Non-asymptotic uniform in time error bounds for new and old numerical schemes for SPDEs

1. Problem Statement

The paper addresses the numerical approximation of Stochastic Partial Differential Equations (SPDEs) with non-globally Lipschitz nonlinearities, specifically focusing on equations like the stochastic Allen-Cahn equation and stochastic advection-diffusion equations.

• The Challenge: Standard explicit or semi-implicit Euler schemes often fail for SPDEs with super-linear drift terms (e.g., $f(u) = u - u^3$). Specifically, these schemes can exhibit finite-time blow-up (divergence to infinity) if the initial data is sufficiently large, rendering them useless for long-time simulations.
• The Gap: While many existing works provide error bounds for finite time intervals (transient phase) or asymptotic bounds (long-time behavior/invariant measure sampling), there is a lack of non-asymptotic, uniform-in-time (naUiT) error bounds. Such bounds are crucial because they guarantee that the numerical error does not grow unbounded as $t \to \infty$, ensuring the scheme faithfully captures both transient dynamics and the long-term invariant measure.

2. Methodology

The authors develop a unified framework to prove naUiT error bounds for full space-time discretizations of SPDEs.

A. General Theoretical Framework (Section 2)

The core contribution is a set of sufficient conditions (Theorem 2.1 and 2.3) to establish naUiT bounds of the form:
$$\sup_{k \in \mathbb{N}} \mathbb{E} \| u(k\tau) - u_h(k\tau) \| \leq C(\tau, h) \bar{C}(u_0)$$
where the constant on the RHS is independent of time $k$. The proof relies on three key hypotheses:

1. [H1] Contractivity of Dynamics: The continuous SPDE solution must exhibit exponential contractivity in a suitable norm (e.g., $L^2$ or $H^1$) with respect to initial data.
2. [H2] Finite-Time Error Bound: The scheme must have a standard error bound over a finite time interval (e.g., $[0, 1]$).
3. [H3] Uniform-in-Time Moment Bounds: The numerical scheme must possess moments (e.g., $\mathbb{E}\|u_h\|^\beta$) that are uniformly bounded in time.

The authors leverage a technique from [55] (originally for SDEs) and adapt it to the infinite-dimensional SPDE setting. The logic is that if the scheme is stable (H3) and accurate for short times (H2), and the true dynamics are contractive (H1), the error cannot accumulate over infinite time.

B. Analysis of Specific Schemes (Section 3)

The paper analyzes three specific full space-time discretizations for SPDEs with non-globally Lipschitz drifts:

1. Fully Implicit Euler (FIE): Treats both the Laplacian and the nonlinearity implicitly.
2. Truncated Pointwise Tamed Scheme: A tamed scheme where the drift is divided by a term involving the pointwise derivative of the nonlinearity ($1 + \tau |f'(u)|$).
3. Tamed Scheme by Gradient Norm: A scheme where the drift is divided by a term involving the global $L^2$ norm of the gradient of the nonlinearity ($1 + \alpha \tau \|\nabla f(u)\|^2$).

C. Demonstration of Failure of Semi-Implicit Schemes (Section 5)

The authors rigorously prove that the Semi-Implicit Euler (SIE) scheme (implicit Laplacian, explicit nonlinearity) suffers from finite-time blow-up for non-globally Lipschitz SPDEs when the initial data is large. This is shown via a lower bound on the second moment of the solution, proving it grows linearly with time steps and diverges.

3. Key Contributions

1. General naUiT Framework: They provide a "black-box" criterion (Theorem 2.1) that simplifies the proof of uniform-in-time error bounds. Instead of deriving complex long-time estimates from scratch, one only needs to verify contractivity, short-time accuracy, and moment stability.
2. Proof of Blow-up for SIE: They establish a rigorous theoretical proof that the semi-implicit Euler scheme is unstable for SPDEs with super-linear drifts, confirming a phenomenon previously observed only in finite-dimensional SDEs.
3. naUiT Bounds for New and Old Schemes:
   • They prove naUiT bounds for the Fully Implicit Euler scheme (previously only finite-time bounds were known).
   • They prove naUiT bounds for a Tamed Scheme (specifically the GTEM variant, Eq 3.18) by leveraging existing literature results via their new framework.
   • They prove naUiT bounds for a Gradient-Norm Tamed Scheme (Eq 3.24).
4. Numerical Validation: Extensive numerical experiments compare FEM (Finite Element Method) and Spectral-Galerkin discretizations.

4. Results

• Theoretical Results:

  • The error bound for the proposed schemes is $O(h + \tau^{1/2})$ uniformly in time.
  • The Fully Implicit Euler scheme is proven to be unconditionally stable and satisfy naUiT bounds.
  • The Semi-Implicit Euler scheme is proven to blow up for large initial data.
  • The Tamed schemes (specifically the pointwise version) are shown to be robust and satisfy the required moment bounds.

• Numerical Findings:

  • Blow-up Confirmation: Simulations confirm that SIE fails (blows up) for large initial data ($u_0 = 100$), while tamed and implicit schemes remain stable.
  • Performance Comparison:
    • Fully Implicit Euler (FIE): Most robust and accurate but computationally expensive (requires solving non-linear systems via Newton's method at each step).
    • Truncated Pointwise Tamed Scheme (Eq 3.13): The authors recommend this as the best practical alternative. It offers a balance of robustness (no blow-up), accuracy (faithful to transient and long-time dynamics), and computational efficiency (explicit evaluation, no root-finding required).
    • Gradient-Norm Taming (Eq 3.24): While theoretically sound, numerical experiments on periodic domains (Spectral-Galerkin) showed instability or large errors. The authors attribute this to the lack of a Poincaré inequality in the periodic setting for this specific discretization, which causes the taming term to vanish initially for constant data, leading to transient blow-up before the gradient grows large enough to tame the drift.

5. Significance

• Theoretical Advancement: This work bridges the gap between finite-time error analysis and long-time invariant measure sampling. It provides a rigorous justification for using specific numerical schemes for SPDEs over infinite time horizons, a critical requirement for statistical physics and climate modeling applications.
• Practical Guidance: The paper moves beyond pure theory to offer concrete recommendations. It identifies the Truncated Pointwise Tamed Scheme as the optimal choice for practitioners: it avoids the computational cost of implicit solvers while preventing the catastrophic blow-up of explicit/semi-implicit methods.
• Unification: By adapting the SDE framework of [55] to SPDEs, the authors unify the analysis of various schemes, showing that many existing results in the literature can be easily extended to uniform-in-time bounds with minimal additional effort.

In summary, the paper provides a robust theoretical foundation and practical validation for simulating SPDEs with non-globally Lipschitz nonlinearities, proving that while standard semi-implicit methods fail, carefully designed tamed or fully implicit schemes can faithfully capture dynamics from the transient phase to the invariant measure.