[Paper Review] On the Asymptotic Normality of Adaptive Multilevel Splitting

Frédéric Cérou, Bernard Delyon+2 more|arXiv (Cornell University)|Apr 23, 2018

Stochastic processes and statistical mechanics14 references26 citations

TL;DR

This paper establishes the first rigorous asymptotic normality and consistency results for the Adaptive Multilevel Splitting (AMS) algorithm in a general Markov process setting. By reformulating AMS as a Fleming-Viot particle system via a level-indexed process (stochastic wave), the authors leverage recent theoretical results on Fleming-Viot systems to prove a Central Limit Theorem (CLT) in the large particle limit (N → ∞), providing theoretical justification for the algorithm's practical efficiency in rare event simulation and conditional sampling.

ABSTRACT

Adaptive Multilevel Splitting (AMS for short) is a generic Monte Carlo method for Markov processes that simulates rare events and estimates associated probabilities. Despite its practical efficiency, there are almost no theoretical results on the convergence of this algorithm. The purpose of this paper is to prove both consistency and asymptotic normality results in a general setting. This is done by associating to the original Markov process a level-indexed process, also called a stochastic wave, and by showing that AMS can then be seen as a Fleming-Viot type particle system. This being done, we can finally apply general results on Fleming-Viot particle systems that we have recently obtained.

Motivation & Objective

• To establish rigorous theoretical convergence guarantees for the Adaptive Multilevel Splitting (AMS) algorithm, which is widely used but lacks theoretical foundation.
• To prove consistency and asymptotic normality of AMS estimators in the large particle limit (N → ∞), not just in idealized cases.
• To bridge the gap between practical implementation and theoretical analysis by connecting AMS to the well-studied class of Fleming-Viot particle systems.
• To provide a general framework applicable to diffusions and other processes, with explicit conditions under which the CLT holds.

Proposed method

• Introduce a level-indexed process (stochastic wave) that reparameterizes the original Markov process Y using the level function ξ as a time-like parameter.
• Reformulate the AMS algorithm as a Fleming-Viot type particle system on this level-indexed process, preserving the dynamics of particle resampling and killing.
• Apply a recently established Central Limit Theorem (CLT) for Fleming-Viot particle systems to the reformulated AMS algorithm.
• Establish the necessary assumptions (Assumptions 1–3) on the underlying process (Y, ξ), including non-degeneracy of the gradient of ξ with respect to the diffusion coefficient.
• Use pathwise continuity and convergence arguments, including Skorokhod embedding and strong Markov property, to verify the required regularity conditions.
• Extend the results to path observables and entrance times by augmenting the state space to include trajectory information.

Experimental results

Research questions

• RQ1 Under what general conditions does the Adaptive Multilevel Splitting algorithm produce consistent and asymptotically normal estimators?
• RQ2 Can the AMS algorithm be rigorously linked to the theory of Fleming-Viot particle systems?
• RQ3 What are the sufficient conditions on the Markov process (Y, ξ) for the CLT to hold in the large N limit?
• RQ4 How can the algorithm be extended to estimate path-dependent functionals, such as first passage times or trajectory distributions?
• RQ5 Does the theoretical CLT match the practical performance of AMS in rare event simulation, particularly for diffusions with non-degenerate reaction coordinates?

Key findings

• The paper proves a Central Limit Theorem (CLT) for the Adaptive Multilevel Splitting algorithm in the large particle limit (N → ∞), establishing asymptotic normality of the estimators.
• The key technical insight is that AMS can be reformulated as a Fleming-Viot particle system via a level-indexed process, enabling the application of existing CLT results.
• The CLT holds under three main assumptions: (1) the process (Y, ξ) is a Feller process, (2) the level function ξ is smooth with compact level sets, and (3) the non-degeneracy condition (∇ξ)ᵀσ ≠ 0 holds almost surely.
• The asymptotic variance of the estimator is explicitly characterized through the CLT, providing a theoretical basis for variance reduction strategies.
• The results are extended to path observables by augmenting the state space of the level-indexed process to include trajectory information, and the CLT is shown to hold in this pathwise setting as well.
• The paper provides a practical variant of Assumption 3 (Assumption 3’) that is easier to verify in applications, particularly in chemical and molecular dynamics simulations.