[Paper Review] New particle representations for ergodic McKean-Vlasov SDEs

Houssam Alrachid, Mireille Bossy+2 more|arXiv (Cornell University)|Jan 16, 2019

Markov Chains and Monte Carlo Methods23 references4 citations

TL;DR

This paper introduces novel particle-based algorithms for efficiently approximating the invariant measure of ergodic McKean-Vlasov SDEs by leveraging self-interacting diffusions and ensemble methods. By exploiting weak convergence analysis and variance reduction techniques, the authors achieve a computational cost of 𝒪(𝜖⁻⁴ log 𝜖⁻¹), improving upon the standard 𝒪(𝜖⁻⁶) cost of naive particle estimators for a given mean-square error tolerance 𝜖.

ABSTRACT

The aim of this paper is to introduce several new particle representations for extit{ergodic} McKean-Vlasov SDEs. We construct new algorithms by leveraging recent progress in weak convergence analysis of interacting particle system. We present detailed analysis of errors and associated costs of various estimators, highlighting key differences between long-time simulations of linear (classical SDEs) versus non-linear (Mckean-Vlasov SDEs) process.

Motivation & Objective

• To address the high computational cost of simulating interacting particle systems for McKean-Vlasov SDEs, especially over long-time horizons.
• To overcome the bias and dependence issues arising from measure approximation in particle systems, which hinder classical variance reduction.
• To develop new particle representations that enable efficient ergodic averaging for non-linear SDEs with invariant measures.
• To analyze the trade-offs between statistical error, time discretization, and particle count in long-time simulations.
• To achieve computational cost reductions compared to standard particle estimators for a given accuracy target 𝜖.

Proposed method

• Propose a new class of particle representations based on self-interacting diffusions, where each particle's dynamics depend on its own empirical measure over time.
• Introduce an ensemble-averaged estimator (ES-AEA) that combines time-averaged particle trajectories with multiple independent realizations to reduce variance.
• Use a time-discretized Euler scheme with step size 1/n to simulate the self-interacting dynamics, ensuring weak convergence to the invariant measure.
• Apply a multilevel Monte Carlo-inspired strategy by balancing time horizon t, particle count N_t, and ensemble size M to minimize total cost.
• Leverage propagation of chaos and exponential ergodicity to bound the bias and statistical error in the invariant measure approximation.
• Analyze the mean-square error decomposition into bias, discretization, and Monte Carlo variance components to derive optimal parameter choices.

Experimental results

Research questions

• RQ1 Can self-interacting diffusions provide a more efficient alternative to standard interacting particle systems for approximating the invariant measure of McKean-Vlasov SDEs?
• RQ2 What is the optimal trade-off between time horizon, number of particles, and ensemble size to minimize computational cost for a given mean-square error tolerance?
• RQ3 How does the variance of the estimator scale with the number of particles and ensemble size in the context of non-linear, measure-dependent SDEs?
• RQ4 Can ensemble-based implementations of self-interacting diffusions achieve better computational complexity than single-trajectory estimators?
• RQ5 What are the theoretical error bounds and convergence rates for the proposed particle representations in the long-time and large-N limits?

Key findings

• The proposed algorithm ES-AEA achieves a mean-square error of order 𝒪(𝜖) with computational cost 𝒪(𝜖⁻⁴ log 𝜖⁻¹), representing a significant improvement over the standard 𝒪(𝜖⁻⁶) cost of naive particle estimators.
• By setting the number of particles N_t = 1 and using M ≈ 𝜖⁻² log⁻¹(𝜖⁻¹) independent ensembles, the method achieves optimal cost scaling while maintaining exponential bias decay.
• The ensemble-averaged estimator (CS-AEA) reduces variance through M independent realizations, leading to a cost of 𝒪(𝜖⁻⁴ log 𝜖⁻¹), which is one order better than the naive estimator.
• The analysis shows that the dominant error terms are the bias (exponentially decaying in time) and the Monte Carlo variance, which can be controlled via ensemble size M.
• The method achieves optimal parameter scaling: t ≈ 𝜆⁻¹ log(𝜖⁻¹), n ≈ 𝜖⁻¹, and N ≈ 𝜖⁻² / log(𝜖⁻¹), balancing all error components.
• The results suggest that self-interacting diffusions with ensemble averaging can bring the cost of simulating non-linear particle systems close to that of independent SDE simulations.