Additive Schwarz Preconditioners

• Additive Schwarz-type preconditioners are domain decomposition techniques that decompose a global PDE system into overlapping subdomains to enable efficient and scalable parallel solutions.
• They employ local subdomain solvers together with a global coarse space to robustly control condition numbers across various discretizations, including nonlinear and nonsymmetric systems.
• Their modular design facilitates parallel implementation and has been extended to handle complex problems such as high-contrast media and adaptive multilevel frameworks.

Additive Schwarz-type Preconditioners

Additive Schwarz-type preconditioners constitute a foundational class of domain decomposition techniques for accelerating the solution of large, sparse systems arising from the discretization of partial differential equations (PDEs). In this family, the solution space is decomposed into overlapping (or nonoverlapping) subdomains, enabling scalable parallel solution strategies and, when appropriately designed, robust condition number control. Additive Schwarz preconditioners have been developed and theoretically analyzed for a range of discretizations—including continuous and discontinuous Galerkin finite elements, variational formulations, nonsymmetric and indefinite systems, and even nonlinear problems. Their role as both solvers and preconditioners for Krylov subspace methods is central in modern computational mathematics and scientific computing.

1. Abstract Form and Key Principles

Let $A \in \mathbb{R}^{n \times n}$ be a linear system arising from the discretization of a PDE, and let $u$ be its solution. In an additive Schwarz setup, the index set of unknowns is covered by $N$ possibly overlapping subdomains, defined by restriction operators $R_i$, and possibly a global "coarse" subspace with restriction $R_0$, as

$$M^{-1} = R_0^T A_0^{-1} R_0 + \sum_{i=1}^N R_i^T A_i^{-1} R_i,$$

where $A_i = R_i A R_i^T$ are local subdomain matrices and $A_0$ is the coarse space Galerkin operator.

Each $A_i^{-1}$ approximates the action of $A^{-1}$ restricted to its subdomain, and the combination ensures stable splitting properties for the global system. For symmetric positive definite (SPD) problems, the coarse space is typically constructed to guarantee robustness and scalability of the preconditioner, i.e., condition number $\kappa(M^{-1}A)$ remains bounded independently of $N$ and (often) mesh parameters $h$ and $H$.

The method is extended to nonsymmetric, indefinite, or nonlinear systems by judicious adaptation, as detailed in subsequent sections.

2. Schwarz Preconditioning for Discontinuous Petrov-Galerkin Methods

In the context of DPG discretizations of second-order elliptic PDEs, a canonical instance is the ultraweak formulation of Poisson's equation. The trial space $U$ consists of $$[\mathrm{L}^2(\Omega)]^d \times \mathrm{L}^2(\Omega) \times H_0^{1/2}(\partial \mathcal{T}_h) \times H^{-1/2}(\partial \mathcal{T}_h)$$, and the operator $A_h$ is SPD with respect to a mesh-dependent energy norm, incorporating both element and trace variables (Barker et al., 2012).

A one-level overlapping additive Schwarz preconditioner is constructed by partitioning the domain into overlapping subdomains $\{\Omega_j\}$ and constructing local subspace injections $I_j: U_j \to U_h$. The preconditioner,

$B_h^{-1} = \sum_{j=1}^J I_j A_j^{-1} I_j^T,$

admits a rigorous energy-norm condition number estimate $\kappa(B_h^{-1}A_h) \leq C \delta^{-2}$, where $\delta$ is the overlap width, leveraging stable partition-of-unity decompositions and discrete trace norm estimates. In practice, this yields significant reduction in Krylov iteration counts—from $\mathcal{O}(h^{-2})$ to $\mathcal{O}(\ln h)$. Extension to multi-level or two-level Schwarz frameworks is natural, with coarse spaces tailored to DPG test-enrichments (Barker et al., 2012).

