pith. sign in

arxiv: 2605.28929 · v1 · pith:SMTMC4MTnew · submitted 2026-05-27 · ❄️ cond-mat.str-el · cond-mat.dis-nn· cond-mat.stat-mech· hep-lat· hep-th

Improving CFT Operators Using Machine Learning

Lior Oppenheim , Snir Gazit , Zohar Ringel This is my paper

Pith reviewed 2026-06-29 09:37 UTC · model grok-4.3

classification ❄️ cond-mat.str-el cond-mat.dis-nncond-mat.stat-mechhep-lathep-th
keywords machine learningCFT operatorslattice improvementcritical phenomenaIsing modelPotts modelfinite-size effectsscaling dimensions
0
0 comments X

The pith

Machine learning constructs improved lattice operators with better overlap to continuum CFT primaries.

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

The paper proposes a data-driven method to refine lattice operators so they overlap more strongly with the primary operators of the corresponding continuum conformal field theory. This targets operator-specific finite-size effects that remain after action improvement. The approach is demonstrated on leading spin and energy operators in the Ising model, the q=3 Potts model, and the dilute q=3 Potts model. In each case the optimized operators produce smaller corrections to scaling and more accurate scaling-dimension estimates than conventional lattice choices. Readers would care because such operator improvement directly increases the precision with which conformal data can be read out from lattice simulations of critical points.

Core claim

We identify improved lattice representations of leading spin and energy operators in three two-dimensional critical systems: the Ising model, the q = 3 Potts model, and the dilute q = 3 Potts model. In all cases, the resulting operators exhibit reduced corrections to scaling and yield more accurate estimates of scaling dimensions compared to conventional lattice choices.

What carries the argument

A data-driven optimization procedure that constructs lattice operators with enhanced overlap with the corresponding primary operators of the continuum conformal field theory.

If this is right

  • The optimized operators reduce corrections to scaling in the Ising, q=3 Potts, and dilute q=3 Potts models.
  • Scaling-dimension estimates extracted from lattice data become more accurate than those obtained with standard operator choices.
  • Operator improvement addresses a class of finite-size effects distinct from those suppressed by action improvement.
  • The method supplies a systematic route to better lattice representations of continuum fields in two-dimensional critical systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same optimization could be applied to other observables such as three-point couplings if the overlap criterion is extended accordingly.
  • Combining the learned operators with existing action-improvement techniques might produce additive gains in overall accuracy.
  • If the procedure generalizes, it offers a route to operator improvement in models where no exact continuum operator expression is known.

Load-bearing premise

The optimization procedure increases genuine overlap with continuum primary operators rather than fitting to the same finite-size artifacts it is meant to suppress.

What would settle it

If the learned operators produce the same magnitude of corrections to scaling as conventional operators when applied to larger lattices or different boundary conditions, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2605.28929 by Lior Oppenheim, Snir Gazit, Zohar Ringel.

Figure 1
Figure 1. Figure 1: FIG. 1. Finite-size scaling of two-point correlation functions using Sandvik’s method (see eq. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Low-order structure of the fitted surrogates within [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows this quantity for the Ising model at L = 24. For the energy operator ϵ, the ratio develops a broad plateau at intermediate and long distances with a value of approximately 3.5, indicating that the learned estima￾tor produces a substantially stronger long-distance sig￾nal than the naive lattice operator. By contrast, for the spin operator σ, the corresponding ratio remains close to unity, around 1.1, … view at source ↗
read the original abstract

Finite-size effects limit the accuracy with which conformal data can be extracted from lattice simulations of critical systems. While action improvement suppresses some corrections to scaling, it does not address operator-dependent effects arising from imperfect lattice representations of continuum conformal fields. In this work, we propose a data-driven method for improving lattice operators themselves, constructing estimators with enhanced overlap with the corresponding primary operators of the continuum conformal field theory. We identify improved lattice representations of leading spin and energy operators in three two-dimensional critical systems: the Ising model, the q = 3 Potts model, and the dilute q = 3 Potts model. In all cases, the resulting operators exhibit reduced corrections to scaling and yield more accurate estimates of scaling dimensions compared to conventional lattice choices. The code and analysis workflows used to produce these results are made available in an accompanying GitHub repository.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a data-driven machine learning procedure to optimize lattice representations of the leading spin and energy operators in three 2D critical models (Ising, q=3 Potts, dilute q=3 Potts). The optimized operators are claimed to have larger overlap with the corresponding continuum CFT primaries, thereby suppressing operator-dependent corrections to scaling and yielding more accurate estimates of scaling dimensions than standard lattice choices. Reproducible code and workflows are provided via GitHub.

