You are on page 1of 23

CORRELATION DIMENSION OF MEASURES

INVARIANT UNDER GROUP ACTIONS

YA. PESIN and A. TEMPELMAN


The Pennsylvania State University

March 27, 1999 2:20pm

Abstract. In this paper we generalize the notion of Grassberger-Hentschel-Procaccia cor-


relation dimension to general actions of semigroups admitting an averaging sequence of mea-
sures. Our approach follows the approach elaborated in [P2]. Our main result claims that
the correlation dimension depends only on the measure invariant under the group action. We
study some properties of the correlation dimension and its relation to other characteristics
of dimension type, in particular the pointwise dimension. We provide some examples which
support a general oppinion that the correlation integral, in general, may not have a regular
asimptotic behaviour when the radius of correlation interaction tends to zero.

1. Introduction

It is well-known that one can get some information about dynamical systems (connected
mainly with invariant ergodic measures) by observing their individual trajectories. Grass-
berger, Procaccia and Hentschel [GPH] introduced the notion of correlation dimension in
attempts to produce a dimension-type characteristic of a dynamical system when observ-
ing one of its individual trajectories. Of course, this trajectory should be representative in
some sense, for example, be typical with respect to an invariant measure. In [GPH], the
authors describe a numerical procedure that leads to the correlation dimension. Namely,
let {f n (x)} be a trajectory of a dynamical system f acting on a metric space (X, ρ). Set
1
C(x, n, r) = 2
card{(i, j) : ρ(f i (x), f j (x)) ≤ r, 1 ≤ i, j ≤ n}
n
where card (A) is the number of elements in the set A. If n is large and r is small then
the correlation dimension α is defined to satisfy
C(x, n, r) ∼ r α
Key words and phrases. Semigroup actions, averaging sequences, correlation dimension, Hentschel-
Procaccia spectrum of dimensions, saturated correlation dimension, pointwise dimension.
The work of the first author was partially supported by a National Science Foundation grant #DMS91-
02887.

Typeset by AMS-TEX
1
(if such a relation can be held at all).
In [P2], the author conducted the rigorous mathematical description of this approach.
Namely, he proved that if f preserves a Borel ergodic measure µ on X then under a rather
mild assumption on µ the following limit exists

def
lim C(x, n, r) = ϕ(r)
n→∞

for µ-almost every x and any sufficiently small r and does not depend on x. In fact,
Z
ϕ(r) = µ(B(x, r))µ(dx)
X

and does not depend on the dynamical system f either but on the measure µ (B(x, r) is the
closed ball of redius r centered at x). One can now define the lower and upper correlation
dimensions as follows

log ϕ(r) log ϕ(r)


α(µ) = lim , ᾱ(µ) = lim .
r→0 log r r→0 log r

If these two quantities coincide the common value represents the correlation dimension
and it characterizes the measure µ on the space X.
One can now imagine two different dynamical systems f and g acting on X and pre-
serving the same measure µ. For a ”typical” pair of points (x, y) the numerical procedure
described above for calculating the lower and upper correlation dimensions along the tra-
jectories {f n (x)} and {g m (y)} leads to the same results: the statement that a priori does
not seem to be obvious.
Such a situation has been recently discovered in the connection to a very interesting
phenomenon known as spatial-temporal chaos (see [BS], [PS], [AP], [APT]). It refers to the
case of d-dimensional lattice models of unbounded media given by an ”evolution” operator
Φ acting on a Banach space X. The spatial-temporal chaos means that Φ preserves a
finite measure µ which is mixing and also invariant under the group of space-translations
{S n , n ∈ Zd } where S n (m) = m + n, m ∈ Zd , d ≥ 1. It is quite surprising that some char-
acteristics of dimension type (including the correlation dimension) calculated with respect
to the group of time-translations (generated by Φ) and the group of space-translations S n
turn out to be the same (see [APT]). This also gives rise to the problem of defining the
correlation dimension with respect to general group actions. We shall give the definition
of the correlation dimension specified by an action τ of a group admitting an averaging
sequence of measures {λn }. We shall show that the correlation dimension depends only
on the measure µ invariant under τ but not on τ, {λn } and an initial point x ∈ X (for
µ-almost every x). Our result (see Theorem 2.1) is more strong and general than the
corresponding result in [P2].
It is easy to see that α(µ) ≤ ᾱ(µ). We provide an example of a measure µ on [0, p], p > 0
equivalent to the Lebesgue measure, for which α(µ) < α(µ). Moreover, it follows drom the
construction and results in [K] that there exists a measure µ on a smooth compact surface
2
invariant under a C 2 −diffeomorphism with non-zero Lyapunov exponents, Bernoulli, and
equivalent to the Lebesgue measure for whichthe strong inequality holds. This complies
with the opinion among specialists in the field that the noncoincidence of the lower and
upper correlation dimensions is a rather ”typical” (or at least ”often”) phenomenon.
We also consider a modification of the notion of correlation dimension, namely the lower
and upper saturated correlation dimension (see [P2] and Section 2.5 below). Let us set
β(µ) and β̄(µ). These quantities have the advantage that β(µ) = β̄(µ) under a rather mild
assumption on µ (see Section 2.5). In fact the common value coincides with many other
dimensionlike characteristics specified by µ.
Unlike other characteristics of dimension type the correlation dimension is widely used
in numerical investigation of dynamical systems since the algorithm of its calculation is
relatively simple and fast. In this connection there are two problems that are important
in applications. The first one is whether the function ϕ(r) is continuous. The answer
depends on the metric on X. We shall show that if X is a smooth finite-dimensional
manifold endowed with a Riemannian metric then the function ϕ(r) is continuous for any
Borel measure µ on X (see Theorem 2.9). On the other hand we shall give a simple
example of a metric and a measure on Rp for which ϕ(r) is discontinuous.
When one observes the trajectory of a point x ∈ X (during numerical or experimental
study of a dynamical system) it usually occurs that the position of the point f n (x) can be
registered only at some particular moments of time but not all. Then the problem comes
up to compute the correlation dimension if only such an information is available. This
leads to the study of the correlation dimension specified by different sequences of measures
on groups that can serve like averaging ones. In general, in this case the correct value can
be obtained under more strong assumptions on the ergodic properties of the group action
(for example, weakly mixing or mixing). Moreover, instead of the pointwise convergence
for µ-almost every x ∈ X one can expect only weaker convergence in Lα (X, F , µ), α ≥ 1.
In Section 2.8 we present some particular results. The detailed description can be found
in [T2].

