Contents lists available atScienceDirect
Journal of Mathematical Analysis and Applications
www.elsevier.com/locate/jmaa
On single-matrix graph-directed iterated function systems ✩
Luo Jun
a, Ya-Min Yang
b, ∗
aMathematics Department, The Chinese University of Hong Kong, Hong Kong bCollege of Sciences, Huazhong Agriculture University, Wuhan, China
a r t i c l e i n f o a b s t r a c t
Article history:
Received 20 March 2009 Available online 21 July 2010 Submitted by C.E. Wayne Keywords:
Graph IFS Weak norm Open set condition
We study graph-directed function systems where each contraction in the system has the form fe(x)=A−1(x+de), where A is an expanding matrix. We show that a certain discreteness implies the open set condition, and the latter implies the strong open set condition. Hausdorff measures and dimensions (w.r.t. a weak norm) of the invariant sets are investigated. The stationary Markov measures of the system are proved to be translation invariant.
©2010 Elsevier Inc. All rights reserved.
1. Introduction
1.1. Graph IFS. Let
(
V, Γ )
be a directed graph with vertex setV= {
1, . . . ,
N}
and edge setΓ
. We call{
fe;
e∈ Γ }
, a collec- tion of contractions fe:
Rd→
Rd, agraph-directed iterated function system(graph IFS).Let
Γ
i j be the set of edges from vertexi to j, then there are unique non-empty compact sets{
Ei}
Ni=1satisfying [27]Ei
=
Nj=1
e∈Γi j
fe
(
Ej),
1iN.
(1.1)We call
(
E1, . . . ,
EN)
theinvariant setsof the graph IFS.The graph IFS is said to satisfy theopen set condition(OSC), if there exist open setsU1
, . . . ,
UN such that Nj=1
e∈Γi j
fe
(
Uj) ⊂
Ui,
1iN,
and the left-hand side are non-overlapping unions [14,27]. In addition, if Ui
∩
Ei= ∅
for all 1iN, then we say the graph IFS satisfies thestrong open set condition(SOSC) [31].Let us define M
= (
mi j)
1i,jN to be the associated matrix of(
V, Γ )
, that is,mi j=
#Γ
ji counts the number of edges from j toi. We say(
V, Γ )
isprimitiveifMis a primitive matrix, i.e.,Mn is a positive matrix for largen. From now on, we will always assume the graph(
V, Γ )
in consideration is primitive.✩ The authors are supported by CNSF 10631040.
*
Corresponding author.E-mail addresses:yangym09@mail.hzau.edu.cn,yangym05@mails.tsinghua.edu.cn(Y.-M. Yang).
0022-247X/$ – see front matter ©2010 Elsevier Inc. All rights reserved.
doi:10.1016/j.jmaa.2010.07.001
1.2. Single-matrix IFS. If a graph
(
V, Γ )
contains only one vertex, then the graph IFS{
fe;
e∈ Γ }
simplifies to aniterated function system{
fj}
Nj=1 ([14]). If all fjhave the formfj
(
x) =
A−1(
x+
dj),
1jN,
(1.2)where A is ad
×
dexpanding matrix anddj∈
Rd, then we call{
fj}
Nj=1 asingle-matrix IFS. (A matrix isexpandingif all its eigenvalues have moduli larger than 1.) Let us denote byq= |
detA|
.Some special cases of system (1.2) define number systems. The study of such number systems goes back as early as 1970’s [18,28,15]. See also a recent survey [6].
Another special case of (1.2) is the so-calledself-affine tiling system, when N
=
q:= |
detA|
and the OSC holds. The self- affine tiling system has been studied by many authors [4,16,17,11,22–24,3].In the above studies, we generally concern the following questions:
(Q1) When does the system satisfy OSC?
(Q2) If the OSC holds, how to compute the Hausdorff dimension and Hausdorff measure of E, the invariant set of(1.2)?
(Q3) Does OSC imply SOSC?
•
If Ais a similitude, these questions have satisfactory answers.LetD
= {
d1, . . . ,
dN}
be the set of translations in (1.2). Define Dn=
An−1D+ · · · +
AD+
D,
n1.
