 182
Roux, Wilhelm 132–134 symmetry-breaking 184, 185, 190, 196,
Russell, Bertrand 1 197
Saffman, P. G. 46 T’ang Chin 2

Saffman-Taylor instability 46, 47 Taylor, Geoffrey 46, 181, 193, 194
Sano, Masaki 79, 80 tearing 81
Sander, Len 30 thermodynamics 185, 189–193
Sapoval, Bernard 122–124 non-equilibrium 185–189, 207–210
Sauvages, Abbé de 46 Thompson, D’Arcy Wentworth 13, 14,
scale-free networks 162–170, 174–178 20, 71, 74, 76, 142, 180
scale-invariance 41, 204 Thoreau, Henry David 25
scaling law, see power law Tjaden, Brett 159
Scheuchzer, Jean-Jacques 28, 29, 45 topology, of networks 152, 153
Schreiner, Wolfgang 137, 138 transport networks 63
Schrödinger, Erwin 210 trees, branching patterns 130–137
Schumm, Stanley 104 Turing patterns 183, 187, 197, 198, 201,
Scoresby, William 10 202, 204
searches, of networks 170–174
sedimentation 114–118 urban growth 59–69
Sekerka, Robert 17 urban planning 61, 62, 69
INDEX j 221
vascular networks 132–148 Whymper, Edward 120

Vella, Dominic 83 Wiesmann, Hans Jurg 85
Vespignani, Alessandro 174 Wilkinson, David 107
Vicsek, Tamás 54, 125 Willemsen, J. F. 107
viscous fingering 44–50, 182, 211 Willis, John 79
Volta, Alessandro 86 Witten, Thomas 30
World Wide Web, topology 162, 165, 167,
Wasson, Glenn 159 168, 170
Watts, Duncan 156–159, 161, 167,
171, 172 Yoffe, Elizabeth 78, 79
Weaire, Denis 98 Yuse, Akifumi 79, 80
West, Geoffrey 138–142
Wettlaufer, John 83 Zucker, Steven 147
Every effort has been made to trace and contact copyright holders and the publisher and
author apologize for any errors or omissions. If notified, the publisher will undertake to
rectify these at the earliest opportunity.
Plate 1 A snowflake displays a
delicate balance of chance (in the
initiation of branches) and determinism
(the sixfold symmetry). (Photo: Ken
Libbrecht, California Institute of
Technology.)
Plate 2 The wispy boundaries of many clouds trace out a fractal form. (Photo: Maciej Szczepaniak.)
Plate 3 Branching patterns in bacterial
colonies. (Images: Eshel Ben-Jacob and
Kinneret Ben Knaan, Tel Aviv University.)
Plate 7 Rivers carve out a complex

topography of hills and valleys as they erode
the land surface. (Photo: Jim Kirchner,
University of California at Berkeley.)
Plate 4 The entrance to Fingal’s Cave on the island of Staffa in Scotland is announced by a colonnade of
natural pillars of roughly hexagonal cross section. (Photo: Lucas Goehring, University of Toronto.)
Plate 5 The hexagonal columns of the Giant’s Causeway, County Antrim, Northern Ireland. (Photo: Stephen
Morris, University of Toronto.)
Plate 6 The fractal character of natural
mountainous terrain is evident in this aerial
photograph of the Himalayas. (Photo:
NASA/Eros Data Center.)
Plate 8 Trees supply some of the most

familiar and beautiful of nature’s branching
patterns. (Photo: Simon Davidson.)

Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)

Uploaded by

Copyright:

Available Formats

Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What patterns does the book discuss?

What patterns does the book discuss?

Nature’s Patterns

This page intentionally left blank

Preface and acknowledgements ix

3: Just For the Crack 71

4: Water Ways 101

5: Tree and Leaf 131

6: Web Worlds 150

Epilogue: The Threads of the Tapestry 180

mathematics as any of their Western counterparts, and by all appear-

By the seventeenth century, Chinese scholars had become more sys-

a commonplace fact. And so, in the words of sinologist Joseph

to be replaced by something more empirical, more discerning and scep-

which he attributes to a ‘formative faculty’ or plant soul. But ‘to imagine

Fig. 1.3: Drawings of snowﬂakes by René Descartes in 1637.

Fig. 1.4: Twelve-pointed snowﬂakes are formed when two normal

We can recognize in this vivid description some of the unusual forms

It was a reasonable enough assertion, but it surely begs the question: if

Flakes frozen on ﬁlm

As Descartes hinted, the shapes of snowﬂakes can evolve and mutate as

Fig. 1.7: The collection of

Fig. 1.9: The ‘morphology diagram’ of snowﬂakes, showing how their

‘ringing her changes on this fundamental form, Nature superadds to

By the time Nakaya’s book appeared, scientists had found a way to

symmetrical, isolated ﬂakes, many other substances may crystallize as

Fig. 1.11: The Mullins–Sekerka instability makes protrusions at the

In principle, this instability will amplify any irregularity on the solid

In 1977 James Langer at the University of California at Santa Barbara

The joy of six

Why hexagons? Remarkably, the Chinese sages were right: there is a

Fig. 1.12: Ice has a structure with hexagonal symmetry at the

crystal structure is dictated not by the shapes of the water molecules

The anisotropy of crystals means that properties like surface tension

The mind-boggling variety of snowﬂake forms is therefore the outcome

the general Christmas-tree shape is preserved in all of them, and their

able to grow a huge variety of snowﬂakes in their scheme (Fig. 1.16a). In

glance, although differing in detail (Fig. 1.16b). Gravner and Griffeath

Fig. 2.1: A mineral dendrite of manganese oxide found in Bavarian

species of a tree in winter simply from its silhouette on a hilltop. How

Keith and Candice Marie were in venerable company. In the early

was probably considered by its author to be the precursor to a (never-

of the Deluge—but not, he said, as fossil plants. Instead, he suggested a

Already, life was deemed dispensable in making ‘organic’ forms.

Deposit at your nearest branch

Today it is a rather simple matter to grow ramiﬁed crystals such as

Fig. 2.3: A branching metal formation produced by electrochemical

the water molecules in the ice crystal stack together in an orderly

Fig. 2.4: A cluster of aggregated particles grown by the diffusion-

into long, narrow ﬁngers. When a bump appears by chance on the

A cluster grown by the DLA mechanism really does resemble a

we cannot be sure that this is because electrodeposition is basically a

Fig. 2.6: A mineral dendrite inside a quartz crystal. (Photo: Tamás

Fig. 2.7: Mineral dendrite patterns generated by a computer model in

a reaction–diffusion model akin to those I described in Book I, where we

such as triangles, squares, and circles—and those that seem merely

the line is inﬁnitely convoluted and it is impossible to say exactly where