Professional Documents
Culture Documents
1 (2002) 77–89
c World Scientific Publishing Company
Abstract
We prove for a subclass of recurrent iterated function systems (also called graph-directed iter-
ated function systems) that the boundary of their attractor is again the attractor of a recurrent
IFS. Our method is constructive and permits computation of the Hausdorff dimension of the
attractor and its boundary.
77
78 F. M. Dekking & P. van der Wal
for d ∈ Dij , e ∈ Dik and 1 ≤ i, j, k ≤ r (except of to a non-empty limit set K(S). Then we can find
course when j = k and d = e). an M -recurrent IFS with attractor (C0 , C1 , . . . , Cr )
The goal of the present work is to construct and such that C0 = K(S).
to characterize the boundaries of self-similar sets Our main results follow from the results about
C, satisfying Eq. (1) and more generally of sets BCA’s and the relation between recurrent IFS’s
C1 , . . . , Cr satisfying Eq. (2), without the tiling con- and BCA’s. Let C be a component of the attrac-
dition. Our approach is to generate the bound- tor (C1 , . . . , Cr ) of an M -recurrent IFS. Then we
ary by iterates of a d-dimensional substitution. An can find another M -recurrent IFS with attractor
advantage of this approach is that the iterates al- (B0 , B1 , . . . , Bl ), such that the boundary of C is
ways converge to the boundary of the tile, since equal to B0 . In addition, we can calculate the Haus-
the d-dimensional substitutions can distinguish in dorff dimension of C and of its boundary.
some sense between boundary points and interior
points of the limiting set (cf. Theorem 3). This
shows that the “well-behavedness” condition on the 2. AN INTRODUCTORY EXAMPLE
boundary which Duvall et al.2 require in their anal-
Consider a two-dimensional tile C defined by
ysis, is in some sense superfluous (cf. their Remark [
on p. 10). 2C = C + D = (C + d)
Let (C1 , . . . , Cr ) be a vector of non-empty com- d∈D
pact sets in RSd that satisfy Eq. (2), where we only where 2C = {2x : x ∈ C}, and
assume that rj=1 Dij is non-empty for each i. For
D = {(−1, −1), (0, 0), (1, 0), (0, 1)} .
reasons of simplicity, we will assume that l is just
an integer scaling, i.e. l(x) = M x, where M ≥ 2 is Since D is a digit set for the lattice Z2 , the dimen-
an integer, and that the sets Dij are subsets of Zd . sion of the boundary of C can be calculated using
However, the results in this paper can be general- the contact matrix as described in Refs. 2 and 5.
ized to the case where l is an expansive similarity However, with this example, we will demonstrate
preserving a lattice R and Dij ⊂ R. We will call the use of branching cellular automata. We approx-
the pair (M, (Dij )1≤i, j≤r ) an M -recurrent IFS. imate the tile C as follows. Let C (0) = {(0, 0)} and
To analyze this M -recurrent IFS, we introduce recursively define sets C (1) , C (2) , . . . by
another way to generate compact sets in Rd , by 1
means of (deterministic) branching cellular au- C (n+1) = C (n) + n+1 D .
2
tomata (BCA), which can also be viewed as d- Then C is the Hausdorff limit of the sequence
dimensional substitutions. Stochastic BCA’s were C (0) , C (1) , . . . (cf. Fig. 1).
already explored in Dekking and van der Wal These approximating sets can also be obtained
(2001).7 A finite set T is associated with each BCA, by means of a branching cellular automaton, i.e. a
a set which we will refer to as the set of types. For two-dimensional substitution with neighbor depen-
any subset S of T , a sequence (Kn (S))n≥0 of closed dence. Let A = {0, 1} be our alphabet and define a
and uniformly bounded subsets of Rd is defined. In map σ : A3×3 → A2×2 by
Ref. 7, results were given concerning convergence of
the sets (Kn (S))n≥0 to a limit set K(S) (w.r.t Haus- ∗ ∗ 0 ∗ ∗ 1
0 0 0 1
dorff metric) and the Hausdorff dimension of K(S). ∗ 0 ∗ → ∗ 0 ∗ →
0 0 0 0
The boundary ∂K(S) of K(S) was shown to be the ∗ ∗ ∗ ∗ ∗ ∗
limit set of a sequence (Kn (∂S))n≥0 , for a specific
subset ∂S of T . In this paper, we re-formulate these ∗ ∗ 0 ∗ ∗ 1
1 0 1 1
results for deterministic BCA’s. ∗ 1 ∗ → ∗ 1 ∗ → .
