Professional Documents
Culture Documents
This content has been downloaded from IOPscience. Please scroll down to see the full text.
(http://iopscience.iop.org/1402-4896/32/4/001)
View the table of contents for this issue, or go to the journal homepage for more
Download details:
IP Address: 93.180.53.211
This content was downloaded on 16/11/2013 at 10:31
Physica Scripta 32
258 Benoit B. Mandelbrot
to have a variant of BH(t) whose rigorous study is elementary; because they lack generality. For self-similar sets, the values
this led me recently to introduce a series of variants of BH(t) yielded by these dimensions were identical. Now that we move
that relate to a random walk one can interpolate. on to self-affine shapes, we shall find that local and global values
First, let me describe the essential properties. MH(t) is must be distinguished for each dimension, and that the different
defined when H is of the form H = log b’/log b“, the integer local values cease to be identical.
bases b’ and b” being such that b’ - b“ is positive and even. The The reason is fundamental: “square,” “distance,” and
idea is that the function BH(t), whose increments are Gaussian “circle,” are vital notions in “isotropic” geometry, but they are
over all At, is replaced by a function MH(t), whose increments meaningless in affine geometry. More precisely, they are mean-
over suitable At’s are binomial with the same mean 0 and the ingful for relief cross-sections, but are meaningless for noises,
same standard deviation. The requirement is that because the units along the t axis and along the B-axis are set
up independently of each other, hence A t and AB cannot be
MH(pb’-’) -MH [(p + l)b’-’] = f (b“)-’ = f (At)H
combined. There being no intrinsic meaning to the notion of
for all k and p . Thus, MH(pb”-’) is a multiple of b”-’. The equal height and width, a square cannot be defined. Similarly, a
linear interpolation between these values is a k-th approximant circle cannot be defined, because its square radius R 2 = A t 2 +
ofMH(t), to be denoted by MH’)(t). AB2 would have to combine the units along both axes. Further-
more, one cannot “walk a compass’ along a self-affine curve,
Actual Construction of M H ( t ) because the distance covered by each step combines a A t and a
The details do not really matter here, but are an interesting AB.
fresh example of the “multiplicative chaos” procedure I had On the other hand, a relief cross-section (while self-affine) is
pioneered in 1972 and in 1974 [ l , pages 278 ff.] . (The earlier a curve in an isotropic plane. And a noise record’s purely affine
uses of this procedure generate certain fractal measures of great plane is always represented on the same graph paper as an
current importance that have been rediscovered in part by I. isotropic plane. This causes the above distinction to be elusive,
Procaccia.) The building blocks are “multiplicative effect and creates the temptation to draw circles and squares to walk
functions” p k ( t ) defined as follows. For all k and t, I/.lk(t)l= compasses, and to evaluate various “prohibited” dimensions
b’/b“, and each interval between the sucessive integers of the “mechanically.” Sections 5 and 6 describe the results thus
form p to p + 1 splits into b’ subintervals; in (1/2) (b‘ + b”) obtained .
subintervals, chosen at random in each interval separately from
the other, one sets & ( t ) > 0 , and in the remaining (1/2) 4. The box dimension is meaningful for the records of B H ( t )
(b’ - b”) subintervals, one sets & ( t ) < 0.This insures that, for and M H ( ~ )and
, its local value is the “correct” 2-H. Its global
all integers k and p , value is 1
After a lattice made of boxes of side r = l / b is made to cover
a set, let N(b) denote the number of boxes in this lattice than
intersect the set. “Box dimension” is my present term for a
For example, if b ’ = 4 and b “ = 2, & ( t ) < 0 over one of notion that applies to sets for which N(b) behaves like
four subintervals. See fig. 1 for illustration. Now pick N(b) a bDB. What does this mean? It may (as it does for most
statistically independent functions p h ( t ) , and form mathematicians) refer to local behavior, and mean that limb -+-
logN(b)/log b = D,. When the set is bounded, one begins by
M”)(t) = n =n‘ - m
pn(b’”t), drawing it within a unit square of the plane. When the set is
unbounded, one considers bounded portions obtained as
then integrate to obtain the approximant intersections with squ ares.
t
f
M”)(t) = M’(’)(A)ds The box argument for the records of B(t) or B H ( t ) , as given
’0 in [ l , bottom left of page 2371 is heuristic and is not readily
and finally take the limit: M ( t ) = limk + ,Mh(t). Though this made rigorous. For the record of M H ( t ) , to the contrary, the
limit depends on b’ and b”, the present discussion only involves exact argument is transparent: To cover our fractal from t = 0
the value of H , hence the notation MH(t). to t = 1 with boxes of side l / b = b‘-‘, one needs b“ = b stacks
of boxes, each with a height between b”-’ and b”-‘
[1/2)(1 + b‘/b“)] . Thus, apart from a multiplying factor of the
3. The fractal dimension of the above self-affine records
-
order 1, one has N b”(b”-’/b’-’) = (b”b”-’)’. From H = log
It is widely known that for the graph of zeros of B(t), the b”/log b’, we have b” = b r H ,hence N = b2-H. The multiplying
Haufsdorff-Besicovitch dimension is 1/2, and almost as widely factor vanishes when taking l i m b . + log N/log b, hence
known that for the graph of B(t) itself the Hausdorff- D B = 2-H. Observe that the scales chosen for t and B do not
Besicovitch dimension is 1.5 = 1/2 + 1. The corresponding matter in this high frequency limit.
