{"id":"7b8d308c-e993-4982-aa22-cc3b7eee4c65","arxiv_id":"2412.10910","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A matrix-free non-nested multigrid method for Lagrange finite elements is implemented in deal.II and shown robust on Poisson and elasticity benchmark problems.","lead":"This paper presents a parallel, matrix-free implementation of non-nested geometric multigrid in the deal.II library, letting each solver level use an independently generated mesh. It is worth reading because it brings multigrid performance to difficult CAD-based geometries without requiring nested mesh hierarchies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The §3.3 curved-boundary transfer heuristic (unstated 'large enough' tolerance plus Eq. (17) nearest-point projection) is untested for sensitivity; if it degrades accuracy, the robustness claim for curved geometries in §4.3 is unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the curved-boundary handling in §3.3 is heuristic and untested. My reading of the full text confirms this is the most consequential gap. The paper provides substantial independent support: a public implementation in deal.II, a tutorial, reproducible meshes and CAD files, nested sanity checks (Tables 1 and 2), and extensive experiments on L-shaped, Fichera, piston, and wrench geometries with roughly level-independent iteration counts. These give real evidence that the default configuration works. However, the central claim of robustness for curved geometries is exactly the setting where the 'large enough' tolerance and nearest-point projection in Eq. (17) intervene, and no experiment isolates their effect. Because the tolerance is a free parameter and its value is not reported, a reviewer cannot tell whether the reported convergence is stable or accidentally favorable. I therefore agree with the conditional verdict: the paper should be accepted only after a sensitivity test of this heuristic or a comparison with a geometry-aware transfer. I do not see a basis for rejection; the numerical evidence and code availability are strong, and the concern is about generalizability rather than an internal inconsistency in the presented experiments.","tokens_in":21381,"tokens_out":4444,"duration_ms":44799,"concrete_test":"On a domain with a curved boundary (e.g., the wrench geometry or a sphere), generate the same non-nested hierarchies as in §4 and solve a problem with a known manufactured solution. Measure the fine-level transfer error ||u - P^{(c,f)} u_c|| and the CG iteration count for the default tolerance, then repeat with the bounding-box tolerance varied over, say, 1e-10, 1e-6, 1e-2 relative to the coarse cell diameter, for p=2 and p=4. Compare against a reference transfer computed by evaluating the coarse field at the true point p when p lies inside the coarse domain, or by propagating the CAD geometry with OpenCASCADE as suggested in §3.3. If iteration counts and errors are insensitive to the tolerance over several orders of magnitude, the heuristic is robust; if they change, the robustness claim in §4.3 must be qualified and a principled choice of tolerance is needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central robustness claim depends on the transfer operator P^{(c,f)} being accurate enough that V-cycle preconditioning keeps iteration counts nearly independent of the number of levels. In §3.3, for Neumann boundaries on curved domains, fine-grid points p that fall outside the coarse mesh are associated to a coarse cell by enlarging local bounding boxes with a tolerance described only as 'large enough', and the coarse shape functions are evaluated at \\hat p := argmin_{\\hat x\\in\\hat K} d(\\hat x, F_K^{-1}(p)) (Eq. 17), i.e., at the nearest point in the reference cell. This substitutes a projected point for the true evaluation point and can, near corners or for larger gaps, even associate p with the wrong coarse element. No error analysis or sensitivity study is given for this projection, and Section 4.3 reports only the default behavior on the piston and wrench. If the projection introduces an O(h) or uncontrolled perturbation in the transfer, the theoretical framework of [15] (which assumes appropriate approximation properties of the intergrid operators) does not apply, and the observed robustness may not extend to other curved geometries, mesh-size ratios, or tolerance settings. The tolerance is effectively a free parameter; its value is never stated or varied. This is the most load-bearing weak point because the paper's novelty and competitiveness claims for non-trivial 3D FEA problems rest