{"id":"6ae4783f-c162-4c31-b848-028a9e78bdf0","arxiv_id":"2507.13066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Benchmarking five solvers for time-harmonic Maxwell equations, the paper finds Hiptmair-Xu and Block Low-Rank preconditioners are the most effective in the tested configurations.","lead":"This paper compares four preconditioners and a direct solver for large-scale time-harmonic Maxwell equations, using test problems with up to 38 million unknowns. The authors report that a Hiptmair-Xu-type preconditioner and a Block Low-Rank factorization perform best in their experiments, but the conclusions are preliminary.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BLR recommendation is supported by only one test matrix (Table 7); without a parameter sweep across k, mesh, and domain, the conclusion that BLR offers a useful speed/memory compromise is not established.","rationale":"The reader's conditional verdict is appropriate. I considered the HX preconditioner's missing sqrt(2) condition-number bound, but the paper explicitly flags that the bound is not recovered, and the HX claim of mesh-independence is supported across several meshes and wavenumbers, albeit with known deterioration at higher k and larger domains. The less secure part of the central claim is the BLR conclusion: it is derived from a single matrix, and Section 4 generalizes it without qualification. The paper's 'temporary conclusions' language is an honest hedge, but a recommendation in a comparative solver paper should not rest on one case when the other leg (HX) has been studied systematically. A concrete multi-case BLR sweep would settle whether the favorable compression/iteration trade-off is a property of the method or of that particular matrix. Since the reader already conditioned on similar points, I leave the verdict unchanged.","tokens_in":109,"tokens_out":8917,"duration_ms":117283,"concrete_test":"Use the Table 7 protocol on the k=10 unit-cube matrices from Table 4 (e.g., 2×172,305 and 2×556,510 dofs) and on the enlarged-domain matrix from Table 6 (k=1, 10 ppw, 32 cores), recording FGMRES iterations and compression ratio at ε=10^-3. If iterations exceed roughly 50 or the compression ratio drops below roughly 2 on these additional cases, the BLR recommendation does not generalize beyond the original test case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two legs: HX mesh-independence and BLR efficiency. The HX leg is supported by multiple tables (Tables 3–6) with varying k, mesh size, and domain. The BLR leg rests on a single matrix: Table 7 (2×2,474,860 dofs, k=1, 10 ppw, 32 cores, unit cube with L-shaped scatterer). Table 8 merely compares that same case. No evidence is given that the favorable ε=10^-3 trade-off (10 FGMRES iterations, factor-size compression 2.05) persists for other wavenumbers, mesh sizes, or scatterer geometries. At higher k, the off-diagonal blocks of the matrix are less likely to be low-rank, so the compression/accuracy trade-off could shift unfavorably. Section 4 states 'with BLR, there is a compromise to be found' as a general conclusion, which is stronger than the single-case evidence warrants. The paper's 'work in progress / temporary conclusions' hedge mitigates but does not remove the inductive gap. The absence of code/data also prevents independent confirmation of the reported compression ratios.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an empirical comparison of four preconditioning strategies—ParaSails sparse approximate inverse, one-level overlapping restricted additive Schwarz (RAS), a Hiptmair-Xu (HX) based block-diagonal preconditioner, and MUMPS Block Low-Rank (BLR) factorization—together with full-rank LU, for time-harmonic Maxwell systems discretized by lowest-order Nédélec elements. The real/imaginary split formulation (11) is used for the hypre-based solvers, and experiments range up to about 38 million degrees of freedom on up to 400 CPU cores. The principal claims are that HX yields mesh-independent FGMRES iteration counts for the unit-cube/L-shaped-scatterer test case, and that BLR provides an attractive compression/accuracy trade-off; the paper explicitly labels these as temporary conclusions.","tokens_in":10257,"tokens_out":7902,"duration_ms":93244,"significance":"If the reported results hold, this is a useful practitioner-oriented comparison because it exercises readily available library preconditioners at scale and honestly reports a negative theoretical result (the sqrt(2) condition-number bound from [26] is not recovered for the indefinite split system). No fitted