Machine-learning techniques as noise reduction strategies in lattice calculations of the muon $g-2$

Alessandro Conigli; Alexander Segner; Hartmut Wittig; Lukas Geyer; Simon Kuberski; Thomas Blum

REVIEW 3 major objections 4 minor 14 references

Machine-learning techniques as noise reduction strategies in lattice calculations of the muon $g-2$

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Machine learning with bias correction halves the cost of isospin corrections to baryon masses, the paper reports.

desk verdict Honest LATTICE2024 methods report: bias-corrected ML gives a clean pseudoscalar result, a clearly reported negative result for the vector rest-eigen correlator, and a promising but not fully quantified factor-two saving for baryon QED corrections. read the letter →

arxiv 2502.10237 v1 pith:O4I5MJV3 submitted 2025-02-14 hep-lat

Thomas Blum , Alessandro Conigli , Lukas Geyer , Simon Kuberski , Alexander Segner , Hartmut Wittig This is my paper · ORCID

classification hep-lat

keywords latticeQCDmuong-2hadronicvacuumpolarizationmachinelearningbiascorrectionisospinbreakingnoisereductionbaryonmasses

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper investigates whether a trained machine-learning model can act as a cheap approximate estimator inside a bias-correction scheme to reduce the numerical cost of lattice QCD calculations relevant to the muon anomalous magnetic moment. For electromagnetic isospin-breaking corrections to octet and decuplet baryon masses, it finds a strong correlation between the QED corrections and quark-mass-detuning contributions, and a simple linear model exploiting this correlation, combined with an exact bias correction on a small sample, reproduces the exact result with comparable errors while reducing computer time to target precision by about 50%. For the rest-eigen contribution to the vector-vector correlator that enters the hadronic vacuum polarization, the same strategy yields a bias-corrected prediction that is consistent with the exact calculation but has a statistical error roughly twice as large at long Euclidean times, so no cost saving is achieved there. The paper concludes that the value of the method is application-dependent and that better models of the correlations among the vector correlator's components are needed.

What carries the argument

The central mechanism is the bias-correction identity of Eq. (4), $\langle O\rangle = \langle O_{\rm appx}\rangle + \langle O - O_{\rm appx}\rangle$, in which a cheap approximate estimate $O_{\rm appx}$ is computed on many configurations and an exact evaluation of $(O - O_{\rm appx})$ on a smaller set removes the bias; the paper replaces the truncated-solver approximation of all-mode averaging with a trained model. For the isospin-breaking application, the load-bearing object is the linear model of Eq. (7), $M(t) = \alpha(t) C^{(0)}(t) + \beta(t) C^{(1)}_{\Delta m_u}(t) + \gamma(t) C^{(1)}_{\Delta m_d}(t) + \delta(t) C^{(1)}_{\Delta m_s}(t) + \epsilon(t)$, whose coefficients are trained on 20 configurations to predict the QED correction from the mass-detuning correlators.

What would settle it

Apply the same 20-configuration training and one-source bias-correction protocol to a second lattice ensemble with a different lattice spacing or pion mass; if the bias-corrected QED mass corrections no longer show competitive statistical errors compared with the exact 32-source evaluation, the claimed 50% cost reduction does not transfer.

Watch

Extended reading notes

Core claim

The paper's central claim is that a trained machine-learning model can serve as the approximate estimator in a bias-correction scheme, $\langle O\rangle = \langle O_{\rm appx}\rangle + \langle O - O_{\rm appx}\rangle$, and thereby lower the numerical cost of some lattice QCD observables. For electromagnetic isospin-breaking corrections to octet and decuplet baryon masses, the paper shows that a simple linear model using the quark-mass-detuning correlators $C^{(1)}_{\Delta m_q}(t)$ as input predicts the QED contribution $C^{(1)}_{e^2}(t)$ accurately enough that, after a bias correction computed on just one of the 32 quark sources, the errors are competitive with the exact calculation. This yields a reduction in computer time of about 50%. For the rest-eigen contribution to the vector-vector correlator $G(t)$ entering $a_\mu^{\rm hvp}$, the bias-corrected machine-learning prediction is consistent with the exact result but its statistical error is about twice as large at large Euclidean times, so the method does not yet reduce cost there.

