On the non-convergence of the Wang-Landau algorithms with multiple random walkers

TL;DR

This paper investigates the convergence behavior of Wang-Landau and $1/t$ algorithms with multiple random walkers, revealing that increasing the number of walkers beyond a critical point does not improve accuracy and that the $1/t$ algorithm is generally more efficient.

Contribution

It demonstrates that adding more walkers beyond a critical number does not enhance convergence, challenging assumptions about parallelization in entropic sampling methods.

Findings

Error scales as 1/√m with the number of walkers

Error saturates beyond a critical number of walkers

The $1/t$ algorithm outperforms Wang-Landau in accuracy and efficiency

Abstract

This paper discusses some convergence properties in the entropic sampling Monte Carlo methods with multiple random walkers, particularly in the Wang-Landau (WL) and $1/t$ algorithms. The classical algorithms are modified by the use of $m$ independent random walkers in the energy landscape to calculate the density of states (DOS). The Ising model is used to show the convergence properties in the calculation of the DOS, as well as the critical temperature, while the calculation of the number $\pi$ by multiple dimensional integration is used in the continuum approximation. In each case, the error is obtained separately for each walker at a fixed time, $t$; then, the average over $m$ walkers is performed. It is observed that the error goes as $1/\sqrt{m}$. However, if the number of walkers increases above a certain critical value $m>m_x$, the error reaches a constant value (i.e. it saturates). This occurs for both algorithms; however, it is shown that for a given system, the $1/t$ algorithm is more efficient and accurate than the similar version of the WL algorithm. It follows that it makes no sense to increase the number of walkers above a critical value $m_x$, since it does not reduces the error in the calculation. Therefore, the number of walkers does not guarantee convergence.