2. Definitions and results

2.1. Averaging sequences on semigroups. Let (X, F , µ) be a probability space, S


a topological semigroup. Let also τ = {τs , s ∈ S} be an action of S on X, i.e.
1) for any s ∈ S, the transformation τs is measurable and preserves the measure µ :
µ(τs−1 A) = µ(A), A ∈ F ;
2) τs1 τs2 = τs1 s2 , s1 , s2 ∈ S;
3) for any measurable function f (x), x ∈ X, the function ϕf (s, x) = f (τs x) is measur-
able with respect to the σ-algebra B × F where B is the Borel σ-algebra on S.
A set A ⊂ X is called invariant (with respect to the action τ ) if µ(τs A4A) = 0 for
any s ∈ S; the action is ergodic if any invariant set is of µ-measure 0 or 1. A sequence
of probability measures {λn } on S is said to be averaging if for any ergodic action τ =
{τs , s ∈ S} and any function f ∈ L1 (X, F , µ) the limit
Z Z
lim f (τs x)λn (ds) = f (x)µ(dx) (1)
n→∞ S X
3
exists for µ-almost every x ∈ X. In particular, when

λ(An ∩ B)
λn (B) =
λ(An )

(i.e. λn are uniformly distributed on sets An ∈ B, with respect to a measure λ) we say


that the sequence of sets {An } is averaging; then one can rewrite (1) in the following form:
Z Z
1
lim f (τs x)λ(ds) = f (x)µ(dx) (2)
n→∞ λ(An ) An X

We say that {λn } is strictly averaging if it is averaging and for any action τ there is a
constant γ > 0 such that for any f ∈ L2 (X, F , µ)
Z  Z 2 Z
2
sup f (τs x)λn (ds) µ(dx) ≤ γ |f (x)| µ(dx).
1≤n≤∞
X S X

The Birkhoff ergodic theorem provides classical examples of averaging sequences: for S =
Z+ = {0, 1, 2, . . . } one can set An = {0, 1, 2, . . . , n}; for S = R+ = [0, ∞) one can set
An = [0, n]. Examples for S = Rd or S = Rd+ are increasing sequences of convex sets {An }
with intrinsic diameters tending to ∞. The intersections of these sets with Zd , Zd+ form
averaging sequences for S = Zd , Zd+ , respectively (see [T1], where many other examples
for general semigroups and groups are given).
M. Lin pointed out to us that any avaraging sequence on a commutative semigroup or
on a compact group is also srictly avaraging; this can be proved using results by S. Sawyer
(see [S] and also [G]); in the case of compact groups one also has to employ the Transfer
Principle (see [T1]).

2.2. Correlation dimension. Let X be a complete separable metric space with metric
ρ and F the Borel σ-algebra on X. Consider an action τ of a semigroup S on (X, F , µ).
From now on we assume that it is ergodic. Let also {λn } be a sequence of measures on S.
Given x ∈ X, n > 0, r > 0 denote

C(x, λn , r) = (λn × λn )({(s, t) : ρ(τs x, τtx) ≤ r, s, t ∈ S})

Define now the lower and upper correlation dimensions at the point x specified by the
sequence of measures {λn }, the action τ , and the measure µ as follows:

log C(x, λn , r)
α(x, µ, {λn}, τ ) = limr→0 lim ,
n→∞ log r
(3)
log C(x, λn , r)
ᾱ(x, µ, {λn}, τ ) = limr→0 lim .
n→∞ log r
4
First we shall consider the existence of the limit in (3) when n → ∞. Denote
Z
ϕ(r) = µ(B(x, r))µ(dx),
X

where B(x, r) is the closed ball of radius r centered at x. It is clear that the function
ϕ(r), r > 0 is nondecreasing and, hence, may have only a finite or countable set of discon-
tinuity points. Since µ(B(x, r)) ≤ 1 for any x ∈ X, we have
Z
ϕ(r + 0) = lim µ(B(x, ρ))µ(dx)
ρ→r+0
ZX
= µ(B(x, r)µ(dx) = ϕ(r),
Z X

ϕ(r − 0) = lim µ(B(x, ρ))µ(dx)


ρ→r−0
ZX
= lim µ(B(x, r) \ S(x, r))µ(dx)
X ρ→r−o
Z
= ϕ(r) − µ(S(x, r))µ(dx)
X

where S(x, r) is the sphere of radius r centered at x. Thus ϕ is right continuous; it is


continuous at r if and only if µ(S(x, r)) = 0 for µ-almost all x ∈ X. We denote by Cϕ the
set of continuity points of ϕ.
Theorem 2.1. Assume that τ is an ergodic action of S in X.
(1) If {λn } is an arbitrary averaging sequence of measures on S then
a) there is a set Y ⊂ X of full µ-measure such that for any x ∈ Y the following
limit exists:
lim C(x, λn , r) = ϕ(r), r ∈ Cϕ ; (4)
n→∞

b) the convergence in (4) is uniform over r on any interval (0, R], 0 < R < ∞
provided ϕ(r) is continuous;
(2) If {λn } is a strictly averaging sequence then for µ-almost all x and for every r > 0

lim C(x, λn , r) = ϕ(r)


n→∞

and the convergence is uniform over r on any interval (0, R], R > 0.

Theorem 2.1 shows that the quantities α(x, µ, {λn}, τ ) and ᾱ(x, µ, {λn }, τ ) do not de-
pend on the averaging sequence {λn } and the action τ and are constant on a set of µ-
measure 1. For points of this set we can write

α(x, µ, {λn}, τ ) = α(µ), ᾱ(x, µ, {λn }, τ ) = ᾱ(µ).

The first statement of Theorem 2.1 in the case when the function ϕ(r) is continuous and
S = Z, An = {0, 1, . . . , n} was proved earlier in another way in [P2]. In the special case
5
when X = Rd , S = Z, An = {0, 1, . . . , n} and ϕ(r) is continuous the uniform convergence
in (4) was announced by R.Serinko (oral communication).
Remark 2.2. One can consider the open balls B− (x, r) and the functions:

C− (x, λn , r) = (λn × λn )({(s, t) : ρ(τs x, τt x) < r, s, t ∈ S}),


Z
ϕ− (r) = µ(B− (x, r))µ(dx).
X

It is easy to verify that ϕ(r − 0) = ϕ− (r) and ϕ− (r + 0) = ϕ(r), ϕ− (r − 0) = ϕ− (r). Thus


ϕ− is nondecreasing, left continuous, and ϕ(r) = ϕ− (r) if and only if r belongs to the
dense set Cϕ . Therefore

log ϕ(r) log ϕ− (r)


α(µ) =limr→0 = limr→0 ,
log r log r
(5)
log ϕ(r) log ϕ− (r)
ᾱ(µ) =limr→0 = limr→0 .
log r log r

Moreover, the following version of statement (2) of Theorem 2.1 is also true: for µ-almost
all x and every r > 0
lim C− (x, λn , r) = ϕ− (r).
n→∞

