{"id":"8f37dac0-b6f7-4086-9185-1688f6f98d11","arxiv_id":"2607.03413","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Cell-wise unassembled multigrid algebraically equals classical local multigrid, so hanging-node machinery vanishes and a structured GPU kernel runs unchanged on adaptive meshes.","lead":"A multigrid solver for high-order finite elements never builds the usual global vector and never assembles hanging-node constraints, yet matches classical adaptive multigrid iterate-by-iterate. On a GPU it hits about 1.1 GDoF/s per V-cycle with a simple Jacobi smoother, competitive with heavier patch smoothers.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged dependence on companion DSS correctness and Assumption 1.","rationale":"The paper's strongest claim is architectural/algebraic rather than spectral: once residuals stay unassembled, hanging-node constraints and edge-block splitting become unnecessary because the geometric transfers plus residual masking already reproduce the classical operators. The proofs (Lemmas 1–5, Theorem 2, Corollary 1) are explicit, use only the characterizing property (20) of C_HN, and close the induction cleanly for the Jacobi smoother that is actually implemented. Numerical results (grid-independent rates, flat iteration counts under adaptive depth, throughput parity with a published patch smoother at tighter tolerance) are consistent with the theory for the Laplace operator. The reader's MODERATE/CONDITIONAL assessment already correctly isolates the two external dependencies (companion DSS correctness and the need for any substituted smoother to satisfy Assumption 1). No stronger internal inconsistency or unstated assumption that would move the verdict was found; the recommended concrete check simply reconfirms the key dual identity on a minimal hanging-node configuration.","tokens_in":28637,"tokens_out":553,"duration_ms":5224,"concrete_test":"Independently re-derive the dual identity of Lemma 4 (Eq. 21) for a single hanging-node interface in 2-D Q2: assemble the classical residual, apply C_HN^T then R_CG, and compare entry-wise against G_{ℓ-1}^T R_cell applied to the unassembled residual; any mismatch larger than machine epsilon falsifies the claim that plain tensor-product restriction on unassembled data reproduces the constrained transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central algebraic claim (Theorem 2) is that the cell-wise V-cycle is iterate-identical to classical local multigrid of Janssen–Kanschat once Assumption 1 (smoother equivalence) holds. Lemmas 1–4 establish the transfer intertwining for unassembled residuals and geometric prolongation without forming hanging-node matrices; Lemma 5 verifies Assumption 1 for the masked damped Jacobi of (26). The induction in Theorem 2 is therefore tight for the smoother studied. The only remaining load-bearing external condition is that the companion DSS operator S of [37] truly realizes the algebraic form (28) (and its averaging variant (30)) on the adaptive level meshes after edge masking. That is already the reader's weakest_assumption; no additional internal gap (e.g., a missing case in the residual-transfer identity (21) or a failure of continuity of the prolonged field across E_ℓ) appears in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper constructs a geometric multigrid V-cycle for high-order continuous FEM that never forms assembled global vectors, working entirely in the redundant cell-wise storage of the companion framework [37]. The central claim is architectural and algebraic: hanging-node constraints are never assembled; plain tensor-product transfers applied to the unassembled residual reproduce the classical constrained restriction (including the transposed constraint matrix); and local-smoothing edge operators reduce to pointwise residual masking. Equivalence, iterate by iterate, to the classical local multigrid method of Janssen and Kanschat [18] is proved by induction (Theorem 2) from intertwining lemmas for prolongation, unassembled restriction, residual transfer under property (20) of CHN, and a smoother Assumption 1 verified for masked damped Jacobi. As a consequence the structured, topologically oblivious DSS kernel can run unchanged on adaptive level meshes. Numerical experiments for the Laplacian (cube and curved ball; uniform and adaptive) show grid-independent contraction essentially unaffected by hanging nodes, with up to 1.1 GDoF/s per V-cycle and end-to-end solve throughput on par with a recent patch-smoother code on a single A100 in fp64.","tokens_in":28830,"tokens_out":1426,"duration_ms":53768,"significance":"If the equivalence and the adaptive DSS argument hold, the paper removes a genuine implementation bottleneck for matrix-free adaptive multigrid on GPUs: constraint matrices, edge/interior operator splits, and topology-dependent residual weights. The contribution is cleanly scoped as architectural rather than spectral (p-robustness is not claimed; Jacobi is the studied smoother), and the inheritance of [18] via exact operator identities is the right way to transfer convergence theory. Strengths include explicit intertwining lemmas, an induction proof of Theorem 2, verification of Assumption 1 for the Jacobi construction (Lemma 5), and adaptive experiments that keep iteration counts flat in refinement depth. The observation that the same edge mask required by local smoothing also licenses a topologically oblivious structured DSS kernel is a useful conceptual point. Competitive throughput with only a masked point-Jacobi smoother, against a patch-smoother baseline driven to a looser tolerance, makes the engineering claim