Langevin diffusions and Metropolis-Hastings algorithms (Q1398011)

Based on the Markov chain Monte Carlo technique, stability and robustness (such as geometric ergodicity) of discretization of diffusions (called Langevin diffusions in this paper) with stationary target distributions are considered. More attention is paid to the so called Langevin tempered diffusions, which is the solution of the stochastic differential equation: \[ dX_t=\frac {1-2d}{2} \pi_u^{-2d}(X_t) \nabla \log\pi_u(X_t)\,dt+\pi_u^{-d}(X_t)\,dB_t, \] where \( 0 \leq d \leq \frac{1}{2} \) and \(\pi_u(x)=k\pi(x)\) with unknown constant \(k\), where \( \frac {1-2d}{2}\) plays the role of the temperature. The Euler discretization and the Ozaki discretization are analyzed. A numerical example shows that the later is more stable than the former for the Metropolis-Hastings algorithm.