Fractal Geometry - Notes

Introduction
There is no hard and fast definition of a fractal set, but just a list of characteristic properties.
Suppose that 𝐹 is a fractal set, then 𝐹 has some (or all) of the following properties:
(i) 𝐹 has fine structure, that is, detail on arbitrary small scales;
(ii) 𝐹 is too irregular to be described in traditional geometrical language, both locally and
globally;
(iii) often 𝐹 has some form of self-similarity, perhaps approximate or statistical;
(iv) usually the fractal dimension of 𝐹 (defined in some way) is greater than its topological
dimension;
(v) in most cases of interest, 𝐹 is defined in a very simple way, perhaps recursively.
 Fractal Geometry - Notes
Mathematical Background
Basic Set Theory
Theorem A set 𝐴 is closed if, and only if:
(i) its complement is open;
(ii) it is equal to its closure, i.e. 𝐴 = 𝐴;
(iii) it contains all of its limit points.
Theorem A set 𝐴 is open if, and only if, it is equal to its interior, i.e. 𝐴 = int 𝐴.
Theorem The union of any collection of open sets is open, as is the intersection of any finite
number of open sets.
Theorem The intersection of any collection of closed sets is closed, as is the union of any
finite number of closed sets.
Theorem A point 𝑥 ∈ ℝ𝑛 is a limit point of a set 𝐴 if, and only if, for every 𝑟 > 0 the open
ball 𝐵 o (𝑥, 𝑟) contains points of 𝐴 other than 𝑥.
Theorem A point 𝑥 ∈ ℝ𝑛 is a boundary point of a set 𝐴 if, and only if, the ball 𝐵(𝑥, 𝑟)
intersects both 𝐴 and its complement for all 𝑟 > 0.
Theorem The diameter of a set 𝐴 is equal to the diameter of its closure, i.e. |𝐴| = |𝐴|.
Theorem Let 𝐴 ⊆ ℝ𝑛 , then 𝐴 is compact if, and only if, 𝐴 is closed and bounded.
Theorem The intersection of any collection of compact sets is compact, as is the union of
any finite number of compact sets.
Theorem Suppose that 𝐴1 ⊃ 𝐴2 ⊃ ⋯ is a decreasing sequence of non-empty compact sets.
Then the intersection 𝐴 = ⋂∞ 𝑖=1 𝐴𝑖 is non-empty. Moreover, if 𝐴 is contained in some open set
𝑉, then the finite intersection ⋂𝑘𝑖=1 𝐴𝑖 is contained in 𝑉 for some 𝑘.
Theorem Suppose that 𝐴 and 𝐵 are compact subsets of ℝ𝑛 . Then 𝐴 and 𝐵 are a positive
distance apart, that is
inf{|𝑥 − 𝑦|: 𝑥 ∈ 𝐴, 𝑦 ∈ 𝐵} > 0.

Functions and Limits

Certain functions from ℝ𝑛 to ℝ𝑛 have a particular geometrical significance; often, in this
context they are referred to as transformations.
A linear transformation may be represented by a matrix 𝐴, that is
𝑇(𝑥) = 𝐴𝑥.
Theorem A non-singular linear transformation is invertible.
Theorem Suppose that 𝑇 is a linear transformation represented by a matrix 𝐴. Then 𝑇 is
invertible if, and only if, 𝐴 is invertible. In which case the inverse of 𝑇 is represented by the
inverse of 𝐴, that is
𝑇 −1 (𝑥) = 𝐴−1 𝑥.
Theorem A bi-Lipschitz function 𝑓: 𝑋 → 𝑌 has an inverse 𝑓 −1 : 𝑓[𝑋] → 𝑋.
To show that a function 𝑓(𝑥) has a limit 𝑦 as 𝑥 tends to 𝑎, we must show that, given 𝜀 > 0,
there exists a 𝛿 > 0 such that
 Fractal Geometry - Notes
|𝑥 − 𝑎| < 𝛿 ⇒ |𝑓(𝑥) − 𝑦| < 𝜀
for all 𝑥 ∈ 𝑋.
A function 𝑓: ℝ𝑛 → ℝ may fluctuate wildly as 𝑥 → 𝑎 and so need not have a limit there. We
use upper and lower limits to describe such fluctuations. The upper and lower limits exist (as
real numbers or as ±∞) for every such function 𝑓 and are indicative of the variation of 𝑓 for 𝑥
close to 𝑎. If the upper and lower limits exist and are equal, then lim 𝑓(𝑥) exists an is equal to
𝑥→𝑎
this common value.
Theorem Let 𝑓: 𝑋 → ℝ be a function where 𝑋 ⊆ ℝ𝑛 and let 𝑎 ∈ 𝑋, then
lim 𝑓(𝑥) ≤ lim 𝑓(𝑥).
𝑥→𝑎 𝑥→𝑎

Theorem Let 𝑓, 𝑔: ℝ𝑛 → ℝ be functions. If there exists 𝛿 > 0 such that 𝑓(𝑥) ≤ 𝑔(𝑥) for 𝑥 ∈
(𝑎, 𝛿), then
lim 𝑓(𝑥) ≤ lim 𝑔(𝑥) and lim 𝑓(𝑥) ≤ lim 𝑔(𝑥).
𝑥→𝑎 𝑥→𝑎 𝑥→𝑎 𝑥→𝑎

Note that, even if 𝑓(𝑥) < 𝑔(𝑥), the upper and lower limits could still be equal.
Holder and Lipschitz functions are continuous functions.
Congruences, similarities and affine transformations on ℝ𝑛 are examples of
homeomorphisms.
Theorem Let 𝑋 ⊆ ℝ𝑛 and 𝑌 ⊆ ℝ𝑚 . If 𝑓: 𝑋 → 𝑌 is a continuous function and 𝐴 is an open
subset of 𝑌, then 𝑓 −1 [𝐴] is an open subset of 𝑋.
Theorem Let 𝑓: 𝑋 → 𝑌 be continuous on 𝑋 and let 𝐴 be a compact subset of 𝑋. Then 𝑓[𝐴] is
also compact.
Mean Value Theorem Let the function 𝑓: ℝ → ℝ be continuous on the closed, finite interval
[𝑎, 𝑏] and differentiable on the open interval (𝑎, 𝑏). Then there exists a point 𝑐 in the open
interval (𝑎, 𝑏) such that
𝑓(𝑏) − 𝑓(𝑎)
𝑓 ′ (𝑐) = .
𝑏−𝑎
Geometrically this states the there exists a point 𝑐 ∈ (𝑎, 𝑏) for which the tangent to the graph
of the function 𝑓 at (𝑐, 𝑓(𝑐)) is parallel to the chord determined by the points (𝑎, 𝑓(𝑎)) and
(𝑏, 𝑓(𝑏)).
Theorem Let {𝑓𝑛 } be a sequence of functions defined on a common domain 𝐷 ⊂ ℝ. If {𝑓𝑛 }
converges pointwise to 𝑓 on 𝐷, then {𝑓𝑛 } converges uniformly to 𝑓 on 𝐷.
Theorem Let {𝑓𝑛 } be a sequence of functions defined on a common domain 𝐷 ⊂ ℝ, and let
𝑥0 ∈ 𝐷. If the sequence {𝑓𝑛 } converges uniformly to some function 𝑓 on 𝐷, and if each of the
functions 𝑓𝑛 is continuous at 𝑥0 , then the function 𝑓 is continuous at 𝑥0 . In particular, if each of
the functions 𝑓𝑛 is continuous on 𝐷, then 𝑓 is continuous on 𝐷.

Measures and Mass Distributions

Measure For the purposes of this course, a measure is defined as follows.
Let ℱ be a collection of subsets of ℝ𝑛 . A function 𝜇: ℱ → [0, ∞] is called a measure on ℝ𝑛 if,
and only if:
(i) 𝜇(∅) = 0;
 Fractal Geometry - Notes
(ii) if 𝐴 ⊆ 𝐵, then 𝜇(𝐴) ≤ 𝜇(𝐵);
(iii) if 𝐴𝑛 ∈ ℱ for all 𝑛 = 1,2, then 𝜇 is countably sub-additive, i.e.
∞ ∞

𝜇 (⋃ 𝐴𝑛 ) ≤ ∑ 𝜇(𝐴𝑛 )
𝑛=1 𝑛=1

with equality
∞ ∞

𝜇 (⋃ 𝐴𝑛 ) = ∑ 𝜇(𝐴𝑛 )
𝑛=1 𝑛=1

if the 𝐴𝑖 are disjoint Borel sets.

Technically, this is the definition of an outer measure on ℝ𝑛 for which the Borel sets are
measurable.
We call 𝜇(𝐴) the measure of 𝐴 and think of 𝜇(𝐴) as the size of 𝐴 measured in some way.
A function 𝜇 is a measure on a set 𝐴 if 𝐴 contains the support of 𝜇.
Mass distribution A measure 𝜇 on a bounded subset of ℝ𝑛 for which 0 < 𝜇(ℝ𝑛 ) < ∞ is
called a mass distribution, and we think of 𝜇(𝐴) as the mass of the set 𝐴.
Thinking intuitively, if we take a finite mass and spread it in some way across a set 𝑋 to get a
mass distribution on 𝑋, then the conditions for a measure will be satisfied.
Theorem If 𝐴 and 𝐵 are Borel sets such that 𝐵 ⊆ 𝐴 and 𝜇(𝐴) is finite, then
𝜇(𝐴 ∖ 𝐵) = 𝜇(𝐴) − 𝜇(𝐵).
Theorem if 𝐴1 ⊂ 𝐴2 ⊂ 𝐴3 ⊂ ⋯ is an increasing sequence of Borel sets, then
∞

𝜇 (⋃ 𝐴𝑛 ) = lim 𝜇(𝐴𝑛 ).
𝑛→∞
𝑛

A simple extension of this is that if, for 𝛿 > 0, 𝐴δ are Borel sets that are increasing as 𝛿
decreases, that is 𝐴𝛿′ ⊂ 𝐴δ for 0 < 𝛿 < 𝛿 ′ , then

𝜇 (⋃ 𝐴𝑛 ) = lim 𝜇(𝐴𝛿 ).
𝛿→0
𝛿>0

The following method is often used to construct a mass distribution on a subset of ℝ𝑛 . It

involves repeated subdivision of a mass between parts of a bounded Borel set 𝐸.
Let ℰ0 consist of the single set 𝐸.
For 𝑘 = 1,2, …, we let ℰ𝑘 be a collection of disjoint Borel subsets of 𝐸 such that each set 𝑈 in ℰ𝑘
is contained in one of the sets in ℰ𝑘−1 and contains a finite number of the sets in ℰ𝑘+1. We
assume that the maximum diameter of the sets tends to 0 as 𝑘 → ∞.
We define a mass distribution on 𝐸 by repeated subdivision. We let 𝜇(𝐸) satisfy 0 < 𝜇(𝐸) <
∞, and we split this mass between the sets 𝑈1 , 𝑈2 , … , 𝑈𝑚 in ℰ1 by defining 𝜇(𝑈𝑖 ) in such a way
that ∑𝑚𝑖=1 𝜇(𝑈𝑖 ) = 𝜇(𝐸). Similarly, we assign masses to the sets of ℰ2 so that, if 𝑈1 , 𝑈2 , … , 𝑈𝑚
are the sets of ℰ2 contained in a set 𝑈 of ℰ1 , then ∑𝑚
𝑖=1 𝜇(𝑈𝑖 ) = 𝜇(𝑈). In general, we assign
masses so that ∑𝑖 𝜇(𝑈𝑖 ) = 𝜇(𝑈) for each set 𝑈 of ℰ𝑘 , where the {𝑈𝑖 } are the disjoint sets in
ℰ𝑘+1 contained in 𝑈. For each 𝑘, we let 𝐸𝑘 be the union of the sets in ℰ𝑘 , and we define 𝜇(𝐴) =
0 for all 𝐴 with 𝐴 ∩ 𝐸𝑘 = ∅.
 Fractal Geometry - Notes
Let ℰ denote the collection of sets that belong to ℰ𝑘 for some 𝑘, together with the subsets of
the ℝ𝑛 ∖ 𝐸𝑘 . Then the above procedure defines the mass 𝜇(𝐴) for every set 𝐴 in ℰ.
Proposition 1.7 Let 𝜇 be defined on a collection of sets ℰ as defined above. Then the
definition of 𝜇 may be extended to all subsets of ℝ𝑛 so that 𝜇 becomes a measure. The value of
𝜇(𝐴) is uniquely determined if 𝐴 is a Borel set. The support of 𝜇 is contained in
∞

