Adaptive Schwarz with Spectral Enrichment

• The paper introduces an adaptive Schwarz method that enriches coarse spaces with locally computed spectral basis functions to achieve mesh-independent convergence.
• It employs eigenproblems on a-harmonic subspaces to construct minimal-dimensional, optimal coarse spaces, balancing accuracy and computational cost.
• Extensive numerical experiments confirm exponential convergence rates in high-contrast, multiscale PDEs for both 2D and 3D applications.

Adaptive Schwarz with local spectral enrichment is a class of domain decomposition algorithms that accelerate the iterative solution of discretized partial differential equations (PDEs) with highly heterogeneous or multiscale coefficients. By augmenting standard Schwarz domain decomposition frameworks with data-driven, spectrally optimal local basis functions, these methods enable robust, mesh-size- and contrast-independent convergence even in the presence of severe multiscale features. The approach, exemplified by the restricted additive Schwarz (RAS) method with multiscale spectral generalized finite element method (MS-GFEM) enrichment, leverages local eigenproblems on a-harmonic subspaces to construct minimal-dimensional optimal coarse spaces (Strehlow et al., 2024).

1. Variational Setting and Discretization

Let $\Omega \subset \mathbb{R}^d$, $d \in \{2,3\}$, be a Lipschitz domain, and let $A(x) \in L^\infty(\Omega)^{d \times d}$ be symmetric and uniformly elliptic, $$\alpha |\xi|^2 \leq A(x)\xi \cdot \xi \leq \beta |\xi|^2$$. The variational problem seeks $u \in H^1_0(\Omega)$ such that

$$a(u,v) := \int_{\Omega} (A \nabla u) \cdot \nabla v\, dx = \ell(v) \quad \forall v\in H^1_0(\Omega)$$

for a given linear functional $\ell\in H_0^1(\Omega)'$. Discretization over a conforming mesh $\mathcal{R}_h$ with mesh size $h$ gives the finite element space $V_h \subset H^1_0(\Omega)$ and the linear algebraic system $A u = f$.

2. Overlapping Decomposition and Discrete Harmonic Spaces

The domain is covered with $n$ overlapping subdomains $\{\Omega_i\}_{i=1}^n$, each a union of fine elements, with extended (oversampled) domains $\Omega_i^* \supset \Omega_i$. Locally, the spaces $V_h(\Omega_i^*)$ (finite elements restricted to $\Omega_i^*$) and $V_{h,0}(\Omega_i^*)$ (with support in $\Omega_i^*$) are defined. The local bilinear form is $$a_{\Omega_i^*}(u,v) = \int_{\Omega_i^*} (A \nabla u) \cdot \nabla v\, dx$$. The discrete a-harmonic space is

$$V_{h,A}(\Omega_i^*) := \left\{ v \in V_h(\Omega_i^*) : a_{\Omega_i^*}(v, w) = 0 \;\forall w\in V_{h,0}(\Omega_i^*) \right\}$$

which consists of local FE functions that are a-harmonic in $\Omega_i^*$.

3. Construction of Local Spectral Basis via Eigenproblems

For each $\Omega_i^*$, a partition-of-unity operator $\chi_{h,i}: V_h(\Omega_i^*) \to V_{h,0}(\Omega_i)$ is constructed (by smooth multiplication and FE interpolation). The key local eigenproblem on the a-harmonic space is:

$$a_{\Omega_i}(\chi_{h,i}(\phi_{i,k}|_{\Omega_i}),\, \chi_{h,i}(v|_{\Omega_i})) = \lambda_{i,k}\; a_{\Omega_i^*}(\phi_{i,k}, v) \quad \forall v \in V_{h,A}(\Omega_i^*)$$

Interpreted as the singular value decomposition (SVD) of a compact transfer operator $P_i: V_{h,A}(\Omega_i^*) \to V_{h,0}(\Omega_i)$, these eigenproblems yield rapidly decaying singular values $\sigma_{i,k} = \lambda_{i,k}^{1/2}$ and corresponding optimal local basis functions. The eigenfunctions for $\lambda_{i,k}$ below a prescribed tolerance $\lambda_\mathrm{tol}$ are selected for enrichment.

