{"id":"6ccbdcd4-e50d-4c82-bc46-257eb56850f0","arxiv_id":"2509.02131","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"In a large reproducible benchmark of two-level domain-decomposition solvers for the heterogeneous Helmholtz equation, harmonic and extended-harmonic spectral coarse spaces outperform DtN and Hk-GenEO in iteration count and size, especially in 3D.","lead":"This paper benchmarks four families of spectral coarse spaces for solving high-frequency wave equations in heterogeneous media with domain decomposition. It finds that harmonic and extended-harmonic spaces give the best trade-off between solver speed and coarse-space size, while DtN and Hk-GenEO spaces struggle in hard 3D cases.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"3D 'breakdown' verdicts for DtN and Hk-GenEO rest on a 200-eigenvector/200-iteration cap; larger budgets are untested.","rationale":"The reader's weakest assumption identifies exactly the premise I would attack. The 3D comparisons in Section 5.3 and 5.4 are run under a uniform cap of 200 coarse vectors per subdomain and 200 GMRES iterations, visible in Tables 6-7 (CSs = 200, entries '>200'). The conclusion that DtN and Hk-GenEO 'breakdown' in 3D and the ✗ entries in Table 8 go beyond what the data show: they demonstrate failure within the chosen budget, not asymptotic non-robustness. Because the headline recommendation — extended-harmonic as best trade-off in high-frequency 3D heterogeneous settings — is explicitly contrasted against DtN and Hk-GenEO, this is load-bearing. A concrete rerun with a larger budget would settle it. I do not see a more serious flaw: the comparison protocol is consistent, the 2D results are extensive, and the paper appropriately disclaims wall-clock time comparisons. The conditional verdict is correct: the paper should disclose the budget and either soften the 'breakdown' language or provide evidence at larger budgets.","tokens_in":27976,"tokens_out":5105,"duration_ms":54235,"concrete_test":"Rerun the two 3D benchmarks (COBRA at N=2916, L=22.9λ, Table 6; GO_3D_OBS at N=2048, Table 7) for DtN and Hk-GenEO with the per-subdomain eigenvector cap removed or raised to at least 1000 (selecting via the same Re(λ)<ηmax / τ criteria that the paper used for 2D) and with the GMRES iteration limit raised to at least 2000. Record iteration counts and coarse-space sizes. If both methods converge to 1e-6 under the enlarged budget, the 'breakdown' verdict is an artifact of the 200-vector/200-iteration cap and the paper's Table 8 should be revised; if they still stall, the cap was not the cause and the 3D ranking is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise for the 3D negative verdicts is a computational budget that is not disclosed as a limiting assumption. In Tables 6 and 7, every method is allotted exactly 200 coarse basis vectors per subdomain (column CSs = 200) and GMRES is permitted at most 200 iterations (entries '>200' indicate non-convergence within this cap). Section 5.3 and 5.4 then conclude that DtN and Hk-GenEO 'breakdown' or provide 'no gain' in 3D, and Table 8 carries these verdicts. But the experiments never test whether DtN or Hk-GenEO would converge with, say, 400 or 1000 vectors per subdomain and a larger iteration limit. Since the number of interface DOFs n∂Ωs is ~5000, such larger spaces are feasible. 'Breakdown within a fixed budget' is not the same as 'inherently not robust'; the paper conflates the two. The headline claim that extended-harmonic spaces are the best trade-off in high-frequency 3D heterogeneous settings depends directly on this comparison, because DtN and Hk are the alternatives that are dismissed. If they merely need a larger budget, the ordering may survive but the 'breakdown' language and the ✗ entries in Table 8 are overstated, and a practitioner could reasonably choose DtN or Hk with a larger coarse space.