The recurrence dimension for piecewise monotonic maps of the interval
FRANZ HOFBAUER
Abstract. We investigate a weighted version of Hausdorff dimension introduced by V. Afraimovich, where the weights are determined by recurrence times. We do this for an ergodic invariant measure with positive entropy of a piecewise mono- tonic transformation on the interval [0,1], giving first a local result and proving then a formula for the dimension of the measure in terms of entropy and charac- teristic exponent. This is later used to give a relation between the dimension of a closed invariant subset and a pressure function.
Mathematics Subject Classification (2000):37E05 (primary); 37C45, 28A80, 37B40, 28A78 (secondary).
1.
Introduction
In the last years there was much interest in generalized notions of dimension. In [14] a theory of dimension structures is developed. Special cases of these structures are weighted versions of Hausdorff dimensions. We define them only on the interval [0,1]. Let|U|be the diameter of a subsetU of [0,1]. Forδ >0 aδ-cover of a set A⊂[0,1] is a finite or countable collection of intervals of length≤δcovering A.
For a set functionwwith values in
R
+ands,t ∈R
we define forA⊂[0,1]νs,t(A)= lim
δ→0inf
C
U∈C
w(U)s|U|t
where the infimum is taken over all δ-covers
C
of A. It follows from the general theory in [14] that for fixeds ∈R
there is a critical valuetc ∈[−∞,∞] such that νs,t(A)= ∞fort <tc andνs,t(A)=0 fort >tc. We denote this critical valuetc byds(A)and call it the dimension of the setA.Ifmis a probability measure on [0,1] and if one setsw(U)=m(U), one gets the dimension introduced by Olsen in [13] and called multifractal generalization of Research for this paper was partially carried out while the author was visiting the Max Planck Institut for Mathematics in Bonn as a participant in the activity on Algebraic and Topological Dynamics.
Pervenuto alla Redazione il 3 febbraio 2005 e in forma definitiva il 2 agosto 2005.
the Hausdorff dimension there. For invariant sets of piecewise monotonic maps on the interval, this dimension is investigated in [9].
If T : [0,1] → [0,1] is a map, one can define the recurrence time of a set U ⊂[0,1] byτ(U)=min{n ≥1 :Tn(U)∩U = ∅}. If one setsw(U)=e−τ(U) one gets the dimension introduced in [1] (see also [2]) and called the spectrum of dimensions for Poincar´e recurrences there. This is the dimension we consider in this paper. We call it recurrence dimension for simplicity.
For topological Markov chains and conformal repellers satisfying a specifica- tion property this recurrence dimension has been computed in terms of topologi- cal entropy or implicitly using a pressure function (see [2]). In this paper we do the same for invariant subsets of piecewise monotonic maps on the interval [0,1].
We say that T : [0,1] → [0,1] is piecewise monotonic if there are numbers 0= c0 <c1 < · · ·< cN = 1, such thatT|(ci−1,ci)is monotone and continuous for 1≤i ≤ N. Set
Z
= {(c0,c1), . . . , (cN−1,cN)}andP = {c0,c1, . . . ,cN}. We assume that the dynamical system([0,1],T)has only finitely many or no attracting periodic orbits. In chapter IV of [12] it shown that this holds forC2-mapsT with non-flat critical points. In the second part of this paper we assume thatT is expand- ing, which implies that there can be no attracting periodic orbits. Furthermore, we assume thatT has a continuous derivativeTon [0,1]\Pand setϕ= −log|T|. We shall need regularity conditions forϕ. To this end letC be the set of all func- tions f : [0,1]\P→R
, such that for each Z ∈Z
the map f|Z can be extended to a continuous real-valued function on the closure Z of Z. Furthermore let D be the set of all functions f : [0,1]\P→R
, such that for each Z∈Z
the map f|Z can be extended to a continuous function fZ : Z →(−∞,∞]. For p ≥1 letDp be the set of all f ∈ Dsuch thate−f is of bounded p-variation. The p-variation of a functiong: [0,1]→R
is defined by varpg:=supnj=1|g(xj)−g(xj−1)|p where the sup is taken over n ≥ 1 and 0 ≤ x0 < x1 < · · · < xn ≤ 1. Notice thatϕ ∈ C excludes the possibility thatT attains the value zero, whereas this is possible on P, ifϕ∈D.
