{"id":"c05d8267-5ca3-48d3-a8a0-05661cdc098b","arxiv_id":"2412.04444","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Block Lanczos is shown to generalize the GEVP method for lattice QCD correlator matrices, providing faster convergence, two-sided error bounds, and improved signal-to-noise.","lead":"This paper introduces a block Lanczos algorithm for analyzing matrices of correlation functions in lattice QCD, extending a previous scalar Lanczos framework to multiple interpolating operators. It offers a principled way to extract hadron energies and matrix elements with faster convergence and two-sided error bounds, potentially replacing the standard GEVP method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ZCW spurious-state filter can discard physical states with small overlaps; its reliability is demonstrated only on one low-statistics example whose two interpolators couple similarly to all low-lying states.","rationale":"The paper's strongest claim—that block Lanczos is a direct extension and superset of GEVP with qualitative and quantitative advantages—rests on the noisy-data pipeline. The GEVP equivalence (Sec. VI A) is mathematically proven, the noiseless mock-data tests validate the recursion algebra, and the residual bounds are rigorous under the stated Hermitian T assumption. However, the practical advantages in real LQCD data are obtained only after spurious-state filtering, and the ZCW filter's adaptive threshold (Eq. 117) is the least secure element. The paper itself acknowledges that physical states below the cut are not extracted, which means the method is not a strict superset of GEVP in the sense of guaranteed variational bounds. The single numerical demonstration uses two interpolators with similar smearing, so it does not exercise the regime where physical states can have small overlaps with all interpolators—precisely the multi-particle scenario motivating correlator-matrix methods. The proposed test directly probes this false-negative risk and would determine whether the conditional verdict should stand or whether the central claim needs to be weakened to a statement about states with sufficiently large overlaps.","tokens_in":64306,"tokens_out":5737,"duration_ms":58056,"concrete_test":"Generate a synthetic noisy correlator matrix from the noiseless mock data of Eq. (72) plus an additional physical state with total normalized overlap |Z|^2 = 10^-4 (below the would-be ε_ZCW from Eq. (117) with F_ZCW = 10) at an energy between E1 and E2, and add Gaussian noise at the signal-to-noise level of the 80-configuration nucleon example. Run the full block Lanczos pipeline (Hermitian subspace + ZCW filtering, F_ZCW = 10) on 200 bootstrap replicates. If the added state is systematically absent from the non-spurious spectrum while its true energy lies within the resolvable range, the filter's false-negative rate is unacceptably high and the central claim of superseding GEVP requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V C 3 defines the ZCW cutoff as ε_ZCW = (1/F_ZCW) min_{k≤m_H} ΔZCW(m_H)_k (Eq. 117), with F_ZCW = 10. The paper explicitly concedes that 'physical states below the cut will not be extracted.' This is load-bearing for the claim that block Lanczos is a direct extension and superset of GEVP: GEVP variational bounds guarantee upper bounds on the lowest r eigenvalues irrespective of overlap, whereas the block Lanczos pipeline can silently drop a true state if its total overlap with the interpolator set is below the threshold. The demonstration uses two nucleon interpolators that differ only in smearing radius, so both have substantial overlap with the same low-lying states; the realistic multi-particle regime, where interpolators have nearly disjoint support, is not tested. On the single 80-configuration dataset, m_H and ε_ZCW are also estimated from noisy data, and the robustness of F_ZCW is only checked over [2,20] on that dataset.