PaperlayerView exact revision on arXiv

Source-linked AI summary

Eliminating the pollution effect in Helmholtz problems by local subscale correction

Daniel Peterseim

arXiv:1411.7512v3math.NAcs.NA

TL;DR

Large-wave-number Helmholtz discretizations suffer pollution, with mesh requirements stricter than the basic representation of oscillatory solutions. The paper builds a Petrov-Galerkin multiscale method using local subscale corrections. Under resolution and oversampling conditions, it suppresses pollution and is stable.

  • Problem

    For large wave numbers, pollution can require meshes much finer than the minimum needed to represent highly oscillatory solutions.

    Source: §1 · p. 1

  • Method

    The method constructs Petrov-Galerkin trial and test spaces from standard finite elements using local subscale corrections computed through coercive cell problems.

    Source: §1 · p. 2

  • Results

    With Hκ ≲ 1 and logarithmic-in-κ oversampling, the method is stable and suppresses pollution for a fairly large class of Helmholtz problems.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Pollution can be eliminated in the stated regime while retaining quasi-minimal resolution at the cost of logarithmically enlarged inter-element communication.

    Source: §1 · p. 2

    In most settings, the required fine-scale resolution is not quantified, and stability of the practical method is shown assuming that scale is sufficiently fine.

    Source: §6.2 · p. 17

Abstract

from arXiv · show

We introduce a new Petrov-Galerkin multiscale method for the numerical approximation of the Helmholtz equation with large wave number $\kappa$ in bounded domains in $\mathbb{R}^d$. The discrete trial and test spaces are generated from standard mesh-based finite elements by local subscale corrections in the spirit of numerical homogenization. The precomputation of the corrections involves the solution of coercive cell problems on localized subdomains of size $\ell H$; $H$ being the mesh size and $\ell$ being the oversampling parameter. If the mesh size and the oversampling parameter are such that $H\kappa$ and $\log(\kappa)/\ell$ fall below some generic constants and if the cell problems are solved sufficiently accurate on some finer scale of discretization, then the method is stable and its error is proportional to $H$; pollution effects are eliminated in this regime.

1. Introduction

Large-wave-number Helmholtz problems face pollution that can demand meshes finer than needed merely to represent oscillations. The paper adapts numerical homogenization into a Petrov-Galerkin method whose local subscale corrections suppress pollution under a quasi-minimal resolution condition, with logarithmic oversampling.

1. Introduction

  • Pollution can impose mesh requirements much stricter than the 5–10 degrees of freedom per wavelength considered minimally necessary to represent oscillations.For less regular solutions, prior pollution-free results were not directly applicable, leaving the existence of a pollution-free scheme open.Source: §1 · p. 1
  • The method adapts numerical homogenization, which approximates problems on coarse meshes that do not resolve coefficient oscillations, using locally supported multiscale basis functions.The approach draws on local orthogonal decompositions into a coarse space and a high-dimensional remainder.Source: §1 · pp. 1–2
  • The Petrov-Galerkin trial and test spaces arise from standard finite elements through local subscale corrections computed by coercive cell problems on subdomains of size ℓH.The precomputation involves O(H^-d) cell problems; H is the mesh size and ℓ the oversampling parameter.Source: §1 · p. 2
  • With Hκ ≲ 1 and logarithmic-in-κ oversampling, the method is stable and suppresses pollution for a fairly large class of Helmholtz problems.The result includes acoustic scattering from convex non-smooth objects; the oversampling enlarges inter-element communication, with complexity overhead comparable to the cited higher-order approach [MS10, MS11].Source: §1 · p. 2
  • The worst-case serial cost can exceed a global fine-scale simulation, although local precomputation can be reused across many force terms or incident directions.Such reuse is relevant to parameter studies, coupled problems, and inverse problems.Source: §1 · p. 2
  • Model-order reduction for unstructured meshes and extension to heterogeneous media remain future prospects rather than analyzed capabilities.The paper says the latter case is not yet covered, though prior work and its analysis indicate potential for treating coefficient oscillations or jumps together with pollution.Source: §1 · pp. 2–3

2. Model Helmholtz problem

The paper formulates a Helmholtz boundary-value problem on a bounded domain with mixed boundary components and recalls conditions for well-posedness. Its quantitative stability analysis assumes polynomial growth of the solution-operator stability constant, an assumption that is not general.