dimensions for the records of both BH(t) and M H ( t ) are 1-H The physicist, however, also thinks of the global limit b + 0
and 2 8 . But the Hausdorff-Besicovitch dimension is a very or r +-, which requires an unbounded record. The portion of
non-intuitive notion. While I was delineating and developing the a self-affine record from 0 to t S 1 is covered by a single box.
new fractal geometry, I used to pay lip service to it, but I was Hence limb+O log N(b)/log b = 1. (The detailed argument re-
careful to describe this as a “tactical” and “tentative” step, quires some care, but we shall not dwell on it). Conclusion: two
because this dimension can be no use in empirical work, and is limits that are identical for self-similar fractals are now found to
unduly complicated in theoretical work, except for self-similar differ!
fractals. Instead, my work introduced (in increasingly formal
fashion) several alternative definitions that are useful precisely t,, defined as being such that B H ( t + t,) -BH(t) -
Thus, a self-affine curve involves a cross-over value o f t , call it
t,. Stated
Physica Scripta 32
SelfAffine Fractals and Fractal Dimension 259
graph paper. The “mass” of the record of B H ( t )between times particular its figurative meaning in mathematics, in science and
t’ and t“ is set to be It‘ - t ” I . in philosophy is profoundly different from its meaning for the
When R 9 t,, the record of B H ( t )is effectively a horizontal architect and the layman. A building’s foundations always come
interval. It occupies a very thin horizontal slice of the square of first, followed by the basic shelter and later by the decorative
side 2R, hence M(R) 0:R , and D M = 1. work. Before expanding a building, one tests its foundations and
When to the contrary, R & t c , the record of B H ( t ) is one strengthens them if necessary. In successful branches of
effectively a collection of vertical intervals, one for each zero of mathematics and of science, to the contrary, concern with
BH(t). Again, the argument is simplified if we replace BH(t) “foundations” tends to come late, after each period of extensive
by MH(t) and consider a square of side R = b“-’, with top and substantive achievement. An endeavor wins little respect as a
bottom ordinates proportional to b“-’. The mass we seek is the science if it favors methodology over substance. Ten years ago,
same as for the kth approximant function MH(‘)(t). Thus, mass laying careful foundations for fractal geometry was not a
is the number of times MH(‘)(t) traverses the ordinate of the priority concern. The new discipline had to gain acceptance, and
center of a square, multiplied by the duration 6 of each - before that - it had to be built almost from scratch (it
traversal. The number of traversal is w ( R / ~ ) ’ -and
~ , 6 = b’-‘ = includes a few essential parts salvaged from the work of
-
(b‘’-k)’’H = R’’H. Hence M(R) R 2 - H ,yielding the familiar
value D M = 2-H in the small R limit.
mathematicians engaged in other pursuits, but to this day - as
exemplified by this Note - many basic issues had been com-
pletely untouched.) Anyhow, events moved fast, and are pres-
ently forcing me to write a systematically organized textbook.
7. Cross-over pitfalls The present Note is based on excerpts from the book’s first
To summarize, self-affine fractals do not involve exponents such draft. A more detailed excerpt will appear in [4].
- - -
thatN(b) b D B , L (7) q - D C for all 7 and M ( R ) RDM for all
R Different exponents are approached on different sides of
the crossover point t,. And the value o f t , is not always intrinsic, Acknowledgement
since in the case of noises it depends on the units chosen along
Richard F. Voss, with whom I discussed the material in this Note
the axes. A truly mechanical estimate of D B ,D M or of D c is repeatedly, has made very useful comments and valuable suggestions.
likely to combine values of 17 or of R that range on both sides of
the cross-over t,, and the estimate will depend upon where
exactly t, lies in the range of 77 or R . Such an estimate will be References
worthless, because it will depend on the units of t and B. Re-
liability is improved by exaggerating the vertical scale. 1. Mandelbrot, B. B., The Fractal Geometry ofNature, Freeman, New
York (1982).
2. Mandelbrot, B. B., Passoja, D., and Pallay, A . Nature 308, 721
8. Discussion (1984).
The notion of fractal dimension brings an unavoidable prolifer- 3. Proceedings of a Symposium on Fractals in the Physical Sciences
US. National Bureau of Standards, Gaithesburg, MD, 21 -23,
ation of distinct quantities, each contributing to a fractal’s November (1983).J. Star. Physics. 36 Numbers 5 / 6 (1984).
overall description. 4. Proceedings o f the Sixth 7kieste International Symposium, “Fractals
Finally, consider the issue of the foundations of fractal in Physics ”, Ed. L. Pietronero, North-Holland, Amsterdam (to
geometry. “Foundations” is a treacherous term, and in appear).
Physica Scripta 32