Load-bearing premise

The method's advertised cost saving rests on the empirical premise, tested on one ensemble with 20 training configurations, that the electromagnetic corrections to baryon masses stay strongly and stably correlated with quark-mass-detuning correlators on the full 991-configuration test set and in all studied baryon channels.

Editorial extensions

If this is right

For isospin-breaking corrections to octet and decuplet baryon masses, the linear model with bias correction reduces the computer time needed to reach target precision by approximately 50%, because the QED contribution accounts for about half the total cost and the correction step uses one quark source instead of 32.
Increasing the number of quark sources used in the bias correction beyond one does not further reduce the statistical error of the bias-corrected result, so the efficiency gain is not diluted by adding sources.
For the rest-eigen contribution to the vector-vector correlator entering the hadronic vacuum polarization, the bias-corrected machine-learning prediction is consistent with the exact result but carries about twice the statistical error at large Euclidean times; reducing that error would require a bias-correction sample large enough that the cost saving disappears.
If a future model increases the correlation between the approximate and exact evaluations of the rest-eigen contribution, the method could cut the CPU time of the HVP calculation by up to 50%, since the rest-eigen term makes up about half of the computational effort and the training overhead is negligible.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The factor-of-32 source reduction in the isospin study was demonstrated on one ensemble with 20 training configurations; on another ensemble or in another baryon channel the linear correlation between QED corrections and mass detuning could be weaker, so the advertised 50% saving should be re-tested before being assumed.
In the vector-correlator case, the paper's variance decomposition suggests the ML prediction did not shrink the rest-eigen variance itself, only replaced its cost profile; an alternative use of the trained model as a control variate, subtracting its prediction from the exact data without a separate bias-correction term, might extract variance reduction even when the bias-corrected estimator does n
The same 'train cheap, correct exact' strategy could be applied to other expensive lattice quantities such as disconnected diagrams or long-distance hadronic observables, but the vector-correlator result warns that the error of the bias correction, not the training quality, is what can dominate the total error.
A straightforward extension would be to assess the isospin model's transferability by cross-validating across time slices or by training on a few configurations and testing the saved parameters on the omitted ones, which the paper does not report.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short take: an honest, workmanlike proceedings paper. The bias-correction-plus-ML approach works cleanly for a pseudoscalar test, fails for the rest-eigen part of the vector correlator (the paper says so plainly), and shows real promise for QED corrections to baryon masses. The advertised factor-of-two saving is not fully nailed down, but the underlying strategy is sound.

What's new: applying the AMA/Machine-Learning-estimator idea to predict the noise-dominated rest-eigen contribution of vector/pseudoscalar correlators, and a linear model that predicts the QED correction to baryon masses from mass-detuning terms. Both are concrete numerical studies, citing the prior bias-correction work properly. The pseudoscalar check is nice: the mismatch A(t) and the bias correction B(t) agree within errors, which is a good internal consistency test. The vector result is a genuinely useful negative result: bias correction alone does not reduce the error, and the paper doesn't oversell it. For baryons, the correlation in Fig. 6 is visually strong, and Fig. 7 shows the bias-corrected result agrees with the exact calculation within comparable errors.

Soft spots: the central cost-saving claim relies on a single-source bias correction whose variance is asserted to be constant when the number of sources increases, but no quantitative evidence is given. That sentence is load-bearing: if the single-source bias error is not subdominant, the 50% saving shrinks or vanishes. The stress-test note lands exactly here. Also, the training set is only 20 configurations on one ensemble (N451); the test set of 991 gives a fair check, but the transferability to other ensembles is unknown. Minor issues: some figures lack explicit error bands, and no code or full error budget is provided. These are addressable in revision.

