{"id":"ee9e0cbd-aee0-473b-b45d-6109dfc2cfd0","arxiv_id":"2505.01719","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new stochastic estimator, blending distillation low modes with reweighted random high modes, gives unbiased all-to-all fermion propagators and enables sub-percent nucleon axial charge and 4-point pion form factor calculations on 40 lattice configurations.","lead":"This paper introduces a \"blending\" algorithm for lattice QCD that estimates the all-to-all quark propagator by combining low-energy spatial modes with random high-frequency modes, making multi-point correlation functions and disconnected diagrams cheaper. The authors demonstrate it by computing nucleon axial charges and the pion form factor from 40 configurations, with results consistent with much larger calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unbiasedness is proven only for a single quark bilinear; two insertions sharing a time slice reuse the same random modes, so the claimed arbitrary-N-point unbiasedness is unsupported and fails in a toy D=2, Nst=1 check.","rationale":"The reader's conditional verdict is well-founded, but the most load-bearing issue is not the gA excited-state model; it is the gap between the proved single-bilinear unbiasedness and the claimed arbitrary-N-point unbiasedness. Supplement A.1 proves Eq. (12) for one quark bilinear and invokes independence across time slices for multi-current insertions. For two insertions at the same time slice, the same random vectors are used in both estimators, so E[XY] differs from E[X]E[Y]; the required higher moments are not provided. The paper itself flags this by saying the proof for more non-trivial cases can be obtained similarly, without giving it. This gap matters because the abstract and summary promise arbitrary-point correlation functions, and the projected propagator S_ij is a single object reused in all contractions. The demonstrated applications avoid the issue (currents at different times, single insertions), so the main physics results are not invalidated, but the advertised generality is conditional on either a proof of same-time product unbiasedness or an explicit restriction to insertions at distinct time slices (or independent noise sets per insertion). The gA excited-state fit concern raised by the reader is real and should also be addressed, but it is an application-level systematic rather than a flaw in the blending construction itself. For these reasons the verdict remains CONDITIONAL, with the condition sharpened to include the missing same-time product proof.","tokens_in":25748,"tokens_out":21390,"duration_ms":231618,"concrete_test":"Perform the analytic toy check that isolates the shared-noise correlation: take a two-dimensional space (D=2), one random mode (Nst=1), no low modes, and define Ĵ(x)=ω0|η(x)|^2 with ω0=D/Nst=2 for each spatial point x. Verify that E[Ĵ(x)]=1 but E[Ĵ(x)Ĵ(y)] for x≠y equals 2/3 instead of 1 when the same η is used for both insertions. If confirmed, the blending construction is not unbiased for products of operators that share a time slice; then test the natural fix of drawing independent noise vectors for each insertion and check that the product expectation becomes 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Supplement A.1 proves Eq. (12): E[Σ_{i,j} Ω_{ij}|φ_i><φ_i|⊗|φ_j><φ_j|] = I⊗I, establishing unbiasedness for one quark-bilinear kernel. The paper then states \"The proof for more non-trivial cases can be obtained using the similar procedure,\" but that extension is not supplied and is not automatic. For a four-point function with two currents at the same time slice, the estimator is a product of two such sums using the same Nst random vectors at that time slice. The expectation does not factorize; it requires fourth moments involving four random vectors (or two pairs), which Eq. (21) does not provide. The paper's multi-current argument relies on independence across time slices (\"since the probabilities are independent on the different time slices\"), and this fails when t1=t2. A toy example isolates the problem: for D=2, Nst=1, no low modes, the single-current estimator Ĵ(x)=ω0|η(x)|^2 is unbiased with ω0=2, but E[Ĵ(x)Ĵ(y)] = 4·E[|η(x)|^2|η(y)|^2] = 4/6 = 2/3, not 1. Thus the