Math. Nachr. 280, No. 12, 152 170 (2007) / DOI 10.1002/mana.

200410471
A notion of Euler characteristic for fractals
Marta Llorente
1
and Steffen Winter
2
1
Universidad Aut onoma de Madrid, Facultad de Ciencias Econ omicas y Empresariales, Departamento de An alisis
Econ omico: Economa Cuantitativa, Campus de Cantoblanco, Carretera de Colmenar no. 16, 28046 Madrid, Spain
2
Mathematisches Institut, Friedrich-Schiller-Universit at Jena, 07737 Jena, Germany
Received 11 May 2004, revised 8 April 2005, accepted 19 May 2005
Published online 19 December 2006
Key words Fractal, self-similar set, Euler characteristic, parallel set, Renewal theorem, convex ring
MSC (2000) Primary: 28A80; Secondary: 60K05, 28A75
A notion of (average) fractal Euler number for subsets of R
d
with innite singular complexes is introduced by
means of rescaled Euler numbers of innitesimal -neighbourhoods. For certain classes of self-similar sets we
calculate the associated Euler exponent and the (average) fractal Euler number with the help of the Renewal
theorem. Examples like the Sierpinski gasket or carpet are provided.
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
1 Introduction
The Euler characteristic of a nite cell complex K in the Euclidean space R
d
is given by
(K) =
d

k=0
(1)
k

k
where
k
denotes the number of k-cells of K. If
k
is the k-th Betti number of K, i.e., the rank of the k-th
homology group of K, the EulerPoincar e formula
(K) =
d

k=0
(1)
k

k
shows that (K) does not depend on the special decomposition of the complex K (for more details see, e.g. Dold
[1]). In this way the Euler characteristic is a homological, homotopic and topological invariant. For appropriate
classes of sets from convex geometry, differential geometry and geometric measure theory is well-dened and
special calculation methods are known (for one of them compare Section 4). For fractal subsets of R
d
, however,
the associated complexes are typically innite and the Euler number in the classical sense is not determined.
Until now, very little is known about the topological structure of self-similar sets, which have been extensively
studied in other respects (see, for example, [2], [10] and the references therein). In the present paper we introduce
the notion of fractal Euler number for some classes of self-similar sets. In particular, it allows to distinguish self-
similar sets with equal fractal dimensions by their different fractal Euler numbers (see Subsection 2.3).
The main idea is the following: for nice classical singular sets M like compact Lipschitz submanifolds the
Euler number of parallel sets M

coincides with that of M for sufciently small > 0 and may be calculated
by a generalized GaussBonnet formula (cf. [12]). It turns out that for more general compact sets F R
d
the
-neighbourhoods
F

:=
_
x R
d
: inf
yF
d(x, y)
_

Corresponding author: e-mail: m.llorente@uam.es, Phone: +34 91 497 2877, Fax: +34 91 497 2991

e-mail: winter@minet.uni-jena.de, Phone: +49 3641 946186, Fax: +49 3641 946186
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
Math. Nachr. 280, No. 12 (2007) 153
admit the classical Euler characteristic (F

). Here d(, ) denotes the Euclidean metric. We suppose now that

(F

) is determined for all > 0. In this case the following notions can be dened. Let b be the diameter of the
compact set F R
d
.
Denition 1.1 The Euler exponent of F is the number
= (F) := inf
_
t 0 :
t
|(F

)| is bounded as 0
_
.
Denition 1.2 If the limit

f
(F) := lim
0
_

b
_

(F

)
exists, then it is called the fractal Euler number of the set F.
The diameter b in the denition is a normalization to ensure scaling invariance of the limit (compare Corol-
lary 2.3). Unfortunately, the existence of such limits is a rare event. From the study of local quantities of
self-conformal sets like densities or tangent measure distributions it is well-known that, in general, average limits
provide better results. It turns out that this is a useful tool for our purposes, too.
Denition 1.3 If the limit

a
f
(F) := lim
0
1
|log |
_
1

b
_

(F

)
d

exists, then it is called the average fractal Euler number of the set F.
In both denitions we could work with upper and lower limits leading to the quantities
f
(F),
f
(F) and

a
f
(F),
a
f
(F), respectively, for all sets F R
d
with unbounded (F

) as 0. In the present paper,

however, we are interested in nding sufcient conditions for the existence of the (average) limits. We restrict
our considerations to the special case of self-similar sets F and assume that the Euler numbers of parallel sets
F

exist. Moreover, if S
1
, . . . , S
N
are the similarity mappings generating F and s its similarity dimension, we
assume that the overlap function
R() = (F

)
N

i=1
((S
i
F)

)
of the set F does not grow too fast as 0, i.e., there are constants c, > 0 such that
|R()| c
s
for all 0 < 1. (1.1)
This and the scaling behaviour of F admit the application of the Renewal theorem from probability theory to
our problem. Appropriate calculations show that in this situation typically the Euler exponent coincides with
the similarity dimension of F. Moreover,
a
f
(F) is calculated in terms of R and the contraction ratios of the
maps S
1
, . . . , S
N
. Under additional conditions on the contraction ratios also the fractal Euler limit
f
(F) exists
(see Theorem 2.1). More tractable sufcient conditions for the existence of (average) fractal Euler numbers are
discussed in the second part of the paper where we specify the results to the convex ring (Theorem 2.4). In this
setting direct calculations of limits are possible. Examples are provided. For self-similar sets on the real line
fractal Euler numbers essentially coincide with the rescaled limit of the classical notion of gap counting function
(cf. [2]). As a corollary of Theorem 2.1 we obtain the existence of average limits for the gap counting function.
The paper is organized as follows. In Section 2 the main results are formulated and illustrated with some
examples. First a general existence theorem of fractal Euler numbers for self-similar sets is discussed, while
in Subsection 2.2 we restrict to self-similar sets with neighbourhoods in the convex ring. After some examples
illustrating the concept and the calculation methods, in 2.4 we treat the special case of self-similar sets in R.
Section 3 is devoted to the proof of Theorem 2.1. For this purpose an appropriate version of the Renewal theorem
is presented and its consequences for average limits are discussed. In Section 4 we recall the notion of Euler
characteristic in the convex ring setting and discuss the consequences for neighbourhoods of self-similar sets to
be in the convex ring. In Section 5 Theorem 2.4 is proved and extensions to a larger class of self-similar sets are
discussed.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
154 Llorente and Winter: A notion of Euler characteristic for fractals
2 Main results and examples
2.1 Existence theorem
Let S
i
: R
d
R
d
, i = 1, . . . , N, be contracting similarities. Denote the contraction ratio of S
i
by r
i
(0, 1)
and let r
max
:= max
i=1,...,N
r
i
. It is a well-known fact in fractal geometry (cf. Hutchinson [5]), that for such a
system {S
1
, . . . , S
N
} of similarities there is a unique, nonempty, compact subset F of R
d
such that S(F) = F,
where S is the set mapping dened by
S(A) =
N
_
i=1
S
i
A, A R
d
.
F is called the self-similar set associated with {S
1
, . . . , S
N
}. Moreover, the unique solution s of