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the Lanczos-based analysis framework for lattice QCD correlation functions from scalar correlators to correlator matrices. It constructs an oblique block Lanczos algorithm, derives recursion relations for block-tridiagonal matrix elements, Ritz values/vectors, overlap factors, and operator matrix elements, and develops a physically motivated reformulation of the Cullum-Willoughby test (the ZCW test) for filtering spurious states. The method is tested on noiseless mock data and on a 2x2 nucleon correlator matrix from 80 lattice configurations, with comparisons to GEVP and scalar Lanczos. The central formal claim is that GEVP is the one-step limit of block Lanczos, making block Lanczos a direct extension and superset of GEVP methods with faster convergence, two-sided residual bounds, and asymptotically constant signal-to-noise.","tokens_in":64672,"tokens_out":3672,"duration_ms":100483,"significance":"If the claims hold, this is a substantial methodological contribution: it provides a parameter-free deterministic analysis pipeline for correlator matrices, with explicit two-sided error bounds and no fits, and it unifies several existing approaches (GEVP, Prony-type methods, scalar Lanczos) in a single framework. The algebraic derivations in Sections II and IV and the noiseless demonstrations are careful and internally consistent, and the explicit construction of residual bounds and the KPS bound generalization are valuable. The main weakness is that the noisy-data claims, especially the advantage over GEVP for excited states, rest on a single 80-configuration example with acknowledged state-identification ambiguities, and on a ZCW spurious-state filter whose overlap-dependence can in principle discard physical states.","major_comments":[{"comment":"The ZCW cutoff defined in Eq. (117) is set by the minimum overlap factor over early iterations, and the text immediately following explicitly concedes that \"physical states below the cut will not be extracted.\" This is load-bearing for the claim, repeated in the abstract and in Sec. VI A, that block Lanczos is a direct extension and superset of GEVP methods. GEVP provides rigorous variational upper bounds for the lowest r eigenvalues irrespective of the overlap of the interpolators with those states, whereas the block Lanczos pipeline can silently drop a true state whose total overlap with all interpolating operators falls below epsilon_ZCW. The paper's only noisy demonstration uses two nucleon interpolators that differ only in the quark smearing radius and therefore couple similarly to the same low-lying states; the realistic multi-particle regime with nearly disjoint interpolator support is not tested. The robustness check over F_ZCW in [2,20] is also performed only on this same dataset. I request either a demonstration with qualitatively different interpolators, or a clear statement that ZCW filtering is a practical heuristic whose guaranteed-state property is weaker than that of GEVP.","section":"Sec. V C 3, Eq. (117)"},{"comment":"The noisy-data energy extraction depends on a filter-and-sort state identification scheme, which the authors themselves describe as having \"some ambiguities\" and \"obvious misidentifications.\" The final block Lanczos result E1 = 0.792(30) is in 1.4-sigma agreement with the high-statistics scalar Lanczos value 0.735(25), while the GEVP fit result is 0.879(25), a 4-sigma difference. With a single 80-configuration dataset and acknowledged state-misidentification issues, the claim of a qualitative improvement in excited-state resolution is not yet convincingly established. A higher-statistics block-Lanczos analysis, or a more robust state-identification protocol, would be needed to support the strength of the present conclusions.","section":"Sec. V D and Tables I-II"},{"comment":"The advertised advantages of block Lanczos for matrix elements are demonstrated only on noiseless mock data; no noisy three-point correlator application is presented. The matrix-element estimator in Eq. (58) inherits the same state-identification and ZCW-filtering dependence as the spectrum, so the abstract's claim that matrix elements are \"straightforward[ly]\" extracted should be qualified as applying to the infinite-statistics or noiseless limit until a noisy three-point demonstration is provided.","section":"Sec. V A and Sec. VI B 4"}],"minor_comments":[{"comment":"The name \"Lanczos\" is misspelled as \"Lanzcos\" in several places, including the caption of Fig. 2 and Sec. V B; please correct this globally.","section":"Throughout"},{"comment":"The word \"arithmetic\" is misspelled as \"arithematic\" in