Density convergence of a fully discrete finite difference method for stochastic Cahn--Hilliard equation

This paper establishes the $L^1(\mathbb{R})$ convergence of the probability density for a fully discrete finite difference method solving the stochastic Cahn--Hilliard equation with multiplicative space-time white noise by introducing a novel localization argument to overcome the non-Lipschitz drift and partially resolving an open problem regarding the numerical computation of the solution's density.

The Gist

Imagine you are trying to predict the weather, but instead of just temperature and wind, you are dealing with a chaotic, bubbling mixture of chemicals that is constantly changing shape and splitting into different phases (like oil and water separating). This is what the Stochastic Cahn–Hilliard Equation models. It describes how materials separate and evolve over time, but with a twist: there is random noise (like static on a radio) constantly jiggling the system, making exact predictions impossible.

In this paper, the authors are trying to answer a very specific question: "If we use a computer to simulate this messy process, can we trust the computer's picture of the 'probability' of what's happening?"

Here is a breakdown of their work using simple analogies:

1. The Problem: The "Wild" Chemical Reaction

Think of the chemical mixture as a pot of soup.

• The Drift (The Recipe): The soup has a natural tendency to separate into two distinct flavors (phases). Mathematically, this is the "drift coefficient." The problem is that this recipe is wild. It's not a gentle, predictable slope; it's a steep cliff that gets steeper the more you move away from the center. In math terms, it's "neither globally Lipschitz nor one-sided Lipschitz."
• The Noise (The Jiggling): Someone is constantly shaking the pot with random, unpredictable jolts (multiplicative space-time white noise).
• The Goal: We want to know the density. Imagine taking a photo of the soup at a specific moment. The "density" is a map showing exactly how likely the soup is to be in any specific state at that moment.

2. The Challenge: The Computer's Struggle

When you try to simulate this on a computer, you have to chop time and space into tiny grid squares (a Finite Difference Method).

• The Trap: Because the "recipe" (drift) is so wild and steep, standard computer methods often crash or give wildly wrong answers. It's like trying to walk up a cliff that gets steeper the higher you go; a normal step-by-step approach will make you fall off.
• The Missing Piece: Previous research could simulate the average behavior, but they couldn't prove that the computer's probability map (the density) actually matched the real world's probability map.

3. The Solution: The "Safety Net" (Localization)

The authors invented a clever trick called Localization.

• The Metaphor: Imagine you are trying to predict the path of a runaway horse (the wild chemical reaction). You can't predict the whole journey because the horse might run off a cliff.
• The Trick: Instead of trying to track the horse forever, you put a giant, invisible fence around a safe area (a "box" of size $R$).
  1. Inside the box: The horse behaves nicely. The "recipe" becomes tame and predictable. You can easily prove your computer simulation works here.
  2. Outside the box: The horse is wild, but the chance of the horse running that far away is incredibly small (like winning the lottery twice in a row).
• The Result: By proving the simulation works perfectly inside the safe box, and showing that the "wild" outside area is so unlikely to happen, they proved the entire simulation is accurate.

4. The "Strong Convergence" (The Speedometer)

Before proving the probability map is right, they had to prove the simulation itself was accurate.

• They showed that as you make the computer's grid finer (smaller steps in time and space), the computer's result gets closer to the true answer at a specific, optimal speed.
• Analogy: If you are drawing a curve with a ruler, the more lines you draw, the smoother the curve looks. They proved exactly how fast the curve becomes smooth as you add more lines.

5. The Big Win: The "Density Convergence"

This is the main achievement.

• The Question: Does the computer's "probability map" (the density) look like the real probability map?
• The Answer: Yes.
• They proved that the difference between the computer's map and the real map shrinks to zero as the simulation gets more detailed. In mathematical language, the "Total Variation Distance" between the two maps goes to zero.

Summary in Plain English

The authors took a notoriously difficult, chaotic mathematical model (a soup that separates while being shaken by random noise) and built a computer program to simulate it.

Because the math was so messy, they couldn't just run the code and hope for the best. They developed a new strategy:

1. They created a "safety zone" where the math behaves nicely.
2. They proved the computer simulation is perfect inside that zone.
3. They proved that the chance of the system escaping that zone is so tiny it doesn't matter.

The Result: They successfully proved that this computer method doesn't just guess the average behavior; it accurately reconstructs the entire probability landscape of the system. This solves a long-standing open problem in the field, giving scientists a reliable tool to predict the behavior of complex materials under random conditions.

Technical Summary

Here is a detailed technical summary of the paper "Density Convergence of a Fully Discrete Finite Difference Method for Stochastic Cahn–Hilliard Equation" by Hong, Jin, and Sheng.

1. Problem Statement

The paper addresses the numerical approximation of the probability density function (PDF) of the solution to the Stochastic Cahn–Hilliard (SCH) equation driven by multiplicative space-time white noise:
$$\partial_t u + \Delta^2 u = \Delta f(u) + \sigma(u) \dot{W}$$
where $f(u) = u^3 - u$ (the derivative of a double-well potential) and $\sigma(u)$ is a multiplicative noise coefficient.

Key Challenges:

1. Non-Globally Lipschitz Drift: The nonlinear term $\Delta f(u)$ involves a cubic polynomial, making the drift coefficient neither globally Lipschitz nor one-sided Lipschitz. This complicates standard error analysis and the existence of densities for numerical schemes.
2. Density Convergence: While strong convergence rates for numerical solutions of SPDEs are well-studied, the convergence of the density of the numerical solution to the density of the exact solution is a much harder problem, particularly for SPDEs with polynomial nonlinearities. Previous work was limited to additive noise or globally Lipschitz coefficients.
3. Open Problem: The authors aim to resolve an open problem posed in [Cui & Hong, 2020] regarding the numerical computation of the density for SCH equations with multiplicative noise.