N
i=1
r
s
i
= 1
is called the similarity dimension of F. Throughout the paper, whenever we refer to a self-similar set, we keep
these notations. The system {S
1
, . . . , S
N
} is said to satisfy the open set condition (OSC) if there exists an open,
nonempty, bounded subset O R
d
with

i
S
i
O O and S
i
O S
j
O = for all i = j. {S
1
, . . . , S
N
} is said
to satisfy the strong separation condition (SSC) if S
i
F S
j
F = for all i = j.
To formulate our main result the following concept is needed. Let h > 0. A nite set of positive real numbers
{y
1
, . . . , y
N
} is called h-arithmetic if h is the largest number such that y
i
hZ for i = 1, . . . , N. If no such
number h exists for {y
1
, . . . , y
N
}, the set is called non-arithmetic.
From now on we assume that the Euler characteristic (F

) of the -neighbourhoods is dened for > 0, i.e.,

we restrict to those self-similar sets F for which all Betti numbers of the singular complex of the parallel sets
F

are nite for all > 0. Note that this assumption is the base of the whole concept. Without it the denitions
of fractal Euler numbers would not make much sense. In this situation the geometric invariance of the Euler
characteristic and the fact that
(S
i
F)

= S
i
(F
/ri
) (2.1)
imply that also ((S
i
F)

) is well-dened. This allows to dene the overlap function R : (0, +) Z of F by

R() = (F

)
N

i=1
((S
i
F)

). (2.2)
The name is motivated by the fact that R() is closely related to the set O() =

i=j
(S
i
F)

(S
j
F)

, called the
overlap of F

(compare Section 2.3 and [11]). It turns out that the overlap function is the crucial object to study
for deciding whether (average) fractal Euler numbers exist or not.
The following theorem gives sufcient conditions for the existence of the (average) fractal Euler number of
a self-similar set F. The stated formula shows that, in case it exists, the (average) fractal Euler number only
depends on the similarity dimension, the contraction ratios and the overlap function of F.
Theorem 2.1 Let F be a self-similar set with diameter b and similarity dimension s. Suppose that the Euler
numbers of F

are dened for all > 0. Furthermore, assume that the overlap function R has a discrete set of
discontinuities and satises condition (1.1). Then the Euler exponent of F satises s and the following
holds:
(i) The limit lim
0
1
|log |
_
1

b
_
s
(F

)
d

exists and equals

X =
1

__
1
0

s1
R(b) d +
N 1
s
_
, (2.3)
where =

N
i=1
r
s
i
log r
i
. If X = 0 then = s and the average fractal Euler number
a
f
(F) equals X.
(ii) If {log r
1
, . . . , log r
N
} is non-arithmetic, the limit lim
0
_

b
_
s
(F

) exists and equals X. In case

X = 0 the fractal Euler number
f
(F) exists and equals
a
f
(F).
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 155
Remark 2.2 Only in the case where X = 0 in Theorem 2.1, it can happen that < s which implies that
the considered limits for exponent s do not coincide with
a
f
(F) and
f
(F), respectively. Simple self-similar
sets, like the square, provide examples where the Euler exponent is zero and thus strictly less than its similarity
dimension (cf. Remark 4.5). However, it is not clear whether the case X = 0 does always correspond to self-
similar sets with < s or even = 0, or if X = 0 and = s is also possible. Further investigation is
needed.
The proof of Theorem 2.1 is given in Section 3. In general, the calculation of Euler numbers is difcult. Also,
condition (1.1) on the overlap function might be difcult to verify. In order to show that Theorem 2.1 allows
explicit calculations of fractal Euler numbers, at least for some classes of self-similar sets, we restrict the setting
to a subclass of cell complexes, namely to the convex ring.
Before doing this we point out some general properties of fractal Euler numbers. First observe that whenever
the fractal Euler number
f
(F) of a set F exists, then the average counterpart
a
f
(F) exists as well and coincides
with
f
(F). t is also immediate from the Denitions 1.2 and 1.3 that some invariance properties of the Euler
characteristic carry over to the limits if they exist, namely the scaling invariance, which is due to the normalization
constant b in the denitions, and the invariance with respect to Euclidean motions. Corollary 2.3 states the
invariance for
a
f
implying the same result for
f
. For > 0, dene F := {x : x F} and for an Euclidean
motion g let gF := {gx : x F}.
Corollary 2.3 (Scaling and motion invariance) Let F be a subset of R
d
. Assume that
a
f
(F) exists. Then,
for > 0 and g an Euclidean motion,
a
f
(F) and
a
f
(gF) exist and coincide with
a
f
(F).
Note that the above corollary is a direct consequence of the existence of the limit in Denition 1.3. It holds, in
particular, under the conditions of Theorem 2.1. We point out that the fractal Euler number is not a topological
invariant like its classical counterpart the Euler characteristic. It also takes the relative size of the structures in a
set into account.
2.2 Results for the convex ring
The convex ring R
d
is the family of all sets that are nite unions of compact convex subsets of R
d
. R
d
is closed
with respect to unions and intersections. Moreover, the Euler characteristic is well-dened for all elements of
R
d
, which are also refered to as polyconvex sets, and the following additivity property holds for A, B R
d
:
(A B) = (A) +(B) (A B). (2.4)
Convex ring and Euler characteristic in this class of sets are discussed in more detail in Section 4. Here we just
point out that the notion of Euler characteristic in the convex ring coincides with that for cell complexes for all
polyconvex sets. One of the advantages of this setting is that it allows to give sufcient geometric conditions easy
to verify for the existence of (average) fractal Euler numbers, as the following theorem shows. Let A denote the
(topological) boundary of a set A R
d
. Note that by A B we mean that A is a proper subset of B.
Theorem 2.4 Let F be a self-similar set. Assume there exists a set M R
d
such that M F M and
S
i
(M) S
j
(M) = S
i
(M) S
j
(M) (2.5)
for all i = j. Then the assertions of Theorem 2.1 hold.
Assume that for a self-similar set F and for some > 0, the -neighbourhood of F is polyconvex. It turns out
that this is equivalent to assuming that all its neighbourhoods are polyconvex (compare Proposition 4.6). This is
a consequence of the self-similarity of F and the fact that neighbourhoods of polyconvex sets are polyconvex.
Hence the Euler characteristic of neighbourhoods as well as the overlap function R of F are well-dened in
this situation. Moreover, it turns out that R has a discrete set of discontinuities (cf. Proposition 4.7). The only
possible accumulation point of discontinuities of R is 0. Thus the assumption F

R
d
for some > 0 already
implies some of the assumptions in Theorem 2.1. If we assume additionally the remaining one, namely that the
overlap function R satises condition (1.1), then the assertions of Theorem 2.1 hold. Therefore, the assumptions
in Theorem 2.1 simplify in the polyconvex setting.
In Theorem 2.4 we assume the existence of a polyconvex set M being a good approximation of the set F
and satisfying certain boundary conditions that prevent strong overlapping in F and make R bounded.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
156 Llorente and Winter: A notion of Euler characteristic for fractals
u

0 1

3
2
6
-
S
1
M S
2
M
S
3
M
M
w
1
w
2
w
3
0 1

3
2
6
-
Fig. 1 The Sierpinski gasket is the self-similar set generated by a system {S1, S2, S3} of similarities each mapping the
triangle M on the right-hand side to one of the smaller triangles SiM with contraction ratio
1
2
.
Sometimes, as for Sierpinski gasket and the Sierpinski carpet, the convex hull [F] of a self-similar set F is a
suitable set M for Theorem 2.4. Recall that the convex hull of a set K R
d
is dened as
[K] =