This paper deserves a serious referee. It is honest, the bias-correction identity is exact, and the negative result for the vector correlator is useful information for the community. I'd recommend accepting after minor revision, with the request that the authors show the error on the baryon results as a function of bias sources, or at least a table. I'd bring it to a reading group as a methods note, and I'd cite it if I were working on g-2 isospin breaking.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a machine-learning-based noise-reduction strategy for lattice QCD calculations relevant to the muon g-2 program. The method follows the all-mode-averaging logic: a cheap approximate estimate O_appx is combined with an exact bias correction (O - O_appx), so the final estimate is unbiased by construction (Eq. (4)). Two applications are reported. First, for the rest-eigen contribution to pseudoscalar and vector two-point correlators, a neural network is trained on the eigen-eigen and rest-rest parts; for the pseudoscalar correlator the bias-corrected prediction agrees with the exact calculation, whereas for the vector correlator the bias-corrected error is approximately twice the exact error at large Euclidean times. Second, for electromagnetic isospin-breaking corrections to baryon masses on ensemble N451, a linear model (Eq. (7)) maps quark-mass-detuning correlators to the QED contribution, and a bias correction evaluated on a single quark source is claimed to yield errors competitive with the exact 32-source calculation, reducing the computer time by about 50%.

Significance. If the claimed factor-of-two saving for the baryon electromagnetic corrections holds, it is practically valuable for scale setting and for the overall precision budget of lattice g-2 calculations. The paper is honest about the less favorable vector-correlator case, reporting that the bias-corrected error is about twice the exact error and that the bias correction is the dominant source of error and cost. The pseudoscalar test, with the comparison of A(t) and B(t) in Fig. 3, is a useful diagnostic of the method. However, the central quantitative claim of a 50% cost reduction rests on an empirical correlation and on a single-source bias correction whose variance and source dependence are not documented. The manuscript does not ship code or data, so the main checkable content is the reported figures and the textual statements about the error budget.

major comments (3)

[Section 4] The statement 'We have studied the effect of increasing the number of sources in the bias correction and found that the errors stayed approximately constant' is not supported by any quantitative evidence in the manuscript. The efficiency condition for Eq. (4) requires the bias-correction term (O - O_appx) to have sub-dominant variance and to be representative of all 32 sources; a single-source estimate may be noisier than the average over sources or may vary with source position. Please provide an error budget: the per-source variance of the correction, a comparison of one-source versus multi-source bias corrections (for example 1, 2, 4, 8, 16, and 32 sources), and a test of source-position dependence. Without this, the claimed factor-of-two gain cannot be assessed.
[Section 4, Eq. (7)] The linear model has five free parameters per timeslice but is trained on only 20 configurations, all on a single ensemble (N451). The paper does not report cross-validation results, regularization, or the stability of the fitted coefficients, nor does it separate the model prediction variance from the bias-correction variance. Since the claimed saving depends on the model generalizing from 20 training configurations to the 991-configuration test set, please quantify the generalization error by varying N_train, reporting residuals in different baryon channels, and ideally testing on another ensemble. If the observed correlation between QED corrections and quark-mass detuning is weaker elsewhere, the advertised transferable cost reduction does not follow.
[Section 5 / Fig. 7] The conclusion states a 50% reduction in computer time 'to reach our targeted precision', but no numerical comparison of the total statistical error of the bias-corrected result with the exact result is given. Fig. 7 displays effective masses with error bars but no correlated difference or error ratio. Please provide a table of the exact and bias-corrected baryon-mass errors for the shown channels, including the contribution of the single-source bias correction, so that the precision-preserving claim can be verified quantitatively.

minor comments (4)

[Section 1] The first sentence of Section 1 reads 'The hadronic vacuum polarization (HVP) contribution, a_hvp_mu, to the muon anomalous magnetic has been a major focus'; 'magnetic' should be followed by 'moment'.
[Figure 5 caption] The phrase 'The upper plot in each panel shows the exact calculation' is unclear; it should say 'upper row' or 'upper panel' to match the arrangement of the figure.
[Section 4] The fact that QED corrections account for about 50% of the computer time is stated twice in Section 4 ('with 50% of the computer time spent...' and 'Given that QED corrections account for about 50% of the total time'); a single cost-breakdown statement would be less repetitive.
[Section 3] The text switches from describing neural-network regression to the linear model of Eq. (7) in Section 4 without an explicit transition; it would help to state clearly that Section 3 uses a neural network while Section 4 uses a linear model, and to explain the choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: ML outputs are always combined with exact bias corrections, and the final estimates do not reduce to the fitted models; the paper's self-citations supply only background formalism.

