University of Aberdeen

Department of Mathematical Sciences

MX4020 Project Report
Mr Cameron J Lawson
2009-2010
Sipervisor: Dr Stephen THeriault
Summary
In this project we dene what is a fractal and give examples of frac-
tals. In particular, we study the examples of the Cantor set and
Sierpinski gasket. We mention a few of the important features of
fractals, and give a detailed discussion of dimension and its relation
to fractals. We dene and discuss the theory behind conventional
notions of dimension, namely Euclidean dimension and topological
dimension, and apply our theory to some examples. For the study of
fractals, we dene and study non-conventional dimensions (or fractal
dimensions), which can take non-integer values. The fractal dimen-
sions we discuss are the Hausdor dimension, similarity dimension,
and box-counting dimension. We develop the sucient theory to en-
able us to calculate the various fractal dimensions of some fractals.
We conclude by comparing dimensions and discussing their relative
merits.
Contents
1
Chapter 1. Introduction 1
Chapter 2. Fractal Examples 3
1. The Cantor Set 3
2. The Cantor Set as an Iterated Function System 6
3. Sierpinski Gasket 7
Chapter 3. Non-Fractal Dimension 9
1. Euclidean Dimension 9
2. Topological Dimension 9
Chapter 4. Fractal Dimension 15
1. The Hausdor Dimension 15
2. Similarity Dimension 18
3. Box-Counting Dimension 20
Chapter 5. Comparison of Dimensions 24
1. Comparison of Fractal Dimensions 24
2. Comparison of Topological Dimension and Fractal Dimension 25
Bibliography 26
0
CHAPTER 1
Introduction
There is no agreed upon denition amongst mathematicians of what is a fractal, and so
when talking about fractals we often consider objects that exhibit features associated
with fractals. Perhaps the two most important features associated with fractals are
the notions of self-similarity and non-integer dimension. A self-similar structure
can be dened as a structure that can be broken down into arbitrary small pieces,
each of which is a small replica of the entire structure.[3, p. 192] This notion of self-
similarity, which is crucial in the construction of fractals, is best studied through
iterated function systems, which we will touch upon but not study in depth. The
main feature of fractals that we will focus on is dimension, and in particular the
notion of non-integer dimension (which we call fractal dimension).
The topic of fractals has been of considerable popular interest since Benoit Mandel-
brot coined the term in 1975, and popularised the topic with books on how fractals
seemingly appear in nature. He also gave a rigorous denition in his book The Frac-
tal Geometry of Nature, which denes a fractal to be a set for which the Hausdor
Besicovitch dimension strictly exceeds the topological dimension.[1, p. 15] It follows
from this that sets with non-integer dimension are fractals. We may consider this
denition the motivation for the material in this project.
Although the term was only coined in 1975, the mathematical roots of fractals go
back to at least 1872, when Karl Weierstrass constructed his famous function that is
continuous and nowhere dierentiable. The graph of this function is today considered
a fractal, and is conjectured to have Hausdor dimension
3
2
. The roots of what we
today call fractal dimension go back to the work of Felix Hausdor and others in
the early 20th century. The discovery of topological dimension is often credited to
Henri Lebesgue, and the theory behind it was also developed around the beginning
of the 20th century.
As already noted, there is no agreement on what exactly constitutes a fractal, and
Mandelbrot himself wrote that his denition is rigorous but tentative and that if
and when a good reason arises, [it] ought to be changed.[1, p. 362]
1
Chapter 1 Page 2
There are many borderline fractals which do not t the criteria of Mandelbrots
denition in that they have integer dimension, but they nonetheless have irregular
self-similar structures that one associates with fractals. Examples of such sets are
the Cantor function and the space-lling curvesof Peano and Hilbert, discussed
in [2].
In this project we will begin with a chapter of examples, where we will give the con-
struction of two fractals: the Cantor set and Sierpinski gasket. We will discuss some
of their basic properties, but leave the discussion of dimension until later. Conven-
tional notions of dimension, which are useful for describing regular sets as well as
fractals, will be discussed in a chapter on non-fractal dimension. The notion of non-
integer dimension will be introduced for describing fractals. The fractal dimensions
we will discuss are the Hausdor dimension, similarity dimension, and box-counting
dimension, and we will use these to calculate the fractal dimensions of some fractals.
Finally we will conclude by comparing the dierent dimensions.
CHAPTER 2
Fractal Examples
Perhaps the best way to introduce fractals is by way of example. In this chapter we
will construct the two fractals called the Cantor set and the Sierpinski gasket, and
describe some of their properties. Often there are several ways to construct the same
fractal. In this chapter we will see that the Cantor set can be constructed either by
repeatedly removing segments from an interval; by considering the set of all numbers
with certain ternary expansions, or by nding the attractor of a system of recursive
functions (which we will call an iterated function system).
1. The Cantor Set
One of the rst fractals to be constructed was published by Georg Cantor (1845-
1918) in 1883 and is known as the Cantor set.[3, p. 65] The set is relatively simple
to construct, but has many interesting properties. Cantor originally constructed it as
an example of a perfect set that is nowhere dense. Much of this section is based on
[7] and [4, Ch. 1].
The Cantor set is dened on the closed interval C
0
= [0, 1] of the real line R, and is
generated by recursively removing the open middle third intervals of C
0
. Thus we
remove
_
1
3
,
2
3
_
to get C
1
=
_
0,
1
3

