FROSch: Fast & Robust Overlapping Schwarz

• FROSch is a scalable parallel domain decomposition framework that implements overlapping Schwarz preconditioners using both one- and two-level methods within Trilinos.
• It employs interface-based coarse spaces such as GDSW and RGDSW to enhance scalability and robustness for high-contrast, multiphysics, and algebraically challenging problems.
• FROSch supports multilevel, GPU, and nonlinear extensions, making it versatile for applications in elasticity, fluid mechanics, and coupled phenomena.

Searching arXiv for FROSch and closely related overlapping Schwarz papers to ground the article in current literature. Fast and Robust Overlapping Schwarz (FROSch) is a parallel domain decomposition framework for overlapping Schwarz preconditioning, described in the literature as a package inside Trilinos and, in particular, inside ShyLU, with support for one- and two-level methods, GDSW- and RGDSW-type coarse spaces, and later extensions to multilevel, monolithic saddle-point, GPU, ORAS, and nonlinear formulations (Yamazaki et al., 2023, Köhler et al., 6 Aug 2025, Heinlein et al., 2024, Heinlein et al., 2024). At its core, FROSch realizes additive overlapping Schwarz operators that combine local overlapping subdomain solves with a coarse correction built from interface-based basis functions extended by discrete harmonic or energy-minimizing solves. Closely related work on algebraic coarse spaces, high-contrast multiscale basis construction, vector-field Schwarz theory, and virtual-element coarse spaces has enlarged the conceptual scope of what counts as a FROSch-style method, even when the name itself is not used explicitly (Alves et al., 2024, Wang et al., 2019, Oh et al., 2024, Aguilar-Pineda et al., 8 Dec 2025).

1. Definition and position within domain decomposition

FROSch is presented as a framework or library for scalable overlapping Schwarz methods rather than as a single fixed preconditioner. In the GPU study, it is introduced as a package in Trilinos implementing GDSW-type two-level overlapping Schwarz preconditioners for CPU and GPU clusters (Yamazaki et al., 2023). In the monolithic fluid study, it is described as a framework inside ShyLU, itself part of Trilinos, providing additive overlapping Schwarz preconditioners, GDSW coarse spaces, multilevel support, and algebraic construction of local matrices, restriction and interpolation operators, coarse problems, and overlap information (Köhler et al., 6 Aug 2025). The deal.II coupling paper makes the same point from the operator side: FROSch is a Schwarz framework able to build one-level and two-level operators and to combine Schwarz operators additively or multiplicatively (Heinlein et al., 2024).

This framing places FROSch within the overlapping branch of domain decomposition rather than the substructuring branch represented by BDDC or FETI-DP. The common ingredients are an overlapping cover of the domain, local restriction and prolongation operators, local subdomain solves, and a low-dimensional coarse correction. In published FROSch work, the coarse level is usually based on extension-based GDSW or RGDSW spaces, but later papers broaden this to monolithic saddle-point coarse spaces, ORAS local operators, nonlinear coarse corrections, and multilevel recursion (Heinlein et al., 19 Jun 2025, Köhler et al., 6 Aug 2025, Heinlein et al., 2024).

A recurring distinction in the literature is between one-level and two-level overlapping Schwarz. One-level methods retain only the sum of local overlapping subdomain solves and are repeatedly described as not scalable as the number of subdomains grows. Two-level methods add a coarse correction and thereby provide the global communication path missing from purely local overlap (Köhler et al., 6 Aug 2025, Heinlein et al., 2024, Cumaru et al., 1 Dec 2025).

2. Algebraic form of the preconditioner

The standard two-level overlapping Schwarz operator used throughout the FROSch literature has the form

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

or equivalently

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

with $A_i=R_i A R_i^T$ the local overlapping subdomain matrices and $A_0=\Phi^T A\Phi$ the coarse operator (Alves et al., 2024, Yamazaki et al., 2023). This form is used for scalar elliptic systems, elasticity, reduced-integration Stokes discretizations, and algebraic coarse-space comparisons (Yamazaki et al., 2023, Cumaru et al., 1 Dec 2025, Alves et al., 2024).

