de Gruyter Textbook

Helmberg Getting Acquainted with Fractals

Gilbert Helmberg

Getting Acquainted
with Fractals

Walter de Gruyter
Berlin New York

Gilbert Helmberg
Kalkofenweg 5
6020 Innsbruck
Austria

Printed on acid-free paper which falls within the guidelines

of the ANSI to ensure permanence and durability.

Library of Congress Cataloging-in-Publication Data

Helmberg, Gilbert.
Getting acquainted with fractals / by Gilbert Helmberg.
p. cm.
Includes bibliographical references.
ISBN 978-3-11-019092-2 (hardcover : alk. paper)
1. Fractals. I. Title.
QA614.86.H45 2007
5141.742dc22
2006102211

Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

ISBN 978-3-11-019092-2
Copyright 2007 by Walter de Gruyter GmbH & Co. KG, 10785 Berlin, Germany.
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced in any form or by any means, electronic or mechanical, including photocopy,
recording, or any information storage and retrieval system, without permission in writing from
the publisher.
Printed in Germany.
Coverdesign: malsy, kommunikation und gestaltung, Willich.
Printing and binding: Hubert & Co. GmbH & Co. KG, Gttingen.

Preface
To someone, having heard about fractals but not yet acquainted with them, they might
seem to be regarded with suspicion: How could real objects accessible by sight and
not only by thought be replicas of arbitrarily small parts of themselves? How could
a continuous path which runs almost everywhere parallel to sea level climb up to any
height? How could a continuous curve pass through every point of a square?
Getting acquainted with fractals opens a glimpse into a world of wonders, but these
wonders are strongly supported by a frame of serious mathematics in which various
of its branches play together: geometry, analysis, linear algebra, topology, measure
theory, functions of complex variables, algebra, . . . .
I have tried to do justice to both aspects: the fascination of geometric objects as well
as the serious mathematical background as far as an advanced undergraduate level.
At some points, where the technicalities would transgress this level, I have at least
indicated where an interested reader could find the whole story. I hope the presentation
adds something worthwhile to the many remarkable books on this topic which also
lead much farther into the world of fractals.
These books also contain something which a reader might miss in the present one: I
have chosen to avoid the possibility of frustrating the reader by expecting him to do exercises; he will find them in abundance in the mentioned books (e.g. [Barnsley, 1988],
[Falconer, 1990]) if he wants to. However, it is at least my intention to make accessible via the internet address http://techmath.uibk.ac.at/helmberg
the programs producing the illustrations, thus enabling the reader to create and play
with fractals according to his own taste.
My thanks are due to the de Gruyter Publishing Company, in particular to Dr. Plato,
for their interest in and support of this book project. My first book has been dedicated
to my parents, my wife, and my two eldest children, but there are more people who
mean very much to me. Therefore this book is dedicated
to Chri, Moni, and Mui.
Innsbruck, Cavalese, August 2006

Gilbert Helmberg

Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1

Fractals and dimension . . . . . . . .

1.1 The game of deleting and replacing
1.2 The box-counting dimension . . .
1.3 The H AUSDORFF dimension . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

. 1
. 1
. 50
. 55

Iterative function systems . . . . . . . . . . . . . . . . . . .

2.1 The space of compact subsets of a complete metric space
2.2 Contractions in a complete metric space . . . . . . . . .
2.3 Affine iterative function systems in R2 . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

63
63
70
74

Iteration of complex polynomials . . . .

3.1 General theory of J ULIA sets . . . .
3.2 J ULIA sets for quadratic polynomials
3.3 The M ANDELBROT set . . . . . . . .
3.4 Generation of J ULIA sets . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

109
111
121
124
150

Bibliography
List of symbols

.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.

.
.
.
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Contents (detailed) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

Fractals and dimension

1.1

The game of deleting and replacing

The word fractal comes from the Latin word frangere (with past participle fractus) which means to break, to destroy. Let us begin with exploring how such a
destruction process may still generate some new mathematical object displaying interesting features.

1.1.1

The C ANTOR set

Let us define an operation f (such an operation is commonly called an operator) working on any closed segment [a, b ] R (= the real line) by deleting the open middle
ba
third ]a + ba
3 , b 3 [ , and let us denote the interval [ 0, 1 ] R by A(0) . Application
of f to A(0) deletes the interval ] 13 , 23 [ and produces a closed set
 


A(1) = 0, 13 23 , 1 ,

the union of the two disjoint closed intervals A0 = [ 0, 13 ] and A1 = [ 23 , 1 ], each of

which has length 13 . If we apply f now to A(1) we get a closed set
A(2) = f (A(1) ) = f (f (A(0) )) A(1)

consisting of four disjoint intervals A0,0 , A0,1 , A1,0 , A1,1 of length 19 = 312 each.
Since we want to continue the application of f , in order to avoid the clumsy notation
f (f (. . .)) let us use the notation
f (0) (A) := A,
f (1) (A) := f (A),
f (k+1) (A) := f (f (k) (A)).

(We shall call the index k the level of the construction.) Applied to our interval A(0)
this allows us to define a sequence of closed sets A(k) (1 k < ) by
A(k) := f (k) (A(0) )

satisfying
A(0) A(1) A(k) A(k+1) .

(1.1)

The set A(k) is the union of 2k closed intervals Aj1 ,...,jk (ji {0, 1}, 1 i k) of
length 31k each. A sequence {A(k) }
k=1 as well behaved as indicated by (1.1) raises the
question whether there exists, in some sense, a limit set A. Indeed, by a well known

Chapter 1 Fractals and dimension

topological theorem, the decreasing sequence of non-empty compact sets {A(k) }

k=1
has the property that the set


A :=
A(k) ,
k=1

called the C ANTOR set [Cantor, 1883], is compact and not empty. See Figure 1.1 for
an illustration of the set A(4) .

Figure 1.1. The set A(4) , pictured in blue, is the union of sixteen closed
component-intervals. The open set [0, 1] \ A(4) , pictured in red, is decomposed
according to the intervals deleted at levels 1, 2, 3 and 4.

Still, as to the size of the set A, we notice that it is contained in all sets A(k)
(1 k < ); as observed above, the total length of the 2k component-intervals of
k
A(k) is 23k = ( 23 )k which approaches zero as k . If A is to have any length in
some sense at all, it therefore must be zero. Indeed, using one-dimensional L EBESGUE
measure L (= L1 ), which on the sets A(k) coincides with their lengths, by a wellknown theorem of measure theory we get

 k


2
L(A) = L
A(k) = lim L(A(k) ) = lim 3
= 0.
k=1

It is not surprising that A is not empty: the countably many end points of all component
intervals Aj1 ,...,jk (0 k < , ji {0, 1}, 1 i k) are never deleted by
any application of f and therefore are all contained in A. But there are more points
surviving all these applications:

xk
Let us write every non-zero x [ 0, 1 ] as an infinite series x =
k=1 3k (xk
{0, 1, 2}), in short x = 0
.x1 x2 . . . with the understanding that any finite sum of the
form x = 0.x1 x2 . . . 1 = nk=1 x3kk (xn = 1) shall be written as a non-ending periodic
triadic fraction
x = 0.x1 x2 . . . xn1 022 . . . =

1
k=1

xk
2
+
.
3k
3k
k=n+1

Section 1.1 The game of deleting and replacing

Application of f to A(0) eliminates all points x for which x1 = 1. The set A(1) therefore
contains none of these. Renewed application of f to A(1) now eliminates all points x
for which x2 = 1 (in both intervals A0 and A1 which are characterized by x1 = 0
and x1 = 2 respectively). Repeated application of f subsequently eliminates all points
x A(0) for which xk = 1 (1 k < ). What remains? Precisely the set of all
points x A(0) whose digits xk are either 0 or 2. It is well known that the points
of the interval [ 0, 1 ], apart from the countably many dyadic rational points, are in
one-to-one correspondence with the points which in dyadic notation may be written as
y = 0.y1 y2 . . . (yk {0, 1}). The conclusion is that our set A is not countable but
contains as many points as the interval [ 0, 1 ], i.e. has the cardinality of the reals.
A mathematician may be tempted to exploit the relation between the set A and a
subset of [ 0, 1 ] even further. Just now we have associated with the point
y =

yk
k=1



yk {0, 1},
1 =
1=

y k =0

the point
a(y) =

2 yk
k=1

3k

y k =1

A.

Denoting by N the set of natural numbers, we may extend this mapping a to the compact topological product {0, 1}N of all {0, 1}-sequences by defining
a
(
y ) :=

2 yk
k=1

3k

for y = {yk }
k=1

(yk {0, 1}),

e.g. if y = {0, 1, 1, . . . }, then a(
y ) = k=1 32k = 13 , while for y = {1, 0, 0, . . . } we get
a
(
y ) = 23 . It is not hard to see that the mapping a
: {0, 1}N A is bijective (every point
of A is the image of exactly one sequence in {0, 1}N ) and continuous. A well-known
topological theorem (cf. [Kelley, 1955, p. 141]) then asserts that A is homeomorphic to
{0, 1}N . In particular, A is completely disconnected (in topology such a space is also
called zero-dimensional) and perfect (i.e. closed without isolated points), but nowhere
dense. Remembering that [ 0, 1 ] may be considered as a subset of {0, 1}N and roughly
speaking, the mapping a furnishes an extended parametrization of the set A (i.e. to
every dyadic rational point of [ 0, 1 ] there correspond two neighbouring points of A).
At this point we may notice one more property of the set A which is important for
us since it will turn up in adapted form repeatedly in sets which we legitimately may
call fractals: suppose we omit the component A1 of the set A(1) and restrict repeated
application of f to the interval A0 . What would we have got? Evidently part of A, to
wit a copy of the set A, only reduced by a factor 13 in size. In fact, every component
set Aj1 ,...,jk of A(k) , treated by itself with successive applications of f , produces a set
which is part of A and, at the same time, a copy similar to A but reduced by a factor 31k .
In other words, one may say that the set A is self-similar in the sense that it consists
of smaller parts which are still similar to A.

Chapter 1 Fractals and dimension

A slight adaption of our construction of the C ANTOR set furnishes a set with strikingly different features. Since we are now going to move from R to R2 , and more
Rn , for any vector x = (x1 , . . . , xn ) Rn we shall use the E UCLIDEAN
generally to
n
n
n
2
norm |x| =
k=1 xk . The (E UCLIDEAN ) distance of two points a R and b R
is then given by |a b|.

1.1.2

The KOCH curve

We shall now modify the operator f considered in Section 1.1.1 by allowing it to work
on any closed segment [a, b ] in the plane R2 , and in the following way: it not only
ba
deletes but replaces the open middle third ]a + ba
3 , b 3 [ by two sides of an equilateral triangle, side length | ba
3 |, located to the left of [a, b ] if this segment is directed
from a to b. We shall apply f also to piecewise linear curves in R2 . Such a curve E is
the graph of a piecewise linear, not necessarily continuous, function g : [ 0, 1 ] R2 . It
consists of finitely many segments [aj , bj ] (1 j n), at most pairwise joined at their
endpoints. The result f (E) of applying f to E is obtained by applying the operator f
to each of the component segments of E .
We start out again with the segment A(0) = [ 0, 1 ] on the x-axis. Application of
f to A(0) produces a continuous piecewise linear curve A(1) = f (A(0) ) consisting of
four segments denoted consecutively by Aj (0 j 3), each of which has length 13
(Figure 1.2).

Figure 1.2. The generator A(1) of the KOCH curve.

Why not apply f again, this time to A(1) , i.e. to each of these four segments? The
result is a continuous piecewise linear curve A(2) = f (2) (A(0) ) consisting of 42 segments Aj1 ,j2 (0 ji 3) of length 312 each. Repetition of this procedure furnishes
a sequence of continuous piecewise curves A(k) consisting of 4k segments Aj1 ,...,jk
(0 ji 3, 1 i k) of length 31k each (Figures 1.31.5). Unfortunately, however, these curves, considered as subsets of R2 , do not anymore satisfy (1.1). Do they
still converge in some sense to some limit? The eye emphatically approves, but does
mathematics support this impression?
In order to investigate the situation, we turn our attention to the sequence of segments Aj1 ,...,jk in R2 constituting the curve A(k) successively, starting at (0, 0) and
ending at (1, 0). Notice that the endpoints of these segments are preserved when f

Section 1.1 The game of deleting and replacing

Figure 1.3. The approximating set A(2) for the KOCH curve. The grey lines
illustrate the open set condition (Definition 1.1.3.2) needed for the computation
of dimS (A).

Figure 1.4. The approximating set A(3) for the KOCH curve.

Figure 1.5. A closer approximation (A(7) ) of the KOCH curve A. The four
colours indicate subsets of A which are similar to the whole of A.

Chapter 1 Fractals and dimension

is applied to A(k) since they only become endpoints of smaller subsegments. Let us
define a map k : [ 0, 1 ] A(k) in the following way: write every x [ 0, 1 ] in its
4-adic expansion
x =

x
i=1

i
i

x
i=1

4i

(xi {0, 1, 2, 3}, 1 i < ).

+ rk (x)

(1.2)

If we agree to use the finite sum expansion when possible, then 0 rk (x) < 41k . Now
define k (x) to be the point of the segment Ax1 ,x2 ,...,xk lying at distance rk (x) from the
starting point of this segment (in the positive direction). Evidently k is a continuous
piecewise linear map of [ 0, 1 ] onto A(k) .
What happens to k if k increases to ? In order to find out about this let us
observe what happens to the point k (x) if x is given as in (1.2) and if we define
x(m) :=

x
i=1

4i

As pointed out above, for the 4-adic rational part x(m) of x and for all k m we get
k (x(m) ) = m (x(m) ) Ax1 ,...,xm ,0,...,0 ; A(k)

(1.3)

(in fact, k (x(m) ) is the starting point of this subsegment of A(k) ). Observing the effect
of consecutive applications of f to Ax1 ,...,xm , we find that every such application moves
the point k (x) Ax1 ,...,xm ,...,xk to its new position k+1 (x) Ax1 ,...,xm ,...,xk ,xk+1
4
about a distance of at most 3k+
1 (four times the length of Ax1 ,...,xk+1 ; a rough estimate

3
since at most 23k+1 would do). Adding this up for k > m we get the estimate
|k (x) m (x)|

k1

4
i=m

3i+1

<

4
i=m

3i+1

2
.
3m

(1.4)

As a consequence, we see that the sequence {k (x)}

k=1 is a C AUCHY sequence (=
fundamental sequence) in the plane R2 and has to converge to a limit point (x). By
(1.4) we even see that the functions k converge uniformly on [ 0, 1 ] and that therefore
the limiting map furnishes a continuous curve in R2 . This curve is called the KOCH
curve [von Koch, 1904].
The fact that endpoints of subsegments Ax1 ,...,xm of A(m) do not change position
under further applications of f , as expressed by (1.3), helps to realize that the KOCH
curve is nowhere differentiable. Non-differentiability is readily seen at such an endpoint itself: For any x [ 0, 1[ consider the point p0 = (x(m) ) as defined above (for
x = 1 the reasoning has to be slightly adapted). For k > m let




1
2
p2 := x(m) + k .
p1 := x(m) + k ,
4

p1 is the endpoint of the subsegment of A(k) beginning at (xm ), p2 is the endpoint of

the following subsegment of A(k) . If k > m is sufficiently large, then the points p1 and

Section 1.1 The game of deleting and replacing

p2 are arbitrarily close to p0 , while the secants p0 p1 and p0 p2 always include the same
positive angle.
It is somewhat more tedious to deal with a point (x) if x is not a 4-adic rational
number. Roughly speaking, if were differentiable in x, then two different points
close to (x) (as which will be taken endpoints of subsegments) would have to define a
secant close to the tangent in (x), and this will be shown to be impossible. Recall that
for complex-valued functions g and h of the argument y one writes h = o(g) as y a
|h(y)|
if limya h(y)
g(y) = 0, while h = O(g) as y a means lim supya |g(y)| < . Correspondingly o(1) (as y x) will denote a function which vanishes as y x, and O(1)
will denote a function which remains bounded as y x. Although not strictly necessary the notation o1 (1), o2 (1), . . . , O1 (1), O2 (1), . . . will be used to indicate different
such functions.

1.1.2.1 Lemma. Suppose is differentiable in x, i.e. there exists a vector q R2 such

that
(x) (y) = (q + o(1)) (x y) as y x,
and let y1 x and y2 x in such a way that
x y1 = O1 (y1 y2 ),

(1.5)

x y2 = O2 (y1 y2 ).

Then

(y1 ) (y2 ) = (q + o(1)) (y1 y2 ).

(1.6)

Proof of the lemma.

(y1 ) (y2 ) = (x) (y2 ) ((x) (y1 ))




= q + o2 (1) (x y2 ) q + o1 (1) (x y1 )
= q (y1 y2 ) + o2 (1) (x y2 ) + o1 (1) (x y1 )

= q + o2 (1) O2 (1) + o1 (1) O1 (1)] (y1 y2 ).

In order to show that cannot be differentiable in x let, for arbitrarily large m,

y1 := x(m) ,

y2 := x(m) +

1
,
4m+1

y2 := x(m) +

Then
|y1 y2 | = 2 |y1 y2 | =

2
,
4m+1

|x y1 |

4
4m+1

= 4 |y1 y2 |,

|x y2 |

3
4m+1

= 3 |y1 y2 |,

|x y2 |

2
4m+1

= |y1 y2 |.

2
.
4m+1

Chapter 1 Fractals and dimension

So the requirements (1.5) are satisfied, but not (1.6) since we have already seen that the
secants (y1 )(y2 ) and (y1 )(y2 ) always include the same small but non-zero angle.
Consequently, the function cannot be differentiable in x.
How do we measure the length of a continuous curve? Take any finite sequence
S = {pj }nj=0 of points corresponding to increasing parameter values and compute
n
lS := j=1 |pj pj1 |. The length of the curve is then by definition the supremum
over all values lS obtained in this way. For the KOCH curve it seems convenient to
k
choose Sk := {( 4jk )}4j=0 , the endpoints of the 4k subsegments in A(k) . Each of these
has length 31k , therefore we get lSk = ( 43 )k . As k this also tends to . We
conclude that the KOCH curve has infinite length, rather a contrast to the C ANTOR set.
One last question (for the time being): what would have happened if we had restricted the action of f to one subsegment Aj (j {0, 1, 2, 3}) of A(0) or, more generally, to a subsegment Aj1 ,...,jk of A(k) ? Obviously we would have got a curve similar
to A but reduced to 13 , resp. 31k , in size. In other words, again the KOCH curve is
self-similar, it consists of parts which are smaller copies of itself.

1.1.3

Heuristics of dimension

We have not yet pinned down any property of a set in R2 or R3 or, more generally,
Rn which might be strange and characteristic enough to make it reasonable to call the
set a fractal. Self-similarity as encountered in the C ANTOR set or the KOCH curve
seems a possible candidate but there are perfectly harmless sets which also are selfsimilar, for instance a square in the plane. Shrinking its sides to half the original length
again produces a square and the original square consists of four copies thereof if we
allow the sides of the small squares to coincide. In fact, this is intimately connected
with the assertion that a full square is a set of dimension 2: reducing the sides to n1
of their original length produces a set, n2 copies of which (allowing sides to coincide)
constitute the original square. Similarly n3 cubes of side length n1 make up the unit
cube corresponding to its three-dimensionality and n1 intervals of length n1 joined
together give the one-dimensional unit interval.
If we had not already been familiar with the concept of dimension we could have
computed the dimension of a square A, say, using the following reasoning: the dimension of A is the exponent d determined by the fact that the set A is a (almost
disjoint, whatever this may mean) union of nd similar copies of A, reduced in size by
the factor n1 (such a similar copy S(A) is congruent with the set n1 A, which originates
by multiplying every vector in A by the factor (S) = n1 ). In other words and roughly
speaking (to be made more precise in later sections), if (S) = n1 and if A happens to
be decomposable into NS (A) sets of the form (S)A, then the self-similarity dimension
dimS (A) may be considered as the solution of the equation

dimS (A)
1
= NS (A),
(S)

i.e.
dimS (A) =

log NS (A)
.
log (S)

(1.7)

Section 1.1 The game of deleting and replacing

Applying this reasoning to the C ANTOR set A we recall that it is indeed the disjoint
union of two similar copies of itself, reduced by the factor (S) = 13 . Formula (1.7)
now gives for its dimension
dimS (A) =

log 2
0.63.
log 3

There is one objection to be dealt with: dimS (A) has not been defined in a unique
way. If A is the (almost) disjoint union of NS (A) copies of (S)A, then (S)A is
the (almost) disjoint union of NS (A) (almost) disjoint copies of ((S))2 A and A is
the (almost) disjoint union of NS (2) (A) = (NS (A))2 copies of (S (2) )A = ((S))2 A.
Should we have been told, before applying formula (1.7), whether to work with S or
with S (2) , or even with the k -fold iteration S (k) of S ? Fortunately this does not matter,
since
log NS (k) (A)
k log NS (A)
log NS (A)
=
=
.
(k)
(S

k log (S)

log (S)

If there is some doubt left, please be patient until dimension is discussed more thoroughly in Section 1.2 and Section 1.3.
The startling fact is that the dimension of the C ANTOR set, with this understanding,
is not 1 but less, to wit approximately 0.63 (also different from its topological dimension as a completely disconnected set, which is zero). Looking now at the KOCH curve,
formula (1.7) tells us that (if the points in which the subsegments join do not do any
4
damage) its self-similarity dimension is log
log 3 1.26, while of course the topological
dimension of each of its approximating sets A(k) is 1.
A theorem (Theorem 1.3.8) to be stated later tells a condition under which this
reasoning is applicable, the so-called open set condition. Let us first state explicitly
what is meant by a similarity.
1.1.3.1 Definition. A map S : Rn Rn is called a similarity with similarity factor s
if |S(x) S(y)| = s |x y| for some positive number s and for all x Rn and y Rn .
1.1.3.2 Definition. The similarities Si (1 i k) satisfy the open set condition if
there exists a bounded non-empty open set V with mutually disjoint image sets Si (V )
k
(1 i k) satisfying i=1 Si (V ) V .
In essence the mentioned theorem states that if the similarities Si (1 i k)

satisfy the open set condition and if A = ki=1 Si (A), then (1.7) and even a more

general formula for the computation of dimS (A = ki=1 Si (A)) may be applied. The
open set condition is obviously satisfied in the case of the C ANTOR set: denoting by S1
and S2 the similarities mapping the unit interval into its first and last third, as the set V
we may take e.g. the open unit interval. It is satisfied also in case of the KOCH curve:
let V be the open isosceles triangle with vertices a = (0, 0), b = (1, 0), c = ( 12 , 21 3 )
(see Figure 1.3). V contains its four images under the similarities mapping the unit
interval into the four line segments constituting the set A(1) . As a consequence, we
may also note that the whole set A is contained in the closure of the triangle V .
Are we now in position to define what is meant by a fractal? Yes and no. Yes,
since the original definition of M ANDELBROT [Mandelbrot, 1982, Section 3] says: A

10

Chapter 1 Fractals and dimension

subset of Rn is a fractal (also fractal set) if its topological dimension (which is always
an integer, zero for the C ANTOR set and one for the KOCH curve) is less than its
fractal dimension (for the C ANTOR set and the KOCH curve as computed above).
According to this definition a set with a non-integral dimension (as discussed more
generally later) is automatically a fractal. No, since it has turned out that there are sets
(as the dragon Section 1.1.5.3 to be discussed later) that one would like to consider
as fractals but are not included by the just mentioned definition. Up to now it has
seemed difficult to find a satisfying definition including all sets which one would like
to consider as fractals.

1.1.4

Initiators and generators

There are evidently two ways to produce more general fractals besides the C ANTOR set
and the KOCH curve: we can start from a set different from the unit interval A = [ 0, 1 ]
and we can (as we have done already) change the definition of the map f . The first
is done by defining a set of segments A(0) , then called the initiator, upon which the
iterates f (k) of f should act. An example is provided as follows.
1.1.4.1 The KOCH island
Let A(0) be the equilateral triangle below the x-axis, one side of which is the unit
interval [ 0, 1 ]. Applying the map f defining the KOCH curve (Section 1.1.2) to the
three segments constituting the set A(0) produces a star with six vertices which we
may also imagine as an equilateral hexagon, each side of which carries an equilateral
triangle of side length 13 (Figure 1.5). The next application of f adds twelve smaller
equilateral triangles of side length 19 . Continuing this procedure eventually produces
the contour of a set looking like a snow flake, consisting of three copies of the KOCH
curve we know from Section 1.1.2. The idea of it being surrounded by water leads to
calling it the KOCH island (Figures 1.6, 1.7).

Figure 1.6. The first approximating set A(1) for the KOCH island.

Section 1.1 The game of deleting and replacing

11

Figure 1.7. A closer approximation (A(7) ) of the KOCH island.

Still, there is more to this: somebody seeing it for the first time and being asked to
estimate the length of the coast line may think: Well, a little bit more than the perimeter of a circle roughly the same size; taking into account the coasts of the peninsulas
and the bays, perhaps twice this perimeter. Asked to look a bit closer and perhaps to
use a compass with a rather small opening he may to his surprise find that his measurement of the coast line becomes longer and longer as he decreases this opening, until we
disclose to him that already one third of the coast line our well-known KOCH curve
has infinite length.
It is in this line of thought that M ANDELBROT [Mandelbrot, 1982, Section 5] points
out that also e.g. the coast line of England, measured with increasing precision, turns
out to have infinite length.
Keeping, for the time being, the unit interval as our initiator A(0) , we may change
the mapping f by requiring that it should act on every segment of any union B of
segments by replacing this segment with a suitably diminished similar copy of a
given union G of segments, called the generator. Let us look at several samples of the
vast family of fractals obtained in this way.
1.1.4.2 A modified KOCH curve
Suppose the generator G consists of five segments of length 13 each, obtained by replacing the two middle segments of the KOCH curve generator with three sides of a
square (Figure 1.8).
We may think of f as employing five similarity maps Si (1 i 5) each with
similarity factor 13 . Now the fractal A = limk f (k) (A(0) ), defined in essentially
the same way as in Section 1.1.2, consists of the union of five similar copies Si A
(1 i 5) joined at the vertices of the original generator G = A(1) , but the first two
and the last two copies having a lot more points (in fact a whole diagonal segment)
in common (Figure 1.9). Still, the open set condition (Definition 1.1.3.2) is satisfied:
the open isosceles right-angled triangle D with A(0) as hypotenuse contains the union

12

Chapter 1 Fractals and dimension

Figure 1.8. The generator of the first modified KOCH curve. The grey lines
delimit the triangle D providing for the open set condition.

Figure 1.9. A closer approximation (A(6) ) of the first modified KOCH curve.

i=1 Si D of its five copies under the similarities Si (1 i 5). According to what
has been said in Section 1.1.3, we may compute the dimension of the set A as being
5
dimS (A) = log
log 3 1.46.
Suppose we still adapt the generator G by shortening its second and fourth segment
to 15 , say, in place of 13 (Figure 1.10).
The resulting fractal A (Figure 1.11) still consists of five similar copies of A, but
the first, third, and fifth similarity employ the similarity factor 13 , while the second and
fourth copy employ the similarity factor 15 . Now the first and second copies, as well as
the fourth and fifth copies, have only the connecting vertices of the original generator
in common, and the open set condition still holds since the former right-angled triangle
above may be replaced by an oblique-angled symmetric triangle through the outer
vertices of the new generator. Our method of determining a dimension of the resulting
fractal A, however, breaks down, since the similar copies of A constituting A do not
have the same size. If we consider A to be eligible for a dimension as seems to
be sensible then we shall have to use a refined reasoning. Fortunately this will be

Section 1.1 The game of deleting and replacing

13

Figure 1.10. The generator of the second modified KOCH curve.

Figure 1.11. A closer approximation (A(6) ) of the second modified KOCH curve.

supplied by Theorem 1.3.8. An application thereof, using the concept of H AUSDORFF

dimension will furnish the result dimH (A) 1.27.
A more troublesome situation happens if we adapt the generator G by enlarging
(instead of shortening) its second and fourth segment to 12 , say (Figures 1.121.14).
The second and fourth copy of the resulting fractal A appearing as part of A then
employ the similarity factor 12 and become entangled with (i.e. intersect) the first and
fifth copy respectively to an extent which is not anymore negligible and no longer allows one to find an open set furnishing the open set condition. But should this prevent
the fractal A from legitimately having a dimension? This would be very unsatisfactory

Figure 1.12. The generator of the third modified KOCH curve.

14

Chapter 1 Fractals and dimension

Figure 1.13. The approximating set A(4) for the third modified KOCH curve.

Figure 1.14. A closer approximation (A(7) ) of the third modified KOCH curve.

15

Section 1.1 The game of deleting and replacing

indeed. In Section 1.2 and Section 1.3 we shall therefore discuss two reasonable concepts of dimension which, although not always computable in practice, will give
every reasonable set a number as its dimension.
1.1.4.3 A second type of modified KOCH curves
Let us consider a whole family of related fractals, starting with the unit interval as
initiator and the generator G1 consisting of eight segments of equal length joining
consecutively the points
p0 = (0, 0)

p1 = (0.25, 0)

p2 = (0.25, 0.25)

p3 = (0.5, 0.25)

p4 = (0.5, 0)

p5 = (0.5, 0.25)

p6 = (0.75, 0.25)

p7 = (0.75, 0)

p8 = (1, 0).

The open parallelogram formed by the straight lines p0 p2 , p3 p8 , p8 p6 , p5 p0 serves to

verify the open set condition; the dimension of the resulting fractal is dimS (A) =
log 8
3
log 4 = 2 (Figures 1.15, 1.16).
Replacing in the generator G1 the two squares above and below the x-axis by equilateral triangles produces a generator G2 consisting of six segments of equal length
joining consecutively the points
p0 = (0, 0)
p3 = (0.5, 0)

p1 = (0.25, 0)
p4 = (0.625,

p2 = (0.375,

3
8 )

3
8

0.2165)

p5 = (0.75, 0)

p6 = (1, 0).

Figure 1.15. The generator G1 . The grey lines delimit the parallelogram providing for the open set condition.

16

Chapter 1 Fractals and dimension

Figure 1.16. The approximating set A(5) for the fractal with generator G1 .

Again the open parallelogram with vertices p0 , p2 , p6 , p4 serves to verify the open set
condition (also its similar images with diagonals p3 p4 and p4 p5 have only the point
6
p4 in common); the dimension of the resulting fractal is dimS (A) = log
log 4 1.29
(Figures 1.17, 1.18).
It would be perfectly legitimate to omit the vertex p3 and to consider p2 p4 as a
single segment, but formula (1.7) could no longer be used for the computation of the
dimension of the resulting fractal.
Finally we contract the two truncated equilateral triangles in G2 above and below
the x-axis to segments of length 13 , positioned perpendicularly to the unit interval at
its points ( 13 , 0) (above the x-axis) and ( 23 , 0) (below the x-axis) respectively (G3 ). In
fact although each approximating curve f (k) (A(0) ) runs twice through these segments,
because of the symmetry properties of the generator we may now envision the resulting

Figure 1.17. The generator G2 . The grey lines delimit the parallelogram providing for the open set condition.

Section 1.1 The game of deleting and replacing

17

Figure 1.18. The approximating set A(5) for the fractal with generator G2 .

fractal A to consist of five similar copies of the form 15 A (Figures 1.19, 1.20). The open
quadrangle (0, 0), ( 13 , 13 ), (1, 0), ( 23 , 13 ) provides the set for the open set condition,
5
therefore the dimension of the corresponding fractal is again dimS (A) = log
log 3 1.46.

Figure 1.19. The generator G3 . The grey lines delimit the parallelogram providing for the open set condition.

18

Chapter 1 Fractals and dimension

Figure 1.20. The approximating set A(6) for the fractal with generator G3 .

1.1.4.4 More modified KOCH curves

A nice-looking fractal is obtained using the unit interval as initiator and a generator G4
(Figure 1.21) consisting only of two segments joining the points
p0 = (0, 0)

1
p1 = (0.5, 2
0.2887)
3

p2 = (1, 0).

Figure 1.21. The generator G4 .

In fact, the second coordinate of p1 may have some other value, but it is convenient
at this point to have an inclination of 30 degrees at both endpoints of the generator.
There is no obvious candidate for an open set admitting the open set condition, so we
stop worrying about the dimension.
Although the first polygonal sets A(k) = f (k) (A(0) ) look somewhat harmless, like
bathing caps, there are strange things going on in the endpoints of the segments constituting the sets A(k) . Recall that, if p is such a point in A(k0 ) , it remains fixed under
further applications of f . Let the left and right neighbouring points of p in A(k) be
called q(k) and q(k) respectively and let us agree to paint everything to the left of p
(with respect to the orientation of the curves A(k) ) blue and everything to the right of

19

Section 1.1 The game of deleting and replacing

p red. In contrast to q(k+1) p and p

q(k+1) the blue segment q(k) q(k+1) and the red segment q(k+1) q(k) are not part of A(k+1) but for sake of supporting our intuition let these
segments for the moment replace the part of A(j) (j > k) between their endpoints. An
application of f to A(k) produces a blue segment qk+1 p by rotating the blue segment
q(k) p clockwise about 30 degrees and shortening it by the factor 13 . The same operq(k) by p
q(k+1) . After at most six applications of f
ation counterclockwise replaces p
the points q(k) and q(k) coincide, causing the corresponding polygon A(k) to produce
a loop (Figure 1.22). The blue segments q(k) q(k+1) and the red segments q(k+1) q(k)
spiral in opposite direction about the point p and the same will be done by the fractal
A. To give intuition the final blow: this happens at a dense set of points on the curve
A = ([ 0, 1 ]) (Figure 1.23).
A weird feature is added if we join two more points
p3 = (0.8, 0)

p4 = (1, 0).

