FRACTALS AND
CHAOS IN
GEOLOGY AND
GEOPHYSICS
FRACTA
CHAOS
GEOLOGY A
GEOPHYSICS
Second Edition
DONALD L. TURCOTTE
Cornell University
CAMBRIDGE
UNIVERSITY PRESS
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
96-3 1558
550'. 1'51474 - dc20
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
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
17
Where do we stand?
References
Appendix A: Glossary of terms
Appendix B: Units and symbols
Answers to selected problems
Index
vii
PREFACE
PREFACE
PREFACETOTHE
SECOND EDITION
A large number of new results on fractals, chaos, and self-organized criticality 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 application of fractal techniques to geological problems. To make this textbook
more complete, a brief introduction to probability and statistics has been included in Chapter 3. Chapter 4 on seismicity and tectonics has been extensively 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 revisions concerns the treatment of self-affine fractals in Chapter 7. An introductory section on time series has been added as well as deterministic examples 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 treatment 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 filtering techniques. Chapters 9-14 remain essentially unchanged as there have
xii
Chapter One
SCALE INVARIANCE
SCALE INVARIANCE
the number of objects larger than a specified size has a power-law dependence on the size. The empirical applicability of power-law statistics to geological phenomena was recognized long before the concept of fractals was
conceived. A striking example is the Gutenberg-Richter relation for the frequency-magnitude statistics of earthquakes (Gutenberg and Richter, 1954).
The proportionality factor in the relationship between the number of earthquakes and earthquake magnitude is known as the b-value. It has been recognized for nearly 50 years that, almost universally, b = 0.9. It is now accepted 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 distributions that have been applied to geological phenomena. Other examples include the normal (Gaussian) distribution and the log-normal distribution.
However, the power-law distribution is the only distribution that does not include 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 rational 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 synthetic 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 complex 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 wavelength An,a fractal distribution may result. For topography and bathymetry it
is found that, to a good approximation, the Fourier amplitudes are proportional 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 increases 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 proportional to wavelength A,,.
SCALE INVARIANCE
Many geophysical data sets have power-law spectra. These include surface 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 threedimensional 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 inventor 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 emphasizing 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 approximate thermal convection in a fluid. This set of equations was the first to be
shown to exhibit chaotic behavior. Infinitesimal variations in initial conditions led to first-order differences in the solutions obtained. This is the definition of chaos. The equations are completely deterministic; however, because 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 comprehensive 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) exhibited 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 example 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 becomes 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
SCALE INVARIANCE
and chaotic behavior results. The logistic map and similar recursion relations 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 behavior can result. This is evidence that the deformation of the crust associated 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 invariant 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 twodimensional 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 critical 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 because 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
ChapterTwo
DEFINITION OF A
FRACTAL SET
where Ni is the number of objects (i.e., fragments) with a characteristic linear dimension ri, C is a constant of proportionality, and D is the fractal dimension. The fractal dimension can be an integer, in which case it is equivalent 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 fractional 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
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
the result is the same if the logarithm to the base e (In) is used or if the logarithm to the base 10 (log) is used. For the example considered in Figure
2.1 (a), In (N2/N,) = In 1 = 0, ln(r,lr,) = In 2, and D = 0, the Euclidean dimension of a point. This construction can be extended to higher and higher orders, but at each order i, i = 1,2, . . . , n, we have In (Ni+,INi)= In 1 = 0. As
the order approaches infinity, n + =, the remaining line length approaches
zero, rn + 0, becoming a point. Thus the Euclidean dimension of a point,
zero, is appropriate. The construction illustrated in Figure 2.l(b) is similar
except that the line segment of unit length at zero order is divided into three
parts at first order so that r, = f ;one part is retained so that N, = 1. At second
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.
In Figure 2.1 (c) the zero-order line segment of unit length is divided into
two parts but both are retained at first order so that r, = $ and N, = 2. The
process is repeated at second order so that r, = and N, = 4. From (2.2) we
see that D = In 2/ln 2 = 1. Similarly for Figure 2.l(d) we have D = 1; in both
cases the fractal dimension is the Euclidean dimension of a line segment.
This is appropriate since the remaining segment will be a line segment of
unit length as the construction is repeated. However, not all constructions
will give integer fractal dimensions; two examples are given in Figures
2.1 (e) and 2.1(f). In Figure 2.1 (e) the zero-order line segment of unit length
is divided into three parts at first order so that r, = 31 ; the two end segments
are retained so that N, = 2. The process is repeated at second order so that
4;
0 order
1 st order
2 nd order
3 rd order
4 th order
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 zeroorder 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
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 , =
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
square is removed; thus at first order r , = , N , = 8, and at second order r2 =
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;
1
thus at first order r , = 51 , N , = 9, and at second order r, = 9,
N2 = 81. From
(2.2) we have D = In 91ln 3 = 2. This is the Euclidean dimension of a square
and is appropriate because when we retain all the blocks we continue to re-
6,
5,
4,
4,
10
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 , = remain 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 Figure 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 remain1
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 construction is scale invariant.
The examples given above illustrate how geometrical constructions can
give noninteger, non-Euclidean dimensions. However, in each case the
i,
i.
11
where ri is the side length at order i and N is the number of sides. Substitution of (2. I) gives
(2.5)
This is the same result that was obtained above using (2.2), as expected. The
perimeter of the triadic Koch island increases as i increases. As i approaches
infinity, the length of the perimeter also approaches infinity, as indicated by
(2.4), since D >I (D is greater than unity). The perimeter of the triadic Koch
island in the limit i -+ is continuous but is not differentiable.
12
13
tonic setting and age. Topography is primarily a result of erosional processes; however, in young terrains topography is being created by active tectonic processes. It is not surprising that many of these processes are scale invariant and generate fractal topography. An interesting question, however, is
whether erosional processes and tectonic processes each generate topographies 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 modeled 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 fractal statistics and are therefore scale invariant.
14
Although the ruler (divider) method was the first used to obtain fractal
dimensions, it is not the most generally applicable method. The box-counting 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
box-counting method to determine the fractal dimension of a rocky coastline. 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 results 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 cumulative 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 distribution 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 correlation 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) taking D = 1.55.
15
16
As we discussed, the term fractal dimension stands for fractional dimension. 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 distributions 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
17
all power-law distributions that satisfy (2. I ) or (2.6) are fractal. In this book
18
;,
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 represent 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 hypothesize 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 occurring, 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 previously deposited sediments would rise above sea level. We assume, however, that erosion is sufficiently rapid that the rising landscape is maintained 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
19
20
Age 5 Myr
21
Age T, M yr
Depth
Y
km
22
There is also an unconformity at a depth of 2 km. The sediments immediately above this unconformity have an age T = 1 Myr and immediately below 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 sediments 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
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
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 deposition 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 function 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
in our model is equivalent to the line segment length riin the fractal sets illustrated in Figure 2. l . For the example given in Figure 2.13 we have T,, T, =
~ ~ 1and
3 , T, = ~ d 9The
. thickness of sediments Li is given by the number of
segments retained at a specified order N so that
T;
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
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
23
24
D E F I N I T I O N OF A FRACTAL S E T
tics of sediment deposition. The erosional processes responsible for the formation of coastlines and the depositional processes responsible for the structure of a sedimentary pile are both extremely complex. But despite the complexity, both examples exhibit fractal behavior to a good approximation. A
simple explanation is that a distribution will be fractal if there is no characteristic length in the problem. The fractal distribution is the only statistical
distribution that is scale invariant. However, a broad class of nonlinear physical problems involving chaotic behavior and/or self-organized critical behavior invariably yield fractal behavior. One objective in succeeding chapters is to describe physically realistic models that generate fractal behavior.
Problems
Problem 2.1. Consider the construction illustrated in Figure 2.l(e). (a) Illustrate 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) Illustrate 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.
(c) Determine the fractal dimension.
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,.
(c) Determine the fractal dimension.
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 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 2.8. A unit square is divided into nine smaller squares of equal size.
The center square and four corner squares are retained and the construction 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.
25
26
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 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 2.12. A unit square is divided into 25 smaller squares of equal size.
All the squares on the boundary and the central square are retained 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 2.13. 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 equilateral triangle with sides of length three. Equilateral triangles with sides
of unit length are placed in the center of each side. (a) Illustrate this construction 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 condition (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
unit square is considered at zero order and the first-order fractal construction is also illustrated. (a) Illustrate the construction at second order. (b) Determine No, N,, N,, r, r,, r,, Po, P I ,P,. (c) Determine the fractal dimension.
Problem 2.16. Assume that the open squares in the Sierpinski carpet illustrated 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 construction 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 unconformities? (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?
27
ChapterThree
FRAGMENTATION
3.1 Background
To illustrate how fractal distributions are applicable to real data sets, we consider 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 fragmentation is of considerable economic importance and many experimental, numerical, and theoretical studies have been camed out on fragmentation, relatively 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 process requiring complex models for even the simplest configuration. Fragmentation 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 material 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).
FRAGMENTATION
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
29
30
FRAGMENTATION
In many cases, each of the values x , , x,, . . . ,xi, . . . ,xn will have a probability of occurring f , , f,, . . . ,&, . . . ,fn. By definition we have
FRAGMENTATION
--
It should be noted that the usual definition of a cumulative distribution function 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
6x) < x 5
f(u)du = F(x,)
F(x2)
(3.14)
From statistical mechanics (Morse, 1969) the x-component of molecular velocities in a room satisfy a Gaussian distribution of the form
where m is the molecular mass, k the Boltzmann constant, and T the temperature.
31
32
FRAGMENTATION
Introducing
Y=-
x-x
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
=
f(x)dx =
1
o (2n)lD
I:m
(3.20)
FRAGMENTATION
33
34
FRAGMENTATION
FRAGMENTATION
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 lognormal 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
35
36
FRAGMENTATION
2 112
+ c,)]
we obtain
The cumulative distribution functions for the log-normal and normal distributions 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 distribution. The probability distribution functionsflx) for the log-normal distribution 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 distributions depending on the appropriate value of c,. This has important impli-
FRAGMENTATION
37
38
FRAGMENTATION
Fb) = (1
y)'
y20
Y a dy
(1
+ y)'+l
for a > 1
This integral does not converge and a mean does not exist for a I1.The variance of the standard form of the Pareto distribution is given by
(a - l)(a - 2)
for a > 2
(3.45)
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 distribution 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
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 association between statistical fractals and the Pareto distribution, there are
FRAGMENTATION
39
many other aspects of fractal concepts. The wide applicability of scale invariance 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 requires 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 complete probability or cumulative distribution functions for earthquakes. However, it is possible to determine the number of earthquakes N that have rupture 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
'
vxv-
f(x)
I[-"):(
7
ex*
xo
x 20
m)
40
FRAGMENTATION
where the power v is generally taken to be an integer. The mean of the exponential distribution is given by
we obtain
where r (v') is the tabulated gamma function (Dwight, 1961, Table 1005). If
v = 2 we have r
= 0.886 so that 2 = 0.88%. The variance of the exponential distribution is given by
(i)
This is known as the Rosin and Rammler (1933) distribution and it is used
extensively in geostatistical applications. We can also write
FRAGMENTATION
1 - F(x) = 1 - exp -
41
(3'1
-
42
FRAGMENTATION
This power-law mass relation can be related to the fractal number relation
FRAGMENTATION
43
When data are obtained by sieve analyses, (3.64) is used to convert mass distributions to number distributions to specify a fractal dimension.
Many experimental studies of the frequency-size distributions of fragments 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 distribution 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 resulting 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 magnitude. 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 shatFigure 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
Object
Reference
Fractal
dimension
D
Hartmann (1969)
Hartmann (1969)
Hartmann (1969)
1.89
2.13
2.22
Schoutens (1979)
2.42
Schoutens (1979)
Bennett (1936)
Klacka (1992)
2.50
2.50
2.50
2.55
2.56
2.60
2.61
2.80
2.82
2.88
3.54
FRAGMENTATION
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 related 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 fractal (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 integration gives
where C is a geometrical factor depending on the average shape of the fragments. For a power-law distribution, substitution of (3.61) into (3.67) and integration 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 distributions of fragments (see Table 3.2) the surface area of the smallest fragments dominates.
45
46
FRAGMENTATION
where Vo is the volume of the zero-order cells. The probability that a zeroorder 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 fragmented, No,, is given by
FRAGMENTATION
for these smaller cubes. The problem is renormalized and the cubes with dimension h/2 are treated in exactly the same way that the cubes with linear dimension 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 divided into eight first-order cubic elements with linear dimensions h/4 as illustrated in Figure 3.6. The volume of each first-order element is
After fragmentation the number of first-order cells that have not been fragmented is
Taking the natural logarithm of both sides we can write (3.75) and (3.76) as
47
48
FRAGMENTATION
FRAGMENTATION
49
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 fragments 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 implication 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 fragmentation. 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 fragments. 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 maximumentropy 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, intergranular porosity and fracture porosity. Based on the discussion given above
it would not be surprising if both types of porosity exhibited fractal behavior. Fractures are directly related to fragmentation, and detridal rocks are
composed of rock grains with a variety of scales. Based on laboratory studies a number of authors have suggested that sandstones have a fractal distribution 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
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
52
FRAGMENTATION
where r is the linear dimension of the sample considered. Similarly, the density of the fractal medium scales with its size according to
FRAGMENTATION
Problems
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
53
54
FRAGMENTATION
a. Show that
FRAGMENTATION
+,
+,
+,
55
Chapter Four
SEISMICITY
AND TECTONICS
4.1. Seismicity
A variety of tectonic processes are responsible for the creation of topography. 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 concern 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 (Turcotte, 1993, 1994a, 1995).