_
2
3
, 1

. Then, by removing the open middle thirds

of C
1
, i.e.
_
1
9
,
2
9
_
and
_
7
9
,
8
9
_
, we get C
2
=
_
0,
1
9

_
2
9
,
1
3

_
2
3
,
7
9

_
8
9
, 1

. Continuing
this process ad innitum gives us the Cantor Set C of closed intervals of [0, 1].
Figure 1. The construction of the Cantor set.
3
Chapter 2 Page 4
Notice that the sequence C
0
C
1
C
2
C
n
is decreasing for any n N, hence
we may formally dene the Cantor set to be the limit of the of the intersection of
these sets,
C =

n=0
C
n
.
The Cantor set C is formed by removing a countable number of open intervals from
[0, 1], the union of which is open. The complement of this union is closed, so C is
closed. C is clearly bounded and hence compact by the Heine-Borel theorem.
It is easy to see that C
n
consists of 2
n
intervals of length
_
1
3
_
n
. Thus the total length
of C
n
is 2
n
_
1
3
_
n
=
_
2
3
_
n
, and hence the total length of the Cantor set C is
lim
n
_
2
3
_
n
= 0.
This implies that the Cantor set has no intervals.
Observe that each endpoint of C
n
is an endpoint of C
n+1
, and that, by induction, each
endpoint of C
n
is an endpoint of C
m
for all m n. One might expect the Cantor set
to consist only of the endpoints of the sets C
n
, but we will see that this is not the
case.
First note that any number x in [0, 1] has a base 3 or ternary representation
x =

j=0
a
j
3
j
with digits a
j
{0, 1, 2}.
We shall denote ternary expansions with subscript 3 to distinguish them from decimal
expansions, and refer to a ternary point rather than a decimal point. Note that
some numbers can have two representations; for example,
1
3
= (0.0

2)
3
= (0.1

0)
3
.
1
Proposition 2.1. A point x [0, 1] belongs to the Cantor set C if and only if x has
a ternary expansion with digits a
j
{0, 2}.
Proof. This proof is based on [4, p. 4]. All ternary expansions of x have digit 1
at the rst place to the right of the ternary point if and only if x is strictly between
(0.1

0)
3
=
1
3
and (0.1

2)
3
=
2
3
. If we remove all such x we remove the open interval
_
1
3
,
2
3
_
and get C
1
, thus C
1
contains no x whose rst ternary digit is 1. Now, for
some x in C
1
, the second value to the right of the ternary point is 1 if and only if x
1
This is not unique to the ternary system; for example, the decimal expansion of
1
10
is 0.1

0 = 0.0

9.
Chapter 2 Page 5
is in
_
1
9
,
2
9
_
or
_
7
9
,
8
9
_
. These are the intervals removed from C
1
to get C
2
. Thus C
2
has no numbers with a ternary expansion with digit 1 at the rst or second place.
Proceeding similarly we can see that for n > 2, C
n
has no ternary digits with value 1
at the rst n places. Thus C =

n=0
C
n
is the set of points x in [0, 1] with a ternary
expansion without ternary digits 1.
It follows from this result that the Cantor set is uncountable (see [9, p.155] for a full
proof).
Proposition 2.2. The point
1
4
belongs to C but is not an endpoint of any C
n
.
Proof. Dividing 1 = (1)
3
by 4 = (11)
3
in ternary shows that
1
4
has ternary
representation (0.02)
3
, so
1
4
C by Proposition 2.1.
It remains to show that
1
4
is not an endpoint of any C
n
. By the construction of C,
it is clear that all endpoints of C
n
are of the form
m
3
k
for m, k non-negative integers.
We show that
1
4
is not of this form. Suppose (for a contradiction) that it is. Then
1
4
=
m
3
k
, and so 4m = 3
k
, and clearly 4m is even for all m. But 3
k
= 3 3
. .
ktimes
, and
since 3 is a prime, this is the unique prime factorisation of 3
k
. Hence 2 is not a prime
factor of 3
k
and so 3
k
is odd for all k, a contradiction.
We say a set X is perfect if it is precisely the set of limit points of X.
Theorem 2.3. The Cantor set C is a perfect set.
Proof. Any x C has a ternary expansion
x =

j=1
a
j
3
j
,
for a
j
{0, 2}. The representation of x up to k digits is given by
x
k
=
k

j=1
a
j
3
j
,
which is in C
k
. Now, to prove the theorem, we show that as k tends to innity, x
k
tends to the point x in C. We have
|x x
k
| =

j=k+1
a
j
3
j

j=k+1
2
3
j
=
1
3
k
.
Now lim
k
1
3
k
= 0, so by the squeezing rule, lim
k
|xx
k
| = 0, and so lim
k
x
k
=
x, as required.
Chapter 2 Page 6
A set X being perfect is equivalent to X being closed with no isolated points.
The Cantor set is totally disconnected space of total length (or Lebesgue measure)
0, so it is a collection of isolated points, each of which has (topological) dimension 0
(see Chapter 3). In this sense, one might expect the Cantor set to have dimension 0
(we see later that 0 is its topological dimension), but this value of dimension does not
seem to do the set justice, since the Cantor set is uncountable, and it has the same
cardinality of the 1-dimensional interval [0,1]. Thus a value between 0 and 1 seems
appropriate. We see later that the Cantor set has Hausdor dimension
log 2
log 3
0.6309.
2. The Cantor Set as an Iterated Function System
A useful way to dene a self-similar set such as the Cantor set is as the attractor of
an iterated function system (IFS) of contraction mappings. We dene these
notions now.
Let X, Y be metric spaces and d be the Euclidean metric. Let > 0. We dene a
similarity with scale factor to be a function f : X Y such that d(f(a), f(b)) =
d(a, b), for all a, b X.
In particular, f is a contraction mapping if it maps from X to itself and < 1.
An iterated function system on a complete metric space X is a set of contraction
mappings {f
i
: X X : i = 1, 2, . . . , k}, for k N, with scale factors {
i
: i =
1, 2, . . . , k}.
We say a non-empty compact set A X is an attractor of the above IFS if A =
f
1
(A) f
2
(A) f
k
(A).
For x [0, 1], the Cantor set can be generated by the following IFS:
f
1
(x) =
1
3
x;
f
2
(x) =
1
3
x +
2
3
.
We know C
0
= [0, 1], so C
1
= f
1
(C
0
) f
2
(C
0
), and so C
2
= f
1
(C
1
) f
2
(C
1
). So by
induction,
C
n
= f
1
(C
n1
) f
2
(C
n1
). (2.1)
The following proposition says what happens when we take the innite intersection
or limit. We show that C is the attractor of the IFS above.
Chapter 2 Page 7
Proposition 2.4.
C = f
1
(C) f
2
(C).
Proof. Note that Equation (2.1) can be written as C
n+1
= f
1
(C
n
) f
2
(C
n
).
Firstly, we want to show that C f
1
(C) f
2
(C). Suppose that x C. Then
x C
1
, and thus x
_
0,
1
3

or x
_
2
3
, 1

. Consider the case x

_
2
3
, 1

. Then for
any n, we know that x C
n+1
= f
1
(C
n
) f
2
(C
n
). But f
1
(C
n
) f
1
([0, 1]) =
_
0,
1
3

