Frequency-explicit convergence analysis of a multiscale finite element method for highly heterogeneous scattering problems

T. Chaumont-Frelet; Z. Kassali

Source-linked AI summary

Frequency-explicit convergence analysis of a multiscale finite element method for highly heterogeneous scattering problems

T. Chaumont-Frelet, Z. Kassali

arXiv:2609.04930v1Sep 4, 2026math.NAmath.AP

TL;DR

The paper addresses high-frequency scattering by penetrable obstacles whose periodic heterogeneities are much smaller than the wavelength. It develops frequency-explicit homogenization and high-order multiscale finite-element analysis, finding that higher polynomial degree mitigates frequency effects in the multiscale error estimates and numerical examples. The theory, however, assumes exact local PDE solves and does not analyze the practical second-level discretization used numerically.

Problem
The problem is high-frequency Helmholtz scattering with periodic heterogeneities in the regime kε ≪ 1 and kℓD ≫ 1.
Method
The method combines frequency-explicit homogenization estimates with analysis of the multiscale hybrid-mixed finite-element method.
Results
High-order elements mitigate growth of the stability constant with frequency, and numerical examples support the value of high-order multiscale methods.
Takeaways & Limitations
Higher polynomial degree is better suited to the analyzed high-frequency multiscale problems.
Takeaways & Limitations
The estimates assume exact element-wise PDE solves, while the numerical implementation uses a second-level mesh with h ∼ ε whose effect is not theoretically analyzed.

Abstract

from arXiv · show

We analyze the numerical approximation of time-harmonic scattering by highly heterogeneous penetrable obstacles. These problems are especially challenging in the high-frequency regime, where the size of the scatterer $L$ is much larger than the wavelength, i.e., the wavenumber $k$ is such that $kL \gg 1$. Here, we further consider the situation where the scatterer contains different materials, with a characteristic size $\varepsilon$ such that $k\varepsilon \ll 1$. We propose a high-order multiscale finite element method, and provide an error analysis that is explicit in both $k$ and $\varepsilon$. Crucially, our error estimates suggest that using a high-order method should reduce the computational cost for large frequencies, which is corroborated by numerical examples.

1. Introduction

The paper studies high-frequency Helmholtz scattering with periodic heterogeneities and develops frequency-explicit homogenization and multiscale finite-element analysis. Its results show that high-order multiscale elements can mitigate frequency-dependent stability growth, while the numerical theory assumes exact local basis-function solves.

Problem setting: The target regime has microscopic heterogeneities and macroscopic high-frequency scattering: kε ≪ 1 and kℓD ≫ 1.The wavelength is smaller than the computational domain but larger than the heterogeneity scale.
Theoretical contributions: The first contribution derives homogenization error estimates explicit in the wavenumber and uniform over all f ∈ L2(Ω).Under appropriate coefficient assumptions, the stability constant is independent of ε and can receive fully explicit bounds in k.
Theoretical contributions: The second contribution analyzes the multiscale hybrid-mixed method, whose elementwise basis functions use Neumann problems with polynomial boundary data of degree ℓ.The methodology may also apply to methods such as MsFEM with Dirichlet basis construction.
Convergence result: Under scale separation and conditions Cst(kε)1/2 ≪ 1 and Cst(kH)^(ℓ+1) ≪ 1, the finite-element solution is quasi-optimal.When ε ∼ H, resonance terms involving ε/H enter the estimate; high-order elements mitigate growth of the stability constant with frequency.
Scope and limitation: The theoretical estimates assume exact element-wise PDE solves for basis functions, although the numerical method uses a second-level discretization with h ∼ ε.The effect of this second-level discretization is not analyzed theoretically.
Numerical evidence: Numerical examples show interest in high-order multiscale methods for high-frequency problems.The paper also notes that modern LOD methods can be robust but may be more expensive because they solve local PDEs on element patches.

2. Setting

The setting consists of a smooth obstacle inside a larger computational domain, with periodic, elliptic heterogeneous coefficients and a high-frequency, scale-separated regime. The notation specifies the function spaces, periodic cells, coefficient regularity, and parameter dependencies used in the analysis.

Geometrical configuration: The obstacle D is a smooth bounded domain in R^d, d = 2, 3, embedded in a larger smooth domain Ω with transparent boundary conditions.The setup assumes D ⋐ Ω and a distance from ∂Ω to D at least ℓD.
Scale regime: Inside D, ε is the coefficient length scale and the analysis focuses on kε ≪ 1 and kℓD ≫ 1.This combines rapidly oscillating material structure with high-frequency scattering at the obstacle scale.
Function spaces: The framework uses L2 and Sobolev spaces, including piecewise H1(Ω\Γ) and H2(Ω\Γ) regularity across the interface Γ.Vector-field and H0(div, U) notation is also introduced for the analysis.
Periodic coefficients: The periodic coefficients are described on the unit cell Y = (0,1)^d using periodic Lipschitz spaces and standard ellipticity and boundedness assumptions.The coefficient bounds include 0 < Amin ≤ bA(y)ξ · ξ ≤ Amax < ∞ for unit vectors ξ.
Assumptions: The authors believe the regularity requirements on the periodic coefficients could be weakened, but doing so would require additional technical work.Thus the stated coefficient regularity is an assumption of the presented analysis, not necessarily an intrinsic requirement.
Parameter dependence: Hidden constants may depend on coefficient and domain geometry, kε and kℓD through simplified bounds, mesh shape regularity, and polynomial degree, but not on ε, k, Hmin, or Hmax.The additional mesh-related dependencies are introduced in the multiscale-method analysis.