According to plate tectonic theory, crustal deformation takes place at
the boundaries between the major surface plates. In the idealized plate tectonic model plate boundaries are spreading centers (ocean ridges), subduction zones (ocean trenches), and transform faults (such as the San Andreas
fault in California). Relative displacements at subduction zones and transform faults would occur on well-defined faults. Displacements across these
faults would be associated with great earthquakes such as the 1906 San Francisco earthquake. However, crustal deformation is more complex and is usually 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 deformation 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 mountain 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 associated 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.
Although the crustal deformation in the western United States may appear 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 frequency-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 reasons 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 radiated 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 monitor earthquakes as small as magnitude two or less. The global seismic network 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 (Gutenberg 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 occurring 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 popular 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 magnitude 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 regional level of seismicity.
57
58
(4.2)
where Es is in Joules. The strain released during an earthquake is directly related 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 established 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 magnitude 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 approximation to relate the moment of an earthquake to the area A of the rupture by
with
bd
log@=C
+ loga-
b
-1oga
C
59
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 fractal dimension of distributed seismicity is
Thus the fractal dimension of regional or worldwide seismic activity is simply 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 earthquakes per year with magnitudes greater than m is given in Figure 4.1 as a
60
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 relative velocity. From (4.1) and (4.4) we have
and
Since c > b the integral diverges for large rn so that the maximum-magnitude earthquake mmaxmust be specified. This is the well-known observation
that a large fraction of the total moment and energy associated with seismicity occurs in the largest events. Integration of (4.15) gives
61
62
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 earthquakes 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 northeast of Los Angeles (Sieh et al., 1989). In addition to historical great earthquakes on January 9, 1857, and December 8, 1812, additional great earthquakes 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 repeat time is 132 years, giving N = 0.0076 yr-1. The most recent in the sequence 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 earthquakes 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-
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 relatively 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 implications. If a great earthquake substantially relieved the regional stress, then it
would be expected that the regional seismicity would systematically increase as the stress increased before the next great earthquake. An alternative 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 earthquakes not associated with aftershocks, have little time dependence. This
hypothesis will be discussed in Chapter 16. Acceptance of this hypothesis allows the regional background seismicity to be used in assessing seismic hazards (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 occur 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 regional 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 exceptions can be attributed to the aftershock sequences of the Whittier (1987),
Landers (1992), and Northridge (1994) earthquakes. Comparing the correlation 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 uniformity 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 velocity 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 seismogenic zone to be 15 km, the length of the seismogenic zone corresponding to
63
64
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.
1.
2.
65
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 transmitted 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)
66
ample was the three great earthquakes that occurred in the MemphisSt. Louis (New Madrid, Missouri) seismic zone during the winter of 18111812. Nuttli (1983), based on historical records, has estimated that the surface wave magnitudes of these earthquakes were 8.5, 8.4, and 8.8, respectively. 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 statistics 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 probability 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 recurrence 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 estimate recurrence times for earthquakes of various magnitudes.
N
yr"
4.2 Faults
There are two end-member models that give fractal distributions of earthquakes. 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 hypothesis. 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 earthquake 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 earthquakes. Equal fractal dimensions would imply that the interval of time between earthquakes is independent of scale. This need not be the case. Tectonic models for a fractal distribution of faults have been proposed by King
(1983, 1986), Turcotte (1986b), King et al. (1988), and Hirata (1989a). Fractal 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. Fractures 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 earthquakes, 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 provide 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 numberlength 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 microfractures in laboratory experiments that stressed unfractured granite.
67
68
Systematic studies of the statistics of exposed joint and fault-trace patterns over many orders of magnitude in length scale have been given by Barton (1995). Bedrock exposures were created on Yucca Mountain, Nevada, by
the removal of soil and debris creating exposures known as pavements. Barton (1995) mapped these exposures to obtain the two-dimensional distribution 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 fracture 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-
69
70
ments that include points (fractures) N(r) is determined and log N(r) is plotted 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 Perseverance Mine, Juneau, Alaska. His results for core hole 7- 18, using the onedimensional box-counting technique, are given in Figure 4.8. A good correlation 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 responsible for the formation of joints and faults and not all would be expected to
yield fractal distributions. Limitations of the fractal approach have been discussed 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-
D = 1.7.
71
veloping. The result is that plate interiors must deform to accommodate the
geometrical incompatibilities. Because of the weaker silicic rocks of the continental 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 comminution model for fragmentation given in Figure 3.7. To fragment the single 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 further fragmented; this requires N, = 3 X 6 = 18 faults of size r, = hl2. The result is 48 blocks of size hl4; 36 of these 48 blocks are further fragmented to
llr (rn-')
0.59.
72
we see that the fractal dimension of the cross section, D2 = In 3lln 2, is related to the fractal dimension of the original construction from Figure 3.3.,
D,= In 6 t h 2, by
74
Similarly we assume that for earthquakes De= 2 so that the number of earthquakes per unit time, in a given area, with a characteristic rupture size
greater than r scales with r according to
75
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 (fractal 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 correlation 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
76
additional parameter, the critical shear stress for fault propagation. These authors correlate fault displacement and length in individual tectonic environments 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 environments. In addition, Gillespie et al. (1992) find a universal power-law
correlation between fault width and total displacement. Jackson and Sanderson (1992) and Pickering et al. (1994) have concluded that in several examples 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, resulting 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 associated with the observed crustal thinning. Typically only 40-70% of the required extension can be associated with the larger faults on which displacements 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
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.
77
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 sequences; 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 dimension of two would be expected if the earthquakes lie on a plane; however, there is considerable three-dimensional structure to the aftershock seshock
78
quences. This led Robertson et al. (1995) to suggest that the earthquakes
form the "backbone" of a percolation cluster. The "backbone" of a threedimensional 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-
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-volume 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 determine 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 recurrence 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 recurrence time for a characteristic earthquake of magnitude six? Take c =
1.5, d = 9.1.
79
80
Chapter Five
ORE GRADE
ANDTONNAGE
82
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 enrichment: 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
The enrichment factor must be in the range 1 < 4, < 2. This model is illustrated 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 enriched 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 concentration
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 logarithmic 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 discussed 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 concentrated 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 fractal (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
83
84
where C,, is the mean ore grade associated with the mass
and
With the density assumed to be constant, M r3, where r is the linear dimension of the ore deposit considered, and we have
Since the allowed range for 4, is 1 < 4, < 2, the allowed range for the fractal dimension is 0 < D < 3. To be fractal the distribution must be scale invariant. 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 divided 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
The mass ratio a can take on the range of values 1 < a < m. The concentration ratio is defined as before and is the ratio of the concentration in the enriched element C , , to the reference concentration C,.
The enriched zero-order element becomes a first-order cell; this cell is divided into two parts with the enriched part having a mass
7- M0
a2
Mll
M21=
85
86
and
and
+,
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 concentration into one-eighth of the original mass, $, must be equivalent to
three orders of the concentration into one-half the original mass, 4,. Thus we
can write
It follows that
where M is the mass of magma, Mm is the mass of the mineral in the magma,
and KR is the solid-liquid partition coefficient. If KR < 1 the remaining
magma is systematically enriched. The allowed range of values for the
87
88
solid-liquid partition coefficient is 0 I KR I 1 . If KR = 1 there is no enrichment of the remaining magma and the concentration of the mineral is constant. If K , = 0 the concentration of the mineral in the solid is zero until the
melt contains only the mineral.
We can write (5.29) as a differential equation in the form
The concentration of the mineral in the enriched residual magma Cmand the
concentration of the mineral in the original magma Cmoare given by
and
c$a=a+(l -a)KR
Using this result the power in (5.24) is given by
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 substitution of (5.42) into (5.25) gives
89
90
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.
where M is the mass of the highest grade ores, which have a mean concentration
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 Harris (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 dependence. These authors based their analyses on records of annual production 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
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.
c.
c,
91
These computations represent the two data points farthest to the left in Figure 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 deposits 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 concentration in the continental crust. The mean crustal concentration of mercury as
= 8 X 10-8 (0.08 ppm). Using this value in
given by Taylor (1964) is
(5.44) we find that the correlation line in Figure 5.2 is given by
c,
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 original 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
92
processes responsible for the enrichment of mercury ore deposits are restricted 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 Figure 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 dimension is somewhat less than the value obtained for mercury, indicating a
smaller enrichment factor.
Again, the mean crustal concentration is introduced as a reference concentration. Taking
= 3 X 10-9 (3 ppb) (Taylor and McLennan, 1985) for
gold, we find the correlation line in Figure 5.3 is given by
co
93
with M in kilograms. According to the fractal model the lode gold in the
94
with M in kilograms. According to the fractal model, the copper ore in the
United States has been concentrated from continental crust with a mass
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 cumulative 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 enrichment 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 underestimation 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 uranium in the upper crust as given by Taylor and McLennan (1981) is C, =
1.25 X
(1.25 ppm). Using this value in (5.44), we find that the correlation line in Figure 5.5 is given by
with M in kilograms. According to the fractal model the uranium ore in the
United States has been concentrated from continental crust with a mass
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 original 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 uranium 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, chromatographic and Rayleigh distillation, both lead directly to fractal distributions.
The examples considered here yield a considerable range of fractal dimensions: 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 concentration 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).
95
96
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 fractal statistics to petroleum reserves can have important implications. Reserve
estimates for petroleum have been obtained by using power-law (fractal) statistics and log-normal statistics. Accepting power-law statistics leads to considerably 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 illustrates the difficulties associated with restrictions on power-law .(fractal)
distributions. As stated previously, we define a power-law statistical distribution 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 generally 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
V km3
Barton and Scholz (1995) have examined the spatial distribution of hydrocarbon accumulations and have concluded that they obey fractal statistics. 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 northeast comer of Colorado and the southwest comer of Nebraska. A 40 X 40mile 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 illustrated 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.
98
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 petroleum traps leads to a fractal spatial distribution of oil pools. Barton and
Scholz (1995) also examined the spatial distribution of hydrocarbon accumulations in the Powder River basin, Wyoming, and found a good correlation 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 , corresponding to an ore deposit with D = 2.
Problem 5.4. Determine the solid-liquid partition coefficient K , corresponding 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 elements? (b) What is the concentration in the seven depleted first-order elements? (c) What is the allowed range for $,? (d) What is the corresponding 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),
= 1 ) would be expected in the United States?
how much pure gold
(c
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) determine 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?
99
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 determine 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 probability that a step of len th r, = 1 will encounter a line segment, p, = 1; at
1
first order we have r , = 2 and p , = 2,
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, =
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 order 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
For the Cantor set the probability that a step of length ri = (f)' encounters a
line segment is pi= ($1' so that D = In 2Iln 3 as was obtained previously.
The Cantor set is both scale invariant and deterministic. Its deterministic
aspect can be eliminated quite easily. A scale-invariant random set is generated
by randomly removing one-third of each line rather than always removing the
FRACTAL CLUSTERING
101
middle third. This process is illustrated in Figure 6.1. The fractal dimension is
unchanged and the probability relations derived above are still applicable.
We will use the examples given above as the basis for studying fractal clustering. We consider a series of point events that occur at specified times. To consider N point events that have occurred in the time interval i0we introduce the
natural period T , = i d N . We then introduce a sequence of intervals defined by
IIIIIIIIIIIIIIIIIIIlllIllllm
11I
102
FRACTAL CLUSTERING
where
FRACTAL CLUSTERING
103
clustering by the "box method," we take intervals of length r,, = 2" and determine the fraction p that include at least one ninth-order element as a function
of rn. An example is given by the open circles in Figure 6.2. The best-fit
straight line has a slope of 0.368 so that p T-0.368 and D = 0.632. The deviation from the exact value D = 0.6309 for the deterministic Cantor set is due
to the reduced rate of curdling in the probabilistic set. If the same number of
ninth-order elements is uniformly distributed (no clustering), the probability
of finding an element with an interval from (6.4) is given by the solid circles
in Figure 6.2. In this case, the slope is unity for r < (;)9 and zero for r > (;)9.
Thus, D = 0 for r < (;)9, that is, a set of isolated points, and D = 1 for r > ($)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 temporal 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 considered. 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 significantly different from the earthquake data and is close to the uniform distribution.
104
FRACTAL CLUSTERING
FRACTAL CLUSTERING
Fractal clustering can also be studied in higher dimensions. The application 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 construction given in Figure 2.3(a). At zero order the probability that a box
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 construction 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
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 previously found. The Sierpinski carpet can be applied to clustering in two dimensions in the same way that the Cantor set was applied in one dimension.