The last segment of the new generator G5 (Figure 1.24) is now counted twice, as p2 p3
and p3 p4 . Correspondingly, similar pictures of the generator appear above and below
this segment. The effect is the apparition of loops, increasing to the right and making
the fractal appear like a strange flower (Figure 1.25).

Figure 1.22. The approximating set A(6) for the fractal with generator G4 .

Figure 1.23. A closer approximation (A(13) ) of the fractal with generator G4 .

20

Chapter 1 Fractals and dimension

Figure 1.24. The generator G5 .

Figure 1.25. The approximating set A(8) for the fractal with generator G5 .

Loops also appear in a fractal as well behaved as produced by the generator G6

(Figure 1.26) with three segments of equal length 0.5 and vertices
p0 = (0, 0)
p2 = (0.75,

p1 = (0.25,

3
4 )

3
4

0.4330)

p3 = (1, 0).

The first few sets A(k) seem to resemble a cauliflower, but eventually they settle down
in form of a crystal with bubbling inclusions (Figure 1.27).

Figure 1.26. The generator G6 .

21

Section 1.1 The game of deleting and replacing

Figure 1.27. The approximating set A(8) for the fractal with generator G6 .

Finally, we can produce any weird fractal we wish by prescribing a suitably weird
generator G7 . It suffices to let an error enter in the coordinates of the second example
of Section 1.1.4.3 above (Figures 1.28, 1.29):
p0 = (0, 0)
p3 = (0.5, 0)

p1 = (0.25, 0)
p4 = (0.875,

p2 = (0.375,

3
8 )

p5 = (0.75, 0)

p6 = (1, 0)

Figure 1.28. The generator G7 .

3
8

0.2165)

22

Chapter 1 Fractals and dimension

Figure 1.29. The approximating set A(6) for the fractal with generator G7 .

1.1.5

Space-filling curves

Surprising as it may sound, by a suitable choice of the generator it may happen that the
resulting fractal, without losing its character as a curve, may pass through all points of
an open set in the plane R2 . Such a curve (here we only consider sets in R2 ) will be
called space-filling.
1.1.5.1 The half square
Let the initiator A(0) again be the unit interval. Suppose the generator (Figure 1.30)
connects the points
p0 = (0, 0)

p1 = (0.5, 0)

p3 = (0.5, 0)

p4 = (1, 0).

p2 = (0.5, 0.5)

Considered as a curve, the polygon A(1) runs twice, in different directions, through
the segment p1 p2 . The open triangle p0 p2 p4 furnishes the open set condition since it

Figure 1.30. The generator of the half-square fractal.

Section 1.1 The game of deleting and replacing

23

4
contains its two open halves p0 p1 p2 and p2 p3 p4 . Formula (1.7) gives dimS (A) = log
log 2 =
2. Indeed, repeated application of the map f produces all segments of rectangular
meshes of width 21k parallel to the coordinate axes ofthe plane R2 which are contained
in the closed triangle p0 p2 p4 (Figure 1.31). Hence
k=1 A(k) is dense in this triangle
and every point therein appears as a limit point, for some t [ 0, 1 ], of the sequence
(k) (t) as described in Section 1.1.2. Figure 1.32 represents the graph of the function
(7) for 0 x 0.675.

Figure 1.31. The approximating set A(5) for the half-square fractal.

Figure 1.32. A closer approximation (A(7) ) of the half-square fractal; the graph
of (7) is drawn for 0 x 0.675; the cross marks the point (7) (0.675).

Although this means that indeed we can pass through every point of the whole
triangle p0 p2 p4 on a continuous curve, and although the construction is surprisingly
simple, from an aesthetic point of view one might consider it not so satisfying that
apart from the unit interval every segment in A(k) has to be passed through twice. Can
we do better than that?

24

Chapter 1 Fractals and dimension

1.1.5.2 The crab

Take an equilateral triangle above the x-axis with basis [ 0, 12 ] and one below the x-axis
with basis [ 12 , 1 ] and delete both bases (Figure 1.33). The resulting generator connects
consecutively the points
p0 = (0, 0)
p3 = (0.75,

p1 = (0.25,

3
4 )

3
4 )

p4 = (1, 0).

p2 = (0.5, 0)

(1.8)

Figure 1.33. The generator A(1) of the crab.

Although there is no candidate set presenting itself for the open set condition, if for
some good reason formula (1.7) would still be applicable, then for the corresponding
fractal A it would give
4
dimS (A) = log
log 2 = 2
which nourishes the suspicion that A might in some sense be genuinely a planar set.
In order to find out more about this, denote by g0 , g1 and g2 the threestraight lines
through the origin (0, 0) of the plane R2 with slopes 0 (the x-axis) and 3. Consider
the triangular mesh T(k) generated by the three families of straight lines parallel to g0 ,

g1 and g2 and having, as pairwise distances, integral multiples of k = 2k+31 . Let us

call the equilateral triangles of side length 21k in T(k) its cells. The generator is incident
with T(1) , i.e. its four segments are sides of cells of T(1) . An application of f not only
to the generator but to each segment of the mesh T(1) maps T(1) onto T(2) . It is readily
seen that more generally we have f (T(k) ) = T(k+1) . As a consequence, we get that
A(k) = f (k) (A(0) ) is incident with T(k) .
Fix a cell B(k) in T(k) (painted red) and surround it by the union C(k) of all outside
adjacent cells of T(k) (painted blue; in fact, B(k) is incident with C(k) ) (Figure 1.34).
Let us call two sets of segments essentially disjoint if the two corresponding unions
of open component-segments are disjoint. An application of f to any segment outside
of C(k) does not in any way afflict B(k) . Inspection of the action of f on the segments
constituting T(k) we find:

Section 1.1 The game of deleting and replacing

25

Figure 1.34. The cell B(k) (red) and the corresponding set C(k) (blue).

(a) the image of a segment does not depend on its orientation;

(b) images of disjoint open segments are essentially disjoint, images of adjacent segments produce adjacent segments in T(k+1) ;
(c) two pairs of adjacent segments which do not traverse each other have images
which do not traverse each other;
(d) the set f (C(k) ) contains all segments of all cells of T(k+1) inside the image of the
perimeter of C(k) ;
(e) the set f (C(k) ) contains the belt C(k+1) of all cells in T(k+1) adjacent to and outside
of B(k) , i.e. its perimeter has distance k+1 from B(k) ;
(f) by induction, assertion (e) implies that the set f (m) (C(k) ) contains all segments of
all cells in C(k+m) , i.e. its perimeter has distance k+m from B(k) ;
(g) as a further consequence of (e), for every cell B(k+1) in T(k+1) within B(k) , a belt
of all neighbouring cells in T(k+1) is again contained in f (C(k) );
(h) again by induction, all assertions (a)(g) hold for every cell B(j) of T(j) (j > k)
within B(k) .
Several conclusions may be drawn from these facts:
() Since A(1) , considered as a curve, passes through its segments only once and does
not traverse itself, by (b) and (c) the same is true for A(k) for all k > 1.
( ) If, for some k N, a set C(k) as described above is incident with A(k) (i.e. A(k) ,
considered as a curve, runs through all segments of C(k) ), then by (d) and (f), for
all j > k the set A(j) contains the segments of all cells of T(j) lying within the
corresponding cell B(k) of T(k) .
( ) In the situation described in ( ) and for j > k , by (h) for every cell B(j) of T(j)
lying within B(k) a set C(j) of all cells of T(j) adjacent to B(j) is again incident
with A(j) .

26

Chapter 1 Fractals and dimension

( ) Recall from Section 1.1.2 how the fractal A may be considered as a curve ([ 0, 1 ])
appearing as set of limit points of the approximating curves A(k) = (k) ([ 0, 1 ]).
Again in the situation described in ( ), every point within B(k) is such a limit
point (possibly for more than one parameter value t) and therefore must belong to
A. Consequently, the curve A runs through every point of the set B(k) , a set with
a non-empty interior.
Back to our special generator (1.8): the set A(5) = f (5) (A(0) ) indeed (in several
places in the area marked yellow) contains a set C(5) of cells in T(5) as described above
(Figure 1.35). Conclusion: the fractal A is space-filling. Still, this fact is not limited
to the central cells of the just identified sets C(5) in T(5) . Since the resulting fractal
is self-similar, it contains open sets in the neighbourhood of each of its points. As a
continuous image of the unit interval it is closed. So it is the connected closure of
a bounded open set in the plane R2 . It is only due to blooming imagination that it
is envisioned as a crab with scissors and legs marching sideways through the picture
(Figure 1.36).

Figure 1.35. The approximating set A(5) for the crab. The areas indicated in
yellow contain sets of type C(5) as described in the text.

The crab displays one more interesting feature related to what has been discussed
in Section 1.1.4.4. Due to the positive slopes of the generator in the points (0, 0) and
(1, 0), again it spirals about each vertex appearing in any set A(k) (k 1), but now
the left (blue) and the right (red) spiral both turn counterclockwise without interfering
with each other, the blue one leading into the vertex, and the red one leading out.

27

Section 1.1 The game of deleting and replacing

Figure 1.36. A closer approximation (A(8) ) of the crab fractal.

1.1.5.3 The H EIGHWAY-H ARTER dragon

A very simple generator consists of only two segments, the small side of a right-angled
triangle with hypotenuse [ 0, 1 ] (Figure 1.37). Its vertices are
p0 = (0, 0)

p1 = (0.5, 0.5)

p2 = (1, 0).

Figure 1.37. The generator A(1) of the dragon.

The initiator is again the interval [ 0, 1 ]. A mapping f as defined so far produces

a fractal A somewhat similar to the one generated in Section 1.1.4.4. It might be
log2
=
confusing that an application of formula (1.7) would seem to give dimS (A) = log
2
2, but on second thought we have not verified the open set condition, and the loops
appearing in A indicate that we shall not be able to do so if we try.
To complicate matters we shall now agree in the definition of the map f that it
should replace each segment by the suitably diminished generator alternatingly to the
right and to the left (cf. [Gardner, 1967], [Gardner, 1977], [Davis and Knuth, 1970]).
Now it turns out that an application of f (2k) to A(0) produces a set A(2k) incident with
the mesh T(k) generated by the two families of lines parallel to the coordinate axes and
distant from each other an integral multiple of 21k (Figures 1.381.40).

28

Chapter 1 Fractals and dimension

Figure 1.38. The approximating set A(2) for the dragon.

Figure 1.39. The approximation set A(3) of the dragon.

Figure 1.40. The approximating set A(4) of the dragon.

Section 1.1 The game of deleting and replacing

29

In fact an inspection similar to the one conducted in Section 1.1.5.2 reveals that
indeed the resulting fractal A is space-filling: the set B is to be a replaced by a square
of side length 21k , the set C(2k) by the set of all segments belonging to the square cells
neighbouring B , but one has to be careful to mark the vertices of A(2k) where the
generator A(2) of f (2) has to start replacing two adjacent segments in order to produce
A(2k+2) . The set A(8) contains a configuration of the form C(8) (Figure 1.41), and
A(2j) (j 4) contains all segments of cells in T(j) lying within the middle cell B .
Therefore every point within the closed square B as a limit point of the sequence of
sets A(2k) belongs to A. Again, for the reasons given in Section 1.1.5.2, the fractal A
is the connected closed hull of its interior.

Figure 1.41. The approximating set A(8) for the dragon. Indicated in yellow is
a set of type C(8) as described in the text.

The tail, the claws and fins, and the spiralling tongue might have inspired the choice
of the name dragon for this fractal (Figure 1.42).

30

Chapter 1 Fractals and dimension

Figure 1.42. A closer approximation (A(16) ) of the dragon fractal. The white
spots mark the screen pixels which are not yet visited by the piecewise linear
curve furnishing the point set A(16) .

1.1.5.4 The P EANO curve

Although the full square (just as the half square) is hardly eligible for the name fractal, there is a way, much more elegant than done in Section 1.1.5.1, by the process of
repeatedly applying a suitable generator to the unit interval as initiator, to produce a
curve A which passes through every point of this full square [Peano, 1890]. (A related
construction avoiding multiple points is due to H ILBERT, cf. [Hilbert, 1891].) This
generator A(1) (Figure 1.43) connects the points
p0 = (0, 0)

p1 = ( 13 , 0)

p2 = ( 13 , 13 )

p3 = ( 23 , 13 )

p4 = ( 23 , 0)

p5 = ( 13 , 0) (= p1 )

p6 = ( 13 , 13 )

p7 = ( 23 , 13 )

p8 = ( 23 , 0) (= p4 )

p9 = (1, 0).

It is readily seen that the open square Q with the diagonal p0 p9 furnishes the open
9
set condition and that for the resulting set A, formula (1.7) gives dimS (A) = log
log 3 = 2.
In fact, it is also readily seen that every point of the closure of Q is a limit point of the
sequence of sets A(k) (Figure 1.44).
So the method employed in the study of the KOCH curve should produce a continuous function on the unit interval passing through every point of the closed square Q.
Let us try to construct it explicitly.
In order to simplify the computation let us turn the two coordinate axes clockwise
about 4 so that the y -axis contains the point p2 , and redefine the side of Q to have unit

Section 1.1 The game of deleting and replacing

Figure 1.43. The generator A(1) for the P EANO curve.

Figure 1.44. The approximating set A(2) of the P EANO curve.

31

32

Chapter 1 Fractals and dimension

length. As in Section 1.1.1 we define the piecewise linear mapping (1) on [ 0, 1 ] to

map each of the intervals [ j9 , j+9 1 ] linearly onto the segment Aj of A(1) (0 j 8).
As a curve, (1) ([ 0, 1 ]) passes through the diagonals of consecutive sub-squares Qj
with side length 13 . Due to the varying inclination of segments in A(1) we attach to
each sub-square Qj an integer mj (mod 4) as inclination coefficient, which indicates
that the segment Aj serving as diagonal of Qj is turned clockwise about a hook of mj 2
with respect to the position of p0 p9 . Thus we have
m0 = 0

m1 = 3

m2 = 0

m3 = 1

m4 = 2

m5 = 1

m6 = 0

m7 = 3

m8 = 0.

Let the point p = (x, y) in the closed square Q be given and consider the (if possible,
finite) triadic decompositions
x =

x
i=1

y =

3i

yi
i=1

3i

= 0.x1 x2 . . . xi . . .

(0 xi 2),

= 0.y1 y2 . . . yi . . .

(0 yi 2).

The digit x1 determines the column of three sub-squares Qj in which p lies, the digit y1
determines the row of three sub-squares in which p lies. In fact, the subsegment Aj1 of
A(1) forming the diagonal of the sub-square Qj1 containing the point p (if p belongs to
more than one sub-square Qj , then j1 is taken to be minimal) has consecutive number

y 1
j1 = 5 y1

6 + y1

if x1 = 0,
if x1 = 1,
if x1 = 2.

On A(1) = (1) ([ 0, 1 ]) and within Qj1 we approximate the point p by the point p(1) =
(1) ( j91 ), the initial point of the segment Aj1 .
An application of f to A(1) replaces the diagonal Aj1 of Qj1 with a copy similar to
A(1) but diminished by a factor 13 , producing inside of Qj1 a string of segments Aj1 ,j
(0 j 8). Again this corresponds to a subdivision of Qj1 into nine sub-squares
Qj1 ,j of side length 312 each. As before, the mapping (2) is defined linearly on every
1
interval [ j91 + 9j2 , j91 + j+
92 ] onto Aj1 ,j . Again we want to approximate the point p Qj1
by the initial point of the segment Aj1 ,j2 which is the diagonal of the sub-square Qj1 ,j2
containing p (if p is contained in more than one of these sub-squares we take the one
with minimal index j2 ). We proceed as before but in order to find j2 in terms of x and y
we have to take into account the inclination of Qj1 indicated by mj1 . Inside of Qj1 the
index j2 of the sub-square Qj1 ,j2 containing p will again be determined by the triadic

33

Section 1.1 The game of deleting and replacing

digits x2 and y2 , but the role played before by x1 and y1 will now be played in case of
mj1 = 0

by

x
2 = x2

and

y2 = y2 ,

mj1 = 1

by

x
2 = 2 y2

and

y2 = x2 ,

mj1 = 2

by

x
2 = 2 x2

and

y2 = 2 y2 ,

mj1 = 3

by

x
2 = y2

and

y2 = 2 x2 .

Correspondingly,

y2
j2 = 5 y2

6 + y2

if x2 = 0,
if x2 = 1,
if x2 = 2.

On A(2) = (2) ([ 0, 1 ]) the point approximating p is taken to be p(2) = (2) ( j91 + 9j22 ),
the initial point of the segment Aj1 ,j2 . The distance between p and p(2) which both

lie in Qj1 ,j2 is at most 322 . Note that the inclination coefficient of Qj1 ,j2 is mj1 + mj2
(mod 4).
Proceeding by induction we obtain a sequence of continuous piecewise linear map

pings (k) mapping the intervals [ ki=1 n9ii , ki=1 n9ii + 91k ] (0 ni 8) linearly
onto An1 ,...,nk ; furthermore a sequence of segments Aj1 ,...,jk with initial points p(k) =

(k) ( ki=1 9jii ) distant from p at most 3k2 , and finally a sequence of corresponding incli


nation coefficients ki=1 mji (mod 4). For l > k and t [ ki=1 n9ii , ki=1 n9ii + 91k ] the

points (k) (t) and (l) (t) are at most a distance of 3k2 apart since both lie in Qn1 ,...,nk .
Consequently, the sequence {(k) }
1 converges uniformly to a continuous mapping

k=
k
k
and ( i=1 9jii ) = limk ( i=1 9jii ) = limk (k) ( i=1 9jii ) = p. In other
words, we have established that the continuous curve ([ 0, 1 ]) indeed passes through
every given point p Q.

34

Chapter 1 Fractals and dimension

Figure 1.45. A closer approximation of the P EANO curve. The graph of (5)
is drawn for 0 x 0.833. The cross indicates the point (5) (0.833). The
argument x = 0.833 has 9-adic expansion 0.7442233153 . . . .

Figure 1.46. A tessellation of the unit square by sub-squares Qj1 ,j2 (0 ji

8) of side length 312 in the order in which they are passed through by the P EANO
curve. In this order the sub-squares are painted consecutively in the colours
blue, red, green, magenta, cyan, dark grey, light blue, light green, light red. The
tessellation starts in the left corner, going up first in north-east direction, and
ends in the right corner.

Section 1.1 The game of deleting and replacing

1.1.6

35

Short fractal curves

Although the KOCH curve is the image under a continuous mapping of the unit interval
into R2 so we may pass through it in finite time it has infinite length. The same
is valid for the fractal curves we have discussed thereafter, not withstanding the fact
that we have not always stated this explicitly. So rather naturally the question came up
in a course: Are there also fractal curves of finite length? There is a sloppy answer
to it: Yes, the C ANTOR set! But this is not quite satisfying since here we meet the
other extreme: if length should have some meaning for the C ANTOR set, it would
have to be measured by L EBESGUE measure which gives it the value zero. So the next
question asks whether there is a fractal curve whatever this may mean which has in
some natural sense a finite but positive measure?
1.1.6.1 A thick C ANTOR set
Why has the C ANTOR set got L EBESGUE measure zero? Because at every stage
A(k) of its construction we have deleted disjoint subintervals from A(k) of total length
1
A(k+1) of L EBESGUE measure L(A(k+1) ) = ( 23 )k+1 .
3 L(A(k) ), thereby leaving a set 
As k this gives L(A) = L(
k=1 A(k) ) = limk L(A(k) ) = 0. It seems prudent
to delete in a more cautious way. Let any < 13 be given and start out again with the
closed unit interval A(0) . At the k -th stage (k 1), delete from every component interval of A(k1) (there are 2k1 of them) a centrally positioned open subinterval of length

. This leaves a closed set A(k) with L EBESGUE measure 1 ki=1 2i1 22i
1 =
22k1

k

1
1
1 i=1 2i = 1 (1 2k ). The set A = k=1 A(k) is closed, totally disconnected
(thereby topologically zero-dimensional), perfect, nowhere dense, topologically homeomorphic with the C ANTOR set, and has L EBESGUE measure 1 > 0, as close to 1
as desired.
Due to the fact that the total length of the deleted subintervals is not proportional
to the total length of the set from which they are deleted, we do not have exact selfsimilarity within A. So as to dimension we better wait until later.
1.1.6.2 A self-similar fractal curve with positive finite length
Self-similarity of a fractal curve is certainly present if we use the method of initiator
and generator. Still, in the cases considered so far, a continuous piecewise linear generator different from the unit interval A(0) but connecting its end points is bound to
have length larger than 1. With every application of f to A(k) (k 0) the length of
A(k) is multiplied with the length of the generator. Consequently, the length of A(k)
for k eventually exceeds all bounds, and the length of the limiting set A, which
cannot be smaller, must be infinite.
If we want to keep the length of A(k) bounded, our only chance is to cut the unit
interval into finitely many pieces and to place them disjointly from each other. This
will destroy continuity in the resulting fractal, but we have already made peace with
this fact when dealing with the C ANTOR set.

36

Chapter 1 Fractals and dimension

In order to keep things simple, consider the generator A(1) (Figure 1.47) consisting
of two segments p0 p1 and p2 p3 with
p0 = (0, 0)

p1 = (0.5, 0)

p2 = (1, 0.5)

p3 = (1, 1).

Figure 1.47. The generator A(1) for the short fractal curve.

Again the application of the mapping f consists in replacing every subsegment of

the set A(k1) with a suitably diminished copy of A(1) . It is readily seen (e.g. by
induction) that A(k) consists of 2k subsegments of length 21k each, so the total length
of the subsegments of A(k) always remains 1. Unfortunately this fact alone does not
allow us to conclude that the limiting set A (Figure 1.48) in some sense also has to have
length 1. First of all, the points of A lie scattered around in the unit square and it is not
clear how to attribute a length to their union. Secondly, even in the continuous case of
a staircase-curve from p0 to p3 with steps of width and height 21k , having total length 2

for every k N, the limiting curve would be the diagonal p0 p3 with length 2.
As an encouraging fact, the open unit square furnishes a set needed for the open
2
set condition, which justifies the computation dimS (A) = log
log 2 = 1. But, as explained
above, the computation of a length by means of L EBESGUE measure breaks down.
Perhaps we can simulate the computation of L EBESGUE measure of a subset of the
real line (by means of coverings by open intervals of arbitrarily small length) and find
coverings of A by means of sets of arbitrarily small diameter and bounded sum of
all diameters. Indeed, the set A(1) is contained in the union P(1) of two closed subsquares Qj (0 j 1)) of the unit square Q sharing its diagonal and having side
length 12 each. Similarly, A(2) is contained in the union P(2) of four closed sub-squares
Qj1 ,j2 (0 ji 1) of Q of side length 14 each, contained in P(1) . By induction we

37

Section 1.1 The game of deleting and replacing

Figure 1.48. A closer approximation (A(9) ) of the short fractal curve.

find that A(k) is contained in the union P(k) of 2k closed sub-squares Qj1 ,...,jk of side
length 21k each and that P(1) P(2) P(k) . The set A consists of limit points
of the sequence of closed sets {A(k) }
k=1 . All of these limits points are contained in


P
and
therefore,
for
each
k

N, in P(k) . In other words, A is covered by the

k=1 (k)

k
2 sub-squares Qj1 ,...,jk mentioned above. The sum of their diameters amounts to 2.
We have just computed an upper bound for the one-dimensional H AUSDORFF measure of A which is defined by
H1 (A) := sup inf
>0

= lim inf
0

diam(Ci ) : A

i=1

i=1


Ci , diam(Ci ) (i N)


i=1

diam(Ci ) : A


Ci , diam(Ci ) (i N) .

(1.9)

i=1

Recall that diam(C) := sup{|x y| : x C, y C}. As will be established later,

on the real line the one-dimensional H AUSDORFF measure coincides with L EBESGUE
measure. So for a legitimate extension H1 of the concepts oflength and onedimensional L EBESGUE measure we have found that H1 (A) 2.
It would be disappointing if it turned out that the H AUSDORFF measure of A would
in fact be zero. In order to obtain a lower bound for H1 (A), observe that each of the
projections px (A(k) ) and py (A(k) ) of the set A(k) upon the x-axis and y -axis respectively consists of disjoint intervals of length at most 21k each (some of them just consist
of one single point), spaced at most a 21k from the neighbouring one. Again this may
be proved by induction: the assertion is true for k = 1; assuming its truth for some k
we note that A(k+1) is obtained by placing a similar copy half the size of A(k) into the

38

Chapter 1 Fractals and dimension

lower left sub-square Q1 and an equal copy, turned about an angle of 2 counterclockwise, into the upper right sub-square Q4 . Together their projections supply the wanted
family of intervals on the x-axis and y -axis. The set of limit points of the sequences of
sets px (A(k) ) and py (A(k) ) as k therefore consists of all points of the unit interval
on the x-axis and y -axis respectively.
As mappings of R2 upon the x-axis and y -axis respectively, the projections px and
py are continuous (the following statements, formulated for px , hold for py as well).
For any two points p and q in the plane, px even satisfies the inequality
|px (p) px (q)| |p q|.


If A i=1 Ci and diam
all i N, then px (A)
i=1 px (Ci ),
(Ci ) for
diam(px (Ci )) and i=1 diam(Ci )
diam
(p
(C
))
.
For
any
set
A R2
x
i
i=1
this implies
inf

diam(Ci ) : A

i=1


Ci , diam(Ci ) (i N)

i=1

inf

diam(Ci ) : px (A)

i=1

sup inf
>0


Ci , diam(Ci ) (i N) ,

i=1

diam(Ci ) : A

i=1


Ci , diam(Ci ) (i N)

i=1

sup inf
>0

diam(Ci ) : px (A)

i=1

H (A) H (px (A)).


Ci , diam(Ci ) (i N) ,

i=1

(1.10)

The sets A(k) converge for k pointwise to the set A. Because of the continuity
of px we get px (A) = limk px (A(k) ) i.e. the set of limit points of the sets pk (A(k) )
as k . We already know this set to be the unit interval. Consequently we obtain
H1 (A) 1.
For anyone already familiar with the concept of H AUSDORFF dimension, our estimates again imply dimH (A) = 1 as an immediate consequence of Definition 1.3.5.
1.1.6.3 The C ANTOR staircase
Deletion of the middle third pieces in the construction of the C ANTOR set does not
mean that these pieces are useless. Recall that in the construction of the C ANTOR
set the (k 1)-th configuration A(k1) consists of 2(k1) disjoint intervals Aj1 ,...,jk1
1
(ji {0, 2}) of length 3k
1 each. The next configuration A(k) is obtained from A(k1)
by deleting in each one of those intervals the open middle third of length 31k . We
shall denote these deleted intervals in their consecutive order B(k),j (1 j 2(k1) ).
We now start with lifting B(1),1 within the unit square to height y = 12 . Connect its
left endpoint b(1),1,0 by a line segment with (0, 0) and its right endpoint b(1),1,1 by a

Section 1.1 The game of deleting and replacing

39

line segment with (1, 1). We shall denote this configuration of three segments C(1)
(Figure 1.49).

Figure 1.49. The approximating set C(1) for the C ANTOR staircase.

In order to construct the next configuration C(2) we now replace the first and the last
segment by the following configurations of three segments each: lift B(2),1 to height
y = 212 and connect its left endpoint b(2),1,0 as before with (0, 0), similarly its right endpoint b(2),1,1 with b(1),1,0 ; lift B(2),2 to height y = 232 and connect its left endpoint b(2),2,0
with b(1),1,1 and its right endpoint b(2),2,1 with (1, 1). The configurations, obtained by
these replacements, like upon C(1) but are no similar copies thereof. Their widths are
1
1
3 of the width of C(1) , but their heights are 2 of the height of C(1) . While the slope of
3
the two connecting segments in C(1) was 2 , the slope of the four connecting segments
in C(2) is ( 32 )2 .
Proceeding by induction, at the k -th step we lift every interval B(k),j (1 j
1
k1
2 ) to height y = 2j
2k and connect its endpoints with the endpoints of the neighbouring lifted x-parallel intervals in C(k1) , thereby arriving at the configuration C(k) .
It consists of 2k 1 x-parallel intervals of total length 1 ( 23 )k and of 2k segments connecting points with horizontal distance 31k and vertical distance 21k ; their slope therefore

is ( 32 )k and their total length is 2k 312k + 212k . The total length of the configuration C(k)

is 1 ( 23 )k + ( 23 )2k + 1 (Figure 1.50).
It is readily seen that C(k) is the graph of a continuous piecewise linear function
k on [ 0, 1 ] which, as k , converges uniformly to a continuous function . The

2k1
graph of has a horizontal tangent at each point of
j=1 B(k),j , an open set
k=1
of L EBESGUE measure 1. On its complement, the C ANTOR set, it has vertical tangents. The function , called L EBESGUEs singular function [Lebesgue, 1905], is almost everywhere differentiable with vanishing derivative, but it is not the indefinite

40

Chapter 1 Fractals and dimension

Figure 1.50. The approximating set C(2) for the C ANTOR staircase.

integral of its derivative (which would be a constant function). The strange feature of
it is that while its tangents are almost everywhere horizontal, it increases, and actually
increases only on the C ANTOR set, a set of L EBESGUE measure zero, from zero to
one. Its graph C is frequently called C ANTORs staircase and, probably because of the
mentioned weird feature, sometimes devils staircase (Figure 1.51).

Figure 1.51. A closer approximation (C(8) ) of the C ANTOR staircase.

It is tame with respect to its length. As sequences of segments connecting points

of C , the configurations C(k) are perfectly legitimate means to compute, in the limit as
k , the length of C as a curve. From what has been computed above this limit is

Section 1.1 The game of deleting and replacing

41




limk 1 ( 23 )k + ( 23 )2k + 1 = 2.
Although C ANTORs staircase tries hard to look self-similar, it is not, due to the
reasons indicated above. This might be an indication that one should not be too strict
with self-similarity when dealing with fractals. In fact we shall meet more general
concepts in the second chapter.

1.1.7

Higher dimensional C ANTOR sets

The construction of the C ANTOR set started with a one-dimensional line segment and
produced a set with a smaller but still positive dimension. We shall now try to see
whether this procedure may be modified so as to be applicable also to initial sets of
greater dimension. If we succeed with some examples in two- or three-dimensional
space then we should have no trouble in trusting that similar fractals may be constructed
in spaces of arbitrary dimension.
1.1.7.1 The S IERPINSKI triangle
Take an equilateral triangle A(0) with side length 1. In fact any other triangle would
do as well, but for symmetry reasons such a setup gives a nice picture. The triangle
A(0) , as all triangles and polygonal sets in Section 1.1.7, will be considered to contain
all points inside its boundary. Connecting the midpoints of the sides by straight line
segments divides A(0) into four equilateral triangles of side length 12 each. Just as done
in the construction of the C ANTOR set, we delete the open middle triangle to obtain
the set A(1) . Continuing in this way with every one of the three equilateral triangles
constituting A(1) , at the k -th step we arrive at a compact set A(k) consisting of 3k

equilateral triangles of side length 21k . The total area of these triangles 
is 43 ( 34 )k .
Taking the intersection of all sets A(k) , we obtain a compact set A =
k=1 A(k) of

3
3 k
two-dimensional L EBESGUE measure limk 4 ( 4 ) = 0 (Figure 1.52). The set
A is called the S IERPINSKI triangle [Sierpinski, 1916]. It is self-similar and the open
triangle A(0) furnishes the set needed for the open set condition. Formula (1.7) gives
3
dimS (A) = log
log 2 1.585.
The construction used above may be applied to any initial triangle in place of A(0) ;
a common version of the S IERPINSKI triangle starts out with an isosceles right-angled
triangle (Figure 1.53).

42

Chapter 1 Fractals and dimension

Figure 1.52. The approximating set A(6) for the S IERPINSKI triangle.

Figure 1.53. The approximating set A(6) for another version of the S IERPINSKI
triangle.

Section 1.1 The game of deleting and replacing

43

1.1.7.2 The L EVY

surface
Recalling how we have come from the C ANTOR set to the KOCH curve, we can try to
imitate this in two dimensions: replace the deleted triangle in A(1) by three faces of a
tetrahedron of side length 12 , thereby obtaining a continuous piecewise linear surface
B(1) consisting of six equilateral triangles.

Figure 1.54. The surface B(1) is depicted in blue, the edges ad, bd and cd of the
pyramid C and its height d d in grey.

Now we repeat what we have just done with every one of these equilateral triangles.
We obtain a continuous polyhedral surface B(2) consisting of 36 equilateral triangles of
side length 14 . Our intention is to continue this process in hope of arriving in the limit
at a self-similar fractal set B . But what about the open set condition, and can we be
sure not to get entangled in some troubling self-intersection of the approximating sets
B(k) ? Fortunately it turns out (and can be checked by elementary geometric considerations as done below) that the open pyramid C , with A(0) as basis and the same top
as the first tetrahedron, satisfies the open set condition: its six copies Cj , contracted
by a similarity factor 12 , built over the three equilateral triangles in A(1) and over the
faces of the middle tetrahedron, have pairwise just a face in common and are all contained in C . Continuation of our construction process therefore concerns each time just
the part of B(k) inside of the contracted copies of C and leaves their interiors disjoint.
By induction we find that the sequence of sets B(k) (which we may also consider as
graphs of continuous functions k defined on A(0) ) converges uniformly to a continuous surface B (we shall discuss convergence of sets in this context in Section 2.1),
cf. [Levy, 1938]. The L E VY surface B is also called a tetrahedral fractal or KOCH
6
pyramid. By formula (1.7) in Section 1.1.3 its dimension is dimS (B) = log
log 2 2.585.

44

Chapter 1 Fractals and dimension

A closer study of the construction of B reveals a story with a surprising conclusion.

