PaperlayerView exact revision on arXiv

Source-linked AI summary

Higher-Order QMC Galerkin FEM for Singular Perturbation: Reaction-Diffusion with Affine-Parametric Reaction

Fabian Rohner, Christoph Schwab

arXiv:2610.39734v1math.NAcs.NA

TL;DR

The paper studies QMC–FE approximation of linear quantities of interest for a singularly perturbed reaction–diffusion problem with an affine-parametric reaction coefficient. It develops ε-uniform stability and parametric regularity estimates and combines them with higher-order QMC and robust spatial approximation. The resulting analysis gives dimension-independent bounds, with a one-dimensional Shishkin-mesh example and numerical experiments illustrating the predicted robustness and rates.

  • Problem

    Boundary layers obstruct standard finite element bounds, and existing QMC analyses lack uniform stability, norm-equivalence, and adjoint-regularity bounds as ε approaches zero.

    Source: §1 · p. 2

  • Method

    The paper combines ε-weighted stability and mixed-parametric regularity estimates with higher-order QMC and an abstract ε-robust Galerkin finite element approximation property.

    Source: §1 · p. 3

  • Results

    The analysis yields a fully discrete QMC–FE error bound uniform in ε, truncation dimension s, QMC point count N, and spatial discretization parameter h.

    Source: §6 · pp. 16–17

  • Takeaways & Limitations

    A single QMC point set can be used for all ε ∈ (0, 1], and the abstract spatial approximation property is verified for a one-dimensional Shishkin-mesh model.

    Source: §1 · p. 3

    The analysis treats single-level QMC–FE methods and affine-parametric reaction coefficients; multilevel and non-affine extensions are left for future work.

    Source: §8 · p. 21

Abstract

from arXiv · show

We analyze higher-order quasi-Monte Carlo (QMC) Galerkin discretizations for a singularly perturbed reaction-diffusion problem with an affine-parametric reaction coefficient. We derive error estimates that are robust with respect to the perturbation parameter $\varepsilon$. To this end, we establish $\varepsilon$-uniform stability and mixed-parametric regularity estimates in an $\varepsilon$-weighted energy norm, with constants independent of the parameter dimension and of $\varepsilon$. These estimates imply robust dimension truncation estimates and dimension-independent higher-order QMC error bounds for linear quantities of interest. Assuming an abstract, $\varepsilon$-robust finite element approximation property, we combine the parametric estimates with the spatial discretization analysis to derive a general fully discrete QMC-FE error estimate with constants independent of both the parameter dimension and $\varepsilon$. We verify this approximation property for a one-dimensional model problem using continuous piecewise affine finite elements on Shishkin meshes. Numerical experiments illustrate the predicted robustness and convergence rates.

1 Introduction

The paper addresses the lack of ε-robust QMC–FE analysis for reaction–diffusion problems with uncertain inputs, where boundary layers obstruct standard regularity and finite element bounds. It develops uniform parametric estimates and combines them with higher-order QMC and robust spatial approximation to obtain a fully discrete error bound.

Motivation

  • Boundary layers obstruct standard finite element bounds, while existing QMC analyses lack uniform bounds for stability, norm equivalences, and adjoint regularity as ε approaches zero.The paper identifies ε-robust spatial and parametric regularity as a central obstacle.Source: §1 · p. 2

Problem Setting

  • The model is a linear elliptic reaction–diffusion problem on a bounded domain, with diffusion scaled by ε^2 and an affine-parametric reaction coefficient.The reaction coefficient is assumed uniformly bounded above and bounded away from zero.Source: §1 · p. 2
  • The target quantity is the expectation of a bounded linear functional of the solution under the uniform parameter measure, approximated by truncating the coefficient expansion and applying QMC quadrature.The QMC points are chosen using a higher-order lattice rule.Source: §1 · pp. 2–3

Assumptions

  • The analysis assumes summability of the coefficient fluctuation terms and sufficient spatial smoothness to support ε-robust finite element approximation.The introduction strengthens the coefficient assumptions with p-summability for some 0 < p < 1.Source: §1 · p. 3

Contributions

  • The main contribution is an ε-robust QMC–FE error analysis for a singularly perturbed reaction–diffusion problem with an uncertain reaction coefficient.The analysis combines ε-uniform stability and mixed-parametric regularity with higher-order QMC and a robust spatial approximation property.Source: §1 · p. 3
  • The QMC weights and point sets can be independent of ε, and the fully discrete error bound is uniform in ε, truncation dimension, QMC point count, and spatial discretization.The spatial approximation property is verified for a one-dimensional model using continuous piecewise affine finite elements on Shishkin meshes.Source: §1 · p. 3

2 Robust parametric regularity

The section establishes well-posedness and ε-uniform stability in a weighted energy norm, then derives mixed-parametric regularity for affine reaction coefficients. The estimates cover ε in (0, 1], with the zero-ε limiting equation treated separately.

