Parallel Overlapping Domain Decomposition

• Parallel Overlapping DDM is a method that decomposes a global PDE into overlapping local subproblems, enabling simultaneous parallel solves with controlled interface communication.
• It employs varied transmission conditions, such as Robin boundary conditions and PML layers, to accurately capture wave propagation and improve preconditioning efficiency for Helmholtz equations.
• The method demonstrates strong scalability on large computational domains, achieving near-ideal performance with metrics like equivalent sweeps to manage convergence across billions of unknowns.

Searching arXiv for the focal paper and closely related overlapping DDM work.

Parallel overlapping domain decomposition methods (DDMs) are numerical solvers and preconditioners that replace a single global problem by a family of local subproblems posed on overlapping subdomains, with the overlap used to transmit information and to reassemble a global approximation. In the modern arXiv literature, the term covers both classical overlapping Schwarz formulations and more specialized wave-propagation algorithms with perfectly matched layers (PMLs), source transfer, trace transfer, and Krylov acceleration. A central example is the one-step overlapping DDM preconditioner for the high-frequency Helmholtz equation, in which the computational domain is decomposed in both coordinate directions, all local subdomain problems are solved simultaneously, and the resulting local solutions are summed, including their PML regions, to form a global preconditioned iterate (Leng et al., 2015).

1. Conceptual and algebraic formulation

In the standard overlapping formulation for a discretized PDE system

$A U = F,$

the domain is decomposed into overlapping subdomains $\{\Omega_i\}_{i=1}^N$, Boolean restriction matrices $R_i$ extract local degrees of freedom, and local operators are defined by

$A_i = R_i A R_i^T,$

possibly with modified boundary conditions such as Dirichlet for ASM/RAS or Robin for ORAS (Dolean et al., 24 May 2026). A discrete partition of unity is introduced through diagonal weights $D_i$ satisfying

$\sum_{i=1}^N R_i^T D_i R_i = I_n,$

with multiplicity scaling as a common choice (Dolean et al., 24 May 2026). This yields the familiar one-level Schwarz preconditioners

$$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$

This algebraic perspective clarifies what “parallel overlapping” means. The local solves are compute-bound and independent once the residual has been restricted, while the overlap enters through weighted prolongation, interface communication, or explicit coupling constraints. In Krylov form, right-preconditioned GMRES solves $A M^{-1} y = F$, with $U=M^{-1}y$, and one advantage of right preconditioning is that the algorithm reports the true residual $\|F-AU\|$ (Dolean et al., 24 May 2026).

For Helmholtz problems, the same pattern is adapted to indefinite operators and artificial absorbing layers. On a truncated domain

$\{\Omega_i\}_{i=1}^N$0

with a uniaxial PML, the operator is written

$\{\Omega_i\}_{i=1}^N$1

with PML-stretched coefficients $\{\Omega_i\}_{i=1}^N$2. The interior region is partitioned into non-overlapping blocks $\{\Omega_i\}_{i=1}^N$3, and each block is enlarged by PML thickness $\{\Omega_i\}_{i=1}^N$4 and optionally extra overlap $\{\Omega_i\}_{i=1}^N$5 to form an overlapping subdomain $\{\Omega_i\}_{i=1}^N$6 (Leng et al., 2015). In this setting, the overlap is not merely a geometric convenience; it is also the medium through which outgoing information is absorbed and transferred between neighboring local problems.

2. One-step parallel overlapping DDM for the Helmholtz equation

The one-step preconditioner in “An Overlapping Domain Decomposition Preconditioner for the Helmholtz equation” is formulated by restricting the source to each subdomain interior, solving local Helmholtz–PML problems independently, and summing the local solutions on the full domain (Leng et al., 2015). With $\{\Omega_i\}_{i=1}^N$7 the restriction to $\{\Omega_i\}_{i=1}^N$8 and $\{\Omega_i\}_{i=1}^N$9 the trivial zero extension, the local right-hand side is

$R_i$0

and the local problem is

$R_i$1

equivalently written in the simplified form

$R_i$2

with PML boundary conditions (Leng et al., 2015). The global preconditioned field is then assembled as

$R_i$3

A defining characteristic of this construction is that there is no sweeping along certain directions. The local solution on each subdomain is updated simultaneously in one iteration, so the method is closer in spirit to overlapping DDMs for Poisson problems than to multiplicative sweeping schemes for wave propagation (Leng et al., 2015). When embedded as a right preconditioner for GMRES, each outer iteration applies the map $R_i$4 by solving all local problems in parallel, and convergence is monitored through the normalized residual

$R_i$5

with $R_i$6–$R_i$7 reported as typical (Leng et al., 2015).