{C : K C, C convex}.
Example 2.5 (Sierpinski gasket) The Sierpinski gasket satises the assumptions of Theorem 2.4 for the
set M = [] (compare Figure 1). Each of the intersections S
i
MS
j
M is a single point w
k
(k = i = j) of their
boundary, implying Equation (2.5). Thus
a
f
() exists.
If a self-similar set F in R
d
can be described by a cut out procedure of an initial set A R
d
, then, as far as no
point from the boundary of A is removed in the cut out process, A is a suitable set M for Theorem 2.4, compare
for instance Example 2.8. However, it is not difcult to nd self-similar sets not tting into the framework of
Theorem 2.4 to which Theorem 2.1 still applies. Example 5.6 and Proposition 5.7 provide a mild generalization.
Note that totally disconnected self-similar sets in R
d
with d 2 are naturally excluded by the conditions of
Theorem 2.4. If a self-similar set F is totally disconnected (and not a singleton) then, under the assumptions of
Theorem 2.4, the same would hold for M. But the polyconvexity of M would imply M, and thus M and F,
to be a nite union of points, contradicting the fact that F is uncountable. For self-similar sets in R we refer to
Section 2.4.
2.3 Examples
Here we present some examples of self-similar sets F, to which Theorem 2.4 can be applied to obtain their
(average) fractal Euler numbers, including the Sierpinski gasket and the Sierpinski carpet. In all the examples
an appropriate set M is easily determined and the conditions on M are easily veried. To calculate the value of

a
f
(F) using formula (2.3), we need to determine the overlap function R of F.
Before turning to the examples we discuss the overlap function R in more detail. In the convex ring we have
the following useful formula for R, which is due to the inclusion-exclusion principle (4.2) and the invariance of
the self-similar set (S(F) = F):
R() =
N

k=2
(1)
k1

1i1<...<i
k
N
((S
i1
F)

. . . (S
i
k
F)

) . (2.6)
Observe that all sets (S
i
F)

and their intersections are polyconvex such that all Euler numbers in this formula
are dened. In examples often very few of the intersections in (2.6) are nonempty and have to be considered
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 157
(S1) (S2)
(S1) (S3)
@
@@
(S2) (S3)

(S1) (S2)
(S1) (S3)
T
T
T
T
(S2) (S3)

Fig. 2 -neighbourhoods of the Sierpinski gasket for u (left) and < u (middle and right). Note that for < u, the sets
(Si) (Sj) remain convex and disjoint as 0.
for R(). Since the number of nonempty intersections cannot increase as 0, this formula proves to be
particularly useful for small > 0.
Since the overlap function R assumes only integer values and since it has a discrete set of discontinuities, it
is a piecewise constant function with integer steps at discontinuity points and constant integer values in between.
In many examples the set of discontinuities is in fact nite, such that R is bounded and condition (1.1) is trivially
satised. As we will see in the examples and, in general, in Section 5, the conditions in Theorem 2.4 imply this
niteness. In the examples we x the diameter b of the self-similar sets to be 1, avoiding this constant in the
formulas. By the scaling invariance (cf. Corollary 2.3) the diameter does not affect the limits.
Example 2.6 (Sierpinski gasket continued) We calculate
a
f
(). The only discontinuity point of R
is u =

3
12
, being the radius of the incircle of the middle triangle of (compare Figure 1). For u, the
neighbourhood

is convex as well as the sets (S

i
)

, implying R() = (

((S
i
)

) = 2. For
< u,

is not convex anymore (cf. Figure 2). But now the intersection (S
i
)

(S
j
)

is a convex set, i = j,
while the intersection of all three sets (S
i
)

is empty. In Equation (2.6) just the terms for k = 2 remain. Thus

R() = (1)
_
((S
1
)

(S
2
)

) +((S
1
)

(S
3
)

) +((S
2
)

(S
3
)

)
_
= 3. Now formula (2.3)
yields the average fractal Euler number of :

a
f
() =
u
s
s
=
u
s
log 3
0.098.
Example 2.7 (Sierpinski carpet) The Sierpinski carpet Q is the self-similar set associated to a system of 8
similarities S
i
each mapping the square H =
_
0,
1

2
to one of the small squares 1, . . . , 8 with contraction ratio
1
3
(cf. Figure 3). Q has similarity dimension s =
log 8
log 3
. As for the Sierpinski gasket, Q satises Theorem 2.4
for the set H = [Q], the convex hull of Q, and, since all contraction ratios are the same, the theorem ensures
the existence of the average fractal Euler number
a
f
(Q). To determine
a
f
(Q) we look at the overlap function
R. Again it has only one point of discontinuity, namely u =
1
6

2
. For u, Q

as well as (S
i
Q)

are
convex, implying R() = 1 8 = 7, by (2.2). For < u we consider the intersection structure of the
sets (S
i
Q)

and use formula (2.6). There are nonempty intersections of 2 and of 3 of these sets, all being
convex, but no intersections of higher order. (S
3
Q)

, for instance, intersects (S

2
Q)

and (S
4
Q)

and also the set

(S
3
Q)

(S
2
Q)

(S
4
Q)

is nonempty. All together there are 12 intersection of two sets and 4 intersections of
three sets, implying R() = 12 + 4 = 8. Integration according to formula (2.3) yields

a
f
(Q) =
u
s
s
=
u
s
log 8
0.019.
In the following example we compare the average fractal Euler numbers of two self-similar sets with equal
dimension but different topological structure.
Example 2.8 (U sets) We modify the construction of the Sierpinski carpet by considering only seven sim-
ilarities, again each mapping the square H =
_
0,
1

2
in Figure 3 (right) to one of the squares 1, . . . , 7. But
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
158 Llorente and Winter: A notion of Euler characteristic for fractals
u
0
1

2
1

2
6
-
Q
3
2
1
4 5
6
7 8
0
1

2
1

2
6
-
H
Fig. 3 Sierpinski carpet Qand the square H =
h
0,
1

2
i
2
subdivided into 9 subsquares each being a copy of H scaled by
1
3
.

u
0
1

2
1

2
6
-
U
1
0
1

2
1

2
6
-
U
2
Fig. 4 Self-similar sets with equal dimension but different fractal Euler numbers.
this time we include a rotation for some of the maps. For the set U
1
we rotate the squares 1 and 7 by angle
, 4 by

2
and 3 by
3
2
. For U
2
we rotate 1, 2, 6 and 7 by , 4 and 5 by

2
and 3 by
3
2
(compare Fig-
ure 4). The so dened self-similar sets U
1
and U
2
both have similarity dimension s =
log 7
log 3
. Their convex hull
[U
1
] = [U
2
] = H is not an appropriate set for Theorem 2.4, but U
1
and U
2
satisfy the conditions of this theorem
for the set U =

i
S
1
i
H =

i
S
2
i
H, where the S
k
i
, i = 1, . . . , 7, are the similarities generating the set U
k
.
Since all contraction ratios are the same (arithmetic case), we calculate the average fractal Euler numbers

a
f
(U
1
) and
a
f
(U
2
). Their values differ, since the overlap functions R
1
of U
1
and R
2
of U
2
are different. Both
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 159
u
2
u
1
Fig. 5 Modied Sierpinski carpet F, an example where the fractal Euler number
f
(F) exists (and not only
a
f
(F))
have the same unique discontinuity point u =
1
18

