Fractal Geometry

Mathematical Foundations and Applications

Third Edition
by Kenneth Falconer

Solutions to Exercises
Acknowledgement: Grateful thanks are due to Gwyneth Stallard for providing
solutions for many of the exercises.

For many of the exercises, drawing a diagram will be found extremely helpful.

Chapter 1

1.1 (i) The triangle inequaility.

Let x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ). Then
n
X n
X n
X n
X
|x + y|2 = (xi + yi )2 = x2i + 2 xi y i + yi2
i=1 i=1 i=1 i=1
n n
!1/2 n
!1/2 n
X X X X
≤ x2i +2 x2i yi2 + yi2
i=1 i=1 i=1 i=1

n
!1/2 n
!1/2 2
X X
=  x2i + yi2  = (|x| + |y|)2
i=1 i=1

where we have used Cauchy’s inequality.

(ii) The reverse triangle inequality.
Write y = z −x so x = z −y. Then (i) becomes |z| ≤ |z −y|+|y| or |z|−|y| ≤ |z −y|.
Interchanging the roles of y and z we also have |y| − |z| ≤ |y − z| = |z − y|. Thus
||z| − |y|| = max{|z| − |y|, |y| − |z|} = |z − y|, which is the desired inequality.
(iii) Triangle inequality - metric form.
We have
|x − y| = |(x − z) + (z − y)| ≤ |x − z| + |z − y|
using triangle inequality (i).

1.2 We may assume A is closed since Aδ = Aδ . Let x ∈ Aδ+δ0 . Then there exists a ∈ A
such that |x − a| ≤ δ + δ 0 . If x = a, then clearly x ∈ (Aδ )δ0 . Otherwise let y be the
point on the line segment [a, x] distance δ from a. Thus y = a + δ(x − a)/|x − a|, so
|y − a| = δ|x − a|/|x − a| = δ, so y ∈ Aδ . Moreover, x − y = x − a − δ(x − a)/|x − a| =
(x − a)[1 − δ/|x − a|], so |x − y| = |x − a| − δ ≤ δ + δ 0 − δ = δ 0 . As y ∈ Aδ , x ∈ (Aδ )δ0 ,
so Aδ+δ0 ⊆ (Aδ )δ0 .

1
 Now let x ∈ (Aδ )δ0 . We may find y ∈ Aδ such that |x − y| ≤ δ 0 , and then we
may find a ∈ A such that |y − a| ≤ δ. By the triangle inequality, Exercise 1.1(iii),
|x − a| ≤ |x − y| + |y − a| ≤ δ 0 + δ, so x ∈ Aδ+δ0 . Thus (Aδ )δ0 ⊆ Aδ+δ0 . We conclude
(Aδ )δ0 = Aδ+δ0 .

1.3 Let A be bounded, that is A has finite diameter, so supx,y∈A |x − y| = d < ∞, where
d is the diameter of A. Let a be any point of A. Then for all x ∈ A, |x − a| ≤ d, so
that |x| = |a+(x−a)| ≤ |a|+|x−a| ≤ |a|+d, using the triangle inequality, Exercise
1.1(i). Thus, setting r = |a| + d, we have x ∈ B(0, r). We conclude A ⊆ B(0, r).
If A ⊆ B(0, r) and x, y ∈ A, then |x − y| ≤ |x| + |y| ≤ r + r = 2r, so diamA ≤ 2r,
and in particular A is of finite diameter.

1.4 (i) A non-empty finite set is closed but not open, with A = A, and intA = ∅.
(ii) The interval (0, 1) is open but not closed, with (0, 1) = [0, 1] and int(0, 1) = (0, 1).
(iii) The interval [0, 1] is closed but not open, with [0, 1] = [0, 1] and int[0, 1] = (0, 1).
(iv) The half-open interval [0, 1) is neither open or closed, with [0, 1) = [0, 1] and
int[0, 1) = (0, 1).
(v) The set A = {0, 1, 21 , 13 , 14 , . . .} is closed but not open, with A = A and intA = ∅.

1.5 T
Following the usual construction, the middle third Cantor set may be written F =
∞ k
k=0 Ek , where Ek consists of the union of 2 disjoint closed intervals in [0, 1], each
of length 3−k . For each k, Ek is closed since it is the union of finitely many closed
sets. Since the intersection of any collection of closed sets is closed (see Exercise
1.6), we conclude that F is closed. F is a subset of [0, 1] so it is bounded, and hence
F is compact.
To show that F is totally disconnected, suppose x, y ∈ F with x < y. Then we can
find an Ek such that x and y belong to different intervals [a, b] and [c, d] of Ek with
b < c. Let b < p < c. Then F is contained in the union of the disjoint open intervals
(−1, p) and (p, 2), with x ∈ (−1, p) and y ∈ (p, 2). Thus F is totally disconnected.
Since F is closed, F = F . Since F contains no open interval, intF = ∅, and thus
∂F = F \ intF = F .

1.6 Let {Ai : i ∈ I} be a collection of open subsets of Rn and let A = i∈I Ai . If x ∈ A,

S
then x belongs to one of the sets, Aj , say. Since Aj is open, there exists r > 0 such
that B(x, r) ⊂ Aj ⊂ A, and hence A is open.
n
Tk let {A1 , A2 , . . . , Ak } be a finite collection of open subsets of R and let A =
Now
i=1 Ai . If x ∈ A, then x belongs to each of the open sets Ai and hence, for each
i = 1, . . . , k, there exists ri > 0 such that B(x, ri ) ⊂ Ai . Letting r = min1≤i≤k ri > 0,
then B(x, r) ⊂ B(x, ri ) ⊂ Ai for all i, so that B(x, r) ⊂ A and hence A is open.
Let A ⊂ Rn and let B = Rn \ A be the complement of A. First assume that B is
not open. Then there exists x ∈ B such that, for every positive integer k, the ball
B(x, 1/k) is not contained in B and we may choose a sequence xk ∈ B(x, 1/k) \ B,

2
 so xk ∈ A and xk → x ∈
/ A, so A is not closed. Thus if A is closed then B must be
open.
Now suppose that A is not closed so that there exists a sequence of points xk ∈ A
with xk → x ∈ B = Rn \ A. It follows that, for every r > 0, there is some
xk ∈ B(x, r) \ B so that B(x, r) 6⊂ B, giving that B is not open. Thus if B is open
then A = Rn \ B must be closed.
Now let {Bi : i ∈ I} be a collection of closed subsets of Rn and let B = i∈I Bi .
T
Each of the sets Ai = Rn \ Bi is open. Thus
[ [ \
A= Ai = (Rn \ Bi ) = Rn \ Ai = Rn \ B
i∈I i∈I i∈I

is open and hence B is closed.

n
Sk if {Bi : i = 1, . . . , k} is a finiten collection of closed subsets of R and
Similarly,
B = i=1 Bi , then each of the sets Ai = R \ Bi is open and hence
k
\ k
\ k
\
n n
A= Ai = (R \ Bi ) = R \ Bi = Rn \ B
i=1 i=1 i=1

is open so that B is closed.

1.7 Recall that a subset of Rn is compact if and only if it is both closed and bounded.
Exercise 1.6 showed that the intersection of any collection of closed subsets of Rn
is closed. Thus, if A1 ⊃TA2 ⊃ · · · is a decreasing sequence of non-empty compact
subsets of Rn then A = ∞ k=1 Ak is certainly closed. It is also bounded, since it is a
subset of A1 which is bounded, so A is compact.
T∞
To show that A is non-empty
T∞ we
S∞ argue by contradiction. Suppose
S∞ that k=1 Ak = ∅
n n n n
so that R = R \ k=1 Ak = k=1 (R \ Ak ). Then A1 ⊂ k=1 (R \ Ak ). Since A1
is compact, it follows that A1 is contained in the union of finitely many of the open
sets Rn \ Ak . Since Rn \ A1 ⊂ Rn \ A2 ⊂ · · · , it follows that A1 ⊂ (Rn \ Ak ) for some
k. This is impossible, since Ak ⊂ A1 and Ak 6= ∅, for each k.

1.8 The half-open interval [0, 1) is a Borel subset of R since, for example,

[0, 1) = [0, 2] ∩ (−1, 1).

where [0, 2] is closed and hence a Borel set and (−1, 1) is open and hence a Borel set.

1.9 Let Ak be the set of numbers in [0, 1] whose kth digit is 5. Then Ak is union of
10k−1 half open intervals, so is Borel. Then
∞ [
\ ∞
F = Ak ,
j=1 k=j

as x ∈ F if and only if x ∈ Ak for arbitraily large k. Thus F is formed as the

countable intersection of a countable union of Borel sets, so is Borel.

3
 1.10 Let x = (x1 , x2 ), y = (y1 , y2 ), a = (a1 , a2 ). We may write the transformation S as
S(x1 , x2 ) = (cx1 cos θ − cx2 sin θ + a1 , cx1 sin θ + cx2 cos θ + a2 )
so
|S(x1 ,x2 ) − S(y1 , y2 )|2
= c2 | (x1 − y1 ) cos θ − (x2 − y2 ) sin θ, (x1 − y1 ) sin θ + (x2 − y2 ) cos θ |2

= c2 (x1 − y1 )2 cos2 θ + (x2 − y2 )2 sin2 θ − 2(x1 − y1 )(x2 − y2 ) sin θ cos θ

+ (x1 − y1 )2 sin2 θ + (x2 − y2 )2 cos2 θ + 2(x1 − y1 )(x2 − y2 ) sin θ cos θ

= c2 ((x1 − y1 )2 + (x2 − y2 )2 ) = c2 |(x1 , x2 ) − (y1 , y2 )|2 ,

using that cos2 θ + sin2 θ = 1. Thus |S(x) − S(y)| = c|x − y|, so S is a similarity of
ratio c.     
cos θ − sin θ x1 x1
Note that gives the vector rotated about the ori-
sin θ cos θ x2 x2
gin by an anticlockwise angle θ. Thus the geometrical effect of the similarity S is a
dilation about the origin of scale
 c, followed
 by a rotation through angle θ, followed
a1
by a translation by the vector .
a2

1.11(i) Since sin x → sin 0 = 0 as x → 0, we have

limx→0 sin x = limx→0 sin x = lim sin x = 0.
x→0

(ii) We know that

−1 ≤ sin(1/x) ≤ 1, for x > 0
so that
−1 ≤ limx→0 sin(1/x) ≤ limx→0 sin(1/x) ≤ 1.
Moreover, for each n = 1, 2, . . .,
sin(1/xn ) = −1, for xn = 1/(2n + 3/2)π → 0
and
sin(1/yn ) = 1, for yn = 1/(2n + 1/2)π.
Thus
limx→0 sin(1/x) ≤ −1 and limx→0 sin(1/x) ≥ 1,
so limx→0 sin(1/x) = −1 and limx→0 sin(1/x) = 1.
(iii) We have
|x2 + x sin(1/x)| ≤ |x2 | + |x| → 0 as x → 0. Thus
limx→0 (x2 + (3 + x) sin(1/x)) = limx→0 (x2 + x sin(1/x)) + limx→0 3 sin(1/x)
= 0 − 3 = −3
using part (ii). Similarly
limx→0 (x2 + (3 + x) sin(1/x)) = limx→0 (x2 + x sin(1/x)) + limx→0 3 sin(1/x)
= 0 + 3 = 3.

4
1.12 If f, g : [0, 1] → R are Lipschitz functions, then there exist c1 , c2 > 0 such that

|f (x) − f (y)| ≤ c1 |x − y| and |g(x) − g(y)| ≤ c2 |x − y| (x, y ∈ [0, 1]).

It follows that

|f (x) + g(x) − (f (y) + g(y))| ≤ |f (x) − f (y)| + |g(x) − g(y)|

≤ (c1 + c2 )|x − y| (x, y ∈ [0, 1])

and so the function defined by f (x) + g(x) is also Lipschitz.

For x, y ∈ [0, 1],

|f (x)g(x) − f (y)g(y)| = |f (x)g(x) − f (y)g(x) + f (y)g(x) − f (y)g(y)|

≤ |f (x)g(x) − f (y)g(x)| + |f (y)g(x) − f (y)g(y)|
= |g(x)||f (x) − f (y)| + |f (y)||g(x) − g(y)|.

Moreover, for x ∈ [0, 1], we have |f (x) − f (0)| ≤ c1 |x| ≤ c1 , so that |f (x)| ≤
|f (0)| + c1 = c01 , say. Similarly |g(x)| ≤ c02 . Thus

|f (x)g(x) − f (y)g(y)| ≤ |c01 ||f (x) − f (y)| + |c02 ||g(x) − g(y)|

≤ (c1 c01 + c2 c02 )|x − y|

so f (x)g(x) is Lipschitz.

1.13 Given x, y ∈ R with y 6= x, it follows from the mean-value theorem that there exists
a ∈ (y, x) with
f (x) − f (y)
= f 0 (a).
x−y
Thus  
 f (x) − f (y)  0
 x − y  = |f (a)| ≤ c
 

and hence
|f (x) − f (y)| ≤ c|x − y| (x, y ∈ R)
so that f is a Lipschitz function.

1.14 If f : X → Y is a Lipschitz function, then

|f (x) − f (y)| ≤ c|x − y| (x, y ∈ R),

for some c > 0. Thus, given  > 0 and y ∈ R, it follows that |f (x) − f (y)| < ,
whenever
c|x − y| < ,
that is, whenever
|x − y| < /c.
So, on taking δ = /c > 0, it follows that f is continuous at y, using the ‘epsilon-
delta’ definition of continuity.

5
1.15 Note that if y = f (x) = x2 + x, then solving the quadratic equation for x, we get
x = 21 (−1 ± (1 + 4y)1/2 ), taking real values only. Thus (i) f −1 (2) = {−2, 1}. (ii)
f −1 (−2) = ∅. (iii) As y increaces from 2 to 6, (1 + 4y)1/2 increases from 3 to 5, so
x runs over two ranges [1, 2] and [−3, −2]. Hence f −1 ([2, 6]) = [−3, −2] ∪ [1, 2].

1.16 For 0 ≤ x, y ≤ 2,

|f (x) − f (y)| = |x2 − y 2 | = |x + y||x − y| ≤ 4|x − y|

so f is also Lipschitz on [0, 2].

Thus f is also Lipschitz on [1, 2], with f ([1, 2]) = [1, 4]. For 1 ≤ x, y ≤ 4,
√ √ x−y 1
|f −1 (x) − f −1 (y)| = | x − y| = | √ √ | ≤ |x − y|
x+ y 2

so f −1 is Lipschitz on [1, 4], so f is bi-Lipschitz on [1, 2].

For x > 0,
|f (2x) − f (x)| 4x2 − x2
= = 3x.
|2x − x| x
Thus |f (x) − f (y)|/|x − y| is not bounded on R so f is not Lipschitz on R.

1.17 We use the ‘openScover’ definition of compactness. Let E be compact, f continu-

ous, and f (E) ⊂ Ui , a cover
S of f (E) by open sets. Since f is continuous, the sets
f −1 (Ui ) are open, so E ⊂ f −1 (Ui ) is aScover of E by open sets. By
Scompactness of
m −1 m
E this has a finite subcover, say E ⊂ r=1 f (Ui(r) ), so f (E) ⊂ r=1 Ui(r) , which
gives a cover of f (E) by a finite subset of the Ui . Thus f (E) is compact.

1.18 We take complements in A1 . Thus A1 \ A2 , A1 \ A3 , . . . is an increasing sequence of

sets, so by (1.6)
∞
\ ∞
[
µ(A1 \ Ai ) = µ( (A1 \ Ai )) = lim µ(A1 \ Ai ).
i→∞
i=1 i=1
T∞
Since µ(A1 ) < ∞, this gives T∞µ(A1 ) − µ( i=1 A1 ) = limi→∞ (µ(A1 ) − µ(Ai )) =
µ(A1 ) − limi→∞ µ(Ai ), so µ( i=1 Ai ) = limi→∞ µ(Ai ).

1.19 We show that µ satisfies conditions (1.1)–(1.4) and is hence a measure.

First, since a ∈
/ ∅, µ(∅) = 0 and thus (1.1) is satisfied.
Secondly, suppose that A ⊂ B. If a ∈ A, then a also belongs to B and hence
µ(A) = µ(B) = 1. If a ∈ / A, then µ(A) = 0 ≤ µ(B). Thus, in both of the two
possible cases, (1.2) is satisfied.
Finally, suppose
S∞ that A1 , A2 , . . . is a sequence of sets. If a ∈
/ Ai , for each i ∈ N,
then a ∈/ i=1 Ai so that
∞
[ ∞
X
µ( Ai ) = 0 = µ(Ai ).
i=1 i=1

6
 S∞
On the other hand, if a ∈ Aj , for some integer j, then a ∈ i=1 Ai so that
∞
[ ∞
X
µ( Ai ) = 1 = µ(Aj ) ≤ µ(Ai ).
i=1 i=1

If the sets Ai are disjoint, then a ∈

/ Ai for i 6= j so that µ(Ai ) = 0 for i 6= j and
hence ∞ ∞
[ X
µ( Ai ) = 1 = µ(Ai ).
i=1 i=1

Thus, in both of the two possible cases, (1.3) and (1.4) are satisfied.

1.20 With the construction of the middle third Cantor set F as indicated in figure 0.1,
k −k
the kth stage of the construction ET
k is the union of 2 intervals each of length 3 ,
with E0 ⊃ E1 ⊃ E2 ⊃ . . . and f = ∞ k=1 Ek .
Define a mass distribution µ by starting with unit mass on E0 = [0, 1], splitting
this equally between the two intervals of E1 , splitting the mass on each of these
intervals equally between the two sub-intervals in E2 , etc. Thus we construct a
mass distribution µ on F by repeated subdivision, splitting the mass in as uniform
a way as possible at each stage. For each interval I in Ek we have µ(I) = 2−k , and
this allows us to calculate the mass of any combination of intervals from the Ek and
defines µ on every subset of R.

1.21 For all  > 0, ∅ ⊂ [0, ] so L1 (∅) ≤ L1 ([0, ]) = . This is true for arbitraily small
 > 0, so L(∅) = 0, as required for (1.1).
Let A ⊂ B. Given S> 0 we may find P∞a countable collection of intervals [ai , bi ]
such that A ⊂ B ⊂ ∞ [a
i=1 i i, b ] with (b
i=1 i − a i ) < L1
(B) + . It follows that
1 1 1 1
L (A) ≤ L (B) +  for all  > 0, so that L (A) ≤ L (B) for (1.2).
For (1.3), assume that L1 (Ai ) < ∞ for each i, since the result is clearly true other-
wise. For each  > 0 and i = 1, 2, . . ., there exist intervals [ai,j , bi,j ] such that
∞ ∞
[ X 
Ai ⊂ [ai,j , bi,j ] and (bi,j − ai,j ) < L1 (Ai ) + .
j=1 j=1
2i
S∞ S∞ S∞
Clearly i=1 Ai ⊂ i=1 j=1 [ai,j , bi,j ] and so
∞ ∞ X
∞ ∞ ∞
[ X X  X
L1 ( Ai ) ≤ (bi,j − ai,j ) ≤ (L1 (Ai ) + i
) = L1 (Ai ) + .
i=1 i=1 j=1 i=1
2 i=1

It follows that (1.3) holds.

1.22 We begin by showing that µ satisfies conditions (1.1)–(1.4) and is hence a measure.
First,
µ(∅) = L1 ({x : (x, f (x)) ∈ ∅}) = L1 (∅) = 0
and so (1.1) is satisfied.

7
 Second, if A ⊂ B, then {x : (x, f (x)) ∈ A} ⊂ {x : (x, f (x)) ∈ B} and so, since L1
is a measure,
µ(A) = L1 ({x : (x, f (x)) ∈ A}) ≤ L1 ({x : (x, f (x)) ∈ B}) = µ(B)
so that (1.2) is satisfied.
Finally, if A1 , A2 , . . . is a sequence of sets, then, since L1 is a measure,
∞
[ ∞
[ ∞
[
1 1
µ( Ai ) = L ({x : (x, f (x)) ∈ Ai }) = L ( {x : (x, f (x)) ∈ Ai })
i=1 i=1 i=1
∞
X ∞
X
1
≤ L ({x : (x, f (x)) ∈ Ai }) = µ(Ai )
i=1 i=1

so that (1.3) is satisfied. If the sets Ai are disjoint Borel sets then, since L1 is a
measure,
∞
[ ∞
[ ∞
[
1 1
µ( Ai ) = L ({x : (x, f (x)) ∈ Ai }) = L ( {x : (x, f (x)) ∈ Ai })
i=1 i=1 i=1
∞
X ∞
X
1
= L ({x : (x, f (x)) ∈ Ai }) = µ(Ai )
i=1 i=1

so that (1.4) is satisfied.

Thus µ is a measure on R2 . We now show that µ is supported by the graph of f .
We begin by noting that, since [0, 1] is compact (that is, closed and bounded) and
the map F defined by F (x) = (x, f (x)) is continuous, then the graph of f which is
equal to F ([0, 1]) is also compact and hence closed. Clearly,
µ(R2 \ graphf ) = L1 ({x : (x, f (x)) ∈ R2 \ graphf }) = L1 (∅) = 0.
Now let a ∈ [0, 1] and let r > 0. Since f is continuous, a belongs to a non-trivial
interval Ir ⊂ [0, 1] such that, for each x ∈ Ir , we have (x, f (x)) ∈ B((a, f (a)), r)
and hence
µ(B((a, f (a)), r)) = L1 ({x : (x, f (x)) ∈ B((a, f (a)), r)}) ≥ L1 (Ir ) > 0.
Thus the graph of f is the smallest closed set X such that µ(R2 \ X) > 0; that is,
the graph of f is the support of µ.
Finally,
µ(graphf ) = L1 ([0, 1]) = 1
so that 0 < µ(graphf ) < ∞ and hence µ is a mass distribution.

1.23 For positive integers m, n define sets

1
for all k ≥ n}.
Am,n = {x ∈ D : |fk (x) − f (x)| <
m
S∞ each m the sequence of sets Am,1 ⊂ Am,2 ⊂ Am,2 ⊂ . . . is increasing with
For
n=1 Am,n = D, so by (1.6) thereTis a positive integer nm such that µ(D \ Am,n ) <
2−m  for all n ≥ nm . Define A = ∞m=1 Am,nm . Then
∞ ∞
[ X 
µ(D \ A) < µ( D \ Am,nm ) ≤ m
≤
m=1 m=1
2

8
 .
1
Let δ > 0 and take m > 1/δ. If x ∈ A, then x ∈ Am,nm , so |fk (x) − f (x)| < m
<δ
for all k ≥ nm , so fk (x) → f (x) uniformly on A.

1.24 For n = 1, 2, . . . let Dn = {x : f (x) ≥ 1/n}. Then

Z Z
1 1
0= f dµ ≥ χDn dµ = µ(Dn ),
D D n n
1
S∞ n χDn is a simple function. Thus µ(Dn ) = 0 for all n. Since {x : f (x) > 0} =
since
n=1 Dn , it follows that µ{x : f (x) > 0} = 0, that is f (x) = 0 for almost all x.

1.25 E((X − E(X))2 ) = E(X 2 − 2XE(X) + E(X)2 ) = E(X 2 ) − 2E(X)E(X) + E(X)2 =

E(X 2 ) − E(X)2 .

1.26 The uniform distribution on [a, b] has p.d.f. f (u) = 1/(a − b) for a ≤ u ≤ b and
f (u) = 0 otherwise. Thus
Z b  b
−1 −1 1 2 1 1
E(X) = (a − b) udu = (a − b) u = (b2 − a2 )/(b − a) = (a + b).
a 2 a 2 2
Z b  b
2 −1 2 −1 1 3
E(X ) = (a − b) u du = (a − b) u
a 3 a
1 1
= (b3 − a3 )/(b − a) = (a2 + ab + b2 ).
3 3
Thus
1 1
var(X) = E(X 2 ) − E(X)2 = (a2 + ab + b2 ) − (a + b)2
3 4
1 1
= (a2 − 2ab + b2 ) = (a − b)2 .
12 12

1.27 Define random variables Xk by Xk = 0 if ω ∈ / Ak and Xk = 1 if ω ∈ Ak . Then

Nk = X1 +. . .+Xk , so by the strong law of large numbers (1.25), Nk /k → E(Xk ) = p.
Thus taking Ak to be the event that the kth trial is sucessful, Nk /k is the proportion
of sucesses, which converges to p, the probability of sucess.

1.28 With Xk = 1 if a six is scored on he kth throw and 0 otherwise, and Sk = X1 +. . .+Xk
as the number of sixes in the first k throws, Xk has mean m = 16 and variance
σ 2 = 16 ( 56 )2 + 56 ( 61 )2 = 36
5
. By (1.26)
! Z
√ ∞  
Sk − 1000 1 1 2
P(Sk ≥ 1050) = P p √ ≥ 3 ' √ √ exp − u du = 0.075.
5/36 6000 3 2π 2

9
 Chapter 2

2.1 Let F be a subset of Rn , let Nδ (F ) denote the smallest number of closed balls
of radius δ that cover F and let Nδ0 (F ) denote the number of δ-mesh cubes that
intersect F .
√
For each δ-mesh cube that intersects F , take a closed ball of radius δ n whose centre
is at the centre
√ of the cube; the ball clearly contains the cube (whose diagonal is
of length δ n) and so Nδ n (F ) ≤ Nδ0 (F ). On the other hand, any closed ball of
√

radius δ intersects at most 4n δ-mesh cubes and so Nδ0 (F ) ≤ 4n Nδ (F ). Combining:

Nδ√n (F ) ≤ Nδ0 (F ) ≤ 4n Nδ (F )
√
so that if δ n < 1, then

log Nδ√n (F ) log Nδ0 (F ) log 4n Nδ (F )

≤ ≤
− log δ − log δ − log δ
so
log Nδ√n (F ) log Nδ0 (F ) log 4n + log Nδ (F )
√ √ ≤ ≤ .
− log δ n + log n − log δ − log δ
√
Taking lower limits as δ → 0, so that also δ n → 0, we get that

log Nδ (F ) log Nδ0 (F ) log Nδ (F )

limδ→0 ≤ limδ→0 limδ→0 ,
− log δ − log δ − log δ
so these terms are equal; in other words the value of the expression for lower box-
counting dimension is the same for both Nδ (F ) (using definition (ii) of lower box
dimension), and Nδ0 (F ) (using definition (iv)).
The correspondance of the two definitions of upper box dimension follows in exactly
the same way but taking upper limits.

2.2 Let f satisfy the Hölder condition

|f (x) − f (y)| ≤ c|x − y|α (x, y ∈ F ).

Suppose that F can be covered by Nδ (F ) sets of diameter at most δ. Then the

Nδ (F ) images of these sets under f form a cover of f (F ) by sets of diameter at
most cδ α . Thus
log Ncδα (f (F )) log Nδ (F )
dimB f (F ) = limcδα →0 α
≤ limδ→0
− log cδ −α log δ − log c
1 log Nδ (F ) 1
= limδ→0 = dimB F.
α − log δ α

A similar argument gives the result for lower box dimensions.

2.3 Let Ek denote those numbers in [0, 1] whoseT expansions do not contain the digit
5 in the fist k decimal places. Then F = ∞ k=1 Ek . Let Nδ (F ) denote the least
number of intervals of length δ that can cover F . Let k be the integer such that

10
 10−k ≤ δ < 10−k−1 . Since Ek may be regarded as the union of 9k intervals of lengths
10−k , we get Nδ (F ) ≤ 9k , so

log Nδ (F ) log 9k k log 9 log 9

dimB F = limδ→0 ≤ limk→∞ −k−1
≤ limk→∞ = .
− log δ − log 10 (k + 1) log 10 log 10

Now let 0 < δ < 1 and let k be the integer such that 10−k+1 ≤ δ < 10−k . Since
any set of diameter δ can intersect at most two of the component intervals of Ek
of length 10−k and each such component interval contains points of F , at least 12 9k
intervals of length δ are needed to cover F . Thus Nδ (F ) ≥ 12 9k , so

log Nδ (F ) log 12 9k k log 9 log 9

dimB F = limδ→0 ≥ limk→∞ −k+1
≥ limk→∞ = .
− log δ − log 10 (k − 1) log 10 log 10

We conclude that the box dimension of F exists, with dimB F = log 9/ log 10.

2.4 Let Nδ (F ) denote the smallest number of squares (that is, 2-dimensional cubes) of
side δ that cover F . We will use the fact (see (2.10)) that, if δk = 4−k , then

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

if this limit exists.

It follows from the construction of F shown in figure 0.4 that Nδk (F ) ≤ 4k and so

log Nδk (F ) log 4k

dimB F = limk→∞ ≤ limk→∞ = 1.
− log δk log 4k

On the other hand, any square of side δk = 4−k intersects at most two of the squares
of side δk in Ek . Since F meets every one of the 4k squares which comprise Ek , it
follows that Nδk (F ) ≥ 21 4k , so

log Nδk (F ) log 12 4k

dimB F = limk→∞ ≥ limk→∞
− log δk log 4k
k log 4 − log 2
= limk→∞ = 1.
k log 4
Thus dimB F = 1.

2.5 Let Nδ (F ) denote the smallest number of sets of diameter at most δ that cover F and
let δk = 3−k . For each of the straight line segments that makes up Ek , take a closed
disc of diameter δk , centred at the midpoint of the line. There are 4k such discs and
they cover F , so that (see Equivalent definitions 3.1 and comments following)

log Nδk (F ) log 4k log 4

dimB F = limk→∞ ≤ limk→∞ k
= .
− log δk log 3 log 3

Now let Nδ (F ) denote the largest number of disjoint balls of radius δ with centres
in F . The 4k straight line segments that make up Ek have 4k + 1 distinct endpoints,

11
 each of which belongs to F . Balls of radius 1/3k+1 centred at these endpoints are
mutually disjoint and so, putting δk = 3−(k+1) , we have by Equivalent definition (v),
that
log Nδk (F ) log(4k + 1)
dimB F = limk→∞ ≥ limk→∞
− log δk log 3k+1
log(4k ) k log 4 log 4
≥ limk→∞ k+1
= limk→∞ = .
log 3 (k + 1) log 3 log 3

2.6 Let Nδ (F ) denote the smallest number of squares (that is, 2-dimensional cubes) of
side δ that cover F . For k = 1, 2, . . ., the Sierpiński triangle F can be covered by 3k
squares of side 2−k and so, putting δk = 2−k , we have
log Nδk (F ) log 3k log 3
dimB F = limk→∞ ≤ limk→∞ k
= .
− log δk log 2 log 2
Now let Nδ (F ) denote the largest number of disjoint balls of radius δ with centres in
F . The top vertex of each of the 3k triangles in Ek belongs to F and balls of radius
1/2k+1 centered at these vertices are mutually disjoint. So, putting δk = 2−(k+1) , we
have
log Nδk (F ) log 3k log 3
dimB F = limk→∞ ≥ limk→∞ k+1
= .
− log δk log 2 log 2
Thus dimB F = log 3/ log 2.