the note about high-precision eigensolves; please fix this typo.","section":"Sec. V F"},{"comment":"The residual-bound identity in Eq. (111) is stated to have been \"numerically verified\" for oblique Lanczos, but no details of the verification are given; since this identity motivates the CW-ZCW equivalence, please state the range of m and the datasets for which it was checked.","section":"Sec. V C 1, Eq. (111)"},{"comment":"In Eq. (55) the right-hand side is proportional to a product of sums over a, with the no-summation-over-a convention stated only in prose; please make this explicit in the notation, e.g., by adding a parenthetical or using a dummy index distinct from the free a.","section":"Eq. (55)"}],"recommendation":"major_revision","confidential_remarks":"The algebraic core of the paper is sound and the noiseless demonstrations are convincing, so this is a potentially important methods paper for hep-lat. My main reservation is that the advertised \"superset of GEVP\" claim is stronger than the evidence: the ZCW filter's overlap-dependence means the block Lanczos pipeline does not inherit GEVP's variational guarantee for the lowest r eigenvalues, and the single noisy dataset with similar interpolators does not test the regime where this distinction matters. I would advise the editor to ask for either a targeted demonstration with multi-particle-like interpolators or a careful softening of the superset claim. The relationship to Refs. [73,74] and to GEVM/PGEVM is addressed in App. G, but the presentation in Sec. VII and the outlook could be interpreted as overstating novelty; this should be checked during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper actually delivers on its main claims. The oblique block Lanczos recursion for correlator matrices is new, the proof that GEVP is a single iteration of block Lanczos is clean, and the reformulation of the Cullum-Willoughby test as an overlap-factor cut (ZCW) gives the method a physical interpretation that extends naturally to the block case. The noiseless examples are convincing: block Lanczos converges faster than GEVP, extracts relative phases of overlaps, and the residual bounds give two-sided uncertainties instead of GEVP's one-sided upper bounds. The paper also includes useful extras: the block KPS bound, a proof that common interpolators produce real symmetric correlators, and an explicit disproof of the claim that block Lanczos is equivalent to GEVM/PGEVM.\n\nThe soft spots are real but not fatal. The ZCW threshold is heuristic and data-dependent (Eq. 117); the paper explicitly concedes that physical states below the cut will not be extracted. That is a genuine limitation for multi-particle spectroscopy, where interpolators can have nearly disjoint overlap with different states. The noisy-data validation is thin: one 80-configuration dataset with two proton interpolators that differ only in smearing radius. The authors admit this is not the most important use case. The claim 'no fits or statistical inference are required' is overstated—m_H and epsilon_ZCW are estimated from the noisy data, and state identification still requires sorting and a bootstrap median. The E1 result agrees with high-statistics scalar Lanczos at 1.4 sigma, but GEVP fits give a higher value; that tension is left unresolved.\n\nOverall, the core mathematics is sound and the method is a direct generalization of GEVP, so the central claim holds up. The paper deserves serious peer review. I would ask the authors for a more realistic noisy demonstration (e.g., N, Npi, Npi-pi interpolators) and a sensitivity study of the ZCW cut to different spectral densities and overlap patterns.","headline":"A solid, genuinely new method—block Lanczos for correlator matrices—that cleanly contains GEVP as a one-step limit, though its noisy-data validation rests on a thin example and the ZCW filter can silently drop physical states.","tokens_in":65001,"tokens_out":2887,"would_cite":true,"duration_ms":30130,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F15","15A18"],"pacs":["12.38.Gc","11.15.Ha","02.70.