www.imstat.org/aihp 2009, Vol. 45, No. 4, 932–942
DOI: 10.1214/08-AIHP302
© Association des Publications de l’Institut Henri Poincaré, 2009
A probabilistic ergodic decomposition result
Albert Raugi
IRMAR, Université de Rennes I, Campus de Beaulieu, 35042 Rennes Cedex, France E-mail: raugi@univ-rennes1.fr
Received 14 April 2008; accepted 1 September 2008
Abstract. Let(X,X, μ)be a standard probability space. We say that a sub-σ-algebraBofXdecomposesμin an ergodic way if any regular conditional probabilityBP with respect toBandμsatisfies, forμ-almost everyx∈X,∀B∈B,BP (x, B)∈ {0,1}. In this case the equalityμ(·)=
XBP (x,·)μ(dx), gives us an integral decomposition in “B-ergodic” components.
For any sub-σ-algebraBofX, we denote byBthe smallest sub-σ-algebra ofXcontainingBand the collection of all setsA inXsatisfyingμ(A)=0. We say thatBisμ-complete ifB=B.
Let{Bi: i∈I}be a non-empty family of sub-σ-algebras which decomposeμin an ergodic way. Suppose that, for any finite subsetJofI,
i∈JBi=
i∈JBi; this assumption is satisfied in particular when theσ-algebrasBi,i∈I, areμ-complete. Then we prove that the sub-σ-algebra
i∈IBidecomposesμin an ergodic way.
Résumé. Soit(X,X, μ)un espace probabilisé standard. Nous disons qu’une sous-tribuBdeXdécompose ergodiquementμsi toute probabilité conditionnelle régulièreBPrelativement àBetμ, vérifie, pourμ-presque toutx∈X,∀B∈B,BP (x, B)∈ {0,1}.
Dans ce cas l’égalitéμ(·)=
XBP (x,·)μ(dx), nous donne une décomposition intégrale en composantes “B-ergodiques.”
Pour toute sous-tribuBdeX, nous notonsBla plus petite sous-tribu deXcontenantBet tous les sous-ensembles mesurables deXdeμ-mesure nulle. Nous disons que la tribuBestμ-complète siB=B.
Soit{Bi:i∈I}une famille non vide de sous-tribus deXdécomposant ergodiquementμ. Supposons que, pour toute partie finie JdeI,
i∈JBi=
i∈JBi; cette hypothèse est satisfaite si les tribusBi,i∈I, sontμ-complètes. Alors la sous-tribu i∈IBi décompose ergodiquementμ.
MSC:28A50; 28D05; 60A10
Keywords:Regular conditional probability; Disintegration of probability; Quasi-invariant measures; Ergodic decomposition
1. Introduction
There are several versions of ergodic decomposition theorems in the literature (cf. [3,4,6–8]) which give an integral decomposition of a probability measureμ, on a standard measurable space, in ergodic components. Most of these decompositions are based on abstract results like Choquet’s theorem. A probabilistic approach which can prove to be more convenient due to the properties of the conditional expectation (see [2]) is the following. Let(X,X, μ)be a standard Borel probability space. LetBbe a sub-σ-algebra ofX. We denote byBP a regular conditional probability ofBandμ. We say thatBdecomposesμin an ergodic way if forμ-almost everyx∈X,∀B∈B,BP (x, B)∈ {0,1}.
In this case, the equalityμ(dx)=
X
BP (x,·)μ(dx)gives us an integral decomposition inB-ergodiccomponents.
In [7,8] Shimomura proves that the intersection of a decreasing sequence of separable sub-σ-algebras ofXdecom- posesμin an ergodic way. He also gives an example of a standard probability space and a suitable sub-σ-algebra for which the above decomposition is not ergodic.
Let{Bi: i∈I}be a non-empty family of sub-σ-algebras which decomposeμin an ergodic way. Suppose that, for any finite subsetJ ofI,
i∈JBi=
i∈JBi. The aim of this paper is to prove that the sub-σ-algebra
i∈IBi decomposesμin an ergodic way.