concrete.","major_comments":[{"comment":"§4.3.1 and Algorithm 9: The argument that incorrect structured-DSS sums on hanging interfaces are harmless is written for residual-side DSS (Algorithm 6): dual mask zeros E_ℓ before DSS, and the correction on E_ℓ is claimed to vanish. The performance-critical realization is the deferred-assembly sweep (Algorithm 9 / identity (32)), which applies averaging DSS S̄ to the updated iterate. If the cascade writes corrupted values onto E_ℓ when exchanging toward an empty pocket (Figure 2: “stale or zero”), S̄ can alter edge DoFs that local smoothing must leave equal to the prolonged coarse field. Please spell out why Algorithm 9 preserves E_ℓ values under the same incorrect cascade—e.g., explicit re-mask after S̄, pocket contents, or a proof that S̄δ vanishes on E_ℓ whenever δ does—so that the central “topologically oblivious kernel on adaptive meshes” claim covers the implemented smoother, not","section":null},{"comment":"§3.2.3 Assumption 1 and §4: Theorem 2 imports the full convergence theory of [18] only through smoother intertwining. Lemma 5 verifies this for the abstract masked Jacobi (26) assuming the algebraic DSS form (28). The manuscript should state as a precise hypothesis what is taken from the companion [37] (that the implemented structured/unstructured cascade realizes (28) and the averaging form (30) on interior interfaces of each level mesh after edge masking) versus what is proved here. Without that boundary, the load-bearing external condition for both equivalence and the adaptive performance claim is harder to audit than the transfer lemmas.","section":null}],"minor_comments":[{"comment":"§6 / Table 1 vs [11]: The comparison is useful but should state more prominently in the table caption or main text that [11] stops at residual reduction 10^{-9} while this work uses energy-norm reduction 10^{-14}, so the near-parity already favors the present solver on tolerance; the current discussion in the prose is easy to miss.","section":null},{"comment":"§6: The fixed under-relaxation (ω=0.7 cube, 0.6 ball) and the exclusion of p=1 because that single choice is “mildly too large” should be flagged earlier when the smoother is introduced (§3.3), not only in the experiments section.","section":null},{"comment":"References [22] and [23] appear to be duplicate entries of the same Kronbichler–Sashko–Munch paper; please deduplicate.","section":null},{"comment":"Figure 4: The caption correctly notes that the energy measure vanishes at iteration zero for u=0; consider starting the solid curves at cycle 1 in the plot itself to avoid a visual gap that readers may misread as missing data.","section":null},{"comment":"Notation: V_cell vs V_{cell,ℓ} vs V^{act}_{cell,ℓ} and the several inclusion maps (ι_ℓ, ι_{Sℓ}, ι^*†_ℓ) are introduced densely in §3.2; a short notation table or a one-paragraph recap before Algorithm 4 would help readers who skip the companion paper.","section":null},{"comment":"Author line: “Micha l Wichrowski” appears to have a spurious space ( Michał ).","section":null},{"comment":"§5 “Detection of the refinement edge”: the one-time DSS-based edge detection is elegant; a sentence on cost relative to a single V-cycle would reassure readers that setup is negligible for the reported solve throughputs.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is tightly coupled to the companion cell-wise/DSS manuscript [37]; if that companion is not already accepted or co-submitted, the editor may want both in view, because the adaptive “incorrect sums are harmless” claim and the GPU kernels live at the interface of the two. I do not see an internal algebraic gap in the transfer theory (Lemmas 1–4, Theorem 2) beyond the Algorithm 9 clarification requested above; empirical adaptive iteration counts being flat is strong supporting evidence that the implementation is doing the right thing on E_ℓ. Scope (Laplace, Jacobi, single GPU) is appropriate for an architectural methods paper and should not be grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real news is architectural. Once you keep residuals unassembled and cell-local, the hanging-node machinery that usually clogs adaptive multigrid simply disappears: plain tensor-product transfers on the raw residual reproduce the classical constrained restriction (including the transposed constraint action), and local-smoothing edge operators collapse to a pointwise residual mask. Theorem 2 then shows the whole cell-wise V-cycle is iterate-identical to Janssen–Kanschat, so the classical theory transfers. That identity is new relative to the deal.II / Kronbichler line, and the paper proves it carefully with intertwining lemmas rather than hand-waving.\n\nWhat works well: the math is explicit (Lemmas 1–5, induction for Theorem 2), the edge-mask argument that licenses a topologically oblivious structured DSS kernel is clean, and the A100 experiments for Laplace are coherent—grid-independent rates essentially unchanged under adaptive refinement, ~1.1 GDoF/s per cycle at p=3, end-to-end throughput on par with a recent patch smoother while driving a tighter tolerance. The free parameter is just a fixed under-relaxation ω; no invented entities.\n\nSoft spots, in proportion: the whole story rests on the companion cell-wise DSS framework [37] actually realizing the algebraic form of S (and its averaging variant) after masking, and on Assumption 1 holding for any smoother you swap in. Both are already flagged by the reader and the stress-test; neither is an internal contradiction. Only Laplace is tested, no code ships, and Jacobi is not p-robust (the paper says so). Those are real but secondary limits for an architecture paper.