This talk presents an explicit, structure-preserving probablistic numerical integrator for Langevin systems, and an efficient, structure-preserving integrator for Langevin systems with holonomic constraints. In a nut-shell, the method does for the Boltzmann-Gibbs measure what symplectic integrators do for energy. More precisely, we prove the method very nearly preserves the Boltzmann-Gibbs measure. As a consequence of its variational design, the algorithm also exactly preserves the symplectic area change associated to Langevin processes. The method with supporting theory enables one to take time-step sizes and friction factors at the limit of stability of the integration scheme (e.g., the time-step size must be smaller than the fastest characteristic frequency in the system).
Related Links
* http://www.acm.caltech.edu/~nawaf - applicant's webpage
* http://www.acm.caltech.edu/~owhadi - postdoc advisor's webpage
view more