2. Preliminaries
It is not necessary to work with standard Borel spaces. We only need probability space for which any sub-σ-algebra has regular conditional probabilities. In this section we recall some results about this property. On a standard Borel space(X,X), it is well known that, for any probability measureμand any sub-σ-algebraBofX, there exists a regular conditional probability with respect toμandB.
Definition 2.1. Aσ-algebra on a setXis called separable if it is generated by a countable sub-algebra.
Proposition 2.2. Let(X,X)be a measurable space with a separableσ-algebraX.Then two positiveσ-finite mea- sures are equal if they coincide on a countable algebra generatingX.
Definition 2.3. A classCof subsets ofXis said to be compact if,for any sequence(Cn)n∈Nof elements ofCwith an empty intersection
n∈NCn,there exists a natural integerpsuch thatp
n=0Cn=∅.
Definition 2.4. Let (X,X, μ) be a probability space. Let C be a compact subclass of X. We say that C is μ- approximating if
∀A∈X μ(A)=sup
μ(C):C∈C, C⊂A .
Definition 2.5. Let(X,X, μ)be a probability space.LetBbe a sub-σ-algebra ofX.We call regular conditional probability with respect toBandμa mapP fromX×Xto[0,1]such that:
(i) for anyx∈X,P (x,·)is a probability measure onX.
(ii) for anyA∈X,the mapx∈X→P (x, A)is a version of the conditional expectationEμ[1A|B];that is,this map isB-measurable and,for anyB∈B,
X
1A(x)1B(x)μ(dx)=
X
P (x, A)1B(x)μ(dx).
Then for any non-negative(or bounded)X-measurable functionf,the functionPf defined byPf (x)=
Xf (y)× P (x,dy)(expectation offwith respect to the probabilityP (x,·))is a version of the conditional expectationEμ[f|B].
Theorem 2.6 ([5], corollaire Proposition V-4-4). Let(X,X, μ)be a probability space with a separableσ-algebra Xcontaining aμ-approximating compact class.
Then,for any sub-σ-algebraBofXthere exists a regular conditional probability with respect toBandμ.
Remarks. Let(X,X, μ)be a probability space with a separableσ-algebraXcontaining aμ-approximating compact class.LetBbe a sub-σ-algebra ofX.
1. IfP andQare two regular conditional probabilities with respect toBandμthen,forμ-almost everyx∈X, the probability measuresP (x,·)andQ(x,·)are equal.
2. IfP is a regular conditional probability with respect toBandμ,then for anyB∈B,we have,forμ-almost everyx∈X,
P (x, B)=Eμ[1B|B](x)=1B(x)=δx(B)∈ {0,1}, whereδxis the Dirac measure at the pointx.
When theσ-algebraBis separable,from Proposition2.2we can permute “for anyB∈B” and “forμ-almost everyx∈X.”
3. LetBbe the smallest sub-σ-algebra containingBand the collection of all setsAinXsatisfyingμ(A)=0.
One sees easily that any regular conditional probability with respect toBandμis a regular conditional probability with respect toBandμ.For two sub-σ-algebrasB1andB2 ofX,the sub-σ-algebraB1∩B2 is not necessarily equal toB1∩B2;consequentlyL2(X,B1∩B2, μ)is not necessarily equal toL2(X,B1∩B2, μ).
3. Main results
Throughout this section, we assume that(X,X, μ)is a probability space with a separableσ-algebraXcontaining a μ-approximating compact class. The preceding Remark 2 leads us to introduce the following definition.
Definition 3.1. We say that a sub-σ-algebraBofXdecomposesμin an ergodic way if one(and thus all)regular conditional probabilityBP with respect toBandμsatisfies,forμ-almost everyx∈X,∀B∈B,BP (x, B)∈ {0,1}.
From the preceding Remarks 2 and 3 it follows that:
• Any separable sub-σ-algebra ofXdecomposesμin an ergodic way.