3. Adaptive, Algebraic, and Multiscale Coarse Spaces

A critical advance in Schwarz theory is robust coarse space construction, especially for problems with high contrast, multiscale, or discontinuous coefficients. Several variants have been developed:

• Adaptive Schwarz with Local Spectral Enrichment: In heterogeneous elliptic problems, local eigenvalue problems are solved in each subdomain, and eigenfunctions corresponding to small Rayleigh quotients are included in the coarse space. The resulting adaptive preconditioner achieves a condition number bound

$$\kappa(PA) \lesssim \left( \max_k \lambda_{M_k+1}^k \right)^{-1} \frac{H}{h},$$

with $\lambda_{M_k+1}^k$ the first omitted local eigenvalue, robust to coefficient jumps (Marcinkowski et al., 2017).

• Algebraic Multiscale and Energy-Minimizing Coarse Spaces: Purely algebraic techniques such as AMS (Algebraic Multiscale Solver), GDSW (Generalized Dryja–Smith–Widlund), and RGDSW variants build prolongation operators and coarse matrices based only on system matrix block structure, often via discrete harmonic extensions and component-wise partition-of-unity. The condition number can be controlled as $\kappa(M^{-1}A) \leq C(1+H/\delta)$, even in high-contrast or multiscale settings (Alves et al., 2024, Cumaru et al., 1 Dec 2025).
• Spectral Schwarz for Nonsymmetric and General Sparse Matrices: For general (SPD, indefinite, nonsymmetric) matrices, local eigenproblems or SVDs define the coarse basis, ensuring controlled convergence in GMRES or CG, with fully algebraic setup and coarse space selection (Daas et al., 2024).

4. Schwarz Preconditioning for Nonsymmetric or Indefinite Problems

The extension of additive Schwarz preconditioners to nonsymmetric or indefinite systems (e.g., finite volume element discretizations, advection-diffusion, or highly oscillatory Helmholtz problems) requires new analyses and variants:

• Nonsymmetric Schwarz and Least Squares Preconditioners: For a nonsymmetric definite $A$, both the direct Schwarz variant $B^{-1}$ and the normal matrix variant $Z = B A_0^{-1} B^T$ (with $A_0$ a symmetric positive definite comparison matrix) are considered. Combined variable preconditioning—using optimal linear combinations of $B^{-1}$ and $Z^{-1}$—achieves convergence rates at least as good as the symmetric case, with only weak dependence on mesh and subdomain count (Dios et al., 2012, Marcinkowski et al., 2014).
• Nonoverlapping and Optimized Schwarz Methods: For Helmholtz-type problems at high frequency, nonoverlapping spectral additive Schwarz (NOSAS) and optimized RAS (ORAS) methods employ local generalized eigenproblems to identify and remove near-kernel components, with adaptive coarse spaces based on spectral thresholding. These methods yield $k$-robust convergence with the average number of coarse modes per subdomain remaining moderate and iteration counts nearly independent of mesh refinement or wavenumber (Yu et al., 1 May 2025, Ma et al., 2024, May et al., 2019).

5. Nonlinear and Trace-Structural Schwarz Solvers

Additive Schwarz principles have been extended to nonlinear PDEs, where nonlinear subdomain solves are used to construct preconditioned nonlinear residual maps for use with Newton methods:

• RASPEN and SRASPEN: The Restricted Additive Schwarz Preconditioned Exact Newton method constructs a global nonlinear residual by assembling local nonlinear Dirichlet solves into a residual map, and its Jacobian can be assembled from local Newton iterations (Dolean et al., 2016). These methods can be formulated both on the volume (full unknown) level and in a substructured (trace/skeleton) fashion, resulting in lower memory cost and communication requirements with the same optimal convergence rate (Chaouqui et al., 2021).
• Schwarz Preconditioning for Partition of Unity and Surface PDEs: Additive Schwarz preconditioners in the context of partition of unity methods (e.g., for the obstacle problem for plates or state-constrained optimal control problems) as well as for surface PDEs discretized by embedding methods (e.g., CPM) have been constructed. These frameworks leverage tailored subdomain covers and partition-of-unity weights, achieving condition number bounds that depend on overlap size and, with appropriate coarse spaces, system size robustness (Brenner et al., 2018, Brenner et al., 2018, May et al., 2019).