2
and their values coincide for u: R
1
() = R
2
() =
1 7 = 6, by Equation (2.2). But for < u, R
1
() and R
2
() assume different values. Analyzing the
intersection structure of the sets (S
k
i
U
k
)

for U
k
, Equation (2.6) yields R
1
() = 10 and R
2
() = 12. By (2.3)
and with = log 3, the average fractal Euler numbers are:

a
f
(U
1
) = 4
u
s
s
= 4
u
s
log 7
0.015,
and

a
f
(U
2
) =
a
f
(U
1
) 2
1

_
u
0

s1
d = 6
u
s
log 7
0.023.
Note that U
1
and U
2
have different numbers of holes of radius u, namely 4 and 6, respectively. This is reected
by the average fractal Euler numbers.
In all the examples considered so far we found a unique discontinuity point u for the overlap function R. We
can give u a geometric meaning as the radius of the largest holes in the corresponding self-similar set F, or,
more precisely, as the maximumdistance to F of points in the union of the bounded connected components of the
complement of F. But this unique discontinuity point is due to the simplicity of the examples, where all holes
were either largest holes or iterated copies of them. The situation is not always as simple as the following
example of a modied Sierpinski carpet shows. Moreover, in contrast to the previous examples in this one the
fractal Euler number
f
exists.
_
The non-existence of the fractal Euler number in the previous examples can be
easily seen by comparing
s
k
(F

k
) for different sequences
k
0, see also Example 2.11.
_
Example 2.9 Starting from the square H =
_
0,
1

2
, the self-similar set F is generated by a system of
27 similarities each mapping H to one of the subsquares described in Figure 5. Three of the similarities have
contraction ratio
1
3
, while twelve of them have ratio
1
6
and the remaining twelve ratio
1
12
. Thus the similarity
dimension s of F is the unique solution of
3
3
s
+
12
6
s
+
12
12
s
= 1 (s 1.818), and the non-arithmetic case applies.
Since F satises the conditions of Theorem 2.4 for M = H,
f
(F) exists. The overlap function of F has two
discontinuity points corresponding to the radii u
1
=
1
6

2
and u
2
=
1
12

2
of the two holes in S(M):
R() =
_

_
28 if 0 < u
1
,
27 if u
1
< u
2
,
26 if u
2
< 1.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
160 Llorente and Winter: A notion of Euler characteristic for fractals
Hence, by formula (2.3),
f
(F) =
(u1)
s
+(u2)
s
s
0.021.
2.4 Self-similar sets on the real line
We consider self-similar sets F in R, generated by contracting similarities S
i
: R R, i = 1, . . . , N, with
similarity dimension s. Without loss of generality we assume that the diameter of F is 1. Suppose the system
{S
1
, . . . , S
N
} satises the SSC. For simplicity we assume that the convex hulls [S
i
F], i = 1, . . . , N, are pairwise
disjoint and ordered from left to right according to the index i. Let b
j
, j = 1, . . . , N 1, denote the length of the
gap between the intervals [S
j
F] and [S
j+1
F]. Observe that the SSC implies b
j
> 0.
Following Falconer [2], the gap counting function of F is dened as
G() = #{complementary intervals of F with length > }.
Moreover, we dene the gap limit and the average gap limit of F by
G(F) := lim
0

s
G() and G
a
(F) := lim
0
1
|log |
_
1

s
G()
d

, (2.7)
respectively, in case these limits exist. Falconer gave sufcient conditions for the existence of the gap limit and
provided a formula in terms of the gaps b
i
(compare [2, Proposition 7.5]). Here we obtain the same result as a
corollary of Theorem 2.1. Moreover, we give sufcient conditions for the average gap limit to exist.
The main point to observe is the close relation between the gap counting function and the Euler characteristic
of neighbourhoods of F. First note, that all neighbourhoods F

are nite unions of compact intervals, hence

polyconvex. Moreover, their Euler characteristic, being the number of their connected components, is determined
by the number of gaps of F of length greater than 2:
(F

) = 1 +G(2). (2.8)
This relation between both expressions results in a similar interdependence of their rescaled limits.
Corollary 2.10 Let F be a self-similar set in R as dened above. Then the Euler exponent of F equals s and
the following holds:
(i) Average gap limit and average fractal Euler number of F always exist and are given by
G
a
(F) = 2
s

a
f
(F) =
1
s
N1

i=1
b
i
s
, (2.9)
where =

N
i=1
r
s
i
log r
i
.
(ii) If {log r
1
, . . . , log r
N
} is non-arithmetic, both, gap limit G(F) and fractal Euler number
f
(F),
exist and coincide with their averaged counterparts.
Pr o o f. Let u
i
=
bi
2
denote the radius of the i-th rst level gap (i = 1, . . . , N 1). It is not difcult to see,
that R() is the number of rst level gaps of radius smaller than , i.e.,
R() = #{u
i
< : i = 1, . . . , N 1} =
N1

i=1
1
[ui,1]
(),
where 1
I
denotes the indicator function of the set I R, i.e., 1
I
() = 1 if I and 1
I
() = 0 else. In particular,
this implies that R() = 0 for small (i.e., for < min
i
u
i
). Therefore, Theorem 2.1 can be applied:
X =
1

N1

i=1
_
1
ui

s1
d +
N 1
s
_
=
1
s
N1

i=1
u
s
i
.
Since u
i
> 0, obviously X = 0. We conclude that = s. Moreover,
a
f
(F) exists and equals X. Now (2.8)
implies the existence of the average gap limit G
a
(F), since
G
a
(F) = lim
0
1
|log |
_
1

s
G()
d

= lim
0
1
|log |
_
1

s
(F
/2
)
d

= 2
s

a
f
(F),
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 161
proving (i). In case {log r
1
, . . . , log r
N
} is non-arithmetic,
f
(F) exists (since X = 0) and coincides with

a
f
(F), implying that the gap limit G(F) exists and is given by
G(F) = lim
0

s
G() = 2
s
lim
0

s
(F

) = 2
s

f
(F).
This completes the proof.
As with other lacunarity parameters, such as the Minkowski content (see [2], [3], [8] and [10]), for many
interesting self-similar sets the fractal Euler number and the gap limit do not exist. This leads to considering their
average counterparts, as in the following example of the middle-third Cantor set.
Example 2.11 Let C be the middle-third Cantor set, which is generated by the similarities S
1
(x) =
1
3
x and
S
2
(x) =
1
3
x +
2
3
. Therefore, r
1
= r
2
=
1
3
, b
1
=
1
3
and s =
log 2
log 3
. Since C is h-arithmetic with h = log 3, the
existence of
f
(C) and G(C) is not guaranteed by Corollary 2.10. In fact, both numbers do not exist: Since the
gap counting function of C is
G() = 2
k
1 for 3
(k+1)
< 3
k
,
the sequences (
k
) =
_
1
2
3
k
_
and (
k
) = 3
k
, k N, provide different limits as k , namely
lim
k

s
k
G(
k
) = 2
s
< 1 = lim
k

s
k
G(
k
)
and similarly
lim
k

s
k
(C

k
) = 2
s
> 2
1
= lim
k

s
k
(C

k
).
On the other hand, formula (2.9) yields G
a
(C) =
1
2 log 2
0.721 and
a
f
(C) =
2
(s+1)
log 2
0.466.
3 Proof of Theorem 2.1
3.1 Renewal theorem
For proving Theorem 2.1 we need the Renewal theorem which we state and discuss now. Let P be a Borel
probability measure with support contained in [0, ) and =
_

0
t P(dt) < . Let z : R R be a function
with a discrete set of discontinuities satisfying
|z(t)| c
1
e
c2|t|
for all t R (3.1)
for some constants 0 < c
1
, c
2
< . It is well-known in probability theory that under the above conditions on z
the equation
Z(t) = z(t) +
_

