{"id":"893a78db-1d83-4cb1-8c52-79456f18147b","arxiv_id":"2506.16875","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"On a realistic 3D geophysical benchmark with multiple sources, the non-overlapping optimized Schwarz method (OSM) outperforms the overlapping ORAS preconditioner by about a factor of two in time and memory when both are tuned.","lead":"This paper compares two domain decomposition methods for solving very large 3D wave (Helmholtz) problems with many sources, as used in seismic imaging. It finds that the non-overlapping method, once its transmission parameters are tuned, solves a 630-million-unknown test case about twice as fast as the overlapping method while using less memory.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transmission coefficients (α, β) and the block-triangular solve variant are never reported; the headline factor-of-two comparison may partly rest on an untracked implementation or tuning choice, so the OSM advantage should be re-confirmed with the published parameters before acceptance.","rationale":"The reader's weakest assumption concerns the transfer of tolerance calibration from 2 Hz/256-subdomain runs to all larger cases; that is a genuine and important concern, and I agree it warrants conditionality. My stress-test pass pushes one level deeper: the actual optimized transmission coefficients are never reported, so the central comparison cannot be reproduced, and the manuscript's own Section 3.4 describes an OSM-only block-triangular local-solve optimization that is not mirrored for ORAS. These are not accusations of unfair comparison—the authors are comparing their implementations, and the conclusion is stated for their tuned implementations—but they are load-bearing for the general claim that 'when appropriately tuned' OSM outperforms ORAS. The memory argument is stronger and less dependent on tuning: Tables 1–3 show smaller factors and smaller GMRES subspaces, which are structural. The timing factor of two is the part most exposed to the missing parameter disclosure. The proposed test—disclosing α, β and rerunning the largest case with a symmetric implementation choice—directly settles whether the headline timing advantage is robust. In my read, the paper's evidence is otherwise credible: the weak-scaling tables are internally consistent, the residual-versus-L2-error calibration is presented (Figures 4 and 6), the reproducibility repository is promised, and the scaling trends are physically sensible (iteration counts grow with the number of subdomains). The concern is not that the comparison is fraudulent; it is that the manuscript does not yet provide enough information to distinguish a methodological advantage from an implementation or tuning advantage. Hence CONDITIONAL, not REJECT.","tokens_in":16481,"tokens_out":1767,"duration_ms":17873,"concrete_test":"Re-run the largest configuration (6.25 Hz, 630.1 MDOFs, 16,384 subdomains, 64 sources) with the exact α and β values used in the paper, disclosed in a supplement, and with OSM's block-triangular local solve replaced by a direct solve of the full 2x2 local system (or with an equivalent optimized solve applied to ORAS). If OSM's total time remains below ORAS's within, say, 20%, the central comparison survives; if the margin collapses or reverses, the reported factor of two is partly an implementation artifact.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that, when appropriately tuned, non-overlapping OSM outperforms ORAS by a factor of about two in time and with lower memory. The paper reports the numerical optimization of α and β (Section 4) and the use of second-order transmission conditions S = α + βΔΣ (Eq. 3.1), but it never reports the optimized values for α, β at any frequency, nor the formula or procedure used to obtain them. Since the entire comparison rests on closing the convergence gap via optimized transmission conditions, the headline result is not reproducible from the manuscript alone: a referee or practitioner cannot determine whether the measured gap is a property of the methods or a property of the particular tuning in the provided implementation. This is the same load-bearing point as the reader's weakest assumption—the calibration is assumed to transfer across all sizes—but narrowed to the even more basic issue that the calibrated quantities themselves are absent. A second, closely related omission is that Section 3.4 states the OSM local solve exploits the block-triangular 2x2 structure of the local problem, replacing a symmetric solve by two symmetric solves with a fieldsplit preconditioner; the ORAS implementation is not reported to use an equivalent optimization. If this block-triangular trick is implemented efficiently only for OSM, part of the measured speedup is implementation-specific rather than intrinsic to the non-overlapping substructured formulation. This does not invalidate the engineering comparison, but it weakens the general claim that OSM inherently outperforms ORAS by the reported margin.