6. Theoretical Condition Number Estimates and Scalability

A recurring theme is the relation of the preconditioned system's condition number to parameters of the decomposition:

• Condition Number Dependence on Overlap and Coarse Space: $$\kappa(M^{-1}A) \leq C\left(1+\frac{H}{\delta}\right),$$ where $H$ is the subdomain diameter and $\delta$ the overlap. Without a coarse space, one-level Schwarz methods typically see a condition number deteriorating as mesh is refined or $N$ grows (e.g., $\kappa \sim H^{-2}$ for DPG (Barker et al., 2012)). Two-level and adaptive coarse spaces restore scalability and robustness, ensuring the number of Krylov iterations is nearly constant.
• Performance in Multilevel and Adaptive Settings: Multilevel local diagonal Schwarz preconditioners for boundary integral equations on locally refined meshes can achieve condition numbers independent of number of elements and refinement levels, $\kappa = O(1)$ (Feischl et al., 2013). In fractional PDEs and DG discretizations, modified Schwarz methods deliver spectral bounds essentially uniform in $h$, $p$, penalization, and coefficient variations (Antonietti et al., 2014, Jiang et al., 2015).

Table: Examples of Schwarz Preconditioners and Condition Number Estimates

Preconditioner Type	Condition Number Bound	Reference
1-level ASM for DPG	$\kappa \leq C\delta^{-2}$	(Barker et al., 2012)
2-level adaptive coarse (SUBD/LAYER)	$\kappa \lesssim (H/h)/\tau$	(Marcinkowski et al., 2017)
Non-overlapping NOSAS/ORAS (Helmholtz)	$\kappa \leq (1+2\eta)/(1-2\eta)$, $\eta<$1/2	(Yu et al., 1 May 2025)
LMLD for BEM (multilevel)	$\kappa = O(1)$	(Feischl et al., 2013)
Classical FVE Schwarz (nonsymmetric)	$\kappa \leq C(1+\log(H/h))^2$	(Marcinkowski et al., 2014)
2-level OAS with RGDSW (Stokes)	$\kappa \leq C(1+H/\delta)(1+\log(H/h))^2$	(Cumaru et al., 1 Dec 2025)

7. Scalability and Parallel Implementation

Practical implementations of additive Schwarz-type preconditioners exploit their locality: local subdomain (or trace) solves are executed independently and in parallel, with communication patterns dictated by overlap and coarse space solution assembly. Large-scale performance is governed by: (1) local factorization costs (usually $O(n_i^2)$ for subdomain size $n_i$), (2) coarse solve cost (scales as $O(n_C^3)$), and (3) communication for assembling global corrections.

Extremely large systems, arising e.g. in fluid flow, multiscale media, or complex geometries, benefit from algebraic coarse space selection and efficient parallel assembly and factorization, as demonstrated for FROSch/NGSolve (Cumaru et al., 1 Dec 2025), as well as algebraic multigrid and substructuring frameworks (Daas et al., 2024, Chaouqui et al., 2021). Iteration counts and total parallel solve times scale very favorably up to hundreds of processors and millions of unknowns, providing robust solution strategies for a wide range of PDE models.

---

References (arXiv IDs):

• DPG Schwarz for Poisson: (Barker et al., 2012)
• Adaptive average Schwarz: (Marcinkowski et al., 2017)
• Uniform Schwarz for DG/hp: (Antonietti et al., 2014)
• General algebraic Schwarz: (Daas et al., 2024)
• Two-level Schwarz for Stokes: (Cumaru et al., 1 Dec 2025)
• Nonoverlapping/optimized Schwarz Helmholtz: (Yu et al., 1 May 2025, May et al., 2019, Ma et al., 2024)
• Multilevel BEM Schwarz: (Feischl et al., 2013)
• Nonsymmetric Schwarz/normal equation: (Dios et al., 2012, Marcinkowski et al., 2014)
• Substructured (trace) Schwarz, nonlinear: (Chaouqui et al., 2021, Dolean et al., 2016)
• Algebraic coarse spaces (AMS/GDSW): (Alves et al., 2024)
• Schwarz for fractional/integral/obstacle/control: (Jiang et al., 2015, Brenner et al., 2018, Brenner et al., 2018)

For further technical and implementation details on domain decomposition, multilevel, adaptive, and nonlinear Schwarz frameworks, the referenced papers provide rigorous analyses, pseudocode, and comprehensive computational experiments.