We present a general method to obtain parametrised models for the drift and diffusion terms of the Fokker-Planck equation of a coarse-grained description of molecular systems. The method is based on the minimisation of the relative entropy defined in terms of the two-time joint probability and thus captures the full dynamics of the coarse-grained description. In addition, we show an alternative Bayesian argument that starts from the path probability of a diffusion process which allows one to obtain the best parametrised model that fits an actual observed path of the coarse-grained variables. Both approaches lead to exactly the same optimisation function giving strong support to the methodology. We provide an heuristic argument that explains how both approaches are connected.
This journal is © the Owner Societies 2011