Non-overlapping Schwarz Alternating Method (NO-SAM)

• Non-overlapping Schwarz Alternating Method (NO-SAM) is a domain decomposition technique that splits a global problem into non-overlapping subdomains solved iteratively via exchanged interface data.
• It employs various transmission conditions such as Dirichlet–Neumann, Robin–Robin, and impedance formulations to ensure stability and enhance convergence across subdomain boundaries.
• The modular NO-SAM framework effectively couples full-order models with reduced or neural surrogates, enabling versatile applications in solid mechanics, wave propagation, and manifold geodesic computations.

The non-overlapping Schwarz Alternating Method (NO-SAM) is a family of domain-decomposition algorithms in which a global problem is split into non-overlapping subproblems, each subproblem is solved with interface data supplied by its neighbors, and interface traces are iteratively updated until compatibility is achieved across subdomain boundaries. In the classical PDE setting, NO-SAM alternates Dirichlet, Neumann, or Robin transmission conditions across interfaces; in more recent formulations it also appears as an abstract scattering/exchange iteration, a substructured interface solver for Helmholtz problems, a coupling mechanism between full-order and reduced-order models, a contact algorithm without explicit contact constraints, a trace iteration for HDG discretizations, a learning-based interface scheme for physics-constrained neural networks, and even a time-domain decomposition method for geodesic computation on manifolds (Pechstein, 2022, Barnett et al., 2022, Sutti et al., 2024).

1. Historical and abstract formulation

Non-overlapping Schwarz methods with generalized Robin transmission conditions were originally introduced by B. Després for time-harmonic wave propagation problems, and the subsequent theory was extended by P. L. Lions, Collino, Ghanemi, Joly, Claeys, Parolin, Pechstein, and others. In the abstract variational framework developed for Robin–Schwarz methods, one starts from a global invertible operator $\widehat A:\widehat U\to\widehat U^*$, decomposes the problem into local spaces $U_i$ and operators $A_i$, and introduces trace spaces $\Lambda_i$ with a global exchange operator $\mathcal X:\Lambda\to\Lambda$ satisfying $\mathcal X^2=I$. Continuity across interfaces is encoded by the identity

$\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$

and the method can be rewritten in scattering form as

$(I-\mathcal X^T S)\lambda=d,$

with $S$ the scattering operator and $\lambda$ the impedance trace. The corresponding Schwarz iteration is then a damped Richardson iteration on the interface problem (Pechstein, 2022).

This abstract view is important because it unifies continuous and discrete formulations, applies to scalar Helmholtz, Maxwell, dual $U_i$0 Helmholtz, conforming finite element discretizations, and coercive problems, and makes clear that NO-SAM is fundamentally an interface algorithm rather than a specific boundary-condition recipe (Pechstein, 2022). A related line of work for heterogeneous Helmholtz scattering problems with arbitrary non-overlapping partitions and cross-points uses the Multi-Trace Formalism. There, the interface unknown is an incoming-Robin skeleton trace $U_i$1, the exchange operator is a nonlocal projector $U_i$2 built from Yukawa layer potentials, and the interface equation takes the form

$U_i$3

This yields an “identity + contraction” formulation and removes the usual restriction excluding cross-points (Claeys, 2019).

A notable conceptual extension appears in geodesic computation on Riemannian manifolds. There, the “domain” is the time interval $U_i$4, subdomains are adjacent time windows $U_i$5, and interface unknowns are points $U_i$6 on the manifold. In that setting, Noakes’ leapfrog algorithm is reinterpreted as a classical non-overlapping Schwarz alternating method in the time domain, showing that the Schwarz viewpoint is not limited to spatial PDE decompositions (Sutti et al., 2024).

2. Core iteration and interface data exchange

At its most concrete, NO-SAM alternates local solves on non-overlapping subdomains while passing interface data between neighbors. For a two-subdomain solid-mechanics decomposition $U_i$7 with interface $U_i$8, a standard Dirichlet–Neumann sweep solves on $U_i$9 with Dirichlet data on $A_i$0, computes the interface traction $A_i$1, solves on $A_i$2 with that traction as Neumann data, updates the Dirichlet guess by relaxation, and checks convergence by relative changes in displacement, velocity, and acceleration (Barnett et al., 2022). In fixed-point notation, the same idea can be written as

$A_i$3

where $A_i$4 is the interface trace, $A_i$5 is assembled from a Dirichlet-to-Neumann map and a Neumann-to-Dirichlet map, and relaxation gives

$A_i$6

This formulation is central in recent analyses of relaxation, Aitken acceleration, and Anderson acceleration (Sambataro et al., 17 Mar 2026).

In the manifold-geodesic formulation, the subdomain solver is the midpoint of the geodesic joining two neighboring interface values,

$A_i$7

and the Jacobi-type NO-SAM update is

$A_i$8

while the Gauss–Seidel variant replaces one argument by the most recent update. The leapfrog algorithm corresponds to the Gauss–Seidel variant (Sutti et al., 2024).

