Pith. sign in

REVIEW 3 major objections 5 minor 6 cited by

This paper claims that a single nonparametric dipole amplitude, extracted by a physics-informed neural network under collinearly improved BK evolution and DIS plus J/psi data, simultaneously fits all fitted observables and yields a non-nega

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 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 18:33 UTC pith:IGZXVY3Y

load-bearing objection Deserves a serious referee: a genuinely useful PINN-based global dipole extraction, but the headline momentum-space positivity is imposed by a huge penalty, so the central claim needs an ablation to be fully convincing. the 3 major comments →

arxiv 2603.08008 v2 pith:IGZXVY3Y submitted 2026-03-09 hep-ph hep-exnucl-exnucl-th

Physics-Informed Global Extraction of the Universal Small-x Dipole Amplitude

Si-Wei Dai , Fu-Peng Li , Long-Gang Pang , Guang-You Qin , Shu-Yi Wei , Han-Zhong Zhang , Wenbin Zhao This is my paper
classification hep-ph hep-exnucl-exnucl-th
keywords dipole amplitudesmall-xColor Glass CondensateBalitsky-Kovchegov evolutionphysics-informed neural networkdeep inelastic scatteringcharm productionJ/psi photoproduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that the universal small-x dipole scattering amplitude, the function controlling many high-energy QCD processes, can be inferred from data as a flexible neural-network function instead of being forced into a standard analytic shape. The network is trained to respect the collinearly improved Balitsky-Kovchegov evolution equation while reproducing measured inclusive and charm deep-inelastic cross sections and exclusive J/psi photoproduction. One single amplitude describes all three channels, which the paper says relaxes the long-standing difficulty of fitting total and charm cross sections together in conventional small-x analyses. The same solution has a smooth, non-negative Fourier-space form, making it a ready input for Color Glass Condensate calculations.

Core claim

On the paper's own terms, the discovery is that a physics-informed neural network can act as a differentiable surrogate for N(r,xB), with the ciBK equation imposed as a loss penalty rather than through a parametric initial condition; the starting profile at x0=0.03 is learned from data, not assumed. With penalties enforcing black-disk saturation, color-transparency slope bounds, and Fourier positivity of the momentum-space dipole, the best solution simultaneously fits the total and charm reduced cross sections and exclusive J/psi photoproduction, and it matches a direct numerical ciBK evolution from the same initial condition. The conclusion: the functional flexibility that rigid analytic an

What carries the argument

The PINN itself is the central object: a residual neural network with softplus hidden layers and a sigmoid output maps (r, Y=ln(x0/xB)) to N in (0,1). Training minimizes a weighted loss with three parts: the ciBK integro-differential residual, a negative log-likelihood data term over the three observable classes, and physical penalties. The Fourier-positivity term penalizes negative values of S(kT,Y)=∫d2r e^{ikT·r}(1-N(r,Y)) and carries the largest physical weight, so it is the mechanism that produces the advertised non-negative momentum-space dipole. Automatic differentiation lets the network supply the Y-derivative in the evolution residual and the small-r logarithmic slope without a grid.

Load-bearing premise

The load-bearing premise is that the heavy penalty forcing the momentum-space dipole to be non-negative corresponds to a genuine physical requirement; if it is only a desirable property, the constraint distorts the extracted amplitude and the advertised smooth, positive S(kT) is guaranteed by construction rather than by the data.

What would settle it

Retrain the same global fit with the Fourier-positivity penalty switched off (lambda_pos=0). If the best-fit S(kT,xB) goes negative inside kT in [0,100] GeV while the inclusive, charm, and J/psi descriptions stay essentially unchanged, then the paper's non-negativity result is an artifact of the penalty; if the data alone keep S non-negative, the positivity requirement is genuine.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • One extracted N(r,xB) can replace separate fitted amplitudes for inclusive DIS, charm, and exclusive J/psi, simplifying CGC global analyses.
  • Because the ansatz is nonparametric, tensions between total and charm channels are attributed to the functional rigidity of standard initial conditions rather than to the evolution dynamics.
  • The smooth, non-negative S(kT,xB) over kT up to 100 GeV provides a stable input for momentum-space CGC computations of particle production.
  • The trained surrogate evaluates N at any xB in a single forward pass, avoiding repeated grid-based BK evolution in later fits.
  • Extending the same scheme to impact-parameter dependent amplitudes and NLO impact factors is the stated path toward next-generation electron-ion collider constraints.