A defining feature of GDSW-type FROSch is that the coarse basis is obtained from interface traces and then extended into subdomain interiors by an energy-minimizing or discrete harmonic solve. After ordering unknowns into interior and interface blocks,

$$A = \begin{bmatrix} A_{II} & A_{I\Gamma}\ A_{\Gamma I} & A_{\Gamma\Gamma} \end{bmatrix},$$

the coarse basis satisfies

$$\Phi = \begin{bmatrix} -A_{II}^{-1}A_{I\Gamma}\ I \end{bmatrix}\Phi_\Gamma,$$

so the coarse space is determined entirely by the choice of interface traces $\Phi_\Gamma$ (Alves et al., 2024, Yamazaki et al., 2023, Heinlein et al., 2023). For scalar diffusion, the traces are restrictions of the null space of the global Neumann problem; for vector problems such as elasticity they are multiplied by translational, and in some settings rotational, null-space modes (Yamazaki et al., 2023, Bevilacqua et al., 27 Feb 2025).

The same algebraic template extends to coupled saddle-point systems. For the monolithic Stokes and Navier–Stokes operators,

$$\mathcal A=\begin{bmatrix} A & B^T\ B & 0 \end{bmatrix},$$

the preconditioner becomes

$$\mathcal{B}_{\mathrm GDSW}=\Phi \mathcal A_0^{-1}\Phi^T+\sum_{i=1}^N \mathcal R_i^T \mathcal A_i^{-1}\mathcal R_i,$$

with block restrictions in velocity and pressure and a coarse basis matrix $\Phi$ containing both components (Köhler et al., 6 Aug 2025, Heinlein et al., 19 Jun 2025). This monolithic formulation is central to later fluid papers because the local and coarse problems preserve the full saddle-point coupling instead of approximating only a velocity block or Schur complement factor.

The ORAS extension modifies the first level rather than the coarse space. In that case the local matrices $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$0 are replaced by optimized local operators $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$1 assembled with Robin or impedance-type transmission conditions: $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$2 The deal.II coupling paper emphasizes that this step is no longer purely algebraic, because the optimized local matrices must be assembled separately from geometric interface data (Heinlein et al., 2024).

3. Coarse-space families

The literature around FROSch is largely a literature on coarse-space design. Standard FROSch papers treat GDSW and RGDSW as the reference family, while later work introduces intermediate, monolithic, algebraic multiscale, and energy-minimizing multiscale variants (Yamazaki et al., 2023, Alves et al., 2024, Heinlein et al., 19 Jun 2025, Wang et al., 2019).

Coarse-space family	Interface construction	Role reported in the literature
GDSW	Vertices, edges, and faces with null-space restrictions	Standard FROSch coarse space (Yamazaki et al., 2023)
RGDSW	Reduced interface components or nodal equivalence classes with partition-of-unity weighting	Smaller coarse space, often weaker robustness (Alves et al., 2024, Cumaru et al., 1 Dec 2025)
GDSW$$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$3	Vertices with adjacent edges grouped, faces kept separate	Intermediate space between GDSW and RGDSW (Heinlein et al., 19 Jun 2025)
AMS / algebraic MsFEM	Recursive edge/face traces, then harmonic extension	Algebraic multiscale alternative in the same two-level architecture (Alves et al., 2024)
Energy-minimizing multiscale basis	Local generalized eigenproblems plus constrained energy minimization	FROSch-related high-contrast coarse-space design (Wang et al., 2019)

In classical GDSW, interface components are faces, edges, and vertices, and the interface basis functions form a partition of unity. RGDSW reduces coarse dimension by merging interface components according to nodal equivalence classes and using inverse-multiplicity-type scaling. The computational study of algebraic coarse spaces interprets GDSW as essentially the standard FROSch coarse space and RGDSW as a reduced-dimension FROSch-type alternative (Alves et al., 2024). The GPU paper and the direct-solver comparison both use reduced GDSW spaces in practice, reflecting the persistent tradeoff between coarse-space richness and coarse-problem cost (Yamazaki et al., 2023, Heinlein et al., 2023).