1 1 1 1
We prove an equivalence result for M -recurrent ∗ ∗ ∗ ∗ ∗ ∗
IFS’s and BCA’s. Consider an M -recurrent IFS Here a ∗ stands for either 0 or 1. We extend σ to a
with attractor (C1 , . . . , Cr ) and fix a set C = Ci for substitution v → σ(v) = w with v, w ∈ AZ×Z in the
some 1 ≤ i ≤ r. Then we can find a BCA with set obvious way (see also Sec. 4). So for example
of types T and a set S ⊆ T such that the associated
sets (Kn (S))n≥0 converge to C. On the other hand, w2i, 2j+1 w2i+1, 2j+1 10
σ(vij ) = =
consider a BCA with set of types T and let S ⊆ T be w2i, 2j w2i+1, 2j 11
such that the associated sets (Kn (S))n≥0 converge if vij = 1 and vi+1, j+1 = 0 .
The Boundary of the Attractor of a Recurrent Iterated Function System 79
Fig. 1 From top left to bottom right, the approximating sets C (0) , . . . , C (5) of the tile C generated by the two-recurrent IFS
in Sec. 2.
Fig. 2 From top left to bottom right, the sets K0 (∂S), . . . , K5 (∂S) approximating the boundary ∂C of the tile C from Sec. 2.
Sr
Let s be a type in T and let v ∈ AZ×Z be such of {1, . . . , m} such that j=1 Qij is non-empty for
that the type of v0 is s. Define the offspring matrix each i. Then the pair
M = (mst )s, t∈T by
((fk )1≤k≤m , (Qij )1≤i,j≤r )
mst = Card ({(i, j) : i, j ∈ {0, 1},
is called a recurrent iterated function system.
the type of (σ(v))ij is t}) . Let H be the set of compact subsets of Rd and
In this example, M is a 16 × 16 matrix. Let M∂S let H0 be H without the empty set. On the space
denote the matrix M restricted to ∂S (which is a Hr , define g by
14 × 14 irreducible matrix), and let λ∂S denote the
g(A1 , . . . , Ar )
largest eigenvalue of M∂S . If we denote the bound-
ary of C by ∂C, then by Theorem 2, we can express [
r [ [
r [
the dimension of ∂C in terms of λ∂S : = fk (Aj ), . . . , fk (Aj ) .
log λ∂S j=1 k∈Q1j j=1 k∈Qrj
dimH ∂C = dimH K(∂S) = .
log 2 In Bandt (1989)8 it is shown that, restricted to H0r ,
A simple computation shows that λ∂S = 3, g has a unique fixed point (C1 , . . . , Cr ), i.e.
hence the dimension of the boundary of C equals
log 3/ log 2. g(C1 , . . . , Cr ) = (C1 , . . . , Cr )
and (C1 , . . . , Cr ) is an attractor in H0r , i.e. for all
(A1 , . . . , Ar ) ∈ H0r
3. RECURRENT ITERATED
FUNCTION SYSTEMS lim gn (A1 , . . . , Ar ) = (C1 , . . . , Cr )
n→∞
Let m, r ≥ 1 be integers and f1 , . . . , fm be contrac- where the limit is with respect to the Hausdorff
tions on Rd . For 1 ≤ i, j ≤ r, let Qij be a subset metric.
The Boundary of the Attractor of a Recurrent Iterated Function System 81
(n)
We shall always approximate the Ci by sets Ci It is easily proved by induction that for n ≥ 0
given by
(n) [
r
(n)
(n)
(C1 , . . . , Cr(n) ) = g ({0 }, . . . , {0 })
n d d
for n ≥ 0 Ci = Ci (j), 1 ≤ i ≤ r.
j=1
where 0d denotes the origin in Rd .
