Control variates with neural networks

Hyunwoo Oh

REVIEW 3 major objections 4 minor 12 references

Control variates with neural networks

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper argues that neural-network Schwinger-Dyson control variates cut variance by hundreds in strong-coupling scalar field theory and exponentially for U(1) Wilson loops, while showing only the vector-valued construction is universal.

desk verdict Promising method, but the key numerical claims need out-of-sample validation and the new universality counterexample is too hastily proved. read the letter →

arxiv 2501.14614 v1 pith:LSJYTIKY submitted 2025-01-24 hep-lat

Hyunwoo Oh This is my paper · ORCID

classification hep-lat

keywords controlvariatesneuralnetworkslatticefieldtheorySchwinger-Dysonequationssignal-to-noiseproblemvariancereductionU(1)gaugeWilsonloops

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 tries to establish that the guesswork in constructing control variates for lattice field theory can be replaced by a neural network, and that this yields concrete noise reductions in theories where the signal-to-noise problem is severe. In the strongly coupled $1+1$-dimensional scalar model, adding hidden layers improves the standard deviation by two orders of magnitude relative to both the raw estimator and the linear control variate. In 2D $U(1)$ gauge theory, the method turns an exponentially degrading Wilson-loop signal into one with essentially constant relative error. A separate structural claim is that among translation-invariant Schwinger-Dyson constructions, the vector-valued one can in principle represent every control variate, while the scalar one cannot represent the perfect control variate for a two-area Wilson loop.

What carries the argument

The central object is the Schwinger-Dyson control variate $f = \sum_i(\partial_i g - g\,\partial_i S)$, whose zero expectation follows from Stein's identity $\int D\phi\, \frac{\delta}{\delta\phi}(g e^{-S}) = 0$; the neural network supplies the parameterized function $g$. Translation covariance is imposed via $g(\phi)_x = g_0(T_x[\phi])$, $Z_2$ oddness via zero biases and odd activations such as $\operatorname{arcsinh}$, and periodicity for $U(1)$ via trigonometric input features. The training minimizes $L(w,\mu) = \frac{1}{N}\sum_i (O(\phi_i) - f(\phi_i) - \mu)^2$, with the auxiliary parameter $\mu$ absorbing the overfitted sample mean so that $\bar{f}$ stays near zero while the final estimator $O-f$ remains unbiased.

What would settle it

Two concrete checks would decide it: compute $\sigma_{\mathrm{Raw}}/\sigma_{\mathrm{CV}}$ on held-out configurations never used for training, and numerically solve Eq. (21) on a torus to test for a periodic $g$. If the first ratio is order one or the second has a periodic solution, the central claims fail.

Watch

Extended reading notes

Core claim

On the paper's own terms, the result is that control variates built from neural networks are practical and, in the right parametrization, universal. The vector-valued construction $f = \sum_i(\partial_i g_i - g_i\,\partial_i S)$ covers all zero-mean functions, so with a sufficiently expressive network the loss should be able to approach the perfect control variate $f^P = O - \langle O\rangle$; the scalar variant Eq. (20) does not, because for the 2-area Wilson loop the defining differential equation Eq. (21) has no periodic solution. Numerically, the paper reports hundreds-fold standard-deviation improvements for the strong-coupling scalar correlator and exponential error reduction for $U(1)$ Wilson loops, with a transfer-learning variant giving the most precise mass extraction in Table 1.

Load-bearing premise

The numerical claim assumes that the variance reduction observed after training persists on configurations not used for training, since the paper estimates errors on the same ensemble that produced the networks.

Editorial extensions

If this is right

If the central claim is correct, Monte Carlo errorbars for strongly coupled scalar observables can be reduced by two orders of magnitude without changing the action or the sampling algorithm.
Wilson loops in 2D $U(1)$ gauge theory can be computed with nearly area-independent statistical errors, removing the exponential signal-to-noise barrier that otherwise limits large loops.
Practitioners should build control variates from vector-valued functions (Eq. (5)) when they seek universality, since the scalar ansatz (Eq. (20)) provably misses the perfect control variate in at least one compact-theory example.
Transfer learning between neighboring time slices is a viable way to train a family of control variates, improving both training time and final variance reduction.