In the first part of the paper we investigate the dimension of an invariant mea- sure. For a probability measureµwe define
ds(µ)=inf{ds(A):µ(A)=1}.
Theorem 2.4 below gives a local result. If ϕ is either in C or in Dp for some p ≥ 1 and if µ is an ergodic T–invariant measure with hµ > 0, then χµ :=
−µ(ϕ)is greater than zero and limr→0τ(−Blogr(xr)) = χ1µ holds forµ-almost all x. In this paper we prove only lim supr→0τ(−Blogr(xr)) ≤ χ1µ using methods similar to those used in [7] and [8] for Hausdorff dimension and local dimension of an ergodic invariant measure. The other part of this theorem, namely lim infr→0τ(−Blogr(xr)) ≥
χ1µ, is already shown in [15]. Using Theorem 2.4 and results from [7] on local dimension we get then thatds(µ) = hµχ−µs for everys ∈[0,hµ). This is stated as
Theorem 2.6 below. Similar results are known for subshifts with positive entropy satisfying a specification property (see [3]).
In the second part of this paper we investigate the dimension of a closed subset Aof [0,1], which is invariant under the piecewise monotonic mapT. We assume thatT is expanding, which means that inf|T|> 1, and thatϕ ∈C. Furthermore, we set
Z
n=n−1j=0T−j
Z
. For a function f ∈C we define p(T|A, f)=lim supn→∞
1
nlog
Z∈Zn,Z∩A=∅
sup
x∈Z∩A
e
n−1 j=0f(Tjx).
This is the definition of pressure for the isomorphic shift space one gets by coding with respect to the partition
Z
. Since the transformationT can have discontinuities on the set P, we use p(T|A, f)as the notion of pressure. Setπ(t)= p(T|A,tϕ) for t ≥ 0. For s ∈ [0,htop(T|A)) we show then that there is a unique solution zs(A) >0 of the equationπ(t) = s, which equalsds(A). This is stated as Theo- rem 3.1 below.2.
Dimension of an invariant measure
In order to investigate the recurrence dimension of an ergodic invariant measure µwithhµ > 0 we use Markov extensions introduced in [5] and [10] and used in [7] and in [8] for the investigation of Hausdorff dimensions. Let
Y
be a finite or countable collection of open pairwise disjoint intervals, such thatT|Y is monotone for allY ∈Y
and such thatµ(Y∈YY)=1. SetR=∞
j=0T−j(
Y∈YY). Since µis invariant, we haveµ(R) = 1. Forn ≥ 1 letYn(x)be the unique element of
Y
n =n−1j=0T−j
Y
, which containsx. Ifx ∈ R, thenYn(x)exists for alln≥1.The Markov extension of the dynamical system([0,1], µ,T)with respect to
Y
is constructed as follows. If D is a subinterval of some Y ∈Y
, we call the nonempty sets amongT(D)∩Y forY ∈Y
the successors ofD. We writeD→C, ifC is a successor of D. The successors are again subintervals of elements ofY
, so we can iterate the formation of successors. SetD
1=Y
andD
k+1 =D
k ∪ {C : there isD ∈D
k withD→C}fork≥1. Finally setD
=∞k=1
D
k.In order to get a dynamical system, take disjoint copies Dˆ of the intervals D ∈
D
and set Xˆ =D∈DD. Letˆ q : Dˆ → [0,1] be the imbedding, such that q : Xˆ → [0,1] is defined. If D ∈
D
andx ∈ D∩Y∈YY thenT(x)is defined and contained in
D→CC. For the corresponding pointxˆ ∈ ˆDwe letTˆ(xˆ)be the point in
D→CCˆ corresponding toT(x). In this way we have definedTˆ on the subsetq−1(
Y∈YY)ofX, such thatˆ q◦ ˆT =T◦qholds. Note thatq−1(
Y∈YY) containsq−1(R). IfHµ(
Y
)= −Y∈Yµ(Y)logµ(Y) <∞, it is shown in [8] (for finite
Y
in [11]) that there is a probability measureµˆ on Xˆ withµ(ˆ q−1(R)) = 1,such thatTˆ is definedµˆ-almost everywhere and has the following properties ˆ
µ is aTˆ-invariant ergodic measure, ˆ
µ(q−1(A))=µ(A) for all Borel subsetsAof [0,1].