Remark 2.3. Let µ be a Borel measure on a separable metric space (X, ρ). The
relations (5) produce the values α(µ) and ᾱ(µ) that one can call the lower and upper
correlation dimensions of µ. This definition does not involve any ”dynamics”. If, in
addition, there is an action τ of a semigroup S on X preserving µ then the lower and
upper correlation dimensions specified by τ coincide respectively with the lower and upper
correlation dimensions of µ, i.e. α(x, µ, {λn}, τ ) = α(µ), ᾱ(x, µ, {λn}, τ ) = ᾱ(µ) for µ-
almost every x ∈ X.
Let us summarize. The lower and upper correlation dimensions defined by (3) depend
neither on the choice of µ-generic point x nor on the averaging sequence {λn }. Moreover,
they do not depend on the semigroup S and its action τ on (X, F , µ) as long as µ is fixed.
Thus they are completely specified by µ and it coincides with the quantities defined by (5).
One can use any ergodic action of arbitrary semigroup on (X, F , µ) (for a given µ) in order
to estimate the lower and upper correlation dimensions. If one can choose an action with
”good” ergodic properties this may simplify the calculation of the correlation dimension
(see [T2]).
We shall also consider a modified approach to define the correlation dimension. Let
(X, ρ) be a separable metric space, µ a Borel probability measure on X and σ and τ
actions of a topological semigroup S on (X, B, µ). Let also κn be a sequence of Borel
probability measures on S × S. Given x, y ∈ X, n > 0, r > 0 set

C(x, y, κn , r) = κn ({(s, t) : ρ(σs x, τt y) ≤ r, s, t ∈ S}),


6
log C(x, y, κn, r)
α(x, y, µ, {κn}, σ, τ ) = limr→0 lim ,
n→∞ log r
(6)
log C(x, y, κn, r)
ᾱ(x, y, µ, {κn}, σ, τ ) = limr→0 lim
n→∞ log r
(assuming that the limit when n → ∞ exists).
Consider the direct product space (Y, ν) where Y = X × X, ν = µ × µ.
Theorem 2.4. Let σ and τ be ergodic actions of S in X and let {κn } be an arbitrary
averaging sequence of measures on S × S. Then for ν-almost every (x, y) ∈ Y
(1) for any r > 0 the following limit exists:

lim C(x, y, κn , r) = ϕ(r);


n→∞

(2) α(x, y, µ, {κn}, σ, τ ) = α(µ), ᾱ(x, y, µ, {κn}, σ, τ ) = ᾱ(µ).

Remark 2.5 Roughly speaking the quantities α(x, µ, {λn }, τ ) and ᾱ(x, µ, {λn}, τ ) de-
fined by (3) represent the asymptotical (when r → 0) ”remoteness” of points on the orbit
starting at x while the quantities α(x, y, µ, {κn}, σ, τ ) and ᾱ(x, y, µ, {κn}, σ, τ ) defined by
(6) reflect the asymptotical ”remoteness” of points on the orbits of two different actions
starting at different points x and y. The amazing phenomenon is that for ”typical initial
data” these quantities coincide, respectively, and do not depend on the data.

2.3. Correlation dimension of higher order. We now extend the notion of corre-
lation dimension by considering correlations of higher order. Namely, given x ∈ X, n ≥
0, r > 0, and integer q ≥ 1 denote

C(x, q, λn , r) = (λn × λn )({(i1 . . . iq ) ∈ S q : ρ(τij x, τik x) ≤ r for any ij , ik ∈ S})

Define now the lower and upper correlation dimensions of order q at point x specified by
the sequence of measures {λn }, the action τ , and the measure µ as follows:
1 log C(x, q, λn , r)
αq (x, µ, {λn}, τ ) = limr→0 lim ,
q n→∞ log r
1 log C(x, q, λn , r)
ᾱq (x, µ, {λn}, τ ) = limr→0 lim .
q n→∞ log r
The existence of the limit when n → ∞ is guaranted by a general version of Theorem
2.1. Namely, consider the function
Z
ϕq (r) = µ(B(x, r))q−1µ(dx),
X

for integers q ≥ 1 and r > 0. It is non-decreasing and, hence, may have only a finite
or countable set of discontinuity points. It is clearly right-continuous. The proof of the
following statement is a modification of the proof of Theorem 2.1.
7
Theorem 2.6. Assume that τ is an ergodic action of S in X.
(1) If {λn } is an arbitrary averaging sequence of measures on S then for any q =
1, 2, . . .
a) there is a set Y ⊂ X of full µ-measure such that for any x ∈ Y the following
limit exists:
lim C(x, q, λn , r) = ϕ(r), r ∈ Cϕq ;
n→∞

b) the convergence in (4) is uniform over r on any interval (0, R], 0 < R < ∞
provided ϕq (r) is continuous; (note that ϕq (r) is continuous if and only if ϕ(r) is
continuous);
(2) If {λn } is a strictly averaging sequence then for µ-almost all x, for every r > 0,
and any q = 1, 2, . . .

lim C(x, q, λn , r) = ϕq (r)


n→∞

and the convergence is uniform over r on any interval (0, R], R > 0.

Similarly to what was mentioned above the lower and upper correlation dimensions of
higher order depend neither on the choice of µ-generic point x nor on the averaging sequence
{λn }. Moreover, they do not depend on the semigroup S and its action τ on (X, F , µ) as
long as µ is fixed. Thus they are completely specified by µ and we can introduce

αq (µ) = α(x, q, µ, {λn}, τ ), ᾱq (µ) = ᾱ(x, q, µ, {λn}, τ )

for µ-almost every x ∈ X.

2.4. Saturated correlation dimension. Following [P2] we introduce a modification


of the notion of correlation dimension of order q. We set for q = 1, 2, . . .
R
log µ(B(x, r))q−1µ(dx)
1 Z
β q (µ) = lim sup lim ,
q − 1 δ→0 r→0 log r
(10)
R
log µ(B(x, r))q−1µ(dx)
1 Z
β q (µ) = lim sup lim
q − 1 δ→0 r→0 log r
where the supremum is taken over all sets Z with µ(Z) ≥ 1 − δ. We call these quantities
the lower and upper saturated correlation dimensions of order q. From the definitions one
can easily derive the relations between correlation dimensions of order q and saturated
correlation dimensions of order q

β q (µ) ≥ αq (µ), β̄q (µ) ≥ ᾱq (µ). (11)

The following theorem shows that β q (µ) and β̄q (µ) are completely specified by the equiv-
alence class of µ.
8
Theorem 2.7. Let µ1 and µ2 be Borel measures on X. If µ1 ∼ µ2 then β q (µ1 ) = β q (µ2 )
and β̄q (µ1 ) = β̄q (µ2 ) for any q > 0.