(n)
Consider the following special case of a recur- For an M -recurrent IFS, the sets Ci (j) satisfy
rent IFS. For 1 ≤ i, j ≤ r, let DijS be finite a nice recursion relation.
subsets of Zd , such that forS each i, rj=1 Dij is
non-empty. Let mS = Card( 1≤i, j≤r Dij ), and let Lemma 1. For i, j ∈ {1, . . . , r} and n = 0, 1, . . . ,
b : {1, . . . , m} → 1≤i, j≤r Dij be a bijection. De- we have
fine contractions f1 , . . . , fm by
[
r
(Ci (h) + M −(n+1) Dhj ) .
−1 (n+1) (n)
fk (x) = M (x + b(k)) Ci (j) =
h=1
where M ≥ 2 is an integer and define
Proof. The proof will be by induction. For n = 0,
Qij = {k : b(k) ∈ Dij } both sides of the equation equal M −1 Dij . Assume
that the lemma is proved for n − 1. Then, using the
to obtain the recurrent IFS ((fk ), (Qij )). The pair
induction hypothesis in the third step,
(M, (Dij )1≤i, j≤r )
(n+1) [
r [ (n)
will be called an M -recurrent IFS. Ci (j) = fk (Cl (j))
l=1 k∈Qil
For an M -recurrent IFS, the fixed point equa-
tions can be written as [r
−1 (n)
= M (Cl (j) + Dil )
[
r [ l=1
Ci = fk (Cj ) [r [
r
M −1
(n−1)
j=1 k∈Qij = (Cl (h)
[
r [ l=1 h=1
!
= M −1 (Cj + b(k))
−n
j=1 k∈Qij +M Dhj ) + Dil
[
r [
= M −1 (Cj + e) [
r [
r
M −1 (Cl
(n−1)
j=1 e∈Dij
= (h) + Dil )
h=1 l=1
!
for 1 ≤ i ≤ r, which is equivalent to −(n+1)
+M Dhj
[
r
M Ci = (Cj + Dij ) [
r
(Ci (h) + M −(n+1) Dhj ) .
(n)
j=1 =
h=1
where Cj + Dij is short for {x + y: x ∈ Cj , y ∈ Dij }.
By convention, Cj + Dij = ∅ when Cj or Dij is
empty.
(n)
To obtain a refinement of the sets Ci (for an 4. DEFINITION OF A BCA
(n)
example see Fig. 3 in Sec. 9), define sets Ci (j) for For notational convenience, the definitions in this
1 ≤ i, j ≤ r and n ≥ 0 by section are for dimension 1, but they can be easily
extended to higher dimensions. Let A be a finite set
(n)
((C1 (1), . . . , Cr(n) (1)) = gn ({0d }, ∅, . . . , ∅) with a designated element denoted 0 and let M, N
(n) be integers with M ≥ 2 and N ≥ 0. We will refer to
((C1 (2), . . . , Cr(n) (2)) = gn (∅, {0d }, ∅, . . . , ∅) A as the alphabet, to M as the substitution length
.. .. .. and to N as the interaction length.
. . .
Let σ be a map from A2N +1 to AM such that
(n)
((C1 (r), . . . , Cr(n) (r)) = gn (∅, . . . , ∅, {0d }) . σ(02N +1 ) = 0M (here 0M denotes the word of length
82 F. M. Dekking & P. van der Wal
for n = 1, 2, . . . . Define a branching cellular automa- The following two results concerning the con-
ton (BCA) as the quintuple vergence of the sequence (Kn (S))n≥0 follow from
Ref. 7, Theorem 2.
(A, M, N, σ, u)
Proposition 1. Let S be a subset of T . If S =
Z
where u ∈ A consists of only finitely many nonzero S≥, S = S≤ or S is a communicating class, then
letters and serves as the starting word of the (Kn (S))n≥0 converges.
substitution.