,
so x / f
1
(C
n
). Thus x f
2
(C
n
) =
C
n
+2
3
, and 3x 2 C
n
(for all n). Thus
3x 2

n=0
C
n
= C, and so x
C+2
3
= f
2
(C). Proceeding similarly for the case
x
_
0,
1
3

shows that x f
1
(C). Thus x C implies thatx f
1
(C) f
2
(C).
The second relationship to prove is C f
1
(C) f
2
(C). Suppose that x f
1
(C)
f
2
(C). Then x f
1
(C) or x f
2
(C). Consider the case x f
2
(C) =
C+2
3
. Then
3x 2 C, so 3x 2 C
n
for all n. Thus x
C
n
+2
3
= f
2
(C
n
) C
n=1
. Thus x

n=0
C
n+1
=

n=0
C
n
= C. The case x f
1
(C) is similar. Thus x f
1
(C) f
2
(C)
implies that x C, and we have proved the proposition.
3. Sierpinski Gasket
Here is another standard fractal. Consider, in the plane R
2
, the area enclosed by
the equilateral triangle S
0
with sides of length 1. Now S
0
can be split into 4 similar
equilateral triangles, with sides of length
1
2
(see Figure 3), by drawing a straight line
between each midpoint of the sides of S
0
. We remove the interior of the middle such
triangle, and call the area enclosed by the remaining 3 triangles S
1
. Splitting the 3
triangles of S
1
up and again removing the interior of the middle fourth triangle gives
us S
2
. Continuing, ad innitum, the process of removing middle fourth triangles
constructs the Sierpinski Gasket S.
We have S
0
S
1
S
2
S
n
. Then the Sierpinski Gasket is
S =

n=0
S
n
.
S
n
clearly consists of 3
n
triangles with sides of length
1
2
n
. Note that half the length
of one side is
1
2
n+1
, so the height of each triangle in S
n
is

_
1
2
n
_
2

_
1
2
n+1
_
2
=
_
1
2
sn

1
2
2n+2
=
_
3
2
2n+2
=

3
2
n+1
,
Chapter 2 Page 8
Figure 2. The construction of the Sierpinski Gasket.
and thus the area of each triangle in S
n
is
_
1
2
__
1
2
n+1
_
_