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares two domain decomposition strategies for large-scale 3D Helmholtz problems with multiple right-hand sides: the overlapping ORAS preconditioner and the non-overlapping substructured Optimized Schwarz Method (OSM) with second-order transmission conditions S = α + βΔΣ. Using the GO 3D OBS geophysical benchmark and meshes discretized with third-order finite elements, the authors calibrate transmission parameters and a common stopping criterion in Section 4, then report memory and timing experiments up to 630.1 million unknowns and 64 sources in Section 5. The central claim is that, once appropriately tuned, OSM reduces the convergence gap relative to ORAS and consistently outperforms ORAS by roughly a factor of two in wall-clock time while using substantially less memory. The paper also discusses weak scaling, batch-size effects, and a comparison with a direct solver.","tokens_in":16823,"tokens_out":4289,"duration_ms":44750,"significance":"If the reported comparison is robust, the paper provides practically important evidence for choosing non-overlapping substructured methods over overlapping preconditioners for frequency-domain full-waveform inversion workloads, where many right-hand sides must be solved on very large 3D meshes. The strengths of the work are its realistic test case, the systematic weak-scaling study, the explicit memory and time breakdowns, the pseudo-block GMRES treatment of multiple sources, and the availability of source code via the reproducibility section. The main qualifications are that the numerically optimized transmission coefficients are never reported, the two methods are compared with different batch sizes in the headline timing tables, and the accuracy equivalence between the two stopping criteria is only calibrated at 256 subdomains and low frequency. These issues do not undermine the plausibility of the qualitative conclusion, but they currently prevent the quantitative factor-of-two claim from being fully reproducible and unambiguous.","major_comments":[{"comment":"The optimized complex coefficients α and β are never reported, and the numerical optimization procedure is not described. Since the entire conclusion that OSM can close the convergence gap relies on these choices, the paper is not self-contained: a reader cannot tell whether the reported speedup is a property of the methods or of the particular tuning in the provided implementation. Please report the optimized values of α and β for the frequencies and partitionings used, the search space and objective (e.g., GMRES iteration count versus transmission parameter), and whether the same values are used for OSM and ORAS.","section":"Section 4, Eq. (3.1)"},{"comment":"The headline timing comparison uses an asymmetric batch-size setting: OSM is run with a full batch of 64 sources while ORAS is run with a batch size of 16, and Figure 8 shows that batch size affects the per-iteration time. The factor-of-two speedup is therefore partly attributable to a more favorable batching choice for OSM rather than to the method itself. Please provide a same-batch-size comparison for both methods, or otherwise quantify and remove the batch-size contribution, before claiming that OSM is intrinsically faster by about a factor of two.","section":"Section 5.2, Tables 4 and 5"},{"comment":"The common stopping criterion is calibrated from the residual-versus-L2-error curves of Figures 4 and 6 only at 256 subdomains, at 1 Hz for the homogeneous case and 2 Hz for the heterogeneous case, in double precision. The large-scale runs in Section 5 use a relative residual tolerance of 1e-4 in single precision over a much wider range of frequencies, partition counts, and subdomain sizes. The manuscript does not report the actual L2 errors achieved in the large runs, so it is not verified that OSM and ORAS stop at equal accuracy in those runs. Please report L2 errors, or at least a representative subset, for the large-scale cases, or verify with a stricter tolerance that the relative timings are unchanged.","section":"Section 4.2 and Section 5"},{"comment":"The OSM implementation exploits a block-triangular 2x2 structure of the local problem using a fieldsplit preconditioner, while the ORAS implementation is not described as using an analogous optimization. If ORAS solves the full local system without this optimization, part of the measured OSM advantage could be implementation-specific rather than intrinsic to the non-overlapping substructured formulation. Please clarify whether ORAS local solves use an equivalent optimized treatment, and if not, estimate the contribution of this implementation choice to the measured speedup.","section":"Section 3.4"}],"minor_comments":[{"comment":"The caption of Figure 5 says \"homogeneous medium,\" but Section 4.2 describes a heterogeneous medium; the caption should be corrected.","section":"Figure 5"},{"comment":"There is a typo: \"One the one hand\" should read \"On the one hand.