Where Pith is reading between the lines

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

  • Beyond the paper: if Fourier positivity is not an exact requirement in the impact-parameter-independent leading-order setup, the very large positivity weight (10000) could bias the extracted N; an ablation with that penalty removed would show whether data alone demand a non-negative S(kT).
  • Beyond the paper: the same universal N should predict other exclusive channels (rho, phi) and diffractive structure functions without retuning; those are cheap tests of the universality claim.
  • Beyond the paper: the method transfers naturally to nuclear targets, where the initial profile is even less constrained and could be inferred from e+A data at a future electron-ion collider.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a physics-informed neural network (PINN) extraction of the universal small-x dipole amplitude N(r,x_B), constrained by the collinearly improved Balitsky–Kovchegov (ciBK) evolution equation and by a global fit to HERA reduced total and charm cross sections and to exclusive J/psi photoproduction. A momentum-space positivity penalty on the Fourier transform S(k_T,Y) is imposed, and physical parameters (R_p, alpha_fr, K_VM) are fitted simultaneously with the network weights. The authors report chi^2/dof = 1.150 for sigma_r, 1.546 for charm, and 0.474 for J/psi, an out-of-fit F_L comparison with chi^2/dof = 0.733, consistency with numerical ciBK evolution, and a smooth, non-negative S(k_T,x_B) for x_B <= 0.01 and k_T in [0,100] GeV. The extracted N is compared to an MV-type parametrization and found not to be faithfully captured by that functional form. Tabulated values are made publicly available.

Significance. If the result holds, this is a useful contribution to small-x phenomenology: it provides a flexible, data-constrained dipole amplitude that could serve as an input to a wide class of CGC calculations and offers a path around the total/charm tension of conventional parametric fits. The main strengths are the simultaneous description of several observables with a single N, the independent F_L cross-check, the consistency check against numerical ciBK evolution, and the public release of tabulated results. However, the advertised momentum-space positivity is enforced by a large penalty rather than shown to be emergent, and the 'nonparametric' nature of the extraction is qualified by network architecture and hand-set loss weights. These issues affect the strength of the central claims but are addressable with additional analysis.

major comments (3)
  1. [Eq. (S3) and Abstract] The claim that the extracted S(k_T,x_B) is 'smooth, non-negative' is guaranteed by construction: Eq. (S3) contains lambda_pos ||min(0,S)||^2 with lambda_pos = 10^4. This penalty drives S>=0 over the fitted domain, so the result is not evidence for a physical property unless Fourier-positivity is actually required in the impact-parameter-independent LO framework. The cited Refs. [21,25,31] are not derived here, and no formal argument is given that the 2D Fourier transform of [1-N(r,Y)] must be non-negative for a unitarity- and color-transparency-respecting N. Please provide either a derivation of the positivity condition in this framework or an ablation study with lambda_pos = 0 (or a substantially reduced value), reporting how chi^2 and the extracted N change. Without such a test, the paper's 'central outcome' is a tautology and the claimed alleviation of the total/charm tension could be
  2. [Fig. 4, 'Results and Discussion'] The abstract and conclusion state that S(k_T,x_B) is non-negative 'within the fitted kinematic domain' and for x_B <= 0.01. However, the global fit is explicitly restricted to experimental data with x_B < 0.004 (section 'Physics-informed global analysis'). Figure 4 displays x_B = 10^-2, which lies between x_0 = 0.03 and the data-constrained region, and the conclusion extends the positivity claim to x_B <= 0.01. This is an extrapolated region, not part of the fitted domain. The authors should either restrict the positivity claim to x_B < 0.004 or provide evidence that the extrapolation to x_B = 0.01 is trustworthy (e.g., by showing that the positivity persists independent of the Y<Y0 boundary choice and of the penalty strength).
  3. [Section 'Results and Discussion'] The central claim that the fit 'alleviates the long-standing tension between total and charm channels' is based on a simultaneous fit in which sigma_r and charm are both in L_data. Agreement with fitted data is expected, so the statement is stronger than the evidence shown. To make the claim compelling, the authors should include a control: for example, a fit with the same PINN framework but a conventional parametric initial condition (MV-type or equivalent), or a fit with lambda_pos = 0, to demonstrate that the improved description of charm is not merely a consequence of the network's flexibility or of the positivity penalty. The MV-type comparison in Fig. 3 is only a profile comparison at x_B = 0.01 and does not serve as a global control.
minor comments (5)
  1. [Abstract, 'nonparametric'] The term 'nonparametric' is used repeatedly, but the method depends on the ResNet architecture, activation functions, number of layers/neurons, and the chosen loss weights. Consider replacing 'nonparametric' with 'flexible' or 'neural-network-based' to avoid overclaiming.
  2. [Section 'Physics-informed global analysis'] The statement that starting at x_0 = 0.03 'reduces sensitivity' to the boundary condition for Y < Y_0 is plausible but not demonstrated. No test with different boundary choices or different x_0 is reported; please add a short sensitivity check.
  3. [Fig. 2] The quoted chi^2/dof = 0.474 for J/psi is remarkably low, which can indicate overfitting or underestimated uncertainties. Please report the number of data points, the number of fitted parameters, and, if possible, a breakdown by experiment to reassure the reader.
  4. [Section 'Results and Discussion'] In the sentence 'Recent analyses [37,38] emphasize that charm production...', the text says 'the short-distance behavior and on the profile entering the evolution' but does not specify which profile; clarify.
  5. [Supplemental Material] There are minor language issues, e.g., 'the the Supplemental Material' in the main text and 'This Supplementary Material section provides the details about a a' in the supplement. A careful proofread is recommended.

