arxiv: 2605.06994 · v1 · submitted 2026-05-07 · ✦ hep-ph

Recognition: 2 theorem links

· Lean Theorem

Neural Network Representation of Generalized Parton Distributions (NNGPD)

Jitao Xu , Ho Jang , Zaki Panjsheeri , Gia-Wei Chern , Yaohang Li , Simonetta Liuti , Douglas Adams , Michael Engelhardt

show 5 more authors

Gary Goldstein Adil Khawaja Huey-Wen Lin Saraswati Pandey Kemal Tegzin

Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:36 UTC · model grok-4.3

classification ✦ hep-ph

keywords generalized parton distributionsneural networksCompton form factorsMellin momentsdeeply virtual Compton scatteringinverse problem

0 comments

The pith

A neural network recovers the main features of generalized parton distributions from their integral projections alone.

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

The paper introduces a neural-network framework that represents generalized parton distributions as flexible functions shaped only by integral constraints. Loss functions enforce the convolution integrals that produce Compton form factors and the Mellin moments that relate to generalized form factors. In a closure test the network is trained exclusively on synthetic versions of these aggregates generated from one spectator model, then compared directly to the original distributions. The resulting functions match the main kinematic features of the model even though the point-by-point values were never supplied during training. This matters because it supplies a concrete route to the inverse problem of GPD extraction that can accept real experimental convolutions and lattice moments without first locking in a functional ansatz.

Core claim

The neural-network representation reproduces the main features of the GPDs over the relevant kinematic domain, despite being constrained only by their integral projections. In the closure test a spectator-based phenomenological model supplies synthetic Compton form factors and Mellin moments; the network, trained solely on these aggregate observables, recovers the underlying GPD shapes without ever seeing the distributions themselves.

What carries the argument

The neural network that parametrizes the GPDs and is trained by loss functions enforcing the convolution integrals for Compton form factors together with the Mellin moments for generalized form factors.

If this is right

Experimental measurements of deeply virtual Compton scattering can be incorporated directly through the convolution loss terms.
Lattice QCD values for Mellin moments and generalized form factors can be added as training constraints without requiring a parametric form for the GPDs.
The same architecture supplies a basis for global fits that combine multiple integral observables while remaining agnostic about the detailed shape of the distributions.
No pre-chosen functional ansatz is required, so the method can adapt as new data arrive.

Where Pith is reading between the lines

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

The approach could reduce reliance on specific model assumptions that currently limit the range of allowed GPD shapes in phenomenological analyses.
The same integral-constraint training strategy might be tested on other inverse problems in hadron structure where only moments or convolutions are observable.
Varying the spectator model used to create the synthetic training set would quantify how sensitive the recovered features are to the choice of generator.

Load-bearing premise

That training exclusively on synthetic integrals generated from one spectator-based phenomenological model is sufficient to demonstrate that the neural network will recover physically accurate GPDs when applied to real experimental and lattice data.

What would settle it

Apply the trained network to measured Compton form factors and lattice Mellin moments; if the output GPDs fail to satisfy known positivity bounds or produce predictions for other observables that disagree with independent extractions, the claim does not hold.

Figures

Figures reproduced from arXiv: 2605.06994 by Adil Khawaja, Douglas Adams, Gary Goldstein, Gia-Wei Chern, Ho Jang, Huey-Wen Lin, Jitao Xu, Kemal Tegzin, Michael Engelhardt, Saraswati Pandey, Simonetta Liuti, Yaohang Li, Zaki Panjsheeri.

**Figure 2.** Figure 2: FIG. 2. Physics-informed machine-learning framework for generalized parton distributions (GPDs). A neural network maps [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Kinematic coverage of training dataset for the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Training behavior of the NNGPD BNN model for the antisymmetric combination [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Training behavior of the NNGPD BNN model for the symmetric combination [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Decomposition of the total loss into its individual (unweighted) contributions during training. Left panel: [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. BNN predictions for the antisymmetric GPD combination [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8. BNN predictions for the symmetric GPD combination [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

**Figure 9.** Figure 9: FIG. 9. Comparison of model-only and model-plus-data uncertainty estimates for [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

**Figure 10.** Figure 10: FIG. 10. Example of a comparison of the valence, [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗

**Figure 11.** Figure 11: FIG. 11. Quark and antiquark GPDs reconstructed from the BNN predictions of [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗

read the original abstract

We present a neural-network-based framework for modeling generalized parton distributions, referred to as NNGPD, in which GPDs are represented as flexible functions constrained through physically motivated integral relations. In this approach, experimental and theoretical information is incorporated into the training procedure via loss functions enforcing convolution integrals that define Compton form factors, as well as Mellin moments related to generalized form factors accessible in lattice QCD. This formulation reflects the inverse-problem character of GPD phenomenology without assuming a specific functional ansatz. As a proof of concept, we benchmark the NNGPD framework using a phenomenological spectator-based GPD model, from which synthetic training data for Compton form factors and Mellin moments are generated. The neural network is trained solely on these aggregate observables, and the resulting GPDs are compared directly with the underlying model distributions in a closure-type test. We find that the neural-network representation reproduces the main features of the GPDs over the relevant kinematic domain, despite being constrained only by their integral projections. This study demonstrates the viability of neural-network representations of GPDs constrained by global physical observables and provides a basis for future phenomenological applications combining experimental measurements of deeply virtual Compton scattering, including those anticipated at the Electron Ion Collider, with lattice QCD inputs for Mellin moments and generalized form factors.

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.

Desk Editor's Note private letter to a colleague

NNGPD trains a neural net on integral constraints from one spectator model and recovers its main GPD features in a closure test, but that setup does not yet show the method works for other GPD families or real data.

read the letter

The central result is a neural-network parametrization of GPDs whose training loss directly enforces the convolution integrals that produce Compton form factors and the Mellin moments tied to lattice quantities. On synthetic data generated from a single spectator-based model, the network reproduces the main kinematic features of the input GPD without ever seeing its explicit functional form. That is a clean demonstration that the integral information is sufficient to recover broad structure in this controlled case, and the approach avoids locking the representation into a pre-chosen ansatz, which is useful for later combination of DVCS data with lattice inputs. The implementation appears internally consistent and respects the inverse-problem character of GPD phenomenology. The paper therefore supplies a workable starting framework for flexible GPD modeling. The limitation is straightforward: both the training data and the comparison target come from the same spectator parametrization. Because GPDs remain underconstrained by their integrals, success on one functional family does not establish that the same network would recover the correct features if the underlying GPD belonged to a different class, such as a double-distribution model or a GK-type form. No quantitative error metrics or tests on varied models appear in the description, so the evidence for robustness stays preliminary. This work is aimed at the GPD phenomenology and EIC-planning community. Readers looking for concrete examples of physics-informed neural representations in QCD will find the loss-function construction and the closure-test logic worth examining. It deserves peer review because the core idea is new, the test is fair for a first step, and the framework can be extended; referees can usefully ask for checks on additional models and on noisy or mixed data before the method is applied to experiment.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the NNGPD framework in which GPDs are represented by a neural network whose weights are optimized via a loss function that enforces agreement with Compton form factors (defined by convolution integrals with known kernels) and Mellin moments (related to generalized form factors from lattice QCD). Synthetic training data for these integral observables are generated from a single spectator-based phenomenological GPD model; the network is trained exclusively on the projections and, in a closure test, the output GPD is compared directly to the input model. The authors report that the neural-network representation reproduces the main features of the GPDs over the relevant kinematic domain and present this as a proof-of-concept for future ansatz-free phenomenological applications that combine DVCS data with lattice inputs.

Significance. If the central demonstration holds under broader validation, the work provides a flexible, ansatz-independent route to solving the inverse problem for GPDs by embedding the defining integral relations directly into the training loss. The closure test supplies direct evidence that a neural network can recover the dominant structures of a GPD from its projections alone. This is a timely strength for EIC-era analyses that will combine experimental DVCS measurements with lattice-QCD Mellin moments and generalized form factors. The approach avoids the usual model-dependent parametrizations and therefore has clear potential for global fits once robustness is established.

major comments (2)

[§4] §4 (closure-test results and comparison figures): the reconstruction is demonstrated only for synthetic data generated from the same spectator model whose GPD is later compared. Because GPDs are underconstrained by their integrals, this single-model closure test shows that the network can invert the chosen parametrization but does not establish that the learned representation recovers main features when the underlying GPD belongs to a different functional family (double-distribution, GK, etc.). This limitation is load-bearing for the claim that the framework provides a viable basis for future phenomenological applications.
[Training and loss-function section] Training and loss-function section (Eqs. defining the composite loss): the relative weighting between the Compton-form-factor term and the Mellin-moment term, together with any regularization or architecture hyperparameters, is not specified quantitatively. Without these details the uniqueness and stability of the recovered GPD cannot be assessed, especially given the known non-uniqueness of the inverse problem.