Let dimE denote the Hausdorff dimension ofE, and letHs
(
E)
be thes-dimensional Hausdorff measure ofE. A setGis said to ber-uniformly discreteif|
x−
y| >
rfor anyx,
y∈
G. Then(i) OSC holds if and only if#Dn
=
NnandDnis r-uniformly discrete for some r>
0independent of n.(ii) If OSC holds, then s
=
dimE=
dloglogqN and0<
Hs(
E) < +∞
. (iii) OSC implies SOSC.Especially in the self-affine tiling system case, many deep results on (Q1) and (Q2) have been obtained by Fourier transformation method [16,11,24]. For (Q3), Schief [31] gives a positive answer for general self-similar IFS.
•
In case of Ais not a similitude andNdoes not equal to|
detA|
, it is much more complicated. Actually, in this case the second assertion does not hold. McMullen’s carpets provide counter-examples [26].To overcome the difficulty that A is not similitude, Lemarié-Rieusset [20] introduce a weak norm
ω
of Rd such thatω (
Ax) =
q1/dω (
x)
. Under the weak norm, Ais a ‘similitude’. He and Lau [12] introduce Hausdorff dimension and Hausdorff measure w.r.t. the weak norm, which will be denoted by dimω andHsw respectively. [12] proved that the above results still hold except the second assertion is replaced by(ii) If OSC holds, then s
=
dimωE=
dlogqlogN and0<
Hsω(
E) < +∞
.1.3. Single-matrix graph IFS. In this paper, we investigate the graph IFS
{
fe;
e∈ Γ }
with the formfe
(
x) =
A−1(
x+
de),
(1.3)where A is a d
×
d expanding matrix andde∈
Rd. The paper is motivated by the questions posed by Professor S. Ito in a conference in Beijing in 2006. To state the questions, we need some notations.Denote by
Γ
i jnthe paths from vertex ito vertex j with lengthn. For I=
e1· · ·
en∈ Γ
i jn, set fI(
x) :=
fe1◦
fe2◦ · · · ◦
fen(
x)
and definedI
:=
An−1de1+
An−2de2+ · · · +
Aden−1+
den,
then fI
(
x)
has the form: fI(
x) =
A−n(
x+
dI)
. Set Dni j:=
dI
;
I∈ Γ
i jn.
(1.4)(P1) Does r-uniformly discreteness ofDni jimply OSC?
(P2) Does OSC imply SOSC?
(P3) Let
μ
ibe the stationary Markov measures on Ei, areμ
i‘translation invariant’ on Ei?Stationary Markov measures will be introduced in Section 2. For a measure
μ
supported by a set E, we sayμ
is translation invarianton E if for anyB1,
B2⊂
E andB1=
B2+
x, it holds thatμ (
B1) = μ (
B2)
.1.4. Main results. In this paper, we generalize the results of [12] to single-matrix graph IFS. These results are worth to be documented, since graph IFS are frequently encountered in practice, for example, in the study of IFS with overlap structures
[30,7,13], in self-similar tiling theory [33,34,5], in Rauzy geometry [29,2,32,19], etc. Our results are in great general form and contain many previous results as special case. Also, they give satisfactory answers to the questions of Professor Ito.
Theorem 1.1.For graph IFS(1.3), the following are equivalent:
(i) OSC.
(ii) #Dni j
=
#Γ
i jnand there is an r>
0such thatDni jis r-uniformly discrete for all1i,
jN and n1.(iii) SOSC.
In Section 6, we show by examples that OSC does not imply SOSC if the system is not a single-matrix system.
Theorem 1.2.For graph IFS(1.3), let
λ
be the maximal eigenvalue of M, the associate matrix of(
V, Γ )
. If OSC holds, then for any 1iN,(i) s
=
dimωEi=
dlogλ/
logq.(ii) 0
<
Hωs(
Ei) < +∞
.(iii) The right-hand side of(1.1)is a disjoint union in sense of the measureHsw.
Remark 1.3. For Theorem 1.1 and Theorem 1.2, the case that A is a similitude has been studied by Li [21]. The case that (1.2) is a tiling system has been settled by Lagarias and Wang [25].
The next theorem answers question (P3). The translation invariance of a measure on a fractal set has never been con- sidered before. Theorem 1.4 seems to be the first result of this type. Moreover, although the result seems very nature, it is hard to prove without using the weak norm technique.
Theorem 1.4.For1iN, the stationary Markov measure
μ
iis equal to a−i1Hsω|
Ei, where ai=
Hsω(
Ei)
. Consequently,μ
i is translation invariant.1.5. Applications. Recently, Furukado, Ito and Rao [9] apply the above results to the study of atomic surfaces of hyperbolic substitutions and obtain some interesting results. According to a substitution
σ
, [9] constructs a single-matrix graph IFS.They define a fractal domain-exchange transformation
Φ
on E=
Ni=1Ei, the union of the invariant sets.