constants enter the reported quantities, and the HX leg is supported by several tables with stable iteration counts across mesh sizes. The BLR leg, however, rests on a single large test case, and no code or data are provided for independent verification. The paper's empirical claims are therefore credible but not yet at the level of a general recommendation for the BLR strategy.","major_comments":[{"comment":"The BLR recommendation is supported by a single matrix (2x2,474,860 dofs, k=1, 10 ppw, 32 cores), and Table 8 only reuses that same case. The general statement in Section 4 that \"with BLR, there is a compromise to be found\" is stronger than this evidence: at higher wavenumbers, different mesh sizes, or different scatterer geometries, the off-diagonal blocks may become less low-rank and the compression/accuracy trade-off could shift unfavorably. I recommend either adding a parameter sweep in k, mesh size, and domain size for BLR, or explicitly limiting the BLR conclusion to the demonstrated regime.","section":"Section 3.4, Table 7"},{"comment":"The reported \"compression\" is the ratio of factor sizes and the reported time is the total solve time, so the memory-savings and speed claims for BLR cannot be separated into factorization/setup cost versus FGMRES iteration cost. Please report factorization time, memory usage, and iteration time as separate quantities, or otherwise substantiate the memory-savings claim.","section":"Section 3.4, Table 7"},{"comment":"The mesh-independence conclusion for HX is empirical, and the only theoretical motivation cited for the block-diagonal preconditioner in (15) is the sqrt(2) condition-number bound from [26], which the authors state is not recovered. Because the preconditioner may be a poor approximation for other wavenumbers, domain sizes, or geometries, I ask for a quantitative diagnostic on at least one additional configuration (for example, estimated condition numbers of the preconditioned system or a second geometry) or a more explicit limitation statement in Section 4, so that the reader can assess how far the mesh-independence observation is expected to generalize.","section":"Section 3.3, Eq. (15)"}],"minor_comments":[{"comment":"The table headers \"HX\" and \"HX as solver\" are not defined in the captions; the text explains them, but a brief caption note would help. Similarly, the column \"#CG\" should state whether it refers to inner CG iterations for both block solves or their total.","section":"Section 3.3, Table 3"},{"comment":"The statement that \"the number of FGMRES iterations is at least proportional to k^0.86\" appears to be based on only a few wavenumber values; please specify the fitting procedure or present the data points supporting this exponent.","section":"Section 3.3"},{"comment":"The claim that \"with more dofs (not shown here), the number of GMRES iterations reaches the upper limit of 1000\" is not supported by any table; either include the data or rephrase this as a qualitative observation.","section":"Section 3.1"},{"comment":"There is a typo in the text \"preconditoner\" which should be \"preconditioner.\"","section":"Section 3.4"},{"comment":"No code or data availability statement is provided. Given that all central conclusions are empirical, a reproducibility statement or a link to the experimental data and scripts would substantially strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is explicitly a work-in-progress report, and the \"temporary conclusions\" hedge mitigates the risk of overclaiming. However, the BLR recommendation is the weaker leg of the central claim, and the absence of any BLR experiments beyond one matrix makes the Section 4 conclusion difficult to evaluate. The HX evidence is considerably stronger, and the honest reporting of the un-recovered condition-number bound is a positive feature. If the journal values broad empirical comparisons, a major revision with additional BLR data or a clearly scoped claim would suffice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a benchmark paper, not a methods paper, and the most useful result is the HX one. The authors take the block-diagonal preconditioner from Grayver/Burg and Chen et al., apply it to a scattering problem with real permittivity and impedance boundary condition, and show empirically that FGMRES iteration counts are essentially flat as the mesh is refined for k=1 and k=10 on the unit cube. They are upfront that they do not recover the sqrt(2) condition-number bound. That honesty carries through the paper: the SAI and RAS results show them struggling to converge, and they say so plainly. The HX tables are the strongest part because they vary mesh size, wavenumber, domain size, and core count. The scaling results and the failure of the direct solver on the largest problems are also useful.