Significance. If the central claim holds and the procedure demonstrably increases overlap with continuum primaries rather than fitting finite-size artifacts, the work would offer a practical, generalizable tool for improving the extraction of conformal data from lattice Monte Carlo simulations, complementing existing action-improvement techniques. The public release of code is a clear strength that facilitates independent verification.

major comments (2)
  1. [§3 (Method) and §4 (Results)] The training procedure is performed exclusively on finite-L Monte Carlo ensembles. No diagnostic is presented that independently confirms increased overlap with the continuum primary (e.g., via three-point function matrix elements or an L→∞ extrapolation of the improvement itself). This leaves open the possibility that the loss is minimized by fitting L-dependent mixing or statistical fluctuations rather than by suppressing operator-dependent corrections to scaling.
  2. [§4.1–4.3 and associated figures/tables] In the reported scaling-dimension estimates (e.g., for the energy operator in the Ising case), the reduction in corrections to scaling is shown, but the manuscript does not quantify the change in operator overlap or demonstrate that the improvement survives after subtracting the leading irrelevant operator contribution. Without this, the claim that the operators are closer to continuum primaries remains under-supported.
minor comments (2)
  1. [§3] Notation for the optimized operators and the precise definition of the loss function should be introduced with an equation in §3 to allow readers to reproduce the optimization step without consulting the repository.
  2. [Figures 2–5] Figure captions should explicitly state the system sizes used for training versus validation and whether error bars include both statistical and systematic uncertainties from the ML procedure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments on our manuscript. We address each major comment below and will revise the manuscript accordingly to strengthen the supporting evidence for our claims.

read point-by-point responses

  1. Referee: [§3 (Method) and §4 (Results)] The training procedure is performed exclusively on finite-L Monte Carlo ensembles. No diagnostic is presented that independently confirms increased overlap with the continuum primary (e.g., via three-point function matrix elements or an L→∞ extrapolation of the improvement itself). This leaves open the possibility that the loss is minimized by fitting L-dependent mixing or statistical fluctuations rather than by suppressing operator-dependent corrections to scaling.

    Authors: We agree that an independent diagnostic would provide stronger confirmation. The loss is explicitly constructed from the L-dependence of two-point correlators to penalize operator-dependent corrections to scaling. In the revised manuscript we will add an explicit L→∞ extrapolation of the fitted correction amplitudes (before and after optimization) to demonstrate that the improvement persists in the continuum limit rather than arising from finite-L artifacts. We will also include a brief discussion of why the chosen loss targets continuum overlap rather than model-specific mixing. revision: yes

  2. Referee: [§4.1–4.3 and associated figures/tables] In the reported scaling-dimension estimates (e.g., for the energy operator in the Ising case), the reduction in corrections to scaling is shown, but the manuscript does not quantify the change in operator overlap or demonstrate that the improvement survives after subtracting the leading irrelevant operator contribution. Without this, the claim that the operators are closer to continuum primaries remains under-supported.

    Authors: The current manuscript presents the improvement through more accurate scaling-dimension estimates, which serve as indirect evidence of better overlap. We acknowledge that a direct quantification of the overlap change and an explicit subtraction of the leading irrelevant operator would make the claim more robust. In the revision we will add a table reporting the fitted amplitudes of the leading corrections for both standard and optimized operators, and we will show scaling-dimension fits performed after explicitly subtracting the leading irrelevant contribution to confirm that the improvement remains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained against benchmarks

full rationale

The paper describes a data-driven ML optimization to construct improved lattice operators for 2D critical models, claiming reduced finite-size corrections and more accurate scaling dimensions versus conventional choices. No equations, self-citations, or load-bearing steps are visible that reduce the central result to a fit or definition by construction. The improvement is presented as an empirical outcome benchmarked externally (scaling dimension accuracy, GitHub code), satisfying the criterion for a self-contained derivation with independent content. No patterns from the enumerated list apply.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the method is described at a high level without equations or implementation details.

pith-pipeline@v0.9.1-grok · 5682 in / 1090 out tokens · 28276 ms · 2026-06-29T09:37:26.516336+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 4 canonical work pages

  1. [1]

    neu- ral

    The corrections to scaling ex- ponents include a non-analytic exponentω= 4/5 from the leading irrelevant operator, and analytic corrections withω= 2,4, . . .. Thedilute Pottsmodel is a variant of the standard q= 3 Potts model, obtained by adding a site-dilution (chemical-potential) term to the Hamiltonian: −βH=K X ⟨i,j⟩ δσi,σj (1−δ σi,0) +D X i δσi,0 whic...

    work page arXiv

  2. [2]

    S. e. a. Navas (Particle Data Group Collaboration), Re- view of particle physics, Phys. Rev. D110, 350 (2024)

    2024

  3. [3]

    J. L. Cardy, ed.,Finite-Size Scaling, Current Physics - Sources and Comments, Vol. 2 (North-Holland (Else- vier), Amsterdam; Oxford; New York; Tokyo, 1988)

    1988

  4. [4]

    Nahum, J

    A. Nahum, J. T. Chalker, P. Serna, M. Ortu˜ no, and A. M. Somoza, Deconfined quantum criticality, scaling violations, and classical loop models, Phys. Rev. X5, 041048 (2015)

    2015

  5. [5]

    Hasenfratz and F

    P. Hasenfratz and F. Niedermayer, Perfect lattice action for asymptotically free theories, Nuclear Physics B414, 785 (1994)