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a unified numerical comparison of four spectral coarse spaces—DtN, Hk-GenEO, harmonic, and extended-harmonic—used in two-level ORAS preconditioners for the heterogeneous Helmholtz equation. The study is conducted in the FreeFEM/ffddm framework across 2D square, medical-imaging, COBRA-cavity, and GO_3D_OBS benchmarks, and reports iteration counts, coarse-space sizes, and scaling behavior. The central claim is that extended-harmonic coarse spaces offer the best trade-off between solver efficiency and coarse-space size, while DtN and Hk-GenEO, though competitive in 2D, break down in large 3D problems.","tokens_in":28264,"tokens_out":5218,"duration_ms":61280,"significance":"If the comparison is accepted, this is a practically valuable benchmark study for a community that lacks definitive guidance among spectral coarse spaces for high-frequency heterogeneous Helmholtz problems. The paper benefits from a single common implementation, large-scale 3D tests, a clear comparison framework, and a publicly available example script. However, the strength of the conclusions depends on the fairness and completeness of the numerical protocol, especially for the 3D negative verdicts.","major_comments":[{"comment":"The 3D 'breakdown' verdicts for DtN and Hk-GenEO rest on runs with a fixed per-subdomain budget of exactly 200 coarse vectors (CSs=200 in every row) and GMRES stopped at 200 iterations (entries '>200'). The manuscript does not test whether these methods would converge with, say, 400 or 1000 vectors per subdomain, even though the reported n∂Ωs ≈ 5000 makes such budgets feasible. 'Breakdown within a fixed budget' is not equivalent to 'inherently not robust'; the paper conflates the two. Since Table 8 and Section 6 carry these verdicts, this is a load-bearing gap: either experiments with larger budgets are needed, or the claims must be reformulated as budget-limited observations.","section":"§5.3, Tables 6–7; §5.4, Table 7; §6, Table 8"},{"comment":"The headline tables report 'the parameters with the minimum number of iterations obtained for each method.' This means every method is evaluated at its individually tuned best. The resulting iteration counts and coarse-space sizes are therefore not produced under a common resource constraint. The threshold-sweep figures provide useful sensitivity information, but the 'best trade-off' conclusion in Section 6 would be more solid if the cross-method comparison were also made at a fixed budget or fixed threshold, rather than only at per-method optima.","section":"§5.1.3; also §5.2–§5.3"}],"minor_comments":[{"comment":"The column header 'CSs' is defined only in the text of §5.1.3. Each table caption should repeat the definition, because the repeated 'It CS CS s' blocks are otherwise hard to parse.","section":"§5.1.3, Tables 2–7"},{"comment":"The typo 'Matèrn' for the point-process name should be corrected to 'Matérn' or 'Matern'.","section":"§5.2"},{"comment":"The symbols ✓, ✓✓, and ✓/✗ are not quantitatively defined. State the criteria (e.g., convergence within a given iteration budget, relative coarse-space size) so the table is reproducible rather than subjective.","section":"Table 8"},{"comment":"The extended-harmonic eigenproblem is dense and introduced quickly. A short paragraph interpreting the action of Rs ˇRT_s − bA^{-1}_s Rs ˇRT_s ˇAs on a candidate vector would help the reader understand why the construction yields Helmholtz-harmonic coarse functions.","section":"§4.3.2, Eq. (4.18)"},{"comment":"The paper explicitly leaves setup and GMRES run times out of scope, yet the conclusion uses the word 'trade-off' repeatedly. A reader cannot fully judge practical trade-offs without at least an indication of setup costs; this should be acknowledged more prominently in the abstract or conclusions.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is authored by researchers who developed several of the compared methods and the ffddm framework. The numerical comparison is careful, but independent validation or full release of all benchmark scripts would strengthen the conclusions. The fixed 200-vector/200-iteration cap in the 3D experiments is the main correctness risk and should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Clear take: this is a genuinely useful benchmark paper, and the central ranking—harmonic and extended-harmonic coarse spaces beat DtN and Hk-GenEO at a given coarse-space size in these tests—is credible as far as iteration counts go. But the 'breakdown' verdicts for DtN and Hk-GenEO in 3D are overstated. Tables 6 and 7 cap every method at 200 eigenvectors per subdomain and stop GMRES at 200 iterations. Within that budget, DtN and Hk-GenEO fail on the largest COBRA cases and on GO_3D_OBS, and the paper then reports them as ✗ in Table 8 and says they 'breakdown in 3D.' That conflates a budget-limited failure with an inherent one. The interface has around 5000 dofs in those tests, so a 400 or 1000-vector space is feasible; the paper never tests it. The right wording is 'did not converge within the tested budget,' and the ranking might survive—extended-harmonic reaching low counts with only 200 vectors is exactly the practical claim—but the collapse language is wrong.