Uniform projection designs (Q1731775)

A computer or physical experiment often involves a large number of factors. Latin hypercube designs (LHDs) are very popular in such situations. If only a few of the large number of factors are active, designs with good projection properties are desirable. Unfortunately, LHDs may have bad projections onto lower dimensions. There are several proposals in the relevant literature to overcome this disadvantage, e.g. by improvements of space-filling properties and consideration of further criteria like orthogonality or maximin distance criteria. In this paper the authors propose a new criterion, called uniform projection criterion, by focussing on projection uniformity. Uniform projection designs have good space-filling properties in terms of distance, uniformity and orthogonality. Some theoretical results are presented, e.g. that maximin equidistant designs are uniform projection designs. A class of asymptotically optimal uniform projection designs based on good lattice point designs is constructed. The results are applied to a multidrug combination experiment.