\"","section":"Section 1"},{"comment":"The text says \"More details are available in section 7,\" but Section 7 is a short reproducibility note and contains no implementation details; either add the details or rephrase the reference.","section":"Section 3.4"},{"comment":"The header \"MDOFs indiciates\" contains a typo; it should read \"MDOFs indicates.\"","section":"Tables 4 and 5"},{"comment":"The statement that OSM and MUMPS solve \"with the same resources\" would benefit from explicit node/process/thread counts for both runs; as written, the comparison is difficult to verify.","section":"Section 5.2, direct solver comparison"},{"comment":"The reproducibility link would be more durable if it included a version number, commit hash, or DOI.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and addresses a practically important question. The reader's weakest-assumption concern about tolerance calibration is real and is reflected in my major comments, as is the additional batch-size asymmetry that I did not see emphasized in the reader's report. The missing α and β values are a straightforward but necessary fix. I do not see evidence of circular reasoning; the issue is incomplete reporting and comparison fairness, not a derived-quantity tautology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a real head-to-head, not a toy: OSM (FETI-2LM substructured) versus ORAS on the GO 3D OBS target, up to 630M unknowns and 64 sources, with a batched multi-RHS implementation. For a frequency-domain FWI practitioner, this is useful. The central finding—when optimized transmission conditions close the convergence gap, OSM is about twice as fast and uses substantially less memory—is credible from the data. The memory tables are particularly convincing: substructuring reduces both factorization and Krylov storage, and the RSS measurements show OSM fitting into memory where ORAS cannot.\n\nWhat's actually new is the scale and the controlled comparison, not the methods. The convergence calibration in Section 4 is a nice touch, and the paper honestly notes that ORAS converges faster per iteration; the advantage comes from cheaper iterations and fewer unknowns, not from magic. The residual/error relationship is shown for homogeneous and heterogeneous cases, which is the right way to define a common stopping criterion.\n\nThe soft spots are real but manageable. The optimized α and β are never reported. Saying 'numerically optimized' without giving the values (or at least the procedure) makes the headline comparison hard to reproduce from the paper alone. The source code is available, so a determined reader can recover them, but the manuscript itself is incomplete on that point. Second, timings are single runs, no variance or repeated measurements; at this scale that's common, but it means the factor of two has unknown noise. Third, the tolerance calibration is done at 2 Hz with 256 subdomains and then transferred to all sizes and frequencies. The paper shows the residual/error curves are similar in the heterogeneous case, so this is reasonable, but it's an assumption worth noting. The block-triangular solve in OSM is an implementation optimization the ORAS path doesn't get; that could inflate the OSM advantage beyond the intrinsic method comparison. None of these flaws sinks the conclusion, but they should be disclosed or addressed in revision.\n\nThe citation pattern is fine, and the experimental discipline is above average for this literature. This paper deserves a serious referee and likely acceptance after a moderate revision that reports the tuning values, adds some repeatability notes, and acknowledges the implementation asymmetries.\n\nFor your reading group: yes, worth a look, especially for the memory analysis and the batched RHS discussion.","headline":"A credible and useful engineering comparison that gives OSM a real factor-of-two edge on large multi-source Helmholtz problems, but the unreported tuning parameters and single-run timings need attention before the numbers become doctrine.","tokens_in":17300,"tokens_out":1760,"would_cite":true,"duration_ms":19161,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35J05","65N55","68W10","35-04","86-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"On a 630-million-unknown seismic problem, a non-overlapping substructured solver beats the overlapping preconditioner by about 2x in time.","keywords":["domain decomposition","Helmholtz equation","optimized Schwarz method","additive Schwarz","full waveform inversion","high-order finite elements","multiple right-hand sides","GMRES"],"falsifier":"Run the 6.25 Hz, 16,384-subdomain case with both methods and compare the true L2 error of the final solution