Fractals and Chaos in Geology and Geophysics

mn CM:9209630051
FRRCTRLS I D CHROS I N GEDLDGY

lUTHUR : CMBRIDGE
336 EMTH SCIEHCE
The fundamental concepts of fractal geometry and chaotic dynamics, along
with the related concepts of multifractals, self-similar time series, wavelets,
and self-organized criticality, are introduced in this book, for a broad range
of readers interested in complex natural phenomena.
Now in a greatly expanded, second edition, this book relates fractals and
chaos to a variety of geological and geophysical applications. These include
drainage networks and erosion, floods, earthquakes, mineral and petroleum
resources, fragmentation, mantle convection, and magnetic field generation.
Many advances have been made in the field since the first edition was pub-
lished. In this new edition coverage of self-organized criticality is expanded
and statistics and time series are included to provide a broad background for
the reader. All concepts are introduced at the lowest possible level of mathe-
matics consistent with their understanding, so that the reader requires only a
background in basic physics and mathematics.
Fractals and Chaos in Geology and Geophysics can be used as a text for
advanced undergraduate and graduate courses in the physical sciences. Prob-
lems are included for the reader to solve.
FRACTALS AND
CHAOS IN
GEOLOGY AND
GEOPHYSICS
FRACTA
CHAOS
GEOLOGY A
GEOPHYSICS
Second Edition
DONALD L. TURCOTTE
Cornell University
CAMBRIDGE
UNIVERSITY PRESS
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, S l o Paulo
Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521561648
O Donald L. Turcotte 1997
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 1992

Second edition 1997
A catalogue record for this publication is available from the British Library
Library of Congress Cataloguing in Publica~iondata

Turcotte, Donald Lawson.
Fractals and chaos in geology and geophysics / Donald L. Turcotte.
- 2nd ed.
p. cm.
Includes bibliographical references (p. 343-70) and index.
ISBN 0-521-56164-7 (hc). -ISBN 0-521-56733-5 (pbk.)
1. Geology - Mathematics. 2. Geophysics - Mathematics.
3. Fractals. 4. Chaotic behavior in systems. I. Title.
QE33.2.M3T87 1997
550'. 1'51474 - dc20 96-3 1558
CIP
ISBN 978-0-521-56164-8 hardback

ISBN 978-0-52 1-56733-6 paperback
Transferred to digital printing (with amendments) 2007
The color figures within this publication have been removed for this digital
reprinting. At the time of going to press the original images were available
in color for download from http://www.cambridge.org/9780521567336
CONTENTS
Preface
Preface to the second edition
1 Scale invariance
2 Definition of a fractal set
2.1 Deterministic fractals
2.2 Statistical fractals
2.3 Depositional sequences
2.4 Why fractal distributions?
3 Fragmentation
3.1 Background
3.2 Probability and statistics
3.3 Fragmentation data
3.4 Fragmentation models
3.5 Porosity
4 Seismicity and tectonics
4.1 Seismicity
4.2 Faults
4.3 Spatial distribution of earthquakes
4.4 Volcanic eruptions
5 Ore grade and tonnage
5.1 Ore-enrichment models
5.2 Ore-enrichment data
5.3 Petroleum data
6 Fractal clustering
6.1 Clustering
6.2 Pair-correlation techniques
6.3 Lacunarity
6.4 Multifractals
vi CONTENTS
7 Self-affine fractals
7.1 Definition of a self-affine fractal
7.2 Time series
7.3 Self-affine time series
7.4 Fractional Gaussian noises and fractional
Brownian walks
7.5 Fractional log-normal noises and walks
7.6 Rescaled-range (WS)analysis
7.7 Applications of self-affine fractals
8 Geomorphology
Drainage networks
Fractal trees
Growth models
Diffusion-limited aggregation (DLA)
Models for drainage networks
Models for erosion and deposition
Floods
Wavelets
9 Dynamical systems
9.1 Nonlinear equations
9.2 Bifurcations
10 Logistic map
10.1 Chaos
10.2 Lyapunov exponent
11 Slider-block models
12 Lorenz equations
13 Is mantle convection chaotic?
14 Rikitake dynamo
15 Renormalization group method
15.1 Renormalization
15.2 Percolation clusters
15.3 Applications to fragmentation
15.4 Applications to fault rupture
15.5 Log-periodic behavior
16 Self-organized criticality
16.1 Sand-pile models
16.2 Slider-block models
16.3 Forest-fire models
CONTENTS vii
17 Where do we stand?
References
Appendix A: Glossary of terms
Appendix B: Units and symbols
Answers to selected problems
Index
PREFACE
I was introduced to the world of fractals and renormalization groups by Bob

Smalley in 198 1. At that time Bob had transferred from physics to geology at
Cornell as a Ph.D. student. He organized a series of seminars and convinced
me of the relevance of these techniques to geological and geophysical prob-
lems. Although his official Ph.D. research was in observational seismology,
Bob completed several renormalization and fractal projects with me. Subse-
quently, my graduate students Jie Huang and Cheryl Stewart have greatly
broadened my views of the world of chaos and dynamical systems. Original
research carried out by these students is included throughout this book.
The purpose of this book is to introduce the fundamental principles of
fractals, chaos, and aspects of dynamical systems in the context of geologi-
cal and geophysical problems. My goal is to introduce the fundamental con-
cepts at the lowest level of mathematics that is consistent with the under-
standing and application of the concepts. It is clearly impossible to discuss
all aspects of applications. I have tried to make the applications reasonably
comprehensible to non-earth scientists but may not have succeeded in all
cases.
After an introduction, the next seven chapters are devoted to fractals.
The fundamental concepts of self-similar fractals are introduced in Chapter
2. Applications of self-similar fractals to fragmentation, seismicity and tec-
tonics, ore grades and tonnage, and clustering are given in the next chapters.
Self-affine fractals are introduced in Chapter 7 and are applied to geomor-
phology in Chapter 8. A brief introduction to dynamical systems is given in
Chapter 9. The fundamental concepts of chaos are introduced through the lo-
gistic map, slider-block models, the Lorenz equations, mantle convection,
and the Rikitake dynamo in the next five chapters. The renormalization
group method is introduced in Chapter 15 and self-organized criticality is
considered in Chapter 16.
Problems are included so that this book can be used as a higher-level un-
dergraduate text or a graduate text, depending upon the background of the
x PREFACE
students and the material used. Little mathematical background is required

for the introduction to self-similar fractals and chaos that includes Chapters
1-6 and 10-1 1. The treatment of self-affine fractals in Chapters 7-8 requires
some knowledge of spectral techniques. Chapters 9 and 12-13 require a
knowledge of differential equations.
I would like to dedicate this book to the memory of Ted Flinn. Until his
untimely death in 1989 Ted was Chief of the Geodynamics Branch at NASA
Headquarters. In this position, over a period of ten years, he supervised a
program that changed routine centimeter-level geodetic position measure-
ments from a dream to a reality. Ted also had the foresight to support re-
search on fractals and chaos applied to crustal dynamics at a time when these
subjects were anything but popular. In particular his enthusiasm was instru-
mental in a conference on earthquakes, fractals, and chaos held at the Asilo-
mar conference facility in January 1989. This conference established a
dialogue between physicists, applied mathematicians, and seismologists fo-
cused on the applications of dynamical systems to earthquake prediction.
I would also like to acknowledge extensive discussions with John Run-
dle, Charlie Sammis, Chris Barton, and Per Bak. Chapter 8 is largely the re-
sult of a collaboration with Bill Newman. And this book could not have been
completed without the diligent manuscript preparation of Maria Petricola.
PREFACETOTHE
SECOND EDITION
A large number of new results on fractals, chaos, and self-organized critical-

ity applied to problems in geology and geophysics have appeared since the
first edition of this book was published in 1992. Evidence for this comes
from the large number of new references included in this edition, increasing
from 160 to over 500. Because of the rapid advances in knowledge it became
evident shortly after the publication of the first edition that a second edition
would be required in a few years.
A large number of additions and some deletions have been made in
preparing this second edition. In Chapter 2 a comprehensive treatment of the
completeness of the sedimentary record has been used to introduce the appli-
cation of fractal techniques to geological problems. To make this textbook
more complete, a brief introduction to probability and statistics has been in-
cluded in Chapter 3. Chapter 4 on seismicity and tectonics has been exten-
sively revised to include the work that has been recently carried out on the
spatial distributions of earthquakes and fractures. In Chapter 5 the elegant
work by Claude Allkgre and his associates explaining power-law (fractal)
distributions of mineral deposits has been added.
A major addition to the second edition is the comprehensive treatment of
multifractals in Chapter 6. Also added to this chapter on fractal clustering
are pair-correlation techniques and lacunarity. One of the most extensive re-
visions concerns the treatment of self-affine fractals in Chapter 7. An intro-
ductory section on time series has been added as well as deterministic exam-
ples of self-affine fractals. Additional techniques for generating fractional
Gaussian noises and fractional Brownian walks have been added as well as a
treatment of rescaled-range (RIS) analyses.
Chapter 8 on geomorphology is almost entirely new. An in-depth treat-
ment of fractal trees has been added, including the Tokunaga taxonomy of
quantifying side branching. Also added to this chapter are treatments of
growth models, models for erosion and deposition, floods, and wavelet fil-
tering techniques. Chapters 9-14 remain essentially unchanged as there have
xii PREFACE TO THE SECOND EDITION
been relatively few new advances concerning the fundamental aspects of

chaos.
A major addition to Chapter 15 on the renormalization group method has
been the introduction of log-periodic behavior. If the fractal exponent is
complex, log-periodic behavior is found. Applications of log-periodic con-
cepts may be one of the most promising avenues for future development.
Chapter 15 on self-organized criticality has been extensively revised. Al-
though the entire concept of self-organized criticality is controversial, the
models developed appear to cross the gap between the chaotic behavior of
low-order systems and the complex, often fractal behavior of high-order sys-
tems. The development of forest-fire models is of particular interest and a
section on these has been added. Systems that exhibit self-organized critical
behavior can satisfy Maxwell-Boltzmann statistics and thus are closely
linked to classical statistical mechanics. Extensive discussions and collabo-
rative work with John Rundle have been particularly valuable in developing
the sections in Chapter 16 on this subject.
Again my graduate students Bruce Malamud, Jon Pelletier, Gleb
Morein, Algis Kucinskas, Lesley Greene, Lygia Gomes Da Silva, and Kirk
Haselton have made major contributions to the preparation of this second
edition. Important contributions were also made by Galena Narkounskaia
and Andrei Gabrielov, visiting scientists from the Institute for Earthquake
Prediction and Theoretical Geophysics in Moscow. The director of the Insti-
tute, Volodya Keilis-Borok, has been an inspiration in providing connections
between dynamical systems and earthquake prediction. His Institute brings a
level of mathematics to geophysical problems that is not available in the
United States.
Once again I would like to acknowledge extensive discussions and col-
laborative research with Bill Newman, John Rundle, Chris Barton, Charlie
Sammis, and Claude Allkgre. This second edition could not have been com-
pleted without the diligent manuscript preparation of Marilyn Grant and Sue
Peterson and the figure preparation of Teresa Howley.
Chapter One
SCALE INVARIANCE
A stone, when it is examined, will be found a mountain in miniature. The

fineness of Nature's work is so great, that, into a single block, a foot or two
in diameter, she can compress as many changes of form and structure, on a
small scale, as she needs for her mountains on a large one; and, taking moss
for forests, and grains of crystal for crags, the sur$ace of a stone, in by far
the plurality of instances, is more interesting than the surface of an ordinary
hill; more fantastic in form, and incomparably richer in colour - the last
quality being most noble in stones of good birth (that is to say, fallen from
the crystalline mountain ranges).
J. Ruskin, Modern Painters, Vol. 5, Chapter 18 (1860)
The scale invariance of geological phenomena is one of the first concepts

taught to a student of geology. It is pointed out that an object that defines the
scale, i.e., a coin, a rock hammer, a person, must be included whenever a
photograph of a geological feature is taken. Without the scale it is often im-
possible to determine whether the photograph covers 10 cm or 10 km. For
example, self-similar folds occur over this range of scales. Another example
would be an aerial photograph of a rocky coastline. Without an object with a
characteristic dimension, such as a tree or house, the elevation of the photo-
graph cannot be determined. It was in this context that Mandelbrot (1967)
introduced the concept of fractals. The length of a rocky coastline is ob-
tained using a measuring rod with a specified length. Because of scale in-
variance, the length of the coastline increases as the length of the measuring
rod decreases according to a power law; the power determines the fractal di-
mension of the coastline. It is not possible to obtain a specific value for the
length of a coastline, owing to all the small indentations down to a scale of
millimeters or less.
Many geological phenomena are scale invariant. Examples include
frequency-size distributions of rock fragments, faults, earthquakes, volcanic
eruptions, mineral deposits, and oil fields. A fractal distribution requires that
2 SCALE INVARIANCE
the number of objects larger than a specified size has a power-law depen-
dence on the size. The empirical applicability of power-law statistics to geo-
logical phenomena was recognized long before the concept of fractals was
conceived. A striking example is the Gutenberg-Richter relation for the fre-
quency-magnitude statistics of earthquakes (Gutenberg and Richter, 1954).
The proportionality factor in the relationship between the number of earth-
quakes and earthquake magnitude is known as the b-value. It has been rec-
ognized for nearly 50 years that, almost universally, b = 0.9. It is now ac-
cepted that the Gutenberg-Richter relationship is equivalent to a fractal
relationship between the number of earthquakes and the characteristic size
of the rupture; the value of the fractal dimension D is simply twice the
b-value; typically D = 1.8 for distributed seismicity.
Power-law distributions are certainly not the only statistical distribu-
tions that have been applied to geological phenomena. Other examples in-
clude the normal (Gaussian) distribution and the log-normal distribution.
However, the power-law distribution is the only distribution that does not in-
clude a characteristic length scale. Thus the power-law distribution must be
applicable to scale-invariant phenomena. If a specified number of events are
statistically independent, the central-limit theorem provides a basis for the
applicability of the Gaussian distribution. Scale invariance provides a ratio-
nal basis for the applicability of the power-law, fractal distribution. Fractal
concepts can also be applied to continuous distributions; an example is
topography. Mandelbrot (1982) has used fractal concepts to generate syn-
thetic landscapes that look remarkably similar to actual landscapes. The
fractal dimension is a measure of the roughness of the features. The earth's
topography is a composite of many competing influences. Topography is
created by tectonic processes including faulting, folding, and flexure. It is
modified and destroyed by erosion and sedimentation. There is considerable
empirical evidence that erosion is scale invariant and fractal; a river network
is a classic example of a fractal tree. Topography often appears to be com-
plex and chaotic, yet there is order in the complexity. A standard approach to
the analysis of a continuous function such as topography along a linear track
is to determine the coefficients An in a Fourier series as a function of the
wavelength An. If the amplitudes A,, have a power-law dependence on wave-
length An,a fractal distribution may result. For topography and bathymetry it
is found that, to a good approximation, the Fourier amplitudes are propor-
tional to the wavelengths. This is also true for a Brownian walk, which can
be generated by the random walk process as follows. Take a step forward
and flip a coin; if tails occurs take a step to the right and if heads occurs take
a step to the left; repeat the process. The divergence of the walk or signal in-
creases in proportion to the square root of the number of steps. A spectral
analysis of the random walk shows that the Fourier coefficients A,, are pro-
portional to wavelength A,,.
SCALE INVARIANCE 3
Many geophysical data sets have power-law spectra. These include sur-
face gravity and magnetics as well as topography. Since power-law spectra
are defined by two quantities, the amplitude and the slope, these quantities
can be used to carry out textural analyses of data sets. The fractal structure
can also be used as the basis for interpolation between tracks where data
have been obtained. A specific example is the determination of the three-
dimensional distribution of porosity in an oil reservoir from a series of well
logs from oil wells.
The philosophy of fractals has been beautifully set forth by their inven-
tor Benoit Mandelbrot (Mandelbrot, 1982). A comprehensive treatment of
fractals from the point of view of applications has been given by Feder
(1988). Vicsek (1992) has also given an extensive treatment of fractals em-
phasizing growth phenomena. Kaye (1989, 1993) covers a broad range of
fractal problems emphasizing those involving particulate matter. Korvin
(1992) has considered many fractal applications in the earth sciences.
Although fractal distributions would be useful simply as a means of
quantifying scale-invariant distributions, it is now becoming evident that
their applicability to geological problems has a more fundamental basis.
Lorenz (1963) derived a set of nonlinear differential equations that approxi-
mate thermal convection in a fluid. This set of equations was the first to be
shown to exhibit chaotic behavior. Infinitesimal variations in initial condi-
tions led to first-order differences in the solutions obtained. This is the defi-
nition of chaos. The equations are completely deterministic; however, be-
cause of the exponential sensitivity to initial conditions, the evolution of a
chaotic solution is not predictable. The evolution of the solution must be
treated statistically and the applicable statistics are often fractal. A compre-
hensive study of problems in chaos has been given by Schuster (1995).
The most universal example of chaotic behavior is fluid turbulence. It
has long been recognized that turbulent flows must be treated statistically
and that the appropriate spectral statistics are fractal. Since the flows in the
earth's core that generate the magnetic field are expected to be turbulent, it is
not surprising that they are also chaotic. The random reversals of the earth's
magnetic field are a characteristic of chaotic behavior. In fact, solutions of a
parameterized set of dynamo equations proposed by Rikitake (1958) exhib-
ited spontaneous reversals and were subsequently shown to be examples of
deterministic chaos (Cook and Roberts, 1970).
Recursion relations can also exhibit chaotic behavior. The classic exam-
ple is the logistic map studied by May (1976). This simple quadratic relation
has an amazing wealth of behavior. As the single parameter in the equation is
varied, the period of the recursive solution doubles until the solution be-
comes fully chaotic. The Lyapunov exponent is the quantitative test of
chaotic behavior; it is a measure of whether adjacent solutions converge or
diverge. If the Lyapunov exponent is positive, the adjacent solutions diverge
4 SCALE INVARIANCE
and chaotic behavior results. The logistic map and similar recursion rela-
tions are applicable to population dynamics and other ecological problems.
The logistic map also produces fractal sets.
Slider-block models have long been recognized as a simple analog for
the behavior of a fault. The block is dragged along a surface with a spring
and the friction between the surface and the block can result in the stick-slip
behavior that is characteristic of faults. Huang and Turcotte (1990a) have
shown that a pair of slider blocks exhibits chaotic behavior in a manner that
is totally analogous to the chaotic behavior of the logistic map. The two
slider blocks are attached to each other by a spring and each is attached to a
constant-velocity driver plate by another spring. As long as there is any
asymmetry in the problem, for example, nonequal block masses, chaotic be-
havior can result. This is evidence that the deformation of the crust associ-
ated with displacements on faults is chaotic and, thus, is a statistical process.
This is entirely consistent with the observation that earthquakes obey fractal
statistics.
Nonlinearity is a necessary condition for chaotic behavior. It is also a
necessary condition for scale invariance and fractal statistics. Historically
continuum mechanics has been dominated by the applications of three linear
partial differential equations. They have also provided the foundations of
geophysics. Outside the regions in which they are created, gravitational
fields, electric fields, and magnetic fields all satisfy the Laplace equation.
The wave equation provides the basis for understanding the propagation of
seismic waves. And the heat equation provides the basis for understanding
how heat is transferred within the earth. All of these equations are linear and
none generates solutions that are chaotic. Also, the solutions are not scale in-
variant unless scale-invariant boundary conditions are applied.
Two stochastic models that exhibit fractal statistics in a variety of ways
are percolation clusters (Stauffer and Aharony, 1992) and diffusion-limited
aggregation (DLA) (Vicsek, 1992). In defining a percolation cluster a two-
dimensional grid of square boxes can be considered. The probability that a
site is permeable p is specified, and there is a sudden onset of flow through
the grid at a critical value of this probability, pc = 0.59275. This is a critical
point and there are a variety of fractal scaling laws valid at and near the crit-
ical point. There is observational evidence that distributed seismicity has a
strong similarity to percolation clusters.
In generating a diffusion-limited aggregation a two-dimensional grid of
square boxes can again be considered. A seed cell is placed in one of the
boxes. Additional cells are added randomly and follow a random-walk path
from box to box until they accrete to the growing cluster of cells by entering
a box adjacent to the growing cluster. A sparse dendritic structure results be-
cause the random walkers are more likely to accrete near the tips of a cluster
rather than in the deep interior. The resulting cluster satisfies fractal statistics
SCALE INVARlANCE 5
in a variety of ways. Diffusion-limited aggregation has been applied to the

dendritic growth of minerals and to the growth of drainage networks.
The renormalization group method can be applied to a wide variety of
problems that exhibit scale invariance. A relatively simple system is mod-
eled at the smallest scale; the problem is then renormalized (rescaled to uti-
lize the same simple model at the next larger scale). The process is repeated
at larger and larger scales. The method can be applied to the analysis of per-
colation clusters and to model fragmentation, fracture, the concentration of
economic ore deposits, and other problems that satisfy fractal statistics.
The concept of self-organized criticality was introduced by Bak et al.
(1988) in terms of a cellular automata model for avalanches on a sand pile. A
natural system is said to be in a state of self-organized criticality if, when
perturbed from this state, it evolves naturally back to the state of marginal
stability. In the critical state there is no natural length scale so that fractal
statistics are applicable. In the sand-pile model the frequency-magnitude
statistics of the sand avalanches are fractal.
Regional seismicity is taken to be a classic example of a self-organized
critical phenomena. Stress is added continuously by the movement of the
surface plates of plate tectonics. The stress is dissipated in earthquakes with
fractal frequency-magnitude statistics. Scholz (1991) argues that the earth's
crust is in a state of self-organized criticality; he cites as evidence the obser-
vation that induced seismicity occurs whenever the reservoir behind a large
dam is filled. This is evidence that the earth's crust is on the brink of failure.
As discussed above, a pair of interacting slider blocks can exhibit
chaotic behavior. Large numbers of driven slider blocks are a classic exam-
ple of self-organized critica!ity. A two-dimensional array of slider blocks is
considered. Each block is attached to its four neighbors and to a constant ve-
locity driver plate by springs. Slip events occur chaotically and the fre-
quency-size statistics of the events are generally fractal. By increasing the
number of blocks considered, the low-order chaotic system is transformed
into a high-order system that exhibits self-organized criticality.
ChapterTwo
DEFINITION OF A
FRACTAL SET
2.1 Deterministic fractals
Since its original introduction by Mandelbrot (1967), the concept of fractals

has found wide applicability. It has brought together under one umbrella a
broad range of preexisting concepts from pure mathematics to the most em-
pirical aspects of engineering. It is not clear that a single mathematical defi-
nition can encompass all these applications, but we will begin our quantita-
tive discussion by defining a fractal set according to
where Ni is the number of objects (i.e., fragments) with a characteristic lin-

ear dimension ri, C is a constant of proportionality, and D is the fractal di-
mension. The fractal dimension can be an integer, in which case it is equiva-
lent to a Euclidean dimension. The Euclidean dimension of a point is zero, of
a line segment is one, of a square is two, and of a cube is three. In general,
the fractal dimension is not an integer but a fractional dimension; this is the
origin of the term fractal.
We now illustrate why it is appropriate to refer to D as a fractal or frac-
tional dimension by using a line segment of unit length. Several examples
of fractals are illustrated in Figure 2.1. In Figure 2.1(a) the line segment of
unit length at zero order is divided into two parts at first order so that r, = $;
one part is retained so that N, = 1. The remaining segment is then divided
1
into two parts at second order so that r, = 4; again one part is retained so that
N, = 1. To determine D, (2.1) can be written as
DEFINITION OF A FRACTAL SET 7
where In is a logarithm to the base e and log is a logarithm to the base 10. In
almost all applications we will require the ratio of logarithms; in this case
Figure 2.1. Illustration of
the result is the same if the logarithm to the base e (In) is used or if the loga- six one-dimensional fractal
rithm to the base 10 (log) is used. For the example considered in Figure constructions. At zero order
2.1 (a), In (N2/N,) = In 1 = 0, ln(r,lr,) = In 2, and D = 0, the Euclidean dimen- a line segment of unit length
sion of a point. This construction can be extended to higher and higher or- is considered. At first order
ders, but at each order i, i = 1,2, . . . , n, we have In (Ni+,INi)= In 1 = 0. As the line segment is divided
the order approaches infinity, n + =, the remaining line length approaches into an integer number of
equal-sized smaller segments
zero, rn + 0, becoming a point. Thus the Euclidean dimension of a point, and a fraction of these
zero, is appropriate. The construction illustrated in Figure 2.l(b) is similar segments is retained. The
except that the line segment of unit length at zero order is divided into three first-order fractal acts as a
order r, = 31 and again N2 = 1. Thus as the order is increased and n + the

construction again tends to a point and D = 0.
-
parts at first order so that r, = f ;one part is retained so that N, = 1. At second generator for higher-order
fractals. Each of the retained
line segments at first order is
further divided into smaller
In Figure 2.1 (c) the zero-order line segment of unit length is divided into segments using the generator
two parts but both are retained at first order so that r, = $ and N, = 2. The to create a second-order
fractal. The first two orders
process is repeated at second order so that r, = and N, = 4. From (2.2) we
are illustrated but the
see that D = In 2/ln 2 = 1. Similarly for Figure 2.l(d) we have D = 1; in both construction can be carried
cases the fractal dimension is the Euclidean dimension of a line segment. to any order desired. (a) A
This is appropriate since the remaining segment will be a line segment of line segment is divided into
unit length as the construction is repeated. However, not all constructions two parts and one is retained;
will give integer fractal dimensions; two examples are given in Figures D = In lfln 2 = 0 (fractal
dimension of a point). (b)
2.1 (e) and 2.1(f). In Figure 2.1 (e) the zero-order line segment of unit length A line segment is divided
is divided into three parts at first order so that r, = 31 ; the two end segments into three parts and one is
are retained so that N, = 2. The process is repeated at second order so that retained; D = In lfln 3 = 0
(fractal dimension of a
point). (c) A line segment is
divided into two parts and
both are retained; D = In 21111
2 = 1 (fractal dimension of a
line). (d) A line segment is
divided into three parts and
all three are retained; D = In
3nn 3 = 1 (fractal dimension
of a line). (e) A line segment
is divided into three parts
and two are retained; D = In
2fln 3 = 0.6309 (noninteger
fractal dimension; this
construction is also known
as a Cantor set). (f) A line
segment is divided into five
parts and three are retained;
D = In 3/ln 5 = 0.6826
(noninteger fractal
dimension).
DEFINITION OF A FRACTAL SET
r2 = 91 and N2 = 4. From (2.2) we find that D = In 2An 3 = 0.6309. This is

known as a Cantor set and has long been regarded by mathematicians as a
pathological construction. In Figure 2.l(f) the zero-order line segment is di-
4;
vided into five parts at first order so that r, = the two end segments and the
center segment are retained, giving N, = 3. The process is repeated at second
9
order so that r2 = and N2 = 9. From (2.2) we find that D = In 31ln 5 =
0.6826. These two examples have fractal dimensions between the limiting
cases of zero and one; thus they have fractional dimensions. Constructions
can be devised to give any fractional dimension between zero and one using
the method illustrated in Figure 2.1.
The iterative process illustrated in Figure 2.1 can be carried out as often
as desired, making the remaining line lengths shorter and shorter. The con-
structions given in Figure 2.1 are scale invariant. Scale invariance is a neces-
sary condition for the applicability of (2.1) since no natural length scale en-
ters this power-law (fractal) relation. As a particular example consider the
Cantor set illustrated in Figure 2.l(e). Iterations of the Cantor set up to fifth
order, i = 5, are illustrated in Figure 2.2. The first-order Cantor set is used as
a "generator" for higher-order sets. Each of the two remaining line segments
at first order is replaced by a scaled-down version of the generator to obtain
the second-order set, and so forth at higher orders. If n iterations are carried
out, then the line length at the nth iteration, rn, is related to the length at the
first iteration, r,, by rJr, = (r,lr,)". Thus, as n + =, rn + 0; in this limit the
Cantor set illustrated in Figure 2.2 is known as a Cantor "dust," an infinite
set of clustered points. The repetitive iteration leading to a dust is known as
"curdling."
0 order
1 st order
2 nd order
Figure 2.2. Illustration of the

3 rd order
Cantor set carried to fifth
order. The first-order Cantor
set acts as a generator; the
straight-line segments at 4 th order
order i are replaced by the

generator to obtain the set at
order i + 1. 5 th order
The fractal concepts applied above to a line segment can also be applied
to a square. A series of examples is given in Figure 2.3. In each case the zero-
order square is divided into nine squares at first order each with r , = f . At
second order the remaining squares are divided into nine squares each with
6,
r2 = and so forth. In Figure 2.3(a) only one square is retained, so that N , =
N2 = . . . = Nn = 1. From (2.2) D = 0,which is the Euclidean dimension of a
point; this is appropriate since as n + w the remaining square will become a
point. In Figure 2.3(b) two squares are retained at first order so that r , = 5,
4,
N , = 2 and at second order r2 = N2 = 4 . Thus from (2.2), D = In 2/ln 3 =
0.6309, the same result that was obtained from Figure 2.l(e), as expected.
Similarly, in Figure 2.3(c) three squares are retained at first order so that
r , = f , N , = 3, and at second order r2 = $, N2 = 9; thus D = In 3An 3 = 1. In the
limit n -+ = the remaining squares will become a line as in Figure 2.l(d).
The Euclidean dimension of a line is found. In Figure 2.3(d), only the center
3
square is removed; thus at first order r , = , N , = 8, and at second order r2 =
4, N2 = 64. From (2.2) we have D = In 81ln 3 = 1.8928. This construction is
known as a Sierpinski carpet. In Figure 2.3(e) all nine squares are retained;
thus at first order r , = 51 , N , = 9, and at second order r, = 9,
1
N2 = 81. From
(2.2) we have D = In 91ln 3 = 2. This is the Euclidean dimension of a square Figure 2.3. Illustration of
and is appropriate because when we retain all the blocks we continue to re- five two-dimensiona1
constructions. At zero order
a square of unit area is
considered. At first order the
unit square is divided into
nine equal-sized smaller
squares with r , = 51 and a
fraction of these squares
is retained. The first-order
fractal acts as a generator
for higher-order fractals.
Each of the retained squares
at first order is divided
into smaller sauares using -
the generator to create a
second-order fractal. The
1
first two orders with r, = 5
1
and r, = g illustrated
but the construction can be
carried to any order desired.
(a)N,=l,N,=l,D=lnlAn
3 = 0. (b) N, = 2, N, = 4, D =
In 2/ln 3 = 0.6309. (c) N, = 3,
N 2 = 9 , D = I n 3 A n 3 = l.(d)
N, = 8, N2 = 64, D = In 81111 3
= 1.8928 (known as a
Sierpinski carpet). (e) N, = 9,
N2=81,D=ln9/ln3=2.
10 DEFINITION OF A FRACTAL SET
tain the unit square at all orders. Iterative constructions can be devised to
yield any fractal dimensions between 0 and 2; again each construction is
scale invariant.
The examples for one and two dimensions given in Figures 2.1 and 2.3
can be extended to three dimensions. Two examples are given in Figure 2.4.
The Menger sponge is illustrated in Figure 2.4(a). A zero-order solid cube of
unit dimensions has square passages with dimensions r , = f cut through the
centers of the six sides. At first order six cubes in the center of each side are
removed as well as the central cube. Twenty cubes with dimensions r , = re- 3
main so that N, = 20. At second order the remaining 20 cubes have square
passages with dimensions r2 = ) cut through the centers of their six sides. In
each case the six cubes in the centers of each side are removed as well as the
center cube. Four hundred cubes with r2 = $ remain so that N2 = 400. From
(2.2) we find that D = In 201ln 3 = 2.7268. The Menger sponge can be used as
a model for flow in a porous media with a fractal distribution of porosity.
Another example of a three-dimensional fractal construction is given in Fig-
ure 2.4(b). Again the unit cube is considered at zero order. At first order it is
divided into eight equal-sized cubes with r , = $, and two diagonally opposite
comer cubes are removed so that N , = 6. At second order each of the remain-
1
ing six cubes are divided into eight equal-sized smaller cubes with r, = 2. In
each case two diagonally opposite corner cubes are removed so that N2 = 36.
From (2.2) we find that D = In 61ln 2 = 2.585. We will use this configuration
for a variety of applications in later chapters. Iterative constructions can be
devised to yield any fractal dimension between 0 and 3; again each construc-
two three-dimensional tion is scale invariant.
fractal constructions. (a) At The examples given above illustrate how geometrical constructions can
first order the unit cube is give noninteger, non-Euclidean dimensions. However, in each case the
divided into 27 equal-sized
smaller cubes with r, = i,
20 cubes are retained so that
N, = 20. At second order r, =
and 400 out of 729 cubes
are retained so that N, = 400;
D = In 20lln 3 = 2.727. This
construction is known as the
Menger sponge. (b) At first
order the unit cube is divided
into eight equal-sized
smaller cubes with r, = i.
Two diagonally opposite
cubes are removed so that six
cubes are retained and N, =
a
6. At second order r, = and
36 out of 64 cubes are
retained so that N, = 36;
D = In 6fln 2 = 2.585.
structure is not continuous. An example of a continuous fractal construction

is the triadic Koch island illustrated in Figure 2.5. At zero order this con-
struction starts with an equilateral triangle with three sides of unit length,
No = 3, r, = 1. At first order equilateral triangles with sides of length r , = 51
are placed in the center of each side; there are now 12 sides so that N, = 12.
This construction is continued to second order by placing equilateral trian-
gles of length r2 = $ in the center of each side; there are 48 sides so that N, =
48. From (2.2) we have D = In 4Iln 3 = 1.26186. The fractal dimension is be-
tween one (the Euclidean dimension of a line) and two (the Euclidean di-
mension of a surface). This construction can be continued to infinite order;
the sides are scale invariant, and a photograph of a side is identical at all
scales. To quantify this we consider the length of the perimeter. The length
of the perimeter Pi of a fractal island is given by
where ri is the side length at order i and N is the number of sides. Substitu-
tion of (2. I) gives
For the triadic Koch island illustrated in Figure 2.5 we have Po = 3, P I = 4,

and P2 = = 5.333. Taking the logarithm of (2.4) and substituting these val-
ues we find that
D = 1 + log(P;+ ,/pi) = 1 + log(413) = 1 + log4 - log3 - -

log4
(2.5)
l o d r j r ;+ I ) log3 log3 log3 Figure 2.5. The triadic Koch
island. (a) An equilateral
This is the same result that was obtained above using (2.2), as expected. The triangle with three sides
perimeter of the triadic Koch island increases as i increases. As i approaches of unit length. @)Three
infinity, the length of the perimeter also approaches infinity, as indicated by F les with sides of length
r, = J are placed in the center
(2.4), since D >I (D is greater than unity). The perimeter of the triadic Koch of each side The is
island in the limit i -+ is continuous but is not differentiable. now made up of 12 sides and
N, = 12. (c) Twelve triangles
4
with sides of length r, = are
placed in the center of each
side. The perimeter is now
made up of 48 sides and N2=
48; D = In 4fln 3 = 1.26186.
The length of the perimeter
in (a) is Po = 3, in (b) is PI =
16
4, and in (c) is P2 = y =
5.333.
2.2 Statistical fractals
The triadic Koch island can be considered to be a model for measuring the
length of a rocky coastline. However, there are several fundamental differ-
ences. The primary difference is that the perimeter of the Koch island is de-
terministic and the perimeter of a coastline is statistical. The perimeter of the
Koch island is identically scale invariant at all scales. The perimeter of a
rocky coastline will be statistically different at different scales but the differ-
ences do not allow the scale to be determined. Thus a rocky coastline is a sta-
tistical fractal. A second difference between the triadic Koch island and a
rocky coastline is the range of scales over which scale invariance (fractal
behavior) extends. Although a Koch island has the maximum scale of the
origin triangle, the construction can be extended over an infinite range of
scales. A rocky coastline has both a maximum scale and a minimum scale.
The maximum scale would typically be 103 to 104 km, the size of the conti-
nent or island considered. The minimum scale would be the scale of the
grain size of the rocks, typically 1 mm. Thus the scale invariance of a rocky
coastline could extend over nine orders of magnitude. The existence of both
upper and lower bounds is a characteristic of all naturally occurring fractal
systems. In addition, the scale invariance of a coastline will be only approxi-
mately scale invariant (fractal), and there will be statistical fluctuations in
any measure of fractality. On the other hand, the triadic Koch island is ex-
actly scale invariant (fractal).
Mandelbrot (1967) introduced the concept of fractals by using (2.4) to
determine the fractal dimension of the west coast of Great Britain. The
length of the coastline Pi was determined for a range of measuring rod
lengths ri. Mandelbrot (1967) used measurements of the length of the coast-
line obtained previously by Richardson (1961). Taking a map of a coastline,
the length is obtained by using dividers of different lengths ri. Using the
scale of the map, the length of the coastline is plotted against the divider
length on log-log paper. If the data points define a straight line, the result is
a statistical fractal. The result for the west coast of Great Britain is given in
Figure 2.6. As shown, the data correlate well with (2.4), taking D = 1.25.
This is evidence that the coastline is a fractal and is statistically scale invari-
ant over this range of scales.
The technique for obtaining the fractal dimension of a coastline is easily
extended to any topography. Contour lines on a topographic map are entirely
equivalent to coastlines; the lengths along specified contours Piare obtained
using dividers of different lengths ri.The fractal relation (2.4) is generally a
good approximation and fractal dimensions can be obtained. As illustrated in
Figure 2.7, the fractal dimensions of topography using the ruler (divider)
+
method are generally in the range D = 1.20 0.05 independent of the tec-
tonic setting and age. Topography is primarily a result of erosional pro-

cesses; however, in young terrains topography is being created by active tec-
tonic processes. It is not surprising that many of these processes are scale in-
variant and generate fractal topography. An interesting question, however, is
whether erosional processes and tectonic processes each generate topogra-
phies with about the same fractal dimension. Bruno et al. (1992, 1994) and
Gaonach et al. (1992) showed that the perimeters of basaltic lava flows are
also fractal with D = 1.12- 1.42. Details on the use of the ruler (divider)
method have been given by Andrle (1992).
It should be emphasized that not all topography is fractal (Goodchild,
1980). Young volcanic edifices are one example. Until modified by erosion,
both shield and strata volcanoes are generally conical in shape and do not
yield well-defined fractal dimensions. Alluvial fans are another example of a
nonfractal geomorphic feature. The morphology of alluvial fans can be mod-
eled using the heat equation (Culling, 1960). Because the heat equation is
linear, it contains a characteristic length (or time) and cannot give solutions
that are scale invariant (fractal). The heat equation can also be used to model
the elevation of mid-ocean ridges. The morphology of ocean trenches can be
modeled by considering the bending of the elastic lithosphere. Again, the
equation governing flexure is linear, introducing a characteristic length, and
solutions are not scale invariant (fractal). However, despite these exceptions,
most of the earth's topography and bathymetry are best modeled using frac-
tal statistics and are therefore scale invariant.
Figure 2.6. Length P of the

west coast of Great Britain as
a function of the length r of
the measuring rod; data from
Mandelbrot (1967). The data
are correlated with (2.4) using
D = 1.25.
Although the ruler (divider) method was the first used to obtain fractal
dimensions, it is not the most generally applicable method. The box-count-
ing method has a much wider range of applicability than the ruler method
(Pfeiffer and Obert, 1989). For example, it can be applied to a distribution of
points as easily as it can be applied to a continuous curve. We now use the
Figure 2.7. The lengths P

of specified topographic
contours in several mountain
belts are given as functions
of the length r of the
measuring rod. (a) 3000 ft
contour of the Cobblestone
Mountain quadrangle,
Transverse Ranges,
California (D = 1.21); (b)
5400 ft contour of the Tatooh
Buttes quadrangle, Cascade
Mountains, Washington
(D = 1.21); (c) 10,000 ft
contour of the Byers Peak
quadrangle, Rocky
Mountains, Colorado (D =
1.15); (d) 1000 ft contour of
the Silver Bay quadrangle,
Adirondack Mountains,
New York (D = 1.19). The
straight-line correlations in
these log-log plots are with
(2.4).
box-counting method to determine the fractal dimension of a rocky coast-

line. As a specific example we consider the coastline in the Deer Island,
Maine, quadrangle illustrated in Figure 2.8(a). The coastline is overlaid with
a grid of square boxes; grids of different-size boxes are used. The number of
boxes Ni of size ri required to cover the coastline is plotted on log-log paper
as a function of ri. If a straight-line correlation is obtained, then (2.2) is used
to obtain the applicable fractal dimension. The box-counting method for the
coastline given in Figure 2.8(a) is illustrated in Figures 2.8(b) and 2.8(c).
The shaded areas are the boxes required to cover the coastline. In Figure
2.8(b) we require 98 boxes with r = 1 km to cover the coastline; in Figure
2.8(c) we require 270 boxes with r = 0.5 km to cover the coastline. The re-
sults for a range of box sizes are given in Figure 2.9. The correlation with
(2.2) yields D = 1.4. This is somewhat higher than the values given above for
other examples. But this is due to the extreme roughness of the coastline
used in this example. When the ruler method is applied to this coastline, the
same fractal dimension is found.
The statistical number-size distribution for a large number of objects
can also be fractal. A specific example is rock fragments. For the distribution
to be fractal, the number of objects N with a characteristic linear dimension
greater than r should satisfy the relation
where D is again the fractal dimension. It is appropriate to use this cumula-

tive relation rather than the set relation (2.1) when the distribution takes on a
continuous rather than a discrete set of values. Another example where (2.6)
is applicable is the frequency-magnitude distribution of earthquakes.
As a statistical representation of a natural phenomenon, (2.6) will be only
approximately applicable, with both upper and lower bounds to the range of
applicability. A specific example of the applicability of (2.6) is the Korcak
(1 940) empirical relation for the number of islands on the earth with an area
greater than a specified value. Taking the characteristic length to be the
square root of the area of the island, Mandelbrot (1975) showed that (2.6) is
a good approximation with D = 1.30. The worldwide frequency-size distri-
bution of lakes is given in Figure 2.10 (Meybeck, 1995). The cumulative
number of lakes N with an area A greater than a specified value is given as a
function of both area A and the square root of the area r. An excellent corre-
lation is obtained with (2.6) taking D = 1.go. There is a considerable regional
variation in this result; Kent and Wong (1982) applied the same approach for
the number of lakes in Canada and found a good correlation with (2.6) tak-
ing D = 1.55.
As we discussed, the term fractal dimension stands for fractional dimen-

sion. The meaning of this is clear in Figures 2.1-2.3; however, the meaning
may be less clear in statistical power-law distributions. Some power-law dis-
tributions fall within the limits associated with fractional dimensions, i.e.,
0 < D < 3, but others do not. The question that must be addressed is whether
Figure 2.8. (a) Illustration of

a rocky coastline: the Deer
Island, Maine, quadrangle.
(b) The shaded area contains
the square boxes with r = 1
km required to cover the
coastline; N = 98. (c) The
shaded area contains the
square boxes with r = 0.5 km
required to cover the
coastline; N = 270.
all power-law distributions that satisfy (2. I ) or (2.6) are fractal. In this book
we define them to be fractal. Such distributions are clearly scale invariant,
even if not directly associated with a fractal dimension. This choice elimi-
nates an ambiguity that can lead to considerable confusion when addressing
measured data sets. We will continually address this question as we consider
specific applications.
Figure 2.9. The number N

of square boxes required to
cover the coastline in Figure
2.8(a) as a function of the
box size r. The correlation
with (2.1) yields D = 1.4.
Figure 2.10. Worldwide

frequency-size distribution
of lakes. The cumulative
number of lakes N with
an area A greater than a
specified value is given as a
function of both area and the
square root of area r. The
straight-line correlation is
with the fractal relation (2.6)
taking D = 1.90.
2.3 Depositional sequences
The relationship between deterministic and statistical fractals will be further

illustrated using a classic problem in geology, the deposition of sediments.
Many mechanisms are associated with the deposition of sediments, and it
certainly can be considered a complex geological process. Gaps in the sedi-
mentary record are recognized on a global basis, and these gaps form the
boundaries of various geological epochs. But these gaps appear at all time
scales and can be attributed to periods dominated by erosion or lack of depo-
sition. We will introduce a simple model, based on fractal concepts, for sedi-
ment deposition. The basis of this model is the devil's staircase. An example
of a devil's staircase based on the third-order Cantor set from Figure 2.2,
given in Figure 2.11, has the same fractal dimension as the Cantor set. In-
stead of removing the middle third of each line segment at each order, it is
retained as a horizontal segment. The vertical segments are equal upward
steps moving from left to right. Taking the total horizontal length to be unity,
there is one horizontal step N, = 1 with a length r , = i, there are two horizon-
;,
tal steps N2 = 2 with a length r2 = and there are four horizontal steps N, = 4
with a length r, = &. Thus from (2.2) we have D = In 2/ln 3 = 0.6309. Note
that there are 24 steps with r, = ii but 16 of these would be further subdi-
vided if the construction was continued to higher order. The devil's staircase
based on the Cantor set can also be obtained as the integral of the Cantor set
from 0 to x.
The devil's staircase has a strong similarity to the age distribution in a
pile of sediments, periods of rapid deposition interspersed with gaps in the
sedimentary record (unconformities). For our simplified model of sediment
deposition, we assume that the rate of subsidence of the earth's crust is a
constant R. Without sediment deposition the water depth yw would increase
linearly with time and would be given by yw = Rt. We further assume that the
Figure 2.11. A fractal devil's

staircase based on the third-
order Cantor set illustrated in
Figure 2.2. The horizontal
step sizes are given by the
Cantor set; the vertical step
sizes are equal. f
sediment supply rate is sufficient to keep the surface of the sediments at sea
level. With this assumption and a constant rate of subsidence R , the rate of
deposition of sediments is also R and the thickness of sediments is ys = Rt.
With this simple model the rate of deposition is constant, and there are no
gaps in the sedimentary record. However, it is well known that sedimentary
sequences are characterized by unconformities (bedding planes), which rep-
resent gaps in the sedimentary record. An unconformity represents a period
of time during which erosion was occurring and/or a period of time during
which no sediment was deposited.
One mechanism for generating sedimentary unconformities is to hypoth-
esize variations in sea level. We will first illustrate how harmonic variations
in sea level with time can generate gaps (unconformities) in the sedimentary
record. Our simple model is illustrated in Figure 2.12. The dashed straight
line in Figure 2.11(a) gives the thickness of sediments ys = Rt with R =
1 m d y r and no variations in sea level. After two million years, t = 2 Myr,
the thickness of sediments is ys = 2 km. Now assume that the variation in sea
level is given by
and we take ysL0 = 400 m and T, = 2 Myr. During the first 500,000 yr sea
level is rising, during the next 1,000,000 yr sea level is falling, and during
the final 500,000 yr sea level is again rising. If no sedimentation was occur-
ring, the depth of water during a cycle 7, would be given by
yw = Rt + y,, sin (2 T tIrJ
and this is the solid line in Figure 2.12(a). We again assume that the rate of
sedimentation is sufficiently high that the actual water depth is zero. At
t = 0 the rate of subsidence is R = 1 mrnlyr; the rate of sea level rise is 1.26
m d y r so that the rate of sediment deposition is 2.26 m d y r . The thickness
of sediments deposited follows the solid curve in Figure 2.12(a). However,
at t = 792,000 yr (point a) the rate of sea level fall becomes equal to the
rate of subsidence. For the period 792,000 < t < 1,208,000 yr (point b) sea
level is falling faster than the subsidence rate. Without erosion the previ-
ously deposited sediments would rise above sea level. We assume, how-
ever, that erosion is sufficiently rapid that the rising landscape is main-
tained at sea level. At t = 1,208,000 yr, 70 m of previously accumulated
sediments have been eroded. The result is an unconformity and a gap in the
sedimentary record. The sediments immediately below the uncomformity
were deposited at t = 577,000 yr (point c) and the sediments immediately
above the unconformity were deposited at t = 1,208,000 yr (point b), a gap
in the sedimentary record results. From t = 1,208,000 yr to t = 2,000,000 yr

sea level is either falling more slowly than the subsidence rate or is rising
so that sedimentation occurs. The entire sequence of deposition is illus-
trated in Figure 2.12a.
At the end of deposition, i.e., at t = 2 Myr, the age of the sediments 7 is
given as a function of depth y in Figure 2.12(b). The age of the sediments at
the base of the sedimentary pile, y = 2 km, is t = 2 Myr. The gap illustrated
in Figure 2.12(a) results in the unconformity illustrated in Figure 2.12(b).
Figure 2.12. Illustration The age of the sediments above the unconformity is 7 = 792,000 yr (point b)
of the development of an and the age of the sediments below the unconformity is T = 1,423,000 yr
unconformity during the (point c).
deposition of sediments. (a)
The thickness of sediments
ys in a sedimentary basin is
given as a function of time t.
The dashed line is the
thickness of sediments with
no sea level change and a
constant rate of subsidence
R = 1 mmlyr. If sea level
varies according to (2.7), the
thickness of sediments is
given by the solid line. From
t = 0 to t = 792,000 yr (point
a) deposition occurs and the
thickness of sediments is
given by (2.8). From t =
792,000 yr to t = 1,208,000
yr (point b) sea level is
falling faster than the rate of
subsidence and erosion is
occumng. Sediments
deposited between t = Age 5 Myr 1
577,000 yr (point c) and t =
792,000 yr (point a) are
eroded, as shown in the
cross-hatched region.
This erosion creates an
unconfonnity; a gap in the
sedimentary record lasts
from t = 577,000 yr (point c)
to t = 1,208,000 yr (point b).
(b) For the model given in
(a) the age of the sediments
T is given as a function of
depth y. The unconfonnity
corresponds to the gap
illustrated in (a).
A simple harmonic variation of sea level will lead to a periodic sequence

of unconformities of equal length. However, there is observational evidence
that variations in sea level obey fractal statistics (Hsui et al., 1993) so that it
should not be surprising that sea level variations could generate a distribu-
tion of unconformities that also obeys fractal statistics. We next develop a
fractal model for sediment deposition based on the devil's staircase given in
Figure 2.1 1 (Plotnick, 1986; Korvin, 1992, pp. 95-113). We consider the
devil's staircase associated with a second-order Cantor set. The age of sedi-
ments in this model is given as a function of depth in Figure 2.13(a). Eight
kilometers of sediments have been deposited in this model sedimentary
basin in a period of 9 Myr so that the mean rate (velocity) of deposition is
= 8 kml9 Myr = 0.89 mrnlyr over this period. However, there is a major
unconformity at a depth of 4 km. The sediments immediately above this un-
conformity have an age T = 3 Myr and the sediments immediately below it
have an age T = 6 Myr. There are no sediments in the sedimentary pile with
ages between T = 3 and 6 Myr. In terms of the Cantor set this is illustrated in
Figure 2.13(b). The line of unit length is divided into three parts, and the
middle third, representing the period without deposition, is removed. The
two remaining parts are placed on top of each other as shown.
During the first three million years of deposition (the lower half of the
sedimentary section) and the last three million years of deposition (the upper
half of the sedimentary section) the mean rates of deposition are = 4 km/3
Myr = 1.33 mdyr. Thus the rate of deposition increases as the period con-
sidered decreases. This is shown in Figure 2.13(c).
Age T, M yr
Depth
Y
km
Figure 2.13. Illustration
of a model for sediment
deposition based on a devil's
staircase associated with a
second-order Cantor set.
(a) Age of sediments T
as a function of depth y.
(b) Illustration of how the
Cantor set is used to
construct the sedimentary
pile. (c) Average rate of
deposition as a function of
the period T considered.
There is also an unconformity at a depth of 2 km. The sediments imme-

diately above this unconformity have an age T = 1 Myr and immediately be-
low have an age T = 2 Myr. Similarly there is an unconformity at a depth of
6 krn;the sediments above this unconformity have an age T = 7 Myr and sed-
iments below an age T = 8 Myr. There are no sediments in the pile with ages
between T = 8 and 7 Myr, between T = 6 and 3 Myr, and between T = 2 and
1 Myr. This is illustrated in Figure 2.13(a). In terms of the Cantor set, Figure
5
2.13(b), the two remaining line segments of length are each divided into
three parts and the middle thirds are removed. The four remaining segments
6
of length are placed on top of each other as shown. During the periods T =
9 to 8, 7 to 6, 3 to 2, and 1 to 0 Myr the rates of deposition are = 2 km/l
Myr = 2 m d y r . This rate is also included in Figure 2.13(c). The rate of dep-
osition clearly has a power-law dependence on the length of the time interval
considered. The results illustrated in Figure 2.13 are based on a second-order
Cantor set, but the construction can be extended to any order desired and the
power-law results given in Figure 2.13(c) would be extended to shorter and
shorter time intervals.
We now generalize the determination of the rate of deposition as a func-
tion of record length and relate it to the fractal dimension of a set. The rate of
deposition Ri for a set of order i is given by
where L, is the thickness of sediments deposited in the period T;. The period
T;in our model is equivalent to the line segment length riin the fractal sets il-
lustrated in Figure 2. l . For the example given in Figure 2.13 we have T,, T, =
3 , T, = ~ d 9The
~ ~ 1and . thickness of sediments Li is given by the number of
segments retained at a specified order N so that
For the example given in Figure 2.13 we have N, = 2 and L, = Ld2 and N, =
4 and L, = Ld4. Noting the equivalence of our T~and ri in the fractal relation
(2. l), we can write the fractal relation
Combining (2.9), (2.10), and (2.1 1) we relate Ri to T~ with the result

which also can be written as
The rate of deposition Ri has a power-law dependence on the time interval

over which the deposition occurs. For the case considered above and illus-
trated in Figure 2.13, we have D = In 2/ln 3 = 0.6309; this is expected since it
is the fractal dimension of a Cantor set. Depending on which set is used to
construct the devil's staircase model for sedimentation, any fractal dimen-
sion between 0 and 1 can be obtained.
The model for sediment deposition given above is deterministic, whereas
actual deposition is clearly stochastic. Nevertheless, the deterministic model
illustrates how fractal variations in sea level or other fractal depositional
mechanisms can give a fractal sedimentary sequence. It is also clear that not
all gaps in the sedimentary record can be attributed to variations in sea level.
Despite these obvious limitations it is of interest to consider the observa-
tional data on rates of sediment deposition. In Figure 2.14 the rates of depo-
sition of fluvial sediments R are given as a function of time span T over
which deposition occurs (Sadler and Strauss, 1990). These data were based
on 5600 deposition rates determined in modern sedimentary basins and an-
cient stratigraphic sections. The straight-line correlation is with (2.13) tak-
ing D = 0.336. A reasonably good correlation is obtained over 8 orders of
magnitude in rate and 12 orders of magnitude in time. Clearly such a correla-
tion is only approximate since important complications have not been con-
sidered. For example, different depositional mechanisms would be expected
to dominate on different time scales. However, the results in Figure 2.14
clearly indicate the strong episodicity of sedimentation and a good correla-
tion with power-law (fractal) statistics.
Gardner et al. (1987) have correlated elevation changes with measured
intervals for rates of tectonic uplift and erosion. In both cases they find good
power-law (fractal) correlations. Snow (1992) has applied the devil's stair-
case model described above to explain their results. For uplift Gardner et al.
(1987) found that the rate of uplift RUis related to the interval T,, by
And from (2.14) we have D = 0.746. For erosion these authors found that the
rate of erosion Re is related to the interval T~ by
And from (2.14) we have D = 0.815.

The devil's staircase model for sedimentary completeness could have
been carried to higher and higher orders and the rates of deposition would
have increased accordingly. Also, the thickness of the layers would have got-
ten smaller and smaller. In the limit of infinite order, the rate of deposition
would be infinity and the layer thickness would be zero, clearly unreason-
able from a physical standpoint. Thus when the mathematical construction
of a Cantor set is applied to a real physical problem, the deposition of sedi-
ments, it is necessary to truncate the model at finite order. Also, it must be
emphasized that the layer thicknesses in an actual sedimentary pile will have
a statistical distribution (a problem we will consider in a later chapter),
rather than having equal thicknesses as in this model. This is a common
problem when a deterministic fractal model (a Cantor set) is applied to a sta-
tistical fractal problem (a layered stack of sediments in a sedimentary pile).
2.4 Why fractal distributions?
In this chapter a number of deterministic fractal sets have been introduced

including the Cantor set, the Sierpinski carpet, the Menger sponge, and the
triadic Koch island. Two geological applications have been considered, the
length of a coastline and the completeness of the sedimentary record. The
Koch triadic island illustrates the fractal statistics of the length of the coast-
line. The devil's staircase based on the Cantor set illustrates the fractal statis-
Figure 2.14. Dependence

of rates of deposition of
sediments R on the time span
T over which deposition
occurs. The circles are mean
rates of deposition of fluvial
sediments from modem
sedimentary basins and
ancient stratigraphic sections
(Sadler and Strauss, 1990).
The straight line is the fractal
correlation from (2.14)
taking D = 0.336.
D E F I N I T I O N OF A FRACTAL S E T 25
tics of sediment deposition. The erosional processes responsible for the for-
mation of coastlines and the depositional processes responsible for the struc-
ture of a sedimentary pile are both extremely complex. But despite the com-
plexity, both examples exhibit fractal behavior to a good approximation. A
simple explanation is that a distribution will be fractal if there is no charac-
teristic length in the problem. The fractal distribution is the only statistical
distribution that is scale invariant. However, a broad class of nonlinear phys-
ical problems involving chaotic behavior and/or self-organized critical be-
havior invariably yield fractal behavior. One objective in succeeding chap-
ters is to describe physically realistic models that generate fractal behavior.
Problems
Problem 2.1. Consider the construction illustrated in Figure 2.l(e). (a) Illus-
trate the construction at third order. (b) Determine N,, N,, r,, and r,.
Problem 2.2. Consider the construction illustrated in Figure 2.l(f). (a) Illus-
trate the construction at third order. (b) Determine N,, N,, r,, and r,.
Problem 2.3. A unit line segment is divided into five equal parts and two are
retained. The construction is repeated. (a) Illustrate the construction to
third order, i.e., consider i = 1, 2, 3. (b) Determine N,, N,, N,, r,, r,, r,.
(c) Determine the fractal dimension.
Problem 2.4. A unit line segment is divided into seven equal parts and three
are retained. The construction is repeated. (a) Illustrate this construction
to second order, i.e., consider i = 1,2. (b) Determine N,, N,, N3, r,, r2, r3.
Problem 2.5. A unit line segment is divided into seven equal parts and four
are retained. The construction is repeated. (a) Illustrate this construction
to second order, i.e., consider i = 1,2. (b) Determine N,, N,, N,, r,, r2, r,.
Problem 2.6. Consider the construction of the Sierpinski carpet illustrated in
Figure 2.3(d) at third order. (a) Illustrate the construction at third order.
(b) Determine N,, N,, r,, and r,.
Problem 2.7. A unit square is divided into four smaller squares of equal size.
Two diagonally opposite squares are retained and the construction is re-
peated. (a) Illustrate the construction to third order, i.e., consider i = 1, 2,
3. (b) Determine N,, N,, N,, r,, r,, r,. (c) Determine the fractal dimen-
sion.
Problem 2.8. A unit square is divided into nine smaller squares of equal size.
The center square and four corner squares are retained and the construc-
tion is repeated. This is known as a Koch snowflake. (a) Illustrate the
construction to second order, i.e., consider i = 1,2. (b) Determine N,, N,,
N,, r,, r,, r,. (c) Determine the fractal dimension.
Problem 2.9. A unit square is divided into nine smaller squares of equal size
and the four corner squares are discarded. The construction is repeated.
(a) Illustrate the construction to second order. (b) Determine N,, N,, N,,
r,, r,, r,. (c) Determine the fractal dimension.
Problem 2.10. A unit square is divided into 16 smaller squares of equal size.
The four central squares are removed and the construction is repeated.
(a) Illustrate this construction to second order, i.e., consider i = 1, 2. (b)
Determine N,, N,, N,, r,, r,, r,. (c) Determine the fractal dimension.
Problem 2.1 1. A unit square is divided into 25 smaller squares of equal size.
All squares are retained except the central one and the construction is re-
peated. (a) Illustrate this construction to second order, i.e., consider i =
1, 2. (b) Determine N,, N,, N,, r,, r,, r,. (c) Determine the fractal dimen-
sion.
All the squares on the boundary and the central square are retained and
the construction is repeated. (a) Illustrate this construction to second or-
der, i.e., consider i = 1, 2. (b) Determine N,, N,, N,, r,, r,, r,. (c) Deter-
mine the fractal dimension.
Problem 2.13. A unit cube is divided into 27 smaller cubes of equal volume.
All the cubes are retained except for the central one. What is the fractal
dimension?
Problem 2.14. Consider a variation on the Koch island illustrated in Figure
2.5. At zero order again consider an equilateral triangle with three sides
of unit length. At first order this triangle is enlarged so that it is an equi-
lateral triangle with sides of length three. Equilateral triangles with sides
of unit length are placed in the center of each side. (a) Illustrate this con-
struction at second order. (b) Determine the areas to second order, i.e.,
obtain A,, A,, A,. (c) Do the areas given in (b) satisfy the fractal condi-
tion (2. I)? If the answer is yes, what is the fractal dimension?
Problem 2.1 5. Consider the fractal construction illustrated in Figure 2.15. A

a fractal construction. (a)
The zero-order unit square.
(b) The first-order fractal
construction.
unit square is considered at zero order and the first-order fractal con-
struction is also illustrated. (a) Illustrate the construction at second or-
der. (b) Determine No, N,, N,, r, r,, r,, Po, P I ,P,. (c) Determine the frac-
tal dimension.
Problem 2.16. Assume that the open squares in the Sierpinski carpet illus-
trated in Figure 2.3(d) represent lakes. (a) Determine the numbers of
lakes to the third order, i.e., obtain N,, N,, N, corresponding to r,, r,, r,.
(b) Do the numbers of lakes given in (a) satisfy the fractal condition
(2. l)? If the answer is yes, what is the fractal dimension?
Problem 2.17. Zipf's law (Zipf, 1949) has been applied in a wide variety of
problems including the size distribution of cities. This law states that the
2nd largest is $ the size of the largest, the 3rd largest is $ the size of the
largest, the 4th largest is the size of the largest, and so forth. Does this
distribution satisfy a cumulative fractal distribution and, if so, what is
the fractal dimension?
Problem 2.18. Construct a second-order devil's staircase based on the fractal
construction given in Figure 2.l(f).
Problem 2.19. Consider the simple deposition model illustrated in Figure
2.11. Assume that no erosion occurs. What are the ages of the sediments
immediately above and below the resulting unconformity?
Problem 2.20. Use the second-order Cantor set based on the fractal construc-
tion given in Figure 2,l(f) as a model for sedimentation. Assume in this
model that 9 km of sediments have been deposited in 25 Myr. (a) At
what depths do the two first-order unconformities occur, and what are
the ages of the sediments just above and just below the unconformities?
(b) At what depths do the six second-order unconformities occur, and
what are the ages of the sediments just above and just below the uncon-
formities? (c) Plot the age of the sediments as a function of depth. (d)
What are the rates of deposition associated with the periods 25 Myr,
5 Myr, 1 Myr?
ChapterThree
FRAGMENTATION
3.1 Background
To illustrate how fractal distributions are applicable to real data sets, we con-
sider fragmentation. Fragmentation plays an important role in a variety of
geological phenomena. The earth's crust is fragmented by tectonic processes
involving faults, fractures, and joint sets. Rocks are further fragmented by
weathering processes. Rocks are also fragmented by explosive processes,
both natural and man made. Volcanic eruptions are an example of a natural
explosive process. Impacts produce fragmented ejecta. Although fragmenta-
tion is of considerable economic importance and many experimental, numer-
ical, and theoretical studies have been camed out on fragmentation, rela-
tively little progress has been made in developing comprehensive theories of
fragmentation. A primary reason is that fragmentation involves the initiation
and propagation of fractures. Fracture propagation is a highly nonlinear pro-
cess requiring complex models for even the simplest configuration. Frag-
mentation involves the interaction between fractures over a wide range of
scales. Fragmentation phenomena have been discussed by Grady and Kipp
(1987) and Clark (1987). If fragments are produced with a wide range of
sizes and if natural scales are not associated with either the fragmented ma-
terial or the fragmentation process, fractal distributions of number versus
size would seem to be expected. Some fractal aspects of fragmentation have
been considered by Turcotte (1986a).
3.2 Probability and statistics
Clearly the distribution of fragment sizes is a statistical problem. This will

also be the case in other applications we will consider. Thus at this point it is
appropriate to introduce some of the fundamental concepts of probability
and statistics. Data that we will be considering can be divided into two types,
FRAGMENTATION 29
discrete data and continuous data. Discrete data are generally characterized
by a set of n data points {x,,x,, . . .xi, . . . ,xn).Examples include the masses
of n fragments and the magnitudes of n earthquakes. It is standard practice to
describe the statistical properties of a discrete data set by defining the mean
and moments of deviations from the mean. The mean value of the xi, i , is
given by
The average squared deviation from the mean is a measure of the spread of
the data; this is the variance V and for a discrete set of n data points it is
given by
The variance is the second-order moment of the distribution. The standard

deviation of the distribution a is simply the square root of the variance
The asymmetry of the data is quantified by the coefficient of skew y, which

is the third-order moment
The factor a 3 makes y a nondimensional number.

Higher-order moments can be defined but they are generally of little use.
It is standard practice in geostatics to fit an empirical statistical distribution
to a discrete set of data. This is often done by equating the mean i ,variance
V, and skew y of the distribution to that of the data. It should be noted that a
variety of definitions for the variance and skew appear in the literature; for
example, the n in (3.2) is sometimes replaced by the factor n - 1 (Barlow,
1989, pp. 10-12).
As a simple example assume that x takes the values x, = 2, x, = 5, x3 = 9,
and x, = 16. From (3.1) the mean is i = 8, from (3.2) the variance is V = 27.5,
from (3.3) the standard deviation is a = 5.244, and from (3.4) the skew is y =
1.311.
30 FRAGMENTATION
In many cases, each of the values x , , x,, . . . ,xi, . . . ,xn will have a prob-
ability of occurring f , , f,, . . . ,&, . . . ,fn. By definition we have
Introducing these probabilities, the mean, variance, standard deviation, and

skew of a distribution are given by
+
As an example consider flipping coins. Assign 1 to a head and - 1 to a tail.
For a single coin there are two values of x, x = + 1 (a head) and x2 = - 1 (a
tail). Since the probabilities of having a head or a tail are equal, we have f , =
f2 = 0.5. Next consider flipping two coins. We now have three values for x,
x , = + 2 (two heads), x, = 0 (one head and one tail), and x, = -2 (two tails).
However, there are two ways to obtain x, = 0, the first coin is a head and the
second a tail or the first is a tail and the second a head, whereas there is only
one way to obtain x, = + 2 (two heads) and x3 = -2 (two tails). Thus we have
f,=0.25,f2=0.5,andf3=0.25.From(3.6) t o ( 3 . 9 ) w e f i n d i = y = 0 , V = 2 ,
and a = fi.Finally consider flipping three coins. In this case we have x, = 3
(3 heads) with f,= 0.125, x, = 1 (2 heads, 1 tail) with f, = 0.375, x, = - 1
(1 head, 2 tails) with f3 = 0.375, and x, = -3 (3 tails) with f, = 0.125. From
(3.6) to (3.9) we find 2 = y = 0, V = 3, and a = fi.
We will next consider continuous data. A particular variable can take any
value over a specified range, say - = < x < =. An example would be the
x-component of the velocity of the gas molecules in a room, vx. It is appro-
priate to consider the range of velocities -= < v < -, and there will be a
X.
statistical probability that a particular molecule will have a velocity greater
than a specified value vx. In terms of our general distribution the cumulative
distribution function F (x,) is the probability Pr that x has a value greater
than x,
FRAGMENTATION 31
--
It should be noted that the usual definition of a cumulative distribution func-
tion in probability and statistics is from to x rather than x to m, i.e.,
For most applications in geology and geophysics we are concerned with the
"number" larger than a specified value and thus the definition of F(x) given
in (3.10) is preferred. The cumulative distribution function is related to the
probability distribution functionflx) by
where fix) 6 x is the probability that x lies in the range (x - 6x) < x 5
(x + 3 6 ~ )We
. also have
and the probability Pr that x lies between x, and x2 is given by
Pr(x, < x 5 x,) = f(u)du = F(x,) - F(x2) (3.14)
Several widely used statistical distributions will now be discussed.

One of the most widely used statistical distributions is the normal distri-
bution; it is also known as the Gaussian distribution or simply as the bell
curve (from its shape). Its probability distribution function takes the form
From statistical mechanics (Morse, 1969) the x-component of molecular ve-

locities in a room satisfy a Gaussian distribution of the form
where m is the molecular mass, k the Boltzmann constant, and T the temper-
ature.
32 FRAGMENTATION
The mean of the normal distribution is obtained using the relation
Introducing
x-x
Y=-
diu
(3.17) becomes
but
so that i is the mean of the normal distribution. The variance of the normal
distribution is obtained from
:v : 1
= - x f(x)dx =
1
o (2n)lD I:m [
( x - i)' exp - -
" 2-
i '-
y ]-
dx (3.20)
Introducing (3.18) gives
so that a is the standard deviation of the normal distribution.

We define the cumulative distribution function for the normal distribu-
tion to be given by (3.13)
Making a substitution of the form of (3.18) into (3.22) gives

FRAGMENTATION 33
It is convenient to introduce the error function
and the complementary error function
And noting that
the cumulative distribution for the normal distribution becomes
and F (- w) = 1 as expected. The probability that x is in the range x, < x Ix,

is given by
The standard form of the normal distribution is obtained by taking i = 0 and

a = 1. The probability distribution functionflx) and the cumulative distribu-
tion function F(x) for the standard form of normal distribution are given in
Figure 3.1. Relevant values are tabulated in Table 3.1. Note that for this sym-
metric distribution F(0) = 0.5 and F(-=) = 1.
It is often necessary to generate a set of random numbers with a speci-
fied distribution of values, say a normal (Gaussian) distribution. One way to
do this is to use the cumulative distribution function F(x). A set of random
numbers with a uniform distribution between 0 and 1 is obtained. Each ran-
dom number is assumed to be Fi(xi) and the corresponding value of xi is
found. If the dependence of F(x) on x given in Figure 3.1 is used, the values
of xi will satisfy a normal (Gaussian) distribution.
34 FRAGMENTATION
Table 3.1. Values relevant to the normal distribution
For an application in which a finite number of observations have been

carried out, it is difficult to apply the probability distribution function di-
rectly. A selection of "bins" is chosen and each data point is put in a bin. The
number of points in the bins then constitute the statistical distribution. For
the normal distribution, (3.28) can be used to establish the probability of a
data point being in a bin. If the cumulative distribution is used it is not nec-
essary to "bin" the data. The number of data points with values greater than x
is plotted against x. This is the preferred way to handle most geological and
geophysical data.
The wide applicability of the normal distribution is based on the central
limit theorem. This theorem states that if a distribution is the sum of a large
Figure 3.1. The probability

distribution functionflx) and
the cumulative distribution
function F(x) for the
standard form of the normal
distribution, from (3.15)
and (3.22) with 2 = 0 and
a = 1.
FRAGMENTATION 35
number of independent random distributions, the distribution will approach

a normal distribution as the number approaches infinity. For example, con-
sider the coin flipping experiment discussed above with x, = + 1 for a head
and x2 = - 1 for a tail. As the number of coins approaches infinity the distri-
bution of values approaches a normal distribution.
A normal distribution is symmetrical about its mean so that its coeffi-
cient of skew is zero, y = 0; also, the independent variable x takes on all val-
ues from -M to +-. In many applications a distribution of only positive
values is required. One of the most widely used distributions of this type is
the log-normal distribution. The log-normal distribution can be obtained di-
rectly from the normal distribution simply by taking the logarithm of the
normally distributed values [replacing the x in the definition of the normal
probability distribution function given in (3.15) with y]; the corresponding
log-normal probability distribution function is given by
where the substitution
has been made noting that dy = dxlx andfix) dx =f(y) dy. The values of y are
normally distributed with a mean j and a standard deviation uy.Using the
definitions of the mean and standard deviation, (3.6) and (3.8), we can relate
i and a to j and uywith the result
Since both the standard deviation and the mean are positive for the log-
normal distribution, the ratio of the two quantities is a measure of the spread
of the distribution
and is known as the coefficient of variation. The coefficient of skew for the
log-normal distribution is
36 FRAGMENTATION
The values of ayand y are related to 2 and c, by
a,,= [ln(l + c,)]

2 112
These relations with (3.29) specify the log-normal probability distribution

function when i and c, have been given.
The cumulative distribution function for the log-normal distribution is
obtained from (3.13)
And making the substitution
we obtain
The cumulative distribution functions for the log-normal and normal distri-
butions have the same forms when x is replaced by In x, i.e. (3.30).
The standard form of the normal distribution was obtained by taking i =
0 and V = 1. All normal distributions have this universal form and can be
obtained simply by rescaling. This is not the case for the log-normal distrib-
ution. The probability distribution functionsflx) for the log-normal distribu-
tion are given in Figure 3.2 for 2 = 1 and c, = 0.25,0.50, 1.00. It is seen that
the shape of the log-normal distribution changes systematically with changes
in c,. As the value of c, becomes smaller the distribution narrows and the
maximum value approaches x = 1. In the limit c, + 0 the distribution is a 6
function centered at x = 1.As c, becomes larger the distribution spreads out
and the maximum value occurs at smaller x. Whereas the normal distribution
is symmetric with a zero coefficient of skew, the asymmetry and coefficient
of skew for the log-normal distribution increase with increasing c,
Log-normal distributions are basically a one-parameter family of distri-
butions depending on the appropriate value of c,. This has important impli-
FRAGMENTATION 37
cations in terms of applications. It is often appropriate to approximate the

distribution of annual rainfalls at a station by a log-normal distribution. A
maritime station, say Seattle, would have little year-to-year variation in rain-
fall and a small value for c , On the other hand, an arid station, say Phoenix,
would have large year-to-year variations in rainfall and a larger value for c,.
This dependence on c , for rainfall and river flows is referred to as the Noah
effect (Mandelbrot and Wallis, 1968).
We will next consider the Pareto distribution, which is closely associated
with fractal distributions. The Pareto probability distribution function is
given by
The corresponding cumulative distribution function is given by
The standard form of the Pareto distribution is obtained by taking k = 1 so

that (3.40) becomes

distribution function f(x) for
the log-normal distribution
with unit mean, % = 1, and
several values of the
coefficient of variation c,.
38 FRAGMENTATION
and (3.41) becomes
Fb) = (1 ; y)'
y20
The mean of the standard form is given by
Y a dy 1
= [ (1 + y)'+l
-
a - 1
for a > 1
This integral does not converge and a mean does not exist for a I1.The vari-
ance of the standard form of the Pareto distribution is given by
a
- for a > 2 (3.45)
(a - l)(a - 2)
This integral does not converge for a 1 2 and the variance does not exist.
The Pareto distribution is widely used in economics and is often a good
approximation for the distribution of incomes (Ijiri and Simon, 1977). The
probability distribution functions for the standard form of the Pareto distri-
bution are given in Figure 3.3 for a = 1, 2, and 3. The power-law tail of the
Pareto distribution dies off much more slowly than the tails of the normal or
log-normal distributions; this is the characteristic of fractal distributions.
For y >> 1we can write (3.43) as
This is clearly quite similar to the statistical fractal relation (2.6)
The power a in the Pareto distribution is equivalent to the fractal dimension

D. This has led some statisticians and others to conclude that fractals are a
trivial extension of the Pareto distribution. While there is clearly a close as-
sociation between statistical fractals and the Pareto distribution, there are
FRAGMENTATION 39
many other aspects of fractal concepts. The wide applicability of scale in-
variance provides a rational basis for fractal statistics just as the central limit
theorem provides a rational basis for Gaussian statistics.
An important distinction between the cumulative Pareto distribution
(3.41) and the fractal distribution (3.47) is that the former is finite as x + 0
whereas the latter diverges to w as r + 0. Scale invariance implicitly re-
quires this divergence. Many geological and geophysical data sets also have
this divergence. As a specific example, consider earthquakes. Data on large
earthquakes are often complete, but data on small earthquakes generally do
not exist. Even the best seismic networks cannot resolve the very smallest
earthquakes that are known to occur. Thus it is impossible to define com-
plete probability or cumulative distribution functions for earthquakes. How-
ever, it is possible to determine the number of earthquakes N that have rup-
ture dimensions greater than r, and we will show in Chapter 4 that the
frequency-magnitude statistics for earthquakes are fractal and do satisfy
(3.47).
The final distribution we will consider is an exponential distribution; its
probability distribution function is given by
f(x) =
vxv-
7ex*
xo
'
I[-"):( x 20

distribution functionm)
for the standard form of the
Pareto distribution with a =
1, 2,3; from (3.42).
40 FRAGMENTATION
where the power v is generally taken to be an integer. The mean of the expo-
nential distribution is given by
Making the substitution
we obtain
where r (v') is the tabulated gamma function (Dwight, 1961, Table 1005). If
v = 2 we have r (i) = 0.886 so that 2 = 0.88%. The variance of the exponen-
tial distribution is given by
Again substituting (3.50) we obtain
If v = 2 we have r (2) = 1 so that V = 0.2146~;.

The cumulative distribution function for the exponential distribution is
obtained from (3.13) with the result
This is known as the Rosin and Rammler (1933) distribution and it is used
extensively in geostatistical applications. We can also write
FRAGMENTATION 41
1 - F(x) = 1 - exp -
[ (3'1
-
This is known as the Weibull distribution. Thus the Weibull distribution is

entirely equivalent to the Rosin and Rammler distribution. The Rosin and
Rammler distribution is the Pr (x' > x) and the Weibull distribution is the Pr
(0 < x' < x) when the probability distribution function is given by (3.48).
The Rosin and Rammler, F b ) , and the Weibull, 1 - F(y), distributions are
given in Figure 3.4 for v = 2 and 4.
If (xIx,)" is small, then the exponential in (3.55) can be expanded in a
Taylor series to give
where higher powers of (xlx,). have been neglected. Substitution of (3.56)

into the Weibull distribution (3.55) gives
Thus for small x the Weibull distribution reduces to a power-law (fractal)

distribution. This power-law approximation is also illustrated in Figure 3.4.
Figure 3.4. The Rosin and

Rammler distribution F ( y )
from (3.54) with y = xlx,
and the Weibull distribution
1 -F(y) from (3.55) are
given for v = 2 and 4. Also
included is the power-law
(fractal) approximation to
the Weibull distribution from
(3.57).
42 FRAGMENTATION
3.3 Fragmentation data
Many of the statistical distributions discussed above have been used to rep-
resent the frequency-size (mass) distributions of fragments; these include
log-normal, Pareto, Rosin and Rammler, Weibull, and power law. In terms of
the concepts developed in Chapter 2, it is clear that we would like to relate
the number of fragments N to their linear dimension r. Since fragments can
occur in a variety of shapes, it is appropriate to define a linear dimension r as
the cube root of volume, r = W 3 . Assuming constant density it follows that
-
m r3, where m is the mass of a fragment. However, it is standard practice to
give the total mass of fragments with a linear dimension r less than a speci-
fied value M (<r) or the total mass of fragments with a linear dimension r
greater than a specified value M (>r). The reason for this is that these
masses are obtained directly from a sieve or screen analysis; the mass of
fragments passing through a sieve with a specified aperture r is M (<r) and
the remaining mass is M (>r). Of course we have
the total mass of fragments.