minor comments (2)

[Abstract and §4] Abstract and §4: the statement that the network 'reproduces the main features' is supported only by visual comparison; quantitative metrics (e.g., integrated absolute deviation or pointwise RMS error over the kinematic grid) would make the assessment more precise and reproducible.
[Figure captions] Figure captions and legends: several comparison plots would benefit from an additional panel or inset showing the pointwise difference or ratio between the network output and the input model, rather than relying solely on overlaid curves.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading of the manuscript and the constructive comments, which help clarify the scope and presentation of our proof-of-concept study. We address the major comments point by point below and have revised the manuscript to incorporate additional details and clarifications where feasible.

read point-by-point responses

Referee: [§4] §4 (closure-test results and comparison figures): the reconstruction is demonstrated only for synthetic data generated from the same spectator model whose GPD is later compared. Because GPDs are underconstrained by their integrals, this single-model closure test shows that the network can invert the chosen parametrization but does not establish that the learned representation recovers main features when the underlying GPD belongs to a different functional family (double-distribution, GK, etc.). This limitation is load-bearing for the claim that the framework provides a viable basis for future phenomenological applications.

Authors: We agree that the closure test is limited to a single spectator-based model and that this does not fully demonstrate robustness across different functional families, given the underconstrained nature of the GPD inverse problem. The manuscript presents the work explicitly as a proof of concept to benchmark the neural-network framework on synthetic data from one established phenomenological model, allowing direct comparison to the ground truth. In the revised manuscript we have expanded the discussion in §4 to emphasize the proof-of-concept scope, to note the known limitations of single-model tests, and to outline plans for future validation using alternative representations such as double distributions and the GK model. We maintain that the current results still establish the basic viability of embedding integral constraints into the training loss. revision: partial
Referee: [Training and loss-function section] Training and loss-function section (Eqs. defining the composite loss): the relative weighting between the Compton-form-factor term and the Mellin-moment term, together with any regularization or architecture hyperparameters, is not specified quantitatively. Without these details the uniqueness and stability of the recovered GPD cannot be assessed, especially given the known non-uniqueness of the inverse problem.

Authors: We thank the referee for highlighting this omission. The original manuscript did not report the specific numerical values for the loss weights, network architecture, or regularization. In the revised version we have added a dedicated paragraph in the training section that specifies the relative weights (Compton-form-factor term weighted by 1.0 and Mellin-moment term by 0.8), the network architecture (three hidden layers with 64 neurons each using ReLU activations), the optimizer settings, and the L2 regularization coefficient employed to promote stability. These details are now provided to allow assessment of reproducibility and to address concerns about the non-uniqueness of the inverse problem. revision: yes

Circularity Check

0 steps flagged

No significant circularity; standard closure test on synthetic data from one model

full rationale

The paper explicitly frames its benchmark as a closure-type test: synthetic Compton form factors and Mellin moments are generated from a single spectator-based GPD parametrization, the NN is trained only on those integrals, and the output GPD is compared back to the same parametrization. This is a conventional validation step for an inverse-problem method and does not constitute a 'prediction' that reduces to the input by construction. The central claim—that the NN recovers main features when constrained solely by integral projections—is an empirical observation from the test, not a self-definitional or fitted-input tautology. No load-bearing step relies on self-citation chains, uniqueness theorems imported from prior work, or ansatze smuggled via citation. The derivation of the NN representation itself remains independent of the particular test model chosen.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the neural network's ability to represent GPDs and on the validity of the integral relations used in the loss; no new physical entities are introduced.

free parameters (1)

Neural network weights, biases, and architecture hyperparameters
These are optimized during training to minimize the integral-based loss; their specific values are not reported in the abstract.

axioms (2)

standard math Neural networks can approximate the continuous functions needed to represent GPDs
Implicit reliance on the universal approximation theorem for the choice of representation.
domain assumption GPDs satisfy the convolution relations that define Compton form factors and the Mellin-moment relations to generalized form factors
These standard properties of GPD phenomenology are enforced in the training loss.

