REVIEW 34 references
Neural networks reconstruct full two-dimensional CFT torus and annulus partition functions from modular or open/closed crossing, a spectral gap, and a single interior anchor.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · grok-4.5
2026-07-31 05:51 UTC pith:J2RVOPZY
load-bearing objection Solid numerical pipeline extending their neural bootstrap to modular/Cardy constraints; the selection-principle claim is weaker than the abstract once you inspect the ansatz and the bimodal WZW runs.
Neural Spectral Bias and Conformal Correlators II: Modular and Annulus Bootstrap
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
When modular S-invariance and the Cardy condition are cast as four-point crossing for twist fields or defect-changing operators, an anchored neural bootstrap—crossing loss plus a gap-weighted ansatz plus one interior anchor—reconstructs the full diagonal reduced correlator, and therefore the corresponding torus or annulus partition function, with high accuracy. Spectral bias in the lazy-training regime systematically selects the physical crossing-symmetric configuration among the many functions allowed by the same sparse input.
What carries the argument
Anchored neural bootstrap for modular and annulus crossing: the partition function is rewritten as a reduced four-point correlator on a line; a gap-dependent prefactor is factored out; a small feed-forward network learns the remainder under a crossing loss and a single anchor value; spectral bias in lazy training supplies the selection principle that picks the physical solution.
Load-bearing premise
The method assumes that neural spectral bias reliably prefers the partition function of a real consistent CFT over the large space of other smooth functions that obey the same crossing equation, gap, and single anchor.
What would settle it
Train the same architecture on a known rational CFT with fixed gap and anchor; if the ensemble mean of the reconstructed reduced correlator systematically deviates from the exact modular-invariant answer by more than the reported sub-percent errors (or lands on a stable wrong peak even after early-stopping filters), the selection claim fails.
If this is right
- Full diagonal torus and annulus partition functions can be rebuilt from modular or Cardy crossing plus minimal spectral input without positivity or unitarity.
- The same sparse protocol applies to non-compact theories, including free non-compact bosons and Liouville ZZ/FZZ annuli, once scales are balanced in the loss or ansatz.
- Modular and annulus constraints become instances of the same anchored four-point neural search used for ordinary correlators.
- Higher-genus sewing constraints and mixed defect systems become natural next targets for the same function-space search.
- When ensembles are bimodal, early stopping can serve as a practical filter that retains the physical low-loss peak.
Where Pith is reading between the lines
- If spectral bias is the real selector, one could deliberately vary network depth, width, or activation to map which function classes are preferred and turn the bias into a diagnostic of ‘physical’ smoothness.
- The method may extend to off-diagonal modular parameter by learning concentric circles in the cross-ratio plane, yielding the full complex-structure dependence from the same sparse anchors.
- Combining the neural reconstruction with independent spectral bounds could produce hybrid bootstraps that output complete partition functions rather than only exclusion plots.
- Failure modes at large central charge, where correlators span many orders of magnitude, suggest that adaptive rescaling of the loss will be needed before the method is routine for holographic or high-c theories.
Editorial analysis
A structured set of objections, weighed in public.
Circularity Check
Mild circularity only: exact anchor plus self-cited spectral-bias premise; full-curve match is still a real test, but WZW early-stop filtering is answer-aware.
specific steps
-
fitted input called prediction
[§3, Eqs. (3.7)–(3.10) and (3.22)–(3.25)]
"Lanc = (ẽG(z0)−Bt(z0)Gexact(z0))2, with one anchor at z0=0.3. … L(ann)anc=(ẽG(o)(z0)−Ba(z0)G(open)exact(z0))2+(ẽG(c)(z0)−Ba(z0)G(closed)exact(z0))2, with a single anchor at z0=0.3 for each of the two channels."
The sole interior data point supplied to the loss is the exact target value of the reduced correlator (or both open/closed channels). That number is therefore fitted by construction; only the off-anchor shape is a genuine reconstruction test. Calling the overall procedure reconstruction from “sparse data” is fair for the curve, but any claim that the method predicts the anchored value itself would be circular.
-
self citation load bearing
[Abstract; §1; §7 (citing [9,10])]
"A key ingredient of this approach is the spectral bias of the neural networks in the lazy training regime, which selects specific crossing-symmetric configurations. … our previous study [9,10] revealed a remarkable fact. In all examples studied there, the network always selected a solution close to a physical correlator. The main goal of the present paper is to present experimental evidence that this bias towards physical correlators extends to modular-invariant partition functions as well."