\n\nThe soft spot is exactly where the stress-test note points. The BLR recommendation rests on Table 7, one matrix at k=1, one geometry, one core count. Table 8 compares BLR to HX only on that same case. The text says \"with BLR, there is a compromise to be found\" as if it were a general observation, but it is an observation about one matrix. At higher wavenumber the off-diagonal blocks are less likely to compress well, so the trade-off may not survive. The other issue is reproducibility: no code or data are included. For a benchmark whose whole value is the numbers, that is a real limitation. I don't think the missing sqrt(2) bound is itself a flaw—the paper never claims it—but it means the HX mesh-independence is empirical, and the range of tested wavenumbers is modest.\n\nThe citation pattern is fine, and the list of untested methods is a plus. No circularity: the parameters are set before the runs and the conclusions are observations.\n\nWho is this for? Practitioners choosing a solver for time-harmonic Maxwell with Nedelec elements, and people working on HX-type preconditioners. It deserves a serious referee. I would send it to review, but I would ask for code/data and at least a second BLR test case—different wavenumber or geometry—before accepting. As is, it is a solid work-in-progress report, not yet a reference result.","headline":"An honest, useful benchmark whose HX mesh-independence result is credible, but whose BLR conclusion rests on a single matrix and no code or data are shipped.","tokens_in":10846,"tokens_out":2612,"would_cite":false,"duration_ms":31499,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F08","65F10","65N22","78M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For large time-harmonic Maxwell systems, Hiptmair-Xu and Block Low-Rank preconditioning emerge as the two viable strategies among those tested.","keywords":["time-harmonic Maxwell equations","preconditioned iterative solvers","Hiptmair-Xu preconditioner","Block Low-Rank factorization","Nédélec finite elements","FGMRES","mesh-size independence","large-scale electromagnetic scattering"],"falsifier":"Run the same FGMRES plus Hiptmair-Xu pipeline on the L-shaped-scatterer geometry at a fixed wavenumber and fixed points per wavelength while refining the mesh, and record the outer iteration count; significant growth with refinement would contradict the claimed mesh independence. Independently, push the wavenumber above 15 on the enlarged domain, where the paper's own scaling already shows iteration growth at least like $k^{0.86}$; a falsifier is observing that growth become superlinear or that FGMRES stalls before reaching the $10^{-8}$ relative residual target.","tokens_in":9851,"feed_emoji":"⚡","tokens_out":6700,"duration_ms":72483,"temperature":0.7,"pith_summary":"The paper tries to show which preconditioning strategies can solve the large, sparse, indefinite, non-Hermitian linear systems that arise when time-harmonic Maxwell equations are discretized with lowest-order Nédélec finite elements. It compares five approaches, including no preconditioner, sparse approximate inverse, restricted additive Schwarz, Hiptmair-Xu, and Block Low-Rank factorization, on problems with up to 38 million degrees of freedom. Its working conclusion is that Hiptmair-Xu preconditioning keeps outer iteration counts independent of mesh refinement for a fixed wavenumber, while Block Low-Rank preconditioning gives fast convergence with compressed memory requirements. This matters because direct LU factorization becomes too slow and memory-hungry for realistic large-scale electromagnetic scattering, and the underlying operator is neither Hermitian nor definite, so standard iterative methods struggle without preconditioning.","feed_headline":"Two preconditioners win for time-harmonic Maxwell systems","feed_subtitle":"Five solvers compared on up to 38 million degrees of freedom; these two converge fastest and scale best.","key_machinery":"The machinery has three parts. First, the complex Maxwell system is split into real and imaginary parts, turning the original equation into a $2 \\times 2$ block system whose diagonal blocks are $C - M$ (a curl-curl term minus a mass term) with boundary coupling $B$. Second, the Hiptmair-Xu preconditioner for the positive Maxwell operator $C + M + B$ is used to approximately invert the diagonal blocks of a