full rationale

The paper's derivation chain is self-contained for its actual claims. Equation (4) is the standard AMA identity <O> = <O_appx> + <O - O_appx>, and every ML result is reported after such a bias correction: for the rest-eigen contribution to the vector correlator, the bias-corrected prediction is compared with the exact correlator, with the paper explicitly stating that the statistical error is about twice as large at large Euclidean times; for the baryon isospin corrections, Fig. 7 shows 'bias-corrected prediction' curves, i.e. the model prediction augmented by the exact QED evaluation on (at least) one quark source, not raw model output. The central factor-two saving is an accounting claim: QED corrections are stated to consume about 50% of the computer time, and using one exact source instead of 32 reduces that part by a factor of 32; whether the resulting error stays competitive is an empirical variance question, not a definitional reduction. The linear model of Eq. (7) is fitted to QED corrections on 20 training configurations and applied to test configurations, with the final result again bias-corrected; no fitted parameter is renamed as an independent prediction. The self-citations [6] and [12] are background references for the LMA decomposition and the RM123 baryon implementation, respectively, and neither is load-bearing for the ML noise-reduction result. The statement 'We have studied the effect of increasing the number of sources in the bias correction and found that the errors stayed approximately constant' lacks a quantitative table or plot, which is an evidence/completeness concern rather than a circular step.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or forces. Its central claims rest on a set of fitted regression parameters, the LMA spectral decomposition, the exact bias-correction identity, the RM123 expansion, and an assumption of generalization from a small training set.

free parameters (5)

Linear model coefficients alpha(t), beta(t), gamma(t), delta(t), epsilon(t) = not tabulated
Eq. (7) parameters fitted to 20 training configurations of N451; the model's predictive quality, and hence the cost saving, depends on them.
Neural network hyperparameters (hidden units, dropout, epochs, regularization) = not provided
Selected by grid search; exact values absent, so the method is not exactly reproducible.
Training set size N_train and bias set size N_bias = Ntrain=100..500 for correlators; Ntrain=20, Nbias=991 for baryons
Chosen by hand; the balance between prediction quality, bias-correction cost, and total error depends on these.
Number of low modes N_low = O(1000)
Defines the LMA split and the cost of the exact rest-eigen calculation.
Number of quark sources used in bias correction = 1 (32 in training)
This choice drives the factor-32 reduction in QED source evaluations and is central to the claimed saving.

assumptions (5)

standard math The decomposition S = S_eigen + S_rest (Eq. 2) and the spectral representation Eq. (3) are exact.
Standard linear algebra of the hermitian Wilson-Dirac operator; invoked in Section 2.
standard math The identity <O> = <O_appx> + <O - O_appx> (Eq. 4) gives an unbiased estimator when both averages are over the same distribution.
Linearity of expectation; used throughout as the basis of bias correction.
domain assumption The RM123 expansion about iso-symmetric QCD gives the leading QED and quark-mass corrections used in Section 4.
Adopted from Refs. [13,14]; not rederived in this paper.
domain assumption Training set is statistically independent of bias/test sets, so bias correction removes model bias without contamination.
Stated in Section 3 that the training set is completely disjoint; this underpins the exactness of the corrected estimate.
ad hoc to paper The model generalizes from 20 training configurations to the 991-configuration test set on N451.
Empirical assumption; the paper presents evidence for one ensemble but no error model for generalization elsewhere.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine-learning techniques as noise reduction strategies in lattice calculations of the muon $g-2$." pith.science (2026). https://pith.science/paper/O4I5MJV3

@misc{pith2026250210237,
  author       = {Pith},
  title        = {Pith review of: Machine-learning techniques as noise reduction strategies in lattice calculations of the muon $g-2$},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4I5MJV3}},
  note         = {Machine review of arXiv:2502.10237}
}

read the original abstract

Lattice calculations of the hadronic contributions to the muon anomalous magnetic moment are numerically highly demanding due to the necessity of reaching total errors at the sub-percent level. Noise-reduction techniques such as low-mode averaging have been applied successfully to determine the vector-vector correlator with high statistical precision in the long-distance regime, but display an unfavourable scaling in terms of numerical cost. This is particularly true for the mixed contribution in which one of the two quark propagators is described in terms of low modes. Here we report on an ongoing project aimed at investigating the potential of machine learning as a cost-effective tool to produce approximate estimates of the mixed contribution, which are then bias-corrected to produce an exact result. A second example concerns the determination of electromagnetic isospin-breaking corrections by combining the predictions from a trained model with a bias correction.