In particular we want to compute the three-dimensional volume of the body B bounded
by the fractal surface B and the ground plane (which, in Figures 1.54 and 1.56, coincides with the [x, y]-plane).
To this end, let us have a look at the surface B(1) . E.g. the face bdc of the pyramid
C consists of two halves bd
a and cd
a. Each of these is an isosceles right-angled tricda are
angle. Therefore, so also is their union bdc. Since the three faces adb, bdc and

2
congruent, they consist just of the three half-squares of a cube of side length 2 joined
at the vertex d (Figure 1.55). The height d d of C is easily computed to be 16 , the area
of the basic triangle abc is

3
4 ,

and the three-dimensional volume of C is

1
.
12 2

Figure 1.55. The pyramid C = abcd is part of a cube which consists of four
pyramids congruent with C, attached to the faces abc, abe, bce and cae of a
central tetrahedron abce.

Let us now have a look at the surface B(2) resulting from replacing each of the
cb, b
ac, cb
a, a
bd, b
d with a similar copy of B(1) , reduced
equilateral triangles a
cd, ca
1
in size by the factor 2 . In particular let us have a look at the two copies with basic
triangles bac and cad as in Figure 1.56. The tops d1 and d2 of the tetrahedra a1 b1 c1 d1
and a2 b2 c2 d2 coincide with the midpoint of the line segment bd. As a consequence,
these tetrahedra intersect just in the edge b1 d1 = b2 d2 , but the corresponding open
images of the open pyramid C are disjoint and contained in C , as already remarked
above. By induction, B(2) and every following surface B(k) is also contained in (the

45

Section 1.1 The game of deleting and replacing

closed pyramid) C and there is no overlapping of any newly constructed tetrahedra in

B(k) apart from common edges.

Figure 1.56. The tetrahedra a1 b1 c1 d1 and a2 b2 c2 d2 , appearing in the transition

from B(1) to B(2) , intersect only in the common edge b1 d1 = b2 d2 ; both of them
are contained in the pyramid a
cbd which is part of the pyramid C = abcd and
consists of two similar copies a
b
cd1 and a
d
cd2 of C, reduced to half of its size.

The content of the first tetrahedron ab

cd is 13 163 16 = 4812 . Transition from B(1)
to B(2) adds the contents of six tetrahedra half the size. By induction we find that the
looked-for volume of B is

 2

1
3
3
1
1
1
1+ +
+
.
.
.
=
== ,
3
4
4
48 2

48 2

12 2

precisely the same as the volume of the pyramid C which entirely contains B . As far
entirely fills up the pyramid
as three-dimensional L EBESGUE measure is concerned, B
C , although the surface B is dramatically different from the side faces of C .
We shall not pause here to consider the question whether B is eligible to have any
surface measure. Still, this is the case for every approximating surface B(k) , consisting of 6k congruent equilateral triangles of side-length 21k and therefore having

two-dimensional surface measure L2 (B(k) ) = 6k 22k+3 2 . So limk L2 (B(k) ) = , in

analogy to the behavior of the length of the curves approximating e.g. the KOCH island.
There is one more surprise to it: let us proceed analogously as in the case of the
KOCH island (Section 1.1.4.1), where we applied the construction of the KOCH curve
to every line segment of an equilateral triangle. If now we perform the construction
above with each of the four faces of the tetrahedron abde in Figure 1.55 in an outward

46

Chapter 1 Fractals and dimension

direction, one might have expected to obtain something like a ball covered by spikes.
Instead, the result will be a body, still contained in the original cube but almost filling
it up, since it has the same three-dimensional L EBESGUE measure as the cube. The
boundary, however, is a closed surface consisting of four copies of the fractal surface B .
1.1.7.3 The C ANTOR dust
Take for A(0) the unit square and let s be any positive real number smaller than 12 . In
order to obtain the first set A(1) , delete everything in A(0) which is outside the four
small squares of side length s situated at the four corners of A(0) . At every step we
proceed in the same way with every square so far obtained, reducing the sides of the
original square by the factor s. As a result, the set A(k) A(k1) consists of 4k
k
2k
squares
of side length s each and has total area (2s) . Obviously the limiting set
A = k=1 A(k) , called C ANTOR dust, is self-similar, consisting of four copies similar
to itself but reduced with similarity factor s. The interior of A(0) satisfies the open set
4
condition, and formula (1.7) gives dimS (A) = log
log s .

Figure 1.57. The approximating set A(4) for the C ANTOR dust A with s = 0.4,
dimS (A) 1.513.

The two-dimensional L EBESGUE measure of A is zero in any case, but a special situation happens for s = 14 : we get dimS (A) = 1. Although there is no indication how A
may be considered as a curve, we can try to find out something about one-dimensional
A is covered by the 4k
H AUSDORFF measure H1 (A) as in Section 1.1.6.2. Obviously

squares constituting A(k) , each of which has diameter 4k2 which becomes arbitrarily

small if k is sufficiently large. By (1.9) in Section 1.1.6.2 we obtain H1 (A) 2.

For an estimate from below, consider a somewhat modified C ANTOR dust: in A(1)
move every small square of side length 14 from its original position in the hook of A(0)
a step of length 14 counter-clockwise, say, along the boundary of A(0) to obtain the set
B(1) . At every step we now proceed in the same way with every square constructed

47

Section 1.1 The game of deleting and replacing

and obtain, as before, the sets B(k) with 

the same properties as stated above for the
sets A(k) . In particular, for the set B =
k=1 B(k) we also have dimS (B) = 1 and

1
H (B) 2. Considering again, as in Section 1.1.6.2, the projections px and py
upon the x-axis and y -axis respectively, we see that, for every k N, px (B(k) ) and
py (B(k) ) are the unit intervals on the x-axis and y -axis respectively. Formula (1.10) in
Section 1.1.6.2 implies H1 (B) 1.

Figure 1.58. The approximating set B(2) for the modified C ANTOR dust B with
s = 14 , dimS (B) = 1.

There is not much to be seen of C ANTORs dust set B (and A, for that matter), so
is there a way to identify, at least analytically, the points of this set? To this end, for
x [ 0, 1 ], denote by xk (0 xk 3) the k -th digit in the 4-adic expansion of x, i.e.
x = 0.x1 x2 . . . =

x
k=1

4k

Let Q4 [ 0, 1 ] be the set of 4-adic rational numbers

x =

x
k=1

4k

(n N),

:= I I. The symbol \ denotes set theoretic subtraction,

and let I := I \ Q4 and Q
i.e. I is the complement of Q4 in I . In order to avoid ambiguities, for x Q4 we could
e.g. prefer the finite expansion x = 0.x1 x2 . . . xk of x = 0.x1 x2 . . . xk1 33 . . . .

48

Chapter 1 Fractals and dimension

The four closed squares Bj (0 j 3) constituting the set B(1) may then be
described in the following way:


= (x, y) Q
: x1 = 0, y1 = 2 ,
B0 Q


= (x, y) Q
: x1 = 1, y1 = 0 ,
B1 Q


= (x, y) Q
: x1 = 2, y1 = 3 ,
B2 Q


= (x, y) Q
: x1 = 3, y1 = 1 .
B3 Q
Correspondingly, for any 4-adic digit xk let xk be defined by

2 for xk = 0,

0 for xk = 1,
x
k :=

3 for xk = 2,

1 for xk = 3,
so that

=
Bj Q



: x1 = j, y1 = j .
(x, y) Q

Inductively, for the 4k closed squares Bj1 ,...,jk constituting the set B(k) , we arrive at the
formula


= (x, y) Q
: x1 = j1 , . . . , xk = jk , y1 = j1 , . . . , yk = jk .
Bj1 ,...,jk Q

For the fractal B =
k=1 B(k) this furnishes the formula


= (x, y) : x I,
yk = x
BQ
k , 1 k < .
Thus, at least on I, the set B turns out to
be the graph of a function , defined on all but
x
xk
k

countably many values in [ 0, 1 ] by x =

k=1 4k I  (x) =
k=1 4k (it would be
possible, if also somewhat cumbersome and not necessary for our purposes, to identify
the points in B with abscissas in Q4 ). This fact, together with dimS (B) = 1, nourishes
the suspicion that the set B may have something to do with a short fractal as met in
Section 1.1.6.2. In fact, here is a solution to this puzzle: let the initiator be the unit
interval, and let the generator consist of the four intervals of length 14 each in which
the set B(1) intersects the boundary of the unit square (this indeed reminds us of the
generator used in Section 1.1.6.2). Then the C ANTOR dust set B coincides with the
fractal produced as done in Section 1.1.6.2.
1.1.7.4 The S IERPINSKI carpet
Again let A(0) be the unit square and let n N be odd and greater than 2. The set
A(1) is obtained by removing from A(0) the central square of side length n1 . To obtain
A(2) , in each of the remaining n2 1 small squares of side length n1 delete the central
square of side length n12 and proceed in the same way by induction to construct A(k)

Section 1.1 The game of deleting and replacing

49


for all k N. The limiting set A =
k=1 A(k) (especially in the case n = 3 often
called the S IERPINSKI carpet) is self-similar, and the open unit square helps to satisfy
the open set condition. The set A has two-dimensional L EBESGUE measure L2 (A) =
2
2
1)
limk L2 (A(k) ) = limk ( nn2 1 )k = 0 and dimension dimS (A) = log(n
log n .

Figure 1.59. The approximating set A(4) for the S IERPINSKI carpet A with
log 2
1.893.
n = 3 and dimS (A) = 3log
3

Especially in the case n = 3 a nice modification of what we called the L E VY surface

may be constructed: Replace the deleted small middle square by the 5 faces, above the
basis, of a cube of side length 13 , to obtain the set A(1) consisting of 13 squares of side
length 13 . If in each of these small squares the middle square of side length 312 is again
replaced by 5 protruding faces of cubes of side length 312 , then eight of these just touch
each other pairwise along one edge. Continue this process inductively in the familiar
way to obtain the set A(k) , a surface consisting of 13k1 copies similar to A(1) but
reduced by a similarity factor ( 13 )k1 . The sets A(k) again converge uniformly, in a
way which is intuitively clear but will again be discussed in detail in Section 2.1, to a
self-similar set A, heuristically only reluctantly considered as a surface of infinite area,
with protuberances and caves of arbitrarily small size. The open pyramid C with basis
A(0) and top (0.5, 0.5, 0.5) helps to satisfy the open set condition. Formula (1.7) gives
13
dimS (A) = log
log 3 2.335.
In contrast to the situation in Section 1.1.7.2 where the body under the L E VY surface measure-theoretically filled up the pyramid covering it, and admitting the open
set condition, this is not the case in the present situation. Since no part of A protrudes above the pyramid C , also no part of A protrudes beyond the similar copies of
C reduced by the factor 13 and attached to each of the 13 squares of side length 13 constituting A(1) . But there is empty space for instance between C and the union of these
13 smaller copies of C .

50

Chapter 1 Fractals and dimension

1.1.7.5 The M ENGER sponge

To keep things short consider the unit cube A(0) = {(x1 , x2 , x3 ) : 0 xi 1, 1
i 3} in R3 . In the direction of each of the coordinate axes, remove a tunnel through
A(0) , with cross-section the middle square of side length 13 in the face of the cube
perpendicular to the coordinate axis. The remains of the unit cube may be divided
into 20 smaller cubes of side length 13 each. With every one of these we proceed as
1
before, only
 reduced by a similarity factor 3 , and so on. In the limit we obtain a
set A = k=1 A(k) , called the M ENGER sponge [Menger, 1926], self similar and of
20
dimension dimS (A) = log
log 3 2.727, but of three-dimensional L EBESGUE measure
k
L3 (A) = limk L3 (A(k) ) = limk ( 20
27 ) = 0.

Figure 1.60. The generating set A(1) for the M ENGER sponge. Partially visible
tunnel faces lying in two planes are marked in colours red and blue respectively.

On each of the faces of the original unit cube there appears a copy of the S IERPIN carpet. Transgressing into R4 one could try again to construct a modified object as
in Section 1.1.7.4, but since help from our intuition stops here, we shall stop, too.
SKI

1.2

The box-counting dimension

It would certainly seem desirable to be able to attach a dimension not only to selfsimilar sets in Rn . There are two lines of thought which seem promising: Since in
general the set A is no longer the (more or less) disjoint union of a finite number N (A)
of smaller but similar -copies of A, we simply count the least number N (A) of sets of
N (A)
diameter at most (> 0) which we need to cover A, and we observe the ratio loglog

as 0. In an alternative approach, less obvious but in fact more powerful, we attach

51

Section 1.2 The box-counting dimension

to each d R+ = [ 0, [ a measure on Rn (defined also by means of coverings of sets)

and investigate for which value of d this measure seems specially appropriate for the
set A.
This section is devoted to a discussion of the first mentioned approach, in which we
follow the exposition in [Falconer, 1990]. We start out with the pertinent definitions.
1.2.1 Definition. Let A be a non-empty bounded subset of Rn . Define N (A) to be the
minimal number of subsets of Rn , of diameter not exceeding (> 0), needed to cover
A. The lower and upper box-counting dimension of A are defined respectively by
log N (A)
,
log

dimB (A) := lim inf

0

dimB (A) := lim sup

0

log N (A)
.
log

If both are equal, then

dimB (A) = lim

0

log N (A)
log

is called the box-counting dimension of A.

Obviously, dimB (A) is independent of rotations and translations of the set A. There
are equivalent definitions of dimB (A) which are handier for the actual computation
since they work with restricted classes of sets used in the coverings.
Denote by Z

the set of all integers. Let us agree to call a cube of the form ni=1 [mi , (mi + 1) ]
(mi Z, 1 i n) a -lattice cube. In the proofs below frequent use is made of the
fact that, for c > 0,
log(c)
lim
= 1.
log
0

1.2.2 Theorem. Let

a set A Rn . Then

N (A)

be the minimal number of -lattice cubes needed to cover

dimB (A) = lim inf

0

log N (A)
,
log

dimB (A) = lim sup

0

log N (A)
.
log

Proof. A -lattice cube Q has diameter n. Let <

and therefore
dimB (A) lim inf
0

Then Nn (A) N (A)

log N (A)
,
log

dimB (A) lim sup

0

1 .
n

log N (A)
.
log

On the other hand, every set C of diameter at most can be covered by 3n -lattice
cubes (take a point in C , cover it by a -lattice cube Q and its two neighbours in

52

Chapter 1 Fractals and dimension

direction of every coordinate axis; fill up this n-dimensional cross to a cube of side
length 3 : it will cover the whole set C ). Therefore N (A) 3n N (A). This gives us
log N (A)
,
log

dimB (A) lim inf

0

dimB (A) lim sup

0

log N (A)
.
log

Note that again there is no requirement as to how the coordinate system (and therefore the lattice) should be positioned with respect to the set A. In Theorem 1.2.2 the
number N (A) may still be replaced e.g. by:
(a) the smallest number of arbitrary cubes of side length covering A,
(b) the smallest number of balls with radius covering A,
(c) the largest number of disjoint balls with radius and center in A.
We omit the proofs since 1) they run along the lines of the proof given for Theorem 1.2.2, and 2) there will be no need for us to use any of the assertions (a), (b), or (c).
However, it will turn out to be useful that one may also restrict the family of numbers
approaching 0:
1.2.3 Theorem. Suppose the sequence of positive real numbers {k }
k=1 converging to
zero satisfies, for some c ]0, 1[ , the inequalities c k > k+1 ck (1 k < ).
Then for any set A Rn one has
dimB (A) = lim inf
k

log Nk (A)
,
log k
log Nk (A)
.
log k

dimB (A) = lim sup

k

Proof. We shall make use of the facts that 1 >

inequalities
dimB (A) lim inf
k

k+1
k

k+1
k

log c. The

log Nk (A)
,
log k

dimB (A) lim sup

k

c and 0 > log

log Nk (A)
log k

are direct consequences of Definition 1.2.1. For the inequalities in the other directions,

53

Section 1.2 The box-counting dimension

given ]0, 2 [ , choose k such that k+1 < k . Then

log Nk (A)
log Nk (A)
log N (A)

1
log
log k+1
log k log k+
k

dimB (A) = lim inf

0

log Nk (A)
,
log k log c

log Nk (A)
log N (A)
lim inf
,
log
log k
k

log Nk+1 (A)

log Nk+1 (A)
log N (A)

1
log
log k
log k+1 + log k+

log Nk+1 (A)

,
log k+1 + log c

log Nk+1 (A)

log N (A)
dimB (A) = lim sup
lim sup
.

log

log k+1
k
0

The hypothesis of Theorem 1.2.3 is satisfied especially if k = ck for some c

]0, 1[ . This allows us to verify that the self-similarity dimensions of the various fractals
which we have got to know so far indeed coincide with their box-counting dimensions:
Let us start with the C ANTOR set A. Choose k = ( 13 )k . Then Nk (A) = 2k and
2
log 2
dimB (A) = limk kk log
log 3 = log 3 .
Let now A denote the KOCH curve as in Section 1.1.2. As pointed out in Section 1.1.3 and illustrated in Figure 1.3, the set A may be covered by a sequence of
k
4k closed copies of triangles {Vk,i = ak,i bk,i ck,i }4i=1 similar to the isosceles triangle
V = abc, reduced by the similarity factor k = 3k , and the interiors of the triangles
Vk,i (1 i 4k ) are disjoint. Since diam(Vk,i ) = k (1 i 4k ) this implies
dimB (A) = lim sup
k

Nk (A)
log 4

.
k
log 3

Of course other covering sets with diameter less than or equal to 3k could compete
with the sets Vk,i , but in any case they would have to cover the vertices of each of the
triangles Vk,i . It is readily seen that a closed disc D of radius k with center ak,i meets
at most the interiors of Vk,i2 , Vk,i1 , Vk,i and Vk,i+1 . Any set of diameter less than or
equal to k containing ak,i is contained in D. As a consequence we get Nk (A) 14 4k
and
log Nk (A)
log 4
k log 4 log 4
= lim
lim inf
= dimB (A).
log 3
k log 3
log
k

log 4
log 3 ,

Therefore dimB (A) =

the same value as the one which has been calculated as
dimS (A).
What has the open set condition been good for? It has provided for Nk to be essentially the number of segments in A(k) . If we consider a modified KOCH curve with
generator G4 as in Section 1.1.4.4, then the loops appearing at the vertices admit coverings with sets of diameter less than or equal to k which cover more than one segment
of an approximating set A(k) at the same time. Let NS k (A), as in Section 1.1.3, denote the number of segments in A(k) , which is also the number of similar copies of A
with similarity factor k = (S)k = ( 13 )k contained in A. The just mentioned observation implies Nk (A) NS k (A). Formula (1.7) is not applicable any more and the
log 2
S (A)
dimension of A may well be smaller than N(S)
= 2log
3 .

54

Chapter 1 Fractals and dimension

Let us have a look at the higher-dimensional fractals considered in Section 1.1.7:

The dimension of the S IERPINSKI triangle A (Section 1.1.7.1) is most conveniently
computed if we consider it with initiating set A(0) not an equilateral triangle but half the
unit square to the left below the diagonal (Figure 1.53). Taking k = 21k and relying on
3
log 3
Theorem 1.2.2, we obtain Nk = 3k and dimB (A) = limk kk log
log 2 = log 2 1.585.
For the C ANTOR dust A (Section 1.1.7.3) we take k = sk < ( 12 )k . Again an
k log 4
log 4
application of Theorem 1.2.2 gives dimB (A) = limk k
log s = log s . In case
of the S IERPINSKI carpet A (Section 1.1.7.4), taking n > 2 odd and k = n1k , we
2

1)
obtain by the same token dimB (A) = log(n
log n . Finally, for the M ENGER sponge A
20
(Section 1.1.7.5), taking k = 31k , we arrive at Nk = 20k and dimB (A) = log
log 3
2.727.
We sum up some important properties of the box-counting dimension. (Recall that
dimB (A) is only defined for bounded subsets A Rn .)

1.2.4 Theorem.
(a) If E F , then dimB (E) dimB (F ) and dimB (E) dimB (F ). (monotony)
(b) If Q is a non-empty cube in Rn , then dimB (Q) = n.
(c) If E Rn is any bounded set, then dimB (E) n.
(d) If E Rn is open, then dimB (E) = n.
(e) dimB (E F ) = max(dimB (E), dimB (F )). (stability)
(f) If the function f : Rn Rm is L IPSCHITZ, i.e. if for some c > 0,
|f (x) f (y)| c|x y|

for all x Rn and all y Rn ,

and if E Rn , then dimB (f (E)) dimB (E) and dimB (f (E)) dimB (E).
(g) If E is the closure of E Rn , then dimB (E) = dimB (E) and dimB (E) =
dimB (E).
Proof.
(a) The proof is straightforward.
log N 

(b) If the side of Q has length s, take k = 2sk . Then Nk = (2k )n and limk log kk
nk log 2
= limk k log
2log s = n.
(c) Take any cube Q containing E and apply (a) and (b).
(d) Take any cube Q contained in E and apply (a) and (b).
(e) The inequality dimB (E F ) max(dimB (E), dimB (F )) follows from (a). In
order to check the inverse inequality we argue as follows:
N (E F ) N (E) + N (F ) 2 max(N (E), N (F ))
log N (E F )
log 2 + max(log N (E), log N (F ))

log
log

dimB (E F ) max(dimB (E), dimB (F )).

55

Section 1.3 The H AUSDORFF dimension

(f) If E
N (E)

N (E)

Ci and diam(Ci ) for 1 i N (E), then one has f (E)

f (Ci ) and diam(f (Ci )) c (1 i N (E)). This implies Nc (f (E))
i=1
N (E) and
i=1

log Nc (f (E))
log N (E)

.
log c
log c log

Taking lim inf and lim sup on both sides furnishes the assertion.
(g) If E is covered by N closed sets of diameter at most , then the same sets cover
E , i.e. N (E) = N (E).
The last assertion (g) implies, e.g. that the countable set of rational numbers in [ 0, 1 ]
has the same dimension as the interval [ 0, 1 ], namely one, while the dimension of one
single point is zero. This is disturbing for everybody familiar with measure theory. It
is one of the reasons for looking for a concept of dimension which coincides as much
as desirable with the box-counting dimension and still satisfies the assertions (a)(f) of
Theorem 1.2.4 but avoids assertion (g).

1.3

The H AUSDORFF dimension

In order to be able to define H AUSDORFF dimension one first has to get acquainted
with H AUSDORFF measure Hs , which depends on a non-negative real parameter s, its
dimensionality. For integral values of s, Hs turns out to coincide with s-dimensional
L EBESGUE measure. We have already encountered the one-dimensional H AUSDORFF
measure in Section 1.1.6.2. Roughly it measures a subset A Rn by trying to cover A
by countably many sets Cj (1 j < ) of diameter not exceeding an arbitrarily small
> 0 in such a way that the sum of the s-dimensional measures of s-dimensional
balls of the same diameter as Ci gets as small as possible. Obviously this imitates the
definition of L EBESGUE measure at least of subsets of R, but the terms between the
quotes have yet to be defined. There is a formula which does this: the quantity
(s) :=

s/2
( s2 + 1)

(s 0)

turns out to furnish, for integral s, precisely the s-dimensional L EBESGUE measure of
the s-dimensionalunit ball ( denoting the gamma function). Using the well-known
formulas ( 12 ) = , (1) = 1 and (t + 1) = t(t) we get
(0) = 1

(1) = 2

(2) =

(3) =

4
3 .

It is not quite clear how we should imagine a 12 -dimensional unit ball, but according to
1/4
the formula above its 12 -dimensional measure would be ( 12 ) = (
5 . The rough de4)
scription above (of Hs ) is made precise in the following way.
1.3.1 Definition. Suppose A Rn , 0 s < and 0 < < . Then







diam(Cj ) s
Hs (A) := inf (s)
:
A

C
,
diam
(C
)

.
j
j
2
j=1

j=1

56

Chapter 1 Fractals and dimension

The s-dimensional H AUSDORFF measure of A is defined by

Hs (A) := lim Hs (A) = sup Hs (A).
0

>0

For convenience, define H (A) = for s < 0.

s

Note again that Hs is invariant under rotation and translation (this is a special case
of the assertion (e) in Theorem 1.3.3 below). Properly speaking, Hs is an outer measure
on Rn , i.e. apart from being non-negative and attributing measure zero to the empty set,
n
for any sequence of subsets {Ak }
k=1 R it satisfies the inequality




Hs
Ak
Hs (Ak ).
k=1

k=1

As usual, the sets A measurable with respect to Hs are those which satisfy C ARATHEO DORY s condition
Hs (E) = Hs (E A) + Hs (E \ A)

E Rn .

(The symbol turning up above and in Theorem 1.3.3 denotes the logical for all.)
It turns out that all B OREL sets are Hs -measurable. A proof of the theorem which
collects these facts (Theorem 1.3.2) may be found e.g. in [Evans and Gariepy, 1992,
p. 61], a text which also inspired the following exposition.
1.3.2 Theorem. For every s 0 and every n N, the H AUSDORFF measure Hs is an
outer B OREL measure on Rn .
1.3.3 Theorem (Properties of the H AUSDORFF measure).

(a) H0 is the counting measure, i.e. H0 (A) = pA 1.
(b)
(c)
(d)
(e)

H1 (A) = L1 (A)
H (A) = 0
s

A R1 .

A Rn if s > n.

Hs (A) = s Hs (A)

for > 0.
H (LA) = H (A) if L is an isometry in Rn , i.e. |Lx Ly| = |x y| for all
x Rn and all y Rn .
(f) Hn (A) = Ln (A) A Rn .
s

Proof.
(a) Let p Rn . Given any > 0 by definition we have H0 ({p}) = 1.
(b) Let A R1 and > 0 be given. We shall first show L1 (A) H1 (A) and then
L1 (A) H1 (A). This will imply L1 (A) = H1 (A) = H1 (A).



L1 (A) = inf
(j j ) : A
]j , j [
j=1

= inf

j=1

H1 (A).

j=1

(j j ) : A


j=1

]j , j [, 0 < j j

57

Section 1.3 The H AUSDORFF dimension


For the converse inequality suppose A
j := inf Cj , j := sup Cj .
j=1 Cj ,

We have diam(Cj ) = j
j , and therefore





j
j
H1 = inf (1)
:
A

[
j , j ], j
j
2
j=1

= inf

j=1

(j
j ) : A

j=1


[
j , j ], j
j .

j=1

Let now also > 0 be given and suppose the cover {Cj }
j=1 of A satisfies

1 (A) + . Let :=
(

)
<
H

:=
j + 2j+
j
j
j
j
1 . Then

j=1 j
2j+1
we have
A

]j , j [,

j=1

L1 (A)

(j j )

j=1

(j
j ) +

j=1

j=1

< H1 (A) + 2.

2j

(c) We cover the n-dimensional unit cube Q = [ 0, 1 ]n Rn by mn smaller cubes

1
Qm,k (1 k mn ) with side length m
(m N) and diameter mn . Given > 0

we choose m large enough to satisfy mn < . For s > n we get

m  s

n
n

Hs (Q)

(s)

k=1

2m

= (s)

ns
.
2 msn
s

For m the last term goes to zero. We conclude

Hs (Q) = 0

for all > 0,

Hs (Q) = 0.

Since Rn is a countable union of translated unit cubes this implies Hs (Rn ) = 0.

(d), (e) If {Cj }

j=1 is a covering of A, then {Cj }j=1 and {LCj }j=1 are coverings of
A and LA respectively.

(f) See e.g. [Evans and Gariepy, 1992, p. 70].

As a function of s, the H AUSDORFF measure of a set A behaves somewhat cholerically.
1.3.4 Theorem. Let A Rn be given and suppose 0 s < t < . Then:
(a) Hs (A) < = Ht = 0.
(b) Ht (A) > 0

Hs = .

58

Chapter 1 Fractals and dimension

Proof.
(a) Let > 0 be given and suppose the covering {Cj }
j=1 satisfies diam(Cj )
(1 j < ) and
(s)




diam(Cj ) s
j=1

Hs (A) + 1 Hs (A) + 1.

We conclude
Ht (A) (t)




diam(Cj ) t
j=1

diam(Cj )
(t)
(s)
2
(s)
j=1

(t)
(Hs (A) + 1)
(s)

s  diam(C ) ts
j

 ts

Letting 0 we obtain Ht (A) = lim0 Ht = 0.

(b) The assertion (b) is an immediate consequence of (a).
On second thought, the assertion of Theorem 1.3.4 is not so surprising. We cover
the set A with infinitely many sets of small diameter, eventually all smaller than < 1,
and essentially add up the s-th powers of these diameters. If s gets smaller, then these
powers rise in direction of the value 1 which evidently makes the sum diverge. If s gets
bigger, then the s-th powers decrease and their sum tends to vanish as 0. Anyway,
the theorem has an interesting consequence: For every set A Rn there exists an
s0 n with the property that

for s < s0 ,
s
H (A) =
0 for s > s0 .
This seems to be the correct value of s for the set A as far as H AUSDORFF measure
is concerned. No wonder we take this for the H AUSDORFF dimension of A.
1.3.5 Definition. The H AUSDORFF dimension dimH (A) of a set A Rn is defined by
dimH (A) := sup{s : Hs (A) = } = inf{s : Hs (A) = 0}.

Let us compare now H AUSDORFF dimension (defined for all subsets A Rn ) with
box-counting dimension.
1.3.6 Theorem (Properties of the H AUSDORFF dimension).
(a) A B = dimH (A) dimH (B). (monotony)
(b) dimH ([ 0, 1 ]n ) = n.

59

Section 1.3 The H AUSDORFF dimension

(c) A Rn

dimH (A) n.

(d) If A R is open, then dimH (A) = n.


(e) dimH (
i=1 Ai ) = sup1i< dimH (Ai ).
n

(countable stability)

(f) If the function f : R R is L IPSCHITZ (cf. Theorem 1.2.4(f)), then

dimH (f (A)) dimH (A) for all A Rn .
n

(g) If A Rn is countable, then dimH (A) = 0.

Proof.
(a) A B implies Hs (A) Hs (B). Therefore we have dimH (A) = sup{s :
Hs (A) = } dimH (B) = sup{s : Hs (B) = }.
(b) Hn ([ 0, 1 ]n ) = Ln ([ 0, 1 ]) = 1. By Theorem 1.3.3(f) we have Hs ([ 0, 1 ])n = 0 for
s > n and therefore dimH ([ 0, 1 ]n ) = inf{s : Hs ([ 0, 1 ]s ) = 0} = n.
(c) For s > n we have Hs (Rn ) = 0. This implies dimH (A) n.
(d) This is a consequence of (a), (b) and (c) in combination with Theorem 1.3.3(f):
Every open subset of Rn contains a coordinate-parallel small cube.
(e) An immediate consequence of (a) is the inequality
dimH




Ai

i=1

sup dimH (Ai ).

1i<

In order to show the reverse inequality

it is enough to prove that for every s >
sup1i< dimH (Ai ) we have dimH ( i=1 Ai ) s. Indeed, the supposition implies Hs (Ai ) = 0 for all i and therefore also Hs (
i=1 Ai ) = 0.
(f) Suppose
|f (x) f (y)| c |x y|

for all (x, y) Rn Rn .

We shall show that for all s 0 and for all A Rn this implies
Hs (f (A)) cs Hs (A).

(1.11)

Consequently, for Hs (A) = 0 we also have Hs (f (A)) = 0 and the assertion will

follow. Let > 0 be given and let {Cj }

j=1 be a -covering of A. Then {f (Cj )}j=1
is a c -covering of f (A) and



diam(f (Cj )) s
j=1

cs




diam(Cj ) s
j=1

s
(f (A)) cs Hs (A).
Hc

Letting 0 we obtain (1.11).


0
s
(g) Suppose A =
i=1 {ai }. We have H ({ai }) = 1 and therefore H ({ai }) = 0
for all s > 0 and i 1. This implies dimH ({ai }) = 0 and by (e) dimH (A) =
sup1i< dimH ({ai }) = 0.

60

Chapter 1 Fractals and dimension

We see that H AUSDORFF dimension competes favorably with box-counting dimension and can distinguish between a countable dense set and its closure. So the next
theorem will not come as a surprise.
1.3.7 Theorem. For every bounded subset A Rn one has dimH (A) dimB (A).
Proof. For dimH (A) = 0 there is nothing to prove. We shall show
s < dimH (A)

The hypothesis implies

s dimB (A).

lim Hs (A) = Hs (A) = .

0

For sufficiently small > 0 we therefore have Hs (A) > 1. Let such a < 1 be given.
We can cover A by N (A) sets of diameter not exceeding . Consequently,
 s

1 < Hs (A) (s) N (A) 2 ,

0 < log Hs (A) log
s log < log
s <

(s)
+ log N (A) + s log ,
2s
(s)
+ log N (A),
2s

log N (a) + log (s) s log 2

,
log

s lim inf
0

log N (A)
= dimB (A).
log

H AUSDORFF and box-counting dimension are peacefully joined in the theorem on

which we have already relied so often when trying to get more insight in the structure
of self-similar fractals. The proof is rather elaborate, resting on non-trivial parts of
measure theory, and is therefore not presented here.
1.3.8 Theorem. Suppose Si (1 i m) are similarity mappings with similarity
factors si < 1. Suppose the open set condition is satisfied and suppose the bounded
non-empty set A Rn satisfies
A =

m


Si A.

(1.12)

i=1

Then its dimension d = dimH (A) = dimB (A) is the solution of the equation
m

sdi = 1,

i=1

and one has 0 < Hd (A) < .

Proof. See e.g. [Falconer, 1990, Section 9.2].

