Explicit error bounds for lazy reversible Markov chain Monte Carlo (Q998975)

For a Markov kernel with stationary distribution \(\pi\), the conductance \(\phi\) is defined by \[ \phi = inf_{0<\pi(A)\leq \frac {1}{2}}\frac {{\int_A }K(x,A^c)\pi(dx)}{\pi(A)}. \] A Markov chain is called lazy, if it stays at least with probability \(\frac {1}{2}\) in the current state. Thus a lazy chain always makes the second eigenvalue positive. In the present paper, for approximating of an integral of a bounded integrand \(f\) with respect to \(\pi\), the mean square error of the time average with large enough burn-in time \(n_0\) of the Markov chain with stationary distribution \(\pi\) and initial distribution \(\nu\) bounded with respect to \(\pi\) is proven bounded by \(\frac {10}{\phi \sqrt{\pi}}||f||_{\infty}\). Then this result is applied to the Metropolis algorithm based on a \(\delta\) ball walk.