2.7 The middle-third Cantor set has 2k gaps of length 3−k−1 for k = 0, 1, 2, . . .. If
1 −k
2
3 < δ ≤ 12 3−k−1 the δ-neighbourhood fills the gaps of lengths 3−k or less, and
has two parts of length δ in the gaps of length 3−k−1 or more. Summing these
lengths over all gaps, and noting that the parts of Fδ at each end of F have length
δ, gives
∞
X k−1
X
i−1 −i
L(Fδ ) = 2 3 + 2δ 2i−1 + 2δ
i=k i=1
 k−1
2
= + 2k δ
3
on summing the geometric series. Hence
 k−1
k k 2
2 δ ≤ L(Fδ ) ≤ 2 δ ≤≤ 4 × 2k δ
3
or
c1 δ 1−log 2/log3 δ ≤ L(Fδ ) ≤ c21 δ 1−log 2/log3 .
Hence Proposition 2.4 gives that dimP F = log 2/log3

2.8 The idea is to construct a set such at some scales a relatively large number of boxes
are needed in a covering and at other scales one can manage with relatively few. We
adapt the middle third Cantor set by deleting the middle 3/5 of intervals at certain
scalesTrather than the middle 1/3. Thus set kn = 10n , for n = 0, 1, 2, . . . and let
E= ∞ k=0 Ek , where E0 = [0, 1], and

12
 • for k0 ≤ k ≤ k1 , k2 < k ≤ k3 , . . ., Ek is obtained by deleting the middle 1/3 of
each interval in Ek−1 ;
• for k1 < k ≤ k2 , k3 < k ≤ k4 , . . ., Ek is obtained by deleting the middle 3/5 of
each interval in Ek−1 .

We estimate the lower and upper box dimensions of E by estimating Nδ (E), the
least number of closed intervals of length δ that can cover E.
(i) If n is even, then Ekn is made up of 2kn intervals of length
 k1  k2 −k1  kn−1 −kn−2  kn −kn−1  kn −kn−1
1 1 1 1 1
δn = ··· < .
3 5 3 5 5

Taking these intervals as a cover

log Nδn (E) log 2kn

dimB E ≤ limn→∞ ≤ limn→∞
− log δn log 5kn −kn−1
kn log 2 10kn−1 log 2 10 log 2
= limn→∞ = limn→∞ = .
(kn − kn−1 ) log 5 9kn−1 log 5 9 log 5

(ii) If n is odd, then Ekn is made up of 2kn intervals of length

 k1  k2 −k1  kn−1 −kn−2  kn −kn−1  kn−1  kn −kn−1
1 1 1 1 1 1
δn = ··· > .
3 5 5 3 5 3

Any interval of length δn meets at most two of the intervals in Ekn and so, since E
has points in every interval in Ekn ,

log Nδn (E) log(2kn /2)

dimB E ≥ limn→∞ ≥ limn→∞
− log δn log(5kn−1 3kn −kn−1 )
kn log 2 − log 2
= limn→∞
kn−1 log 5 + (kn − kn−1 ) log 3
10kn−1 log 2 − log 2
= limn→∞
kn−1 log 5 + 9kn−1 log 3
10 log 2 10 log 2
= ≥ .
log 5 + 9 log 3 11 log 3

Since
10 log 2 10 log 2
<
9 log 5 11 log 3

dimB E < dimB E,

as required.

2.9 The idea is to construct sets E and F such that at every scale one of E or F looks
‘large’ and the other looks ‘small’. Let E be the set described in the Solution to
Exercise 3.8. We construct a set F in a similar way, except that the scaling of
intervals is complementary and the set is positioned to be disjoint from E. Thus
set kn = 10n , for n = 0, 1, 2, . . . and let F = ∞
T
F
k=0 k , where F0 = [2, 3], and

13
 • for k0 ≤ k ≤ k1 , k2 < k ≤ k3 , . . ., Fk is obtained by deleting the middle 3/5 of
each interval in Fk−1 ;
• for k1 < k ≤ k2 , k3 < k ≤ k4 , . . ., Fk is obtained by deleting the middle 1/3 of
each interval in Fk−1 .

As in the solution to Exercise 2.8 we get that

10 log 2
dimB E, dimB F ≤ .
9 log 5

For each k = 1, 2, . . ., let δk denote the length of the longest intervals in Ek ∪ Fk —

there are 2k such intervals, each of which meets E ∪ F . Since any other interval of
length δk meets at most two of these intervals, it follows that the smallest number
of closed intervals of length δk that cover E ∪ F satisfies Nδk (E ∪ F ) ≥ 2k /2. Now

k/2 1
k/2
δk ≥ 31 5
and δk ≥ (1/5)δk−1 , so by the note after Definitions 3.1
log Nδk (E ∪ F ) log 2k /2
dimB E ∪ F = limδk →0 ≥ limk→∞
− log δk log 5k/2 log 3k/2
k log 2 − log 2 2 log 2 10 log 2
= limk→∞ = > .
(k/2) log 5 + (k/2) log 3 log 5 + log 3 9 log 5

2.10 Since F is a countable set, dimH F = 0.

The box dimension calculation is similar to Example 2.7. Let Nδ (F ) be the smallest
number of sets of diameter at most δ that cover F . If |U | = δ < 1/2 and k is the
integer satisfying
2k − 1 1 1 1 1 2k + 1
= − 2 >δ≥ 2 − = ,
k 2 (k
− 1)2 (k − 1) 2 k k (k + 1) 2 (k + 1)2 k 2
then U can cover at most one of the points {1, 41 , . . . , k12 }. Thus Nδ (F ) ≥ k and
hence
log Nδ (F ) log k
dimB F = limδ→0 ≥ limk→∞ 2 2
− log δ log (k+1) k
2k+1
log k 1
= limk→∞ = .
2 log(k + 1) + 2 log k − log 2 − log(k + 1/2) 3
On the other hand, if
2k − 1 2k + 1
>δ≥ ,
k 2 (k
− 1)2 (k + 1)2 k 2
then k + 1 intervals of length δ cover [0, 1/k 2 ] leaving k − 1 points of F which can
be covered by k − 1 intervals of length δ. Thus
log Nδ (F ) log 2k
dimB F = limδ→0 ≤ limk→∞ 2 2
− log δ log (k−1) k
2k−1
log 2k 1
= limk→∞ = .
2 log(k − 1) + 2 log k − log 2 − log(k − 1/2) 3
Thus dimB F = 1/3.

14
2.11 The von Koch curve F has (upper and lower) box dimensions equal to log 4/ log 3.
Moreover, by virtue of the self-similarity of F , dimB (F ∩ V ) = log 4/ log 3 for every
open set V that intersects F . By Proposition 2.8, dimMB F = dimB F = log 4/ log 3.

2.12 Recall that the divider dimension of a curve C is defined as limδ→0 log Mδ (C)/−log δ
(assuming that this limit exists), where Mδ (C) is the maximum number of points
x0 , x1 , . . . , xm on C, in that order, with |xi − xi−1 | ≥ δ for i = 1, 2, . . . , m.
By inspection of the von Koch curve C, taken to have of baselength 1, (see Figure
0.2), we have that if k is the integer such that 3−k−1 ≤ δ < 3−k , then 4k < 4k + 1 ≤
Mδ (C) ≤ 4k+1 + 1 < 4k+2 . Then
k log 4 log(4k ) log Mδ (C) log(4k+2 ) (k + 2) log 4
= ≤ ≤ .
(k + 1) log 3 − log(3−k−1 ) − log δ − log(3−k ) log 3
As δ → 0, k → ∞, so taking limits gives that
log Mδ (C) log 4
divider dimension = lim =
δ→0 − log δ − log 3
(which, of course equals the Hausdorff and box dimensions of C).

2.13 Recall that the divider dimension of a curve C is defined as limδ→0 log Mδ (C)/−log δ
(assuming that this limit exists), where Mδ (C) is the maximum number of points
x0 , x1 , . . . , xm on C, in that order, with |xi − xi−1 | ≥ δ for i = 1, 2, . . . , m.
Consider Equivalent definition 3.1(v) of box dimension, taking Nδ (C) to be the
greatest number of disjoint balls of radius δ with centres on C. Then if B1 , . . . , BNδ (C)
is a maximal collection of disjoint balls of radii δ with centres on C, every ball Bi
must contain at least one point xj in any maximal sequence of points x0 , x1 , . . . , x1
for the divider dimension, otherwise the centre of Bi may be added to the sequence
to increase its length. Thus Nδ (C) ≤ Mδ (C), so
Nδ (C) log Mδ (C)
≤ ,
− log δ − log δ
and taking limits as δ → 0 gives that the box dimension is less than or equal to
the divider dimension, assumming both exist. (If not a similar inequality holds for
lower and upper box and divider dimensions.)

2.14 The middle λ Cantor set F may be constructed from the unit interval by removing
2k open intervals of lengths λ( 12 (1 − λ))k for k = 0, 1, 2, . . .. Thus, denoting these
complementary intervals by Ii , we have
∞  ks
X
s
X
k s 1
|Ii | = 2 λ (1 − λ) .
i k=0
2

This is a geometric series which converges if and only if the common ratio 2( 21 (1 −
λ))s < 1, that is if s > log 2/ log(2/(1 − λ)), a number equal to box dimensions (and
Hausdorff dimension) of F , see also Exercise 3.14.

15
 Chapter 3

3.1 Put
s X
Hδ (F ) = inf{ |Ui |s : {Ui } is a δ-cover of F by closed sets}.
i

Since we have reduced the class of permissible covers by restricting to covers by

s
closed sets, we must have Hδ (F ) ≥ Hδs (F ). Now suppose that {Ui } is a δ-cover of
F . Since the closure U i of
PUi satisfiesP |U i | = |Ui |, it follows that {U i } is a δ-cover
of F by closed sets with i |U i |s = i |Ui |s . Since this is true for every δ-cover of
s s
F , it follows that Hδ (F ) ≤ Hδs (F ). Thus Hδ (F ) = Hδs (F ) for all δ > 0 and so the
value of Hs (F ) = limδ→0 Hδs (F ) is unaltered if we only consider δ-covers by closed
sets.

3.2 SupposeP that {Ui } is a δ-cover of F . For any set Ui in the cover we have |Ui |0 = 1
and so i |Ui |0 is equal to the number of sets in the cover. Thus Hδ0 (F ) is the
smallest number of sets that form a δ-cover of F .
If F has k points, x1 , x2 , . . . , xk , then the k balls of radius δ/2 with centers at
x1 , x2 , . . . , xk form a δ-cover of F and so Hδ0 (F ) ≤ k. Moreover, if δ > 0 is so
small that |xi − xj | > δ for all i 6= j, then any δ-cover of F must contain at least
k sets and so Hδ0 (F ) ≥ k. So, for δ small enough, we have Hδ0 (F ) = k and hence
H0 (F ) = limδ→0 Hδ0 (F ) = k.
Finally, if F has infinitely many points, then for each positive integer k, we can take
a set Fk ⊂ F such that Fk has k points. Then H0 (F ) ≥ H0 (Fk ) = k for all k, so
H0 (F ) = ∞.

3.3 Clearly, for every 0 <  ≤ δ, we may cover the empty set with a single set of diameter
, so 0 ≤ Hδs (∅) ≤ s for all  > 0, giving Hδs (∅) = 0. Thus Hs (F ) = limδ→0 Hδs (F ) =
0.
If E ⊂ F , every δ-cover of F is also a δ-cover of E, so taking infima over all δ-covers
gives Hδs (E) ≤ Hδs (F ) for all δ > 0. Letting δ → 0 gives Hδ (E) ≤ Hδ (F ).
n
Now
P∞ letsF1 , F2 , . . . be subsets of R . Without loss of generality, we may assume that
P∞i=1 Hδ (Fi ) < ∞. For  > 0 let {Ui,j : j = 1, 2, . . .} be a δ-cover of Fi such that
s s −i
S∞j=1 |Ui,j | ≤ Hδ (Fi ) + 2 . Then {Ui,j : i = 1, 2, . . . , j = 1, 2, . . .} is a δ-cover of
i=1 Fi and

∞
! ∞ X ∞ ∞  ∞ ∞
s
[ X
s
X
s  X
s
X
Hδ Fi ≤ |Ui,j | ≤ Hδ (Fi ) + i = + Hδ (Fi ) ≤ + Hs (Fi ).
i=1 i=1 j=1 i=1
2 i=1 i=1

Since this is true for every  > 0, it follows that

∞
[ [∞ ∞
X
s s
H( Fi ) = lim Hδ ( Fi ) ≤ Hs (Fi )
δ→0
i=1 i=1 i=1

as required.

16
3.4 Note that in calculating Hδs ([0, 1]) it is enough to consider coverings by intervals.
If 0 ≤ s < 1 and {Ui } is a δ-cover of [0, 1] by intervals, then
X X X
1≤ |Ui | = |Ui |1−s |Ui |s ≤ δ 1−s |Ui |s .
i i i

Hence Hδs ([0, 1]) ≥ δ s−1 , so letting δ → 0 gives Hs ([0, 1]) = ∞.

For s > 1, we may cover [0, 1] by at most (1 + 1/δ) intervals of length δ, so
Hδs ([0, 1]) ≤ (1 + 1/δ)δ s → 0
as δ → 0, so Hs ([0, 1]) = 0.
|Ui |, so Hδ1 ([0, 1]) ≥ 1,
P
For s = 1, if {Ui } is a δ-cover of [0, 1] by intervals, then 1 ≤ i
and letting δ → 0 gives H1 ([0, 1]) ≥ 1.
Taking a cover [0, 1] by at most (1 + 1/δ) intervals of length δ,
Hδ1 ([0, 1]) ≤ (1 + 1/δ)δ → 1
as δ → 0, so H1 ([0, 1]) ≤ 1. We conclude that H1 ([0, 1]) = 1.

3.5 First suppose that F is bounded, say F ⊂ [−m, m]. By the mean value theorem,
for some z ∈ [−m, m],
|f (x) − f (y)| = |f 0 (z)||x − y| ≤ ( sup |f 0 (z)|)|x − y|
z∈[−m,m]

Since f 0 (z) is continuous it is bounded on [−m, m]. Thus f is LipschitzS on F , so

dimH f (F ) ≤ dimH (F ) by Proposition 3.3. For arbitrary F ⊂ R, f (F ) = ∞
m=1 f (Fn ∩
[−m, m]), so by countable stability
dimH f (F ) = sup dimH f (Fn ∩ [−m, m]) ≤ sup dimH (Fn ∩ [−m, m]) ≤ dimH F,
m m

by the bounded case.

3.6 Let Fk = F ∩ [1/k, k]. If x, y ∈ Fk , then

|f (x) − f (y)| = |x2 − y 2 | = |x − y||x + y|
and so
2
|x − y| ≤ |f (x) − f (y)| ≤ 2k|x − y|.
k
Thus f is a bi-Lipschitz map on Fk and so, by Proposition 3.3, dimH f (Fk ) =
dimH Fk . Similarly, if Gk = F ∩ [−k, −1/k], then dimH f (Gk ) = dimH Gk .
Now F = (F ∩ {0}) ∪ ∞
S S∞
k=1 (Fk ∪ Gk ) and f (F ) = f (F ∩ {0}) ∪ k=1 (f (Fk ) ∪ f (Gk )).
Since F ∩ {0} and f (F ∩ {0}) contain at most one point, they both have zero
dimension. Thus, by countable stability,
dimH F = sup{dimH Fk , dimH Gk : k = 1, 2, . . .}
= sup{dimH f (Fk ), dimH f (Gk ) : k = 1, 2, . . .} = dimH f (F ).

[Note that this result is not true for box dimension. For example, using Example
2.7 and Exercise 2.10 we see that dimB f (F ) 6= dimB F when F = {0, 1, 21 , 13 , . . .}.]

17
 3.7 Define g : [0, 1] → graphf by g(x) = (x, f (x)). We claim that g is bi-Lipschitz. For:

|g(x) − g(y)|2 = |x − y|2 + |f (x) − f (y)|2

so
|x − y|2 ≤ |g(x) − g(y)|2 ≤ |x − y|2 + c2 |x − y|2 = (1 + c2 )|x − y|2
since |f (x) − f (y)| ≤ c|x − y| for some c > 0. Thus g is bi-Lipschitz, so 1 =
dimH ([0, 1]) = dimH g([0, 1]) = dimH graphf .

3.8 Both {0, 1, 2, 3, . . .} and {0, 1, 21 , 13 , . . .} are countable sets, so have Hausdorff dimen-
sion 0.

3.9 Note that F splits into 9 parts Fi = F ∩ [i/10, (i + 1)/10] for i = 0, 1, 2, 3, 4, 6, 7, 8, 9,

these parts disjoint except possibly for endpoints which have s-dimensional measure
0 if s > 0. It follows from Scaling Property 2.1 that, for s > 0, Hs (Fi ) = 10−s Hs (F )
for all i. Summing, and using that F is essentially a disjoint union of the Fi , it
follows that, for s > 0,
X
Hs (F ) = Hs (Fi ) = 9 × 10−s Hs (F ).
i=0,1,2,3,4,6,7,8,9

If we assume that when s = dimH F we have 0 < Hs (F ) < ∞, then, for this
value of s, we may divide through by Hs (F ) to obtain 1 = 9 × 10−s and hence
s = dimH F = log 9/ log 10.

3.10 Note that, for i, j = 0, 1, 2, 3, 4, 6, 7, 8, 9 the sets F ∩ ([i/10, (i + 1)/10] × [j/10, (j +

1)/10]) are scale 1/10 similar copies of F . By the addition and scaling properties of
Hausdorff measure,
X
Hs (F ) = Hs (F ∩ ([i/10, (i + 1)/10] × [j/10, (j + 1)/10]) = 92 10−s Hs (F ),
i,j6=5

provided 0 < Hs (F ) < ∞ when s = dimH F , in which case 1 = 92 10−s , giving

s = 2 log 9/ log 10.

3.11 The set F comprises one similar copy of itself at scale 21 , say F0 , and four simi-
lar copies at scale 14 , say F1 , F2 , F3 , F4 . By the additive and scaling properties of
Hausdorff measure, noting that the Fi intersect only in single points,
4  s  s
s s
X
s 1 s 1
H (F ) = H (F0 ) + H (Fi ) = H (F ) + 4 Hs (F )
i=1
2 4

for s > 0. Provided 0 < Hs (F ) < ∞ when s = dimH F , we have 1 = ( 12 )s + 4( 21 )s .

s
Thus 4( 21 )2 + ( 12 )s − 1 = 0; solving this quadratic equation in ( 12 )s gives ( 12 )s =
√ √
(−1 + 17)/8 as the positive solution, so s = (log 8 − log( 17 − 1))/ log 2.

18
3.12 F is the union of countably many translates of the middle third Cantor set, all of
wich have Hausdorff dimension log 2/ log 3, so dimH F = log 2/ log 3 using countable
stability.

3.13 F is the union, over all finite sequences a1 , a2 , . . . , ak of the digits 0, 1, 2, of similar
copies of the middle-third Cantor set scaled by a factor 3−k and translated to have
left hand end at 0.a1 a2 . . . ak , to base 3. Thus F is the union of countably many
similar copies of the Cantor set, so dimH F = log 2/ log 3 using countable stability.

3.14 The set F is the union of two disjoint similar copies of itself, FL , FR , say, at scales
1
2
(1 − λ). By the additive and scaling properties of Hausdorff measure
 s
s s s 1
H (F ) = H (FL ) + H (FR ) = 2 (1 − λ) Hs (F )
2
for s ≥ 0. Provided 0 < Hs (F ) < ∞ when s = dimH F , we have 1 = 2( 12 (1 − λ))s ,
giving that dimH F = log 2/ log(2/(1 − λ)).

The set E is the union of four disjoint similar copies of itself, E1 , E2 , E3 , E4 , say, at
scales 21 (1 − λ). By the additive and scaling properties of Hausdorff measure
4  s
s
X
s 1
H (F ) = H (Fi ) = 4 (1 − λ) Hs (F )
i=1
2

for s ≥ 0. Provided 0 < Hs (F ) < ∞ when s = dimH F , we have 1 = 4( 12 (1 − λ))s ,

giving that dimH F = log 4/ log(2/(1 − λ)) = 2 log 2/ log(2/(1 − λ)).

3.15 Take the unit square E0 and divide it into 16 squares of side 1/4. Now take 0 < r <
1/4, put a square of side r in the middle of each of the 16 small squares and discard
everything that is not inside one of these squares, to get a set E1 .
Keep on repeating this process so that, at the T k-th stage, there is a collection Ek
k k
of 16 disjoint squares of side r . Then Fr = k Ek is a totally disconnected subset
of R2 . (If two points x, y are in the same component of Fr , then they √ must belong
k
to the same square in Ek , for all k = 1, 2, . . .. Thus |x − y| ≤ 2r , for each
k = 1, 2, . . ., and hence |x − y| = 0 so that x = y.)
The set Fr is made up of 16 disjoint similar copies of itself, each scaled by a factor
r, denote these sets as Fr,1 , . . . , Fr,16 . It follows from Scaling property 2.1 that, for
s ≥ 0,
X16 X16
s s
H (Fr ) = H (Fr,i ) = rs Hs (Fr ).
i=1 i=1

Assuming that when s = dimH Fr we have 0 < Hs (Fr ) < ∞ (using the heuristic
method), then, for this value of s we may divide by Hs (Fr ) to obtain 1 = 16rs and
so s = dimH Fr = − log 16/ log r. As r increases from 0 to 1/4, dimH Fr increases
from 0 to 2, taking every value in between.
A set consisting of a single point gives a totally disconnected subset of R2 of Haus-
dorff dimension zero. It remains to show that there exists a totally disconnected

19
 subsetS∞of R2 of Hausdorff dimension two. For one of many ways to do this, let
G = k=5 Gk , where Gk = F1/4−1/k + (k, 0). The sets Gk are disjoint and hence G
is a totally disconnected subset of R2 . By countable stability, we have
dimH G = sup dimH Gk = sup dimH (F1/4−1/k ) = 2,
5≤k≤∞ 5≤k≤∞

using that Gk is congruent to F1/4−1/k , whose dimension tends to 2 as k → ∞.

1
3.16 Note that F is just a copy of the middle-third Cantor set scaled by 3
π. Thus
dimH F = log 2/ log 3.

√
3.17 Let E = [0, 1] ∩ Q and F = {x ∈ [0, 1] : x − 2 ∈ Q}, so that E and F are disjoint
dense subsets of [0, 1]. If {Bi } is any collection of disjoint balls (i.e. intervals) with
centres
P in E and radii at most δ, then, by considering the lengths of the Bi , we see
that i |BiP| ≤ 1 + δ. Moreover, taking Bi as nearly abutting intervals of lengths 2δ
we can get i |Bi | ≥ 1. Thus, since
( )
X
1
Pδ (E) = sup |Bi | : {Bi }are disjoint balls of radii ≤ δ with centres in F ,
i

we get 1 ≤ Pδ1 (E) ≤ 1 + δ. Letting δ → 0 gives P01 (E) = 1. In a similar way,

P01 (E) = 1 and P01 (E ∪ F ) = 1. In particular P01 (E ∪ F ) 6= P01 (E) + P01 (F ).

3.18 We use the notation of section 2.5. Let Ui be a δ-cover of F . Then

X X X
h(|Ui |) = (h(|Ui |)/g(|Ui |))g(|Ui |) ≤ η(δ) g(|Ui |)
where η(δ) = sup0<t≤δ h(t)/g(t). Taking infima, Hδh (F ) ≤ η(δ)Hδg (F ). Letting
δ → 0, then η(δ) → 0, and Hδg (F ) → Hg (F ) < ∞, so Hδh (F ) → 0, that is
Hh (F ) = 0.

3.19 If F1 ⊂ F2 then any δ-cover of F2 by rectangles is also a δ-cover of F1 , so that from

the definition, Hδs,t (F1 ) ≤ Hδs,t (F2 ), and letting δ → 0 gives Hs,t (F1 ) ≤ Hs,t (F2 ).
In particular, if (s, t) ∈ printF1 then 0 < Hs,t (F1 ) so 0 < Hs,t (F2 ) giving (s, t) ∈
printF2 . Thus printF1 ⊂ printF2 .
S∞ S∞
S∞ at once that printFk ⊂ print( i=1 Fi ) for all k, so that i=1 printFi ⊂
It follows
print( i=1 Fi ).
/ printFi for all i. Then Hs,t (Fi ) = 0 S ( ∞
for all i, so Hs,tS
S
Now suppose (s, t) ∈ i=1 Fi ) =
s,t ∞ ∞
0 sinceSH is a measure. We conclude that (s, t) ∈ / print( i=1 Fi ). Thus i=1 printFi =
∞
print( i=1 Fi ).
Suppose now that s0 + t0 ≤ s + t and t0 ≤ t. For any δ-cover of a set F by rectangles
Ui with sides a(Ui ) ≥ b(Ui ), we have
0 0 0 0
X X
a(Ui )s b(Ui )t ≤ a(Ui )s−s a(Ui )s b(Ui )t−t b(Ui )t
i i
0 0 0 0
X
≤ a(Ui )s b(Ui )t a(Ui )s−s +t−t
i
(s+t)−(s0 +t0 ) 0 0
X
≤ δ a(Ui )s b(Ui )t .
i

20
 0 0
It follows from the definition that if 0 < δ < 1 then Hδs,t (F ) ≤ Hδs ,t (F ), so Hs,t (F ) ≤
0 0 0 0
Hs ,t (F ). Thus if (s, t) ∈ printF then 0 < Hs,t (F ) ≤ Hs ,t (F ), so (s0 , t0 ) ∈ printF .
Since print(F1 ∪ F2 ) = printF1 ∪ printF2 , taking F1 and F2 such that the union
of their dimension prints is not convex will give a set F1 ∪ F2 with non-convex di-
mension print. Taking F1 a circle and F2 the product of uniform Cantor sets of
dimensions 31 and 34 , will achieve this, see figure 3.4.

3.20 Let F be the Sierpiński triangle of side 1. Let x ∈ F and let

√ √
2−k−1 3 ≤ r ≤ 2−k 3

for some positive integer

√ k ≥ 1. From the geometry of the Sierpiński triangle
−k−1
B(x, r) ⊇ B(x, 2 3 contains an empty √ equilateral triangle of side 2−k−1 which
contains an inscribed disc of radius 2−k−1 3/4 ≥ r/8. Thus por(F, x, r) ≥ 18 , so F
is 18 -uniformly porous (and so 91 -uniformly porous).

21
Chapter 4

4.1 We begin by noting that the Cantor tartan is equal to (F × R) ∪ (R × F ). In

Example 4.3 it is shown that dimH (F × [0, 1]) = 1 + log 2/ log 3. Now F × [n, n + 1]
is a translate of F ×[0, 1], so for each integer n, dimH (F ×[n, n+1]) = 1+log 2/ log 3.
Thus, by countable stability, dimH (F × R) = 1 + log 2/ log 3. As R × F is congruent
to F × R under a 90 degree rotation, dimH (R × F ) = 1 + log 2/ log 3. Finally,

log 2
dimH ((F × R) ∪ (R × F )) = max{dimH (F × R), dimH (R × F )} = 1 + .
log 3

4.2 Let F be the set of numbers in [0, 1] containing only even digits. Writing Ek for the
set of numbers T
in [0, 1] containing only even digits in the first k decimal places, we
have that F = ∞ k
k=0 Ek . For each positive integer k, the 5 intervals in Ek of length
log 5/ log 10
10−k form a 10−k -cover of F and so H10−k (F ) ≤ 5k (10−k )log 5/ log 10 = 5k 5−k = 1.
Letting k → ∞ gives Hlog 5/ log 10 (F ) ≤ 1.
Now let µ be the natural mass distribution on F obtained by repeated subdivision
of mass into 5 equal parts, so that each of the 5k intervals of length 10−k in Ek
carries a mass of 5−k . If 10−(k+1) ≤ |U | < 10−k for some k ≥ 1, then U can intersect
at most one of the intervals in Ek and so

µ(U ) ≤ 5−k = (10−k )log 5/ log 10 ≤ (10|U |)log 5/ log 10 = 5|U |log 5/ log 10 .

It follows from the Mass distribution principle 4.2 that Hlog 5/ log 10 (F ) ≥ 1/5. Com-
bining these results, we see that Hlog 5/ log 10 (F ) is positive and finite, so that dimH F =
log 5/ log 10.

4.3 Let F be the Cantor dust depicted in√figure 0.4. For each positive integer k, the
nk = 4k squares of diameter δk = 4−k 2 in Ek form a cover of F and so it follows
from Proposition 4.1 that

log 4k k log 4
dimH F ≤ limk→∞ √ = limk→∞ √ = 1.
− log(4 −k 2) k log 4 − log 2

Now let µ be the natural mass distribution on F , so that each of the 4k squares of
side 4−k in Ek carries a mass of 4−k . If 4−(k+1) ≤ |U | < 4−k for some k ≥ 1, then U
can intersect at most one of the squares in Ek and so

µ(U ) ≤ 4−k ≤ 4|U |.

It follows from the Mass distribution principle 4.2 that dimH F ≥ 1. Combining
these results, we deduce that dimH F = 1 as required.

1
4.4 If λ = 2
then F = [0, 1] so dimH F = dimB F = 1.
Thus assume 0 < λ < 12 . It is easy to see that F is a subset of the interval
[0, λ/(1 − λ)] and moreover F is the union of two similar copies of itself at scale λ,
that is F = (F ∩[0, λ2 /(1−λ)])∪(F ∩[λ, λ/(1−λ)]). Thus F may be constructed by
a Cantor-type construction, repeatedly replacing intervals by a pair of subintervals

1
 each of length λ times that of the parent interval. Let Ek be the set of 2k intervals
k+1
T∞lengths λ /(1 − λ) obtained at the kthk stage of this construction. Then F =
of
k=0 Ek . For each positive integer k, the 2 intervals in Ek form a cover of F and
so it follows from Proposition 4.1 that
log 2k k log 2 log 2
dimB F ≤ lim k+1
= lim = .
k→∞ − log λ /(1 − λ) k→∞ −(1 + k) log λ + log(1 − λ) − log λ

Now let µ be the natural mass distribution on F so that each of the 2k intervals of
lengths λk+1 /(1 − λ) in Ek carries a mass of 2−k . If λk+1 (1 − 2λ)/(1 − λ) ≤ |U | <
λk (1 − 2λ)/(1 − λ) for some k ≥ 1, then U can intersect at most one of the intervals
in Ek and so
 − log 2/ log λ
−k k − log 2/ log λ 1−λ
µ(U ) ≤ 2 ≤ (λ ) ≤ |U | ≤ c|U |− log 2/ log λ .
λ(1 − 2λ)
where c > 0 is independent of U . It follows from the Mass distribution princi-
ple 4.2 that dimH F ≥ − log 2/ log λ. Combining these results, we deduce that
dimH F = dimB F = − log 2/ log λ.

4.5 Let Ek0 be the set of 2k intervals of lengths 3−k obtained at the kth stage of the
usual construction of the middle third Cantor √ set. Then Ek = Ek0 T × Ek0 consists
of 4k squares of sides 3−k and diameters 3−k 2, and F × √ F = ∞ k=0 Ek . For
each positive integer k,√ the 4k squares in Ek form a 3−k 2-cover of F and so
log 4/ log 3
H3−k √2 (F ) ≤ 4k (3−k 2)log 4/ log 3 = 2log 2/ log 3 4k 4−k = 2log 2/ log 3 . Letting k → ∞
gives Hlog 4/ log 3 (F ) ≤ 2log 2/ log 3 .
Now let µ be the natural mass distribution on F × F obtained by repeated subdivi-
sion of mass into 4 equal parts, so that each of the 4k squares in Ek carries a mass
of 4−k . If 3−(k+1) ≤ |U | < 3−k for some k ≥ 1, then U intersects at most one of the
squares in Ek and so
µ(U ) ≤ 4−k = (3−k )log 4/ log 3 ≤ (3|U |)log 4/ log 3 = 4|U |log 4/ log 3
It follows from the Mass distribution principle 4.2 that Hlog 4/ log 3 (F ) ≥ 1/4. Com-
bining these results, we see that Hlog 4/ log 3 (F ) is positive and finite, so that dimH F =
log 4/ log 3.