(1.13)

61

Section 1.3 The H AUSDORFF dimension

Two remarks may help to make this assertion plausible. The first is a heuristic one:
the open set condition suggests that the union in (1.12) behaves as a disjoint one, so we
expect the equations
H (A) =
s

H (Si (A)) =
s

i=1

ssi Hs (A)

i=1

to hold. If we knew that this reasoning was legitimate and that Hs (A) is finite and
positive we could cancel Hs (A) and would obtain equation (1.13). The trouble is that
we do not know this without a considerable amount of mathematical details.
The second remark is more rewarding. If we are willing to accept Theorem 1.3.8,
then it helps us to determine the dimension of a fractal even if it is self-similar with
similarity factors of different sizes. An example has been discussed in Section 1.1.4.2.
There is some trouble to be overcome here, too: equation (1.13) can only be solved
numerically, e.g. by N EWTONs method. We look for a zero of the function
g(x) =

(si )x 1

i=1

with derivative
g  (x) =

log si sxi .

i=1

The looked-for dimension d is then given by

x0 = 0,

xk+1 = xk




n
xk
1
i=1 (si )
n
log
s

sxi k
i
i=1

d = lim xk .
k

Let us try this for the modified KOCH curve in Section 1.1.4.2 with initiator A(0) =
[ 0, 1 ] and generator given by the line segments joining consecutively the points
p0 = (0, 0)

p1 = ( 13 , 0)

p2 = ( 13 , 15 )

p3 = ( 23 , 15 )

p4 = ( 23 , 0)

p5 = (0, 0).

The fractal A satisfies the equation (1.12) with s1 = s3 = s5 = 13 and s2 = s4 = 15 .

N EWTONs method gives as a solution of equation (1.13), which now has the form
 d
 d
3 13 + 2 15 = 1,

the value d = dimH (A) 1.2719. For the modified KOCH curve studied in Section 1.1.4.2 with a generator consisting of five line segments of equal length 13 , we get
5
(dimS (A) =) dimH (A) = log
log 3 1.465.

Iterative function systems

In the first chapter our way of creating fractals relied on iterations of mappings. At the
beginning dealing with initiators and generators we applied an operator f , replacing
every line segment of a given piecewise linear curve by a suitably reduced similar copy
of the generator, to an initial set A(0) . Repeated application of f produced the sets
A(k) which converged point-wise and uniformly to the fractal curve A. We invoked
the help of functions k on the unit interval of which A(k) were the graphs and which
converged uniformly to a function of which A was the graph. These functions came
in handy since they allowed us to pass through the fractal in time 1 even if it was
space-filling but for higher dimensional fractals (Section 1.1.7) the handling would
have become complicated and we did not pursue this line of thought any further.
This procedure could also be described by the iteration of a finite set of similarities
Sj (1 j n), connected with the operator f and arising from its actions on the
initiator. Since application of f replaced each segment of the generator by a similar
copy of the whole generator, the fractal resulting from infinite repetition of this process
became similar to finitely many parts of it, namely the fractals having as initiators
the segments of the generator, and thesein their turn could be considered as similar
copies Si A of A. In short, we have A = nj=1 Sj A (suggestive of a snail biting into its
own tail).
In other words, the fractal A is invariant under application of the operator

F = nj=1 Sj . Putting the right side of the above formula for A in place of the last A,
 
and repeating this k times, we get A = ki=1 nji =1 Sj1 . . . Sjk A. This formula does not
 
quite hold for the approximating sets A(k) = ki=1 nji =1 Sj1 . . . Sjk A(0) = F (k) A(0) ,
but the similarity of the two formulas nourishes the suspicion that they have something
to do with the approximation of A by A(k) .
Our aim in this chapter (inspired by the exposition in [Barnsley, 1988]) is (a) to
substantiate this suspicion by analyzing the concept of approximation used in this context and (b) to use the corresponding construction to create an even bigger family of
fractals by also allowing, in place of Sj , mappings more general than similarities. The
first section is devoted to topic (a), the second one to topic (b). In the third section we
use the wisdom acquired for creating fractals to our liking in the plane.

2.1

The space of compact subsets of a complete metric space

In practice we shall deal later with R2 considered as a complete metric space, but we
may as well develop our tools in the general setup of an arbitrary complete metric space
(X, d). Our intention is to provide the set K(X) of all compact subsets of X with a
metric which makes it again a complete metric space.
2.1.1 Definition. Let K(X) be defined by K(X) := {A X : A compact, A = } (=
set of all non-empty compact subsets of X ).

64

Chapter 2 Iterative function systems

2.1.2 Definition. For a X , B K(X) let the distance of a from B be defined by

d(a, B) := min{d(a, b) : b B}.

2.1.3 Definition. For A K(X), B K(X) let the distance of the set A from the set
B be defined by
d(A, B) := max{d(a, B) : a A}.

Note that in general we have d(A, B) = d(B, A). For instance, if B  A, then
d(A, B) > 0, d(B, A) = 0. So d is not yet suitable for a metric in K(X).
2.1.4 Definition. For A K(X), B K(X) the H AUSDORFF distance of the sets A
and B is defined by
h(A, B) := max(d(A, B), d(B, A)).

2.1.5 Theorem (Properties of the H AUSDORFF distance).

(a) h(A, A) = 0.
(b) A = B h(A, B) > 0.
(c) h(A, B) = h(B, A).
(d) h(A, C) h(A, B) + h(B, C).
Proof. Assertions (a) and (c) are immediate consequences of the definition.
(b) Without loss of generality suppose there is an a A \ B . This implies d(A, B)
d(a, B) > 0. The converse implication is clear.
(d) It suffices to show d(A, C) d(A, B) + d(B, C) ( h(A, B) + h(B, C)). This
results from the following inequalities, valid for all a A and all b B :
d(a, C) = min{d(a, c) : c C}
min{d(a, b) + d(b, c) : c C}
= d(a, b) + min{d(b, c) : c C}
= d(a, b) + d(b, C)
d(a, b) + d(B, C),
d(a, C) d(a, B) + d(B, C).

Taking the maximum over all a A on both sides produces the assertion.

Section 2.1 The space of compact subsets of a complete metric space

65

Theorem 2.1.5 guarantees that (K(X), h) is a metric space. We shall therefore for
h also use the term H AUSDORFF metric. There is a rather illustrative way to envision
the H AUSDORFF distance of two subsets of X : it measures the extent to which they
embrace each other (the symbol denotes the logical and, the symbol denotes the
logical there exists).
2.1.6 Definition. For B K(X) and > 0 let the -hull B{} of B be defined by
B{} := {x : d(x, B) } = {x : b B d(x, b) }.

2.1.7 Theorem. h(A, B) A B{} B A{} .

Proof. By definition the inequality h(A, B) is equivalent to d(A, B)
d(B, A) . It will therefore suffice to show d(A, B) A B{} . This
is a consequence of the equivalence of the following assertions:
d(A, B) ,
d(a, B)

a A,

a B{}

a A,

A B{} .

2.1.8 Lemma. For A K(X) and > 0 the set A{} is closed.
Proof. Let the sequence {bn }
n=1 A{} converge to b = limn bn . For every
n N there exists an an A satisfying d(an , bn ) . Since A is compact we
may without loss of generality suppose that the sequence {an }
n=1 converges to some
limit a = limn an . By the continuity of the distance function d we get d(a, b) =
limn d(an , bn ) . This implies b A{} .
Now comes the final but somewhat tedious part: we shall convince ourselves that
the metric space (K(X), h) is complete. As is to be expected, we shall have to use the
completeness of the space (X, d). Definition 2.1.9 and Theorem 2.1.10 are tools which
we shall use. They belong to standard topology. A proof of the theorem is included
here in order to increase self-containedness of this presentation.
2.1.9 Definition. A subset A X is called totally bounded if given any > 0 there
exists a finite set E X satisfying A E{} .
2.1.10 Theorem. A subset A X is compact iff A is closed and totally bounded.
Proof. =: Every compact set A is closed. Suppose A is not totally bounded. Then,
for some > 0, every finite set {a1 , . . . , an } A admits an element an+1 A satisfying d(ai , an+1 ) > (1 i n). We may therefore construct inductively a sequence
{ai }
n=1 A satisfying i < j = d(ai , aj ) . This sequence has no converging
subsequence. This contradicts the compactness of A.

66

Chapter 2 Iterative function systems

=: Suppose A is closed and totally bounded. Let {an }

n=1 A be a sequence in A.
We shall show that this sequence admits a converging subsequence. This will imply
compactness of A.
Let E (0) X be a finite subset of X satisfying


(0)
(0)
{an }
)1 .
n=1 A E{1} = x X : d(x, E

There has to be a closed ball B0 with center in E (0) and with radius 1 containing

an infinite subsequence {a0,n }

n=1 {an }n=1 . Any two elements a0,n , a0,m in this
subsequence have distance d(a0,n , a0,m ) 2. By the same token there is a finite subset
E (1) X satisfying


(1)
(1)
{a0,n }
) 21
n=1 A B0 E{21 } = x X : d(x, E
and a closed ball B1 with center in E {1} and with radius 21 , containing an infinite
subsequence

{a1,n }
n=1 {a0,n }n=1 {an }n=1 .
Any two elements a1,n , a1,m in this subsequence have distance d(a1,n , a1,m ) 1.
In an inductive continuation of this procedure the inductive step starts out with an

infinite subsequence {ak1,n }

n=1 {an }n=1 contained in a closed ball Bk1 of radius
(k1)
2
; the inductive step produces a finite set E (k) such that
(k)

A Bk1 E{2k }

and a closed ball Bk with center in E (k) and radius 2k containing an infinite subsequence

{ak,n }
n=1 {ak1,n }n=1 .
Any two elements ak,n , ak,m in this subsequence have distance d(ak,n , ak,m )
2(k1) .

Now consider the subsequence {ak,1 }

k=0 {an }n=1 . Any two elements ak,1 , aj,1
(k < j) in this subsequence have distance d(ak,1 , aj,1 ) 2(k1) . The subsequence
therefore is fundamental and has to converge.
We shall have to convince ourselves that any fundamental sequence {An }
n=1 in
K(X) admits a limit A K(X). It will turn out that this limit A is the set of all
limit points in X of fundamental sequences {an }
n=1 for which an An n N.
In preparation thereof we shall show that any fundamental sequence {anj }
n=1 X
satisfying anj Anj for a subsequence of indices {nj }

N
can
be
replenished
to
j=1
a fundamental sequence {an }
satisfying
a

A
for
all
n

N
.
n
n
n=1
2.1.11 Lemma. Let {An }
n=1 be a fundamental sequence in (K(X), h), and let
{xnj }
(n
<
n
<

<
nj < nj+1 < ) be a fundamental sequence in (X, d)
1
2
j=1
satisfying xnj Anj (j N). Then there is a fundamental sequence {
xn }
n=1 satisfying
x
n An n N
and
x
nj = xnj j N.

67

Section 2.1 The space of compact subsets of a complete metric space

Proof. We shall explicitly construct the looked-for sequence {

xn }
n=1 . Let
j N.

x
nj := xnj

For nj1 < n < nj we choose xn An in such a way that d(xnj , xn ) = d(xnj , An ).
We now have to show: given > 0 there is an index N with the property that the
inequalities
N nj1 < n nj ,
N nk1 < m nk

(2.1)

imply d(
xn , x
m ) < . To this end choose N such as to satisfy

n N, m N,

nj N, nk N.

d(An , Am ) < 3
d(xnj , xnk ) < 3

If n and m satisfy (2.1), then

d(
xn , x
m ) d(xnj , xn ) + d(xnj , xnk ) + d(xnk , x
m )

d(xnj , An ) + 3 + d(xnk , Am )

d(Anj , An ) + 3 + d(Ank , Am )
.

2.1.12 Theorem. Let {An }

n=1 be a fundamental sequence in (K(X), h) and define
A := {x X : x = lim xn , xn An (n N)}.
n

Then A = limn An with respect to the H AUSDORFF distance in K(X). Consequently, the metric space (K(X), h) is complete.
Proof. We conduct the proof in five steps.
Claim 1: A = .
Proof: For simplification of the notation let k := 2k . Choose a monotone sequence
of indices {Nj }
j=1 in such a way that
h(An , Am ) < j

n Nj , m Nj .

By Theorem 2.1.7 the conditions n Nj , m Nj imply An (Am ){j } . Starting

with any xN1 AN1 we construct a fundamental sequence of elements xNj ANj as
follows: because of AN1 (AN2 ){1 } we may choose an element xN2 AN2 in such
a way that d(xN1 , xN2 ) 1 . If xNj ANj (1 j k) have been chosen in such a
way that d(xNj , xNj+1 ) j for 1 j k 1, then because of ANk (ANk+1 ){k }
we may again choose xNk+1 ANk+1 in such a way that also d(xNk , xNk+1 ) k is
satisfied.

68

Chapter 2 Iterative function systems

Let now > 0 be given. We choose k so large that k < . For k < i < j we get
d(xNi , xNj ) d(xNi , xNi+1 ) + + d(xNj1 , xNj )
j1

l = k < .

l=k+1

l=i

We see that the sequence {xNj }

j=1 is fundamental in X . By Lemma 2.1.11 we can
extend it to a fundamental sequence {
xn }
n An (n N) and
n=1 which satisfies x
xNj = xNj (j N). Its limit point belongs to A which consequently is not empty.
Claim 2: A is closed.
Proof: Suppose the sequence {am }
m=1 A converges to the limit a X . Each
element am in turn is the limit of a sequence {xm,n }
n=1 , xm,n An (n N). Let
{Mj }
be
an
increasing
sequence
of
indices
satisfying
j=1
d(a, am ) j

m Mj .

Let, furthermore, {Nj }

j=1 be a corresponding increasing sequence of indices satisfying
d(aMj , xMj ,Nj ) j

(1 j < );

then
d(a, xMj ,Nj ) d(a, aMj ) + d(aMj , xMj ,Nj ) 2j ,
xMj ,Nj ANj ,
a =

lim xMj ,Nj .

By Lemma 2.1.11 we may extend the sequence {xMj ,Nj }

xn }
j=1 to a sequence {
n=1
(
xn An , n N) which converges to a. By definition we have a A. This proves
that A is closed.
Claim 3: Given any > 0 there is an N () N such that n N () = A (An ){} .
Proof: Consider any a A and a corresponding sequence {am }
m=1 X satisfying
a = lim am
m

and

am Am

m N.

Choose N () in such a way that m N (), n N () implies h(Am , An ) < . Given

any n N (), for all m N () by Theorem 2.1.7 we then have
am Am (An ){} .

Since the set (An ){} is closed (Lemma 2.1.8) this implies also a (An ){} .
Claim 4: The set A is totally bounded and consequently compact.
Proof: If A were not totally bounded, then, as in the first part of the proof of Theorem 2.1.10, for a suitable > 0 we could construct a sequence {aj }
j=1 A satisfying

Section 2.1 The space of compact subsets of a complete metric space

69

d(ai , aj ) for i = j . By claim 3 we have A (An ){/3} for some sufficiently

large n. Therefore we could associate with every aj A an element bj An such
that d(aj , bj ) 3 . An being compact we may suppose without loss of generality
that {bj }
j=1 converges towards an element b An . For some suitable N () and all
i N (), j N () we would get

d(bi , bj ) < 3 ,
d(ai , aj ) d(ai , bi ) + d(bi , bj ) + d(bj , aj ) < ,

a contradiction to our assumption.

Claim 5: A = limn An with respect to H AUSDORFF distance h.
Proof: By Theorem 2.1.7 the claim is equivalent with the following claim. Given any
> 0 there is an index N () such that for all n N ()
A An,{}

and

An A{} .

Claim 3 has taken care of the first assertion. In order to check the second one let N ()
be so large that

h(Am , An ) 2
m N (), n N ().
Let now any y An (n N ()) be given. We are going to show y A{} .
In order to simplify the notation we put now k := 2k . Since {An }
n=1 is a fundamental sequence in (K(X), h), there is an increasing sequence of indices {Nj }
j=1
(N1 n) satisfying (due to Theorem 2.1.7)
Ak (Al ){j+1 }

k Nj , l Nj .

The scheme of implications below serves to construct a fundamental sequence

{xNj }
j=1 satisfying xNj ANj (1 j < ):
N () n N1

An (AN1 ){1 }

N1 N2

AN1 (AN2 ){2 }

xN1 AN1 : d(y, xN1 ) 1 ,

xN2 AN2 : d(xN1 , xN2 ) 2 ,

..
.
Nj1 Nj

ANj1 (ANj ){j }

xNj ANj : d(xNj1 , xNj ) j ,

..
.
By Lemma 2.1.11 we can extend the fundamental sequence {xnj }
j=1 to a fundamental
sequence {
xn }
satisfying
x

A
(n

N)
and
x

=
x
(j
N). The limit x =
n
n
N
N
j
j
n=1

n A of this sequence satisfies d(y, x)

=
and consequently
limn x
j
j=1
y A{} .

70

2.2

Chapter 2 Iterative function systems

Contractions in a complete metric space

As already indicated at the beginning of this chapter the key to the construction of
fractals, generalizing the oneswe have got to know so far, is the iterated application of
n
an operator of the form F = m
j=1 fj to compact sets in R , where fj (1 j m) are
n
n
suitable mappings K(R ) K(R ). Of course one has to impose some condition on
the mappings fj which guarantee that the sequence {F (k) (A(0) )}
k=1 converges in the
sense of the H AUSDORFF metric, as has seemingly been the case with the similarities
Sj employed in the initiator-generator procedure. Again we might as well develop the
same convenient means in the general setup of mappings in K(X) as in Section 2.1.
We therefore suppose again that (X, d) is a complete metric space.
2.2.1 Definition. A mapping f : X X is called a contraction if there is a constant
c [ 0, 1[ (called a contraction constant) such that
d(f (x), f (y)) c d(x, y)

(x, y) X X.

Note that, as an immediate consequence of the definition, a contraction is a continuous mapping. In general we do not require the constant c to be optimal. For a similarity
in Rn as in Definition 1.1.3.1, however, a similarity factor s < 1 is automatically the
smallest contraction constant possible.
2.2.2 Definition. The sequence {f (k) (x)}
k=0 is called the orbit of the element x X
under f . A point x X is called a fixed point for f if f (x) = x.
2.2.3 Theorem. A contraction f : X X has exactly one fixed point and every orbit
converges to it.
Proof. Let c be the contraction constant of f . Let any x X and the non-negative
integers m < n be given. Then
d(f (m) (x), f (n) (x)) = d(f (m) (x), f (m) f (nm) (x))
cm d(x, f (nm) (x)),
d(x, f (n) (x)

d(f (k1) (x), f (k) (x))

k=1

ck1 d(x, f (x))

k=1

1 c d(x, f (x)),
cm

d(f (m) (x), f (n) (x)) 1 c d(x, f (x)).

(2.2)

The last inequality shows that the sequence {f (k) (x)}

k=1 is fundamental and converges
to some limit xf = limk f (k) (x). At the same time we get
f (xf ) = f ( lim f (k) (x)) = lim f (k+1) (x) = xf .
k

71

Section 2.2 Contractions in a complete metric space

For any y X satisfying f (y) = y we get d(xf , y) = d(f (xf ), f (y)) c d(xf , y).
Because of 0 c < 1 this implies y = xf .
2.2.4 Corollary. Let f : X X be a contraction with contraction constant c and let
xf be the fixed point of f . Then for every x X one has
cm

d(f (m) (x), xf ) 1 c d(x, f (x)),

in particular

d(x, xf ) 1 c d(x, f (x)).

Proof. In formula (2.2) let n .

It is well known that a continuous image of a compact set is again compact. For a
contraction f we therefore have f (E) := {f (x) : x E} K(X) for every E K(X).
We may therefore consider f also as a map from K(X) into itself, and we shall do so
without extra notation. Fortunately, into its elevated position f also carries along its
contraction property.
2.2.5 Theorem. Let f be a contraction in (X, d) with contraction constant c. Then, as
a function on the complete metric space (K(X), h) into itself, it is again a contraction
with contraction constant c.
Proof. By the definition of the H AUSDORFF metric h it suffices to check the following
for A K(X), B K(X):
d(f (A), f (B)) = max{d(f (a), f (B) : a A}
= max{min{d(f (a), d(f (b) : b B} : a A}
max{min{cd(a, b) : b B} : a A}
= c d(A, B).

If we only apply the iterates of a single contraction f to a set A K(X), then

the images simply contract into the fixed point of f . In order to get more than that
we therefore have to invoke the help of some more contractions fj (1 j J) and

combine the single image sets fj (A) to the bigger set Jj=1 fj (A), also called collage
(of the set A under the contractions fj (1 j J); the term may also be used for the

mapping F = Jj=1 fj ). This procedure turns out to be very effective and has therefore
given rise to an extra definition (2.2.6), but in order to see this we have to provide some
intermediate results.
2.2.6 Definition. A finite set of contractions fj (1 j J) in a complete metric space
(X, d) is called an iterative function system (IFS); the functions fj are called collage
mappings.
2.2.7 Lemma. For compact subsets A, B and C of X one has


d(A B, C) = max d(A, C), d(B, C) .

72

Chapter 2 Iterative function systems

Proof.



d(A B, C) = max d(x, C) : x A B


= max max{d(x, C) : x A}, max{d(x, C) : x B}


= max d(A, C), d(B, C) .

2.2.8 Lemma. For compact subsets A, B , C and D of X one has



h(A B, C D) max h(A, C), h(B, D) .
Proof.
h(A B, C D)


= max d(A B, C D), d(C D, A B)


= max max{d(A, C D), d(B, C D)}, max{d(C, A B), d(D, A B)}

(by Lemma 2.2.7)



max max{d(A, C), d(B, D)}, max{d(C, A), d(D, B)}


= max h(A, C), h(B, D) .
There is no need for confusion as to whether in Lemma 2.2.8 one has to combine
A with C or D and B with D or C ; simply take your choice. We can now harvest the
fruits of our endeavors so far.
2.2.9 Theorem. Let {fj }Jj=1 be an IFS in (X, d) and let the map F : K(X) K(X)
be defined by
J

F (A) :=
fj (A)
(A K(X)).
(2.3)
j=1

Then F is a contraction in (K(X), h) with contraction constant c = max{cj : 1 j

J}, where cj is a contraction constant of fj for 1 j J .
Proof. It suffices to prove the assertion for J = 2 and to proceed by induction. We
find
h(F (A), F (B)) = h(f1 (A) f2 (A), f1 (B) f2 (B))


max h(f1 (A), f1 (B)), h(f2 (A), f2 (B))


max c1 h(A, B), c2 h(A, B)
c h(A, B).

(by Lemma 2.2.8)

73

Section 2.2 Contractions in a complete metric space

2.2.10 Theorem. Let {fj }Jj=1 be an IFS in (X, d) and let F : K(X) K(X) be given
by (2.3). Then there is a unique fixed point A K(X) of F , called the attractor of
F , and with respect to the H AUSDORFF metric h in X one has
for all B K(X)

lim F (k) (B)

A =

= F (A)
J


fj (A).

j=1

Proof. Apply Theorem 2.2.3 to the contraction F in the complete metric space
(K(X), h).
2.2.11 Theorem (Collage Theorem). Let {fj }Jj=1 be an IFS in (X, d) and let F :
K(X) K(X) be given by (2.3) and let c = max{cj : 1 j J}, where cj is a
contraction constant of fj for 1 j J . Let A be the attractor of F and let B K(X)
be given. Then
1
h(B, A) 1 c h(B, F (B)).
Proof. Apply Corollary 2.2.4 to the contraction F in the complete metric space
(K(X), h).
It will be convenient to have one more piece of knowledge furnishing insight in the
continuous dependence of an attractor on changes in the IFS producing it.
2.2.12 Theorem. Let (T, dT ) (the parameter space) be a metric space and let

{ft,j }Jj=1 : t T be a family of IFS in (X, d) with contraction constants ct c < 1
and attractors At respectively. Suppose for some constant C and for every x X one
has
d(ft,j (x), fs,j (x)) C dT (t, s)
(t T, s T, 1 j J).
Then

h(At , As ) 1 c dT (t, s).

Proof. For t T let Ft be defined as in (2.3). By Corollary 2.2.4 we have

1

h(At , As ) 1 c h(At , Fs (At )) = 1 c h(Ft (At ), Fs (At )).

It suffices to show that for every B K(X) we have

h(Ft (B), Fs (B)) C dT (t, s).

By Lemma 2.2.8 we know

h

J

j=1

ft,j (B),

J

j=1




fs,j (B) max h(ft,j (B), fs,j (B)) : 1 j J .

74

Chapter 2 Iterative function systems

Therefore it suffices to show

d(ft,j (B), fs,j (B)) C dT (t, s).

To this end we argue as follows: given x ft,j (B) let xt,j B be chosen such that
ft,j (xt,j ) = x. Then
d(x, fs,j (xt,j )) = d(ft,j (xt,j ), fs,j (xt,j )) C dT (t, s),
d(x, fs,j (B)) C dT (t, s),
d(ft,j (B), fs,j (B)) C dT (t, s).

In the examples which we shall meet (Sections 2.3.4.5, 2.3.4.6 and 2.3.7) the parameter space is some bounded interval T R and X is some bounded subset of R2 , as
e.g. the unit square, while d(x, y) = |x y| and dT (t, s) = |t s|. The coordinate functions gj,i (i = 1, 2) of the mappings ft,j : x ft,j (x) = (gj,1 (t, x), gj,2 (t, x)) have
the property that the partial derivatives gtj,i are continuous and therefore uniformly
bounded functions on the compact space T X . By the mean value theorem one has
gj,i (t, x) gj,i (s, x)
gj,i 
=
(t , x)
ts
t

(t ]s, t[ T ).

It is readily seen that this guarantees that the hypotheses of Theorem 2.2.12 are satisfied.

2.3

Affine iterative function systems in R2

As M ANDELBROT has pointed out, fractals are found abound and seem to be the result
of a favorite construction principle in nature. This is reflected in the two-dimensional
images with which to a large extent we perceive the world around us. Striking examples
are provided by the BARNSLEY fern and other fractals hardly distinguishable from
images of objects in nature.
Our intention in this section is to experience the interdependence between IFSs and
the fractals produced by them, and then following the visible traces laid out by nature
and with the help of mathematics to construct fractals which look approximately like
a given compact set B R2 . As a tool we use suitable IFSs, preferentially made up
by contractions which mathematically are easy to handle, as were the similarities in
Chapter 1. A much wider and in fact sufficiently wide family of contractions are those
provided by affine maps in R2 .
2.3.1 Definition. An affine map f in R2 is given by
f (x) := Lx + b

(x R2 ),

where L is a linear map, given by a matrix also denoted by L, and b R2 .

(2.4)

Section 2.3 Affine iterative function systems in R2

75

When computations require a more detailed notation we shall write vectors as

columns and use the well-known matrix operations transposition (indicated by t ), multiplication and addition. E.g. formula (2.4) is then written in the form



   
f (x)1
l1,1 l1,2
x1
b1
=
+
.
f (x)2
l2,1 l2,2
x2
b2
A rotation about an angle is given by an orthogonal matrix


cos sin
O =
.
sin cos
If in (2.4) we have L = s O, then f is a similarity with similarity factor s.
In order to decide whether an affine map is a contraction, we have to compute a
contraction constant. Here we are even able to get the smallest and thereby optimal
one.
2.3.2 Theorem. The smallest contraction constant of the affine map f as in (2.4) is the
square root of the largest eigenvalue of the symmetric matrix Lt L.
Proof. As is well known in linear algebra, there exists an orthogonal matrix O diagonalizing the symmetric matrix Lt L, i.e. satisfying


c1 0
t
t
O (L L)O =
= D.
0 c2
The coefficients c1 and c2 are the (necessarily real and non-negative) eigenvalues of
Lt L. Without loss of generality we assume c1 c2 . The 
distance in R2 of two vectors
2
2
2
2
x
 R and y R is given by d(x, y) = |x y| = (x1 y1 ) +t(x2 y2 ) =
(x y)t (x y). In order to simplify the notation we shall write O (x y) = u.
Note that |u|2 = |x y|2 . We get
|f (x) f (y)|2 = (L(x y))t (L(x y)) = (x y)t Lt L(x y)
= (x y)t ODOt (x y) = ut Du
= c1 u21 + c2 u22 c1 |u|2 = c1 |x y|2 .

Finally note that for x y = O(1, 0)t we have equality in place of , so

the optimal contraction constant.

c1 is indeed

In order that f be a contraction in the sense of Definition 2.3.1 this value c1 must
be smaller than 1. In this case the matrix L cannot have the eigenvalue 1. The next
assertion is then an immediate consequence of formula (2.4).
2.3.3 Theorem. If the affine map f as in (2.4) is a contraction, then its fixed point is
xf = (I L)1 b.

76

Chapter 2 Iterative function systems

As a consequence, formula (2.4) may be reformulated incorporating the fixed point

to
f (x) = Lx + (I L)xf .

Before making use of the broadened possibilities to construct fractals in R2 let

us see whether there is any reward in applying our newly gained knowledge to the
construction of the fractals we are already familiar with. In fact we now obtain a
comfortable quantification of the degree of approximation of the resulting fractals.
Consider first the version of the S IERPINSKI triangle A discussed in Section 1.1.7.1
and Theorem 1.2.3 (Figure 1.53). Knowing its self-similarity properties we see that it

is mapped onto itself by the operator F = 3j=1 fj where the affine mappings fj are
given as in (2.4) by (I denoting the unit matrix)



0.5 0
,
0 0.5




0.5
0
b2 =
, b3 =
.
0
0.5

L1 = L2 = L3 =
 
0
b1 =
,
0

1
2I

All mappings fj are similarities with similarity factor sj = 12 which therefore is also

a contraction constant for F = 3j=1 fj . Applying the iterations of F e.g. to the unit
square B as initial set, we may estimate the approximation of A by F (m) (B) by means
of Corollary 2.2.4. The set F (B) consists of three squares with side length 12 each. For
h(F (m) (B), A) we get


h(B, F (B)) = max d(B, F (B)), d(F (B), B) = d(B, F (B))


1
= max d(b, F (B)) : b B = 2 ,
sm

h(F (m) (B), A) 1 s h(B, F (B)) = m .

2

The original version

of the S IERPINSKI triangle as in Section 1.1.7.1 is obtained by

1
3 t
changing b3 to ( 4 , 4 ) . In the last approximation estimate, only h(B,
F (B)) changes

from 12 to the length of a diagonal of a rectangle with sides 14 and 12 43 , approximately

equal to 0.26.
The collage mappings fj for the C ANTOR dust (Section 1.1.7.2) and the S IERPIN carpet (Section 1.1.7.3) are as easily identified. They all are similarities with equal
similarity factors which leave the directions of the coordinate axes unchanged.

The KOCH curve A is obtained by means of the mapping F = 4j= fj where the fj
SKI

Section 2.3 Affine iterative function systems in R2

77

are similarities with common similarity factor sj = 1/3 given by


 
 

1
2
0
0
3
3
L1 = L4 = 13 I =
=
=
,
b
,
b
,
1
4
0 13
0
0

 

1
1
63
6
3

=
,
b
,
L2 =
2
1
3
0
6
6




L3 =

1
6
63

3
6
1
6

b3 =

1
2
3
6

If we choose as initial set B the unit interval on the x-axis, then d(B, F (B)) = 41 3 ,
1
h(B, F (B)) = d(F (B), B) = 2
. By Corollary 2.2.4 we get h(F (m) (B), A)
3
1
1
3m (11/3) h(B, F (B)) = 43m1 3 . For this concept of approximation we do not need
the help of a function as in Section 1.1.2. On the other hand, the IFS method does
not give any information on how to consider A as a curve.
In case of the short fractal set A in Section 1.1.6.2 with H AUSDORFF dimension 1,
it comes in handy that we do not need to consider it as the graph of a function (which
would have to be highly discontinuous) in order to have it properly defined as a limit
set. It is the product of the IFS consisting of two similarities {f1 , f2 } with common
similarity factor 12 , given as in (2.4) by


 
1
0
0
2
,
b1 =
,
L1 =
1
0 2
0

 

0 12
1
L2 =
,
b2 =
.
1
1
0
2
2
Taking again as initial set B the unit interval on the x-axis we obtain d(B, F (B) = 12 ,
1
h(B, F (B)) = d(F (B), B) = 1 and h(F (m) (B), A) 2m
1.

2.3.4

Some examples of fractals constructed by means of IFS

Some information concerning the illustrations in the present Chapter 2, produced by

means of a Turbo Pascal program, seems to be in order. For every screen pixel (i, k)
carrying
n a colour value, indicating a point z(i, k) of the set to which the IFS mapping
F = j=1 fj is applied, this point z(i, k) is mapped by every one of the mappings fj
(1 j n) into its image fj (z(i, k)). The corresponding pixels (ij , kj ) are stored
consecutively in a file together with the index j , indicating at the same time the colour
corresponding to the mapping fj . These colours are blue (1), green (2), cyan (3), red
(4), magenta (5), brown (6), light gray (7), dark gray (8), and light blue (9). Finally the
file entries are recalled consecutively to produce the new screen content. This implies
that the images of the original set under the mappings fj with higher index j in general
cover those for smaller indices.

78

Chapter 2 Iterative function systems

In the figures we shall adopt the terminology

initial set for the set A(0) , the orbit

of which under the IFS mapping F = nj=1 fj converges to the attracting fractal A,
and generating set for the set A(1) = F (A(0) ) which characterizes the mapping F .
This corresponds with the terminology initiator and generator in Chapter 1.
2.3.4.1 A cross
Consider the five similarities fj reducing the unit square by the factor 13 and placing
the images in form of a cross. As in (2.4) these affine mappings are given by matrices
Lj and translation vectors bj (1 j 5) as follows:


1
0
3
Lj =
(1 j 5),
0 13
 
 
 
 
 
1
1
2
1
0
3
3
3
3
=
=
=
, b2 =
,
b
,
b
,
b
.
b1 =
3
4
5
1
1
2
1
0
3
3
3
3
If we choose the boundary of the unit square as the initial set A(0) , then A(1) =
F (A(0) ) looks as in Figure 2.1, A(2) = F (2) (A(0) ) looks as in Figure 2.2, and A(6) =
F (6) (A(0) ) looks as in Figure 2.3, already rather close to the expected attractor. But
Theorem 2.2.10 has stated that any other initial set A(0) would do as well. So Figures 2.4 and 2.5 illustrate the sets A(1) and A(2) if A(0) is a full circle, while the next
Figures 2.62.9 illustrate these sets if A(0) is a quarter circle or a triangle respectively.
In every case, the set A(6) already looks as in Figure 2.3.
The present IFS satisfies the open set condition (just consider the images of the
open unit square). Therefore Theorem 1.3.8 applies and the dimension of the resulting
5
fractal is dimH (A) = dimB (A) = log
log 3 1.465.