The same alternating pattern persists in specialized formulations. In contact mechanics, each body is a non-overlapping domain, non-penetration is enforced iteratively rather than by contact constraints, and one alternates a Dirichlet step on one body with a Neumann step on the other, using transfer operators for displacement, velocity, acceleration, and traction (Mota et al., 2023). In HDG-based Poisson solvers, two NO-SAM variants are described: a Neumann–Neumann algorithm driven by a single interface trace and a trace–flux alternating algorithm driven by both trace and flux interface unknowns (Oikawa, 2023). In large-scale Helmholtz computations, subdomains solve local systems with impedance-type transmission data $A_i$9 and update outgoing interface fields through a global operator $\Lambda_i$0, yielding a fixed-point problem $\Lambda_i$1 (Martin et al., 20 Jun 2025).

3. Transmission conditions and operator choices

The performance of NO-SAM is strongly shaped by the transmission condition used on interfaces. Alternating Dirichlet–Neumann transmission remains the canonical form in dynamic solid mechanics, ROM–FOM coupling, and contact (Barnett et al., 2022, Mota et al., 2023). For linear elastic wave coupling with full-order and Operator Inference models, the same pattern is written in $\Lambda_i$2 notation, with Dirichlet–Neumann corresponding to $\Lambda_i$3 and $\Lambda_i$4 (Rodriguez et al., 6 Sep 2025).

Robin–Robin transmission generalizes this by imposing mixed conditions on both sides. In the OpInf coupling study, each subdomain solves

$\Lambda_i$5

and the coupled trace satisfies

$\Lambda_i$6

That study reports that optimized Robin–Robin coupling can dramatically accelerate convergence relative to alternating Dirichlet–Neumann, but that poor parameter choices can induce localized interface oscillations in acceleration and velocity (Rodriguez et al., 6 Sep 2025).

For Helmholtz problems, impedance conditions play the role of Robin transmission. In substructured non-overlapping methods one imposes

$\Lambda_i$7

with zeroth-order $\Lambda_i$8 and second-order $\Lambda_i$9 among the practical options. Parameter tuning of $\mathcal X:\Lambda\to\Lambda$0 is used to balance wave-angle modes and improve convergence (Martin et al., 20 Jun 2025). In closed cavities, however, transmission operators optimized for unbounded domains neglect back-propagating waves and stagnate on propagating modes. Cavity-adapted DtN operators instead involve cotangent or coth symbols, and practical local approximations are built from Taylor, Mittag-Leffler, or Padé expansions (Marsic et al., 2022).

Learning-based NO-SAM variants enlarge the interface model further. In PECANN-based domain decomposition, the generalized interface condition on $\mathcal X:\Lambda\to\Lambda$1 is

$\mathcal X:\Lambda\to\Lambda$2

with $\mathcal X:\Lambda\to\Lambda$3 treated as learnable parameters (Hu et al., 2024). In an earlier physics-constrained neural-network formulation, each subdomain learns its own Robin weight $\mathcal X:\Lambda\to\Lambda$4 in

$\mathcal X:\Lambda\to\Lambda$5

so that the method interpolates between Dirichlet and Neumann exchange and adapts the balance online (Basir et al., 2023).

Formulation	Interface data	Representative setting
Alternating Dirichlet–Neumann	Displacement/velocity/acceleration on one side, traction on the other	Nonlinear solid mechanics, contact (Barnett et al., 2022, Mota et al., 2023)
Robin–Robin / impedance	Mixed trace–flux data with tunable parameters	Wave propagation and FOM–ROM/OpInf coupling (Martin et al., 20 Jun 2025, Rodriguez et al., 6 Sep 2025)
Abstract scattering/exchange	Impedance traces $\mathcal X:\Lambda\to\Lambda$6 or Robin traces $\mathcal X:\Lambda\to\Lambda$7	Robin–Schwarz theory and Multi-Trace Helmholtz (Pechstein, 2022, Claeys, 2019)
Generalized learned interface law	Learnable $\mathcal X:\Lambda\to\Lambda$8 or $\mathcal X:\Lambda\to\Lambda$9	PECANN and physics-constrained neural networks (Hu et al., 2024, Basir et al., 2023)
Midpoint map on manifold traces	Interface points $\mathcal X^2=I$0 on $\mathcal X^2=I$1	Time-domain geodesic computation (Sutti et al., 2024)

4. Convergence theory, relaxation, and acceleration

Several convergence mechanisms coexist in the NO-SAM literature. In the abstract Robin–Schwarz theory, convergence can be established under compactness, under strong absorption or SPD-type assumptions, or through an inf–sup/coercivity argument yielding the geometric rate

$\mathcal X^2=I$2

For the Multi-Trace Helmholtz formulation, $\mathcal X^2=I$3 is a contraction and coercivity of $\mathcal X^2=I$4 yields linear convergence of relaxed Richardson iteration, with optimal relaxation $\mathcal X^2=I$5 in the stated estimate (Pechstein, 2022, Claeys, 2019).