advertised \"arbitrary-point correlation functions\" claim is not established for same-time insertions; the demonstrated applications (3pt, and 4pt with t1≠t2) are unaffected, but the central claim's generality is conditional on a missing proof or on using independent noise per insertion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the 'blending' method for estimating all-to-all fermion propagators in lattice QCD. At each time slice, the spatial Dirac-color-site space L is split into a low-mode subspace L1 spanned by Ne eigenvectors of the Laplace operator and its complement L2, and the identity on L is represented in expectation by the Ne exact low modes plus Nst random vectors sampled from L2, with reweighting factors ωn = ([L2]−n)/(Nst−n). Propagators projected onto this blending space compress the all-to-all propagator from (4NcNTN_L^3)^2 to (4NT(Ne+Nst))^2 degrees of freedom, and quark-bilinear operators are reweighted to be unbiased in expectation; the supplement proves this single-bilinear unbiasedness (Eq. 12) using the Haar-random moments of the noise vectors. Numerical anchors include: the pion conserved vector charge R_V^π = 1.0002(29) on C24P29; pion form factors from 3-point and 4-point functions that agree (charge radii 0.332(29) and 0.329(33) fm^2); RI/MOM Z_q from the blended propagators matching the point-source result; and, on F48P30 (a = 0.077 fm, mπ ≈ 300 MeV, 40 configurations), a joint 3-state fit over eight Ne values giving g_A^{u−d} = 1.2339(43) and flavor-decomposed g_A^u = 0.895(15), g_A^d = −0.338(15), g_A^s = −0.0245(72), g_A^{u+d+s} = 0.533(28). The authors argue the method makes disconnected diagrams, excited-state control with distillation-like smearing, and N-point functions (QED/isospin corrections, hadronic tensors) practical.","tokens_in":26162,"tokens_out":24474,"duration_ms":212434,"significance":"Assuming the central claim survives, this is a significant methodological contribution: it restores the high-frequency spatial modes that distillation discards, in an unbiased stochastic way, while retaining distillation's excited-state control and multi-operator flexibility; if it generalizes as claimed, it makes disconnected and multi-current correlators with excited-state systematics tractable on large volumes. The strengths are concrete: the unbiasedness proof for a single bilinear is an explicit moment construction (Supplemental Eqs. 12–23), not a fitted result; the numerical anchors are independent and quantitative (conserved vector charge 1.0002(29); RI/MOM Z_q matching point-source propagators in Supp. Fig. 9; 3pt/4pt form-factor agreement); and g_A^{u−d} = 1.2339(43) agrees with high-statistics external determinations (RQCD 1.238(24), CalLAT 1.235(11)), which is nontrivial evidence that the estimator is not tuned to its own constants. The scaling claims (statistical advantage ∝ N_L^{−3/2}, noise saturation at Nst ~ 100–400) are supported by the tests shown.","major_comments":[{"comment":"The claim that the method enables 'arbitrary-point correlation functions' is not established for multiple quark-bilinear insertions that share a time slice. The proof in Supplemental Sec. A.1 establishes only the two-leg unbiasedness of Eq. (12) for a single kernel, and the fourth-moment factorization in Eq. (25) is valid only for independent time slices. For two insertions at the same time slice, the estimator is a product of two sums over the same Nst random vectors, and its expectation requires contracting the same-time fourth moments of Eq. (21) with the product of two Ω^(2) weight tensors, a computation the paper does not perform; the sentence 'The proof for more non-trivial cases can be obtained using the similar procedure' is therefore not justified. A minimal check isolates the problem: for Ne = 0, D = 2, Nst = 1, the single-current estimator Ĵ(x) = ω0|η(x)|^2 with ω0 = 2 is unbiased (E[Ĵ] = 1), but E[Ĵ(x)Ĵ(y)] = 4E[|η(x)|^2|η(y)|^2] = 4/(D(D+1)) = 2/3 ≠ 1 for x ≠ y. The demonstrated applications are not affected — the 4pt pion form factor uses t2 − t1 ≥ 2a, and the RI/MOM Z_q check in Fig. 9 validates the three-leg weight Ω^(3) empirically — so the required repair is either a proof for coincident insertions (e.g., a four-leg analogue of Eq. (12) with a weight