Modified Overlapping Schwarz Decomposition

• Modified Overlapping Schwarz decomposition is a modular domain decomposition technique that improves convergence by tuning local subproblems, interface exchanges, and coarse corrections.
• It employs innovations such as energy-minimizing coarse spaces and optimized transmission operators to address challenges in sparse linear solvers, high-frequency problems, and complex system discretizations.
• The approach is scalable and versatile, finding applications in PDEs, optimal control, meshless methods, and neural-network systems while leveraging modern hardware for improved performance.

Searching arXiv for recent and relevant papers on modified overlapping Schwarz decomposition. Modified overlapping Schwarz decomposition denotes a family of domain decomposition methods built on overlapping subdomains but altered relative to the classical Schwarz iteration in at least one essential component: the local subproblem, the transmission condition, the interface variables, the coarse correction, or the algebraic assembly. In the cited literature, such modifications appear in sparse linear solvers, high-frequency Helmholtz problems, vector-field discretizations, constrained quadratic programs, optimal control, space–time refinement for the material point method, surrogate modeling via proper generalized decomposition, meshless finite differences, independent-grid spectral collocation, and randomized neural-network least-squares systems (Tran, 2011, Yamazaki et al., 2023, Shin et al., 2020, Luo et al., 9 May 2026).

1. Classical Schwarz baseline and the rationale for modification

The classical overlapping Schwarz method starts from a cover

$\Omega=\bigcup_{l=1}^I \Omega_l$

by overlapping subdomains and solves local problems while exchanging interface data between neighboring subdomains. In the Dirichlet version for semilinear parabolic or elliptic equations, the iterate on each subdomain uses the previous iterate as boundary data on the overlap interfaces. In the notation of the convergence study for semilinear equations, this is the transmission condition

$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$

A weighted maximum-principle argument yields geometric convergence for the Dirichlet method in both the semilinear parabolic and semilinear elliptic settings (Tran, 2011).

The need for modification appears when the classical transmission is either too reflective, too weakly scalable, or too expensive. The semilinear analysis shows that Robin transmission conditions converge for semilinear parabolic equations, but a one-dimensional counterexample demonstrates that fixed Robin parameters may lead to divergence for semilinear elliptic equations. Convergence can be restored by scaling the Robin parameter,

$$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$

with $\rho$ sufficiently large (Tran, 2011).

This baseline already identifies the central theme of modified overlapping Schwarz decomposition: the overlap alone is not the defining design variable. Robustness and efficiency depend on how overlap is combined with boundary exchange, local solvers, coarse representations, and algebraic restrictions. The later literature preserves the parallel subdomain structure of Schwarz while changing precisely these ingredients.

2. Coarse spaces, two-level structure, and multilevel extensions

A major line of modification replaces the one-level method by a two-level or multilevel preconditioner. In the generalized Dryja–Smith–Widlund formulation, the additive two-level preconditioner is

$$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$

where the coarse basis is not piecewise constant but energy-minimizing and nullspace-enriched. After the interface–interior splitting

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

the coarse basis has the form

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

so that the coarse matrix is $A_0=\Phi^\top A\Phi$. This modification is designed to provide robustness and scalability for sparse linear systems arising from the discretization of a wide range of partial different equations (Yamazaki et al., 2023).

For elliptic problems, the resulting condition-number estimate is

$$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$

independent of the number of subdomains. The same study reports that iteration counts remain essentially constant as $N\to$ large in weak scaling, and that robustness with respect to coefficient jumps can be improved further by enrichment of the coarse space (Yamazaki et al., 2023).

The GPU implementation in FROSch uses Trilinos/Tpetra for distributed sparse matrices and Kokkos for performance-portable kernels. A novel decomposition runs multiple MPI processes on each GPU through NVIDIA Multi-Process Service on Summit with NVIDIA V100 GPUs. In the reported experiments, solve time was reduced by factors of about $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$0 using GPUs, while inexact local solvers based on FastILU and FastSpTRSV yielded up to $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$1–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$2 solve-time speed-ups on GPUs, with iteration count $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$3–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$4 larger; building $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$5, $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$6, and local factors in single precision while running GMRES in double precision gave a $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$7–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$8 setup speed-up with identical iteration counts (Yamazaki et al., 2023).