Reading between the lines

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

Beyond the paper: if the scalar-ansatz obstruction is generic, searches for perfect control variates in compact theories should restrict to vector-valued $g$; scalar ansätze will systematically cap variance reduction.
The factorization trick in Eq. (18) looks specific to Abelian plaquettes; a testable extension is to build neural control variates on character or group-element features for non-Abelian links and check whether exponential suppression returns.
A cheap robustness check the author does not report: train on one half of the ensemble and report $\sigma_{\mathrm{Raw}}/\sigma_{\mathrm{CV}}$ on the other half; if the ratio persists, the method is ready for production scale.
The no-periodic-solution claim could be checked independently by Fourier-expanding Eq. (21); a nonzero Fourier mode would settle whether the scalar construction misses the perfect control variate without relying on the closed-form solution.

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 version: this is a clean proceedings writeup of an ongoing control-variate program; the genuinely new piece is Section 5's claim that the scalar-valued construction cannot produce the perfect control variate for a two-plaquette Wilson loop. That claim is interesting but weakly supported.

The paper does good things. The Stein-identity construction is standard, and the paper gives a clear translationally-equivariant parametrization. The 2D U(1) results are checked against exact values, which is strong evidence the estimator is unbiased and works in that setting. The transfer learning trick is a practical nicety.

The soft spot is the numerical evidence. In Sec 3.1, Fig 2 uses 1000 configurations for training and then estimates the observable on all 2000, including the training set. So the error bars for the 'CV at t=20' and 'Transfer learning' rows are in-sample. The mu-shift and L2 regularization do not prevent overfitting entirely. Fig 1 has the same ambiguity — it uses 10^4 for training and 10^3 for the std estimate, but does not say they're disjoint. The U(1) section does not clarify. Unbiasedness from the Stein identity is solid, but variance reduction measured on the training set is not yet evidence of population-level variance reduction.

Section 5's main assertion — that Eq (21) has no periodic solution — is given as 'easy to verify' with no proof. Given it's a counterexample to universality for scalar constructions, it deserves at least a sketch.

Novelty is modest; the abstract and intro frame this as a summary of [4]. The Section 5 result is the actual new contribution. I'd send this to a referee, not desk-reject, but the referee should require out-of-sample validation and a real proof of the non-periodicity claim. The paper is honest and the method looks promising, but the current numerical claims are not fully established.

Referee Report

3 major / 4 minor

Summary. The paper proposes parametrizing Schwinger-Dyson (Stein-identity) control variates with neural networks for lattice field theory. For a 1+1 dimensional scalar field theory it reports substantial variance reduction for the momentum-zero two-point correlator, including a transfer-learning variant, and for two-dimensional U(1) gauge theory it reports exponential error reduction for Wilson loops, with results checked against the exact values (I_1(β)/I_0(β))^A. A final section argues that control variates constructed from a scalar function via Eq. (20) are not universal, because the differential equation for the perfect control variate for a 2-area Wilson loop, Eq. (21), has no periodic solution.

Significance. If the numerical claims are robust, this is a useful contribution to the ongoing effort to use machine-learned control variates for signal-to-noise problems in lattice field theory. The work has several strengths: the Stein-identity construction is standard and correctly guarantees unbiasedness; the U(1) check against exact values is strong evidence that the implementation preserves the required identities; and the comparison of scalar versus vector constructions in Section 5 raises a conceptually interesting point. The main weakness is that the reported variance reductions are not demonstrably out-of-sample, so the central numerical claims are not yet established. The theoretical non-periodicity claim also lacks a proof, despite being load-bearing for the universality discussion.

major comments (3)

[§3.1, Fig. 2, Table 1] The central numerical claim is not yet established because the variance reductions are computed on configurations used for training. The Fig. 2 caption states that 10^3 of the 2×10^3 samples are used for training the control variate and "the entire samples used for estimating the observables," while the loss in Eq. (13) is minimized on exactly those 10^3 configurations. The Stein identity guarantees unbiasedness of O−f, but it does not prevent the trained network from fitting noise on the training set, so the in-sample standard deviation can substantially underestimate the true variance of the estimator. The auxiliary parameter μ and the L2 regularization in Eq. (13) limit the in-sample loss but do not by themselves make the training-sample variance representative of the population variance. Please report variance ratios computed on held-out configurations (or on independently generated ensembles) for all numerical claims, and state explicitly in Figs. 1 and 3 whether the configurations used for the error estimate are disjoint from those used for training.
[§5, Eq. (22)] The claim that the general solution in Eq. (22) has no periodic solution is load-bearing for the conclusion that the scalar construction in Eq. (20) cannot represent the perfect control variate, but it is only asserted as "easy to verify." Please provide a proof or a precise reference for this non-periodicity statement. In particular, state the periodicity condition in terms of the original variables (x,y) or equivalently (s,d), and show that no choice of C(d) in Eq. (22) satisfies the required condition such as g(s+π,d+π)=g(s,d).
[§4, Fig. 3] The U(1) result is checked against the exact values (I_1(β)/I_0(β))^A, which verifies unbiasedness, but the variance-ratio improvement shown in Fig. 3 is a separate claim. The text does not specify how many configurations were used, whether the same configurations were used for training and for the variance estimate, or how the standard deviations in the right panel were obtained. Please add these experimental details; if the variance is evaluated on the training sample, provide the corresponding held-out variance ratios.