-c"],"model":"deepseek-v4-flash","headline":"This paper establishes that the generalized eigenvalue problem (GEVP) spectroscopy method is exactly one iteration of block Lanczos, so running more iterations yields strictly better lattice QCD spectra, two-sided bounds, and matrix…","keywords":["block Lanczos","generalized eigenvalue problem","lattice QCD spectroscopy","spurious eigenvalue filtering","Cullum-Willoughby test","correlation-function matrices","two-sided error bounds","matrix elements"],"falsifier":"Apply block Lanczos to a lattice correlator matrix built from interpolators that deliberately exclude a known low-lying state, such as a Roper-like level with a small overlap on all operators, and give the data enough statistics; if the $F_{\\mathrm{ZCW}}=10$ filter does not return that level, or if the residual-bound windows for the surviving neighboring levels exclude its known energy, the claim that noisy-data spectra remain reliable for small-overlap physical states fails.","tokens_in":64137,"feed_emoji":"⚛️","tokens_out":8310,"duration_ms":74575,"temperature":0.7,"pith_summary":"The paper tries to replace the standard variational analysis of lattice QCD correlation-function matrices with a block Lanczos algorithm. Its central claim is that GEVP, the usual method for extracting energy levels from several interpolating operators, is only the first step of block Lanczos; additional steps explore a larger Krylov space and therefore converge faster to ground and excited states. A reformulated Cullum–Willoughby filter labels states with near-zero overlap with all interpolators as spurious noise, and computable residual bounds turn one-sided variational bounds into two-sided error bars. If correct, block Lanczos gives lattice QCD spectroscopy and matrix elements without fits, with constant signal-to-noise.","feed_headline":"Block Lanczos supersedes GEVP for lattice QCD spectra","feed_subtitle":"One GEVP solve is a single Lanczos step; further steps converge faster and give two-sided error bounds.","key_machinery":"The carrying mechanism is the oblique block Lanczos recursion applied to the $r\\times r$ correlator matrix $C_{ab}(t)=\\langle\\chi_a|T^t|\\psi_b\\rangle$. At each of $m$ iterations the algorithm generates new left and right Lanczos vectors and the block-tridiagonal matrix $T^{(m)}$; diagonalizing $T^{(m)}$ gives Ritz values $\\lambda_k^{(m)}$ that estimate transfer-matrix eigenvalues, with energies $E_k^{(m)}=-\\ln\\lambda_k^{(m)}$. Spurious states from noise are filtered by the block ZCW test, which cuts on the total overlap sum $\\sum_a \\omega_{1ak}^{(m)}(\\omega^{-1})^{(m)}_{k1a}$, with threshold $\\varepsilon_{\\mathrm{ZCW}}$ fixed by the minimum physical overlap at the last clean iteration and the hyperparameter $F_{\\mathrm{ZCW}}=10$. Residual bounds computed from Ritz-vector residuals give two-sided intervals for the distance between each Ritz value and the closest true eigenvalue.","core_discovery":"On the paper's own terms, the discovery is that GEVP is the one-step limit of block Lanczos. Concretely, with a reindexed correlation function $\\tilde C(t)=C(t_0+t_d t)$, the GEVP eigenvalues $\\lambda_k(t_d,t_0)$ equal the Ritz values of a single block Lanczos step, $\\lambda_k(t_d,t_0)=\\tilde\\lambda_k^{(1)}$, and the GEVP eigenvectors are related to one-step Lanczos eigenvectors by an explicit change of basis. Because block Lanczos continues orthogonalizing against more Lanczos vectors, its Ritz values converge exponentially faster, at a rate governed by $e^{-2t\\sqrt{E_r-E_0}}$ rather than $e^{-t(E_r-E_0)}$. The method also provides two-sided residual bounds on every Ritz value, overlap factors with relative signs, and current matrix elements from three-point functions by multiplication with Ritz coefficients, with no fits or statistical inference beyond bootstrap resampling. Proof-of-principle tests on noiseless mock data and a two-by-two proton correlator matrix show the promised faster convergence, improved state resolution, and constant signal-to-noise.","pith_inferences":["If the one-step coincidence is as general as stated, block Lanczos should also replace GEVP in any non-lattice setting where a matrix time series $C(t)$ comes from an underlying transfer matrix, such as nuclear structure problems, with the same recursions and bounds applied with no QCD-specific input.","The ZCW filter's overlap cut implies a practical diagnostic: when the number of surviving states plateaus below expectation, the operator set is likely