0
Z(t ) P(d) (3.2)
has a unique solution Z(t) in the class of functions satisfying lim
t
Z(t) = 0. Equation (3.2) is called re-
newal equation and the asymptotic behaviour of its solution as t is given by the so-called Renewal theorem.
We quote the following discrete version of the Renewal theorem from Levitin and Vassiliev [9, p. 198], which is
adapted to the fractal setting and where a complete proof is provided (compare also Falconer [2, Corollary 7.3, p.
122]). Here P is a probability measure supported by a nite set of points {y
1
, . . . , y
N
} such that P({y
i
}) = p
i
for i = 1, . . . , N and therefore
=
N

i=1
y
i
p
i
. (3.3)
A function g : R R is said to be asymptotic to a function f : R R, g f, if for all > 0 there exists a
number D = D() such that
(1 )f(t) g(t) (1 +)f(t) for all t > D. (3.4)
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
162 Llorente and Winter: A notion of Euler characteristic for fractals
Theorem 3.1 (Renewal Theorem) Let 0 < y
1
y
2
. . . y
N
and p
1
, . . . , p
N
be positive real numbers
such that

N
i=1
p
i
= 1. For a function z which fullls (3.1), let Z : R R be the unique solution of the renewal
equation
Z(t) = z(t) +
N

i=1
p
i
Z(t y
i
) (3.5)
satisfying lim
t
Z(t) = 0.
(i) If the set {y
1
, . . . , y
N
} is non-arithmetic, then
lim
t
Z(t) =
1

z() d.
(ii) If {y
1
, . . . , y
N
} is h-arithmetic for some h > 0 then
Z(t)
h

k=
z(t kh).
Moreover, Z is uniformly bounded in R.
Theorem 3.1 implies that in the non-arithmetic case the limit lim
t
Z(t) exists, while in the h-arithmetic
case Z is asymptotic to some periodic function of period h (i.e., to some function f with f(t + h) = f(t) for
all t R). The latter is sufcient for the limit lim
T
1
T
_
T
0
Z(t) dt to exist as the following lemma shows.
Although this is standard analysis, we include a short proof for the convenience of the reader.
Lemma 3.2 Let f be a locally integrable periodic function with period h > 0, i.e., f(t + h) = f(t) for all
t R, and let L :=
_
h
0
f(t) dt.
(i) Then the limit lim
T
1
T
_
T
0
f(t) dt exists and equals h
1
L.
(ii) If g : R R is a function such that g f, then also the limit lim
T
1
T
_
T
0
g(t) dt exists and equals h
1
L.
Pr o o f. (i) For T > 0 choose n N such that nh < T (n + 1)h. Then
1
T
_
T
0
f(t) dt =
1
T
_
n1

i=0
_
(i+1)h
ih
f(t) dt +
_
T
nh
f(t) dt
_
.
By the periodicity of f, the right-hand side is bounded from above by
1
T
_
nL +
_
h
0
|f(t)| dt
_
and from below by
1
T
_
nL
_
h
0
|f(t)| dt
_
. Since f is locally integrable, the stated limit follows by letting T .
(ii) Fix some > 0. On one hand (3.4) implies
1
T
_
T
0
g(t) dt
1
T
_
_
D
0
g(t) dt + (1 +)
_
T
D
f(t) dt
_
for T > D and thus
limsup
T
1
T
_
T
0
g(t) dt (1 +)h
1
L.
On the other hand (3.4) yields
(1 )h
1
L liminf
T
1
T
_
T
0
g(t) dt.
The statement follows by letting 0.
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 163
As a direct consequence of the Renewal theorem and Lemma 3.2 we have the following
Corollary 3.3 Under the assumptions of Theorem 3.1 the following limit always exists and equals to the
expression on the right-hand side:
lim
T
1
T
_
T
0
Z(t) dt =
1

z() d.
Pr o o f. For {y
1
, . . . , y
N
} being h-arithmetic, just note, that the function f(t) =
h

k=
z(t kh) in
Theorem 3.1 (ii) is uniformly bounded and periodic, and apply Lemma 3.2 (ii) to g(t) = Z(t). In the non-
arithmetic case the limit lim
t
Z(t) exists and the assertion follows by applying Lemma 3.2 to g(t) = Z(t)
which is asymptotic to the constant function f(t) lim
t
Z(t).
3.2 Proof of Theorem 2.1
The following statement describes the structure of -neighbourhoods of general compact sets K R
d
for large
. A consequence is that the Euler characteristic (K

) of the neighbourhoods exists and equals 1 for these . We

make use of this fact in the proof of formula (2.3).
Recall that a set A K is called a strong deformation retract of K if there is a homotopy : K[0, 1] K
such that (, 0) = id, (K, 1) = A and (a, t) = a for all a A and t [0, 1]. The map (, 1) : K A is
called a retraction.
Proposition 3.4 Let b be the diameter of the compact set K R
d
. Then, for b, the convex hull [K] of K
is a strong deformation retract of the parallel set K

and (K

) = 1.
Pr o o f. Fix b. By construction, [K] K

. Let
[K]
be the metric projection from R
d
onto [K]. It is not
difcult to see that for all x K

the line segment connecting x and

[K]
(x) is contained in K

. Therefore, the
mapping : K

[0, 1] K

dened by
(x, t) := (1 t)x +t
[K]
(x)
is a homotopy satisfying (, 0) = id, (x, 1) =
[K]
(x) [K] for all x K

, and (x, t) = x for all x [K]

and t [0, 1]. Hence, [K] is a strong deformation retract of K

. The latter implies that the homology groups

of the singular complexes of [K] and K

are isomorphic, and (K

) = 1 follows from the contractability of the

convex set [K].
Considering the denition of the overlap function R and Equation (2.1), which implies ((S
i
F)

) = (F
/ri
)
for i = 1, . . . , N, we can write the following for the Euler characteristic of F

:
(F

) =
N

i=1
(F
/ri
) +R(). (3.6)
Dene
_
Z(t) = be
st
(F
be
t ) for t 0,
0 for t < 0.
(3.7)
Using (3.6) we obtain
Z(t) = be
st
_
N

i=1
(F
be
(t+log r
i
) ) +R
_
be
t
_
_
=
N

i=1
r
s
i
be
s(t+log ri)
(F
be
(t+log r
i
) ) +be
st
R
_
be
t
_
.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
164 Llorente and Winter: A notion of Euler characteristic for fractals
The terms of the sum over i satisfying t + log r
i
0 can be replaced by r
s
i
Z(t + log r
i
). Then we have
Z(t) =

t+log ri0
r
s
i
Z(t + log r
i
) +z(t) (3.8)
where
_
z(t) = be
st
R
_
be
t
_
+

t+log ri<0
be
st
(F
be
(t+log r
i
) ) for t 0,
0 for t < 0.
(3.9)
Since Z(t) = 0 for t < 0, we can add the zeros for t + log r
i
< 0 in the sum in (3.8) to obtain
Z(t) =
N

i=1
r
s
i
Z(t + log r
i
) +z(t). (3.10)
The function z(t) can be simplied, to see that the term containing the overlap function really is the important
one. Noting that be
(t+log ri)
> b for t +log r
i
< 0, by Proposition 3.4, the Euler characteristic (F
be
(t+log r
i
) )
equals 1 whenever it occurs in z(t). Using indicator functions 1
[ri,1]
we can write (3.9) as
z(t) = be
st
_
R
_
be
t
_
+
N

i=1
1
[ri,1]
_
e
t
_
_
. (3.11)
Now observe that (3.10) is a renewal equation for Z and that, trivially, lim
t
Z(t) = 0. The assumptions
on the overlap function R ensure that z has a discrete set of discontinuities. Moreover, condition (1.1) implies
the existence of positive constants c
1
:= b(cb
s
+N) and c
2
:= > 0 such that
|z(t)| be
st
_

