This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Poisson-Pinsker factor and infinite measure preserving group actions

Emmanuel Roy Laboratoire Analyse Géométrie et Applications, UMR 7539, Université Paris 13, 99 avenue J.B. Clément, F-93430 Villetaneuse, France roy@math.univ-paris13.fr
Abstract.

We solve the question of the existence of a Poisson-Pinsker factor for conservative ergodic infinite measure preserving action of a countable amenable group by proving the following dichotomy: either it has totally positive Poisson entropy (and is of zero type), or it possesses a Poisson-Pinsker factor. If GG is abelian and the entropy positive, the spectrum is absolutely continuous (Lebesgue countable if G=G=\mathbb{Z}) on the whole L2L^{2}-space in the first case and in the orthocomplement of the L2L^{2}-space of the Poisson-Pinsker factor in the second.

Key words and phrases:
Poisson suspensions, infinite ergodic theory, joinings
2000 Mathematics Subject Classification:
Primary 37A40, 37A35, 60G51; secondary 37A15, 37A50
This paper has been written during the MSRI program “Ergodic Theory and Additive Combinatorics” in Berkeley. The author is very grateful to this institution and to the organizers of this program for funding his research during this semester.

1. Introduction

Poisson entropy for a σ\sigma-finite measure preserving \mathbb{Z}-action was introduced in [10] as the Kolmogorov entropy of its Poisson suspension (see Section 2 for the definition of a Poisson suspension), it is a non-trivial invariant which coincides with Kolmogorov entropy in the finite measure case and retains most of its basic features. The definition readily extends to more general group actions as soon as it is well defined in the finite measure preserving case. For conservative \mathbb{Z}-action, there exist other definitions of entropy due to Krengel and Parry (see [8] and [9]) and it is proved in [7] that these two entropies equal Poisson entropy for a large family of systems called “quasi-finite”, however it has been proved recently in [3] that Krengel entropy can be different from Poisson and Parry entropies. The question of the existence of a Pinsker factor for the Krengel entropy (a Krengel-Pinsker factor) was already addressed (and unsolved) in the original Krengel’s paper. The first partial answer was due to Aaronson and Park in [1] where they proved the existence of a Krengel-Pinsker factor for LLB systems (a subfamily of quasi-finite systems). In [7], it is proved that as soon as there exists a (σ\sigma-finite) factor of zero Poisson entropy, there exists a maximal factor with zero Poisson entropy (a Poisson-Pinsker factor).

The method used to obtain the existence of Poisson-Pinsker factor was very particular to \mathbb{Z}-actions and was only partially satisfactory since it was not possible to prove that a system generated by a square integrable function with singular spectral measure had zero Poisson entropy nor a conservative system with finite spectral multiplicity, as it is well known in the finite measure case. In this short paper, we will get, with completely different methods, a much more satisfactory picture of the situation (extended to countable amenable group actions) by showing, more generally, that the Pinsker factor of a Poisson suspension, if not trivial, is always (isomorphic to) a Poisson suspension.

These results strongly rely on the structure of joinings of Poisson suspensions (see [4], [10] and [11]) on the one hand, and entropy results for countable amenable group actions on the other (see [6], [5] and [2]), in particular the relative disjointness results of Thouvenot [12] for \mathbb{Z}-actions, generalized in [6] for countable amenable group actions.

2. Background on Poisson suspensions

We recall a few facts about Poisson suspensions that can be found in [11]. Note that in this section, we mention structural features and results that were once proved in the case of \mathbb{Z}-actions, however, it can be seen that the extension to more general actions is immediate. In the following GG is a countable amenable group with identity element ee.

2.1. Definition of Poisson suspensions

Let (X,𝒜,μ)\left(X,\mathcal{A},\mu\right) a σ\sigma-finite Lebesgue space with an infinite continuous measure μ\mu. Let us define the probability space (X,𝒜,μ)\left(X^{*},\mathcal{A}^{*},\mu^{*}\right) where:

  • XX^{*} stands for the space of measures of the form ν=nδxi\nu=\sum_{n\in\mathbb{N}}\delta_{x_{i}} where xiXx_{i}\in X.

  • 𝒜\mathcal{A}^{*} is the smallest σ\sigma-algebra such that, for any A𝒜A\in\mathcal{A}, the map NA:νν(A)N_{A}:\;\nu\mapsto\nu\left(A\right) from (X,𝒜)\left(X,\mathcal{A}\right) to (¯,𝒫(¯))\left(\overline{\mathbb{N}},\mathcal{P}\left(\overline{\mathbb{N}}\right)\right) is measurable.

  • μ\mu^{*} is the only probability measure on (X,𝒜)\left(X^{*},\mathcal{A}^{*}\right) such that, for any integer kk, and any collection of disjoint sets {A1,,Ak}\left\{A_{1},\dots,A_{k}\right\} of finite measure, the random variables NA1,,NAkN_{A_{1}},\dots,N_{A_{k}} are independent and follow a Poisson distribution with parameter μ(A1),,μ(Ak)\mu\left(A_{1}\right),\dots,\mu\left(A_{k}\right) respectively.