Φ
preserve the stationary measureμ
by our results. [9] shows that(
E, Φ, μ )
is (measure theoretically) isomorphic to the substitution dynamical systemdefined byσ
, i.e., there exists a measure-preserving bijection between two systems except a measure zero set.Akiyama and Loridant [1] apply our results to study the parametrization of boundaries of self-affine tiles.
The paper is organized as follows: In Section 2, we recall some known results on Markov measures. In Section 3, we give a brief introduction to weak norm. Theorem 1.1 is proved in Section 4, Theorem 1.2 and Theorem 1.4 are proved in Section 5. In Section 6 we give some remarks on SOSC.
2. SOSC and Markov measures
In this section, we consider general graph IFS (1.1).
2.1. Markov measures. LetM(Rd
)
denote the collection of probability Borel measures onRdwith bounded support.Let p
: Γ → (
0,
1]
be a function satisfying Nj=1
e∈Γi j
pe
=
1,
1iN.
We shall call paprobability weighton graph
(
V, Γ )
.It is well known that there is a unique vector
( μ
1, . . . , μ
N) ∈ (
M(
Rd))
N satisfying the equationsμ
i=
Nj=1
e∈Γi j
pe
· μ
j◦
fe−1,
1iN,
(2.1)we call
μ
1, . . . , μ
N theMarkov measuresdetermined by the weights{
pe;
e∈ Γ }
. Remark 2.1. For an IFS, Eq. (2.1) simplifies toμ =
Ni=1pi
· μ ◦
fi−1.
The measureμ
is called a self-similar measurewhen the mappings fiare similitudes [14].2.2. Measures on symbolic space and their projections. Let
Σ
i∗=
N j=1
k1
Γ
i jk denote the collection of all finite paths with initial state (or vertex) i andΣ
∗=
Ni=1
Σ
i∗. DenoteΣ
iN to be the collection of infinite paths with initial state i, denoteΣ =
Ni=1
Σ
iN be the set of all infinite paths.For I
=
e1e2· · ·
ek∈ Γ
i jk, we define b(
I) =
i be the initial state of I, and t(
I) =
j be the terminate state of I. Denote EI:=
fI(
Ej)
where j=
t(
I)
.Denote
[
I] := {
e1e2· · ·
ek· · · ∈ Σ :
e1e2· · ·
ek=
I}
, and call it a cylinder ofΣ
. GivenI,
J∈ Σ
∗ witht(
I) =
b(
J)
, denote by I J the concatenation of Iand J.LetS be the shift operator on
Σ
whereS(
e1e2e3· · · ) =
e2e3· · ·
. LetPi be the probability measure onΣ
iN satisfying the relationsP
i[
e1· · ·
en] =
pe1· · ·
pen,
e1· · ·
en∈ Σ
i∗.
(2.2) According to formula (1.1), it is seen that{
fe1···en(
Et(en)) }
n1 is a decreasing sequence of compact sets and their inter- section is a single point in Eb(e1). Define a projectionπ : (Σ
1N, . . . , Σ
NN) → (
Rd, . . . ,
Rd),
whereπ
i: Σ
iN→
Rd is defined byπ
i(
e1e2· · ·
en· · ·)
=
n1
fe1···en
(
Et(en)).
Then fore
∈ Γ
i j and J∈ Σ
∗j,fe−1
◦ π
i(
e J) = π
j◦
S(
e J).
(2.3)Set
μ
i=
Pi◦ π
i−1,
thenμ
iis a probability measure supported by Ei.Proposition 2.2.The projection measures
( μ
1, . . . , μ
N)
satisfy Eq.(2.1)and thus are Markov measures on Ei. Proof. Pick any Borel set AinRd. For any edgee∈ Γ
i j, setHe
= π
−j1◦
fe−1(
A).
Then He
⊂ Σ
Nj and thuse J∈ Σ
iN for any J∈
He. Denote e He:= {
e J: J∈
He}
, thenPi(
e He) =
pePj(
He)
by the definition ofPi’s.First we show that
π
i−1(
A) =
Nj=1
e∈Γi j
e He
.
