🔢 mathematics

Adaptive Time-Stepping Euler--Maruyama Scheme for SDEs with Non-Globally Lipschitz Coefficients: Uniform Convergence, Stability and Ergodicity

This paper proposes an adaptive time-stepping Euler–Maruyama scheme for stochastic differential equations with non-globally Lipschitz coefficients that ensures moment boundedness, achieves a uniform $1/2$ -order strong convergence rate, preserves exponential stability, and guarantees polynomial ergodicity with optimal convergence of the invariant measure.

Original authors: Xueqi Wen, Shan Huang, Xiaoyue Li

Published 2026-08-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xueqi Wen, Shan Huang, Xiaoyue Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the path of a leaf floating down a turbulent river. The water isn't just flowing; it's churning, swirling, and occasionally hitting a rock that sends the leaf spinning wildly. In the world of science, this is how we model everything from the spread of diseases to the movement of stock markets and the behavior of atoms. We use mathematical recipes called "Stochastic Differential Equations" (SDEs) to describe these journeys. The "stochastic" part just means there's a dash of randomness thrown in, like the unpredictable gusts of wind or the random bumps in the water.

To solve these equations on a computer, we can't watch the leaf move in a smooth, continuous line. Instead, we have to take tiny snapshots, or "steps," to see where it is at specific moments. This is called "time-stepping." For a long time, scientists used a fixed-size step, like taking a stride of exactly one foot no matter what. But if the river suddenly turns into a waterfall (a mathematical "explosion"), a fixed step might skip right over the danger, sending your simulation flying off the map into nonsense. This paper tackles the tricky problem of how to take these steps when the river gets too wild for standard rules, ensuring our digital leaf stays on the water and actually tells us something true about the real world.

The Paper's Story: A Smart, Adaptive Hiker

The authors of this paper, Xueqi Wen, Shan Huang, and Xiaoyue Li, are like a team of hikers trying to cross a mountain range where the terrain changes unpredictably. Sometimes the path is flat and easy; other times, it's a steep, jagged cliff that could send you tumbling if you aren't careful. In the language of math, this "terrain" is the behavior of the equations, and the "cliffs" are areas where the rules of smoothness break down (what they call "non-globally Lipschitz coefficients").

The Problem with the Old Way
For years, hikers (mathematicians) used a "fixed-step" method. They decided, "I will take a step of size 0.1 every time." On flat ground, this works fine. But if the ground suddenly turns into a vertical wall, a fixed step of 0.1 might be too big to climb, causing the hiker to fall off the edge of the simulation. The computer crashes, or the numbers blow up to infinity.

The New Solution: The Adaptive Hiker
This paper introduces a new kind of hiker: the Adaptive Time-Stepping Euler–Maruyama (AEM) scheme. Imagine a hiker who carries a smart watch. This watch doesn't just count steps; it looks at the ground ahead.

If the path is smooth and gentle, the watch says, "Great! Take a big, confident step."
If the path gets steep or rocky, the watch screams, "Danger! Slow down! Take a tiny, cautious step!"

By dynamically adjusting the size of the step based on how wild the equation is at that exact moment, this new method prevents the simulation from "exploding." It keeps the numbers stable, even when the math gets messy.

What They Proved (The "Receipts")
The authors didn't just build this smart hiker; they put it through rigorous tests to prove it works. Here is what they found:

It Never Gets Lost: They proved that no matter how long you walk (even if you walk forever), this hiker will eventually reach any destination you set. It won't get stuck in an infinite loop of tiny steps.
It's Accurate: They showed that the path the hiker takes is very close to the "true" path of the leaf in the river. Specifically, they proved the error shrinks at a rate of 1/2. In math-speak, this is a "strong convergence rate." It means if you make your steps smaller, the answer gets significantly better, and they proved this holds true even over very long periods of time.
It Keeps Its Cool: One of the biggest worries with these wild equations is that the numbers might grow out of control (become unstable). The authors proved that their adaptive method keeps the numbers bounded, just like the real system does. If the real leaf settles down, the digital leaf settles down too.
It Finds the Right "Home": Many of these equations describe systems that eventually settle into a specific pattern or "invariant measure" (think of it as the most likely place the leaf ends up after floating for a long time). The authors proved that their method doesn't just wander aimlessly; it eventually finds this correct "home" and stays there. Furthermore, they showed that the digital "home" is extremely close to the real one, converging at that same 1/2 rate.

The Proof in the Pudding
To make sure their theory wasn't just pretty math, they ran computer experiments on three different types of "rivers":

Stiff Systems: These are like rivers with sudden, violent rapids. The adaptive method handled them better and faster than older, fixed-step methods.
Non-Stiff Systems: These are calmer rivers. The new method was still faster and more accurate.
Langevin Systems: These are used to model things like molecules or data in machine learning. Here, the goal is to sample the "invariant measure" (the long-term distribution). The authors showed their method could reconstruct the target distribution with high accuracy while using less computer time than competing methods.

The Bottom Line
This paper doesn't just suggest a new way to solve these equations; it provides a solid mathematical guarantee that the new way works. It bridges a gap between theory and practice, showing that by letting the computer "think" about how big a step to take, we can solve problems that were previously too dangerous or unstable to handle. It's a win for anyone trying to simulate the chaotic, beautiful, and unpredictable nature of the real world.