This is directly analogous to the box-counting algorithm discussed in Chapter 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 probability that a cube of size ro = 1 will include material is p - 1;at first order we have
20
01400
r , = 51 and p , = 2
, and at second order we have r, = 9 and p, = m. The probability that a cube of size ri includes retained material can be generalized to
105
106
FRACTAL CLUSTERING
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-correlation distribution C(r),which is defined to be the number of pairs of points
whose separation is between r - $Ar and r + i h r per unit area (Vicsek,
1992). One point is picked and the distances to all other points are determined. The same thing is done for the second point and for all other points.
The number of pairs with separations between r - khr and r + t h r is divided by Ar. This result is the pair-conelation distribution C(r) in one dimension. For a two-dimensional distribution the number in each interval A r
is divided by r Ar to obtain C(r);for a three-dimensional distribution the
number in each interval Ar is divided by r2 Ar to obtain C(r).
Two simple deterministic examples illustrate how pair-correlation distributions are determined. First consider the one-dimensional example of four
equally s aced points on a line of unit length. The pair-correlation distribution is C ( J )= 6 , C($)= 4, C ( l )= 2. Next consider the two-dimensional example of four points on the corners of a unit square. The pair-correlation distribution is C(1)= 8, ~ ( f =i 4)1 f i = 2 f i .
If the points are randomly distributed in space, the pair-correlation distribution is exponential
FRACTAL CLUSTERING
107
We will consider two examples of scale-invariant (fractal) pair-correlation 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 correlation 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 snowflake 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 distribution 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.
,-.
-
slope=-0 369
a=O 369
u m
QQ0
4
w
o
\
L.p-L
D=l-a=O 631
.
L
o9
12
'
18
log r
2 1
74
108
FRACTAL CLUSTERING
0.0
20.0
40.0
60.0
80.0
col
0.2
0.4
0.6
1 .O
0.8
log r
1.2
1 4
1.6
FRACTAL CLUSTERING
109
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 illustrated 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. Mandelbrot (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 distribution (near zero lacunarity) and a totally clumped distribution (high lacunarity). 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 lacunarity, 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 Allain and Cloitre (1991) is illustrated in Figure 6.8. We consider the thirdorder Cantor set given in Figure 6.7(b). The total length is r, = 27 and individual 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
110
FRACTAL CLUSTERING
FRACTAL CLUSTERING
111
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
0 Uniform
x Cantor
0
0
Random Cantor
o Clumped
112
FRACTAL CLUSTERING
+ 1)2
(6.21)
T.
FRACTAL CLUSTERING
113
6.4 Multifractals
Our analysis of lacunarity introduced higher-order moments to our considerations of self-similarity and fractals. The utilization of higher-order moments 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
i=l
:.
a.
114
FRACTAL CLUSTERING
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 illustrated 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,
f3= i. With n = 9 ( r = ,$)we have i = 1, 2, 3, 4, 5, 6, 7, 8, 9; from Figure
2
6.11(d) we obtain L,= &, L2 = 0,L3 = $, L4 = L5 = L6 = 0,L, = 27,
L8 = 0,
L, = &; and from (6.22) find f, f2= 0,f3=:, f4=f5=f6= 0,f, = $, f, = 0,
f ='.
9 4
4,
=a,
where the sum is taken over the set of fractions and q is the order of the moment; (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 zeroorder 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
r = 1, and ,$ with the result:
i,
M, ( I ) = 12 = 1
FRACTAL CLUSTERING
= ffq-I
=f
= f exp
[(q - 1 ) ln f ]
+E
(6.25)
+ 1 (q
= f [ l + ( 9 - 1) l n f l
xfi
n
and since
= 1 we have
i=l
where the values of Dq form the multifractal spectrum. For q = 0, Do is referred to as the "box" dimension; it is entirely equivalent to the fractal dimensions considered in previous chapters. This is because the probabilityp
115
116
FRACTAL CLUSTERING
(3)
(i)
1
, ( f )= i, and M2(g)
= 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
(6.30).We first consider the quantity In [Mq (rj)lMq(rk)]
in the limit q + 1 ;
using (6.28)we have
FRACTAL CLUSTERING
where
s(f)= -
117
118
FRACTAL CLUSTERING
(i)
= - (a)
In ($) - 0 In 0 - ($)In
(a) - 0 In 0 - 0 In 0
0 In 0
FRACTAL CLUSTERING
where A, is the area in box i and A is the total retained area. Again
Ai = A
so that
Z J= 1
i=l
g.
119
120
FRACTAL CLUSTERING
S -
1n
1n
1n
(f) 1n
(3
+,
+,
FRACTAL CLUSTERING
121
where Ciis the mean concentration in segment i and Cois the overall mean
concentration.
n=4
We now determine the values off;. for n = l(r = I),n = 2 (r=
(r =
and n = 8 (r =
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,
i),
a),
k).
+,
(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 =
(1)+2 [l - (4) +2]2, fa= [l - (1)+2]3. These results are illustrated in Figure
c2= c3= +2
+;
+;9
We can now determine the generalized moments for the De Wijs multiwe find for q = 0 that
plicative cascade. From (6.23)
a.
122
FRACTAL CLUSTERING
+ [(1
- ;+2)3]0
+ 3 + 3 + 1= 8
0 and # 2. For q = 2
FRACTAL CLUSTERING
+,
s(1)
= -
llnl =O
123
124
FRACTAL CLUSTERING
which is valid for both integer and noninteger values of q except q = 1. For
q = 1 we find from (6.37) that
FRACTAL CLUSTERING
2.
3.
4.
5.
125
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).
The generalized moments Mq (r) are obtained using (6.23) if q ;t 1
and (6.28) if q = 1.
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.
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-invariant 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
+,
126
FRACTAL CLUSTERING
N, =
(7)
n!
j!(n - j)!
+:
x4
n
i=O
x
n
N, = 1 rather than
f ; = 1 as in the
i=l
FRACTAL CLUSTERING
f@,>
- j)!
+,.
= n log n -
logj!
=j l o g j
-j
log (n
- j)! =
(n
(6.55)
(6.56)
- j)
(6.57)
log (n - j) - (n - j)
+, x 1% ( 2 -
a = 1 - (1 - x)----log
log 2
Aa)
= &[x
log
(T)1 - x
$2)
log 2
- log ( 1
x)
x < 1.
The dependence offla) on a from (6.58) and (6.59) is given in Figure
= 1.2, 1.4, and 1.6. This result is clearly analogous to the depen6.15 for
dence of a probability distributionflx) on x, for example the log-normal distribution 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.
+,
+,
127
128
FRACTAL CLUSTERING
FRACTAL CLUSTERING
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 PyrakNolte 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) applied multifractal scaling to time series, and Muller (1992) and Saucier and
Muller (1993) have applied the technique to the characterization of geological formations using well logs.
The applicability of multifractal statistics to a natural phenomenon may
provide important clues to the underlying physical processes. However, multifractal 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,
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 retained. 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
of length r includes a line segment for r = &,
5,
3, A?
129
130
FRACTAL CLUSTERING
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 repeated. What is the probability that a square box with dimensions r includes a retained square when r = 1, $, i?
Problem 6.8. A unit square is divided into 25 smaller squares of equal size.
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, $,
,,
:,
I I?
49
'
Problem 6.10. A unit cube is divided into 27 smaller cubes of equal volume.
All the cubes are retained except for the central one and the construction
is repeated. What is the probability that a cube with dimensions r includes solid when r = 1, &,
;?
Problem 6.11. What is the pair-correlation distribution for three equally
spaced particles on a line of unit length?
Problem 6.12. What is the pair-correlation distribution for three particles on
the comers of an equilateral triangle with sides of unit length?
Problem 6.13. What is the pair correlation distribution for the eight particles
on the corners of a unit cube?
Problem 6.14. Consider the third-order Cantor set illustrated in Figure
6.1 l(a). Determine M3 (I), M3 ($)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 Cantor 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 ($1,Do, Dl, andD2.
Problem 6.18. Consider the second-order Sierpinski carpet illustrated in Figure 6.12a. Determine M3 (1) and M3 ($);determine D,.
Problem 6.19. A unit square is divided into four smaller squares of equal
size. Two diagonally opposite squares are retained, A = Determine Ai
4.
FRACTAL CLUSTERING
(i),
(i),
(a),
+, = 1.5.
4.
+,
:+,
+,.
+,.
(i),
131
Chapter Seven
SELF-AFFINE
FRACTALS
SELF-AFFINE FRACTALS
133
To illustrate this scaling we consider the deterministic example of a selfaffine 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 defined 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
54 --
V.E. = 1:270
394.5
4
0
4
X,
(4
10
134
SELF-AFFINE FRACTALS
fractal illustrated in Figure 7.2(c) each straight-line segment in the firstorder fractal has been replaced by the generator. At second order the horizontal 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-
SELF-AFFINE FRACTALS
135
i.
find^=+.
136
SELF-AFFINE FRACTALS
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 variable that changes with time. Our consideration of time series is a direct extension 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 continuous 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 continuous time series over a specified time interval T. Values are usually specified 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 maximum 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 entirely 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 exceptions, 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 applications (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 included 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
SELF-AFFINE FRACTALS
the stochastic component is persistent. If adjacent values are negatively correlated, 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 realization. In general, the ergodicity of practical time series is assumed, although it is often difficult to prove.
The values of the stochastic component of a time series will have a statistical distribution of values and the discussion given in Chapter 3 is applicable. Typical distributions are Gaussian (normal) and log normal. The mean,
variance, and coefficient of variation of the distribution of values can be determined. 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
137
138
SELF-AFFINE FRACTALS
and
Note that neither the mean j nor the variance V is used in this definition. For
a discontinuous time series we have
SELF-AFFINE FRACTALS
139
structed with a set of y, = E ~ i, = 1, 2, . . . , n, then adjacent values are uncorrelated 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 numbers has been used. The classic example of a nonstationary time series is the
140
SELF-AFFINE FRACTALS
Brownian walk. A Brownian walk is simply obtained by summing a Gaussian white-noise sequence; the values in a Brownian-walk time series yBw are
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 sequence 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 contaminant 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
SELF-AFFINE FRACTALS
rk = - for
1 + 0;
r, = 0
for
k = l
k >1
(7.18)
The correlation is very short since only the adjacent point has a non-zero autocorrelation 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
The mean for this correlated noise is again and its variance is
The AR model has longer range correlations than the MA model, but the correlations remain short range. Examples of this time series with 4, = 0,0.2,0.5,
141
142
SELF-AFFINE FRACTALS
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 correlations increase systematically with increasing values of as expected.
+,
2
Yn 0
-2
-4
-6
0
128
256
(4
384
512
SELF-AFFINE FRACTALS
143
256
384
512
n
6
4
2
Yn
0
-2
-4
-6
0
128
256
n
fd)
384
512
144
SELF-AFFINE FRACTALS
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
SELF-AFFINE FRACTALS
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). Topography is an excellent example of a self-affine fractal.
145
146
SELF-AFFINE FRACTALS
;--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
SELF-AFFINE FRACTALS
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 =
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 standard 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
147
148
SELF-AFFINE FRACTALS
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 contribution to the total energy of y(t) from those components with frequencies between 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 spectral density of y(t) is defined by
in the limit T + -. The product S( f)df is the power in the time series associated 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:
SELF-AFFINE FRACTALS
Substituting (7.42) and making the change of variable t' = rt, we obtain
From the definition of the power spectral density given in (7.40) we obtain
149
150
SELF-AFFINE FRACTALS
(2)
(3)
(4)
This transform maps N real numbers (the y n ) into N complex numbers (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.
The resulting Fourier coefficients Ynlare filtered using the relation
The power PI2 is used because the power spectral density is proportional to the amplitude squared. The amplitudes of the small-m coefficients correspond to short wavelengths Am and large wave numbers
km = 2n/Am. The large-m coefficients correspond to long wavelengths and small wave numbers.
An inverse discrete Fourier transform is taken of the filtered Fourier
coefficients. The sequence of points is given by
SELF-AFFINE FRACTALS
151
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 identical 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
152
SELF-AFFINE FRACTALS
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 shortwavelength 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). Consider 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
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 additions 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
i,
our example we take Ha = so that V2 = $.The five resulting random additions are given in Figure 7.9(c). After addition to the five values of y, in Figure 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,
and g7 as illustrated in
Figure 7.9(e). All nine points are now given random additions using a Gaussian probability distribution (3.15) with zero mean but a further reduced vari1
ance from (7.28) V3 =
Again taking Ha = f we have V3 = 4.
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 Figure 7.9(g). The process is repeated until the desired number of points is obtained. Our example with 4097 points is given in Figure 7.9(h). With Ha =
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 fractional 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.
i,i,
SELF-AFFINE FRACTALS
153
- .
0.0
(a)
-2
0.000
I
(f) 0.125
0.250
0.375
0.500
0.625
0.750
0.875
1.000
i,
154
SELF-AFFINE FRACTALS
SELF-AFFINE FRACTALS
155
Just as the fractional Gaussian noises generated using the filtering technique 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 fractional 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 values of p outside this range. Just as naturally occurring self-similar powerlaw distributions may or may not fall within the range of D values prescribed
by mathematical constraints, so too naturally occurring self-affine time series may or may not fall within the range of f3 values prescribed by mathematical constraints.