𝐸∞ = ⋂ 𝐸𝑘 .
𝑘=1
 Fractal Geometry - Notes
Box-counting Dimension
A dimension indicates, in some way, how much space a set occupies near to each of its points.
Fundamental to most definitions of dimension is the idea of ‘measurement of a set at scale 𝛿’.
For each 𝛿 we measure a set in a way that detects irregularities of size 𝛿, and we see how
these measurements behave as 𝛿 → 0.
The following are desirable properties of a dimension;
Monotonicity - if 𝐸 ⊂ 𝐹, then dim 𝐸 ≤ dim 𝐹;
Range of values - if 𝐹 ⊂ ℝ𝑛 , then 0 ≤ dim 𝐹 ≤ 𝑛;
Stability - dim(𝐸 ∪ 𝐹) = max(dim 𝐸 , dim 𝐹);
Countable stability - dim(⋃∞
𝑖=1 𝐹𝑖 ) = sup𝑖 (dim 𝐹𝑖 );

Lipschitz invariance - dim 𝑓[𝐹] = dim 𝐹 if 𝑓 is a bi-Lipschitz transformation;

Geometric invariance - dim 𝑓[𝐹] = dim 𝐹 if 𝑓 is a transformation of ℝ𝑛 such as a
translation, rotation, similarity, or affinity;
Countable sets - dim 𝐹 = 0 if 𝐹 is finite or countable;
Open sets - dim 𝐹 = 𝑛 if 𝐹 is an open subset of ℝ𝑛 ;
Smooth manifolds - dim 𝐹 = 𝑚 if 𝐹 is a smooth 𝑚-dimensional manifold.
Similarity dimension A set made up of 𝑚 copies of itself scaled by a factor 𝑟 has a
similarity dimension of
log 𝑚
− .
log 𝑟

Divider Dimension
Divider dimension Let 𝐶 be a curve and let 𝛿 > 0. Furthermore, let 𝑀𝛿 (𝐶) be the maximum
number of points 𝑥0 , 𝑥1 , … , 𝑥𝑚 on the curve, in that order, such that |𝑥𝑘 − 𝑥𝑘−1 | = 𝛿 for 𝑘 =
1,2, … , 𝑚. Then, the divider dimension is defined as
log 𝑀𝛿 (𝐶)
lim
𝛿→0 − log 𝛿

assuming that the limit exists (otherwise we may define upper and lower divider dimensions
using upper and lower limits).
The quantity (𝑀𝛿 (𝐶) − 1)𝛿 may be thought of as the ‘length’ of the curve 𝐶 measured using a
pair of dividers with points set at a distance 𝛿 apart. Hence the name ‘divider dimension’.
The divider dimension of a line is one.
The divider dimension can be estimated as the gradient of a log-log graph plotted over a
suitable range of values of 𝛿.
 Fractal Geometry - Notes

Box-counting Dimension
Given a subset 𝐹 of ℝ𝑛 , for each 𝛿 > 0 we find the smallest number of sets of diameter at most
𝛿 that can cover the set 𝐹 and call this number 𝑁𝛿 (𝐹), indicating the number of ‘clumps’ of
size about 𝛿 into which 𝐹 may be divided. The dimension of 𝐹 reflects the way in which 𝑁𝛿 (𝐹)
grows as 𝛿 → 0. If 𝑁𝛿 (𝐹) obeys, at least approximately, a power law
𝑁𝛿 (𝐹) ≃ 𝑐𝛿 −𝑠
for positive constants 𝑐 and 𝑠, we say that 𝐹 has box dimension 𝑠.
Taking logarithms gives
log 𝑁𝛿 (𝐹) log 𝑐
log 𝑁𝛿 (𝐹) ≃ log 𝑐 − log 𝛿 ⇒ 𝑠≃ + .
− log 𝛿 log 𝛿
We hope to obtain 𝑠 as
log 𝑁𝛿 (𝐹)
𝑠 = lim
𝛿→0 − log 𝛿

with the second term disappearing in the limit.

Box-counting dimension The upper and lower box-counting dimensions of a non-empty
bounded subset 𝐹 of ℝ𝑛 are given by
log 𝑁𝛿 (𝐹)
dimB 𝐹 = lim
𝛿→0 − log 𝛿

and
log 𝑁𝛿 (𝐹)
dimB 𝐹 = lim
𝛿→0 − log 𝛿

and the box-counting dimension of 𝐹 by

log 𝑁𝛿 (𝐹)
dimB 𝐹 = lim
𝛿→0 − log 𝛿

(if this limit exists), where 𝑁𝛿 (𝐹) is any of the following

(i) the smallest number of sets of diameter at most 𝛿 that cover 𝐹;
(ii) the smallest number of closed balls of radius 𝛿 that cover 𝐹;
(iii) the smallest number of cubes of side 𝛿 that cover 𝐹;
(iv) the number of 𝛿-mesh cubes that intersect 𝐹;
 Fractal Geometry - Notes
(v) the largest number of disjoint balls of radius 𝛿 with centres in 𝐹.
Theorem If dimB 𝐹 = dimB 𝐹 = 𝑑, then dimB 𝐹 = 𝑑, otherwise dimB 𝐹 does not exist.
Typically, box-counting dimension calculations involve finding a lower bound and an upper
bound for 𝑁𝛿 (𝐹) separately, each bound depending on a geometric observation, followed by
an analytic stage involving taking logarithms and passing to the limit.
Any covering of 𝐹 by sets of diameter at most 𝛿, closed balls of radius 𝛿, or cubes of side 𝛿
gives an upper bound for 𝑁𝛿 (𝐹). Whereas, any collection of disjoint balls of radius 𝛿 with
centres in 𝐹 gives a lower bound for 𝑁𝛿 (𝐹).
Box-counting dimensions are often much easier to calculate than other forms of dimension,
and the box-counting version of the definition (i.e. (iii)) is convenient for empirical estimation.
Moreover, the interplay between box dimensions and other definitions of dimension can be
used to powerful effect.
Theorem If 𝛿 tends to zero through any decreasing sequence 𝛿𝑘 such that 𝛿𝑘+1 ≥ 𝑐𝛿𝑘 for
some constant 0 < 𝑐 < 1, in particular for 𝛿𝑘 = 𝑐 𝑘 , then
log 𝑁𝛿𝑘 (𝐹)
dimB 𝐹 = lim
𝑘→∞ − log 𝛿𝑘

and
log 𝑁𝛿𝑘 (𝐹)
dimB 𝐹 = lim .
𝑘→∞ − log 𝛿𝑘

Theorem (Monotonicity) Let 𝐸, 𝐹 ⊂ ℝ𝑛 . If 𝐸 ⊂ 𝐹, then dimB 𝐸 ≤ dimB 𝐹 and dimB 𝐸 ≤

dimB 𝐹.
Theorem (Range) Suppose that 𝐹 is a non-empty subset of ℝ𝑛 , then
0 ≤ dimB 𝐹 ≤ dimB 𝐹 ≤ 𝑛.
Hence, to show that dimB 𝐹 = 𝑑, it is enough to show that 𝑑 ≤ dimB 𝐹 and dimB 𝐹 ≤ 𝑑.
Theorem (Finite Stability) Let 𝐸, 𝐹 ⊂ ℝ𝑛 , then
dimB (𝐸 ∪ 𝐹) = max{dimB 𝐸 , dimB 𝐹}.
That is, dimB is finitely stable.
Theorem (Open sets) If 𝐹 ⊂ ℝ𝑛 is an open set, then dimB 𝐹 = 𝑛.
Theorem (Finite sets) If 𝐹 ⊂ ℝ𝑛 is non-empty and finite, then dimB 𝐹 = 0.
Proposition 2.5 If 𝐹 ⊂ ℝ𝑛 and 𝑓: 𝐹 → ℝ𝑚 is a Lipschitz transformation, that is
|𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐|𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐹),
then dimB 𝑓[𝐹] ≤ dimB 𝐹 and dimB 𝑓[𝐹] ≤ dimB 𝐹.
Furthermore, if 𝑓: 𝐹 → ℝ𝑚 is a bi-Lipschitz transformation, that is
𝑐1 |𝑥 − 𝑦| ≤ |𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐2 |𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐹)
where 0 < 𝑐1 ≤ 𝑐2 < ∞, then dimB 𝑓[𝐹] = dimB 𝐹 and dimB 𝑓[𝐹] = dimB 𝐹.
Corollary 2.5.1 (Geometric invariance) Let 𝑓: 𝐹 → ℝ𝑚 be a congruence, similarity or affine
transformation, then dimB 𝑓[𝐹] = dimB 𝐹 and dimB 𝑓[𝐹] = dimB 𝐹.
 Fractal Geometry - Notes
Corollary 2.5.2 (Smooth sets) If 𝐹 is a smooth (i.e. continuously differentiable) bounded 𝑚-
dimensional sub-manifold (i.e. 𝑚-dimensional surface) of ℝ𝑛 , then dimB 𝐹 = 𝑚. In particular,
smooth curves have dimension 1 and smooth surfaces have dimension 2.
Corollary 2.5.3 (Smooth curves) Let 𝑔: [0,1] → ℝ be Lipschitz. Then dimB (graph 𝑔) = 1,
where graph 𝑔 = {(𝑥, 𝑔(𝑥)): 𝑥 ∈ [0,1]} is the graph of the function 𝑔.
Corollary 2.5.3 (Projections) Let proj denote orthogonal projection from ℝ2 onto some
given line through the origin. Then, for 𝐹 ⊂ ℝ2 , dimB (proj 𝐹) ≤ min{1, dimB 𝐹} and
dimB (proj 𝐹) ≤ min{1, dimB 𝐹}.
Theorem (Generalisation of Proposition 2.5) If 𝐹 ⊂ ℝ𝑛 and 𝑓: 𝐹 → ℝ𝑚 is a mapping that
satisfies a Holder condition, that is
|𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐|𝑥 − 𝑦|𝛼 (𝑥, 𝑦 ∈ 𝐹)
for constants 0 < 𝛼 ≤ 1 and 𝑐 > 0. Then
1 1
dimB 𝑓[𝐹] ≤ dimB 𝐹 and dimB 𝑓[𝐹] ≤ dimB 𝐹 .
𝛼 𝛼
Furthermore
1
dimB 𝑓[𝐹] ≤ dimB 𝐹
𝛼
if dimB 𝐹 is well defined.
Proposition 2.6 Let 𝐹 denote the closure of 𝐹 ⊂ ℝ𝑛 (i.e. the smallest closed subset of ℝ𝑛
containing 𝐹). Then dimB 𝐹 = dimB 𝐹 and dimB 𝐹 = dimB 𝐹.
Corollary 2.6.1 If 𝐹 is a dense subset of an open region of ℝ𝑛 , then dimB 𝐹 = dimB 𝐹 = 𝑛.
As a consequence of Proposition 2.6
(i) countable sets may have non-zero box dimension; and
(ii) it is not generally true that
∞

dimB (⋃ 𝐹𝑖 ) = sup𝑖 {dimB 𝐹𝑖 }.

𝑖=1

These features are not desirable for a dimension, and severely limit the usefulness of box
dimension.
 Fractal Geometry - Notes
Hausdorff Measure and Dimension
Hausdorff measure Suppose that 𝐹 is a subset of ℝ𝑛 and 𝑠 ≥ 0 is a real number. For 𝛿 > 0
define
∞

ℋ𝛿𝑠 (𝐹) = inf {∑|𝑈𝑖 |𝑠 : {𝑈𝑖 } is a 𝛿-cover of 𝐹}.

