Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)
Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)
Philip Ball - Branches - Nature's Patterns - A Tapestry in Three Parts (2009, Oxford University Press)
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.
This page intentionally left blank
Contents
1: A Winter’s Tale 1
The Six-Pointed Snowflake
2: Tenuous Monsters 27
Shapes between Dimensions
Appendix 210
Bibliography 212
Index 217
This page intentionally left blank
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
This page intentionally left blank
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
The ruddy clouds float in the four quarters of the caerulean sky
And the white snowflakes show forth their six-petalled flowers.
Fig. 1.1: The snowflake displays an urge for branching growth played
out with exquisite hexagonal symmetry. (Photo: Ken Libbrecht,
California Institute of Technology.)
4 j NATURE’S PATTERNS: BRANCHES
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
A WINTER’S TALE j 5
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.
6 j NATURE’S PATTERNS: BRANCHES
After this storm cloud, there came another, which produced only little roses or
wheels with six rounded semicircular teeth . . . which were quite transparent and
A WINTER’S TALE j 7
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.
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
8 j NATURE’S PATTERNS: BRANCHES
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.
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.
Fig. 1.6: In 1820, explorer William Scoresby made accurate drawings of the snowflakes he
observed during a trip to the Arctic.
A WINTER’S TALE j 11
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
12 j NATURE’S PATTERNS: BRANCHES
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
14 j NATURE’S PATTERNS: BRANCHES
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
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.)
A WINTER’S TALE j 17
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.
18 j NATURE’S PATTERNS: BRANCHES
How the right arm knows what the left arm is doing
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
A WINTER’S TALE j 23
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
Institute of Technology.)
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.
24 j NATURE’S PATTERNS: BRANCHES
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.)
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).
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).
26 j NATURE’S PATTERNS: BRANCHES
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.
28 j NATURE’S PATTERNS: BRANCHES
Organic rocks
Fig. 2.2: Mineral dendrites drawn by Jean-Jacques Scheuchzer, in his book Herbarium diluvianum
(1709).
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.
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. 2.5: 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.
TENUOUS MONSTERS j 35
typically have the same kind of shape, making them as light and airy as
soufflé.
What’s in a branch?
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
38 j NATURE’S PATTERNS: BRANCHES
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.
TENUOUS MONSTERS j 39
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
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
TENUOUS MONSTERS j 41
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.
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,
TENUOUS MONSTERS j 43
Fig. 2.10: A fractal object called the Koch snowflake is produced by introducing identical kinks into
a line at successively smaller scales.
*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.
44 j NATURE’S PATTERNS: BRANCHES
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.
*I have discussed the aesthetics of ‘fractal landscapes’ in Nature 438 (2005): 915.
TENUOUS MONSTERS j 45
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 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
TENUOUS MONSTERS j 47
*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.
48 j NATURE’S PATTERNS: BRANCHES
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.)
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.)
54 j NATURE’S PATTERNS: BRANCHES
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.)
TENUOUS MONSTERS j 57
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.
Ben-Jacob and his co-workers not only studied bacterial growth theor-
etically; they also learnt the microbiological techniques needed to grow
58 j NATURE’S PATTERNS: BRANCHES
Fig. 2.18: A mutant colony with a new growth pattern sprouting from a dense cluster of Bacillus
subtilis. (Photo: Eshel Ben-Jacob.)
TENUOUS MONSTERS j 59
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.
60 j NATURE’S PATTERNS: BRANCHES
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.)
TENUOUS MONSTERS j 61
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.
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
TENUOUS MONSTERS j 63
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
64 j NATURE’S PATTERNS: BRANCHES
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.)
TENUOUS MONSTERS j 65
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.)
TENUOUS MONSTERS j 67
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.)
70 j NATURE’S PATTERNS: BRANCHES
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
72 j NATURE’S PATTERNS: BRANCHES
Fig. 3.1: The natural pillars of Fingal’s Cave on the Scottish island
of Staffa. (Photo: Lucas Goehring, University of Toronto.)
