Fully Algebraic Alternating Schwarz Method

• The method is a domain decomposition approach built entirely from algebraic data like sparse matrices, restriction operators, and graph adjacency information.
• It employs sequential local corrections and algebraic coarse spaces to enhance convergence, functioning as both a solver and a preconditioner in various Krylov frameworks.
• The framework supports overlapping, non-overlapping, and multilevel variants, balancing contraction, symmetry, and parallel efficiency in different SPD, nonsymmetric, and indefinite settings.

A fully algebraic alternating Schwarz method is a multiplicative domain decomposition procedure in which local corrections are applied sequentially and every ingredient of the method is constructed from algebraic data such as the assembled sparse matrix, its sparsity graph, Boolean restriction operators, algebraic interface transfers, and optionally algebraic coarse spaces. In overlapping form, the method updates the residual after each local subdomain solve; in non-overlapping form, it appears as an interface fixed-point iteration, typically of Dirichlet–Neumann type. The framework encompasses one-level, two-level, and multilevel variants, and it is used both as a stationary solver and as a preconditioner for CG, Bi-CGstab, GMRES, or FGMRES, depending on symmetry and definiteness properties (Holst et al., 2010, Daas et al., 2024, Sambataro et al., 17 Mar 2026).

1. Algebraic formulation of alternating Schwarz

For the SPD setting emphasized in the symmetrization theory, let $A \in L(\mathcal H,\mathcal H)$ be SPD on a finite-dimensional real Hilbert space $\mathcal H$ with inner product $(\cdot,\cdot)$. The associated $A$-inner product and $A$-norm are

$(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$

and the $A$-adjoint of an operator $M$ is

$M^* = A^{-1} M^T A.$

Subdomains or blocks are represented purely algebraically by restriction operators $R_i$, with prolongations $\mathcal H$0. Local matrices are built either variationally, $\mathcal H$1, or by direct algebraic assembly on the local index set (Holst et al., 2010).

The additive reference operator is

$\mathcal H$2

whereas the alternating, or multiplicative, Schwarz iteration is described by a product error propagator. In its generic two-level form,

$\mathcal H$3

This representation makes explicit that alternating Schwarz applies the most recent residual after each local correction, unlike additive Schwarz, which sums independent local contributions (Holst et al., 2010).

In the classical overlapping two-subdomain picture, the sequential character is visible in the updates

$\mathcal H$4

For $\mathcal H$5 subdomains, one sweeps $\mathcal H$6 analogously. This typically converges in fewer iterations than additive Schwarz used inside Krylov, but the sequential nature reduces parallelism; parallel variants use coloring to update independent subdomains concurrently (Kiefer et al., 2022).

2. Fully algebraic construction from a sparse matrix

A fully algebraic construction begins from the adjacency graph of the global sparse matrix. Vertices are partitioned into subdomains, and overlap is defined by expanding each partition by graph-distance layers of neighbors. For each subdomain $\mathcal H$7, one forms a restriction $\mathcal H$8, a prolongation $\mathcal H$9, and a local matrix $(\cdot,\cdot)$0 or a directly assembled local operator. The local inverse may be exact or approximate: the recorded choices include exact factorization, ILU, Gauss–Seidel sweeps, and symmetric variants when SPD conditions are desired (Holst et al., 2010).

A standard one-application multiplicative sweep is purely residual-driven. Starting from a residual $(\cdot,\cdot)$1, one initializes $(\cdot,\cdot)$2 and $(\cdot,\cdot)$3, then for $(\cdot,\cdot)$4 computes the local residual $(\cdot,\cdot)$5, the local correction $(\cdot,\cdot)$6, injects $(\cdot,\cdot)$7 into the global vector, and updates the residual by subtracting $(\cdot,\cdot)$8. Optional coarse correction and optional backward sweep are inserted in the same pattern. This implementation realizes the product operator without ever forming the product explicitly (Holst et al., 2010).

The same algebraic ingredients support two-level methods. In the sparse-matrix literature, a coarse restriction $(\cdot,\cdot)$9 is assembled algebraically, the coarse operator is taken as $A$0, and multiplicative variants insert the coarse correction either before, after, or between local sweeps. For general sparse matrices, the construction is explicitly designed to apply to indefinite and non-self-adjoint operators as well as to HPD cases (Daas et al., 2022, Daas et al., 2024).

The phrase “fully algebraic mode” is also used in large-scale coupled applications. In the FROSch implementation within Trilinos, the preconditioner is constructed from the monolithic system matrix without making explicit use of the problem structure. In this mode, only minimal nullspace information is provided and interior/interface partitions are detected algebraically; the same $A$1, $A$2, $A$3, and $A$4 can then be sequenced to obtain an alternating Schwarz solver rather than an additive Schwarz preconditioner (Kiefer et al., 2022).