2.2 Parametric regularity for affine reaction coefficients

  • For affine reaction coefficients, the parameter-to-solution map has ε-robust mixed-derivative bounds in the weighted energy norm.The proof obtains higher mixed derivatives by iterating the variational argument.Source: §2 · p. 4, §2.2 · pp. 6–7

2.1 Parametric weak formulation

  • The weighted-norm variational problem is uniformly coercive and continuous, yielding a unique solution with an ε-uniform stability bound.For L2 forcing, the embedding into the dual space preserves a uniform bound.Source: §2.1 · p. 5

2.2 Parametric regularity for affine reaction coefficients

  • The derivative construction uses affine perturbations in each parameter coordinate, with perturbations restricted so the reaction coefficient remains positive.At parameter-domain boundaries, derivatives are interpreted through an affine extension that preserves uniform positivity for sufficiently small perturbations.Source: §2.2 · p. 6
  • At ε = 0, the limiting equation is algebraic in the solution, and the same derivative bound follows from the positive lower bound on the reaction coefficient.The limiting solution is given by u0(x; y) = f(x)/b(x; y).Source: §2.2 · pp. 8–9

3 Dimension truncation

This section applies a smooth-function dimension-truncation result to the parametric solution and linear quantities of interest. Under the stated p-summability assumptions, the resulting bounds are independent of truncation dimension and ε.

Dimension truncation

  • The truncation analysis uses a Taylor-expansion result based on parametric regularity, rather than a bound that depends on the affine structure alone.The authors choose this approach with possible nonlinear parameter dependence in view.Source: §3 · p. 9
  • The solution map is truncated by retaining the first s parameters and setting the remaining parameters to zero.The general result assumes a non-increasing coefficient bound sequence in ℓp for some 0 < p < 1.Source: §3 · p. 9
  • The resulting truncation estimate for the solution has a constant independent of both s and ε under assumptions A.1–A.5.The derivation applies the abstract result to the solution in the ε-weighted space and uses the uniform derivative bounds.Source: §3 · pp. 9–10
  • The dimension-truncation result also applies to bounded linear functionals of the solution.The same abstract proposition is applied to the functional in the dual space.Source: §3 · p. 10

4 Higher-order QMC

The section sets up higher-order QMC quadrature for integrals over a finite-dimensional parameter cube, including those arising from linear quantities of interest. It uses weighted higher-order Sobolev regularity assumptions to obtain the dimension-independent rate used later.

4 Higher-order QMC

  • The QMC approximation uses deterministic lattice-rule points in the s-dimensional unit cube, shifted to the parameter domain for the model problem.The integrals include evaluations of linear functionals of the parametric solution.Source: §4 · p. 10

4.1 Weighted spaces and SPOD weights

  • The analysis uses higher-order unanchored Sobolev spaces with weights to measure the integrand’s mixed derivatives.The paper uses the case r = ∞.Source: §4.1 · p. 11
  • The integrands are assumed to satisfy factorial derivative bounds controlled by a non-increasing ℓp sequence, for some 0 < p < 1.These bounds provide the regularity setting for the subsequent higher-order QMC error result.Source: §4.1 · p. 11

SPOD weights

The section selects SPOD weights to place derivative-controlled integrands in weighted function spaces and support efficient lattice construction. Interlaced polynomial lattice rules with these weights yield a QMC rate independent of dimension.

SPOD weights

  • SPOD weights are chosen so derivative bounds place the integrand in the corresponding weighted space while supporting efficient lattice-rule construction.The weights are smoothness-driven, product- and order-dependent weights.Source: SPOD weights · pp. 11–12

4.2 Interlaced polynomial lattice rules

  • Digit interlacing maps blocks of α coordinates into points for an order-α rule in s dimensions.The construction starts from a polynomial lattice point set in αs dimensions.Source: §4.2 · p. 12
  • Interlaced polynomial lattice rules can be constructed by a component-by-component algorithm, with a worst-case error bound in the weighted function space.The construction uses N = bm points for prime b.Source: §4.2 · pp. 12–13
  • Combining the lattice-rule result with SPOD weights gives a QMC convergence rate whose constant is independent of dimension s and point count N.The integrand-specific dependence enters through its partial-derivative bounds.Source: §4.2 · p. 13

5 Spatial approximation by Galerkin methods

The spatial analysis isolates an approximation property whose rate and constant must remain uniform in the perturbation parameter and parametric coefficient. It verifies that property for one-dimensional piecewise-linear finite elements on Shishkin meshes, including an improved L2 rate.

