Pith. sign in

REVIEW 3 major objections 4 minor 20 references

HMC and gradient flow with machine-learned classically perfect fixed-point actions

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read This paper claims that a machine-learned fixed-point lattice action for SU(3) gauge theory removes tree-level artifacts from the gradient flow, so continuum results emerge from coarse lattices with only mild O(a^2) corrections.

desk verdict A promising progress report on machine-learned FP actions for SU(3), whose scaling test is worth taking seriously, but whose central C_FP=1 argument is asserted rather than derived. read the letter →

arxiv 2502.03315 v2 pith:SIE4PTMS submitted 2025-02-05 hep-lat

classification hep-lat MSC 81T2581T13 PACS 11.15.Ha12.38.Gc
keywords fixed-pointactiongradientflowlatticegaugetheorySU(3)gauge-equivariantneuralnetworkhybridMonteCarloartifactsscalesetting
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

This paper tries to establish that a classically perfect fixed-point (FP) lattice action for SU(3) gauge theory can be parameterized by a gauge-equivariant convolutional neural network, simulated with Hybrid Monte Carlo, and used to define a gradient flow free of tree-level lattice artifacts. The payoff would be a practical route to continuum physics from significantly coarser lattices than the Wilson action allows, reducing critical slowing down and topological freezing. The scaling evidence is a measured ratio of gradient-flow scales, $t_{0.3}/w^2_{0.3}$, whose approach to the continuum is mild and compatible with an $O(a^2)$ correction for lattice spacings $a \lesssim 0.13$ fm. A sympathetic reader would take the central assertion to be that both the exact FP action and its learned approximation remove the dominant discretization errors.

What carries the argument

The central object is the fixed-point action defined by the FP equation $A_{\mathrm{FP}}[V] = \min_U (A_{\mathrm{FP}}[U] + T[U,V])$, where $T$ is an RG blocking kernel. The machinery that carries the argument is a lattice gauge-covariant convolutional neural network (L-CNN), a gauge-equivariant network that builds complicated Wilson loops through convolutions with parallel transports and preserves exact gauge invariance, and whose backpropagation yields the action derivatives needed for HMC. For the gradient-flow claim, the key mechanism is that repeated RG blocking drives the propagator to have poles at $(p + 2\pi l)^2$ for all integers $l$, so the tree-level lattice coefficient $C(a^2/t)$ becomes identically 1.

What would settle it

Directly substitute the L-CNN parameterized action's propagator into the tree-level coefficient formula (Eq. (7)/(8)) and evaluate $C(a^2/t)$; if it deviates from 1 by more than a few percent for $t/a^2 \sim 1$, the simulated action is not classically perfect in the flow, independent of any scaling fit.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the exact fixed-point action has no tree-level artifacts in the gauge-field gradient flow: because iterated RG transformations extend the effective momentum range to $\pm\infty$, the FP propagator has the continuum dispersion relation, so the tree-level coefficient in Eq. (8) satisfies $C_{\mathrm{FP}}(a^2/t) = 1$ at any finite lattice spacing. For the parameterized action actually simulated, the paper reports that discretization effects are very mild, with an $O(a^2)$ extrapolation of data at $a \lesssim 0.13$ fm sufficient for a reliable continuum limit, in contrast to the Wilson action. The parameterization itself is done with an L-CNN trained on FP action values and their derivatives, and the paper shows it outperforms earlier traced-loop and fat-link parameterizations.

Load-bearing premise

The whole argument rests on the assumption that the trained neural network approximates the true fixed-point action closely enough on the simulated ensembles that whatever error the approximation introduces is too small to affect the measured scaling.

Editorial extensions

If this is right

  • Continuum quantities such as $t_0/w^2$ could be extracted from lattices with $a \simeq 0.11\text{--}0.13$ fm without the large systematic errors of Wilson-type extrapolations.
  • Gradient-flow scale setting and a flow-based $\beta$-function could be computed with much smaller lattice artifacts when flow action, observable, and simulation action are all the FP action.
  • The L-CNN parameterization makes the FP approach practical for HMC simulation, since exact derivatives are available through backpropagation and standard symplectic integrators apply.
  • If the small-artifact behavior persists in full QCD, the same strategy would mitigate critical slowing down and topological freezing in the approach to the continuum.