The paper reports overall cost

$R_i$8

where $R_i$9 is the total number of unknowns and $A_i = R_i A R_i^T,$0 is the number of GMRES iterations, and observes empirically that $A_i = R_i A R_i^T,$1 grows only proportionally to $A_i = R_i A R_i^T,$2, the number of subdomains along one direction (Leng et al., 2015). To normalize this behavior, the quantity

$A_i = R_i A R_i^T,$3

is introduced and called the “equivalent sweeping”; it remains $A_i = R_i A R_i^T,$4 as the mesh is refined or the domain is enlarged (Leng et al., 2015). This makes the method notable within the Helmholtz literature: it is a parallel overlapping preconditioner whose effectiveness is described in terms usually associated with sweeping solvers, yet its local updates are additive and simultaneous rather than directional.

The numerical experiments summarize this balance between parallelism and iteration growth. In constant medium with fixed $A_i = R_i A R_i^T,$5 subdomains, local subdomain grid sizes range from $A_i = R_i A R_i^T,$6 up to $A_i = R_i A R_i^T,$7 points, frequencies from $A_i = R_i A R_i^T,$8 up to $A_i = R_i A R_i^T,$9, and equivalent sweeping is approximately $D_i$0 for all tests. In constant-medium weak scaling with $D_i$1 points per subdomain, subdomains range from $D_i$2 up to $D_i$3, giving a total grid up to $D_i$4 unknowns, while equivalent sweeps remain around $D_i$5–$D_i$6. In the Marmousi model, subdomains increase from $D_i$7 up to $D_i$8, corresponding to $D_i$9 unknowns on $\sum_{i=1}^N R_i^T D_i R_i = I_n,$0 cores, with equivalent sweeps $\sum_{i=1}^N R_i^T D_i R_i = I_n,$1 and time to solution $\sum_{i=1}^N R_i^T D_i R_i = I_n,$2 at the largest scale (Leng et al., 2015).

3. Additive, restricted, optimized, and sweeping variants

Parallel overlapping DDM is not a single algorithm but a family of related constructions distinguished by their transmission conditions, assembly rules, and information-propagation mechanisms. Classical additive Schwarz and restricted additive Schwarz use overlapping local solves plus weighted prolongation; ORAS and SORAS replace Dirichlet local operators by Robin-modified operators (Dolean et al., 24 May 2026). In high-frequency wave problems, these distinctions become especially consequential because the choice of overlap and interface condition directly affects reflection, absorption, and long-range phase propagation.

The additive overlapping DDM for the Helmholtz equation developed by Leng and Ju partitions the computational domain along all spatial directions, extends each block by a PML layer, and performs fully parallel local solves at each iteration while transferring horizontal, vertical, and corner directional residuals to neighboring subdomains as sources for the next iteration (Leng et al., 2018). In $\sum_{i=1}^N R_i^T D_i R_i = I_n,$3, for constant $\sum_{i=1}^N R_i^T D_i R_i = I_n,$4, Theorem 2.2 states that the DDM solution equals the full-space PML solution exactly, whereas in a bounded truncated domain an error estimate is proved for the modified algorithm (Leng et al., 2018). The method can be used either as an iterative solver or as a preconditioner to FGMRES, with convergence tolerance $\sum_{i=1}^N R_i^T D_i R_i = I_n,$5 described as typical and preconditioner depth $\sum_{i=1}^N R_i^T D_i R_i = I_n,$6 suggested for near-optimal performance (Leng et al., 2018).

Diagonal sweeping methods reorganize this information transfer. The source-transfer diagonal sweeping DDM uses overlapping checkerboard subdomains and $\sum_{i=1}^N R_i^T D_i R_i = I_n,$7 sweeps along coordinate diagonals rather than $\sum_{i=1}^N R_i^T D_i R_i = I_n,$8 directional passes. In $\sum_{i=1}^N R_i^T D_i R_i = I_n,$9 this means four sweeps $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$0, and in $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$1 eight sweeps $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$2, with only $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$3 sequential anti-diagonal steps per sweep under quasi-uniform decomposition (Leng et al., 2020). The trace-transfer version replaces transferred sources by transferred traces on cardinal interfaces, reducing the number of transfer directions from $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$4 to $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$5, and proves exact recovery of the global PML solution in constant media after $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$6 sweeps and in two-layered media with at most one extra round (Leng et al., 2020).

These variants address a frequent misconception: overlap alone does not determine algorithmic behavior. A purely additive parallel method, a restricted Schwarz preconditioner, a source-transfer scheme, and a diagonal-sweeping trace-transfer method may all be “parallel overlapping DDMs,” but they differ fundamentally in whether information moves by residual injection, partition-of-unity averaging, source transfer, or impedance/trace transmission. The literature therefore distinguishes additive overlap from multiplicative sweeping even when both employ PML overlaps and local direct solvers (Leng et al., 2015).