In many cases the power-law approximation to the Weibull distribution
(3.57) can be used to approximate sieve analyses in the form
This power-law mass relation can be related to the fractal number relation
by taking incremental values (Redner, 1990). Taking the derivative of (3.58)

gives
Taking the derivative of (3.59) gives
However, the incremental number is related to the incremental mass by

FRAGMENTATION 43
Substitution of (3.60) and (3.61) into (3.62) gives
When data are obtained by sieve analyses, (3.64) is used to convert mass dis-
tributions to number distributions to specify a fractal dimension.
Many experimental studies of the frequency-size distributions of frag-
ments have been carried out. Several examples of power-law fragmentation
are given in Figure 3.5. A classic study of the frequency-size distribution for
broken coal was carried out by Bennett (1936). The frequency-size distribu-
tion for the chimney rubble above the PILEDRIVER nuclear explosion in
Nevada has been given by Schoutens (1979). This was a 61 kt event at a
depth of 457 m in granite. The frequency-size distribution for fragments re-
sulting from the high-velocity impact of a projectile on basalt has been given
by Fujiwara et al. (1977). In each of the three examples a good correlation
with the fractal relation (2.6) is obtained over two to four orders of mag-
nitude. In each example the fractal dimension for the distribution is near
D = 2.5.
Further examples of power-law distributions for fragments are given in
Table 3.2. It will be seen that a great variety of fragmentation processes can
be interpreted in terms of a fractal dimension. Examples include impact shat-
Figure 3.5. Since fragments
have a variety of shapes, the
cube root of volume is an
objective measure of size.
The number N of fragments
with cube root of volume
greater than r is given as a
function of r for broken coal
(Bennett, 1936), broken
granite from a 61 kt
underground nuclear
detonation (Schoutens,
1979), and impact ejecta
due to a 2.6 km s-1
polycarbonate projectile
impacting on basalt
(Fujiwara et al., 1977). The
best-fit fractal distribution
from (3.59) is shown for
each data set.
44 FRAGMENTATION
tering, explosive disruption, crushed materials, weathered materials, and

volcanic ejecta. The fact that fault gouge has a fractal frequency-size distri-
bution is a particularly striking example of how a natural geological process
can result in fractal fragmentation (Sammis et al., 1986; An and Sammis,
1994; Sammis and Steacy, 1995). The relative displacement across a fault
zone results in the fragmentation of the wall rock to form a zone of frag-
mented rock known as fault gouge. This is referred to as comminution since
it strongly resembles the fragmentation that takes place in a grinding mill.
Sammis et al. (1987) and Sammis and Biegel (1989) have shown that the
fault gouge obtained from the Lopez fault zone, San Gabriel Mountains,
California, has a fractal dimension D = 2.60 + 0.11 on scales from 0.5 pm to
10 mm. Synthetic fault gauge has also been shown to have a fractal dimen-
sion D = 2.60 (Biegel et al., 1989). Sammis et al. (1987) have also suggested
that the comminution of the earth's crust has resulted in fractal tectonic frag-
mentation on scales from millimeters to hundreds of kilometers. We will re-
turn to this concept in the next chapter.
Studies of the frequency-size distribution of asteroids show that they fit
a power-law (fractal) relation to a good approximation taking D I3 2.5
(Klacka, 1992). Since asteroids are responsible for the impact craters on the
surface of the moon, it is not surprising that the frequency-size distribution
of lunar craters is also fractal with D e1.4 (Greeley and Gault, 1970).
Table 3.2. Fractal dimensions for a variety of fragmented objects
Fractal
dimension
Object Reference D
Artificially crushed quartz Hartmann (1969) 1.89
Disaggregated gneiss Hartmann (1969) 2.13
Disaggregated granite Hartmann (1969) 2.22
FLAT TOP I (chemical
explosion, 0.2 kt) Schoutens (1979) 2.42
PILEDRIVER (nuclear
explosion, 62 kt) Schoutens (1979) 2.50
Broken coal Bennett (1936) 2.50
Asteroids Klacka (1992) 2.50
Projectile fragmentation of
quartzite Curran et al. (1977) 2.55
Projectile fragmentation of
basalt Fujiwara et al. (1977) 2.56
Fault gouge Sammis and Biegel(1989) 2.60
Sandy clays Hartmann (1969) 2.61
Soils Wu et al. (1993) 2.80
Terrace sands and gravels Hartmann (1969) 2.82
Glacial till Hartmann (1969) 2.88
Ash and pumice Hartmann (1969) 3.54
FRAGMENTATION 45
It is seen that the values of the fractal dimension vary considerably, but
most lie in the range 2 < D < 3. This range of fractal dimensions can be re-
lated to the total volume of fragments and to their surface area.
The total volume (mass) of fragments is given by
since r has been defined to be the cube root of the volume. In all cases it is
expected that there will be upper and lower limits to the validity of the frac-
tal (power-law) relation for fragmentation. The upper limit rmaxis generally
controlled by the size of the object or region that is being fragmented. The
lower limit rminis likely to be controlled by the scale of the heterogeneities
responsible for fragmentation, for example the grain size. For a power-law
(fractal) distribution of sizes, substitution of (3.61) into (3.65) and integra-
tion gives
If 0 < D < 3 it is necessary to specify rmax but not rminto obtain a finite vol-
ume (mass) of fragments. The volume (mass) of fragments is predominantly
in the largest fragments. This is the case for most observed distributions of
fragments (see Table 3.2). If D > 3 it is necessary to specify rminbut not rmax.
The volume (mass) of the small fragments dominates.
The total surface area A of the fragments is given by
where C is a geometrical factor depending on the average shape of the frag-

ments. For a power-law distribution, substitution of (3.61) into (3.67) and in-
tegration gives
If 0 < D < 2 it is necessary to specify rmaxbut not rminto obtain a finite total
surface area for the fragments. But if D > 2 it is necessary to specify rminto
constrain the total surface area to a finite value. Thus for most observed dis-
tributions of fragments (see Table 3.2) the surface area of the smallest frag-
ments dominates.
46 FRAGMENTATION
3.4 Fragmentation models
A simple model illustrates how fragmentation can result in a fractal distribu-

tion. This model is illustrated in Figure 3.6; it is based on the concept of
renormalization, which will be considered in greater detail in Chapter 15. A
cube with a linear dimension h is referred to as a zero-order cell; there are No
of these cells. Each zero-order cell may be divided into eight equal-sized,
zero-order cubic elements with dimensions h/2. The volume V, of each of
these elements is given by
where Vo is the volume of the zero-order cells. The probability that a zero-
order cell will fragment to produce eight zero-order elements is taken to be f.
The number of zero-order elements produced by fragmentation is
After fragmentation the number of zero-order cells that have not been frag-
mented, No,, is given by
Each of the zero-order elements is now taken to be a first-order cell.

Each first-order cell may be fragmented into eight equal-sized, first-order
cubic elements with dimensions h/4. The fragmentation process is repeated
Figure 3.6. Idealized model

for fractal fragmentation. A
zero-order cubic cell with
dimensions h is divided
into eight zero-order cubic
elements each with
dimensions hl2. The
probability that a zero-order
cell will be fragmented into
eight zero-order elements
is$ The fragments with
dimension hl2 become first-
order cells; each of these
has a probability f of being
fragmented into first-order
elements with dimensions
h/4. The process is repeated
to higher orders. The basic
structure is fractal.
FRAGMENTATION 47
for these smaller cubes. The problem is renormalized and the cubes with di-
mension h/2 are treated in exactly the same way that the cubes with linear di-
mension h were treated above. Each of the fragmented cubic elements with
linear dimension h/2 is taken to be a first-order cell; each of these cells is di-
vided into eight first-order cubic elements with linear dimensions h/4 as il-
lustrated in Figure 3.6. The volume of each first-order element is
The probability that a first-order cell will fragment is again taken to be f to

preserve scale invariance. The number of fragmented first-order elements is
After fragmentation the number of first-order cells that have not been frag-
mented is
This process is repeated at successively higher orders. The volume of the

nth-order cell Vnis given by
After fragmentation the number of nth-order cells is
Taking the natural logarithm of both sides we can write (3.75) and (3.76) as
Elimination of n from (3.77) and (3.78) gives

48 FRAGMENTATION
Comparison with (2.1) shows that this is a fractal distribution with
Although this model is very idealized and non-unique, it illustrates the

basic principles of how scale-invariant fragmentation leads to a fractal dis-
tribution. It also illustrates the principle of renormalization. The division
into eight fragments is an arbitrary choice, however; other choices such as
the division into two or 16 fragments will give the same result. This model is
deterministic rather than statistical. Actual distributions of fragments are
continuous rather than discrete but the deterministic model can be related to
a "bin" analysis of a statistical distribution. Also, this model relates the prob-
ability of fragmentation f to the fractal dimension D but does not place con-
straints on the value of the fractal dimension.
It is of interest to discuss this model in terms of the allowed range of
6
D. The allowed range off is < f < 1 and the equivalent range of D is 0 <
D < 3. Thus the concept of fractional dimension introduced in Figure 2.4
appears to be appropriate for fragmentation. However, the data for ash and
pumice given in Table 3.2 fall outside this allowed range. Since such distri-
butions are not precluded physically, we will consider this a fractal (scale-
invariant) distribution even though it lies outside the geometrically allowed
range. We accept the physical view rather than the mathematical view.
Fragmentation is a process with a wide range of applications. Thus many
studies have been carried out to prescribe size distributions in terms of basic
physics; however, fragmentation is a very complex problem. The results
given above indicate that in many cases fragmentation is a scale-invariant
process that leads to a fractal distribution. We now turn to a discrete model
of fragmentation that does yield a specific fractal dimension. We will con-
sider the fractal cube illustrated in Figure 2.4(b) and use it as a basis for a
fragmentation model. This model is illustrated in Figure 3.7. Although the
geometry and fractal dimension are the same, the concepts of the two models
are quite different. The models given in Figure 2.4 are essentially for a
porous (Swiss cheese) configuration. At each scale blocks are removed to
create void space. In this chapter we consider fragmentation such that some
blocks are retained at each scale but others are fragmented. In the model
given in Figure 3.7 two diagonally opposed blocks are retained at each scale.
No two blocks of equal size are in direct contact with each other. This is the
comminution model for fragmentation proposed by Sammis et al. (1987). It
is based on the hypothesis that direct contact between two fragments of near
equal size during the fragmentation process will result in the breakup of one
of the blocks. It is unlikely that small fragments will break large fragments
or that large fragments will break small fragments.
FRAGMENTATION 49
For the configuration illustrated in Figure 3.7 we have N, = 2 for r , =

h/2,N2 = 12 for r, = h/4, and N, = 72 for r, = h/8. From (2.2) we find that D =
In 6/ln 2 = 2.5850. This is the fractal distribution of a discrete set but we
wish to compare it with statistical fractals obtained from the actual fragmen-
tation observations. It is therefore of interest to consider also the cumulative
statistics for the comminution model. The cumulative number of blocks
larger than a specified size for the three highest orders are N I c = 2 for r , =
h/2, N2= = 14 for r, = hl4, and N,c = 86 for r, = h/8; Nnc is the cumulative
number of the fragments equal to or larger than r,,. The cumulative statistics
for the model illustrated in Figure 3.7 are given in Figure 3.8; excellent

of a fractal model for
fragmentation. Two
diagonally opposite cubes
are retained at each scale.
With r, = h/2, N , = 2 and r, =
h/4, N , = 12 we have D =
In 6 t h 2 = 2.5850.
Figure 3.8. Cumulative

statistics for the
fragmentation model
illustrated in Figure 3.7.
Correlation with (2.6)
gives D = 2.60.
50 FRAGMENTATION
agreement with the fractal relation (2.6) is obtained taking D = 2.60. Thus
the fractal dimensions for the discrete set and the cumulative statistics are
nearly equal.
This comminution model was originally developed for fault gouge. The
derived fractal dimension for the model D = 2.60 is in excellent agreement
with the measured values for fault gouge described in the last section. It is
seen from Figure 3.5 and Table 3.2 that many observed distributions of frag-
ments have fractal distributions near this value. This is evidence that the
comminution model may be widely applicable to rock fragmentation. This
model may also be applicable to tectonic zones in the earth's crust. The im-
plication is that there is a fractal distribution of tectonic blocks over a wide
range of scales.
A number of other models have been proposed to explain fractal frag-
mentation. Steacy and Sammis (1991) developed an automaton that modeled
nearest neighbor fragmentation. Palmer and Sanderson (1991) developed a
model for crushing ice that accounts for the relative size of contacting frag-
ments. In their model, D = 2.5 has the special significance that fragments of
all sizes make equal contributions to the crushing force. Englman et al.
(1987, 1988) have obtained a power-law distribution utilizing a maximum-
entropy model.
3.5 Porosity
Most rock has a natural porosity. This porosity often provides the necessary
permeability for fluid flow. There are generally two types of porosity, inter-
granular porosity and fracture porosity. Based on the discussion given above
it would not be surprising if both types of porosity exhibited fractal behav-
ior. Fractures are directly related to fragmentation, and detridal rocks are
composed of rock grains with a variety of scales. Based on laboratory stud-
ies a number of authors have suggested that sandstones have a fractal distri-
bution of porosity (Katz and Thompson, 1985; Krohn and Thompson, 1986;
Daccord and Lenormand, 1987; Krohn, 1988a, b; Thompson et al., 1987).
Hansen and Skjeltorp (1988) carried out two-dimensional box counting of
the pore space in a sandstone and found D = 1.73. Brakensiek et al. (1992)
carried out similar studies of the two-dimensional porosity of soils and
found D = 1.8. Soils can be considered both in terms of fractal distributions
of particle sizes and in terms of fractal distributions of void spaces (Rieu and
Sposito, 1991a, b; 5 l e r and Wheatcraft, 1992). Fractal distributions of
voids have also been suggested to be applicable to caves (Curl, 1986), karst
regions (Laverty, 1987), and sinkholes (Reams, 1992).
We previously introduced models with scale-invariant porosity in Figure
2.4. The Menger sponge, Figure 2.4(a), can be taken as a simple model for a
FRAGMENTATION 51
porous medium. However, this model is conceptually somewhat different

from that considered in Chapter 2. The model is constructed from solid
cubes of density p, and size r,. We construct a first-order Menger sponge
from these cubes; the size of the first-order cube is r, = 3r0. The first-order
sponge is made up of 20 solid zero-order cubes so that the first-order poros-
ity is 4, = 7/27 and the first-order density is p , = 2 0 ~ 4 2 7 Continuing
. the
construction to second order, the size of the cube is r, = 9r, and there are 400
solid cubes of size r, with density p,. Thus the porosity of the second-order
Menger sponge is 4, = 3291729 and its density is p, = 4 0 0 ~ 4 7 2 9The
. poros-
ity of the nth-order Menger sponge is
which is not a power-law (fractal) relation. The density of the nth-order

Menger sponge is
This is a fractal relation and is illustrated in Figure 3.9. For the Menger
sponge the fractal dimension is D = In 201111 3 = 2.727. Generalizing (3.81),
the porosity 4 for a fractal medium can be related to its fractal dimension by
Figure 3.9. Density

dependence p/p, of a Menger
sponge as a function of the
size of the sponge rlr,. A
fractal decrease in the
density is found with D =
2.727 from (3.82).
52 FRAGMENTATION
where r is the linear dimension of the sample considered. Similarly, the den-
sity of the fractal medium scales with its size according to
The density of a fractal solid systematically decreases with the increasing

size of the sample considered.
A number of studies of the densities of soil aggregates as a function of
size have been carried out. These studies show a systematic decrease in den-
sity as the size of the aggregate increases. A sieve analysis is carried out on a
soil, and the mean density of each aggregate is found. The results for a sandy
loam obtained by Chepil (1950) are given in Figure 3.10. Although there is
scatter, the results agree reasonably well with the fractal soil porosity from
(3.84) using as the fractal dimension D = 2.869.
Figure 3.10. Density of soil

aggregates as a function of
their size (Chepil, 1950). The
solid line is from (3.84) with
D = 2.869.
FRAGMENTATION 53
Problems
Problem 3.1. Assume that x takes the values xl = 1, x2 = 4, x, = 7, and x, = 8.

Determine i , V, u, and y.
Problem 3.2. Assume that x takes the values x, = 3, x, = 6, x, = 14, and x, =
17. Determine 2, V, a,and y.
Problem 3.3. Flip four coins assigning + 1 to a head and - 1 to a tail. What
are the probabilities of obtaining +4, +2,0, -2, -4? What are the vari-
ance and standard deviation of this distribution?
Problem 3.4. Flip five coins assigning + 1 to a head and - 1 to a tail. What
are the probabilities of obtaining +5, +3, +1, -1, -3, -5? What are
the variance and standard deviation of this distribution?
Problem 3.5. Assume that x takes the following valuesxi with probabilitiesf;::
x, = 2 ,fl =0.25,x2 = 4 , f2=0.5,x,= 8, f3=0.25. Determinei,V,u, and y.
Problem 3.6. Assume that x takes the following values xi with probabilities
fi: x1 = - 1' f1 = $, x2 = 1,f2 = i. Determine 2, V, u, and y.
Problem 3.7. For the standard form of the normal distribution, determine the
probability that a value lies in the "bins" -0.141 to 0.141, 0.141 to
0.424 to 0.707, and 0.707 to 0.990.
Problem 3.8. For the standard form of the normal distribution, what is the
probability that a value is greater than 0.707; greater than -0.707?
Problem 3.9. For the standard form of the normal distribution, what are the
mean, standard deviation, and coefficient of variation of the correspond-
ing log-normal distribution?
Problem 3.10. For a log-normal distribution with i = 1 and cV= 0.5, for what
value of x is F(x) = 0.5?
Problem 3.11. For a log-normal distribution with 2 = 1 and cV= 1, for what
value of x is F(x) = 0.5?
Problem 3.12. Derive an expression for the coefficient of variation cVfor the
standard form of the Pareto distribution. What is its value for a = 3?
Problem 3.13. Derive an expression for the coefficient of variation cVfor the
exponential distribution. What is its value for v = 2?
Problem 3.14. The distribution function for the power-law mass distribution
given by (3.58) is
Assume that the maximum fragment size is r, and that v > 1. Determine
the mean fragment radius i and the variance V about this mean.
Problem 3.15. Consider a bar of unit length that has a probability f2 of being
fragmented into two bars of equal length i. The two smaller bars have
54 FRAGMENTATION
the same probability of being fragmented into bars of length a. Show that
this process leads to a fractal distribution with
Show that this result is equivalent to (3.80).

Problem 3.16. Consider a cube with a linear dimension h that is divided into
64 cubic elements with a dimension of h/4. The probability of fragmen-
tation is f,,. The smaller cubes have the same probability of being frag-
mented into cubes with dimensions of hI16. Show that this process leads
to a fractal distribution with
Show that this result is equivalent to (3.80).

Problem 3.17. Consider a model for fragmentation based on the Menger
sponge illustrated in Figure 2.4(a). Seven cubic elements are retained at
each scale. Determine N , for r , = h/3, N2 for r, = h/9, and N3 for r, =
h/27. What is the fractal dimension of the fragments?
Problem 3.18. Consider the fragmentation model illustrated in Figure 3.11.
Determine N, for r , = h/2,N, for r2 = h/4, and N3 for r, = h/8. What is the
fractal dimension of the fragments?
Problem 3.19. A model for fragmentation is constructed from a solid cube
that is divided into 27 equal-sized cubes at each scale and six of these
cubes are retained; i.e., the cubes in the center of each face are retained.

of a fractal model of
fragmentation. Four cubic
elements are retained at each
scale.
FRAGMENTATION 55
Determine N , for r , = h/3, N , for r, = h/9, and N, for r, = h/27. What is

the fractal dimension of the fragments?
Problem 3.20. A model for a porous medium is constructed from solid cubes
of density p, and size r,. At first order 26 of these cubes are used to con-
struct a cube with r , = 3r,; the central cube is missing. Determine the
porosity +, and the density p , . Continue the construction and determine
+, and p,. What is the fractal dimension?
Problem 3.21. A model for a porous medium is constructed from solid cubes
of density p, and size r,. At first order 21 of these cubes are used to con-
struct a cube with r , = 3r0; the cubes in the center of each face are miss-
ing (i.e., a Menger sponge with a central cube). Determine porosity +,
and the density p , . Continue the construction and determine 4, and p,.
What is the fractal dimension?
Chapter Four
SEISMICITY
AND TECTONICS
4.1. Seismicity
A variety of tectonic processes are responsible for the creation of topogra-

phy. These include discontinuous processes such as displacements on faults
and continuous processes such as folding. Tectonic processes are extremely
complex but they satisfy fractal statistics. Earthquakes are of particular con-
cern because of the serious hazard they present; earthquakes also satisfy
fractal statistics in a variety of ways. Seismicity is a classic example of a
complex phenomenon that can be quantified using fractal concepts (Tur-
cotte, 1993, 1994a, 1995).
According to plate tectonic theory, crustal deformation takes place at
the boundaries between the major surface plates. In the idealized plate tec-
tonic model plate boundaries are spreading centers (ocean ridges), subduc-
tion zones (ocean trenches), and transform faults (such as the San Andreas
fault in California). Relative displacements at subduction zones and trans-
form faults would occur on well-defined faults. Displacements across these
faults would be associated with great earthquakes such as the 1906 San Fran-
cisco earthquake. However, crustal deformation is more complex and is usu-
ally associated with relatively broad zones of deformation. Take the western
United States as an example: Although the San Andreas fault is the primary
boundary between the Pacific and the North American plates, significant de-
formation takes place as far east as the Wasatch Front in Utah and the Rio
Grande Graben in New Mexico. Active tectonics is occumng throughout the
western United States. Distributed seismicity is associated with this moun-
tain building. Even the displacements associated with the San Andreas fault
system are distributed over many faults. South of San Francisco and north of
Los Angeles the San Andreas fault has significant bends. Deformation asso-
ciated with these bends is responsible for considerable mountain building
and the 1956 Kern County earthquake, the 1971 San Fernando earthquake,
the 1989 Loma Prieta earthquake, the 1992 Landers earthquake, and the
1994 Northridge earthquake.
SEISMICITY AND TECTONICS 57
Although the crustal deformation in the western United States may ap-
pear to be complex, it does obey fractal statistics in a variety of ways. This is
true of all zones of tectonic deformation. We will first consider the fre-
quency-magnitude statistics of earthquakes. Several quantities can be used
to specify the size of an earthquake; these include the strain associated with
the earthquake and the radiated seismic energy. However, for historical rea-
sons the most commonly used measure of earthquake size is its magnitude.
Unfortunately, a variety of different magnitude scales have been proposed;
but to a first approximation, the magnitude is the logarithm of the energy ra-
diated and dissipated in an earthquake. Typically great earthquakes have a
magnitude m = 8 or larger. The 1992 Landers earthquake had a magnitude
m = 7.6 and was the largest earthquake in California since the great 1906 San
Francisco earthquake. The 1989 Loma Prieta earthquake had m = 7.1 and the
1994 Northridge earthquake had m = 6.6.
Many regions of the world have dense seismic networks that can moni-
tor earthquakes as small as magnitude two or less. The global seismic net-
work is capable of monitoring earthquakes that occur anywhere in the world
with a magnitude greater than about four. Various statistical correlations
have been used to relate the frequency of occurrence of earthquakes to their
magnitude, but the most generally accepted is the log-linear relation (Guten-
berg and Richter, 1954)
where b and a are constants, the logarithm is to the base 10, and N is the
number of earthquakes per unit time with a magnitude greater than m occur-
ring in a specified area. The Gutenberg-Richter law (4.1) is often written in
terms of N, the number of earthquakes in a specified time interval (say 30
years), and the corresponding constant a.
The magnitude scale was originally defined in terms of the amplitude of
ground motions at a specified distance from an earthquake. Typically the
surface wave magnitude was based on the motions generated by surface
waves (Love and Rayleigh waves) with a 20-s period, and the body wave
magnitude was based on the motions generated by body waves (P and S
waves) having periods of 6.8 seconds. The magnitude scale became a popu-
lar measure of the strength of earthquakes because of the logarithmic basis,
which allows essentially all earthquakes to be classified on a scale of 0-10.
Alternative magnitude definitions include the local magnitude and the mag-
nitude determined from the earthquake moment.
The frequency-magnitude relation (4.1) is found to be applicable over a
wide range of earthquake sizes both globally and locally. The constant b or
"b-value" varies from region to region but is generally in the range 0.8 <
b < 1.2 (Frohlich and Davis, 1993). The constant a is a measure of the re-
gional level of seismicity.
58 SEISMICITY AND TECTONICS
The magnitude is an empirical measure of the size of an earthquake. It

can be related to the total energy in the seismic waves generated by the
earthquake, Es, using the relation
log Es = 1.44m + 5.24 (4.2)
where Es is in Joules. The strain released during an earthquake is directly re-

lated to the moment M of the earthquake by the definition
where is the shear modulus of the rock in which the fault is embedded, A is
the area of the fault break, and 6e is the mean displacement across the fault
during the earthquake. The moment of an earthquake can be related to its
magnitude by
where c and d are constants. Kanamori and Anderson (1975) have estab-
lished a theoretical basis for taking c = 1.5. Kanamori (1978) and Hanks and
Kanamori (1979) have argued that (4.4) can be taken as a definition of mag-
nitude with c = 1.5 and d = 9.1 (M in joules). This definition is consistent
with the definitions of local magnitude and surface wave magnitude but not
with the definition of body wave magnitude. It is standard practice today to
use long-period (50-200 s) body and/or surface waves to directly determine
the scalar moment M and (4.4) is used to obtain a moment magnitude.
Kanamori and Anderson (1975) have also shown that it is a good ap-
proximation to relate the moment of an earthquake to the area A of the rup-
ture by
where a is a constant. Combining (4. l), (4.4), and (4.5) gives
with
bd b
log@=-
C
+ loga- -1oga
C
and (4.6) can be written as
In a specified region the number of earthquakes N per unit time with rupture
areas greater than A has a power-law dependence on the area. A comparison
-
with the definition of a fractal given in (2.6) with A r2 shows that the frac-
tal dimension of distributed seismicity is
Taking the theoretical relation c = 1.5 we have
Thus the fractal dimension of regional or worldwide seismic activity is sim-

ply twice the b-value. The empirical frequency-magnitude relation given in
(4.1) is entirely equivalent to a fractal distribution (Aki, 1981).
The Gutenberg-Richter frequency-magnitude relation (4.1) has been
found to be applicable under a great variety of circumstances. We will first
consider its validity on a worldwide basis. The worldwide number of earth-
quakes per year with magnitudes greater than m is given in Figure 4.1 as a
Figure 4.1. Worldwide

number of earthquakes per
year, N,with magnitudes
greater than m as a function
of m. The square root of
the rupture area A is also
given. The solid line is the
cumulative distribution of
moment magnitudes from the
Harvard Centroid Moment
Tensor Catalog for the period
January 1977 to June 1989
(Frohlich and Davis, 1993).
The dashed line represents
(4.1)with b = 1.11 ( D =
2.22) and a = 6 X 108 yr-1.
function of m. This data consists of 8719 earthquakes that occurred between

January 1977 and June 1989. The data are from the Harvard Centroid Mo-
ment Tensor Catalog (Dziewonski et al., 1989) and seismic moments have
been converted to moment magnitudes using (4.4). Comparisons of fre-
quency-magnitude statistics from various catalogs have been given by
Frohlich and Davis (1993). The worldwide data correlate with (4.1) taking
b = 1.11 (D = 2.22) and a = 6 X 108 yr-1. Also given in Figure 4.1 is the
equivalent characteristic length All2 obtained from (4.5) taking a = 3.27 X
106 Pa. The data given in Figure 4.1 can be used to estimate the frequency of
occurrence of earthquakes of various magnitudes on a worldwide basis. For
example, about 10 magnitude-seven earthquakes are expected each year and
a single magnitude-eight earthquake can be expected in a year.
The deviation of the data from the Gutenberg-Richter law (4.1) at mag-
nitudes less than m = 5.2 can be attributed to the resolution limits of the
global seismic network. Regional studies indicate that good correlations are
obtained down to at least m = 2. The deviation of the data from the Guten-
berg-Richter law at magnitudes greater than m = 7.5 is more controversial.
Clearly there must be an upper limit to the size of an earthquake; but the de-
viations in Figure 4.1 can be attributed either to a real deviation from the
correlation line or to the small number of very large earthquakes in the rela-
tively short time span considered. Frequency-magnitude statistics for older
earthquakes have been given by Abe (1981) for the period 1904 to 1980 and
by Purcaru and Berckhemer (1982) for the period 1920 to 1979. The results
given by Abe (198 1) support a systematic reduction of large earthquakes rel-
ative to the correlation curve in Figure 4.1, whereas the results given by Pur-
cam and Berckhemer (1982) support the direct extrapolation of the correla-
tion curve to larger earthquakes. Pacheco et al. (1992) have considered the
extrapolation problem in detail and favor a systematic reduction of the large
earthquakes.
There is a physical basis for a change in scaling for large earthquakes.
Smaller earthquakes are expected to be nearly equidimensional so that r -
Al12. However, the depth of large earthquakes is confined by the thickness of
-
the seismogenic zone, say 20 lun,whereas the length can increase virtually
-
without limit. Thus for large earthquake r A. The transition would be ex-
pected to occur for All2 = 25 km or m = 7.
It is of interest to consider the distribution of seismicity associated with
the relative velocity v across a plate boundary. We consider a specified
length of the fault zone, which also has a specified depth of rupture. Thus the
two plates are assumed to interact over an area Ap. The relative plate velocity
v and interaction area A are related to the rupture area A and mean slip dis-
placement 6, in an indiddual earthquake by
where the integral is camed out over the entire distribution of seismicity and
&is the number of earthquakes per unit time with magnitudes between rn
and rn + dm. The earthquake moment has been introduced from (4.3). We
hypothesize that a fractal distribution of seismicity accommodates this rela-
tive velocity. From (4.1) and (4.4) we have
and
Taking the derivative of (4.12) gives
Since c > b the integral diverges for large rn so that the maximum-magni-
tude earthquake mmaxmust be specified. This is the well-known observation
that a large fraction of the total moment and energy associated with seismic-
ity occurs in the largest events. Integration of (4.15) gives
A large value of regional strain vA implies either a high level of regional

P
seismicity (large a) or a large magnitude for the maximum-magnitude earth-
quake (large rnm,,).
This type of relation has been derived by several authors (Smith, 1976;
Molnar, 1979; Anderson and Luco, 1983) and has been used to estimate re-
gional strain (Anderson, 1986; Young and Coppersmith, 1985) and to com-
pare levels of seismicity with known strain rates (Hyndman and Weichert,
1983; Singh et al., 1983).
As a specific application, we consider the regional seismicity in south-
e m California. The frequency-magnitude distribution of seismicity in south-
ern California is given in Figure 4.2. The data from the southern California
earthquake network are for the period 1932 to 1994, the number of earth-
quakes per year N with magnitudes greater than rn are given as a function of
rn. Over the entire range of 4 < rn < 7.5 the data are in excellent agreement
with (4. l), taking b = 0.923 (D = 1.846) and a = 1.4 X 105 yr-1. In terms of
the linear dimension of the fault rupture, this magnitude range corresponds
to a linear size range 0.7 < All2 < 40 km.
Also included in Figure 4.2 is the value of N associated with great earth-
quakes on the southern section of the San Andreas fault. Dates for 10 large
earthquakes on this section of the fault have been obtained from radiocarbon
dating of faults, folds, and liquifaction features within the marsh and stream
deposits on Pallett Creek where it crosses the San Andreas fault 55 km north-
east of Los Angeles (Sieh et al., 1989). In addition to historical great earth-
quakes on January 9, 1857, and December 8, 1812, additional great earth-
+
quakes were estimated to have occurred in 1480 15, 1346 2 17, 1100 +
+ +
65, 1048 33,997 t 16,797 22,734 2 13, and 67 1 2 13. The mean re-
peat time is 132 years, giving N = 0.0076 yr-1. The most recent in the se-
quence of earthquakes occurred in 1857, and the observed offset across the
fault associated with this earthquake was 12 m (Sieh and Jahns, 1984). Sieh
(1978) estimates that the magnitude of the 1857 earthquake was m = 8.25.
Taking the values given above, the recurrence statistics for these large earth-
quakes are shown by the solid circle in Figure 4.2. An extrapolation of the
fractal relation for regional seismicity appears to make a reasonable predic-
Figure 4.2. Number of

earthquakes per year N
occurring in southern
California with magnitudes
greater than m as a function
of m. The solid line is the
data from the southern
California earthquake
network for the period
1932-1994. The straight
dashed line is the correlation
with (4.1) taking b = 0.923
(D = 1.846) and a = 1.4 X
105. The solid circle is the
observed rate of occurrence
of great earthquakes in
southern California (Sieh
et al., 1989).
tion of great earthquakes on this section of the San Andreas fault. Since this
extrapolation is based on the 40 years of data between 1932 and 1972, a rel-
atively large fraction of the main interval of 132 years, it suggests that the
value of a for this region may not have a strong dependence on time during
the earthquake cycle. This conclusion has a number of important implica-
tions. If a great earthquake substantially relieved the regional stress, then it
would be expected that the regional seismicity would systematically in-
crease as the stress increased before the next great earthquake. An alterna-
tive hypothesis is that an active tectonic zone is continuously in a critical
state and that the fractal frequency-magnitude statistics are evidence for this
critical behavior. In the critical state the background seismicity, small earth-
quakes not associated with aftershocks, have little time dependence. This
hypothesis will be discussed in Chapter 16. Acceptance of this hypothesis al-
lows the regional background seismicity to be used in assessing seismic haz-
ards (Turcotte, 1989b). The regional frequency-magnitude statistics can be
extrapolated to estimate recurrence times for larger magnitude earthquakes.
Unfortunately, no information is provided on the largest earthquake to be
expected.
An important question in seismology is whether the occurrence of large
plate-boundary earthquakes can be estimated by extrapolating the regional
seismicity as was done above for southern California. This is a subject of
considerable controversy. Some authors argue that the large earthquakes oc-
cur more often than would be predicted by an extrapolation.
To further consider the time dependence of regional seismicity (the time
dependence of a), we consider the frequency-magnitude statistics of the re-
gional seismicity in southern California on a yearly basis. Again the number
of earthquakes N in each year between 1980 and 1994 with magnitudes
greater than m are given in Figure 4.3 as a function of m. In general there is
good agreement with (4. I), taking b = 1.05 and a = 2.06 X 105 yr-1. The ex-
ceptions can be attributed to the aftershock sequences of the Whittier (1987),
Landers (1992), and Northridge (1994) earthquakes. Comparing the correla-
tion lines in Figures 4.2 and 4.3 shows that the correlation line in Figure 4.2
lies somewhat above those in Figure 4.3. This is because the data given in
Figure 4.2 include aftershocks. With aftershocks removed, the near uni-
formity of the background seismicity in southern California illustrated in
Figure 4.3 is clearly striking. This is strongly suggestive of a thermodynamic
behavior. We will return to this point in Chapter 16.
We now relate the seismicity in southern California to the relative veloc-
ity across the plate boundary. The data given in Figure 4.2 can be used to
predict the regional strain using (4.16). Substituting p. = 3 X 1010 Pa, b =
0.89, c = 1.5,d=9.1, v = 4 8 mmyr-',mmax=8.05, a n d a = 1.4 X lO5yr-1 we
find from (4.16) that Ap = 1.5 X lo4 km2. Taking the depth of the seismo-
genic zone to be 15 km, the length of the seismogenic zone corresponding to
this area is 730 km. This is about a factor of two larger than the actual length
of the San Andreas fault in southern California. This is reasonably good
agreement considering the uncertainties in the parameters. However, there
are two other factors that can contribute to this discrepancy.
Figure 4.3. The cumulative

number of earthquakes N
with magnitudes greater than
rn for each year between
1980 and 1994 is given as a
function of m; the region
considered is southern
California. (a) 1980-
1984; (b) 1985-1989; (c)
1990-1994. The straight-
line correlation is with the
Gutenberg-Richter relation
(4.1) with b = 1.05 and
a = 2.06 X 105 yr-1. The
relatively large numbers of
earthquakes in 1987,1992,
and 1994 can be attributed
to the aftershocks of the
Whittier, Landers, and
Northridge earthquakes,
respectively. If aftershocks
are excluded, the background
seismicity in southern
California is nearly uniform
in time.
1. Southern California is an area of active compressional tectonics. The

Transverse Ranges are in this region and are associated with the dis-
placements on the San Andreas fault. The strains associated with the
formation of the Transverse Range should be added to the strains as-
sociated with strike-slip displacements on the San Andreas fault
system.
2. South of San Bernardino, displacements on the San Andreas fault
are associated with small and moderate earthquakes; no great earth-
quakes are believed to occur on this section. With a maximum mag-
nitude of about seven, the expected level of seismicity would be
about a factor of five greater than with a maximum magnitude eight
earthquake. Thus a higher level of seismicity on this section of the
San Andreas could contribute to the high observed level.
Since the eastern United States is a plate interior, the concept of rigid plates
would preclude seismicity in the region. However, the plates act as stress
guides. The forces that drive plate tectonics are applied at plate boundaries.
The negative buoyancy force acting on the descending plate at a subduction
zone acts as a "trench pull." Gravitational sliding off an ocean ridge acts as a
"ridge push." Because the plates are essentially rigid, these forces are trans-
mitted through their interiors. However, the plates have zones of weakness
that will deform under these forces and earthquakes result. Thus earthquakes
occur within the interior of the surface plates of plate tectonics, although the
frequencies of occurrence are much lower than at plate boundaries. An ex-
M~ (c) Figure 4.3. (con?.)

ample was the three great earthquakes that occurred in the Memphis-
St. Louis (New Madrid, Missouri) seismic zone during the winter of 1811-
1812. Nuttli (1983), based on historical records, has estimated that the sur-
face wave magnitudes of these earthquakes were 8.5, 8.4, and 8.8, respec-
tively. This area remains the most active seismic zone in the United States
east of the Rocky Mountains. Based on both instrumental and historical
records Johnston and Nava (1985) have given the frequency-magnitude sta-
tistics for earthquakes in this area for the period 1816-1983. Their results are
given in Figure 4.4. The data correlate well with (4.1), taking b = 0.90 (D =
1.80) and a = 2.24 X 103 yr-1. Comparing the data in Figure 4.4 with the
data in Figure 4.2 indicates that the probability of having a moderate-sized
earthquake in the Memphis-St. Louis seismic zone is about 1/50 of the prob-
ability in southern California. Assuming that it is valid to extrapolate the
data in Figure 4.4 to larger earthquakes, a magnitude m = 8 would have a re-
currence time of about 7000 yr.
Although there is certainly a significant range of errors, the results given
above indicate that the measured frequency-magnitude statistics associated
with the Gutenberg-Richter frequency-magnitude relation (4.1) can be used
to assess seismic hazards. The regional b(D) and a values can be used to esti-
mate recurrence times for earthquakes of various magnitudes.
Figure 4.4. The cumulative

number of earthquakes per
year N occurring in the
Memphis-St. Louis (New
Madrid, Missouri) seismic
zone with magnitudes N
greater than m as a function yr"
of m (Johnston and Nava,
1985). The data are for the
period 1816-1983. The
open circles represent
instrumental data and the
solid circles historical data.
The dashed line represents
(4.1) with b = 0.90 (D =
1.80) and a = 2.24 X 103
4.2 Faults
There are two end-member models that give fractal distributions of earth-
quakes. The first is that there is a fractal distribution of faults and each fault
has its own characteristic earthquake. The second is that each fault has a
fractal distribution of earthquakes. Observations strongly favor the first hy-
pothesis. On the northern and southern locked sections of the San Andreas
fault, there is no evidence for a fractal distribution of earthquakes. Great
earthquakes and their associated aftershock sequences occur, but between
great earthquakes seismicity is essentially confined to secondary faults.
A similar statement can be made about the Parkfield section of the San
Andreas fault, where moderate-sized earthquakes occurred in 1881, 1901,
1924, 1934, and 1966. There is no evidence for a fractal distribution of
events on this section of the San Andreas fault. We therefore conclude that a
reasonable working hypothesis is that each fault has a characteristic earth-
quake and a fractal distribution of earthquakes implies a fractal distribution
of faults.
Although we can conclude that the frequency-size distribution of faults
is fractal, the fractal dimension is not necessarily the same as that for earth-
quakes. Equal fractal dimensions would imply that the interval of time be-
tween earthquakes is independent of scale. This need not be the case. Tec-
tonic models for a fractal distribution of faults have been proposed by King
(1983, 1986), Turcotte (1986b), King et al. (1988), and Hirata (1989a). Frac-
tal distributions of faults that give well-defined b-values have been proposed
by Huang and Turcotte (1988) and Hirata (1989b).
Before discussing the observational data on spatial distributions of
faults, we will discuss the definitions of faults, joints, and fractures. Frac-
tures are generally any crack in a rock. If there is a lateral offset across the
fracture, it is a fault; if there is no lateral offset, it is a joint. Because of the
grinding (comminution) effect of creating offsets on faults during earth-
quakes, a zone of brecciated rock (fault gouge) generally develops on the
fault. The larger the total offset on the fault, the wider the disrupted zone.
It is generally difficult to quantify the frequency-size distributions of
faults. This is because the surface exposure is generally limited. Many faults
are not recognized until earthquakes occur on them. Coal mining areas pro-
vide access to faults and fractures at depth. The cumulative distributions of
the number of faults N with lengths greater than r are given in Figure 4.5 for
two coal mining areas (Villemin et al., 1995). Correlations with the fractal
relation (2.6) are given with D = 1.6. Other compilations of the number-
length statistics of faults and comparisons with power-law correlations have
been given by Gudmundson (1987), Hirata (1989a), and Main et al. (1990).
Hirata et al. (1987) and Velde et al. (1993) found a fractal distribution of mi-
crofractures in laboratory experiments that stressed unfractured granite.
Systematic studies of the statistics of exposed joint and fault-trace pat-

terns over many orders of magnitude in length scale have been given by Bar-
ton (1995). Bedrock exposures were created on Yucca Mountain, Nevada, by
the removal of soil and debris creating exposures known as pavements. Bar-
ton (1995) mapped these exposures to obtain the two-dimensional distribu-
tion of joints and faults. His map of pavement 1000 is reproduced in Figure
4.6. This was located in the densely welded orange brick unit of the Topopah
Spring Member of the Miocene Paintbrush Tuff. He used the box-counting
method illustrated in Figure 2.8 to determine the fractal dimension of the frac-
ture traces. His result for the pavement illustrated in Figure 4.6 is given in
Figure 4.7; the mean fractal dimension is D = 1.7. Barton (1995) analyzed 17
fracture maps over scales ranging from 0.5 mm (microfractures) to 5000 km
(transform faults) and found good correlations with fractal statistics with
values of D ranging from 1.4 to 1.7. Davy et al. (1990) and Sornette et al.
(1993) carried out laboratory simulations of brittle crustal deformation and
found D = 1.7 -+ 0.1 for fracture trace networks.
The study of the distribution of faults and joints can also be carried out
by one-dimensional sampling. Drilling cores provide an excellent data base
for this type of study. The intersections of fractures with the core can be rep-

number of faults N,with
lengths r greater than r. The
boxes are measurements in
the Lorraine coal basin
and the circles from the
Vernejoul coal field
(Villemin et al., 1995).
Correlations with the fractal
relation (2.6) are given with
D = 1.6.
resented a s a series of points o n a line, the drill core. A one-dimensional b o x -

counting method is applied in direct analogy to the two-dimensional box
counting illustrated in Figure 2.8. The line is divided up into 2, 4, 8, 16, . . .
segments so that r = 1, 2,1 a,
1 1 1
8, G, . . . . In each case the number of line seg-
Figure 4.6. Map of the faults

and joints exposed on
pavement 1000, Yucca
Mountain, Nevada (Barton,
1995). This exposure was
located in the densely
welded orange brick unit of
the Topopah Spring Member
of the Miocene Paintbrush
Tuff.
ments that include points (fractures) N(r) is determined and log N(r) is plot-
ted against log r(or log llr). If a linear or near-linear dependence is found,
the slope gives the fractal dimension using (2.2).
Barton (1995) analyzed the distribution of gold-bearing, quartz-filled
fractures (veins) intersecting exploratory drill holes from tunnels in the Per-
severance Mine, Juneau, Alaska. His results for core hole 7- 18, using the one-
dimensional box-counting technique, are given in Figure 4.8. A good correla-
tion with the fractal relation (2.2) is obtained taking D = 0.59. For the 23 drill
holes studied by Barton (1995), good correlations with fractal statistics were
obtained, with D ranging from 0.41 to 0.62. Similar studies have been carried
out by La Pointe (1988). Velde et al. (1990, 1991), Ledesert et al. (1993),
Manning (1994), Boadu and Long (1994a, b), and Magde et al. (1995).
It should be emphasized that a wide variety of mechanisms are responsi-
ble for the formation of joints and faults and not all would be expected to
yield fractal distributions. Limitations of the fractal approach have been dis-
cussed by Harris et al. (1991) and Gillespie et al. (1993).
The fractal model for fragmentation illustrated in Figure 3.7 can also be
applied to tectonic fragmentation (Sammis et al., 1987). As the surface plates
of plate tectonics evolve in time, geometrical incompatibilities develop
(Dewey, 1975). Simple plate boundaries consisting of ocean ridges, subduction
zones, and transform faults cannot evolve in time without overlaps or holes de-
Figure 4.7. Statistics using

the box-counting algorithm
on the exposed fracture
network at Yucca Mountain,
Nevada (Barton, 1995), given
in Figure 4.6. A correlation
with (2.1) is used to obtain
the fractal dimension
D = 1.7.
veloping. The result is that plate interiors must deform to accommodate the
geometrical incompatibilities. Because of the weaker silicic rocks of the conti-
nental crust, and the many ancient faults pervading the continental lithosphere,
continental parts of surface plates deform much more readily than oceanic
parts. This can be easily seen at the boundary between the Pacific and North
American plates in the western United States. The adjacent continental North
American plate consisting of the western states deforms extensively whereas
there is little internal deformation in the adjacent oceanic Pacific plate.
Just as the comminution model can be applied to fragmentation, it can
also be applied to the deformation of the continental crust. The tectonic
forces break the continental crust into a fractal distribution of interacting
crustal blocks over a wide range of scales. The crustal blocks are bounded by
faults so that a fractal distribution of block sizes can be related to a fractal
distribution of faults. To illustrate this we consider the deterministic com-
minution model for fragmentation given in Figure 3.7. To fragment the sin-
gle zero-order block of size h requires three orthogonal faults (No = 3) of size
r, = h; the result is eight blocks of size hl2. Six of these eight blocks are fur-
ther fragmented; this requires N, = 3 X 6 = 18 faults of size r, = hl2. The re-
sult is 48 blocks of size hl4; 36 of these 48 blocks are further fragmented to
Figure 4.8. Distribution

of quartz-filled fractures
(veins) in core hole 7- 18,
Perseverance Mine, Juneau,
Alaska (Barton, 1995). The
number of line segments N
of length r that intersect
fractures along the core is
given as a function of r.
This is the one-dimensional
box-counting method. The
correlation is with the fractal
1
relation (2.2) taking D =
llr (rn-') 0.59.
give blocks of size h/8.This further fragmentation requires N2 = 3 X 36 =

108 faults of size r2 = h/4.From (2.2) we have D = In 6/ln 2 = 2.5850 for the
frequency-size distribution of faults in the volume of the block, which is
equal to the value for the frequency-size distribution of blocks obtained in
Chapter 3. However, for a statistical model we require cumulative statistics.
The cumulative frequency-size distribution of faults is essentially identical
to the cumulative statistics for fragment volumes given in Figure 3.8. Thus
we have D,= 2.60 for this model.
This deterministic fragmentation model can also be used to relate one-
dimensional and two-dimensional measurements to the three-dimensional
size-frequency distribution of faults. A surface projection of the comminu-
tion model given in Figure 3.7 is illustrated in Figure 4.9.Also illustrated in
this figure is the box-counting method applied to the deterministic fault expo-
sure. At zero-order there is one box with ro = h and it covers faults, No = 1. At
first order there are four boxes with r , = h/2 and three of them cover faults,
N , = 3. At second order there are 16 boxes with r, = h/4 and nine of them
cover faults, N2 = 9. From (2.2) we find that D = In 3/ln 2 = 1.5850. Noting that
Figure 4.9. (a) Illustration

of the surface exposure of
the deterministic fractal
fragmentation model given
in Figure 3.7. The lines
represent a fractal
distribution of linear faults
that separate a fractal
distribution of blocks.
method applied to this

exposure. (b) At zero order
the one box with r = h covers
faults, No = 1. (c) At first
order three of the four boxes
with r = hl2 cover faults,
N, = 3. (d) At second order
nine of the 16 boxes with r =
h/4 cover faults, N, = 9. And
from (2.2) we have D = In
3lln 2 = 1S8SO.
SEISMICITY AND TECTONICS
we see that the fractal dimension of the cross section, D2 = In 3lln 2, is re-
lated to the fractal dimension of the original construction from Figure 3.3.,
D,= In 6 t h 2, by
Thus we expect, for a statistically self-similar distribution of fault lengths,

that the two-dimensional fault dimension of a surface D, will be related to
the three-dimensional fractal dimension D,by (4.18).
We next consider the fractal dimension of one-dimension transects (drill
holes) through the three-dimensional deterministic fracture model given in
Figure 3.7. In this case different transects intersect different fault distribu-
tions. The four different distributions are given in Figure 4.10. For each of
these we carry out a one-dimensional "boxw-countingmethod using line seg-
ments of length h, h12, and hl4. For the four distributions we find No = 1, 1,
1, 1 so on average No = 1, we find N , = 1, 1 , 2 , 2 so on average N , = 1.5, and
we find N2 = 1 , 2 , 2 , 4 so on average N2 = 2.25. Using the average values we
find from (2.2) that D = In 1Slln 2 = 0.5850. Noting that
we see that the average fractal dimension of the linear transects, D l= In

1.5Iln 2 is related to the fractal dimensions of the cross sections D,= In 31ln
2 and the original construction D,= In 6 t h 2 by

of four different one-
Thus we expect for a statistically self-similar distribution of faults that the dimensional transects
(drill holes) across the
one-dimensional fractal dimension from a bore hole D l will be related to the deterministic fractal
three-dimensional fractal dimension D, by (4.20). These results are quite fragmentation model given
consistent with the values given in Figures 4.7 and Figure 4.8 where fractal in Figure 3.7. In each case
dimensions D, = 1.7 and D,= 0.59 were obtained. Barton (1995) found that the one-dimensional box-
counting method is applied.
For transect (a) we have No =
1 (rO= h), N , = 1 (ro = h/2),
and N2 = 1 (ro = h14); for
transect (b) we have No = 1,
N , = 1 , N2 = 2 ; for transect
(c) we have No = 1, N , = 2,
N, = 2, and for transect (d)
we have No = 1, N , = 2, N, =
4. On average No = 1 , N , =
1.5, N, = 2.25 so that from
(2.2) we have D = In 1.51
In 2 = 0.5850.
a three-dimensional vein network at the Perseverance mine sampled in one

dimension with drill cores and in two dimensions by surface mapping
yielded fractal dimensions of 0.5 and 1.48, respectively. Watanabe and Taka-
hashi (1995) have proposed a more general statistical approach to the deter-
mination of the three-dimension fractal distribution of faults from a one-
dimensional data base.
In many oil fields the permeability of the rock is dominated by faults and
joints. To determine the feasibility of producing from a field, accurate esti-
mates of the permeability are required. The application of fractal statistics as
discussed above can provide the basis for making such estimates. Data on
fault and joint spacings intersecting a single well can be used to estimate the
full three-dimensional permeability of a field.
Although the discussion given above has been based on the applicability
of a comminution-based fragmentation model, it should be noted that Sor-
nette et al. (1990) and Sornette and Davy (1991) have offered an alternative
model for fractal fracture networks. These authors have suggested that the
networks are random growth networks similar to diffusion-limited aggrega-
tion (see Chapter 8). Termonia and Meakin (1986) have given another ap-
proach to fractal fracturing.
Taking the fractal dimension for the distribution of faults in a volume of
rock to be D,= 2.6, the number of faults with a characteristic linear dimen-
sion greater than r, in a given area, scales with r according to
Similarly we assume that for earthquakes De= 2 so that the number of earth-
quakes per unit time, in a given area, with a characteristic rupture size
greater than r scales with r according to
The average interval between earthquakes T~ on a fault with a characteristic

dimension r is given by
Thus the interval between earthquakes on a specified fault is longer for

smaller faults. This is generally consistent with observations. If we further
assume that faults remain active for a time 7,then the total displacement 6 on
a fault of scale r is given by
where ae is the displacement in a single event. However
and substitution of (4.23) and (4.25) into (4.24) gives
Walsh and Wattersen (1988) and Marrett and Allmendinger (1991) have
compiled measurements of the dependence of total displacement on a fault 6
as a function of fault length r and concluded that there is a power-law (frac-
tal scaling). The results obtained by Marrett and Allmendinger (1991) are
given in Figure 4.11. Data from a wide variety of tectonic environments are
included. Although there is considerable scatter, a reasonably good correla-
tion with (4.26) is found. It should be emphasized that this correlation must
be to some extent fortuitous since T, is unlikely to be a constant in different
tectonic settings. Also, Scholz and Cowie (1990), Cowie and Scholz (1992a,
b), Scholz et al. (1993), and Dawers et al. (1993) conclude that 6 - r with an
Figure 4.11. Dependence

of total fault displacement
on fault length; data from
Marrett and Allmendinger
(1991). The straight line
correlation is with (4.26),
-
8 rl.6.
additional parameter, the critical shear stress for fault propagation. These au-
thors correlate fault displacement and length in individual tectonic environ-
ments and find for each environment a reasonably good correlation with 6 r. -
They argue that it is misleading to include data from a variety of tectonic en-
vironments. In addition, Gillespie et al. (1992) find a universal power-law
correlation between fault width and total displacement. Jackson and Sander-
son (1992) and Pickering et al. (1994) have concluded that in several exam-
ples the number of faults with displacements greater than a specified value
satisfy fractal statistics with D = 0.7-1.4.
Sedimentary basins are often formed by horizontal extension on suites
of normal faults. The horizontal extension thins the continental crust, result-
ing in the subsidence of the surface and the deposition of a sedimentary pile
on the subsiding "basement." A common observation is that the amount of
extension associated with "visible" normal faults (for example, on seismic
reflection profiles) is significantly less than the amount of extension associ-
ated with the observed crustal thinning. Typically only 40-70% of the re-
quired extension can be associated with the larger faults on which displace-
ments can be determined.
Using a fractal distribution for the number of faults as a function of size and
the displacements of these faults as described above, the displacements on small
unobserved faults can be determined from the displacements on the larger faults.
Walsh et al. (1991) and Marrett and Allmendinger (1992) have argued that this
approach can explain the discrepancy. The total strain E in a volume V, is related
to the number of faults N,, fault area A,, and total fault displacement 6 by
Taking A, - r2 along with (4.21) and (4.26) we have
If these statistics are valid in a region, the larger faults dominate in terms of
regional strain, but the smaller faults do make a significant contribution.
4.3 Spatial distribution of earthquakes
The box-counting method in three dimensions has been applied to the spatial
distribution of earthquake aftershocks by Robertson et al. (1995). After-
shocks are particularly well located because extensive arrays of seismome-
ters have been deployed following the main shock. These authors considered
the aftershock sequence of the m = 6.1 Joshua Tree earthquake of April 23,
1992 (2600 events in a 20 X 20 X 19 km volume in 160 days), and the after-
shock sequence of the m = 6.2 Big Bear earthquake of June 28, 1992 (818
events in a 20 X 20 X 17 km volume in 375 days). The spatial distributions
of these aftershocks are given in Figure 4.12(a, b).
The numbers of cubes occupied by one or more earthquakes are given as
a function of the cube size in Figure 4.12(c) for the two aftershock se-
quences; cubes with linear dimensions between 500 m and 20 km were used.
The data are in quite good agreement with (2.2) taking D = 2. A fractal di-
mension of two would be expected if the earthquakes lie on a plane; how-
ever, there is considerable three-dimensional structure to the aftershock se-
Figure 4.12. Box-counting

method of cluster analysis
applied to the three-
dimensional distributions of
the aftershocks of the Joshua
Tree earthquake (JTS) and the
Big Bear earthquake (BBS).
The spatial distributions of
these earthquakes are given
in (a) and (b). The number of
cubes N in which one or
more earthquakes occur is
given as a function of the
linear dimension of the cube
r in (c). The straight line is
the fractal correlation (2.2)
with D = 2.
quences. This led Robertson et al. (1995) to suggest that the earthquakes
form the "backbone" of a percolation cluster. The "backbone" of a three-
dimensional percolation cluster has a fractal dimension near two. A detailed
discussion of percolation clusters and the meaning of the "backbone" will be
given in Chapter 15. The box-counting technique has been applied to both
the temporal and two-dimensional spatial distribution of earthquakes in
Japan by Bodri (1993).
4.4Volcanic eruptions
We next turn to volcanic eruptions. It is considerably more difficult to

quantify a volcanic eruption than it is to quantify an earthquake. There are
a variety of types of eruption, and the various types are quantified in dif-
Figure 4.13. Number of

volcanic eruptions per year
N=with a tephra volume
greater than V as a function
of V for the period
1975-1 985 (squares) and for
the last 200 years (circles)
(McClelland et al., 1989).
The line represents the
correlation with (2.6) taking
D = 2.14.
ferent ways. Some eruptions produce primarily magma (liquid rock) while
others produce primarily tephra (ash). Utilizing the volume of tephra as a
measure of size McClelland et al. (1989) have published frequency-vol-
ume statistics for volcanic eruptions. Their results for eruptions during the
period 1975-1985 and for historic (last 200 years) eruptions are given in
Figure 4.13. The number of eruptions with a volume of tephra greater than
a specified value is given as a function of the volume. A reasonably good
correlation is obtained with the fractal relation (2.6) by taking D = 2.14. It
appears that volcanic eruptions are scale invariant over a significant range
of sizes.
A single volcano can produce eruptions with a wide spectrum of sizes.
Also, volcanoes have a wide spectrum of sizes. The circumstances that de-
termine the volume of tephra in an eruption are poorly understood. Thus
models that would provide an explanation of the observed value of D are not
available.
Problems
Problem 4.1. Determine Es, M, A , and se for an m = 7 earthquake (take c =

1.5, d = 9.1, a = 3.27 X 106Pa, p = 3 X 1OlOPa).
Problem 4.2. Determine Es, M, A , and se for an m = 6 earthquake (take c =
1.5, d = 9.1, a = 3.27 X 106Pa, p = 3 X IOIOPa).
Problem 4.3. On a worldwide basis how many magnitude-six earthquakes
are expected in a year?
Problem 4.4. In a region, the recurrence interval -re for a magnitude-six
earthquake is 18 months; if b = 0.9 what is the recurrence interval T~ for
a magnitude-seven earthquake?
Problem 4.5. In a region, the recurrence interval T~ for a magnitude-five
earthquake is 10 years; if b = 1 what is the recurrence interval r efor a
magnitude-seven earthquake?
Problem 4.6. In a subduction zone the length of the seismogenic zone is
1000 km and its depth is 30 km. The convergence velocity is 100 mm
yr-1. (a) Determine a if b = 1, c = 1.5, d = 9.1, p= 3 X IO'OPa, and
mmax= 8.5. (b) Determine the recurrence time for the magnitude-8.5
earthquake.
Problem 4.7. The length of a seismogenic zone on a strike-slip fault is 100
km and its depth is 15 km. (a) Determine a if b = 1, c = 1.5, d = 9.1, p =
3 X 10"JPa, v = 50 mm yr-1, and mmax= 6.2. (b) Determine the recur-
rence time for the magnitude-6.2 earthquake.
Problem 4.8. The characteristic earthquake of magnitude-seven on a fault
has a recurrence interval of 200 years; using (4.23), what is the recur-
rence time for a characteristic earthquake of magnitude six? Take c =
1.5, d = 9.1.
Problem 4.9. The characteristic earthquake of magnitude six on a fault has a

recurrence interval of 400 years; using (4.23), what is the recurrence
time for a characteristic earthquake of magnitude four? Take c = 1.5 and
d = 9.1.
Problem 4.10. Assume that the distribution of rupture zone sizes on a fault is
modeled by the Sierpinski carpet, Figure 2.3(d). (a) Determine N, for
r, = h/3, N2 for r2 = h/9, and N3 for r3 = h/27. Further assume that the
earthquake displacement associated with each rupture zone is propor-
tional to its linear dimension. (b) If the slip velocities on all the zones are
equal, determine the b-value for the earthquakes.
Problem 4.11. Assume that the distribution of rupture zone sizes on a fault is
modeled by the fractal distribution given in Figure 4.9(a). Further as-
sume that the earthquake displacement associated with each rupture
zone is proportional to its linear dimension. If the slip velocities on all
the zones are equal, determine the b-value for the earthquakes.
Problem 4.12. Consider the fragmentation model given in Figure 3.11 as a
model for the distribution of fracture and joints in the earth's crust. What
are the fractal dimensions for a two-dimensional surface exposure and
for a one-dimensional transect?
Problem 4.13. Ash eruptions with a volume greater than 1000 km3 are ex-
pected to have a profound influence on the global climate. What is the
expected recurrence interval for such eruptions?
Problem 4.14. The Pinatuba, Philippines, eruption had a tephra volume of
18 km3. What is the expected recurrence interval for this eruption?
Chapter Five
ORE GRADE
ANDTONNAGE
5.1 Ore-enrichment models
Statistical treatments of ore grade and tonnage for economic ore deposits
have provided a basis for estimating ore reserves. The objective is to deter-
mine the tonnage of ore with grades above a specified value. The grade is de-
fined as the ratio of the mass of the mineral extracted to the mass of the ore.
Evaluations can be made on either a global or a regional basis. Much of the
original work on this problem was carried out by Lasky (1950). He argued
that ore grade and tonnage obey log-normal distributions.
Other authors, however, have suggested that a linear relation is obtained
if the logarithm of the tonnage of ore with grades above a specified value is
plotted against the logarithm of the grade. The latter is a fractal relation. A
fractal relation would be expected if the concentration mechanism is scale
invariant. Many different mechanisms are responsible for the concentrations
of minerals that lead to economic ore deposits. Probably the most widely ap-
plicable mechanisms are associated with hydrothermal circulations.
We first consider two simple models that illustrate the log-normal and
power-law distributions for tonnage versus grade. De Wijs (1951, 1953) pro-
posed the model for mineral concentration that is illustrated in Figure 5.l(a).
In this model an original mass of rock Mo is divided into two equal parts
each with a mass M , = M d 2 . The original mass of the rock has a mean min-
eral concentration Co, which is the ratio of the mass of mineral to mass of
rock. As in Chapter 3 we refer to this mass as a zero-order cell. It is hypothe-
sized that the mineral is concentrated into one of the two zero-order ele-
ments such that one element is enriched and the other element is depleted.
The zero-order elements then become first-order cells, each of which is di-
vided into two first-order elements with mass M , = M d 4 .
The mean mineral concentration in the enriched zero-order element C,,
is given by
82 ORE GRADE AND TONNAGE
where 4, is the enrichment factor. The first subscript on C refers to the order
of cell being considered. The second subscript refers to the amount of en-
richment: the lower the number the more the enrichment and the higher the
concentration. The subscript on the enrichment factor refers to the fact that
each cell is divided into two equal elements; the enrichment factor 4, is
greater than unity since C,, must be greater than C,,.A simple mass balance
shows that the concentration in the depleted zero-order element is

two models for the
concentration of economic
ore deposits. In both models
a mass of rock is first
divided into two equal parts,
then four equal parts, etc. (a)
De Wijs (1951, 1953)
proposed a successive
concentration of minerals
into smaller and smaller
volumes. Four orders of
concentration are illustrated.
At each order one-half of
each cell is enriched by the
ratio 42and the other half is
depleted by the ratio 2 -
4,. This model gives a
binominal distribution for
tonnage versus grade and in
the limit of very small
volumes gives a log-normal
relation. (b) Turcotte ( 1 9 8 6 ~ )
proposed a similar model,
but with the further
concentration limited to the
highest-grade ores. This
leads to a power-law (fractal)
distribution of tonnage
versus grade.
ORE GRADE AND TONNAGE 83
The enrichment factor must be in the range 1 < 4, < 2. This model is illus-
trated in Figure 5.l(a). The process of concentration is then repeated at the
next order as illustrated in Figure 5.l(a). The zero-order elements become
first-order cells and each cell is again divided into two elements of equal
mass M, = Md4. The mineral is again concentrated by the same ratio into
each first-order element. The enriched first-order element in the enriched
first-order cell has a concentration
The depleted first-order element of the enriched first-order cell and the en-
riched first-order element of the depleted first-order cell both have the same
concentrations:
The depleted first-order element of the depleted first-order cell has a concen-
tration
This result is also illustrated in Figure 5.l(a) along with two higher-order
cells. This model gives a binomial distribution of ore grades, and in the limit
of infinite order reduces to the log-normal distribution given in (3.29). The
resulting distribution is not scale invariant; the reason is that the results are
dependent on the size of the initial mass of ore chosen and this mass enters
into the tonnage-grade relation. We will show in Chapter 6 that the resulting
distribution is a multifractal.
Cargill et al. (1980, 1981) and Cargill (1981) disagreed with the loga-
rithmic dependence and suggested that a linear relationship is obtained if the
logarithm of the tonnage is plotted against the logarithm of the mean grade.
A simple model that gives this result was proposed by Turcotte ( 1 9 8 6 ~ and
)
is illustrated in Figure 5.l(b). This model follows very closely the model dis-
cussed above. Again, an original mass of rock M, is divided into two parts
each with a mass M, = M,/2, and it is hypothesized that the mineral is con-
centrated into one of the two zero-order elements so that (5.1) and (5.2) are
applicable. However, at the next step only the enriched element is further
fractionated. The problem is renormalized so that the enriched element is
treated in exactly the same way at every scale (order). This results in a frac-
tal (scale-invariant) distribution. The concentration of ore into one or the
two elements in the enriched first-order cell results in the concentrations
given by (5.3) and (5.4). However, the depleted first-order cell continues to
have the concentration given by (5.2). This result is illustrated in Figure

5.l(b) along with two higher-order cells.
The results given in Figure 5.l(b) can be generalized to the nth order
with the result
where C,, is the mean ore grade associated with the mass
Taking the natural logarithms of (5.6) and (5.7) gives
and
The elimination of n between (5.8) and (5.9) gives
-
With the density assumed to be constant, M r3, where r is the linear dimen-
sion of the ore deposit considered, and we have
Comparison with (2.6) shows that this is a power-law or fractal distribution

with
Since the allowed range for 4, is 1 < 4, < 2, the allowed range for the frac-
tal dimension is 0 < D < 3. To be fractal the distribution must be scale in-
variant. The scale invariance is clearly illustrated in Figure 5.l(b). The con-
centration of ore could be started at any order and the same result would be
obtained. The left half at order two looks like order one, the left half at order
three looks like order two, etc. This is not true for the distribution illustrated
in Figure 5.1 (a).
We now generalize this model so that the original mass of rock is di-
vided into two parts, but the masses of the two parts are not equal. The mass
of the enriched element M , , is related to the original mass Mo by
and the mass of the depleted element M I , is given by
The mass ratio a can take on the range of values 1 < a < m. The concentra-
tion ratio is defined as before and is the ratio of the concentration in the en-
riched element C , , to the reference concentration C,.
A mass balance shows that the concentration in the depleted element C , , is

given by
The enriched zero-order element becomes a first-order cell; this cell is di-
vided into two parts with the enriched part having a mass
M21= 7- M0
Mll
a2
And the concentration in this enriched mass is
The above results are generalized to the nth order to give

and
Taking the natural logarithms of (5.19) and (5.20) yields
and
Elimination of n between (5.21) and (5.22) gives
Comparison with (2.6) shows that this is a power-law or fractal distribution

with
+,
It is clear from (5.25) that depends upon a.It is easy to show that this is
reasonable. The case a = 2 was considered above. For a = 8 we have from
(5.25)
We now show that (5.26) is entirely equivalent to (5.12). The first-order con-
centration into one-eighth of the original mass, $, must be equivalent to
three orders of the concentration into one-half the original mass, 4,. Thus we
can write
It follows that
Thus (5.26) is equivalent to (5.12) and can be derived independently of the

mass ratio chosen.
Two classic models for the generation of ore deposits lead directly to the
fractal distribution derived above. The first is the chromatographic model
and the second is the Rayleigh distillation model (Allbgre et al., 1975; All&-
gre and Lewin, 1995). The chromatographic model can be directly applied to
the dissolution-reprecipitation process that occurs during fluid percolation
through cooling porous intrusions. Such a mechanism has been applied to
explain hydrothermal, epithermal, and skarn mineral deposits.
Consider a segment of the crust with an average trace element concen-
tration Co. Fluid circulation occurs enriching a mass fraction a-I with an en-
richment factor +a so that the new concentration is given by (5.15). Assume
that this enrichment process again affects the already enriched region with
the same enrichment factor. The doubly enriched mass fraction is a - 2 and
the concentration is given by (5.18). The process is repeated to produce
successive enrichments of smaller and smaller segments of the crust. This is
the chromatographic model and is identical to the fractal model described
above.
We next consider Rayleigh distillation. This is the classic model used in
geochemistry to explain the extreme enrichment of trace elements observed
in some crystalline rocks (Allkgre and Minster, 1978). The basic model con-
siders the solidification of a magma to form the crystalline rock; trace ele-
ments are partitioned between the remaining magma and the crystallizing
solid. If a trace element is incompatible with the crystalline solid, the resid-
ual magma will become progressively more enriched.
If an incremental mass of magma 6 M crystallizes, the incremental mass
of mineral 6Mmtransferred from the magma to the solid is given by
where M is the mass of magma, Mm is the mass of the mineral in the magma,
and KR is the solid-liquid partition coefficient. If KR < 1 the remaining
magma is systematically enriched. The allowed range of values for the
solid-liquid partition coefficient is 0 I KR I 1 . If KR = 1 there is no enrich-

ment of the remaining magma and the concentration of the mineral is con-
stant. If K , = 0 the concentration of the mineral in the solid is zero until the
melt contains only the mineral.
We can write (5.29) as a differential equation in the form
Integrating with the initial condition that Mm = Mmowhen M = Mo gives
The concentration of the mineral in the enriched residual magma Cmand the
concentration of the mineral in the original magma Cmoare given by
and
Substitution of (5.32) and (5.33) into (5.31) yields
This is the classic result for Rayleigh distillation. If KR = 1 then Cm/Cmo= 1

and there is no enrichment as expected. If KR = 0 then Cm= 1 when M/Mo =
Cmo;the melt contains only the mineral. Note that Cmis constrained to the
range CmoI CmI 1.
We now show that this classic result for Rayleigh distillation is entirely
equivalent to the fractal relation (5.24). In terms of the fractal model consid-
ered above, the enriched element has a mass M - 6M with concentration C , ,
and the depleted element has a mass 6M with concentration C,,. Thus we
have Co = MmIM and C = 6Mm/6M so that the definition of the solid-
1.2
liquid partition function given in (5.29) becomes
Substitution from (5.16) gives
c$a=a+(l -a)KR
Using this result the power in (5.24) is given by
We assume that at each renormalization the fraction of solid is very small so

that
with E < < 1. In this limit we can write
In [ a + (1 - a)KR] = In [I +~ ( -1 K,)] = ~ ( -1 K,) (5.41)
And the substitution of (5.42) into (5.24) gives (5.34). In the limit a + 1 the
fractal model is identical to Rayleigh distillation. Furthermore, the substitu-
tion of (5.42) into (5.25) gives
The fractal dimension of the ore deposit is simply related to the solid-liquid
partition coefficient of the Rayleigh distillation process. For the allowed
range of values for KR,0 I KR I 1, the allowed range for D is 0 ID I 3. In
the limit KR + 1 there is little enrichment and D + 0; in the limit KR + 0
there is very strong enrichment and D + 3.
5.2 Ore-enrichment data
In each of the enrichment steps in our fractal model the concentration C,, is
the mean concentration in the mass of ore M,,. For applications to actual ore
deposits we generalize the fractal relation between ore grade and tonnage to
where M is the mass of the highest grade ores, which have a mean concentra-
tion c. The reference mass Mo is the mass of rock from which the ore was
derived, which has a mean concentration Co.
As in the previous examples of naturally occurring fractal distributions,
there are limits to the applicability of (5.44). The lower limit on the ore grade
is clearly the regional background grade C , that has been concentrated.
However, there is also an upper limit: the grade C cannot exceed unity,
which corresponds to pure mineral.
The entire subject of tonnage-grade relations has been reviewed by Har-
ris (1984). There is clearly a controversy in the literature between Lasky's
law, which gives a log-normal dependence of tonnage on grade, and the
power-law or fractal dependence. Lasky (1950) and Musgrove (1965) have
argued in favor of the log-normal relation. On the other hand, Cargill et al.
(1980, 1981) and Cargill (1981) have argued in favor of the power-law de-
pendence. These authors based their analyses on records of annual produc-
tion and mean grade. Their results for mercury production in the United
States are given in Figure 5.2. The cumulative tonnage of mercury mined
prior to a specified date is divided by the cumulative tonnage of ore from
which the mercury was obtained to give the cumulative average grade. The
data points in Figure 5.2 represent the five-year cumulative average grade
(in weight ratio) versus the cumulative tonnage of ore. Using Bureau of
Mines records Cargill et al. (1981) found that the total amount of mercury
mined between 1890 and 1895 was Mm, and the tonnage of ore from which
this mercury was obtained was M , ; the mean grade for this period was c, =
Mm,IM,.The cumulative amount of mercury mined between 1890 and 1900
was Mm, and the cumulative tonnage of ore from which the mercury was
mined was M,; the mean cumulative grade for this period was c, = Mm21M2.
These computations represent the two data points farthest to the left in Fig-
ure 5.2. The other data points represent the inclusion of additional five-year
periods in the computations. Cargill et al. (1980, 1981) and Cargill (1981)
further hypothesized that the highest-grade ores are usually mined first so
that the cumulative ratio of mineral tonnage to ore tonnage at a given time is
a good approximation to the mean ore grade of the highest-grade ores. Thus
it is appropriate to compare their data directly with the fractal relation (5.44).
Excellent agreement is obtained taking D = 2.01. This is strong evidence
that the enrichment processes leading up to the formation of mercury de-
posits are scale invariant.
It is also of interest to introduce a reference concentration of mercury
into the fractal relation. An appropriate choice is the mean measured concen-
tration in the continental crust. The mean crustal concentration of mercury as
given by Taylor (1964) is c, = 8 X 10-8 (0.08 ppm). Using this value in
(5.44) we find that the correlation line in Figure 5.2 is given by
with M in kilograms. According to the fractal model the mercury ore in the
United States has been concentrated from continental crust with a mass
M, = 4.05 X 10'7 kg. Assuming a mean crustal density of 2.7 X 103 kg m-3,
the mercury resources of the United States were concentrated from an origi-
nal crustal volume of 1.5 X 105 km3. Since the total crustal volume of the
United States is approximately 2.7 X 108 km3, the source volume for the
mercury deposits is about 0.05 percent of the total. It is concluded that the
Figure 5.2. Dependence of

cumulative ore tonnage M on
mean grade C for mercury
production in the United
States (Cargill et al., 1981).
Correlation with (5.44) gives
a fractal dimension D = 2.01.
processes responsible for the enrichment of mercury ore deposits are re-
stricted to a relatively small fraction of the crustal volume.
It is seen from Figure 5.2 that the cumulative production of 1.2 X 108 kg
of mercury has been obtained from 2 X 1010kg of ore of volume 7.4 X 106 m3.
Since the source region has a volume of 1.5 X 105 km3, the fraction of the
source region that has been mined is only 5 X 10-8. The results given in Fig-
ure 5.2 can also be used to determine how much mercury ore must be mined
in the future to produce a specified amount of mercury. To produce the next
1.2 X lo8 kg of mercury will require the processing of about 1.6 X 1011 kg
of ore.
Using production records of lode gold, Cargill (1981) gave cumulative
tonnage-grade data for lode gold production in the United States. The data
points in Figure 5.3 represent the five-year cumulative average grade versus
the cumulative tonnage of ore for the period 1906-1976. A good correlation
with the fractal relation (5.44) is obtained taking D = 1.55. This fractal di-
mension is somewhat less than the value obtained for mercury, indicating a
smaller enrichment factor.
Again, the mean crustal concentration is introduced as a reference con-
centration. Taking co = 3 X 10-9 (3 ppb) (Taylor and McLennan, 1985) for
gold, we find the correlation line in Figure 5.3 is given by

mean grade for lode gold
States (Cargill, 1981).
with M in kilograms. According to the fractal model the lode gold in the
United States has been concentrated from a continental crustal mass of 3 X
1018 kg. Assuming a mean crustal density of 2.7 X 103 kg m-3, the gold was
concentrated from a crustal volume of 106 km3 or about 0.4 percent of the to-
tal crustal volume.
Using copper production records in the same way, Cargill et al. (1981)
have also given cumulative grade-tonnage data for copper production in the
United States. Their results are given in Figure 5.4. The cumulative grade is
again given as a function of cumulative ore tonnage at five-year intervals.
The data obtained prior to 1920 fall systematically low compared to the later
data. Cargill et al. (1981) attributed this systemic deviation from a fractal
correlation to the adoption of an improved metallurgical technology for the
extraction of copper in the 1920s. A smaller fraction of the available copper
was extracted prior to this time so that the data points are low. It is again ap-
propriate to compare these data with the fractal relation (5.44). Assuming the
early data to be systematically low, excellent agreement is obtained taking
D = 1.16. This fractal dimension is almost a factor of two less than the frac-
tal dimension obtained for mercury ore. This indicates that the applicable en-
richment processes concentrate copper less strongly than they do mercury.
We again relate the fractal relation for the enrichment to the mean crustal
concentration. The mean concentration of copper in the upper crust as given
by Taylor and McLennan (1981) is C, = 2.5 X 10-5 (25 ppm). Using this
value in (5.44), we find that the correlation line in Figure 5.4 is given by

mean grade for copper
States (Cargill et al., 1981).
with M in kilograms. According to the fractal model, the copper ore in the
M,, = 3.22 X 1019 kg. Assuming a mean upper crustal density of 2.7 X 103
kg m-3, the copper resources of the United States were concentrated from an
original crustal volume of 1.19 X 107 km3. This represents about 4 percent
of the total crustal volume of the United States. The crustal volume from
which copper is enriched is nearly 100 times larger than the volume from
which mercury is enriched. It is concluded that the processes responsible for
the enrichment of copper are much more widely applicable than those for
mercury.
As our final example we consider data on the relationship between cu-
mulative tonnage and grade for uranium in the United States. Data for the
preproduction inventory as given by the US Department of Energy have
been tabulated by Harris (1984, p. 228) in terms of cumulative tonnage and
the average grade of this tonnage; these data are tabulated in Figure 5.5. The
high-grade data are based on production records and the lower-grade data
are based on estimates of reserves. The higher-grade data are in excellent
agreement with the fractal relation (5.44) taking D = 1.48. Thus the enrich-
ment of uranium is intermediate between the enrichment of copper and
mercury. The predicted cumulative tonnage at lower grades falls below the
extrapolation of the fractal relation; this can be attributed to an underestima-
tion of the preproduction inventory at low grades.
It is again instructive to relate the fractal relation for the enrichment of
uranium to the mean crustal concentration. The mean concentration of ura-
nium in the upper crust as given by Taylor and McLennan (1981) is C, =

mean grade for uranium
States (Harris, 1984, p. 228).
1.25 X (1.25 ppm). Using this value in (5.44), we find that the correla-
tion line in Figure 5.5 is given by
with M in kilograms. According to the fractal model the uranium ore in the
M, = 6.4 X 10'7 kg. Assuming a mean crustal density of 2.7 X 103 kg m-3,
the uranium resources of the United States were concentrated from an origi-
nal crustal volume of 2.4 X lo5 km3. This represents about 0.09 percent of
the crustal volume of the United States. The crustal volume from which ura-
nium is enriched is about a factor of two larger than the crustal volume for
mercury but is a factor of 50 less than the crustal volume for copper.
In several examples the statistics on ore tonnage versus ore grade have
been shown to be fractal to a good approximation. This is not surprising
since two of the classic models for the generation of ore deposits, chromato-
graphic and Rayleigh distillation, both lead directly to fractal distributions.
The examples considered here yield a considerable range of fractal dimen-
sions: 2.01 for mercury, 1.55 for gold, 1.48 for uranium, and 1.16 for copper.
If Rayleigh distillation were applicable then from (5.43), the applicable
liquid-solid partition functions would be 0.33 for mercury, 0.48 for gold,
0.49 for uranium, and 0.61 for copper. It should be emphasized, however,
that the chromatographic model is a more likely explanation for the concen-
tration of these minerals.
Not all mineral deposits and related statistical data satisfy power-law
(fractal) distributions. A specific example is the frequency-size distribution
of diamonds (Deakin and Boxer, 1986).
5.3 Petroleum data
There is also evidence that the frequency-size distribution of oil fields obeys
fractal statistics (Barton and Scholz, 1995). Drew et al. (1982) used the rela-
tion Ni-, = 1.67N to estimate the number of fields of order i, N j , in the west-
ern Gulf of Mexico. Since the volume of oil in a field of order i is a factor of
two greater than the volume of oil in a field of order i - 1, their relation is
equivalent to a fractal distribution with D = 2.22. Barton and Scholz (1995)
find D = 2.49 for the Frio Strandplain play, onshore Texas. The number-size
statistics for oil fields worldwide as compiled by Carmalt and St. John
(1984) are given in Figure 5.6. A reasonably good correlation with the fractal
relation (2.6) is obtained taking D = 3.3. The large differences between
these values for the fractal dimension may be attributed to differences in the
regional geology, but it may also be due to difficulties in the data. It is often
difficult to determine whether adjacent fields are truly separate, and data on
reserves are often poorly constrained. Nevertheless, the applicability of frac-
tal statistics to petroleum reserves can have important implications. Reserve
estimates for petroleum have been obtained by using power-law (fractal) sta-
tistics and log-normal statistics. Accepting power-law statistics leads to con-
siderably higher estimates for available reserves (Barton and Scholz, 1995;
La Pointe, 1995; Crovelli and Barton, 1995).
The model for the concentration of economic ore deposits given above
leads to a range of geometrically acceptable fractal dimensions. However,
the observed distribution for oil fields falls outside this range. This again il-
lustrates the difficulties associated with restrictions on power-law .(fractal)
distributions. As stated previously, we define a power-law statistical distrib-
ution as a fractal distribution.
It should not be surprising that the frequency-size statistics of oil pools
and oil fields are fractal; it was shown in Chapter 2 that topography is gener-
ally fractal. One consequence is that the frequency-size statistics of lakes
has been found to be fractal (Maybeck, 1995). Because traps for oil involve
topography on impermeable sedimentary layers, it is expected that this
topography will also be fractal. Thus it is reasonable that the frequency-size
distribution of oil pools is fractal.
V lo6 bbl oil
Figure 5.6. The number N of

oil fields worldwide with a
volume of oil greater than V
as a function of V. The
equivalent number of barrels
is also given. The circles
represent the data given by
Carmalt and St. John (1984)
and the line represents the
D = 3.3. V km3
ORE GRADE A N D TONNAGE
Barton and Scholz (1995) have examined the spatial distribution of hy-
drocarbon accumulations and have concluded that they obey fractal statis-
tics. Their results for production from the J sandstone of the Denver basin
are given in Figure 5.7. Production from this basin is primarily in the north-
east comer of Colorado and the southwest comer of Nebraska. A 40 X 40-
mile section of the basin is considered and this section is divided into 80 X
80 square cells of size 0.5 miles. The cells with one or more wells are illus-
trated with black dots in Figure 5.7 as drilled cells. The cells with one or
more wells that are either producing or had a show of hydrocarbons but at
quantities too small to produce are illustrated with black dots in Figure 5.7 as
producing or showing cells.
Figure 5.7. A 40 X 40-mile

section of the Denver basin
is considered. This section is
divided into 80 X 80 cells of
dimension 0.5 mile. Wells
that penetrated the J
sandstone reservoir were
considered (Barton and
Scholz, 1995) and were
listed as dry holes, producing
oil or gas, or showing oil or
gas. Cells with one or more
drilled wells are illustrated
as drilled cells by black dots;
cells with one or more
producing or showing wells
are illustrated as producing
and showing cells by black
dots. The box-counting
technique was applied and
the number of occupied
boxes is given as a function
of box size for the two
distributions. Good
correlations are obtained
taking D = 1.80 for the
drilled cells and D = 1.43 for
the producing and showing
cells.
The box-counting technique was applied to both the drilling data and the
producing and showing data. The number of occupied boxes as a function of
the reciprocal of the box size is given in Figure 5.7 for both data sets. In both
cases good correlations were obtained with the fractal relation (2.2). For the
drilled cells the fractal dimension was D = 1.80; if every cell had been
drilled the fractal dimension would have been D = 2.0. For the producing
and showing cells the derived fractal dimension was D = 1.43. This result
indicates that the complex processes responsible for the generation of petro-
leum traps leads to a fractal spatial distribution of oil pools. Barton and
Scholz (1995) also examined the spatial distribution of hydrocarbon accu-
mulations in the Powder River basin, Wyoming, and found a good correla-
tion with fractal statistics taking D = 1.49.
Carlson (1991) examined the spatial distribution of 4775 hydrothermal
precious-metal deposits in the western United States and found that the
probability-density distribution for these deposits is fractal. Blenkinsop
(1994) found similar results for gold deposits in the Zimbabwe Archean
craton.
Problems
Problem 5.1. Determine the concentration factor $* for an ore deposit with
D = 2.
Problem 5.2. Determine the concentration factor $, for an ore deposit with
D = 1.
Problem 5.3. Determine the solid-liquid partition coefficient K , correspond-
ing to an ore deposit with D = 2.
Problem 5.4. Determine the solid-liquid partition coefficient K , correspond-
ing to an ore deposit with D = 1.
Problem 5.5. Consider the cubic model for mineral concentration illustrated
in Figure 3.6. (a) In terms of the enrichment factor $, defined by (5.26)
and C,, what is the concentration in the seven depleted zero-order ele-
ments? (b) What is the concentration in the seven depleted first-order el-
ements? (c) What is the allowed range for $,? (d) What is the corre-
sponding allowed range for D?
Problem 5.6. From the correlation for mercury production given in (5.45),
how much pure mercury ( c = 1) would be expected?
Problem 5.7. From the correlation for mercury production given in (5.45),
determine the total production of mercury when the mean grade of ore
that has been mined reaches C = 0.001.
Problem 5.8. From the correlation for lode gold production given in (5.46),
how much pure gold (c = 1 ) would be expected in the United States?
Problem 5.9. From the correlation for lode gold production given in (5.46),
determine the total amount of lode gold mined to date. Assume that the
mean grade of ore mined prior to the present is C = 9 ppm.
Problem 5.10. From the correlation for copper production given in (5.47) de-
termine the total production of copper to date. Assume that the mean
grade of ore mined prior to the present time is = 0.008.
Problem 5.1 1. From the correlation for copper production given in (5.47),
how much pure copper (C = 1) would be expected?
Problem 5.12. The fractal dimension for the distribution of areas of lakes has
been found to be D = 1.55 (Kent and Wong, 1982). Assuming that the
mean depth of a lake is proportional to the square root of its area, what is
the fractal dimension for the distribution of water volumes in lakes?
Problem 5.13. Consider the data for the 40 X 40-mile section of the Denver
basin given in Figure 5.7. What fraction of 1 X 1-mile sections would be
expected to contain oil?
Chapter Six
FRACTAL
CLUSTERING
6.1 Clustering
We next relate fractal distributions to probability. This can be done using the
sequence of line segments illustrated in Figure 2.1. The objective is to deter-
mine the probability that a step of length r will include a line segment. First
consider the construction illustrated in Figure 2.1 (a). At zero order the prob-
ability that a step of len th r, = 1 will encounter a line segment, p, = 1; at
f
first order we have r , = 2 and p , = 2, 1
and at second order r, = 31 and p, = 41 .
Next consider the construction illustrated in Figure 2.l(c). At zero order the
probability that a step of len th r, = 1 will encounter a line segment is p, =
f
1; at first order we have r , = 2 and p , = 1, and at second order r, = 41 and p, =
1. Finally we consider the Cantor set illustrated in Figure 2.l(e). At zero or-
der the probability that a set of length r, = 1 will encounter a line segment is
2
p, = 1 ; at first order we have r , = 31 and p , = 3, and at second order r, = 31 and
4
PZ = 9 .
The probability that a step of length ri will include a line segment can be
generalized to
where N is the number of line segments of length ri. Taking C = 1 in (2.1)

the number N is related to ri so that we obtain
For the Cantor set the probability that a step of length ri = (f)' encounters a
line segment is pi= ($1' so that D = In 2Iln 3 as was obtained previously.
The Cantor set is both scale invariant and deterministic. Its deterministic
aspect can be eliminated quite easily. A scale-invariant random set is generated
by randomly removing one-third of each line rather than always removing the
FRACTAL CLUSTERING 101
middle third. This process is illustrated in Figure 6.1. The fractal dimension is
unchanged and the probability relations derived above are still applicable.
We will use the examples given above as the basis for studying fractal clus-
tering. We consider a series of point events that occur at specified times. To con-
sider N point events that have occurred in the time interval i0we introduce the
natural period T , = i d N . We then introduce a sequence of intervals defined by
Our measure of clustering will be the probability pn that an event occurs in

an interval of length i n .
As a specific example, consider a uniform (equally spaced in time) se-
ries of N events that occur in an interval i OThe . first event occurs at t, =
i,,/2N, the second event at t, = 3 i d 2 N , the third event at t, = 5rO/2N,and so
forth. The probability p,, that an event will occur in an interval is given by
IIIIIIIIIIIIIIIIIIIlllIllllm Figure 6.1. Illustration of the

first six orders of a random
Cantor set. At each step, a
11I random third of each solid

line is removed.
102 FRACTAL CLUSTERING
If the number of events N is greater than the number of intervals n , we have

