Variance reduction for the equation-free simulation of multiscale stochastic systems

We study the problem of simulating the slow observable of a multiscale diffusion process. In particular, we extend previous algorithms to the case where the simulation of the different scales cannot be uncoupled and we have no explicit knowledge of the drift or the variance of the multiscale diffusion. This is the case when the simulation data come from a black box "legacy code," or possibly from a fine scale simulator (e.g., MD, kMC) which we want to effectively model as a diffusion process. We improve the algorithm, using the past simulations as control variates, in order to reduce the variance of the subsequent simulations.