Set
Y
ˆ = { ˆD: D∈D
}. By the above construction,Y
ˆ is a finite or countable Markov partition of the dynamical system (Xˆ,Tˆ,µ)ˆ . This explains the name Markov ex- tension. Forn≥1 letYˆn(xˆ)be that element ofY
ˆn =n−1j=0Tˆ−j
Y
ˆ, which contains ˆx. Ifxˆ∈q−1(R)thenYˆn(xˆ)exists for alln≥1.
We investigate recurrence on the Markov extension.
Lemma 2.1. There is a setMˆ ⊂ ˆX withµ(ˆ Mˆ)=1such thatlim sup
n→∞
1
nτ(Yˆn(xˆ))≤1 for allxˆ ∈ ˆM.
Proof. Choose D ∈
D
with µ(ˆ Dˆ) > 0. By the ergodic theorem there is a set MˆD ⊂ ˆX withµ(ˆ MˆD)= 1 such that limn→∞ 1nn−1j=01Dˆ(Tˆj(xˆ))= ˆµ(Dˆ)holds for allxˆ ∈ ˆMD.
Letxˆ ∈ ˆD∩ ˆMDand letn0=0<n1<n2< . . . be those integersn, for which Tˆn(xˆ)∈ ˆD holds. The above result says that limj→∞nj
j = ˆµ(Dˆ) > 0. Since
Y
ˆ is a Markov partition, we have either Dˆ ⊂ ˆTnj(Yˆnj(xˆ))or Dˆ ∩ ˆTnj(Yˆnj(xˆ))= ∅. SinceTˆnj(xˆ) ∈ ˆD∩ ˆTnj(Yˆnj(xˆ))we conclude that Dˆ ⊂ ˆTnj(Yˆnj(xˆ))for j ≥ 1.For arbitraryn ≥ 1 choose j such thatnj−1 < n ≤ nj. Then Yˆnj(xˆ) ⊂ ˆYn(xˆ) andxˆ ∈ ˆD ⊂ ˆTnj(Yˆn(xˆ))follows. Sincexˆ ∈ ˆYn(xˆ)we getτ(Yˆn(xˆ)) ≤ nj and
1
nτ(Yˆn(xˆ))≤ nnj−1j . Furthermore, limj→∞ nnj
j−1 = limj→∞ j−j1njjnj−1
j−1 = 1. This implies lim supn→∞ 1nτ(Yˆn(xˆ))≤1.
Set
D
∗= {D∈D
:µ(ˆ Dˆ) >0}andMˆ =D∈D∗(Dˆ ∩ ˆMD). We have shown, that lim supn→∞n1τ(Yˆn(xˆ))≤1 holds for allxˆ ∈ ˆM. Sinceµ(ˆ Dˆ ∩ ˆMD)= ˆµ(Dˆ) we get
ˆ
µ(Mˆ)=
D∈D∗
ˆ
µ(Dˆ ∩ ˆMD)=
D∈D∗
ˆ
µ(Dˆ)=
D∈D
ˆ
µ(Dˆ)= ˆµ(Xˆ)=1 and the proof is finished.
We bring down this result to the original dynamical system.
Lemma 2.2. We havelim supn→∞n1τ(Yn(x))≤1forµ-almost all x ∈[0,1].
Proof. Let Mˆ be as in Lemma 2.1 and set M = q(Mˆ)∩ R. Then q−1(M) ⊃ Mˆ ∩q−1(R)and henceµ(M)= ˆµ(q−1(M))=1.
