Central limit theorem for temporal average of backward Euler--Maruyama method

This paper establishes the central limit theorem for the temporal average of the backward Euler–Maruyama method applied to stochastic ordinary differential equations with super-linearly growing drift coefficients, deriving the result through direct analysis for sub-optimal deviation orders and via a Poisson equation approach for the optimal strong order.

The Gist

Imagine you are trying to predict the long-term weather of a chaotic, stormy planet. You can't see the future, and the weather changes randomly every second. However, you know that over a very long time, the planet settles into a "normal" climate pattern. This pattern is called the Ergodic Limit.

In the world of mathematics and physics, scientists use equations (called Stochastic Ordinary Differential Equations, or SODEs) to model these random systems. But these equations are often too messy to solve exactly. So, we use computers to take small steps forward in time, simulating the weather step-by-step. This is called a numerical method.

One popular method is the Backward Euler–Maruyama (BEM) method. Think of it as a very sturdy, cautious hiker. Unlike other hikers who might trip over a rock (mathematical instability) if the terrain gets too steep (super-linear growth), the BEM hiker looks ahead and adjusts their footing to ensure they don't fall, even on the steepest, most chaotic mountains.

The Big Question: How "Normal" is the Average?

The paper asks a specific question: If we let this hiker walk for a long time and take the average of their path, how close is that average to the true "climate" (the ergodic limit)?

We already know the average gets closer as the steps get smaller. But the authors wanted to know something deeper: How does the average fluctuate?

Imagine you ask 1,000 different people to walk the same path and calculate their averages. Will their results be identical? No. Some will be slightly higher, some slightly lower. The Central Limit Theorem (CLT) is a famous rule in statistics that says if you look at enough of these averages, they will form a perfect "Bell Curve" (a normal distribution).

The Goal of this Paper:
The authors wanted to prove that even for these difficult, steep mountains (systems with "super-linear" coefficients), the BEM hiker's long-term average still follows this beautiful Bell Curve rule.

The Two Scenarios: Walking Slowly vs. Walking Fast

The paper splits the proof into two scenarios based on how fast the hiker is walking (the "deviation order"):

1. The Slow Walk (Deviation Order < 1/2):
Imagine the hiker is taking tiny, cautious steps. In this case, the error in their path is so small compared to the natural randomness of the wind that we can just borrow the known rules of the "real" weather.

• The Analogy: It's like saying, "If the hiker is walking so slowly that their mistakes are invisible compared to the wind, then their average path behaves exactly like the real weather."
• The Result: The math is straightforward. The average follows the Bell Curve.

2. The Fast Walk (Deviation Order = 1/2):
Now, imagine the hiker is walking faster. The mistakes they make (due to the steep terrain) are now significant enough to matter. We can't just borrow the rules anymore; we have to build a new bridge.

• The Analogy: The hiker is now fighting the wind and the steep rocks. To understand their average path, the authors used a special tool called the Poisson Equation. Think of this as a "map of corrections." It calculates exactly how much the hiker's path deviates from the ideal path at every single step.
• The Result: By using this map, they broke the problem down into a series of small, random "jumps" (a martingale). They proved that even with these jumps, the final average still settles into that perfect Bell Curve.

Why is this a Big Deal?

1. Handling the "Impossible": Most previous math papers could only prove this for smooth, gentle hills (Lipschitz coefficients). This paper proves it works for the jagged, steep, "super-linear" mountains that appear in real-world physics and biology (like population explosions or chemical reactions).
2. Confidence in Simulations: If you are a scientist simulating a complex system, this paper tells you: "You can trust your computer's long-term average. Not only is it accurate, but you can also calculate the probability of your result being off by a certain amount."
3. The "Infinite Horizon" Secret: To prove this, the authors had to show that the BEM method doesn't just work for a short time, but stays stable and bounded even if you let it run forever. This was a new mathematical discovery for this specific type of difficult equation.

The Experiment: The Computer Test

To prove they weren't just dreaming, the authors ran computer simulations.

• They created a fake chaotic system.
• They let the BEM method run for thousands of steps.
• They checked the distribution of the averages.
• The Result: The data points lined up perfectly with the theoretical Bell Curve, confirming their math was correct.

In a Nutshell

This paper is like a guidebook for a very brave hiker (the BEM method). It proves that even when the mountain is incredibly steep and dangerous, if you walk long enough, your average position will follow a predictable, beautiful statistical pattern. This gives scientists the confidence to use these methods to model complex, real-world chaos with high precision.

Technical Summary

Here is a detailed technical summary of the paper "Central Limit Theorem for Temporal Average of Backward Euler–Maruyama Method" by Diancong Jin.

1. Problem Statement

The paper addresses the numerical approximation of the ergodic limit (long-time statistical behavior) of Stochastic Ordinary Differential Equations (SODEs) with super-linearly growing drift coefficients.

• The Equation: The study focuses on SODEs of the form $dX(t) = b(X(t))dt + \sigma(X(t))dW(t)$, where the drift $b$ satisfies a dissipative condition but is allowed to grow super-linearly (non-Lipschitz), while the diffusion $\sigma$ is Lipschitz continuous.
• The Challenge: While the convergence of numerical temporal averages to the ergodic limit in the moment sense (strong convergence) is well-established for such systems using the Backward Euler–Maruyama (BEM) method, the distributional asymptotics (fluctuations around the limit) remain unexplored for non-Lipschitz coefficients.
• The Goal: To establish a Central Limit Theorem (CLT) for the temporal average of the BEM method, characterizing the asymptotic distribution of the error between the numerical average and the true ergodic limit.