on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a matrix-free, parallel implementation of the non-nested geometric multigrid method for continuous Lagrange finite elements on independently generated and partitioned triangulations. The transfer operator is built from pointwise nodal interpolation, supported by a distributed geometric search and tensor-product evaluation at arbitrary reference points. The implementation is integrated into the deal.II multigrid framework. Numerical experiments cover Poisson and linear elasticity problems in 2D and 3D, polynomial degrees p=1 to 4, several mesh hierarchies (L-shaped, Fichera corner, piston, wrench), and comparisons with algebraic multigrid, polynomial multigrid, and the nested global-coarsening variant. The central claim is that the method is robust, in the sense that iteration counts remain nearly independent of the number of levels, and that the matrix-free realization is competitive in time to solution.","tokens_in":21656,"tokens_out":4442,"duration_ms":41412,"significance":"If the claims hold, this is a useful contribution: it provides the first matrix-free non-nested geometric multigrid implementation in a public finite-element library, with code integrated into deal.II and reproducible experiments. The paper has notable strengths: the nested sanity check reproduces global-coarsening iteration counts, the experiments cover a wider range of geometries and polynomial degrees than is common for non-nested multigrid, the performance breakdown is informative, and the strong-scaling study gives a clear picture of efficiency. The main weakness is that the treatment of curved boundaries, which is essential for the piston and wrench cases, relies on an unquantified heuristic projection in Eq. (17). Because the robustness claim is demonstrated experimentally and no theory is supplied for this projection, the curved-geometry results are not yet fully supported.","major_comments":[{"comment":"The curved-boundary transfer is the only component of the algorithm whose accuracy is not verified. The text states that local bounding boxes are expanded by a tolerance only described as 'large enough', and that points outside the coarse mesh are mapped to the nearest point in the reference cell via Eq. (17). The tolerance value is never stated or varied, and no experiment compares Eq. (17) with an alternative such as propagating CAD geometry or a projection onto the true boundary. Since the piston and wrench results of Section 4.3 are obtained with this default setting, the claim that the method is robust for curved geometries is not yet supported. Please report the tolerance used, study its sensitivity, and comment on the possible association of boundary points with the wrong coarse cell near corners or for large gaps.","section":"3.3, Eq. (17)"},{"comment":"The convergence framework of Bramble, Pasciak, and Xu [15] requires the intergrid transfer operators to satisfy appropriate approximation properties. The pointwise interpolation P^{(c,f)} combined with the projection in Eq. (17) replaces the evaluation point by a nearby point in the reference cell, and the paper provides no verification that the resulting transfer still meets the assumptions of that theory. Please either give a bound on the projection error in a suitable norm, or state explicitly that the robustness in the curved-boundary case is demonstrated only experimentally and may not be covered by the cited analysis.","section":"Section 2, [15]"}],"minor_comments":[{"comment":"The mesh-size condition is written as C h_l <= h_{l+1} <= h_l; for C > 1 this is contradictory, and the standard condition should be h_l <= C h_{l+1} (or h_{l+1} <= h_l <= C h_{l+1}). Please correct the inequality.","section":"Eq. (3)"},{"comment":"The phrase 'large enough tolerance' should be replaced with a concrete value or rule, for example a multiple of the local mesh size, so that the experiments are reproducible.","section":"3.3"},{"comment":"The caption says 'Gauss-Lobatto points for a quadrature rule of order p=4', but Section 4 states that Gauss-Legendre quadrature is used; clarify whether the figure shows support points or quadrature points.","section":"Figure 5"},{"comment":"The novelty statement 'To the best of the authors' knowledge, no matrix-free realizations...' is an assertion; consider phrasing it more cautiously or supporting it with a more systematic literature search.","section":"Section 1"},{"comment":"The sub-column layout in these tables is difficult to parse because the same header names are reused across polynomial degrees; use separate table blocks or clearer column