Escaping from an attractor: Importance sampling and rest points. I. (Q748325)

This paper is concerned with the improvement of Monte Carlo schemes for the approximation of finite time exit probabilities of small noise diffusions involving the escape from an equilibrium by importance sampling in the one-dimensional case. More precisely, the question is whether a given diffusion process \(X^\varepsilon\) solving a stochastic differential equation of the form \[ d X^\varepsilon(s) = b(X^\varepsilon(s))ds + \sqrt{\varepsilon}\sigma(X^\varepsilon(s))dB(s) \] defined on an open set \(\mathcal{D}\subset\mathbb{R}\) exits \(O\subset\mathcal{D}\) before time \(T\). When there are no rest points in the domain of interest, an estimator using importance sampling with the Freidlin-Wentsell quasipotential as weight function leads not only to the optimal decay rate as \(\varepsilon\to0\), but also to control of the pre-exponential terms. When there is a rest point, the exponential decay rate dominates for \(\varepsilon\to0\), but the pre-exponential terms destroy the convergence improvements for \(\varepsilon>0\) small but not too small. This issue is overcome by using piecewise solutions joint together by mollification as the importance sampling weight. Near the rest point, an explicit solution to an approximating, time dependent problem is needed and can be found because the diffusion process can be approximated by a Gauss-Markov process, i.e., a process with linear drift and constant diffusion. Away, the time-independent quasipotential still yields the required properties. The theoretical properties of the proposed scheme are proved. Results of numerical experiments demonstrating the non-degradation of the good performance of the scheme in the cases of small noise coefficient \(\varepsilon\) and large time \(T\) are presented.