Forx ∈ M there isxˆ ∈ ˆM withq(xˆ)= x. Suppose thatDn ∈
D
is such that Tˆn(xˆ)∈ ˆDnfor alln≥0. We haveq(Dˆn)= Dn. By the definition of successor andbecause ofq◦ ˆT =T◦qthere areU0,U1,· · · ∈
Y
, such thatTn(x)∈ Dn ⊂Unfor n ≥0. It follows thatYn(x)= n−1j=0T−jUj andYˆn(xˆ)=n−1
j=0Tˆ−jDˆj. We get q(Yˆn(xˆ))=n−1
j=0T−jDj ⊂Yn(x). Ifτ(Yˆn(xˆ))=k, thenTˆk(Yˆn(xˆ))∩ ˆYn(xˆ)= ∅, and sinceq◦ ˆTk =Tk◦qholds we getTk(Yn(x))∩Yn(x)= ∅. Henceτ(Yn(x))≤ τ(Yˆn(xˆ)). It follows from Lemma 2.1 that lim supn→∞ 1nτ(Yn(x))≤1 holds for all x∈ M, which means forµ-almost allx.
Lemma 2.3. Let T be a piecewise monotonic transformation, which has only finitely many or no attracting periodic orbits. Suppose thatϕ= −log|T|is either in C or in Dp for some p ≥ 1. Letµ be an ergodic T –invariant measure with hµ>0. Thenµ(ϕ)≤0andϕ∈L1(µ).
Proof. Suppose thatµ(ϕ) >0 holds. There is a continuous functionψ : [0,1]→
R
satisfyingψ ≤ ϕ andµ(ψ) > 0. Let (Xˆ,Tˆ,µ)ˆ be the Markov extension of ([0,1],T, µ)with respect to the partitionZ
and letqbe the corresponding projec- tion. By the ergodic theorem, there isxˆ ∈ ˆX, such that 1nn−1j=0δTˆj(xˆ) converges weakly toµˆ. Sinceµˆ is ergodic and has nonzero entropy, the set of ally, for whichˆ ∞
j=1Yˆn(yˆ)is a nondegenerate interval, has µˆ-measure zero, and hence we can assume that∞
j=1Yˆn(xˆ) = { ˆx}. There is a sequence D0D1D2. . . in
D
such that Tˆk(xˆ)∈ ˆDkfork≥0. SinceY
ˆ is a Markov partition, we haveTˆ(Dˆk)⊃ ˆDk+1for k ≥0. There isD ∈D
withµ(ˆ Dˆ) >0, and hence Doccurs infinitely often in this sequence. Choosen1 < n2 <n3 < . . . such thatDnk = D. For everym there is a point yˆ of periodnm−n1withTˆj(yˆ)∈ ˆDn1+j for 0≤ j ≤nm−n1. Letνˆm be the periodic-orbit-measure supported by the periodic orbit ofy. Then the sequenceˆˆ
νm converges weakly toµˆ. Since q : Xˆ → [0,1] is continuous, the measures νm = ˆνm◦q−1are periodic-orbit-measures on([0,1],T), which converge weakly toµ. Sinceψ is continuous, there ism0withνm(ψ) >0 form ≥m0. Because of ϕ ≥ψwe have alsoνm(ϕ) >0 form ≥m0. This implies, that the periodic orbits supporting the measuresνm form ≥ m0are attracting, contradicting the assump- tion, that at most finitely many attracting periodic orbits exist. This contradiction shows thatµ(ϕ)≤0. Since we assume thatϕis inC or inDp, it follows thatϕis bounded below and hence inL1(µ).
Theorem 2.4. Let T be a piecewise monotonic transformation, which has only finitely many or no attracting periodic orbits. Suppose thatϕ= −log|T|is either in C or in Dp for some p ≥ 1. Letµ be an ergodic T –invariant measure with hµ > 0. Thenχµ := −µ(ϕ)is greater than zero andlimr→0τ(−Blogr(xr)) = χ1µ holds forµ-almost all x.