(2.4)Suppose I
∈ π
i−1(
A)
and the initial edge ofIise∈ Γ
i j, then I∈ π
i−1(
A) ⇐⇒
fe−1◦ π
i(
I) ∈
fe−1(
A)
⇐⇒ π
j◦
S(
I) ∈
fe−1(
A)
by formula
(
2.
3)
⇐⇒
S(
I) ∈ π
−j1◦
fe−1(
A) =
He⇐⇒
I∈
e He.
So formula (2.4) holds. Applying Pito both sides of (2.4), we obtain
μ
i(
A) =
Nj=1
e∈Γi j
P
i(
e He) =
Nj=1
e∈Γi j
pe
◦ P
j(
He)
=
Nj=1
e∈Γi j
pe
◦ P
j◦ π
−j 1◦
fe−1(
A)
=
Nj=1
e∈Γi j
pe
◦ μ
j◦
fe−1(
A). 2
Two paths are said to becomparableif one of them is a prefix of the other one; otherwise, they are incomparable. The following measure separation property, proved by Fan and Lau [10, Theorem 2.2], illustrates the importance of SOSC.
Proposition 2.3.Suppose a graph IFS
{
fe;
e∈ Γ }
satisfies the strong open set condition with open sets{
Ui}
1iN. Letμ
ibe Markov measures on Ei,1iN. Then(i)
μ
i(
Ui) =
1,(ii)
μ
i(
EI∩
EJ) =
0for any incomparable I,
J∈ Σ
i∗.Question.Does the inverse of Proposition2.3hold? That is, does the measure separation property imply SOSC?
2.3. Stationary Markov measures. Recall thatM
= (
mi j)
1i,jN is the associated matrix of graph(
V, Γ )
. ThenMis a non- negative primitive matrix. Let(
c1, . . . ,
cN)
be the left eigenvector of M corresponding to the maximal eigenvalueλ
. By Perron–Frobenius Theorem,λ >
0 and(
c1, . . . ,
cN)
is a positive vector. Let us assume thatc1+ · · · +
cN=
1.Fore
∈ Γ
i j, set pe=
cjci
λ
−1.
It is easy to see that
b(e)=ipe
=
1 for eachi. Therefore{
pe;
e∈ Γ }
is a probability weight. LetP1, . . . ,
PN be the measures onΣ
1N, . . . , Σ
NNdefined by formula (2.2), and letμ
1, . . . , μ
N be the Markov measures on E1, . . . ,
EN defined by the weight{
pe;
e∈ Γ }
. We shall callμ
1, . . . , μ
N thestationary Markov measuresof the graph IFS{
fe;
e∈ Γ }
.Lemma 2.4.Suppose the SOSC holds for a graph IFS
{
fe;
e∈ Γ }
. Then for any I∈ Γ
i jk, we haveμ
i fI(
Ej) =
cj ciλ
−k.
Proof. It is obvious that
μ
i(
EI)
Pi( [
I] )
.The strong open set condition implies that
μ
i(
EI∩
EJ) =
0 for any incomparable I,
J∈ Σ
i∗ (Proposition 2.3). Henceμ
i(
EI) = μ
i EI\ ∪
EJ
; |
J| = |
I|,
J=
I,
J∈ Σ
i∗P
i[
I] .
Therefore
μ
i fI(
Ej) = μ
i(
EI) = P
i[
I] =
pe1· · ·
pek=
cj ciλ
−k. 2
3. Pseudo-norm
Let A be a d
×
dreal expanding matrix with|
detA| =
q. With respect to A, [20] defines a pseudo normω
onRd as follows (see also [12]).Denote B
(
x,
r)
the open ball with centerxand radiusr. Then V=
A(
B(
0,
1)) \
B(
0,
1)
is an annular region. Choose any 0< δ <
12 and anyC∞ functionφ
δ(
x)
with support inB(
0, δ)
such thatφ
δ(
x) = φ
δ( −
x)
andφ
δ(
x)
dx=
1, define a pseudo normω (
x)
inRd byω (
x) =
n∈Z q−n/dh
Anx
,
(3.1)whereh
(
x) = χ
V∗ φ
δ(
x)
is the convolution of the characteristic functionχ
V andφ
δ(
x)
. We list some basic properties ofω (
x)
.Proposition 3.1.(See [12].) The
ω (
x)
is a C∞function onRdand satisfies (i)ω (
x)
0;ω (
x) =
0if and only if x=
0.(ii)
ω (
x) = ω (−
x)
.(iii)
ω (
Ax) =
q1/dω (
x) ω (
x)
for all x∈
Rd.(iv) There exists
β
such thatω (
x+
y) β
max{ ω (
x), ω (
y)}
for any x,
y∈
Rd.Item (iii) says that matrix A is a similitude in this weak norm, a fact plays a central role in [12] as well as in our paper.