Figure 2.1. The generating set A(1) for the cross, consisting of five images of
the initial set A(0) which consists of the four sides of the unit square.

Section 2.3 Affine iterative function systems in R2

Figure 2.2. The approximating set A(2) for the cross, with initial set as above.

Figure 2.3. The approximating set A(6) for the cross.

79

80

Chapter 2 Iterative function systems

Figure 2.4. The generating set A(1) for the cross, if the initial set A(0) is a
circular disc.

Figure 2.5. The approximating set A(2) for the cross, with initial set a circular
disc.

Section 2.3 Affine iterative function systems in R2

Figure 2.6. The generating set A(1) for the cross, if the initial set A(0) is a
quarter circle.

Figure 2.7. The approximating set A(2) for the cross, with initial set as above.

81

82

Chapter 2 Iterative function systems

Figure 2.8. The generating set A(1) for the cross, if the initial set A(0) is a
triangle.

Figure 2.9. The approximating set A(2) for the cross, with initial set as above.

Section 2.3 Affine iterative function systems in R2

83

2.3.4.2 A decoration
The IFS now consists of the affine mappings fj (1 j


1
0
3
L1 =
,
0 13


0 13
,
L2 =
1
0
3


0 13
,
L3 =
13 0


1
0
3
L4 =
,
0 23

4) given by
 
b1 =

1
3

 
b2 =

1
3

 
b3 =

2
3
1
3

1
3
1
3

 
b4 =

Using the half of the unit square below the diagonal as initial set A(0) one sees (Figure 2.10) that f1 , f2 and f3 are similarities with similarity factor 13 , but f2 turns the
image a right angle counterclockwise while f3 turns it a right angle clockwise. The
mapping f4 is not any more a similarity since it squeezes everything in direction of
the x-axis by a factor 13 , but in direction of the y -axis by a factor of 23 . Although the
open set condition is satisfied, Theorem 1.3.8 is no longer applicable. It is still possible to compute the box-counting dimension of the resulting fractal (Figure 2.11). A
closer inspection, using as A(0) the boundary of the unit square (Figure 2.12), reveals
that the set A(k) is minimally covered by 5k lattice squares of side length 3k . By
5
Theorem 1.2.2 we get again dimB (A) = log
log 3 1.465.

Figure 2.10. The generating set A(1) for the decoration, consisting of four images of the initial set A(0) which consists of the half of the unit square below the
diagonal off the origin.

84

Chapter 2 Iterative function systems

Figure 2.11. The approximating set A(7) for the decoration, with initial set the
boundary of the unit square.

Figure 2.12. The approximating set A(3) for the decoration, with initial set A(0)
the boundary of the unit square.

Section 2.3 Affine iterative function systems in R2

85

2.3.4.3 An antenna
The IFS consists of the affine mappings fj (1 j 4) given by

 

1
1
0
3
3
L1 =
=
,
b
,
1
0 13
0

 

1
0 13
3
L2 =
=
,
b
,
2
1
13 0
3


 
2
0 13
3
,
b
,
=
L3 =
3
1
1
0
3
3

 

1
1
0
3
3
=
,
b
.
L4 =
4
1
0 23
3
There is too much overlap of the images of the unit square as initial set (Figure 2.13) to
hope for an open set condition, and f4 is certainly not a similarity. In order not to get
involved in too much complication it will suffice to notice that the approximating set
A(k) may be covered by less than 5k lattice squares of side length 3k (Figure 2.14a).
5
The dimension of the resulting fractal will therefore not exceed log
log 3 1.465. Starting
from the boundary A(0) of the unit square the set A(1) looks simple, but no longer the
set A(7) approximating the fractal A which resembles an intricate TV-antenna (Figure 2.14b).
The data for the next fractals to be presented here have been taken from or have been
at least inspired by examples in [Barnsley, 1988] and [Peitgen, Jurgens, Saupe, 1992].

Figure 2.13. The approximating set A(1) for the antenna, consisting of four
overlapping images of the initial set A(0) which consists of the boundary of the
unit square.

86

Chapter 2 Iterative function systems

(a)

(b)

Figure 2.14. The approximating sets A(3) (Figure 2.14a) and A(7) (b) for the
antenna, with initial set as above.

2.3.4.4 A flower garden

Consider the three affine mappings defined as follows (Figure 2.15, cf. Figure 5.12 in
[Peitgen, Jurgens, Saupe, 1992]):

 

1
1
0
3
3
L1 =
=
,
b
,
1
2
0 13
3

 

1
0 13
3
,
b2 =
,
L2 =
1 0
0

 

2
0 13
3
L3 =
,
b3 =
.
1 0
1
At a first glance there seems to be some confusion: the mappings f2 and f3 are
no contractions. Still, looking at the set A(2) (obtained just as the set A(1) from the
boundary of the unit square as initial set) everything has indeed been contracted (Figure 2.16). Why? The reason becomes clearer if as initial set A(0) we take a triangle
below the diagonal of the unit square (Figure 2.17): the mapping f2 not only squeezes
everything to 13 its breadth but does so only after having turned everything about a
right angle counterclockwise; similarly the mapping f3 squeezes everything to 13 its
breadth after having it turned about a right angle clockwise. This causes the square of

the collage F = 3j=1 fj to be a contraction and its fixed point A to exist as an attractor
(Figure 2.18).

Section 2.3 Affine iterative function systems in R2

Figure 2.15. The generating set A(1) for the flower garden, starting with the
boundary of the unit square as initial set A(0) .

Figure 2.16. The approximating set A(2) , again starting with the boundary of
the unit square as initial set A(0) .

87

88

Chapter 2 Iterative function systems

Figure 2.17. The generating set A(1) for the flower garden, starting with half
the unit square below the diagonal as initial set A(0) .

Figure 2.18. The approximating set A(11) for the flower garden, starting with
either of the mentioned initial sets.

Section 2.3 Affine iterative function systems in R2

89

If the mappings f2 and f3 would have been replaced by the mappings f2 and f3
given e.g. by

 

1 0
0

L2 =
,
b2 =
,
1
0 3
0

 

2
1 0

3
=
,
b
,
L3 =
3
0 13
0
then the set A(1) (starting with the boundary of the unit square as A(0) ) would again
have looked as in Figure 2.15 and the sequence {F (k) (A(0) )}
k=1 would have converged

in the H AUSDORFF metric to a limiting set A , but the limiting set would depend on the
initial set: taking e.g. A(0) in the unit square disjoint from its lower half would produce
a limiting set A again disjoint from this lower half.
Again a closer inspection of the sets A(k) reveals that they are minimally covered
by 7k lattice squares of side length 3k . The dimension of the fractal A (enforcing
some fantasy one might think of it as the ground map of some flower garden) by The7
orem 1.2.2 is dimB = log
log 3 1.771.
2.3.4.5 A pentagon snowflake
In order to define a parameter family of IFS let us start out with a disc A(0) with center
m = ( 12 , 12 ) and radius 12 , inscribed into the unit square, and five smaller discs A1,j
(1 j 5) with centers mj and radius r, tangentially touching their neighbours and
the boundary of A(0) (Figure 2.19). For d = |m mj | and = 5 36 the equations
1

r + d = 2,

r
= sin
d

give us
r =

0.1851,

2(1 + sin )

0.3149,


1
2 d sin 2(j + 1)
=
1
2 + d cos 2(j + 1)

d =
mj

sin

2(1 + sin )

(1 j 5).

Using the parameter t [ 1, dr ] we define fj to be the similarity with similarity

factor ts = 2tr 0.3702 t mapping A(0) onto the disc with center mj and radius tr,
given as in (2.4) by


s 0
Lj = t
(1 j 5),
0 s


1
ts

sin
2
(j
+
1
)

2
2
bj =
(1 j 5).
1
ts
+
d

cos
2
(j
+
1
)

2
2

90

Chapter 2 Iterative function systems

Figure 2.19. The generating set (A1 )(1) for the pentagon snowflake, starting
with a disc of radius 0.5, inscribed in the unit square as initial set A(0) .

For t = 1 the open set condition is satisfied and the dimension of the resulting fractal
5
A (Figure 2.20) by Theorem 1.3.8 is dimH (A) = log
log s 1.62.
As the variable t increases, the similarity factor for the mappings fj increases to
ts and the radius of the five discs constituting A(1) increases to ts
2 , but their centers
mj remain fixed. The attractor At changes correspondingly in a continuous way as
already stated in Theorem 2.2.12. The open set condition is no longer satisfied from
the beginning, but as long as the five sets fj (At ) lie apart from each other we can still
5
apply our heuristic reasoning of Section 1.1.3 to obtain dimS (At ) = loglogtlog
s . In
fact, for t 1.0319 the fractal becomes a connected set and one obtains ts 0.382,
dimS (At ) 1.672 (Figure 2.21, cf. Figure 5.15 in [Peitgen, Jurgens, Saupe, 1992]).

Figure 2.20. The approximating set (A1 )(9) for the pentagon snowflake with
similarity factor s 0.37019, dimH (A1 ) 1.62.

Section 2.3 Affine iterative function systems in R2

91

Figure 2.21. The approximating set (At )(9) for pentagon snowflake, obtained
with similarity factor ts 0.382, dimS (At ) 1.672.

2.3.4.6 Evolution of a triplicate continent

Let us modify the setup of Section 2.3.4.5 by replacing the five smaller discs by only
three discs Aj (1 j 3) with centers mj and radius r, again tangentially touching
their neighbours and the boundary of A(0) (Figure 2.22). For d = |m mj | and =

3 60 , the same equations as in Section 2.3.4.5 give us

r =

sin
3

=
0.2321,
2(1 + sin )
2 (2 + 3 )

d =

2(1 + sin )


mj =

0.2679,

1

sin
2
j
2
(1 j 3).
1
2 + d cos 2j

Using the parameter t [ 1, dr ] we now define ft,j to be the similarity with similarity
factor ts = 2tr 0.4642 t mapping A(0) onto the disc with center mj and radius tr,
given as in (2.4) by


0 s
Lj = t
,
s 0


1
ts

sin
2
j

2
2
(1 j 3).
bj =
ts
1
+
d

cos
2
j
+
2
2
Again, for t = 1, the open set condition is satisfied and the dimension of the resulting
3
fractal A1 by Theorem 1.3.8 is dimH (A1 ) = log
log s 1.4311. Still, every mapping ft,j
also turns the image about a right angle clockwise. As a result the fractal A1 looks like

92

Chapter 2 Iterative function systems

Figure 2.22. The set (A1 )(1) for the island archipelago if the initial set A(0) is a
disc of radius 0.5 inscribed in the unit square.

Figure 2.23. The approximating set (A1 )(9) for the island archipelago A1 corresponding to the similarity factor s 0.4642, with dimH (A1 ) 1.4311.

a triplicate spiraling archipelago of tiny islands (Figure 2.23) which, in fact, are tiny
archipelagoes in their own right.
What happens if we let the parameter t increase? The similarity factor increases
(Figure 2.24), and with it the size of the archipelagoes, but again, by the construction
of the mappings, the images ft,j ( 12 , 12 ) of the center of A(0) remain constant. The open
set condition evaporates in the mist, but since the archipelagoes at the beginning are
lying apart one might feel encouraged to stick to formula (1.7) providing dimS (At ) =
log 3
log tlog s .

Section 2.3 Affine iterative function systems in R2

93

Figure 2.24. The generating set (At )(1) for the island archipelago if the initial
set (At )(0) = A(0) is a disc of radius 0.5 inscribed in the unit square, and
t = 1.1.

As t increases so does dimS (At ), and whatever mass might be concentrated

in the

islands seems to drift together (Figures 2.25, 2.26), until for t = 2+3 3 1.244 one
3
gets loglog
tlog s = 2 and all archipelagoes are joined in a new continent consisting of
three countries similar in size to the whole continent (Figure 2.27, cf. Figure 5.11 in
[Peitgen, Jurgens, Saupe, 1992]).

Figure 2.25. The approximating set (At )(9) for the island archipelago At corresponding to t = 1.1 and similarity factor ts 0.51051, with dimS (Ar )
1.6340.

94

Chapter 2 Iterative function systems

Figure 2.26. The approximating set (At )(9) for the island archipelago At corresponding to t = 1.2 and similarity factor ts 0.55692, with dimS (At )
1.8769.

Figure 2.27. The approximating set (At )(9) for the triplicate continent At corresponding to t =

2+ 3
3

1.2440, with dimS (At ) = 2.

Section 2.3 Affine iterative function systems in R2

95

2.3.4.7 A twig
While the fractals presented so far originate from geometric objects, it is also possible
to imitate nature. Or is nature governed by hidden geometric laws?
One map (f3 ) of the next IFS is almost degenerate and helps to draw some sticks
(Figures 2.28 and 2.29, cf. Figure 5.13 in [Peitgen, Jurgens, Saupe, 1992]):

L1 =


0.387 0.430
0.430 0.387


,


0.441 0.091
L2 =
,
0.009 0.322


0.468 0.020
L3 =
,
0.113 0.015

b1 =

b2 =

b3 =

0.176
0.522
0.342
0.506
0.320
0.400


,

,

.

Figure 2.28. The generating set A(1) for the twig, if A(0) is the boundary of the
unit square.

Figure 2.29. The approximating set A(10) for the twig.

96

Chapter 2 Iterative function systems

2.3.4.8 A tree (cf. Figure 5.16 in [Peitgen, Jurgens, Saupe, 1992])


L1 =

L2 =

L3 =


0.195 0.488
0.344 0.443
0.462 0.414
0.252 0.361


b1 =

0.058 0.070
0.453 0.111

b2 =


,


0.035 0.070
L4 =
,
0.469 0.022


0.637
0
L5 =
,
0
0.501

b3 =

b4 =

b5 =

0.443
0.245
0.251
0.569
0.598
0.097
0.488
0.507
0.856
0.251


,

,

,

,

.

Figure 2.30. The generating set A(1) for the tree, if A(0) is the boundary of the
unit square.

Section 2.3 Affine iterative function systems in R2

Figure 2.31. The approximating set A(7) for the tree.

2.3.4.9 A leaf
The IFS here consists of four similarities (cf. Figure 3.10.10 in [Barnsley, 1988]):






0.6 0
0.18
0.18
L1 = L2 =
,
b1 =
, b2 =
,
0 0.6
0.36
0.12




0.4 0.3
0.27
L3 =
,
b3 =
,
0.36
0.3 0.4




0.4 0.3
0.27
L4 =
,
b4 =
.
0.3 0.4
0.09
It is instructive to study in more detail a spectacular fractal.

Figure 2.32. The generating set A(1) for the leaf, if A(0) is the boundary of the
unit square.

97

98

Chapter 2 Iterative function systems

Figure 2.33. The approximating set A(9) for the leaf.

2.3.4.10 The BARNSLEY fern

The IFS consists of four affine contractions fj (1 j 4), f2 and f3 being similarities (Figure 2.34, cf. Figure 5.25 in [Peitgen, Jurgens, Saupe, 1992] and Table 3.8.3 in
[Barnsley, 1988]):




0
0
0.500
L1 =
,
b1 =
,
0 0.160
0




0.849 0.037
0.075
L2 =
,
b2 =
,
0.037 0.849
0.183




0.197 0.226
0.400
L3 =
,
b3 =
,
0.226 0.197
0.049




0.15 0.283
0.575
L4 =
,
b4 =
.
0.084
0.260 0.237
Looking at the resulting fractal A (Figure 2.35) one wonders how this comes about.
It may help to study the individual mappings fj more in detail (we round numerical
results to two decimals). f1 maps the whole unit square Q with contraction constant
c1 = 0.16 and fixed point (0.5, 0) onto the interval Iu = [ 0, 0.16 ] on the y -parallel
symmetry-axis u of Q. Every time f1 is applied to a subset of R2 this set is projected
orthogonally upon u and reduced to 0.16 of its original height; in particular, a set
reaching from height 0 to height 1 has as its image the interval Iu on u. The mapping
f1 thus provides for the bottom stem of the fern. The map f2 has determinant (=area
reducing factor) 0.72, contraction constant c2 = 0.85 and fixed point z2 = (0.75, 1.03)
(note that the matrices L2 and L3 are similarity matrices i.e. constant multiples of
orthogonal matrices). It takes the unit square, reduces it to 0.85 of its size (and therefore
to 0.72 0.852 of its area), turns it a little bit down to the right and moves it with the
midpoint of its lower side up into the upper endpoint of the interval Iu . Every time
it is applied to something contained in Q it takes this something along on its journey
towards the fixed point near the upper side of Q, reducing and turning and moving it,
and providing so for the typical shrinking of the leaves close to the top of the fern in the
fixed point of f2 . The map f3 has determinant 0.09, contraction constant c3 = 0.30 and
fixed point (0.45, 0.19). It maps Q into a square f3 (Q) with vertices A3 = (0.40, 0.05),

Section 2.3 Affine iterative function systems in R2

Figure 2.34. The generating set A(1) for the BARNSLEY fern, if A(0) is the
boundary of the unit square.

Figure 2.35. The approximating set A(29) for the BARNSLEY fern.

99

100

Chapter 2 Iterative function systems

B3 = (0.60, 0.28), C3 = (0.37, 0.58), D3 = (0.17, 0.38). So whatever happens in Q

is going to be similarly happening in f3 (Q), reduced in size by the factor 0.30, and
turned and moved along from Q to f3 (Q), and as time k passes by, being contracted
into z3 . This provides for the leaves to be similar images of the whole fern and hanging
on to the left of it. Finally, f4 has determinant 0.11 (the minus due to change in
orientation), contraction constant c4 = 0.38 and fixed point z4 = (0.52, 0.07). In a
similar way as f3 it provides for the leaves to the right of the fern, which are, however,
not similar but only affine images of the whole fern, and while the left leaves follow
the turning of the whole fern to the right, the right leaves now turn to the left.
It is also instructive to observe the action of the IFS upon a polygon E (the symbol
denoting the boundary of the set E ) roughly shaped as the contour of the limiting
fractal A, such as e.g. the quadrangle with vertices a = (0.50, 0), b = (0.75, 0.30),
c = (0.70, 1.05), d = (0.30, 0.40) (Figure 2.36). What happens is that (if E is the
interior of the quadrangle abcd) the H AUSDORFF distance h(E, F (E)) becomes small
(Figure 2.37). This is not surprising, since if E is close to A then so is F (E).

Figure 2.36. A quadrangle E approximating the BARNSLEY fern.

Figures 2.38, 2.39 and 2.40 show how the structure of F (k) (E) develops as k increases: the region towards the top inherits more and more of the lower part, while the
lower branches profit more and more from the entire structure.
The relation between the sets E and F (E) just mentioned above works the other
way too: if h(E, F (E)) is small, then by the Collage Theorem 2.2.11 the set E must be
close to the limiting fractal A. This opens the possibility to construct a fractal looking
approximately like a given set A. The recipe is as follows: construct (a) a polygon E
approximating as well as desired in the H AUSDORFF metric the prospective fractal
set A; construct (b) affine contractions fj (1 j J) with (c) contraction constants as
small as desired (all of them less than c < 1, say) such that (d) the set E is approximated

as well as desired by F (E) := Jj=1 fj (E). Then by the Collage Theorem 2.2.11
for the fractal AF , the fixed point of F , we get h(A, AF ) h(A, E) + h(E, AF )

Section 2.3 Affine iterative function systems in R2

Figure 2.37. The generating set E(1) for the BARNSLEY fern, and the quadrangle E used as initial set.

Figure 2.38. The approximating set E(2) for the BARNSLEY fern, if E(0) is the
quadrangle E.

101

102

Chapter 2 Iterative function systems

Figure 2.39. The approximating set E(4) for the BARNSLEY fern, if E(0) is the
quadrangle E.

Figure 2.40. The approximating set E(8) for the BARNSLEY fern, if E(0) is the
quadrangle E.

Section 2.3 Affine iterative function systems in R2

103

1
h(A, E) + 1c
h(E, F (E)), which then becomes as small as desired.
Of course we also need a recipe for how to satisfy all these desires. Problem (a)
does not offer any special difficulties. The problems (b), (c) and (d) may be tackled
with the help of the following theorem.

2.3.5 Theorem. Let pi = (pi,1 , pi,2 ) (1 i 3) be the vertices of a non-degenerated

triangle. Given any three points qi = (qi,1 , qi,2 ) there is an affine map f satisfying
f (pi ) = qi (1 i 3).
Proof. We look for a 2 2-matrix L and a vector b R2 ,

L =

l11
l21

l12
l22


,

b =

b1
b2


,

satisfying
qi,j = lj,1 pi,1 + lj,2 pi,2 + bj

(1 i 3, j = 1, 2).

This provides for two systems of three inhomogeneous linear equations each for the
unknown variables (lj,1 , lj,2 , bj ). It is uniquely solvable if (bars | | and det denoting,
as usual, determinants)
0 =
=

p1,1
p2,1
p3,1

p1,2
p2,2
p3,2

p2,1 p1,1
p3,1 p1,1

1
1
1

p1,1
p2,1 p1,1
p3,1 p1,1

p2,2 p1,2
p3,2 p1,2

p1,2
p2,2 p1,2
p3,2 p1,2

1
0
0

= det(p2 p1 , p3 p1 ).

This condition is satisfied since the triangle p1 , p2 , p3 is supposed to be non-degenerate.

2.3.6

Construction of a blossom fractal

Let us try an example. Suppose we want to picture by means of a fractal a blossom

with five leaflets fitting roughly into the polygon with vertices a = (0.50, 0.30), b =
(0.70, 0.40), c = (0.65, 0.60), d = (0.45, 0.55), e = (0.35, 0.40) (Figure 2.41a). To
this end we cover the region in the interior P of the polygon, which is supposed to
display eventually the desired fractal, by five affine images of P (Figure 2.42a).
How have we arrived at these images? We have chosen the images of the triangle
ace in a suitable way (Figure 2.42b); this choice requires a little bit of geometric intuition, but everything not to our taste may simply be discarded and replaced by another
triangle aj cj ej until we are satisfied the result complies with requirement (d). These

104

Chapter 2 Iterative function systems

(a)

(b)

Figure 2.41. A pentagon abcde (2.41a) as initial set for a prospective blossom
(2.41b).

(a)

(b)

Figure 2.42. The pentagon abcde (2.42a) and the triangle ace (2.42b), each with
its five images under the affine mappings fj (1 j 5).

choices require the affine mappings fj (1 j 5) to furnish the following images:

f1 :

a1 = (0.50, 0.40),

c1 = (0.38, 0.45),

e1 = (0.43, 0.34),

f2 :

a2 = (0.50, 0.45),

c2 = (0.55, 0.57),

e2 = (0.42, 0.50),

f3 :

a3 = (0.55, 0.50),

c3 = (0.68, 0.50),

e3 = (0.60, 0.59),

f4 :

a4 = (0.60, 0.50),

c4 = (0.60, 0.35),

e4 = (0.69, 0.45),

f5 :

a5 = (0.55, 0.40),

c5 = (0.45, 0.33),

e5 = (0.60, 0.35).

This produces mappings fj (1 j 5) given by the following matrices Lj and

translation vectors bj , with fixed points zfj , determinants dfj and contraction constants

Section 2.3 Affine iterative function systems in R2

105

cfj (all entries rounded to the given decimals):

j

1


Lj

0.15 0.48
0.38 0.02

0.48
0.05

0.03
3
0.45

0.45
4
0.00

0.42
5
0.13
2


0.08
0.43

0.45
0.22

0.22
0.50

0.13
0.30







bj

0.57
0.22
0.28
0.35
0.43
0.66
0.76
0.65
0.80
0.42







zfj

0.455
0.381
0.462
0.564
0.628
0.484
0.590
0.433
0.528
0.380

dfj

cfj

0.18

0.53

0.20

0.52

0.19

0.59

0.23

0.60

0.14

0.44






The resulting fractal is illustrated in Figure 2.41b.

2.3.7

A moving grass (cf. Figure 3.11.3 in [Barnsley, 1988])

According to Theorem 2.2.12 a small change in the maps fj of a given IFS will result
in a small change of the attractor. In fact, it is possible to insert such a changing
mechanism already in the definition of the IFS, as e.g. already done in the examples of
Section 2.3.4.5 and Section 2.3.4.6. Let us conclude this section with the fractal image
of a grass with a variable upper part.




0 0
0.5
L1 =
,
b1 =
,
0 0.3
0.11


0.60 cos 0.032 sin 0.698 sin
,
L2 =
0.032 cos + 0.60 sin 0.698 cos


0.3 cos + 0.1 sin + 0.5
b2 =
,
0.04 cos 0.24 sin + 0.3




0.322 0.414
0.404
L3 =
,
b3 =
,
0.278 0.414
0.008




0.246 0.437
0.312
L4 =
,
b4 =
.
0.368
0.223 0.349
As increases from its initial value 0 the grass begins to bow in the wind until it
finally breaks.

106

Chapter 2 Iterative function systems

Figure 2.43. The generating set A(1) for the grass if the initial set is the boundary
of the unit square.

Figure 2.44. The approximating set A(12) for the grass, with the boundary of
the unit square as initial set.

Figure 2.45. The generating set A(1) for the grass with an inclination angle
= 5 if the initial set is the boundary of the unit square.

Section 2.3 Affine iterative function systems in R2

107

(a) = 5

(b) = 10

(c) = 15

(d) = 20

(e) = 30

(f) = 45

Figure 2.46. The approximating set A(12) for the grass, with different values of
and the boundary of the unit square as initial set.

Iteration of complex polynomials

After having discussed the M ENGER sponge we have already decided to stay in R2 . We
still have the possibility to identify R2 with the complex plane C and to make use of the
algebraic operations defined for complex numbers z = x + iy . For instance we could
use linear maps of the form f (z) = cz + b (c = c1 + ic2 , b = b1 + ib2 , cj R, bj R).
If we would limit ourselves to such maps we would lose a good part of our freedom
to create fractals since all of these maps are similarities. If c is written as c = |c|ei ,
then the map f turns everything in C about the angle (in radials) counterclockwise,
multiplies the distances from the origin 0 by |c| and finally moves the result a distance
of |b| in direction of the vector b. If |c| < 1, then f contracts with similarity factor
b
s = |c| and has the fixed point zf = 1c
. Every orbit {f (k) (z)}
k=0 converges to zf .
Something new happens if we decide to study orbits under more general maps than
linear ones, and a natural next choice would be a polynomial of the form f (z) =

n
j
j=0 cj z (cj C, 0 j n, cn = 0). We have to think about only where to look
for the equivalent of an attractor as in Chapter 2, which should be a set invariant under
the action of f . Funnily enough it turns out that the interesting object connected with
the iterations of a polynomial could be rather called a repeller.
In order to get an impression of what will be waiting for us, let us consider a particularly simple example, namely the polynomial f (z) = z 2 . This polynomial even has
two fixed points: 0 and 1, the two solutions of the equation f (zf ) = zf . But f behaves
differently in the neighbourhood of 0 and 1. In general this behavior at a point z0 is
regulated by the value of the derivative of f , as becomes clear when a (small) complex
increment h is added to z0 :
f (z0 + h) = f (z0 ) + h f  (z0 ) + o(h)
as h 0,

< |h| if |f  (z0 )| < 1,
o(h)
|f (z0 + h) f (z0 )| = |h| f  (z0 ) +
h
> |h| if |f  (z0 )| > 1.

Taking z0 = zf we get

o(h)
|f (zf + h) zf )| = |h| f (zf ) +
h


< |h|
> |h|

if |f  (zf )| < 1,
if |f  (zf )| > 1.

Since f  (z) = 2z we have f  (0) = 0 and f  (1) = 2. So zf = 0 is an attractive fixed

point, while zf = 1 is a repellent one.
A generalization of a fixed point would be a point z0 which the orbit {f (k) (z0 )}
k=1
visits again: f (p) (z0 ) = z0 . Such a point is called periodic and the smallest positive
integer p which does the trick is called the period of z0 . Looking for the periodic points
k
of our map f we first note that f (2) (z) = z 4 , f (3) (z) = z 8 and f (k) (z) = z 2 . So, apart
from z0 = 0, we have to look for points z0 satisfying
p

z02 = z0 ,

z02

= 1,

110

Chapter 3 Iteration of complex polynomials

which implies |z0 | = 1. Writing z0 = e2i0 we get

 2i 2p 1
p
0
e
= e2i0 (2 1) = 1.

Given p N and asking which values of 0 satisfy this equation we get as an answer:
0 (2p 1) must be an integer
n. Asking then for which integers n we get different
n
periodic points zn = e2i 2p 1 , we see that n {0, . . . , 2p 2} does it. All of the
corresponding periodic points lie on the unit circle, and since the points 2pn1 (p N,
0 n < 2p 1) form a dense subset of [ 0, 1 ] the non-zero periodic points form a
dense subset of the unit circle.
What can be said about the behavior of f in the neighbourhood of a periodic point
z0 ? Suppose its period is p and zj := f (j) (z0 ) (0 j < p). Then
[f (p) ] (z) = [f (f (p1) )] (z) = f  (f (p1) (z)) [f (p1) ] (z)
=

p
!1

f  (f (j) (z)),

j=0

[f (p) ] (z0 ) =

p
!1

f  (zj ) = [f (p) ] (zj )

(0 j < p)

j=0

=
[f (p) ] (zj )

p
!1

p
!1

j=0

j=0

(2zj ) = 2p

zj ,

= 2p > 1.

The conclusion is that all periodic points with period p (as, in fact, also all other points)
on the unit circle are repellent with respect to f (p) . What, then, happens to the orbits
of the various points z ? For |z| = 1 they are bound to stay and move around on the unit
circle. For |z| < 1 the orbits converge to the attractive fixed point 0, and for |z| > 1
they diverge to . Adjoining the point to C one obtains a compactification C
of C. Since C may also be obtained by projecting every point of C from the north
pole of a sphere, which touches the complex plane tangentially at the south pole, upon
the sphere, and adjoining the north pole, C is called the R IEMANN sphere. With this
understanding the unit circle forms the boundary between the attraction areas under
the iteration of f of the points 0 and of the R IEMANN sphere.
This peaceful and beautiful picture is roughly disturbed if instead of f (z) = z 2 we
consider the function f (z) = z 2 + c for a non-zero complex number c, even if |c| may
be arbitrarily small. If |c| is sufficiently small there will still be two fixed points an
attractive and a repellent one, za and zr , say, and there will still be a boundary
between the points with orbits converging to za and the points with orbits converging
to . This boundary will be invariant under the action of f and will necessarily have
to contain all periodic points unequal to za . But it seems hard to imagine how this
boundary called a J ULIA set might look. This will be what we are after.

111

Section 3.1 General theory of J ULIA sets

3.1

General theory of J ULIA sets

In this section we follow the expositions given by [Blanchard, 1984], [Falconer, 1990].
As pointed out before we shall be concerned with the behavior of orbits {f (k) (z)}
k=1
under a polynomial given by
f (z) =

cj z j

(cj C, 0 j n, cn = 0)

(3.1)

j=0

of degree n 2. Although we eventually shall apply our knowledge to quadratic

polynomials only, there is nothing to be gained in the general theory by limiting oneself
to n = 2, and therefore in this section we shall not do so. Still, we shall require some
facts from the theory of functions of a complex variable. To this end, apart from the
better known ones, we shall collect some important theorems here in advance. We shall
denote the closed disc with radius centered at z by B(z, ) and its interior by U (z, ).
The closure of a set V will be denoted by V , not to be confused with the complex
conjugate z of a complex number z .
3.1.1 Theorem (Mapping theorems from the theory of functions of a complex variable).
(a) (Open mappings) Let f be a holomorphic, non-constant function on a complex
domain D (i.e. a connected open subset of C). Then the image f (E) of an open
subset E of D is again an open set. [Cartan, 1966, p. 178]
(b) (Uniform convergence of holomorphic functions) The limit function f of a sequence {fn }
n=1 of functions, holomorphic on a complex domain D , which converges uniformly on compact subsets of D is holomorphic in D and the sequence

of derivatives {fn }
n=1 converges uniformly on compact subsets of D to f .
[Cartan, 1966, Theorem 1 and 2, p. 145]
(c) (Inverse function) Suppose the function f is holomorphic in a neighbourhood of
z0 C and suppose f  (z0 ) = 0. Then there exists a neighbourhood U of z0 , a
neighbourhood V of the point w0 = f (z0 ), and a function g which is uniquely
determined and holomorphic in V , such that g(w0 ) = z0 and w = f (z) (z
U ) z = g(w) (w V ). [Cartan, 1966, Theorem 1.1, p. 175]
(d) (Monodromy theorem) The analytic continuation of a function holomorphic in
some open subset of a simply connected domain to this simply connected domain
is unique. [Nevanlinna and Paatero, 1965, 2, p. 246]
3.1.2 Definition. A fixed point z0 C of f (f (z0 ) = z0 ) is called
super-attractive

if f  (z0 ) = 0,

attractive

if |f  (z0 )| < 1,

neutral

if |f  (z0 )| = 1,

repellent

if |f  (z0 )| > 1.

