Wrap-around \(L_2\)-discrepancy of random sampling Latin hypercube and uniform designs (Q1347847)

The aim of the present paper is to compare several important types of designs, viz. the simple random sampling or simple random desin (SRD), the Latin hypercube design (LHD), and the uniform design (UD). For a proper discrepancy among various designs, the authors propose to be used the wrap-around \(L_2\)-discrepancy (denoted as WD) since it is invariant to reordering the runs, relabeling coordinates, and to coordinate shift. The paper derives the expectation and variance for square WD of SRDs and LHDs, and provides comparisons between these two designs. It turns out that LHDs have a smaller expectation and variance compared to that of SRDs. The expectation and variance of LHDs can provide an upper and lower bound of the square WD for searching UDs. For high-dimensional UDs, the threshold accepting heuristic is applied in order to find low discrepancy designs. A conjecture of \textit{K. T. Fang}, \textit{D. K. J. Lin}, \textit{P. Winker} and \textit{Y. Zhang} [Technometrics 42, No. 3, 237-248 (2000; Zbl 0996.62073)] considers that a UD under a proper measure of uniformity is also an orthogonal design if its corresponding orthogonal design exists. The paper shows that this conjecture is true for complete designs, if WD is chosen as the measure of uniformity.