Proof. By Lemma 2.3 we haveϕ ∈ L1(µ)and by Theorem 1 in [6] we getχµ =
−µ(ϕ) >0. Chooseε∈(0,χ2µ). By Lemma 1 in [8] there is a finite or countable
collection
Y
of pairwise disjoint intervals, such that µ
Y∈Y
Y
=1;
sup
x∈Y
ϕ(x)− inf
x∈Yϕ(x) < εfor everyY ∈
Y
; Hµ(Y
)= −Y∈Y
µ(Y)logµ(Y) <∞.
Now let Mε be a set ofµ-measure one, such that lim supn→∞ 1nτ(Yn(x))≤1 and that limn→∞1nn−1
j=0ϕ(Tjx) = µ(ϕ) = −χµ for all x ∈ Mε. This is possible because of Lemma 2.2 and the ergodic theorem. For x ∈ Mε we get by the mean value theorem and asTn(Yn(x))⊂[0,1] that
log|Yn(x)| ≤ sup
y∈Yn(x) n−1
j=0
ϕ(Tjy)+log|Tn(Yn(x))| ≤ sup
y∈Yn(x) n−1
j=0
ϕ(Tjy).
Since supUϕ−infUϕ <εfor everyU∈
Y
we getnj=−10ϕ(Tjy)≤n−1j=0ϕ(Tjx)+
nε, if y∈Yn(x), and hence log|Yn(x)| ≤n−1
j=0ϕ(Tjx)+nε. It follows that
nlim→∞
1
nlog|Yn(x)| ≤ −χµ+ε . In particular we have limn→∞|Yn(x)| =0, as−χµ+ε <0.
For x ∈ Mε andr > 0 let n be such that |Yn(x)| ≤ r < |Yn−1(x)|. This impliesYn(x)⊆ Br(x), and therefore
τ(Br(x))
−logr ≤ τ(Yn(x))
−log|Yn−1(x)| =
1
nτ(Yn(x))
−1nlog|Yn−1(x)|.
There is n0, such that 1nτ(Yn(x)) < 1+ε and−1nlog|Yn−1(x)| > χµ −2ε for n≥n0. Forr <|Yn0(x)|we get τ(−Blogr(xr)) < χ1µ+ε−2ε and hence also
lim sup
r→0
τ(Br(x))
−logr ≤ 1+ε χµ−2ε. Now setM =∞
k=1M1/k. Then we haveµ(M)=1 and lim supr→0τ(−Blogr(xr)) ≤ χ1µ for allx∈ M, which means forµ-almost allx.
Theorem 2 in [15] says that lim infr→0τ(−Blogr(xr)) ≥ χ1µ holds for µ-almost all x. It is proved there first for cylinder sets with respect to a partition for general dynamical systems and then an approximation procedure as above is used.
In the proof of the result on the dimension of a measure we shall use centered covers. Forε > 0 we call
F
a centered δ–cover of A, if A ⊆C∈FC and for everyC ∈
F
there exists anx ∈Aand anα∈(0, δ] withC =(x−α,x+α). The next lemma shows that we can use centered covers instead of covers.Lemma 2.5. For A ⊂[0,1]letν˜s,t(A)be defined in the same way asνs,t(A)but with centered covers instead of covers. Then νs,t(A) ≤ ˜νs,t(A) ≤ 2νs,t(A)for s,t ≥0.
