On the exchange of intersection and supremum of \({\sigma}\)-fields in filtering theory (Q1936815): Difference between revisions

The author deals with applications of the ergodic theory to nonlinear filtering problems. More precisely, let \(E\) and \(F\) be Polish spaces, and let \((X_k, Y_k)\), \(k\in \mathbb Z\), be an \(E\times F\)-valued stochastic process (stationary hidden Markov model) such that (i) \((X_k, Y_k)\) is a stationary Markov process and (ii) there exist transition kernels \(P\) from \(E\) to \(E\) and \(\Phi\) from \(E\) to \(F\) such that \(\operatorname{P}[(X_n, Y_n)\in A|X_{n-1}, Y_{n-1}] =\int \operatorname{1}_A(x,y)P(X_{n-1}, dx)\Phi(x, dy)\). In applications, \((X_k)\), \(k\in \mathbb Z\), normally represents a ``hidden process'' which is not directly observable, while the observable process \((Y_k)\), \(k\in \mathbb Z\), represents ``noisy observations'' of the hidden process. The nonlinear filter \((\pi_k)_{k\geq0}\), \(\pi_n =\operatorname{P}[X_n\in\cdot |Y_1,\dotsc, Y_n]\), is defined as the conditional distribution of the current state of the hidden process given the observations \(\{Y_1,\dotsc, Y_n\}\) of the process \((Y_k)\), \(k\in \mathbb Z\). The filtering process \((\pi_k)_{k\geq0}\) is itself a Markov process taking values in the space \({\mathcal P}(E)\) of probability measures on \(E\), whose transition kernel \(\Pi\) can be expressed in terms of the transition kernels \(P\) and \(\Phi\) that determine the model. Following \textit{H. Kunita} [J. multivariate Anal. 1, 365--393 (1971; Zbl 0245.93027)], the author is interested in the structure of the space of \(\Pi\)-invariant probability measures in \({\mathcal P}({\mathcal P}(E))\). For every \(\Pi\)-invariant measure \(m\in{\mathcal P}({\mathcal P}(E))\) the barycenter \(\mu\in{\mathcal P}(E)\) of \(m\) must be invariant for the transition kernel \(P\) of the hidden process. Conversely, for every \(\Pi\)-invariant measure \(\mu\in{\mathcal P}(E)\), there exists at least one \(\Pi\)-invariant measure \(m\in{\mathcal P}({\mathcal P}(E))\) whose barycenter is \(\mu\); however, the latter need not be unique. Theorem 1 (Kunita). Let \(\operatorname{P}[X_0\in\cdot]:=\mu\) be the \(\Pi\)-invariant measure defined by the stationary hidden Markov model \((X_k, Y_k)\), \(k\in \mathbb Z\), as above. If \[ \bigcap_{n\leq0}\left( {\mathcal F}^Y_{-\infty,0}\vee {\mathcal F}^X_{-\infty,n}\right)= {\mathcal F}^Y_{-\infty,0}\quad \operatorname{P}\text{-a.s.}, \] then there exists a unique \(\Pi\)-invariant measure with barycenter \(\mu\). The converse holds if, in addition, \(\Phi\) possesses a transition density with respect to some \(\sigma\)-finite reference measure. Though the main ideas of the proof are implicitly contained in [Zbl 0245.93027], this simple and general statement does not appear in the literature without various additional simplifying assumptions. For completeness, and in order to make this paper self-contained, the author includes the proof of the statement in the appendix. Kunita's main theorem states that, if the hidden process \((X_k)\), \(k\in \mathbb Z\), is purely non-deterministic (that is, \(\bigcap_{n\leq0}{\mathcal F}^X_{-\infty,n}\) is \(\operatorname{P}\)-a.s. trivial), then there exists a unique \(\Pi\)-invariant measure with barycenter \(\mu\). (In fact, Kunita's paper is written in the context of a continuous time model with white noise observations. None of these specific features are used in the proofs, however.) Kunita's proof, however, does not establish this claim. Indeed, at the crucial point in the proof, Kunita implicitly takes for granted that an exchange of intersection and supremum is permitted, i.e., \[ \bigcap_{n\leq0}\left( {\mathcal F}^Y_{-\infty,0}\vee {\mathcal F}^X_{-\infty,n}\right)= {\mathcal F}^Y_{-\infty,0}\;\bigcap_{n\leq0}{\mathcal F}^X_{-\infty,n}\quad \operatorname{P}\text{-a.s.} \] If this exchange were justified, then Kunita's result would indeed follow immediately from Theorem 1. However, in general, such an exchange of intersection and supremum is not permitted, as it is shown in this article. In his article [Ann. Probab. 37, No. 5, 1876--1925 (2009; Zbl 1178.93142)], the author conjectured that Kunita's claim would hold true whenever the observations are nondegenerate. The goal of the paper is to settle, in the negative, this natural conjecture on the validity of the exchange of intersection and supremum of \(\sigma\)-fields in filtering theory.