{"id":"da889cd9-8db3-4d74-b57b-052e96a9e2f4","arxiv_id":"2501.17988","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Combining distillation with two-level sampling reduces the statistical error of disconnected diagrams in quenched QCD, with variance scaling as 1/N1^2 when the quark loops are in different regions.","lead":"This lattice QCD paper combines two advanced techniques, distillation and two-level sampling, to compute quark-loop (disconnected) contributions to glueball two-point functions in quenched QCD. It shows the statistical error falls like the square of the number of inner measurements when the two quark loops lie in separate regions, which would make future glueball spectroscopy cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bias correction in Eq. (19) is a one-level estimator, so its variance has an N1-independent floor; the final estimator Eq. (20) therefore cannot exhibit the advertised 1/N1^2 scaling, and the gain shown in Fig. 3 is only established for the biased estimator Eq. (18).","rationale":"The reader's weakest-assumption analysis correctly identified the bias correction as the load-bearing point. The present stress-test sharpens that concern: it is not only that δ has 1/√N1 noise, but that δ, as a one-level average over the full quark loops, retains an N1-independent level-0 variance. This makes the claimed 1/N1^2 scaling mathematically impossible for the corrected estimator unless the level-0 component of δ vanishes. The empirical evidence in Fig. 3 (filled circles above empty circles) is consistent with this floor, and the paper's 'no significant bias' check does not address it. Because the paper is an interim proceedings report rather than a final physics result, the appropriate disposition is unchanged from the reader's CONDITIONAL verdict: the method is promising, but the final unbiased estimator's variance behavior must be demonstrated and the role of δ clarified. No ad hominem is involved; the critique is purely about the estimator's variance budget. The proposed test is a straightforward re-analysis of existing data and would settle whether the floor is present, and whether the paper's central claim should be attributed to the biased or unbiased estimator.","tokens_in":7854,"tokens_out":9820,"duration_ms":107405,"concrete_test":"From the stored N0×N1 data, compute δ (Eq. 19) and \\tilde{C}^{2-lvl} (Eq. 18) separately for Δt/a > 11. Estimate Var(δ) with the level-0 Γ-method for N1 = 1, 50, 200 and, if feasible, generate N1 = 400 and 800 submeasurements. Fit Var(δ) = c0 + c1/N1. If c0 differs from zero at more than 2σ, the corrected estimator has an N1-independent floor and the 1/N1^2 scaling in Section 4.2 cannot apply to Eq. (20). Additionally, check whether dropping δ from Eq. (20) shifts the central values by more than the statistical error of \\tilde{C}^{2-lvl}; if the shift is negligible, the unbiased correction is unnecessary and the paper should be reworded to state the claim for the biased estimator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The corrected estimator C^{2-lvl} in Eq. (20) is not a two-level estimator of the disconnected correlator; it is the two-level estimator of the approximated correlator plus the one-level correction δ defined in Eq. (19). Because δ is an average of full-quark-loop products over the same N0×N1 configurations, its variance contains an N1-independent contribution from level-0 fluctuations: for fixed i, averaging over j does not erase fluctuations of the full loop inherited from the frozen/level-0 gauge field. Thus Var(δ) has a floor that is not reduced by increasing N1, and Var(C^{2-lvl}) cannot scale as 1/N1^2 unless the level-0 component of δ vanishes exactly, which is not argued or tested. Fig. 3 shows the empirical consequence: the filled circles (corrected error) lie above the empty circles (uncorrected two-level error) over much of the time range. The paper's statement that 'we do not observe a significant bias ... at least for this particular observable' (Section 4.2) addresses the expectation of δ, not its variance. Therefore the central claim that two-level integration 'reduces substantially the statistical error' with 1/N1^2 variance scaling is established only for the biased estimator in Eq. (18), not for the unbiased estimator in Eq. (20) that is used in the physics results of Fig. 4. A remedy would require factorizing the correction itself along the lines of Ref. [9], or demonstrating that δ is negligible so that Eq. (18) can be used without