{"id":"02c1b680-96b9-46fe-b2e3-ba7898224d8c","arxiv_id":"2412.06347","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A multigrid extension of low-mode averaging keeps the number of Dirac low modes fixed while suppressing stochastic variance on increasingly large lattices.","lead":"The paper presents a new variance-reduction method for lattice QCD, using block-projected low quark modes in a multigrid hierarchy to make translation averaging of quark propagators nearly volume-independent. If it holds, it could make precise hadronic vacuum polarization calculations for the muon g-2 anomaly much cheaper at large volumes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Volume-independence claim rests on local-coherence deficits measured on one configuration; no evidence that Eq. (19) stays uniformly small across ensembles and for the exact Nc=50, b=8 setting used in Fig. 7.","rationale":"The reader's weakest_assumption correctly identifies the local-coherence extrapolation as the load-bearing point: the algebraic decomposition is exact, but the physical statement that a fixed number of modes suffices at fixed block size independently of volume relies on Eq. (19) being small uniformly. The paper's own Fig. 1 is limited to one configuration and to Nc=20/100, not the Nc=50 used in the scaling plot, and no error bars are given for the variance estimates in Fig. 7. My stress-test therefore agrees with the conditional verdict: the method is promising and internally consistent, but the headline volume-independence claim is broader than the currently demonstrated data. I do not see a separate algebraic or technical flaw that would require rejection; the concern is empirical support, and it is best addressed by the deficit and error-bar checks described above.","tokens_in":22620,"tokens_out":5216,"duration_ms":63448,"concrete_test":"Using the stored gauge fields (or a random subset of at least 20 configurations per ensemble), compute the exact local-coherence deficits epsilon_c = ||(1-P) phi_c|| of Eq. (19) for the actual setup of Fig. 7: Nc=50, block size 8^4, on E7, F7, G7 and H7, for c = 51..500, and report the ensemble mean and maximum. In addition, recompute the L0 variances of Fig. 7 with jackknife or bootstrap errors. If the maximum deficit at c ~ 350 stays below roughly 0.2 on all ensembles and the G7 L0 variance is more than 2 sigma below the F7 value, the volume-independence claim is supported; if deficits grow with volume or the variance trend is not statistically significant, the claim should be weakened to a demonstration at the three investigated volumes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that, at fixed block size and fixed Nc, the fine-level variance decreases with volume because block-projected low modes efficiently span the low-mode space (Sec. 3, Eq. (19)). The numerical support for this mechanism is Fig. 1, which shows deficits on a single thermalized F7 configuration for Nc=20 and 100 and for two block sizes. The scaling result in Fig. 7, however, uses Nc=50 with block size 8^4 on E7, F7 and G7, and no deficit data are presented for that parameter choice, for the other ensembles, or across the ensemble. If local coherence degrades with volume or fluctuates between configurations, the remainder S0 will retain a larger low-mode component and the observed decrease of the L0 variance with L could fail at larger volumes or with different statistics. Compounding this, the variance points in Fig. 7 are shown without error bars; with 100 configurations, the relative statistical error on a variance estimate is of order 14%, which may be comparable to the differences between the three volumes. The method's algebraic decomposition is exact, so the concern is not correctness of the identity but the breadth of the empirical support for the volume-independence statement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes multigrid low-mode averaging (MG LMA), a hierarchical variance-reduction scheme for quark-line connected correlation functions. The method decomposes the quark propagator into a telescoping sum over block-decomposed low-mode subspaces, Eq. (29), and evaluates the different levels with tailored stochastic estimators, with the coarsest level often evaluated exactly. The authors test the method on the isovector vector current correlator using N_f=2 O(a)-improved Wilson fermions on ensembles with L approximately 2.1, 3.2, and 4.2 fm. The central numerical claim is that, unlike ordinary low-mode averaging, the fine-level variance of MG LMA decreases as the physical volume grows when the block size and the number of low modes are held fixed, so that a constant variance reduction can be maintained with O(10-100) low modes.","tokens_in":22890,"tokens_out":4682,"duration_ms":48539,"significance":"The proposed algebraic decomposition is exact, and the numerical results are encouraging: if the volume-independence claim survives scrutiny, the method directly addresses a well-known bottleneck in large-volume low-mode averaging for observables such as the hadronic vacuum polarization. The paper is honest about its limitations, explicitly stating that the gauge variance is poorly determined and that the coarse-grid solver is suboptimal, and it avoids circularity