A quasi-ergodic theorem for evanescent processes (Q1613660)

Motivated by applications to Markov chain Monte Carlo methods, the authors prove a conditioned version of the ergodic theorem for Markov processes. Let \((X,\mathbf P_{x})\) be a strong Markov process with right continuous trajectories with left limits, evolving on some state space \(E\). Let \(\tau \) be its lifetime. A nonnegative measurable function \(f\) on \(E\) is called \(\lambda \)-invariant (for some \(\lambda \in \mathbb R\)) if \(\int _{E}\mathbf 1_{\{\tau >t\}}f(X_{t}) d\mathbf P_{x} = e^{\lambda t}f(x)\) for every \(t>0\). A \(\lambda \)-invariant measure is defined analogously. If \(f\) is a \(\lambda \)-invariant function, one may consider \(X\) as a process with a new state space \(E_{f}=\{0<f<\infty \}\) and under a new measure \(\mathbf Q\) defined on the path space by \[ f(x)\int \mathbf 1_{\{\tau >t\}} F d\mathbf Q_{x} = \int \mathbf 1_{\{\tau >t\}} e^{-\lambda t}f(X_{t})F d\mathbf P_{x} \] for all \(t>0\), \(x\in E_{f}\) and all nonnegative \({\mathcal F}_{t}\)-measurable functions \(F\). We say that \(X\) is positive \(\lambda \)-recurrent provided for some \(\lambda \leq 0\) there exists a \(\lambda \)-invariant function \(f\) such that \((X,\mathbf Q_{x})\) is positive Harris recurrent. Suppose that \(X\) is irreducible positive \(\lambda \)-recurrent, with associated \(\lambda \)-invariant function \(f\) and measure \(\mu \). It is shown that if \(\mu (E_{f})<\infty\), then \[ \lim _{t\to \infty }\mathbf P_{x}\left [\frac 1{t}\int ^{t}_{0} g(X_{s}) ds\biggm |\tau >t\right ] = \frac {\int gf d\mu }{\int f d\mu } \] holds for every bounded measurable function \(g\) and each \(x\in E_f\).