From now on, 𝒜\mathcal{A}^{*} is assumed complete with respect to μ\mu^{*}, the probability space (X,𝒜,μ)\left(X^{*},\mathcal{A}^{*},\mu^{*}\right) is a Lebesgue space called the Poisson measure over (X,𝒜,μ)\left(X,\mathcal{A},\mu\right).

Let GG be a countable group acting on (X,𝒜,μ)\left(X,\mathcal{A},\mu\right) by measure preserving automorphisms TgT^{g}, gGg\in G. If TT is a measure preserving automorphism of (X,𝒜,μ)\left(X,\mathcal{A},\mu\right), then TT_{*} defined on (X,𝒜)\left(X^{*},\mathcal{A}^{*}\right) by T(ν)(A)=ν(T1A)T_{*}\left(\nu\right)\left(A\right)=\nu\left(T^{-1}A\right), A𝒜A\in\mathcal{A}, is a measure preserving automorphism of (X,𝒜,μ)\left(X^{*},\mathcal{A}^{*},\mu^{*}\right). (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) is called the Poisson suspension over the base (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right).

2.2. Poisson entropy

The Poisson entropy of a system (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) is defined as the usual (Kolmogorov) entropy of its Poisson suspension (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right). (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) will be said to have totally positive Poisson entropy if, for any invariant set KXK\subset X of positive measure the Poisson entropy of the restricted system (K,𝒜K,μK,Tg)\left(K,\mathcal{A}_{\mid K},\mu_{\mid K},T^{g}\right) is positive on ANY factor. Note that we reserve the terminology factor to invariant sub-σ\sigma-algebras on which the measure is still σ\sigma-finite (the trivial σ\sigma-algebra {X,}𝒜\left\{X,\emptyset\right\}\subset\mathcal{A} is NOT a factor of (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right)), therefore, complete positive entropy (cpe) is a different notion which is not adapted to the infinite measure context. Of course a probability measure preserving system never has totally positive entropy since the trivial algebra is a factor.

2.3. Fock space structure

It is classical to see L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right) as the Fock space over L2(𝒜)L^{2}\left(\mathcal{A}\right), that is:

L2(𝒜)L2(𝒜)L2(𝒜)2L2(𝒜)nL^{2}\left(\mathcal{A}^{*}\right)\simeq\mathbb{C}\oplus L^{2}\left(\mathcal{A}\right)\oplus L^{2}\left(\mathcal{A}\right)^{\odot 2}\otimes\cdots\otimes L^{2}\left(\mathcal{A}\right)^{\odot n}\otimes\cdots

where L2(𝒜)nL^{2}\left(\mathcal{A}\right)^{\odot n} stands for the symmetric tensor product of L2(𝒜)L^{2}\left(\mathcal{A}\right) and is called the nn-th chaos. Within L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right), the subspace correcponding to the first chaos is noted \mathfrak{C} and is the closure of the linear span of vectors N(A)μ(A)N\left(A\right)-\mu\left(A\right), A𝒜A\in\mathcal{A}, μ(A)<\mu\left(A\right)<\infty. A linear operator UU on L2(𝒜)L^{2}\left(\mathcal{A}\right) of norm less than 11 induces an operator U~\widetilde{U} on L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right) when defined on L2(𝒜)nL^{2}\left(\mathcal{A}\right)^{\odot n} by U~(ff)=UfUf\widetilde{U}\left(f\otimes\cdots\otimes f\right)=Uf\otimes\cdots\otimes Uf. We have in particular UTg~=UTg\widetilde{U_{T^{g}}}=U_{T_{*}^{g}}.

There is a distinguished family of vectors, linearly dense in L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right); namely, for fL2(𝒜)f\in L^{2}\left(\mathcal{A}\right) which are finite linear combination of indicator functions,

f:=1f12!ff1n!ff,\mathcal{E}_{f}:=1\oplus f\oplus\frac{1}{\sqrt{2!}}f\otimes f\oplus\dots\oplus\frac{1}{\sqrt{n!}}f\otimes\cdots\otimes f,

which corresponds, through the above identification, to:

f(ν)=exp(Xf𝑑μ)x,ν(x)=1(1+f(x)),νX.\mathcal{E}_{f}\left(\nu\right)=\exp\left(-\int_{X}fd\mu\right){\displaystyle\prod_{x,\nu\left(x\right)=1}}\left(1+f\left(x\right)\right),\quad\nu\in X^{*}.

In case GG is abelian, if σ\sigma is the maximal spectral type of UTgU_{T^{g}} on L2(𝒜)L^{2}\left(\mathcal{A}\right) then the reduced maximal spectral type of UTgU_{T_{*}^{g}} is k=11k!σk\sum_{k=1}^{\infty}\frac{1}{k!}\sigma^{*k} since the maximal spectral type on the kk-th chaos is σk\sigma^{*k}.