4.6 With F the middle third Cantor set, it is easily checked that
F0 ≡ (F ∩ [ 32 , 1]) × [0, 13 ] ⊂ {(x, y) ∈ R2 : x ∈ F and 0 ≤ y ≤ x2 } ⊂ F × [0, 1] ≡ F1 .
We showed in Example 4.3 that dimH F1 = 1 + log 2/ log 3, and F0 is a similar copy
of F1 at scale 13 , so dimH F0 = 1 + log 2/ log 3. It is immediate that the dimension of
the intermediate set is also 1 + log 2/ log 3.

4.7 Let F be the set described in Example 4.5. For each positive integer k, the nk = mk
intervals of length δk = λk in Ek form a cover of F and so Proposition 4.1 gives that
log mk log m
dimH F ≤ dimB F ≤ limk→∞ k
= .
− log λ − log λ

2
 log m/(− log λ)
Since nk δk = mk λk log m/(− log λ) = 1, it also follows from Proposition 4.1
that Hlog m/(− log λ) (F ) < ∞.
Now let µ be the natural mass distribution on F so that each of the mk intervals of
length λk in Ek carries a mass of m−k . If λ(k+1) ≤ |U | < λk for some k ≥ 1, then U
can intersect at most two of the intervals in Ek , so

µ(U ) ≤ 2m−k = 2mm−(k+1) = 2mλ(k+1) log m/−log λ ≤ 2m|U |log m/−log λ .

It follows from the Mass distribution principle 4.2 that Hlog m/−log λ (F ) ≥ µ(F )/2m =
1/(2m) > 0 and dimH F ≥ log m/ − log λ. Combining these results, we deduce that
dimH F = dimB F = log m/ − log λ and 0 < Hlog m/−log λ (F ) < ∞.

4.8 We show inductively that there are integers m and a2 , a3 , . . . with 0 ≤ aj ≤ j − 1,

such that for all k = 1, 2, . . ., x may be expressed in the form
a2 ak
x=m+ + ... + + yk (∗)
2! k!
where 0 ≤ yk < 1/k! . This is clear when k = 1 (expressing x as in integer plus
a fractional part), so suppose inductively that (∗) holds for some k ≥ 1. Since
0 ≤ k!yk < 1, we may write k!yk = ak+1 /(k + 1) + zk+1 where ak+1 is an integer and
0 ≤ zk+1 < 1/(k + 1). Thus yk = ak+1 /(k + 1)! + yk+1 , where 0 ≤ yk+1 = zk+1 /k! <
1/(k + 1)!. Substituting into (∗) gives the same formula with k replaced by k + 1,
completing the inductive step.
Letting k → ∞ in (∗) gives
a2 a3
x=m+ + + ...,
2! 3!
convergence of the series following from comparison with the exponential series.
To find the dimension of
a2 a3
F = {x = m + + + . . . : m = 0 and ak is even for k = 2, 3, . . .}
2! 3!
we use the result of Example 4.6. Writing
a2 a3
Ek = {x = + + . . . : a2 , . . . , ak are even},
2! 3!
we have F = ∞
T
k=1 Ek as in the general construction (4.3). Each interval of Ek−1
contains mk = d 12 ke intervals of Ek , with each interval of Ek of length 1/k! and
separated by at least k = 1/k!. The formula of Example 4.6 gives

log m1 . . . mk−1 log 2.2.3.3 . . . d 12 (k − 1)e

dimH F ≥ limk→∞ = limk→∞
− log(mk k ) − log(d 12 ke/k!)
log(d 12 (k − 2)e!2 ) 2( 14 k log k)
≥ limk→∞ ≥ limk→∞ 1 =1
log k! − log 21 k 2
k log k − log 1
2
k

where we have used Stirling’s formula in the form log n! ∼ 21 log 2π+ 12 (n+ 21 ) log n−n.
We conclude that dimH F = 1.

3
 4.9 An easy way to do this is as follows. For each 0 < s < 1, there is a compact set
Es ⊂ [0, 1] such that dimH Es = s. (For example, the ‘middle λ Cantor set’, see
after Example 4.5, has Hausdorff dimension log 2/ log(2/(1 − λ)) for 0 < λ < 1,
so taking λ = 1 − 21−1/s gives a suitable set Es .) For n = 1, 2, . . . let Fn be
Sn∞⊂ [1/n, 1/(n + 1)]. Then
a similar copy of E1−1/n , scaled and translated so F
dimH Fn = 1 − 1/n and H1 (Fn ) = 0. Set F = {0} ∪
P∞ n=1 Fn . Then F is compact,
1 1
dimH F = sup1≤n<∞ dimH Fn = 1, and H (F ) = n=1 H (Fn ) = 0, as required.

4.10 First we show that if Hs (F ) = ∞ then for every c with 0 < c < ∞ there is a
Borel E ⊂ F such that c < Hs (E) < ∞. For suppose to the contrary. Then a ≡
sup{Hs (E) : E is a Borel subset of F with Hs (E) < ∞} < ∞. There is a Borel set
E with Hs (E) = a (for if we take a sequence of Borel sets Ek with Hs (Ek ) % a then
Hs ( ∞ s s s
S
k=1 Ek ) = a). Thus F \E is a Borel set with H (F \E) ≥ H (F )−H (E) = ∞.
By Theorem 4.10, F \ E has a Borel subset G with Hs (G) > 0, so E ∪ G is a union
of disjoint Borel sets, so is Borel, with ∞ > Hs (E ∪ G) = Hs (E) + Hs (G) > a, a
contradiction.
Thus, given F with Hs (F ) = ∞ and 0 <Sc < ∞ let E0 be a Borel subset of F such
that c < Hs (E0 ) < ∞. Since E0 = E0 ∩ ∞ n=1 [−n, n] there is an integer n such that
c ≤ H (E0 ∩ [−n, n]) < ∞. The function φ given by φ(x) = Hs (E0 ∩ [−n, x]) is
s

continuous and increasing in x, by continuity of finite measures, see Exercise 1.18.

Since φ(−n) = 0 and φ(n) > c, the intermediate value theorem gives that there is
an x, −n < x < n such that φ(x) = c. Thus E = E0 ∩ [−n, x] is a Borel subset of
F with Hs (E) = c, as required.

4.11 Consider the usual construction of the middle third Cantor set F , the kth stage of
the approximation Ek comprising 2k intervals of lengths 3−k . Let µ be the natural
mass distribution on F , so that each interval I of Ek has µ(I) = 2−k . For each
distinct x, y ∈ F there is an integer k ≥ 1 such that x and y are in the same interval
of Ek−1 but with x ∈ I and y ∈ I 0 with I and I 0 distinct intervals of Ek . We
decompose the energy integral into a sum of pairs of intervals of this type. Thus for
s>0
Z Z ∞ Z Z
dµ(x)dµ(y) X X dµ(x)dµ(y)
s
≤
|x − y| k=1 I6=I 0 ;I,I 0 ∈E x∈I y∈I 0 |x − y|s
k
∞ −k −k
X 2 2
≤ 2k
k=1
3−ks
∞  s k
X 3
= .
k=1
2

This series converges if and only if s < log 2/ log 3 (in which case the energy is at
most (3s/2 )/(1 − (3s/2 )), so by Theorem 4.13, dimH F ≥ log 2/ log 3.

4
 Chapter 5
5.1 Let F be a Borel subset of R2 with 0 < L2 (F ) < ∞, so F is a 2-set. Assume
for the time being that F is bounded, say F ⊂ B for some disc B. Noting that
H2 (A) = cL2 (A) for a constant c > 0, (5.3) and Proposition 5.1(a) applied to F
and then B \ F gives
L2 (F ∩ B(x, r))
lim =0
r→0 L2 (B(x, r))
for almost all x ∈
/ F and
L2 (F ∩ B(x, r)) L2 (B ∩ B(x, r)) L2 ((B \ F ) ∩ B(x, r))
lim = lim − lim
r→0 L2 (B(x, r)) r→0 L2 (B(x, r)) r→0 L2 (B(x, r))
=1−0
for almost all x ∈ / B \ F so for almost all x ∈ F . This is the Lebesgue density
theorem for bounded F . For unbounded F , we have the result for F ∩ B for every
disc B, so since the Lebesgue density at x depends only on F in an neighbourhood
of x, the result follows for all Borel sets F .

5.2 Let x ∈ R with f 0 (x) = c. Given 0 <  < c, there exists δ > 0 such that if |y−x| < δ,
then c −  < f 0 (y) < c +  by continuity of f 0 . By the mean value therom
|f (y) − f (z)| = |f 0 (w)||y − z|
for y, z ∈ B(x, δ), for some w ∈ B(x, δ), so f is bi-Lipschitz on B(x, δ) with
(c − )|z − y| ≤ |f (y) − f (z)| ≤ (c + )|y − z|.
Thus if 0 < r < δ, we have
B(f (x), (c − )r) ⊂ f (B(x, r)) ⊂ B(f (x), (c + )r)
and by (2.9)
(c − )s Hs (F ∩ B(x, r)) ≤ Hs (f (F ∩ B(x, r))) ≤ (c + )s Hs (F ∩ B(x, r)).
Hence
Hs (f (F ) ∩ B(f (x), (c − )r)) s Hs (f (F ) ∩ B(f (x), (c + )r))
≤ H (F ∩ B(x, r)) ≤
(c + )s (c − )s
so s
Hs (f (F ) ∩ B(f (x), (c − )r)) Hs (F ∩ B(x, r))

c−
≤
c+ (2(c − )r)s (2r)s
 −s s
c+ H (f (F ) ∩ B(f (x), (c + )r))
≤ .
c− (2(c + )r)s
Taking limits as r & 0 gives
 s  s
c− s s c+
D (f (F ), f (x)) ≤ D (F, x) ≤ Ds (f (F ), f (x)).
c+ c−
This is true for all  > 0, so Ds (f (F ), f (x)) = Ds (F, x). Similarly, taking upper
s s
limits gives D (f (F ), f (x)) = D (F, x).

5
5.3 Let F be the middle third Cantor set. If x ∈
/ F then, since F is closed, F ∩B(x, r) = ∅
s s
for sufficiently small r, so D (F, x) = D (F, x) = 0.
Let x ∈ F . For k = 1, 2, . . . , the interval B(x, 3−k ) intersects just one kth level
interval in Ek in the usual construction of the Cantor set, see figure 0.1. Thus, with
s = log 2/ log 3, Hs (F ∩ B(x, 3−k )) ≤ 2−k (the Hs -measure of a kth level interval),
so
Hs (F ∩ B(x, 3−k )) 2−k
≤ = 2−s .
(2.3−k )s 2s 2−k
It follows that
Hs (F ∩ B(x, r)) Hs (F ∩ B(x, 3−k ))
s
D (F, x) = limr→0 s
≤ limk→∞ −k s
= 2−s .
(2r) (2.3 )

Since Ds (F, x) < 1 for all x, F is irregular.

5.4 Let F be the dust of figure 0.4, so at the kth stage of construction, E k consists of
4k squares of sides 4−k each with H1 -measure 4−k . For k = 1, 2, . . . , and x ∈ F ,
we have H 1 (F ∩ B(x, 4−k )) ≤ 4−k , since B(x, 4−k ) intersects just one square of E k .
Thus
H1 (F ∩ B(x, r)) H1 (F ∩ B(x, 4−k )) 4−k 1
D1 (F, x) = limr→0 ≤ limk→∞ −k
≤ lim −k
≤ .
2r 2.4 k→∞ 2.4 2
In particular, F is irregular.
√ √
Similarly, H1 (F ∩B(x, 4−k 2)) ≥ 4−k , since B(x, 4−k 2) contains a complete square
of E k . Thus
√
0 H1 (F ∩ B(x, r)) H1 (F ∩ B(x, 4−k 2))
D (F, x) = limr→0 ≥ limk→∞ √
2r 2.4−k 2
4−k 1
≥ lim √ ≥ √ .
k→∞ 2.4−k 2 2 2
In fact for almost all x ∈ F , the point x lies arbitrarily√close to the centre of squares
in E k for large k, so that, given  > 0, B(x, 4−k ( 12 2 + )) contains a complete
square of E k for infinitely many k. Proceeding just as above gives that

1 4−k 1
D (F, x) ≥ lim √ =√ .
k→∞ 4−k ( 2 + 2) 2 + 2
1
Thus for H1 -almost all x ∈ F , D (F, x) ≥ √1 .
2

5.5 Let F be an irregular s-set with 0 < s < 1. Suppose for some 0 < d < 1 there
is a subset F1 ⊂ F with Hs (F1 ) > 0 such that Ds (F, x) > d for all x ∈ F1 . By
s
Proposition 5.1(b) D (F, x) ≤ 1 for almost all x ∈ F . Given  > 0, Egoroff’s therom
guarantees that there exists r0 > 0 and a Borel set E ⊂ F ⊂ F1 with Hs (E) > 0
and such that
d(1 − ) ≤ (2r)−s Hs (F ∩ B(x, r)) ≤ (1 + )
for all x ∈ E and r < r0 .

6
 Let y be a cluster point of E. Let η be a number with 0 < η < 1 and let Ar,η be the
annulus B(y, r(1 + η)) \ B(y, r(1 − η)). Then for r < 12 r0 ,

Hs (F ∩ Ar,η ) Hs (F ∩ B(y, r(1 + η))) Hs (F ∩ B(y, r(1 − η)))

= −
(2r)s (2r)s (2r)s

≤ (1 + )(1 + η)s − (1 − )d(1 − η)s .

For a sequence of r & 0 we may find x ∈ E with |x − y| = r. For all  > 0,
B(x, rη(1 − )) ⊂ Ar,η , so since d(1 − ) ≤ (2rη(1 − ))−s Hs (B(x, rη(1 − ))), we get

d(1 − )1+s η s ≤ (1 + )(1 + η)s − (1 − )d(1 − η)s .

Since this is true for all  > 0 and 0 < η < 1, we conclude that

(1 + η)s
d≤
η s + (1 − η)s

for all 0 < η < 1. We minimise this expression using elementry calculus. Differenti-
ating, and equating to 0 gives

0 = s(η s + (1 − η)s )(1 + η)s−1 − s(1 + η)s (η s−1 − (1 − η)s−1 )

which simplifies to
η s−1 = 2(1 − η)s−1
so η = 21/(s−1) /(1 + 21/(s−1) ) gives the minimum. Thus

(1 + η)s (1 + 2.21/(s−1) )s
d≤ = = (1 + 2s/(s−1) )s−1 .
η s + (1 − η)s 2s/(s−1) + 1

We conclude that D(F, x) ≤ (1 + 2s/(s−1) )s−1 for almost all x ∈ F .

5.6 The simplest example is F = [0, 1] \ Q, which has L1 (F ) = 1 (since Q is countable)

and is totally disconnected, since there is a rational between any two distinct real
numbers.
A more interesting example is a ‘fat fractal’ which may be obtained by a Cantor
set construction, with E0 = [0, 1] and with Ek obtained from Ek−1 by removing
the middle proportion 2−k from each of the 2k−1 intervals of Ek−1 . As usual, F =
∩∞k=0 Ek . Clearly F is totally disconnected. Calculating the lengths of the intervals,
and noting that (1 − x) ≥ exp(−2x) for 0 < x < 21 ,
    
1 k 1 1 1 1
L (Ek ) = 2 × k 1 − 1 − 2 ... 1 − k
2 2 2 2

≥ exp(−1) exp(−1/2) · · · exp(−1/2k−1 ) ≥ exp(−2)

for all k. Thus L1 (F ) ≥ exp(−2) > 0.

7
 s s
5.7 Since Hs (E∩B(x, r)) ≤ Hs (F ∩B(x, r)) for all x and r, we have D (E, x) ≤ D (F, x)
s
for all x. For Hs -almost all x ∈ E \ F , we have D (F, x) = 0 by Proposition 5.1(a),
s
and so D (E, x) = 0. On the other hand, for almost all x ∈ E \ F we have
s s
0 < D (E \ F, x) ≤ D (E, x), by Proposition 5.1(b). We conclude Hs (E \ F ) = 0.
We cannot conclude E ⊂ F : for a counter example, take F to be an s-set where
s > 0 and E = F ∪ {x} for some x ∈
/ F.

5.8 The simplest approach is to use Therom 5.9. If the Fk are all regular, each Fk may
be covered by a countable union of rectifiable curves except for a set of H1 - measure
0. Thus ∪∞ k=1 Fk may be covered by a countable union of rectifiable curves, except
for a coutable union of sets of H1 -measure 0, a set which has H1 -measure 0. (Recall
that a countable union of countable sets is countable.) Since ∪∞ k=1 Fk in a 1-set, it
fufills the criteria of Therom 5.9 to be regular.
Now suppose that the Fk are all irregular. If C is a rectifiable curve, H1 (C ∩Fk ) = 0,
so
H1 (C ∩ ∪∞ 1 ∞ ∞ 1
k=1 Fk ) = H (∪k=1 (C ∩ Fk )) ≤ Σk=1 H (C ∩ Fk ) = 0.

By Therom 5.9, ∪∞
k=1 Fk is irregular.

5.9 Suppose H1 (E ∩ F ) > 0, so E ∩ F is a 1-set. Then E ∩ F is a subset of a regular

1-set E so is regular, and also a subset of an irregular 1-set F so is irregular, by the
remark after (5.4). Thus E ∩ F is both regular and irregular, so H1 (E ∩ F ) = 0.

8
 Chapter 6
1
6.1 (a) If 2
< λ < 1 then
2 log 2
1> = dimH projθ E
log(2/(1 − λ))
1
for almost all θ, by Theorem 6.1. If 0 < λ ≤ 2
then dimH E ≥ 1, so dimH projθ E = 1
for almost all θ, by Theorem 6.1.
(b) We have proj0 E = projπ/2 E = F , so
log 2
dimH proj0 E = dimH projπ/2 E = dimH F = .
log(2/(1 − λ))

6.2 With C as the unit circle in the complex plane, f : [0, 1] → C given by f (φ) = e2πiφ
is Lipschitz, since |f (φ1 ) − f (φ2 )| = |e2πiφ1 − e2πiφ2 | ≤ 2π|φ1 − φ2 |. Thus dimH E =
dimH f (F ) ≤ dimH F = log 2/ log 3, so dimH projθ E ≤ log 2/ log 3 for all θ.
On the other hand, given θ, we may choose a basic interval I of the Cantor set
such that the arc A = {f (φ) : φ ∈ I} has all its tangents making angles at most
ψ < 21 π with the line Lθ in direction θ. Then projθ : A → Lθ is bi-Lipschitz, so
dimH projθ E ≥ dimH projθ (E ∩ A) = log 2/ log 3. Hence dimH projθ E = log 2/ log 3
for all θ.

6.3 Let E be the middle-λ Cantor set, with λ chosen so that s = log 2/ log(2/(1 − λ)) =
dimH E. Then E is an s-set. Let F = {(x, y) : (x = 0 and y ∈ E) or (x ∈ E and y =
0)}. Then for all θ 6= 0, π2 , the projection projθ F is the union of two similar copies
of E, so is an s-set. Also proj0 E and proj π2 E are congruent to E and so are s-sets.

6.4 We have by Therom 6.1 that, for almost all θ,

dimH projθ (E × F ) = min{1, dimH (E × F )} ≥ min{1, dimH E + dimH F }

> max{dimH E, dimH F },

since 0 < dimH E, dimH F < 1. But dimH proj0 (E ×F ) = dimH E and dimH proj π2 (E ×
F ) = dimH F , so the projections onto the coordinate axes have exceptionally small
dimensions in the sense of Therom 6.1.

6.5 Assume without loss of generality that θ = 0 and that F is bounded, say F ⊂
[a, b]×[a, b] (transforming by a congruence if necessary). Then F ⊂ (proj0 F )×[a, b],
so
dimH F ≤ dimH ((proj0 F ) × [a, b]) = dimH (proj0 F ) + 1
by a direct covering argument, or see Corollary 7.4. This formula extends to un-
bounded sets
S F in the usual way, expressing F as a countable union of bounded
sets F = ∞ j=1 (F ∩ ([−j, j] × [−j, j])) and using countable stability of Hausdorff
dimension dimH F = sup∞ j=1 dimH F (F ∩ ([−j, j] × [−j, j])).

9
 6.6 Let x 6= y ∈ F ⊂ R2 where F is an irregular 1-set. By Theorem 6.4 we may
choose a direction θ such that length(projθ F ) = H1 (projθ F ) = 0 and such that
projθ x 6= projθ y. Thus we may choose z ∈ Lθ , the line through the origin in direc-
tion θ, with z between projθ x and projθ y such that z ∈ / projθ F . Thus the line L
through z and perpendicular to Lθ does not intersect F , so if U and V are the two
open half-planes bounded by L, we have F = (F ∩ U ) ∪ (F ∩ V ) with x ∈ F ∩ U
and y ∈ F ∩ V . Thus x and y are in different connected components of F . This is
true for all x 6= y, so F is totally disconnected.

6.7 Let x 6= y ∈ F ⊂ R2 and let θ be a direction other than that of the segment [x, y],
so projθ x 6= projθ y. If L1 (projθ F ) = 0, then we may find z ∈ Lθ , the line through
the origin in direction θ, with z between projθ x and projθ y such that z ∈ / projθ F .
Thus the line L through z and perpendicular to Lθ does not intersect F , so if U and
V are the two open half-planes bounded by L, we have F = (F ∩ U ) ∪ (F ∩ V ) with
x ∈ F ∩ U and y ∈ F ∩ V . Thus x and y are in different connected components
of F , contradicting that F is connected. We conclude that L1 (projθ F ) > 0 for all
directions except the direction of the segment [x, y].
In fact, we may conclude that if F is connected then L1 (projθ F ) > 0, and indeed
that projθ F is an interval of positive length, for all θ unless F is a subset of a straight
line, in which case this is true for all but one direction θ.

6.8 Write Lθ for the line through the origin in direction θ. Then projθ (x, y) is the point
on Lθ at distance x cos θ + y sin θ = (x + yλ) cos θ from the origin, where λ = tan θ.
Thus for all θ such that cos θ 6= 0, the set E + λF is similar to projθ (E × F ), so for
almost all θ, that is for almost all λ,

dimH (E + λF ) = dimH projθ (E × F ) = min{1, dimH E × F }

by Therom 6.1.

6.9 Let F = ∞
S
i=1 Fi be a countable union of the irregular 1-sets Fi . By Theorem 6.4,
L1 (projθ Fi ) = 0 for almost all θ, say for all θ ∈
/ Θi , where L1 (Θi ) = 0. Thus
∞
[ ∞
X
1 1
L (projθ F ) = L (projθ Fi ) ≤ L1 (projθ Fi ) = 0
i=1 i=1

/ ∞
S 1
S∞ P∞ 1 1
for all θ ∈ i=1 Θi , where L ( i=1 Θi ) ≤ i=1 L (Θi ) = 0. Thus L (projθ F ) = 0
for almost all θ.

6.10 Suppose H1 (C ∩ (E × F )) > 0 for some rectifiable curve C. Then C ∩ (E × F )

is a regular 1-set by Proposition 5.6, so L1 (projθ (C ∩ (E × F ))) > 0 for all except
at most one direction θ. Thus either 0 < L1 (proj0 (C ∩ (E × F ))) ≤ L1 (E) or
0 < L1 (projπ/2 (C ∩ (E × F ))) = L1 (F ), a contradiction.

6.11 This result of this exercise may be obtained by transforming Projection theorem 6.1
under a projective transformation.

10
In the plane, let C be the ‘line at infinity’, that is the set of directions of lines in
the plane. We claim that, given a line L in R2 , there exists a natural bijection
ψ : R2 ∪ C → R2 ∪ C such that ψ(L) = C which has nice geometrical properties
regarding straight lines, projections and dimensions.
Regard R2 as the ‘x-y’coordinate plane in R3 , let C be its line at infinity, and let L
be a given line in R2 . Let (a, b, 0) be some point of L and let p be the point (a, b, 1).
Let P be a plane which is perpendicular to R2 and parallel to L but not containing
L, and let C 0 be the line at infinity of P . Define ψ : R2 ∪ C → P ∪ C 0 by taking ψ(x)
to be the point of intersection of P with the line through x and p. If x ∈ L then we
take this to be the point of the ‘line at infinity’ C 0 corresponding to the direction
of the line through x and p. The map ψ extends to C by mapping a direction in
C onto the point of intersection of P with the line through p in that direction. By
identifying P with R2 we get the desired mapping ψ : R2 ∪ C → R2 ∪ C.
It is immediate that (i) ψ maps the set of straight lines (including C) bijectively
onto the set of straight lines, (ii) ψ(L) = C. For our purposes we note the following
particular properties which follow easily form the construction: (iii) ψ is bi-Lipschitz
on every set B such that B and ψ(B) are bounded subsets of R2 , so in particular
F and ψ(F ) have equal Hausdorff dimension (provided they both avoid L and C),
(iv) for E ⊂ L we have L1 (ψ(E)) > 0 (thinking of ψ(E) as a set of directions) if
and only if L1 (E) > 0, (v) for each x ∈ L, a set of lines {Lθ : θ ∈ Θ} through x
has positive angular measure, i.e. L1 (Θ) > 0, if and only if the set of parallel image
lines {ψ(Lθ ) : θ ∈ Θ} in direction ψ(x) have displacements of positive L1 -measure.
We may now deduce the required results by transforming Projection theorem 6.1
under ψ and using the above properties. Let L be a straight line and let ψ be as
above, so that ψ(x) ∈ C if x ∈ L. Let F be a Borel subset of R2 \L with dimH F > 1.
Then dimH (ψ(F )) > 1. By Projection theorem 6.1(b), for almost all directions ψ
the lines in direction ψ(x) that intersect ψ(F ) have displacements of positive L1 -
measure. Transforming back using (iv) and (v) above, this becomes that for almost
all x ∈ L, L1 { directions of lines through x that intersect F } > 0.
We may show in a similar manner that if dimH F = s ≤ 1 then for almost all x ∈ L,
dimH projx F = s; here we replace (v) above by: (vi) for each x ∈ L, a set of lines
{Lθ : θ ∈ Θ} through x has angular dimension s, i.e. dimH (Θ) = s, if and only if
the set of parallel image lines {ψ(Lθ ) : θ ∈ Θ} in direction ψ(x) have displacements
of Hausdorff dimension s.

11
 Chapter 7

7.1 We have that dimH [0, 1] = dimB [0, 1] = 1, so by Corollary 7.4

dimH (F × [0, 1]) = dimH F + dimH [0, 1] = dimH F + 1.

7.2 From Example 4.5 or Exercise 3.14, dimH Fλ = dimB Fλ = log 2/ log(2/(1 − λ)), so
by Formula 7.5 and Corollary 7.4
log 2 log 2
+ = dimB Fλ + dimB Fµ ≥ dimB (Fλ × Fµ )
log(2/(1 − λ)) log(2/(1 − µ))
log 2 log 2
≥ dimH (Fλ × Fµ ) = dimH Fλ + dimH Fµ = + .
log(2/(1 − λ)) log(2/(1 − µ))
Hence
log 2 log 2
dimH (Fλ × Fµ ) = dimB (Fλ × Fµ ) = + .
log(2/(1 − λ)) log(2/(1 − µ))

7.3 This is a slight variant on Example 7.7. First assume that F ⊂ [a, b] for some
0 < a < b < ∞. The mapping f : [a, b] × R → R2 given by f (x, y) = (x cos y, x sin y)
is Lipschitz, with F 0 = f (F × [0, 2π]). Thus

dimH F 0 = dimH f (F × [0, 2π]) ≤ dimH (F × [0, 2π])

= dimH F + dimH [0, 2π] = dimH F + 1

by Proposition 3.3(a) and Example 7.6.

On the other hand, for 0 < a < b < ∞, the restriction f : [a, b] × [0, π] → R2 is a
bi-Lipschitz function, with F 0 ⊃ f (F × [0, π]). Thus

dimH F 0 ≥ dimH f (F × [0, π]) ≥ dimH (F × [0, π])

= dimH F + dimH [0, π] = dimH F + 1

by Proposition 3.3(b) and Example 7.6. We conclude that dimH F 0 = dimH F + 1.

Finally, for F ⊂ [0, ∞], we have, by the above, that dimH F 0 ∩((B(0, n)\B o (0,
S∞1/n)) =
dimH (F ∩ [1/n, n]) + 1 for each integer n. Noting that F \ {0} = n=2 (F ∩
[1/n, n]) and using countable stability of Hausdorff dimension, see Section 3.2, we
get dimH (F 0 \ {0}) = dimH (F \ {0}) + 1, so adding in the origin if necessary gives
dimH F 0 = dimH F + 1.
Note further, that by using the Lipschitz mapping properties of Hausdorff measures
(Proposition 3.1) in a similar way, if F ⊂ [a, b] for some 0 < a < b < ∞ and
0 < Hs (F ) < ∞ then 0 < Hs+1 (F 0 ) < ∞.

7.4 Let E be any Borel subset of S [0, 1] such that dimH E = 1 and length(E) = 0. (For
example, we might take E = ∞ k=1 Ek where Ek are Borel sets with dimH Ek % 1,
see Exercise 4.9.) Setting F = E × E ⊂ R2 , we get, using Product formula 7.2,
that 2 ≥ dimH F = dimH (E × E) ≥ 2dimH E = 2, so dimH F = 2. Moreover, the
projection of F onto each of the coordinate axes is just E, so these projections have
zero length.

12
 If F0 is a 1-set with F0 ⊂ F , then the projection of F0 onto each coordinate axis is a
subset of E, and so has length 0. Thus F0 is a 1-set with projections of zero length
in two directions, so is irregular by Corollory 6.6.
If C is a rectifiable curve, then C ∩F is both irregular and regular so H1 (C ∩F ) = 0,
see the remark after (5.4).

7.5 Let {Ui } and {Vj } be δ-covers

√ of E and F by Nδ (E) and Nδ (F ) cubes respectively.
Then {Ui × Vj }i,j is a δ n-cover of E × F . hence

Nδ√n (E × F ) ≤ Nδ (E)Nδ (F ),

so
log Nδ√n (E × F ) log Nδ (E) log Nδ (F )
√ ≤ √ + √ .
− log δ n − log δ − log n − log δ − log n
Taking upper limits we get
log Nδ√n (E × F )
dimB (E × F ) ≤ limδ→0 √
− log δ n
log Nδ (E) log Nδ (F )
≤ limδ→0 √ + limδ→0 √
− log δ − log n − log δ − log n
log Nδ (E) log Nδ (F )
≤ limδ→0 + limδ→0
− log δ − log δ
= dimB E + dimB F.
√
7.6 Note that√ (x, y) 7→ (x + y)/ 2 is projection onto the line ‘y = x’ and (x, y) 7→
(x − y)/ 2 is projection onto the line ‘y = −x’. Thus, with F the middle third
2
Cantor set, E = {(x, y)√ ∈ R : x + y ∈ F and x − y ∈ F } is just F × F scaled
down by a factor of 1/ 2 and rotated by −π/2 about the origin. In particular, E
is similar to F × F . By Example 7.6 and Formulae 7.3 and 7.5, dimH (F × F ) =
dimB (F × F ) = 2 log 2/ log 3, so as the dimensions are preserved under similarity
transformations, dimH E = dimB E = 2 log 2/ log 3.

√
7.7 Recall that (x, y) 7→ (x − y)/ 2 is projection onto the line L given by ‘y = −x’.
Hence the difference set D = {x − y : x, y ∈ F } is similar to projL (F × F ), so

dimH D = dimH projL (F × F ) ≤ dimH (F × F ) = 2dimH F,

by (6.1) and Corollary 7.4. Since D is a subset of a line, dimH D ≤ min{1, 2dimH F }.