by comparing variances against the gauge variance rather than against the method's own outputs. The main weaknesses are statistical: the central scaling plot has no error bars, and the local-coherence mechanism is demonstrated on a single configuration and with parameter sets that do not exactly match the scaling test. These issues are fixable with additional analysis rather than being fundamental flaws in the derivation.","major_comments":[{"comment":"The central claim that the fine-level variance decreases with volume is supported by three points at L approximately 2.1, 3.2, and 4.2 fm, but the variance estimates are plotted without error bars. With N=100 configurations, the relative statistical error on a variance estimate is of order 14%, which is comparable to the differences between the three volumes shown in the right panel of Fig. 7. Please add jackknife or bootstrap uncertainties to the variance estimates in Fig. 7 (and to the corresponding points in Figs. 5 and 6), and state whether the observed decrease is statistically significant.","section":"Sec. 5, Fig. 7"},{"comment":"The local-coherence mechanism underlying the volume-independence claim is tested in Fig. 1 on a single thermalized configuration of F7, for Nc=20 and 100 and for two block sizes, while the scaling test in Fig. 7 uses Nc=50 and a block size of 8^4. No deficit data are shown for that parameter set, for the other ensembles, or across the gauge ensemble. Please either provide the deficit Eq. (19) for the exact parameter set used in Fig. 7 and show its ensemble spread, or argue explicitly why the one-configuration test is sufficient to establish the uniformity needed for the volume-scaling conclusion.","section":"Sec. 3, Fig. 1"},{"comment":"The cost comparison is expressed in terms of the number of stochastic sources needed to 'reach the gauge noise,' but the text concedes that the gauge variance is 'fairly poorly determined' and that the quoted costs should be taken as indicative only. Because the measured costs in Tab. 3 (e.g., 557.8 versus 80.7 for G7) depend directly on that threshold, the uncertainty in sigma_G should be propagated into the quoted costs, or the cost claims should be reformulated as ranges rather than single numbers.","section":"Sec. 5.1 and Tab. 3"}],"minor_comments":[{"comment":"The symbol Nl in Eq. (34) should be N_ell for consistency with the rest of the text.","section":"Eq. (34)"},{"comment":"The H7 entry '192 x 96 3' is ambiguous; it should read 192 x 96^3 (or a similar explicit notation for the spatial extent).","section":"Table 1"},{"comment":"The legend entries '4x4x4x4' and '48x8x8x8' do not match the text's statement that only spatial block sizes of b/a=4 and 8 are varied; please clarify the block geometry used in each curve.","section":"Fig. 1"},{"comment":"The performance model would benefit from a short justification of why mem(K) can be neglected in the asymptotic limit N_rhs -> infinity; Eq. (50) relies on this drop-out but the text does not state the assumption explicitly.","section":"App. B, Eq. (47)"},{"comment":"In the sentence 'unpreconditioned BiCGSTAB 3 solve,' the superscript 3 appears to be an artifact; please remove it.","section":"App. E"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a useful and potentially impactful variance-reduction method for lattice QCD. The algebraic core is sound and the numerical test is relevant. My main concern is evidentiary: the volume-independence claim needs error bars and a broader demonstration of local coherence across the ensemble. I see no novelty or disclosure issues, and the paper fits the journal's scope. With the statistical support added, I would expect it to be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a good methods paper that deserves a serious referee. The new thing is the multigrid low-mode averaging decomposition applied to a quark-line connected correlator, with a hierarchical telescoping sum that pushes most of the stochastic variance onto cheap coarse-grid inversions. The algebraic decomposition in Eq. (29) is exact, and the paper is honest about its numerical limitations. The main caveat is that the headline volume-independence claim is backed by thinner evidence than the text implies.\n\nWhat's genuinely new: prior block-projected LMA work focused on single-propagator traces and disconnected loops. Here they apply block-projected low modes to the connected isovector vector correlator, and the numerical results in Figs. 5-7 show the fine-level variance decreasing as L goes from 2.1 to 4.2 fm at fixed Nc=50. That is a real step for translation averaging in the long-distance window relevant to HVP. The paper also does useful collateral work: the chiral-preserving coarsening discussion (App. C) and the spectra/condition number study (App. D) explain a practical detail that would otherwise be a black box, and the performance model separates algorithmic potential from their admittedly suboptimal coarse solver.