Circularity Check

0 steps flagged

No significant circularity: the extraction is a transparent global fit with explicitly imposed constraints, and the claimed consistency is fit quality rather than a hidden prediction or self-referential derivation.

full rationale

The paper's derivation chain is a constrained global fit, not a claim to derive its output from first principles alone. The dipole amplitude N(r,Y) is a trainable surrogate whose loss explicitly combines ciBK residuals, data likelihoods, and physical penalties (Eqs. S1–S3). The data terms in L_data include all three observables used in the headline comparisons (σ_r, σ_ccbar_r, σ_VM), so the quoted χ² values are honestly fit qualities, not out-of-sample predictions; the paper even says "all fitted observables" in the abstract. The J/ψ comparison is therefore not a circular prediction: fitting the same N to all channels can fail, and the claim of universality is that one amplitude accommodates all of them. The positivity property is imposed by the explicit term λ_pos‖min(0,S)‖² with λ_pos=10000 in Eq. S3, and the text states plainly: "In our PINN global fit we impose a positivity regularization by penalizing regions where S(k_T,x_B)<0." The subsequent statement that S is non-negative is a constraint-satisfaction result, not an emergent prediction, and the conditional wording ("provided sufficient functional flexibility and an explicit positivity regularization are included") explicitly disclaims emergence. There is also an external cross-check: the longitudinal structure function F_L comparison in Fig. S3 uses data outside the fitted observable set O ∈ {σ_r, σ_ccbar_r, σ_VM}, and the numerical ciBK evolution check verifies solver consistency using the same initial condition. No load-bearing argument reduces to a self-citation; the ciBK formalism is cited from the established literature, and no uniqueness theorem is imported from the authors' prior work. The main scientific caveat — that the positivity penalty is very strong and no λ_pos=0 ablation is shown — is a correctness/bias concern about whether the constraint is physical, not a circularity in the derivation. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No fundamentally new physical entities are introduced. The model adds fitted parameters (R_p, K_VM, alpha_fr, network weights, loss weights, x0) and several domain assumptions. The main caveat is that 'nonparametric' is relative: the network architecture and the strong Fourier-positivity penalty are implicit priors that shape the solution.