𝑖=1

That is, ℋ𝛿𝑠 (𝐹) is

the greatest lower bound of the sum of the 𝑠th powers of the diameters of all
covers of 𝐹 by sets of diameter at most 𝛿. Then (since ℋ𝛿𝑠 (𝐹) is non-decreasing as 𝛿 → 0)
ℋ 𝑠 (𝐹) = lim ℋ𝛿𝑠 (𝐹) ≥ ℋ𝛿𝑠 (𝐹) (0 ≤ ℋ 𝑠 ≤ ∞).
𝛿→0

exists for all 𝐹, and is called the 𝑠-dimensional Hausdorff measure of 𝐹.

Thus, we can evaluate ℋ 𝑠 (𝐹) by finding
lim ℋ𝛿𝑠𝑛 (𝐹)
𝑛→∞

for any sequence 𝛿𝑛 such that 𝛿𝑛 → 0 as 𝑛 → ∞.

By definition
∞

ℋ𝛿𝑠 (𝐹) ≤ ∑|𝑈𝑖 |𝑠

𝑖=1

for all 𝛿-covers {𝑈𝑖 } of 𝐹. So, given an arbitrary 𝜀 > 0, we can find a 𝛿-cover of 𝐹 such that
∞

∑|𝑈𝑖 |𝑠 ≤ ℋ𝛿𝑠 (𝐹) + 𝜀.

𝑖=1

This is an important fact that is often used in proofs.

Theorem The 𝑠-dimensional Hausdorff measure ℋ 𝑠 (𝐹) of a subset 𝐹 of ℝ𝑛 is a measure on
ℝ𝑛 . That is
(i) ℋ 𝑠 (∅) = 0;
(ii) if 𝐸 ⊆ 𝐹, then ℋ 𝑠 (𝐸) ≤ ℋ 𝑠 (𝐹);
(iii) if 𝐹𝑖 ∈ ℝ𝑛 for all 𝑖 = 1,2, then ℋ 𝑠 is countably sub-additive, i.e.
∞ ∞

ℋ (⋃ 𝐹𝑖 ) ≤ ∑ ℋ 𝑠 (𝐹𝑖 )
𝑠

𝑖=1 𝑖=1

with equality
∞ ∞

ℋ 𝑠 (⋃ 𝐹𝑖 ) = ∑ ℋ 𝑠 (𝐹𝑖 )
𝑖=1 𝑛=1

if the 𝐹𝑖 are disjoint Borel sets.

Hausdorff measures generalise the familiar notion of length, area, volume, and so on.
Some simple examples of Hausdorff measure:
(i) for any point 𝑥 ∈ ℝ𝑛
 Fractal Geometry - Notes
1 if 𝑠 = 0
ℋ 𝑠 ({𝑥}) = {
0 if 𝑠 > 0
(ii) If 𝐴 is a countable subset of ℝ𝑛 , then ℋ 𝑠 (𝐴) = 0 unless 𝑠 = 0;
(iii) If 𝐹 = [0,1] then
∞ if 0 ≤ 𝑠 < 1
ℋ 𝑠 (𝐹) = { 1 if 𝑠 = 1
0 if 𝑠 > 1
Theorem If 𝐹 is a Borel subset of ℝ𝑛 , then
ℋ 𝑛 (𝐹) = 𝑐𝑛−1 vol𝑛 (𝐹),
where 𝑐𝑛 is the volume of an 𝑛-dimensional ball of radius 1. That is
𝜋 𝑛⁄2
𝑛 if 𝑛 is even
2𝑛 (2) !
𝑐𝑛 =
𝑛−1
(𝑛−1)⁄2( )!
𝜋 2
{ if 𝑛 is odd.
𝑛!
Hence, for the lower dimensional subsets of ℝ𝑛 we have:
 ℋ 0 (𝐹) is the number of points in 𝐹;
 ℋ 1 (𝐹) gives the length of a smooth curve 𝐹;
 ℋ 2 (𝐹) = (4⁄𝜋) × area(𝐹) if 𝐹 is a smooth surface;
 ℋ 3 (𝐹) = (6⁄𝜋) × vol(𝐹).
Proposition 3.1 If 𝐹 ⊂ ℝ𝑛 and 𝑓: 𝐹 → ℝ𝑚 is a mapping that satisfies a Holder condition, that
is
|𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐|𝑥 − 𝑦|𝛼 (𝑥, 𝑦 ∈ 𝐹)
for constants 0 < 𝛼 ≤ 1 and 𝑐 > 0. Then for each 𝑠
ℋ 𝑠⁄𝛼 (𝑓[𝐹]) ≤ 𝑐 𝑠⁄𝛼 ℋ 𝑠 (𝐹).
In particular, if 𝑓 is a Lipschitz mapping, i.e. if 𝛼 = 1, then
ℋ 𝑠 (𝑓[𝐹]) ≤ 𝑐 𝑠 ℋ 𝑠 (𝐹).
Corollary 3.1.1 (scaling property) Let 𝐹 ⊂ ℝ𝑛 and let 𝑓: 𝐹 → ℝ𝑛 be a similarity
transformation of scale factor 𝜆 > 0. Then
ℋ 𝑠 (𝑓[𝐹]) = 𝜆𝑠 ℋ 𝑠 (𝐹).
In particular, if 𝑓 is a congruence or isometry, that is
|𝑓(𝑥) − 𝑓(𝑦)| = |𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐹).
Then ℋ 𝑠 (𝑓[𝐹]) = ℋ 𝑠 (𝐹). Thus, Hausdorff measures are translation invariant, and rotation
invariant.
This scaling property is a fundamental property of Hausdorff measures.
 Fractal Geometry - Notes
Hausdorff Dimension
For any set 𝐹 ⊂ ℝ and 𝛿 < 1, it should be clear from the definition that ℋ𝛿𝑠 (𝐹) is non-
𝑛

increasing with 𝑠, so ℋ 𝑠 (𝐹) is also non-increasing. Furthermore, if 𝑡 > 𝑠 and {𝑈𝑖 } is a 𝛿-cover
of 𝐹, then
∞ ∞ ∞

∑|𝑈𝑖 |𝑡 ≤ ∑|𝑈𝑖 |𝑡−𝑠 |𝑈𝑖 |𝑠 ≤ 𝛿 𝑡−𝑠 ∑|𝑈𝑖 |𝑠 .

𝑖=1 𝑖=1 𝑖=1

Taking infima over all 𝛿-covers gives

ℋ𝛿𝑡 (𝐹) ≤ 𝛿 𝑡−𝑠 ℋ𝛿𝑠 (𝐹).
Taking limits as 𝛿 → 0 gives
ℋ 𝑡 (𝐹) ≤ (lim 𝛿 𝑡−𝑠 ) ℋ 𝑠 (𝐹).
𝛿→0

Therefore, for 𝑡 > 𝑠

 if ℋ 𝑠 (𝐹) < ∞ then ℋ 𝑡 (𝐹) = 0; and
 if ℋ 𝑡 (𝐹) > 0 then ℋ 𝑠 (𝐹) = ∞.
Thus, a graph of ℋ 𝑠 (𝐹) against 𝑠 shows that there is a critical value at which ℋ 𝑠 (𝐹) jumps
from ∞ to zero (since ℋ 𝑠 (𝐹) is non-increasing with 𝑠). This critical value is called the
Hausdorff dimension of 𝐹, written as dimH 𝐹.
ℋ 𝑠 (𝐹)
∞

0 𝑠
0 dimH 𝐹 𝑛

Hausdorff dimension Let 𝐹 ⊂ ℝ𝑛 , then the Hausdorff dimension is given by

dimH 𝐹 = inf{𝑠 ≥ 0: ℋ 𝑠 (𝐹) = 0}
= sup{𝑠: ℋ 𝑠 (𝐹) = ∞}
taking the supremum of the empty set to be zero, so that
∞ if 0 ≤ 𝑠 < dimH 𝐹
ℋ 𝑠 (𝐹) = {
0 if 𝑠 > dimH 𝐹 .
If 𝑠 = dimH 𝐹, then ℋ 𝑠 (𝐹) may be zero or infinite, or may satisfy 0 < ℋ 𝑠 (𝐹) < ∞.
A Borel set satisfying this condition is called an 𝑠-set.
It is possible to define Hausdorff dimension without directly using Hausdorff measures.
Hausdorff dimension II For 𝐹 ⊂ ℝ𝑛 , let
∞

ℋ∞𝑠 (𝐹) = inf {∑|𝑈𝑖 |𝑠 : {𝑈𝑖 } is a countable cover of 𝐹}.

𝑖=1
 Fractal Geometry - Notes
Then the Hausdorff dimension of 𝐹 is given by
dimH 𝐹 = inf{𝑠 ≥ 0: ℋ∞𝑠 (𝐹) = 0}.
Theorem (Monotonicity) Let 𝐸, 𝐹 ⊂ ℝ𝑛 . If 𝐸 ⊂ 𝐹, then dimH 𝐸 ≤ dimH 𝐹.
Theorem (Range) Suppose that 𝐹 is a non-empty subset of ℝ𝑛 , then
0 ≤ dimB 𝐹 ≤ 𝑛.
Theorem (Countable Stability) Let 𝐹1 , 𝐹2 , … be a countable sequence of sets in ℝ𝑛 . Then
∞

dimH ⋃ 𝐹𝑖 = sup {dimH 𝐹𝑖 }.

1≤𝑖≤∞
𝑖=1

Theorem (Countable sets) If 𝐹 ⊂ ℝ𝑛 is a countable set, then dimH 𝐹 = 0.

Theorem (Open sets) If 𝐹 ⊂ ℝ𝑛 is an open set, then dimH 𝐹 = 𝑛.
Proposition 3.3 If 𝐹 ⊂ ℝ𝑛 and 𝑓: 𝐹 → ℝ𝑚 is a mapping that satisfies a Holder condition, that
is
|𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐|𝑥 − 𝑦|𝛼 (𝑥, 𝑦 ∈ 𝐹)
for constants 0 < 𝛼 ≤ 1 and 𝑐 > 0. Then
1
dimH 𝑓[𝐹] ≤ dimH 𝐹 .
𝛼
Furthermore, if 𝑓: 𝐹 → ℝ𝑚 is a bi-Lipschitz transformation, that is
𝑐1 |𝑥 − 𝑦| ≤ |𝑓(𝑥) − 𝑓(𝑦)| ≤ 𝑐2 |𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐹)
where 0 < 𝑐1 ≤ 𝑐2 < ∞, then dimH 𝑓[𝐹] = dimH 𝐹.
Corollary 3.3.1 (Geometric invariance) Let 𝑓: 𝐹 → ℝ𝑚 be a congruence, similarity or affine
transformation. Then dimH 𝑓[𝐹] = dimH 𝐹.
Corollary 3.3.2 (Smooth sets) If 𝐹 is a smooth (i.e. continuously differentiable) bounded 𝑚-
dimensional sub-manifold (i.e. 𝑚-dimensional surface) of ℝ𝑛 , then dimH 𝐹 = 𝑚.
Corollary 3.3.3 (Projections) Let proj denote orthogonal projection from ℝ2 onto some
given line through the origin. Then, for 𝐹 ⊂ ℝ2 , dimH (proj 𝐹) ≤ min{1, dimH 𝐹}.
The Hausdorff dimension can be seen as a refinement of the definition of box dimension.
Assuming that dimB 𝐹 exists, then
log 𝑁𝛿 (𝐹)
dimB 𝐹 = lim
𝛿→0 − log 𝛿

where 𝑁𝛿 (𝐹) is the smallest number of sets that form a 𝛿-cover of 𝐹.

Hence

𝑁𝛿 (𝐹)𝛿 𝑠 = inf {∑ 𝛿 𝑠 : {𝑈𝑖 }𝑖∈𝐼 is a 𝛿-cover of 𝐹}.

