Convergence of tamed Euler schemes for a class of stochastic evolution equations

István Gyöngy, Sotirios Sabanis, David Šiška

Abstract

We prove stability and convergence of a full discretization for a class of stochastic evolution equations with super-linearly growing operators appearing in the drift term. This is done using the recently developed tamed Euler method, which uses a fully explicit time stepping, coupled with a Galerkin scheme for the spatial discretization.

Introduction

Stochastic Evolution Equations (SEEs) are essential for modeling complex systems subject to random perturbations, such as fluid dynamics or signal propagation in biological neurons. A significant challenge in the numerical analysis of these equations arises when the deterministic "drift" term exhibits super-linear growth—for example, cubic terms like $u^3$. In these scenarios, standard numerical tools often fail to provide stable or efficient approximations.

The Numerical Dilemma: Explicit vs. Implicit

The primary difficulty in simulating SEEs with super-linear drift is the trade-off between computational efficiency and numerical stability. The standard explicit Euler scheme is computationally inexpensive but is known to diverge to infinity in finite time when applied to operators with super-linear growth. To maintain stability, researchers traditionally turn to implicit schemes. While stable, implicit methods require solving a complex non-linear equation at every single time step, which is computationally expensive, particularly for high-dimensional problems.

This research addresses whether an explicit method can be designed to be as fast as an explicit scheme but as stable as an implicit one for these difficult equations.

The Tamed Euler Approach

The authors propose a discretization strategy that combines two key techniques:

  1. Spatial Discretization (Galerkin Method): The infinite-dimensional problem is projected onto a finite-dimensional subspace ($V_m$).
  2. Time Discretization (Tamed Euler): The core innovation involves "taming" the super-linear operator $A_2$. Instead of using the raw operator, they apply a taming factor: $$A_{2,\ell}v := \frac{A_2v}{1 + n_\ell^{-1/2}|\Pi_{m_\ell} A_2v|}$$ Where $n$ is the number of time steps. When the operator $A_2$ becomes very large—threatening a numerical explosion—the denominator grows, effectively scaling down the step to keep it within stable bounds. As the number of time steps $n$ goes to infinity, this taming factor vanishes, recovering the original equation.

This framework relies on the Gelfand Triple ($V \hookrightarrow H \hookrightarrow V^*$) and uses monotonicity and coercivity arguments to prove that the numerical approximations converge to the unique solution of the SEE.

Theoretical Insights and Key Ideas

  • Preserving Coercivity: While taming prevents numerical explosions, it can potentially destroy the mathematical properties (like monotonicity) required to prove convergence. The authors demonstrate that a specific form of "degenerate" coercivity is maintained, which is sufficient for convergence.
  • Space-Time Coupling: A critical insight is the relationship between spatial resolution ($m$) and temporal steps ($n$). For the scheme to converge, the authors identify specific constraints (e.g., $m^2/n \to 0$ for certain spectral methods). Notably, they find this requirement is no more onerous than what is required for simpler equations with linear growth.
  • Existence via Discretization: The paper does more than prove the scheme works; it uses the convergence of the scheme to prove that a unique solution to this specific class of SEEs exists. This result was previously unproven for these specific "degenerate" coercivity conditions.

Key Results and Applications

The authors demonstrate that their tamed scheme converges for several high-profile stochastic partial differential equations (SPDEs):

  • Stochastic Ginzburg–Landau Equation: Used in physics to describe phase transitions.
  • Stochastic Swift–Hohenberg Equation: Used to model pattern formation in fluids and chemicals.
  • Stochastic FitzHugh–Nagumo System: A famous model for signal propagation in nerve cells.

The primary quantitative result is Theorem 2.3, which rigorously proves that as the number of time steps and spatial dimensions go to infinity, the numerical approximation $u_\ell$ converges to the true solution $u$ in the appropriate mathematical spaces ($L^2$ and $L^p$).

Limitations and Future Work

  • Coupling Constraints: Convergence depends on a specific relationship between the time step size and the spatial discretization. If the spatial grid is refined too quickly relative to the time step, the scheme may still fail.
  • Specific Growth Rates: The proofs are tailored for operators satisfying certain polynomial growth conditions. Extending this to more extreme non-linearities or different noise types (e.g., jump processes) remains a path for future research.
  • Rate of Convergence: While the paper proves the scheme does converge, it does not provide a specific "rate" (e.g., the error reduction per doubling of time steps). Determining the optimal convergence rate is a logical next step for the field.

Summary updated

Abstract

We prove stability and convergence of a full discretization for a class of stochastic evolution equations with super-linearly growing operators appearing in the drift term. This is done using the recently developed tamed Euler method, which uses a fully explicit time stepping, coupled with a Galerkin scheme for the spatial discretization.
BibTeX
@inproceedings{gyongy2015convergence,
  author = {István Gyöngy and Sotirios Sabanis and David Šiška},
  title = {{Convergence of tamed Euler schemes for a class of stochastic evolution equations}},
  booktitle = {Stochastic},
  year = {2015},
  doi = {10.1007/s40072-015-0057-7},
}
Figure