FRACTALS AND SELF SIMILARITY

JOHN E. HUTCHINSON
This is a retyped (TeXd) version of the article from Indiana University Math-
ematics Journal 30 (1981), 713747 (with some minor formatting changes, a few
old typos corrected, and hopefully few new ones introduced).
The original preprint appeared as Research Report No. 31-1979, Department of
Pure Mathematics, Faculty of Science, Australian National University.
Contents
1. Introduction 2
2. Preliminaries 3
2.1. Sequences of Integers 3
2.2. Maps in Metric Spaces 4
2.3. Similitudes 4
2.4. Hausdor Metric 6
2.5. Measures 6
2.6. Hausdor Measure 7
2.7. Geometric Measure Theory 8
3. Invariant Sets 10
3.1. Elementary Proof of Existence and Uniqueness, and Discussion of
Properties 10
3.2. Convergence in the Hausdor Metric 12
3.3. Examples 13
3.4. Remark 14
3.5. Parametrised Curves 14
4. Invariant Measures 15
4.1. Motivation 15
4.2. Denitions 16
4.3. The L metric 16
4.4. Existence and Uniqueness 16
4.5. Dierent Sets of Similitudes Generating the Same Set 17
5. Similitudes 18
5.1. Self-Similar Sets 18
5.2. Open Set Condition 18
5.3. Existence of Self Similar Sets. 19
5.4. Purely Unrectiable Sets. 21
5.5. Parameter Space 24
6. Integral Flat Chains 24
6.1. The T-metric. 24
6.2. The (-metric 25
6.3. Invariant Chains 26
References 27
1
2 JOHN E. HUTCHINSON
1. Introduction
Sets with non-integral Hausdor dimension (2.6) are called fractals by Mandel-
brot. Such sets, when they have the additional property of being in some sense
either strictly or statistically self-similar, have been used extensively by Mandel-
brot and others to model various physical phenomena (c.f. [MB] and the references
there). However, these notions have not so far been studied in a general framework.
In this paper we set up a theory of (strictly) self-similar objects, in a subsequent
paper we analyse statistical self-similarity.
We now proceed to indicate the main results. The reader should refer to the
examples in 3.3 for motivation. We say the compact set K R
n
is invariant if
there exists a nite set o = S
1
, . . . , S
N
of contraction maps on K R
n
such
that
K =
N
_
i=1
S
i
K.
In such a case we say K is invariant with respect to o. Often, but not always, the
S
i
will be similitudes, i.e. a composition of an isometry and a homothety (2.3).
In [MB], and in the case the S
i
are similitudes, such sets are constructed by an
iterative procedure using an initial and a standard polygon. However, here we
need to consider instead the set o.
It turns out, somewhat surprisingly at rst, that the invariant set K is deter-
mined by o. In fact, for given o there exists a unique compact set K invariant
with respect to o. Furthermore, K is the limit of various approximating sequences
of sets which can be constructed from o.
More precisely we have the following result from 3.1(3), 3.2.
(1) Let X = (X, d) be a complete metric space and o = S
1
, . . . , S
N
be a
nite set of contraction maps (2.2) on X. Then there exists a unique closed bounded
set K such that K =

N
i=1
S
i
K. Furthermore, K is compact and is the closure of
the set of xed points s
i1...ip
of nite compositions S
i1
S
ip
of members of o.
For arbitrary A X let o(A) =

N
i=1
S
i
A, o
p
(A) = o(o
p1
(A)). Then for
closed bounded A, o
p
(A) K in the Hausdor metric (2.4).
The compact set K in (1) is denoted [o[. [o[ supports various measures in a
natural way. We have the following from 4.4.
(2) In addition to the hypotheses of (1), suppose
1
, . . . ,
N
(0, 1) and

N
i=1

i
= 1. Then there exists a unique Borel regular measure of total mass 1
such that =

N
i=1

i
S
i#
(). Furthermore spt() = [o[.
The measure is denoted |o, |.
The set [o[ will not normally have integral Hausdor dimension. However, in
case (X, d) is R
n
with the Euclidean metric, [o[ can often be treated as an m-
dimensional object, m an integer, in the sense that there is a notion of integration
of C

m-forms over [o[. In the language of geometric measure theory (2.7), [o[
supports an m-dimensional integral at chain. The main result here is 6.3(3).
Now suppose (X, d) is R
n
with the Euclidean metric, and the S
i
o are simil-
itudes. Let Lip S
i
= r
i
(2.2) and let D be the unique positive number for which

N
i=1
r
D
i
= 1. Then D is called the similarity dimension of o, a term coined by
Mandelbrot. In case a certain separation condition holds, namely the open set
condition of 5.2(1), one has the following consequences from 5.3(1) (see 2.6(1), (3)
for notation).
(3)
(i) D=Hausdor dimension of [o[ and 0 < H
D
([o[) < ,
(ii) H
D
(S
i
[o[ S
j
[o[) = 0 if i ,= j,
FRACTALS AND SELF SIMILARITY 3
(iii) there exist
1
,
2
such that for all k [o[,
0 <
1

D

([o[, k)
D
([o[, k)
2
< ,
(iv) |o, | = [H
D
([o[)]
1
H
D
[o[ if
i
= r
D
i
.
A result equivalent to (3)(i) was rst proved by Moran in [MP].
With a stronger separation condition we prove in 5.4(1) that for suitable m, [o[
meets no m-dimensional C
1
manifold in a set of strictly positive H
m
measure. In
the notation of [FH], [o[ is purely (H
m
, m) unrectiable.
In the case of similitudes in R
n
, it is possible to parametrise invariant sets by
points in a C

manifold (5.5).
My special thanks to F. J. Almgren, Jr., for making possible my stay at Princeton
University, for suggesting I begin this study, and for his continuing comments and
enthusiasm. I would also like to thank L. Simon for advice and for his invitation
to Melbourne University. Helpful suggestions came from members of the Princeton
and Melbourne seminars, especially R. Hardt, V. Scheer and B. White.
Finally I wish to thank Benoit Mandelbrot, from whose ideas this paper devel-
oped.
2. Preliminaries
(X, d) is always a complete metric space, often Euclidean space R
n
with the
Euclidean metric.
B(a, r) = x X : d(a, x) r,
U(a, r) = x X : d(a, x) < r.
If A X, then A is the closure of A, A

is the interior, A is the boundary, and

A
c
is the complement X A.
A C
1
function is one whose rst partial derivatives exist and are continuous. A
C

function is a function having partial derivatives of all orders.

A C
1
manifold in R
n
will mean a continuously dierentiable embedded subman-
ifold having the induced topology from R
n
.
A proper function is a function for which the inverse of every compact set is
compact.
2.1. Sequences of Integers. P = 1, 2, . . . is the set of positive integers. N P,
N 2 is usually xed.
(1) Ordered p-tuples are denoted i
1
, . . . , i
p
), where usually each i
j
1, . . . , N.
We write if , are p-tuples with an initial segment of , i.e. =
i
1
, . . . , i
p
) and = i
1
, . . . , i
p
, i
p+1
, . . . , i
p+q
) for some q 0. means
and ,= .
(2) C(N), the Cantor set on N symbols, is the set of maps (i.e. sequences)
: P 1, . . . , N. Thus C(N) =

p=1
1, . . . , N. We write
p
for (p). A
typical element of C(N) is often written
1
. . .
p
. . . , or i
1
. . . i
p
. . . . We extend the
notation to the case = i
1
, . . . , i
p
) and = i
1
. . . i
p
i
p+1
. . . i
q
. . . C(N).
(3) If i 1, . . . , N and = i
1
, . . . , i
p
) is a p-tuple, then i = i, i
1
, . . . , i
p
) is
just concatenation of i and . Similarly if C(N) then i = i
1
. . .
p1

p
. . . .
Likewise if is a q-tuple and is a p-tuple or C(N) we form in the obvious
way.
The ith shift operator
i
= C(N) C(N) is given by
i
() = i.
(4) C(N) is given the product topology (also called the weak topology) in-
duced from the discrete topology on each factor 1, . . . , N. Thus a sub-basis of
open sets is given by sets of the form :
p
= i where p P, i 1, . . . , N.
C(N) is compact.
4 JOHN E. HUTCHINSON
(5) By

i
1
. . .

i
p
we mean the innite sequence i
1
. . . i
p
i
1
. . . i
p
. . . i
1
. . . i
p
. . .
C(N). Thus i
1
. . . i
p

i
p+1
. . .

i
p+q
may be regarded as the general rational element
of C(N).
(6) The set I will always be a nite set of nite ordered tuples (of not neces-
sarily equal length) from 1, . . . , N.

