Beyond exponential tilting: sharp asymptotics and improved importance sampling for rare events

Source abstract

A classical approach for importance sampling (IS) of a rare event is to tilt the measure using its large deviation minimizer, thereby shifting the proposal mean towards the event set. While this tilt is log-efficient, we show that the coefficient of variation of the resulting estimator still grows like $\varepsilon^{-1/4}$ in the small-noise limit $\varepsilon\downarrow0$. Motivated by sharp large deviation asymptotics and the local geometry of the event set at the minimizer, we propose improvements using a control variate or an anisotropic, non-Gaussian proposal. We prove, up to truncations that hold with high probability, and confirm empirically in high-dimensional examples, that the former bounds the coefficient of variation and the latter makes it vanish, so that a bounded number of samples suffices. Our analysis addresses both the coefficient of variation of the different proposals and their necessary and sufficient sample size based on the Kullback--Leibler divergence of the IS target from the proposal; both yield the same rates in $\varepsilon$. We demonstrate the resulting samplers, including a variant that blends the classical and the new proposal across noise levels, for a stochastic Korteweg--de Vries equation with thousands of random inputs.