free parameters (6)
  • R_p (effective proton transverse size) = 0.809 +/- 0.010 fm
    Controls the overall normalization of DIS and J/psi total cross sections through the proton transverse area; fitted simultaneously with the network.
  • K_VM factor = 0.902 +/- 0.008
    Multiplicative normalization for J/psi wave-function uncertainties; fitted to exclusive data, so the J/psi comparison is not fully parameter-free.
  • alpha_fr (infrared-frozen strong coupling) = 0.702 +/- 0.070
    Sets the IR behavior of the running coupling in the ciBK kernel; a free parameter that directly affects evolution speed and the small-r behavior of N.
  • Network weights theta = not reported (tens of thousands)
    The full functional form of N is determined by these trained parameters; the 'nonparametric' extraction is a high-dimensional parametric fit with implicit smoothness priors.
  • Loss and regularization weights = w = (50000,100,1); lambda_data=(1,1.2,1); lambda_IR,pos,UV=(1,10000,2000)
    Hand-chosen weights control the balance between physics residual, data, and constraints; central outcomes such as positivity and ciBK enforcement are sensitive to these choices.
  • x0 starting rapidity = 0.03
    Boundary/initial-evolution point chosen by hand; the nonlocal rapidity shift requires a boundary assumption for Y<ln(1/x0), and sensitivity is stated to be reduced but not quantified.
axioms (6)
  • domain assumption The ciBK evolution equation with the specific collinearly improved kernel and running-coupling prescription is the correct evolution for N in the fitted domain.
    Invoked in Eq. (1) and throughout; assumes the large-Nc/mean-field limit and the resummation scheme of Refs [25,48,51].
  • domain assumption Impact-parameter-independent dipole amplitude with an effective proton area sigma0/2 is sufficient.
    Used for sigma_r, charm, and J/psi; replaces the impact-parameter integral with R_p-dependent area; stated in Supplemental B.
  • domain assumption Leading-order photon and vector-meson impact factors with effective quark masses are adequate in the fitted xB, Q2 region.
    Used in Eqs. (S7)-(S9) and (S14)-(S15); missing NLO corrections are acknowledged as a limitation and motivate the xB<=0.004 cut.
  • domain assumption Fourier positivity of S(kT,Y) is a required physical constraint.
    Penalized in Eq. (S3); if this is only an artifact of the impact-parameter-independent LO approximation, the regularization distorts the extracted N.
  • ad hoc to paper The boundary condition for Y<Y0 in the nonlocal rapidity-shift kernel has negligible effect when starting at x0=0.03.
    The paper states common frozen/vanishing choices after Eq. (1) and chooses x0=0.03 to reduce sensitivity, but no quantitative sensitivity scan is shown.
  • domain assumption The ResNet surrogate's smoothness and expressiveness are a sufficiently unbiased prior for the true N.
    Seven residual blocks of 32 neurons with softplus and sigmoid activations are used; no architecture ablation is shown, so the claimed advantage over MV-type fits could partly reflect network bias.

pith-pipeline@v1.3.0-alltime-deepseek · 15488 in / 16496 out tokens · 167337 ms · 2026-08-02T18:33:45.858005+00:00 · methodology

0 comments
read the original abstract

We extract the universal small-$x$ dipole scattering amplitude $N(r,x_B)$ from a global analysis based on a physics-informed neural network (PINN), without imposing a priori MV-type parametrization of the initial condition. The network provides a smooth and differentiable surrogate for $N(r,x_B)$, whose rapidity dependence is constrained by the collinearly improved Balitsky--Kovchegov evolution equation, while its functional form is simultaneously constrained by Deep Inelastic Scattering (DIS) data for the reduced total and charm cross sections, exclusive $J/\psi$ photoproduction measurements, and a positivity requirement for the momentum-space dipole amplitude. The resulting single universal amplitude consistently describes all fitted observables within a unified framework, alleviating the long-standing tension between total and charm channels encountered in conventional small-$x$ fits based on rigid parametric ans\"atze. Within the fitted kinematic domain, the best extracted PINN solution yields a smooth, non-negative momentum-space dipole over the full transverse-momentum range examined. Our results provide a robust and well-behaved input for Color Glass Condensate phenomenology across a broad class of high-energy processes.

discussion (0)

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

Forward citations

Cited by 6 Pith papers

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

  1. Sub-eikonal stress and model dependence of the small-$x$ gluon D-term

    hep-ph 2026-07 unverdicted novelty 7.0

    The gluon D-term at small x is a next-to-eikonal stress observable whose sign is not determined by the dipole or saturation profile.

  2. CLVisc Agent for autonomous relativistic hydrodynamics studies

    nucl-th 2026-07 conditional novelty 6.0

    An LLM agent autonomously created a CLVisc skill and ran two hydrodynamic studies, finding that the high-temperature branch of η/s dominates flow suppression and that PGCM-uniform 16O decouples ellipticity from size.

  3. Unbiased Data-Driven Determination of the Nuclear Dipole Amplitude in the Color Glass Condensate

    hep-ph 2026-07 conditional novelty 6.0

    The 208Pb dipole amplitude is learned from R_pPb and coherent J/ψ photoproduction data with the BK equation embedded in training, giving Q²_s0(Pb)/Q²_s0(p) = 3.17 and an MV-type initial condition.

  4. Simultaneous Color Glass Condensate fit to deep inelastic scattering and forward hadron production at HERA, RHIC, and the LHC

    hep-ph 2026-07 conditional novelty 6.0

    A simultaneous LO CGC/BK fit to HERA DIS and RHIC/LHC forward hadron production reaches χ²/dof≈1, with complementary constraints and RHIC K-factors roughly twice those at the LHC.