I =
1
. . .
q
. . . :
i
I C(N), where we are concatenating nite ordered
tuples in the obvious way. Thus if I = 1), . . . , N) then

I = C(N). If I =
1, 2), 1) then 2
2
. . .
p
. . . /

I, etc.
(7) For an ordered tuple, let

= C(N) : .
We say I is secure if for every C(N) there exists I such that .
This is equivalent to: for every p-tuple with p = maxlength : I, there
exists I such that . Since I is nite there is an obvious algorithm to
check if I is secure.
We say I is tight if for every C(N) there exists exactly one I such that
. Again one can always check, in a nite number of steps, if I is tight.
(8) Proposition
(i) The following are equivalent
(1)

I = C(N),
(2) C(N) =

I

,
(3) I is secure.
(ii) The following are equivalent:
(1) Each member of C(N) has a unique decomposition of the form
1
. . .
q
. . . ,
with
i
I,
(2) C(N) =
_
I

(disjoint union),
(3) I is tight.
Proof. In both cases the implications (i)(ii)(iii)(i) are clear.
(9) One can check that I is tight i I is essential and satises the tree condi-
tion, in the sense of [OP, III].
2.2. Maps in Metric Spaces. If F : X X, then we dene the Lipschitz con-
stant of F by
LipF = sup
x=y
d(F(x), F(y))
d(x, y)
.
Of course if LipF = , then d(F(x), F(y)) d(x, y) for all x, y X, and moreover
LipF is the least such . We say F is Lipschitz if LipF < and F is a contraction
if LipF < 1.
(1) It is a standard fact that every contraction map (in a complete metric
space) has a unique xed point.
(2) Denition. Suppose o = S
1
, . . . , S
N
is a nite family of maps S
i
: X
X. Then S
i1...ip
= S
i1
S
ip
.
2.3. Similitudes. S : X X is a similitude if d(S(x), S(y)) = rd(x, y) for all
x, y X and some xed r.

r
: R
n
R
n
is the homothety
r
(x) = rx (r 0).

b
: R
n
R
n
is the translation
b
(x) = x b.
(1) Proposition. S : R
n
R
n
is a similitude i S =
r

b
O for some
homothety
r
, translation
b
, and orthonormal transformation O.
FRACTALS AND SELF SIMILARITY 5
Proof. The only if is clear.
Conversely, let S be a similitude, LipS = r ,= 0. Let g(x) = r
1
(S(x) S(0)).
Then g is an isometry xing 0.
Since
(x, y) =
1
2
_
|x|
2
+|y|
2
|x y|
2

=
1
2
_
[d(0, x)]
2
+ [d(0, y)]
2
[d(x, y)]
2

,
it follows g preserves inner products.
Let e
i
: 1 i N be an orthonormal basis for R
n
. Then g(e
i
) : 1 i N
is also on orthonormal basis, and hence
g(x) =
n

i=1
_
g(x), g(e
i
)
_
g(e
i
) =
n

i=1
(x, e
i
) g(e
i
),
since g preserves inner products. It follows g is linear and so is an orthonormal
transformation.
Since
S(x) = rg(x) +S(0) = r
_
g(x) +r
1
S(0)
_
,
it follows
S =
r

r
1
S(0)
g,
and we are done.
(2) Remark. The same proof works in a Hilbert space to show that S is a
similitude i S =
r

b
O, where now O is a unitary transformation.
(3) Convention. For the rest of this paper, unless mentioned otherwise, all
similitudes are contractions.
(4) Returning to the case (R
n
, d), let the similitude S have xed point a,
let LipS = r, and let O be the orthonormal transformation given by O(x) =
r
1
[S(x +a) a] (orthonormal since the origin is clearly xed and O is clearly an
isometry, now use (1)).
Then
S(x +a) = rO(x) +a,
so
S(x) = rO(x a) +a,
and hence
S =
1
a

r
O
a
= (
1
a

r

a
) (
1
a
O
a
),
so that S may be conveniently thought of as an orthonormal transformation about
a followed by a homothety about a. We write
S = (a, r, O)
and say that S is in canonical form. a and r are uniquely determined by S, and so
is O if r ,= 0.
If S
1
= (a
1
, r
1
, O
1
), S
2
= (a
2
, r
2
, O
2
), then S
1
S
2
= (a, r, O) where r = r
1
r
2
and O = O
1
O
2
. However the expression for a is not as simple, a calculation gives
a = a
2
+ (I r
1
r
2
O
1
O
2
)
1
(I r
1
O
1
)(a
2
a
1
).
6 JOHN E. HUTCHINSON
2.4. Hausdor Metric. If x X, A X, dene the distance between x and A
by
d(x, A) = infd(x, a) : a A.
If A X, > 0, dene the -neighbourhood of A by
A

= x X : d(x, A) < .
Thus A A

.
Let B be the class of non-empty closed bounded subsets of X. Let ( be the class
of non-empty compact subsets.
Dene the Hausdor metric on B by
(A, B) = supd(a, B), d(b, A) : a A, b B.
Thus (A, B) < i A B

and B A

. It is easy to check that is a metric

on B.
It follows from [FH, 2.10.21] that (B, ) is a complete metric space. It also follows
that if K X is compact, then ( A : A K is compact.
Some elementary properties of which we will use are: let F : X X, then
(i) (F(A), F(B)) Lip(F) (A, B),
(ii)
_
iI
A
i
,

iI
B
i
_
sup
iI
(A
i
, B
i
).
2.5. Measures.
(1) A measure on a set X is a map : T(X) = A : A X [0, ] such
that
(i) () = 0,
(ii)
_

i=1
E
i
_

i=1
(E
i
), E
i
X.
It follows A B implies (A) (B). Thus is what is often called an outer
measure. One says A is measurable i (T) = (T A) +(T A) for all T X.
The family of measurable sets forms a -algebra. is a nite measure if (X) < .
If A X, A is the measure dened by A(E) = (A E).
From now on, X = (X, d) is a complete metric space. One says that is Borel
regular i all Borel sets are measurable and for each A X there exists a Borel
set B A with (A) = (B). If is nite and Borel regular, it follows from [FH,
2.2.2.] that for arbitrary Borel sets E X,
(i) (E) = sup (K) : E K closed,
(ii) (E) = inf (V ) : E V open.
(2) We dene the support of to be the closed set
spt = X
_
V : V open, (V ) = 0.
Dene the mass of by
M() = (X).
Dene / to be the set of Borel regular measures having bounded support and
nite mass.
Dene
/
1
= / : M() = 1.
For a X dene
a
= [[a]] /
1
by
a
(A) = 1 if a A,
a
(A) = 0 if a / A.
(3) Let B((X) = f : X R : f is continuous and bounded on bounded subsets.
For /, B((X), dene () =
_
d. Then : B( [0, ), is linear,
and is positive (i.e. (x) 0 for all x implies () 0).
If f : X X is continuous and sends bounded sets to bounded sets (e.g. if f
is Lipschitz), then we dene f
#
: / / by f
#
(E) = (f
1
(E)). Equivalently
f
#
() = ( f). Notice that M(f
#
) = M().
FRACTALS AND SELF SIMILARITY 7
We dene the weak topology on / by taking as a sub-basis all sets of the form
: a < () < b, for arbitrary real a < b and arbitrary B((X). It follows

i
in the weak topology i
i
() () for all B((X).
2.6. Hausdor Measure.
(1) Let the real number k 0 be xed. For every > 0 and E X we dene
H
k

(E) = inf
_

i=1

k
2
k
(diam E
i
)
k
: E

_
i=1
E
i
, diam E
i

_
H
k
(E) = lim
0
H
k

(E) = sup
0
H
k

(E).
H
k
(E) is called the Hausdor k-dimensional measure of E. A reference is [FH,
2.10.3].
k
is a suitable normalising constant. If k is an integer,
k
= L
k
x R
k
:
[x[ 1. For arbitrary k we dene
k
= (
1
2
)
k
/(
k
2
+ 1). The particular value of

k
for non-integer k will not be important. The value of H
k
(E), but not that of
H
k

(E), remains unchanged if we restrict the E

i
to be open (or closed, or convex).
H
k
is a Borel regular measure, but H
k
is not normally nite on bounded sets.
If X = R
n
then H
n
= L
n
. H
0
is counting measure. If f : A R
m
R
n
is C
1
and one-one, then H
m
(f(A)) =
_
A
J(f) dL
m
, where J(f) is the Jacobian. Thus
H
k
agrees with usual notions of k-dimensional volume on nice sets in case k is
an integer.
If F : X X is Lipschitz, then H
k
(F(A)) (LipF)
k
H
k
(A). If F is a similitude,
F
#
H
k
= (LipF)
k
H
k
.
For each E X there is a unique real number k, called the Hausdor dimension
of E, written dimE, such that H

(E) = if < k, H

(E) = 0 if > k. H
k
(E)
can take any value in [0, ].
(2) Suppose S : X X is a similitude with LipS = r. Then H
k
S(A) =
r
k
S
#
(H
k
A). For (H
k
S(A))(E) = H
k
(S(A) E) = H
k
(S(A S
1
(E))) =
r
k
H
k
(A S
1
(E)) = r
k
(H
k
A)(S
1
(E)) = r
k
S
#
(H
k
A)(E).
(3) The lower (upper) k-dimensional density of the set A at the point x is
dened respectively to be

(A, x) = liminf
r0
H
k
(A B(x, r))

k
r
k

k
(A, x) = limsup
r0
H
k
(A B(x, r))

k
r
k
.
If they are equal, their common value is called the k-dimensional density of A at
x, and is written
k
(A, x).
Likewise, for a measure on X we dene

(, x) = liminf
r0
(B(x, r))

k
r
k
,

k
(, x) = limsup
r0
(B(x, r))

k
r
k
,
and
k
(, x) to be their common value if both are equal. Thus
k

(A, x) =
k

(H
k
A, x),
and similarly for
k
,
k
.
Upper densities turn out to be more important than lower densities. The main
results we will need are that for /,
(i)
k
(, a) for all a A implies H
k
(A)
1
(A),
(ii)
k
(, a) for all a A implies H
k
(A) 2
k

1
(A).
8 JOHN E. HUTCHINSON
In particular if 0 < (A) < , and the upper density is bounded away from 0 and
, this enables us to establish that 0 < H
k
(A) < . For a reference see [FH,
2.10.19 (1), (3)].
2.7. Geometric Measure Theory. We will briey sketch the ideas from geo-
metric measure theory needed for 6. A complete treatment is [FH], in particular
Chapter 4, and a good exposition of the main results is in [FH1]. At a number of
places we have found it convenient to abbreviate the standard notation.
(1) Suppose m 0 is a positive integer. A set E R
n
is m-rectiable i
E is H
m
-measurable, H
m
(E) < , and there exist m-dimensional C
1
manifolds
M
i

i=1
in R
n
such that H
m
_
E

i1
M
i
_
= 0. (Here we dier somewhat from
the convention of [FH]). For H
m
a.a. x E the tangent spaces at x to distinct M
i
containing x are equal. Let