Remark 2.8. It is also easy to see that ᾱq (µ1 ) ≤ ᾱq (µ2 ), αq (µ1 ) ≤ αq (µ2 ) if dµ2
dµ1 (x) ≤
C < ∞, x ∈ X; thus ᾱq (µ1 ) = ᾱq (µ2 ), αq (µ1 ) = αq (µ2 ) if 0 < c ≤ dµ2
dµ1
(x) ≤ C < ∞, x ∈
X.
We study relations between pointwise dimension and correlation dimension Let µ be a
Borel probability measure on X. The lower and upper pointwise dimensions of µ at x ∈ X
are defined as follows

log µ(B(x, r)) ¯ log µ(B(x, r))


dµ (x) = lim , dµ (x) = lim .
¯ r→0 log r r→0 log r

For any x ∈ X the function ψ(x, r) = log µ(B(x,r))


log r is right-continuous and hence the ran-
dom process ψ(x, r) over X with time r > 0 is separable; this implies that the functions
dµ (x), d¯µ (x) are measurable (see [B]). The following theorem is a refined version of Theo-
rem 4 in [P3].
Theorem 2.9. For any q > 0

essinf dµ (x) ≤ β q (µ) ≤ β̄q (µ) ≤ essinf d¯µ (x).


x∈X x∈X

Theorem 2.9 implies the following statement.


Theorem 2.10. If
dµ (x) = d¯µ (x) = dµ (x)
µ-almost everywhere, then

β q (µ) = β̄q (µ) = essinf dµ (x)


x∈X

for any q > 0.

If the measure µ is invariant under a smooth diffeomorphism of a smooth manifold


then dµ (x) = d and d¯µ (x) = d almost everywhere. In [C], Cutler showed that this may
not be true for continuous maps (see also a refined version of her result in [PW]). In
[LM], Ledrappier and Misiurewicz constructed an example of a diffeomorphism for which
d 6= d. In [PY], the authors proved that d = d for a broad class of measures with non-zero
Lyapunov exponents (see also [ER]).
We now provide an example that establishes the difference between lower and upper
correlation dimension and saturated correlation dimension. It illustrates that, in general,
α(µ) < α(µ) < β(µ) even in the ”good” cases when dµ (x) = dµ (x) = 1. This example
shows also that the analog of Theorem 2.10 for the correlation dimensions α and α is false.
9
Example 2.11. For any two numbers 0 < α < α < 1 there exists a probability Borel
measure µ on the interval [0, p], for some p > 0, such that
(1) µ is equivalent to the Lebesgue measure;
(2) dµ (x) = dµ (x) = 1 for almost every x ∈ [0, p];
(3)
α = α(µ) < ᾱ(µ) = ᾱ < β(µ) = β̄(µ) = 1.

The proof, including the construction of the measure µ, is given in Section 3. Moreover,
one can show that for any q ≥ 1

αq (µ) < ᾱq (µ) < β q (µ) = β q (µ) = 1

for any q ≥ 1.
It is well-known that there exists a continuous mapping on [0, p] for which µ is an
invariant Bernoulli measure. In fact, the construction of Example 2.1 and results in [K] can
be used to construct a diffeomorphism of a compact surface which preserves an absolutely
continuous Bernoulli measure with non-zero Lyapunov exponents satisfying condition 3 in
Example 2.11.
Example 2.12. Let σ be the full-shift on the space Σp of two-sided infinite sequences
of symbols 0, 1, . . . , p − 1. We assume that Σp is endowed with the standard metric
X∞ (1) (2)
(1) (2) |ωi − ωi |
ρ(ω ,ω )=
i=−∞
a|i|

where a > 1. Then for any σ−invariant measure µ on Σp and µ−almost every x ∈ Σp
h(µ)
dµ (x) = dµ (x) =
log a
where h(µ) denotes the measure-theoretic entropy of µ; hence,

β(µ) = β(µ) = h(µ).

To see this, we note that for any ω ∈ Σp the ball B(ω, r) coincides with the cylinder set
∆n (ω) = {ω : ωi = ωi for i = −n, . . . , n} which contains ω, where n is chosen such that
p−1
an+1
≤ Cr < a1n for some C > 0 independent of n and r. The result now follows from the
Shannon-McMillan-Breiman Theorem.

2.5. Continuity of the function ϕ(r). In this section we shall discuss the question,
that is rather interesting in applications: when is the function ϕ(r) continuous? In general,
the answer is negative (see Example 2.15 below). In fact this answer depends on the
properties of the metric ρ on X. We shall present two equivalent metrics on the plane
such that ϕ(r) calculated with respect to one of them is continuous and it is not when
calculated with respect to the other one. The following theorem shows that ϕ(r) turns out
to be continuous under rather mild conditions on the metric ρ.
10
Theorem 2.13. Let X be a finite-dimensional Riemannian manifold with Riemannian
metric <, >x , x ∈ X, and µ a nonatomic Borel measure on X. Then the function ϕ(r) is
continuous at any r > 0.

Remark 2.14. If µ has a finite number of atoms then under assumption of Theorem
2.13 the function ϕ(r) is still continuous on some interval (0, R), R > 0.
If µ is a purely atomic measure on a measurable set X the function ϕ(r) may be
discontinuous. However, in this case the correlation dimension α = α = ᾱ = 0.
We shall show that Theorem 2.13 becomes false if we do not impose some restrictions
on the metric ρ on X.
Example 2.15. Let X = R2 with the metric ρ((x1 , y1 ), (x2 , y2 )) = max{|x1 −x2 |, |y1 −
y2 |}. The ρ-sphere Sr ((x0 , y0 )) of radius r centered at (x0 , y0 ) is the square {(x, y) : x0 −r ≤
x ≤ x0 + r, y0 − r ≤ y ≤ y0 + r}.
For a Borel set A in R we define
1
µ(A × {k})= `(A ∩ [−2, 2]), k = 0, 1,
8
where ` is the Lebesgue measure in R, i.e. µ is uniformly distributed on the segments
Ik = {(x, y) : −2 ≤ x ≤ 2, y = k}, k = 0, 1. It is clear that µ{S((x, y), 1)} = 14 if
−1 ≤ x ≤ 1, y = 0 and hence
1 1
ϕ(1 + 0) − ϕ(1 − 0) > µ{(x, y) : −1 ≤ x ≤ 1, y = 0} =
4 16
According to Theorem 2.13 the function ϕ(r) is continuous if ρ is the Euclidean metric in
R2 .

2.6. Estimation of ϕ(r) based on finite sets. Theorem 2.1 claims that the quantities
C(x, r, An ) form a consistent sequence of estimates of ϕ(r) if {An } is an averaging sequence
of sets and if τ is ergodic. In reality, only a finite number of observations {τg x, g ∈ An }
can be performed and the corresponding estimate is

card{(s, t) : ρ(τs x, τtx) ≤ r, s, t ∈ An }


C(x, r, An ) =
(card(An ))2

(so in this case we consider the discrete topology in the semigroup S and the ”basic”
measure λ (see Section 2.1) is the counting measure). There are no averaging sequences
of finite sets, unless S is countable, and Theorem 2.1 is not applicable for finite sets An .
Nevertheless, under some additional assumptions on the action τ the class of sequences
of finite sets {An }, producing consistent sequences of estimates C(x, r, An ), is sufficiently
large. We state here two theorems in this direction (see [T2] for proofs and other results).
Let S be a group. A set D ⊂ S is called a Tquasi-lattice if there exists a bounded
neighborhood V of the identity such that sup |D V x| < ∞. Of course, any subset of a
x∈D
quasi-lattice is also a quasi-lattice.
11
An action τ of S in X is said to be mixing if
Z Z Z
lim f (τs x)g(x)µ(dx) = f dµ g dµ, f, g ∈ L2 (XF , µ).
s→∞
X X X

