Asymptotics of large deviations of finite difference method for stochastic Cahn--Hilliard equation

This paper establishes the Freidlin--Wentzell large deviations principle for the stochastic Cahn--Hilliard equation with small noise and proves the convergence of the one-point large deviations rate function for its spatial finite difference method by utilizing $\Gamma$-convergence of objective functions and overcoming non-Lipschitz drift challenges through discrete interpolation inequalities.

The Gist

Imagine you are watching a pot of molten metal cool down. As it cools, it doesn't just turn into a solid block; it separates into distinct regions, like oil and water separating, creating a complex, swirling pattern of two different phases. This is the Cahn-Hilliard equation in action—a mathematical model describing how materials separate and evolve over time.

Now, imagine this process isn't perfectly smooth. There are tiny, random jitters, like a gentle breeze or microscopic vibrations, disturbing the cooling metal. This is the Stochastic Cahn-Hilliard equation. The "noise" represents these random disturbances.

The Big Question: What are the odds of a "Weird" Outcome?

Usually, if you run this experiment a thousand times, the metal will settle into a very predictable, average pattern. But sometimes, by pure chance, the metal might settle into a very strange, unlikely pattern.

The authors of this paper are interested in Large Deviations. They want to know: How likely is it that the metal will end up in a weird state, and how does that probability change as the random jitters (noise) get smaller?

They found that the probability of these rare, weird events drops off incredibly fast (exponentially) as the noise gets quieter. The speed of this drop-off is measured by something called a Rate Function. Think of the Rate Function as a "difficulty score" for a specific weird outcome. A high score means it's extremely hard (rare) for the metal to end up there; a low score means it's easier (more likely).

The Problem: Computers Can't Solve Everything Perfectly

Mathematicians can calculate this "difficulty score" for the real, perfect physical system. But in the real world, we use computers to simulate these systems. Computers can't handle continuous, smooth curves perfectly; they have to chop the problem up into tiny little blocks (pixels, if you will) and solve it step-by-step. This is called the Finite Difference Method (FDM).

The big worry for scientists is: Does the computer simulation preserve the "difficulty scores" of the real world?

If the real world says a weird pattern is "impossible" (infinite difficulty), but the computer says it's "easy" (low difficulty), then the simulation is lying to us. We need to know if the computer's "difficulty score" gets closer and closer to the real "difficulty score" as we make the computer's blocks smaller and smaller.

The Solution: A Mathematical "Zoom-In"

The authors of this paper proved that yes, the computer simulation does get it right.

Here is the analogy they used to prove it:

1. The Skeleton Equation: Imagine the random noise is removed entirely. The system follows a strict, deterministic path based on the "best" possible route to reach a specific outcome. This path is called the "skeleton." The difficulty score is essentially the "energy" required to push the system along this skeleton path.
2. The Digital vs. The Real: The authors compared the "skeleton" of the real world with the "skeleton" of the computer simulation.
3. The Challenge: The math behind the Cahn-Hilliard equation is tricky. The forces involved aren't "nice" and smooth; they can get wild and unpredictable (mathematicians call this "non-Lipschitz"). This usually breaks standard computer proofs.
4. The Trick: The authors used a clever mathematical tool called $\Gamma$-convergence.
   • Analogy: Imagine you are trying to find the lowest point in a mountain range (the easiest path). The real mountain has a jagged, complex shape. The computer model approximates this mountain with a staircase.
   • $\Gamma$-convergence is a rigorous way of proving that as you make the steps of the staircase smaller and smaller, the "lowest point" you find on the staircase converges to the true lowest point of the real mountain.

The Main Takeaway

The paper proves that as you refine your computer simulation (make the grid finer and finer), the "difficulty score" for rare events in the simulation converges to the true difficulty score of the real physical system.

Why does this matter?
It means that when scientists use computers to predict rare, catastrophic, or unusual events in materials science (like a metal suddenly cracking in a weird way due to tiny vibrations), they can trust the computer's prediction of how rare that event is. The simulation doesn't just look right; it captures the fundamental statistical laws of the universe, even for the most unlikely scenarios.

In short: They proved that the digital map of "rare events" becomes indistinguishable from the real map as the map gets more detailed, giving us confidence in our computer models of complex, noisy materials.

Technical Summary

Here is a detailed technical summary of the paper "Asymptotics of Large Deviations of Finite Difference Method for Stochastic Cahn–Hilliard Equation" by Diancong Jin and Derui Sheng.

1. Problem Statement

The paper addresses the numerical analysis of the Stochastic Cahn–Hilliard (SCH) equation driven by space-time white noise with small noise intensity $\varepsilon$. The equation models phase separation and coarsening in alloys. The specific focus is on the Large Deviation Principle (LDP) and the convergence of the Large Deviation Rate Function (LDRF) when the equation is approximated using a Spatial Finite Difference Method (FDM).

Key challenges identified:

• Non-Lipschitz Drift: The nonlinearity $b(u) = u^3 - u$ in the SCH equation is not one-sided Lipschitz, making standard LDP proofs for numerical schemes difficult.
• Asymptotic Preservation: A practical numerical method should not only converge in the mean-square sense but also asymptotically preserve the exponential decay rate of rare events (hitting probabilities) as the noise intensity $\varepsilon \to 0$ and the discretization parameter $n \to \infty$.
• Gap in Literature: While LDPs for the continuous SCH equation were known, the convergence of the LDRF for numerical discretizations of SPDEs with non-Lipschitz coefficients had not been established.

