A new method of randomization of lattice rules for multiple integration

TL;DR

This paper introduces a new practical randomization method for lattice rules in multiple integration, reducing bias compared to traditional approaches that use limited random bits.

Contribution

The paper proposes a novel randomization technique using extended lattice rules with more points, significantly decreasing bias in integral estimation.

Findings

New randomization method reduces estimator bias

Extended lattice rules improve accuracy of integration

Practical implementation avoids need for infinite random bits

Abstract

Cranley and Patterson put forward the following randomization as the basis for the estimation of the error of a lattice rule for an integral of a one-periodic function over the unit cube in s dimensions. The lattice rule is randomized using independent random shifts in each coordinate direction that are uniformly distributed in the interval [0,1]. This randomized lattice rule results in an unbiased estimator of the multiple integral. However, in practice, random variables that are independent and uniformly distributed on [0,1] are not available, since this would require an infinite number of random independent bits. A more realistic practical implementation of the Cranley and Patterson randomization uses rs independent random bits, in the following way. The lattice rule is randomized using independent random shifts in each coordinate direction that are uniformly distributed on {0, 1/2^r, ... ,(2^r-1)/2^r}, where r may be large. For a rank-1 lattice rule with 2^m quadrature points and r >= m, we show that this randomized lattice rule leads to an estimator of the multiple integral that typically has a large bias. We therefore propose that these rs independent random bits be used to perform a new randomization that employs an extension, in the number of quadrature points, to a lattice rule with 2^(m+sr) quadrature points (leading to embedded lattice rules).This new randomization is shown to lead to an estimator of the multiple integral that has much smaller bias.