Convergence and Refinement of the Wang-Landau Algorithm

TL;DR

This paper analyzes the convergence behavior of the Wang-Landau algorithm, revealing how fluctuations stabilize and how tuning parameters influence the rate of convergence, thereby deepening understanding of its efficiency.

Contribution

It provides an error analysis of the Wang-Landau algorithm, identifying the factors that determine its convergence rate and the role of tuning parameters.

Findings

Fluctuations in the energy histogram saturate at a level proportional to [log(f)]^{-1/2}

Steady state is reached where error corrections are closely related to this fluctuation level

Different tuning parameters significantly affect the convergence rate

Abstract

Recently, Wang and Landau proposed a new random walk algorithm that can be very efficiently applied to many problems. Subsequently, there has been numerous studies on the algorithm itself and many proposals for improvements were put forward. However, fundamental questions such as what determines the rate of convergence has not been answered. To understand the mechanism behind the Wang-Landau method, we did an error analysis and found that a steady state is reached where the fluctuations in the accumulated energy histogram saturate at values proportional to $[\log(f)]^{-1/2}$. This value is closely related to the error corrections to the Wang-Landau method. We also study the rate of convergence using different "tuning" parameters in the algorithm.