Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

Machine learning automorphic forms for black holes

T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Neural networks recover modular weights of black-hole counting functions from truncated Fourier series.

desk verdict The eta and Jacobi experiments give a credible proof of concept, but the E2 weight labels are off by two, so the quasi-modular headline claim does not hold as stated. read the letter →

arxiv 2505.05549 v1 pith:GOBVFFEH submitted 2025-05-08 hep-th cs.LGmath.NT

classification hep-thcs.LGmath.NT MSC 11F1111F5068T07
keywords modularformsautomorphicJacobimock-modularBPSblackholesmachinelearningFouriercoefficientsweight
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 asks whether a machine can recover the modular symmetry of a counting function from a handful of its Fourier coefficients. The authors train feed-forward neural networks on coefficients of powers of the Dedekind eta function, the Eisenstein series $E_2$, and products of Jacobi theta functions, and find that the networks predict the modular weight accurately for negative-weight modular and quasi-modular forms, including the weakly holomorphic forms that appear in exact BPS black hole counting. Performance degrades sharply for positive weights and for products of several theta functions. The stated payoff is a proof of concept for using machine learning to detect and verify modular symmetries in quantum gravity when only finitely many coefficients are available.

What carries the argument

The central object is the pair consisting of a truncated Fourier coefficient vector and the modular weight it maps to, learned by a regression network. The input coefficients come from $\eta^n$, $E_2\eta^{2w}$, and forms built from $\theta_1,\theta_2,\theta_3,\theta_4$, with powers of $\eta$ chosen to control the leading power of $q$; inputs are $L^2$-normalized or passed through a log-absolute-value encoder. The network is a deep feed-forward net with ReLU and GELU activations, trained with ADAM on mean squared error. The load-bearing identity is the modular transformation law that defines the weight $k$, together with the Rademacher and Kloosterman expansions that generate coefficient data for $E_2$ powers from polar data.

What would settle it

Take a negative-weight weakly holomorphic modular form outside the trained families and weight range (for example, an arbitrary eta-quotient with weight below the training minimum), compute its first 30 Fourier coefficients, and pass them to the trained network; if the predicted weight is no better than a random guess, the claim that the network has learned to identify modular weights from truncated expansions is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fully connected feed-forward neural network, given the first few Fourier coefficients of a modular, quasi-modular, or Jacobi-derived modular form, can predict its modular weight $k$ --- the exponent in the transformation law $\phi((a\tau+b)/(c\tau+d)) = (c\tau+d)^k \phi(\tau)$. On powers of $\eta$ with negative weights, test errors are below one percent (for example, $0.16\%$ for half-integer negative powers and $0.21\%$ for random real negative powers), and the same holds for negative powers of $E_2$; positive-weight examples fail badly, with test errors of $35.6\%$ and $383\%$ in the $\eta$ and $E_2$ experiments. For Jacobi $\theta$ functions, accuracy is good for positive powers of $\theta_3,\theta_4$ and for products of $\theta$ functions divided by $\eta$, but worsens for negative powers of $\theta_1,\theta_2$ and for mixed-sign products constrained by $\mathrm{sgn}(k+l+m+n)$. The authors also note that the trained networks perform poorly on weights outside the training range, so the demonstrated ability is interpolation within a known family rather than extrapolation to arbitrary modular forms.

Load-bearing premise

The load-bearing premise is that finite Fourier-coefficient data drawn from the same parametric families and weight ranges used in training represent the modular forms that actually occur in black-hole counting; the paper itself reports poor performance outside those weight ranges and filters datasets to forms with at least $n_c$ nonzero coefficients.

Editorial extensions

If this is right

  • Given only the first few Fourier coefficients of a negative-weight modular or quasi-modular form, the method identifies its weight with sub-percent accuracy, narrowing the search for the exact counting function.
  • The method's failure on positive-weight and mixed-sign theta-function products marks a clear boundary: it is currently reliable for the negative-weight families that appear in exact BPS counting, not for arbitrary automorphic forms.
  • Applying the same protocol to congruence subgroups of $\mathrm{SL}(2,\mathbb{Z})$ is a next step the authors identify for detecting automorphic forms in CHL models.
  • For the $N=2$ STU model, the method is suited to determining the weight of the putative Jacobi form in the approximate counting function, thereby reducing the space of candidate forms.
  • Success on these families makes automated detection of modular symmetries in gravitational data a concrete possibility, with applications in AdS/CFT comparisons.

