Temporal coarse-graining and elimination of slow dynamics with the generalized Langevin equation for time-filtered observables

Abstract: By exact projection in phase space we derive the generalized Langevin equation (GLE) for time-filtered observables. We employ a general convolution filter that directly acts on arbitrary phase-space observables and can involve low-pass, high-pass, band-pass or band-stop components. The derived filter GLE has the same form and properties as the ordinary GLE but exhibits modified potential, mass and memory friction kernel. Our filter-projection approach has diverse applications and can be used to i) systematically derive temporally coarse-grained models by low-pass filtering, ii) undo data smoothening inherent in any experimental measurement process, iii) decompose data exactly into slow and fast variables that can be analyzed separately and each obey Liouville dynamics. The latter application is suitable for removing slow transient or seasonal (i.e. periodic) components that do not equilibrate over simulation or experimental observation time scales and constitutes an alternative to non-equilibrium approaches. We derive integral formulas for the GLE parameters of filtered data for general systems. For the special case of a Markovian system we derive the filter GLE memory kernel in closed form and show that low-band pass smoothening of data induces exponentially decaying memory.