On the Discrepancy of Jittered Sampling

TL;DR

This paper analyzes the discrepancy of jittered sampling in high dimensions, establishing bounds on the expected star-discrepancy and suggesting potential improvements over purely random sampling.

Contribution

It provides new bounds on the expected discrepancy of jittered sampling sets and introduces a partition principle that reduces expected squared $L^2$-discrepancy.

Findings

Established bounds on expected star-discrepancy for jittered sampling

Used sharp inequalities like Dvoretzky-Kiefer-Wolfowitz and Bernstein

Partition principle shows jittered sampling reduces $L^2$-discrepancy

Abstract

We study the discrepancy of jittered sampling sets: such a set $\mathcal{P} \subset [0,1]^d$ is generated for fixed $m \in \mathbb{N}$ by partitioning $[0,1]^d$ into $m^d$ axis aligned cubes of equal measure and placing a random point inside each of the $N = m^d$ cubes. We prove that, for $N$ sufficiently large, $$ \frac{1}{10}\frac{d}{N^{\frac{1}{2} + \frac{1}{2d}}} \leq \mathbb{E} D_N^*(\mathcal{P}) \leq \frac{\sqrt{d} (\log{N})^{\frac{1}{2}}}{N^{\frac{1}{2} + \frac{1}{2d}}},$$ where the upper bound with an unspecified constant $C_d$ was proven earlier by Beck. Our proof makes crucial use of the sharp Dvoretzky-Kiefer-Wolfowitz inequality and a suitably taylored Bernstein inequality; we have reasons to believe that the upper bound has the sharp scaling in $N$. Additional heuristics suggest that jittered sampling should be able to improve known bounds on the inverse of the star-discrepancy in the regime $N \gtrsim d^d$. We also prove a partition principle showing that every partition of $[0,1]^d$ combined with a jittered sampling construction gives rise to a set whose expected squared $L^2-$discrepancy is smaller than that of purely random points.