𝑖∈𝐼

If 𝑡 > 𝑠, then
𝑁𝛿 (𝐹)𝛿 𝑡 = 𝛿 𝑡−𝑠 𝑁𝛿 (𝐹)𝛿 𝑠 .
Taking limits as 𝛿 → 0 gives
 Fractal Geometry - Notes
lim 𝑁𝛿 (𝐹)𝛿 𝑡 = (lim 𝛿 𝑡−𝑠 ) (lim 𝑁𝛿 (𝐹)𝛿 𝑠 ).
𝛿→0 𝛿→0 𝛿→0

Therefore, for 𝑡 > 𝑠

 if lim 𝑁𝛿 (𝐹)𝛿 𝑠 < ∞ then lim 𝑁𝛿 (𝐹)𝛿 𝑡 = 0; and
𝛿→0 𝛿→0

 if lim 𝑁𝛿 (𝐹)𝛿 𝑡 > 0 then lim 𝑁𝛿 (𝐹)𝛿 𝑠 = ∞.

𝛿→0 𝛿→0

The value of 𝑠 at which lim 𝑁𝛿 (𝐹)𝛿 𝑠 jumps from ∞ to 0 is equal to the box dimension of 𝐹.
𝛿→0

If lim 𝑁𝛿 (𝐹)𝛿 𝑠 = ∞, then taking logs gives

𝛿→0

lim [log(𝑁𝛿 (𝐹)𝛿 𝑠 )] = lim [log(𝑁𝛿 (𝐹)) + 𝑠 log 𝛿] = ∞

𝛿→0 𝛿→0

It follows that there exists a 𝛿0 < 1 such that, for 0 < 𝛿 < 𝛿0
log(𝑁𝛿 (𝐹)) + 𝑠 log 𝛿 log(𝑁𝛿 (𝐹)) 𝑠 log 𝛿 log(𝑁𝛿 (𝐹))
= + >0 ⇒ > 𝑠.
− log 𝛿 − log 𝛿 − log 𝛿 − log 𝛿
Hence
log(𝑁𝛿 (𝐹))
dimB 𝐹 = lim ≥ 𝑠.
𝛿→0 − log 𝛿
Furthermore, if lim 𝑁𝛿 (𝐹)𝛿 𝑠 = 0, then taking logs gives
𝛿→0

lim [log(𝑁𝛿 (𝐹)𝛿 𝑠 )] = lim [log(𝑁𝛿 (𝐹)) + 𝑠 log 𝛿] = −∞.

𝛿→0 𝛿→0

It follows that there exists a 𝛿1 < 1 such that, for 0 < 𝛿 < 𝛿1
log(𝑁𝛿 (𝐹)) + 𝑠 log 𝛿 log(𝑁𝛿 (𝐹)) 𝑠 log 𝛿 log(𝑁𝛿 (𝐹))
= + <0 ⇒ < 𝑠.
− log 𝛿 − log 𝛿 − log 𝛿 − log 𝛿
Hence
log(𝑁𝛿 (𝐹))
dimB 𝐹 = lim ≤ 𝑠.
𝛿→0 − log 𝛿
This gives a definition of box dimension that looks very similar to the definition of Hausdorff
dimension. For sets that are 'reasonably regular' these two dimensions turn out to be equal. If,
however, a set can be covered more efficiently by sets of very different sizes (rather than by
sets of exactly the same size), then the expressions ℋ𝛿𝑡 (𝐹) and 𝑁𝛿 (𝐹)𝛿 𝑠 can be very different,
which leads to differing values for the Hausdorff and box dimensions.
Proposition 3.4 For every non-empty bounded 𝐹 ⊂ ℝ𝑛
dimH 𝐹 ≤ dimB 𝐹 ≤ dimB 𝐹.
A fundamental property of dimension is that Hausdorff, lower box and upper box dimensions
are invariant under bi-Lipschitz transformations. Thus, if two sets have different dimensions
there cannot be a Lipschitz mapping from one onto the other. Therefore, one approach to
fractal geometry is to regard two sets as equivalent if there is a bi-Lipschitz mapping between
them.
Proposition 3.5 Every set 𝐹 ⊂ ℝ𝑛 with dimH 𝐹 < 1 is totally disconnected.
 Fractal Geometry - Notes
Calculation of Hausdorff Dimension
Typically, Hausdorff dimension calculations involve finding an upper, and a lower estimate for
dimH 𝐹, which will hopefully give the same values. Each of these estimates usually involves a
geometric observation followed by a calculation.
Calculation of Hausdorff measures and dimension can be rather complicated, even for fairly
simple fractal sets. For the upper estimate, there are often natural coverings that can be used.
It is the lower estimate that is usually more difficult since, according to the definition of
Hausdorff measure, all possible 𝛿-covers have to be considered to obtain the infimum value.
There is a simple 'heuristic' method that often gives the correct result when considering the
Hausdorff dimension of a self-similar fractal set. If 𝐹 = ⋃∞ 𝑖=1 𝐹𝑖 , where each 𝐹𝑖 is geometrically
similar to 𝐹, but scaled by a factor 𝑟𝑖 , then, provided that the 𝐹𝑖 are disjoint Borel sets, we can
write
∞ ∞ ∞

ℋ 𝑠 (𝐹) = ℋ 𝑠 (⋃ 𝐹𝑖 ) = ∑ ℋ 𝑠 (𝐹𝑖 ) = ∑ 𝑟𝑖𝑠 ℋ 𝑠 (𝐹)

𝑖=1 𝑖=1 𝑖=1

by the scaling property of Hausdorff measures.

Assuming that at the critical value 𝑠 = dimH 𝐹 we have 0 < ℋ 𝑠 (𝐹) < ∞, we may divide by
ℋ 𝑠 (𝐹) to get
∞

∑ 𝑟𝑖𝑠 = 1,
𝑖=1

which we may solve for 𝑠.

Equivalent Definitions
The definition of Hausdorff measure involves coverings by arbitrary sets of diameter at most
𝛿. However, there are some restricted classes of covering sets that lead to the same values of
Hausdorff dimension.
The following covering sets give the same values of Hausdorff dimension:
 arbitrary sets of diameter at most 𝛿;
 balls of diameter at most 𝛿;
 open, or closed sets of diameter at most 𝛿.
Binary interval A binary interval is an interval of the form [𝑟2−𝑘 , (𝑟 + 1)2−𝑘 ] where 𝑘 =
0,1,2, … and 𝑟 = 0,1, … , 2𝑘 − 1.
Net measures Let 𝐹 be a subset of the interval [0,1]. Define
∞

ℳ𝛿𝑠 (𝐹) = inf {∑|𝑈𝑖 |𝑠 : {𝑈𝑖 } is a 𝛿-cover of 𝐹 by binary intervals}.

𝑖=1

Then
ℳ 𝑠 (𝐹) = lim ℳ𝛿𝑠 (𝐹)
𝛿→0

is called the net measure of 𝐹.

For certain purposes, net measures are much more convenient that Hausdorff measures. This
is because two binary intervals are either disjoint, or one of them is contained in the other,
 Fractal Geometry - Notes
allowing any cover of a set by binary intervals to be reduced to a cover by disjoint binary
intervals.
 Fractal Geometry - Notes
Techniques for Calculating Dimensions
Calculating Hausdorff dimension (and indeed any other form of dimension) directly from its
definition can be quite complicated and laborious.

Basic Methods
It is often relatively simple to obtain an upper bound for the dimension of a set. However, it is
usually much more difficult to find a lower bound for the dimension of a set.
To obtain an upper bound for the Hausdorff dimension of a set 𝐹 it is enough to evaluate sums
of the form ∑|𝑈𝑖 |𝑠 for specific coverings of 𝐹.
Proposition 4.1 Suppose that a set 𝐹can be covered by 𝑛𝑘 sets of diameter at most 𝛿𝑘 with
𝛿𝑘 → 0 as 𝑘 → ∞. Then
log 𝑛𝑘
dimH 𝐹 ≤ dimB 𝐹 ≤ lim .
𝑘→∞ − log 𝛿𝑘

Moreover, if 𝑛𝑘 𝛿𝑘𝑠 remains bounded as 𝑘 → ∞, then ℋ 𝑠 (𝐹) < ∞. If 𝛿𝑘 → 0 but 𝛿𝑘+1 ≥ 𝑐𝛿𝑘 for
some 0 < 𝑐 < 1, then
log 𝑛𝑘
dimB 𝐹 ≤ lim .
𝑘→∞ − log 𝛿𝑘

To obtain a lower bound for the Hausdorff dimension of a set 𝐹 we must show that sums of
the form ∑|𝑈𝑖 |𝑠 are greater than some positive constant for all 𝛿-coverings of 𝐹.
Proposition 4.2 (mass distribution principle) Let 𝜇 be a mass distribution on a set 𝐹, and
suppose that for some 𝑠 > 0 there exist numbers 𝑐 > 0 and 𝜀 > 0 such that
𝜇(𝑈) ≤ 𝑐|𝑈|𝑠
for all sets 𝑈 with |𝑈| ≤ 𝜀. Then
𝜇(𝐹)
ℋ 𝑠 (𝐹) ≥
𝑐
and
𝑠 ≤ dimH 𝐹 ≤ dimB 𝐹 ≤ dimB 𝐹.
The following theorem is useful when working with Proposition 4.1 and the mass distribution
principle.
Theorem Let 𝑎 and 𝑏 be positive real numbers, then
1
𝑎log 𝑏⁄log 𝑎 = 𝑏 and 𝑎log 𝑏⁄− log 𝑎 = .
𝑏

Generalised Cantor Sets

Let [0,1] = 𝐸0 ⊃ 𝐸1 ⊃ 𝐸2 ⊃ ⋯ be a decreasing sequence of sets, with each 𝐸𝑘 a union of a
finite number of disjoint closed intervals, called 𝑘th level basic intervals, with each interval of
𝐸𝑘 containing at least two intervals of 𝐸𝑘+1 which may be of different lengths, and the
maximum length of the 𝑘th level intervals tending to 0 as 𝑘 → ∞. Then the set
∞

𝐹 = ⋂ 𝐸𝑘
𝑘=0
 Fractal Geometry - Notes
is a totally disconnected subset of [0,1] that is generally a fractal.
The above general construction of fractal subsets of ℝ may be thought of as a generalisation of
the middle third Cantor set.
If the intervals in 𝐸𝑘 may have different lengths, it is not possible to use Proposition 4.1 to
obtain an upper bound for dimH 𝐹. The following generalisation of proposition 4.1 may be
used instead.
Generalisation of Proposition 4.1 Suppose that 𝐸𝑘 is a cover of 𝐹 by sets 𝑈𝑖 of diameter at
most 𝛿𝑘 with 𝛿𝑘 → 0 as 𝑘 → ∞. If there exists 0 < 𝑀 < ∞ such that

∑ |𝑈𝑖 |𝑠 ≤ 𝑀
𝑈𝑖 ∈𝐸𝑘

for each 𝑘, then

ℋ𝛿𝑠𝑘 (𝐹) ≤ 𝑀
so that
ℋ 𝑠 (𝐹) = lim ℋ𝛿𝑠𝑘 (𝐹) ≤ 𝑀 < ∞.
𝑘→∞

It follows from the Hausdorff dimension graph that dimH 𝐹 ≤ 𝑠.