*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.)
74 j NATURE’S PATTERNS: BRANCHES
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?
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
78 j NATURE’S PATTERNS: BRANCHES
Jagged edge
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.
JUST FOR THE CRACK j 81
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.
82 j NATURE’S PATTERNS: BRANCHES
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.)
84 j NATURE’S PATTERNS: BRANCHES
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
JUST FOR THE CRACK j 85
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
86 j NATURE’S PATTERNS: BRANCHES
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.)
JUST FOR THE CRACK j 87
*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.
88 j NATURE’S PATTERNS: BRANCHES
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.
JUST FOR THE CRACK j 91
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.)
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
JUST FOR THE CRACK j 93
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
94 j NATURE’S PATTERNS: BRANCHES
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.)
*I know of no other paper with an acknowledgement like this one: ‘I thank my wife for her
patience with my kitchen interferences.’
98 j NATURE’S PATTERNS: BRANCHES
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.)
JUST FOR THE CRACK j 99
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.
102 j NATURE’S PATTERNS: BRANCHES
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 ).
104 j NATURE’S PATTERNS: BRANCHES
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
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.
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
108 j NATURE’S PATTERNS: BRANCHES
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.)
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
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
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
116 j NATURE’S PATTERNS: BRANCHES
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
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
118 j NATURE’S PATTERNS: BRANCHES
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
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
120 j NATURE’S PATTERNS: BRANCHES
Fig. 4.12: Self-affine fractal landscapes generated by computer look very like real mountainous
terrain. (Image: John Beale.)
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
Carried away
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
128 j NATURE’S PATTERNS: BRANCHES
Inverted icicles
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.
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).
130 j NATURE’S PATTERNS: BRANCHES
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
132 j NATURE’S PATTERNS: BRANCHES
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).
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
134 j NATURE’S PATTERNS: BRANCHES
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.
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.)
138 j NATURE’S PATTERNS: BRANCHES
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.)
*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.
142 j NATURE’S PATTERNS: BRANCHES
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
144 j NATURE’S PATTERNS: BRANCHES
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.)
146 j NATURE’S PATTERNS: BRANCHES
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
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,
154 j NATURE’S PATTERNS: BRANCHES
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
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
‘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
160 j NATURE’S PATTERNS: BRANCHES
*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.
162 j NATURE’S PATTERNS: BRANCHES
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 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.
164 j NATURE’S PATTERNS: BRANCHES
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
168 j NATURE’S PATTERNS: BRANCHES
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
170 j NATURE’S PATTERNS: BRANCHES
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.
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
172 j NATURE’S PATTERNS: BRANCHES
Communication breakdown
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
178 j NATURE’S PATTERNS: BRANCHES
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
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
Competing forces
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.)
THE THREADS OF THE TAPESTRY j 185
Symmetry breaking
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
THE THREADS OF THE TAPESTRY j 187
*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.
188 j NATURE’S PATTERNS: BRANCHES
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
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.
THE THREADS OF THE TAPESTRY j 191
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
192 j NATURE’S PATTERNS: BRANCHES
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 THREADS OF THE TAPESTRY j 193
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
194 j NATURE’S PATTERNS: BRANCHES
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 THREADS OF THE TAPESTRY j 195
Pattern selection
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. 7.8: 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
200 j NATURE’S PATTERNS: BRANCHES
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.
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
THE THREADS OF THE TAPESTRY j 201
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.
202 j NATURE’S PATTERNS: BRANCHES
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).
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.)
THE THREADS OF THE TAPESTRY j 203
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.
204 j NATURE’S PATTERNS: BRANCHES
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.)
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
206 j NATURE’S PATTERNS: BRANCHES
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,
THE THREADS OF THE TAPESTRY j 207
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
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
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
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
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 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.)