Proof. Since a centeredδ-cover is also aδ-cover, it follows thatνs,t(A)≤ ˜νs,t(A). Let I be an interval of length≤δwith I ∩A= ∅. LetI1be the left half and I2the right half of I. If I1∩ A= ∅, there isx1 ∈ I1∩ AwithI1⊂ Br1(x1)⊂ I, where r1 is the distance from x1 to the left endpoint of I. If I1 ∩ A = ∅, set Br1(x1)= ∅. Similarly, ifI2∩A= ∅, there isx2 ∈I2∩ AwithI2⊂ Br2(x2)⊂ I, where r2 is the distance from x2 to the right endpoint of I. If I2∩ A = ∅, set Br2(x2)= ∅. We haveA∩(Br1(x1)∪Br2(x2))= A∩I. If
C
is aδ-cover ofA, for eachI ∈C
choose Br1(x1)andBr2(x2)as above. The nonempty sets among them form a centeredδ-coverC
˜ofAsuch thatC˜∈ ˜Ce−sτ(C˜)| ˜C|t ≤2C∈Ce−sτ(C)|C|t, sinceU ⊂ V impliesτ(U) ≥ τ(V)and|U| ≤ |V|. From this inequality we get
˜
νs,t(A)≤2νs,t(A).
Theorem 2.6. Let T be a piecewise monotonic transformation, which has only finitely many or no attracting periodic orbits. Suppose thatϕ= −log|T|is either in C or in Dp for some p ≥ 1. Letµ be an ergodic T –invariant measure with hµ>0. Then for every s∈[0,hµ)we have ds(µ)= hµχ−µs, whereχµ= −µ(ϕ). Proof. Setα=hχµµ andβ=χ1µ. By Theorem 2.4 we haveχµ>0 and lim
r→0 τ(Br(x))
−logr = β forµ-almost allx. By Theorem 1 in [7] we have limr→0logµ(logBrr(x)) = αforµ- almost allx.
In order to prove ds(µ) ≤ α−sβ choose an arbitrary t > α −sβ and set ε= 12(t −α+sβ). Sinces <hµ, we havet >0. Fork∈
N
defineMk={x∈[0,1] :µ(Br(x))t−εα ≥1krt andµ(Br(x))sβ−εα ≥1ke−sτ(Br(x))for allr∈(0,1)}
Forµ-almost allx∈[0,1] there exists anr0(x) >0 such that t−ε
α
logµ(Br(x))
logr ≤t and sβ−ε
α
−logµ(Br(x))
τ(Br(x)) = sβ−ε α
logµ(Br(x)) logr
−logr τ(Br(x)) ≤s
for allr ∈(0,r0(x)). These are the inequalities in the definition ofMk withk=1.
Hence forµ-almost all x ∈ [0,1] there is ak ∈
N
with x ∈ Mk, and therefore µ(∞k=1Mk)=1.
Fixk ∈
N
andδ∈ (0,1), and letC
be a centeredδ–cover ofMk. Since each C ∈C
is a ball with center inMk we get by the definition ofMk
C∈C
e−sτ(C)|C|t ≤2tk2
C∈C
µ(C)sβ−εα µ(C)t−εα =2tk2
C∈C
µ(C)≤2t+1k2. The last inequality follows, since we can assume that each x ∈[0,1] is contained in at most two element of
C
. Otherwise we can cancel elements fromC
and have still a cover. This implies that ν˜s,t(Mk) < 2t+1k2. By Lemma 2.5 we get also νs,t(Mk) < 2t+1k2. By the definition of dimension we haveds(Mk) ≤ t. For u > t we have thereforeνs,u(Mk) = 0 for allk, and asνs,u is an outer measure, we getνs,u(∞k=1Mk)=0. It follows thatds(∞
k=1Mk)≤t and henceds(µ)≤t, since µ(∞
k=1Mk) = 1. As t can be chosen arbitrarily close toα−sβ, we get ds(µ)≤α−sβ.
