On a partition with a lower expected \(\mathcal{L}_2\)-discrepancy than classical jittered sampling (Q2121490)

Let \([0;1]^d=\Omega_1\sqcup\ldots\sqcup \Omega_N\) be a general partition of \(d\)-dimensional unit cube into \(N\) subsets of positive volume. A stratified sample \(\mathcal{P}\) based on this partition is is a set of \(N\) random points, where the \(i\)-th point in \(\mathcal{P}\) is chosen uniformly in the \(i\)-th set of the partition (and stochastically independent of the other points). If \(N=m^d\) and the partition consists of the axis-aligned congruent cubes, we obtain so called jittered sampling. The main result states that the set of \(m^d\) random points in \([0;1]^d\) obtained from jittered sampling does not have the minimal expected \(L_2\)-discrepancy among all stratified samples from convex equivolume partitions with the same number of points.