  5. Sub-eikonal stress and model dependence of the small-$x$ gluon D-term

    hep-ph 2026-07 conditional novelty 6.0

    The small-x gluon D-term is a next-to-eikonal stress probe and is not fixed by the leading-eikonal dipole or saturation profile alone.

  6. Confronting Color Glass Condensate at next-to-leading order with HERA data

    hep-ph 2026-04 unverdicted novelty 6.0

    A Bayesian global fit at full NLO+NLL accuracy extracts the posterior distribution for the non-perturbative initial condition of the NLO Balitsky-Kovchegov equation from HERA inclusive and charm data.

Reference graph

Works this paper leans on

83 extracted references · 1 canonical work pages · cited by 5 Pith papers

  1. [1]

    Y. V. Kovchegov and E. Levin,Quantum Chromodynam- ics at High Energy, vol. 33 (Oxford University Press, 2013), ISBN 978-1-009-29144-6, 978-1-009-29141-5, 978- 1-009-29142-2, 978-0-521-11257-4, 978-1-139-55768-9

    2013

  2. [2]

    McLerran, Nucl

    L. McLerran, Nucl. Phys. A702, 49 (2002)

    2002

  3. [3]

    Iancu and R

    E. Iancu and R. Venugopalan,The Color glass conden- sate and high-energy scattering in QCD(2003), pp. 249– 3363, hep-ph/0303204

    Pith/arXiv arXiv 2003

  4. [4]

    Weigert, Prog

    H. Weigert, Prog. Part. Nucl. Phys.55, 461 (2005), hep- ph/0501087

    arXiv 2005

  5. [5]

    Gelis, E

    F. Gelis, E. Iancu, J. Jalilian-Marian, and R. Venu- gopalan, Ann. Rev. Nucl. Part. Sci.60, 463 (2010), 1002.0333

    Pith/arXiv arXiv 2010

  6. [6]

    J. L. Albacete and C. Marquet, Prog. Part. Nucl. Phys. 76, 1 (2014), 1401.4866

    Pith/arXiv arXiv 2014

  7. [7]

    Morreale and F

    A. Morreale and F. Salazar, Universe7, 312 (2021), 2108.08254

    Pith/arXiv arXiv 2021

  8. [8]

    Jalilian-Marian, A

    J. Jalilian-Marian, A. Kovner, L. D. McLerran, and H. Weigert, Phys. Rev. D55, 5414 (1997), hep- ph/9606337

    arXiv 1997

  9. [9]

    Jalilian-Marian, A

    J. Jalilian-Marian, A. Kovner, A. Leonidov, and H. Weigert, Nucl. Phys. B504, 415 (1997), hep- ph/9701284

    arXiv 1997

  10. [10]

    Jalilian-Marian, A

    J. Jalilian-Marian, A. Kovner, A. Leonidov, and H. Weigert, Phys. Rev. D59, 014014 (1998), hep- ph/9706377

    arXiv 1998

  11. [11]

    Iancu and L

    E. Iancu and L. D. McLerran, Phys. Lett. B510, 145 (2001), hep-ph/0103032

    Pith/arXiv arXiv 2001

  12. [12]

    Ferreiro, E

    E. Ferreiro, E. Iancu, A. Leonidov, and L. McLerran, Nucl. Phys. A703, 489 (2002), hep-ph/0109115

    Pith/arXiv arXiv 2002

  13. [13]

    Iancu, A

    E. Iancu, A. Leonidov, and L. D. McLerran, Phys. Lett. B510, 133 (2001), hep-ph/0102009. 6

    Pith/arXiv arXiv 2001

  14. [14]

    Iancu, A

    E. Iancu, A. Leonidov, and L. D. McLerran, Nucl. Phys. A692, 583 (2001), hep-ph/0011241

    Pith/arXiv arXiv 2001

  15. [15]

    Balitsky, Nucl

    I. Balitsky, Nucl. Phys. B463, 99 (1996), hep- ph/9509348

    arXiv 1996

  16. [16]

    Y. V. Kovchegov, Phys. Rev. D60, 034008 (1999), hep- ph/9901281

    arXiv 1999

  17. [17]

    Kowalski, T

    H. Kowalski, T. Lappi, and R. Venugopalan, Phys. Rev. Lett.100, 022303 (2008), 0705.3047

    Pith/arXiv arXiv 2008

  18. [18]