For monolithic incompressible fluid problems, the same structural idea is extended recursively. The two-level monolithic Schwarz preconditioner for the Stokes saddle-point system couples velocity and pressure at every level, and the GDSW coarse basis is constructed from interface entities together with the local nullspace given by rigid-body velocity modes plus the constant pressure mode. The three-level extension applies the same procedure recursively to the coarse operator and thereby avoids a single massive coarse-level factorization. Reported parallel results reach $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$9 MPI ranks; for Poiseuille flow on the unit cube, the three-level method handled $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$0 million unknowns with coarse sizes $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$1 and $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$2 FMRES iterations, while the two-level method failed when the coarse solve became too large (Köhler et al., 6 Aug 2025).

3. Transmission operators, impedance conditions, PML, and interface modifications

Another major class of modifications changes the transmission operator itself. For Helmholtz problems, Dirichlet exchange is too reflective, and the restricted additive Schwarz method is modified by first-order absorbing or impedance conditions. In the variational formulation of ORAS, the local sesquilinear form on $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$3 is

$$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$4

and the finite-element iteration is exactly the preconditioned Richardson method with preconditioner

$$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$5

The associated PDE-level and discrete error-propagation operators exhibit power-contraction properties under geometric assumptions on aspect ratio, overlap, and neighbor count (Gong et al., 2021).

For high-frequency Helmholtz, the modification goes further by combining overlap with local perfectly matched layers. In the RAS-PML method, both the overlap width and the local PML thickness are chosen as

$$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$6

while the domain is decomposed into $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$7 overlapping Cartesian subdomains. The reported 2D constant-speed experiments show roughly $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$8 Schwarz iterations and $$\mathfrak B_{l,l'}^\rho v=\partial_{n_{l,l'}}v+\rho\,p_{l,l'}\,v,$$9 wall time, with one processor per subdomain and total parallel scalability $\rho$0. The hybrid PML+impedance boundary condition is reported as more robust than a pure PML+Dirichlet choice when $\rho$1 is small (Xie et al., 31 Jan 2026).

In surface PDEs discretized by the closest point method, ORAS is modified again near cross-points, where three or more subdomains meet. Standard ORAS may stagnate unless the Robin weight is increased locally from $\rho$2 to $\rho$3 in a neighborhood of those nodes. Numerical experiments on the sphere, torus, and Stanford Bunny show that ORAS preconditioners outperform RAS, and that the modified cross-point treatment further reduces Krylov iterations; on the sphere with $\rho$4, $\rho$5, and $\rho$6, the preconditioned iteration counts reported are $\rho$7 for RAS, $\rho$8 for ORAS without cross-point modification, and $\rho$9 for ORAS with the modified treatment (May et al., 2019).

A different interface modification appears in the Schwarz–Fourier method on two overlapping discs. There, each local Laplace solve is replaced by projection of the trace onto the first $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$0 Fourier modes, followed by harmonic extension. The projected Poisson kernel is nonnegative only on a safe subdisc, which leads to a maximum-principle argument on that subregion and to the contraction estimate

$$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$1

with $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$2 and $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$3 as $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$4. The resulting inexact Schwarz method therefore retains geometric convergence in the max-norm and, for moderate $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$5, converges nearly as fast as the exact Schwarz iteration (Reusken, 3 Nov 2025).

4. Primal–dual and time-domain formulations in optimization and control

Modified overlapping Schwarz decomposition has also been transferred from PDE linear systems to structured optimization. For constrained quadratic programs on graphs, the global variable is a primal–dual block $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$6, the graph is partitioned into disjoint sets $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$7, and each is expanded by an overlap radius $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$8 to form $$M_{\mathrm{GDSW}}^{-1} = R_0^\top A_0^{-1}R_0 + \sum_{i=1}^N R_i^\top A_i^{-1}R_i,$$9. The local subproblem is a parametric QP whose right-hand side is shifted by the coupling term $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$0, and the local solution is then trimmed back to the original non-overlapping block $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$1. Relative to classical Schwarz for PDE solvers, the modifications are explicit: the method works on the KKT system of a constrained QP, passes full primal–dual blocks, and uses linear forcing terms rather than Dirichlet boundary values alone (Shin et al., 2020).