In classical alternating Dirichlet–Neumann settings, theory and numerics emphasize the role of relaxation. For the coupling-focused fixed-point formulation, the relaxed update parameter $\mathcal X^2=I$6 controls the spectral radius of the linearized iteration. Aitken’s $\mathcal X^2=I$7 process adapts $\mathcal X^2=I$8 from secant information, while Anderson acceleration uses the last $\mathcal X^2=I$9 residuals to solve a constrained least-squares problem and form a multi-secant interface update. The reported conclusion is that Aitken-accelerated Schwarz is the best approach in terms efficiency and robustness for two-subdomain decompositions, whereas Anderson-accelerated Schwarz is preferable in larger multi-domain settings; an adaptive Anderson variant freezes memory when the relative interface error stagnates (Sambataro et al., 17 Mar 2026).

Problem-specific accelerations are equally prominent. In manifold geodesic computation, the fixed-point residual is

$\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$0

with block-tridiagonal Jacobian $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$1. A RASPEN-style modified Newton–Schwarz method computes a Newton step in the embedding space, projects it onto the tangent space, and retracts back to the manifold. The expected local behavior is quadratic once iterates enter a sufficiently small neighborhood of the true solution (Sutti et al., 2024). In wave problems, optimized Robin parameters or higher-order transmission operators play the role of acceleration. For 3D Helmholtz with multiple sources, tuned second-order transmission conditions reduce GMRES iterations by 30–50% and bring non-overlapping OSM close to overlapping ORAS while using less memory (Martin et al., 20 Jun 2025). For cavity Helmholtz, the exact cavity DtN operator yields $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$2 modewise in the two-subdomain setting, while cavity-adapted Padé approximations substantially outperform operators designed for unbounded problems (Marsic et al., 2022).

Reported numerical evidence is correspondingly diverse. On the unit sphere $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$3 with $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$4, preconditioned leapfrog converges quadratically in 3–4 iterations and remains essentially independent of endpoint distance for the tested cases, whereas the original leapfrog slows dramatically as the distance approaches $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$5; on the Stiefel manifold, preconditioned leapfrog iteration count is essentially constant in $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$6 (Sutti et al., 2024). In nonlinear solid mechanics, 2–3 Schwarz iterations per time step are observed even for strongly nonlinear problems (Barnett et al., 2022). In contact, 2–5 Schwarz iterations are reported as typical, with no relaxation parameter required in the stated Dirichlet–Neumann formulation (Mota et al., 2023). In HDG, the Neumann–Neumann variant converges for $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$7 with an observed optimum near $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$8, while the trace–flux method shows iteration counts growing like $\mathrm{range}\,R=\ker\bigl((I-\mathcal X)T\bigr),$9 for an $(I-\mathcal X^T S)\lambda=d,$0-strip decomposition (Oikawa, 2023).

5. Major application domains

NO-SAM now spans a notably broad range of applications. In nonlinear geometry, it is used to compute geodesics on Riemannian manifolds by decomposing the endpoint boundary-value problem into smaller geodesic problems over adjacent time windows. Preliminary tests were reported on the unit sphere and on the Stiefel manifold, using closed-form Exp/Log on the sphere and single-shooting plus finite-difference Jacobian–vector products on the Stiefel manifold (Sutti et al., 2024).

In computational solid mechanics, NO-SAM is a modular outer coupling loop for full-order models (FOMs), projection-based reduced-order models (ROMs), and hyper-reduced models. One formulation couples FOMs and ROMs in nonlinear solid mechanics while preserving the Hamiltonian structure through POD/Galerkin projection and using ECSW hyper-reduction to reduce nonlinear evaluation cost from $(I-\mathcal X^T S)\lambda=d,$1 to $(I-\mathcal X^T S)\lambda=d,$2 with $(I-\mathcal X^T S)\lambda=d,$3. The numerical test case is a 1D traveling wave in a Henky hyperelastic bar, with reported reproductive and predictive results for FOM–FOM, FOM–ROM, FOM–HROM, ROM–ROM, and HROM–HROM couplings (Barnett et al., 2022). A later study extends non-overlapping coupling to Operator Inference models and compares alternating Dirichlet–Neumann against Robin–Robin transmission in a 1D linear elastic wave benchmark, concluding that Robin–Robin often yields faster convergence and lower error when parameters are properly tuned (Rodriguez et al., 6 Sep 2025).