R
_
be
t
_

+N
_
c
1
e
c2t
for t 0. Thus condition (3.1) is satised and we can apply Theorem 3.1 with p
i
= r
s
i
and y
i
= log r
i
. Since,
by Theorem 3.1, Z(t) is uniformly bounded,
s
|(F

)| is bounded as 0 and therefore, s for the Euler

exponent of F. Furthermore, there are two cases to consider.
The non-arithmetic case. If {log r
1
, . . . , log r
N
} is non-arithmetic, the limit
lim
t
Z(t) = b lim
t
e
st
(F
be
t ) = b lim
0
_

b
_
s
(F

)
exists and equals the integral
1

_

0
z() d =
1

_
1
0
z(log r)
dr
r
(3.12)
where the right-hand side is due to the substitution r = e

and, by (3.3), =

N
i=1
r
s
i
log r
i
. Inserting the
integrand according to (3.11) and dividing by b yields
lim
0
_

b
_
s
(F

) =
1

_
_
1
0
r
s1
R(br) dr +
N

i=1
_
1
0
r
s1
1
[ri,1]
(r) dr
_
=
1

__
1
0
r
s1
R(br) dr +
N 1
s
_
= X,
as proposed in (ii) of Theorem2.1. If nowX = 0 then for each t < s, it is clear that
t
|(F

)| cannot be bounded
as 0, implying s, and thus = s and
f
(F) = X. For the non-arithmetic case, (i) easily follows from
(ii), since the existence of a limit always implies existence and identicalness of the corresponding average limit.
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 165
The h-arithmetic case. If {log r
1
, . . . , log r
N
} is h-arithmetic, Corollary 3.3 states that the average limit
lim
T
1
T
_
T
0
Z(t) dt = b lim
0
1
|log |
_
1

b
_
s
(F

)
d

(3.13)
exists and equals the expression in (3.12). Hence, repeating the calculations from the non-arithmetic case, we
obtain formula (2.3) also for the h-arithmetic case. Now condition X = 0 means that the average limit in (3.13)
is non-zero, implying that also
limsup
0

s
|(F

)| > 0.
Therefore, as in the non-arithmethic case, we obtain = s and
a
f
(F) = X. This completes the proof of
Theorem 2.1.
4 Convex ring and fractal Euler number
4.1 Euler characteristic in R
d
Here we briey introduce the Euler characteristic as a valuation in the convex ring R
d
. We follow the lattice
theoretic approach of Klee [7] and Rota [13]. For more detailed explanations compare Klain and Rota [6],
Schneider [15] and the references therein. A set C R
d
is said to be convex if for any two points x, y C the
line segment connecting them is a subset of C. Denote by K
d
the family of all compact convex sets in R
d
and
recall the denition of the conex ring R
d
from Section 2.2. Note that K
d
R
d
.
Denition 4.1 A valuation on R
d
is a function : R
d
R such that () = 0 and, for all A, B R
d
,
(A B) = (A) +(B) (A B). (4.1)
By iterating the identity (4.1) the so called inclusion-exclusion principle is obtained for :

_
n
_
i=1
A
i
_
=
n

k=1
(1)
k1

1i1<...<i
k
n
(A
i1
. . . A
i
k
) for A
1
, . . . , A
n
R
d
. (4.2)
On K
d
the Euler characteristic is the function which assigns value 1 to each nonempty convex set. Hadwiger
[4] considered extensions of this trivial valuation on K
d
to the convex ring R
d
and showed the existence and
uniqueness of such an extension, as the following theorem states.
Theorem 4.2 [15, Thm. 3.4.12] There exists a unique valuation on the convex ring R
d
, such that (C) = 1
whenever C K
d
\ {}.
This valuation on R
d
is called the Euler characteristic. We emphazise again that for polyconvex sets this
notion of Euler characteristic coincides with the more general notion of Euler characteristic for cell complexes.
For any set A R
d
, (A) can be determined by choosing a decomposition into compact convex sets and
using the inclusion-exclusion principle. According to Theorem 4.2, (A) is independent of the choice of this
decomposition. Note that the convex ring is closed with respect to Euclidean motions and scaling and that is a
motion and scaling invariant valuation.
The following lemma collects some basic properties of polyconvex sets and their -neighbourhoods, which we
use frequently in the sequel. For convenience we sometimes use the 0-neighbourhood A
0
to denote the set A
itself. Observe that (iii) is a special case of Proposition 3.4, however, we include a very simple proof of this fact
for the convex ring setting.
Lemma 4.3 For a set A R
d
the following holds:
(i) A

R
d
for all > 0.
(ii) (A

), as a function of , has a nite set of discontinuities in (0, ) and lim

0
(A

) = (A).
(iii) (A

) = 1 for all b, where b denotes the diameter of A.

www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
166 Llorente and Winter: A notion of Euler characteristic for fractals
Pr o o f. Assertion (i) follows immediately from the fact that neighbourhoods of convex sets are convex. For
a proof of (ii) and (iii), let C
1
, . . . , C
n
K
d
be sets such that A =

n
k=1
C
k
. Then by the inclusion-exclusion
principle,
(A

) =
_
n
_
k=1
C
k

_
=
n

k=1
(1)
k1

1i1<...<i
k
n

_
C
i1

. . . C
i
k

_
. (4.3)
Now observe that each of the sets C
i1

. . . C
i
k

in the sum is convex and has either Euler characteristic 0

or 1. If C
i1
. . . C
i
k
= then
_
C
i1

. . . C
i
k

_
, as a function of , is constant equal to 1, while in case
C
i1
. . . C
i
k
= ,
_
C
i1

. . . C
i
k

_
, as a function of , has exactly one discontinuity point > 0 where the
Euler characteristic jumps from0 to 1. Thus each of the terms in the nite sum (4.3) has at most one discontinuity
point, implying that (A

), as a function of , has a nite number of discontinuities in (0, ). (A

) (A)
as 0 follows since (A

) is continuous (from the right) at 0, proving (ii).

Observe that for b the sets C
i1

. . .C
i
k

are nonempty and convex and therefore have Euler characteristic

one. Applying this to (4.3) and noting that

1i1<...<i
k
n
1 = (
n
k
) and

n
i=1
(1)
k1
(
n
k
) = 1, by the binomial
formula, assertion (iii) follows.
Remark 4.4 Lemma 4.3 (i) implies that for any polyconvex set Athe Euler exponent equals 0, since |(A

)|
is bounded. Thus
f
(A) = lim
0
(A

) and so, by Lemma 4.3 (ii),

f
(A) = (A).
Hence for polyconvex sets the fractal Euler number coincides with the classical Euler characteristic, a further
justication of the concepts introduced in this paper.
Remark 4.5 There are trivial examples of polyconvex self-similar sets like the square. By the above, their
Euler exponent is zero and thus strictly less than their similarity dimension s, implying X = 0 in Theorem 2.1.
Such examples clarify why the restriction to the case X = 0 in Theorem 2.1 is necessary and natural.
4.2 Self-similar sets with -neighbourhoods in R
d
Given a self-similar set F, it is very easy to decide, whether it ts into the convex ring setting, i.e., whether its
-neighbourhoods are polyconvex sets, or not. Either all neighbourhoods are polyconvex or none.
Proposition 4.6 Let F be a self-similar set. Then the following statements are equivalent:
(i) For all > 0, F