N > n or T,,> rO/N.In this case an event occurs in every interval so that pn =
1. If the number of events is less than the number of intervals, we have
N < n or T,, < T,/N. In this case only N of the n intervals have events so that
pn= N/n. Because there is no clustering no interval T~ contains more than one
event.
A more realistic model for a series of events in time is that their occur-
rence is completely uncorrelated. The time at which each individual event
occurs is random. An example would be telephone calls placed in a city dur-
ing a given hour. If N point events occur randomly in a time interval T,, it is a
Poisson distribution. In the limit of a very large number of events (N + =J),
the distribution of intervals between events is given by
where f ( ~ ) dis~the probability that an event will occur after an interval of

time between T and T + d r in length. This distribution is clearly not scale in-
variant since the natural time scale T ~ enters
N (6.5).
We will next determine the probability p that an interval of length T~ will
include an event if N events occur randomly in an interval T,. This is the
classic problem of the random distribution of N balls into n boxes. We intro-
duce Pm= m/n where m is the number of intervals that include events and
n = T ~ T is
, the total number of intervals. We assume that both n and m are in-
tegers. The probability that Pmhas a specified value is given by
n
where fm(pm)= 1; the binomial coefficient is defined by
m=O
It is often appropriate to take Nand n to be large numbers. In this limit, fn(Pm)

from (6.6) has a strong maximum at a specific value of P,, p.
To illustrate the probabilistic approach to clustering we consider a ninth-
order random Cantor set. First we rescale so that unit length is the length of a
ninth-order element. Thus the length of the zero-order element is 3 9 and
there are 29 ninth-order elements. To determine the fractal dimension of the
clustering by the "box method," we take intervals of length r,, = 2" and deter-
mine the fraction p that include at least one ninth-order element as a function
of rn. An example is given by the open circles in Figure 6.2. The best-fit
-
straight line has a slope of 0.368 so that p T-0.368 and D = 0.632. The devia-
tion from the exact value D = 0.6309 for the deterministic Cantor set is due
to the reduced rate of curdling in the probabilistic set. If the same number of
ninth-order elements is uniformly distributed (no clustering), the probability
of finding an element with an interval from (6.4) is given by the solid circles
in Figure 6.2. In this case, the slope is unity for r < (;)9 and zero for r > (;)9.
Thus, D = 0 for r < (;)9, that is, a set of isolated points, and D = 1 for r > ($)9,
that is, a line.
Fractz! clustering has been applied to seismicity by Sadovskiy et al.
(1985) and by Smalley et al. (1987). The latter authors considered the tem-
poral variation of seismicity in several regions near Efate Island in the New
Hebrides island arc for the period 1978-1984. One of their examples is
given in Figure 6.3. During the period under consideration 49 earthquakes
that exceeded the minimum magnitude required for detection occurred in
the region. Time intervals T such that 8 min 1 T I524,288 min were con-
sidered. The fraction of intervals with earthquakes p as a function of interval
length T is given in Figure 6.3(a) as the open circles. The solid line shows
the correlation with the fractal relation (6.2) with D = 0.255. The dashed
line is the result for uniformly spaced events. The results of a simulation
for a random distribution of 49 events in the time interval studied is given
in Figure 6.3(b). The random simulation (Poisson distribution) is signifi-
cantly different from the earthquake data and is close to the uniform distri-
bution.
Figure 6.2. The fraction p of

steps of length r that include
solid lines for a ninth-order
random Cantor set is given
by the open circles. The unit
length is the length of the
shortest line; the original line
length is 39. The solid circles
correspond to a uniform
distribution of the same
number of lines as given by
(6.4). The line corresponds
to (6.2) with D = 0.632.
Fractal clustering of seismic activity in time for California has been

found by Lee and Schwarcz (1995). For the historical record of activity in
the San Andreas fault zone of central California and the paleoseismic record
of the San Gabriel Fault, they found D = 0.43-0.46. For the San Andreas
fault zone in southern California, they found D = 0.67. Velde et al. (1990)
and Meceron and Velde (1991) have carried out studies of the one-dimen-
sional clustering of joints and faults and found good fractal correlations.
They utilized the intersections of the joints and faults with a well. Manning
(1994) has studied the one-dimension clustering of metamorphic veins.
He found D = 0.46 for wollastonite-quartz veins in marbles, D = 0.81 for
actinolite-chlorite veins in oceanic diabases, and D = 0.25-0.63 for epi-
dote-quartz veins in basalts.
Figure 6.3. Fractal cluster

analysis of 49 earthquakes
that occurred near Efate
Island, New Hebrides, in the
period 1978-1984 (Smalley
er al., 1987). (a) The circles
give the fraction of intervals
p of length 7 that include an
earthquake as a function of 7 .
The solid line represents the
D = 0.255. The broken line is
the result for uniformly
spaced events. (b) The
results for 49 randomly
distributed events (Poisson
process).
Fractal clustering can also be studied in higher dimensions. The applica-

tion to two dimensions is illustrated by the sequence of constructions given
in Figure 2.3. The objective is to determine the probability that a square box
of size r encounters a square that has been retained. First consider the con-
struction given in Figure 2.3(a). At zero order the probability that a box
5
of size ro = 1 will include a square is p, = 1; at first order we have r, = and
p , = $, and at second order we have r2 = and p, = &. Next consider the con-
6
struction illustrated in Figure 2.3(b). At zero order the probability that a box
of size r,, = 1 will include a retained square is p, = 1; at first order we have
r , = $ and p , = $, and at second order we have r2 = $ and p2 = &. Finally we
consider the Sierpinski carpet illustrated in Figure 2.3(d). At zero order the
probability that a box of size ri = 1 will include a retained square is po = 1; at
first order we have r , = f and p, = 8, and at second order we have r, = and
64
P2=81e
The probability that a square box of size ri will include a retained square
can be generalized to
and substitution of (2.1) gives
For the Sierpinski carpet the probability that a square box of size ri = (fY
will include a retained square is pi = ($)' so that D = In 8nn 3, as was previ-
ously found. The Sierpinski carpet can be applied to clustering in two di-
mensions in the same way that the Cantor set was applied in one dimension.
This is directly analogous to the box-counting algorithm discussed in Chap-
ter 2 and illustrated in Figure 2.8. The two-dimensional spatial clustering of
intraplate hot spot volcanism (i.e., Hawaii, etc.) has been studied by Jurdy
and Stefanick (1990). They found a fractal correlation with D = 1.2.
This approach can be extended to three dimensions using cubes of various
sizes. The application to three dimensions is illustrated using the Menger
sponge given in Figure 2.4(a). The objective is to determine the probability
that a cube with size r encounters retained material. At zero order the probabil-
ity that a cube of size ro = 1 will include material is p - 1;at first order we have
20
r , = 51 and p , = 2 01- 400
, and at second order we have r, = 9 and p, = m. The proba-
bility that a cube of size ri includes retained material can be generalized to
and substitution of (2.1) gives

For the Menger sponge the probability that a cube of size ri = (f)' encounters
retained material is pi= (E)' so that D = In 20nn 3 as was previously found.
The generalization of (6.2),(6.9)and (6.11)is
6.2 Pair-correlationtechniques
Another approach to the clustering of point events is to use the pair-correla-

tion distribution C(r),which is defined to be the number of pairs of points
whose separation is between r - $Ar and r + i h r per unit area (Vicsek,
1992). One point is picked and the distances to all other points are deter-
mined. The same thing is done for the second point and for all other points.
The number of pairs with separations between r - khr and r + t h r is di-
vided by Ar. This result is the pair-conelation distribution C(r) in one di-
mension. For a two-dimensional distribution the number in each interval A r
is divided by r Ar to obtain C(r);for a three-dimensional distribution the
number in each interval Ar is divided by r2 Ar to obtain C(r).
Two simple deterministic examples illustrate how pair-correlation distri-
butions are determined. First consider the one-dimensional example of four
equally s aced points on a line of unit length. The pair-correlation distribu-
P
tion is C ( J )= 6 , C($)= 4, C ( l )= 2. Next consider the two-dimensional exam-
ple of four points on the corners of a unit square. The pair-correlation distrib-
ution is C(1)= 8, ~ ( f =i 4)1 f i = 2 f i .
If the points are randomly distributed in space, the pair-correlation dis-
tribution is exponential
If the points exhibit scale-invariant clustering, a power-law dependence is

obtained
where a is related to the fractal dimension of the distribution by
It is seen that (6.14)and (6.15)are entirely equivalent to (6.12).

We will consider two examples of scale-invariant (fractal) pair-correla-

tion distributions. Our first example is a sixth-order Cantor set with 64
points. The pair-correlation distribution is given in Figure 6.4. A good corre-
lation with (6.14) is obtained taking a = 0.369. From (6.15) we find with d =
1 that D = 0.631, the fractal dimension for a Cantor set is D = In 2/ln 3 =
0.6309. As our second example we consider the fourth-order Koch snow-
flake illustrated in Figure 6.5, which has 625 points. The pair-correlation
distribution is given in Figure 6.6. A good correlation with (6.14) is obtained
taking a = 0.58. From (6.15) we find with d = 2 that D = 1.42, and the fractal
dimension of the Koch snowflake is D = In 5/ln 3 = 1.465. Again reasonably
good agreement is found.
Kagan and Knopoff (1980) have determined the pair-correlation distrib-
ution for the two-dimensional spatial distribution of worldwide seismicity
and found that a = 1 for shallow seismicity so that D = 1 . This is consistent
with the D = 2 found for the spatial distribution of aftershocks in California
by Robertson et al. (1995) illustrated in Figure 4.12. The pair-correlation
technique is entirely equivalent to the box-counting technique for point
events, and both methods give the same fractal dimension for scale-invariant
distributions.
Figure 6.4. The pair-

,-.
-L slope=-0 369 correlation distribution C(r)
u m a=O 369
is given as a function of r for
\
QQ0
D=l-a=O 631
4
a sixth-order (n = 64) Cantor

set. A good correlation with
the fractal relation (6.14) is
w
o obtained taking a = 0.369.
From (6.15) the
*
0
L.p-L
o9 12 ' 5
log r
18
.L
2 1 74
corresponding fractal
dimension is D = 0.631
fractal dimension for the(the
Cantor set is D = In 2nn 3 =
0.6309).
Figure 6.5. Illustration of a

fourth-order Koch snowflake
(n = 625). This deterministic
fractal has a fractal
dimension D = In 5An 3 = 0.0 20.0 40.0 60.0 80.0
1A65. col
Figure 6.6. The pair-

correlation distribution C(r)
is given as a function of r for
the fourth-order Koch
snowflake illustration in
Figure 6.5. A good
correlation with the fractal
relation (6.14) is obtained
taking a = 0.58. From (6.15)
we have D = 1.42 (D = 1.465 0.2 0.4 0.6 0.8 1 .O 1.2 1 4 1.6
for the Koch snowflake). log r
6.3 Lacunarity
It is clear that fractal constructs with identical fractal dimensions can have
quite different appearances. One example is the deterministic Cantor set il-
lustrated in Figure 2. l (e) compared with the random Cantor set illustrated in
Figure 6.1. Third-order examples of these sets are given in Figure 6.7. The
difference between these two sets is the distribution of the size of gaps. Man-
delbrot (1982) introduced the concept of lacunarity as a quantitative measure
of the distribution of gap sizes. Large lacunarity implies large gaps and a
clumping of points; small lacunarity implies a more uniform distribution of
gap sizes. Also included in Figure 6.7 are examples of a near uniform distri-
bution (near zero lacunarity) and a totally clumped distribution (high lacu-
narity). In each case a line segment with a length of 27 is divided into 27
equal parts, each of unit length, and 8 are retained.
Allain and Cloitre (1991) have introduced a quantitative measure of la-
cunarity, which we will use below. Alternative measures have been given by
Gefen et al. (1984) and by Lin and Yang (1986). The technique given by Al-
lain and Cloitre (1991) is illustrated in Figure 6.8. We consider the third-
order Cantor set given in Figure 6.7(b). The total length is r, = 27 and indi-
vidual segments have unit length ( r = I). We consider a moving window of
length r, which is translated in unit increments. The total number of steps
considered is given by
For the example considered in Figure 6.8, we take r = 9 and N(9) = 19 as

shown. The number of remaining line segments covered by a step is s and for
the first steps, = 4, second steps, = 3, and so on, as shown. We denote the num-
ber of steps of length r that contain s segments by n(s, r). For the example
Figure 6.7. A line segment
with a length of 27 is divided
into 27 equal-sized segments
of unit length and 8 are
retained. (a) A near uniform
distribution of equally
spaced elements. (b) A
Cantor set as in Figure
2.l(e). (c) A random Cantor
set as in Figure 6.1. (d) A
clumped distribution with
the 8 segments adjacent to
each other. The lacunarity
increases from top to bottom
(a to d).
considered we have n(0, 9 ) = 1 , n(1, 9 ) = 4, n(2, 9 ) = 8 , n(3, 9 ) = 4 , and

n ( 4 , 9 ) = 2. We next define a sequence of probabilities by
a n d ~ e f i n d p ( 0 , 9 ) = & , p ( l , 9 ) = $ , ~ ( 2 , 9 ) = & , ~ (=3&

, 9, a) n d p ( 4 , 9 ) =
2 for the example considered. The first- and second-order moments of this
19.
distribution are defined according to

sliding-window method used
to determine the lacunarity
of a one-dimensional set.
This third-order Cantor set
contains eight line segments
of unit length and a total
length r, = 27. The moving
window has a length r = 9.
The window moves unit
steps to the right and there
are N(9) = 19 steps as
illustrated. The values of s
for each step are given. The
number of steps that contain
s line segments n(s, r) are
n(0,9)=1,n(l,9)=4,
n(2,9)=8,n(3,9)=4,and
n ( 4 , 9 )= 2 .
For the example illustrated in Figure 6.7 we have M, (9) = $ and M, (9) =
%. The lacunarity L is defined in terms of the moments by
The lacunarity for the example considered is L (9) = 1.235.

The lacunarities for the four distributions illustrated in Figure 6.7 are
given in Figure 6.9 as a function of the window length r ( r = l , 2 , . . . , 27).
For r = 1 all the distributions have L (1) = 3.375 [n (0, 1) = 19, n (1, 1) = 8 , p
19 8 8 8
(0, 1) = 27, p (I, 1) = 3 ,M1(l)= 27, M2 (1) = 3 1.AS the value of r increases,
the lacunarity decreases toward unity. The near uniform distribution illus-
trated in Figure 6.7(a) has the lowest lacunarity, near zero for r > 2, as ex-
pected. Also, the clumped distribution illustrated in Figure 6.7(d) has a sin-
gle large gap and has the highest lacunarity, as expected. The random Cantor
set has a significantly higher lacunarity than the deterministic Cantor set
even though both have the same fractal dimension. Thus lacunarity can be
used as a measure of a distribution in addition to its fractal dimension.
The method described above for determining the lacunarity in one di-
mension is easily extended to higher dimensions. As a specific example we
consider the Sierpinski carpet illustrated in Figure 2.3(d). The moving-box
0
0 0 Uniform
x Cantor
0
+ Random Cantor
0 o Clumped
Figure 6.9. Lacunarities L as

a function of step length r for
the four distributions given
in Figure 6.7.
method applied to a second-order Sierpinski carpet is illustrated in Figure

6.10. The carpet is a 9 X 9 square, which has been divided into 81 unit
squares of which 64 have been retained. The moving box is a r X r square
that moves at unit increments until the entire carpet has been covered. The
total number of steps required for a two-dimensional problem is
Nr = (r,- r + 1)2 (6.21)
For the example considered in Figure 6.10 we have r = 9 and r = 3 so that N

(3)= 49;the 49 steps are illustrated. For the first step in the upper, left-hand
corner, the 3 X 3 box covers 8 remaining unit squares so that s = 8. The val-
ues of s for the 49 steps are given. Again we denote the number of steps with
a r X r box that contains s unit squares by n (s, r). For the example consid-
ered we have n (0,3)= 1, n (3,3)= 4,n (5,3)= 8,n (6,3)= 8,and n (7,3) =
4,n = (8,3) = 24. The corresponding sequence of probabilities is p (0,3)=
&,P(3,3)=~,p(5,3)=$,p(6,3)=$,p(7,3)=~,andp(8,3)=~.~rom
(6.18)and (6.19)the moments are M, (3)= % and M, (3)= T. And from
(6.20)the lacunarity is L = 1.080.
Lacunarity studies can be used in a variety of applications involving tex-
ture analysis. Plotnick et al. (1993)give applications to landscape textures.
The concept of extending fractal analyses to higher-order moments leads us
the moving-box method used
naturally to the introduction of multifractal methods.
to determine the lacunarity
of a two-dimensional
distribution. This second-
order Sierpinski carpet
contains 81 unit squares in a
9 X 9 square carpet. The
moving box is a 3 X 3
square. The box moves unit
steps to the right until the
sweep is completed with
seven steps. The box is then
moved down a unit step and
a second sweep is carried
out. In all there are seven
sweeps with a total of N
( 3 ) = 49 steps as illustrated.
The number of unit squares
included in the box at each
step s is given for each step.
The number of steps that
contain s unit squares n ( s , r)
are n ( 0 , 3 ) = 1 , n ( 3 , 3 ) = 4,
n (5, 3 ) = 8, n (6, 3 ) = 8, n
(7, 3 ) = 4 , and n (8, 3 ) = 24.
6.4 Multifractals
Our analysis of lacunarity introduced higher-order moments to our consider-

ations of self-similarity and fractals. The utilization of higher-order mo-
ments of statistical distributions can be generalized utilizing the concept of
multifractals (Halsey et al., 1986; Mandelbrot, 1989). We again begin by
considering the Cantor set illustrated in Figure 2.l(e). A third-order Cantor
set is illustrated in Figure 6.1 l(a). A line segment of unit length is divided
into 27 equal parts and 8 line segments are retained. To define a multifractal
set, the original line of unit length is divided into n equal segments, and the
segments are denoted by i = 1, 2, . . . , n; and the length of each segment is
given by r = n-1. The fractionfi of the remaining line in segment i is given by
where Liis the length of line in segment i and L is the total length of line.
Since
n
i=l
Li= L,we have xfi = 1. The quantityfi is the probability that the re-
n
i=l
maining line segment is found in "box" i.

the multifractal fractions for
a third-order Cantor set. (a)
Third order Cantor set. (b)
The single fractionf , is given
for r = 1 . (c) The three
fractions f,-f, are given for
:.
r = (d) The nine fractions
f,-f,are given for r = a.
For the third-order Cantor set of unit length illustrated in Figure 6.11(a),
we have L = $. We now determine the values off, for three cases, n = 1
(r = I), n = 3 ( r = 3). and n = 9 (r = i).
Taking n = 1 (r = 1) we have i = 1; in
this one segment we have L,= $ and from (6.22) obtain f,= 1. This is illus-
trated in Figure 6.1 l(b). For n = 3 ( r = i) we have i = 1, 2, 3; from Figure
6.11(c) we obtain L,= A, L2 = 0,L3 = $ and from (6.22) find f,= f2= 0,4,
f3= i. With n = 9 ( r = ,$)we have i = 1, 2, 3, 4, 5, 6, 7, 8, 9; from Figure
6.11(d) we obtain L,= &, L2 = 0,L3 = $, L4 = L5 = L6 = 0,L, = 27, 2
L8 = 0,
=a,
L, = &; and from (6.22) find f, f2= 0,f3=:, f4=f5=f6= 0,f, = $, f, = 0,
f ='.
9 4
It is next necessary to define generalized moments M4(r) of the set of
fractionsf,(r). This is done using the relation
where the sum is taken over the set of fractions and q is the order of the mo-
ment; (6.23) is valid for both integer and noninteger values of q as long as
q # 1. The special case q = 1 will be considered below. For the example
given in Figure 6.11, we can obtain the moments of the distribution for any
order q (except q = 1) using (6.23). We first take q = 0 and find the zero-
order moments for r = I , $,and ,$ with the result:
Note that any finite number raised to the power 0 is 1, but 0 raised to the
power of 0 is 0. We next take q = 2 and find the second-order moments for
i,
r = 1, and ,$ with the result:
M, ( I ) = 12 = 1
Moments of other orders q can be obtained in a similar manner.

For the first-order moment M, (r), q = 1, a modified definition of the mo-
ment is required. We first write
fq = ffq-I =f exp (ln fq-1) = f exp [(q - 1 ) ln f ]
In the limit E < < 1 we can write
exp=
~ 1 +E (6.25)
plus higher-order terms. Thus in the limit q + 1 (q 1 << 1) (6.24) be-

comes
fq = f [ l + ( 9 - 1) l n f l
Substitution of this result into the definition of the generalized moment

(6.23) gives
and since xfi

n
i=l
= 1 we have
This result is valid in the limit q + 1.

The generalized multifractal dimension Dq is defined by
where the values of Dq form the multifractal spectrum. For q = 0, Do is re-

ferred to as the "box" dimension; it is entirely equivalent to the fractal di-
mensions considered in previous chapters. This is because the probabilityp
is unity if the box is occupied and is zero if it is not. For q = 1 , D, is referred

to as the "information" or "entropy" dimension and for q = 2, D, is referred
to as the "correlation" dimension. These specific cases will be discussed in
more detail below. If all values of Dq in (6.29) are equal, then a "homoge-
neous fractal" is defined. If (6.29) gives well-defined power laws, but the
values of D, are different, then a "multifractal" is obtained.
The definition (6.29)can be rewritten as
This result can be used directly as long as q # 1 . Substitution of (6.23)gives
for q # 1 . The zero-order box dimension is ob~tainedfrom (6.30)taking q = 0

with the result
For the third-order Cantor set illustrated in Figure 6.11 we have M, ( 1 ) = 1 ,

M, (3)= 2 and Mo (i)
= 4 . Taking either rj = 1 and rk = j1 or r.J = j1 and rk = g1 ,
we find Do = In 2/ln 3, which is identical to the value obtained in Chapter 2.
The correlation dimension is obtained from (6.30) taking q = 2 with the
result
For the third-order Cantor set we have M,(1) = 1 , ~ , ( f )= i, and M2(g)

1
= 21 .
Taking either rj = 1 and rk= 31 or rj = 51 and rk= y1 we find that D,= In 21ln 3,
which is again identical to the value obtained in Chapter 2.
To obtain an expression for D, it is necessary to combine (6.28) and
in the limit q + 1 ;
(6.30).We first consider the quantity In [Mq (rj)lMq(rk)]
using (6.28)we have
However, in the limit E << 1 we have
plus higher-order terms. Thus in the limit q 4 1 (q - 1 << 1) (6.34) be-

comes
Substitution of this limit into the definition of the multifractal dimension

(6.30) while taking the limit q + 1 gives
where
For the third-order Cantor set we have
s(f)= - (i)In (f)- OlnO - (f)In (f)= In2

s = - (a)
(i) In ($) - 0 In 0 - ($)In (a) - 0 In 0 - 0 In 0 - 0 In 0
where by definition 0 In 0 = 0. Taking either r.J = 1 and r, = 31 or r.J = 51 and

i
r, = in (6.37), we find that again D, = In 2An 3. We have found that Do=
D, = D,= In 2lln 3 for a third-order Cantor set. In fact this value is obtained
for any value of q and the third-order Cantor set is a homogeneous fractal.
Any other order Cantor set would have given the same result and would also
be a homogeneous fractal.
To interpret the first-order fractal dimension D,it is appropriate to note
that the classical statistical mechanics definition of the entropy S of a ther-
modynamic system is (Morse, 1969, p. 248)
with x & = 1 where is the distribution function, k is the Boltzmann con-

stant, ind the summation is carried out over the allowed states of the system.
The similarity between (6.39) and (6.38) is the basis for calling D,the infor-
mation or entropy dimension.
As a second example of a homogeneous fractal we consider the Sierpin-
ski carpet illustrated in Figure 2.3(d). A second-order Sierpinski carpet is il-
lustrated in Figure 6.12(a). The unit square is divided into 81 equal-sized
squares with r, = $ and 64 squares are retained. To define a multifractal in
this case we divide the unit area into n equal-sized squares with linear di-
mensions r = n- '2; the squares are denoted by i = 1,2, . . . , n. The fraction&
of the remaining area in each square box i is given by

a second-order Sierpinski
carpet. (a) Second-order
Sierpinski carpet. (b) The
single fractionf,is given
for r = 1. (c) The nine
fractionsf,-f,are given for
n
where A, is the area in box i and A is the total retained area. Again Ai = A
n
so that Z J= 1
i=l
For the second-order Sierpinski carpet illustrated in Figure 6.12(a) we
have A = g. We now determine the values of fn for two cases, n = 1 (r = 1)
1
and n = 9 ( r = j). Taking n = 1 we have i = 1; in this one area we have A, =
3 and from (6.40) f , = 1. This is illustrated in Figure 6.12(b). For n = 9 we
have i = 1, 2, 3, 4, 5, 6, 7, 8, 9 as illustrated in Figure 6.12(c); we have A , =
8 8
A, = A3 = A4 = 8 , A5 = 0, A6 = A, = A8 = A9 = 81. From (6.40) we have
1
f , = f , = f 3 = f , = 8 , f 5 = 0 , f6=f,=f8=f9=$.From(6.23) we find
And from (6.38) we find
(i) (3 (f) (3 (f) (3 (f)

S - - - 1n - 1n - 1n - (f) 1n (3
And from (6.32), (6.33), and (6.37) we obtain Do = Dl = D, = In 8An 3, the

same value obtained in Chapter 2. The Sierpinski carpet is also a homoge-
neous fractal.
Many examples similar to the above can be found. If a problem consists
of two regions, one occupied and one not occupied, with a scale-invariant
boundary between them, then the box-counting method introduced in Chap-
ter 2 is generally applicable and there is no need for multifractal concepts.
However, there are many geological problems in which there is a continuous
variation in a quantity rather than a discontinuous variation. The concentra-
tion of a mineral has a continuous variation so that a straightforward box-
counting approach is not applicable to its spatial distribution; every box
would be occupied so that the fractal dimension would be equal to the spatial
dimension.
We will now consider a variation on the De Wijs (1951, 1953) model for
the concentration variability in ore deposits, which was illustrated in Figure
5.l(a). At each order the mass was divided into two parts and the enriched
half was given a concentration C,, = +, C, (1 < 4, < 2) and the depleted
half was given a concentration C,, = (2 - +,)C0. The third-order distribution
is given in Figure 6.13(a). A line of unit length is divided into eight equal
parts, each segment having been enriched or depleted as indicated, with 4,
the enrichment factor. Since the concentration in each part is determined by
a multiplication by powers of +, and (2 - +,), this process is known as a
multiplicative cascade.
To carry out a multifractal analysis of the third-order distribution, the
line of unit length is divided into n equal segments, which are denoted by i =
1,2, . . . , n and r = n-1. The fraction4 is defined by
where Ciis the mean concentration in segment i and Cois the overall mean
concentration.
We now determine the values off;. for n = l(r = I),n = 2 (r= i),
n=4
(r = a), and n = 8 (r = k).
Taking n = 1 we have i = 1, C, = C,,
and f,= 1; for n = 2 we have i = 1, 2, C,= +,C,,C2= (2 - 4,) C,,
f,= (1)+2, f2= 1 - (i)+2; for n = 4 we have i = 1, 2, 3, 4,C,= +,2 Co,
c2= c3= +2 (2 - +J cO, c4= (2- +,)2 cO, f,= ($1$2 , f2= f3= (1) +2
[l - (1)+,I, f4= [l - (1)+,I2; for n = 8 we have i = 1, 2, 3,4,5 , 6,7,8,
+;
c,= c,,c2= c3= c4= +; (2 - +,) C,,C5= C6= C,= +,(2 - +,)2Co,Cg
= (2- +,I3, fl= (i) f2= f3 = f4 = ( f ) +$ - ($1 $21 f 5 = f6 = f7 =
+;9
(1)+2 [l - (4) +2]2, fa= [l - (1)+2]3. These results are illustrated in Figure
We can now determine the generalized moments for the De Wijs multi-
we find for q = 0 that
plicative cascade. From (6.23)
Figure 6.13. Illustrations of

a third-order De Wijs
multiplicative cascade. (a)
Concentrations in the eight
equal masses at third order.
(b) The single fraction f,is
given for r = 1. (c) The two
fractions f,,f, are given for
r = i. (d) The four fractions
a.
fl-f,are given for r = (e)
The eight fractions fl-f,are
given for r = i.
+ [(1 - ;+2)3]0 = 1 + 3 + 3 + 1= 8
where it is assumed that 0 < +, < 2; in particular +, # 0 and # 2. For q = 2

we have
This result can be easily generalized to a De Wijs multiplicative cascade of

arbitrary order n, r,, = ($)", and to a moment of arbitrary order q giving
with q # 1 and 0 < +, < 2. Note that this result is also valid for noninteger
values of q.
For q = 1 from (6.38) we find
s(1) = - llnl =O
And again this result can be generalized to a cascade of arbitrary order n

with the result
We can now determine the generalized multifractal dimensions for the

De Wijs multiplicative cascade. For q = 0 we find from (6.32) that Do = 1
and this result is valid at all orders n. Each of the linear array of boxes con-
sidered is occupied by some ore so that we find the box-counting dimension
of a line, Do = 1. For q = 2 we find from (6.33) that
And again this result is independent of the order n considered.

The fractal dimension for arbitrary moment order q is given by
which is valid for both integer and noninteger values of q except q = 1. For
q = 1 we find from (6.37) that
again independent of the order n. The De Wijs multiplicative cascade is a

perfect multifractal in that for all values of q, D, is independent of order n.
As a specific example we take +, = 1.5 and find that Do = 1, D, .= 0.81 1,
and D, = 0.678. Values of Dq for a range of values of 4, are given in Figure
6.14. As 4, + 1 the multifractal dimension Dq approaches unity for all val-
ues of q. In this limit each box has the initial concentration of ore Co and the
dimension of a straight line is obtained. As +, + 2 the multifractal dimen-
sion Dq approaches zero for all values of q except q = 0. In this limit all the
ore becomes concentrated into one box so that the fractal dimension of zero
is obtained. However, since minute quantities of ore remain in the other
boxes, the box-counting dimension, q = 0, remains unity.
We now review the steps necessary to carry out a multifractal analysis.
As a specific example consider a spatial distribution of faults and joints such
as that illustrated in Figure 4.6. The total length of the fractures is L. The re-
quired steps are:
1. The standard box-counting method is applied. The two-dimensional

array of joints and faults is overlaid with a grid of square boxes;
grids of different size boxes are used.
2. The fractions& for each box of size r,, is determined. The length of
faults and joints in box i is L,. The fraction of the faults and joints in
box i,f,,is determined using (6.22).
3. The generalized moments Mq (r) are obtained using (6.23) if q ;t 1
and (6.28) if q = 1.
4. If the Mq ( r ) have a power-law dependence on r for specified values
of q, then the fractal dimensions Dq are obtained using (6.3) if q # 1
and (6.37)if q = 1.
5. The Dq are given as a function of q over the range 0 5 q < -.
It is common practice in multifractal analysis to introduce a scale-invari-

ant probability distribution analysis. This is known as thefla), a curve. To
illustrate this approach we will again use the De Wijs multiplicative cascade.
Our analysis will be carried out in the limit of very high order n + -. In
terms of the order n of the system, the length of the line segment is given by
Figure 6.14. Generalized

multifractal dimension D y
as a function of q for the
De Wijs multiplicative
cascade; enrichment factors
+, = 1.25, 1S O , and 1.75 are
considered. The results are
valid for any order n of the
cascade.
For the De Wijs multiplicative cascade the range of concentrations is given

by
with j = 0 , 1, . . . , n. Note that this nomenclature is different from that used

above because we assign each concentration a subscript j rather than assign-
ing each line segment a subscript i. We also use a reference concentration Cor
to avoid confusion with C, corresponding to j = 0 . For the De Wijs multi-
plicative cascade the number of line segments with the concentration C,, Ni.
is given by
N, = (7) =
n!
j!(n - j)!
where the binomial coefficient has been defined in (6.7).

To illustrate this nomenclature consider the third-order, n = 3, example
given in Figure 6.13. From (6.46) we have r = i.From (6.47) and (6.48)
+:
we have f o r j = 0: C, = C,,, No = 1 , f o r j = 1 : C, = +;(2 - +,) C , , N, = 3,
for j = 2: C, = ( 2 - +2)2 CoryN2 = 3; and for j = 3: C, = ( 2 - +,)3 C,,,
N3 = 1 . The probability has been previously introduced in (6.41) and here
takes the form
Note, however, that we now have x4

n
i=O
N, = 1 rather than xn
i=l
f ; = 1 as in the
previous analysis. Substituting (6.46)and (6.47)into (6.49)gives
We can consider the4 from (6.50)to be analogous to a concentration and the

N, from (6.48) to be analogous to the probability of having that concentra-
tion.
In order to define the flor,), or, curve, we introduce the following defini-
tions o f f i y ) and or,
For the De Wijs multiplicative cascade, substitution of (6.46), (6.48), and

(6.50) into (6.51) and (6.52) gives
log n! - log j! - log ( n - j)!

f@,> =
n log 2
Values of a.andfla,) can be determined for specified values of n, j, and +,.

A contihuous dependence offla) on a can be obtained by assuming both
j and n to be large integers. In this limit the Stirling approximation gives
log n! = n log n - n (6.55)
logj! =j l o g j -j (6.56)
log (n - j)! = (n - j) log (n - j) - (n - j) (6.57)
Introducing x = jln along with (6.55)-(6.57), (6.53) and (6.54) become
a = 1 - (1 - x)----log +, x 1% ( 2 -
-
$2)
log 2 log 2
Aa) = &[x log (T)1 - x - log ( 1 - x) 1

The dependence of fla) on a is obtained by varying x over the range 0 <
x < 1.
The dependence offla) on a from (6.58) and (6.59) is given in Figure
6.15 for +,
= 1.2, 1.4, and 1.6. This result is clearly analogous to the depen-
dence of a probability distributionflx) on x, for example the log-normal dis-
tribution in Figure 3.2. Also included in this figure are the measured values
for one-dimensional sections through the energy-dissipation field in several
fully developed turbulent flows (grid turbulence, wake of a circular cylinder,
atmospheric turbulence) as given by Meneveau and Sreenivasan (1987).
+,
Good agreement with the case = 1.4 ( p , = 0.7) is found.
It is seen that the De Wijs multiplicative cascade generates a perfect

multifractal. However, as was shown in Chapter 4, the resulting distribution
of ore grade on tonnage is not fractal. In fact, as the construction is carried to
infinite order, the tonnage-grade distribution becomes log normal. A log-
normal distribution is a multifractal. Neither the log-normal distribution nor
the De Wijs construction is scale invariant. Thus a multifractal is not neces-
sarily scale invariant, and in general it will not be scale invariant.
Multifractals represent a particular type of scaling that can be associated
with multiplicative cascades. Multifractal scaling can be found in a wide va-
riety of geological and geophysical applications. A number of authors have
shown that the spatial distribution of earthquakes may obey multifractal sta-
tistics (Hirabayashi et al., 1992; Hooge et al., 1994a, b; Blanter and Shnir-
man, 1994). Hirata and Imoto (1991) found the spatial distribution of micro-
earthquakes in the Kanto region of Japan to be multifractal, and Eneva
Figure 6.15. Theflcx) vs.

a curve for the De Wijs
multiplicative cascade model
in the limit of infinite order.
Curves are given for 4, =
1.2, 1.4, 1.6 ( p , = 0.6,0.7,
0.8). The data points are
measured values for the
energy-dissipation field in
fully developed turbulence
(Meneveau and Sreenivasan,
1987).
(1994) found the spatial distribution of mining-induced seismicity to be mul-

tifractal. Dongsheng et al. (1994) have suggested that there was a systematic
variation in the multifractal spectra of the spatial distribution of regional
seismicity prior to the great Tanshan (China) earthquake in July 1976.
Belfield (1994) found that the clustering of fracture networks can be
multifractal, and Sornette et al. (1993) found similar results for fractures
generated in a laboratory experiment. Muller (1994) found the distribution
of pore spaces in North Sea chalk formations to be multifractal, and Pyrak-
Nolte et al. (1992) found that the distribution of contact areas and void
spaces in single fractures in granite were multifractal. Perfect et al. (1993)
applied multifractal statistics to soil aggregate fragmentation. Saucier
(1992) determined the effective permeability of a multifractal porous media.
Lovejoy et al. (1995) applied multifractal characterization to topography
and Ijjasz-Vasquez et al. (1992) to river basins. Klement et al. (1993) ap-
plied multifractal scaling to time series, and Muller (1992) and Saucier and
Muller (1993) have applied the technique to the characterization of geologi-
cal formations using well logs.
The applicability of multifractal statistics to a natural phenomenon may
provide important clues to the underlying physical processes. However, mul-
tifractal statistics are much less useful than monofractal statistics from a
practical point of view. A monofractal distribution is characterized by two
constants, for example in (2.2). But a full multifractal spectral in principle
requires an infinite number of constants. Monofractal applicability implies
scale invariance, but multifractal applicability does not.
Problems
Problem 6.1. Consider the construction given in Figure 2.l(b). What is the
probability that a step of length r includes a line segment for r = 1, i,
1 L7
99 2 7 '
Problem 6.2. Consider the construction given in Figure 2.l(d). What is the
probability that a step of length r includes a line segment for r = 1, 5,
1 I?
97 27
Problem 6.3. Consider the construction given in Figure 2.l(f). What is the
probability that a step length r includes a line segment for r = 1, $, &,?
Problem 6.4. A line segment is divided into seven equal parts and four are re-
tained. The construction is repeated. What is the probability that a step
of length r includes a line segment for r = i , &, & ?
Problem 6.5. A line segment is divided into seven equal parts and three are
retained. The construction is repeated. What is the probability that a step
3, A?
of length r includes a line segment for r = &,
Problem 6.6. Consider the construction given in Figure 2.3(c). What is the
probability that a square box with dimensions r includes a retained
square when r = 1 1 ?,1,,
Problem 6.7. A unit square is divided into four smaller squares of equal size.
Two diagonally opposite squares are retained and the construction is re-
peated. What is the probability that a square box with dimensions r in-
cludes a retained square when r = 1, $, i?:,
All the squares on the boundary and the central square are retained and
the construction is repeated. What is the probability that a square box
with dimensions r includes a retained square when r = 1, i , &?
Problem 6.9. Consider the construction given in Figure 2.4(b). What is the
probability that a cube with dimensions r includes solid when r = 1, $,
I I?
49 8 '
Problem 6.10. A unit cube is divided into 27 smaller cubes of equal volume.
All the cubes are retained except for the central one and the construction
is repeated. What is the probability that a cube with dimensions r in-
cludes solid when r = 1, &, ;?
Problem 6.11. What is the pair-correlation distribution for three equally
spaced particles on a line of unit length?
Problem 6.12. What is the pair-correlation distribution for three particles on
the comers of an equilateral triangle with sides of unit length?
Problem 6.13. What is the pair correlation distribution for the eight particles
on the corners of a unit cube?
Problem 6.14. Consider the third-order Cantor set illustrated in Figure
6.1 l(a). Determine M3 (I), M3 ($)and M3 (6). Write an expression for D3
in terms of ri and r, and determine its value for the third-order Cantor set.
Prob!em 6.15. Consider the third-order Cantor set illustrated in Figure
6.11(a). Determine MI, (I), M,,,($), and M,,,($). Write an expression for
Dl,, in terms of ri and rj and determine its value for the third-order Can-
tor set.
Problem 6.16. Consider the second-order set illustrated in Figure 2.l(f) ( L =
&). Determine L, and fi for n = 1 and n = 5, determine Mo(l), M&),
M2(f),411, ~ ( 3Do,
) ~Dl, and D,.
Problem 6.17. A line segment is divided into seven equal parts and four are
retained (L = $).Determine Li and& for n = 1 and n = 7. Determine Mo
(11, Mo ($),M2(lI9M2(+I, s (I), s ($1,Do, Dl, andD2.
Problem 6.18. Consider the second-order Sierpinski carpet illustrated in Fig-
ure 6.12a. Determine M3 (1) and M3 ($);determine D,.
Problem 6.19. A unit square is divided into four smaller squares of equal
4.
size. Two diagonally opposite squares are retained, A = Determine Ai
a n d 4 for n = 1 (r = 1) and n = 4 (r = $), determine M,,(I), M,, (i),

M2
(I), 4($1, s (I), s (;I, Do,Dl, D2.
Problem 6.20.Consider the third-order De Wijs model for ore concentration
illustrated in Figure 6.13.Determine expressions for M 3(I), M , (i),
M3
(a),M, (Q), and D,in terms of +,. Obtain values of these quantities for
+, = 1.5.
Problem 6.21.Consider a unit cube, r,, = 1, which is divided into 8 smaller
4.
cubes of equal size, r, = Four of the cubes have been enriched by the
factor +,and four have been depleted by the factor 2 - +,. Show that
f,=f, = f, = f, = :+,and f, = f, = f, = f, = (2 - +,). Determine
expressions for Mo(I), M o (i),
M, ( I ) , M, ($1, s (11, s ($1, Do, D l ,
and D,
in terms of +,.
Chapter Seven
SELF-AFFINE
FRACTALS
7.1 Definition of a self-affine fractal
Up to this point we have considered self-similar fractals; we now turn to

self-affine fractals (Mandelbrot, 1985). Topography is an example of both.
In the two horizontal directions topography is often self-similar; the ruler
method can be applied to a coastline or to a contour on a topographic map to
define a fractal dimension. The box-counting method can also be applied to
a coastline, and square boxes are used to determine a fractal dimension.
These are examples of self-similar fractals. Consider next the elevation of
topography. Three examples of elevation h as a function of distance x along
linear tracks are given in Figure 7.1. The vertical coordinate is statistically
related to the horizontal coordinate but systematically has a smaller magni-
tude. Vertical cross sections of this type are often examples of self-affine
fractals (Dubuc et al., 1989a).
A statistically self-similar fractal is by definition isotropic. In two di-
mensions defined by x and y coordinates the results do not depend on the
geometrical orientation of the x- and y-axes. This principle was illustrated in
Figure 2.8, where the box-counting method was introduced. The fractal di-
mension of a rocky coastline is independent of the orientation of the boxes.
A formal definition of a self-similar fractal in a two-dimensional xy-space is
thatflrx, ry) is statistically similar toflx, y) where r is a scaling factor. This
result is quantified by applications of the fractal relation (2.6). The num-
ber of boxes with dimensions x,, y, required to cover a rocky coastline is N,;
the number of boxes with dimensions x, = rx,, y, = ry, required to cover a
rocky coastline is N,. If the rocky coastline is a self-similar fractal, we have
N2/N1= r-D.
A statistically self-affine fractal is not isotropic. A formal definition of a
self-affine fractal in a two-dimensional xy-space is that f(rx, Fay) is statisti-
cally similar to f(x, y) where Ha is known as the Hausdorff measure; we shall
relate Ha to the fractal dimension in what follows. In the box-counting method
square boxes become more and more rectangular as their size is increased.
SELF-AFFINE FRACTALS 133
To illustrate this scaling we consider the deterministic example of a self-

affine fractal in Figure 7.2. A rectangular region with width ro and arbitrary
height h, is defined. At zero order we consider a straight line between x = 0,
y = 0, and x = r,, y = h,. At all orders our fractal construction will begin at
x = 0, y = 0, and end at x = r, y = h,. Our first-order, self-affine fractal is de-
fined in Figure 7.2(b). The horizontal coordinate is divided into four equal
parts so that r, = rd4, and the vertical coordinate is divided into two equal
parts so that h, = hd2. We connect the points (0, O), (r0/4, hd2), (r0/2, O), and
(r,, h,). This is the generator for our fractal construction. In the second-order
5-
4 -- V.E. = 1:270
Figure 7.1. Elevation h as a

function of position x along
linear tracks. Three examples
are given: (a) Earth's
topography along a 7,500-
km track from 70W.55s to
394.5 4 70W,12N. (b) Venus
0 2 4 6 8 10 topography in Ovda Regio.
X, m (c) Elevation across a small
(4 gully on the earth.
134 SELF-AFFINE FRACTALS
fractal illustrated in Figure 7.2(c) each straight-line segment in the first-

order fractal has been replaced by the generator. At second order the hori-
zontal coordinate has been divided into 16 equal parts so that r, = rd16 and
the vertical coordinate is divided into four equal parts so that h, = hd4. The
use of the generator in this construction is entirely equivalent to the genera-

deterministic, self-affine
fractal. (a) At zero order a
rectangular region of width
r, and height h, is
considered. A straight-line
segment extends from (0,O)
to (r,, h,). (b) The first-order
self-affine fractal is given.
This construction also serves
as the generator for higher-
order fractals. (c) Each
straight-line segment in (b) is
replaced by the scaled
generator from (b) to give
the second-order fractal
construction. (d) Each
straight-line segment in (c) is
replaced by the scaled
generator from (b) to give
the third-order fractal. The
construction can be carried
out as far as desired.
tion of the self-similar fractals in Chapter 2. In terms of the formal definition

of a self-affine fractal given above, we havef(xl4, yI2) statistically similar to
f(x, y). Thus the Hausdorff measure can be obtained from r = 114, rHa = 112,
and (114)Ha = 112 or
Our construction of a self-affine fractal is extended to third order in Figure

7.2(d), where each straight line segment of the second-order fractal has been
replaced by the generator.
We next wish to determine the fractal-dimension of our self-affine frac-
tal. To do this we will use the box-counting method. For self-similar fractals
we used square boxes, but for self-affine fractals it will be necessary to use
rectangular boxes. At zero order consider a single box with width ro and
height h,; thus No = 1. At first order we have r, = rd4 so that we will use rec-
tangular boxes with width rd4 and height ho14. We wish to determine how
many of these boxes are required to cover the first-order fractal illustrated in
Figure 7.2(b). This determination is illustrated in Figure 7.3, where we find
N, = 8. Using the standard technique given in Chapter 2 we find from (2.2)
that
We find that the fractal dimension is between 1 and 2, intermediate between

D = 1 for the straight line at zero order and D = 2 for the entire box consid-
ered. For the higher-order constructions illustrated in Figure 7.2(b) and (c)
1 1
we have r, = 16, N2 = 64 and r3 = 64, N3 = 512. For this example the Haus-
i.
dorff measure Ha = and the fractal dimension D = This is consistent with
the relation Ha = 2 - D, which we will derive in general terms.
Figure 7.3. Box-counting
technique applied to the first-
order self-affine fractal given
in Figure 7.2(b). To
determine its fractal
dimension, rectangular boxes
with width r , = r d 4 and
height h , = h d 4 are used. We
find that the shaded N, = 8
boxes (out of 16) cover the
fractal construction. Noting
that No = 1 for the single box
of width ro and height h,, we
find^=+.
7.2Time series
Before continuing our discussion of self-affine fractals, it is appropriate to

introduce some of the fundamental concepts of time series (Chatfield, 1989).
As the name implies, a time series is the set of numerical values of any vari-
able that changes with time. Our consideration of time series is a direct ex-
tension of our consideration of probability and statistics in Chapter 3. Just as
we had continuous and discrete data, we have continuous and discrete time
series. A continuous time series would be a set of values y (t) that are contin-
uous in time over the interval T. An example would be a continuous record
of the atmospheric temperature at a specified point. Another example would
be the discharge down a river measured at a gauging station (Salas,'1993). A
discrete or noncontinuous time series consists of a set of values that are not
continuous. A discrete time series can be obtained from a continuous time
series by sampling it at specified time intervals T, or by integrating the con-
tinuous time series over a specified time interval T. Values are usually speci-
fied at equal increments of time T so that we have values of y (t) given at t =
nT, n = 1, 2, . . . ,N with T = NT.An example of a discrete time series would
be the sequence of daily rainfall totals at a measuring station or the maxi-
mum flood discharge at a gauging station each year.
Time series are good representations of other data sets. For example, the
heights of topography along linear tracks as illustrated in Figure 7.1 are en-
tirely equivalent to continuous time series. Another example would be the
concentration of a mineral (i.e., gold) as a function of depth in a drill core.
The actual concentration would be continuous, with possibly a few excep-
tions, but from practical considerations measurements of concentrations
would be carried out at discrete intervals, giving a discrete time series. Well
logs are another example of a time series in a geological context. Digitized
measurements of density, porosity, and/or permeability at prescribed depth
intervals represent discrete time series. Time series have a wide range of ap-
plications (Box et al., 1994).
Time series may be characterized by discontinuities, a trend component,
one or more periodic components, and a stochastic component. The trend
component is a long-term increase or decrease in the series. The rotational
period of the earth (length of day) can be considered to be a time series. Over
long periods of time the length of day is increasing due to tidal friction,
which is a trend component. Many time series have periodic components;
the atmospheric temperature time series will have strong daily and yearly
periodicities. The stochastic component includes the fluctuations not in-
cluded in either the trend or periodic components. An important aspect of the
stochastic component is whether it is persistent, random, or antipersistent. If
adjacent values in the time series are uncorrelated with each other, then the
stochastic component is random. If adjacent values are positively correlated,
then adjacent values are on average closer than for a random time series, and
the stochastic component is persistent. If adjacent values are negatively cor-

related, then adjacent values are on average further apart than for a random
time series, and the stochastic component is antipersistent. A continuous
time series is, by definition, persistent.
An important question regarding time series is ergodicity. If a time series
is ergodic, an average at a given time over a large number of realizations of
the time series is entirely equivalent to the average in time of a single real-
ization. In general, the ergodicity of practical time series is assumed, al-
though it is often difficult to prove.
The values of the stochastic component of a time series will have a sta-
tistical distribution of values and the discussion given in Chapter 3 is applic-
able. Typical distributions are Gaussian (normal) and log normal. The mean,
variance, and coefficient of variation of the distribution of values can be de-
termined. The persistence (or antipersistence) of the stochastic component
can be quantified by using the autocorrelation function r. For a continuous
time series we have
with
and
The time s is the lag; with s = 0 we have cS= co = V (the variance) and rs = 1.
As s increases, rs generally decreases as the statistical correlations of y (t +
s) with y(t) decrease. The plot of rs versus s is known as a correlogram. A
rapid decay of the correlogram indicates weak persistence (short memory),
and a slow decay indicates strong persistence (long memory). Since the time
series is continuous, it is required that rs + 1 as s + 0.
For a discontinuous time series the autocorrelation function is given by
with
and
with increasing values of k corresponding to increasing lag. For a random

stochastic time series the values of r, will be near zero. Positive values of r,
indicate persistence and negative values indicate antipersistence.
In terms of self-affine fractals, one of the most important aspects of time
series is the question of stationarity. A time series is stationary if the statisti-
cal properties of the time series are independent of its length T. If either the
mean or the variance is a function of T, then the time series is nonstationary.
A measure of long-range correlations that is valid for both stationary and
nonstationary time series is the semivariance y. For a continuous time series
the semivariance is given by
Note that neither the mean j nor the variance V is used in this definition. For
a discontinuous time series we have
with increasing values of k corresponding to increasing lag. For a random

stochastic time series the values of y, will approach the variance V. The plot
of ys versus s or y, versus k is known as a semivariogram (Gallant et al.,
1994). For a stationary time series the autocorrelation function r and the
semivariance y are related by
An alternative measure of long-range correlations is rescaled-range (WS)

analysis, which will be introduced after we consider some examples of time
series.
The classic example of a stationary, discontinuous time series is a white
noise. Consider a random variable E~with a Gaussian (normal) distribution;
the distribution has zero mean and a variance VE= a. If a time series is con-
:
structed with a set of y, = E ~ i, = 1, 2, . . . , n, then adjacent values are uncor-

related and this is known as a white noise. Four examples of white noises are
given in Figure 7.4(a). In each case V , = 1 and a different set of random num-
bers has been used. The classic example of a nonstationary time series is the
Figure 7.4. (a) Four

examples of Gaussian
white noises. Adjacent
values are not correlated.
(b) The four white noises in
(a) are summed to give four
Brownian walks. In each
case the standard deviation
after n steps, given by (7.13),
is included.
Brownian walk. A Brownian walk is simply obtained by summing a Gauss-

ian white-noise sequence; the values in a Brownian-walk time series yBw are
given by
The white-noise sequences shown in Figure 7.4(a) have been summed to

give the four Brownian walks illustrated in Figure 7.4(b). The variance of a
Brownian walk after n values have been summed is given statistically by
where u,2 is the variance of the white-noise sequence. The corresponding

standard deviation of the walk is given by
This result is compared with each of the four Brownian walks illustrated in
Figure 7.4(b). A Brownian walk is an example of a statistical self-affine
fractal.
The association of white noises and Brownian walks is the basis for the
kinetic theory of gases. The distribution of distances in a specified direction
that a molecule in a gas travels between collisions is Gaussian. Thus the se-
quence of distances that a molecule travels in a gas is a Gaussian white
noise. The sum of these distances, the distance the molecule diffuses in the
gas, is a Brownian walk. This is the basic reason that the distance that a con-
taminant diffuses in a gas scales with the square root of time.
Several empirical models have been developed to produce persistent
(correlated) noises (Bras and Rodriguez-Iturbe, 1993). We first consider the
moving average model (MA). In this model the discrete times series is given
by
where E~ is again the random variable described above and the 0, (O,, 0,, . . . ,
Oq) are q prescribed coefficients relating yi to the q previous values of ei.The
parameters in this model are the mean j, the variance of the white noise u,2,
and O,, I,.,. . , Oq.Taking q = 1 the MA model simplifies to
The mean for this correlated noise is j and its variance is
The autocorrelation function is
rk = - for k = l
1 + 0;
r, = 0 for k >1 (7.18)
The correlation is very short since only the adjacent point has a non-zero au-
tocorrelation function. Examples of this time series with 0, = 0,0.2, 0.5,0.9
are given in Figure 7.5.
An alternative empirical model for a persistent (correlated) time series is
the autoregressive model (AR). This time series is given by
+,
where ei is again the random variable and ( j= 1,2, . . . ,p) are prescribed
coefficients relating y i to the p previous values of yi - y. Clearly the MA and
AR models are closely related. In the MA model q previous values of the
random variable ei are included, and in the AR model p previous values of
the deviation of the time series from the mean y, - j are included. Taking
p = 1 the AR model simplifies
L
The mean for this correlated noise is again and its variance is
The autocorrelation function is
The AR model has longer range correlations than the MA model, but the corre-
lations remain short range. Examples of this time series with 4, = 0,0.2,0.5,
and 0.9 are given in Figure 7.6. There is clearly much greater smoothing of the
time series in the AR than in the MA model. Correlograms for these time series
are given in Figure 7.7. The agreement with (7.22) is excellent; the correla-
+,
tions increase systematically with increasing values of as expected.
2
Yn 0
Figure 7.5. Examples of -2

moving average (MA)time -4
series from (7.15) with (a) -6
8, = 0, (b) 8, = 0.2, (c) 0 = 0 128 256 384 512
0.5,and (d) 0, = 0.9; in each n
casej=Oandu,2= 1 . (4
The autoregressive moving average model (ARMA) combines the two

models given above. The ARMA time series can be written as
0 256 384 512

n
6
4
2
Yn 0
-2
Figure 7.6. Examples of
-4
autoregressive (AR) time
-6 series from (7.20) with
0 128 256 384 512
(a) 4, = 0, (b) 4, = 0.2,
n (c) 4, = 0.5, and (d) 4, = 0.9;
fd) in each case j = 0 and o: = 1 .
The variables in this model have been discussed above. With p = 0 the
ARMA model reduces to the moving average (MA) model, with q = 0, the
ARMA model reduces to the autoregressive (AR) model. Taking p = q = 1
the ARMA model simplifies to
Figure 7.7. Correlograms

for the four AR time series
given in Figure 7.6; excellent
agreement with the
theoretical prediction (7.22)
is found.
The mean of this noise is again j and the variance is given by
The autocorrelation function is given by
The time-series models considered above are based on normal (Gaussian)

statistics with both positive and negative values. Each of these time series
can be converted into a skewed, positive only, time series by using the trans-
formation from a normal (Gaussian) distribution to a log-normal distribution
given in (3.30). Just as a stationary white noise can be summed to give a
nonstationary Brownian walk, a stationary ARMA model can be summed to
give a nonstationary autoregressive integrated moving average model
(ARIMA).
7.3 Self-affine time series
Although time series are defined to be sets of values as a function of a single

variable, that is, y(t), time series concepts can be extended to two, three, and
four dimensions. An example in two dimensions is topography h(x,y) as a
function of the two horizontal coordinates x and y. Topography has a statisti-
cal distribution (the hypsometric curve), a mean value relative to sea level,
and a variance. Topography clearly has horizontal persistence (i.e., adjacent
values of topography are correlated). Consider the difference in elevation Ah
between two points separated by a horizontal distance L. Self-similarity of
topography implies that
where Ha is again the Hausdorff measure. Ahnert (1984) found that actual
topography is in excellent agreement with (7.28), taking Ha = 0.635 +
0.105. Similar results were obtained by Dietler and Zhang (1992). Topogra-
phy is an excellent example of a self-affine fractal.
An example of a "time series" in three dimensions is the surface of the

ocean. The surface height h (x, y, t) is a function of the horizontal coordi-
nates x and y but is also a function of time t. Another example would be min-
eral grade C(x, y, z) in an ore deposit as a function of position x, y, z. An ex-
ample in four dimensions is the temperature in the atmosphere T (x, y, z, t) as
a function of position coordinates x, y, z and time t.
In terms of the time-series analysis introduced above, a time series is a
self-affine fractal if its variance scales according to
where Ha is again the Hausdorff measure. Comparing (7.29) with (7.13) we

find that Ha = for a Brownian walk. Thus the four Brownian walks given in
Figure 7.2(b) are self-affine fractals. An equivalent definition of a self-affine
fractal is the requirement that the semivariogram scales such that
Another expression for a self-affine fractal is
where F(y) is the cumulative distribution function of the basic distribution

as defined in (3.10).
The dependence of y on t in a Brownian walk is similar to the shape of
the Koch island illustrated in Figure 2.5. There are variations in the time se-
ries at all scales. Neither the length of the track defined by the time series
nor the local derivatives (slopes) are defined. Thus it is appropriate to con-
sider y(() to be a fractal. Just as in the case of the triadic Koch island and
other fractal constructions on a plane, the fractal dimension is between 1
and 2 and the appropriate Euclidean dimension is 2. With d = 2, (6.12) can
be written
Pr
;--D = constant
where the appropriate scale for the time series is T. In (7.31) the time series
diverges with the interval T according to the power law THO. Comparing
(7.31) and (7.32) we define
This is the basic definition of the fractal dimension for a self-affine fractal.
Below we give an alternative derivation that gives the same result. For the
deterministic self-affine fractal illustrated in Figure 7.2 we found Ha = $ and
D = in agreement with (7.33). For a Brownian walk we also have Ha = 4
a n d ~ = i . ~ 1o<r D < 2 w e r e q u i r e t h a t O < H a < 1.
An alternative derivation of the fractal dimension of a time series can be
obtained by using the box-counting method illustrated in Figure 7.3. We first
introduce a rectangular reference "box" with a width T and height a, = a(T).
Note that since the units of the signal y, and therefore the units of the stan-
dard deviation a , can differ from the unit of time t, the aspect ratio (width/
height) of the box can have arbitrary units. If we measure an electric current
as a function of time, the width of the box is in seconds and its height is in
amperes.
We next divide the time interval T into n smaller time intervals with the
length Ttl = T/n. We also introduce scaled smaller boxes of width T,, and
height a,, = a,/n. These boxes have the same aspect ratio as the reference
box. However, the standard deviation associated with the interval T,, a(Tn)=
a(T/n), is not equal to a,,. We determine the number of scaled smaller boxes
N,, of size T,, X anthat are required to cover the area of width T and height
aTtl.This is given by
However, from (7.29) we have
so that we can write
and combining (7.34) and (7.36) gives
This is basically a fractal relation if we associate T,, with r,,. Comparing

(7.37) with the definition of the fractal dimension in (2.1) we have the rela-
tion 2 - Ha = D, which is identical to (7.33).
A standard approach to the analysis of a time series is to carry out a Fourier

transform of the series. A time series can be prescribed either in the physical
domain as y(t) or in the frequency domain in terms of the spectrum Y ( j T ) ,
where f is the frequency. The quantity Y ( j T ) is generally a complex number
indicating both the amplitude and the phase of the signal. To obtain Y ( j T ) we
use the Fourier transform of y(t) in the interval 0 < t < T; it is given by
where i = fi. The complementary equation relating y(t) to Y ( j 7') is the

inverse Fourier transform
A time series with a single periodic component will have a single spike in its
spectrum at that frequency. A time series with several components will have
spikes in its spectrum at those frequencies. A white noise has no embedded
frequencies and its spectrum is flat. The quantity I Y ( j T)12 df is the contri-
bution to the total energy of y(t) from those components with frequencies be-
tween f and f + d$ The vertical bars in I YI refer to the absolute value of the
complex quantity. The power is obtained by dividing by T. The power spec-
tral density of y(t) is defined by
in the limit T + -. The product S( f)df is the power in the time series associ-
ated with the frequency range between f and f + d$ For a time series that is a
self-affine fractal, the power spectral density has a power-law dependence
on frequency:
We now obtain a relationship between the power P, the Hausdorff measure

Ha, and the fractal dimension D.
We consider two time series y , ( t ) and y2(t) that are related by
The fundamental property of a self-affine fractal time series is that y , ( t )

has the same statistical properties as y,(t). The Fourier transform of y,(t) is
given by
Substituting (7.42) and making the change of variable t' = rt, we obtain
and comparing (7.44) with (7.38) we have
From the definition of the power spectral density given in (7.40) we obtain
Since y2 is a properly rescaled version of y,, their power spectral densities

must also be properly scaled. Thus we can write
From (7.41) and (7.33) it follows that
p=2Ha+1=5-20
For a self-affine fractal (0 < Ha < 1, 1 < D < 2) we have 1 < P < 3. For a
Brownian walk with Ha = $ (D = 23 ) we have P = 2.
7.4 Fractional Gaussian noises and fractional Brownian walks
We next introduce the concept of fractional Gaussian noises. These will be

generated synthetically from Gaussian white noises using the following steps:
(1) A Gaussian white noise sequence is generated. Four examples have

been given in Figure 7.4(a).
(2) A discrete Fourier transform is taken of the random values. The

Fourier coefficients are given by
This transform maps N real numbers (the y n ) into N complex num-

bers (the Ym) Because the transform is taken of a Gaussian white
noise sequence the Fourier spectrum will be flat, that is p = 0 in
(7.41). Except for the statistical scatter the amplitudes of the lYml
will be equal.
(3) The resulting Fourier coefficients Ynlare filtered using the relation
The power PI2 is used because the power spectral density is propor-
tional to the amplitude squared. The amplitudes of the small-m coef-
ficients correspond to short wavelengths Am and large wave numbers
km = 2n/Am. The large-m coefficients correspond to long wave-
lengths and small wave numbers.
(4) An inverse discrete Fourier transform is taken of the filtered Fourier
coefficients. The sequence of points is given by
These points constitute the fractional Gaussian noise. To remove edge

effects (periodicities)only the central portion should be retained.
Several examples of fractional Gaussian noises are given in Figure 7.8.

In each case the Gaussian white noise sequence with p = 0 has been filtered
using the steps given above. Fractional Gaussian noises are given for P =
- 1.0, -0.5, 0.5, and 1.0. Note that the range of ps corresponding to frac-
tional Gaussian noises is - 1 < p < 1.
Just as a Gaussian white noise (p = 0) can be summed to give a Brown-
ian walk (P = 2), a fractional Gaussian noise can be summed to give frac-
tional Brownian walks. In each case we have Pmw= 2 + P,". Fractional
Brownian walks are self-affine fractals and are restricted to the range 1 <
p < 3 as discussed above. The white and fractional Gaussian noises in Figure
7.8(a) with (3 = - 1.0, -0.5, 0, 0.5, and 1.0 have been summed using (7.1 1)
to give the fractional Brownian walks illustrated in Figure 7.8(b) with P =
1.0, 1.5, 2.0, 2.5, and 3.0. Each fractional noise and walk given in Figure 7.8
has been rescaled to have zero mean y = 0 and unit variance V = 1. The
fractional Gaussian noise in Figure 7.8(a) with P = 1.0 is statistically identi-
cal to the fractional Brownian walk in Figure 7.8(b) with P = 1.O.
As the value of p is increased the contribution of the short wavelength
(large wave number) terms is reduced. The result is that adjacent values in
Figure 7.8. (a) The white

Gaussian noise, p = 0,
has been filtered to give
fractional Gaussian noises
with p = - 1.0, -0.5,0.5,
and 1 .O. (b) Each of these
with pGnhas been summed
using (7.11) to give
fractional Brownian walks
with PfBu,= 2 + PfGn,P =
1.0, 1.5, 2.0,2.5, 3.0. The
are self-affine fractals with
D = (5 - P,,")/2.
the time series become increasingly correlated and profiles are smoothed.
The persistence in the time series is increased. This is clearly illustrated in
Figure 7.8 as P is increased from 0 to 3.0. With P = -0.5 and - 1.0 the short-
wavelength contributions dominate over the long-wavelength contributions.
These time series are antipersistent, and adjacent values are less correlated
than for the random white noise (p = 0).
An alternative method for the direct generation of fractional Brownian
walks is the method of successive random additions (Voss 1985a, 1988). Con-
sider the time interval 0 5 t I 1 as illustrated in Figure 7.9. Random
values of y are generated based on the Gaussian probability distribution given
in (3.15) with zero mean j = 0 and unit variance V, = 1. Three of these
i,
values are placed at t = 0, 1 as shown in Figure 7.9(a). Two straight lines are
drawn between these three points. The midpoints of these two line segments are
taken as initial values of y at t = $ and as illustrated in Figure 7.9(b).
The five points are now given random additions. These random addi-
tions are also based on the Gaussian probability distribution (3.15) with zero
mean 7= 0 but with a reduced variance given by (7.29). Since the interval
has been reduced by a factor of two, the variance is given by V2 = ( f ) ~ . . For
our example we take Ha = so that V2 = $.The five resulting random addi-
tions are given in Figure 7.9(c). After addition to the five values of y, in Fig-
ure 7.9(b), the resulting five values of y, are given in Figure 7.9(d). Again
the five points are connected by four straight-line segments and the four
midpoints are taken as initial values of y at t = Q, i,i, and g7 as illustrated in
Figure 7.9(e). All nine points are now given random additions using a Gauss-
ian probability distribution (3.15) with zero mean but a further reduced vari-
ance from (7.28) V3 = Again taking Ha = f we have V3 = 4. 1
The nine
random additions are given in Figure 7.9(f). After addition to the nine values
of y2 given in Figure 7.9(e), the resulting nine values of y, are given in Fig-
ure 7.9(g). The process is repeated until the desired number of points is ob-
tained. Our example with 4097 points is given in Figure 7.9(h). With Ha = i
and (3 = 2 this is a Brownian walk and strongly resembles the Brownian
walks given in Figures 7.4 and 7.8. A sequence of fractional Brownian walks
generated by the method of successive random additions is given in Figure
7.10. Fractional Brownian walks are given for Ha = 0 (P = l), Ha = 0.25
(P = 1.5)- Ha = 0.50 (P = 2), same as Figure 7.9(h), Ha = 0.75 (P = 2.5) and
Ha = 1.00 ( p = 3); in each case 4097 points are given. As expected, these
noises closely resemble those generated by the filtering technique given in
Figure 7.8. A detailed comparison of fractional Gaussian noises and frac-
tional Brownian walks using the Fourier filtering technique and the method
of successive random additions has been given by Gallant et al. (1994).
These authors also considered a third method using Weierstrass- Mandelbrot
functions.
- .
0.0 (a)

generation of a fractional
Brownian walk using the
method of successive
random additions. (a) Three
random numbers are
generated using a Gaussian
distribution with zero mean
and unit variance; these
areplacedatt=O,f, 1.
(b) Values at t = $ and
are obtained by linear
interpolation. (c) Assuming
5
Ha = five random numbers
are generated using a
Gaussian distribution with
zero mean and V = (: )2Ha = f .
(d) The random numbers in
-2 4 I (c) are added to the values in
0.000 (f) 0.125 0.250 0.375 0.500 0.625 0.750 0.875 1.000 i,
(b). (e) values at t = i,{,
t are obtained by linear
interpolation. ( f ) Nine
random numbers are
generated using a Gaussian
distribution with zero mean
and v=(:)~H.=:. (g) The
random numbers in (f ) are
added to the values in (d).
(h) The construction has
been continued to 4097
points; the result is a
Brownian walk.
Figure 7.10. A sequence of

generated by the method of
successive random additions.
(a) H a = 0 (p = 1). (b) H a =
0.25 (P = 1.5). (c) H a = 0.5
(p = 2, a Brownian walk). (d)
H a = 0.75 (P = 2.5). (e) H a =
1.OO (p = 3).
Just as the fractional Gaussian noises generated using the filtering tech-
nique with - I < p I 1 can be summed to give fractional Brownian walks
with 1 I p I 3, the fractional Brownian walks generated using the method of
successive random additions with 1 < P < 3 can be differenced to give frac-
tional Gaussian noises with - 1 6 p I 1. Extended fractional Gaussian noises
with -3 I p I - 1 can be obtained by differencing the fractional Gaussian
noises with - 1 I p < 1. Similarly extended fractional Brownian walks with
3 I p I 5 can be obtained by summing fractional Brownian walks with 1 I
(3 I 3. Although the mathematical definition of self-affine fractals restricts
the applicable range of P to 1 I P I 3, naturally occurring time series with a
power-law dependence of the power spectral density on frequency have val-
ues of p outside this range. Just as naturally occurring self-similar power-
law distributions may or may not fall within the range of D values prescribed
by mathematical constraints, so too naturally occurring self-affine time se-
ries may or may not fall within the range of f3 values prescribed by mathe-
matical constraints.
Using the definition of the semivariance y, given in (7.9), semivari-
ograms for several of the fractional Gaussian noises and fractional Brownian
Figure 7.11. Semivariograms

for several of the fractional
Gaussian noises and
illustrated in Figure 7.8,
T = 4 , 8 , 16,. . . ,2048.The
triangles given for P = 0, 1
are for the filtered fractional
Gaussian noises from Figure
7.8(a). The circles given
for p = I, 2, 3 are for the
summed fractional Brownian
walks given in Figure 7.8(b).
The squares given for P = 3
are for the summed
fractional Brownian walk
1 10 100 1000 10000 1 10 100 1000 10000 with p = 1. The straight-line
Z Z correlations are with (7.30).
walks illustrated in Figure 7.8 are given in Figure 7.11. For the uncorrelated
Gaussian white noise, p = 0, the semivariance scatters statistically about y =
1 as expected since V = 1. For P = 1, 2, and 3 excellent correlations are ob-
tained with the fractal relation (7.30). For P = 2 we find H a = 0.47 compared
to the expected value Ha = 0.50.
The values of Ha obtained for the best fit of (7.30) to the semivari-
ograms in the range - 1 I p 5 5 are given in Figure 7.12. The straight-line
correlation is with the self-affine fractal relation (7.48). Quite good agree-
ment is found in the range 1 < P < 3, where the fractional Brownian walks
are expected to be self-affine fractals.
From Figure 7.12 it is seen that Ha .= 0 for fractional Gaussian noises in
the range - 1 < p < 1. From (7.29) and (7.35) we conclude that the variance
V and standard deviation a are not dependent on the length of the signal T.
Thus these fractional noises are stationary even though adjacent values may
be correlated or anticorrelated. The fractional Brownian walks in the range
1 < p < 3 are clearly nonstationary from (7.29) and (7.35) since H a varies
from 0 to 1 and V and a have a power-law dependence on the length of the
signal T.
Figure 7.12. The dependence

of the Hausdorff measure Ha
on p as obtained from the
best fit of (7.30) to
semivariograms. The
triangles given for - 1 5
p < 1 are for the filtered
given in Figure 7.8(a). The
circles given for 1 5 p 5 3
are for the summed
given in Figure 7.8(b). The
squares given for 3 5 p < 5
are for the extended
fractional walks obtained by
summing the fractional
Brownian walks with 1 I P I
3. The straight-line
correlation is with the self-
affine fractal relation (7.48)
for 1 5 p S 3 .
7.5 Fractional log-normal noises and walks
The fractional Gaussian noises and fractional Brownian walks we have con-
sidered have both been based on a Gaussian distribution of values. Thus the
resulting time series have both positive and negative values. Many naturally
occurring time series have only positive values. For example, the volumetric
flow in a river Q ( t ) is always positive. Another example is the density or
porosity in a well, which is also always positive. The coefficient of variation
cVis the ratio of the standard deviation of the signal to its mean (3.33). If
(a) c , = 0.2
0 128 256 t 384 512 0 128 256 t 384 512
(b) C , = 0.5
Figure 7.13. Examples of

fractional log-normal noises
and walks. In each case a
fractional Gaussian noise or
a fractional Brownian walk
has been converted to a
(c) c , = 1.0 fractional log-normal noise
or walk, examples are given
for p = 0, 1.2. The
conversions were made
using (3.31) to (3.33).
Examples are given for: (a)
~,,=0.2,(b)c,,=O.5,(c)c~=
1 .O, and (d) cV= 2.0. In all
cases the mean of the time
(d) c , = 2.0 series is unity f = 1.
cV< < 1 it may be appropriate to consider Gaussian statistics. In many cases,

however, this will be a poor approximation.
The most widely used positive distribution is the log-normal distribution
discussed in Chapter 3. A normal distribution can be converted to a log-nor-
ma1 distribution using the relation y = In x where x is a log-normal distribu-
tion and y is a normal distribution as given in (3.30). However, a log-normal
distribution can be specified only if its coefficient of variation is given as de-
fined in (3.33). A Gaussian white-noise sequence can then be converted to a
log-normal, white-noise sequence using (3.31) and (3.32). Log-normal
white-noise sequences, p = 0, with unit mean, 2 = 1, are given in Figure 7.13
for cv = 0.2, 0.5, 1.0, and 2.0. With cv = 0.2 the standard deviation is small
compared with the mean, the distribution is nearly symmetric, and it closely
resembles a Gaussian white noise. With cv = 2 the variance is large compared
with the mean and the distribution is strongly asymmetrical.
Just as a Gaussian white-noise sequence can be converted to a log-nor-
mal, white-noise sequence, so too can fractional Gaussian noises and frac-
tional Brownian walks be converted to fractional log-normal noises and
walks using (3.31) to (3.33). Several examples are given in Figure 7.13. In
each case the mean is unity i = I . This is a two-parameter family of noises
and walks. The values of (3 are a measure of the persistence of the time se-
ries. The values of cv are a measure of the skew of the distribution of values.
Extensive studies of fractional log-normal noises and walks have been given
by Mandelbrot and Wallis (1968). These authors referred to the dependence
on cV as the Noah effect and the dependence on P as the Joseph effect. The
time series in Figure 7.13 resemble typical river flow time series. Increasing
cV,the Noah effect, is indicative of a climate where there is large variability
in river flow. Increasing P, the Joseph effect, is indicative of more strongly
correlated values. With higher values of P a year of flood is more likely to
follow a previous year of flood, and a year of drought is more likely to fol-
low a previous year of drought.
7.6 Resealed-range (R/S)analysis
An alternative approach to the quantification of correlations in time series

was developed by Henry Hurst (Hurst et al., 1965). Hurst spent his life
studying the hydrology of the Nile River, in particular the record of floods
and droughts. He considered a river flow as a time series and determined the
storage limits in an idealized reservoir. On the basis of these studies he intro-
duced empirically the concept of rescaled-range ( R B ) analysis. His method
is illustrated in Figure 7.14. Consider a reservoir behind a dam that never
overflows or empties; the flow into the reservoir is the flow in the river up-
stream of the dam Q(t). The flow out of the reservoir Q(T) is assumed to be
the mean of the flow into the reservoir for a period T
The volume of water in the reservoir as a function of time V(t) is given by
V(t) = V(0) +
1Q(tf)dt' - tQ (7.53)
where V(0) is the volume of water at t = 0. Taking t = T and substituting

(7.52) into (7.53), we have V(T) = V(0). The range R(T) is defined to be the
difference between the maximum volume of water Vmaxand the minimum
volume of water Vminduring the period T
The rescaled range is defined to be R/S where S is the standard deviation of

the flow Q(t) during the period T
This is identical to the definition of a introduced in (3.3). We use S here to

maintain the standard R/S nomenclature.
Hurst et al. (1965) found empirically that many data sets in nature sat-
isfy the power-law relation

how resealed-range (RB)
analysis is camed out. The
flow into a reservoir is Q(t)
a
and the flow out is The
maximum volume of water
in the reservoir during the
period T is Vmax( T ) and the
minimum Vmln ( T ) ;the
difference is the range
R (7') = Vmax( T ) - Vmin (TI.
where Hu is known as the Hurst exponent. Examples included river dis-

charges and lake levels, thicknesses of tree rings and vanes, atmospheric
temperature and pressure, and sunspot numbers. They generally found that
0.70 < Hu < 0.80.
The WS analysis is easily extended to discrete time series. The running
sum of the series relative to its mean is
The range is defined by
with
and and a, are obtained from (3.1) and (3.3). The Hurst exponent, Hu,is
obtained from
For example, if 64 values of yi (i = 1, 2, . . . , 64) are available for a time se-

ries, the R, and S, for N = 64 are obtained. Then the data are broken into
two parts (i = 1, 2, . . . , 32 and i = 33, 34, . . . , 64) and values for R,, and
S,, are obtained for the two parts. The two values of R,,IS,, are then aver-
aged to give (R,,IS,,),,. The data set is then broken into four parts (i = 1,
2 , . . . , l 6 ; i = l 7 , 1 8 , . . . , 3 2 ; i = 3 3 , 3 4 , . . . , 4 8 ; a n d i = 4 9 , 5 0 , . . . ,64)the
values of R,gS,, are obtained for the four parts and are averaged to give
(RIJSl,)aV.This process is continued for N = 8 and N = 4 to give (R,ISJaV and
(R41S4)aV. For N = 2 we have R, = S, so that R21S2= 1. The values of log
(R,IS,),, are plotted against log (Nl2) and the best-fit straight line gives Hu
from (7.60). In practice there is generally some curvature of values for small
N and the values of (R,IS,),, for small values of N are omitted.
The running sum of a Gaussian white noise with P = 0 is a Brownian
walk with P = 2 and Ha = 0.5. This would imply that
And from (7.48) that
Since a white noise p = 0 is a random process it would be appropriate to con-

clude that Hu = 0.5 implies randomness. It follows that 0.5 < Hu < 1.0 im-
plies persistence and that 0 < Hu < 0.5 implies antipersistence.
The Hurst resealed-range analysis was first applied to Gaussian fractional
noises and fractional Brownian walks by Mandelbrot and Wallis (1969a). The
dependence of log (WS)," on log 7 (log N) for several of the fractional noises
and walks illustrated in Figure 7.8 are given in Figure 7.15. For P = 0, 1, and
2 excellent correlations with the Hurst relation (7.60) are obtained. For P = 0
we find H u = 0.56 compared with the expected value of 0.5 for the random
white Gaussian noise.
Figure 7.15. Hurst WS

analyses for several of the
and fractional Brownian
walks illustrated in Figure
7.8. Average va'lues of WS
are given as a function of
the interval N for N = 4, 8,
16, . . . ,4096. The triangles
given for p = - 1,0, 1 are
for the filtered fractional
Gaussian noises from Figure
7.8(a). The diamonds given
for p = - 1 are for the
extended fractional Gaussian
noise obtained by
differencing the filtered
with p = 1. The circles given
for p = 1.2areforthe
summed fractional Brownian
walks given in Figure 7.8(b).
The straight-line correlation
is with (7.60).
The values of Hu obtained for the best fit to the Hurst relation (7.60) in
the range -3 I 5 3 are given in Figure 7.16. The straight-line correlation is
with (7.61). Reasonably good agreement is found in the range - 1 < P < 1.
The Hurst exponent provides a quantitative measure of persistence and an-
tipersistence for fractional Gaussian noises. Extensive R/S analyses of frac-
tional Gaussian noises and fractional Brownian walks have been carried out
by Bassingthwaighte and Raymond (1994).
Several closely related techniques have been introduced to quantify the
self-affine properties of observed time series. Malinverno (1990) introduce
the roughness-length method. In this method the local trend is determined as
a function of window length. Ivanov (1994a, b) introduced counterscaling.
Two types of counterscaling were considered. In the first the variance was
determined for different window lengths. In the second the means were ob-
tained for various window lengths and the variances of these means were ob-
tained as a function of window length. Gomes Da Silva and Turcotte (1994)
have applied the counterscaling technique to fractional noises and walks.

of the Hurst exponent Hu on
p as obtained for the best-fit
correlations with the Hurst
relation (7.60). The triangles
given for - 1 I p I 1 are for
the filtered fractional
Gaussian noises given in
Figure 7.8(a). The circles
given for 1 < fiI 3 are for
the summed fractional
Brownian walks given in
Figure 7.8(b). The diamonds
given for -3 I p S1 are for
the extended fractional
Gaussian noises obtained by
differencing the filtered
with - 1 I p I l . T h e
straight-line correlation is
with (7.61).
7.7 Applications of self-affine fractals
Three examples of elevation along linear tracks were given in Figure 7.1.
These are equivalent to time series and are examples of naturally occurring
self-affine fractals. Many authors have carried out Fourier spectral analyses
of topography and bathymetry along linear tracks (Bell, 1975, 1979; Berk-
son and Mathews, 1983; Barenblatt et al., 1984; Fox and Hayes, 1985;
Gilbert and Malinverno, 1988; Fox, 1989; Gilbert, 1989; Malinverno, 1989,
1995; Mareschal, 1989). In general a power-law dependence of the power
spectral density on wave number was found with P = 2 (D = 1.5). Similar
results were obtained for Venus by Kucinskas et al. (1992). Thus the eleva-
tion of topography is approximately a Brownian walk. Twenty-four exam-
ples of the dependence of the power spectral density on wave number for lin-
ear topographic profiles from three different parts of Oregon are given in
Figure 7.17. One-dimensional Fourier spectral analyses were obtained using
Figure 7.17. Plots of one-

dimensional power spectral
density versus wave number
selected from three regions
with different tectonic and
geomorphic settings in
Oregon. The profiles are
offset vertically so as not to
overlap. (a) Willamette
lowland, (b) Wallowa
Mountains, (c) Klamath
Falls.
the periodogram method. Three different regions were considered with dif-
ferent geomorphic and tectonic settings. The Willamette lowland is domi-
nated by sedimentary processes, the Wallowa Mountains are associated with
a major tectonic uplift, and the Klamath Falls area belongs to the basin and
range tectonic regime. The topography was digitized along lines of latitude
and longitude at seven points per kilometer. For each of the three regions, 20
equally spaced one-dimensional profiles of length 5 12 points were analyzed
in both the latitudinal and longitudinal directions. Log-log plots of the spec-
tral power density versus wave number show a good power-law dependence
in all three regions, as shown in Figure 7.17. Eight typical examples are given
for each of the three regions. The best-fit fractal dimension for each profile
is obtained using (7.41). The mean fractal dimensions for three regions are
given in Table 7.1. The mean values are close to D = 1.5, indicating that the
spectral power density corresponds to a Brownian walk to a good approxi-
mation. A variety of previous studies have found values for D near 1.5.
Power Spectra
track 8
track 7
track 6
track 5
track I
Klarnath
Figure 17. (cont.)