5 Spatial approximation by Galerkin methods

  • The abstract spatial assumption requires a Galerkin approximation rate κ(h) → 0 with a constant independent of ε, y, and h.Its validity depends on domain shape, right-hand-side regularity, and reaction-coefficient smoothness.Source: §5 · p. 13
  • Standard Galerkin methods on quasi-uniform meshes generally lack ε-uniform error estimates because they inadequately resolve boundary and corner layers.Layer-adapted Shishkin or Bakhvalov meshes and, under analyticity assumptions, hp-FE methods can provide robust estimates.Source: §5 · p. 14

5.1 A one-dimensional model problem

  • For the one-dimensional model, the cited robust FE estimates depend on the reaction coefficient only through bounds b0 and b1, yielding error estimates uniform in y.The model assumes b and f belong to H2(I).Source: §5.1 · p. 14

5.2 Robust finite element approximation

  • Galerkin orthogonality and coercivity give a best-approximation estimate with a constant depending only on b0 and b1.The discretization uses continuous piecewise-linear functions on a partition of I.Source: §5.2 · p. 14

5.3 Shishkin meshes

  • On Shishkin meshes, continuous piecewise-linear FE approximations satisfy the stated ε- and mesh-independent bound for 4 ≤ Mh ≤ 1/ε; with H2 data, it also holds for Mh > 1/ε.The mesh allocates Mh/4, Mh/2, and Mh/4 intervals to the left layer, interior, and right layer regions.Source: §5.3 · p. 15

5.3.1 Improved error bounds in the L2-norm

  • The improved L2 estimate avoids relying on a standard Aubin–Nitsche argument, whose needed dual-problem layer decomposition is not readily available.The resulting improved rate is the spatial approximation rate used in the one-dimensional fully discrete QMC-FE estimate.Source: §5.3.1 · pp. 15–16

6 Combined error analysis

The combined analysis brings dimension truncation, QMC quadrature, and spatial approximation into a fully discrete error bound. Its constants are independent of ε, parameter dimension, point count, and spatial discretization; the one-dimensional Shishkin result supplies a concrete specialization.

6 Combined error analysis

  • The fully discrete QMC-FE error bound has a constant independent of ε, s, N, and h.It combines dimension truncation, QMC quadrature, and spatial discretization estimates.Source: §6 · pp. 16–17
  • For the one-dimensional model, the QMC-FE corollary uses an interlaced polynomial lattice rule and a continuous piecewise-linear FE approximation on a Shishkin mesh.Its error-bound constant is independent of ε, s, N, and h.Source: §6 · p. 17
  • Evaluating the QMC-FE approximation costs O(sNMh) operations; balancing s, N, and Mh to target error O(δ) gives the stated overall cost order.The cost expression's order is given in the cited passage.Source: §6 · p. 17

7 Numerical experiments

The experiments examine finite-element, dimension-truncation, and QMC–FE convergence for several perturbation parameters. They report similar convergence rates across ε values, with one QMC point set serving all values.

7 Numerical experiments

  • The numerical example combines P1 finite elements on Shishkin meshes with interlaced polynomial lattice QMC and fluctuation amplitudes decaying as j^-ζ, with ζ = 21/8.The QMC points use SPOD weights and are generated using a fast CBC algorithm.Source: §7 · p. 18

7.1 FE convergence

  • The Shishkin-mesh FE study tracks L2, 1, ε, and balanced-norm errors for four ε values against reference convergence rates.The study addresses the lower regularity of its forcing term relative to assumptions common in the literature.Source: §7.1 · p. 18, Figure 1 · p. 19

7.2 Dimension truncation convergence

  • Dimension-truncation errors decay at O(s^-17/4) after the first two data points, matching the expected rate.The truncation experiment fixes QMC and finite-element discretization errors to be negligible.Source: §7.2 · p. 19, Figure 2 · p. 20

7.3 QMC–FE convergence

  • QMC–FE convergence is faster than the rate predicted by (4.4), while essentially the same rates are observed across ε values and one QMC point set serves all values.The faster-rate explanation uses a fixed truncation dimension and entails a constant depending on that dimension.Source: §7 · p. 18, §7.3 · p. 20

8 Conclusion

The conclusion presents QMC–FE error analysis for a linear singularly perturbed reaction–diffusion model with affine-parametric uncertain reaction. Its estimates give algebraic fully discrete convergence under a parameter-robust Galerkin FE assumption, while identifying several scope boundaries.

Conclusion

  • Under a parameter-robust Galerkin FE assumption, the fully discrete method has algebraic convergence rates uniform in the uncertain parameter sequence and robust for ε ∈ (0, 1].The analysis establishes solution parametric regularity with explicit dependence on both the parameter sequence and ε.Source: §8 · pp. 20–21
  • The analysis is developed for a linear scalar model problem to verify ε-independence of its error-analysis constants.This states the scope chosen to develop the framework and check robustness.Source: §8 · p. 21
  • Only single-level QMC–FE methods are analyzed; robust multilevel bounds would require refining the ε-explicit regularity analysis.Multilevel discretizations are proposed as a possible extension.Source: §8 · p. 21
  • The error analysis treats PDE regularity and FE approximation error only in an ε-weighted energy norm; balanced-norm details are deferred.The conclusion notes that uniform-in-y and ε-robust bounds in balanced norms also hold, with details to be reported elsewhere.Source: §8 · p. 21
  • The analysis assumes affine-parametric reaction coefficients; non-affine, including Gevrey-regular, extensions are left for future work.The paper identifies these as further generalizations.Source: §8 · p. 21
Loading 2610.39734v1…