7.8 With F the middle third Cantor set, the set E = {(x, y) : y − x2 ∈ F } is the union
of the parabolae {y = x2 + a : a ∈ F }, that is a stack of homothetic (ie translates
of each other) parabolae, that intersect the y-axis in the points of F .
Locally, E is the product of a line segment and the Cantor set, so we would expect
E to have dimension 1 + log 2/ log 3. More formally, defining φ(x, y) = (x, y + x2 ),
it is easy to see that, for each k, the mapping φ : [−k, k] × F → E ∩ ([−k, k] × R)
is a bi-Lipschitz bijection, so

dimH (E ∩ ([−k, k] × R)) = dimH ([−k, k] × F ) = 1 + log 2/ log 3.

13
 Since E = ∪∞
k=1 E ∩ ([−k, k] × R), we conclude that dimH E = log 2/ log 3 + 1.

Technically, the box dimension of E is not defined since E is unbounded. Any

bounded portion has box dimension log 2/ log 3 + 1.

7.9 Write Lx for the line through (x, 0) parallel to the y-axis, and let Es = {x ∈ R :
dimH (F ∩ Lx ) ≥ s}. By Corollary 7.12, dimH F ≥ s + dimH Es for all 0 ≤ s ≤ 1, so
dimH F ≥ sup0≤s≤1 {s + dimH Es }.

7.10 Writing Ek for the kth stage of the iterative construction of F in the usual way, we
note that Ek consists of 12k rectangles of size 3−k × 5−k . Each of these rectangles
may be covered by at most (5/3)k + 1 ≤ 2(5/3)k squares of side 5−k by dividing
the rectangles using a series of vertical cuts. Thus Ek may √be covered by 12k ×
2 × (5/3)k = 2 × 20k squares of side 5−k i.e. of diameter 5−k 2. In the usual way
(see Proposition 4.1) this gives that dimH F ≤ log 20/ log 5 = (log 5 + log 4)/ log 5 =
1 + log 4/ log 5.
For the lower bound, let Lx be the line through (x, 0) parallel to the y-axis. Then,
except for x of the form j3−k where j and k are integers, we have that Ek ∩ Lx con-
sists of 4k intervals of length 5−k . A standard application of the mass distribution
principle (considering a mass such that each of these intervals has mass 4−k ) gives
that dimH (F ∩ Lx ) ≥ log 4/ log 5. By Corollary 7.12 dimH F ≥ 1 + log 4/ log 5, so
dimH F = 1 + log 4/ log 5.

7.11 Writing Ek for the kth stage of the iterative construction of F in the usual way, we
note that Ek consists of 8k rectangles of size 3−k ×5−k . For a given positive integer k,
let q be the integer such that 5−k−1 < 3−q ≤ 5−k . Then dividing the rectangles of Ek
horizontally into nearly square rectangles of size 3−q × 5−k and selecting those above
the set Ck , the kth stage of the usual middle third Cantor set construction on the
x-axis, we get that F may be covered by 2q 4k = 3q log 2/ log 3 4k ≤ 5(k+1)
√
log 2/ log 3 k
4 rect-
−q −k −k
angles of size 3 × 5 , each contained in a square of diameter 5 2. In the usual
way (see theorem 4.1) this gives that dimH F ≤ ((log 2/ log 3) log 5 + log 4)/ log 5 =
log 2/ log 3 + log 4/ log 5.

The lower bound is similar to Exercise 7.10. Let Lx be the line through (x, 0) paral-
lel to the y-axis. For all x ∈ C where C is the middle third Cantor set, except those
x of the form j3−k where j and k are integers, we have that Ek ∩ Lx consists of 4k
intervals of length 5−k . The mass distribution principle (considering a mass such
that each of these intervals has mass 4−k ) gives that dimH (F ∩ Lx ) ≥ log 4/ log 5.
By Corollary 7.12 dimH F ≥ dimH C + log 4/ log 5 = log 2/ log 3 + log 4/ log 5, so
dimH F = log 2/ log 3 + log 4/ log 5.

14
 Chapter 8
8.1 Let E and F be line segments of lengths L(E) and L(F ) making an angle θ 6= 0
with each other. Then E and F + x intersect if and only if x lies in a parallelogram
that is a translate of that formed by the vectors along E and F . Thus
Z
#(E ∩ (F + x))dx = area of parallelogram = L(E)L(F )| sin θ|.

Now letting the direction of F vary before translating,

Z Z 2π
#(E ∩ σ(F ))dσ = L(E)L(F ) | sin θ|dθ = 4L(E)L(F ). (∗)
0

Now let E = m
S Sn
i=1 Ei and F = j=1 Fj be polygonal curves with Ei and Fj line
segments of lengths L(Ei ) and L(Fj ). Then by (∗)
Z Xm Xn Z
#(E ∩ σ(F ))dσ = #(Ei ∩ σ(Fj ))dσ
i=1 j=1

m X
X n
= 4L(Ei )L(Fj ) = 4L(E)L(F ), (∗∗)
i=1 j=1

where L(E) and L(F ) are the total lengths of E and F .

We proceed from polygonal curves to rectifiable curves by approximation. Let E and
F be rectifiable curves, and let En and Fn be sequences of polygonal curves giving
closer and closer approximations to E and F , such that each En is a refinement of
En−1 (ie the pologonal curve En is obtained from En−1 by adding further vertices)
and such that each Fn is a refinement of Fn−1 , and such that limn→∞ L(En ) = L(E)
and limn→∞ L(Fn ) = L(F ). From (∗∗)
Z
#(En ∩ σ(Fn ))dσ = 4L(En )L(Fn ).
R R
We now take the limit as n → ∞. Provided that #(En ∩ σ(Fn ))dσ → #(E ∩
σ(F ))dσ we get that
Z
#(E ∩ σ(F ))dσ = 4L(E)L(F ). (∗)

for rectifiable E and F . For most specific curves E and F , this will follow from the
bounded convergence theorem or the monotone convergence theorem. Justification
of this step is more involved for general rectifiable curves, see Santaló (2004).

8.2 Let let L be a unit segment in the plane oriented perpendicular to direction θ.
Assume that C contains no line segment parallel to L (this can only happen for a
set of directions of L1 -measure 0), and let C + and C − be the ‘upper’ and ‘lower’
parts of the curve C with respect to the direction of L. Then (L + x) ∩ C + is a
single point if x lies in a region congruent to that swept out by translating C + unit
distance in the direction of L, otherwise (L + x) ∩ C + = ∅. Thus
Z Z
#(C ∩ (L + x))dx = #(C − ∩ (L + x))dx
+

15
 = L1 (projθ C) × 1
so Z
#(C ∩ (L + x)) = 2L1 (projθ C).

Integrating with respect to θ for 0 ≤ θ < 2π and using Exercise 8.1 gives
Z Z 2π
1
4L (C) = #(C ∩ σ(F ))dσ = 2L1 (projθ C)
0

giving the required formula.

8.3 We recall that, with E the product of two middle third Cantor sets, dimH E =
2 log 2/ log 3 = log 4/ log 3.
(i) With F a circle dimH F = dimB F = 1. By Theorem 8.1 and Corollary 7.4

dimH (E ∩ (F + x)) ≤ max{0, dimH (E × F ) − 2} = max{0, dimH E + dimH F − 2}

= log 4/ log 3 + 1 − 2 = log 4/ log 3 − 1

for almost all x ∈ R2 , and thus

dimH (E ∩ σ(F )) ≤ log 4/ log 3 − 1

for almost all congruence transformations σ.

By Theorem 8.3(b)

dimH (E ∩ σ(F )) ≥ dimH E + dimH F − 2 = log 4/ log 3 + 1 − 2 = log 4/ log 3 − 1

for a set of congruence transformations of positive measure.

(ii) With F the von Koch curve dimH F = dimB F = log 4/ log 3. By Theorem 8.1
and Corollary 7.4

dimH (E ∩ (F + x)) ≤ max{0, dimH (E × F ) − 2}

= max{0, dimH E + dimH F − 2} = 2 log 4/ log 3 − 2

for almost all x ∈ R2 , and thus

dimH (E ∩ σ(F )) ≤ 2 log 4/ log 3 − 2

for almost all congruence transformations σ.

By Theorem 8.3(a)

dimH (E ∩ σ(F )) ≥ dimH E + dimH F − 2

= log 4/ log 3 + log 4/ log 3 − 2 = 2 log 4/ log 3 − 2

for a set of similarity transformations of positive measure. Note that we cannot
apply Theorem 8.3(c) since dimH E, dimH F ≤ 3/2.
(iii) With F the product of two middle third Cantor sets, dimH F = log 4/ log 3. By
Theorem 8.1 and Corollary 7.4

dimH (E ∩ (F + x)) ≤ max{0, dimH (E × F ) − 2}

16
 = max{0, dimH E + dimH F − 2} = 2 log 4/ log 3 − 2
for almost all x ∈ R2 , and thus

dimH (E ∩ σ(F )) ≤ 2 log 4/ log 3 − 2

for almost all congruence transformations σ.

By Theorem 8.3(a)

dimH (E ∩ σ(F )) ≥ dimH E + dimH F − 2 = 2 log 4/ log 3 − 2

for a set of similarity transformations of positive measure. Again, we cannot apply

Theorem 8.3(c) since dimH E, dimH F ≤ 3/2.

8.4 As in Theorem 8.1, we prove this when n = 1. Let Lc be the line x = y + c. If

dimH (E × F ) < 1, the projection of E × F onto the line x + y = 0 has zero length
by (6.1), in other words, (E × F ) ∩ Lc = ∅ for almost all c ∈ R. But (E × F ) ∩ Lc is
similar to E ∩ (F + c) (since (x, x − c) ∈ (E × F ) ∩ Lc if and only if x ∈ E ∩ (F + c)).
Thus E ∩ (F + c) = ∅ for almost all c ∈ R.

8.5 Let E be the set of points (x, y) in the unit square A such that both coordinates x
and y are rational. Then dimB E = dimB E = dimB A = 2, by Proposition 2.6. Let
F be a line segment. Since E is a countable set (or since its projection in every
direction has length 0), E ∩ σ(F ) = ∅ so that dimB (E ∩ σ(F )) = 0 for almost all
similarities σ. However, dimB E + dimB F − 2 = 2 + 1 − 2 = 1, so (8.7) fails for
almost all similarities σ.

8.6 Let C be a middle-λ Cantor set with λ chosen so that dimH C = s − 1 (see after
example 4.5) and transformed by a similarity so that the end points are 21 and 1. Let
F be the ‘target’ given in polar coordinates by F = {(r, θ) : r ∈ C, 0 ≤ θ < 2π}. By
Exercise 7.3, dimH F = 1 + dimH C = s, and by the note at the end of the solution
of Exercise 7.3, 0 < Hs (F ) < ∞, so F is an s-set.
It is easy to see that any line segment E that intersects the interior of the unit disc
cuts a set of the rings of the target F corresponding to a similar subset of C, that
is E ∩ F contains a subset that is bi-Lipschitz equivalent to C and so has positive
s-dimensional Hausdorff measure. On the other-hand, provided that E is not tan-
gential to one of the rings of F , E ∩ F is contained in a finite union of Lipschitz
images of C, and so E ∩ F has finite s-dimensional Hausdorff measure and so is an
s-set.

8.7 Writing SLk for the line segment {(x, k −1/2 ) : 0 ≤ x ≤ k −1/2 }, we see that E =
{(0, 0)} ∞k=1 Lk . Given small enough δ, let k be the integer such that

1
2
(k + 1)−3/2 ≤ k −1/2 − (k + 1)−1/2 ≤ δ < (k − 1)−1/2 − k −1/2 (∗)

where the left hand inequality follows using the mean value theorem. Then any set
of diameter δ or less can intersect at most one of the segments L1 , . . . , Lk , and the

17
 segment Lj requires at least j −1/2 /δ sets of diameter δ in any covering. Thus, using
an ‘integral test’ estimate for the sum,
k k
j −1/2 2k 1/2
Z
X 1
Nδ (E) ≥ ≥ x−1/2 dx = ≤ 2δ −1 cδ −1/3 ≤ 2cδ −4/3 ,
j=1
δ δ 0 δ

using (∗), where c is independent of δ. It follows immediately that

log Nδ (E) log 2cδ −4/3 4

dimB E ≥ dimB E = limδ→0 ≥ = .
− log δ − log δ 3

Every line L, that does not pass through (0, 0) and that does not contain one
of the line segments Lk , intersects E in a finite set of points, so in particular
dimB (L ∩ E) = 0.

8.8 Let 0 < s < 1, let I be an interval and let  > 0 be given. We use a mass
s
distribuition method to estimate H∞ (I ∩ Ek ) for large k. We may find η > 0 such
1−s 2
that η ≤ ( 3 + ). For given k, let µ be the mass distribution on Ek given by the
restriction of Lebesgue measure to Ek . Note that if U is a ‘not too small’ subinterval
of I and k is large enough then µ(U ) is close to 23 |U |, since two-thirds of the ternary
intervals of length 3−k are present in Ek . In particular, provided that k is large
enough, k ≥ k0 , say, we can ensure that ( 23 − )|I| ≤ µ(I) and
      s
2 2 |U | 2 |U |
µ(U ) ≤ +  |U | = + |I| ≤ + |I| if |U | ≥ η|I|,
3 3 |I| 3 |I|

and
s
|U |s 1−s
 
s 1−s 2 |U |
µ(U ) ≤ |U | = |U | |U | ≤ η |I| ≤ + |I| if |U | ≤ η|I|.
|I|s 3 |I|

Thus if {Ui } is any cover of I ∩ Ek , then

|Ui |s
    P
2 X 2
−  |I| ≤ µ(I) ≤ µ(Ui ) ≤ +  |I| i s
3 i
3 |I|

giving that
2
X
s s (3 − )
|Ui | ≥ |I| 2 ≥ |I|s (1 − 2).
i
(3 + )
s
Thus H∞ (I ∩ Ek ) ≥ |I|s (1 − 2).
s
It follows that limk→∞ H∞ (I ∩ Ek ) = |I|s , so (8.10) is satisfied and F = limk→∞ Ek
is in class C s (−∞, ∞) for all 0 < s < 1.
It is immediate from Proposition 8.6 that dimH F ≥ s for all 0 < s < 1, so dimH F =
1. Moreover, for all 0 < s < 1 and any x1 , x2 , . . ., we have that F +xi ∈ C s (−∞, ∞),
by Proposition 8.9 or by repeating the argument above. Thus, by Proposition 8.7
and Corollary 8.8, ∩∞ s ∞
i=1 (F + xi ) ∈ C (−∞, ∞) and dimH ∩i=1 (F + xi ) ≥ s, so we
∞
conclude that dimH ∩i=1 (F + xi ) = 1.

18
Chapter 9

9.1 The Hausdorff metric is given by

d(A, B) = inf{δ : A ⊂ Bδ and B ⊂ Aδ },

where Aδ , Bδ are the δ-neighbourhoods of A and B.

To show that d satisfies condition (i) for a metric, we first note that d(A, B) ≥ 0
with d(A, A) = 0. Now suppose that d(A, B) = 0. Taking x ∈ B, we note that, for
each positive integer n, we must have x ∈ B ⊂ A1/n and so there exists xn ∈ A such
that |x − xn | ≤ 1/n. Thus x ∈ A = A since A is compact and hence closed. Thus
B ⊂ A; similarly A ⊂ B so A = B as required.
Clearly by the symmetry of the definition, d(A, B) = d(B, A), which is condition
(ii) for a metric.
To show that d satisfies condition (iii), we suppose that d(A, C) = 1 and d(C, B) =
2 . Then, for each δ1 > 1 and each δ2 > 2 , we have

A ⊂ Cδ1 , C ⊂ Aδ1 , C ⊂ Bδ2 , B ⊂ Cδ2 .

So
A ⊂ Cδ1 ⊂ Bδ1 +δ2 and B ⊂ Cδ2 ⊂ Aδ2 +δ1 .
Thus d(A, B) ≤ δ1 + δ2 for all δ1 , δ2 with δ1 > 1 and δ2 > 2 }, so d(A, B) ≤ 1 + 2 =
d(A, C) + d(C, B) as claimed.

9.2 Let c be any real number for which 0 < c < 1. Then the interval [0, 1] is the at-
tractor for the similarity transformations defined on R by S1 (x) = cx and S2 (x) =
(1 − c)x + c, since S1 ([0, 1]) ∪ S2 ([0, 1]) = [0, c] ∪ [c, 1] = [0, 1]. Clearly S1 and S2
are both contractions, so from Theorem 9.1 [0, 1] is the unique non-empty compact
attractor for S1 and S2 .

9.3 We begin by noting that the middle third Cantor set is non-empty and compact. The
middle third Cantor set is therefore the attractor for the following four similarity
transformations on R which map the interval E0 = [0, 1] onto the four intervals in
E2 :

S1 (x) = x/9, S2 (x) = x/9 + 2/9, S3 (x) = x/9 + 2/3, S4 (x) = x/9 + 8/9.

The ratios of these similarities are all 1/9 so that equation (9.14) is 4(1/9)s = 1.
Taking the log of both sides gives log 4 − s log 9 = 0 so that

log 4 log 22 2 log 2 log 2

s= = 2
= =
log 9 log 3 2 log 3 log 3

The middle third Cantor set is also the attractor for the following three similarity
transformations on R which map the interval E0 = [0, 1] onto the first two intervals
in E2 and the second interval in E1 :

S1 (x) = x/9, S2 (x) = x/9 + 2/9, S3 (x) = x/3 + 2/3.

1
 The ratios of the similarities are 1/9, 1/9 and 1/3 respectively, so that equation
(9.13) is 2(1/9)s + (1/3)s = 1. Putting x = (1/3)s , we can write this as 2x2 + x = 1
or, equivalently, (2x − 1)(x + 1) = 0 which has solutions x = 1/2 and x = −1. Since
x = (1/3)s > 0, it follows that (1/3)s = 1/2. Taking logarithms of both sides gives
−s log 3 = − log 2 so that s = log 2/ log 3.

9.4 Recall
 that the matrix which represents a rotation about the origin through angle
cos θ − sin θ
θ is . Note that the generator of the von Koch curve (see figure
sin θ cos θ
√
0.2) has vertices (0, 0),( 31 , 0),( 12 , 63 ), ( 23 , 0) and (1, 0). Regarding the similarities
that maps the line segment joining (0, 0) and (1, 0) onto the intermediate segments
as a composition of a rotation by ±π/3 (if necessary), a scaling by factor of 31 and
a translation, we see that an IFS {S1 , S2 , S3 , S4 } that has the von Koch curve as
attractor is given by
√ !
1 3
        1 
x 1 x x 1 2
− 2
x
S1 = S2 = √
3 1
+ 3
y 3 y y 3 y 0
2 2

√ !
  1 3
  1       2

x 1 2√ 2
x √2
x 1 x 3
S3 = 3 1
+ 3 S4 = + .
y 3 − y 6
y 3 y 0
2 2

The open set condition holds, taking √ the open set V to be the interior of the isosce-
les triangle with vertices (0, 0),( 2 , 63 ) and (1, 0). This open triangle is maped by
1

S1 , . . . , S4 to four similar open triangles at scale 13 with bases on the four segments of
the generator, with their union containd in V . Theorem 9.3 immediately P4 givess that
the box and Hausdorff dimension s of the von Koch curve is given by i=1 (1/3) = 1,
that is 4 × 3−s = 1 or s = log 4/ log 3.

9.5 The set F in figure 0.5 is the attractor of the following five similarities on R2 which
map E0 onto the five squares in E1 :

S1 (x, y) = (x/4, y/4), S2 (x, y) = (x/4 + 3/4, y/4),

S3 (x, y) = (x/4 + 3/4, y/4 + 3/4), S4 (x, y) = (x/4, y/4 + 3/4)

S5 (x, y) = (x/2 + 1/4, y/2 + 1/4).
The ratios of these similarities are 1/4, 1/4, 1/4, 1/4, 1/2 respectively. Each of the
sets in Ek is compact. Thus F is the intersection of a decreasing sequence of compact
sets and is hence compact. So F is the attractor satisfying F = 5i=1 Si (F ). The
S
open set condition holds, taking V as the interior of the initial square E0 , and
so it follows from Theorem 9.3 that dimH F = dimB F = s, where s is given by
1 = 4(1/4)s + (1/2)s .
Note that this can be solved by putting x = (1/2)s to give

4x2 + x − 1 = 0

2
 so that √
x = (−1 ± 17)/8.
s s
√
Since x = (1/2)
√ > 0, it follows that (1/2) = −1/8 + 17/8 and so −s log 2 =
log(−1/8 + 17/8). Thus
√
− log(−1/8 + 17/8)
s= = 1.357 . . . .
log 2

9.6 The set F is the attractor for the three similarities on R2 :

S1 (x, y) = (x/2, y/2), S2 (x, y) = (x/2 + 1/2, y/2), S3 (x, y) = (y/4 + 1/2, −x/4).

These have ratios 1/2, 1/2, 1/4 respectively. The open set condition holds, taking
V to be the interior of the triangle formed by the three free ends of the segments.
From Theorem 9.3 dimH F = dimB F = s, where

1 = 2(1/2)s + (1/4)s = 2(1/2)s + (1/2)2s .

Putting x = (1/2)s , we have
x2 + 2x − 1 = 0
and so √
x = −1 ± 2.
s s
√
Since
√ x = (1/2) > 0, it follows that (1/2) = −1 + 2 and so −s log 2 = log(−1 +
2). Thus √
− log(−1 + 2)
s= = 1.271 . . . .
log 2

9.7 The set F is the attractor for the following similarities on R which map [0, 1] onto
the intervals [0, 1/4] and [1/2, 1] respectively:

S1 (x) = x/4 and S2 (x) = x/2 + 1/2,

with ratios 1/4 and 1/2. The set F is the intersection of a decreasing sequence of non-
empty compact sets and is hence non-empty and compact. The open set condition
holds, taking V as the open interval (0, 1) and so, by Theorem 9.3, dimH F =
dimB F = s, where s is given by

1 = (1/4)s + (1/2)s = (1/2)2s + (1/2)s .

√
Putting x = (1/2)s , we have x2 + x − 1 = 0√and so x = (−1 ± 5)/2. Since
x = (1/2)s > 0, it follows that (1/2)s = (−1 + 5)/2 = 0.6942 . . . .
and so √ √
− log( −1+2 5
) log(−1 + 5)
s= =1− .
log 2 log 2

3
 9.8 In each case it may be verified trivially that the stated attractor is compact and
satisfies F = S1 (F ) ∪ S2 (F ).
(i) The middle half Cantor set (i.e. the Cantor-type construction with the (open)
middle half of intervals removed at each stage).
(ii) The interval [0, 1].
(iii) The interval [0, 1]. (Notice that in this case the two parts S1 ([0, 1]) and S2 ([0, 1])
overlap non-trivially).

9.9 The open set condition holds for the IFS {S1 , . . . , Sm }, taking V as the open unit
square,
Sm so that the Si (V ) are the interiors of the squares selected in E1 , with V ⊃
i=1 Si (V ) and the union disjoint.
Thus by Theorem 9.3, the box and Hausdorff dimension s of F is given by m s
P
i=1 (1/p) =
−s
1, that is m × p = 1 or s = log m/ log p.

9.10 This is similar to Example 9.8. Here S1 and S2 are contractions on the closed set
D = [0, 1] ⊂ R. We note that
2 −2
S10 (x) = 2
> 0 and S20 (x) = < 0.
(2 + x) (2 + x)2
Thus S1 is increasing on D and S2 is decreasing on D so that
S1 (D) = [S1 (0), S1 (1)] = [0, 1/3] and S2 (D) = [S2 (1), S2 (0)] = [2/3, 1].

Since F ⊂ D, it follows that S1 (F ) ⊂ [0, 1/3] and S2 (F ) ⊂ [2/3, 1]. Thus F is the
disjoint union of S1 (F ) and S2 (F ) and so we can apply Propositions 9.6 and 9.7 to
estimate dimH F . (We could use Proposition 9.6 even if the union was not disjoint.)
For x ∈ D, i = 1, 2, we have
2 2 2 2 1
= 2 ≤ |Si0 (x)| = ≤ = .
9 3 (2 + x)2 22 2
It follows from the mean-value theorem that, for x, y ∈ D, i = 1, 2,
2 1
|x − y| ≤ |Si (x) − Si (y)| ≤ |x − y|.
9 2
By Propositions 9.6 and 9.7, t ≤ dimH F ≤ s, where
2(2/9)t = 1 = 2(1/2)s .
Clearly s = 1 and taking logs gives
log 1/2
t= = 0.46
log 2/9
to two decimal places.
These estimates are rather poor, and so we use the fact that F is also the attractor
of the four contractions defined by
x 2
2+x x 2+x 1
S1 ◦ S1 (x) = x = S1 ◦ S2 (x) = 2 =
2 + 2+x 4 + 3x 2+ 2+x
3+x

4
 2 4 + 2x 2 2+x
S2 ◦ S1 (x) = x = S2 ◦ S2 (x) = 2 = .
2 + 2+x 4 + 3x 2 + 2+x 3+x
Thus for x ∈ D  
4 0
 4  1
≤ |(S1 ◦ S1 ) (x)| =  2
≤
49 (4 + 3x)  4
 
1 0
 −1  1
≤ |(S1 ◦ S2 ) (x)| = 
 ≤
16 (3 + x)2  9
 
4 0
 −4  1
≤ |(S2 ◦ S1 ) (x)| = 
 ≤
49 (4 + 3x)2  4
 
1 0
 1  1
≤ |(S2 ◦ S2 ) (x)| = 
 ≤ .
16 (3 + x)2  9
By the mean-value theorem, for each x, y ∈ D, i = 1, 2,
4 1
|x − y| ≤ |Si ◦ S1 (x) − Si ◦ S1 (y)| ≤ |x − y|
49 4
and
1 1
|x − y| ≤ |Si ◦ S2 (x) − Si ◦ S2 (y)| ≤ |x − y|.
16 9
Since S1 (F ) and S2 (F ) are disjoint, the sets S1 ◦ S1 (F ), S1 ◦ S2 (F ), S2 ◦ S1 (F ) and
S2 ◦ S2 (F ) are also disjoint and so it follows from Propositions 9.6 and 9.7 that
t ≤ dimH F ≤ s, where

2(4/49)t + 2(1/16)t = 1 = 2(1/4)s + 2(1/9)s .

To two decimal places, this is satisfied by s = 0.80 and t = 0.53 and so 0.52 <
dimH F < 0.81.

T∞ the notation of Theorem 9.3. If x ∈ F then, as in (9.8), we have that

9.11 We use
x = k=1 Si1 ◦ · · · ◦ Sik (E). Given 0 < r < |F |, with Q1 as in the proof of Theorem
9.3, we have [
F ∩ B(x, r) ⊂ V i1 ,...,ik ,
i1 ,...,ik ∈Q1
so X X
Hs (F ∩ B(x, r)) ≤ Hs (F ∩ V i1 ,...,ik ) = Hs (Fi1 ,...,ik )
Q1 Q1
X
≤ (ci1 · · · cik )s |F |s ≤ qrs |F |s
Q1
so
s Hs (F ∩ B(x, r)) qrs |F |s
D (F, x) = limr→0 ≤ lim r→0 ≤ q2−s |F |s .
(2r)s (2r)s

On the other hand, if x = ∞

T
k=1 Si1 ◦ · · · ◦ Sik (E), then choosing k such that
(mini ci )r ≤ ci1 · · · cik |F | ≤ r, we have Fi1 ,...,ik ⊂ B(x, r), so that

Hs (F ∩ B(x, r)) ≥ Hs (Fi1 ,...,ik ) ≥ (ci1 · · · cik )s Hs (F ) ≥ (min ci )s |F |−s rs ≡ brs

i

5
 Thus
Hs (F ∩ B(x, r)) brs
Ds (F, x) = limr→0 ≥ limr→0 ≥ b2−s .
(2r)s (2r)s

9.12 Since F ∩ V ⊂ F we have dimH (F ∩ V ) ≤ dimH F .

Since V is an open subset intersecting F , there is a bi-Lipschitz mapping S : F → V ,
so that F ∩ V contains a bi-Lipschitz image T∞of F and thus dimH (F ∩ V ) ≥ dimH F .
(To see this, let x ∈ F ∩ V so that x = k=1 Si1 ◦ · · · ◦ Sik (E), as in (9.8). Then
Si1 ◦ · · · ◦ Sik (F ) ⊂ F ∩ V if k is large enough, so we may take S = Si1 ◦ · · · ◦ Sik as
the bi-Lipschitz mapping.)
Thus dimH (F ∩ V ) ≤ dimH F , so dimH (F ∩ V ) = dimH F . An identical argument
shows that dimB (F ∩V ) = dimB F and dimB (F ∩V ) = dimB F since these dimensions
are also preserved under bi-Lipschitz mappings.
By Corollary 3.10, dimP F = dimB F .

9.13 Note that this is a generalisation of Example 7.13 and Exercises 7.10 and 7.11.
(a) The formula of Example 9.11 gives
p
!
X log p/ log q 1 log(pN log p/ log q ) log N
dimH F = log Nj = =1+ .
j=1
log p log p log q

To check this, write Ek for the kth stage of the iterative construction of F in the
usual way, and note that Ek consists of (pN )k rectangles of size p−k × q −k . Each of
these rectangles may be covered by at most (q/p)k + 1 ≤ 2(q/p)k squares of side q −k
by dividing the rectangles using a series of vertical cuts. Thus E√k may be covered by
(pN )k 2(q/p)k = 2(N q)k squares of side q −k i.e. of diameter q −k 2. In the usual way
(see Theorem 4.1) this gives that dimH F ≤ log(N q)/ log q = (log N + log q)/ log q =
1 + log N/ log q.
For the lower bound, let Lx be the line through (x, 0) parallel to the y-axis. Then,
except for x of the form jp−k where j and k are integers, we have that Ek ∩Lx consists
of N k intervals of length q −k . A standard application of the mass distribution
principle (considering a mass such that each of these intervals has mass N −k ) gives
that dimH (F ∩ Lx ) ≥ log N/ log q. By Corollary 7.12 dimH F ≥ 1 + log N/ log q, so
dimH F = 1 + log N/ log q.
(b) The formula of Example 9.11 gives
p
!
X log p/ log q 1 log(p1 N log p/ log q ) log p1 log N
dimH F = log Nj = = + .
j=1
log p log p log p log q

To check this, write Ek for the kth stage of the iterative construction of F in the
usual way. Note that Ek consists of (p1 N )k rectangles of size p−k × q −k . Let C
be the projection of F onto the x-axis, so that C is a self-similar set subset of
the x-axis formed by p1 similarities of ratios 1/p. Let Ck be the projection of
Ek onto the x-axis, so that Ck is the kth stage of the construction of C under

6
 the usual process. For a given positive integer k, let s be the integer such that
q −k−1 < p−s ≤ q −k . Considering the part of F above the set Cs , we get that F may
s log p / log p k
be covered by ps1 N k = p1 1 N ≤ q (k+1)
√ N rectangles of size p−s ×q −k ,
log p1 / log p k

each contained in a square of diameter q −k 2. In the usual way (see Theorem 4.1)
this gives that dimH F ≤ log(q log p1 / log p N )/ log q = log N/ log q + log p1 / log p.
The lower bound is similar to part (a). Let Lx be the line through (x, 0) parallel to
the y-axis. Let C be the projection of F onto the x-axis, as above, so that C is a
self-similar set subset of the x-axis formed by p1 similarities of ratios 1/p.
For all x ∈ C , except those x of the form jp−k where j and k are integers, we have
that Ek ∩ Lx consists of N k intervals of length q −k . The mass distribution princi-
ple (considering a mass such that each of these intervals has mass N −k ) gives that
dimH (F ∩ Lx ) ≥ log N/ log q. By Corollary 7.12 dimH F ≥ dimH C + log N/ log q =
log p1 / log p + log N/ log q, so dimH F = log p1 / log p + log N/ log q.