Using the definition of the semivariance y, given in (7.9), semivariograms for several of the fractional Gaussian noises and fractional Brownian
10
100
1000
10000
10
100
1000
10000
156
SELF-AFFINE FRACTALS
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 obtained 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 semivariograms 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 agreement 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.
SELF-AFFINE FRACTALS
157
The fractional Gaussian noises and fractional Brownian walks we have considered 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
128
256 t 384
512
128
256 t 384
512
(b) C , = 0.5
(c) c , = 1.0
(d) c , = 2.0
158
SELF-AFFINE FRACTALS
SELF-AFFINE FRACTALS
159
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
Q(tf)dt'
tQ
(7.53)
160
SELF-AFFINE FRACTALS
where Hu is known as the Hurst exponent. Examples included river discharges 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
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 series, 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 averaged 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
SELF-AFFINE FRACTALS
161
Since a white noise p = 0 is a random process it would be appropriate to conclude that Hu = 0.5 implies randomness. It follows that 0.5 < Hu < 1.0 implies 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
on log 7 (log N) for several of the fractional noises
dependence of log (WS),"
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.
162
SELF-AFFINE FRACTALS
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 antipersistence for fractional Gaussian noises. Extensive R/S analyses of fractional 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 obtained for various window lengths and the variances of these means were obtained as a function of window length. Gomes Da Silva and Turcotte (1994)
have applied the counterscaling technique to fractional noises and walks.
SELF-AFFINE FRACTALS
163
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; Berkson 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 elevation of topography is approximately a Brownian walk. Twenty-four examples of the dependence of the power spectral density on wave number for linear topographic profiles from three different parts of Oregon are given in
Figure 7.17. One-dimensional Fourier spectral analyses were obtained using
164
SELF-AFFINE FRACTALS
the periodogram method. Three different regions were considered with different geomorphic and tectonic settings. The Willamette lowland is dominated 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 spectral 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 approximation. 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
SELF-AFFINE FRACTALS
Two implications of this result will be discussed. The first is the comparison 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 reason 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 corresponding 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 measure 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 roughness of the topography. The mean intercepts for the latitudinal and longitudinal 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 electrical conductivity, and neutron activation. The measured quantities are obtained as a function of depth so that they are equivalent to a time series and
spectral techniques can be applied.
Fractal Dimension
Roughness
1.436
1SO7
5.948
6.354
1.499
1.485
6.549
6.830
1.492
1SO0
5.825
5.963
165
166
SELF-AFFINE FRACTALS
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 obtained 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 geophysics that yield power-law spectra. Brown and Scholz (1985) have carried
out spectral studies of natural rock surfaces. They generally find fractal behavior with a relatively large range of variability between 1 < D < 1.6. Similar studies have been carried out by Power and Tullis (1991, 1995) and by
Pyrak-Nolte et al. (1995). The observation that fracture surfaces are selfaffine fractals can be used to scale the fluid permeability associated with
fractures.
An interesting question is whether climate obeys fractal statistics (Nicolis and Nicolis, 1984). Fluigeman and Snow (1989) have shown that the spa-
SELF-AFFINE FRACTALS
tial distribution of oxygen isotope ratios in sea floor cores obey fractal spectral statistics. Since it is generally accepted that the isotope ratios are proportional 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 between measured data sets. Consider the bathymetry of the oceans. Bathymetry is typically measured from ships along linear tracks and must be interpolated 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 bathymetry. This method can also be used to interpolate airborne magnetic surveys.
Hewett (1986) has used fractal techniques to interpolate well-log porosity 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 welllog data to predict the ground water movement and pollutant dispersion adjacent to a waste disposal site.
In some cases a time series x(t) will have a well-defined correlation dimension (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 uncorrelated, the delay T can be small. This process is known as forming ntuples 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 embedding 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 caldera (Campi Flegrei, Italy) as a self-affine time series and were able to make
successful forward predictions. Nicholl et al. (1994) have applied this approach 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
167
168
SELF-AFFINE FRACTALS
where ko is the wave number and A, = Ilko is the wavelength over which data
are included in the expansion. With A, = 2naOwe have
SELF-AFFINE FRACTALS
169
170
SELF-AFFINE FRACTALS
lo8
10'
Sg
m2
cycles
km
lo6
1 o5
Figure 7.20. Power spectral
density of the earth's geoid
as a function of wave
number. The circles
represent a compilation of
the data of Reigber et al.
(1985). The solid line
represents (7.41) with P =
3.5 (D = 0.75).
1o4
103
1 o4
cycles
1o
-~
SELF-AFFINE FRACTALS
171
18
28
38
48
58
68
78
88
17
27
37
47
57
67
77
87
16
26
36
46
56
66
76
86
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
172
SELF-AFFINE FRACTALS
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 derivation of the relationship between P and Ha must be reexamined but
SELF-AFFINE FRACTALS
Table 7.2. Summary of mean fractal dimensions estimated by onedimensional and two-dimensional spectral analysis for the topography
of Oregon and for synthetic images
Average D
Data
Oregon topography
Synthetic topography
Two-dimensional
analysis
One-dimensional
analysis
173
174
SELF-AFFINE FRACTALS
At the time of going to press a color version of this figure was available for download from http:llwww.cambridge.org,978052 1567336
SELF-AFFINE FRACTALS
175
At the time of going to press a color version of this figure was available far download from http://www.cambridge.0'g/978052l567336
176
SELF-AFFINE FRACTALS
Fractal dimensions and roughness amplitudes are obtained using suhregions of 32 X 32 data points. Thus fractal dimensions and roughness amplitudes 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 welldefined 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 roughness amplitude for each subregion.
A 32 X 32 set of digitized elevations is chosen to form each subregion (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).
is assigned an equivalent radial wave number r
Each coefficient Hs,
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 leastsquares regression yields a fractal dmension D, using (7.72) and
-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
SELF-AFFINE FRACTALS
177
At the time of going to press a color version of this figure was available for download from hftp://ww.cambridgeeorg/s78O52I567336
178
SELF-AFFINE FRACTALS
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 studied the self-similar fractal behavior of topographic contours. They argue that
Thus for two-dimensional Brown topography with D2 = 2.5, they would obtain Dss = 1.25, in good agreement with many observations.
SELF-AFFINE FRACTALS
179
Problems
Problem 7.1. Consider the deterministic first-order self-affine fractal construction 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 construction? What is the corresponding fractal dimension?
Problem 7.2. Consider the deterministic first-order self-affine fractal construction 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.
(a) Extend the construction to second order. (b) How many boxes with
180
SELF-AFFINE FRACTALS
dimensions hd5 by rd5 are required to cover the first-order construction? What is the corresponding fractal dimension?
Problem 7.3. Consider the deterministic first-order self-affine fractal construction 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 construction? What is the corresponding fractal dimension?
Problem 7.4. Consider the deterministic first-order self-affine fractal construction 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.
(a) Extend the construction to second order. (b) How many boxes with
dimensions hd6 by rd6 are required to cover the first-order construction? What is the corresponding fractal dimension?
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 deviation, 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.
New York
Seattle
Phoenix
SELF-AFFINE FRACTALS
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, standard deviation, and coefficient of variation for these values. (b) Determine (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 deviation, 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
E,E,-,
--x
1"
&Zi
n .i Z.l
= uzand
-x,
1 "
n,,.
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 distribution 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"
+,
181
182
SELF-AFFINE FRACTALS
Chapter Eight
GEOMORPHOLOGY
184
GEOMORPHOLOGY
GEOMORPHOLOGY
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 dimension 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.
185
186
GEOMORPHOLOGY
GEOMORPHOLOGY
187
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 ratio R , and the angle of divergence 0. And these three quantities are independent 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 Figure 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 Figure 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 implication of D = 2, the dimension of a plane.
Comparing the drainage network in Figure 8.1 with the fractal trees illustrated in Figure 8.5 shows an important discrepancy. The drainage network 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 intersect third-order, fourth-order, and all higher-order streams.
188
GEOMORPHOLOGY
GEOMORPHOLOGY
189
This class of fractal trees can also be quantified in terms of branching ratios 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
This is now a two-parameter family of fractal trees. For the fractal tree illustrated 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
2
n-i
Rb
k=l
Ck-l
1=-+ax--
8 0 0 0
4 0 0
2 0
1
(a)
Rkb
189 0 0
6 3 0
2 1
22116 4
6 3 2
2 1
(b)
(c)
(dl
190
GEOMORPHOLOGY
And if n
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
Tk = ( R , - 1)R y
(8.16)
GEOMORPHOLOGY
191
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-order 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 determined 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 bifurcation 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 correlate well with the fractal relation (8.4) taking D = 1.85. The branching-ratio matrices for the two river basins are given in Figure 8.10. We now deterover i
mine values for Tkby averaging the values of
lo5 lo4 -
0 Kentucky River
x Powder River
lo3 N
lo2
10
8
v n
1 6'
10
r
lo2
I
lo3
192
GEOMORPHOLOGY
1
Tk =
n-k
(n - k ) i = l
Ti,i
+k
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
0 Kentucky River
x Powder River
DLA synthetic network
GEOMORPHOLOGY
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 agreement 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. Another 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 reaching 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 ascending magma and has considered its implications on magma composition.
The concepts of fractal trees also have a wide variety of other applications. 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 northeastern United States are given in Figure 8.12 (Hack, 1957). The applicable
fractal relation is
Rr
2
4
4.6
2
2
2.5
1
2
1.67
0
1
1.2
0
2
2.5
4.7
2.4
1.77
1.2
2.5
10
7
5.15
3.98
2.87
2.09
1.56
1.87
1.5
1.2
2.70
2.5
Order
Idealized tree Figure 8.6(a)
Idealized tree Figure 8.6(c)
Kentucky River basin
(Peckham, 1995)
Powder River basin
(Peckham, 1995)
DLAa cluster (Ossadnick, 1992)
DLA drainage network model
00
193
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 mainstream 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 address 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 evolution of landscapes is a problem that has fascinated natural scientists for
centuries. The forms of mature landscapes evolve through processes of erosion and deposition. An essential question is whether it is possible to develop 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 introduced by Shreve (1966, 1967) and Schreidegger (1967) in which drainage
networks were considered as infinite topologically random networks (i.e., no
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 number-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 little 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 followed 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 gradual expansion (accretion). Howard (1971) introduced an accretionary headward growth model, a site adjacent to the existing network was chosen randomly, and the network propagated to this site. Thus, all sites on the network
had an equal probability for growth.
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 firstorder 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 particles 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 phenomena, it is generally preferable to use the "radius of gyration" rather than
the radius of a circle (sphere) that covers the growing aggregate. The definition 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 relation 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 distance fifrom the seed particle. Thus from (8.21) we have
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)
"Killing Circle"
Random Path
"Launching Circle"
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 result of a DLA random growth. Two-dimensional surface exposures of fractures and joints are generally fractal with D = 1.7 in agreement with the values 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
Seed Cells
I I I
1 1 1 1
1 1 1
I I 1 I I I I
I I IWI IU-hI
I I I I I
Accreted Cells
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 selected site. The iteration of this basic procedure results in a branching network composed of linked drainage cells. This "self-avoiding" algorithm prevents 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 accretionary nature of network growth produces a headward evolving drainage
GEOMORPHOLOGY
201
202
GEOMORPHOLOGY
where
h = h, at t = 0 the solution is
where J is the transport coefficient and we consider only the one-dimensional problem h = h(x,t).With the conservation of mass relation
this gives
GEOMORPHOLOGY
i).
203
204
GEOMORPHOLOGV
b)
GEOMORPHOLOGY
256.0
51 2.0
1024.0
768.0
lattice site
-2.8
-2.4
-2.0
-1.6
log k
- 1.2
-0.8
-0.4
205
206
GEOMORPHOLOGY
12.0
15.0
18.0
timesteps ( ~ 1 0 0 0 0 0 )
-6.3
-6.0
-5.7
-5.4
-5.1
log f
-4.8
-4.5
-4.2
GEOMORPHOLOGY
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 Brownianwalk 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 reasonably 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 topography 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 simulations and real braided rivers have a fractal distribution of island sizes with
D = 1.
Stark (1991) used an invasion percolation technique, in which the growing network was superposed on a fixed random field (analogous to a substrate with variable erodibility). At each time step, the network propagated to
the adjacent site having the highest erodibility value over the entire perimeter. 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 distribution of drainage basin sizes. Nikora and Sapozhnikov (1993) have presented 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). Sornette et al. (1994) have presented a model for the tectonic generation of fractal topography utilizing statistical distributions of displacements on fault
arrays.