2.4. Poissonian factors

A Poissonian factor is a sub-σ\sigma-algebra of 𝒜\mathcal{A}^{*} of the form (K):=σ{NA,AK}\left(\mathcal{B}_{\mid K}\right)^{*}:=\sigma\left\{N_{A},\;A\in\mathcal{B}_{\mid K}\right\} where KXK\subset X is a TgT^{g}-invariant measurable set of positive μ\mu-measure and K\mathcal{B}_{\mid K} is a factor of the restricted system (K,𝒜K,μK,Tg)\left(K,\mathcal{A}_{\mid K},\mu_{\mid K},T^{g}\right). In terms of systems, the factor (K)\left(\mathcal{B}_{\mid K}\right)^{*} corresponds to the Poisson suspension (K,(K),(μK),Tg)\left(K^{*},\left(\mathcal{B}_{\mid K}\right)^{*},\left(\mu_{\mid K}\right)^{*},T_{*}^{g}\right). The trivial factor {X,}𝒜\left\{X^{*},\emptyset\right\}\subset\mathcal{A}^{*} is also considered as Poissonian.

2.5. Infinite divisibility and Poissonian joinings

Addition is well defined on (X,𝒜)\left(X^{*},\mathcal{A}^{*}\right) as the usual sum of measures and so is convolution of probability measures on (X,𝒜)\left(X^{*},\mathcal{A}^{*}\right). A probability measure pp on (X,𝒜)\left(X^{*},\mathcal{A}^{*}\right) such that, for any integer kk, there exists a probability measure pkp_{k} satisfying p=pkpkp=p_{k}*\cdots*p_{k} (kk terms) is said to be infinitely divisible. It is well known that μ\mu^{*} is infinitely divisible as μ=(1kμ)(1kμ)\mu^{*}=\left(\frac{1}{k}\mu\right)^{*}*\cdots*\left(\frac{1}{k}\mu\right)^{*}.

Addition on the product space (X×X,𝒜𝒜)\left(X^{*}\times X^{*},\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right) is defined coordinate wise and so is convolution and infinite divisibility. A self-joining of a Poisson suspension (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) is said to be a Poissonian self-joining if its distribution on (X×X,𝒜𝒜)\left(X^{*}\times X^{*},\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right) is infinitely divisible. In [11], it is proved:

Proposition 1.

A self-joining, determined by a Markov operator Ψ\Psi on L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right), is a Poissonian self-joining if and only if there exists a sub-Markov operator Φ\Phi on L2(𝒜)L^{2}\left(\mathcal{A}\right) (i.e. a positive operator such that Φ(1)1\Phi\left(1\right)\leq 1 and Φ(1)1\Phi^{*}\left(1\right)\leq 1 ) such that Φ~=Ψ\widetilde{\Phi}=\Psi (see [4] and [11]).

Poissonian self-joinings were originally introduced independently through their Markov operator characterization in [4], and their infinitely divisible one in [10].

The next proposition is found in [11], however, we give a proof as the original one is too sketchy:

Proposition 2.

If \mathcal{F} is a non-trivial factor of (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) and if the relatively independent product of (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) over \mathcal{F} is a Poissonian joining, then \mathcal{F} is a Poissonian factor, that is there exists a TgT^{g}-invariant subset KXK\subset X and a factor K𝒜K\mathcal{B}_{\mid K}\subset\mathcal{A}_{\mid K} such that (K)=\left(\mathcal{B}_{\mid K}\right)^{*}=\mathcal{F}.

Proof.

Let π\pi_{\mathcal{F}} the projection on L2()L^{2}\left(\mathcal{F}\right) that corresponds to the relatively independent joining over \mathcal{F}. Since this joining is Poissonian, there exists a sub-Markov Φ\Phi operator on L2(𝒜)L^{2}\left(\mathcal{A}\right) such that Φ~=π\widetilde{\Phi}=\pi_{\mathcal{F}}. Observe that Φ\Phi is also an orthogonal projection since it is induced by π\pi_{\mathcal{F}} restricted to the first chaos, we therefore have:

XΦ(1A)𝑑μ=X1AΦ(1)𝑑μ.\int_{X}\Phi\left(1_{A}\right)d\mu=\int_{X}1_{A}\Phi\left(1\right)d\mu.

But we also have

XΦ(1A)𝑑μ=XΦ(Φ(1A))𝑑μ=XΦ(1A)Φ(1)𝑑μ,\int_{X}\Phi\left(1_{A}\right)d\mu=\int_{X}\Phi\left(\Phi\left(1_{A}\right)\right)d\mu=\int_{X}\Phi\left(1_{A}\right)\Phi\left(1\right)d\mu,

and this implies

XΦ(1A)(1Φ(1))𝑑μ=0.\int_{X}\Phi\left(1_{A}\right)\left(1-\Phi\left(1\right)\right)d\mu=0.

Therefore, as Φ\Phi is positive and Φ(1)1\Phi\left(1\right)\leq 1, μ\mu-a.e.