E
x
be this tangent space where it exists.
(2) Suppose now we are given
(1) a bounded m-rectiable set E with m 1,
(2) a multiplicity function , i.e. an H
m
-measurable function with domain E
and range a subset of the positive integers, such that
_
E
dH
m
< ,
(3) an orientation

T , i.e. an H
m
-measurable function

T with domain E such
that for H
m
a.a. x E,

T (x) is one of the two simple unit m-vectors
associated with

E
x
.
With the above ingredients we dene a linear operator on C

m-forms by
T() =
_
E
(x)

T (x), (x)
_
dH
m
.
This generalises the notion of integration over an oriented manifold. The set
of all such operators is called the set of m-dimensional rectiable currents. A 0-
dimensional rectiable current is dened to be a linear operator T on C

functions
(i.e. 0-forms) such that
T() =
r

i=1

i
(a
i
),
where r 0,
1
, . . . ,
r
are integers, and a
1
, . . . , a
r
R
n
. Thus T corresponds to
a nite number of points with integer multiplicities. T is written

r
i=1

i
[[a
i
]].
The set of m-dimensional rectiable currents forms an abelian group in a natural
way. It is denoted
m
.
(3) For each T
m
, m 1, we dene a linear operator T on C

(m1)-
forms by Stokes formula:
T() = T(d).
If T corresponds to a compact oriented manifold with boundary, then T corre-
sponds to the oriented boundary. Clearly T = 0. However it is not necessarily
true that T
m1
. Accordingly we dene the abelian group of m-dimensional
integral currents by
I
m
= T
m
: T
m1
if m 0,
I
0
=
0
.
Clearly I
m

m
. For m 1, : I
m
I
m1
, and is a group homomorphism.
We can also enlarge
m
to the abelian group of m-dimensional integral at
chains, or m-chains for short, dened by
T
m
= R+S : R
m
, S
m+1
.
In the natural way is extendible to a group homomorphism : T
m
T
m1
if
m 1.
FRACTALS AND SELF SIMILARITY 9
(4) For T
m
we dene the mass of T by
M(T) =
_
E
dH
m
if m 1,
M
_
r

i=1

i
[[a
i
]]
_
=
r

i=1
[
i
[.
One can extend the denition of M to T
m
, but then one has [FH, 4.2.16].

m
= T
m
T : M(T) < ,
I
m
=
m
T : M(T) < .
One now denes the integral at norm on T
m
by
T(T) = infM(R) +M(S) : T = R +S,
and the integral at metric by
T(T
1
, T
2
) = T(T
1
T
2
).
Figure 2.1 T(T
1
, T
2
) = T(T
1
T
2
).
Thus T
1
and T
2
of Figure 2.1 are close in the T-metric since there exist R and S
of small mass such that T
1
T
2
= R+S. is continuous in the T-metric, indeed
T(T
1
, T
2
) T(T
1
, T
2
).
(5) The m-dimensional integral at chains generalise the notion of an oriented
m-dimensional C
1
manifold, but retain many of the desirable properties and at the
same time are closed under various useful operations.
Thus T
m
is a complete metric space under T [FH, 4.1.24]. The inmum in the
denition of T is always realised [FH, 4.2.18].
Convergence in M implies convergence in T, but certainly not conversely. If
T
j
T in T, then M(T) liminf M(T
j
).
For integral currents there is the important compactness theorem [FH, 4.2.17]:
if K R
n
is compact and c < , then
T I
m
: M(T) < c, M(T) < c, spt T K
is compact in the T-topology. For T T
m
we dene spt T, the support of T, to be
the intersection of all closed sets C such that spt C = implies T() = 0.
If T T
m
, m 1, and T = 0 (or if T T
0
), we say T is an m-dimensional
integral at cycle or m-cycle for short. If m 1, it follows by a cone construction
[FH, 4.1.11] that T = S for some S T
m+1
. Furthermore, one has the isoperi-
metric inequality [FH, 4.2.10]: for m 1 there is a constant = (m, n) depending
only on m and n, such that if T I
m
and T = 0, then T = S for some S I
m+1
with M(S) M(T)
m+1/m
.
If T T
m
and T = S for some S T
m+1
, we say T is an m-dimensional
integral at boundary, or m-boundary for short. Thus if m 1, every m-cycle is an
m-boundary.
(6) If T T
m
and f : R
n
R
n
is Lipschitz and proper, then one denes
f
#
T T
m
[FH, 4.1.14, 4.1.24]. In case T corresponds to an oriented manifold and
10 JOHN E. HUTCHINSON
f is C
1
, then f
#
T corresponds to the oriented image of T under f, with appropriate
multiplicities if f is not one-one.
The properties of f
#
T we will need are:
(a) f
#
T = f
#
T;
(b) f : T
m
T
m
, is linear, and is continuous in the T-metric;
(c) if Lipf = r and T T
m
, then M(f
#
T) r
m
M(T) and T(f
#
T)
maxr
m
, r
m+1
T(T);
(d) spt f
#
T f(sptT).
(7) One can generalise from the integral at chains to the so-called at chains,
and even more generally to the currents of de Rham. However, one loses the useful
geometric properties of the integral at chains. For a full treatment of all these
subjects see [FH].
3. Invariant Sets
We follow the notation of 2.2, and the other subsections of 2 as necessary. o =
S
1
, . . . , S
N
is a set of contraction maps on the complete metric space (X, d).
Lip S
i
= r
i
. s
i1...ip
is the xed point of S
i1...ip
.
We show the existence and uniqueness of a compact set invariant with respect
to S, and discuss its properties.
We suggest the reader considers Examples 3.3 for motivation.
3.1. Elementary Proof of Existence and Uniqueness, and Discussion of
Properties.
(1) For arbitrary A X let o(A) =

N
i=1
S
i
(A). Let o
0
(A) = A, o
1
(A) = o(A),
o
p
(A) = o(o
p1
(A)) for p 2. We will often use the notation A
i1...ip
= S
i1...ip
(A).
Notice o
p
(A) =

i1,...,ip
A
i1...ip
. Notice also that diam (A
i1...ip
) r
i1
. . . r
ip
diam
(A) 0 as p , provided A is bounded.
(2) Denition. A is invariant (with respect to o) if A = o(A).
(3) Theorem and Denitions.
(i) There is a unique closed bounded set K which is invariant with respect to
o. Thus K =

N
i=1
K
i
. Moreover K is compact.
(ii) K
i1...ip
=

N
ip+1=1
K
i1...ipip+1
.
(iii) K K
i1
K
i1...ip
, and

p=1
K
i...ip
is a singleton whose
member is denoted k
i1...ip...
. K is the union of these singletons.
(iv) k

i1...

ip
= s
i1...ip
, and in particular s
i1...ip
K (recall 2.1(5)). Also k
i1...ip...
=
lim
p
s
i1...ip
, and in particular this limit exists.
(v) K is the closure of the set of xed points of the S
i1...ip
.
(vi) S
j1...jq
(K
i1...ip
) = K
j1...jqi1...ip
, S
j1...jq
(k
i1...ip...
) = k
j1...jqi1...ip...
.
(vii) The coordinate map : C(N) K given by () = k

is a continuous
map onto K.
(viii) If A is a non-empty bounded set, then d(A
i1...ip
, k
i1...ip...
) 0 uniformly as
p . In particular o
p
(A) K in the Hausdor metric.
(4) Proof of Uniqueness. We rst remark that (i) and (viii) are established
in 3.2 independently of the following.
FRACTALS AND SELF SIMILARITY 11
Assume now that K is a closed bounded set invariant with respect to o, and
observe the following consequences.
K =
N
_
i=1
S
i
(K) =
_
i,j
S
i
(S
j
K) =
_
i,j
S
ij
(K) =
_
i,j
K
ij
. . .
=
_
i1,...,ip
K
i1...ip
.
Similarly
K
i1...ip
= S
i1...ip
(K) = S
i1...ip
_
N
_
ip+1=1
S
ip+1
(K)
_
=
N
_
ip+1=1
S
i1...ip+1
(K) =
N
_
ip+1=1
K
i1...ipip+1
.
Thus K K
i1
K
i1i2
K
i1...ip
, and since diam (K
i1...ip
) as p ,

p
K
i1...ip
is a singleton (by completeness of X) whose unique member we denote
k
i1...ip...
. Thus we have established (ii) and (iii) under the given assumptions on K.
The rst part of (vi) is immediate, since
S
j1...jq
(K
i1...ip
) = S
j1...jq
(S
i1...ip
(K)) = S
j1...jqi1...ip
(K) = K
j1...jqi1...ip
.
The second part follows, since
S
j1...jq
(k
i1...ip...
) S
j1...jq

p=1
K
j1...ip
=

p=1
K
j1...jqi1...ip
= k
j1...jqi1...ip...
.
Since S
i1...ip
(k

i1...

ip
) = k

i1...

ip
by the above, it follows k

i1...

ip
is the unique xed
point s
i1...ip
of S
i1...ip
. It follows both s
i1...ip
, k
i1...ip...
K
i1...ip
, and hence since
lim
p
diam (K
i1...ip
) = 0, that lim
p
s
i1...ip
= k
i1...ip...
. This establishes (iv),
and (v) follows from (iv). Notice we have established the uniqueness of K (since K
is the union of singletons, each of which is the limit of a certain sequence of xed
points of the S
i1...ip
).
To establish (vii), and hence that K is compact (being the continuous image
of a compact set), let be as in (vii). Suppose =
1
. . .
p
. . . ) C(N) and
> 0. Then () = k
1...p...
and so there is a q such that K
1...q
x K :
d(x, ()) < . Since K
1...q
is the image of the open set :
i
=
i
if i q,
it follows is continuous.
To prove (viii) suppose A is non-empty and bounded. Then
d(A
i1...ip
, k
i1...ip...
) = d(S
i1...ip
(A), S
i1...ip
(k
ip+1...
))
r
i1
. . . r
ip
d(A, k
ip+1...
)
r
i1
. . . r
ip
supd(a, b) : a A, b K
Constant(max
1iN
r
i
)
p
0 as p .
All that remains now is to prove the existence of a closed bounded invariant set.
But notice that we know from (v) what this set must be.
(5) Proof of Existence. First we need to establish the following lemma.
Lemma. If S
1
, . . . , S
N
is a set of contraction maps on a complete metric
(X, d), and s
i1...ip
is the xed point of S
i1...ip
= S
i1
S
ip
then for each sequence
i
1
. . . i
p
. . ., lim
p
s
i1...ip
exists.
12 JOHN E. HUTCHINSON
Proof. Let = max
1i,jN
d(s
i
, s
j
), and let R = (1 r)
1
where r = maxr
i
=
Lip(S
i
) : 1 i N.
Then

