Tuesday, November 14, 2017 at 11:40am to 1:10pm
Uris Hall, 498
University of Pennsylvania
Abstract: The paper describes a new approximate nonlinear filtering technique. Strengths of the technique include: (1) it can be used as long as one can simulate from the model, without the need to evaluate transition densities, (2) it can be used with some infinite-dimensional state variables, and (3) computational speed thanks to the embarrassingly parallel nature of its computationally intensive part. The main theoretical result of the paper is that the approximation error of the technique goes to zero with computational power, and that it does so uniformly with respect to the time horizon of the data.