tensor that is not the product of two Ω^(2) factors) or an explicit restriction of the claim to insertions on distinct time slices.","section":"Supplemental Sec. A.1, Eqs. (12), (21), (25); Abstract/Introduction/Summary"},{"comment":"The headline results (g_A^{u−d} = 1.2339(43), g_A^u = 0.895(15), g_A^d = −0.338(15), g_A^s = −0.0245(72), g_A^{u+d+s} = 0.533(28)) are quoted with jackknife statistical errors only. The 3-state ansatz of Eqs. (49)–(50) represents the entire excited-state contamination by two effective states with Δ1 ≈ 0.49 GeV and Δ2 ≈ 1.4 GeV, and the fit parameter c11 = −0.08(53) is effectively unconstrained; no variation of the number of states, fit ranges, or the set of Ne values is reported to bound the resulting systematic uncertainty. Moreover, the C24P29 ensemble at a = 0.105 fm yields g_A^{u−d} = 1.1690(72), about 5% lower than the F48P30 result, an offset the paper attributes to 'discretization errors and finite volume effects' without quantifying it. For claims such as 'sub-percent determinations' and 'first principle high accuracy' in the title, Introduction, and Abstract, the paper needs a systematic error budget (fit-ansatz variation, renormalization-scale uncertainty from Z_s,MS_A/Z_V = 1.115(15), and an explicit statement that the result is a single-lattice-spacing, fixed-pion-mass demonstration), or the claims must be scoped accordingly.","section":"Sec. C.2, Eqs. (49)–(50), Table IV"}],"minor_comments":[{"comment":"The excited-state term appears as 'ci0di(e^{−Δi(tf−t)} + e^{−Δi t}) ciid^2_i e^{−Δi tf}', which is missing the '+' before the cii term; compare with the correct form in Supplemental Eq. (49).","section":"Main text Eq. (8)"},{"comment":"The definition 'ωn = ([L2]−n)/)/(Nst−n)' contains a stray '/)' and should read (([L2]−n)/(Nst−n)); the superscripts on Ω (Ω^(1), Ω^(2)) are also used before being defined.","section":"Main text Eq. (1)"},{"comment":"The LaTeX '{\\color{black} ...}' markup around the g_A values should be removed before submission.","section":"Abstract"},{"comment":"The conditions defining Ω^(3) are incomplete: 'i≠j≠k>Ne' does not specify whether the indices are pairwise distinct, and 'all the permutation of i,j,k' does not enumerate the weights for mixed cases such as i=j>Ne with k≤Ne; additionally, 'Ω(2)ij/V' in Eq. (36) contains a dangling '/V'.","section":"Supplemental Eq. (37) and Eq. (36)"},{"comment":"The text refers to 'Ne = 80 (left lower panel)', but the panels are labeled Ne = 40, 55, 70, 100; the text should match the figure labels.","section":"Fig. 13 and Sec. C.2"},{"comment":"The two-state fit at the 'optimal' Ne value is reported to give exactly g_A^{u−d} = 1.2339(43), the same value as the three-state joint fit; please clarify whether this is an independent fit and how its uncertainty was obtained.","section":"Sec. C.2"},{"comment":"The generic reference to 'Supplemental materials [33]' in the main text should be replaced by pointers to the specific supplement sections (A.1, A.4, B, C) and equation numbers.","section":"Main text references"},{"comment":"The sentence 'the statistical uncertainty of ⟨V4^cc⟩π scales approximately as Nst^{−1}, as the expectation value of RVπ is exactly one' conflates the mean and the variance of the ratio; please rephrase (the ratio's expectation value is exactly one, and its variance is what scales as Nst^{−2} for the quoted Nst^{−1} uncertainty).","section":"Supplemental Sec. A.5"},{"comment":"The floor notation '⌊L2⌋' is used for the dimension of L2, which is already an exact integer [L2] = NcN_L^3 − Ne; please use [L2] consistently.","section":"Supplemental Eqs. (10)–(13)"}],"recommendation":"major_revision","confidential_remarks":"This is a methods paper with demonstration results at a single lattice spacing and pion mass; the title's 'first principle high accuracy' and the abstract's 'sub-percent determinations' overstate what the manuscript currently supports, and I would recommend the editor require the authors to temper those claims or supply a systematic error budget. The main risk is the same-time multi-insertion gap, which is a real technical hole in the advertised 'arbitrary N-point' capability, but the empirical anchors (conserved charge, Z_q, 3pt/4pt agreement, external g_A