2.6. Highly heterogeneous scattering problem.

This section formulates the heterogeneous Helmholtz scattering problem and its homogenized counterpart. The heterogeneous coefficients are generated from periodic patterns, while homogenized coefficients are obtained by cell averaging and corrector problems.

Heterogeneous problem: The model seeks uε ∈ H1(Ω) solving the Helmholtz problem with heterogeneous coefficients in D and identity coefficients in Ω\D.The outer boundary satisfies the absorbing condition ∇uε · n − ikuε = 0 on ∂Ω.
Homogenization: The homogenized scalar coefficient µH is the mean of its periodic pattern over the cell Y.The homogenized matrix coefficient AH is constructed through periodic correctors and cell averaging.
Homogenization: Periodic correctors χ̂j solve auxiliary cell problems with zero mean, and their collection forms the corrector used to define AH.The corrector is stated to have W1,∞ regularity under the coefficient assumptions.
Homogenized problem: The homogenized problem replaces the oscillating material coefficients with piecewise-constant effective coefficients while retaining the scattering geometry and boundary condition.Its weak formulation is posed for all v ∈ H1(Ω).
Weak formulation: The homogenized variational formulation uses a bilinear form b0 obtained from the heterogeneous form bε by replacing µε and Aε with µ0 and A0.The formulation is interpreted in weak form rather than only through the displayed differential equation.

2.9. Elementary stability results.

The homogenized problem is uniquely solvable, with stability behavior in frequency characterized under geometric and coefficient assumptions. For the heterogeneous problem, uniform stability in ε is established only under broader conditions than earlier results.

The homogenized problem has a unique solution for every f ∈ L2(Ω).
For star-shaped obstacles with µ0 ≤ 1 and A0 ≤ I, the homogenized stability constant satisfies C^H_st ∼ kℓD.
The heterogeneous stability constant C^ε_st is uniformly bounded in ε for fixed k under the stated coefficient assumptions.
Earlier available heterogeneous estimates required coefficient independence in one variable, faster wave speed inside D, and a geometric obstacle condition.
The present estimate additionally applies under relaxed assumptions, including piecewise constant periodic patterns and absorbing boundary conditions.

3. Uniform stability estimates for the highly heterogeneous scattering problem

This section proves that the heterogeneous scattering stability constant remains uniformly bounded as ε varies for fixed frequency, using a contradiction argument and homogenization through two-scale convergence.

The main result establishes that C^ε_st is uniformly bounded in ε for any fixed k.The authors identify this as a result not previously rigorously established for general configurations.
The stability constant is also shown to be continuous as a function of ε.
The proof proceeds by contradiction, considering normalized right-hand sides and a sequence εn that must converge to zero.
Well-posedness of the homogenized problem forces the limiting function u0 to vanish, contradicting the normalized L2 behavior.
Two-scale convergence yields a homogenized limit relation, bεn(ûn,v) → b0(u0,v).

4. Homogenization of the scattering problem

The homogenization analysis compares heterogeneous and homogenized scattering solutions with estimates explicit in frequency and heterogeneity scale. It then derives a corresponding comparison of stability constants.

Theorem 4.1 provides frequency-explicit homogenization error estimates for the heterogeneous solution uε and homogenized solution u0.
The proof adapts prior estimates by replacing the layer stability factor with the heterogeneous stability constant C^ε_st.
The homogenization estimate directly yields a comparison between the heterogeneous and homogenized stability constants.

5. The multiscale hybrid-mixed method

The multiscale hybrid-mixed method relaxes primal continuity and reduces the scattering problem to a skeleton formulation for a multiplier. Local Helmholtz problems define the operators, while high-order Raviart–Thomas traces provide the discrete space.

Mesh and regularity: The method uses a curved simplicial mesh aligned so each element lies entirely inside or outside the obstacle.
Mesh and regularity: Its analysis assumes element-wise Neumann problems admit a uniform L2(Ω)-to-H2(Ω) regularity shift.
Hybrid formulation: The hybrid formulation relaxes primal continuity and introduces a Lagrange multiplier on the mesh skeleton.
Hybrid formulation: Local operators Tε and bTε are defined through element-wise Helmholtz problems with Neumann boundary conditions.
Skeletal formulation: When kHmax ≤ τ, the local operators are well-defined and the primal unknown can be eliminated through a skeletal variational formulation.
Discretization: The discrete multiplier space is spanned by normal traces of Raviart–Thomas elements, with polynomial traces that are single-valued on each face.
Discretization: The method requires solving local Helmholtz problems numerically through a second-level discretization in practice.