N
i=1
B(s
i
, rR)

N
i=1
B(s
i
, R) = C, say. For if d(s
i
, x) rR then
d(s
j
, x) + rR = + r(1 r)
1
= (1 r)
1
= R. Thus S
i
C C for
i = 1, . . . , N, and so C S
i1
(C) S
i1i2
(C) S
i1...ip
(C) , i.e. C
C
i1
C
i1i2
C
i1...ip
. But the xed point s
i1...ip
must lie in S
i1...ip
(C),
and so since diam(S
i1...ip
(C)) 0 as p and the S
i1...ip
(C) are closed, it
follows lim
p
s
i1...ip
exist and is the unique member of

p=1
S
i1...ip
(C).
For C(N) let s = lim
p
s
1...p
, and let K = s

: C(N). Then
S
i
(s

) = s
i
, since S
i
(s

) S
i
(

p=1
C
1...p
) =

p=1
C
i1...p
s
i
(notice that it
is not normally true that S
i
(s
1...p
) = (s
i1...p
)). Thus K =

N
i=1
S
i
(K) = o(K),
i.e. K is invariant with respect to o.
It remains to prove that K is compact. Dene : C(N) K by () = s

.
Since diam (K) is bounded (being a subset of C in the previous lemma) it follows
precisely as in the proof of (vii) that is continuous and hence that K is compact.
This gives the existence of (i) and completes the proof of the theorem.
(6) Denition. The compact set invariant under o is denoted by [o[.
(7) Non-compact invariant sets. There are always non-bounded invariant
sets, R
n
being a trivial example.
For any A, o(A) = A implies o(

A) =

A. Thus if A is bounded and invariant,
then so is

A, and hence

A = [o[ by (3)(i). For example, o1
2
(0, 1) = (0, 1) where o1
2
is as in 3.3(1).
(8) The following observation is useful. Suppose A is a set such that o(A) A.
Then clearly A o(A) o
2
(A) o
p
(A) . If furthermore A is closed
and non-empty, then K(= [o[) A, and K
i1...ip
A
i1...ip
for all i
1
, . . . , i
p
.
To see this latter, choose a A. Then by (3)(viii) for each xed i
1
, . . . , i
p
, . . .,
k
i1...ip...
= lim
p
S
i1...ip
(a) A. Hence K A. Applying S
i1...ip
to both sides,
K
i1...ip
A
i1...ip
.
(9) If

N
i=1
r
i
< 1, then K is totally disconnected. For given a, b K select p
such that
_
i1,...,ip
r
i
. . . r
ip
_
=
_
N
i=1
r
i
_
p
< d(a, b), where = diamK. Since
K =

ip,...,ip
K
i1...ip
, and diamK
i1...ip
= r
i
. . . r
ip
, it follows by an elementary
argument that a and b are in distinct components of K.
3.2. Convergence in the Hausdor Metric. We remark that this Section is
independent of 3.1(3), (4).
Let B be the family of closed bounded subsets of X, ( the family of compact
subsets. Clearly o : B B and o : ( (. We have
(1) Theorem. o is a contraction map on B (respectively () in the Hausdor
metric.
Proof.

_
o(A), o(B)
_
= (
_
i
S
i
(A),
_
i
S
i
(B))
max
1iN
(S
i
(A), S
i
(B))
(max
1iN
r
i
)(A, B).
Existence and uniqueness of a closed bounded invariant set [o[ follow from the
contraction mapping principle. Since ( is a closed subset of B, it follows that
[o[ (.
FRACTALS AND SELF SIMILARITY 13
Remark added subsequent to publication: As pointed out by a number of people,
since A B does not imply S
i
(A) closed, one should replace S
i
(A) by its closure
in the denition of o(A), when o operates on B. This new map o has a unique
xed point in B, which must then agree by uniqueness with the xed point of o
operating on (.
3.3. Examples.
(1) Cantor Set. In the notation of 2.3 let
o
r
= S
1
(r), S
2
(r), S
i
(r) : R R,
S
1
(r) = (0, r, I), S
2
(r) = (1, r, I),
where I is the identity map.
Figure 3.1 The classical Cantor set C.
If r =
1
3
, then o
r
(C) = C where C is the classical Cantor set, and so [o1
3
[ = C.
We have sketched C, more precisely o
3
([0, 1]), in Figure 3.1. Notice the numbering
system for the various components C
i1...ip
.
If 0 < r <
1
2
, then [o
r
[ is a generalised Cantor set. It is standard, and a
consequence of 5.3(1)(ii), that dim [o
r
[ = log 2/log(
1
r
).
If
1
2
r < 1, then o
r
([0, 1]) = [0, 1], and hence [o
r
[ = [0, 1]. Thus dierent o
r
can generate the same set. In this connection see 4.1.
(2) Koch Curve. We refer to Figure 3.2. Let a
1
, a
2
, a
3
, a
4
, a
5
be as shown.
Let o = S
1
, S
2
, S
3
, S
4
where S
i
: R
2
R
2
is the unique similitude mapping

a
1
a
5
to

a
i
a
i+1
and having positive determinant (i.e. no reections).
Figure 3.2 The Koch curve K.
Let K = [o[. Actually Figure 3.2 shows the approximation o
4
([a
1
, a
5
]) to K.
Notice how one nds the components of K, e.g. K
331
. S
i
has the xed point s
i
= k

i
;
14 JOHN E. HUTCHINSON
s
1
= a
1
, s
4
= a
5
, and s
2
, s
3
are shown. Similarly S
ij
= S
i
S
j
has the xed point
s
ij
= k

j
, where s
23
is shown.
One can visualise o
p
(A) =

i1,...,ip
A
i1...ip
K as p , for arbitrary
bounded A (e.g. A a singleton).
Now let o

= S

1
, S

2
, where S

i
is the unique similitude mapping

a
1
a
5
to

a
1
a
3
(i = 1),

a
3
a
5
(i = 2), having negative determinant (i.e. the S

i
include a reection
component). Then it follows o

(K) = K and hence [o

[ = K. For S

1
S

1
= S
1
,
S

1
S

2
= S
2
, S

2
S

1
= S
3
, S

1
S

2
= S
4
, hence (o

)
2
= o, hence [o

[ is xed by o,
hence [o

[ = [o[ by uniqueness. Thus as in (2), dierent o can generate the same

set.
(3) Let M R
n
be an oriented m-dimensional manifold with oriented bound-
ary N as in Figure 3.3. Let o = S
1
, . . . , S
N
where S
i
: R
n
R
n
are contraction
maps such that

N
i=1
S
i
(N) = N, taking into account orientation and after allow-
ing cancellation of portions of manifolds having opposite orientation. Obviously
such o are easy to nd.
Figure 3.3 M R
3
, M has the boundary N. Consider the case
where M and N do not lie in a plane.
[o[ will normally have dimension > m, and so cannot be an m-dimensional
manifold, yet in some sense [o[ is an m-dimensional object with oriented boundary
N. We make this precise in 6, where under mild restrictions on the S
i
, [o[ becomes
an integral at m-chain having N as its boundary.
3.4. Remark. The following gives a curious characterisation of line segments:
A compact connected set A R
n
is a line segment i A = o(A) for some o =
S
1
, . . . , S
N
where N 2 , the S
i
are similitudes, LipS
i
= r
i
, and

N
i=1
r
i
= 1.
Proof. One direction is trivial.
Conversely, suppose A = o(A) with o as above. Let diamA = d(p, q) where
p, q A. By projecting A onto the line segment pq, one sees that H
1
(A) diamA.
If A ,= pq, by taking the nearest point retraction of A onto a suitably thin solid
ellipsoid having p and q as extremal points, one nds H
1
(A) H
1
((A)). But
H
1
((A)) diamA as we just saw, and so H
1
(A) diamA unless A = pq. One
can check H
1
(A) diamA by a covering argument, and so the required result
follows.
The above was in response to a query of B. Mandelbrot concerning characteri-
sations of the line his query in turn arose from some rather vague remarks in a
work of Liebniz. F. J. Almgren Jr. suggested a shortening of the original proof.
FRACTALS AND SELF SIMILARITY 15
3.5. Parametrised Curves.
(1) Suppose o = S
1
, . . . , S
N
has the property that
a = s
1
= xed point of S
1
,
b = s
N
= xed point of S
N
,
S
i
(b) = S
i+1
(a) if 1 i N 1,
(for example, 3.3(2)). Then one can dene a continuous f : [0, 1] [o[, with Image
(f) = [o[, in a natural way.
For this purpose, x 0 = t
1
< t
2
< < t
N+1
= 1. Dene g
i
: [t
i
, t
i+1
] (0, 1)
for 1 i N by
g
i
(x) =
x t
i
t
i+1
t
i
.
Let
T = T(a, b) = f : [0, 1] X : f is continuous, f(0) = a, f(1) = b.
Dene o(f) for f T by
o(f)(x) = S
i
f g
i
(x) for x [t
i
, t
i+1
], 1 i N.
Dene a metric T on T by
T(f
1
, f
2
) = sup[f
1
(x) f
2
(x)[ : x [0, 1].
P is clearly a metric, and is furthermore complete since the uniform limit of con-
tinuous functions is continuous.
(2) Proposition. o is well-dened, o : T T, and o is a contraction map in
the metric T.
Proof. o is well-dened and o(f) T if f T, since S
i
f g
i
(t
i+1
) = S
i