Φ(1A)(1Φ(1))=0.\Phi\left(1_{A}\right)\left(1-\Phi\left(1\right)\right)=0.

Replacing AA by a sequence AnA_{n} of finite measure sets increasing to XX, we get:

Φ(1)(1Φ(1))=0\Phi\left(1\right)\left(1-\Phi\left(1\right)\right)=0

and

Φ(1)=(Φ(1))2.\Phi\left(1\right)=\left(\Phi\left(1\right)\right)^{2}.

So Φ(1)\Phi\left(1\right) is the indicator function of a set KK and if we consider the restricted system (K,𝒜K,μK,Tg)\left(K,\mathcal{A}_{\mid K},\mu_{\mid K},T^{g}\right), Φ\Phi is a Markov operator and an orthogonal projection of L2(𝒜K)L^{2}\left(\mathcal{A}_{\mid K}\right), thus a conditional expectation on a factor K𝒜K\mathcal{B}_{\mid K}\subset\mathcal{A}_{\mid K}. Now if fL2(𝒜)f\in L^{2}\left(\mathcal{A}\right), Φf\mathcal{E}_{\Phi f} is (K)\left(\mathcal{B}_{\mid K}\right)^{*}-measurable and, for any gL2(K)g\in L^{2}\left(\mathcal{B}_{\mid K}\right)

πf,gL2(𝒜)=Φf,gL2(𝒜)=expΦf,gL2(K)\left\langle\pi_{\mathcal{F}}\mathcal{E}_{f},\mathcal{E}_{g}\right\rangle_{L^{2}\left(\mathcal{A}^{*}\right)}=\left\langle\mathcal{E}_{\Phi f},\mathcal{E}_{g}\right\rangle_{L^{2}\left(\mathcal{A}^{*}\right)}=\exp\left\langle\Phi f,g\right\rangle_{L^{2}\left(\mathcal{B}_{\mid K}\right)}
=expf,gL2(K)=expf,gL2(𝒜)=f,gL2(𝒜)=πKf,gL2(𝒜)=\exp\left\langle f,g\right\rangle_{L^{2}\left(\mathcal{B}_{\mid K}\right)}=\exp\left\langle f,g\right\rangle_{L^{2}\left(\mathcal{A}\right)}=\left\langle\mathcal{E}_{f},\mathcal{E}_{g}\right\rangle_{L^{2}\left(\mathcal{A}^{*}\right)}=\left\langle\pi_{\mathcal{B}_{\mid K}^{*}}\mathcal{E}_{f},\mathcal{E}_{g}\right\rangle_{L^{2}\left(\mathcal{A}^{*}\right)}

with some slight abuses in notation. Therefore =(K)\mathcal{F}=\left(\mathcal{B}_{\mid K}\right)^{*}. ∎

3. The main results

Proposition 3.

Let (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) be the Poisson suspension of the infinite measure preserving system (X,𝒜,μ,G)\left(X,\mathcal{A},\mu,G\right). Then Π\Pi, the Pinsker factor of the system, is a Poissonian factor.

Proof.

Take α,β>0\alpha,\beta>0 and consider the direct product

(X×X,𝒜𝒜,(αμ)(βμ),Tg×Tg)\left(X^{*}\times X^{*},\mathcal{A}^{*}\otimes\mathcal{A}^{*},\left(\alpha\mu\right)^{*}\otimes\left(\beta\mu\right)^{*},T_{*}^{g}\times T_{*}^{g}\right)

of Poisson suspensions. Thanks to the classical formula (αμ)(βμ)=(α+β)μ\left(\alpha\mu\right)^{*}*\left(\beta\mu\right)^{*}=\left(\alpha+\beta\right)\mu^{*}, the map φ:(ν1,ν2)ν1+ν2\varphi:\,\left(\nu_{1},\nu_{2}\right)\mapsto\nu_{1}+\nu_{2} induces a factor map from

(X×,X𝒜𝒜,(αμ)(βμ),Tg×Tg)\left(X^{*}\times,X^{*}\mathcal{A}^{*}\otimes\mathcal{A}^{*},\left(\alpha\mu\right)^{*}\otimes\left(\beta\mu\right)^{*},T_{*}^{g}\times T_{*}^{g}\right)

to

(X,𝒜,((α+β)μ),Tg).\left(X^{*},\mathcal{A}^{*},\left(\left(\alpha+\beta\right)\mu\right)^{*},T_{*}^{g}\right).