minor comments (4)

[Fig. 1 caption] The caption contains typos: "observabale" should be "observable" and "deivation" should be "deviation."
[§2.2, Eq. (9)] The text says that since ⟨O−f⟩=⟨O⟩ is independent of the network parameters, the constant term can be omitted and the loss simplifies to Eq. (9). It would be clearer to state explicitly that the objective is the second moment rather than the variance, since the ⟨O⟩^2 term is dropped; this is valid for optimization but should be distinguished from the variance itself.
[References] Reference [8] lists the venue as "San Diega, CA, USA"; this should be "San Diego." There are also formatting artifacts in Eq. (12) and the surrounding text (extra spaces and a stray "and").
[§5 and Conclusion] The conclusion states that scalar-valued constructions "are not universal," while Section 5 demonstrates non-universality only through the 2-area Wilson loop example. Because one counterexample is sufficient for the non-universality claim, the wording is acceptable, but the logical structure could be made explicit.

Circularity Check

1 steps flagged · score 6.0 of 10

In-sample variance reduction in Fig. 2/Table 1 is the minimized training loss; otherwise the Stein-based derivation is self-contained.

fitted input called prediction [Section 3.1, Fig. 2 caption and Table 1; also Figs. 1 and 3]
"A total of 2×10^3 samples are used, with 10^3 samples reserved for training the control variate, and the entire samples used for estimating the observables."

The network is trained by minimizing the sample mean-square residual in Eq. (10)/(13) on the 10^3 training configurations. The paper then reports the CV estimator's standard deviation from all 2×10^3 configurations, i.e., it includes the very configurations whose residuals were minimized. The variance reduction in Fig. 2 and the error bars in Table 1 are therefore the minimized training objective, not an independent estimate. The μ-shift and L2 regularization do not remove this: a flexible network can still drive training residuals down, so the reported improvement is statistically forced by the fit. Moreover, because f depends on the same configurations on which it is evaluated, the Stein zero-mean identity (2) does not by itself make the in-sample estimator unbiased. Fig. 1 and Fig.

full rationale

The paper's core derivation is self-contained: the Stein identity (2) shows that any control variate of the form ∂g − g∂S has zero mean for fixed g, so the estimator O − f is unbiased independently of how g is parametrized. The neural-network ansatz and symmetry constraints are not defined in terms of the observables being predicted. The self-citations are not load-bearing: Ref. [4] is the author's own prior work but is used as a pointer ('summarizing and extending the work in [4]'), while the universality statement in Sec. 5 rests on Ref. [3], which is not co-authored by Oh. The one genuine circular step is the numerical evaluation in Sec. 3.1: the control variate is trained by minimizing the mean-square residual on the training configurations, and the reported variance reduction is then computed on the full sample that includes those same configurations. The improvement is thus the minimized training objective rather than an out-of-sample estimate, and the unbiasedness guarantee from Eq. (2) does not apply to data-dependent f evaluated on its own training data. Fig. 1 and Fig. 3 lack explicit disjoint test sets, so the same concern attaches to the paper's strongest quantitative claims. Separately, Sec. 5's assertion that the general solution (22) has no periodic solution is stated without proof ('It is easy to verify'); this is an omitted support rather than a circular argument, but it should be checked before relying on the claimed non-universality of the scalar construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central method uses standard Stein identities and universal approximation; the only new postulates are the neural network ansatz and the unproved non-periodicity of Eq. (22). The U(1) construction relies on the exact factorization of the gauge-fixed measure. No new physical particles or forces are introduced.

free parameters (3)