4. Parallel implementation and scalability patterns

Across the cited works, the dominant implementation model is MPI- or thread-based subdomain ownership, with one or a few subdomains assigned to each rank or thread, local matrices assembled and factored once, and only nearest-neighbor overlap communication during iterative application (Leng et al., 2015). In the Helmholtz one-step preconditioner, each MPI rank owns one $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$7, including its PML layers and local right-hand side; each subdomain assembles and factors its local discrete operator once, with factorization cost $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$8 in $$(M_{AS})^{-1} r = \sum_{i=1}^N R_i^T A_i^{-1} R_i r, \qquad (M_{RAS})^{-1} r = \sum_{i=1}^N R_i^T D_i A_i^{-1} R_i r.$$9D for a direct solver and reuse cost $A M^{-1} y = F$0 per solve (Leng et al., 2015). Communication is purely nearest-neighbour in the $A M^{-1} y = F$1 directions, and there are no global directional sweeps (Leng et al., 2015).

This communication pattern recurs in other settings. For general overlapping Schwarz methods, one preconditioner application consists of residual computation, restriction, local solve, weighted prolongation by $A M^{-1} y = F$2, and possibly a coarse solve $A M^{-1} y = F$3 assembled either redundantly or on a coarse communicator (Dolean et al., 24 May 2026). In the Closest Point Method on surfaces, each processor owns one or a few overlapping node sets $A M^{-1} y = F$4, exchanges overlap/interface values on $A M^{-1} y = F$5, and solves the local correction system with UMFPACK; the communication cost is $A M^{-1} y = F$6 (May et al., 2019). In the ptychographic OD2P algorithm, only the overlapping-layer arrays and multipliers on $A M^{-1} y = F$7 are exchanged, while all other ADMM updates are local and diagonally solvable (Chang et al., 2020).

Performance results show that overlapping DDM can be either algorithmically scalable, parallel scalable, or both, depending on the problem class and the solver architecture. In the Helmholtz one-step preconditioner, weak scaling extends to over $A M^{-1} y = F$8 subdomains, and nearly a billion unknowns are solved on several thousand cores (Leng et al., 2015). In the additive overlapping Helmholtz DDM on Tianhe-2, one MPI rank per subdomain with local MUMPS solves yields nearly ideal parallel efficiency up to $A M^{-1} y = F$9 cores, and communication occurs only over $U=M^{-1}y$0-layer overlaps (Leng et al., 2018). In the XFEM crack simulation preconditioner, the tailored overlapping additive Schwarz strategy scales up to $U=M^{-1}y$1 processor cores with over $U=M^{-1}y$2 parallel efficiency for problems with more than $U=M^{-1}y$3 degrees of freedom (Wei et al., 2021). A plausible implication is that overlap is most effective in practice when local factorizations dominate runtime and the inter-subdomain exchange remains thin, structured, and nearest-neighbor.

The modern survey literature also emphasizes the limits of one-level overlap. One-level Schwarz is not weakly scalable in general because the condition number of $U=M^{-1}y$4 grows with the number of subdomains, whereas two-level methods restore scalability by augmenting the local solves with a coarse correction, using spaces such as Nicolaides, GenEO, or DtN-based coarse spaces (Dolean et al., 24 May 2026). For diffusion problems, GenEO yields contrast-robust coarse spaces of moderate dimension; for Helmholtz or Maxwell, DtN-based coarse spaces are used to address wave-specific global modes (Dolean et al., 24 May 2026).

5. Convergence theory and interface operators

The theoretical analysis of parallel overlapping DDM often proceeds by formulating the iteration as a fixed-point map on a product of local solution spaces. For the Helmholtz equation with first-order absorbing transmission conditions, the error satisfies

$U=M^{-1}y$5

where the block operator $U=M^{-1}y$6 is defined by solving a homogeneous Helmholtz problem on $U=M^{-1}y$7 with inhomogeneous Robin data induced by $U=M^{-1}y$8 on interior interfaces (Beuchler et al., 2021). The method is shown to be well defined in dimensions $U=M^{-1}y$9 in a tensor product of appropriate local function spaces, and norm bounds are expressed through impedance-to-impedance maps arising from local interactions between subdomains (Beuchler et al., 2021).

For strip decompositions, the analysis introduces quantities $\|F-AU\|$0 and $\|F-AU\|$1 derived from norms of these interface maps and proves estimates such as

$\|F-AU\|$2

which imply power contractivity under appropriate assumptions (Beuchler et al., 2021). The same operator-theoretic structure extends to heterogeneous time-harmonic Maxwell equations with impedance boundary conditions. There, the local error-transfer operators are characterized by tangential impedance-to-impedance maps, and the discrete Nédélec-edge formulation is shown to inherit the convergence behavior of the continuous method under an approximation assumption for the discrete interface maps (Cen et al., 3 Jun 2026).