against a reference obtained on a coarser grid or with a different solver. If OSM's solution error is more than a small factor worse than ORAS's at the same 1e-4 residual tolerance, the reported time comparison is biased; alternatively, if ORAS with matched true accuracy and a large enough batch size finishes no slower than OSM, the central claim fails.","tokens_in":16295,"feed_emoji":"🌊","tokens_out":4472,"duration_ms":43152,"temperature":0.7,"pith_summary":"This paper aims to show that for large-scale, multiple-source Helmholtz problems, the non-overlapping substructured Optimized Schwarz Method (OSM) can be made competitive enough—through numerically tuned second-order transmission conditions—to consistently beat the overlapping ORAS preconditioner by roughly a factor of two in time and with much smaller memory. The authors test this on a realistic 3D geophysical model (GO 3D OBS) with up to 630.1 million unknowns and 64 sources, using third-order finite elements on adapted tetrahedral meshes. The practical stakes are frequency-domain full waveform inversion, where many right-hand sides must be solved and memory per process is a hard limit. If the comparison holds, OSM is the better one-level method for these problems, and the case for adding a coarse level is strengthened.","feed_headline":"A non-overlapping DDM beats ORAS by 2x on 630M unknowns","feed_subtitle":"On the GO 3D seismic benchmark, tuned optimized Schwarz method halves time and memory versus optimized restricted additive Schwarz.","key_machinery":"The load-bearing object is the substructured interface problem $(I - T)g = b$, a Schur complement of the coupled subdomain Helmholtz systems with respect to the interface fields $g$, on which GMRES is applied directly without a preconditioner. The operator $T$ is the exchange operator that maps incoming waves on each interface to outgoing waves, with local solves built in, and the right-hand side $b$ encodes the sources. For both methods the second-order transmission operator $S = \\alpha + \\beta\\Delta_{\\Sigma}$, with complex coefficients $\\alpha$ and $\\beta$ tuned numerically to minimize Krylov iterations, is the main convergence lever; in ORAS the same transmission conditions appear inside the preconditioner $M^{-1}_{\\mathrm{ORAS}} = \\sum_i R_i^T D_i A_i^{-1} R_i$.","core_discovery":"On a realistic geophysical benchmark with up to 630.1 million degrees of freedom and 64 simultaneous sources, the non-overlapping substructured Optimized Schwarz Method with numerically optimized second-order transmission conditions $S = \\alpha + \\beta\\Delta_{\\Sigma}$ outperforms the Optimized Restricted Additive Schwarz preconditioner by roughly a factor of two in total wall time (205.5 s vs 502.6 s for the largest case) while consuming substantially less memory. The advantage comes from three compounding effects: smaller subdomains (no overlap), a substructured interface problem over which GMRES iterates (typically 5 to 10 times fewer unknowns), and replacement of the global sparse matrix-vector product by local interface operations. The paper also shows that with basic zeroth-order conditions ORAS converges faster, but the second-order conditions shrink that gap enough that OSM's cheaper iterations dominate.","pith_inferences":["The residual-to-L2-error calibration was only checked at 2 Hz with 256 subdomains; if the OSM error-to-residual curve drifts at higher frequencies, the same 1e-4 GMRES tolerance may give OSM a looser effective accuracy, so the timing advantage could be partially an accuracy advantage. A direct L2-error comparison at 6.25 Hz would settle this.","Since OSM's savings come from smaller subdomains and interface-sized Krylov spaces, the same qualitative advantage should appear for Maxwell and elasticity equations, provided optimized transmission conditions keep the iteration gap small; the paper notes this extension is open.","The memory headroom OSM frees could be reinvested in larger GMRES restarts or a coarse space rather than just larger batches; the paper does not test this trade, and it may be the fastest path to a two-level OSM.","The break-even with direct solvers depends on source count and factorization reuse; in FWI loops with many adjoint solves, direct solves may win at low frequencies, so hybrid strategies remain attractive."],"forward_implications":["On the tested GO 3D OBS configuration, OSM's total solve time stays below half of ORAS's up to 6.25 Hz and 16,384 subdomains; the factor-of-two gap persists across the whole weak-scaling sweep.","Memory per process is low enough with OSM that a 20-million-unknown problem fits on two LUMI nodes, and the largest 630-million-unknown case fits in the available per-process budget, whereas ORAS requires finer partitioning or smaller batches.","The substructured formulation makes batch solves of 64 right-hand sides affordable: OSM can use the full