Geometric inequalities for the eigenvalues of concentrated Markov chains

From MaRDI portal
Publication:4503207





The article is devoted to the estimation of the spectral gap of an ergodic reversible Markov chain defined on a finite set. Let \(P(i,j)\) be the transition matrix, \(\pi\) be the stationary distribution of the chain. Then, by Theorem 3.1, if for some \(i^*\), \(\pi(i^*)>v^2\), then the second eigenvalue of \(P\) is \(\lambda_2^2\leq 1-K^2\varphi^2\), where \(v\approx 0.701\) is an absolute constant, \[ K=1-\sqrt{1-\pi(i^*)}(1+\sqrt{\pi(i^*)}), \] \(\varphi\) is the global symmetric isoperimetric constant of the chain. As a corollary one has \(\forall i\) \[ \|P^k(i,\cdot)-\pi(\cdot)\|^2_{\text{TV}}\leq {1-\pi(i)\over 4\pi(i)} (1-K^2\varphi^2)^k, \] where \(\|\cdot\|_{\text{TV}}\) is the total variance norm. This and other similar results are applied to estimate the confidence level of Metropolis-Hastings algorithm approximations to the absolute minimum of some function \(H\). It is noted that these results provide much slow convergence than estimates based on Poincaré's inequalities. But they works even if \(\min_{H(i)\not=H(j)}|H(i)-H(j)|\to 0\) when Poincaré's inequalities fail.











This page was built for publication: Geometric inequalities for the eigenvalues of concentrated Markov chains

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4503207)