Breakdown of the Finite-Time and -Population Scalings of the Large Deviation Function in the Large-Size Limit of a Contact Process

TL;DR

This study extends the analysis of large deviation function estimator scalings in contact processes to larger system sizes, revealing that traditional $t^{-1}$ and $N_c^{-1}$ scalings break down and can even become unreliable with increasing system size.

Contribution

The paper introduces generalized exponents to characterize LDF estimator behavior across various system sizes, highlighting limitations of standard scalings in large systems.

Findings

Standard scalings fail for large system sizes.

Negative exponents indicate unreliable estimations.

Scaling laws depend on system size and can break down.

Abstract

In a recent study, the finite-time ($t$) and -population size ($N_c$) scalings in the evaluation of a large deviation function (LDF) estimator were analyzed by means of the cloning algorithm. These scalings provide valuable information about the convergence of the LDF estimator in the infinite-$t$ and infinite-$N_c$ limits. For the cases analyzed in that study, the scalings of the systematic errors of the estimator were found to behave as $t^{-1}$ and $N_c^{-1}$ in the large-$t$ and large-$N_c$ asymptotics. Moreover, it was shown how this convergence speed can be used in order to extract an asymptotic limit which resulted to render a better LDF estimation in comparison to the standard estimator. However, the validity of these scaling laws and thus, the convergence of the estimator was proved only in systems for which the number of sites $L$ (where the dynamics occurs) was small. In this paper, the analysis is extended to a wider range of system sizes $L$. We show how the introduction of the exponents $\gamma_{t}$ and $\gamma_{N_c}$ allows to characterize the behavior of the LDF estimator for any system size. From these generalized $t^{-\gamma_{t}}$- and $N_{c}^{-\gamma_{N_c}}$- scalings, we verify that in the large-$L$ limit the $t^{-1}$- and $N_c^{-1}$-scalings are no longer valid. Moreover, as the convergence of the estimator relies on the positivity of these exponents, we show how for some cases $\gamma_{N_c}$ can be negative implying that the estimation provided by the cloning algorithm is no longer reliable.