2. Model Helmholtz problem

  • The model is posed on a bounded Lipschitz domain in dimensions 1–3 with mixed Dirichlet, Neumann, and Robin boundary components, and homogeneous boundary data.The Robin component is assumed to have positive surface measure.Source: §2 · p. 3
  • The variational problem is uniquely solvable, with an a priori bound for its solution.The impedance boundary condition ensures well-posedness; the sesquilinear form also satisfies an inf-sup condition for the primal and adjoint problems.Source: §2 · pp. 4–5
  • The error analysis assumes the stability constant grows at most polynomially with κ, a condition that does not hold for all domains.Trapping domains can exhibit at least exponential growth; admissible examples include Robin conditions on Lipschitz domains and certain star-shaped scattering configurations.Source: §2 · p. 4

3. Standard finite elements and quasi-interpolation

The method is built on mesh-based finite-element spaces and a local quasi-interpolation framework. Coarse patches localize subsequent constructions, while interpolation and its local right inverse provide key structural tools.

3.1. Meshes, patches and spaces.

  • The framework uses regular simplicial meshes and standard conforming P1 spaces, with nodal basis functions associated with vertices outside the Dirichlet boundary.The paper focuses on triangles or tetrahedra but notes that higher-order elements and other mesh types are possible; the basis has a partition-of-unity property.Source: §3.1 · p. 5, §3 · p. 5
  • Element patches grow by adding successive layers of neighboring elements and have finite overlap for fixed patch order.This patch structure supplies localized coarse regions for the method's constructions and analysis.Source: §3.1 · p. 5

3.2. Quasi-interpolation.

  • A bounded Clément-type quasi-interpolation operator maps V to the finite-element space, and a bounded local right inverse is available even though the inverse restricted to the finite-element space is nonlocal.The interpolation choice is not unique and can significantly affect practical performance; one alternative uses local L2 projections.Source: §3.2 · p. 6

4. Global wave number adapted approximation

The section builds non-polynomial, mesh-based approximation spaces from standard finite elements using subscale corrections, then analyzes an ideal Petrov–Galerkin method and its localizable correctors. Under a quasi-minimal resolution condition, the ideal method is stable and accurate, and its element correctors decay exponentially.

4. Global wave number adapted approximation

  • Corrected basis functions are associated with mesh vertices and have quasi-local support, with moduli that decay rapidly away from their elements.The resolution condition is quasi-minimal: it requires a certain number of degrees of freedom per wavelength for meaningful approximation of highly oscillatory waves.Source: §4 · pp. 6–7

4.1. An ideal method.

  • The ideal method splits the space into corrected trial and test spaces plus the remainder W, using correctors defined by variational problems on W.The ideal method is not used for actual computations because its corrected basis functions generally have global support and its corrector problems are infinite-dimensional; localized correctors yield feasible sparse nearby spaces.Source: §4.1 · pp. 7–9

4.2. Stability and accuracy of the ideal method.

  • When Hκ ≲ 1, the ideal method is stable and accurate independently of κ, and its error is quasi-optimal in the corrected trial space.The corrected solution also gives the L2-best approximation in the standard finite element space.Source: §4.2 · pp. 9–10

4.3. Exponential decay of element correctors.

  • Under the resolution condition, element correctors decay exponentially outside their elements, with decay constants independent of H and κ.The paper reports decay rates between 1/3 and 2/3 in numerical experiments, while noting that the theoretical bound may be pessimistic.Source: §4.3 · pp. 10–11

5. Localized approximation

The section localizes corrector problems to element patches and uses them to form localized trial and test spaces. Under resolution and logarithmic-oversampling conditions, the localized method is stable and has a controlled error.