9.14 We apply the formulae in Example 9.11 with:

p = 3, q = 6, N1 = 4, N2 = 1, N3 = 3, p1 = 3.

Thus, writing α = log p/ log q = log 3/ log 6,

p
!
X log p/ log q 1 log(4α + 1α + 3α )
dimH F = log Nj = = 1.518 . . .
j=1
log p log 3

and
p
!
log p1 1 X 1 log( 31 (4 + 1 + 3))
dimB F = + log Nj =1+ = 1.627 . . . .
log p p1 j=1 log q log 5

9.15 The programming is left to the reader.

9.16 Let M (t) = #{lj : lj ≥ t}. For J ∈ N and s > 0

Z ∞ J−1 Z
X lj J−1 Z
X lj
s−1 s−1
M (t)t dt = M (t)t dt = jts−1 dt
J j=1 lj+1 j=1 lj+1

J−1 J−1
X 1 1X s 1
= [ljs − lj+1
s
]= l − (J − 1)lJs . (1)
j=1
s s j=1 j s

Letting J → ∞ gives
Z ∞ ∞
X
s−1
s M (t)t dt = ljs = ζ(s). (2)
0 j=1

Clearly neither this integral or sum converge for real s ≤ 0, so this identity is valid
for all real s > σ where σ is the abscissa of convergence of ζ.

7
Since (1) remains valid for all non-zero s ∈ C and the series is absolutely convergent
for s > σ, the identity (2) holds for Re(s) > σ, noting that |ljs | = |lj |Re(s) .
R∞
Since s 0 M (t)ts−1 dt and ζ(s) are analytic and equal on Re(s) > σ, they have the
same analytic continuation to C.

8
 Chapter 10
1
10.1 F is the (non-empty compact) attractor of the IFS {S1 , . . . , S5 } where Si = 10 x+
i−1
5
(i = 1, . . . , 5).
Since 5i=1 Si (0, 1) = 5i=1 Si ( i−1 , i−1 1
S S
5 5
+ 10 ) ⊂ (0, 1), the open set condition holds
with open set (0, 1), so by Theorem 9.3 dimH F = s, where 5 × (1/10)s = 1, that is
dimH F = log 5/ log 10.

10.2 Let {S1 , . . . , Sm } be given by Si = m1 x + i−1

m
(i = 1, . . . , m). Then, since member-
ship of F (p0 , . . . , pm−1 ) is determined by the base-m digits of a number after any
given place,
m
[ m
[
Si (F (p0 , . . . , pm−1 )) = [0, 1) ∩ F (p0 , . . . , pm−1 ) = F (p0 , . . . , pm−1 ),
i=1 i=1

that is F (p0 , . . . , pm−1 ) is a (non-compact) attractor of the Si .

10.3 With the notation of Section 10.1, the numbers in F (1 − 3p, p, 2p) have twice as
many 2s as 1s for all 0 < p < 31 . Thus we must find the maximum value of
dimH F (1 − 3p, p, 2p) over such p. Proposition 10.1 gives
1
φ(p) ≡ dimH F (p, 2p, 1 − 3p) = − [(1 − 3p) log(1 − 3p) + p log p + 2p log 2p]
log 3

1
=− [p(log p + 2 log 2p − 3 log(1 − 3p)) + log(1 − 3p)].
log 3
Then
dφ 1 1 4p3
=− [log p + 2 log 2p − 3 log(1 − 3p)] = − log .
dp log 3 log 3 (1 − 3p)3

Thus a maximum occurs when 4p3 = (1 − 3p)3 or p = 1/(3 + 41/3 ), that is when(1 −
3p) = 41/3 /(3 + 41/3 ). The value of the maximum is
 
1/3 1 1/3 2
φ(1/(3 + 4 )) = log(3 + 4 ) − log 2 = 0.9660 . . . .
log 3 3

Thus the required Hausdorff dimension is 0.9660 . . ..

10.4(i) To find the continued fraction expansion of 41/9, we first note that
41 5 1
=4+ =4+
9 9 9/5

so a0 = 4 and x1 = 9/5. Now

9 4 1
=1+ =1+
5 5 5/4

9
 so a1 = 1 and x2 = 5/4. Now
5 1
=1+
4 4
so a2 = 1 and a3 = 4. So
41 1 1 1
=4+ .
9 1+ 1+ 4
√ √
(ii) To find the continued fraction expansion of 5, we first note that 2 < 5 < 3 and
so
√ 1
5=2+ ,
x1
where x1 > 1 (that is, a0 = 2). Now
√
1 1 5+2 √
x1 = √ =√ √ = 5+2
5−2 5−2 5+2
and so
1
x1 = 4 + ,
x2
where x2 > 1 (that is, a1 = 4). Now
1 1
x2 = =√ = x1
x1 − 4 5−2
and so
1
,
x2 = 4 +
x3
where x3 = x2 > 1 (that is, a2 = 4). This process now repeats itself giving 4 = a3 =
a4 = · · · . Thus
√ 1 1 1
5=1+ .
4+ 4+ 4 + · · ·

1 1 1 1 1
10.5 Letting x = 1 + 1+ 1+ 1+··· √
we see that x = 1 + 1/(1 + 1+ 1+···
) = 1 + 1/x. Thus
2 1+ 5
x − x − 1 = 0, so x = 2 , the golden mean. (We take the positive root of the
quadratic equation since x is clearly positive.)

√ 1 1 1
10.6 We have 2 = 1 + 2+ so that curtailing after each term gives successive
√ 2+ 2+···
approximations to 2 = 1.41421 . . . of:
3 7 17 41 99
= 1.5, = 1.4, = 1.41666 . . . , = 1.41379 . . . , = 1.41428 . . .
2 5 12 29 70

10.7 This similar to Example 10.2, using Example 9.8. Let F denote the set of positive
numbers with infinite continued fraction expansions which have all partial quotients
equal to 2 or 3. Then each x ∈ F can be written as
1
x = a0 + ,
x1

10
 where a0 is equal to 2 or 3 and x1 > 1, so 2 < x < 4. Now let S1 , S2 : [2, 4] → [2, 4]
be given by
1 1
S1 (x) = 2 + and S2 (x) = 3 + .
x x
We claim that F is the attractor of S1 and S2 . To see this, we note that from the
definition of F and the continued fractions, we have that x ∈ F if and only if either
S1 (x) ∈ F or S2 (x) ∈ F . Thus F = S1 (F ) ∪ S2 (F ). Clear F ⊂ [2, 4] is bounded,
1 1 1
and non-empty, since 1 + 2+ 2+ 2+···
∈ F . To see that F is closed, note that its
complement is open, since if

x = a0 + 1/(a1 + 1/(a2 + 1/(a3 + · · · ))) ∈

/F

then ak 6= 2, 3 for some k, so numbers whose continued fraction expansion start

a0 + 1/(a1 + 1/(a2 + 1/(· · · + 1/(ak )))),

that is numbers close enough to x, are not in F .

Noting that S1 ([2, 4]) = [2 41 , 2 21 ] and S2 ([2, 4]) = [3 14 , 3 12 ] are disjoint, we may use
Propositions 9.6 and 9.7 to obtain estimates for the dimensions of F .
For x ∈ [2, 4], i = 1, 2,
1 −1 1
≤ |Si0 (x)| = | 2 | ≤ .
16 x 4
It follows from the mean-value theorem that, for x, y ∈ [2, 4], i = 1, 2,
1 1
|x − y| ≤ |Si (x) − Si (y)| ≤ |x − y|,
16 4
so in particular S1 and S2 are contractions. It follows from Propositions 9.6 and 9.7
that t ≤ dimH F ≤ s, where

2(1/16)t = 1 = 2(1/4)s ;

that is,
21−4t = 1 = 21−2s .
Thus t = 1/4 and s = 1/2; that is

1/4 ≤ dimH F ≤ 1/2.

10.8 For a real number x and a positive integer Q, the set {rx (mod1) : r = 0, 1, . . . , Q}
contains Q + 1 distinct numbers in the interval [0, 1], so two of these numbers will
differ by ≤ 1/Q; thus there are integers 0 ≤ r 6= s ≤ Q such that 0 < (s −
r)x (mod1) ≤ 1/Q. Letting q = |s − r| we have 0 < q ≤ Q and 0 ≤ qx (mod1) ≤ 1/q
so that kqxk ≤ 1/Q.
If x is rational, then kqxk = 0 is an integer for infinitely many q, and so kqxk ≤ q −1
infinitely often. If x is irrational, then for each K = 1, 2, . . . we may find, by
the above, positive integers qK ≤ K such that 0 < kqK xk ≤ 1/K ≤ 1/qK . Since
kqK xk = 6 0 and 1/K → 0 as K → ∞, there must be infinitely many distinct such qK .

11
 10.9 If xn − dy n = 1, then
1/n 1/n
xn − 1
 
1/n x 1
d = = 1− n .
yn y x

Both d and y are positive integers so that x > 1 and hence 0 < 1 − 1/xn < 1. Thus
 1/n
1 1
1− n < 1− n <1
x x

and so    1/n 
 1/n
d − x  x  1  x 1
= 1 − − 1 < n = n−1 .

y y x n  yx yx
  

Since xn = 1 + dy n , and d ≥ 1, we have x > y and so

 
 1/n x 
d −  < 1 .
 y  yn

If xn − dy n = 1 has infinitely many solutions (x, y), where x and y are positive
integers, then, since it can have only one solution for each value of y, it follows that
there are infinitely many positive integers y such that |d1/n − x/y| < 1/y n for some
integer x, that is, d1/n is n-well approximable.

10.10 For m and n integers, (x, y) ∈ F if and only if (x + m, y + n) ∈ F . Thus

dimH F = dimH G, where G = F ∩ ([0, 1] × [0, 1]).

For each integer q, let Gq denote the set of (x, y) ∈ [0, 1]×[0, 1] such that kqxk ≤ q 1−α
and kqyk ≤ q 1−α . Then Gq can be covered by the (q +1)2 boxes of side 2/q α centered
at the points (p/q, p0 /q), 0
S∞ where 0 ≤ p, p ≤ q. We denote this S∞ collection
S of boxes
by Cq . Clearly G ⊂ q=k Gq and so G can be covered by q=k U ∈Cq U . If k is
√
sufficiently large to ensure that 2 2/k α ≤ δ, then each of the boxes in Cq for q ≥ k
has diameter at most δ and so
∞
X √
Hδs (G) ≤ (q + 1)2 (2 2/q α )s .
q=k

If s = 3/α +  for some  > 0, then

∞ ∞ ∞
X √ X X
(q + 1)2 (2 2/q α )s < 41+s q 2 /q αs = 41+s 1/q 1+α < ∞
q=1 q=1 q=1

P∞ √
and so q=k (q + 1)2 (2 2/q α )s → 0 as k → ∞. Since k → ∞ as δ → 0, it follows
that
Hs (G) = lim Hδs (G) = 0,
δ→0

if s > 3/α. Thus dimH F = dimH G ≤ 3/α.

12
10.11 We use the sets of large intersection of Section 8.2. Let

F = {x : x is α-well approximable } = {x : kqyk ≤ q 1−α for infinitely many q }.

+ − + −
Define fm , fm : [0, ∞) → R by fm (x) = x1/2 −m, fm (x) = −x1/2 −m. By Proposition
−
10.4, F ∈ C s [0, ∞) for all s < 2/α. Since fm +
, fm are differentiable with derivative
bounded away from 0 on [m, M ] for all 0 < m < M , it follows from Proposition 8.9,
by taking a countable union, that

Fm = {x : (x + m)2 is α-well approximable}

+ −
= fm (F ∩ [0, ∞)) ∪ fm (F ∩ [0, ∞)) ∈ C s (−∞, ∞).
By Corollary 8.8,
∞
!
\
dimH {x : (x + m)2 is α-well approximable for all m} = dimH Fm ≥s
m=−∞

for all s < 2/α.

On the other hand,

dimH {x : (x + m)2 is α-well approximable for all m}

≤ {x : x2 is α-well approximable } = 2/α

as in Proposition 10.4, so

dimH {x : (x + m)2 is α-well approximable for all m} = 2/α.

13
 Chapter 11

11.1 If f 0 is continuous on [0, 1] then f 0 ([0, 1]) is bounded. Thus there exists 0 < c < ∞
such that |f 0 (t)| ≤ c, for each t ∈ [0, 1]. It follows from the mean-value theorem
that, for 0 ≤ t, u ≤ 1,
|f (t) − f (u)| ≤ c|t − u|.
Thus (11.2) is satisfied with s = 1 and so it follows from Corollary 11.2(a) that
H1 (graphf ) < ∞.
The graph of f is a continuous curve joining the points in the plane (0, f (0))
and (1, f (1)). The projection of this curve onto the x-axis is the interval [0, 1],
so H1 (graphf ) ≥ H1 ([0, 1]) = 1, by (6.1). Thus 0 < H1 (graphf ) < ∞.
To show that the graph is a regular set, we show that the graph is a rectifiable
curve and apply Lemma 5.5. For 0 = t0 < t1 < . . . < tm = 1 we have polygonal
approximations to the length of the graph given by
m m
X X
1/2
|(ti , f (ti )) − (ti−1 , f (ti−1 ))| = (ti − ti−1 )2 + (f (ti ) − f (ti−1 ))2
i=1 i=1

m m
X
1/2 X
≤ (ti − ti−1 )2 + c2 (ti − ti−1 )2 = (1 + c2 )1/2 |ti − ti−1 | = (1 + c2 )1/2 < ∞.
i=1 i=1

Hence the supremum of the lengths of the polygonal approximations to the graph
is finite, so graph f is a rectifiable curve in the plane, and so by Lemma 5.5 is a
regular 1-set.
An alternative approach would be to use the fact that there is a tangent at each
point on the curve so the 1-density equals 1 everywhere except at its endpoints.

11.2 Assume that |f (t) − f (u)| ≤ c|t − u| for 0 ≤ t, u ≤ 1. Define ψ : graphg →

graph(f + g) by ψ(t, g(t)) = (t, f (t) + g(t)). Then

|ψ(t, g(t)) − ψ(u, g(u))|2 = |(t, f (t) + g(t)) − (u, f (u) + g(u))|2
= |t − u|2 + |f (t) − f (u) + g(t) − g(u)|2
≤ |t − u|2 + 2|f (t) − f (u)|2 + 2|g(t) − g(u)|2
≤ c|t − u|2 + 2|g(t) − g(u)|2
≤ c1 (|t − u|2 + |g(t) − g(u)|2 )
= |(t, g(t)) − (u, g(u))|2 .

Thus ψ is Lipschitz. On the other hand,

|ψ(t, g(t)) − ψ(u, g(u))| = |(t, f (t) + g(t)) − (u, f (u) + g(u))|
≥ max{|t − u|, |f (t) − f (u) + g(t) − g(u)|}
≥ max{|t − u|, |g(t) − g(u)| − |f (t) − f (u)|}
≥ max{|t − u|, |g(t) − g(u)| − c|t − u|}
≥ ((c + 1)|t − u| + (|g(t) − g(u)| − c|t − u|)) /(c + 2)
= (|t − u| + |g(t) − g(u)|)/(c + 2)
≥ |(t, g(t)) − (u, g(u))|/(c + 2),

14
 using that max{a, b} ≥ ((c + 1)a + b)/(c + 2). Thus ψ is biLipschitz, so that
dimH graphg = dimH graph(f + g), with similar equality for box dimensions.

11.3 If the box dimensions of graph f and graphg exist, then it follows from Proposition
11.1 that

− log δ + log m−1

P
log Nδ i=0 Rf [iδ, (i + 1)δ]
dimB graphf = lim = lim
δ→0 − log δ δ→0 − log δ
Pm−1
log i=0 Rf [iδ, (i + 1)δ]
= 1 + lim
δ→0 − log δ

and Pm−1
log
Rg [iδ, (i + 1)δ] i=0
dimB graphg = 1 + lim .
δ→0 − log δ
If dimB graphf = dimB graphg + 2, for some  > 0, then it follows that there exists
δ0 > 0 such that, for all δ < δ0 ,

log m−1 log m−1

P P
i=0 Rf [iδ, (i + 1)δ] i=0 Rg [iδ, (i + 1)δ]
> +
− log δ − log δ
and hence
m−1
X m−1
X m−1
X
(− log δ) 
Rf [iδ, (i + 1)δ] > e Rg [iδ, (i + 1)δ] = (1/δ) Rf [iδ, (i + 1)δ].
i=0 i=0 i=0

Now, for any interval [t1 , t2 ] ⊂ [0, 1],

Rf [t1 , t2 ] − Rg [t1 , t2 ] ≤ Rf +g [t1 , t2 ] = sup |f (t) + g(t) − f (t) − g(t)|

t1 <t,u<t2
≤ Rf [t1 , t2 ] + Rg [t1 , t2 ].

So, for δ < δ0 ,

m−1
X m−1
X m−1
X
 
(1 − δ ) Rf [iδ, (i + 1)δ] ≤ Rf +g [iδ, (i + 1)δ] ≤ (1 + δ ) Rf [iδ, (i + 1)δ]
i=0 i=0 i=0

and hence
log m−1 log m−1 
P P
i=0 Rf +g [iδ, (i + 1)δ] i=0 Rf [iδ, (i + 1)δ] + log(1 + δ )
≤
− log δ − log δ
Pm−1
log i=0 Rf [iδ, (i + 1)δ] + δ 
≤ .
− log δ
Similarly,
Pm−1 Pm−1
log i=0 Rf +g [iδ, (i + 1)δ] log i=0 Rf [iδ, (i + 1)δ] − δ 
≥ .
− log δ − log δ
Thus Pm−1 Pm−1
log i=0 Rf +g [iδ, (i + 1)δ] log i=0 Rf [iδ, (i + 1)δ]
lim = lim
δ→0 − log δ δ→0 − log δ

15
 and hence dimB graphf + g = dimB graphf .
To see why we must require dimB graphf and dimB graphg to be unequal, consider
the case when dimB graphf > 1 and g(t) = −f (t) so that graph f + g is a straight
line and hence has box dimension one.

11.4 Given that (11.3) holds with 1 < s < 2, we have that, for all t ∈ [0, 1] and 0 < δ ≤ δ0 ,
there exists u with |t − u| ≤ δ such that
2−s
c|t − u|2−s
 
 f (u) − f (t) 
  ≥ cδ ≥ = c|t − u|1−s .
 t − u  |t − u| |t − u|

Hence  
 f (u) − f (t) 
lim   ≥ limc|t − u|1−s = ∞
u→t  t − u  u→t
and so the derivative f 0 (t) at t does not exist.
This condition (11.3) is satisfied by the Weierstrass function, see the penultimate
line of the Calculation of Example 11.3, so the Weierstrass function is nowhere
differentiable.
For the self-affine functions f of Example 11.4, note that from (11.9) there is a
number 0 <  < 1 such that m−1+ ≤ ci for all i. Thus, from the calculation of
Example 11.4, dm(−1+)k ≤ dci1 · · · cik ≤ Rf [Ii1 ,...,ik ] for each interval Ii1 ,...,ik , this
interval having length m−k . Thus given t ∈ [0, 1] and 0 < δ < 1 we may find an
interval I ≡ Ii1 ,...,ik containing t and with length |I| = m−k ≤ δ < m−k+1 . There
are points u1 , u2 ∈ I with

|f (u2 ) − f (u1 )| = Rf [Ii1 ,...,ik ] ≥ dm(−1+)k ≥ (δm−1 )1− ,

so since either |f (t) − f (u1 )| ≥ 21 |f (u2 ) − f (u1 )| or |f (t) − f (u2 )| ≥ 21 |f (u2 ) − f (u1 )|,
we conclude that there is u with |t − u| ≤ δ such that |f (t) − f (u)| ≥ 21 m−1 δ 1− .
This is condition (11.3) with s = 1 + , so by the first part of the question, the
self-affine function f is nowhere diferentiable.

11.5 The calculation is similar to that for Example 11.3. Given 0 < h < λ−1 , let N be
the integer such that
λ−(N +1) ≤ h < λ−N .
Then
N
X
|f (t + h) − f (t)| ≤ λ(s−2)k | sin(λk (t + h) + θk ) − sin(λk t + θk )|
k=1
X∞
+ λ(s−2)k | sin(λk (t + h) + θk ) − sin(λk t + θk )|.
k=N +1

Let g(t) = sin(λk t + θk ), then

|g 0 (t)| = λk | cos(λk t + θk ))| ≤ λk

16
and so, by the mean-value theorem,
N
X N
X
(s−2)k k k
λ | sin(λ (t + h) + θk ) − sin(λ t + θk )| ≤ λ(s−2)k λk h.
k=1 k=1

Since | sin t| ≤ 1 for all real values of t, we have

∞
X ∞
X
(s−2)k k k
λ | sin(λ (t + h) + θk ) − sin(λ t + θk )| ≤ 2λ(s−2)k .
k=N +1 k=N +1

So,
N
X ∞
X
|f (t + h) − f (t)| ≤ λ(s−2)k λk h + 2λ(s−2)k
k=1 k=N +1
(s−1)N (s−2)(N +1)
hλ 2λ
≤ 1−s
+ .
1−λ 1 − λs−2
Since λ−(N +1) ≤ h < λ−N , it follows that
hh1−s 2h2−s
|f (t + h) − f (t)| ≤ 1−s
+ s−2
≤ ch2−s ,
1−λ 1−λ
where c is independent of h. It now follows from Corollary 11.2(a) that dimB graphf ≤
s.

Similar arguments show that,

|f (t + h) − f (t) − λ(s−2)N (sin(λN (t + h) + θk ) − sin(λN t + θk ))|

N
X −1 ∞
X
(s−2)k k
≤ λ λ h+ 2λ(s−2)k
k=1 k=N +1

hλ(s−1)(N −1) 2λ(s−2)(N +1)

≤ +
1 − λ1−s 1 − λs−2
λ(s−2)N −s+1 2λ(s−2)(N +1)
≤ + ,
1 − λ1−s 1 − λs−2
if λ−(N +1) ≤ h < λ−N .
We now observe that, since sin is a periodic function with period 2π and is strictly
increasing on (−π/2, π/2) and strictly decreasing on (π/2, 3π/2), then there exists
c > 0 such that, for each T ∈ R we may choose H with 1/2 ≤ H < 1 such that
| sin(T + H) − sin T | > c.
We note that, if h = λ−N , then

λN (t + h) + θk − (λN t + θk ) = λN h = 1,

and, if h = λ−(N +1) , then

λN (t + h) + θk − λN t + θk = λN h = λ−1 < 1/2,

17
 provided that λ > 10. Thus, if λ > 10, then, for each t ∈ (0, 1) and each N , we may
choose h with λ−(N +1) ≤ h < λ−N such that | sin λN (t+h)+θk −sin λN (t+h)+θk | > c.
If λ is sufficiently large, then

λ(s−2)N −s+1 2λ(s−2)(N +1) λ(s−2)N

+ < ,
1 − λ1−s 1 − λs−2 2c

for all N , and so, for each t ∈ (0, 1) and each N , we may choose h with λ(−N +1) ≤
h < λ−N such that

|f (t + h) − f (t)| ≥ λ(s−2)N /(2c) ≥ λs−2 δ s−2 /(2c).

It now follows from Corollary 11.2(b) that dimB graphf ≥ s.

11.6 The calculation is similar to that for Example 11.3 and the solution to Exercise 11.5.
Given 0 < h < λ−1 , let N be the integer such that

λ−(N +1) ≤ h < λ−N .

Then
N
X
|f (t + h) − f (t)| ≤ λ(s−2)k |g(λk (t + h)) − g(λk t)|
k=1
X∞
+ λ(s−2)k |g(λk (t + h)) − g(λk t)|.
k=N +1

We note from its zig-zag form that g is a Lipschitz function with

|g(t1 ) − g(t2 )| ≤ |t1 − t2 | for all t1 , t2 ∈ R

and so
N
X N
X
(s−2)k k k
λ |g(λ (t + h)) − g(λ t)| ≤ λ(s−2)k λk h.
k=1 k=1

Since |g(t)| ≤ 1 for all real values of t, we have

∞
X ∞
X
(s−2)k k k
λ |g(λ (t + h)) − g(λ t)| ≤ 2λ(s−2)k .
k=N +1 k=N +1

Thus
N
X ∞
X
|f (t + h) − f (t)| ≤ λ(s−2)k λk h + 2λ(s−2)k ≤ ch2−s ,
k=1 k=N +1

where c is independent of h. It now follows from Corollary 11.2(a) that dimB graphf ≤
s.

In the same way,

|f (t + h) − f (t) − λ(s−2)N (g(λN (t + h)) − g(λN t))|

18
 λ(s−2)N −s+1 2λ(s−2)(N +1)
≤ + ,
1 − λ1−s 1 − λs−2
if λ−(N +1) ≤ h < λ−N .
We now observe that, since g is a periodic function with period 4 and is strictly
increasing on (−1, 1) and strictly decreasing on (1, 3), then there exists c > 0 such
that, for each t ∈ (0, 1) and each N ∈ N, we may choose h with 1/2 ≤ λN h < 1
such that |g(λN (t + h)) − g(λN t)| > c. If λ > 2, then this implies that, for each
t ∈ (0, 1) and each positive integer N , we may choose h with λ(−N +1) ≤ h < λ−N
such that |g(λN (t + h)) − g(λN t)| > c. If λ is sufficiently large, then the right-hand
side of the last displayed inequality above will be less than λ(s−2)N /(2c) for all N .
It then follows from the same inequality that, for each t ∈ (0, 1) and each N , we
may choose h with λ(−N +1) ≤ h < λ−N such that

|f (t + h) − f (t)| ≥ λ(s−2)N /(2c) ≥ λs−2 δ s−2 /(2c).

It now follows from Corollary 11.2(b) that dimB graphf ≥ s.

11.7 From Proposition 3.3(a) and (11.2) wesee that dimH f (F ) ≤ min{1, dimH F/(2 − s)}.
More interesting is the dimension of the subset of graphf given by E = {(t, f (t)) :
t ∈ F }. Suppose that F ⊂ [0, 1] intersects Nδ (F ) of the δ-mesh intervals. For such
an interval I the maximum range R[I] ≤ cδ 2−s by (11.2). Thus the portion of E
above the interval I may be covered by cδ 2−s δ −1 + 1 = cδ 1−s + 1 squares of side δ,
so the number of squares of side δ needed to cover E is Nδ (E) ≤ Nδ (F )(cδ 1−s + 1).
Hence
log Nδ (E) log Nδ (F ) log(cδ 1−s + 1)
≤ +
− log δ − log δ − log δ
so taking lower and upper limits as δ → 0 gives dimB E ≤ dimB F + (s − 1) and
dimB E ≤ dimB F + (s − 1).

11.8 Define a measure µ on graphf by µ(A) = LR1 {t ∈ [0, 1] : (t,Rf (t)) ∈ A} for A ⊂ R2 ,
1
so that for measurable g : R2 → R we have g(x)dµ(x) = 0 g(t, f (t))dt. Then
Z Z Z Z
dµ(x)dµ(y)
= |(t, f (t)) − (u, f (u))|−s dtdu
|x − y|s
Z Z
= [|t − u|2 + |f (t) − f (u)|2 ]−s/2 dtdu < ∞

by the given condition. Since µ is supported by graphf , it follows from Theorem

4.13(a) that dimH (graphf ) ≥ s.

11.9 Let S be a δ-mesh square of √ D. Then the maximum range over the square R[S] =
supt,u∈S |f (t) − f (u)| ≤ c(δ 2)3−s where 2 ≤ s < 3. Thus the portion of the surface
√
F = {(t, f (t)) : t ∈ D} above the square S may be covered by c(δ 2)3−s δ −1 + 1 =
c2(3−s)/2 δ 2−s + 1 cubes of side δ, so the number of mesh cubes of side δ needed to
cover F is Nδ (F ) ≤ (c2(3−s)/2 δ 2−s + 1)(δ −1 + 1)2 ≤ 20cδ −s for small δ. Hence
log Nδ (F ) log(20cδ −s ) log 20c − s log δ
≤ =
− log δ − log δ − log δ

19
 so taking upper limits as δ → 0 gives dimB F ≤ dimB F ≤ s.
We now give a surface analogue to Corollary 11.2(b). Suppose that there are num-
bers c > 0, δ0 > 0 and 2 ≤ s < 3 with the property: for each t ∈ S and 0 < δ ≤ δ0
there exists u such that |t − u| ≤ δ and

|f (t) − f (u)| ≥ cδ 3−s .

Then s ≥ dimB F where F = {(t, f (t)) : t ∈ S}.

To prove this, we note that for a square S of side δ the maximum range R[S] =
supt,u∈S |f (t) − f (u)| ≥ c(δ/2)3−s , so at least c(δ/2)3−s δ −1 = c2s−3 δ 2−s cubes of side
δ are needed to cover the portion of the surface F above a square S of side δ. Thus
the number of mesh cubes of side δ needed to cover F is at least (c2s−3 δ 2−s )(δ −2 ) ≥
c2s−3 δ −s . Hence
log Nδ (F ) log(c2s−3 δ −s ) log(c2s−3 ) − s log δ
≥ = ,
− log δ − log δ − log δ
so taking lower limits as δ → 0 gives dimB F ≤ s.

11.10 The transformations S1 and S2 are of the form (11.8) with

m = 2, a1 = 1/4, b1 = 0, r1 = 5/6, a2 = −1/4, b2 = 1/4, r2 = 5/6.

To verify that the attractor F of S1 and S2 is the graph of a continuous function,

we must check that conditions (11.9) and (11.10) are satisfied with S1 (p1 ), S2 (p1 )
and p2 not all collinear. We begin by noting that

1/m = 1/2 < 5/6 = r1 = r2

so that (11.9) is satisfied. Also, p1 = (0, b1 /(1 − r1 )) = (0, 0) and p2 = (1, (a2 +
b2 )/(1−r2 )) = (1, 0), so that S1 (p1 ) = p1 = (0, 0), S2 (p1 ) = (1/2, 1/4) and p2 = (1, 0)
are not all collinear.

We must now check that the fixed points p1 and p2 of S1 and S2 satisfy S1 (p2 ) =
S2 (p1 ). We have S2 (1, 0) = (1/2, 1/4), so S1 (p2 ) = S2 (p1 ) = (1/2, 1/4). Thus F is
the graph of a self-affine continuous fractal curve.
We calculate that S1 (q) = (1/4, 27/8) and S2 (q) = (3/4, 5/2) so that E2 may be
sketched.
The box dimension of F is given by the formula in Example 11.4:

log(r1 + r2 ) log(5/3)
dimB F = 1 + =1+ = 1.737
log m log 2
to three decimal places.