f(1) = S
i
(b) = S
i+1
(a) = S
i+1
f(0) = S
i+1
f g
i+1
(t
i+1
) for 1 i N 1,
S
1
f g
1
(0) = S
1
f(0) = S
1
(a) = a, and S
N
f g
N
(1) = S
N
f(1) = S
N
(b) = b.
Now suppose x [t
i
, t
i+1
] and f
1
, f
2
T. Then
[o(f
1
)(x) o(f
2
)(x)[ = [S
i
f g
i
(x) S
i
f
2
g
i
(x)[
Lip S
i
[f
1
(g
i
(x)) f
2
(g
i
(x))[
Lip S
i
T(f
1
, f
2
).
Hence T(o(f
1
), o(f
2
)) rT(f
1
, f
2
), where r = maxLipS
i
: 1 i N. It follows
o is a contraction map.
(3) Theorem. Under the hypotheses on o in (1), there is a unique g T such
that o(g) = g. Furthermore Image (g) = [S[.
Proof. The existence of a unique such g follows from (2).
By construction, Image o(f) = o(Imagef) for every f T. If o(g) = g, this
implies Image g = o(Image g), and hence Image g = [o[ by 3.1(3)(i).
(4) It is often possible to parametrise other invariant sets [o[ R by maps
g : x R
m
: [x[ 1 R
n
for suitable m, for example m = 2 in 3.3(3). But
if m > 1 there is a lot of arbitrariness in the selection of the particular parametric
map g. It is often better to treat [o[ as an intrinsic m-dimensional object in R
n
via the notion of an m-dimensional integral at chain, c.f. 2.7, and 6.
16 JOHN E. HUTCHINSON
4. Invariant Measures
4.1. Motivation. A motivation for this section is the following. In 3.3(1), (2) we
saw examples of dierent families of contractions generating the same set. Yet the
o
r
of 3.3(1) seem to be dierent from one another in a way that S and o

of 3.3(2)
are not. We make this precise in 4.4(6).
Another motivation is that it will be easier to use invariant sets if we can
impose additional natural structure on them, in this case a measure.
4.2. Denitions. (X, d) is a complete metric space, o = S
1
, . . . , S
N
is a family of
contraction maps. Additionally, we assume the existence of a set =
1
, . . . ,
N

with
i
(0, 1) and

N
i=1

i
= 1. In 5 we will see that in case the S
i
are similitudes
with Lip S
i
= r
i
, it is natural to take
i
= r
D
i
, where D is the similarity dimension
of S, 5.1(3).
We refer back to 2.5 for terminology on measure.
(1) Denition. If / let (o, )() =

N
i=1

i
S
i#
. Thus (o, )()(A) =

N
i=1

i
(S
1
i
(A)). Let (o, )
0
() = , (o, )
1
() = (o, )(), and (o, )
p
() =
(o, )((o, )
p1
()) for p 2. Let

i1...ip
=
i1
. . .
ip
S
i1...ip#
().
(2) Notice (o, )
p
() =

i1,...,ip

i1...ip
. Also M((o, )()) = M() and so
M((o, )
p
()) = M() for all p. In particular, (o, ) : /
1
/
1
.
(3) Denition. is invariant (with respect to (o, )) if (o, )() = .
4.3. The L metric. We introduce a metric L on /
1
(see 2.5(2)) similar to the
one introduced by Almgren in [AF, 2.6], but modied in a way which enables 4.4(1)
to hold.
(1) Denition. For , /
1
let
L(, ) = sup() () [ : X R, Lip 1.
Notice that of the denition is a member of B((X), and notice also that there
is no restriction on sup(x) : x X.
In checking that L is indeed a metric, the only part which is not completely
straightforward is verifying L(, ) < . So suppose spt spt B(a, R).
Then for Lip 1,
() () =
_
(a) +(a)
_

_
(a) +(a)
_
=
_
(a)
_

_
(a)
_
, since ((a)) = ((a))
(R) +(R)
= 2R.
One can check that the L metric topology and the weak topology coincide on
H
1
: spt is compact.
Finally notice that L(
a
,
b
) = d(a, b).
4.4. Existence and Uniqueness.
(1) Theorem.
(i) (o, ) : /
1
/
1
is a contraction map in the L metric.
(ii) There exists a unique /
1
such that (o, ) = . If /
1
then
(o, )
p
() is the L metric, and hence in the topology of convergence
with respect to each compactly supported continuous function.
FRACTALS AND SELF SIMILARITY 17
Proof. (ii) follows immediately from (i).
To establish (i), suppose Lip 1 and let r = max
1iN
r
i
. Then for ,
/
1
,
(o, )()() (o, )()() =
N

i=1
(
i
S
i#
)()
N

i=1
(
i
S
i#
)()
=
N

i=1

i
(( S
i
) ( S
i
))
=
N

i=1

i
r((r
1
S
i
) (r
1
S
i
))

i=1

i
rL(, ) = rL(, ),
since Lip (r
1
S
i
) r
1
1 r
i
1.
(2) Denition. The unique measure invariant with respect to (o, ) is denoted
|S, |.
(3) Denition. Let be the product measure on ((N) induced by the measure
(i) =
i
on each factor 1, . . . , N.
(4) Theorem.
(i) |o, | =
#
, where : ((N) K is the coordinate map of 3.1(3)(vii).
(ii) spt |o, | = [o[.
Proof. (ii) follows from (i) and 3.1(3)(vii).
To establish (i) let
i
: C(N) C(N) be the ith shift operator 2.1(3). Clearly

i
= S
i
and is (
1
, . . . ,
N
, ) invariant. Hence

i
S
i#
(
#
) =

#
(
i#
) =
#

i
(
i#
) =
#
, and so by uniqueness
#
= |o, |.
(5) Remarks.
(1) It follows from 4(ii) that |o, | has compact support.
(2) There are unbounded invariant measures, in particular and trivially, L
n
on
R
n
.
(3) If is invariant, so is for any positive constant . Requiring |o, | /
1
is simply a normalisation requirement.
(6) Example. Referring back to 3.3(1), let =
1
2
,
1
2
and write
r
for |o
r
, |.
We will show
r
,=
s
for
1
2
r < s < 1. In 5.3(iii) we see that
r
= H
r
|o
r
| for
0 < r
1
2
.
Suppose
1
2
r < s < 1. Take A [0, 1 s). Then S
2
(r)
1
(A) [0, 1] = and
hence
r
(S
2
(r)
1
(A)) = since spt
r
[0, 1]. It follows
r
(A) =
1
2

r
(S
1
(r)
1
(A)) =
1
2

r
(rA). Similarly
s
(A) =
1
2

s
(sA). If
r
=
s
= , say, it follows (sA rA) =
0. Choosing A = [0, 1 s), this contradicts spt = [0, 1].
4.5. Dierent Sets of Similitudes Generating the Same Set. For I a nite
set as in 2.1(6), let o
I
= S

: I. Then [o
I
[ = k

:

I, as follows by
applying 3.1(3)(iii), (iv) to o
I
. From 2.1(8)(i),

I = I i I is secure. Thus [o
I
[ = [o[
if I is secure, and if the coordinate map is one-one, then [o
I
[ = [o[ i I is secure.
A similar result was rst shown in [OP].
Let
I
: I (0, 1) be given by
I
(i
1
, . . . , i
p
)) = (i
1
) . . . (i
p
). Then if
I is tight, by using 2.1(8)(ii), one can check

I

I
() = 1, and furthermore
18 JOHN E. HUTCHINSON
|o
I
,
I
| = |o, | as follows from 4.4(4) and 3.1(3). Finally, if is one-one, then
|o
I
,
I
| = |o, | i I is tight.
5. Similitudes
5.1. Self-Similar Sets. We continue the notation of 3 and 4.
(1) Denition. K is self-similar (with respect to o) if
(1) K is invariant with respect to o, and
(2) H
k
(K) > 0, H
k
(K
i
K
j
) = 0 for i ,= j, where k = dimK.
Thus (ii) is a kind of minimal overlap condition, and rules out the examples in
3.3(1) with
1
2
< r < 1. However, it is still rather weak. For example, if o = S
1
, S
2
,
S
i
: R
2
R
2
, S
1
(re
i
) = (

2)
1
re
i(/4)
and S
2
(1+re
i
) = 1+(

2)
1
re
i(+/4)
,
then [o[ is a continuous image of [0, 1] with a dense set of self-intersections, see [OP,
D=2, L=0] or [LP, Figure 3]. In [LP, page 374] it is shown H
2
([o[) =
1
4
, and so
dim[o[ = 2 and [o[ is self-similar in the above sense by (4)(ii).
A more useful notion of self-similarity may be a condition analogous to the open
set condition below, which we will see implies [o[ is self-similar, but allows us to
separate out the components [o[
i
.
(2) Convention. For the rest of this section we restrict to the case o is a
family of similitudes. (X, d) will be R
n
with the Euclidean metric, although other
conditions suce. H
k
is Hausdor k-dimensional measure. Lip (S
i
) = r
i
.
Let (t) =

N
i=1
r
t
i
. Then (0) = N and (t) 0 as t , and hence there is
a unique D such that