Reading between the lines

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

  • The sharp drop in performance outside the trained weight range suggests the networks are interpolating coefficient-growth patterns rather than learning the modular transformation law itself; a control experiment with randomly shuffled coefficient vectors matched to the same statistics would test this directly.
  • Because the datasets keep only forms with at least $n_c$ nonzero Fourier coefficients, the reported accuracy may overstate performance on sparse expansions, which are common when only a few terms of an unknown counting function are known.
  • The strong signal for negative weights may reflect the exponential coefficient growth of weakly holomorphic forms, so the method might also detect mock-modular and other rapidly growing families, a transfer that the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper trains feed-forward neural networks to predict the weight (or exponent) of modular, quasi-modular, and Jacobi forms from a finite number of Fourier coefficients. The datasets are generated from powers of the Dedekind eta function, products E2(τ)η^{2w}, and products of Jacobi theta functions. The authors report mean relative test errors below 1% for negative-weight eta and E2 families and for several Jacobi families, with larger errors for positive weights and more complicated products. They position this as a proof of concept for identifying modular symmetries in black hole counting functions.

Significance. If the reported results were valid, the paper would provide a useful proof of concept: a finite Fourier expansion suffices to identify the modular weight of a generating function, which could assist in pinning down unknown counting functions in string theory. The manuscript is honest about the failure on positive weights and outside the training range, and it provides code and data on GitHub, which is a strength. However, as detailed below, the labeling of the training targets is off for the E2 and Jacobi experiments, so the central claim is not currently supported.

major comments (3)
  1. [Section 2, Table 2, Appendix A (Eqs. (A.12)–(A.16))] The E2 experiments train on the exponent w in E2(τ)η^{2w} rather than the actual quasi-modular weight k = w+2. Since E2 has weight 2 and η^{2w} has weight w, the product is a quasi-modular form of weight w+2 under the paper's own definition (1.7). Table 2 labels the range (−200, −1/2) as 'negative', but for w = −1/2 the true weight is k = +3/2, so the 'negative-weight' dataset contains positive-weight examples. The networks are therefore learning to predict the exponent w, not the modular weight. This is not a harmless shift: the claim in Section 3 that the networks learn weights of 'negative powers of E2' (which are actually E2 times negative powers of η) is unsupported, and a user applying the trained model to E2/η^24, whose true weight is −10, would obtain a prediction near −12. The same issue propagates into the Kloosterman-sum experiments because (A.12)–(A.16) are written in terms of w. The tables and text must be revised with the correct target k = w+2, or the claims must be rescaled accordingly.
  2. [Section 2, Tables 3–6, Appendix A] The Jacobi experiments use the power k of θ_a as the prediction target, but the modular weight of θ^k_a is k/2, since θ_a is defined as a Jacobi form of weight 1/2 in Appendix A. Moreover, the statement in Section 2 that 'the coefficient of u^l in θ^k_a is a modular form of weight k + l/2' is inconsistent with this: for l=0 it would give weight k, not k/2. The simple u-expansion coefficient of a Jacobi form is not generally a modular form of SL(2,Z) (e.g., the u^1 coefficient of θ_3 is q^{1/2}, which is not a modular form). The tables report errors on predicting k, so the abstract's claim of predicting 'modular weights' from Jacobi data is not demonstrated. The authors need to either (a) define the target as the true modular weight (k/2 adjusted for the u^l coefficient and any η factor) and retrain, or (b) explicitly restrict the claim to predicting the exponent k and justify why that is the relevant physical quantity.
  3. [Section 2, Tables 1–6, Appendix B] The evaluation provides no baselines and no error bars. All reported errors are single mean relative errors without standard deviation across random seeds or comparisons to simple alternatives (e.g., linear regression on the same Fourier coefficients, or a constant predictor equal to the mean training weight). Because the labels are known analytically, a low error could simply reflect an easy regression problem rather than the network 'learning modular structure.' For a proof-of-concept claim, the authors should include at least one baseline and repeat the experiments a few times to report the spread.
minor comments (7)
  1. [Section 3] 'Negative powers of E2' is a misnomer; the experiments use E2 multiplied by negative powers of η, not negative powers of E2.
  2. [Section 2] The statement that imaginary-part experiments are not reported should be quantified (e.g., give the actual poor error) so readers can judge the omission.
  3. [Table 1] The notation '(−200, 30)' in the third column is ambiguous; it should clarify whether the first entry is the minimum weight or the negative of the maximum, and specify the step size in the half-integer and random sampling.
  4. [Equation (B.2)] The encoder Log[N[Abs[·]]] is undefined for zero Fourier coefficients; the paper should explain how zeros are handled in the data preprocessing.
  5. [Throughout (abstract, Section 2, Appendix A)] The paper uses 'mock-modular' and 'quasi-modular' interchangeably for E2; this is imprecise and should be corrected to 'quasi-modular' throughout.
  6. [Appendix B] The GitHub repository is a strength, but the paper should list the random seed(s), software versions, and any preprocessing steps that affect the exact numerical results.
  7. [Section 2, Tables 3–6] The n_c cut selects only forms with at least n_c non-zero coefficients; a sensitivity analysis (e.g., varying n_c) would clarify how the reported errors depend on this filtering.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: supervised regression on known labels; self-citations are background only, and the E2 off-by-two is a labeling/correctness risk rather than a circular step.