Theorem 2.16. Suppose {An } is a sequence of finite sets in S with the following proper-
ties:
S

(1) An is a quasi-lattice;
n=1
(2) |An | → ∞.
If the action τ is mixing, then for any α ≥ 1

(Lα ) lim Cn (x, An , r)=ϕ(r), r ∈ Cϕ .


n→∞

A lattice in S is the set S0 a = {s = s0 a, s ∈ S0 } where S0 is an infinite discrete subgroup


of S and a ∈ S. Of course, T any lattice is a quasi-lattice: if V contains only the identity
element of S0 then |S0 a V x| = 1, x ∈ S0 a.
Theorem 2.17. Let {Bn } be an averaging sequence in S0 and An = Bn a. If τ is mixing
then µ-almost everywhere
lim Cn (x, An , r)=ϕ(r)
n→∞

and the convergence is uniform over r ∈ (0, R), 0 < R < ∞.

3. Proofs

Proof of Theorem 2.1. I. First suppose S∞ that r ∈ Cϕ . We fix an arbitrary ε > 0


and consider a countable partition X = k=1 Dkε of X into disjoint sets of diameter
d(Dkε ) < ε, k = 1, 2, . . . . Define the sets
[ \
ε ε
Fk,+ (r) = B(u, r), Fk,− (r) = B(u, r)
ε
u∈Dk ε
u∈Dk

in X and the sets


ε
Ek,+ (r) = Dkε × Fk,+
ε ε
(r), Ek,− (r) = Dkε × Fk,−
ε
(r),
U+ (r) = {(x, y) : ρ(x, y) ≤ r} = U+ (Λ, r),
U− (r) = {(x, y) : ρ(x, y) < r} = U− (Λ, r),
[∞ ∞
[
ε ε ε ε
U+ (r) = Ek,+ (r), U− (r) = Ek,− (r)
k=1 k=1

in X × X. Recall that ν = µ × µ. By the Fubini theorem ϕ(r) = ν(U+ (r)), ϕ− (r) =


ν(U− (r)).
12
Since {λn } is an averaging sequence, in X there exist two F -sets, M1 and M2 , with
µ-measure 1 such that Z
lim IDkε (τs x)λn (ds) = µ(Dkε )
n→∞ S

and Z
lim IFk,+ (r) (τs y)λn (ds) = µ(Fk,+ (r))
n→∞ S

for all x ∈ MT1 , y ∈ M2 where IA denotes the indicator of the set A. Therefore for all
x ∈ M = M1 M2
Z Z
lim IEk,+
ε (τs x, τt x)λn (ds)λn (dt)
n→∞ S S
Z Z !
= lim IDkε (τs x)λn (ds) · IFk,+
ε (r) (τt x)λn (dt)
n→∞ S S

= µ(Dkε )µ(Fk,+
ε ε
(r)) = ν(Ek,+ (r)).

Note that µ(M ) = 1. Obviously,



X
IU+ε (r) (x, y) = IEk,+
ε (r) (x, y).
k=1

P

Since µ(Dkε ) = 1, for any η > 0 there exists a natural number Nε such that
k=1
P
∞ S

ε (r) ≤ 1 and hence ν(E
µ(Dkε ) < η. Denote LεNε = Dkε . Since IFk,+ k,+ (r)) ≤
ε
k=Nε+1 Nε+1
µ(Dkε ), we have for all x ∈ M :
Z Z
lim | IU+ε (r) (τs x, τt x)λn (ds)λn (dt) − ν(U+
ε
(r))| =
n→∞ S S
∞ Z Z
X ∞
X
lim | IEk,+
ε (r) (τs x, τt x)λn (ds)λn (dt) −
ε
ν(Ek,+ (r))| ≤
n→∞ S S
k=1 k=1
Nε Z Z
X X

lim | IEk,+
ε (r) (τs x, τt x)λn (ds)λn (dt) −
ε
ν(Ek,+ (r)|+
n→∞ S S
k=1 k=1

X Z
ε
ν(Ek,+ (r)) + lim ILεNε (τs x)λn (ds) ≤ 2µ(LεNε ) < 2η.
n→∞ S
k=Nε +1

Since η is chosen arbitrarily, for all x ∈ M the following limit exists:


Z Z
ε
lim IU+ε (r) (τs x, τt x)λn (ds)λn (dt) = ν(U+ (r)). (7)
n→∞ S S
13
Similarly one can prove that for all x ∈ M
Z Z
ε
lim IU−ε (r) (τs x, τt x)λn (ds)λn (dt) = ν(U− (r)). (8)
n→∞ S S

It is easy to check that for every x̄ ∈ Dkε


B(x̄, r − ε) ⊂ Ek,+
ε
(r) ⊂ B(x̄, r + ε)
and
εm
U+ (r) ↓ U+ (r), U−
εm
(r) ↑ U− (r)
S

if X = Dkεm is a sequence of refining subpartitions of X with εm → 0. This implies
k=1
ν(U+εm
(r)) ↓ ν(U+ (r)) = ϕ(r) and ν(U−εm
(r)) ↑ ν(U− (r)) = ϕ− (r). If ϕ is continuous at r,
then ν(U− (r)) = ν(U+ (r)) and
 
lim ν(U+ (r)) − ν(U− (r)) = 0.
εm εm
m→∞

Since IU−ε (r) ⊂ IU− (r) ⊂ IU+ (r) ⊂ IU+ε (r) , relations (7) and (8) imply the existence of the
limits
Z Z
lim IU+ (r) (τs x, τt x)λn (ds)λn (dt) = ν(U (Λ, r)) = ϕ(r),
n→∞ S S
Z Z
lim IU− (r) (τs x, τt x)λn (ds)λn (dt) = ν(U (Λ, r)) = ϕ(r).
n→∞ S S

II. Since the action τ is ergodic µ is either nonatomic or its support is finite. In both
cases we can construct a sequence of subpartitoins of X: X = ∪k Dkεm with the additional
2
property: µ(Dkεm ) ≤ 4ak where a is a positive constant. Define:
Z Z
m
fn (x) = IU+εm (r) (τs x, τt x)λn (ds)λn (dt)
S S
m,l
and denote: F̂k,+ εm
(r) = Fu,+ (r) if Dkεl ⊂ Duεm . Note that if l > m then
ZZ X

fnm (x) = IDεl (τs x)IF̂ m,l (r) (τt x)λn (ds)λn (dt).
k k,+
SS k=1

Therefore

fnm (x) − fnl (x) =


ZZ X

IDεl (τs x)IF̂ m,l (r)\F εl (r) (τt x)λn (ds)λn (dt)
k k,+ k,+
SS k=1
X∞ Z Z
= IDεl (τs x)λn (ds) IF̂ m,l (r)\F εl (τt x)λn (dt)
k k,+ k,+ (r)
k=1 S S
14
and hence,
Z
sup |fnm (x) − fnl (x)|µ(dx) ≤
n
 X

