NOSAS: Nonoverlapping Spectral Additive Schwarz

• NOSAS is a domain decomposition preconditioner that uses nonoverlapping subdomains and local spectral analysis to build adaptive coarse spaces.
• It constructs coarse spaces by selecting eigenmodes from local interface eigenproblems and harmonic extensions, ensuring mesh-independent convergence.
• NOSAS demonstrates robust performance and parallel scalability on challenging PDEs, including high-contrast diffusion and Helmholtz problems.

The Nonoverlapping Spectral Additive Schwarz (NOSAS) methodology constitutes a family of domain decomposition-based preconditioners utilizing nonoverlapping subdomains and adaptively selected coarse spaces defined by local spectral problems. These preconditioners target symmetric positive-definite (SPD), indefinite, and highly oscillatory systems—including those with strong coefficient heterogeneity or arising from challenging PDEs such as the Helmholtz equation. NOSAS achieves robust coarse space representation and parallel scalability by combining local eigenproblem solves, low-rank discrete harmonic extensions, and nonoverlapping basis construction, resulting in convergence rates and condition numbers that are largely insensitive to coefficient contrast or local geometry (Yu et al., 2020, Yu et al., 1 May 2025).

1. Mathematical Framework and Problem Setting

NOSAS is posed in the context of PDE discretizations (finite element or algebraic) on a domain $\Omega\subset\mathbb{R}^d$, such as

• Elliptic model problems: $a(u,v) = \int_\Omega p(x)\nabla u\cdot\nabla v\,dx$ with $p(x)$ possibly highly heterogeneous (Yu et al., 2020).
• The Helmholtz equation: $-\Delta u - k^2 u = f$ with complex boundary conditions (Yu et al., 1 May 2025).

The global degrees of freedom (DoFs) are partitioned into $N$ nonoverlapping subdomains $\{\Omega_i\}$, each further splitting into interior and interface sets. After discretization, the global linear system is $A u = b$ (SPD) or $B u = \ell$ (indefinite/complex). Domain decomposition leads to the formulation of local subdomain problems and a global Schur complement/interface problem.

NOSAS constructs the coarse space adaptively by extracting local eigenmodes from Schur complements representing the “energy” of discrete harmonic extensions from each interface into the subdomain interior. This low-rank spectral approach distinguishes NOSAS from classical additive Schwarz and average Schwarz variants.

2. Construction of the Spectral Coarse Space

The adaptive coarse space in NOSAS is defined through the solution of generalized eigenvalue problems on the interface of each nonoverlapping subdomain. The standard sequence of steps is:

1. Subdomain Partitioning and Matrix Assembly: The global matrix (e.g., the SPD stiffness matrix $A$ or the complex Helmholtz matrix $B$) is split into subdomain blocks, isolating interface and interior DoFs.
2. Harmonic Extension and Schur Complement Formation: For each subdomain $\Omega_i$, the Neumann or Dirichlet local matrix $A^{(i)}$ (or $B^{(i)}$) is partitioned:

$$A^{(i)} = \begin{bmatrix} A^{(i)}_{\Gamma\Gamma} & A^{(i)}_{\Gamma I} \ A^{(i)}_{I\Gamma} & A^{(i)}_{II} \end{bmatrix}$$

1. Eigenproblem Definition: The interface Schur complement is $$S_i = A^{(i)}_{\Gamma\Gamma} - A^{(i)}_{\Gamma I} (A^{(i)}_{II})^{-1} A^{(i)}_{I\Gamma}$$. The key eigenproblem is

$$S_i \varphi_j^{(i)} = \lambda_j^{(i)} B_i \varphi_j^{(i)}$$

where $B_i$ is typically $A^{(i)}_{\Gamma\Gamma}$.

2. Selection and Extension of Modes: All eigenvectors with $\lambda_j^{(i)}\leq \lambda_{\text{cut}}$ (with $\lambda_{\text{cut}}=O(h/H)$ typical for mesh size $h$ and subdomain diameter $H$) are chosen. Each selected eigenvector $\varphi_j^{(i)}$ is extended harmonically into $\Omega_i$ by

$\Psi_j^{(i)} := H_i \varphi_j^{(i)},$

where $H_i$ denotes the harmonic extension operator.

This process defines the global NOSAS coarse space $$V_0 = \operatorname{span} \{ R_i^T \Psi_j^{(i)} : i, j \}$$, with overall dimension $\sum_i k_i$. The resulting basis is nonoverlapping in support and tailored to represent high-energy or high-contrast features—such as multiple high-conductivity channels intersecting $\Gamma_i$ (Yu et al., 2020).