Let R be the maximum of 1 and d M −1 e and
MN Let u be the starting word of the BCA, u having
define T = A 2R+1 to be the set of types. A let- only finitely many nonzero letters. Define
ter wk in a word w ∈ AZ is said to be of type V = {t ∈ T : there is a type t letter in u}
t ∈ T if wk−R · · · wk+R = t. The type of wk in w
determines the types of the letters (σ(w))kM , . . . , and
(σ(w))(k+1)M −1 in σ(w). By 0̄, we denote the type U(S) = {U ⊆ T : U is a nontrivial communicating
02R+1 ∈ T .
Fix a set S ⊆ T which does not contain the type class in S ≥ ∩ V ≤ } .
0̄ and define for n ≥ 0
Theorem 1. Assume that (Kn (S))n≥0 converges
−n
Kn (S) = M {k ∈ Z: type of (σ (u))k
n to K(S). Then
[
is an element of S} . K(S) = K(U ) .
U ∈U (S)
Note that the sets Kn (S) are closed and uniformly
bounded. Fix a set S ⊆ T such that (Kn (S))n≥0 converges.
Let s ∈ T and let w ∈ AZ be such that the type For a communicating class U ⊆ U(S), let λU denote
of w0 is s. Define for t ∈ T the Perron-Frobenius eigenvalue of U . If U(S) = ∅,
then K(S) = ∅. If U(S) 6= ∅, define
J(s, t) = {k ∈ {0, . . . , M − 1}: λS = max λU .
U ∈U (S)
the type of (σ(w))k is t} .
The following theorem concerning the dimension of
The offspring matrix M = (mst )s,t∈T is defined by the limit set K(S) follows from Ref. 7, Theorem 4.
6. FROM M-RECURRENT IFS {1, . . . , r} and with designated element ∅ such that
TO BCA C = K(S), where S = {t ∈ T : t 6= ¯ ∅}.
We consider the attractor (C1 , . . . , Cr ) of a d- The theorem will be proved in two steps. First,
dimensional M -recurrent IFS (M, (Dij )1≤i, j≤r ), i.e. we show that K(S) = K(S̃), where
(C1 , . . . , Cr ) is the unique non-empty solution of
[
r S̃ = {t ∈ T : t0d 6= ∅}
M Ci = (Cj + Dij ), 1 ≤ i ≤ r.
j=1 t0d denoting the central letter of type t. Then we
prove that C = K(S̃).
Fix an i0 ∈ {1, . . . , r} and write C for Ci0 . In
this section, we will construct a BCA B with al- Lemma 2. The sequences (Kn (S))n≥0 and
phabet A the set of all subsets of {1, . . . , r}, with (Kn (S̃))n≥0 converge to the same limit set.
designated element ∅, and with set of types T such
that C = K(S), where S = {t ∈ T : t 6= ¯∅}. Here, ¯ ∅ Proof. (cf. the introduction to Corollary 1 in
denotes the type in T with all entries equal to ∅. Ref. 7) By Proposition 1, (Kn (S))n≥0 converges to
For x = (x1 , . . . , xd ) ∈ Zd , define K(S). Since S̃ ⊆ S, we have for all n ≥ 0 that
Kn (S̃) ⊆ Kn (S). If M −n x is a point in Kn (S),
x mod M = (x1 mod M, . . . , xd mod M ) . then (σ n (u))x has a type s in S. Since s 6= 0̄, we
can find a y ∈ BR (σ n (u), x) such that
√ (σ n (u))y 6= 0.
Construction 1. Define a BCA as follows. Let the So M −n y ∈ Kn (S̃) and kx − yk ≤ dR. Therefore,
alphabet A be the set of all subsets of {1, . . . , r} and
√
let ∅ be the designated element. Choose S
the substi- Kn (S) ⊆ Kn (S̃) + {x ∈ Rd : kxk ≤ dRM −n } .
tution length to be M . Define D = 1≤i, j≤r Dij .
Let the interaction length N be the smallest integer It thus follows from the definition of the Hausdorff
l which satisfies metric that (Kn (S̃))n≥0 converges to K(S).