In order to prove thatα−sβ ≤ds(µ), it suffices to showµ(L)=0 for every Borel setL ⊆[0,1] withds(L) < α−sβ. LetLbe such a set. Sinces <hµ, we haveα−sβ >0. Chooset >0 withds(L) <t < α−sβand setε= 12(α−sβ−t). Fork∈
N
defineLk={x∈L :µ(Br(x))t+εα ≤krt andµ(Br(x))sβ+εα ≤ke−sτ(Br(x))for allr∈(0,1)}
Forµ-almost allx∈Lthere is anr0(x) >0 such that t+ε
α
logµ(Br(x))
logr ≥t and sβ+ε
α
−logµ(Br(x))
τ(Br(x)) = sβ+ε α
logµ(Br(x)) logr
−logr τ(Br(x)) ≥s
holds for allr ∈(0,r0(x)). These are the inequalities in the definition ofLk with k = 1. Hence forµ-almost allx ∈L there is ak ∈
N
withx ∈ Lk, and therefore µ(L)=µ(∞k=1Lk). It remains to showµ(Lk)=0 for allk∈
N
.To this end fix k ∈
N
and η > 0. Since ds(Lk) ≤ ds(L) < t we have νs,t(Lk)=0 and hence alsoν˜s,t(Lk)=0 by Lemma 2.5. There exists aδ∈(0,1) and a centeredδ–coverC
ofLkwith
C∈C
e−sτ(C)|C|t < 2t k2η .
Since eachC ∈
C
is a ball with center inLk we get by the definition ofLk that µ(Lk)≤C∈C
µ(C)=
C∈C
µ(C)sβ+εα µ(C)t+εα ≤ k2 2t
C∈C
e−sτ(C)|C|t < η . Asη >0 was arbitrary we getµ(Lk)=0, which completes the proof.
Remark 2.7. The statement in Theorem 2.6 holds also fors < 0. This can be shown by the method developed in chapter 5 of [16], which is based on a result in [4]. But it seems that the proof fors >0 cannot be given by this method. One gets only an inequality in this case.
3.
Dimension of an invariant set
In order to investigate the dimension of a closed invariant set Ain the dynamical system([0,1],T), we assume that T is expanding, this means that inf|T| > 1.
Furthermore, we assume thatT ∈C. This implies thatϕ= −log|T|is also inC.
For a function f ∈Cwe define p(T|A, f)=lim sup
n→∞
1
nlog
Z∈Zn(A)
sup
x∈Z∩A
eSnf(x)
where Snf(x) = n−1
j=0 f(Tjx) and
Z
n(A) = {Z ∈Z
n : Z ∩ A = ∅}. If we consider the dynamical system([0,1],T)as a shift space, this is the usual definition of pressure. In particular, the pressure does not change, if we refineZ
, and the variational principle holds. These and other properties of the pressure can be found in [17]. We define the functionπ : [0,∞)→R
byπ(t)= p(T|A,tϕ) .
For t1 < t2 we have t2ϕ ≤ t1ϕ +(t2− t1)supϕ and hence π(t2) ≤ π(t1)+ (t2 −t1)supϕ. Since supϕ < 0 this implies that π is strictly decreasing and that limt→∞π(t) = −∞. Furthermore, π(0) = htop(T|A) ≥ 0. Hence for s ∈ [0,htop(T|A)] there is a unique t˜ ≥ 0 withπ(t˜) = s, with π(t) > s for t < t˜and withπ(t) < s fort > t˜. We denote thist˜byzs(A). Ifs < htop(T|A) thenzs(A) >0.
Theorem 3.1. Let T be an expanding piecewise monotonic transformation on[0,1], such that T ∈ C, and A a closed invariant subset with htop(T|A) > 0. For s ∈[0,htop(T|A))we have then ds(A)=zs(A).
Proof. Sinces < htop(T|A)we havezs(A) > 0. Chooset ∈ (0,zs(A)). Since t < zs(A), we get p(T|A,tϕ) = π(t) > s. By the variational principle there is an ergodic invariant measure µon Asatisfying hµ +tµ(ϕ) > s. Because of µ(ϕ) ≤supϕ < 0 we gethµ > s. Sinceϕ ∈ C we can apply Theorem 2.6 and get ds(µ) = hµχ−µs. Since hµ −s > −tµ(ϕ) = tχµ we get ds(µ) > t. This implies ds(A) > t. As t can be chosen arbitrarily close to zs(A), the inequality ds(A)≥zs(A)is shown.