Reading between the lines

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

  • A testable extension not in the paper is to compute the tree-level coefficient $C(a^2/t)$ for the learned L-CNN action directly, not just for the exact FP propagator, to see how much parameterization error re-introduces artifacts.
  • The paper's comparison of parameterizations reports no error bars on the training data; an independent check would be to generate several HMC streams from different random starts and compare the spread of $t_{0.3}/w^2_{0.3}$ at fixed $\beta_{\mathrm{FP}}$.
  • One could extend the same learned-action logic to the Dirac operator: if a classically perfect FP fermion action can be parameterized in a similar gauge-equivariant network, coarser lattices might also control chiral symmetry breaking and index-theorem effects.
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.

Referee Report

3 major / 4 minor

Summary. The paper reports on the construction and first use of a machine-learned parameterization of the fixed-point (FP) action for four-dimensional SU(3) gauge theory. The FP action is represented by a gauge-equivariant convolutional network (L-CNN), trained on FP action values and derivatives obtained from the FP equation. The authors show that this parameterization reproduces FP data better than previous loop-based or fat-link parameterizations, and they simulate the resulting action with HMC, checking the Hamiltonian constraint. They then argue that the exact FP action leads to a classically perfect gradient flow with tree-level coefficient C_FP(a^2/t)=1, and they present a scaling study of the ratio t_{0.3}/w^2_{0.3} on coarse lattices, reporting much milder discretization effects than for the Wilson action. The paper is a proceedings contribution presenting initial results rather than a complete analysis.

Significance. If the central claims hold, the work opens a practical route to continuum physics from coarser lattices, potentially reducing critical slowing down and topological freezing. The L-CNN parameterization with exact derivatives through backpropagation is a genuine technical advance over earlier FP-action parameterizations, and the explicit HMC implementation with a Hamiltonian-constraint check is a valuable step. The scaling results, while preliminary, are suggestive and would be significant if backed by controlled statistics. However, the paper's central theoretical claim about the exact FP gradient flow is not proven, and the scaling evidence is presented without the ensemble-level detail needed to support the stated conclusion.

major comments (3)
  1. [Section 4, Eq. (8)] The claim that the exact FP action gives C_FP(a^2/t)=1 is asserted rather than derived. Equation (8) is an integral over the first Brillouin zone of an integrand built from the lattice quadratic form A, with no sum over momenta p+2πl. Even if the FP propagator has poles at (p+2πl)^2 for all integer vectors l, that does not by itself imply that the first-Brillouin-zone integral in Eq. (8) equals the continuum all-momentum integral, because A_f and A_e in that equation are still functions of the Brillouin-zone momentum p. A concrete counterexample is any quadratic form equal to p^2 on the Brillouin zone, for which the integral in Eq. (8) would differ from the continuum result by O(a^2/t) contributions. The authors need to provide an explicit free-field solution of the FP equation (3) or a direct calculation of Eq. (8) for the exact FP action before the statement 'This results in C_FP(a^2/t)=1' can be accepted. This point is load-bearing because Section 5 attributes all residual artifacts to 'an imperfect parameterization or to quantum effects', which presupposes that the exact FP flow has no tree-level artifacts.
  2. [Section 5, Fig. 8] The scaling study is presented without the information needed to assess its reliability. The right panel of Fig. 8 shows three FP lattice sizes (10^4, 14^4, 16^4), but no table is given with the corresponding beta values, statistics, autocorrelation times, or the number of configurations used in the gradient-flow measurement. The left panel shows a single example ensemble without stating how many configurations are averaged. With only three lattice sizes and no statistical errors or finite-volume checks, the claim that 'an O(a^2) extrapolation of the data with a ≲ 0.13 fm is sufficient for a reliable continuum limit' is not supported. The authors should provide ensemble details and a quantitative estimate of the parameterization error on the actual HMC ensembles used in the scaling test.
  3. [Section 5, discussion of residual artifacts] The sentence attributing any remaining lattice artifacts to 'an imperfect parameterization or to quantum effects' depends critically on the unproved result C_FP=1 from Section 4. If the exact FP action is not shown to yield a tree-level perfect gradient flow, then the observed mild scaling violations could equally be caused by flow-action artifacts. The attribution should either be removed or made conditional on a completed free-field derivation. This is not a mere wording issue; it changes the interpretation of the scaling data.
minor comments (4)
  1. [Figure 5] The comparison of parameterization errors for the L-CNN and previous constructions would be much more informative if the figure included uncertainties or at least the number of configurations and the range of beta values used for each ensemble.
  2. [Section 3, around Eq. (2)] The notation N_beta_mu[U] is introduced but not defined in detail; a short explanation of how this term fixes the normalization of the partition function would help the reader.
  3. [Figure 8] The plot uses t_{0.3}/w^2_{0.3} where both scales are extracted from the same flowed observables; the statistical correlation between them should be discussed when errors are reported.
  4. [Section 4, Figure 7] It would be useful to state explicitly that the hard-cutoff and Wilson curves in Fig. 7 are obtained from known analytic expressions, and to give the corresponding formulas or references in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned FP action is validated against the FP equation and the scaling test is a genuine simulation, not a fitted prediction.