112

Chapter 3 Iteration of complex polynomials

We shall also consider the point C as a fixed point of f . As a consequence of

the assertion of Lemma 3.1.10 it should be considered as attractive.
3.1.3 Definition. A point z0 C is called p-periodic for f (p N) if f (p) (z0 ) = z0 .
The minimal number p N satisfying f (p) (z0 ) = z0 is called the period of z0 .
3.1.4 Theorem. Let z0 be a p-periodic point for f and let zk := f (k) (z0 ) (0 k < p).
Then
p
!1

f (p) (zk ) =
f  (zn ).
n=0

Proof. Without loss of generality we suppose k = 0.

f

(p) 

(z0 ) = f (f

(p1)

(z0 )) f

(p1) 

(z0 ) =

p
!1

f  (zk ).

k=0

3.1.5 Definition. A periodic orbit {zk = f

and each of its points zk is called
super-attractive
attractive

(k)

1
(z0 )}p
k=0

(f

(p)

(zk ) = zk , 0 k p 1)

if (f (p) ) (z0 ) = 0,
if |(f (p) ) (z0 )| < 1,

neutral if |(f (p) ) (z0 )| = 1,

repellent

if |(f (p) ) (z0 )| > 1.

3.1.6 Definition. The J ULIA set J(f ) is the closure of the set of all repellent periodic
points for f . The complement F (f ) = C \ J(f ) of the J ULIA set J(f ) is called the
FATOU set of f .
The content of the rest of this section is to make sure that (just as in the case of
f (z) = z 2 ) the J ULIA set is not empty, compact and perfect (i.e. it does not contain
any isolated points), nowhere dense, invariant under the action of f and f (1) , and that
it is the boundary of the set of complex numbers with unbounded orbits. In order to
get to know the J ULIA set it is advisable to make a detour. The set J0 (f ) which we
are going to define will turn out to be identical with J(f ). But the definition of J0 (f ),
complicated as it may seem, allows an easier access to the properties of J0 (f ).
3.1.7 Definition. A sequence {gn }
n=0 of complex-valued functions gn which are holomorphic in an open set U C is called normal in U if every subsequence of {gn }
n=0
again admits a subsequence {gnk }
k=0 which converges uniformly on all compact subsets K U either towards a function holomorphic on U or towards , i.e. which
either satisfies
lim sup |gnk (z) gnl (z)| = 0

k,l zK

or

lim inf |gnk (z)| = .

k zK

The sequence {gn }

n=0 is called normal in w C if it is normal in some open neighbourhood V of w.

113

Section 3.1 General theory of J ULIA sets

3.1.8 Theorem (M ONTEL

). Suppose the sequence {gn }
n=0 is not normal

 in an open
set U C. Then either n=0 gn (U ) = C or there is a w C such that
n=0 gn (U ) =
C \ {w}.
Proof. See [Saks and Zygmund, 1971, Chapter VII 13, Exercise 2].
3.1.9 Definition.

z C : {f (n) }
n=1 is not normal in z .


F0 (f ) := z C : {f (n) }
= C \ J0 (f ).
n=1 is normal in z
J0 (f ) :=

Evidently the set F0 (f ) is open; consequently J0 (f ) is closed. We shall simplify

the notation by writing J0 = J0 (f ), F0 = F0 (f ).
3.1.10 Lemma. For f as in (3.1) let


r0 (f ) := max 1,

2+

n1
j=0

|cj | 

|cn |

Then for |z| > r0 (f ) one has

|f (z)| > 2|z|.

Consequently
lim f (k) (z) =

uniformly on
V := {z : |z| > r0 (f )} = C \ B(0, r0 (f )).

(3.2)

Proof. For |z| > r0 (f ) we have

|cn | |z|

|cj | |z|jn+1 |cn | |z|

j=0

|cj | > 2,

j=0
n1

|f (z)| |z|

n



1
|cn | |z|
|cj | |z|jn+1 > 2|z|.
j=0

3.1.11 Theorem. J0 is compact.

Proof. The sequence {f (k) }
k=1 is normal on the set V as defined in (3.2). Consequently J0 B(0, r0 (f )) is bounded.

114

Chapter 3 Iteration of complex polynomials

3.1.12 Theorem. J0 = .
Proof. The fixed points of f are the n solutions zj (1 j n) of the equation
f (z) = z . Let r1 > max(r0 (f ), |z1 |). The open disc U (0, r1 ) with center 0 and radius
r1 contains a point z0 satisfying
k

|f (k) (z0 )| 2k |z0 | .

Suppose J0 = . Then the sequence {f (k) }

k=1 would have to be normal in U (0, r1 ).
But it is impossible for any subsequence of {f (k) }
k=1 to converge uniformly on a
compact subset of U (0, r1 ) containing both z0 and z1 .
3.1.13 Theorem. J0 = f 1 (J0 ) = f (J0 ).
Proof. Note that the map f : C C is surjective. We shall show F0 = f 1 (F0 ) =
f (F0 ). This implies f 1 (J0 ) = f 1 (C \ F0 ) = C \ F0 = J0 and f (J0 ) = J0 . To
simplify the notation again we shall write {f (n) }n for {f (n) }
n=1 .
Claim 1: f (1) (F0 ) F0 . Since f is surjective this implies F0 f (F0 ).
Proof: Let z0 f (1) (F0 ) be given. We shall exhibit an open neighbourhood U of
z0 on which the sequence {f (n) }n is normal. To this end, let V be an open neighbourhood of f (z0 ) F0 on which {f (n) }n is normal, and let U := f (1) (V ). Given



a subsequence {f (n ) }n of {f (n) }n let {f (n 1) }n be a subsequence of {f (n 1) }n ,
uniformly converging on compact subsets of V . Let K be a compact subset of U . Then

f (K) V is again compact. The sequence {f (n 1) }n converges uniformly on f (K)

and does so by continuity {f (n ) }n on K .
Claim 2: f (F0 ) F0 . This will imply F0 f (1) (f (F0 )) f (1) (F0 ).
Proof: We have to show: Given z0 F0 there is an open neighbourhood U of f (z0 ) on
which the sequence {f (n) }n is normal. Let V0 and V1 be bounded open neighbourhoods
of z0 satisfying z0 V1 V1 V0 on which {f (n) }n is normal. The map f is open
(Theorem 3.1.1(a)). Therefore U := f (V1 ) is an open neighbourhood of f (z0 ). Given



a subsequence {f (n ) }n of {f (n) }n , let {f (n +1) }n be a subsequence of {f (n +1) }n
which converges uniformly on compact subsets of V0 . Let K be a compact subset of
U = f (V1 ). Then K0 := f (1) (K) V1 is a compact subset of V0 . The sequence


{f (n +1) }n converges uniformly on K0 , therefore so does the sequence {f (n ) }n on
f (K0 ) = f (f (1) (K) V1 ) = K .
3.1.14 Theorem. F0 contains every attractive periodic orbit of f .
Proof. Let p be the period of z0 under f and let zj := f (j) (z0 ) (0 j < p). It suffices
to show z0 F0 . Choose any s ] |(f (p) ) (z0 )|, 1 [ and any d > max{|(f (j) ) (z0 )| : 0
j < p}. Let be so small that for all z V = U (z0 , ) one has
|f (j) (z) zj | < d |z z0 |
|f

(p)

(0 j < p),

(z) z0 | < s |z z0 | (< )

f (p) (z) V.

and therefore

115

Section 3.1 General theory of J ULIA sets



Let {f (k ) }k be a subsequence of {f (k) }k . For some j (0 j < p) there has to be



a subsequence {f (k ) }k of {f (k ) }k satisfying k  j (mod p) for all k  . If z V
then


|f (k ) (z) zj | = |f (mp+j) (z) zj |

= |f (j) (f (mp) (z)) f (j) (z0 )|
< d |f (mp) (z) z0 |
d sm |z z0 |.


Therefore the sequence {f (k ) }k converges uniformly on V to the constant function

zj .
3.1.15 Theorem. J0 contains every repellent periodic orbit of f . Consequently one
has J(f ) J0 .
Proof. Let z0 be a repellent periodic point for f with period p. z0 is a repellent fixed
point of the polynomial g = f (p) . Suppose the sequence {g (k) }k is normal in z0 .

Then there would be a neighbourhood V of z0 and a subsequence {g (k ) }k of {g (k) }k
converging uniformly on every compact subset K V . Since z0 is a fixed point
of g , the limit function g0 has to be finite and holomorphic in V . The same holds

for g0 = limk (g (k ) ) (Theorem 3.1.1(b)). But z0 is a repellent fixed point of g .


Therefore one has limk (g (k ) ) (z0 ) = limk (g  (z0 ))k = . Consequently the
(k)
sequence {g }k is not normal in z0 and z0 J0 . Finally, since J0 is closed, we also
have J(f ) J0 .
3.1.16 Definition. Let z0 be an attractive fixed point for f (either z0 C or z0 = ).
The set A(f, z0 ) := {z C : limk f (k) (z) = z0 } is called the attractive basin of z0 .
Note that as an immediate consequence of the definition one gets A(f, z0 ) =
f (1) (A(f, z0 )).
3.1.17 Theorem. If z0 is an attractive fixed point for f , then A(f, z0 ) is open and
A(f, z0 ) F0 .
Proof. Let first z0 = . Choose s > 0 so as to satisfy |f  (z0 )| < s < 1. Then for a
suitable > 0 and the open disc V = U (z0 , ) one has
zV

|f (z) z0 |
< s,
|z z0 |

and consequently V A(f, z0 ). As in the proof of Theorem 3.1.14 the sequence

{f (k) }k converges on V uniformly to the constant z0 . Consequently one has V F0 .
Given z A(f, z0 ) there is a k N satisfying f (k) (z) V . We obtain
A(f, z0 ) =


k=0

f (k) (V )


k=0

f (k) (F0 ) = F0 .

116

Chapter 3 Iteration of complex polynomials

Suppose now z0 = . Choose V F0 as in (3.2). Then, as before, the set

A(f, ) =


k=0

f (k) (V )

f (k) (F0 ) = F0

k=0

is open.
3.1.18 Definition. For z F0 let C(f, z) be the connected component of F0 containing
z.
1
(p)
(z0 ) = z0 ) is an attractive periodic orbit, then by TheoIf {zj = f (j) (z0 )}p
j=0 (f
rem 3.1.17 for 0 j < p one has C(f, zj ) A(f (p) , zj ). Otherwise there would be a
point z A(f (p) , zj ) C(f, zj ) F0 ; for every neighbourhood V C(f, zj ) of z
the sequence {f (kp) (z)}k would converge to zj on V A(f (p) , zj ) but remain outside
of A(f (p) , zj ) on V \ A(f (p) , zj ); this contradicts z F0 . As a consequence, the sets
C(f, zj ) A(f (p) , zj ) have to be disjoint.
1
3.1.19 Theorem. Let {zj = f (j) (z0 )}p
(f (p) (z0 ) = z0 C) be an attractive periodic
p1 j=0
 1
(p)
, zj ) F0 contains at least
orbit for f . Then the set E = j=0 C(f, zj ) p
j=0 A(f
one critical value of f , i.e. a number w = f (z), the pre-image z of which satisfies
f  (z) = 0.

Proof. Let us assume the converse hypothesis. We shall lead it to a contradiction in a

series of steps.
Claim 1: f  (z) = 0 for all z E .
Proof: For 0 j < p we shall write j  := j + 1 (mod p). By Theorem 3.1.1(a) the
set f (C(f, zj )) is open. It contains zj and, by continuity, it is connected. Therefore
one has f (C(f, zj )) C(f, zj ) and f (E) E = C. If for some z C(f, zj ) we had
f  (z) = 0, then f (z) C(f, zj  ) would be a critical value of f in E , the possibility of
which we have excluded.
Claim 2: The functions gk = f (kp) (1 k < ) are well defined and form a normal
family on a sufficiently small simply connected neighbourhood of z0 .
 1  (j)
(z)) = 0 for all k N and all
Proof: Note that, by claim 1, [f (k) ] (z) = k
j=0 f (f
z E . By Theorem 3.1.1(c) the function f (kp) = (f (p) )(k) (k N) is locally invertible
at each z C(f, z0 ). Let U C(f, z0 ) be a simply connected neighbourhood of
z0 . By Theorem 3.1.1(d) the inverse function gk of f (kp) , obtained by starting with
gk (z0 ) = z0 and analytic continuation to all of U , is uniquely defined and holomorphic.
By M ONTELs Theorem 3.1.8 the sequence of holomorphic functions {gk }k on U must
be normal since gk (U ) C(f, z0 ) E for all k N.
(k)
Claim 3: z0 is a repellent fixed point for the sequence {gk }
k=1 = {g1 }k=1 .
(kp)

(p) 
Proof: We have gk (z0 ) = f
(z0 ) = z0 and s = |g1 (z0 )| = 1/|(f ) (z0 )| > 1.
As in the proof of Theorem 3.1.15, claim 3 implies that the sequence {gk }
k=1 cannot be normal in z0 , in contradiction with claim 2.

Section 3.1 General theory of J ULIA sets

117

3.1.20 Theorem. The number of different attractive periodic orbits is less than n.
p2
1
Proof. Suppose {z1,j }pj=
1 and {z2,j }j=1 are two different attractive periodic orbits.
Without loss of generality we may suppose that C(f, z1,0 ) and C(f, z2,0 ) contain the
critical values w1 and w2 respectively of f . If w1 = w2 then C(f, z1,0 ) = C(f, w1 ) =
C(f, z2,0 ). This implies z1,0 = limk f (kp1 ) (z2,0 ) = limk f (kp1 p2 ) (z2,0 ) = z2,0
which is impossible. Therefore different attractive periodic orbits must correspond to
different critical values of f . But there are at most n 1 of those.

At this point one wonders about the number N of neutral periodic points. The idea
is that maybe a small variation of the function f may induce also a small change of
the periodic points in such a way as to keep the attractive periodic points attractive
but change at least some of the originally neutral periodic points into attractive ones
and then to apply Theorem 3.1.20. In fact it may be shown that changing the function f to fw given by fw (z) = (1 w)f (z) + w, taking w C sufficiently small
and with a suitable argument, this is accomplished with at least N/2 originally neutral
periodic points changing into attractive ones. So the number of non-repellent periodic
points cannot exceed 2n 2. Unfortunately the argument is too elaborate to be reproduced here. It may be found in [Blanchard, 1984, Theorem 5.12, p. 111]. Furthermore,
D OUADY [Douady, 1983] has shown that the total number of non-repelling periodic
points is even less than n. So we have to be content here with an assertion to be proved
elsewhere:
3.1.21 Theorem. The number of neutral periodic orbits is finite.
Proof. See [Blanchard, 1984, p. 111].
3.1.22 Theorem. If the orbits of all critical points of f are unbounded, then there are
no attractive periodic orbits of f .
Proof. Our hypothesis implies that limk f (k) (w) = for all critical values w of
1
f . Suppose {zj }p
j=0 is an attractive periodic orbit of f . Without loss of generality we
suppose that C(f, z0 ) contains a critical value w of f . Then limk f (kp) (w) = z0 in
contradiction to our hypothesis.
3.1.23 Theorem. J0 (f (m) ) = J0 (f ) for all m N.
Proof. We shall show F0 (f (m) ) = F0 (f ).
Claim 1: F0 (f ) F0 (f (m) ).
Proof: Suppose the sequence {f (k) }k is normal on the open neighbourhood U of z0


F0 . A subsequence {(f (m) )(k ) }k = {(f (mk ) }k of the sequence {(f (m) )(k) }k is also

(k)
a subsequence of the sequence {f }k . Therefore there is a subsequence {f (mk ) }k
(mk ) 
of {f
}k , which converges uniformly on all compact subsets of U .
Claim 2: F0 (f (m) ) F0 (f ).
Proof: Suppose the sequence {(f (m) )(k) }k is normal on the open neighbourhood U of

z0 F0 (f (m) ), and let {f (k ) }k be a subsequence of the sequence {f (k) }k . At least

118

Chapter 3 Iteration of complex polynomials





for one r (0 r < m) there is a subsequence of {f (k ) }k of the form {f (mk +r) }k .


The sequence {f (mk ) }k contains a subsequence {f (mk ) }k , which converges uni
formly on all compact subsets of U . Consequently, also the sequence {f (mk +r) }k
converges uniformly on all compact subsets of U .
3.1.24 Theorem. Let f be a polynomial of degree n > 1 and let U be a neighbourhood
of a point w J0 . Then


f (k) (U ) J0
k=1

and either

f (k) (U ) = C

k=1

or there exists a number c C and a fixed point z0 F0 (f ) of f satisfying

f (k) (U ) = C \ {z0 }

f (z) = z0 + c (z z0 )n .

and

k=1

 (k)
Proof. The sequence {f (k) }
(U ).
k=1 is not normal on U . Suppose C = W =
k=1 f
Then, by the M ONTELs Theorem 3.1.8, there exists a point z0 C satisfying W =
C \ {z0 }. Let zi (1 i n) be the roots of the equation f (z) z0 = 0. If for
some i (1 i n) we had zi W , then z0 = f (zi ) f (W ) W , the possibility
of which we have excluded. As a consequence we get zi = z0 (1 i n) and
f (z) z0 = c (z z0 )n . Suppose now |z z0 | r := (2|c|)1/(n1) . We get
|f (z) z0 | = |c| |z z0 |n
|f (k) (z) z0 |

|c|
1
|z z0 | = 2 |z z0 |,
2|c|

1
r
|z z0 | k .
2k
2

Therefore, on the closed disc B(z0 , r) the sequence f (k) (z) converges uniformly to the
constant z0 . As a consequence, the sequence {f (k) }
k=1 is normal in z0 .

(k)
3.1.25 Theorem. Let U be an open set satisfying U J0 = , and let w
(U )
k=1 f
(k)
be given. Then f
(w) U = for infinitely many k N.

(k)
(U ) = C let z0 be defined as in Theorem 3.1.24. Let k1 :=
Proof. If W =
k=1 f
(k)
min{k N : w f (U )}. There exists a point w1 U such that f (k1 ) (w1 ) = w. In
case W = C we have w1 = z0 since f (k1 ) (z0 ) = z0 = w. Consequently, in any case,
we have w1 W . By induction we may therefore construct a monotone increasing

(kj )
(w) U
sequence {kj }
j=1 N and a sequence {wj }j=1 C satisfying wj f
(1 j < ).
3.1.26 Theorem. For every point w J0 one has
J0 =


k=1

f (k) (w).

119

Section 3.1 General theory of J ULIA sets

Proof. For w J0 Theorem 3.1.13 furnishes f (k) (w) J0 and

f (k) (w) J0 .

k=1

Let U be an arbitrary open set meeting J0 . By Theorem 3.1.24 we have

w

f (k) (U )

k=1

and therefore

f (k) (w) U =

k=1

and

f (k) (w) J0 .

k=1

3.1.27 Theorem. J0 is nowhere dense.


(k)
Proof. If J0 contained an open set U , Theorem 3.1.13 would yield
(U ) J0 .
k=1 f
By Theorem 3.1.24 this is impossible since J0 is compact (Theorem 3.1.11).

3.1.28 Theorem. J0 is perfect, i.e. does not contain any isolated points.
Proof. Let U be an open neighbourhood of a point w0 J0 . We want to show that U
contains a point w J0 different from w0 . We distinguish three different cases.
(a) w0 is not a periodic point of f . Then all pre-images of w0 under f are mutually different. Let
w0 = f (w1 ). Then w1 = w0 and Theorems 3.1.13 and 3.1.24
(k)
(U ) and w0
/ f (k) (w1 ) J0 . However, by Theoyield w1 J0
k=1 f
(k)
(w1 ) U = for infinitely many k N.
rem 3.1.25, one has f
(b) w0 = f (w0 ) is a fixed point of f . If all pre-images f (1) (w0 ) coincide with
w0 , then as shown in the proof of Theorem 3.1.24 f (z) = w0 + c(z w0 )n and
w0 F0 which is impossible. If w1 = w0 and f (w1 ) = w0 , then as in case (a) we
/ f (k) (w1 ) J0 and f (k) (w1 ) U = for infinitely many k N.
get w0
(c) w0 = f (p) (w0 ) (p > 1). The reasoning already employed in case (b) applies for
the polynomial f (p) and the set J0 (f (p) ) = J0 (f ) (Theorem 3.1.23).
3.1.29 Theorem. J(f ) = J0 .
Proof. In Theorem 3.1.15 it has already been shown that J(f ) J0 . So it remains to
show J(f ) J0 .
There are only finitely many fixed points and critical values of f . Eliminating these
points from J0 we obtain a set


J1 = w J0 : (f (w) = w) (f  (z) = 0 z f (1) (w)) .
(3.3)
Since J0 does not contain any isolated points we have J1 = J0 . We shall substantiate
our claim by showing J1 J(f ) = J(f ).

120

Chapter 3 Iteration of complex polynomials

Given w0 J1 , by (3.3) there is a z0 J0 \ {w0 } satisfying the relations f (z0 ) =

w0 , f  (z0 ) = 0. Then by Theorem 3.1.1(c) there are two disjoint arbitrarily small
open neighbourhoods U of z0 and W of w0 such that f maps U bijectively upon W .
Denote this bijection by f. The union P of all non-repellent periodic orbits is finite by
Theorems 3.1.20 and 3.1.21. Therefore the sets U and W may be chosen disjoint from
P . Define a sequence of functions hk which are holomorphic on W by
f (k) (w) w

hk (w) := (1)
(w) w
f

(w W, k N).

(k)
If the sequence {hk }
}k=1
k=1 were normal on W , then so would the sequence {f
have tobe, which is impossible because of w0 J0 . By Theorem 3.1.8 of M ONTEL
the set
k=1 hk (W ) contains at least one of the numbers 0 and 1. Consequently, either
there is a w1 W and a k1 N satisfying hk1 (w1 ) = 0, or there is a w2 W and a
k2 N satisfying hk2 (w2 ) = 1. In the first case we obtain f (k1 ) (w1 ) = w1 , in the second
case we obtain f (k2 +1) (w2 ) = w2 . We conclude that the point w0 may be approximated
arbitrarily close by repellent periodic points. It therefore must belong to J(f ).

3.1.30 Theorem. If z0 is an attractive fixed point for f or if z0 = , then A(f, z0 ) =

J(f ).
Proof. In any case we have z0
/ J(f ).
Claim 1: J(f ) A(f, z0 ).
Proof: Let z J(f ) be given. By Theorem 3.1.13 its orbit is contained in J(f )
and cannot converge to z0 . We conclude z
/ A(f, z0 ). Let U
 be an(k)arbitrary open
(U ) = and
neighbourhood of z . By Theorem 3.1.24 we have A(f, z0 )
k=1 f
consequently U f (k) A(f, z0 ) = U A(f, z0 ) = for at least one (in fact infinitely
many) k N. We conclude z A(f, z0 ).
Claim 2: J(f ) A(f, z0 ).
Proof: Suppose there were a z A(f, z0 ) \ J0 . Then there would also be an open

neighbourhood V of z and a subsequence {f (k ) }k converging uniformly on every
compact subset of V either to z0 = or to a holomorphic function g . On the open

set V A(f, z0 ) = the sequence {f (k ) }k certainly converges to z0 . Therefore it
has to converge on all of V , including on z , to z0 . This would imply z A(f, z0 ) in
contradiction to our hypothesis.
Let us try to visualize the different situations that occur corresponding to the number of attractive periodic orbits. If there are none of them apart from the attractive fixed
point , then J(f ) must consist of a nowhere dense but perfect set of gaps in the
complex plane. If there is exactly one attractive periodic orbit consisting of a fixed
point, then one can visualize J(f ) as the boundary between its attractive basin and the
attractive basin of . As soon as there are more than one attractive fixed points then
J(f ) is the boundary of each of the corresponding attractive basins which is going
hard on our intuition. The same situation already occurs if there is at least one attractive periodic orbit with period p > 1: Each point of this orbit is a fixed point of f (p) ,
and J(f ) and J(f (p) ) coincide by Theorem 3.1.23.

Section 3.2

121

J ULIA sets for quadratic polynomials

Everything stated in this section holds in adapted form also for rational functions, i.e. the quotient of two polynomials [Blanchard, 1984]. We, however, shall be
happily content to apply its assertions to the simplest situation, the case of a polynomial
f of second order.

3.2

J ULIA sets for quadratic polynomials

Properly speaking, a polynomial of second order would look like the equation g(w) =
c2 w2 + c1 w + c0 (c2 = 0). In order to simplify the situation let us try the substitution
w = h(z) = az+b with the inverse z = h(1) (w) = wb
a , which amounts to a similarity
transformation in the complex plane. Using the notation g h(z) = g(h(z)) we get
g h(z) = a2 c2 z 2 + a(2bc2 + c1 )z + c2 b2 + c1 b + c0 ,
h(1) g h(z) = ac2 z 2 + (2bc2 + c1 )z +

By choosing a =

1
c2 ,

b = 2cc12 , and c = c2 b

c2 b2 + (c1 1)b + c0
.
a

+(c1 1)b+c0
a

we obtain

ac2 = 1,

2bc2 + c1 = 0,
f (z) := h(1) g h(z)
= z 2 c,
f

(k)

= h

(1)

(3.4)

(k)

for all k N.

The form (3.4) for f which we shall use in the sequel is handy since it provides
immediate information on fixed points, critical values, and points with periodic orbits
of period 2. For fixed points we only have to solve the equation
z 2 z c = 0,

which furnishes the following answer:

3.2.1 Theorem. The fixed points of f as in (3.4) are
z1,2 =

1 + 4c
.
2

In order to find out whether a fixed point is attractive we investigate the derivative
of f .
f  (z) = 2z,
|f  (z1,2 )| = |1

1 + 4c |.

By Theorem 3.1.20 there can be at most one attractive fixed point.

Suppose we choose
the complex square root in such a way that |f  (z2 )| = |1 + 1 + 4c | 1. Let us

122

Chapter 3 Iteration of complex polynomials

first
locate the values c for which |1
1 1 + 4c = ei . We obtain

1 + 4c | = 1. To this end we write f  (z1 ) =

1 ei =

1 + 4c,

1 2ei + e2i = 1 + 4c,

4c = e2i 2ei .
Elementary geometric arguments show that for the points
w() = 2ei e2i = 2 cos cos 2 + i(2 sin sin 2)

lie on a cycloid, the path of the point 1 on the circle with center 2 and radius 1 rolling on
the circle with center 0 and
radius 1. Elementary analysis provides the points w(0) = 1

and w( 23 ) = 12 i 3 2 3 with tangents parallel to the real axis, and the points w() =

3 and w( 3 ) = 32 i 23 with tangents parallel to the imaginary axis.

The looked-for values of c are given byc = w()
4 , evidently the boundary between
those values of c for which |f  (z1 )| = |1 1 + 4c | < 1 and those for which |f  (z1 )| >

1. Since c = 0 lying inside
p given by
the cycloid provides for |f (z1 )| = 0 the function

p(c) := f (z1 (c)) = 1 1 + 4c maps the interior of the cycloid C0 = { 14 (e2i 2ei ) :

0 2} onto the interior of the unit circle. Since the points f  (z1,2 ) = 1 1 + 4c
lie symmetric with respect to the point 1 on the unit circle, if z1 = 1 lies in the unit
disc, then z2 has to lie in the exterior of it, and z2 has to be a repellent fixed point.

3.2.2 Theorem. The function f = z 2 c admits an attractive fixed point z1 = 1 21+4c

if and only if c lies inside the cycloid C0 = { 14 (e2i 2ei ) : 0 2}, and a
i
neutral fixed point z1 = e2 with f  (z1 ) = ei if and onlyif c = 14 (e2i 2ei ). In both
cases (except for c = 14 , = 0) the fixed point z2 = 1+ 21+4c is repellent and belongs
to the J ULIA set J(f ).
Points z on a periodic orbit of period 2 have to satisfy the equation
f (2) (z) z = 0,

(3.5)

where
f (2) (z) = f (z 2 c) = (z 2 c)2 c = z 4 2cz 2 + c2 c,
f (2) (z) z = z 4 2cz 2 z + c2 c.

(3.6)

Since also fixed points have to satisfy equation (3.5), the polynomial (3.6) has to be
dividable by the polynomial z 2 z c, the roots of which are the fixed points. Indeed
one has
z 4 2cz 2 z + c2 c = (z 2 z c)(z 2 + z + 1 c).
There evidently exists precisely one periodic orbit with period 2. Its points are the roots
z3,4 of the polynomial z 2 + z + 1 c. Whether it is attractive or not is decided by the
value
|[f (2) ] (z3,4 )| = |f  (z3 )| |f  (z4 )| = 4 |z3 z4 | = 4 |1 c|.

Section 3.2

123

J ULIA sets for quadratic polynomials

2
3.2.3 Theorem. For the polynomial f (z) =
z c there is precisely one periodic orbit
1
{z3 , z4 } with period 2, and z3,4 = 2 (1 4c 3). This orbit is attractive if and only
if |1 c| < 14 .

The last inequality is satisfied precisely if the point c lies in the interior of a circle
with center 1 and radius 14 touching the cycloid C0 in the point 34 .
There is only one critical point z0 = 0 and one critical value c. According to
Theorem 3.1.22 it is useless to look for attractive periodic orbits if the orbit of 0 is
unbounded. We can be more precise about that. For c = 2 we still get f (2) = 2, so 2
is a fixed point for f and f (k) (0) = 2 for all k 2. But a slightly larger value of |c|
changes the situation.
3.2.4 Theorem. If |z| = 2 + > 2 and |z| |c|, then
|f (z)| (1 + ) |z|

and

lim f (k) (z) = .

Consequently, for |c| > 2 the sequence {f (k) (0)}

k=0 = is unbounded.
Proof. For |z| |c| we have
|f (z)| =



c
z z z

= |z| z z



|c|
|z| |z|
|z|

|z| (2 + 1),
k

|f (k) (z)| (1 + )k |z| ,

|f (k) (0)| = |f (k1) (c)|
(1 + )k1 (2 + )

(if |c| 2 + )

While inf{Re (c) > 0 : limk f (k) (0) = } = 2 we can get an absolutely
smaller lower bound for the set {Re (c) < 0 : limk f (k) (0) = }.
3.2.5 Theorem. If Re (c) < 1, then limk f (k) (0) = .
Proof. Suppose Re (c) < 1 and |c| 2. We have
|f (2) (0)| = |f (c)| = |c2 c| = |c| |c 1| > 1 2,
|f (2) (0)| = 2 +

( > 0),
k

|f (k) (0)| = |f (k2) (f (2) (0))| (1 + )(k2) (2 + ) .

Theorems 3.2.4 and 3.2.5 furnish bounds for the famous set of all c with bounded
orbits {f (k) (0)}
k=1 [Mandelbrot, 1980].
3.2.6 Definition. The set M of all parameter values c for which the orbit {f (k) (0)}
k=1
is bounded is called the M ANDELBROT set.

124

Chapter 3 Iteration of complex polynomials

Figure 3.1. The bounds for the M ANDELBROT set established in Theorems
3.2.23.2.5.

3.3

The M ANDELBROT set

The M ANDELBROT set serves as a lexicon for the J ULIA sets. As already to be expected in view of its definition and of Theorems 3.2.2 and 3.2.3, the location of the
parameter c within the M ANDELBROT set furnishes information on properties of the
corresponding J ULIA set. One reason for the importance of this set lies in the assertion
of Theorem 3.3.2, for the proof of which we need a lemma. In order to comprehend
and visualize its statement it may help to imagine a radar located in the point c, scanning the set G, and a second one, located in the point 0, moved by the apparatus f (1) ,
turning with half the speed and recording the radial data of the first radar in coded form
by taking square roots.
3.3.1 Lemma. Suppose a simply connected domain G has a smooth boundary G.
Both if c G and if c G, the set f (1) Ghas a smooth boundary and may be
G0 = {w = r ei/2 : r = 0, 0 < 2, rei
represented as the union of two
sets
i/2
: r = 0, 0 < 2, rei c G} = G0 .
c G} and G1 = {w = r e
The sets G0 and G1 are central symmetric to each other and f maps each of them
one-to-one onto G. If c G, then the closure of f (1) (G) is simply connected. If
c G, then G0 and G1 are open and disjoint, and the boundary G0 G1 of
f (1) (G) contains precisely one double point at z = 0.
Proof. In both cases considered the boundary G may be written as a differentiable
curve {z(t) = r(t)ei(t) c : 0 t 1} where : [ 0, 1 ] [ 0, 2 ] is a differentiable
function and z(0) = z(1). In the first case (c G) a sufficiently small disc U (c, )
is contained in G; we have r(t) for all t [ 0, 1 ] and without loss of generality
(0) = 0, (1) = 2 . If c G then z(t) starts and ends at c = z(0) = z(1) and
is measured counterclockwise starting from the tangent to G in c. Without loss of
generality we may suppose (0) = 0, ( 12 ) = and r(0) = r( 12 ) = 0.

Section 3.3 The M ANDELBROT set

125

In the first case the set G0 contains inverse images of all points of G, but only half of
all inverse images, namely those with argument 2 [ 0, [ . The set G1 again contains
inverse images of all points of G, to wit the remaining half of the inverse images,
namely those with argument 2 in [, 2[ . The inverse image of the boundary G is

i(t)/2
the connected union of thetwo arcs { r(t)
: 0 t < 1} obtained by passing
e
i(t)/2
= r(t) ei((t)/2+) : 0 t < 1} obtained by
once through G, and { r(t) e
passing once more through G. As a closed continuous curve without multiple points
it is the boundary of a simply connected domain consisting of the union of the sets G0
and G1 .