∞ Z
X Z Z
sup IDεl (τs x)λn (ds) sup IF̂ m,l (r)\F εl (τt x)λn (dt) µ(dx)
n k
n k,+ k,+ (r)
k=1 X S S
 !2  12

X Z Z
≤  sup IDεl (τs x)λn (ds) µ(dx) ×
k
n
k=1 X S
 !2  12
Z Z
 sup IF̂ m,l (r)\F εl (τt x)λn (ds) µ(dx) .
n k,+ k,+ (r)
X S

Since {An } is a stricly averaging sequence,


Z
sup |fnm (x) − fnl (x)|µ(dx) ≤
n
X
  12

X Z Z
γ  I εl (x)µ(dx) I m,l ε (x)µ(dx) =
D k F̂ (r)\F l k,+ k,+ (r)
k=1 X X

X 1
m,l 1
γ (µ(Dkεl )) 2 (µ(F̂k,+ (r) \ Fk,+
εl
(r))) 2 .
k=1

! 12
The k th summand here does not exceed both a2−k and µ(U+ (r +εm ))−µ(U+ (r +εl )) ,

and since U+ (r + εm ) ↓ U+ (r) we get:


Z
lim sup |fnm (x) − fnl (x)|µ(dx) = 0. (9)
m→∞ X n
l→∞

Let C be the Banach space of all bounded sequences α = {αn } with the norm kαk =
sup |αn | and let L1c (X, F , µ) denote the Banach space of all integrable C-valued functions.
n
(m)
Equality (9) means that the sequence of C-valued functions f (m) (x) = {fn (x)} is a
Cauchy sequence in L1C (X, F , µ). Since the space L1C (X, F , µ) is complete there is a C-
valued integrable function f˜(x) = {f˜n (x)} such that
Z
lim sup |fnm (x) − f˜n (x)|µ(dx) = 0.
m→∞ X n
15
Therefore there exist a subsequence {mk } such that

lim sup |fnmk (x) − f˜n (x)| = 0


k→∞ n

for µ-almost all x ∈ X. On the other hand, it is obvious that


Z Z
m def
lim fn (x) = IU+ (r) (τs x, τt x)λn (ds)λn (dt) = fn (x).
m→∞ S S

Thus for almost all x ∈ X we have f˜n (x) = fn (x) and

lim sup |fnmk (x) − fn (x)| = 0. (10)


k→∞ n

We have now

|fn (x) − ϕ+ (r)| ≤ sup|flmk (x) − fl (x)| + |fnmk (x)−


l
εm εmk
ν(U+ k (r))| + |ν(U+ (r)) − ϕ(r)|.

It was proved earlier that the second summand tends to 0 as n → ∞. Therefore on a set
Qr of µ-measure 1
εmk
lim |fn (x) − ϕ(r)| ≤ sup|flmk (x) − fl (x)| + |ν(U+ (r)) − ϕ(r)|. (11)
n→∞ l

εmk
Since limk→∞ ν(U+ (r)) = ν(U+ (r)) = ϕ+ (r) relations (10) and (11) imply that for
x ∈ Qr
lim fn (x) = ϕ(r). (12)
n→∞

Thus lim C(x, λn , r) = ϕ(r) for µ-almost all x ∈ X. Similarly, we can prove that
n→∞
lim C− (x, λn , r) = ϕ− (r) for µ-almost all x ∈ X.
n→∞
The uniform convergence follows from (12) by the usual Glivenko-Cantelli type argu-
ments (see for example, [B], p.275).

Proof of Theorem 2.4. Consider the following action τ̃ of the group S × S in the
space (X × X, F × F , µ × µ) : τ̃s,t (x, y) = (σs x, τty).
It is easy to see that this action is ergodic. Using the notation U+ (r) = {(x, y) : ρ(x, y) ≤
r} we can write: Z

C(x, y, κn, r)= IU+ (r) τ̃s,t (x, y) κn (ds dt)
S×S

and since {κn } is an averaging system in S × S we have for µ × µ-almost all pairs (x, y) :
Z Z
lim C(x, y, κn , r)= IU+ (r) (x, y)µ (dx)µ (dy)=ϕ(r).
n→∞
X X
16
Proof of Theorem 2.7. Suppose that µ1 is absolutely continuous with respect to
µ2 . Let us fix some δ > 0. Denote f (x) = dµ 1
dµ2 (x). Let Cδ be a number such that
µ1 {x : f (x) < Cδ } ≥ 1 − 3δ and Xδ = {x : f (x) < Cδ }. By Luzin’s theorem there is
a closed set Γδ ⊂ X such that µ1 (Γδ ) > 1 − 3δ and f (x) is continuous on Γδ . Denote
Aδ = Xδ ∩ Γδ . We have µ1 (Aδ ) > 1 − 2δ 3
. Note that Aδ is an open set in the induced
topology of Γδ . We denote by Aδ,r the set of all points x ∈ Aδ such that B(x, r) ∩ Γδ ⊂ Aδ .
It is evident that Aδ,r ↑ Aδ as r ↓ 0 and, hence, there exists a set Qδ = Aδ,rδ such that
µ1 (Qδ ) > 1 − δ. If x ∈ Qδ and r ≤ rδ then µ1 (B(x, r)) ≤ Cδ µ2 (B(x, r)). For Y ∈ Qδ and
i = 1, 2 denote
Z
(i)
ϕY,q (r) = (µi (B(x, r)))q−1µi (dx),
Y
(i)
1 log ϕY,q (r)
ᾱY,q (µi ) = lim .
q − 1 r→0 log r

For any Y ⊂ Qδ we have


Z
(1) Cδ (2)
ϕY,q (r) ≤ (µ2 (B(x, r)))q−1f (x)µ2 (dx) ≤ Cδq ϕY,q (r).
q−1
Y

Hence, ᾱY,q (µ1 ) > ᾱY,q (µ2 ). For any γ > 0 we can find a number δγ > 0 such that δγ ↓ 0
if γ ↓ 0 and µ1 (E) > 1 − γ if µ2 (E) > 1 − δγ for a measurable E. Then µ2 (Qδγ ) > 1 − γ
(1) (1)
and, since Y1 ⊂ Y2 implies ϕY1 ,q (r) ≤ ϕY2 ,q (r) and ᾱY1 ,q (µ1 ) ≥ ᾱY2 ,q (µ1 ), we have

β̄q (µ1 ) = lim sup{ᾱZ,q (µ1 )|Z : Z ∈ F , µ1 (Z) > 1 − δγ }


γ→0

≤ lim sup{ᾱZ∩Qδγ ,q (µ1 )|Z : Z ∈ F , µ1 (Z) > 1 − δγ }


γ→0

≤ lim sup{ᾱZ∩Qδγ ,q (µ2 )|Z : Z ∈ F , µ2 (Z) > 1 − γ}


γ→0

≤ lim sup{ᾱW,q (µ2 )|W : W ∈ F , µ2 (W ) > 1 − 2γ} = β̄q (µ2 ).