Two implications of this result will be discussed. The first is the compar-
ison with the value of D for topography obtained in Chapter 2 using the ruler
method. As illustrated in Figure 2.7 the ruler method generally gives fractal
dimensions near D = 1.2. These are systematically lower than the values near
D = 1.5 obtained using the spectral method. Fundamentally there is no rea-
son why the two fractal dimensions should be equal. Elevation profiles are
not necessarily related to the shape of contours.
The correspondence of topography and bathymetry to a Brownian walk
also implies, importantly, that they are truly self-similar. For a Brownian
walk the amplitude coefficients are directly proportional to the correspond-
ing wavelengths. Thus the height-to-width ratios of mountains and hills are
the same at all scales.
It should also be noted that the power-law spectra given in Figure 7.17
provide further information beyond the fractal dimension. The spectra are
characterized by the amplitude in addition to the slope. A quantitative mea-
sure of the amplitude is the intercept (value of S) at a specified wave number
(k = 1 cycle km-1).These reference amplitudes are a measure of the rough-
ness of the topography. The mean intercepts for the latitudinal and longitudi-
nal directions for the three regions in Oregon are given in Table 7.1.
Another application of spectral techniques is to well logs. It is common
practice to make a variety of measurements as a function of depth in oil
wells. Typical measurements include the local acoustic velocity, the electri-
cal conductivity, and neutron activation. The measured quantities are ob-
tained as a function of depth so that they are equivalent to a time series and
spectral techniques can be applied.
Table 7.1. Regional averages over one-dimensional profiles of Oregon

topography
Area Fractal Dimension Roughness

Willarnette lowland
Latitude 1.436 5.948
Longitude 1SO7 6.354
Wallowa Mountains
Longitude 1.485 6.830
Klamath Falls
Longitude 1SO0 5.963
The power spectral densities obtained from porosity logs for eight wells
in the Gulf of Mexico are given in Figure 7.18 (Pelletier and Turcotte, 1996).
At spatial scales greater than 10 ft a good correlation is obtained with the
fractal relation (7.41) taking P in the range 1.31-1.6. Below this scale the
variability decreases significantly in most wells. This may be attributed to
increased homogeneity within beds. Todoeschuck et al. (1990) have also
considered the fractal behavior of well logs. Leary and Abercrombie (1994)
attribute the shear-wave source and coda-wave displacement spectra ob-
tained from seismograms in the Cajon Pass borehole to scattering that was
observed to obey power-law spectra from well logs.
There are many other examples of measurements in geology and geo-
physics that yield power-law spectra. Brown and Scholz (1985) have carried
out spectral studies of natural rock surfaces. They generally find fractal be-
havior with a relatively large range of variability between 1 < D < 1.6. Sim-
ilar studies have been carried out by Power and Tullis (1991, 1995) and by
Pyrak-Nolte et al. (1995). The observation that fracture surfaces are self-
affine fractals can be used to scale the fluid permeability associated with
fractures.
An interesting question is whether climate obeys fractal statistics (Nico-
lis and Nicolis, 1984). Fluigeman and Snow (1989) have shown that the spa-

the spectral power density on
wave number for porosity
logs from eight wells in the
Gulf of Mexico (Pelletier
and Turcotte, 1996). The
straight-line correlations are
with (7.41).
tial distribution of oxygen isotope ratios in sea floor cores obey fractal spec-
tral statistics. Since it is generally accepted that the isotope ratios are propor-
tional to the local temperatures, these results can be taken as evidence that
climate obeys fractal statistics. Hsui et al. (1993) have shown that variations
of sea level with time are a self-affine fractal. This is consistent with the
fractal distribution of sedimentary hiatuses discussed in Chapter 2.
An important application of power-law spectra is in interpolating be-
tween measured data sets. Consider the bathymetry of the oceans. Bathyme-
try is typically measured from ships along linear tracks and must be interpo-
lated to make bathymetric charts. This interpolation can make use of the fact
that the bathymetry has a power-law spectrum. The amplitude coefficients
are determined from the applicable fractal relation and the data are used to
determine the phases in a two-dimensional Fourier expansion of the bathym-
etry. This method can also be used to interpolate airborne magnetic surveys.
Hewett (1986) has used fractal techniques to interpolate well-log poros-
ity data from production wells to obtain the full three-dimensional porosity
distribution in an oil field. The horizontal variations in porosity are treated
as a Brownian walk in analogy to the fractal behavior of topography, and the
fractal behavior of the vertical variations are obtained directly from the well
logs. Molz and Boman (1993) have used this technique to interpolate well-
log data to predict the ground water movement and pollutant dispersion ad-
jacent to a waste disposal site.
In some cases a time series x(t) will have a well-defined correlation di-
mension (Grassberger and Procaccia, 1983a, b). A vector for the time series
at t = t, is defined by the quantities x(t,), x(t, + T), x (t, + 27), . . . , x (t, +
n ~ )At
. a later time t = t, another vector is defined by the quantities x (t,), x
(t, + T),x (t, + 27), . . . ,x (t2 + "7). As long as the signals at t, and t, are un-
correlated, the delay T can be small. This process is known as forming n-
tuples and n is the embedding dimension. The cumulative number of pairs of
points N separated by a distance less than r is plotted against r for embed-
ding dimensions n = 2, 3, 4, . . . . If a straight-line correlation is obtained
-
such that N (r, n) rd and if d becomes independent of n, for all values of d
greater than dc,then dc is the correlation dimension. Smith and Shaw (1990)
have applied this technique to sea-floor bathymetry. Cortini and Barton
(1993) analyzed the inflation-deflation patterns of an active volcanic cal-
dera (Campi Flegrei, Italy) as a self-affine time series and were able to make
successful forward predictions. Nicholl et al. (1994) have applied this ap-
proach to the sequence of intervals between eruptions of Old Faithful geyser
in Yellowstone National Park, Wyoming. They conclude that the sequence of
eruption intervals is a chaotic time series. Osbourne and Provenzale (1989)
have obtained values of the correlation dimension for fractional Brownian
walks. These authors find a systematic dependence with dc = 1.140 0.005 +
for p = 3, d c = 2.053 + 0.009 for f3 = 2, dc = 3.749 5 0.015 for P = 1.5, and

dc> 10forp=1.
We now turn to two-dimensional spectral studies. Of particular interest
is the earth's topography and bathymetry.
It is common practice to expand data sets on the surface of the earth in
terms of spherical harmonics; examples include topography and geoid. Us-
ing topography as an example the appropriate expansion for the radius r of
the earth is
+
where a, is a reference earth radius, 8 is latitude, is longitude, Cc, and Cs,
are coefficients, and P
, are associated Legendre functions fully normalized
so that
The variance of the spectra for order 1 is defined by
and the power spectral density is defined by
where ko is the wave number and A, = Ilko is the wavelength over which data
are included in the expansion. With A, = 2naOwe have
A fractal dependence can be defined if S, has a power-law relation to the

wave number k,.
The power spectral density of the earth's topography and bathymetry as
a function of wave number is given in Figure 7.19. This is based on the
Figure 7.19. Power spectral

density S as a function of
wave number k for spherical
harmonic expansions of
topography (degree I ) for (a)
the earth (Rapp, 1989) and
(b) Venus (Kucinskas and
Turcotte, 1994). Correlations
are with (7.41) taking P = 2.
spherical harmonic expansion to order 1 = 180 given by Rapp (1989). Except

for the low-degree harmonics an excellent correlation is obtained with (7.41)
taking P = 2 (D = 1.5). The spectral dependence of topography corresponds
to a Brownian walk as previously noted.
The power spectral density of topography on Venus (Kucinskas and Tur-
cotte, 1994) is given in Figure 7.19(b), and again a good correlation with a
Brownian walk (P = 2) is obtained. The spectral power density on Venus is
about an order of magnitude less than on the earth for the same wave num-
ber. This is attributed to the higher surface temperature and, thus, weaker
lithosphere on Venus. Nevertheless, it is somewhat surprising that topogra-
phy on both the earth and Venus are Brownian walks since erosion and depo-
sition are dominant in the evolution of many landforms on the earth, whereas
these processes are essentially absent on Venus so that tectonic processes are
dominant. This suggests that the tectonic processes that build topography
and the erosional and depositional processes that destroy topography both
give Brownian walk statistics. Once again the fractal dimension of topogra-
phy does not appear to be diagnostic.
The power spectral density of the earth's geoid is given as a function of
wave number in Figure 7.20; this was compiled by Turcotte (1987) from the
data given by Reigber et al. (1985). Except for the low-degree harmonics an
lo8
10'
Sg
m2 lo6
cycles
km
1 o5
Figure 7.20. Power spectral

density of the earth's geoid
as a function of wave
1o4
number. The circles
represent a compilation of
the data of Reigber et al.
103 J 1 1 1 1 1 I 1 I I , , ,
(1985). The solid line
represents (7.41) with P =
1 o4 . cycles 1o -~
3.5 (D = 0.75).
excellent correlation is obtained with (7.41) taking P = 3.5 (D = 0.75). This

is another example that falls outside the range of fractional Brownian walks
and the formal limits for fractal behavior.
The Fourier spectral approach to fractal analysis for one-dimensional
profiles discussed previously can be extended to two-dimensional image
analysis (Dubuc et al., 1989b). Consider an N X N grid of equally spaced
data points in a square with linear size L. The W data points are denoted by
hrImwith (n,m)specifying the position in the x- and y-directions respectively.
A case with N = 8 is illustrated in Figure 7.2 1.
The first step is to carry out a two-dimensional discrete Fourier trans-
form on the W set of data points hnm.An N X N array of complex coeffi-
cients HF,is obtained by the usual definition
18 28 38 48 58 68 78 88
17 27 37 47 57 67 77 87
16 26 36 46 56 66 76 86
m 15 25 35 45 55 65 75 85
14 24 34 44 54 64 74 84
13 23 33 43 53 63 73 83
12 22 32 42 52 62 72 82
11 21 31 41 51 61 71 81
n
(a) The 64 nm coefficients for an 8 x 8 sub-set of raw data.
8 8 8
7 7 8 8
6 7 7 7 8
5 5 6 7 7 8
4 5 5 6 7 8 8
3 4 5 5 7 7 8
2 3 4 5 6 7 8
2 3 4 5 6 7 8
2 3 4 5 6 7 8
S Figure 7.21. Illustration of
subscript arrangement in
( b ) Equivalent radial coefficients r for various coefficients two-dimensional spectral
s and r in spatial frequency space. analysis.
where s denotes the transform in the x-direction (s = 0, 1,2, . . . ,N - 1) and

t denotes the transform in the y-direction (t = 0, 1, 2, . . . ,N - 1). Then each
transform coefficient Hstis assigned an equivalent radial number using the
relation
The two-dimensional mean power spectral density S2j for each radial wave
number k, is given by
where N,is the number of coefficients that satisfy the condition j < r <j + 1
and the summation is carried out over the coefficients Hs, in this range. The
coefficients assigned to each interval for the example given in Figure 7.21(a)
are illustrated in Figure 7.21(b).
The dependence of the mean power spectral density on the radial wave
number k, for a fractal distribution is (Voss, 1988)
instead of (7.41). The addition of minus one to the power is required because
of the radial coordinates that are used in phase space. The dependence of
V(L) on L given in (7.29) is still valid but with the additional dimension the
"box" derivation that follows now gives
for the fractal dimension of the surface instead of (7.33). Similarly, the de-
rivation of the relationship between P and Ha must be reexamined but
remains valid. Combining (7.73) and (7.74) gives
for the two-dimensional case.

Synthetic images can be generated using the same technique used to

generate a synthetic fractional Gaussian noise. The method used to generate
images is as follows.
We consider an N X N square grid consisting of NZ equally spaced

points. Each point is given a random value hmnbased on the Gauss-
ian probability distribution defined in (3.15). A typical example is il-
lustrated in Figure 7.22a. This is Gaussian white noise so that adja-
cent points are totally uncorrelated and P = 0.
A two-dimensional discrete Fourier transform is taken using (7.69)
generating an N X N array of complex coefficients Hs,.
A fractal dimension D, is specified and the corresponding value for
p is obtained from (7.75). A new set of complex coefficients are ob-
tained from the relation
An inverse two-dimensional discrete Fourier transform is carried out

to generate a new image.
Five examples of synthetically generated images are given in Figure

7.22. In (b) P = 1.2 (D2=2.9), in (c) P = 1.6 (D,= 2.7), in (d) P =2.0(D2=
2.5). in (e) P = 2.4 (D, = 2.3), and in (f) P = 2.8 (D, = 2.1). The synthetic re-
sult for D, = 2.5 looks quite realistic for a typical topographic map.
We have carried out one-dimensional spectral decompositions of linear
profiles of our synthetic data. The results for synthetic topography with D, =
2.6,2.7 and 3.0 are given in Table 7.2. For realistic topography with D, = 2.6
we find that the corresponding one-dimensional profiles give D, = 1.58. This
is consistent with the previously published results for one-dimensional ba-
thymetric and topographic profiles where values near D = 1.5 have been
Table 7.2. Summary of mean fractal dimensions estimated by one-

dimensional and two-dimensional spectral analysis for the topography
of Oregon and for synthetic images
Average D
Two-dimensional One-dimensional
Data analysis analysis
Oregon topography
Synthetic topography
Figure 7.22. Synthetic fractal

images on a 256 X 256 grid.
(a) White noise without
fractal filtering. (b) Filtered
white noise with f3 = 1.2 and
D = 2.9. (c) Filtered white
noise with p= 1.6 and D =
2.7. (d) Filtered with p = 2.0
and D = 2.5. (e) Filtered with
p=2.4andD=2.3 (f)
Filtered with p = 2.8 and D =
2.1.
At the time of going to press a color version of this figure was available for download from http:llwww.cambridge.org,978052 1567336
found as discussed above. In general the relation D,= D, + 1 is a good ap-

proximation.
The results given above can be extended to carry out fractal mapping of
digitized images. As a specific example we consider the digitized topogra-
phy of the state of Oregon (Huang and Turcotte, 1990b). Combining Defense
Mapping Agency (DMA) 1" X 1" data with topographic maps, the U.S. Geo-
logical Survey (Flagstaff) has produced digitized topography on a grid scale
of about seven points per kilometer. The topography for Oregon is illustrated
in Figure 7.23.
The fact that fractal statistics are a good approximation for topography
allows us to make fractal maps of a region of diverse tectonics. Using the dig-
itized topography of Oregon, plots of power spectral density versus wave num-
ber are made for subregions. From these plots a fractal dimension (slope) and
unit wave number amplitude are obtained for each subregion. The amplitude
is a measure of roughness. We are basically carrying out a texture analysis us-
ing the fractal statistics as a basis.
Figure 7.23. Map of the

digitized topography for
Oregon. Data resolution is
about seven points per
kilometer. The width of the
state of Oregon is about 375
miles.
At the time of going to press a color version of this figure was available far download from http://www.cambridge.0'g/978052l567336
Fractal dimensions and roughness amplitudes are obtained using suhre-

gions of 32 X 32 data points. Thus fractal dimensions and roughness am-
plitudes are obtained for each 4.5 km X 4.5 km subregion in the state; maps
are generated. The 32 X 32 set was chosen because it generally gives well-
defined fractal spectra; for smaller regions the errors in fractal dimension
and roughness become substantially larger. For larger regions, the spatial
resolution of the map is degraded.
The following technique is used to obtain a fractal dimension and rough-
ness amplitude for each subregion.
A 32 X 32 set of digitized elevations is chosen to form each subre-

gion (N = 32).
The mean and linear trends for each subset of data are removed.
A two-dimensional discrete Fourier transform is carried out, and an
N X N array of complex Fourier coefficients H,,is obtained using
(7.69).
Each coefficient Hs, is assigned an equivalent radial wave number r
using (7.70). The two-dimensional mean spectra energy density SZj
is obtained for each radial integer wave number k, using (7.72).
The mean slope on a log-log plot of +j versus kj obtained by a least-
squares regression yields a fractal dmension D, using (7.72) and
Figure 7.24. Plots of mean

power spectral density
versus radial wave number
for four typical 32 x 32
point subregions in Oregon.
The linear trend on the
Log-log plot indicates a
power-law (fractal)
distribution.
-0.- -w.s V.V V.L v.- J.6
log(k, krn-l)
At the time of going to press a color version of this figure was available for download fmm http:l/www.cambridge.org/9780521567336
(7.75); the intercept at kj = 1 cycle kmm' yields a roughness ampli-

tude.
Examples for four randomly selected subregions in Oregon are given in

Figure 7.24. The mean two-dimensional fractal dimension for all of Oregon
is D, = 2.586 (Table 7.2). This is remarkably close to the mean value D, =
2.59 that was obtained for the state of Arizona (Huang and Turcotte, 1989).
It is seen that the results are in good agreement with the relation D,,= 1 D. +
Maps of fractal dimension and roughness amplitude are given in Figure
7.25. As expected there is relatively little variation in the fractal dimension
about the mean value, although the range is from about 2.4 < D, < 2.9. The
variation in the roughness amplitude in Figure 7.25(b) is much more impres-
sive. The sedimentary Willamette lowland shows low overall roughness
while the erosion system associated with the nearby mountain ranges and the
Wallowa Mountains in the northeast stand out as regions of high roughness.
Figure 7.25. Maps of (a)

fractal dimension and (b)
roughness amplitude for
Oregon. There is generally
limited systematic variation
in the fractal dimension:
however, the roughness
amplitude is sensitive to
texture changes.
At the time of going to press a color version of this figure was available for download from hftp://ww.cambridgeeorg/s78O52I567336
The roughness contrasts in the southern basin and range region are also quite
remarkable. The fractal analysis gives a quantitative measure of roughness.
In this chapter we have shown that we typically have D, = 1.5 and D, =
2.5 for self-affine topography. In Chapter 2 we found that the self-similar
fractal dimension of topography is near Dm = 1.25. Kondev and Henley
(1995) have generated synthetic two-dimensional topography and have stud-
ied the self-similar fractal behavior of topographic contours. They argue that
Thus for two-dimensional Brown topography with D2 = 2.5, they would ob-
tain Dss = 1.25, in good agreement with many observations.
Figure 25. (conr.)

Problems
Problem 7.1. Consider the deterministic first-order self-affine fractal con-

struction illustrated in Figure 7.26(a). The vertical scale is divided into
three equal parts and the horizontal scale is divided into nine equal parts.
(a) Extend the construction to second order. (b) How many boxes with
dimensions hd9 by rd9 are required to cover the first-order construc-
tion? What is the corresponding fractal dimension?
struction illustrated in Figure 7.26(b). The vertical scale is divided into
three equal parts and the horizontal scale is divided into five equal parts.
Figure 7.26. Four examples

of deterministic self-affine
fractal constructions.

struction illustrated in Figure 7.26(c). The vertical scale is divided into
three equal parts and the horizontal scale is divided into seven equal
parts. (a) Extend the construction to second order. (b) How many boxes
with dimensions hd7 by rd7 are required to cover the first-order con-
struction? What is the corresponding fractal dimension?
struction illustrated in Figure 7.26(d). The vertical scale is divided into
three equal parts and the horizontal scale is divided into six equal parts.
Problem 7.5. Consider the 16-year record of annual rainfall totals for Miami,
FL, given in Table 7.3. (a) Determine the mean, variance, standard de-
viation, and coefficient of variation for these values. (b) Determine
(RNISN)aV for N = 4, 8, and 16.
Problem 7.6. Consider the 16-year record of annual rainfall totals for New
York, NY, given in Table 7.3. (a) Determine the mean, variance, standard
deviation, and coefficient of variation for these values. (b) Determine
(RNISN),V for N = 4, 8, and 16.
Table 7.3. Annual rainfall totals in millimeterslyear for the 16 years

between 1969 and 1984 for Miami, New York, Seattle, and Phoenix
Miami New York Seattle Phoenix

Problem 7.7. Consider the 16-year record of annual rainfall totals for
Phoenix, AR, given in Table 7.3. (a) Determine the mean, variance, stan-
dard deviation, and coefficient of variation for these values. (b) Deter-
mine (RNISN),,for N = 4,8, and 16.
Problem 7.8. Consider the 16-year record of annual rainfall totals for Seattle,
WA, given in Table 7.3. (a) Determine the mean variance, standard devi-
ation, and coefficient of variation for these values. (b) Determine
(RNISN)aV for N = 4, 8, and 16.
Problem 7.9. Derive (7.16). It is appropriate to assume --x

1"
&Zi
n .i Z.l
= uzand -x,
1 "
n,,. -
E,E,-, = 0 since adjacent values of ei are uncorrelated.
Problem 7.10. Derive (7.17) and (7.18), use the relations given in Problem
7.9.
Problem 7.1 1. Obtain an expression for the semivariance y, for the MA
model given in (7.15).
Problem 7.12. The following set of random numbers have a Gaussian distri-
bution with zero mean: -0.4287, -0.0541,0.6224, -0.9545, -0.3745,
0.0455, -1.0512, 0.3431, 0.1318, -0.6346, 0.4436, 0.3743, 0.4589,
1.3667, -0.403 1,O.1154. Use (7.15) to determine a MA time series with
0 = 0.5, j = 0, using these random numbers. Determine the variance of
the time series and compare the result with the predicted value from
(7.16). 1"
Problem 7.13. Derive (7.21). It is appropriate to assume that -CE~(Y;-, -
ni,l
j ) = 0 and -x
1"
ni,l
(y,- 1 - j)2 = 02 as well as the relations given in
Problem 7.9.
Problem 7.14. Obtain an expression for the semivariance y, for the AR
model given in (7.20).
Problem 7.15. Using the set of random numbers given in Problem 7.13, de-
termine an AR time series from (7.20) with +,
= 0.5 and j = 0. Deter-
mine the variance of the time series and compare the result with the pre-
dicted value from (7.21). Determine the auto correlation function for k =
1 , 2 and compare the results with the predicted values from (7.22).
Problem 7.16. Consider the 16 random numbers given in Problem 7.13. De-
termine (RNISJaVfor N = 4, 8, and 16. Determine the best fit value of
Hu from (7.60).
Problem 7.17. The definition of red noise is P = 1. What is the fractal dimen-
sion? How do the variance V and standard deviation a depend upon the inter-
val T? Is red noise an example of a fractional Brownian walk; if so, why?
Problem 7.18. Determine the aspect ratio (height-to-width ratio of the moun-
tains and valleys) using the correlation line from Figure 7.18. From this
case the spectral power density is defined as the amplitude coefficient

squared times the circumference of the earth.
Problem 7.19. Determine the ratio of the height of topography on Venus to
earth using the data in Figure 7.18. Note that the spectral power density
is defined as the product of the square of the amplitude and the circum-
ference of the body.
Problem 7.20. In W S analysis for N = 2 show that R, = S,.
Chapter Eight
GEOMORPHOLOGY
8.1 Drainage networks
In the previous chapters we concluded that landscapes generally obey fractal

statistics. Analyses of shore lines and topographic contours using ruler or
box-counting methods provide statistical correlations with the self-similar
fractal relation (2.1). Spectral studies of topography and bathymetry corre-
late well with the self-affine fractal relation (7.41). We will show in this
chapter that drainage networks are fractal trees. Landforms evolve as a result
of the tectonic processes that produce them and the erosional processes that
destroy them. Landforms are a classic example of a complex phenomenon
that can be quantified using fractal concepts (Turcotte 1993, 1994b, 1995).
We will begin our study of geomorphic processes by considering
drainage networks. Drainage networks are a universal feature of landscapes
on the earth. Over a large fraction of land areas, water eventually flows into
an ocean. The surface area that drains into an ocean through a river defines
the drainage basin of that river. The streams and rivers that drain into the
Mississippi River and eventually drain into the Gulf of Mexico define the
Mississippi River basin. Small streams merge to form larger streams, large
streams merge to form rivers, and so forth. A typical example of a drainage
network is given in Figure 8.1, from the Volfe and Bell Canyons in the San
Gabriel Mountains near Glendora, California. We will show that drainage
networks are classic examples of fractal trees (Tarboton et al., 1988; Beer
and Borgas, 1993; Garcia-Ruiz and Otalora, 1992).
Long before the concept of either fractals or fractal trees was introduced,
a quantitative stream-ordering system was introduced by Horton (1945) and
Strahler (1957). The Strahler (1957) ordering system is illustrated in Figure
8.2. Streams on a standard topographic map with no upstream tributaries are
defined to be first-order (i = 1) streams. When two first-order streams com-
bine they form a second-order (i = 2) stream. When two second-order
streams combine they form a third-order (i = 3) stream, and so forth. Note
that first-order (i = 1) streams can also join second-order (i = 2), third-order
184 GEOMORPHOLOGY
Figure 8.1. The drainage

network in the Volfe and Bell
Canyons, San Gabriel
Mountains, near Glendora,
California, obtained from
field mapping (Maxwell,
1960).

Strahler ( 1957) stream-
ordering system.
GEOMORPHOLOGY 185
(i = 3), and all other higher-order streams. Similarly second-order (i = 2)

streams can join third-order (i = 3) streams and so forth. We will refer to
these as side tributaries.
Horton (1945) defined the bifurcation ratio Rb according to
where N represents the number of streams of order i. It is the ratio of the

number of streams of one order to the number of streams of the next higher
order. For example, if there are 20 third-order streams and 5 fourth-order
streams we have Rb = ?$! = 4. Horton also introduced the length-order ratio
defined by
where ri represents the mean length of streams of order i. Empirically, both

R, and Rr are found to be nearly constant for a range of stream orders in any
given drainage basin. These are known as Horton's laws. Taking the defini-
tion of the fractal dimension D to be
the substitution of (8.1) and (8.2) gives the fractal dimension of a drainage
network as
Standard stream-ordering parameters are directly related to the fractal di-

mension of the network. The validity of Horton's laws implies that drainage
networks are fractal trees.
The dependence of the number of streams of various orders on their mean
length for the drainage network illustrated in Figure 8.1 is given in Figure 8.3.
The smallest streams on this map based on field studies are one order lower
than the smallest streams on the topographic map of this region; we refer to
these streams as 0-order streams. The highest order streams in this region are
fourth order. It is seen that the results correlate well with (8.3) taking D =
1.81. The equivalent number-length statistics for the entire United States are
given in Figure 8.4. The single tenth order river is the Mississippi. Again a
good correlation with the fractal relation (8.3) is obtained with D = 1.83.
186 GEOMORPHOLOGY

the number of streams N of
various orders 0-4 on their
mean length r for the
example illustrated in Figure
8.1. The power-law, straight-
line correlation is with (8.3)
takingD= 1.81.

the number N of rivers with a
specified order on their mean
length r. The correlation with
(8.3) gives D = 1.83.
GEOMORPHOLOGY 187
It is seen that stream networks are, to a good approximation, fractal and

have fractal dimensions near 1.80, clearly somewhat less than the space fill-
ing D = 2. This is consistent with the example illustrated in Figure 8.1. Over
short distances, fluxes of water act diffusively, flowing over the surface or
through the near-surface without incising channels (streams).
8.2 Fractal trees
We now turn our attention to deterministic fractal trees. Three examples are
given in Figure 8.5. To specify the geometry of a deterministic fractal tree,
three quantities must be given: the bifurcation ratio R,, the length-order ra-
tio R , and the angle of divergence 0. And these three quantities are indepen-
dent of order. For the example given in Figure 8.5(a), R, = 3, Rr = 3 , O = 30";
and from (8.4) D = 1 for this fractal tree. For the example given in Figure
8.5(b), R, = 2, Rr = 2, 0 = 60, and again D = 1. And for the example in Fig-
ure 8.5(c), R, = 2, Rr = fi,0 = 90, and D = 2. In all cases the constructions
can be extended to infinite order without overlap. If the construction in Fig-
ure 8.5(c) is extended to infinite order, the plane is entirely covered by the
construction but with no overlap. Thus, this construction is an example of a
self-similar (identical at all scales), deterministic network that can drain
every point on a surface at as small a scale as is specified. This is the impli-
cation of D = 2, the dimension of a plane.
Comparing the drainage network in Figure 8.1 with the fractal trees il-
lustrated in Figure 8.5 shows an important discrepancy. The drainage net-
work has side tributaries whereas the fractal trees do not. First-order streams
intersect other first-order streams to form second-order streams. But other
first-order streams intersect second-order, third-order, and all higher-order
streams. Similarly second-order streams intersect other second-order
streams to form third-order streams. But other second-order streams inter-
sect third-order, fourth-order, and all higher-order streams.
Figure 8.5. Three examples

of fractal trees. (a) R, =
N,INi+I= 3, R,= ri+llri= 3.
8=30,D= l.(b)R,=
N,INitI =2,R,=ri+l/ri=2,
8=60, D = l . ( c ) R , =
NiIN,,, = 2, Rr= ri+,lri=
fi,0 = 90, D = 2.
188 GEOMORPHOLOGY
To account for side tributaries Tokunaga (1978, 1984) introduced the

family of fractal trees illustrated in Figure 8.6. His approach has been con-
sidered in detail by Peckham (1995). In Figure 8.6(a) we have a standard bi-
nary fractal tree with R, = 2, Rr = 2, and 0 = 45". From (8.4) we have D = 1.
There are no side branches on this tree. In Figure 8.6(b) a first-order branch
has been added to each second-order branch, a second-order branch to each
third-order branch, and a third-order branch to the fourth-order branch. We
have N , = 27, N, = 9, N , = 3, N4 = 1 so that R, = 3 and D = In 3fln 2 = 1.585.
This is clearly also a self-similar fractal tree. In Figure 10.6(c) further side
branches have been added. l h o first-order branches have been added to each
third-order branch and two second-order and four first-order branches have
been added to the fourth order branch. We now have N , = 43, N2 = 11, N, = 3,
N4 = 1. This tree is not a deterministic fractal tree but becomes statistically
self-similar in the limit of infinite order; we will show that in this limit R, +
4 and D + 2.
To quantify this more general class of fractal trees, a first-order branch
joining a first-order branch is denoted "1 1" and the number of such branches
is N , , , a first-order branch joining a second-order branch is denoted "12"and
the number of such branches is N,,, a second-order branch joining a second-
order branch is denoted "22" and the number of such branches is N2,, and so
forth. This classification of branches is illustrated in Figure 8.6.
Figure 8.6. (a) Binary, self- The branch numbers Nu for a fractal tree of order n constitute a square,
similar fractal tree with N , = upper-triangular matrix. The branch-number matrices for the three fractal
8 ( N , , = 8 , N I 2= N I 3= trees illustrated in Figure 8.6 are given in Figure 8.7. The total number of
N,, = O), N2 = 4 (N22= 4 , streams of order i, N is related to the Ni by
N23= N2, = 0 ) . N3 = 2 (N33=
2 , N3, = O), and N, = 1 (N,, =
1). Thus we have R, = R, = 2
and D = 1. (b) Self-similar
fractal tree with side
branches. We have N , = 27
for a fractal tree of order n.
( N , , = 18, N12= 9, N13=
N,, = 0 ) . N2 = 9 (Nz2= 6 ,
NZ3= 3, N2, = 0 ) . N3 = 3
(N,, = 2 , N3, = l ) , and N, = 1
( N , = I). This gives R, = 3,
R, = 2, and D = In 3lln 2 =
1 S 8 5 . (c) Addition of more
side branches. We have N , =
43 ( N , , = 22, N12= 1 1 , N I 3=
6 , N l , = 4 ) , N 2 = 11 (N2,=6,
N23= 3, N2, = 2)' N 3 = 3
(N,, = 2 , N,, = 1) and N, = 1
( N , = 1). This fractal tree
becomes statistically self-
similar in the limit of
infinite order with R, 4 4
and D + 2.
GEOMORPHOLOGY 189
This class of fractal trees can also be quantified in terms of branching ra-
tios To. These are the average numbers of branches of order i joining
branches of order j. Branching ratios are related to branch numbers by
If the primary branching is binary, (8.5) and (8.6) can be combined to give
Again the branching ratios Ti constitute a square, upper-triangular matrix.

The branching-ratio matrices for the three fractal trees illustrated in Figure
8.6 are given in Figure 8.8.
We now define self-similar trees to be the subset of trees for which Ti,i + k =
Tk where Tkis a branching ratio that depends on k but not on i. From Figure
8.8 we see that the three fractal trees illustrated in Figure 8.6 have (a) Tl = T2=
T3 = 0, (b) T , = 1, T2 = T3 = 0, and (c) Tl = 1, T2 = 2, T, = 4 . Tokunaga(1978,
1984) introduced a more restricted class of self-similar fractal trees by re-
quiring
This is now a two-parameter family of fractal trees. For the fractal tree illus-
trated in Figure 8.6(b) we have a = 1, c = 0 and for the fractal tree illustrated
in Figure 8.6(c) we have a = 1 and c = 2. Substitution of (8.8) into (8.7) gives
n-i n -i
If we divide (8.9) by N and introduce the branching ratios from (8.2) we obtain
n-i Ck-l
2
1=-+ax--
Rb k=l Rkb
8 0 0 0 189 0 0 22116 4 N1l N12 N13 N14 Figure 8.7. The

4 0 0 6 3 0 6 3 2 branch-number matrices for
N22 N23 N24 the fractal trees illustrated
2 0 2 1 2 1 N33 N34 in Figure 8.6(a), (b), (c) are
1 1 1 N44 given in (a), (b), and (c).
The nomenclature is
(a) (b) (c) (dl illustrated in (d).
190 GEOMORPHOLOGY
which can be written as
And if n - i - 1 is large we can approximate (8.11) with
which becomes the quadratic
with the solution
Thus large Tokunaga trees have branching ratios that are independent of
order. For the tree illustrated in Figure 8.6(c) with a = 1 and c = 2 we have
Rb + 4 as n + =. And when the length-order ratio Rr is specified the fractal
dimension is given by
For the tree illustrated in Figure 8.6(c) we have Rr = 2 and D = 2.

The Tokunaga model can be further constrained if the number of side
branches is assumed to be proportional to the length of the branch they enter.
For this case a = Rr- 1 and c = R , thus the Tokunaga branching ratio is given by
Figure 8.8. The branching- Tk = ( R , - 1)R y (8.16)

ratio matrices for the fractal
trees illustrated in Figure The fractal tree illustrated in Figure 8.6(c) is a deterministic example of this
10.6(a), (b), (c) are given in class of tree with Rr = 2, a = 1, and c = 2. Substitution of (8.16) into (8.14)
(a), (b), and (c). The
nomenclature of the
gives
branching ratios is given in
(d). In (a) we have TI = T, =
T3 = 0, in (b) we have TI = 1,
T2=T3=Oora=l , c = O
from (8.8), and in (c) we
have T, = 1, T, = 2, T3 = 4,
or a = 1, c = 2 from (8.8).
GEOMORPHOLOGY 191
Substitution of (8.16) into (8.15) gives
For this subclass of Tokunaga fractal trees the branching ratio R, and the
fractal dimension D can be obtained from (8.17) and (8.18) if the length-or-
der ratio Rr is specified. A related quantification of side branching has been
given by Vannimenus and Viennot (1989) and Ossadnik (1992).
We now address the question whether the statistics of actual drainage
networks are represented by Tokunaga fractal trees. Peckham (1995) has de-
termined branching-ratio matrices for the Kentucky River basin in Kentucky
and the Powder River basin in Wyoming. Both are eighth-order basins with
the Kentucky River basin having an area of 13,500 km2 and the Powder
River basin an area of 20,18 1 km2. For the Kentucky River basin the bifurca-
tion ratio is R, = 4.6 and the length-order ratio is Rr = 2.5; for the Powder
River basin the bifurcation ratio is R, = 4.7 and the length-order ratio is Rr =
2.4. The dependence of the number of streams of various orders on their
mean length for the two basins are given in Figure 8.9. Again the results cor-
relate well with the fractal relation (8.4) taking D = 1.85. The branching-ra-
tio matrices for the two river basins are given in Figure 8.10. We now deter-
mine values for Tkby averaging the values of over i
lo5 -
lo4 -
0 Kentucky River
x Powder River
lo3 -
N
lo2
-
10 - mean length r for the
Kentucky River basin in
8 Kentucky and the Powder
1 v n
V
I River basin in Wyoming.
1 6' 1 10 lo2 lo3 The straight-line correlation
r is with (8.4) taking D = 1.85.
192 GEOMORPHOLOGY
1 n-k
Tk = Ti,i +k
(n - k ) i = l
For example, we find that T, = 18 for the Kentucky River basin by taking the
average of T,, = 15.6, T,, = 20.3, T,, = 16.0, and T,, = 20.0. The values of T,
for the two basins are given in Figure 8.11 as a function of k. It is seen that
Figure 8.10. Branching-ratio

matrices for (a) the Kentucky
River basin and (b) the
Powder River basin as
obtained by Peckham (1995).
0 Kentucky River
x Powder River
DLA synthetic network
Figure 8.1 1. Dependence of

the mean branching ratios T,
on k for the Kentucky River
basin and the Powder River
basin. The straight-line
correlation is with (8.8)
taking a = 1.2 and c = 2.5.
Also included are the mean
branching ratios for a fifth-
order diffusion limited
aggregation (DLA) network.
GEOMORPHOLOGY 193
the results correlate well with (8.8) taking a = 1.2 and c = 2.5. These results
are also tabulated in Table 8.1. At least for these two basins, in quite different
geological settings, good agreement with Tokunaga fractal trees is obtained
with these values of the parameters a and c. It is also of interest to compare
these values with those given in (8.16). With Rr = 2.5 for the Kentucky River
basin and Rr = 2.4 for the Powder River basin, the values from (8.16) are a =
1.5, c = 2.5 and a = 1.4, c = 2.4 respectively. These are in quite good agree-
ment with the best-fit values of a = 1.2 and c = 2.5.
Empirically, actual drainage basins appear to be well approximated by
Tokunaga fractal trees. There are a number of other applications of fractal
trees in geology and geophysics. River deltas are one obvious choice. An-
other is the upward migration of magma beneath a volcano. Partial melting
in the earth's mantle occurs on grain boundaries. Because the melt, the
magma, is lighter than the residual solid, it drains upward eventually reach-
ing the earth's surface, resulting in a volcanic eruption. One approach to the
magma ascent problem is to treat it as a flow in a uniform porous media
(Turcotte and Schubert, 1982, pp. 413-416). An alternative is to treat the
magma paths like a drainage network. Rivelets of magma combine to form
ascending magma streams, and ascending streams of magma combine to
form magma rivers. Hart (1993) has proposed a fractal tree model for the as-
cending magma and has considered its implications on magma composition.
The concepts of fractal trees also have a wide variety of other applica-
tions. Examples include the growth of actual trees and other plants, as well as
the cardiovascular distribution of veins and arteries and the bronchial system.
Returning to drainage networks, another fractal correlation to drainage
patterns is obtained if the length of the principal river in a drainage basin P is
plotted against the area of the basin A. Data for several basins in the north-
eastern United States are given in Figure 8.12 (Hack, 1957). The applicable
fractal relation is
Table 8.1 .Branching statistics for a variety of fractal trees
Order R, Rr D a c
Idealized tree Figure 8.6(a) - 2 2 1 0 0

Idealized tree Figure 8.6(c) 00 4 2 2 1 2
Kentucky River basin 8 4.6 2.5 1.67 1.2 2.5
(Peckham, 1995)
Powder River basin 8 4.7 2.4 1.77 1.2 2.5
(Peckham, 1995)
DLAa cluster (Ossadnick, 1992) 10 5.15 2.87 1.56 1.5 2.70
DLA drainage network model 7 3.98 2.09 1.87 1.2 2.5
"DLA, diffusion limited aggregation.

194 GEOMORPHOLOGY
and good agreement with the data in Figure 8.12 is obtained taking D = 1.22.
Robert and Roy (1990) have discussed this fractal relation between main-
stream length and drainage area.
It is clear that drainage networks are self-similar and fractal to a good
approximation. But this is basically an empirical statement that does not ad-
dress how drainage networks evolve. It is clear that in most terrains drainage
networks are a direct consequence of erosion. In young terrains tectonic
processes play an important role; however, erosional processes may still be
dominant. Consider the Hawaiian chain of volcanic islands. A young island
such as Hawaii is made up of deterministic conical structures associated
with shield volcanoes. These are not fractal. However, sufficient erosion has
occurred on Maui and Oahu in a few million years to develop an irregular,
scale-invariant morphology that exhibits fractal statistics. The erosional evo-
lution of landscapes is a problem that has fascinated natural scientists for
centuries. The forms of mature landscapes evolve through processes of ero-
sion and deposition. An essential question is whether it is possible to de-
velop a basic theory of landscapes or whether it is necessary to consider only
statistical aspects of the problem.
A variety of models were proposed in the 1960s to describe the statistics
and origins of drainage networks (Smart, 1972). Descriptive models were in-
troduced by Shreve (1966, 1967) and Schreidegger (1967) in which drainage
networks were considered as infinite topologically random networks (i.e., no

the length P of the principle
river on the area A of the
drainage basin for several
drainage basins in the
northeastern United States
(Hack, 1957).
GEOMORPHOLOGY 195
one distribution of network links is preferred over any other). They showed
that the statistics of real drainage networks matched the most probable num-
ber-order distribution of a topologically random network. Snow (1989) has
shown that the sinuosity of streams exhibits fractal behavior, and Nagatani
(1993) has shown that meander patterns are fractal. Although these models
have proven useful as a way to describe drainage networks, they contain lit-
tle information on the dynamical processes that form them.
Other workers have proposed random growth models to explain the
planform organization of drainage networks. Leopold and Langbein (1962)
and Schenck (1963) proposed models in which the streams themselves fol-
lowed random walks. Thus the network was not headward growing, but
propagated laterally from the most central "trunk" stream. In addition, the
network grew by the addition of entire stream segments, rather than by grad-
ual expansion (accretion). Howard (1971) introduced an accretionary head-
ward growth model, a site adjacent to the existing network was chosen ran-
domly, and the network propagated to this site. Thus, all sites on the network
had an equal probability for growth.
8.3 Growth models
A growth model that has been applied to stream networks as well as a wide
variety of other applications is diffusion limited aggregation (DLA). How-
ever, before considering DLA we illustrate a deterministic fractal growth
model based on the Koch snowflake. This model is illustrated in Figure 8.13.
An initial unit square at zero order [(Figure 8.13(a)] grows at first order by

deterministic fractal growth
(accretion) model based on
the Koch snowflake. (a) A
single unit-square seed
particle at zero order. (b) At
first-order, four unit-square
particles are added to the
corners of the seed particle.
This is the generator for the
higher-order fractal
constructions. (c) At second-
order, four of the first-order
structures are added at the
comers of the first order
structure. (d) At third-order,
four of the second-order
structures are added at the
comers of the second-order
structure.
196 GEOMORPHOLOGY
the addition of four unit squares at the four corners of the original "seed"
particle, as illustrated in Figure 8.13(b). At second order, four of the first-
order structures are added as shown in Figure 8.13(c). At third order, four of
the second-order structures are added as shown in Figure 8.13(d). We have
No= 1, r o = 1; N, = 5 , r, = 3; N2=25, r 2 = 9 ;N3= 125, r,=27 and from (2.2)
we have D = In 5Iln 3 = 1.465. One approach to quantifying the growth of an
aggregate such as that illustrated in Figure 8.13 is to determine the number
of particles as a function of size. At zero order the number of particles is No =
1 and a circle with radius ro = l l f i coven the particle, at first order the
number of particles is N, = 5 and a circle with radius r, = 3 1 f i coven the
particles, at second order the number of particles is N2 = 25 and a circle with
radius r, = 9 1 f i covers the particles, and at third order the number of parti-
cles is N3 = 125 and a circle with radius r, = 2 7 1 f i . Noting that in this case
-
N rD we again find D = In 51ln 3 = 1.465 just as above.
However, in applications to statistical growth models and to natural phe-
nomena, it is generally preferable to use the "radius of gyration" rather than
the radius of a circle (sphere) that covers the growing aggregate. The defini-
tion of the radius of gyration for an aggregation of N particles growing from
a seed particle in two dimensions is
where ri is the radial distance of particle i from the seed particle. A fractal re-
lation is defined by
where a is the particle size. For the example given in Figure 8.13 we find
that at first order we have the centers of the four accreted particles at a dis-
tance fifrom the seed particle. Thus from (8.21) we have
Similarly at second order we have
And from (8.22) the corresponding fractal dimension is given by

GEOMORPHOLOGY 197
slightly less than the value D = 1.46 obtained for the basic construction. For
the third-order construction illustrated in Figure 8.13(d) we obtain rg4/a=
12.06 and from (8.22)
indicating a fractal dependence.
8.4 Diffusion-limited aggregation (DLA)
The concept of diffusion-limited aggregation was introduced by Witten and

Sander (1981). They considered a two-dimensional grid of cells and placed a
seed cell near the center of the grid. An accreting cell was randomly intro-
duced on a "launching circle" as illustrated in Figure 8.14. The accreting cell
was allowed to follow a random path, that is, it was a random walker, until
(1) it accreted to the growing cluster of cells by entering a cell adjacent to
the cluster or (2) it wandered across the "killing circle" in which case a new
cell (random walker) was introduced on the "launching circle." An illustra-
tion of a cluster created by DLA is given in Figure 8.15. The resulting sparse
dendritic structure results because the walkers are more likely to accrete near
the tips of the cluster rather than in the deep interior. Halsey and Leibig
"Killing Circle"
n Random Path
"Launching Circle"
Figure 8.14. illustration of

diffusion-limited aggregation
(DLA) growth of a cluster.
An accreting cell is
introduced at a random point
on the "launching circle."
The cell carries out a random
walk until it either accretes
to the growing cluster (as
shown) or crosses the
"killing circle," in which
case the walk is terminated.
GEOMORPHOLOGY 199
sulting fractal structure often resembles DLA clusters (Chen and Wilkinson,
1985; Malay et al., 1985; Nitmann et al., 1985, 1986; Van Damme et al.,
1986; Feder and Jossang, 1995). Sornette et al. (1990) have suggested that
the fractal distributions of faults and joints discussed in Chapter 4 are the re-
sult of a DLA random growth. Two-dimensional surface exposures of frac-
tures and joints are generally fractal with D = 1.7 in agreement with the val-
ues obtained for DLA clusters. DLA models for crack propagation have also
been used to model fragmentation, and power-law number-size statistics for
fragments were obtained (Gomes and Sales, 1993).
8.5 Models for drainage networks
A modified version of the diffusion-limited aggregation (DLA) technique

has been used to generate realistic drainage networks (Masek and Turcotte,
1993). To apply DLA to drainage networks several modifications are neces-
sary. Random walkers are introduced randomly over the entire grid and are
allowed to walk until they either intersect the evolving network or are lost
from the grid. The random walkers can be viewed as unit water fluxes (rain-
fall and overland flow) that migrate over a relatively flat surface until they
find a gully (network) in which to flow. When the flux joins the gully, it
erodes and expands the network.
The growth of a synthetic drainage network based on the DLA model is
illushated in Figure 8.16. A square grid of 15 X 15 cells is used in this illus-
hation. Five seed cells are introduced at random points on thc lower bound-
ary. The evolving network must grow from these seed cells. From the exam-

Seed Cells the mechanism for network
growth in the DLA model for
drainage network evolution.
Accreted Cells
I I I
1 1 1 1
I
1 1 1 I I 1 I I I I A random walker is
-
I I IWI IU-hI I I I I I randomly introduced to an
unoccupied cell. The random
Newly Added Cell walk proceeds until a cell is
encountered with one (and
only one) of the four nearest
Random Walk neighbors occupied
(hatched). The new cell is
+ Prohibited Sltes accreted to the drainage
network. If a random walker
enters a prohibited cell or
0 Other allowed sites wanders off the grid it is
for accretion terminated.
200 GEOMORPHOLOGY
ple shown, 16 cells have been accreted to the seed cells. Cells are allowed to
accrete if one (and only one) of the four nearest neighbor cells is part of the
preexisting network. Prohibited sites that already have two neighboring sites
occupied are identified by stars. Sites available for accretion to the network
are indicated by open circles. A random walker is introduced at a random
cell on the grid and the resulting random-walk path is traced by the solid
line. After 28 random walks it accretes to the network at the cross-hatched
cell. A random walk proceeds until the walker (1) accretes to the network,
(2) exits the grid, or (3) lands on a prohibited cell. In cases (2) and (3) the
walk is terminated and a new walker is introduced on a new, randomly se-
lected site. The iteration of this basic procedure results in a branching net-
work composed of linked drainage cells. This "self-avoiding" algorithm pre-
vents local clumping of drainage cells.
Although the model is highly schematic, the mechanics outlined here are
analogous to the mechanics operating in real drainage systems. The accre-
tionary nature of network growth produces a headward evolving drainage

synthetic drainage network
generated using a diffusion
limited aggregation (DLA)
model on a 256 X 256 grid
using 20,000random
walkers.
GEOMORPHOLOGY 201
pattern similar to patterns of headward erosion seen in nature. The accretion

process itself is analogous to a flux of water (e.g., overland flow) intersect-
ing an existing drainage, thereby initiating a new first-order channel. The
self-avoiding algorithm prevents drainages from becoming locally space
filling at the finest scales. In nature, this limit may be controlled by a thresh-
old transition from diffusive slope processes to advective channelization
processes.
An illustration of a DLA drainage network simulation grown on a 256 X
256 grid of cells is illustrated in Figure 8.17. One-third of the bottom row of
cells were tagged as seed cells and 20,000 random walkers were introduced.
There is considerable visual similarity between this simulated network and
the actual drainage network illustrated in Figure 8.1. The dependence of the
number of streams of various orders in the simulation on their mean length is
given in Figure 8.18. The highest-order streams in the simulation are seventh
order. It is seen that the results correlate well with (8.3) taking D = 1.85. This
result is in good agreement with actual drainage networks. The branch-num-
ber matrix and the branching-ratio matrix for a fifth order network obtained
using the DLA model are given in Figure 8.19. Using (8.19) the correspond-

mean length r for the DLA
simulation given in Figure
8.17. The correlation is with
(8.3) taking D = 1.85.
Figure 8.19. (a) Branch-

number matrix and (b)
branching-ratio matrix for a
fifth-order synthetic drainage
network generated using the
DLA model illustrated in
Figure 8.17.
202 GEOMORPHOLOGY
ing values of mean-branching ratios T, are determined and are compared

with those for actual drainage networks in Figure 8.1 1. Good agreement be-
tween the synthetic DLA network and the actual drainage networks is found.
These results are also tabulate in Table 8.1.
8.6 Models for erosion and deposition
To totally model landforms it is also necessary to include erosion with the

evolution of drainage networks.
A number of theories have been proposed for erosion and have been
summarized by Scheidegger (1991).The simplest theory hypothesizes that
erosion is proportional to the elevation h so that
where T is a characteristic time for erosion. With the initial condition that
h = h, at t = 0 the solution is
This gives an exponential decay of topography with a characteristic time T

but no information on the form of the evolving landscapes.
Culling (1960, 1963, 1965) hypothesized that the horizontal flux of
eroded material m, is proportional to the slope:
where J is the transport coefficient and we consider only the one-dimen-

sional problem h = h(x,t).With the conservation of mass relation
this gives
which is the linear diffusion equation.

GEOMORPHOLOGY 203
Solutions to the diffusion equation take a variety of forms but in many

cases reduce to the error function. The form of many alluvial fans and pro-
grading deltas can be approximated quite accurately with the error function
(Kenyon and Turcotte, 1985). In addition, a number of authors have used the
Culling model to estimate the age of faults and shoreline scarps (Wallace,
1977; Buckman and Anderson, 1979; Nash, 1980a, b; Mayer, 1984; Hanks
et al., 1984; Hanks and Wallace, 1985; Andrews and Hanks, 1985). Typical
values for J are in the range 10-2-10-3 m2 yr-I. However, solutions to a lin-
ear equation such as the diffusion equation cannot produce fractal, self-
similar solutions.
A variety of stochastic models have been proposed for deposition. The
simplest model of random deposition was presented by Schwarzacher (1976).
In his model the rate of deposition on a two-dimensional cross-section of the
landscape is Gaussian white noise. Once sediment is deposited, it cannot be
transported along the surface. Although this model produces a rough surface,
the model does not produce self-similar (self-affine) topography.
A two-dimensional model may be considered to be an idealized model of
deposition in three dimensions with channels that are parallel and of uniform
capacity in the direction of flow. Because the depositional channels that
evolve at the surface are uniform in the flow direction, the function repre-
senting the surface will be uniform in that direction. As such, we can de-
scribe the evolution of the two-dimensional surface by the one-dimensional
cross section. We will assume that the surface of the landscape has a constant
mean growth rate. With this assumption, the variations in the surface in time
and depth are equivalent.
Pelletier and Turcotte (1997) proposed an extension of Schwarzacher's
model. A site on a one-dimensional lattice is chosen at random. A unit of sed-
iment is deposited at that site or at one of its nearest neighbors depending on
which site has the lowest elevation. This model is illustrated in Figure 8.20.
The local elevation is the total number of units of sediment that have been
deposited at the site. This model of surface growth was first analyzed by
Family (1986) with applications to the growth of atomic surface layers. He
reported the results of computer simulations that showed that the model pro-
duces scale-invariant variations of the surface in space and time. He found
that the standard deviation, a , of the surface obeys the relation
where L is a length scale and T is a time scale. Surfaces with scale-invariant

standard deviations a(L,T) a LHa TK have a power-law dependence of the
power spectral density, S(k), on wave number k of the form S(k) a k-2Ha-I
(i.e., a k-2 for Ha = $,) and a power-law dependence on frequency of the
i).
form S(f) ~ l f - 2 ~ - 1 (i.e., cx f-3i2 for K =
204 GEOMORPHOLOGV
Equilibrium topography produced by this model with 1024 lattice sites

is illustrated in Figure 8.21. The resulting dependence of the spectral power
density on wave number is given in Figure 8.22. The data points are the av-
erage of 50 simulations. The correlation with the self-affine fractal relation
(7.41) is excellent taking fl = 2.01, and the topography is basically a Brown-
ian walk.
Variations in the rate of growth at a specified lattice point are given in
Figure 8.23. Deviations of the elevation from the mean elevation are given
as a function of time. The resulting dependence of the spectral power density
on frequency is given in Figure 8.24. The data points are again the average
of 50 simulations. The correlation with the self-affine fractal relation (7.41)
is excellent taking fl = 1.5 1. This value is in reasonable agreement with the
values given for the porosity logs in Figure 7.10.
Although the analysis given above is for deposition, it can be easily
modified to model erosion. Instead of randomly adding units, units are ran-
domly removed. A random unit is eroded either at the chosen site or on one
of its nearest neighbors depending on which site has the highest elevation.

the sediment deposition
model. The dotted block is
the unit of sediment being
added randomly to the
surface. The arrows point
toward the site upon which
the unit of sediment will be
deposited. In (a) the
randomly chosen site has a b)
lower elevation than either of
its nearest neighbors, so the
sediment is deposited at the
chosen site. In (b) one of the
nearest neighboring sites has
a lower elevation and the
sediment is deposited at that
lower site. If both adjacent
sites have lower but equal
elevations, one of the two
sites is chosen randomly, as
illustrated in (c).
GEOMORPHOLOGY 205
Figure 8.21. Depositional

topography obtained using
the model illustrated in
256.0 51 2.0 768.0 1024.0 Figure 8.20 with 1024 lattice
lattice site sites.

the spectral power density S
on wave number k for the
synthetic depositional
topography illustrated in
Figure 8.21. The data points
are the average of 50
simulations. An excellent
I
t
correlation with the self-
-2.8 -2.4 -2.0 -1.6 - 1.2 -0.8 -0.4 affine fractal relation (7.41)
log k is obtained taking P = 2.01.
206 GEOMORPHOLOGY
Figure 8.23. Deviations from

the mean rate of growth as a
function of time at a chosen 12.0 15.0 18.0
site. timesteps ( ~ 1 0 0 0 0 0 )

the spectral power density S
on frequency f for the time
dependence of the deposition
illustrated in Figure 8.23.
Again the data points are the
average of 50 simulations.
An excellent correlation with
the self-affine fractal relation I
(7.41) is obtained taking -6.3 -6.0 -5.7 -5.4 -5.1 -4.8 -4.5 -4.2
p = 1.51. log f
GEOMORPHOLOGY 207
Family (1 986) showed that a continuous version of this discrete model is

provided by the diffusion equation with a Gaussian white noise term
where q ( X J ) is the Gaussian white noise. This is one form of the linear
Langevin equation and is also known as the Edward-Wilkinson equation.
Comparing (8.29) and (8.27) shows that the Culling model gives Brownian-
walk topography if erosion or deposition occurs randomly. An extensive
discussion of growth processes has been given by Barabasi and Stanley
(1995).
A wide variety of models have been proposed to simulate scale-invariant
(fractal) topography and/or river networks. Chase (1992) has combined a
cellular-automata advection model with diffusion and has generated reason-
ably realistic topography with drainage networks. Similar models have been
given by Takayasu and Inaoki (1992) and Lifton and Chase (1992). Leheny
and Nagel (1993) introduced an avalanche model and derived both topogra-
phy and river networks. Meakin et al. (1991) have applied a DLA approach,
and Willgoose et al. (1991) an advective-diffusive model. Kramer and
Marder (1992) have modeled the development of drainage networks on a
water-covered landscape assuming that the erosion is proportional to the
product of velocity and pressure. Barzini and Ball (1993) use a similar
model to develop synthetic braided rivers. They point out that both the simu-
lations and real braided rivers have a fractal distribution of island sizes with
D = 1.
Stark (1991) used an invasion percolation technique, in which the grow-
ing network was superposed on a fixed random field (analogous to a sub-
strate with variable erodibility). At each time step, the network propagated to
the adjacent site having the highest erodibility value over the entire perime-
ter. Although all sites on the network had differing probabilities for growth,
these probabilities did not change through time since the random field was
fixed from the start.
Stark (1994) modeled patterns of erosion using invasion percolation,
Eden growth, and DLA models. Liu (1992) has utilized percolation clusters.
Takayasu (1993) used random self-affme tiling to explain the power-law dis-
tribution of drainage basin sizes. Nikora and Sapozhnikov (1993) have pre-
sented a model based on random walk simulations. Minimum energy
dissipation and/or entropy methods have been applied to landforms and river
networks by Rinaldo et al. (1993) and by Sun et al. (1993,1994,1995). Sor-
nette et al. (1994) have presented a model for the tectonic generation of frac-
tal topography utilizing statistical distributions of displacements on fault
arrays.
208 GEOMORPHOLOGY
It is also desirable that models of topography exhibit the Brownian-walk

spectrum discussed in the previous chapter (Newman and Turcotte, 1990). It
has been shown that the three-dimensional nonlinear Langevin equation can
yield topography with the observed Brownian-walk spectrum (Sornette and
Zhang, 1993).
8.7 Floods
The volumetric flow Q(t) in a river constitutes a time series. However, the
flow is strongly asymmetric so that Gaussian statistics are generally a poor
approximation. Large values of Q for relatively short periods of time consti-
tute floods, whereas low values of Q for relatively long periods constitute
droughts. Most rivers also have a strong annual periodic component.
Floods present a severe natural hazard; to assess the hazard and to allo-
cate resources for its mitigation it is necessary to make flood-frequency haz-
ard assessments. The integral of the flow in a river is required for the design
of reservoirs and to assess available water supplies .during periods of
drought. An important question in geomorphology concerns which floods
dominate erosion. Is erosion dominated by the 10 year, the 100 year, or the
very largest floods? The answer to this question depends upon whether ex-
treme flood probabilities have an exponential or power-law dependence on
time. One estimate of the severity of a flood is the peak discharge at a station
Q,. The magnitude of the peak discharge is affected by a variety of circum-
stances including (1) the amount of rainfall produced by the storm or storms
in question, (2) the upstream drainage area, (3) the saturation of the soil in
the drainage area, (4) the topography, soil type, and vegetation in the
drainage area, and (5) whether snow melt is involved. In addition dams,
stream channelization, and other man-made modifications can affect the
severity of floods.
To estimate the severity of future floods, historical records are used to
provide flood-frequency estimates. Unfortunately, this record generally cov-
ers a relatively short time span and no general basis has been accepted for its
extrapolation. Quantitative estimates of peak discharges associated with pa-
leofloods are generally not sufficiently accurate to be of much value. A wide
variety of geostatistical distributions have been applied to flood-frequency
forecasts, often with quite divergent predictions. Examples of distributions
used include power law (fractal), log normal, gamma, Gumbel, log Gumbel,
Hazen, and log Pearson.
It is standard practice to use the annual peak discharges in flood-fre-
quency analyses. In the United States this is the peak discharge during a wa-
ter year, which extends from October 1 of the preceding year to September
30. There are serious problems with this approach and alternatives will be
GEOMORPHOLOGY 209
discussed. However, the basic question we wish to consider is whether

floods satisfy power-law (fractal) scaling (Turcotte and Greene, 1993; Tur-
cotte, 1994c; Wu and Lye, 1994).
If floods are fractal, then it would be expected that the peak discharge
Q, in a time interval T is related to the interval by
where the Hausdorff measure H a plays a role similar to that in (7.29). The
Hausdorff measure is related to the fractal dimension by (7.33). Since river
discharges have a strong annual variability, the interval T is generally taken
as an integer number of years when floods are considered. This scale-invari-
ant distribution can also be expressed in terms of a flood-frequency factor F,
which is the ratio of the peak discharge over a 10-year period to the peak dis-
charge over a 1-year period. With self-similarity the flood-frequency factor
F is also the ratio of the 100-year peak discharge to the 10-year peak dis-
charge and the ratio of the 1000-year peak discharge to the 100-year peak
discharge. In terms of H a and D we have
The flood-frequency factor is a measure of the severity of great floods.

We now turn to the analysis of flood-frequency records. As our first ex-
ample, the ten benchmark stations considered by Benson (1968) will be
studied. Benson (1968) applied a variety of geostatistical distributions to the
data from these stations; these will be compared with the fractal approach
discussed above. The maximum annual floods for two stations are given in
Figure 8.25. Values for station 1- 1805 on the Middle Branch of the Westfield
River in Goss Heights, Massachusetts, are given in Figure 8.25(a) for the pe-
riod 1911-1960 (Green, 1964) and values for station l 1-0980 in the Arroyo
Seco near Pasadena, California, are given in Figure 8.25(b) for the period
1914-1965 (Young and Cruff, 1967).
Carrying out statistical studies on an annual flood series systematically
underestimates the number of smaller floods. The reason is that the second
or third largest flood in some water years is larger than the largest flood in
other water years. The alternative is to consider a partial-duration series,
which may include several floods in a given water year. However, the defin-
ition of two statistically independent floods is arbitrary. We assume that two
floods are statistically independent if they are separated by more than two
months.
The largest floods for each record are ordered: The largest flood is as-
signed a period equal to the length of the record T,,, the second largest flood a
period td2, the third largest flood a period 743, and so forth. The log of the
210 GEOMORPHOLOGY
peak discharge for each flood is plotted against the log of its assigned period.
This is the same technique that was used for earthquakes in Chapter 4. Re-
sults for station 1-1805 (Goss Heights, MA) are given in Figure 8.26(a). The
solid line is the least square fit of (8.30) with the data over the range 50 <
Q, < 200 m31s; large floods are omitted from the fit because of their small
number. The solid line corresponds to Ha = 0.5 1 and from (8.3 1) we have
F = 3.3. Results for station 11-0980 (Pasadena, CA) are given in Figure
8.26(b), the solid line is the best fit of (8.30) with the data over the range
10 < Q, < 100 m31s. The solid line corresponds to H a = 0.87, and from
(8.31) we have F = 7.4. In both cases the fit to the power-law (fractal) rela-
tion is quite good. The values of Ha and F in California are considerably
larger than in Massachusetts. Large floods are relatively more probable in
the arid climate than in the temperate climate.
Many statistical distributions have been applied to historical records of
floods. Benson (1968) has given six statistical correlations for each of his
Figure 8.25. Maximum

annual floods Qm for (a)
station 1 - 1805 on the Middle
Branch of the Westfield
River, Goss Heights,
Massachusetts, and (b)
station 11 -0980 in the
Arroyo Seco near Pasadena,
California.
GEOMORPHOLOGY 211
ten benchmark stations. His results for the 2-parameter gamma (Ga), Gum-
be1 (Gu), log Gumbel (LGu), log-normal (LN), Hazen (H) and log Pearson
type I11 (LP) are given in Figure 8.26(a) for station 1-1805 and in Figure
8.26(b) for station 11-0980. For large floods the fractal prediction (F)corre-
lates best with the log Gumbel (LGu), whereas the other statistical tech-
niques predict longer recurrence times for very serious floods. The fractal
and log Gumbel are essentially power-law correlations, whereas the others
are essentially exponential.
Figure 8.26. The points are

the observed floods at the
measuring stations during
the periods considered (a) for
Station 1-1805 and (b) for
Station 11-0980. The peak
discharge Qm is given as a
function of recurrence
intervals T. The power-law
(fractal) prediction, F, is
compared with the six
statistical predictions given
by Benson (1 968). 2-
parameter gamma (Ga),
Gumbel (Gu), log Gumbel
(LGu), log-normal (LN),
Hazen (H), and log Pearson
type I11 (LP). For the fractal
correlation (F)the
corresponding values of Ha
and Fare given.
212 GEOMORPHOLOGY
The values of Ha,D, and F are given for all ten benchmark stations in
Table 8.2. The correlations with the fractal relation (8.30) in Figure 8.26 are
typical of the ten stations. The parameter F is a measure of the relative sever-
ity of flooding. The higher the value of F, the more likely that severe floods
will occur. Our results show that there are clear regional trends in values of
F. The values in the southwest including Nevada (F = 4.13) and New Mexico
( F = 4.27) as well as California ( F = 7.4) are systematically high. These high
values can be attributed to the arid conditions and the rare tropical (mon-
soonal) storm that causes severe flooding. Central Texas (F = 5.24) is also
high and Georgia (F = 3.47) is intermediate. These areas are influenced by
hurricanes. The northern tier of states including Massachusetts (F = 3.26),
Minnesota (F = 2.95), Nebraska ( F = 3.47), and Wyoming ( F = 3.31) range
from low values in the east to intermediate values in the west. Washington
(F = 2.04) has the lowest value of the stations considered; this low value is
consistent with the maritime climate where extremes of climate are rare.
We have also determined the Hurst exponent Hu for the ten benchmark
stations. Values of R/S for T = 5, 10,25, and 50 years (WS = 1 for T = 2 by
definition) are given in Figure 8.27(a) for station 1-1805 (Goss Heights,
MA) and in Figure 8.27(b) for station 11-0980 (Pasadena, CA). Good corre-
lations are obtained with (7.59) taking Hu = 0.67 for station 1-1805 and Hu =
0.68 for station 11-0980. Values of Hu for all ten stations are given in Table
8.1. The values are nearly constant, with a range from 0.66 to 0.73, indicat-
ing moderate persistence. It is not surprising that the values of the Hausdorff
measure Ha differ from the values of the Hurst exponent Hu since the former
refers to the statistics of the flood events and the latter to the statistics of the
running sum.
The results indicate that there is considerable variation of F but very lit-
tle variation in Hu. Mandelbrot and Wallis (1968, 1969b) introduced the
Table 8.2. Values of the Hausdorff measure Ha, fractal dimension D, flood
intensity factor F, and Hurst exponent Hu for the ten benchmark stations
Station River (State) Ha D F Hu

1-1805 Westfield (MA) 0.513 1.39 3.26 0.67
2-2 185 Oconee (GA) 0.540 1.46 3.47 0.72
5-3310 Mississippi (MN) 0.470 1.53 2.95 0.72
6-3440 Little Missouri (wy) 0.520 1.48 3.31 0.72
6-8005 Elkhorn (NE) 0.540 1.46 3.47 0.67
7-2 165 Mora (NM) 0.630 1.37 4.27 0.73
8-1500 Llano (TX) 0.719 1.28 5.24 0.70
10-3275 Humboldt (NV) 0.616 1.38 4.13 0.66
1 1-0980 Arroyo Seco (CAI 0.870 1.13 7.40 0.68
12-1570 Wenatchee (WA) 0.310 1.69 2.04 0.72
GEOMORPHOLOGY 213
Noah and Joseph effects. The Noah effect is the skewness of the distribution
of flows in a river (or of a non-Gaussian distribution) and the Joseph effect is
the persistence of the flows. It is reasonable to conclude that the variations in
F can be attributed to the Noah effect and the constancy of the Hurst expo-
nent can be attributed to the Joseph effect. An important conclusion is that
R/S analysis is not relevant to flood-frequency hazard assessments.
'T yrs
lo r-
-
-
-
5 -
-
R -
s
-
2 -
Figure 8.27. The rescaled
range (WS) for several
intervals T. (a) Station
1 - 1805. (b) Station 11 -0980.
1 I I I I I I I I I I I The correlations are with
2 5 10 20 50
(7.59) and the Hurst
'$ yrs exponents Hu are given.
214 GEOMORPHOLOGY
8.8 Wavelets
Fourier transforms have a long history of applications to a wide variety of

problems. They have great utility in terms of obtaining the frequency content
of a time series. Despite the many advantages of Fourier transforms, there
are also disadvantages. To overcome some of these disadvantages Grossman
and Morlet (1984) introduced the wavelet transform. This transform has a
fractal basis and is particularly useful when applied to local, nonperiodic,
multiscaled phenomena such as stream flows.
The wavelet transform is essentially a filter that is passed over a time se-
ries, the width of the filter being generally increased by powers of two. The
generalized form of the wavelet transform is given by
whereflt') is the time series and g [(t' - t)la] is the filter. The filter is cen-
tered at t and a is a measure of the width of the filter. The quantity g (t') is
known as the "mother wavelet." Other wavelets are rescaled versions of the
mother wavelet. The area of each wavelet must sum to zero so that
When a is increased by powers of two, a suite of wavelets is generated that

can accommodate a wide range of scales in the signalflt).
A commonly used mother wavelet is the "Mexican hat" wavelet, which
is the second derivative of the Gaussian distribution and takes the form
Figure 8.28. Mother

Mexican hat wavelet g ( r )
from (8.34).
GEOMORPHOLOGY 215
This wavelet satisfied the conditions given in (8.33). Substitution of (8.34)

into (8.32) gives
W ,a ) = ( ) Jrn[I - ( 1 (
e x - 2 a 1 t )dt (8.35)
The mother Mexican hat wavelet is illustrated in Figure 8.28. For more de-
tails of the wavelet transform, the reader is referred to Schiff (1992),
Daubechies (1988), and Young (1992).
To illustrate the application of the wavelet transform we will consider
two streamflow time series (Smith et al., 1997). We first consider a seven-
year, daily discharge record for the Ammonoosuc River, Maine. The record
of daily discharges is given in Figure 8.29(a). The record is characterized by
Figure 8.29. Wavelet

transforms for a seven-year,
daily discharge record for
the Ammonoosuc River,
Maine. (a) Discharge record.
(b) Wavelet scalogram.
0 500 1000 1500 2000 2 500 (c) Wavelet transform
magnitudes for a = 1, 2.4, 8,
t (days) 16.32.64, and 128 days.
216 GEOMORPHOLOGY
long-period high flows associated with the annual spring snow-melt events
and by short-period high flows associated with rainstorm events throughout
the year. Wavelet-transform magnitudes of the time series W(a,t) are given in
Figure 8.29(c) for scales a = 1, 2, 4, 8, 16, 32, 64, and 128 days. The trans-
form magnitudes are contoured using a single threshold W= 5.5 X 10-6 m3/s
in Figure 8.29(b) giving a wavelet scalogram. Regions of this a versus t plot
in which the magnitude threshold is exceeded appear in black. The seven an-
nual snow-melt events are clearly illustrated.
As a second example we consider the hourly discharge record for the
Forrest Kerr Creek in northwestern British Columbia, Canada. The 31 1 km2
drainage area contains several large glaciers, and we consider the 100-day
summer 1992 record given in Figure 8.30(a). The record has a strong diurnal
variation associated with the daily melting cycle. Wavelet transform magni-
tudes of the time series W(t,a) are given in Figure 8.30(c) for scales a = 1,2,
4,8, 16,32,64, and 128 hours. The corresponding contoured wavelet scalo-
gram is given in Figure 8.30(b). The strong diurnal signal appears in the a =
8 and 16 hour wavelet transform magnitudes and are clearly illustrated in the
wavelet scalogram. In addition there are melt episodes with a 5-20-day peri-
ods associated with the regional climate.
Figure 8.30. Wavelet

transforms for a 100-day,
hourly discharge record
for the Forrest Kerr Creek,
British Columbia, Canada.
(a) Discharge record.
(b) Wavelet scalogram.
(c) Wavelet transform
magnitudes for a = 1 , 2 , 4 , 8 ,
16,32,64, and 128 hours. t (hours)
GEOMORPHOLOGY 217
Wavelet transforms have been applied to a variety of geophysical time

series including seismograms (Lou and Rial, 1995), the length of day (Gam-
bis, 1992), the geomagnetic field (Alexandrescu et al., 1995), bathymetric
profiles (Little et al., 1993; Little, 1994), the earth's geoid (Cazenave et al.,
1995), and solar irradiance (Gambis, 1992). Flandrin (1993) has applied
wavelet analysis to fractional Brownian walks. Wavelet analyses are particu-
larly useful when a time series has strong fluctuations over a wide range of
time scales. The application of Fourier transforms to this class of problems
can easily lead to spurious results.
Problems
Problem 8.1. Consider the fractal tree illustrated in Figure 8.31(a). The
length-order ratio is Rr = 2. (a) Determine the bifurcation ratios for 1st to
2nd, 2nd to 3rd, and 3rd to 4th order branches. What are the correspond-
ing fractal dimensions? (b) Write the branch-number matrix and the
branch-ratio matrix for this tree. (c) Determine T, for k = 1, 2, 3. (d) De-
termine a and c as defined in (8.8). (e) Determine the asymptotic bifur-
cation ratio and fractal dimension for large trees of this form from (8.14)
and (8.15).
Problem 8.2. Consider the fractal tree illustrated in Figure 8.31(b). The
ing fractal dimensions? (b) Write the branch-number matrix and
branch-ratio matrix for this tree. (c) Determine T, for k = 1,2, 3. (d) De-
termine a and c as defined in (8.8).
Figure 8.3 1. Illustration of

several fractal trees.
218 GEOMORPHOLOGY
Problem 8.3. The asymptotic branching relation (8.14) is not valid for the
fractal tree illustrated in Figure 8.31(b) because it has triple branching
rather than double (binary) branching. (a) Show that the correct relation
is
(b) Determine the asymptotic branching ratio for the fractal tree illus-
trated in Figure 8.3 1(b) and the corresponding fractal dimension.
Problem 8.4. Consider the fractal tree illustrated in Figure 8.31(c). The
ing fractal dimensions? (b) Write the branch-number matrix and
branch-ratio matrix for this tree. (c) Determine T, fork = 1 , 2 , 3 . (d) De-
termine a and c as defined in (8.8). (e) Determine the asymptotic bifur-
cation ratio and fractal dimension for large trees of this form using the
result given in Problem 8.3.
Problem 8.5. Consider the deterministic fractal growth model illustrated in
Figure 8.32. An initial equilateral triangle at zero order (a) grows at first
order (b) by the addition of three triangles at the three corners of the
original seed particle. At second order (c), three of the first-order con-
struction are added as shown.
(a) What is the fractal dimension of this construction?
(b) What is the radius of gyration at first and second order? What is the
corresponding fractal dimension from (8.22)?

deterministic fractal growth
model based on equilateral
triangles.
Chapter Nine
DYNAMICAL
SYSTEMS
9.1 Nonlinear equations
We now turn our attention to some examples of deterministic chaos that have
applications in geology and geophysics. There are two requirements for a so-
lution that exhibits deterministic chaos. The first requirement is that we are
solving deterministic equations with specified initial andlor boundary condi-
tions. Thus the applicable equations are deterministic, not statistical. The
second requirement is that solutions that have initial conditions that are in-
finitesimally close diverge exponentially as they evolve. However, before
we consider solutions that are chaotic, some necessary introductory material
is presented. Chaotic behavior is found only for nonlinear systems of equa-
tions. In this chapter some of the standard nomenclature for the study of non-
linear equations is presented (Verhulst, 1990).
Probably the simplest nonlinear total differential equation is the logistic
equation
This equation is a simple model for population dynamics with x representing

population and t time; the first term on the right-hand side accounts for
births and the second deaths. The parameter T is the characteristic time and
xe is a representative population. It is standard procedure in treating nonlin-
ear equations to introduce nondimensional variables. This procedure reduces
the number of parameters that must be varied to study the solutions. For the
logistic equation it is appropriate to introduce
220 DYNAMICAL SYSTEMS
and there are no governing parameters. An exact solution of (9.3) is
where Z = Z, at t = 0 is the initial condition. Solutions of (9.4) for various

values of Xo are given in Figure 9.1.
Although (9.1) is solved by writing (9.4), it is illustrative to examine the
solutions in somewhat more detail. We first examine the fixed points ip of
(9.3). As the solutions of (9.3) evolve in time, the time dependence dies out
and i approaches a particular value. This value is known as a stable fixed
point. The fixed points are formally obtained by setting dildl = 0 in (9.3)
with the result that
If the initial condition is either of these values, then no time dependence is

obtained. It is also of interest to examine the stability of the fixed points. To
do this the applicable equation is linearized about the fixed point. In the
vicinity of the fixed point ip = 0, it is appropriate to neglect the quadratic
term in (9.3) with the result
Figure 9.1. Solutions of the

logistic equation (9.1) for
several initial conditions go.
All solutions converge to the
fixed point X = 1.
DYNAMICAL SYSTEMS 221
d x- -
- -X
dt
The solution is
T = &,el
where i , is the value of 2 at t = 0 and is assumed to be finite but small. Thus

solutions in the immediate vicinity of the fixed point 3, = 0 diverge with
time and are unstable; this fixed point is unstable.
To examine the fixed point = 1 it is appropriate to introduce the new
variable 2,:
Substitution into (9.3) gives
In the vicinity of the fixed point Z, = 0 it is appropriate to neglect the qua-

dratic term in (9.9) with the result
which has the solution
where i,,is again assumed to be small but finite. As time evolves i , ap-
proaches zero. Thus the solutions in the immediate vicinity of the fixed point
-
x = 1 are stable. The stability of the fixed point 3p= 1 is clearly illustrated in
dgure 9.1. For all initial conditions iothe solut~ons"flow" in time toward
the stable fixed point 2, = 1. Also, adjacent solutions tend to converge to-
ward each other. These solutions are not chaotic.
To discuss further the behavior of nonlinear equations we consider the
van der Pol equation
If a = p = 0 this is the equation of motion of a spring-mass oscillator system

with M the mass, k the spring constant, and x the extension of the spring. The
coefficients ci and p represent the linear and nonlinear damping terms, re-
spectively. Again it is appropriate to introduce nondimensional variables.
The frequency of the harmonic oscillator o = (klM)lQintroduces a natural
time to the problem. The relative amplitudes of the damping terms a and P
introduce a natural length scale (a/P)112/o to the problem. Using these time
and length scales we introduce the nondimensional variables t and i ;we also
define the nondimensional parameter E according to
Substitution of (9.13)into (9.12)gives
with E the only parameter governing the behavior. In considering the solu-
tions of this second-order nonlinear equation, it is standard practice to intro-
duce the definition of velocity
and to rewrite (9.14)as
Dividing (9.16)by (9.15)gives
The E-space is known as the phase space or phase plane. Solutions of (9.17)
follow phase trajectories in this two-dimensional plane with time t as a para-
meter.
We first consider the solution of (9.17) when E = 0. In this case it be-
comes
DYNAMICAL SYSTEMS
which integrates to give
Simple harmonic motion is a circle in the phase plane. The radius of the cir-
cle is determined by the initial nondimensional amplitude 2, (or the initial
velocity). The relation (9.19) also represents conservation of energy in this
nondissipative system; it is the sum of the potential and kinetic energies. For
this system the fixed point at x = j = 0 is known as a center. The behavior is
illustrated in Figure 9.2(a).
For finite values of E it is necessary to solve (9.17) numerically. The re-
sult for E = 1 is given in Figure 9.2(b). solutions for all initial conditions
converge toward a limit cycle; this limit cycle is independent of the initial
conditions. The physical reason for this behavior can be seen in the original
van der Pol equation (9.12). For small amplitudes the negative linear damp-
ing term dominates and the amplitude increases. For large amplitudes the
positive cubic damping term dominates and the amplitude decreases. The re-
sult is that all solutions converge on the same limit cycle at large times.
Many sets of equations that produce deterministic chaos for a range of para-
meter values produce limit cycles for other parameter values.
Before considering chaotic solutions we will present some further intro-
ductory material on singular points. Consider the pair of linear total differen- Figure 9.2. (a) Solutions of
the nondimensional van der
tial equations Pol equation (9.14) in the
phase plane with E = 0.The
solution is a circle
representing simple
harmonic motion. The
position of the circle is
dependent upon initial
conditions. (b) Solution of
the van der Pol equation
(9.14) in the phase plane
with E = 1. The solutions
approach a limit cycle
independent of initial
conditions. Solutions for
initial conditions inside the
limit cycle spiral out to it and
solutions for initial
conditions outside the limit
cycle spiral into it.
where a, b, c, and f are constants. Dividing (9.20) by (9.21) gives
If b = - c and a =f = 0 this reduces to (9.18); the solution is given by (9.19)

and is a circle in the xy-plane as illustrated in Figure 9.2(a). If a = a/f and b =
c = 0 we can write (9.22) as
This equation has the solution
y = yx"
where y is the constant of integration. If a > 0 the fixed pointy = x = 0 is a

node. The behavior for a = 1 is illustrated in Figure 9.3(a) and for a = 2 in
Figure 9.3(b). If a > 0 and f < 0, the solutions converge on y = x = 0 and the
fixed point is a stable node. If a > 0 and f > 0, the solutions diverge from
y = x = 0 and the fixed point is an unstable node.

singular point behavior; (a)
and (b) are nodes and (c) is a
saddle point.
If a < 0 in (9.23) and (9.24), the fixed pointy = x = 0 is a saddle point.

