Conditional Quasi-Monte Carlo with Constrained Active Subspaces

TL;DR

This paper introduces a constrained active subspace method for conditional quasi-Monte Carlo, enabling efficient pre-integration in high-dimensional integrals, leading to reduced variance and improved accuracy in applications like finance and chemistry.

Contribution

It develops a novel constrained active subspace approach that ensures tractable pre-integration, enhancing variance reduction techniques in QMC methods.

Findings

Achieves smaller errors than previous methods in numerical experiments.

Provides a computationally-efficient dimension reduction technique.

Demonstrates effectiveness in finance, density estimation, and chemistry applications.

Abstract

Conditional Monte Carlo or pre-integration is a powerful tool for reducing variance and improving the regularity of integrands when using Monte Carlo and quasi-Monte Carlo (QMC) methods. To select the variable to pre-integrate, one must consider both the variable's importance and the tractability of the conditional expectation. For integrals over a Gaussian distribution, any linear combination of variables can potentially be pre-integrated. Liu and Owen (2022) propose to select the linear combination based on an active subspace decomposition of the integrand. However, pre-integrating the selected direction might be intractable. In this work, we address this issue by finding the active subspace subject to constraints such that pre-integration can be easily carried out. The proposed algorithm also provides a computationally-efficient alternative to dimension reduction for pre-integrated functions. The method is applied to examples from computational finance, density estimation, and computational chemistry, and is shown to achieve smaller errors than previous methods.