R
d
.
(ii) There is an > 0 such that F

R
d
.
Pr o o f. We prove (ii) (i), the other direction is trivial. Assume F

R
d
for some > 0. Then, by
Lemma 4.3 (i), F

R
d
for all > . Let now < and n N such that > r
n
max
. We can write F as the
union of its n-th level cylinder sets implying
F

=
_
i
k
{1,...,N}
(S
i1
. . . S
in
F)

. (4.4)
Observe that (S
i1
. . . S
in
F)

= S
i1
. . . S
in
(F
/ri
1
...rin
). Since

ri
1
...rin

r
n
max
> , we have F
/ri
1
...rin

R
d
. Thus F

, being the nite union of polyconvex sets in (4.4), is an element of R

d
.
According to Proposition 4.6, it is sufcient to investigate an arbitrarily chosen single neighbourhood F

of
F. If it is polyconvex all neighbourhoods of F are, else none is. The value can be chosen large (i.e., b)
such that the -neighbourhood is simply connected and has no inner structure, simplifying the investigation.
However, there are self-similar sets with non-polyconvex neighbourhoods, e.g. the von Koch curve or the Cantor
set in Figure 6 (left).
In view of Proposition 4.6 it is clear that the overlap function R is well-dened if we assume that F

R
d
for some > 0. But with this assumption we can even say more about the overlap function, namely that R has a
discrete set of discontinuities:
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 167
C
0
1

2
1

2
6
-
U
3
Fig. 6 The Cantor set C has no polyconvex neighbourhoods. The self-similar set U3 does not satisfy the assumptions of
Theorem 2.4 but of Proposition 5.7.
Proposition 4.7 Let F be a self-similar set with polyconvex neighbourhoods. Then the overlap function
R : (0, ) R has a discrete set of discontinuities.
Pr o o f. Let > 0. Since F

and (S
i
F)

are in R
d
, by Lemma 4.3 (ii), (F
+
) and ((S
i
F)
+
) have,
as functions of , a nite set of discontinuities in (0, ). Therefore, the overlap function R() = (F

N
i=1
((S
i
F)

) has a nite set of discontinuities in (, ). Since this holds for each > 0, R has a discrete set
of discontinuities in (0, ).
Note that 0 is the only possible accumulation point of discontinuities of the overlap function R.
5 Classes of self-similar sets with (average) fractal Euler number
Here we discuss the class of self-similar sets F satisfying the conditions of Theorem 2.4 and some extension of
this class. First observe that in all examples in Section 2.3 the -neighbourhoods of F can be described in terms
of the set M of Theorem 2.4. Depending on we can nd a level n = n() such that F

coincides with the

union of the -neighbourhoods of all n-th level iterated images S
i1
. . . S
in
M of the set M. This holds for
all self-similar sets satisfying the conditions of Theorem 2.4, in fact even for a larger class of sets, as Lemma 5.1
shows. Let F be a self-similar set with diameter b. We dene the function
n() := min
_
n N
0
: r
n
max
b
_
.
For A R
d
, let A
0
:= A and for n = 1, 2, . . . , dene A
n
:=

N
i=1
S
i
(A
n1
). With A and

A we denote the
closure and interior of the set A, respectively.
Lemma 5.1 Assume there exists a compact set M R
d
such that M F M. Then for all b
F

= M

. (5.1)
More generally,
F

= (M
m
)

(5.2)
for all integers m n() and > 0. In particular, if M R
d
, then F

R
d
for all > 0.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
168 Llorente and Winter: A notion of Euler characteristic for fractals
Pr o o f. Let M R
d
be given as above and b. Then it is clear that M F

. On the other hand, for

x M

\ M we have d(x, M) and, since M F, x F

. Therefore, M

= M (M

\ M) F

. The
reversed inclusion is clear since F M, thus (5.1) holds.
To prove (5.2) we rst note that, by denition of n(),

ri
1
...rim

r
n()
max
b for any m n() and i
j

{1, . . . , N} (j = 1, . . . , m). Therefore, by (5.1), M
/(ri
1
...rim
)
= F
/(ri
1
...rim
)
implying
(M
m
)

=
N
_
i1=1
. . .
N
_
im=1
S
i1
. . . S
im
_
M
/(ri
1
...rim
)
_
=
N
_
i1=1
. . .
N
_
im=1
S
i1
. . . S
im
_
F
/(ri
1
...rim
)
_
= F

.
Finally, if M R
d
, then F

R
d
follows directly from (5.1) and Proposition 4.6.
Remark 5.2 Clearly, if the set M in Lemma 5.1 is convex, then M = [F]. As pointed out before, for many
examples of self-similar sets the set M = [F] satises M F M. However, as Example 2.8 shows, there
are self-similar sets satisfying the conditions of Lemma 5.1 with M = [F].
If a set M R
d
of Lemma 5.1 exists for a self-similar set F, then Proposition 4.7 implies that the overlap
function R has a discrete set of discontinuities. The following lemma shows that for the class of self-similar sets
satisfying the conditions of Theorem 2.4 the set of discontinuities of the (piecewise constant) function R is even
nite implying that R is bounded.
Lemma 5.3 If there exists a set M R
d
such that M F M and
S
i
(M) S
j
(M) = S
i
(M) S
j
(M) (5.3)
for all i = j, then the overlap function R has a nite set of discontinuities and is bounded. Moreover,
lim
0
R() =
N

k=2
(1)
k1

1i1<...<i
k
N
(S
i1
M . . . S
i
k
M). (5.4)
Pr o o f. It sufces to show that
(S
i
F)

(S
j
F)

= (S
i
M)

(S
j
M)

(5.5)
for > 0 and i = j. By the inclusion-exclusion principle, this implies, that the overlap function can be written as
R() =
N

k=2
(1)
k1

1i1<...<i
k
N
((S
i1
M)

. . . (S
i
k
M)

)
=
__
N
_
i=1
S
i
M
_

i=1
((S
i
M)

).
By Lemma 4.3 (ii), each term in the last sum has, as a function of , a nite set of discontinuities in (0, ),
implying the same for R. Therefore, since R is a piecewise constant function, we conclude that it is bounded.
Formula (5.4) follows immediately.
To complete the proof we show Equation (5.5). Since S
i
F S
i
M, one of the inclusions follows immediately.
To prove the remaining one, let x (S
i
M)

(S
j
M)

. We show that d(x, S

i
M) . Assume the contrary.
Then, x S
i
M (x (S
i
M)

\ S
i
M implies d(x, S
i
M) .) Since x (S
j
M)

we can nd y S
j
M such
that d(x, y) . Now the assumption implies that on one hand y S
i
MS
j
M since d(x, (S
i
M)
c
) > . On the
other hand y / S
i
M, a contradiction to (5.3). Thus d(x, S
i
M) , implying that x (S
i
M)

(S
i
F)

.
The same arguments apply to j. Hence x (S
i
F)

(S
j
F)

, completing the proof of (5.5).

c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com
Math. Nachr. 280, No. 12 (2007) 169
Remark 5.4 Condition (5.3) in Lemma 5.3 cannot be removed. Consider, for example, a modied Sierpinski
gasket with strong overlap as dened by the following system of similarities in R
2
: S
1
x =
2
3
x, S
2
x =
2
3
x +
_
1
3
, 0
_
and S
3
x =
1
3
x +
_
1
3
,
1