Technical Summary: Adaptive Time-Stepping Euler–Maruyama Scheme for SDEs with Non-Globally Lipschitz Coefficients

Problem Statement
The paper addresses the numerical approximation of stochastic differential equations (SDEs) where the drift and diffusion coefficients are locally Lipschitz but not globally Lipschitz. Such equations frequently arise in applications like statistical physics (e.g., the overdamped Langevin equation) and machine learning, where the goal is often to sample from a target probability distribution (the invariant measure).

Standard explicit methods, such as the Euler–Maruyama (EM) scheme, often fail for these systems. Specifically, explicit EM schemes can exhibit numerical instability (explosions) or fail to preserve the ergodic properties of the underlying continuous-time dynamics, even when the exact solution is exponentially ergodic. While adaptive time-stepping methods have been proposed to mitigate instability by adjusting the step size based on the current state, existing theoretical results for SDEs with fully locally Lipschitz coefficients remain sparse. Key gaps include the lack of uniform-in-time strong convergence analysis, limited understanding of moment exponential stability in adaptive settings, and unresolved questions regarding the preservation of ergodicity and the convergence rate of the numerical invariant measure.

Methodology
The authors propose a continuous-time adaptive time-stepping Euler–Maruyama (EM) scheme. The core mechanism involves a dynamically adjusted timestep function $\delta(x)$ , which is continuous and strictly positive. This function is designed to satisfy specific growth conditions relative to the drift $f(x)$ and diffusion $g(x)$ coefficients to prevent numerical explosions.

Key methodological components include:

Adaptive Timestep Construction: The step size $\delta_n$ at iteration $n$ is determined by the state $Y_{t_n}$ . The function $\delta(\cdot)$ is chosen such that $|f(x)|(\delta(x))^\gamma \leq 1 + |x|$ and $|g(x)|^2(\delta(x))^\gamma \leq (1+|x|)^2$ for some $\gamma \in (0, 1/3]$ .
Auxiliary K-Scheme: To prove that the terminal time $T$ is almost surely attainable (i.e., the simulation does not get stuck in infinite steps before reaching $T$ ) and to establish moment bounds, the authors employ an auxiliary truncated scheme (the "K-scheme") introduced by Fang and Giles. This involves truncating the solution at a radius $K$ and analyzing the limit as $K \to \infty$ .
Continuous-Time Interpolation: The discrete numerical solution is extended to continuous time via piecewise linear and piecewise constant interpolations ( $\hat{Y}(t)$ and $Y(t)$ ) to facilitate the application of Itô's formula and martingale techniques.
Stability and Ergodicity Analysis: The analysis utilizes Lyapunov functions, the optional stopping theorem, and Harris's theorem for Markov chains. Specifically, the authors establish that the discrete numerical solution forms a time-homogeneous, irreducible, and aperiodic Markov chain.

Key Contributions and Results

Attainability and Moment Boundedness:
- The paper proves that the terminal time $T > 0$ is almost surely attainable, ensuring the adaptive iteration reaches $T$ in a finite number of steps.
- Moment boundedness of the numerical solution is established for both finite-time intervals and uniformly in time, provided the coefficients satisfy specific polynomial growth and coercivity conditions.
Strong Convergence:
- Finite-Time: The scheme achieves a strong convergence rate of order $1/2$ on finite time intervals.
- Uniform-in-Time: Under slightly strengthened assumptions (specifically regarding the one-sided Lipschitz condition and growth of coefficients), the paper establishes a uniform-in-time strong convergence rate of order $1/2$ . This result holds without imposing a uniform positive minimum step size.
- For SDEs with additive noise, a first-order strong convergence rate is attainable under stronger assumptions on the drift.
Stability Preservation:
- The adaptive scheme is shown to faithfully inherit the $p$ -th moment exponential stability of the exact solution. This fills a theoretical gap, as moment exponential stability in adaptive settings had rarely been investigated previously.
Ergodicity and Invariant Measure Convergence:
- The numerical solution is proven to be ergodic, possessing a unique invariant probability measure.
- The paper establishes polynomial ergodicity for the numerical scheme.
- Crucially, the authors derive a convergence rate of $1/2$ in the $L_q$ -Wasserstein distance between the numerical invariant measure and the exact invariant measure of the underlying SDE. This confirms that the scheme accurately approximates the target distribution in the long-time limit.
Numerical Validation:
- Numerical experiments on stiff SDEs, non-stiff SDEs (Ginzburg-Landau), and high-dimensional overdamped Langevin systems (including Bayesian inference examples) validate the theoretical results.
- Comparisons with fixed-step schemes (backward EM, tamed EM, truncated EM) and existing adaptive methods (LMS, Lemaire's method) demonstrate that the proposed scheme achieves superior accuracy and computational efficiency (lower CPU time for comparable error).

Significance
The paper claims to bridge critical gaps in the numerical analysis of SDEs with non-globally Lipschitz coefficients. By developing an explicit adaptive scheme that guarantees uniform-in-time strong convergence and preserves both moment exponential stability and ergodicity, the work provides a robust theoretical foundation for sampling from complex distributions. The derivation of the optimal $1/2$ -order convergence rate for the invariant measure in the Wasserstein distance is highlighted as a significant theoretical advancement, offering a reliable tool for applications in computational statistics, statistical physics, and machine learning where long-time behavior and distribution sampling are paramount.

The Paper's Story: A Smart, Adaptive Hiker

More like this