The central selection principle—that lazy-training spectral bias picks the physical CFT partition function among many crossing-symmetric functions obeying the same gap and anchor—is not derived in this paper. It is load-bearing narrative imported from the same authors’ prior neural-bootstrap papers and then re-illustrated on modular/annulus examples. Without that self-cited premise, crossing+gap+anchor alone do not uniquely determine the reported curves.
-
other
[§4.4 ŝu(2)2 (and App. A.1 ŝu(2)1)]
"In this case, we observed that the full ensemble of runs is bimodal, exhibiting two peaks at z=0.5 cleanly separated. Among the seed-level diagnostics we surveyed, the only feature that correlated with the peak a run landed on was whether the run triggered early stopping. Runs that terminated via the 5×10^3-epoch stagnation criterion … fell almost exclusively into the peak tracking the exact answer, while runs that ran out the full epoch budget clustered around the spurious peak. … Retaining only the early-stopped runs selected 92 of 1000 seeds."
When selection is nontrivial, the majority of seeds miss the physical peak. The early-stopping cut that retains the physical minority is chosen because it correlates with the already-known exact answer. That is post-hoc, answer-aware filtering: the reported “NN prediction” for these models is the filtered sub-ensemble that matches the target, not an unsupervised output of spectral bias alone. For an unknown partition function the same cut would lack independent justification.
full rationale
The modular/Cardy-to-crossing reformulation and the gap exponents are independent external inputs, not defined from the target curves. The network is trained on crossing plus one interior anchor that is literally the exact reduced correlator at z0 (Eqs. 3.10, 3.25), so that single number is fitted, not predicted; the nontrivial claim is recovery of the rest of the function. That recovery is checked against closed-form or character-sum partition functions the authors already know, which is a legitimate reconstruction benchmark rather than a self-definitional loop. Load-bearing selection language (“spectral bias … selects specific crossing-symmetric configurations”) is imported from the authors’ companion papers [9,10] and treated as an empirical fact, not re-derived here—self-citation that supports the narrative but does not algebraically force the curves. In the ŝu(2)1/2 cases the unfiltered ensemble is bimodal and only the early-stopped minority is retained because it tracks the known answer; that filter is answer-aware and would not be independently justified for an unknown theory. None of these steps make the reported full-interval accuracy equivalent to the inputs by construction, so the circularity remains mild (score 3).
Axiom & Free-Parameter Ledger
free parameters (4)
- anchor weight λ_anc =
100
- anchor location z0 =
0.3 (default)
- MLP architecture and training schedule =
2×64 GELU; schedule as in §3
- gap Δ_gap and leading channel exponents =
theory-dependent
axioms (5)
- domain assumption Modular S-invariance of the torus partition function is equivalent to crossing of four identical Z2 twist fields on CP1, reducing on the diagonal to G(z)=(z/(1-z))^{c/4} G(1-z).
- domain assumption Open/closed duality of the annulus equals crossing of a mixed four-point function of defect-changing operators with Δ=c/16 endpoints.
- ad hoc to paper In the lazy-training regime, feed-forward nets exhibit spectral bias that preferentially selects smooth, physically realized crossing-symmetric correlators among many solutions compatible with gap and anchor data.
- domain assumption Cardy boundary states and Verlinde open-channel multiplicities correctly give the annulus spectra for diagonal rational CFTs.
- ad hoc to paper Crossing-symmetric functions satisfying the same gap exponents and single anchor are highly non-unique, so successful full-interval reconstruction is a nontrivial selection effect.
read the original abstract
We develop a neural network bootstrap framework for reconstructing partition functions of two-dimensional conformal field theories (CFTs) based on modular invariance and the Cardy condition, which are recast as crossing equations for four-point correlators. For torus partition functions, we use the twist-field representation in the symmetric-orbifold description to map modular S-invariance to four-point crossing and focus on the diagonal kinematics of four insertions on a line. For annulus partition functions, we formulate open/closed channel duality as crossing symmetry for mixed four-point functions of defect-changing operators in interface CFT. In both cases, the reconstruction problem is formulated in the anchored-bootstrap form, where the crossing constraints are supplemented by minimal spectral input (a gap) and anchor data. We solve this under-determined problem by using lightweight feed-forward neural networks to parametrise the correlators and their corresponding partition functions. A key ingredient of this approach is the spectral bias of the neural networks in the lazy training regime, which selects specific crossing-symmetric configurations. This reformulation unifies standard modular and annulus constraints in two dimensions with the anchored neural approach for CFT correlators, providing a new way to reconstruct full partition functions from sparse data with remarkable accuracy.