Figures

Figures reproduced from arXiv: 2502.10237 by the authors.

**Figure 2.** Left: The rest-eigen part of the pseudoscalar correlator on ensemble A654 as predicted by the model compared with the exact calculation for 𝑁train = 200. Right: Deviation between prediction and exact calculation for different sizes of the training set. Hartmut Wittig Test quality of predic3on for “rest-eigen” contribu3on — bias-corrected (A654 ensemble) Pseudoscalar correlator: rest-eigen contribu3on 2 12 18 24 30 3… view at source ↗

**Figure 3.** Left: The bias-corrected rest-eigen part of the pseudoscalar correlator on ensemble A654 with the exact calculation for 𝑁train = 200, 𝑁bias = 300. Right: Bias correction 𝐵(𝑡) for different values of 𝑁bias. The deviation 𝐴(𝑡) between prediction and exact evaluation is shown as the black star. sets yield a better prediction, the values of 𝐴(𝑡) are quite stable when 𝑁train ≳ 200. The left panel of [PITH_FULL_IMAGE:fig… view at source ↗

**Figure 5.** Fractional contributions of the eigen-eigen (ee), rest-eigen (re) and rest-rest (rr) parts to the total error of the correlator as a function of Euclidean time for ensemble D450. The upper plot in each panel shows the exact calculation with the machine-learning model shown in the lower plots. The vector correlator is shown in the two panels on the left, while the pseudoscalar case is shown on the right. shows that t… view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: The contributions from the strange quark mass detuning (left panel) and electromagnetic corrections (right panel) to the effective mass of the Ω− baryon, in lattice units. Hartmut Wittig Results 5 • Increasing has no effect on the uncertainty in the bias-corrected resu…

**Figure 7.** Figure 7: Electromagnetic corrections to the effective mass of the Ω− (left panel) and Ξ − (right panel) baryons plotted in lattice units. The insets show the enlarged region for 𝑡/𝑎 = 16 − 21. 𝐺(𝑡) in the long-distance regime is less straightforward. Although our machine-learne…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages

[1]

Bernecker and H.B

D. Bernecker and H.B. Meyer,Vector Correlators in Lattice QCD: Methods and applications,Eur. Phys. J. A47(2011) 148 [1107.4388]

arXiv 2011
[2]

Parisi,The Strategy for Computing the Hadronic Mass Spectrum, Phys

G. Parisi,The Strategy for Computing the Hadronic Mass Spectrum, Phys. Rept.103 (1984) 203

work page 1984
[3]

Lepage,The Analysis of Algorithms for Lattice Field Theory, inBoulder ASI 1989:97-120, pp

G.P. Lepage,The Analysis of Algorithms for Lattice Field Theory, inBoulder ASI 1989:97-120, pp. 97–120, 1989

work page 1989
[4]

Giusti, P

L. Giusti, P. Hernandez, M. Laine, P. Weisz and H. Wittig,Low-energy couplings of QCD from current correlators near the chiral limit, JHEP 04(2004) 013 [hep-lat/0402002]

arXiv 2004
[5]

DeGrand and S

T.A. DeGrand and S. Schaefer,Improving meson two point functions in lattice QCD, Comput. Phys. Commun.159(2004) 185 [hep-lat/0401011]

arXiv 2004
[6]

Djukanovic, G

D. Djukanovic, G. von Hippel, S. Kuberski, H.B. Meyer, N. Miller, K. Ottnad et al.,The hadronicvacuumpolarizationcontributiontothemuon 𝑔− 2atlongdistances , 2411.07969

arXiv
[7]