    Abdul Khalek et al., Nucl

    R. Abdul Khalek et al., Nucl. Phys. A1026, 122447 (2022), 2103.05419

    Pith/arXiv arXiv 2022

  19. [19]

    E. C. Aschenauer, S. Fazio, J. H. Lee, H. Mantysaari, B. S. Page, B. Schenke, T. Ullrich, R. Venugopalan, and P. Zurita, Rept. Prog. Phys.82, 024301 (2019), 1708.01527

    Pith/arXiv arXiv 2019

  20. [20]

    Accardi et al., Eur

    A. Accardi et al., Eur. Phys. J. A52, 268 (2016), 1212.1701

    Pith/arXiv arXiv 2016

  21. [21]

    J. L. Albacete, N. Armesto, J. G. Milhano, and C. A. Salgado, Phys. Rev. D80, 034031 (2009), 0902.1112

    Pith/arXiv arXiv 2009

  22. [22]

    J. L. Albacete, N. Armesto, J. G. Milhano, P. Quiroga- Arias, and C. A. Salgado, Eur. Phys. J. C71, 1705 (2011), 1012.4408

    Pith/arXiv arXiv 2011

  23. [23]

    Lappi and H

    T. Lappi and H. M¨ antysaari, Phys. Rev. D88, 114020 (2013), 1309.6963

    Pith/arXiv arXiv 2013

  24. [24]

    M¨ antysaari and B

    H. M¨ antysaari and B. Schenke, Phys. Rev. D98, 034013 (2018), 1806.06783

    Pith/arXiv arXiv 2018

  25. [25]

    Duclou´ e, E

    B. Duclou´ e, E. Iancu, G. Soyez, and D. N. Triantafyl- lopoulos, Phys. Lett. B803, 135305 (2020), 1912.09196

    Pith/arXiv arXiv 2020

  26. [26]

    M¨ antysaari, B

    H. M¨ antysaari, B. Schenke, C. Shen, and W. Zhao, Phys. Lett. B833, 137348 (2022), 2202.01998

    Pith/arXiv arXiv 2022

  27. [27]

    M¨ antysaari and J

    H. M¨ antysaari and J. Penttala, JHEP08, 247 (2022), 2204.14031

    Pith/arXiv arXiv 2022

  28. [28]

    M¨ antysaari, J

    H. M¨ antysaari, J. Penttala, F. Salazar, and B. Schenke, Phys. Rev. D111, 054033 (2025), 2411.13533

    Pith/arXiv arXiv 2025

  29. [29]

    Hentschinski and R

    M. Hentschinski and R. R. Ram ´ ırez, Phys. Rev. D113, 014043 (2026), 2508.11545

    Pith/arXiv arXiv 2026

  30. [30]

    M¨ antysaari, H

    H. M¨ antysaari, H. Roch, F. Salazar, B. Schenke, C. Shen, and W. Zhao, Phys. Rev. D113, 014038 (2026), 2507.14087

    arXiv 2026

  31. [31]

    B. G. Giraud and R. Peschanski, Phys. Lett. B760, 26 (2016), 1604.01932

    Pith/arXiv arXiv 2016

  32. [32]

    Gelis and R

    F. Gelis and R. Venugopalan, Phys. Rev. D69, 014019 (2004), hep-ph/0310090

    Pith/arXiv arXiv 2004

  33. [33]

    J. P. Blaizot, F. Gelis, and R. Venugopalan, Nucl. Phys. A743, 13 (2004), hep-ph/0402256

    Pith/arXiv arXiv 2004

  34. [34]

    J. P. Blaizot, F. Gelis, and R. Venugopalan, Nucl. Phys. A743, 57 (2004), hep-ph/0402257

    Pith/arXiv arXiv 2004

  35. [35]

    Dumitru, A

    A. Dumitru, A. Hayashigaki, and J. Jalilian-Marian, Nucl. Phys. A765, 464 (2006), hep-ph/0506308

    Pith/arXiv arXiv 2006

  36. [36]

    Caucal, Z.-B

    P. Caucal, Z.-B. Kang, P. Korcyl, F. Salazar, B. Schenke, T. Stebel, R. Venugopalan, and W. Zhao (2025), 2512.21466

    arXiv 2025

  37. [37]

    H¨ anninen, H

    H. H¨ anninen, H. M¨ antysaari, R. Paatelainen, and J. Penttala, Phys. Rev. Lett.130, 192301 (2023), 2211.03504

    Pith/arXiv arXiv 2023

  38. [38]