Denote by Πα\Pi_{\alpha} and Πβ\Pi_{\beta} and Πα+β\Pi_{\alpha+\beta} the Pinsker algebra of the systems (X,𝒜,(αμ),Tg)\left(X^{*},\mathcal{A}^{*},\left(\alpha\mu\right)^{*},T_{*}^{g}\right), (X,𝒜,(βμ),Tg)\left(X^{*},\mathcal{A}^{*},\left(\beta\mu\right)^{*},T_{*}^{g}\right) and (X,𝒜,((α+β)μ),Tg)\left(X^{*},\mathcal{A}^{*},\left(\left(\alpha+\beta\right)\mu\right)^{*},T_{*}^{g}\right) respectively and set :=φ1𝒜\mathcal{B}:=\varphi^{-1}\mathcal{A}^{*} and Πα+β~:=φ1Πα+β\widetilde{\Pi_{\alpha+\beta}}:=\varphi^{-1}\Pi_{\alpha+\beta}. Thanks to a classical result (generalized to countable amenable group actions in Theorem 4 in [6]), the Pinsker algebra of the product (X×X,𝒜𝒜,(αμ)(βμ),Tg×Tg)\left(X^{*}\times X^{*},\mathcal{A}^{*}\otimes\mathcal{A}^{*},\left(\alpha\mu\right)^{*}\otimes\left(\beta\mu\right)^{*},T_{*}^{g}\times T_{*}^{g}\right) is ΠαΠβ\Pi_{\alpha}\otimes\Pi_{\beta}. The extension Πα+β~\mathcal{B}\to\widetilde{\Pi_{\alpha+\beta}} is a cpe extension and ΠαΠβΠα+β~\Pi_{\alpha}\otimes\Pi_{\beta}\to\widetilde{\Pi_{\alpha+\beta}} is a zero entropy extension. Therefore, by Lemma 3 in [12] (once again generalized to countable amenable group actions in Theorem 1 in [6]), they are relatively disjoint over Πα+β~\widetilde{\Pi_{\alpha+\beta}}. As a consequence, L2()L2(Πα+β~)L^{2}\left(\mathcal{B}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right) and L2(ΠαΠβ)L2(Πα+β~)L^{2}\left(\Pi_{\alpha}\otimes\Pi_{\beta}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right) are orthogonal in L2(𝒜𝒜)L^{2}\left(\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right). Indeed if fL2()L2(Πα+β~)f\in L^{2}\left(\mathcal{B}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right) and gL2(ΠαΠβ)L2(Πα+β~)g\in L^{2}\left(\Pi_{\alpha}\otimes\Pi_{\beta}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right), we have:

𝔼[fg]\displaystyle\mathbb{E}\left[fg\right]
=𝔼[𝔼[fgΠα+β~]]\displaystyle=\mathbb{E}\left[\mathbb{E}\left[fg\mid\widetilde{\Pi_{\alpha+\beta}}\right]\right]
=𝔼[𝔼[fΠα+β~]𝔼[gΠα+β~]]\displaystyle=\mathbb{E}\left[\mathbb{E}\left[f\mid\widetilde{\Pi_{\alpha+\beta}}\right]\mathbb{E}\left[g\mid\widetilde{\Pi_{\alpha+\beta}}\right]\right]
=0\displaystyle=0

We can therefore decompose L2(𝒜𝒜)L^{2}\left(\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right) into the following orthogonal sum:

L2(𝒜𝒜)\displaystyle L^{2}\left(\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right)
=L2(Πα+β~)(L2()L2(Πα+β~))(L2(ΠαΠβ)L2(Πα+β~))H\displaystyle=L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right)\oplus\left(L^{2}\left(\mathcal{B}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right)\right)\oplus\left(L^{2}\left(\Pi_{\alpha}\otimes\Pi_{\beta}\right)\ominus L^{2}\left(\widetilde{\Pi_{\alpha+\beta}}\right)\right)\oplus H

where HH is the orthocomplement of everything else. Now write fL2(𝒜𝒜)f\in L^{2}\left(\mathcal{A}^{*}\otimes\mathcal{A}^{*}\right) as f=f1+f2+f3+f4f=f_{1}+f_{2}+f_{3}+f_{4} according to the decomposition, we have:

𝔼[𝔼[f]ΠαΠβ]=𝔼[(f1+f2)ΠαΠβ]=f1\mathbb{E}\left[\mathbb{E}\left[f\mid\mathcal{B}\right]\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]=\mathbb{E}\left[\left(f_{1}+f_{2}\right)\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]=f_{1}

thus 𝔼[fΠα+β~]=𝔼[𝔼[f]ΠαΠβ]\mathbb{E}\left[f\mid\widetilde{\Pi_{\alpha+\beta}}\right]=\mathbb{E}\left[\mathbb{E}\left[f\mid\mathcal{B}\right]\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right].