correction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This proceedings contribution reports an exploratory combination of two-level sampling with distillation for disconnected quark-loop correlators in quenched QCD. On a single 16^3x64 ensemble at beta=6.0 with m_pi ~ 760 MeV, the authors compare one-level, approximate two-level, and bias-corrected two-level estimators (Eqs. 16-20) for scalar disconnected two-point functions. They find that the approximate two-level estimator reduces statistical errors when the quark loops lie in different dynamical regions, with variance reportedly scaling as 1/N1^2, and they show a modest extension of the disconnected-signal effective mass to larger separations. The central caveat is that the bias-corrected estimator (Eq. 20) uses a correction delta (Eq. 19) whose variance has an N1-independent one-level component, so the advertised variance reduction is not established for the estimator actually used in the physics results.","tokens_in":8208,"tokens_out":7464,"duration_ms":76728,"significance":"If the variance reduction for disconnected diagrams is confirmed, the method would be a useful step toward affordable glueball and singlet-meson spectroscopy with fermions. The paper is a concrete first demonstration that distillation and domain-decomposed propagators can be combined with multilevel integration, and the error analysis uses the Gamma-method with explicit configuration counts. However, the central quantitative claim currently rests on the uncorrected two-level estimator, while the corrected estimator used for the effective-mass comparison has not been shown to retain the 1/N1^2 scaling. The issue is load-bearing rather than cosmetic, and it needs to be addressed before the stated conclusions can be considered supported.","major_comments":[{"comment":"The bias correction delta in Eq. (19) is the difference of two one-level estimators, C^{1-lvl} and \\tilde C^{1-lvl}, both averaged over the N0 x N1 configurations. For a fixed level-0 configuration i, the full quark loop in Eq. (14) depends on the frozen-region gauge fields, so the average over j does not remove the level-0 fluctuation of delta. Consequently Var(delta) has an N1-independent component of order 1/N0, and adding delta to \\tilde C^{2-lvl} introduces a noise floor into C^{2-lvl} that prevents the advertised 1/N1^2 scaling in general. The empirical consequence is visible in Fig. 3, where the filled circles (corrected estimator) lie above the empty circles (uncorrected estimator) over much of the time range. The sentence in Section 4.2 stating that 'the variance scales with 1/N1^2' is therefore only supported for \\tilde C^{2-lvl} in Eq. (18), not for the corrected estimator used in Fig. 4. The statement 'we do not observe a significant bias' addresses the expectation value of delta, not its variance, so it does not resolve this concern.","section":"Section 4.2, Eqs. (19)-(20), Fig. 3"},{"comment":"The conclusion that 'the two-level integration reduces substantially the statistical error of the disconnected contributions' is demonstrated for the uncorrected estimator in Eq. (18), but the corrected estimator in Eq. (20) is the one used for the physical results in Fig. 4. Until either delta is shown to be negligible in the relevant Delta t range, or delta is recomputed with a factorized estimator whose variance is also reduced by the two-level integration, the final estimator's gain is not established. A concrete test would be to plot Var(C^{2-lvl}) versus N1 alongside Var(\\tilde C^{2-lvl}) and compare the slopes, including a fit to a floor term.","section":"Section 5, Conclusion"},{"comment":"The claim that the variance scales as 1/N1^2 is made without a quantitative fit. With only N1 values of 1, 50, and 200 and N0=101, the data are sufficient to fit Var(N1) = a/N1^2 + b, which would directly reveal whether the correction floor b is significant. Adding such a fit, or at least showing the ratio of the filled-to-empty circle errors, would make the scaling statement precise and would quantify how much of the two-level gain survives in the corrected estimator.","section":"Section 4.2, Fig. 3"}],"minor_comments":[{"comment":"The text refers to 'C_I(t)' for the scalar correlator, while Eqs. (16)-(20) use the subscript Gamma; please define the relation between these notations.","section":"Section 4.2"},{"comment":"The orange data points are said to be shifted along the x-axis for visibility, but the shift amount is not stated; this makes the comparison of the full and