A more specialized development is GDSW$$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$4, introduced for monolithic Navier–Stokes preconditioning. It keeps faces separate, but groups each vertex together with its adjacent edges, yielding a space larger than RGDSW and smaller than full GDSW. The reported motivation is to preserve much of the iteration-count benefit of GDSW while reducing setup and coarse-solve cost. In the cited Navier–Stokes experiments, the combination “velocity: GDSW$$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$5, pressure: RGDSW” is repeatedly identified as particularly effective (Heinlein et al., 19 Jun 2025).

A different line of development replaces classical null-space traces by coefficient-aware algebraic or multiscale traces. The algebraic coarse-space study places AMS, GDSW, and RGDSW inside the same two-level additive Schwarz architecture, differing only in the choice of interface traces $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$6 before harmonic extension (Alves et al., 2024). The multiscale coarse-basis paper goes further: local generalized eigenproblems identify low-energy modes, and the final basis functions are obtained by constrained energy minimization rather than by direct partition-of-unity injection. Although that paper does not use the term FROSch, its overlap, partition-of-unity, spectral selection, and robustness philosophy is explicitly described as closely aligned with modern fast and robust overlapping Schwarz methods (Wang et al., 2019).

For coupled multiphysics, several papers argue that different fields need different coarse spaces. Thermo-elastoplastic laser beam welding uses different GDSW variants for temperature and displacement and reports that richer mechanics coarse spaces with reduced thermal coarse spaces are advantageous (Bevilacqua et al., 27 Feb 2025). Monolithic Navier–Stokes preconditioners likewise use different GDSW-type spaces for velocity and pressure (Heinlein et al., 19 Jun 2025). This suggests that, in FROSch practice, the decisive question is often not whether a coarse level exists, but how its interface objects are chosen for each field.

4. Theory, robustness, and known failure modes

The classical theoretical estimate recalled in the GPU paper is

$$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$7

with $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$8 the maximum subdomain diameter, $$M^{-1}=R_0^T A_0^{-1}R_0+\sum_{i=1}^N R_i^T A_i^{-1}R_i,$$9 the overlap width, and $A_i=R_i A R_i^T$0 the fine mesh size (Yamazaki et al., 2023). This is the canonical explanation for why a two-level overlapping Schwarz method can be robust and scalable while a one-level method cannot. Closely related three-dimensional vector-field analysis for Nédélec and Raviart–Thomas discretizations establishes bounds of the form

$A_i=R_i A R_i^T$1

with robustness to topology and nonconvexity rather than to coefficient contrast (Oh et al., 2024).

For high-contrast scalar diffusion, the standard piecewise-polynomial coarse problem is explicitly described as non-robust. The multiscale energy-minimization paper addresses exactly this failure mode by selecting local generalized eigenvectors below a threshold $A_i=R_i A R_i^T$2, constructing energy-minimizing basis functions constrained by these auxiliary modes, proving exponential decay on oversampled patches, and obtaining stable-decomposition constants independent of the coefficient contrast and overlap width, provided oversampling is large enough (Wang et al., 2019). This paper is particularly significant because it isolates a design principle that reappears in later FROSch-adjacent work: detect problematic local modes spectrally, then realize the actual coarse correction by an energy-minimizing extension.