Its behavior for a = - 1 is illustrated in Figure 9.3(c). Only the singular solu-
tions y = 0 or x = 0 enter or leave the fixed pointy = x = 0. If x = 0 we have
and the fixed point is stable for a < 0. If y = 0 we have
and the fixed point is stable for f < 0. Since a and f must have opposite signs,
one singular solution will be stable and the other singular solution will be
unstable.
We next substitute b = 1, c = - 1, and a =f = a in (9.22) with the result
dy-- x + ay
-
du ax-y
Changing to polar coordinates p and 0 we substitute the variables
x = p cos 0
y = p sin 0
giving
With p = p, at 0 = 0 this is integrated to give
p=
This solution is a logarithmic spiral and the fixed point at y = x = 0 in (9.27)

is known as a spiral.
9.2 Bifurcations
We now turn to the subject of bifurcations. Solutions to a set of nonlinear

equations generally experience a series of bifurcations as they approach
chaotic behavior. These bifurcations occur when a parameter of the system is
varied. We first consider the equation
where p is considered to be a parameter. The fixed points of this equation

obtained by setting dxldt = 0 are
When p is negative there are no real fixed points, and when p is positive
there are two real fixed points. The transition at p = 0 from no solutions to
two solutions is known as a turning point bifurcation. We examine the stabil-
ity of the two real roots by linearization. We substitute
into (9.32); after dropping the quadratic term in x , we have
Thus the fixed point x = pl/2 is stable: solutions as they evolve in time con-
verge to it. The fixed point x = - p"2 is unstable: solutions as they evolve in
time diverge from it. The corresponding bifurcation diagram is given in Fig-
ure 9.4(a). This figure illustrates the meaning of the word bifurcation, to
split into two branches. This figure also shows that for x < -p1I2all solu-
tions diverge to x = - = and for - pl/2 < x < + = all solutions converge to
the stable fixed point x = ~ 1 1 2 .
We now turn to a modified form of the logistic equation (9.1)
where p is again considered to be a parameter. There are fixed points at x = 0

and x = p for all values of p. For p < 0 there is an unstable fixed point at x =
(I, and a stable fixed point at x = 0. As p increases the unstable fixed point ap-
proaches the origin and coalesces with it when p = 0. For p > 0 the fixed
point at x = p is now stable. This is known as a transcritical bifurcation, and
an exchange of stabilities between the two fixed points has taken place at
p = 0. The corresponding bifurcation diagram is given in Figure 9.4(b).
We next turn to a third class of bifurcations, the pitchfork bifurcations.
We first consider the supercritical pitchfork bifurcation, an example of
which is given by the equation
Note that this equation is invariant under the transformation x' = - x . Thus
solutions are symmetric in x and fixed point must appear or disappear in
pairs. The fixed points of (9.37) obtained by setting dxldt = 0 are
When IJ, is negative there is a single real fixed point x = 0, and when p, is
positive there are three fixed points x = 0, 2 ~ ' 1 2 The
. transition at p = 0
from one to three solutions is known, for obvious reasons, as a pitchfork bi-
furcation (although the word trifurcation would be more appropriate). A sta-
bility analysis shows that for (I < 0 the solution x = 0 is stable. For b > 0
this solution is unstable but the other solutions are stable. The corresponding
bifurcation diagram is shown in Figure 9.4(c). For < 0 all solutions con- Figure 9.4. (a) Illustration of
a turning point bifurcation
occurring at y = 0. The
stable and unstable fixed
points of (9.32) are given as
a function of y. (b)
Illustration of a transcritical
bifurcation and exchange of
stabilities occurring at y = 0.
no solutions The stable and unstable fixed
points of (9.36) are given as
a function of y. (c)
Illustration of a supercritical
I ... _
, UnSlrble branch
pitchfork bifurcation
1 ----_ occurring at y = 0. The
stable and unstable fixed
points of (9.36) are given.
The transition is from a
single stable branch for
y < 0 to three branches, two
stable and one unstable, for
y > 0. (d) Illustration of a
subcritical pitchfork
bifurcation occurring at y =
unnUble branch 0. The stable and unstable
fixed points of (9.39) are
given. The transition is from
three branches, one stable
and two unstable, for y < 0
to a single unstable branch
for y > 0.
verge to the stable fixed point x = 0. For p > 0 all solutions for x > 0 con-
verge to the stable fixed point x = pllz and all solutions for x < 0 converge to
the stable fixed point x = - p"2.
An example of a subcritical pitchfork bifurcation is given by the equa-
tion
When p is negative ( p < 0) there are two unstable fixed points at x = 2

(-p)'/z and a stable fixed point at x = 0. For positive p ( p > 0) the only
fixed point is at x = 0 and it is unstable. In this region all solutions diverge to
+ w. Since solutions converge to a finite value of x only for p < 0, the term
subcritical is used. The applicable bifurcation diagram is shown in Figure
9.4(d).
Finally we consider the pair of equations
As in (9.28) and (9.29) it is again appropriate to introduce polar coordinates

p and 0 in the xy phase plane:
x = p cos 0 (9.42)
y = p sin 0 (9.43)
Substitution into (9.40) and (9.41) gives

Hopf bifurcation at p = 0. stable
periodic
The limiting solutions of orbit
(9.40) and (9.41) are given
for various values of p. For
-----
P
p < 0 the origin x = y = 0 is a unstoble
stable branch. For p > 0 branch
there are stable limit cycles,
which are circles in the
xy-plane with radius p = p ' / Z .
These equations have the fixed point solution p = 0 (x = y = 0); it is stable for
y < 0 and unstable for y > 0. In addition, for y > 0, solutions of (9.44) and
(9.45) converge to a circular limit cycle given by
These solutions are illustrated in Figure 9.5. For y < 0, all solutions spiral
into the stable fixed point p = 0. For p > 0, solutions for p > p1/2 spiral into
circular limit cycle given by (9.46): solutions for p < y112 spiral outward to
this circular limit cycle. The transition from a stable branch for y < 0 to a
stable limit cycle for y > 0 is a Hopf bifurcation. The van der Pol equation
(9.14) also undergoes a Hopf bifurcation at E = 0.
Problems
Problem 9.1. For b = c = 0 in (9.20) and (9.21) solve for y(t) and x(t) directly.
Show that these solutions reduce to (9.24).
Problem 9.2. Derive (9.30) from (9.27)-(9.29).
Problem 9.3. Solve (9.36) in the vicinity of the three fixed points.
Problem 9.4. Derive (9.44) and (9.45) from (9.40) and (9.4 1).
Problem 9.5. Consider the equation
(a) What are the fixed points?
(b) Are they stable or unstable?
(c) Show that x =

(1 + xo)e2' - 1 + xo is a solution if x = x, at t = 0.
(1 + x,)e2' + 1 - xo
(d) Sketch solutions for xo = -2, 0, 2 and discuss in terms of the fixed
points.
Problem 9.6. Consider the nondimensional logistic equation (9.3). Deter-
dx
mine the solution in the E phase plane, where i = -.
dt
dx
Problem 9.7. Consider the equation -= x - x3
dt
(c) Show that x =

xie2'
X: e2'
+ 1- xi I is a solution if x = x, at t = 0.
(d) Sketch solutions for x, = - 1.5, -0.5,0.5, and 1.5.

dx
Problem 9.8. Consider the equation - = sin x
dt
(c) Show that x = 2 tan-'

[e'tan (31-
IT 3IT
if x = x, at t = 0.
Sketch the solution for x, = - -.

2' 2
dx
Problem 9.9. Consider the equation - = px + x2. Determine the fixed
dt
points and sketch the bifurcation diagram. What type of bifurcation is
this?
dx
Problem 9.10. Consider the equation = x - px3.Determine the fixed
-
dt
this? dx
Problem 9.11. Consider the equation - = x + px3.Determine the fixed
dt
this?
ChapterTen
LOGISTIC M A P
10.1 Chaos
The concept of deterministic chaos is a major revolution in continuum me-

chanics (Berg6 et al., 1986). Its implications may turn out to be equivalent to
the impact of quantum mechanics on atomic and molecular physics. Solu-
tions to problems in solid and fluid mechanics have generally been thought
to be deterministic. If initial and boundary conditions on a region are speci-
fied, then the time evolution of the solution is completely determined. This
is in fact the case for linear equations such as the Laplace equation, the heat
conduction equation, and the wave equation.
However, the problem of fluid turbulence has remained one of the major
unsolved problems in physics. Turbulent flows govern the behavior of the
oceans and atmosphere. The appropriate Navier-Stokes equations can be
written down, but solutions yielding fully developed turbulence cannot be
obtained. It is necessary to treat turbulent flows statistically and to carry out
spectral analyses.
The concept of deterministic chaos bridges the gap between stable deter-
ministic solutions to equations and deterministic solutions that are unstable to
infinitesimal disturbances. Chaotic solutions must also be treated statisti-
cally; they evolve in time with exponential sensitivity to initial conditions. A
deterministic solution is defined to be chaotic if two solutions that initially
differ by a small amount diverge exponentially as they evolve in time. The
evolving solutions are predictable only in a statistical sense. A necessary con-
dition that a solution be chaotic is that the governing equations be nonlinear.
As our first example of deterministic chaos we consider the logistic map
This is a recursive relation that determines the sequence of values x,, x,,
x,, . . . . An initial value, x,, is chosen; this value is substituted into (10.1) as
232 LOGISTIC M A P
x, and x, is determined as xn+,.This value of x, is then substituted as x, and

x, is determined as x,+ 1. The process is continued iteratively. This is re-
ferred to as a map because the algebraic relation maps out a sequence of val-
ues of x,; xorx,, x2, . . . . The procedure is best illustrated by taking a specific
example. Assume that a = 1 and x, = 0.5 and substitute these values into
(10.1), giving x, = 0.25. This value is then substituted as x, and we find that
x, = 0.1875. The iterations of (10.1) were studied by May (1976) and have a
remarkable range of behavior depending upon the value of a that is chosen.
This is in striking contrast to the rather dull behavior of the logistic differen-
tial equation given in (9.1).
The logistic map is a simple representation of the population dynamics
of a specie with an annual breeding cycle (May and Oster, 1976). The quan-
tity x, is the population of the specie in year n and the parameter a can be in-
terpreted as the average net reproductive rate of a population.
To better understand the behavior of the logistic map, it is appropriate to
study the functional relation
fix) = ax(1 - x) (10.2)
The fixed points x, of this equation are obtained by settingfix,) = x, with the
result
This is equivalent to setting xn+,= x,. The two fixed points obtained by solv-
ing (10.3) are
An essential question is whether the iterative mapping given by (10.1) will

evolve to these fixed points. Depending on the behavior offix) in the vicinity
of the fixed point, the fixed point can be either stable or unstable. Solutions
will iterate toward stable fixed points and will iterate away from unstable
fixed points. We introduce
LOGISTIC M A P 233
This is the slope of the functionfix) evaluated at the fixed point x,. If Irl < 1,
where (r( is the absolute value of r, the fixed point is attracting (stable), but
if Irl > 1 the fixed point is repelling (unstable). For the logistic map from
(10.2) we find that
For positive values of a we find that the fixed point at x, = 0 is stable for 0 <
a < 1 and unstable for a > 1. The fixed point x, = 1 - a-1 is unstable for 0 <
a < 1, stable for 1 < a < 3, and unstable for a > 3.
We next examine a sequence of iterations of the logistic map (10.1). As
our first example we consider the iteration for a = 0.8 as illustrated in Figure
10.1. The curve represents the functionflx) given by (10.2) for a = 0.8. Tak-
ing x, = 0.5 we draw a vertical line; its intersection with the parabolic curve
gives x, = 0.2. A horizontal line drawn from this intersection to the diagonal
line of unit slope transfers xn+, to xn.A vertical line is drawn to the parabola
giving x, = 0.128. Further iterations give x, = 0.0892928, x, = 0.06505567,
etc. The sequence iterates to the stable fixed point xf = 0. All iterations converge
to x, = 0 for 0 < x, < 1. As our next example we consider two iterations for

the iteration of the logistic
map (10.1) for a = 0.8. The
iteration from x, = 0.5
converges on the stable fixed
point x, = 0.
23 4 LOGISTIC M A P
a = 2.5 as illustrated in Figure 10.2. The parbolic curveflx) given by (10.2)

now intersects the diagonal at the two fixed points given by (10.4) and
(10.5),x, = 0 and x, = 0.6. For x, = 0, A = 2.5 and the singular point is unsta-
ble; for x, = 0.6, A = 0.5 and the singular point is stable. For x, = 0.1 we find
x , = 0.225, x, = 0.43594, x, = 0.61474 and the iteration converages on the
fixed point x, = 0.6. For x, = 0.8 we find x , = 0.4 and x, = 0.6 with no further
iteration required. All iterations converge to x, = 0.6 for 0 < x, < 1 . This is
consistent with the stability of the fixed points discussed above.
At a = 3 a flip bifurcation occurs. Both singular points are unstable and
the iteration converges on a limit cycle oscillating between x,, and x,. The
period of the oscillation doubles from one iteration, n = 1, for a C 3, to two
iterations, n = 2, for a > 3. The values of x,, and x, are obtained from the lo-
gistic map (10.1) by writing
The limit cycle oscillates between x,, and x,. As an example of the period
n = 2 limit cycle, we consider the iteration for a = 3.1 given in Figure 10.3.
The iteration from x, = 0.1 approaches the limit cycle that oscillates between
x,, = 0.558 and x, = 0.765. The n = 2 limit cycle occurs in the range 3 < a <
3.449479. At a = 3.449479 another flip bifurcation occurs and the period

map (10.1) for a = 2.5. The
iterations from xo = 0.1 and
xo = 0.8 converge on the
stable fixed point at x, = 0.6.
LOGISTIC M A P 235
doubles again so that n = 4. As an example of the period n = 4 limit cycle, we

consider the iteration for a = 3.47, which is given in Figure 10.4. The itera-
tion from x, = 0.1 approaches the n = 4 limit cycle that oscillates between x,, =
0.403, xo = 0.835, x, = 0.479, and x, = 0.866. The n = 4 limit cycle occurs in

map(lO.l)fora= 3.1.The
converges to the n = 2 limit
cycle between x,, = 0.558
and x, = 0.765.

map (10.1) for a = 3.47. The
converges to the n = 4 limit
cycle between xf, = 0.403,
x, = 0.835, x, = 0.479, and
x, = 0.866.
236 LOGISTIC M A P
the range 3.449499 < a < 3.544090.At larger values of a higher-order limit
cycles are found. They are summarized as follows:
where n is the period of the limit cycle and k is the number of flip bifurca-
tions that have occurred. Period-doubling flip bifurcations occur at a se-
quence of values a,, where a , = 3, a, = 3.449499, a3 = 3.544090, a, =
3.564407, a, = 3.568759, a, = 3.569692, a, = 3.569891, a, = 3.569934, etc.
In the region 3.569946 < a < 4 windows of chaos and multiple cycles occur.
The values of a, approximately satisfy the Feigenbaum relation
Where F = 4.669202 is the Feigenbaum constant. This becomes a better ap-

proximation as k becomes larger. This relation indicates a fractal-like, scale-
invariant behavior for the period-doubling sequence of bifurcations. The
Feigenbaum relation can also be written in the form
Thus the initial values of the period-doubling sequence can be used to pre-
dict the onset of chaotic behavior at a=. Taking a , = 3 and a, = 3.449499, we
find that a_ = 3.572005 from (10.12).Taking a, and a, = 3.544090, we find
a_ = 3.569870. Taking a, and a, = 3.564407, we find a_ = 3.569944. Taking
a, and a, = 3.568759, we find a_ = 3.569945. These are clearly converging
on the observed value of am= 3.569946.
We now turn to the behavior of the logistic map in the region of cha-
otic behavior. An example illustrating chaotic behavior is given in Figure
10.5 with a = 3.9; one thousand iterations are shown and no convergence
to a limit cycle is observed. The behavior is space filling (chaotic) but the
range of values of xn is well defined. The maximum value is obtained taking
xn = 0.5 with the result xn+, = 0.975. The minimum value is obtained tak-
LOGISTIC M A P 237
ing x,, = 0.975 with the result x n + , = 0.0950625. Thus we have for a = 3.9 that
0.0950625 < x,, < 0.975. For a = 4 the logistic map (10.1) becomes
This iteration can be expressed analytically by taking
The nth iteration of this relation is
Provided P is not a rational number, the values of x,, jump around randomly
and fully chaotic behavior is obtained.
The route to chaos and the windows of chaotic behavior of the logistic
map are illustrated in the bifurcation diagram given in Figure 10.6. The
asymptotic, large n, behavior of the map is illustrated for 2.9 < a < 4.0. At

map (10.1) for a = 3.9. The
iteration from 0.9 gives
chaotic behavior; 1000
iterations are shown.
238 LOGISTIC M A P
a = 2.9 the fixed point x, = 0.655 is shown. At a = 3 the fixed point is

x, = 0.66767 and the period-doubling flip bifurcation to the n = 2 limit cycle
is shown. In the interval 3 < a < 3.449499 the two values of x, correspond-
ing to the n = 2 limit cycle are given. At a = 3.449499 the period-doubling
flip bifurcation to the n = 4 limit cycle is shown. In the interval 3.449499 <
a < 3.544090 the four values of x, corresponding to the n = 4 limit cycle are
given. In the interval 3.544090 < a < 3.569946 an infinite sequence of pe-
riod-doubling flip bifurcations occurs as n + m. For the higher values of a
the windows of chaotic behavior are illustrated by the cloud of points.
Chaotic behavior results in an infinite set of random values of x, with a
well-defined range of values; this range is clearly illustrated in Figures 10.5
and 10.6. The maximum value of x, is the maximum value offlx) from (10.2)
and this maximum is at x = 0.5; thus, we have
Taking a = 3.9 we have xfmax = 0.975, which is in agreement with the exam-
ple given in Figure 10.5. The minimum value of x, is obtained by substitut-
ing (10.17) into (10.1) with the result
Figure 10.6. Bifurcation

diagram showing the
asymptotic behavior for
large n of the logistic map
(10.1) as a function of a.
LOGISTIC MAP 239
Taking a = 3.9 we have xhin = 0.0950625, which is again in agreement with

the example given in Figure 10.5.
10.2 Lyapunov exponent
Chaotic behavior can be quantified in terms of the Lyapunov exponent A.

The definition of the Lyapunov exponent is
where dxn is the incremental difference after the nth iteration if dxo is the in-
cremental difference in the initial value. If the Lyapunov exponent is nega-
tive, adjacent solutions converge and deterministic solutions are obtained. If
the Lyapunov exponent is positive, adjacent solutions diverge exponentially
and chaos ensues. To determine the Lyapunov exponent, we consider the in-
cremental divergence in a single iteration by writing (10.1) in the form
where f(x) is the functional form of the mapping; for the logistic map it is
given by (10.2). Since
Xn+1 = f ( x n )
by definition, (10.20) can be written
And for the logistic map from (10.2) we find
From ( l o . 19) and (10.22) the definition of Lyapunov exponent is
1 "
A = lim
rn+ m
-
m n=O
log,
1H.1
240 LOGISTIC M A P
where log, is the logarithm to the base 2. The Lyapunov exponents A for the
logistic map (10.1) are given in Figure 10.7 for a range of values for a. The
windows of chaotic behavior for 3.569946 < a < 4 where A is positive are
clearly illustrated. The Lyapunov exponent goes to zero at each flip bifurca-
tion as shown.
Consider as a particular example the iteration for a = 4 given by (10.16).
For this case we find
dxn = 2 sin ( ~ " I T Pcos

) (2%p)2nndp (10.25)
and
dx, = 2 sin I$ cos TTPIT~P (10.26)
Combining (10.19), (10.25), and (10.26) gives
I
2hn = [sin (TITP)cos ( T n P ) 2n
sin IT^ cos I$
-
Although the coefficient is variable, the growth with n as n + requires
that A = 1. Thus the Lyapunov exponent for this special case is unity and the
iteration is fully chaotic.
The role of the Lyapunov exponent is clearly illustrated by the simple
linear map
Figure 10.7. Lyapunov

exponents A from (10.24) for
the logistic map (10.1) as a
function of the parameter a.
LOGISTIC M A P 241
The only singular point is at x, = 0 and it is stable if a < 1 and unstable if a >
1. Illustrations of the iteration of this linear map for a = 0.6, 1.2 and x, = 0.8
are given in Figure 10.8. With a = 0.6 we have x , = 0.48, x, = 0.288, x, =
0.1728, and x, = 0.10368, and the solution iterates to the stable fixed point
x, = 0. With a = 1.2 we have x , = 0.96, x, = 1.152, x, = 1.3824, and x, =
1.65888, and the solution iterates to x + oo. If there is an incremental differ-
ence in x,, 6x0, the incremental difference in x , , 6 x , , is 6 x , = a 6x0; similarly
the incremental difference in x,, sx,, is ax, = a s x , = a2 6xo. This can be gen-
eralized to give
Combining (10.19) and (10.29) gives
A = - log a
log 2
Thus the Lyapunov exponent A is positive for a > 1 and adjacent solutions
diverge; the Lyapunov exponent is negative for 0 < a <1 and adjacent solu-
tions converge.
We next consider the triangular or tent map defined by

iterations of the linear map
(10.28). (a) With a = 0.6 and
x,, = 0.8, the iteration
converges on stable fixed
point x, = 0. (b) With a = 1.2
and x,, = 0.8, the iteration
diverges to infinity.
242 LOGISTIC M A P
with 0 < a < 1 and 0 < x < 1 . This map can also be defined by
1
x n + , = 2a xn for 0 < xn < -
2
(10.33)
For 0 < a < $ the only fixed point is at x, = 0 and it is stable. For $ < a < 1
2a
there are fixed points at x, = 0, x, = ;both fixed points are unstable.
(1 + 2a)
Illustrations of the iteration of this map for a = 0.4, 0.8, and x, = 0.8 are
given in Figure 10.9. With a = 0.4 we have x , = 0.16, x2 = 0.128, x3 = 0.1024,
and x4 = 0.08192, and the solution iterates to the stable fixed point x, = 0.
With a = 0.8 we have x , = 0.32, x2 = 0.512, x, = 0.7808, and x4 = 0.35072.
The iteration for a = 0.8 is chaotic. All iterations with 0.5 < a < 1 are

iterations of the triangular
map (10.33). (a) With a = 0.4
and x,, = 0.8, the iteration
converges to the stable fixed
point x, = 0. (b) With a = 0.8
and x, = 0.8, the iteration is
chaotic in the range 0.32 c
x < 0.8.

diagram for the triangular
map (10.33) as a function of
the parameter a.
LOGISTIC M A P 243
chaotic; this can be demonstrated by noting that the Lyapunov exponents for
the triangular map are easily obtained from the Lyapunov exponents for the
linear map given in (10.31); the result is
A = - log 2a
log 2
Thus A is negative for x < a <0.5 and positive for 0.5 < a < 1. In the chaotic
regime, the range of values of x is 2a (1 - a ) < x < a. The resulting bifurca-
tion diagram for the triangular map is given in Figure 10.10. For 0 < a < 0.5
the solutions converge to the stable fixed point x, = 0. For 0.5 < a < 1.0
chaotic behavior is found between the limits given above.
The recursive maps considered in this chapter are clearly very simple
models with limited direct applicability to problems in geology and geo-
physics. However, the complex chaotic behavior exhibited by the simple
models is strongly indicative that many natural systems can be expected to
behave chaotically. Some natural systems exhibit behavior that closely re-
sembles the behavior of recursive maps. As an example Sornette et al.
( 1991) and Dubois and Cheminee (1991) have treated the return periods for
eruptions of the volcanoes Piton de la Fournaise on Reunion Island and
Mauna Loa and Kilauea in Hawaii as return maps. The results appear to re-
semble the chaotic maps considered in this chapter.
Problems
Problem 10.1. Determine x i , x,, x,, and x, for the logistic map (10.1) taking
a = 0.5 and xo = 0.5. What is the value of x,?
Problem 10.2. Determine x , , x,, x,, and x, for the logistic map (10.1) taking
a = 0.9 and x, = 0.75. What is the value of x,?
a = 2 and xo = 0.2. What is that value of x,?
a = 2.5 and xo = 0.3. What is that value of x,?
Problem 10.5. Determine x,, and xn for the logistic map (10.1) taking a =
3.2.
Problem 10.6. Determine x,, and x, for the logistic map (10.1) taking a =
3.4.
Problem 10.7. For a = 3.7 the logistic map (10.1) is fully chaotic. What are
the maximum and minimum values of xn?
Problem 10.8. For a = 3.8 the logistic map (10.1) is fully chaotic. What are
the maximum and minimum values of xn?
244 LOGISTIC M A P
Problem 10.9. Determine x,, x,, x,, x,, and x, for the logistic map (10.1) tak-
inga = 4 and p = ( 2 ~ ) - 1 .
Problem 10.10. Determine x,, x,, x,, x,, and x, for the logistic map (10.1) tak-
inga = 4 and p = ( 3 ~ ) - 1 .
Problem 10.1 1. Show that x, = 0 is a fixed point of the linear map (10.28).
Determine the value of r defined in (10.6) for this map and determine
the stability of the fixed point.
Problem 10.12. Determine the fixed points for the triangular points for the
triangular map (10.32). Determine the values of r defined in (10.6) for
the fixed points and determine their stability.
Problem 10.13. Consider the iterative map
This map is also used for population dynamics (May and Oster, 1976).
(a) Determine the fixed points and the range of positive values for a that
are stable.
(b) Determine x,, x,, x,, x,, and x, taking a = 3 and x, = 0.5.
(c) For a = 3, what are the maximum and minimum values of xn?
Chapter Eleven
SLIDER-BLOCK
MODELS
We next turn to a low-order example of deterministic chaos that has some-

what more direct applications to geology and geophysics. As discussed in
Chapter 4, we accept the hypothesis that earthquakes occur repetitively on
preexisting faults. A simple model for the behavior of a fault is a slider block
pulled by a spring as illustrated in Figure 11.1 (Burridge and Knopoff,
1967). The block is constrained to move smoothly along the surface. It inter-
acts with the surface through friction; this friction prevents sliding of the
block until a critical value of the pulling force is reached. The block sticks
and the force in the spring increases until it equals the frictional resistance to
sliding on the surface, and then slip occurs. The extension of the spring is
analogous to the elastic strain in the rock adjacent to a fault. The slip is anal-
ogous to an earthquake on a fault. This is stick-slip behavior. The stored elastic
strain in the spring is relieved; this is known as elastic rebound on a fault.
The behavior of this simple spring-block model will now be studied
quantitatively. A constant velocity driver moving at velocity v extends the
spring with spring constant k until the pulling force ky equals the frictional
static resisting force Fs. The static condition for the onset of sliding is thus
Figure 1 1 . 1 . Illustration of
the slider-block model for
fault behavior. The constant
velocity driver extends the
spring until the force ky
exceeds the static friction
force Fs.
246 SLIDER-BLOCK MODELS
Once sliding begins the equation of motion for the block is
where rn is the mass of the block and Fd is the dynamic or sliding friction.
The sliding is analogous to an earthquake and relieves the stress in the spring
in analogy to elastic rebound. The further assumption is made that the load-
ing velocity of the driver, v, is sufficiently slow so that we may assume it to
be zero during the sliding of the block. This is reasonable since an earth-
quake lasts only a few tens of seconds, whereas the interval between earth-
quakes on a fault is typically hundreds of years or more.
The static-dynamic friction law is the simplest that generates stick-slip
behavior. A necessary and sufficient condition for stick-slip behavior is that
the static friction exceeds the dynamic friction, F, > F,. A variety of empiri-
cal velocity-weakening friction laws are in agreement with laboratory obser-
vations and also generate stick-slip behavior. Dynamic instabilities associ-
ated with complicated friction laws are well known from single-block
models (Byerlee, 1978; Dieterich, 1981; Ruina, 1983; Rice and Tse, 1986).
Slider-block models have been used to simulate foreshocks, aftershocks,
pre- and post-seismic slip, and earthquake statistics (Dieterich, 1972; Run-
dle and Jackson, 1977; Cohen, 1977; Cao and Aki, 1984, 1986). Gu et al.
(1984) found some chaotically bounded oscillations; Nussbaum and Ruina
(1987) used a two-block model with spatial symmetry and found periodic
behavior. Huang and Turcotte (1990a, 1992) and McCloskey and Bean
(1992) studied the same system without spatial symmetry and obtained clas-
sic chaotic behavior.
We first consider the solution for the behavior of the single block shown
in Figure 11.1. It is convenient to introduce the nondimensional variables
In terms of these variables, the sliding condition (11.1) becomes
and the equation of motion (1 1.2) becomes

SLIDER-BLOCK MODELS 247
We assume sliding starts at T = 0 with Y = 1 (1 1.4) and dY1d.r = 0. The applic-

able solution is
Y = - +
4 (1 - -t) COST
Sliding ends at T = .rr when dY1d.r is again zero. When the velocity is zero the
friction jumps to its static value, preventing further sliding. The position of
the block at the end of sliding is Y = (2/+) - 1 so that the slip during sliding
is
The dependence of Y and d Y l d ~on T during sliding are given in Figure 11.2
for 4 = 1.25. For this case Y drops from 1 to 0.6 during sliding and AY = -0.4.
After sliding is completed, the spring extends due to the velocity of the driver
until Y again equals unity and the cycle repeats. With a single slider block
periodic behavior is obtained. The variables Y and dY1d.r define a phase
plane for the solution.
We next consider the behavior of a pair of slider blocks as illustrated in
Figure 11.3. We will show that the behavior of the blocks can be a classical
example of deterministic chaos. The blocks are an analog of two interacting
faults or two interacting segments of a single fault. A constant velocity driver
drags the blocks over the surface at a mean velocity v. The two blocks are
coupled to each other and to the constant velocity driver with springs whose
constants are kc, k , , and k,. Other model parameters are the block masses m ,
and m, and the frictional forces F, and F,. The position coordinates of the
blocks relative to the constant velocity driver are y , and y,. The static condi-
tions for the onset of sliding are the force balances

Y and dY1d.r on T during a
sliding episode assuming
+ = 1.25.
Once sliding begins the applicable equations of motion are
(11.10)
(11.11)
To simplify the model we assume rn, = rn2 = rn, k, = k2 = k, and Fs,IFD, =

Fs21FD2 = +'
In addition to the nondimensional variables introduced in (1 1.3), we
write the nondimensional parameters
(11.12)
The parameter a is a stiffness parameter. If a = 0 the blocks are completely
-
decoupled and each will exhibit the periodic behavior described above for a
single block. As a + the blocks become locked together and act as a sin-
gle block that again exhibits the periodic behavior described above. If P = 1
there is complete symmetry between the two blocks, P # 1 introduces an
asymmetry. In terms of the nondimensional variables (1 1.3) and parameters
(1 1.12), the sliding conditions (1 1.8) and (1 1.9) become
and the equations of motion (1 1.10) and (11.11) become

the two-block model. The
constant velocity driver
extends the springs until
sliding of a block
commences. In some cases
sliding of one block induces
the sliding of the second
block.
The blocks are expected to exhibit stick-slip behavior for c$ > 1. The first
block will begin sliding if (11.13) is satisfied, and the second block will be-
gin to slide if (11.14) is satisfied. Together (11.13) and (11.14) define a fail-
ure envelope in the Y, Y,-plane. The sliding behavior is governed by (1 1.15)
and (1 1.16). In some cases the sliding of one block induces the sliding of the
second block.
The solutions for the behavior of the blocks can be represented in a four-
dimensional phase plane consisting of Y,, Y2, dY,ld~,and dY,ld~.For sim-
plicity we consider the projection of the solution onto the Y, Y2 plane.
We first consider the symmetric case in which both blocks have the same
frictional behavior, that is, P = 1. An example with a = 3 and y = 1.25 is
given in Figure 11.4. The diagonal lines converging at Y, = Y2 = 1 are the
failure envelope given by (1 1.13) and (1 1.14). A periodic orbit is given by
abcd in Figure 11.4. At point a block 2 fails with Y, = 0.780 and Y2 = 0.835.
During the slip of block 2, block 1 remains fixed and the slip of block 2 is
represented by the vertical line ab in the Y,Y2 phase plane. The termination
of the sliding of block 2 is obtained from (11.16). Sliding of block 2 termi-
Figure 11.4. Behavior of the

symmetric p = 1 two-block
model illustrated in Figure
1 1 . 3 w i t h a = 3 a n d + = 1.25.
Cyclic behavior is obtained
with alternating slip of the
two blocks.
nates at point b where Y, = 0.780 and Y2 = 0.735. From point b to point c the
blocks stick and the springs extend due to the movement of the constant-ve-
locity driver. The increments in Y, and Y, are equal and the strain accumula-
tion phase is represented by the diagonal line bc, which has unit slope. The
termination of this strain occurs when this line intercepts the failure enve-
lope. This occurs at point c, where Y, = 0.865 and Y2 = 0.820. During the slip
of block 1, block 2 remains fixed and the slip of block 1 is represented by the
horizontal line cd. The termination of the sliding of block 1 is obtained from
(1 1.15). Sliding of block 1 terminates at point d, where Y, = 0.765 and Y2 =
0.820. Between points d and a the blocks stick and the springs again extend
at equal rates due to the movement of the constant-velocity driver. The incre-
ments in Y, and Y2 are equal and the strain accumulation phase is represented
by the diagonal line ad, which has unit slope. The termination of this strain-
accumulation phase occurs when this line intercepts the failure envelope.
This occurs at point a, and the cycle repeats. The behavior of this symmetri-
cal two-block model is periodic, with first one block sliding and then the
other.
We next consider an asymmetric case with p = 2.5, a = 3.49, and = +
1.25. The results are given in Figure 11.5. The behavior is similar to that
given in Figure 10.5 and is fully chaotic. The curves that fall outside the fail-
ure envelope are cases in which both blocks are sliding simultaneously.
When a diagonal strain-accumulation line intercepts the upper failure enve-
lope, block 2 begins to slide. Because P is relatively large, the failure force
Figure 11.5. Behavior of an

asymmetrical two-block
model with f3 = 2.25, a =
3.49, and 4 = 1.25. Chaotic
behavior is obtained.
for block 2 is considerably larger than the failure force for block 1. Thus the
vertical failure path for block 2 crosses the failure envelope of block 1 and it
begins to slide. The sliding of both blocks results in S-shaped curves. The
next strain-accumulation phase intercepts the lower failure envelope and
block 1 begins to slide. Because of the large force required to induce the fail-
ure of block 2, these horizontal failure paths do not cross the upper failure
envelope. The result is a chaotic sequence of failures of both blocks together,
followed by a failure of the weaker block 1 .
To study this behavior further a bifurcation diagram is given in Figure
11.6. The values of Y, - Y , at the termination of slip are given for various
+
values of a with p = 2.25 and = 1.25. A detailed illustration of the behav-
ior in the range 3.2 < a < 3.5 is given in Figure 11.7. Solutions evolve to an
asymptotic, large-time behavior independent of the initial conditions. As il-
lustrated in Figures 11.6 and l l .7, the system may evolve to limit cycle be-
havior or chaotic behavior. A series of period-doubling pitchfork bifurca-
tions are clearly illustrated in Figure 11.7. The cyclic behavior for the n = 2
limit cycle obtained for a= 3.25 is given in Figure 11.8. The cyclic behavior
for the n = 4 limit cycle obtained for a = 3.38 is given in Figure 11.9. These
limit cycles evolve into the type of chaotic behavior illustrated in Figure
11.5. The behavior of the asymmetric two-block model is remarkably similar
to that of the logistic map.
The values of the Lyapunov exponents corresponding to the points given
on the bifurcation diagram in Figure 11.6 are given in Figure 1 1.10. The win-
dows of chaotic behavior are clearly illustrated.

'I.
.4l::...l!:... diagram for an asymmetric
two-block model with f3 =
+
2.25 and = 1.25. The
values of Y, = Y, are given
as a function of a. Singular
point, limit cycle, and
chaotic types of behavior are
found.
A modification to the analysis of a pair of slider blocks is to allow only

one slider block to slip at a time. When a slider block becomes unstable, it is
allowed to move according to (1 1.6). The stability of the second slider block
is determined, and if it is unstable, it is allowed to move according to (1 1.6).
The primary advantage of this modification is that slider-block displace-
ments are given by an analytical expression (11.6) and numerical solutions
Figure 11.7. Details of the

bifurcation diagram given in
Figure 11.6 for the region
3.2 < a < 3.5. The pitchfork
bifurcation illustrated
evolves to deterministic
chaos.

asymmetric two-block model
with p = 2.25, a = 3.25, and
c$ = 1.25. An n = 2 limit
cycle is obtained.
of differential equations are not required. Studies of this modified slider-

block problem (Narkounskaia and Turcotte, 1992; Narkounskaia et al.,
1992) yield a variety of behaviors including periodic solutions, limit cycles,
and chaotic solutions.
Huang and Turcotte ( 1 9 9 0 ~ )have applied the chaotic behavior of the
asymmetric two-block system to two examples of interacting fault segments.
The Pacific plate descends beneath the Asian plate, resulting in the forma-

asymmetric two-block model
with p = 2.25, a = 3.38, and
+ = 1.25. Ann = 4 limit
cycle is obtained.

the Lyapunov exponent on
a corresponding to the
bifurcation diagram given in
Figure 11.6.
tion of the Nankai trough along the coast of southwestern Japan. The relative
motion between the plates has resulted in a sequence of great earthquakes
that have been documented through historical records for the period AD
684-1946. The sequence is marked by an irregular but somewhat repetitive
pattern in which whole section failures occur following several alternate
failures of single segments. In the two-block model the simultaneous slip of
both blocks corresponds to an earthquake that ruptures the entire section,
and single-block failures correspond to an earthquake on a single segment.
Taking P = 1.05 and a = 0.81, Huang and Turcotte ( 1 9 9 0 ~ found) chaotic
model behavior that strongly resembled the observed sequence of earth-
quakes in the Nankai trough.
Another example is the interaction between the Parkfield segment and
the rest of the south central locked segment of the San Andreas fault in Cali-
fornia. A sequence of magnitude-six earthquakes occurred on the Parkfield
segment in 1881, 1901, 1922, 1934, and 1966. The last great earthquake on
the locked segment to the south occurred in 1857 and is also associated with
a rupture on the Parkfield segment. Taking P = 2 and a = 1.2, Huang and
Turcotte ( 1 9 9 0 ~ )found chaotic model behavior similar to that described
above. A sequence of slip events on the weaker block often preceded the si-
multaneous slip of the weaker and stronger blocks. The model simulation
suggested two alternative scenarios for a great southern California earth-
quake following a sequence of Parkfield earthquakes. In the first case a
Parkfield earthquake will transfer sufficient stress to trigger the great south-
ern California earthquake; the Parkfield earthquake is thus essentially a fore-
shock for the great earthquake. In the second case a small additional strain
after a Parkfield earthquake will trigger an earthquake on the southern sec-
tion and this will result in an additional displacement on the Parkfield sec-
tion. The evolution of the system is chaotic: its evolution is not predictable
except in a statistical sense.
Spring-block models are a simple analogy to the behavior of faults in the
earth's crust. However, the chaotic behavior of low-dimensional analog sys-
tems often indicates that natural systems will also behave chaotically. Thus it
is reasonable to conclude that the interaction between faults that leads to the
fractal frequency-magnitude statistics discussed in Chapter 4 is an example
of deterministic chaos. The prediction of earthquakes is not possible in a de-
terministic sense. Only a probabilistic approach to the occurrence of earth-
quakes will be possible.
Problems
+
Problem 11.1. Consider a single slider block with = 1.5. (a) At what value
of Y does slip occur? (b) What is the value of Y after slip?
+
Problem 11.2. Consider a single slider block with = 3. (a) At what value of
Y does slip occur? (b) What is the value of Y after slip?
Problem 11.3. For a single slider block determine the dependence of V =
dYldT on Y during slip.
+
Problem 11.4. Consider a pair of slider blocks with a = 0, P = 1, and = 2.
Assume that initially Y, = 0.5, Y2 = 0. (a) What are the values of Y, and
Y, when block 1 first slips? (b) What are the values of Y, and Y, after
block 1 slips? (c) What are the values of Y, and Y2 when block 2 first
slips? (d) What are the values of Y, and Y2 after block 2 slips? (e) Draw
the behavior of the system in the Y, Y2 phase plane.
Problem 11.5. Consider a pair of slider blocks with a = 0, P = 1, and 4 = 413.
Assume that initially Y, = 0.75, Y2 = 0.5. (a) What are the values of Y,
and Y, when block 1 first slips? (b) What are the values of Y, and Y2 after
block 1 slips? (c) What are the values of Y, and Y, when block 2 first
slips? (d) What are the values of Y, and Y2 after block 2 slips? (e) Draw
the behavior of the system in the Y, Y2 phase plane.
EQUATIONS
Sets of coupled nonlinear differential equations can also yield solutions that
are examples of deterministic chaos. The classic example is the Lorenz
equations. Lorenz (1963) derived a set of three coupled total differential
equations as an approximation for thermal convection in a fluid layer heated
from below. He showed that the solutions in a particular parameter range had
exponential sensitivity to initial conditions and were thus an example of de-
terministic chaos. This was the first demonstration of chaotic behavior. The
Lorenz equations have been studied in detail by Sparrow (1982).
Because of their historical significance and because thermal convection
in the earth's mantle drives plate tectonics, we will consider the Lorenz
equations in some detail. When a fluid is heated its density generally de-
creases because of thermal expansion. We consider a fluid layer of thickness
h that is heated from below and cooled from above; the cool fluid near the
upper boundary is dense and the fluid near the lower boundary is light. This
situation is gravitationally unstable. The cool fluid tends to sink and the hot
fluid tends to rise. This is thermal convection.
Appropriate forms of the continuity, force balance, and energy balance
equations are required for a quantitative study of thermal convection. We
will restrict our attention to two-dimensional flows in which the velocities
are confined to the xy-plane. Continuity of fluid requires that
where u is the horizontal component of velocity in the x-direction and v is

the vertical component of velocity in the y-direction, measured downward
from the upper boundary. In writing (12.1) the density is assumed to be con-
stant. Force balances in the x and y directions require that
LORENZ EQUATIONS 257
where p is density, Ap the density difference, p pressure, CI, viscosity, and g

the acceleration due to gravity. The fractional density difference Aplp is as-
sumed to be small so that Ap can be neglected except in the buoyancy term
Ap g of the vertical force equation (12.3). This is known as the Boussinesq
approximation. The terms on the left-hand side of (12.2) and (12.3) are the
inertial forces associated with the acceleration of the fluid. The first terms on
the right-hand sides of the equations are the pressure forces and the second
terms are the viscous forces. The energy balance requires that
where T is temperature, k thermal conductivity, and cpis the specific heat at

constant pressure. The terms on the left-hand side of (12.4) account for the
convection of heat and the terms on the right-hand side account for the con-
duction of heat. A full derivation of these equations has been given by Tur-
cotte and Schubert (1982, pp. 240-74).
In the absence of convection, that is, u = v = 0, the temperature Tc is a
linear conduction profile given by
(12.5)
where T , is the constant temperature of the upper boundary ( y = 0) and T, is

the constant temperature of the lower boundary ( y = h). The incompressible
continuity equation in two dimensions (12.1) is satisfied if we introduce a
stream function I+!J defined by
It is also convenient to introduce the temperature difference 0 between the

actual temperature T and the temperature Tc in the absence of convection:
The density difference Ap in the buoyancy term of the vertical force the
equation (12.3) is related to this temperature difference by
258 LORENZ EQUATIONS
where a is the volumetric coefficient of thermal expansion.

Substitution of (12.6)-(12.8) into (12.1)-(12.4) gives
The problem has been reduced to the solution of two partial differential equa-
tions for q!~and 8.
To better understand the roles of various terms, it is appropriate to intro-
duce the nondimensional variables
where K = k/(pcp) is the thermal diffusivity. With these nondimensional vari-

ables two nondimensional parameters govern the behavior of the equations:
where Ra is the Rayleigh number and Pr is the Prandtl number. The Rayleigh
number is a measure of the strength of the buoyancy forces that drive con-
vection relative to the viscous forces that damp convection. The higher the
Rayleigh number the stronger the convection. The Prandtl number is the ra-
tio of the momentum diffusion to the heat diffusion. It is instructive to esti-
mate these two parameters for the earth's mantle. Due to solid-state creep,
the earth's mantle has a mean viscosity of around = 1021 Pa s, its thickness
is h = 2880 krn, and the temperature increase across it is estimated to be T2 -
TI = 3000 K. For the rock properties we taken K = 1 mm2s-1 and a = 3 X
10-5K-1. We assume g = 10 m s-2 and an average density p = 4000 kg m-3
and find Ra = 8.6 X lo6 and Pr = 2.5 X 1023, both very large values. The be-
havior of the earth's mantle will be discussed further in the next chapter.
We now return to the basic equations. Substitution of (12.11) to (12.13)
into (12.9) and (12.10) gives
The solution is determined by the two parameters Ra and Pr and the bound-
ary conditions. For small values of the Rayleigh number, the viscous forces
are sufficiently strong to prevent any flow. Thus there is a critical minimum
value of the Rayleigh number for the onset of thermal convection.
We next consider a linearized stability analysis for the onset of convec-
tion as given by Rayleigh (1916). Only terms linear in 8 and 3 are retained
and the marginal stability problem is solved by setting dlat = 0. Thus (12.14)
and (12.15) become
It is appropriate to assume solutions that are periodic in the horizontal coor-

dinate 3. Solutions that satisfy (12.17) are
275
t j= sin (T) sin
0= sin rrj
The flow is assumed to be periodic in the horizontal direction with a wave-

length A; the nondimensional wavelength is h = hlh. The flow consists of
counter-rotating cells; each cellular flow has width A12, as illustrated in Fig-
ure 12.1. The temperature boundary conditions e = 0 at 7 = 0, 1 are satisfied

by (12.19). The requirement of no flow through the walls corresponds to i =
= 0 at 7 = 0, 1 and these conditions are satisfied by (12.18). If the
boundaries of the layer are solid surfaces, we would require E = - a@ay = 0
at j = 0, 1. These are the no-slip conditions requiring that there be no relative
motion between a viscous fluid and a bounding solid surface at the solid-
fluid interface. If the boundaries of the layer are free surfaces, that is, if there
is nothing at the boundaries to exert a sheer stress on the fluid, we would re-
quire that the shear stress be zero at L = 0, 1. These free surface boundary
conditions can be written as aiilaj = -a2$/ay2 = 0 at 7 = 0, 1. The stream
function given in (12.18) satisfies the free surface boundary conditions.
In order that (12.18) and (12.19) also satisfy (12.16) we require that
When the nondimensional wavelength h is specified, this is the minimum

value of the Rayleigh number at which convection will occur. As the
Rayleigh number for a fluid layer is increased, flow will occur at the wave-
length for which the Rayleigh number given by (12.20) is a minimum. This
value of the Rayleigh number is given by
27 7r4
Ra, = ---- = 657.5
4
This is the critical Rayleigh number for the onset of thermal convection in a
fluid layer heated from below. At Rayleigh numbers less than that given by
(12.21), thermal convection will not occur. The nondimensional wavelength
corresponding to (12.21) is

two-dimensional cellular
convection in a fluid layer
heated from below.
This is the wavelength of the initial convective flow that takes the form of
counter-rotating, two-dimensional cells. Each cell has a width 21/2h, one-half
the wavelength of the initial disturbance.
A pitchfork bifurcation occurs at the critical Rayleigh number. If Ra <
Rac the only solution is the conduction solution, which is stable. Above the
critical Rayleigh number the conduction solution remains a solution of the
governing equations but it is now unstable. Above the critical Rayleigh num-
ber there are two stable convective solutions corresponding to cellular rolls
rotating either clockwise or counterclockwise. This is identical to the pitch-
fork bifurcation illustrated in Figure 9.4(b).
Because the applicable equations are linear, the stability analysis does
not predict the amplitude of the convection. It is not possible to specify the
value of $, in (12.18) and (12.19). To determine the amplitude of the thermal
convection, it is necessary to retain nonlinear terms. One approach to the so-
lutions of the full nonlinear equations (12.14) and (12.15) is to expand the
variables and 8 in double Fourier series in i and j with coefficients that
are functions of time. Lorenz (1963) strongly truncated these series and re-
tained only three terms of the form
1
$ = j q (4 + i 2 ) ~ ( rsin
) (T)
2 6
sin ~7
7r3
= -(4
4~~
+ X2l3 C ( I ) sin 27rJ - 21i2Z3(r) cos (9)
2 -
sin ~ j ] (12.24)
where
with Rac given by (12.20). These equations satisfy the same set of boundary
conditions that (12.18) and (12.19) satisfy. The expansion of the stream
function (12.23) is essentially identical to the form used in the linear stabil-
ity analysis (12.18). However, the expansion of the temperature (12.24) in-
cludes an additional term that is not dependent on i .
It is necessary to derive differential equations for the time dependence of
the coefficients A(T), B(r), C(T). This is done by substituting the expansions
(12.23) and (12.24) into the governing equations (12.14) and (12.15). Coef-
ficients of the Fourier terms are equated to obtain the necessary equations.
All nonlinear terms in the stream function equation (12.14) are ne-
glected. When (12.23) and (12.24) are substituted into this equation the re-
sult is
As expected this is a linear equation and no further approximations have

been made in writing it. To maintain a consistent approximation in the en-
ergy equation (12.15), it is necessary to retain several nonlinear terms. Sub-
stitution of (12.23) and (12.24) into (12.15) gives several terms that are not
in the form required. These products of trigonometric functions are reduced
to the standard form using the multiple angle formulas
sin .rrj cos (271~1%

cos (271~1%) ) cos T%= sin 2- (12.29)
with higher-order terms neglected.

Equating coefficients gives
where
The three first-order total differential equations (12.27), (12.30), and (12.31)
are the Lorenz equations. These equations would be expected to give accu-
rate solutions to the full equations when the Rayleigh number is slightly su-
percritical, but large errors would be expected for strong convection because
of the extreme truncation.
Solutions of the Lorenz equations represent cellular, two-dimensional
convection. Because only one term is retained in the expansion of the stream
function, the particle paths are closed and represent streamlines even when
the flow is unsteady. The time dependence of the coefficient A determines
the velocity of a fluid particle. But the fluid particle follows the same closed
trajectory independent of its time variation. The coefficient B represents
temperature variations associated with the stream function mode A. The co-
efficient C represents a horizontally averaged temperature mode. A detailed

discussion of the behavior of the Lorenz equations has been given by Spar-
row (1982).
To examine the behavior of the Lorenz equations, we first determine the
allowed steady-state solution. Obviously the steady-state solution A = B =
C = 0 corresponds to heat conduction without flow. An additional pair of al-
lowed solutions is
These solutions correspond to an infinite set of two-dimensional cells as il-

lustrated in Figure 12.1. Adjacent cells rotate in opposite directions; the
choice of sign given in (12.33) determines whether a specified cell rotates
clockwise or counterclockwise. A stability analysis for the conduction solu-
tions shows that it is stable for r < 1 and unstable for r > 1. Thus the Lorenz
equations exhibit the same type of pitchfork bifurcation at r = 1 (Ra = Rat)
that the full equations do. This is expected since the linearized form of the
Lorenz equations is identical to the linearized form of the full equations.
The stability of the steady solution given in (12.33) and (12.34) can also
be examined. Expanding about this solution with
and substituting into (12.27), (12.30), and (12.31) with linearization gives
the characteristic equation
A3 + (Pr + b + l)A2 + ( r + Pr)bA + 2b Pr(r - 1) = 0 (12.38)
This equation has one real negative root and two complex conjugate roots
when r > 1. If the product of the coefficients of A2 and A equals the constant
term we obtain
At this value of r the complex roots of (12.38) have a transition from nega-
tive to positive real parts. This is the critical value of r for the instability of
steady convection and represents a subcritical Hopf bifurcation. If Pr > b + 1
the steady solutions given by (12.33) and (12.34) are unstable for Rayleigh
numbers larger than those given by (12.39).
To examine further the behavior of the Lorenz equations it is necessary
to carry out numerical solutions. Following Lorenz (1963)we consider h, =
i.
8112, the critical value from (12.22), so that b = For these values the steady-
state solution given by (12.23), (12.24), (12.33), and (12.34) becomes
= 2 [24(r - I)]''~sin
(Z)
- sin 71.y
3 = E 4{ ( r - I) sin 2?rj
[y
T -
(
r - 1)
21/2
cOs (.Dli) }
Ty (12.41)
This steady-state solution is valid if r > 1 and is less than the critical value
given by (12.39).
As the Rayleigh number increases above r = 1, the strength of the con-
vection increases, as indicated by (12.40). This results in larger transport of
heat by convection, and as a result the thermal gradients at the upper and
lower boundaries increase. The Nusselt number is a measure of the effi-
ciency of the convective heat transfer across the layer. The Nusselt number
Nu is the ratio of the heat transferred by convection to the conductive value
without convection. In terms of our nondimensional variables it is given by
where ( ), indicates an average across either the upper or lower surface.

These averages must be equal since the mean heat flux into the layer across
the lower boundary must be equal to the mean heat flux out of the layer
through the upper boundary. For the steady-state solution given by (12.41)
the Nusselt number is
This relation is not in good agreement with experiment when r is signifi-

cantly larger than unity. This is clearly due to the extreme nature of the trun-
cation, which is expected to be valid only near the stability limit r = 1. Nev-
ertheless, it is of interest to explore the behavior of the Lorenz equations for
larger values of r.
The critical Rayleigh number for stability of the steady-state solution is
r = 24.74 for h = 8112 from (12.39). For values of r greater than this, unsteady
LOREN2 EQUATIONS 265
solutions are exposed. A numerical solution of the Lorenz equations for r =

28 is given in Figure 12.2. The time dependence of the three variables A(T),
B(T), C(T) is obtained. It is convenient to study the solution in the three-
dimensional ABC phase space; the time T is a parameter. The projection of
the solution onto the BA-plane is given in Figure 12.2(a) and the projection
onto the BC-plane is given in Figure 12.2(b). These are known as phase por-
traits. The dependence of the variable B on time is given in Figure 12.2(c).
The solution randomly oscillates between cellular rolls with clockwise rota-
tion for B > 0 and with counterclockwise rotation for B < 0. The unstable
fixed points from (12.33) and (12.34), A = B = 272'12, C = 27, are the crosses
in Figures 12.2(a) and 12.2(b).
This solution is chaotic in that adjacent solutions diverge exponentially in
time. The solution oscillates about a fixed point with growing amplitude until
it flips into the other quadrant, where it oscillates about the other fixed point.
The fixed points behave as "strange attractors." In Figure 12.2(d) the fixed
points are projected onto the rB-plane. The solid lines represent stable fixed
points and the dashed lines represent unstable fixed points; the solid circle is
a pitchfork bifurcation and the open circles are Hopf bifurcations. The pitch-
fork bifurcation at r = 1 corresponds to the onset of thermal convection; the
Hopf bifurcations at r = 24.74 correspond to the onset of chaotic flows.
An essential feature of the solution illustrated in Figure 12.2(b) is that
the value of C is always positive and oscillates aperiodically around a posi-
tive value. It is the C term that gives an approximation to a thermal boundary
layer structure. The growth and decay of C around a positive value is an ap-
proximation to the growth and separation of the thermal boundary layers at
the top and bottom of the convection cell. It is the buoyancy forces in the
boundary layers that drive the flow. The direction of the trajectory near B = 0
is always toward C = 0; the direction of the trajectory for large B is always in
the positive C-direction. When C is at its largest the thermal boundary layers
are thin, the buoyancy forces are small, and the flow decelerates (B de-
creases). This quiescence of the convective flow allows the thermal bound-
ary layers to thicken and C decreases. The thickening thermal boundary lay-
ers become unstable and the resultant buoyancy accelerates the flow,
increasing B; depending on where the instability occurs, either clockwise or
counterclockwise flow results. As the convective motion increases, the ther-
mal boundary layers again thin, and the value of C grows. The amplitude of
the convective motion consequently slows down, and the absolute values of
A and B once again move toward zero. Once convective motion has slowed
enough, the boundary layer can grow again. Every time the convective mo-
tion stops to let the thermal boundary layer grow, the cell is presented with a
choice of whether to convect in a clockwise or a counterclockwisedirection.
It is this freedom of choice in a completely deterministic system that pro-
duces the chaotic, aperiodic behavior of the Lorenz attractor.
The essential feature of the solution of the Lorenz equations in this param-
eter range is deterministic chaos. One consequence of the deterministic
chaos of the Lorenz equations is that solutions that begin a small distance
apart in phase space diverge exponentially. With essentially infinite sensitiv-
Figure 12.2. A numerical

solution of the Lorenz
equations (12.27), (12.30).
(12.31) with Pr = 10, b = 813,
r = 28. The solution in the
ABC phase space is shown
projected (a) into the BA-
plane and (b) onto the BC-
plane.
40 60
time
(c)
Figure 12.2. (conr.) (c) Time

dependence of the coefficient
B. (d) The loci of the fixed
points are projected onto the
&plane. The solid lines
represent stable fixed points,
the broken lines represent
unstable fixed points, the
solid circle is a pitchfork
bifurcation, and the open
circles are Hopf bifurcations.
268 LOREN2 EQUATIONS
ity to initial conditions, the zero-order behavior of a solution is not pre-

dictable.
It should be emphasized that the numerical solution of the Lorenz equa-
tions given above is not a valid solution for thermal convection in a fluid
layer. The Rayleigh number is outside the range of validity of the Lorenz
truncation. Nevertheless, this solution has a great significance in that it was
the first solution to exhibit the mathematical conditions for chaotic behavior.
But the significance goes beyond this. Experimental studies and numerical
simulations of thermal convection in a fluid layer heated from below at high
Rayleigh numbers and intermediate values of Prandtl number are generally
unsteady and "turbulent." There is good reason to accept that the solutions to
the full set of equations are also chaotic in this range of parameters. Thus
weather is taken to be chaotic, and deterministic predictions are not possible.
It is essential to treat weather as a statistical problem with uncertainties in-
creasing with forward extrapolations.
Problems
Problem 12.1. Show that the critical Rayleigh number given by (12.20) has
the minimum value as given by (12.21) and (12.22).
Problem 12.2. For the steady-state solution of the Lorenz equations given in
(12.40) and (12.41), determine an expression for the mean horizontal ve-
locities on the boundaries at y = 0, 1.
ChapterThirteen
IS MANTLE
CONVECTION
CHAOTIC?
The Lorenz equations are a low-order expansion of the full equations applic-
able to thermal convection in a fluid layer heated from below. For the range
of parameters in which chaotic behavior is obtained, the low-order expan-
sion is not valid; higher-order terms should be retained. Nevertheless, the
chaotic behavior of the low-order analog is taken as a strong indication that
the full equations will also yield chaotic solutions. Numerical solutions of
the full equations are strongly time dependent for high Rayleigh number
flows; these solutions appear to be turbulent or chaotic.
It is generally accepted that thermal convection is the primary means of
heat transport in the earth's mantle. Heat is produced in the mantle due to the
decay of the radioactive isotopes of uranium, thorium, and potassium. Heat
is also lost due to the cooling of the earth. The surface plates of plate tecton-
ics are the thermal boundary layers of mantle convection cells. The plates
are created by ascending mantle flows at ocean ridges. The plates become
gravitationally unstable and founder into the mantle at ocean trenches (sub-
duction zones). Intraplate hot spots such as Hawaii are attributed to mantle
plumes that ascend from the hot unstable thermal boundary layer at the base
of the convection mantle.
An important question with regard to the earth is whether mantle con-
vection is chaotic. The earth's solid mantle behaves as a fluid on geological
time scales because of thermally activated creep. The discussion in the pre-
vious chapter considered only a constant viscosity. This is a poor approxima-
tion for the earth's mantle because the dependence of strain on stress is al-
most certainly nonlinear and is an exponential function of temperature and
pressure. Also, the Boussinesq approximation is not applicable because of
the significant increase in density with depth (i.e., pressure). Nevertheless,
calculations assuming a linear stress-strain relation and constant fluid prop-
erties can provide important insights. In the previous chapter we estimated
that the Rayleigh number for mantle convection is near 107 and the Prandtl
number is larger than 1023. The latter is such a large value that it is appropri-
ate to assume that the mantle has an infinite Prandtl number. Because the
270 IS MANTLE CONVECTION CHAOTIC?
Prandtl number for the mantle is so large, the momentum terms on the left-
hand side of the momentum equations (12.2) and (12.3) can be neglected.
Thus the only nonlinear terms are those in the energy equation (12.4). The
question is whether these terms can generate chaotic behavior and thermal
turbulence.
The first question we address is whether the Lorenz equations yield
chaotic solutions in the limit Pr -. In this limit (12.27) requires
The substitution of this result into (12.30) and (12.31) gives
Again these equations have the steady-state, d / d = ~ 0, solution A = B = C = 0

corresponding to conduction. These equations also have the same fixed
points as the Lorenz equations; these fixed points are given in (12.33) and
(12.34) and correspond to cellular rolls rotating either in the clockwise or
counterclockwise directions. A stability analysis again shows that the con-
duction solution is stable for r < 1 and unstable for r > 1. However, the
steady solution consisting of cellular rolls is now stable for the entire range
r > 1. In the limit of infinite Prandtl number the Lorenz equations do not
yield chaotic solutions. This has been taken as evidence that mantle convec-
tion is not chaotic.
To study this problem further Stewart and Turcotte (1989) considered a
3
higher-order expansion than the Lorenz equations. The variables and 8-are
expanded in double Fourier series in F and j. The coefficients qm, and Om,n
represent the terms sin (2rrrnilh) or cos (2-/h) and sin ( m y ) in the ex-
- the left-hand
pansions. In the limit of infinite Prandtl number, - side of (12.14)
is zero, resulting in a linear equation between @m,n and Om,, that can be written
This result is then substituted into (12.15). The lowest consistent order of
truncation beyond that used by Lorenz is m = 2 for the expansion in x (m = 0,
1,2) and n = 4 for the expansion in y (n = 1,2,3,4). This truncation yields a
set of 12 ordinary differential equations for the time dependence of the tem-
perature coefficients,., that can be written
IS MANTLE CONVECTION CHAOTIC? 271
It is necessary to take the resolution in the vertical direction twice compared with
the horizontal direction to resolve the convection terms in the energy equation.
,,,, ,,,, ,,,,
The time evolution of the 12 coefficients 8 8 8,,,, 8,,,, 8 8,,,, el,,,
8 a,,,, 8,,,, 8,,? is found by integrating numerically the 12 equations
given by (13.5). The time evolution can be thought of as trajectories in a 12-
dimensional phase space. It is convenient to project the 12-dimensional tra-
jectories onto the two-dimensional phase space consisting of 8,,, and 8,,,;
these correspond to the fundamental mode and the first subharmonic. There
are two parameters in this problem, the Rayleigh number, Ra or r, and the
wavelength. In this discussion solutions are given only for the critical value
of the wavelength h = 2312.
At the subcritical Rayleigh numbers 0 < r < 1 (0 C Ra < 657.512), the
only fixed point of the solution is at the origin and it is stable; there is no
flow. For higher Rayleigh numbers, the two fixed points corresponding to
clockwise and counterclockwise rotations in the fundamental model o , , , be-
Table 13.1. Numerical values of Fourier coefficients of the fixed points of

the 12-mode equations (13.5)
come stable. The steady-state solution for Ra = lo4 (r = 15.21) is given in

Table
-
13.1. It is seen that only six of the 12 coefficients are nonzero: go,?,
O,,, a,,,, 8 j,,,, and 0,:,. This solution was obtained by specifying an in;-
tial condition near the origin and studying the time evolution-of - the 12 coef-
ficients using (13.5). This time evolution projected onto the 02JOl,l-planeis
given in Figure 13.l(a). Although the subharmonic coefficient O,,, is zero at
the fixed point, it is nonzero during the time evolution.
The steady-state solution for Ra = 3 X 104 (r = 45.62) is also given in
Table
- 13.1. It is seen that only four of the 12 coefficients are nonzero: e,,?,
O,,, , , ! . At this Rayleigh number the fundamental mode and its associ-
ated coefficients 8,,,, 8,,,, and 8,, are zero at the stable fixed point. The
time evolution of the solution projected onto the ~ , , 8 , , , - p l a nis
e given in
Figure 13.l(b). Finally, the steady-state solution for Ra = 4.3 X 104 (r =
65.39) is given- in Table
- - 13.1.
- It -is seen
- that
- eight
- of the 12 -coefficients are
now nonzero: O,,,, O,,, O,,,, O,,,, O,,,, O,,,,&,,, 19,,~.All of the Ole coefficients
are zero including the fundamental mode O,,,.The time evolution of the solu-
tion projected onto the 8,,,e,,,-plane is given in Figure 13.l(c). It is seen that
the evolution prior to entering the stable - fixed point is much more complex;
the solution oscillates in the positive O,., quadrants before entering the nega-
tive quadrants.
The time evolution of the solution for Ra = 4.5 X lo4(r = 68.44) is given
in Figure 13.l(d); it is fully chaotic and no fixed points are stable. The flow
alternates between aperiodic oscillations about the two fundamental modes
(clockwise and counterclockwise) and the two subharmonic modes (clock-
wise and counterclockwise). All 12 coefficients are nonzero and are time de-
pendent. The time dependences of the a,,, al,,
and coefficients are shown in
Figure 13.2. The resemblance to the time behavior of the Lorenz attractor il-
lustrated in Figure 12.2(c) is striking. Oscillatory behavior of the 8,,, mode
amplifies until the flow undergoes a turbulent burst in the fundamental el,,
mode, where it is briefly trapped before flipping into the 8,,, mode with the
opposite
- - sense of rotation.
To better understand the transitions in the behavior of the time-depen-
dent solutions, Figure 13.3 gives two projections of the loci of fixed points
as a function of the Rayleigh number of the system. The solid lines denote
stable fixed points, the dashed lines denote unstable fixed points, and the
open circles denote Hopf bifurcations. Figure 13.3(a) shows a projection
onto the ~a,8,,,-plane. The origin representing the conduction solution be-
comes unstable and bifurcates at Ra = 657.5 12, giving two stable symmetric
solutions that do not contain the o,,, mode. One branch of this solution ap-
pears in the positive quadrant and is labeled "8,,,(pure)" in Figure 13.3(a) to
distinguish it from the mixed-mode solution, which contains a contribution
from the 8,,, mode. Each branch becomes unstable and undergoes a subcriti-
cal pitchfork bifurcation at Ra = 3.802 X 104, producing four unstable
mixed-mode solutions, labeled "a,,,(mixed)". Each 8,,,mixed-mode branch
sweeps back to a saddle bifurcation at Ra = 1.909 X lo4.This type of bifur-

cation configuration (subcritical pitchfork plus two saddles) typically pro-
duces hysteresis effects when the saddle has one stable branch and one un-
stable branch. Here, the %,,,(mixed)solution has one stable manifold (out of
12) on one side of the saddle, and two unstable manifolds on the other.
The second bifurcation of the conduction solution is at Ra = 1315.023,
where two unstable symmetric fixed points dominant in the subharmonic 8,,,
mode appear. Since these fixed points contain no component in the 8,,,
mode, we call these unstable solution branches. Each of these
branches becomes stable and undergoes a pitchfork bifurcation at Ra =
2041.918, resulting in the branching solution labeled "%2,,(mixed)'' in Fig-
ures 13.3(a) and 13.3b(b). The 8,,, mixed-mode branches nearly connect
with the %,,,mixed-mode branches.
Both the fundamental and the subharmonic pure-mode solutions are sta-
ble between Ra = 2369 and Ra = 3.802 X lo4. The trajectories in Figures
13.1(a) and 13.1(b) have the same initial condition, yet the trajectory in Fig-
ure 13.l(a) converges to the fundamental subharmonic stable fixed point.
Presumably this is because the unstable mixed-mode branches disrupt the
separatrix between the basin of the attraction of the fundamental and subhar-
monk pure-mode solutions. Note that the transition from Figure 13.l(a) to
Figure 13.1(b) occurs at a Rayleigh number above the stability limit of the
fundamental mode.
The third bifurcation of the conduction solution is at Ra = 4.140 X 104,
where two unstable symmetric solutions in the fundamental mode 8,,, ap-
pear. These are labeled "new 8,.,(pure)" in Figure 13.3(a). Each of these un-
dergo Hopf bifurcations at Ra = 5.23 X lo4 and Ra = 5.53 X 104. At no point
does the origin itself undergo a Hopf bifurcation, nor does the conduction
solution bifurcate to mixed-mode solutions.
No Hopf bifurcations were detected for the conduction or fundamental
harmonic solutions; however, the stable subharmonic branch undergoes two
Hopf bifurcations (Figure 13.3(b)), one at Ra = 4.37 X lo4 and one at Ra =
6.36 X 104. Each mixed-mode saddle (%,,,(mixed))undergoes two Hopf bi-
furcations, at Ra = 4.491 X 104 and at Ra = 5.039 X 104. Each mixed-mode
saddle (j,,,(mixed)) undergoes six Hopf bifurcations. Each of these Hopf bi-
furcations sheds stable or unstable periodic orbits that are responsible for the
oscillations of the trajectory at Ra = 4.5 X lo4 shown in Figure 13.l(d). In
- 7000 points of the trajectory at Ra = 4.5 X 104 are pro-
Figure 13.4, the-first
jected onto the 8,,,8,,,-plane (dotted line) and are shown superimposed on the
central portion of the branches of the fixed points (solid lines). Note that the tra-
jectory weaves aperiodically around several Hopf bifurcations (open circles).
Physically, infinite Prandtl number, high Rayleigh number convection
becomes time dependent through boundary layer instabilities that generate
thermal plumes. In terms of spectral expansions, these instabilities result
from the nonlinear coupling in the convective terms of the heat equation.
Rayleigh Number = 40000
Figure 13.1. Numerical

solutions of the 12-mode
infinite Prandtl number
equations projected onto
thea,,, 8,,,-plane of the
12-dimensional phase space.
(a) Ra = 10 000(r = 15.21),
(b) Ra = 30 000(r = 45.62),
--
r
I--
0 --
0
'-4000.0 - 2000.0 0.0 2000.0 4000.0

Figure 13.1. (cont.)
4.1
(c) Ra = 43 000(r = 65.39),
(4 (d) Ra = 45 000(r = 68.44).
Figure 13.2. Time

dependences of the
coefficients (a)a,,, and (b)
8,., for the solution given in
Figure l3.l(d); Ra = 45 000.
_ 'V
'
/
d
-'-- ' -- new I
M e - - - - - - - - --- - ' 4 8
$ 1 (pure
'--
Q2, (mixed)
I
\
-.
%
RAYLElGH NUMBER (x 10')
(a)

diagram for the 12-mode
infinite Prandtl number
equations. The fixed points
of (13.5) are projected (a)
onto the Ra, a,,,-plane,(b)
onto the Ra, a,,,-plane.
Stable branches are shown as
solid lines, unstable branches
20 4.0
as dashed lines, pitchfork
RAYLEIGH NUMBER (x 1P) bifurcations as solid circles,
and Hopf bifurcations as
(b) open circles.
Figure 13.4. Loci of the

fixed points from Figure
13.3 projected onto the
8,,,8,,,-plane.Superimposed
as a dotted line is the time
evolution of the chaotic
solution from Figure 13.l(d).
Problems
Problem 13.1. Show that the temperature coefficient B in the Lorenz expan-
sion (12.24) is related to TI,,, by
For Ra = 104 and h = 8112 compare the value of B from the three-mode
(Lorenz) expansion with the value from the 12-mode expansion.
Problem 13.2. Show that the temperature coefficient C in the Lorenz expan-
sion (12.24) is related to %o,2 by
A
For Ra = 104 and = 8112 compare the value of C from the three-mode
(Lorenz) expansion with the value from the 12-mode expansion.
Chapter Fourteen
RIKITAKE DYNAMO
The earth's magnetic field is attributed to the electrically conducting outer

core, which acts as a dynamo. The liquid outer core is primarily composed of
iron, which is an excellent electrical conductor at core conditions. Electrical
currents in the core generate a magnetic field. Buoyancy forces in the core,
due to either temperature or composition, drive a fluid flow. The flowing
electrical conductor in the magnetic field induces an electric field. This is a
self-excited dynamo.
Paleornagnetism is the study of the earth's past magnetic field from the
records preserved in magnetized rocks. Rocks containing small amounts of
ferromagnetic minerals such as magnetite and hematite can acquire a weak
permanent magnetism when they are formed. This fossil magnetism in a
rock is referred to as natural remanent magnetism.
Many volcanic rocks at the surface of the earth can be magnetized be-
cause of the presence of minerals such as magnetite. When these volcanic
rocks were cooled through the Curie temperature, they acquired a permanent
magnetism from the earth's field at the time of cooling. Paleomagnetic stud-
ies of remanent magnetism have provided a variety of remarkable conclu-
sions. These studies have traced the movement of the rocks due to plate tec-
tonics and continental drift over periods of hundreds of millions of years.
They have shown that the magnetic field at the surface of the earth has been
primarily a dipole, as it is today, and has remained nearly aligned to the
earth's axis of rotation. These studies have also shown that the earth's mag-
Figure 14.1. Observed
netic field has been subject to random reversals in which the north magnetic
polarity of the earth's
pole becomes the south magnetic pole and vice versa. The observed polari- magnetic field for the last 10
ties of the earth's magnetic field for the last 10 million years (Myr) are given Myr; the scale is in Myr
in Figure 14.1. Measurements indicate that for the last 720,000 years the before the present (BP). The
solid bands are the normal
(present) polarity and the
open bands are reversed
polarity. The last polarity
reversal occurred 720,000
years ago.
280 RlKlTAKE DYNAMO
magnetic field has been in its present (normal) orientation; between 0.72 and
2.5 Myr ago there was a period during which the orientation of the field was
predominantly reversed. Clearly one characteristic of the core dynamo is
that it is subject to spontaneous reversals.
A question that can be asked is whether the reversals give a fractal distri-
bution of polarity intervals. This question has been addressed by Gaffin
(1989) and his results are given in Figure 14.2. The number of polarity inter-
vals N ( T ) of length greater than T is given as a function of T. For intervals
between 300,000 yr and 50 Myr a good correlation with the fractal relation
(2.6) is obtained taking D = 1.43.
No detailed theory exists for the behavior of the core dynamo. The vis-
cosity of the liquid outer core is sufficiently small that the flow is undoubt-
edly turbulent. Thus the patterns of flow, electrical currents, and magnetic
fields are very complex. Because of this complexity, relatively simple disk
dynamos have been proposed as analog models. Rikitake (1958) proposed
the symmetric two-disk dynamo illustrated in Figure 14.3. It is composed of
two symmetric disk dynamos in which the current produced by one dynamo
energizes the other. Equal torques G are applied to the two dynamos in order
to overcome ohmic losses. Rikitake (1958) found that this dynamo was sub-
ject to random reversals of the magnetic field, but it was much later (Cook
and Roberts, 1970) that it was demonstrated that the Rikitake dynamo be-
haved in a chaotic manner.

number of reversal polarity
intervals N ( T ) with length
greater than T as a function
of T as given by Gaffin
(1989).The Cretaceous
superchron has been
excluded. The correlation
line is the fractal relation
(2.6)with D = 1.43.
RlKlTAKE DYNAMO 281
The steady-state behavior of the Rikitake dynamo is relatively easy to

understand, as illustrated in Figure 14.3. The current I, passes through the
current loop on the right in a clockwise direction. This current loop gener-
ates a magnetic field B, that is positive in the downward direction. This mag-
netic field also passes through the rotating, electrically conducting disk on
the right. Because of the applied torque G, the disk is rotating in the counter-
clockwise direction with an angular velocity R,. The induced electric field
in the disk Ei = u X B is the inward radial direction as illustrated. This in-
duced electrical field drives the electric current I,. In the steady state the ap-
plicable circuit equation is
where R is the resistance in either circuit and M is the mutual inductance be-
tween the current loops and the electrically conducting disks.
An electrical current I in a magnetic field B results in the electromotive
force F = I X B per unit length of current path. The interaction between the
magnetic field B, and the radially inward electric current I, results in a
torque in the clockwise direction. In the steady state this torque balances the
applied torque G and is given by
The discussion given above is also applicable to the current loop and rotat-
ing disk on the left in Figure 14.3.

the Rikitake two-disk
dynamo. The applied torques
G drive the shafts and
electrically conducting
disks to rotate in the
counterclockwise direction.
The currents in the coils
generate magnetic fields that
interact with the rotating
disks to generate the
currents.
282 RlKlTAKE DYNAMO
The equations governing the unsteady behavior of the two-disk Rikitake

dynamo are
where L is the self-inductance in either circuit and C the moment of inertia of

each dynamo. Subtracting (14.5) and (14.6) and integrating gives
where ROis a constant. This can replace either (14.5) or (14.6).