consistency) indicate the estimator is fundamentally sound and the missing four-leg proof is likely obtainable; I therefore view this as a fixable major revision rather than a rejection. The citation practice and the comparison with FLAG, PDG, RQCD, and CalLAT results are appropriate, and the paper does not conceal the large C24P29 versus F48P30 g_A difference, which is reported in the text. I would also ask the editor to encourage a clean-up of the supplement's notation (Ω^(1)/Ω^(2)/Ω^(3) and the [L2]/⌊L2⌋ mix), since these small inconsistencies currently make the proof harder to check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe blending estimator is a genuinely new construction. It combines exact low-lying Laplace eigenmodes with reweighted random vectors in the orthogonal complement, giving an unbiased estimator of quark bilinears while keeping the smearing benefits of distillation. That construction is not in the cited distillation, low-mode substitution, or field sparsening papers. The numerical checks are strong: the conserved vector charge comes out 1.0002(29), the renormalization constant Zq matches point-source propagators, and the 3pt and 4pt pion form factors agree. For a 40-configuration calculation, the gA values are consistent with high-statistics results, which is a good sanity check.\n\nThe soft spot is the 'arbitrary N-point' claim. Supplement A.1 proves unbiasedness for one quark bilinear. The extension to multiple insertions is asserted without proof. For insertions on distinct time slices, the independence argument works, and the demonstrated 3pt and 4pt applications are fine. But for two currents on the same time slice, the same random vectors are reused, the expectation does not factorize, and the claim fails: a toy D=2, Nst=1 example gives E[J(x)J(y)] = 2/3, not 1. So the title's 'arbitrary-point' overreaches. The paper should either prove the multi-insertion case, restrict the claims to time-separated insertions, or use independent noise vectors for insertions sharing a time slice.\n\nThe other soft spots are more conventional: no systematic error budget for gA, a three-state fit whose second effective excited state at ~1.42 GeV is doing a lot of work, and per-Ne fit range choices in Table IV that are not justified. No code or data are released. These are fixable in revision.\n\nWho this is for: lattice QCD practitioners working on disconnected diagrams, N-point functions, and excited-state control. It deserves peer review. I would send it, with a request to fix the generality claim and add systematics. If those are addressed, this could be a widely used method.","headline":"Clever, useful blend of distillation and stochastic high modes that checks out for the demonstrated cases, but the arbitrary-N-point claim outruns the proof.","tokens_in":26688,"tokens_out":4580,"would_cite":true,"duration_ms":47401,"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","02.70.Uu"],"model":"deepseek-v4-flash","headline":"The paper claims that a \"blending\" of a few low-frequency spatial modes with random high-frequency modes gives an unbiased, low-cost estimator of the all-to-all quark propagator, and demonstrates sub-percent nucleon axial charges and a…","keywords":["all-to-all fermion propagator","lattice QCD","blending method","distillation","stochastic estimation","nucleon axial charge","N-point correlation functions","disconnected diagrams"],"falsifier":"Set $N_{st}$ equal to the full complement dimension $[L_2]$ on a small lattice, compute the pion conserved-charge ratio of Eq. (5) with the blending estimator, and compare it with the exact point-source value at every time slice; exact agreement would confirm Eq. (12), while any $t$-dependent deviation would falsify the unbiasedness claim.","tokens_in":25537,"feed_emoji":"⚛️","tokens_out":12453,"duration_ms":118018,"temperature":0.7,"pith_summary":"Lattice QCD computations usually approximate the all-to-all quark propagator by expensive sequential or stochastic sources; this paper claims a way to estimate it from a small \"blending space\" made of low-frequency spatial modes plus random high-frequency modes at each time slice. The paper proves that a reweighted sum over these modes reproduces the exact operator in