The mass distribution principle may be used to find a lower bound for dimH 𝐹.
Non-uniform Cantor sets Let 𝑠 be a number with 0 < 𝑠 < 1. Assume that in the general
construction of a fractal subset of ℝ, for each 𝑘th level interval 𝐼, the (𝑘 + 1)th level intervals
𝐼1 , 𝐼2 , … , 𝐼𝑚 with 𝑚 ≥ 2 contained in 𝐼 are of equal length and equally spaced, the lengths being
given by
1 𝑠
|𝐼𝑖 |𝑠 = |𝐼| (1 ≤ 𝑖 ≤ 𝑚)
𝑚
with the left-hand ends of 𝐼1 and 𝐼 coinciding, and the right-hand ends of 𝐼𝑚 and 𝐼 coinciding.
Then
dimH 𝐹 = 𝑠 and 0 < ℋ 𝑠 (𝐹) < ∞.
Note: 𝑚 is allowed to be different for different intervals 𝐼 in the construction so that the 𝑘th
level intervals may have widely differing lengths. If 𝑚 is kept constant throughout the general
construction for non-uniform Cantor sets, the sets obtained are called uniform Cantor sets.
Uniform Cantor sets Let 𝑚 ≥ 2 be an integer and let 0 < 𝑟 < 1⁄𝑚. Furthermore, let 𝐹 be
the sets obtained by the general construction of a fractal subset of ℝ in which each basic
interval 𝐼 is replaced by 𝑚 equally spaced subintervals of lengths 𝑟|𝐼| where 𝑚 is kept
constant throughout the construction, and the ends of 𝐼 coincide with the ends of the extreme
subintervals. Then
log 𝑚
dimH 𝐹 = dimB 𝐹 = and 0 < ℋ log 𝑚⁄− log 𝑟 (𝐹) < ∞.
− log 𝑟
Proposition (Example 4.6) Suppose that in the general construction of a fractal subset of ℝ
each (𝑘 − 1)th level interval contains at least 𝑚𝑘 ≥ 2 𝑘th level intervals (𝑘 = 1,2,3, … ) which
are separated by gaps of at least 𝜀𝑘 , where 0 < 𝜀𝑘+1 < 𝜀𝑘 for each 𝑘. Then
log(𝑚1 ⋯ 𝑚𝑘−1 )
dimH 𝐹 ≥ lim .
𝑘→∞ − log(𝑚𝑘 𝜀𝑘 )
 Fractal Geometry - Notes
Suppose further that, in addition, the 𝑘th level intervals are all of length 𝛿𝑘 and that each
(𝑘 − 1)th level interval contains exactly 𝑚𝑘 𝑘th level intervals which are roughly evenly
spaced in the sense that 𝑚𝑘 𝜀𝑘 ≥ 𝑐𝛿𝑘−1 for some constant 𝑐 > 0. Then
log(𝑚1 ⋯ 𝑚𝑘−1 )
dimH 𝐹 = lim .
𝑘→∞ − log 𝛿𝑘−1
Proposition (Example 4.7) Fix 0 < 𝑠 < 1 and let 𝑛0 , 𝑛1 , 𝑛2 , … be a rapidly increasing
⁄
sequence of integers, say 𝑛𝑘+1 ≥ max{𝑛𝑘𝑘 , 4𝑛𝑘1 𝑠 } for each 𝑘. For each 𝑘, let 𝐻𝑘 ⊂ ℝ consist of
⁄
equally spaced equal intervals of lengths 𝑛𝑘−1 𝑠 with the mid-points of consecutive intervals a
distance of 𝑛𝑘−1 apart. Let
∞

𝐹 = ⋂ 𝐻𝑘 ,
𝑘=0

then dimH 𝐹 = 𝑠.
Jarnik's theorem Suppose that 𝛼 > 2. Let 𝐹 be the set of real numbers 𝑥 ∈ [0,1] for which
the inequality
‖𝑞𝑥‖ ≤ 𝑞1−𝛼
is satisfied by infinitely many positive integers 𝑞. Then
2
dimH 𝐹 = .
𝛼
It should be noted that ‖𝑞𝑥‖ ≤ 𝑞1−𝛼 if, and only if
𝑛 𝑛
𝑥 ∈ [ − 𝑞 −𝛼 , + 𝑞 −𝛼 ]
𝑞 𝑞
for some integer 𝑛.
 Fractal Geometry - Notes
Fractals Defined by Transformations
Many fractals are made up of parts that are, in some way, similar to the whole. These self-
similarities are not only properties of the fractals: they may actually be used to define them.
Iterated function systems do this in a unified way and, moreover, often lead to a simple way of
finding dimensions.
Iterated function system A finite family of contractions {𝑆1 , 𝑆2 , … , 𝑆𝑚 } with 𝑚 ≥ 2 is called
an iterated function system, or IFS (technically, this is the definition of a hyperbolic, or
contracting IFS).
Invariant set (attractor) Let 𝐷 be a closed subset of ℝ𝑛 . A nonempty compact subset 𝐹 of
𝐷 is an invariant set, or attractor, for the iterated function system {𝑆1 , 𝑆2 , … , 𝑆𝑚 } if, and only if
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹).
𝑖=1

Theorem 9.1 Consider the iterated function system given by the contractions {𝑆1 , 𝑆2 , … , 𝑆𝑚 }
on 𝐷, which is a closed subset of ℝ𝑛 , so that
|𝑆𝑖 (𝑥) − 𝑆𝑖 (𝑦)| ≤ 𝑐𝑖 |𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐷)
with 𝑐𝑖 < 1 for each 𝑖. Then there is a unique attractor 𝐹, i.e. a non-empty compact subset 𝐹 of
𝐷 such that
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹).
𝑖=1

Moreover, if we define a transformation 𝑆 on the class 𝒮of non-empty compact sets by

𝑚

𝑆(𝐸) = ⋃ 𝑆𝑖 (𝐸)
𝑖=1

for 𝐸 ∈ 𝒮, and write 𝑆 𝑘 for the 𝑘th iterate of 𝑆 (so 𝑆 0 = 𝐸 and 𝑆 𝑘 (𝐸) = 𝑆(𝑆 𝑘−1 (𝐸)) for 𝑘 ≥ 1),
then
∞

𝐹 = ⋂ 𝑆 𝑘 (𝐸).
𝑘=0

for every set 𝐸 ∈ 𝒮 such that 𝑆𝑖 (𝐸) ⊂ 𝐸 for all 𝑖.

Theorem 9.1 states that, for any family of contractions {𝑆1 , 𝑆2 , … , 𝑆𝑚 } there is exactly one non-
empty compact set 𝐹 satisfying
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹)
𝑖=1

However, there may be other sets satisfying this equation which are not compact.
There are two main problems that arise in connection with iterated function systems:
(i) to determine an IFS for a given unique attractor;
(ii) to determine a unique attractor for a given IFS.
 Fractal Geometry - Notes
Finding an IFS that has a given 𝐹 as its unique attractor can often be done by inspection, at
least if 𝐹 is self-similar or self-affine. For example, the Cantor set is easily seen to be the
attractor of the two similarities which give the basic self-similarities of the set
1 1 2
𝑆1 (𝑥) = 𝑥 and 𝑆1 (𝑥) = 𝑥 + .
3 3 3
However, there is no general method that can be applied to any 𝐹.
The transformation 𝑆 introduced in Theorem 9.1 is the key to computing the attractor of an
IFS.
The sequence of iterates 𝑆 𝑘 (𝐸) converges to the attractor 𝐹 for any initial set 𝐸 ∈ 𝒮 in the
sense that the Hausdorff metric 𝑑(𝑆 𝑘 (𝐸), 𝐹) → 0. Thus the 𝑆 𝑘 (𝐸) provide increasingly good
approximations to 𝐹. If 𝐹 is a fractal, these approximations may be called pre-fractals for 𝐹.
For each 𝑘, we have

𝑆 𝑘 (𝐸) = ⋃ 𝑆𝑖1 ∘ 𝑆𝑖2 ∘ ⋯ ∘ 𝑆𝑖𝑘 (𝐸) = ⋃ 𝑆𝑖1 (𝑆𝑖2 ⋯ (𝑆𝑖𝑘 (𝐸)))

𝐼𝑘 𝐼𝑘

where the union is over the set 𝐼𝑘 of all 𝑘-term sequences (𝑖1 , 𝑖2 , … , 𝑖𝑘 ). If 𝑆𝑖 (𝐸) is contained in
𝐸 for each 𝑖, and 𝑥 is a point of 𝐹, it follows from the fact that
∞

𝐹 = ⋂ 𝑆 𝑘 (𝐸)
𝑘=0

that there is a (not necessarily unique - see point 10 in the course notes) sequence (𝑖1 , 𝑖2 , … 𝑖𝑘 )
such that
𝑥 ∈ 𝑆𝑖1 ∘ 𝑆𝑖2 ∘ ⋯ ∘ 𝑆𝑖𝑘 (𝐸)
for all 𝑘. This sequence provides a natural coding for 𝑥, with
∞

𝑥 = 𝑥𝑖1 ,𝑖2,… = ⋂ 𝑆𝑖1 ∘ 𝑆𝑖2 ∘ ⋯ ∘ 𝑆𝑖𝑘 (𝐸)

𝑘=1

so that 𝐹 = ⋃{𝑥𝑖1 ,𝑖2 ,… }. This expression for 𝑥𝑖1 ,𝑖2 ,… is independent of 𝐸 provided that 𝑆𝑖 (𝐸) is
contained in 𝐸 for all 𝑖.
For example, the Cantor set has the following IFS
1 1 2
𝑆1 (𝑥) = 𝑥 and 𝑆1 (𝑥) = 𝑥 + .
3 3 3
Now, if 𝑥𝑖1 ,𝑖2 ,… is a point in the Cantor set with base-3 expansion 0. 𝑎1 𝑎2 ⋯, then
𝑎𝑘 = 0 if 𝑖𝑘 = 1
and
𝑎𝑘 = 2 if 𝑖𝑘 = 2.
Note that, if the union
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹)
𝑖=1

is disjoint, then 𝐹 must be totally disconnected (provided that the 𝑆𝑖 are injections).
 Fractal Geometry - Notes
The pre-fractals 𝑆 𝑘 (𝐸) provide the usual construction of many fractals for a suitably chosen
initial set 𝐸, in which case, the 𝑆𝑖1 ∘ 𝑆𝑖2 ∘ ⋯ ∘ 𝑆𝑖𝑘 (𝐸) are called the 𝑘-level sets of the
construction.

Dimensions of Self-similar Sets

Self-similar set The attractor of an IFS {𝑆1 , 𝑆2 , … , 𝑆𝑚 } where the mappings 𝑆𝑖 are
contracting similarities, that is,
|𝑆𝑖 (𝑥) − 𝑆𝑖 (𝑦)| = 𝑐|𝑥 − 𝑦|
with 0 < 𝑐 < 1 for all 1 ≤ 𝑖 ≤ 𝑚, is called a (strictly) self-similar set.
A self-similar set is the union of a number of smaller similar copies of itself, e.g., the von Koch
curve.
Open set condition An IFS {𝑆1 , 𝑆2 , … , 𝑆𝑚 } satisfies the open set condition if, and only if,
there exists a non-empty bounded open subset 𝑉 of ℝ𝑛 such that
𝑚

𝑉 ⊃ ⋃ 𝑆𝑖 (𝑉)
𝑖=1

with the union disjoint.

Lemma 9.2 Let {𝑉𝑖 } be a collection of disjoint open subsets of ℝ𝑛 such that each 𝑉𝑖 contains
a ball of radius 𝑎1 𝑟 and is contained in a ball of radius 𝑎2 𝑟. Then any ball 𝐵 of radius 𝑟
intersects at most ((1 + 2𝑎2 )⁄𝑎1 )𝑛 of the closures 𝑉𝑖 .
Theorem 9.3 Suppose that the open set condition holds for the similarities 𝑆𝑖 on ℝ𝑛 with
ratios 0 < 𝑐𝑖 < 1 for 1 ≤ 𝑖 ≤ 𝑚. If 𝐹 is the attractor of the IFS {𝑆1 , 𝑆2 , … , 𝑆𝑚 }, that is
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹)
𝑖=1

then dimH 𝐹 = dimB 𝐹 = 𝑠, where 𝑠 is given by

𝑚

∑ 𝑐𝑖𝑠 = 1.
𝑖=1

Moreover, for this value of 𝑠 we have 0 < ℋ 𝑠 (𝐹) < ∞.

If the open set condition is not assumed in Theorem 9.3, it may be shown that we still have
dimH 𝐹 = dimB 𝐹, although this value may be less than 𝑠.
Theorem Let 𝐹 be a set generated by a polygonal curve, then 𝐹 is compact.