full rationale

This paper performs supervised regression: the target weight labels are known from the generating formulas, and the held-out test errors measure how well a feed-forward network can invert the Fourier-coefficient map within the sampled parametric families. No step feeds the predicted weight back into the inputs, and no parameter is fitted to a subset and then 'predicted' on a closely related quantity, so the central ML claim is not circular by construction. The self-citations (Refs [18], [23], [26]) appear in background, motivation, and future-work contexts; they are not load-bearing for the reported experiments and do not raise the circularity score. The main limitations are external-validity and labeling concerns, not circularity: the paper itself acknowledges in Section 2 that 'the performance of the trained neural networks is poor on modular forms with weights outside of these ranges,' and the n_c filtering in Tables 3-6 selects forms with enough non-zero coefficients. In addition, for E2-based forms, Table 2 labels the exponent w in 2 E2(q)Delta(q)^{w/12} as the 'weight,' whereas under the paper's own quasi-modular transformation (1.7) the total weight is w+2, so the reported negative-weight accuracy mixes positive-weight cases; this is a correctness/labeling error, not a circular reduction, because the network still predicts a well-defined label from its inputs. Overall, the derivation is self-contained and no circular step is exhibited; the score of 2 reflects the minor non-load-bearing self-citations and the acknowledged generalization limits.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central ML claim rests on standard modular form definitions and Rademacher theory, plus the physical domain assumption that black hole counting functions are automorphic forms. The free parameters are experimental choices (coefficient truncation, data filtering, weight range, architecture, normalization) that bound the claimed capability. No new particles, forces, or entities are introduced.

free parameters (5)
  • Fourier coefficient truncation (n_q, n_u) = n_q 20-80, n_u 50-80 depending on experiment
    The number of input coefficients is hand-picked per experiment and directly controls the input representation, with different truncations changing reported errors.
  • Minimum non-zero coefficient threshold (n_c) = n_c = 22 to 53 across experiments
    Datasets retain only modular forms with at least n_c non-zero coefficients in the truncated expansion; this post hoc filter changes dataset size and may bias toward forms whose weight is easier to infer.
  • Training weight range (w_min, w_max) = examples: (-200,-1/2), (1/2,100), (-40,40)
    The range is hand-chosen; the paper acknowledges poor generalization to weights outside the range, so the claimed capability is bounded by these choices.
  • Network architecture and hyperparameters = Net1 (Eq. B.3) or Net2 (Eq. B.4), ADAM learning rate 0.001, batch size 64, 75/15/10 split
    These hand-picked choices affect performance; the authors state the architecture was not optimized, so accuracies are tied to these settings.
  • Input normalization scheme = L2 norm for eta/E2; log-abs encoder (Eq. B.2) for Jacobi theta data
    The choice of normalization (L2 vs elementwise log|.|) is a modeling decision that changes the features available to the network.
assumptions (4)
  • standard math Dedekind eta, Eisenstein E2, and Jacobi theta functions have the stated Fourier expansions and modular transformation properties (Eqs. A.1, A.8, A.17-A.20).
    Used to generate all training and test data; these are standard results in the theory of modular forms.
  • standard math The Rademacher expansion (Eq. A.5) correctly expresses Fourier coefficients of negative-weight modular forms in terms of polar coefficients and Kloosterman sums.
    Used in Section 2 to generate E2-related training coefficients and to split into I_(-w-1) and I_(-w) pieces.
  • domain assumption BPS black hole degeneracies in N=4 and N=8 string compactifications are generated by modular, Jacobi, or Siegel modular forms (Eq. 1.1, refs. [4,6,7,9,16]).
    This motivates the study, though the ML experiments themselves do not test black hole degeneracies.
  • standard math Quasi-modular forms of depth 1 transform as in Eq. (1.7), with E2 as a weight-2 example.
    Used to frame the E2 experiments as mock/quasi-modular data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine learning automorphic forms for black holes." pith.science (2026). https://pith.science/paper/GOBVFFEH