expectation, so any quark-bilinear current, including disconnected loops and four-point functions, can be evaluated from one set of projected propagators. Because the low modes are treated exactly, hadron interpolating fields retain the excited-state suppression of distillation, while the random modes make local current insertions unbiased. As demonstrations, the method yields nucleon axial charges $g_A^{u-d}=1.2339(43)$, $g_A^{u+d+s}=0.533(28)$ at $m_\\pi\\simeq300$ MeV with 40 configurations, and a pion charge radius from four-point functions consistent with the three-point result. If the claim holds, high-precision strong-interaction quantities, including QED and isospin-breaking corrections, become accessible from first principles.","feed_headline":"Blending trick makes all-to-all quark propagators unbiased and cheap","feed_subtitle":"Low modes plus random high modes reproduce exact correlators; nucleon axial charge and pion form factor follow.","key_machinery":"The load-bearing object is the blending space: the span of $N_e$ low-lying eigenvectors of the spatial Laplace operator (the distillation subspace) plus $N_{st}$ orthonormal random vectors sampled uniformly from its orthogonal complement. The identity that carries the argument is Eq. (12), $\\mathbb{E}\\left[\\sum_{i,j=1}^{N_e+N_{st}}\\Omega_{ij}\\,\\lvert\\phi_i\\rangle\\langle\\phi_i\\rvert\\otimes\\lvert\\phi_j\\rangle\\langle\\phi_j\\rvert\\right]=\\hat{I}\\otimes\\hat{I}$, with the reweighting factors $\\omega_n=([L_2]-n)/(N_{st}-n)$ and $\\Omega_{ij}^{(2)}$ given in Eq. (3). The proof fixes the second and fourth moments of the random vectors by unitary invariance, which is exactly what converts the projected operator into an unbiased estimator while keeping the low-mode part exact. This mechanism is what lets hadron states be constructed by distillation and local currents be inserted without bias.","core_discovery":"The central discovery is that the all-to-all fermion propagator does not need to be stored or inverted in full. The paper constructs a blending space from $N_e$ low-lying eigenvectors of the spatial Laplacian and $N_{st}$ random orthonormal vectors in the orthogonal complement, and assigns a reweighting tensor $\\Omega_{ij}$ so that the expectation of the projected identity is $\\hat{I}\\otimes\\hat{I}$. This identity means that every quark-bilinear operator $O=\\bar q M_O q$ evaluated in the blending space has expectation equal to the exact operator, with the low-mode sector exact and the high-mode sector unbiased. Consequently the all-to-all propagator is compressed from $(4N_cN_TN_L^3)^2$ to $(4N_T(N_e+N_{st}))^2$ degrees of freedom. Numerical checks include the pion conserved charge, the pion form factor from three- and four-point functions, and sub-percent nucleon axial charges from a joint $N_e$-dependent fit.","pith_inferences":["If the unbiasedness identity holds at finite $N_{st}$ with the quoted scaling, the same blending-space propagators can be reused for many operators and momenta, so the marginal cost of a new $N$-point observable should become a contraction cost rather than new fermion inversions, an economy the paper only partially exploits.","The effective second excited state at $\\Delta_2\\simeq1.42$ GeV is likely a mixture of single-particle and multiparticle states; a GEVP analysis with $\\pi N$ and $\\pi\\pi N$ interpolators on the same blending-space propagators would show whether the three-state model has hidden state-dependent bias in $g_A$.","The paper notes a difference between its coarse and fine ensemble results for $g_A^{u-d}$; a dedicated continuum-extrapolation study would determine how much of that gap is discretization error versus finite-volume effects."],"forward_implications":["Any quark-bilinear current, including disconnected quark loops and four-point currents, can be evaluated from the same projected propagators instead of new sequential inversions.","The compression of the all-to-all propagator from $(4N_cN_TN_L^3)^2$ to $(4N_T(N_e+N_{st}))^2$ degrees of freedom makes large spatial volumes and small lattice spacings comparatively cheaper than in standard all-to-all approaches.","Varying $N_e$ changes the smearing of the external hadron, so excited-state contamination can be mapped