3. Algorithmic Realization

The application of the NOSAS preconditioner within Krylov/substructuring solvers follows an additive Schwarz paradigm with specializations:

1. Local Eigenproblem Setup: Each subdomain solves for its local interface eigenpairs in the preprocessing stage.
2. Coarse Matrix Assembly and Inversion: The global coarse matrix $A_0=R_0^T A R_0$ (or the analogous assembly for the Helmholtz system) is constructed and factored, where $R_0$ collects local harmonic extensions.
3. Preconditioner Application (per iteration):
   • Local Correction: Solve $(A^{(i)})^{-1}$ or $(B^{(i)})^{-1}$ on $R_i r$, where $r$ is the current residual.
   • Coarse Correction: Solve on the spectral coarse space, $A_0^{-1} (R_0^T r)$, and prolong back to the fine space via $R_0$.
   • Combination: The preconditioned vector is $\sum_{i=1}^N R_i^T y_i + R_0 y_0$ (Yu et al., 2020).

For indefinite problems, variants handle real and imaginary parts separately, extracting coarse modes from both components of the interface Schur complements as needed (Yu et al., 1 May 2025).

4. Theoretical Properties and Robustness

NOSAS achieves rigorous and mesh-independent convergence properties characterized by:

• For elliptic problems, the condition number of the preconditioned system obeys:

$$\kappa(M^{-1}A) \leq C = (2 + 1/\lambda_{\text{cut}})(1 + \rho(E)) = O(H/h),$$

where $\rho(E)$ is the spectral radius of the subdomain overlap matrix, and crucially, this is independent of contrast $p_{\max}/p_{\min}$ (Yu et al., 2020).

• In the Helmholtz setting, provided the real and imaginary part thresholds ($\theta_{\rm Re},\,\theta_{\rm Im}$) are chosen such that $\theta_{\rm Re} + h k \theta_{\rm Im} < \gamma/2$ (with $\gamma$ the inf-sup constant of the sesquilinear form), the preconditioned spectrum collapses around 1, yielding efficient convergence (Yu et al., 1 May 2025).
• The adaptive selection of spectral modes guarantees that local unresolved interface energy is bounded, preventing “high-energy” artifacts from dominating convergence.

A comparison with additive average Schwarz (AAS) and minimum energy Schwarz (MES) indicates that NOSAS is robust to coefficient heterogeneity and attains optimal coarse space dimensioning in the presence of multiple high-permeability islands, where AAS and MES may fail to ensure bounded condition number unless strong global information is included (Yu et al., 2020).

5. Multilevel Extensions and Algebraic Generalizations

NOSAS admits fully algebraic and recursively multilevel generalizations, especially within the algebraic multigrid (AMG) setting for SPD systems of the form $A=G^T G$:

• The method applies to arbitrary sparse matrices, requiring no geometric information or explicit element matrices (Southworth et al., 7 Jan 2026).
• Recursive aggregation, local generalized eigenproblems, and block-diagonal interpolation build a hierarchy of coarse grids with nonoverlapping support at each level.
• Smoothers (e.g., restricted additive Schwarz, RAS) and local coarse solves are naturally parallelizable per aggregate.
• Ultimate performance is maintained as problem size increases, with operator complexity and iteration counts largely preserved in the face of strong anisotropy, coefficient jumps, or challenging geometry.

For Helmholtz and other indefinite problems, the NOSAS framework includes specific algebraic adaptations—such as saddle-point corrections and the splitting of real and imaginary parts—retaining the spectral adaptivity intrinsic to the coarse selection (Yu et al., 1 May 2025).

6. Parallel and Computational Aspects

• Setup Phase: The costliest component is the local spectral decomposition, scaling as $O(n_i^3)$ for subdomain size $n_i$. However, as these are independent, strong parallel scaling is observed.
• Iterative Phase: Each iteration (in PCG or GMRES) involves:
  • $N$ local solves (small systems),
  • A single coarse solve of size $\sum_i k_i$,
  • Minimal communication (all-reduce for coarse residual), enabling excellent parallel efficiency (Yu et al., 2020, Southworth et al., 7 Jan 2026).
• The global coarse space communicates only interface data, further enhancing scalability.
• For practical system sizes, e.g., $10^6$ DoFs and hundreds of subdomains, the coarse space remains tractable due to its adaptive size, and wall time is dominated by local operations.

7. Applications and Numerical Results

NOSAS demonstrably excels in domains with

• High-contrast diffusion (heterogeneous elliptic problems),
• Oscillatory or indefinite PDEs (Helmholtz, wave propagation),
• Strong anisotropy (e.g., magnetic confinement fusion heat conduction with $\kappa_\parallel/\kappa_\perp\gg1$) (Southworth et al., 7 Jan 2026).

Numerical experiments confirm:

• Uniform iteration counts ($\sim 10$–$50$) as the number of subdomains or anisotropy parameter increases.
• Operator complexity that can be tuned by coarsening and threshold selection, balancing computational work and convergence rates.
• Robustness compared to classical AMG and aggregation methods, especially where those approaches fail due to lack of global spectral adaptivity (Southworth et al., 7 Jan 2026, Yu et al., 2020, Yu et al., 1 May 2025).

The architecture supports broad generalization to algebraic and geometric scenarios and exhibits strong parallel scaling properties suitable for large-scale scientific computing environments.