• IfBdecomposesμin an ergodic way then so doesB. (Any regular conditional probabilityBP with respect toB andμis a regular conditional probability with respect toBandμ. If Bdecomposesμin an ergodic way then, forμ-almost everyx∈X,BP (x, B)∈ {0,1}for anyB∈Band a fortiori for anyB∈B. Which proves thatB decomposesμin an ergodic way.)
Lemma 3.2. LetBbe a sub-σ-algebra ofX,and letBP be a regular conditional probability with respect toBandμ.
Then,forμ-almost everyx∈X,we have the probability equalities
BP (y,·)=BP (x,·) forBP (x,·)-almost everyy∈X and consequently for anyB∈B,we have forμ-almost everyx∈X,
∀A∈X
X
1A(y)1B(y)BP (x,dy)=
X
BP (y, A)1B(y)BP (x,dy).
The following proposition tells us that the sub-σ-algebraBofXdecomposesμin an ergodic way if and only if, in the last equalities, we can permute “for anyB∈B” and “forμ-almost everyx∈X.”
Proposition 3.3. LetBbe a sub-σ-algebra ofX,and letBP be a regular conditional probability with respect toB andμ.
Then the two following assertions are equivalent:
(i) Bdecomposesμin an ergodic way;
(ii) Forμ-almost everyx∈X,BP is a regular conditional probability with respect toBandBP (x,·).
In this case,for any sub-σ-algebraCofBand any regular conditional probabilityCP with respect toCandμ,for μ-almost everyx∈X,BP is a regular conditional probability with respect toBandCP (x,·);that is,forμ-almost everyx∈X,
∀(A, B)∈X×B
X
BP (y, A)1B(y)CP (x,dy)=
X
1A(y)1B(y)CP (x, dy).
Moreover,this assertion is true for any sub-σ-algebraCofXsuch that,forμ-almost everyx∈X,
CPBP (x,·)def=
X
CP (x,dy)BP (y,·)=CP (x,·).
This last property is satisfied whenCis a sub-σ-algebra ofB.
Theorem 3.4. Let (X,X, μ)be a probability space with a separable σ-algebraX containing aμ-approximating compact class.Let{Bi: i∈I}be a non-empty family of sub-σ-algebras which decomposeμin an ergodic way.We suppose that,for any finite subsetJ ofI,
i∈JBi=
i∈JBi. Then the sub-σ-algebra
i∈IBi decomposesμin an ergodic way.
4. Proof of the results
Throughout this section, we assume that(X,X, μ)is a probability space with a separableσ-algebraXcontaining a μ-approximating compact class. For any sub-σ-algebraBofX, we denote byBP a regular conditional probability with respect toBandμ.
4.1. Proof of Lemma3.2
LetA∈X. The functionsg(x)=BP (x, A)and(g(x))2areB-measurable. Therefore, forμ-almost everyx∈X,
BP g(x)=Eμ[g|B](x)=g(x) and BP g2(x)=g2(x)=B P g(x)2
.
From the Cauchy–Schwarz equality it follows that, forμ-almost everyx∈X,
BP (x, A)=g(x)=g(y)=BP (y, A) forBP (x,·)-almost everyy∈X.
The first assertion of the lemma is then a consequence of Proposition 2.2.
For anyB∈Band forμ-almost everyy∈X,
BP (y, B)=Eμ[1B|B](y)=1B(y).
Asμ(dy)=
X
BP (x,dy)μ(dx), forμ-almost everyx∈X,
1B(y)=BP (y, B) forBP (x,·)-almost everyy∈X.
Hence, for anyA∈Xand forμ-almost everyx∈X,
X
1A(y)1B(y)BP (x,dy)=
X
1A(y)BP (y, B)BP (x,dy)
=
X
1A(y)BP (x, B)BP (x,dy) (first assertion)
=BP (x, A)BP (x, B)
=
X
BP (x, A)1B(y)BP (x,dy)
=
X
BP (y, A)1B(y)BP (x,dy) (first assertion). (1)
The Proposition 2.2 allows us to permute “for anyA∈X” and “forμ-almost everyx∈X.”