\n\nWhat's new and good: this is the first common-framework comparison that includes the harmonic and extended-harmonic families from 2024 preprints alongside DtN and Hk-GenEO, with honest 3D weak/strong scaling up to 44M dofs. Protocol is consistent (fixed overlap, one subdomain per MPI process, GMRES tol 1e-6), and the paper is refreshingly explicit about what it does not measure: wall-clock times. That missing piece matters because the abstract and conclusions keep saying 'trade-off,' but trade-off between iteration count and coarse-space size is not the full cost picture, especially for methods with different eigenproblem setups.\n\nOther soft spots: the headline tables pick each method at its individually best threshold (Section 5.1.3), which is standard but flatters methods with sensitive thresholds; the authors' own ties to ffddm and to the extended-harmonic and Hk-GenEO papers are a mild conflict worth noting, though the benchmarks are external; and only the 2D homogeneous script is public, with no parameter files for the 3D cases.\n\nBottom line: practitioners choosing a coarse space for high-frequency heterogeneous Helmholtz get real value from this paper. It deserves peer review, with a referee pushing the authors to soften the breakdown language and ideally to add wall-clock times or larger-budget experiments. I'd send it back for minor revision.","headline":"Solid benchmark with a credible ranking of Helmholtz coarse spaces, but the 3D 'breakdown' of DtN and Hk-GenEO is a fixed-budget artifact, not an established failure.","tokens_in":28863,"tokens_out":4856,"would_cite":true,"duration_ms":53246,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65N35","65F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that harmonic and extended-harmonic spectral coarse spaces give the best balance of GMRES efficiency and coarse-space size for high-frequency heterogeneous Helmholtz solvers, while DtN and Hk-GenEO spaces break down in the","keywords":["Helmholtz equation","domain decomposition","coarse space","GenEO","harmonic coarse space","DtN coarse space","wavenumber robustness","two-level preconditioner"],"falsifier":"Rerun the two 3D benchmarks where DtN and Hk-GenEO break down—the COBRA cavity at k=360 m^-1 with 2916 subdomains and the GO_3D_OBS test with 1024 subdomains—allowing 400 or 1000 eigenvectors per subdomain and up to 1000 GMRES iterations. If either method then reaches the 1e-6 residual tolerance with moderate iteration counts, the reported 3D breakdown is an artifact of the 200-dimensional budget rather than a genuine robustness failure.","tokens_in":27809,"feed_emoji":"🌊","tokens_out":11308,"duration_ms":109380,"temperature":0.7,"pith_summary":"This paper compares four families of spectral coarse spaces for two-level overlapping Schwarz solvers for the heterogeneous Helmholtz equation, asking which construction delivers wavenumber robustness at acceptable coarse-space cost. Its central claim is that harmonic and extended-harmonic coarse spaces offer the best trade-off: they keep GMRES iteration counts low, even under weak scaling and in high-frequency three-dimensional heterogeneous settings, while using comparatively compact coarse spaces. The paper also reports that the established DtN and Hk-GenEO coarse spaces remain competitive in two dimensions and at moderate problem sizes, but fail to converge in its three-dimensional benchmarks. If correct, the practical consequence is a clear ordering: for challenging 3D problems, prefer extended-harmonic or harmonic coarse spaces, and reserve DtN and Hk-GenEO for cheaper 2D or moderate-scale settings.","feed_headline":"Harmonic coarse spaces win 3D high-frequency Helmholtz comparison","feed_subtitle":"They keep GMRES iterations low with compact coarse spaces, even