Again it is appropriate to introduce nondimensional variables and para-
meters according to
~'2
y2 = ( g ) 1 / 2 R 2 A
, = ( g ) 1 / 2 R o ,A
, = (---)
GLM
1/2
Substitution into (14.3)-(14.5) and (14.7) gives

RlKlTAKE DYNAMO 283
This is a set of three coupled, nonlinear differential equations that determine

the time evolution of the Rikitake dynamo. Setting the time derivatives equal
to zero the steady-state solutions are obtained:
where
The plus and minus signs refer respectively to the normal and reversed states
of the magnetic field.
Stability calculations (Cook and Roberts, 1970) have shown that the sin-
gular points given above are unstable for all parameter values. Their numeri-
cal solutions for p. = 1 and K = 2 are given in Figures 14.4 and 14.5. The sin-
Figure 14.4. Numerical

solution of the Rikitake two-
disk dynamo equations (14.9)-
(14.11) in the X,X,Y, phase
space projected onto the
X,Y, phase plane. The singular
points corresponding to
normal and reversed
polarizations of the magnetic
field, X , = 50.5, Y , = 4
are shown.
284 RlKlTAKE DYNAMO
gular points X2 = ?;, Y, = 4 are shown in the X2Y, phase plane illustrated in
Figure 14.4. The strange-attractor behavior of the solution is very similar to
that of the Lorenz equations given in Figures 12.l(a), (b). The time evolution
of the solutions, given in Figure 14.5, is also similar to that of the Lorenz
equations given in Figure 12.l(c). Oscillations grow in one polarity of the
field until it flips into the other polarity.
Extensive studies of the behavior of the Rikitake dynamo equations have
been carried out by Ito (1980) and by Hoshi and Kono (1988). The behavior
is found to be periodic or chaotic. A map of these two behaviors in the K - p
parameter space is given in Figure 14.6. The transition from periodic to
chaotic behavior follows the period-doubling route to chaos previously ob-
tained for the logistic map. A bifurcation diagram for K = 2 illustrating the
period doubling is given in Figure 14.7. Marzocchi et al. (1995) have studied
the reversal statistics of the Rikitake dynamo and found that they are not
similar to the reversals of the earth's magnetic field. This is not surprising
since the Rikitake dynamo is a low-order system and the earth's dynamo
must be a very high-order system.
Figure 14.5. Time evolution

of X, and X, for the solution
given in Figure 14.4.
RlKlTAKE DYNAMO 285
An alternative low-order model for the generation of the earth's mag-

netic field was proposed by Robbins (1977). The governing equations take
the form
CHAOS
Figure 14.6. Map of the

K - p parameter space
giving regions in which
periodic and chaotic
behaviors are found (Ito,
1980).

diagram illustrating the
period doubling route to
chaos of the Rikitake
dynamo equations as the
parameter p, is varied with
K = 2 (Ito, 1980).
286 RlKlTAKE DYNAMO
where a, R, and v are adjustable parameters. Chillingworth and Holmes

(1980) have obtained extensive solutions to these equations and found be-
havior similar to that of the Rikitake dynamo. Cortini and Barton (1994)
have determined the correlation dimensions from the reversal sequences of
the Rikitake and Chillingworth-Holmes models and have compared the re-
sults with values obtained from the actual reversal sequence.
Krause and Schmide (1988) have proposed a third-order recursive map
as a model for the reversals for the earth's magnetic field
The fixed points of this equation are x, = 0 and x, = 2 (1 - a-1)lt2.The fixed

point at x, = 0 is stable for 0 < a < 1 and unstable for a > 1. The fixed point
at x, = 2 (1 - a-l)"2 is real only for a > 1; it is stable for 1 < a < 2 and un-
stable for a > 2. If 0 < a < 3 f i 1 2 , values of x do not change sign and the
behavior is similar to that of the logistic map considered in Chapter 10. If
a > 3, the map is not confined and iterations cannot be carried out. In the
range 3 f i 1 2 < a < 3 reversals in the sign of x can occur. Chaotic oscilla-
tions occur within the range -2013 fi < x < 2 a 1 3 f i .
As an illustration of the behavior of this iterative map with reversals we
consider an example with a = 2.75 as shown in Figure 14.8. The curve repre-
sents the function
Taking x, = 0.2 we draw a vertical line; its intersection withflx) gives x, =

0.528. A horizontal line drawn from this intersection to the diagonal line of
unit slope transfers x ~ to+ x,,.
~ A vertical line is drawn to flx) giving x, =
1.0472056. Because x > 1 the next iteration results in a change of sign (a re-
versal) and we obtain x, = 0.2783043. Further iterations give x, =
-0.7060590 and x, = - 0.9737061 as illustrated. One hundred iterations of
this map are given in Figure 14.9. The sequence of chaotic reversals is
clearly illustrated.
The models considered above are clearly gross simplifications of the
complex fluid flows that occur in the earth's core. Nevertheless, the models
produce patterns of random reversals that are remarkably similar to the re-
RlKlTAKE DYNAMO 287
versals of the earth's magnetic field. Again this can be taken as evidence that
the dynamo action in the core is chaotic. It is certainly desirable to consider
higher-order systems that better simulate the "turbulent" interactions be-
tween electrical currents and flows of the electrically conducting fluid. A
start in this direction has been given by Glatzmaier and Roberts (1995).

of the iteration of the
third-order recursive map
(14.21) with a = 2.75 and
x, = 0.2. A reversal in sign
occurs on the third iteration.
yo
1
of100 iterations of the
third-order recursive map
(14.21) with a = 2.75 and
x, = 0.2. The sequence of
chaotic reversals is clearly
shown.
288 RlKlTAKE DYNAMO
Problems
Problem 14.1. Consider a single-disk dynamo. Determine the steady-state

current I and angular velocity i2 in terms of resistance R, mutual induc-
tance M, and applied torque G.
Problem 14.2. Determine the fixed points of the Chillingworth-Homes dy-
namo equations given in (14.18)-(14.20).
Problem 14.3. Determine x,, x,, x,, x,, and x, for the third-order recursive
map (14.21) with a = 3 and xo = 0.8. What is the range of values of x?
Chapter Fifteen
RENORMALIZATION
GROUP METHOD
15.1 Renormalization
In the first eight chapters of this book we considered the fractal behavior of
natural systems. This behavior was generally statistical and the physical
causes were generally inaccessible. In the six chapters that followed we con-
sidered low-order dynamical systems that exhibit chaotic behavior. Because
of the low order, the examples are generally quite far removed from natural
systems of interest. In this chapter and the next we take a collective view of
natural phenomena and consider some applications in geology and geo-
physics.
Thermodynamics represents the standard approach to collective phe-
nomena. System variables are defined, that is, temperature, pressure, den-
sity, entropy; and the evolution of these variables is determined from the first
law of thermodynamics (conservation of energy) and the second law of ther-
modynamics (variation of entropy). Statistical mechanics provides the ratio-
nal microscopic basis for much of thermodynamics.
In general, neither thermodynamics nor statistical mechanics yields frac-
tal statistics or chaotic behavior. Exceptions include critical points and phase
changes. A characteristic feature of a phase change is a discontinuous (cata-
strophic) change of macroscopic parameters of the system under a continu-
ous change in the system's state variables. For example, when water freezes
its viscosity changes from a very small value to a very large value with no
change in temperature.
The renormalization group method has been used successfully in treat-
ing a variety of phase change and critical-point problems (Wilson and
Kagut, 1974). This method often produces fractal statistics and explicitly
utilizes scale invariance. A relatively simple system is considered at the
smallest scale; the problem is then renormalized (rescaled) to utilize the
same system at the next larger scale. The process is repeated at larger and
larger scales. This is very similar to our renormalization models for frag-
RENORMALIZATION GROUP METHOD
mentation and the concentration of ores given in Chapters 3 and 5 respec-

tively.
We consider three specific applications to illustrate the use of the renor-
malization group method. We first consider a model for the flow of a fluid
through a porous medium. Such problems have wide applicability in ground-
water hydrology and petroleum engineering. The porous media may either
be a granular material such as a sandstone or a matrix of rock fractures. In ei-
ther case Darcy's law is generally applicable and the fluid velocity is propor-
tional to the pressure gradient; Darcy's law is a linear relation.
15.2 Percolation clusters
We apply the renormalization group method to understand the onset of fluid

flow through a porous medium. We first consider the two-dimensional
model illustrated in Figure 15.l(a). A square array is made up of a matrix of
square boxes that are referred to as elements; each box (element) may be ei-
ther permeable or impermeable. For the example given in Figure 15.l(a)
there are n = 256 elements. The probability that an element is permeable is
p,, the probability it is impermeable is 1 - p,. The question is whether the
square array is permeable or impermeable. The array is defined to be perme-
able if there is a continuous permeable path from the left side of the array to
the right side of the array. This is clearly a statistical question since the ac-
tual distribution of permeable and impermeable elements is random. For a
specified value of p, there is a probability P that the array with n elements is
permeable. For large arrays it is found that P is very small ( P 4 1) if 0 <
p, < p* where p* is the critical probability for the percolation threshold and
P is near unity (P = 1) if p* < p, < 1 (Stinchcombe and Watson, 1976).
Figure 15.1.(a)A16 X 16
array of square elements.
The probability p, that an
element is permeable is 0.5;
either the dark or the light
elements can be assumed to
be permeable. For either
case, no permeable path
across the array is found.
(b) Illustration of the
renormalization group
method; four square
elements are considered at
each of the four scales.
RENORMALIZATION GROUP METHOD 291
Thus there is a critical value of p,, p*, for the onset of flow through the grid
of elements. If p, is less than this critical value p*, a large square grid will al-
most certainly be impermeable to flow. If p, is greater than this critical value
p*, a large square grid will almost certainly be permeable to flow.
It may be easier to visualize this problem if one considers a forest made
up of a square grid of trees. The probability that a grid point has a tree is p,.
The question is whether a forest fire can burn through the forest if a tree can
ignite only its nearest neighbors. If there are no nearest neighbors the fire
does not spread. This forest-fire problem is mathematically identical to the
percolation problem considered above.
We now turn to the percolation problem and consider in some detail the
16 X 16 array of square elements illustrated in Figure 15.l(a). The total
number of elements n = 256. Taking p, = 0.5, it has been randomly deter-
mined whether each element is permeable or impermeable. With p, = 0.5 ei-
ther the dark squares or the light squares can be taken to be permeable. In ei-
ther case no continuous permeable path is found either horizontally or
vertically. Using the Monte Carlo approach a large number of random real-
izations would be carried out and the probability P(p,) that the array is per-
meable would be determined. For the two-dimensional square array with
large n, numerical simulations find that the critical probability for the onset
of flow in the array is p* = 0.59275 (Stauffer and Aharony, 1992).
It is also of interest to consider the number-size statistics for percolation
clusters at the critical limit p, = p*. The size of a percolation cluster is de-
fined to be the number of permeable elements in contact with each other
when the array first becomes permeable. The number of elements in the per-
colation cluster n: has been determined numerically as a function of the ar-
ray size n. For two-dimensional arrays it is found that (Stauffer and Aharony,
1992)
This result can be compared with the deterministic Sierpinski carpet assum-
ing that the remaining squares illustrated in Figure 2.3 represent a percola-
tion cluster. For the Sierpinski carpet ne = 8 when n = 9 and ne = 64 when n =
8 1, thus
Comparison of (15.1) and (15.2) shows that the fractal dimension for the
percolation cluster at the critical limit is D = 91/48 = 1.896, which is very
close to the value for a Sierpinski carpet D = In 81111 3 = 1.893.
At the onset of percolation, the sites through which flow takes place are
known as the percolation backbone. Sahimi et al. (1992, 1993) explained the
292 RENORMALIZATION GROUP METHOD
fractal distribution of seismicity illustrated in Figure 4.12 in terms of the

percolation backbone of a three-dimensional percolation cluster.
The direct statistical approach to the percolation problem becomes ex-
tremely time consuming if the number of elements is large since many ran-
dom realizations must be considered. An alternative approximate approach is
illustrated in Figure 15.l(b). At the lowest order a square array of four ele-
ments is considered. The probability p, that the first-order cell is permeable is
determined in terms of the probability p, that an individual first-order ele-
ment is permeable. The cell is defined to be permeable if there is a continuous
permeable path from left to right. The problem is then renormalized and four
first-order cells become the four second-order elements in a second-order
cell. The probability p, that the second-order cell is permeable is then deter-
mined in terms of the probability p, that a second-order element (first-order
cell) is permeable. The process is repeated at larger and larger scales (higher
and higher orders). This is the renormalization group method and is essen-
tially the same as the fractal fragmentation model illustrated in Figure 3.6.
We now consider the first-order cell and determine the probability that it
is permeable. All possible configurations are illustrated in Figure 15.2. The
probability that all four elements are impermeable is (1 - p0)4 and there is
only one configuration, as shown in Figure 15.2(a). This configuration is
clearly not permeable (-). The probability that one element is permeable
and three elements are not is p, (1 - p0)3 and there are four configurations
Figure 15.2. A cell consisting for the cell as illustrated in Figure 15.2(b). The permeable element can be in
of four elements is
any of the four positions shown but the cell is not permeable in any of them
considered. A cell is defined
to be permeable if there is a (-). The probability that two elements are permeable and two elements are
continuous permeable path not is pi(1 - po)2 and there are six independent configurations as shown in
from left to right. In (a) all Figure 15.2(c). The first and sixth configurations result in the cell being per-
four elements are meable for flows from left to right (+) and the other four configurations are
impermeable, in (b) one impermeable (-). The probability that three elements are permeable and one
element is permeable and the
four possible configurations
element is not is pi(l - p,) and there are four independent configurations as
are given, in (c) two shown in Figure 15.2(d).All four cell configurations are permeable (+). The
elements are permeable and
the six possible
configurations are given, in
(d) three elements are
permeable and the four
possible configurations are
given, and in (e) all four
elements are permeable. The
configurations that are
permeable from left to right
are denoted by (+) and the
configurations that are
impermeable from left to
right are denoted by (-).
probability that all four elements are permeable is p; and there is only one
configuration as shown in Figure 15.2(e): it is permeable (+).
Taking into account all possible configurations the probability that the
first-order cell is permeable is given by
The first-order probability includes the two configurations with two perme-
able elements, the four configurations with three permeable elements, and
the single configuration with four permeable elements. Renormalization is
carried out and four first-order cells become second-order elements. After
renormalization exactly the same statistics are applicable to the second-
order cell with the result
This result can be applied to the nth-order cell with the result
This recursive relation for the probability is quite similar to the logistic map
considered in Chapter 10.
,
Figure 15.3 shows the dependence of p,, on pn given in (15.5). To con-
sider the fixed points it is appropriate to rewrite (15.5) as

of the probability p,,, that
an ( n + 1)th-order cell is
permeable on the probability
p,, that an nth-order cell is
permeable. The critical
probability for the onset of
permeability is p* = 0.618.
The iteration for p, = 0.4 is
illustrated; this iteration is to
the impermeable fixed point
p_ = 0.The iteration for p, =
0.8 is also illustrated; this
iteration is to the permeable
fixed point p_ = 1.
The fixed points are obtained by settingf ( x ) = x with the result
In the range 0 < x < 1 there are three fixed points x = 0,0.618, 1. The corre-
sponding values of A = dfldx are 0, 1S28, 0. The fixed points at x = 0 and 1
are stable since IAl < 1 but the fixed point at x = 0.618 is unstable.
To illustrate further the iteration of the probabilities given by (15.5) we
consider two specific cases. For po = 0.4 we find p , = 0.294, p, = 0.166, and
p, = 0.054 as illustrated in Figure 15.3. The construction is the same as that
used in Figures 10.1-10.5. As the iteration is continued to large n the proba-
bility p,, approaches the stable fixed point p,, = 0. A large two-dimensional
array is impermeable forp, = 0.4. Forp, = 0.8 we findp, = 0.870, p, = 0.941,
and p, = 0.987 as illustrated in Figure 15.3. As the iteration is continued to
large n the probability p,, approaches the stable fixed point pn = 1. A large
two-dimensional array is permeable for p, = 0.8.
The unstable fixed point at p* = 0.618 is a critical point. At the critical
point, p,, = p* for all values of n; the probability that a cell is permeable is
scale invariant. For probabilities smaller than the critical value 0 < p, < p*
the iteration is to the impermeable limit pn = 0. For probabilities greater than
the critical value p* < p, < 1 the iteration is to the permeable limit pn = 1.
The value p* = 0.618 obtained by the renormalization group method com-
pares with p* = 0.59275 obtained by the direct numerical simulations for
large arrays.
The two-dimensional renormalization group method can be based on a
larger lowest-order array. Consider a 3 X 3 array with nine elements. Taking
into account all possible configurations, the probability p,,, that the (n +
1)th-order cell is permeable is related to the probability pn that the nth-order
cell is permeable by
The behavior of this relation is essentially similar to the behavior of (15.5).

The critical value for the onset of permeability is p* = 0.609. This is essen-
tially identical to the result obtained using the 2 X 2 renormalization group
result given above.
Extensive studies have also been carried out on three-dimensional cubic
arrays made up of n cubic elements. In this case the individual cubic ele-
ments are taken to be either permeable or impermeable. The array is defined
to be permeable if there is a continuous permeable path from one side of the
array to the other. Numerical studies of large cubic arrays show that the crit-
ical probability for the onset of flow in the array is p* = 0.3 117 (Stauffer and
Aharony, 1992). Again, a fractal relation of the form (15.2) is obtained be-
tween the number of permeable elements ne in the critical percolation cluster
and the total number of elements n with D = 2.5. This result can be compared
with the deterministic Menger sponge illustrated in Figure 2.4(a), assuming
that the remaining cubes represent a percolation cluster. For the Menger
sponge ne = 20 when n = 27 and ne = 400 when n = 729; the value D = 2.727
for the Menger sponge is somewhat higher than the value for three-dimen-
sional percolation clusters.
The simplest renormalization group model for the array of cubic ele-
ments is a 2 X 2 X 2 cubic array of eight elements. Taking into account
all possible configurations, the probability that the (n + 1)th-order cell
is permeable is related to the probability that the nth-order cell is perme-
able by
The critical value for the onset of permeability is p* = 0.282. This is in rea-
sonably good agreement with the numerical results considering the simplic-
ity of the model.
For fluid flow through rocks the two measurable quantities are the
porosity (the degree to which void space becomes filled with fluid) and the
permeability (the ability of the fluid to flow through the rock under fluid
pressure). The highly idealized model considered above predicts that there
will be a sudden onset of permeability at a critical value of the porosity. Al-
though rocks with low porosity have essentially zero permeability, the sud-
den onset of permeability at a universal critical value of the porosity is not
observed. This is attributed to the variety of aperture sizes and lengths occur-
ring in natural systems, which is not fully described by the idealized model.
Problems associated with electrical conduction through a matrix of ele-
ments are essentially identical to the percolation problems considered
above. Madden (1 983) has applied the renormalization group method to the
onset of electrical conductance through a grid of electrical conductors and
insulators.
15.3 Applications to fragmentation
Our second application of the renormalization group method is to fragmen-

tation (All&greet al., 1982; Turcotte, 1986a). We again consider fragmenta-
tion in terms of the scale-invariant model illustrated in Figure 3.6 with frag-
ments that differ in size by factors of two. However, there is a fundamental
difference. In Chapter 3 we considered a large zero-order cell and made

smaller and smaller elements. Here we consider the first-order elements to
be the smallest fragments and make larger and larger cells following the
renormalization group method. At the lowest order a cubic array of eight ele-
ments is assumed to constitute a first-order cell. Following Allbgre et al.
(1982), each element in a cell is hypothesized to be either fragile (f) if it is
permeated with microfractures or sound (s) if it is not. The probability that a
first-order cell is fragile is p, and is determined in terms of the probability
that an individual element is fragile, p,. It is necessary to specify a condition
for the fragility (soundness) of the first-order cell. Allkgre et al. (1982) hy-
pothesized that a cell was sound if a "pillar" of sound elements linked two
faces of the cell. Having prescribed the probability p, that an element is frag-
ile, it is necessary to determine the probability p, that a cell is fragile. To do
that it is necessary to consider all alternative configurations. In each cell
there can be zero to eight fragile elements so that there are 28 = 256 possible
combinations. Excluding multiplicities, there are 22 topologically different
configurations; these are illustrated in Figure 15.4. The numbers in parenthe-
ses are the multiplicities of each configuration. The fragile elements are in-
dicated by solid dots at the comers. Configurations 4f, 5c. 6b, 6c, 7, and 8
are fragile and are indicated by solid underlining in Figure 15.4. Examples
of sound and fragile cells are illustrated in Figure 15.5; 5b is a sound cell
with the pillar of strength illustrated by heavy lines, and 5c is a fragile cell.
The probability that all eight elements are fragile is pi, the probability
that seven elements are fragile and one is sound is p:(l - p,), and so forth.
Taking into account the configurations that are fragile and their multiplic-

22 topologically different
configurations. The "fragile"
elements are indicated by
solid dots on the comers.
The first number beneath a
configuration indicates the
number of fragile elements,
the letters indicate the
various configurations with
the same number of fragile
elements, and the number in
parentheses gives the
multiplicity of that
configuration. Using the
AlDgre er al. (1982) "pillar
of strength" condition, the
fragile cells are underlined.
ities the probability p , that a first-order cell is fragile is related to the proba-
bility po that a first-order element is fragile by
After renormalization exactly the same statistics are applicable at higher or-
ders. Thus we can write
If the characteristic size of the first-order cell is 2h, then the characteristic
size of the nth-order cell is 2nh.
Figure 15.6 shows the dependence of pn+!on pn given in (15.11). To con-
sider the fixed points it is appropriate to rewrite (15.1 1) as
The fixed points are obtained by settingflx) = x with the result
In the range 0 < x < 1 there are three fixed points x = 0,0.896, 1. The corre-
sponding values of A = dfldx are 0, 1.766, 0. The fixed points at x = 0 and 1
are stable since IAl < 1 but the fixed point at x = 0.896 is unstable.
Figure 15.5. Eight cubic

elements with dimension h
form a cell with dimension
2h. Configurations 5b and 5c
are illustrated; the five
fragile elements are shaded,
and the three sound elements
are unshaded. Cell 5b is
sound because the unshaded
(outlined) "pillar" of sound
elements links two faces.
Cell 5c is fragile because no
pillar of sound elements
links two faces.
To illustrate further the iteration given by (15.1l), we consider a specific

case. Forpo=0.6 wefindp, =0.2736,p,=0.0118,andp,=3.91 X 10-gas
illustrated in Figure 15.6. The construction is again the same as that used in
Figures 10.1-10.5 and in Figure 15.3.As the iteration is continued to large n,
the probability p,, approaches the stable fixed point pm = 0. Fragmentation
does not occur forp, = 0.6. The unstable fixed point atp* = 0.896 is a critical
point corresponding to catastrophic fragmentation. For probabilities smaller
than the critical value 0 < po < p* the iteration is to pm= 0 and fragmenta-
tion does not occur. At the critical limit p* the probability of fragmentation
is the same at all orders. Thus it is appropriate to set p* equal to the probabil-
ity f that a cell will fragment into eight elements, which was introduced in
(3.70). Settingf = p * = 0.896 and substituting into (3.76) we find that the re-
sulting fractal dimension is D = 2.84.
It is of interest to relate the renormalization group approach to fragmen-
tation as given above to the discussion of fragmentation given in Chapter 3.
The renormalization group approach provides a rational basis for assuming
that the fragmentation probability f is scale invariant. The applicable value
off is model dependent. For the "pillar of strength" model the derived fractal
dimension for fragments is D = 2.84. For the comminution model proposed
in Chapter 3 the derived fractal dimension for fragments was D = 2.6. The
latter is in somewhat better agreement with observed distributions.
Figure 15.6. Probability pn+,

of fragility at order n + 1 as
a function of the probability
p,, of fragility at order n from
(15.1 1). The critical
probability p* corresponding
to catastrophic fragmentation
is 0.896. The renormalization
group iteration is illustrated
for p, = 0.6, giving p , =
0.2736, p, = 0.01 18, and
p, = 3.91 X 10-8.
15.4 Application to fault rupture
As our final application of the renormalization group method we consider a

fractal-tree model for the rupture of a fault. When a fault ruptures the result is
an earthquake. An earthquake is in many ways analogous to a critical phenome-
non. At a critical value of the stress a catastrophic event, the earthquake, occurs.
As a result of the earthquake there is a discontinuity in the stress and strain.
A fault is generally considered to be a planar structure and can be mod-
eled using a velocity-weakening coefficient of friction. An actual fault in the
earth is generally more complex, being imbedded in a matrix that is frag-
mented by faults and joints. It appears appropriate to assume that a fault con-
tains asperities on a wide range of scales. At a small scale an asperity can be
an interacting roughness on the fault. At a large scale an asperity can be a bend
in the fault. An asperity has a limiting strength and when this limiting strength
is reached it will rupture. As a result of the rupture the stress on the asperity
will be transferred to adjacent asperities and they may or may not rupture.
In many ways this rupture is analogous to the failure of a stranded cable.
If one strand fails, the stress carried by that strand is transferred to adjacent
unbroken strands. Extensive numerical calculations of this failure problem
have been carried out (Harlow and Phoenix, 1982). Two observations are of
interest. The first is that a stranded cable will fail after only a few strands
have failed. The second is that the load that can be carried by a stranded ca-
ble is, on average, less than the load that could have been carried on a single-
strand cable of the same mass. Both of these observations appear to be ap-
plicable to faults. There is seldom any seismic indication of small ruptures
on a fault that is about to fail producing a large earthquake. Also, the failure
stress on faults is generally less than that predicted by simple frictional con-
siderations. Numerical simulations of the failure of stranded cables and the
problems relation to fault rupture have been given by Newman and Gab-
rielov (1991), Gabrielov and Newman (1994), and Newman et al. (1994).
To model fault rupture Smalley et al. (1985) considered the failure of a
fractal tree; they applied the renormalization group approach. Their basic
model is illustrated in Figure 15.7(a). The force F is carried by each of the
two first-order elements that make up the first-order cell. If one element
fails, the force on it is transferred to the other element and it must carry the
force 2F. A third-order fractal tree is illustrated in Figure 15.7(b). It includes
four first-order cells each of which carries a force 2F. These each become
second-order elements in the two second-order cells each of which carries a
force 4F. The two second-order cells become third-order elements in the sin-
gle third-order cell, which carries a force SF.
The objective is to determine the probability of failure of a cell in terms of
the probabilities of failure of its elements. We assume that the probability p,(F)
of failure of an element is given by the quadratic Weibull distribution (3.55):
where Fo is a reference strength.

For a cell containing two elements that are either broken (failed) or un-
broken, four states are possible: (1) [bb],(2) [bu],(3) [ub],(4) [uu],where b
represents a broken element and u represents an unbroken element. Note that
states (2) and (3) are equivalent and can be combined into a single state with
a multiplicity of two. The probabilities of these states in terms of the proba-
bility of failure po are:
In writing (15.15) to (15.17) the transfer of the force (stress) when an ele-
ment fails has not been considered.
If one element fails and the other is unbroken it is necessary to determine
whether the second element will fail when the stress from the first element is
transferred to it. We introduce the conditional probability p,, that an unbro-
ken element already supporting a force F will fail when an additional force F
is transferred to it. This mechanism for stress transfer leads to induced fail-
ures. The probabilities that the [ub] state will be broken or unbroken under
stress transfer are given by

the fractal tree model for the
renormalization group
approach to an earthquake
rupture. (a) The basic model;
(b) a third-order
construction.
From (15.14) and (15.17) the probability that a zero-order cell fails, p , , is
given by
It is now necessary to determine the conditional probabilityp,,; it is given by
where p0(2F) is the probability of failure under a force 2F. For the quadratic
Weibull distribution given in (15.14) we have
po(2F) = 1 - exp
[ - (31
The substitution of (15.14) into (15.22) gives
Combining (15.21) and (15.23) the conditional probability for the quadratic
Weibull distribution is given by
Substitution of (15.24) into (15.20) gives the probability that a cell fails, p , ,
in terms of the probability that the element fails, p,:
After renormalization exactly the same statistics are applicable to the sec-
ond-order cell, with the result
This result can be applied to the nth-order cell, with the result
Again this recursive relation resembles the logistic map considered in Chap-
ter 10.
Figure 15.8 shows the dependence of p,,, on pn given in (15.27). To
consider the fixed points it is appropriate to rewrite (15.27) as
The fixed points are obtained by settingf(x) = x with the result
In the range 0 < x < 1 there are three fixed points x = 0,0.2063, 1. The cor-
responding values of A = dfldx are 0, 1.6 l9,O. The fixed points at x = 0 and 1
are stable since IAl = 1 but the fixed point at x = 0.2063 is unstable.
To illustrate further the iteration of the probabilities given by (15.27), we
consider two specific cases. For p, = 0.1 we find p , = 0.05878, p, = 0.02184,
and p, = 0.00322 as illustrated in Figure 15.8. The construction is the same
as used in Figures 10.1-10.5. As the iteration is continued to large n, the
probability p,, approaches the stable fixed point p _ = 0. The large fractal tree
does not fail for p, = 0.1. For p, = 0.6 we find p , = 0.8093, p, = 0.9615, and
p, = 0.9985 as illustrated in Figure 15.6. As the iteration is continued to large
n, the probability p,, approaches the stable fixed point pm = 1. The large frac-
tal tree fails for p, = 0.6.
The unstable fixed point at p* = 0.2063 is a critical point. For probabili-
ties smaller than the critical value 0 < p, < p* the iteration is to the unbro-
ken limit pm = 0. For probabilities greater than the critical value p* < p, < 1
the iteration is to the broken limit p_ = 1 . The value p* = 0.2063 corresponds
to the catastrophic failure of the fractal tree.
Because of the transfer of stress, all elements fail when the probability of
failure of an individual element is only 0.2063. From (15.14) this corre-
the probability pn+,of failure
at order n + 1 on the
probability pn of failure at
order n from (15.27) for cells
containing two asperities
with a quadratic Weibull
distribution of strengths. The
procedure described in the
text for determining the
probability of cell failure for
successive iterations is
illustrated for p, = 0.6,O. 1.
The critical probability of
failure p* gives the
bifurcation point for
catastrophic failure of the
system. If 0 < p, < p*, the
solution iterates to pn = 0
and no failure occurs. If
p* < p, < 1 , the solution
iterates to pn = 1, and the
system has failed.
sponds to FIFO = 0.4807. Considered individually half the elements will

have failed when p, = 0.5 and FIFO = 0.8326. Thus the transfer of stress re-
sults in a lower failure stress than if a single element was considered.
Clearly this renormalization group approach is an idealization of actual
faults. However, some of the results are directly related to field observations.
A failure of a few elements cascades into a catastrophic total failure. Actual
faults do not have enhanced seismicity prior to a major rupture. This obser-
vation coupled with the model studies indicates that stress transfer is impor-
tant in fault ruptures. The relatively low failure stress is also in agreement
with field observations.
15.5 Log-periodic behavior
In the renormalization group approach to failure considered in the previous

section, each element was assigned a prescribed probability of failure as a
function of the stress on the element. For many materials it is preferable to
assign a statistical distribution of times to failure as a function of the stress
on the element. If a wire is placed under a stress a it does not fail instanta-
neously, but has a statistical distribution of lifetimes (Coleman, 1958). The
mean lifetime decreases as the applied stress is increased. In many cases an
exponential cumulative distribution of a failure times rf is assumed
where p(r,) is the probability that failure will occur in a time less than tf and
v(u) is known as the "hazard rate" under stress a. One-half of a large collec-
tion of wires under stress a will have failed when t,,, = (In 2)lv. It is often ap-
propriate to assume that the dependence of the hazard rate on stress is given by
where v, is the hazard rate under stress a, and p is typically in the range 2-5.
Combining (15.30) and (15.31) gives
This is a modified form of the Weibull distribution given in (3.55).

To illustrate the application of the time-to-failure approach, we consider
the failure of a stranded cable composed of N, elements each initially carry-
ing a load a,. When an element fails, the load on that element is transferred
uniformly to all the remaining elements; this is known as a global load-shar-

ing model. This approach is also an example of a mean-field approximation.
When n, elements have failed, the stress s on the n = No - n, surviving ele-
ments is given by
The rate at which elements fail is assumed to be given by the rate law
The solution of this deterministic equation with v a constant and n, = 0 at t =

0 basically corresponds to the probabilistic distribution given in (15.30).
Substituting the dependence of the hazard rate on stress from (15.31), we
combine (15.33) and (15.34) to obtain
Integrating with the condition that n, = No when t = t, we obtain
The number of surviving elements has a power-law dependence on the time

to failure with the exponent p-1. The route to failure is scale invariant and
exhibits fractal behavior. The time to failure is given by t, = (p vo)-1. An in-
crease in the power p decreases the time to failure as expected.
The occurrence of aftershock sequences following earthquakes can be
explained in terms of the time-to-failure approach discussed above. The dis-
placement on the primary fault associated with the initial earthquake results
in an increased stress in some adjacent regions. Earthquakes subsequently
occur on preexisting faults in these regions, resulting in aftershocks. The de-
lay in the occurrence of the aftershocks can be attributed to the applicability
of (15.32). It is interesting to note that a power-law decay in the rate of after-
shocks R of the form
is found to be universally applicable with p == 1. This relation is known as

Omori's law (Utsu, 1961).
It is also of interest to consider whether there is a precursory change in

seismicity that occurs prior to a major earthquake that can be associated with
the time-to-failure approach. Varnes (1989), Bufe and Varnes (1993), and
Bufe et al. (1994) have argued that there is a power-law increase in the re-
gional cumulative Benioff strain release prior to an earthquake. The Benioff
strain B is the square root of the seismic moment M defined in (4.3):
The cumulative Benioff strain in northern California prior to the October 17,
1989, Loma Prieta earthquake is given in Figure 15.9(a) (Bufe and Varnes,
1993). The increase in the Benioff strain illustrated in this figure fits the ex-
ponential scaling given in (15.36) quite well, but there also appears to be a
periodic component. Sornette and Sammis (1995) have also considered these
data and concluded that there is an excellent fit to a log-periodic increase in
seismic activity.
A simple power-law (fractal) increase in the cumulative Benioff strain B
is given by
where t, - t is the time prior to the earthquake and the constant a is negative.
To obtain log-periodic behavior, we assume that exponent a is complex: a =
5 + iq. In this case we obtain
= (1 - 8' [. $1
cos 1.(1 -
where 2 stands for the real part. This is log-periodic behavior. Combining
(15.39) and (15.40) a generalized self-similar expression for the cumulative
Benioff strain takes the form
where C, specifies the amplitude and 6 the phase of the log-periodic compo-
nent. This result is fully self-similar.
The sequence of positive maxima and/or minima in a log-periodic se-

quence can be used to predict the time to failure t,. The positive maxima in
( 15.40)correspond to the sequence
1920 1930 1940 1950 1960 1970 1980 1990

Date
(a)
Figure 15.9. (a) The data

points are the cumulative
Benioff strain release in
magnitude 5 and greater
earthquakes in the San
Francisco Bay area prior to
the October 17, 1989, Lorna
Prieta earthquake (Bufe
and Varnes, 1993). The solid
line is the log-periodic
correlation with the data.
(b) Predicted dates for the
Loma Prieta earthquake
based on the data applicable
prior to the cutoff date
(Sornette and Sammis,
1995).
where n = 1, 2, 3, . . . If three successive values of the maxima are specified,

t , , t,, t,, the failure time t, is given by
This result is also valid for the more general form of log-periodic behavior
given in (15.41). Because of the basic scale invariance, the value of t, ob-
tained from (15.43) is independent of the origin of the time scale used.
Sornette and Sammis (1995) used the generalized log-periodic relation
(15.41) to fit the strain accumulation data of Bufe and Varnes (1993); the re-
sult is given in Figure 15.9(a). The data appear to exhibit a series of log-peri-
odic fluctuations. An important question is whether this type of strain accu-
mulation data can be used to predict earthquakes. Sornette and Sammis
(1995) used the log-periodic fit to the data available prior to a cutoff date in
Figure 15.9(a) to predict when an earthquake would be expected. Their re-
sults are given in Figure 15.9(b). The predictions became increasingly accu-
rate as the cutoff times approached the date of the Loma Prieta earthquake,
October 17, 1989. Although the ability to make this retrospective prediction
is encouraging, it remains to be demonstrated that this technique can be used
successfully to predict earthquakes.
As a further illustration of log-periodic behavior, we now consider a hi-
erarchical model for failure similar to that discussed above but including the
time-to-failure approach (Newmann et al., 1995). An array of stress-carrying
elements is considered analogous to the strands of an ideal, frictionless ca-
ble. Each element has a time-to-failure that is dependent on the stress the
element carries and has a statistical distribution of values. When an element
fails, the stress on the element is transferred to a neighboring element; if two
adjacent elements fail, stress is transferred to two neighboring elements; if
four adjacent elements fail, stress is transferred to four neighboring ele-
ments; and so forth. The hierarchical model for failure is illustrated in Figure
15.10. At the lowest order in this example there are 128 zero-order elements.
These elements are paired to give 64 first-order elements, the 64 first-order
elements are paired to give 32 second-order elements, and so forth. A statis-
tical distribution of lifetimes is assigned to the lowest-order elements. When
one of these elements fails, the stress on the element is transferred to the
neighboring element, increasing the stress on it. If a pair of zero-order ele-
ments fail, that is, a first-order element, the stress is transferred to the adja-
cent pair of zero-order elements, that is, to the adjacent first-order element,
and so forth.
To illustrate the stress transfer consider the second-order (n = 4) exam-
ple given in Figure 15.1 1. Each element is given a probabilistic "lifetime"
and two examples of failure are illustrated. At time t = 0 the stress a, is ap-
plied to the four elements. In both examples element "2" has the shortest
lifetime and it is the first to fail. The stress uo on element "2" is transferred
to element "1" placing a stress 2u0 on this element as illustrated in Figure
15.11 (ii). The question now is whether the enhanced stress on element "1"
will cause it to fail prior to elements "3" or "4." In example (a) element "1"
is the next to fail and the stress 2u, on this element is transferred to elements

seventh-order (N = 128)
example of the hierarchical
model of failure.
Figure 15.1 1. Illustration of

stress transfer in a second-
order (N = 4) example of the
hierarchical model. Each
element is given a statistical
"lifetime." In example (a)
element "2" fails,
transfemng stress to element
"1"; element "1" then fails
and stress is transferred to
elements "3" and "4";
element "4" fails,
transferring stress to element
"3," which subsequently
fails. In example (b) element
"2" fails, transfemng stress
to element "1"; element "4"
fails, transfemng stress to
element "3"; element "3"
fails, transferring stress to
element "1 ," which (ii) (iii)
subsequently fails. (b)
"3" and "4" placing a stress 2a0 on both of these elements. Element "4" is
the next to fail and the stress 20, on it is transferred to the last surviving ele-
ment "3," which has a stress 4a0. In example (b) element "4" is the second
element to fail and the stress a, on this element is transferred to element "3"
placing stress 2a0 on this element. Element "3" is the next to fail and the
stress 20, is again transferred to the last surviving element "1," which has a
stress 4a0.
The zone of stress transfer is equal in size to the zone of failure. This lo-
cal load-sharing model simulates the elastic redistribution of stress adjacent
to a rupture. If elements are subjected to a constant stress a at t = 0, (15.32)
gives the statistical distribution of failure times t,. However, with stress
transfer the stress is not necessarily constant. To accommodate the increase
in stress caused by local load sharing from failed elements, we introduce a
reduced time to failure for each element Ti, given by
Each element i is assigned a random time to failure ti, under stress a, based
on (15.30). The actual time to failure of element i, namely Ti, is reduced be-
low t, if stress is transferred to the element. The time Ti, is obtained by re-
quiring that (15.44) be satisfied.
Consider the example illustrated in Figure 15.11(a). The four elements
i = 1, 2, 3, 4, carrying stress a, are assigned failure times t,,, t,,, t,,, and t4,
using the probability distribution (15.30). Element "2" has the shortest fail-
ure time so that
Upon failure the stress a, carried by this element is transferred to element

"1 ," as illustrated in Figure 15.11(a) (ii). Element "1" is the next element to
fail, and its failure time TI, is given by
Upon failure of "1," the stress 2ao is transferred to elements "3" and "4," as
illustrated in Figure 15.11(a) (iii). Element "4" is the next element to fail and
its failure time T4, is given by
Upon the failure of "4," the stress a, is transferred to element "3," as illus-
trated in Figure 15.11(a) (iv). The time to failure of element "3" is given by
Figure 15.12. Failure

sequence for a 16th-order
(N = 65,536) realization of
the model. (a) Entire failure
sequence (failure is
completed at t = 0.0480266).
(b) Expansion of the final
sequence of partial failures.
(c) Further expansion of two
partial failures.
Alternative failure sequences are also possible, one example of which is il-
lustrated in Figure 15. l l(b). Again "2" is the first element to fail; however,
in this case the second element to fail is "4," then "3" fails, and finally "1"
fails.
Results of a numerical calculation using a 16th-order (n = 65,536) real-
ization of this model are given in Figure 15.12. The total failure sequence is
given in Figure 15.12(a). The nondimensional time is taken to be T = v,t and
failure in this case occurs at T = 0.048027. It is interesting that failure occurs
at a nondimensional time that is more than an order of magnitude shorter
than the mean time to failure of an individual element T,,, = 0.61315. The
lifetime of the composite material is much shorter than the mean lifetime of
individual elements. This is in agreement with the results obtained above us-
ing the renormalization group approach.
The failure sequence between T = 0.0445 and total failure is expanded in
Figure 15.12(b). There is a well-defined sequence of partial failures prior to
the total failure at T~ = 0.048027. Well-defined partial failures occur at T , =
0.047965, T~ = 0.047799, T~ = 0.047487, T~ = 0.047 162, and T , = 0.046124.
The failure sequence between T = 0.04745 and T = 0.04785 is further ex-
panded in Figure 15.12(c) to show the structure of the partial failures at T =
0.047792 and T = 0.047487. In each case there is a nested sequence of
higher-order partial failures. Further expansion would show higher orders of
nesting. The structure is basically self-similar or fractal. There is a scale-
invariant sequence of precursory failures at all levels. Because of the sto-
chastic nature of the model, the embedding is not always clear, and a partic-
Figure 15.12. (cont.)

ular partial failure may be a part of a sequence or may be precursory to an-

other failure in the sequence. But this is also the problem with distributed
seismicity.
It is also of interest to determine whether the sequence of partial failures
can be inserted into the predictive log-periodic relation (15.43) to predict the
time of the total failure. Taking the sequence of partial failures T,, T, and T,,
we obtain the prediction T, = 0.047636 from (15.43); taking the sequence T,,
r3, and T,, we obtain T, = 0.054774 from (15.43); and taking the sequence T,,
r2, and T,, we obtain T, = 0.048154 from (15.43). The results are summarized
in Table 15.1. There is clearly considerable scatter in the predictions. Other
realizations give similar results. Although the embedded sequences of pre-
cursory failures are a ubiquitous feature of all realizations, there is consider-
able stochastic variability of the timing.
In Figure 15.13 the logarithm of the number of unfailed elements is
given as a function of the logarithm of the time to failure for a realization
with 4096 points and p = 4. The power-law fit shown by the dashed line has
a slope of 0.24, which compares with the power p = 0.25 predicted by the
global load-sharing relation (15.36). Although there is considerable scatter,
the power-law relation does appear to be a reasonable predictor for this
model, just as Bufe and Varnes (1993) found for regional seismicity. It is im-
Table 15.1. Successive precursory failure events illustrated in Figure 15.12

are used to predict the final total failure time in the simulation. Three
successive precursory failures are substituted into (15.43) to obtain the
estimated failure times. The actual failure time is also given
0.047636
3
1st estimate
0.048627
0.054775 True
3
2nd estimate failure
0.047799
time
0.047487
0.048154
0.047799 3
3rd estimate
0.047965
portant to note, however, that the quality of the fit deteriorates as complete
failure is approached. The global analysis employed in the derivation of
(15.36) deteriorates owing to the increasing importance of localization in the
evolution of the cascade of failures.
The sequence of failures as a function of position on the linear array of
elements is shown in Figure 15.14 for the above realization. The precursory
I o4 . , . , .... . . , ,.,, . . .... . , ,..,, . , .,..,. . . .,.,. . .....,.

,, , , ,,,.
1
--------
i i I
y = 1.391e+04 * xA(0.2445)R= 0.913 1
I I
' .........I
.... .
Figure 15.13. Failure

sequence for a 12th-order
(N= 4096) realization of the
model. The dashed line is a
power-law fit to the failure
sequence based on ( 1 5.36).
Figure 15.14. Sequence of

failures as a function of
position. First 512 elements.
3 14 RENORMALIZATION GROUP METHOD
cascades of failure are clearly illustrated. This figure illustrates the growing
importance of localization in failure events as criticality is approached.
Problems
Problem 15.1. A unit square is divided into 16 smaller squares of equal size
and the four central squares are removed; the construction is repeated.
Assume that the remaining squares represent a percolation cluster and
determine ne for n = 16 and 256.
Problem 15.2. Determine the equivalent expression to (15.2) for a cubic ar-
ray; use the Menger sponge as an example.
Problem 15.3. Consider the 3 X 3 renormalization group approach to the
two-dimensional array of square elements. For this array the tabulation
(permeable elements, impermeable elements, alternative configurations,
and permeable configurations) is: (0,9, 1, O), (1,8,9, O), (2,7,36, O), (3,
6, 8431, (4,5, 126,22), (5,4, 126,591, ( 6 3 , 84,67), (7,2,36,36), (8,
1,9,9), (9,0, 1, 1). Derive equation (15.8).
Problem 15.4. Consider the 2 X 2 X 2 renormalization group approach to
the three-dimensional array of cubic elements. For this array the tabula-
tion (permeable elements, impermeable elements, alternative configura-
tions, and permeable configurations) is: (0,8, 1, O), (1,7,8, O), (2,6,28,
41, (3,5,56,24), (4,4,70,54), (5,3,56,56), (6,2,28,28), (7, 1, 8, 81,
(8,0, 1, 1). Derive equation (15.9).
Problem 15.5. Assuming the dark elements in Figure 15.1 are permeable,
what is the size of the largest percolation cluster?
Problem 15.6. Assuming the light elements in Figure 15.1 are permeable,
what is the size of the largest percolation cluster?
Problem 15.7. Derive the conditional probability given by (15.21).
Problem 15.8. Derive (15.23) from (15.14) and (15.22).
Problem 15.9. Consider the third-power Weibull distribution given by
instead of (15.14). Show that (15.23) should be replaced by
Show that the recursive failure relation becomes

Find the value of the unstable fixed point p* and the corresponding value
of FIFO.
Problem 15.10. Consider the fourth-power Weibull distribution given by
instead of (15.14). Show that (15.23) should be replaced by
Show that the recursive failure relation becomes
Find the value of the unstable fixed point p* and the corresponding value
of FIFO.
Problem 15.11. Consider the rate law for failure given in (15.34). Assume
v = v, a constant. Show that the time where n, = 31 NOis T,,, = (In 2)lv0.
Problem 15.12. Consider the failure relation given in (15.36). Show that the
time to failure is given by t, = (pvo)-P.
Problem 15.13. Derive (15.43) from (15.42).
Problem 15.14. Show that (15.43) is invariant to a change in the origin of
time. Substitute t, = ti + to, t, = t; + to, etc. and show that the primed
times also satisfy (15.43).
Problem 15.15. If the first three maxima in a log-periodic sequence are t,, t,
and t,, (a) show that the fourth maxima in the sequence t, is given by
(b) Show that this result is invariant to a change in the origin of time,
substitute t, = ti + to, t, = ti + to, etc. and obtain the same result.
Problem 15.16. Assume that a series of events satisfy log-periodic behavior
leading to a catastrophic event. The first three events occur at t, = 0, t, =
15 days, and t, = 25 days. Determine t, and t, (use the result obtained in
Problem 15.15).
Problem 15.17. Assume that three earthquakes that occurred in 1956.2,
1980.7, and 1994.5 are precursors to a great earthquake. Also assume
that log-periodic behavior is applicable. Determine when the fourth
earthquake in the precursory sequence will occur and when the great
earthquake will occur.
Chapter Sixteen
SELF-ORGANIZED
CRITICALITY
16.1 Sand-pile models
In the last chapter we considered the renormalization group method for treat-
ing large interactive systems. By assuming scale invariance a relatively
small system could be scaled upward to a large interactive system. The ap-
proach is often applicable to systems that have critical point phenomena. In
this chapter we consider the alternative approach to large interactive sys-
tems. This approach is called self-organized criticality. A system is said to be
in a state of self-organized critically if it is maintained near a critical point
(Bak et al., 1988). According to this concept a natural system is in a margin-
ally stable state; when perturbed from this state it will evolve naturally back
to the state of marginal stability. In the critical state there is no longer a nat-
ural length scale so that fractal statistics are applicable.
The simplest physical model for self-organized criticality is a sand pile.
Consider a pile of sand on a circular table. Grains of sand are randomly
dropped on the pile until the slope of the pile reaches the critical angle of re-
pose. This is the maximum slope that a granular material can maintain with-
out additional grains sliding down the slope. One hypothesis for the behavior
of the sand pile would be that individual grains could be added until the
slope is everywhere at an angle of repose. Additional grains would then sim-
ply slide down the slope. This is not what happens. The sand pile never
reaches the hypothetical critical state. As the critical state is approached ad-
ditional sand grains trigger landslides of various sizes. The frequency-size
distribution of landslides is fractal. The sand pile is said to be in a state of
self-organized criticality. On average the number of sand grains added bal-
ances the number that slide down the slope and off the table. But the actual
number of grains on the table fluctuates continuously.
The principles of self-organized criticality are illustrated using a simple
cellular-automata model. As in the previous chapter we again consider a
square grid of n boxes. Particles are added to and lost from the grid using the
following procedure.
SELF-ORGANIZED CRITICALITY 317
(1) A particle is randomly added to one of the boxes. Each box on the
grid is assigned a number and a random-number generator is used to
determine the box to which a particle is added. This is a statistical
model.
(2) When a box has four particles it is unstable and the four particles are
redistributed to the four adjacent boxes. If there is no adjacent box
the particle is lost from the grid. Redistributions from edge boxes re-
sult in the loss of one particle from the grid. Redistributions from the
corner boxes result in the loss of two particles from the grid.
(3) If after a redistribution of particles from a box any of the adjacent
boxes has four or more particles, it is unstable and one or more fur-
ther redistributions must be carried out. Multiple events are common
occurrences for large grids.
(4) The system is in a state of marginal stability. On average, added par-
ticles must be lost from the sides of the grid.
This is a nearest neighbor model. At any one step a box interacts only
with its four immediate neighbors. However, in a multiple event interactions
can spread over a large fraction of the grid.
The behavior of the system is characterized by the statistical fre-
quency-size distribution of events. The size of a multiple event can be quan-
tified in several ways. One measure is the number of boxes that become un-
stable in a multiple event. Another measure is the number of particles lost
from the grid during a multiple event.
When particles are first added to the grid there are no redistributions and
no particles are lost from the grid. Eventually the system reaches a quasi-
equilibrium state. On average the number of particles lost from the edges of
the grid is equal to the number of particles added. Initially, small redistribu-
tion events dominate, but in the quasi-equilibrium state the frequency-size
distribution is fractal. This is the state of self-organized criticality. There is a
strong resemblance to the renormalization group approach considered in the
last chapter. In the renormalization group approach the frequency-size sta-
tistics are fractal only at the critical point. In the cellular automata model the
frequency-size statistics are fractal only in the state of self-organized criti-
cality.
The behavior of a sand pile and the behavior of the cellular automata
model have remarkable similarities to the seismicity associated with an ac-
tive tectonic zone. The addition of particles to the grid is analogous to the
addition of stress caused by the relative displacement between two surface
plates, say, across the San Andreas fault. The multiple events in which parti-
cles are transferred and are lost from the grid are analogous to earthquakes in
which some accumulated stress is transferred and some is lost. There is a
strong similarity between the frequency-magnitude statistics of multiple
events and the Gutenberg-Richter statistics for earthquakes. Before consid-
318 SELF-ORGANIZED CRITICALITY
ering the analogy further, we will describe the behavior of the cellular au-
tomata model in some detail.
As a specific example we consider the 3 x 3 grid illustrated in Figure
16.1. The nine boxes are numbered sequentially from left to right and top to
bottom as illustrated in Figure l6.l(a). The cellular automata model has
been run for some time to establish a state of self-organized criticality. The
further evolution of the model is as follows and is illustrated in Figure
16.1(b).
Step I A particle has been randomly added to box 8. The number of parti-
cles in this box has been increased from two to three.
Step 2 A particle has been randomly added to box 6, increasing the number
of particles from one to two. This addition is illustrated in the
change between steps 1 and 2 in Figure 16.l(b).
Step 3a A particle has been randomly added to box 5, increasing the number
of particles from three to four and making it unstable; the four parti-
cles are redistributed to the four adjacent boxes, increasing the num-
ber of particles in box 2 from three to four, the number of particles in
box 4 from three to four, the number of particles in box 6 from two
to three, and the number of particles in box 8 from three to four.
Boxes 2, 4, and 8 are now unstable. No particles are lost from the

the cellular automata model
for a 3 X 3 grid of boxes.
The boxes are numbered 1 to
9 as shown in (a). Particles 1 0 2 1 3 3 4 4 5 5 6 7
are randomly added to boxes
in (b) as shown in steps 1
and 2. In step 3a an added
particle in box 5 gives four
7 8 8 10 9 12
particles and these are
9 10 lla
redistributed to the adjacent
boxes. Nine more
redistributions are required
in steps 3b to 3j before the 1 0 2 1 3 2 4 3
grid is stabilized. The first
number below the grid is the
number of boxes that have
been unstable in the 1 2 2 3 3 3 4 5 1 2
sequence of redistributions.
The second number is the
cumulative number of
particles that have been lost
from the grid in the sequence
of redistributions.
grid. This redistribution is illustrated in step 3a in Figure 16.l(b).

The numbers below the grid are, on the left, the cumulative numbers
of boxes subject to redistribution and, on the right, the cumulative
number of particles lost from the grid.
Step 3b Since several boxes are now unstable, an arbitrary choice must be
made about which box will be considered first for further redistribu-
tion. The choice does not have a significant effect on the statistical
evolution of the system. The four particles in box 2 are redistributed.
One is lost from the grid and box 3 becomes unstable with four parti-
cles. Boxes 3, 4, and 8 remain unstable. In this sequence of redistri-
butions two boxes have been made unstable and one particle has
been lost from the grid.
Step 3 c The four particles in box 3 are redistributed. Two are lost from the
grid and box 6 becomes unstable with four particles. Boxes 4 , 6 , and
8 remain unstable. In this sequence of redistributions three boxes
have been made unstable and three particles have been lost from the
grid.
Step 3d The four particles in box 4 are redistributed. One is lost from the
grid and box 1 becomes unstable with four particles. Boxes 1,6, and
8 remain unstable. In this sequence of redistributions four boxes
have been made unstable and four particles have been lost from the
grid.
Step 3e The four particles on grid point 8 are redistributed. One is lost from
the grid and boxes 7 and 9 become unstable with four particles.
Boxes 1, 6, 7, and 9 remain unstable. In this sequence of redistribu-
tions five boxes have been made unstable and five particles have
been lost from the grid.
Step 3f The four particles in box 9 are redistributed. Two are lost from the
grid and box 6 is now unstable with five particles. Grid points 1, 6,
and 7 remain unstable. In this sequence of redistributions six boxes
have been made unstable and seven particles have been lost from the
grid.
Step 3g Four of the five particles in box 6 are redistributed. One is lost from
the grid and box 5 is now unstable. Boxes 1, 5, and 7 remain unsta-
ble. In this sequence of redistributions seven boxes have been made
unstable and eight particles have been lost from the grid.
Step 3h The four particles in box 5 are redistributed for the second time. No
particles are lost and no boxes are made unstable. Boxes 1 and 7 re-
main unstable. In this sequence of redistributions seven boxes have
been made unstable and eight particles have been lost from the grid.
Step 3i The four particles in box 7 are redistributed and two are lost from the
grid. No boxes are made unstable so that 1 is the only remaining un-
stable box. In this sequence of redistributions eight boxes have been
made unstable and ten particles have been lost from the grid.
Step 3j The four particles in box 1 are redistributed and two are lost from the
grid. No boxes remain unstable so that the sequence of 10 redistribu-
tions has completed step 3. During step 3 all nine boxes were unsta-
ble and 12 particles were lost from the grid.
of particles from zero to one.
of particles from two to three.
This relatively simple example illustrates how the cellular automata

model works. To develop significant statistics larger grids must be consid-
ered. Kadanoff et al. (1989) have carried out extensive studies of the behav-
ior of this model as a function of grid size. One statistical measure of the size
of an event is the number of grid points that become unstable. The results for
a 50 X 50 grid of boxes are given in Figure 16.2. The number of events N in
which a specified number of boxes A participated is given as a function of
the number of boxes. A good correlation with a fractal power law is ob-
tained, with a slope of 1.03. Since the number of grid points is equivalent to
an area, the equivalent fractal dimension is D = 2.06. This statistical behav-
ior appears to resemble that of distributed seismicity. However, the statistics
in Figure 16.2 are not cumulative. In fact a fractal relation is not obtained for
the cumulative statistics.
:
C9
r(
2
V
bD
0
-e 4
y=2.99-1.03~
Figure 16.2. Statistics for a

cellular-automatamodel on a
2
50 X 50 grid. The number Incremental Distribution
N of events in which a Grid Size: 50x50
specified number A of boxes 8
became unstable is given as a 0.0 0.4 0.8 1.2 1.6 2.0 2.4 2.8
function of A. log(Area)
A number of groups have studied the frequency-size statistics of

avalanches on real sand piles, and in some cases fractal distributions have
been found (Evesque, 1991; Nagel, 1992; Puhl, 1992). Held et al. (1990)
found fractal statistics applicable for small avalanches but not for large
avalanches. Bretz et al. (1992) and Rosendahl et al. (1993) found near-peri-
odic large avalanches and a fractal distribution of small avalanches. Segre
and Deangeli (1995) have developed a more realistic cellular-automata
model for actual landslides. The fractal statistics of actual landslides have
been considered by Yokoi et al. (1995).
Turbidite deposits are associated with slumps off the continental margin.
These avalanche-like events can be considered a natural analog for sand
slides and thus for the cellular-automata model considered above. Turbidite
deposits are generally composed of a sequence of layers, each layer repre-
senting a distinct event (slump). Each layer is composed of an upward grada-
tion from coarse-grained sediments to fine-grained sediments, and individ-
ual layers are generally separated by well-defined bedding planes.
Several studies of the thickness statistics of turbidite deposits have been
carried out. Rothman et al. (1994) carried out direct measurements on an
outcrop of the Kingston Peak Formation near the southern end of Death Val-
ley, California. Their results are given in Figure 16.3(a); an excellent corre-
lation with the fractal relation (2.6) is obtained taking D = 1.39. Hiscott et al.
(1992) have studied a volcaniclastic turbidity deposit in the Izu-Bonin fore-
arc basin off the shore of Japan. Layer thicknesses were obtained from for-
mation-microscanner images from well logs in the middle to upper
Oligocene part of the section. Results for two DSDP holes located 75 km
apart are given in Figure 16.3(b); a good correlation with (2.6) is obtained
taking D = 1.12.
It is difficult to make a direct comparison between the thickness statis-
tics of the sedimentary layers and the volume statistics of sand piles. How-
ever, the layer statistics appear to be scale invariant to a good approximation.
It is interesting to note that the fractal dimensions of the thickness statistics
are greater than one. For such a one-dimensional sequence it would appear
that this would be impossible considering the examples given in Figure 2.1.
The constructions illustrated in Figure 16.4 show that D can in fact be
greater than one. The standard Cantor set is illustrated in Figure 16.4(a); one
4,
layer, N, = 1, with thickness r , = two layers, N2 = 2, with thickness r2 = $,
four layers, N3 = 4, with thickness r3 = &. From (2.1) D = In 2lln 3 = 0.6309.
In Figure 16.4(b) a stretched Cantor set is illustrated. At each step the re-
maining segments are stretched by a factor of two before being further sub-
divided. This gives N, = 1 with r , = 3, N, = 2 with r2 = $,N3 = 4 with r3 = &.
From (2.1) D = In 2lln (312) = 1.710. The length L of the set is unbounded,
L + = as r + 0. For real data sets with both upper and lower bounds on r,
this construction illustrates that values of D greater than one are acceptable.
To model crustal seismicity, Barriere and Turcotte (1994) introduced a

cellular-automata model in which the boxes have a scale-invariant distribu-
tion of sizes. The objective was to model a scale-invariant distribution of
fault sizes. When a redistribution from a box occurs, it is equivalent to a
characteristic earthquake on the fault. A redistribution from a small box (a
Log layer thickness h (m)

(a)

frequency-thickness
statistics for turbidite
sequences of sedimentary
layers. (a) Kingston Peak
Formation near the southern
end of Death Valley, CA.
(b) Izu-Bonin forearc basin
off the shore of Japan. The
roll-off for thin layers
is attributed to loss of
resolution. The straight-line -2.5 kL _ I- _. - _ 1 L .
J
';4
J
correlations with the fractal -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0
relation (2.6) give D = 1.39 Log layer thickness h (m)
in (a) and D = 1.12 in (b). (b)
foreshock) may trigger an instability in a large box (the main shock). A re-
distribution from a large box always triggers many instabilities in the
smaller boxes (aftershocks).
As a specific example we again consider the surface exposure of the frac-
tal fragmentation model given in Figure 3.3. A fifth-order realization of this
construction is given in Figure 16.5. We have N , = 1 box with r , = $, N2 = 3
boxes with r, = $, N , = 9 boxes with r, = N4 = 27 boxes with r4 = h, and
i,
N, = 108 boxes with r, = 2. Except for N, the N are related to the ri by the
fractal relation (2. I) with D = In 3lln 2 = 1.5850.
Figure 16.4. (a) Cantor set

N,=l,r,=:;~,=2,r,=a;
N,=4,r,=h;~=ln2nn
3 = 0.6309. (b) Stretched
CantorsetN, = l , r , = f ;
i,
N, = 2, r,= ~ , = 4 ,
r, = $, ; D = In 21111(312) =
1.710.

the fractal cellular model
corresponding to the discrete
model for comminution
illustrated in Figure 3.7
carried to fifth order, D = In
311112 = 1 S85.
The standard cellular-automata rules are applied to this model:
(1) Particles are added one at a time to randomly selected boxes. The
probability that a particle is added to a box is proportional to the area
A, = r;of the box.
(2) A box becomes unstable when it contains 4A. particles.
(3) Particles are redistributed to immediately acfjacent boxes or are lost
from the grid. The number of particles redistributed to an adjacent
box is proportional to the linear dimension ri of that box.
(4) If, after a redistribution of particles from a box, any of the adjacent
boxes are unstable, one or more further redistributions are carried
out. In any redistribution, the critical number of particles is redis-
tributed. Redistributions are continued until all boxes are stable.
The cumulative frequency-magnitude statistics for main shocks of a

seventh-order (1 28 X 128) version of the model are given in Figure 16.6. We
find an excellent correlation with the fractal relation (2.6) taking D = 2.50
(b = 1.25). This is significantly higher than the observed values for distrib-
uted seismicity. Evernden (1970) has obtained b-values for regional seismic-
+
ity and concludes that b = 0.85 0.20. It was also found that 31.5% of the
largest events had foreshocks. This is in reasonable agreement with studies
of actual earthquakes; von Seggern et al. (1981) found that 21%of the earth-
quakes studied had foreshocks and Jones and Molnar (1979) found that 44%

frequency-magnitude
statistics for unstable events.
The number of events Nc in
boxes equal to or smaller
than r is divided by the total
number of events N, and
given as a function of r. The
correlation is with (2.6)
taking D = 2.5. Box Size r
of larger shallow earthquakes that could be recorded teleseismically had

foreshocks. The aftershocks also correlate well with (2.6) taking D = 2.02
(b = 1.01). A similar model has been proposed by Henderson et al. (1994).
16.2 Slider-block models
Slider-block models that exhibit self-organized criticality can also be con-

structed. In Chapter 11 we showed that a pair of interacting slider blocks can
exhibit deterministic chaos. This model is easily extended to include large
numbers of slider blocks. Carlson and Langer (1989) considered long linear
arrays of slider blocks with each block connected by springs to the two
neighboring blocks and to a constant-velocity driver. They used a velocity-
weakening friction law and considered up to 400 blocks. Slip events involv-
ing large numbers of blocks were observed, the motion of all blocks in-
volved in a slip event were coupled, and the applicable equations of motion
had to be solved simultaneously. Although the system is completely deter-
ministic, the behavior was apparently chaotic. Frequency-size statistics
were obtained for slip events and the events fell into two groups: smaller
events obeyed a power-law (fractal) relationship but there was an anom-
alously large number of large events that included all the slider blocks. This
model was considered to be a model for the behavior of a single fault, not a
model for distributed seismicity. The large events were associated with char-
acteristic earthquakes on the fault and smaller events with background seis-
micity on the fault between characteristic earthquakes.
Nakanishi (1990, 1991) proposed a model that combined features of the
cellular-automata model and the slider-block model. A linear array of slider
blocks was considered but only one block was allowed to move in a slip
event. The slip of one block could lead to the instability of either or both of
the adjacent blocks, which would then be allowed to slip in a subsequent
step or steps, until all blocks were again stable. Brown et al. (1991) proposed
a modification of this model involving a two-dimensional array of blocks.
Other models of this type have been considered by Bak and Tang (1989),
Takayasu and Matsuzaki (1988), Ito and Matsuzaki (1990), Sornette and
Sornette (1989, 1990), Langer and Tang (1991), Carlson (1991a, b), Carlson
et al. (1991, 1993a, b, 1994), Matsuzaki and Takayasu (1991), Rundle and
Brown (1991), Feder and Feder (1991), Chen et a1. (1991), Shaw et al.
(1992), Huang et al. (1992), Christensen and Olami (1992), Olami and
Christensen (1992), Olami et al. (1992), Vasconcelos et al. (1992), de Sousa
Vieira (1992), Cowie et al. (1993), Shaw (1993a, b, 1994, 1995), Rundle and
Klein (1993, 1995), de Sousa Vieira et al. (1993), Schmittbuhl et al. (1993),
Knopoff et al. (1993), Ding and Yu (1993), Lu et al. (1994), Senatorski
(1994), Xu and Knopoff (1994), Pepke and Carlson (1994), Pepke et al.
(1994), Rubio and Galeano (1994), Robinson (1994), Espanol (1994), and
Lin and Taylor (1994). McCloskey (1993), and McCloskey and Bean (1994)
considered arrays of slider blocks connected to two driver plates, and these
driver plates were treated as a pair of interacting slider blocks.
The standard two-dimensional array of slider blocks is illustrated in Fig-
ure 16.7. In the cellular-automata approximation it is assumed that during
the sliding of one block, all other blocks are stationary; this requirement lim-
its the system to nearest neighbor interactions, which is characteristic of cel-
lular-automata systems. To minimize the complexity we considered a dis-
continuous static-dynamic friction law. After non-dimensionalization of the
governing equations, the governing parameters are a = kclkl (kc is the spring
constant of the connector springs, k, is the spring constant of the puller
+
springs), a is a measure of the stiffness of the system, = FJF, (the ratio of
the static friction F, to dynamic friction F,), and N the number of blocks
+
considered. In this model the parameter can be eliminated by rescaling.
Thus for large systems (N very large) the only scaling parameter is the stiff-
ness a. Frequency-size statistics for a 50 X 50 (N = 2500) array are given in
Figure 16.8 for several values of the stiffness parameter a . A good correla-
tion is obtained with the fractal relation (2.6) with D = 2.72. The fre-
quency-size relation shows a roll-off from the power law near the larger end
of the scaling region. This deviation is reduced as the parameter a increases.
Frequency-size statistics for several different size arrays are given in Figure
16.9. When the parameter alNlJ2 is greater than one, we observe an excess
number of catastrophic events that include the failure of all blocks. The fail-
ure statistics of these multiple-block systems clearly indicate a self-orga-
nized critical behavior and are remarkably similar to distributed seismicity.

the two-dimensional slider
block model. An array of
blocks each with mass m is
pulled across a surface by a
driver plate at a constant
velocity V. Each block is
coupled to the adjacent
blocks with either leaf or coil
springs with constant kc, and
to the driver plate with leaf
springs with spring constant
k,. The extension of the (i, j )
pulling spring is xii
The frequency-size distribution of events associated with self-organized

criticality certainly resembles the regional distribution of earthquakes in a
zone of active tectonics. This suggests that interactions between faults play
an essential role in the behavior df such zones.
Figure 16.8. The ratio of the

number of events N with size
Nf to the total number of
events No is plotted against
Nf (Nf is the number of
blocks that participate in an
event and is a measure of the
area of an event). Results are
given for 4 = 1.5 and a = 10,
15,20, 30, and 40. The solid
line is the correlation with
the power-law (fractal)
relation (2.6); the
corresponding fractal
dimension is D = 2.72.
Figure 16.9. The ratio of the

number of events N with size
Nf to the total number of
events No is plotted against
N,. Results are given for
systems of size 20 X 20.30
X 30.40 X 40, and 50 X 50
with parameters 4 = 1.5 and
a = 50. The peaks at log
I
- N, = 2.60,2.95, and 3.20
correspond to catastrophic
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 events involving the entire
log(N~) system.
An important consequence of a critical state in the crust is the large

range of interactions. A basic question is whether an earthquake on one part
of the planet, say Mexico, is correlated with an earthquake at a large dis-
tance, say Japan. The classical approach to earthquakes would say that this is
impossible. The stresses associated with seismic waves are too small to trig-
ger an earthquake and there is no conclusive observational evidence for cor-
related events on this spatial scale. The stress changes associated with the
fault displacement are localized and are damped by the athenospheric vis-
cosity. However, interactions at large distances are a characteristic of critical
phenomena. The interactions are not through the direct transmission of stress
but through the interactions of faults with each other. Scholz (1991) has ar-
gued that the entire earth's crust is in a state of self-organized criticality. He
sites as direct supporting evidence the induced seismicity associated with
dams and other sources. Whenever a reservoir is filled behind a large new
dam, extensive swarms of earthquakes are generally triggered. This is evi-
dence that the crust is at the brink of failure even at large distances from
plate boundaries.
This action at a distance may help to explain the apparent success of the
earthquake-prediction algorithms developed at the International Institute for
the Theory of Earthquake Prediction and Theoretical Geophysics in Moscow
under the direction of Academician V. I. Keilis-Borok. This approach is based
on pattern recognition of distributed regional seismicity (Keilis-Borok, 1990;
Keilis-Borok and Rotwain, 1990; Keilis-Borok and Kossobokov, 1990). The
pattern recognition includes quiescence (Schreider, 1990), increases in the
clustering of events, and changes in aftershock statistics (Molchan et al.,
1990). In reviewing regional seismicity after a major earthquake it is often
observed that the regional seismicity in the vicinity of the fault rupture was
anomalously low for several years prior to the earthquake (Kanamori, 1981;
Wyss and Haberman, 1988). This is known as seismic quiescence. The prob-
lem has been to provide quantitative measures of quiescence prior to the ma-
jor earthquake. We discussed the fractal clustering of earthquakes in Chapter
6. The clustering of regional seismicity appears to become more fractal-like
prior to a large earthquake. There also appears to be a systematic reduction
in the number of aftershocks associated with regional intermediate-sized
earthquakes prior to a major earthquake. Pattern-recognition algorithms
were developed to search earthquake catalogs for anomalous recursory be-
havior. Premonitory seismicity patterns were found for strong earthquakes in
California and Nevada (algorithm "CN") and for earthquakes with M > 8
worldwide (algorithm "M8"). When a threshold of the anomalous behavior
was reached, a warning of the time of increased probability (TIP) of an
earthquake was issued.
On a worldwide basis TIPS were triggered prior to 42 of 47 events. TIPS
were released prior to the Armenian earthquake on December 7, 1988, and to
the Loma Prieta earthquake on October 17, 1989. These are illustrated in
Figure 16.10. The TIP issued for region 3 in the Caucasus during January
1987 was still in effect when the Armenian earthquake occurred in this re-
gion on December 7, 1988. TIPS were issued for region 5 in California dur-
ing October 1984 and for region 6 during January 1985. These warnings
were still in effect when the Loma Prieta earthquake occurred within these
overlapping regions on October 17, 1989.
The fault rupture of the Loma Prieta earthquake extended over about 40
km. However, the prediction algorithms detected anomalous seismic behav-
ior over two regions with diameters of 500 km. Self-organizedcriticality can
explain anomalous correlated behavior over large distances.
This approach is certainly not without its critics. Independent studies
have established the validity of the TIP for the Loma Prieta earthquake;
however, the occurrence of recognizable precursory patterns prior to the
Landers earthquake are questionable. Also, the statistical significance of the
size and time intervals of warnings in active seismic areas has been ques-
tioned. Nevertheless, seismic activation prior to a major earthquake cer-
tainly appears to be one of the most promising approaches to earthquake pre-
diction.
I . THE CAUCASUS. M 2 6.5
44
42
Figure 16.10. Illustrations of

40
the Armenian (December 7,
1988) and Lorna Prieta
38 (October 18, 1989, Moscow
time) earthquakes by 'Keilis-
Borok (1990). In (a) the
Caucasus region is broken up
b. CALIFORNIA-NEVADA,M 2 7
into 10 areas with diameters
50
of 500 km; two warnings (for
1975 1980 1985 region 3 and 9) are shown on
45 1
!he right. The locations and
2 times of four earthquakes are
40
3 also given. In (b) the
35
'4 - California-Nevada region is
broken up into eight areas
5 with diameters of 500 km.
30 6- Four warnings (for regions
4-6.8) and the locations and
times of four earthquakes are
25
-135 -130 -125 -120 -115 -110 given.
Although long-distance correlations between earthquakes are a subject

of considerable controversy, such correlations have been widely accepted in
China and Russia (as well as the former Soviet Union). A striking example
was a sequence of five earthquakes that occurred in China between 1966 and
1976. These were the m = 7.2 Shentai (1966), m = 6.3 Hijien (1967), m = 7.4
Bo Sea (1969), m = 7.3 Haicheng (1975), and the m = 7.8 Tangshan (1976)
earthquakes. These earthquakes spanned a distance of some 700 km, and the
Haicheng earthquake was successfully predicted by the Chinese, at least par-
tially on the basis of seismic activation (Scholz, 1977). However, the Tang-
shan earthquake was not predicted and estimates of fatalities in this earth-
quake range from 250,000 to 450,000.
Seismic activation has been previously recognized in association with an
increase in seismicity that occurred in the San Francisco Bay area prior to
the 1906 earthquake (Sykes and JaumC, 1990). Earthquakes with estimated
magnitudes between 6.5 and 7.0 occurred in 1865 (Santa Cruz Mountains),
1868 (Hayward), 1892 (Vacaville), and 1898 (Mare Island). There is a seri-
ous concern that a similar seismic activation is now underway in southern
California. A number of intermediate-size earthquakes have occurred in
southern California in the last 45 years. These include the m = 7.4 Kern
County earthquake on July 21, 1952, the m = 6.4 San Fernando earthquake
on February 9, 1971, the m = 7.6 Landers earthquake on June 28, 1992, and
the m = 6.6 Northridge earthquake on January 17, 1994.
The Landers earthquake provided direct evidence that faults interact
with each other over large distances (Hill et al., 1993). The Landers earth-
quake triggered earthquakes at 14 distant sites scattered over the western
United States. The farthest site was Yellowstone National Park in Wyoming,
1250 km from Landers. Just how information is transmitted over these dis-
tances is uncertain. One hypothesis is that the surface waves of the Landers
earthquake were responsible. However, the stress levels associated with sur-
face waves at this distance are no larger than the daily variations in stress as-
sociated with the earth tides.
It is known from statistical mechanics that near a critical point spatial
correlations extend to large distances. To better understand the statistical
mechanics of slider-block models, we consider a two-dimensional array of
slider blocks without a driver plate. Each block is connected to its four
neighbors with springs (spring constant k) and is confined to move in the
x-direction. The slider blocks interact frictionally with a surface; however, to
conserve energy the dynamic friction is assumed to be zero. The problem is
specified by the static friction and the initial total energy in the system. The
force on a block (i, j) is
A block slides if lF;,.,l > Fs,where Fsis the prescribed static friction force. To
simplify the analysis and simulations, only one block in the array is updated
during each microscopic time step. Before the update there are two possible
states:
(I) The block was stuck after the previous update IF;, ,In+
< Fs. How-
ever, the forces on the block have changed because of subsequent
updates on neighboring blocks, there are now two possibilities:
(a) The block is still stable, l ~ ~ , ~-, ,<, ,Fs, and the update is termi-
nated.
(b) The block is still unstable, IF, ,In+,
- > Fs. In this case motion of
the single slider block is given by
The slipping block executes one-half of an harmonic cycle and

sticks when the velocity is again zero. The change in the posi-
tion of block (i, j), A x ,,, is related to the initial force on the block
(FO;,,), + 1 - by
The new net force on the block (Firj)n+l+is determined, again

there are two possibilities:
(i) If IF;, < F, the block remains stuck until the next up-
date,
(ii) If IF^,^^,, > Fsthe block slips until the next update.
+
(2) The block was slipping after the previous update IF;, ,In+
> Fs. But
again the forces on the block have changed because of subsequent
updates on neighboring blocks. There are two possibilities:
(a) The block is now stable, IF;,]^+^- < Fs, and the step is termi-
nated.
(b) The block is still unstable IF;, , I n + , -
> Fs, and then (16.3) is used
to determine the new position of the block and the new net force
,+
on the block (F,,j)n+ is determined. Again there are two possi-
bilities:
(i) If (F;,]~+, < Fs the block remains stuck until the next up-
+
date.
(ii) If IF^, jln+
I > Fs, the block slips until the next update.
+
The slider blocks are considered sequentially using a checker-board algo-

rithm to sweep across the two-dimensional array.
It is convenient to introduce the nondimensional variables Ti, = Fi,,IFs.
zk
Z = kxlF,, i = t m , and = k E k l qwhere E, is the energy in spring k. The
nondimensional force on a block is
and the motion of a block is given by
If IFi]- > 1, block (i, j) is unstable and its nondimensional slip is given by
At t = 0 the blocks are given a random distribution of displacements of the

nondimensional energy in spring k is &.
The only parameter in this problem
is the mean energy per spring introduced at t = 0, E, which is given by
Since no energy is dissipated, this value remains constant and we use it as a

control parameter for the model. If E is large, very few blocks will stick and
we would expect that the system should behave like a set of harmonic oscil-
lators with a Maxwell-Boltzmann (Gaussian) distribution of displacements.
If the distribution of displacements of individual block (i,j) is Gaussian,
the distribution of spring displacements will also be Gaussian. If this is the
case, the probability distribution function for the energies in the springs &
will be given by
One of the questions we address is whether the system evolves to this

Maxwell-Boltzmann distribution. The corresponding probability distribu-
tion function for the forces on the springs is
However, the slip condition for a block is determined by the statistical distri-
bution of forces on the blocks. From (16.4) it is seen that the random force
on a block is the sum of four random forces on the neighbor springs. These
forces are not independent as one can see from (16.4) and the Gaussian dis-
tribution of forces on the block is
A block can slip if [FBI> 1. Using (16.10) the probability that a block will be
slipping Psis
We will show that our results satisfy this condition.

We have carried out a series of simulations on square arrays of up to
2000 X 2000 blocks. Springs on the boundaries of the array are attached to
fixed walls. Random initial displacements were given to the blocks corre-
sponding to specified values of E. Various initial distributions of energy
(non-Gaussian) were used, and in all cases the system evolved to the
Maxwell-Boltzmann distribution (16.8). A typical example is given in Fig-
ure 16.11 with E = 1. The statistical distribution of forces on the blocks was
also determined and was found to be in excellent agreement with (16.10).
The fraction of the blocks that are slipping Ps are given for several values of
the mean energy E in Figure 16.12. Good agreement with the equilibrium
prediction (16.11) is found.
As the fraction of slipping blocks increases with increasing values of E, a
path of slipping blocks across the array is eventually established. This strongly
resembles the percolation threshold for the site-percolation model considered
in Chapter 15. Both are critical points and the critical value of E is 0.213 with
the corresponding fraction of slipping blocks Ps = 0.583. This value can be
compared with the critical point for the site percolation model where the prob-
ability that a lattice percolates is p* = 0.5927. The small discrepancy between
the two values is attributed to correlations between blocks in the slider-block

model. Figure 16.13 shows the frequency-size distribution at this critical
point (E = 0.213) and, as a reference, the frequency-size distribution for the
site-percolation model with p = 0.5927; the two distributions are virtually
identical power laws. A typical slider-block configuration with a continuous
path of slipping blocks across the array is shown in Figure 16.14. It is clearly
very similar to the site-percolation distribution given in Figure 15.1l(a).
Figure 16.11. T_he probability

distribution p(E,)of the
nondimensional energies E k
in the springs of a multiple
slider-block model. The
crosses are the result for a
2000 X 2000 array of slider
blocks with E = 1. The
solid line is the Maxwell-
Boltzmann distribution of
energies given in (16.8).
Figure 16.12. The fraction of

the blocks that are slipping
PI (E)is given as a function
of the mean energy E. The
crosses are results for a
1000 X 1000 array of slider
blocks, and the soiid line is
the prediction from (16.1 1).
This simple energy-conserving system exhibits behavior that is quite

similar to the seismicity in active tectonic regions such as California. In
southern California the seismic activity level in the magnitude range 2 <
M < 5 not only satisfies the fractal Gutenberg-Richter frequency-magni-
p=0.213
site percolatm I
i Figure 16.13. Number of
clusters ns of size s as a
function of s. The solid line
is the distribution of
percolating clusters for a
2000 X 2000 array with the
critical percolation
probability p = 0.5927. The
dashed line is the distribution
of slipping clusters of blocks
on our 2000 X 2000 array of
slider blocks at the critical
point E = 0.2 13.

a typical configuration of
sliding blocks at the critical
point E = 0.2 13 for a 64 X
64 array. White blocks are
stuck and black blocks are
sliding. A continuous path of
sliding blocks across the
array is present.
tude scale, but also, the level of activity does not vary from year to year (see
Figure 4.3). Earthquakes in this magnitude range strongly resemble the sta-
tistical fluctuations of the slider-block array near its critical point.
Further evidence supporting the applicability of the "percolationw-like
model comes from the spatial distribution of seismicity in southern Califor-
nia. The distributions given in Figure 4.12 appear to correspond to the fractal
dimension of the percolation "backbone" of a critical three-dimensional per-
colation cluster. It appears that the earthquakes on a complex array of faults
form a connected path across the zone of crustal deformation in direct anal-
ogy to the "percolation backbone" of slipping blocks in the array. Rundle et
al. (1995) found that the block energy distribution for a driven slider-block
model is a Maxwell-Boltzmann distribution as the model approaches the
mean field where fluctuations are minimal.
16.3 Forest-fire models
We next consider a class of models that are referred to as forest-fire models.