\n\nThis is for people who build high-order matrix-free GPU multigrid and who care about adaptive meshes without constraint infrastructure. The citation pattern is honest. I would send it to peer review; a serious referee can check the companion dependence and the numerical claims. Worth engaging.","headline":"Solid architectural result: cell-wise unassembled residuals + edge masks make hanging-node multigrid algebraically identical to Janssen–Kanschat without ever building constraints, with clean proofs and competitive A100 numbers.","tokens_in":29432,"tokens_out":489,"would_cite":true,"duration_ms":5089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N55","65N30","65F08","65Y10","65Y20"],"pacs":[],"model":"grok-4.5","headline":"A geometric multigrid V-cycle that never assembles vectors or hanging-node constraints is algebraically identical to classical local multigrid.","keywords":["geometric multigrid","adaptive mesh refinement","hanging nodes","matrix-free","finite elements","GPU","cell-wise storage","local smoothing"],"falsifier":"On a sequence of adaptively refined meshes with hanging nodes, run both the cell-wise V-cycle and a classical local multigrid code with the same smoother and check whether the iteration matrices (or successive residual vectors) differ by more than round-off; any systematic discrepancy falsifies the claimed equivalence.","tokens_in":29518,"feed_emoji":"⚡","tokens_out":708,"duration_ms":5471,"temperature":0.7,"pith_summary":"This paper shows that if every field in a high-order finite-element solver is kept in a redundant, cell-local representation, the usual obstacles of adaptive multigrid simply disappear. Hanging-node constraints need never be built: plain tensor-product transfers applied to the raw, unassembled residual already reproduce the classical constrained restriction, including the action of the transposed constraint matrix. Edge operators of local smoothing reduce to a pointwise mask on the residual, so the level operator never has to be split into interior and edge blocks. The single inter-cell primitive that remains is one structured direct-stiffness summation inside the smoother; that kernel can stay topologically oblivious even on arbitrarily refined meshes. The author proves that the resulting cell-wise V-cycle produces exactly the same iterates as the classical local multigrid method of Janssen and Kanschat, and therefore inherits its level-independent convergence theory. Numerical tests for the Laplace equation confirm that convergence is essentially unaffected by hanging nodes, and that a masked point-Jacobi smoother already delivers GPU throughput competitive with more elaborate patch smoothers.","feed_headline":"Multigrid that never builds hanging-node constraints","feed_subtitle":"Cell-wise storage makes adaptive V-cycles algebraically identical to classical local multigrid, with GPU throughput to match.","key_machinery":"Three intertwining identities (prolongation commuting with gather, unassembled residual restriction reproducing the transposed constraint matrix, and smoother equivalence under masking) that, by induction on levels, make the cell-wise V-cycle identical to the classical local multigrid V-cycle.","core_discovery":"When fields live permanently in redundant cell-wise storage, hanging-node constraints and edge-operator splitting are unnecessary: plain geometric transfers on the unassembled residual, together with a pointwise residual mask, already realize the classical constrained local multigrid V-cycle. The cell-wise algorithm is therefore algebraically identical, iterate by iterate, to the classical method and inherits its convergence theory without any new spectral analysis.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Cell-wise multigrid never assembles hanging-node constraints","Unassembled residuals make adaptive multigrid constraint-free","Plain transfers on cell storage match classical local multigrid","Pointwise residual masks replace edge splitting in V-cycles","Cell-wise V-cycle algebraically identical to constrained multigrid"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The cell-wise smoother, when fed a continuous iterate and an unassembled residual, must produce a continuous correction that matches exactly what the classical local smoother would produce on the assembled system; if that matching fails, the inheritance of classical convergence theory collapses.","fun_headline_variants_meta":{"raw":{"variants":["Cell-wise multigrid never assembles hanging-node constraints","Unassembled residuals make adaptive multigrid constraint-free","Plain transfers on cell storage match classical local multigrid","Pointwise residual masks replace edge splitting in V-cycles","Cell-wise V-cycle algebraically identical to constrained multigrid"]},"model":"grok-4.5","effort":"low","cost_usd":0.00329,"raw_usage":{"total_tokens":1153,"prompt_tokens":811,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":32900000,"prompt_tokens_details":{"text_tokens":811,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":258,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":811,"tokens_out":84,"duration_ms":9538,"temperature":1.0,"reasoning_tokens":258,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T02:39:18.674865+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a sequence of adaptively refined meshes with hanging nodes, run both the cell-wise V-cycle and a classical local multigrid code with the same smoother and check whether the iteration matrices (or successive residual vectors) differ by more than round-off; any systematic discrepancy falsifies the claimed equivalence.","supporting_citations":[],"review_version":1}