Neural network weights W_i and biases b_i = Optimized by ADAM; numerical values not reported
These parameters define the control variate g and are fitted to training configurations; the reported variance reduction depends on them.
Auxiliary parameter mu = Trained jointly; not used in final estimator
Introduced in Eq. (10) to absorb the sample mean of O during training; changes the optimization trajectory.
Network hyperparameters: hidden layers, neurons, L2 strength delta, learning rate schedule = 5 hidden layers x 4 neurons (scalar), 4 x 16 (U(1)); delta not specified; LR 10^-3 to 10^-6
Architecture and regularization are chosen by hand; no sensitivity study is provided, and delta is never given a value.

assumptions (5)

domain assumption Stein identity: integral Dphi delta/delta phi (g e^{-S}) = 0 for suitable g with proper boundary conditions (Eq. 2).
Used to guarantee <f>=0 for control variates; requires boundary terms to vanish, e.g. periodic boundary conditions.
standard math Neural networks are universal approximators (Refs. [5,6]).
Justifies replacing hand-built g with a network representation.
domain assumption Maximal-tree gauge fixing with open boundary conditions makes 2D U(1) plaquette variables independent with measure exp(beta cos theta_P) (Eq. 17).
Required for the factored Wilson-loop control variate in Eq. (18) to have zero mean.
domain assumption The vector-valued Schwinger-Dyson construction is universal (Ref. [3]).
The paper uses this cited topological result as the benchmark against which the scalar construction is shown non-universal.
ad hoc to paper The general solution in Eq. (22) is correct and has no periodic solution, as asserted after Eq. (22).
The non-universality conclusion rests on this unproved verification; the paper only states 'It is easy to verify'.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Control variates with neural networks." pith.science (2026). https://pith.science/paper/LSJYTIKY

@misc{pith2026250114614,
  author       = {Pith},
  title        = {Pith review of: Control variates with neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSJYTIKY}},
  note         = {Machine review of arXiv:2501.14614}
}

abstract

The precision of lattice QCD calculations is often hindered by the stochastic noise inherent in these methods. The control variates method can provide an effective noise reduction but are typically constructed using heuristic approaches, which may be inadequate for complex theories. In this work, we introduce a neural network-based framework for parametrizing control variates, eliminating the reliance on manual guesswork. Using $1+1$ dimensional scalar field theory as a test case, we demonstrate significant variance reduction, particularly in the strong coupling regime. Furthermore, we extend this approach to gauge theories, showcasing its potential to tackle signal-to-noise problems in diverse lattice QCD applications.

Figures

Figures reproduced from arXiv: 2501.14614 by the authors.

**Figure 1.** Training histories of the small (𝜆 = 0.5) and large (𝜆 = 24.0) couplings on 20 × 20 lattice with 𝑚 2 = 0.1 are shown. The figure displays the improvement of standard deviation using control variates relative to the raw observabale (𝜎Raw/𝜎CV). The dashed line corresponds to the zero hidden layers (linear transformation), while the solid line represents the result with 5 hidden layers, each containing 4 neurons. 104 c… view at source ↗

**Figure 2.** Correlation functions with 𝑚 2 = 0.01 and 𝜆 = 0.1 on 40×10 lattice are shown. The raw result and the result with control variates are shown. A total of 2 × 103 samples are used, with 103 samples reserved for training the control variate, and the entire samples used for estimating the observables. The left plot shows the correlation functions along with their fitting curves, as described in Eq. (14), with the results… view at source ↗

**Figure 3.** Wilson loops on 8 × 8 lattice, with the coupling 𝛽 = 5.555 are shown. The left panel represents the expectation values of Wilson loops with 1𝜎 error bars, while the right plot illustrates the improvement in standard deviation achieved using control variates, expressed as the ratio (𝜎Raw/𝜎CV). Here, 𝑆1 (𝜃 𝑃) = 𝛽(1 − cos 𝜃 𝑝 ) is the action for a single plaquette. Note that since the observable in Eq. (16) is complex … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages

[4]

P.F.Bedaque,andH.Oh., Leveragingneuralcontrolvariatesforenhancedprecisioninlattice field theory, Phys. Rev. D109, (2024) 094519 [2312.08228]

arXiv 2024
[1]

Bhattacharya, S

T. Bhattacharya, S. Lawrence, and J.-S. Yoo,Control variates for lattice field theory, Phys. Rev. D109, (2024) 3, L031505 [2307.14950]

arXiv 2024
[2]