Now form the relatively independent joining of

(X×,X𝒜𝒜,(αμ)(βμ),Tg×Tg)\left(X^{*}\times,X^{*}\mathcal{A}^{*}\otimes\mathcal{A}^{*},\left(\alpha\mu\right)^{*}\otimes\left(\beta\mu\right)^{*},T_{*}^{g}\times T_{*}^{g}\right)

over ΠαΠβ\Pi_{\alpha}\otimes\Pi_{\beta} and remark that it is just the direct product of the relatively independent joinings of (X,𝒜,(αμ),Tg)\left(X^{*},\mathcal{A}^{*},\left(\alpha\mu\right)^{*},T_{*}^{g}\right) and (X,𝒜,(βμ),Tg)\left(X^{*},\mathcal{A}^{*},\left(\beta\mu\right)^{*},T_{*}^{g}\right) over their respective Pinsker factors Πα\Pi_{\alpha} and Πβ\Pi_{\beta}. Let’s now compute the distribution of the self-joining of (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) induced by the preceding joining, through the factor map φ×φ\varphi\times\varphi. Take AA and BB in \mathcal{B} and compute:

𝔼[𝔼[1AΠαΠβ]𝔼[1BΠαΠβ]]\displaystyle\mathbb{E}\left[\mathbb{E}\left[1_{A}\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]\mathbb{E}\left[1_{B}\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]\right]
=𝔼[𝔼[𝔼[1A]ΠαΠβ]𝔼[𝔼[1B]ΠαΠβ]]\displaystyle=\mathbb{E}\left[\mathbb{E}\left[\mathbb{E}\left[1_{A}\mid\mathcal{B}\right]\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]\mathbb{E}\left[\mathbb{E}\left[1_{B}\mid\mathcal{B}\right]\mid\Pi_{\alpha}\otimes\Pi_{\beta}\right]\right]
=𝔼[𝔼[1AΠα+β~]𝔼[1BΠα+β~]]\displaystyle=\mathbb{E}\left[\mathbb{E}\left[1_{A}\mid\widetilde{\Pi_{\alpha+\beta}}\right]\mathbb{E}\left[1_{B}\mid\widetilde{\Pi_{\alpha+\beta}}\right]\right]

Therefore, the joining induced is nothing else than the relatively independent joining over its Pinsker factor. We have just proved that the image measure of

((αμ)Πα(αμ))((βμ)Πβ(βμ))\left(\left(\alpha\mu\right)^{*}\otimes_{\Pi_{\alpha}}\left(\alpha\mu\right)^{*}\right)\otimes\left(\left(\beta\mu\right)^{*}\otimes_{\Pi_{\beta}}\left(\beta\mu\right)^{*}\right)

by the sum application φ×φ\varphi\times\varphi is the measure μΠα+βμ\mu^{*}\otimes_{\Pi_{\alpha+\beta}}\mu^{*} , that is, we have proved the following formula:

((αμ)Πα(αμ))((βμ)Πβ(βμ))=μΠα+βμ\left(\left(\alpha\mu\right)^{*}\otimes_{\Pi_{\alpha}}\left(\alpha\mu\right)^{*}\right)*\left(\left(\beta\mu\right)^{*}\otimes_{\Pi_{\beta}}\left(\beta\mu\right)^{*}\right)=\mu^{*}\otimes_{\Pi_{\alpha+\beta}}\mu^{*}

and we can deduce, for any integer kk:

((1kμ)Π1k(1kμ))k=μΠ1μ.\left(\left(\frac{1}{k}\mu\right)^{*}\otimes_{\Pi_{\frac{1}{k}}}\left(\frac{1}{k}\mu\right)^{*}\right)^{*k}=\mu^{*}\otimes_{\Pi_{1}}\mu^{*}.

This means that the distribution of this relatively independent joining is infinitely divisible, i.e. it is a Poissonian joining. But according to Proposition 2, Π1=Π\Pi_{1}=\Pi is a Poissonian factor.∎

Proposition 4.

Let (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) be a dynamical system. There exists a (possibly trivial) partition into TgT^{g}-invariant sets AA and AcA^{c} such that:

  1. (1)

    (A,𝒜A,μA,Tg)\left(A,\mathcal{A}_{\mid A},\mu_{\mid A},T^{g}\right) possesses a Poisson-Pinsker factor.

  2. (2)

    for any TgT^{g}-invariant set BAcB\subset A^{c} of positive measure, (B,𝒜B,μB,Tg)\left(B,\mathcal{A}_{\mid B},\mu_{\mid B},T^{g}\right) has totally positive Poisson entropy.

Proof.

If the Pinsker factor of (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) is trivial, the system has complete positive entropy, therefore, for any TgT^{g}-invariant set BXB\subset X of positive measure, (B,𝒜B,μB,Tg)\left(B,\mathcal{A}_{\mid B},\mu_{\mid B},T^{g}\right) has totally positive Poisson entropy since any factor B\mathcal{B}_{\mid B} corresponds to the non-trivial Poissonian factor (B)\left(\mathcal{B}_{\mid B}\right)^{*} on which the Poisson suspension has positive entropy. If the Pinsker factor Π\Pi of (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) is not trivial, from Proposition 3, there exists a TgT^{g}-invariant set AXA\subset X of positive measure and a factor, say 𝒫A\mathcal{P}_{\mid A}, of the restricted system (A,𝒜A,μA,Tg)\left(A,\mathcal{A}_{\mid A},\mu_{\mid A},T^{g}\right) such that (𝒫A)=Π\left(\mathcal{P}_{\mid A}\right)^{*}=\Pi. 𝒫A\mathcal{P}_{\mid A} is clearly the Poisson-Pinsker factor of the system (A,𝒜A,μA,Tg)\left(A,\mathcal{A}_{\mid A},\mu_{\mid A},T^{g}\right). Indeed, if 𝒞A𝒜A\mathcal{C}_{\mid A}\subset\mathcal{A}_{\mid A} is a factor with zero Poisson entropy, (𝒞A)(𝒫A)\left(\mathcal{C}_{\mid A}\right)^{*}\subset\left(\mathcal{P}_{\mid A}\right)^{*} as the latter is the Pinsker factor of the suspension, and this implies 𝒞A𝒫A\mathcal{C}_{\mid A}\subset\mathcal{P}_{\mid A} (the fact that a factor \mathcal{R} is the Poisson-Pinsker factor if the associated Poissonian factor \mathcal{R}^{*} is the Pinsker factor of the suspension was already observed in [7]).