block-diagonal preconditioner $$P = \\mathrm{diag}(C + M + B,\\; C + M + B),$$ with FGMRES as the outer solver and either preconditioned conjugate gradient or Hiptmair-Xu itself as the inner solver. Third, the Block Low-Rank preconditioner is a compressed LU factorization that replaces eligible off-diagonal dense blocks by low-rank approximations, controlled by a threshold $\\varepsilon$, and then applies the compressed factors inside FGMRES.","core_discovery":"On the paper's own terms, the discovery is a practical ranking of solver strategies for time-harmonic Maxwell scattering: among the tested methods, the Hiptmair-Xu preconditioner, adapted by splitting the field into real and imaginary parts and preconditioning each block by the positive Maxwell operator, and the Block Low-Rank compressed factorization, used as a preconditioner, are the two candidate strategies that deserve further development. Hiptmair-Xu keeps the outer FGMRES iteration count essentially constant as the mesh is refined for a fixed wavenumber, but the count grows with wavenumber and with the physical size of the domain. Block Low-Rank, applied inside FGMRES, converges in very few outer iterations at moderate compression thresholds and reduces both factorization time and memory relative to full-rank LU. The paper presents these as temporary conclusions from a work in progress, not as a proven theory.","pith_inferences":["The paper's real/imaginary split doubles the number of unknowns, so its time and memory comparisons would shift even further in favor of the iterative solvers if the formulation were taken into account in the accounting.","A natural untested combination is to use the Block Low-Rank factorization to invert the diagonal blocks of the positive Maxwell preconditioner inside the Hiptmair-Xu approach, potentially reducing both outer iterations and memory at once.","The strong deterioration on the 50-times-larger domain suggests that Hiptmair-Xu in this form may be limited to moderate electrical sizes, and that sweeping or absorption-based domain decomposition strategies, which the paper lists as untested, would be needed to extend it.","A concrete prediction from the paper's scaling is that for fixed points per wavelength the Hiptmair-Xu outer iteration count stays flat under mesh refinement but grows roughly like $k^{0.86}$ with wavenumber; testing other scatterer geometries would show whether that exponent is universal or geometry-dependent."],"forward_implications":["If the observed mesh-size independence holds more broadly, users can refine the mesh to control discretization error without increasing the outer iteration count for a fixed wavenumber, making large simulations more predictable.","Block Low-Rank preconditioning lets practitioners keep some of the robustness of direct solvers while using less memory, and in the paper it succeeds on the 38-million-degree-of-freedom case where full-rank LU fails.","The parallel scaling results suggest that iterative approaches benefit more from additional CPU cores than the direct solver, so they are the more natural route to very large electromagnetic problems.","The growth of Hiptmair-Xu iterations with wavenumber and physical domain size signals that neither method will automatically handle high-frequency regimes, and the Block Low-Rank compression threshold sets a tunable trade-off among accuracy, speed, and memory."],"supporting_citations":[{"why":"Defines the Hiptmair-Xu auxiliary-space preconditioner for H(curl) problems that is the core of the positive-Maxwell block inversion.","marker":"[7]"},{"why":"Provides the lemma bounding the condition number by $\\sqrt{2}$ for block-diagonal positive Maxwell preconditioning of a split time-harmonic system, the theoretical template the paper adapts.","marker":"[26]"},{"why":"Motivates the split-system block-diagonal preconditioning strategy for time-harmonic Maxwell equations through large-scale geoelectromagnetic modeling.","marker":"[24]"},{"why":"Offers a related robust finite-element formulation for time-harmonic electromagnetic modeling that supports the block preconditioning design.","marker":"[25]"},{"why":"Provides the direct LU solver used as the comparison baseline and the implementation of the Block Low-Rank factorization.","marker":"[20]"},{"why":"Supplies the auxiliary-space Maxwell solver implementation used for the Hiptmair-Xu preconditioning experiments.","marker":"[21]"},{"why":"Describes the parallel auxiliary-space AMG method for H(curl) on which the implementation of the Hiptmair-Xu preconditioner relies.","marker":"[22]"},{"why":"Establishes