In the second case the closed continuous curve G0 = { r(t) ei(t)/2 : 0 t 1}
starts at 0 with the real axis as a tangent and passes again (without loss of generality
for t = 12 and ( 12 ) = ) through 0 with the imaginary axis as a tangent. The set
G1 = G0 is bounded by the symmetric image of G0 with respect to the origin,
which continues with the same aforementioned tangent at 0 and ends tangentially on
the real axis in 0. Since G does not contain any multiple point, this also holds for its
one-to-one pre-images G0 and G1 . Suppose G0 and G1 would intersect in a point
p = 0. Then p = z(t1 ) = z(t2 ) where z(t1 ) G0 , z(t2 ) G0 , and t1 = t2 ; the
two different points z(t1 ) and z(t2 ) on G0 would have the same image on G under f
which is impossible. Consequently, the two domains G0 and G1 , bounded by G0 and
G1 respectively, are disjoint.
3.3.2 Theorem. The J ULIA set for the function f (z) = z 2 c is connected if and only
if c M .
Proof. Claim 1: If {f (k) (0)}
k=1 is bounded, then J(f ) is connected.
Proof: Let D = B(0, 2) be the closed disc with center 0 and radius 2. By Theorem 3.2.4 its complement V0 in C is part of the attractive basin A(f, ) of and
)
(D) D. Let Vk :=
satisfies f (V0 ) V0 . This implies V0 f (1) (V0 ) and f (1

(k)
(k)
(D) = f  (V0 ). Then Vk Vk+1 , and A(f, ) = k=1 Vk . Consequently
C\f

(k)
(D) = C\A(f, ). By Lemma 3.3.1 the sets f (k) (D)
we have c K := k=1 f
(which are the closures of the open sets f (k) U (0, 2)) are simply connected and so is
K . Therefore its boundary K = A(f, ) = J(f ) (Theorem 3.1.30) is connected.
Claim 2: If {f (k) (0)}
k=1 is unbounded, then J(f ) is disconnected.
Proof: Let D = U (0, r) be an open disc with center 0 and radius r > 2 such that J(f )
D and for some p N we have f (p) (0) D and f (k) (0) D for 1 k < p. Then
c = f (0) f (p1) (D). By Lemma 3.3.1 f (p) (D) is disconnected and consists
of two open components D0 and D1 , each of which is mapped onto f (p1) (D). The
inclusion J(f ) f (p1) (D) implies that J(F ) D1 = and J(F ) D2 = .
Consequently J(f ) is disconnected.
3.3.3 Theorem. The M ANDELBROT set is closed.
Proof. Let c0 M be given. Then |c0 | 2 by Theorem 3.2.4. The orbit of 0 under
f = fc0 may be approximated point-wise arbitrarily close by orbits of 0 under functions
f = fc for c M . These orbits by Theorem 3.2.4 must be contained in B(0, 2).

126

3.3.4

Chapter 3 Iteration of complex polynomials

The escape time approximation for the M ANDELBROT set

With the information acquired so far about bounds and structure of the M ANDELBROT
set M , one would like to be able to draw an at least approximative picture of M . There
is a convenient way to do so with the help of a computer. Establishing the real and
imaginary axes with the corresponding unit distances on the screen, we associate with
every pixel a complex number z . In particular we want to find out whether, given a
c C, the orbit of 0 under the corresponding function (3.4) diverges. This happens
if and only if for some k N we get |f (k) (0)| > 2. So we decide to wait at most
until k reaches a certain number n (e.g. n = 100 or n = 500) and to leave the pixel
corresponding to c white if |f (k) (0)| > 2 for some k n (the corresponding c certainly
does not belong to M ) or to paint it e.g. blue if this has not happened so far (the
corresponding c is under suspicion of belonging to M ) (Figure 3.2). In the resulting
blueprint of M we may have enriched M with some points c not belonging to M , since
the orbit of 0 under the iterates of the corresponding functions fc diverges too slowly
and its escape time exceeds n, but we can adjust the time limit n if we are not satisfied
with the resulting picture. Sometimes this has a surprising effect: Where there seemed
to be an open subset of M with a strange looking boundary, with an increase of the
allowed escape time limit n this subset all but disappears, leaving barely visible traces.
These traces, however, signal the existence of more points of M , since it has been
shown that M is connected [Douady and Hubbard, 1982]. The reader is kindly asked
to bear with the computer which in several of the following figures has difficulty
catching pixels of all connecting threads belonging to M .

Figure 3.2. An approximating set for the M ANDELBROT set: all points c C
for which |f (k) (0)| r = 2 for all k n = 100. The grid may help to locate
points c with respect to the M ANDELBROT set.

There is an additional feature we can incorporate in this procedure: having e.g. 16

different colours at our disposal we can observe to which residue class k modulo 16
the first exit index k belongs for which |f (k) (0)| > 2. If we paint the corresponding

Section 3.3 The M ANDELBROT set

127

pixel with colour number k, we partition the complement of M into layers indicating how close to M the corresponding values of c lie (the closer to M , the larger
the escape time k ). In the following figures featuring escape time colour charts these
colours appear in the following order: blue (1), green (2), cyan (3), red (4), magenta (5),
brown (6), light grey (7), dark grey (8), light blue (9), light green (10), light cyan (11),
light red (12), light magenta (13), yellow (14), white (15), black (160). So the blue
area in Figure 3.3 indicates that for these values of c one already has |f (0)| = |c| > 2,
while for the values of c in the neighbouring green area the orbit {f (k) (0)}
k=0 is leaving the disc B(0, 2) already for k = 2, and so on. The inner black area (not followed by
blue) indicates approximately the proper M ANDELBROT set, namely the set of points
for which the partial orbit {f (k) (0)}100
k=0 still lies in the disc B(0, 2).

Figure 3.3. An escape time colour chart for the M ANDELBROT set M as described in the text.

3.3.5

The escape time approximation for J ULIA sets

Essentially the same procedure may be used for an approximate representation of a

J ULIA set J . In contrast to the former situation, however, now c and with it the function
f = fc is fixed and we test every available z as to whether its orbit diverges. Since
J = A(f, ) (Theorem 3.1.30), the points of the invariant and compact set J are
characterized by the fact that there are neighbouring points whose orbits escape to .
According to Lemma 3.1.10 this will certainly be the case as soon as the orbit leaves the
closed disc with center 0 and radius 2 +|c|, so this value may serve as a safe escape time
bound in place of 2 for the M ANDELBROT set. There is, however, a practical difficulty:
If c belongs to the M ANDELBROT set, then J is the connected boundary of A(f, )

128

Chapter 3 Iteration of complex polynomials

(Theorems 3.1.30, 3.3.2), and the pixels corresponding to all points z inside of J are
marked blue by our procedure. So we get the set J only as boundary of the complement
J of A(f, ) which is painted blue (and is sometimes also called the filled J ULIA
set). The situation is even less satisfying if c does not belong to the M ANDELBROT
set: As we know, the set J is nowhere dense (Theorem 3.1.27); as a consequence, the
chance for a pixel to represent a point of J is small, and since now all neighbouring
points of J have orbits diverging to , the pixels of points not extraordinarily close to
J will be marked white and a good part of J will remain invisible. A colour chart for
the escape time as above may help a little bit in this case.
As an example Figure 3.4 shows an approximation of the J ULIA set J for c =
0.3 + 0.3 i (sorry, not the blue area but only its boundary; c is marked by a grey
cross). J is connected and contains the repellent fixed point z2 0., 8565 + 0.4208 i
(marked by a red cross) and the repellent periodic orbit consisting of the two points
z3 0.3551 + 1.0349 i, z4 0.6449 1.0349 i (marked by green crosses). There
is an attractive fixed point z1 0.1435 0.4208 i with |f  (z1 )| 0.9. The blue area
approximates the attractive basin A(f, z1 ). The series of cyan crosses mark the orbit of
a point inside of J (0.87 + 0.43 i) but close to z2 which of course converges to z1 . Note
the five channels along which this convergence takes place; we shall come back to this
feature in Section 3.3.7.

Figure 3.4. The filled J ULIA set for c = 0.3 + 0.3 i. Marked by cyan crosses
the orbit converging to the attractive fixed point z1 0.1435 0.4208 i of
a point inside of J (0.87 + 0.43 i) and close to the repellent fixed point z2
0.8565 + 0.4208 i.

Figure 3.5 displays in case c = 0.4 + 0.3 i (marked by a grey cross) what the
computer is able to catch of the now disconnected J ULIA set J : evidently only some
dust marks. Two red crosses mark two repellent fixed points z1 0.1955 0.4927 i,
z2 0.8045 + 0.4927 i, two green crosses mark a repellent periodic orbit of period 2
consisting of the points z3 0.3613 + 1.0813 i, z4 0.6387 1.0813 i. The pixel
representing (only approximately) the fixed point z1 has been used for the initial point

Section 3.3 The M ANDELBROT set

129

Figure 3.5. The sparse blue points represent pixels which succeed in catching
points of the J ULIA set for c = 0.4 + 0.3 i. The light blue crosses mark the
first 160 points of the orbit of a point close to the repellent fixed point z1
0.1955 0.4927 i.

of an orbit which now diverges to . The five channels harboring this orbit furnish
a striking counterpart to the picture in Figure 3.4 where the attractive fixed point has
occupied approximately the same position as here the repellent fixed point z1 .
Figure 3.6 displays the escape time colour chart for the present J ULIA set J but
where is J to be found? The answer is: hardly in a recognizable set of pixels; the
nowhere dense set J escapes the too coarse grid of pixels. But its presence is indicated
in the accumulating sequence of consecutive colours indicating higher and higher escape times, as in the neighbourhood of the point z1 . There one even recognizes the five
channels mentioned above. Although now J is (even completely) disconnected and
forms the boundary of the attractive basin A(f, ) only as a compact and topologically perfect set of gaps in C, there seems to be a conspicuous relationship to the filled
J ULIA set in Figure 3.4, evidently caused by the closeness of the two corresponding
values of c on both sides of the boundary of M .
At this point one starts to wonder how the picture can develop from Figure 3.4 to
Figures 3.5 and 3.6, in particular, what happens if the parameter c, wandering from
c1 = 0.3 + 0.3 i to c2 = 0.4 + 0.3 i, crosses the boundary of the M ANDELBROT set.
Traveling on the straight line connecting these two points there certainly must happen
something at its intersection c 0.364852 + 0.3 i with the cycloid z = 14 (e2i 2ei )
(for 69.0512 , Figure 3.7).

130

Chapter 3 Iteration of complex polynomials

Figure 3.6. An escape time colour chart for the J ULIA set in Figure 3.5, corresponding to c = 0.4 + 0.3 i.

Figure 3.7. The filled J ULIA set for the parameter c 0.3649 + 0.3 i lying on the cycloid C0 (Theorem 3.2.2) for 69.0512 . The cyan crosses
mark the orbit of a point (0.85 + 0.48 i) close to the repellent fixed point
z2 0.8212 + 0.4670 i, visiting consecutively and clockwise limbs, tail
and trunk of the dragon and finally rotating around the neutral fixed point
z1 0.1788 0.4670 i.

Section 3.3 The M ANDELBROT set

131

Evidently the corresponding filled J ULIA set J is on the point of breaking up into
pieces, and surprisingly enough the pieces display our well-known feature of selfsimilarity all over; in fact, we seem to encounter a relative of the dragon in Figure 1.42.
According to Theorem 3.2.2 there is a neutral fixed point z1 0.1788 0.4670 i in
which f  (z1 ) e69.0512 i , and another one z2 0.8212 + 0.4670 i with |f  (z2 )|
1.8894. The periodic orbit of period 2 consists of the two points z3 0.3592 +
1.0652 i), z4 0.6408 1.0652 i. If we let the computer start an orbit {f (k) (z0 )}
k=0
(marked by light blue crosses) at the pixel representing z2 (it is unable to catch the
precise fixed point and belongs to some neighbouring point z0 at the tip of the tongue
of the dragon), then after some first steps, which hardly amount to any visible movement, the points f (k) (z0 ) jump to the side of the tongue (k = 11) and to the lower lip
(k = 12), from there to the right front leg on the trunk (k = 13), then consecutively to
the four right legs connected to the belly (14 k 17), to the tail (k = 18), and to
the heart in the trunk (k = 19). Each of these points assumes in its corresponding part
of J roughly the position which z1 assumes in the main lower body. From there on the
following 10000 points of the orbit circle the fixed point z1 roughly in vertices of pentagons (rotating about angles 69 ), filling up a closed curve about z1 without coming
close to z1 in Figure 3.7 the center line of the cyan belt produced by the cross-marks.
If we check this unexpected behavior by observing the orbit {f (k) (0)}10200
k=0 (Figure 3.8), the computer produces, after the first 10000 points, a string of points for the
eye a closed curve, again the center line in the cyan belt of cross-marks following the
outskirts of the lower main body of the filled J ULIA set without caring for the encircled fixed point z1 or even the smaller annulus in Figure 3.7. The action of f in this
region seems to be topologically equivalent to a rotation (cf. [Blanchard, 1984, 7]).

Figure 3.8. The filled J ULIA set as in Figure 3.7 with the orbit of the critical
point 0, the blue line in the center of the cyan belt.

The special features encountered in the last studied J ULIA set suggest that the
boundary of the M ANDELBROT set deserves special interest. We will focus on this
topic in the next section.

132

3.3.6

Chapter 3 Iteration of complex polynomials

Zooming into M

As soon as we magnify parts of the M ANDELBROT set we encounter the practical

difficulty mentioned above: we want to picture the set M = {c C : |fc(k) (0)|
2 k N}, but not being able to wait infinitely long we must content ourselves with a
picture of the larger approximating set M (r, n) := {c C : |fc(k) (0)| r k n}
M . We still have the choice of r and n, and as r 2 gets smaller and n gets larger,
the set M (r, n) shrinks towards M . Without trouble we can always take r = 2, but
magnifying smaller and smaller cuts at the boundary of M we meet more and more
points close to M with large escape time. If we want to reasonably distinguish details
we are bound to increase n (and thereby the effort and time of computing). Still, if
we simply want to enjoy aesthetically whatever pictures are provided for us by the
dynamics of the mapping fc , it seems perfectly legitimate also to care for pictures of
sets M (r, n) which avoid fuzzing too much with details. (The size of the figures in this
section is, in fact, a little bit smaller than indicated since a margin originally carrying
some data information has been eliminated.)
Let us have a closer look at the cut indicated by a red rectangle in Figure 3.9 and
depicted magnified in Figure 3.10. There appear armies of buds magnified again in
Figure 3.11 which ask for to be looked at closer distance.

Figure 3.9. The red rectangle marks the area shown in Figure 3.10.

Section 3.3 The M ANDELBROT set

Figure 3.10. 0.1 x 0.4, 0.575 y 0.95; r = 5, n = 100. The red

rectangle marks the area shown in Figure 3.11.

Figure 3.11. 0 x 0.1, 0.625 y 0.7; r = 10, n = 100. The red

rectangle marks the area shown in Figure 3.12.

133

134

Chapter 3 Iteration of complex polynomials

Figure 3.12. 0.044 x 0.06, 0.669 y 0.681; r = 10, n = 100.

Figure 3.13. Same figure as Figure 3.11, but now the red rectangle marks the
area shown in Figure 3.14.

Section 3.3 The M ANDELBROT set

135

Figure 3.14. 0.08 x 0.084, 0.648 y 0.651; r = 10, n = 200.

The hanging ones already appear rather fantastic (Figure 3.12), but how about
the standing ones indicated again by a red rectangle in Figure 3.13? With your
indulgence, my first impression of Figure 3.14 was: three cows are curious about the
spiral above their heads (we shall meet other members of this population in later figures
again). Here the spell is lifted as soon as we increase the escape time bound from 200 to
500 as done in Figure 3.15: the heads are dissolved, together with the other large blue
areas of Figure 3.14, into radiating flowers, the centers of which would still dissolve
further if we would further increase the escape time bound. Still, the structure of the
spiral seems to be interesting. Magnifying the red rectangle again produces a strip
resembling a crafty bracelet (the blue discs standing for gems, Figure 3.16). The middle
piece magnified again (Figure 3.17) furnishes an area suspiciously looking like the
original M ANDELBROT set, but now in a luxurious decoration (Figure 3.18) in which
spirals and rosettes abound (Figure 3.19). We stop zooming here since we have to stop
somewhere, but there seems to be no end to beauty in the abysses of this microcosm.
(This series of pictures has been inspired by similar ones in [Peitgen and Richter, 1986,
maps 3450] and [Peitgen, Jurgens, Saupe, 1992, 14.3].)

136

Chapter 3 Iteration of complex polynomials

Figure 3.15. Same rectangle as in Figure 3.14, but r = 10, n = 500. The red
rectangle marks the area shown in Figure 3.16.

Figure 3.16. 0.0812 x 0.0815, 0.649775 y 0.65; r = 2, n = 500.

The red rectangle marks the area shown in Figure 3.17.

Section 3.3 The M ANDELBROT set

Figure 3.17. 0.08137 x 0.08141, 0.649902 y 0.649932; r = 2,

n = 500. The red rectangle marks the area shown in Figure 3.18.

Figure 3.18. 0.081389 x 0.081398, 0.6499092 y 0.649916; r = 2,

n = 700. The red rectangle marks the area shown in Figure 3.19.

137

138

Chapter 3 Iteration of complex polynomials

Figure 3.19. 0.0813936 x 0.0813944, 0.6499115 y 0.6499121;

r = 2, n = 700.

The curiosity of the cows lined up at the frontier of the set M (2, 100) (Figures 3.20,
3.21) seems to be concentrated on strange growths on the opposite slope of the trench
between different buds as e.g. in Figures 3.22 and 3.243.26 (the last ones give an
impression as if coming from somewhere in East Asia). The strange objects start to
dissolve if the escape time bound n = 100 is increased to e.g. n = 200 (Figures 3.23
and 3.27), as has already happend with the bovine portraits in Figure 3.15.

Figure 3.20. The red rectangle marks the area shown in Figure 3.21.

Section 3.3 The M ANDELBROT set

Figure 3.21. 0.74 x 0.77, 0.0775 y 1.0000; r = 2, n = 100. The

red rectangle marks the area shown in Figure 3.22.

Figure 3.22. 0.757 x 0.764, 0.08095 y 0.08620; r = 2, n = 100.

139

140

Chapter 3 Iteration of complex polynomials

Figure 3.23. Same limits as in Figure 3.22, but r = 2, n = 200.

Figure 3.24. The red rectangle marks the area shown in Figure 3.25.

Section 3.3 The M ANDELBROT set

Figure 3.25. 1.2465 x 1.2650, 0.030125 y 0.044000; r = 2,

n = 100. The red rectangle marks the area shown in Figure 3.26.

Figure 3.26. 1.2565 x 1.261, 0.035625 y 0.039000; r = 10,

n = 100.

141

142

Chapter 3 Iteration of complex polynomials

Figure 3.27. Same limits as in Figure 3.26, but now r = 10, n = 200.

We resist the temptation to dwell further on fantastic pictures provided lavishly by

the boundary of M an extensive sample is displayed in [Peitgen and Richter, 1986,
maps 2654]. But let us still explore at least two regions at the outskirts of M up
north and in the east, where M ends at the point c = 2. The outer red region in Figure 3.3 informs us that the set M does not extend beyond Im (c) = 1.13, and since
the points of M actually depicted by pixels in this area seem to be sparse, we again
resort to an escape time colour chart of the region indicated by the red rectangle in Figure 3.28 (Figure 3.29). Again as in Figures 3.17 and 3.18 we meet a curious feature
of M : it seems to contain again and again smaller and smaller replicas of itself (commonly called secondary M ANDELBROT sets). Up north an example is demonstrated in
Figures 3.30 and 3.31, and in the far east (Figure 3.32) in Figure 3.33.

Section 3.3 The M ANDELBROT set

Figure 3.28. The red rectangle marks the area shown in Figure 3.29.

Figure 3.29. 0.1 x 0.4, 0.775 y 1.150; r = 2, n = 100. The red

rectangle marks the area shown in Figure 3.30.

143

144

Chapter 3 Iteration of complex polynomials

Figure 3.30. 0.18 x 0.21, 1.0875 y 1.1100; r = 2, n = 200. The

red rectangle marks the area shown in Figure 3.31.

Figure 3.31. 0.1973 x 0.2001, 1.0993 y 1.1014; r = 2, n = 400.

Section 3.3 The M ANDELBROT set

Figure 3.32. The red rectangle marks the area shown in Figure 3.33.

Figure 3.33. 1.70 x 2.05, 0.1325 y 1.3000; r = 2, n = 100.

145

146

3.3.7

Chapter 3 Iteration of complex polynomials

Regions with attractive periodic orbits

We already know that for c

/ M there are no attractive periodic orbits of f (Theorem 3.1.22). So one gets curious about attractive periodic orbits in the case c M .
There can be at most one by Theorem 3.1.20, and we have located the parameter values
c for which there is an attractive fixed point or an attractive orbit of period 2 (Theorems 3.2.2 and 3.2.3). It becomes harder to do so for larger periods, but there is an odd
way, avoiding any deep analysis, to obtain evidence about attractive periodic orbits for
a given c with the help of a computer: start with an initial point z0 inside of J (e.g.
z0 = 0) and let the computer hurry through the first, say, 10000 values of the orbit
{zk = f (k) (z0 )}
k=0 . By that time there is a good chance that this orbit will be so close
to any attractive periodic orbit that the next values will practically coincide with it and
signal a coincidence of the points z10000+k = z10000+k+p after p further applications of
f.
Let us try out this procedure first for real values of c and plot the (real) coordinates
of the periodic points obtained in this way against the abscissas c on the real line. From
Theorems 3.2.2 and 3.2.3 we already know that for 0.25 < c < 0.75 we should get a
single value for an attractive fixed point; for 0.75 < c < 1.25 we should jump to two
values as members of an attractive periodic orbit of period 2, but what happens if we
continue to send the value c further up to the last sensible value c = 2?
The picture produced by the computer (Figure 3.34) is somewhat surprising: as c
increases, the number of periodic points is doubled again and again up to a certain
point; there seemingly some chaos starts, but still a certain structure prevails. The
phenomenon of this bifurcation process connected with period-doubling has been
studied already by V ERHULST (see [Verhulst, 1845], [Verhulst, 1847]) in the context
of population dynamics (see also [May, 1976]). An attentive observer may recognize
self-similarities in the produced graph which, due to the insight into the relevance
of this phenomenon for dynamics revealed by [Feigenbaum, 1978], is also called a
F EIGENBAUM diagram.
We still have no information about attractive periodic orbits with e.g. period 3.
We can try the procedure having led up to Theorem 3.2.3: find roots of the equation
f (3) (z) = z .
f (3) (z) =

(z 2 c)2 c

2

= z 8 4cz 6 + 2c(3c 1)z 4 4c2 (c 1)z 2 + c4 2c3 + c2 c

= z.

The equation f (3) (z) z = 0 is certainly satisfied by every fixed point, so the left side
must be divisible by (z 2 z c). Indeed we have
z 8 4cz 6 + 2c (3c 1)z 4 4c2 (c 1)z 2 z + c (c3 2c2 + c 1)

= (z 2 z c) z 6 + z 5 + (1 3c)z 4 + (1 2c)z 3 + (1 3c + 3c2 )z 2

+ (1 2c + c2 )z + 1 c + 2c2 c3 .
(3.7)

147

Section 3.3 The M ANDELBROT set

Figure 3.34. For values of c [ 14 , 2] on the real axis there are real nonrepellent periodic orbits {xk = f (k) (x0 )}p1
k=0 ; the graph in brown displays the
corresponding values xk (0 k < p) (measured with the brown line segment
01 as unit) as ordinates.

If w1 is a root of the polynomial (3.7), then so are w2 = f (w1 ) and w3 = f (w2 ), and
{w1 , w2 , w3 } must be an orbit of period 3. But there must be six roots, and since in
general the polynomial (3.7) will not be the square of a cubic polynomial, there must
be another orbit {w4 , w5 = f (w4 ), w6 = f (w5 )} of period 3. The solution of this, at a
first glance maybe puzzling, situation is that in general there are two different periodic
orbits with period 3 belonging to the J ULIA set corresponding to the parameter c, at
most one of them being attractive. Since by Theorem 3.1.4,


f (3) (w1 ) = 8w1 w2 w3 ,

and

6
!

f (3) (w4 ) = 8w4 w5 w6 ,

wk = 1 c + 2c2 c3 ,

k=1

this is certainly the case if

1 c + 2c2 c3 = 0.

(3.8)

In this case the attractive orbit is even super-attractive since it must contain the point 0;
consequently it has the form {0, c, c(c 1)}. The three solutions of (3.8) are
c1,2 0.12256117 0.74486177 i,

c3 1.75487767.

The J ULIA set corresponding to c3 is not spectacular (Figure 3.35). The conjugate
complex parameters c1 and c2 lie in the big buds at the top and the bottom of the M AN DELBROT set. The corresponding J ULIA sets (Figure 3.36) for c1 also shows an orbit

148

Chapter 3 Iteration of complex polynomials

Figure 3.35. The escape time colour chart for the J ULIA set corresponding to
c = 1.75487767 for r = 2, n = 200. The super-attractive orbit of period 3
consists of the points 0, c and c(c 1) 1.32471797.

Figure 3.36. The escape time approximation of D OUADYs rabbit, i.e. for the
filled J ULIA set corresponding to c 0.12256117 + 0.74486177 i (marked
by a grey cross) for r = 2, n = 200. The super-attractive orbit of period
3 consists of the points 0, c, and c(c 1) 0.66236 0.56228 i. The
orbit marked by light cyan crosses starts close to the repellent fixed point z1
0.2763 0.4797 i. The second repellent fixed point (marked by a red cross) is
z2 1.2763 + 0.4797 i, the periodic orbit with period 2 consists of the points
z3 0.0838 + 0.8948 i and z4 0.9162 0.8948 i (marked by green
crosses).

Section 3.3 The M ANDELBROT set

149

converging to the super-attractive orbit with period 3) are called D OUADYs rabbit. A
continuous change of the parameter c should not change much of the situation, so there
ought to be an open set, containing c1 , of parameter values c which admit an attractive periodic orbit with period 3, and the bud of M containing c1 seems to be a good
candidate for it.
It would be tiresome to pursue in general the same procedure in the search for
attractive periods of period k N. There is a heuristic reasoning shedding some light
on the location of the corresponding values of c: Suppose kl is a reduced fraction, i.e.
(l, k) Z N and l and k are relatively prime; if l,k = 2kl and cl,k = 14 (e2il,k
2li/k
2eil,k ), then Theorem 3.2.2 informs us that the point z1 = e 2 is a neutral fixed
point with f  (z1 ) = e2li/k . For kl = 13 this limiting case is illustrated in Figure 3.37.
The action of f in the neighbourhood of z1 consists approximately in a rotation about
the angle 2kl which tries hard to produce periodic orbits of period k . Traversing the
cycloid C0 at cl,k to the outside of it makes z1 repellent and hands attractiveness indeed
over to one of the orbits of period k , a situation already illustrated in Figure 3.36. In
fact it turns out that the points cl,k are the points at which buds are attached to C0 ,
for l = 1 decreasing in size as k , which harbor parameter values associated
with J ULIA sets admitting attractive periodic orbits with period k . An example for
k = 20 is depicted in Figure 3.38. There are more laws governing the army of buds
connected directly or via other buds to the interior of the cycloid C0 , but we leave the
mathematics behind this structure to [Blanchard, 1984] and the sources cited there, as
well as the assertions justifying the subsumption of the M ANDELBROT and J ULIA sets
under the family of fractals and identifying their H AUSDORFF dimension.

Figure 3.37. The escape time approximation of the filled J ULIA set corresponding to c1,3 = 14 (e4i/3 2e2i/3 ) 0.125 + 0.6495 i on the cycloid C0 ,
for r = 1.5, n = 1000. The point z1 = e2i/3 0.2500 0.4330 i is
a neutral fixed point, the point z2 1.2500 + 0.4330 i a repellent one (both
marked by red crosses). The periodic orbit with period 2 consists of the points
z3 0.1283 + 0.8736 i and z4 0.8717 0.8736 i marked by green
crosses.

150

Chapter 3 Iteration of complex polynomials

Figure 3.38. The escape time approximation of the filled J ULIA set corresponding to c = 0.2733 + 0.0074 i for r = 2, n = 200. The two fixed points
z1 0.4761 0.1545 i, z2 = 0.5239 + 0.1545 i, marked by red crosses, are
repellent, if also not very strongly: |f  (z1 )| 1.0010, |f  (z2 )| 1.0925. The
two points z3 0.4963 + 1.0116 i and z4 0.5037 1.0116 i, marked

by green crosses, constitute the periodic orbit with period 2 and |f (2) (z3 )| =

|f (2) (z4 )| 5.0933. There is an attractive periodic orbit with period 20,
marked by light cyan crosses, one point of which is 0.27066 0.01203 i.

In the illustrations of J ULIA sets to follow, as already done in Figures 3.363.38,

the fixed points will be marked by red crosses and the points of the orbit of period 2 by
green crosses, while the location of c will be marked by a grey cross.

3.4
3.4.1

Generation of J ULIA sets

The sets Jc(r, n)

We have already used the method of escape time and corresponding colour charts in
order to construct with the help of a computer an approximate picture of the J ULIA
set J = Jc for the function f = fc in (3.4). As done with the M ANDELBROT set in
the beginning of Section 3.3.4, in practice we are mostly bound to be content with a
larger set J(r, n) = Jc (r, n) = {z : |fc(k) (z)| r k n}. This set might contain
some point (or even an open set) where there is none in J , depicted e.g. in blue, but it
is a mathematically well defined, bona fide set closely related to the JULIA set J (since

J(r, n) J(r, n + 1) and for the filled J ULIA set J we have J = n=1 J(r, n)) and
sometimes aesthetically pleasing. So we may legitimately be interested in such sets,
examples of which are given in Figures 3.393.43. Again, a beautiful collection of
J ULIA sets is presented in [Peitgen and Richter, 1986].
There are at least two other ways to approximately depict J ULIA sets, suggested by
what we have found out in Section 3.1 and Section 3.2.

Section 3.4 Generation of J ULIA sets

Figure 3.39. The set J(2, 100) for c = 0.7454 + 0.1130 i. Fixed points are
z1 0.4993 0.0565 i (almost neutral) and z2 1.4993 + 0.0565 i. The
periodic orbit with period 2 consists of the points z3 0.2670 + 0.2426 i and
z4 0.7330 0.2426 i.

Figure 3.40. The set J(2, 100) for c = 0.1940 + 0.6557 i. Fixed points are
z1 0.2861 0.4171 i (almost neutral) and z2 1.2851 + 0.4171 i. The
periodic orbit with period 2 consists of the points z3 0.1103 + 0.8413 i and
z4 0.8897 0.8413 i.

151

152

Chapter 3 Iteration of complex polynomials

Figure 3.41. The set J(2, 100) for c = 1.3000+0.0500 i. Fixed points are z1
0.7452 0.0201 i and z2 1.7452 + 0.0201 i. The periodic orbit with period
2 consists of the points z3 1.2424 + 0.0337 i and z4 0.2424 + 0.0337 i.

Figure 3.42. The set J(2, 100) for c = 1.1500+0.2500 i. Fixed points are z1
0.6879 0.1052 i and z2 1.6879 + 0.1052 i. The periodic orbit with period
2 consists of the points z3 1.1602 + 0.1893 i and z4 01.1602 + 0.1893 i.

Section 3.4 Generation of J ULIA sets

153

Figure 3.43. The set J(2, 100) for c = 0.1103 + 0.6300 i. Fixed points are
z1 0.1265 0.5028 i and z2 1.1265 + 0.5028 i. The periodic orbit with
period 2 consists of the points z3 0.8209 0.9815 i and z4 0.1791 +
0.9815 i.

3.4.2

Pre-images of repellent periodic orbits

Theorem 3.2.2 furnishes at least one repellent fixed point z2 for f which certainly
belongs to J(f ). Since J(f ) is forward and backward invariant (Theorem 3.1.13), it
also contains all inverse image sets f (k) ({z2 }) containing 2k points which again can
be visualized on the screen of a computer. In fact, everymember of a repellent periodic
(k)
({z2 }) fills up J(f )
orbit would do as well. The good news is that the set
k=0 f
densely from the inside by Theorem 3.1.26, but the bad news is that this may happen
in a rather irregular way, preferring certain parts of J(f ) (where then a whole bunch of
points of J(f ) are represented by a single pixel), and for a long time rather neglecting
other parts of J(f ).
For parameter values c M this procedure has the advantage to show the proper
J ULIA set, not only as the boundary of the set inside of it at least as long as the last
mentioned drawback remains bearable (Figure 3.44).
Still, also in this case, an escape time picture may appear more satisfactory than
the backward orbit of a fixed point which can be deceptive if parts of J are visited by
pre-images too sparsely (Figures 3.453.48).

154

Chapter 3 Iteration of complex polynomials

Figure 3.44. The set f (14) ({z2 }) for c = 0.4 + 0.3 i. Fixed points are z1
0.3264 0.1815 i (attractive) and z2 1.3264 + 0.1815 i (repellent). The
periodic orbit with period 2 consists of the points z3 0.2644 + 0.6368 i and
z4 0.7356 0.6368 i.

Figure 3.45. The set f (17) ({z2 }) for c = 0.1100 + 0.6557 i. Fixed points
are z1 0.2443 0.4405 i (almost neutral) and z2 1.2443 + 0.4405 i
(repellent). The periodic orbit with period 2 consists of the points z3
0.1283 + 0.8821 i and z4 0.8717 0.8821 i.

Section 3.4 Generation of J ULIA sets

Figure 3.46. The set J(2, 200) for c = 0.11000 + 0.6557 i as in Figure 3.45.

Figure 3.47. The set f (18) ({z1 }) for c = 0.3000 + 0.5500 i. Fixed points
are z1 0.0011 0.5488 i and z2 1.0011 + 0.5488 i (both repellent). The
periodic orbit with period 2 consists of the points z3 0.7601 1.072 i and
z4 0.2399 + 1.0572 i.