If AcA^{c} has positive measure, the Poisson suspension (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) splits into the direct product

(A×(Ac),(𝒜A)(𝒜Ac),(μA)(μAc),Tg×Tg)\left(A^{*}\times\left(A^{c}\right)^{*},\left(\mathcal{A}_{\mid A}\right)^{*}\otimes\left(\mathcal{A}_{\mid A^{c}}\right)^{*},\left(\mu_{\mid A}\right)^{*}\otimes\left(\mu_{\mid A^{c}}\right)^{*},T_{*}^{g}\times T_{*}^{g}\right)

which implies that the Pinsker factor Π\Pi also splits accordingly. But as (𝒫A)=Π\left(\mathcal{P}_{\mid A}\right)^{*}=\Pi, the Pinsker factor is concentrated in the first side of the product, that is, ((Ac),(𝒜Ac),(μAc),Tg)\left(\left(A^{c}\right)^{*},\left(\mathcal{A}_{\mid A^{c}}\right)^{*},\left(\mu_{\mid A^{c}}\right)^{*},T_{*}^{g}\right) has complete positive entropy and we conclude as in the first part of the proof. ∎

In the ergodic case, the result takes the following more pleasant form:

Theorem 5.

Let (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) be an ergodic infinite measure preserving system. Either it has totally positive Poisson entropy, or it possesses a Poisson-Pinsker factor.

As in the \mathbb{Z}-action case, we can observe the behaviour of Poisson entropy with respect to joinings.

Proposition 6.

Zero Poisson entropy is stable under taking joinings. Totally positive entropy systems are strongly disjoint from systems possessing a Poisson-Pinsker factor.

Proof.

The first statement is obvious and the proof of the second is identical to the \mathbb{Z}-action case which can be found in [7]. ∎

4. Spectral properties

We first recall that a system (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) is of zero type if for any measurable sets AA and BB in 𝒜\mathcal{A} of finite measure, μ(ATgB)\mu\left(A\cap T^{g}B\right) tends to zero as gg tends to infinity.

Proposition 7.

If (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) has totally positive entropy, then it is of zero type.

Proof.

From Proposition 3 (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) has complete positive entropy and is therefore mixing. But, thanks to the classical isometry formula

μ(AB)=𝔼μ[(NAμ(A))(NBμ(B))]\mu\left(A\cap B\right)=\mathbb{E}_{\mu^{*}}\left[\left(N_{A}-\mu\left(A\right)\right)\left(N_{B}-\mu\left(B\right)\right)\right]

we have

μ(ATgB)\displaystyle\mu\left(A\cap T^{g}B\right)
=𝔼μ[(NAμ(A))(NTgBμ(TgB))]\displaystyle=\mathbb{E}_{\mu^{*}}\left[\left(N_{A}-\mu\left(A\right)\right)\left(N_{T^{g}B}-\mu\left(T^{g}B\right)\right)\right]
=𝔼μ[(NAμ(A))(NBμ(B))Tg]\displaystyle=\mathbb{E}_{\mu^{*}}\left[\left(N_{A}-\mu\left(A\right)\right)\left(N_{B}-\mu\left(B\right)\right)\circ T_{*}^{g}\right]

which goes to zero as gg tends to infinity.∎

Proposition 8.

If GG is abelian and (X,𝒜,μ,Tg)\left(X^{*},\mathcal{A}^{*},\mu^{*},T_{*}^{g}\right) has positive entropy, then it has absolutely continuous spectrum on (L2(Π))\mathfrak{C}\cap\left(L^{2}\left(\Pi\right)^{\perp}\right).

Proof.

