Successive Umbrella Sampling

Abstract

We propose an extension of umbrella sampling in which the pertinent range of states is subdivided in windows that are sampled consecutively and linked together. Extrapolating results from one window we estimate a weight function for the neighboring window. We present a detailed analysis and demonstrate that the error is controlled and independent from the window sizes. The analysis also allows us to detect sampling difficulties. The efficiency of the algorithm is comparable to a multicanonical simulation with an ideal weight function. We exemplify the computational scheme by simulating the liquid-vapor coexistence in a Lennard--Jones system.