3
2
n+1
_
=

3
2
2n+3
,
and so the area of S is
lim
n

3
2
2n+3
= 0.
Each triangle in S
n
has perimeter 3
_
1
2
n
_
=
3
2
n
, and since the sides of each triangle do
not overlap, the total length of S is
lim
n
3
n
_
3
2
n
_
= lim
n
3
_
3
2
_
n
= .
As we have just seen, the Sierpinski Gasket has no area, and so it would make little
sense to talk of it being a two-dimensional object. In the sense that it has no area, it
is like the one-dimensional line, but unlike the line, it has innite length in a bounded
region, so in this way it is bigger than the one-dimensional line. If one were to allow
for non-integer dimension, then one would guess the dimension of S to be between 1
and 2. We shall see later that S has fractal dimension
log 3
log 2
1.5849.
CHAPTER 3
Non-Fractal Dimension
1. Euclidean Dimension
The Euclidean dimension dim
E
of a set is simply the number of co-ordinates re-
quired to describe each point in that set.[5, p. 11] It follows that dim
E
is always an
integer and that any set dened in the Euclidean space E
n
(or R
n
) has dim
E
= n. This
notion of dimension makes sense for describing a space as a whole, but is less accurate
for describing sets in that space. For example, both a cube and an isolated point in
E
3
have Euclidean dimension 3. Also it does not adequately describe a shape such as
(the boundary of) a circle in R
2
. In one sense, the circle is 2-dimensional since the
points in the set are occupying 2-dimensions, but the circle itself is a length without
breadth, which makes one think of it as being a one-dimensional object. This latter
notion of dimension is what we will dene later as topological dimension.
Example 3.1. Trivially, dim
E
R
n
= n.
The Cantor set C has dim
E
C = 1, since it is dened in R.
The Seirpinksi Gasket S has dim
E
S = 2, since it dened in R
2
.
2. Topological Dimension
If we consider length, breadth, and depth to be the 3 dimensions of R
3
, then
we can see that the intuition for topological dimension goes back to Euclid, who says
that a point is that of which there is no part; a line is a length without breadth;
a surface is that which has length and breadth only, and a solid is a (gure)
having length, breadth and depth.[8] It was not, however, until the 20th century
that a formal denition of topological dimension was arrived at.
Unlike Euclidean dimension, Topological dimension aims to dene the dimension of
a set in a way that is irrespective of the dimension of the space that the set is dened
in. Naturally, any sensible denition of the topological dimension of a set will not
allow the dimension of a set to be greater than the dimension of the space it is dened
in. Also, it will not allow for topologically equivalent (homeomorphic) sets to have
dierent dimensions. The two most important denitions of topological dimension
9
Chapter 3 Page 10
Figure 1. A covering of a line by discs, and a covering of a line where
each point is not covered by more than 2 discs.
are Lebesgue covering dimension and inductive dimension. We shall restrict
out discussion to the former.
2.1. Lebesgue Covering Dimension. The main idea behind the Lebesgue cov-
ering dimension is that of covering an object with an open covering of arbitrarily
small open sets (for two-dimensional objects, we often use discs; in three-dimensions,
spheres). If one tries to cover a line with a number of small open discs, one nds
that to cover the whole line there must be some overlapping of discs, but that each
point in the line need not be covered by more than n = 2 discs. Similarly, if one tries
to cover the area enclosed by a rectangle, one nds again that there is an overlapping,
but that each point in the rectangle need not be covered by more than n = 3 discs.
In fact, we will see that this corresponds to a set having topological dimension n 1
(which gives the line and the area of the rectangle dimensions 1 and 2 respectively).
In this section we will develop enough theory to enable us to calculate the topological
dimension of certain fractals. The material for this section up to Denition 3.13 is
based on [6, 50, 27].
Definition 3.2. We say a collection A of subsets of a metric space X has order
n+1 if some point of X lies in n+1 elements of A, and all points in X lie in at most
n + 1 elements of A.
Definition 3.3. Let A and B be two collections of subsets of a metric space X.
Then we say B renes A if for every B B there exists A A such that B A.
Definition 3.4. We say a space X has nite dimension if there exists an integer n
such that for every open covering A of X, there is an open covering B of X such that
B renes A, and B has order at most n+1. We call the smallest such n the Lebesgue
covering dimension (or topological dimension) of X, which we denote dim
T
X.
Definition 3.5. We denote by d the familiar Euclidean metric on R
n
given by
d(x, y) = ||x y|| =
_
(x
1
y
1
)
2
+ (x
1
y
1
)
2
+ + (x
n
y
n
)
2
,
where x = (x
1
, x
2
, . . . , x
n
) and y = (y
1
, y
2
, . . . , y
n
) are dened in R
n
.
Chapter 3 Page 11
Definition 3.6. Let (X, d) be a metric space, and let A be a nonempty subset of X.
Then for x X we dene the distance between x and A to be
d(x, A) = inf{d(x, a) : a A}.
Lemma 3.7. Let A be a subset of X, and set g(x) = d(x, A). Then g is continuous.
Proof. Let x, y X. Then for each a A we have (by the triangle inequality)
d(x, A) d(x, a) d(x, y) + d(y, a),
and so
d(x, A) d(x, y) d(y, a) = inf d(y, a) = d(y, A).
Thus
d(x, A) d(y, A) d(x, y).
A similar argument gives
d(y, A) d(x, A) d(x, y).
This implies that g(x) is continuous.
The following lemma is known as Lebesgues number lemma, and is useful for
nding a renement of an open covering of a compact metric space.
Lemma 3.8. Let A be an open covering of the metric space (X, d), where X is compact.
Then there exists a > 0 such that for every subset B X with diameter less than
, there is an A A such that B A. We call a Lebesgue number of the
covering A.
Proof. In the case where X = A, for some element A A, we have that every
subset of X is contained in A, hence any > 0 is a Lebesgue number for A.
In the case where X is not an element of some A A, consider the nite subcover
{A
1
, . . . , A
n
} A of X, which exists by the compactness of X.
Set C
i
= X A
i
, and dene f : X R by
f(x) =
1
n
n

i=1
d(x, C
i
).
That is, for each x X, f(x) takes the value of the average of the minimum distances
between x and each C
i
.
Chapter 3 Page 12
We show that f(x) > 0 for all x X. For some x X, choose i such that x A
i
.
Now choose > 0 such that B

(x) A
i
. (We know that such an exists since A
i
is
open). Now we have d(x, C
i
) (by the denition of C
i
), and so f(x)

n
. Now
by Lemma 3.7 is is clear that f is continuous, and we know that f is greater than 0,
hence f has a minimum value . Let S be a subset of X, with diameter < . Then
there exists some x
0
S such that B

(x
0
) S. Now choose m such that d(x
0
, C
m
)
is the largest of all the numbers d(x
0
, C
i
). Then
f(x
0
) d(x
0
, C
m
).
Thus B

(x
0
) A
m
, and since S B

(x
0
) we have S A
m
, and so is a Lebesgue
number for X.
Theorem 3.9. Let X be a compact subspace of R. Then dim
T
X 1.
Proof. We shall show that there exists an open covering B of X with order 2
that renes every open covering A of X, thus proving the theorem.
Dene B
1
to be the collection of all open intervals in R of the form (k, k+1), for k some
integer. Let B
2
be the collection of all open intervals in R of the form (k
1
2
, k +
1
2
),
for k an integer. Then B = B
1
B
2
is an open covering of R. Note that all sets in B
1
and B
2
have diameter 1, and are disjoint. Thus a point in B can be in at most one
set of B
1
and one set of B
2
, hence the open covering B has order 2.
Now let X be a compact subspace of R. Consider an open covering C of X. Then this
covering has a Lebesgue number > 0. This means that any collection of subsets of
X with diameter less than is a renement of C. Now consider the map f : R R
given by f(x) =
1
2
x, where the domain is the set of all elements x in A. Then the
image of f provides an open covering, call it D, of order 2 with each element in D
having diameter
1
2
, which is less than , as required. The intersection of D with X
provides a covering that renes C and has order 2.
Example 3.10. The interval [0, 1] has dim
T
= 1.
Let X = [0, 1]. Then by Theorem 3.9, dim
T
X 1. Now let A be a covering of X
by the intervals [0, 1) and (0, 1]. Then A has order 2. Note that any arbitrary open
covering of X must be a renement of A. We show that if B is an open covering of
X that renes A, then it has order at least 2.
Let U be an element of B and let V be the union of the other elements in B. Suppose
that B has order 1. Then, by denition, no point in X lies in more than 1 element of
B. This implies that U and V do not overlap and are disjoint. But then B is not a
covering of X, a contradiction. Thus B has order at least 2, and so dim
T
X 1.
Chapter 3 Page 13
Theorem 3.11. Let X be a nite dimensional space, and let Y be a closed subspace
of X. Then Y is nite dimensional and dim
T
Y dim
T
X.
Proof. Let dimX = n and let A be a collection of sets open in Y that covers Y
(i.e. an arbitrary open covering of Y ). For each A A, choose an open set A