full rationale

The derivation chain is not circular. The L-CNN parameterization is trained to reproduce the FP action values given by the RHS of Eq. (3), and Fig. 5 checks the parameterization against those same numerical-minimization targets; this is a fit-quality test, not the prediction of an independent quantity from a fitted input. The HMC evolution and the scaling ratio t0.3/w2_0.3 in Fig. 8 come from actual Monte Carlo simulations with the parameterized FP action, and no parameter of the scaling extrapolation is adjusted to force agreement with the Wilson comparison. Self-citations [2,3,5] provide the L-CNN architecture and earlier parameterizations, but they are not load-bearing for the new numerical results. The Section 4 claim that C_FP(a^2/t)=1 is not itself circular, but it is under-derived: Eq. (8) remains an integral over the first Brillouin zone, and the statement that poles at (p+2πl)^2 extend the momentum range to ±∞ does not by itself turn that integral into the continuum integral. That is a mathematical-support gap, not a reduction of the output to the input; the scaling result is not fitted to the C_FP=1 claim.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The ledger shows the construction rests on the perfect-action framework from [4] and on the L-CNN machinery from [5]; the one genuinely unsupported input is the pole-structure assertion for the FP propagator, which carries the tree-level GF claim. The only fitted numbers internal to this paper are the L-CNN weights and the hand-chosen blocking parameters, neither of which is released.

free parameters (2)
  • L-CNN weights (3 layers, 12/24/24 channels, kernel sizes 2,2,1) = Trained on FP-equation action values and derivatives; values not published
    The simulated FP action is this network; all numerical results depend on the trained weights, which are fit to training data, not derived.
  • Blocking kernel parameters s0, s_pl, s_d, s_hd, kappa in Eq. (2) = Not reported
    These are chosen by hand to optimize locality and define the RGT that generates the target FP action; different choices yield different FP actions and hence different scaling behavior.
assumptions (5)
  • domain assumption The RGT defined by Eq. (2) yields a well-defined, local fixed-point action solving Eq. (3).
    The existence, uniqueness and locality of the FP action are taken from the perfect-action literature [4]; the paper does not derive them.
  • domain assumption For asymptotically free theories, actions on the critical surface flow to the FP under repeated RGTs, and FP actions suppress lattice artifacts at weak coupling.
    This RG picture (Fig. 3) motivates the whole approach and is cited from [4].
  • domain assumption L-CNN layers preserve exact gauge covariance and backpropagation yields exact derivatives of the action.
    Needed for training on derivatives and for HMC forces; attributed to [5], not re-proven.
  • ad hoc to paper The exact FP propagator has poles at (p+2πl)^2 for all integers l, so its dispersion is continuum-like on the lattice.
    Asserted in Section 4 without derivation; it is the critical step from Eq. (8) to C_FP=1.
  • standard math Tree-level discretization artifacts of gradient flow are captured by C(a^2/t) in Eq. (7) from perturbation theory [16].
    The GF artifact analysis uses this formula; accepted perturbative result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HMC and gradient flow with machine-learned classically perfect fixed-point actions." pith.science (2026). https://pith.science/paper/SIE4PTMS

@misc{pith2026250203315,
  author       = {Pith},
  title        = {Pith review of: HMC and gradient flow with machine-learned classically perfect fixed-point actions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIE4PTMS}},
  note         = {Machine review of arXiv:2502.03315}
}
read the original abstract