4.2. Proof of Proposition3.3
LetX0be a measurable subset ofXsuch thatμ(X0)=1 and for anyx∈X0,
BP (y,·)=BP (x,·) forBP (x,·)-almost everyy∈X.
(i)⇒(ii) If theσ-algebraBdecomposesμin an ergodic way, then there exists a measurable subsetX1ofXsuch thatμ(X1)=1 and for anyx∈X1,
∀B∈B BP (x, B)∈ {0,1}.
Forx∈X0∩X1, we have for any(A, B)∈X×B,
X
1A(y)1B(y)BP (x,dy)=BP (x, A)BP (x, B)=
X
BP (y, A)1B(y)BP (x,dy),
which shows that, for anyx∈X0∩X1,BP is a regular conditional probability with respect toBandBP (x,·).
(ii)⇒(i) Assume there exists a measurable subsetX2ofXsuch thatμ(X2)=1 and for anyx∈X2,
∀A∈X BP (y, A)=EBP (x,·)[1A|B](y) forBP (x,·)-almost everyy∈X.
Then forx∈X0∩X2, we have, for anyB∈B,
BP (x, B)=BP (y, B)=EBP (x,·)[1B|B](y)=1B(y) forBP (x,·)-almost everyy∈X.
Hence the assertion (i).
To prove the last assertion of the proposition we need the following lemma.
Lemma 4.1. LetBandCbe two sub-σ-algebras ofXsuch thatC⊂B.Then forμ-almost everyx∈X,we have the probability equalities
CPBP (x,·)=BPCP (x,·)=CP (x,·).
Proof. For anyA∈X, we have the classicalμ-almost everywhere equalities Eμ Eμ[1A|B]|C
=Eμ Eμ[1A|C]|B
=Eμ[1A|C].
Moreover, iff andgare non-negative (or bounded) measurable functions, we know that:f =g μ-a.e.⇒Eμ[f|B] = Eμ[g|B]μ-a.e. It follows that, for anyA∈X,
CPBP (x, A)=BPCP (x, A)=CP (x, A) forμ-almost everyx∈X.
Then the result follows from Proposition 2.2.
Assume (ii), forμ-almost everyz∈X, we have: for any(A, B)∈X×B,
X
1A(y)1B(y)BP (z,dy)=
X
BP (y, A)1B(y)BP (z,dy).
Asμ(dz)=
X
CP (x,dz)μ(dx), forμ-almost everyx∈XandCP (x,·)-almost everyz∈X, for any(A, B)∈X×B,
X
1A(y)1B(y)BP (z,dy)=
X
BP (y, A)1B(y)BP (z,dy).
Integration byCP (x,dz)gives us, forμ-almost everyx∈X,
∀(A, B)∈X×B
X
1A(y)1B(y)CPBP (x,dy)=
X
BP (y, A)1B(y)CPBP (x,dy).
Then the result follows from Lemma 4.1.
4.3. Proof of Theorem3.4 Case of twoσ-algebras We need the following result.
Theorem 4.2. Let(Ω,F,P)be a probability space.LetF1[resp.F2]be a sub-σ-algebra ofF;we callP1[resp.P2] the operator of conditional expectation relative toF1[resp.F2]on the spaceL1(Ω,F,P).
Then,forf ∈L1(Ω,F,P),the sequences of functions
1 n
n−1
k=0
(P1P2)kf
n≥1
and
1 n
n−1
k=0
(P2P1)kf
n≥1
convergeP-almost everywhere and in normL1(P)towardsEP[f|F1∩F2].
Proof. It’s a consequence of the classical ergodic theorem of E. Hopf (see [5], Proposition V-6-3). To identify the limit we note that:P2P1is the dual operator ofP1P2and, as the operatorsP1andP2are idempotent, the common
limit isP1- andP2-invariant (see also [1]).
LetB1andB2be two sub-σ-algebras ofXdecomposingμin an ergodic way which satisfyB1∩B2=B1∩B2. We setC=B1∩B2.