The next proposition shows that the pseudo-norm
ω (
x)
is comparable with the Euclidean norm|
x|
throughλ
max andλ
min, the maximal and minimal modulus of the eigenvalues of A.Proposition 3.2.(See [12].) For any0
< ε < λ
min−
1, there exists C>
0 (depends onε
)such that C−1|
x|
lnq/dln(λmax+ε)ω (
x)
C|
x|
lnq/dln(λmin−ε),
if|
x| >
1,
C−1
|
x|
lnq/dln(λmin−ε)ω (
x)
C|
x|
lnq/dln(λmax+ε),
if|
x|
1.
Let E be a subset of Rd. Define diamωE
=
sup{ ω (
x−
y)
: x,
y∈
E}
be theω
-diameter of E. Now we can define a Hausdorff measure of E with respect to the pseudo normω (
x)
.Hsω,δ
(
E) =
inf ∞i=1
(
diamωEi)
s: E⊂
i
Ei
,
diamωEiδ
.
SinceHsω,δ
(
E)
is increasing whenδ
tends to 0, we can define Hsω(
E) =
limδ→0Hωs,δ
(
E).
It is shown thatHsω
(
E)
is an outer measure and is a regular measure on the family of Borel subsets onRd. It is translation invariant and has the scaling property; precisely,Hsω
(
E+
x) =
Hsω(
E)
and HsωA−1E
=
q−s/dHsω(
E).
(3.2)A Hausdorff dimension with respect to the pseudo norm thus can be defined to be dimωE
=
infs
;
Hsω(
E) =
0=
sups
;
Hsω(
E) = ∞ .
The relation of dimωE and the classical Hausdorff dimension dimH
(
E)
has been studied in [12].4. Uniform discreteness and OSC
We prove Theorem 1.1 in this section. Our proof is an analogue of [12], where the basic idea belongs to [31].
Proof of Theorem 1.1. (i)
⇒
(ii). Suppose the open set condition holds andU1, . . . ,
UN are open sets such that Nj=1
e∈Γi j
fe
(
Uj) ⊂
Ui,
1iN,
(4.1)and the left-hand side of (4.1) is a disjoint union.
First we note that fI
(
Uj) ⊂
Uiholds forI∈ Γ
i jn.Let I
=
e1· · ·
en and I=
e1· · ·
en be two elements ofΓ
i jn. Using (4.1) repeatedly, one can show that fI(
Uj)
and fI(
Uj)
belong toUiand they are disjoint. Hence A−n(
Uj+
dI) ∩
A−n(
Uj+
dI) = ∅
and so that(
Uj+
dI) ∩ (
Uj+
dI) = ∅.
It follows thatdI−
dI=
0 and thus #Dni j=
#Γ
i jn.Set
η
j=
infx∈Rd{|
x|
:(
Uj+
x) ∩
Uj= ∅}
, thenη
j>
0 since Uj is an open set. Put r=
min{ η
j: 1 jN}
. Since(
Uj+
dI) ∩ (
Uj+
dI) = ∅
, we conclude that|
dI−
dI|
rforI,
I∈
Dni j.(ii)
⇒
(iii). This is the difficult part of this theorem.We define the
δ
-parallel body of a set E with respect to the weak normω
as follows:[
E]
δ= {
x∈
Rd:ω (
x−
y) < δ
for somey∈
E} .
(Usually[
E]
δ stands for theδ
-parallel body of E. Here we use an unusual notation to avoid confusion.)Pick any
δ >
0 and fix it. For any J∈ Γ
i jn, define GJ=
fJ[
Ej]
δ,
thenGJ
= [
EJ]
q−n/dδ by Proposition 3.1(iii). We define V(
J)
, neighbors of J, to be a collection of paths giving by V(
J) =
I
∈ Σ
i∗:|
I| =
nandEI∩
GJ= ∅ .
Note thatEI
∩
GJ= ∅
if and only if(
Et(I)+
dI) ∩ ( [
Ej]
δ+
dJ) = ∅
. We shall show that #V(
J)
have a uniform bound, namely,γ =
sup#V
(
J)
: J∈ Σ
∗< +∞ .