N
i=1
r
D
i
= 1.
(3) Denition. If

r
D
i
= 1, D is called the similarity dimension of o.
We will see in 5.3(1) that D often equals the Hausdor dimension of [o[.
For the rest of this section =
1
, . . . ,
N
where
i
= r
D
i
, and so is de-
termined by o. We write |o| for |o, |, and often write K for [o[ and for
|o|.
Note that

i1,...,ip
r
D
i1
. . . r
D
i1
=
_

N
i=1
r
D
i
_
p
= 1, a fact we use frequently.
Finally we take r
1
r
2
r
N
, so that r
1
= min r
i
: 1 i N,
r
N
= max r
i
: 1 i N.
(4) Proposition. Let K = [o[, dim K = k. Then
(i) H
D
(K) < and so k D (this is true for arbitrary contractions S
i
).
(ii) 0 < H
k
(K) < implies (K is self-similar i k = D).
Proof. (i) K =

i1,...,ip
K
i1...ip
and

i1,...,ip
(diam K
i1...ip
)
D
=

i1,...,ip
r
D
i1
. . .
r
D
ip
(diam K)
D
= (diam K)
D
. Since diam K
i1...ip
r
p
N
diam K 0 as p , we
are done.
(ii) Suppose 0 < H
k
(K) < and K is self-similar, so that H
k
(K
i
K
j
) = 0
if i ,= j. Then H
k
(K) =

N
i=1
H
k
(K
i
) =

N
i=1
r
k
i
H
k
(K), hence

r
k
i
= 1, hence
D = k.
Conversely, suppose 0 < H
D
(K) < . Then H
D
(K)

N
i=1
H
D
(K
i
) =

N
i=1
r
D
i
H
D
(K). Since

N
i=1
r
D
i
= 1, H
D
(K) =

N
i=1
H
D
(K
i
) and so it is standard
measure theory that H
D
(K
i
K
j
) = 0 if i ,= j.
5.2. Open Set Condition. Recall convention 5.1(2).
(1) Denition. o satises the open set condition if there exists a non-empty
open set O such that
(i)

N
i=1
S
i
O O,
(ii) S
i
O S
j
O = if i ,= j.
FRACTALS AND SELF SIMILARITY 19
(2) Examples. (a) Suppose we already have a non-empty closed set C satis-
fying (i) and (ii) of (1) with O replaced by C. Let d = min
i=j
d(S
i
C, S
j
C), and
select so r
i
< d/2 for i = 1, . . . , N. Then o satises the open set condition with
O =

xC
B(x, ). To see this observe that
S
i
O =
_
xC
S
i
B(x, ) =
_
xC
B(S
i
x, r
i
) =
_
ySiC
B(y, r
i
).
Hence S
i
O S
j
O = if i ,= j and furthermore S
i
O O. This situation applies to
3.3(1), 0 < r <
1
2
, with the non-empty closed set [0, 1].
(b) Suppose there is a closed set C with non-empty interior such that
(1) S
i
C C if i = 1, . . . , N,
(2) (S
i
C)

(S
j
C)

= if i ,= j.
Then the open set condition holds with O = C

. This situation applies in 3.3(1),

(2) with C the closed convex hull of [o[, i.e. C = [0, 1] for 3.3(1) and C is the
triangle (a
1
, a
5
, a
3
) for 3.3(2).
(3) Elementary consequences. Suppose S satises the open set condition
with O. Note that S
i1...ip
commutes with the topological operators

,

, ,
c
. In
particular (O
i1...ip
)

= (O

)
i1...ip
, and so we can write

O
i1...ip
unambiguously.
Then
(i) O O
i1
O
i1i2
O
i1i2...ip
;
(ii) K
i1...ip


O
i1...ip
;
(iii) K
j1...jp
O
i1...ip
= if (j
1
, . . . , j
p
) ,= (i
1
, . . . , i
p
);
(iv) if I is tight (2.1(7)), then the O

, I, are mutually disjoint.

Thus (ii) and (iii) say that O
i1...ip
isolates K
i1...ip
from the K
j1...jp
for (j
1
, . . . , j
p
) ,=
(i
1
, . . . , i
p
).
Proof. (i) and (ii) follow immediately from 3.1(8).
For (iii), suppose (j
1
, . . . , j
p
) ,= (i
1
, . . . , i
p
). But K
j1...jp


O
j1...jp
, and

O
j1...jp

O
i1...ip
= since O
j1...jp
O
i1...ip
= .
For (iv), suppose I is tight, , I, and ,= . Let p be the greatest integer
(perhaps 0) for which there is a sequence i
1
, . . . , i
p
) with i
1
, . . . , i
p
) and
i
1
, . . . , i
p
) . Since I is tight there exist i
p+1
,= j
p+1
such that i
1
, . . . , i
p
, i
p+1
)
, i
1
, . . . , i
p
, j
p+1
) . But then O

O
i1...ipip+1
, O

O
i...ipjp+1
by (1), and so
O

S
i1...ip
(O
ip+1
O
jp+1
) = .

5.3. Existence of Self Similar Sets.

(1) Theorem. Suppose o satises the open set condition. Then
(i) there exist
1
,
2
such that
0 <
1

D

(K, k)
D
(K, k)
2
< for all k K,
(ii) 0 < H
D
(K) < and so K is self-similar by 5.1(4)(ii). In particular dim
K = D,
(iii) |o| = [H
D
(K)]
1
H
D
K.
Proof.
(a) Lemma. Suppose 0 < c
1
< c
2
< and 0 < < . Let U
i
be a family of
disjoint open sets. Suppose each U
i
contains a ball of radius c
1
and is contained
in a ball of radius c
2
. Then at most (1 + 2c
2
)
n
c
n
1
of the

U
i
meet B(0, ).
20 JOHN E. HUTCHINSON
For suppose

U
i
, . . . ,

U
k
meet B(0, ). Then each of

U
i
, . . . ,

U
k
is a subset of
B(0, (1 + 2c
2
)). Summing the volumes of the k corresponding disjoint spheres of
radius c
1
, we see that
k
n

n
c
n
1

n
(1 + 2c
2
)
n

n
,
and hence k (1 + 2c
2
)
n
c
n
1
.
(b) For the rest of the proof let O be the open set asserted to exist by the
open set condition.
Let = |o|. We will rst prove that there exist constants
1
,
2
such that
0 <
1

D

(, k)
D
(, k)
2
<
for all k K.
First note that
(K
i1...ip
)
i1...ip
(K
i1...ip
) = r
D
i1
. . . r
D
ip
(S
1
i1...ip
K
i1...ip
)
= r
D
i1
. . . r
D
ip
(K) = r
D
i1
. . . r
D
ip
.
Let k = k
i1...ip...
and consider B(k, ). Choose the least p such that K
i1...ip

B(k, ). Then r
i1
. . . r
ip
(diam K) r
1
(recalling r
1
r
N
). Hence
B(k, )

D

(K
i1...i
)

D

r
D
i1
. . . r
D
ip

D

r
D
1

D
(diamK)
D
.
Hence
D

(, k) r
D
1

1
D
(diamK)
D
for k K.
We now show that
D
(, k) is uniformly bounded away from for k K.
Suppose O contains a ball of radius c
1
and is contained in a ball of radius c
2
. For
each sequence j
1
. . . j
q
. . . C(N) select the least q such that r
1
r
j1
. . . r
jq
.
Let I be the set of j
1
, . . . , j
q
) thus selected, and notice that I is tight (2.1(7)).
From 5.2(3) it follows O
j1...jq
: j
1
, . . . , j
q
) I is a collection of disjoint open
sets. Moreover, each such O
j1...jq
contains a ball of radius r
j1
. . . r
jq
c
1
and hence
of radius r
1
c
1
and is contained in a ball of radius r
j1
. . . r
jq
c
2
and hence of
radius c
2
. It follows from (a) that at most (1 + 2c
2
)
n
(r
1
c
1
)
n
of the

O
j1...jq
,
j
1
, . . . , j
q
) I, meet B(k, ). Hence at most (1 + 2c
2
)
n
(r
1
c
1
)
n
of the K
j1...jq
,
j
1
, . . . , j
q
) I, meet B(k, ).
Now spt(
j1...jq
) = K
j1...jq
by 4.4(4)(ii). By 4.5
=

j1,...,jqI

j1...jq
.
Finally M(
j1...jq
) = r
D
j1
. . . r
D
jq

D
for j
1
, . . . , j
q
) I.
Hence
B(k, )

D

(1 + 2c
2
)
n
r
n
1
c
n
1

D
=
(1 + 2c
2
)
n

D
r
n
1
c
n
1
.
It follows
D
(, k) (1 + 2c
2
)
n
(
D
r
n
1
c
n
1
)
1
.
(c) (ii) now follows from 2.6(3).
(d) Since K is self-similar, H
D
(K
i
K
J
) = 0 if i ,= j, and so
H
D
K =
N

i=1
H
D
K
i
=
N

i=1
r
D
i
S
i#
(H
D
K)
by 2.6(2).
Letting = [H
D
(K)]
1
H
D
K, it follows that =

N
i=1
r
D
i
S
i#
(), and that
M() = 1. By uniqueness, = , proving (iii).
(e) From (iii),
D

(K, k) =
D

(H
D
K, k) = [H
D
(K)]
1

(, k), and similarly

for
D
. (i) now follows from (b).
FRACTALS AND SELF SIMILARITY 21
(2) Remarks. Result (i) says that K is rather uniformly spread out in the
dimension k. But on the other hand, by a result of Marstrand [MJ], the inequality
between the upper and lower densities cannot be replaced by an equality if D is
non-integral.
Result (ii) is due to Moran [MP].
5.4. Purely Unrectiable Sets. We continue our assumptions that (X, d) is R
n
with the Euclidean metric and that o is a family of similitudes.
(1) Theorem. Suppose o satises the open set condition with both the open set
O and the open set U, where O U. Suppose furthermore that whenever A is an
m-dimensional ane subspace of R
n
for which A

