P1: GLQ Final Qu: 00, 00, 00, 00

Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals
Benoit B. Mandelbrot
Michael Frame
Yale University

I. Scale Invariance VI. Lacunarity

II. The Generic Notion of Fractal Dimension and a VII. Fractal Graphs and Self-Affinity
Few Specific Implementations VIII. Fractal Attractors and Repellers of Dynamical
III. Algebra of Dimensions and Latent Dimensions Systems
IV. Methods of Computing Dimension in IX. Fractals and Differential or Partial Differential
Mathematical Fractals Equations
V. Methods of Measuring Dimension in Physical X. Fractals in the Arts and in Teaching
Systems

GLOSSARY ematics had two effects. It led to the development of tools

like fractal dimensions, but marked a turn toward abstrac-
Dimension An exponent characterizing how some tion that contributed to a deep and long divide between
aspectmass, number of boxes in a covering, etc.of mathematics and physics. Quite independently from fun-
an object scales with the size of the object. damental mathematical physics as presently defined, frac-
Lacunarity A measure of the distribution of hole sizes tal geometry arose in equal parts from an awareness of
of a fractal. The prefactor in the massradius scaling is past mathematics and a concern for practical, mundane
one such measure. questions long left aside for lack of proper tools.
Self-affine fractal A shape consisting of smaller copies The mathematical input ran along the lines described
of itself, all scaled by affinities, linear transformations by John von Neumann: A large part of mathematics
with different contraction ratios in different directions. which became useful developed with absolutely no de-
Self-similar fractal A shape consisting of smaller copies sire to be useful . . . This is true for all science. Successes
of itself, all scaled by similitudes, linear transforma- were largely due to . . . relying on . . . intellectual elegance.
tions with the same contraction ratios in every direction. It was by following this rule that one actually got ahead
in the long run, much better than any strictly utilitarian
course would have permitted . . . The principle of laissez-
FRACTALS have a long history: after they became the faire has led to strange and wonderful results.
object of intensive study in 1975, it became clear that they The influence of mundane questions grew to take on
had been used worldwide for millenia as decorative pat- far more importance than was originally expected, and re-
terns. About a century ago, their appearance in pure math- cently revealed itself as illustrating a theme that is common

185
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

186 Fractals

in science. Every science started as a way to organize a Euclidean or locally Euclidean, and observed nature is
large collection of messages our brain receives from our written in the language of noisy Euclidean geometry.
senses. The difficulty is that most of these messages are Fractal geometry was invented to approach roughness in
very complex, and a science can take off only after it suc- a very different way. Under magnification, smooth shapes
ceeds in identifying special cases that allow a workable are more and more closely approximated by their tangent
first step. For example, acoustics did not take its first step spaces. The more they are magnified, the simpler (bet-
with chirps or drums but with idealized vibrating strings. ter) they look. Over some range of magnifications, look-
These led to sinusoids and constants or other functions in- ing more closely at a rock or a coastline does not reveal
variant under translation in time. For the notion of rough- a simpler picture, but rather more of the same kind of
ness, no proper measure was available only 20 years ago. detail. Fractal geometry is based on this ubiquitous scale
The claim put forward forcibly in Mandelbrot (1982) is invariance. A fractal is an object that doesnt look any
that a workable entry is provided by rough shapes that are better when you blow it up. Scale invariance is also called
dilation invariant. These are fractals. symmetry under magnification.
Fractal roughness proves to be ubiquitous in the works A manifestation is that fractals are sets (or measures)
of nature and man. Those works of man range from mathe- that can be broken up into pieces, each of which closely
matics and the arts to the Internet and the financial markets. resembles the whole, except it is smaller. If the pieces scale
Those works of nature range from the cosmos to carbon isotropically, the shape is called self-similar; if different
deposits in diesel engines. A sketchy list would be use- scalings are used in different directions, the shape is called
less and a complete list, overwhelming. The reader is re- self-affine.
ferred to Frame and Mandelbrot (2001) and to a Panorama There are deep relations between the geometry of fractal
mentioned therein, available on the web. This essay is or- sets and the renormalization approach to critical phenom-
ganized around the mathematics of fractals, and concrete ena in statistical physics.
examples as illustrations of it.
To avoid the need to discuss the same topic twice, math- B. Examples of Self-Similar Fractals
ematical complexity is allowed to fluctuate up and down.
The reader who encounters paragraphs of oppressive dif- 1. Exact Linear Self-Similarity
ficulty is urged to skip ahead until the difficulty becomes A shape S is called exactly (linearly) self-similar if
manageable. the whole S splits into the union of parts Si : S =
S1 S2 . . . Sn . The parts satisfy two restrictions: (a)
each part Si is a copy of the whole S scaled by a linear
I. SCALE INVARIANCE contraction factor ri , and (b) the intersections between
parts are empty or small in the sense of dimension. An-
A. On Choosing a Symmetry Appropriate ticipating Section II, if i = j, the fractal dimension of the
to the Study of Roughness intersection Si S j must be lower than that of S. The
roughness of these sets is characterized by the sim-
The organization of experimental data into simple theoret- ilarity dimension d. In the special equiscaling case
ical models is one of the central works of every science; r1 = = rn = r , d = log(n)/ log(1/r ). In general, d is
invariances and the associated symmetries are powerful the solution of the Moran equation
tools for uncovering these models. The most common in- n
variances are those under Euclidean motions: translations, rid = 1.
rotations, reflections. The corresponding ideal physics is i=1
that of uniform or uniformly accelerated motion, uniform More details are given in Section II.
or smoothly varying pressure and density, smooth subman- Exactly self-similar fractals can be constructed by sev-
ifolds of Euclidean physical or phase space. The geometric eral elegant mathematical approaches.
alphabet is Euclidean, the analytical tool is calculus, the
statistics is stationary and Gaussian. a. Initiator and generator. An initiator is a starting
Few aspects of nature or man match these idealizations: shape; a generator is a juxtaposition of scaled copies of the
turbulent flows are grossly nonuniform; solid rocks are initiator. Replacing the smaller copies of the initiator in the
conspicuously cracked and porous; in nature and the stock generator with scaled copies of the generator sets in mo-
market, curves are nowhere smooth. One approach to this tion a process whose limit is an exactly self-similar frac-
discrepancy, successful for many problems, is to treat ob- tal. Stages before reaching the limit are called protofrac-
served objects and processes as roughened versions of tals. Each copy is anchored by a fixed point, and one may
an underlying smooth ideal. The underlying geometry is have to specify the orientation of each replacement. The
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 187

as the limit set of the orbit O+ (x0 ) of any point x0 under

the action of the semigroup generated by {T1 , . . . , Tn }.
The formal definition of IFS, which is delicate and
technical, proceeds as follows. Denoting by K the set
of nonempty compact subsets of Rn and by h the Haus-
FIGURE 1 Construction of the Sierpinksi gasket. The initiator
dorff metric on K [h(A, B) = inf{: A B and B A },
is a filled-in equilateral triangle, and the generator (on the left)
is made of N = 3 triangles, each obtained from the initiator by a where A = {x Rn : d(x, y) for some y A} is the -
contraction map Ti of reduction ratio r = 1/2. The contractions thickening of A, and d is the Euclidean metric], the Ti
fixed points are the vertices of the initiator. The middle shows the together
n define a transformation T : K K by T (A) =
second stage, replacing each copy of the initiator with a scaled {T
i=1 i (x): x A}, a contraction in the Hausdorff met-
copy of the generator. On the right is the seventh stage of the
ric with contraction ratio r = max{r1 , . . . , rn }. Because
construction.
(K, h) is complete, the contraction mapping principle
guarantees there is a unique fixed point C of T . This fixed
Sierpinski gasket (Fig. 1) is an example. The eye sponta- point is the attractor of the IFS {T1 , . . . , Tn }. Moreover, for
neously splits the whole S into parts. The simplest split any K K, the sequence K , T (K ), T 2 (K ), . . . converges
yields N = 3 parts Si , each a copy of the whole reduced by to C in the sense that limn h(T n (K ), C) = 0.
a similitude of ratio 1/2 and with fixed point at a vertex of The IFS inverse problem is, for a given compact set A
the initiator. In finer subdivisions, Si S j is either empty and given tolerance > 0, to find a set of transformations
or a single point, for which d = 0. In this example, but {T1 , . . . , Tn } with attractor C satisfying h(A, C) < . The
not always, it can be made empty by simply erasing the search for efficient algorithms to solve the inverse prob-
topmost point of every triangle in the construction. lem is the heart of fractal image compression. Detailed
Some of the most familiar fractals were orignally con- discussions can be found in Barnsley and Hurd (1993)
structed to provide instances of curves that exemplify and Fischer (1995).
properties deemed counterintuitive: classical curves may
have one multiple point (like Fig. 8) or a few. To the con- 2. Exact Nonlinear Self-Similarity
trary, the Sierpinski gasket (Fig. 2, far left) is a curve with
dense multiple points. The Sierpinski carpet (Fig. 2, mid A broader class of fractals is produced if the decomposi-
left) is a universal curve in the sense that one can embed tion of S into the union S = S1 S2 . . . Sn allows the
in the carpet every plane curve, irrespective of the collec- Si to be the images of S under nonlinear transformations.
tion of its multiple points. The Peano curve [initiator the
diagonal segment from (0, 0) to (1, 1), generator in Fig. 2 a. Quadratic Julia sets. For fixed complex number
mid right] is actually not a curve but a motion. It is plane- c, the quadratic orbit of the starting complex number z
filling: a continuous onto map [0, 1] [0, 1] [0, 1]. is a sequence of numbers that begins with f c (z) = z 2 + c,
then f c2 (z) = ( f c (z))2 + c and continues by following the
b. Iterated function systems. Iterated function sys- rule f cn (z) = f c ( f cn1 (z)). The filled-in (quadratic) Julia
tems (IFS) are a formalism for generating exactly self- set consists of the starting points that do not iter-
similar fractals based on work of Hutchinson (1981) and ate to infinity, formally, the points {z: f cn (z) remains
Mandelbrot (1982), and popularized by Barnsley (1988). bounded as n }. The (quadratic) Julia set Jc is the
IFS are the foundation of a substantial industry of im- boundary of the filled-in Julia set. Figure 3 shows the
age compression. The basis is a (usually) finite collection Julia set Jc for c = 0.4 + 0 i and the filled-in Julia set
{T1 , . . . , Tn } of contraction maps Ti : Rn Rn with con- for c = 0.544 + 0.576 i. The latter has an attracting
traction ratios ri < 1. Each Ti is assigned a probability pi 5-cycle, the black region is the basin of attraction of the
that serves, at each (discrete) instant of time, to select the
next map to be used. An IFS attractor also can be viewed

FIGURE 2 The Sierpinski gasket, Sierpinski carpet, the Peano FIGURE 3 The Julia set of z 2 + 0.4 (left) and the filled-in Julia
curve generator, and the fourth stage of the Peano curve. set for z 2 0.544 + 0.576 i (right).
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