in X
such that A

Y = A. Now consider a covering B of X composed of each A

set and
the (open) set X Y . Since dim
T
X = n, by denition there is an open covering C
of X that renes B and has order at most n + 1. Then {C Y : C C} provides a
covering of Y of open sets in Y that renes A and has order at most n +1. Hence Y
has nite dimension, and dimension at most n.
Definition 3.12. From [4, p. 41]. Let A be a subset of a metric space (X, d), where
d is the Euclidean metric. Then we say that the diameter of A is
diam A = sup{d(x, y) : x, y A}
Definition 3.13. If A is a covering of the metric space (X, d) then the mesh of A
is the supremum of the diameters of of all elements A A.
Theorem 3.14. Let X be a compact space. Then dim
T
X n if and only if for every
> 0 there exists an open cover of X with order n and mesh
Proof. Proof based on [4, p. 98]. Suppose dim
T
X n, and let > 0. Then the
collection A of all open subsets of diameter is a covering of X. Since dim
T
X n,
by denition there is a renement B of A of order at most n. Since B is a renement
of A, clearly mesh B mesh A = .
To prove the other direction, suppose that for every > 0 there exists an open cover
of X with order n and mesh . Let A be an arbitrary open cover of X. Then,
by Lemma 3.8, A has a positive Lebesgue number . Let B be an open cover of X
with order at most n and mesh smaller than both and . Then by Lemma 3.8, B is
a renement of A. Hence dim
T
X n.
Example 3.15. The Cantor set C has dim
T
C = 0.
Recall that for an any n, C
n
is made up of 2
n
intervals of length
1
3
n
. Note that any two
of these intervals are apart by a distance of at least
1
3
n
. Now let r > 0 be a number
such that r is strictly less than
1
2

1
3
n
. Then the collection A of rneighbourhoods of
each of the 2
n
intervals of C
n
forms an open covering of C
n
with order 1 and mesh at
most
1
3
n
+ 2r.
Chapter 3 Page 14
Now for any > 0 we can choose n large enough such that
1
3
n
+ 2r , and we have
that A is a covering of C
n
(and C) with order 1 and mesh . Thus by Theorem
3.14 we have dim
T
S 0.
Example 3.16. The Sierpinksi Gasket S has dim
T
S = 1.
Recall that each S
n
is composed of the interior and boundary of 3
n
triangles of side
length
1
2
n
. Any two two triangles in S
n
are either adjacent, in which case they meet
at a point, or they are not adjacent, in which case they are apart by a distance of at
least
1
2
n

2

3
(this value is the height of one triangle). Now let r be a number such
that 0 < r <
1
2
n

1

3
(so r is less than half the height of one triangle). Now consider
the collection A of rneighbourhoods of every triangle in S
n
. Then A forms an open
covering of S
n
(and S) that has order 2 and mesh at most
1
2
n
+ 2r.
Now for any > 0 we can choose n large enough so that
1
2
n
+ 2r , and we have
that A is a covering of S
n
(and hence S) of order 2 and mesh at most . Thus by
Theorem 3.14 we have dim
T
S 1. But S contains the compact subspace [0, 1],
which has topological dimension 1 (by Example 3.10), so by Theorem 3.11 we have
1 dim
T
S.
CHAPTER 4
Fractal Dimension
We discussed briey, in chapter one, why one might expect the dimension of some
sets, such as the Cantor set and Sierpinksi Gasket, to take a non-integer value. It was
supposed that the dimension of the Cantor Set should be between 0 and 1. We have
seen in the previous chapter that 0 and 1 are the values of the Cantor sets topological
and Euclidean dimension respectively. In this section we allow for the Cantor Set to
take a dimensional value strictly between 0 and 1. In general, we call a dimension
that can take a non-integer value a fractal dimension. Mathematicians have come
up with many fractal dimensions, each with advantages and disadvantages. Perhaps
the most important one is the one that Mandelbrot singled out in his denition of a
fractal, namely the Hausdor dimension (or Hausdor Besicovitch dimen-
sion). Other useful fractal dimensions, which we shall discuss in this chapter, are the
similarity dimension, lower box-counting dimension, upper box-counting
dimension, and box-counting dimension. For many sets, the values of these di-
mensions coincide, and one may talk loosely of the fractal dimension of the set.
Note that any set of non-integer Hausdor dimension is certainly a fractal, by Man-
delbrots denition, since topological dimension is always an integer. The following
section is based on [10].
1. The Hausdor Dimension
If U is a non-empty subset of R
n
then the diameter of U is the greatest distance
between any two points in U, i.e. |U| = sup {|x y| : x, y U}. We dene a -cover
of a set X to be a countable collection of sets {U
i
} of diameter at most such that
X

i=1
for all i. Now consider the sum of the s-th powers of each |U
i
|. The
minimal such cover of X, for a given > 0, is dened as the Hausdor content
of X
H
s

(X) = inf
_

i=1
|U
i
|
s
: U
i
is a -cover of X
_
.
The limit as 0 is the Hausdor measure
H
s
(X) = lim
0
H
s