γ→0

We proved that β̄q (µ1 ) ≤ β̄q (µ2 ) if µ1 is absolutely continuous with respect to µ2 . By
symmetry β̄q (µ1 ) ≥ β̄q (µ2 ) if µ2 is absolutely continuous with respect to µ1 and hence,
β̄q (µ1 ) = β̄q (µ2 ) if µ1 is equivalent to µ2 . The equality β q (µ1 ) = β q (µ2 ) can be proved in
the same way.

Proof of Theorem 2.9. Fix δ > 0 and ε > 0. Denote

log µ(B(x, r))


Xδ,ε = {x : dµ (x) − ε ≤ ≤ d¯µ (x) + ε if r < δ}.
log r
17
For any Y ∈ F we set
Z
1 log ϕY (r) 1 log ϕY (r)
ϕY (r) = µ(B(x, r)µ(dx), ᾱq (Y ) = lim , αq (Y ) = lim ,
q r→0 log r q r→0 log r
Y

d¯∗ (Y ) = essinf d¯µ (x), d∗ (Y ) = essinf dµ (x).


x∈Y x∈Y

For x ∈ Xδ,ε and r < δ we have:


¯
r dµ (x)+ε ≤ µ(B(x, r)) ≤ r dµ (x)−ε .

For any Y ∈ F and any η > 0 consider the set

Yη = {x : x ∈ Y, d̄µ (x) ≤ d¯∗ (Y ) + η}.

It is obvious that µ(Yη ) > 0 for every η > 0. If Y ⊂ Xδ,ε we have


R R
log (µ(B(x, r)))q−1µ(dx) log Yη
(µ(B(x, r)))q−1µ(dx)
Y
≤ ≤
log r log r
log[r (d̄∗ (Y )+ε+η)(q−1) µ(Yη )] log µ(Yη )
= (q − 1)(d¯∗ (Y ) + ε + η) +
log r log r

and hence, ᾱq (Y ) ≤ d¯∗ (Y ) + ε + η. Therefore, ᾱq (Y ) ≤ d¯∗ (Y ) + ε if Y ⊂ Xδ,ε .


It is clear that if Y1 ⊂ Y2 then ϕY1 ≤ ϕY2 and hence ᾱq (Y1 ) ≥ ᾱq (Y2 ). It is also obvious
that, for any ε, Xδ,ε ↑ X and µ(Xδ,ε ) ↑ µ(X) as δ ↓ 0. Therefore

sup{ᾱq (Z)|Z : µ(Z) ≥ 1 − γ}


≤ sup{ᾱq (Z ∩ Xδ,ε )|Z, δ : µ(Z) ≥ 1 − γ, µ(Xδ,ε) ≥ 1 − γ}
≤ sup{αq (Y )|Y : Y ⊂ Xδ,ε , µ(Y ) ≥ 1 − 2γ}
≤ sup{d∗ (Y ) + ε|Y : µ(Y ) ≥ 1 − 2γ}
≤ d∗ (X) + ε + η

if γ < 13 µ(X). This implies for any ε > 0, η > 0

β̄q (µ) = lim sup{αq (Y )|Y : µ(Y ) > 1 − γ} ≤ d∗ (X) + ε + η


γ→0

We have also
αq (Y ) ≥ d∗ (Y ) − ε, Y ⊂ Xδ,ε
and

β q (µ) ≥ lim sup{d∗ (Y )|Y : µ(Y ) > 1 − γ} − ε


γ→0
≥ sup{d∗ (Y )|Y : µ(Y ) = 1} − ε = d∗ (X) − ε.
18
Since ε and η are arbitrary positive numbers our statement is proved.

Proof of statements in Example 2.11. Choose arbitrary 0 < α < ᾱ < 1. Let
an = αn , where 0 < α < 1, and {nk } an increasing sequence of integers. Consider a
sequence {bn } that satisfies the following conditions: 1) bn = β n if n3k ≤ n ≤ n3k+1 ; 2)
bn = γ n if n3k+2 ≤ n ≤ n3k+3 ; 3) bn ≤ bn+1 , bn+1 ≤ Cbn for all n ≥ 0 (C is a constant).
We choose numbers α, β, γ such that α < 14 , 0 < β < γ < α1 and β 2 > α1 . Thus
def P
p = an bn < ∞.
One can choose the sequence nk growing so fast that

X
n ∞
X
log(αγ 2 )/ log α = lim log b2i ai / log ai <
n→∞
i=1 i=n
X
n ∞
X
< lim log b2i ai / log ai = log(αβ 2 )/ log α.
n→∞
i=1 i=n

One can now choose α, β, γ such that

α − 1 = log(αγ 2 )/ log α, ᾱ − 1 = log(αβ 2 )/ log α

P

αn
(see [P1] for detailed arguments). Set rn = ai = 1−α .
i=n
Let µ be the measure on the interval [0, p] that is absolutely continuous with respect to
the Lebesgue measure with the density function f (x) given as follows

f (x) = bn if rn ≤ x < rn−1 .

We first compute ϕ(rn ). Consider an interval Ii = [ri , ri−1 ] of length |Ii | = ai−1 where
1 ≤ i ≤ n. We decompose Ii into three subintervals

(1) (2) (3)


Ii = [ri , ri + rn ], Ii = [ri + rn , ri−1 − rn ], Ii = [ri−1 − rn , ri−1 ].

It is easy to see that


 (1)
 rn bi + (x − ri )bi + (ri − x + rn )bi+1
 if x ∈ Ii
(2)
µ(B(x, rn )) = 2rn bi if x ∈ Ii

 (3)
rn bi + (ri−1 − x)bi + (x + rn − ri−1 )bi−1 if x ∈ Ii

Integrating this over x ∈ Ii with respect to the measure µ we have


Z Xn
bi bi+1 + bi bi−1
µ(B(x, rn))µ(dx) = rn (2b2i ai + rn − rn b2i ).
2 i=1
x∈Ii
19
Taking the sum over i = 1, . . . , n we obtain
Z X
n X
n X
bi bi+1 + bi bi−1
pµ(B(x, rn ))µ(dx) = rn (2 b2i ai + rn − rn b2i ).
i=1 i=1
2
rn

One can easily check that


X Zrn ∞
X ∞
X
( ak bk ) ≤
2
µ(B(x, rn ))µ(dx) ≤ ( ak bk )( ak bk )
k=n 0 k=n−1 k=n

P
n
αn
Denote An = b2i ai . Since rn = 1−α
, bi+1 < bi , and bi+1 < Cbi it follows that
i=1

Zp
rn C2 An ≤ µ(B(x, rn))µ(dx) ≤ rn C1 An
0

where C1 > 0, C2 > 0 are constants. This implies that

log ϕ(rn ) log ϕ(rn )


α = lim < lim = ᾱ.
n→∞ log rn log rn

Let ρk be any decreasing sequence of numbers, ρk → 0. Given k > 0 one can find n = n(k)
such that
rn(k) ≤ ρk ≤ rn(k)−1 .
Since the function ϕ(r) is non-decreasing and rn(k)−1 ≤ const × rn(k) it follows that

log ϕ(ρk ) log ϕ(ρk )


α = lim , ᾱ = lim .
k→∞ log ρk k→∞ log ρk

This implies that α = α(µ), ᾱ = ᾱ(µ). Let λ be the Lebesgue measure on [0,p]. By
Theorem 2.6, β(µ) = β(µ) = β(λ) = β(λ) = 1. Obviously, the same is true for the
µ(A)
probability measure µ1 (A) = µ([0,p]) .

