Some large-sample results on a modified Monte Carlo integration method (Q2573522): Difference between revisions

In \textit{S. Haber}'s locally antithetic variates integration method [Math. Comput. 21, 388--397 (1967; Zbl 0298.65022)] the integral \(\int_{[0,1]^d}f(u)du\) is approximated by \[ \tilde I_n={1\over 2n}\sum_{j=1}^n[f(V_j)+f(2a_j-V_j)], \] where the sampling points \(V_j\) are derived as follows: \([0,1]^d\) is partitioned by uniform grid on small hypercubes \(K_j\), \(j=1,\dots,m^d=n\) with centers \(a_j\) and \(V_j\) is sampled randomly from uniform distribution on \(K_j\). (Note that \(2a_j-V_j\) is a point symmetric to \(V_j\) w.r.t. \(a_j\)). Haber has shown asymptotic normality of \((\tilde I_n-I)\) with asymptotic variance \(\sigma^2_n\sim n^{-(1+4/d)}\sigma^2\) if \(f\in C^2([0,1]^d)\). The authors demonstrate this under the milder assumption that \(\nabla f\) is a Lipschitz function. An estimator for \(\sigma^2\) is proposed and its asymptotic properties are investigated. These results are applied to environmental sampling design.