out systematically; the joint three-state fit gives $g_A^{u-d}=1.2339(43)$ with 40 configurations at $m_\\pi\\simeq300$ MeV.","The consistency of three- and four-point pion form factors demonstrates that $N$-point correlation functions with two current insertions are practical, opening a route to QED corrections and hadronic-tensor calculations."],"supporting_citations":[{"why":"This reference supplies the distillation low-mode projection that the blending space extends to all-to-all propagation.","marker":"[23]"},{"why":"This reference provides the C24P29 ensemble used for the pion conserved-charge and form-factor demonstrations.","marker":"[32]"},{"why":"This reference contains the mathematical proof of Eq. (12) and the numerical scaling checks for the method.","marker":"[33]"},{"why":"This reference provides the F48P30 ensemble used for the nucleon axial-charge and disconnected-loop calculations.","marker":"[34]"},{"why":"This reference supplies the singlet axial renormalization factor used to convert the flavor-singlet $g_A$ to the $\\overline{\\mathrm{MS}}$ scheme at 2 GeV.","marker":"[2]"},{"why":"This reference provides the summed-ratio fitting approach and the comparable $g_A$ benchmark used in the excited-state analysis.","marker":"[7]"},{"why":"This reference gives the z-expansion parameterization used to extract the pion charge radius from both three- and four-point form-factor data.","marker":"[35]"},{"why":"This reference supplies the high-precision pion charge radius with which the blending-based results are compared.","marker":"[36]"},{"why":"This reference defines the RI/MOM scheme used to test the blending estimator on off-shell Green's functions.","marker":"[56]"}],"fun_headline_variants":["All-to-all quark propagator compressed via blending low and random modes","Unbiased all-to-all fermion propagator from low plus stochastic high modes","Blending eigenmodes makes arbitrary-point QCD correlators cheap and exact","Low modes exact, high modes random: unbiased all-to-all propagator","Compress all-to-all propagator: low eigenvectors plus random vectors for QCD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The $g_A$ extraction assumes that two effective excited states, one near the nucleon-pion threshold and one near 1.4 GeV, whose masses and $N_e$-dependent weights are fitted to the same data, capture all excited-state contamination in the chosen fit ranges.","fun_headline_variants_meta":{"raw":{"variants":["All-to-all quark propagator compressed via blending low and random modes","Unbiased all-to-all fermion propagator from low plus stochastic high modes","Blending eigenmodes makes arbitrary-point QCD correlators cheap and exact","Low modes exact, high modes random: unbiased all-to-all propagator","Compress all-to-all propagator: low eigenvectors plus random vectors for QCD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1856,"prompt_tokens":1013,"completion_tokens":843,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":744}},"tokens_in":629,"tokens_out":843,"duration_ms":7368,"temperature":1.0,"reasoning_tokens":744,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:13:14.859877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set $N_{st}$ equal to the full complement dimension $[L_2]$ on a small lattice, compute the pion conserved-charge ratio of Eq. (5) with the blending estimator, and compare it with the exact point-source value at every time slice; exact agreement would confirm Eq. (12), while any $t$-dependent deviation would falsify the unbiasedness claim.","supporting_citations":[{"cited_title":"Investigation of $\\pi N$ contributions to nucleon matrix elements","cited_arxiv_id":"2412.07263","evidence_quote":"This reference provides the C24P29 ensemble used for the pion conserved-charge and form-factor demonstrations."},{"cited_title":"blending space","cited_arxiv_id":null,"evidence_quote":"This reference supplies the singlet axial renormalization factor used to convert the flavor-singlet $g_A$ to the $\\overline{\\mathrm{MS}}$ scheme at 2 GeV."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference gives the z-expansion parameterization used to extract the pion charge radius from both three- and four-point form-factor data."}],"review_version":1}