Rare tail approximation using asymptotics and $L^1$ polar coordinates

TL;DR

This paper introduces importance sampling estimators for rare tail probabilities of sums of continuous variables using $L^1$ polar coordinates, focusing on practical pre-asymptotic performance across various tail and dependence settings.

Contribution

It proposes a novel importance sampling method leveraging asymptotics and $L^1$ polar coordinates, applicable to both heavy- and light-tailed sums, with demonstrated advantages over existing methods in certain regimes.

Findings

Estimator performs well for independent summands in light- and heavy-tailed cases.

Performance degrades with increasing dimension for dependent subexponential summands.

Focus on practical pre-asymptotic tail probability estimation between 10^{-3} and 10^{-7}.

Abstract

In this work, we propose a class of importance sampling (IS) estimators for estimating the right tail probability of a sum of continuous random variables based on a change of variables to $L^1$ polar coordinates in which the radial and angular components of the IS distribution are considered separately. When the asymptotic behaviour of the sum is known we exploit it for the radial change of measure, and the resulting estimator has the appealing form of the (known) asymptotic multiplied by a random multiplicative correction factor. Given we assume knowledge of the asymptotic behaviour of the sum in this framework, traditional notions of efficiency that appear in the rare-event literature hold little practical meaning here. Instead, we focus on the practical behaviour of the proposed estimator in the pre-asymptotic regime for right tail probabilities between roughly $10^{-3}$ and $10^{-7}$. The proposed estimator and procedure are applicable in both the heavy- and light-tailed settings, as well as for independent and dependent summands. In the case of independent summands, we find that our estimator compares favourably with exponential tilting (iid light-tailed summands) and the Asmussen--Kroese method (independent subexponential summands). However, for dependent subexponential summands using the same simple angular distribution as for the independent case, the performance of our estimator rapidly degenerates with increasing dimension, suggesting an open avenue for further research.