Strong error analysis and first-order convergence of Milstein-type schemes for McKean-Vlasov SDEs with superlinear coefficients

Jingtao Zhu, Yuying Zhao, Siqing Gan · 2025

Read Strong error analysis and first-order convergence of Milstein-type schemes for McKean-Vlasov SDEs with superlinear coefficients by Jingtao Zhu, Yuying Zhao, Siqing Gan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-16801.txt.

Introduction

Jingtao Zhu, Yuying Zhao, and Siqing Gan strengthen simulation math for interacting particles in a 2025 paper. Their equations are McKean-Vlasov equations, where each particle feels the whole crowd. Simulating them needs step-by-step numerical schemes. Standard schemes can break when growth terms rise faster than linearly.

The team builds one family of Milstein-type schemes covering those cases. Named variants tame the wild terms in different gentle ways. They prove the schemes still reach first-order accuracy overall. The proof needs only mild smoothness from the equation's coefficients. Earlier work had tested one such scheme without proof. Numerical experiments in the paper back the theory. The results give simulators a safer toolkit for crowded particle systems.

Catalog record

CREATORJingtao Zhu, Yuying Zhao, Siqing Gan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE124 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16801

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Strong error analysis and first-order convergence of Milstein-type schemes for McKean-Vlasov SDEs with superlinear coefficients Jingtao Zhu, Yuying Zhao, Siqing Gan Abstract In the study of McKean-Vlasov stochastic differential equations (MV-SDEs), numerical approximation plays a crucial role in understanding the behavior of interacting particle systems (IPS). Classical Milstein schemes provide strong convergence of order one under globally Lipschitz coefficients. Nevertheless, many MV-SDEs arising from applications possess super-linearly growing drift and diffusion terms, where classical methods may diverge and particle corruption can occur. In the present work, we aim to fill this gap by developing a unified class of Milstein-type discretizations handling both super-linear drift and diffusion coefficients. The proposed framework includes the tamed-, tanh-, and sine-Milstein methods as special cases and establishes order-one strong convergence for the associated interacting particle system under mild regularity assumptions, requiring only once differentiable coefficients. In particular, our results complement Chen et al. (Electron. J. Probab., 2025), where a taming-based Euler scheme was only tested numerically without theoretical guarantees, by providing a rigorous convergence theory within a broader Milstein-type framework. The analysis relies on discrete-time arguments and binomial-type expansions, avoiding the continuous-time Itô approach that is standard in the literature. Numerical experiments are presented to illustrate the convergence behavior and support the theoretical findings. 2020 MSC 65C30 , 60H35 1 Introduction As a class of important mathematical models, McKean-Vlasov stochastic differential equations (MV-SDEs), also known as distribution-dependent SDEs, play a central role in modeling interacting particle systems (IPS) and have been extensively studied in fields such as statistical physics Sznitman 1991 , Braun and Hepp 1977 , mathematical finance and mean-field games Lasry and Lions 2007 , Carmona and Delarue 2018 , control theory Bensoussan et al. 2013 , and neuroscience Baladron et al. 2012 . In general, closed-form solutions to MV-SDEs are not available and the development and analysis of numerical methods for their simulation are of significant practical interest. In this paper, we concern the design and convergence analysis of numerical approximations for the following MV-SDEs: \text{d}X_{t}=f(X_{t},\mathscr{L}_{X_{t}})\,\text{d}t+g(X_{t},\mathscr{L}_{X_{t}})\,\text{d}W_{t},\quad t\in[0,\mathcal{T}], (1.1) where \mathscr{L}_{X_{t}} denotes the law of X_{t} , f\colon\mathbb{M}\to\mathbb{R}^{d} and g\colon\mathbb{M}\to\mathbb{R}^{d\times m} , with \mathbb{M}:=\mathbb{R}^{d}\times\mathscr{P}_{2}(\mathbb{R}^{d}) as defined in Section 2.1 . The process \{W_{t}\}_{0\leq t\leq\mathcal{T}} is an m -dimensional standard Brownian motion on a complete probability space (\Omega,\mathcal{F},\{\mathcal{F}_{t}\}_{0\leq t\leq\mathcal{T}},\mathbb{P}) with a filtration \{\mathcal{F}_{t}\}_{t\in[0,\mathcal{T}]} . As the law of X_{t} is known, the underlying MV-SDEs reduce to the usual SDEs, numerical methods for which have been extensively studied for the past decades (consult monographs Milstein and Tretyakov 2021 , Hutzenthaler and Jentzen 2015 , Wang et al. 2020 , Li et al. 2018 and references therein), in the context of both numerical convergence and numerical stability. When the law of X_{t} is unknown, a widely used approach is to approximate the solution via an IPS and then simulate the particle system using time discretization schemes. Under suitable assumptions, the principle of propagation of chaos (PoC) guarantees that the empirical distribution of the particle system converges to the law of the solution