Overlapping Additive Schwarz Methods

• Overlapping additive Schwarz methods are domain decomposition techniques that split a computational domain into overlapping subdomains and combine local solutions additively with a coarse correction for global error control.
• They employ adaptive coarse spaces, energy-minimizing extensions, and optimized transmission conditions to enhance convergence for elliptic, Helmholtz, saddle-point, and nonlinear problems.
• These methods are implemented in monolithic, recursive, and science-driven frameworks to improve scalability and efficiency in applications like incompressible flow, isogeometric analysis, and machine learning optimization.

Overlapping additive Schwarz methods are domain decomposition methods for the iterative solution of partial differential equations in which a computational domain is split into overlapping subdomains, local subproblems are solved independently, and the resulting corrections are combined additively. In their scalable forms, these methods include a coarse correction that propagates global information and damps error components that one-level local solves leave largely untouched. Contemporary work treats this framework not only for classical elliptic operators, but also for Helmholtz problems with optimized transmission conditions, saddle-point systems for incompressible flow, vector-field discretizations, fourth-order variational inequalities, boundary element formulations, and even nonlinear systems and scientific machine learning optimization problems (Alves et al., 2024, Köhler et al., 31 May 2026).

1. Core formulation and algorithmic structure

The standard overlapping construction starts from a nonoverlapping partition $\Omega_1,\dots,\Omega_N$ and enlarges each $\Omega_i$ to an overlapping subdomain $\Omega_i'$. At the algebraic level, with restriction operators $R_i$ and local matrices $A_i=R_i A R_i^T$, the canonical two-level preconditioner has the form

$$M_{\mathrm{OAS},2}^{-1}=\Phi A_0^{-1}\Phi^T+\sum_{i=1}^N R_i^T A_i^{-1}R_i, \qquad A_0=\Phi^T A\Phi,$$

so that local overlapping solves are complemented by a coarse solve (Alves et al., 2024, Cumaru et al., 1 Dec 2025). This local-plus-coarse additive structure recurs throughout the literature, including scalar elliptic systems, Stokes and Navier–Stokes saddle-point operators, and algebraic multiscale formulations.

At the PDE level, the same idea appears as a parallel overlapping iteration. For Helmholtz with an overlapping cover $\{\Omega_\ell\}$ and a partition of unity $\{\chi_\ell\}$, each subdomain problem is solved independently and the global iterate is reconstructed by weighted averaging,

$u^{n+1}=\sum_\ell \chi_\ell u_\ell^{n+1}.$

In correction form, this becomes $u^{n+1}=u^n+\sum_\ell \chi_\ell \delta_\ell^n$, which makes the interpretation as a preconditioned Richardson step explicit (Gong et al., 2021). The same weighted recombination is central in restricted variants, where the aim is to use overlap in the local problems while preventing redundant global updates.

A basic structural fact, emphasized repeatedly, is that one-level overlap alone does not yield scalability. Local subdomain solves damp high-frequency and localized error components efficiently, but global modes remain poorly controlled. The coarse level is therefore not an optional refinement of the method; it is the mechanism by which information is transmitted across all subdomains, and it is what prevents deterioration as the number of subdomains grows (Köhler et al., 31 May 2026, Alves et al., 2024).

2. Coarse spaces and interface representations

The decisive design problem in overlapping additive Schwarz methods is the coarse space. Much of the modern literature formulates coarse basis functions as energy-minimizing or discrete harmonic extensions of data prescribed on the subdomain interface $\Omega_i$0. With the standard interior/interface splitting

$\Omega_i$1

the extension takes the form

$\Omega_i$2

Thus the essential distinction among coarse spaces is the choice of interface values $\Omega_i$3 (Alves et al., 2024, Köhler et al., 31 May 2026).