The above theorem tells us that, for any f ∈L1(X,X, μ)the sequence of functions(1nn−1
k=0(B1PB2P )kf )n≥1
convergesμ-a.e. and inL1(μ)-norm towardsCPf =CPf. Asμ(·)=
X
CP (x,·)μ(dx), it follows that, for anyf ∈L1(X,X, μ)and forμ-almost everyx∈X, the sequence of functions(n1n−1
k=0(B1PB2P )kf )n≥1convergesCP (x,·)-a.e. towardsCPf.
We callX0 a measurable subset ofX, such that μ(X0)=1 and for any x∈X0, fori∈1,2, BiP is a regular conditional probability with respect toBiandCP (x,·)(Proposition 3.3).
The same theorem tells us that, for any x∈X0 and for any f ∈L1(X,X,CP (x,·)), the sequence of functions (1nn−1
k=0(B1PB2P )kf )n≥1convergeCP (x,·)-a.e. and inL1(CP (x,·))-norm towardsECP (x,·)[f|B1∩B2]where, for i=1 or 2,Bi is theCP (x,·)-completedσ-algebra ofBi.
LetX be a countable subalgebra ofXgeneratingX. From above and Lemma 3.2, it follows that, forμ-almost any x∈X, for anyA∈X,
forCP (x,·)-almost everyy∈X, CP (x, A)=CP (y, A)=ECP (x,·)[1A|B1∩B2](y).
We deduce that, forμ-almost everyx∈X,
∀(A, C)∈X×(B1∩B2)
X
1A(y)1C(y)CP (x,dy)=
X
CP (y, A)1C(y)CP (x,dy).
These equalities extend to the couples(A, C)∈X×(B1∩B2)(Proposition 2.2).
SinceC=B1∩B2⊂B1∩B2, the above equalities show that, forμ-almost everyx∈X,CP is a regular con- ditional probability with respect toC andCP (x,·). From the Proposition 3.3, the σ-algebraCdecomposesμin an ergodic way.
Case of a sequence ofσ-algebras
Let(Bn)n≥1be a sequence of sub-σ-algebras ofXwhich decomposeμin an ergodic way and satisfy the hypothesis of Theorem 3.4.
For anyn≥2, we have
1≤i≤n
Bi⊂
1≤i≤n−1
Bi∩Bn⊂
1≤i≤n
Bi.
From our hypothesis, it follows that
1≤i≤n
Bi=
1≤i≤n
Bi
and consequently
1≤i≤n
Bi=
1≤i≤n−1
Bi∩Bn=
1≤i≤n
Bi.
We set
∀n≥1 Cn= n
k=1
Bk and C=
k≥1
Bk.
From the case treated previously, we prove by induction that, for anyn≥1, theσ-algebraCn decomposesμin an ergodic way. From Proposition 3.3, forμ-almost everyx∈X,
∀(A, C)∈X×Cn
X
1A(y)1C(y)CP (x,dy)=
X Cn
P (y, A)1C(y)CP (x,dy).
The decreasing martingale theorem implies that, for anyA∈Xand forμ-almost everyx∈X,Eμ[1A|Cn](x) −→
n→+∞
Eμ[1A|C](x). Consequently, for anyA∈Xand forμ-almost everyx∈X,CnP (x, A) −→
n→+∞
CP (x, A).
Asμ(·)=
XCP (x,·)μ(dx), it follows that: for anyA∈Xand forμ-almost everyx∈X, forCP (x,·)-almost everyy∈X, CnP (y, A) −→
n→+∞
CP (y, A).
While limiting itself to elementsC ofC, the dominated convergence theorem implies that, for anyA∈Xand for μ-almost everyx∈X,
∀C∈C
X
1A(y)1C(y)CP (x,dy)=
X
CP (y, A)1C(y)CP (x,dy).
Now from Proposition 2.2, we can permute “for anyA∈X” and “forμ-almost everyx∈X,” which shows thatC decomposesμin an ergodic way.
The preceding proof shows the following corollary which improves Shimomura’s result.