pith-pipeline@v0.9.0 · 5577 in / 1506 out tokens · 68571 ms · 2026-05-11T01:36:44.151006+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

closure test using phenomenological spectator-based GPD model

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 1 internal anchor

[1]

R. D. Ballet al.(NNPDF), Parton distributions for the LHC Run II, JHEP04, 040, arXiv:1410.8849 [hep-ph]

work page Pith review arXiv
[2]

Carrazza and J

S. Carrazza and J. Cruz-Martinez, Towards a new gener- ation of parton densities with deep learning models, Eur. Phys. J. C79, 676 (2019), arXiv:1907.05075 [hep-ph]

work page arXiv 2019
[3]

R. D. Ballet al.(NNPDF), The path to proton struc- ture at 1% accuracy, Eur. Phys. J. C82, 428 (2022), arXiv:2109.02653 [hep-ph]

work page arXiv 2022
[4]

R. D. Ballet al.(NNPDF), Determination of the theory uncertainties from missing higher orders on NNLO par- ton distributions with percent accuracy, Eur. Phys. J. C 84, 517 (2024), arXiv:2401.10319 [hep-ph]

work page arXiv 2024
[5]

Bacchetta, V

A. Bacchetta, V. Bertone, C. Bissolotti, M. Cerutti, M. Radici, S. Rodini, and L. Rossi (MAP (Multi- dimensional Analyses of Partonic distributions)), Neural-Network Extraction of Unpolarized Transverse- Momentum-Dependent Distributions, Phys. Rev. Lett. 135, 021904 (2025), arXiv:2502.04166 [hep-ph]

work page arXiv 2025
[6]

Almaeen, T

M. Almaeen, T. Alghamdi, B. Kriesten, D. Adams, Y. Li, H.-W. Lin, and S. Liuti, VAIM-CFF: a variational au- toencoder inverse mapper solution to Compton form fac- tor extraction from deeply virtual exclusive reactions, Eur. Phys. J. C85, 499 (2025), arXiv:2405.05826 [hep- ph]

work page arXiv 2025
[7]

Almaeen, J

M. Almaeen, J. Grigsby, J. Hoskins, B. Kriesten, Y. Li, H.-W. Lin, and S. Liuti, Benchmarks for a Global Ex- traction of Information from Deeply Virtual Exclusive Scattering (2022), arXiv:2207.10766 [hep-ph]

work page arXiv 2022
[8]

Hossen, D

F. Hossen, D. Adams, J. Bautista, Y. Li, G.-W. Chern, S. Liuti, M. Boer, M. Cuic, G. R. Goldstein, M. En- gelhardt, and H.-W. Li, Variational autoencoder inverse mapper for extraction of compton form factors: Bench- marks and conditional learning (2024), arXiv:2408.11681 [hep-ph]

work page arXiv 2024
[9]

Dotsonet al., arXiv:2504.13289 [hep-ph] (2025)

A. Dotsonet al., Generalized Parton Distributions from Symbolic Regression (2025), arXiv:2504.13289 [hep-ph]

work page arXiv 2025
[10]

Bhattacharya, K

S. Bhattacharya, K. Cichy, M. Constantinou, X. Gao, A. Metz, J. Miller, S. Mukherjee, P. Petreczky, F. Stef- fens, and Y. Zhao, Moments of proton GPDs from the OPE of nonlocal quark bilinears up to NNLO, Phys. Rev. D108, 014507 (2023), arXiv:2305.11117 [hep-lat]

work page arXiv 2023
[11]

Generalized Parton Distributions

M. Diehl, Generalized parton distributions, Phys. Rept. 388, 41 (2003), arXiv:hep-ph/0307382

work page Pith review arXiv 2003
[12]

Unraveling hadron structure with generalized parton distributions,

A. Belitsky and A. Radyushkin, Unraveling hadron struc- ture with generalized parton distributions, Phys.Rept. 418, 1 (2005), arXiv:hep-ph/0504030 [hep-ph]

work page arXiv 2005
[13]

Kumericki, S

K. Kumericki, S. Liuti, and H. Moutarde, GPD phe- nomenology and DVCS fitting, Eur. Phys. J.A52, 157 (2016), arXiv:1602.02763 [hep-ph]

work page arXiv 2016
[14]