3. Symmetry, positive definiteness, and minimal symmetrization

A central issue for alternating Schwarz is whether the induced preconditioner is self-adjoint positive definite in the $A$5-inner product. For the product form

$A$6

sufficient conditions for the corresponding $A$7 to be symmetric and positive are: $A$8, $A$9, $A$0, and $A$1 non-negative. In multiplicative domain decomposition, the analogous sufficient conditions are $A$2 with $A$3, $A$4, $A$5, contraction of the forward sweep in the $A$6-norm, and non-negativity of $A$7. Under these hypotheses, $A$8 is $A$9-self-adjoint and positive, so CG is applicable (Holst et al., 2010).

The same theory provides the standard norm-based condition-number estimate

$(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$0

together with the CG contraction bound

$(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$1

These formulas explain why multiplicative Schwarz can serve as a preconditioner even when the underlying stationary method is non-variational or non-convergent, provided the sufficient positivity conditions for the preconditioned operator are met (Holst et al., 2010).

The same paper also isolates the penalty associated with excessive symmetrization. For any error propagator $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$2,

$(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$3

This shows that the symmetrized propagator $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$4 can have strictly worse contraction than the nonsymmetric $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$5. On that basis, the paper conjectures that enforcing minimal symmetry gives the best results when combined with CG: the nonsymmetric factors are retained, and symmetry is imposed only at the product level when needed to satisfy the CG requirement (Holst et al., 2010).

The same analysis distinguishes Krylov accelerators. For SPD $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$6, CG requires $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$7 to be $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$8-self-adjoint and positive. By contrast, Bi-CGstab does not require application of the adjoint of the preconditioner, and the paper argues and demonstrates empirically that nonsymmetric linear preconditioners are advantageous with Bi-CGstab. In practical terms, forward-only multiplicative sweeps or nonsymmetric local smoothers may be preferable when Bi-CGstab or GMRES is available (Holst et al., 2010).

4. Two-level and multilevel algebraic coarse spaces

The development of fully algebraic alternating Schwarz is inseparable from the development of fully algebraic coarse spaces. One important construction is the spectral coarse space for general sparse matrices built from an algebraic decomposition of the graph into nonoverlapping sets $(u,v)_A = (Au,v), \qquad \|u\|_A = (u,u)_A^{1/2},$9, one-ring overlaps $A$0, local matrices $A$1, and a partition of unity $A$2. A local SPSD splitting $A$3 is obtained by overlap lumping, and the coarse basis is selected from the generalized eigenproblem

$A$4

With threshold $A$5, the coarse restriction is assembled as

$A$6

For HPD diagonally dominant matrices, the paper proves

$A$7

and numerical experiments report efficiency especially for non-self-adjoint operators (Daas et al., 2022).

A related fully algebraic two-level framework for sparse matrices defines local harmonic projections $A$8, local SPSD splittings $A$9, and reduced coarse spaces from the generalized eigenproblem

$M$0

In the SPD case, the resulting additive preconditioner satisfies explicit bounds such as

$M$1

and, for the reduced coarse space,

$M$2

For general matrices, the coarse space is constructed from the image of the local operator $M$3, using left singular vectors associated with large singular values. The paper analyzes the additive form, then states that the same local and coarse solvers can be reused in a multiplicative sweep with error propagator

$M$4

(Daas et al., 2024).

For high-contrast SPD diffusion, another fully algebraic construction replaces standard AGDSW ingredients by edge-local transfer and Dirichlet eigenvalue problems that rely solely on local Dirichlet matrices extracted from the assembled global matrix. The coarse space

$M$5

combines vertex functions, constant edge functions, Dirichlet edge modes, and transfer edge modes. The resulting two-level Schwarz preconditioner has the contrast-independent bound

$M$6

with

$M$7

The same algebraic coarse space is then used unchanged in a symmetric multiplicative Schwarz variant (Heinlein et al., 2022).

In multilevel least-squares AMG, local spectral information is coupled directly to overlapping Schwarz smoothers. For systems $M$8, the coarse basis is defined by local generalized eigenproblems derived from an SPSD splitting of $M$9, while smoothing is carried out by one restricted additive Schwarz sweep and one transpose-RAS sweep. The method is explicitly designed to remain fully algebraic, naturally recursive, and compatible with nonoverlapping coarse basis functions (Southworth et al., 7 Jan 2026).

5. Interface formulations, relaxation, and algebraic acceleration

A non-overlapping alternating Schwarz method is naturally expressed as an interface fixed-point iteration. For two subdomains, with trace operators $M^* = A^{-1} M^T A.$0, interface projections $M^* = A^{-1} M^T A.$1, and discrete Neumann operators $M^* = A^{-1} M^T A.$2, the Dirichlet-to-Neumann and Neumann-to-Dirichlet maps are

$M^* = A^{-1} M^T A.$3

and the multiplicative Dirichlet–Neumann iteration is rewritten as

$M^* = A^{-1} M^T A.$4

In the one-dimensional Laplace model, the relaxed iteration reduces to

$M^* = A^{-1} M^T A.$5

so convergence requires $M^* = A^{-1} M^T A.$6, and the optimal parameter is $M^* = A^{-1} M^T A.$7, yielding convergence in two iterations in the reported experiments (Sambataro et al., 17 Mar 2026).