207
208
GEOMORPHOLOGY
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 constitute 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 allocate resources for its mitigation it is necessary to make flood-frequency hazard 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 extreme 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 circumstances 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 covers a relatively short time span and no general basis has been accepted for its
extrapolation. Quantitative estimates of peak discharges associated with paleofloods 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-frequency analyses. In the United States this is the peak discharge during a water 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
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-invariant 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 discharge 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 discharge and the ratio of the 1000-year peak discharge to the 100-year peak
discharge. In terms of H a and D we have
209
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. Results 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) relation 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
GEOMORPHOLOGY
211
ten benchmark stations. His results for the 2-parameter gamma (Ga), Gumbe1 (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)correlates best with the log Gumbel (LGu), whereas the other statistical techniques predict longer recurrence times for very serious floods. The fractal
and log Gumbel are essentially power-law correlations, whereas the others
are essentially exponential.
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 severity 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 (monsoonal) 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 correlations 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, indicating 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 little 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
Hu
1-1805
2-2 185
5-3310
6-3440
6-8005
7-2 165
8-1500
10-3275
1 1-0980
12-1570
Westfield
Oconee
Mississippi
Little Missouri
Elkhorn
Mora
Llano
Humboldt
Arroyo Seco
Wenatchee
(MA)
(GA)
(MN)
(wy)
(NE)
(NM)
0.513
0.540
0.470
0.520
0.540
0.630
0.719
0.616
0.870
0.310
1.39
1.46
1.53
1.48
1.46
1.37
1.28
1.38
1.13
1.69
3.26
3.47
2.95
3.31
3.47
4.27
5.24
4.13
7.40
2.04
0.67
0.72
0.72
0.72
0.67
0.73
0.70
0.66
0.68
0.72
(TX)
(NV)
(CAI
(WA)
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 exponent 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-
R
s
I
5
10
20
'$ yrs
I
50
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 series, 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 centered 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
GEOMORPHOLOGY
215
W ,a ) =
[I
Jrn
( 1
e x - 2
)dt
(8.35)
The mother Mexican hat wavelet is illustrated in Figure 8.28. For more details 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 sevenyear, 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
500
1000
1500
t (days)
2000
2 500
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 transform 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 annual 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 magnitudes 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 scalogram 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 periods associated with the regional climate.
t (hours)
GEOMORPHOLOGY
217
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 corresponding 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) Determine a and c as defined in (8.8). (e) Determine the asymptotic bifurcation 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
length-order ratio is Rr = 3. (a) Determine the bifurcation ratios for 1st to
2nd, 2nd to 3rd, and 3rd to 4th order branches. What are the corresponding fractal dimensions? (b) Write the branch-number matrix and
branch-ratio matrix for this tree. (c) Determine T, for k = 1,2, 3. (d) Determine a and c as defined in (8.8).
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 illustrated 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
length-order ratio is Rr = 3. (a) Determine the bifurcation ratios for 1st to
2nd, 2nd to 3rd, and 3rd to 4th order branches. What are the corresponding fractal dimensions? (b) Write the branch-number matrix and
branch-ratio matrix for this tree. (c) Determine T, fork = 1 , 2 , 3 . (d) Determine a and c as defined in (8.8). (e) Determine the asymptotic bifurcation 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 construction 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)?
Chapter Nine
DYNAMICAL
SYSTEMS
220
DYNAMICAL SYSTEMS
DYNAMICAL SYSTEMS
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,:
In the vicinity of the fixed point Z, = 0 it is appropriate to neglect the quadratic term in (9.9) with the result
where i,,is again assumed to be small but finite. As time evolves i , approaches 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 toward each other. These solutions are not chaotic.
To discuss further the behavior of nonlinear equations we consider the
van der Pol equation
221
222
DYNAMICAL SYSTEMS
with E the only parameter governing the behavior. In considering the solutions of this second-order nonlinear equation, it is standard practice to introduce the definition of velocity
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 parameter.
We first consider the solution of (9.17) when E = 0. In this case it becomes
DYNAMICAL SYSTEMS
Simple harmonic motion is a circle in the phase plane. The radius of the circle 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 result 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 damping term dominates and the amplitude increases. For large amplitudes the
positive cubic damping term dominates and the amplitude decreases. The result 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 parameter values produce limit cycles for other parameter values.
Before considering chaotic solutions we will present some further introductory material on singular points. Consider the pair of linear total differential equations
224
DYNAMICAL SYSTEMS
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.
DYNAMICAL SYSTEMS
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
= p sin 0
giving
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
225
226
DYNAMICAL SYSTEMS
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 stability of the two real roots by linearization. We substitute
Thus the fixed point x = pl/2 is stable: solutions as they evolve in time converge 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 Figure 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 solutions 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)
DYNAMICAL SYSTEMS
227
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
.
transition at p = 0
positive there are three fixed points x = 0, 2 ~ ' 1 2 The
from one to three solutions is known, for obvious reasons, as a pitchfork bifurcation (although the word trifurcation would be more appropriate). A stability 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-
no solutions
... _
----_
,
UnSlrble branch
unnUble branch
228
DYNAMICAL SYSTEMS
verge to the stable fixed point x = 0. For p > 0 all solutions for x > 0 converge 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 equation
(9.42)
(9.43)
= p sin
stable
periodic
orbit
P
-----
unstoble
branch
DYNAMICAL SYSTEMS
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
(1
(1
+ xo)e2' - 1 + xo is a solution if x = x, at t = 0.
+ 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). Deterdx
mine the solution in the E phase plane, where i = -.
dt
229
230
DYNAMICAL SYSTEMS
dx
Problem 9.7. Consider the equation -= x
dt
- x3
X:
xie2'
e2'
+ 1- xi
is a solution if x = x, at t = 0.
dx
-
dt
= sin x
[e'tan (31
-
if x = x, at t = 0.
IT 3IT
dt
ChapterTen
LOGISTIC M A P
10.1 Chaos
The concept of deterministic chaos is a major revolution in continuum mechanics (Berg6 et al., 1986). Its implications may turn out to be equivalent to
the impact of quantum mechanics on atomic and molecular physics. Solutions to problems in solid and fluid mechanics have generally been thought
to be deterministic. If initial and boundary conditions on a region are specified, 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 deterministic solutions to equations and deterministic solutions that are unstable to
infinitesimal disturbances. Chaotic solutions must also be treated statistically; 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 condition 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
= 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 solving (10.3) are
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. Taking 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
23 4
LOGISTIC M A P
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
LOGISTIC M A P
235
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 bifurcations that have occurred. Period-doubling flip bifurcations occur at a sequence 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 approximation as k becomes larger. This relation indicates a fractal-like, scaleinvariant 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 predict 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 chaotic 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
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
238
LOGISTIC M A P
LOGISTIC MAP
where dxn is the incremental difference after the nth iteration if dxo is the incremental difference in the initial value. If the Lyapunov exponent is negative, 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 incremental 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 )
A = lim
rn+
1 "
log,
m n=O
1H.1
239
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 bifurcation as shown.
Consider as a particular example the iteration for a = 4 given by (10.16).
For this case we find
(10.25)
(10.26)
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 difference 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 generalized to give
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 solutions converge.
We next consider the triangular or tent map defined by
242
LOGISTIC M A P
with 0 < a < 1 and 0 < x < 1 . This map can also be defined by
x n + , = 2a xn
for
< xn < 2
(10.33)
For 0 < a < $ the only fixed point is at x, = 0 and it is stable. For $ < a < 1
2a
;both fixed points are unstable.
there are fixed points at x, = 0, x, =
(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
LOGISTIC M A P
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 bifurcation 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 geophysics. 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 resembles 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 resemble 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,?
Problem 10.3. Determine x , , x,, x,, and x, for the logistic map (10.1) taking
a = 2 and xo = 0.2. What is that value of x,?
Problem 10.4. Determine x , , x,, x,, and x, for the logistic map (10.1) taking
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?
243
244
LOGISTIC M A P
Problem 10.9. Determine x,, x,, x,, x,, and x, for the logistic map (10.1) takinga = 4 and p = ( 2 ~ ) - 1 .
Problem 10.10. Determine x,, x,, x,, x,, and x, for the logistic map (10.1) takinga = 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 somewhat 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 interacts 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 analogous 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
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 loading 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 earthquake lasts only a few tens of seconds, whereas the interval between earthquakes 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 empirical velocity-weakening friction laws are in agreement with laboratory observations and also generate stick-slip behavior. Dynamic instabilities associated 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; Rundle 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 classic 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
SLIDER-BLOCK MODELS
247
We assume sliding starts at T = 0 with Y = 1 (1 1.4) and dY1d.r = 0. The applicable solution is
Y = - +
(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 conditions for the onset of sliding are the force balances
248
SLIDER-BLOCK MODELS
(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 single 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
SLIDER-BLOCK MODELS
249
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 begin to slide if (11.14) is satisfied. Together (11.13) and (11.14) define a failure 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 fourdimensional phase plane consisting of Y,, Y2, dY,ld~,and dY,ld~.For simplicity 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-
250
SLIDER-BLOCK MODELS
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-velocity driver. The increments in Y, and Y, are equal and the strain accumulation phase is represented by the diagonal line bc, which has unit slope. The
termination of this strain occurs when this line intercepts the failure envelope. 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 increments 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 strainaccumulation phase occurs when this line intercepts the failure envelope.
This occurs at point a, and the cycle repeats. The behavior of this symmetrical 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 failure envelope are cases in which both blocks are sliding simultaneously.
When a diagonal strain-accumulation line intercepts the upper failure envelope, block 2 begins to slide. Because P is relatively large, the failure force
SLIDER-BLOCK MODELS
251
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 failure 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 behavior 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 illustrated in Figures 11.6 and l l .7, the system may evolve to limit cycle behavior or chaotic behavior. A series of period-doubling pitchfork bifurcations 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 windows of chaotic behavior are clearly illustrated.
'I.
.4l::...l!:...
252
SLIDER-BLOCK MODELS
SLIDER-BLOCK MODELS
253
of differential equations are not required. Studies of this modified sliderblock 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-
254
SLIDER-BLOCK MODELS
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.
)
chaotic
Taking P = 1.05 and a = 0.81, Huang and Turcotte ( 1 9 9 0 ~ found
model behavior that strongly resembled the observed sequence of earthquakes 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 California. 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 simultaneous slip of the weaker and stronger blocks. The model simulation
suggested two alternative scenarios for a great southern California earthquake following a sequence of Parkfield earthquakes. In the first case a
Parkfield earthquake will transfer sufficient stress to trigger the great southern California earthquake; the Parkfield earthquake is thus essentially a foreshock for the great earthquake. In the second case a small additional strain
after a Parkfield earthquake will trigger an earthquake on the southern section and this will result in an additional displacement on the Parkfield section. 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 systems 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 deterministic sense. Only a probabilistic approach to the occurrence of earthquakes 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?
SLIDER-BLOCK MODELS
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.
255
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 deterministic 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 decreases 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
LORENZ EQUATIONS
The density difference Ap in the buoyancy term of the vertical force the
equation (12.3) is related to this temperature difference by
257
258
LORENZ EQUATIONS
The problem has been reduced to the solution of two partial differential equations for q!~and 8.
To better understand the roles of various terms, it is appropriate to introduce the nondimensional variables
where K = k/(pcp) is the thermal diffusivity. With these nondimensional variables 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 convection relative to the viscous forces that damp convection. The higher the
Rayleigh number the stronger the convection. The Prandtl number is the ratio of the momentum diffusion to the heat diffusion. It is instructive to estimate 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
LORENZ EQUATIONS
and find Ra = 8.6 X lo6 and Pr = 2.5 X 1023, both very large values. The behavior 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 boundary 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 convection 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 coordinate 3. Solutions that satisfy (12.17) are
t j=
0=
sin
275
(T)
sin
sin rrj
The flow is assumed to be periodic in the horizontal direction with a wavelength A; the nondimensional wavelength is h = hlh. The flow consists of
counter-rotating cells; each cellular flow has width A12, as illustrated in Fig-
259
260
LORENZ EQUATIONS
Ra,
27 7r4
4
= ---- =
657.5
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
LORENZ EQUATIONS
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 number there are two stable convective solutions corresponding to cellular rolls
rotating either clockwise or counterclockwise. This is identical to the pitchfork 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 solutions 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 retained only three terms of the form
$ = j q (4 + i 2 ) ~ ( rsin
)
7r3
4~~
= -(4
2 6
(T)
sin ~7
(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 stability analysis (12.18). However, the expansion of the temperature (12.24) includes an additional term that is not dependent on i .
It is necessary to derive differential equations for the time dependence of
This is done by substituting the expansions
the coefficients A(T), B(r), C(T).
(12.23) and (12.24) into the governing equations (12.14) and (12.15). Coefficients of the Fourier terms are equated to obtain the necessary equations.
261
262
LORENZ EQUATIONS
All nonlinear terms in the stream function equation (12.14) are neglected. When (12.23) and (12.24) are substituted into this equation the result is
cos (271~1%)
sin .rrj cos (271~1%
) cos T%= sin 2-
(12.29)
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 accurate solutions to the full equations when the Rayleigh number is slightly supercritical, 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-
LORENZ EQUATIONS
These solutions correspond to an infinite set of two-dimensional cells as illustrated 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 solutions 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
(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 negative 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
263
264
LORENZ EQUATIONS
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, =
8112, the critical value from (12.22), so that b = For these values the steadystate solution given by (12.23), (12.24), (12.33), and (12.34) becomes
i.