11.11 The transformations S1 , S2 and S3 are of the form (11.8) with

m = 3, a1 = 1/3, b1 = 0, r1 = 1/2, a2 = −2/3, b2 = 1/3, r2 = 1/2

20
 a3 = 1/3, b3 = −1/3, r3 = 1/2.
To verify that the attractor F of S1 , S2 and S3 is the graph of a continuous function,
we must check that conditions (11.9) and (11.10) are satisfied with, say, S1 (p1 ),
S2 (p1 ) and p2 not all collinear. We begin by noting that

1/m = 1/3 < 1/2 = r1 = r2 = r3

so that (11.9) is satisfied. Also, p1 = (0, b1 /(1 − r1 )) = (0, 0) and p3 = (1, (a3 +
b3 )/(1 − r3 )) = (1, 0), so that S1 (p1 ) = p1 = (0, 0), S2 (p1 ) = S2 (0, 0) = (1/3, 1/3)
and p2 = (1, 0) are not all collinear.
Now note that S1 (p3 ) = (1/3, 1/3) = S2 (p1 ) and S2 (p3 ) = (2/3, −1/3) = S3 (p1 ), so
(11.10) is satisfied and so F is the graph of a self-affine continuous fractal curve.
The points on the polygon E2 may be calculated as: (0, 0), (1/9, 5/18), (2/9, 1/18),
(1/3, 1/3), (4/9, 5/18), (5/9, −5/18), (2/3, −1/3), (7/9, −1/18), (8/9, −1/18), (1, 0).
The box dimension of F is given by the formula in Example 11.4:

log(r1 + r2 + r3 ) log(3/2)
dimB F = 1 + =1+ = 1.369
log m log 3
to three decimal places.

11.12 Let f : [0, 1] → R be the Weierstrass function. The calculation in Example 11.3
shows that there is a constant c such that |f (t + h) − f (t)| ≤ ch2−s if 0 < h ≤ 1, so
Z T
1
(f (t + h) − f (t))2 dt ≤ c2 h4−2s .
2T −T

On the other hand, the end of the calculation in Example 11.3 shows that for some
constant r1 there exist numbers h > 0 arbitraily close to 0 such that |f (t+h)−f (t)| ≥
r1 h2−s for all t, so for such h
Z T
1
(f (t + h) − f (t))2 dt ≥ r1 h4−2s .
2T −T

By (11.13)
 Z T 
1 1 2
log(C(0) − C(h)) = log limT →∞ (f (t + h) − f (t)) dt .
2 2T −T

Hence
log(C(0) − C(h)
limh→0 = 4 − 2s.
log h

21
 Chapter 12
12.1 In a similar way to Proposition 12.2, let F be an irregular 1-set in the unit square
such that proj0 F contains the interval [0, 1] of the x-axis. Define a mapping ψ :
[0, 1] × [0, 1] → R2 by ψ(x, y) = (x(1 + y 2 )1/2 , y). It is easy to see that ψ is bi-
Lipschitz and continuously differentiable, so ψ(F ) is an irregular 1-set. (Such maps
preserve irregularity, see for example, Exercise 5.2.) For all 0 ≤ d ≤ 1, there is a
point (d, b) ∈ F for some b, so there is a point (d(1 + b2 )1/2 , b) ∈ ψ(F ), that is a
point (a, b) ∈ ψ(F ) with a = d(1 + b2 )1/2 , that is a/(1 + b2 )1/2 = d.
By Proposition 12.1(b) the line set L(ψ(F )) has area 0. However, the line y = a+bx
is at perpendicular distance a/(1 + b2 )1/2 from the origin, so since there are points
in ψ(F ) for which this expression takes all values in [0, 1], the set L(ψ(F )) contains
lines at all perpendicular distances between 0 and 1 from the origin, as required.

12.2 The mapping φ : R2 → R2 given by φ(r, θ) = (1/r, θ) transforms a line at per-

pendicular distance R from the origin to a circle of radius 1/R through the origin.
Thus, taking E = L(ψ(F )) to be the set of the last exercise, φ(E) contains a circle
of radius 1/R through the origin for all 0 < R ≤ 1. Clearly, φ maps sets of area
0 to sets of area 0, so S∞φ(E) is a set of area 0 containing a circle of every radius
≥ 1. Taking a union n=1 n φ(E), where n1 φ(E) is the set φ(E) scaled about the ori-
1

gin by a factor n1 , we get a set of zero area containing a circle of every positive radius.

12.3 This is a variation of Proposition 12.2. Let F be any irregular 1-set such that the
projection onto the x-axis contains the unit interval [0, 1]. By Proposition 12.1(b)
the line set L(F ) has area 0. The line L(a, b) given by y = a + bx cuts the y axis at
a = proj0 (a, b), so since [0, 1] ⊂ proj0 F , the line set L(F ) contains lines cutting the
y-axis
S∞ at every point of the interval [0, 1]. Taking a countable union of translates
n=−∞ (L(F ) + (0, n)) gives a set of area 0 containing a line cutting the y-axis at
each of its point, which is essentially the required set.

12.4 This is an extension of the second part of Proposition 12.2. Writing L(a, b) for the
set of points in the plane on the line y = a+bx, let E = {(a, b) : L(a, b) ⊂ F } so that
F ⊃ L(E). Then, since F contains a line in every direction θ for θ ∈ A, projπ/2 E ⊃
{tan θ : θ ∈ A}. Thus dimH E ≥ dimH projπ/2 E ≥ dimH A. By Proposition 12.1(a),
dimH F ≥ dimH L(E) ≥ min{2, 1 + dimH E} ≥ min{2, 1 + dimH A} = 1 + dimH A.

12.5 Let A be a Borel subest of R2 of area a. For each θ ∈ [0, π) define Aθ to be the set
(1 + c2 )−1 (Aθ ∩ Lc ), where c = tan θ, Lc is the line x = c, and we have scaled the set
Aθ ∩ Lc by a factor (1 + c2 )−1 . By theorem 6.9 there is a compact set F ⊂ R2 such
that projθ F ⊃ Aθ for all θ and L1 (projθ F ) = L1 (Aθ ) for almost all θ. By duality,
writing L(F ) for the line set of F , we have L(F ) ∩ Lc is congruent to projθ F . It
follows that for all c we have L(F ) ∩ Lc ⊃ A ∩ Lc with L1 (L(F ) ∩ Lc ) = L1 (A ∩ Lc )
for almost all c, so we have L(F ) ⊃ A and on integrating L2 (L(F )) = L2 (A), as
required.

22
12.6 Note that if µ is supported by F then f (z) = F (z − w)−1 dµ(w) is analytic at
R

z ∈ C \ F . Thus for F to be removable, there would have to be an analytic function

f˜(z) with f (z) = f˜(z) for z ∈ C \ F . In particular, by Cauchy’s identity, for every
contour C we would require
Z Z Z Z Z Z
˜ dµ(w) dz
f (z)dz = f (z)dz = dz = dµ(w)
C C C F (z − w) F C (z − w)
Z
= 2πidµ(w) = 2πiµ(F ) > 0,
F

provide C encloses F . By Cauchy’s theorem, f˜(z) cannot be analytic on any domain

containing C, so F is not removable.
If 1 < dimRH F , Theorem 4.13(b) gives a mass distribution µ on F and a constant M
such that |z − w|−1 dµ(w) ≤ M for all z ∈ R2 . Identifying R2 with C gives
Z  Z
 dµ(w)  dµ(w)
 (z − w)  ≤ ≤ M,

|z − w|
so by the first part, F is not removable.

12.7 Let F = {x1 , . . . , xk } be a finite subset of C. Let V be an open domain containing

F and let f be a bounded analytic function on V \ F , say |f (z)| ≤ M . Let C be
a contour in V enclosing F . Given  > 0 let C1 , . . . , Ck be contours with centres
x1 , . . . , xk and radii ; we may assume that  is small enough so that the contours
are disjoint. By Cauchy’s integral formula,
Z Z
1 f (w)dw X 1 f (w)dw
−
2πi C z − w j
2πi Cj z − w

for z inside C but outside all of the Cj (to see this make cuts to join the contours
Cj to C to form a single contour). Thus
 Z  X Z 

f (z) − 1 f (w)dw 
≤ 1  f (w)dw  X 1 M 2π
≤
 
2πi C z − w 2π  Cj z − w  2π d(z, Cj )
  
j j

where d(z, Cj ) is the distance from z to the contour Cj . Letting  → 0 gives that
Z
1 f (w)dw
f (z) = .
2πi C z − w
But this defines an analytic function throughout the interior of V including at the
points xj , so this formula defines the required analytic extension of f .
Note that a slight modification of this proof shows that any compact set F with
H1 (F ) = 0 is removable.
To see that the unit circle F is not removable, consider the function on C \ F given
by f (z) = 2 if |z| < 1 and f (z) = 1/z if |z| > 1. Then f is analytic and bounded on
C \ F , but clearly has no analytic extension to any region containing F since such
an extension would be discontinuous on F . Thus F is not removable.

23
 12.8 Define a function g by taking g(t) to be the point x such that graphf has a line
of support with slope t at x; then g is defined on some maximal sub-interval
I of (−π/2, π/2). For each x ∈ R, we have that g −1 (x) is a closed interval,
which is a single point if and only if graphf has a unique tangent at x, that
is if and only if f is differentiable at x. For each k define the set Ak = {x :
g −1 (x) is an interval of length ≥ 1/k}. Then since the intervals g −1 (x) and g −1 (y)
are disjoint if x 6= y and g −1 (R) = I ⊂ (−π/2, π/2), it follws that Ak contains
at most π/k points,S so in particular Ak is finite. The set of x at which f is not
differentiable is ∞ k=1 Ak which is therefore either finite or countable.

12.9 The function f : R2 → R given in coordinate form by f (x, y) = |x| is easily seen to
be convex, with set of non-differentiability the y-axis, which has Hausdorff dimension
1.
The Weierstrass function f : R → R given by (11.4) is continuous but nowhere
differentiable, see Exercise 11.4, so the function g : R2 → R given by g(x, y) = f (x)
is nowhere differentiable on the plane, i.e. the non-differentiability set has Hausdorff
dimension 2.

12.10 Let G be a subgroup of (R, +). We have two cases:

(i) For all  > 0 there exists x ∈ G ∩ (0, ). Then for all y ∈ R and  > 0 we have
nx ∈ (y − , y + ) for some integer n, and also nx ∈ G, as an n-fold sum of x or −x.
Hence the set of elements of G is dense in R. Thus for every interval [a, b] we have
dimB (G ∩ [a, b]) = dimB (G ∩ [a, b]) = dimB [a, b] = 1, using Proposition 2.6. (In any
meaningful sense dimB G = 1 also, though we have not defined box dimension for
unbounded sets.)
(ii) There exists  > 0 such that G ∩ (0, 2) = ∅. Then for all y ∈ R the interval
(y − , y + ) contains at most one element of G (if it contained two such elements
their difference would be in (0, 2). It follows that for every interval [a, b] the set
G ∩ [a, b] is finite, so that dimB (G ∩ [a, b]) = 0, using Proposition 2.6. (In any
meaningful sense dimB G = 1 also, though we have not defined box dimension for
unbounded sets.)

12.11 For 0 < t < 2 let F be the set in Example 12.4 with s = t/2, so that F = ∞
S
r=1 Fr
is a subgroup of R with dimH F = s and dimH Fr = dimB Fr = s (from Exam-
ple 4.7). Consider F × F . Then F × F is a group, with (0, 0) ∈ F × F , with
(x1 , y1 ) + (x2 , y2 ) = (x1 + x2 , y1 + y2 ) ∈ F × F whenever x1 , y1 , x2 , y2 ∈ F , us-
ing the group properties of F , and with −(x, y) = (−x, −y) ∈ F × F whenever
x, y ∈ F .SBy Product formulae 7.2 and 7.3 dimH (Fr × Fr ) = 2dimH F = t; then
F ×F = ∞ r=1 (Fr × Fr ) since this is a union of increasing sets, so dimH (F × F ) = t.
Thus F × F is a subgroup of R2 with dimH (F × F ) = t.

24
 Chapter 13
13.1 Let f (x) = 2(1 − |2x − 1|), so if x ≤ 1/2, then f (x) = 2(1 + 2x − 1) = 4x, and
if x ≥ 1/2, then f (x) = 2(1 − 2x + 1) = 4 − 4x. Note that f has a maximum at
x = 1/2 with f (1/2) = 2.
We note f that there are two branches of f −1 defined on [0, 1]. It follows from the
definition of f that these are the two functions S1 , S2 : [0, 1] → R defined by
x x
S1 (x) = , S2 (x) = 1 − .
4 4
These functions satisfy f (S1 (x)) = f (S2 (x)), for x ∈ [0, 1]. Also, S1 and S2
are both contractions, since |Si (x) − Si (y)| = |x − y|/4, for x, y ∈ [0, 1] and
i = 1, 2, and so it follows from TheoremT9.1 that there is a compact set F sat-
isfying F = S1 (F ) ∪ S2 (F ) given by F = ∞ k
k=0 S ([0, 1]). Clearly F is the attractor
for f since f (F ) = f (S1 (F )) ∪ f (S2 (F )) = F ∪ F = F .

We now show that F is a repeller for f . We begin by noting that, if x < 0, then
f n (x) → −∞ as n → ∞. Also, if x > 1, then f (x) < 0 and so f n (x) → −∞ as
n → ∞. Thus any repeller for f must be contained in [0, 1]. If x ∈ [0, 1] \ F , then,
for some k ∈ Z+ , x ∈
/ S k ([0, 1]) and so f k (x) ∈
/ [0, 1]. Thus f n (x) → −∞ as n → ∞,
for any x ∈
/ F , and so F is indeed a repeller for f .

We now show that f is chaotic on F by denoting the points of F by xi1 ,i2 ,... =
T ∞
k=1 Si1 ◦Si2 · · · Sik ([0, 1]) with ij = 1, 2. We begin by noting that |xi1 ,i2 ,... −xi1 ,i2 ,... | ≤
0 0

4−k if (i1 , . . . , ik ) = (i01 , . . . , i0k ) and that f (xi1 ,i2 ,... ) = xi2 ,i3 ,... . Now suppose that the
sequence (i1 , i2 , . . .) is an infinite sequence with every finite sequence of 1s and 2s
appearing as a consecutive block of terms. In this case, the orbit {f k (xi1 ,i2 ,... )} is
dense in F since, if xi01 ,i02 ,... ∈ F and q ∈ Z+ , then there exists k ∈ Z+ such that
(i01 , . . . , i0q ) = (ik+1 , . . . , ik+q ) and hence

|f k (xi1 ,i2 ,... ) − xi01 ,i02 ,... | = |xik+1 ,ik+2 ,... − xi01 ,i02 ,... | ≤ 4−q .

Now let xi1 ,i2 ,... denote any point in F . The point xi1 ,...,ik ,i1 ,...,ik ,i1 ,... is a periodic point
of f in F and
|xi1 ,i2 ,... − xi1 ,...,ik ,i1 ,...,ik ,i1 ,... | ≤ 4−k .
Thus the periodic points of f are dense in F .

Finally we show that the iterates of f have sensitive dependence on initial conditions.
Again, let xi1 ,i2 ,... denote any point in F and let xi01 ,i02 ,... be another point in F with
(i1 , . . . , ik ) = (i01 , . . . , i0k ) but ik+1 6= i0k+1 . One of f k (xi1 ,i2 ,... ), f k (xi01 ,i02 ,... ) belongs to
[0, 1/4] whilst the other one belongs to [3/4, 1]. Thus

|f k (xi1 ,i2 ,... ) − f k (xi01 ,i02 ,... )| ≥ 1/2

even though
|xi1 ,i2 ,... − xi01 ,i02 ,... | ≤ 4−k .

F is the attractor of the similarities S1 and S2 given above, which have ratios
r1 = r2 = 1/4 and satisfy the open set condition (9.12) with V = (0, 1) (since the

25
 sets S1 (V ) = (0, 1/4) and S2 (V ) = (3/4, 1) are disjoint and contained in V ). Thus
it follows from Theorem 9.3 that dimH F = dimB F = s, where s is given by
2
X
1= ris = 2(1/4)s .
i=1

Thus s log(1/4) = log(1/2) and so

log(1/2) log(1/2) 1
s= = 2
= .
log(1/4) log(1/2) 2

13.2 Inverting each of the three parts of the mapping defining f we get an associated
IFS {S1 , S2 , S3 } on [0, 5] by taking
1 1 1
S1 (x) = x, S2 (x) = 2 − x, S3 (x) = 2 + x.
5 5 5
Then f (Si (x)) = x for i = 1, 2, 3 and x ∈ [0, 5].
The Si are contracting similarities,T so the IFS has an attractor F satisfying F =
S1 (F )∪S2 (F )∪S3 (F ), with F = ∞ k=0 S k
([0, 5]). From the definition of the Si as the
−1
branches of f , we see that f (F ) = F . To see that F is a repeller for f , note that if
x > 5 then f (x) = 5x − 10 = 3x + 2x − 10 > 3x, so f k (x) ≥ 3k x → T∞∞, and if x < 0
k k
then f (x) = 5x so f (x) ≤ 5 x → −∞. If x ∈ [0, 5] \ F then x ∈ / k=0 S k ([0, 5]), so
k
that f (x) ∈/ [0, 5] for some positive integer k, so either f (x) → ∞ or f k (x) → −∞.
k

Thus all points outside F are iterated to ±∞, so F is a repeller.

The open set condition holds for the IFS {S1 , S2 , S3 } taking (0, 5) as the open set,
with S1 (0, 5) = (0, 1), S2 (0, 5) = (1, 2), S3 (0, 5) = (2, 3); since each Si is a similar-
ity of ratio 1/5 Theorem 9.3 gives that dimH F = s where 3 × 5−s = 1, that is
s = log 3/ log 5.

13.3 This is similar to the argument for the logistic map for large λ. Assuming λ >
1, Write a = π1 sin−1 λ1 so that fλ maps each of the intervals [0, a] and [1 − a, 0]
monotonically onto [0, 1]. Inverting the restriction of fλ to each of these intervals
we get an associated IFS {S1 , S2 } on [0, 1] givne by
1 x 1 x
S1 (x) = sin−1 , S2 (x) = 1 − sin−1 .
π λ π λ
Then f (Si (x)) = x for i = 1, 2 and x ∈ [0, 1].
Differentiating, |Si0 (x)| = π −1 (λ2 − x2 )−1/2 , for i = 1, 2, so using the mean value
theorem.
1 |Si (x) − Si (y)| 1
= inf |Si0 (x)| ≤ ≤ sup |Si0 (x)| = √ (∗)
πλ x∈[0,1] |x − y| x∈[0,1] π λ2 − 1

for x 6= y. In particular, if (π + 1)1/2 /π < λ, S1 and S2 are contractions,

T∞ k so the IFS
has an attractor F satisfying F = S1 (F ) ∪ S2 (F ) with F = k=0 S ([0, 5]). From
the definition of the Si as the branches of the inverse of fλ , we see that fλ (F ) = F .

26
 / ∞ k
T
To see that F is a repeller for fλ , note that if x ∈ [0, 1] \ F then x ∈ k=0 S ([0, 1]),
so that fλk (x) ∈
/ [0, 1] for some positive integer k. Thus all points outside F are
iterated to outside [0, 1], so F is a repeller.
Since S1 ([0, 1]) and S2 ([0, 1]) are disjoint, Propositions 9.6 and 9.7 give s ≤ dimH F ≤
t, where
2(π(λ2 − 1)1/2 )−s = 1 = 2(πλ)−t
by (∗), giving

log 2/ log(π(λ2 − 1)1/2 ) ≤ dimH F ≤ log 2/ log(πλ)

. Thus when λ is large, dimH F ' log 2/ log(πλ).

13.4 Suppose that fλn (x) → l as n → ∞. Then fλn+1 (x) → l as n → ∞. Since

fλn+1 (x) = λfλn (x)(1 − fλn (x)),

it follows that l = λl(1 − l). Thus either l = 0 or 1 = λ(1 − l); that is, either l = 0
or l = 1 − 1/λ.
If x ∈ (0, 1), then
1
0 < f1/2 (x) = x(1 − x) < x/2.
2
n
Thus f1/2 (x) is a decreasing sequence converging to 0.
Note that f2 (x) = 2x(1 − x). Thus, if x ∈ (0, 1/2), then

x < f2 (x) < 1/2.

Thus f2n (x) is an increasing sequence which is bounded above. It follows that f2n (x)
converges, so from the first part of the question, that f2n (x) converges to 1 − 1/λ =
1/2. If x ∈ (0, 1/2), then 0 < f2 (x) < x and fλn (x) increases to 1/2 = 1 − 1/λ.
If x ∈ (1/2, 1), then f2n (x) ∈ (0, 1/2) and so f2n (x) also converges to 1/2. Finally,
f2 (1/2) = 1/2 and so f2n (1/2) trivially converges to 1/2.
Finally, we consider f4 . We note that 0 is an unstable fixed point of f4 (since
f4 (0) = 0 and f40 (0) = 4 > 1) and so, if f4n (x) converges to 0, then there must
be some integer m for which f4m (x) = 0. Now f4 (1) = 0 and f4 (1/2) = 1 so that
f42 (1/2) = 0. There are no other non-zero preimages of 0 and 1 under f4 and hence
1/2 is the only non-zero preimage of 0 under f42 . If 0 < x < 1, then there are
exactly two points in (0, 1) which map to x under f4 . Thus, for each k ∈ N, there
are exactly 2k points in (0, 1) which map to x under f4k and hence to 0 under f4k+2 .
These are the only points which converge to 0 under iteration.
Similarly, 3/4 = 1 − 1/λ is an unstable fixed point of f4 and so, if f4n (x) converges
to 3/4, then there must be some integer m for which f4m (x) = 3/4. For each k ∈ R,
there are exactly 2k points in (0, 1) which map to 3/4 under f4k and hence converge
to 3/4 under iteration.
We have shown that there are countably many points x ∈ (0, 1) for which f4n (x) → 0
as n → ∞ and countably many points x ∈ (0, 1) for which f4n (x) → 3/4 as n → ∞.
This leaves infinitely many points x ∈ (0, 1) for which f4n (x) cannot converge (since
any convergent sequence of iterates must converge to either 0 and 3/4).

27
13.5 (i) We will use proof by induction. If k = 0, then xk = 12 (1−exp(2k a)) = 12 (1−exp a) =
x. Thus the formula is correct for k = 0. If the formula is correct for k, then

xk+1 = f1/2 (xk ) = 2xk (1 − xk )

= 2(1/2)(1 − exp(2k a))(1 − (1 − exp(2k a))/2)
= (1 − exp(2k a))(1/2 + exp(2k a)/2)
= (1/2)(1 − exp(2k a))(1 + exp(2k a))
= (1/2)(1 − exp(2k+1 a)).

Thus, if the formula is correct for k, then it is also correct for k + 1. Since we have
shown that the formula is correct for k = 0, it follows by induction that the formula
is correct for all k.

(ii) We will use proof by induction. If k = 0, then xk = sin2 (πa) = x. Thus the
formula is correct for k = 0. If the formula is correct for k, then

xk+1 = f4 (xk ) = 4xk (1 − xk )

= 4 sin2 (2k πa)(1 − sin2 (2k πa))
= 4 sin2 (2k πa) cos2 (2k πa)
= [2 sin(2k πa) cos(2k πa)]2
= sin2 (2k+1 πa).

Thus, if the formula is correct for k, then it is also correct for k + 1. Since we have
shown that the formula is correct for k = 0, it follows by induction that the formula
is correct for all k.

If a = 0 · a1 a2 . . . in binary form, then

xk = sin2 (2k πa) = sin2 (a1 a2 . . . ak · ak+1 . . . π)

= sin2 (a1 . . . ak π + 0 · ak+1 . . . π)
= cos2 (a1 . . . ak π) sin2 (0 · ak+1 . . . π)
= sin2 (0 · ak+1 . . . π).

So, if a = 0 · a1 . . . ap a1 . . . ap a1 . . ., then

xp = sin2 (0 · a1 . . . ap a1 . . . ap a1 . . . π) = x;

that is, f4p (x) = x. We must now show that it is possible to choose a1 , . . . , ap in
such a way as to ensure that the periodic point x is unstable; that is, |(f4p )0 (x)| > 1.
Now
p−1
Y
(f4p )0 (x) = f40 (f4i (x)),
i=0

by the chain rule, and, for any x̂, we have f4 (x̂) = 4x̂(1 − x̂) so that f40 (x̂) = 4 − 8x̂.
If x̂ = sin2 (πâ0 ), then

f40 (x̂) = 4(1 − 2 sin2 (πâ)) = 4 cos(2πâ).

28
 Thus p
Y
(f4p )0 (x) = 4p cos(0 · ai . . . ap a1 . . . ap a1 . . . π).
i=1

If p is odd, then we put

a1 . . . ap = 1010 . . . 101
and, if p is even, then we put

a1 . . . ap = 11010 . . . 101.

If
0 · ai . . . ap a1 . . . ap a1 . . . π > 0 · 101π = 5π/8
or
0 · ai . . . ap a1 . . . ap a1 . . . π < 0 · 011π = 3π/8,
then

| cos(0 · ai . . . ap a1 . . . ap a1 . . . π)| ≥ | cos(3π/8)| = | cos(5π/8)| > 0.38.

Otherwise, 0 · ai . . . ap a1 . . . ap a1 . . . π is close to π/2 and

| cos(0 · ai . . . ap a1 . . . ap a1 . . . π) cos(0 · ai+1 . . . ap a1 . . . ap a1 . . . π)|

≥ | cos(0 · 011101 . . . π) cos(0 · 11101π)|
≥ | cos(0 · 01111π) cos(0 · 111π)|
= | cos(15π/32) cos(7π/8)| > 0.098 × 0.92 > 0.09.

Now (0 · 38)2 > 0 · 09 and hence

|(f4p )0 (x)| > (0.09)p/2 × 4p = (1 · 2)p > 1

as required.
Finally, we show that f4 has a dense orbit. Suppose that (a1 , a2 , . . .) is an infinite
sequence with every finite sequence of 0s and 1s appearing as a consecutive block
of terms. We claim that the orbit {f4k (x)} is dense in [0, 1], if x = sin2 (πa), where
a = 0 · a1 a2 . . .. To show that this is true, we take another point x0 ∈ [0, 1] with
x0 = sin2 (πa0 ), where a0 = 0 · a01 a02 . . . in binary form. For each q ∈ N, there exists
k ∈ N such that (a01 , . . . , a0q ) = (ak+1 , . . . , ak+q ). Thus

|f4k (x) − x0 | = | sin2 (0 · ak+1 ak+2 . . . π) − sin2 (0 · a01 a02 . . . π)|

= | sin2 ((a0 + )π) − sin2 (a0 π)|,

where || ≤ 2−q . Since we can choose q to be arbitrarily large, it follows that the
orbit {f4k (x)} comes arbitrarily close to the point x0 .

13.6 Let f : E → E be given by

(0 ≤ x ≤ 21 )

(2x, λy)
f (x, y) =
(2x − 1, µy + 1/2) ( 21 < x ≤ 1)

where 0 <Tλ, µ < 1/2, and let Ek = f k (E). Then Ek is a decreasing sequence of sets
and F = ∞ k
k=0 Ek satisfies f (F ) = F . Each set Ek is made up of 2 horizontal strips

29
 of height at most M k , where M = max(λ, µ). Thus F is made up of horizontal lines
with at least one line in each strip of Ek . If (x, y) ∈ E, then f k (x, y) ∈ Ek and so
the distance of f k (x, y) from F is at most M k . Since M < 1/2, it follows that every
point in E is attracted to F .
We now find the Hausdorff dimension of F . We begin by noting that (0, 1] × F1 ⊂
F ⊂ [0, 1] × F1 , where F1 is the attractor of the mappings S1 , S2 : [0, 1] → R defined
by
S1 (y) = λy, S2 (y) = µy + 1/2.
Now S1 and S2 are contractions with ratios r1 = λ and r2 = µ. Since they also
satisfy the open set condition (9.12) with V = (0, 1), it follows from Theorem 9.3
that dimH F1 = dimB F1 = s, where s is given by 1 = r1s + r2s = λs + µs . It follows
from Corollary 7.4 that

dimH ((0, 1] × F1 ) = dimH ([0, 1] × F1 ) = 1 + s

and hence, by monotonicity, dimH F = 1 + s, where λs + µs = 1.

13.7 It may be verified computationally that the four sides of the quadrilateral specified
are mapped onto parabolae which lie inside the quadrilateral.
Iterates of a typical point give a plot similar to Figure 13.8.

13.8 This is very similar to the argument for the solenoid in Section 13.4. With

f (φ, w) = (3φ(mod2π), aw + 21 φ̂),

k k k
we see
T∞thatkf (D) is a solid tube of radius a going round D 3 times. The set
F = k=1 f (D) is compact and invariant under f and attracts all points of D.
To find the dimensions of F , let Pφ be the half-plane bounded by the central axis L
and cutting the central circle of D at (φ, 0). Then f k (C) is a smooth closed curve
traversing the torus 3k times with total length at most 3k c, where c is independent
of k. The set f k (D) is a fattening of the curve f k (C) to a tube of radius ak , so
it may be covered by a collection of balls of radii 2ak spaced at intervals ak along
f k (C). Then 2 × 3k ca−k such balls will suffice, so applying Proposition 4.1 in the
usual way, we get dimH F ≤ dimB F ≤ 1 − log 3/ log a.
For a lower estimate, we consider the sections F ∩ Pφ for each φ. The set f (D) ∩ Pφ
contains three discs of radius a situated symmetrically with centres at least 14 apart.
Each of these discs contains three discs of f 2 (D) ∩ Pφ of radius a2 with centres at
least 41 a apart, and so on. Thus we may regard F ∩ Pφ as formed by a standard
nested construction, the kth stage consisting of 3k discs of radius ak with centres
separated by at least 41 ak−1 . We may define a mass distribution µ on F ∩ Pφ such
that each of the 3k level-k discs has mass 3−k . A standard application of the mass
distribution principle gives that dimH (F ∩ Pφ ) ≥ − log 3/ log a. Since F is built up
from sections F ∩ Pφ (0 ≤ φ < 2π, a higher dimensional version of Proposition 7.9
gives that dimH F ≥ 1 − log 3/ log a, so that dimH F = dimB F = 1 − log 3/ log a.
The chaotic behaviour of f may be examined by noting that if φ/2π = 0.a1 a2 . . .
to base 3, then f k (φ, w) = (φk , vk ), where φk /2π = 0.ak+1 ak+2 . . . and where the

30
 integer with base 3 representation ak ak−1 . . . ak−d+1 determines which of the 3d discs
of f d (D) ∩ Pφk the point vk belongs to for d ≤ k. By choosing digits a1 , a2 , . . .
suitably, it is easy to produce orbits that are dense in f or which are periodic.

13.9 Suppose that x = f (t) +  for some  6= 0. Then

h(t, x) = (λt, λ2−s (x − g(t)))
= (λt, λ2−s ( + f (t) − g(t)))
X∞
2−s
= (λt, λ ( + λ(s−2)k g(λk t)))
k=1
∞
X
= (λt, λ2−s  + λ(s−2)(k−1) g(λk t))
k=1
X∞
= (λt, λ2−s  + λ(s−2)k g(λk+1 t))
k=0
2−s
= (λt, λ  + f (λt)).

So, if (t, x) is at a vertical distance  from graph f , then h(t, x) is at a vertical

distance λ2−s  >  from graph f . Since g is periodic and differentiable, there exists
M > 0 such that |g(t)| ≤ M for all t ∈ R and hence there exists M 0 > 0 such that
|f (t)| ≤ M 0 for all t ∈ R. It follows that the distance of hn (t, x) from graph f tends
t0 infinity as n → ∞ and so graph f is indeed a repeller for h as claimed.