The same fixed-point framework supports algebraic acceleration. The paper derives a vector Aitken parameter

$M^* = A^{-1} M^T A.$8

and an Anderson step based on the constrained least-squares problem

$M^* = A^{-1} M^T A.$9

It further introduces “Anderson with memory adaptation.” The reported conclusion is that Aitken-accelerated Schwarz is the best approach in terms efficiency and robustness when considering two sub-domain DDs, whereas Anderson-accelerated Schwarz is the method of choice in larger multi-domain setting (Sambataro et al., 17 Mar 2026).

In contact dynamics, the same non-overlapping Dirichlet–Neumann logic is recast as an algebraic interface problem. After Newmark discretization, each body has effective tangent $R_i$0, and static condensation yields the interface Schur complement

$R_i$1

The interface projections are built algebraically through matrices of the form $R_i$2, so different meshes, time integrators, and time steps can be coupled without explicit contact constraints. In the reported benchmark, Schwarz DN-ASM achieves near-constant total energy with maximum error $R_i$3, whereas conventional methods show up to $R_i$4 loss immediately after impact; the reported 3D runs require on average $R_i$5 DN iterations depending on the coupling (Mota et al., 2023).

A further algebraic acceleration of interface traces is provided by ARAS. Starting from a RAS iteration, interface snapshots are collected, an SVD basis $R_i$6 is formed, and a reduced trace-transfer operator $R_i$7 is estimated algebraically. The preconditioner is then

$R_i$8

Its spectral analysis gives

$R_i$9

The same reduced interface acceleration is stated to be embeddable in a sequential alternating Schwarz sweep, yielding a fully algebraic alternating variant with learned interface correction (Dufaud et al., 2013).

6. Implementations, applications, and limitations

In large-scale multiphysics computations, fully algebraic Schwarz components are often available first in additive form and only then reorganized multiplicatively. In the hydrogel swelling study based on deal.II, p4est, and FROSch, the preconditioner is assembled in fully algebraic mode from the monolithic coupled matrix. Strong and weak scaling are reported up to 512 cores. For the 691,635-DOF mechanical-induced diffusion case with fully algebraic “black-box” GDSW and overlap $\mathcal H$00, the average number of Krylov iterations per Newton step grows from $\mathcal H$01 at 64 cores to $\mathcal H$02 at 512 cores. The same source states that numerical scalability cannot be expected, and is not observed, for the fully algebraic mode, because the coarse space is built without knowledge of the block physics; nevertheless, the fully algebraic mode is still preferable since a faster time to solution is achieved (Kiefer et al., 2022).

For sparse linear systems more generally, fully algebraic two-level Schwarz constructions have been reported on SPD, symmetric indefinite, nonsymmetric, and saddle-point matrices. The sparse-matrix coarse spaces of the 2022 and 2024 papers are compared against multigrid and domain decomposition baselines on SuiteSparse-type tests, convection–diffusion, Stokes, elasticity, biharmonic, and heterogeneous diffusion. The common practical rule is that right-preconditioned GMRES or FGMRES is preferred for general $\mathcal H$03, while CG is reserved for HPD or SPD cases with symmetric application of the preconditioner (Daas et al., 2022, Daas et al., 2024).

A different implementation trade-off appears in least-squares AMG with overlapping Schwarz smoothers. There, multiplicative Schwarz is explicitly avoided in order to preserve parallelism and avoid ordering effects. Instead, one uses one RAS pre-smoothing sweep and one RAS–T post-smoothing sweep, which the paper describes as analogous to forward/backward Gauss–Seidel in producing an SPD preconditioner. On extremely anisotropic heat conduction operators arising in magnetic confinement fusion, AMG and smoothed aggregation fail to reduce the residual by $\mathcal H$04 in 1000 iterations, while the least-squares AMG-DD method remains robust up to $\mathcal H$05 (Southworth et al., 7 Jan 2026).

The main limitations of fully algebraic alternating Schwarz follow directly from the algebraic design. Multiplicative sweeps strengthen local contraction but reduce concurrency; coloring partly restores parallelism, but the number of sequential stages remains tied to the coloring number. Robustness depends on overlap depth, local solver quality, sweep ordering, and spectral thresholds such as $\mathcal H$06, $\mathcal H$07, $\mathcal H$08, and $\mathcal H$09. The strongest theory is still concentrated on SPD or HPD settings, high-contrast scalar elliptic models, and particular least-squares factorizations, whereas nonsymmetric and indefinite cases are often supported primarily by numerical experiments rather than universal bounds (Holst et al., 2010, Heinlein et al., 2022, Daas et al., 2024).

Taken together, these developments define a broad class of fully algebraic alternating Schwarz methods: sequential local or interface corrections derived solely from matrix-level data, enriched when necessary by algebraic coarse spaces or reduced interface models, and combined with the minimum symmetry required by the intended Krylov accelerator.