188 Fractals

5-cycle, and the Julia set is the boundary of the black vation by Mandelbrot, Tan Lei (1984) proved the conver-
region. Certainly, Jc is invariant
under f c and under
the gence of appropriate magnifications of Julia sets and the M
1 1
inverses of f c , f c+ (z) = z c and f c (z) = z c. set at certain points named after Misiurewicz. Shishikura
Polynomial functions allow several equivalent character- (1994) proved Mandelbrots (1985) and Milnors (1989)
izations: Jc is the closure of the set of repelling periodic conjecture that the boundary of the M set has Hausdorff
points of f c (z) and Jc is the attractor of the nonlinear IFS dimension 2. Lyubich proved that the boundary of the M
1 1
{ f c+ , f c }. set is asymptotically self-similar about the Feigenbaum
Much is known about Julia sets of quadratic functions. point.
For example, McMullen proved that at a point whose ro- Mandelbrots first conjecture, that the interior of the M
tation number has periodic continued-fraction expansion, set consists entirely of components (called hyperbolic) for
the J set is asymptotically self-similar about the critical which there is a stable cycle, remains unproved in general,
point. though McMullen (1994) proved it for all such compo-
The J sets are defined for functions more general nents that intersect the real axis. Mandelbrots notion that
than polynomials. Visually striking and technically in- M may be the closure of M 0 is equivalent to the assertion
teresting examples correspond to the Newton function that the M set is locally connected. Despite intense ef-
N f (z) = z f (z)/ f  (z) for polynomial families f (z), or forts, that assertion remains a conjecture, though Yoccoz
entire functions like sin z, cos z, or exp z (see Sec- and others have made progress.
tion VIII.B). Discussions can be found in Blanchard Other developments include the theory of quadratic-like
(1994), Curry et al. (1983), Devaney (1994), Keen (1994), maps (Douady and Hubbard, 1985), implying the univer-
and Peitgen (1989). sality and ubiquity of the M set. This result was presaged
by the discovery (Curry et al., 1983) of a Mandelbrot set
b. The Mandelbrot set. The quadratic orbit f cn (z) in the parameter space of Newtons method for a family
always converges to infinity for large enough values of z. of cubic polynomials.
Mandelbrot attempted a computer study of the set M 0 of The recent book by Tan Lei (2000) surveys current re-
those values of c for which the orbit does not converge to sults and attests to the vitality of this field.
infinity, but to a stable cycle. This approach having proved
unrewarding, he moved on to a set that promised an easier c. Circle inversion limit sets. Inversion IC in a cir-
calculation and proved spectacular. Julia and Fatou, build- cle C with center O and radius r transforms a point P
ing on fundamental work of Montel, had shown that the into the point P  lying on the ray OP and with d(O, P)
Julia set Jc of f c (z) = z 2 + c must be either connected or d(O, P  ) = r 2 . This is the orientation-reversing involution
totally disconnected. Moreover, Jc is connected if, and defined on R2 {} by P IC (P) = P  . Inversion in C
only if, the orbit O+ (0) of the critical point z = 0 re- leaves C fixed, and interchanges the interior and exterior
mains bounded. The set M defined by {c: f cn (0) remains of C. It contracts the outer component not containing
bounded} is now called the Mandelbrot set (see the left O, but the contraction ratio is not bounded by any r < 1.
side of Fig. 4). Mandelbrot (1980) performed a computer Poincare generalized from inversion in one circle to a
investigation of its structure and reported several observa- collection of more than one inversion. As an example,
tions. As is now well known, small copies of the M set are consider a collection of circles C1 , . . . , C N each of which
infinitely numerous and dense in its boundary. The right is external to all the others. That is, for all j = i, the disks
side of Fig. 4 shows one such small copy, a nonlinearly bounded by Ci and C j have disjoint interiors. The limit set
distorted copy of the whole set. Although the small copy (C1 , . . . , C N ) of inversion in these circles is the set of
on the right side of Fig. 4 appears to be an isolated island, limit points of the orbit O+ (P) of any point P, external to
Mandelbrot conjectured and Douady and Hubbard (1984) C1 , . . . , C N , under the group generated by IC1 , . . . , IC N .
proved that the M set is connected. Sharpening an obser- Equivalently, it is the set left invariant by every one of the
inversions IC1 , . . . , IC N .
The limit set is nearly always fractal but the nonlin-
earity of inversion guarantees that is nonlinearly self-
similar. An example is shown in Fig. 5: the part of the limit
set inside C1 is easily seen to be the transform by I1 of the
part of the limit set inside C2 , C3 , C4 , and C5 .
How can one draw the limit set when the arrangement
FIGURE 4 Left: The Mandelbrot set. Right: A detail of the Man- of the circles C1 , . . . , C N is more involved? Poincares
delbrot set showing a small copy of the whole. Note the nonlinear original algorithm converges extraordinarily slowly. The
relation between the whole and the copy. first alternative algorithm was advanced in Mandelbrot
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 189

(Keen et al., 1993). In fact, the limit set of the Kleinian

groups that are in the Maskit embedding (Keen and Series,
1993) of the Teichmuller space of any finite-type Riemann
surface are Apollonian packings. These correspond to hy-
perbolic 3-manifolds having totally geodesic boundaries.
McShane et al. (1994) used automatic group theory to
produce efficient pictures of these limit sets, and Parker
(1995) showed that in many cases the Hausdorff dimen-
FIGURE 5 Left: The limit set generated by inversion in the five sion of the limit set equals the circle packing exponent,
circles C1 , . . . , C5 . Right: A magnification of the limit set. easily estimated as the slope of the loglog plot of the
number of circles of radius r (y axis) versus r (x axis).
Limit sets of Kleinian group actions are an excellent
(1982, Chapter 18); it is intuitive and the large-scale fea- example of a deep, subtle, and very active area of pure
tures of appear very rapidly, followed by increasingly mathematics in which fractals play a central role.
fine features down to any level a computers memory can
support. 3. Statistical Self-Similarity

d. Kleinian group limit sets. A Kleinian group A trees branches are not exact shrunken copies of that
(Beardon, 1983; Maskit, 1988) is a discrete group of tree, inlets in a bay are not exact shrunken copies of that
Mobius transformations bay, nor is each cloud made up of exact smaller copies
of that cloud. To justify the role of fractal geometry as
az + b
z a geometry of nature, one must take a step beyond ex-
cz + d act self-similarity (linear or otherwise). Some element of
randomness appears to be present in many natural objects
acting on the Riemann sphere C, the sphere at infinity
and processes. To accommodate this, the notions of self-
of hyperbolic 3-space H3 . The isometries of H3 can be
similarity and self-affinity are made statistical.
represented by complex matrices
 
a b a. Wiener brownian motion: its graphs and trails.
. The first example is classical. It is one-dimensional
c d
Brownian motion, the random process X (t) defined by
[More precisely, by their equivalence classes in these properties: (1) with probability 1, X (0) = 0 and X (t)
P S L 2 (C).] Sullivans side-by-side dictionary (Sullivan, is continuous, and (2) the increments X (t +
t) X (t) of
1985) between Kleinian groups and iterates of rational X (t) are Gaussian with mean 0 and variance
t. That is,
maps is another deep mathematical realm informed, 1
at least in part, by fractal geometry. Thurstons ge- Pr {X (t +
t) X (t) x} =
2
t
ometrization program for 3-manifolds (Thurston, 1997)  x  2
involves giving many 3-manifolds hyperbolic structures u
exp du.
by viewing them as quotients of H3 by the action of a 2
t
Kleinian group G (Epstein, 1986). The corresponding
An immediate consequence is independence of incre-
action of G on C determines the limit set (G), defined
ments over disjoint intervals. A fundamental property of
as the intersection of all nonempty G-invariant subsets
Brownian motion is statistical self-affinity: for all s > 0,
of C. For many G, the limit set is a fractal. An example
gives the flavor of typical results: the limit set of a finitely Pr {X (s(t +
t)) X (st) sx} = Pr {X (t +
t)
generated Kleinian group is either totally disconnected, a
X (t) x}.
circle, or has Hausdorff dimension greater than 1 (Bishop
and Jones, 1997). The Hausdorff dimension of the limit That is, rescaling t by a factor of s, and of x by a fac-
set has been studied by Beardon, Bishop, Bowen, Canary, tor of s, leaves the distribution unchanged. This correct
Jones, Keen, Mantica, Maskit, McMullen, Mumford, rescaling is shown on the left panel of Fig. 6: t (on the
Parker, Patterson, Sullivan, Tricot, Tukia, and many horizontal axis) is scaled by 4, x (on the vertical axis) is
others. Poincare exponents, eigenvalues of the Laplacian, scaled by 2 = 41/2 . Note that this magnification has about
and entropy of geodesic flows are among the tools used. the same degree of roughness as the full picture. In the
Figure 5 brings forth a relation between some limit sets center panel, t is scaled by 4, x by 4/3; the magnification
of inversions or Kleinian groups and Apollonian packings is flatter than the original. In the right panel, both t and
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

190 Fractals

FIGURE 6 Left panel: Correct rescaling illustrating the self-

affinity of Brownian motion. Center and right panels: Two incorrect
rescalings. FIGURE 8 A Brownian trail. Right: The first quarter of the left
trail, magnified and with additional turns interpolated so the left
and right pictures have about the same number of turns.
x are scaled by 4; the magnification is steeper than the
original.
A sequence of increments of Brownian motion is called r For E 2 a Brownian trail B: [0, 1] R E has
Gaussian white noise. Even casual inspection of the graph Hausdorff and box dimensions dH = dbox = 2,
reveals some fundamental features. The width of an old respectively.
pen-plotter line being equal to the spacing between suc- r The graph of one-dimensional Brownian motion
cessive difference values, the bulk of the difference plot B: [0, 1] R has dH = dbox = 3/2.
merges into a band with the following properties (see
Fig. 7):
Some related constructions have been more resistant
r The bands width is approximately constant. to theoretical analysis. Mandelbrots planar Brownian
r The values beyond that band stay close to it (this is due cluster is the graph of the complex B(t) constrained
to satisfy B(0) = B(1). It can be constructed by lin-
to the fact that the Gaussian has short tails).
r The values beyond that band do not cluster. early detrending the x- and y-coordinate functions:
(X (t) t X (1), Y (t) tY (1)). See Fig. 9. The cluster is
Positioning E independent one-dimensional Brownian known to have dimension 2. Visual inspection supported
motions along E coordinate axes gives a higher dimen- by computer experiments led to the 4/3 conjecture, which
sional Brownian motion: B(t) = {X 1 (t), . . . , X E (t)}. Plot- asserts that the boundary of the cluster has dimension
ted as a curve in E-dimensional space, the collection of 4/3 (Mandelbrot, 1982, p. 243). This has been proved
points that B(t) visits between t = 0 and t = 1 defines a by Lawler et al. (2000).
Brownian trail. Brownian motion is the unique stationary random pro-
When E > 2, this is an example of a statistically self- cess with increments independent over disjoint intervals
similar fractal. To split the Brownian trail into N reduced- and with finite variance. For many applications, these con-
scale parts, pick N 1 arbitrary instants tn with 0 = t0 < ditions are too restrictive, drawing attention to other ran-
t1 < < t N 1 < t N = 1. The Brownian trail for 0 t 1 dom processes that retain scaling but abandon either in-
splits into N subtrails Bn for the interval tn1 < t < tn . dependent increments or finite variance.
The parts Bn follow the same statistical distribution as
the whole, after the size is expanded by (tn tn1 )1/2 in b. Fractional Brownian motion. For fixed 0 < H <
every direction. 1, fractional Brownian motion (FBM) of exponent H is
Due to the definition of self-similarity, this example a random process X (t) with increments X (t +
t) X (t)
reveals a pesky complication: for i = j, Bi B j must be following the Gaussian distribution with mean 0 and stan-
of dimension less than B. This is indeed the case if E > 2, dard deviation (
t) H . Statistical self-affinity is straight-
but not in the plane E = 2. However, the overall idea can be forward: for all s > 0
illustrated for E = 2. The right side of Fig. 8 shows B1 (t) Pr {X (s(t +
t)) X (st) s H x}
for 0 t 1/4, expanded by a factor of 2 = (1/4 0)1/2
and with additional points interpolated so the part and = Pr {X (t +
t) X (t) x}.
the whole exhibit about the same number of turns. The
The correlation is the expected value of the product of
details for E = 2 are unexpectedly complex, as shown in
successive increments. It equals
Mandelbrot (2001b, Chapter 3).
The dimensions (see Sections II.B and II.E) of some
Brownian constructions are well-known, at least in most
cases. For example, with probability 1:

FIGURE 7 Plot of 4000 successive Brownian increments. FIGURE 9 A Brownian cluster.

P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 191

FIGURE 12 Levy flight on the line. Left: the graph as a function

of time. Right, the increments.