A different analytical route appears in the overlapping Schwarz-type method for the Monge–Ampère equation. Starting from a monotone, proper, continuous finite-difference scheme

$\|F-AU\|$3

one defines subdomain solution operators $\|F-AU\|$4, chooses weights $\|F-AU\|$5 with $\|F-AU\|$6, and iterates

$\|F-AU\|$7

Using the discrete comparison principle, the global map $\|F-AU\|$8 is monotone, and Theorem 3.1 states that for any initial guess $\|F-AU\|$9, the iterates converge uniformly to the unique discrete solution (Boubendir et al., 2023). This is notable because it shows that parallel overlapping DDM is not restricted to linear or weakly nonlinear settings.

These analyses also temper an overly broad claim sometimes associated with overlapping methods, namely that overlap automatically guarantees convergence. The Maxwell and Helmholtz theories instead express convergence through the norms of local interface operators, the amount of overlap, and the form of the transmission condition (Beuchler et al., 2021). The Monge–Ampère result likewise depends on monotonicity, properness, and a comparison principle rather than on overlap alone (Boubendir et al., 2023).

6. Application domains beyond classical Helmholtz solvers

Although the terminology originates in PDE linear algebra, parallel overlapping DDM has been adapted to a wide range of discretizations, operators, and optimization problems. In surface PDEs discretized by the Closest Point Method, overlapping RAS and ORAS are built on graph partitions of active grid nodes, with Robin transmission conditions defined through a discrete conormal direction on the subdomain boundary. On the Stanford Bunny with $\{\Omega_i\}_{i=1}^N$00 and $\{\Omega_i\}_{i=1}^N$01, ORAS yields roughly a two-fold reduction in iterations compared with RAS when $\{\Omega_i\}_{i=1}^N$02, while GMRES preconditioned with block-Jacobi requires $\{\Omega_i\}_{i=1}^N$03 iterations (May et al., 2019).

In variational imaging, overlap is defined through “essential domains,” namely the minimal enlarged sets on which a local integrand depends. The original minimization

$\{\Omega_i\}_{i=1}^N$04

is reformulated as

$\{\Omega_i\}_{i=1}^N$05

and a decoupled augmented-Lagrangian iteration replaces the coupling penalty by $\{\Omega_i\}_{i=1}^N$06, so that all local minimizations decouple and can be solved in parallel (Park, 2020). The framework is applied to total variation minimization and higher-order models, with near-ideal speedup up to $\{\Omega_i\}_{i=1}^N$07 subdomains on a $\{\Omega_i\}_{i=1}^N$08 test image (Park, 2020).

In ptychographic imaging, the OD2P method decomposes the image into overlapping subdomains, enforces continuity across overlaps, and uses an ADMM formulation with a smooth-truncated amplitude-Gaussian metric whose proximal operator has a closed-form coordinate-wise solution (Chang et al., 2020). For a $\{\Omega_i\}_{i=1}^N$09 object, the algorithm requires $\{\Omega_i\}_{i=1}^N$10 iterations for $\{\Omega_i\}_{i=1}^N$11 subdomains, respectively, while virtual runtimes shrink nearly inversely in $\{\Omega_i\}_{i=1}^N$12 (Chang et al., 2020). This suggests that overlap-based decomposition can remain effective even in large-scale nonconvex inverse problems, provided the local subproblems preserve closed-form or diagonal structure.

Additional examples reinforce the breadth of the paradigm. In XFEM crack simulation, the overlap is tailored to crack-tip enriched regions and regular regions, and the preconditioner is applied in two stages, crack-tip first and regular subdomains second, to match the processor layout (Wei et al., 2021). In adaptive moving mesh generation, the DDPMA method solves local parabolic Monge–Ampère equations on overlapping slabs or blocks with one halo exchange per pseudo-time step and reports second-order spatial accuracy and first-order pseudo-time convergence (Sulman et al., 2020). In large-scale photonic and electromagnetic modeling, a finite-difference frequency-domain overlapping DDM uses PML overlaps and equivalent sources

$\{\Omega_i\}_{i=1}^N$13

to transfer fields between neighboring subdomains, with a multi-frontal preconditioner and OpenMP parallelization (Wang et al., 25 Sep 2025).

Taken together, these applications show that “parallel overlapping DDM” refers less to a single transmission formula than to a recurrent architecture: decompose, enlarge by overlap, solve locally, exchange or average information on the overlap, and accelerate globally by fixed-point iteration, GMRES, ADMM, or augmented-Lagrangian updates. The common technical challenge is always the same: the overlap must be large enough and the interface model sufficiently accurate that local parallelism does not destroy global propagation or global consistency.