4. Adaptive Coarse Space Enrichment Strategy

In each subdomain, eigenfunctions with $\lambda_{i,k}\leq\lambda_\mathrm{tol}$ are retained to form the local coarse space $$S_i = \mathrm{span}\{ \phi_{i,k}|_{\Omega_i} : k\in I_i \}$$, $$I_i := \{k: \lambda_{i,k}\leq \lambda_\mathrm{tol}\}$$. The global coarse (MS-GFEM) space is assembled as

$$V_H = \left\{ \sum_{i=1}^n \chi_{h,i}(\phi_i) : \phi_i \in S_i \right\}$$

As $\lambda_\mathrm{tol}\rightarrow 0$, the approximation becomes arbitrarily accurate but the coarse space dimension increases. In practice, $\lambda_\mathrm{tol}$ is set so that the resultant dimension achieves rapid convergence, balancing iteration count and coarse problem size.

5. Two-Level Restricted Additive Schwarz Algorithm

The two-level RAS algorithm is formally defined as:

Notation:

• $R_i^T$: zero-extension from $V_{h,0}(\Omega_i^*)$ to $V_h$
• $A_i = R_i A R_i^T$: local matrix
• $\chi_i$: matrix corresponding to $\chi_{h,i}$
• $R_H^T$, $A_H$: embedding and matrix for the global coarse space

Preconditioner Action on Residual $r$:

1. Local solves (all $i$ in parallel): $d_i = \chi_i\, A_i^{-1} R_i\, r$
2. Sum: $d = \sum_i R_i^T d_i$
3. Coarse correction: $g = R_H^T A_H^{-1} R_H (r - A d)$
4. Return: $B r = d + g$

Iteration advances as $u^{j+1} = u^j + B (f - A u^j)$, or $B$ is used as a preconditioner for GMRES.

6. Convergence Properties and Exponential Decay

Define $$\Lambda = \left( \xi\,\xi^*\max_i \lambda_{i, m_i+1} \right)^{1/2}$$, where $m_i = \#\text{basis retained}$, and $\xi, \xi^*$ are overlap-coloring constants. The main convergence results are:

• Richardson iteration: $\|u^{j+1} - u_h\|_a \leq \Lambda \|u^j-u_h\|_a$.
• GMRES: $\|BA(u - u^j)\|_b \leq \Lambda^j C \|BA(u-u^0)\|_b$, $C$ depends only on norm equivalence.

Theoretical justification is provided by the GFEM best-approximation property and spectral estimates from the eigenproblem:

$\|v - G(v)\|_a \leq \Lambda \|v\|_a$

Importantly, it is proved that the eigenvalues decay exponentially with oversampling and basis number:

$$\lambda_{i,k}^{1/2} \leq C_i \exp(-b_i k^{1/(d+1)})$$

independent of $h$; thus, only $O((\log 1/\epsilon)^{d+1})$ local basis functions are necessary to ensure rapid convergence, uniformly in $h$ and regardless of coefficient contrast.

7. Performance: Numerical Experiments and Practical Considerations

Extensive numerical experiments substantiate the theoretical results:

• 2D high-contrast “skyscraper” problem: With mesh-size $h=1/700$ and $7\times 7$ subdomains, iteration count decreases exponentially as either basis size or oversampling increases. Optimal total time is achieved for moderate enrichment (e.g., oversampling=8, $12$ eigenfunctions).
• 3D composite aero-structure elasticity: On up to $10^6$ DOF and $1024$ cores, $30$ eigenfunctions and a single oversampling layer yield $10$–$20$ GMRES iterations, independent of both mesh size and problem size.

The hybrid (multiplicative coarse) RAS variant consistently outperforms additive approaches. MS-GFEM coarse spaces display significant gains in coarse-space dimension versus iteration count compared to classical GenEO. Rapid, $h$-independent convergence is achieved with very compact coarse spaces (Strehlow et al., 2024).

8. Connections and Theoretical Context

Adaptive Schwarz with local spectral enrichment operates at the intersection of domain decomposition, multiscale methods, and spectral approximation theory. The key properties—contrast- and mesh-size-independent iteration bounds, exponential decay of spectral errors, and minimal coarse space dimension for target accuracy—generalize to DG settings (Eikeland et al., 2017), Helmholtz equations with impedance transmission (Ma et al., 2024), and elliptic systems with high-contrast or oscillatory coefficients in both $H^1$ and $H(\text{curl})$ formulations. Algebraic variants have been developed for robust preconditioning of linear systems where geometric information is inaccessible (Heinlein et al., 2022).

Theoretical advances in exponential localization and stable decomposition underpin the robustness of the method, with the convergence rate fully characterized in terms of the spectral decay of local eigenproblems. This yields a practical criterion for enrichment: select all eigenfunctions with eigenvalues below a computable threshold dictated by the desired global convergence rate. The minimal set of localized spectral modes guarantees both theoretical and observed scalability as mesh size vanishes and coefficient contrast increases.