FIGURE 10 Top: Fractional Brownian motion simulations with and a (one-dimensional) Levy stable process is defined as
H = 0.25, H = 0.5, and H = 0.75. Bottom: Difference plots X(t +
a sum
1) X(t) of the graphs above.


f (t) = k (t tk ),
k=1
E((X (t) X (0)) (X (t + h) X (t)))
where the pulse times tn and amplitudes n are cho-
= 12 ((t + h)2H t 2H h 2H ). sen according to the following Levy measure: given t
and , the probability of choosing (ti , i ) in the rectan-
If H = 1/2, this correlation vanishes and the increments
gle t < ti < t + dt , < i < + d is C 1 d dt. Fig-
are independent. In fact, FBM reduces to Brownian mo-
ure 12 shows the graph of a Levy process or flight, and a
tion. If H > 1/2, the correlation is positive, so the incre-
graph of its increments.
ments tend to have the same sign. This is persistent FBM.
Comparing Figs. 7, 10, and 12 illustrates the power of
If H < 1/2, the correlation is negative, so the increments
the increment plot for revealing both global correlations
tend to have opposite signs. This is antipersistent FBM.
(FBM) and long tails (Levy processes).
See Fig. 10. The exponent determines the dimension of
The effect of large excursions in Levy processess is
the graph of FBM: with probability 1, dH = dbox = 2 H .
more visible in the plane. See Fig. 13. These Levy flights
Notice that for H > 1/2, the central band of the difference
were used in Mandelbrot (1982, Chapter 32) to mimic the
plot moves up and down, a sign of long-range correlation,
statistical properties of galaxy distributions.
but the outliers still are small. Figure 11 shows the trails
Using fractional Brownian motion and Levy processes,
of these three flavors of FBM. FBM is the main topic of
Mandelbrot (in 1965 and 1963) improved upon Bache-
Mandelbrot (2001c).
liers Brownian model of the stock market. The former
corrects the independence of Brownian motion, the lat-
c. Levy stable processes. While FBM introduces
ter corrects its short tails. The original and corrected pro-
correlations, its increments remain Gaussian and so have
cesses in the preceding sentence are statistically self-affine
small outliers. The Gaussian distribution is characterized
random fractal processes. This demonstrates the power
by its first two moments (mean and variance), but some
of invariances in financial modeling; see Mandelbrot
natural phenomena appear to have distributions for which
(1997a,b).
these are not useful indicators. For example, at the critical
point of percolation there are clusters of all sizes and the
d. Self-affine cartoons with mild to wild random-
expected cluster size diverges.
ness. Many natural processes exhibit long tails or global
Paul Levy studied random walks for which the jump dis-
dependence or both, so it was a pleasant surprise that both
tributions follow the power law Pr {X > x} x . There
can be incorporated in an elegant family of simple car-
is a geometrical approach for generating examples of Levy
toons. (Mandelbrot, 1997a, Chapter 6; 1999, Chapter N1;
processes.
2001a). Like for self-similar curves (Section I.B.1.a), the
The unit step function (t) is defined by
basic construction of the cartoon involves an initiator and a

0 for x < 0
(t) =
1 for x 0

FIGURE 11 Top: Fractional Brownian motion simulations with

H = 0.25, H = 0.5, and H = 0.75. Bottom: Difference plots X(t + FIGURE 13 Left: Trail of the Levy flight in the plane. Right: The
1) X(t) of the graphs above. Levy dust formed by the turning points.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

192 Fractals

spanning cluster connects opposite sides of the lattice.

For large L there is a critical probability or percolation
threshold pc ; spanning clusters do not arise for p < pc .
Numerical experiments suggest pc 0.59275. In Fig. 17,
FIGURE 14 The initiator (left), generator (middle), and first gen- p = 0.4, 0.6, and 0.8. Every lattice has its own pc .
eration (right) of the Brownian cartoon.
At p = pc the masses of the spanning clusters scale
with the lattice size L as L d , independently of the lattices.
generator. The process used to generate the graph consists Experiment yields d = 1.89 0.03, and theory yields
in replacing each copy of the initiator with an appropriately d = 93/49. This d is the mass dimension of Section II.C.
rescaled copy of the generator. For a Brownian cartoon, In addition, spanning clusters have holes of all sizes; they
the initiator can be the diagonal of the unit square, and the are statistically self-similar fractals.
generator, the broken line with vertices (0, 0), (4/9, 2/3), Many fractals are defined as part of a percolation clus-
(5/9, 1/3), and (1, 1). Pictured in Fig. 14 are the initiator ter. The backbone is the subset of the spanning cluster
(left), generator (middle), and first iteration of the process that remains after removing all parts that can be separated
(right). from both spanned sides by removing a single filled cell
To get an appreciation for how quickly the local rough- from the spanning cluster. Numerical estimates suggest
ness of these pictures increases, the left side of Fig. 15 the backbone has dimension 1.61. The backbone is the
shows the sixth iterate of the process. path followed by a fluid diffusing through the lattice.
Self-affinity is built in because each piece is an appro- The hull of a spanning cluster is its boundary. It was
priately scaled version of the whole. In Fig. 14, the scal- observed by R. F. Voss in 1984 and proven by B. Duplantier
ing ratios have been selected to match the square root that the hulls dimension is 7/4.
property of Brownian motion: for each segment of the A more demanding definition of the boundary yields the
generator we have |
xi | = (
ti )1/2 . perimeter. It was observed by T. Grossman and proven by
More generally, a cartoon is called unifractal if there B. Duplantier that the perimeters dimension is 4/3.
is a constant H with |
xi | = (
ti ) H for each generator Sapoval et al. (1985) examined discrete diffusion and
segment, where 0 < H < 1. If different H are needed for showed that it involves a fractal diffusion front that can
different segments, the cartoon is multifractal. be modeled by the hull and the perimeter of a percolation
The left side of Figure 15 is too symmetric to mimic cluster.
any real data, but this problem is palliated by shuffling the
order in which the three pieces of the generator are put f. Diffusion-limited aggregation (DLA; Vicsek,
into each scaled copy. The right side of Fig. 15 shows a 1992). DLA was proposed by Witten and Sander (1981,
Brownian cartoon randomized in this way. 1983) to simulate the aggregates that carbon particles
Figure 16 illustrates how the statistical properties of form in a diesel engine. On a grid of square cells, a cartoon
the increments can be modified by adjusting the genera- of DLA begins by occupying the center of the grid with a
tor in a symmetrical fashion. Keeping fixed the endpoints seed particle. Next, place a particle in a square selected
(0, 0) and (1, 1), the middle turning points are changed at random on the edge of a large circle centered on the seed
into (a, 2/3) and (1 a, 1/3) for 0 < a 1/2. square and let it perform a simple random walk. With each
tick of the clock, with equal probabilities it will move to an
e. Percolation clusters (Stauffer and Aharony, adjacent square, left, right, above, or below. If the moving
1992). Given a square lattice of side length L and a particle wanders too far from the seed, it falls off the edge
number p [0, 1], assign a random number x [0, 1] to of the grid and another wandering particle is started at a
each lattice cell and fill the cell if x p. A cluster is a randomly chosen edge point. When a wandering particle
maximal collection of filled cells, connected by sharing reaches one of the four squares adjacent to the seed, it
common edges. Three examples are shown in Fig. 17. A sticks to form a cluster of two particles, and another mov-
ing particle is released. When a moving particle reaches a
square adjacent to the cluster, it sticks there. Continuing
in this way builds an arbitrarily large object called a
diffusion-limited aggregate (DLA) because the growth of
the cluster is governed by the particles diffusing across
the grid. Figure 18 shows a moderate-size DLA cluster.
Early computer experiments on clusters of up to the 104
FIGURE 15 Left: The sixth iterate of the process of Fig. 14. Right: particles showed the mass M(r ) of the part of the cluster
A sixth iterate of a randomized Brownian cartoon. a distance r from the seed point scales as M(r ) k r d ,
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 193

FIGURE 16 Generators, cartoons and difference graphs for symmetric cartoons with turning points (a, 2/3) and
(1 a, 1/3), for a = 0.333, 0.389, 0.444, 0.456, and 0.467. The same random number seed is used in all graphs.

with d 1.71 for clusters in the plane and d 2.5 for most general cases but coincide for exactly self-similar
clusters in space. This exponent d is the mass dimen- fractals. Many other dimensions cannot be mentioned
sion of the cluster. (See Sections II.C and V.) These val- here.
ues match measured scalings of physical objects moder- A more general approach to quantifying degrees of
ately, but not terribly well. A careful examination of much roughness is found in the article on Multifractals.
larger clusters revealed discrepancies that added in due
time to a very complex picture of DLA. Mandelbrot et al. A. Similarity Dimension
(1995) investigated clusters in the 107 range; careful mea-
surement reveals an additional dimension of 1.65 0.01. The definition of similarity dimension is rooted in the fact
This suggests the clusters become more compact as they that the unit cube in D-dimensional Euclidean space is
grow. Also, as the cluster grows, more arms develop and self-similar: for any positive integer b the cube can be de-
the largest gaps decrease in size; i.e., the lacunarity de- composed into N = b D cubes, each scaled by the similar-
screases. (See Section VI.) ity ratio r = 1/b, and overlapping at most along (D 1)-
dimensional cubes.
The equiscaling or isoscaling case. Provided the
II. THE GENERIC NOTION OF FRACTAL pieces do not overlap significantly, the power-law relation
DIMENSION AND A FEW SPECIFIC N = (1/r ) D between the number N and scaling factor r of
IMPLEMENTATIONS the pieces generalizes to all exactly self-similar sets with
all pieces scaled by the factor r . The similarity dimension
The first, but certainly not the last, step in quantifying dsim is
fractals is the computation of a dimension. The notion of log(N )
dsim = .
Euclidean dimension has many aspects and therefore ex- log(1/r )
tends in several fashions. The extensions are distinct in the

FIGURE 17 Percolation lattices well below, near, and well above

the percolation threshold. FIGURE 18 A moderate-size DLA cluster.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

194 Fractals

The pluriscaling case. More generally, for self-similar M(r ) = V (d) r d

sets where each piece is scaled by a possibly different

d 

factor ri , the similarity dimension is the unique positive with V (d) =  12  d + 12 ,
root d of the Moran equation where V (d) is the volume of the d-dimensional unit

N sphere. That is, for constant-density Euclidean objects,
rid = 1. the ordinary dimensionamong many other rolesis the
i=1 exponent relating mass to size. This role motivated the
The relation 0 dsim E. If the fractal is a subset of definition of mass dimension for a fractal. The definition
E-dimensional Euclidean space, E is called the embed- of mass is delicate. For example, the mass of a Sierpin-
ding dimension. ski gasket cannot be defined by starting with a triangle of
So long as the overlap of the parts is not too great uniform density and removing middle triangles; this pro-
(technically, under the open set condition), we have cess would converge to a mass reduced to 0. One must,
dsim E. If at least two of the ri are positive, we have instead, proceed as on the left side of Fig. 1: take as ini-
dsim > 0. However, Section III.F shows that some circum- tiator a triangle of mass 1, and as generator three triangles
stances introduce a latent dimension d, related indirectly each scaled by 1/2 and of mass 1/3. Moreover, two very
to the similarity dimension, and that can satisfy d < 0 new facts come up.
or d > E. Firstly, the = sign in the formula for M(r ) must be re-
placed by . That is, M(r ) fluctuates around a multiple
of r d . For example, as mentioned in Sections I.B.3.e and
B. Box Dimension I.B.3.f, the masses of spanning percolation clusters and
The similarity dimension is meaningful only for exactly diffusion-limited aggregates scale as a power-law func-
self-similar sets. For more general sets, including exper- tion of size. Consequently, the exponent in the relation
imental data, it is often replaced by the box dimension. M(r ) k r dmass is called the mass dimension.
For any bounded (nonempty) set A in E-dimensional Second, in the Euclidean case the center is arbitrary
Euclidean space, and for any > 0, a -cover of A is a but in the fractal case it must belong to the set un-
collection of sets of diameter whose union contains A. der consideration. As an example, Fig. 19 illustrates
Denote by N (A) the smallest number of sets in a -cover attempts to measure the mass dimension of the Sier-
of A. Then the box dimension dbox of A is pinski gasket. Suppose we take circles centered at the
lower left vertex of the initiator, and having radii 1/2,
log(N (A)) 1/4, 1/8, . . . . We obtain M(1/2i ) = 1/3i = (1/2i )d , where
dbox = lim
0 log(1/) dmass = log 3/ log 2. See the left side of Fig. 19. That is,
when the limit exists. When the limit does not exist, the the mass dimension agrees with the similarity dimension.
replacement of lim with lim sup and lim inf defines the On the other hand, if the circles center is randomly
upper and lower box dimensions: selected in the interior of the initiator, the passage to the
limit r 0 almost surely eventually stops with circles
log(N (A))
dbox = lim sup , bounding no part of the gasket. See the middle of Fig. 19.
0 log(1/) Taking a family of circles with center c a point of the
log(N (A)) gasket, the massradius relation becomes M(r ) = k(r, c)
dbox = lim inf . r d , where the prefactor k(r, c) fluctuates and depends on
0 log(1/)
both r and c. See the right side of Fig. 19. Even in this
The box dimension can be thought of as measuring how case, the exponent is the mass dimension. The prefactor
well a set can be covered with small boxes of equal size, is no longer a constant density, but a random variable,
because the limit (or lim sup and lim inf) remain un- depending on the choice of the origin.
changed if N (A) is replaced by the smallest number of
E-dimensional cubes of side needed to cover A, or even
the number of cubes of a lattice that intersect A.
Section V describes methods of measuring the box di-
mension for physical datasets.