under weak scaling in heterogeneous 3D media.","key_machinery":"The central object is the spectral coarse space: a set of global basis functions assembled by partition of unity from eigenvectors of local eigenproblems. The decisive difference between families is which local problem is diagonalized and which vectors are admitted. DtN diagonalizes an interface Dirichlet-to-Neumann map and extends the eigenfunctions by a Helmholtz extension. GenEO-type methods diagonalize volumetric operators; Hk-GenEO uses a positive-definite surrogate operator close to the indefinite Helmholtz operator. Harmonic and extended-harmonic spaces diagonalize self-adjoint, coercive eigenproblems whose eigenfunctions are constrained to be local Helmholtz-harmonic fields (solution","core_discovery":"The discovery is comparative and empirical. Within a common two-level ORAS (optimized restricted additive Schwarz) preconditioner and a non-restarted GMRES solver, the paper tests DtN (Dirichlet-to-Neumann), GenEO-type (Generalized Eigenproblems in the Overlap, specifically Hk-GenEO), harmonic, and extended-harmonic coarse spaces on problems from a homogeneous square to the COBRA cavity and the GO_3D_OBS crustal geomodel. Every two-level method beats the one-level baseline by a wide margin, but the balance of robustness and coarse-space size shifts with dimensionality: harmonic and extended-harmonic spaces give the lowest iteration counts for a given coarse-space size and keep those counts u","pith_inferences":["Because the paper's 3D 'breakdown' verdicts are made under a fixed budget of 200 eigenvectors per subdomain and 200 GMRES iterations, an immediate extension is to rerun those benchmarks with larger budgets; convergence there would demote the failure from structural to budget-limited.","The fact that the robust harmonic spaces stay roughly interface-sized suggests the information needed for wavenumber robustness lives on subdomain interfaces, so a hybrid that combines DtN-style boundary extraction with Helmholtz-harmonic projection might produce even cheaper robust spaces.","If coarse-space solve cost dominates each GMRES iteration, total wall-clock time in 2D could still favor DtN despite slightly higher iteration counts; the paper leaves setup and solve timing for future work, so the runtime ranking is still open.","The fixed first-level preconditioner, solver, meshes, and threshold sweeps make this comparison directly reusable: any new coarse space can be ranked on the same coarse-space-size-versus-iteration curves without new theory."],"forward_implications":["For high-frequency 3D heterogeneous Helmholtz problems, practitioners should build two-level ORAS preconditioners with extended-harmonic or harmonic coarse spaces rather than DtN or Hk-GenEO spaces.","Two-level coarse correction is not optional: in these tests every spectral coarse space dramatically outperforms the one-level ORAS baseline as frequency or domain size grows.","Wavenumber robustness is achievable with compact coarse spaces whose dimension tracks the number of subdomain-interface unknowns, provided the selected modes are genuinely Helmholtz-harmonic.","Working eigenvalue thresholds are roughly 5–10 for extended-harmonic, about 10 for harmonic, 500–1000 for DtN, and 0.8–1 for Hk-GenEO.","In 3D, DtN and Hk-GenEO should be used with caution on large domains: the paper reports breakdown on the COBRA cavity and GO_3D_OBS benchmarks."],"supporting_citations":[{"why":"Introduces the DtN coarse space for heterogeneous Helmholtz problems and the eigenvalue threshold criterion used in the comparison.","marker":"[14]"},{"why":"Introduces the harmonic coarse space built from local Helmholtz-harmonic functions via a self-adjoint eigenproblem.","marker":"[38]"},{"why":"Introduces the multiscale spectral generalized finite element (MS-GFEM) based harmonic coarse space and proves the GMRES convergence rate follows the exponential MS-GFEM approximation error.","marker":"[42]"},{"why":"Introduces the extended-harmonic coarse space in a fully algebraic setting, providing the construction that performs best in 3D.","marker":"[43]"},{"why":"Introduces the Hk-GenEO coarse space with theoretical estimates, the GenEO variant whose 2D robustness and 3D breakdown are evaluated.","marker":"[13]"},{"why":"Provides