headers to distinguish AMG, NN, and PMG entries for each p.","section":"Tables 6 and 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the numerical results are extensive. The main load-bearing issue is the unstudied curved-boundary projection in Eq. (17); if the authors provide a sensitivity study and a clearer statement of the theoretical scope, I would support acceptance. The novelty claim is plausible but difficult to verify from the text alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThis is the first matrix-free non-nested geometric multigrid implementation for continuous Lagrange elements I've seen, and it's a real piece of engineering. The authors extend deal.II's global coarsening framework with a transfer operator that evaluates coarse basis functions at arbitrary fine-grid points, using distributed geometric search and matrix-free point evaluation. The code is open source, integrated with a tutorial, and tested across Poisson and linear elasticity in 2D/3D, polynomial degrees 1-4, several independent mesh hierarchies, and two CAD-based geometries (piston, wrench). The nested sanity check reproduces global coarsening iteration counts exactly, which is a clean consistency check. The performance breakdowns are honest: non-nested transfers are roughly 10x more expensive than nested sum-factorization transfers, yet the method stays competitive with AMG and polynomial multigrid, especially for Q2.\n\nThe main soft spot is exactly what you flagged: the curved-boundary handling in Section 3.3. For Neumann boundary points that fall outside the coarse mesh, the code expands bounding boxes by a tolerance described only as 'large enough' and projects to the nearest reference point via Eq. (17). No value for the tolerance is given, and no sensitivity study is run. The paper's own tests include curved geometry with Neumann boundaries (ball, piston, wrench) and report stable iteration counts, so the heuristic is not obviously broken, but the generalization claim is conditional on that projection being accurate. A sensitivity study varying the tolerance, or at least a statement of the chosen value, would close this. The reader's second condition—a direct comparison against a matrix-based non-nested transfer baseline—is also reasonable. The paper compares against nested global coarsening and against AMG/PMG, but not against, say, the Krause-Zulian variational transfer with explicit matrices. The claimed arithmetic-intensity advantage is plausible but not directly measured.\n\nOverall, this is a solid paper worth serious refereeing. The core contribution is clear, the evidence is substantial, and the code is public. I'd ask for the tolerance sensitivity and a matrix-based baseline as revisions, but neither undermines the existence of the contribution.","headline":"A genuinely useful, open-source matrix-free non-nested multigrid implementation; the curved-boundary tolerance is an unquantified heuristic but the paper's own experiments give it empirical support.","tokens_in":22194,"tokens_out":3228,"would_cite":true,"duration_ms":26621,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65F10","65N30","65Y05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that non-nested geometric multigrid can be implemented entirely matrix-free with pointwise interpolation, yielding nearly level-independent iteration counts on Poisson and linear elasticity problems.","keywords":["non-nested multigrid","matrix-free methods","geometric multigrid","transfer operators","pointwise interpolation","finite element method","parallel computing","linear elasticity"],"falsifier":"Vary the bounding-box tolerance in the piston or wrench experiments on the same hierarchy and record conjugate-gradient iteration counts; a strong dependence on the tolerance would show that the curved-boundary projection, not the matrix-free transfer itself, is carrying the reported stability.","tokens_in":21186,"feed_emoji":"⚙️","tokens_out":8982,"duration_ms":77872,"temperature":0.7,"pith_summary":"This paper argues that geometric multigrid does not need nested meshes to be practical: each level can be an independently generated triangulation of the same domain, and the transfer between levels can be performed without ever assembling the interpolation matrix. The authors implement this for continuous Lagrange finite elements in a widely used C++ finite-element library, using pointwise interpolation as the prolongation and restriction operator and a search-and-evaluate procedure that works level by level. Across Poisson and linear elasticity problems in 2D and 3D, with polynomial degrees 