3
_
. The convex hull of this self-similar set satises the conditions of M in
Lemma 5.3 except (5.3) but its overlap function does not have a nite set of discontinuities.
Remark 5.5 If, in addition to the conditions of Lemma 5.3, we have S(M) M (which is always true for
convex M), then {S
1
, . . . , S
N
} satises the OSC for the set

M: condition (5.3) implies S

i
_

M
_
S
j
_

M
_
=
and S(M) M obviously implies S
_

M
_

M.
Observe that the set M satisfying the conditions in Lemma 5.3 is not uniquely determined for a self-similar set
F. In fact, if there is one, there are many such sets. It is not difcult to see that, for instance, with M, also the set
S(M) satises the conditions. In the examples it is easy to nd such sets M with the assumption S(M) M not
satised. For the Sierpinski gasket , for example, the set M

being the union of the set SS[] and the triangle

w
1
w
2
w
3
(compare Figure 1) satises the conditions of Lemma 5.3 but not S(M

) M

. But this condition

must be satised, if we want to use the interior of M

as the open set of the OSC. It is, however, not clear if this
assumption is really needed. We conjecture that, whenever a set M of Lemma 5.3 exists, one can also nd such
a set M with S(M) M. Note that, if S(M) M, the fact

M F = implies that

M is also a suitable open

set for the strong open set condition, which in R
d
is equivalent to the OSC (cf. [14]).
Pr o o f o f Th e o r e m 2.4. By Lemma 5.1, the existence of a set M for F as assumed in Theorem2.4 implies
that F

R
d
for > 0. Moreover, Lemma 5.3 states that under the assumptions of Theorem 2.4 the overlap
function Rof F has a nite set of discontinuities and that Ris bounded. Therefore, the conditions of Theorem2.1
are satised and the assertions follow directly.
Although the conditions in Theorem 2.4 provide a reasonable class of self-similar sets for which the (average)
fractal Euler numbers exist, the following example shows that there are still interesting self-similar sets satisfying
the conditions of Theorem 2.1 that are not included in the class covered by Theorem 2.4. In particular, the
condition M F can be relaxed (cf. Proposition 5.7).
Example 5.6 Let U
3
be the self-similar set in Figure 6 (right) obtained in the same way as U
1
and U
2
in Ex-
ample 2.8 with the obvious changes in the choice of the similarities. Proceeding as in the examples in Section 2.3,
we obtain
R() =
_
1 7 = 6 if u,
12 + 2 = 10 if 0 < < u,
where u =
1
18

2
. Therefore,

a
f
(U
3
) =
1

__
1
0

s1
R() d +
N 1
s
_
=
4u
s
s
=
4u
s
log 7
0.015.
The above example motivates the following proposition that gives some sufcient geometric conditions for
R() to be bounded in the case when M is not a subset of F.
Proposition 5.7 Let F be a self-similar set. Assume there is a set M R
d
such that F M and
S
i
M S
j
M = S
i
A S
j
A (5.6)
for all i = j, where A = M F. Let V =
_
y R
d
\ M : d(y, M) < d(y, A)
_
. If
(S
i
M)

(S
j
M)

S
i
V = (S
i
M S
j
M)

S
i
V (5.7)
for all i = j and all > 0, then the overlap function R has a nite set of discontinuities. Moreover, Equation
(5.4) holds.
www.mn-journal.com c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
170 Llorente and Winter: A notion of Euler characteristic for fractals
Pr o o f. We only need to show that the assumptions imply Equation (5.5) as in the proof of Lemma 5.3. Then
the assertions follow by the arguments given there.
Since S
i
F S
i
M, one of the inclusions in (5.5) follows immediately. To prove the remaining one, let
x (S
i
F)

(S
j
F)

. Then either x S
i
V or not.
(i) For x S
i
V , conditions (5.6) and (5.7) imply x (S
i
AS
i
A)

and, since A F, x (S
i
F S
j
F)

(S
i
F)

(S
j
F)

, proving (5.5) in this case.

(ii) For x / S
i
V , we showthat d(x, S
i
A) . Assume the contrary. Then, clearly, x S
i
M. (x (S
i
M)

\
S
i
M implies d(x, S
i
A) d(x, S
i
M) since x / S
i
V.) Since x (S
j
M)

we can nd y S
j
M such that
d(x, y) . Nowthe assumption implies that on one hand y S
i
MS
j
M since d(x, (S
i
M)
c
) > , on the other
hand y / S
i
A, a contradiction to (5.6). Thus d(x, S
i
A) , implying that x (S
i
A)

(S
i
F)

. The same
arguments apply to j (we can assume that x / S
j
V , otherwise we are in case (i)). Hence x (S
i
F)

(S
j
F)

,
completing the proof of (5.5) and thus of the proposition.
Acknowledgements The authors would like to thank Martina Z ahle for directing our interest towards these problems and
for many fruitful discussions and invaluable suggestions.
Note added in proof. During the revision process we managed to prove that for self-similar sets (with neighbourhoods
in R
d
) the OSC is sufcient for Theorem 2.1 to hold, i.e., the OSC implies condition (1.1). Since the proofs are technical
and long, we do not include this result here. It will be included in a more general treatment of the rescaled limits of intrinsic
volumes of neighbourhoods of self-similar sets.
References
[1] A. Dold, Lectures on Algebraic Topology (Springer-Verlag, Berlin New York, 1980).
[2] K. J. Falconer, Techniques in Fractal Geometry (Wiley, New York, 1997).
[3] D. Gatzouras, Lacunarity of self-similar and stochastically self-similar sets, Trans. Amer. Math. Soc. 352, 19531983
(2000).
[4] H. Hadwiger, Eulers Charakteristik und kombinatorische Geometrie, J. Reine Angew. Math. 194, 101110 (1955).
[5] J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30, 713747 (1981).
[6] D. A. Klain and G.-C. Rota, Introduction to Geometric Probability (Cambridge University Press, Cambridge, 1997).
[7] V. Klee, The Euler characteristic in combinatorial geometry, Amer. Math. Monthly 70, 119127 (1963).
[8] M. Lapidus and C. Pomerance, The Riemann Zeta function and the one-dimensional WeylBerry conjecture for fractal
drums, Proc. London Math. Soc. (3) 66, 4169 (1993).
[9] M. Levitin and D. Vassiliev, Spectral assymptotics, Renewal theorem, and the Berry conjecture for a class of fractals,
Proc. London. Math. Soc. (3) 72, 188214 (1996).
[10] B. Mandelbrot, Measures of fractal lacunarity: Minkowski content and alternatives, in: Fractal Geometry and Stochas-
tics, edited by C. Bandt, S. Graf and M. Z ahle, Progress in Probability Vol. 37 (Birkh auser, Basel and Boston, 1995),
pp. 1542.
[11] M. Mor an, Dynamical boundary of a self-similar set, Fund. Math. 160, 114 (1999).
[12] J. Rataj and M. Z ahle, Normal cycles of Lipschitz manifolds by approximation with parallel sets, Differential Geom.
Appl. 19, 113126 (2003).
[13] G.-C. Rota, On the combinatorics of the Euler characteristic, in: Studies in Pure Mathematics (Academic Press, London,
1971), pp. 221233.
[14] A. Schief, Separation properties of self-similar sets, Proc. Amer. Math. Soc. 122, 111115 (1994).
[15] R. Schneider, Convex Bodies: The BrunnMinkowski Theory (Cambridge University Press, New York, 1993).
c 2007 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.mn-journal.com