5.1. Localized correctors.

  • Exponential decay of element correctors motivates approximating them on patches ΩT,ℓ, and the section establishes an error bound for this localization.The corrector problems are localized to patches using the oversampling parameter ℓ.Source: §5 · p. 12, §5.1 · p. 12

5.2. Localized trial and test spaces.

  • The localized trial space is the image of the classical finite element space under 1 − Cℓ, with a corresponding localized test space used in the Petrov-Galerkin method.The section also notes that on uniform meshes, translated or rotated equivalent patches have equivalent correctors, reducing the number of distinct cell problems.Source: §5.1 · p. 12, §5.2 · p. 14

5.3. Stability of the localized method.

  • Under the same mesh and oversampling conditions, the localized trial and test spaces pair stably, ensuring a unique discrete solution for any f ∈ V′.The localized pairing is exponentially close to the ideal pairing.Source: §5.3 · p. 14
  • When Hκ is sufficiently small and ℓ is at least of order log κ, the localized Petrov-Galerkin method satisfies the stated error estimate.The supplied theorem states these conditions but does not include the estimate's formula.Source: §5.3 · p. 15

6. Fully discrete localized approximation

The fully discrete method approximates localized correctors with finer-scale finite elements and uses them to build modified coarse basis functions. Its error analysis is relative to a fine-scale reference solution, with a separate boundary on what is quantified for general settings.

6.1. The fully discrete method.

  • Fine-scale cell problems produce correctors that modify the coarse nodal basis, yielding a discrete trial space with the same dimension as the classical finite element space.The cell problems may use independent local discretizations, and their constraints are implemented with Lagrange multipliers.Source: §6.1 · p. 16

6.2. Error analysis of the fully discrete method.

  • Quantifying a feasible fine scale h is open in most settings discussed, although hκ^3/2 ≲ 1 suffices for a pure Robin problem on a convex domain with P1 elements.The paper's error analysis of the general approach is deferred because estimating the additional corrector-discretization error is nontrivial and requires, for instance, regularity results.Source: §6.2 · pp. 16–17
  • With a sufficiently rich fine-scale space, Hκ ≲ 1, and ℓ ≳ log κ, the fully discrete localized method satisfies the stated stability and error estimate.The estimate compares the localized approximation with the reference solution uh.Source: §6.2 · p. 17
  • For a pure Robin problem on a convex domain, with h ≈ κ^-2, ∥u − uH,ℓ,h∥V ≲ (H + β^ℓκ + hκ)∥f∥Ω; H ≈ κ^-1 and ℓ ≈ 2 log(κ) yield overall error proportional to κ^-1 [Mel95].The full error also includes the reference error ∥u − uh∥V, whose further estimation relies on additional regularity of u.Source: §6.2 · p. 18
  • For the stated example, precomputation and coarse assembly cost roughly O(log(κ)^(d+1)κ^d), no more than solving the resulting O(κ^d)-dimensional coarse Helmholtz problem.The corrections are independent of the right-hand side and boundary data, a feature the paper identifies as useful in large-scale and inverse problems.Source: §6.2 · p. 18

7. Numerical Experiments

Two numerical examples examine the method’s theoretical predictions in one dimension and for scattering from a triangle. With appropriate oversampling, both the multiscale method and its Petrov–Galerkin variant eliminate pollution effects.

7.1. Illustration of the theoretical results in 1d.

  • With ℓ = |log2 H|, optimal convergence appears once Hκ ≲1; the observed rate N_dof^(3/(2d)) exceeds the predicted rate.In that regime, errors coincide with best approximation in the multiscale space.Source: §7.1 · p. 19
  • In one dimension, precomputed test functions cure pollution: the Petrov–Galerkin errors are comparable to best approximation, while coarse P1-FEM shows pollution.This example illustrates the theory but is not a proper benchmark, since local pollution-free generalized finite element methods exist in one dimension.Source: §7.1 · p. 19
  • Using Q_H instead of I_H produces faster corrector decay in the one-dimensional example, achieving the same accuracy with more localized basis functions.The improved performance is not quantified by existing theory and requires further investigation.Source: §7.1 · pp. 19–20

7.2. Scattering from a triangle.

  • For triangular scattering, ℓ = |log2 H| eliminates pollution for both methods; low solution regularity leaves multiscale convergence rates unchanged but reduces Petrov–Galerkin rates.The regularity of the reference solution also limits overall accuracy.Source: §7.2 · p. 23
Loading 1411.7512v3…