(X).
15
Chapter 4 Page 16
Figure 1. A -covering of a surface by balls.
It can be checked that the Hausdor measure is indeed a measure. Another fact that
we will state and not prove, is that all Borel sets are H
s
measurable. All the sets
considered in this project are Borel sets, and so we only deal with H
s
measurable
sets.
We shall now work towards dening the Hausdor dimension.
Consider a -cover {U
i
} of X. Then |U
i
| by denition. It follows that for some
t > s,
|U
i
|
ts

ts
,
and so
|U
i
|
t

ts
|U
i
|
s
.
Then taking sums

i=1
|U
i
|
t

i=1

ts
|U
i
|
s
Then taking inma
H
t

(X)
ts
H
s

(X).
Now taking the limits as 0, we have H
t
(X) = 0 if H
s
(X) < . The contra-
positive implies that if H
t
(X) = 0, then H
t
(X) > 0 and H
s
(X) = . So there is a
critical value of s, call it s
0
, that is between 0 and ; all s less than this value will
have H
s
= , and all s greater than this value will have H
s
= 0. We call this critical
value s
0
the Hausdor dimension of X, written formally as
dim
H
X = inf{s : H
s
(X) = 0} = sup{s : H
s
(X) = }.
Chapter 4 Page 17
We shall now state and prove two fundamental properties of Hausdor dimension.
Theorem 4.1. Let A and B be Borel sets. If A B, then dimA dimB.
Proof. From [4, p. 168]. A B implies that H
s
(A) H
s
(B) for all s (by a
property of measures). Thus sup{s : H
s
(A) = } sup{s : H
s
(B) = }; that is,
dimA dimB.
Theorem 4.2. Let X R
n
, and > 0. Then
H
s
(X) =
s
H
s
(X).
Proof. We write

as shorthand for

i=1
. Let {U
i
} be a -cover of X. Then
{U
i
} is a -cover of X. Hence
H
s

(X)

|U
i
|
s
=
s

|U
i
|
s

s
H
s

(X).
As 0, we have H
s
(X)
s
H
s
(X). To prove the inequality holds the other way,
we proceed similarly. Let {U
i
} be a -cover of X. Then {
1
U
i
} is a
1
-cover of
X. Thus
H
s

(X)

|
1
U
i
|
s
=
s

|U
i
|
s

s
H
s

(X).
So as 0, we have
s
H
s
(X) H
s
(X) as required.
Example 4.3. We will show that the Cantor set C has dim
H
C =
log 2
log 3
0.6309.
We will rst guess the value of s for the dimension of the Cantor set C, and then
prove dim
H
C = s.
The Cantor set C can be split into a two parts, C
L
= C [0,
1
3
], and C
R
= C [
2
3
, 1].
These are similarities of C with scaling factor
1
3
. Now C is the disjoint union C
L
C
R
,
so by a basic property of measures, and then by Proposition (4.2), we have, for any s
H
s
(C) = H
s
(C
L
) H
s
(C
R
) =
_
1
3
_
s
H
s
(C) +
_
1
3
_
s
H
s
(C).
Then if H
s
(C) is nite and non-zero at the critical value (we will prove it is soon),
we can divide by H
s
(C) to get 1 =
_
1
3
_
s
+
_
1
3
_
s
. Solving for s, rst note that
_
1
3
_
s
=
_
1
2
_
. (1.1)
Then taking logs we see that s log
_
1
3
_
= log
_
1
2
_
, and so s =
log
(
1
2
)
log
(
1
3
)
=
log 2
log 3
0.6309.
Now that we have guessed the critical value of s, we must show that 0 < H
s
(C) <
for s =
log 2
log 3
. This will imply that dim
H
C = s.
Chapter 4 Page 18
Consider a covering {U
i
} of C. Recall that C consists of 2
n
intervals of length
1
3
n
.
Then for a minimum covering of C, we have sum consisting of 2
n
sets of diameter
|
1
3
n
|. Hence for s =
log 2
log 3
(and writing

as shorthand for

i=1
) we have
H
s
3
n(C)

|U
i
|
s
= 2
n
3
ns
= 2
_
1
3
_
s
= 2
_
1
2
_
= 1.
Letting n , we get H
s
(C) 1 < .
It now remains to show that H
s
(C) > 0. We will show that H
s
(C)
1
2
=
_
1
3
_
s
. For
each U
i
, we have some n such that
1
3
n+1
|U
i
| <
1
3
n
(1.2)
Then U
i
can intersect at most one interval of C
n
, since the gaps between intervals
of C
n
are of length
1
3
n
. If m n, the number of intervals of C
m
that U
i
intersects is
at most 2
mn
= 2
m
3
sn
2
m
3
s
|U
i
|
s
. (The rst equality is by (1.1). To see the last
inequality, observe that
1
3
n
3|U
i
| by (1.2), then take powers).
If we choose m such that
1
3
m+1
|U
i
| for all U
i
, then the set {U
i
} will intersect all 2
m
intervals of length
1
3
m
. Hence, counting each interval, we get
2
m

2
m
3
s
|U
i
|
s
So 1

3
s
|U
i
|
s
, hence 3
s

|U
i
|
s
, as required.
2. Similarity Dimension
In this section we will dene similarity dimension, which is an easy to calculate
dimension, but is unfortunately only applicable to a relatively small number of sets,
and not all fractals. In fact we will see later that it gives an exact estimate of Hausdor
dimension for self-similar for which Morans open set condition holds (dened later).
When dealing with self-similar sets, similarity dimension has the advantages of being
somewhat more intuitive and easier to deal with than the Hausdor dimension. Before
we give the denition, we give the following motivation for it, based on Mandelbrots
exposition in [1, p. 37].
Start by considering the one dimensional Euclidean space E (or just the real line R).
It is clear that any interval [0, a) on this line can be partitioned into a number N = b
Chapter 4 Page 19
Figure 2. The construction of a line, square, and cube as described
on this page.
of equal parts. Then for x [0, a) we have 0 x < a, or
(k 1)a
b
x <
ka
b
,
for k = 1, . . . , b. Each part is a scaling, or similarity, of the whole interval with
similarity ratio r =
1
b
=
1
N
. If we expand along a perpendicular axis, we can also split
this line into b parts and we will have an area enclosed by a square in the Euclidean
plane E
2
that is split into N = b
2
parts. These parts are then a similarity of the
whole square with similarity ratio r =
1
b
=
1
N
1
2
. Similarly, in E
3
we have a cube and
r =
1
N
1
3
. So all shapes we dene like this are of the form
r =
1
N
1
s
, (2.3)
where S corresponds to the shapes Euclidean dimension, and s is a solution of
Nr
s
= 1. (2.4)
Taking the logarithm of both sides of equation (2.3), we get log r = log
_
1
N
1
s
_
=
(log N)
s
, and so
s =
log N
log r
=
log N
log
_
1
r
_.
Definition 4.4. We can now dene the similarity dimension of self-similar objects
as
dim
S
=
log N
log
_
1
r
_,
Chapter 4 Page 20
where N is the number of parts the object is split into, and r is the ratio of the
similarities.
We can also dene the similarity dimension of a set X, where X is the attractor of
an IFS (as described in Chapter 2). Consider the set of scale factors of the IFS. Note
that these scale factors (ratios) do not all have to be equal, unlike the single similarity
ratio in (3.8). Thus for an IFS, s is the solution of
k