O
i
,= and A

O
j
,= for some
i ,= j, then A
_
U

N
i=1

O
i
_
,= . Then for any m-dimensional C
1
manifold M
in R
n
, H
m
(M [o[) = 0.
Proof. We proceed in stages.
(a) Let
/ = A : A is an m-dimensional ane space with
A

O
i
,= , A

0
j
,= for some i ,= j.
(see Figure 5.1.)
Figure 5.1
Let g : / (0, ) be dened by
g(A) = sup
_
r : B(a, r) U
N
_
i=1

O
i
for some a A
_
.
By the hypotheses of the theorem, 0 < g(A) < . We want to show that:
g is uniformly bounded away from 0.
To do this we dene a topology on / and prove that / is compact and g is lower
semi-continuous (i.e. g(A
0
) > implies g(A) > for all A suciently close to A
0
).
The required result then follows.
Let O
m
be the set of m-dimensional subspaces through the origin, give O
m
its
usual compact topology as a subset of R
n
2
, and let R
n
O
m
have the product
22 JOHN E. HUTCHINSON
topology. The map (a, O) a + O is a map from R
n
O
m
onto the set of m-
dimensional ane spaces in R
n
, we give this set the induced topology. Since / is
the image of a closed bounded (hence compact) set, it is compact.
Next suppose g(A
0
) > , A
0
/. Select a
0
A
0
and
0
> such that
B(a
0
,
0
) U

N
i=1

O
i
. For all A suciently close to A
0
, d(a
0
, A) <
0
.
Select a A such that d(a
0
, a) <
0
. Then B(a, ) B(a
0
,
0
) U

N
i=1

O
i
.
Hence g(A) > . Thus g is lower semi-continuous. The required result follows.
(b) For each > 0, let
(

=
_
C : C is an m-dimensional C
1
manifold in R
n
, and for some
A / there exists a C
1
map f : A U C such that
(i) f is one-one,
(ii) Lip f 1 +, Lip f
1
1 +,
(iii) d(f(x), x) < for all x A U
_
.
We will show:
there exist > 0, > 0 such that
H
m
_
C
_
U
N
_
i=1

O
i
_
_
> for all C (

.
Suppose by (a) that g(A) > > 0 for all A /. Fix so 0 < < /3.
Suppose C (

with f and A as in the denition of (

. Select a A such that

B(a, ) U

N
i=1

O
i
.
Since d(f(a), a) < it follows B(f(a), ) B(a, ) and hence B(f(a), )
U

N
i=1

O
i
. But one can check that f(A B(a, 3)) C B(f(a), ). It
follows that
H
m
_
C
_
U
N
_
i=1

O
i
_
_

_
AB(a,3)
J(f) dH
m

m
( 3)
m
(1 +)
m
,
where J(f) is the Jacobian of f. This gives the required result.
(c) Now assume that the hypotheses of the theorem hold and that H
m
(M
K) ,= 0 for some C
1
manifold M, where K = [o[. We will deduce a contradiction.
First note that
m
(M K, k) = 1 for some k M K [FH, 3.2.19] since M K
is (H
m
, m)-rectiable. Alternatively the corresponding result for
m
in R
m
[MM,
page 184] can readily be lifted back to the manifold M be means of the area formula
H
m
(f(A)) =
_
A
J(f)dL
m
for C
1
dieomorphisms f : A R
n
, A R
m
.
In the following we will need to be a little careful, since due to overlap it is
possible that for xed p, an arbitrary member of K may belong to more than one
K
i1...ip
.
Since k is a point of non-zero m-dimensional density for M K, it follows that
there is a sequence k
j
k as j , k ,= k
j
MK. By passing to a subsequence
we may suppose all k
j
K
i1
for some i
1
, which we x. By passing to a subsequence
again we may suppose all k
j
K
i1i2
for some i
2
which we also x. Repeating this
argument and then diagonalising, we extract a subsequence k
j
k and a sequence
i
1
. . . i
j
. . . , such that k
j
K
i1...ij
for all j. Moreover k = k
i1...ij...
.
For each j let p(j) be the least integer p such that k
j
K
i1...ip
, k
j
/ K
i1...ipip+1
,
and notice p(j) j, so p(j) as j .
Now
m
(M K, k) = 1, and
m
(M, k) = 1 since M is a C
1
manifold, hence

m
(M K, k) = 0. Select R diam U, so that diam U
i1...i
p(j)
< Rr
i1
. . . r
i
p(j)
.
FRACTALS AND SELF SIMILARITY 23
We have
() lim
j
H
m
_
(M K) B(k, Rr
i1
. . . r
i
p(j)
)
_

m
(Rr
i1
. . . r
i
p(j)
)
m
= 0.
To simplify notation we write p for p(j). See Figure 5.2.
Figure 5.2
Now
(M K) B(k, Rr
i1
. . . r
ip
)
(M K) U
i1...ip
= M (U
i1...ip
K)
= M (U
i1...ip
K
i1...ip
) by 5.2(3)(iii)
M (U
i1...ip

N
_
=1

O
i1...i
p

) by 5.2(3)(ii).
Hence
()
H
m
_
(M K) B(k, Rr
i1
. . . r
i
p(j)
)
_

m
(Rr
i1
. . . r
i
p(j)
)
m

H
m
_
M
_
U
i1...i
p(j)

N
=1

O
i1...i
p(j)

_
_

m
(Rr
i1
. . . r
i
p(j)
)
m
=
H
m
_
f
j
(M)
_
U

N
=1

_
_

m
R
m
,
where f
j
= S
1
i
p(j)
S
1
i1
is an explosion map. Here we are using the fact that
f
j
(U
i1...i
p(j)
) = U, f
j
(

O
i1...i
p(j)

) =

O

, and H
m
(f
j
(A)) = r
m
i
p(j)
. . . r
m
i1
H
m
(A)
for arbitrary A.
But for suciently large j we will show that f
j
(M) (

. From (b) this shows the

expression in () is bounded away from 0 for all suciently large j, contradicting
(). Thus our original assumption that H
m
(M K) ,= 0 is false.
To see that f
j
(M) (

for all suciently large j we rst observe that in analogy

with the denition of (

, we have for all suciently large j a C

1
map g
j
: f
j
(T)V
f
j
(M) where
(i) V is a xed open neighbourhood of

U,
(ii) T is the tangent plane to M at k,
(iii) g
j
is one-one,
(iv) Lip g
j
1 +, Lip g
1
j
1 +,
24 JOHN E. HUTCHINSON
(v) d(g
j
(x), x) < /2 for all x f
j
(T) V .
See for example [FH, 3.1.23]. But we do not know if f
j
M /. However, we
can select an ane space A
j
through k and k
j
such that for all suciently large
j, f
j
(A
j
) U and f
j
(T) V are arbitrarily close in the topology on ane spaces
introduced in (a). In particular there will exist
j
: f
j
(A
j
) U f
j
(T) V such
that Lip
j
= Lip
1
j
= 1 and d(
j
(x), x) < /2 for all x f
j
(A
j
) U.
Notice that f
j
(A
j
) /, since f
j
(k) f
j
(A
j
) K
i
p(j)+1
A

O
ip
and f
j
(k
j
)
f
j
(A
j
) K

A

O

for some ,= i
p(j)+1
.
Finally f
j
(M) (

, where in the denition of (

, A is replaced by f
j
(A
j
), f is
replaced by g
j

j
, and C is replaced by f
j
(M).
(2) Remark. In the terminology of [FH, 3.2.14], K is purely (H
m
, m) unrec-
tiable. In this respect the interest of the present theorem lies in the fact that it
establishes pure (H
m
, m) unrectiability for sets K such that H
m
(K) = (pro-
vided m < D). Unlike the examples in [FH, 3.3.19, 3.3.20] one cannot argue by
using the structure theorems for sets having nite H
m
measure.
(3) Example.
(a) If K = UK
i
with the K
i
disjoint, then the hypotheses of the theorem are
easily seen to be satised if m = 1, where O is as in 5.2(2)(a), and U = O. For
A
_
U

N
i=1

O
i
_
= i A
_
U

N
i=1

O
i
_
c
= U
c

N
i=1

O
i
. But this latter
cannot be true if A /, since then A can be split into two disjoint non-empty
components A U
c
and A

N
i=1

O
i
.
(b) From Example 3.3(2) let O be the interior of the triangle (a
1
, a
5
, a
3
). Let
U O be a slightly larger open set also satisfying the open set condition and such
that U O = a
1
, a
5
. Suppose A / where m = 1. If A
_
U

4
i=1

O
i
_
=
then it is straightforward to show a
1
, a
5
A. But then (a
2
, a
4
) A, which
contradicts A
_
U