2. Methodology

The author employs a combination of probabilistic analysis, stochastic numerical analysis, and partial differential equation (PDE) techniques.

• Numerical Scheme: The Backward Euler–Maruyama (BEM) method is used due to its stability properties for stiff or super-linearly growing SODEs.
  $$\bar{X}_{n+1} = \bar{X}_n + b(\bar{X}_{n+1})\tau + \sigma(\bar{X}_n)\Delta W_n$$
• Temporal Average Definition: The study analyzes the normalized temporal average:
  $$\Pi_{\tau, \alpha}(h) = \frac{1}{\tau^{-\alpha}} \sum_{k=0}^{\tau^{-\alpha}-1} h(\bar{X}_k)$$
  where $\tau$ is the time step, $\alpha \in (1, 2]$ is a parameter controlling the deviation scale, and $h$ is a test function.
• Proof Strategies (Split by Deviation Order):
  1. Case $\alpha \in (1, 2)$ (Deviation order < Optimal Strong Order):
     • The proof relies on the classical CLT for the exact solution of the SODE and the uniform strong convergence order of the BEM method ($O(\tau^{1/2})$).
     • Since the deviation scaling $\tau^{(\alpha-1)/2}$ is "slower" than the strong error order, the numerical error is negligible in the limit, allowing the CLT of the numerical solution to inherit the CLT of the exact solution directly via Slutsky's theorem.
  2. Case $\alpha = 2$ (Deviation order = Optimal Strong Order):
     • This is the critical case where the numerical error is of the same order as the fluctuation.
     • The author utilizes the Poisson equation associated with the generator $\mathcal{L}$ of the SODE: $\mathcal{L}\phi = h - \pi(h)$.
     • The normalized error is decomposed into a martingale difference series (which converges to a normal distribution) and a negligible remainder (which converges to zero in probability).
     • This decomposition requires precise control over the regularity of the solution $\phi$ to the Poisson equation and the moment bounds of the numerical scheme.

3. Key Contributions

The paper makes two primary theoretical contributions:

1. Establishment of the CLT for Non-Lipschitz SODEs:

   • It extends existing CLT results (which were limited to Lipschitz coefficients) to SODEs with super-linearly growing drift coefficients.
   • It proves that the normalized temporal average converges in distribution to a Gaussian distribution $N(0, \pi(|\sigma^\top \nabla \phi|^2))$ for both $\alpha \in (1, 2)$ and $\alpha = 2$.

2. Moment Boundedness of the BEM Method:

   • A crucial technical hurdle for the $\alpha=2$ case is the need for $p$-th moment boundedness ($p > 2$) of the BEM method over an infinite time horizon.
   • The paper provides a rigorous proof that the BEM method satisfies $\sup_{n \ge 0} E|\bar{X}_n|^p \le K(1 + |x|^p)$ for any $p \ge 1$, even with super-linear drift. This result was previously unreported for SODEs with non-Lipschitz coefficients.

4. Key Results

• Theorem 3.2 & 3.4: For $h \in C_b^4(\mathbb{R}^d)$, as $\tau \to 0$:
  $$\frac{1}{\tau^{\frac{\alpha-1}{2}}} \left( \Pi_{\tau, \alpha}(h) - \pi(h) \right) \xrightarrow{d} \mathcal{N}\left(0, \pi(|\sigma^\top \nabla \phi|^2)\right)$$
  where $\phi$ is the solution to the Poisson equation $\mathcal{L}\phi = h - \pi(h)$.
• Regularity of Poisson Solution: The paper establishes that under the given assumptions, the solution $\phi$ to the Poisson equation possesses sufficient regularity (up to 4th order derivatives) and polynomial growth bounds, which are essential for the error analysis.
• Numerical Verification:
  • Experiments on SODEs with both Lipschitz and non-Lipschitz diffusion coefficients confirm the theoretical convergence rates.
  • The numerical results show that the expectation of the normalized error converges to a constant (the variance of the limiting normal distribution) as the step size $\tau$ decreases, validating the CLT even for super-linear growth cases.

5. Significance

• Theoretical Advancement: This work bridges a gap in the numerical analysis of ergodic SODEs. While strong convergence and invariant measure approximation were known for BEM with super-linear drift, the statistical distribution of the error (fluctuations) was unknown. This paper provides the necessary theoretical foundation for uncertainty quantification in long-time simulations of such systems.
• Practical Implication: In applications like physics, biology, and chemistry where SODEs often feature super-linear drifts (e.g., in chemical reaction networks or fluid dynamics), practitioners can now rely on the CLT to construct confidence intervals for ergodic averages computed via the BEM method.
• Robustness: The proof techniques, particularly the derivation of infinite-horizon moment bounds for the BEM method, offer a robust framework for analyzing other numerical schemes applied to non-Lipschitz stochastic systems.

In summary, Diancong Jin's work rigorously proves that the Backward Euler–Maruyama method not only converges to the correct ergodic limit for super-linear SODEs but also exhibits Gaussian fluctuations around that limit, provided the time step is sufficiently small. This validates the method's reliability for statistical sampling in complex, non-Lipschitz stochastic environments.