Corollary 4.3. Let(Bi)i∈Nbe a sequence of sub-σ-algebras ofX.If for anyn∈Ntheσ-algebran
i=0Bi decom- posesμin an ergodic way,then the intersection
i∈NBi decomposesμin an ergodic way.
Case of an uncountable family ofσ-algebras We need the following lemmas.
Lemma 4.4. Let (H,·,·) be a separable Hilbert space.Let {Vi: i∈I} a noncountable family of closed vector subspaces ofH.Then there exists a countable subsetJ ofI such that
i∈IVi=
i∈JVi. Proof. We easily see that the orthogonal complementV⊥ of V =
i∈IVi inH is equal to Vect(
i∈IVi⊥)(the closure of the subspace generated by
i∈IVi⊥). We choose a dense sequence of vectors(un)n≥1of Vect(
i∈IVi⊥)in V⊥. The Schmidt orthonormalization process allows us to extract a maximal orthonormal system; that is, an Hilbert basis(en)n≥1deV⊥. For anyp≥1,epis a finite linear combination of vectors from{un: n≥1}; eachun is itself a finite linear combination of vectors of
i∈IVi⊥. Therefore, there exists a countable subsetJ ofI such that,∀p≥ 1, ep∈Vect(
i∈JVi⊥). HenceV⊥=Vect(
i∈JVi⊥)andV =
i∈JVi.
Lemma 4.5. LetBbe a sub-σ-algebra ofXwhich decomposesμin an ergodic way.LetCbe a sub-σ-algebra of Bsuch that,for any boundedB-measurable functionf,there exists a boundedC-measurable functiongsatisfying f =g μ-a.e.
ThenCdecomposesμin an ergodic way.
Proof. From Proposition 3.3, forμ-almost everyx∈X,BP is a regular conditional probability with respect toBand
CP (x,·), that is,
∀(A, B)∈X×B
X
1A(y)1B(y)CP (x,dy)=
X
BP (y, A)1B(y)CP (x,dy).
In these equalities, we have to replaceBP (y, A)byCP (y, A). From Proposition 2.2, we can restrict our equalities toA∈X a separable sub-algebra ofXgeneratingX.
Letf be a boundedB-measurable function. There exists a bounded C-measurable functiong such thatf =g μ-a.e. Then we have,μ-a.e.,
Eμ[f|B] =f=g=Eμ[g|C] =Eμ[f|C]. It follows that, forμ-almost everyy∈X,
∀A∈X BP (y, A)=CP (y, A).
Now fromμ(dy)=
X
CP (x,dy)μ(dx)we deduce that, forμ-almost everyx∈X, forCP (x,·)-almost everyy∈X,
∀A∈X BP (y, A)=CP (y, A) and consequently, forμ-almost everyx∈X,
∀(A, B)∈X×B
X
1A(y)1B(y)CP (x,dy)=
X
CP (y, A)1B(y)CP (x,dy).
Hence the result.
Lemma 4.6. Let {Bn: n∈N∗} be a sequence of sub-σ-algebras of X satisfying, for any n≥2,
1≤k≤nBk =
1≤k≤nBk. Then
n≥1Bk=
n≥1Bk.
Proof. Letf ∈L1(X,X, μ). From the decreasing martingale theorem, we have,μ-almost everywhere, Eμ
f
k≥1
Bk
= lim
n→+∞Eμ
f
1≤k≤n
Bk
= lim
n→+∞Eμ
f
1≤k≤n
Bk
= lim
n→+∞Eμ
f
1≤k≤n
Bk
=Eμ
f
k≥1
Bk
. (2)
Hence the result.
Consider the separable Hilbert space L2(X,X, μ). It is well known that, for each sub-σ-algebra B of X, the spaceL2(X,B, μ)is identified to a closed subspace ofL2(X,X, μ)and the conditional expectation relative toBis identified to the orthogonal projection onto this closed subspace.