1 through 4 and parallel runs, they report iteration counts that stay roughly constant as levels are added. If correct, this removes the main obstacle to using geometric multigrid on unstructured and CAD-derived meshes, where building a nested hierarchy is the hard part. The efficiency advantage over algebraic and polynomial multigrid grows with polynomial degree, while keeping the flexibility of freely chosen coarse levels.","feed_headline":"Non-nested multigrid goes matrix-free, keeping iterations flat","feed_subtitle":"Independent remeshing at each level plus pointwise interpolation keeps solver iterations nearly constant in tests","key_machinery":"The load-bearing object is the matrix-free two-level transfer operator. It combines (i) a distributed geometric search that associates every fine-grid degree-of-freedom point with an owning coarse cell and reference coordinate, using axis-aligned bounding boxes or a distributed tree search; (ii) a tensor-product evaluation routine that computes $\\sum_{i,j,k} \\varphi^{1D}_i(\\hat x_1)\\varphi^{1D}_j(\\hat x_2)\\varphi^{1D}_k(\\hat x_3)u_{ijk}$ at arbitrary points, with complexity $O((N^{1D}_{DoFs})^d)$ per point and no assembled interpolation matrix; and (iii) pointwise nodal interpolation $P^{(c,f)}_{ij}=\\varphi^{l-1}_j(x_i)$ defining prolongation, with restriction as its transpose. For curved boundaries, points outside the coarse mesh are projected to the nearest point in the reference cell according to $\\hat p := \\arg\\min_{\\hat x \\in \\hat K} d(\\hat x, F_K^{-1}(p))$; this is the heuristic step whose tolerance must be 'large enough'. The structure is generic: it extends an existing global-coarsening multigrid framework by a new two-level transfer base class, so the same smoothers, coarse solvers, and matrix-free kernels are reused.","core_discovery":"The paper's central claim is that non-nested geometric multigrid can be competitive with the best nested and algebraic alternatives if the transfer operator is evaluated matrix-free. For each pair of levels, the prolongation operator is the pointwise nodal interpolant from the coarse space $V_{l-1}$ to the fine space $V_l$: $P^{(c,f)}_{ij} = \\varphi^{l-1}_j(x_i)$, where $x_i$ are the fine-grid support points and $\\varphi_j^{l-1}$ are coarse basis functions. Restriction is the transpose of this operator. The implementation locates each fine point in an independently partitioned coarse mesh by a distributed geometric search, then evaluates the coarse field at the reference coordinates using tensor-product sum-factorization, never assembling the interpolation matrix. With this machinery, the paper reports nearly level-independent iteration counts on 2D and 3D Poisson problems and linear elasticity tests (piston, wrench) for polynomial degrees 1 through 4, and competitive or better time-to-solution than algebraic multigrid and polynomial multigrid, especially at higher orders. It is claimed to be the first matrix-free realization of multigrid for non-nested finite-element levels.","pith_inferences":["The new vertical communication efficiency metric (share of fine-grid points whose coarse owner is the same process) could be used proactively to choose partitionings that minimize transfer communication; the reported matching-partition policy raises it from a few percent to roughly 90 percent in the piston case.","A natural next experiment is to compare the nearest-point projection for curved boundaries against a CAD-aware transfer that propagates geometry information; the paper's own remarks suggest the latter is more reliable and less heuristic.","For very high polynomial degrees, the complexity of arbitrary-point evaluation, $O((N^{1D}_{DoFs})^d)$ per point, will eventually dominate; exploiting tensor-product structure of point sets within a coarse cell, as already done for nested transfers, is a concrete optimization path.","If the reliability claim extends to simplex meshes and discontinuous nodal spaces, which the implementation already supports but the paper does not analyze, the method would apply to a much wider class of unstructured finite-element workflows."],"forward_implications":["Geometric multigrid no longer requires a nested hierarchy, so coarse levels can be produced by independent remeshing or from CAD models; this removes the main obstacle for unstructured and complex-geometry meshes.","The reported iteration counts stay essentially constant as the number of levels grows, for polynomial degrees 1 through 4 and for 2D and 3D Poisson and elasticity problems, so the method behaves reliably in the tested regimes.","For higher-order