[
r
Lemma 3. Let n ≥ 0 and 1 ≤ j ≤ r. Then
Hj = (M Gi + Dij ) .
i=1
[
r
Kn+1 (Sj ) = (Kn (Si ) + M −(n+1) Dij ) .
The word σ(v) in AM ×···×M , indexed by {0, . . . , i=1
M − 1}d , is then defined by
S
(σ(v))x = {i ∈ {1, . . . , r} : x ∈ Hi } Proof. To prove that ri=1 (Kn (Si ) + M −(n+1) Dij )
is contained in Kn+1 (Sj ), let M −(n+1) x be an ele-
x ∈ {0, . . . , M − 1}d . ment of Kn (Si ) + M −(n+1) Dij for some 1 ≤ i ≤ r.
Write
Let the starting word u be defined by u0d = {i0 }
and ux = ∅ for x 6= 0d . Denote the constructed x = M y + e with y ∈ M n Kn (Si ) and e ∈ Dij .
BCA (A, M, N, σ, u) by B.
Let
Theorem 4. Let B be the BCA constructed in Con-
struction 1 and denote its set of types by T . Then 1
B is a BCA with alphabet A the set of all subsets of z= (x − x mod M )
M
84 F. M. Dekking & P. van der Wal
Hence [
r
M −(n+1) x ∈ (Kn (Si ) + M −(n+1) Dij ) .
y − z ∈ Gi = {a ∈ {−N, . . . , N } : i ∈ va }
d
i=1
= K1 (Sj ) . ∂C = B0 .
The Boundary of the Attractor of a Recurrent Iterated Function System 87
D21 = {(−1, 0), (0, −1), (1, 0), (1, 1)} . where λ = max(λU : U ∈ U(∂S)).
Since Card(T ) = 425 , we will first project the set
Let (C1 , C2 ) denote its attractor. In Fig. 4, the first of types on the smaller set T̃ = A2×2 , which consists
eight approximating sets for C1 are shown. We will of only 256 elements. Let π be the projection
calculate the dimension of the boundary of C1 . By
means of Construction 1, we obtain a BCA with π : T → T̃
t(−2,2) ... t(2,2)
A = {∅, {1}, {2}, {1, 2}}, M = 2, N =1 .. .. t(0,1) t(1,1)
. . 7→
2 t(0,0) t(1,0) .
and the starting word u ∈ AZ is such that u(0,0) = t(−2,−2) ··· t(2,−2)
{1} and ux = ∅ for x 6= (0, 0). Hence R = d M
MN
−1 e =
2 and T = A5×5 . Let M = (mst )s, t∈T denote the offspring matrix
We will denote ¯∅, the type in T with all entries and fix s̃, t̃ ∈ T̃ . Then it can be checked that
equal to ∅, by ∅5×5 . If X
mst
S = {t ∈ T : t 6= ∅5×5 } t:π(t)=t̃
(0) (5)
Fig. 4 From top left to bottom right, the sets C1 , . . . , C1 approximating the first component C1 of the attractor generated
by the two-recurrent IFS in Example 2.
The Boundary of the Attractor of a Recurrent Iterated Function System 89
( )
does not depend on s for all s ∈ T with π(s) = s̃. {1, 2} {1, 2}
Ũ4 =
Define a matrix M̃ = (m̃s̃t̃ )s̃, t̃∈T̃ by {1, 2} {1}
X
m̃s̃t̃ = mst Ũ5 = Ṽ ≤ \(Ũ1 ∪ Ũ2 ∪ Ũ3 ∪ Ũ4 ) .
t:π(t)=t̃
where s ∈ T is such that π(s) = s̃. Define a set We find that U(∂ S̃) = {Ũ5 } and λ = λ̃ = λŨ5 =
Z ⊆ ∂S by 3.0491 . . . . Hence
Z = {t ∈ T : t → ∅5×5 , π(t) 6= ∅2×2 }
log λ
where ∅2×2 is short for dimH ∂C1 = = 1.6084 . . . .
log 2
∅ ∅
∈ A2×2 .