Its convergence theory is based on exponential decay of sensitivity. The solution map on a subgraph has a block-Lipschitz constant decaying exponentially with graph distance, which yields a contraction factor

$A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$2

for sufficiently large overlap. The numerical DC optimal power flow test on a $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$3-node network, partitioned into $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$4 parts by METIS and solved with IPOPT on $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$5 cores, reports $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$6 iterations and $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$7 s for $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$8, $A=\begin{bmatrix}A_{II}&A_{I\Gamma}\A_{\Gamma I}&A_{\Gamma\Gamma}\end{bmatrix},$9 iterations and $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$0 s for $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$1, and $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$2 iterations and $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$3 s for $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$4 (Shin et al., 2020).

The same exponential-decay mechanism underlies overlapping Schwarz methods for nonlinear optimal control. In the discrete-time OCP setting, the time horizon is partitioned into overlapping intervals, all local subproblems are solved in parallel, and convergence is enforced by updating primal–dual information at the boundaries of overlapping subdomains. Under uniform SOSC, controllability, and boundedness assumptions, both primal and dual solutions satisfy exponential decay of sensitivity, and the Schwarz iteration converges locally linearly with factor

$$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$5

where $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$6 is the overlap size. On the quadrotor motion-planning example with horizon $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$7 and $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$8 subdomains, the reported OSD performance is $$\Phi= \begin{bmatrix} -\,A_{II}^{-1}A_{I\Gamma}\ I_\Gamma \end{bmatrix}\Phi_\Gamma,$$9 iterations and $A_0=\Phi^\top A\Phi$0 s for $A_0=\Phi^\top A\Phi$1 overlap, versus more than $A_0=\Phi^\top A\Phi$2 ADMM iterations and about $A_0=\Phi^\top A\Phi$3 s; for the thin-plate PDE control problem, OSD with full relative overlap requires about $A_0=\Phi^\top A\Phi$4 iterations and $A_0=\Phi^\top A\Phi$5 s, comparable to Ipopt and much faster than ADMM (Na et al., 2020).

In the continuous-time linear–quadratic case, the same idea is formulated directly from the Pontryagin minimum principle. The time interval is partitioned into overlapping subintervals, and each local problem is a Hamiltonian two-point boundary-value problem on an extended interval with boundary data

$A_0=\Phi^\top A\Phi$6

The global iterate is rebuilt by discarding the overlap padding and concatenating the interior pieces. The resulting Schwarz map is a contraction provided the overlap $A_0=\Phi^\top A\Phi$7 is large enough, with geometric estimate

$A_0=\Phi^\top A\Phi$8

The numerical example on a linearized biochemical reaction network shows strictly linear convergence in iteration count, exponential acceleration as $A_0=\Phi^\top A\Phi$9 grows, and no change in Schwarz-iteration rate when explicit Euler, implicit Euler, RK45, or MATLAB ode23 are used; the higher-order or adaptive solvers only reduce discretization error (Zhao et al., 6 Oct 2025).

5. Independent grids, surrogate models, meshless discretizations, and neural-network bases

Several modifications keep the Schwarz logic but change the discretization model itself. In the independent-grid Chebyshev formulation, the original global PDE is replaced by an equivalent multidomain system in which each subdomain carries its own copy of the solution, and matching is imposed only on interfaces. There is therefore no global shared discretization and no need for restrictive-type overlap updates. Communication is reduced to interface interpolation operators $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$0, and the nonlinear Schwarz preconditioner is built from

$$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$1

On driven-cavity flow at $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$2, Newton–Krylov–Schwarz fails to converge from the smooth initial guess, whereas Schwarz–Newton–Krylov still converges robustly; on the Burgers test, the nonlinear residual evaluations parallelize nearly ideally up to $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$3 cores (Aiton et al., 2019).

For parametric elliptic problems, the DD–PGD formulation treats the unknown interface trace $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$4 as an additional parametric field. Local solutions are decomposed into a source-only part and boundary-driven parts, each represented by proper generalized decomposition modes. Elimination of interior degrees of freedom produces a purely algebraic interface system

$$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$5

and continuity across overlaps is enforced without Lagrange multipliers. In the online phase, no further high-fidelity PDE solves are performed; the computation reduces to interpolation of stored PGD modes and matrix-free Krylov iterations on the interface system. The reported speed-ups are about $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$6 for the synthetic diffusion problem, about $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$7 for the Poiseuille–Graetz convection–diffusion example, and $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$8 to $$\kappa(M_{\mathrm{GDSW}}^{-1}A)\le C\,(1+H/\delta)\,[1+\log(H/h)]^2,$$9 in the $N\to$0-block thermal problem, with global errors of order $N\to$1 to $N\to$2 and interface mismatch around $N\to$3 (Discacciati et al., 2023).