i=1

s
i
, (2.5)
and we call S the similarity dimension of X.
It can be checked that the solution of (2.5) exists and is unique (see [7]).
Example 4.5. The Cantor set C has
dim
S
C =
log 2
log 3
0.6309.
This is immediate from Denition 4.4, with N = 2 and r =
1
3
.
Example 4.6. The Seirpinksi gasket S has
dim
S
S =
log 3
log 2
1.585,
since we have 3 similarities with scaling factor
1
2
.
3. Box-Counting Dimension
The third kind of fractal dimension we will discuss is the box-counting dimension. It
has the advantage of being easier to calculate than the Hausdor dimension, whilst
also being applicable to a wider variety of sets than the similarity dimension.
The idea is to split an area, for example the unit square R
2
, into a mesh of boxes
(or cubes) of length . Observe that if the unit line is split into boxes (in this
case, intervals suce) of length , the number of boxes covering the line will be
1

.
Extending to the unit square, the number of boxes of side length required to cover
the square is (
1

)
2
. Similarly, any unit interval in R
n
(i.e. the area enclosed by a box
of Euclidean dimension n) is covered by (
1

)
n
boxes (or cubes, etc) of side length .
The value of n corresponds to the dimension of the unit interval. We can study the
dimensions of non-regular objects (such as fractals) in this way by replacing n with s,
where s can take the value of a non-negative real number (not necessarily an integer).
Chapter 4 Page 21
If we split the plane into a mesh of boxes of length , then any set X in the plane
can be completely covered by a number of boxes, we call the minimum such number
N. The value of N clearly depends on the the size of the boxes. Notice that as
decreases, the number of boxes required to cover X increases. We write N

(X) for
the minimum number of boxes of length required to completely cover
X. As we have seen above, we can expect an object to be covered by N

(X) = (
1

)
s
boxes, where s corresponds to the dimension of the object. Solving for s, we take logs
of both sides to get
s =
log N

(X)
log
1

=
log N

(X)
log
.
We are interested in what happens as becomes arbitrarily small, so we dene the
box-counting dimension of a set X in R
n
to be
dim
B
X = lim
0
log N

(X)
log
,
where N

(X) is the smallest number of sets of diameter not exceeding , required

to cover X. In general, this limit may not exist, in which case the box-counting
dimension is not dened. However, the limit exists if (and only if) the upper limit
and lower limit are equal, then the (ordinary) limit takes the same value as the upper
and lower limit. We dene then, the lower box-counting dimension dim
B
X and
upper box-counting dimension dim
B
X respectively by
dim
B
X = liminf
0
log N

(X)
log
,
dim
B
X = limsup
0
log N

(X)
log
.
Note the general inequality dim
B
X dim
B
X dim
B
X. So if dim
B
X = dim
B
X,
then dim
B
X = dim
B
X = dim
B
X.
Note that by the above construction its clear that dim
B
X dim
E
X, since X is
simply the amount of co-ordinates of the space that X is dened in.
We can take the limit as tends to 0 directly, or through a decreasing sequence like
in the following lemma.
Chapter 4 Page 22
Lemma 4.7. Let X be a subset of R
n
, and let
n
be a decreasing sequence such that
c

n+1

n
, for some 0 < c < 1, and c
n
=
k
. Then
dim
B
X = lim
n
log N

n
(X)
log
n
.
Proof. If
n+1
<
n
then
log N

(X)
log

log N

n+1
(X)
log
n

log N

n+1
(X)
log
n+1
+ log(
n+1
/
n
)

log N

n+1
(X)
log
n+1
+ log c
.
Taking upper limits we have,
limsup
0
log N

(X)
log
limsup
n
log N

n
(X)
log
n
.
Also, taking lower limits, we have
liminf
0
log N

(X)
log
limsup
n
log N

n
(X)
log
n
.
A similar argument shows that the opposite inequalities hold.
Example 4.8. The Cantor set C has dim
B
C =
log 2
log 3
0.6309
For any C
n
, we require a covering of the 2
n
intervals of length
1
3
n
. Consider > 0
such that
1
3
n
<
1
3
n1
.
Then it takes at most 2
n
intervals of length to cover C
n
(and C), so we have
N

(C) 2
n
. Thus (making use of Lemma 4.7), we have
dim
B
C = limsup
0
log N

(C)
log
= limsup
n
log N

n
(C)
log
n
limsup
n
log 2
n
log 3
n1
=
log 2
log 3
.
To prove the opposite inequality, consider intervals of length where
1
3
n+1
<
1
3
n
.
Then each interval intersects at most one of the 2
n
intervals of length
1
3
n
of C
n
. Thus
at least 2
n
intervals of length are needed to cover C
n
(and C), so N