the complexity estimates for Block Low-Rank multifrontal factorization that justify the claimed memory and FLOP savings.","marker":"[31]"},{"why":"Explains how the compression threshold $\\varepsilon$ controls the low-rank approximation quality and the resulting factor size.","marker":"[32]"}],"fun_headline_variants":["Hiptmair-Xu and block low-rank lead Maxwell solver race","Maxwell solvers: Hiptmair-Xu and low-rank LU top list","Two preconditioners edge out three others for Maxwell","Hiptmair-Xu and low-rank factorization win for Maxwell"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is empirical: the positive Maxwell operator $C + M + B$ used in the block-diagonal preconditioner is close enough to the indefinite split system that a few inner iterations still make the outer FGMRES converge. The paper does not prove this and explicitly notes that the $\\sqrt{2}$ condition-number bound is not recovered, so higher wavenumbers, larger domains, or different scatterer geometries could break the mesh-independence and speed claims.","fun_headline_variants_meta":{"raw":{"variants":["Hiptmair-Xu and block low-rank lead Maxwell solver race","Maxwell solvers: Hiptmair-Xu and low-rank LU top list","Two preconditioners edge out three others for Maxwell","Hiptmair-Xu and low-rank factorization win for Maxwell"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001149,"raw_usage":{"total_tokens":4748,"prompt_tokens":914,"completion_tokens":3834,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":3758}},"tokens_in":530,"tokens_out":3834,"duration_ms":30489,"temperature":1.0,"reasoning_tokens":3758,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:31:10.546830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same FGMRES plus Hiptmair-Xu pipeline on the L-shaped-scatterer geometry at a fixed wavenumber and fixed points per wavelength while refining the mesh, and record the outer iteration count; significant growth with refinement would contradict the claimed mesh independence. Independently, push the wavenumber above 15 on the enlarged domain, where the paper's own scaling already shows iteration growth at least like $k^{0.86}$; a falsifier is observing that growth become superlinear or that FGMRES stalls before reaching the $10^{-8}$ relative residual target.","supporting_citations":[{"cited_title":"Nodal auxiliary space preconditioning in H(curl) and H(div) spaces","cited_arxiv_id":null,"evidence_quote":"Defines the Hiptmair-Xu auxiliary-space preconditioner for H(curl) problems that is the core of the positive-Maxwell block inversion."},{"cited_title":"An adaptive finite element method for the eddy current model with circuit/field couplings","cited_arxiv_id":null,"evidence_quote":"Provides the lemma bounding the condition number by $\\sqrt{2}$ for block-diagonal positive Maxwell preconditioning of a split time-harmonic system, the theoretical template the paper adapts."},{"cited_title":"Large-scale 3d geoelectromagnetic modeling using parallel adaptive high-order finite element method","cited_arxiv_id":null,"evidence_quote":"Motivates the split-system block-diagonal preconditioning strategy for time-harmonic Maxwell equations through large-scale geoelectromagnetic modeling."},{"cited_title":"Grayver and Markus B¨ urg","cited_arxiv_id":null,"evidence_quote":"Offers a related robust finite-element formulation for time-harmonic electromagnetic modeling that supports the block preconditioning design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the direct LU solver used as the comparison baseline and the implementation of the Block Low-Rank factorization."},{"cited_title":"hypre: High Performance Preconditioners","cited_arxiv_id":null,"evidence_quote":"Supplies the auxiliary-space Maxwell solver implementation used for the Hiptmair-Xu preconditioning experiments."},{"cited_title":"Kolev and Panayot S","cited_arxiv_id":null,"evidence_quote":"Describes the parallel auxiliary-space AMG method for H(curl) on which the implementation of the Hiptmair-Xu preconditioner relies."},{"cited_title":"On the complex- ity of the block low-rank multifrontal factorization","cited_arxiv_id":null,"evidence_quote":"Establishes the complexity estimates for Block Low-Rank multifrontal factorization that justify the claimed memory and FLOP savings."},{"cited_title":"Improving multifrontal methods by means of block low-rank representations","cited_arxiv_id":null,"evidence_quote":"Explains how the compression threshold $\\varepsilon$ controls the low-rank approximation quality and the resulting factor size."}],"review_version":1}