= 2 [24(r -
3 = E 4{
( r
I)]''~sin
(Z)
[y
-
I) sin 2?rj
sin 71.y
T -
r
(
1)
cOs (.Dli)
21/2
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 convection 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 efficiency 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
This relation is not in good agreement with experiment when r is significantly larger than unity. This is clearly due to the extreme nature of the truncation, which is expected to be valid only near the stability limit r = 1. Nevertheless, 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
266
LORENZ EQUATIONS
The essential feature of the solution of the Lorenz equations in this parameter 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-
LORENZ EQUATIONS
267
60
40
time
(c)
268
LOREN2 EQUATIONS
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 velocities on the boundaries at y = 0, 1.
ChapterThirteen
IS MANTLE
CONVECTION
CHAOTIC?
The Lorenz equations are a low-order expansion of the full equations applicable to thermal convection in a fluid layer heated from below. For the range
of parameters in which chaotic behavior is obtained, the low-order expansion 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 tectonics 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 (subduction 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 convection is chaotic. The earth's solid mantle behaves as a fluid on geological
time scales because of thermally activated creep. The discussion in the previous chapter considered only a constant viscosity. This is a poor approximation for the earth's mantle because the dependence of strain on stress is almost 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 properties 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 appropriate to assume that the mantle has an infinite Prandtl number. Because the
270
Prandtl number for the mantle is so large, the momentum terms on the lefthand 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
~ 0, solution A = B = C = 0
Again these equations have the steady-state, d / d =
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 conduction 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 convection is not chaotic.
To study this problem further Stewart and Turcotte (1989) considered a
higher-order expansion than the Lorenz equations. The variables and 8-are
and Om,n
expanded in double Fourier series in F and j. The coefficients
represent the terms sin (2rrrnilh) or cos (2-/h)
and sin ( m y ) in the ex- the left-hand
side of (12.14)
pansions. In the limit of infinite Prandtl number,
is zero, resulting in a linear equation between @m,n and Om,, that can be written
3
qm,
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 temthat can be written
perature coefficients,.,
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 12dimensional phase space. It is convenient to project the 12-dimensional trajectories 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-
,,,, ,,,,
,,,,
271
272
a,,, al,,
sweeps back to a saddle bifurcation at Ra = 1.909 X lo4.This type of bifurcation configuration (subcritical pitchfork plus two saddles) typically produces hysteresis effects when the saddle has one stable branch and one unstable 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 Figures 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 stable 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 Figure 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 subharmonk 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,,, appear. These are labeled "new 8,.,(pure)" in Figure 13.3(a). Each of these undergo 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 bifurcations, 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 bifurcations 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
Figure 13.4, the-first
- 7000 points of the trajectory at Ra = 4.5 X 104 are projected 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 trajectory 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.
273
274
a,,,
275
I-0
0
--
'-4000.0
- 2000.0
0.0
2000.0
4.1
(4
4000.0
276
a,,,
_
new
'
4
- --- '-'V
'
Q2, (mixed)
-'-- ' M e - - - - - - -
277
$ 1 (pure
-.
(a)
20
4.0
RAYLEIGH NUMBER (x 1P)
(b)
278
Problems
Problem 13.1. Show that the temperature coefficient B in the Lorenz expansion (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 expansion (12.24) is related to %o,2 by
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
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 distribution of polarity intervals. This question has been addressed by Gaffin
(1989) and his results are given in Figure 14.2. The number of polarity intervals 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 viscosity of the liquid outer core is sufficiently small that the flow is undoubtedly 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 subject to random reversals of the magnetic field, but it was much later (Cook
and Roberts, 1970) that it was demonstrated that the Rikitake dynamo behaved in a chaotic manner.
RlKlTAKE DYNAMO
281
where R is the resistance in either circuit and M is the mutual inductance between 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 rotating disk on the left in Figure 14.3.
282
RlKlTAKE DYNAMO
y2 = ( g ) 1 / 2 R 2 A
, = ( g ) 1 / 2 R o ,A
,
~'2
(---)
GLM
1/2
RlKlTAKE DYNAMO
283
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 singular points given above are unstable for all parameter values. Their numerical solutions for p. = 1 and K = 2 are given in Figures 14.4 and 14.5. The sin-
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 obtained 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.
RlKlTAKE DYNAMO
285
netic field was proposed by Robbins (1977). The governing equations take
the form
CHAOS
286
RlKlTAKE DYNAMO
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 between electrical currents and flows of the electrically conducting fluid. A
start in this direction has been given by Glatzmaier and Roberts (1995).
yo
1
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 inductance M, and applied torque G.
Problem 14.2. Determine the fixed points of the Chillingworth-Homes dynamo 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 considered 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 geophysics.
Thermodynamics represents the standard approach to collective phenomena. System variables are defined, that is, temperature, pressure, density, entropy; and the evolution of these variables is determined from the first
law of thermodynamics (conservation of energy) and the second law of thermodynamics (variation of entropy). Statistical mechanics provides the rational microscopic basis for much of thermodynamics.
In general, neither thermodynamics nor statistical mechanics yields fractal statistics or chaotic behavior. Exceptions include critical points and phase
changes. A characteristic feature of a phase change is a discontinuous (catastrophic) change of macroscopic parameters of the system under a continuous 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 treating 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-
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.
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 almost 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 determined whether each element is permeable or impermeable. With p, = 0.5 either the dark squares or the light squares can be taken to be permeable. In either case no continuous permeable path is found either horizontally or
vertically. Using the Monte Carlo approach a large number of random realizations would be carried out and the probability P(p,) that the array is permeable 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 defined to be the number of permeable elements in contact with each other
when the array first becomes permeable. The number of elements in the percolation cluster n: has been determined numerically as a function of the array size n. For two-dimensional arrays it is found that (Stauffer and Aharony,
1992)
This result can be compared with the deterministic Sierpinski carpet assuming that the remaining squares illustrated in Figure 2.3 represent a percolation 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
291
292
293
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 permeable 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 secondorder 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 consider the fixed points it is appropriate to rewrite (15.5) as
294
In the range 0 < x < 1 there are three fixed points x = 0,0.618, 1. The corresponding 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 probability 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 compares 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
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 between 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-dimensional percolation clusters.
The simplest renormalization group model for the array of cubic elements 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 permeable by
The critical value for the onset of permeability is p* = 0.282. This is in reasonably good agreement with the numerical results considering the simplicity 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. Although rocks with low porosity have essentially zero permeability, the sudden 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 occurring in natural systems, which is not fully described by the idealized model.
Problems associated with electrical conduction through a matrix of elements 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.
295
296
297
ities the probability p , that a first-order cell is fragile is related to the probability po that a first-order element is fragile by
After renormalization exactly the same statistics are applicable at higher orders. 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 consider the fixed points it is appropriate to rewrite (15.1 1) as
In the range 0 < x < 1 there are three fixed points x = 0,0.896, 1. The corresponding 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.
298
299
300
In writing (15.15) to (15.17) the transfer of the force (stress) when an element 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 unbroken 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 failures. The probabilities that the [ub] state will be broken or unbroken under
stress transfer are given by
From (15.14) and (15.17) the probability that a zero-order cell fails, p , , 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
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 second-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 Chapter 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
301
302
In the range 0 < x < 1 there are three fixed points x = 0,0.2063, 1. The corresponding 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 fractal tree fails for p, = 0.6.
The unstable fixed point at p* = 0.2063 is a critical point. For probabilities smaller than the critical value 0 < p, < p* the iteration is to the unbroken 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-
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 collection of wires under stress a will have failed when t,,, = (In 2)lv. It is often appropriate 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
303
304
uniformly to all the remaining elements; this is known as a global load-sharing 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 elements is given by
The rate at which elements fail is assumed to be given by the rate law
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 exponential 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 component. This result is fully self-similar.
305
306
The sequence of positive maxima and/or minima in a log-periodic sequence 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
Date
(a)
1970
1980
1990
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, obtained 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 result is given in Figure 15.9(a). The data appear to exhibit a series of log-periodic fluctuations. An important question is whether this type of strain accumulation 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 results are given in Figure 15.9(b). The predictions became increasingly accurate 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 hierarchical 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 cable. 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 elements; 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 statistical 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 elements fail, that is, a first-order element, the stress is transferred to the adjacent 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) example 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-
307
308
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
(ii)
(iii)
(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 element "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 local 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 below t, if stress is transferred to the element. The time Ti, is obtained by requiring 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 failure time so that
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
309
310
Upon the failure of "4," the stress a, is transferred to element "3," as illustrated in Figure 15.11(a) (iv). The time to failure of element "3" is given by
311
Alternative failure sequences are also possible, one example of which is illustrated 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) realization 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 using 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 expanded 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 scaleinvariant sequence of precursory failures at all levels. Because of the stochastic nature of the model, the embedding is not always clear, and a partic-
312
ular partial failure may be a part of a sequence or may be precursory to another 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 precursory failures are a ubiquitous feature of all realizations, there is considerable 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-
0.047636
1st estimate
0.048627
0.054775
2nd estimate
0.047799
0.047487
0.047799
0.047965
0.048154
3rd estimate
True
3
failure
time
313
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 .
,,
'
,.
, ,,
.........I
.... .
3 14
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 array; 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 tabulation (permeable elements, impermeable elements, alternative configurations, 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
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
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.
315
Chapter Sixteen
SELF-ORGANIZED
CRITICALITY
In the last chapter we considered the renormalization group method for treating large interactive systems. By assuming scale invariance a relatively
small system could be scaled upward to a large interactive system. The approach is often applicable to systems that have critical point phenomena. In
this chapter we consider the alternative approach to large interactive systems. 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 marginally 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 natural 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 repose. This is the maximum slope that a granular material can maintain without 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 simply slide down the slope. This is not what happens. The sand pile never
reaches the hypothetical critical state. As the critical state is approached additional 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 balances 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
(1)
(2)
(3)
(4)
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 frequency-size distribution of events. The size of a multiple event can be quantified in several ways. One measure is the number of boxes that become unstable 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 quasiequilibrium state. On average the number of particles lost from the edges of
the grid is equal to the number of particles added. Initially, small redistribution 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 statistics 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 criticality.
The behavior of a sand pile and the behavior of the cellular automata
model have remarkable similarities to the seismicity associated with an active 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 particles 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-
317
318
SELF-ORGANIZED CRITICALITY
ering the analogy further, we will describe the behavior of the cellular automata 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 particles 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 particles are redistributed to the four adjacent boxes, increasing the number 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
8
9
10
10
12
lla
SELF-ORGANIZED CRITICALITY
319
320
SELF-ORGANIZED CRITICALITY
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 redistributions has completed step 3. During step 3 all nine boxes were unstable and 12 particles were lost from the grid.
Step 4 A particle has been randomly added to box 5, increasing the number
of particles from zero to one.
Step 5 A particle has been randomly added to box 6, increasing the number
of particles from two to three.
:
C9
r(
2
V
bD
-e
y=2.99-1.03~
2
8
Incremental Distribution
Grid Size: 50x50
0.0
0.4
0.8
1.2
1.6
log(Area)
2.0
2.4
2.8
SELF-ORGANIZED CRITICALITY
4,
321
322
SELF-ORGANIZED CRITICALITY
-2.5
kL _
-2.0
I-
-1.5
_
1
L
-1.0
-0.5
0.0
0.5
Log layer thickness h (m)
_.
(b)
.
J
1.0
';4
J
SELF-ORGANIZED CRITICALITY
323
foreshock) may trigger an instability in a large box (the main shock). A redistribution 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 fractal 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
N, = 108 boxes with r, =
Except for N, the N are related to the ri by the
fractal relation (2. I) with D = In 3lln 2 = 1.5850.
2.
i,
i,
324
SELF-ORGANIZED CRITICALITY
(2)
(3)
(4)
Box Size r
SELF-ORGANIZED CRITICALITY
325
326
SELF-ORGANIZED CRITICALITY
(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 Figure 16.7. In the cellular-automata approximation it is assumed that during
the sliding of one block, all other blocks are stationary; this requirement limits the system to nearest neighbor interactions, which is characteristic of cellular-automata systems. To minimize the complexity we considered a discontinuous 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 stiffness 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 correlation is obtained with the fractal relation (2.6) with D = 2.72. The frequency-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 failure statistics of these multiple-block systems clearly indicate a self-organized critical behavior and are remarkably similar to distributed seismicity.
SELF-ORGANIZED CRITICALITY
327
0.0
0.5
1.0
1.5
2.0
log(N~)
2.5
3.0
3.5
328
SELF-ORGANIZED CRITICALITY
SELF-ORGANIZED CRITICALITY
329
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 region on December 7, 1988. TIPS were issued for region 5 in California during 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 behavior 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 questioned. Nevertheless, seismic activation prior to a major earthquake certainly appears to be one of the most promising approaches to earthquake prediction.