Proof of Theorem 2.13. According to [N] there exists an isometric imbedding i of


X into a finite-dimensional Eucledian space Rm . In particular, this means that for any
x ∈ X, r > 0 the intersection of the Rm -ball of radius r centered at x ∈ X with i(X) is a
ball in X; vice versa any ball in i(X) can be obtained in this way. Thus, it is enough to
consider the case when X is the Eucledian space of a finite dimension n. If the function
ϕ(r) is discontinuous for some r > 0 then there exists a set A ⊂ X of positive measure
such that for any x ∈ A 
µ S (n−1) (x, r) > 0
20
(S (k) (x, r) denotes the sphere of dimension k centered at x and of radius r). It is easy to
see that for any distinct points x1 , . . . , xk ∈ X, 0 < k ≤ n, the intersection

\
k
S(x1 , . . . , xk ) = S (n−1) (xi , r) (13)
i=1

is a sphere of radius q = q(x1 , . . . , xk ) ≤ r centered at a point x ∈ X, of dimension


l = l(x1 , . . . , xk ) ≤ n − k. Let Sk be the class of all spheres of positive measure and of the
form (13) (k = 1, . . . , n). The class S1 is the set of all spheres S (n−1) (x, r) with x ∈ A,
and thus the sets A and S1 , have the same cardinality. Since the measure µ is non-atomic
the desired result follows from the following lemma.
Lemma. For any k, 1 ≤ k ≤ n, the class Sk is at most countable.
Proof. Any sphere in Sn is of dimension 0 and actually this is a pair of points y1
and y2 with µ({y1 })+µ({y2 }) > 0. But this is impossible since µ is non-atomic. This
means that Sn = ∅. Assume now that for any l, n ≥ l ≥ k + 1 the set Sl is at most
countable. Let T Tk be the set of all pairs of spheres S (l1 ) (x, g1 ), S (l2 ) (x, g2 ) ∈ Sk such that
µ(S (l1 ) (x, q1 ) S (l2 ) (x, q2 )) = 0. It is clear that Tk is at most countable. On Tthe other hand
if S (l1 ) (x, q1 ) and ST(l2 ) (x, q2 ) are different spheres in Sk with µ(S l1 ) (x, q1 ) S (l2 ) (x, q2 )) >
0 then S (l1 ) (x, g1 ) S (l2 ) (x, q2 ) ∈ Si where k + 1 ≤ i ≤ n; according to our assumption,
the class Si is countable. Thus the set of all different pairs of spheres of Sk is at most
countable and therefore Sk is at most countable, too. The lemma is proved.

21
REFERENCES

[AP] Afraimovich V., Pesin Ya., Traveling waves in lattice models of multi-dimensional
and multi-component media 1. General hyperbolic properties, Nonlinearity (1993).
[APT] Afraimovich V., Pesin Ya., Tempelman A., Traveling waves in lattice models of
multi-dimensional and multi-component media 2. Ergodic properties and dimension,
Chaos (1993).
[B] Billingsley P., Probability and Measure; 2nd edition, Wiley and sons, 1986.
[BS] Bunimovich L., Sinai Ya., Space-time chaos in coupled map lattices. Physica., vol. D37,
1989, pp. 60-82.
[C] Cutler C.D., Connecting Ergodicity and Dimension in Dynamical Systems, Ergod.Th.
and Dyn.Syst. 10 (1990), 451-462.
[ER] Eckman J.P., Ruelle D., Ergodic Theory of Chaos and Strange Attractors, Rev. Mod.
Phys., 57 3 (1985), 617-656.
[GPH] Grassberger P., Procaccia I., Hentschel G.B., On the characterization of chaotic
motions, Lect.Notes Physics 179 (1983), 212-221.
[G] Garsia A., Topics in almost everywhere convergence, Markham Publishing Co., 1970.
[HP] Hentschel G.B., Procaccia I., Physica, vol. 8D, 1983.
[K] Katok A., Bernoulli diffeomorphisms on serfaces., Ann. of Math. 110 (1979).
[LM] Ledrappier F., Misiurewicz, Dimension of invariant measures for maps with exponent
zero, Ergod.Th. and Dyn.Syst. 5 (1985), 595-610.
[N] Nash J., The imbedding problem for Riemannian manifolds, Ann. of Math. 63 # 1
(1956).
[P1] Pesin Ya., Dimension type characteristics for invariant sets of dynamical systems,
Russian Math.Surveys 43:4 (1988), 111-151.
[P2] Pesin Ya., On rigorous mathematical definition of correlation dimension and general-
ized spectrum for dimensions., Journal of Stat.Phys. 7 (1993).
[P3] Pesin Ya., Generalized spectrum for dimensions: the approach based on Carathéodory
construction, Constantin Carathéodory, an International Tribute (ed.by Themistokles
Rassias), vol. 2, 1991, pp. 1108-1119.
[PS] Pesin Ya., Sinai Ya., Space-time chaos in chains of weakly hyperbolic mappings, Ad-
vances in Soviet Math. 3 (1991).
[PY] Pesin Ya., Yue C., Hausdorff Dimension of Measures With Nonzero Lyapunov Expo-
nents and local Product Structure, Preprint, PSU (1993).
[PW] Pesin Ya., Weiss H., On the Dimension of a General Class of Geometrically Defined
Deterministic and Random Cantor-like Sets, Preprint PSU (1993).
[S] Sawyer S., Maximal inequalities of weak type, Ann. Math. 84(1) (1966), 157–173.
[T1] Tempelman A., Ergodic Theorems For Group Actions, Kluwer Acad.Publ., 1992.
[T2] Tempelman A., Consistent estimators of the correlation dimension (to appear).
[Y] Yong L.-S., Dimension, Entropy, and Lyapunov Exponents, Ergod.Th. and Dyn.Syst.
2 (1982), 109-124.

Ya. Pesin A. Tempelman


Department of Mathematics Departments of Mathematics and Statistics
The Pennsylvania State University The Pennsylvania State University
22
University Park, PA 16802 University Park, PA 16802
U.S.A. U.S.A.
Email: pesin@math.psu.edu Email:axt12@psuvm.psu.edu

23

You might also like