Source-linked AI summary
Optimal Error Estimates of Galerkin Finite Element Methods for Stochastic Partial Differential Equations with Multiplicative Noise
Raphael Kruse
TL;DR
Strong-error estimates support pathwise approximation and efficient multilevel Monte Carlo, while prior semilinear-SPDE rates were suboptimal or included logarithmic terms. The analysis combines Galerkin finite-element methods, spatial and temporal regularity results, and error estimates for the deterministic homogeneous equation. For a standard finite-element semidiscretization, approximation Xh converges to the mild solution X with order 1 + r, which is optimal because it matches X’s spatial regularity.
Problem
Strong-error estimates support pathwise approximation and efficient multilevel Monte Carlo, while prior semilinear-SPDE rates were suboptimal or included logarithmic terms.
Source: §1 · p. 3Method
The analysis combines Galerkin finite-element methods, spatial and temporal regularity results, and error estimates for the deterministic homogeneous equation.
Source: §1 · pp. 1, 4Results
For a standard finite-element semidiscretization, approximation Xh converges to the mild solution X with order 1 + r, which is optimal because it matches X’s spatial regularity.
Source: §1 · p. 3Takeaways & Limitations
The error analysis applies in a unified setting to standard finite-element and spectral Galerkin methods.
Source: §1 · p. 1The analysis discretizes time and H but not U, and assumes a regularity parameter r∈[0,1).
Source: §1 · pp. 2–3
Abstract
from arXiv · showhide
We consider Galerkin finite element methods for semilinear stochastic partial differential equations (SPDEs) with multiplicative noise and Lipschitz continuous nonlinearities. We analyze the strong error of convergence for spatially semidiscrete approximations as well as a spatio-temporal discretization which is based on a linear implicit Euler-Maruyama method. In both cases we obtain optimal error estimates. The proofs are based on sharp integral versions of well-known error estimates for the corresponding deterministic linear homogeneous equation together with optimal regularity results for the mild solution of the SPDE. The results hold for different Galerkin methods such as the standard finite element method or spectral Galerkin approximations.
1. Introduction
The introduction analyzes strong-error convergence for Galerkin approximations of multiplicative-noise SPDEs, establishing optimal spatial and spatio-temporal estimates under stated regularity assumptions. Its discretization covers time and the solution space H, but not the noise space U.
1. Introduction
- The analysis combines Galerkin finite-element theory with optimal spatial and temporal regularity results, covering finite-element and spectral Galerkin methods in one framework.The finite-dimensional spaces may consist of spatially regular functions, including standard finite-element spaces or spans of eigenfunctions.Source: §1 · pp. 1–2
- The paper discretizes time and H but leaves U undiscretized, deferring a fully discrete approximation that includes U.Both H and U may be high- or infinite-dimensional.Source: §1 · p. 2
- Strong convergence measures pathwise approximation and is also essential for efficient multilevel Monte Carlo when weak approximation is of interest.The paper distinguishes strong convergence from weak approximation of expectations of smooth observables.Source: §1 · p. 3
- For standard finite elements, the spatial semidiscretization converges with order 1+r, matching the mild solution's spatial regularity.Here r∈[0,1) controls solution regularity, while h denotes the mesh's maximal edge length in the finite-element example.Source: §1 · p. 3
- Earlier semilinear-SPDE estimates were, to the authors' knowledge, suboptimal by ǫ or included a logarithmic error term.The cited examples either imposed stronger Lipschitz assumptions or included a log(t/h) term.Source: §1 · p. 3
- The spatio-temporal approximation also has optimal convergence order, with implicit Euler's maximum order of 1/2 because it uses only Wiener increments.A Milstein-like scheme can overcome this barrier [19].Source: §1 · p. 4
- Sharp integral error estimates for the deterministic homogeneous equation, extended to nonsmooth initial data, support the main proofs.These estimates address both spatially semidiscrete and fully discrete approximations.Source: §1 · p. 4
2. Preliminaries
The preliminaries state assumptions on the drift, noise, and initial data that ensure a unique mild solution and provide regularity results used in the error analysis. They also collect analytic-semigroup estimates used in the proofs.
Solution assumptions and regularity
- The assumptions on f, g, and X0 yield a unique mild solution, with regularity results available for s in [0, r + 1] when r is in [0, 1).The initial value is required to be F0-measurable and valued in the stated fractional space.Source: §2 · p. 5
Nonlinearity assumptions
- Assumption 2.1 is slightly weaker than requiring f to be Lipschitz from H to H, while retaining the same numerical order for the spatial errors in the drift and noise integrals.The assumption also covers equations with a fractional power of A before a Lipschitz map, such as the Cahn-Hilliard equation.Source: §2 · p. 6
Initial-data condition
- Relaxing the initial-data assumption to an F0-measurable random variable with finite pth moment is possible, but leads to a singularity at t = 0 in the error estimates.Source: §2 · p. 6
Semigroup estimates
- The analytic semigroup E(t) is accompanied by estimates for its smoothing and integral behavior that are used in the analysis.The cited passage identifies the smoothing property and Lemma 2.4(ii) as classical results.Source: §2 · p. 6
3. A short review of Galerkin finite element methods
The section sets out Galerkin spaces, approximation assumptions, and associated discrete operators. It illustrates the framework with standard finite element and spectral Galerkin methods.