The generalized Dryja–Smith–Widlund (GDSW) coarse space uses interface entities such as vertices, edges, and, in three dimensions, faces. Basis functions are defined from null-space traces on those entities and then extended into subdomain interiors. In monolithic saddle-point formulations for incompressible flow, the interface basis is block diagonal between velocity and pressure null spaces,

$\Omega_i$4

while the interior extension may couple velocity and pressure. This preserves the saddle-point structure on every level rather than statically condensing pressure (Köhler et al., 31 May 2026).

Reduced-dimension GDSW (RGDSW) compresses the interface description through nodal equivalence classes and partition-of-unity scaling. In the reduced-integration Stokes setting, coarse basis functions are assembled from coarse nodes, offspring relations among equivalence classes, and null-space restrictions corresponding to rigid-body translations,

$\Omega_i$5

The motivation is to retain the global low-energy content of GDSW while reducing coarse dimension and coarse-solve cost (Cumaru et al., 1 Dec 2025).

A computational study of algebraic coarse spaces places GDSW, RGDSW, and AMS in a common energy-minimizing framework. AMS is interpreted as an algebraic multiscale or MsFEM-type coarse space in which interface values are not merely constant on entities, but are computed recursively through reduced edge and face problems. This gives AMS a more adaptive interface representation in heterogeneous media (Alves et al., 2024).

Adaptive coarse spaces add another layer of selectivity. For three-dimensional elliptic problems with high coefficient contrast, face and edge generalized eigenvalue problems are solved, and eigenfunctions with eigenvalues below a threshold $\Omega_i$6 are added to the coarse space. Two variants are described: a wire-basket coarse space enriched by face eigenfunctions, and a vertex-based space enriched by both face and edge eigenfunctions. The resulting condition-number bounds are independent of coefficient contrast once sufficiently many low-eigenvalue interface modes are included (Eikeland et al., 2016).

Coarse space	Interface construction	Main objective
GDSW	Null-space traces on vertices, edges, faces	Standard scalable coarse correction
RGDSW	Nodal equivalence classes with partition-of-unity scaling	Reduced coarse dimension
AMS	Recursive algebraic edge/face/vertex problems	Multiscale adaptation in heterogeneous media
Adaptive spectral enrichment	Interface eigenfunctions below threshold $\Omega_i$7	Robustness to coefficient contrast

3. Restriction, transmission conditions, and Helmholtz-specific variants

For indefinite wave problems, overlap is typically paired with nontrivial transmission conditions. In the variational interpretation of ORAS, each Helmholtz subproblem is posed with impedance transmission conditions

$\Omega_i$8

and the preconditioner becomes

$\Omega_i$9

A technically important point is that the prolongation is a nodewise extension operator $\Omega_i'$0, since naive zero extension is generally not $\Omega_i'$1-conforming (Gong et al., 2021).

Restricted variants are motivated by the fact that unrestricted additive overlap counts shared corrections multiple times. In ORAS and related formulations, a discrete partition of unity is built into the prolongation:

$\Omega_i'$2

This uses overlap in the local problems while restricting the output so that the global update has partition-of-unity consistency (Martin et al., 20 Jun 2025).

For high-frequency Helmholtz, recent work replaces Robin transmission by local PMLs. The resulting one-level RAS-PML preconditioner has weighted extension

$\Omega_i'$3

and numerical experiments indicate that both overlap width and local PML thickness can decrease like $\Omega_i'$4 while maintaining good convergence. In the reported two-dimensional constant-wave-speed experiments, the method exhibits $\Omega_i'$5 parallel scalability under Cartesian decomposition and approximately $\Omega_i'$6 iteration counts and convergence time as $\Omega_i'$7 increases (Xie et al., 31 Jan 2026). Related experiments with Schwarz methods using PML transmission show that additive and multiplicative variants remain robust even when overlap shrinks to the mesh scale, provided the PML is about one wavelength wide, $\Omega_i'$8 (Galkowski et al., 2024).