    1994

  6. [6]

    Campostrini, A

    M. Campostrini, A. Pelissetto, P. Rossi, and E. Vi- cari, Improved high-temperature expansion and critical equation of state of three-dimensional ising-like systems, Phys. Rev. E60, 3526 (1999)

    1999

  7. [7]

    Hasenbusch, K

    M. Hasenbusch, K. Pinn, and S. Vinti, Critical expo- nents of the three-dimensional ising universality class from finite-size scaling with standard and improved ac- tions, Phys. Rev. B59, 11471 (1999)

    1999

  8. [8]

    Segall, S

    G. Segall, S. Gazit, and D. Podolsky, Improved actions using the renormalization group, Phys. Rev. B111, 184413 (2025)

    2025

  9. [9]

    Hasenbusch, Eliminating leading and subleading cor- rections to scaling in the three-dimensional xy universal- ity class, arXiv:2507.19265 [cond-mat.stat-mech] (2025)

    M. Hasenbusch, Eliminating leading and subleading cor- rections to scaling in the three-dimensional xy universal- ity class, arXiv:2507.19265 [cond-mat.stat-mech] (2025)

    work page arXiv 2025

  10. [10]

    L¨ uscher, Advanced lattice qcd (1998)

    M. L¨ uscher, Advanced lattice qcd (1998)

    1998

  11. [11]

    Heatlie, C

    G. Heatlie, C. Sachrajda, G. Martinelli, C. Pittori, and G. Rossi, The improvement of hadronic matrix elements in lattice qcd, Nuclear Physics B352, 266 (1991)

    1991

  12. [12]

    A. W. Sandvik, Using operator covariance to disentangle scaling dimensions in lattice models (2025)

    2025

  13. [13]

    Koch-Janusz and Z

    M. Koch-Janusz and Z. Ringel, Mutual information, neural networks and the renormalization group, Nature Physics14, 578 (2018)

    2018

  14. [14]

    Gordon, A

    A. Gordon, A. Banerjee, M. Koch-Janusz, and Z. Ringel, Relevance in the Renormalization Group and in Informa- tion Theory, Phys. Rev. Lett.126, 240601 (2021)

    2021

  15. [15]

    Oppenheim, M

    L. Oppenheim, M. Koch-Janusz, S. Gazit, and Z. Ringel, Machine learning the operator content of the critical self- dual ising-higgs lattice gauge theory, Phys. Rev. Res.6, 043322 (2024)

    2024

  16. [16]

    Symanzik, Continuum limit and improved action in lattice theories: (i)

    K. Symanzik, Continuum limit and improved action in lattice theories: (i). principles and phi 4 theory, Nuclear Physics B226, 187 (1983)

    1983

  17. [17]

    Ramos and S

    A. Ramos and S. Sint, Symanzik improvement of the gra- dient flow in lattice gauge theories, The European Phys- ical Journal C76, 15 (2016)

    2016

  18. [18]

    J. L. Cardy, Operator content of two-dimensional con- formally invariant theories, Nuclear Physics B270, 186 (1986)

    1986

  19. [19]

    X. Qian, Y. Deng, and H. W. J. Bl¨ ote, Dilute potts model in two dimensions, Phys. Rev. E72, 056132 (2005)

    2005

  20. [20]

    X. Qian, Y. Deng, Y. Liu, W. Guo, and H. W. J. Bl¨ ote, Equivalent-neighbor potts models in two dimen- sions, Phys. Rev. E94, 052103 (2016)

    2016

  21. [21]

    A. W. Sandvik and B. Zhao, Consistent scaling expo- nents at the deconfined quantum-critical point, Chinese Physics Letters37, 057502 (2020)

    2020

  22. [22]

    D. E. G¨ okmen, Z. Ringel, S. D. Huber, and M. Koch- Janusz, Symmetries and phase diagrams with real-space mutual information neural estimation, Phys. Rev. E104, 064106 (2021)

    2021

  23. [23]

    D. E. G¨ okmen, Z. Ringel, S. D. Huber, and M. Koch- Janusz, Statistical Physics through the Lens of Real- Space Mutual Information, Phys. Rev. Lett.127, 240603 (2021)

    2021

  24. [24]

    D. E. G¨ okmen, S. Biswas, S. D. Huber, Z. Ringel, F. Flicker, and M. Koch-Janusz, Compression theory for inhomogeneous systems, arXiv:2301.11934 [cond- mat.stat-mech] (2023)

    work page arXiv 2023

  25. [25]

    3 FC×w, ReLU→d

    D. Tong, Comments on symmetric mass generation in 2d and 4d, Journal of High Energy Physics2022, 10.1007/jhep07(2022)001 (2022). 6 I. APPENDIX A - DET AILS OF THE RSMI-NE PROCEDURE For each model, we generate Monte Carlo snapshots at criticality for a fixed linear sizeL. From every snapshot we extract a single training example (V, E). The visible regionVi...

    work page doi:10.1007/jhep07(2022)001 2022