@misc{pith2026250505549,
  author       = {Pith},
  title        = {Pith review of: Machine learning automorphic forms for black holes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GOBVFFEH}},
  note         = {Machine review of arXiv:2505.05549}
}
read the original abstract

Modular, Jacobi, and mock-modular forms serve as generating functions for BPS black hole degeneracies. By training feed-forward neural networks on Fourier coefficients of automorphic forms derived from the Dedekind eta function, Eisenstein series, and Jacobi theta functions, we demonstrate that machine learning techniques can accurately predict modular weights from truncated expansions. Our results reveal strong performance for negative weight modular and quasi-modular forms, particularly those arising in exact black hole counting formulae, with lower accuracy for positive weights and more complicated combinations of Jacobi theta functions. This study establishes a proof of concept for using machine learning to identify how data is organized in terms of modular symmetries in gravitational systems and suggests a pathway toward automated detection and verification of symmetries in quantum gravity.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Black Hole Black Boxes: Numerical Black Hole Metrics via AInstein Neural Networks

    gr-qc 2026-07 conditional novelty 6.0 of 10

    Unsupervised Lorentzian PINNs with embedded S^{2} topology recover maximally extended Schwarzschild and yield candidate Petrov type-I vacuum black-hole metrics with genuinely trapped interiors.

Reference graph

Works this paper leans on

27 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Sen, Black Hole Entropy Function, Attractors and Precision Coun ting of Microstates , Gen

    A. Sen, Black Hole Entropy Function, Attractors and Precision Coun ting of Microstates , Gen. Rel. Grav. 40 (2008) 2249 [0708.1270]

  2. [2]

    Black holes and modular forms in string theory

    S. Murthy, Black holes and modular forms in string theory , 2305.11732

  3. [3]

    Alexandrov, Mock modularity at work, or black holes in a forest , 2505.02572

    S. Alexandrov, Mock modularity at work, or black holes in a forest , 2505.02572

  4. [4]

    Dabholkar, Exact counting of black hole microstates , Phys

    A. Dabholkar, Exact counting of black hole microstates , Phys. Rev. Lett. 94 (2005) 241301 [hep-th/0409148]

  5. [5]

    Exact and Asymptotic Degeneracies of Small Black Holes

    A. Dabholkar, F. Denef, G.W. Moore and B. Pioline, Exact and asymptotic degeneracies of small black holes , JHEP 08 (2005) 021 [hep-th/0502157]

  6. [6]

    Jatkar and A

    D.P. Jatkar and A. Sen, Dyon spectrum in CHL models , JHEP 04 (2006) 018 [hep-th/0510147]

  7. [7]

    Dabholkar, S

    A. Dabholkar, S. Murthy and D. Zagier, Quantum Black Holes, Wall Crossing, and Mock Modular Forms, 1208.4074

  8. [8]

    Dabholkar, J

    A. Dabholkar, J. Gomes and S. Murthy, Nonperturbative black hole entropy and Kloosterman sums, JHEP 03 (2015) 074 [1404.0033]. – 15 –

Show all 27 references
  1. [9]

    Maldacena, G.W

    J.M. Maldacena, G.W. Moore and A. Strominger, Counting BPS black holes in toroidal Type II string theory, hep-th/9903163

