A combinatorial central limit theorem for randomized orthogonal array sampling designs (Q1816980)

Let \(X\) be a random vector uniformly distributed on \([0,1]^3\) and \(f\) be a measurable function from \([0,1]^3\) to the real line. An objective of many computer experiments is to estimate \(\mu= E(f \circ x)\) by computing \(f\) at a fixed number of points. \textit{A. B. Owen} [ibid. 22, No. 2, 930-945 (1994; Zbl 0807.62059); Stat. Sinica 2, No. 2, 439-452 (1992; Zbl 0822.62064)] and \textit{B. Tang} [J. Am. Stat. Assoc. 88, No. 424, 1392-1397 (1993; Zbl 0792.62066)] independently suggested the use of randomized orthogonal arrays in sampling designs for computer experiments. The main attraction of these designs is that they, in contrast to simple random sampling, stratify on all \(t\)-variate margins simultaneously. \textit{C. M. Stein}'s method [see ``Approximate computation of expectations.'' (1986; Zbl 0721.60016)] is used by the author to prove asymptotic normality of \(\widehat\mu\) which is the usual average of the orthogonal array-based sample of size \(q^2\) of \(f\) values. In fact, he shows that under the finiteness of \(r\)-th moments the corresponding error bound is of the order \(O(q^{-(r-2)/(2r-2)})\), where \(r\) is an even integer greater than or equal to 4.