At the same time, large three-dimensional geophysical experiments show the trade-off attached to overlap. ORAS may require fewer iterations than nonoverlapping substructured methods, but overlap enlarges local Helmholtz problems and raises memory and Krylov-storage costs. When optimized second-order transmission conditions are used in the nonoverlapping method, the convergence gap can shrink enough that the nonoverlapping approach becomes faster and much more memory efficient in multiple-right-hand-side regimes (Martin et al., 20 Jun 2025). This is a recurrent caution in the Helmholtz literature: overlap improves local communication, but it does not remove coarse-level or memory bottlenecks by itself.

4. Multilevel, monolithic, and recursive Schwarz hierarchies

Large saddle-point systems in incompressible flow have motivated monolithic overlapping Schwarz methods in which the coupled velocity–pressure structure is retained on every level. For the discrete Stokes operator

$\Omega_i'$9

the monolithic two-level GDSW preconditioner is

$R_i$0

In this setting, the local and coarse problems are themselves saddle-point systems, and the pressure is not condensed out (Köhler et al., 31 May 2026, Köhler et al., 6 Aug 2025).

The main recent extension is recursive. Instead of solving the coarse problem by a sparse direct method at the first coarse level, the same two-level construction is applied again to the coarse operator:

$R_i$1

This produces nested three-level and four-level preconditioners in which second- and third-level coarse problems are themselves treated by overlapping Schwarz, and only the smallest top-level problem is solved directly (Köhler et al., 31 May 2026).

The numerical message is nuanced rather than monotone. In a three-dimensional incompressible stationary Stokes problem with Carreau-type viscosity on a complex extrusion-die geometry, the three-level monolithic method is consistently better than the two-level method, and the four-level method is always faster than the two-level method. However, on the tested problem sizes the four-level method is slightly slower than the three-level method, because the third-level coarse problem is not yet large enough for a fourth level to offset its additional recursion overhead (Köhler et al., 31 May 2026). Earlier large-scale experiments with monolithic two- and three-level GDSW solvers showed the same coarse-bottleneck mechanism in practice: on the unit-cube Poiseuille problem, the two-level method fails at $R_i$2 subdomains because the coarse factorization becomes infeasible, whereas the three-level method remains viable (Köhler et al., 6 Aug 2025).

For incompressible Navier–Stokes, overlapping Schwarz appears both as a monolithic preconditioner for the full Jacobian and as an approximate inverse inside block preconditioners such as PCD, LSC, and SIMPLE. The paper comparing these choices finds that monolithic preconditioners are typically more robust than incomplete block factorizations, and that the new intermediate coarse space GDSW$R_i$3 often gives a favorable compromise between GDSW robustness and RGDSW coarse size. In the reported tests, the monolithic approach is markedly less sensitive to Reynolds number and CFL number than the block alternatives, while PCD is the strongest block preconditioner among those examined (Heinlein et al., 19 Jun 2025).

5. Theory, condition numbers, and spectral interpretation

A substantial part of the literature is devoted to proving that two-level overlap restores scalability and to quantifying how the bounds depend on overlap width, subdomain size, discretization order, or coefficient contrast. For fourth-order variational inequalities, one-level overlap is proved non-scalable, with stable decomposition constant $R_i$4, whereas the two-level method has convergence depending on $R_i$5 and $R_i$6 only. The crucial new tool is a nonlinear positivity-preserving coarse interpolation operator, needed because standard linear coarse interpolation does not preserve feasibility under inequality constraints (Park, 2023).

For three-dimensional vector field problems discretized with lowest-order Nédélec and Raviart–Thomas elements, a new analysis based on Hiptmair–Pechstein regular decompositions yields condition-number bounds linear in the relative overlap,

$R_i$7

without topological assumptions on the domain or subdomains. This replaces older arguments based on more topology-sensitive Helmholtz decompositions and sharpens the overlap dependence (Oh et al., 2024).

In the $R_i$8 boundary element setting for the three-dimensional hypersingular operator, the decomposition into a global lowest-order space plus overlapping vertex-patch high-order spaces yields uniform bounds with respect to both mesh size $R_i$9 and polynomial order $A_i=R_i A R_i^T$0:

$A_i=R_i A R_i^T$1

When the global lowest-order solve is replaced by a local multilevel diagonal scaling preconditioner on adaptively refined meshes, the bound remains uniform in $A_i=R_i A R_i^T$2, $A_i=R_i A R_i^T$3, and the number of newest-vertex-bisection refinement steps (Führer et al., 2014).