elements (Q3, Q4), the matrix-free non-nested solver matches or beats algebraic multigrid and polynomial multigrid in time to solution, and the gap in favor of the non-nested method widens with polynomial degree.","The transfer cost is comparable to the smoother cost rather than negligible, so the method trades some per-iteration cost for the flexibility of arbitrary level generation; future optimizations of arbitrary-point evaluation would directly reduce total time.","Since the implementation reuses an existing multigrid framework, adding non-nested transfer capability makes the whole machinery available to applications that already use global coarsening."],"supporting_citations":[{"why":"Provides the convergence theory for multigrid algorithms with non-nested spaces and non-inherited bilinear forms, the theoretical basis for the method.","marker":"[15]"},{"why":"Supplies the distributed global-coarsening multigrid framework and matrix-free kernels that the non-nested implementation extends.","marker":"[45]"},{"why":"Describes parallel variational transfer between arbitrarily distributed unstructured meshes, the basis for the geometric search and ownership strategy.","marker":"[38]"},{"why":"Provides optimized matrix-free evaluation at arbitrarily located reference points, used by the transfer operator.","marker":"[12]"},{"why":"Supplies the distributed tree search optionally used in the coarse geometric search phase.","marker":"[44]"},{"why":"Compares prolongation operators between non-nested meshes, supporting the choice of pointwise interpolation.","marker":"[21]"},{"why":"Describes the flexible parallel adaptive geometric multigrid method whose infrastructure is reused.","marker":"[18]"}],"fun_headline_variants":["Matrix-free non-nested multigrid keeps iterations flat","First matrix-free non-nested multigrid for finite elements","Independent mesh levels via matrix-free multigrid","Matrix-free transfer operator enables non-nested multigrid","Non-nested multigrid without assembled matrices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's reliability on curved geometries rests on the heuristic that a fine-grid boundary point falling outside the coarse mesh can be safely associated with the nearest point of a nearby coarse cell, using a bounding-box tolerance described only as 'large enough'; if that projection is inaccurate, the transfer operator loses accuracy and iteration counts may stop being level-independent.","fun_headline_variants_meta":{"raw":{"variants":["Matrix-free non-nested multigrid keeps iterations flat","First matrix-free non-nested multigrid for finite elements","Independent mesh levels via matrix-free multigrid","Matrix-free transfer operator enables non-nested multigrid","Non-nested multigrid without assembled matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001429,"raw_usage":{"total_tokens":5758,"prompt_tokens":930,"completion_tokens":4828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":4750}},"tokens_in":546,"tokens_out":4828,"duration_ms":29961,"temperature":1.0,"reasoning_tokens":4750,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:28:54.292076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Vary the bounding-box tolerance in the piston or wrench experiments on the same hierarchy and record conjugate-gradient iteration counts; a strong dependence on the tolerance would show that the curved-boundary projection, not the matrix-free transfer itself, is carrying the reported stability.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the convergence theory for multigrid algorithms with non-nested spaces and non-inherited bilinear forms, the theoretical basis for the method."},{"cited_title":"Munch, T","cited_arxiv_id":null,"evidence_quote":"Supplies the distributed global-coarsening multigrid framework and matrix-free kernels that the non-nested implementation extends."},{"cited_title":"Krause and P","cited_arxiv_id":null,"evidence_quote":"Describes parallel variational transfer between arbitrarily distributed unstructured meshes, the basis for the geometric search and ownership strategy."},{"cited_title":"Lebrun-Grandi ´e, A","cited_arxiv_id":null,"evidence_quote":"Supplies the distributed tree search optionally used in the coarse geometric search phase."},{"cited_title":"Dickopf and R","cited_arxiv_id":null,"evidence_quote":"Compares prolongation operators between non-nested meshes, supporting the choice of pointwise interpolation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the flexible parallel adaptive geometric multigrid method whose infrastructure is reused."}],"review_version":1}