approximated loops at a given time slice ambiguous.","section":"Fig. 2"},{"comment":"The symbol m_pi denotes the mass of the lightest pseudoscalar state in the quenched approximation; this is a valence-quark mass parameter, not the physical pion mass, and the text should state this explicitly.","section":"Eq. (10) and Section 4.1"},{"comment":"The footnote states that autocorrelation effects are suppressed, but no integrated autocorrelation times are reported; providing these estimates for the loop observables would strengthen the error analysis.","section":"Section 4.1, footnote 1"},{"comment":"The truncation to 10 Laplacian eigenvectors is introduced without a convergence study; a short check showing the stability of the disconnected correlator as a function of the number of eigenvectors would help assess the systematic error from distillation.","section":"Section 2, Eq. (7)-(8)"},{"comment":"The figure legend is difficult to parse: the labels include both \\tilde sigma_C and sigma_C for several N1 values, and the correspondence between symbols and estimators is not fully specified in the caption. Please make the caption self-contained.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a proceedings contribution, and the incremental nature relative to Refs. [4,5,9] is acceptable for its venue. The main problem is substantive: the corrected estimator used for the physics results has a one-level noise floor from Eq. (19), and the paper does not quantify how much of the two-level gain survives. This is fixable within the scope of the manuscript by adding a variance-scaling analysis of the corrected estimator or by recomputing the correction with a factorized estimator. I would support acceptance after that issue is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know: this is a genuinely new combination—distillation plus two-level sampling for disconnected quark-loop correlators—and it is written clearly and honestly. But the headline error-reduction claim is only rigorously demonstrated for the biased estimator in Eq. (18). Once the bias correction δ (Eq. 19) is added, the variance floor from that one-level correction prevents the 1/N1^2 scaling advertised, and Fig. 3 shows it: the filled circles sit above the empty circles for much of the time range.\n\nWhat is actually new: the paper shows, on one quenched ensemble (β=6.0, 16^3×64, mπ~760 MeV), that approximated quark propagators from domain decomposition can be combined with distillation and two-level sampling to reduce the error of disconnected scalar two-point functions when the loops are in different dynamical regions. The unbiased estimator (Eq. 20) does still gain over standard sampling, but the gain is downgraded relative to the biased one. The authors are admirably candid that no effective-mass plateau is reached; they do not oversell this as a physics result.\n\nThe soft spots, in proportion: (1) The bias correction is a one-level estimator. Its variance has an N1-independent piece from the level-0 fluctuations, so the final estimator cannot scale as 1/N1^2 unless δ is negligible or itself factorized. The paper only tests bias, not variance, and only for one observable. This is a load-bearing gap in the central claim. (2) The demonstration is a single ensemble, single channel, no release of data or code mentioned. That is fine for a proceedings, but it limits the strength of the conclusion. (3) The references and context are fine; self-citation is to prior pure-gauge work on the same method and is appropriate.\n\nWho this is for: lattice practitioners working on multilevel algorithms and glueball/fermionic disconnected diagrams. It deserves a serious referee, but the referee should push on the bias-correction variance. I would recommend the authors either show that δ is numerically negligible where the two-level gain matters (and use Eq. 18 alone), or redesign the correction so it is also two-level. Conditionally acceptable.","headline":"A clean, honest methods progress report with a real statistical gap: the corrected estimator's error does not achieve the advertised 1/N1^2 scaling because the bias correction is one-level.","tokens_in":8762,"tokens_out":4816,"would_cite":true,"duration_ms":46866,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["11.15.Ha","12.38.Gc","02.70.Uu"],"model":"deepseek-v4-flash","headline":"Two-level sampling, combined with distillation, cuts the statistical error of disconnected quark-loop correlation functions in quenched QCD, giving a variance that falls as $1/N_1^2$ when the loops lie in different dynamical regions.","keywords":["two-level sampling","multilevel algorithm","distillation","disconnected diagrams","quark loops","quenched QCD","glueball observables","signal-to-noise ratio"],"falsifier":"Increase $N_1$ with $N_0$ fixed on the same ensemble (for example to $N_1=1000$) and plot the error of the corrected estimator $C^{2\\text{-lvl}}$; if the error stops following the $1/N_1^2$ line and instead flattens toward the $1/\\sqrt{N_1}$ behavior of the correction term, then the bias correction is the bottleneck and the exponential error reduction claimed for fermionic disconnected diagrams fails in that regime.","tokens_in":7638,"feed_emoji":"⚛️","tokens_out":9971,"duration_ms":86195,"temperature":0.7,"pith_summary":"This paper combines distillation with a two-level sampling algorithm to reduce the statistical error of disconnected quark-loop correlation functions — the pieces of glueball and singlet-meson observables that are the most demanding to compute — in quenched QCD. By factorizing the quark propagator into overlapping domains with Dirichlet boundary conditions, the authors sub-average quark loops living in different temporal regions independently, then correct for the factorization bias with a one-level term. On a single $16^3 \\times 64$ ensemble at $\\beta=6.0$ ($m_\\pi \\approx 760$ MeV) they show the variance of the corrected two-level estimator falls as $1/N_1^2$ when the loops lie in different dynamical regions, compared with $1/N_1$ for standard sampling. The result extends the pure-gauge glueball speedup to fermionic disconnected diagrams and pushes the effective-mass signal of the would-be $f_0$ state to larger time separations.","feed_headline":"Two-level sampling cuts disconnected glueball error in QCD","feed_subtitle":"Statistical error of quark-loop correlators falls as the inverse square of sub-update count when loops lie in separate domains.","key_machinery":"The load-bearing mechanism is the factorization of the quark propagator by domain decomposition: the temporal extent is split into four regions $\\Lambda=\\Lambda_0\\oplus\\Lambda_1\\oplus\\Lambda_2\\oplus\\Lambda_3$, with $\\Lambda_0$ and $\\Lambda_2$ updated (dynamical) and $\\Lambda_1$ and $\\Lambda_3$ frozen, and the propagator is approximated on the overlapping domains $\\Omega_0$ and $\\Omega_1$ with Dirichlet boundary conditions. This approximation is justified by exponential locality, $\\|D^{-1}(y,x)\\|\\sim e^{-\\frac12 m_\\pi |y-x|}$, which makes distant gauge-field fluctuations irrelevant for a given quark loop. Distillation enters through the elementals $\\phi$ and perambulators $\\tau$ that compute the quark-loop traces, and the two-level estimator together with the bias correction $\\delta$ carries the variance reduction: the $1/N_1^2$ scaling appears because sub-averages over independent level-1 updates in different dynamical regions decorrelate, as in the pure-gauge analysis.","core_discovery":"The paper establishes that a two-level estimator built from an approximated, domain-decomposed quark propagator reduces the statistical error of disconnected quark-loop two-point functions in quenched QCD. The central object is the corrected estimator $C^{2\\text{-lvl}}(t_1,t_0) = \\tilde{C}^{2\\text{-lvl}}(t_1,t_0) + \\delta(t_1,t_0)$, where $\\tilde{C}^{2\\text{-lvl}}$ sub-averages the quark loops in separate dynamical regions using the approximated propagator, and $\\delta = C^{1\\text{-lvl}} - \\tilde{C}^{1\\text{-lvl}}$ removes the bias of the approximation using full-propagator one-level statistics. The paper reports that when the two loops are placed in different dynamical regions the variance of the two-level part scales as $1/N_1^2$, with corrections that fall exponentially with distance from the frozen regions, and that adding $\\delta$ does not visibly bias this particular observable. The net effect is a reduction in the error of the disconnected scalar correlator and a longer usable window in the effective mass.","pith_inferences":["A direct consequence of the reported $1/N_1^2$ scaling is that any disconnected diagram built from quark loops should inherit the same variance reduction, not just the scalar channel, provided the loops sit in distinct dynamical regions and $\\delta$ stays subdominant.","A testable extension would be to estimate $\\delta$ from two independent one-level samples; if $\\delta$'s variance dominates at larger $N_1$, the method would need a multi-level estimator for the correction itself, which the paper