the earlier comparison of DtN and GenEO coarse spaces whose methodology this study extends to harmonic spaces and 3D benchmarks.","marker":"[11]"},{"why":"Calibrates DtN and Hk-GenEO eigenvalue thresholds and supplies empirical wavenumber-scaling formulas used to set parameters.","marker":"[8]"},{"why":"Supplies the GO_3D_OBS crustal geomodel benchmark used for the heterogeneous 3D strong-scaling test where DtN and Hk-GenEO fail.","marker":"[31]"},{"why":"Supplies the COBRA cavity geometry used for the 3D weak-scaling experiments.","marker":"[41]"}],"fun_headline_variants":["Harmonic coarse spaces lead in 3D Helmholtz solver efficiency","Spectral coarse spaces: harmonic wins on 3D heterogeneous Helmholtz","3D Helmholtz: harmonic coarse spaces keep GMRES iterations low","Best coarse space for high-frequency Helmholtz? Harmonic in 3D","Harmonic coarse spaces dominate GMRES efficiency in 3D"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that failing to converge within 200 eigenvectors per subdomain and 200 GMRES iterations counts as genuine loss of robustness; the paper reports no experiments with larger budgets that would rule out a budget-limited failure.","fun_headline_variants_meta":{"raw":{"variants":["Harmonic coarse spaces lead in 3D Helmholtz solver efficiency","Spectral coarse spaces: harmonic wins on 3D heterogeneous Helmholtz","3D Helmholtz: harmonic coarse spaces keep GMRES iterations low","Best coarse space for high-frequency Helmholtz? Harmonic in 3D","Harmonic coarse spaces dominate GMRES efficiency in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1325,"prompt_tokens":741,"completion_tokens":584,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":500}},"tokens_in":485,"tokens_out":584,"duration_ms":6161,"temperature":1.0,"reasoning_tokens":500,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:51:01.990703+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the two 3D benchmarks where DtN and Hk-GenEO break down—the COBRA cavity at k=360 m^-1 with 2916 subdomains and the GO_3D_OBS test with 1024 subdomains—allowing 400 or 1000 eigenvectors per subdomain and up to 1000 GMRES iterations. If either method then reaches the 1e-6 residual tolerance with moderate iteration counts, the reported 3D breakdown is an artifact of the 200-dimensional budget rather than a genuine robustness failure.","supporting_citations":[{"cited_title":"Analytical and numerical studies of a finite element PML for the Helmholtz equation","cited_arxiv_id":null,"evidence_quote":"Introduces the harmonic coarse space built from local Helmholtz-harmonic functions via a self-adjoint eigenproblem."},{"cited_title":"Karypis and V","cited_arxiv_id":null,"evidence_quote":"Introduces the multiscale spectral generalized finite element (MS-GFEM) based harmonic coarse space and proves the GMRES convergence rate follows the exponential MS-GFEM approximation error."},{"cited_title":"Coarse spaces for non-symmetric two-level preconditioners based on local extended generalized eigenproblems","cited_arxiv_id":"2404.02758","evidence_quote":"Introduces the extended-harmonic coarse space in a fully algebraic setting, providing the construction that performs best in 3D."},{"cited_title":"Schwarz preconditioner with $H_k$-GenEO coarse space for the indefinite Helmholtz problem","cited_arxiv_id":"2406.06283","evidence_quote":"Introduces the Hk-GenEO coarse space with theoretical estimates, the GenEO variant whose 2D robustness and 3D breakdown are evaluated."},{"cited_title":"Can DtN and GenEO Coarse Spaces Be Sufficiently Robust for Heteroge- neous Helmholtz Problems?","cited_arxiv_id":null,"evidence_quote":"Calibrates DtN and Hk-GenEO eigenvalue thresholds and supplies empirical wavenumber-scaling formulas used to set parameters."},{"cited_title":"GO_3D_OBS: the multi-parameter benchmark geomodel for seismic imaging method assessment and next-generation 3D survey design (version 1.0)","cited_arxiv_id":null,"evidence_quote":"Supplies the GO_3D_OBS crustal geomodel benchmark used for the heterogeneous 3D strong-scaling test where DtN and Hk-GenEO fail."},{"cited_title":"Scattering analysis of a large body with deep cavities","cited_arxiv_id":null,"evidence_quote":"Supplies the COBRA cavity geometry used for the 3D weak-scaling experiments."}],"review_version":1}