155

156

Chapter 3 Iteration of complex polynomials

Figure 3.48. The set J(2, 200) for c = 0.3000 + 0.5500 i as in Figure 3.47.

For parameter values c

/ M the graph of the backward orbit of a fixed point may
give a more informative picture than the escape time (Figures 3.49, 3.50) or at least
help to interpret better an escape time colour chart (Figures 3.51, 3.5, 3.6); it may
even appear more satisfactory than an escape time picture (Figures 3.52, 2.5). So this
procedure theoretically perfect with all of its shortcomings in practice appears
certainly to be useful.

Figure 3.49. The set f (20) ({z2 }) for c = 0.8 i. This parameter value does not
belong to M as demonstrated in Figure 3.50. Fixed points are z1 0.2376
0.5423 i and z2 1.2376 + 0.5423 i. The (repellent) periodic orbit with period
2 consists of the points z3 0.9163 0.9609 i and z4 0.0837 + 0.9609 i.

Section 3.4 Generation of J ULIA sets

Figure 3.50. An escape time colour chart of the subset of the M ANDELBROT
set for 0.009 x 0.009, 0.796 y 0.809, r = 2, n = 200. The
point c = 0.8 i (Figure 3.49) belongs to the green escape-area in the center of
(18)
the rectangle; in fact one has |fc (0)| > 2.

Figure 3.51. The set f (20) ({z1 }) for c = 0.4000 + 0.3000 i as in Figures 3.5
and 3.6.

157

158

Chapter 3 Iteration of complex polynomials

Figure 3.52. The set f (17) ({z1 }) for c = 0.1981 + 1.1002 i in the center of the
secondary M ANDELBROT set in Figure 3.31. Fixed points are z1 0.4044
0.6082 i and z2 1.4044 + 0.6082 i (both repellent). The periodic orbit with
period 2 consists of the points z3 1.0827 0.9441 i and z4 0.0827 +
0.9441 i.

Figure 3.53. The set J(2, 13) for c = 0.1981 + 1.1002 i as in Figure 3.52.

Section 3.4 Generation of J ULIA sets

3.4.3

159

The J ULIA set as attractor

Lemma 3.1.10 also asserts that the exterior E of the closed disc B(0, r0 ) is mapped
by f into the exterior of B(0, 2r0 ). As a consequence we have f (B(0, r0 )) B(0, r0 )
and f (1) (B(0, r0 )) B(0, r0 ). In other words, on B(0, r0 )the inverse map f (1)

(k)
behaves like a contraction and
(B(0, r0 )) = C \ k=1 f (k) (E) = C \
k=1 f
A(f, ) (in fact, every disc containing J in its interior would do as well). Therefore,
constructing the intersection produces a set with boundary J(f ). This construction
is still supported e.g. if, in case there exists an attractive fixed point z1 , a sufficiently
small disc B(z1 , r1 ) with center z1 , lying inside of the connected set J , is deleted from
B(0, r0 ) (Figure 3.54): since f works as a contraction in A(f, z1 ) with J = A(f, z1 ),
the inverse map f (1) contracts the set B(0, r0 ) \ B(z1 , r1 ) eventually to J .

Figure 3.54. The initial set F , the complement of B(z1 , 0.25) in B(0, 1.5), for
the contracting inverse map fc (1) for c = 0.7 + 0.2 i (marked by a grey cross)
and the attractive fixed point z1 0.4800 0.1020 i.

This construction at the same time gives a supplementary answer to the question
why a J ULIA set should be considered as a fractal. It is the attractor of a sequence of
sets formed by applying the iterates of the map f (1) to a set F J . This looks very
much like an iterativefunction system since f(1) (F ) is produced by two functions
defined by g1 (w) = w + c and g2 (w) = w + c (after having chosen a branch
of the complex square root) applied to F , and f as well as f (1) = g1 g2 furnish
conformal mappings. However,
the twomappings gj are not contractions in the sense
of Definition 2.2.1 since e.g. 0.25 0.16 = 0.5 0.4 = 0.1 > 0.25 0.16. As
a practical consequence, in the neighbourhood of c the images of w under g1 and g2
expand and in the screen graphics there appear lines of pixels not carrying images of
points of the set F (Figures 3.55 and 3.61), which lines are proliferated through the
further applications of f (1) (Figures 3.56, 3.57, 3.62).

160

Chapter 3 Iteration of complex polynomials

Figure 3.55. The set f (1) (F ) for c as in Figure 3.54. The functions g1 and g2
constituting the inverse map f (1) are chosen so as to provide the square roots

w + c with non-negative resp. non-positive real part. The white lines signal
the lack of image-pixels as f (1) is applied to the neighbourhood of w = c.

Figure 3.56. The set f (3) (F ) for c as in Figure 3.54.

Figure 3.57. The set f (8) (F ) for c as in Figure 3.54.

Section 3.4 Generation of J ULIA sets

Figure 3.58. The set f (47) (F ) for c as in Figure 3.54. The repellent fixed point
is z2 1.4800 + 0.1020 i, the periodic orbit of period 2 consists of the points
z3 0.7794 0.3579 i, z4 0.2206 + 0.3579 i. The circles delimit the
original set F .

Figure 3.59. The set f (21) ({z2 }) for c as in Figures 3.543.58 and z2 as in
Figure 3.58.

161

162

Chapter 3 Iteration of complex polynomials

Figure 3.60. The set J(2, 20) for c as in Figures 3.543.59.

Figure 3.61. The set f (1) (F ) for c = 0.8 + 0.3 i

/ M and F = B(0, 1.6)
with red circumference. The repellent fixed points are z1 0.53490.1449 i,
z2 1.5349 + 0.1449 i. The periodic orbit of period 2 consists of the points
z3 0.9208 0.3565 i, z4 0.0792 + 0.3565 i.

Section 3.4 Generation of J ULIA sets

163

Figure 3.62. The set f (5) (F ) for c and F as in Figure 3.61.

A comparison of the various methods of approximately computer-picturing a J ULIA

set may be carried out on the basis of Figures 3.583.60 in case of c M and of
Figures 3.63 and 3.64 in case of c
/ M.

Figure 3.63. The set f (30) (F ) for c and F as in Figure 3.61.

164

Chapter 3 Iteration of complex polynomials

Figure 3.64. An escape time colour chart for the set J(2, 200) and c as in
Figures 3.613.63.

Good luck now for playing with fractals, and for learning more about them in the
references cited below!

Bibliography
The pages on which the author is cited are indicated in italics. Much more literature on fractals is listed in the references below, in particular in [Falconer, 1990],
[Mandelbrot, 1982], [Peitgen, Jurgens, Saupe, 1992].

Literature on fractals
[Addison, 1997] P. S. Addison. Fractals and Chaos: An Illustrated Course. Institute of Physics
Publishing, Bristol, 1997.
[Barnsley, 1988] M. Barnsley. Fractals Everywhere. Academic Press, Inc., Toronto, 1988. (v, 63,
85, 97, 98, 105)
[Becker and Dorfler, 1989] K.-H. Becker and M. Dorfler. Dynamical Systems and Fractals. Computer Graphics Experiments in Pascal. Cambridge University Press, Cambridge, 1989. German
Edition: Computergraphische Experimente in Pascal. Vieweg, Braunschweig 1986.
[Blanchard, 1984] P. K. Blanchard. Complex analytic dynamics on the Riemann sphere. Bull. Am.
Math. Soc., 11:85141, 1984. (111, 117, 121, 131, 149)

[Cantor, 1883] G. Cantor. Uber

unendliche, lineare Punktmannigfaltigkeiten, V. Math. Ann.,
21:545591, 1883. (2)
[Collet and Eckmann, 1980] P. Collet and J.-P. Eckmann. Iterated Maps of the Interval as Dynamical Systems. Birkhauser, Basel, 1980.
[Davis and Knuth, 1970] C. Davis and D. E. Knuth. Number representations and dragon curves.
Journal of Recreational Mathematics, 3:6681, 133149, 1970. (27)
[Douady and Hubbard, 1982] A. Douady and J. H. Hubbard. Iteration des polynomes quadratiques
complexes. C. R. Acad. Sci., Paris, Ser. I, 294:123126, 1982. (126)
[Douady, 1983] A. Douady. Syst`emes dynamiques holomorphes. Seminaire Bourbaki, 35e annee,
1980/81, Exp. 599, Asterisque, 105106:3963, 1983. (117, 148, 149)
[Falconer, 1985] K. J. Falconer. The Geometry of Fractal Sets. Cambridge University Press, Cambridge, 1985.
[Falconer, 1990] K. J. Falconer. Fractal Geometry. Mathematical Foundations and Applications.
Wiley, New York, 1990. German Edition: Fraktale Geometrie. Spektrum Akademischer Verlag,
Heidelberg, 1993. (v, 51, 60, 111)
[Feder, 1988] P. Feder. Fractals. Plenum Press, New York, 1988.
[Frame and Mandelbrot, 2002] M. L. Frame and B. B. Mandelbrot. Fractals, Graphics, and Mathematics Education. The Mathematical Association of America, Washington, DC, 2002.
[Gardner, 1967] M. Gardner. An array of problems that can be solved with elementary mathematical
techniques. Sci. Amer., 216:3:124129, 4:116123, 6, 1967. (27)
[Gardner, 1977] M. Gardner. Mathematical Magic Show. Knopf, New York, 1977. (27)
[Gleick, 1987] J. Gleick. Chaos. Making a New Science. Penguin Books, New York, 1987.

[Hilbert, 1891] D. Hilbert. Uber

die stetige Abbildung einer Linie auf ein Flachenstuck. Math.
Ann., 38:459460, 1891. (30)

166

Bibliography

[Hoveijn and Scholtmeijer, 2001] I. Hoveijn and J. Scholtmeijer. Fractals. Epsilon Uitgaven,
Utrecht, 2001.
[Hutchinson, 1981] J. Hutchinson. Fractals and self-similarity. Indiana Univ. Math. J., 30:713747,
1981.
[Julia, 1918] G. Julia. Memoire sur literation des fonctions rationnelles. J. de Math. Pures et
Appliquees, 8:47245, 1918. (112)
[Lebesgue, 1905] H. Lebesgue. Sur les fonctions representables analytiquement. Journ. de Math.
(6), 1:139216, 1905. (39)
[Levy, 1938] P. Levy. Les courbes planes ou gauches et les surfaces composees de parties sem
blables au tout. J. de lEcole
Polytechnique III, 78:227247, 249291, 1938. (43)
[Mandelbrot, 1977] B. B. Mandelbrot. Fractals: Form, Chance, and Dimension. W. H. Freeman
and Co., New York, 1977.
[Mandelbrot, 1980] B. B. Mandelbrot. Fractal aspects of the iteration z z(1 z) for complex
and z. Ann. New York Acad. Sci., 357:249259, 1980. (123)
[Mandelbrot, 1982] B. B. Mandelbrot. The Fractal Geometry of Nature. W. H. Freeman and Co.,
New York, 1982. German Edidion: Die fraktale Geometrie der Natur. Birkhauser, Basel, 1991.
(9, 11, 74, 123)
[Menger, 1926] K. Menger. Allgemeine Raume und charakteristische Raume, Zweite Mitteilung:

Uber
umfassenste n-dimensionale Mengen. Proc. Acad. Amsterdam, 29:11251128, 1926. (50)
[Peano, 1890] G. Peano. Sur une courbe, qui remplit toute une aire plane. Math. Ann., 36:157160,
1890. (30)
[Peitgen and Jurgens, 1990] H.-O. Peitgen and H. Jurgens. Fraktale: Gezahmtes Chaos. Carl
Friedrich von Siemens Stiftung, Munchen, 1990.
[Peitgen and Richter, 1986] H.-O. Peitgen and P. H. Richter. The Beauty of Fractals. Springer,
Berlin, 1986. (135, 142, 150)
[Peitgen, Jurgens, Saupe, 1992] H.-O. Peitgen, H. Jurgens, and D. Saupe. Fractals for the Classroom. Part 1: Introduction to Fractals and Chaos. Part 2: Complex Systems and Mandelbrot
Set. Springer, Berlin, 1992. German Edition of Part 1: Bausteine des Chaos. Fraktale. RoRoRo,
Hamburg, 1998. (85, 86, 90, 93, 95, 96, 98, 135)
[Schroder, 1991] M. Schroder. Fractals, Chaos, Power Laws. W. H. Freeman and Co., New York,
1991.
[Sierpinski, 1915] W. Sierpinski. Sur une courbe cantorienne dont tout point est un point de ramification. C.R. Acad. Paris, 160:302, 1915. (41)
[Sierpinski, 1916] W. Sierpinski. Sur une courbe cantorienne qui contient une image biunivoque et
continue de toute courbe donnee. C.R. Acad. Paris, 162:629632, 1916. (41)
[Vicsek, 1989] T. Vicsek. Fractal Growth Phenomena. World Scientific, London, 1989.
[von Koch, 1904] H. von Koch. Sur une courbe continue sans tangente, obtenue par une construction geometrique e lementaire. Arkiv for Matematik, 1:681704, 1904. (6)
[Zeitler and Pagon, 2000] H. Zeitler and D. Pagon. Fraktale Geometrie. Eine Einfuhrung. Vieweg,
Braunschweig, 2000.

Bibliography

167

Literature on general topics

[Cartan, 1966] H. Cartan. Elementare Theorie der analytischen Funktionen einer oder mehrerer
komplexen Veranderlichen. Hochschultaschenbucher Bibl. Inst., Mannheim, 1966. (111)
[Evans and Gariepy, 1992] L. C. Evans and R. F. Gariepy. Measure Theory and Fine Properties of
Functions. CRC Press, New York, 1992. (56, 57)
[Feigenbaum, 1978] M. J. Feigenbaum. Quantitative universality for a class of nonlinear transformations. J. Stat. Phys., 19:2552, 1978. (146)
[Kelley, 1955] J. L. Kelley. General Topology. D. Van Nostrand Company, New York, 1955. (3)
[May, 1976] R. M. May. Simple mathematical models with very complicated dynamics. Nature,
261:459467, 1976. (146)
[Montel, 1927] P. Montel. Lecons sur les familles normales. Gauthier-Villars, Paris, 1927. (113)
[Nevanlinna and Paatero, 1965] R. Nevanlinna and V. Paatero. Einfuhrung in die Funktionentheorie. Birkhauser, Basel, 1965. (111)
[Saks and Zygmund, 1971] S. Saks and A. Zygmund. Analytic Functions. Elsevier Publ. Comp.,
Amsterdam, 3rd edition, 1971. (113)
[Verhulst, 1845] P. F. Verhulst. Recherches mathematiques sur la loi daccroissement de la population. Nouveaux memoires de lAcademie Royale des Sciences et Belles-Lettres de Bruxelles,
18:138, 1845. (146)
[Verhulst, 1847] P. F. Verhulst. Deuxi`eme memoire sur la loi daccroissement de la population.
Memoires de lAcademie Royale des Sciences, des Lettres et des Beaux-Arts de Belgique, 20:1
32, 1847. (146)

List of symbols
The symbols appear for the first time on the indicated page.
N
Z
Q4
R
R+
Rn
C
C
Re
Im
||
o(g)
O(g)
f g(z)
f (k) (z)
A(k)
A
A
L, L1
L2
L3
dimS
dimB
dimB
dimB
dimH
H1
Hs
Hs
(s)

diam
N
K(X)
d(a, b)
d(a, B)
d(A, B)
h(A, B)
B{}

natural numbers
integers
4-adic rational numbers in [ 0, 1 ]
real line
non-negative reals
n-dimensional E UCLIDEAN space
complex plane
R IEMANN sphere
real part
imaginary part
absolute value, E UCLIDEAN norm
lim |o(g)(x)|
|g(x)| = 0
lim sup |O(g)(x)|
|g(x)| <
f (g(z))
f (f (k1) (z))
f (k) (A)
closure of A
boundary of A
L EBESGUE measure in R
L EBESGUE measure in R2
L EBESGUE measure in R3
self-similarity dimension
upper box counting dimension
lower box counting dimension
box-counting dimension
H AUSDORFF dimension
one-dimensional H AUSDORFF measure
-approximation of Hs
s-dimensional H AUSDORFF measure
Hs of s-dimensional unit ball
gamma function
diameter
minimal cardinality of a -covering
space of compact subsets of X
distance between a and b
distance of a from B
distance of A from B
H AUSDORFF distance
closed -hull of B

3
51
47
1
51
4
109
110
123
142
4
7
7
121
1
1
54
100
2
45
50
8
51
51
51
58
37
56
56
55
55
37
50
63
64
64
64
64
65

170

dT (t, s)
IFS
F
L
Lt
|L|
det
O
J(f ) = Jc (f )
F (f )
J0 (f )
F0 (f )
J
J(r, n) = Jc (r, n)
A(f, z0 )
C(f, z0 )
M
M (r, n)
B(z, )
U (z, )

List of symbols

distance in parameter space T

iterative
function system
n
f
j=1 j
(matrix of) linear mapping in Rn
transpose of L
determinant of L
determinant
orthogonal mapping
J ULIA set
FATOU set
{f (k) }k not normal on J0
{f (k) }k normal on F0
filled J ULIA set
escape time approximation of Jc
attractive basin of fixed point z0
connected component of F0 containing z0
M ANDELBROT set
escape time approximation of M
closed -disc about z
open -disc about z
end of proof
logical and
logical there exists
logical for all
set theoretic substraction

73
71
72
74
75
103
103
75
112
112
113
113
128
150
115
116
123
132
111
111
7
65
65
56
47

Index
A
-adic
dy-, 3
tri-, 2, 32
4-, 6, 7, 47, 48
9-, 34
affine map, 7476, 78, 83, 85, 86,
103, 104
attractive basin, 115, 120, 125, 128,
129
attractor, 73, 78, 86, 90, 105, 109,
159

B
BARNSLEY fern, 74, 98102
bifurcation, 146
blossom, 103, 104
boundary, 41, 46, 89, 91, 100, 110,
112, 122, 124126, 128, 153,
159
of the unit square, 48, 78, 83
87, 89, 9597, 99, 106, 107
of A(f, ) (= J(f )), 120, 127,
128, 129
of A(f, z1 ), 120
of M , 129, 131, 132, 142
buds, 132, 138, 147, 149

C
C ANTOR
dust, 4648, 54, 76
set, 1, 2, 4, 810, 35, 41, 43, 53
set, thick, 35
staircase, 3841
C AUCHY sequence, see fundamental sequence
cell, 2426, 29
collage, 71, 86
mappings, 71, 76
theorem, 73, 100
colour, 5, 34, 50, 77, 126, 127

chart, 127130, 142, 148, 150,

156, 157, 164
complete metric space, 63, 70, 71,
73
completely disconnected, 3, 9, 35,
129
connected, 26, 29, 90, 111, 116,
125128, 159
component, 116
simply, 111, 116, 124, 125
contraction, 7075, 86, 98, 100,
159
constant, 7073, 75, 76, 98, 100,
104
crab, 24, 26, 27
critical value, 116, 117, 119, 121,
123, 132
cross, 23, 34, 52, 7882, 128131,
148150, 159
cycloid, 122, 123, 129, 130, 149

D
decoration, 83, 84, 135
devils staircase, see C ANTOR staircase
diameter, 36, 37, 46, 50, 51, 53, 55,
57, 58, 60
dimension, 1, 810, 12, 13, 1518,
35, 41, 43, 49, 50, 5355, 60,
61, 78, 85, 8991
box-counting, 50, 51, 5355,
58, 60, 83
H AUSDORFF, 13, 38, 55, 58, 60,
77, 149
lower box-counting, 51
self-similarity, 8, 9, 53
topological, 9, 10
upper box-counting, 51
distance, 6, 24, 25, 33, 39, 64-66,
75, 109, 126
E UCLIDEAN, 4
H AUSDORFF, 64, 65, 67, 69,

172
100
of a from B , 64
of A from B , 64
D OUADYs rabbit, 148, 149
dragon, 10, 2730, 130, 131

E
-hull, 65
escape time, 126130, 132, 135,
138, 142, 148150, 153, 156,
157, 164
essentially disjoint, 24, 25
E UCLIDEAN
distance, 4
norm, 4
expansion, 6, 34, 47

F
FATOU set, 112
F EIGENBAUM diagram, 146
filled J ULIA set, see J ULIA set
finite length, 35
fixed point, 70,71, 73, 75, 76, 86,
98, 100, 104, 109, 112 114,
115, 118123, 128, 131, 146,
150154, 156
attractive, 109112, 115, 120
122, 128, 129, 146, 159
neutral, 111, 122, 130, 131, 149
repellent, 109, 110, 111, 115,
116, 122, 128130, 148, 155,
158, 161, 162
super-attractive, 111, 148
flower, 19, 8689, 135
fractal, 1, 3, 8, and frequently thereafter
fundamental sequence, 6, 6670

Index

H
half square, 22, 23, 30, 44
H AUSDORFF
dimension, 13, 38, 55, 58, 60,
77, 149
distance, 64, 65, 67, 69, 100
measure, 37, 46, 5558
metric, 65, 70, 71, 73, 89, 100
H EIGHWAY-H ARTER dragon, 10,
2730, 130, 131

I
IFS, see iterative function system
infinite length, 8, 11, 35
initiator, 10, 11, 15, 18, 22, 27, 30,
35, 48, 61, 63, 70, 78
inverse function, 111, 116
iterative function system, 63, 71
74, 77, 78, 83, 85, 89, 95, 97,
98, 100, 105, 159

J
J ULIA set, 111, 112, 121, 122, 124,
125, 127131, 147150, 153,
159, 163
filled, 128131, 148150

K
KOCH
curve, 46, 811, 30, 35, 43, 45,
53, 76
curve, modified, 1115, 18, 53,
61
island, 10, 11, 45
pyramid, see L E VY surface

L
G
gamma function, 55
generator, 4, 1013, 1522, 24, 26,
27, 2931, 35, 36, 48, 53, 61,
63, 70, 78

lattice, 51, 52, 83, 85, 89

leaf, 97, 98
L EBESGUE
measure, 2, 35, 36, 37, 39, 40,
41, 45, 46, 49, 50, 55

173

Index

singular function, 39
length, 8
finite, 35
infinite, 8, 11, 35
level, 1, 2
L E VY surface, 43
L IPSCHITZ function, 54, 59
loop, 19, 20, 27, 53

M
M ANDELBROT set, 123129, 131,
132, 135, 147, 149, 150, 157,
158
secondary, 142
M ENGER sponge, 50, 54, 109
mesh, 23, 24, 27
monodromy theorem, 111, 116
monotony, 54, 58
M ONTELs theorem, 113, 116, 118,
120
moving grass, 105

N
normal sequence, 112118, 120
in U , 112
in w, 112
nowhere
dense, 3, 35, 112, 119, 120, 128,
129
differentiable, 6

P
parameter, 8, 26, 55, 73, 74, 89, 91,
92, 123, 124, 129, 130, 146,
147, 149, 153, 156
P EANO curve, 30, 31, 34
pentagon, 8991, 104, 131
perfect, 3, 35, 112, 119, 120, 129
period, 109, 110, 112, 114, 115,
121123, 128, 131, 146156,
158, 161, 162,
periodic orbit, 112, 114117, 120
123, 128, 131, 146155, 158,
161, 162
attractive, 112, 114, 116, 117,
120, 123, 146, 147, 149, 150
neutral, 112, 117, 120
repellent, 112, 115, 128, 153,
156
super-attractive, 112, 147, 148,
149
periodic point 109, 110, 112, 119,
146
p-, 112
repellent, 112, 115, 117, 120

R
rational, 55
dyadic, 3
4-adic, 6, 7, 47
function, 121
R IEMANN sphere, 110
rotation, 19, 51, 56, 75, 131, 149

O
open mappings, 111, 114, 116
open set condition, 5, 9, 1113, 15
18, 22, 24, 27, 30, 36, 41, 43,
46, 49, 53, 60, 61, 78, 83, 85,
90, 91
orbit, 70, 78, 109112, 117, 125
129
backward, 153, 156
bounded, 120, 123, 128, 130,
131
periodic, see periodic orbit

S
self-similar, 3, 8, 9, 26, 35, 41, 43,
46, 49, 53, 60, 61, 76, 146
S IERPINSKI
carpet, 4850, 54, 76
triangle, 41, 42, 54, 76
similarity, 9, 11, 12, 35, 70, 7577,
83, 85, 89, 91, 98, 121
factor, 9, 1113, 43, 45, 49, 50,
53, 60, 61, 70, 7577, 83, 89
94, 109

174
simply connected, 111, 116, 124,
125
snowflake, 8991
space-filling, 22, 26, 29, 63
spiral, 19, 26, 29, 92, 135
stability, 54
countable, 59

T
tessellation, 34
tetrahedral fractal, see L E VY surface
theorem
of M ONTEL, 113, 116, 118, 120
on inverse functions, 111, 116,
120
on monodromy, 111, 116
on open mappings, 111, 116
on uniform convergence of
holomorphic functions, 111,
115
thick, see C ANTOR
totally bounded 65, 66, 68
tree, 96, 97
triplicate continent, 9194
twig, 95

U
uniform convergence of holomorphic functions, see theorem
unit ball, 55
unit disc, 122

Z
zero-dimensional, 3, 35

Index

Contents (detailed)
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
1

Fractals and dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 The game of deleting and replacing . . . . . . . . . . . . . . . . . . . . 1
1.1.1 The C ANTOR set . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 The KOCH curve . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2.1 Theorem . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.3 Heuristics of dimension . . . . . . . . . . . . . . . . . . . . . . 8
1.1.3.1 Definition (similarity) . . . . . . . . . . . . . . . . . 9
1.1.3.2 Definition (open set condition) . . . . . . . . . . . . 9
1.1.4 Initiators and generators . . . . . . . . . . . . . . . . . . . . . . 10
1.1.4.1 The KOCH island . . . . . . . . . . . . . . . . . . . . 10
1.1.4.2 A modified KOCH curve . . . . . . . . . . . . . . . . 11
1.1.4.3 A second type of modified KOCH curves . . . . . . . 15
1.1.4.4 More modified KOCH curves . . . . . . . . . . . . . 18
1.1.5 Space-filling curves . . . . . . . . . . . . . . . . . . . . . . . . 22
1.1.5.1 The half square . . . . . . . . . . . . . . . . . . . . . 22
1.1.5.2 The crab . . . . . . . . . . . . . . . . . . . . . . . . 24
1.1.5.3 The H EIGHWAY-H ARTER dragon . . . . . . . . . . . 27
1.1.5.4 The P EANO curve . . . . . . . . . . . . . . . . . . . 30
1.1.6 Short fractal curves . . . . . . . . . . . . . . . . . . . . . . . . 35
1.1.6.1 A thick C ANTOR set . . . . . . . . . . . . . . . . . . 35
1.1.6.2 A self-similar fractal curve with positive finite length 35
1.1.6.3 The C ANTOR staircase . . . . . . . . . . . . . . . . . 38
1.1.7 Higher dimensional C ANTOR sets . . . . . . . . . . . . . . . . 41
1.1.7.1 The S IERPINSKI triangle . . . . . . . . . . . . . . . . 41
1.1.7.2 The L E VY surface . . . . . . . . . . . . . . . . . . . 43
1.1.7.3 The C ANTOR dust . . . . . . . . . . . . . . . . . . . 46
1.1.7.4 The S IERPINSKI carpet . . . . . . . . . . . . . . . . 48
1.1.7.5 The M ENGER sponge . . . . . . . . . . . . . . . . . 50
1.2 The box-counting dimension . . . . . . . . . . . . . . . . . . . . . . . 50
1.2.1 Definition (box-counting dimension) . . . . . . . . . . . . . . . 51
1.2.21.2.4 Theorems . . . . . . . . . . . . . . . . . . . . . . . . 5154
1.3 The H AUSDORFF dimension . . . . . . . . . . . . . . . . . . . . . . . . 55
1.3.1 Definition (H AUSDORFF measure) . . . . . . . . . . . . . . . . 55
1.3.21.3.4 Theorems . . . . . . . . . . . . . . . . . . . . . . . . 5657
1.3.5 Definition (H AUSDORFF dimension) . . . . . . . . . . . . . . . 58
1.3.61.3.8 Theorems . . . . . . . . . . . . . . . . . . . . . . . . 5860

176

Contents (detailed)

Iterative function systems . . . . . . . . . . . . . . . . . . . . . . .

2.1 The space of compact subsets of a complete metric space . . . .
2.1.1 Definition (K(X)) . . . . . . . . . . . . . . . . . . . . .
2.1.2 Definition (d(a, B)) . . . . . . . . . . . . . . . . . . . .
2.1.3 Definition (d(A, B)) . . . . . . . . . . . . . . . . . . . .
2.1.4 Definition (h(A, B)) . . . . . . . . . . . . . . . . . . . .
2.1.5 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6 Definition (B{} ) . . . . . . . . . . . . . . . . . . . . .
2.1.72.1.8 Theorems . . . . . . . . . . . . . . . . . . . . . .
2.1.9 Definition (totally bounded) . . . . . . . . . . . . . . . .
2.1.102.1.12 Theorems . . . . . . . . . . . . . . . . . . . . .
2.2 Contractions in a complete metric space . . . . . . . . . . . . .
2.2.1 Definition (contraction) . . . . . . . . . . . . . . . . . .
2.2.2 Definition (orbit, fixed point) . . . . . . . . . . . . . . .
2.2.32.2.5 Theorems . . . . . . . . . . . . . . . . . . . . . .
2.2.6 Definition (iterative function system, IFS) . . . . . . . .
2.2.72.2.12 Theorems . . . . . . . . . . . . . . . . . . . . . .
2.3 Affine iterative function systems in R2 . . . . . . . . . . . . . .
2.3.1 Definition (affine map) . . . . . . . . . . . . . . . . . .
2.3.22.3.3 Theorems . . . . . . . . . . . . . . . . . . . . . .
2.3.4 Some examples of fractals constructed by means of IFS
2.3.4.1 A cross . . . . . . . . . . . . . . . . . . . . .
2.3.4.2 A decoration . . . . . . . . . . . . . . . . . .
2.3.4.3 An antenna . . . . . . . . . . . . . . . . . . .
2.3.4.4 A flower garden . . . . . . . . . . . . . . . .
2.3.4.5 A pentagon snowflake . . . . . . . . . . . . .
2.3.4.6 Evolution of a triplicate continent . . . . . . .
2.3.4.7 A twig . . . . . . . . . . . . . . . . . . . . . .
2.3.4.8 A tree . . . . . . . . . . . . . . . . . . . . . .
2.3.4.9 A leaf . . . . . . . . . . . . . . . . . . . . . .
2.3.4.10 The BARNSLEY fern . . . . . . . . . . . . . .
2.3.5 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.6 Construction of a blossom fractal . . . . . . . . . . . . .
2.3.7 A moving grass . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

. 63
. 63
. 63
. 64
. 64
. 64
. 64
. 65
. 65
. 65
6567
. . 70
. . 70
. . 70
7071
. . 71
7173
. . 74
. . 74
. . 75
. . 77
. . 78
. . 83
. . 85
. . 86
. . 89
. . 91
. . 95
. . 96
. . 97
. . 98
. . 103
. . 103
. . 105

Iteration of complex polynomials . . . . . . . . . . . . . . . . . . . . . .

3.1 General theory of J ULIA sets . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Theorem (Mapping theorems from the theory of functions of a
complex variable) . . . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Definition (attractive, neutral, repellent fixed point) . . . . . .
3.1.3 Definition (periodic orbit) . . . . . . . . . . . . . . . . . . . .
3.1.4 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.5 Definition (attractive, neutral, repellent periodic orbit) . . . .
3.1.6 Definition (J ULIA and FATOU sets) . . . . . . . . . . . . . . .
3.1.7 Definition (normal sequence) . . . . . . . . . . . . . . . . . .

. 109
. 111
.
.
.
.
.
.
.

111
111
112
112
112
112
112

177

Contents (detailed)

3.2
3.3

3.4

3.1.8 Theorem (M ONTEL) . . . . . . . . . . . . . . . . . . . . .

3.1.9 Definition (J0 (f ), F0 (f )) . . . . . . . . . . . . . . . . . .
3.1.103.1.15 Theorems . . . . . . . . . . . . . . . . . . . . . .
3.1.16 Definition (attractive basin) . . . . . . . . . . . . . . . . .
3.1.17 Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.18 Definition (C(f, z)) . . . . . . . . . . . . . . . . . . . . .
3.1.193.1.30 Theorems . . . . . . . . . . . . . . . . . . . . . .
J ULIA sets for quadratic polynomials . . . . . . . . . . . . . . . .
3.2.13.2.5 Theorems . . . . . . . . . . . . . . . . . . . . . . .
3.2.6 Definition (M ANDELBROT set) . . . . . . . . . . . . . . .
The M ANDELBROT set . . . . . . . . . . . . . . . . . . . . . . . .
3.3.13.3.3 Theorems . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 The escape time approximation for the M ANDELBROT set
3.3.5 The escape time approximation for J ULIA sets . . . . . . .
3.3.6 Zooming into M . . . . . . . . . . . . . . . . . . . . . .
3.3.7 Regions with attractive periodic orbits . . . . . . . . . . .
Generation of J ULIA sets . . . . . . . . . . . . . . . . . . . . . . .
3.4.1 The sets Jc (r, n) . . . . . . . . . . . . . . . . . . . . . . .
3.4.2 Pre-images of repellent periodic orbits . . . . . . . . . . .
3.4.3 The J ULIA set as attractor . . . . . . . . . . . . . . . . . .

Bibliography
List of symbols

. . . 113
. . . 113
113115
. . . 115
. . . 115
. . . 116
116120
. . . 121
121123
. . . 123
. . . 124
124125
. . . 126
. . . 127
. . . 132
. . . 146
. . . 150
. . . 150
. . . 153
. . . 159

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171