13.10 With the notation of Section 10.1, take µ to be the probability measure P defined
in (10.2). With F P = F (p0 , . . . , pm−1 ), Proposition 10.1 showed that P(F ) = 1 and
that dimH F = −( m−1 i=0 pi log pi )/ log m which is strictly < 1 provided that the pi
are not all equal, giving by (13.16) dimH P ≤ dimH F < 1. However, F is dense in
[0, 1], so that dimH sptP = 1, giving dimH P < dimH sptP, as required.

13.11 Differentiating f we see that the Jacobian matrix of f is

 
∂f 1 1
J≡ =
∂(x, y) 1 2
at all (x, y) such that f (x, y) is not on the boundary of the unit square E. Since
the determinant of J is 1, the map is area preserving.
√
The eigenvalues of J are (3 ± 5)/2, so we may choose orthogonal  axes √such that, 
(3 + 5)/2 0
√
with respect to these axes, J is represented by the diagonal matrix .
0 (3 − 5)/2
By the chain rule, the derivative of the kth iterate f k is J k , which with respect to
these axes is  √ 
((3 + 5)/2)k √0
0 ((3 − 5)/2)k
at all (x, y) such that f j (x, y) is not on the boundary of the unit square E for all
j = 1, 2, . . . , k, which is the case for L2 -almost all (x, y) ∈ E. It follows from (13.9)
that, for almost all (x, y), the Liapunov exponents are
√ √
log(3 + 5)/2 and log(3 − 5)/2,

31
and thus these values are the Liapounov exponents of the cat map.

32
Chapter 14
14.1 We may choose distinct points x1 , x2 ∈ E and a number 0 < r1 ≤ 2−1 such that
the discs B(x1 , r1 ) and B(x2 , r1 ) are disjoint. Since x1 and x2 are not isolated, we
may, for i = 1, 2, choose xi,1 , xi,2 ∈ E ∩ B(xi , r1 ) and a number 0 < r2 ≤ 2−2
such that the discs B(xi,1 , r2 ) and B(xi,2 , r2 ) are contained in B(xi , r1 ) and are
disjoint. Proceeding in this way, we may find distinct points xi1 ,i2 ,...,ik ∈ E and discs
B(xi1 ,i2 ,...,ik , rk ) with rk ≤ 2−k such that B(xi1 ,i2 ,...,ik ,1 , rk+1 ) and B(xi1 ,i2 ,...,ik ,2, rk+1 )
are disjoint subdiscs of B(xi1 ,i2 ,...,ik , rk ).
For every infinite sequence i1 , i2 , . . . of 1s and 2s, let xi1 ,i2 ,... = ∞
T
i=0 B(xi1 ,i2 ,...,ik , rk ),
which is a single point as the intersection of closed discs of radii tending to 0. Since
E is closed, xi1 ,i2 ,... ∈ E as the limit as k → ∞ of xi1 ,i2 ,...,ik ∈ E. Moreover, the xi1 ,i2 ,...
are distinct, for if i1 , i2 , . . . , ik−1 , ik , ik+1 , . . . and i1 , i2 , . . . , ik−1 , jk , jk+1 , . . . are dis-
tinct sequences of 1s and 2s with ik 6= jk , then xi1 ,i2 ,...,ik−1 ,ik ,ik+1 ,... ∈ B(xi1 ,i2 ,...,ik , rk )
and xi1 ,i2 ,...,ik−1 ,jk ,jk+1 ,... ∈ B(xi1 ,i2 ,...,ik−1 ,jk , rk ) are in distinct level k discs. There are
uncountably many sequences i1 , i2 , . . . (such sequences are in correspondance with
real numbers in [0, 1] numbers expressed to base 2), so there are uncountably many
points xi1 ,i2 ,... ∈ E.

14.2 We use the fact that, if f (z) = (α2 z 2 +2αβz +β 2 +c−β)/α, then J(f ) = h−1 (J(fc )),
where fc (z) = z 2 + c and h(z) = αz + β, see Section 14.2. Here f (z) = z 2 + 4z + 2
and so α2 /α = 1, 2αβ/α = 4 and (β 2 + c − β)/α = 2. Thus α = 1, β = 2 and
c = 2 + β − β 2 = 0. So fc (z) = z 2 and hence J(fc ) is the unit circle. Since
h−1 (z) = (z − β)/α = z − 2, it follows that J(f ) is the circle of radius 1 whose
centre is at −2.

14.3 We recall from Section 14.2 that if h(z) = αz + β and f (z) = h−1 (fc (h(z))) then
J(f ) = h−1 (J(fc )). Taking h(z) = z + i so h−1 (z) = z − i, this gives that the Julia
set of f (z) = (z + i)2 + c − i = z 2 + 2iz − 1 + c − i is J(fc ) − i, which is congruent
to J(fc ). This Julia set is connected if and only if c ∈ M , thus the Julia set of
f (z) = z 2 + 2iz + b is connected if and only if c ∈ M where b = −1 + c − i, that is
if and only if b + 1 + i ∈ M .

1
14.4 If |c| ≤ 4
and |z| ≤ 12 , then

1 1 1
|fc (z)| = |z 2 + c| ≤ |z|2 + |c| ≤ ( )2 + = .
2 4 2

Thus if |z| ≤ 12 then, applying this inductively, |fck (z)| ≤ 21 for all k ∈ Z+ . In
particular, |fck (z)| 6→ ∞, so z ∈ K(fc ), the filled in Julia set. Thus B(0, 12 ) ⊂ K(fc ).
On the other hand, if |z| ≥ 2, then
1 3
|fc (z)| = |z 2 + c| ≥ |z|2 − |c| ≥ 2|z| − |c| ≥ 2|z| − |z| = |z|,
2 2
since |z| ≥ 2 ≥ 2|c|. Applying this repeatedly, if |z| ≥ 2 then |fck (z)| ≥ 23 |fck−1 (z)| ≥
. . . ≥ ( 23 )k |z| ≥ 2( 32 )k → ∞. We conclude that if |z| ≥ 2 then z ∈ / K(fc ), so
K(fc ) ⊂ B(0, 2).

1
 From B(0, 12 ) ⊂ K(fc ) ⊂ B(0, 2) we conclude that the Julia set, which is the bound-
ary of K(fc ), lies in the annulus B(0, 2) \ B O (0, 21 ).

14.5 The fixed points of f are given by f (z) = z 2 − 2 = z, so are z = −1, 2. Since
f 0 (z) = 2z we have that |f 0 (2)| = 4 > 1, so 2 is a repelling fixed point.
By
S∞ Theorem 14.10, 2 ∈ J(f ), so by Corollary 14.8(b), J(f ) is the closure of
−k
k=1 f (2). If w ∈ [−2, 2] then f −1 (w) = (w + 2)1/2 ∈ [−2, 2], so f −k (w) ∈ [−2, 2],
and in particular f −k (2) ∈ [−2, 2], for all k ∈ Z+ . Thus J(f ) is contained in the
closure of [−2, 2], that is J(f ) ⊂ [−2, 2].
Now observe tha f (0) = −2 and f 2 (0) = f (−2) = 2. Since 2 is a fixed point of
f , it follows that f k (0) = 2 for k = 2, 3, . . ., so f k (0) 6→ ∞. Thus −2 ∈ M by
Theorem 14.14, so the Julia set J(f ) is connected. But 2 ∈ J(f ) and −2 ∈ J(f )
(since f −1 (2) = {2, −2}), and the Julia set J(f ) is a connected subset of the interval
[−2, 2] containing its endpoints −2 and 2, so this requires J(f ) = [−2, 2].

14.6 For z ∈ C we have fc (−z) = fc (z), so fck (−z) = fck (z) for all k ∈ Z+ . Thus
fck (−z) → ∞ if and only if fck (z) → ∞, so that −z ∈ K(fc ) if and only if z ∈ K(fc ).
Thus the filled-in Julia set K(fc ) is symmetric about the origin, and its boundary,
the Julia set J(fc ), is also symmetric about the origin.

14.7 Let c be real with c > 41 . Then fck (z) is real for all k ∈ Z+ and real z. In particular,
for real z,
1 1 1
fc (z) − z = z 2 + c − z = (z − )2 + (c − ) ≥ (c − ).
2 4 4
Applying this repeatedly,
1
fck (z) ≥ z + k(c − ) → ∞.
4
We conclude that z ∈
/ J(fc ) if z is real.
The Julia set J(fc ) is non-empty (by Proposition 14.2) and symmetric about the
origin (by Exercise 14.6), so there exists z ∈ C with z, −z ∈ J(fc ). Since J(fc )
contains no points on the real axis and z and −z lie on opposite sides of the real
axis, J(fc ) is not connected. Thus by (14.4) c ∈
/ M.

14.8 The fixed points of f = fc are given by z 2 + c = z, that is z 2 − z + c = 0 or

z = 21 (1±(1−4c)1/2 ). The sum of these two distinct roots is 1, so we may choose one
of the fixed points w = 21 (1+(1−4c)1/2 ), say, with |w| > 21 . Thus |f 0 (w)| = 2|w| > 1,
so w is a repelling fixed point. The number w is real if and only if (1 − 4c)1/2 is real,
which is not that case if c is non-real. Hence f 0 (w) = 2w is not real.
We know from Theorem 14.16 that if |c| < 14 then the Julia set J = J(f ) is a simple
closed curve. Suppose that this curve J has a tangent at w. Since f (w) = w and
f (J) = J, and f is analytic near w, f maps a neighbourhood of J containing w to a

2
 neighbourhood of J containing w. To a first order approximation (considering the
Taylor series expansion of f around w), we have

f (w + z) = f (w) + f 0 (w)z + O(z 2 ) = w + f 0 (w)z + O(z 2 ).

In particular, f maps the tangent to J at w, which may be written parametricly as

{w + tz0 : t ∈ R} near w, onto a smooth curve that is tangential to J at w of the
form {w + tf 0 (w)z0 + O(t2 ) : t ∈ R}. This is only possible if this curve is tangential
to the original tangent at w, that is if f 0 (w) is real.
1
We conclude from the first part that if c is a non-real number with |c| < 4
then
f 0 (w) is not real, so J does not have a tangent at w.
If J contains an arc A, then we may find n ∈ Z+ such that f n (w) is a interior point
of the arc A, by Corollary 14.8(a). But f n is a smooth locally bijective mapping
that maps a non-differentiable arc of J containing w into A, so that A cannot be
differentiable.

1
14.9 If |c| ≤ 4
and |z| ≤ 12 , then using the triangle inequality
 2
2 2 1 1 1
|fc (z)| = |z + c| ≤ |z| + |c| ≤ + = .
2 4 2

If |c| ≤ 14 , applying this inductively gives |fck (0)| ≤ 1

2
for all k ∈ Z+ . Thus
fck (0) 6→ ∞, so c ∈ M . Thus B(0, 41 ) ⊂ M .

1 21 1
14.10 If |c + 1| ≤ 20
, we note that |c| ≤ 20
. Thus if |z| ≤ 10
, then using the triangle
inequality

|fc (fc (z))| = |(z 2 + c)2 + c| = |z 4 + 2cz 2 + c(c + 1)| ≤ |z|4 + 2|c||z|2 + |c||c + 1|
 4  2
1 21 1 21 1 1
≤ +2 + < .
10 20 10 20 20 10
1 1
Thus if |c + 1| ≤ 20 then, applying this inductively, |fc2k (0)| ≤ 10
for all k ∈ Z+ .
1
Thus fck (0) 6→ ∞, so c ∈ M . Thus B(−1, 20 ) ⊂ M.

14.11 We have
 
2 2 |c|
|fc (z)| = |z + c| ≥ |z| − |c| = |z| |z| − ≥ |z|(2 +  − 1) ≥ |z|(1 + ),
|z|

provided that |z| ≥ max{2 + , |c|}.

If |c| > 2 we may choose  > 0 such that |c| > 2 + , so noting that fc (0) = c and
applying the above estimate inductively,

|fck (0)| ≥ (1 + )k → ∞.

Thus c ∈
/ M by Theorem 14.14.

3
14.12 If c is such that |fck (0)| > 2 for some k, then either
(i) |c| > 2 so c ∈
/ M by Exercise 14.11, or
(ii) |c| ≤ 2, and we may choose  > 0 such that |fck (0)| > 2 +  > |c|. Then applying
the first part of Exercise 14.11 repeatedly to z = fck (0) and its iterates under fc ,
gives |fck+n (0)| = |fcn (fck (0))| ≥ (1 + )n |fck (0)| → ∞, so c ∈
/M .
/ M then |fck (0)| → ∞, so |fck (0)| > 2 for some k.
On the other hand, if c ∈

√
14.13 Let f (z) = z 3 +√cz, so f 0 (z) = 3z 2 + c. The fixed points of f are 0, ± 1 − c, with
f 0 (0) = 0, f 0 (± 1 − c) = 3 − 2c. Hence, provided |c| < 1, 0 is an attractive fixed
point of f . But the Julia set of a polynomial f is a closed curve precisely when f
has an attractive fixed point, see note before Theorem 14.16 (a proof along the lines
of Theorem 14.16 works when this is the case).

√
14.14 The solution is similar to the proof of Theorem 14.15. Let |c| > 2. Let C be the
circle |z| = |c| and let D be its interior |z| < |c|. If we define S1 , S2 , S3 : D → D to
be the branches of f −1 , then S1 (D), S2 (D) and S3 (D) are the interiors of the three
loops of the curve f −1 (C).
We now let V√be the disc {z : |z| < |2c|1/3 } so that V just contains f −1 (D). Note
that, if |c| > 2, then V ⊂ D and so the sets S1 (V ), S2 (V ), S3 (V ) are all in V and
the sets S1 (V ), S2 (V ), S3 (V ) are disjoint. For i = 1, 2, 3, we have
1
|Si0 (z)| = |z − c|−2/3
3
and so, if z ∈ V ,
1 1
(|c| + |2c|1/3 )−2/3 ≤ |Si0 (z)| ≤ (|c| − |2c|1/3 )−2/3 .
3 3
It now follows from a complex mean-value theorem that
1 |Si (z1 ) − Si (z2 )| 1
(|c| + |2c|1/3 )−2/3 ≤ ≤ (|c| − |2c|1/3 )−2/3 ,
3 |z1 − z2 | 3

for i = 1, 2, 3 and z1 , z2 ∈ V .
If |c| > 2 (for example), then the upper bound is less than 1 and so S1 , S2 and S3
are contractions on the disc V . It follows from Theorem 9.1 that there is a unique
non-empty compact invariant set F ⊂ V satisfying

S1 (F ) ∪ S2 (F ) ∪ S3 (F ) = F.

It follows from Propositions 9.6 and 9.7 that lower and upper bounds for dimH F
are provided by the solutions of
1
3( (|c| ± |2c|1/3 )−2/3 )s = 1,
3
that is, by
log 3
s= .
log 3 + (2/3) log(|c| ± |2c|1/3 )

4
 So, when |c| is large,
3 log 3
dimH F ∼ .
2 log |c|
√
It remains to show that F = J(f ). If |c| > 2 and |z| ∈/ V , i.e. |z| > |2c|1/3 then
1 1 1 1
|f (z)| = |z 3 + c| ≥ |z|3 − |c| ≥ |z|3 − |z|3 = |z|3 = |z| |z|2 ≥ |z| |2c|2/3 = λ|z|
2 2 2 2
where λ = 12 |2c|2/3 > 21 |2|2/3 |2|1/3 = 1. Iterating, it follows that |f k (z)| ≥ λk |z| →
∞, so z ∈ / J(f ). Thus J(f ) ⊂ K(f ) ⊂ V . It follows from Propositions 14.2 and
14.3 that J(f ) is the non-empty compact subset of V satisfying J(f ) = f −1 (J(f )),
that is, J(f ) = S1 (J(f )) ∪ S2 (J(f )) ∪ S3 (J(f )) and so J(f ) = F as claimed.

14.15 If fc (z) = z 2 + c, then the fixed points of fc are the solutions of z 2 + c = z. Since
fc0 (z) = 2z, we see that z = reiθ is an attractive fixed point of fc if and only if
0 ≤ r < 1/2, 0 ≤ θ < 2π and

c = z − z 2 = z(1 − z) = reiθ (1 − reiθ ).

Thus fc has an attractive fixed point precisely when c lies inside the main cardioid
of the Mandelbrot set.

14.16 Since fc (z) = z 2 + c and fc2 (z) = fc (fc (z)), we have

fc2 (z) − z = (z 2 + c)2 + c − z

= z 4 + 2cz 2 − z + c2 + c
= (z 2 − z + c)(z 2 + z + c + 1).

Now z is a periodic point of fc of period 2 if and only if fc2 (z) = z and fc (z) 6= z.
Thus we are looking for the solutions of fc2 (z) − z = 0 which are not solutions of
fc (z) − z = z 2 − z + c = 0. In other words, we are interested in the solutions of
z 2 + z + c + 1 = 0. By the chain rule

(fc2 )0 (z) = fc0 (fc (z))fc0 (z) = 2(z 2 + c)2z

and so z is an attractive fixed point of fc2 if and only if

|(z 2 + c)z| < 1/4 and z 2 + z + c + 1 = 0.

Using the second of these conditions to rewrite the first, we find that fc2 has an
attractive fixed point if and only if

|(z 2 + c)z| = |(z + 1)z| = |z 2 + z| = |c + 1| < 1/4.

Thus fc2 has an attractive fixed point precisely when c belongs to the main bud of
the Mandelbrot set; this is the region labelled 2 in Figure 14.8.

5
14.17 Assume, for a contradiction, that c is not in the basin of attraction A(w) of the
(finite) attractive fixed point w of fc . Let U be an open disc with w ∈ U ⊂ A(w).
Then fck (c) ∈/ U for all k = 0, 1, 2, . . .. Thus for each k we may select a branch of
the inverse fc−k on U to be a continuous analytic function with fc−k (w) = w. If
z ∈ fc−k (U ) then fck (z) ∈ U ⊂ A(w), so z ∈ A(w); thus fc−k (U ) ⊂ A(w) for all
k. Since A(w) is a bounded subset of C, Montel’s theorem implies that {fc−k }∞ k=0
is a normal family on U . However, since w is a repelling fixed point of fc−1 , no
subsequence of fc−k (z) can be uniformly convergent to an analytic function near w
(since (fc−k )0 (w) = ((fc−1 )0 (w))k → ∞ by the chain rule), so {fc−k }∞k=0 cannot be
normal by the definition of a normal family. We conclude that c must be in the
basin of attraction A(w) of w.
Since c cannot be in the basin of attraction of more than one point, it follows that
fc has at most one (finite) attractive fixed point.
Now let f be any polynomial on C and let A(w) be the basin of attraction of some
(finite) attractive fixed point w of f . Assume, for a contradiction, that c ∈ / A(w)
0
for all c ∈ C such that f (c) = 0. Let U be an open disc with w ∈ U ⊂ A(w).
Then f k (c) ∈/ U for all k = 0, 1, 2, . . . and all c such that f 0 (c) = 0. This enables
us to choose, for each k, a branch of the inverse f −k on U that is a continuous
analytic function with f −k (w) = w. If z ∈ f −k (U ) then f k (z) ∈ U ⊂ A(w), so
z ∈ A(w); thus f −k (U ) ⊂ A(w) for all k. Since A(w) is a bounded subset of C,
Montel’s theorem implies that {f −k }∞ k=0 is a normal family on U . However, since w
is a repelling fixed point of f , no subsequence of f −k (z) can have a subsequence
−1

uniformly convergent to an analytic function near w , so {f −k }∞ k=0 cannot be normal

by the definition of a normal family. We conclude that c ∈ A(w) for some c with
f 0 (c) = 0.

14.18 Let f (z) = az 2 + bz + d with a 6= 0. Then f p is a polynomial of order 2p . Suppose

that f p has an attractive fixed point w. By Exercise 14.17, the basin of attraction
A(w) of w under iteration of f p contains a point c such that

0 = (f p )0 (c) = f 0 (c)f 0 (f (c)) . . . f 0 (f p−1 (c)) = (2ac + b)(2af (c) + b) . . . (2af p−1 (c) + b)

using the chain rule. It follows that for some 0 ≤ r ≤ p − 1 we have 0 = 2af r (c) + b,
that is f r (c) = −b/2a for some c in A(w).
If f has an attractive periodic orbit of order p then f p has some attractive fixed
point w, so −b/2a is attracted to this periodic orbit under iteration of f . We con-
clude that there can beat most one attractive periodic orbit.

6
 Chapter 15

15.1 This is a particular case of the random Cantor set described in Section 15.1. In this
case
C1 = C2 = 13 with probability 12
.
C1 = C2 = 16 with probability 12
By Theorem 15.1, the Hausdorff dimension is given by the solution s of the expec-
tation equation
1 1
1 = E(C1s + C2s ) = (3−s + 3−s ) + (6−s + 6−s ) = 3−s + 6−s ,
2 2
giving s = 0.4895 . . . .

15.2 This is a particular case of the random construction considered in Theorem 15.2.
Since the number of segments at least doubles at each step, the probability of
extinction occurring in the construction is 0. Writing C1 , C2 , C3 , C4 for the length
ratios of each of the four subsegments at each step, we have

C1 = C2 = C3 = C4 = 31 with probability 1
2 .
C1 = C4 = 13 and C2 = C3 = 0 with probability 1
2

By Theorem 15.2, the Hausdorff dimension is given by the solution s of the expec-
tation equation
4
X 1 1
1 = E( Cis ) = (4 × 3−s ) + (2 × 3−s ) = 3 × 3−s .
i=1
2 2

Thus s = 1.

√
15.3 Let E0 √be the (closed) parallelogram with vertices (0, 0), (1/2, 3/6), (1, 0) and
(1/2, − 3/6), so that E0 has diameter 1. Let F be any ‘random’ von Koch curve
constructed by substituting an upwards or downwards figure at each stage. Let
E1 be the set consisting of 4 similar parallelograms of diameter 3−1 with axes
on the 4 segments of the first stage of the construction of F ; all these parallel-
ograms are contained in E0 . Let E2 be the set consisting of similar parallelo-
grams of diameter 3−2 with axes on the 42 segments of the second stage of the
construction of F , each contained in a parallelogram of E1 , and so on. Thus Ek
consists of 4k parallelograms of diameters 3−k withT∞disjoint interiors, each con-
tained in a parallelogram of Ek−1 . Then F = k=0 Ek . By Proposition 4.1,
dimH F ≤ dimB F ≤ limk→∞ log 4k / − log 3−k = log 4/ log 3.
For the lower bound, let µ be the mass distribution obtained by repeated subdivision
so that µ(P ) = 4−k for each parallelogram P of Ek . Then if |U | < 1, and k is the
integer such that 2−1 3−k−1 ≤ |U | < 2−1 3−k , it is easy to see from the geometry of
the parallelograms that U intersects at most 6 parallelograms of Ek . Thus

µ(U ) ≤ 6 × 4−k = 6 × 3−k log 4/ log 3 ≤ 6 × (6|U |)log 4/ log 3 ,

so it follows from the Mass distribution principle 4.2 that dimH F ≥ log 4/ log 3.

7
15.4 This fits into the context of Theorem 15.2. The expected number of sub-triangles
at each stage of the construction is 3p, so extinction of the construction, leading to
the empty set, will occur if and only if 3p ≤ 1, that is if and only if p ≤ 1/3 (note:
not p ≤ 2/3).
 
3
If p > 1/3, we have P(N = j) = pj (1 − p)3−j , and by the binomial theorem,
j
equation (15.8) becomes (tp + (1 − p))3 = t. By Theorem 15.2 the smallest positive
solution t of this cubic equation gives the probability that F is empty.
To find the dimension of F when it is non-empty, write C1 , C2 , C3 for the ratios of
the three similarities, so that Ci = 21 with probability p and Ci = 0 with probability
1 − p, independently for i = 1, 2, 3. Thus by (15.9), the Hausdorff dimension is given
by the solution s of the expectation equation
X3
1 = E( Cis ) = 3(p2−s + (1 − p)0) = 3p2−s .
i=1

Taking logarithms gives dimH F = s = log 3p/ log 2.

15.5 Let v be a vertex of a triangle T of the kth stage of the construction of the stan-
dard Sierpinski triangle. We claim that in the random Serpinski triangle F , with
probability 1 the set F ∩ T does not contain some neighbourhood of v. To see
this, let T ⊃ Tk+1 ⊃ Tk+2 ⊃ . . . be the nested triangles at the subsequent stages of
the standard Sierpinski triangle construction that contain the vertex v. There is a
probability p that each of these triangles is retained in the random construction, so
the probability that Tk+1 , . . . , Tk+n are all retained is pn . Using continuity of proba-
bilities, since pn → 0 as n → ∞, the probability that all the triangles Tk+1 , Tk+2 , . . .
are retained is 0. Thus with probability 1, one of the triangles Tk+1 , Tk+2 , . . . is
removed, so (F ∩ T ) ⊂ (T \ Tk+n ) for some n.
Thus, if x, y ∈ F , let T be a triangle of the standard Sierpinski triangle construction
with x ∈ T and y ∈ / T . By the above, there is, with probability 1, a neighbourhood
of each of the three vertices of T that is not in F , and these neighbourhoods dis-
connect x from the parts of F outside T , so x and y are in different components of F .

15.6 Since the set F may be covered with either a single unit interval, or the aggregate
of coverings of F ∩ I1 and F ∩ I2 , we have that
s s s
H∞ (F ) ≤ min{1, H∞ (F ∩ I1 ) + H∞ (F ∩ I2 )}
s s s s
= min{1, C1 H∞ (F1 ) + C2 H∞ (F2 )},
where F1 and F2 are independent realisations of F and independent of {C1 , C2 }.
Then
s
E(H∞ (F )) = E(min{1, C1s H∞
s
(F1 ) + C2s H∞
s
(F2 )})
≤ E(C1 H∞ (F ) + C2 H∞ (F )) ≤ E(C1s + C2s )E(H∞
s s s s s s
(F )) = E(H∞ (F )).
s s
Thus equality holds, so since the terms are finite, either E(H∞ (F )) = 0, or H∞ (F ) =
s s s s
C1 H∞ (F1 ) + C2 H∞ (F2 ) almost surely. In the latter case
s
esssupH∞ (F ) = esssup(C1s + C2s )esssupH∞
s
(F ).

8
 Hence either esssup(C1s + C2s ) = 1, which would imply that C1s + C2s = 1 almost
surely, since E(C1s + C2s ) = 1, or else H∞ s
(F ) = 0 almost surely. If H∞ s
(F ) = 0
k s
then given δ > 0, by scaling, if |Ii1 ,...,ik | ≤ b ≤ δ then Hδ (F ∩ Ii1 ,...,ik ) = 0, so by
taking unions of such basic intervals Hδs (F ) = 0. Letting δ → 0 we conclude that
Hs (F ) = 0.

15.7 Writing CI,j = |Ii1 ,...,ik , j|/|Ii1 ,...,ik | where I = Ii1 ,...,ik , we have that
 2 
X
2
|I|s  |Fk 


E Xk+1 |Fk = E 
I∈Ek+1
 !2 
X
= E |I|s (CI,1
s s
+ CI,2 ) |Fk 
I∈Ek

X
= E |I|s |J|s (CI,1
s s
+ CI,2 s
)(CJ,1 s
+ CJ,2 )
I6=J∈Ek
! !
X
+ |I|2s (CI,1
s s 2
+ CI,2 ) |Fk
I∈Ek
X X
s s
|I|2s E (C1s + C2s )2


= |I| |J| +
I6=J∈Ek I∈Ek
X X X
≤ ( |I|s )2 + a |I|2s = Xk2 + a |I|2s ,
I∈Ek I∈Ek I∈Ek

where a = E((C1s + C2s )2 ) and we have used (15.2). Taking unconditional expecta-
tions, we get
X
2
E(Xk+1 ) ≤ E(Xk2 ) + aE( |I|2s ) = E(Xk2 ) + aE((C12s + C22s )k ) = E(Xk2 ) + aγ k ,
I∈Ek

where γ = E((C12s + C22s )k ) < 1, using (15.3) repeatedly. Applying this (k − 1) times,
E(Xk2 ) ≤ E(X12 ) + a(γ + γ 2 + . . . + γ k−1 ) ≤ E(X12 ) + aγ/(1 − γ)
for all k. Thus Xk2 is an L2 bounded martingale.

15.8 Fix p > p0 , and let r > 0 be the probability that the random fractal Fp constructed
by the percolation construction is non-empty and therefore, with the same probabil-
ity, contains a non-trivial connected component. Let I be a square that is retained
at some kth level of the 3 × 3 fractal percolation construction. There is a probabil-
ity p(1 − p)8 that at the (k + 1)th stage of the construction the middle (k + 1)th
level subsquare of I is selected and the other 8 sub-squares are removed, and by
self-similarity there is a probability of r that this middle square intersects Fp in a
non-trivial connected component. Thus there is a probability at least s ≡ rp(1 − p)8
that such a square I contains a non-trivial connected component of Fp that does
not extend outside I.
Given  > 0 and a probability s, the laws of large numbers imply that there is an
integer N0 such that, if N ≥ N0 , in N independent trials each with probability s of

9
success, there is a probability of at least 1 −  that at least N s/2 of the trials will be
successful. Thus, suppose that, for some k there are at least N squares in Ek . There
is a probability of at least s that, independently, each of these squares contains a
distinct non-trivial connected component, so there is a probability of at least 1 − 
that at least N s/2 of these squares contain a non-trivial connected component.
Finally, with r the probability of Fp being non-empty, given  > 0 and N , there
is an integer k such that with probability at least r −  the kth level stage of the
construction, Ek , contains at least N squares. (This follows from extinction prop-
erties of branching processes.) Thus, given  > 0 there is a probability of at least
(r −)(1−) > r −2 that Fp contains at least N s/2 distinct connected components.
Since this is true for all  > 0 and N ≥ N0 , the random set Fp contains infinitely
many distinct connected components with probability r.

10
 Chapter 16

16.1 Let Nδ (X[0, 1]) denote the (random) least number of sets of diameter δ that can
cover the Brownian trail X[0, 1]. Then for each n, since the Brownian trails
X[(i−1)2−n , i2−n ] for i = 1, 2, . . . , 2n have the same distribution as X[0, 1] but under
a similarity scaling by a factor 2−n/2 , we conclude that N2−n/2 (X[(i−1)2−n , i2−n ]) has
the same statistical distribution as N1 (X[0, 1]). In particular, taking expectations,

E(N2−n/2 (X[(i − 1)2−n , i2−n ])) = E(N1 (X[0, 1])),

so taking the aggregate of such coverings,

E(N2−n/2 (X[0, 1])) ≤ 2n E(N1 (X[0, 1])).

Thus, for all  > 0, we have E(2−n(1+) N2−n/2 (X[0, 1])) ≤ 2−n E(N1 (X[0, 1])). Sum-
ming,
X∞ ∞
X
E( 2−n(1+) N2−n/2 (X[0, 1])) ≤ 2−n E(N1 (X[0, 1])) < ∞.
n=1 n=1
P∞ −n(1+)
Thus, with probability 1, n=1 2 N2−n/2 (X[0, 1]) < ∞, implying that for some
random number C we have N2−n/2 (X[0, 1]) ≤ C2n(1+) for all n, and so by Proposi-
tion 4.1 that dimB X[0, 1] ≤ 2(1+) for all  > 0. We conclude that dimB X[0, 1] ≤ 2.