C. Mass Dimension
The mass M(r ) of a d-dimensional Euclidean ball of con- FIGURE 19 Attempts at measuring the mass dimension of a
stant density and radius r is given by Sierpinski gasket using three families of circles.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 195

Section V describes methods of measuring the mass more difficult because the inf is taken over the collection of
dimension for physical datasets. all -covers. Because of the inf that enters in its definition,
the HausdorffBesicovitch dimension cannot be measured
for any physical object.
D. MinkowskiBouligand Dimension
Note: If A can be covered by N (A) sets of diame-
Given a set A R E and > 0, the Minkowski sausage of ter at most , then Hs (A) N (A) s . From this it fol-
A, also called the -thickening or -neighborhood of A, lows dH (A) dbox (A), so dH (A) dbox (A) if dbox (A) ex-
is defined as A = {x R E: d(x, y) for some y A}. ists. This inequality can be strict. For example, if A is
(See Section I.B.b.) In the Euclidean case when A is a any countable set, dH (A) = 0 and yet dbox (rationals in
smooth m-dimensional manifold imbedded in R E , one has [0, 1]) = 1.
vol(A ) Em . That is, the E-dimensional volume
of A scales as to the codimension of A. This concept
F. Packing Dimension
extends to fractal sets A: if the limit exists,
log(vol(A )) Hausdorff dimension measures the efficiency of covering
E lim a set by disks of varying radius. Tricot (1982) introduced
0 log()
packing dimension to measure the efficiency of packing
defines the MinkowskiBouligand dimension, dMB (A) (see a set with disjoint disks of varying radius. Specifically,
Mandelbrot, 1982, p. 358). In fact, it is not difficult to see for > 0 a -packing of A is a countable collection of
that dMB (A) = dbox (A). If the limit does not exist, lim sup disjoint disks {Bi } with radii ri < and with centers in A.
gives dbox (A) and lim inf gives dbox (A). In analogy with Hausdorff measure, define
In the privileged case when the limit 

vol(A ) P (A) = sup
s
|Bi | : {Bi } is a -packing of A .
s
lim
Em
0 i

exists, it generalizes the notion of Minkowski content for As decreases, so does the collection of -packings of A.
smooth manifolds A. Section VI will use this prefactor to Thus Ps (A) decreases as decreases and the limit
measure lacunarity.
P0s (A) = lim Ps (A)
0

E. HausdorffBesicovitch Dimension exists. A technical complication requires an additional

step. The s-dimensional packing measure of A is defined
For a set A in Euclidean space, given s 0 and > 0, as
consider the quantity 
  

 P (A) = inf
s
P0 (Ai ): A
s
Ai .
H (A) = inf
s
|Ui | : {Ui } is a -cover of A .
s i i=1
i
Then the packing dimension dpack (A) is
A decrease of reduces the collection of -covers of
dpack (A) = inf{s: P s (A) = 0} = sup{s: P s (A) = }.
A, therefore Hs (A) increases as 0 and Hs (A) =
lim0 Hs (A) exists. This limit defines the s-dimensional Packing, Hausdorff, and box dimensions are related:
Hausdorff measure of A. For t > s, Ht (A) ts Hs (A).
dH (A) dpack (A) dbox (A).
It follows that a unique number dH has the property that
For appropriate A, each inequality is strict.
s < dH implies Hs (A) =
and
III. ALGEBRA OF DIMENSIONS
s > dH implies Hs (A) = 0.
AND LATENT DIMENSIONS
That is,
The dimensions of ordinary Euclidean sets obey sev-
dH (A) = inf{s: Hs (A) = 0} = sup{s: Hs (A) = }.
eral rules of thumb that are widely used, though rarely
This quantity dH is the HausdorffBesicovitch dimension stated explicitly. For example, the union of two sets of
of A. It is of substantial theoretical significance, but in dimension d and d  usually has dimension max{d, d  }.
most cases is quite challenging to compute, even though it The projection of a set of dimension d to a set of di-
suffices to use coverings by disks. An upper bound often is mension d  usually gives a set of dimension min{d, d  }.
relatively easy to obtain, but the lower bound can be much Also, for Cartesian products, the dimensions usually add:
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

196 Fractals

dim(A B) = dim(A) + dim(B). For the intersection of The obvious generalization holds for fractals A R E
subsets A and B of R E , it is the codimensions that usually and projections to k-dimensional hyperplanes through the
add: E dim(A B) = (E dim(A)) + (E dim(B)), origin.
but only so long as the sum of the codimensions is non- Projections of fractals can be very complicated. There
negative. If this sum is negative, the intersection is empty. are fractal sets A R3 with the surprising property that
Mandelbrot (1984, Part II) generalized those rules to frac- for almost every plane P through the origin, the projection
tals and (see Section III.G) interpreted negative dimen- proj P (A) is any prescribed shape, to within a set of area 0.
sions as measures of degree of emptiness. Consequently, as Falconer (1987) points out, in principle
For simplicity, we restrict our attention to generaling we could build a fractal digital sundial.
these properties to the Hausdorff and box dimensions of
fractals. D. Subordination and Products of Dimension
We have already seen operations realizing the sum, max,
A. Dimension of Unions and Subsets and min of dimensions, and in the next subsection we shall
Simple applications of the definition of Hausdorff dimen- examine the sum of codimensions. For certain types of
sion give fractals, multiplication of dimensions is achieved through
subordination, a process introduced in Bochner (1955)
AB implies dH (A) dH (B) and elaborated in Mandelbrot (1982). Examples are con-
and structed easily from the Koch curve generator (Fig. 20a).
The initiator (the unit interval) is unchanged, but the new
dH (A B) = max{dH (A), dH (B)}. generator is a subset of the original generator. Figure 20
Replacing max with sup, this property holds for countable shows three examples.
collections of sets. The subset and finite union properties In Fig. 20, generator (b) gives a fractal dust (B) of
hold for box dimension, but the countable union property dimension log 3/log 3 = 1. Generator (c) gives the stan-
fails. dard Cantor dust (C) of dimension log 2/log 3. Generator
(d) gives a fractal dust (D) also of dimension log 2/log 3.
Thinking of the Koch curve K as the graph of a func-
B. Product and Sums of Dimensions tion f : [0, 1] K R2 , the fractal (B) can be obtained
For all subsets A and B of Euclidean space, by restricting f to the Cantor set with initiator [0, 1] and
dH (A B) dH (A) + dH (B). Equality holds if one generator the intervals [0, 1/4], [1/4, 1/2], and [1/2, 3/4].
of the sets is sufficiently regular. For example, if dH (A) = In this case, the subordinand is a Koch curve, the subor-
dH (A), then dH (A B) = dH (A) + dH (B). Equality does dinator is a Cantor set, and the subordinate is the fractal
not always hold: Besicovitch and Moran (1945) give an (B). The identity
example of subsets A and B of R with dH (A) = dH (B) = 0, log 3 log 4 log 3
yet dH (A B) = 1. =
log 3 log 3 log 4
For upper box dimensions, the inequality is reversed:
dbox (A B) dbox (A) + dbox (B). expresses that the dimensions multiply,
dim(subordinate) = dim(subordinand)
C. Projection dim(subordinator).
Denote by proj P (A) the projection of a set A R to a3
Figure 20, (C) and (D) give other illustrations of this
plane P R3 through the origin. If A is a one-dimensional multiplicative relation. The seeded universe model
Euclidean object, then for almost all choices of the plane of the distribution of galaxies (Section IX.D.1) uses
P, projP (A) is one-dimensional. If A is a two- or three- subordination to obtain fractal dusts; see Mandelbrot
dimensional Euclidean object, then for almost all choices (1982, plate 298).
of the plane P, proj P (A) is two-dimensional of positive
area. That is, dim(proj P (A)) = min{dim(A), dim(P)}.
The analogous properties hold for fractal sets A. If
dH (A) < 2, then for almost all choices of the plane
P, dH (proj P (A)) = dH (A). If dH (A) 2, then for al-
most all choices of the plane P, dH (proj P (A)) = 2 and FIGURE 20 The Koch curve (A) and its generator (a); (b), (c),
proj P (A) has positive area. So again, dH (proj P (A)) = and (d) are subordinators, and the corresponding subordinates of
min{dH (A), dH (P)}. the subordinand (A) are (B), (C), and (D).
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 197

