Convergence rates and moments of Markov chains associated with the mean of Dirichlet processes. (Q2574512)

Let \(\alpha \) be a finite Borel measure on \(\mathbb R\) such that \(\int _ {\mathbb R} \log (1+| y| )\,\text d\alpha (y) <\infty \). Let \(P_ \alpha \) be the Dirichlet process on \(\mathbb R\) with parameter \(\alpha \), that is, a random Borel probability measure on \(\mathbb R\) such that the random vector \((P_ \alpha (A_ 1),\ldots ,P_ \alpha (A_ {k}))\) has a Dirichlet distribution with parameters \((\alpha ( A_ {1}),\ldots ,\alpha (A_ {k}))\) whenever \(\{A_ 1,\ldots ,A_ {k}\}\) is a measurable partition of \(\mathbb R\). Let us denote by \(M_ \alpha \) the distribution of the random variable \(\int _ {\mathbb R} x\,\text dP_ \alpha (x)\). \textit{A.\ Guglielmi} and \textit{R.\ L.\ Tweedie} [Bernoulli 7, 573-592 (2001; Zbl 1005.62073)] proposed a Markov chain Monte Carlo method, based on a certain Markov chain \(\mathbf X\), to obtain approximate samples from \(M_ \alpha \). In the paper under review, necessary and sufficient conditions for geometric and polynomial ergodicity of the chain \(\mathbf X\) are obtained. In particular, if \(\alpha \) is not a point mass, then \(\mathbf X\) is geometrically ergodic if and only if \(\int _ {\mathbb R} | y| ^ {s}\,\text d\alpha (y) <\infty \) for some \(s>0\).