Scrambling Sobol' and Niederreiter-Xing points (Q1279911): Difference between revisions

Integration of a function \(f\) over the unit cube of a large dimension \(s\) by a hybrid of Monte Carlo and equidistribution methods is considered. The hybrid technique is based on scrambling the digits in a \((t,m,s)\)-net or \((t,s)\)-sequence in base \(b\). The variance of scrambled net integration over nets with \(t>0\) such as those of I. M. Sobol' and \textit{H. Niederreiter} and \textit{C. Xing} [Finite Fields Appl. 2, No. 3, 241-273 (1996; Zbl 0893.11029)] is studied. This variance under scrambled \((t,m,s)\)-net sampling in base \(b\) is never more than \(b^t[(b+ 1)/(b- 1)]^s\) times as large as the Monte Carlo variance. Scrambled nets with \(t>0\) achieve a variance that is \(O(n^{-3}(\log n)^{s-1})\) for smooth integrands. Numerical results are shown.