E. Intersection and Sums of Codimension Embedding. A problem that concerns R2 can often be
reinterpreted as a problem that really concerns R E , with
The dimension of the intersection of two sets obviously
E > 2, but must be approached within planar intuitions by
depends on their relative placement. When A B = , the
R2 . Conversely, if a given problem can be embedded into a
dimension vanishes. The following is a typical result. For
problem concerning R E , the question arises, which is the
Borel subsets A and B of R E , and for almost all x R E ,
critical value of E 2, defined as the smallest value for
dH (A (B + x)) max{0, dH (A B) E}. which the intersection ceases to be empty, and precisely
If dH (A B) = dH (A) + dH (B), this reduces to reduces to a point? In the example of a line and a point,
the critical E 2 is precisely 1: once embedded in R3 , the
dH (A (B + x)) max{0, dH (A) + dH (B) E}. problem transforms into the intersection of a plane and a
This is reminiscent of the transversality relation for inter- line, which is a point.
sections of smooth manifolds. Approximation and pre-asymptotics in mathematics
Corresponding lower bounds are known in more re- and the sciences. Consider a set defined as the limit of
stricted circumstances. For example, there is a positive a sequence of decreasing approximations. When the limit
measure set M of similarity transformations of R E with is not empty, all the usual dimensions are defined as be-
ing properties of the limit, but when the limit is empty
dH (A T (B)) dH (A) + d H (B) E and all the dimensions vanish, it is possible to consider
for all T M. Note dH (A T (B)) = dH (A) + dH (B) E instead the limits of the properties of the approximations.
is equivalent to the addition of codimensions: E The MinkowskiBouligand formal definition of dimen-
dH (A T (B)) = (E dH (A)) + (E dH (B)). sion generalizes to fit the naive intuitive values that may
be either positive or negative.
F. Latent Dimensions below 0 or above E
2. Latent Dimensions That Exceed
A blind application of the rule that codimensions are addi- That of the Embedding Space
tive easily yields results that seem nonsensical, yet become
useful if they are properly interpreted and the Hausdorff For a strictly self-similar set in R E , the Moran equation
dimension is replaced by a suitable new alternative. defines a similarity dimension that obeys dsim E. On
the other hand, a generator that is a self-avoiding broken
line can easily yield log(N )/ log(1/r ) = dsim > E. Recur-
1. Negative Latent Dimensions as Measures
sive application of this generator defines a parametrized
of the Degree of Emptiness
motion, but the union of the positions of the motion is
Section E noted that if the codimension addition rule gives neither a self-similar curve nor any other self-similar set.
a negative dimension, the actual dimension is 0. This ex- It is, instead, a set whose points are covered infinitely of-
ception is an irritating complication and hides a feature ten. Its box dimension is E, which a fortiori is <dsim .
worth underlining. However, one can load a mass on this set by following the
As background relative to the plane, consider the fol- route that applies in the absence of multiple points. Mass
lowing intersections of two Euclidean objects: two points, is distributed on the generators intervals in proportion to
a point and a line, and two lines. Naive intuition tells the values of ridsim . By infinite recursion, the difference
us that the intersection of two points is emptier than between the times t  and t  when points P  and P  are
the intersection of a point and a line, and that the lat- visited is defined as the mass supported by the portion of
ter in turn is emptier than the intersection of two lines the curve that links these points.
(which is almost surely a point). This informal intu- If so and dsim > E, the similarity dimension acquires a
ition fails to be expressed by either a Euclidean or a useful role as a latent dimension. For example, consider
Hausdorff dimension. On the other hand, the formal ad- the multiplication of dimensions in Section III.D. Suppose
dition of codimensions suggests that the three intersec- that our recursively constructed set is not lighted for all
tions in question have the respective dimensions 2, 1, instants of time, but only intermittently when time falls
and 0. The inequalities between those values conform within a fractal dust of dimension d  . Then, the rule of
with the above-mentioned naive intuition. Therefore, they thumb is that the latent dimension of the lighted points
ushered in the search for a new mathematical defini- is dsim d  . When dsim d  < E, the rule of thumb is that the
tion of dimension that can be measured and for which true dimension is also dsim d  .
negative values are legitimate and intuitive. This search Figure 21 shows an example. The generator has N = 6
produced several publications leading to Mandelbrot segments, each with scaling ratio r = 1/2, hence latent
(1995). Two notions should be mentioned. dimension dsim = log 6/log 2 > 2. Taking as subordinator
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

198 Fractals

FIGURE 21 Left: Generator and limiting shape with latent dimension exceeding 2. Right: generator and limiting
shape of a subordinate with dimension <2. For comparison, this limiting shape is enclosed in the outline of the left
limiting shape.


a Cantor set with generator having N = 3 segments, each d(x)
s (x) = .
with scaling ratio r = 1/2, yields a self-similar fractal with |x y|s
dimension log 3/log 2.
 there is a mass distribution on a set A with
If
G. Mapping s (x) d(x) < , then dH (A) s. Potential theory has
been useful for computing dimension of many sets, for
Recall f satisfies the Holder condition with exponent example, Brownian paths.
H if there is a positive constant c for which | f (x)
f (y)| c|x y| H . For such functions, dH ( f (A))
(1/H )dH (A). If H = 1, f is called a Lipschitz function; C. Implicit Methods
f is bi-Lipschitz if there are constants c1 and c2 with
c1 |x y| | f (x) f (y)| c2 |x y|. Hausdorff dimen- McLaughlin (1987) introduced a geometrical method,
sion is invariant under bi-Lipschitz maps. The analogous based on local approximate self-similarities, which suc-
properties hold for box-counting dimension. ceeds in proving that dH (A) = dbox (A), without first deter-
mining dH (A). If small parts of A can be mapped to large
parts of A without too much distortion, or if A can be
IV. METHODS OF COMPUTING DIMENSION mapped to small parts of A without too much distortion,
IN MATHEMATICAL FRACTALS then dH (A) = dbox (A) = s and Hs (A) > 0 (in the former
case) or Hs (A) < (in the latter case). Details and ex-
Upper bounds for the Hausdorff dimension can be rel- amples can be found in Falconer (1997, Section 3.1).
atively straightforward: it suffices to consider a specific
family of coverings of the set. Lower bounds are more
delicate. We list and describe briefly some methods for D. Thermodynamic Formalism
computing dimension.
Sinai (1972), Bowen (1975), and Ruelle (1978) adapted
methods of statistical mechanics to determine the dimen-
A. Mass Distribution Methods
sions of fractals arising from some nonlinear processes.
A mass distribution on a set A is a measure with Roughly, for a fractal defined as the attractor A of a fam-
supp() A and 0 < (A) < . The mass distribution ily of nonlinear contractions Fi with an inverse function f
principle (Falconer, 1990, p. 55) establishes a lower bound defined on A, the topological pressure P() of a Lipschitz
for the Hausdorff dimension: Let be a mass distribution function : A R is
on A and suppose for some s there are constants c > 0
1 
and > 0 with (U ) c |U |s for all sets U with |U | . P() = lim log exp[(x) + ( f (x))
Then dH (A). k k
x Fix( f k )
Suitable choice of mass distribution can show that no

individual set of a cover can cover too much of A. This
can eliminate the problems caused by covers by sets of a + ( f k1
(x))] ,
wide range of diameters.
where Fix( f k ) denotes the set of fixed points of f k . The
B. Potential Theory Methods
sum plays the role of the partition function in statistical
Given a mass distribution , the s-potential is defined by mechanics, part of the motivation for the name ther-
Frostman (1935) as modynamic formalism. There is a unique s for which
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 199

P(s log | f  |) = 0, and s = dH (A). Under these condi-

tions, 0 < Hs (A) < , Hs (A) is a Gibbs measure on A,
and many other results can be deduced. Among other
places, this method has been applied effectively to the
study of Julia sets.

FIGURE 22 Two Sierpinski carpet fractals with the same

V. METHODS OF MEASURING DIMENSION dimension.
IN PHYSICAL SYSTEMS

For shapes represented in the planefor example, coast- another step in characterizing fractals through associated
lines, rivers, mountain profiles, earthquake faultlines, frac- numbers. How can the distribution of a fractals holes or
ture and cracking patterns, viscous fingering, dielectric gaps (lacunae) be quantified?
breakdown, growth of bacteria in stressed environments
box dimension is often relatively easy to compute. Select a
A. The Prefactor
sequence 1 > 2 > > n of sizes of boxes to be used to
cover the shape, and denote by N (i ) the number of boxes Suppose A is either carpet in Fig. 22, and let A de-
of size i needed to cover the shape. A plot of log(N (i )) note the -thickening of A. As mentioned in Section II.D,
against log(1/i ) often reveals a scaling range over which area(A ) 2log 40/ log 7 . One measure of lacunarity is
the points fall close to a straight line. In the presence of 1/ , if the appropriate limit exists.
other evidence (hierarchical visual complexity, for exam- It is well known that for the box dimension, the limit as
ple), this indicates a fractal structure with box dimension  0 can be replaced by the sequential limit n 0, for
given by the slope of the line. Interpreting the box di- n satisfying mild conditions. For these carpets, natural
mension in terms of underlying physical, chemical, and choices are those n just filling successive generations of
biological processes has yielded productive insights. holes. Applied to Fig. 22, these n give 1/ 0.707589
For physical objects in three-dimensional spacefor and 0.793487, agreeing with the notion that higher lacu-
example, aggregates, dustballs, physiological branch- narity corresponds to a more uneven distribution of holes.
ings (respiratory, circulatory, and neural), soot parti- Unfortunately, the prefactor is much more sensitive than
cles, protein clusters, terrain mapsit is often easier the exponent: different sequences of n give different lim-
to compute mass dimension. Select a sequence of radii its. Logarithmic averages can be used, but this is work in
r1 > r2 > > rn and cover the object with concentric progress.
spheres of those radii. Denoting by M(ri ) the mass of the
part of the object contained inside the sphere of radius ri ,
a plot of log(M(ri )) against log(ri ) often reveals a scaling B. The Crosscuts Structure
range over which the points fall close to a straight line. In
An object is often best studied through its crosscuts by
the presence of other evidence (hierarchical arrangements
straight lines, concentric circles, or spheres. For a fractal
of hole sizes, for example), this indicates a fractal struc-
of dimension d in the plane, the rule of thumb is that the
ture with mass dimension given by the slope of the line.
crosscuts are Cantor-like objects of dimension d 1. The
Mass dimension is relevant for calculating how density
case when the gaps between points in the crosscut are
scales with size, and this in turn has implications for how
statistically independent was singled out by Mandelbrot
the object is coupled to its environment.
as defining neutral lacunarity. If the crosscut is also self-
similar, it is a Levy dust.
Hovi et al. (1996) studied the intersection of lines (linear
VI. LACUNARITY crosscuts) with two- and three-dimensional critical per-
colation clusters, and found the gaps are close to being
Examples abound of fractals sharing the same dimension statistically independent, thus a Levy dust.
but looking quite different. For instance, both Sierpin- In studying very large DLA clusters, Mandelbrot et al.
ski carpets in Fig. 22 have dimension log 40/log 7. The (1995) obtained a crosscut dimension of dc = 0.65 0.01,
holes distribution is more uniform on the left than on different from the value 0.71 anticipated if DLA clus-
the right. The quantification of this difference was un- ters were statistically self-similar objects with mass di-
dertaken in Mandelbrot (1982, Chapter 34). It introduced mension dmass = 1.71. The difference can be explained by
lacunarity as one expression of this difference, and took asserting the number of particles Nc (r/l) on a crosscut
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

200 Fractals

of radius r scales as Nc (r/l) = (r )(r/l)dc . Here l is

the scaling length, and the lacunarity prefactor varies
with r . Assuming slow variation of (r ) with r , the
observed linear loglog fit requires (r ) r d , where
d = dmass 1 dc = 0.06 0.01. Transverse crosscut
analysis reveals lacunarity decreases with r for large DLA
clusters.
FIGURE 23 The effect of H on Weierstrass graph roughness. In
all pictures, b = 1.5 and H has the indicated value.

C. Antipodal Correlations
Select an occupied point p well inside a random fractal known was constructed by Weierstrass in 1872. The Weier-
cluster, so that the R R square centered at p lies within strass sine function is
the cluster. Now select two vectors V and W based at p 

and separated by the angle . Finally, denote by x and y the W (t) = bH n sin(2 bn t),
n=0
number of occupied sites within the wedges with apexes
at p, apex angles much less than , and centered about and the complex Weierstrass function is
the vectors V and W . The angular correlation function 

is W0 (t) = bH n exp(2ibn t).
x y xy n=0
C() = , Hardy (1916) showed W (t) is continuous and nowhere-
x 2  xx
differentiable if and only if b > 1 and 0 < H < 1.
where   denotes an average over many realizations of As shown in Fig. 23, the parameter H determines the
the random fractal. Antipodal correlations concern = . roughness of the graph. In this case, H is not a perspicuous
Negative and positive antipodal correlations are inter- roughness exponent. Indeed, as b increases, the ampli-
preted as indicating high and low lacunarity; vanishing tudes of the higher frequency terms decrease and the graph
correlation is a weakened form of neutral lacunarity. is more clearly dominated by the lowest frequency terms.
Mandelbrot and Stauffer (1994) used antipodal correla- This effect of b is a little-explored aspect of lacunarity.
tions to study the lacunarity of critical percolation clusters.
On smaller central subclusters, they found the antipodes
are uncorrelated. B. WeierstrassMandelbrot Functions
Trema random fractals. These are formed by remov- The Weierstrass function revolutionized mathematics but
ing randomly centered discs, tremas, with radii obeying a did not enter physics until it was modified in a series
power-law scaling. For them, C() 0 with because of steps described in Mandelbrot (1982, pp. 388390;
a circular hole that overlaps a sector cannot overlap the (2001d, Chapter H4). The step from W0 (t) to W1 (t) added
opposite sector. But nonconvex tremas introduce positive low frequencies in order to insure self-affinity. The step
antipodal correlations. For close to , needle-shaped from W1 (t) to W2 (t) added to each addend a random phase
tremas, though still convex, yield C() much higher than n uniformly distributed on [0, 1]. The step from W1 (t) to
for circular trema sets. From this more refined viewpoint, W3 (t) added a random amplitude An = 2 log V , where
needle tremas lacunarity is much lower. V is uniform on [0, 1]. A function W4 (t) that need not be
written down combines a phase and an amplitude. The lat-
est step leads to another function that need not be written
VII. FRACTAL GRAPHS down: it is W5 (t) = W4 (t) + W4 (t), where the two ad-
AND SELF-AFFINITY dends are statistically independent. Contrary to all earlier
extensions, W5 (t) is not chiral. We have
A. Weierstrass Functions 