∅ ∅
Analogously to Lemma 2, one can prove that ACKNOWLEDGEMENT
K(Z) = K(∂S). Hence λ = max(λU : U ∈ U(Z)).
Let We thank the referees for their remarks which have
S̃ = {t̃ ∈ T̃ : t̃ 6= ∅ 2×2
} led to several improvements in this paper.
and
∂ S̃ = {t̃ ∈ T̃ : t̃ → ∅2×2 , t̃ 6= ∅2×2 } REFERENCES
then since s → t for some s, t ∈ T implies π(s) →
1. P. Duvall and J. Keesling, “The Hausdorff Dimen-
π(t), we have that ∂ S̃ = π(Z).
sion of the Boundary of the Lévy Dragon,” in Ge-
We claim that λ̃ = λ. To see this, let s ∈ T ometry and Topology in Dynamics (Winston-Salem,
be a type such that U (s), the communicating class NC, 1998/San Antonio, TX, 1999), pp. 87–97, Amer.
containing s, is an element of U(Z). Let λπ(U (s)) be Math. Soc., Providence, RI, 1999.
the Perron-Frobenius eigenvalue of M̃ restricted to 2. P. Duvall, J. Keesling and A. Vince, “The Hausdorff
π(U (s)). Then by a lemma of Furstenberg, λU (s) = Dimension of the Boundary of a Self-Similar Tile,”
λπ(U (s)) . Since π(U (s)) ⊆ U (π(s)) ∈ U(∂ S̃), we J. London Math. Soc. (2) 61(3), 748–760 (2000).
3. X. He, Ph.D. Thesis, Chinese University of Hong
have
Kong.
λU (s) = λπ(U (s)) ≤ λU (π(s)) 4. K.-S. Lau and Y. Xu, “On the Boundary of At-
tractors with Non-Void Interior,” Proc. Amer. Math.
and hence λ ≤ λ̃. If we consider a non-trivial com- Soc. 128(6), 1761–1768 (2000).
municating class Ũ ∈ U(∂ S̃), then we can find a 5. R. S. Strichartz and Y. Wang, “Geometry of Self-
communicating class U ∈ U(Z) such that π(U ) = Affine Tiles. I,” Indiana Univ. Math. J. 48(1), 1–23
Ũ . Again by Furstenberg’s Lemma, λŨ = λU and (1999).
hence λ̃ ≤ λ. We may conclude that λ̃ = λ. 6. A. Vince, “Self-Replicating Tiles and Their Bound-
The reduced set of types T̃ has only 256 states. ary,” Discrete Comput. Geom. 21(3), 463–476
Constructing and analyzing M̃ with a computer (1999).
package yields the following. Let V ⊆ T denote 7. F. M. Dekking and P. van der Wal, “Fractal Perco-
lation and Branching Cellular Automata,” Probab.
the set of types appearing in the starting word
Theory Rel. Fields, to appear (2001).
u and define Ṽ = π(V ). The set Ṽ ≤ contains
8. C. Bandt, “Self-Similar Sets. III. Constructions With
110 elements and five non-trivial aperiodic classes Sofic Systems,” Monatsh. Math. 108(2-3), 89–102
Ũ1 , . . . , Ũ5 , where (1989).
( )
∅ ∅ 9. K.-S. Lau and S.-M. Ngai, “Multifractal Measures
Ũ1 = and a Weak Separation Condition,” Adv. Math.
∅ ∅ 141(1), 45–96 (1999).
( ) 10. M. P. W. Zerner, “Weak Separation Properties for
{1, 2} {1, 2}
Ũ2 = Self-Similar Sets,” Proc. Amer. Math. Soc. 124(11),
{1, 2} {1, 2} 3529–3539 (1996).
( ) 11. H. Rao and Z. Wen, “A Class of Self-Similar Fractals
{2} {1, 2} with Overlap Structure,” Adv. Appl. Math. 20(1),
Ũ3 =
{1, 2} {1, 2} 50–72 (1998).