16.2 This is a variation on the proof of Theorem 16.5. Consider Brownian motion
X : [0, 1] → R3 . Let 0 < λ < 1/2. By an obvious modification of Proposition 16.1,
there is with probability 1 a (random) number B such that

|X(t) − X(u)| ≤ B|t − u|λ (t, u ∈ [0, 1]),

so by Proposition 2.3 dimH X(F ) ≤ (1/λ)dimH F = log 2/λ log 3. This is true for all
0 < λ < 1/2, so dimH X(F ) ≤ 2 log 2/ log 3 = log 4/ log 3.
For the lower bound, we define a measure by transfering Hausdorff measure on the
Cantor set F to the trail. With q = log 2/ log 3, define a random measure µ on
q
R = H {t : t ∈RF and X(t) ∈
X(F ) by µ(A) A} for A ⊂ R3 . Thus for a function g on
3 q
R we have g(x)dµ(x) = F g(X(t))dH (t). Then for s < 2 log 2/ log 3
Z Z  Z Z 
−s −s
E |x − y| dµ(x)dµ(y) = E |X(t) − X(u)| dtdu
F F
Z Z
= E(|X(t) − X(u)|−s )dtdu
FZ FZ

= c1 |t − u|−s/2 dtdu < ∞

F F

using Exercise 4.11, with c1 as in (16.11). It follows from Theorem 4.13(a) that
dimH X(F ) ≥ s for all s < 2 log 2/ log 3, so dimH X(F ) ≥ 2 log 2/ log 3 = log 4/ log 3.
If X is index-α fractional Brownian motion, we get, in a similar way, that dimH X(F )
= log 2/α log 3, using that, with probability 1, X satisfies a Hölder condition of in-
dex λ for all 0 < λ < α (Proposition 16.7). For the lower bound we replace (16.11)
by E(|X(t + h) − X(t)|−s ) = c1 h−sα .

11
16.3 The approach here is similar to that in Theorem 16.6. Suppose, for a contradic-
tion, that X[0, 1] ∩ F = ∅ with probability 1. Using the isotropy and scaling of
Brownian trails in R3 , it follows that for every similarity σ, the probability that
σ(X[0, 1]) ∩ F = ∅ is also 1. By Fubini’s theorem, with probability 1 we have
that σ(X[0, 1]) ∩ F = ∅ for almost all similarities σ. But with, probability 1,
dimH X[0, 1] = 2, so since dimH X[0, 1] + dimH F − 3 > 2 + 1 − 3 = 0, this contradicts
Theorem 8.3(a). We conclude that X[0, 1] intersects F with positive probability.

2 2
 f : R → R given
16.4 The esiest way to see this is to note that the affine transformation

1 0
by f (t, u) = (t, u+ct) is bi-Lipschitz (since the matrix is invertible). Then
c 1
f (graphX(t)) = graph(X(t) + ct), so dimH (graph(X(t) + ct)) = dimH (graphX(t)) =
1 21 almost surely, using Proposition 3.3(b) and Theorem 16.2.

16.5 Let t = t0 < t1 < . . . < tn = u. Then X(ti−1 ) ≤ X(ti ) with probability 12 ,
independently for i = 1, 2, . . . , n. Hence the probability that X(ti−1 ) ≤ X(ti ) for
all i = 1, 2, . . . , n is 2−n and similarly the probability that X(ti−1 ) ≥ X(ti ) for all
i = 1, 2, . . . , n is 2−n . If X(t) is monotonic on [t, u] then one of these possibilities
must occur, so P(X(t) is monotonic on [t, u]) ≤ 2 × 2−n . This is true for all positive
integers n, so P(X(t) is monotonic on [t, u]) = 0.
Since there are countably many rational numbers, there are countably many ‘ra-
tional intervals’, i.e. intervals with rational endpoints. Since a countable union of
events each of probability 0 has probability 0, we conclude that P(X(t) is monotonic
on every rational interval [t, u]) = 0. Since every non-degenerate interval contains
a rational interval, this implies that P(X(t) is monotonic on every interval [t, u]) = 0.

16.6 First we show that there is a number γ > 0 such that, given X[0, 1], the probability
that X(t) = 0 for some t > 1 is at least γ, i.e.
P(X(t) = 0 for some t > 1|X(t) (0 ≤ t ≤ 1)) ≥ γ. (1)
To see this, suppose that X(1) = −M < 0. Then, conditional on this, using (16.2),
P(X(t) = 0 for some t > 1) ≥ P(X(1 + M 2 ) − X(1) > M )
Z ∞
−x2
 
1
= √ exp dx
M 2π M 2M 2
Z 2M
1
≥ √ exp(−2)dx
M 2π M
exp(−2)
= √ ≡ γ,
2π
with similar estimates if X(1) = M ≥ 0. In particular, it follows from (1) that
P(X(t) = 0 for some t > 1|X(t) 6= 0 (0 < t ≤ 1)) ≥ γ.

Now set p = P(X(t) = 0 for some 0 < t ≤ 1). By statistical self-similarity of

Brownian motion, for every N > 0,
p = P(X(t) = 0 for some 0 < t ≤ N }, (2)

12
 so taking a union over all positive integers N , and noting that these are decreasing
events (i.e. if N1 < N2 and X(t) 6= 0 for 0 < t ≤ N2 then X(t) 6= 0 for 0 < t ≤ N1 ),
we conclude that p = P(X(t) = 0 for some t > 0). But
p = P(X(t) = 0 for some t > 0)
≥ P(X(t) = 0 for some 0 < t ≤ 1)
+ P(X(t) 6= 0 for 0 < t ≤ 1 and X(t) = 0 for some t > 1)
= p + P(X(t) = 0 for some t > 1|X(t) 6= 0 (0 < t ≤ 1))P(X(t) 6= 0 (0 < t ≤ 1))
≥ p + γ(1 − p)
using the definition of conditional probability (1.15). Thus p ≥ p + γ(1 − p) which
requires 1 − p = 0, that is p = 1. Thus with probability 1, X(t) = 0 for some t > 0.
By (2), with probability 1, X(t) = 0 for some 0 < t ≤ 1/n, for all n = 1, 2, . . .. This
can only happen if X(t) = 0 infinitely often in every interval (0, a) with a > 0.

16.7 Let X be Brownian motion. As in (16.10),

r
−x2
Z  
1
p(r) ≡ P(0 ≤ X(t + h) − X(t) ≤ r) = √ exp dx.
2πh 0 2h
Thus
Z ∞
q
E(|X(t + h) − X(t)| ) = 2 rq dp(r)
0
Z ∞  2
2 q −r
= √ r exp dr
2πh 0 2h
2 q/2 ∞ q
 2
−u
Z
= √ h u exp du,
2π 0 2
√
on substituting u = r/ h, as required.

16.8 Let λ > α. Suppose that, for a given t and b, there almost surely exists H0 such
that
|X(t + h) − X(t)| ≤ b|h|λ for all |h| ≤ H0 . (1)
Then, by Egoroff’s theorem, there exists h0 > 0 such that with probability at least
1/2 we have |X(t + h) − X(t)| ≤ b|h|λ for all |h| ≤ h0 .
On the other hand, from (16.13), we have
Z b|h|λ
λ −1/2 −α
P(|X(t + h) − X(t)| ≤ b|h| ) = 2(2π) |h| exp(−u2 /2|h|2α )du
0
Z b|h|λ
−1/2 −α
≤ 2(2π) |h| 1du
0
= 2(2π)−1/2 b|h| λ−α
,
so by taking h sufficiently small we get a contradiction.
We conclude that for all t, with probability 1 there is no number H0 such that (1)
holds. Thus, by Fubini’s theorem, with probability 1, there is, for almost all t, no
H0 such that (1) holds.

13
16.9 Write X(t) = (X1 (t), X2 (t)), where X1 and X2 are independent index-α1 and index-
α2 fractional Browninan motions, with 1/2 ≤ α1 ≤ α2 < 1. To get an upper bound
for the dimension of X[0, 1] we use the Hölder estimate of Proposition 16.7. Thus,
given  > 0, there are, with probability 1, (random) constants 0 < B1 , B2 < ∞ such
that

|X1 (t) − X1 (u)| ≤ B1 |t − u|α1 − and |X2 (t) − X2 (u)| ≤ B2 |t − u|α2 − (t, u ∈ [0, 1]).

Thus, for k = 1, 2, . . ., if |t − u| ≤ 2−k then X[t, u] is contained in a rectangle with

sides B1 2−k(α1 −) and B2 2−k(α2 −) . Dividing this rectangle into approximate squares
of sides at most B2 2−k(α2 −) , the rectangle, and thus X[t, u], may be covered by
2B1 2−k(α1 −) /B2 2−k(α2 −) = (2B1 /B2 )2k(α2 −α1 ) sets of diameter at most 2B2 2−k(α2 −) ,
provided k is sufficiently large. Thus, dividing the interval [0, 1] into 2k subinter-
vals of length 2−k , the trail X[0, 1] may be covered by (2B1 /B2 )2k(1+α2 −α1 ) sets of
diameter at most 2B2 2−k(α2 −) . It follows by Proposition 4.1 that

log((2B1 /B2 )2k(1+α2 −α1 ) ) 1 + α2 − α1

dimH X[0, 1] ≤ dimB X[0, 1] ≤ lim −k(α −)
= ,
k→∞ − log(2B2 2 2 ) α2 − 

for all  > 0, so dimH X[0, 1] ≤ (1 + α2 − α1 )/α2 .

For the lower bound we use the potential theoretic method. We need to estimate
the integral E((|X1 (t + h) − X1 (t)|2 + |X2 (t + h) − X2 (t)|2 )−s/2 ), so we first consider
the X1 part. Write
Z r
1
p(r) = P(0 ≤ X1 (t + h) − X1 (t) ≤ r) = α1 exp(−x2 /2h2α1 )dx, (1)
h (2π)1/2 0

by (16.13). For fixed y,

E((|X1 (t + h) − X1 (t)|2 + y 2 )−s/2 )

Z ∞
= 2 (r2 + y 2 )−s/2 dp(r)
0
Z ∞
−α1
= ch (r2 + y 2 )−s/2 exp(−r2 /2h2α1 )dr
Z ∞ 0
= c (u2 h2α1 + y 2 )−s/2 exp(−u2 /2)du
0
Z y/hα1 Z ∞
−s
≤ c 2
y exp(−u /2)du + c (uhα1 )−s exp(−u2 /2)du
0 y/hα1
Z y/hα1 Z ∞
≤ c y −s du + c (uhα1 )−s du
0 y/hα1
1−s −α1
≤ c1 y h

where c, c1 do not depend on h, and we have substituted u = r/hα1 .

Using the analogue of (1) for indx-α2 fractional Brownian motion, we get in a similar

14
 way,

E((|X1 (t + h) − X1 (t)|2 + |X2 (t + h) − X2 (t)|2 )−s/2 )

= E((c1 |X2 (t + h) − X2 (t)|1−s h−α1 )
Z ∞
−α2
= c2 h y 1−s h−α1 exp(−y 2 /2h2α2 )dy
0
Z ∞
−α1 +(1−s)α2
= c2 h u1−s exp(−u2 /2)du
0
= c3 h−α1 +(1−s)α2

where we have substituted u = y/hα2 .

Define a random measure
2
R R 1 1] by µ(A) = L{t : 0 ≤ t ≤ 1 and X(t) ∈ A}, so
on X[0,
that for f : R → R, f dµ = 0 f (X(t))dt. Then
Z Z 
−s
E |x − y| dµ(x)dµ(y)
Z 1 Z 1 
2 2 −s/2
= E (|X1 (t) − X1 (u)| + |X2 (t) − X2 (u)| ) )dt du
0 0
Z 1Z 1
= E((|X1 (t) − X1 (u)|2 + |X2 (t) − X2 (u)|2 )−s/2 ))dt du
Z0 1 Z0 1
≤ c3 |t − u|−α1 +(1−s)α2 dt du.
0 0

This is finite if −α1 + (1 − s)α2 > −1, that is if s < (α2 − α1 + 1)/α2 . Thus if
s < (α2 − α1 + 1)/α2 then almost surely the trail X[0, 1] supports a measure µ with
finite s-energy, so by Proposition 4.13(a), dimH X[0, 1] ≥ (α2 − α1 + 1)/α2 almost
surely.

16.10 We take as our starting point that for index-α fractional Brownian motion E(X(t)2 ) =
|t|2α and E((X(t) − X(u))2 ) = |t − u|2α . Expanding the latter,

|t − u|2α = E(X(t)2 ) + E(X(u)2 ) − 2E(X(t)X(u)) = |t|2α + |u|2α − 2E(X(t)X(u))

so
2E(X(t)X(u)) = |t|2α + |u|2α − |t − u|2α .
Thus

2E((X(t) − X(0))(X(t + h) − X(t)))

= 2(E(X(t)X(t + h)) − E(X(0)X(t + h)) − E(X(t)X(t)) + E(X(0)X(t)))
= (|t|2α + |t + h|2α − |h|2α ) − 0 − (2|t|2α ) + 0
= |t + h|2α − |t|2α − |h|2α ,

giving
1
E((X(t) − X(0))(X(t + h) − X(t))) = (|t + h|2α − |t|2α − |h|2α ).
2

15
For t, h 6= 0, we have, by elementary calculus, that if 12 < α < 1 then |t + h|2α >
|t|2α +|h|2α , so E((X(t)−X(0))(X(t+h)−X(t))) > 0, and the increments X(t)−X(0)
and X(t + h) − X(t) are positively correlated. Thus if the sample path has increased
after a certain time, there is a tendency for it to continue to increase, and if it has
decreased there is a tendency for it to decrease further.
Similarly, if 0 < α < 21 then |t + h|2α < |t|2α + |h|2α , so E((X(t) − X(0))(X(t + h) −
X(t))) < 0, and the increments X(t) − X(0) and X(t + h) − X(t) are negatively
correlated. Thus if the sample path has increased after a certain time, there is a
tendency for it to decrease.

16
 Chapter 17

17.1 The Legendre transform is inf q {e−q + qα}. Writing g(q) = e−q + qα we have

dg d2 g
= −e−q + α, = e−q ,
dq dq 2
so g takes a minimum at q = − log α, so the Legendre transform is elog α − α log α =
α(1 − log α).

17.2 If x ∈ sptµ1 then µ2 (B(x, r)) = 0 is r is small enough, since the supports of µ1 and µ2
are disjoint, so ν(B(x, r) = µ1 (B(x, r)) for small r, giving dimloc ν(x) = dimloc µ1 (x).
Similarly, if x ∈ sptµ2 then dimloc ν(x) = dimloc µ2 (x). Thus

{x : dimloc ν(x) = α} = {x : dimloc µ1 (x) = α} ∪ {x : dimloc µ2 (x) = α}.

We get the fine spectra by taking the Hausdorff dimensions of these sets, so

fHν (α) = max{fH1 (α), fH2 (α)}.

Taking µ1 and µ2 to be self-similar measures, such that fH1 and fH2 are different
concave functions with graphs that cross, fHν will fail to be a convex function.

17.3 Suppose that b|x − y| ≤ |g(x) − g(y)| ≤ c|x − y| where 0 < b ≤ c. Then for x ∈ Rn
we have g(B(x, r/c)) ⊂ B(g(x), r) ⊂ g(B(x, r/b)), or B(x, r/c) ⊂ g −1 (B(g(x), r)) ⊂
B(x, r/b). Thus µ(B(x, r/c)) ≤ ν(B(g(x), r)) ≤ µ(B(x, r/b)). Taking logarithms,
for small enough r.
log µ(B(x, r/c)) log ν(B(g(x), r)) log µ(B(x, r/b))
≥ ≥ ,
log r/c + log c log r log r/b + log b
so letting r → 0, we get dimloc µ(x) ≥ dimloc ν(g(x)) ≥ dimloc µ(x) assuming these
limits exist, so dimloc ν(g(x)) = dimloc µ(x).
It follows that for all α

g{x : dimloc µ(x) = α} = {y : dimloc ν(y) = α};

since g is bi-Lipschitz the sets {x : dimloc µ(x) = α} and {y : dimloc ν(y) = α} have
the same dimension, that is the fine (Hausdorff) multifractal spectra for µ and ν
are identical.

17.4 We have
β(q) β(q)
1 = pq1 r1 + pq2 r2 = (pq1 + pq2 )4−β(q) .
Taking logarithms,
log(pq1 + pq2 )
β(q) = .
log 4
For each q,
dβ pq log p1 + pq2 log p2
α=− =− 1 q ,
dq (p1 + pq2 ) log 4

17
 so
dβ q(pq1 log p1 + pq2 log p2 ) log(pq1 + pq2 )
f (α) = −q + β(q) = − + .
dq (pq1 + pq2 ) log 4 log 4

17.5 First, take r = 4−k . With the intervals Ii1 ,...,ik in the construction of the middle-half
Cantor F set indexed in the usual way, see (17.22), we get from (17.6):
X X q
M4−k (q) = µ(Ii1 ,...,ik )q = pi1 . . . pqik = (pqi1 + pqi2 )k .
i1 ,...,ik i1 ,...,ik

Now suppose that 4−k−1 ≤ r < 4−k . Then each mesh interval of length r intersects
at most one of the k-th level component intervals Ii1 ,...,ik of F , and each Ii1 ,...,ik
intersects at most 3 mesh intervals of length r. Thus
X
3−q (pqi1 + pqi2 )k = ( 13 µ(Ii1 ,...,ik ))q ≤ Mr (q)
i1 ,...,ik

X
≤3 µ(Ii1 ,...,ik )q = 3(pqi1 + pqi2 )k .
i1 ,...,ik

Hence
log 3−q (pqi1 + pqi2 )k log Mr (q) log 3(pqi1 + pqi2 )k
≤ ≤
− log 4−k−1 − log r − log 4−k
or
−q log 3 + k log(pqi1 + pqi2 ) log Mr (q) log 3 + k log(pqi1 + pqi2 )
≤ ≤ .
(k + 1) log 4 − log r k log 4
log Mr (q) log(pqi1 + pqi2 )
Letting r → 0, so k → ∞, and (17.7) gives β(q) = lim = .
r→0 − log r log 4

17.6 By (17.26), pq1 ( 21 )−β + pq2 ( 14 )−β = 1. This is a quadratic equation in x = ( 12 )−β , that
is pq2 x2 + pq1 x − 1 = 0. Thus x = (−pq1 + (p2q q 1/2
1 + 4p2 ) )/2pq2 (taking the positive
solution since x > 0). Hence

log x log(2pq2 ) − log((p2q q 1/2

1 + 4p2 ) ) − pq1 )
β(q) = − = .
log 2 log 2

17.7 From (17.13)

log(pq1 + pq2 ) − log(p−q −q

1 + p2 )
β(q) − β(−q) =
log 3
log p1 p2 (p1 + p−q
q q −q −q −q
2 ) − log(p1 + p2 )
=
log 3
q q
log p1 p2 q log(p1 p2 )
= = .
log 3 log 3

18
 Thus
  
log(p1 p2 )
f (α) = inf {β(q) + qα} = inf β(−q) + q α +
q q log 3
    
log(p1 p2 ) log(p1 p2 )
= inf β(q) − q α + = f −α − .
q log 3 log 3

17.8 Since p1 < p2 ,

log(pq1 + pq2 ) log pq2 (1 + (p1 /p2 )q ) q log p2 + log(1 + (p1 /p2 )q )
β(q) = = =
log 3 log 3 log 3
q log p2 q log p2
= + O((p1 /p2 )q )) = + o(1)
log 3 log 3
as q → ∞. Similarly
q log p1
β(q) = + o(1)
log 3
as q → −∞.
Thus if the β curve approaches the line β = aq + b as q → ∞, a = log p2 / log 3
and b = 0. Thus β = q log p2 / log 3 is the asymptote as q → ∞, and similarly
β = q log p1 / log 3 is the asymptote as q → −∞, both of these lines passing through
the origin.
The slopes of the asymptotes give the extreme values of α, so αmin = − log p2 / log 3
and αmax = − log p1 / log 3. Moreover, f (αmin ) = f (αmax ) = 0, since these values are
given by the intercepts of the asymptotes with the vertical axis.

17.9 From (17.34) df /dα = q, so d2 f /dα2 = dq/dα = 1/(dα/dq) = 1/(−d2 β/dq 2 ) < 0,
since β(q) is convex, using that α = −dβ/dq.

17.10 Clearly, β(1) = 0. For 0 < q < 1, we have, by Hölder’s inequality, that
m m
!q m !1−q
X q 1−q
X X
pi ri ≤ pi ri ≤ 1.
i=1 i=1 i=1

Pm β(q)
Since i=1 pqi ri = 1, we have β(q) ≤ 1 − q.
For q > 1, by Hölder’s inequality
m
X m
X (1−q)/q (q−1)/q
1= pi = (pi ri )(ri )
i=1 i=1

m
!1/q m
!(q−1)/q
X X
≤ pqi ri1−q ri .
i=1 i=1
Pm
Hence 1 ≥ i=1 pqi ri1−q , so β(q) ≥ 1 − q.

19
17.11 Note that for all x ∈ R2 and r > 0 we have proj(B(x, r)) = BL (projx, r), so

(projµ)(BL (projx, r)) = µ{y ∈ R2 : projy ∈ BL (projx, r)} ≥ µ(B(x, r)).

Thus
(log(projµ)(BL (projx, r))) log µ(B(x, r))
limr→0 ≤ limr→0 .
log r log r

17.12 Take  > 0. Let Qk denote those k-th level sequences i = (i1 , . . . , ik ) ∈ Ik such that
µ(Ii ) ≤ |Ii |α− . For q < 0:
X X X
|Ii |β+q(α−) ≤ |Ii |β µ(Ii )q ≤ |Ii |β µ(Ii )q
i∈Qk i∈Qk i∈Ik
X
= (ci1 ci2 . . . cik )β (pi1 pi2 . . . pik )q
i1 ,...,ik
m
!k
X
= pqi cβi = 1,
i=1

using a multinomial expansion and (17.26).

For each integer K, write

FαK = {x ∈ F : µ(Ik (x)) ≤ |Ik (x)|α− for all k ≥ K},

where Ik (x) is the k-th level interval containing x. Then for all k ≥ K, the set
β+q(α−)
F K ⊂ i∈Qk Ii , so Hck (F K ) ≤ 1, since for a k-th level interval, |Ii | ≤ ck where
S

c = max1≤i≤m ci . Letting k S → ∞ gives Hβ+q(α−) (F K ) ≤ 1, so that dimH (F K ) ≤

β + q(α − ). But Fα ⊂ ∞ K
K=1 F , since if log µ(Ik (x))/ log |Ik (x)| → α then
α−
µ(Ik (x)) ≤ |Ik (x)| for all k sufficiently large. Thus dimH (Fα ) ≤ β + q(α − )
for all  > 0, giving dimH (Fα ) ≤ β + qα.

17.13 In the partial proof of Theorem 17.4, if we set

F K = {x ∈ F : µ(Ik (x)) ≥ |Ik (x)|α+ for some k ≥ K}

then, as before, dimH ∞ K

S
K=1 F ≤ β + q(α + ). But
∞
[
S ≡ {x ∈ F : lim log µ(Ik (x))/ log |Ik (x)| ≤ α} ⊂ F K,
k→∞
K=1

so dimH S ≤ β + q(α + ) for all  > 0, so dimH S ≤ β + qα.

17.14 By a similar argument to those of Proposition 17.17 and Theorem 17.4, we see that,
for all α > αmin ,

dimH (Fαmin ) ≤ dimH {x : lim inf log µ(B(x, r))/ log r ≤ α} ≤ f (α).
r→0

20
Since the Legendre transform f (α) of β(q) is continuous on [αmin , αmax ] and increas-
ing near αmin (corresponding to q → ∞), we conclude that

dimH (Fαmin ) ≤ lim f (α) = f (αmin ). (1)

α&αmin

Now consider the asymptote of β(q) as q → ∞. If this is the line y = b − aq, then
β(q) = b − aq + o(1) for large q. From the definition of β
m
X m
X b+o(1)
1= pi riβ (q) = (pi ri−a )q ri , (2)
i=1 i=1

and, on inspecting the dominant term of the sum, this requires maxi (pi ri−a ) = 1,
i.e. a = mini log pi / log ri = αmin .
Writing I0 = {i : log pi / log ri = a}, if I0 contains a single index i0 , then from (2)
b = 0 so that the asymptote of β passes through the origin. Thus considering the
Legendre transform f of β as q → ∞ we get that limα&αmin f (α) = f (αmin ) = 0, so
from (1) dimH (Fαmin ) = f (αmin ) = 0.
b
P
If I0 contains more than one index, then from (2), 1 = i∈I0 ri . Writing E for
the self-similar attractor of the IFS {Si : i ∈ I0 }, since strong separation holds for
E (as it does for F ), dimH E = b. Moreover, if x ∈ E, it may be checked that
dimloc µ(x) = a = αmin , so E ⊆ Fαmin . But the asymptote of β(q) as q → ∞ cuts
the vertical axis at b, so the Legendre transform f (αmin ) = b. Thus dimH (Fαmin ) ≥
dimH (E) = f (αmin ), with the opposite inequality following from (1).

21
 Chapter 18
18.1 From the way that densities of s-dimensional sets behave, see Chapter 5, we might
heuristically expect that µ(B(0, r)) ∼ crs where µ is the mass of the deposit and c is
a constant. Thus the number of shaded squares k might be considered proportional
to this mass ∼ c1 rs , so r ∼ c2 k 1/s .

18.2 Again arguing heuristically, a rate of mass deposition proportional to r(t) implies
that
dm
= cr(t) ∼ c0 m1/s
dt
dm
using m(t) ∼ cr(t)s . Thus, m−1/s ∼ c0 , so solving this differential equation
dt
1−1/s
with m(0) = 0 gives m ∼ c0 t, so m(t) ∼ c1 ts/(s−1) and by Exercise 18.1,
r(t) ∼ c2 m(t)1/s ∼ t1/(s−1) .

18.3 We may express u(y, t + h) in terms of u(y, t) by

−(x − y)2
Z  
1
u(y, t + h) = exp u(x, t)dx.
2πh 2h
Differentiating with respect to h
−(x − y)2 (x − y)2
Z    
∂u 1 1
= exp u(x, t) − 2 dx.
∂h 2π 2h 2h3 h
Now, with x = (x1 , x2 ) and y = (y1 , y2 ), differentiating with respect to yi gives
−(x − y)2
   
xi − y i
Z
∂u 1
= exp u(x, t) dx,
∂yi 2πh 2h h
∂ 2u −(x − y)2 (xi − yi )2
Z    
1 1
= exp u(x, t) − dx.
∂yi2 2πh 2h h2 h
Thus
∂ 2u ∂ 2u −(x − y)2 (x − y)2
Z    
2 1 2
∇ u= 2 + 2 = exp u(x, t) − 2 dx,
∂y1 ∂y2 2π 2h h3 h
∂u 1
so = ∇2 u.
∂h 2

18.4 We first establish the identity

∇2 ψ(x)
Z
dx = −4πψ(y)
|y − x|
for a smooth (twice continuously differentiable, say) ψ(x) that is zero for all suffici-
1
etly large x. To see this, note that for small  > 0, since ∇2 |y−x| = 0 for x 6= y,

∇2 ψ(x)dx ∇2 ψ(x)
Z Z  
1
= − ψ(x)∇2 dx
|y−x|≥ |y − x| |y−x|≥ |y − x| |y − x|
 
∇ψ
Z
1
= − − ψ(x)∇ .dn(x) (1)
|y−x|= |y − x| |y − x|

22
 where n(x) denotes the outwards pointing unit normal at x on the sphere |y −x| = .
Here we have used Green’s theorem for a region between a sphere of radius  and
a large sphere on and outside which ψ(x) = 0. Differentiating with respect to each
1 (x − y)
coordinate gives ∇x =− . Hence (1) gives
|y − x| |y − x|3

∇2 ψ(x) ∇2 ψ(x)
Z Z
dx = lim dx
|y − x| →0 |y−x|≥ |y − x|
 
∇ψ (x − y)
Z
= − lim + ψ(x) .dn(x)
→0 |y−x|=  3
= −4πψ(y), (2)

since φ is continuous at y, and |y−x|= −1 ∇2 ψ(x)dn(x) = O(), and |y−x|= (x −

R R

y).dn(x) = 4π2 .
Z
3 f (y)
Now let f be continuous and integrable on R and let φ(x) = dy. Let ψ(x)
|y − x|
be any smooth function vanishing for all sufficiently large x. Using Green’s formula,
f (y)∇2 ψ(x)
Z Z Z Z
2 2
ψ∇ φdx = φ∇ ψdx = dx dy
|y − x|
∇2 ψ(x)
Z Z
= f (y) dx dy
|y − x|
Z
= − 4πf (y)ψ(y)
R
using (2). Hence, (∇2 φ + 4πf )ψ = 0 for all smooth ψ vanishing for large x, so by
an orthogonality argument, ∇2 φ + 4πf = 0, as required.

18.5 Assume that f (x) = 0 for x ∈/ B(0, r). Using the Cauchy-Schwartz inequality and
setting R = r + |x|, we have
Z 2
2 f (y)dy
φ(x) =
B(0,r) |x − y|
Z Z
2 dy
≤ f (y) dy 2
B(0,r) B(0,r) |x − y|
Z
dy
≤ c 2
B(0,R) |x − y|
Z Z R 2
r drdθ
= c < ∞,
θ∈S r=0 r2
changing to polar coordinates, with S the unit sphere. Thus the singularity set of
x, that is x such that φ(x) = ∞, is empty.

18.6 Let µ = ν × m be the measure on F = D × L where ν is t-dimensional Hausdorff

measure restricted to a self-similar Cantor dust of dimension t (satisfying the open
set condition) and m is the restriction of 1-dimensional Lebesgue measure to the line
segment L. Then, as in Exercise 9.11, there are constants c1 , c2 such that if x ∈ D

23
 and r < 1/2, we have c1 rt ≤ ν(B(x, r)) ≤ c2 rt . Writing C(w, r) for the cylinder
with axis L(w) parallel to L, center w = (x, z) ∈ D × L, radius r and height 2r, it
follows that

c1 rt+1 = c1 rt r ≤ ν(B(x, r)) × m(L(w)) = µ(C(w, r)) ≤ c2 rt 2r = 2c2 rt+1 .

Since C(w, 2−1/2 r) ⊂ B(w, r) ⊂ C(w, r), it follows that

2−(t+1)/2 c1 rt+1 ≤ µ(B(w, r)) ≤ 2c2 rt+1 .

With s = t + 1 and redefining c1 and c2 , we are in the situation of Section 18.3,

again with Z
h(x)(x + h)dhi ∼ rs ,
|h|≤r

but this time with the dissipation occurs around the stratified set F .

18.7 For a von Koch curve of base length a, the resonant wavelengths are given by the ba-
sic similarity sizes: a, 3−1/2 a, 3−1 a, 3−3/2 . . . (remembering that the von Koch curve
comprises two similar copies at scale 3−1/2 ). Thus the resonant frequencies are pro-
portional to the reciprocals ω, 31/2 ω, 3ω, 33/2 , . . ..

18.8 With B α index-α fractional Brownian motion, we have from (16.10)

Z r
−x2
 
α α 1
p(r) ≡ P(0 ≤ B (t + h) − B (t) ≤ r) = √ exp dx,
2πhα 0 2h2α

In particular
Z ∞
α α q
E(|B (t + h) − B (t)| ) = 2 rq dp(r)
0
Z ∞  2
2 q −r
= √ r exp dr
α
2πh 0 2h2α
Z ∞  2
qα q −u
= ch u exp du
0 2

on setting u = r/hα . Thus

E(|X(t + h) − X(t)|q ) = E(|B α (T (t + h)) − B α (T (t))|q )

= c1 |T (t + h) − T (t)|qα
= c1 µ[t, t + h]qα ∼ hγqα

for t ∈ Eγ .

24