W1 (t) = bH n (exp(2ibn t) 1),
Smooth functions graphs, as seen under sufficient mag- n=
nification, are approximated by their tangents. Unless the
function itself is linear, the existence of a tangent con- 

W2 (t) = bH n (exp(2ibn t) 1) exp(in ),
tradicts the scale invariance that characterizes fractals. n=
The early example of a continuous, nowhere-differentiable


function devised in 1834 by Bolzano remained unpub- W3 (t) = An bH n (exp(2ibn t) 1).
lished until the 1920s. The first example to become widely n=
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 201

C. The Holder Exponent the dynamics to a shift map on a Cantor set under general
conditions. In this sense, chaos often equivalent to simple
A function f: [a, b] R has Holder exponent H if there
dynamics on an underlying fractal.
is a constant c > 0 for which
| f (x) f (y)| c |x y| H
B. Fractal Basin Boundaries
for all x and y in [a, b] (recall Section III.G). If f is con- For any point c belonging to a hyperbolic component of the
tinuous and has Holder exponent H satisfying 0 < H 1, Mandelbrot set, the Julia set is the boundary of the basins
then the graph of f has box dimension dbox 2 H . of attraction of the attracting cycle and the attracting fixed
The Weierstrass function W (t) has Holder exponent point at infinity. See the right side of Fig. 3.
H , hence its graph has dbox 2 H . For large enough Another example favored by Julia is found in New-
b, dbox = 2 H , so one can think of the Holder exponent tons method for finding the roots of a polynomial f (z)
as a measure of roughness of the graph. of degree at least 3. It leads to the dynamical system
z n+1 = N f (z n ) = z n f (z n )/ f  (z n ). The roots of f (z) are
attracting fixed points of N f (z), and the boundary of the
VIII. FRACTAL ATTRACTORS AND basins of attraction of these fixed points is a fractal; an ex-
REPELLERS OF DYNAMICAL ample is shown on the left side of Fig. 24. If contaminated
SYSTEMS by even small uncertainties, the fate of initial points near
the basin boundary cannot be predicted. Sensitive depen-
The modern renaissance in dynamical systems is asso- dence on initial conditions is a signature of chaos, but here
ciated most often with chaos theory. Consequently, the we deal with something different. The eventual behavior
relations between fractal geometry and chaotic dynam- is completely predictable, except for initial points taken
ics, mediated by symbolic dynamics, are relevant to our exactly on the basin boundary, usually of two-dimensional
discussion. In addition, we consider fractal basin bound- Lebesgue measure 0.
aries, which generalize Julia sets to much wider contexts The same complication enters mechanical engineer-
including mechanical systems. ing problems for systems with multiple attractors. Moon
(1984) exhibited an early example. Extensive theoretical
and computer studies by Yorke and coworkers are de-
A. The Smale Horseshoe
scribed in Alligood and Yorke (1992). The driven har-
If they exist, intersections of the stable and unstable monic oscillator with two-well potential
manifolds of a fixed point are called homoclinic points.
d2x dx 1
Poincare (1890) recognized that homoclinic points cause + f x(1 x 2 ) = A cos(t)
great complications in dynamics. Yet much can be un- dt 2 dt 2
derstood by labeling an appropriate coarse-graining of a is a simple example. The undriven system has two
neighborhood of a homoclinic point and translating the equilibria, x = 1 and x = +1. Initial values (x , x  ) are
corresponding dynamics into a string of symbols (the painted white if the trajectory from that point eventually
coarse-grain bin labels). The notion of symbolic dynam- stays in the left basin, black if it eventually stays in the right
ics first appears in Hadamard (1898), and Birkhoff (1927) basin. The right side of Fig. 24 shows the initial condition
proved every neighborhood of a homoclinic point contains portrait for the system with f = 0.15, = 0.8, and A =
infinitely many periodic points. 0.094.
Motivated by work of Cartwright and Littlewood (1945)
and Levinson (1949) on the forced van der Pol oscillator,
Smale (1963) constructed the horseshoe map. This is a IX. FRACTALS AND DIFFERENTIAL OR
map from the unit square into the plane with completely PARTIAL DIFFERENTIAL EQUATIONS
invariant set a Cantor set , roughly the Cartesian product
of two Cantor middle-thirds sets. Restricted to , with the The daunting task to which a large portion of Mandelbrot
obvious symbolic dynamics encoding, the horseshoe map (1982) is devoted was to establish that many works of
is conjugate to the shift map on two symbols, the archetype nature and man [as shown in Mandelbrot (1997), the lat-
of a chaotic map. ter includes the stock market!] are fractal. New and often
This construction is universal in the sense that it oc- important examples keep being discovered, but the hard-
curs in every transverse homoclinic point to a hyperbolic est present challenge is to discover the causes of fractal-
saddle point. The ConleyMoser theorem (see Wiggins, ity. Some cases remain obscure, but others are reasonably
1990) establishes the existence of chaos by conjugating clear.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

202 Fractals

FIGURE 24 Left: The basins of attraction of Newtons method for finding the roots of z 3 1. Right: The basins of
attraction for a damped, driven two-well harmonic oscillator.

Thus, the fractality of the physical percolation clusters nection with lasers. The sensitivity to initial conditions
(Section I.B.3.e) is the geometric counterpart of scaling common to chaotic dynamics is mediated by the intricate
and renormalization: the analytic properties of those ob- fractal interleaving of the multiple layers of the attractor.
jects follow a wealth of power-law relations. Many math- In addition, Birman and Williams (1983) showed an abun-
ematical issues, some of them already mentioned, remain dance of knotted periodic orbits embedded in the Lorenz
open, but the overall renormalization framework is firmly attractor, though Williams (1983) showed all such knots
rooted. Renormalization and the resulting fractality also are prime. Grist (1997) constructed a universal template,
occur in the structure of attractors and repellers of dy- a branched 2-manifold in which all knots are embedded.
namical systems. Best understood is renormalization for Note the interesting parallel with the universal aspects of
quadratic maps. Feigenbaum and others considered the the Sierpinski carpet (Section I.B.1.a). It is not yet known
real case. For the complex case, renormalization estab- if the attractor of any differential equation contains a uni-
lishes that the Mandelbrot set contains infinitely many versal template. The PoincareBendixson theorem pro-
small copies of itself. hibits fractal attractors for differential equations in the
Unfortunately, additional examples of fractality proved plane, but many other classical ordinary differential equa-
to be beyond the scope of the usual renormalization. A tions in at least three dimensions exhibit similar fractal
notorious case concerns DLA (Section I.B.3.f). attractors in certain parameter ranges.

A. Fractal Attractors of Ordinary

Differential Equations B. Partial Differential Equations on Domains
with Fractal Boundaries (Can One Hear
The Lorenz equations for fluid convection in a two- the Shape of a Fractal Drum?)
dimensional layer heated from below are
Suppose D Rn is an open region with boundary D.
dx dy dz Further, suppose the eigenvalue problem 2 u = u
= (y x), = x z +r x y, = x y bz.
dt dt dt with boundary conditions u(x) = 0 for all x D has
Here x denotes the rate of convective overturning, y the real eigenvalues 0 < 1 < 2 < . For D with suffi-
horizontal temperature difference, and z the departure ciently smooth boundary, a theorem of Weyl (1912)
from a linear vertical temperature gradient. For the pa- shows N () n/2 , where the eigenvalue counting func-
rameters = 10, b = 8/3, and r = 28, Lorenz (1963) sug- tion N () = {the number of i for which i }. If the
gested that trajectories in a bounded region converge to boundary D is a fractal, Berry (1979, pp. 5153) postu-
an attractor that is a fractal, with dimension about 2.06, lated that some form of the dimension of D appears in
as estimated by Liapunov exponents. The Lorenz equa- the second term in the expansion of N (), therefore can
tions are very suggestive but do not represent weather be recovered from the eigenvalues. This could not be the
systems very well. However, Haken established a con- Hausdorff dimension, but Lapidus (1995) showed that it
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 203

With this Laplacian, the heat and wave equations can be

defined on fractals. Among other things, the wave equation
on domains with fractal boundaries admits localized solu-
tions, as we saw for the wave equation on fractal drums. A
major challenge is to extend these ideas to fractals more
FIGURE 25 Perimeter of an extensively studied fractal drum. complicated than the Sierpinski gasket and its relatvives.

is the MinkowskiBouligard dimension. The analysis is D. How Partial Differential Equations

subtle, involving some deep number theory. Generate Fractals
For regions with fractal boundaries, the heat equation A quandary: It is universally granted that physics is ruled
2 u = (/t)u shows heat flow across a fractal bound- by diverse partial differential equations, such as those of
ary is related to the dimension of the boundary. Sapoval Laplace, Poisson, and NavierStokes. A differential equa-
(1989) and Sapoval et al. (1991) conducted elegant exper- tion necessarily implies a great degree of local smooth-
imenths to study the modes of fractal drums. The perimeter ness, even though close examination shows isolated
of Fig. 25 has dimension log 8/log 4 = 3/2. A membrane singularities or catastrophes. To the contrary, fractal-
stretched across this fractal curve was excited acoustically ity implies everywhere-dense roughness or fragmentation.
and the resulting modes observed by sprinkling powder This is one of the several reasons that fractal models in di-
on the membrane and shining laser light transverse to the verse fields were initially perceived as being anomalies
surface. Sapoval observed modes localized to bounded contradicting one of the firmest foundations of science.
regions A , B , C , and D shown in Fig. 25. By carefully A conjecturechallenge responding to the preceding
displacing the acoustic source, he was able to excite each quandry. There is no contradiction at all: fractals arise
separately. unavoidably in the long-time behavior of the solution of
Theoretical and computer-graphic analyses of the wave very familiar and innocuous-looking equations. In partic-
equation on domains with fractal boundaries have been ular, many concrete situations where fractals are observed
carried out by Lapidus et al. (1996), among others. involve equations that allow free and moving boundaries,
interfaces, or singularities. As a suggestive principle,
C. Partial Differential Equations on Fractals Mandelbrot (1982, Chapter 11) described the following
possibility: under broad conditions that largely remain to
The problem is complicated by the fact that a fractal is not be specified, these free boundaries, interfaces, and singu-
a smooth manifold. How is the Laplacian to be defined on larities converge to suitable fractals. Many equations have
such a space? One promising approach was put forward been examined from this viewpoint, but we limit ourselves
by physicists in the 1980s and made rigorous in Kigami to two examples of central importance.
(1989): approximate the fractal domain by a sequence of
graphs representing successive protofractals, and define
the fractal Laplacian as the limit of a suitably renormalized 1. The Large-Scale Distribution of Galaxies
sequence of Laplacians on the graphs. Figure 26 shows the Chapters 9 and 3335 of Mandelbrot (1982) conjecture
first four graphs for the equilateral Sierpinski gasket. The that the distribution of galaxies is fractal. This conjec-
values at the boundary vertices are specified by the bound- ture results from a search for invariants that was central
ary conditions at any nonboundary vertex x0 . The mth to every aspect of the construction of fractal geometry.
approximate Laplacian of a function f (x) is the product Granted that the distribution of galaxies certainly deviates
of a renormalization factor by ( f (y) f (x0 )), where from homogeneity, one broad approach consists in correct-
the sum is taken over all vertices y in the mth protofrac- ing for local inhomogeneity by using local patches. The
tal graph corresponding to the mth-stage reduction of the next simplest global assumption is that the distribution is
whole graph. nonhomogeneous but scale invariant, therefore fractal.
Excluding the strict hierarchies, two concrete construc-
tions of random fractal sets were subjected to detailed
mathematical and visual investigation. These construc-
tions being random, self-similarity can only be statistical.
But a strong counteracting asset is that the self-similarity
ratio can be chosen freely, is not restricted to powers of a
FIGURE 26 Graphs corresponding to protofractal approxima- prescribed r0 . A surprising and noteworthy finding came
tions of the equilateral Sierpinski gasket. forth. These constructions exhibited a strong hierarchical
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