These models generally exhibit self-organized criticality. We consider a
square grid of sites, with each site designated by two numbers ij, where i is
the row and j is the column. At each time step either a tree is randomly
planted on a site or a match is dropped on the site. The sparking frequency f
indicates how many trees are planted before a match is dropped. If r = &, 99
trees are planted (or are attempted to be planted) before a match is dropped.
If a match is dropped on an empty site, nothing happens. But if a match is
dropped on a tree, it ignites and all immediately adjacent trees bum.
As a specific example of our forest-fire model we consider the 10 X 10
grid illustrated in Figure 16.15. The model has been run for some time to es-
tablish a state of self-organized criticality and its initial state is given in Fig-
:
ure 16.15(a). We take f = so that four trees are planted before a match is
dropped. Between Figures 16.15(a) and (b) there are 5 time steps and the
randomly selected grid points were 71, 76, 56, 81, and e5L Trees were
planted on 7 1,56, and 8 1; 76 already had a tree, and a match was dropped on
95. This tree ignited and 35 adjacent trees also burned. Note that only trees
immediately above, below, or to the sides of a burning tree also ignite. Fol-
lowing this forest fire 10 additional time steps are camed out to reach the
distribution illustrated in Figure 1 6 . 1 3 ~ ) The
. 10 randomly selected grid
points are 72, 36, 00, 88, 48, 65, 44, 30, 45, and 44. Trees were planted on
72, 36, 88, 65, 44, 30, and 45; 00 already had a tree. The match dropped on
08 did not ignite because there was no tree on the grid point. The match
dropped on 44 ignited this tree and burned the adjacent tree on 45. Following
this small fire 25 additional time steps are carried out to reach the distribu-
tion illustrated in Figure 16.15(d). The 25 randomly selected grid points are
56,36,05, 15,25,68,40,52, 1 8 , 8 1 , 0 3 , 7 9 , 3 5 , 3 5 , 9 5 , 5 6 , 5 9 , 8 0 , 5 1 , 0 7 ,
20,56,86,46, and 30. Trees were already planted on 36,05,79,35,56, and
56. The matches dropped on 25, 81, 95, and 07 did not ignite because there
were no trees on these grid points. The match dropped on 30 ignited this tree
and burned six adjacent trees.
Frequency-size statistics for forest fires can be determined. Two exam-
ples for a 100 X 100 tree forest are given in Figure 16.16. The number of
1
burning clusters N is given as a function of their size A, forf = &j and f = m.
For the larger value f = &
j , fires consume the forest before large clusters can
form. A reasonably good correlation with the fractal relation (2.2) is ob-
tained taking D = 2.00. The roll-off from the power law near the larger end
of the scaling region is very similar to that illustrated for the slider-block
model in Figure 16.8. When the sparking frequency f is reduced to &, we
observe an excess number of catastrophic fires that consume all or nearly all
of the 10,000 trees. Again this is very similar to the behavior found for slider
blocks when the stiffness parameter is large as illustrated in Figure 16.9.

the forest-fire model on a
10 X 10 grid of points. Each
0 0 X X point is identified by its ij
0 X 0 X coordinates (i-row,j-column).
X X 0 X We take f = &.(a) Grid points
X X X with trees are indicated by
0 X X X X X X
circles. Between (a) and (b)
X X
trees have been planted on
X X X X X X X 0
points 71.56, and 81; a
X X X X X
match was then dropped on
X X x x m x 0
point 95, igniting the tree
(indicated by m) and burning
35 adjacent trees (indicated
by xs). Between (b) and (c)
trees were planted on points
72,36, 88,65,44, 31, and
45; a match was dropped on
point 44, igniting this tree
(indicated by m ) and burning
the adjacent tree at 45
(indicated by an x). Between
(c) and (d) trees were planted
on points 56, 15.68.40.52.
18,03,35,59, 80,20, 86,
and 46; a match was dropped
on 30 (indicated by m) and
burned 6 adjacent trees
(indicated by xs).
i Best fit line
1 Best fit line \

lo'
N
io3
1'
0
Figure 16.16. The number of
forest fires N with size Af
(Af is the number of trees that
bum in a fire) is given as a 10'
function of A, Results are for
a 100 X 100~orest-grid with
sparking parameters f = &j 10
o +-I
and A. 1o0
*-. Best fit line)

,&lope =--0.59)
- -

number per year N o f fires 0.1
with burned areas greater
than Af given as a function
of AT The data is for the
Australian Capital Temtory
for the period 1926-199 1.
The best-fit straight line is
. .
0.01
taking D = 1.18. 0.01 0.1 1 A,, km2 10 100 loo0
It is also of interest to compare the results of forest-fire models with the

frequency-size statistics for actual forest fires. Data for forest and brush
fires in the Australian Capital Territory for the period 1926-1991 are given
in Figure 16.17. A reasonably good straight-line fit with the fractal relation
(2.6) is obtained taking D = 1.18. It should be emphasized that this is the cu-
mulative number, whereas the model results given in Figure 16.16 are non-
cumulative.
The model results given in Figure 16.16 clearly illustrate the "Yellow-
stone Park" effect. After a massive forest fire covered a significant fraction
of Yellowstone National Park, it was argued that if smaller fires had been al-
lowed to burn, the massive forest fire could have been prevented. Allowing
small fires to burn is equivalent to having a larger sparking frequency. The
results given in Figure 16.16 illustrate how the small fires prevent the occur-
rence of catastrophic fires that burn essentially the entire model forest.
A variety of authors have studied forest-fire models, including Drossel
and Schwabl (1992a, b, 1993a, b, 1994), Mosner et al. (1992), Bak et al.
(1992), Drossel et al. (1993), Henley (1993), Christensen et al. (1993), and
Clar et al. (1994), and Strocka et al. (1995). Johansen (1994) has applied the
forest-fire model to the spread of diseases.
Problems
Problem 16.1. Consider the evolution of the cellular-automata model illus-

trated in Figure 16.l(b). (a) Which boxes have an additional particle in
steps, 6 , 7 , 9 and lo? (b) Which boxes are unstable and how many parti-
cles are lost from the grid in steps 8, I la, I lb, I lc, and 1Id?
steps 12, 13, 14, and 15? (b) Which boxes are unstable and how many
particles are lost from the grid in steps 16a, 16b, 16c, 16d, 17a, and 17b?
steps 18 and 19? (b) Which boxes are unstable and how many particles
are lost from the grid in steps 20a, 20b, 20c, 20d, and 20e?
Problem 16.4.
Consider a 2 X 2 grid of four boxes as illustrated above in (a). Also

given above in ( 6 ) is a sequence of random numbers in the range 1-4.
Use the random numbers to assign particles to boxes and carry out the
cellular automata model described in this chapter.
Problem 16.5.
Consider the linear grid of four boxes illustrated above. Use the se-
quence of random numbers given in Problem 16.4 to assign particles to
the four boxes. Use the following rules: When a box has two particles it
is unstable and they are redistributed to the two adjacent boxes. If either
of these boxes has two elements, they are again redistributed. Particles
are lost from the ends of the linear grid.
Problem 16.6. Consider the evolution of the forest-fire model illustrated in
Figure 16.15. Consider the configuration given in (d) and determine its
subsequent evolution using the random number sequence 96,09,35,67,
-13, 33, 94,44, 66, 37. (a) How many trees are planted? (b) How many
forest fires occur and how many trees are burned in them?
Problem 16.7. Consider the evolution of the forest-fire model illustrated in
Figure 16.15. Consider the configuration given in (d) and determine its
subsequent evolution using the random number sequence 15,81,55,25,
53, 65, 29, 17, 7 3 , s . (a) How many trees are planted? (b) How many
-
forest fires occur and how many trees are burned in them?
Problem 16.8. Consider a linear (one-dimensional) forest-fire model using a
grid of 10 points numbered sequentially from 0 to 9. Consider p = 4 so
that after three trees are planted on random points, a match is dropped on
a random point. Assume initially that trees are planted on points 1,3, and
5 and consider the random sequence 0, 1,7,2,3,2,6,4,0,7,7,4,9,4,7,
6. (a) Which points have trees after these 16 time steps? (b) How many
-
forest fires occurred and how many trees burned in each fire?
Chapter Seventeen
WHERE DO WE
STAND?
The concepts of fractals and chaos were introduced in 1967 (Mandelbrot,

1967) and in 1963 (Lorenz, 1963), respectively. Unlike many advances in
science, the attribution is in both cases quite clear. In both cases it took more
than 10 years before either concept received wide attention. Today, most sci-
entists take these concepts for granted, although their utility may be ques-
tioned. The concept of self-organized criticality was introduced in 1988
(Bak et al., 1988). Again the attribution is not open to question, but the defi-
nition of the concept remains somewhat unclear, particularly in.regard to
classical problems in criticality.
There is no question that fractals are a useful empirical tool. They pro-
vide a rational means for the extrapolation and interpolation of observations.
The normal distribution and the power-law (fractal) distribution have very
wide applicability. Just as the central limit theorem provides a broad base for
the application of the normal distribution, scale invariance provides a basis
for the application of the power-law distribution. Some statisticians argue
that fractals are a trivial extension of the Pareto distribution. But in its gen-
eral form, the Pareto distribution is not scale invariant and historically it has
been applied empirically without justification.
A strong case can certainly be made for the wide applicability of power-
law (fractal) distributions, but does this applicability have a more fundamen-
tal basis? Fractality appears to be fundamentally related to chaos and self-
organized criticality.
The chaotic behavior of deterministic nonlinear maps and sets of differ-
ential equations has been demonstrated beyond question. Chaotic behavior
has been quantified in terms of the Lyapunov exponent. Toy models have
been constructed that experimentally verify the basic concepts of chaos. It
can be argued convincingly that the chaotic behavior of low-order systems
demonstrate that weather, climate, seismicity, and so on are chaotic phenom-
ena. But the low-order examples of chaos are far removed from practical
problems. There are large gaps between the Lorenz equations and fluid tur-
342 WHERE DO WE STAND?
bulence, and between the Rikitate dynamo and the generation of the earth's
magnetic field.
The introduction of models that exhibit self-organized criticality has
been a major advance in extending concepts of chaos to higher-order sys-
tems. In this regard slider-block models play a central role. Slider-block
models were introduced as simple analog models for earthquakes. Distrib-
uted seismicity is taken to be a type example of self-organized critical be-
havior. The behavior of slider-block models is deterministic. Two slider
blocks exhibit the classical chaotic behavior of a low-order system. Large
numbers of slider blocks are self-organized critical. By systematically in-
creasing the number of blocks, the transition from chaotic to self-organized
critical behavior can be studied.
One of the present frontiers of research is to examine the relationship be-
tween models that exhibit self-organized criticality and the basic aspects of
statistical mechanics. For example, can earthquakes be better understood in
terms of the statistical fluctuations of a quasi-equilibrium system? Another
recent development is the recognition that complex fractal dimensions lead
to log-periodic behavior. It has been suggested that log-periodic behavior
may lead to a viable earthquake-prediction strategy.
Some would argue that the concepts covered in this book fall under the
broad umbrella of "complexity." But complexity is so broad a term that it de-
fies any all-encompassing definition. Certainly, many aspects of geology
and geophysics are complex; just as many problems in biology, economics,
and human behavior are complex. There are also links between important
problems in all these areas. This has led a number of scientists to propose a
new science of complexity. The science would include fractals, chaos, and
self-organized criticality. This is a major feature of the activities at the Santa
Fe Institute. But there has also been a strong reaction against "complexity"
with regard to its generality and a failure to deliver on promises made by
some of its practitioners. The entire area of fractals, chaos, self-organized
criticality, and complexity remains extremely active, and it is impossible to
predict with certainly what the future holds.
REFERENCES
Abe, K. (1981). Magnitudes of large shallow earthquakes from 1904 to 1980,

Phys. Earth Planet. Intr. 27,72-92.
Ahnert, F. (1984). Local relief and the height limits of mountain ranges, Am.
J. Sci. 284, 1035-55.
Aki, K. (1981). A probabilistic synthesis of precursory phenomena, in Earth-
quake Prediction, D . W . Simpson & P. G. Richards, eds., pp. 566-74,
American Geophysical Union, Washington, D.C.
Alexandrescu, M., Gibert, D., Hulot, G., Le Mouel, J. L. & Saracco, G. (1995).
Detection of geomagnetic jerks using wavelet analysis, J. Geophys. Res.
100, 12,557-72.
Allain, C. & Cloitre, M. (1991). Scaling rules in rock fracture and possible
implications for earthquake prediction, Nature 297,47-9.
Allbgre, C. J., Le Mouel, J. L. & Provost, A. (1982). Scaling rules in rock
fracture and possible implications for earthquake prediction, Nature
297,47-9.
Allkgre, C. J. & Lewin, E. (1995). Scaling laws and geochemical distribu-
tions, Earth Planet. Sci. Lett. 132, 1 - 1 3.
Alkgre, C. J. & Minster, J. F. (1978). Quantitative models of trace element
behavior, Earth Planet. Sci. Lett. 38, 1-25.
Allkgre, C. J., Treuil, M., Minster, J. F. Minster, B. & Albarede, F. (1975).
Systematic use of trace element in igneous process, Contrib. Min. Petrol.
60,57-75.
An, L. J. & Sammis, C. G. (1994). Particle size distribution of cataclastic fault
materials from southern California, Pure Appl. Geophys. 143,203-27.
Anderson, J . G. (1986). Seismic strain rates in the central and eastern United
States, Seis. Soc. Am. Bull. 76, 273-90.
Anderson, J. G. & Luco, L. E. (1983). Consequences of slip rate constraints
on earthquake occurrence relations, Seis. Soc. Am. Bull. 73,47 1-96.
Andrews, D. J. & Hanks, T. C. (1985). Scarp degraded by linear diffusion:
inverse solution for age, J. Geophys. Res. 90, 10193-208.
344 REFERENCES
Andrle, R. (1992). Estimating fractal dimension with the divider method in

geomorphology, Geomorph. 5,13 1 4 1 .
Bak, P., Chen, K. & Tang, C. (1992). A forest-fire model and some thoughts
on turbulence, Phys. Lett. A147,297-300.
Bak, P. & Tang, C. (1989). Earthquakes as a self-organized critical phenom-
enon, J. Geophys. Res. 94,15635-7.
Bak, P., Tang, C. & Wiesenfeld, K. (1988). Self-organized criticality, Phys.
Rev. A38,364-74.
Barabasi, A. L. & Stanley, H. E. (1995). Fractal Concepts in Sudace Growth,
Cambridge University Press, Cambridge, 366 pp.
Barenblatt, G. I., Zhivago, A. V., Nephrochnov, Y. P. & Ostrovskiy, A. A.
(1984). The fractal dimension: A quantitative characteristic of ocean-
bottom relief, Oceanology 34,695-7.
Barlow, R. (1989). Statistics, John Wiley & Sons, Chichester, 204 pp.
Barriere, B., & Turcotte, D. L. (1994). Seismicity and self-organized criti-
cality, Phys. Rev. E49, 11 5 1-60.
Barton, C. C. (1995). Fractal analysis of scaling and spatial clustering of
fractures, Fractals in the Earth Sciences, C. C. Barton & P. R. La Pointe,
eds., pp. 141-78, Plenum Press, New York.
Barton, C. C. & Scholz, C. H. (1995). The fractal size and spatial distribu-
tion of hydrocarbon accumulation, in Fractals in Petroleum Geology
and Earth Processes, C. C. Barton & P. R. La Pointe, eds., pp. 13-34,
Plenum Press, New York.
Barzini, G. N. & Ball, R. C. (1993). Landscape evolution in flood - a mathe-
matical model, J. Phys. A26,6777-87.
Bassingthwaighte, J. B. & Raymond, G. M. (1994). Evaluating rescaled
range analysis for time series, Ann. Biomedical Eng. 22,432-44.
Beer, T. & Borgas, M. (1993). Horton's laws and the fractal nature of streams,
Water Resoul: Res. 29,1475-87.
Belfield, W. C. (1994). Multifractal characteristics of natural fracture aper-
tures, Geophys. Res. Lett. 2 1 , 2 6 4 1 4
Bell, T. H. (1975). Statistical features of sea-floor topography, Deep-sea
Res. 22,883-92.
Bell, T. H. (1979). Mesoscale sea floor roughness, Deep-sea Res. 26A,
65-76.
Bennett, J. G. (1936). Broken coal, J. Inst. Fuel 10,22-39.
Benson, M. A. (1968). Uniform flood-frequency estimating methods for fed-
eral agencies, Water Resoul: Res. 4,891-908.
BergC, P., Poneau, Y. & Vidal, C. (1986). Order with Chaos, John Wiley &
Sons, New York, 329 pp.
Berkson, J. M. & Mathews, J. E. (1983). Statistical properties of sea-floor
roughness, in Acoustics and the Sea-bed, N . G . Pace, ed., pp. 215-23,
Bath University Press, Bath.
REFERENCES 345
Biegel, R. L., Sammis, C. G. & Dieterich, J. H. (1989). The frictional prop-

erties of a simulated gouge having a fractal particle distribution, J. Struct.
Geol. 11,82746.
Blanter, E. M. & Shnirman, M. G. (1994). On multifractal theory approach
to the clustering of epicenters, Comp. Seis. Geodynamics, 2,3745.
Blenkinsop, T. (1994). The fractal distribution of gold deposits: nkro examples
from the Zimbabwe Archean Craton, in Fractals and Dynamic Systems in
Geoscience, J. H . Kruhl, ed., pp. 247-58, Springer-Verlag,Berlin.
Boadu, F. K. & Long, L. T. (1994a). Statistical distribution of normal frac-
tures and the possible physical generating mechanism, Pure Appl. Geo-
phys., 142,273-93.
Boadu, F. K. & Long, L. T. (1994b). The fractal character of fracture spacing
and RQD, Znt. J. Rock Mech. Min. Sci., 31, 127-34.
Bodri, B. (1993). A fractal model for seismicity at Izu-Tokai region, central
Japan, Fractals 1,53946.
Box, G. E. P., Jenkins, G. M. & Reinsel, G. C. (1994). Time Series Analysis,
Prentice-Hall, Englewood Cliffs, N.J., 598 pp.
Brakensiek, D. L., Rawls, W. J., Logsdon, S. D. & Edwards, W. M.
(1992). Fractal description of macroporosity, Soil Sci. Soc. Am. J. 56,
1721-23.
Bras, R. L. & Rodriguez-Iturbe, I. (1993). Random Functions and Hydrol-
ogy, Dover, New York, 559 pp.
Bretz, M., Cunningham, J. B., Kurczynski, P. L. & Nori, F. (1992). Imaging
of avalanches in granular materials, Phys. Rev. Lett. 69,243 1 4 .
Brown, S. R. & Scholtz, S. H. (1985). Broad bandwidth study of the topog-
raphy of natural rock surfaces, J. Geophys. Res. 90,12575-82.
Brown, S. R., Scholz, C. H. & Rundle, J. B. (1991). A simplified spring-block
model of earthquakes, Geophys. Res. Lett. 18,2 15-8.
Bruno, B. C., Taylor, G. J., Rowland, S. K. & Baloga, S. M. (1994). Quantify-
ing the effect of rheology on lava-flow margins using fractal geometry,
Bull. Volcan. 56, 193-206.
Bruno, B. C., Taylor, G. J., Rowland, S. K., Lucey, P. G., & Self, S. (1992).
Lava flows are fractals, Geophys. Res. Lett. 19,305-8.
Bucknam, R. C. & Anderson, R. E. (1979). Estimation of fault-scarp ages
from a scarp-height-slope-angle relationship, Geology 7,ll-14.
Bufe, C. G., Nishenko, S. P. & Varnes, D. J. (1994). Seismicity trends and
potential for large earthquakes in the Alaska-Aleutian region, Pageoph.
142,83-99.
Bufe, C. G. & Varnes, D. J. (1993). Predictive modeling of the seismic cycle
of the greater San Francisco Bay region, J. Geophys. Res. 98,9871-83.
Burridge, R. & Knopoff, L. (1967). Model and theoretical seismicity, Seis.
Soc. Am. Bull. 57,34 1-7 1.
Byerlee, J. (1978). Friction of rocks, Pure Appl. Geophys. 116,615-26.
346 REFERENCES
Cao, T. & Aki, K. (1984). Seismicity simulation with a mass-spring model

and a displacement hardening-softening friction law, Pure Appl. Geophys.
122, 10-24.
Cao, T. & Aki, K. (1986). Seismicity simulation with a rate-and-state friction
law, Pure Appl. Geophys. 124,487-5 13.
Cargill, S. M. (1981). United States gold resource profile, Econ. Geol. 76,
937-43.
Cargill, S. M., Root, D. H. & Bailey, E. H. (1980). Resources estimation
from historical data: Mercury, a test case, J. Int. Assoc. Math. Geol. 12,
489-522.
Cargill, S. M., Root, D. H. & Bailey, E. H. (1981). Estimating unstable re-
sources from historical industrial data, Econ. Geol. 76, 1081-95.
Carlson, C. A. (1991). Spatial distribution of ore deposits, Geology 19, 111-4.
Carlson, J. M. (1991a). Two-dimensional model of a fault, Phys. Rev. A44,
6226-32.
Carlson, J. M. (1991b). Time intervals between characteristic earthquakes
and correlations with smaller events: An analysis based on a mechanical
model of a fault, J. Geophys. Res. 96,4255-67.
Carlson, J. M., Grannan, E. R. & Swindle, G. H. (1993a). Self-organizing
systems at finite driving rates, Phys. Rev. E47,93-105.
Carlson, J . M., Grannan, E. R., Singh, C. & Swindle, G. H. (1993b). Fluctu-
ations in self-organizing systems, Phys. Rev. E48,688-98.
Carlson, J. M. & Langer, J. S. (1989). Mechanical model of an earthquake
fault, Phys. Rev. A40,6470-84.
Carlson, J . M., Langer, J. S. & Shaw, B. E. (1994). Dynamics of earthquake
faults, Rev. Mod. Phys. 66,657-70.
Carlson, J . M., Langer, J. S., Shaw, B. E. &Tang, C. (1991). Intrinsic proper-
ties of a Burridge-Knopoff model of an earthquake fault, Phys. Rev.
A44,884-97.
Carmalt, S. W. & St. John, B. (1984). Giant oil and gas fields, in Future Pe-
troleum Provinces of the World, M . T. Halbouty, ed., pp. 11-53, Ameri-
can Association of Petroleum Geologists, Tulsa, Memoir 40.
Cazenave, A., Parsons, B., & Calcagno, P. (1995). Geoid lineations of I000 km
wavelength over the central Pacific, Geophys. Res. Lett. 22,97-100.
Chase, C. G. (1992). Fluvial landsculpting and the fractal dimension of
topography, Geomorphology 5,39-57.
Chatfield, C. (1989). The Analysis of Time Series, Chapman and Hall, Lon-
don, 241 pp.
Chen, J. D. & Wilkinson, D. (1985). Pore-scale viscous fingering in porous
media, Phys. Rev. Lett. 55, 1892-5.
Chen, K., Bak, P. & Obukhov, S. P. (1991). Self-organized criticality in a
crack propagation model of earthquakes, Phys. Rev. A43,625-30.
Chepil, W. S. (1950). Methods of estimating apparent density of discrete soil
grains and aggregates, Soil Sci. 70,35 1-62.
REFERENCES 347
Chillingworth, R. J. & Holmes, P. J. (1980). Dynamical systems and models

for reversals of the earth's magnetic field. Math. Geol. 12,41-59.
Chopard, B., Hermann, H. J. & Vicsek, T. (1991). Structure and growth mech-
anism of mineral dendrites, Nature 353,409-12.
Christensen, K., Flyvbjerb, H. & Olami, Z. (1993). Self-organized critical
forest-fire model: Mean-field theory and simulation results in 1 to 6 di-
mensions, Phys. Rev. Lett. 71,2737-40.
Christensen, K . & Olami, Z. (1992). Variation of the Gutenberg-Richter b
values and nontrivial temporal correlations in a spring-block model for
earthquakes, J. Geophys. Res. 97,8729-35.
Clar, S., Drossel, B. & Schwabl, F. (1994). Scaling laws and simulation re-
sults for the self-organized forest-fire model, Phys. Rev. E50, 1009-1 8.
Clark, G . B. (1987). Principles of Rock Fragmentation, John Wiley & Sons,
New York, 610 pp.
Cohen, S. (1977). Computer simulation of earthquakes, J. Geophys. Res. 82,
378 1-96.
Coleman, B. D. (1958). Statistics and time-dependence of mechanical break-
down in fibers, J. Appl. Phys. 29,968-83.
Cook, A. E. & Roberts, P. H. (1970). The Rikitake two-disc dynamo system,
Proc. Camb. Phil. Soc. 68,547-69.
Cortini, M. & Barton, C. C. (1993). Nonlinear forecasting analysis of infla-
tion-deflation patterns of an active caldera (Campi Flegrei, Italy), Geol-
ogy 21,239-42.
Cortini, M. & Barton, C. C. (1994). Chaos in geomagnetic reversal records:
A comparison between Earth's magnetic field data and model disk dy-
namo data, J. Geophys. Res. 99, l8,O2 1-33.
Cowie, P. A. & Scholz, C. H. (1992a). Displacement-length scaling relation-
ship for faults: Data synthesis and discussion. J. Struct. Geol. 14, 1149-56.
Cowie, P. A. & Scholz, C. H. (1992b). Growth of faults by accumulation of
seismic slip, J. Geophys. Res. 97, 11,085-95.
Cowie, P. A., Vanneste, C. & Sornette, D. (1993). Statistical physics model
for the spatiotemporal evolution of faults. J. Geophys. Res. 98, 21,
809-2 1.
Crovelli, R. A. & Barton, C. C. (1995). Fractals and the Pareto distribution
applied to petroleum accumulation-size distributions, in Fractals in Pe-
troleunz Geology and Earth Processes, C. C. Barton & P. R. La Pointe,
eds., pp. 59-72, Plenum Press, New York.
Culling, W. E. H. (1960). Analytical theory of erosion. J. Geol. 68,336-44.
Culling, W . E. H. (1963). Soil creep and the development of hillside slopes,
J. Geol. 71, 127-6 1 .
Culling, W. E. H. (1965). Theory of erosion on soil-covered slopes, J. Geol.
73,230-54.
Curl, R. L. (1986). Fractal dimensions and geometries of caves, Math. Geol.
18,765-83.
348 REFERENCES
Curran, D. R., Shockey, D. A., Seaman, L. & Austin, M. (1977). Mecha-

nisms and models of cratering in earth media, in Impact and Explosion
Cratering, D. J . Roddy, R. 0.Pepin & R. P. Merrill, eds., pp. 1057-87,
Pergamon Press, New York.
Daccord, G. & Lenormand, R. (1987). Fractal patterns from chemical disso-
lution, Nature 325,41-3.
Daubechies, I. (1990). The wavelet transform, time-frequency localization
and signal analysis, IEEE Trans. Info. Theory, 36,96 1-1005.
Davy, P., Sornette, A. & Sornette, D. (1990). Some consequences of a pro-
posed fractal nature of continental faulting, Nature 348,568.
Dawers, N. H., Anders, M. H. & Scholz, C. H. (1993). Growth of normal
faults: Displacement-length scaling, Geology 21, 1107-10.
Deakin, A. S. & Boxer, G. L. (1986). Argyle AK1 diamond size distribution:
The use of fine diamonds to predict the occurrence of commercial sized
diamonds, Geol. Soc. Aust. Special Publ. 24,1117-22.
de Sousa Vieira, M. (1992). Self-organized criticality in a deterministic me-
chanical model, Phys. Rev. A46,6288-93.
de Sousa Vieira, M., Vasconcelos, G. L. & Nagel, S. R. (1993). Dynamics of
spring-block models: Tuning to criticality, Phys. Rev. E47, R2221-4.
Dewey, J. F. (1975). Finite plate implications: Some implications for the
evolution of rock masses at plate margins, Am. J. Sci. 275A, 260-84.
De Wijs, H. J. (1951). Statistics of ore distribution. Part I: Frequency distrib-
ution of assay values, Geol. Mijnb. 13,365-75.
De Wijs, H. J. (1953). Statistics of ore distribution. Part 11: Theory of bino-
mial distribution applied to sampling and engineering problems, Geol.
Mijnb. 15,12-24.
Dieterich, J . H. (1972). Time-dependent friction as a possible mechanism for
aftershocks, J. Geophys. Res. 77,3771-81.
Dieterich, J. H. (1981). Constructive properties of faults with simulated
gouge, in Mechanical Behavior of Crustal Rocks, N. L. Carter, N. Fried-
man, J. M. Logan & D. W. Steans, eds., pp. 103-20, American Geophys-
ical Union, Geophysical Monograph 24.
Dietler, G. & Zhang, Y. C. (1992). Fractal aspects of the Swiss landscape,
Physica A191,213-19.
Ding, E. J. & Lu, Y. N. (1993). Analytical treatment for a spring-block model,
Phys. Rev. Lett. 70,3627-30.
Dongsheng, L., Zhaobi, Z. & Binghong, W. (1994). Research into the multi-
fractal of earthquake spatial distribution. Tectonophys. 233,91-7.
Drew, L. J., Schuenemeyer, J. H. & Bawiec, W. J. (1982). Estimation of the
future rates of oil and gas discoveries in the western gulf of Mexico,
U.S. Geological Survey Professional Paper 1252,26 pp.
Drossel, B., Clar, S. & Schwabl, F. (1993). Exact results for the one-di-
mensional self-organized critical forest-fire model, Phys. Rev. Lett. 71,
3739-42.
REFERENCES 349
Drossel, B. & Schwabl, F. (1992a). Self-organized criticality in a forest-fire

model, Phys. Rev. Lett. 69, 1629-32.
Drossel, B. & Schwabl, F. (1992b). Self-organized criticality in a forest-fire
model, Physica A191,47-50.
Drossel, B. & Schwabl, F. (1993a). Self-organization in a forest-fire model,
Fractals 1, 1022-9.
Drossel, B. & Schwabl, F. (1993b). Forest-fire model with immune trees,
Physica A199,183-97.
Drossel, B. & Schwabl F. (1994). Formation of space-time structure in a
forest-fire model, Physica A204,212-29.
Dubois, J. & Cheminee, J. L. (1991). Fractal analysis of eruptive activity of
some basaltic volcanoes, J. Volcan. Geotherm. Res. 45, 197-208.
Dubuc, B., Quiniou, J. F., Roques-Carmes, C., Tricot, C. & Zucker, S. W.
(1989a). Evaluating the fractal dimension of profiles, Phys. Rev. A39,
1500-12.
Dubuc, B., Zucker, S. W., Tricot, C., Quiniou, J. F. & Wehbi, D. (1989b). Eval-
uating the fractal dimension of surfaces, Proc. Roy. Soc. London A425,
113-27.
Dwight, H. B. (1961). Table of Integrals and Other Mathematical Data, 4th
ed., Macmillan, New York, 336 pp.
Dziewonski, A. M., Ekstrom, G., Woodhouse, J. H. & Zwart, G. (1989).
Centroid-moment tensor solutions for October-December 1987, Phys.
Earth Planet. Int. 54, 10-2 1.
Eneva, M. (1994). Monofractal or multifractal: A case study of spatial distri-
bution of mining induced seismic activity, Nonlinear Proc. Geophys. 1,
182-90.
Englman, R., Rivier, N. & Jaeger, Z. (1987). Fragment-size distribution in
disintegration by maximum-entropy formalism, Phil. Mag. B56, 75 1-
69.
Englman, R., Rivier, N. & Jaeger, Z. (1988). Size-distribution in sudden
breakage by the use of entropy maximization, J. Appl. Phys. 63,4766-8.
Espanol, P. (1994). Propagative slipping modes in a spring-block model,
Phys. Rev. E50,227-35.
Evernden, J . F. (1970). Study of regional seismicity and associated problems,
Seis. Soc. Am. Bull. 60,393-446.
Evesque, P. (1991). Analysis of statistics of sandpile avalanches using soil-
mechanics results and concepts, Phys. Rev. A43,2720-40.
Family, F. (1986). Scaling of rough surfaces: Effects of surface diffusion,
J. Phys. A19, L441-6.
Feder, J. (1988). Fractals, Plenum Press, New York, 283 pp.
Feder, H. J. S. & Feder, J. (1991). Self-organized criticality in a stick-slip
process, Phys. Rev. Lett. 66,2669-72.
Feder, J. & Jossang, T. (1995). Fractal patterns in porous media flow-model-
ing of laboratory experiments, in Fractals in Petroleum Geology and
350 REFERENCES
Earth Processes, C. C. Barton & P. R. La Pointe, eds., pp. 179-226.

Plenum Press, New York.
Flandrin, P. (1993). Fractional Brownian motion and wavelets, in Wavelets,
Fractals, and Fourier Transforms, M . Farge, J. C. R. Hunt & J. C. Vas-
silicos, eds., pp. 109-23, Clarendon Press, Oxford.
Fluigeman, R. H. & Snow, R. S. (1989). Fractal analysis of long-range pa-
leoclimatic data: Oxygen isotope record of Pacific Core V28-3 13, Pure
Appl. Geophys. 131,307-13.
Fowler, A. D. (1990). Self-organized mineral textures of igneous rocks: The
fractal approach, Earth-Sci. Rev. 29,47-55.
Fowler, A. D. (1995). Mineral crystallinity in igneous rocks - fractal method,
in Fractals in the Earth Sciences, C. Barton & P. R. La Pointe, eds., pp.
237-49, Plenum Press, New York.
Fowler, A. D., Stanley, H. E. & Daccord, G. (1989). Disequilibrium silicate
mineral textures: Fractal and non-fractal features, Nature 341, 134-8.
Fox, C. G. (1989). Empirically derived relationships between fractal dimen-
sion and power law from frequency spectra, Pure Appl. Geophys. 131,
307-13.
Fox, C. G. & Hayes, D. E. (1985). Quantitative methods for analysing the
roughness of the seafloor, Rev. Geophys. 23, 1-48.
Frohlich, C. & Davis, S. D. (1993). Teleseismic b values; or, much ado about
1.O, J. Geophys. Res. 98,63 1-44.
Fujiwara, A., Kamimoto, G. & Tsukamoto, A. (1977). Destruction of basaltic
bodies by high-velocity impact, Icarus 31,277-88.
Gabrielov, A. & Newman, W. I. (1994). Seismicity modeling and earthquake
prediction: A review, in Nonlinear Dynamics and Predictability of Geo-
physic Phenomena, W. I . Newman, A. Gabrielov & D. L. Turcotte, eds.,
pp. 7-13, Geophys. Mono. 83, American Geophysical Union, Washing-
ton, D.C.
Gaffin, S. (1989). Analysis of scaling in the geomagnetic polarity reversal
record, Phys. Earth Planet. Int. 57,284-90.
Gallant, J . C., Moore, I. D., Hutchinson, M. F. & Gessler, P. (1994). Estimat-
ing fractal dimensions of profiles: A comparison of methods, Math. Geol.
26,455-8 1.
Gambis, D. (1992). Wavelet transform analysis of the length of the day and
the El-NinoISouthern Oscillation variations at intraseasonal and inter-
annual time scales, Ann. Geophys., 10,429-37.
Gaonach, H., Lovejoy, S. & Stix, J. (1992). Scale invariance of basaltic lava
flows and their fractal dimensions, Geophys. Res. Lett. 19,785-88.
Garcia-Ruiz, J. M., Checa, A. & Rivas, P. (1990). On the origin of the am-
monite sutures, Paleobiology 16,349-54.
Garcia-Ruiz, J . M. & Otalora, F. (1992). Fractal trees and Horton's laws,
Math. Geol. 24,6 1-7 1 .
REFERENCES 351
Gardner, T. W., Jorgensen, D. W., Shuman, C. & Lemieux, C. R. (1987). Geo-

morphic and tectonic process rates: Effects of measured time interval,
Geology 15,259-6 1.
Gefen, Y., Aharony, A. & Mandelbrot, B. B. (1984). Phase transitions on
fractals: 111. Infinitely ramified lattices, J. Phys. A17, 1277-89.
Gilbert, L. E. (1989). Are topographic data sets fractal? Pure Appl. Geophys.
131,241-54.
Gilbert, L. E. & Malinverno, A. (1988). A characterization of the spectral
density of residual ocean floor topography, Geophys. Res. Lett. 15,
1401-4.
Gillespie, P. A., Walsh, J. J. & Watterson, J. (1992). Limitations of dimen-
sion and displacement data from single faults and the consequences for
data analysis and interpretations, J. Struct. Geol. 14, 1157-72.
Gillespie, P. A., Howard, C. B., Walsh, J. J. & Watterson, J. (1993). Measure-
ment and characterisation of spatial distributions of fractures, Tectono-
phys., 226, 113-4 1.
Glatzmaier, G. A. & Roberts, P. H. (1995). A three-dimensional self-consis-
tent computer simulation of a geomagnetic field reversal, Nature 377,
203-9.
Gomes, M. A. F. & Sales, T. R. M. (1993). Diffusive fragmentation model,
Phys. Lett. A177,2 15-9.
Gomes Da Silva, L. M. & Turcotte, D. L. (1994). A comparison between
Hurst and Hausdorff measures derived from fractional time series,
Chaos Solitons Fractals, 4,2 18 1-92.
Goodchild, M. F. (1980). Fractals and the accuracy of geographical mea-
sures, Math. Geol. 12,85-98.
Grady, D. E. & Kipp, M. E. (1987). Dynamic rock fragmentation, in Frac-
ture Mechanics of Rock, B. K . Atkinson, ed., pp. 429-75, Academic
Press, London.
Grassberger, P. & Procaccia, I. (1983a). Characterization of strange attrac-
tors, Phys. Rev. Lett. 50,346-9.
Grassberger, P. & Procaccia, I. (1983b). Measuring the strangeness of
strange attractors, Physica 9d, 189-208.
Greeley, R. & Gault, D. E. (1970). Precision size-frequency distribution of
craters for 12 selected areas of the lunar surface, The Moon 2, 10-77.
Green, A. R. (1964). Magnitude and frequency of floods, Part 1-A, U.S.
Geological Survey, Water Supply Paper 1671,2 12-3.
Grossman, A. & Morlet, J. (1984). Decomposition of Hardy functions into
square integrable wavelets of constant shape, SIAM J. Math. Anal., 15,
723-36.
Gu, J. C., Rice, J. R., Ruina, A. L. & Tse, S. T. (1984). Slip motion and sta-
bility of a single degree of freedom elastic system with rate and state de-
pendent friction, J. Phys. Solids 32, 167-96.
352 REFERENCES
Gudmundsson, A. (1987). Geometry, formation and development of tectonic

fractures on the Reykjanes Peninsula, southwest Iceland, Tectonophys.
139,295-08.
Gutenberg, B. & Richter, C. F. (1954). Seismicify of the Earth and Associ-
ated Phenomenon, 2nd ed., Princeton University Press, Princeton.
Hack, J. T. (1957). Studies of longitudinal stream profiles in Virginia and
Maryland, U.S. Geological Survey Professional Paper 294B, 97 pp.
Halsey, T. C., Jensen, M. H., Kadanoff, L. P., Procaccia, I. & Shraiman, B. I.
(1986). Fractal measures and their singularities: The characterization of
strange sets, Phys. Rev. A33,1141-51.
Halsey, T. C. & Leibig, M. (1992). Theory of branched growth, Phys. Rev.
A46,7793-09.
Hanks, T. C., Buckman, R. C., Lajoie, K. R. & Wallace, R. E. (1984). Modi-
fication of wave-cut and faulting-controlled landforms, J. Geophys. Res.
89,577 1-90.
Hanks, T. C. & Kanamori, H. (1979). A moment-magnitude scale, J. Geo-
phys. Res. 84,2348-50.
Hanks, T. C. & Wallace, R. E. (1985). Morphological analysis of the Lake
Lahontan shoreline and beachfront fault scarps, Pershing County, Nevada,
Seis. Soc. Am. Bull. 75,835-46.
Hansen, J. P. & Skjeltorp, A. T. (1988). Fractal pore space and rock perme-
ability implications, Phys. Rev. B38,2635-8.
Harlow, D . G. & Phoenix, S. L. (1982). Probability distributions for the
strength of fibrous materials under load sharing, 1. Two-level failure and
edge effects, Adv. Appl. Prob. 14,68-94.
Harris, C . , Franssen, R. & Loosveld, R. (1991). Fractal analysis of frac-
tures in rocks: The Cantor dust method - comment, Tectonophys., 198,
107-15.
Harris, D. P. (1984). Mineral Resources Appraisal, Oxford University Press,
Oxford, 445 pp.
Hart, S. R. (1993). Equilibration during mantle melting: A fractal tree model,
Proc. Natl. Acad. Sci. 90, 11,914-8.
Hartmann, W. K. (1969). Terrestrial, lunar, and interplanetary rock fragmen-
tation, kurus 10,201-13.
Held, G. A., Solina, D. H., Keane, D. T., Haag, W. J., Horn, P. M. & Grim-
stein, G. (1990). Experimental study of critical-mass fluctuations in an
evolving sandpile, Phys. Rev. Lett. 65, 1120-3.
Henderson, J. R., Main, I. G., Maclean, C. & Norman, M. G. (1994). A frac-
ture-mechanical cellular automation model of seismicity, Pure Appl.
Geophys. 142,545-65.
Henley, C. L. (1993). Statics of a "self-organized percolation model, Phys.
Rev. Lett. 7 1 , 2 7 4 1 4
REFERENCES 353
Hewett, T. A. (1986). Fractal distributions of reservoir heterogeneity and

their influence on fluid transport, Society of Petroleum Engineers Paper
15386, 15 pp.
Hill, D. P., Reasenberg, P. A., Michael, A., Arabaz, W. J., Beroza, G., Brum-
baugh, D., Brune, J. N., Castro, R., Davis, S., dePolo, D., Ellsworth, W. L.,
Gomberg, J., Harmsen, S., House, L., Jackson, S. M., Johnston, M. J. S.,
Jones, L., Keller, R., Malone, S., Munguia, L., Nava, S., Pechmann, J. C.,
Sanford, A., Simpson, R. W., Smith, R. B., Stark, M., Stickney, M., Vidal,
A., Walter, S., Wong, V. & Zollweg, J. (1993). Seismicity remotely trig-
gered by the magnitude 7.3 Landers, California, earthquake, Science 260,
1617-23.
Hirabayashi, T., Ito, K. & Yoshii, T. (1992). Multifractal analysis of earth-
quakes, Pure Appl. Geophys. 138,591-6 10.
Hirata, T. (1989a). Fractal dimension of fault systems in Japan: Fractal struc-
ture in rock fracture geometry at various scales, Pure Appl. Geophys.
131, 157-70.
Hirata, T. (1989b). A correlation between the b value and the fractal dimen-
sion of earthquakes, J. Geophys. Res. 94,7507-14.
Hirata, T. & Imoto M. (1991). Multifractal analysis of spatial distribution
of microearthquakes in the Kanto region, Geophys. J. Int. 107, 155-
62.
Hirata T., Satoh, T. & Ito, K. (1987). Fractal structure of spatial distribution
of microfracturing in rock, Geophys. J. Roy. Astron. Soc. 90,369-74.
Hiscott, R. N., Colella, A., Pezard, P., Lovell, M. A. & Malinverno, A. (1992).
Sedimentology of deep-water, volcaniclastics, Oligocene Izu-Bonin fore-
arc basin, based on formation microscanner images, Proc. Ocean Dril.
Prog. Sci. Results 126,75-96.
Hooge, C., Lovejoy, S., Pecknold, S., Malouin, J. F. & Schertzer, D. (1994a).
Universal multifractals in seismicity, Fractals 2,445-9.
Hooge, C., Lovejoy, S., Schertzer, D., Pecknold, S., Malouin, J. F. &
Schmitt, F. (1994b). Multifractal phase transitions: The origin of self-
organized criticality earthquakes, Nonlinear Proc. Geophys. 1, 191-7.
Horton, R. E. (1945). Erosional development of streams and their drainage
basins: Hydrophysical approach to quantitative morphology, Geol. Soc.
Am. Bull. 56,275-370
Hoshi, M. & Kono, M. (1988). Rikitake two-disk dynamo system: Statistical
properties and growth of instability, J. Geophys. Res. 93, 11,643-54.
Howard, A. D. (1971). Simulation model of stream capture, Geol. Soc. Am.
Bull. 82, 1355-76.
Hsui, A. T., Rust, K. A. & Klein, G. D. (1993). A fractal analysis of quater-
nary, Cenozoic-Mesozoic, and Late Pennsylvanian sea level changes,
J. Geophys. Res. 98,2 1,963-67.
354 REFERENCES
Huang, J., Narkounskaia, G., & Turcotte, D. L. (1992). A cellular auto-

maton, slider-block model for earthquakes. 2. Demonstration of self-
organized criticality for a two-dimensional system, Geophys. J . Int.
111,259-69.
Huang, J. & Turcotte, D. L. (1988). Fractal distributions of stress and strength
and variations of b-value, Earth Planet. Sci. Lett. 91,223-30.
Huang, J. & Turcotte, D. L. (1989). Fractal mapping of digitized images:
Application to the topography of Arizona and comparisons with syn-
thetic images, J. Geophys. Res. 94,7491-5.
Huang, J. & Turcotte, D. L. (1990a). Are earthquakes an example of deter-
ministic chaos? Geophys. Res. Lett. 17,223-6.
Huang, J. & Turcotte, D. L. (1990b). Fractal image analysis: Application to the
topography of Oregon and synthetic images, J. Opt. Soc. Am. A7,1124-30.
Huang, J. & Turcotte, D. L. (1990~).Evidence for chaotic fault interaction in
the seismicity of the San Andreas fault and Nakai trough, Nature 348,
234-6.
Huang, J. & Turcotte, D. L. (1992). Chaotic seismic faulting with a mass-
spring model and velocity-weakening friction, Pure Appl. Geophys. 138,
569-89.
Hurst, H. E., Black, R. P. & Simaika, Y. M. (1965). Long-term Storage, Con-
stable, London, 145 pp.
Hyndman, R. D. & Weichert, D. H. (1983). Seismicity and rates of relative
motion on the plate boundaries of western North America, Geophys. J.
Roy. Astron. Soc. 72,59-82.
Ijiri, Y. & Simon, H. A. (1977). Skew Distributions and the Sizes of Business
Firms, North-Holland, Amsterdam, 23 1 pp.
Ijjasz-Vasquez, E. J., Rodriguez-Iturbe, I. & Bras, R. L. (1992). On the mul-
tifractal characterization of river basins, Geomorph. 5,297-3 10.
Ito, K. (1980). Chaos in the Rikitake two-disk dynamo system, Earth Planet.
Sci. Lett. 51,45 1-6.
Ito, K. & Matsuzaki, M. (1990). Earthquakes as self-organized critical phe-
nomena, J. Geophys. Res. 95,6853-68.
Ivanov, S. S. (1994a). Global relief: Evidence of fractal geometry, in Frac-
tals and Dynamic Systems in Geoscience, J. H. Kruhl, ed., pp. 221-30,
Springer-Verlag, Berlin.
Ivanov, S. S. (1994b). "Counter-scaling'' method for estimation of fractal
properties of self affine objects, in Fractals and Dynamic Systems in
Geoscience, J. H . Kruhl, ed., pp. 391-7, Springer-Verlag, Berlin.
Jackson, P. & Sanderson, D. J. (1992). Scaling of fault displacements from
the Badajoz-Cordoba shear zone, SW Spain, Tectonophys. 210, 179-90.
Johansen, A. (1994). Spatio-temporal self-organization in a model of disease
spreading, Physica D78, 186-93.
REFERENCES 355
Johnston, A . C. & Nava, S. J. (1985). Recurrence rates and probability esti-

mates for the New Madrid seismic zone, J. Geophys. Res. 90,6737-53.
Jones, L. M. & Molnar, P. (1979). Some characteristics of foreshocks and
their possible relationship to earthquake prediction and premonitory slip
of faults, J. Geophys. Res. 84,3596-608.
Jurdy, D. M. & Stefanick, M. (1990). Models for the hotspot distribution,
Geophys. Res. Lett. 17, 1965-8.
Kadanoff, L. P., Nagel, S. R., Wu, L. & Zhou, S. M. (1989). Scaling and uni-
versality in avalanches, Phys. Rev. A39,6524-33.
Kagan, Y. Y. & Knopoff, L. (1980). Spatial distribution of earthquakes: The
two-point distribution problem, Geophys. J. Roy. Astron. Soc. 62,303-20.
Kanamori, H. (1978). Quantification of earthquakes, Nature 271,411-14.
Kanamori, H . (1981). The nature of seismicity patterns before large earth-
quakes, in Earthquake Prediction, D. W. Simpson and P. G. Richards,
eds., pp. 1-19, American Geophysical Union, Washington, D.C.
Kanamori, H. & Anderson, D. L. (1975). Theoretical basis of some empirical
relations in seismology, Seis. Soc. Am. Bull. 65, 1073-96.
Katz, A. J. & Thompson, A. H. (1985). Fractal sandstone pores: Implications
for conductivity and pore formation, Phys. Rev. Lett. 54, 1325-8.
Kaye, B. H. (1989). A Random Walk Through Fractal Dimensions, VCH,
Weinheim, 42 1 pp.
Kaye, B. (1993). Chaos and Complexity, VCH, Weinheim, 593 pp.
Keilis-Borok, V. I. (1990). The lithosphere of the earth as a nonlinear system
with implications for earthquake prediction. Rev. Geophys. 28, 19-34.
Keilis-Borok, V. I. & Kossobokov, V. G. (1990). Premonitory activation of
earthquake flow. Algorithm M8, Phys. Earth Planet. Int. 61,73-83.
Keilis-Borok, V. I. & Rotwain, I. M. (1990). Diagnosis of time of increased
probability of strong earthquakes in different regions of the world: Algo-
rithm CN, Phys. Earth Planet. Int. 61,57-72.
Kent, C. & Wong, J. (1982). An index of littoral zone complexity and its
measurement, Can. J. Fish. Aquat. Sci. 39,847-63.
Kenyon, P. M. & Turcotte, D. L. (1985). Morphology of a delta prograding
by bulk sediment transport, Geol. Soc. Am. Bull. 96, 1457-65.
King, G. (1983). The accommodation of large strains in the upper litho-
sphere of the earth and other solids by self-similar fault systems: The
geometrical origin of b-value, Pure Appl. Geophys. 121,76 1-8 15.
King, G. C. P. (1986). Speculations on the geometry of the initiation and ter-
mination processes of earthquake rupture and its relation to morphology
and geological structure, Pure Appl. Geophys. 124,567-85.
King, G. C. P., Stein, R. S. & Rundle, J. B. (1988). The growth of geological
structures by repeated earthquakes 1. Conceptual framework, J. Geo-
phys. Res. 93, 13307-1 8.
356 REFERENCES
Klacka, J. (1992). Mass distribution in the asteroid belt, Earth Moon Planets
56,47-52.
Klement, S., Kratky, K. W. & Nittmann J. (1993). Practical time-series
analysis with multifractal methods, Fractals 1,735-43.
Knopoff, L., Landoni, J. A. & Abinante, M. S. (1993). Dynamic model of an
earthquake fault with localization, Phys. Rev. A46,7445-9.
Kondev, J. & Henley, C. L. (1995). Geometrical exponents of contour loops
on random Gaussian surfaces, Phys. Rev. Lett. 74,4580-3.
Korcak, J. (1940). Deux types fondamentaux de distribution statistique, Bull.
Insti. Int. Stat. 30,295-9.
Korvin, G. (1992). Fractal Models in the Earth Sciences, Elsevier, Amsterdam,
381 pp.
Kramer, S. & Marder, M. (1992). Evolution of river networks, Phys. Rev.
Lett. 68,205-8.
Krause, F. & Schmidt, J. J. (1988). A low-dimensional attractor for model-
ling the reversals of the earth's magnetic field, Phys. Earth Planet. Int.
52,23-9.
Krohn, C. E. (1988a). Fractal measurements of sandstones, shales, and car-
bonates, J. Geophys. Res. 93,3297-305.
Krohn, C. E. (1988b). Sandstone fractal and Euclidean pore volume distribu-
tions, J. Geophys. Res. 93,3286-96.
Krohn, C. E. & Thompson, A. H. (1986). Fractal sandstone pores: Auto-
mated measurements using scanning-electron microscope images, Phys.
Rev. B33,6366-74.
Kucinskas, A. B. & Turcotte, D. L. (1994). Isostatic compensation of equato-
rial highlands on Venus, Icarus, 112, 104-16.
Kucinskas, A. B., Turcotte, D. L., Huang, J. & Ford, P. G. (1992). Fractal analy-
sis of Venus topography in Tinatin Planitia and Ovda Regio, J. Geophys.
Res. 97, 13,635-4 1.
Langer, J. S. & Tang, C. (1991). Rupture propagation in a model of an earth-
quake fault, Phys. Rev. Lett. 67, 1043-6.
La Pointe, P. R. (1988). A method to characterize fracture density and connec-
tivity through fractal geometry, Int. J. Rock Mech. Min. Sci. 25,421-9.
La Pointe, P. R. (1995). Estimation of undiscovered hydrocarbon potential
through fractal geometry, in Fractals in Petroleum Geology and Earth
Process, C . C . Barton and P. R. La Pointe, eds., pp. 35-37, Plenum
Press, New York.
Lasky, S. G. (1950). How tonnage and grade relations help predict ore re-
serves, Eng. Mining J. 151,8 1-5.
Laverty, M. (1987). Fractals in karst, Earth Surface Proc. Landfomzs 12,475-80.
Leary, P. & Abercrombie, R. (1994). Fractal fracture scattering origin of
S-wave coda: Spectral evidence from recordings at 2.5 km, Geophys.
Res. Lett. 21, 1683-6.
REFERENCES 357
Ledesen, B., Dubois, J., Genter, A. & Meunier, A. (1993). Fractal analysis of
fractures applied to Soultz-sous-Forets hot dry rock geothermal pro-
gram, J. Volcan. Geotherm. Res. 57, 1-17.
Lee, H. K. & Schwarcz, H. P. (1995). Fractal clustering of fault activity in
California, Geology 23,377-80.
Leheny, R. L. & Nagel, S. R. (1993). Model for the evolution of river net-
works, Phys. Rev. Lett. 71, 1470-3.
Leopold, L. B. & Langbein, W. B. (1962). The concept of entropy in land-
scape evolution, U.S. Geological Survey Paper 500A, 20 pp.
Lifton, N. A. & Chase, C. G. (1992). Tectonic, climatic and lithologic influ-
ences on landscape fractal dimension and hypsometry: Implications for
landscape evolution in the San Gabriel Mountains, California, Geo-
morph. 5,77-114.
Lin, B. & Taylor, P. L. (1994). Model of spatiotemporal dynamics of stick-
slip motion, Phys. Rev. E49,3940-7.
Lin, B . & Yang, Z. R. (1986). A suggested lacunarity expression for Sierpin-
ski carpets, J. Phys. A19, L49-52.
Little, S. A. (1994). Wavelet analysis of seafloor bathymetry: An example, in
Wavelets in Geophysics, E. Foufoula-Georgiou & P. Kumar, eds., pp.
167-82, Academic Press, New York.
Little, S. A., Carter, P. H. & Smith, D. K. (1993). Wavelet analysis of
bathymetric profile reveals anomalous crust, Geophys. Res. Lett. 20,
1915-8.
Liu, T. (1992). Fractal structure and properties of stream networks, Water
Resoul: Res. 28,298 1-8.
Lorenz, E. N. (1963). Deterministic nonperiodic flow, J. Atmos. Sci. 20,
130-41.
Lou, M. & Rial, J. A. (1995). Applications of the wavelet transform in de-
tecting multiple events of microearthquake seismograms, Geophys. Res.
Lett. 22,2 199-202.
Lovejoy, S., Lavallee, D., Schertzer, D. & Ladoy, P. (1995). The 1'12 law and
multifractal topography: Theory and analysis, Nonlinear Proc. Geophys.
2, 16-22.
Lu, Y. N., Liu, W. S. & Ding, E. J. (1994). Hysteresis in a theoretical spring-
block model, Phys. Rev. Lett. 72,4005-8.
Madden, T. R. (1983). Microcrack connectivity in rocks: A renormalization
group approach to the critical phenomena of conduction and failure in
crystalline rocks, J. Geophys. Res. 88, 585-92.
Magde, L. S., Dick, H. J. B. & Hart, S. R. (1995). Tectonics, alteration and
the fractal distribution of hydrothermal veins in the lower ocean crust,
Earth Planet Sci. Lett. 129, 103-19.
Main, I., Peacock, S. & Meredith, P. G. (1990). Scattering attenuation and the
fractal geometry of fracture systems, Pure Appl. Geophys. 133,283-304.
35 8 REFERENCES
Malinverno, A. (1989). Testing linear models of sea-floor topography, Pure

Malinverno, A. (1990). A simple method to estimate the fractal dimension of
a self-affine series, Geophys. Res. Lett. 17, 1953-6.
Malinverno, A. (1995). Fractals and ocean floor topography: A review and a
model, in Fractals in the Earth Sciences, C . Barton & P. R. La Pointe,
eds., pp. 107-30. Plenum Press, New York.
Maloy, K. J., Feder, J. & Jossang, T. (1985). Viscous fingering fractals in
porous media, Phys. Rev. Lett. 55,2688-9 1.
Mandelbrot, B. B. (1967). How long is the coast of Britain? Statistical self-
similarity and fractional dimension, Science 156,636-8.
Mandelbrot, B. B. (1975). Stochastic models for the earth's relief, the shape
and the fractal dimension of the coastlines, and the number-area rule for
islands, Proc. Nut. Acad. Sci. USA 72,3825-8.
Mandelbrot, B. B. (1982). The Fractal Geometry of Nature, Freeman, San
Francisco.
Mandelbrot, B. B. (1985). Self-affine fractals and fractal dimension, Physica
Scripta 32,257-60.
Mandelbrot, B. B. (1989). Multifractal measures, especially for the geo-
physicist, Pure Appl. Geophys. 1 3 1 , 5 4 2 .
Mandelbrot, B. B. & Wallis, J. R. (1968). Noah, Joseph, and operational hy-
drology, Water Resour: Res. 4,909-18.
Mandelbrot, B. B. &Wallis, J. R. (1969a). Computer experiments with frac-
tional Gaussian noises, Parts I, 11,111, Water Resour. Res. 5,228-67.
Mandelbrot, B. B. & Wallis, J. R. (1969b). Robustness of the rescaled range
R/S in the measurement of noncyclic long run statistical dependence,
Water Resour: Res. 5,967-88.
Manning, C . E. (1994). Fractal clustering of metamorphic veins, Geology 22,
335-8.
Mareschal, J. C. (1989). Fractal reconstruction of sea-floor topography, Pure
Marrett, R. & Allmendinger, R. W. (1991). Estimates of strain due to brittle
faulting: Sampling of fault populations, J. Struct. Geol. 13,735-8.
Marrett, R. & Allmendinger, R. W. (1992). Amount of extension on "small"
faults: An example from the Viking graben, Geology 20,47-50.
Marzocchi, W., Gonzato, G. & Mulargia, F. (1995). Rikitake's geodynamo
model analysed in terms of classical time series statistics, Phys. Earth
Planet. Int. 88, 83-8.
Masek, J. G. & Turcotte, D. L. (1993). A diffusion-limited aggregation
model for the evolution of drainage networks, Earth Planet. Sci. Lett.
119,379-86.
Matsuzaki, M. & Takayasu, H. (1991). Fractal features of the earthquake
phenomenon and a simple mechanical model, J. Geophys. Res. 96, 19,
925-3 1.
REFERENCES 359
Maxwell, J. C. (1960). Quantitative geomorphology of the San Dimas exper-

imental forest, California, Project NR 389-042, Columbia Univ. Dep.
Geol. Tech. Rep. 19.
May, R. M. (1976). Simple mathematical models with very complicated dy-
namics, Nature 261,459-67.
May, R. M. & Oster, G. F. (1976). Bifurcation and dynamic complexity in
simple ecological models, Am. Natur: 110,573-99.
Mayer, L. (1984). Dating quaternary fault scarps formed in alluvium using
morphologic parameters, Quat. Res. 22,300- 13.
McClelland, L., Simkin, T., Summers, M., Nielson, E. & Stein, T. C., eds.,
(1989). Global Volcanism 1975-1985, Prentice-Hall, Englewood Cliffs,
N.J., 655 pp.
McCloskey, J. (1993). A hierarchical model for earthquake generation on
coupled segments of a transform fault, Geophys. J. Int. 115,538-5 1.
McCloskey, J. & Bean, C. J. (1992). Time and magnitude predictions in
shocks due to chaotic fault interactions, Geophys. Res. Lett. 19, 119-22.
McCloskey, J. & Bean, C. J. (1994). Temporally unstable recurrence of earth-
quakes due to breaks in fractal scaling, Science 266,410-2.
Meakin, P., Feder, J. & Jossang, T. (1991). Simple statistical models for river
networks, Physica A176,409-29.
Meneveau, C. & Sreenivasan, K. R. (1987). Simple multifractal cascade
model for fully developed turbulence, Phys. Rev. Lett. 59, 1424-7.
Merceron, T. & Velde, B. (1991). Application of Cantor's method for fractal
analysis of fractures in the Toyoha Mine, Hokkaido, Japan, J. Geophys.
Res. 96, l6,64 1-50.
Meybeck, M. (1995). Global distribution of lakes, in Physics and Chemistry
of Lakes, 2nd ed., A. Lerman, D. M. Imboden & J. R. Gat, eds., pp. 1-35,
Molchan, G. M., Dmitrieva, 0 . E., Rotwain, I. M. & Dewey, J. (1990). Sta-
tistical analysis of the results of earthquake prediction, based on bursts
of aftershocks, Phys. Earth Planet. Int. 61, 128-39.
Molnar, P. (1979). Earthquake recurrence intervals and plate tectonics, Seis.
Soc. Am. Bull. 69, 115-33.
Molz, F. J. & Bowman, G. K. (1993). A fractal-based stochastic interpolation
scheme in subsurface hydrology, Water Resour. Res. 29,3769-74.
Morse, P. M. (1969). Thermal Physics, 2nd ed., W .A. Benjamin, New York.
Mosner, W. K., Drossel, B. & Schwabl, F. (1992). Computer simulations of
the forest-fire model, Physica A190,205-17.
Muller, J. (1992). Multifractal characterization of petrophysical data, Phys-
ica A191,284-8.
Muller, J. (1994). Characterization of the North Sea chalk by multifractal
analysis, J. Geophys. Res. 99,7275-80.
Musgrove, P. A. (1965). Lead: Grade-tonnage relation, Mining Mag. 112,
249-5 1.
360 REFERENCES
Nagatani, T. (1993). Dynamic scaling of river-size distribution in the Schei-

degger's river network model, Fractals 1,247-52.
Nagel, S. R. (1992). Instabilities in a sandpile, Rev. Mod. Phys., 64,321-5.
Nakanishi, H . (1990). Cellular-automaton model of earthquakes with deter-
ministic dynamics, Phys. Rev. A41,7086-9.
Nakanishi, H. (1991). Statistical properties of the cellular automata model
for earthquakes, Phys. Rev. A43,6613-2 1.
Narkounskaia, G., Huang, J. & Turcotte, D. L. (1992). Chaotic and self-or-
ganized critical behavior of a generalized slider-block model, J. Stat.
Phys. 67,115 1-83.
Narkounskaia, G . & Turcotte, D. L. (1992). A cellular-automata, slider block
model for earthquakes, I. Demonstration of chaotic behavior for a low
order system, Geophys. J. Int. 111,250-8.
Nash, D. B. (1980a). Morphologic dating of degraded normal fault scarps,
J. Geol. 88,353-60.
Nash, D. (1980b).Forms of bluffs degraded for different lengths of time in Em-
met County, Michigan, USA, Earth Sugace Proc. Landforms 5,33145.
Newman, W. I. & Gabrielov, A. M. (1991). Failure of hierarchical distribu-
tions of fibre bundles, I, Int. J. Fracture 50, 1-14.
Newman, W. I., Gabrielov, A. M., Durand, T. A., Phoenix, S. L. & Turcotte,
D. L. (1994). An exact renormalization model for earthquakes and mate-
rial failure. Statics and dynamics, Physica D77,200-16.
Newman, W. I. & Turcotte, D. L. (1990). Cascade model for fluvial geomor-
phology, Geophys. J. Int. 100,433-9.
Newman, W. I., Turcotte, D. L. & Gabrielov, A. M. (1995). Log-periodic be-
havior of a hierarchical failure model with applications to precursory
seismic activation, Phys. Rev. E52,4827-35.
Nicholl, M. J., Wheatcraft, S. W., Tyler, S. W. & Berkowitz, B. (1994). Is
Old Faithful a strange attractor?, J. Geophys. Res. 99,4495-503.
Nicolis, C . & Nicolis, G. (1984). Is there a climatic attractor? Nature 311,
529-32.
Nikora, V. I. & Sapozhnikov, V. B. (1993). River network fractal geometry
and its computer simulation, Water Resour. Res. 29,3569-75.
Nittmann, J., Daccord, G. & Stanley, H. E. (1985). Fractal growth of viscous
fingers: Quantitative characterization of a fluid instability phenomenon,
Nature 314, 141-4.
Nittmann, J., Daccord, G. & Stanley, H. E. (1986). When are viscous fingers
fractal?, in Fractals in Physics, L. Pietronero & E. Tosatt, eds., pp.
193-202, Elsevier, Amsterdam.
Nussbaum, J. & Ruina, A. (1987). A two degree-of-freedom earthquake
model with staticldynamic friction, Pure Appl. Geophys. 124,629-56.
Nuttli, 0. W. (1983). Average seismic source-parameter relations for mid-
plate earthquakes, Seis. Soc. Am. Bull. 73,5 19-35.
REFERENCES 361
Olami, Z. & Christensen, K. (1992). Temporal correlations, universality, and

multifractality in a spring-block model of earthquakes, Phys. Rev. A46,
R1720-3.
Olami, Z., Feder, H. J. S. & Christensen, K. (1992). Self-organized critical-
ity in a continuous, nonconservative cellular automaton modeling earth-
quakes, Phys. Rev. Lett. 68, 1244-7.
Osborne, A. R. & Provenzale, A. (1989). Finite correlation dimension for
stochastic systems with power-law spectra, Physica D35,357-8 1.
Ossadnik, P. (1992). Branch order and ramification analysis of large diffu-
sion limited aggregation clusters, Phys. Rev. A45, 1058-66.
Pacheco, J. F., Scholz, C. H. & Sykes, L. R. (1992). Changes in frequency-
size relationship from small to large earthquakes, Nature 355,7 1-3.
Palmer, A. C. & Sanderson, T. J. 0. (1991). Fractal crushing of ice and brittle
solids, Proc. Roy. Soc. A433,469-77.
Peckham, S. D. (1995). New results for self-similar trees with applications
to river networks, Water Resoul: Res. 31,1023-9.
Pelletier, J. D. & Turcotte, D. L. (1997). Scale-invariant porosity variations
in sedimentary basins, J. Geophys. Res., in press.
Pepke, S. L. & Carlson, J. M. (1994). Predictability of self-organizing sys-
tems, Phys. Rev. E50, 236-42.
Pepke, S. L., Carlson, J. R. & Shaw, B. E. (1994). Prediction of large events
on a dynamical model of a fault, J. Geophys. Res. 99,6769-88.
Perfect, E., Kay, B. D. & Rasiah, V. (1993). Multifractal model for soil ag-
gregate fragmentation, Soil Sci. Soc. Am. J. 57,896-900.
Pfeiffer, P. & Obert, M. (1989). Fractals: Basic concepts and terminology, in
The Fractal Approach to Heterogeneous Chemistry, D. Avnir, ed., pp.
11-43, John Wiley & Sons, Chichester.
Pickering, G., Bull, J. M., Sanderson, D. J. & Harrison, P. V. (1994). Fractal
fault displacements: A case study from the Moray Firth, Scotland, in
Fractals and Dynamic Systems in Geoscience, J . H. Kruhl, ed., pp.
105-19, Springer-Verlag, Berlin.
Plotnick, R. E. (1986). A fractal model for the distribution of stratigraphic
hiatuses, J. Geol. 94,885-90.
Plotnick, R. E., Gardner, R. H. & O'Neill, R. V. (1993). Lacunarity indices
as measures of landscape texture, Landscape Ecol. 8,201-1 1 .
Power, W. L. & Tullis, T. E. (1991). Euclidean and fractal models for the de-
scription of rock surface roughness, J. Geophys. Res. 96,415-24.
Power, W. L. & Tullis, T. E. (1995). Review of the fractal character of nat-
ural fault surfaces with implications for friction and the evolution of
fault zones, in Fractals in the Earth Sciences, C . Barton & P. R. La
Pointe, eds., pp. 89-105, Plenum Press, New York.
Puhl, H. (1992). On the modelling of real sand piles, Physica A182,
295-3 19.
362 REFERENCES
Purcaru, G. & Berckhemer, H. (1982). Quantitative relations of seismic

source parameters and a classification of earthquakes, Tectonophys. 84,
57-128.
Pyrak-Nolte, L. J., Myer, L. R. & Nolte, D. D. (1992). Fractures: Finite-size
scaling and multifractals, Pure Appl. Geophys. 138,679-706.
Pyrak-Nolte, L. J., Nolte, D. D. & Cook, N. G. W. (1995). Hierarchical cas-
cades and the single fracture-percolation and seismic detection, in Frac-
tals in Petroleum Geology and Earth Processes, C . C . Barton & P. R. La
Pointe, eds., pp. 143-78, Plenum Press, New York.
Rapp, R. H. (1989). The decay of the spectrum of the gravitational potential
and the Earth, Geophys. J. Int. 99,449-55.
Rayleigh, Lord (1916). On convection currents in a horizontal layer of fluid
when the higher temperature is on the underside, Phil. Mag. 32,529-46.
Reams, R. W. (1992). Fractal dimensions of sinkholes, Geomorph. 5,
159-65.
Redner, S. (1990). Fragmentation, in Statistical Models for the Fracture of
Disordered Media, H. J . Herman & S. Roux, eds., pp. 321-48. North-
Holland, Amsterdam.
Reigber, C. Balmino, G., Muller, H., Bosch, W. & Moynot, B. (1985). GRIM
gravity model improvement using LAGEOS (GRIM 3-Li), J. Geophys.
Res. 90,9285-99.
Rice, J . R. & Tse, S. T. (1986). Dynamic motion of a single degree of free-
dom system following a rate and state dependent friction law, J. Geo-
phys. Res. 91,521-30.
Richardson, L. F. (1961). The problem of contiguity: An appendix of statis-
tics of deadly quarrels, General Systems Yearbook, 6, 139-87.
Rieu, M. & Sposito, G. (1991a). Fractal fragmentation, soil porosity, and
soil water properties, I. Theory, Soil Sci. Soc. Am. J. 55, 1231-8.
Rieu, M. & Sposito, G. (1991b). Fractal fragmentation, soil porosity, and
soil water properties, 11. Applications, Soil Sci. Soc. Am. J. 55, 1239-44.
Rikitake, T. (1958). Oscillations of a system of disc dynamos, Proc. Cam-
bridge Phil. Soc. 54,89-105.
Rinaldo, A., Rodriguez-Iturbe, I., Rigon, R., Ijjasz-Vasquez, E. & Bras, R. L.
(1993). Self-organized river networks, Phys. Rev. Lett. 70,822-5.
Robbins, K . A. (1977). A new approach to subcritical instability and turbu-
lent transitions in a simple dynamo, Math. Proc. Cambridge Phil. Soc.
82,309-25.
Robert, A. & Roy, A. G. (1990). On the fractal interpretation of the main-
stream length-drainage area relationship, Water Resour: Rex 26,839-42.
Robertson, M. C., Sammis, C. G., Sahimi, M. & Martin, A. J. (1995). Fractal
analysis of three-dimensional spatial distributions of earthquakes with a
percolation interpretation. J. Geophys. Res. 100,609-20.
REFERENCES 363
Robinson, P. A. (1994). Scaling properties of self-organized criticality, Phys.