Figures
Reference graph
Works this paper leans on
-
[1]
The Conformal Bootstrap: Theory, Numerical Techniques, and Applications
D. Poland, S. Rychkov & A. Vichi,“The Conformal Bootstrap: Theory, Numerical Techniques, and Applications”, Rev. Mod. Phys.91, 015002 (2019),arXiv:1805.04405 [hep-th]
Pith/arXiv arXiv 2019
-
[2]
New developments in the numerical conformal bootstrap
S. Rychkov & N. Su,“New developments in the numerical conformal bootstrap”, Rev. Mod. Phys. 96, 045004 (2024),arXiv:2311.15844 [hep-th]
arXiv 2024
-
[3]
Conformal bootstrap: From Polyakov to our times
S. Rychkov,“Conformal bootstrap: From Polyakov to our times”, Int. J. Mod. Phys. A40, 2530021 (2025),arXiv:2509.02779 [hep-th]
arXiv 2025
-
[4]
Bounding scalar operator dimensions in 4D CFT
R. Rattazzi, V. S. Rychkov, E. Tonni & A. Vichi,“Bounding scalar operator dimensions in 4D CFT”, JHEP0812, 031 (2008),arXiv:0807.0004 [hep-th]
Pith/arXiv arXiv 2008
-
[5]
Bootstrapping Mixed Correlators in the 3D Ising Model
F. Kos, D. Poland & D. Simmons-Duffin,“Bootstrapping Mixed Correlators in the 3D Ising Model”, JHEP1411, 109 (2014),arXiv:1406.4858 [hep-th]
Pith/arXiv arXiv 2014
-
[6]
The Lightcone Bootstrap and the Spectrum of the 3d Ising CFT
D. Simmons-Duffin,“The Lightcone Bootstrap and the Spectrum of the 3d Ising CFT”, JHEP 1703, 086 (2017),arXiv:1612.08471 [hep-th]
Pith/arXiv arXiv 2017
-
[7]
Solving the 3D Ising Model with the Conformal Bootstrap
S.El-Showk, M.F.Paulos, D.Poland, S.Rychkov, D.Simmons-Duffin&A.Vichi,“Solving the 3D Ising Model with the Conformal Bootstrap”, Phys. Rev. D86, 025022 (2012),arXiv:1203.6064 [hep-th]
Pith/arXiv arXiv 2012
-
[8]
Bootstrapping the 3d Ising stress tensor
C.-H. Chang, V. Dommes, R. S. Erramilli, A. Homrich, P. Kravchuk, A. Liu, M. S. Mitchell, D. Poland & D. Simmons-Duffin,“Bootstrapping the 3d Ising stress tensor”, JHEP2503, 136 (2025),arXiv:2411.15300 [hep-th]
Pith/arXiv arXiv 2025
-
[9]
Neural networks reveal a universal bias in conformal correlators
K. Ghosh, S. Kumar, V. Niarchos & A. Stergiou,“Neural networks reveal a universal bias in conformal correlators”, Phys. Lett. B880, 140708 (2026),arXiv:2604.18673 [hep-th]
Pith/arXiv arXiv 2026
-
[10]
Neural Spectral Bias and Conformal Correlators I: Introduction and Applications
K. Ghosh, S. Kumar, V. Niarchos & A. Stergiou,“Neural Spectral Bias and Conformal Correlators I: Introduction and Applications”,arXiv:2604.18686 [hep-th]
-
[11]
Classical and Quantum Conformal Field Theory
G. W. Moore & N. Seiberg,“Classical and Quantum Conformal Field Theory”, Commun. Math. Phys.123, 177 (1989)
1989
-
[12]
Descending into the Modular Bootstrap
N. Benjamin, A. L. Fitzpatrick, W. Li & J. Thaler,“Descending into the Modular Bootstrap”, arXiv:2604.01275 [hep-th]
-
[13]
andstergiou/nn-cft: Zenodo archive for Neural Spectral Bias and Conformal Correlators
K. Ghosh, S. Kumar, V. Niarchos & A. Stergiou,“andstergiou/nn-cft: Zenodo archive for Neural Spectral Bias and Conformal Correlators”,https://doi.org/10.5281/zenodo.20343436
-
[14]
A Universal Inequality for CFT and Quantum Gravity
S. Hellerman,“A Universal Inequality for CFT and Quantum Gravity”, JHEP1108, 130 (2011), arXiv:0902.2790 [hep-th]
Pith/arXiv arXiv 2011
-
[15]
Constraints on 2d CFT partition functions
D. Friedan & C. A. Keller,“Constraints on 2d CFT partition functions”, JHEP1310, 180 (2013), arXiv:1307.6562 [hep-th]. 49
Pith/arXiv arXiv 2013
-
[16]
S. Collier, Y.-H. Lin & X. Yin,“Modular Bootstrap Revisited”, JHEP1809, 061 (2018), arXiv:1608.06241 [hep-th]
Pith/arXiv arXiv 2018
-
[17]