204 Fractals

structure that is not a deliberate and largely arbitrary input. extended to turbulence, and circa 1964 led to the following
Details are given in Mandelbrot (1982). conjecture.
The first construction is the seeded universe based on Conjecture. The property of being turbulently dissipa-
a Levy flight. Its Hausdorff-dimensional properties were tive should not be viewed as attached to domains in a fluid
well known. Its correlation properties (Mandelbrot 1975) with significant interior points, but as attached to fractal
proved to be nearly identical to those of actual galaxy sets. In a first approximation, those sets intersection with
maps. The second construction is the parted universe a straight line is a Cantor-like fractal dust having a dimen-
obtained by subtracting from space a random collection sion in the range from 0.5 to 0.6. The corresponding full
of overlapping tremas. Either construction yields sets that sets in space should therefore be expected to be fractals
are highly irregular and involve no special center, yet, with with Hausdorff dimension in the range from 2.5 to 2.6.
no deliberate design, exhibit a clear-cut clustering, fila- Actually, Cantor dust and Hausdorff dimension are not
ments and walls. These structures were little known the proper notions in the context of viscous fluids because
when these constructions were designed. viscosity necessarily erases the fine detail essential to frac-
Conjecture: Could it be that the observed clusters, tals. Hence the following conjecture (Mandelbrot, 1982,
filaments, and walls need not be explained separately? Chapter 11; 1976). The dissipation in a viscous fluid oc-
They may not result from unidentified features of spe- curs in the neighborhood of a singularity of a nonviscous
cific models, but represent unavoidable consequences of approximation following Eulers equations, and the mo-
a variety of unconstrained forms of random fractality, as tion of a nonviscous fluid acquires singularities that are
interpreted by a human brain. sets of dimension about 2.52.6. Several numerical tests
A problem arose when careful examination of the sim- agree with this conjecture (e.g., Chorin, 1981).
ulations revealed a clearly incorrect prediction. The sim- A related conjecture, that the NavierStokes equations
ulations in the seeded universe proved to be visually far have fractal singularities of much smaller dimension, has
more lacunar than the real world. That is, the simula- led to extensive work by V. Scheffer, R. Teman, and
tions holes are larger than in reality. The parted universe C. Foias, and many others. But this topic is not exhausted.
model fared better, since its lacunarity can be adjusted at Finally, we mention that fractals in phase space entered
will and fit to the actual distribution. A lowered lacunarity the transition from laminar to turbulent flow through the
is expressed by a positive correlation between masses in work of Ruelle and Takens (1971) and their followers. The
antipodal directions. Testing this specific conjecture is a task of unifying the real- and phase-space roles of fractals
challenge for those who analyze the data. is challenging and far from being completed.
Does dynamics make us expect the distribution of galax-
ies to be fractal? Position a large array of point masses
in a cubic box in which opposite sides are identified to X. FRACTALS IN THE ARTS
form a three-dimensional torus. The evolution of this ar- AND IN TEACHING
ray obeys the Laplace equation, with the novelty that the
singularities of the solution are the positions of the points, The Greeks asserted art reflects nature, so it is little sur-
therefore movable. All simulations we know (starting with prise that the many fractal aspects of nature should find
those performed at IBM around 1960) suggest that, even their way into the artsbeyond the fact that a representa-
when the pattern of the singularities begins by being uni- tional painting of a tree exhibits the same fractal branching
form or Poisson, it gradually creates clusters and a sem- as a physical tree. Voss and Clarke (1975) found fractal
blance of hierarchy, and appears to tend toward fractality. It power-law scaling in music, and self-similarity is designed
is against the preceding background that the limit distribu- in the music of the composers Gyorgy Ligeti and Charles
tion of galaxies is conjectured to be fractal, and fractality Wuorinen. Pollard-Gott (1986) established the presence of
is viewed as compatible with Newtons equations. fractal repetition patterns in the poetry of Wallace Stevens.
Computer artists use fractals to create both abstract aes-
thetic images and realistic landscapes. Larry Poons paint-
2. The NavierStokes Equation
ings since the 1980s have had rich fractal textures. The
The first concrete use of a Cantor dust in real spaces is decalcomania of the 1830s and the 1930s and 1940s
found in Berger and Mandelbrot (1963), a paper on noise used viscous fingering to provide a level of visual com-
records. This was nearly simultaneous with Kolmogorovs plexity. Before that, Giacomettis Alpine wildflower paint-
work on the intermittence of turbulence. After numerous ings are unquestionably fractal. Earlier still, relatives of the
experimental tests designed to create an intuitive feeling Sierpinski gasket occur as decorative motifs in Islamic and
for this phenomenon (e.g., listening to turbulent velocity Renaissance art. Fractals abound in architecture, for exam-
records that were made audible), the fractal viewpoint was ple, in the cascades of spires in Indian temples, Bramantes
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 205

plan for St. Peters, Malevichs Architektonics, and some SEE ALSO THE FOLLOWING ARTICLES
of Frank Lloyd Wrights designs. Fractals occur in the
writing of Clarke, Crichton, Hoag, Powers, Updike, and CHAOS PERCOLATION TECTONOPHYSICS
Wilhelm, among others, and in at least one play, Stoppards
Arcadia. Postmodern literary theory has used some con- BIBLIOGRAPHY
cepts informed by fractal geometry, though this applica-
tion has been criticized for its overly free interpretations Alligood, K., and Yorke, J. (1992). Ergodic Theory Dynam. Syst. 12,
of precise scientific language. Some have seen evidence 377400.
of power-law scaling in historical records, the distribu- Alligood, K., Sauer, T., and Yorke, J. (1997). Chaos. An Introduction
tion of the magnitudes of wars and of natural disasters, to Dynamical Systems, Springer-Verlag, New York.
Barnsley, M. (1988). Fractals Everywhere, 2nd ed., Academic Press,
for example. In popular culture, fractals have appeared Orlando, FL.
on t-shirts, totebags, book covers, MTV logos, been men- Barnsley, M., and Demko, S. (1986). Chaotic Dynamics and Fractals,
tioned on public radios A Prairie Home Companion, and Academic Press, Orlando, FL.
been seen on television programs from Nova and Murphy Barnsley, M., and Hurd, L. (1993). Fractal Image Compression, Peters,
Brown, through several incarnations of Star Trek, to The X- Wellesley, MA.
Batty, M., and Longley, P. (1994). Fractal Cities, Academic Press,
Files and The Simpsons. While Barnsleys (1988) slogan, London.
fractals everywhere, is too strong, the degree to which Beardon, A. (1983). The Geometry of Discrete Groups, Springer-
fractals surround us outside of science and engineering is Verlag, New York.
striking. Beck, C., and Schlogl, F. (1993). Thermodynamics of Chaotic Systems:
A corollary of this last point is a good conclusion to An Introduction, Cambridge University Press, Cambridge.
Berger, J., and Mandelbrot, B. (1963). IBM J. Res. Dev. 7, 224236.
this high-speed survey. In our increasingly technologi- Berry, M. (1979). Structural Stability in Physics, Springer-Verlag,
cal world, science education is very important. Yet all New York.
too often humanities students are presented with limited Bertoin, J. (1996). Levy Processes, Cambridge University Press,
choices: the first course in a standard introductory se- Cambridge.
quence, or a survey course diluted to the level of journal- Besicovitch, A., and Moran, P. (1945). J. Lond. Math. Soc. 20, 110120.
Birkhoff, G. (1927). Dynamical Systems, American Mathematical
ism. The former builds toward major points not revealed Society, Providence, RI.
until later courses, the latter discusses results from science Birman, J., and Williams, R. (1983). Topology 22, 4782.
without showing how science is done. In addition, many Bishop, C., and Jones, P. (1997). Acta Math. 179, 139.
efforts to incorporate computer-aided instruction attempt Blanchard, P. (1994). In Complex Dynamical Systems. The Mathemat-
to replace parts of standard lectures rather than engage ics behind the Mandelbrot and Julia Sets (Devaney, R., ed.), (pp.
139154, American Mathematical Society, Providence, RI.
students in exploration and discovery. Bochner, S. (1955). Harmonic Analysis and the Theory of Probability,
Basic fractal geometry courses for non-science students University of California Press, Berkeley, CA.
provide a radical departure from this mode. The subject Bowen, R. (1975). Equilibrium States and the Ergodic Theory of
of fractal geometry operates at human scale. Though new Anosov Diffeomorphisms, Springer-Verlag, New York.
to most, the notion of self-similarity is easy to grasp, Bunde, A., and Havlin, S. (1991). Fractals and Disordered Systems,
Springer-Verlag, New York.
and (once understood) handles familiar objects from a Cartwright, M., and Littlewood, L. (1945). J. Lond. Math. Soc. 20, 180
genuinely novel perspective. Students can explore frac- 189.
tals with the aid of readily available software. These in- Cherbit, G. (1987). Fractals. Non-integral Dimensions and Applica-
stances of computer-aided instruction are perfectly nat- tions, Wiley, Chichester, UK.
ural because computers are so central to the entire field Chorin, J. (1981). Commun. Pure Appl. Math. 34, 853866.
Crilly, A., Earnshaw, R., and Jones, H. (1991). Fractals and Chaos,
of fractal geometry. The contemporary nature of the field Springer-Verlag, New York.
is revealed by a supply of mathematical problems that Crilly, A., Earnshaw, R., and Jones, H. (1993). Applications of Fractals
are simple to state but remain unsolved. Altogether, many and Chaos, Springer-Verlag, New York.
fields of interest to non-science students have surprising Curry, J., Garnett, L., and Sullivan, D. (1983). Commun. Math. Phys. 91,
examples of fractal structures. Fractal geometry is a pow- 267277.
Dekking, F. M. (1982). Adv. Math. 44, 78104.
erful tool for imparting to non-science students some of Devaney, R. (1989). An Introduction to Chaotic Dynamical Systems,
the excitement for science often invisible to them. Sev- 2nd ed., Addison-Wesley, Reading, MA.
eral views of this are presented in Frame and Mandelbrot Devaney, R. (1990). Chaos, Fractals, and Dynamics. Computer Exper-
(2001). iments in Mathematics, Addison-Wesley, Reading, MA.
The importance of fractals in the practice of science and Devaney, R. (1992). A First Course in Chaotic Dynamical Systems.
Theory and Experiment, Addison-Wesley, Reading, MA.
engineering is undeniable. But fractals are also a proven Devaney, R. (ed.). (1994). Complex Dynamical Systems. The Mathe-
force in science education. Certainly, the boundaries of matics Behind the Mandelbrot and Julia Sets, American Mathemati-
fractal geometry have not yet been reached. cal Society, Providence, RI.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