Fixed-point (FP) lattice actions are classically perfect, i.e., they have continuum classical properties unaffected by discretization effects and are expected to have suppressed lattice artifacts at weak coupling. Therefore they provide a possible way to extract continuum physics with coarser lattices, allowing to circumvent problems with critical slowing down and topological freezing towards the continuum limit. We use machine-learning methods to parameterize a FP action for four-dimensional SU(3) gauge theory using lattice gauge-covariant convolutional neural networks. The large operator space allows us to find superior parameterizations compared to previous studies and we show how such actions can be efficiently simulated with the Hybrid Monte Carlo algorithm. Furthermore, we argue that FP lattice actions can be used to define a classically perfect gradient flow without any lattice artifacts at tree level. We present initial results for scaling of the gradient flow with the FP action.

Figures

Figures reproduced from arXiv: 2502.03315 by the authors.

Figure 1
Figure 1. Illustration of the continuum limit towards the left with lattice spacings 𝑎/𝜉 < 𝑎′ /𝜉 < 𝑎′′/𝜉 at fixed physical scale 𝜉 (upper row), and renormalization group transformation (RGT) steps yielding effective couplings {𝑐𝛼} → {𝑐 ′ 𝛼 } → {𝑐 ′′ 𝛼 } (lower row). In these proceedings we describe the status of our project to construct an RG-improved lattice action which is classically perfect, i.e., which has no lattice dis… view at source ↗
Figure 2
Figure 2. Illustration of the RGT blocking employed here. with weight 𝑠𝑝𝑙, diagonal staples with weight 𝑠𝑑, and hyper-diagonal staples with weight 𝑠ℎ𝑑, as illustrated in the top row of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the RG flow for asymptotically free gauge theories. the resulting effective gauge action. For asymptotically free theories, there is one marginally relevant opera￾tor parameterized by 𝛽 and hence the critical surface, where 𝜉/𝑎 = ∞, is located at 𝛽 = ∞. Under repeated RGT steps, any lattice ac￾tion defined on the criti￾cal surface is driven to the Fixed Point (FP), cf [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Example of an L-CNN that processes lattice gauge data while preserving gauge symmetry through specialized convolutional (L-Conv), bilinear (L-Bilin), and activation layers [5]. FP Eq. (3) implicitly provides data for each coarse configuration 𝑉, finding the explicit fo…
Figure 5
Figure 5. Figure 5: Comparison of the new parameterization of the FP action (L-CNN) and previous parameterizations using either powers of traced loops up to length six (IIIc-4) or powers of traced plaquettes containing APE￾smeared fat links (APE 431, APE 444), from Ref. [2]. While an accu…
Figure 6
Figure 6. Figure 6: HMC evolution of the plaquette value for the FP action at 𝛽FP = 2.95 corresponding to 𝑎 ≃ 0.11 fm. at a range of 𝛽 values, to make quantitative com￾parison with other lattice actions and the con￾tinuum limit of appropriately chosen quantities. Given that the derivative…
Figure 7
Figure 7. Figure 7: The tree-level coefficient 𝐶(𝑎 2 /𝑡) in the gauge field gradient flow for different regulators, with a hard momentum cutoff (left) or Wilson lattice action (right). The exact FP propagator extends the momentum range to ±∞ even at finite lattice spacing, resulting in 𝐶 …
Figure 8
Figure 8. Figure 8: (Left) Example of gradient flow on one FP gauge ensemble, the light blue curves are from individual configurations, the black curve is the ensemble average. (Right) Scaling test of FP simulation results for the combination of scales 𝑡0.3/𝑤 2 0.3 . lattice spacing. Disc…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 5 canonical work pages

  1. [1]

    Schaefer, R

    ALPHA collaboration, S. Schaefer, R. Sommer and F. Virotta,Critical slowing down and error analysis in lattice QCD simulations, Nucl. Phys. B845 (2011) 93, [1009.5228]

  2. [2]

    Holland, A

    K. Holland, A. Ipp, D. I. Müller and U. Wenger,Machine learning a fixed point action for SU(3) gauge theory with a gauge equivariant convolutional neural network,Phys. Rev. D110 (2024) 074502, [2401.06481]

  3. [3]

    Holland, A

    K. Holland, A. Ipp, D. I. Müller and U. Wenger,Fixed point actions from convolutional neural networks,PoSLATTICE2023(2024) 038, [2311.17816]

  4. [4]