Sphere Packing and Quantum Gravity
T. Hartman, D. Mazáč & L. Rastelli,“Sphere Packing and Quantum Gravity”, JHEP1912, 048 (2019),arXiv:1905.01319 [hep-th]
Pith/arXiv arXiv 2019
-
[18]
Bootstrapping boundaries and branes
S. Collier, D. Mazac & Y. Wang,“Bootstrapping boundaries and branes”, JHEP2302, 019 (2023),arXiv:2112.00750 [hep-th]
Pith/arXiv arXiv 2023
-
[19]
High-dimensional sphere packing and the modular bootstrap
N. Afkhami-Jeddi, H. Cohn, T. Hartman, D. de Laat & A. Tajdini,“High-dimensional sphere packing and the modular bootstrap”, JHEP2012, 066 (2020),arXiv:2006.02560 [hep-th]
Pith/arXiv arXiv 2020
-
[20]
Upgrading Extremal Flows in the Space of Derivatives
R. S. Erramilli,“Upgrading Extremal Flows in the Space of Derivatives”,arXiv:2604.24839 [hep-th]
-
[21]
Correlation functions forM N /SN orbifolds
O. Lunin & S. D. Mathur,“Correlation functions forM N /SN orbifolds”, Commun. Math. Phys. 219, 399 (2001),hep-th/0006196
Pith/arXiv arXiv 2001
-
[22]
Form factors of branch-point twist fields in quantum integrable models and entanglement entropy
J. L. Cardy, O. A. Castro-Alvaredo & B. Doyon,“Form factors of branch-point twist fields in quantum integrable models and entanglement entropy”, J. Statist. Phys.130, 129 (2008), arXiv:0706.3384 [hep-th]
Pith/arXiv arXiv 2008
-
[23]
J. Maldacena, D. Simmons-Duffin & A. Zhiboedov,“Looking for a bulk point”, JHEP1701, 013 (2017),arXiv:1509.03612 [hep-th]
Pith/arXiv arXiv 2017
-
[24]
The Boundary and Crosscap States in Conformal Field Theories
N. Ishibashi,“The Boundary and Crosscap States in Conformal Field Theories”, Mod. Phys. Lett. A4, 251 (1989)
1989
-
[25]
Boundary Conditions, Fusion Rules and the Verlinde Formula
J. L. Cardy,“Boundary Conditions, Fusion Rules and the Verlinde Formula”, Nucl. Phys. B 324, 581 (1989)
1989
-
[26]
Fusion Rules and Modular Transformations in 2D Conformal Field Theory
E. P. Verlinde,“Fusion Rules and Modular Transformations in 2D Conformal Field Theory”, Nucl. Phys. B300, 360 (1988)
1988
-
[27]
Adam: A Method for Stochastic Optimization
D. P. Kingma & J. Ba,“Adam: A Method for Stochastic Optimization”, arXiv:1412.6980 [cs.LG]
-
[28]
The A-D-E Classification of Minimal and A1(1) Conformal Invariant Theories
A. Cappelli, C. Itzykson & J. B. Zuber,“The A-D-E Classification of Minimal and A1(1) Conformal Invariant Theories”, Commun. Math. Phys.113, 1 (1987)
1987
-
[29]
Infinite-dimensional Lie algebras, theta functions and modular forms
V. G. Kac & D. H. Peterson,“Infinite-dimensional Lie algebras, theta functions and modular forms”, Adv. Math.53, 125 (1984)
1984
-
[30]
Conformal Field Theory
P. Di Francesco, P. Mathieu & D. Senechal,“Conformal Field Theory”, Springer (1997), New York
1997
-
[31]
Liouville field theory on a pseudosphere
A. B. Zamolodchikov & A. B. Zamolodchikov,“Liouville field theory on a pseudosphere”, hep-th/0101152 [hep-th]. 50
-
[32]
Boundary Liouville field theory. 1. Boundary state and boundary two point function
V. Fateev, A. B. Zamolodchikov & A. B. Zamolodchikov,“Boundary Liouville field theory. 1. Boundary state and boundary two point function”,hep-th/0001012 [hep-th]
-
[33]
Remarks on Liouville theory with boundary
J. Teschner,“Remarks on Liouville theory with boundary”, PoStmr2000, 041 (2000), hep-th/0009138
Pith/arXiv arXiv 2000
-
[34]
King’s Computational Research, Engineering and Technology Environ- ment (CREATE)
King’s College London,“King’s Computational Research, Engineering and Technology Environ- ment (CREATE)”,https://doi.org/10.18742/rnvf-m076. 51
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.