Some Variations
The calculations underlying Theorem 9.3 may be adapted to estimate the dimension of the
attractor of an IFS consisting of contractions that are not similarities.
Proposition 9.6 Let 𝐹 be the attractor of an IFS consisting of contractions 𝑆1 , 𝑆2 , … , 𝑆𝑚 on a
closed subset 𝐷 of ℝ𝑛 such that
|𝑆𝑖 (𝑥) − 𝑆𝑖 (𝑦)| ≤ 𝑟𝑖 |𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐷)
with 0 < 𝑟𝑖 < 1 for each 𝑖. Then dimH 𝐹 ≤ dimB 𝐹 ≤ 𝑠, where 𝑠 is given by
 Fractal Geometry - Notes
𝑚

∑ 𝑟𝑖𝑠 = 1.
𝑖=1

Proposition 9.7 Let 𝐹 be the attractor of an IFS consisting of contractions 𝑆1 , 𝑆2 , … , 𝑆𝑚 on a

closed subset 𝐷 of ℝ𝑛 such that
𝑟𝑖 |𝑥 − 𝑦| ≤ |𝑆𝑖 (𝑥) − 𝑆𝑖 (𝑦)| (𝑥, 𝑦 ∈ 𝐷)
with 0 < 𝑟𝑖 < 1 for each 𝑖. Furthermore, suppose that the (non-empty compact) attractor 𝐹
satisfies
𝑚

𝐹 = ⋃ 𝑆𝑖 (𝐹)
𝑖=1

with this union disjoint. Then 𝐹 is totally disconnected, and dimH 𝐹 ≥ 𝑠, where 𝑠 is given by
𝑚

∑ 𝑟𝑖𝑠 = 1.
𝑖=1

Note that the union ⋃𝑚 𝑖=1 𝑆𝑖 (𝐹) will certainly be disjoint if there is some non-empty compact
set 𝐸 with 𝑆𝑖 (𝐸) ⊂ 𝐸 for all 𝑖 and with the 𝑆𝑖 (𝐸) disjoint.

Applications to Encoding Images

The following theorem, sometimes known as the collage theorem, gives an idea of how close
an approximation the attractor of an IFS is to a given set. The corollary that follows shows
that, in principle, any set can be represented arbitrarily closely by an IFS attractor.
Theorem 9.1.3 Let 𝐹 be the attractor of an IFS consisting of contractions 𝑆1 , 𝑆2 , … , 𝑆𝑚 on a
closed subset 𝐷 of ℝ𝑛 such that
|𝑆𝑖 (𝑥) − 𝑆𝑖 (𝑦)| ≤ 𝑐|𝑥 − 𝑦| (𝑥, 𝑦 ∈ 𝐷)
with 0 < 𝑐 < 1 for each 𝑖. Furthermore, let 𝐸 ⊂ 𝐷 be any non-empty compact set. Then
1 𝑚
𝑑(𝐸, 𝐹) ≤ 𝑑 (𝐸, ⋃ 𝑆𝑖 (𝐸))
1−𝑐 𝑖=1

where 𝑑 is the Hausdorff metric.

Corollary 9.1.4 Let 𝐸 be any non-empty compact subset of ℝ𝑛 . Given 𝛿 > 0, there exist IFSs
consisting of similarities 𝑆1 , 𝑆2 , … , 𝑆𝑚 with attractor 𝐹 satisfying 𝑑(𝐸, 𝐹) < 𝛿.
 Fractal Geometry - Notes
Graphs of Functions
Graph of a function the graph of a function 𝑓: [𝑎, 𝑏] → ℝ is defined as
graph 𝑓 = {(𝑡, 𝑓(𝑡)): 𝑎 ≤ 𝑡 ≤ 𝑏}.
Maximum range of a graph of a function the maximum range 𝑅𝑓 of a function 𝑓: [𝑎, 𝑏] →
ℝ over an interval [𝑡1 , 𝑡2 ] is given by
𝑅𝑓 [𝑡1 , 𝑡2 ] = sup |𝑓(𝑡) − 𝑓(𝑢)|.
𝑡1 ≤𝑡,𝑢≤𝑡2

Dimensions of Graphs
Proposition 11.1 Let 𝑓: [0,1] → ℝ be a continuous function. Let 0 < 𝛿 < 1, and let 𝑚 be the
least integer greater or equal to 1⁄𝛿 . Then, if 𝑁𝛿 is the number of squares of the 𝛿-mesh that
intersect graph 𝑓
𝑚−1 𝑚−1
1 1
∑ 𝑅𝑓 [𝑖𝛿, (𝑖 + 1)𝛿] ≤ 𝑁𝛿 ≤ 2𝑚 + ∑ 𝑅𝑓 [𝑖𝛿, (𝑖 + 1)𝛿].
𝛿 𝛿
𝑖=0 𝑖=0

The right-hand inequality remains true even if 𝑓 is not continuous.

Corollary 11.2 Let 𝑓: [0,1] → ℝ be a continuous function. Suppose that
|𝑓(𝑡) − 𝑓(𝑢)| ≤ 𝑐|𝑡 − 𝑢|2−𝑠 (0 ≤ 𝑡, 𝑢 ≤ 1)
where 𝑐 > 0 and 1 ≤ 𝑠 ≤ 2. Then ℋ 𝑠 (graph 𝑓) < ∞ and
dimH (graph 𝑓) ≤ dimB (graph 𝑓) ≤ dimB (graph 𝑓) ≤ 𝑠
This remains true if, for some 𝛿 > 0, the Holder condition holds when |𝑡 − 𝑢| < 𝛿.
Now suppose that for each 𝑡 ∈ [0,1] and 0 < 𝛿 ≤ 𝛿0 there exists 𝑢 such that |𝑡 − 𝑢| < 𝛿 and
|𝑓(𝑡) − 𝑓(𝑢)| ≥ 𝑐𝛿 2−𝑠
with 𝑐 > 0 and 1 ≤ 𝑠 ≤ 2. Then
𝑠 ≤ dimB (graph 𝑓).
Lower bounds for the Hausdorff dimension of graphs are generally very much more awkward
to find than box dimensions.
In order to calculate the box dimension of 𝐹 = graph 𝑓 and 𝐺 we require the smallest number
of sets of diameter at most 𝛿 that can cover the set 𝐹. This quantity is given by the left-hand
inequality above. Therefore, if the box dimension of graph 𝑓 exists, then it follows from
Proposition 11.1 that
1
log ( ∑𝑚−1
𝑖=0 𝑅𝑓 [𝑖𝛿, (𝑖 + 1)𝛿])
dimB (graph 𝑓) = lim 𝛿
𝛿→0 − log 𝛿
log(∑𝑚−1
𝑖=0 𝑅𝑓 [𝑖𝛿, (𝑖 + 1)𝛿])
= 1 + lim ( ).
𝛿→0 − log 𝛿
Weierstrass function the function 𝑓: [0,1] → ℝ defined by
∞

𝑓(𝑡) = ∑ 𝜆(𝑠−2)𝑘 sin(𝜆𝑘 𝑡)

𝑘=1
 Fractal Geometry - Notes
with 𝜆 > 1 and 1 < 𝑠 < 2 is called the Weierstrass function.
The box dimension of the Weierstrass function is equal to 𝑠 provided that 𝜆 is large enough.
The Weierstrass function is representative of a much wider class of functions to which the
method shown in Example 11.3 may apply. If 𝑔 is a suitable periodic function, a similar
method can often show that
∞

𝑓(𝑡) = ∑ 𝜆(𝑠−2)𝑘 𝑔(𝜆𝑘 𝑡)

𝑘=1

has dimB (graph 𝑓) = 𝑠.

Self-affine Curves
Affine transformation A transformation defined by 𝑆(𝑥) = 𝑇(𝑥) + 𝑎 where 𝑇 is a non-
singular linear transformation and 𝑎 is a vector in ℝ𝑛 is called an affine transformation, or
affinity.
An affinity may be thought of as a shearing transformation; its contracting or expanding effect
need not be the same in every direction.
Let {𝑆1 , 𝑆2 , … , 𝑆𝑚 } be affine transformations represented in matrix notation with respect to
(𝑡, 𝑥) coordinates by
𝑡 1⁄𝑚 0 𝑡 (𝑖 − 1)⁄𝑚
𝑆𝑖 [ ] = [ ][ ] + [ ]
𝑥 𝑎𝑖 𝑟𝑖 𝑥 𝑏𝑖
that is
𝑡+𝑖−1
𝑆𝑖 (𝑡, 𝑥) = ( , 𝑎𝑖 𝑡 + 𝑟𝑖 𝑥 + 𝑏𝑖 ).
𝑚
Then, for each 𝑖, 𝑆𝑖 transforms vertical lines {(𝑡0 , 𝑥): 𝑥 ∈ ℝ} to vertical lines {(𝑡0 +
(𝑖 − 1)⁄𝑚 , 𝑥 ′ ): 𝑥 ′ ∈ ℝ}, with the vertical strip 0 ≤ 𝑡 ≤ 1 mapped onto the vertical strip
(𝑖 − 1)⁄𝑚 ≤ 𝑡 ≤ 𝑖 ⁄𝑚. We suppose that
1
≤ 𝑟𝑖 < 1
𝑚
for each 𝑖 so that contractions in the 𝑡 direction are stronger than in the 𝑥 direction.
Now
0 1⁄𝑚 0 0 0 0
𝑆1 [ ⁄(1 ] = [ ] [ ⁄(1 ] + [ ] = [ ]
𝑏1 − 𝑟1 ) 𝑎1 𝑟1 𝑏1 − 𝑟1 ) 𝑏1 𝑏1 ⁄(1 − 𝑟1 )
and
1 1⁄𝑚 0 1 (𝑚 − 1)⁄𝑚
𝑆𝑚 [(𝑎 ) ⁄(1 ) ]=[ ][ ]+[ ]
𝑚 + 𝑏𝑚 − 𝑟1 𝑎𝑚 𝑟𝑚 (𝑎𝑚 + 𝑏𝑚 )⁄(1 − 𝑟1 ) 𝑏𝑚
1
= [(𝑎 ].
𝑚 + 𝑏𝑚 )⁄(1 − 𝑟1 )