For highly heterogeneous scalar elliptic problems, adaptive coarse-space enrichment by local interface eigenfunctions gives robustness with respect to coefficient jumps once all sufficiently small generalized eigenvalues are represented in the coarse space (Eikeland et al., 2016). This complements the algebraic coarse-space results for GDSW, RGDSW, and AMS, which show that coarse-space design is the decisive mechanism by which two-level overlap acquires robustness to heterogeneity (Alves et al., 2024).

A different theoretical perspective comes from generalized locally Toeplitz analysis of additive Schwarz iterations. In that framework, unrestricted overlapping BAS has a distorted symbol,

$A_i=R_i A R_i^T$4

so that the preconditioned operator no longer clusters ideally at $A_i=R_i A R_i^T$5. Restricted additive Schwarz restores the partition-of-unity structure and recovers

$A_i=R_i A R_i^T$6

This provides a spectral explanation for the common observation that unrestricted additive overlap is not an effective stand-alone iteration in the overlapping case, whereas restricted variants behave as proper preconditioners (Rifqui et al., 4 Feb 2026).

6. Generalizations, implementations, and emerging application domains

Overlapping additive Schwarz methods now appear far beyond their classical linear finite element setting. One nonlinear reformulation replaces the original global PDE by a Schwarz-overlapping problem posed directly on independently discretized subdomains with no shared unknowns:

$A_i=R_i A R_i^T$7

This yields Newton–Krylov–Schwarz (NKS) when linearization precedes preconditioning, and Schwarz–Newton–Krylov (SNK) when the nonlinear problem itself is preconditioned through

$A_i=R_i A R_i^T$8

The formulation avoids restrictive updates, requires communication only through interface interpolation, and extends naturally to a two-level FAS correction (Aiton et al., 2019).

In isogeometric analysis, overlapping additive Schwarz is used on overlapping subdomains with nonmatching meshes, motivated by CSG assemblies, multi-patch CAD geometries, and local refinement. Trace, approximation, and extension operators transfer interface data across independently meshed spline patches, and a Chimera-style zooming strategy provides local refinement without globally refining the tensor-product grid. The reported computations include distorted geometries, singular corner problems, and multi-patch three-dimensional examples (Bercovier et al., 2015).

A more recent generalization reinterprets scientific machine learning training as a nonlinear system preconditioned by a two-level overlapping additive Schwarz method. Here the “subdomains” are overlapping groups of network layers or parameters, the coarse space is built from the first layer of each subdomain, and recombination uses a sequential synchronization strategy rather than simple averaging. The reported experiments cover physics-informed neural networks and operator learning, including Burgers, diffusion-advection, Allen–Cahn, anisotropic Poisson, Helmholtz, and advection problems (Lee et al., 2024).

On the implementation side, overlapping additive Schwarz is represented strongly in algebraic software stacks. FROSch in Trilinos/ShyLU is used for two-level, three-level, and monolithic GDSW-type methods and is coupled to FEATFLOW for incompressible flow and to NGSolve for reduced-integration Stokes discretizations (Köhler et al., 6 Aug 2025, Cumaru et al., 1 Dec 2025). Helmholtz implementations combine PETSc and HPDDM with local factorizations by MUMPS, while geophysical ORAS and substructured optimized Schwarz methods are built on Gmsh, GmshFEM, and GmshDDM workflows (Martin et al., 20 Jun 2025). These implementations reinforce a practical conclusion repeated across problem classes: overlap is valuable for local coupling, but scalability ultimately depends on coarse-space design, restricted assembly when double counting matters, and, at extreme scale, the ability to push Schwarz recursion deeper into the hierarchy rather than relying on a single direct coarse factorization.