C. Stein,A bound for the error in the normal approximation to the distribution of a sum of dependentrandomvariables ,ProceedingsoftheSixthBerkeleySymposiumonMathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, 583–602

work page 1970
[3]

Lawrence,Schwinger-Dyson control variates for lattice fermions, 2404.10707

S. Lawrence,Schwinger-Dyson control variates for lattice fermions, 2404.10707

arXiv
[5]

Hornik, M

K. Hornik, M. Stinchcombe, and H. White,Multilayer feedforward networks are universal approximators, Neural Networks2, 359 (1989)

work page 1989
[6]

Cybenko,Approximation by superpositions of a sigmoidal function, Math

G. Cybenko,Approximation by superpositions of a sigmoidal function, Math. Control Signal Systems2, 303–314 (1989)

work page 1989
[7]

R. Wan, M. Zhong, H. Xiong, and Z. Zhu,Neural control variates for Monte Carlo variance reduction, In: Brefeld, U., Fromont, E., Hotho, A., Knobbe, A., Maathuis, M., Robardet, C. (eds)MachineLearningandKnowledgeDiscoveryinDatabases.ECMLPKDD2019.Lecture Notes in Computer Science(), vol 11907. Springer, Cham

work page
[8]

Kingma, and J

D. Kingma, and J. Ba,Adam: A Method for Stochastic Optimization, in International Confer- ence on Learning Representations (ICLR) (San Diega, CA, USA, 2015)

work page 2015

Show all 12 references

[9]

Bozinovski, and A

S. Bozinovski, and A. Fulgosi,The influence of pattern similarity and transfer learning on the base perceptron training (original in Croatian), Proceedings of Symposium Informatica 3-121-5, Bled (1976)

1976
[10]

S.Lawrence, H.Oh., and Y.Yamauchi,Latticescalarfieldtheoryatcomplexcoupling , Phys. Rev. D106, (2022) 114503 [2205.12303]

2022 arXiv
[11]

Detmold, G

W. Detmold, G. Kanwar, M. L. Wagman, and N. C. Warrington,Path integral contour deformations for noisy observables, Phys. Rev. D102, (2020) 014514 [2003.05914]

2020 arXiv
[12]

Lawrence, and Y

S. Lawrence, and Y. Yamauchi,Lattice scalar field theory at complex coupling, Phys. Rev. D 110, (2024) 014508 [2311.13002]. 9

2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [4]

P.F.Bedaque,andH.Oh., Leveragingneuralcontrolvariatesforenhancedprecisioninlattice field theory, Phys. Rev. D109, (2024) 094519 [2312.08228]

arXiv 2024

[2] [1]

Bhattacharya, S

T. Bhattacharya, S. Lawrence, and J.-S. Yoo,Control variates for lattice field theory, Phys. Rev. D109, (2024) 3, L031505 [2307.14950]

arXiv 2024

[3] [2]

C. Stein,A bound for the error in the normal approximation to the distribution of a sum of dependentrandomvariables ,ProceedingsoftheSixthBerkeleySymposiumonMathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, 583–602

work page 1970

[4] [3]

Lawrence,Schwinger-Dyson control variates for lattice fermions, 2404.10707

S. Lawrence,Schwinger-Dyson control variates for lattice fermions, 2404.10707

arXiv

[5] [5]

Hornik, M

K. Hornik, M. Stinchcombe, and H. White,Multilayer feedforward networks are universal approximators, Neural Networks2, 359 (1989)

work page 1989

[6] [6]

Cybenko,Approximation by superpositions of a sigmoidal function, Math

G. Cybenko,Approximation by superpositions of a sigmoidal function, Math. Control Signal Systems2, 303–314 (1989)

work page 1989

[7] [7]

R. Wan, M. Zhong, H. Xiong, and Z. Zhu,Neural control variates for Monte Carlo variance reduction, In: Brefeld, U., Fromont, E., Hotho, A., Knobbe, A., Maathuis, M., Robardet, C. (eds)MachineLearningandKnowledgeDiscoveryinDatabases.ECMLPKDD2019.Lecture Notes in Computer Science(), vol 11907. Springer, Cham

work page

[8] [8]

Kingma, and J

D. Kingma, and J. Ba,Adam: A Method for Stochastic Optimization, in International Confer- ence on Learning Representations (ICLR) (San Diega, CA, USA, 2015)

work page 2015