Choose t > zs(A). Then t > 0 and π(t) < s. Fix ε ∈ (0,s−π(t t)). We choose points 0 = d0 < d1 < · · · < dK = 1 containingc0,c1, . . . ,cN and set
Z
= {(d0,d1), . . . , (dK−1,dK)}, such thatT|Z is monotone and supx∈Zϕ(x)− infx∈Zϕ(x) < ε for every Z ∈Z
. If we define the pressure with respect to this finer partitionZ
we get the same as for the originalZ
. Setα= s−π(2t)−εt >0 andpn =en(tε−s)
Z∈Zn(A)supZ∩Aet Snϕforn≥1. Then lim sup
n→∞
1
nlogpn =tε−s+lim sup
n→∞
1
nlog
Z∈Zn(A)
sup
Z∩A
et Snϕ =tε−s+π(t)= −2α . Hence there is a constantc>0, such that pn≤ce−nαholds for alln≥1.
Set
Z
nk(A)= {Z ∈Z
n(A):τ(Z)=k}andZ
nn(A)= {Z ∈Z
n(A): τ(Z)≥ n}. Suppose thatZ =n−1j=0T−jZj ∈
Z
nk(A), where Z0,Z1, . . . ,Zn−1 ∈Z
and k < n. Then Tk(Z)∩ Z = ∅ and hence Zj+k = Zj for 0 ≤ j ≤ n−k −1.Set Z˜ = k−1
j=0T−jZj. ThenZ ⊂ ˜Z and Z˜ ∈
Z
k(A), as A∩ Z = ∅and hence A∩ ˜Z= ∅. Furthermore, the mapZ→ ˜ZfromZ
nk(A)toZ
k(A)is injective, since Zk, . . . ,Zn−1are determined byZ1, . . . ,Zk−1. Fork =nwe setZ˜ = Zand these results are trivial.For each Z∈
Z
k(A)we get by the mean value theorem that|Z| ≤ |TkZ|sup
x∈Z
|T(k)(x)| −1= |TkZ|sup
x∈Z
eSkϕ(x).
Since TkZ ⊂ [0,1] we have|TkZ| ≤ 1. Since supUϕ−infUϕ < ε for every U ∈
Z
we haveSkϕ(x)−Skϕ(y)≤kεwheneverxandyare in Z. As A∩Z = ∅ it follows that|Z| ≤ sup
x∈A∩Z
eSkϕ(x)+kε.
If Z ∈
Z
nk(A)andZ˜ as above, we have|Z| ≤ | ˜Z|since Z⊂ ˜Z. Hence we getZ∈Zn(A)
e−sτ(Z)|Z|t ≤ n k=1
e−sk
Z∈Znk(A)
|Z|t ≤ n k=1
e−sk
˜ Z∈Zk(A)
| ˜Z|t
≤ n
k=1
e−sk
Z˜∈Zk(A)
sup
A∩ ˜Z
et Skϕ+tkε
=n
k=1
pk ≤∞
k=1
pk ≤c ∞ k=1
e−kα ≤ ce−α 1−e−α
Now
Z
n(A)covers the setAexcept for possibly finitely many points. We can cover these finitely many points by intervalsI1,I2, . . . ,Ij of arbitrary small length, such thatji=1e−sτ(Ii)|Ii|t <1 holds. SinceT is expanding, for everyδ >0 there isn such that the diameter of all intervals in
Z
n(A)is less thanδ. Hence for everyδ >0 we have found aδ-coverC
ofAwithC∈Ce−sτ(C)|C|t < 1ce−−αe−α+1. This implies thatνs,t(A)≤ 1ce−−αe−α +1 andds(A)≤t follows. We can choosetarbitrarily close tozs(A)so thatds(A)≤zs(A)is shown.References
[1] V. AFRAIMOVICH,Pesins dimension for Poincar´e recurrences, Chaos7(1997), 12–20.
[2] V. AFRAIMOVICHand J. URIAS,Dimension–like characteristics of invariant sets in dy- namical systems, In: “Dynamics and randomness”, A. Maass, S. Martinez and J. San Martin (eds.), Kluwer Academic Publishers, Dordrecht, 2002, pp. 1–30.