I . THE CAUCASUS. M 2 6.5
44
42
40
38
b. CALIFORNIA-NEVADA,M 2 7
50
1975
1
45
1980
2
40
3
'4
35
6-
30
25
-135 -130
-125
-120
-115 -110
1985
330
SELF-ORGANIZED CRITICALITY
SELF-ORGANIZED CRITICALITY
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)
IF;, ,In+
IF, ,In+,
IF;,
(2)
IF;, ,In+
IF;, , I n + , -
,+
jln+
331
332
SELF-ORGANIZED CRITICALITY
The slider blocks are considered sequentially using a checker-board algorithm to sweep across the two-dimensional array.
It is convenient to introduce the nondimensional variables Ti, = Fi,,IFs.
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
zk
If
IFi]- > 1, block (i, j) is unstable and its nondimensional slip is given by
&.
SELF-ORGANIZED CRITICALITY
However, the slip condition for a block is determined by the statistical distribution 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 distribution 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
333
334
SELF-ORGANIZED CRITICALITY
SELF-ORGANIZED CRITICALITY
335
p=0.213
site percolatm
336
SELF-ORGANIZED CRITICALITY
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 statistical 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 California. The distributions given in Figure 4.12 appear to correspond to the fractal
dimension of the percolation "backbone" of a critical three-dimensional percolation cluster. It appears that the earthquakes on a complex array of faults
form a connected path across the zone of crustal deformation in direct analogy 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.
&,
SELF-ORGANIZED CRITICALITY
337
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 examples 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 obtained 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
we
model in Figure 16.8. When the sparking frequency f is reduced to
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.
&,
X
X
X
X
X
X
X
X
338
SELF-ORGANIZED CRITICALITY
lo'
N
io3
1'
0
A.
10'
10
+-I
1o0
*-.
Best fit line)
,&lope =--0.59)
- -
0.1
0.01
0.01
0.1
A,, km2
10
100
loo0
SELF-ORGANIZED CRITICALITY
Problems
Problem 16.1. Consider the evolution of the cellular-automata model illustrated 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 particles are lost from the grid in steps 8, I la, I lb, I lc, and 1Id?
Problem 16.2. Consider the evolution of the cellular-automata model illustrated in Figure 16.l(b). (a) Which boxes have an additional particle in
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?
Problem 16.3. Consider the evolution of the cellular-automata model illustrated in Figure 16.l(b). (a) Which boxes have an additional particle in
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.
339
340
SELF-ORGANIZED CRITICALITY
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 sequence 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?
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 systems. In this regard slider-block models play a central role. Slider-block
models were introduced as simple analog models for earthquakes. Distributed seismicity is taken to be a type example of self-organized critical behavior. 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 increasing 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 between 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 defies 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
344
REFERENCES
REFERENCES
Biegel, R. L., Sammis, C. G. & Dieterich, J. H. (1989). The frictional properties 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 fractures and the possible physical generating mechanism, Pure Appl. Geophys., 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 Hydrology, 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 topography 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). Quantifying 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.
345
346
REFERENCES
REFERENCES
347
348
REFERENCES
Curran, D. R., Shockey, D. A., Seaman, L. & Austin, M. (1977). Mechanisms 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 dissolution, 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 proposed 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 mechanical 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 distribution of assay values, Geol. Mijnb. 13,365-75.
De Wijs, H. J. (1953). Statistics of ore distribution. Part 11: Theory of binomial 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. Friedman, J. M. Logan & D. W. Steans, eds., pp. 103-20, American Geophysical 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 multifractal 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-dimensional self-organized critical forest-fire model, Phys. Rev. Lett. 71,
3739-42.
REFERENCES
349
350
REFERENCES
REFERENCES
Gardner, T. W., Jorgensen, D. W., Shuman, C. & Lemieux, C. R. (1987). Geomorphic 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 dimension 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). Measurement and characterisation of spatial distributions of fractures, Tectonophys., 226, 113-4 1.
Glatzmaier, G. A. & Roberts, P. H. (1995). A three-dimensional self-consistent 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 measures, Math. Geol. 12,85-98.
Grady, D. E. & Kipp, M. E. (1987). Dynamic rock fragmentation, in Fracture Mechanics of Rock, B. K . Atkinson, ed., pp. 429-75, Academic
Press, London.
Grassberger, P. & Procaccia, I. (1983a). Characterization of strange attractors, 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 stability of a single degree of freedom elastic system with rate and state dependent friction, J. Phys. Solids 32, 167-96.
351
352
REFERENCES
REFERENCES
353
354
REFERENCES
Huang, J., Narkounskaia, G., & Turcotte, D. L. (1992). A cellular automaton, slider-block model for earthquakes. 2. Demonstration of selforganized 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 synthetic images, J. Geophys. Res. 94,7491-5.
Huang, J. & Turcotte, D. L. (1990a). Are earthquakes an example of deterministic 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 massspring model and velocity-weakening friction, Pure Appl. Geophys. 138,
569-89.
Hurst, H. E., Black, R. P. & Simaika, Y. M. (1965). Long-term Storage, Constable, 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 multifractal 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 phenomena, J. Geophys. Res. 95,6853-68.
Ivanov, S. S. (1994a). Global relief: Evidence of fractal geometry, in Fractals 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
Johnston, A . C. & Nava, S. J. (1985). Recurrence rates and probability estimates 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 universality 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 earthquakes, 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: Algorithm 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 lithosphere 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 termination 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. Geophys. Res. 93, 13307-1 8.
355
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 modelling 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 carbonates, J. Geophys. Res. 93,3297-305.
Krohn, C. E. (1988b). Sandstone fractal and Euclidean pore volume distributions, J. Geophys. Res. 93,3286-96.
Krohn, C. E. & Thompson, A. H. (1986). Fractal sandstone pores: Automated measurements using scanning-electron microscope images, Phys.
Rev. B33,6366-74.
Kucinskas, A. B. & Turcotte, D. L. (1994). Isostatic compensation of equatorial highlands on Venus, Icarus, 112, 104-16.
Kucinskas, A. B., Turcotte, D. L., Huang, J. & Ford, P. G. (1992). Fractal analysis 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 earthquake fault, Phys. Rev. Lett. 67, 1043-6.
La Pointe, P. R. (1988). A method to characterize fracture density and connectivity 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 reserves, 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
Ledesen, B., Dubois, J., Genter, A. & Meunier, A. (1993). Fractal analysis of
fractures applied to Soultz-sous-Forets hot dry rock geothermal program, 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 networks, Phys. Rev. Lett. 71, 1470-3.
Leopold, L. B. & Langbein, W. B. (1962). The concept of entropy in landscape evolution, U.S. Geological Survey Paper 500A, 20 pp.
Lifton, N. A. & Chase, C. G. (1992). Tectonic, climatic and lithologic influences on landscape fractal dimension and hypsometry: Implications for
landscape evolution in the San Gabriel Mountains, California, Geomorph. 5,77-114.
Lin, B. & Taylor, P. L. (1994). Model of spatiotemporal dynamics of stickslip motion, Phys. Rev. E49,3940-7.
Lin, B . & Yang, Z. R. (1986). A suggested lacunarity expression for Sierpinski 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 detecting 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 springblock 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.
357
35 8
REFERENCES
REFERENCES
Maxwell, J. C. (1960). Quantitative geomorphology of the San Dimas experimental forest, California, Project NR 389-042, Columbia Univ. Dep.
Geol. Tech. Rep. 19.
May, R. M. (1976). Simple mathematical models with very complicated dynamics, 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 earthquakes 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,
Springer-Verlag, Berlin.
Molchan, G. M., Dmitrieva, 0 . E., Rotwain, I. M. & Dewey, J. (1990). Statistical 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, Physica 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.
359
360
REFERENCES
Nagatani, T. (1993). Dynamic scaling of river-size distribution in the Scheidegger'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 deterministic 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-organized 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 Emmet County, Michigan, USA, Earth Sugace Proc. Landforms 5,33145.
Newman, W. I. & Gabrielov, A. M. (1991). Failure of hierarchical distributions 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 material failure. Statics and dynamics, Physica D77,200-16.
Newman, W. I. & Turcotte, D. L. (1990). Cascade model for fluvial geomorphology, Geophys. J. Int. 100,433-9.
Newman, W. I., Turcotte, D. L. & Gabrielov, A. M. (1995). Log-periodic behavior 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 midplate earthquakes, Seis. Soc. Am. Bull. 73,5 19-35.
REFERENCES
361
362
REFERENCES
REFERENCES
363
364
REFERENCES
Sammis, C., King, G. & Biegel, R. (1987). The kinematics of gouge deformation, 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
Appl. Geophys. 123,53-78.
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 characterization 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., SpringerVerlag, 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 wavelet transform, Optical Eng. 31(11), 2492-5.
Schmittbuhl, J., Vilotte, J. P. & Roux, S. (1993). Propagative macrodislocation modes in an earthquake fault model, Europhys. Lett. 21,375-80.
Schoeny, P. A. & Saunders, J. A. (1993). Natural gold dendrites from hydrothermal 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 phenomena 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 quiescence, Phys. Earth Planet. Int. 61, 113-27.
Schuster, H. G. (1995). Deterministic Chaos, 3rd ed., VCH, Weinheim, 291 pp.
REFERENCES
Schwarzacher, W. (1976). Stratigraphic implications of random sedimentation, 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 activity 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 earthquakes and frequency of occurrence of large earthquakes along the Mexican 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. 90,18841900.
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 characterization and feature detection using a discrete wavelet transform, submitted for publication.
365
366
REFERENCES
REFERENCES
Stinchcombe, R. B. & Watson, B. P. (1976). Renormalization group approach 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-organized 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 threshold elements, Phys. Lett. A131,244-7.
Tarboton, D. G . , Bras, R. L. & Rodriguez-Iturbe, I. (1988). The fractal nature 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 networks and their evolution, Geographical Rep. Tokyo Metro. Univ. 13,
1-27.
Tokunaga, E. (1984). Ordering of divide segments and law of divide segment numbers, Jap. Geomorph. Un. 5 , 7 1-7.
367
368
REFERENCES
REFERENCES
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 Systems, 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 parameters 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 displacements and dimensions of faults, J. Struct. Geol. 10,329-47.
Walsh, J., Watterson, J. & Yielding, G. (1991). The importance of smallscale 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 aggregation 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 expansion, 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.
369
370
REFERENCES
Appendix A
GLOSSARY
OFTERMS
372
A P P E N D I X A: GLOSSARY OF T E R M S
A P P E N D I X A: GLOSSARY OF T E R M S
373
Appendix B
UNITS AND
SYMBOLS
Table B 1. SI units
Quantity
Unit
Basic units
Length
Time
Mass
Temperature
Electric current
meter
second
kilogram
kelvin
ampere
Derived units
Force
Energy
Power
Pressure
Frequency
Charge
Electric potential
Magnetic field
newton
joule
watt
pascal
hertz
coulomb
volt
tesla
Multiples of 10
lo-'
10-6
lo-'
10-l2
I 0'
1O6
I o9
10l2
milli
micro
nano
pic0
kilo
mega
giga
tera
Symbol
m
s
kg
K
A
N
J
W
Pa
Hz
C
V
Equivalent
APPENDIX
B: U N I T S A N D S Y M B O L S
Quantity
Equation
introduced
Power
Parameter
Radius of earth
Frequency of earthquakes
Area
Lorenz variable
b-value for earthquakes
Lorenz variable
Benioff strain
Constant
Specific heat at constant pressure
Coefficient of variation
Constant
Concentration
Pair correlation distribution
Lorenz variable
Moment of inertia
Constant
Euclidean dimension
Fractal dimension
Energy
Probability
Probability distribution function
Probability of fragmentation
Fraction
Frequency
Cumulative distribution function
Flood frequency factor
Feigenbaum constant
Force
Wavelet filter
Acceleration due to gravity
Applied torque
Elevation
Layer thickness
Hausdorff measure
Hurst exponent
Electrical current
Transport coefficient
Boltzmann constant
Constant
Wave number
Spring constant
Thermal conductivity
Partition coefficient
Rikitake parameter
Length
Lacunarity
Self-inductance
(3.40)
(10.1)
(7.63)
(4.1)
(3.67)
(12.23)
(4.1)
(12.24)
(15.38)
(4.4)
(12.4)
(3.33)
(2.1)
(5.1)
(6.13)
(12.24)
(14.3)
(4.4)
(6.12)
(2.1)
(4.2)
(3.5)
(3.12)
(3.70)
(6.22)
(7.41)
(3.10)
(8.31)
(10.11)
(11.1)
(8.32)
(12.3)
(14.2)
(7.28)
(12.5)
(7.1)
(7.56)
(14.1)
(8.25)
(3.16)
(3.40)
(7.66)
(9.12)
(12.4)
(5.29)
(14.13)
(2.9)
(6.20)
(14.3)
SI units
m
s-I
m2
Jl/?
J kg-' K
kg m2
-- I
N
ms- ?