Let{Bi: i∈I}be an uncountable family ofμ-complete sub-σ-algebras which decomposeμin an ergodic way and satisfy the hypothesis of Theorem 3.4. From the Lemmas 4.4 and 4.6, there exists a countable subsetJ ofI such that
L2
X,
i∈I
Bi, μ
⊂
i∈I
L2(X,Bi, μ)=
i∈J
L2(X,Bi, μ)=L2
X,
i∈J
Bi, μ
.
It follows that for anyf ∈L2(X,
i∈IBi, μ)there exists a functiong∈L2(X,
i∈JBi, μ) such thatf =g, μ-a.e. According to the case treated previously, we know that the sub-σ-algebra
i∈JBidecomposesμin an ergodic way. Then the result follows from Lemma 4.5.
5. Examples and applications
1. Let(X,X, μ)be a probability space with a separableσ-algebraXcontaining a μ-approximating compact class.
Letτ be an invertible bi-measurable transformation of(X,X)such thatμis quasi-invariant for the action ofτ. We consider the sub-σ-algebraJ=Jτ of Xdefined by J= {B∈X: τ−1(B)=B}. Then the following result is well known.
Proposition 5.1. Theσ-algebraJdecomposesμin an ergodic way.
Proof. An idea of the proof is the following. We consider the contractionT ofL1(X,X, μ)defined by Tf (x)=f◦τ−1(x)d(τ (μ))
d(μ) (x).
Replacingτ byτ−1we obtain the inverse operatorT−1.
From the Chacon–Ornstein ergodic theorem, one proves [2] that, with obvious notations: for anyf∈L1(X,X, μ) and forμ-almost everyx∈X,
n
k=−n
Tkf (x) n
k=−n
Tk1(x) −→
n→+∞
JPf (x).
One sees easily that there exists a measurable subsetX0ofXsuch thatμ(X0)=1 and for anyx∈X0the proba- bilityJP (x,·)isτ-quasi-invariant withd(τdJJP (x,·)P (x,·)) =d(τ μ)dμ .
Then the same ergodic theorem tells us that, for anyx∈X0, for anyf ∈L1(X,X,JP (x,·))and forJP (x,·)-almost everyy∈X,
n
k=−n
Tkf (y) n
k=−n
Tk1(y) −→
n→+∞EJP (x,·)[f|J](x).
As in the first case of Theorem 3.4, we prove that forμ-almost everyx∈X,JP is a regular conditional probability with respect toJandJP (x,·). The result follows from Proposition 3.3.
In [3], Greshchonig and Schmidt consider the case of a Borel action of a locally compact second countable group Gon a standard probability space(X,X, μ); that is, a group homomorphismg→τgfromGinto the group Aut(X) of Borel automorphisms ofXsuch that the map (g, x)→τgx fromG×XtoX is Borel andμ is quasi-invariant under eachτg,g∈G. They prove that theσ-algebra
g∈GJτg decomposesμin an ergodic way.
The Theorem 3.4 makes it possible to find and improve this result.
Corollary 5.2. Let{τi: i∈I}be a non-empty family of Borel automorphisms ofX.Then the sub-σ-algebra
i∈IJτi ofXdecomposesμin an ergodic way.
Proof. Taking into account the Proposition 5.1, it is enough to show that, for any finite subsetJ ofI,
i∈JJτi=
i∈JJτi. Letf be a
i∈JJτi-measurable function. We setX0=
i∈J{f◦τi=f}; we haveμ(X0)=1.
We callGthe algebraic subgroup of Aut(X)generated by the Borel automorphisms{τi: i∈J};Gis a countable subset of Aut(X). The subsetX1=
s∈GsX0ofX0belongs to
i∈JJτiandμ(X1)=1. Then the functiong=f1X1
is
i∈JJτi-measurable andf=gμ-a.e. Which shows thatf is
i∈JJτi-measurable.
2. Let(X,X, μ, τ )be a dynamical system with a polish space and a not necessarily invertible transformation. We denote by (Y,F, λ, η) the natural extension of our dynamical and byπ the natural projection of Y ontoX. With obvious notations, one sees easily thatf isJη∩π−1(X)-measurable (resp.Jη∩π−1(X)-measurable) if and only if there existsg∈Jτ such thatf =g◦π (resp.f =g◦π λ-a.e.). It follows that
Jη∩π−1(X)=Jη∩π−1(X).