  2. [10]

    Kunduri, J

    H.K. Kunduri, J. Lucietti and H.S. Reall, Near-horizon symmetries of extremal black holes , Class. Quant. Grav. 24 (2007) 4169 [0705.4214]

  3. [11]

    Ferrara and R

    S. Ferrara and R. Kallosh, Supersymmetry and attractors , Phys. Rev. D 54 (1996) 1514 [hep-th/9602136]

  4. [12]

    Ferrara and R

    S. Ferrara and R. Kallosh, Universality of supersymmetric attractors , Phys. Rev. D 54 (1996) 1525 [hep-th/9603090]

  5. [13]

    Sen, Black hole entropy function and the attractor mechanism in h igher derivative gravity , JHEP 09 (2005) 038 [hep-th/0506177]

    A. Sen, Black hole entropy function and the attractor mechanism in h igher derivative gravity , JHEP 09 (2005) 038 [hep-th/0506177]

  6. [14]

    Goldstein, N

    K. Goldstein, N. Iizuka, R.P. Jena and S.P. Trivedi, Non-supersymmetric attractors, Phys. Rev. D 72 (2005) 124021 [hep-th/0507096]

  7. [15]

    Dabholkar, A

    A. Dabholkar, A. Sen and S.P. Trivedi, Black hole microstates and attractor without supersymmetry, JHEP 01 (2007) 096 [hep-th/0611143]

  8. [16]

    Dijkgraaf, E.P

    R. Dijkgraaf, E.P. Verlinde and H.L. Verlinde, Counting dyons in N=4 string theory , Nucl. Phys. B 484 (1997) 543 [hep-th/9607026]

  9. [17]

    Lopes Cardoso, B

    G. Lopes Cardoso, B. de Wit, J. Kappeli and T. Mohaupt, Asymptotic degeneracy of dyonic N = 4 string states and black hole entropy , JHEP 12 (2004) 075 [hep-th/0412287]

  10. [18]

    Lopes Cardoso, S

    G. Lopes Cardoso, S. Nampuri and M. Rossell´ o, Rademacher Expansion of a Siegel Modular Form for N = 4 Counting, Annales Henri Poincare 25 (2024) 4065 [2112.10023]

  11. [19]

    Cvetic and D

    M. Cvetic and D. Youm, Dyonic BPS saturated black holes of heterotic string on a six torus, Phys. Rev. D 53 (1996) 584 [hep-th/9507090]

  12. [20]

    He, K.-H

    Y.-H. He, K.-H. Lee and T. Oliver, Machine-learning the Sato–Tate conjecture , J. Symb. Comput. 111 (2022) 61 [2010.01213]

  13. [21]

    He, K.-H

    Y.-H. He, K.-H. Lee and T. Oliver, Machine learning invariants of arithmetic curves , J. Symb. Comput. 115 (2023) 478 [2012.04084]

  14. [22]

    He, K.-H

    Y.-H. He, K.-H. Lee, T. Oliver and A. Pozdnyakov, Murmurations of elliptic curves , 2204.10140

  15. [23]

    Cardoso, S

    G.L. Cardoso, S. Nampuri and D. Polini, An approach to BPS black hole microstate counting in an N = 2 STU model , JHEP 08 (2020) 057 [1903.07586]

  16. [24]

    Hashimoto, S

    K. Hashimoto, S. Sugishita, A. Tanaka and A. Tomiya, Deep learning and the AdS/CFT correspondence, Phys. Rev. D 98 (2018) 046019 [1802.08313]

  17. [25]

    Hashimoto, AdS/CFT correspondence as a deep Boltzmann machine , Phys

    K. Hashimoto, AdS/CFT correspondence as a deep Boltzmann machine , Phys. Rev. D 99 (2019) 106017 [1903.04951]

  18. [26]

    Jejjala, S

    V. Jejjala, S. Mondkar, A. Mukhopadhyay and R. Raj, Learning holographic horizons , Phys. Rev. D 111 (2025) 026016 [2312.08442]

  19. [27]

    ml modular forms

    V. Jejjala, S. Nampuri, D. Nxumalo, P. Roy and A. Swain, “ml modular forms.” https://github.com/abinash7s/ml_modular_forms, 2025. – 16 –

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.