On the stability and ergodicity of adaptive scaling Metropolis algorithms (Q645599): Difference between revisions

The Markov chain Monte Carlo (MCMC) algorithm is a general method used to approximate integrals of the type \(I = \int_{\mathbb R^d}f(x) \pi(x) dx\), where \(f\) is the integrand and \(\pi\) is a probability density function. The MCMC method is based on a Markov chain \((X_n)_{n \geq 1}\), and the average \(I_n = n^{-1} \sum_{k=1}^n f(X_k)\) converges to the integral \(I\) as \(n \to \infty\). In the present paper the stability and the ergodicity properties of two adaptive random walk Metropolis algorithms are established. In the Introduction a short survey of the adaptive MCMC algorithms is made. The idea of using a scaling parameter \(\theta\) and its optimization in the so-called adaptive scaling Metropolis algorithms (ASM) is developed. Results about the influence of the density \(\pi\) on the correct ergodic properties and the stability of the adaptive Metropolis algorithms are discussed. In Section 2 the stability and the ergodicity of the ASM algorithm are considered. By using a special function \(\phi,\) the process of scaling adaption is realized. In Assumption 1 the conditions that the function \(\phi\) must satisfy are presented. In Assumption 2 conditions about the density are given. In Assumption 3 the conditions for the adaption step size sequence are shown. In Definition 4 the computing scheme of the adaptive scaling Metropolis algorithm is presented. In Definition 5 the computing scheme of the adaptive scaling within adaptive Metropolis (ASWAM) algorithm is presented. Under given assumptions, Theorem 7 states that either the ASM or the ASWAM processes for any bounded function \(f\) satisfy the strong law of large numbers, so the MCMC algorithm converges almost surely to the integral. In Assumption 8 conditions about boundedness and differentiability of the density function \(\pi\) are exposed. In Theorem 9 the density function \(\pi\) satisfies the Assumption 8 and the function \(f\) is uniformly estimated in terms of the function \(\pi\). It is shown that the ASM and the ASWAM processes satisfy the strong law of large numbers. In Section 3 a general framework for scale adaptation covering simultaneously both the ASM and the ASWAM algorithms is described. In Section 4 stability results for general adaptive scaling processes are developed. Theorem 14 is a general stability result. Under conditions about the filtration, the adaption step size sequence and the adapted random variable \(H_{n}\), two stability results are proved. The problem for targets \(\pi\) with an unbounded support is considered. In Proposition 20 the controlled growth for the adaptive scaling process is considered. Section 5 considers ergodicity results for ASN and ASWAM processes. Theorem 22, under the given conditions, states the strong law of large numbers. In Appendices A and B some useful lemmas are proved.