Therefore, 𝑝1 = (0, 𝑏1 ⁄(1 − 𝑟1 )) and 𝑝𝑚 = (1, (𝑎𝑚 + 𝑏𝑚 )⁄(1 − 𝑟1 )) are fixed points of 𝑆1 and
𝑆𝑚 , respectively.
The transformation 𝑆𝑖 maps the straight line segment from 𝑝1 to 𝑝𝑚 to the straight line
segment from 𝑆𝑖 (𝑝1 ) to 𝑆𝑖 (𝑝𝑚 ), which we denote [𝑆𝑖 (𝑝1 ), 𝑆𝑖 (𝑝𝑚 )]. We let 𝐸0 = [𝑝1 , 𝑝𝑚 ] and
choose the matrix entries so that
 Fractal Geometry - Notes
𝑆𝑖 (𝑝𝑚 ) = 𝑆𝑖+1 (𝑝1 )
to ensure that the line segments [𝑆𝑖 (𝑝1), 𝑆𝑖 (𝑝𝑚 )] join up to form a polygonal curve 𝐸1 . To avoid
trivial cases, we assume that the points 𝑆1 (𝑝1 ) = 𝑝1 , 𝑆2 (𝑝1), … , 𝑆𝑚 (𝑝1 ), 𝑝𝑚 are not all collinear.
The attractor 𝐹 of the iterated function system {𝑆1 , 𝑆2 , … , 𝑆𝑚 } may be constructed by
repeatedly replacing line segments by affine images of the ‘generator’ set 𝐸1 . The condition
𝑆𝑖 (𝑝𝑚 ) = 𝑆𝑖+1 (𝑝1 ) ensures that the segments join up with the result that 𝐹 is the graph of
some continuous function 𝑓: [0,1] → ℝ.
It should be noted that the above conditions do not necessarily imply that the transformations
𝑆𝑖 are contractions with respect to Euclidean distance. However, it is possible to redefine
distance in the (𝑡, 𝑥) plane in such a way that the 𝑆𝑖 become contractions, and, in this context,
the IFS theory guarantees a unique attractor.
Let 𝐹 = graph 𝑓 be the self-affine curve described above, then
log(𝑟1 + ⋯ + 𝑟𝑚 )
dimB (graph 𝑓) = 1 + .
log 𝑚
Self-affine functions are useful for fractal interpolation. Suppose we wish to find a fractal
curve of a given dimension passing through the points with coordinates (𝑖 ⁄𝑚 , 𝑥𝑖 ) for 𝑖 =
0,1, … , 𝑚. By choosing transformations in such a way that 𝑆𝑖 maps the segment [𝑝1 , 𝑝𝑚 ] onto
the segment [((𝑖 − 1)⁄𝑚 , 𝑥𝑖−1 ), (𝑖 ⁄𝑚 , 𝑥𝑖 )] for each 𝑖, the construction described above gives a
self-affine function with graph passing through the given points. By adjusting the values of the
matrix entries, we can ensure that the curve has the required box dimension.
Self-affine curves can be generalised so that the 𝑆𝑖 do not have the same contraction in the 𝑡
direction. This leads to fractal interpolation between points at unequally spaced intervals of 𝑡.
The calculation of Example 11.4 may be extended to give the box dimension of such curves.
 Fractal Geometry - Notes
Dynamical Systems
Discrete dynamical system Let 𝐷 ⊆ ℝ𝑛 , and let 𝑓: 𝐷 → 𝐷 be a continuous mapping. A
discrete dynamical system is an iterative scheme {𝑓 𝑘 } where 𝑓 𝑘 denotes the 𝑘th iterate of 𝑓.
Orbit The orbit of a discrete dynamical system about a point 𝑥 is the sequence of iterates
𝑘=1 .
{𝑓 𝑘 (𝑥)}∞
Dense orbit Suppose that 𝐹 is an attractor, or repeller of a function 𝑓 and that 𝑦 ∈ 𝐹. Then
the orbit {𝑓 𝑘 (𝑦)} is said to be dense in 𝐹 if, given any point 𝑥 ∈ 𝐹 and any 𝜀 > 0, there exists a
𝑘 ∈ ℕ such that the distance between 𝑓 𝑘 (𝑦) and 𝑥 is at most 𝜀.
The behaviour 𝑓 𝑘 (𝑥)as 𝑘 → ∞ may be one of the following:
(i) convergence to a fixed point 𝑤 ∈ 𝐷, i.e. 𝑓(𝑤) = 𝑤;
(ii) convergence to a period-𝑝 orbit, i.e. {𝑤, 𝑓(𝑤), … , 𝑓 𝑝−1 (𝑤)} where 𝑝 is the least positive
integer with 𝑓 𝑝 (𝑤) = 𝑤 in the sense that |𝑓 𝑘 (𝑥) − 𝑓 𝑘 (𝑤)| → 0 as 𝑘 → ∞;
(iii) random sequence of iterates that remain bounded, these often remain close to a certain
set which may be a fractal;
(iv) divergence to infinity.
Attractor Let 𝐷 ⊆ ℝ𝑛 and let 𝑓: 𝐷 → 𝐷 be a continuous mapping. A subset 𝐹 of 𝐷 is called
an attractor for 𝑓 if 𝐹 is a closed set that is invariant under 𝑓 (i.e. with 𝑓(𝐹) = 𝐹) such that the
distance from 𝑓 𝑘 (𝑥) to 𝐹 converges to zero as 𝑘 tends to infinity for all 𝑥 in an open set 𝑉
containing 𝐹, and 𝐹 has no proper subset satisfying these conditions. The largest such open
set 𝑉 is called the basin of attraction of 𝐹.
Repeller Let 𝐷 ⊆ ℝ𝑛 and let 𝑓: 𝐷 → 𝐷 be a continuous mapping. A subset 𝐹 of 𝐷 is called a
repeller for 𝑓 if 𝐹 is a closed set that is invariant under 𝑓 (i.e. with 𝑓(𝐹) = 𝐹) such that there
exists some 𝑘 for which 𝑓 𝑘 (𝑥) ∉ 𝑉 for all 𝑥 ∈ 𝑉 ∖ 𝐹 in an open set 𝑉 containing 𝐹, and 𝐹 has no
proper subset satisfying these conditions.
Roughly speaking, an attractor is a set to which all nearby orbits converge, whereas a repeller
is a set from which all nearby orbits diverge.
Since 𝑓(𝐷) ⊂ 𝐷, repeatedly applying 𝑓 gives
𝑓 𝑘 (𝐷) ⊂ 𝑓 𝑘−1 (𝐷) ⊂ ⋯ ⊂ 𝑓(𝐷) ⊂ 𝐷
so that
𝑘

⋂ 𝑓 𝑖 (𝐷) = 𝑓 𝑘 (𝐷).
𝑖=1
𝑘
Thus, the set 𝐹 = ⋂∞ 𝑖=1 𝑓 (𝐷) is invariant under 𝑓. Since 𝑓 (𝑥) ∈ ⋂𝑖=1 𝑓 (𝐷) for all 𝑥 ∈ 𝐷, the
𝑖 𝑘 𝑖

iterates 𝑓 𝑘 (𝑥) approach 𝐹 as 𝑘 → ∞, and 𝐹 is frequently an attractor of 𝑓.

If 𝐹 is an attractor, or repeller of 𝑓, then, if 𝑥 ∈ 𝐹, the iterates 𝑓 𝑘 (𝑥) remain in 𝐹 for all 𝑘.
Furthermore, if 𝐹 is a fractal they often exhibit ‘chaotic’ or apparently random behaviour
around 𝐹.
Periodic Point The point 𝛼 is a periodic point, with period 𝑝, of a function 𝑓 if
𝑓 𝑝 (𝛼) = 𝛼 but 𝑓 𝑘 (𝛼) ≠ 𝛼
for 𝑘 = 1,2,3, … , 𝑝 − 1.
The 𝑝 + 1 points
 Fractal Geometry - Notes
𝛼, 𝑓(𝛼), 𝑓 2 (𝛼), … , 𝑓 𝑝 (𝛼)
are said to form a period 𝑝 orbit of 𝑓.
Chaos Let 𝐷 ⊆ ℝ𝑛 , 𝑓: 𝐷 → 𝐷 be a continuous mapping, and 𝐹 be an attractor, or repeller of
𝑓. Then, 𝑓 may certainly be regarded as chaotic on 𝐹 if the following are all true:
(i) for some 𝑥 ∈ 𝐹, the orbit {𝑓 𝑘 (𝑥)} is dense in 𝐹;
(ii) the periodic points of 𝑓 in 𝐹 (points for which 𝑓 𝑝 (𝑥) = 𝑥 for some positive integer 𝑝) are
dense in 𝐹;
(iii) 𝑓 has sensitive dependence on initial conditions, that is, there is a number 𝛿 > 0 such
that for every 𝑥 in 𝐹 there are points 𝑦 in 𝐹 arbitrarily close to 𝑥 such that |𝑓 𝑘 (𝑥) −
𝑓 𝑘 (𝑦)| ≥ 𝛿 for some 𝑘. Thus, points that are arbitrarily close to each other do not remain
so under iterates of 𝑓.
Condition (i) implies that 𝐹 cannot be decomposed into smaller closed invariant sets.
Condition (ii) suggests a skeleton of regularity in the structure of 𝐹. Condition (iii) reflects the
unpredictability of iterates of points on 𝐹.

Repellers and Iterated Function Systems

Under certain circumstances, a repeller in a dynamical system coincides with the attractor of a
related iterated function system.
Tent map The mapping 𝑓: ℝ → ℝ given by
3
𝑓(𝑥) = (1 − |2𝑥 − 1|)
2
is called the tent map (because its graph resembles a tent).
The repeller of the dynamical system {𝑓 𝑘 (𝑥)} where 𝑓 is the tent map is the middle third
Cantor set. It coincides with the attractor of the IFS 𝑆1 , 𝑆2 : [0,1] → [0,1] defined by
𝑥 𝑥
𝑆1 (𝑥) = and 𝑆2 (𝑥) = 1 − .
3 3
If 𝑆1 , 𝑆2 , … , 𝑆𝑚 is a set of bijective contractions on a domain 𝐷 with attractor 𝐹 such that
𝑆1 (𝐹), 𝑆2 (𝐹), … , 𝑆𝑚 (𝐹) are disjoint, then 𝐹 is a repeller for any mapping 𝑓 such that 𝑓(𝑥) =
𝑆𝑖−1 (𝑥) when 𝑥 is near 𝑆𝑖 (𝐹), 𝑖 = 1,2, … , 𝑚. By examining the effect of 𝑓 on the point 𝑥𝑖1 ,𝑖2,… in
𝐹, it may be shown that 𝑓 acts chaotically on 𝐹. The study of 𝑓 by its effect on points of 𝐹
represented by sequences (𝑖1 , 𝑖2 , … ) is known as symbolic dynamics.

The Logistic Map

Logistic map The mapping 𝑓: ℝ → ℝ given by
𝑓𝜆 (𝑥) = 𝜆𝑥(1 − 𝑥)
where 𝜆 is a positive constant, is called the logistic map.
The logistic map is an archetypal 1-dimensional dynamical system that undergoes
bifurcations as the parameter 𝜆 increases. If 𝜆 > 2 + √5, then the dynamics of the logistic map
can be analysed in the same way as the dynamics of the tent map.
 Fractal Geometry - Notes
Stretching and Folding Transformations
One of the simplest planar dynamical systems with a fractal attractor is the baker’s
transformation, so-called because it resembles the process of repeatedly stretching a piece of
dough and folding it in two.
Baker’s transformation Let 𝐸 = [0 × 1] → [0 × 1] be the unit square. The mapping 𝑓: 𝐸 →
𝐸 given by
1
(2𝑥, 𝜆𝑦) 0≤𝑥≤
2
𝑓(𝑥, 𝑦) =
1 1
(2𝑥 − 1, 𝜆𝑦 + ) ≤𝑥≤1
{ 2 2
where 0 < 𝜆 < 1⁄2 is a positive constant, is called the baker’s transformation.
This transformation may be thought of as stretching 𝐸 into a 2 × 𝜆 rectangle, cutting it into
two 1 × 𝜆 rectangles and placing these one above the other with a gap of 1⁄2 − 𝜆 in between.
As such, it is a stretching and cutting transformation, rather than a stretching and folding one.
The fractal attractor 𝐹 of the baker’s transformation is chaotic. It is essentially the product
[0,1] × 𝐹1 , where 𝐹1 is a Cantor set, that is, the IFS attractor of
1
𝑆1 (𝑥) = 𝜆𝑥 and 𝑆2 (𝑥) = + 𝜆𝑥.
2
Furthermore
log 2 log 2
dimH 𝐹1 = and dimH 𝐹 = 1 − .
− log 𝜆 log 𝜆
A specific example of a stretching and folding transformation is the Hénon map.
Hénon map The mapping 𝑓: ℝ2 → ℝ2 given by
𝑓(𝑥) = (𝑦 + 1 − 𝑎𝑥 2 , 𝑏𝑥)
where 𝑎 and 𝑏 are constants, is called the Hénon map.
The transformation may be decomposed into an area preserving bend, a contraction and a
reflection, the net effect being ‘horse-shoe’ like.
 Fractal Geometry - Notes
Iteration of Complex Functions - Julia Sets
Julia sets arise in connection with the iteration of a function of a complex variable 𝑓: ℂ → ℂ , so
are related to discrete dynamical systems.
The study of sequences 𝑓 𝑘 (𝑧) for various initial 𝑧 is known as complex dynamics.