Nm
m
m
A
m2s-I
JK-'
m-I
Nm-I
Wm-I K-'
m
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
Quantity
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
Equation
introduced
SI units
m
(continued )
APPENDIX B: U N I T S A N D SYMBOLS
Symbol
Quantity
Variable
Vertical coordinate
Fourier transform
Nondimensional position
Constant
Mass ratio
Stiffness parameter
Volume coefficient of thermal
expansion
Power
Constant
Symmetry parameter
Constant
Coefficient of skew
Semivariance
Slope
Displacement across fault
Strain
Small quantity
Parameter
Latitude
Polar coordinate
Temperature difference
Thermal diffusivity
Wavelength
Lyapunovexponent
Shear modulus
Viscosity
Rikitake parameter
Power
Hazard rate
Density
Polar coordinate
Standard deviation
Stress
Time interval
Nondimensional time
Porosity
Enrichment factor
Longitude
Friction parameter
Stream function
Angular velocity
Y
a
B
Y
IS
E
CL
v
P
u
T
dJ
Equation
introduced
(3.18)
(2.7)
(7.38)
(11.3)
(4.5)
(5.13)
(11.12)
(12.8)
(7.41)
(9.12)
(11.12)
(4.6)
(3.4)
(7.8)
(10.6)
(4.3)
(4.27)
(5.39)
(9.13)
(7.63)
(9.28)
(12.7)
(12.11)
(7.66)
(10.19)
(4.3)
(12.2)
(14.8)
(3.48)
(15.30)
(3.82)
(9.28)
(3.3)
(15.44)
(4.23)
(11.3)
(3.81)
(5.1)
(7.63)
(1 1.3)
(12.6)
(14.1)
SI units
m
K-'
K
m2 s-I
m
Pa
Pa s
t-I
kg m-'
Pa
s
m2s-I
s-I
377
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.
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.
379
380
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 .
381
382
3i.
383
INDEX
386
INDEX
chaos, 3,219,223,225,231,236,237,238,240,
245,25 1,253,256,265,266,269,
272,289,341,37 1,372
deterministic, 219. 223, 23 1 , 245, 256, 266
route to, 237
windows of, 236,237,238,240,25 1
chaotic mantle convection, 269
characteristic earthquake, 67, 322. 325
chromatographic model, 87.95
circulations, hydrothermal, 81
climate, 166, 341
clustering, 100, 103, 104,328, 371, 372
fractal, 100, 103, 104, 328
lacunarity of, 109,372
clustering of faults, 104
clustering of joints, 104
clustering of metamorphic veins, 104
clustering of seismicity, 103, 328
clusters, 4,78, 100, 103, 104,290, 291, 372
percolation, 4.78.290, 291,372
backbone of, 78,291,336
number-size statistics of, 291
coal, 43.44
coal mines, 67
coastline, 1 , 12, 15, 132
fractal dimension of, 1, 12, 15, 132
length of, 1, 12
rocky, 1, 12, 15, 132
roughness of, 15
coefficient of skew, 29,30,35
coefficient of skew for a log-normal distribution,
35
coefficient of thermal expansion, 258
coefficient of variation, 35, 137, 158
coefficient of variation for a log-normal distribution, 35, 158
comminution, 44,48,49,50,67,71,74, 298
comminution model, 48,49,50,71,74,298
complementary error function, 33
complexity, 342
component, periodic, 136
stochastic, 136, 137
trend, 136
concentration, mineral, 8 1.90, 120, 136
ore, 8 1.90, 120, 136
conditional probability, 301
conduction, electrical, 295
heat, 257
conduction solution. 261
conductivity, electrical, 165
thermal, 257
conservation of energy, 223,289
conservation of mass, 202
contact areas, multifractal analysis of, 129
continental drift. 279
continental margin, 321
continuity equation, 256
continuous data, 29.30
continuous processes, 56
continuous time series, 136, 137
contour, topographic, 12, 132
fractal dimensions of, 12
INDEX
387
388
INDEX
INDEX
303
age of, 203
clustering of, 104
distribution of, 1.67,68,71, 199.322
fractal distribution of, 1.67.71.322
interacting, 247
length of, 67
normal, 76
number-length statistics, 67, 322
pre-existing. 245
transform, 56.70
Feigenbaum constant, 236,371
Feigenbaum relation, 236
fields, electric, 4
gravitational,4
induced electric, 28 1
magnetic, 4,217,279,342
filling, space, 236
filtering, 214
Fourier, 152
fingering, viscous, 198
first law of thermodynamics, 289
fixed point, 220,221,223,224,225,226,227,
232,242,265,270,294,297,302,372
stability of, 220, 221
stable, 220,225,232,265,294,297,302
unstable, 232,265,294,297,302
flexure, 13
flip bifurcation, 234,236,238,240
flood discharge, 136,208
flood frequency factor, 209
flood hazard, 208
floods, 158,208,209
annual, 208,209
frequency of, 208
partial duration series of. 209
fluid flow, 50
fluid layer. 256,268,269
fluid turbulence, 3, 127,231,268,341
folds, 1, 56
force balance equations, 256
forces, buoyancy, 257,265
inertial, 257
pressure, 257
viscous, 257
foreshocks. 246.323.324
forest-fire models, 336, 337, 339
forest fires, 291.339
distribution of, 337,339
fossil magnetism, 279
Fourier coefficients, 2, 150
Fourier filtering technique, 152
Fourier series, 2. 148
amplitudes of. 2, 148
phases of, 148
Fourier spectrum, 2, 148, 163
of bathmetry, 2, 163
of topography, 2, 163
Fourier transform, 148, 150, 171.214
discrete, 150, 171
inverse, 148, 150
two-dimensional, 17 1, 172
390
INDEX
INDEX
392
INDEX
INDEX
207
mining induced seismicity, 129
model, advective-diffusion,207
avalanche, 5,207,316,317
cellular-automata. 5,316,317,320,322,324,
325,326
comminution, 48,49,50,71,74,298
Culling, 202,207
diffusion limited aggregation, 4,74, 195, 197,
199,201,207
forest fire, 336,337,339
sandpile, 5,316,317
slider-block, 4,245,247,254,325,330,336,
342
stochastic, 4
molecular velocities, 30, 3 1
moment magnitude, 58,60
moment of inertia, 282
moments, 29, 110, 114,372
earthquake, 57,58,59,60,61,305
generalized, 114
seismic, 60, 305
Monte Carlo, 291
mother wavelet, 214
motion, equation of, 222,246,325
moving-average model, 140
autocorrelation function for, 141
autoregressive, 143
autoregressive integrated, 145
variance of, 141
moving-box method, 111
moving-window method, 109
multifractal, 83, 113,372
perfect, 124, 128
multifractal analysis, 120, 125, 129
multifractal analysis of fragmentation, 129
multifractal analysis of well logs, 129
multifractal dimension, 115, 117, 124
multifractal distribution of contact areas, 129
multifractal distribution of earthquakes, 128
multifractal distribution of fractures, 129
multifractal distribution of mining induced
seismicity, 129
multifractal distribution of pore spaces, 129
multifractal distribution of void spaces, 129
multifractal scaling, 128
multifractal spectrum, 115
multifractal time series, 129
multifractal topography, 129
multiplicative cascade, 120, 121, 124, 126, 128
mutual inductance, 281
394
INDEX
INDEX
INDEX
spiral, 225
logarithmic, 225
sponge, Menger, 10,50,105,295
spreading centers, 56
spring-mass oscillator, 222
stabilities, exchange of, 226
stability analysis, 220, 221, 227.259.261
linearized, 220,259
marginal, 259, 3 16
stable fixed point, 220,225,232,265,294,297,
302
stable node, 224
staircase, devil's, 18,21,23,24
standard deviation, 29.30
Brownian walk, 140
exponential distribution, 40
Gaussian distribution, 32
log-normal distribution, 35
normal distribution, 32
Pareto distribution, 38
standard form, 33,37,38
state, critical, 5,316
static friction, 246, 326, 330
stationarity, 138
statistical fractal, 12, 38, 289
statistical mechanics, 289,330,342
statistical self-affine fractal, 140
statistics, 28
stick-slip behavior, 245, 246, 249
stiffness parameter, 248,326
Stirling approximation, 127
stochastic, 136, 137, 373
stochastic component, 136, 137
stochastic models, 4
storage, reservoir, 158
strain, 76,305
Benioff, 305
stranded cable, 299,303,307
strange attractor. 265,284,373
stratigraphichiatuses, 18, 19,20, 167
stream function, 257, 260
stream order. 181
stream flow time series, 215
stress, shear, 260
strong persistence, 137
subcritical pitchfork bifurcation, 228,272
subduction zones, 13.56.70.269
subsidence, 76
successive random additions, 152
sunspot numbers. 160
supercritical pitchfork bifurcation, 226
surface, free, 260
ocean, 146
surface wave magnitude, 57,58
surfaces, rock, 166
symmetry. 248
synthetic landscapes, 2, 173
Tangshan earthquake, 330
tectonic fragmentation, 44, 70
tectonic processes, 13.56, 181
tectonic uplift. 23
INDEX
tectonics, 13,50, 56
plate, 5.56.60, 70,256, 269, 279
temperature, 31, 136, 146, 160,289
atmospheric, 136, 146, 160
Curie, 279
tent map, 24 1
tephra, 79
textural analysis, 3, 112
landscape, 112
thermal boundary layer, 265,269
thermal conductivity, 257
thermal convection, 3,256,261,268,269
thermal diffusivity, 258
thermal expansion, 256,258
coefficient of, 258
thermally activated creep, 269
thermodynamics, 289
first law of, 289
second law of, 289
thickness statistics, 321
thinning, crustal, 76
threshold, percolation, 333
tiling, self-affine, 207
time series, 136, 137, 140, 146, 148, 158, 167,
214,373
atmospheric temperature, 136, 146
box-counting method for, 147
Brownian walk. 140
continuous, 136, 137
correlation dimension of, 167
correlations of, 158
discontinuous, 137
discrete, 136
fractal dimension of, 147, 148
multifractal dimension of, 129
river flow, 158,215
self-affine, 145
stream flow, 158,215
time to failure, 303,304,305,306
Tokunaga fractal tree, 188, 190, 198
tonnage, ore, 8 1
tonnage-grade, 8 1,90,92
copper, 93
gold, 92
mercury, 90
uranium, 94
topographic contour, 12, 132
topography, 2, 12, 13,56, 132, 145, 163, 165,
168, 170, 175, 177,207
Arizona, 177
Brownian walk, 163,207,208
elevation of, 132, 136
Fourier spectral analysis of, 163
fractal, 132, 208
height of, 132, 136
multifractal analysis of, 129
Oregon, 163, 165, 175, 177
power spectral density of, 168, 170
roughness of, 2, 165. 176
self-affine, 145, 178
Venus, 170
torque, 28 1
trace elements, 87
trajectory, phase, 222
transcritical bifurcation, 226
transform, discrete Fourier, 150, 171
Fourier, 148, 150, 171,214
inverse discrete Fourier, 150
inverse Fourier, 148
two-dimensional Fourier, 171, 172
wavelet, 214
transform faults. 56, 70
Transverse Ranges, 65
traps, petroleum, 98
trees, fractal, 2, 181, 187, 188, 299, 302
binary, 188
branch numbers for, 188
branching ratios for, 189
deterministic, 187
length-order ratio of, 185
side branching, 188
Tokunaga, 188, 190,198
tree rings, 160
trench, oceanic, 13, 56, 70, 269
trench pull. 65
trend component, 136
triadic Koch island, 11.24
triangular map, 241
tributaries, 185
truncations, 262,268, 270
tuples, 167
turbidite deposits, 321
turbulence, 3, 127, 23 1,268, 34 1
turning-point bifurcation, 226
two-dimensional Fourier transform, 171, 172
unconformities, 18, 19, 20.2 1, 167
unstable. gravitationally, 256
unstable fixed point, 232,265,294, 294,302
unstable node, 224
uplift, 23
uranium, 94.95
van der Pol equation, 221,223,229
variables, nondimensional, 219,222,246,248,
258,282,332
variance, 29.30, 137, 140, 141, 145
autoregressive model. 141
autoregressive moving average model, 145
Brownian walk, 140
exponential, 40
Gaussian, 32
log-normal. 35
moving average model, 141
normal, 32
Pareto, 38
self-affine fractal, 146
two-dimensional spectra, 168
variation, coefficient of, 35, 137, 158
varves, 160
veins, 70, 104, 193
clustering of, 104
velocities, molecular, 30.3 1
velocity, angular, 281
397
398
INDEX
wave equation, 4. 23 1
wavelength, 2
wavelet transform, 214, 373
bathymetric profile. 2 17
fractional Brownian walk, 2 17
geiod, 2 17
length of day, 217
seismogram, 2 17
wavelets, 214, 2 17, 373
Mexican hat. 214
mother, 2 14
weak persistence, 137
weather, 268, 341
weathering processes, 44
Weibull distribution, 41,42,299,301,303
Weirstrass-Mandelbrot functions, 152, 165
well logs, 3, 136, 165, 166, 168
acoustic, 165
electrical conductivity, 165
multifractal analysis of, 129
neutron activation, 165
porosity, 136, 166, 167
white noise, 138, 139, 140, 146, 148, 149, 150,
156,207,373
Gaussian, 140, 149, 150, 160, 207
Whittier earthquake, 63
window method, moving, 109
windows of chaos, 236,237,238,240.25 1
Zipf's law, 26
UNIVERSITY PRESS
wwwcambridge.org
a.,
'
Ill"