Approximation framework
- The spatial approximation spaces must satisfy an error bound for the Ritz projection, of order h^s for s in {1, 2}.The two examples presented are the standard finite element method and the spectral Galerkin method.Source: §3 · pp. 6–7
Ritz projection
- The Ritz projection coincides with the Galerkin approximation of the associated elliptic problem, making Assumption 3.1 an order-of-convergence condition for that approximation.Source: §3 · p. 7
Discrete operator
- The discrete operator Ah is self-adjoint and positive definite, and generates an analytic semigroup of contractions with a smoothing property.Ah is defined through the discrete bilinear form a on Sh.Source: §3 · p. 7
Standard finite element method
- For the standard finite element example, Sh consists of continuous piecewise-linear functions on a regular simplex partition, with zero boundary values; quasi-uniform meshes also satisfy Assumption 3.2.The stated setting is a bounded convex domain in dimensions 1, 2, or 3, with Dirichlet boundary conditions.Source: §3 · p. 8
Spectral Galerkin method
- For the spectral Galerkin example on (0, 1), Sh is built from Laplacian eigenfunctions, and the Ritz projection is the restriction of the orthogonal L2 projector.The passage states the spectral approximation property for ρ = 1, 2.Source: §3 · pp. 8–9
4. Error estimates of Galerkin methods for deterministic equations
This section develops deterministic Galerkin error estimates, extending homogeneous-equation bounds to nonsmooth initial data and establishing integral estimates used to derive SPDE error bounds. It also completes a fully discrete approximation lemma by estimating its two summands for ρ = 0 and ρ = 1.
4. Error estimates of Galerkin methods for deterministic equations
- The section extends estimates from to nonsmooth initial data x ∈ ˙H−1 and presents integral versions crucial for optimal error estimates.These estimates concern discretizations of the deterministic linear homogeneous equation.Source: §4 · p. 9
4.1. Error estimates for a spatially semidiscrete approximation.
- Under Assumption 3.1, Lemmas 4.1–4.2 provide error-operator estimates for the spatially semidiscrete approximation.The operator Fh represents the difference between the exact and spatially semidiscrete solutions.Source: §4.1 · pp. 9–10
4.2. Error estimates for a fully discrete approximation.
- For the fully discrete approximation, implicit Euler is unconditionally stable and has discrete smoothing, while Lemmas 4.3–4.4 estimate Fkh; Lemma 4.4(ii) requires Assumption 3.2.Source: §4.2 · pp. 11, 12, 13, 15
- Parseval’s identity and a split at kλh,m > 1 are used to bound terms in the fully discrete approximation proof, with a bound independent of n.For modes above the threshold, the proof applies (4.17).Source: §4.2 · p. 19
- For the second summand, Parseval’s identity and 1 − e^−kλh,m ≤ kλh,m yield a bound by 1.The proof then combines this estimate with (4.20)–(4.23) to complete the case ρ = 0.Source: §4.2 · p. 20
- For ρ = 1, the proof additionally uses (3.6) and Assumption 3.2, completing the lemma for all x ∈ ˙H1.Source: §4.2 · p. 20
5. Proof of Theorem 1.1
The proof targets strong convergence of the spatially semidiscrete approximation by decomposing the mild-solution error and estimating its deterministic and stochastic integral terms. It uses operator error bounds, a stochastic-integral estimate, and a generalized Gronwall lemma.
Proof of Theorem 1.1
- The proof targets strong convergence of the spatially semidiscrete approximation to the mild solution.It begins by expressing their difference using the mild-solution representations.Source: §5 · p. 21
Error estimates
- The error decomposition separates deterministic and stochastic integral terms, which are estimated using bounds for the operator Fh and the discrete semigroup.The deterministic estimates invoke Lemma 4.1 and Lemma 4.2; the stochastic terms use Lemma 5.1 and Assumption 2.2.Source: §5 · pp. 21–23
Regularity parameter
- The generic constant in one estimate depends on r and blows up as r approaches 1.Source: §5 · p. 23
Proof completion
- A generalized Gronwall lemma completes the proof after the separate integral estimates are combined.Source: §5 · pp. 21, 23, 24
6. Proof of Theorem 1.2
The proof of Theorem 1.2 uses a discrete variation-of-constants formula to decompose the error, then bounds the terms and applies a discrete Gronwall lemma. The stochastic-integral term is handled using predictability and stochastic-integral estimates.
6. Proof of Theorem 1.2
- The discrete variation-of-constants formula decomposes the error into terms for the homogeneous-equation error and the remaining contributions.The proof introduces an auxiliary process for the latter contributions.Source: §6 · pp. 24–25
- Predictability of the integrand permits the stochastic-integral estimate used to bound the error contribution.The process is described as adapted and left-continuous, hence predictable.Source: §6 · pp. 25–27
- A discrete Gronwall lemma completes the estimate, whose constant does not depend on the time step k.The proof states that the discrete version of Gronwall’s lemma is the final step.Source: §6 · pp. 24, 28
7. Additive noise
For additive noise, the paper gives error estimates for spatially semidiscrete and fully discrete approximations. The convergence results extend to r = 1, and the comparison with [21] notes removal of a logarithmic singularity.
7. Additive noise
- Under Assumptions 2.1, 2.3, and 7.1, for r ∈ [0, 1] and p ∈ [2, ∞), a mild solution exists in H˙^(1+r); when r = 0, Tr(Q) < ∞.Source: §7 · pp. 28–29
- The additive-noise results give error estimates for spatially semidiscrete and fully discrete approximations, with constants independent of h and of k and h, respectively.The corresponding approximations are Xh and the spatio-temporal discrete approximation.Source: §7 · p. 29
- Both additive-noise convergence results extend to r = 1 because the term I5 vanishes, and the restriction r < 1 is not needed.For the fully discrete result, the restriction appears only in estimate (6.11).Source: §7 · p. 29
- A similar convergence result appears in [21, Prop. 2.3], but that result incorporates a singularity of the form max(0, log(t/h²)), which is now removed.Source: §7 · p. 29