Variance reduction using antithetic variables for a nonlinear convex stochastic homogenization problem (Q479065)

When treating the homogenization of the random variational problem \[ \inf\Big\{ \int_{D} W(\frac{x}{\epsilon} , \omega , \nabla u(x)) dx - \int_{D} f(x) u(x) dx, u \in W_{0}^{1,p}(D)\Big\} \] where \(D \subset {\mathbb R}^{d}\) is an open bounded domain, \(2 < p,\infty\) and \(W\) is a stationary random smooth field, displaying convexity properties in \(\nabla u\), it can be shown that as \(\epsilon \to 0\) this problem tends to a deterministic homogenized problem of the form \(\inf\{ \int_{D} W^{*}( \nabla u(x)) dx - \int_{D} f(x) u(x) dx, u \in W_{0}^{1,p}(D)\}\), however this requires in general the determination of a properly defined corrector function which has to be determined as the solution of a nonlinear problem (the corrector problem) typically defined on the entire space \({\mathbb R}^{d}\). A common approach for the calculation of the homogenized limit is to generate realizations of the energy density \(W\) over a finite (but large volume) \(Q_{N}\) and then approach the homogenized energy density by some empirical means using approximate correctors computed on \(Q_{N}\). This paper considers the problem of efficiently averaging over these realizations requires to understand how variance affects the result. In particular, it is shown that the technique of antithetic variables can be used to reduce the variance of the computed quantities, thus decreasing the computational cost at equal accuracy and providing an efficient approach for approximating expectations of the apparent homogenized energy density and of related quantities.