    Casuga, H

    C. Casuga, H. H¨ anninen, and H. M¨ antysaari, Phys. Rev. D112, 034003 (2025), 2506.00487

    Pith/arXiv arXiv 2025

  39. [39]

    Casuga, M

    C. Casuga, M. Karhunen, and H. M¨ antysaari, Phys. Rev. D109, 054018 (2024), 2311.10491

    Pith/arXiv arXiv 2024

  40. [40]

    Gao, Z.-B

    M. Gao, Z.-B. Kang, J. Penttala, and D. Y. Shao (2025), 2510.26779

    arXiv 2025

  41. [41]

    Kou and X

    W. Kou and X. Chen (2026), 2601.16391

    Pith/arXiv arXiv 2026

  42. [42]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, J. Com- put. Phys.378, 686 (2019), 1711.10561

    Pith/arXiv arXiv 2019

  43. [43]

    von Rueden, S

    L. von Rueden, S. Mayer, K. Beckh, B. Georgiev, S. Gies- selbach, R. Heese, B. Kirsch, J. Pfrommer, A. Pick, R. Ramamurthy, et al., IEEE Transactions on Knowl- edge and Data Engineering35, 614 (2023)

    2023

  44. [44]

    Cuomo, V

    S. Cuomo, V. S. D. Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli, CoRRabs/2201.05624 (2022), 2201.05624

    Pith/arXiv arXiv 2022

  45. [45]

    Beuf, Phys

    G. Beuf, Phys. Rev. D89, 074039 (2014), 1401.0313

    Pith/arXiv arXiv 2014

  46. [46]

    Iancu, J

    E. Iancu, J. D. Madrigal, A. H. Mueller, G. Soyez, and D. N. Triantafyllopoulos, Phys. Lett. B744, 293 (2015), 1502.05642

    Pith/arXiv arXiv 2015

  47. [47]

    Iancu, J

    E. Iancu, J. D. Madrigal, A. H. Mueller, G. Soyez, and D. N. Triantafyllopoulos, Phys. Lett. B750, 643 (2015), 1507.03651

    Pith/arXiv arXiv 2015

  48. [48]

    Duclou´ e, E

    B. Duclou´ e, E. Iancu, A. H. Mueller, G. Soyez, and D. N. Triantafyllopoulos, JHEP04, 081 (2019), 1902.06637

    Pith/arXiv arXiv 2019

  49. [49]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell,Sim- ple and scalable predictive uncertainty estimation using deep ensembles(2017), 1612.01474

    Pith/arXiv arXiv 2017

  50. [50]

    Iancu, J

    E. Iancu, J. D. Madrigal, A. H. Mueller, G. Soyez, and D. N. Triantafyllopoulos, Nucl. Phys. A956, 557 (2016), 1601.06525

    Pith/arXiv arXiv 2016

  51. [51]

    Iancu, A

    E. Iancu, A. H. Mueller, D. N. Triantafyllopoulos, and S. Y. Wei, JHEP07, 196 (2021), 2012.08562

    Pith/arXiv arXiv 2021

  52. [52]

    J. L. Albacete, J. G. Milhano, P. Quiroga-Arias, and J. Rojo, Eur. Phys. J. C72, 2131 (2012), 1203.1043

    Pith/arXiv arXiv 2012

  53. [53]

    Agostini et al

    P. Agostini et al. (LHeC, FCC-he Study Group), J. Phys. G48, 110501 (2021), 2007.14491

    arXiv 2021

  54. [54]

    K. J. Golec-Biernat and M. Wusthoff, Phys. Rev. D59, 014017 (1998), hep-ph/9807513

    Pith/arXiv arXiv 1998

  55. [55]

    Kowalski and D

    H. Kowalski and D. Teaney, Phys. Rev. D68, 114005 (2003), hep-ph/0304189

    Pith/arXiv arXiv 2003

  56. [56]

    J. R. Forshaw, R. Sandapen, and G. Shaw, Phys. Rev. D 69, 094013 (2004), hep-ph/0312172

    Pith/arXiv arXiv 2004

  57. [57]

    Kowalski, L

    H. Kowalski, L. Motyka, and G. Watt, Phys. Rev. D74, 074016 (2006), hep-ph/0606272

    Pith/arXiv arXiv 2006

  58. [58]