Contact mechanics provides a distinct use case. Here each body is treated as its own non-overlapping domain, non-penetration is enforced iteratively through zero normal gap rate rather than Lagrange multipliers or penalty terms, and naïve-stabilized Newmark sets interface normal acceleration to zero to suppress artificial high-frequency inertia while preserving total energy. On a one-dimensional impact benchmark, release-time error is reported as penalty/Lagrange $(I-\mathcal X^T S)\lambda=d,$4–$(I-\mathcal X^T S)\lambda=d,$5 versus NO-SAM $(I-\mathcal X^T S)\lambda=d,$6, total-energy drift as penalty/Lagrange up to $(I-\mathcal X^T S)\lambda=d,$7 versus NO-SAM $(I-\mathcal X^T S)\lambda=d,$8, and moderate chatter is removed by naïve stabilization (Mota et al., 2023).

Wave propagation is another central domain. For heterogeneous Helmholtz scattering, non-overlapping OSM and multi-trace interface formulations target large-scale indefinite systems and arbitrary partitions with cross-points (Claeys, 2019, Martin et al., 20 Jun 2025). For rectangular cavities, transmission operators that account for standing and back-propagating waves are essential; in the reported 3D helium-vessel model, a cavity-adapted Padé operator reduces the iteration count from $(I-\mathcal X^T S)\lambda=d,$9 to $S$0, a 46% reduction (Marsic et al., 2022). In HDG discretizations of Poisson problems, both Neumann–Neumann and trace–flux non-overlapping algorithms are formulated directly at the level of HDG traces and numerical fluxes (Oikawa, 2023).

Recent work also integrates NO-SAM with machine learning surrogates. PECANN-based and physics-constrained neural-network approaches solve Poisson, Helmholtz, and inverse conductivity problems on non-overlapping subdomains while learning interface parameters from augmented-Lagrangian losses (Hu et al., 2024, Basir et al., 2023). A hybrid finite element–neural operator framework replaces overlap by a non-overlapping Neumann–Dirichlet coupling, where the neural operator supplies traction and the FE solver returns displacement. This is reported to reduce inner Schwarz iterations relative to the earlier overlapping FE–NO framework, while extending the neural subdomain to irregular geometries through Point-DeepONet operating directly on unstructured FE point clouds (Wang et al., 7 Jun 2026).

6. Interpretation, common misconceptions, and current limitations

A common misconception is that NO-SAM is synonymous with a single Dirichlet–Neumann algorithm for elliptic PDEs. The current literature shows a much broader picture. NO-SAM may be formulated with alternating Dirichlet–Neumann, Robin–Robin, trace–flux, Neumann–Neumann, impedance, generalized learned interface conditions, or abstract scattering traces; the interface unknown can be a displacement, traction, impedance trace, HDG hybrid trace, neural interface residual, or a manifold point in time-domain geodesic computation (Barnett et al., 2022, Rodriguez et al., 6 Sep 2025, Oikawa, 2023, Sutti et al., 2024).

A second misconception is that non-overlap precludes difficult interface topologies. In fact, arbitrary non-overlapping partitions with cross-points are a central concern of the Multi-Trace and unified Robin–Schwarz theories, which introduce global exchange operators precisely to recover continuity and solvability in such settings (Claeys, 2019, Pechstein, 2022). Likewise, non-matching meshes and heterogeneous time integration are explicitly accommodated in the contact formulation via transfer operators, and independently meshed subdomains are part of the motivation in OpInf and learning-based couplings (Mota et al., 2023, Rodriguez et al., 6 Sep 2025).

The main limitations reported across the literature are iterative overhead and sensitivity to interface design. Robin–Robin schemes can outperform Dirichlet–Neumann, but improper $S$1 choices may induce spurious oscillations localized near the interface (Rodriguez et al., 6 Sep 2025). In contact, high-frequency chatter may persist without stabilization (Mota et al., 2023). In cavity Helmholtz, local approximations degrade as geometry departs from the ideal rectangular setting, even though they remain convergent (Marsic et al., 2022). For HDG formulations, the cited paper does not provide a full spectral convergence proof (Oikawa, 2023). In the geodesic setting, quadratic convergence is stated as an expectation once iterates are sufficiently close to the true solution, and the reported experiments are preliminary (Sutti et al., 2024). In the FE–neural-operator framework, convergence stability is empirical rather than theorem-backed (Wang et al., 7 Jun 2026).

Taken together, these results depict NO-SAM less as a single algorithm than as a general interface-centered paradigm. Its defining feature is the replacement of a monolithic global solve by repeated local solves and explicit interface reconciliation on non-overlapping subdomains. This suggests that its enduring value lies in modularity: it can couple heterogeneous discretizations, reduced models, neural surrogates, and even non-PDE subproblems while preserving a mathematically recognizable Schwarz structure (Pechstein, 2022, Barnett et al., 2022, Sutti et al., 2024).