Let R be a real number such that
1jNEj
⊂
B(
0,
R),
and letη
be a positive number such that Nj=1
[
Ej]
δ⊂
B(
0, η )
. Then for J∈ Γ
i jn,#V
(
J) =
#I
∈ Σ
i∗:|
I| =
n, (
Et(I)+
dI) ∩
[
Ej]
δ+
dJ= ∅
#I
∈ Σ
i∗:|
I| =
n, (
Et(I)+
dI) ∩
B
(
0, η ) +
dJ= ∅
supx∈Rd
#
I
∈ Σ
i∗:|
I| =
n, (
Et(I)+
dI) ∩
B(
x, η ) = ∅
N
j=1
sup
x∈Rd
#
dI:dI
∈
B(
x, η +
R) ∩
Dni j< ∞,
by the uniform discreteness of Dni j. Therefore
γ < +∞
.Let J
∈ Σ
∗ be a path such that #V(
J) = γ
attains the maximum. We claim that if I J∈ Σ
∗, then V(
I J) =
I V(
J) :=
I L: L
∈
V(
J) .
For any L
∈
V(
J)
, we have EL∩
GJ= ∅
, hence EI L∩
GI J=
fI(
EL) ∩
fI(
GJ) = ∅
. It follows that I L∈
V(
I J)
and so that I V(
J) ⊂
V(
I J)
. By the maximality of #V(
J)
, we obtainI V(
J) =
V(
I J)
. The claim is proved.Let us assume that J
∈ Γ
hjn. Letδ
=
βδ, where the constantβ
is as in Proposition 3.1(iv). DefineUi
=
∞n=1
I∈Γihn fI J
[
Ej]
δ,
1iN.
(4.2)ClearlyUi
∩
Ei= ∅
. We shall show these open sets satisfy SOSC.Fore
∈ Γ
ik, we have fe(
Uk) =
∞n=1
I∈Γkhn feI J
[
Ej]
δ⊂
∞n=2
I∈Γihn fIJ
[
Ej]
δ⊂
Ui.
So
Nk=1
e∈Γik
fe
(
Uk) ⊂
Ui.
(4.3)It remains to show that the left-hand side of (4.3) is a disjoint union.
Suppose this is false. Then fe
(
Uk) ∩
fe(
Uk) = ∅
for somee∈ Γ
ik, e∈ Γ
ik,
e=
e (it may happen that k=
k). By the construction of the open sets, there exist two paths I1 fromktohandI2fromk tohsuch thatfeI1J
[
Ej]
δ∩
feI2J[
Ej]
δ= ∅.
Suppose yis a point in this intersection, then y
=
feI1J(
y1) =
feI2J(
y2)
for some y1
,
y2∈ [
Ej]
δ. Choosez1,
z2∈
Ej such thatω (
y1−
z1) < δ
, ω (
y2−
z2) < δ
.
Without loss of generality, we assume that
|
I1| |
I2|
. On one hand, as fI(
x) =
A−|I|(
x+
dI),
we haveω
feI1J(
z1) −
feI2J(
z2) = ω
feI1J(
z1) −
y+
y−
feI2J(
z2)
= ω
A−|eI1J|(
z1−
y1) +
A−|eI2J|(
y2−
z2)
< β
maxδ
q−|eId1J|, δ
q−|eI 2J|
d by Proposition 3.1(iv)
= βδ
q−|eId1J|=
q−|eId1J|δ.
On the other hand, for any L
∈ Σ
k∗ with length|
L| = |
I1J|
, clearlyeL∈ /
V(
eI1J) =
eI1V(
J)
; so we have feL(
Et(L)) ∩
feI1J[
Ej]
δ= ∅ .
(4.4)As
|
eL| = |
eI1J| |
eI2J|
, we have feI2J(
Ej) ⊂
L
feL
(
Et(L)),
whereLruns over the paths in
Σ
k∗ with length|
L| = |
I1J|
. Therefore, the point feI2J(
z2) ∈
feI2J(
Ej)
must belong to some cylinder of the form feL(
Et(L))
; as a result, it does not belong to feI1J([
Ej]
δ)
by (4.4). It follows thatω
feI1J(
z1) −
feI2J(
z2)
q−|eId1J|δ,
which is a contradiction.
(iii)