    A. G. Shuvaev, K. J. Golec-Biernat, A. D. Martin, and M. G. Ryskin, Phys. Rev. D60, 014015 (1999), hep- ph/9902410

    arXiv 1999

  59. [59]

    Lappi and H

    T. Lappi and H. Mantysaari, Phys. Rev. C87, 032201 (2013), 1301.4095

    Pith/arXiv arXiv 2013

  60. [60]

    Abramowicz et al

    H. Abramowicz et al. (H1, ZEUS), Eur. Phys. J. C75, 580 (2015), 1506.06042

    Pith/arXiv arXiv 2015

  61. [61]

    F. D. Aaron et al. (H1, ZEUS), JHEP01, 109 (2010), 0911.0884

    Pith/arXiv arXiv 2010

  62. [62]

    Abramowicz et al

    H. Abramowicz et al. (H1, ZEUS), Eur. Phys. J. C78, 473 (2018), 1804.01019

    Pith/arXiv arXiv 2018

  63. [63]

    B. B. Abelev et al. (ALICE), Phys. Rev. Lett.113, 232504 (2014), 1406.7819

    Pith/arXiv arXiv 2014

  64. [64]

    Acharya et al

    S. Acharya et al. (ALICE), Eur. Phys. J. C79, 402 (2019), 1809.03235

    Pith/arXiv arXiv 2019

  65. [65]

    Aktas et al

    A. Aktas et al. (H1), Eur. Phys. J. C46, 585 (2006), hep-ex/0510016

    Pith/arXiv arXiv 2006

  66. [66]

    Alexa et al

    C. Alexa et al. (H1), Eur. Phys. J. C73, 2466 (2013), 1304.5162

    Pith/arXiv arXiv 2013

  67. [67]

    Chekanov et al

    S. Chekanov et al. (ZEUS), Eur. Phys. J. C24, 345 (2002), hep-ex/0201043

    Pith/arXiv arXiv 2002

  68. [68]

    Aaij et al

    R. Aaij et al. (LHCb), JHEP10, 167 (2018), 1806.04079

    Pith/arXiv arXiv 2018

  69. [69]

    Aaij et al

    R. Aaij et al. (LHCb), SciPost Phys.18, 071 (2025), 2409.03496

    arXiv 2025

  70. [70]

    J. L. Albacete, N. Armesto, J. G. Milhano, C. A. Salgado, and U. A. Wiedemann, Phys. Rev. D71, 014003 (2005), hep-ph/0408216

    Pith/arXiv arXiv 2005

  71. [71]

    Marquet, R

    C. Marquet, R. B. Peschanski, and G. Soyez, Phys. Lett. B628, 239 (2005), hep-ph/0509074. 7

    Pith/arXiv arXiv 2005

  72. [72]

    Munier and R

    S. Munier and R. B. Peschanski, Phys. Rev. Lett.91, 232001 (2003), hep-ph/0309177

    Pith/arXiv arXiv 2003

  73. [73]

    Loshchilov and F

    I. Loshchilov and F. Hutter, inInternational Conference on Learning Representations(2019)

    2019

  74. [74]

    Loshchilov and F

    I. Loshchilov and F. Hutter, CoRRabs/1608.03983 (2016), 1608.03983

    Pith/arXiv arXiv 2016

  75. [75]

    Y. Wang, Y. Yao, J. Guo, and Z. Gao, Journal of Com- putational Physics510, 113112 (2024), ISSN 0021-9991

    2024

  76. [76]

    S. Wang, X. Yu, and P. Perdikaris, Journal of Computa- tional Physics449, 110768 (2022), ISSN 0021-9991

    2022

  77. [77]

    L. D. McLerran and R. Venugopalan, Phys. Rev. D49, 2233 (1994), hep-ph/9309289

    Pith/arXiv arXiv 1994

  78. [78]

    L. D. McLerran and R. Venugopalan, Phys. Rev. D49, 3352 (1994), hep-ph/9311205

    Pith/arXiv arXiv 1994

  79. [79]

    S. Dai, F. Li, L.-G. Pang, G. Qin, S.-y. Wei, H. Zhang, and W. Zhao,Physics-informed global extraction of the universal small-xdipole amplitude(2026), URLhttps: //doi.org/10.5281/zenodo.18826120

    work page doi:10.5281/zenodo.18826120 2026

  80. [80]

    K. He, X. Zhang, S. Ren, and J. Sun, CoRR abs/1512.03385(2015), 1512.03385

    Pith/arXiv arXiv 2015

Showing first 80 references.