We know that theσ-algebraJηdecomposesλin an ergodic way. Theσ-algebraπ−1(X)is separable. Therefore theσ-algebraC=Jη∩π−1(X)decomposesλin an ergodic way.
LetP be a regular conditional probability with respect toJτ andμ. LetQbe a regular conditional probability with respect toCandλ. For anyA∈XandC∈Jτ we have:
X
P (x, A)1C(x)μ(dx)=
X
1A(x)1C(x)μ(dx) and therefore
Y
P
π(y), A 1C
π(y)
λ(dy)=
Y
1A
π(y) 1C
π(y) λ(dy)
=
Y
Eλ[1A◦π|C](y)1C
π(y) λ(dy)
=
Y
Q
y, π−1(A) 1C
π(y)
λ(dy). (3)
Which proves, via the Proposition 2.2, that forλ-almost everyy∈Y, P
π(y),·
=Q
y, π−1(·) and theσ-algebraJτ decomposesμin an ergodic way.
3. Let P be a transition probability on a measurable space (X,X) with a separable σ-algebra Xcontaining a μ-approximating compact class.
We denote byΠ the set ofP-invariant probability measures on(X,X):
π∈Π ⇔
X
f (x)π P (dx)=
X
Pf (x)π(dx)=
X
f (x)π(dx)
for any non-negative or bounded measurable functionf onX. We assume thatΠ=∅.
For anyπ∈Π, we denote byBπ the sub-σ-algebra ofXdefined by:
Bπ= {A∈X: P1A=1Aπ-a.e.} and we setB=
π∈ΠBπ.
Letπ∈Π. Letf be a boundedBπ-measurable function onX. The functiong, defined by g(x)=lim inf1
n
n−1
k=0
Pkf (x),
satisfiesg=f π-a.e. andP g≤g. From the latter inequality, it follows that, for anyσ ∈Π,P g=g σ-a.e. andgis B-measurable. We deduce thatBπ=Bπ-a.e.
From the Hopf theorem ([3], Proposition V-6-3), for anyf ∈L1(X,X, π ), the sequences of functions
1 n
n−1
k=0
d((f π )Pk) dπ
n≥1
and
1 n
n−1
k=0
Pkf
n≥1
convergeπ-almost everywhere and in norm L1(π )towards BπPf =BPf. As in Example 1, one deduces thatB decomposesπin an ergodic way.
Acknowledgments
I wish to thank B. Bekka and S. Gouëzel for helpful discussions and the referees for their careful reading and valuable comments.
References
[1] D. L. Burkholder and Y. S. Chow. Iterates of conditional expectations operators.Proc. Amer. Math. Soc.12(1961) 490–495. MR0142144 [2] J.-P. Conze and A. Raugi. On the ergodic decomposition for a cocycle. Preprint, 2007. (pdf version “ReducErg.pdf” in personal university
site.)
[3] G. Greschonig and K. Schmidt. Ergodic decomposition of quasi-invariant probability measures.Colloq. Math. 84/85 (2000) 495–514.
MR1784210
[4] J. Kerstan and A. Wakolbinger. Ergodic decomposition of probability laws.Z. Wahrsch. Verw. Gebiete56(1981) no. 3 399–414. MR0621119 [5] J. Neveu. Bases Mathématiques du Calcul des Probabilités. Masson, Paris, 1964. MR0198504
[6] K. Schmidt. A probabilistic proof of ergodic decomposition.Sankhy¯a Ser. A40(1978) no. 1 10–18. MR0545459
[7] H. Shimomura. Ergodic decomposition of quasi-invariant measures.Publ. Res. Inst. Math. Sci.14(1978) no. 2 359–381. MR0509194 [8] H. Shimomura. Remark to the ergodic decomposition of measures.Publ. Res. Inst. Math. Sci.26(1990) no. 5 861–865. MR1082320