Analysis of Variational Integrators for Langevin Processes
Nawaf Bou-Rabee
Free University of Berlin
This talk analyzes generalizations of variational integrators from
simple mechanical systems to Langevin processes that are relevant in,
e.g., computations of molecular dynamics trajectories. An introduction
to variational integrators and discrete mechanics is provided. Special
attention is paid towards operating these integrators in multiscale
systems at the verge of linear stability where autonomous backward
error analysis often breaks down.
Against this backdrop stochastic variational integrators are presented
based on Lie-Trotter and symmetric Strang splittings of the
Focker-Planck operator. These discretizations of Langevin processes are
quite natural, but seem to have only recently been proposed in the
literature by Vanden-Eijnden & Ciccotti [2006] and Bussi &
Parrinello [2007]. An analysis of these integrators reveals the
necessity of rejection sampling for nonlinear stability. Geometric
ergodicity of the resulting Metropolis-stabilized, discrete Markov
chain is demonstrated. Numerics computing equilibrium statistics and
autocorrelation functions in conformation changes of certain alkanes is
presented.
This work is joint with Christof Schutte (FUB) and Eric Vanden-Eijnden
(NYU).
1