Since, thanks to Proposition 3, Π\Pi has the structure of a Poissonian factor, its associated L2L^{2}-space is a Fock space compatible with the underlying one, in particular =(L2(Π))(L2(Π))\mathfrak{C}=\left(L^{2}\left(\Pi\right)^{\perp}\cap\mathfrak{C}\right)\begin{array}[b]{c}\perp\\ \oplus\end{array}\left(L^{2}\left(\Pi\right)\cap\mathfrak{C}\right). Since, by assumption L2(Π)L2(𝒜)L^{2}\left(\Pi\right)\neq L^{2}\left(\mathcal{A}^{*}\right), L2(Π)\mathfrak{C}\not\subset L^{2}\left(\Pi\right) as σ()=𝒜\sigma\left(\mathfrak{C}\right)=\mathcal{A}^{*}. Therefore (L2(Π))\left(L^{2}\left(\Pi\right)^{\perp}\cap\mathfrak{C}\right) is not empty and we get the result, as the maximal spectral type on L2(Π)L^{2}\left(\Pi\right)^{\perp} is Lebesgue by a Theorem proved in [5] and independently by Thouvenot (unpublished). ∎

As a direct application, we get:

Proposition 9.

Assume GG is abelian. If (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) has totally positive Poisson entropy, its maximal spectral type is absolutely continuous.

If (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) has positive Poisson entropy and possesses a Poisson-Pinsker factor, then the maximal spectral type on the orthocomplement of the Poisson-Pinsker factor is absolutely continuous.

Proof.

This follows directly from Proposition 8 and from the unitary isomorphism between the first chaos of L2(𝒜)L^{2}\left(\mathcal{A}^{*}\right) and L2(𝒜)L^{2}\left(\mathcal{A}\right). ∎

We therefore can deduce the following corollary, well known in the finite measure case, the proof being almost the same:

Corollary 10.

Assume (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) is the dynamical system associated to a square integrable stationary GG-process {Xg}gG\left\{X_{g}\right\}_{g\in G} where GG is a countable abelian group. If the spectral measure of XeX_{e} is singular, then (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) has zero Poisson entropy.

In particular, if (X,𝒜,μ,Tg)\left(X,\mathcal{A},\mu,T^{g}\right) has singular maximal spectral type, then it has zero Poisson entropy.

For \mathbb{Z}-action, we can be more precise.

Proposition 11.

If (X,𝒜,μ,T)\left(X,\mathcal{A},\mu,T\right) has totally positive Poisson entropy, its maximal spectral type is Lebesgue countable.

If (X,𝒜,μ,T)\left(X,\mathcal{A},\mu,T\right) has positive Poisson entropy and possesses a Poisson-Pinsker factor, then the maximal spectral type on the orthocomplement of the Poisson-Pinsker factor is Lebesgue countable.

Proof.

The first statement follows from Proposition 10.2 in [7], combined with Proposition 9. The second can also be deduced from an adaptation of Proposition 10.2 in [7] but is also a direct application of Theorem 3.2 in [2] combined with the fact that conditional Poisson entropy coincides with conditional Krengel entropy as proved in [7]. ∎

Examples of totally positive Poisson entropy transformations are given by shift associated to null recurrent Markov chains (see [11]).

Corollary 12.

If (X,𝒜,μ,T)\left(X,\mathcal{A},\mu,T\right) has finite multiplicity, then it has zero Poisson entropy.

The conclusion of Corollary 12 is not true if μ\mu is not continuous (think of the shift on \mathbb{Z}).

References

  • [1] J. Aaronson and K. K. Park. Predictability, entropy and information of infinite transformations. preprint.
  • [2] I. A. Danilenko and D. J. Rudolph. Conditional entropy theory in infinite measure and a question of Krengel. Israel J. Math. To appear.
  • [3] T. de la Rue and E. Janvresse. Krengel entropy does not kill Poisson entropy. In preparation.
  • [4] Y. Derriennic, K. Fra̧czek, M. Leman´\acute{\mathrm{n}}czyk, and F. Parreau. Ergodic automorphisms whose weak closure of off-diagonal measures consists of ergodic self-joinings. Colloq. Math., 110:81–115, 2008.
  • [5] A. N. Dooley and V. Ya. Golodets. The spectrum of completely positive entropy actions of countable amenable groups. J. Funct. Anal., 196:1–18, 2002.
  • [6] E. Glasner, J. P. Thouvenot, and B. Weiss. Entropy without a past. Ergodic Theory Dynam. Systems, 20:1355–1370, 2000.
  • [7] E. Janvresse, T. Meyerovitch, T. de la Rue, and E. Roy. Poisson suspensions and entropy for infinite transformations. T.A.M.S., 2008. To appear.
  • [8] U. Krengel. Entropy of conservative transformations. Z.Wahrscheinlichkeitstheorie verw. Geb., 7:161–181, 1967.
  • [9] W. Parry. Entropy and generators in ergodic theory. W. A. Benjamin, Inc., New York-Amsterdam, 1969.
  • [10] E. Roy. Mesures de Poisson, infinie divisibilité et propriétés ergodiques. PhD thesis, Université Paris 6, 2005.
  • [11] E. Roy. Poisson suspensions and infinite ergodic theory. Ergodic Theory Dynam. Systems, 29(2):667–683, 2009.
  • [12] J.-P. Thouvenot. Une classe de systèmes pour lesquels la conjecture de Pinsker est vraie. Israel J. Math., 21(2–3):208–214, 1975.