Ergodic properties of the nonlinear filter. (Q1765991): Difference between revisions

The following filtering model is considered: The signal process is a càdlàg homogeneous Markov process \((X_ {t})\) on a Polish space \(E\) with a transition semigroup \((P_ {t})\) which is Feller, has a unique invariant probability measure \(\mu \) and such that \(P_ {t}f\to \langle \mu ,f\rangle \) in \(L^ 1(\mu )\) as \(t\to \infty \) for all bounded continuous real functions \(f\) on \(E\). The observation process is given by \(Y_ {t} = \int ^ {t}_ {0} h(X_ {s})\,ds + W_ {t}\), where \(h:E\to \mathbb R^ {d}\) is a bounded continuous function and \(W\) is a \(d\)-dimensional Wiener process independent of \(X\). Denote by \(P(E)\) the set of all Borel probability measures on \(E\), and let \((\Pi _ {t})\) be the nonlinear filter, i.e.\ the \(P(E)\)-valued process defined by \(\Pi _ {t}(\cdot ) = \mathbf P(X_ {t}\in \cdot \mid \sigma (Y_ {s}, \, 0\leq s\leq t))\). Rather general sufficient conditions for the pair \((X_ {t},\Pi _ {t})\) to be a Feller-Markov process on \(E\times P(E)\) and to have a unique invariant probability measure are found, these conditions being closely related to asymptotic stability properties of the filter.