T. Blum, T. Izubuchi and E. Shintani,New class of variance-reduction techniques using lattice symmetries, Phys. Rev. D88 (2013) 094503 [1208.4349]

arXiv 2013
[8]

B. Yoon, T. Bhattacharya and R. Gupta,Machine Learning Estimators for Lattice QCD Observables, Phys. Rev. D100 (2019) 014504 [1807.05971]

work page Pith review arXiv 2019

Show all 14 references

[9]

Hinton, N

G.E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever and R. Salakhutdinov,Improving neural networks by preventing co-adaptation of feature detectors, 1207.0580

arXiv
[10]

LeCun, L

Y.A. LeCun, L. Bottou, G.B. Orr and K.-R. Müller,Efficient backprop, inNeural Networks: Tricks of the Trade: Second Edition, G. Montavon, G.B. Orr and K.-R. Müller, eds., (Berlin, Heidelberg), pp. 9–48, Springer Berlin Heidelberg (2012), DOI

2012
[11]

Della Morte, A

M. Della Morte, A. Francis, V. Gülpers, G. Herdoíza, G. von Hippel, H. Horch et al.,The hadronic vacuum polarization contribution to the muon𝑔− 2 from lattice QCD,JHEP 10 (2017) 020 [1705.01775]

2017 arXiv
[12]

Segner, A

A.M. Segner, A. Risch and H. Wittig,Precision Determination of Baryon Masses including Isospin-breaking,PoS LATTICE2023(2023) 044 [2312.09065]

2023 arXiv
[13]

de Divitiis et al.,Isospin breaking effects due to the up-down mass difference in Lattice QCD, JHEP 04(2012) 124 [1110.6294]

G.M. de Divitiis et al.,Isospin breaking effects due to the up-down mass difference in Lattice QCD, JHEP 04(2012) 124 [1110.6294]

2012 arXiv
[14]

de Divitiis, R

G.M. de Divitiis, R. Frezzotti, V. Lubicz, G. Martinelli, R. Petronzio, G.C. Rossi et al., Leading isospin breaking effects on the lattice,Phys. Rev. D87(2013) 114505 [1303.4896]. 9

2013 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Bernecker and H.B

D. Bernecker and H.B. Meyer,Vector Correlators in Lattice QCD: Methods and applications,Eur. Phys. J. A47(2011) 148 [1107.4388]

arXiv 2011

[2] [2]

Parisi,The Strategy for Computing the Hadronic Mass Spectrum, Phys

G. Parisi,The Strategy for Computing the Hadronic Mass Spectrum, Phys. Rept.103 (1984) 203

work page 1984

[3] [3]

Lepage,The Analysis of Algorithms for Lattice Field Theory, inBoulder ASI 1989:97-120, pp

G.P. Lepage,The Analysis of Algorithms for Lattice Field Theory, inBoulder ASI 1989:97-120, pp. 97–120, 1989

work page 1989

[4] [4]

Giusti, P

L. Giusti, P. Hernandez, M. Laine, P. Weisz and H. Wittig,Low-energy couplings of QCD from current correlators near the chiral limit, JHEP 04(2004) 013 [hep-lat/0402002]

arXiv 2004

[5] [5]

DeGrand and S

T.A. DeGrand and S. Schaefer,Improving meson two point functions in lattice QCD, Comput. Phys. Commun.159(2004) 185 [hep-lat/0401011]

arXiv 2004

[6] [6]

Djukanovic, G

D. Djukanovic, G. von Hippel, S. Kuberski, H.B. Meyer, N. Miller, K. Ottnad et al.,The hadronicvacuumpolarizationcontributiontothemuon 𝑔− 2atlongdistances , 2411.07969

arXiv

[7] [7]

T. Blum, T. Izubuchi and E. Shintani,New class of variance-reduction techniques using lattice symmetries, Phys. Rev. D88 (2013) 094503 [1208.4349]

arXiv 2013

[8] [8]

B. Yoon, T. Bhattacharya and R. Gupta,Machine Learning Estimators for Lattice QCD Observables, Phys. Rev. D100 (2019) 014504 [1807.05971]

work page Pith review arXiv 2019