6. Frequency-explicit convergence analysis

The analysis establishes quasi-optimality and frequency-explicit convergence for the MHM method, combining approximation, local homogenization, and stability estimates. The resulting bounds clarify how mesh scales, heterogeneity, frequency, and polynomial degree affect accuracy.

Abstract stability and convergence: The MHM method is quasi-optimal when its approximation factor is bounded by a constant independent of ε, k, and H.Under this condition, the discrete problem has a unique solution.
Approximation and homogenization: The analysis combines homogenized-problem approximability with local homogenization estimates under the condition kHmax ≤ τ.The local estimates focus on mesh elements intersecting the heterogeneous scatterer.
Scope and limitations: The local homogenization analysis relies on Lipschitz coefficient regularity, although weaker assumptions may be possible, especially in two dimensions.The authors note that sharper arguments could potentially relax the three-dimensional requirement.
Final error estimate: The final MHM error estimate provides a unique discrete solution under explicit conditions involving ε, k, Hmax, and Hmin.The result follows by combining abstract MHM stability with the approximability theorem.

7. Numerical examples

The numerical examples compare multiscale and standard methods across frequencies, heterogeneity scales, mesh sizes, and polynomial degrees. They show that multiscale discretization is beneficial, while higher order becomes especially effective at the larger frequency.

Experimental setting: The experiments use Cartesian coarse grids with N from 16 to 1024 and MHM polynomial degrees ℓ = 0 and 1, supported by fine-scale element solves.Reference errors are computed against a solution using N = 2048 and ℓ = 1.
Results at ω = 2π: At ω = 2π, multiscale results outperform the standard method for all considered ε values.The standard method is limited mainly by coefficient approximation, whereas multiscale basis functions retain k-oscillatory behavior.
Results at ω = 2π: At ω = 2π, resonance appears for smaller ε, while the ℓ = 0 curves can plateau when the error is dominated by H-dependent terms.The observed decrease with ε can be linear even though the theory predicts a √ε rate, possibly because the computations are pre-asymptotic.
Results at ω = 2π: At ω = 2π, the ℓ = 1 multiscale method is more performant, but its gain is relatively small.This contrasts with the stronger high-frequency advantage observed for the higher-order method.
Results at ω = 10π: At ω = 10π, the ℓ = 0 method barely reaches 10% accuracy, whereas ℓ = 1 delivers good accuracy across cases on reasonable meshes.The higher-order advantage agrees with the frequency-explicit analysis.

Appendix A. Homogenization of local problems

This appendix establishes local homogenization results used in the main analysis, under the mesh resolution condition kHmax ≤ τ. It introduces the variational framework and corrector-based homogenization estimate for local problems.

The appendix assumes kHmax ≤ τ and uses this mesh-resolution condition in its local homogenization analysis.
An inf-sup condition derived from a Poincaré inequality provides the stability needed for the local problems.
For given volume data f and boundary data λ, unique heterogeneous and homogenized local solutions wε and w0 exist.
The homogenization result compares wε with w0 plus a first-order corrector εχε · ∇w0 through the corresponding variational forms.
Correctors are introduced as specialized functions on the physical domain and periodic cell to perform the homogenization analysis.

A.1. Correctors.

This section develops first- and second-order correctors, trace and vector-potential tools, and regularity estimates needed to complete the local homogenization proof.

Correctors: The first-order corrector is defined from the cell function bχ and the gradient of the homogenized solution w0.
Correctors: A second-order corrector bη is defined as a zero-mean element of the periodic H1 space and is shown to be well posed.
Correctors: Under the stated coefficient regularity, the second-order corrector gains W 1,∞ regularity through a Sobolev-embedding argument.
Auxiliary estimates: A uniform trace inequality and a vector potential q are introduced to control boundary terms in the homogenization estimates.
Auxiliary estimates: These ingredients are assembled to conclude the proof of Theorem A.2 for the homogenized solution w0.

A.2. End of the proof.

The proof concludes by decomposing the homogenization error into terms controlled with first- and second-order correctors, integration by parts, and a vector-potential argument.

End of the proof: The coefficient difference bμε − μH is rewritten as a scaled divergence involving the second-order corrector bη.
End of the proof: Integration by parts transforms the coefficient-error contribution into volume terms involving ∇w0 and the flux generated by bη.
End of the proof: The argument begins by setting w1 := bχ · ∇w0 and expanding the relevant bilinear-form expression according to equation (2.1).
End of the proof: A vector potential q is used to rewrite the first term, after which the divergence theorem and ∇·∇×qε = 0 simplify the expression.
End of the proof: The proof estimates four right-hand-side terms using the first-order corrector, second-order corrector, and vector-potential results.