206 Fractals

Devaney, R., and Keen, L. (1989). Chaos and Fractals. The Mathemat- Lapidus, M., Neuberger, J., Renka, R., and Griffith, C. (1996). Int. J.
ics Behind the Computer Graphics, American Mathematical Society, Bifurcation Chaos 6, 11851210.
Providence, RI. Lasota, A., and Mackey, M. (1994). Chaos, Fractals, and Noise. Stochas-
Douady, A., and Hubbard, J. (1984). Etude dynamique des polynomes tic Aspects of Dynamics, 2nd ed., Springer-Verlag, New York.
complexes. I, II, Publications Mathematiques dOrsay, Orsay, Lawler, G., Schramm, O., and Warner, W. (2000). Acta Math., to appear
France. [xxx.lanl.gov./abs/math. PR/0010165].
Douady, A., and Hubbard, J. (1985). Ann. Sci. Ecole Norm. Sup. 18, Lei, T. (2000). The Mandelbrot Set, Theme and Variations, Cambridge
287343. University Press, Cambridge.
Edgar, G. (1990). Measure, Topology, and Fractal Geometry, Springer- Le Mehaute, A. (1990). Fractal Geometries. Theory and Applications,
Verlag, New York. CRC Press, Boca Raton, FL.
Edgar, G. (1993). Classics on Fractals, Addison-Wesley, Reading, MA. Levinson, N. (1949). Ann. Math. 50, 127153.
Edgar, G. (1998). Integral, Probability, and Fractal Measures, Springer- Lorenz, E. (1963). J. Atmos. Sci. 20, 130141.
Verlag, New York. Lu, N. (1997). Fractal Imaging, Academic Press, San Diego.
Eglash, R. (1999). African Fractals. Modern Computing and Indigenous Lyubich, M. (2001). Ann. Math., to appear.
Design, Rutgers University Press, New Brunswick, NJ. Mandelbrot, B. (1975). C. R. Acad. Sci. Paris 280A, 10751078.
Encarncacao, J., Peitgen, H.-O., Sakas, G., and Englert, G. (1992). Frac- Mandelbrot, B. (1975, 1984, 1989, 1995). Les objects fractals, Flam-
tal Geometry and Computer Graphics, Springer-Verlag, New York. marion, Paris.
Epstein, D. (1986). Low-Dimensional Topology and Kleinian Groups, Mandelbrot, B. (1976). C. R. Acad. Sci. Paris 282A, 119120.
Cambridge University Press, Cambridge. Mandelbrot, B. (1980). Ann. N. Y. Acad. Sci. 357, 249259.
Evertsz, C., Peitgen, H.-O., and Voss, R. (eds.). (1996). Fractal Geom- Mandelbrot, B. (1982). The Fractal Geometry of Nature, Freeman,
etry and Analysis. The Mandelbrot Festschrift, Curacao 1995, World New York.
Scientific, Singapore. Mandelbrot, B. (1984). J. Stat. Phys. 34, 895930.
Falconer, K. (1985). The Geometry of Fractal Sets, Cambridge Uni- Mandelbrot, B. (1985). In Chaos, Fractals, and Dynamics (Fischer, P.,
versity Press, Cambridge. and Smith, W., eds.), pp. 235238, Marcel Dekker, New York.
Falconer, K. (1987). Math. Intelligencer 9, 2427. Mandelbrot, B. (1995). J. Fourier Anal. Appl. 1995, 409432.
Falconer, K. (1990). Fractal Geometry. Mathematical Foundations and Mandelbrot, B. (1997a). Fractals and Scaling in Finance. Discontinuity,
Applications, Wiley, Chichester, UK. Concentration, Risk, Springer-Verlag, New York.
Falconer, K. (1997). Techniques in Fractal Geometry, Wiley, Chich- Mandelbrot, B. (1997b). Fractales, Hasard et Finance, Flammarion,
ester, UK. Paris.
Family, F., and Vicsek, T. (1991). Dynamics of Fractal Surfaces, World Mandelbrot, B. (1999). Multifractals and 1/ f Noise. Wild Self-Affinity
Scientific, Singapore. in Physics, Springer-Verlag, New York.
Feder, J. (1988). Fractals, Plenum Press, New York. Mandelbrot, B. (2001a). Quant. Finance 1, 113123, 124130.
Feder, J., and Aharony, A. (1990). Fractals in Physics. Essays in Honor Mandelbrot, B. (2001b). Gaussian Self-Affinity and Fractals: Globality,
of B. B. Mandelbrot, North-Holland, Amsterdam. the Earth, 1/ f Noise, & R/S, Springer-Verlag, New York.
Fisher, Y. (1995). Fractal Image Compression. Theory and Application, Mandelbrot, B. (2001c). Fractals and Chaos and Statistical Physics,
Springer-Verlag, New York. Springer-Verlag, New York.
Flake, G. (1998). The Computational Beauty of Nature. Computer Ex- Mandelbrot, B. (2001d). Fractals Tools, Springer-Verlag, New York.
plorations of Fractals, Chaos, Complex Systems, and Adaptation, Mandelbrot, B. B., and Stauffer, D. (1994). J. Phys. A 27, L237L242.
MIT Press, Cambridge, MA. Mandelbrot, B. B., Vespignani, A., and Kaufman, H. (1995). Europhy.
Fleischmann, M., Tildesley, D., and Ball, R. (1990). Fractals in the Lett. 32, 199204.
Natural Sciences, Princeton University Press, Princeton, NJ. Maksit, B. (1988). Kleinian Groups, Springer-Verlag, New York.
Frame, M., and Mandelbrot, B. (2001). Fractals, Graphics, and Massopust, P. (1994). Fractal Functions, Fractal Surfaces, and
Mathematics Education, Mathematical Association of America, Wavelets, Academic Press, San Diego, CA.
Washington, DC. Mattila, P. (1995). Geometry of Sets and Measures in Euclidean Space.
Frostman, O. (1935). Meddel. Lunds. Univ. Math. Sem. 3, 1118. Fractals and Rectifiability, Cambridge University Press, Cambridge.
Gazale, M., (1990). Gnomon. From Pharaohs to Fractals, Princeton McCauley, J. (1993). Chaos, Dynamics and Fractals. An Algorith-
University Press, Princeton, NJ. mic Approach to Deterministic Chaos, Cambridge University Press,
Grist, R. (1997). Topology 36, 423448. Cambridge.
Gulick, D. (1992). Encounters with Chaos, McGraw-Hill, New York. McLaughlin, J. (1987). Proc. Am. Math. Soc. 100, 183186.
Hadamard, J. (1898). J. Mathematiques 5, 2773. McMullen, C. (1994). Complex Dynamics and Renormalization,
Hardy, G. (1916). Trans. Am. Math. Soc. 17, 322323. Princeton University Press, Princeton, NJ.
Hastings, H., and Sugihara, G. (1993). Fractals. A Users Guide for the McShane, G., Parker, J., and Redfern, I. (1994). Exp. Math. 3, 153170.
Natural Sciences, Oxford University Press, Oxford. Meakin, P. (1996). Fractals, Scaling and Growth Far from Equilibrium,
Hovi, J.-P., Aharony, A., Stauffer, D., and Mandelbrot, B. B. (1996). Cambridge University Press, Cambridge.
Phys. Rev. Lett. 77, 877880. Milnor, J. (1989). In Computers in Geometry and Topology (Tangora,
Hutchinson, J. E. (1981). Ind. Univ. J. Math. 30, 713747. M., ed.), pp. 211257, Marcel Dekker, New York.
Keen, L. (1994). In Complex Dynamical Systems. The Mathematics Moon, F. (1984). Phys. Rev. Lett. 53, 962964.
behind the Mandelbrot and Julia Sets (Devaney, R., ed.), pp. 139 Moon, F. (1992). Chaotic and Fractal Dynamics. An Introduction for
154, American Mathematical Society, Providence, RI. Applied Scientists and Engineers, Wiley-Interscience, New York.
Keen, L., and Series, C. (1993). Topology 32, 719749. Parker, J. (1995). Topology 34, 489496.
Keen, L., Maskit, B., and Series, C. (1993). J. Reine Angew. Math. 436, Peak, D., and Frame, M. (1994). Chaos Under Control. The Art and
209219. Science of Complexity, Freeman, New York.
Kigami, J. (1989). Jpn. J. Appl. Math. 8, 259290. Peitgen, H.-O. (1989). Newtons Method and Dynamical Systems,
Lapidus, M. (1995). Fractals 3, 725736. Kluwer, Dordrecht.
P1: GLQ Final
Encyclopedia of Physical Science and Technology EN006H-259 June 28, 2001 20:1

Fractals 207

Peitgen, H.-O., and Richter, P. H. (1986). The Beauty of Fractals, Shlesinger, M., Zaslavsky, G., and Frisch, U. (1995). Levy Flights and
Springer-Verlag, New York. Related Topics in Physics, Springer-Verlag, New York.
Peitgen, H.-O., and Saupe, D. (1988). The Science of Fractal Images, Sinai, Y. (1972). Russ. Math. Surv. 27, 2170.
Plenum Press, New York. Smale, S. (1963). In Differential and Combinatorial Topology
Peitgen, H.-O., Jurgens, H., and Saupe, D. (1992). Chaos and Fractals: (Cairns, S., ed.), pp. 6380, Princeton University Press, Princeton,
New Frontiers of Science, Springer-Verlag, New York. NJ.
Peitgen, H.-O., Rodenhausen, A., and Skordev, G. (1998). Fractals 6, Stauffer, D., and Aharony, A. (1992). Introduction to Percolation The-
371394. ory, 2nd ed., Taylor and Francis, London.
Pietronero, L. (1989). Fractals Physical Origins and Properties, North- Strogatz, S. (1994). Nonlinear Dynamics and Chaos, with Applica-
Holland, Amsterdam. tions to Chemistry, Physics, Biology, Chemistry, and Engineering,
Pietronero, L., and Tosatti, E. (1986). Fractals in Physics, North- Addison-Wesley, Reading, MA.
Holland, Amsterdam. Sullivan, D. (1985). Ann. Math. 122, 410418.
Poincare, H. (1890). Acta Math. 13, 1271. Tan, Lei (1984). In Etude dynamique des polynomes complexes
Pollard-Gott, L. (1986). Language Style 18, 233249. (Douardy, A., and Hubbard, J., eds.), Vol. II, pp. 139152, Publi-
Rogers, C. (1970). Hausdorff Measures, Cambridge University Press, cations Mathematiques dOrsay, Orsay, France.
Cambridge. Tan, Lei. (2000). The Mandelbrot Set, Theme and Variations,
Ruelle, D. (1978). Thermodynamic Formalism: The Mathematical Cambridge University Press, Cambridge.
Structures of Classical Equilibrium Statistical Mechanics, Addison- Thurston, W. (1997). Three-Dimensional Geometry and Topology,
Wesley, Reading, MA. Princeton University Press, Princeton, NJ.
Ruelle, D., and Takens, F. (1971). Commun. Math. Phys. 20, 167192. Tricot, C. (1982). Math. Proc. Camb. Philos. Soc. 91, 5474.
Samorodnitsky, G., and Taqqu, M. (1994). Stable Non-Gaussian Ran- Vicsek, T. (1992). Fractal Growth Phenomena, 2nd ed., World Scien-
dom Processes. Stochastic Models with Infinite Variance, Chapman tific, Singapore.
and Hall, New York. Voss, R., and Clarke, J. (1975). Nature 258, 317318.
Sapoval, B. (1989). Physica D 38, 296298. West, B. (1990). Fractal Physiology and Chaos in Medicine, World
Sapoval, B., Rosso, M., and Gouyet, J. (1985). J. Phys. Lett. 46, L149 Scientific, Singapore.
L156. Weyl, H. (1912). Math. Ann. 71, 441479.
Sapoval, B., Gobron, T., and Margolina, A. (1991). Phys. Rev. Lett. 67, Williams, R. (1983). Ergodic Theory Dynam. Syst. 4, 147163.
29742977. Wiggins, S. (1990). Introduction to Applied Nonlinear Dynamical Sys-
Scholz, C., and Mandelbrot, B. (1989). Fractals in Geophysics, tems and Chaos, Springer-Verlag, New York.
Birkhauser, Basel. Witten, T., and Sander, L. (1981). Phys. Rev. Lett. 47, 14001403.
Shishikura, M. M. (1994). Asterisque 222, 389406. Witten, T., and Sander, L. (1983). Phys. Rev. B 27, 56865697.