\n\nThe soft spots are where the stress-test lands. The local coherence mechanism is verified in Fig. 1 on a single configuration of F7, for Nc=20 and 100 and two block sizes, but the scaling claim in Fig. 7 uses Nc=50, b=8 on E7/F7/G7, and no deficit data are shown for that setting or across the ensembles. The variance estimates are shown without error bars; with 100 configurations, a variance has ~14% relative error, which is comparable to the level differences between volumes in Fig. 7. The gauge variance is admittedly 'fairly poorly determined' (Sec. 5.1), and H7 has only 5 configurations. None of this is fatal — the decomposition is exact and the trend is consistent — but it means the volume-independence statement is a well-motivated conjecture from limited data, not a demonstrated scaling law.\n\nFor a methods paper this is a solid contribution. I would bring it to a reading group and would cite it if working on large-volume correlation functions. For peer review: send it, but require the authors to either add error bars, show local coherence deficits for the actual parameter settings, or soften the scaling claim.","headline":"A well-executed methods paper extending LMA to connected correlators; the volume-independence claim is promising but backed by thinner numerics than the text suggests.","tokens_in":23436,"tokens_out":2604,"would_cite":true,"duration_ms":23898,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["12.38.Gc","11.15.Ha"],"model":"deepseek-v4-flash","headline":"By projecting a small fixed set of low quark modes onto local spacetime blocks, multigrid low-mode averaging keeps stochastic variance under control as the lattice volume grows, without requiring more low modes.","keywords":["lattice QCD","low-mode averaging","multigrid","variance reduction","local coherence","hadronic vacuum polarization","Dirac operator","translation averaging"],"falsifier":"Extend the two-level multigrid LMA measurement of Fig. 7 to $L\\approx 6$-$8$ fm with the same $N_c=50$ and block size $8^4$, averaging the fine-level variance over many configurations: if the fine-level contribution stops decreasing with $L$, or the coarse-level inversion count grows so that total cost no longer stays flat, the central claim fails. A cheaper check is to measure the deficits of Eq. (19) on an ensemble average at larger volume and look for a systematic rise.","tokens_in":22402,"feed_emoji":"📉","tokens_out":8564,"duration_ms":81690,"temperature":0.7,"pith_summary":"This paper develops a variance-reduction method for lattice QCD that keeps the required number of Dirac-operator low modes fixed as the physical volume grows. Ordinary low-mode averaging splits the quark propagator into an exactly computed low-mode part and a stochastically estimated remainder, but the remainder becomes noisier at larger volumes because the density of low modes grows. The new scheme additionally projects the low modes onto local spacetime blocks, creating a much larger effective subspace, and splits the propagator into a telescoping multigrid sum. In tests on the isovector vector current correlator with $N_f=2$ $\\mathrm{O}(a)$-improved Wilson fermions on lattices from 2.1 to 4.2 fm, the variance of the fine-level piece decreases as the volume increases when the block size is fixed, whereas standard low-mode averaging loses its benefit. If correct, this makes full translation averaging affordable for large-volume, high-precision hadronic observables without generating hundreds or thousands of low modes.","feed_headline":"QCD noise falls as lattice grows, using fixed low modes","feed_subtitle":"Block-projected low modes make full translation averaging affordable on large lattices.","key_machinery":"The key object is the block-projected low-mode subspace: each of the $N_c$ low eigenmodes of the Hermitian Dirac operator $Q=\\gamma_5 D$ is restricted to each block of a lattice decomposition and re-orthonormalized, so that $N_c N_s V_1$ fields span a space approximating the whole low-mode band, exploiting local coherence (small deficits in Eq. (19)). Restriction and prolongation operators between nested grids define coarse-grid operators $Q_{l+1}=R_l Q_l T_l$; iterating the identity $Q_l^{-1}=\\{Q_l^{-1}-T_l Q_{l+1}^{-1}R_l\\}+T_l Q_{l+1}^{-1}R_l$ gives a telescoping sum for the quark propagator, $S=S_0+\\cdots+S_{N_\\ell-1}$, whose levels can be estimated with different numbers of stochastic sources. Preserving chiral degrees of freedom on the coarse grids ($N_s=2$) keeps the coarse operators well conditioned and makes the lowest $N_c$ eigenvalues match the fine-grid ones.","core_discovery":"The paper's central claim is that local coherence of the low quark modes converts low-mode averaging from a method whose mode count must grow with volume into one whose mode count can stay fixed. Concretely, with $N_c=50$ exact low modes block-projected onto cubes of side roughly $0.5\\,\\mathrm{fm}$, the variance of the fine-level contribution to the translation-averaged isovector vector correlator at $t\\simeq 1.3\\,\\mathrm{fm}$ falls as $L$ goes from 2.1 to 4.2 fm, while ordinary LMA's fine-level variance rises until it equals the plain stochastic estimator. The coarser levels carry most of the stochastic variance, but their Dirac operators act on much smaller spaces, so many stochastic sources there are cheap; the coarsest level can be evaluated exactly. The result is an estimator that reaches the gauge-noise floor with one stochastic source on the fine grid and a modest number on the coarse grids, at a cost in