At the same time, published FROSch work makes clear that robustness is not automatic. In the algebraic coarse-space comparison, GDSW and RGDSW are not robust for one high-conductivity inclusion-at-node configuration, while AMS is robust and scalable there; for parallel channel configurations all three coarse spaces yield condition numbers of the order of the contrast (Alves et al., 2024). The chemo-mechanics study reaches a different but related conclusion: fully algebraic black-box FROSch can be used successfully on a monolithic coupled problem, but numerical scalability should not be expected and is not observed in that mode (Kiefer et al., 2022). The direct-solver comparison emphasizes a further limitation: fully algebraic use with an incomplete nullspace sacrifices the theoretical GDSW guarantee, and iteration counts rise with core count accordingly (Heinlein et al., 2023).

These papers collectively correct two recurrent misconceptions. The first is that any GDSW-type coarse space is automatically coefficient-robust; the literature instead shows strong dependence on coefficient geometry, interface traces, and possible adaptive enrichment (Alves et al., 2024, Wang et al., 2019). The second is that “fully algebraic” is synonymous with “numerically scalable”; the chemo-mechanics and direct-solver studies explicitly state the opposite for their settings (Kiefer et al., 2022, Heinlein et al., 2023).

5. Software realization and parallel performance

FROSch is embedded in the Trilinos ecosystem. The deal.II coupling paper describes it as accessible through Stratimikos and compatible with both Epetra and Tpetra through Xpetra, with Tpetra identified as the forward-looking path because it enables Kokkos and KokkosKernels (Heinlein et al., 2024). Several application papers emphasize that the implementation is algebraic in the sense that only the global matrix and the parallel distribution of the degrees of freedom are needed, after which FROSch builds overlap information, local subdomain matrices, restriction and prolongation operators, and coarse matrices (Köhler et al., 6 Aug 2025, Cumaru et al., 1 Dec 2025).

The implementation footprint has broadened substantially. FROSch has been coupled to FEATFLOW for monolithic fluid solvers (Köhler et al., 6 Aug 2025), to deal.II through a Tpetra-based interface supporting both Lagrange and Nédélec elements (Heinlein et al., 2024), to NGSolve through an interface from NGSolve’s Python frontend to Trilinos’ C++ implementation (Cumaru et al., 1 Dec 2025), and to FEDDLib for nonlinear two-level Schwarz methods (Heinlein et al., 2024, Ho et al., 25 Mar 2026). The chemo-mechanics paper uses deal.II and FROSch in fully algebraic mode for monolithic nonlinear deformation–diffusion problems (Kiefer et al., 2022).

Parallel performance studies show both the strength and the cost structure of the framework. On Summit with NVIDIA V100 GPUs, the GPU study reports that solve time was reduced by factors of about $A_i=R_i A R_i^T$3 using GPUs, and that GPU-friendly iterative ILU variants yielded solve-time speedups around $A_i=R_i A R_i^T$4 to $A_i=R_i A R_i^T$5 in weak scaling and up to about $A_i=R_i A R_i^T$6 in a single-node study (Yamazaki et al., 2023). A central implementation result there is that multiple MPI processes per GPU, enabled through NVIDIA MPS, improved both numerical setup and solve times by shrinking local problems and reducing fill-heavy factorization costs (Yamazaki et al., 2023).

For CPU-only direct-solver studies, the picture is different. The comparison of direct solvers in FROSch on elasticity and chemo-mechanics reports good strong scaling from 64 to 1728 cores, with deterioration beyond that because the single-core coarse solve becomes dominant; among the tested solvers, MUMPS is consistently the fastest overall (Heinlein et al., 2023). This coarse-bottleneck diagnosis reappears in the monolithic multilevel fluid work, where the two-level method fails at 32,768 subdomains because the coarse factorization becomes infeasible, while the three-level recursive extension remains solvable and reduces total time once the second-level coarse problem has become too large (Köhler et al., 6 Aug 2025).

A further software development is the addition of ORAS to FROSch. The deal.II interface paper introduces new geometric classes for O(R)AS and emphasizes that optimized local operators $A_i=R_i A R_i^T$7 must be assembled separately from finite-element and interface data. In that sense, ORAS marks a shift from the purely algebraic local-operator model toward a geometry-aware first level, while retaining FROSch’s operator-composition viewpoint (Heinlein et al., 2024).