    Hasenfratz and F

    P. Hasenfratz and F. Niedermayer,Perfect lattice action for asymptotically free theories, Nucl. Phys. B414 (1994) 785, [hep-lat/9308004]

  5. [5]

    Favoni, A

    M. Favoni, A. Ipp, D. I. Müller and D. Schuh,Lattice Gauge Equivariant Convolutional Neural Networks,Phys. Rev. Lett.128 (2022) 032003, [2012.12901]

  6. [6]

    Blatter and F

    M. Blatter and F. Niedermayer,New fixed point action for SU(3) lattice gauge theory, Nucl. Phys. B482 (1996) 286, [hep-lat/9605017]

  7. [7]

    Niedermayer, P

    F. Niedermayer, P. Rufenacht and U. Wenger,Fixed point gauge actions with fat links: Scaling and glueballs, Nucl. Phys. B597 (2001) 413, [hep-lat/0007007]

  8. [8]

    Duane, A

    S. Duane, A. D. Kennedy, B. J. Pendleton and D. Roweth,Hybrid Monte Carlo, Phys. Lett. B 195 (1987) 216

Show all 20 references
  1. [9]

    S. A. Gottlieb, W. Liu, D. Toussaint, R. L. Renken and R. L. Sugar,Hybrid Molecular Dynamics Algorithms for the Numerical Simulation of Quantum Chromodynamics,Phys. Rev. D35(1987) 2531

  2. [10]

    Takaishi and P

    T. Takaishi and P. de Forcrand,Testing and tuning new symplectic integrators for hybrid Monte Carlo algorithm in lattice QCD, Phys. Rev. E73 (2006) 036706, [hep-lat/0505020]

  3. [11]

    I. P. Omelyan, I. M. Mryglod and R. Folk,Symplectic analytically integrable decomposition algorithms: classification, derivation, and application to molecular dynamics, quantum and celestial mechanics simulations, Comput. Phys. Commun.151 (2003) 272

  4. [12]

    Lüscher,Properties and uses of the Wilson flow in lattice QCD, JHEP08(2010) 071, [1006.4518]

    M. Lüscher,Properties and uses of the Wilson flow in lattice QCD, JHEP08(2010) 071, [1006.4518]

  5. [13]

    Luscher and P

    M. Luscher and P. Weisz,Perturbative analysis of the gradient flow in non-abelian gauge theories, JHEP02(2011) 051, [1101.0963]

  6. [14]

    9 HMC and gradient flow with machine-learned classically perfect FP actions U

    Z.Fodor,K.Holland,J.Kuti,D.NogradiandC.H.Wong, Anewmethodforthebetafunction in the chiral symmetry broken phase, EPJ Web Conf.175 (2018) 08027, [1711.04833]. 9 HMC and gradient flow with machine-learned classically perfect FP actions U. Wenger

  7. [15]

    Hasenfratz and O

    A. Hasenfratz and O. Witzel,Continuous renormalization group𝛽 function from lattice simulations,Phys. Rev. D101 (2020) 034514, [1910.06408]

  8. [16]

    Fodor, K

    Z. Fodor, K. Holland, J. Kuti, S. Mondal, D. Nogradi and C. H. Wong,The lattice gradient flow at tree-level and its improvement, JHEP09(2014) 018, [1406.0827]

  9. [17]

    T. A. DeGrand, A. Hasenfratz, P. Hasenfratz and F. Niedermayer,The Classically perfect fixed point action for SU(3) gauge theory,Nucl. Phys. B454 (1995) 587, [hep-lat/9506030]

  10. [18]

    Hasenfratz,Lattice QCD without tuning, mixing and current renormalization, Nucl

    P. Hasenfratz,Lattice QCD without tuning, mixing and current renormalization, Nucl. Phys. B 525 (1998) 401, [hep-lat/9802007]

  11. [19]

    Hasenfratz, V

    P. Hasenfratz, V. Laliena and F. Niedermayer,The Index theorem in QCD with a finite cutoff, Phys. Lett. B427 (1998) 125, [hep-lat/9801021]

  12. [20]

    Borsányi, S

    BMWcollaboration, S. Borsányi, S. Dürr, Z. Fodor, C. Hoelbling, S. D. Katz, S. Krieg et al., High-precision scale setting in lattice QCD, JHEP09(2012) 010, [1203.4469]. 10

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.