as the number of particles tends to infinity Sznitman 1991 , the time discretization of the IPS becomes a powerful tool to understand the behavior of the underlying problems. To analyze time discretization approximations of the IPS, a global Lipschitz condition is often imposed on the coefficient functions of MV-SDEs Li et al. 2022 , Bao and Huang 2022 . Nevertheless, MV-SDEs arising from applications rarely obey such a traditional but restrictive condition. Notable examples of MV-SDEs with non-globally Lipschitz continuous coefficients include numerous models, such as the double-well dynamics (see ( 4.1 )), the Cucker-Smale flocking model (see ( 4.2 ) or Erban et al. 2016 , Chen and dos Reis 2022 ), and the FitzHugh-Nagumo model (see Example 4.3 or Baladron et al. 2012 ). Evidently, the coefficients of these three models violate the globally Lipschitz condition. It is well established that even for SDEs Hutzenthaler et al. 2011 , Mattingly et al. 2002 with superlinearly growing coefficients, classical schemes such as Euler-Maruyama and Milstein methods may diverge. In the case of MV-SDEs, analogous divergence phenomena commonly referred to as “particle corruption" have also been observed and analyzed dos Reis et al. 2021 . Specifically, a single particle may become excessively influential over all others, thereby destroying the weakly interacting particle structure underlying the MV-SDE system. To address this, a growing body of work has focused on the construction and analysis of convergent Euler-type schemes in non-globally Lipschitz settings (see dos Reis et al. 2021 , Liu et al. 2023a , Liu et al. 2023b , Chen and dos Reis 2022 , Chen and dos Reis 2023 , Kumar et al. 2022 , Cui et al. 2025 , Guo et al. 2024 , Chen et al. 2025 , Jian et al. 2025 , Reisinger and Stockinger 2022 , Neelima et al. 2020 , Khue Tran et al. 2025 , Biswas et al. 2025 , Zhu et al. 2025 and references therein). Nevertheless, Euler-type schemes for MV-SDEs with multiplicative noise inherently achieve at most a strong convergence rate of order \frac{1}{2} , which often results in high computational costs when higher accuracy is required. In contrast, Milstein-type schemes offer order-one strong convergence and can be efficiently combined with multilevel Monte Carlo techniques to reduce computational complexity further Giles 2006 , Giles 2008 . Recently, tamed Milstein methods have been successfully applied to MV-SDEs with superlinear drift and Lipschitz diffusion coefficients Bao et al. 2021 , Kumar and Neelima 2021 , Bao et al. 2023 . A natural yet unresolved question thus arises as to whether such methods can be extended to settings where both the drift and diffusion coefficients exhibit superlinear growth. To fill this gap, we develop a general framework for a family of Milstein-type schemes designed to handle superlinear drift and diffusion coefficients and establish their strong convergence theory. Compared with Bao et al. 2021 , our analysis only requires once differentiability of the coefficients in the state and measure components. Specifically, our scheme modifies the drift and diffusion coefficients through bounded operators that scale with a negative power of the time step size \Delta t , leading to the following general form: \displaystyle Y_{t_{k+1}}^{i,N}= \displaystyle~Y_{t_{k}}^{i,N}+\Gamma_{1}\!\left(f\!\left(Y_{t_{k}}^{i,N},\rho_{t_{k}}^{Y,N}\right),\Delta t\right)\Delta t+\sum_{j_{1}=1}^{m}\Gamma_{2}\!\left(g_{j_{1}}\!\left(Y_{t_{k}}^{i,N},\rho_{t_{k}}^{Y,N}\right),\Delta t\right)\Delta W^{i,j_{1}}_{t_{k}} \displaystyle~+\sum_{j_{1},j_{2}=1}^{m}\Gamma_{3}\!\left(\mathcal{L}_{y}^{j_{2}}~g_{j_{1}}\!\left(Y_{t_{k}}^{i,N},\rho_{t_{k}}^{Y,N}\right),\Delta t\right)\!\int_{t_{k}}^{t_{k+1}}\!\!\int_{t_{k}}^{s}\mathrm{d}W_{r}^{i,j_{2}}\mathrm{d}W^{i,j_{1}}_{s} (1.2) \displaystyle~+\frac{1}{N}\sum_{j_{1},j_{2}=1}^{m}\sum_{k_{1}=1}^{N}\Gamma_{4}\!\left(\mathcal{L}_{\rho}^{j_{2}}~g_{j_{1}}\!\left(Y_{t_{k}}^{i,N},\rho_{t_{k}}^{Y,N},Y_{t_{k}}^{k_{1},N}\right),\Delta t\right)\!\int_{t_{k}}^{t_{k+1}}\!\int_{t_{k}}^{s}\mathrm{d}W_{r}^{k_{1},j_{2}}\mathrm{d}W^{i,j_{1}}_{s}. Here, the operators \Gamma_{i}(\cdot,\Delta t) , i=1,\dots,4 , are uniformly bounded and designed to control the superlinear growth of the coefficients through their dependence on \Delta t . Moreover, this framework encompasses several schemes as special cases, including the tamed, sine-, and tanh-type Milstein methods (see Examples ( 3.14 )–( 3.16 )) and also recovers the taming-in Euler scheme in Chen et al. 2025 as a degenerate instance, where the scheme was only tested numerically without theoretical guarantees. In particular, the tamed Milstein scheme considered in Kumar et al. 2022 for MV-SDEs with common noise under similar growth conditions is recovered as a specific example within our framework, corresponding to special choices of the operators \Gamma_{i} . Indeed, when treating MV-SDEs with superlinearly growing coefficients by taming strategy, it becomes a standard way in the literature to work with continuous-time extensions of the schemes and carry out the analysis based on the Itô formula (see, e.g. Neelima et al. 2020 , Kumar et al. 2022 ,

Read the complete text