4
i=1

0
i
_
= since (a
2
, a
4
) U

N
i=1

O
i
.
(4) Remark. Let us strengthen the hypotheses in (1) by taking A to be a one-
dimensional ane subspace. Then Mattila [MaP] has shown the existence of an
> 0, depending only on o, such that for any m-dimensional C
1
manifold M, dim
(M [o[) m.
In the same paper, Mattila also shows that under the hypotheses of 5.3(1), if
m D, then there are only two possibilities; either K lies in an m-dimensional
ane subspace or H
D
([o[ M) = 0 for every m-dimensional C
1
manifold M.
5.5. Parameter Space. The orthogonal group O(n) of orthonormal transforma-
tions of R
n
is an n(n 1)/2 dimensional manifold [FH, 3.2.28(1)], and hence the
set of similitudes S = (a, r, O) in R
n
corresponds to an n(n 1)/2 + (n + 1) =
(n
2
+n + 2)/2 dimensional manifold. Thus every invariant set generated by some
o = S
1
, . . . , S
N
of similitudes in R
n
corresponds to a point in an N(n
2
+n+2)/2
dimensional manifold, which we call the parameter space. Oppenheimer [OP] has
made a systematic computer analysis of a part of N = n = 2.
6. Integral Flat Chains
In this section we will see how integral at chains, which will not normally be
rectiable, arise naturally in the context of self-similarity. In particular, the Koch
curve of 3.3(2) supports a 1-dimensional integral at chain in a natural way, and
[o[ in Example 3.3(3) supports a 2-dimensional integral at chain provides D < 3
(with D dened in 6.2(2)).
We make the convention that all currents we consider are integral at chains, or
chains for short.
We need to introduce a new metric, but rst we need a lemma on the T-metric.
FRACTALS AND SELF SIMILARITY 25
6.1. The T-metric.
(1) Lemma. Suppose 1 m n1, T is a (not necessarily rectiable) m-cycle,
and = (m, n) is the isoperimetric constant of 2.7(5).
(i) If T(T) <
m
, T = A+R, and T(T) = M(A) +M(R), then R = 0.
(ii) If T(T)
m
, then T = A for some A such that M(A) = T(T).
Proof. (i) We rst remark that any T can be written as T = A+R with T(T) =
M(A) +M(R), as noted in 2.7(5).
Assume the hypotheses of (i). If R = 0 we are done, so suppose R ,= 0. Since
R = T = 0, and M(R) T(T) <
m
, there is an m-cycle D such that
R = D and M(D) [M(R)]
m+1/m
by 2.7(5). Hence M(D) < M(R), since
[M(R)]
1/m
[T(T)]
1/m
<
1
. But then T = (A + D) and M(A + D)
M(A) +M(D) < M(A) +M(R) = T(T), a contradiction.
(ii) Suppose T(T)
m
and let T = A + R with T(T) = M(A) + M(R).
Then the same argument as for (i) shows that R = D with M(D) M(R). Hence
T = (A + D) and M(A + D) M(A) + M(D) M(A) + M(R) = T(T). Thus
M(A+D) = T(T), and so we are done.
(2) We see the necessity of the condition T(T)
m
in the following exam-
ple. Let T
r
be a 1-cycle supported on x : [x[ = r R
2
with M(T
r
) = 2r. Let
A
r
be the rectiable 2-current supported on x : [x[ r R
2
such that A
r
= T
and M(A
r
) = r
2
. Using the fact (1, 2) = 4 [FH 4.5.14] and the constancy
theorem [FH 4.1.7], one can show that if r 2, then r
2
= M(A
r
) = T(T
r
). If
r > 2, then again by [FH 4.1.7] C = T
r
implies C = A
r
and so M(C) = r
2
(unless we allow C to have non-bounded support, in which case M(C) = ). But
r
2
> 2r = M(T
r
) T(T
r
).
6.2. The (-metric.
(1) Denition. Let B be an (m 1)-boundary, m 1. Then (
B
is the set of
m-chains given by
(
B
= R T
m
: R = B.
(2) Denition. For R, S (
B
let
((R, S) = infM(A) : A
m+1
, R S = A.
We now see that ( is a complete metric on (
B
with the useful transformation
result (3)(i).
(3) Lemma. Let B be an (m1)-boundary, m 1.
(i) If f : R
n
R
n
is a proper Lipschitz map and Lip f = r, then ((f
#
R, f
#
S)
r
m+1
((R, S).
(ii) T(R, S) ((R, S), and T(R, S) = ((R, S) if T(R, S)
m
.
(iii) ( is a complete metric on (
B
. The (- and T-topologies agree on (
B
.
(iv) The inmum in the denition of ((R, S) is realised for some A
m+1
.
Proof. (i) is immediate from 2.7(6)(c).
The rst assertion in (ii) is immediate, and the second follows from 6.1(1)(i).
To see that ((R, S) < let T(R, S) = < and by 2.7(6)(c) choose f =
r
such that T(
r#
R,
r#
S)
m
(
r
is dened in 2.3). Then ((
r#
R,
r#
S)

m
and so by (i) ((R, S) r
(m+1)

m
. The other properties of a metric are
easily veried, noting in particular that if ((R, S) = 0 then T(R, S) = 0 and so
R = S.
The (- and T-topologies clearly agree on (
B
. Since a sequence is (
B
-Cauchy i
it is T-Cauchy, (
B
is closed in T
m
in the T-metric, and T is a complete metric on
T
m
, it follows ( is a complete metric on (
B
.
26 JOHN E. HUTCHINSON
To prove (iv) suppose ((R, S) = and let T
j

m+1
, T
j
= RS, M(T
j
) .
Let B(0, r) be some ball large enough to include spt (R S), and let f : R
n

B(0, r) be a retraction map with Lipschitz constant 1. Let T

j
= f
#
T
j
. Then
spt T

j
B(0, r) and M(T

j
) M(T
j
) by 2.7(6). We can apply the compactness
theorem of 2.7(5) to T

j
T

1
and extract a convergent subsequence with limit AT

1
,
say. From 2.7(6) it follows M(A) and hence M(A) = . Furthermore A =
R S, and so A is the required current.
6.3. Invariant Chains. Suppose o = S
1
, . . . , S
N
are proper contraction maps
on R
n
, not necessarily similitudes.
(1) Denition. For any k-chain T, we let o(T) =

N
i=1
S
i#
(T); also o
0
(T) =
T, o
1
(T) = o(T), o
p+1
(T) = S(S
p
(T)) if p 1.
From 2.7(6), o : T
k
T
k
and is a continuous linear operator which commutes
with .
(2) Suppose now that Lip S
i
= r
i
, and let D be specied by

N
i=1
r
D
i
= 1
as in 5.1(3), but recall that here the S
i
are not necessarily similitudes. Let m
be the unique integer given by m D < m + 1. Now suppose m 1, B is an
(m 1)-boundary, and o(B) = B. As examples consider 3.3(2) with m = 1 and
B = [[(1, 0)]][[(0, 0)]], or 3.3(3) with m = 2 and B = N. Finally let =

N
i=1
r
m+1
i
and note that < 1.
(3) Theorem. Under the hypotheses of (2) the following hold
(i) o is a contraction map on (
B
in the (-metric.
(ii) There is a unique m-chain T (
B
such that o(T) = T.
(iii) If R (
B
, then o
p
(R) T in the T-metric (and (-metric).
(iv) If R (
B
and o(R) R = A with a
m+1
(which is always possible by
6.2(3)) then A
0
=

p=0
o
p
(A)
m+1
with convergence in the M-norm,
and T = R +A
0
.
Proof. First note that for any D T
m+1
,
M(o(D)) = M
N

i=1
S
i#
D
N

i=1
M(S
i#
D)

i=1
r
m+1
i
M(D) (by 2.7(6)(c)) = M(D).
We now show (i). If R (
B
then o(R) (
B
since (o(R)) = o(R) = o(B) =
B. Next suppose by 6.2(3)(iv) that R
1
R
2
= C with C
B
(R
1
, R
2
) = M(C). Then
o(R
1
R
2
) = o(C) = (o(C)),
and M(o(C)) M(C). Hence (
B
(o(R
1
), o(R
2
)) (
B
(R
1
, R
2
).
(ii) and (iii) follow immediately, using 6.2(3).
To establish (iv), suppose R (
B
, o(R) R = A, A
m+1
. Then
M(o
p
(A)) M(o
p1
(A)) and so M(o
p
(A))
p
M(A). Thus A
0
=

p=0
o
p
(A)
converges in the M-norm. Thus A
0
is a chain of nite mass and hence rectiable
by 2.7(4). Finally
A
0
=

p=0
o
p
(A) =

p=0
o
p
(A) =

p=0
o
p
(o(R) R) = lim
p
(o
p
(R) R) = T R.

(4) We can often take particularly simple chains for R and A in (3)(iv). For
example in 3.3(2) we can take R = [[a
1
, a
5
]] and A = [[a
2
, a
3
, a
4
]] to be the obvious
oriented simplices [FH, 4.1.11].
FRACTALS AND SELF SIMILARITY 27
(5) Again taking the hypotheses of (2), let T (
B
be given by (3). Now spt
T = spt o(T)

N
i=1
spt S
i
#T

N
i=1
S
i
(spt T) = o(spt T). Thus o
p
(spt T)
as p , and since the limit in the Hausdor metric is [o[ it follows spt T [o[.
It is easy to construct examples, where due to cancellation, spt T [o[.
References
[AF] F. J. Almgren, Jr., Existence and regularity almost everywhere of solutions to elliptic
variational problems among surfaces of varying topological type and singularity structure,
Ann. of Math. 87 (1968), 321391.
[FH] H. Federer, Geometric Measure Theory, Springer-Verlag, New York, 1969.
[FH1] H. Federer, Colloquium lectures on geometric measure theory, Bull. Am. Math. Soc 84
(1978), 291338.
[LP] P. Levy, Les courbes planes ou gauches et les surfaces composees de parties semblables au
tout, Journal de lEcole Polytechnique Serie III 7-8 (1938), 227-291. Reprinted in Oeuvres
de Paul Levy Vol. II, (D. Dugue, P. Deheuvels & M. Ibero, Eds.) pp. 331394, Gauthier
Villars, Paris, Bruxelles and Montreal, 1973.
[MB] B. Mandelbrot, Fractals, Form, Chance, and Dimension, Freeman, San Francisco, 1977.
[MJ] J. M. Marstrand, The (, s) regular subsets of n space, Trans. Am. Math. Soc. 113 (1964),
369392.
[MM] M. E. Munroe, Measure and Integration, 2nd edition, Addison-Wesley, Reading, MA, 1971.
[MP] P. A. P. Moran, Additive functions of intervals and Hausdor measure, Proc. Camb. Phil.
Soc. 42 (1946), 1523.
[MaP] P. Matilla, On the structure of self-similar fractals, (preprint).
[OP] P. E. Oppenheimer, Constructing an Atlas of Self Similar Sets, B.A. Thesis, Princeton,
1979.
The Australian National University, Canberra, ACT, 2600, Australia
Received March 3, 1980; Revised October 16, 1980