Rev. E49,39 19-26.
Rosendahl, J., Vekie, M. & Kelley, J. (1993). Persistent self-organization of
sandpiles, Phys. Rev. E47, 1401-4.
Rosin, P. & Rammler, E. (1933). Laws governing the fineness of powdered
coal, J. Inst. Fuel 7,29-36.
Rothman, D. H., Grotzinger, J. & Flemings, P. (1994). Scaling in turbidite
deposition, J. Sed. Res. A64,59-67.
Rubio, M. A. & Galeano, J. (1994). Stick-slip dynamics in the relaxation of
stresses in a continuous elastic medium, Phys. Rev. E50, 1004-4.
Ruina, A. (1983). Slip instability and state variability laws, J. Geophys. Res.
88, 10359-70.
Rundle, J. B. & Brown, S. R. (1991). Origin of rate dependence in frictional
sliding. J. Stat. Phys. 65,403-12.
Rundle, J. B. & Jackson, D. D. (1977). Numerical simulation of earthquake
sequences, Seis. Soc. Am. Bull. 67, 1363-77.
Rundle, J. B. & Klein, W. (1993). Scaling and critical phenomena in a cel-
lular automaton slider-block model for earthquakes, J. Stat. Phys. 72,
405-12.
Rundle, J. B. & Klein, W. (1995). Dynamical segmentation and rupture pat-
terns in a "toy" slider-block model for earthquakes, Nonlinear Proc.
Geophys. 2,61-79.
Rundle, J . B., Klein, W., Gross, S. & Turcotte, D. L. (1995). Boltzmann fluc-
tuations in numerical simulations of nonequilibrium lattice threshold
systems, Phys. Rev. Letr. 75, 1658-6 1.
Sadler, P. M. & Strauss, D. J. (1990). Estimation of completeness of strati-
graphical sections using empirical data and theoretical models, J. Geol.
Soc. London 147,47 1-85.
Sadovskiy, M. A., Golubeva, T. V., Pisarenko, V. F. & Shnirman, M. G. (1985).
Characteristic dimensions of rock and hierarchical properties of seismic-
ity, Phys. Solid Earth 20, 87-95.
Sahimi, M., Robertson, M. C. & Sammis, C. G. (1992). Relation between the
earthquake statistics and fault patterns, and fractals and percolation,
Physica A191,57-68.
Sahimi, M., Robertson, M. C. & Sammis, C. G. (1993). Fractal distribution
of earthquake hypocenters and its relation to fault patterns and percola-
tion, Phys. Rev. Lett. 70,2186-9.
Salas, J. D. (1993). Analysis and modeling of hydrologic time series, in
Handbook ofHydrology, D. R. Maidment, ed., pp. 19.1-19.72, McGraw-
Hill, New York.
Sammis, C. G. & Biegel, R. L. (1989). Fractals, fault-gouge, and friction,
Pure Appl. Geophys. 131,255-27 1.
364 REFERENCES
Sammis, C., King, G. & Biegel, R. (1987). The kinematics of gouge defor-
mation, Pure Appl. Geophys. 125,777-8 12.
Sammis, C. G., Osborne, R. H., Anderson, J. L., Banerdt, M. & White, P.
(1986). Self-similar cataclasis in the formation of fault gouge, Pure
Sammis, C. G. & Steacy, S. J. (1995). Fractal fragmentation in crustal shear
zones, in Fractals in the Earth Sciences, C. Barton & P. R. La Pointe,
eds. pp. 179-204, Plenum Press, New York.
Saucier, A. (1992). Scaling of the effective permeability in multifractal
porous media, Physica A191,289-94.
Saucier, A. & Muller, J. (1993). Use of multifractal analysis in the character-
ization of geological formations, Fractals 1 , 6 17-28.
Scheidegger, A. E. (1967). A stochastic pattern for drainage patterns into an
intermontane trench, Int. Assoc. Sci. Hydrol. Bull. 12, 15-20.
Scheidegger, A. E. (1991). Theoretical Geomorphology, 3rd ed., Springer-
Verlag, Berlin, 434 pp.
Schenck, H. (1963). Simulation of the evolution of drainage-basin networks
with a digital computer, J. Geophys. Res. 68,5739-45.
Schiff, S . J. (1992). Resolving time-series structure with a controlled wave-
let transform, Optical Eng. 31(11), 2492-5.
Schmittbuhl, J., Vilotte, J. P. & Roux, S. (1993). Propagative macrodisloca-
tion modes in an earthquake fault model, Europhys. Lett. 21,375-80.
Schoeny, P. A. & Saunders, J. A. (1993). Natural gold dendrites from hy-
drothermal Au-Ag deposits: Characteristics and computer simulation,
Fractals 1,585-93.
Scholz, C. H. (1977). A physical interpretation of the Haicheng earthquake
prediction, Nature 267, 121-4.
Scholz, C. H. (1991). Earthquakes and faulting: Self-organized critical phe-
nomena with a characteristic dimension, in Spontaneous Formation of
Space-Time Structures and Criticality, T. Riste & D. Sherrington, eds.,
pp. 41-56, Kluwer Academic Publishers, Netherlands.
Scholz, C. H. & Cowie, P. A. (1990). Determination of total strain from
faulting using slip measurements, Nature 346,837-9.
Scholz, C. H., Dawers, N. H., Yu, J. Z. & Anders, M. H. (1993). Fault growth
and fault scaling laws: Preliminary results. J. Geophys. Res. 98, 21,
951-61.
Schoutens, J. E. (1979). Empirical analysis of nuclear and high-explosive
cratering and ejecta, in Nuclear Geoplosics Sourcebook, Vol. 55, part 2,
section 4, Rep. DNA OIH-4-2, Def. Nuclear Agency, Bethesda, MD.
Schreider, S. Yu (1990). Formal definition of premonitory seismic quies-
cence, Phys. Earth Planet. Int. 61, 113-27.
Schuster, H. G. (1995). Deterministic Chaos, 3rd ed., VCH, Weinheim, 291 pp.
REFERENCES 365
Schwarzacher, W. (1976). Stratigraphic implications of random sedimenta-

tion, in Random Processes in Geology, D. F. Merriam, ed., pp. 96-1 11,
Springer-Verlag, New York.
Segre, E. & Deangeli, C. (1995). Cellular automaton for realistic modelling
of landslides, Nonlinear Proc. Geophys. 2, 1-15.
Senatorski, P. (1994). Spatio-temporal evolution of faults: Deterministic
model. Physica D76,420-35.
Shaw, B. E. (1993a). Moment spectra in a simple model of an earthquake
fault, Geophys. Res. Lett. 20,643-6.
Shaw, B. E. (1993b). Generalized Omori law for aftershock and foreshocks
from a simple dynamics, Geophys. Res. Lett. 20,907-10.
Shaw, B. E. (1994). Complexity in a spatially uniform continuum fault
model, Geophys. Res. Lett. 21, 1983-6.
Shaw, B. E. (1995). Frictional weakening and slip complexity in earthquake
faults, J. Geophys. Res. 100, 18,239-5 1.
Shaw, B. E., Carlson, J. M. & Langer, J. S. (1992). Patterns of seismic activ-
ity preceding large earthquakes, J. Geophys. Res. 97,479-88.
Shreve, R. L. (1967). Infinite topologically random channel networks, J. Geol.
75,178-86.
Shreve, R. L. (1996). Statistical law of stream numbers, J. Geol. 74, 17-37.
Sieh, K. E. (1978). Slip along the San Andreas fault associated with the great
1857 earthquake, Seis. Soc. Am. Bull. 68, 1421-48.
Sieh, K. E. & Jahns, R. H. (1984). Holocene activity of the San Andreas fault
at Wallace Creek, California, Geol. Soc. Am. Bull. 95,883-96.
Sieh, K., Stuiver, M. & Brillinger, D. (1989). A more precise chronology of
earthquakes produced by the San Andreas fault in southern California,
J. Geophys. Res. 94,603-23.
Singh, S. K., Rodriguez, M. & Esteva, L. (1983). Statistics of small earth-
quakes and frequency of occurrence of large earthquakes along the Mex-
ican subduction zone, Seis. Soc. Am. Bull. 73, 1779-96.
Smalley, R. F., Chatelain, J. L., Turcotte, D. L. & Prevot, R. (1987). A fractal
approach to the clustering of earthquakes: Applications to the seismicity
of the New Hebrides, Seis. Soc. Am. Bull 77,1368-8 1.
Smalley, R. F., Turcotte, D. L. & Sola, S. A. (1985). A renormalization group
approach to the stick-slip behavior of faults, J. Geophys. Res. 90,1884-
1900.
Smart, J. S. (1972). Channel networks, Adv. Hydrosci. 8,305-46.
Smith, D. K. & Shaw, P. R. (1990). Seafloor topography: A record of a
chaotic dynamical system?, Geophys. Res. Lett. 17, 1541-44.
Smith, L. C., Turcotte, D. L. & Isacks, B. L. (1997). Streamflow characteri-
zation and feature detection using a discrete wavelet transform, submit-
ted for publication.
366 REFERENCES
Smith, S. W. (1976). Determination of maximum earthquake magnitude,

Geophys. Res. Lett. 3,351-4.
Snow, R. S. (1989). Fractal sinuosity of stream channels, Pure Appl. Geo-
phys. 131,99-109.
Snow, R. S. (1992). The Cantor dust model for discontinuity in geomorphic
process rates, Geomorph. 5, 185-94.
Sornette, A., Davey, P. & Sornette D. (1993). Fault growth in brittle-ductile
experiments and the mechanics of continental collisions, J. Geophys.
Res. 98, 12,111-39.
Sornette, A., Dubois, J., Cheminee, J. L. & Sornette D. (1991). Are sequences
of volcanic eruptions deterministically chaotic?, J. Geophys. Res. 96, 1 1,
93145.
Sornette, A. & Sornette, D. (1989). Self-organized criticality and earth-
quakes, Europhys. Lett. 9, 197-202.
Sornette, A. & Sornette, D. (1990). Earthquake rupture as a critical point:
Consequences for telluric precursors, Tectonophys. 179,327-34.
Sornette, D. & Davy, P. (1991). Fault growth model and the universal fault
length distribution, Geophys. Res. Lett. 18, 1079-8 1.
Sornette, D., Davy, P. & Sornette, A. (1990), Structuration of the lithosphere
in plate tectonics as a self-organized critical phenomenon, J. Geophys.
Res. 95, 17,353-6 1.
Sornette, D., Miltenberger, P. & Vanneste, C. (1994). Statistical physics of
fault patterns self-organized by repeated earthquakes, Pure Appl. Geo-
phys. 142,49 1-527.
Sornette, D. & Sammis C. G. (1995). Complex critical exponents from
renormalization group theory of earthquakes: Implications for earth-
quake predictions, J. Phys. 1 France 5,607-19.
Sornette, D. & Zhang, Y. C. (1993). Non-linear model of geomorphic ero-
sion processes, Geophys. J. Int. 113,382-6.
Sparrow, C. (1982). The Lorenz Equations: Bifurcations, Chaos, und
Strange Attractors, Springer-Verlag, New York, 269 pp.
Stark, C. (1991). An invasion percolation model of drainage network evolu-
tion, Nature 352,423-5.
Stark, C . P. (1994). Cluster growth modeling of plateau erosion, J. Geophys.
Res. 99, 13,957-69.
Stauffer, D. & Aharony, A. (1992). Introduction to Percolation Theory, 2nd
ed., Taylor and Francis, London, 181 pp.
Steacy, S. J. & Sammis, C. G . (1991). An automaton for fractal patterns of
fragmentation, Nature 353,250-2.
Stewart, C. A. & Turcotte, D. L. (1989). The route to chaos in thermal con-
vection at infinite Prandtl number, 1. Some trajectories and bifurcations,
J. Geophys. Res. 94, 13,707-17.
REFERENCES 367
Stinchcombe, R. B. & Watson, B. P. (1976). Renormalization group ap-

proach for percolation conductivity, J. Phys. C9,3221-47.
Strahler, A. N. (1957). Quantitative analysis of watershed geomorphology,
Trans. Am. Geophys. Un. 38,9 1 3-20.
Strocka, B., Duarte, J. A. M. S. & Schreckenberg, M. (1995). The self-orga-
nized critical forest-fire model with trees and bushes, J. Phys. 1 France,
5,1233-8.
Sun, T., Meakin, P. & Jossang, T. (1993). A minimum energy dissipation
model for river networks and their associated topographies, Fractals 1,
576-84.
Sun, T., Meakin, P. & Jossang, T. (1994). The topography of optimal
drainage basins, Water Resour: Res. 30,2599-6 10.
Sun, T., Meakin, P. & Jossang, T. (1995). Minimum energy dissipation river
networks with fractal boundaries, Phys. Rev. E51,5353-9.
Sykes, L. R. & JaumC, S. C. (1990). Seismic activity on neighbouring faults
as a long-term precursor to large earthquakes in the San Francisco Bay
area, Nature 348,595-9.
Takayasu, H. (1993). Power-law distribution of river basin sizes, Fractals 1,
521-8.
Takayasu, H. & Inaoka, H. (1992). New type of self-organized criticality in a
model for erosion, Phys. Rev. Lett. 68,966-9.
Takayasu, H. & Matsuzaki, M. (1988). Dynamical phase transition in thresh-
old elements, Phys. Lett. A131,244-7.
Tarboton, D. G . , Bras, R. L. & Rodriguez-Iturbe, I. (1988). The fractal na-
ture of river networks, Watel: Res. Research 24, 13 17-22.
Taylor, S. R. (1964). Abundance of chemical elements in the continental
crust: A new table, Geochem. Cosmochim. Acta. 28, 1273-85.
Taylor, S. R. & McLennan, S. M. (1981). The composition and evolution of
the continental crust: Rare earth element evidence from sedimentary
rocks, Phil. Trans. Roy. Soc. London 301A, 381-99.
Termonia, Y. & Meakin, P. (1986). Formation of fractal cracks in a kinetic
fracture model, Nature 320,429-3 1.
Thompson, A. H., Katz, A. J. & Krohn, C. E. (1987). The microgeometry
and transport properties of sedimentary rock, Adv. Phys. 36,625-94.
Todoeschuck, J. P., Jensen, 0. G. & Labonte, S. (1990). Gaussian scaling
noise model of seismic reflection sequences: Evidence from well logs,
Geophys. 55,480-4.
Tokunaga, E. (1978). Consideration on the composition of drainage net-
works and their evolution, Geographical Rep. Tokyo Metro. Univ. 13,
1-27.
Tokunaga, E. (1984). Ordering of divide segments and law of divide seg-
ment numbers, Jap. Geomorph. Un. 5 , 7 1-7.
368 REFERENCES
Turcotte, D. L. (1986a). Fractals and fragmentation, J. Geophys. Res. 91,

1921-6.
Turcotte, D. L. (1986b). A fractal model for crustal deformation, Tectono-
phys. 132,261-9.
Turcotte, D. L. (1986~).A fractal approach to the relationship between ore
grade and tonnage, Econ. Geol. 81,1528-32.
Turcotte, D. L. (1987). A fractal interpretation of topography and geoid spec-
tra on the earth, moon, Venus, and Mars, J. Geophys. Res. 92, E597-E6Ol.
Turcotte, D. L. (1989a). Fractals in geology and geophysics, Pure Appl.
Geophys. 131,171-96.
Turcotte, D. L. (1989b). A fractal approach to probabilistic seismic hazard
assessment, Tectonophys. 167,17 1-7.
Turcotte, D. L. (1993). Fractal tectonics and erosion, Fractals 1,491-512.
Turcotte, D. L. (1994a). Crustal deformation and fractals, a review, in Frac-
tals and Dynamic Systems in Geoscience, J. H . Kruhl, ed., pp. 7-23,
Turcotte, D. L. (1994b). Modeling geomorphic processes, Physica D77,
229-37.
Turcotte, D. L. (1994~).Fractal theory and the estimation of extreme floods,
J. Res. Natl. Inst. Stand. Technol. 99,377-89.
Turcotte, D. L. (1995). Scaling in geology: Landforms and earthquakes,
Proc. Natl. Acad. Sci. USA, 92,6697-704.
Turcotte, D. L. & Greene, L. (1993). A scale-invariant approach to flood-
frequency analysis, Stochastic Hydrol. Hydraul. 7,33-40.
Turcotte, D. L. & Schubert, G. (1982). Geodynamics, John Wiley & Sons,
New York, 450 pp.
Tyler, S. W. & Wheatcraft, S. W. (1992). Fractal scaling of soil particle-size
distributions: Analysis and limitations, Soil Sci. Soc. Am. J. 56,362-9.
Utsu, T. (1961). A statistical study on the occurrence of aftershocks, Geo-
phys. Mag. 30,52 1-605.
Van Damme, H., Obrecht, F., Levitz, P., Gatineau, L. & Laroche, C. (1986).
Fractal viscous fingering in clay slurries, Nature 320,73 1-3.
Vannimenus, J. & Viennot, X. G. (1989). Combinatorial tool for the analysis
of ramified patterns, J. Stat. Phys. 54, 1529-38.
Varnes, D. J. (1989). Predicting earthquakes by analyzing accelerating pre-
cursory seismic activity, Pure Appl. Geophys. 130,66 1-86.
Vasconcelos, G. L., de Sousa Vieire, M. & Nagel, S. R. (1992). Phase transi-
tions in a spring-block model of earthquakes, Physica A191,69-74.
Velde, B., Dubois, J., Moore, D. & Touchard, G. (1991). Fractal patterns of
fractures in granites, Earth Planet. Sci. Lett. 104,25-35.
Velde, B., Dubois, J., Touchard, G. & Badri, A. (1990). Fractal analysis of
fractures in rocks: The Cantor's dust method, Tectonophys. 179,345-52.
REFERENCES 369
Velde, B., Moore, D., Badri, A. & Ledesert, B. (1993). Fractal and length
analysis of fractures during brittle to ductile changes, J. Geophys. Res.
98,11,935-40.
Verhulst, F. (1990). Nonlinear Differential Equations and Dynamical Sys-
tems, Springer-Verlag, Berlin, 277 pp.
Vicsek, T. (1992). Fractal Growth Phenomena, 2nd ed., World Scientific,
Singapore, 488 pp.
Villemin, T., Angelier, J. & Sunwoo, C. (1995). Fractal distribution of fault
length and offsets: Implications of brittle deformation evaluation - the
Lorraine coal basin, in Fractals in the Earth Sciences, C . Barton & P. R.
La Pointe, eds., pp. 205-26, Plenum Press, New York.
von Seggern, D., Alexander, S. S. & Baag, C. E. (1981). Seismicity parame-
ters preceding moderate to major earthquakes, J. Geophys. Res. 86,
9325-5 1 .
Voss, R. F. (1985a). Random fractals: Characterization and measurement, in
Scaling Phenomena in Disordered Systems, R. Pynn & A. Skejeltorp,
eds., pp. 1-1 1, Plenum Press, New York.
Voss, R. F. (1985b). Random fractal forgeries, in Fundamental Algorithms
for Computer Graphics, NATO AS1 Series, Vol. F17, R. A. Earnshaw,
ed., pp. 805-35, Springer-Verlag, Berlin.
Voss, R. F. (1988). Fractals in nature: From characterization to simulation, in
The Science of Fractal Images, H . 0 . Peitgen & D. Saupe, eds., pp.
2 1-70, Springer-Verlag, New York.
Wallace, R. E. (1977). Profiles and ages of young fault scarps, north-central
Nevada, Geol. Soc. Am. Bull. 88, 1267-81.
Walsh, J. J. & Watterson, J. (1988). Analysis of the relationship between dis-
placements and dimensions of faults, J. Struct. Geol. 10,329-47.
Walsh, J., Watterson, J. & Yielding, G. (1991). The importance of small-
scale faulting in regional extension, Nature 351,391-3.
Watanabe, K. & Takahashi, H. (1995). Fractal geometry characterization of
geothermal reservoir fracture networks, J. Geophys. Res. 100,521-8.
Weitz, D. A. & Oliveria, M. (1984). Fractal structures formed by kinetic ag-
gregation of aqueous gold colloids, Phys. Rev. Lett. 52, 1433-6.
Willgoose, G., Bras, R. L. & Rodriquez-Iturbe, I. (1991). A coupled channel
network growth and hillslope evolution model, I: Theory, Water Resour.
Res. 27, 1671-84.
Wilson, K. G. & Kogut, J. (1974). The renormalization group and the E ex-
pansion, Phys. Rev. C12,75-200.
Witten, T. & Sander, L. M. (1981). Diffusion-limited aggregation, a kinetic
critical phenomenon, Phys. Rev. Lett. 47, 1400-3.
Wu, B. & Lye, L. M. (1994). Identification of temporal scaling behavior of
flood: A study of fractals, Fractals 2,283-6.
370 REFERENCES
Wu, Q., Borkovee, M. & Sticher, H. (1993). On particle-size distributions in

soils, Soil Sci. Soc. Am. J. 57,883-90.
Wyss, M. & Haberman, R. E. (1988). Precursory seismic quiescence, Pure
Appl. Geophys. 126,3 19-32.
Xu, H. J. & Knopoff, L. (1994). Periodicity and chaos in a one-dimensional
dynamical model of earthquakes, Phys. Rev. E50,3577-8 1 .
Yokoi, Y., Carr, J. R. & Watters, R. J. (1995). Fractal character of landslides,
Env. Eng. Geosci. 1,75-8 1 .
Young, I. E. & Cruff, R. W. (1967). Magnitude and frequency of floods, Part
1 1 , U.S. Geological Survey, Water Supply Paper 1685, pp. 714-5.
Young, R. K. (1992). Wavelet Theory and Its Applications, Kluwer, Dordrecht,
223 pp.
Youngs, R. R. & Coppersmith, K. J. (1985). Implications of fault slip rates
and earthquake recurrence models to probabilistic seismic hazard esti-
mates, Seis. Soc. Am. Bull. 75,939-64.
Zipf, G. P. (1949). Human Behavior and the Principle of Least EfSort, Addi-
son-Wesley, Reading, Mass.
Appendix A
GLOSSARY
OFTERMS
ATTR ACTOR A point in phase space toward which a time history evolves
as transients die out.
BAS IN O F ATT R ACT1 ON Some dynamical systems have more than
one fixed point. The region in phase space in which solutions approach a
particular fixed point is known as the basin of attraction of that fixed
point. The boundaries of a basin of attraction are often fractal.
B 1 FUR C AT I 0 N A change in the dynamical behavior of a system when a
parameter is varied.
B R 0 W N I A N WALK The running sum of a sequence of random values
usually obtained from a normal distribution.
C A N TOR D U S T A fractal set generated by subdividing a line into parts.
C H A O S Solutions to deterministic equations are chaotic if adjacent solu-
tions diverge exponentially in phase space; this requires a positive Lya-
punov exponent.
C L U S T E R A group of particles with nearest-neighbor links to other parti-
cles in the cluster.
D E T E R M I N IS T I C A dynamical system whose equations and initial con-
ditions are fully specified and are not stochastic or random.
D I F F E R E N C E E Q U AT I 0 N An equation that relates a value of a func-
tion x,, to a previous value x,,. A difference equation generates a dis-
crete set of values of the function x.
D I F F U S I O N - L I M I T E D A G G R E G A T I O N ( D L A ) Diffusing (random-
walking) particles accrete to a seed particle to form a dendritic structure.
D I M E N S I O N The usual definition of dimension is the topological di-
mension. The dimension of a point is zero, of a line is one, of a square is
two, of a cube is three. In this book we have introduced the concept of
fractional (noninteger) dimensions, or fractals.
F E I G EN B A U M C 0 N S TA NT The ratios of successive differences be-
tween period-doubling bifurcation parameters approach this number
(F = 4.699 202).
372 A P P E N D I X A: GLOSSARY OF T E R M S
FIXED POINT A point in phase space toward which a dynamical system

approaches as transients die out.
F R A C T A L An object that is quantified by a fractional dimension. A mea-
sure of a scale-invariant object. In some cases a power-law relation be-
tween the number of objects and their linear size.
F R A C T A L D I M E N S I O N The fractional dimension of a fractal. The
power in the power-law fractal scaling.
HA U S D 0 R F F M E AS U R E The power law in the scaling of a self-affine
fractal.
H 0 P F B I FU R C AT I 0 N A bifurcation from a fixed point to a limit cycle.
H U R S T E X P 0 N E NT Scale-invariant correlation used in rescaled range
(WS)analysis.
L A C U N A R I T Y A quantitative measure of clustering.
L I M I T C Y C L E A periodic orbit in phase space toward which a dynamic
system approaches as transients die out.
L 0 G - PER I 0 D I C B E H AV I 0 R Power-law (fractal) behavior when the
power is complex.
L 0 R E N Z E Q U AT I 0 N S A set of three first-order differential equations
derived from the equations governing thermal convection. Historically
this was the first example of deterministic chaos.
LYA P U N 0 V E X P O N E N T Solutions to deterministic equations are cha-
otic if adjacent solutions diverge exponentially in phase space; the expo-
nent is known as the Lyapunov exponent. Solutions are chaotic if the
Lyapunov exponent is positive.
M A P A mathematical relation that translates one or more points into other
points.
M U LT I FR A CTA L S An infinite sequence of fractal dimensions obtained
from the moments of a statistical distribution.
N O D E A fixed point toward which solutions evolve.
P E R C O L AT1 ON C LU S T E R A grid of sites in two or three dimensions.
The probability that a site is permeable is specified, and there is a sudden
onset of flow through the grid at a critical value of the probability.
P E R I O D D O U B L I N G A sequence of periodic oscillations in which the
period doubles as a parameter is varied.
P H A S E S PACE A coordinate space defined by the state variables of a
dynamical system.
P I T C H F O R K B I F U R C A T I O N A bifurcation in which the period
doubles.
P 0 I N c A ~ f S iE C T I 0 N The sequence of points in phase space generated
by the penetration of the evolving trajectory through a specified planar
surface.
PR 0 B A B I L I T Y The likelihood that a particular event will occur. The
probability that the next flip of a coin will be heads is 0.5.
A P P E N D I X A: GLOSSARY OF T E R M S 373
RANDOM A choice that is determined by pure chance. For example, a

flip of a coin.
RENORMALIZATION The transformation of a set of equations from
one scale to another by a change of variables.
R E S C A L E D R A N G E ( R / S ) A N A L Y S I S An empirical technique for
quantifying correlations in time series.
S A D D L E POINT A fixed point that attracts only a singular set of trajec-
tories.
S C A L E IN VA R I A N C E The phenomenon whereby an object appears iden-
tical at a variety of scales.
S ELF - A FFI N E Under an affine transformation the different coordinates
are scaled by different factors. If an object is scaled using an affine
transformation, the object is described as being self-affine.
S ELF - S I M I L A R I T Y A property of a set of points if their geometrical
structure at one length scale is the same as at another length scale.
S TOC HA ST1 C A process determined by chance.
S T R A N G E ATTRACTOR A fixed point in a phase space in which the
orbits are chaotic.
T I M E S E R I E S Values of a quantity as a function of time.
WAVELET A scale-invariant filter used to quantify time series.
WHITE N 0IS E A sequence of random values, often obtained from a nor-
mal distribution.
Appendix B
UNITS AND
SYMBOLS
Table B 1. SI units
Quantity Unit Symbol Equivalent
Basic units
Length meter m
Time second s
Mass kilogram kg
Temperature kelvin K
Electric current ampere A
Derived units
Force newton N
Energy joule J
Power watt W
Pressure pascal Pa
Frequency hertz Hz
Charge coulomb C
Electric potential volt V
Magnetic field tesla T
Multiples of 10
lo-' milli
10-6 micro
lo-' nano
10-l2 pic0
I 0' kilo
1O6 mega
I o9 giga
10l2 tera
APPENDIX B: U N I T S A N D S Y M B O L S
Table B2. Symbols
Equation
Symbol Quantity introduced SI units
Power (3.40)
Parameter (10.1)
Radius of earth (7.63) m
Frequency of earthquakes (4.1) s-I
Area (3.67) m2
Lorenz variable (12.23)
b-value for earthquakes (4.1)
Jl/?
Benioff strain (15.38)
Constant (4.4)
Specific heat at constant pressure (12.4) J kg-' K
Coefficient of variation (3.33)
Constant (2.1)
Concentration (5.1)
Pair correlation distribution (6.13)
Moment of inertia (14.3) kg m2
Constant (4.4)
Euclidean dimension (6.12)
Fractal dimension (2.1)
Energy (4.2) J
Probability (3.5)
Probability distribution function (3.12)
Probability of fragmentation (3.70)
Fraction (6.22)
-- I
Frequency (7.41)
Cumulative distribution function (3.10)
Flood frequency factor (8.31)
Feigenbaum constant (10.11)
Force (11.1) N
Wavelet filter (8.32)
Acceleration due to gravity (12.3) ms- ?
Applied torque (14.2) Nm
Elevation (7.28) m
Layer thickness (12.5) m
Hausdorff measure (7.1)
Hurst exponent (7.56)
Electrical current (14.1) A
Transport coefficient (8.25) m2s-I
Boltzmann constant (3.16) JK-'
Constant (3.40)
Wave number (7.66) m-I
Spring constant (9.12) Nm-I
Thermal conductivity (12.4) Wm-I K-'
Partition coefficient (5.29)
Rikitake parameter (14.13)
Length (2.9) m
Lacunarity (6.20)
Self-inductance (14.3) VsA-1
(continued)
376 A P P E N D I X B: U N I T S A N D S Y M B O L S
Table B2. (cont.)
Equation
Mass
Earthquake magnitude
Mass
Earthquake moment
Moment
Mutual inductance
Number
Number
Nusselt number
Number per unit time
Pressure
Probability
Perimeter
Prandtl number
Power
Volumetric flow
Linear dimension
Autocorrelation function
Radius
Ratio of Rayleigh numbers
Rate
Range
Resistance
Rayleigh number
Bifurcation ratio
Length-order ratio
Information entropy
Entropy
Power spectral density
Standard deviation
Time
Temperature
Time interval
Branching ratio
Variable
Horizontal velocity coordinate
Velocity
Vertical velocity
Volume
Variance
Wavelet transform
Variable
Horizontal coordinate m
(continued )
APPENDIX B: U N I T S A N D SYMBOLS 377
Table B2. (cont.)
Equation
Y Variable (3.18)
Vertical coordinate (2.7) m
Y Fourier transform (7.38)
Nondimensional position (11.3)
a Constant (4.5)
Mass ratio (5.13)
Stiffness parameter (11.12)
Volume coefficient of thermal
expansion (12.8) K-'
P Power (7.41)
Constant (9.12)
Symmetry parameter (11.12)
B Constant (4.6)
Y Coefficient of skew (3.4)
Semivariance (7.8)
I- Slope (10.6)
S Displacement across fault (4.3) m
E Strain (4.27)
Small quantity (5.39)
Parameter (9.13)
0 Latitude (7.63)
Polar coordinate (9.28)
Temperature difference (12.7) K
K Thermal diffusivity (12.11) m2 s-I
A Wavelength (7.66) m
Lyapunovexponent (10.19)
CL Shear modulus (4.3) Pa
Viscosity (12.2) Pa s
Rikitake parameter (14.8)
v Power (3.48)
Hazard rate (15.30) t-I
P Density (3.82) kg m-'
Polar coordinate (9.28)
u Standard deviation (3.3)
Stress (15.44) Pa
T Time interval (4.23) s
Nondimensional time (11.3)
dJ Porosity (3.81)
Enrichment factor (5.1)
Longitude (7.63)
Friction parameter (1 1.3)
$ Stream function (12.6) m2s-I
a Angular velocity (14.1) s-I
ANSWERS
TO SELECTED
PROBLEMS
N3 = 8 , r 3 = 1/27,N4 = 16,r4 = 1181.

N3 = 27, r3 = 11125, N4 = 81, r, = 11625.
( b ) N l = 2, N2 = 4 , N , = 8 , r , = 115,r2 = 1125,r, = 11125.
( c ) D = 0.4307.
( b ) N , = 3 , N 2 = 9 , N 3 = 27, r , = 117, r2 = 1149, r, = 11343.
(c) D = 0.5646.
( b ) N l = 4 , N, = 16, N3 = 6 4 , r , = 117,r2 = 1149,r, = 11343.
( c ) D = 0.7124.
N3 = 512, N4 = 4096, r3 = 1/27, r4 = 1/81.
( b )N , = 2, N, = 4 , N, = 8, r , = 112,r2 = 114, r3 = 118.
( c )D = 1.
( b )N , = 5 , N, = 25, N3 = 125, r , = 113, r2 = 119, r, = 1/27.
(e) D = 1.46.
( b ) N , = 5 , N2 = 25,N3 = 125,rl = 113,r2 = 119,r3 = 1/27.
( c ) D = 1.46.
( b )N , = 12, N, = 144, N, = 1728, r , = 114, r2 = 1116, r3 = 1164.
( c ) D = 1.79.
( b )N , = 24, N2 = 576, N3 = 13 824, r , = 115, r2 = 1/25, r, = 11125.
( c ) D = 1.9746.
( b ) N , = 17,N2 = 289, N, = 4 9 1 3 , r , = 115,r2 = 1125,r3 = 11125.
( c ) D = 1.76.
D = 2.975.
( b )A , = 112, A, = 1212, A, = 12012. ( c ) No.
( b )No = 4 , N l = 32, N2 = 256, ro = 1, rl = 114, r2 = 1/16,
Po = 4 , P I = 8, P, = 16.
( c ) D = 1.5.
( a ) N , = 1,N2 = 8, N, = 6 4 , r l = 113,r2 = 119, r, = 1127.
( b )Yes. ( c ) D = 1.89.
Yes, 1.
0.577 Myr, 1.208 Myr.
ANSWERS TO SELECTED PROBLEMS 379
(a) 3 km, 6 km, 5 Myr, 15 Myr, 10 Myr, 20 Myr.

(b) 1 km, 2 km, 4 km, 5 km, 7 km, 8 km, 1 Myr, 3 Myr, 11 Myr,
13 Myr, 2 1 Myr, 23 Myr, 2 Myr, 4 Myr, 12 Myr, 14 Myr, 22 Myr,
24 Myr.
(d) R (25 Myr) = 0.36 m d y r , R (5 Myr) = 0.6 m d y r , R (1 Myr) =
0.112, 0.108,0.096, 0.079.

0.240,0.760.
1.65, 2.16, 1.31.
0.894.
0.707.
vz.
0.523.
+
7 = vrO/(v I), V = v(v + 3)(v - I)ril(v + 2)(v + 1)2.
7, 140,2800; 2.727.
Nl=4,N,= 16,N3=64,D=2.
6, 126,2646; 2.77 12.
4, = 0.0385, p, = 0.9627p0, 4, = 0.0727, p, = 0.9273p0,
D = 2.9656.
4, = 0.2857, p, = 0.7778p0, 4, = 0.395 1, p, = 0.6049p0,
D = 2.7712.
E, = 2.1 X 1015 J, M = 4 X lO19J,A = 530km2, ae = 2.5 m.
E, = 7.1 X 1013J, M = 1.3 X 1018J,A = 54km2, ae = 0.8m.
100.
11.9 yr.
103 yr.
(a) a = 2 X 106yr-1. (b) re = 158 yr.
(a) a = 7 X lo4 yr-1. (b) re = 22 yr.
400 yr.
1600 yr.
1, 8,64; 1.446.
1.292.
1,o.
1000 yr.
113 yr.
380 ANSWERS TO SELECTED PROBLEMS
1.57.
1.26.
113.
213.
(a) (8 - 4,)Cd7. (b) (8 - $8)+8Cd7' ( c ) 1 < < 8.
(d)O < D < 3.
107 kg.
3.18 X 108 kg.
88 kg.
4.85 X 106 kg.
8.37 X 1010 kg.
3.9 X 107 kg.
2.55.
0.5.
1, 113, 119, 1/27.
1, 1, 1, 1.
1,315,9125.
1,417, 16149,641343.
1,317,9149,271343.
1, 113, 119, 1127.
1, 112, 114, 118.
1, 17125,2891625.
1, 618, 36164, 2161512.
1,26127, 6761729.
c (112) = 4, C (1) = 2.
C (1) = 6.
C (1) = 24, C ( f i ) = 2 4 1 f i , C ( f i ) = 8 1 f i .
382 ANSWERS TO SELECTED PROBLEMS
1.513; 0.75, 2.0194, 1.40.

y = clear,x = c2efi.
X, = clept(diverges), x, = ~ ~ e - (converges),
2~' x, = c,e-2~'
(converges).
x = 1 (stable), x = - l (unstable).
(a) - 1, 0, 1. (b) stable, unstable, stable.
x = 2n7r (stable), x = 1 + 2nn (stable), n = 0, 1, 2, . . .
0, - p, p < 0: 0 stable, - p unstable.
p > 0: 0 unstable, - p, stable.
transcritical.
0 (unstable), + k-112 for p > 0 (stable) supercritical pitchfork.
+
0 (unstable), (-p)-112 for p, < 0 (stable) subcritical pitchfork.
x, = 0 . 1 2 5 , ~=~0 . 0 5 4 7 , ~=~0 . 0 2 5 8 , ~=~0 . 0 1 2 6 , ~=~0.
x, = 0.16875, x, = 0.12625, x, = 0.09928, x, = 0.08048, xf = 0.
x, = 0 . 3 2 , ~=0.4352,x3
~ =0.4916,x4= 0.4999,xf=0.5.
x, = 0 . 5 2 5 , ~=~0 . 6 2 3 4 , ~=~0 . 5 8 6 9 , ~=~0 . 6 0 6 1 , ~=~0.6.
x,, = 0.513045, x, = 0.799455.
x,, = 0 . 4 5 1 9 7 , ~=~0.84216.
xfmax = 0.925, xfmin = 0.2567.
xfmax= 0.95, xfmin= 0.1805.
x, = 0.2298, x, = 0.7081, x, = 0.8268, x, = 0.5728, x4 = 0.9788.
x, = 0.1070, x, = 0 . 3 8 2 4 , ~=~0.9447, x, = 0 . 2 0 9 1 , ~=~0.6615.
r = a.
(a) x, = 0 unstable, x, = 1 stable for 0 < a < 2, unstable for a > 2.

(b) 2.2408, 0.0542,0.9248, 1.1589,0.7196.
(c) 2.46302,0.03057.
(a) Y = 1. (b) Y = 1.333.
(a) Y = 1. (b) Y = -0.3333.
(a) Y, = 1, Y, = 0.5. (b) Y, = 0, Y, = 0.5. (c) Y, = 0.5, Y2 = 1.

(d) Y, = 0.5, Y2 = 0.
(a) Y, = 1, Y2 = 0.75.(b) Y, = 0.5, Y2 = 0.75.(c) Y, = 0.75, Y, = 1.
(d) Y, = 0.75, Y, = 0.5.
-6.16, -1.714.
14.21, 7.20.
x = y = t ( R - v ) ~ / * ,=z 1.
0.864,0.6023, 1.0555, -0.3309, -0.8103; - 1.1547 < x < 1.1547.
12, 144.
= nD/3.
3i.
INDEX
accretionary headward growth, 195 autoregressive model, 141

acoustic well log, 165 autocorrelation function for, 141
activation, seismic, 329,330 comelogram for, 142
advection-diffusionmodel, 207 variance of, 141
aftershock statistics, 304 autoregressive moving average model, 143
aftershocks, 76.77, 107,246,304,323,325,328 autocorrelation function for, 145
Omori's law for, 304 variance of, 145
spacial distribution of, 77, 107 avalanche model, 5,207,3 l6,3 17
age distribution of sediments, 18 avalanches, 3 16,317,321
age of faults, 203
age of shorelines, 203 backbone, percolation, 78,291,336
aggregation, diffusion-limited,4,74, 195, 197, Bak, Per, 5,316,341
199,201,207,371 basement, 76
alluvial fans, 13, 203 basin, drainage, 181, 193.207
analysis, screen, 42 sedimentary, 76
sieve, 42 basin of attraction, 37 1
angle of repose, 3 16 bathymetry, 2, 163, 165, 167, 168,217
angular velocity, 28 1 correlation dimension of, 167
annual floods, 208,209 Fourier spectral analysis of, 2. 163
annual peak discharge, 208 wavelet analysis of, 217
antipersistent, 136, 137, 138, 152, 161, 162 bedding planes, 19,321
area, drainage, 193 bell curve, 31
rupture, 39.58 Benioff strain, 305
Arizonia topography, 177 bifurcation, 185, 187,225,226,227,229,
Armenian earthquake, 328 234,236,237,240,243,251,
arrays, slider-block, 325, 330, 336
arteries, 193
ash, volcanic, 44.79 flip, 234,236,238,240
fractal dimension of, 44 Hopf, 229,263,265,272,273,372
asperities, 299 pitchfork, 226,227,228,251,261,263,265,
asteroids, 44 272,372
atmosphere. 23 1 subcritical, 228, 272
atmospheric pressure, 160 supercritical,226
atmospheric temperature, 136, 146, 160 transcritical, 226
attraction, basin of, 371 turning point, 226
attractor, 265,272,284,371,373 bifurcation diagram, 226,227,237,243,251,
Lorenz, 265,272,284 284
strange, 265,284,373 bifurcation ratio, 185, 187
autocomelation function, 137, 138, 141, 145 Big Bear earthquake. 77
automata, cellular, 5,316,317,320,322,324, bin analysis, 34, 37
325,326 binomial coefficient, 102, 126
automaton. 50 binomial distribution, 83
autoregressive integrated moving average model, block models, slider, 4, 245,247, 254, 325, 330,
145 336,342
386 INDEX
blocks, pair of slider, 247, 342 chaos, 3,219,223,225,231,236,237,238,240,

multiple slider, 325, 330, 334, 336, 342 245,25 1,253,256,265,266,269,
slider, 245,247,325,330,334,342 272,289,341,37 1,372
body-wave magnitude, 57.58 deterministic, 219. 223, 23 1 , 245, 256, 266
Boltzmann constant, 31, 118 route to, 237
boundary layer, thermal, 265,269 windows of, 236,237,238,240,25 1
Boussinesq approximation. 257,269 chaotic mantle convection, 269
box-counting method, 14, 15,50,68,69,70,72, characteristic earthquake, 67, 322. 325
76.78.98, 107, 120, 124, 132, 135, chromatographic model, 87.95
147 circulations, hydrothermal, 81
moving, 111 climate, 166, 341
one-dimensional, 69, 70 clustering, 100, 103, 104,328, 371, 372
three-dimensional, 76 fractal, 100, 103, 104, 328
box-counting method for earthquakes, 76 lacunarity of, 109,372
box-counting method for oil pools, 98 clustering of faults, 104
box-counting method for pore spaces, 50 clustering of joints, 104
box-counting method for self-affine fractals, 132, clustering of metamorphic veins, 104
135, 147 clustering of seismicity, 103, 328
box-counting method for time series, 147 clusters, 4,78, 100, 103, 104,290, 291, 372
box dimension, 115, 124 percolation, 4.78.290, 291,372
braided rivers, 207 backbone of, 78,291,336
branch numbers, 188 number-size statistics of, 291
branches, side, 188 coal, 43.44
branching ratios, 189 coal mines, 67
bronchial system, 193 coastline, 1 , 12, 15, 132
Brownian topography, 163,207,208 fractal dimension of, 1, 12, 15, 132
Brownian walk, 2, 138, 140, 146, 147, 149, 150, length of, 1, 12
152, 155, 157, 158, 160, 163, 165, rocky, 1, 12, 15, 132
167,207,208,371 roughness of, 15
correlation dimension of, 167 coefficient of skew, 29,30,35
fractional, 149, 150, 152, 155, 156, 158, 161. coefficient of skew for a log-normal distribution,
162, 166 35
as a self-affine fractal, 140, 146, 150 coefficient of thermal expansion, 258
correlation dimension of, 166 coefficient of variation, 35, 137, 158
method of successive random additions, coefficient of variation for a log-normal distribu-
152 tion, 35, 158
rescaled range analysis of, 161, 162 comminution, 44,48,49,50,67,71,74, 298
log normal, 157, 158 comminution model, 48,49,50,71,74,298
semivariagrams of, 155 complementary error function, 33
standard deviation of, 140 complexity, 342
time series, 140 component, periodic, 136
variance of, 140 stochastic, 136, 137
buoyancy force, 257,265 trend, 136
b-value, 2, 57.59.324 concentration, mineral, 8 1.90, 120, 136
ore, 8 1.90, 120, 136
cable, stranded, 299,303, 307 conditional probability, 301
California earthquakes, 61.63.330.335 conduction, electrical, 295
southern, 61.63.330, 335 heat, 257
Cantor dust, 8.37 1 conduction solution. 261
Cantor set, 8, 18.21.24, 100, 102, 105, 107, 109, conductivity, electrical, 165
111, 113, 116, 118,321 thermal, 257
pair correlation of, 107 conservation of energy, 223,289
random, 100, 102, 109, 1 11 conservation of mass, 202
cardiovascular, 193 contact areas, multifractal analysis of, 129
carpet, Sierpinski, 9.24, 105, 1 11, 118, 119, 29 1 continental drift. 279
cascade, multiplicative, 120, 12 1, 124, 126, 128 continental margin, 321
caves, 50 continuity equation, 256
cell, 46.81.83.292.299 continuous data, 29.30
cellular-automata model, 5, 3 16.3 17, 320, 322, continuous processes, 56
324,325,326 continuous time series, 136, 137
center, 223 contour, topographic, 12, 132
central-limit theorem, 2, 34, 39 fractal dimensions of, 12
INDEX 387
convection. 3, 256, 257, 261, 268, 269 delta, prograding, 203

chaotic mantle, 269 river, 193
heat, 257 dendritic gold, 198
mantle, 256,269 dendritic growth, 197
thermal, 3, 256,261,268, 269 dendritic structure, 4, 37 1
copper, 93.95 density, 136
core, 279,280 power spectral, 148, 149, 150, 166, 168. 170,
core dynamo, 279, 280 203
cores, drilling, 68, 136 soil. 52
correlated noises, 140 depleted element, 85, 120
correlation, pair, 106, 107 deposition of sedimens, 18, 22, 202
of Cantor set, 107 rate of, 22
of Koch snowflake, 107 depositional sequence, 18
of seismicity, 107 deposits. epithermal, 87
correlation dimensions, 116, 167, 286 hydrothermal, 87
of bathymetry, I67 mineral, I, 5, 8 1,87,90, I46
of Brownian walks. 167 ore, 1, 5,81. 87,90, 146
of dynamo models. 286 skarn, 87
of geysers, 167 turbidite, 321
of time series, 167 deterministic, 37 1
of volcanic eruptions. 167 deterministic chaos, 2 19,223. 23 1 , 245, 256.
correlations, 138. 158 266,372
long range, 138 deterministic fractal, 6
time series, 158 determinisstic self-affine fractal, 133
correlogram, 137, 142 deviation, standard, 29, 30
autoregressive model, 142 devil's staircase, 18, 2 1, 23, 24
counter scaling, 162 diamonds, 95
craters, 44 difference equation, 37 1
impact, 44 diffusion equation, 202, 203, 207
lunar, 44 diffusion-limited aggregation, 4, 74, 195, 197,
creep, thermally activated, 269 l99,2OI,207,37l
solid state, 258 diffusivity, thermal. 258
critical permeability, 4, 290, 295 dimension, 115, 167, 371
critical phenomena, 63, 316, 328 box, 115, 124
critical point, 4, 289, 294, 3 16, 333 correlation, 116. 167, 286
critical porosity, 295 embedding, I67
critical probability, 4, 290, 295. 372 entropy, 1 16, 1 18
critical Rayleigh number, 260,264 Euclidian, 6
critical state, 5, 316 fractal, I, 2 , 6 , 14, 15, 16,22,59,90,92, 1 15,
criticality, self-organized, 5, 3 16, 3 17, 325, 326, 132, 135,372
328,329,336,341 of drainage networks, 185
crushed materials, 44 of earthquakes. 59
crustal deformation, 56 of time series, 147
crustal seismicity, 322 fractional, 6, 16, 372
crustal thinning. 76 information, 116, 1 18
Culling model, 202, 207 rnultifractal, 115, 117, 124
cumulative distribution function, 30, 3 1, 32, 36, self-affine fractal, 135
37.40. 146 discharge, annual, 208
exponential, 40. 303 flood, 136,208
log-normal. 36 river, 37, 136, 158, 160,215
normal, 32 discontinuous process, 56
Pareto, 37 discontinuous time series, 137
curdling, 8, 103 discrete data, 29
Curie temperature. 279 discrete Fourier transform. 150, 17 1
cycle, limit, 223,229,234,236,238.25 I, 253,372 inverse, 150
discrete time series, 136
damping, linear, 222 disk dynamo, 280
nonlinear, 222 displacement, earthquake, 75
dams, induced seismicity by, 328 fault, 56, 74
Darcy's law, 290 disruptions, explosive. 44
day, length of, 136, 217 dissipation, minimum energy, 207
deformation, crustal, 56 dissolution, 87
388 INDEX
distillation, Rayleigh, 87, 88, 89.95 drilling cores, 68, 136

distribution, binomial, 83 droughts, 158, 208
exponential, 39.40.303 dust, Cantor, 8, 37 1
fractal, 3, 15,21,28,38,39,41,84,87,90, dynamic friction, 246,326
96, 100,208,211,317,341 dynamical systems, 219, 371
frequency-magnitude, 15,57,66 dynamics, population, 4, 219,232,244
gamma, 208.2 1 1 dynamo, 3,279,280,282,285
Gaussian, 2.31.33, 137, 138, 140, 145,214, core, 279,280
332 disk, 280
Gumbel, 208,211 Rikitaki, 3,279,280,282,342
Hazen, 208.2 1 1 self-excited, 279
log-Gumbel, 208,211 two-disk, 280,282
log-normal, 2,35,38,42,81,83,90,96, 137, dynamo equations, 3,282,285
145, 158,208,211
log-Pearson, 208,211 earthquake, 67,246,299,317,322
Maxwell-Boltzmann, 332,333,336 aftershocks of, 76,77, 107,246,304,323,
multifractal, 128, 129 325,328
normal,2,31,33,38, 137, 138, 145,341, Omori's law for, 304
371,373 spacial distribution of, 77. 107
number-size, 1, 15, 79, 291 Armenian, 328
Pareto, 37.38.42.341 Big Bear, 77
Poisson, 102, 103 characteristic, 67,322, 325
power-law, 2, 16,38,41,42,50,81,84,90, Haicheng, 330
96,208,211,341 Joshua Tree, 76
probability, 3 1 Kern County, 56,330
Rosin-Rammler, 40,42 Landers, 56,57,63.329,330
scale-invariant, 3 Loma Prieta, 56,57,305,329
Weibull, 4 1,42,299,301,303 Northridge, 56,57,63,330
distribution function, 118 San Fernando, 56,330
cumulative, 30,31,32,36,37,40, 146 San Francisco, 56,57,330
distribution of aftershocks, 77, 107,304,323 Tangshan. 330
distribution of earthquakes, 1.76, 107, 128,320, Whittier, 63
324,325,326,327 earthquake displacement, 75
distribution of faults, 1, 67, 68, 71, 199, 322 earthquake energy, 57,58,61
distribution of floods, 208,209 earthquake magnitude, 57.58.61
distribution of forest fires, 337,339 bodywave, 57,58
distribution of fractures, 1.67.68, 129, 199 local, 57.58
distribution of fragments, 1, 15.42 moment, 57,58,59,60,61,305
distribution of gaps, 109 surface wave, 57.58
distribution of incomes, 38 earthquake moment, 57,58,59,60,61,305
distribution of islands, 15 earthquake precursors, 305,328
distribution of joints, 68, 199 earthquake prediction, 254,307,328,329,342
distribution of lakes, 15,96 earthquake rupture area, 39.58
distribution of landslides, 316,321 earthquakes, 1.4, 15,39,56,57,58,59,61,65,
distribution of lifetimes, 303 74, 103, 107, 128, 145,317,320,324,
distribution of mineral deposits, 1, 81 325,326,327,328,330
distribution of oil fields, 1.95 clustering of, 103,328
distribution of ore, 1.81 distribution of, 1.76, 107, 128,320,324,
distribution of pore spaces. 129 325,326,327
distribution of porosity, 3, 166, 167 foreshocksof, 246,323,324
distribution of sand slides, 321 fractal dimension of, 59.76
distribution of sediment ages, 18 frequency-magnitudestatistics of, 1, 15, 39,
distribution of slider block slip events, 325, 326 57,59,60,66,317,324,327
distribution of topography, 129 intervals between, 74
distribution of volcanic eruptions, 1.79 Memphis-St. Louis, 66
divider method, 12, 13, 14 multifractal distribution of, 128
doubling, period, 234,236,238,251,371,372 New Madrid, 66
drainage area, 193 pair correlations for, 107
drainage basins, 181, 193, 207 Parkfield, 67.254
drainage networks, 2.5, 181, 185, 191, 194, 199, rupture dimensions of, 39,58
207 southern California, 61.63.330, 336
fractal dimension of, 185 spatial distribution of, 76, 128
drift, continental, 279 earth's mantle, viscosity of. 258
INDEX
economic ore deposits, 5, 8 1,90,120, 146 faults, 1, 28.56.67.70. 104,203,245,247,299,

economics, 38 303
Eden growth, 207 age of, 203
Edward-Wilkinson equation, 207 clustering of, 104
ejecta, volcanic, 44 distribution of, 1.67,68,71, 199.322
elastic lithosphere, 13 fractal distribution of, 1.67.71.322
elastic rebound, 245,246 interacting, 247
electric fields, 4 length of, 67
induced, 28 1 normal, 76
electrical conduction, 295 number-length statistics, 67, 322
electrical conductivity, 165 pre-existing. 245
element, 46,81,85,290,292,299,303,307 transform, 56.70
depleted, 85, 120 Feigenbaum constant, 236,371
enriched, 85, 120 Feigenbaum relation, 236
impermiable, 290 fields, electric, 4
permiable, 290 gravitational,4
elevation, 163 induced electric, 28 1
elevation changes, 23 magnetic, 4,217,279,342
embedding dimension, 167 filling, space, 236
energy, 223,289 filtering, 214
conservation of, 223,289 Fourier, 152
earthquake, 57.58.61 fingering, viscous, 198
energy dissipation, minimum, 207 first law of thermodynamics, 289
energy equation, 256 fixed point, 220,221,223,224,225,226,227,
enriched element, 85, 120 232,242,265,270,294,297,302,372
enrichment factor, 82,87, 120 stability of, 220, 221
entropy, 118,289 stable, 220,225,232,265,294,297,302
maximum, 50 unstable, 232,265,294,297,302
entropy dimension. 116.11 8 flexure, 13
entropy method, 207 flip bifurcation, 234,236,238,240
episodic sedimentation, 23 flood discharge, 136,208
epithermal mineral deposits, 87 flood frequency factor, 209
epochs, geological, 18 flood hazard, 208
ergodic, 137 floods, 158,208,209
erosion, 2, 13, 18, 19.23, 194,202,208 annual, 208,209
characteristic time for, 202 frequency of, 208
error function, 33,203 partial duration series of. 209
complementary, 33 fluid flow, 50
eruptions, volcanic, 1.28.78.79. 167, 243 fluid layer. 256,268,269
chaotic behavior of, 243 fluid turbulence, 3, 127,231,268,341
correlation dimension of, 167 folds, 1, 56
Euclidian dimension, 6 force balance equations, 256
exchange of stabilities, 226 forces, buoyancy, 257,265
expansion, thermal, 256,258 inertial, 257
coefficient of, 258 pressure, 257
explosion, nuclear, 43.44 viscous, 257
explosive processes. 44 foreshocks. 246.323.324
exponential, distribution, 39, 303 forest-fire models, 336, 337, 339
cumulative, 40,303 forest fires, 291.339
mean. 40 distribution of, 337,339
probability, 39 fossil magnetism, 279
variance, 40 Fourier coefficients, 2, 150
Fourier filtering technique, 152
failure, time to, 303, 304, 305, 306 Fourier series, 2. 148
fan, alluvial, 13,203 amplitudes of. 2, 148
fault, 4.56.74 phases of, 148
San Andreas, 56.62.65.67. 104 Fourier spectrum, 2, 148, 163
San Gabriel, 104 of bathmetry, 2, 163
fault displacement, 56.74 of topography, 2, 163
fault gouge, 44,50 Fourier transform, 148, 150, 171.214
fault rupture, 299 discrete, 150, 171
fault scarps, 203 inverse, 148, 150
age of, 203 two-dimensional, 17 1, 172
390 INDEX
fractal, 1.2.6, 12,89,304,341,372 correlation dimension of, 166

deterministic, 6 fractal dimension of, 149
self-affine, 133 rescaled range analysis of, 161. 162
homogeneous, 116, 118 self-affine fractal, 150
self-affine, 132, 135, 138, 140, 145, 146, 148, semivariograms of, 155
150, 163, 166,372 successive random additions, 152
deterministic, 133 wavelet analysis of, 217
statistical, 140 fractional dimension, 6. 16, 372
self-similar, 132 fractional Gaussian noise, 149, 150, 152, 155,
statistical, 12, 38, 289 158, 161, 162
fractal clustering, 100, 103, 104, 328 Fourier filtering techniques, 152
fractal dimension, 1.2.6, 12, 14, 15, 16, 22, 59, rescaled range analysis of, 161, 162
90.92, 115, 132, 135, 372 semivariograms of, 155
box, 115, 124 fractional log-normal noises, 157, 158
correlation, 116, 167,286 fractional log-normal walks, 157, 158
entropy, 116, 118 fracture networks, 129
information, 116, 1 18 multifractal analysis of, 129
methods of determining, 12, 14 fracture surfaces, 166
box counting, 14, 15,50,68,69,70, 72, fractures, 5,28,50,67,68,70,72, 120, 166, 199
76,78,98, 105. 120, 132, 135, 147 distribution of, 1.67.68, 129, I99
divider, 12, 13, 14 fractal dimension of, 68
ruler, 12, 13, 14 permability of, 50, 166
multifractal, 115, 117, 124 porosity of, 50
of coastlines, 1, 12, 15, 132 fragmentation, 5, 28,42,48, 71, 292, 295, 323
of drainage networks, 185 fractal, 42, 292, 323
of earthquakes, 59,76 models for, 46,48,49,50,71, 74,292, 323
of fractures, 68 multifractal analysis of, 129
of mineral deposits, 5, 81, 87, 90 renormalization group analysis of, 295
of oil fields, 95 tectonic, 44, 70
of ore deposits, 5, 8 1,87,90 fragmentation probability, 46, 292, 298
of sedimentary sequences, 22.32 1 fragments, rock, 1, 15. 28-42
of time series, 147, 148 mass of, 42
of topographic contours, 12 number of, 42
self-affine, 135 size distribution of, 1, 15.42
fractal distribution, 3, 15, 2 1, 28, 38, 39,4 1, 84, free surface, 260
87.90.95, 100,208, 211,317, 341 frequency-magnitude statistics, 15, 57, 66
of earthquakes, 59.76 for earthquakes, 1, 2, 15, 39,57, 59,60,66,
of faults, 1, 67, 68, 71, 322 317,324,335
of floods, 208 for floods, 208
of fragments, 28,323 Gutenberg-Richter, 2,57,59,60, 66, 3 17,
of landslides, 32 1 335
of mineral deposits, 5, S 1, 87.90 frequency-size distribution, 1, 15, 57, 67. 79,
of oil fields, 95 322
of ore concentrations, 5.81.87.90 for earthquakes, 1, 15,57, 3 17
of sedimentary layer thicknesses, 321 for faults, 1.67, 322
of volcanic ash, 44 for floods, 208
fractal fracture surfaces, 166 for forest fires. 337, 339
fractal fragmentation, 42,292, 323 for islands, 15
fractal island, 11. 24 for lakes, 15.96
fractal landscapes, 181 for landslides, 3 16, 32 1
fractal set, 6, 371 for mineral deposits, 1, 8 1
fractal statistics, 12, 38, 181, 289 for oil fields, 1,95,96
fractal topography, 132, 208 for rock fragments, 1, 15.42
fractal tree, 2, 181, 187, 188, 299, 302 for slider blocks, 325,326
binary, 188 for volcanic eruptions, I, 79
branch numbers for, 188 friction, 245, 246,299, 326, 330
branching ratios of, 189 dynamic, 246,326
deterministic, 187 static, 246, 326, 330
side branching of, 188 velocity weakening, 299,325
Tokunaga, 188, 190, 198
fractional Brownian walk, 149, 150, 152, 155, gamma distribution, 208.21 1
156, 158, 161, 162, 166 gamma function, 40
INDEX
gaps, distribution of, 109 heat convection, 257

sedimentary record, 18, 19,20, 167 heat equation, 4, 13,257
Gaussian distribution, 2, 31, 33. 137, 138, 140, height of topography, 132, 136
145,214,332 hiatuses. sedimentary, 18, 19.20, 167
Gaussian white noise, 140, 149, 150, 160, 207 homogeneous fractal, 1 1 6, 118
fractional, 149. 150, 152, 155, 158, 161, 162 Hopf bifurcation, 229, 263, 265, 272, 273, 372
Fourier filtering technique, 152 Horton's laws, 185
semivariograms of, 155 hot spots, 106,269
geoid, 170, 217 Hurst exponent, 160, 161. 162,212,372
power spectral density of, 170 hydrology, 167,290
wavelet analysis of, 217 hydrothermal circulations, 8 1
geological epochs, 18 hydrothermal mineral deposits, 87
geomagnetic field, 2 17 hypsometric curve, 145
geometric incompatibility, 70
geomorphology, 13, 181,208 impact craters, 44
geysers, correlation dimension of. 167 impacts. 28,43,44
glaciers, 2 16 impermeable element, 290
global load sharing, 304 incomes, distribution of, 38
global seismicity, 59 incompatibilities. geometrical, 70
gold, 92.95.98. 198 induced electric field, 281
dendritic, 198 induced seismicity. 5, 129. 328
lode, 92 inductance, mutual, 281
spatial distribution of, 98 self, 282
gouge, fault, 44,50 inertia, moment of, 282
grade, ore, 81, 84,87,90, 136, 146 inertial force, 256
granular, material, 3 16 information dimension, 116, 118
gravitational fields. 4 interacting faults, 247
gravitationally unstable, 256 interactive systems. 3 16
gravity, 3, 170 intergranular porosity, 50
ground water hydrology, 167,290 intervals between earthquakes, 74
group, renormalization, 5,289, 290,292,295, invariance, scale, 1, 11,39,84,91,289,304, 316,
299,316,317 372,373
growth, dendritic, 197 invasion percolation, 207
Eden, 207 inverse Fourier transform, 148
growth models, 195 discrete, 150
diffusion-limited aggregation, 4, 74, 195, island, fractal, 11, 24
197, 199,201,207 Koch, 11,24, 146
growth networks, 74. 195 volcanic, 194
accretionary, 195 islands, frequency-size distribution of, 15
growth phenomena, 3,207 isotropic, 132
Gumbel distribution, 208, 21 1 iterative map, 232
log, 208,2 11
Gutenberg-Richter relation, 2,57,59,60, 66, joints, 28.67.68, 70, 104,299
317,335 clustering of, 104
gyration, radius of, 196 distribution of, 68, 199
Joseph effect, 158.2 13
Haicheng earthquake. 330 Joshua Tree earthquake, 76
harmonic motion, 223
harmonic oscillator, 222 karst regions, 50
harmonics, spherical, 168 Kern county earthquake, 56,330
Hausdorff measure, 132, 135, 145, 146, 148,209, kinetic theory of gases, 140
372 Koch island, 11, 24, 146
Hawaii, 194,243,269 Koch snowflakes, 25, 107, 195
hazard rate, 303 pair correlation of, 107
hazards, flood, 208 Korcak relation, I5
seismic, 63, 66, 254, 307, 328, 329, 342
volcanic, 79 lacunarity, 109, 11I, 112, 372
Hazen distribution, 208. 21 1 moving box method for, 111
headward growth, 195 moving window method for, 109
heat, specific, 257 lag, 137, 138
heat conduction, 257 lake levels, 160
heat conduction equation. 23 1 lakes, number-size statistics of, 15, 96
392 INDEX
Landers earthquake, 56,57,63,329,330 magma migration, 193

landforms, 181 magnetic field, 4,217,279,342
landscapes, 2, 181 magnetic field polarity, 279
fractal, 181 magnetic field reversals, 3.279
roughness of, 2, 165, 176 correlation dimension of, 286
synthetic, 2. 173 magnetic surveys, 167
textures of. 112 magnetics, 3
landslides, 3 16,32 1 magnetism, fossil, 279
frequency-size distribution, of 316, 321 natural remanent, 279
Langevin equation, 207,208 magnitude, body wave, 57.58
Laplace equation, 4,23 1 earthquake, 57,58,61
layer, fluid, 256,268,269 local, 57,58
thermal boundary, 265,269 moment, 58
Legendre functions, 168 surface wave, 57,58
length of coastline, 1, 12 Mandelbrot, Benoit, 1.6, 15,341
length of day, 136,2 17 mantle convection, 256,269
length of faults, 67 chaotic, 269
length of perimeter, 11 mantle plume, 269,273
length-order ratio, 185, 187 mantle viscosity, 258
levels, lake, 160 map, 232,372
lifetimes. distribution of. 303 iterative, 232
limit cycle, 223,229,234,236,238,251,253, linear, 240
372 logistic, 4, 231,232,237, 293
linear damping, 222 recursive, 243,286
linear map, 240 tent, 241
linearization, 220,226,259 triangular, 24 1
linearized stability analysis, 220, 259 margin, continental, 321
lithosphere, elastic, 13 marginal stability, 259,3 16
load sharing, global. 304 mass conservation,202
local earthquake magnitude, 57.58 mass distribution of fragments, 42
lode gold, 92 maximum entropy. 50
logarithmic spiral, 225 Maxwell-Boltzmann distribution, 332, 333, 336
logistic equation, 219,226 mean-field approximation. 304
logistic map, 4.23 1,232,237,293 mean value, 29.30, 137, 138, 141
log-normal distribution, 2,35,38,42, 81, 83,90, exponential, 40
96, 137,145, 158,208,211 Gaussian, 32
coefficient of skew for, 35 log-normal, 35
coefficient of variation for, 35, 158 normal, 32
cumulative distribution function for, 36 Pareto, 38
for ore deposits, 83 meanders, 195
mean of, 35 measure, Hausdorff, 132, 135, 145, 146, 148,
probability distribution function for, 35 209,372
standard deviation of, 35 measuring rod, 1, 12
variance of, 35 mechanics, statistical, 289, 330,342
log-normal noises. 157, 158 medium, porous. 290
log-normal walks, 157, 158 memory, long, 137
log-Pearson distribution, 208.21 1 short, 137
log-periodic behavior, 303,305,307,342,372 Memphis earthquakes, 66
logs, well, 3, 136, 165, 166, 168 Menger sponge, 10,50, 105,295
Loma Prieta earthquake, 56,57,305, 329 mercury, 90.9 I. 95
long memory, 137 metamorphic veins, clustering of, 104
long-range correlation, 138 Mexican hat wavelet, 214
Lorenz, Ed, 3,256,341 migration, magma, 193
Lorenz attractor, 265,272,284 mineral concentration, 8 1.90, 120, 136
Lorenz equations, 3,256,262,263,264,266, mineral deposits, 1,5,81, 87,90, 146
269,270,284,341,372 epithermal, 87
Lorenz truncation, 268 fractal dimension of, 90
lunar craters, 44 frequency-sizedistribution of, 1.81
Lyapunov exponent, 3,239,240,241,25 1.37 1, hydrothermal, 87
372 skarn, 87
spatial distribution of, 98
magma, 87 mines, 67
INDEX
minimum energy dissipation, 207 side branching, 188

mining induced seismicity, 129 neutron activation well logs, 165
model, advective-diffusion,207 New Madrid earthquakes, 66
avalanche, 5,207,316,317 Nile River, 158
cellular-automata. 5,316,317,320,322,324, Noah effect, 37, 158,213
325,326 node, 224,372
comminution, 48,49,50,71,74,298 stable, 224
Culling, 202,207 unstable, 224
diffusion limited aggregation, 4,74, 195, 197, noises, 140, 149, 150, 152, 155, 157, 158, 161,
199,201,207 162,373
forest fire, 336,337,339 antipersistent, 152
sandpile, 5,316,317 correlated, 140
slider-block, 4,245,247,254,325,330,336, fractional Gaussian, 149, 150, 152, 155, 158,
342 161, 162
stochastic, 4 fractional log-normal, 157, 158
molecular velocities, 30, 3 1 Gaussian white, 140, 149, 150, 160,207
moment magnitude, 58,60 persistent, 140, 152
moment of inertia, 282 white, 138, 139, 140, 146, 148, 149, 150,
moments, 29, 110, 114,372 156,207.373
earthquake, 57,58,59,60,61,305 nondimensional parameters, 222, 248,258, 282
generalized, 114 nondimensional variables.. 219.222.246.248.
. . . .
seismic, 60, 305 258,282,332
Monte Carlo, 291 nondimensionalization, 219, 222,246, 248,258,
mother wavelet, 214 282,326,332
motion, equation of, 222,246,325 nonlinear damping, 222
moving-average model, 140 nonlinear equations, 4,219,23 1
autocorrelation function for, 141 normal distribution, 2,31,33, 38, 137, 138, 145,
autoregressive, 143 341,371.373
autoregressive integrated, 145 cumulative, 32
variance of, 141 mean of. 32
moving-box method, 111 probability distribution for, 3 1
moving-window method, 109 standard deviation of, 32
multifractal, 83, 113,372 variance of, 32
perfect, 124, 128 normal fault, 76
multifractal analysis, 120, 125, 129 no slip boundary condition, 260
multifractal analysis of fragmentation, 129 Northridge earthquake, 56.57.63.330
multifractal analysis of well logs, 129 nuclear explosion, 43,44
multifractal dimension, 115, 117, 124 number-size distribution, 1, 15,79, 291,317, 337
multifractal distribution of contact areas, 129 earthquake, 1,15,57,59,60,66,317,335
multifractal distribution of earthquakes, 128 fault. 67.322
multifractal distribution of fractures, 129 flood, 208
multifractal distribution of mining induced forest fire, 337,339
seismicity, 129 island, 15
multifractal distribution of pore spaces, 129 lake, 15,96
multifractal distribution of void spaces, 129 oil field, 1,95
multifractal scaling, 128 percolation cluster, 291
multifractal spectrum, 115 rock fragment, 1, 15.42
multifractal time series, 129 volcanic eruption, 1.79
multifractal topography, 129 numerical solutions, 264
multiplicative cascade, 120, 121, 124, 126, 128 Nusselt number. 264
mutual inductance, 281
ocean ridge, 13.56.70.269
Nankai Trough, 254 ocean surface, 146
natural remanent magnetism, 279 ocean trench, 13.56.70.269
Navier-Stokes equations, 23 1 oceans, 231
nearest neighbor model, 317 oil, spatial distribution of, 97,98
networks, branch numbers, 188 oil fields, 74,95,96,97
branch ratio, 189 fractal dimension of, 95
drainage, 2.5, 181, 185, 191, 194, 199, frequency-size distribution of, 1.95.96
207 oil reservoirs, 3,96,97
fractal dimension of, 185 Old Faithful Geyser, correlation dimension of,
river.2.5, 181, 185, 191, 194, 199,207 167
394 INDEX
Omori's law, 304 subcritical, 228, 272

order, 6 supercritical, 226
of rivers, 181 plane, phase, 249
ore concentrations, 1 , 5 , 8 1, 87,90, 120, 136, plate tectonics, 5, 56.60, 70, 256, 269, 279
I46 plumes, 269, 273
ore deposits, 1, 5, 81, 87,90, 120, 136, 146 Poincare section, 372
ore grade, 81.84.87.90. 136, 146 Poisson distribution, 102, 103
ore reserves, 8 1 polarity, magnetic field, 279
ore tonnage, 8 1,90,92 population dynamics, 4, 219, 232, 244
Oregon topography, 163, 165, 175, 177 pore spaces, 50
oscillator, harmonic. 222 pore spaces, multifractal distribution of, 129
oxygen isotope ratios, 167 porosity, 3,50, 136, 166, 167,295
critical value of, 295
pair-correlation technique, 106, 107 fracture, 50
Cantor set, 107 intergranular, 50
Koch snowflake, I07 porosity distribution, 3, 166, 167
seismicity, 107 porosity logs, 166, 167
paleomagnetism, 279 porous media, 290
parameters, nondimensional, 222, 248, 258, portrait, phase, 265
282 power law, 1, 372
Pareto distribution, 37, 38.42.341 power-law distribution, 2, 16,38,41,42, 50.81,
cumulative, 37 84,90,96,208,2 11, 34 1
mean of, 38 power-law spectra, 3, 166, 168
probability distribution for, 37 power spectral density, 148, 149, 150, 166, 168,
standard deviation of, 38 170,203
standard form of, 37.38 geoid, 170
variance of, 38 topography, 168, 170
Parkfield earthquakes, 67,254 two-dimensional, 168, 172
partial duration series, 209 Prandtl number, 258,268,269,270,273
particulate matter, 3 precursor, seismic, 305,328
partition coefficient, 87, 90 prediction of earthquakes, 254, 307,328,329,
pattern recognition, 328 342
pavements, 68 pre-existing faults, 245
peak discharge, 136,208 pressure, 160,289
percolation, invasion, 207 atmospheric, 160
percolation backbone, 78,291,336 pressure forces, 257
percolation clusters, 4.78, 290, 291, 372 probability, 4, 28.30, 100, 102, 113, 301, 372
number-size statistics of, 291 conditional, 301
percolation model, 333 critical, 4, 290, 295
percolation threshold, 333 distribution, 3 1
perimeter, length of, 11 exponential, 39
period, rotational, 136 Gaussian, 3 1
period doubling, 234,236,238, 25 1.37 1, 372 log-normal, 35
periodic behavior, 248,250,253,305 normal, 31
periodic component, I36 Pareto, 37
periodic oscillation, 222, 372 of fragmentation, 46, 292, 298
permeability, 4.50.74, 136, 166, 295,372 prograding delta, 203
critical, 4, 290, 295, 372 pull, trench, 65
fracture, 50, 166 pumice, 44
permeable element, 290 push, ridge, 65
persistence, 136, 137, 138, 140, 152, 161, 162
strong, 137 quiescence, seismic. 328
weak, 137
petroleum, spatial distribution of, 97, 98 radius of gyration, 196
petroleum reserves, 96 rainfall, 37, I36
petroleum traps, 98 random, 102, 136,373
phase plane, 249 random additions, successive, 152
phase portrait, 265 random Cantor set, 100, 102, 109, 111
phase space, 222,372,373 random growth networks, 74, 195
phase trajectory, 222 random simulation, 103
pitchfork bifurcation, 226,227, 228.25 I, 261, random walk, 2, 195, 197, 199,207, 371
263,265,272,372 Rayleigh distillation model, 87, 88, 89.95
INDEX
Rayleigh number, 258, 260, 264, 268, 269, 271. scale invariant distribution, 3
272,273 scaling, counter, 162
critical, 260, 264 multifractal, 128
rebound, elastic, 245, 246 scarpes, earthquake, 203
recursion relations. 3, 293 shoreline, 203
recursive map, 243,286 screen analysis, 42
regional seismicity, 5, 63, 324 sea-floor bathymetry, 2, 163, 165, 167, 168, 217
remanent magnetism, 279 sea level, 19, I67
renormalization, 47.83.89.289. 373 second law of thermodynamics, 289
renormalization group method, 5,289,290,292, sedimentary basement, 76
295, 299,316,317 sedimentary basin, 76
applied to fault rupture, 299 sedimentary bedding planes, 19, 321
applied to fragmentation, 295 sedimentary completeness, 24
repose, angle of, 3 16 sedimentary hiatuses, 18, 19.20, 167
rescaled range analysis, 138, 158, 161, 162,373 sedimentary pile, 25, 76
of fractional Gaussian noises, 161, 162 sedimentary record, 18, 19
reserves, ore, 8 1 gaps in, 18, 19.20, 167
petroleum, 96 sedimentary sequence, 22.23, 321
reservoir, oil, 3.96.97 fractal dimension of, 22,321
reservoir storage, 158 power-law correlation of, 23
resistance, 28 1 thickness statistics of, 321
reversals, magnetic field, 3,279 sedimentary unconformities, 18, 19,20,2 1
ridge, ocean, 13.56.70.269 sedimentation, episodic, 23
ridge push, 65 sediments, age distribution of, 18
Rikitaki dynamo, 3,279,280,282, 342 deposition of, 18.22.202. 321
rings, tree, 160 erosion of, 18, 19.23.202
river deltas, 193 subsidence of, 76
river discharge, 37, 136, 158, 160, 215 seismic activation, 329, 330
time series of, 158,215 seismic hazards, 63,66,254,307,328,329,342
river meanders, 195 seismic moment, 60, 305
river networks, 2, 5, 18 1, 185, 19 1, 194, 199, 207 seismic precursors, 305, 328
river sinuosity, 195 seismic quiescence, 328
rivers, braided, 207 seismicity, 1.4, 15, 39.56, 57.58, 59, 61, 65, 74,
multifractal analysis of, 129 103, 107, 128, 145, 317,320, 324,
order of, 18 1 325,326,327,328.329,330,341
rock fragments, 1, 15, 28.42 clustering of, 103, 328
size distribution of, 1, 15.42 crustal, 322
rock surfaces, 166 distributed, 1, 76, 107, 128, 320, 324, 325,
rocky coastlines, 1, 12, 15, 132 326,327
length of, 1, 12 fractal dimension, of 59, 76
rod. measuring. 1, 12 global, 59
Rosin and Rammler distribution, 40.42 induced, 5, 129,328
rotational period, 136 Memphis-St. Louis. 66
roughness, 2, 15, 165, 176 pair correlations of, 107
of coastline, 15 regional, 5.63.324
of topography, 2, 165, 176 southern California, 61.63.330, 336
roughness-range method, 162 seismograms, 166,217
route to chaos, 237 self affine. 373
ruler method, 12, 13, 14 self-affine fractal, 132. 135, 138, 140, 145, 146,
rupture, fault, 299 148, 150, 163, 166,372
rupture area, earthquake, 39.58 box counting method for, 132, 135, 147
Brownian walk as a, 140
saddle point, 225, 373 deterministic, 133
St. Louis earthquakes, 66 fractional Brownian walk as a, 140, 146, 150
San Andreas fault, 56,62,65,67, 104 fracture surfaces as a, 166
San Fernando earthquake, 56.330 sea level as a. 167
San Francisco earthquake, 56.57.330 statistical, 140
San Gabriel fault, 104 topography as a, 145, 178
sand pile model, 5.3 16.3 17 variance of, 146
sand piles, 316, 317, 321 self-affine fractal dimension, 135
scale invariance, 1, 11, 39, 84, 91, 289, 304, 3 16. self-affine tiling, 207
372,373 self-affine time series, 145
INDEX
self-excited dynamo, 279 spiral, 225

self inductance. 282 logarithmic, 225
self-organizedcriticality, 5, 316, 317, 325, 326, sponge, Menger, 10,50,105,295
328,329,336,341 spreading centers, 56
self similar, 1,373 spring-mass oscillator, 222
self-similar fractal, 132 stabilities, exchange of, 226
semivariance, 138, 155 stability analysis, 220, 221, 227.259.261
semivariogram, 138, 146, 155 linearized, 220,259
fractional Brownian walk, 155 marginal, 259, 3 16
fractional Gaussian noise, 155 stable fixed point, 220,225,232,265,294,297,
sequencies, depositional, 18 302
shear stress, 260 stable node, 224
shoreline scarps, 203 staircase, devil's, 18,21,23,24
short memory, 137 standard deviation, 29.30
side branches, 188 Brownian walk, 140
Sierpinski carpet, 9.24, 105, 111, 118, 119,291 exponential distribution, 40
sieve analysis, 42 Gaussian distribution, 32
simulation, random, 103 log-normal distribution, 35
singular point, 241 normal distribution, 32
sinkholes, 50 Pareto distribution, 38
sinuosity, 195 standard form, 33,37,38
skarn mineral deposits, 87 state, critical, 5,316
skew, coefficient of, 29.30.35 static friction, 246, 326, 330
for log-normal distribution, 35 stationarity, 138
slider-block array, 325,330,336 statistical fractal, 12, 38, 289
slider-block model, 4,245,247, 254, 325,330, statistical mechanics, 289,330,342
336,342 statistical self-affine fractal, 140
slider blocks, 245,247,325,330,334,342 statistics, 28
pair of, 247, 342 stick-slip behavior, 245, 246, 249
multiple, 325,330,334,336,342 stiffness parameter, 248,326
slip events, frequency-size statistics of, 325, Stirling approximation, 127
326 stochastic, 136, 137, 373
slumps, 321 stochastic component, 136, 137
snowflake, Koch, 25, 107, 195 stochastic models, 4
soil density, 52 storage, reservoir, 158
soils, 44.50.52 strain, 76,305
fractal dimension of, 44 Benioff, 305
multifractal fragmentation of, 129 stranded cable, 299,303,307
solid-state creep, 258 strange attractor. 265,284,373
southern California seismicity, 61, 63, 330, 335 stratigraphichiatuses, 18, 19,20, 167
space, phase, 222,373 stream function, 257, 260
space filling, 236 stream order. 181
spatial distribution of aftershocks, 77, 107 stream flow time series, 215
of earthquakes, 76, 128 stress, shear, 260
of gold, 98 strong persistence, 137
of mineral deposits, 98 subcritical pitchfork bifurcation, 228,272
of oil, 97 subduction zones, 13.56.70.269
specific heat, 257 subsidence, 76
spectra, power-law, 3, 166, 167 successive random additions, 152
spectral analysis, 2, 163 sunspot numbers. 160
of bathymetry, 2, 163 supercritical pitchfork bifurcation, 226
of topography, 2, 163 surface, free, 260
two-dimensional, 168 ocean, 146
spectral density, power, 148, 149, 150, 166, 168, surface wave magnitude, 57,58
170,203 surfaces, rock, 166
of geoid, 170 symmetry. 248
of topography, 168, 170 synthetic landscapes, 2, 173
of topography on Venus, 170
two-dimensional, 168, 172 Tangshan earthquake, 330
spectrum, Fourier, 2, 148, 163 tectonic fragmentation, 44, 70
multifractal, 115 tectonic processes, 13.56, 181
spherical harmonics, 168 tectonic uplift. 23
INDEX 397
tectonics, 13,50, 56 trace elements, 87

plate, 5.56.60, 70,256, 269, 279 trajectory, phase, 222
temperature, 31, 136, 146, 160,289 transcritical bifurcation, 226
atmospheric, 136, 146, 160 transform, discrete Fourier, 150, 171
Curie, 279 Fourier, 148, 150, 171,214
tent map, 24 1 inverse discrete Fourier, 150
tephra, 79 inverse Fourier, 148
textural analysis, 3, 112 two-dimensional Fourier, 171, 172
landscape, 112 wavelet, 214
thermal boundary layer, 265,269 transform faults. 56, 70
thermal conductivity, 257 Transverse Ranges, 65
thermal convection, 3,256,261,268,269 traps, petroleum, 98
thermal diffusivity, 258 trees, fractal, 2, 181, 187, 188, 299, 302
thermal expansion, 256,258 binary, 188
coefficient of, 258 branch numbers for, 188
thermally activated creep, 269 branching ratios for, 189
thermodynamics, 289 deterministic, 187
first law of, 289 length-order ratio of, 185
second law of, 289 side branching, 188
thickness statistics, 321 Tokunaga, 188, 190,198
thinning, crustal, 76 tree rings, 160
threshold, percolation, 333 trench, oceanic, 13, 56, 70, 269
tiling, self-affine, 207 trench pull. 65
time series, 136, 137, 140, 146, 148, 158, 167, trend component, 136
214,373 triadic Koch island, 11.24
atmospheric temperature, 136, 146 triangular map, 241
box-counting method for, 147 tributaries, 185
Brownian walk. 140 truncations, 262,268, 270
continuous, 136, 137 tuples, 167
correlation dimension of, 167 turbidite deposits, 321
correlations of, 158 turbulence, 3, 127, 23 1,268, 34 1
discontinuous, 137 turning-point bifurcation, 226
discrete, 136 two-dimensional Fourier transform, 171, 172
fractal dimension of, 147, 148
multifractal dimension of, 129 unconformities, 18, 19, 20.2 1, 167
river flow, 158,215 unstable. gravitationally, 256
self-affine, 145 unstable fixed point, 232,265,294, 294,302
stream flow, 158,215 unstable node, 224
time to failure, 303,304,305,306 uplift, 23
Tokunaga fractal tree, 188, 190, 198 uranium, 94.95
tonnage, ore, 8 1
tonnage-grade, 8 1,90,92 van der Pol equation, 221,223,229
copper, 93 variables, nondimensional, 219,222,246,248,
gold, 92 258,282,332
mercury, 90 variance, 29.30, 137, 140, 141, 145
uranium, 94 autoregressive model. 141
topographic contour, 12, 132 autoregressive moving average model, 145
topography, 2, 12, 13,56, 132, 145, 163, 165, Brownian walk, 140
168, 170, 175, 177,207 exponential, 40
Arizona, 177 Gaussian, 32
Brownian walk, 163,207,208 log-normal. 35
elevation of, 132, 136 moving average model, 141
Fourier spectral analysis of, 163 normal, 32
fractal, 132, 208 Pareto, 38
height of, 132, 136 self-affine fractal, 146
multifractal analysis of, 129 two-dimensional spectra, 168
Oregon, 163, 165, 175, 177 variation, coefficient of, 35, 137, 158
power spectral density of, 168, 170 varves, 160
roughness of, 2, 165. 176 veins, 70, 104, 193
self-affine, 145, 178 clustering of, 104
Venus, 170 velocities, molecular, 30.3 1
torque, 28 1 velocity, angular, 281
398 INDEX
velocity weakening friction, 299,325 wave equation, 4. 23 1

Venus, 163. 170 wavelength, 2
spectral density of topography, 170 wavelet transform, 214, 373
viscosity, 257, 258 bathymetric profile. 2 17
mantle, 258 fractional Brownian walk, 2 17
viscous fingering, 198 geiod, 2 17
viscous forces, 257 length of day, 217
void spaces, multifractal analysis of, 129 seismogram, 2 17
volcanic ash, 44.79 wavelets, 214, 2 17, 373
fractal dimension of. 44 Mexican hat. 214
volcanic edifices, 13 mother, 2 14
volcanic ejecta, 44 weak persistence, 137
volcanic eruptions, 1,28,78,79, 167,243 weather, 268, 341
correlation dimension of, 167 weathering processes, 44
frequency-volume statistics of, 1, 79 Weibull distribution, 41,42,299,301,303
volcanic hazard, 79 Weirstrass-Mandelbrot functions, 152, 165
volcanic islands. 194 well logs, 3, 136, 165, 166, 168
acoustic, 165
walk, Brownian, 2, 138, 140, 146, 147, 149, 150, electrical conductivity, 165
152, 155, 157, 158, 160, 163, 165, multifractal analysis of, 129
167,207,208,371 neutron activation, 165
correlation dimension of, 167 porosity, 136, 166, 167
fractional, 149, 150, 152, 155, 156, 158, 161, white noise, 138, 139, 140, 146, 148, 149, 150,
162, 166 156,207,373
log-normal, 157, 158 Gaussian, 140, 149, 150, 160, 207
random, 2, 195, 197, 199,207,371 Whittier earthquake, 63
self-affine fractal, 140, 146 window method, moving, 109
standard deviation of, 140 windows of chaos, 236,237,238,240.25 1
time series, 140
variance of, 140 Zipf's law, 26
UNIVERSITY PRESS
wwwcambridge.org
a.,
Ill"
a
'

Fractals and Chaos in Geology and Geophysics

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fractals and Chaos in Geology and Geophysics

Uploaded by

Copyright:

Available Formats

mn CM:9209630051

FRRCTRLS I D CHROS I N GEDLDGY

Cambridge University Press

O Donald L. Turcotte 1997

This publication is in copyright. Subject to statutory exception

First published 1992

Library of Congress Cataloguing in Publica~iondata

ISBN 978-0-521-56164-8 hardback

Transferred to digital printing (with amendments) 2007

I was introduced to the world of fractals and renormalization groups by Bob

students and the material used. Little mathematical background is required

A large number of new results on fractals, chaos, and self-organized critical-

been relatively few new advances concerning the fundamental aspects of

A stone, when it is examined, will be found a mountain in miniature. The

The scale invariance of geological phenomena is one of the first concepts

in a variety of ways. Diffusion-limited aggregation has been applied to the

2.1 Deterministic fractals

Since its original introduction by Mandelbrot (1967), the concept of fractals

where Ni is the number of objects (i.e., fragments) with a characteristic lin-

order r, = 31 and again N2 = 1. Thus as the order is increased and n + the

r2 = 91 and N2 = 4. From (2.2) we find that D = In 2An 3 = 0.6309. This is

Figure 2.2. Illustration of the

order i are replaced by the

structure is not continuous. An example of a continuous fractal construction

For the triadic Koch island illustrated in Figure 2.5 we have Po = 3, P I = 4,

D = 1 + log(P;+ ,/pi) = 1 + log(413) = 1 + log4 - log3 - -

2.2 Statistical fractals

tonic setting and age. Topography is primarily a result of erosional pro-

Figure 2.6. Length P of the

Figure 2.7. The lengths P

box-counting method to determine the fractal dimension of a rocky coast-

where D is again the fractal dimension. It is appropriate to use this cumula-

As we discussed, the term fractal dimension stands for fractional dimen-

Figure 2.8. (a) Illustration of

Figure 2.9. The number N

Figure 2.10. Worldwide

2.3 Depositional sequences

The relationship between deterministic and statistical fractals will be further

Figure 2.11. A fractal devil's

yw = Rt + y,, sin (2 T tIrJ

in the sedimentary record results. From t = 1,208,000 yr to t = 2,000,000 yr

A simple harmonic variation of sea level will lead to a periodic sequence

There is also an unconformity at a depth of 2 km. The sediments imme-

Combining (2.9), (2.10), and (2.1 1) we relate Ri to T~ with the result

which also can be written as

The rate of deposition Ri has a power-law dependence on the time interval

And from (2.14) we have D = 0.815.

2.4 Why fractal distributions?

In this chapter a number of deterministic fractal sets have been introduced

Figure 2.14. Dependence

Figure 2.15. Illustration of

3.2 Probability and statistics

Clearly the distribution of fragment sizes is a statistical problem. This will

The variance is the second-order moment of the distribution. The standard

The asymmetry of the data is quantified by the coefficient of skew y, which

The factor a 3 makes y a nondimensional number.

Introducing these probabilities, the mean, variance, standard deviation, and

and the probability Pr that x lies between x, and x2 is given by

Pr(x, < x 5 x,) = f(u)du = F(x,) - F(x2) (3.14)

Several widely used statistical distributions will now be discussed.

From statistical mechanics (Morse, 1969) the x-component of molecular ve-

The mean of the normal distribution is obtained using the relation

Introducing (3.18) gives

so that a is the standard deviation of the normal distribution.