does not provide.","Because the factorization rests on exponential locality, the speedup should strengthen at heavier quark masses and weaken near the chiral limit; the present $m_\\pi\\approx760$ MeV setup is a favorable case, so the low-mass regime remains open.","The Dirichlet-boundary effects visible near the interfaces are repaired only through the one-level $\\delta$; a fully local two-level scheme would need to control these boundary effects directly, for instance by enlarging the overlap regions."],"forward_implications":["With $N_0=101$ and $N_1=200$ sub-updates per level-0 configuration, two-level sampling extends the disconnected scalar correlator signal by a few more time slices than standard sampling; the paper notes the statistics are not yet sufficient for a reliable effective-mass plateau.","Because the variance scales as $1/N_1^2$ when the quark loops are in different dynamical regions, increasing the sub-measurements by a factor of 5 would reduce the error by a similar factor for $t>11a$, a route the authors identify toward a visible plateau.","The same corrected two-level estimator applies to the other singlet channels $\\Gamma=1,\\gamma_5,\\gamma_\\mu,\\gamma_4\\gamma_5,\\gamma_i\\gamma_j$, with the connected pieces still computed at one level.","For full QCD the method requires a local factorization of the fermion determinant, which the paper frames as the next step toward glueball studies in full QCD."],"supporting_citations":[{"why":"Supplies the domain-decomposition factorization of the quark propagator that makes fermionic observables amenable to multi-level integration.","marker":"[9]"},{"why":"Provides the pure-gauge two-level algorithm and the exponential-error-reduction scaling result (including the weighted-average construction) that this work extends to fermionic disconnected diagrams.","marker":"[5]"},{"why":"Supplies the distillation framework (elementals and perambulators) used to compute the quark-loop traces and contractions.","marker":"[13]"},{"why":"Gives the domain-decomposition Dirac solver with Dirichlet boundary conditions used to build the approximated propagator $D^{-1}_{\\Omega_r}$.","marker":"[16]"},{"why":"Provides the $\\Gamma$-method used to estimate statistical errors on the level-0 configurations.","marker":"[17]"},{"why":"Supplies the pure-gauge scalar glueball mass ($m_{0^{++}}\\approx 1560$ MeV) used as a reference scale for the would-be $f_0$ signal.","marker":"[15]"},{"why":"Establishes locality and exponential error reduction in lattice gauge theory, the theoretical basis for two-level sampling.","marker":"[3]"}],"fun_headline_variants":["Two-level sampling cuts disconnected glueball error","Disconnected glueball error drops with two-level scheme","Two-level method reduces quark-loop noise in QCD","Error scaling N^-2 improves glueball correlators","Two-level estimator shrinks glueball statistical error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the bias correction $\\delta$, computed with one-level full-propagator statistics, stays small enough and accurate enough that it does not re-introduce the one-level $1/\\sqrt{N_1}$ noise; the paper checks this only by observing no significant bias for this particular observable.","fun_headline_variants_meta":{"raw":{"variants":["Two-level sampling cuts disconnected glueball error","Disconnected glueball error drops with two-level scheme","Two-level method reduces quark-loop noise in QCD","Error scaling N^-2 improves glueball correlators","Two-level estimator shrinks glueball statistical error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1600,"prompt_tokens":823,"completion_tokens":777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":704}},"tokens_in":439,"tokens_out":777,"duration_ms":7964,"temperature":1.0,"reasoning_tokens":704,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:25:32.811080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Increase $N_1$ with $N_0$ fixed on the same ensemble (for example to $N_1=1000$) and plot the error of the corrected estimator $C^{2\\text{-lvl}}$; if the error stops following the $1/N_1^2$ line and instead flattens toward the $1/\\sqrt{N_1}$ behavior of the correction term, then the bias correction is the bottleneck and the exponential error reduction claimed for fermionic disconnected diagrams fails in that regime.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the distillation framework (elementals and perambulators) used to compute the quark-loop traces and contractions."}],"review_version":1}