Implicit discretization schemes for langevin dynamics

We explore here several numerical schemes for Langevin dynamics in the general implicit discretization framework of the Langevin/implicit-EuIer scheme, LI. Specifically, six schemes are constructed through different discretization combinations of acceleration, velocity, and position. Among them, the explicit BBK method (LE in our notation) and LI are recovered, and the other four (all implicit) are named LIM1, LIM2, MIDI, and MID2. The last two correspond, respectively, to the well-known implicit-midpoint scheme and the trapezoidal rule. LI and LIM1 are first-order accurate and have intrinsic numerical damping. LIM2, MIDI, and MID2 appear to have large-timestep stability as LI but overcome numerical damping. However, numerical results reveal limitations on other grounds. From simulations on a model butane, we find that the nondamping methods give similar results when the timestep is small; however, as the timestep increases, LIM2 exhibits a pronounced rise in the potential energy and produces wider distributions for the bond lengths. MIDI and MID2 appear to be the best among those implicit schemes for Langevin dynamics in terms of reasonably reproducing distributions for bond lengths, bond angles and dihedral angles (in comparison to 1 fs timestep explicit simulations), as well as conserving the total energy reasonably. However, the minimization subproblem (due to the implicit formulation) becomes difficult when the timestep increases further. In terms of computational time, all the implicit schemes are very demanding. Nonetheless, we observe that for moderate timesteps, even when the error is large for the fast motions, it is relatively small for the slow motions. This suggests that it is possible by large timestep algorithms to capture the slow motions without resolving accurately the fast motions.