missing a physical state nearly orthogonal to all interpolators, and adding a distinct operator should be the first response.","Because block Lanczos reconstructs the correlator exactly for all $t \\le 2m-1$, the difference between reconstructed and raw correlators at fixed $m$ could serve as a fit-free diagnostic of residual noise artifacts."],"forward_implications":["Any analysis currently using GEVP eigenvalues can be rerun with extra block Lanczos iterations on the same correlator matrices, and the GEVP answers appear as the $m=1$ output.","Energies of ground and excited states carry rigorous two-sided error bars from residual bounds, replacing one-sided variational upper bounds.","Overlap factors and external-current matrix elements are available from the same run by contracting three-point functions with Ritz coefficients, including relative signs between overlaps.","When the gap to the first state outside the block is small, the expected speedup is exponential: block Lanczos converges like $e^{-2t\\sqrt{E_r-E_0}}$ whereas GEVP converges like $e^{-t(E_r-E_0)}$."],"supporting_citations":[{"why":"Introduces the scalar Lanczos method for correlators, including Cullum–Willoughby spurious-state filtering and residual bounds, which this paper extends to matrices.","marker":"[1, 2]"},{"why":"Defines GEVP variational methods for correlator matrices; the paper shows these are one-step block Lanczos.","marker":"[3–8]"},{"why":"Provides GEVP convergence rates, effective-energy estimators, and the moving-pivot scheme used in numerical comparisons.","marker":"[5, 6]"},{"why":"Establishes the interlacing and variational bounds for GEVP that block Lanczos two-sided residual bounds generalize.","marker":"[7]"},{"why":"Supplies the block Lanczos Kaniel–Paige–Saad convergence bound adapted and derived in Sec. IV.","marker":"[21]"},{"why":"Standard reference for Lanczos eigenvalue convergence and residual bounds used to derive the block residual bound.","marker":"[22]"},{"why":"Original Cullum–Willoughby test, reformulated here as an overlap-based ZCW cut and generalized to block Lanczos.","marker":"[44, 45]"},{"why":"Eigenvalue–eigenvector identity that connects the Cullum–Willoughby and ZCW test quantities.","marker":"[57]"},{"why":"Higher-statistics nucleon-correlator dataset and reference fit results used as comparison values in the noisy block Lanczos demonstration.","marker":"[37, 46]"}],"fun_headline_variants":["GEVP is one step of block Lanczos","Block Lanczos extends GEVP with error bounds","Faster QCD spectra via block Lanczos","GEVP as Lanczos step one: faster convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The filter that removes spurious states assumes that every physical state in the chosen quantum-number sector has an overlap with at least one interpolating operator that is not pathologically small; the threshold is set with a single hyperparameter $F_{\\mathrm{ZCW}}=10$, and states below it are treated as noise even though they could be genuine but weakly coupled states.","fun_headline_variants_meta":{"raw":{"variants":["GEVP is one step of block Lanczos","Block Lanczos extends GEVP with error bounds","Faster QCD spectra via block Lanczos","GEVP as Lanczos step one: faster convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2947,"prompt_tokens":1003,"completion_tokens":1944,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":1884}},"tokens_in":619,"tokens_out":1944,"duration_ms":16324,"temperature":1.0,"reasoning_tokens":1884,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:23:24.653781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply block Lanczos to a lattice correlator matrix built from interpolators that deliberately exclude a known low-lying state, such as a Roper-like level with a small overlap on all operators, and give the data enough statistics; if the $F_{\\mathrm{ZCW}}=10$ filter does not return that level, or if the residual-bound windows for the surviving neighboring levels exclude its known energy, the claim that noisy-data spectra remain reliable for small-overlap physical states fails.","supporting_citations":[{"cited_title":"Edwards, R","cited_arxiv_id":null,"evidence_quote":"Eigenvalue–eigenvector identity that connects the Cullum–Willoughby and ZCW test quantities."}],"review_version":1}