In the meshless RBF–FD setting, Schwarz is reformulated entirely algebraically on unstructured node sets. Subdomains are built by k-means clustering and growth by nearest neighbors; two partitions of unity are tested, a discontinuous dPoU and a continuous bump-based cPoU. Two preconditioners are compared: RAS, which applies the weight after the local solve, and ASH, which applies it to the right-hand side before the solve. The experiments show that a small but nonzero overlap, typically near $N\to$4 or $N\to$5, gives the best trade-off between iteration count and local solve cost in 2D Poisson and Stokes problems, while too little overlap yields stagnation and too much increases LU fill-in. ASH outperforms RAS in almost all configurations, and no disjoint partitioning technique is required (Westermann et al., 1 Jul 2026).

A related algebraic reformulation appears for randomized neural networks with domain decomposition. Local basis functions are initialized randomly on each subdomain, multiplied by a partition of unity, and combined into a global approximation

$N\to$6

Boundary conditions are handled exactly through a constraining operator, local feature matrices are compressed by PCA, and the global normal equations are solved with additive Schwarz or restricted additive Schwarz preconditioners. The construction is reported to reduce the condition number from values as large as $N\to$7 to $N\to$8–$N\to$9, with empirical condition-number reductions of $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$00–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$01 and wall-clock speed-ups of $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$02–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$03 (Shang et al., 2024).

6. Specialized PDE systems, space–time refinement, and recurring convergence patterns

For vector-field problems in three dimensions, the modification is primarily analytical rather than algorithmic. In the Nédélec and Raviart–Thomas settings, the two-level Schwarz preconditioner has the standard additive form

$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$04

but the new analysis replaces classical orthogonal Helmholtz decompositions by Hiptmair–Pechstein regular stable decompositions. This yields condition-number bounds

$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$05

without convexity or Betti-number assumptions. The numerical results on cube and cavity-cube geometries show computed condition numbers near $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$06–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$07 and CG iteration counts near $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$08–$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$09, with no degradation due to the cavity (Oh et al., 2024).

A distinct space–time modification appears in OS-MPM for the material point method. The domain is decomposed into overlapping coarse and fine subdomains,

$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$10

with heterogeneous spatial resolutions $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$11 and temporal subcycling ratio $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$12. Coarse-to-fine coupling uses interpolation

$$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$13

while fine-to-coarse exchange uses the mass-weighted projection $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$14. The multiplicative Schwarz loop alternates one Backward-Euler MPM step on the coarse domain with $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$15 implicit fine-domain substeps and checks convergence by normalized interface residuals. In the elastic inclusion benchmark, both single-domain and OS-MPM exhibit roughly first-order convergence in stress, the finest OS-MPM errors are equal to or slightly lower than the single-domain errors, and the speed-up reaches up to $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$16 for stiffness ratio $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$17; the abstract reports computational-cost reduction by up to $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$18 times (Luo et al., 9 May 2026).

Across these diverse formulations, one recurring conclusion is that “modification” does not identify a single algorithmic recipe. In the cited work it may mean an energy-minimizing coarse space, a damped Robin operator, a local PML, a cross-point-dependent impedance weight, a Fourier projection, a primal–dual exchange rule, an independent-grid reformulation, a partition-of-unity interface system, or a space–time transfer operator. A second recurring conclusion is that overlap is important but not monotone in its effect. Larger overlap accelerates convergence exponentially in graph-structured QPs and optimal control, yet intermediate overlap can minimize total runtime because local subproblems become more expensive; in meshless RBF–FD a small but non-disjoint overlap is optimal, and in high-frequency Helmholtz the overlap and PML layers can decrease like $$u_l^k|_{\Gamma_{l,l'}} = u_{l'}^{\,k-1}|_{\Gamma_{l,l'}}.$$19 while maintaining good observed convergence (Shin et al., 2020, Westermann et al., 1 Jul 2026, Xie et al., 31 Jan 2026).

Taken together, the modern literature presents modified overlapping Schwarz decomposition as a modular framework rather than a single method. The invariant core is the decomposition into overlapping local problems; the decisive advances come from how one represents coarse information, how one transmits interface data, and how one matches the Schwarz mechanism to the algebraic, variational, temporal, and hardware structure of the target problem.