2. Methodology

The authors employ a rigorous analytical framework combining Large Deviation Theory and $\Gamma$-convergence (variational convergence).

A. Mathematical Framework

1. Freidlin–Wentzell LDP: The authors first establish the LDP for the continuous solution $u^\varepsilon$ on the space $C(O_T; \mathbb{R})$ (continuous functions on space-time). This relies on the weak convergence method (Budhiraja-Dupuis-Maroulas approach), analyzing the "skeleton equation" (the deterministic control equation) associated with the stochastic system.
2. One-Point LDP: By applying the contraction principle to the pathwise LDP, they derive the LDP for the solution at a specific point $(T, \bar{x})$, characterized by a rate function $I(y)$.
3. Numerical Scheme: They construct a spatial FDM for the SCH equation, converting the SPDE into a system of Stochastic Ordinary Differential Equations (SODEs). They prove this discrete system also satisfies an LDP with a discrete rate function $I_n(y)$.

B. Core Analytical Strategy: $\Gamma$-Convergence

The main result (convergence of $I_n$ to $I$) is proven by analyzing the convergence of the underlying objective functions (variational problems) used to define the rate functions.

• The rate functions are defined as infima of action functionals:
  $$I(y) = \inf_{f: f(T, \bar{x})=y} J(f), \quad I_n(y) = \inf_{f: f(T, \bar{x})=y} J_n(f)$$
• The authors prove that the sequence of functionals $\{J_n\}$ $\Gamma$-converges to $J$.
• Technical Route:
  1. Equi-coerciveness: Proving that minimizing sequences for $J_n$ are pre-compact.
  2. $\Gamma$-liminf inequality: Showing lower semicontinuity of the limit.
  3. $\Gamma$-limsup inequality: Constructing a "recovery sequence" to show the upper bound.

C. Overcoming Non-Lipschitz Difficulties

The drift coefficient $b(u) = u^3 - u$ lacks global Lipschitz continuity. To handle this in the discrete setting:

• Equivalent Characterization: They utilize an equivalent characterization of the discrete skeleton equation.
• Discrete Interpolation Inequalities: They derive uniform boundedness estimates for the solutions of the discrete skeleton equations using discrete interpolation inequalities (specifically relating discrete $L^\infty$, $L^2$, and discrete Laplacian norms).
• Discrete Neumann Laplacian Properties: They leverage specific properties of the discrete Neumann Laplacian $\Delta_n$ to handle the nonlinearity $\Delta_n b$ in the $\Gamma$-limsup proof.

3. Key Contributions

1. Validation of Pathwise LDP: The paper establishes the Freidlin–Wentzell LDP for the continuous SCH equation on the space of continuous functions $C(O_T; \mathbb{R})$, strengthening previous results that were limited to $L^p$ or Hölder spaces.
2. Discrete One-Point LDP: It proves that the spatial FDM approximation satisfies a one-point LDP with a well-defined rate function $I_n$.
3. Convergence of Rate Functions: The primary contribution is the proof that the discrete rate function $I_n$ converges pointwise to the continuous rate function $I$ as the mesh size $n \to \infty$.
   $$\lim_{n \to \infty} I_n(y) = I(y)$$
4. Handling Non-Lipschitz Coefficients: This is the first work to establish the convergence of one-point LDRFs for numerical methods applied to SPDEs with non-Lipschitz coefficients (specifically the cubic nonlinearity in Cahn–Hilliard). Previous works were largely restricted to Lipschitz coefficients.

4. Main Results

• Theorem 2.8: The continuous solution $\{u^\varepsilon\}$ satisfies the LDP on $C(O_T; \mathbb{R})$ with a good rate function $J$. Consequently, the one-point process $\{u^\varepsilon(T, \bar{x})\}$ satisfies the LDP on $\mathbb{R}$ with rate function $I$.
• Theorem 3.2: The spatial FDM solution $\{u^{\varepsilon, n}\}$ satisfies the LDP on $\mathbb{R}$ with a discrete rate function $I_n$.
• Theorem 4.14: Under Assumptions 1–3 (boundedness, Lipschitz $\sigma$, non-vanishing $\sigma$, and regularity of initial data $u_0 \in C^3$), the discrete rate function converges to the continuous one:
  $$\lim_{n \to \infty} I_n(y) = I(y), \quad \forall y \in \mathbb{R}.$$
  This implies that the numerical method asymptotically preserves the exponential decay rate of the hitting probability $P(u^\varepsilon(T, \bar{x}) \in A)$.

5. Significance

• Theoretical Advancement: It bridges the gap between large deviation theory and numerical analysis for SPDEs with non-Lipschitz nonlinearities, a class of equations that is notoriously difficult to analyze due to potential blow-up and lack of global Lipschitz bounds.
• Numerical Reliability: The result provides a theoretical guarantee that the spatial FDM is a reliable tool for simulating rare events in phase separation processes. It ensures that the probability of rare deviations is not distorted by the discretization in the limit of small noise.
• Methodological Innovation: The use of $\Gamma$-convergence combined with discrete interpolation inequalities offers a robust template for analyzing the asymptotic properties of other numerical schemes for complex SPDEs.
• Application: The findings are crucial for materials science and physics simulations where understanding the probability of phase transitions (rare events) is as important as the average behavior.