Analysis of an exponential integrator for stochastic PDEs driven by Riesz noise

1. Problem Statement

The paper addresses the numerical approximation of semilinear stochastic partial differential equations (SPDEs) of parabolic type (specifically stochastic heat equations) in spatial dimensions $d \geq 1$. The equations are driven by a Gaussian noise that is white in time but possesses spatial correlations defined by a Riesz kernel.

The specific equation under study is:
$$\frac{\partial u}{\partial t}(t, x) = \Delta u(t, x) + b(t, x, u(t, x)) + \sigma(t, x, u(t, x)) \dot{F}_\alpha(t, x)$$
where:

• $\Delta$ is the Laplacian operator on the domain $Q = (0, 1)^d$ with homogeneous Dirichlet boundary conditions.
• $\dot{F}_\alpha$ represents the noise with spatial correlation given by the Riesz kernel $|x-y|^{-\alpha}$, where the parameter $\alpha \in (0, 2 \wedge d)$.
• $b$ and $\sigma$ are nonlinear coefficients satisfying Lipschitz and linear growth conditions.

The Challenge: While convergence results for SPDEs in 1D are well-established, results for dimensions $d > 1$ with random field settings (specifically Riesz noise) are scarce. Existing schemes often suffer from strict CFL (Courant–Friedrichs–Lewy) stability restrictions or require high regularity of the initial data.

2. Methodology

The authors propose and analyze an explicit stochastic exponential integrator. The methodology involves three main pillars:

A. Mathematical Framework

• Noise Definition: The Riesz noise is defined via a Hilbert space $H_\alpha$ constructed using the inner product induced by the Riesz kernel. The stochastic integral is interpreted in the Walsh–Dalang sense.
• Mild Solution: The SPDE is reformulated as a mild solution involving the heat kernel $G_d$ (the Green's function for the Laplacian with Dirichlet BCs).
• Regularity Analysis: The authors establish precise temporal and spatial Hölder regularity estimates for the mild solution. Crucially, they prove that the solution's time regularity is limited by the noise parameter $\alpha$, specifically being Hölder continuous with exponent $\gamma < \frac{1}{2} - \frac{\alpha}{4}$.

B. The Numerical Scheme

The proposed Stochastic Exponential Integrator is defined in continuous time as:
$$u_m(t, x) = \int_Q G_d(t, x, y)u_0(y) dy + \int_0^t \int_Q G_d(t - \kappa_m(s), x, y) b(\dots) dy ds + \int_0^t \int_Q G_d(t - \kappa_m(s), x, y) \sigma(\dots) F_\alpha(ds, dy)$$
where $\kappa_m(s)$ is the step function mapping time $s$ to the previous grid point.
In discrete form (at time steps $t_\ell = \ell \tau$), the scheme is:
$$U_m^{\ell+1} = e^{\tau \Delta} \left( U_m^\ell + \tau b(t_\ell, \cdot, U_m^\ell) + \sigma(t_\ell, \cdot, U_m^\ell) \delta_\ell F_\alpha \right)$$
Here, $e^{\tau \Delta}$ is the semigroup operator (convolution with the heat kernel), and $\delta_\ell F_\alpha$ represents the increment of the Riesz noise over the time step.

C. Error Analysis

The proof of convergence relies on:

1. Decomposition of Error: The error between the numerical solution and the mild solution is decomposed into drift and diffusion terms, further split into components related to time discretization, spatial regularity, and noise approximation.
2. Heat Kernel Estimates: The authors derive sharp estimates for the heat kernel $G_d$ and its derivatives in the norms associated with the Riesz noise covariance. These estimates are critical for bounding the stochastic integrals.
3. Burkholder–Davis–Gundy (BDG) Inequality: Adapted for the Riesz noise setting to bound the moments of the stochastic convolution.
4. Gronwall Inequality: Used to close the error estimates and derive the final convergence rate.

3. Key Contributions

1. Optimal Strong Convergence Rate: The paper proves that the explicit exponential integrator achieves a strong convergence rate of $\gamma$ for any $\gamma < \frac{1}{2} - \frac{\alpha}{4}$. This rate is shown to be optimal, as it coincides with the intrinsic temporal regularity of the mild solution itself.
2. Relaxed Initial Data Assumptions: Unlike many previous works that require the initial condition $u_0$ to be in specific Sobolev spaces, this scheme only requires $u_0$ to be a continuous function ($u_0 \in C_0(Q)$).
3. Uniform Error Bounds: The error estimates are uniform with respect to both time and space, a significant improvement over pointwise estimates found in some literature.
4. Implementation Advantages:
   • No CFL Restriction: Unlike explicit Euler–Maruyama schemes, the exponential integrator does not suffer from severe time-step restrictions imposed by the stiffness of the Laplacian operator.
   • Efficiency: It avoids the need to solve linear systems at each step (unlike implicit schemes), making it computationally cheaper per step while maintaining stability.
5. Extension to Higher Dimensions: The work provides the first numerical confirmation of convergence rates for SPDEs driven by Riesz noise in spatial dimension $d=2$ within a random field setting.

4. Results

• Theoretical: Theorem 8 establishes the strong error bound:
  $$\sup_{(t,x) \in [0,T] \times Q} \mathbb{E}[|u_m(t,x) - u(t,x)|^{2p}]^{1/2p} \leq C \tau^\gamma$$
  where $\tau$ is the time step size.
• Numerical Experiments:
  • Dimension 1: Experiments confirm the theoretical rate. For $\alpha=0.7$, the observed rate is $\approx 0.65$ ($1 - \alpha/2$ in the specific test case, consistent with the theoretical bound). For smoother noise ($\alpha=0.2$), the rate improves to $\approx 0.9$.
  • Dimension 2: The authors successfully simulate the SPDE on a 2D grid. The results confirm that the convergence rate depends on $\alpha$ as predicted ($\approx 1/2 - \alpha/4$) and is independent of the spatial grid size $n$.
  • Comparison: The exponential integrator (SEXP) demonstrates superior performance compared to the explicit Euler–Maruyama (which fails or requires tiny steps due to CFL) and the semi-implicit Euler–Maruyama (which requires solving linear systems). The SEXP scheme achieves lower errors for the same computational time.

5. Significance

This paper bridges a critical gap in the numerical analysis of SPDEs in higher dimensions with correlated noise.

• Theoretical Impact: It provides a rigorous framework for analyzing exponential integrators with Riesz noise, extending Dalang's theory to bounded domains and establishing sharp error bounds that depend explicitly on the noise correlation parameter $\alpha$.
• Practical Impact: The proposed scheme offers a robust, efficient, and stable alternative for simulating stochastic heat equations in 2D and potentially higher dimensions. The removal of the CFL condition makes it particularly suitable for problems where the diffusion term is stiff, a common scenario in physical applications.
• Novelty: The demonstration of convergence in 2D with Riesz noise is a significant step forward, as most existing literature is restricted to 1D or space-time white noise.

In summary, the authors have developed a theoretically sound and practically efficient numerical method for a class of SPDEs that are difficult to simulate using standard schemes, proving that the exponential integrator is optimal for this specific noise structure.