2. Methodology

The authors propose a Fully Discrete Finite Difference Method (FDM) combining spatial discretization and a backward Euler time-stepping scheme.

A. Numerical Schemes

1. Spatial Discretization: The domain is discretized with step size $h = \pi/n$. The spatial FDM transforms the SPDE into a system of $(n-1)$-dimensional Stochastic Differential Equations (SDEs).
2. Temporal Discretization: A backward Euler scheme with time step $\tau = T/m$ is applied to the spatial semi-discrete system, resulting in a fully discrete scheme.

B. Strong Convergence Analysis

To handle the non-globally Lipschitz drift, the authors introduce a novel auxiliary process $\tilde{U}(t)$ (and $\tilde{U}^i$ for the fully discrete case).

• Strategy: Instead of comparing the numerical solution $U$ directly to the exact solution $u$, they compare $U$ to $\tilde{U}$, and $\tilde{U}$ to $u$.
• Error Decomposition:
  • $u - \tilde{u}$: Analyzed using properties of the discrete Green's function and the regularity of the exact solution.
  • $\tilde{u} - u_{num}$: Analyzed using the local weak monotonicity of the drift in a negative Sobolev norm and the local Lipschitz continuity in a positive Sobolev norm.
• Key Technical Tool: The use of discrete negative Sobolev norms $\|(-A_n)^{-1/2} \cdot \|_{l^2_n}$ allows the authors to exploit the one-sided Lipschitz property of the linear part to control the error, despite the super-linear growth of the nonlinearity.

C. Density Convergence Analysis (The Core Innovation)

To prove the convergence of the density in $L^1(\mathbb{R})$, the authors develop a localization argument:

1. Localization: They introduce a smooth cut-off function $K_R$ to define a localized version of the equation where the nonlinearity $f_R = f \cdot K_R$ becomes globally Lipschitz.
2. Criterion for Total Variation (TV) Distance: They establish a criterion (Proposition 6.1) to reduce the TV distance between the original random variables to the TV distance between their localizations.
   $$\limsup_{N \to \infty} d_{TV}(X_N, X) \leq \limsup_{R \to \infty} \limsup_{N \to \infty} d_{TV}(X_{N,R}, X_R)$$
3. Malliavin Calculus: For the localized (globally Lipschitz) system, they utilize Malliavin calculus (specifically the Bouleau-Hirsch criterion and convergence in the Malliavin-Sobolev space $D^{1,2}$) to prove that the numerical solution admits a density and that the TV distance converges to zero.
4. Synthesis: By combining the strong convergence rates of the numerical method with the density convergence of the localized system, they deduce the density convergence for the original non-globally Lipschitz problem.

3. Key Contributions

1. Optimal Strong Convergence Rates:

   • Spatial FDM: Proved a strong convergence order of 1 (coinciding with the spatial Hölder continuity of the exact solution).
   • Fully Discrete FDM: Proved a strong convergence order of nearly 3/8 in time (coinciding with the temporal Hölder continuity).
   • These results are optimal and hold for polynomial nonlinearities and multiplicative noise.

2. First Density Convergence for Polynomial Nonlinearities:

   • This is the first work to establish the $L^1(\mathbb{R})$ convergence of the density for numerical approximations of SPDEs with polynomial nonlinearities (specifically the Cahn-Hilliard equation).
   • It partially answers the open problem raised in [Cui & Hong, 2020] regarding the numerical computation of densities for such systems.

3. Novel Localization Argument:

   • The authors propose a generalizable criterion to reduce the total variation distance of random variables to that of their localizations. This technique bypasses the difficulty of proving global Malliavin differentiability for the exact solution of non-globally Lipschitz SPDEs, which is often unknown or difficult to establish.

4. Main Results

• Theorem 4.7 & 5.7: Establish the strong convergence rates:
  $$\mathbb{E}[|u(t,x) - u_n(t,x)|^\zeta] \leq C n^{-\zeta}$$
  $$\mathbb{E}[|u_n(t,x) - u_{n,\tau}(t,x)|^\zeta] \leq C \tau^{(3/8 - \epsilon)\zeta}$$
  for $\zeta \in [1, 2)$.

• Theorem 6.4: Establishes the convergence of the probability density functions in $L^1(\mathbb{R})$:
  $$\lim_{n \to \infty} \int_{\mathbb{R}} |p^n_{T, kh}(\xi) - p_{T, kh}(\xi)| d\xi = 0$$
  $$\lim_{n \to \infty} \lim_{\tau \to 0} \int_{\mathbb{R}} |p^{n,\tau}_{T, kh}(\xi) - p_{T, kh}(\xi)| d\xi = 0$$
  where $p$ denotes the density of the exact solution and $p^n, p^{n,\tau}$ denote the densities of the spatial and fully discrete numerical solutions, respectively.

5. Significance

• Theoretical Advancement: The paper bridges a significant gap in the numerical analysis of SPDEs. While strong convergence is standard, density convergence is crucial for probabilistic applications (e.g., calculating probabilities of rare events, risk assessment) but remains elusive for complex SPDEs.
• Methodological Innovation: The localization argument combined with Malliavin calculus provides a robust framework for handling non-globally Lipschitz coefficients. This approach is expected to be applicable to other SPDEs with similar difficulties, such as the stochastic Allen-Cahn equation.
• Practical Implication: It validates the use of fully discrete finite difference methods for simulating not just the sample paths, but also the statistical distribution (density) of phase separation processes modeled by the stochastic Cahn-Hilliard equation. This is vital for materials science applications where understanding the probability distribution of concentration phases is essential.