batch, while ORAS memory constraints force smaller batches, widening the time gap further.","For very large problems the remaining bottleneck is one-level scalability; the authors expect two-level (coarse-space) extensions to be needed beyond the tested sizes.","A direct solver (MUMPS with BLR) on the 1.6 Hz case takes 74 s factorization plus 6 s for 64 sources versus 41 s for OSM with the same resources; the paper estimates the break-even point near 150 sources, beyond which direct solves lose on this problem size."],"supporting_citations":[{"why":"Gander, Magoulès and Nataf: the optimized non-overlapping Schwarz method with transmission conditions whose second-order variant OSM is built on.","marker":"[18]"},{"why":"St-Cyr, Gander and Thomas: the optimized restricted additive Schwarz preconditioner that the paper compares against.","marker":"[38]"},{"why":"Dolean, Jolivet and Nataf: the textbook derivation unifying ORAS and the substructured OSM interface problem $(I-T)g=b$.","marker":"[15]"},{"why":"Antoine and Geuzaine: optimized Schwarz conditions for Helmholtz that motivate the numerically tuned $S=\\alpha+\\beta\\Delta_\\Sigma$.","marker":"[6]"},{"why":"Górszczyk and Operto: the GO 3D OBS velocity model and target region used as the test case.","marker":"[22]"},{"why":"Jolivet and Tournier: block pseudo-block Krylov methods with multiple right-hand sides that determine the batching strategy.","marker":"[24]"},{"why":"Tournier et al.: prior large-scale FWI with optimized additive Schwarz, the main baseline showing that DDM can beat direct solvers and giving the 130-source observation.","marker":"[39]"},{"why":"Mary: complexity analysis of multifrontal solvers used to motivate the memory scaling expectations and the direct-solver comparison.","marker":"[26]"}],"fun_headline_variants":["OSM halves run time vs ORAS on 630M unknowns","Non-overlapping Schwarz beats overlapping ORAS 2x at scale","630M Helmholtz dof: OSM wins in time and memory vs ORAS","Tuned OSM outperforms ORAS by 2x on geophysical benchmark","Substructured OSM speeds up 630M-unknown seismic inversion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimized transmission coefficients and the relationship between residual and true L2 error are calibrated on a 2 Hz, 256-subdomain case; the entire large-scale comparison assumes those settings transfer to every frequency, mesh, and partitioning, and that the same 1e-4 relative GMRES tolerance gives both methods equal true accuracy.","fun_headline_variants_meta":{"raw":{"variants":["OSM halves run time vs ORAS on 630M unknowns","Non-overlapping Schwarz beats overlapping ORAS 2x at scale","630M Helmholtz dof: OSM wins in time and memory vs ORAS","Tuned OSM outperforms ORAS by 2x on geophysical benchmark","Substructured OSM speeds up 630M-unknown seismic inversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000931,"raw_usage":{"total_tokens":3974,"prompt_tokens":925,"completion_tokens":3049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2951}},"tokens_in":541,"tokens_out":3049,"duration_ms":23016,"temperature":1.0,"reasoning_tokens":2951,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:37:33.451991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the 6.25 Hz, 16,384-subdomain case with both methods and compare the true L2 error of the final solution against a reference obtained on a coarser grid or with a different solver. If OSM's solution error is more than a small factor worse than ORAS's at the same 1e-4 residual tolerance, the reported time comparison is biased; alternatively, if ORAS with matched true accuracy and a large enough batch size finishes no slower than OSM, the central claim fails.","supporting_citations":[{"cited_title":"St-Cyr, M","cited_arxiv_id":null,"evidence_quote":"St-Cyr, Gander and Thomas: the optimized restricted additive Schwarz preconditioner that the paper compares against."},{"cited_title":"Antoine and C","cited_arxiv_id":null,"evidence_quote":"Antoine and Geuzaine: optimized Schwarz conditions for Helmholtz that motivate the numerically tuned $S=\\alpha+\\beta\\Delta_\\Sigma$."},{"cited_title":"G ´orszczyk and S","cited_arxiv_id":null,"evidence_quote":"Górszczyk and Operto: the GO 3D OBS velocity model and target region used as the test case."},{"cited_title":"Jolivet and P.-H","cited_arxiv_id":null,"evidence_quote":"Jolivet and Tournier: block pseudo-block Krylov methods with multiple right-hand sides that determine the batching strategy."},{"cited_title":"Tournier, P","cited_arxiv_id":null,"evidence_quote":"Tournier et al.: prior large-scale FWI with optimized additive Schwarz, the main baseline showing that DDM can beat direct solvers and giving the 130-source observation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mary: complexity analysis of multifrontal solvers used to motivate the memory scaling expectations and the direct-solver comparison."}],"review_version":1}