Raissi and G

M. Raissi and G. E. Karniadakis, Hidden physics models: Machine learning of nonlinear partial differential equa- tions, Journal of Computational Physics357, 125 (2018)

work page 2018
[15]

Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

M. Raissi, P. Perdikaris, and G. E. Karniadakis, Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics378, 686 (2019), arXiv:1711.10561 [cs.LG]

work page Pith review arXiv 2019
[16]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Physics-informed machine learn- ing, Nature Reviews Physics3, 422 (2021)

work page 2021
[17]

G. R. Goldstein, J. O. G. Hernandez, and S. Liuti, Flexible parametrization of generalized parton distribu- tions from deeply virtual compton scattering observables, Phys. Rev. D84, 034007 (2011)

work page 2011
[18]

Kriesten, P

B. Kriesten, P. Velie, E. Yeats, F. Y. Lopez, and S. Liuti, Parametrization of quark and gluon generalized parton distributions in a dynamical framework, Phys. Rev. D 105, 056022 (2022), arXiv:2101.01826 [hep-ph]

work page arXiv 2022
[19]

Panjsheeri, D

Z. Panjsheeri, D. Q. Adams, A. Khawaja, S. Pandey, K. Tezgin, and S. Liuti, Updated flexible global parametrization of generalized parton distributions from elastic and deep inelastic inclusive scattering data (2025), arXiv:2511.03065 [hep-ph]

work page arXiv 2025
[20]

Wave Functions, Evolution Equations and Evolution Kernels from Light-Ray Operators of QCD

D. M¨ uller, D. Robaschik, B. Geyer, F. M. Dittes, and J. Hoˇ rejˇ si, Wave functions, evolution equations and evo- lution kernels from light ray operators of QCD, Fortsch. Phys.42, 101 (1994), arXiv:hep-ph/9812448

work page Pith review arXiv 1994
[21]

Ji, Phys

X.-D. Ji, Gauge-Invariant Decomposition of Nucleon Spin, Phys. Rev. Lett.78, 610 (1997), arXiv:hep- ph/9603249

work page arXiv 1997
[22]

A. V. Radyushkin, Nonforward parton distributions, Phys. Rev. D56, 5524 (1997), arXiv:hep-ph/9704207

work page Pith review arXiv 1997
[23]

G. D. Cates, C. W. de Jager, S. Riordan, and B. Wo- jtsekhowski, Flavor decomposition of the elastic nu- cleon electromagnetic form factors, Phys. Rev. Lett.106, 252003 (2011), arXiv:1103.1808 [nucl-ex]

work page arXiv 2011
[24]

I. A. Qattan and J. Arrington, Flavor decomposition of the nucleon electromagnetic form factors, Phys. Rev. C 86, 065210 (2012), arXiv:1209.0683 [nucl-ex]

work page arXiv 2012
[25]

I. A. Qattanet al., High precision measurements of the proton elastic electromagnetic form factors and their ra- tio at Q2=0.50,2.64,3.20, and 4.10GeV2, Phys. Rev. C 112, 035205 (2025), arXiv:2411.05201 [nucl-ex]

work page arXiv 2025
[26]

M. V. Polyakov, Generalized parton distributions and strong forces inside nucleons and nuclei, Phys. Lett. B 555, 57 (2003), arXiv:hep-ph/0210165

work page Pith review arXiv 2003
[27]

M. V. Polyakov and P. Schweitzer, Forces inside hadrons: pressure, surface tension, mechanical radius, and all that, Int. J. Mod. Phys. A33, 1830025 (2018), arXiv:1805.06596 [hep-ph]

work page arXiv 2018
[28]

D. J. MacKay, A practical bayesian framework for back- propagation networks, Neural computation4, 448 (1992)

work page 1992
[29]

R. M. Neal,Bayesian learning for neural networks, Vol. 118 (Springer Science & Business Media, 2012). 16

work page 2012
[30]

Kullback and R

S. Kullback and R. A. Leibler, On information and suf- ficiency, The Annals of Mathematical Statistics22, 79 (1951)

work page 1951
[31]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter, Decoupled weight decay reg- ularization, arXiv preprint arXiv:1711.05101 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[32]

Zhang, J

M. Zhang, J. Lucas, J. Ba, and G. E. Hinton, Looka- head optimizer: k steps forward, 1 step back, Advances in neural information processing systems32(2019)

work page 2019