General Theory of Julia Sets

By specialising to functions that are analytic on the complex plane, the powerful techniques of
complex analysis can be used to obtain detailed information about the structure of Julia sets.
For convenience, we take 𝑓: ℂ → ℂ to be a polynomial of degree 𝑛 ≥ 2 with complex
coefficients
𝑓(𝑧) = 𝑎𝑛 𝑧 𝑛 + 𝑎𝑛−1 𝑧 𝑛−1 + ⋯ + 𝑎1 𝑧 + 𝑎0 .
Julia set The filled-in Julia set 𝐾(𝑓) of the polynomial 𝑓 is defined as
𝐾(𝑓) = {𝑧 ∈ ℂ: 𝑓 𝑘 (𝑧) ↛ ∞}.
The Julia set 𝐽(𝑓) of 𝑓 is the boundary of the filled-in Julia set, that is
𝐽(𝑓) = 𝜕𝐾(𝑓).
Thus 𝑧 ∈ 𝐽(𝑓) if, in every neighbourhood of 𝑧, there are points 𝑤 and 𝑣 with 𝑓 𝑘 (𝑤) → ∞ and
𝑓 𝑘 (𝑣) ↛ ∞.
Fatou (or Stable) set The complement of the Julia set is called the Fatou set 𝐹(𝑓), that is
𝐹(𝑓) = ℂ ∖ 𝐽(𝑓).
Fixed point A point 𝛼 is a fixed point of a function 𝑓 if 𝑓(𝛼) = 𝛼.
Attracting fixed point The fixed point 𝛼 of an analytic function 𝑓 is attracting (or
attractive) if
|𝑓 ′ (𝛼)| < 1.
Repelling fixed point The fixed point 𝛼 of an analytic function 𝑓 is repelling (or repulsive)
if
|𝑓 ′ (𝛼)| > 1.
Points close to an attracting fixed point move towards it under iteration by 𝑓, whereas, points
close to a repelling fixed point move away from it under iteration by 𝑓.
Lemma 14.1 Given a polynomial 𝑓(𝑧) = 𝑎𝑛 𝑧 𝑛 + 𝑎𝑛−1 𝑧 𝑛−1 + ⋯ + 𝑎1 𝑧 + 𝑎0 with 𝑎0 ≠ 0,
there exists a number 𝑟 such that, if |𝑧| ≥ 𝑟 then |𝑓(𝑧)| ≥ 2|𝑧|. In particular, if |𝑓 𝑚 (𝑧)| ≥ 𝑟 for
some 𝑚 ≥ 0, then 𝑓 𝑘 (𝑧) → ∞ as 𝑘 → ∞. Thus, either 𝑓 𝑘 (𝑧) → ∞, or {𝑓 𝑘 (𝑧): 𝑘 = 0,1,2, … } is a
bounded set.
Proposition 14.2 Let 𝑓(𝑧) be a polynomial. Then the filled-in Julia set 𝐾(𝑓) and the Julia set
𝐽(𝑓) are non-empty and compact, with 𝐽(𝑓) ⊂ 𝐾(𝑓). Furthermore, 𝐽(𝑓) has an empty interior.
Proposition 14.3 Let 𝑓(𝑧) be a polynomial. The Julia set 𝐽 = 𝐽(𝑓) of 𝑓 is forward and
backward invariant under 𝑓, that is
𝐽 = 𝑓(𝐽) = 𝑓 −1 (𝐽).
Proposition 14.4 Let 𝑓(𝑧) be a polynomial, then
𝐽(𝑓 𝑃 ) = 𝐽(𝑓)
for every positive integer 𝑝.
 Fractal Geometry - Notes
Normal family of analytic functions Let 𝑈 be an open subset of ℂ, and let 𝑔𝑘 : 𝑈 → ℂ, 𝑘 =
1,2,2 … be a family of complex analytic functions. The family 𝒢 = {𝑔𝑘 } is said to be normal on
𝑈 if every sequence of functions in 𝒢 has a subsequence which converges uniformly on every
compact subset of 𝑈, either to a bounded analytic function or to ∞. The family 𝒢 is normal at
the point 𝑤 of 𝑈 if there is some open subset 𝑉 of 𝑈 containing 𝑤 such that 𝒢 is a normal
family on 𝑉.
Theorem 14.5 (Montel's Theorem) Let {𝑔𝑘 } be a family of complex analytic functions on an
open domain 𝑈. If {𝑔𝑘 } is not a normal family, then for all 𝑤 ∈ ℂ with at most one exception
we have 𝑔𝑘 (𝑧) = 𝑤 for some 𝑧 ∈ 𝑈 and some 𝑘.
The contrapositive of Montel’s theorem is often useful and can be stated as follows. If a family
of complex analytic functions {𝑔𝑘 } on an open domain 𝑈 fails to take two or more values on 𝑈,
then the family is not normal in 𝑈.
Montel’s Theorem (contrapositive definition) Let 𝑈 be an open subset of the complex
plane and let 𝑎, 𝑏 ∈ ℂ such that 𝑎 ≠ 𝑏. Suppose that
𝑔𝑘 : 𝑈 → ℂ ∖ {𝑎, 𝑏}
is analytic for 𝑘 = 1,2,3, …, then the family of functions {𝑔𝑘 } is normal on 𝑈.
Proposition 14.6 Let 𝑓(𝑧) be a polynomial. The Julia set 𝐽(𝑓) is the set of points 𝑧 ∈ ℂ such
that the family {𝑓𝑘 } is not a normal at 𝑧, that is
𝐽(𝑓) = {𝑧 ∈ ℂ: the family {𝑓𝑘 } is not a normal at 𝑧}.
Proposition 14.6 is used as the definition of the Julia set for general complex functions.
Corollary of Montel’s Theorem Suppose that 𝑈 an open subset of the complex plane and at
least two points of ℂ lie outside of 𝑈. If 𝑓 is a polynomial such that 𝑓(𝐷) ⊆ 𝐷, then 𝐷 is a
subset of the Fatou set of 𝑓.
Lemma 14.7 Let 𝑓(𝑧) be a polynomial, let 𝑤 ∈ 𝐽(𝑓) and let 𝑈 be any neighbourhood of 𝑤.
Then, for each 𝑗 = 1,2, …, the set
∞

𝑊 ≡ ⋃ 𝑓 𝑘 (𝑈)
𝑘=𝑗

is the whole of ℂ, except possibly for a single point. Any such exceptional point is not in 𝐽(𝑓),
and is independent of 𝑤 and 𝑈.
Corollary 14.8 Let 𝑓(𝑧) be a polynomial.
(a) The following holds for all 𝑧 ∈ ℂ with at most one exception: if 𝑈 is an open set
intersecting 𝐽(𝑓), then 𝑓 −𝑘 (𝑧) intersects 𝑈 for infinitely many values of 𝑘.
(b) If 𝑧 ∈ 𝐽(𝑓), then 𝐽(𝑓) is the closure of ⋃∞
𝑘=1 𝑓
−𝑘 (𝑧)
.
Proposition 14.9 Let 𝑓(𝑧) be a polynomial, then 𝐽(𝑓) is a perfect set (i.e. closed and with no
isolated points) and is therefore uncountable.
Theorem 14.10 Let 𝑓(𝑧) be a polynomial, then 𝐽(𝑓) is the closure of the repelling periodic
points of 𝑓.
Basin of attraction If 𝛼 is an attracting fixed point of an analytic function 𝑓, then the basin
of attraction 𝐴(𝛼) of 𝛼 under 𝑓 is the set
{𝑧: 𝑓 𝑘 (𝑧) → 𝛼 as 𝑘 → ∞}.
Lemma 14.11 Let 𝑤 be an attractive fixed point of the polynomial 𝑓. Then
 Fractal Geometry - Notes
𝜕𝐴(𝑤) = 𝐽(𝑓).
The same is true if 𝑤 = ∞.
In summary then, the Julia set 𝐽(𝑓) of a polynomial 𝑓 is the boundary of the set of points 𝑧 ∈ ℂ
such that 𝑓 𝑘 (𝑧) → ∞. It is an uncountable non-empty compact set containing no isolated
points and is invariant under 𝑓 and 𝑓 −1 . Furthermore, 𝐽(𝑓) = 𝐽(𝑓 𝑃 ) for every positive integer
𝑝. If 𝑧 ∈ 𝐽(𝑓), then 𝐽(𝑓) is the closure of ⋃∞
𝑘=1 𝑓
−𝑘 (𝑧)
. The Julia set is the boundary of the basin
of attraction of each attractive fixed point of 𝑓, including ∞, and is the closure of the repelling
periodic points of 𝑓.

Quadratic Functions – the Mandelbrot Set

Let
𝑓𝑐 = 𝑧 2 + 𝑐 (𝑐 ∈ ℂ),
and
ℎ(𝑧) = 𝛼𝑧 + 𝛽.
Then
𝛼 2 𝑧 2 + 2𝛼𝛽𝑧 + 𝛽 2 − 𝛽 + 𝑐
𝑓(𝑧) = ℎ−1 (𝑓𝑐 (ℎ(𝑧))) =
𝛼
is a quadratic function.
The mapping ℎ is called the conjugacy between 𝑓 and 𝑓𝑐 .
Since ℎ−1 ∘ 𝑓𝑐𝑘 ∘ ℎ = 𝑓 𝑘 , the sequence of iterates {𝑓 𝑘 (𝑧)} of a point 𝑧 under 𝑓 is just the image
under ℎ−1 of the sequence of iterates {𝑓𝑐𝑘 (ℎ(𝑧))} of the point ℎ(𝑧) under 𝑓𝑐 . The mapping ℎ
transforms the dynamical picture of 𝑓 to that of 𝑓𝑐 . In particular, 𝑓 𝑘 (𝑧) → ∞ if, and only if,
𝑓𝑐𝑘 (𝑧) → ∞; thus the Julia set of 𝑓 is the image under ℎ−1 of the Julia set of 𝑓𝑐 .
Since we are free to choose the value of 𝛼 and 𝛽 in the definition of ℎ, any quadratic function is
conjugate to 𝑓𝑐 for some 𝑐. So, by studying the Julia sets of 𝑓𝑐 for 𝑐 ∈ ℂ, we effectively study the
Julia sets of all quadratic polynomials. Since ℎ is a similarity transformation, the Julia set of
any quadratic polynomial is geometrically similar to that of 𝑓𝑐 for some 𝑐.
Mandelbrot set The Mandelbrot set 𝑀 is defined to be the set of parameters 𝑐 for which the
Julia set is connected, that is
𝑀 = {𝑐 ∈ ℂ: 𝐽(𝑓𝑐 ) is connected}.
Loop A smooth (i.e. differentiable) closed simple (i.e. non-self-intersecting) curve in the
complex plane.
Lemma 14.13 Let 𝐶 be a loop in the complex plane.
(a) If 𝑐 is inside 𝐶, then 𝑓𝑐−1 (𝐶) is a loop. Moreover, 𝑓𝑐 maps the interior of 𝑓𝑐−1 (𝐶) onto the
interior of 𝐶, and the exterior of 𝑓𝑐−1 (𝐶) onto the exterior of 𝐶.
(b) If 𝑐 lies on 𝐶, then 𝑓𝑐−1 (𝐶) is a figure of eight with self-intersection at 𝑧 = 0. Moreover, 𝑓𝑐
maps the interior of the two loops of 𝑓𝑐−1 (𝐶) onto the interior of 𝐶, and the exterior of
𝑓𝑐−1 (𝐶) onto the exterior of 𝐶.
(c) If 𝑐 is outside 𝐶, then 𝑓𝑐−1 (𝐶) comprises of two distinct loops. Moreover, 𝑓𝑐 maps the
interior of each loop of 𝑓𝑐−1 (𝐶) onto the interior of 𝐶, and the region outside these two
loops onto the exterior of 𝐶.
 Fractal Geometry - Notes
Theorem 14.14 For 𝑐 ∈ ℂ, the Julia set 𝐽(𝑓𝑐 ) is connected if the sequence of iterates
𝑘=1 is bounded and is totally disconnected otherwise. Thus
{𝑓𝑐𝑘 (0)}∞
𝑀 = {𝑐 ∈ ℂ: 𝐽(𝑓𝑐 ) is connected}
= {𝑐 ∈ ℂ: {𝑓𝑐𝑘 (0)}∞
𝑘=1 is bounded}

= {𝑐 ∈ ℂ: 𝑓𝑐𝑘 (0) ↛ ∞ as 𝑘 → ∞}.

It should be noted that the origin 𝑧 = 0 is the critical point of 𝑓𝑐 for each 𝑐, and is the point at
which 𝑓𝑐 fails to be a local bijection.

Julia Sets of Quadratic Functions

Theorem 14.15 Suppose
1
|𝑐| > (5 + 2√6) = 2.475 ….
4
Then, 𝐽(𝑓𝑐 ) is totally disconnected, and is the attractor of the contractions given by the two
branches of
𝑓 −1 (𝑧) = ±(𝑧 − 𝑐)1⁄2
for 𝑧 near 𝐽. When |𝑐| is large
2 log 2
dimB 𝐽(𝑓𝑐 ) = dimH 𝐽(𝑓𝑐 ) ≃ .
log(4|𝑐|)
Theorem 14.16 If |𝑐| < 1⁄4, then 𝐽(𝑓𝑐 ) is a simple closed curve.