Bounds on regeneration times and convergence rates for Markov chains (Q1593619)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Bounds on regeneration times and convergence rates for Markov chains |
scientific article |
Statements
Bounds on regeneration times and convergence rates for Markov chains (English)
0 references
17 January 2001
0 references
Let \(\{X_{n}\}_{n\geq 0}\) be a \(\pi \)-irreducible Markov chain on a general state space \(S\), \(\pi \) being an invariant measure. Denote by \(P^{n}\) the \(n\)-step transition function and let \(C\) be a small set, satisfying the minorisation condition \(P(x,\cdot) \geq \varepsilon \varphi \), \(x\in C\), for some \(\varepsilon >0\) and a probability measure \(\varphi \). In many applications, it is of primary importance to establish the geometric ergodicity of the chain, that is, to prove that \[ \|P^{n}(x,\cdot) -\pi \|\leq M(x)\rho ^{n}, \qquad x\in S,\tag{1} \] for some \(M(x)<\infty \) and \(\rho < 1\), where \(\|\cdot \|\) denotes the total variation norm of a signed measure. Many methods of obtaining (1) are based on the knowledge of the tail properties of the time \(T\) of the first regeneration according to the distribution \(\varphi \). New bounds on the tails of \(T\) are obtained in terms of the first hitting time of the set \(C\) and of Foster-Lyapunov functions \(V\) satisfying \(E_{\bullet} V(X_{1}) \leq \lambda V + b\mathbf 1_{C}\), \(b<\infty \), \(\lambda <1\). This leads to new bounds on \(\rho \) in (1), which are shown to be in many cases better than those obtained previously by \textit{J. S. Rosenthal} [J. Am. Stat. Assoc. 90, No. 430, 558-566 (1995; Zbl 0824.60077)] and \textit{S. P. Meyn} and \textit{R. L. Tweedie} [Ann. Appl. Probab. 4, No. 4, 981-1011 (1994; Zbl 0812.60059)].
0 references
geometric ergodicity
0 references
renewal times
0 references
rates of convergence
0 references
Markov chain Monte Carlo
0 references