(C) 2
n
.
Proceeding as before and taking lower limits gives dim
B
C
log 2
log 3
as required.
The box-counting dimension has some undesirable properties. One is that box-
counting dimension is not countably stable. Note that any singleton point has box-
counting dimension 0 (as one would expect), since it takes only 1 box of of side length
to cover a point, and lim
0
log 1
log
= 0. However, a countable collection of points
does not necessarily have box-counting dimension 0. We shall illustrate this with an
example.
Chapter 4 Page 23
Example 4.9. Let S be the closure of the set of all points of the form
1
n
, for all
positive integers n; that is, S = {0, 1,
1
2
,
1
3
, . . . }. We show that dim
B
S =
1
2
.
We want to consider coverings of S by sets of arbitrarily small diameter. So let
0 < <
1
2
, and let n be the integer such that
1
n(n + 1)
<
1
n(n 1)
(3.6)
Let U be an interval such that |U| . Then
|U| <
1
n(n 1)
=
1
(n 1)

1
n
.
So |U| is less than the distance between the n-th point and the (n 1)-th point in
the set, and so it takes at least n sets of diameter to cover S. Thus N

(S) n,
and hence log N

(S) log n. Taking logs of the (non-strict) inequality in (3.6) gives

log log n(n + 1). Taking quotients of these last two inequalities gives
N

(S)
log

log n
log n(n + 1)
. (3.7)
It is clear by (3.6) that as 0, n . Thus by letting 0 in (3.7), and then
applying lHopitals rule twice, we get
dim
B
S lim
n
log n
log n(n + 1)
= lim
n
n + 1
2n + 1
= lim
n
1
2
=
1
2
.
Now exploiting the other side of (3.6),

1
n(n + 1)
=
1
n

1
n + 1
,
so it takes (k + 1) intervals of length to cover [0,
1
n
], in particular, 2 intervals of
length to cover [0, 1], so N

(S) 2n. Taking logs of this, and dividing by the logs

of (3.6), we get
N

(S)
log

log 2n
log n(n 1)
. (3.8)
Thus as 0, we get
dim
B
S lim
n
log 2n
log n(n 1)
= lim
n
n 1
(2n 1)
= lim
n
1
2
=
1
2
.
Hence dim
B
S = dim
B
S = dim
B
S =
1
2
.
CHAPTER 5
Comparison of Dimensions
As we have seen in the previous chapter, some dimensions are easier to calculate than
others, and we have seen with the fractal examples that dimensions often coincide.
We will now compare dimensions for sets in general, and show under what conditions
these dimensions coincide. Hausdor dimension is the most important denition of
dimension and we will tend to compare others to it. We will rst compare fractal
dimensions, and then compare fractal dimension to topological dimension. Note that
the Euclidean dimension is of little interest; it is always greater than or equal to the
fractal dimension or topological dimension of any set.
1. Comparison of Fractal Dimensions
Theorem 5.1. If X R
n
is bounded, then
dim
H
X dim
B
X dim
B
X.
Proof. From [10]. Let X be a set covered by N

(X) sets of diameter . Now,

consider a -covering of X; that is, a covering of sets of diameter at most . It is
clear, by the denition of the Hausdor content of X, that we have
H
s

(X) N

(X)
s
. (1.1)
Now as gets small enough such that H
s

(X) > 1, we have (by 1.1) 1 < N

(X)
s
.
Taking logarithms, 0 < log N

(X)
s
= log N

(X) + s log . Thus

s <
log N

(X)
log
,
and so taking the lower limit we have
s liminf
0
log N

(X)
log
= dim
B
X dim
B
X.

The key dierence between the two dimensions is that Hausdor dimension allows for
coverings of variable diameter, whilst the box-counting dimension demands coverings
24
Chapter 5 Page 25
be of same diameter. We have seen, with the Cantor set and Sierpinski gasket, an
example where the upper box-counting dimension is equal to the Hausdor dimension.
Now, we say that an IFS satises Morans open set condition if and only if there
exists a nonempty open set U such that f
i
(U) f
j
(U) = for all i not equal to j and
U f
i
(U). Then a theorem says that if a set X is satised, then dim
S
X = dim
H
X
(for a proof, see [4, p. 191]).
Its easy to check (without using the theorem) that the Cantor set and Sierpinski
gasket satisfy Morans open set condition.
2. Comparison of Topological Dimension and Fractal Dimension
In the case of a set that is a fractal, according to Mandelbrots denition we always
have that dim
T
< dim
H
.
We shall now state some results, from [4, p. 217], about the relationship between
topological dimension and Hausdor dimension for sets in general (the relationship to
other fractal dimensions follows from the previous section). To proofs of the following
results are beyond the scope of this project, for they require a background in algebraic
topology and sometimes Lebesgue integration.
In general, it is the case that for any metric space X, we have dim
T
X dim
H
X.
This allows one to automatically deduce that any set with non-integer Hausdor
dimension is a fractal.
It is also true that if X is homeomorphic to Y , then dim
T
X = dim
T
Y . We conclude
this chapter by stating a nice theorem about the relationship between topological
dimension and Hausdor dimension: if X is a separable metric space then
dim
T
X = inf{dim
H
Y : Y is homeomorphic to X}.
Bibliography
[1] Mandelbrot, B.B. 1982 The Fractal Geometry of Nature (W.H. Freeman and Company)
[2] Sagan, Hans 1994 Space-Filling Curves (Springer-Verlag)
[3] Jrgens, Hartmut, Heins-Otto Peitgen, and Dietmar Saupe 1992 Chaos and Fractals: New
Frontiers of Science (Springer-Verlag: New York)
[4] Edgar, Gerald A. 2008 Measure, Topology, and Fractal Geometry. 2nd. ed. (Springer-Verlag:
New York)
[5] Addison, Paul S. 1997 Fractals and Chaos: An Illustrated Course (Institute of Physics Pub-
lishing)
[6] Munkres, James. 1999 Topology (Prentice Hall)
[7] Theriault, Stephen Lecture Notes: Chaos and Fractals
[8] Euclid Elements of Geometry Trans. Richard Fitzpatrick
[9] Alligood, K. T. 1996 Chaos: an introduction to dynamical systems (Springer-Verlag: New
York)
[10] Falconer, Kenneth 1990 Fractal Geometry (John Wiley and Son Ltd)
26