6. Application domains and ongoing extensions

The application range of FROSch now spans scalar elliptic problems, elasticity, incompressible flow, thermo-mechanics, chemo-mechanics, fluid–structure interaction, Maxwell equations, vector-field problems, and virtual element discretizations. In incompressible fluid mechanics, monolithic two-level and three-level overlapping Schwarz solvers implemented with FROSch and FEATFLOW are reported up to 32,768 MPI ranks for Poiseuille flow and an extrusion die geometry (Köhler et al., 6 Aug 2025). A broader Navier–Stokes study uses FROSch both as a monolithic preconditioner for the full saddle-point operator and as an inner solver inside block preconditioners such as PCD, SIMPLEC, and LSC, concluding that monolithic FROSch preconditioners are generally more robust, especially with respect to viscosity, Reynolds number, and CFL number (Heinlein et al., 19 Jun 2025).

For reduced-integration Stokes discretizations, FROSch is used with an RGDSW coarse space on an SPD velocity-only system, and the reported weak-scaling experiments show that the two-level RGDSW preconditioner has good numerical scalability compared with the one-level method, while setup time stays nearly constant over the tested penalty-parameter range (Cumaru et al., 1 Dec 2025). In thermo-elastoplastic laser beam welding, monolithic two-level overlapping Schwarz with different combinations of GDSW, GDSW$A_i=R_i A R_i^T$8, and RGDSW coarse spaces is implemented in PETSc rather than Trilinos, but the method is explicitly described as a GDSW-family overlapping Schwarz construction for multiphysics saddle-point systems and is therefore directly relevant to FROSch-style design (Bevilacqua et al., 27 Feb 2025). The earlier monolithic chemo-mechanics study, by contrast, uses FROSch in fully algebraic mode and shows that it remains applicable on the tested core counts even though numerical scalability is not observed (Kiefer et al., 2022).

Nonlinear two-level Schwarz is the most substantial recent conceptual extension. The 2024 FROSch paper introduces one-level, additive two-level, and hybrid two-level nonlinear Schwarz solvers on top of FROSch, with local nonlinear corrections $A_i=R_i A R_i^T$9, a nonlinear coarse correction $A_0=\Phi^T A\Phi$0, and exact tangents of the nonlinear Schwarz map (Heinlein et al., 2024). The 2026 follow-up reports that the same framework, using FROSch GDSW-type coarse spaces and modifications of those coarse spaces, scales exceptionally well up to 9,000 subdomains and is more robust than standard Newton–Krylov–Schwarz for high-Reynolds-number lid-driven cavity flow and large-deformation Neo-Hookean elasticity (Ho et al., 25 Mar 2026). One of the striking conclusions of that paper is that apparently small modifications of RGDSW or MsFEM interface functions can materially change nonlinear robustness, especially near Dirichlet boundaries (Ho et al., 25 Mar 2026).

Related research continues to expand the conceptual perimeter of FROSch. The vector-field analysis for $A_0=\Phi^T A\Phi$1 and $A_0=\Phi^T A\Phi$2 problems supplies topology-robust stable-decomposition theory with linear dependence on relative overlap (Oh et al., 2024). The VEM paper develops a two-level additive overlapping Schwarz preconditioner on irregular three-dimensional polyhedral decompositions, numerically observing weak dependence on the number of subdomains and on $A_0=\Phi^T A\Phi$3, and approximately linear dependence on $A_0=\Phi^T A\Phi$4, with a geometric vertex-based interface-harmonic coarse space that is close in spirit to reduced GDSW constructions (Aguilar-Pineda et al., 8 Dec 2025). This suggests that FROSch is best understood not only as a concrete Trilinos package, but also as a continuing research program centered on overlapping local solves, interface-based coarse spaces, and the search for coarse corrections that remain robust under stronger coupling, more irregular discretizations, and more demanding parallel scales.