fine-grid inversion units that is orders of magnitude below plain stochastic sampling on the larger volumes.","pith_inferences":["Beyond the tested range, the same local-coherence argument suggests the variance suppression persists at $L\\gtrsim 6$ fm, but only a measurement with error bars on the variance estimates at those volumes would confirm it.","Because the coarse subspace dimension grows with the lattice volume while its operators stay cheap, the scheme should pair naturally with master-field style analysis on very large lattices; this connection is not explored in the paper.","A direct extension to baryonic correlators (three quark propagators) is plausible since low modes dominate at large separations, but the cross-term variance structure is untested.","Replacing exact low modes with inexact ones in the coarse operators could remove most of the mode-generation overhead, since the construction does not require exact eigenvectors; the paper notes this possibility but does not test it."],"forward_implications":["A fixed set of order 10-100 low modes suffices for constant variance reduction as the lattice volume grows, so the cost and storage of low-mode generation no longer scale with volume.","Each level of the multigrid split can be estimated independently: one stochastic source on the fine grid, more on coarser grids, and an exact evaluation on the coarsest level, so the total cost is set by the small coarse-grid operators.","The variance reduction applies to quark-line connected diagrams at large separations, directly targeting the isovector hadronic vacuum polarization contribution to the muon $g-2$ and baryonic correlators.","The measured and modelled costs in fine-grid inversion units improve by orders of magnitude over plain stochastic estimators on the larger volumes, with further gains expected from a multiple right-hand-side coarse solver.","Retaining chiral spin structure on the coarse grid is required for well-conditioned coarse operators; without it, spurious low eigenvalues appear."],"supporting_citations":[{"why":"Supplies the local-coherence property of low modes that the method exploits.","marker":"[24]"},{"why":"Introduces adaptive multigrid for the Wilson-Dirac operator, providing the coarse-grid construction inherited here.","marker":"[25]"},{"why":"Establishes that low eigenmodes dominate long-distance quark propagation and defines the standard low-mode averaging baseline.","marker":"[13–15]"},{"why":"Gives the mode density that scales with volume, explaining why ordinary LMA requires more modes as volume grows.","marker":"[16]"},{"why":"Documents that high-precision long-distance HVP needs roughly 1000 low modes in practice, the problem this work addresses.","marker":"[19]"},{"why":"Provides the stochastic one-end trick estimator used for the level contributions and the variance definitions.","marker":"[9, 10]"},{"why":"Shows that block-projected low modes accelerate and compress low-mode computation, supporting the practical overhead assumptions.","marker":"[34]"},{"why":"Presents earlier multilevel or block-decomposed variance reductions for single-propagator traces, which this work extends to connected correlators.","marker":"[35–37]"},{"why":"Supplies the ensembles and the O(a)-improved Wilson action parameters used in the numerical tests.","marker":"[43, 44]"}],"fun_headline_variants":["Multigrid trick keeps QCD noise flat as lattice grows","Fixed low modes tame QCD noise on big lattices","Volume-independent noise with multigrid low-mode averaging","Coarse grids let low-mode averaging scale to big volumes","Local coherence keeps mode count fixed as QCD lattices enlarge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is local coherence: a fixed small set of block-projected low modes spans almost the entire low-mode space (small deficits in Eq. (19)) uniformly across gauge configurations and volumes; if that uniformity fails as the volume or the gauge field changes, the volume-independence of the variance suppression collapses.","fun_headline_variants_meta":{"raw":{"variants":["Multigrid trick keeps QCD noise flat as lattice grows","Fixed low modes tame QCD noise on big lattices","Volume-independent noise with multigrid low-mode averaging","Coarse grids let low-mode averaging scale to big volumes","Local coherence keeps mode count fixed as QCD lattices enlarge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3077,"prompt_tokens":987,"completion_tokens":2090,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2008}},"tokens_in":603,"tokens_out":2090,"duration_ms":13897,"temperature":1.0,"reasoning_tokens":2008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:45:24.688083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Extend the two-level multigrid LMA measurement of Fig. 7 to $L\\approx 6$-$8$ fm with the same $N_c=50$ and block size $8^4$, averaging the fine-level variance over many configurations: if the fine-level contribution stops decreasing with $L$, or the coarse-level inversion count grows so that total cost no longer stays flat, the central claim fails. A cheaper check is to measure the deficits of Eq. (19) on an ensemble average at larger volume and look for a systematic rise.","supporting_citations":[],"review_version":1}